1 #ifdef CONFIG_CPU_SUP_INTEL
3 /* The maximal number of PEBS events: */
4 #define MAX_PEBS_EVENTS 4
6 /* The size of a BTS record in bytes: */
7 #define BTS_RECORD_SIZE 24
9 #define BTS_BUFFER_SIZE (PAGE_SIZE << 4)
10 #define PEBS_BUFFER_SIZE PAGE_SIZE
13 * pebs_record_32 for p4 and core not supported
15 struct pebs_record_32 {
23 struct pebs_record_core {
28 u64 r12, r13, r14, r15;
31 struct pebs_record_nhm {
36 u64 r12, r13, r14, r15;
37 u64 status, dla, dse, lat;
41 * A debug store configuration.
43 * We only support architectures that use 64bit fields.
48 u64 bts_absolute_maximum;
49 u64 bts_interrupt_threshold;
52 u64 pebs_absolute_maximum;
53 u64 pebs_interrupt_threshold;
54 u64 pebs_event_reset[MAX_PEBS_EVENTS];
57 static void init_debug_store_on_cpu(int cpu)
59 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
64 wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
65 (u32)((u64)(unsigned long)ds),
66 (u32)((u64)(unsigned long)ds >> 32));
69 static void fini_debug_store_on_cpu(int cpu)
71 if (!per_cpu(cpu_hw_events, cpu).ds)
74 wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
77 static int alloc_pebs_buffer(int cpu)
79 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
80 int node = cpu_to_node(cpu);
81 int max, thresh = 1; /* always use a single PEBS record */
87 buffer = kmalloc_node(PEBS_BUFFER_SIZE, GFP_KERNEL | __GFP_ZERO, node);
88 if (unlikely(!buffer))
91 max = PEBS_BUFFER_SIZE / x86_pmu.pebs_record_size;
93 ds->pebs_buffer_base = (u64)(unsigned long)buffer;
94 ds->pebs_index = ds->pebs_buffer_base;
95 ds->pebs_absolute_maximum = ds->pebs_buffer_base +
96 max * x86_pmu.pebs_record_size;
98 ds->pebs_interrupt_threshold = ds->pebs_buffer_base +
99 thresh * x86_pmu.pebs_record_size;
104 static void release_pebs_buffer(int cpu)
106 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
108 if (!ds || !x86_pmu.pebs)
111 kfree((void *)(unsigned long)ds->pebs_buffer_base);
112 ds->pebs_buffer_base = 0;
115 static int alloc_bts_buffer(int cpu)
117 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
118 int node = cpu_to_node(cpu);
125 buffer = kmalloc_node(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_ZERO, node);
126 if (unlikely(!buffer))
129 max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
132 ds->bts_buffer_base = (u64)(unsigned long)buffer;
133 ds->bts_index = ds->bts_buffer_base;
134 ds->bts_absolute_maximum = ds->bts_buffer_base +
135 max * BTS_RECORD_SIZE;
136 ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
137 thresh * BTS_RECORD_SIZE;
142 static void release_bts_buffer(int cpu)
144 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
146 if (!ds || !x86_pmu.bts)
149 kfree((void *)(unsigned long)ds->bts_buffer_base);
150 ds->bts_buffer_base = 0;
153 static int alloc_ds_buffer(int cpu)
155 int node = cpu_to_node(cpu);
156 struct debug_store *ds;
158 ds = kmalloc_node(sizeof(*ds), GFP_KERNEL | __GFP_ZERO, node);
162 per_cpu(cpu_hw_events, cpu).ds = ds;
167 static void release_ds_buffer(int cpu)
169 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
174 per_cpu(cpu_hw_events, cpu).ds = NULL;
178 static void release_ds_buffers(void)
182 if (!x86_pmu.bts && !x86_pmu.pebs)
186 for_each_online_cpu(cpu)
187 fini_debug_store_on_cpu(cpu);
189 for_each_possible_cpu(cpu) {
190 release_pebs_buffer(cpu);
191 release_bts_buffer(cpu);
192 release_ds_buffer(cpu);
197 static void reserve_ds_buffers(void)
199 int bts_err = 0, pebs_err = 0;
202 x86_pmu.bts_active = 0;
203 x86_pmu.pebs_active = 0;
205 if (!x86_pmu.bts && !x86_pmu.pebs)
216 for_each_possible_cpu(cpu) {
217 if (alloc_ds_buffer(cpu)) {
222 if (!bts_err && alloc_bts_buffer(cpu))
225 if (!pebs_err && alloc_pebs_buffer(cpu))
228 if (bts_err && pebs_err)
233 for_each_possible_cpu(cpu)
234 release_bts_buffer(cpu);
238 for_each_possible_cpu(cpu)
239 release_pebs_buffer(cpu);
242 if (bts_err && pebs_err) {
243 for_each_possible_cpu(cpu)
244 release_ds_buffer(cpu);
246 if (x86_pmu.bts && !bts_err)
247 x86_pmu.bts_active = 1;
249 if (x86_pmu.pebs && !pebs_err)
250 x86_pmu.pebs_active = 1;
252 for_each_online_cpu(cpu)
253 init_debug_store_on_cpu(cpu);
