qemu-e2k/cpu-exec.c

874 lines
34 KiB
C
Raw Normal View History

/*
* emulator main execution loop
*
* Copyright (c) 2003-2005 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/>.
*/
#include "config.h"
#include "cpu.h"
#include "trace.h"
#include "disas/disas.h"
#include "tcg.h"
#include "qemu/atomic.h"
#include "sysemu/qtest.h"
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 11:56:31 +02:00
#include "qemu/timer.h"
/* -icount align implementation. */
typedef struct SyncClocks {
int64_t diff_clk;
int64_t last_cpu_icount;
int64_t realtime_clock;
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 11:56:31 +02:00
} SyncClocks;
#if !defined(CONFIG_USER_ONLY)
/* Allow the guest to have a max 3ms advance.
* The difference between the 2 clocks could therefore
* oscillate around 0.
*/
#define VM_CLOCK_ADVANCE 3000000
#define THRESHOLD_REDUCE 1.5
#define MAX_DELAY_PRINT_RATE 2000000000LL
#define MAX_NB_PRINTS 100
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 11:56:31 +02:00
static void align_clocks(SyncClocks *sc, const CPUState *cpu)
{
int64_t cpu_icount;
if (!icount_align_option) {
return;
}
cpu_icount = cpu->icount_extra + cpu->icount_decr.u16.low;
sc->diff_clk += cpu_icount_to_ns(sc->last_cpu_icount - cpu_icount);
sc->last_cpu_icount = cpu_icount;
if (sc->diff_clk > VM_CLOCK_ADVANCE) {
#ifndef _WIN32
struct timespec sleep_delay, rem_delay;
sleep_delay.tv_sec = sc->diff_clk / 1000000000LL;
sleep_delay.tv_nsec = sc->diff_clk % 1000000000LL;
if (nanosleep(&sleep_delay, &rem_delay) < 0) {
sc->diff_clk -= (sleep_delay.tv_sec - rem_delay.tv_sec) * 1000000000LL;
sc->diff_clk -= sleep_delay.tv_nsec - rem_delay.tv_nsec;
} else {
sc->diff_clk = 0;
}
#else
Sleep(sc->diff_clk / SCALE_MS);
sc->diff_clk = 0;
#endif
}
}
static void print_delay(const SyncClocks *sc)
{
static float threshold_delay;
static int64_t last_realtime_clock;
static int nb_prints;
if (icount_align_option &&
sc->realtime_clock - last_realtime_clock >= MAX_DELAY_PRINT_RATE &&
nb_prints < MAX_NB_PRINTS) {
if ((-sc->diff_clk / (float)1000000000LL > threshold_delay) ||
(-sc->diff_clk / (float)1000000000LL <
(threshold_delay - THRESHOLD_REDUCE))) {
threshold_delay = (-sc->diff_clk / 1000000000LL) + 1;
printf("Warning: The guest is now late by %.1f to %.1f seconds\n",
threshold_delay - 1,
threshold_delay);
nb_prints++;
last_realtime_clock = sc->realtime_clock;
}
}
}
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 11:56:31 +02:00
static void init_delay_params(SyncClocks *sc,
const CPUState *cpu)
{
if (!icount_align_option) {
return;
}
sc->realtime_clock = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 11:56:31 +02:00
sc->diff_clk = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) -
sc->realtime_clock +
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 11:56:31 +02:00
cpu_get_clock_offset();
sc->last_cpu_icount = cpu->icount_extra + cpu->icount_decr.u16.low;
if (sc->diff_clk < max_delay) {
max_delay = sc->diff_clk;
}
if (sc->diff_clk > max_advance) {
max_advance = sc->diff_clk;
}
/* Print every 2s max if the guest is late. We limit the number
of printed messages to NB_PRINT_MAX(currently 100) */
print_delay(sc);
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 11:56:31 +02:00
}
#else
