qemu-e2k/accel/tcg/translate-all.c

2633 lines
76 KiB
C
Raw Normal View History

/*
* Host code generation
*
* Copyright (c) 2003 Fabrice Bellard
*
* This library is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2 of the License, or (at your option) any later version.
*
* This library is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with this library; if not, see <http://www.gnu.org/licenses/>.
*/
#ifdef _WIN32
#include <windows.h>
#endif
#include "qemu/osdep.h"
#include "qemu-common.h"
#define NO_CPU_IO_DEFS
#include "cpu.h"
#include "trace.h"
#include "disas/disas.h"
#include "exec/exec-all.h"
#include "tcg.h"
#if defined(CONFIG_USER_ONLY)
#include "qemu.h"
#if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
#include <sys/param.h>
#if __FreeBSD_version >= 700104
#define HAVE_KINFO_GETVMMAP
#define sigqueue sigqueue_freebsd /* avoid redefinition */
#include <sys/proc.h>
#include <machine/profile.h>
#define _KERNEL
#include <sys/user.h>
#undef _KERNEL
#undef sigqueue
#include <libutil.h>
#endif
#endif
#else
#include "exec/ram_addr.h"
#endif
#include "exec/cputlb.h"
#include "exec/tb-hash.h"
#include "translate-all.h"
#include "qemu/bitmap.h"
#include "qemu/error-report.h"
#include "qemu/timer.h"
tcg: drop global lock during TCG code execution This finally allows TCG to benefit from the iothread introduction: Drop the global mutex while running pure TCG CPU code. Reacquire the lock when entering MMIO or PIO emulation, or when leaving the TCG loop. We have to revert a few optimization for the current TCG threading model, namely kicking the TCG thread in qemu_mutex_lock_iothread and not kicking it in qemu_cpu_kick. We also need to disable RAM block reordering until we have a more efficient locking mechanism at hand. Still, a Linux x86 UP guest and my Musicpal ARM model boot fine here. These numbers demonstrate where we gain something: 20338 jan 20 0 331m 75m 6904 R 99 0.9 0:50.95 qemu-system-arm 20337 jan 20 0 331m 75m 6904 S 20 0.9 0:26.50 qemu-system-arm The guest CPU was fully loaded, but the iothread could still run mostly independent on a second core. Without the patch we don't get beyond 32206 jan 20 0 330m 73m 7036 R 82 0.9 1:06.00 qemu-system-arm 32204 jan 20 0 330m 73m 7036 S 21 0.9 0:17.03 qemu-system-arm We don't benefit significantly, though, when the guest is not fully loading a host CPU. Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com> Message-Id: <1439220437-23957-10-git-send-email-fred.konrad@greensocs.com> [FK: Rebase, fix qemu_devices_reset deadlock, rm address_space_* mutex] Signed-off-by: KONRAD Frederic <fred.konrad@greensocs.com> [EGC: fixed iothread lock for cpu-exec IRQ handling] Signed-off-by: Emilio G. Cota <cota@braap.org> [AJB: -smp single-threaded fix, clean commit msg, BQL fixes] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Pranith Kumar <bobby.prani@gmail.com> [PM: target-arm changes] Acked-by: Peter Maydell <peter.maydell@linaro.org>
2017-02-23 19:29:11 +01:00
#include "qemu/main-loop.h"
#include "exec/log.h"
#include "sysemu/cpus.h"
/* #define DEBUG_TB_INVALIDATE */
/* #define DEBUG_TB_FLUSH */
/* make various TB consistency checks */
/* #define DEBUG_TB_CHECK */
#ifdef DEBUG_TB_INVALIDATE
#define DEBUG_TB_INVALIDATE_GATE 1
#else
#define DEBUG_TB_INVALIDATE_GATE 0
#endif
#ifdef DEBUG_TB_FLUSH
#define DEBUG_TB_FLUSH_GATE 1
#else
#define DEBUG_TB_FLUSH_GATE 0
#endif
#if !defined(CONFIG_USER_ONLY)
/* TB consistency checks only implemented for usermode emulation. */
#undef DEBUG_TB_CHECK
#endif
#ifdef DEBUG_TB_CHECK
#define DEBUG_TB_CHECK_GATE 1
#else
#define DEBUG_TB_CHECK_GATE 0
#endif
/* Access to the various translations structures need to be serialised via locks
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
* for consistency.
* In user-mode emulation access to the memory related structures are protected
* with mmap_lock.
* In !user-mode we use per-page locks.
*/
#ifdef CONFIG_SOFTMMU
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
#define assert_memory_lock()
#else
#define assert_memory_lock() tcg_debug_assert(have_mmap_lock())
#endif
#define SMC_BITMAP_USE_THRESHOLD 10
typedef struct PageDesc {
/* list of TBs intersecting this ram page */
uintptr_t first_tb;
#ifdef CONFIG_SOFTMMU
/* in order to optimize self modifying code, we count the number
of lookups we do to a given page to use a bitmap */
unsigned long *code_bitmap;
unsigned int code_write_count;
#else
unsigned long flags;
#endif
#ifndef CONFIG_USER_ONLY
QemuSpin lock;
#endif
} PageDesc;
/**
* struct page_entry - page descriptor entry
* @pd: pointer to the &struct PageDesc of the page this entry represents
* @index: page index of the page
* @locked: whether the page is locked
*
* This struct helps us keep track of the locked state of a page, without
* bloating &struct PageDesc.
*
* A page lock protects accesses to all fields of &struct PageDesc.
*
* See also: &struct page_collection.
*/
struct page_entry {
PageDesc *pd;
tb_page_addr_t index;
bool locked;
};
/**
* struct page_collection - tracks a set of pages (i.e. &struct page_entry's)
* @tree: Binary search tree (BST) of the pages, with key == page index
* @max: Pointer to the page in @tree with the highest page index
*
* To avoid deadlock we lock pages in ascending order of page index.
* When operating on a set of pages, we need to keep track of them so that
* we can lock them in order and also unlock them later. For this we collect
* pages (i.e. &struct page_entry's) in a binary search @tree. Given that the
* @tree implementation we use does not provide an O(1) operation to obtain the
* highest-ranked element, we use @max to keep track of the inserted page
* with the highest index. This is valuable because if a page is not in
* the tree and its index is higher than @max's, then we can lock it
* without breaking the locking order rule.
*
* Note on naming: 'struct page_set' would be shorter, but we already have a few
* page_set_*() helpers, so page_collection is used instead to avoid confusion.
*
* See also: page_collection_lock().
*/
struct page_collection {
GTree *tree;
struct page_entry *max;
};
/* list iterators for lists of tagged pointers in TranslationBlock */
#define TB_FOR_EACH_TAGGED(head, tb, n, field) \
for (n = (head) & 1, tb = (TranslationBlock *)((head) & ~1); \
tb; tb = (TranslationBlock *)tb->field[n], n = (uintptr_t)tb & 1, \
tb = (TranslationBlock *)((uintptr_t)tb & ~1))
#define PAGE_FOR_EACH_TB(pagedesc, tb, n) \
TB_FOR_EACH_TAGGED((pagedesc)->first_tb, tb, n, page_next)
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
#define TB_FOR_EACH_JMP(head_tb, tb, n) \
TB_FOR_EACH_TAGGED((head_tb)->jmp_list_head, tb, n, jmp_list_next)
/* In system mode we want L1_MAP to be based on ram offsets,
while in user mode we want it to be based on virtual addresses. */
#if !defined(CONFIG_USER_ONLY)
#if HOST_LONG_BITS < TARGET_PHYS_ADDR_SPACE_BITS
# define L1_MAP_ADDR_SPACE_BITS HOST_LONG_BITS
#else
# define L1_MAP_ADDR_SPACE_BITS TARGET_PHYS_ADDR_SPACE_BITS
#endif
#else
# define L1_MAP_ADDR_SPACE_BITS TARGET_VIRT_ADDR_SPACE_BITS
#endif
/* Size of the L2 (and L3, etc) page tables. */
#define V_L2_BITS 10
#define V_L2_SIZE (1 << V_L2_BITS)
/* Make sure all possible CPU event bits fit in tb->trace_vcpu_dstate */
QEMU_BUILD_BUG_ON(CPU_TRACE_DSTATE_MAX_EVENTS >
sizeof_field(TranslationBlock, trace_vcpu_dstate)
* BITS_PER_BYTE);
/*
* L1 Mapping properties
*/
static int v_l1_size;
static int v_l1_shift;
static int v_l2_levels;
/* The bottom level has pointers to PageDesc, and is indexed by
* anything from 4 to (V_L2_BITS + 3) bits, depending on target page size.
*/
#define V_L1_MIN_BITS 4
#define V_L1_MAX_BITS (V_L2_BITS + 3)
#define V_L1_MAX_SIZE (1 << V_L1_MAX_BITS)
static void *l1_map[V_L1_MAX_SIZE];
/* code generation context */
TCGContext tcg_init_ctx;
tcg: enable multiple TCG contexts in softmmu This enables parallel TCG code generation. However, we do not take advantage of it yet since tb_lock is still held during tb_gen_code. In user-mode we use a single TCG context; see the documentation added to tcg_region_init for the rationale. Note that targets do not need any conversion: targets initialize a TCGContext (e.g. defining TCG globals), and after this initialization has finished, the context is cloned by the vCPU threads, each of them keeping a separate copy. TCG threads claim one entry in tcg_ctxs[] by atomically increasing n_tcg_ctxs. Do not be too annoyed by the subsequent atomic_read's of that variable and tcg_ctxs; they are there just to play nice with analysis tools such as thread sanitizer. Note that we do not allocate an array of contexts (we allocate an array of pointers instead) because when tcg_context_init is called, we do not know yet how many contexts we'll use since the bool behind qemu_tcg_mttcg_enabled() isn't set yet. Previous patches folded some TCG globals into TCGContext. The non-const globals remaining are only set at init time, i.e. before the TCG threads are spawned. Here is a list of these set-at-init-time globals under tcg/: Only written by tcg_context_init: - indirect_reg_alloc_order - tcg_op_defs Only written by tcg_target_init (called from tcg_context_init): - tcg_target_available_regs - tcg_target_call_clobber_regs - arm: arm_arch, use_idiv_instructions - i386: have_cmov, have_bmi1, have_bmi2, have_lzcnt, have_movbe, have_popcnt - mips: use_movnz_instructions, use_mips32_instructions, use_mips32r2_instructions, got_sigill (tcg_target_detect_isa) - ppc: have_isa_2_06, have_isa_3_00, tb_ret_addr - s390: tb_ret_addr, s390_facilities - sparc: qemu_ld_trampoline, qemu_st_trampoline (build_trampolines), use_vis3_instructions Only written by tcg_prologue_init: - 'struct jit_code_entry one_entry' - aarch64: tb_ret_addr - arm: tb_ret_addr - i386: tb_ret_addr, guest_base_flags - ia64: tb_ret_addr - mips: tb_ret_addr, bswap32_addr, bswap32u_addr, bswap64_addr Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-20 00:57:58 +02:00
__thread TCGContext *tcg_ctx;
TBContext tb_ctx;
bool parallel_cpus;
static void page_table_config_init(void)
{
uint32_t v_l1_bits;
assert(TARGET_PAGE_BITS);
/* The bits remaining after N lower levels of page tables. */
v_l1_bits = (L1_MAP_ADDR_SPACE_BITS - TARGET_PAGE_BITS) % V_L2_BITS;
if (v_l1_bits < V_L1_MIN_BITS) {
v_l1_bits += V_L2_BITS;
}
v_l1_size = 1 << v_l1_bits;
v_l1_shift = L1_MAP_ADDR_SPACE_BITS - TARGET_PAGE_BITS - v_l1_bits;
v_l2_levels = v_l1_shift / V_L2_BITS - 1;
assert(v_l1_bits <= V_L1_MAX_BITS);
assert(v_l1_shift % V_L2_BITS == 0);
assert(v_l2_levels >= 0);
}
void cpu_gen_init(void)
{
tcg_context_init(&tcg_init_ctx);
}
/* Encode VAL as a signed leb128 sequence at P.
Return P incremented past the encoded value. */
static uint8_t *encode_sleb128(uint8_t *p, target_long val)
{
int more, byte;
do {
byte = val & 0x7f;
val >>= 7;
more = !((val == 0 && (byte & 0x40) == 0)
|| (val == -1 && (byte & 0x40) != 0));
if (more) {
byte |= 0x80;
}
*p++ = byte;
} while (more);
return p;
}
/* Decode a signed leb128 sequence at *PP; increment *PP past the
decoded value. Return the decoded value. */
static target_long decode_sleb128(uint8_t **pp)
{
uint8_t *p = *pp;
target_long val = 0;
int byte, shift = 0;
do {
byte = *p++;
val |= (target_ulong)(byte & 0x7f) << shift;
shift += 7;
} while (byte & 0x80);
if (shift < TARGET_LONG_BITS && (byte & 0x40)) {
val |= -(target_ulong)1 << shift;
}
*pp = p;
return val;
}
/* Encode the data collected about the instructions while compiling TB.
Place the data at BLOCK, and return the number of bytes consumed.
The logical table consists of TARGET_INSN_START_WORDS target_ulong's,
which come from the target's insn_start data, followed by a uintptr_t
which comes from the host pc of the end of the code implementing the insn.
Each line of the table is encoded as sleb128 deltas from the previous
line. The seed for the first line is { tb->pc, 0..., tb->tc.ptr }.
That is, the first column is seeded with the guest pc, the last column
with the host pc, and the middle columns with zeros. */
static int encode_search(TranslationBlock *tb, uint8_t *block)
{
uint8_t *highwater = tcg_ctx->code_gen_highwater;
uint8_t *p = block;
int i, j, n;
for (i = 0, n = tb->icount; i < n; ++i) {
target_ulong prev;
for (j = 0; j < TARGET_INSN_START_WORDS; ++j) {
if (i == 0) {
prev = (j == 0 ? tb->pc : 0);
} else {
prev = tcg_ctx->gen_insn_data[i - 1][j];
}
p = encode_sleb128(p, tcg_ctx->gen_insn_data[i][j] - prev);
}
prev = (i == 0 ? 0 : tcg_ctx->gen_insn_end_off[i - 1]);
p = encode_sleb128(p, tcg_ctx->gen_insn_end_off[i] - prev);
/* Test for (pending) buffer overflow. The assumption is that any
one row beginning below the high water mark cannot overrun
the buffer completely. Thus we can test for overflow after
encoding a row without having to check during encoding. */
if (unlikely(p > highwater)) {
return -1;
}
}
return p - block;
}
/* The cpu state corresponding to 'searched_pc' is restored.
icount: fix cpu_restore_state_from_tb for non-tb-exit cases In icount mode, instructions that access io memory spaces in the middle of the translation block invoke TB recompilation. After recompilation, such instructions become last in the TB and are allowed to access io memory spaces. When the code includes instruction like i386 'xchg eax, 0xffffd080' which accesses APIC, QEMU goes into an infinite loop of the recompilation. This instruction includes two memory accesses - one read and one write. After the first access, APIC calls cpu_report_tpr_access, which restores the CPU state to get the current eip. But cpu_restore_state_from_tb resets the cpu->can_do_io flag which makes the second memory access invalid. Therefore the second memory access causes a recompilation of the block. Then these operations repeat again and again. This patch moves resetting cpu->can_do_io flag from cpu_restore_state_from_tb to cpu_loop_exit* functions. It also adds a parameter for cpu_restore_state which controls restoring icount. There is no need to restore icount when we only query CPU state without breaking the TB. Restoring it in such cases leads to the incorrect flow of the virtual time. In most cases new parameter is true (icount should be recalculated). But there are two cases in i386 and openrisc when the CPU state is only queried without the need to break the TB. This patch fixes both of these cases. Signed-off-by: Pavel Dovgalyuk <Pavel.Dovgaluk@ispras.ru> Message-Id: <20180409091320.12504.35329.stgit@pasha-VirtualBox> [rth: Make can_do_io setting unconditional; move from cpu_exec; make cpu_loop_exit_{noexc,restore} call cpu_loop_exit.] Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2018-04-09 11:13:20 +02:00
* When reset_icount is true, current TB will be interrupted and
* icount should be recalculated.
*/
static int cpu_restore_state_from_tb(CPUState *cpu, TranslationBlock *tb,
icount: fix cpu_restore_state_from_tb for non-tb-exit cases In icount mode, instructions that access io memory spaces in the middle of the translation block invoke TB recompilation. After recompilation, such instructions become last in the TB and are allowed to access io memory spaces. When the code includes instruction like i386 'xchg eax, 0xffffd080' which accesses APIC, QEMU goes into an infinite loop of the recompilation. This instruction includes two memory accesses - one read and one write. After the first access, APIC calls cpu_report_tpr_access, which restores the CPU state to get the current eip. But cpu_restore_state_from_tb resets the cpu->can_do_io flag which makes the second memory access invalid. Therefore the second memory access causes a recompilation of the block. Then these operations repeat again and again. This patch moves resetting cpu->can_do_io flag from cpu_restore_state_from_tb to cpu_loop_exit* functions. It also adds a parameter for cpu_restore_state which controls restoring icount. There is no need to restore icount when we only query CPU state without breaking the TB. Restoring it in such cases leads to the incorrect flow of the virtual time. In most cases new parameter is true (icount should be recalculated). But there are two cases in i386 and openrisc when the CPU state is only queried without the need to break the TB. This patch fixes both of these cases. Signed-off-by: Pavel Dovgalyuk <Pavel.Dovgaluk@ispras.ru> Message-Id: <20180409091320.12504.35329.stgit@pasha-VirtualBox> [rth: Make can_do_io setting unconditional; move from cpu_exec; make cpu_loop_exit_{noexc,restore} call cpu_loop_exit.] Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2018-04-09 11:13:20 +02:00
uintptr_t searched_pc, bool reset_icount)
{
target_ulong data[TARGET_INSN_START_WORDS] = { tb->pc };
uintptr_t host_pc = (uintptr_t)tb->tc.ptr;
CPUArchState *env = cpu->env_ptr;
translate-all: use a binary search tree to track TBs in TBContext This is a prerequisite for supporting multiple TCG contexts, since we will have threads generating code in separate regions of code_gen_buffer. For this we need a new field (.size) in struct tb_tc to keep track of the size of the translated code. This field uses a size_t to avoid adding a hole to the struct, although really an unsigned int would have been enough. The comparison function we use is optimized for the common case: insertions. Profiling shows that upon booting debian-arm, 98% of comparisons are between existing tb's (i.e. a->size and b->size are both !0), which happens during insertions (and removals, but those are rare). The remaining cases are lookups. From reading the glib sources we see that the first key is always the lookup key. However, the code does not assume this to always be the case because this behaviour is not guaranteed in the glib docs. However, we embed this knowledge in the code as a branch hint for the compiler. Note that tb_free does not free space in the code_gen_buffer anymore, since we cannot easily know whether the tb is the last one inserted in code_gen_buffer. The next patch in this series renames tb_free to tb_remove to reflect this. Performance-wise, lookups in tb_find_pc are the same as before: O(log n). However, insertions are O(log n) instead of O(1), which results in a small slowdown when booting debian-arm: Performance counter stats for 'build/arm-softmmu/qemu-system-arm \ -machine type=virt -nographic -smp 1 -m 4096 \ -netdev user,id=unet,hostfwd=tcp::2222-:22 \ -device virtio-net-device,netdev=unet \ -drive file=img/arm/jessie-arm32.qcow2,id=myblock,index=0,if=none \ -device virtio-blk-device,drive=myblock \ -kernel img/arm/aarch32-current-linux-kernel-only.img \ -append console=ttyAMA0 root=/dev/vda1 \ -name arm,debug-threads=on -smp 1' (10 runs): - Before: 8048.598422 task-clock (msec) # 0.931 CPUs utilized ( +- 0.28% ) 16,974 context-switches # 0.002 M/sec ( +- 0.12% ) 0 cpu-migrations # 0.000 K/sec 10,125 page-faults # 0.001 M/sec ( +- 1.23% ) 35,144,901,879 cycles # 4.367 GHz ( +- 0.14% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,758,252,643 instructions # 1.87 insns per cycle ( +- 0.33% ) 10,871,298,668 branches # 1350.707 M/sec ( +- 0.41% ) 192,322,212 branch-misses # 1.77% of all branches ( +- 0.32% ) 8.640869419 seconds time elapsed ( +- 0.57% ) - After: 8146.242027 task-clock (msec) # 0.923 CPUs utilized ( +- 1.23% ) 17,016 context-switches # 0.002 M/sec ( +- 0.40% ) 0 cpu-migrations # 0.000 K/sec 18,769 page-faults # 0.002 M/sec ( +- 0.45% ) 35,660,956,120 cycles # 4.378 GHz ( +- 1.22% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,095,366,607 instructions # 1.83 insns per cycle ( +- 1.73% ) 10,803,480,261 branches # 1326.192 M/sec ( +- 1.95% ) 195,601,289 branch-misses # 1.81% of all branches ( +- 0.39% ) 8.828660235 seconds time elapsed ( +- 0.38% ) Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-06-24 01:00:11 +02:00
uint8_t *p = tb->tc.ptr + tb->tc.size;
int i, j, num_insns = tb->icount;
#ifdef CONFIG_PROFILER
TCGProfile *prof = &tcg_ctx->prof;
int64_t ti = profile_getclock();
#endif
searched_pc -= GETPC_ADJ;
if (searched_pc < host_pc) {
return -1;
}
/* Reconstruct the stored insn data while looking for the point at
which the end of the insn exceeds the searched_pc. */
for (i = 0; i < num_insns; ++i) {
for (j = 0; j < TARGET_INSN_START_WORDS; ++j) {
data[j] += decode_sleb128(&p);
}
host_pc += decode_sleb128(&p);
if (host_pc > searched_pc) {
goto found;
}
}
return -1;
found:
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
if (reset_icount && (tb_cflags(tb) & CF_USE_ICOUNT)) {
assert(use_icount);
icount: fix cpu_restore_state_from_tb for non-tb-exit cases In icount mode, instructions that access io memory spaces in the middle of the translation block invoke TB recompilation. After recompilation, such instructions become last in the TB and are allowed to access io memory spaces. When the code includes instruction like i386 'xchg eax, 0xffffd080' which accesses APIC, QEMU goes into an infinite loop of the recompilation. This instruction includes two memory accesses - one read and one write. After the first access, APIC calls cpu_report_tpr_access, which restores the CPU state to get the current eip. But cpu_restore_state_from_tb resets the cpu->can_do_io flag which makes the second memory access invalid. Therefore the second memory access causes a recompilation of the block. Then these operations repeat again and again. This patch moves resetting cpu->can_do_io flag from cpu_restore_state_from_tb to cpu_loop_exit* functions. It also adds a parameter for cpu_restore_state which controls restoring icount. There is no need to restore icount when we only query CPU state without breaking the TB. Restoring it in such cases leads to the incorrect flow of the virtual time. In most cases new parameter is true (icount should be recalculated). But there are two cases in i386 and openrisc when the CPU state is only queried without the need to break the TB. This patch fixes both of these cases. Signed-off-by: Pavel Dovgalyuk <Pavel.Dovgaluk@ispras.ru> Message-Id: <20180409091320.12504.35329.stgit@pasha-VirtualBox> [rth: Make can_do_io setting unconditional; move from cpu_exec; make cpu_loop_exit_{noexc,restore} call cpu_loop_exit.] Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2018-04-09 11:13:20 +02:00
/* Reset the cycle counter to the start of the block
and shift if to the number of actually executed instructions */
cpu->icount_decr.u16.low += num_insns - i;
}
restore_state_to_opc(env, tb, data);
#ifdef CONFIG_PROFILER
atomic_set(&prof->restore_time,
prof->restore_time + profile_getclock() - ti);
atomic_set(&prof->restore_count, prof->restore_count + 1);
#endif
return 0;
}
icount: fix cpu_restore_state_from_tb for non-tb-exit cases In icount mode, instructions that access io memory spaces in the middle of the translation block invoke TB recompilation. After recompilation, such instructions become last in the TB and are allowed to access io memory spaces. When the code includes instruction like i386 'xchg eax, 0xffffd080' which accesses APIC, QEMU goes into an infinite loop of the recompilation. This instruction includes two memory accesses - one read and one write. After the first access, APIC calls cpu_report_tpr_access, which restores the CPU state to get the current eip. But cpu_restore_state_from_tb resets the cpu->can_do_io flag which makes the second memory access invalid. Therefore the second memory access causes a recompilation of the block. Then these operations repeat again and again. This patch moves resetting cpu->can_do_io flag from cpu_restore_state_from_tb to cpu_loop_exit* functions. It also adds a parameter for cpu_restore_state which controls restoring icount. There is no need to restore icount when we only query CPU state without breaking the TB. Restoring it in such cases leads to the incorrect flow of the virtual time. In most cases new parameter is true (icount should be recalculated). But there are two cases in i386 and openrisc when the CPU state is only queried without the need to break the TB. This patch fixes both of these cases. Signed-off-by: Pavel Dovgalyuk <Pavel.Dovgaluk@ispras.ru> Message-Id: <20180409091320.12504.35329.stgit@pasha-VirtualBox> [rth: Make can_do_io setting unconditional; move from cpu_exec; make cpu_loop_exit_{noexc,restore} call cpu_loop_exit.] Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2018-04-09 11:13:20 +02:00
bool cpu_restore_state(CPUState *cpu, uintptr_t host_pc, bool will_exit)
{
TranslationBlock *tb;
bool r = false;
uintptr_t check_offset;
/* The host_pc has to be in the region of current code buffer. If
* it is not we will not be able to resolve it here. The two cases
* where host_pc will not be correct are:
*
* - fault during translation (instruction fetch)
* - fault from helper (not using GETPC() macro)
*
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
* Either way we need return early as we can't resolve it here.
