diff options
author | Max Filippov <jcmvbkbc@gmail.com> | 2022-04-15 13:05:31 +0300 |
---|---|---|
committer | Max Filippov <jcmvbkbc@gmail.com> | 2022-05-02 05:51:23 +0300 |
commit | 11e969bc964a0e50ae64cdba092048e3937d2389 (patch) | |
tree | 99303eb470e622faf5397649005d6f221f46bdb0 /arch/xtensa | |
parent | f29cab2906346fa93831376256e01f5a6629f979 (diff) | |
download | linux-11e969bc964a0e50ae64cdba092048e3937d2389.tar.xz |
xtensa: support coprocessors on SMP
Current coprocessor support on xtensa only works correctly on
uniprocessor configurations. Make it work on SMP too and keep it lazy.
Make coprocessor_owner array per-CPU and move it to struct exc_table for
easy access from the fast_coprocessor exception handler. Allow task to
have live coprocessors only on single CPU, record this CPU number in the
struct thread_info::cp_owner_cpu. Change struct thread_info::cpenable
meaning to be 'coprocessors live on cp_owner_cpu'.
Introduce C-level coprocessor exception handler that flushes and
releases live coprocessors of the task taking 'coprocessor disabled'
exception and call it from the fast_coprocessor handler when the task
has live coprocessors on other CPU.
Make coprocessor_flush_all and coprocessor_release_all work correctly
when called from any CPU by sending IPI to the cp_owner_cpu. Add
function coprocessor_flush_release_all to do flush followed by release
atomically. Add function local_coprocessors_flush_release_all to flush
and release all coprocessors on the local CPU and use it to flush
coprocessor contexts from the CPU that goes offline.
Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>
Diffstat (limited to 'arch/xtensa')
-rw-r--r-- | arch/xtensa/include/asm/coprocessor.h | 4 | ||||
-rw-r--r-- | arch/xtensa/include/asm/thread_info.h | 7 | ||||
-rw-r--r-- | arch/xtensa/include/asm/traps.h | 6 | ||||
-rw-r--r-- | arch/xtensa/kernel/asm-offsets.c | 8 | ||||
-rw-r--r-- | arch/xtensa/kernel/coprocessor.S | 122 | ||||
-rw-r--r-- | arch/xtensa/kernel/entry.S | 12 | ||||
-rw-r--r-- | arch/xtensa/kernel/process.c | 112 | ||||
-rw-r--r-- | arch/xtensa/kernel/ptrace.c | 3 | ||||
-rw-r--r-- | arch/xtensa/kernel/signal.c | 3 | ||||
-rw-r--r-- | arch/xtensa/kernel/smp.c | 7 | ||||
-rw-r--r-- | arch/xtensa/kernel/traps.c | 13 |
11 files changed, 230 insertions, 67 deletions
diff --git a/arch/xtensa/include/asm/coprocessor.h b/arch/xtensa/include/asm/coprocessor.h index a360efced7e7..3b1a0d5d2169 100644 --- a/arch/xtensa/include/asm/coprocessor.h +++ b/arch/xtensa/include/asm/coprocessor.h @@ -142,10 +142,12 @@ typedef struct { XCHAL_CP6_SA_LIST(2) } xtregs_cp6_t typedef struct { XCHAL_CP7_SA_LIST(2) } xtregs_cp7_t __attribute__ ((aligned (XCHAL_CP7_SA_ALIGN))); -extern struct thread_info* coprocessor_owner[XCHAL_CP_MAX]; +struct thread_info; void coprocessor_flush(struct thread_info *ti, int cp_index); void coprocessor_release_all(struct thread_info *ti); void coprocessor_flush_all(struct thread_info *ti); +void coprocessor_flush_release_all(struct thread_info *ti); +void local_coprocessors_flush_release_all(void); #endif /* XTENSA_HAVE_COPROCESSORS */ diff --git a/arch/xtensa/include/asm/thread_info.h b/arch/xtensa/include/asm/thread_info.h index f6fcbba1d02f..52974317a6b6 100644 --- a/arch/xtensa/include/asm/thread_info.h +++ b/arch/xtensa/include/asm/thread_info.h @@ -52,12 +52,17 @@ struct thread_info { __u32 cpu; /* current CPU */ __s32 preempt_count; /* 0 => preemptable,< 0 => BUG*/ - unsigned long cpenable; #if XCHAL_HAVE_EXCLUSIVE /* result of the most recent exclusive store */ unsigned long atomctl8; #endif + /* + * If i-th bit is set then coprocessor state is loaded into the + * coprocessor i on CPU cp_owner_cpu. + */ + unsigned long cpenable; + u32 cp_owner_cpu; /* Allocate storage for extra user states and coprocessor states. */ #if XTENSA_HAVE_COPROCESSORS xtregs_coprocessor_t xtregs_cp; diff --git a/arch/xtensa/include/asm/traps.h b/arch/xtensa/include/asm/traps.h index 514376eff58c..6f74ccc0c7ea 100644 --- a/arch/xtensa/include/asm/traps.h +++ b/arch/xtensa/include/asm/traps.h @@ -27,6 +27,10 @@ struct exc_table { void *fixup; /* For passing a parameter to fixup */ void *fixup_param; +#if XTENSA_HAVE_COPROCESSORS + /* Pointers to owner struct thread_info */ + struct thread_info *coprocessor_owner[XCHAL_CP_MAX]; +#endif /* Fast user exception handlers */ void *fast_user_handler[EXCCAUSE_N]; /* Fast kernel exception handlers */ @@ -35,6 +39,8 @@ struct exc_table { xtensa_exception_handler *default_handler[EXCCAUSE_N]; }; +DECLARE_PER_CPU(struct exc_table, exc_table); + xtensa_exception_handler * __init trap_set_handler(int cause, xtensa_exception_handler *handler); diff --git a/arch/xtensa/kernel/asm-offsets.c b/arch/xtensa/kernel/asm-offsets.c index 37278e2785fb..e3b9cf4c2289 100644 --- a/arch/xtensa/kernel/asm-offsets.c +++ b/arch/xtensa/kernel/asm-offsets.c @@ -91,10 +91,12 @@ int main(void) /* struct thread_info (offset from start_struct) */ DEFINE(THREAD_RA, offsetof (struct task_struct, thread.ra)); DEFINE(THREAD_SP, offsetof (struct task_struct, thread.sp)); - DEFINE(THREAD_CPENABLE, offsetof (struct thread_info, cpenable)); #if XCHAL_HAVE_EXCLUSIVE DEFINE(THREAD_ATOMCTL8, offsetof (struct thread_info, atomctl8)); #endif + DEFINE(THREAD_CPENABLE, offsetof(struct thread_info, cpenable)); + DEFINE(THREAD_CPU, offsetof(struct thread_info, cpu)); + DEFINE(THREAD_CP_OWNER_CPU, offsetof(struct thread_info, cp_owner_cpu)); #if XTENSA_HAVE_COPROCESSORS DEFINE(THREAD_XTREGS_CP0, offsetof(struct thread_info, xtregs_cp.cp0)); DEFINE(THREAD_XTREGS_CP1, offsetof(struct thread_info, xtregs_cp.cp1)); @@ -137,6 +139,10 @@ int main(void) DEFINE(EXC_TABLE_DOUBLE_SAVE, offsetof(struct exc_table, double_save)); DEFINE(EXC_TABLE_FIXUP, offsetof(struct exc_table, fixup)); DEFINE(EXC_TABLE_PARAM, offsetof(struct exc_table, fixup_param)); +#if XTENSA_HAVE_COPROCESSORS + DEFINE(EXC_TABLE_COPROCESSOR_OWNER, + offsetof(struct exc_table, coprocessor_owner)); +#endif DEFINE(EXC_TABLE_FAST_USER, offsetof(struct exc_table, fast_user_handler)); DEFINE(EXC_TABLE_FAST_KERNEL, diff --git a/arch/xtensa/kernel/coprocessor.S b/arch/xtensa/kernel/coprocessor.S index 