static void align_clocks(SyncClocks *sc, const CPUState *cpu)
{
}
static void init_delay_params(SyncClocks *sc, const CPUState *cpu)
{
}
#endif /* CONFIG USER ONLY */
void cpu_loop_exit(CPUState *cpu)
{
cpu->current_tb = NULL;
siglongjmp(cpu->jmp_env, 1);
}
/* exit the current TB from a signal handler. The host registers are
restored in a state compatible with the CPU emulator
*/
#if defined(CONFIG_SOFTMMU)
void cpu_resume_from_signal(CPUState *cpu, void *puc)
{
/* XXX: restore cpu registers saved in host registers */
cpu->exception_index = -1;
siglongjmp(cpu->jmp_env, 1);
}
#endif
/* Execute a TB, and fix up the CPU state afterwards if necessary */
static inline tcg_target_ulong cpu_tb_exec(CPUState *cpu, uint8_t *tb_ptr)
{
CPUArchState *env = cpu->env_ptr;
uintptr_t next_tb;
#if defined(DEBUG_DISAS)
if (qemu_loglevel_mask(CPU_LOG_TB_CPU)) {
#if defined(TARGET_I386)
log_cpu_state(cpu, CPU_DUMP_CCOP);
#elif defined(TARGET_M68K)
/* ??? Should not modify env state for dumping. */
cpu_m68k_flush_flags(env, env->cc_op);
env->cc_op = CC_OP_FLAGS;
env->sr = (env->sr & 0xffe0) | env->cc_dest | (env->cc_x << 4);
log_cpu_state(cpu, 0);
#else
log_cpu_state(cpu, 0);
#endif
}
#endif /* DEBUG_DISAS */
next_tb = tcg_qemu_tb_exec(env, tb_ptr);
trace_exec_tb_exit((void *) (next_tb & ~TB_EXIT_MASK),
next_tb & TB_EXIT_MASK);
if ((next_tb & TB_EXIT_MASK) > TB_EXIT_IDX1) {
/* We didn't start executing this TB (eg because the instruction
* counter hit zero); we must restore the guest PC to the address
* of the start of the TB.
*/
CPUClass *cc = CPU_GET_CLASS(cpu);
TranslationBlock *tb = (TranslationBlock *)(next_tb & ~TB_EXIT_MASK);
if (cc->synchronize_from_tb) {
cc->synchronize_from_tb(cpu, tb);
} else {
assert(cc->set_pc);
cc->set_pc(cpu, tb->pc);
}
}
if ((next_tb & TB_EXIT_MASK) == TB_EXIT_REQUESTED) {
/* We were asked to stop executing TBs (probably a pending
* interrupt. We've now stopped, so clear the flag.
*/
cpu->tcg_exit_req = 0;
}
return next_tb;
}
/* Execute the code without caching the generated code. An interpreter
could be used if available. */
static void cpu_exec_nocache(CPUArchState *env, int max_cycles,
TranslationBlock *orig_tb)
{
CPUState *cpu = ENV_GET_CPU(env);
TranslationBlock *tb;
/* Should never happen.
We only end up here when an existing TB is too long. */
if (max_cycles > CF_COUNT_MASK)
max_cycles = CF_COUNT_MASK;
tb = tb_gen_code(cpu, orig_tb->pc, orig_tb->cs_base, orig_tb->flags,
max_cycles);
cpu->current_tb = tb;
/* execute the generated code */
trace_exec_tb_nocache(tb, tb->pc);
cpu_tb_exec(cpu, tb->tc_ptr);
cpu->current_tb = NULL;
tb_phys_invalidate(tb, -1);
tb_free(tb);
}
static TranslationBlock *tb_find_slow(CPUArchState *env,
target_ulong pc,
target_ulong cs_base,
uint64_t flags)
{
CPUState *cpu = ENV_GET_CPU(env);
TranslationBlock *tb, **ptb1;
unsigned int h;
tb_page_addr_t phys_pc, phys_page1;
target_ulong virt_page2;
tcg_ctx.tb_ctx.tb_invalidated_flag = 0;
/* find translated block using physical mappings */
phys_pc = get_page_addr_code(env, pc);
phys_page1 = phys_pc & TARGET_PAGE_MASK;
h = tb_phys_hash_func(phys_pc);
ptb1 = &tcg_ctx.tb_ctx.tb_phys_hash[h];
for(;;) {
tb = *ptb1;
if (!tb)
goto not_found;
if (tb->pc == pc &&
tb->page_addr[0] == phys_page1 &&