*
* We are using unsigned arithmetic so if host_pc <
* tcg_init_ctx.code_gen_buffer check_offset will wrap to way
* above the code_gen_buffer_size
*/
check_offset = host_pc - (uintptr_t) tcg_init_ctx.code_gen_buffer;
if (check_offset < tcg_init_ctx.code_gen_buffer_size) {
tcg: track TBs with per-region BST's This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-26 22:58:05 +02:00
tb = tcg_tb_lookup(host_pc);
if (tb) {
icount: fix cpu_restore_state_from_tb for non-tb-exit cases In icount mode, instructions that access io memory spaces in the middle of the translation block invoke TB recompilation. After recompilation, such instructions become last in the TB and are allowed to access io memory spaces. When the code includes instruction like i386 'xchg eax, 0xffffd080' which accesses APIC, QEMU goes into an infinite loop of the recompilation. This instruction includes two memory accesses - one read and one write. After the first access, APIC calls cpu_report_tpr_access, which restores the CPU state to get the current eip. But cpu_restore_state_from_tb resets the cpu->can_do_io flag which makes the second memory access invalid. Therefore the second memory access causes a recompilation of the block. Then these operations repeat again and again. This patch moves resetting cpu->can_do_io flag from cpu_restore_state_from_tb to cpu_loop_exit* functions. It also adds a parameter for cpu_restore_state which controls restoring icount. There is no need to restore icount when we only query CPU state without breaking the TB. Restoring it in such cases leads to the incorrect flow of the virtual time. In most cases new parameter is true (icount should be recalculated). But there are two cases in i386 and openrisc when the CPU state is only queried without the need to break the TB. This patch fixes both of these cases. Signed-off-by: Pavel Dovgalyuk <Pavel.Dovgaluk@ispras.ru> Message-Id: <20180409091320.12504.35329.stgit@pasha-VirtualBox> [rth: Make can_do_io setting unconditional; move from cpu_exec; make cpu_loop_exit_{noexc,restore} call cpu_loop_exit.] Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2018-04-09 11:13:20 +02:00
cpu_restore_state_from_tb(cpu, tb, host_pc, will_exit);
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
if (tb_cflags(tb) & CF_NOCACHE) {
/* one-shot translation, invalidate it immediately */
tb_phys_invalidate(tb, -1);
tcg: track TBs with per-region BST's This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-26 22:58:05 +02:00
tcg_tb_remove(tb);
}
r = true;
}
}
return r;
}
static void page_init(void)
{
page_size_init();
page_table_config_init();
#if defined(CONFIG_BSD) && defined(CONFIG_USER_ONLY)
{
#ifdef HAVE_KINFO_GETVMMAP
struct kinfo_vmentry *freep;
int i, cnt;
freep = kinfo_getvmmap(getpid(), &cnt);
if (freep) {
mmap_lock();
for (i = 0; i < cnt; i++) {
unsigned long startaddr, endaddr;
startaddr = freep[i].kve_start;
endaddr = freep[i].kve_end;
if (h2g_valid(startaddr)) {
startaddr = h2g(startaddr) & TARGET_PAGE_MASK;
if (h2g_valid(endaddr)) {
endaddr = h2g(endaddr);
page_set_flags(startaddr, endaddr, PAGE_RESERVED);
} else {
#if TARGET_ABI_BITS <= L1_MAP_ADDR_SPACE_BITS
endaddr = ~0ul;
page_set_flags(startaddr, endaddr, PAGE_RESERVED);
#endif
}
}
}
free(freep);
mmap_unlock();
}
#else
FILE *f;
last_brk = (unsigned long)sbrk(0);
f = fopen("/compat/linux/proc/self/maps", "r");
if (f) {
mmap_lock();
do {
unsigned long startaddr, endaddr;
int n;
n = fscanf(f, "%lx-%lx %*[^\n]\n", &startaddr, &endaddr);
if (n == 2 && h2g_valid(startaddr)) {
startaddr = h2g(startaddr) & TARGET_PAGE_MASK;
if (h2g_valid(endaddr)) {
endaddr = h2g(endaddr);
} else {
endaddr = ~0ul;
}
page_set_flags(startaddr, endaddr, PAGE_RESERVED);
}
} while (!feof(f));
fclose(f);
mmap_unlock();
}
#endif
}
#endif
}
static PageDesc *page_find_alloc(tb_page_addr_t index, int alloc)
{
PageDesc *pd;
void **lp;
int i;
/* Level 1. Always allocated. */
lp = l1_map + ((index >> v_l1_shift) & (v_l1_size - 1));
/* Level 2..N-1. */
for (i = v_l2_levels; i > 0; i--) {
void **p = atomic_rcu_read(lp);
if (p == NULL) {
void *existing;
if (!alloc) {
return NULL;
}
p = g_new0(void *, V_L2_SIZE);
existing = atomic_cmpxchg(lp, NULL, p);
if (unlikely(existing)) {
g_free(p);
p = existing;
}
}
lp = p + ((index >> (i * V_L2_BITS)) & (V_L2_SIZE - 1));
}
pd = atomic_rcu_read(lp);
if (pd == NULL) {
void *existing;
if (!alloc) {
return NULL;
}
pd = g_new0(PageDesc, V_L2_SIZE);
#ifndef CONFIG_USER_ONLY
{
int i;
for (i = 0; i < V_L2_SIZE; i++) {
qemu_spin_init(&pd[i].lock);
}
}
#endif
existing = atomic_cmpxchg(lp, NULL, pd);
if (unlikely(existing)) {
g_free(pd);
pd = existing;
}
}
return pd + (index & (V_L2_SIZE - 1));
}
static inline PageDesc *page_find(tb_page_addr_t index)
{
return page_find_alloc(index, 0);
}
static void page_lock_pair(PageDesc **ret_p1, tb_page_addr_t phys1,
PageDesc **ret_p2, tb_page_addr_t phys2, int alloc);
/* In user-mode page locks aren't used; mmap_lock is enough */
#ifdef CONFIG_USER_ONLY
#define assert_page_locked(pd) tcg_debug_assert(have_mmap_lock())
static inline void page_lock(PageDesc *pd)
{ }
static inline void page_unlock(PageDesc *pd)
{ }
static inline void page_lock_tb(const TranslationBlock *tb)
{ }
static inline void page_unlock_tb(const TranslationBlock *tb)
{ }
struct page_collection *
page_collection_lock(tb_page_addr_t start, tb_page_addr_t end)
{
return NULL;
}
void page_collection_unlock(struct page_collection *set)
{ }
#else /* !CONFIG_USER_ONLY */
#ifdef CONFIG_DEBUG_TCG
static __thread GHashTable *ht_pages_locked_debug;
static void ht_pages_locked_debug_init(void)
{
if (ht_pages_locked_debug) {
return;
}
ht_pages_locked_debug = g_hash_table_new(NULL, NULL);
}
static bool page_is_locked(const PageDesc *pd)
{
PageDesc *found;
ht_pages_locked_debug_init();
found = g_hash_table_lookup(ht_pages_locked_debug, pd);
return !!found;
}
static void page_lock__debug(PageDesc *pd)
{
ht_pages_locked_debug_init();
g_assert(!page_is_locked(pd));
g_hash_table_insert(ht_pages_locked_debug, pd, pd);
}
static void page_unlock__debug(const PageDesc *pd)
{
bool removed;
ht_pages_locked_debug_init();
g_assert(page_is_locked(pd));
removed = g_hash_table_remove(ht_pages_locked_debug, pd);
g_assert(removed);
}
static void
do_assert_page_locked(const PageDesc *pd, const char *file, int line)
{
if (unlikely(!page_is_locked(pd))) {
error_report("assert_page_lock: PageDesc %p not locked @ %s:%d",
pd, file, line);
abort();
}
}
#define assert_page_locked(pd) do_assert_page_locked(pd, __FILE__, __LINE__)
void assert_no_pages_locked(void)
{
ht_pages_locked_debug_init();
g_assert(g_hash_table_size(ht_pages_locked_debug) == 0);
}
#else /* !CONFIG_DEBUG_TCG */
#define assert_page_locked(pd)
static inline void page_lock__debug(const PageDesc *pd)
{
}
static inline void page_unlock__debug(const PageDesc *pd)
{
}
#endif /* CONFIG_DEBUG_TCG */
static inline void page_lock(PageDesc *pd)
{
page_lock__debug(pd);
qemu_spin_lock(&pd->lock);
}
static inline void page_unlock(PageDesc *pd)
{
qemu_spin_unlock(&pd->lock);
page_unlock__debug(pd);
}
/* lock the page(s) of a TB in the correct acquisition order */
static inline void page_lock_tb(const TranslationBlock *tb)
{
page_lock_pair(NULL, tb->page_addr[0], NULL, tb->page_addr[1], 0);
}
static inline void page_unlock_tb(const TranslationBlock *tb)
{
PageDesc *p1 = page_find(tb->page_addr[0] >> TARGET_PAGE_BITS);
page_unlock(p1);
if (unlikely(tb->page_addr[1] != -1)) {
PageDesc *p2 = page_find(tb->page_addr[1] >> TARGET_PAGE_BITS);
if (p2 != p1) {
page_unlock(p2);
}
}
}
static inline struct page_entry *
page_entry_new(PageDesc *pd, tb_page_addr_t index)
{
struct page_entry *pe = g_malloc(sizeof(*pe));
pe->index = index;
pe->pd = pd;
pe->locked = false;
return pe;
}
static void page_entry_destroy(gpointer p)
{
struct page_entry *pe = p;
g_assert(pe->locked);
page_unlock(pe->pd);
g_free(pe);
}
/* returns false on success */
static bool page_entry_trylock(struct page_entry *pe)
{
bool busy;
busy = qemu_spin_trylock(&pe->pd->lock);
if (!busy) {
g_assert(!pe->locked);
pe->locked = true;
page_lock__debug(pe->pd);
}
return busy;
}
static void do_page_entry_lock(struct page_entry *pe)
{
page_lock(pe->pd);
g_assert(!pe->locked);
pe->locked = true;
}
static gboolean page_entry_lock(gpointer key, gpointer value, gpointer data)
{
struct page_entry *pe = value;
do_page_entry_lock(pe);
return FALSE;
}
static gboolean page_entry_unlock(gpointer key, gpointer value, gpointer data)
{
struct page_entry *pe = value;
if (pe->locked) {
pe->locked = false;
page_unlock(pe->pd);
}
return FALSE;
}
/*
* Trylock a page, and if successful, add the page to a collection.
* Returns true ("busy") if the page could not be locked; false otherwise.
*/
static bool page_trylock_add(struct page_collection *set, tb_page_addr_t addr)
{
tb_page_addr_t index = addr >> TARGET_PAGE_BITS;
struct page_entry *pe;
PageDesc *pd;
pe = g_tree_lookup(set->tree, &index);
if (pe) {
return false;
}
pd = page_find(index);
if (pd == NULL) {
return false;
}
pe = page_entry_new(pd, index);
g_tree_insert(set->tree, &pe->index, pe);
/*
* If this is either (1) the first insertion or (2) a page whose index
* is higher than any other so far, just lock the page and move on.
*/
if (set->max == NULL || pe->index > set->max->index) {
set->max = pe;
do_page_entry_lock(pe);
return false;
}
/*
* Try to acquire out-of-order lock; if busy, return busy so that we acquire
* locks in order.
*/
return page_entry_trylock(pe);
}
static gint tb_page_addr_cmp(gconstpointer ap, gconstpointer bp, gpointer udata)
{
tb_page_addr_t a = *(const tb_page_addr_t *)ap;
tb_page_addr_t b = *(const tb_page_addr_t *)bp;
if (a == b) {
return 0;
} else if (a < b) {
return -1;
}
return 1;
}
/*
* Lock a range of pages ([@start,@end[) as well as the pages of all
* intersecting TBs.
* Locking order: acquire locks in ascending order of page index.
*/
struct page_collection *
page_collection_lock(tb_page_addr_t start, tb_page_addr_t end)
{
struct page_collection *set = g_malloc(sizeof(*set));
tb_page_addr_t index;
PageDesc *pd;
start >>= TARGET_PAGE_BITS;
end >>= TARGET_PAGE_BITS;
g_assert(start <= end);
set->tree = g_tree_new_full(tb_page_addr_cmp, NULL, NULL,
page_entry_destroy);
set->max = NULL;
assert_no_pages_locked();
retry:
g_tree_foreach(set->tree, page_entry_lock, NULL);
for (index = start; index <= end; index++) {
TranslationBlock *tb;
int n;
pd = page_find(index);
if (pd == NULL) {
continue;
}
if (page_trylock_add(set, index << TARGET_PAGE_BITS)) {
g_tree_foreach(set->tree, page_entry_unlock, NULL);
goto retry;
}
assert_page_locked(pd);
PAGE_FOR_EACH_TB(pd, tb, n) {
if (page_trylock_add(set, tb->page_addr[0]) ||
(tb->page_addr[1] != -1 &&
page_trylock_add(set, tb->page_addr[1]))) {
/* drop all locks, and reacquire in order */
g_tree_foreach(set->tree, page_entry_unlock, NULL);
goto retry;
}
}
}
return set;
}
void page_collection_unlock(struct page_collection *set)
{
/* entries are unlocked and freed via page_entry_destroy */
g_tree_destroy(set->tree);
g_free(set);
}
#endif /* !CONFIG_USER_ONLY */
static void page_lock_pair(PageDesc **ret_p1, tb_page_addr_t phys1,
PageDesc **ret_p2, tb_page_addr_t phys2, int alloc)
{
PageDesc *p1, *p2;
tb_page_addr_t page1;
tb_page_addr_t page2;
assert_memory_lock();
g_assert(phys1 != -1);
page1 = phys1 >> TARGET_PAGE_BITS;
page2 = phys2 >> TARGET_PAGE_BITS;
p1 = page_find_alloc(page1, alloc);
if (ret_p1) {
*ret_p1 = p1;
}
if (likely(phys2 == -1)) {
page_lock(p1);
return;
} else if (page1 == page2) {
page_lock(p1);
if (ret_p2) {
*ret_p2 = p1;
}
return;
}
p2 = page_find_alloc(page2, alloc);
if (ret_p2) {
*ret_p2 = p2;
}
if (page1 < page2) {
page_lock(p1);
page_lock(p2);
} else {
page_lock(p2);
page_lock(p1);
}
}
#if defined(CONFIG_USER_ONLY)
/* Currently it is not recommended to allocate big chunks of data in
user mode. It will change when a dedicated libc will be used. */
/* ??? 64-bit hosts ought to have no problem mmaping data outside the
region in which the guest needs to run. Revisit this. */
#define USE_STATIC_CODE_GEN_BUFFER
#endif
/* Minimum size of the code gen buffer. This number is randomly chosen,
but not so small that we can't have a fair number of TB's live. */
#define MIN_CODE_GEN_BUFFER_SIZE (1024u * 1024)
/* Maximum size of the code gen buffer we'd like to use. Unless otherwise
indicated, this is constrained by the range of direct branches on the
host cpu, as used by the TCG implementation of goto_tb. */
#if defined(__x86_64__)
# define MAX_CODE_GEN_BUFFER_SIZE (2ul * 1024 * 1024 * 1024)
#elif defined(__sparc__)
# define MAX_CODE_GEN_BUFFER_SIZE (2ul * 1024 * 1024 * 1024)
#elif defined(__powerpc64__)
# define MAX_CODE_GEN_BUFFER_SIZE (2ul * 1024 * 1024 * 1024)
#elif defined(__powerpc__)
# define MAX_CODE_GEN_BUFFER_SIZE (32u * 1024 * 1024)
#elif defined(__aarch64__)
# define MAX_CODE_GEN_BUFFER_SIZE (2ul * 1024 * 1024 * 1024)
#elif defined(__s390x__)
/* We have a +- 4GB range on the branches; leave some slop. */
# define MAX_CODE_GEN_BUFFER_SIZE (3ul * 1024 * 1024 * 1024)
#elif defined(__mips__)
/* We have a 256MB branch region, but leave room to make sure the
main executable is also within that region. */
# define MAX_CODE_GEN_BUFFER_SIZE (128ul * 1024 * 1024)
#else
# define MAX_CODE_GEN_BUFFER_SIZE ((size_t)-1)
#endif
#define DEFAULT_CODE_GEN_BUFFER_SIZE_1 (32u * 1024 * 1024)
#define DEFAULT_CODE_GEN_BUFFER_SIZE \
(DEFAULT_CODE_GEN_BUFFER_SIZE_1 < MAX_CODE_GEN_BUFFER_SIZE \
? DEFAULT_CODE_GEN_BUFFER_SIZE_1 : MAX_CODE_GEN_BUFFER_SIZE)
static inline size_t size_code_gen_buffer(size_t tb_size)
{
/* Size the buffer. */
if (tb_size == 0) {
#ifdef USE_STATIC_CODE_GEN_BUFFER
tb_size = DEFAULT_CODE_GEN_BUFFER_SIZE;
#else
/* ??? Needs adjustments. */
/* ??? If we relax the requirement that CONFIG_USER_ONLY use the
static buffer, we could size this on RESERVED_VA, on the text
segment size of the executable, or continue to use the default. */
tb_size = (unsigned long)(ram_size / 4);
#endif
}
if (tb_size < MIN_CODE_GEN_BUFFER_SIZE) {
tb_size = MIN_CODE_GEN_BUFFER_SIZE;
}
if (tb_size > MAX_CODE_GEN_BUFFER_SIZE) {
tb_size = MAX_CODE_GEN_BUFFER_SIZE;
}
return tb_size;
}
#ifdef __mips__
/* In order to use J and JAL within the code_gen_buffer, we require
that the buffer not cross a 256MB boundary. */
static inline bool cross_256mb(void *addr, size_t size)
{
return ((uintptr_t)addr ^ ((uintptr_t)addr + size)) & ~0x0ffffffful;
}
/* We weren't able to allocate a buffer without crossing that boundary,
so make do with the larger portion of the buffer that doesn't cross.
Returns the new base of the buffer, and adjusts code_gen_buffer_size. */
static inline void *split_cross_256mb(void *buf1, size_t size1)
{
void *buf2 = (void *)(((uintptr_t)buf1 + size1) & ~0x0ffffffful);
size_t size2 = buf1 + size1 - buf2;
size1 = buf2 - buf1;
if (size1 < size2) {
size1 = size2;
buf1 = buf2;
}
tcg_ctx->code_gen_buffer_size = size1;
return buf1;
}
#endif
#ifdef USE_STATIC_CODE_GEN_BUFFER
static uint8_t static_code_gen_buffer[DEFAULT_CODE_GEN_BUFFER_SIZE]
__attribute__((aligned(CODE_GEN_ALIGN)));
static inline void *alloc_code_gen_buffer(void)
{
void *buf = static_code_gen_buffer;
void *end = static_code_gen_buffer + sizeof(static_code_gen_buffer);
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
size_t size;
/* page-align the beginning and end of the buffer */
buf = QEMU_ALIGN_PTR_UP(buf, qemu_real_host_page_size);
end = QEMU_ALIGN_PTR_DOWN(end, qemu_real_host_page_size);
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
size = end - buf;
/* Honor a command-line option limiting the size of the buffer. */
if (size > tcg_ctx->code_gen_buffer_size) {
size = QEMU_ALIGN_DOWN(tcg_ctx->code_gen_buffer_size,
qemu_real_host_page_size);
}
tcg_ctx->code_gen_buffer_size = size;
#ifdef __mips__
if (cross_256mb(buf, size)) {
buf = split_cross_256mb(buf, size);
size = tcg_ctx->code_gen_buffer_size;
}
#endif
if (qemu_mprotect_rwx(buf, size)) {
abort();
}
qemu_madvise(buf, size, QEMU_MADV_HUGEPAGE);
return buf;
}
#elif defined(_WIN32)
static inline void *alloc_code_gen_buffer(void)
{
size_t size = tcg_ctx->code_gen_buffer_size;
return VirtualAlloc(NULL, size, MEM_RESERVE | MEM_COMMIT,
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
PAGE_EXECUTE_READWRITE);
}
#else
static inline void *alloc_code_gen_buffer(void)
{
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
int prot = PROT_WRITE | PROT_READ | PROT_EXEC;
int flags = MAP_PRIVATE | MAP_ANONYMOUS;
uintptr_t start = 0;
size_t size = tcg_ctx->code_gen_buffer_size;
void *buf;
/* Constrain the position of the buffer based on the host cpu.
Note that these addresses are chosen in concert with the
addresses assigned in the relevant linker script file. */
# if defined(__PIE__) || defined(__PIC__)
/* Don't bother setting a preferred location if we're building
a position-independent executable. We're more likely to get
an address near the main executable if we let the kernel
choose the address. */
# elif defined(__x86_64__) && defined(MAP_32BIT)
/* Force the memory down into low memory with the executable.
