diff --git a/arch/tile/include/asm/processor.h b/arch/tile/include/asm/processor.h index b3f104953da2..cda27243fb09 100644 --- a/arch/tile/include/asm/processor.h +++ b/arch/tile/include/asm/processor.h @@ -247,6 +247,13 @@ unsigned long get_wchan(struct task_struct *p); #define KSTK_EIP(task) task_pc(task) #define KSTK_ESP(task) task_sp(task) +/* Fine-grained unaligned JIT support */ +#define GET_UNALIGN_CTL(tsk, adr) get_unalign_ctl((tsk), (adr)) +#define SET_UNALIGN_CTL(tsk, val) set_unalign_ctl((tsk), (val)) + +extern int get_unalign_ctl(struct task_struct *tsk, unsigned long adr); +extern int set_unalign_ctl(struct task_struct *tsk, unsigned int val); + /* Standard format for printing registers and other word-size data. */ #ifdef __tilegx__ # define REGFMT "0x%016lx" diff --git a/arch/tile/include/asm/ptrace.h b/arch/tile/include/asm/ptrace.h index fd412260aff7..73b681b566f7 100644 --- a/arch/tile/include/asm/ptrace.h +++ b/arch/tile/include/asm/ptrace.h @@ -79,8 +79,7 @@ extern void single_step_execve(void); struct task_struct; -extern void send_sigtrap(struct task_struct *tsk, struct pt_regs *regs, - int error_code); +extern void send_sigtrap(struct task_struct *tsk, struct pt_regs *regs); #ifdef __tilegx__ /* We need this since sigval_t has a user pointer in it, for GETSIGINFO etc. */ diff --git a/arch/tile/include/asm/sections.h b/arch/tile/include/asm/sections.h index 7d8a935a9238..cc95276ef9c9 100644 --- a/arch/tile/include/asm/sections.h +++ b/arch/tile/include/asm/sections.h @@ -28,7 +28,9 @@ extern char __w1data_begin[], __w1data_end[]; /* Not exactly sections, but PC comparison points in the code. */ extern char __rt_sigreturn[], __rt_sigreturn_end[]; -#ifndef __tilegx__ +#ifdef __tilegx__ +extern char __start_unalign_asm_code[], __end_unalign_asm_code[]; +#else extern char sys_cmpxchg[], __sys_cmpxchg_end[]; extern char __sys_cmpxchg_grab_lock[]; extern char __start_atomic_asm_code[], __end_atomic_asm_code[]; diff --git a/arch/tile/include/asm/thread_info.h b/arch/tile/include/asm/thread_info.h index d1733dee98a2..b8aa6df3e102 100644 --- a/arch/tile/include/asm/thread_info.h +++ b/arch/tile/include/asm/thread_info.h @@ -39,6 +39,11 @@ struct thread_info { struct restart_block restart_block; struct single_step_state *step_state; /* single step state (if non-zero) */ + int align_ctl; /* controls unaligned access */ +#ifdef __tilegx__ + unsigned long unalign_jit_tmp[4]; /* temp r0..r3 storage */ + void __user *unalign_jit_base; /* unalign fixup JIT base */ +#endif }; /* @@ -56,6 +61,7 @@ struct thread_info { .fn = do_no_restart_syscall, \ }, \ .step_state = NULL, \ + .align_ctl = 0, \ } #define init_thread_info (init_thread_union.thread_info) diff --git a/arch/tile/include/asm/traps.h b/arch/tile/include/asm/traps.h index e28c3df4176a..5f172b2403a6 100644 --- a/arch/tile/include/asm/traps.h +++ b/arch/tile/include/asm/traps.h @@ -15,6 +15,7 @@ #ifndef _ASM_TILE_TRAPS_H #define _ASM_TILE_TRAPS_H +#ifndef __ASSEMBLY__ #include /* mm/fault.c */ @@ -69,6 +70,16 @@ void gx_singlestep_handle(struct pt_regs *, int fault_num); /* kernel/intvec_64.S */ void fill_ra_stack(void); + +/* Handle unalign data fixup. */ +extern void do_unaligned(struct pt_regs *regs, int vecnum); +#endif + +#endif /* __ASSEMBLY__ */ + +#ifdef __tilegx__ +/* 128 byte JIT per unalign fixup. */ +#define UNALIGN_JIT_SHIFT 7 #endif #endif /* _ASM_TILE_TRAPS_H */ diff --git a/arch/tile/kernel/Makefile b/arch/tile/kernel/Makefile index 5334be8e2538..6846c4ef5bf1 100644 --- a/arch/tile/kernel/Makefile +++ b/arch/tile/kernel/Makefile @@ -5,7 +5,8 @@ extra-y := vmlinux.lds head_$(BITS).o obj-y := backtrace.o entry.o irq.o messaging.o \ pci-dma.o proc.o process.o ptrace.o reboot.o \ - setup.o signal.o single_step.o stack.o sys.o sysfs.o time.o traps.o \ + setup.o signal.o single_step.o stack.o sys.o \ + sysfs.o time.o traps.o unaligned.o \ intvec_$(BITS).o regs_$(BITS).o tile-desc_$(BITS).o obj-$(CONFIG_HARDWALL) += hardwall.o diff --git a/arch/tile/kernel/asm-offsets.c b/arch/tile/kernel/asm-offsets.c index 8fff4757fffe..8652b0be4685 100644 --- a/arch/tile/kernel/asm-offsets.c +++ b/arch/tile/kernel/asm-offsets.c @@ -60,6 +60,12 @@ void foo(void) offsetof(struct thread_info, homecache_cpu)); DEFINE(THREAD_INFO_STEP_STATE_OFFSET, offsetof(struct thread_info, step_state)); +#ifdef __tilegx__ + DEFINE(THREAD_INFO_UNALIGN_JIT_BASE_OFFSET, + offsetof(struct thread_info, unalign_jit_base)); + DEFINE(THREAD_INFO_UNALIGN_JIT_TMP_OFFSET, + offsetof(struct thread_info, unalign_jit_tmp)); +#endif DEFINE(TASK_STRUCT_THREAD_KSP_OFFSET, offsetof(struct task_struct, thread.ksp)); diff --git a/arch/tile/kernel/intvec_32.S b/arch/tile/kernel/intvec_32.S index 25966af74a28..388061319c4c 100644 --- a/arch/tile/kernel/intvec_32.S +++ b/arch/tile/kernel/intvec_32.S @@ -1420,7 +1420,6 @@ handle_ill: { lw r0, r0 /* indirect thru thread_info to get task_info*/ addi r1, sp, C_ABI_SAVE_AREA_SIZE /* put ptregs pointer into r1 */ - move r2, zero /* load error code into r2 */ } jal send_sigtrap /* issue a SIGTRAP */ diff --git a/arch/tile/kernel/intvec_64.S b/arch/tile/kernel/intvec_64.S index 85d483957027..884af9ea5bed 100644 --- a/arch/tile/kernel/intvec_64.S +++ b/arch/tile/kernel/intvec_64.S @@ -17,11 +17,13 @@ #include #include #include +#include #include #include #include #include #include +#include #include #include #include @@ -98,6 +100,189 @@ } .endm + /* + * Unalign data exception fast handling: In order to handle + * unaligned data access, a fast JIT version is generated and stored + * in a specific area in user space. We first need to do a quick poke + * to see if the JIT is available. We use certain bits in the fault + * PC (3 to 9 is used for 16KB page size) as index to address the JIT + * code area. The first 64bit word is the fault PC, and the 2nd one is + * the fault bundle itself. If these 2 words both match, then we + * directly "iret" to JIT code. If not, a slow path is invoked to + * generate new JIT code. Note: the current JIT code WILL be + * overwritten if it existed. So, ideally we can handle 128 unalign + * fixups via JIT. For lookup efficiency and to effectively support + * tight loops with multiple unaligned reference, a simple + * direct-mapped cache is used. + * + * SPR_EX_CONTEXT_K_0 is modified to return to JIT code. + * SPR_EX_CONTEXT_K_1 has ICS set. + * SPR_EX_CONTEXT_0_0 is setup to user program's next PC. + * SPR_EX_CONTEXT_0_1 = 0. + */ + .macro int_hand_unalign_fast vecnum, vecname + .org (\vecnum << 8) +intvec_\vecname: + /* Put r3 in SPR_SYSTEM_SAVE_K_1. */ + mtspr SPR_SYSTEM_SAVE_K_1, r3 + + mfspr r3, SPR_EX_CONTEXT_K_1 + /* + * Examine if exception comes from user without ICS set. + * If not, just go directly to the slow path. + */ + bnez r3, hand_unalign_slow_nonuser + + mfspr r3, SPR_SYSTEM_SAVE_K_0 + + /* Get &thread_info->unalign_jit_tmp[0] in r3. */ + mm r3, zero, LOG2_THREAD_SIZE, 63 +#if THREAD_SIZE < 65536 + addli r3, r3, -(PAGE_SIZE - THREAD_INFO_UNALIGN_JIT_TMP_OFFSET) +#else + addli r3, r3, -(PAGE_SIZE/2) + addli r3, r3, -(PAGE_SIZE/2 - THREAD_INFO_UNALIGN_JIT_TMP_OFFSET) +#endif + + /* + * Save r0, r1, r2 into thread_info array r3 points to + * from low to high memory in order. + */ + st_add r3, r0, 8 + st_add r3, r1, 8 + { + st_add r3, r2, 8 + andi r2, sp, 7 + } + + /* Save stored r3 value so we can revert it on a page fault. */ + mfspr r1, SPR_SYSTEM_SAVE_K_1 + st r3, r1 + + { + /* Generate a SIGBUS if sp is not 8-byte aligned. */ + bnez r2, hand_unalign_slow_badsp + } + + /* + * Get the thread_info in r0; load r1 with pc. Set the low bit of sp + * as an indicator to the page fault code in case we fault. + */ + { + ori sp, sp, 1 + mfspr r1, SPR_EX_CONTEXT_K_0 + } + + /* Add the jit_info offset in thread_info; extract r1 [3:9] into r2. */ + { + addli r0, r3, THREAD_INFO_UNALIGN_JIT_BASE_OFFSET - \ + (THREAD_INFO_UNALIGN_JIT_TMP_OFFSET + (3 * 8)) + bfextu r2, r1, 3, (2 + PAGE_SHIFT - UNALIGN_JIT_SHIFT) + } + + /* Load the jit_info; multiply r2 by 128. */ + { + ld r0, r0 + shli r2, r2, UNALIGN_JIT_SHIFT + } + + /* + * If r0 is NULL, the JIT page is not mapped, so go to slow path; + * add offset r2 to r0 at the same time. + */ + { + beqz r0, hand_unalign_slow + add r2, r0, r2 + } + + /* + * We are loading from userspace (both the JIT info PC and + * instruction word, and the instruction word we executed) + * and since either could fault while holding the interrupt + * critical section, we must tag this region and check it in + * do_page_fault() to handle it properly. + */ +ENTRY(__start_unalign_asm_code) + + /* Load first word of JIT in r0 and increment r2 by 8. */ + ld_add r0, r2, 8 + + /* + * Compare the PC with the 1st word in JIT; load the fault bundle + * into r1. + */ + { + cmpeq r0, r0, r1 + ld r1, r1 + } + + /* Go to slow path if PC doesn't match. */ + beqz r0, hand_unalign_slow + + /* + * Load the 2nd word of JIT, which is supposed to be the fault + * bundle for a cache hit. Increment r2; after this bundle r2 will + * point to the potential start of the JIT code we want to run. + */ + ld_add r0, r2, 8 + + /* No further accesses to userspace are done after this point. */ +ENTRY(__end_unalign_asm_code) + + /* Compare the real bundle with what is saved in the JIT area. */ + { + cmpeq r0, r1, r0 + mtspr SPR_EX_CONTEXT_0_1, zero + } + + /* Go to slow path if the fault bundle does not match. */ + beqz r0, hand_unalign_slow + + /* + * A cache hit is found. + * r2 points to start of JIT code (3rd word). + * r0 is the fault pc. + * r1 is the fault bundle. + * Reset the low bit of sp. + */ + { + mfspr r0, SPR_EX_CONTEXT_K_0 + andi sp, sp, ~1 + } + + /* Write r2 into EX_CONTEXT_K_0 and increment PC. */ + { + mtspr SPR_EX_CONTEXT_K_0, r2 + addi r0, r0, 8 + } + + /* + * Set ICS on kernel EX_CONTEXT_K_1 in order to "iret" to + * user with ICS set. This way, if the JIT fixup causes another + * unalign exception (which shouldn't be possible) the user + * process will be terminated with SIGBUS. Also, our fixup will + * run without interleaving with external interrupts. + * Each fixup is at most 14 bundles, so it won't hold ICS for long. + */ + { + movei r1, PL_ICS_EX1(USER_PL, 1) + mtspr SPR_EX_CONTEXT_0_0, r0 + } + + { + mtspr SPR_EX_CONTEXT_K_1, r1 + addi r3, r3, -(3 * 8) + } + + /* Restore r0..r3. */ + ld_add r0, r3, 8 + ld_add r1, r3, 8 + ld_add r2, r3, 8 + ld r3, r3 + + iret + ENDPROC(intvec_\vecname) + .endm #ifdef __COLLECT_LINKER_FEEDBACK__ .pushsection .text.intvec_feedback,"ax" @@ -118,15 +303,21 @@ intvec_feedback: * The "processing" argument specifies the code for processing * the interrupt. Defaults to "handle_interrupt". */ - .macro int_hand vecnum, vecname, c_routine, processing=handle_interrupt - .org (\vecnum << 8) + .macro __int_hand vecnum, vecname, c_routine,processing=handle_interrupt intvec_\vecname: /* Temporarily save a register so we have somewhere to work. */ mtspr SPR_SYSTEM_SAVE_K_1, r0 mfspr r0, SPR_EX_CONTEXT_K_1 - andi r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */ + /* + * The unalign data fastpath code sets the low bit in sp to + * force us to reset it here on fault. + */ + { + blbs sp, 2f + andi r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */ + } .ifc \vecnum, INT_DOUBLE_FAULT /* @@ -176,7 +367,7 @@ intvec_\vecname: } .endif - +2: /* * SYSTEM_SAVE_K_0 holds the cpu number in the low bits, and * the current stack top in the higher bits. So we recover @@ -1223,10 +1414,31 @@ STD_ENTRY(_sys_clone) j sys_clone STD_ENDPROC(_sys_clone) -/* The single-step support may need to read all the registers. */ + /* + * Recover r3, r2, r1 and r0 here saved by unalign fast vector. + * The vector area limit is 32 bundles, so we handle the reload here. + * r0, r1, r2 are in thread_info from low to high memory in order. + * r3 points to location the original r3 was saved. + * We put this code in the __HEAD section so it can be reached + * via a conditional branch from the fast path. + */ + __HEAD +hand_unalign_slow: + andi sp, sp, ~1 +hand_unalign_slow_badsp: + addi r3, r3, -(3 * 8) + ld_add r0, r3, 8 + ld_add r1, r3, 8 + ld r2, r3 +hand_unalign_slow_nonuser: + mfspr r3, SPR_SYSTEM_SAVE_K_1 + __int_hand INT_UNALIGN_DATA, UNALIGN_DATA_SLOW, int_unalign + +/* The unaligned data support needs to read all the registers. */ int_unalign: push_extra_callee_saves r0 - j do_trap + j do_unaligned +ENDPROC(hand_unalign_slow) /* Fill the return address stack with nonzero entries. */ STD_ENTRY(fill_ra_stack) @@ -1240,6 +1452,11 @@ STD_ENTRY(fill_ra_stack) 4: jrp r0 STD_ENDPROC(fill_ra_stack) + .macro int_hand vecnum, vecname, c_routine, processing=handle_interrupt + .org (\vecnum << 8) + __int_hand \vecnum, \vecname, \c_routine, \processing + .endm + /* Include .intrpt1 array of interrupt vectors */ .section ".intrpt1", "ax" @@ -1272,7 +1489,7 @@ STD_ENTRY(fill_ra_stack) int_hand INT_SWINT_1, SWINT_1, SYSCALL, handle_syscall int_hand INT_SWINT_0, SWINT_0, do_trap int_hand INT_ILL_TRANS, ILL_TRANS, do_trap - int_hand INT_UNALIGN_DATA, UNALIGN_DATA, int_unalign + int_hand_unalign_fast INT_UNALIGN_DATA, UNALIGN_DATA int_hand INT_DTLB_MISS, DTLB_MISS, do_page_fault int_hand INT_DTLB_ACCESS, DTLB_ACCESS, do_page_fault int_hand INT_IDN_FIREWALL, IDN_FIREWALL, do_hardwall_trap diff --git a/arch/tile/kernel/proc.c b/arch/tile/kernel/proc.c index dafc447b5125..681100c59fda 100644 --- a/arch/tile/kernel/proc.c +++ b/arch/tile/kernel/proc.c @@ -113,7 +113,6 @@ arch_initcall(proc_tile_init); * Support /proc/sys/tile directory */ -#ifndef __tilegx__ /* FIXME: GX: no support for unaligned access yet */ static ctl_table unaligned_subtable[] = { { .procname = "enabled", @@ -160,4 +159,3 @@ static int __init proc_sys_tile_init(void) } arch_initcall(proc_sys_tile_init); -#endif diff --git a/arch/tile/kernel/process.c b/arch/tile/kernel/process.c index 8d6c51d55762..25678b83b747 100644 --- a/arch/tile/kernel/process.c +++ b/arch/tile/kernel/process.c @@ -33,6 +33,7 @@ #include #include #include +#include #ifdef CONFIG_HARDWALL #include #endif @@ -147,6 +148,14 @@ int copy_thread(unsigned long clone_flags, unsigned long sp, */ task_thread_info(p)->step_state = NULL; +#ifdef __tilegx__ + /* + * Do not clone unalign jit fixup from the parent; each thread + * must allocate its own on demand. + */ + task_thread_info(p)->unalign_jit_base = NULL; +#endif + /* * Copy the registers onto the kernel stack so the * return-from-interrupt code will reload it into registers. @@ -205,6 +214,18 @@ int copy_thread(unsigned long clone_flags, unsigned long sp, return 0; } +int set_unalign_ctl(struct task_struct *tsk, unsigned int val) +{ + task_thread_info(tsk)->align_ctl = val; + return 0; +} + +int get_unalign_ctl(struct task_struct *tsk, unsigned long adr) +{ + return put_user(task_thread_info(tsk)->align_ctl, + (unsigned int __user *)adr); +} + /* * Return "current" if it looks plausible, or else a pointer to a dummy. * This can be helpful if we are just trying to emit a clean panic. diff --git a/arch/tile/kernel/ptrace.c b/arch/tile/kernel/ptrace.c index 0f83ed4602b2..bac187498d61 100644 --- a/arch/tile/kernel/ptrace.c +++ b/arch/tile/kernel/ptrace.c @@ -272,7 +272,7 @@ void do_syscall_trace_exit(struct pt_regs *regs) trace_sys_exit(regs, regs->regs[0]); } -void send_sigtrap(struct task_struct *tsk, struct pt_regs *regs, int error_code) +void send_sigtrap(struct task_struct *tsk, struct pt_regs *regs) { struct siginfo info; @@ -288,5 +288,5 @@ void send_sigtrap(struct task_struct *tsk, struct pt_regs *regs, int error_code) /* Handle synthetic interrupt delivered only by the simulator. */ void __kprobes do_breakpoint(struct pt_regs* regs, int fault_num) { - send_sigtrap(current, regs, fault_num); + send_sigtrap(current, regs); } diff --git a/arch/tile/kernel/single_step.c b/arch/tile/kernel/single_step.c index 27742e87e255..5ef2e9eae5c5 100644 --- a/arch/tile/kernel/single_step.c +++ b/arch/tile/kernel/single_step.c @@ -12,42 +12,31 @@ * more details. * * A code-rewriter that enables instruction single-stepping. - * Derived from iLib's single-stepping code. */ -#ifndef __tilegx__ /* Hardware support for single step unavailable. */ - -/* These functions are only used on the TILE platform */ +#include +#include #include #include #include #include #include #include +#include #include +#include +#include #include #include +#include #include + +#ifndef __tilegx__ /* Hardware support for single step unavailable. */ + #define signExtend17(val) sign_extend((val), 17) #define TILE_X1_MASK (0xffffffffULL << 31) -int unaligned_printk; - -static int __init setup_unaligned_printk(char *str) -{ - long val; - if (strict_strtol(str, 0, &val) != 0) - return 0; - unaligned_printk = val; - pr_info("Printk for each unaligned data accesses is %s\n", - unaligned_printk ? "enabled" : "disabled"); - return 1; -} -__setup("unaligned_printk=", setup_unaligned_printk); - -unsigned int unaligned_fixup_count; - enum mem_op { MEMOP_NONE, MEMOP_LOAD, @@ -56,12 +45,13 @@ enum mem_op { MEMOP_STORE_POSTINCR }; -static inline tile_bundle_bits set_BrOff_X1(tile_bundle_bits n, s32 offset) +static inline tilepro_bundle_bits set_BrOff_X1(tilepro_bundle_bits n, + s32 offset) { - tile_bundle_bits result; + tilepro_bundle_bits result; /* mask out the old offset */ - tile_bundle_bits mask = create_BrOff_X1(-1); + tilepro_bundle_bits mask = create_BrOff_X1(-1); result = n & (~mask); /* or in the new offset */ @@ -70,10 +60,11 @@ static inline tile_bundle_bits set_BrOff_X1(tile_bundle_bits n, s32 offset) return result; } -static inline tile_bundle_bits move_X1(tile_bundle_bits n, int dest, int src) +static inline tilepro_bundle_bits move_X1(tilepro_bundle_bits n, int dest, + int src) { - tile_bundle_bits result; - tile_bundle_bits op; + tilepro_bundle_bits result; + tilepro_bundle_bits op; result = n & (~TILE_X1_MASK); @@ -87,13 +78,13 @@ static inline tile_bundle_bits move_X1(tile_bundle_bits n, int dest, int src) return result; } -static inline tile_bundle_bits nop_X1(tile_bundle_bits n) +static inline tilepro_bundle_bits nop_X1(tilepro_bundle_bits n) { return move_X1(n, TREG_ZERO, TREG_ZERO); } -static inline tile_bundle_bits addi_X1( - tile_bundle_bits n, int dest, int src, int imm) +static inline tilepro_bundle_bits addi_X1( + tilepro_bundle_bits n, int dest, int src, int imm) { n &= ~TILE_X1_MASK; @@ -107,15 +98,26 @@ static inline tile_bundle_bits addi_X1( return n; } -static tile_bundle_bits rewrite_load_store_unaligned( +static tilepro_bundle_bits rewrite_load_store_unaligned( struct single_step_state *state, - tile_bundle_bits bundle, + tilepro_bundle_bits bundle, struct pt_regs *regs, enum mem_op mem_op, int size, int sign_ext) { unsigned char __user *addr; int val_reg, addr_reg, err, val; + int align_ctl; + + align_ctl = unaligned_fixup; + switch (task_thread_info(current)->align_ctl) { + case PR_UNALIGN_NOPRINT: + align_ctl = 1; + break; + case PR_UNALIGN_SIGBUS: + align_ctl = 0; + break; + } /* Get address and value registers */ if (bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK) { @@ -160,7 +162,7 @@ static tile_bundle_bits rewrite_load_store_unaligned( * tilepro hardware would be doing, if it could provide us with the * actual bad address in an SPR, which it doesn't. */ - if (unaligned_fixup == 0) { + if (align_ctl == 0) { siginfo_t info = { .si_signo = SIGBUS, .si_code = BUS_ADRALN, @@ -209,14 +211,14 @@ static tile_bundle_bits rewrite_load_store_unaligned( if (err) { siginfo_t info = { - .si_signo = SIGSEGV, - .si_code = SEGV_MAPERR, + .si_signo = SIGBUS, + .si_code = BUS_ADRALN, .si_addr = addr }; - trace_unhandled_signal("segfault", regs, - (unsigned long)addr, SIGSEGV); + trace_unhandled_signal("bad address for unaligned fixup", regs, + (unsigned long)addr, SIGBUS); force_sig_info(info.si_signo, &info, current); - return (tile_bundle_bits) 0; + return (tilepro_bundle_bits) 0; } if (unaligned_printk || unaligned_fixup_count == 0) { @@ -285,7 +287,7 @@ void single_step_execve(void) ti->step_state = NULL; } -/** +/* * single_step_once() - entry point when single stepping has been triggered. * @regs: The machine register state * @@ -304,20 +306,31 @@ void single_step_execve(void) */ void single_step_once(struct pt_regs *regs) { - extern tile_bundle_bits __single_step_ill_insn; - extern tile_bundle_bits __single_step_j_insn; - extern tile_bundle_bits __single_step_addli_insn; - extern tile_bundle_bits __single_step_auli_insn; + extern tilepro_bundle_bits __single_step_ill_insn; + extern tilepro_bundle_bits __single_step_j_insn; + extern tilepro_bundle_bits __single_step_addli_insn; + extern tilepro_bundle_bits __single_step_auli_insn; struct thread_info *info = (void *)current_thread_info(); struct single_step_state *state = info->step_state; int is_single_step = test_ti_thread_flag(info, TIF_SINGLESTEP); - tile_bundle_bits __user *buffer, *pc; - tile_bundle_bits bundle; + tilepro_bundle_bits __user *buffer, *pc; + tilepro_bundle_bits bundle; int temp_reg; int target_reg = TREG_LR; int err; enum mem_op mem_op = MEMOP_NONE; int size = 0, sign_ext = 0; /* happy compiler */ + int align_ctl; + + align_ctl = unaligned_fixup; + switch (task_thread_info(current)->align_ctl) { + case PR_UNALIGN_NOPRINT: + align_ctl = 1; + break; + case PR_UNALIGN_SIGBUS: + align_ctl = 0; + break; + } asm( " .pushsection .rodata.single_step\n" @@ -390,7 +403,7 @@ void single_step_once(struct pt_regs *regs) if (regs->faultnum == INT_SWINT_1) regs->pc -= 8; - pc = (tile_bundle_bits __user *)(regs->pc); + pc = (tilepro_bundle_bits __user *)(regs->pc); if (get_user(bundle, pc) != 0) { pr_err("Couldn't read instruction at %p trying to step\n", pc); return; @@ -627,9 +640,9 @@ void single_step_once(struct pt_regs *regs) /* * Check if we need to rewrite an unaligned load/store. - * Returning zero is a special value meaning we need to SIGSEGV. + * Returning zero is a special value meaning we generated a signal. */ - if (mem_op != MEMOP_NONE && unaligned_fixup >= 0) { + if (mem_op != MEMOP_NONE && align_ctl >= 0) { bundle = rewrite_load_store_unaligned(state, bundle, regs, mem_op, size, sign_ext); if (bundle == 0) @@ -668,9 +681,9 @@ void single_step_once(struct pt_regs *regs) } /* End with a jump back to the next instruction */ - delta = ((regs->pc + TILE_BUNDLE_SIZE_IN_BYTES) - + delta = ((regs->pc + TILEPRO_BUNDLE_SIZE_IN_BYTES) - (unsigned long)buffer) >> - TILE_LOG2_BUNDLE_ALIGNMENT_IN_BYTES; + TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES; bundle = __single_step_j_insn; bundle |= create_JOffLong_X1(delta); err |= __put_user(bundle, buffer++); @@ -698,9 +711,6 @@ void single_step_once(struct pt_regs *regs) } #else -#include -#include -#include static DEFINE_PER_CPU(unsigned long, ss_saved_pc); @@ -743,10 +753,10 @@ void gx_singlestep_handle(struct pt_regs *regs, int fault_num) } else if ((*ss_pc != regs->pc) || (!(control & SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK))) { - ptrace_notify(SIGTRAP); control |= SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK; control |= SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK; __insn_mtspr(SPR_SINGLE_STEP_CONTROL_K, control); + send_sigtrap(current, regs); } } diff --git a/arch/tile/kernel/unaligned.c b/arch/tile/kernel/unaligned.c new file mode 100644 index 000000000000..b425fb6a480d --- /dev/null +++ b/arch/tile/kernel/unaligned.c @@ -0,0 +1,1609 @@ +/* + * Copyright 2013 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program 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, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + * + * A code-rewriter that handles unaligned exception. + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + + +/* + * This file handles unaligned exception for tile-Gx. The tilepro's unaligned + * exception is supported out of single_step.c + */ + +int unaligned_printk; + +static int __init setup_unaligned_printk(char *str) +{ + long val; + if (kstrtol(str, 0, &val) != 0) + return 0; + unaligned_printk = val; + pr_info("Printk for each unaligned data accesses is %s\n", + unaligned_printk ? "enabled" : "disabled"); + return 1; +} +__setup("unaligned_printk=", setup_unaligned_printk); + +unsigned int unaligned_fixup_count; + +#ifdef __tilegx__ + +/* + * Unalign data jit fixup code fragement. Reserved space is 128 bytes. + * The 1st 64-bit word saves fault PC address, 2nd word is the fault + * instruction bundle followed by 14 JIT bundles. + */ + +struct unaligned_jit_fragment { + unsigned long pc; + tilegx_bundle_bits bundle; + tilegx_bundle_bits insn[14]; +}; + +/* + * Check if a nop or fnop at bundle's pipeline X0. + */ + +static bool is_bundle_x0_nop(tilegx_bundle_bits bundle) +{ + return (((get_UnaryOpcodeExtension_X0(bundle) == + NOP_UNARY_OPCODE_X0) && + (get_RRROpcodeExtension_X0(bundle) == + UNARY_RRR_0_OPCODE_X0) && + (get_Opcode_X0(bundle) == + RRR_0_OPCODE_X0)) || + ((get_UnaryOpcodeExtension_X0(bundle) == + FNOP_UNARY_OPCODE_X0) && + (get_RRROpcodeExtension_X0(bundle) == + UNARY_RRR_0_OPCODE_X0) && + (get_Opcode_X0(bundle) == + RRR_0_OPCODE_X0))); +} + +/* + * Check if nop or fnop at bundle's pipeline X1. + */ + +static bool is_bundle_x1_nop(tilegx_bundle_bits bundle) +{ + return (((get_UnaryOpcodeExtension_X1(bundle) == + NOP_UNARY_OPCODE_X1) && + (get_RRROpcodeExtension_X1(bundle) == + UNARY_RRR_0_OPCODE_X1) && + (get_Opcode_X1(bundle) == + RRR_0_OPCODE_X1)) || + ((get_UnaryOpcodeExtension_X1(bundle) == + FNOP_UNARY_OPCODE_X1) && + (get_RRROpcodeExtension_X1(bundle) == + UNARY_RRR_0_OPCODE_X1) && + (get_Opcode_X1(bundle) == + RRR_0_OPCODE_X1))); +} + +/* + * Check if nop or fnop at bundle's Y0 pipeline. + */ + +static bool is_bundle_y0_nop(tilegx_bundle_bits bundle) +{ + return (((get_UnaryOpcodeExtension_Y0(bundle) == + NOP_UNARY_OPCODE_Y0) && + (get_RRROpcodeExtension_Y0(bundle) == + UNARY_RRR_1_OPCODE_Y0) && + (get_Opcode_Y0(bundle) == + RRR_1_OPCODE_Y0)) || + ((get_UnaryOpcodeExtension_Y0(bundle) == + FNOP_UNARY_OPCODE_Y0) && + (get_RRROpcodeExtension_Y0(bundle) == + UNARY_RRR_1_OPCODE_Y0) && + (get_Opcode_Y0(bundle) == + RRR_1_OPCODE_Y0))); +} + +/* + * Check if nop or fnop at bundle's pipeline Y1. + */ + +static bool is_bundle_y1_nop(tilegx_bundle_bits bundle) +{ + return (((get_UnaryOpcodeExtension_Y1(bundle) == + NOP_UNARY_OPCODE_Y1) && + (get_RRROpcodeExtension_Y1(bundle) == + UNARY_RRR_1_OPCODE_Y1) && + (get_Opcode_Y1(bundle) == + RRR_1_OPCODE_Y1)) || + ((get_UnaryOpcodeExtension_Y1(bundle) == + FNOP_UNARY_OPCODE_Y1) && + (get_RRROpcodeExtension_Y1(bundle) == + UNARY_RRR_1_OPCODE_Y1) && + (get_Opcode_Y1(bundle) == + RRR_1_OPCODE_Y1))); +} + +/* + * Test if a bundle's y0 and y1 pipelines are both nop or fnop. + */ + +static bool is_y0_y1_nop(tilegx_bundle_bits bundle) +{ + return is_bundle_y0_nop(bundle) && is_bundle_y1_nop(bundle); +} + +/* + * Test if a bundle's x0 and x1 pipelines are both nop or fnop. + */ + +static bool is_x0_x1_nop(tilegx_bundle_bits bundle) +{ + return is_bundle_x0_nop(bundle) && is_bundle_x1_nop(bundle); +} + +/* + * Find the destination, source registers of fault unalign access instruction + * at X1 or Y2. Also, allocate up to 3 scratch registers clob1, clob2 and + * clob3, which are guaranteed different from any register used in the fault + * bundle. r_alias is used to return if the other instructions other than the + * unalign load/store shares same register with ra, rb and rd. + */ + +static void find_regs(tilegx_bundle_bits bundle, uint64_t *rd, uint64_t *ra, + uint64_t *rb, uint64_t *clob1, uint64_t *clob2, + uint64_t *clob3, bool *r_alias) +{ + int i; + uint64_t reg; + uint64_t reg_map = 0, alias_reg_map = 0, map; + bool alias; + + *ra = -1; + *rb = -1; + + if (rd) + *rd = -1; + + *clob1 = -1; + *clob2 = -1; + *clob3 = -1; + alias = false; + + /* + * Parse fault bundle, find potential used registers and mark + * corresponding bits in reg_map and alias_map. These 2 bit maps + * are used to find the scratch registers and determine if there + * is register alais. + */ + if (bundle & TILEGX_BUNDLE_MODE_MASK) { /* Y Mode Bundle. */ + + reg = get_SrcA_Y2(bundle); + reg_map |= 1ULL << reg; + *ra = reg; + reg = get_SrcBDest_Y2(bundle); + reg_map |= 1ULL << reg; + + if (rd) { + /* Load. */ + *rd = reg; + alias_reg_map = (1ULL << *rd) | (1ULL << *ra); + } else { + /* Store. */ + *rb = reg; + alias_reg_map = (1ULL << *ra) | (1ULL << *rb); + } + + if (!is_bundle_y1_nop(bundle)) { + reg = get_SrcA_Y1(bundle); + reg_map |= (1ULL << reg); + map = (1ULL << reg); + + reg = get_SrcB_Y1(bundle); + reg_map |= (1ULL << reg); + map |= (1ULL << reg); + + reg = get_Dest_Y1(bundle); + reg_map |= (1ULL << reg); + map |= (1ULL << reg); + + if (map & alias_reg_map) + alias = true; + } + + if (!is_bundle_y0_nop(bundle)) { + reg = get_SrcA_Y0(bundle); + reg_map |= (1ULL << reg); + map = (1ULL << reg); + + reg = get_SrcB_Y0(bundle); + reg_map |= (1ULL << reg); + map |= (1ULL << reg); + + reg = get_Dest_Y0(bundle); + reg_map |= (1ULL << reg); + map |= (1ULL << reg); + + if (map & alias_reg_map) + alias = true; + } + } else { /* X Mode Bundle. */ + + reg = get_SrcA_X1(bundle); + reg_map |= (1ULL << reg); + *ra = reg; + if (rd) { + /* Load. */ + reg = get_Dest_X1(bundle); + reg_map |= (1ULL << reg); + *rd = reg; + alias_reg_map = (1ULL << *rd) | (1ULL << *ra); + } else { + /* Store. */ + reg = get_SrcB_X1(bundle); + reg_map |= (1ULL << reg); + *rb = reg; + alias_reg_map = (1ULL << *ra) | (1ULL << *rb); + } + + if (!is_bundle_x0_nop(bundle)) { + reg = get_SrcA_X0(bundle); + reg_map |= (1ULL << reg); + map = (1ULL << reg); + + reg = get_SrcB_X0(bundle); + reg_map |= (1ULL << reg); + map |= (1ULL << reg); + + reg = get_Dest_X0(bundle); + reg_map |= (1ULL << reg); + map |= (1ULL << reg); + + if (map & alias_reg_map) + alias = true; + } + } + + /* + * "alias" indicates if the unalign access registers have collision + * with others in the same bundle. We jsut simply test all register + * operands case (RRR), ignored the case with immidate. If a bundle + * has no register alias, we may do fixup in a simple or fast manner. + * So if an immidata field happens to hit with a register, we may end + * up fall back to the generic handling. + */ + + *r_alias = alias; + + /* Flip bits on reg_map. */ + reg_map ^= -1ULL; + + /* Scan reg_map lower 54(TREG_SP) bits to find 3 set bits. */ + for (i = 0; i < TREG_SP; i++) { + if (reg_map & (0x1ULL << i)) { + if (*clob1 == -1) { + *clob1 = i; + } else if (*clob2 == -1) { + *clob2 = i; + } else if (*clob3 == -1) { + *clob3 = i; + return; + } + } + } +} + +/* + * Sanity check for register ra, rb, rd, clob1/2/3. Return true if any of them + * is unexpected. + */ + +static bool check_regs(uint64_t rd, uint64_t ra, uint64_t rb, + uint64_t clob1, uint64_t clob2, uint64_t clob3) +{ + bool unexpected = false; + if ((ra >= 56) && (ra != TREG_ZERO)) + unexpected = true; + + if ((clob1 >= 56) || (clob2 >= 56) || (clob3 >= 56)) + unexpected = true; + + if (rd != -1) { + if ((rd >= 56) && (rd != TREG_ZERO)) + unexpected = true; + } else { + if ((rb >= 56) && (rb != TREG_ZERO)) + unexpected = true; + } + return unexpected; +} + + +#define GX_INSN_X0_MASK ((1ULL << 31) - 1) +#define GX_INSN_X1_MASK (((1ULL << 31) - 1) << 31) +#define GX_INSN_Y0_MASK ((0xFULL << 27) | (0xFFFFFULL)) +#define GX_INSN_Y1_MASK (GX_INSN_Y0_MASK << 31) +#define GX_INSN_Y2_MASK ((0x7FULL << 51) | (0x7FULL << 20)) + +#ifdef __LITTLE_ENDIAN +#define GX_INSN_BSWAP(_bundle_) (_bundle_) +#else +#define GX_INSN_BSWAP(_bundle_) swab64(_bundle_) +#endif /* __LITTLE_ENDIAN */ + +/* + * __JIT_CODE(.) creates template bundles in .rodata.unalign_data section. + * The corresponding static function jix_x#_###(.) generates partial or + * whole bundle based on the template and given arguments. + */ + +#define __JIT_CODE(_X_) \ + asm (".pushsection .rodata.unalign_data, \"a\"\n" \ + _X_"\n" \ + ".popsection\n") + +__JIT_CODE("__unalign_jit_x1_mtspr: {mtspr 0, r0}"); +static tilegx_bundle_bits jit_x1_mtspr(int spr, int reg) +{ + extern tilegx_bundle_bits __unalign_jit_x1_mtspr; + return (GX_INSN_BSWAP(__unalign_jit_x1_mtspr) & GX_INSN_X1_MASK) | + create_MT_Imm14_X1(spr) | create_SrcA_X1(reg); +} + +__JIT_CODE("__unalign_jit_x1_mfspr: {mfspr r0, 0}"); +static tilegx_bundle_bits jit_x1_mfspr(int reg, int spr) +{ + extern tilegx_bundle_bits __unalign_jit_x1_mfspr; + return (GX_INSN_BSWAP(__unalign_jit_x1_mfspr) & GX_INSN_X1_MASK) | + create_MF_Imm14_X1(spr) | create_Dest_X1(reg); +} + +__JIT_CODE("__unalign_jit_x0_addi: {addi r0, r0, 0; iret}"); +static tilegx_bundle_bits jit_x0_addi(int rd, int ra, int imm8) +{ + extern tilegx_bundle_bits __unalign_jit_x0_addi; + return (GX_INSN_BSWAP(__unalign_jit_x0_addi) & GX_INSN_X0_MASK) | + create_Dest_X0(rd) | create_SrcA_X0(ra) | + create_Imm8_X0(imm8); +} + +__JIT_CODE("__unalign_jit_x1_ldna: {ldna r0, r0}"); +static tilegx_bundle_bits jit_x1_ldna(int rd, int ra) +{ + extern tilegx_bundle_bits __unalign_jit_x1_ldna; + return (GX_INSN_BSWAP(__unalign_jit_x1_ldna) & GX_INSN_X1_MASK) | + create_Dest_X1(rd) | create_SrcA_X1(ra); +} + +__JIT_CODE("__unalign_jit_x0_dblalign: {dblalign r0, r0 ,r0}"); +static tilegx_bundle_bits jit_x0_dblalign(int rd, int ra, int rb) +{ + extern tilegx_bundle_bits __unalign_jit_x0_dblalign; + return (GX_INSN_BSWAP(__unalign_jit_x0_dblalign) & GX_INSN_X0_MASK) | + create_Dest_X0(rd) | create_SrcA_X0(ra) | + create_SrcB_X0(rb); +} + +__JIT_CODE("__unalign_jit_x1_iret: {iret}"); +static tilegx_bundle_bits jit_x1_iret(void) +{ + extern tilegx_bundle_bits __unalign_jit_x1_iret; + return GX_INSN_BSWAP(__unalign_jit_x1_iret) & GX_INSN_X1_MASK; +} + +__JIT_CODE("__unalign_jit_x01_fnop: {fnop;fnop}"); +static tilegx_bundle_bits jit_x0_fnop(void) +{ + extern tilegx_bundle_bits __unalign_jit_x01_fnop; + return GX_INSN_BSWAP(__unalign_jit_x01_fnop) & GX_INSN_X0_MASK; +} + +static tilegx_bundle_bits jit_x1_fnop(void) +{ + extern tilegx_bundle_bits __unalign_jit_x01_fnop; + return GX_INSN_BSWAP(__unalign_jit_x01_fnop) & GX_INSN_X1_MASK; +} + +__JIT_CODE("__unalign_jit_y2_dummy: {fnop; fnop; ld zero, sp}"); +static tilegx_bundle_bits jit_y2_dummy(void) +{ + extern tilegx_bundle_bits __unalign_jit_y2_dummy; + return GX_INSN_BSWAP(__unalign_jit_y2_dummy) & GX_INSN_Y2_MASK; +} + +static tilegx_bundle_bits jit_y1_fnop(void) +{ + extern tilegx_bundle_bits __unalign_jit_y2_dummy; + return GX_INSN_BSWAP(__unalign_jit_y2_dummy) & GX_INSN_Y1_MASK; +} + +__JIT_CODE("__unalign_jit_x1_st1_add: {st1_add r1, r0, 0}"); +static tilegx_bundle_bits jit_x1_st1_add(int ra, int rb, int imm8) +{ + extern tilegx_bundle_bits __unalign_jit_x1_st1_add; + return (GX_INSN_BSWAP(__unalign_jit_x1_st1_add) & + (~create_SrcA_X1(-1)) & + GX_INSN_X1_MASK) | create_SrcA_X1(ra) | + create_SrcB_X1(rb) | create_Dest_Imm8_X1(imm8); +} + +__JIT_CODE("__unalign_jit_x1_st: {crc32_8 r1, r0, r0; st r0, r0}"); +static tilegx_bundle_bits jit_x1_st(int ra, int rb) +{ + extern tilegx_bundle_bits __unalign_jit_x1_st; + return (GX_INSN_BSWAP(__unalign_jit_x1_st) & GX_INSN_X1_MASK) | + create_SrcA_X1(ra) | create_SrcB_X1(rb); +} + +__JIT_CODE("__unalign_jit_x1_st_add: {st_add r1, r0, 0}"); +static tilegx_bundle_bits jit_x1_st_add(int ra, int rb, int imm8) +{ + extern tilegx_bundle_bits __unalign_jit_x1_st_add; + return (GX_INSN_BSWAP(__unalign_jit_x1_st_add) & + (~create_SrcA_X1(-1)) & + GX_INSN_X1_MASK) | create_SrcA_X1(ra) | + create_SrcB_X1(rb) | create_Dest_Imm8_X1(imm8); +} + +__JIT_CODE("__unalign_jit_x1_ld: {crc32_8 r1, r0, r0; ld r0, r0}"); +static tilegx_bundle_bits jit_x1_ld(int rd, int ra) +{ + extern tilegx_bundle_bits __unalign_jit_x1_ld; + return (GX_INSN_BSWAP(__unalign_jit_x1_ld) & GX_INSN_X1_MASK) | + create_Dest_X1(rd) | create_SrcA_X1(ra); +} + +__JIT_CODE("__unalign_jit_x1_ld_add: {ld_add r1, r0, 0}"); +static tilegx_bundle_bits jit_x1_ld_add(int rd, int ra, int imm8) +{ + extern tilegx_bundle_bits __unalign_jit_x1_ld_add; + return (GX_INSN_BSWAP(__unalign_jit_x1_ld_add) & + (~create_Dest_X1(-1)) & + GX_INSN_X1_MASK) | create_Dest_X1(rd) | + create_SrcA_X1(ra) | create_Imm8_X1(imm8); +} + +__JIT_CODE("__unalign_jit_x0_bfexts: {bfexts r0, r0, 0, 0}"); +static tilegx_bundle_bits jit_x0_bfexts(int rd, int ra, int bfs, int bfe) +{ + extern tilegx_bundle_bits __unalign_jit_x0_bfexts; + return (GX_INSN_BSWAP(__unalign_jit_x0_bfexts) & + GX_INSN_X0_MASK) | + create_Dest_X0(rd) | create_SrcA_X0(ra) | + create_BFStart_X0(bfs) | create_BFEnd_X0(bfe); +} + +__JIT_CODE("__unalign_jit_x0_bfextu: {bfextu r0, r0, 0, 0}"); +static tilegx_bundle_bits jit_x0_bfextu(int rd, int ra, int bfs, int bfe) +{ + extern tilegx_bundle_bits __unalign_jit_x0_bfextu; + return (GX_INSN_BSWAP(__unalign_jit_x0_bfextu) & + GX_INSN_X0_MASK) | + create_Dest_X0(rd) | create_SrcA_X0(ra) | + create_BFStart_X0(bfs) | create_BFEnd_X0(bfe); +} + +__JIT_CODE("__unalign_jit_x1_addi: {bfextu r1, r1, 0, 0; addi r0, r0, 0}"); +static tilegx_bundle_bits jit_x1_addi(int rd, int ra, int imm8) +{ + extern tilegx_bundle_bits __unalign_jit_x1_addi; + return (GX_INSN_BSWAP(__unalign_jit_x1_addi) & GX_INSN_X1_MASK) | + create_Dest_X1(rd) | create_SrcA_X1(ra) | + create_Imm8_X1(imm8); +} + +__JIT_CODE("__unalign_jit_x0_shrui: {shrui r0, r0, 0; iret}"); +static tilegx_bundle_bits jit_x0_shrui(int rd, int ra, int imm6) +{ + extern tilegx_bundle_bits __unalign_jit_x0_shrui; + return (GX_INSN_BSWAP(__unalign_jit_x0_shrui) & + GX_INSN_X0_MASK) | + create_Dest_X0(rd) | create_SrcA_X0(ra) | + create_ShAmt_X0(imm6); +} + +__JIT_CODE("__unalign_jit_x0_rotli: {rotli r0, r0, 0; iret}"); +static tilegx_bundle_bits jit_x0_rotli(int rd, int ra, int imm6) +{ + extern tilegx_bundle_bits __unalign_jit_x0_rotli; + return (GX_INSN_BSWAP(__unalign_jit_x0_rotli) & + GX_INSN_X0_MASK) | + create_Dest_X0(rd) | create_SrcA_X0(ra) | + create_ShAmt_X0(imm6); +} + +__JIT_CODE("__unalign_jit_x1_bnezt: {bnezt r0, __unalign_jit_x1_bnezt}"); +static tilegx_bundle_bits jit_x1_bnezt(int ra, int broff) +{ + extern tilegx_bundle_bits __unalign_jit_x1_bnezt; + return (GX_INSN_BSWAP(__unalign_jit_x1_bnezt) & + GX_INSN_X1_MASK) | + create_SrcA_X1(ra) | create_BrOff_X1(broff); +} + +#undef __JIT_CODE + +/* + * This function generates unalign fixup JIT. + * + * We fist find unalign load/store instruction's destination, source + * reguisters: ra, rb and rd. and 3 scratch registers by calling + * find_regs(...). 3 scratch clobbers should not alias with any register + * used in the fault bundle. Then analyze the fault bundle to determine + * if it's a load or store, operand width, branch or address increment etc. + * At last generated JIT is copied into JIT code area in user space. + */ + +static +void jit_bundle_gen(struct pt_regs *regs, tilegx_bundle_bits bundle, + int align_ctl) +{ + struct thread_info *info = current_thread_info(); + struct unaligned_jit_fragment frag; + struct unaligned_jit_fragment *jit_code_area; + tilegx_bundle_bits bundle_2 = 0; + /* If bundle_2_enable = false, bundle_2 is fnop/nop operation. */ + bool bundle_2_enable = true; + uint64_t ra, rb, rd = -1, clob1, clob2, clob3; + /* + * Indicate if the unalign access + * instruction's registers hit with + * others in the same bundle. + */ + bool alias = false; + bool load_n_store = true; + bool load_store_signed = false; + unsigned int load_store_size = 8; + bool y1_br = false; /* True, for a branch in same bundle at Y1.*/ + int y1_br_reg = 0; + /* True for link operation. i.e. jalr or lnk at Y1 */ + bool y1_lr = false; + int y1_lr_reg = 0; + bool x1_add = false;/* True, for load/store ADD instruction at X1*/ + int x1_add_imm8 = 0; + bool unexpected = false; + int n = 0, k; + + jit_code_area = + (struct unaligned_jit_fragment *)(info->unalign_jit_base); + + memset((void *)&frag, 0, sizeof(frag)); + + /* 0: X mode, Otherwise: Y mode. */ + if (bundle & TILEGX_BUNDLE_MODE_MASK) { + unsigned int mod, opcode; + + if (get_Opcode_Y1(bundle) == RRR_1_OPCODE_Y1 && + get_RRROpcodeExtension_Y1(bundle) == + UNARY_RRR_1_OPCODE_Y1) { + + opcode = get_UnaryOpcodeExtension_Y1(bundle); + + /* + * Test "jalr", "jalrp", "jr", "jrp" instruction at Y1 + * pipeline. + */ + switch (opcode) { + case JALR_UNARY_OPCODE_Y1: + case JALRP_UNARY_OPCODE_Y1: + y1_lr = true; + y1_lr_reg = 55; /* Link register. */ + /* FALLTHROUGH */ + case JR_UNARY_OPCODE_Y1: + case JRP_UNARY_OPCODE_Y1: + y1_br = true; + y1_br_reg = get_SrcA_Y1(bundle); + break; + case LNK_UNARY_OPCODE_Y1: + /* "lnk" at Y1 pipeline. */ + y1_lr = true; + y1_lr_reg = get_Dest_Y1(bundle); + break; + } + } + + opcode = get_Opcode_Y2(bundle); + mod = get_Mode(bundle); + + /* + * bundle_2 is bundle after making Y2 as a dummy operation + * - ld zero, sp + */ + bundle_2 = (bundle & (~GX_INSN_Y2_MASK)) | jit_y2_dummy(); + + /* Make Y1 as fnop if Y1 is a branch or lnk operation. */ + if (y1_br || y1_lr) { + bundle_2 &= ~(GX_INSN_Y1_MASK); + bundle_2 |= jit_y1_fnop(); + } + + if (is_y0_y1_nop(bundle_2)) + bundle_2_enable = false; + + if (mod == MODE_OPCODE_YC2) { + /* Store. */ + load_n_store = false; + load_store_size = 1 << opcode; + load_store_signed = false; + find_regs(bundle, 0, &ra, &rb, &clob1, &clob2, + &clob3, &alias); + if (load_store_size > 8) + unexpected = true; + } else { + /* Load. */ + load_n_store = true; + if (mod == MODE_OPCODE_YB2) { + switch (opcode) { + case LD_OPCODE_Y2: + load_store_signed = false; + load_store_size = 8; + break; + case LD4S_OPCODE_Y2: + load_store_signed = true; + load_store_size = 4; + break; + case LD4U_OPCODE_Y2: + load_store_signed = false; + load_store_size = 4; + break; + default: + unexpected = true; + } + } else if (mod == MODE_OPCODE_YA2) { + if (opcode == LD2S_OPCODE_Y2) { + load_store_signed = true; + load_store_size = 2; + } else if (opcode == LD2U_OPCODE_Y2) { + load_store_signed = false; + load_store_size = 2; + } else + unexpected = true; + } else + unexpected = true; + find_regs(bundle, &rd, &ra, &rb, &clob1, &clob2, + &clob3, &alias); + } + } else { + unsigned int opcode; + + /* bundle_2 is bundle after making X1 as "fnop". */ + bundle_2 = (bundle & (~GX_INSN_X1_MASK)) | jit_x1_fnop(); + + if (is_x0_x1_nop(bundle_2)) + bundle_2_enable = false; + + if (get_Opcode_X1(bundle) == RRR_0_OPCODE_X1) { + opcode = get_UnaryOpcodeExtension_X1(bundle); + + if (get_RRROpcodeExtension_X1(bundle) == + UNARY_RRR_0_OPCODE_X1) { + load_n_store = true; + find_regs(bundle, &rd, &ra, &rb, &clob1, + &clob2, &clob3, &alias); + + switch (opcode) { + case LD_UNARY_OPCODE_X1: + load_store_signed = false; + load_store_size = 8; + break; + case LD4S_UNARY_OPCODE_X1: + load_store_signed = true; + /* FALLTHROUGH */ + case LD4U_UNARY_OPCODE_X1: + load_store_size = 4; + break; + + case LD2S_UNARY_OPCODE_X1: + load_store_signed = true; + /* FALLTHROUGH */ + case LD2U_UNARY_OPCODE_X1: + load_store_size = 2; + break; + default: + unexpected = true; + } + } else { + load_n_store = false; + load_store_signed = false; + find_regs(bundle, 0, &ra, &rb, + &clob1, &clob2, &clob3, + &alias); + + opcode = get_RRROpcodeExtension_X1(bundle); + switch (opcode) { + case ST_RRR_0_OPCODE_X1: + load_store_size = 8; + break; + case ST4_RRR_0_OPCODE_X1: + load_store_size = 4; + break; + case ST2_RRR_0_OPCODE_X1: + load_store_size = 2; + break; + default: + unexpected = true; + } + } + } else if (get_Opcode_X1(bundle) == IMM8_OPCODE_X1) { + load_n_store = true; + opcode = get_Imm8OpcodeExtension_X1(bundle); + switch (opcode) { + case LD_ADD_IMM8_OPCODE_X1: + load_store_size = 8; + break; + + case LD4S_ADD_IMM8_OPCODE_X1: + load_store_signed = true; + /* FALLTHROUGH */ + case LD4U_ADD_IMM8_OPCODE_X1: + load_store_size = 4; + break; + + case LD2S_ADD_IMM8_OPCODE_X1: + load_store_signed = true; + /* FALLTHROUGH */ + case LD2U_ADD_IMM8_OPCODE_X1: + load_store_size = 2; + break; + + case ST_ADD_IMM8_OPCODE_X1: + load_n_store = false; + load_store_size = 8; + break; + case ST4_ADD_IMM8_OPCODE_X1: + load_n_store = false; + load_store_size = 4; + break; + case ST2_ADD_IMM8_OPCODE_X1: + load_n_store = false; + load_store_size = 2; + break; + default: + unexpected = true; + } + + if (!unexpected) { + x1_add = true; + if (load_n_store) + x1_add_imm8 = get_Imm8_X1(bundle); + else + x1_add_imm8 = get_Dest_Imm8_X1(bundle); + } + + find_regs(bundle, load_n_store ? (&rd) : NULL, + &ra, &rb, &clob1, &clob2, &clob3, &alias); + } else + unexpected = true; + } + + /* + * Some sanity check for register numbers extracted from fault bundle. + */ + if (check_regs(rd, ra, rb, clob1, clob2, clob3) == true) + unexpected = true; + + /* Give warning if register ra has an aligned address. */ + if (!unexpected) + WARN_ON(!((load_store_size - 1) & (regs->regs[ra]))); + + + /* + * Fault came from kernel space, here we only need take care of + * unaligned "get_user/put_user" macros defined in "uaccess.h". + * Basically, we will handle bundle like this: + * {ld/2u/4s rd, ra; movei rx, 0} or {st/2/4 ra, rb; movei rx, 0} + * (Refer to file "arch/tile/include/asm/uaccess.h" for details). + * For either load or store, byte-wise operation is performed by calling + * get_user() or put_user(). If the macro returns non-zero value, + * set the value to rx, otherwise set zero to rx. Finally make pc point + * to next bundle and return. + */ + + if (EX1_PL(regs->ex1) != USER_PL) { + + unsigned long rx = 0; + unsigned long x = 0, ret = 0; + + if (y1_br || y1_lr || x1_add || + (load_store_signed != + (load_n_store && load_store_size == 4))) { + /* No branch, link, wrong sign-ext or load/store add. */ + unexpected = true; + } else if (!unexpected) { + if (bundle & TILEGX_BUNDLE_MODE_MASK) { + /* + * Fault bundle is Y mode. + * Check if the Y1 and Y0 is the form of + * { movei rx, 0; nop/fnop }, if yes, + * find the rx. + */ + + if ((get_Opcode_Y1(bundle) == ADDI_OPCODE_Y1) + && (get_SrcA_Y1(bundle) == TREG_ZERO) && + (get_Imm8_Y1(bundle) == 0) && + is_bundle_y0_nop(bundle)) { + rx = get_Dest_Y1(bundle); + } else if ((get_Opcode_Y0(bundle) == + ADDI_OPCODE_Y0) && + (get_SrcA_Y0(bundle) == TREG_ZERO) && + (get_Imm8_Y0(bundle) == 0) && + is_bundle_y1_nop(bundle)) { + rx = get_Dest_Y0(bundle); + } else { + unexpected = true; + } + } else { + /* + * Fault bundle is X mode. + * Check if the X0 is 'movei rx, 0', + * if yes, find the rx. + */ + + if ((get_Opcode_X0(bundle) == IMM8_OPCODE_X0) + && (get_Imm8OpcodeExtension_X0(bundle) == + ADDI_IMM8_OPCODE_X0) && + (get_SrcA_X0(bundle) == TREG_ZERO) && + (get_Imm8_X0(bundle) == 0)) { + rx = get_Dest_X0(bundle); + } else { + unexpected = true; + } + } + + /* rx should be less than 56. */ + if (!unexpected && (rx >= 56)) + unexpected = true; + } + + if (!search_exception_tables(regs->pc)) { + /* No fixup in the exception tables for the pc. */ + unexpected = true; + } + + if (unexpected) { + /* Unexpected unalign kernel fault. */ + struct task_struct *tsk = validate_current(); + + bust_spinlocks(1); + + show_regs(regs); + + if (unlikely(tsk->pid < 2)) { + panic("Kernel unalign fault running %s!", + tsk->pid ? "init" : "the idle task"); + } +#ifdef SUPPORT_DIE + die("Oops", regs); +#endif + bust_spinlocks(1); + + do_group_exit(SIGKILL); + + } else { + unsigned long i, b = 0; + unsigned char *ptr = + (unsigned char *)regs->regs[ra]; + if (load_n_store) { + /* handle get_user(x, ptr) */ + for (i = 0; i < load_store_size; i++) { + ret = get_user(b, ptr++); + if (!ret) { + /* Success! update x. */ +#ifdef __LITTLE_ENDIAN + x |= (b << (8 * i)); +#else + x <<= 8; + x |= b; +#endif /* __LITTLE_ENDIAN */ + } else { + x = 0; + break; + } + } + + /* Sign-extend 4-byte loads. */ + if (load_store_size == 4) + x = (long)(int)x; + + /* Set register rd. */ + regs->regs[rd] = x; + + /* Set register rx. */ + regs->regs[rx] = ret; + + /* Bump pc. */ + regs->pc += 8; + + } else { + /* Handle put_user(x, ptr) */ + x = regs->regs[rb]; +#ifdef __LITTLE_ENDIAN + b = x; +#else + /* + * Swap x in order to store x from low + * to high memory same as the + * little-endian case. + */ + switch (load_store_size) { + case 8: + b = swab64(x); + break; + case 4: + b = swab32(x); + break; + case 2: + b = swab16(x); + break; + } +#endif /* __LITTLE_ENDIAN */ + for (i = 0; i < load_store_size; i++) { + ret = put_user(b, ptr++); + if (ret) + break; + /* Success! shift 1 byte. */ + b >>= 8; + } + /* Set register rx. */ + regs->regs[rx] = ret; + + /* Bump pc. */ + regs->pc += 8; + } + } + + unaligned_fixup_count++; + + if (unaligned_printk) { + pr_info("%s/%d. Unalign fixup for kernel access " + "to userspace %lx.", + current->comm, current->pid, regs->regs[ra]); + } + + /* Done! Return to the exception handler. */ + return; + } + + if ((align_ctl == 0) || unexpected) { + siginfo_t info = { + .si_signo = SIGBUS, + .si_code = BUS_ADRALN, + .si_addr = (unsigned char __user *)0 + }; + if (unaligned_printk) + pr_info("Unalign bundle: unexp @%llx, %llx", + (unsigned long long)regs->pc, + (unsigned long long)bundle); + + if (ra < 56) { + unsigned long uaa = (unsigned long)regs->regs[ra]; + /* Set bus Address. */ + info.si_addr = (unsigned char __user *)uaa; + } + + unaligned_fixup_count++; + + trace_unhandled_signal("unaligned fixup trap", regs, + (unsigned long)info.si_addr, SIGBUS); + force_sig_info(info.si_signo, &info, current); + return; + } + +#ifdef __LITTLE_ENDIAN +#define UA_FIXUP_ADDR_DELTA 1 +#define UA_FIXUP_BFEXT_START(_B_) 0 +#define UA_FIXUP_BFEXT_END(_B_) (8 * (_B_) - 1) +#else /* __BIG_ENDIAN */ +#define UA_FIXUP_ADDR_DELTA -1 +#define UA_FIXUP_BFEXT_START(_B_) (64 - 8 * (_B_)) +#define UA_FIXUP_BFEXT_END(_B_) 63 +#endif /* __LITTLE_ENDIAN */ + + + + if ((ra != rb) && (rd != TREG_SP) && !alias && + !y1_br && !y1_lr && !x1_add) { + /* + * Simple case: ra != rb and no register alias found, + * and no branch or link. This will be the majority. + * We can do a little better for simplae case than the + * generic scheme below. + */ + if (!load_n_store) { + /* + * Simple store: ra != rb, no need for scratch register. + * Just store and rotate to right bytewise. + */ +#ifdef __BIG_ENDIAN + frag.insn[n++] = + jit_x0_addi(ra, ra, load_store_size - 1) | + jit_x1_fnop(); +#endif /* __BIG_ENDIAN */ + for (k = 0; k < load_store_size; k++) { + /* Store a byte. */ + frag.insn[n++] = + jit_x0_rotli(rb, rb, 56) | + jit_x1_st1_add(ra, rb, + UA_FIXUP_ADDR_DELTA); + } +#ifdef __BIG_ENDIAN + frag.insn[n] = jit_x1_addi(ra, ra, 1); +#else + frag.insn[n] = jit_x1_addi(ra, ra, + -1 * load_store_size); +#endif /* __LITTLE_ENDIAN */ + + if (load_store_size == 8) { + frag.insn[n] |= jit_x0_fnop(); + } else if (load_store_size == 4) { + frag.insn[n] |= jit_x0_rotli(rb, rb, 32); + } else { /* = 2 */ + frag.insn[n] |= jit_x0_rotli(rb, rb, 16); + } + n++; + if (bundle_2_enable) + frag.insn[n++] = bundle_2; + frag.insn[n++] = jit_x0_fnop() | jit_x1_iret(); + } else { + if (rd == ra) { + /* Use two clobber registers: clob1/2. */ + frag.insn[n++] = + jit_x0_addi(TREG_SP, TREG_SP, -16) | + jit_x1_fnop(); + frag.insn[n++] = + jit_x0_addi(clob1, ra, 7) | + jit_x1_st_add(TREG_SP, clob1, -8); + frag.insn[n++] = + jit_x0_addi(clob2, ra, 0) | + jit_x1_st(TREG_SP, clob2); + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_ldna(rd, ra); + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_ldna(clob1, clob1); + /* + * Note: we must make sure that rd must not + * be sp. Recover clob1/2 from stack. + */ + frag.insn[n++] = + jit_x0_dblalign(rd, clob1, clob2) | + jit_x1_ld_add(clob2, TREG_SP, 8); + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_ld_add(clob1, TREG_SP, 16); + } else { + /* Use one clobber register: clob1 only. */ + frag.insn[n++] = + jit_x0_addi(TREG_SP, TREG_SP, -16) | + jit_x1_fnop(); + frag.insn[n++] = + jit_x0_addi(clob1, ra, 7) | + jit_x1_st(TREG_SP, clob1); + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_ldna(rd, ra); + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_ldna(clob1, clob1); + /* + * Note: we must make sure that rd must not + * be sp. Recover clob1 from stack. + */ + frag.insn[n++] = + jit_x0_dblalign(rd, clob1, ra) | + jit_x1_ld_add(clob1, TREG_SP, 16); + } + + if (bundle_2_enable) + frag.insn[n++] = bundle_2; + /* + * For non 8-byte load, extract corresponding bytes and + * signed extension. + */ + if (load_store_size == 4) { + if (load_store_signed) + frag.insn[n++] = + jit_x0_bfexts( + rd, rd, + UA_FIXUP_BFEXT_START(4), + UA_FIXUP_BFEXT_END(4)) | + jit_x1_fnop(); + else + frag.insn[n++] = + jit_x0_bfextu( + rd, rd, + UA_FIXUP_BFEXT_START(4), + UA_FIXUP_BFEXT_END(4)) | + jit_x1_fnop(); + } else if (load_store_size == 2) { + if (load_store_signed) + frag.insn[n++] = + jit_x0_bfexts( + rd, rd, + UA_FIXUP_BFEXT_START(2), + UA_FIXUP_BFEXT_END(2)) | + jit_x1_fnop(); + else + frag.insn[n++] = + jit_x0_bfextu( + rd, rd, + UA_FIXUP_BFEXT_START(2), + UA_FIXUP_BFEXT_END(2)) | + jit_x1_fnop(); + } + + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_iret(); + } + } else if (!load_n_store) { + + /* + * Generic memory store cases: use 3 clobber registers. + * + * Alloc space for saveing clob2,1,3 on user's stack. + * register clob3 points to where clob2 saved, followed by + * clob1 and 3 from high to low memory. + */ + frag.insn[n++] = + jit_x0_addi(TREG_SP, TREG_SP, -32) | + jit_x1_fnop(); + frag.insn[n++] = + jit_x0_addi(clob3, TREG_SP, 16) | + jit_x1_st_add(TREG_SP, clob3, 8); +#ifdef __LITTLE_ENDIAN + frag.insn[n++] = + jit_x0_addi(clob1, ra, 0) | + jit_x1_st_add(TREG_SP, clob1, 8); +#else + frag.insn[n++] = + jit_x0_addi(clob1, ra, load_store_size - 1) | + jit_x1_st_add(TREG_SP, clob1, 8); +#endif + if (load_store_size == 8) { + /* + * We save one byte a time, not for fast, but compact + * code. After each store, data source register shift + * right one byte. unchanged after 8 stores. + */ + frag.insn[n++] = + jit_x0_addi(clob2, TREG_ZERO, 7) | + jit_x1_st_add(TREG_SP, clob2, 16); + frag.insn[n++] = + jit_x0_rotli(rb, rb, 56) | + jit_x1_st1_add(clob1, rb, UA_FIXUP_ADDR_DELTA); + frag.insn[n++] = + jit_x0_addi(clob2, clob2, -1) | + jit_x1_bnezt(clob2, -1); + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_addi(clob2, y1_br_reg, 0); + } else if (load_store_size == 4) { + frag.insn[n++] = + jit_x0_addi(clob2, TREG_ZERO, 3) | + jit_x1_st_add(TREG_SP, clob2, 16); + frag.insn[n++] = + jit_x0_rotli(rb, rb, 56) | + jit_x1_st1_add(clob1, rb, UA_FIXUP_ADDR_DELTA); + frag.insn[n++] = + jit_x0_addi(clob2, clob2, -1) | + jit_x1_bnezt(clob2, -1); + /* + * same as 8-byte case, but need shift another 4 + * byte to recover rb for 4-byte store. + */ + frag.insn[n++] = jit_x0_rotli(rb, rb, 32) | + jit_x1_addi(clob2, y1_br_reg, 0); + } else { /* =2 */ + frag.insn[n++] = + jit_x0_addi(clob2, rb, 0) | + jit_x1_st_add(TREG_SP, clob2, 16); + for (k = 0; k < 2; k++) { + frag.insn[n++] = + jit_x0_shrui(rb, rb, 8) | + jit_x1_st1_add(clob1, rb, + UA_FIXUP_ADDR_DELTA); + } + frag.insn[n++] = + jit_x0_addi(rb, clob2, 0) | + jit_x1_addi(clob2, y1_br_reg, 0); + } + + if (bundle_2_enable) + frag.insn[n++] = bundle_2; + + if (y1_lr) { + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_mfspr(y1_lr_reg, + SPR_EX_CONTEXT_0_0); + } + if (y1_br) { + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_mtspr(SPR_EX_CONTEXT_0_0, + clob2); + } + if (x1_add) { + frag.insn[n++] = + jit_x0_addi(ra, ra, x1_add_imm8) | + jit_x1_ld_add(clob2, clob3, -8); + } else { + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_ld_add(clob2, clob3, -8); + } + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_ld_add(clob1, clob3, -8); + frag.insn[n++] = jit_x0_fnop() | jit_x1_ld(clob3, clob3); + frag.insn[n++] = jit_x0_fnop() | jit_x1_iret(); + + } else { + /* + * Generic memory load cases. + * + * Alloc space for saveing clob1,2,3 on user's stack. + * register clob3 points to where clob1 saved, followed + * by clob2 and 3 from high to low memory. + */ + + frag.insn[n++] = + jit_x0_addi(TREG_SP, TREG_SP, -32) | + jit_x1_fnop(); + frag.insn[n++] = + jit_x0_addi(clob3, TREG_SP, 16) | + jit_x1_st_add(TREG_SP, clob3, 8); + frag.insn[n++] = + jit_x0_addi(clob2, ra, 0) | + jit_x1_st_add(TREG_SP, clob2, 8); + + if (y1_br) { + frag.insn[n++] = + jit_x0_addi(clob1, y1_br_reg, 0) | + jit_x1_st_add(TREG_SP, clob1, 16); + } else { + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_st_add(TREG_SP, clob1, 16); + } + + if (bundle_2_enable) + frag.insn[n++] = bundle_2; + + if (y1_lr) { + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_mfspr(y1_lr_reg, + SPR_EX_CONTEXT_0_0); + } + + if (y1_br) { + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_mtspr(SPR_EX_CONTEXT_0_0, + clob1); + } + + frag.insn[n++] = + jit_x0_addi(clob1, clob2, 7) | + jit_x1_ldna(rd, clob2); + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_ldna(clob1, clob1); + frag.insn[n++] = + jit_x0_dblalign(rd, clob1, clob2) | + jit_x1_ld_add(clob1, clob3, -8); + if (x1_add) { + frag.insn[n++] = + jit_x0_addi(ra, ra, x1_add_imm8) | + jit_x1_ld_add(clob2, clob3, -8); + } else { + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_ld_add(clob2, clob3, -8); + } + + frag.insn[n++] = + jit_x0_fnop() | + jit_x1_ld(clob3, clob3); + + if (load_store_size == 4) { + if (load_store_signed) + frag.insn[n++] = + jit_x0_bfexts( + rd, rd, + UA_FIXUP_BFEXT_START(4), + UA_FIXUP_BFEXT_END(4)) | + jit_x1_fnop(); + else + frag.insn[n++] = + jit_x0_bfextu( + rd, rd, + UA_FIXUP_BFEXT_START(4), + UA_FIXUP_BFEXT_END(4)) | + jit_x1_fnop(); + } else if (load_store_size == 2) { + if (load_store_signed) + frag.insn[n++] = + jit_x0_bfexts( + rd, rd, + UA_FIXUP_BFEXT_START(2), + UA_FIXUP_BFEXT_END(2)) | + jit_x1_fnop(); + else + frag.insn[n++] = + jit_x0_bfextu( + rd, rd, + UA_FIXUP_BFEXT_START(2), + UA_FIXUP_BFEXT_END(2)) | + jit_x1_fnop(); + } + + frag.insn[n++] = jit_x0_fnop() | jit_x1_iret(); + } + + /* Max JIT bundle count is 14. */ + WARN_ON(n > 14); + + if (!unexpected) { + int status = 0; + int idx = (regs->pc >> 3) & + ((1ULL << (PAGE_SHIFT - UNALIGN_JIT_SHIFT)) - 1); + + frag.pc = regs->pc; + frag.bundle = bundle; + + if (unaligned_printk) { + pr_info("%s/%d, Unalign fixup: pc=%lx " + "bundle=%lx %d %d %d %d %d %d %d %d.", + current->comm, current->pid, + (unsigned long)frag.pc, + (unsigned long)frag.bundle, + (int)alias, (int)rd, (int)ra, + (int)rb, (int)bundle_2_enable, + (int)y1_lr, (int)y1_br, (int)x1_add); + + for (k = 0; k < n; k += 2) + pr_info("[%d] %016llx %016llx", k, + (unsigned long long)frag.insn[k], + (unsigned long long)frag.insn[k+1]); + } + + /* Swap bundle byte order for big endian sys. */ +#ifdef __BIG_ENDIAN + frag.bundle = GX_INSN_BSWAP(frag.bundle); + for (k = 0; k < n; k++) + frag.insn[k] = GX_INSN_BSWAP(frag.insn[k]); +#endif /* __BIG_ENDIAN */ + + status = copy_to_user((void __user *)&jit_code_area[idx], + &frag, sizeof(frag)); + if (status) { + /* Fail to copy JIT into user land. send SIGSEGV. */ + siginfo_t info = { + .si_signo = SIGSEGV, + .si_code = SEGV_MAPERR, + .si_addr = (void __user *)&jit_code_area[idx] + }; + + pr_warn("Unalign fixup: pid=%d %s jit_code_area=%llx", + current->pid, current->comm, + (unsigned long long)&jit_code_area[idx]); + + trace_unhandled_signal("segfault in unalign fixup", + regs, + (unsigned long)info.si_addr, + SIGSEGV); + force_sig_info(info.si_signo, &info, current); + return; + } + + + /* Do a cheaper increment, not accurate. */ + unaligned_fixup_count++; + __flush_icache_range((unsigned long)&jit_code_area[idx], + (unsigned long)&jit_code_area[idx] + + sizeof(frag)); + + /* Setup SPR_EX_CONTEXT_0_0/1 for returning to user program.*/ + __insn_mtspr(SPR_EX_CONTEXT_0_0, regs->pc + 8); + __insn_mtspr(SPR_EX_CONTEXT_0_1, PL_ICS_EX1(USER_PL, 0)); + + /* Modify pc at the start of new JIT. */ + regs->pc = (unsigned long)&jit_code_area[idx].insn[0]; + /* Set ICS in SPR_EX_CONTEXT_K_1. */ + regs->ex1 = PL_ICS_EX1(USER_PL, 1); + } +} + + +/* + * C function to generate unalign data JIT. Called from unalign data + * interrupt handler. + * + * First check if unalign fix is disabled or exception did not not come from + * user space or sp register points to unalign address, if true, generate a + * SIGBUS. Then map a page into user space as JIT area if it is not mapped + * yet. Genenerate JIT code by calling jit_bundle_gen(). After that return + * back to exception handler. + * + * The exception handler will "iret" to new generated JIT code after + * restoring caller saved registers. In theory, the JIT code will perform + * another "iret" to resume user's program. + */ + +void do_unaligned(struct pt_regs *regs, int vecnum) +{ + tilegx_bundle_bits __user *pc; + tilegx_bundle_bits bundle; + struct thread_info *info = current_thread_info(); + int align_ctl; + + /* Checks the per-process unaligned JIT flags */ + align_ctl = unaligned_fixup; + switch (task_thread_info(current)->align_ctl) { + case PR_UNALIGN_NOPRINT: + align_ctl = 1; + break; + case PR_UNALIGN_SIGBUS: + align_ctl = 0; + break; + } + + /* Enable iterrupt in order to access user land. */ + local_irq_enable(); + + /* + * The fault came from kernel space. Two choices: + * (a) unaligned_fixup < 1, we will first call get/put_user fixup + * to return -EFAULT. If no fixup, simply panic the kernel. + * (b) unaligned_fixup >=1, we will try to fix the unaligned access + * if it was triggered by get_user/put_user() macros. Panic the + * kernel if it is not fixable. + */ + + if (EX1_PL(regs->ex1) != USER_PL) { + + if (align_ctl < 1) { + unaligned_fixup_count++; + /* If exception came from kernel, try fix it up. */ + if (fixup_exception(regs)) { + if (unaligned_printk) + pr_info("Unalign fixup: %d %llx @%llx", + (int)unaligned_fixup, + (unsigned long long)regs->ex1, + (unsigned long long)regs->pc); + return; + } + /* Not fixable. Go panic. */ + panic("Unalign exception in Kernel. pc=%lx", + regs->pc); + return; + } else { + /* + * Try to fix the exception. If we can't, panic the + * kernel. + */ + bundle = GX_INSN_BSWAP( + *((tilegx_bundle_bits *)(regs->pc))); + jit_bundle_gen(regs, bundle, align_ctl); + return; + } + } + + /* + * Fault came from user with ICS or stack is not aligned. + * If so, we will trigger SIGBUS. + */ + if ((regs->sp & 0x7) || (regs->ex1) || (align_ctl < 0)) { + siginfo_t info = { + .si_signo = SIGBUS, + .si_code = BUS_ADRALN, + .si_addr = (unsigned char __user *)0 + }; + + if (unaligned_printk) + pr_info("Unalign fixup: %d %llx @%llx", + (int)unaligned_fixup, + (unsigned long long)regs->ex1, + (unsigned long long)regs->pc); + + unaligned_fixup_count++; + + trace_unhandled_signal("unaligned fixup trap", regs, 0, SIGBUS); + force_sig_info(info.si_signo, &info, current); + return; + } + + + /* Read the bundle casued the exception! */ + pc = (tilegx_bundle_bits __user *)(regs->pc); + if (get_user(bundle, pc) != 0) { + /* Probably never be here since pc is valid user address.*/ + siginfo_t info = { + .si_signo = SIGSEGV, + .si_code = SEGV_MAPERR, + .si_addr = (void __user *)pc + }; + pr_err("Couldn't read instruction at %p trying to step\n", pc); + trace_unhandled_signal("segfault in unalign fixup", regs, + (unsigned long)info.si_addr, SIGSEGV); + force_sig_info(info.si_signo, &info, current); + return; + } + + if (!info->unalign_jit_base) { + void __user *user_page; + + /* + * Allocate a page in userland. + * For 64-bit processes we try to place the mapping far + * from anything else that might be going on (specifically + * 64 GB below the top of the user address space). If it + * happens not to be possible to put it there, it's OK; + * the kernel will choose another location and we'll + * remember it for later. + */ + if (is_compat_task()) + user_page = NULL; + else + user_page = (void __user *)(TASK_SIZE - (1UL << 36)) + + (current->pid << PAGE_SHIFT); + + user_page = (void __user *) vm_mmap(NULL, + (unsigned long)user_page, + PAGE_SIZE, + PROT_EXEC | PROT_READ | + PROT_WRITE, +#ifdef CONFIG_HOMECACHE + MAP_CACHE_HOME_TASK | +#endif + MAP_PRIVATE | + MAP_ANONYMOUS, + 0); + + if (IS_ERR((void __force *)user_page)) { + pr_err("Out of kernel pages trying do_mmap.\n"); + return; + } + + /* Save the address in the thread_info struct */ + info->unalign_jit_base = user_page; + if (unaligned_printk) + pr_info("Unalign bundle: %d:%d, allocate page @%llx", + raw_smp_processor_id(), current->pid, + (unsigned long long)user_page); + } + + /* Generate unalign JIT */ + jit_bundle_gen(regs, GX_INSN_BSWAP(bundle), align_ctl); +} + +#endif /* __tilegx__ */ diff --git a/arch/tile/mm/fault.c b/arch/tile/mm/fault.c index 6152819e555b..7863298dad4d 100644 --- a/arch/tile/mm/fault.c +++ b/arch/tile/mm/fault.c @@ -722,8 +722,49 @@ void do_page_fault(struct pt_regs *regs, int fault_num, { int is_page_fault; +#ifdef __tilegx__ + /* + * We don't need early do_page_fault_ics() support, since unlike + * Pro we don't need to worry about unlocking the atomic locks. + * There is only one current case in GX where we touch any memory + * under ICS other than our own kernel stack, and we handle that + * here. (If we crash due to trying to touch our own stack, + * we're in too much trouble for C code to help out anyway.) + */ + if (write & ~1) { + unsigned long pc = write & ~1; + if (pc >= (unsigned long) __start_unalign_asm_code && + pc < (unsigned long) __end_unalign_asm_code) { + struct thread_info *ti = current_thread_info(); + /* + * Our EX_CONTEXT is still what it was from the + * initial unalign exception, but now we've faulted + * on the JIT page. We would like to complete the + * page fault however is appropriate, and then retry + * the instruction that caused the unalign exception. + * Our state has been "corrupted" by setting the low + * bit in "sp", and stashing r0..r3 in the + * thread_info area, so we revert all of that, then + * continue as if this were a normal page fault. + */ + regs->sp &= ~1UL; + regs->regs[0] = ti->unalign_jit_tmp[0]; + regs->regs[1] = ti->unalign_jit_tmp[1]; + regs->regs[2] = ti->unalign_jit_tmp[2]; + regs->regs[3] = ti->unalign_jit_tmp[3]; + write &= 1; + } else { + pr_alert("%s/%d: ICS set at page fault at %#lx: %#lx\n", + current->comm, current->pid, pc, address); + show_regs(regs); + do_group_exit(SIGKILL); + return; + } + } +#else /* This case should have been handled by do_page_fault_ics(). */ BUG_ON(write & ~1); +#endif #if CHIP_HAS_TILE_DMA() /*