tb->cs_base == cs_base &&
tb->flags == flags) {
/* check next page if needed */
if (tb->page_addr[1] != -1) {
tb_page_addr_t phys_page2;
virt_page2 = (pc & TARGET_PAGE_MASK) +
TARGET_PAGE_SIZE;
phys_page2 = get_page_addr_code(env, virt_page2);
if (tb->page_addr[1] == phys_page2)
goto found;
} else {
goto found;
}
}
ptb1 = &tb->phys_hash_next;
}
not_found:
/* if no translated code available, then translate it now */
tb = tb_gen_code(cpu, pc, cs_base, flags, 0);
found:
/* Move the last found TB to the head of the list */
if (likely(*ptb1)) {
*ptb1 = tb->phys_hash_next;
tb->phys_hash_next = tcg_ctx.tb_ctx.tb_phys_hash[h];
tcg_ctx.tb_ctx.tb_phys_hash[h] = tb;
}
/* we add the TB in the virtual pc hash table */
cpu->tb_jmp_cache[tb_jmp_cache_hash_func(pc)] = tb;
return tb;
}
static inline TranslationBlock *tb_find_fast(CPUArchState *env)
{
CPUState *cpu = ENV_GET_CPU(env);
TranslationBlock *tb;
target_ulong cs_base, pc;
int flags;
/* we record a subset of the CPU state. It will
always be the same before a given translated block
is executed. */
cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);
tb = cpu->tb_jmp_cache[tb_jmp_cache_hash_func(pc)];
if (unlikely(!tb || tb->pc != pc || tb->cs_base != cs_base ||
tb->flags != flags)) {
tb = tb_find_slow(env, pc, cs_base, flags);
}
return tb;
}
static CPUDebugExcpHandler *debug_excp_handler;
void cpu_set_debug_excp_handler(CPUDebugExcpHandler *handler)
{
debug_excp_handler = handler;
}
static void cpu_handle_debug_exception(CPUArchState *env)
{
CPUState *cpu = ENV_GET_CPU(env);
CPUWatchpoint *wp;
if (!cpu->watchpoint_hit) {
QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
wp->flags &= ~BP_WATCHPOINT_HIT;
}
}
if (debug_excp_handler) {
debug_excp_handler(env);
}
}
/* main execution loop */
volatile sig_atomic_t exit_request;
int cpu_exec(CPUArchState *env)
{
CPUState *cpu = ENV_GET_CPU(env);
#if !(defined(CONFIG_USER_ONLY) && \
(defined(TARGET_M68K) || defined(TARGET_PPC) || defined(TARGET_S390X)))
CPUClass *cc = CPU_GET_CLASS(cpu);
#endif
#ifdef TARGET_I386
X86CPU *x86_cpu = X86_CPU(cpu);
#endif
int ret, interrupt_request;
TranslationBlock *tb;
uint8_t *tc_ptr;
uintptr_t next_tb;
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 11:56:31 +02:00
SyncClocks sc;
/* This must be volatile so it is not trashed by longjmp() */
volatile bool have_tb_lock = false;
if (cpu->halted) {
if (!cpu_has_work(cpu)) {
return EXCP_HALTED;
}
cpu->halted = 0;
}
current_cpu = cpu;
/* As long as current_cpu is null, up to the assignment just above,
* requests by other threads to exit the execution loop are expected to
* be issued using the exit_request global. We must make sure that our
* evaluation of the global value is performed past the current_cpu
* value transition point, which requires a memory barrier as well as
* an instruction scheduling constraint on modern architectures. */
smp_mb();
if (unlikely(exit_request)) {
cpu->exit_request = 1;
}
#if defined(TARGET_I386)
/* put eflags in CPU temporary format */
CC_SRC = env->eflags & (CC_O | CC_S | CC_Z | CC_A | CC_P | CC_C);
env->df = 1 - (2 * ((env->eflags >> 10) & 1));
CC_OP = CC_OP_EFLAGS;
env->eflags &= ~(DF_MASK | CC_O | CC_S | CC_Z | CC_A | CC_P | CC_C);
#elif defined(TARGET_SPARC)
#elif defined(TARGET_M68K)