95412409c49e..ef33e76e07d8 100644 --- a/arch/xtensa/kernel/coprocessor.S +++ b/arch/xtensa/kernel/coprocessor.S @@ -19,6 +19,26 @@ #include <asm/current.h> #include <asm/regs.h> +/* + * Rules for coprocessor state manipulation on SMP: + * + * - a task may have live coprocessors only on one CPU. + * + * - whether coprocessor context of task T is live on some CPU is + * denoted by T's thread_info->cpenable. + * + * - non-zero thread_info->cpenable means that thread_info->cp_owner_cpu + * is valid in the T's thread_info. Zero thread_info->cpenable means that + * coprocessor context is valid in the T's thread_info. + * + * - if a coprocessor context of task T is live on CPU X, only CPU X changes + * T's thread_info->cpenable, cp_owner_cpu and coprocessor save area. + * This is done by making sure that for the task T with live coprocessor + * on CPU X cpenable SR is 0 when T runs on any other CPU Y. + * When fast_coprocessor exception is taken on CPU Y it goes to the + * C-level do_coprocessor that uses IPI to make CPU X flush T's coprocessors. + */ + #if XTENSA_HAVE_COPROCESSORS /* @@ -101,9 +121,37 @@ ENTRY(fast_coprocessor) + s32i a3, a2, PT_AREG3 + +#ifdef CONFIG_SMP + /* + * Check if any coprocessor context is live on another CPU + * and if so go through the C-level coprocessor exception handler + * to flush it to memory. + */ + GET_THREAD_INFO (a0, a2) + l32i a3, a0, THREAD_CPENABLE + beqz a3, .Lload_local + + /* + * Pairs with smp_wmb in local_coprocessor_release_all + * and with both memws below. + */ + memw + l32i a3, a0, THREAD_CPU + l32i a0, a0, THREAD_CP_OWNER_CPU + beq a0, a3, .Lload_local + + rsr a0, ps + l32i a3, a2, PT_AREG3 + bbci.l a0, PS_UM_BIT, 1f + call0 user_exception +1: call0 kernel_exception +#endif + /* Save remaining registers a1-a3 and SAR */ - s32i a3, a2, PT_AREG3 +.Lload_local: rsr a3, sar s32i a1, a2, PT_AREG1 s32i a3, a2, PT_SAR @@ -117,6 +165,9 @@ ENTRY(fast_coprocessor) s32i a5, a1, PT_AREG5 s32i a6, a1, PT_AREG6 s32i a7, a1, PT_AREG7 + s32i a8, a1, PT_AREG8 + s32i a9, a1, PT_AREG9 + s32i a10, a1, PT_AREG10 /* Find coprocessor number. Subtract first CP EXCCAUSE from EXCCAUSE */ @@ -139,51 +190,66 @@ ENTRY(fast_coprocessor) addx8 a7, a3, a7 addx4 a7, a3, a7 - /* Retrieve previous owner. (a3 still holds CP number) */ + /* Retrieve previous owner (a8). */ - movi a0, coprocessor_owner # list of owners + rsr a0, excsave1 # exc_table addx4 a0, a3, a0 # entry for CP - l32i a4, a0, 0 + l32i a8, a0, EXC_TABLE_COPROCESSOR_OWNER - beqz a4, 1f # skip 'save' if no previous owner + /* Set new owner (a9). */ - /* Disable coprocessor for previous owner. (a2 = 1 << CP number) */ + GET_THREAD_INFO (a9, a1) + l32i a4, a9, THREAD_CPU + s32i a9, a0, EXC_TABLE_COPROCESSOR_OWNER + s32i a4, a9, THREAD_CP_OWNER_CPU - l32i a5, a4, THREAD_CPENABLE - xor a5, a5, a2 # (1 << cp-id) still in a2 - s32i a5, a4, THREAD_CPENABLE + /* + * Enable coprocessor for the new owner. (a2 = 1 << CP number) + * This can be done before loading context into the coprocessor. + */ + l32i a4, a9, THREAD_CPENABLE + or a4, a4, a2 /* - * Get context save area