263 static struct event_constraint bts_constraint =
264 EVENT_CONSTRAINT(0, 1ULL << X86_PMC_IDX_FIXED_BTS, 0);
266 static void intel_pmu_enable_bts(u64 config)
268 unsigned long debugctlmsr;
270 debugctlmsr = get_debugctlmsr();
272 debugctlmsr |= DEBUGCTLMSR_TR;
273 debugctlmsr |= DEBUGCTLMSR_BTS;
274 debugctlmsr |= DEBUGCTLMSR_BTINT;
276 if (!(config & ARCH_PERFMON_EVENTSEL_OS))
277 debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
279 if (!(config & ARCH_PERFMON_EVENTSEL_USR))
280 debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
282 update_debugctlmsr(debugctlmsr);
285 static void intel_pmu_disable_bts(void)
287 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
288 unsigned long debugctlmsr;
293 debugctlmsr = get_debugctlmsr();
296 ~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
297 DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
299 update_debugctlmsr(debugctlmsr);
302 static int intel_pmu_drain_bts_buffer(void)
304 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
305 struct debug_store *ds = cpuc->ds;
311 struct perf_event *event = cpuc->events[X86_PMC_IDX_FIXED_BTS];
312 struct bts_record *at, *top;
313 struct perf_output_handle handle;
314 struct perf_event_header header;
315 struct perf_sample_data data;
321 if (!x86_pmu.bts_active)
324 at = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
325 top = (struct bts_record *)(unsigned long)ds->bts_index;
330 ds->bts_index = ds->bts_buffer_base;
332 perf_sample_data_init(&data, 0);
333 data.period = event->hw.last_period;
337 * Prepare a generic sample, i.e. fill in the invariant fields.
338 * We will overwrite the from and to address before we output
341 perf_prepare_sample(&header, &data, event, ®s);
343 if (perf_output_begin(&handle, event, header.size * (top - at)))
346 for (; at < top; at++) {
350 perf_output_sample(&handle, &header, &data, event);
353 perf_output_end(&handle);
355 /* There's new data available. */
356 event->hw.interrupts++;
357 event->pending_kill = POLL_IN;
364 static struct event_constraint intel_core2_pebs_event_constraints[] = {
365 INTEL_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
366 INTEL_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
367 INTEL_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
368 INTEL_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
369 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED.* */
373 static struct event_constraint intel_atom_pebs_event_constraints[] = {
374 INTEL_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
375 INTEL_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
376 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED.* */
380 static struct event_constraint intel_nehalem_pebs_event_constraints[] = {
381 INTEL_EVENT_CONSTRAINT(0x0b, 0xf), /* MEM_INST_RETIRED.* */
382 INTEL_EVENT_CONSTRAINT(0x0f, 0xf), /* MEM_UNCORE_RETIRED.* */
383 INTEL_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
384 INTEL_EVENT_CONSTRAINT(0xc0, 0xf), /* INST_RETIRED.ANY */
385 INTEL_EVENT_CONSTRAINT(0xc2, 0xf), /* UOPS_RETIRED.* */
386 INTEL_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
387 INTEL_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
388 INTEL_EVENT_CONSTRAINT(0xc7, 0xf), /* SSEX_UOPS_RETIRED.* */
389 INTEL_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
390 INTEL_EVENT_CONSTRAINT(0xcb, 0xf), /* MEM_LOAD_RETIRED.* */
391 INTEL_EVENT_CONSTRAINT(0xf7, 0xf), /* FP_ASSIST.* */
395 static struct event_constraint intel_westmere_pebs_event_constraints[] = {
396 INTEL_EVENT_CONSTRAINT(0x0b, 0xf), /* MEM_INST_RETIRED.* */
397 INTEL_EVENT_CONSTRAINT(0x0f, 0xf), /* MEM_UNCORE_RETIRED.* */
398 INTEL_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
399 INTEL_EVENT_CONSTRAINT(0xc0, 0xf), /* INSTR_RETIRED.* */
400 INTEL_EVENT_CONSTRAINT(0xc2, 0xf), /* UOPS_RETIRED.* */
401 INTEL_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
402 INTEL_EVENT_CONSTRAINT(0xc5, 0xf), /* BR_MISP_RETIRED.* */
403 INTEL_EVENT_CONSTRAINT(0xc7, 0xf), /* SSEX_UOPS_RETIRED.* */
404 INTEL_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
405 INTEL_EVENT_CONSTRAINT(0xcb, 0xf), /* MEM_LOAD_RETIRED.* */