env->cc_op = CC_OP_FLAGS;
env->cc_dest = env->sr & 0xf;
env->cc_x = (env->sr >> 4) & 1;
#elif defined(TARGET_ALPHA)
#elif defined(TARGET_ARM)
#elif defined(TARGET_UNICORE32)
#elif defined(TARGET_PPC)
PPC: Fix sync instructions problem in SMP In the current emulation of the load-and-reserve (lwarx) and store-conditional (stwcx.) instructions, the internal reservation mechanism is taken into account, however each CPU has its own reservation information and this information is not synchronized between CPUs to perform proper synchronization. The following test case with 2 CPUs shows that the semantics of the "lwarx" and "stwcx." instructions are not preserved by the emulation. The test case does the following : - CPU0: reserve a memory location - CPU1: reserve the same memory location - CPU0: perform stwcx. on the location The last store-conditional operation succeeds while it is supposed to fail since the reservation was supposed to be lost at the second reserve operation. This (one line) patch fixes this problem in a very simple manner by removing the reservation of a CPU every time it is scheduled (in cpu_exec()). While this is a harsh workaround, it does not affect the guest code much because reservations are usually held for a very short time, that is an lwarx is almost always followed by an stwcx. a few instructions below. Therefore, in most cases, the reservation will be taken and consumed before a CPU switch occurs. However in the rare case where a CPU switch does occur between the lwarx and its corresponding stwcx. this patch solves a potential erroneous behavior of the synchronization instructions. Signed-off-by: Elie Richa <richa@adacore.com> Signed-off-by: Alexander Graf <agraf@suse.de>
2011-07-22 07:58:39 +02:00
env->reserve_addr = -1;
#elif defined(TARGET_LM32)
#elif defined(TARGET_MICROBLAZE)
#elif defined(TARGET_MIPS)
#elif defined(TARGET_MOXIE)
#elif defined(TARGET_OPENRISC)
#elif defined(TARGET_SH4)
#elif defined(TARGET_CRIS)
#elif defined(TARGET_S390X)
#elif defined(TARGET_XTENSA)
#elif defined(TARGET_TRICORE)
/* XXXXX */
#else
#error unsupported target CPU
#endif
cpu->exception_index = -1;
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 11:56:31 +02:00
/* Calculate difference between guest clock and host clock.
* This delay includes the delay of the last cycle, so
* what we have to do is sleep until it is 0. As for the
* advance/delay we gain here, we try to fix it next time.
*/
init_delay_params(&sc, cpu);
/* prepare setjmp context for exception handling */
for(;;) {
if (sigsetjmp(cpu->jmp_env, 0) == 0) {
/* if an exception is pending, we execute it here */
if (cpu->exception_index >= 0) {
if (cpu->exception_index >= EXCP_INTERRUPT) {
/* exit request from the cpu execution loop */
ret = cpu->exception_index;
if (ret == EXCP_DEBUG) {
cpu_handle_debug_exception(env);
}
break;
} else {
#if defined(CONFIG_USER_ONLY)
/* if user mode only, we simulate a fake exception
which will be handled outside the cpu execution
loop */
#if defined(TARGET_I386)
cc->do_interrupt(cpu);
#endif
ret = cpu->exception_index;
break;
#else
cc->do_interrupt(cpu);
cpu->exception_index = -1;
#endif
}
}
next_tb = 0; /* force lookup of first TB */
for(;;) {
interrupt_request = cpu->interrupt_request;