and call save routine. - * (a4 still holds previous owner (thread_info), a3 CP number) + * Make sure THREAD_CP_OWNER_CPU is in memory before updating + * THREAD_CPENABLE */ + memw # (2) + s32i a4, a9, THREAD_CPENABLE - l32i a2, a7, CP_REGS_TAB_OFFSET - l32i a3, a7, CP_REGS_TAB_SAVE - add a2, a2, a4 - callx0 a3 + beqz a8, 1f # skip 'save' if no previous owner - /* Note that only a0 and a1 were preserved. */ + /* Disable coprocessor for previous owner. (a2 = 1 << CP number) */ - rsr a3, exccause - addi a3, a3, -EXCCAUSE_COPROCESSOR0_DISABLED - movi a0, coprocessor_owner - addx4 a0, a3, a0 + l32i a10, a8, THREAD_CPENABLE + xor a10, a10, a2 - /* Set new 'owner' (a0 points to the CP owner, a3 contains the CP nr) */ + /* Get context save area and call save routine. */ -1: GET_THREAD_INFO (a4, a1) - s32i a4, a0, 0 + l32i a2, a7, CP_REGS_TAB_OFFSET + l32i a3, a7, CP_REGS_TAB_SAVE + add a2, a2, a8 + callx0 a3 + /* + * Make sure coprocessor context and THREAD_CP_OWNER_CPU are in memory + * before updating THREAD_CPENABLE + */ + memw # (3) + s32i a10, a8, THREAD_CPENABLE +1: /* Get context save area and call load routine. */ l32i a2, a7, CP_REGS_TAB_OFFSET l32i a3, a7, CP_REGS_TAB_LOAD - add a2, a2, a4 + add a2, a2, a9 callx0 a3 /* Restore all registers and return from exception handler. */ + l32i a10, a1, PT_AREG10 + l32i a9, a1, PT_AREG9 + l32i a8, a1, PT_AREG8 l32i a7, a1, PT_AREG7 l32i a6, a1, PT_AREG6 l32i a5, a1, PT_AREG5 @@ -233,12 +299,4 @@ ENTRY(coprocessor_flush) ENDPROC(coprocessor_flush) - .data - -ENTRY(coprocessor_owner) - - .fill XCHAL_CP_MAX, 4, 0 - -END(coprocessor_owner) - #endif /* XTENSA_HAVE_COPROCESSORS */ diff --git a/arch/xtensa/kernel/entry.S b/arch/xtensa/kernel/entry.S index f2c789a5a92a..3255d4f61844 100644 --- a/arch/xtensa/kernel/entry.S +++ b/arch/xtensa/kernel/entry.S @@ -2071,8 +2071,16 @@ ENTRY(_switch_to) #if (XTENSA_HAVE_COPROCESSORS || XTENSA_HAVE_IO_PORTS) l32i a3, a5, THREAD_CPENABLE - xsr a3, cpenable - s32i a3, a4, THREAD_CPENABLE +#ifdef CONFIG_SMP + beqz a3, 1f + memw # pairs with memw (2) in fast_coprocessor + l32i a6, a5, THREAD_CP_OWNER_CPU + l32i a7, a5, THREAD_CPU + beq a6, a7, 1f # load 0 into CPENABLE if current CPU is not the owner + movi a3, 0 +1: +#endif + wsr a3, cpenable #endif #if XCHAL_HAVE_EXCLUSIVE diff --git a/arch/xtensa/kernel/process.c b/arch/xtensa/kernel/process.c index e8bfbca5f001..7e38292dd07a 100644 --- a/arch/xtensa/kernel/process.c +++ b/arch/xtensa/kernel/process.c @@ -47,6 +47,7 @@ #include <asm/asm-offsets.h> #include <asm/regs.h> #include <asm/hw_breakpoint.h> +#include <asm/traps.h> extern void ret_from_fork(void); extern void ret_from_kernel_thread(void); @@ -63,52 +64,114 @@ EXPORT_SYMBOL(__stack_chk_guard); #if XTENSA_HAVE_COPROCESSORS -void coprocessor_release_all(struct thread_info *ti) +void local_coprocessors_flush_release_all(void) { - unsigned long cpenable; - int i; + struct thread_info **coprocessor_owner; + struct thread_info *unique_owner[XCHAL_CP_MAX]; + int n = 0; + int i, j; - /* Make sure we don't switch tasks during this operation. */ + coprocessor_owner = this_cpu_ptr(&exc_table)->coprocessor_owner; + xtensa_set_sr(XCHAL_CP_MASK, cpenable); - preempt_disable(); + for (i = 0; i < XCHAL_CP_MAX; i++) { + struct thread_info *ti = coprocessor_owner[i]; - /* Walk through all cp owners and release it for the requested one. */ + if (ti) { + coprocessor_flush(ti, i); - cpenable = ti->cpenable; + for (j = 0; j < n; j++) + if (unique_owner[j] == ti) + break; + if (j == n) + unique_owner[n++] = ti; - for (i = 0; i < XCHAL_CP_MAX; i++) { - if (coprocessor_owner[i] == ti) { - coprocessor_owner[i] = 0; - cpenable &= ~(1 << i); + coprocessor_owner[i] = NULL; } } + for (i = 0; i < n; i++) { + /* pairs with memw (1) in fast_coprocessor and memw in switch_to */ + smp_wmb(); + unique_owner[i]->cpenable = 0; + } + xtensa_set_sr(0, cpenable); +} - ti->cpenable = cpenable; +static void local_coprocessor_release_all(void *info) +{ + struct thread_info *ti = info; + struct thread_info **coprocessor_owner; + int i; + + coprocessor_owner = this_cpu_ptr(&exc_table)->coprocessor_owner; + + /* Walk through all cp owners and release it for the requested one. */ + + for (i = 0; i < XCHAL_CP_MAX; i++) { + if (coprocessor_owner[i] == ti) + coprocessor_owner[i] = NULL; + } + /* pairs with memw (1) in fast_coprocessor and memw in switch_to */ + smp_wmb(); + ti->cpenable = 0; if (ti == current_thread_info()) xtensa_set_sr(0, cpenable); +} - preempt_enable(); +void coprocessor_release_all(struct thread_info *ti) +{ + if (ti->cpenable) { + /* pairs with memw (2) in fast_coprocessor */ + smp_rmb(); + smp_call_function_single(ti->cp_owner_cpu, + local_coprocessor_release_all, + ti, true); + } } -void coprocessor_flush_all(struct thread_info *ti) +static void local_coprocessor_flush_all(void *info) { - unsigned long cpenable, old_cpenable; + struct thread_info *ti = info; + struct thread_info **coprocessor_owner; + unsigned long old_cpenable; int i; - preempt_disable(); - - old_cpenable = xtensa_get_sr(cpenable); - cpenable = ti->cpenable; - xtensa_set_sr(cpenable, cpenable); + coprocessor_owner = this_cpu_ptr(&exc_table)->coprocessor_owner; + old_cpenable = xtensa_xsr(ti->cpenable, cpenable); for (i = 0; i < XCHAL_CP_MAX; i++) { - if ((cpenable & 1) != 0 && coprocessor_owner[i] == ti) + if (coprocessor_owner[i] == ti) coprocessor_flush(ti, i); - cpenable >>= 1; } xtensa_set_sr(old_cpenable, cpenable); +} - preempt_enable(); +void coprocessor_flush_all(struct thread_info *ti) +{ + if (ti->cpenable) { + /* pairs with memw (2) in fast_coprocessor */ + smp_rmb(); + smp_call_function_single(ti->cp_owner_cpu, + local_coprocessor_flush_all, + ti, true); + } +} + +static void local_coprocessor_flush_release_all(void *info) +{ + local_coprocessor_flush_all(info); + local_coprocessor_release_all(info); +} + +void coprocessor_flush_release_all(struct thread_info *ti) +{ + if (ti->cpenable) { + /* pairs with memw (2) in fast_coprocessor */ + smp_rmb(); + smp_call_function_single(ti->cp_owner_cpu, + local_coprocessor_flush_release_all, + ti, true); + } } #endif @@ -140,8 +203,7 @@ void flush_thread(void) { #if