406 INTEL_EVENT_CONSTRAINT(0xf7, 0xf), /* FP_ASSIST.* */
410 static struct event_constraint intel_snb_pebs_events[] = {
411 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
412 INTEL_UEVENT_CONSTRAINT(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
413 INTEL_UEVENT_CONSTRAINT(0x02c2, 0xf), /* UOPS_RETIRED.RETIRE_SLOTS */
414 INTEL_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
415 INTEL_EVENT_CONSTRAINT(0xc5, 0xf), /* BR_MISP_RETIRED.* */
416 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.* */
417 INTEL_UEVENT_CONSTRAINT(0x11d0, 0xf), /* MEM_UOP_RETIRED.STLB_MISS_LOADS */
418 INTEL_UEVENT_CONSTRAINT(0x12d0, 0xf), /* MEM_UOP_RETIRED.STLB_MISS_STORES */
419 INTEL_UEVENT_CONSTRAINT(0x21d0, 0xf), /* MEM_UOP_RETIRED.LOCK_LOADS */
420 INTEL_UEVENT_CONSTRAINT(0x22d0, 0xf), /* MEM_UOP_RETIRED.LOCK_STORES */
421 INTEL_UEVENT_CONSTRAINT(0x41d0, 0xf), /* MEM_UOP_RETIRED.SPLIT_LOADS */
422 INTEL_UEVENT_CONSTRAINT(0x42d0, 0xf), /* MEM_UOP_RETIRED.SPLIT_STORES */
423 INTEL_UEVENT_CONSTRAINT(0x81d0, 0xf), /* MEM_UOP_RETIRED.ANY_LOADS */
424 INTEL_UEVENT_CONSTRAINT(0x82d0, 0xf), /* MEM_UOP_RETIRED.ANY_STORES */
425 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
426 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
427 INTEL_UEVENT_CONSTRAINT(0x02d4, 0xf), /* MEM_LOAD_UOPS_MISC_RETIRED.LLC_MISS */
431 static struct event_constraint *
432 intel_pebs_constraints(struct perf_event *event)
434 struct event_constraint *c;
436 if (!event->attr.precise_ip)
439 if (x86_pmu.pebs_constraints) {
440 for_each_event_constraint(c, x86_pmu.pebs_constraints) {
441 if ((event->hw.config & c->cmask) == c->code)
446 return &emptyconstraint;
449 static void intel_pmu_pebs_enable(struct perf_event *event)
451 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
452 struct hw_perf_event *hwc = &event->hw;
454 hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
456 cpuc->pebs_enabled |= 1ULL << hwc->idx;
457 WARN_ON_ONCE(cpuc->enabled);
459 if (x86_pmu.intel_cap.pebs_trap && event->attr.precise_ip > 1)
460 intel_pmu_lbr_enable(event);
463 static void intel_pmu_pebs_disable(struct perf_event *event)
465 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
466 struct hw_perf_event *hwc = &event->hw;
468 cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
470 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
472 hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
474 if (x86_pmu.intel_cap.pebs_trap && event->attr.precise_ip > 1)
475 intel_pmu_lbr_disable(event);
478 static void intel_pmu_pebs_enable_all(void)
480 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
482 if (cpuc->pebs_enabled)
483 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
486 static void intel_pmu_pebs_disable_all(void)
488 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
490 if (cpuc->pebs_enabled)
491 wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
494 #include <asm/insn.h>
496 static inline bool kernel_ip(unsigned long ip)
499 return ip > PAGE_OFFSET;
505 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
507 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
508 unsigned long from = cpuc->lbr_entries[0].from;
509 unsigned long old_to, to = cpuc->lbr_entries[0].to;
510 unsigned long ip = regs->ip;
513 * We don't need to fixup if the PEBS assist is fault like
515 if (!x86_pmu.intel_cap.pebs_trap)
519 * No LBR entry, no basic block, no rewinding
521 if (!cpuc->lbr_stack.nr || !from || !to)
525 * Basic blocks should never cross user/kernel boundaries
527 if (kernel_ip(ip) != kernel_ip(to))
531 * unsigned math, either ip is before the start (impossible) or
532 * the basic block is larger than 1 page (sanity)
534 if ((ip - to) > PAGE_SIZE)
538 * We sampled a branch insn, rewind using the LBR stack
547 u8 buf[MAX_INSN_SIZE];
551 if (!kernel_ip(ip)) {
552 int bytes, size = MAX_INSN_SIZE;
554 bytes = copy_from_user_nmi(buf, (void __user *)to, size);
562 kernel_insn_init(&insn, kaddr);
563 insn_get_length(&insn);
573 * Even though we decoded the basic block, the instruction stream
574 * never matched the given IP, either the TO or the IP got corrupted.