if (unlikely(interrupt_request)) {
if (unlikely(cpu->singlestep_enabled & SSTEP_NOIRQ)) {
/* Mask out external interrupts for this step. */
interrupt_request &= ~CPU_INTERRUPT_SSTEP_MASK;
}
if (interrupt_request & CPU_INTERRUPT_DEBUG) {
cpu->interrupt_request &= ~CPU_INTERRUPT_DEBUG;
cpu->exception_index = EXCP_DEBUG;
cpu_loop_exit(cpu);
}
#if defined(TARGET_ARM) || defined(TARGET_SPARC) || defined(TARGET_MIPS) || \
defined(TARGET_PPC) || defined(TARGET_ALPHA) || defined(TARGET_CRIS) || \
defined(TARGET_MICROBLAZE) || defined(TARGET_LM32) || \
defined(TARGET_UNICORE32) || defined(TARGET_TRICORE)
if (interrupt_request & CPU_INTERRUPT_HALT) {
cpu->interrupt_request &= ~CPU_INTERRUPT_HALT;
cpu->halted = 1;
cpu->exception_index = EXCP_HLT;
cpu_loop_exit(cpu);
}
#endif
#if defined(TARGET_I386)
if (interrupt_request & CPU_INTERRUPT_INIT) {
cpu_svm_check_intercept_param(env, SVM_EXIT_INIT, 0);
do_cpu_init(x86_cpu);
cpu->exception_index = EXCP_HALTED;
cpu_loop_exit(cpu);
}
#else
if (interrupt_request & CPU_INTERRUPT_RESET) {
cpu_reset(cpu);
}
#endif
#if defined(TARGET_I386)
#if !defined(CONFIG_USER_ONLY)
if (interrupt_request & CPU_INTERRUPT_POLL) {
cpu->interrupt_request &= ~CPU_INTERRUPT_POLL;
apic_poll_irq(x86_cpu->apic_state);
}
#endif
if (interrupt_request & CPU_INTERRUPT_SIPI) {
do_cpu_sipi(x86_cpu);
} else if (env->hflags2 & HF2_GIF_MASK) {
if ((interrupt_request & CPU_INTERRUPT_SMI) &&
!(env->hflags & HF_SMM_MASK)) {
cpu_svm_check_intercept_param(env, SVM_EXIT_SMI,
0);
cpu->interrupt_request &= ~CPU_INTERRUPT_SMI;
do_smm_enter(x86_cpu);
next_tb = 0;
} else if ((interrupt_request & CPU_INTERRUPT_NMI) &&
!(env->hflags2 & HF2_NMI_MASK)) {
cpu->interrupt_request &= ~CPU_INTERRUPT_NMI;
env->hflags2 |= HF2_NMI_MASK;
do_interrupt_x86_hardirq(env, EXCP02_NMI, 1);
next_tb = 0;
} else if (interrupt_request & CPU_INTERRUPT_MCE) {
cpu->interrupt_request &= ~CPU_INTERRUPT_MCE;
do_interrupt_x86_hardirq(env, EXCP12_MCHK, 0);
next_tb = 0;
} else if ((interrupt_request & CPU_INTERRUPT_HARD) &&
(((env->hflags2 & HF2_VINTR_MASK) &&
(env->hflags2 & HF2_HIF_MASK)) ||
(!(env->hflags2 & HF2_VINTR_MASK) &&
(env->eflags & IF_MASK &&
!(env->hflags & HF_INHIBIT_IRQ_MASK))))) {
int intno;
cpu_svm_check_intercept_param(env, SVM_EXIT_INTR,
0);
cpu->interrupt_request &= ~(CPU_INTERRUPT_HARD |
CPU_INTERRUPT_VIRQ);
intno = cpu_get_pic_interrupt(env);
qemu_log_mask(CPU_LOG_TB_IN_ASM, "Servicing hardware INT=0x%02x\n", intno);
do_interrupt_x86_hardirq(env, intno, 1);
/* ensure that no TB jump will be modified as
the program flow was changed */
next_tb = 0;
#if !defined(CONFIG_USER_ONLY)
} else if ((interrupt_request & CPU_INTERRUPT_VIRQ) &&
(env->eflags & IF_MASK) &&
!(env->hflags & HF_INHIBIT_IRQ_MASK)) {
int intno;
/* FIXME: this should respect TPR */
cpu_svm_check_intercept_param(env, SVM_EXIT_VINTR,
0);
intno = ldl_phys(cpu->as,
env->vm_vmcb
+ offsetof(struct vmcb,
control.int_vector));
qemu_log_mask(CPU_LOG_TB_IN_ASM, "Servicing virtual hardware INT=0x%02x\n", intno);
do_interrupt_x86_hardirq(env, intno, 1);
cpu->interrupt_request &= ~CPU_INTERRUPT_VIRQ;
next_tb = 0;
#endif
}
}
#elif defined(TARGET_PPC)
if (interrupt_request & CPU_INTERRUPT_HARD) {
ppc_hw_interrupt(env);