XTENSA_HAVE_COPROCESSORS struct thread_info *ti = current_thread_info(); - coprocessor_flush_all(ti); - coprocessor_release_all(ti); + coprocessor_flush_release_all(ti); #endif flush_ptrace_hw_breakpoint(current); } diff --git a/arch/xtensa/kernel/ptrace.c b/arch/xtensa/kernel/ptrace.c index 323c678a691f..22cdaa6729d3 100644 --- a/arch/xtensa/kernel/ptrace.c +++ b/arch/xtensa/kernel/ptrace.c @@ -171,8 +171,7 @@ static int tie_set(struct task_struct *target, #if XTENSA_HAVE_COPROCESSORS /* Flush all coprocessors before we overwrite them. */ - coprocessor_flush_all(ti); - coprocessor_release_all(ti); + coprocessor_flush_release_all(ti); ti->xtregs_cp.cp0 = newregs->cp0; ti->xtregs_cp.cp1 = newregs->cp1; ti->xtregs_cp.cp2 = newregs->cp2; diff --git a/arch/xtensa/kernel/signal.c b/arch/xtensa/kernel/signal.c index 6f68649e86ba..c9ffd42db873 100644 --- a/arch/xtensa/kernel/signal.c +++ b/arch/xtensa/kernel/signal.c @@ -162,8 +162,7 @@ setup_sigcontext(struct rt_sigframe __user *frame, struct pt_regs *regs) return err; #if XTENSA_HAVE_COPROCESSORS - coprocessor_flush_all(ti); - coprocessor_release_all(ti); + coprocessor_flush_release_all(ti); err |= __copy_to_user(&frame->xtregs.cp, &ti->xtregs_cp, sizeof (frame->xtregs.cp)); #endif diff --git a/arch/xtensa/kernel/smp.c b/arch/xtensa/kernel/smp.c index 1254da07ead1..4dc109dd6214 100644 --- a/arch/xtensa/kernel/smp.c +++ b/arch/xtensa/kernel/smp.c @@ -30,6 +30,7 @@ #include <linux/thread_info.h> #include <asm/cacheflush.h> +#include <asm/coprocessor.h> #include <asm/kdebug.h> #include <asm/mmu_context.h> #include <asm/mxregs.h> @@ -272,6 +273,12 @@ int __cpu_disable(void) */ set_cpu_online(cpu, false); +#if XTENSA_HAVE_COPROCESSORS + /* + * Flush coprocessor contexts that are active on the current CPU. + */ + local_coprocessors_flush_release_all(); +#endif /* * OK - migrate IRQs away from this CPU */ diff --git a/arch/xtensa/kernel/traps.c b/arch/xtensa/kernel/traps.c index 62c497605128..138a86fbe9d7 100644 --- a/arch/xtensa/kernel/traps.c +++ b/arch/xtensa/kernel/traps.c @@ -57,6 +57,9 @@ static void do_nmi(struct pt_regs *regs); static void do_unaligned_user(struct pt_regs *regs); #endif static void do_multihit(struct pt_regs *regs); +#if XTENSA_HAVE_COPROCESSORS +static void do_coprocessor(struct pt_regs *regs); +#endif static void do_debug(struct pt_regs *regs); /* @@ -69,7 +72,8 @@ static void do_debug(struct pt_regs *regs); #define USER 0x02 #define COPROCESSOR(x) \ -{ EXCCAUSE_COPROCESSOR ## x ## _DISABLED, USER|KRNL, fast_coprocessor } +{ EXCCAUSE_COPROCESSOR ## x ## _DISABLED, USER|KRNL, fast_coprocessor },\ +{ EXCCAUSE_COPROCESSOR ## x ## _DISABLED, 0, do_coprocessor } typedef struct { int cause; @@ -327,6 +331,13 @@ static void do_unaligned_user(struct pt_regs *regs) } #endif +#if XTENSA_HAVE_COPROCESSORS +static void do_coprocessor(struct pt_regs *regs) +{ + coprocessor_flush_release_all(current_thread_info()); +} +#endif + /* Handle debug events. * When CONFIG_HAVE_HW_BREAKPOINT is on this handler is called with * preemption disabled to avoid rescheduling and keep mapping of hardware |