Leave the choice of exact location with the kernel. */
flags |= MAP_32BIT;
/* Cannot expect to map more than 800MB in low memory. */
if (size > 800u * 1024 * 1024) {
tcg_ctx->code_gen_buffer_size = size = 800u * 1024 * 1024;
}
# elif defined(__sparc__)
start = 0x40000000ul;
# elif defined(__s390x__)
start = 0x90000000ul;
# elif defined(__mips__)
# if _MIPS_SIM == _ABI64
start = 0x128000000ul;
# else
start = 0x08000000ul;
# endif
# endif
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
buf = mmap((void *)start, size, prot, flags, -1, 0);
if (buf == MAP_FAILED) {
return NULL;
}
#ifdef __mips__
if (cross_256mb(buf, size)) {
/* Try again, with the original still mapped, to avoid re-acquiring
that 256mb crossing. This time don't specify an address. */
size_t size2;
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
void *buf2 = mmap(NULL, size, prot, flags, -1, 0);
switch ((int)(buf2 != MAP_FAILED)) {
case 1:
if (!cross_256mb(buf2, size)) {
/* Success! Use the new buffer. */
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
munmap(buf, size);
break;
}
/* Failure. Work with what we had. */
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
munmap(buf2, size);
/* fallthru */
default:
/* Split the original buffer. Free the smaller half. */
buf2 = split_cross_256mb(buf, size);
size2 = tcg_ctx->code_gen_buffer_size;
if (buf == buf2) {
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
munmap(buf + size2, size - size2);
} else {
munmap(buf, size - size2);
}
size = size2;
break;
}
buf = buf2;
}
#endif
/* Request large pages for the buffer. */
qemu_madvise(buf, size, QEMU_MADV_HUGEPAGE);
return buf;
}
#endif /* USE_STATIC_CODE_GEN_BUFFER, WIN32, POSIX */
static inline void code_gen_alloc(size_t tb_size)
{
tcg_ctx->code_gen_buffer_size = size_code_gen_buffer(tb_size);
tcg_ctx->code_gen_buffer = alloc_code_gen_buffer();
if (tcg_ctx->code_gen_buffer == NULL) {
fprintf(stderr, "Could not allocate dynamic translator buffer\n");
exit(1);
}
}
static bool tb_cmp(const void *ap, const void *bp)
{
const TranslationBlock *a = ap;
const TranslationBlock *b = bp;
return a->pc == b->pc &&
a->cs_base == b->cs_base &&
a->flags == b->flags &&
(tb_cflags(a) & CF_HASH_MASK) == (tb_cflags(b) & CF_HASH_MASK) &&
a->trace_vcpu_dstate == b->trace_vcpu_dstate &&
a->page_addr[0] == b->page_addr[0] &&
a->page_addr[1] == b->page_addr[1];
}
tb hash: track translated blocks with qht Having a fixed-size hash table for keeping track of all translation blocks is suboptimal: some workloads are just too big or too small to get maximum performance from the hash table. The MRU promotion policy helps improve performance when the hash table is a little undersized, but it cannot make up for severely undersized hash tables. Furthermore, frequent MRU promotions result in writes that are a scalability bottleneck. For scalability, lookups should only perform reads, not writes. This is not a big deal for now, but it will become one once MTTCG matures. The appended fixes these issues by using qht as the implementation of the TB hash table. This solution is superior to other alternatives considered, namely: - master: implementation in QEMU before this patchset - xxhash: before this patch, i.e. fixed buckets + xxhash hashing + MRU. - xxhash-rcu: fixed buckets + xxhash + RCU list + MRU. MRU is implemented here by adding an intermediate struct that contains the u32 hash and a pointer to the TB; this allows us, on an MRU promotion, to copy said struct (that is not at the head), and put this new copy at the head. After a grace period, the original non-head struct can be eliminated, and after another grace period, freed. - qht-fixed-nomru: fixed buckets + xxhash + qht without auto-resize + no MRU for lookups; MRU for inserts. The appended solution is the following: - qht-dyn-nomru: dynamic number of buckets + xxhash + qht w/ auto-resize + no MRU for lookups; MRU for inserts. The plots below compare the considered solutions. The Y axis shows the boot time (in seconds) of a debian jessie image with arm-softmmu; the X axis sweeps the number of buckets (or initial number of buckets for qht-autoresize). The plots in PNG format (and with errorbars) can be seen here: http://imgur.com/a/Awgnq Each test runs 5 times, and the entire QEMU process is pinned to a single core for repeatability of results. Host: Intel Xeon E5-2690 28 ++------------+-------------+-------------+-------------+------------++ A***** + + + master **A*** + 27 ++ * xxhash ##B###++ | A******A****** xxhash-rcu $$C$$$ | 26 C$$ A******A****** qht-fixed-nomru*%%D%%%++ D%%$$ A******A******A*qht-dyn-mru A*E****A 25 ++ %%$$ qht-dyn-nomru &&F&&&++ B#####% | 24 ++ #C$$$$$ ++ | B### $ | | ## C$$$$$$ | 23 ++ # C$$$$$$ ++ | B###### C$$$$$$ %%%D 22 ++ %B###### C$$$$$$C$$$$$$C$$$$$$C$$$$$$C$$$$$$C | D%%%%%%B###### @E@@@@@@ %%%D%%%@@@E@@@@@@E 21 E@@@@@@E@@@@@@F&&&@@@E@@@&&&D%%%%%%B######B######B######B######B######B + E@@@ F&&& + E@ + F&&& + + 20 ++------------+-------------+-------------+-------------+------------++ 14 16 18 20 22 24 log2 number of buckets Host: Intel i7-4790K 14.5 ++------------+------------+-------------+------------+------------++ A** + + + master **A*** + 14 ++ ** xxhash ##B###++ 13.5 ++ ** xxhash-rcu $$C$$$++ | qht-fixed-nomru %%D%%% | 13 ++ A****** qht-dyn-mru @@E@@@++ | A*****A******A****** qht-dyn-nomru &&F&&& | 12.5 C$$ A******A******A*****A****** ***A 12 ++ $$ A*** ++ D%%% $$ | 11.5 ++ %% ++ B### %C$$$$$$ | 11 ++ ## D%%%%% C$$$$$ ++ | # % C$$$$$$ | 10.5 F&&&&&&B######D%%%%% C$$$$$$C$$$$$$C$$$$$$C$$$$$C$$$$$$ $$$C 10 E@@@@@@E@@@@@@B#####B######B######E@@@@@@E@@@%%%D%%%%%D%%%###B######B + F&& D%%%%%%B######B######B#####B###@@@D%%% + 9.5 ++------------+------------+-------------+------------+------------++ 14 16 18 20 22 24 log2 number of buckets Note that the original point before this patch series is X=15 for "master"; the little sensitivity to the increased number of buckets is due to the poor hashing function in master. xxhash-rcu has significant overhead due to the constant churn of allocating and deallocating intermediate structs for implementing MRU. An alternative would be do consider failed lookups as "maybe not there", and then acquire the external lock (tb_lock in this case) to really confirm that there was indeed a failed lookup. This, however, would not be enough to implement dynamic resizing--this is more complex: see "Resizable, Scalable, Concurrent Hash Tables via Relativistic Programming" by Triplett, McKenney and Walpole. This solution was discarded due to the very coarse RCU read critical sections that we have in MTTCG; resizing requires waiting for readers after every pointer update, and resizes require many pointer updates, so this would quickly become prohibitive. qht-fixed-nomru shows that MRU promotion is advisable for undersized hash tables. However, qht-dyn-mru shows that MRU promotion is not important if the hash table is properly sized: there is virtually no difference in performance between qht-dyn-nomru and qht-dyn-mru. Before this patch, we're at X=15 on "xxhash"; after this patch, we're at X=15 @ qht-dyn-nomru. This patch thus matches the best performance that we can achieve with optimum sizing of the hash table, while keeping the hash table scalable for readers. The improvement we get before and after this patch for booting debian jessie with arm-softmmu is: - Intel Xeon E5-2690: 10.5% less time - Intel i7-4790K: 5.2% less time We could get this same improvement _for this particular workload_ by statically increasing the size of the hash table. But this would hurt workloads that do not need a large hash table. The dynamic (upward) resizing allows us to start small and enlarge the hash table as needed. A quick note on downsizing: the table is resized back to 2**15 buckets on every tb_flush; this makes sense because it is not guaranteed that the table will reach the same number of TBs later on (e.g. most bootup code is thrown away after boot); it makes sense to grow the hash table as more code blocks are translated. This also avoids the complication of having to build downsizing hysteresis logic into qht. Reviewed-by: Sergey Fedorov <serge.fedorov@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1465412133-3029-15-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2016-06-08 20:55:32 +02:00
static void tb_htable_init(void)
{
unsigned int mode = QHT_MODE_AUTO_RESIZE;
qht_init(&tb_ctx.htable, tb_cmp, CODE_GEN_HTABLE_SIZE, mode);
tb hash: track translated blocks with qht Having a fixed-size hash table for keeping track of all translation blocks is suboptimal: some workloads are just too big or too small to get maximum performance from the hash table. The MRU promotion policy helps improve performance when the hash table is a little undersized, but it cannot make up for severely undersized hash tables. Furthermore, frequent MRU promotions result in writes that are a scalability bottleneck. For scalability, lookups should only perform reads, not writes. This is not a big deal for now, but it will become one once MTTCG matures. The appended fixes these issues by using qht as the implementation of the TB hash table. This solution is superior to other alternatives considered, namely: - master: implementation in QEMU before this patchset - xxhash: before this patch, i.e. fixed buckets + xxhash hashing + MRU. - xxhash-rcu: fixed buckets + xxhash + RCU list + MRU. MRU is implemented here by adding an intermediate struct that contains the u32 hash and a pointer to the TB; this allows us, on an MRU promotion, to copy said struct (that is not at the head), and put this new copy at the head. After a grace period, the original non-head struct can be eliminated, and after another grace period, freed. - qht-fixed-nomru: fixed buckets + xxhash + qht without auto-resize + no MRU for lookups; MRU for inserts. The appended solution is the following: - qht-dyn-nomru: dynamic number of buckets + xxhash + qht w/ auto-resize + no MRU for lookups; MRU for inserts. The plots below compare the considered solutions. The Y axis shows the boot time (in seconds) of a debian jessie image with arm-softmmu; the X axis sweeps the number of buckets (or initial number of buckets for qht-autoresize). The plots in PNG format (and with errorbars) can be seen here: http://imgur.com/a/Awgnq Each test runs 5 times, and the entire QEMU process is pinned to a single core for repeatability of results. Host: Intel Xeon E5-2690 28 ++------------+-------------+-------------+-------------+------------++ A***** + + + master **A*** + 27 ++ * xxhash ##B###++ | A******A****** xxhash-rcu $$C$$$ | 26 C$$ A******A****** qht-fixed-nomru*%%D%%%++ D%%$$ A******A******A*qht-dyn-mru A*E****A 25 ++ %%$$ qht-dyn-nomru &&F&&&++ B#####% | 24 ++ #C$$$$$ ++ | B### $ | | ## C$$$$$$ | 23 ++ # C$$$$$$ ++ | B###### C$$$$$$ %%%D 22 ++ %B###### C$$$$$$C$$$$$$C$$$$$$C$$$$$$C$$$$$$C | D%%%%%%B###### @E@@@@@@ %%%D%%%@@@E@@@@@@E 21 E@@@@@@E@@@@@@F&&&@@@E@@@&&&D%%%%%%B######B######B######B######B######B + E@@@ F&&& + E@ + F&&& + + 20 ++------------+-------------+-------------+-------------+------------++ 14 16 18 20 22 24 log2 number of buckets Host: Intel i7-4790K 14.5 ++------------+------------+-------------+------------+------------++ A** + + + master **A*** + 14 ++ ** xxhash ##B###++ 13.5 ++ ** xxhash-rcu $$C$$$++ | qht-fixed-nomru %%D%%% | 13 ++ A****** qht-dyn-mru @@E@@@++ | A*****A******A****** qht-dyn-nomru &&F&&& | 12.5 C$$ A******A******A*****A****** ***A 12 ++ $$ A*** ++ D%%% $$ | 11.5 ++ %% ++ B### %C$$$$$$ | 11 ++ ## D%%%%% C$$$$$ ++ | # % C$$$$$$ | 10.5 F&&&&&&B######D%%%%% C$$$$$$C$$$$$$C$$$$$$C$$$$$C$$$$$$ $$$C 10 E@@@@@@E@@@@@@B#####B######B######E@@@@@@E@@@%%%D%%%%%D%%%###B######B + F&& D%%%%%%B######B######B#####B###@@@D%%% + 9.5 ++------------+------------+-------------+------------+------------++ 14 16 18 20 22 24 log2 number of buckets Note that the original point before this patch series is X=15 for "master"; the little sensitivity to the increased number of buckets is due to the poor hashing function in master. xxhash-rcu has significant overhead due to the constant churn of allocating and deallocating intermediate structs for implementing MRU. An alternative would be do consider failed lookups as "maybe not there", and then acquire the external lock (tb_lock in this case) to really confirm that there was indeed a failed lookup. This, however, would not be enough to implement dynamic resizing--this is more complex: see "Resizable, Scalable, Concurrent Hash Tables via Relativistic Programming" by Triplett, McKenney and Walpole. This solution was discarded due to the very coarse RCU read critical sections that we have in MTTCG; resizing requires waiting for readers after every pointer update, and resizes require many pointer updates, so this would quickly become prohibitive. qht-fixed-nomru shows that MRU promotion is advisable for undersized hash tables. However, qht-dyn-mru shows that MRU promotion is not important if the hash table is properly sized: there is virtually no difference in performance between qht-dyn-nomru and qht-dyn-mru. Before this patch, we're at X=15 on "xxhash"; after this patch, we're at X=15 @ qht-dyn-nomru. This patch thus matches the best performance that we can achieve with optimum sizing of the hash table, while keeping the hash table scalable for readers. The improvement we get before and after this patch for booting debian jessie with arm-softmmu is: - Intel Xeon E5-2690: 10.5% less time - Intel i7-4790K: 5.2% less time We could get this same improvement _for this particular workload_ by statically increasing the size of the hash table. But this would hurt workloads that do not need a large hash table. The dynamic (upward) resizing allows us to start small and enlarge the hash table as needed. A quick note on downsizing: the table is resized back to 2**15 buckets on every tb_flush; this makes sense because it is not guaranteed that the table will reach the same number of TBs later on (e.g. most bootup code is thrown away after boot); it makes sense to grow the hash table as more code blocks are translated. This also avoids the complication of having to build downsizing hysteresis logic into qht. Reviewed-by: Sergey Fedorov <serge.fedorov@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1465412133-3029-15-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2016-06-08 20:55:32 +02:00
}
/* Must be called before using the QEMU cpus. 'tb_size' is the size
(in bytes) allocated to the translation buffer. Zero means default
size. */
void tcg_exec_init(unsigned long tb_size)
{
tcg_allowed = true;
cpu_gen_init();
page_init();
tb hash: track translated blocks with qht Having a fixed-size hash table for keeping track of all translation blocks is suboptimal: some workloads are just too big or too small to get maximum performance from the hash table. The MRU promotion policy helps improve performance when the hash table is a little undersized, but it cannot make up for severely undersized hash tables. Furthermore, frequent MRU promotions result in writes that are a scalability bottleneck. For scalability, lookups should only perform reads, not writes. This is not a big deal for now, but it will become one once MTTCG matures. The appended fixes these issues by using qht as the implementation of the TB hash table. This solution is superior to other alternatives considered, namely: - master: implementation in QEMU before this patchset - xxhash: before this patch, i.e. fixed buckets + xxhash hashing + MRU. - xxhash-rcu: fixed buckets + xxhash + RCU list + MRU. MRU is implemented here by adding an intermediate struct that contains the u32 hash and a pointer to the TB; this allows us, on an MRU promotion, to copy said struct (that is not at the head), and put this new copy at the head. After a grace period, the original non-head struct can be eliminated, and after another grace period, freed. - qht-fixed-nomru: fixed buckets + xxhash + qht without auto-resize + no MRU for lookups; MRU for inserts. The appended solution is the following: - qht-dyn-nomru: dynamic number of buckets + xxhash + qht w/ auto-resize + no MRU for lookups; MRU for inserts. The plots below compare the considered solutions. The Y axis shows the boot time (in seconds) of a debian jessie image with arm-softmmu; the X axis sweeps the number of buckets (or initial number of buckets for qht-autoresize). The plots in PNG format (and with errorbars) can be seen here: http://imgur.com/a/Awgnq Each test runs 5 times, and the entire QEMU process is pinned to a single core for repeatability of results. Host: Intel Xeon E5-2690 28 ++------------+-------------+-------------+-------------+------------++ A***** + + + master **A*** + 27 ++ * xxhash ##B###++ | A******A****** xxhash-rcu $$C$$$ | 26 C$$ A******A****** qht-fixed-nomru*%%D%%%++ D%%$$ A******A******A*qht-dyn-mru A*E****A 25 ++ %%$$ qht-dyn-nomru &&F&&&++ B#####% | 24 ++ #C$$$$$ ++ | B### $ | | ## C$$$$$$ | 23 ++ # C$$$$$$ ++ | B###### C$$$$$$ %%%D 22 ++ %B###### C$$$$$$C$$$$$$C$$$$$$C$$$$$$C$$$$$$C | D%%%%%%B###### @E@@@@@@ %%%D%%%@@@E@@@@@@E 21 E@@@@@@E@@@@@@F&&&@@@E@@@&&&D%%%%%%B######B######B######B######B######B + E@@@ F&&& + E@ + F&&& + + 20 ++------------+-------------+-------------+-------------+------------++ 14 16 18 20 22 24 log2 number of buckets Host: Intel i7-4790K 14.5 ++------------+------------+-------------+------------+------------++ A** + + + master **A*** + 14 ++ ** xxhash ##B###++ 13.5 ++ ** xxhash-rcu $$C$$$++ | qht-fixed-nomru %%D%%% | 13 ++ A****** qht-dyn-mru @@E@@@++ | A*****A******A****** qht-dyn-nomru &&F&&& | 12.5 C$$ A******A******A*****A****** ***A 12 ++ $$ A*** ++ D%%% $$ | 11.5 ++ %% ++ B### %C$$$$$$ | 11 ++ ## D%%%%% C$$$$$ ++ | # % C$$$$$$ | 10.5 F&&&&&&B######D%%%%% C$$$$$$C$$$$$$C$$$$$$C$$$$$C$$$$$$ $$$C 10 E@@@@@@E@@@@@@B#####B######B######E@@@@@@E@@@%%%D%%%%%D%%%###B######B + F&& D%%%%%%B######B######B#####B###@@@D%%% + 9.5 ++------------+------------+-------------+------------+------------++ 14 16 18 20 22 24 log2 number of buckets Note that the original point before this patch series is X=15 for "master"; the little sensitivity to the increased number of buckets is due to the poor hashing function in master. xxhash-rcu has significant overhead due to the constant churn of allocating and deallocating intermediate structs for implementing MRU. An alternative would be do consider failed lookups as "maybe not there", and then acquire the external lock (tb_lock in this case) to really confirm that there was indeed a failed lookup. This, however, would not be enough to implement dynamic resizing--this is more complex: see "Resizable, Scalable, Concurrent Hash Tables via Relativistic Programming" by Triplett, McKenney and Walpole. This solution was discarded due to the very coarse RCU read critical sections that we have in MTTCG; resizing requires waiting for readers after every pointer update, and resizes require many pointer updates, so this would quickly become prohibitive. qht-fixed-nomru shows that MRU promotion is advisable for undersized hash tables. However, qht-dyn-mru shows that MRU promotion is not important if the hash table is properly sized: there is virtually no difference in performance between qht-dyn-nomru and qht-dyn-mru. Before this patch, we're at X=15 on "xxhash"; after this patch, we're at X=15 @ qht-dyn-nomru. This patch thus matches the best performance that we can achieve with optimum sizing of the hash table, while keeping the hash table scalable for readers. The improvement we get before and after this patch for booting debian jessie with arm-softmmu is: - Intel Xeon E5-2690: 10.5% less time - Intel i7-4790K: 5.2% less time We could get this same improvement _for this particular workload_ by statically increasing the size of the hash table. But this would hurt workloads that do not need a large hash table. The dynamic (upward) resizing allows us to start small and enlarge the hash table as needed. A quick note on downsizing: the table is resized back to 2**15 buckets on every tb_flush; this makes sense because it is not guaranteed that the table will reach the same number of TBs later on (e.g. most bootup code is thrown away after boot); it makes sense to grow the hash table as more code blocks are translated. This also avoids the complication of having to build downsizing hysteresis logic into qht. Reviewed-by: Sergey Fedorov <serge.fedorov@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1465412133-3029-15-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2016-06-08 20:55:32 +02:00
tb_htable_init();
code_gen_alloc(tb_size);
#if defined(CONFIG_SOFTMMU)
/* There's no guest base to take into account, so go ahead and
initialize the prologue now. */
tcg_prologue_init(tcg_ctx);
#endif
}
/*
* Allocate a new translation block. Flush the translation buffer if
* too many translation blocks or too much generated code.
*/
static TranslationBlock *tb_alloc(target_ulong pc)
{
TranslationBlock *tb;
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
assert_memory_lock();
tb = tcg_tb_alloc(tcg_ctx);
tcg: allocate TB structs before the corresponding translated code Allocating an arbitrarily-sized array of tbs results in either (a) a lot of memory wasted or (b) unnecessary flushes of the code cache when we run out of TB structs in the array. An obvious solution would be to just malloc a TB struct when needed, and keep the TB array as an array of pointers (recall that tb_find_pc() needs the TB array to run in O(log n)). Perhaps a better solution, which is implemented in this patch, is to allocate TB's right before the translated code they describe. This results in some memory waste due to padding to have code and TBs in separate cache lines--for instance, I measured 4.7% of padding in the used portion of code_gen_buffer when booting aarch64 Linux on a host with 64-byte cache lines. However, it can allow for optimizations in some host architectures, since TCG backends could safely assume that the TB and the corresponding translated code are very close to each other in memory. See this message by rth for a detailed explanation: https://lists.gnu.org/archive/html/qemu-devel/2017-03/msg05172.html Subject: Re: GSoC 2017 Proposal: TCG performance enhancements Message-ID: <1e67644b-4b30-887e-d329-1848e94c9484@twiddle.net> Suggested-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Pranith Kumar <bobby.prani@gmail.com> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1496790745-314-3-git-send-email-cota@braap.org> [rth: Simplify the arithmetic in tcg_tb_alloc] Signed-off-by: Richard Henderson <rth@twiddle.net>
2017-06-07 01:12:25 +02:00
if (unlikely(tb == NULL)) {
return NULL;
}
return tb;
}
/* call with @p->lock held */
static inline void invalidate_page_bitmap(PageDesc *p)
{
assert_page_locked(p);
#ifdef CONFIG_SOFTMMU
g_free(p->code_bitmap);
p->code_bitmap = NULL;
p->code_write_count = 0;
#endif
}
/* Set to NULL all the 'first_tb' fields in all PageDescs. */
static void page_flush_tb_1(int level, void **lp)
{
int i;
if (*lp == NULL) {
return;
}
if (level == 0) {
PageDesc *pd = *lp;
for (i = 0; i < V_L2_SIZE; ++i) {
page_lock(&pd[i]);
pd[i].first_tb = (uintptr_t)NULL;
invalidate_page_bitmap(pd + i);
page_unlock(&pd[i]);
}
} else {
void **pp = *lp;
for (i = 0; i < V_L2_SIZE; ++i) {
page_flush_tb_1(level - 1, pp + i);
}
}
}
static void page_flush_tb(void)
{
int i, l1_sz = v_l1_size;
for (i = 0; i < l1_sz; i++) {
page_flush_tb_1(v_l2_levels, l1_map + i);
}
}
static gboolean tb_host_size_iter(gpointer key, gpointer value, gpointer data)
{
const TranslationBlock *tb = value;
size_t *size = data;
*size += tb->tc.size;
return false;
}
/* flush all the translation blocks */
static void do_tb_flush(CPUState *cpu, run_on_cpu_data tb_flush_count)
{
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
mmap_lock();
/* If it is already been done on request of another CPU,
* just retry.
*/
if (tb_ctx.tb_flush_count != tb_flush_count.host_int) {
goto done;
}
if (DEBUG_TB_FLUSH_GATE) {
tcg: track TBs with per-region BST's This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-26 22:58:05 +02:00
size_t nb_tbs = tcg_nb_tbs();
size_t host_size = 0;
translate-all: use a binary search tree to track TBs in TBContext This is a prerequisite for supporting multiple TCG contexts, since we will have threads generating code in separate regions of code_gen_buffer. For this we need a new field (.size) in struct tb_tc to keep track of the size of the translated code. This field uses a size_t to avoid adding a hole to the struct, although really an unsigned int would have been enough. The comparison function we use is optimized for the common case: insertions. Profiling shows that upon booting debian-arm, 98% of comparisons are between existing tb's (i.e. a->size and b->size are both !0), which happens during insertions (and removals, but those are rare). The remaining cases are lookups. From reading the glib sources we see that the first key is always the lookup key. However, the code does not assume this to always be the case because this behaviour is not guaranteed in the glib docs. However, we embed this knowledge in the code as a branch hint for the compiler. Note that tb_free does not free space in the code_gen_buffer anymore, since we cannot easily know whether the tb is the last one inserted in code_gen_buffer. The next patch in this series renames tb_free to tb_remove to reflect this. Performance-wise, lookups in tb_find_pc are the same as before: O(log n). However, insertions are O(log n) instead of O(1), which results in a small slowdown when booting debian-arm: Performance counter stats for 'build/arm-softmmu/qemu-system-arm \ -machine type=virt -nographic -smp 1 -m 4096 \ -netdev user,id=unet,hostfwd=tcp::2222-:22 \ -device virtio-net-device,netdev=unet \ -drive file=img/arm/jessie-arm32.qcow2,id=myblock,index=0,if=none \ -device virtio-blk-device,drive=myblock \ -kernel img/arm/aarch32-current-linux-kernel-only.img \ -append console=ttyAMA0 root=/dev/vda1 \ -name arm,debug-threads=on -smp 1' (10 runs): - Before: 8048.598422 task-clock (msec) # 0.931 CPUs utilized ( +- 0.28% ) 16,974 context-switches # 0.002 M/sec ( +- 0.12% ) 0 cpu-migrations # 0.000 K/sec 10,125 page-faults # 0.001 M/sec ( +- 1.23% ) 35,144,901,879 cycles # 4.367 GHz ( +- 0.14% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,758,252,643 instructions # 1.87 insns per cycle ( +- 0.33% ) 10,871,298,668 branches # 1350.707 M/sec ( +- 0.41% ) 192,322,212 branch-misses # 1.77% of all branches ( +- 0.32% ) 8.640869419 seconds time elapsed ( +- 0.57% ) - After: 8146.242027 task-clock (msec) # 0.923 CPUs utilized ( +- 1.23% ) 17,016 context-switches # 0.002 M/sec ( +- 0.40% ) 0 cpu-migrations # 0.000 K/sec 18,769 page-faults # 0.002 M/sec ( +- 0.45% ) 35,660,956,120 cycles # 4.378 GHz ( +- 1.22% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,095,366,607 instructions # 1.83 insns per cycle ( +- 1.73% ) 10,803,480,261 branches # 1326.192 M/sec ( +- 1.95% ) 195,601,289 branch-misses # 1.81% of all branches ( +- 0.39% ) 8.828660235 seconds time elapsed ( +- 0.38% ) Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-06-24 01:00:11 +02:00
tcg: track TBs with per-region BST's This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-26 22:58:05 +02:00
tcg_tb_foreach(tb_host_size_iter, &host_size);
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
printf("qemu: flush code_size=%zu nb_tbs=%zu avg_tb_size=%zu\n",
tcg_code_size(), nb_tbs, nb_tbs > 0 ? host_size / nb_tbs : 0);
}
CPU_FOREACH(cpu) {
tcg: consistently access cpu->tb_jmp_cache atomically Some code paths can lead to atomic accesses racing with memset() on cpu->tb_jmp_cache, which can result in torn reads/writes and is undefined behaviour in C11. These torn accesses are unlikely to show up as bugs, but from code inspection they seem possible. For example, tb_phys_invalidate does: /* remove the TB from the hash list */ h = tb_jmp_cache_hash_func(tb->pc); CPU_FOREACH(cpu) { if (atomic_read(&cpu->tb_jmp_cache[h]) == tb) { atomic_set(&cpu->tb_jmp_cache[h], NULL); } } Here atomic_set might race with a concurrent memset (such as the ones scheduled via "unsafe" async work, e.g. tlb_flush_page) and therefore we might end up with a torn pointer (or who knows what, because we are under undefined behaviour). This patch converts parallel accesses to cpu->tb_jmp_cache to use atomic primitives, thereby bringing these accesses back to defined behaviour. The price to pay is to potentially execute more instructions when clearing cpu->tb_jmp_cache, but given how infrequently they happen and the small size of the cache, the performance impact I have measured is within noise range when booting debian-arm. Note that under "safe async" work (e.g. do_tb_flush) we could use memset because no other vcpus are running. However I'm keeping these accesses atomic as well to keep things simple and to avoid confusing analysis tools such as ThreadSanitizer. Reviewed-by: Paolo Bonzini <pbonzini@redhat.com> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1497486973-25845-1-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2017-06-15 02:36:13 +02:00
cpu_tb_jmp_cache_clear(cpu);
}
qht_reset_size(&tb_ctx.htable, CODE_GEN_HTABLE_SIZE);
page_flush_tb();
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
tcg_region_reset_all();
/* XXX: flush processor icache at this point if cache flush is
expensive */
atomic_mb_set(&tb_ctx.tb_flush_count, tb_ctx.tb_flush_count + 1);
done:
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
mmap_unlock();
}
void tb_flush(CPUState *cpu)
{
if (tcg_enabled()) {
unsigned tb_flush_count = atomic_mb_read(&tb_ctx.tb_flush_count);
async_safe_run_on_cpu(cpu, do_tb_flush,
RUN_ON_CPU_HOST_INT(tb_flush_count));
}
}
/*
* Formerly ifdef DEBUG_TB_CHECK. These debug functions are user-mode-only,
* so in order to prevent bit rot we compile them unconditionally in user-mode,
* and let the optimizer get rid of them by wrapping their user-only callers
* with if (DEBUG_TB_CHECK_GATE).