if (env->pending_interrupts == 0) {
cpu->interrupt_request &= ~CPU_INTERRUPT_HARD;
}
next_tb = 0;
}
#elif defined(TARGET_LM32)
if ((interrupt_request & CPU_INTERRUPT_HARD)
&& (env->ie & IE_IE)) {
cpu->exception_index = EXCP_IRQ;
cc->do_interrupt(cpu);
next_tb = 0;
}
#elif defined(TARGET_MICROBLAZE)
if ((interrupt_request & CPU_INTERRUPT_HARD)
&& (env->sregs[SR_MSR] & MSR_IE)
&& !(env->sregs[SR_MSR] & (MSR_EIP | MSR_BIP))
&& !(env->iflags & (D_FLAG | IMM_FLAG))) {
cpu->exception_index = EXCP_IRQ;
cc->do_interrupt(cpu);
next_tb = 0;
}
#elif defined(TARGET_MIPS)
if ((interrupt_request & CPU_INTERRUPT_HARD) &&
cpu_mips_hw_interrupts_pending(env)) {
/* Raise it */
cpu->exception_index = EXCP_EXT_INTERRUPT;
env->error_code = 0;
cc->do_interrupt(cpu);
next_tb = 0;
}
#elif defined(TARGET_TRICORE)
if ((interrupt_request & CPU_INTERRUPT_HARD)) {
cc->do_interrupt(cpu);
next_tb = 0;
}
#elif defined(TARGET_OPENRISC)
{
int idx = -1;
if ((interrupt_request & CPU_INTERRUPT_HARD)
&& (env->sr & SR_IEE)) {
idx = EXCP_INT;
}
if ((interrupt_request & CPU_INTERRUPT_TIMER)
&& (env->sr & SR_TEE)) {
idx = EXCP_TICK;
}
if (idx >= 0) {
cpu->exception_index = idx;
cc->do_interrupt(cpu);
next_tb = 0;
}
}
#elif defined(TARGET_SPARC)
if (interrupt_request & CPU_INTERRUPT_HARD) {
if (cpu_interrupts_enabled(env) &&
env->interrupt_index > 0) {
int pil = env->interrupt_index & 0xf;
int type = env->interrupt_index & 0xf0;
if (((type == TT_EXTINT) &&
cpu_pil_allowed(env, pil)) ||
type != TT_EXTINT) {
cpu->exception_index = env->interrupt_index;
cc->do_interrupt(cpu);
next_tb = 0;
}
}
}
#elif defined(TARGET_ARM)
if (interrupt_request & CPU_INTERRUPT_FIQ
&& !(env->daif & PSTATE_F)) {
cpu->exception_index = EXCP_FIQ;
cc->do_interrupt(cpu);
next_tb = 0;
}
/* ARMv7-M interrupt return works by loading a magic value
into the PC. On real hardware the load causes the
return to occur. The qemu implementation performs the
jump normally, then does the exception return when the
CPU tries to execute code at the magic address.
This will cause the magic PC value to be pushed to
the stack if an interrupt occurred at the wrong time.
We avoid this by disabling interrupts when
pc contains a magic address. */
if (interrupt_request & CPU_INTERRUPT_HARD
&& ((IS_M(env) && env->regs[15] < 0xfffffff0)
|| !(env->daif & PSTATE_I))) {
cpu->exception_index = EXCP_IRQ;
cc->do_interrupt(cpu);
next_tb = 0;
}
#elif defined(TARGET_UNICORE32)
if (interrupt_request & CPU_INTERRUPT_HARD
&& !(env->uncached_asr & ASR_I)) {
cpu->exception_index = UC32_EXCP_INTR;
cc->do_interrupt(cpu);
next_tb = 0;
}
#elif defined(TARGET_SH4)
if (interrupt_request & CPU_INTERRUPT_HARD) {
cc->do_interrupt(cpu);
next_tb = 0;
}
#elif defined(TARGET_ALPHA)
{
int idx = -1;
/* ??? This hard-codes the OSF/1 interrupt levels. */
switch (env->pal_mode ? 7 : env->ps & PS_INT_MASK) {
case 0 ... 3:
if (interrupt_request & CPU_INTERRUPT_HARD) {
idx = EXCP_DEV_INTERRUPT;
}
/* FALLTHRU */
case 4:
if (interrupt_request & CPU_INTERRUPT_TIMER) {
idx = EXCP_CLK_INTERRUPT;
}
/* FALLTHRU */
case 5:
if (interrupt_request & CPU_INTERRUPT_SMP) {
idx = EXCP_SMP_INTERRUPT;
}
/* FALLTHRU */
case 6:
if (interrupt_request & CPU_INTERRUPT_MCHK) {
idx = EXCP_MCHK;
}
}
if (idx >= 0) {
cpu->exception_index = idx;