579 static int intel_pmu_save_and_restart(struct perf_event *event);
581 static void __intel_pmu_pebs_event(struct perf_event *event,
582 struct pt_regs *iregs, void *__pebs)
585 * We cast to pebs_record_core since that is a subset of
586 * both formats and we don't use the other fields in this
589 struct pebs_record_core *pebs = __pebs;
590 struct perf_sample_data data;
593 if (!intel_pmu_save_and_restart(event))
596 perf_sample_data_init(&data, 0);
597 data.period = event->hw.last_period;
600 * We use the interrupt regs as a base because the PEBS record
601 * does not contain a full regs set, specifically it seems to
602 * lack segment descriptors, which get used by things like
605 * In the simple case fix up only the IP and BP,SP regs, for
606 * PERF_SAMPLE_IP and PERF_SAMPLE_CALLCHAIN to function properly.
607 * A possible PERF_SAMPLE_REGS will have to transfer all regs.
614 if (event->attr.precise_ip > 1 && intel_pmu_pebs_fixup_ip(®s))
615 regs.flags |= PERF_EFLAGS_EXACT;
617 regs.flags &= ~PERF_EFLAGS_EXACT;
619 if (perf_event_overflow(event, &data, ®s))
620 x86_pmu_stop(event, 0);
623 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
625 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
626 struct debug_store *ds = cpuc->ds;
627 struct perf_event *event = cpuc->events[0]; /* PMC0 only */
628 struct pebs_record_core *at, *top;
631 if (!x86_pmu.pebs_active)
634 at = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
635 top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
638 * Whatever else happens, drain the thing
640 ds->pebs_index = ds->pebs_buffer_base;
642 if (!test_bit(0, cpuc->active_mask))
645 WARN_ON_ONCE(!event);
647 if (!event->attr.precise_ip)
655 * Should not happen, we program the threshold at 1 and do not
661 __intel_pmu_pebs_event(event, iregs, at);
664 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
666 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
667 struct debug_store *ds = cpuc->ds;
668 struct pebs_record_nhm *at, *top;
669 struct perf_event *event = NULL;
673 if (!x86_pmu.pebs_active)
676 at = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
677 top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
679 ds->pebs_index = ds->pebs_buffer_base;
686 * Should not happen, we program the threshold at 1 and do not
689 WARN_ON_ONCE(n > MAX_PEBS_EVENTS);
691 for ( ; at < top; at++) {
692 for_each_set_bit(bit, (unsigned long *)&at->status, MAX_PEBS_EVENTS) {
693 event = cpuc->events[bit];
694 if (!test_bit(bit, cpuc->active_mask))
697 WARN_ON_ONCE(!event);
699 if (!event->attr.precise_ip)
702 if (__test_and_set_bit(bit, (unsigned long *)&status))
708 if (!event || bit >= MAX_PEBS_EVENTS)
711 __intel_pmu_pebs_event(event, iregs, at);
716 * BTS, PEBS probe and setup
719 static void intel_ds_init(void)
722 * No support for 32bit formats
724 if (!boot_cpu_has(X86_FEATURE_DTES64))
727 x86_pmu.bts = boot_cpu_has(X86_FEATURE_BTS);
728 x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
730 char pebs_type = x86_pmu.intel_cap.pebs_trap ? '+' : '-';
731 int format = x86_pmu.intel_cap.pebs_format;
735 printk(KERN_CONT "PEBS fmt0%c, ", pebs_type);
736 x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
737 x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
741 printk(KERN_CONT "PEBS fmt1%c, ", pebs_type);
742 x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
743 x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
747 printk(KERN_CONT "no PEBS fmt%d%c, ", format, pebs_type);
753 #else /* CONFIG_CPU_SUP_INTEL */
755 static void reserve_ds_buffers(void)
759 static void release_ds_buffers(void)
763 #endif /* CONFIG_CPU_SUP_INTEL */