*/
#ifdef CONFIG_USER_ONLY
tb hash: track translated blocks with qht Having a fixed-size hash table for keeping track of all translation blocks is suboptimal: some workloads are just too big or too small to get maximum performance from the hash table. The MRU promotion policy helps improve performance when the hash table is a little undersized, but it cannot make up for severely undersized hash tables. Furthermore, frequent MRU promotions result in writes that are a scalability bottleneck. For scalability, lookups should only perform reads, not writes. This is not a big deal for now, but it will become one once MTTCG matures. The appended fixes these issues by using qht as the implementation of the TB hash table. This solution is superior to other alternatives considered, namely: - master: implementation in QEMU before this patchset - xxhash: before this patch, i.e. fixed buckets + xxhash hashing + MRU. - xxhash-rcu: fixed buckets + xxhash + RCU list + MRU. MRU is implemented here by adding an intermediate struct that contains the u32 hash and a pointer to the TB; this allows us, on an MRU promotion, to copy said struct (that is not at the head), and put this new copy at the head. After a grace period, the original non-head struct can be eliminated, and after another grace period, freed. - qht-fixed-nomru: fixed buckets + xxhash + qht without auto-resize + no MRU for lookups; MRU for inserts. The appended solution is the following: - qht-dyn-nomru: dynamic number of buckets + xxhash + qht w/ auto-resize + no MRU for lookups; MRU for inserts. The plots below compare the considered solutions. The Y axis shows the boot time (in seconds) of a debian jessie image with arm-softmmu; the X axis sweeps the number of buckets (or initial number of buckets for qht-autoresize). The plots in PNG format (and with errorbars) can be seen here: http://imgur.com/a/Awgnq Each test runs 5 times, and the entire QEMU process is pinned to a single core for repeatability of results. Host: Intel Xeon E5-2690 28 ++------------+-------------+-------------+-------------+------------++ A***** + + + master **A*** + 27 ++ * xxhash ##B###++ | A******A****** xxhash-rcu $$C$$$ | 26 C$$ A******A****** qht-fixed-nomru*%%D%%%++ D%%$$ A******A******A*qht-dyn-mru A*E****A 25 ++ %%$$ qht-dyn-nomru &&F&&&++ B#####% | 24 ++ #C$$$$$ ++ | B### $ | | ## C$$$$$$ | 23 ++ # C$$$$$$ ++ | B###### C$$$$$$ %%%D 22 ++ %B###### C$$$$$$C$$$$$$C$$$$$$C$$$$$$C$$$$$$C | D%%%%%%B###### @E@@@@@@ %%%D%%%@@@E@@@@@@E 21 E@@@@@@E@@@@@@F&&&@@@E@@@&&&D%%%%%%B######B######B######B######B######B + E@@@ F&&& + E@ + F&&& + + 20 ++------------+-------------+-------------+-------------+------------++ 14 16 18 20 22 24 log2 number of buckets Host: Intel i7-4790K 14.5 ++------------+------------+-------------+------------+------------++ A** + + + master **A*** + 14 ++ ** xxhash ##B###++ 13.5 ++ ** xxhash-rcu $$C$$$++ | qht-fixed-nomru %%D%%% | 13 ++ A****** qht-dyn-mru @@E@@@++ | A*****A******A****** qht-dyn-nomru &&F&&& | 12.5 C$$ A******A******A*****A****** ***A 12 ++ $$ A*** ++ D%%% $$ | 11.5 ++ %% ++ B### %C$$$$$$ | 11 ++ ## D%%%%% C$$$$$ ++ | # % C$$$$$$ | 10.5 F&&&&&&B######D%%%%% C$$$$$$C$$$$$$C$$$$$$C$$$$$C$$$$$$ $$$C 10 E@@@@@@E@@@@@@B#####B######B######E@@@@@@E@@@%%%D%%%%%D%%%###B######B + F&& D%%%%%%B######B######B#####B###@@@D%%% + 9.5 ++------------+------------+-------------+------------+------------++ 14 16 18 20 22 24 log2 number of buckets Note that the original point before this patch series is X=15 for "master"; the little sensitivity to the increased number of buckets is due to the poor hashing function in master. xxhash-rcu has significant overhead due to the constant churn of allocating and deallocating intermediate structs for implementing MRU. An alternative would be do consider failed lookups as "maybe not there", and then acquire the external lock (tb_lock in this case) to really confirm that there was indeed a failed lookup. This, however, would not be enough to implement dynamic resizing--this is more complex: see "Resizable, Scalable, Concurrent Hash Tables via Relativistic Programming" by Triplett, McKenney and Walpole. This solution was discarded due to the very coarse RCU read critical sections that we have in MTTCG; resizing requires waiting for readers after every pointer update, and resizes require many pointer updates, so this would quickly become prohibitive. qht-fixed-nomru shows that MRU promotion is advisable for undersized hash tables. However, qht-dyn-mru shows that MRU promotion is not important if the hash table is properly sized: there is virtually no difference in performance between qht-dyn-nomru and qht-dyn-mru. Before this patch, we're at X=15 on "xxhash"; after this patch, we're at X=15 @ qht-dyn-nomru. This patch thus matches the best performance that we can achieve with optimum sizing of the hash table, while keeping the hash table scalable for readers. The improvement we get before and after this patch for booting debian jessie with arm-softmmu is: - Intel Xeon E5-2690: 10.5% less time - Intel i7-4790K: 5.2% less time We could get this same improvement _for this particular workload_ by statically increasing the size of the hash table. But this would hurt workloads that do not need a large hash table. The dynamic (upward) resizing allows us to start small and enlarge the hash table as needed. A quick note on downsizing: the table is resized back to 2**15 buckets on every tb_flush; this makes sense because it is not guaranteed that the table will reach the same number of TBs later on (e.g. most bootup code is thrown away after boot); it makes sense to grow the hash table as more code blocks are translated. This also avoids the complication of having to build downsizing hysteresis logic into qht. Reviewed-by: Sergey Fedorov <serge.fedorov@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1465412133-3029-15-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2016-06-08 20:55:32 +02:00
static void
do_tb_invalidate_check(struct qht *ht, void *p, uint32_t hash, void *userp)
{
tb hash: track translated blocks with qht Having a fixed-size hash table for keeping track of all translation blocks is suboptimal: some workloads are just too big or too small to get maximum performance from the hash table. The MRU promotion policy helps improve performance when the hash table is a little undersized, but it cannot make up for severely undersized hash tables. Furthermore, frequent MRU promotions result in writes that are a scalability bottleneck. For scalability, lookups should only perform reads, not writes. This is not a big deal for now, but it will become one once MTTCG matures. The appended fixes these issues by using qht as the implementation of the TB hash table. This solution is superior to other alternatives considered, namely: - master: implementation in QEMU before this patchset - xxhash: before this patch, i.e. fixed buckets + xxhash hashing + MRU. - xxhash-rcu: fixed buckets + xxhash + RCU list + MRU. MRU is implemented here by adding an intermediate struct that contains the u32 hash and a pointer to the TB; this allows us, on an MRU promotion, to copy said struct (that is not at the head), and put this new copy at the head. After a grace period, the original non-head struct can be eliminated, and after another grace period, freed. - qht-fixed-nomru: fixed buckets + xxhash + qht without auto-resize + no MRU for lookups; MRU for inserts. The appended solution is the following: - qht-dyn-nomru: dynamic number of buckets + xxhash + qht w/ auto-resize + no MRU for lookups; MRU for inserts. The plots below compare the considered solutions. The Y axis shows the boot time (in seconds) of a debian jessie image with arm-softmmu; the X axis sweeps the number of buckets (or initial number of buckets for qht-autoresize). The plots in PNG format (and with errorbars) can be seen here: http://imgur.com/a/Awgnq Each test runs 5 times, and the entire QEMU process is pinned to a single core for repeatability of results. Host: Intel Xeon E5-2690 28 ++------------+-------------+-------------+-------------+------------++ A***** + + + master **A*** + 27 ++ * xxhash ##B###++ | A******A****** xxhash-rcu $$C$$$ | 26 C$$ A******A****** qht-fixed-nomru*%%D%%%++ D%%$$ A******A******A*qht-dyn-mru A*E****A 25 ++ %%$$ qht-dyn-nomru &&F&&&++ B#####% | 24 ++ #C$$$$$ ++ | B### $ | | ## C$$$$$$ | 23 ++ # C$$$$$$ ++ | B###### C$$$$$$ %%%D 22 ++ %B###### C$$$$$$C$$$$$$C$$$$$$C$$$$$$C$$$$$$C | D%%%%%%B###### @E@@@@@@ %%%D%%%@@@E@@@@@@E 21 E@@@@@@E@@@@@@F&&&@@@E@@@&&&D%%%%%%B######B######B######B######B######B + E@@@ F&&& + E@ + F&&& + + 20 ++------------+-------------+-------------+-------------+------------++ 14 16 18 20 22 24 log2 number of buckets Host: Intel i7-4790K 14.5 ++------------+------------+-------------+------------+------------++ A** + + + master **A*** + 14 ++ ** xxhash ##B###++ 13.5 ++ ** xxhash-rcu $$C$$$++ | qht-fixed-nomru %%D%%% | 13 ++ A****** qht-dyn-mru @@E@@@++ | A*****A******A****** qht-dyn-nomru &&F&&& | 12.5 C$$ A******A******A*****A****** ***A 12 ++ $$ A*** ++ D%%% $$ | 11.5 ++ %% ++ B### %C$$$$$$ | 11 ++ ## D%%%%% C$$$$$ ++ | # % C$$$$$$ | 10.5 F&&&&&&B######D%%%%% C$$$$$$C$$$$$$C$$$$$$C$$$$$C$$$$$$ $$$C 10 E@@@@@@E@@@@@@B#####B######B######E@@@@@@E@@@%%%D%%%%%D%%%###B######B + F&& D%%%%%%B######B######B#####B###@@@D%%% + 9.5 ++------------+------------+-------------+------------+------------++ 14 16 18 20 22 24 log2 number of buckets Note that the original point before this patch series is X=15 for "master"; the little sensitivity to the increased number of buckets is due to the poor hashing function in master. xxhash-rcu has significant overhead due to the constant churn of allocating and deallocating intermediate structs for implementing MRU. An alternative would be do consider failed lookups as "maybe not there", and then acquire the external lock (tb_lock in this case) to really confirm that there was indeed a failed lookup. This, however, would not be enough to implement dynamic resizing--this is more complex: see "Resizable, Scalable, Concurrent Hash Tables via Relativistic Programming" by Triplett, McKenney and Walpole. This solution was discarded due to the very coarse RCU read critical sections that we have in MTTCG; resizing requires waiting for readers after every pointer update, and resizes require many pointer updates, so this would quickly become prohibitive. qht-fixed-nomru shows that MRU promotion is advisable for undersized hash tables. However, qht-dyn-mru shows that MRU promotion is not important if the hash table is properly sized: there is virtually no difference in performance between qht-dyn-nomru and qht-dyn-mru. Before this patch, we're at X=15 on "xxhash"; after this patch, we're at X=15 @ qht-dyn-nomru. This patch thus matches the best performance that we can achieve with optimum sizing of the hash table, while keeping the hash table scalable for readers. The improvement we get before and after this patch for booting debian jessie with arm-softmmu is: - Intel Xeon E5-2690: 10.5% less time - Intel i7-4790K: 5.2% less time We could get this same improvement _for this particular workload_ by statically increasing the size of the hash table. But this would hurt workloads that do not need a large hash table. The dynamic (upward) resizing allows us to start small and enlarge the hash table as needed. A quick note on downsizing: the table is resized back to 2**15 buckets on every tb_flush; this makes sense because it is not guaranteed that the table will reach the same number of TBs later on (e.g. most bootup code is thrown away after boot); it makes sense to grow the hash table as more code blocks are translated. This also avoids the complication of having to build downsizing hysteresis logic into qht. Reviewed-by: Sergey Fedorov <serge.fedorov@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1465412133-3029-15-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2016-06-08 20:55:32 +02:00
TranslationBlock *tb = p;
target_ulong addr = *(target_ulong *)userp;
if (!(addr + TARGET_PAGE_SIZE <= tb->pc || addr >= tb->pc + tb->size)) {
printf("ERROR invalidate: address=" TARGET_FMT_lx
" PC=%08lx size=%04x\n", addr, (long)tb->pc, tb->size);
}
}
/* verify that all the pages have correct rights for code
*
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
* Called with mmap_lock held.
*/
tb hash: track translated blocks with qht Having a fixed-size hash table for keeping track of all translation blocks is suboptimal: some workloads are just too big or too small to get maximum performance from the hash table. The MRU promotion policy helps improve performance when the hash table is a little undersized, but it cannot make up for severely undersized hash tables. Furthermore, frequent MRU promotions result in writes that are a scalability bottleneck. For scalability, lookups should only perform reads, not writes. This is not a big deal for now, but it will become one once MTTCG matures. The appended fixes these issues by using qht as the implementation of the TB hash table. This solution is superior to other alternatives considered, namely: - master: implementation in QEMU before this patchset - xxhash: before this patch, i.e. fixed buckets + xxhash hashing + MRU. - xxhash-rcu: fixed buckets + xxhash + RCU list + MRU. MRU is implemented here by adding an intermediate struct that contains the u32 hash and a pointer to the TB; this allows us, on an MRU promotion, to copy said struct (that is not at the head), and put this new copy at the head. After a grace period, the original non-head struct can be eliminated, and after another grace period, freed. - qht-fixed-nomru: fixed buckets + xxhash + qht without auto-resize + no MRU for lookups; MRU for inserts. The appended solution is the following: - qht-dyn-nomru: dynamic number of buckets + xxhash + qht w/ auto-resize + no MRU for lookups; MRU for inserts. The plots below compare the considered solutions. The Y axis shows the boot time (in seconds) of a debian jessie image with arm-softmmu; the X axis sweeps the number of buckets (or initial number of buckets for qht-autoresize). The plots in PNG format (and with errorbars) can be seen here: http://imgur.com/a/Awgnq Each test runs 5 times, and the entire QEMU process is pinned to a single core for repeatability of results. Host: Intel Xeon E5-2690 28 ++------------+-------------+-------------+-------------+------------++ A***** + + + master **A*** + 27 ++ * xxhash ##B###++ | A******A****** xxhash-rcu $$C$$$ | 26 C$$ A******A****** qht-fixed-nomru*%%D%%%++ D%%$$ A******A******A*qht-dyn-mru A*E****A 25 ++ %%$$ qht-dyn-nomru &&F&&&++ B#####% | 24 ++ #C$$$$$ ++ | B### $ | | ## C$$$$$$ | 23 ++ # C$$$$$$ ++ | B###### C$$$$$$ %%%D 22 ++ %B###### C$$$$$$C$$$$$$C$$$$$$C$$$$$$C$$$$$$C | D%%%%%%B###### @E@@@@@@ %%%D%%%@@@E@@@@@@E 21 E@@@@@@E@@@@@@F&&&@@@E@@@&&&D%%%%%%B######B######B######B######B######B + E@@@ F&&& + E@ + F&&& + + 20 ++------------+-------------+-------------+-------------+------------++ 14 16 18 20 22 24 log2 number of buckets Host: Intel i7-4790K 14.5 ++------------+------------+-------------+------------+------------++ A** + + + master **A*** + 14 ++ ** xxhash ##B###++ 13.5 ++ ** xxhash-rcu $$C$$$++ | qht-fixed-nomru %%D%%% | 13 ++ A****** qht-dyn-mru @@E@@@++ | A*****A******A****** qht-dyn-nomru &&F&&& | 12.5 C$$ A******A******A*****A****** ***A 12 ++ $$ A*** ++ D%%% $$ | 11.5 ++ %% ++ B### %C$$$$$$ | 11 ++ ## D%%%%% C$$$$$ ++ | # % C$$$$$$ | 10.5 F&&&&&&B######D%%%%% C$$$$$$C$$$$$$C$$$$$$C$$$$$C$$$$$$ $$$C 10 E@@@@@@E@@@@@@B#####B######B######E@@@@@@E@@@%%%D%%%%%D%%%###B######B + F&& D%%%%%%B######B######B#####B###@@@D%%% + 9.5 ++------------+------------+-------------+------------+------------++ 14 16 18 20 22 24 log2 number of buckets Note that the original point before this patch series is X=15 for "master"; the little sensitivity to the increased number of buckets is due to the poor hashing function in master. xxhash-rcu has significant overhead due to the constant churn of allocating and deallocating intermediate structs for implementing MRU. An alternative would be do consider failed lookups as "maybe not there", and then acquire the external lock (tb_lock in this case) to really confirm that there was indeed a failed lookup. This, however, would not be enough to implement dynamic resizing--this is more complex: see "Resizable, Scalable, Concurrent Hash Tables via Relativistic Programming" by Triplett, McKenney and Walpole. This solution was discarded due to the very coarse RCU read critical sections that we have in MTTCG; resizing requires waiting for readers after every pointer update, and resizes require many pointer updates, so this would quickly become prohibitive. qht-fixed-nomru shows that MRU promotion is advisable for undersized hash tables. However, qht-dyn-mru shows that MRU promotion is not important if the hash table is properly sized: there is virtually no difference in performance between qht-dyn-nomru and qht-dyn-mru. Before this patch, we're at X=15 on "xxhash"; after this patch, we're at X=15 @ qht-dyn-nomru. This patch thus matches the best performance that we can achieve with optimum sizing of the hash table, while keeping the hash table scalable for readers. The improvement we get before and after this patch for booting debian jessie with arm-softmmu is: - Intel Xeon E5-2690: 10.5% less time - Intel i7-4790K: 5.2% less time We could get this same improvement _for this particular workload_ by statically increasing the size of the hash table. But this would hurt workloads that do not need a large hash table. The dynamic (upward) resizing allows us to start small and enlarge the hash table as needed. A quick note on downsizing: the table is resized back to 2**15 buckets on every tb_flush; this makes sense because it is not guaranteed that the table will reach the same number of TBs later on (e.g. most bootup code is thrown away after boot); it makes sense to grow the hash table as more code blocks are translated. This also avoids the complication of having to build downsizing hysteresis logic into qht. Reviewed-by: Sergey Fedorov <serge.fedorov@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1465412133-3029-15-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2016-06-08 20:55:32 +02:00
static void tb_invalidate_check(target_ulong address)
{
address &= TARGET_PAGE_MASK;
qht_iter(&tb_ctx.htable, do_tb_invalidate_check, &address);
tb hash: track translated blocks with qht Having a fixed-size hash table for keeping track of all translation blocks is suboptimal: some workloads are just too big or too small to get maximum performance from the hash table. The MRU promotion policy helps improve performance when the hash table is a little undersized, but it cannot make up for severely undersized hash tables. Furthermore, frequent MRU promotions result in writes that are a scalability bottleneck. For scalability, lookups should only perform reads, not writes. This is not a big deal for now, but it will become one once MTTCG matures. The appended fixes these issues by using qht as the implementation of the TB hash table. This solution is superior to other alternatives considered, namely: - master: implementation in QEMU before this patchset - xxhash: before this patch, i.e. fixed buckets + xxhash hashing + MRU. - xxhash-rcu: fixed buckets + xxhash + RCU list + MRU. MRU is implemented here by adding an intermediate struct that contains the u32 hash and a pointer to the TB; this allows us, on an MRU promotion, to copy said struct (that is not at the head), and put this new copy at the head. After a grace period, the original non-head struct can be eliminated, and after another grace period, freed. - qht-fixed-nomru: fixed buckets + xxhash + qht without auto-resize + no MRU for lookups; MRU for inserts. The appended solution is the following: - qht-dyn-nomru: dynamic number of buckets + xxhash + qht w/ auto-resize + no MRU for lookups; MRU for inserts. The plots below compare the considered solutions. The Y axis shows the boot time (in seconds) of a debian jessie image with arm-softmmu; the X axis sweeps the number of buckets (or initial number of buckets for qht-autoresize). The plots in PNG format (and with errorbars) can be seen here: http://imgur.com/a/Awgnq Each test runs 5 times, and the entire QEMU process is pinned to a single core for repeatability of results. Host: Intel Xeon E5-2690 28 ++------------+-------------+-------------+-------------+------------++ A***** + + + master **A*** + 27 ++ * xxhash ##B###++ | A******A****** xxhash-rcu $$C$$$ | 26 C$$ A******A****** qht-fixed-nomru*%%D%%%++ D%%$$ A******A******A*qht-dyn-mru A*E****A 25 ++ %%$$ qht-dyn-nomru &&F&&&++ B#####% | 24 ++ #C$$$$$ ++ | B### $ | | ## C$$$$$$ | 23 ++ # C$$$$$$ ++ | B###### C$$$$$$ %%%D 22 ++ %B###### C$$$$$$C$$$$$$C$$$$$$C$$$$$$C$$$$$$C | D%%%%%%B###### @E@@@@@@ %%%D%%%@@@E@@@@@@E 21 E@@@@@@E@@@@@@F&&&@@@E@@@&&&D%%%%%%B######B######B######B######B######B + E@@@ F&&& + E@ + F&&& + + 20 ++------------+-------------+-------------+-------------+------------++ 14 16 18 20 22 24 log2 number of buckets Host: Intel i7-4790K 14.5 ++------------+------------+-------------+------------+------------++ A** + + + master **A*** + 14 ++ ** xxhash ##B###++ 13.5 ++ ** xxhash-rcu $$C$$$++ | qht-fixed-nomru %%D%%% | 13 ++ A****** qht-dyn-mru @@E@@@++ | A*****A******A****** qht-dyn-nomru &&F&&& | 12.5 C$$ A******A******A*****A****** ***A 12 ++ $$ A*** ++ D%%% $$ | 11.5 ++ %% ++ B### %C$$$$$$ | 11 ++ ## D%%%%% C$$$$$ ++ | # % C$$$$$$ | 10.5 F&&&&&&B######D%%%%% C$$$$$$C$$$$$$C$$$$$$C$$$$$C$$$$$$ $$$C 10 E@@@@@@E@@@@@@B#####B######B######E@@@@@@E@@@%%%D%%%%%D%%%###B######B + F&& D%%%%%%B######B######B#####B###@@@D%%% + 9.5 ++------------+------------+-------------+------------+------------++ 14 16 18 20 22 24 log2 number of buckets Note that the original point before this patch series is X=15 for "master"; the little sensitivity to the increased number of buckets is due to the poor hashing function in master. xxhash-rcu has significant overhead due to the constant churn of allocating and deallocating intermediate structs for implementing MRU. An alternative would be do consider failed lookups as "maybe not there", and then acquire the external lock (tb_lock in this case) to really confirm that there was indeed a failed lookup. This, however, would not be enough to implement dynamic resizing--this is more complex: see "Resizable, Scalable, Concurrent Hash Tables via Relativistic Programming" by Triplett, McKenney and Walpole. This solution was discarded due to the very coarse RCU read critical sections that we have in MTTCG; resizing requires waiting for readers after every pointer update, and resizes require many pointer updates, so this would quickly become prohibitive. qht-fixed-nomru shows that MRU promotion is advisable for undersized hash tables. However, qht-dyn-mru shows that MRU promotion is not important if the hash table is properly sized: there is virtually no difference in performance between qht-dyn-nomru and qht-dyn-mru. Before this patch, we're at X=15 on "xxhash"; after this patch, we're at X=15 @ qht-dyn-nomru. This patch thus matches the best performance that we can achieve with optimum sizing of the hash table, while keeping the hash table scalable for readers. The improvement we get before and after this patch for booting debian jessie with arm-softmmu is: - Intel Xeon E5-2690: 10.5% less time - Intel i7-4790K: 5.2% less time We could get this same improvement _for this particular workload_ by statically increasing the size of the hash table. But this would hurt workloads that do not need a large hash table. The dynamic (upward) resizing allows us to start small and enlarge the hash table as needed. A quick note on downsizing: the table is resized back to 2**15 buckets on every tb_flush; this makes sense because it is not guaranteed that the table will reach the same number of TBs later on (e.g. most bootup code is thrown away after boot); it makes sense to grow the hash table as more code blocks are translated. This also avoids the complication of having to build downsizing hysteresis logic into qht. Reviewed-by: Sergey Fedorov <serge.fedorov@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1465412133-3029-15-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2016-06-08 20:55:32 +02:00
}
static void
do_tb_page_check(struct qht *ht, void *p, uint32_t hash, void *userp)
{
TranslationBlock *tb = p;
int flags1, flags2;
flags1 = page_get_flags(tb->pc);
flags2 = page_get_flags(tb->pc + tb->size - 1);
if ((flags1 & PAGE_WRITE) || (flags2 & PAGE_WRITE)) {
printf("ERROR page flags: PC=%08lx size=%04x f1=%x f2=%x\n",
(long)tb->pc, tb->size, flags1, flags2);
}
}
/* verify that all the pages have correct rights for code */
static void tb_page_check(void)
{
qht_iter(&tb_ctx.htable, do_tb_page_check, NULL);
}
#endif /* CONFIG_USER_ONLY */
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
/*
* user-mode: call with mmap_lock held
* !user-mode: call with @pd->lock held
*/
static inline void tb_page_remove(PageDesc *pd, TranslationBlock *tb)
{
TranslationBlock *tb1;
uintptr_t *pprev;
unsigned int n1;
assert_page_locked(pd);
pprev = &pd->first_tb;
PAGE_FOR_EACH_TB(pd, tb1, n1) {
if (tb1 == tb) {
*pprev = tb1->page_next[n1];
return;
}
pprev = &tb1->page_next[n1];
}
g_assert_not_reached();
}
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
/* remove @orig from its @n_orig-th jump list */
static inline void tb_remove_from_jmp_list(TranslationBlock *orig, int n_orig)
{
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
uintptr_t ptr, ptr_locked;
TranslationBlock *dest;
TranslationBlock *tb;
uintptr_t *pprev;
int n;
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
/* mark the LSB of jmp_dest[] so that no further jumps can be inserted */
ptr = atomic_or_fetch(&orig->jmp_dest[n_orig], 1);
dest = (TranslationBlock *)(ptr & ~1);
if (dest == NULL) {
return;
}
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
qemu_spin_lock(&dest->jmp_lock);
/*
* While acquiring the lock, the jump might have been removed if the
* destination TB was invalidated; check again.
*/
ptr_locked = atomic_read(&orig->jmp_dest[n_orig]);
if (ptr_locked != ptr) {
qemu_spin_unlock(&dest->jmp_lock);
/*
* The only possibility is that the jump was unlinked via
* tb_jump_unlink(dest). Seeing here another destination would be a bug,
* because we set the LSB above.
*/
g_assert(ptr_locked == 1 && dest->cflags & CF_INVALID);
return;
}
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
/*
* We first acquired the lock, and since the destination pointer matches,
* we know for sure that @orig is in the jmp list.
*/
pprev = &dest->jmp_list_head;
TB_FOR_EACH_JMP(dest, tb, n) {
if (tb == orig && n == n_orig) {
*pprev = tb->jmp_list_next[n];
/* no need to set orig->jmp_dest[n]; setting the LSB was enough */
qemu_spin_unlock(&dest->jmp_lock);
return;
}
pprev = &tb->jmp_list_next[n];
}
g_assert_not_reached();
}
/* reset the jump entry 'n' of a TB so that it is not chained to
another TB */
static inline void tb_reset_jump(TranslationBlock *tb, int n)
{
uintptr_t addr = (uintptr_t)(tb->tc.ptr + tb->jmp_reset_offset[n]);
tb_set_jmp_target(tb, n, addr);
}
/* remove any jumps to the TB */
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
static inline void tb_jmp_unlink(TranslationBlock *dest)
{
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
TranslationBlock *tb;
int n;
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
qemu_spin_lock(&dest->jmp_lock);
TB_FOR_EACH_JMP(dest, tb, n) {
tb_reset_jump(tb, n);
atomic_and(&tb->jmp_dest[n], (uintptr_t)NULL | 1);
/* No need to clear the list entry; setting the dest ptr is enough */
}
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
dest->jmp_list_head = (uintptr_t)NULL;
qemu_spin_unlock(&dest->jmp_lock);
}
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
/*
* In user-mode, call with mmap_lock held.
* In !user-mode, if @rm_from_page_list is set, call with the TB's pages'
* locks held.