env->error_code = 0;
cc->do_interrupt(cpu);
next_tb = 0;
}
}
#elif defined(TARGET_CRIS)
if (interrupt_request & CPU_INTERRUPT_HARD
&& (env->pregs[PR_CCS] & I_FLAG)
&& !env->locked_irq) {
cpu->exception_index = EXCP_IRQ;
cc->do_interrupt(cpu);
next_tb = 0;
}
if (interrupt_request & CPU_INTERRUPT_NMI) {
unsigned int m_flag_archval;
if (env->pregs[PR_VR] < 32) {
m_flag_archval = M_FLAG_V10;
} else {
m_flag_archval = M_FLAG_V32;
}
if ((env->pregs[PR_CCS] & m_flag_archval)) {
cpu->exception_index = EXCP_NMI;
cc->do_interrupt(cpu);
next_tb = 0;
}
}
#elif defined(TARGET_M68K)
if (interrupt_request & CPU_INTERRUPT_HARD
&& ((env->sr & SR_I) >> SR_I_SHIFT)
< env->pending_level) {
/* Real hardware gets the interrupt vector via an
IACK cycle at this point. Current emulated
hardware doesn't rely on this, so we
provide/save the vector when the interrupt is
first signalled. */
cpu->exception_index = env->pending_vector;
do_interrupt_m68k_hardirq(env);
next_tb = 0;
}
#elif defined(TARGET_S390X) && !defined(CONFIG_USER_ONLY)
if ((interrupt_request & CPU_INTERRUPT_HARD) &&
(env->psw.mask & PSW_MASK_EXT)) {
cc->do_interrupt(cpu);
next_tb = 0;
}
#elif defined(TARGET_XTENSA)
if (interrupt_request & CPU_INTERRUPT_HARD) {
cpu->exception_index = EXC_IRQ;
cc->do_interrupt(cpu);
next_tb = 0;
}
#endif
/* Don't use the cached interrupt_request value,
do_interrupt may have updated the EXITTB flag. */
if (cpu->interrupt_request & CPU_INTERRUPT_EXITTB) {
cpu->interrupt_request &= ~CPU_INTERRUPT_EXITTB;
/* ensure that no TB jump will be modified as
the program flow was changed */
next_tb = 0;
}
}
if (unlikely(cpu->exit_request)) {
cpu->exit_request = 0;
cpu->exception_index = EXCP_INTERRUPT;
cpu_loop_exit(cpu);
}
spin_lock(&tcg_ctx.tb_ctx.tb_lock);
have_tb_lock = true;
tb = tb_find_fast(env);
/* Note: we do it here to avoid a gcc bug on Mac OS X when
doing it in tb_find_slow */
if (tcg_ctx.tb_ctx.tb_invalidated_flag) {
/* as some TB could have been invalidated because
of memory exceptions while generating the code, we
must recompute the hash index here */
next_tb = 0;
tcg_ctx.tb_ctx.tb_invalidated_flag = 0;
}
if (qemu_loglevel_mask(CPU_LOG_EXEC)) {
qemu_log("Trace %p [" TARGET_FMT_lx "] %s\n",
tb->tc_ptr, tb->pc, lookup_symbol(tb->pc));
}
/* see if we can patch the calling TB. When the TB
spans two pages, we cannot safely do a direct
jump. */
if (next_tb != 0 && tb->page_addr[1] == -1) {
tb_add_jump((TranslationBlock *)(next_tb & ~TB_EXIT_MASK),
next_tb & TB_EXIT_MASK, tb);
}
have_tb_lock = false;
spin_unlock(&tcg_ctx.tb_ctx.tb_lock);
/* cpu_interrupt might be called while translating the
TB, but before it is linked into a potentially
infinite loop and becomes env->current_tb. Avoid
starting execution if there is a pending interrupt. */
cpu->current_tb = tb;
barrier();
if (likely(!cpu->exit_request)) {
trace_exec_tb(tb, tb->pc);
tc_ptr = tb->tc_ptr;
/* execute the generated code */
next_tb = cpu_tb_exec(cpu, tc_ptr);
switch (next_tb & TB_EXIT_MASK) {
case TB_EXIT_REQUESTED:
/* Something asked us to stop executing
* chained TBs; just continue round the main
* loop. Whatever requested the exit will also
* have set something else (eg exit_request or
* interrupt_request) which we will handle
* next time around the loop.
*/
tb = (TranslationBlock *)(next_tb & ~TB_EXIT_MASK);
next_tb = 0;
break;
case TB_EXIT_ICOUNT_EXPIRED:
{
/* Instruction counter expired. */
int insns_left;
tb = (TranslationBlock *)(next_tb & ~TB_EXIT_MASK);
insns_left = cpu->icount_decr.u32;
if (cpu->icount_extra && insns_left >= 0) {
/* Refill decrementer and continue execution. */
cpu->icount_extra += insns_left;
if (cpu->icount_extra > 0xffff) {
insns_left = 0xffff;
} else {
insns_left = cpu->icount_extra;
}
cpu->icount_extra -= insns_left;
cpu->icount_decr.u16.low = insns_left;
} else {
if (insns_left > 0) {
/* Execute remaining instructions. */
cpu_exec_nocache(env, insns_left, tb);
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 11:56:31 +02:00
align_clocks(&sc, cpu);
}
cpu->exception_index = EXCP_INTERRUPT;
next_tb = 0;
cpu_loop_exit(cpu);
}
break;
}
default:
break;
}
}
cpu->current_tb = NULL;
cpu-exec: Add sleeping algorithm The goal is to sleep qemu whenever the guest clock is in advance compared to the host clock (we use the monotonic clocks). The amount of time to sleep is calculated in the execution loop in cpu_exec. At first, we tried to approximate at each for loop the real time elapsed while searching for a TB (generating or retrieving from cache) and executing it. We would then approximate the virtual time corresponding to the number of virtual instructions executed. The difference between these 2 values would allow us to know if the guest is in advance or delayed. However, the function used for measuring the real time (qemu_clock_get_ns(QEMU_CLOCK_REALTIME)) proved to be very expensive. We had an added overhead of 13% of the total run time. Therefore, we modified the algorithm and only take into account the difference between the 2 clocks at the begining of the cpu_exec function. During the for loop we try to reduce the advance of the guest only by computing the virtual time elapsed and sleeping if necessary. The overhead is thus reduced to 3%. Even though this method still has a noticeable overhead, it no longer is a bottleneck in trying to achieve a better guest frequency for which the guest clock is faster than the host one. As for the the alignement of the 2 clocks, with the first algorithm the guest clock was oscillating between -1 and 1ms compared to the host clock. Using the second algorithm we notice that the guest is 5ms behind the host, which is still acceptable for our use case. The tests where conducted using fio and stress. The host machine in an i5 CPU at 3.10GHz running Debian Jessie (kernel 3.12). The guest machine is an arm versatile-pb built with buildroot. Currently, on our test machine, the lowest icount we can achieve that is suitable for aligning the 2 clocks is 6. However, we observe that the IO tests (using fio) are slower than the cpu tests (using stress). Signed-off-by: Sebastian Tanase <sebastian.tanase@openwide.fr> Tested-by: Camille Bégué <camille.begue@openwide.fr> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2014-07-25 11:56:31 +02:00
/* Try to align the host and virtual clocks
if the guest is in advance */
align_clocks(&sc, cpu);
/* reset soft MMU for next block (it can currently
only be set by a memory fault) */
} /* for(;;) */
} else {
/* Reload env after longjmp - the compiler may have smashed all
* local variables as longjmp is marked 'noreturn'. */
cpu = current_cpu;
env = cpu->env_ptr;
#if !(defined(CONFIG_USER_ONLY) && \
(defined(TARGET_M68K) || defined(TARGET_PPC) || defined(TARGET_S390X)))
cc = CPU_GET_CLASS(cpu);
#endif
#ifdef TARGET_I386
x86_cpu = X86_CPU(cpu);
#endif
if (have_tb_lock) {
spin_unlock(&tcg_ctx.tb_ctx.tb_lock);
have_tb_lock = false;
}
}
} /* for(;;) */
#if defined(TARGET_I386)
/* restore flags in standard format */
env->eflags = env->eflags | cpu_cc_compute_all(env, CC_OP)
| (env->df & DF_MASK);
#elif defined(TARGET_ARM)
/* XXX: Save/restore host fpu exception state?. */
#elif defined(TARGET_UNICORE32)
#elif defined(TARGET_SPARC)
#elif defined(TARGET_PPC)
#elif defined(TARGET_LM32)
#elif defined(TARGET_M68K)
cpu_m68k_flush_flags(env, env->cc_op);
env->cc_op = CC_OP_FLAGS;
env->sr = (env->sr & 0xffe0)
| env->cc_dest | (env->cc_x << 4);
#elif defined(TARGET_MICROBLAZE)
#elif defined(TARGET_MIPS)
#elif defined(TARGET_TRICORE)
#elif defined(TARGET_MOXIE)
#elif defined(TARGET_OPENRISC)
#elif defined(TARGET_SH4)
#elif defined(TARGET_ALPHA)
#elif defined(TARGET_CRIS)
#elif defined(TARGET_S390X)
#elif defined(TARGET_XTENSA)
/* XXXXX */
#else
#error unsupported target CPU
#endif
/* fail safe : never use current_cpu outside cpu_exec() */
current_cpu = NULL;
return ret;
}