*/
static void do_tb_phys_invalidate(TranslationBlock *tb, bool rm_from_page_list)
{
CPUState *cpu;
PageDesc *p;
tb hash: hash phys_pc, pc, and flags with xxhash For some workloads such as arm bootup, tb_phys_hash is performance-critical. The is due to the high frequency of accesses to the hash table, originated by (frequent) TLB flushes that wipe out the cpu-private tb_jmp_cache's. More info: https://lists.nongnu.org/archive/html/qemu-devel/2016-03/msg05098.html To dig further into this I modified an arm image booting debian jessie to immediately shut down after boot. Analysis revealed that quite a bit of time is unnecessarily spent in tb_phys_hash: the cause is poor hashing that results in very uneven loading of chains in the hash table's buckets; the longest observed chain had ~550 elements. The appended addresses this with two changes: 1) Use xxhash as the hash table's hash function. xxhash is a fast, high-quality hashing function. 2) Feed the hashing function with not just tb_phys, but also pc and flags. This improves performance over using just tb_phys for hashing, since that resulted in some hash buckets having many TB's, while others getting very few; with these changes, the longest observed chain on a single hash bucket is brought down from ~550 to ~40. Tests show that the other element checked for in tb_find_physical, cs_base, is always a match when tb_phys+pc+flags are a match, so hashing cs_base is wasteful. It could be that this is an ARM-only thing, though. UPDATE: On Tue, Apr 05, 2016 at 08:41:43 -0700, Richard Henderson wrote: > The cs_base field is only used by i386 (in 16-bit modes), and sparc (for a TB > consisting of only a delay slot). > It may well still turn out to be reasonable to ignore cs_base for hashing. BTW, after this change the hash table should not be called "tb_hash_phys" anymore; this is addressed later in this series. This change gives consistent bootup time improvements. I tested two host machines: - Intel Xeon E5-2690: 11.6% less time - Intel i7-4790K: 19.2% less time Increasing the number of hash buckets yields further improvements. However, using a larger, fixed number of buckets can degrade performance for other workloads that do not translate as many blocks (600K+ for debian-jessie arm bootup). This is dealt with later in this series. Reviewed-by: Sergey Fedorov <sergey.fedorov@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1465412133-3029-8-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2016-06-08 20:55:25 +02:00
uint32_t h;
tb_page_addr_t phys_pc;
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
assert_memory_lock();
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
/* make sure no further incoming jumps will be chained to this TB */
qemu_spin_lock(&tb->jmp_lock);
atomic_set(&tb->cflags, tb->cflags | CF_INVALID);
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
qemu_spin_unlock(&tb->jmp_lock);
/* remove the TB from the hash list */
phys_pc = tb->page_addr[0] + (tb->pc & ~TARGET_PAGE_MASK);
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
h = tb_hash_func(phys_pc, tb->pc, tb->flags, tb_cflags(tb) & CF_HASH_MASK,
tb->trace_vcpu_dstate);
if (!(tb->cflags & CF_NOCACHE) &&
!qht_remove(&tb_ctx.htable, tb, h)) {
return;
}
/* remove the TB from the page list */
if (rm_from_page_list) {
p = page_find(tb->page_addr[0] >> TARGET_PAGE_BITS);
tb_page_remove(p, tb);
invalidate_page_bitmap(p);
if (tb->page_addr[1] != -1) {
p = page_find(tb->page_addr[1] >> TARGET_PAGE_BITS);
tb_page_remove(p, tb);
invalidate_page_bitmap(p);
}
}
/* remove the TB from the hash list */
h = tb_jmp_cache_hash_func(tb->pc);
CPU_FOREACH(cpu) {
if (atomic_read(&cpu->tb_jmp_cache[h]) == tb) {
atomic_set(&cpu->tb_jmp_cache[h], NULL);
}
}
/* suppress this TB from the two jump lists */
tb_remove_from_jmp_list(tb, 0);
tb_remove_from_jmp_list(tb, 1);
/* suppress any remaining jumps to this TB */
tb_jmp_unlink(tb);
atomic_set(&tcg_ctx->tb_phys_invalidate_count,
tcg_ctx->tb_phys_invalidate_count + 1);
}
static void tb_phys_invalidate__locked(TranslationBlock *tb)
{
do_tb_phys_invalidate(tb, true);
}
/* invalidate one TB
*
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
* Called with mmap_lock held in user-mode.
*/
void tb_phys_invalidate(TranslationBlock *tb, tb_page_addr_t page_addr)
{
if (page_addr == -1) {
page_lock_tb(tb);
do_tb_phys_invalidate(tb, true);
page_unlock_tb(tb);
} else {
do_tb_phys_invalidate(tb, false);
}
}
#ifdef CONFIG_SOFTMMU
/* call with @p->lock held */
static void build_page_bitmap(PageDesc *p)
{
int n, tb_start, tb_end;
TranslationBlock *tb;
assert_page_locked(p);
p->code_bitmap = bitmap_new(TARGET_PAGE_SIZE);
PAGE_FOR_EACH_TB(p, tb, n) {
/* NOTE: this is subtle as a TB may span two physical pages */
if (n == 0) {
/* NOTE: tb_end may be after the end of the page, but
it is not a problem */
tb_start = tb->pc & ~TARGET_PAGE_MASK;
tb_end = tb_start + tb->size;
if (tb_end > TARGET_PAGE_SIZE) {
tb_end = TARGET_PAGE_SIZE;
}
} else {
tb_start = 0;
tb_end = ((tb->pc + tb->size) & ~TARGET_PAGE_MASK);
}
bitmap_set(p->code_bitmap, tb_start, tb_end - tb_start);
}
}
#endif
/* add the tb in the target page and protect it if necessary
*
* Called with mmap_lock held for user-mode emulation.
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
* Called with @p->lock held in !user-mode.
*/
static inline void tb_page_add(PageDesc *p, TranslationBlock *tb,
unsigned int n, tb_page_addr_t page_addr)
{
#ifndef CONFIG_USER_ONLY
bool page_already_protected;
#endif
assert_page_locked(p);
tb->page_addr[n] = page_addr;
tb->page_next[n] = p->first_tb;
#ifndef CONFIG_USER_ONLY
page_already_protected = p->first_tb != (uintptr_t)NULL;
#endif
p->first_tb = (uintptr_t)tb | n;
invalidate_page_bitmap(p);
#if defined(CONFIG_USER_ONLY)
if (p->flags & PAGE_WRITE) {
target_ulong addr;
PageDesc *p2;
int prot;
/* force the host page as non writable (writes will have a
page fault + mprotect overhead) */
page_addr &= qemu_host_page_mask;
prot = 0;
for (addr = page_addr; addr < page_addr + qemu_host_page_size;
addr += TARGET_PAGE_SIZE) {
p2 = page_find(addr >> TARGET_PAGE_BITS);
if (!p2) {
continue;
}
prot |= p2->flags;
p2->flags &= ~PAGE_WRITE;
}
mprotect(g2h(page_addr), qemu_host_page_size,
(prot & PAGE_BITS) & ~PAGE_WRITE);
if (DEBUG_TB_INVALIDATE_GATE) {
printf("protecting code page: 0x" TB_PAGE_ADDR_FMT "\n", page_addr);
}
}
#else
/* if some code is already present, then the pages are already
protected. So we handle the case where only the first TB is
allocated in a physical page */
if (!page_already_protected) {
tlb_protect_code(page_addr);
}
#endif
}
/* add a new TB and link it to the physical page tables. phys_page2 is
* (-1) to indicate that only one page contains the TB.
*
* Called with mmap_lock held for user-mode emulation.
*
* Returns a pointer @tb, or a pointer to an existing TB that matches @tb.
* Note that in !user-mode, another thread might have already added a TB
* for the same block of guest code that @tb corresponds to. In that case,
* the caller should discard the original @tb, and use instead the returned TB.
*/
static TranslationBlock *
tb_link_page(TranslationBlock *tb, tb_page_addr_t phys_pc,
tb_page_addr_t phys_page2)
{
PageDesc *p;
PageDesc *p2 = NULL;
assert_memory_lock();
/*
* Add the TB to the page list, acquiring first the pages's locks.
* We keep the locks held until after inserting the TB in the hash table,
* so that if the insertion fails we know for sure that the TBs are still
* in the page descriptors.
* Note that inserting into the hash table first isn't an option, since
* we can only insert TBs that are fully initialized.
*/
page_lock_pair(&p, phys_pc, &p2, phys_page2, 1);
tb_page_add(p, tb, 0, phys_pc & TARGET_PAGE_MASK);
if (p2) {
tb_page_add(p2, tb, 1, phys_page2);
} else {
tb->page_addr[1] = -1;
}
if (!(tb->cflags & CF_NOCACHE)) {
void *existing_tb = NULL;
uint32_t h;
/* add in the hash table */
h = tb_hash_func(phys_pc, tb->pc, tb->flags, tb->cflags & CF_HASH_MASK,
tb->trace_vcpu_dstate);
qht_insert(&tb_ctx.htable, tb, h, &existing_tb);
/* remove TB from the page(s) if we couldn't insert it */
if (unlikely(existing_tb)) {
tb_page_remove(p, tb);
invalidate_page_bitmap(p);
if (p2) {
tb_page_remove(p2, tb);
invalidate_page_bitmap(p2);
}
tb = existing_tb;
}
}
if (p2 && p2 != p) {
page_unlock(p2);
}
page_unlock(p);
#ifdef CONFIG_USER_ONLY
if (DEBUG_TB_CHECK_GATE) {
tb_page_check();
}
#endif
return tb;
}
/* Called with mmap_lock held for user mode emulation. */
TranslationBlock *tb_gen_code(CPUState *cpu,
target_ulong pc, target_ulong cs_base,
uint32_t flags, int cflags)
{
CPUArchState *env = cpu->env_ptr;
TranslationBlock *tb, *existing_tb;
tb_page_addr_t phys_pc, phys_page2;
target_ulong virt_page2;
tcg_insn_unit *gen_code_buf;
int gen_code_size, search_size;
#ifdef CONFIG_PROFILER
TCGProfile *prof = &tcg_ctx->prof;
int64_t ti;
#endif
assert_memory_lock();
phys_pc = get_page_addr_code(env, pc);
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
buffer_overflow:
tb = tb_alloc(pc);
if (unlikely(!tb)) {
/* flush must be done */
tb_flush(cpu);
mmap_unlock();
/* Make the execution loop process the flush as soon as possible. */
cpu->exception_index = EXCP_INTERRUPT;
cpu_loop_exit(cpu);
}
gen_code_buf = tcg_ctx->code_gen_ptr;
tb->tc.ptr = gen_code_buf;
tb->pc = pc;
tb->cs_base = cs_base;
tb->flags = flags;
tb->cflags = cflags;
tb->trace_vcpu_dstate = *cpu->trace_dstate;
tcg_ctx->tb_cflags = cflags;
#ifdef CONFIG_PROFILER
/* includes aborted translations because of exceptions */
atomic_set(&prof->tb_count1, prof->tb_count1 + 1);
ti = profile_getclock();
#endif
tcg_func_start(tcg_ctx);
tcg_ctx->cpu = ENV_GET_CPU(env);
gen_intermediate_code(cpu, tb);
tcg_ctx->cpu = NULL;
trace_translate_block(tb, tb->pc, tb->tc.ptr);
/* generate machine code */
tb->jmp_reset_offset[0] = TB_JMP_RESET_OFFSET_INVALID;
tb->jmp_reset_offset[1] = TB_JMP_RESET_OFFSET_INVALID;
tcg_ctx->tb_jmp_reset_offset = tb->jmp_reset_offset;
if (TCG_TARGET_HAS_direct_jump) {
tcg_ctx->tb_jmp_insn_offset = tb->jmp_target_arg;
tcg_ctx->tb_jmp_target_addr = NULL;
} else {
tcg_ctx->tb_jmp_insn_offset = NULL;
tcg_ctx->tb_jmp_target_addr = tb->jmp_target_arg;
}
#ifdef CONFIG_PROFILER
atomic_set(&prof->tb_count, prof->tb_count + 1);
atomic_set(&prof->interm_time, prof->interm_time + profile_getclock() - ti);
tcg: fix corruption of code_time profiling counter upon tb_flush Whenever there is an overflow in code_gen_buffer (e.g. we run out of space in it and have to flush it), the code_time profiling counter ends up with an invalid value (that is, code_time -= profile_getclock(), without later on getting += profile_getclock() due to the goto). Fix it by using the ti variable, so that we only update code_time when there is no overflow. Note that in case there is an overflow we fail to account for the elapsed coding time, but this is quite rare so we can probably live with it. "info jit" before/after, roughly at the same time during debian-arm bootup: - before: Statistics: TB flush count 1 TB invalidate count 4665 TLB flush count 998 JIT cycles -615191529184601 (-256329.804 s at 2.4 GHz) translated TBs 302310 (aborted=0 0.0%) avg ops/TB 48.4 max=438 deleted ops/TB 8.54 avg temps/TB 32.31 max=38 avg host code/TB 361.5 avg search data/TB 24.5 cycles/op -42014693.0 cycles/in byte -121444900.2 cycles/out byte -5629031.1 cycles/search byte -83114481.0 gen_interm time -0.0% gen_code time 100.0% optim./code time -0.0% liveness/code time -0.0% cpu_restore count 6236 avg cycles 110.4 - after: Statistics: TB flush count 1 TB invalidate count 4665 TLB flush count 1010 JIT cycles 1996899624 (0.832 s at 2.4 GHz) translated TBs 297961 (aborted=0 0.0%) avg ops/TB 48.5 max=438 deleted ops/TB 8.56 avg temps/TB 32.31 max=38 avg host code/TB 361.8 avg search data/TB 24.5 cycles/op 138.2 cycles/in byte 398.4 cycles/out byte 18.5 cycles/search byte 273.1 gen_interm time 14.0% gen_code time 86.0% optim./code time 19.4% liveness/code time 10.3% cpu_restore count 6372 avg cycles 111.0 Reviewed-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 00:22:49 +02:00
ti = profile_getclock();
#endif
/* ??? Overflow could be handled better here. In particular, we
don't need to re-do gen_intermediate_code, nor should we re-do
the tcg optimization currently hidden inside tcg_gen_code. All
that should be required is to flush the TBs, allocate a new TB,
re-initialize it per above, and re-do the actual code generation. */
gen_code_size = tcg_gen_code(tcg_ctx, tb);
if (unlikely(gen_code_size < 0)) {
goto buffer_overflow;
}
search_size = encode_search(tb, (void *)gen_code_buf + gen_code_size);
if (unlikely(search_size < 0)) {
goto buffer_overflow;
}
translate-all: use a binary search tree to track TBs in TBContext This is a prerequisite for supporting multiple TCG contexts, since we will have threads generating code in separate regions of code_gen_buffer. For this we need a new field (.size) in struct tb_tc to keep track of the size of the translated code. This field uses a size_t to avoid adding a hole to the struct, although really an unsigned int would have been enough. The comparison function we use is optimized for the common case: insertions. Profiling shows that upon booting debian-arm, 98% of comparisons are between existing tb's (i.e. a->size and b->size are both !0), which happens during insertions (and removals, but those are rare). The remaining cases are lookups. From reading the glib sources we see that the first key is always the lookup key. However, the code does not assume this to always be the case because this behaviour is not guaranteed in the glib docs. However, we embed this knowledge in the code as a branch hint for the compiler. Note that tb_free does not free space in the code_gen_buffer anymore, since we cannot easily know whether the tb is the last one inserted in code_gen_buffer. The next patch in this series renames tb_free to tb_remove to reflect this. Performance-wise, lookups in tb_find_pc are the same as before: O(log n). However, insertions are O(log n) instead of O(1), which results in a small slowdown when booting debian-arm: Performance counter stats for 'build/arm-softmmu/qemu-system-arm \ -machine type=virt -nographic -smp 1 -m 4096 \ -netdev user,id=unet,hostfwd=tcp::2222-:22 \ -device virtio-net-device,netdev=unet \ -drive file=img/arm/jessie-arm32.qcow2,id=myblock,index=0,if=none \ -device virtio-blk-device,drive=myblock \ -kernel img/arm/aarch32-current-linux-kernel-only.img \ -append console=ttyAMA0 root=/dev/vda1 \ -name arm,debug-threads=on -smp 1' (10 runs): - Before: 8048.598422 task-clock (msec) # 0.931 CPUs utilized ( +- 0.28% ) 16,974 context-switches # 0.002 M/sec ( +- 0.12% ) 0 cpu-migrations # 0.000 K/sec 10,125 page-faults # 0.001 M/sec ( +- 1.23% ) 35,144,901,879 cycles # 4.367 GHz ( +- 0.14% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,758,252,643 instructions # 1.87 insns per cycle ( +- 0.33% ) 10,871,298,668 branches # 1350.707 M/sec ( +- 0.41% ) 192,322,212 branch-misses # 1.77% of all branches ( +- 0.32% ) 8.640869419 seconds time elapsed ( +- 0.57% ) - After: 8146.242027 task-clock (msec) # 0.923 CPUs utilized ( +- 1.23% ) 17,016 context-switches # 0.002 M/sec ( +- 0.40% ) 0 cpu-migrations # 0.000 K/sec 18,769 page-faults # 0.002 M/sec ( +- 0.45% ) 35,660,956,120 cycles # 4.378 GHz ( +- 1.22% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,095,366,607 instructions # 1.83 insns per cycle ( +- 1.73% ) 10,803,480,261 branches # 1326.192 M/sec ( +- 1.95% ) 195,601,289 branch-misses # 1.81% of all branches ( +- 0.39% ) 8.828660235 seconds time elapsed ( +- 0.38% ) Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-06-24 01:00:11 +02:00
tb->tc.size = gen_code_size;
#ifdef CONFIG_PROFILER
atomic_set(&prof->code_time, prof->code_time + profile_getclock() - ti);
atomic_set(&prof->code_in_len, prof->code_in_len + tb->size);
atomic_set(&prof->code_out_len, prof->code_out_len + gen_code_size);
atomic_set(&prof->search_out_len, prof->search_out_len + search_size);
#endif
#ifdef DEBUG_DISAS
if (qemu_loglevel_mask(CPU_LOG_TB_OUT_ASM) &&
qemu_log_in_addr_range(tb->pc)) {
qemu_log_lock();
qemu_log("OUT: [size=%d]\n", gen_code_size);
if (tcg_ctx->data_gen_ptr) {
size_t code_size = tcg_ctx->data_gen_ptr - tb->tc.ptr;
size_t data_size = gen_code_size - code_size;
size_t i;
log_disas(tb->tc.ptr, code_size);
for (i = 0; i < data_size; i += sizeof(tcg_target_ulong)) {
if (sizeof(tcg_target_ulong) == 8) {
qemu_log("0x%08" PRIxPTR ": .quad 0x%016" PRIx64 "\n",
(uintptr_t)tcg_ctx->data_gen_ptr + i,
*(uint64_t *)(tcg_ctx->data_gen_ptr + i));
} else {
qemu_log("0x%08" PRIxPTR ": .long 0x%08x\n",
(uintptr_t)tcg_ctx->data_gen_ptr + i,
*(uint32_t *)(tcg_ctx->data_gen_ptr + i));
}
}
} else {
log_disas(tb->tc.ptr, gen_code_size);
}
qemu_log("\n");
qemu_log_flush();
qemu_log_unlock();
}
#endif
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
atomic_set(&tcg_ctx->code_gen_ptr, (void *)
ROUND_UP((uintptr_t)gen_code_buf + gen_code_size + search_size,
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
CODE_GEN_ALIGN));
/* init jump list */
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
qemu_spin_init(&tb->jmp_lock);
tb->jmp_list_head = (uintptr_t)NULL;
tb->jmp_list_next[0] = (uintptr_t)NULL;
tb->jmp_list_next[1] = (uintptr_t)NULL;
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
tb->jmp_dest[0] = (uintptr_t)NULL;
tb->jmp_dest[1] = (uintptr_t)NULL;
/* init original jump addresses which have been set during tcg_gen_code() */
if (tb->jmp_reset_offset[0] != TB_JMP_RESET_OFFSET_INVALID) {
tb_reset_jump(tb, 0);
}
if (tb->jmp_reset_offset[1] != TB_JMP_RESET_OFFSET_INVALID) {
tb_reset_jump(tb, 1);
}
/* check next page if needed */
virt_page2 = (pc + tb->size - 1) & TARGET_PAGE_MASK;
phys_page2 = -1;
if ((pc & TARGET_PAGE_MASK) != virt_page2) {
phys_page2 = get_page_addr_code(env, virt_page2);
}
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
/*
* No explicit memory barrier is required -- tb_link_page() makes the
* TB visible in a consistent state.
*/
existing_tb = tb_link_page(tb, phys_pc, phys_page2);
/* if the TB already exists, discard what we just translated */
if (unlikely(existing_tb != tb)) {
uintptr_t orig_aligned = (uintptr_t)gen_code_buf;
orig_aligned -= ROUND_UP(sizeof(*tb), qemu_icache_linesize);
atomic_set(&tcg_ctx->code_gen_ptr, (void *)orig_aligned);
return existing_tb;
}
tcg: track TBs with per-region BST's This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-26 22:58:05 +02:00
tcg_tb_insert(tb);
return tb;
}
/*
* @p must be non-NULL.
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
* user-mode: call with mmap_lock held.
* !user-mode: call with all @pages locked.
*/
static void
tb_invalidate_phys_page_range__locked(struct page_collection *pages,
PageDesc *p, tb_page_addr_t start,
tb_page_addr_t end,
int is_cpu_write_access)
{
TranslationBlock *tb;
tb_page_addr_t tb_start, tb_end;
int n;
#ifdef TARGET_HAS_PRECISE_SMC
CPUState *cpu = current_cpu;
CPUArchState *env = NULL;
int current_tb_not_found = is_cpu_write_access;
TranslationBlock *current_tb = NULL;
int current_tb_modified = 0;
target_ulong current_pc = 0;
target_ulong current_cs_base = 0;
uint32_t current_flags = 0;
#endif /* TARGET_HAS_PRECISE_SMC */
assert_page_locked(p);
#if defined(TARGET_HAS_PRECISE_SMC)
if (cpu != NULL) {
env = cpu->env_ptr;
}
#endif
/* we remove all the TBs in the range [start, end[ */
/* XXX: see if in some cases it could be faster to invalidate all
the code */
PAGE_FOR_EACH_TB(p, tb, n) {
assert_page_locked(p);
/* NOTE: this is subtle as a TB may span two physical pages */
if (n == 0) {
/* NOTE: tb_end may be after the end of the page, but
it is not a problem */
tb_start = tb->page_addr[0] + (tb->pc & ~TARGET_PAGE_MASK);
tb_end = tb_start + tb->size;
} else {
tb_start = tb->page_addr[1];
tb_end = tb_start + ((tb->pc + tb->size) & ~TARGET_PAGE_MASK);
}
if (!(tb_end <= start || tb_start >= end)) {
#ifdef TARGET_HAS_PRECISE_SMC
if (current_tb_not_found) {
current_tb_not_found = 0;
current_tb = NULL;
if (cpu->mem_io_pc) {
/* now we have a real cpu fault */
tcg: track TBs with per-region BST's This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-26 22:58:05 +02:00
current_tb = tcg_tb_lookup(cpu->mem_io_pc);
}
}
if (current_tb == tb &&
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
(tb_cflags(current_tb) & CF_COUNT_MASK) != 1) {
/* If we are modifying the current TB, we must stop
its execution. We could be more precise by checking
that the modification is after the current PC, but it
would require a specialized function to partially
restore the CPU state */
current_tb_modified = 1;
icount: fix cpu_restore_state_from_tb for non-tb-exit cases In icount mode, instructions that access io memory spaces in the middle of the translation block invoke TB recompilation. After recompilation, such instructions become last in the TB and are allowed to access io memory spaces. When the code includes instruction like i386 'xchg eax, 0xffffd080' which accesses APIC, QEMU goes into an infinite loop of the recompilation. This instruction includes two memory accesses - one read and one write. After the first access, APIC calls cpu_report_tpr_access, which restores the CPU state to get the current eip. But cpu_restore_state_from_tb resets the cpu->can_do_io flag which makes the second memory access invalid. Therefore the second memory access causes a recompilation of the block. Then these operations repeat again and again. This patch moves resetting cpu->can_do_io flag from cpu_restore_state_from_tb to cpu_loop_exit* functions. It also adds a parameter for cpu_restore_state which controls restoring icount. There is no need to restore icount when we only query CPU state without breaking the TB. Restoring it in such cases leads to the incorrect flow of the virtual time. In most cases new parameter is true (icount should be recalculated). But there are two cases in i386 and openrisc when the CPU state is only queried without the need to break the TB. This patch fixes both of these cases. Signed-off-by: Pavel Dovgalyuk <Pavel.Dovgaluk@ispras.ru> Message-Id: <20180409091320.12504.35329.stgit@pasha-VirtualBox> [rth: Make can_do_io setting unconditional; move from cpu_exec; make cpu_loop_exit_{noexc,restore} call cpu_loop_exit.] Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2018-04-09 11:13:20 +02:00
cpu_restore_state_from_tb(cpu, current_tb,
cpu->mem_io_pc, true);
cpu_get_tb_cpu_state(env, &current_pc, &current_cs_base,
&current_flags);
}
#endif /* TARGET_HAS_PRECISE_SMC */
tb_phys_invalidate__locked(tb);
}
}
#if !defined(CONFIG_USER_ONLY)
/* if no code remaining, no need to continue to use slow writes */
if (!p->first_tb) {
invalidate_page_bitmap(p);
tlb_unprotect_code(start);
}
#endif
#ifdef TARGET_HAS_PRECISE_SMC
if (current_tb_modified) {
page_collection_unlock(pages);
/* Force execution of one insn next time. */
cpu->cflags_next_tb = 1 | curr_cflags();
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
mmap_unlock();
cpu_loop_exit_noexc(cpu);
}
#endif
}
/*
* Invalidate all TBs which intersect with the target physical address range
* [start;end[. NOTE: start and end must refer to the *same* physical page.
* 'is_cpu_write_access' should be true if called from a real cpu write
* access: the virtual CPU will exit the current TB if code is modified inside
* this TB.
*
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
* Called with mmap_lock held for user-mode emulation
*/
void tb_invalidate_phys_page_range(tb_page_addr_t start, tb_page_addr_t end,
int is_cpu_write_access)
{
struct page_collection *pages;
PageDesc *p;
assert_memory_lock();
p = page_find(start >> TARGET_PAGE_BITS);
if (p == NULL) {
return;
}
pages = page_collection_lock(start, end);
tb_invalidate_phys_page_range__locked(pages, p, start, end,
is_cpu_write_access);
page_collection_unlock(pages);
}
/*
* Invalidate all TBs which intersect with the target physical address range
* [start;end[. NOTE: start and end may refer to *different* physical pages.
* 'is_cpu_write_access' should be true if called from a real cpu write
* access: the virtual CPU will exit the current TB if code is modified inside
* this TB.
*
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
* Called with mmap_lock held for user-mode emulation.
*/
#ifdef CONFIG_SOFTMMU
void tb_invalidate_phys_range(ram_addr_t start, ram_addr_t end)
#else
void tb_invalidate_phys_range(target_ulong start, target_ulong end)
#endif
{
struct page_collection *pages;
tb_page_addr_t next;
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
assert_memory_lock();
pages = page_collection_lock(start, end);
for (next = (start & TARGET_PAGE_MASK) + TARGET_PAGE_SIZE;
start < end;
start = next, next += TARGET_PAGE_SIZE) {
PageDesc *pd = page_find(start >> TARGET_PAGE_BITS);
tb_page_addr_t bound = MIN(next, end);
if (pd == NULL) {
continue;
}
tb_invalidate_phys_page_range__locked(pages, pd, start, bound, 0);
}
page_collection_unlock(pages);
}
#ifdef CONFIG_SOFTMMU
/* len must be <= 8 and start must be a multiple of len.
* Called via softmmu_template.h when code areas are written to with
tcg: drop global lock during TCG code execution This finally allows TCG to benefit from the iothread introduction: Drop the global mutex while running pure TCG CPU code. Reacquire the lock when entering MMIO or PIO emulation, or when leaving the TCG loop. We have to revert a few optimization for the current TCG threading model, namely kicking the TCG thread in qemu_mutex_lock_iothread and not kicking it in qemu_cpu_kick. We also need to disable RAM block reordering until we have a more efficient locking mechanism at hand. Still, a Linux x86 UP guest and my Musicpal ARM model boot fine here. These numbers demonstrate where we gain something: 20338 jan 20 0 331m 75m 6904 R 99 0.9 0:50.95 qemu-system-arm 20337 jan 20 0 331m 75m 6904 S 20 0.9 0:26.50 qemu-system-arm The guest CPU was fully loaded, but the iothread could still run mostly independent on a second core. Without the patch we don't get beyond 32206 jan 20 0 330m 73m 7036 R 82 0.9 1:06.00 qemu-system-arm 32204 jan 20 0 330m 73m 7036 S 21 0.9 0:17.03 qemu-system-arm We don't benefit significantly, though, when the guest is not fully loading a host CPU. Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com> Message-Id: <1439220437-23957-10-git-send-email-fred.konrad@greensocs.com> [FK: Rebase, fix qemu_devices_reset deadlock, rm address_space_* mutex] Signed-off-by: KONRAD Frederic <fred.konrad@greensocs.com> [EGC: fixed iothread lock for cpu-exec IRQ handling] Signed-off-by: Emilio G. Cota <cota@braap.org> [AJB: -smp single-threaded fix, clean commit msg, BQL fixes] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Pranith Kumar <bobby.prani@gmail.com> [PM: target-arm changes] Acked-by: Peter Maydell <peter.maydell@linaro.org>
2017-02-23 19:29:11 +01:00
* iothread mutex not held.
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
*
* Call with all @pages in the range [@start, @start + len[ locked.
*/
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
void tb_invalidate_phys_page_fast(struct page_collection *pages,
tb_page_addr_t start, int len)
{
PageDesc *p;
#if 0
if (1) {
qemu_log("modifying code at 0x%x size=%d EIP=%x PC=%08x\n",
cpu_single_env->mem_io_vaddr, len,
cpu_single_env->eip,
cpu_single_env->eip +
(intptr_t)cpu_single_env->segs[R_CS].base);
}
#endif
assert_memory_lock();
p = page_find(start >> TARGET_PAGE_BITS);
if (!p) {
return;
}
assert_page_locked(p);
if (!p->code_bitmap &&
++p->code_write_count >= SMC_BITMAP_USE_THRESHOLD) {
build_page_bitmap(p);
}
if (p->code_bitmap) {
unsigned int nr;
unsigned long b;
nr = start & ~TARGET_PAGE_MASK;
b = p->code_bitmap[BIT_WORD(nr)] >> (nr & (BITS_PER_LONG - 1));
if (b & ((1 << len) - 1)) {
goto do_invalidate;
}
} else {
do_invalidate:
tb_invalidate_phys_page_range__locked(pages, p, start, start + len, 1);
}
}
#else
/* Called with mmap_lock held. If pc is not 0 then it indicates the
* host PC of the faulting store instruction that caused this invalidate.
* Returns true if the caller needs to abort execution of the current
* TB (because it was modified by this store and the guest CPU has
* precise-SMC semantics).
*/
static bool tb_invalidate_phys_page(tb_page_addr_t addr, uintptr_t pc)
{
TranslationBlock *tb;
PageDesc *p;
int n;
#ifdef TARGET_HAS_PRECISE_SMC
TranslationBlock *current_tb = NULL;
CPUState *cpu = current_cpu;
CPUArchState *env = NULL;
int current_tb_modified = 0;
target_ulong current_pc = 0;
target_ulong current_cs_base = 0;
uint32_t current_flags = 0;
#endif
assert_memory_lock();
addr &= TARGET_PAGE_MASK;
p = page_find(addr >> TARGET_PAGE_BITS);
if (!p) {
return false;
}
#ifdef TARGET_HAS_PRECISE_SMC
if (p->first_tb && pc != 0) {
tcg: track TBs with per-region BST's This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-26 22:58:05 +02:00
current_tb = tcg_tb_lookup(pc);
}
if (cpu != NULL) {
env = cpu->env_ptr;
}
#endif
assert_page_locked(p);
PAGE_FOR_EACH_TB(p, tb, n) {
#ifdef TARGET_HAS_PRECISE_SMC
if (current_tb == tb &&
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
(tb_cflags(current_tb) & CF_COUNT_MASK) != 1) {
/* If we are modifying the current TB, we must stop
its execution. We could be more precise by checking
that the modification is after the current PC, but it
would require a specialized function to partially
restore the CPU state */
current_tb_modified = 1;
icount: fix cpu_restore_state_from_tb for non-tb-exit cases In icount mode, instructions that access io memory spaces in the middle of the translation block invoke TB recompilation. After recompilation, such instructions become last in the TB and are allowed to access io memory spaces. When the code includes instruction like i386 'xchg eax, 0xffffd080' which accesses APIC, QEMU goes into an infinite loop of the recompilation. This instruction includes two memory accesses - one read and one write. After the first access, APIC calls cpu_report_tpr_access, which restores the CPU state to get the current eip. But cpu_restore_state_from_tb resets the cpu->can_do_io flag which makes the second memory access invalid. Therefore the second memory access causes a recompilation of the block. Then these operations repeat again and again. This patch moves resetting cpu->can_do_io flag from cpu_restore_state_from_tb to cpu_loop_exit* functions. It also adds a parameter for cpu_restore_state which controls restoring icount. There is no need to restore icount when we only query CPU state without breaking the TB. Restoring it in such cases leads to the incorrect flow of the virtual time. In most cases new parameter is true (icount should be recalculated). But there are two cases in i386 and openrisc when the CPU state is only queried without the need to break the TB. This patch fixes both of these cases. Signed-off-by: Pavel Dovgalyuk <Pavel.Dovgaluk@ispras.ru> Message-Id: <20180409091320.12504.35329.stgit@pasha-VirtualBox> [rth: Make can_do_io setting unconditional; move from cpu_exec; make cpu_loop_exit_{noexc,restore} call cpu_loop_exit.] Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2018-04-09 11:13:20 +02:00
cpu_restore_state_from_tb(cpu, current_tb, pc, true);
cpu_get_tb_cpu_state(env, &current_pc, &current_cs_base,
&current_flags);
}
#endif /* TARGET_HAS_PRECISE_SMC */
tb_phys_invalidate(tb, addr);
}
p->first_tb = (uintptr_t)NULL;
#ifdef TARGET_HAS_PRECISE_SMC
if (current_tb_modified) {
/* Force execution of one insn next time. */
cpu->cflags_next_tb = 1 | curr_cflags();
return true;
}
#endif
return false;
}
#endif
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
/* user-mode: call with mmap_lock held */
void tb_check_watchpoint(CPUState *cpu)
{
TranslationBlock *tb;
tcg: remove tb_lock Use mmap_lock in user-mode to protect TCG state and the page descriptors. In !user-mode, each vCPU has its own TCG state, so no locks needed. Per-page locks are used to protect the page descriptors. Per-TB locks are used in both modes to protect TB jumps. Some notes: - tb_lock is removed from notdirty_mem_write by passing a locked page_collection to tb_invalidate_phys_page_fast. - tcg_tb_lookup/remove/insert/etc have their own internal lock(s), so there is no need to further serialize access to them. - do_tb_flush is run in a safe async context, meaning no other vCPU threads are running. Therefore acquiring mmap_lock there is just to please tools such as thread sanitizer. - Not visible in the diff, but tb_invalidate_phys_page already has an assert_memory_lock. - cpu_io_recompile is !user-only, so no mmap_lock there. - Added mmap_unlock()'s before all siglongjmp's that could be called in user-mode while mmap_lock is held. + Added an assert for !have_mmap_lock() after returning from the longjmp in cpu_exec, just like we do in cpu_exec_step_atomic. Performance numbers before/after: Host: AMD Opteron(tm) Processor 6376 ubuntu 17.04 ppc64 bootup+shutdown time 700 +-+--+----+------+------------+-----------+------------*--+-+ | + + + + + *B | | before ***B*** ** * | |tb lock removal ###D### *** | 600 +-+ *** +-+ | ** # | | *B* #D | | *** * ## | 500 +-+ *** ### +-+ | * *** ### | | *B* # ## | | ** * #D# | 400 +-+ ** ## +-+ | ** ### | | ** ## | | ** # ## | 300 +-+ * B* #D# +-+ | B *** ### | | * ** #### | | * *** ### | 200 +-+ B *B #D# +-+ | #B* * ## # | | #* ## | | + D##D# + + + + | 100 +-+--+----+------+------------+-----------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/HwmBHXe debian jessie aarch64 bootup+shutdown time 90 +-+--+-----+-----+------------+------------+------------+--+-+ | + + + + + + | | before ***B*** B | 80 +tb lock removal ###D### **D +-+ | **### | | **## | 70 +-+ ** # +-+ | ** ## | | ** # | 60 +-+ *B ## +-+ | ** ## | | *** #D | 50 +-+ *** ## +-+ | * ** ### | | **B* ### | 40 +-+ **** # ## +-+ | **** #D# | | ***B** ### | 30 +-+ B***B** #### +-+ | B * * # ### | | B ###D# | 20 +-+ D ##D## +-+ | D# | | + + + + + + | 10 +-+--+-----+-----+------------+------------+------------+--+-+ 1 8 16 Guest CPUs 48 64 png: https://imgur.com/iGpGFtv The gains are high for 4-8 CPUs. Beyond that point, however, unrelated lock contention significantly hurts scalability. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-05 05:46:31 +02:00
assert_memory_lock();
tcg: track TBs with per-region BST's This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-26 22:58:05 +02:00
tb = tcg_tb_lookup(cpu->mem_io_pc);
if (tb) {
/* We can use retranslation to find the PC. */
icount: fix cpu_restore_state_from_tb for non-tb-exit cases In icount mode, instructions that access io memory spaces in the middle of the translation block invoke TB recompilation. After recompilation, such instructions become last in the TB and are allowed to access io memory spaces. When the code includes instruction like i386 'xchg eax, 0xffffd080' which accesses APIC, QEMU goes into an infinite loop of the recompilation. This instruction includes two memory accesses - one read and one write. After the first access, APIC calls cpu_report_tpr_access, which restores the CPU state to get the current eip. But cpu_restore_state_from_tb resets the cpu->can_do_io flag which makes the second memory access invalid. Therefore the second memory access causes a recompilation of the block. Then these operations repeat again and again. This patch moves resetting cpu->can_do_io flag from cpu_restore_state_from_tb to cpu_loop_exit* functions. It also adds a parameter for cpu_restore_state which controls restoring icount. There is no need to restore icount when we only query CPU state without breaking the TB. Restoring it in such cases leads to the incorrect flow of the virtual time. In most cases new parameter is true (icount should be recalculated). But there are two cases in i386 and openrisc when the CPU state is only queried without the need to break the TB. This patch fixes both of these cases. Signed-off-by: Pavel Dovgalyuk <Pavel.Dovgaluk@ispras.ru> Message-Id: <20180409091320.12504.35329.stgit@pasha-VirtualBox> [rth: Make can_do_io setting unconditional; move from cpu_exec; make cpu_loop_exit_{noexc,restore} call cpu_loop_exit.] Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2018-04-09 11:13:20 +02:00
cpu_restore_state_from_tb(cpu, tb, cpu->mem_io_pc, true);
tb_phys_invalidate(tb, -1);
} else {
/* The exception probably happened in a helper. The CPU state should
have been saved before calling it. Fetch the PC from there. */
CPUArchState *env = cpu->env_ptr;
target_ulong pc, cs_base;
tb_page_addr_t addr;
uint32_t flags;
cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);
addr = get_page_addr_code(env, pc);
tb_invalidate_phys_range(addr, addr + 1);
}
}
#ifndef CONFIG_USER_ONLY
/* in deterministic execution mode, instructions doing device I/Os
tcg: drop global lock during TCG code execution This finally allows TCG to benefit from the iothread introduction: Drop the global mutex while running pure TCG CPU code. Reacquire the lock when entering MMIO or PIO emulation, or when leaving the TCG loop. We have to revert a few optimization for the current TCG threading model, namely kicking the TCG thread in qemu_mutex_lock_iothread and not kicking it in qemu_cpu_kick. We also need to disable RAM block reordering until we have a more efficient locking mechanism at hand. Still, a Linux x86 UP guest and my Musicpal ARM model boot fine here. These numbers demonstrate where we gain something: 20338 jan 20 0 331m 75m 6904 R 99 0.9 0:50.95 qemu-system-arm 20337 jan 20 0 331m 75m 6904 S 20 0.9 0:26.50 qemu-system-arm The guest CPU was fully loaded, but the iothread could still run mostly independent on a second core. Without the patch we don't get beyond 32206 jan 20 0 330m 73m 7036 R 82 0.9 1:06.00 qemu-system-arm 32204 jan 20 0 330m 73m 7036 S 21 0.9 0:17.03 qemu-system-arm We don't benefit significantly, though, when the guest is not fully loading a host CPU. Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com> Message-Id: <1439220437-23957-10-git-send-email-fred.konrad@greensocs.com> [FK: Rebase, fix qemu_devices_reset deadlock, rm address_space_* mutex] Signed-off-by: KONRAD Frederic <fred.konrad@greensocs.com> [EGC: fixed iothread lock for cpu-exec IRQ handling] Signed-off-by: Emilio G. Cota <cota@braap.org> [AJB: -smp single-threaded fix, clean commit msg, BQL fixes] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Pranith Kumar <bobby.prani@gmail.com> [PM: target-arm changes] Acked-by: Peter Maydell <peter.maydell@linaro.org>
2017-02-23 19:29:11 +01:00
* must be at the end of the TB.
*
* Called by softmmu_template.h, with iothread mutex not held.
*/
void cpu_io_recompile(CPUState *cpu, uintptr_t retaddr)
{
#if defined(TARGET_MIPS) || defined(TARGET_SH4)
CPUArchState *env = cpu->env_ptr;
#endif
TranslationBlock *tb;
uint32_t n;
tcg: track TBs with per-region BST's This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-26 22:58:05 +02:00
tb = tcg_tb_lookup(retaddr);
if (!tb) {
cpu_abort(cpu, "cpu_io_recompile: could not find TB for pc=%p",
(void *)retaddr);
}
icount: fix cpu_restore_state_from_tb for non-tb-exit cases In icount mode, instructions that access io memory spaces in the middle of the translation block invoke TB recompilation. After recompilation, such instructions become last in the TB and are allowed to access io memory spaces. When the code includes instruction like i386 'xchg eax, 0xffffd080' which accesses APIC, QEMU goes into an infinite loop of the recompilation. This instruction includes two memory accesses - one read and one write. After the first access, APIC calls cpu_report_tpr_access, which restores the CPU state to get the current eip. But cpu_restore_state_from_tb resets the cpu->can_do_io flag which makes the second memory access invalid. Therefore the second memory access causes a recompilation of the block. Then these operations repeat again and again. This patch moves resetting cpu->can_do_io flag from cpu_restore_state_from_tb to cpu_loop_exit* functions. It also adds a parameter for cpu_restore_state which controls restoring icount. There is no need to restore icount when we only query CPU state without breaking the TB. Restoring it in such cases leads to the incorrect flow of the virtual time. In most cases new parameter is true (icount should be recalculated). But there are two cases in i386 and openrisc when the CPU state is only queried without the need to break the TB. This patch fixes both of these cases. Signed-off-by: Pavel Dovgalyuk <Pavel.Dovgaluk@ispras.ru> Message-Id: <20180409091320.12504.35329.stgit@pasha-VirtualBox> [rth: Make can_do_io setting unconditional; move from cpu_exec; make cpu_loop_exit_{noexc,restore} call cpu_loop_exit.] Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2018-04-09 11:13:20 +02:00
cpu_restore_state_from_tb(cpu, tb, retaddr, true);
/* On MIPS and SH, delay slot instructions can only be restarted if
they were already the first instruction in the TB. If this is not
the first instruction in a TB then re-execute the preceding
branch. */
n = 1;
#if defined(TARGET_MIPS)
if ((env->hflags & MIPS_HFLAG_BMASK) != 0
&& env->active_tc.PC != tb->pc) {
env->active_tc.PC -= (env->hflags & MIPS_HFLAG_B16 ? 2 : 4);
cpu->icount_decr.u16.low++;
env->hflags &= ~MIPS_HFLAG_BMASK;
n = 2;
}
#elif defined(TARGET_SH4)
if ((env->flags & ((DELAY_SLOT | DELAY_SLOT_CONDITIONAL))) != 0
&& env->pc != tb->pc) {
env->pc -= 2;
cpu->icount_decr.u16.low++;
env->flags &= ~(DELAY_SLOT | DELAY_SLOT_CONDITIONAL);
n = 2;
}
#endif
/* Generate a new TB executing the I/O insn. */
cpu->cflags_next_tb = curr_cflags() | CF_LAST_IO | n;
translate-all: protect TB jumps with a per-destination-TB lock This applies to both user-mode and !user-mode emulation. Instead of relying on a global lock, protect the list of incoming jumps with tb->jmp_lock. This lock also protects tb->cflags, so update all tb->cflags readers outside tb->jmp_lock to use atomic reads via tb_cflags(). In order to find the destination TB (and therefore its jmp_lock) from the origin TB, we introduce tb->jmp_dest[]. I considered not using a linked list of jumps, which simplifies code and makes the struct smaller. However, it unnecessarily increases memory usage, which results in a performance decrease. See for instance these numbers booting+shutting down debian-arm: Time (s) Rel. err (%) Abs. err (s) Rel. slowdown (%) ------------------------------------------------------------------------------ before 20.88 0.74 0.154512 0. after 20.81 0.38 0.079078 -0.33524904 GTree 21.02 0.28 0.058856 0.67049808 GHashTable + xxhash 21.63 1.08 0.233604 3.5919540 Using a hash table or a binary tree to keep track of the jumps doesn't really pay off, not only due to the increased memory usage, but also because most TBs have only 0 or 1 jumps to them. The maximum number of jumps when booting debian-arm that I measured is 35, but as we can see in the histogram below a TB with that many incoming jumps is extremely rare; the average TB has 0.80 incoming jumps. n_jumps: 379208; avg jumps/tb: 0.801099 dist: [0.0,1.0)|▄█▁▁▁▁▁▁▁▁▁▁▁ ▁▁▁▁▁▁ ▁▁▁ ▁▁▁ ▁|[34.0,35.0] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-08-03 02:34:06 +02:00
if (tb_cflags(tb) & CF_NOCACHE) {
if (tb->orig_tb) {
/* Invalidate original TB if this TB was generated in
* cpu_exec_nocache() */
tb_phys_invalidate(tb->orig_tb, -1);
}
tcg: track TBs with per-region BST's This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-26 22:58:05 +02:00
tcg_tb_remove(tb);
}
/* TODO: If env->pc != tb->pc (i.e. the faulting instruction was not
* the first in the TB) then we end up generating a whole new TB and
* repeating the fault, which is horribly inefficient.
* Better would be to execute just this insn uncached, or generate a
* second new TB.
*/
cpu_loop_exit_noexc(cpu);
}
tcg: consistently access cpu->tb_jmp_cache atomically Some code paths can lead to atomic accesses racing with memset() on cpu->tb_jmp_cache, which can result in torn reads/writes and is undefined behaviour in C11. These torn accesses are unlikely to show up as bugs, but from code inspection they seem possible. For example, tb_phys_invalidate does: /* remove the TB from the hash list */ h = tb_jmp_cache_hash_func(tb->pc); CPU_FOREACH(cpu) { if (atomic_read(&cpu->tb_jmp_cache[h]) == tb) { atomic_set(&cpu->tb_jmp_cache[h], NULL); } } Here atomic_set might race with a concurrent memset (such as the ones scheduled via "unsafe" async work, e.g. tlb_flush_page) and therefore we might end up with a torn pointer (or who knows what, because we are under undefined behaviour). This patch converts parallel accesses to cpu->tb_jmp_cache to use atomic primitives, thereby bringing these accesses back to defined behaviour. The price to pay is to potentially execute more instructions when clearing cpu->tb_jmp_cache, but given how infrequently they happen and the small size of the cache, the performance impact I have measured is within noise range when booting debian-arm. Note that under "safe async" work (e.g. do_tb_flush) we could use memset because no other vcpus are running. However I'm keeping these accesses atomic as well to keep things simple and to avoid confusing analysis tools such as ThreadSanitizer. Reviewed-by: Paolo Bonzini <pbonzini@redhat.com> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1497486973-25845-1-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2017-06-15 02:36:13 +02:00
static void tb_jmp_cache_clear_page(CPUState *cpu, target_ulong page_addr)
{
tcg: consistently access cpu->tb_jmp_cache atomically Some code paths can lead to atomic accesses racing with memset() on cpu->tb_jmp_cache, which can result in torn reads/writes and is undefined behaviour in C11. These torn accesses are unlikely to show up as bugs, but from code inspection they seem possible. For example, tb_phys_invalidate does: /* remove the TB from the hash list */ h = tb_jmp_cache_hash_func(tb->pc); CPU_FOREACH(cpu) { if (atomic_read(&cpu->tb_jmp_cache[h]) == tb) { atomic_set(&cpu->tb_jmp_cache[h], NULL); } } Here atomic_set might race with a concurrent memset (such as the ones scheduled via "unsafe" async work, e.g. tlb_flush_page) and therefore we might end up with a torn pointer (or who knows what, because we are under undefined behaviour). This patch converts parallel accesses to cpu->tb_jmp_cache to use atomic primitives, thereby bringing these accesses back to defined behaviour. The price to pay is to potentially execute more instructions when clearing cpu->tb_jmp_cache, but given how infrequently they happen and the small size of the cache, the performance impact I have measured is within noise range when booting debian-arm. Note that under "safe async" work (e.g. do_tb_flush) we could use memset because no other vcpus are running. However I'm keeping these accesses atomic as well to keep things simple and to avoid confusing analysis tools such as ThreadSanitizer. Reviewed-by: Paolo Bonzini <pbonzini@redhat.com> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1497486973-25845-1-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2017-06-15 02:36:13 +02:00
unsigned int i, i0 = tb_jmp_cache_hash_page(page_addr);
tcg: consistently access cpu->tb_jmp_cache atomically Some code paths can lead to atomic accesses racing with memset() on cpu->tb_jmp_cache, which can result in torn reads/writes and is undefined behaviour in C11. These torn accesses are unlikely to show up as bugs, but from code inspection they seem possible. For example, tb_phys_invalidate does: /* remove the TB from the hash list */ h = tb_jmp_cache_hash_func(tb->pc); CPU_FOREACH(cpu) { if (atomic_read(&cpu->tb_jmp_cache[h]) == tb) { atomic_set(&cpu->tb_jmp_cache[h], NULL); } } Here atomic_set might race with a concurrent memset (such as the ones scheduled via "unsafe" async work, e.g. tlb_flush_page) and therefore we might end up with a torn pointer (or who knows what, because we are under undefined behaviour). This patch converts parallel accesses to cpu->tb_jmp_cache to use atomic primitives, thereby bringing these accesses back to defined behaviour. The price to pay is to potentially execute more instructions when clearing cpu->tb_jmp_cache, but given how infrequently they happen and the small size of the cache, the performance impact I have measured is within noise range when booting debian-arm. Note that under "safe async" work (e.g. do_tb_flush) we could use memset because no other vcpus are running. However I'm keeping these accesses atomic as well to keep things simple and to avoid confusing analysis tools such as ThreadSanitizer. Reviewed-by: Paolo Bonzini <pbonzini@redhat.com> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1497486973-25845-1-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2017-06-15 02:36:13 +02:00
for (i = 0; i < TB_JMP_PAGE_SIZE; i++) {
atomic_set(&cpu->tb_jmp_cache[i0 + i], NULL);
}
}
void tb_flush_jmp_cache(CPUState *cpu, target_ulong addr)
{
/* Discard jump cache entries for any tb which might potentially
overlap the flushed page. */
tcg: consistently access cpu->tb_jmp_cache atomically Some code paths can lead to atomic accesses racing with memset() on cpu->tb_jmp_cache, which can result in torn reads/writes and is undefined behaviour in C11. These torn accesses are unlikely to show up as bugs, but from code inspection they seem possible. For example, tb_phys_invalidate does: /* remove the TB from the hash list */ h = tb_jmp_cache_hash_func(tb->pc); CPU_FOREACH(cpu) { if (atomic_read(&cpu->tb_jmp_cache[h]) == tb) { atomic_set(&cpu->tb_jmp_cache[h], NULL); } } Here atomic_set might race with a concurrent memset (such as the ones scheduled via "unsafe" async work, e.g. tlb_flush_page) and therefore we might end up with a torn pointer (or who knows what, because we are under undefined behaviour). This patch converts parallel accesses to cpu->tb_jmp_cache to use atomic primitives, thereby bringing these accesses back to defined behaviour. The price to pay is to potentially execute more instructions when clearing cpu->tb_jmp_cache, but given how infrequently they happen and the small size of the cache, the performance impact I have measured is within noise range when booting debian-arm. Note that under "safe async" work (e.g. do_tb_flush) we could use memset because no other vcpus are running. However I'm keeping these accesses atomic as well to keep things simple and to avoid confusing analysis tools such as ThreadSanitizer. Reviewed-by: Paolo Bonzini <pbonzini@redhat.com> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1497486973-25845-1-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2017-06-15 02:36:13 +02:00
tb_jmp_cache_clear_page(cpu, addr - TARGET_PAGE_SIZE);
tb_jmp_cache_clear_page(cpu, addr);
}
static void print_qht_statistics(FILE *f, fprintf_function cpu_fprintf,
struct qht_stats hst)
{
uint32_t hgram_opts;
size_t hgram_bins;
char *hgram;
if (!hst.head_buckets) {
return;
}
cpu_fprintf(f, "TB hash buckets %zu/%zu (%0.2f%% head buckets used)\n",
hst.used_head_buckets, hst.head_buckets,
(double)hst.used_head_buckets / hst.head_buckets * 100);
hgram_opts = QDIST_PR_BORDER | QDIST_PR_LABELS;
hgram_opts |= QDIST_PR_100X | QDIST_PR_PERCENT;
if (qdist_xmax(&hst.occupancy) - qdist_xmin(&hst.occupancy) == 1) {
hgram_opts |= QDIST_PR_NODECIMAL;
}
hgram = qdist_pr(&hst.occupancy, 10, hgram_opts);
cpu_fprintf(f, "TB hash occupancy %0.2f%% avg chain occ. Histogram: %s\n",
qdist_avg(&hst.occupancy) * 100, hgram);
g_free(hgram);
hgram_opts = QDIST_PR_BORDER | QDIST_PR_LABELS;
hgram_bins = qdist_xmax(&hst.chain) - qdist_xmin(&hst.chain);
if (hgram_bins > 10) {
hgram_bins = 10;
} else {
hgram_bins = 0;
hgram_opts |= QDIST_PR_NODECIMAL | QDIST_PR_NOBINRANGE;
}
hgram = qdist_pr(&hst.chain, hgram_bins, hgram_opts);
cpu_fprintf(f, "TB hash avg chain %0.3f buckets. Histogram: %s\n",
qdist_avg(&hst.chain), hgram);
g_free(hgram);
}
translate-all: use a binary search tree to track TBs in TBContext This is a prerequisite for supporting multiple TCG contexts, since we will have threads generating code in separate regions of code_gen_buffer. For this we need a new field (.size) in struct tb_tc to keep track of the size of the translated code. This field uses a size_t to avoid adding a hole to the struct, although really an unsigned int would have been enough. The comparison function we use is optimized for the common case: insertions. Profiling shows that upon booting debian-arm, 98% of comparisons are between existing tb's (i.e. a->size and b->size are both !0), which happens during insertions (and removals, but those are rare). The remaining cases are lookups. From reading the glib sources we see that the first key is always the lookup key. However, the code does not assume this to always be the case because this behaviour is not guaranteed in the glib docs. However, we embed this knowledge in the code as a branch hint for the compiler. Note that tb_free does not free space in the code_gen_buffer anymore, since we cannot easily know whether the tb is the last one inserted in code_gen_buffer. The next patch in this series renames tb_free to tb_remove to reflect this. Performance-wise, lookups in tb_find_pc are the same as before: O(log n). However, insertions are O(log n) instead of O(1), which results in a small slowdown when booting debian-arm: Performance counter stats for 'build/arm-softmmu/qemu-system-arm \ -machine type=virt -nographic -smp 1 -m 4096 \ -netdev user,id=unet,hostfwd=tcp::2222-:22 \ -device virtio-net-device,netdev=unet \ -drive file=img/arm/jessie-arm32.qcow2,id=myblock,index=0,if=none \ -device virtio-blk-device,drive=myblock \ -kernel img/arm/aarch32-current-linux-kernel-only.img \ -append console=ttyAMA0 root=/dev/vda1 \ -name arm,debug-threads=on -smp 1' (10 runs): - Before: 8048.598422 task-clock (msec) # 0.931 CPUs utilized ( +- 0.28% ) 16,974 context-switches # 0.002 M/sec ( +- 0.12% ) 0 cpu-migrations # 0.000 K/sec 10,125 page-faults # 0.001 M/sec ( +- 1.23% ) 35,144,901,879 cycles # 4.367 GHz ( +- 0.14% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,758,252,643 instructions # 1.87 insns per cycle ( +- 0.33% ) 10,871,298,668 branches # 1350.707 M/sec ( +- 0.41% ) 192,322,212 branch-misses # 1.77% of all branches ( +- 0.32% ) 8.640869419 seconds time elapsed ( +- 0.57% ) - After: 8146.242027 task-clock (msec) # 0.923 CPUs utilized ( +- 1.23% ) 17,016 context-switches # 0.002 M/sec ( +- 0.40% ) 0 cpu-migrations # 0.000 K/sec 18,769 page-faults # 0.002 M/sec ( +- 0.45% ) 35,660,956,120 cycles # 4.378 GHz ( +- 1.22% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,095,366,607 instructions # 1.83 insns per cycle ( +- 1.73% ) 10,803,480,261 branches # 1326.192 M/sec ( +- 1.95% ) 195,601,289 branch-misses # 1.81% of all branches ( +- 0.39% ) 8.828660235 seconds time elapsed ( +- 0.38% ) Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-06-24 01:00:11 +02:00
struct tb_tree_stats {
tcg: track TBs with per-region BST's This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-26 22:58:05 +02:00
size_t nb_tbs;
size_t host_size;
translate-all: use a binary search tree to track TBs in TBContext This is a prerequisite for supporting multiple TCG contexts, since we will have threads generating code in separate regions of code_gen_buffer. For this we need a new field (.size) in struct tb_tc to keep track of the size of the translated code. This field uses a size_t to avoid adding a hole to the struct, although really an unsigned int would have been enough. The comparison function we use is optimized for the common case: insertions. Profiling shows that upon booting debian-arm, 98% of comparisons are between existing tb's (i.e. a->size and b->size are both !0), which happens during insertions (and removals, but those are rare). The remaining cases are lookups. From reading the glib sources we see that the first key is always the lookup key. However, the code does not assume this to always be the case because this behaviour is not guaranteed in the glib docs. However, we embed this knowledge in the code as a branch hint for the compiler. Note that tb_free does not free space in the code_gen_buffer anymore, since we cannot easily know whether the tb is the last one inserted in code_gen_buffer. The next patch in this series renames tb_free to tb_remove to reflect this. Performance-wise, lookups in tb_find_pc are the same as before: O(log n). However, insertions are O(log n) instead of O(1), which results in a small slowdown when booting debian-arm: Performance counter stats for 'build/arm-softmmu/qemu-system-arm \ -machine type=virt -nographic -smp 1 -m 4096 \ -netdev user,id=unet,hostfwd=tcp::2222-:22 \ -device virtio-net-device,netdev=unet \ -drive file=img/arm/jessie-arm32.qcow2,id=myblock,index=0,if=none \ -device virtio-blk-device,drive=myblock \ -kernel img/arm/aarch32-current-linux-kernel-only.img \ -append console=ttyAMA0 root=/dev/vda1 \ -name arm,debug-threads=on -smp 1' (10 runs): - Before: 8048.598422 task-clock (msec) # 0.931 CPUs utilized ( +- 0.28% ) 16,974 context-switches # 0.002 M/sec ( +- 0.12% ) 0 cpu-migrations # 0.000 K/sec 10,125 page-faults # 0.001 M/sec ( +- 1.23% ) 35,144,901,879 cycles # 4.367 GHz ( +- 0.14% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,758,252,643 instructions # 1.87 insns per cycle ( +- 0.33% ) 10,871,298,668 branches # 1350.707 M/sec ( +- 0.41% ) 192,322,212 branch-misses # 1.77% of all branches ( +- 0.32% ) 8.640869419 seconds time elapsed ( +- 0.57% ) - After: 8146.242027 task-clock (msec) # 0.923 CPUs utilized ( +- 1.23% ) 17,016 context-switches # 0.002 M/sec ( +- 0.40% ) 0 cpu-migrations # 0.000 K/sec 18,769 page-faults # 0.002 M/sec ( +- 0.45% ) 35,660,956,120 cycles # 4.378 GHz ( +- 1.22% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,095,366,607 instructions # 1.83 insns per cycle ( +- 1.73% ) 10,803,480,261 branches # 1326.192 M/sec ( +- 1.95% ) 195,601,289 branch-misses # 1.81% of all branches ( +- 0.39% ) 8.828660235 seconds time elapsed ( +- 0.38% ) Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-06-24 01:00:11 +02:00
size_t target_size;
size_t max_target_size;
size_t direct_jmp_count;
size_t direct_jmp2_count;
size_t cross_page;
};
static gboolean tb_tree_stats_iter(gpointer key, gpointer value, gpointer data)
{
const TranslationBlock *tb = value;
struct tb_tree_stats *tst = data;
tcg: track TBs with per-region BST's This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-26 22:58:05 +02:00
tst->nb_tbs++;
tst->host_size += tb->tc.size;
translate-all: use a binary search tree to track TBs in TBContext This is a prerequisite for supporting multiple TCG contexts, since we will have threads generating code in separate regions of code_gen_buffer. For this we need a new field (.size) in struct tb_tc to keep track of the size of the translated code. This field uses a size_t to avoid adding a hole to the struct, although really an unsigned int would have been enough. The comparison function we use is optimized for the common case: insertions. Profiling shows that upon booting debian-arm, 98% of comparisons are between existing tb's (i.e. a->size and b->size are both !0), which happens during insertions (and removals, but those are rare). The remaining cases are lookups. From reading the glib sources we see that the first key is always the lookup key. However, the code does not assume this to always be the case because this behaviour is not guaranteed in the glib docs. However, we embed this knowledge in the code as a branch hint for the compiler. Note that tb_free does not free space in the code_gen_buffer anymore, since we cannot easily know whether the tb is the last one inserted in code_gen_buffer. The next patch in this series renames tb_free to tb_remove to reflect this. Performance-wise, lookups in tb_find_pc are the same as before: O(log n). However, insertions are O(log n) instead of O(1), which results in a small slowdown when booting debian-arm: Performance counter stats for 'build/arm-softmmu/qemu-system-arm \ -machine type=virt -nographic -smp 1 -m 4096 \ -netdev user,id=unet,hostfwd=tcp::2222-:22 \ -device virtio-net-device,netdev=unet \ -drive file=img/arm/jessie-arm32.qcow2,id=myblock,index=0,if=none \ -device virtio-blk-device,drive=myblock \ -kernel img/arm/aarch32-current-linux-kernel-only.img \ -append console=ttyAMA0 root=/dev/vda1 \ -name arm,debug-threads=on -smp 1' (10 runs): - Before: 8048.598422 task-clock (msec) # 0.931 CPUs utilized ( +- 0.28% ) 16,974 context-switches # 0.002 M/sec ( +- 0.12% ) 0 cpu-migrations # 0.000 K/sec 10,125 page-faults # 0.001 M/sec ( +- 1.23% ) 35,144,901,879 cycles # 4.367 GHz ( +- 0.14% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,758,252,643 instructions # 1.87 insns per cycle ( +- 0.33% ) 10,871,298,668 branches # 1350.707 M/sec ( +- 0.41% ) 192,322,212 branch-misses # 1.77% of all branches ( +- 0.32% ) 8.640869419 seconds time elapsed ( +- 0.57% ) - After: 8146.242027 task-clock (msec) # 0.923 CPUs utilized ( +- 1.23% ) 17,016 context-switches # 0.002 M/sec ( +- 0.40% ) 0 cpu-migrations # 0.000 K/sec 18,769 page-faults # 0.002 M/sec ( +- 0.45% ) 35,660,956,120 cycles # 4.378 GHz ( +- 1.22% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,095,366,607 instructions # 1.83 insns per cycle ( +- 1.73% ) 10,803,480,261 branches # 1326.192 M/sec ( +- 1.95% ) 195,601,289 branch-misses # 1.81% of all branches ( +- 0.39% ) 8.828660235 seconds time elapsed ( +- 0.38% ) Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-06-24 01:00:11 +02:00
tst->target_size += tb->size;
if (tb->size > tst->max_target_size) {
tst->max_target_size = tb->size;
}
if (tb->page_addr[1] != -1) {
tst->cross_page++;
}
if (tb->jmp_reset_offset[0] != TB_JMP_RESET_OFFSET_INVALID) {
tst->direct_jmp_count++;
if (tb->jmp_reset_offset[1] != TB_JMP_RESET_OFFSET_INVALID) {
tst->direct_jmp2_count++;
}
}
return false;
}
void dump_exec_info(FILE *f, fprintf_function cpu_fprintf)
{
translate-all: use a binary search tree to track TBs in TBContext This is a prerequisite for supporting multiple TCG contexts, since we will have threads generating code in separate regions of code_gen_buffer. For this we need a new field (.size) in struct tb_tc to keep track of the size of the translated code. This field uses a size_t to avoid adding a hole to the struct, although really an unsigned int would have been enough. The comparison function we use is optimized for the common case: insertions. Profiling shows that upon booting debian-arm, 98% of comparisons are between existing tb's (i.e. a->size and b->size are both !0), which happens during insertions (and removals, but those are rare). The remaining cases are lookups. From reading the glib sources we see that the first key is always the lookup key. However, the code does not assume this to always be the case because this behaviour is not guaranteed in the glib docs. However, we embed this knowledge in the code as a branch hint for the compiler. Note that tb_free does not free space in the code_gen_buffer anymore, since we cannot easily know whether the tb is the last one inserted in code_gen_buffer. The next patch in this series renames tb_free to tb_remove to reflect this. Performance-wise, lookups in tb_find_pc are the same as before: O(log n). However, insertions are O(log n) instead of O(1), which results in a small slowdown when booting debian-arm: Performance counter stats for 'build/arm-softmmu/qemu-system-arm \ -machine type=virt -nographic -smp 1 -m 4096 \ -netdev user,id=unet,hostfwd=tcp::2222-:22 \ -device virtio-net-device,netdev=unet \ -drive file=img/arm/jessie-arm32.qcow2,id=myblock,index=0,if=none \ -device virtio-blk-device,drive=myblock \ -kernel img/arm/aarch32-current-linux-kernel-only.img \ -append console=ttyAMA0 root=/dev/vda1 \ -name arm,debug-threads=on -smp 1' (10 runs): - Before: 8048.598422 task-clock (msec) # 0.931 CPUs utilized ( +- 0.28% ) 16,974 context-switches # 0.002 M/sec ( +- 0.12% ) 0 cpu-migrations # 0.000 K/sec 10,125 page-faults # 0.001 M/sec ( +- 1.23% ) 35,144,901,879 cycles # 4.367 GHz ( +- 0.14% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,758,252,643 instructions # 1.87 insns per cycle ( +- 0.33% ) 10,871,298,668 branches # 1350.707 M/sec ( +- 0.41% ) 192,322,212 branch-misses # 1.77% of all branches ( +- 0.32% ) 8.640869419 seconds time elapsed ( +- 0.57% ) - After: 8146.242027 task-clock (msec) # 0.923 CPUs utilized ( +- 1.23% ) 17,016 context-switches # 0.002 M/sec ( +- 0.40% ) 0 cpu-migrations # 0.000 K/sec 18,769 page-faults # 0.002 M/sec ( +- 0.45% ) 35,660,956,120 cycles # 4.378 GHz ( +- 1.22% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,095,366,607 instructions # 1.83 insns per cycle ( +- 1.73% ) 10,803,480,261 branches # 1326.192 M/sec ( +- 1.95% ) 195,601,289 branch-misses # 1.81% of all branches ( +- 0.39% ) 8.828660235 seconds time elapsed ( +- 0.38% ) Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-06-24 01:00:11 +02:00
struct tb_tree_stats tst = {};
translate-all: add tb hash bucket info to 'info jit' dump Examples: - Good hashing, i.e. tb_hash_func5(phys_pc, pc, flags): TB count 715135/2684354 [...] TB hash buckets 388775/524288 (74.15% head buckets used) TB hash occupancy 33.04% avg chain occ. Histogram: [0,10)%|▆ █ ▅▁▃▁▁|[90,100]% TB hash avg chain 1.017 buckets. Histogram: 1|█▁▁|3 - Not-so-good hashing, i.e. tb_hash_func5(phys_pc, pc, 0): TB count 712636/2684354 [...] TB hash buckets 344924/524288 (65.79% head buckets used) TB hash occupancy 31.64% avg chain occ. Histogram: [0,10)%|█ ▆ ▅▁▃▁▂|[90,100]% TB hash avg chain 1.047 buckets. Histogram: 1|█▁▁▁|4 - Bad hashing, i.e. tb_hash_func5(phys_pc, 0, 0): TB count 702818/2684354 [...] TB hash buckets 112741/524288 (21.50% head buckets used) TB hash occupancy 10.15% avg chain occ. Histogram: [0,10)%|█ ▁ ▁▁▁▁▁|[90,100]% TB hash avg chain 2.107 buckets. Histogram: [1.0,10.2)|█▁▁▁▁▁▁▁▁▁|[83.8,93.0] - Good hashing, but no auto-resize: TB count 715634/2684354 TB hash buckets 8192/8192 (100.00% head buckets used) TB hash occupancy 98.30% avg chain occ. Histogram: [95.3,95.8)%|▁▁▃▄▃▄▁▇▁█|[99.5,100.0]% TB hash avg chain 22.070 buckets. Histogram: [15.0,16.7)|▁▂▅▄█▅▁▁▁▁|[30.3,32.0] Acked-by: Sergey Fedorov <sergey.fedorov@linaro.org> Suggested-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1465412133-3029-16-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2016-06-08 20:55:33 +02:00
struct qht_stats hst;
translate-all: use a binary search tree to track TBs in TBContext This is a prerequisite for supporting multiple TCG contexts, since we will have threads generating code in separate regions of code_gen_buffer. For this we need a new field (.size) in struct tb_tc to keep track of the size of the translated code. This field uses a size_t to avoid adding a hole to the struct, although really an unsigned int would have been enough. The comparison function we use is optimized for the common case: insertions. Profiling shows that upon booting debian-arm, 98% of comparisons are between existing tb's (i.e. a->size and b->size are both !0), which happens during insertions (and removals, but those are rare). The remaining cases are lookups. From reading the glib sources we see that the first key is always the lookup key. However, the code does not assume this to always be the case because this behaviour is not guaranteed in the glib docs. However, we embed this knowledge in the code as a branch hint for the compiler. Note that tb_free does not free space in the code_gen_buffer anymore, since we cannot easily know whether the tb is the last one inserted in code_gen_buffer. The next patch in this series renames tb_free to tb_remove to reflect this. Performance-wise, lookups in tb_find_pc are the same as before: O(log n). However, insertions are O(log n) instead of O(1), which results in a small slowdown when booting debian-arm: Performance counter stats for 'build/arm-softmmu/qemu-system-arm \ -machine type=virt -nographic -smp 1 -m 4096 \ -netdev user,id=unet,hostfwd=tcp::2222-:22 \ -device virtio-net-device,netdev=unet \ -drive file=img/arm/jessie-arm32.qcow2,id=myblock,index=0,if=none \ -device virtio-blk-device,drive=myblock \ -kernel img/arm/aarch32-current-linux-kernel-only.img \ -append console=ttyAMA0 root=/dev/vda1 \ -name arm,debug-threads=on -smp 1' (10 runs): - Before: 8048.598422 task-clock (msec) # 0.931 CPUs utilized ( +- 0.28% ) 16,974 context-switches # 0.002 M/sec ( +- 0.12% ) 0 cpu-migrations # 0.000 K/sec 10,125 page-faults # 0.001 M/sec ( +- 1.23% ) 35,144,901,879 cycles # 4.367 GHz ( +- 0.14% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,758,252,643 instructions # 1.87 insns per cycle ( +- 0.33% ) 10,871,298,668 branches # 1350.707 M/sec ( +- 0.41% ) 192,322,212 branch-misses # 1.77% of all branches ( +- 0.32% ) 8.640869419 seconds time elapsed ( +- 0.57% ) - After: 8146.242027 task-clock (msec) # 0.923 CPUs utilized ( +- 1.23% ) 17,016 context-switches # 0.002 M/sec ( +- 0.40% ) 0 cpu-migrations # 0.000 K/sec 18,769 page-faults # 0.002 M/sec ( +- 0.45% ) 35,660,956,120 cycles # 4.378 GHz ( +- 1.22% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,095,366,607 instructions # 1.83 insns per cycle ( +- 1.73% ) 10,803,480,261 branches # 1326.192 M/sec ( +- 1.95% ) 195,601,289 branch-misses # 1.81% of all branches ( +- 0.39% ) 8.828660235 seconds time elapsed ( +- 0.38% ) Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-06-24 01:00:11 +02:00
size_t nb_tbs;
tcg: track TBs with per-region BST's This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-26 22:58:05 +02:00
tcg_tb_foreach(tb_tree_stats_iter, &tst);
nb_tbs = tst.nb_tbs;
/* XXX: avoid using doubles ? */
cpu_fprintf(f, "Translation buffer state:\n");
/*
* Report total code size including the padding and TB structs;
* otherwise users might think "-tb-size" is not honoured.
* For avg host size we use the precise numbers from tb_tree_stats though.
*/
tcg: introduce regions to split code_gen_buffer This is groundwork for supporting multiple TCG contexts. The naive solution here is to split code_gen_buffer statically among the TCG threads; this however results in poor utilization if translation needs are different across TCG threads. What we do here is to add an extra layer of indirection, assigning regions that act just like pages do in virtual memory allocation. (BTW if you are wondering about the chosen naming, I did not want to use blocks or pages because those are already heavily used in QEMU). We use a global lock to serialize allocations as well as statistics reporting (we now export the size of the used code_gen_buffer with tcg_code_size()). Note that for the allocator we could just use a counter and atomic_inc; however, that would complicate the gathering of tcg_code_size()-like stats. So given that the region operations are not a fast path, a lock seems the most reasonable choice. The effectiveness of this approach is clear after seeing some numbers. I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark. Note that I'm evaluating this after enabling per-thread TCG (which is done by a subsequent commit). * -smp 1, 1 region (entire buffer): qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357 qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363 qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364 qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373 qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373 qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360 qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370 qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367 That is, 8 flushes. * -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]: qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356 qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361 qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361 qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375 qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375 qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360 qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365 qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368 Again, 8 flushes. Note how buffer utilization is not 100%, but it is close. Smaller region sizes would yield higher utilization, but we want region allocation to be rare (it acquires a lock), so we do not want to go too small. * -smp 8, static partitioning of 8 regions (10 MB per region): qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354 qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370 qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365 qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377 qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358 qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367 qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364 qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358 qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362 qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372 qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374 qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376 qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374 qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372 qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359 qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362 qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368 qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378 qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367 qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364 That is, 20 flushes. Note how a static partitioning approach uses the code buffer poorly, leading to many unnecessary flushes. Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-08 01:24:20 +02:00
cpu_fprintf(f, "gen code size %zu/%zu\n",
tcg_code_size(), tcg_code_capacity());
translate-all: use a binary search tree to track TBs in TBContext This is a prerequisite for supporting multiple TCG contexts, since we will have threads generating code in separate regions of code_gen_buffer. For this we need a new field (.size) in struct tb_tc to keep track of the size of the translated code. This field uses a size_t to avoid adding a hole to the struct, although really an unsigned int would have been enough. The comparison function we use is optimized for the common case: insertions. Profiling shows that upon booting debian-arm, 98% of comparisons are between existing tb's (i.e. a->size and b->size are both !0), which happens during insertions (and removals, but those are rare). The remaining cases are lookups. From reading the glib sources we see that the first key is always the lookup key. However, the code does not assume this to always be the case because this behaviour is not guaranteed in the glib docs. However, we embed this knowledge in the code as a branch hint for the compiler. Note that tb_free does not free space in the code_gen_buffer anymore, since we cannot easily know whether the tb is the last one inserted in code_gen_buffer. The next patch in this series renames tb_free to tb_remove to reflect this. Performance-wise, lookups in tb_find_pc are the same as before: O(log n). However, insertions are O(log n) instead of O(1), which results in a small slowdown when booting debian-arm: Performance counter stats for 'build/arm-softmmu/qemu-system-arm \ -machine type=virt -nographic -smp 1 -m 4096 \ -netdev user,id=unet,hostfwd=tcp::2222-:22 \ -device virtio-net-device,netdev=unet \ -drive file=img/arm/jessie-arm32.qcow2,id=myblock,index=0,if=none \ -device virtio-blk-device,drive=myblock \ -kernel img/arm/aarch32-current-linux-kernel-only.img \ -append console=ttyAMA0 root=/dev/vda1 \ -name arm,debug-threads=on -smp 1' (10 runs): - Before: 8048.598422 task-clock (msec) # 0.931 CPUs utilized ( +- 0.28% ) 16,974 context-switches # 0.002 M/sec ( +- 0.12% ) 0 cpu-migrations # 0.000 K/sec 10,125 page-faults # 0.001 M/sec ( +- 1.23% ) 35,144,901,879 cycles # 4.367 GHz ( +- 0.14% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,758,252,643 instructions # 1.87 insns per cycle ( +- 0.33% ) 10,871,298,668 branches # 1350.707 M/sec ( +- 0.41% ) 192,322,212 branch-misses # 1.77% of all branches ( +- 0.32% ) 8.640869419 seconds time elapsed ( +- 0.57% ) - After: 8146.242027 task-clock (msec) # 0.923 CPUs utilized ( +- 1.23% ) 17,016 context-switches # 0.002 M/sec ( +- 0.40% ) 0 cpu-migrations # 0.000 K/sec 18,769 page-faults # 0.002 M/sec ( +- 0.45% ) 35,660,956,120 cycles # 4.378 GHz ( +- 1.22% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,095,366,607 instructions # 1.83 insns per cycle ( +- 1.73% ) 10,803,480,261 branches # 1326.192 M/sec ( +- 1.95% ) 195,601,289 branch-misses # 1.81% of all branches ( +- 0.39% ) 8.828660235 seconds time elapsed ( +- 0.38% ) Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-06-24 01:00:11 +02:00
cpu_fprintf(f, "TB count %zu\n", nb_tbs);
cpu_fprintf(f, "TB avg target size %zu max=%zu bytes\n",
nb_tbs ? tst.target_size / nb_tbs : 0,
tst.max_target_size);
cpu_fprintf(f, "TB avg host size %zu bytes (expansion ratio: %0.1f)\n",
nb_tbs ? tst.host_size / nb_tbs : 0,
tst.target_size ? (double)tst.host_size / tst.target_size : 0);
translate-all: use a binary search tree to track TBs in TBContext This is a prerequisite for supporting multiple TCG contexts, since we will have threads generating code in separate regions of code_gen_buffer. For this we need a new field (.size) in struct tb_tc to keep track of the size of the translated code. This field uses a size_t to avoid adding a hole to the struct, although really an unsigned int would have been enough. The comparison function we use is optimized for the common case: insertions. Profiling shows that upon booting debian-arm, 98% of comparisons are between existing tb's (i.e. a->size and b->size are both !0), which happens during insertions (and removals, but those are rare). The remaining cases are lookups. From reading the glib sources we see that the first key is always the lookup key. However, the code does not assume this to always be the case because this behaviour is not guaranteed in the glib docs. However, we embed this knowledge in the code as a branch hint for the compiler. Note that tb_free does not free space in the code_gen_buffer anymore, since we cannot easily know whether the tb is the last one inserted in code_gen_buffer. The next patch in this series renames tb_free to tb_remove to reflect this. Performance-wise, lookups in tb_find_pc are the same as before: O(log n). However, insertions are O(log n) instead of O(1), which results in a small slowdown when booting debian-arm: Performance counter stats for 'build/arm-softmmu/qemu-system-arm \ -machine type=virt -nographic -smp 1 -m 4096 \ -netdev user,id=unet,hostfwd=tcp::2222-:22 \ -device virtio-net-device,netdev=unet \ -drive file=img/arm/jessie-arm32.qcow2,id=myblock,index=0,if=none \ -device virtio-blk-device,drive=myblock \ -kernel img/arm/aarch32-current-linux-kernel-only.img \ -append console=ttyAMA0 root=/dev/vda1 \ -name arm,debug-threads=on -smp 1' (10 runs): - Before: 8048.598422 task-clock (msec) # 0.931 CPUs utilized ( +- 0.28% ) 16,974 context-switches # 0.002 M/sec ( +- 0.12% ) 0 cpu-migrations # 0.000 K/sec 10,125 page-faults # 0.001 M/sec ( +- 1.23% ) 35,144,901,879 cycles # 4.367 GHz ( +- 0.14% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,758,252,643 instructions # 1.87 insns per cycle ( +- 0.33% ) 10,871,298,668 branches # 1350.707 M/sec ( +- 0.41% ) 192,322,212 branch-misses # 1.77% of all branches ( +- 0.32% ) 8.640869419 seconds time elapsed ( +- 0.57% ) - After: 8146.242027 task-clock (msec) # 0.923 CPUs utilized ( +- 1.23% ) 17,016 context-switches # 0.002 M/sec ( +- 0.40% ) 0 cpu-migrations # 0.000 K/sec 18,769 page-faults # 0.002 M/sec ( +- 0.45% ) 35,660,956,120 cycles # 4.378 GHz ( +- 1.22% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 65,095,366,607 instructions # 1.83 insns per cycle ( +- 1.73% ) 10,803,480,261 branches # 1326.192 M/sec ( +- 1.95% ) 195,601,289 branch-misses # 1.81% of all branches ( +- 0.39% ) 8.828660235 seconds time elapsed ( +- 0.38% ) Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-06-24 01:00:11 +02:00
cpu_fprintf(f, "cross page TB count %zu (%zu%%)\n", tst.cross_page,
nb_tbs ? (tst.cross_page * 100) / nb_tbs : 0);
cpu_fprintf(f, "direct jump count %zu (%zu%%) (2 jumps=%zu %zu%%)\n",
tst.direct_jmp_count,
nb_tbs ? (tst.direct_jmp_count * 100) / nb_tbs : 0,
tst.direct_jmp2_count,
nb_tbs ? (tst.direct_jmp2_count * 100) / nb_tbs : 0);
translate-all: add tb hash bucket info to 'info jit' dump Examples: - Good hashing, i.e. tb_hash_func5(phys_pc, pc, flags): TB count 715135/2684354 [...] TB hash buckets 388775/524288 (74.15% head buckets used) TB hash occupancy 33.04% avg chain occ. Histogram: [0,10)%|▆ █ ▅▁▃▁▁|[90,100]% TB hash avg chain 1.017 buckets. Histogram: 1|█▁▁|3 - Not-so-good hashing, i.e. tb_hash_func5(phys_pc, pc, 0): TB count 712636/2684354 [...] TB hash buckets 344924/524288 (65.79% head buckets used) TB hash occupancy 31.64% avg chain occ. Histogram: [0,10)%|█ ▆ ▅▁▃▁▂|[90,100]% TB hash avg chain 1.047 buckets. Histogram: 1|█▁▁▁|4 - Bad hashing, i.e. tb_hash_func5(phys_pc, 0, 0): TB count 702818/2684354 [...] TB hash buckets 112741/524288 (21.50% head buckets used) TB hash occupancy 10.15% avg chain occ. Histogram: [0,10)%|█ ▁ ▁▁▁▁▁|[90,100]% TB hash avg chain 2.107 buckets. Histogram: [1.0,10.2)|█▁▁▁▁▁▁▁▁▁|[83.8,93.0] - Good hashing, but no auto-resize: TB count 715634/2684354 TB hash buckets 8192/8192 (100.00% head buckets used) TB hash occupancy 98.30% avg chain occ. Histogram: [95.3,95.8)%|▁▁▃▄▃▄▁▇▁█|[99.5,100.0]% TB hash avg chain 22.070 buckets. Histogram: [15.0,16.7)|▁▂▅▄█▅▁▁▁▁|[30.3,32.0] Acked-by: Sergey Fedorov <sergey.fedorov@linaro.org> Suggested-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1465412133-3029-16-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2016-06-08 20:55:33 +02:00
qht_statistics_init(&tb_ctx.htable, &hst);
print_qht_statistics(f, cpu_fprintf, hst);
translate-all: add tb hash bucket info to 'info jit' dump Examples: - Good hashing, i.e. tb_hash_func5(phys_pc, pc, flags): TB count 715135/2684354 [...] TB hash buckets 388775/524288 (74.15% head buckets used) TB hash occupancy 33.04% avg chain occ. Histogram: [0,10)%|▆ █ ▅▁▃▁▁|[90,100]% TB hash avg chain 1.017 buckets. Histogram: 1|█▁▁|3 - Not-so-good hashing, i.e. tb_hash_func5(phys_pc, pc, 0): TB count 712636/2684354 [...] TB hash buckets 344924/524288 (65.79% head buckets used) TB hash occupancy 31.64% avg chain occ. Histogram: [0,10)%|█ ▆ ▅▁▃▁▂|[90,100]% TB hash avg chain 1.047 buckets. Histogram: 1|█▁▁▁|4 - Bad hashing, i.e. tb_hash_func5(phys_pc, 0, 0): TB count 702818/2684354 [...] TB hash buckets 112741/524288 (21.50% head buckets used) TB hash occupancy 10.15% avg chain occ. Histogram: [0,10)%|█ ▁ ▁▁▁▁▁|[90,100]% TB hash avg chain 2.107 buckets. Histogram: [1.0,10.2)|█▁▁▁▁▁▁▁▁▁|[83.8,93.0] - Good hashing, but no auto-resize: TB count 715634/2684354 TB hash buckets 8192/8192 (100.00% head buckets used) TB hash occupancy 98.30% avg chain occ. Histogram: [95.3,95.8)%|▁▁▃▄▃▄▁▇▁█|[99.5,100.0]% TB hash avg chain 22.070 buckets. Histogram: [15.0,16.7)|▁▂▅▄█▅▁▁▁▁|[30.3,32.0] Acked-by: Sergey Fedorov <sergey.fedorov@linaro.org> Suggested-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Richard Henderson <rth@twiddle.net> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1465412133-3029-16-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net>
2016-06-08 20:55:33 +02:00
qht_statistics_destroy(&hst);
cpu_fprintf(f, "\nStatistics:\n");
cpu_fprintf(f, "TB flush count %u\n",
atomic_read(&tb_ctx.tb_flush_count));
cpu_fprintf(f, "TB invalidate count %zu\n", tcg_tb_phys_invalidate_count());
cpu_fprintf(f, "TLB flush count %zu\n", tlb_flush_count());
tcg_dump_info(f, cpu_fprintf);
}
void dump_opcount_info(FILE *f, fprintf_function cpu_fprintf)
{
tcg_dump_op_count(f, cpu_fprintf);
}
#else /* CONFIG_USER_ONLY */
void cpu_interrupt(CPUState *cpu, int mask)
{
tcg: drop global lock during TCG code execution This finally allows TCG to benefit from the iothread introduction: Drop the global mutex while running pure TCG CPU code. Reacquire the lock when entering MMIO or PIO emulation, or when leaving the TCG loop. We have to revert a few optimization for the current TCG threading model, namely kicking the TCG thread in qemu_mutex_lock_iothread and not kicking it in qemu_cpu_kick. We also need to disable RAM block reordering until we have a more efficient locking mechanism at hand. Still, a Linux x86 UP guest and my Musicpal ARM model boot fine here. These numbers demonstrate where we gain something: 20338 jan 20 0 331m 75m 6904 R 99 0.9 0:50.95 qemu-system-arm 20337 jan 20 0 331m 75m 6904 S 20 0.9 0:26.50 qemu-system-arm The guest CPU was fully loaded, but the iothread could still run mostly independent on a second core. Without the patch we don't get beyond 32206 jan 20 0 330m 73m 7036 R 82 0.9 1:06.00 qemu-system-arm 32204 jan 20 0 330m 73m 7036 S 21 0.9 0:17.03 qemu-system-arm We don't benefit significantly, though, when the guest is not fully loading a host CPU. Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com> Message-Id: <1439220437-23957-10-git-send-email-fred.konrad@greensocs.com> [FK: Rebase, fix qemu_devices_reset deadlock, rm address_space_* mutex] Signed-off-by: KONRAD Frederic <fred.konrad@greensocs.com> [EGC: fixed iothread lock for cpu-exec IRQ handling] Signed-off-by: Emilio G. Cota <cota@braap.org> [AJB: -smp single-threaded fix, clean commit msg, BQL fixes] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Pranith Kumar <bobby.prani@gmail.com> [PM: target-arm changes] Acked-by: Peter Maydell <peter.maydell@linaro.org>
2017-02-23 19:29:11 +01:00
g_assert(qemu_mutex_iothread_locked());
cpu->interrupt_request |= mask;
cpu->icount_decr.u16.high = -1;
}
/*
* Walks guest process memory "regions" one by one
* and calls callback function 'fn' for each region.
*/
struct walk_memory_regions_data {
walk_memory_regions_fn fn;
void *priv;
target_ulong start;
int prot;
};
static int walk_memory_regions_end(struct walk_memory_regions_data *data,
target_ulong end, int new_prot)
{
if (data->start != -1u) {
int rc = data->fn(data->priv, data->start, end, data->prot);
if (rc != 0) {
return rc;
}
}
data->start = (new_prot ? end : -1u);
data->prot = new_prot;
return 0;
}
static int walk_memory_regions_1(struct walk_memory_regions_data *data,
target_ulong base, int level, void **lp)
{
target_ulong pa;
int i, rc;
if (*lp == NULL) {
return walk_memory_regions_end(data, base, 0);
}
if (level == 0) {
PageDesc *pd = *lp;
for (i = 0; i < V_L2_SIZE; ++i) {
int prot = pd[i].flags;
pa = base | (i << TARGET_PAGE_BITS);
if (prot != data->prot) {
rc = walk_memory_regions_end(data, pa, prot);
if (rc != 0) {
return rc;
}
}
}
} else {
void **pp = *lp;
for (i = 0; i < V_L2_SIZE; ++i) {
pa = base | ((target_ulong)i <<
(TARGET_PAGE_BITS + V_L2_BITS * level));
rc = walk_memory_regions_1(data, pa, level - 1, pp + i);
if (rc != 0) {
return rc;
}
}
}
return 0;
}
int walk_memory_regions(void *priv, walk_memory_regions_fn fn)
{
struct walk_memory_regions_data data;
uintptr_t i, l1_sz = v_l1_size;
data.fn = fn;
data.priv = priv;
data.start = -1u;
data.prot = 0;
for (i = 0; i < l1_sz; i++) {
target_ulong base = i << (v_l1_shift + TARGET_PAGE_BITS);
int rc = walk_memory_regions_1(&data, base, v_l2_levels, l1_map + i);
if (rc != 0) {
return rc;
}
}
return walk_memory_regions_end(&data, 0, 0);
}
static int dump_region(void *priv, target_ulong start,
target_ulong end, unsigned long prot)
{
FILE *f = (FILE *)priv;
(void) fprintf(f, TARGET_FMT_lx"-"TARGET_FMT_lx
" "TARGET_FMT_lx" %c%c%c\n",
start, end, end - start,
((prot & PAGE_READ) ? 'r' : '-'),
((prot & PAGE_WRITE) ? 'w' : '-'),
((prot & PAGE_EXEC) ? 'x' : '-'));
return 0;
}
/* dump memory mappings */
void page_dump(FILE *f)
{
const int length = sizeof(target_ulong) * 2;
(void) fprintf(f, "%-*s %-*s %-*s %s\n",
length, "start", length, "end", length, "size", "prot");
walk_memory_regions(f, dump_region);
}
int page_get_flags(target_ulong address)
{
PageDesc *p;
p = page_find(address >> TARGET_PAGE_BITS);
if (!p) {
return 0;
}
return p->flags;
}
/* Modify the flags of a page and invalidate the code if necessary.
The flag PAGE_WRITE_ORG is positioned automatically depending
on PAGE_WRITE. The mmap_lock should already be held. */
void page_set_flags(target_ulong start, target_ulong end, int flags)
{
target_ulong addr, len;
/* This function should never be called with addresses outside the
guest address space. If this assert fires, it probably indicates
a missing call to h2g_valid. */
#if TARGET_ABI_BITS > L1_MAP_ADDR_SPACE_BITS
assert(end <= ((target_ulong)1 << L1_MAP_ADDR_SPACE_BITS));
#endif
assert(start < end);
assert_memory_lock();
start = start & TARGET_PAGE_MASK;
end = TARGET_PAGE_ALIGN(end);
if (flags & PAGE_WRITE) {
flags |= PAGE_WRITE_ORG;
}
for (addr = start, len = end - start;
len != 0;
len -= TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
PageDesc *p = page_find_alloc(addr >> TARGET_PAGE_BITS, 1);
/* If the write protection bit is set, then we invalidate
the code inside. */
if (!(p->flags & PAGE_WRITE) &&
(flags & PAGE_WRITE) &&
p->first_tb) {
tb_invalidate_phys_page(addr, 0);
}
p->flags = flags;
}
}
int page_check_range(target_ulong start, target_ulong len, int flags)
{
PageDesc *p;
target_ulong end;
target_ulong addr;
/* This function should never be called with addresses outside the
guest address space. If this assert fires, it probably indicates
a missing call to h2g_valid. */
#if TARGET_ABI_BITS > L1_MAP_ADDR_SPACE_BITS
assert(start < ((target_ulong)1 << L1_MAP_ADDR_SPACE_BITS));
#endif
if (len == 0) {
return 0;
}
if (start + len - 1 < start) {
/* We've wrapped around. */
return -1;
}
/* must do before we loose bits in the next step */
end = TARGET_PAGE_ALIGN(start + len);
start = start & TARGET_PAGE_MASK;
for (addr = start, len = end - start;
len != 0;
len -= TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
p = page_find(addr >> TARGET_PAGE_BITS);
if (!p) {
return -1;
}
if (!(p->flags & PAGE_VALID)) {
return -1;
}
if ((flags & PAGE_READ) && !(p->flags & PAGE_READ)) {
return -1;
}
if (flags & PAGE_WRITE) {
if (!(p->flags & PAGE_WRITE_ORG)) {
return -1;
}
/* unprotect the page if it was put read-only because it
contains translated code */
if (!(p->flags & PAGE_WRITE)) {
if (!page_unprotect(addr, 0)) {
return -1;
}
}
}
}
return 0;
}
/* called from signal handler: invalidate the code and unprotect the
* page. Return 0 if the fault was not handled, 1 if it was handled,
* and 2 if it was handled but the caller must cause the TB to be
* immediately exited. (We can only return 2 if the 'pc' argument is
* non-zero.)
*/
int page_unprotect(target_ulong address, uintptr_t pc)
{
unsigned int prot;
bool current_tb_invalidated;
PageDesc *p;
target_ulong host_start, host_end, addr;
/* Technically this isn't safe inside a signal handler. However we
know this only ever happens in a synchronous SEGV handler, so in
practice it seems to be ok. */
mmap_lock();
p = page_find(address >> TARGET_PAGE_BITS);
if (!p) {
mmap_unlock();
return 0;
}
/* if the page was really writable, then we change its
protection back to writable */
page_unprotect(): handle calls to pages that are PAGE_WRITE If multiple guest threads in user-mode emulation write to a page which QEMU has marked read-only because of cached TCG translations, the threads can race in page_unprotect: * threads A & B both try to do a write to a page with code in it at the same time (ie which we've made non-writeable, so SEGV) * they race into the signal handler with this faulting address * thread A happens to get to page_unprotect() first and takes the mmap lock, so thread B sits waiting for it to be done * A then finds the page, marks it PAGE_WRITE and mprotect()s it writable * A can then continue OK (returns from signal handler to retry the memory access) * ...but when B gets the mmap lock it finds that the page is already PAGE_WRITE, and so it exits page_unprotect() via the "not due to protected translation" code path, and wrongly delivers the signal to the guest rather than just retrying the access In particular, this meant that trying to run 'javac' in user-mode emulation would fail with a spurious guest SIGSEGV. Handle this by making page_unprotect() assume that a call for a page which is already PAGE_WRITE is due to a race of this sort and return a "fault handled" indication. Since this would cause an infinite loop if we ever called page_unprotect() for some other kind of fault than "write failed due to bad access permissions", tighten the condition in handle_cpu_signal() to check the signal number and si_code, and add a comment so that if somebody does ever find themselves debugging an infinite loop of faults they have some clue about why. (The trick for identifying the correct setting for current_tb_invalidated for thread B (needed to handle the precise-SMC case) is due to Richard Henderson. Paolo Bonzini suggested just relying on si_code rather than trying anything more complicated.) Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Message-Id: <1511879725-9576-3-git-send-email-peter.maydell@linaro.org> Signed-off-by: Laurent Vivier <laurent@vivier.eu>
2017-11-28 15:35:25 +01:00
if (p->flags & PAGE_WRITE_ORG) {
current_tb_invalidated = false;
page_unprotect(): handle calls to pages that are PAGE_WRITE If multiple guest threads in user-mode emulation write to a page which QEMU has marked read-only because of cached TCG translations, the threads can race in page_unprotect: * threads A & B both try to do a write to a page with code in it at the same time (ie which we've made non-writeable, so SEGV) * they race into the signal handler with this faulting address * thread A happens to get to page_unprotect() first and takes the mmap lock, so thread B sits waiting for it to be done * A then finds the page, marks it PAGE_WRITE and mprotect()s it writable * A can then continue OK (returns from signal handler to retry the memory access) * ...but when B gets the mmap lock it finds that the page is already PAGE_WRITE, and so it exits page_unprotect() via the "not due to protected translation" code path, and wrongly delivers the signal to the guest rather than just retrying the access In particular, this meant that trying to run 'javac' in user-mode emulation would fail with a spurious guest SIGSEGV. Handle this by making page_unprotect() assume that a call for a page which is already PAGE_WRITE is due to a race of this sort and return a "fault handled" indication. Since this would cause an infinite loop if we ever called page_unprotect() for some other kind of fault than "write failed due to bad access permissions", tighten the condition in handle_cpu_signal() to check the signal number and si_code, and add a comment so that if somebody does ever find themselves debugging an infinite loop of faults they have some clue about why. (The trick for identifying the correct setting for current_tb_invalidated for thread B (needed to handle the precise-SMC case) is due to Richard Henderson. Paolo Bonzini suggested just relying on si_code rather than trying anything more complicated.) Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Message-Id: <1511879725-9576-3-git-send-email-peter.maydell@linaro.org> Signed-off-by: Laurent Vivier <laurent@vivier.eu>
2017-11-28 15:35:25 +01:00
if (p->flags & PAGE_WRITE) {
/* If the page is actually marked WRITE then assume this is because
* this thread raced with another one which got here first and
* set the page to PAGE_WRITE and did the TB invalidate for us.
*/
#ifdef TARGET_HAS_PRECISE_SMC
tcg: track TBs with per-region BST's This paves the way for enabling scalable parallel generation of TCG code. Instead of tracking TBs with a single binary search tree (BST), use a BST for each TCG region, protecting it with a lock. This is as scalable as it gets, since each TCG thread operates on a separate region. The core of this change is the introduction of struct tcg_region_tree, which contains a pointer to a GTree and an associated lock to serialize accesses to it. We then allocate an array of tcg_region_tree's, adding the appropriate padding to avoid false sharing based on qemu_dcache_linesize. Given a tc_ptr, we first find the corresponding region_tree. This is done by special-casing the first and last regions first, since they might be of size != region.size; otherwise we just divide the offset by region.stride. I was worried about this division (several dozen cycles of latency), but profiling shows that this is not a fast path. Note that region.stride is not required to be a power of two; it is only required to be a multiple of the host's page size. Note that with this design we can also provide consistent snapshots about all region trees at once; for instance, tcg_tb_foreach acquires/releases all region_tree locks before/after iterating over them. For this reason we now drop tb_lock in dump_exec_info(). As an alternative I considered implementing a concurrent BST, but this can be tricky to get right, offers no consistent snapshots of the BST, and performance and scalability-wise I don't think it could ever beat having separate GTrees, given that our workload is insert-mostly (all concurrent BST designs I've seen focus, understandably, on making lookups fast, which comes at the expense of convoluted, non-wait-free insertions/removals). Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-07-26 22:58:05 +02:00
TranslationBlock *current_tb = tcg_tb_lookup(pc);
page_unprotect(): handle calls to pages that are PAGE_WRITE If multiple guest threads in user-mode emulation write to a page which QEMU has marked read-only because of cached TCG translations, the threads can race in page_unprotect: * threads A & B both try to do a write to a page with code in it at the same time (ie which we've made non-writeable, so SEGV) * they race into the signal handler with this faulting address * thread A happens to get to page_unprotect() first and takes the mmap lock, so thread B sits waiting for it to be done * A then finds the page, marks it PAGE_WRITE and mprotect()s it writable * A can then continue OK (returns from signal handler to retry the memory access) * ...but when B gets the mmap lock it finds that the page is already PAGE_WRITE, and so it exits page_unprotect() via the "not due to protected translation" code path, and wrongly delivers the signal to the guest rather than just retrying the access In particular, this meant that trying to run 'javac' in user-mode emulation would fail with a spurious guest SIGSEGV. Handle this by making page_unprotect() assume that a call for a page which is already PAGE_WRITE is due to a race of this sort and return a "fault handled" indication. Since this would cause an infinite loop if we ever called page_unprotect() for some other kind of fault than "write failed due to bad access permissions", tighten the condition in handle_cpu_signal() to check the signal number and si_code, and add a comment so that if somebody does ever find themselves debugging an infinite loop of faults they have some clue about why. (The trick for identifying the correct setting for current_tb_invalidated for thread B (needed to handle the precise-SMC case) is due to Richard Henderson. Paolo Bonzini suggested just relying on si_code rather than trying anything more complicated.) Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Message-Id: <1511879725-9576-3-git-send-email-peter.maydell@linaro.org> Signed-off-by: Laurent Vivier <laurent@vivier.eu>
2017-11-28 15:35:25 +01:00
if (current_tb) {
current_tb_invalidated = tb_cflags(current_tb) & CF_INVALID;
}
#endif
page_unprotect(): handle calls to pages that are PAGE_WRITE If multiple guest threads in user-mode emulation write to a page which QEMU has marked read-only because of cached TCG translations, the threads can race in page_unprotect: * threads A & B both try to do a write to a page with code in it at the same time (ie which we've made non-writeable, so SEGV) * they race into the signal handler with this faulting address * thread A happens to get to page_unprotect() first and takes the mmap lock, so thread B sits waiting for it to be done * A then finds the page, marks it PAGE_WRITE and mprotect()s it writable * A can then continue OK (returns from signal handler to retry the memory access) * ...but when B gets the mmap lock it finds that the page is already PAGE_WRITE, and so it exits page_unprotect() via the "not due to protected translation" code path, and wrongly delivers the signal to the guest rather than just retrying the access In particular, this meant that trying to run 'javac' in user-mode emulation would fail with a spurious guest SIGSEGV. Handle this by making page_unprotect() assume that a call for a page which is already PAGE_WRITE is due to a race of this sort and return a "fault handled" indication. Since this would cause an infinite loop if we ever called page_unprotect() for some other kind of fault than "write failed due to bad access permissions", tighten the condition in handle_cpu_signal() to check the signal number and si_code, and add a comment so that if somebody does ever find themselves debugging an infinite loop of faults they have some clue about why. (The trick for identifying the correct setting for current_tb_invalidated for thread B (needed to handle the precise-SMC case) is due to Richard Henderson. Paolo Bonzini suggested just relying on si_code rather than trying anything more complicated.) Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Message-Id: <1511879725-9576-3-git-send-email-peter.maydell@linaro.org> Signed-off-by: Laurent Vivier <laurent@vivier.eu>
2017-11-28 15:35:25 +01:00
} else {
host_start = address & qemu_host_page_mask;
host_end = host_start + qemu_host_page_size;
prot = 0;
for (addr = host_start; addr < host_end; addr += TARGET_PAGE_SIZE) {
p = page_find(addr >> TARGET_PAGE_BITS);
p->flags |= PAGE_WRITE;
prot |= p->flags;
/* and since the content will be modified, we must invalidate
the corresponding translated code. */
current_tb_invalidated |= tb_invalidate_phys_page(addr, pc);
#ifdef CONFIG_USER_ONLY
if (DEBUG_TB_CHECK_GATE) {
tb_invalidate_check(addr);
}
#endif
}
mprotect((void *)g2h(host_start), qemu_host_page_size,
prot & PAGE_BITS);
}
mmap_unlock();
/* If current TB was invalidated return to main loop */
return current_tb_invalidated ? 2 : 1;
}
mmap_unlock();
return 0;
}
#endif /* CONFIG_USER_ONLY */
/* This is a wrapper for common code that can not use CONFIG_SOFTMMU */
void tcg_flush_softmmu_tlb(CPUState *cs)
{
#ifdef CONFIG_SOFTMMU
tlb_flush(cs);
#endif
}