2 * Performance counter x86 architecture code
4 * Copyright(C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2008 Red Hat, Inc., Ingo Molnar
6 * Copyright(C) 2009 Jaswinder Singh Rajput
7 * Copyright(C) 2009 Advanced Micro Devices, Inc., Robert Richter
9 * For licencing details see kernel-base/COPYING
12 #include <linux/perf_counter.h>
13 #include <linux/capability.h>
14 #include <linux/notifier.h>
15 #include <linux/hardirq.h>
16 #include <linux/kprobes.h>
17 #include <linux/module.h>
18 #include <linux/kdebug.h>
19 #include <linux/sched.h>
20 #include <linux/uaccess.h>
23 #include <asm/stacktrace.h>
26 static bool perf_counters_initialized __read_mostly;
27 static u64 perf_counter_mask __read_mostly;
29 struct cpu_hw_counters {
30 struct perf_counter *counters[X86_PMC_IDX_MAX];
31 unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
32 unsigned long active[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
33 unsigned long interrupts;
39 * struct x86_pmu - generic x86 pmu
44 int (*handle_irq)(struct pt_regs *, int);
45 u64 (*save_disable_all)(void);
46 void (*restore_all)(u64);
47 void (*enable)(struct hw_perf_counter *, int);
48 void (*disable)(struct hw_perf_counter *, int);
51 u64 (*event_map)(int);
52 u64 (*raw_event)(u64);
55 int num_counters_fixed;
60 static struct x86_pmu x86_pmu __read_mostly;
62 static DEFINE_PER_CPU(struct cpu_hw_counters, cpu_hw_counters) = {
67 * Intel PerfMon v3. Used on Core2 and later.
69 static const u64 intel_perfmon_event_map[] =
71 [PERF_COUNT_CPU_CYCLES] = 0x003c,
72 [PERF_COUNT_INSTRUCTIONS] = 0x00c0,
73 [PERF_COUNT_CACHE_REFERENCES] = 0x4f2e,
74 [PERF_COUNT_CACHE_MISSES] = 0x412e,
75 [PERF_COUNT_BRANCH_INSTRUCTIONS] = 0x00c4,
76 [PERF_COUNT_BRANCH_MISSES] = 0x00c5,
77 [PERF_COUNT_BUS_CYCLES] = 0x013c,
80 static u64 intel_pmu_event_map(int event)
82 return intel_perfmon_event_map[event];
85 static u64 intel_pmu_raw_event(u64 event)
87 #define CORE_EVNTSEL_EVENT_MASK 0x000000FFULL
88 #define CORE_EVNTSEL_UNIT_MASK 0x0000FF00ULL
89 #define CORE_EVNTSEL_COUNTER_MASK 0xFF000000ULL
91 #define CORE_EVNTSEL_MASK \
92 (CORE_EVNTSEL_EVENT_MASK | \
93 CORE_EVNTSEL_UNIT_MASK | \
94 CORE_EVNTSEL_COUNTER_MASK)
96 return event & CORE_EVNTSEL_MASK;
100 * AMD Performance Monitor K7 and later.
102 static const u64 amd_perfmon_event_map[] =
104 [PERF_COUNT_CPU_CYCLES] = 0x0076,
105 [PERF_COUNT_INSTRUCTIONS] = 0x00c0,
106 [PERF_COUNT_CACHE_REFERENCES] = 0x0080,
107 [PERF_COUNT_CACHE_MISSES] = 0x0081,
108 [PERF_COUNT_BRANCH_INSTRUCTIONS] = 0x00c4,
109 [PERF_COUNT_BRANCH_MISSES] = 0x00c5,
112 static u64 amd_pmu_event_map(int event)
114 return amd_perfmon_event_map[event];
117 static u64 amd_pmu_raw_event(u64 event)
119 #define K7_EVNTSEL_EVENT_MASK 0x7000000FFULL
120 #define K7_EVNTSEL_UNIT_MASK 0x00000FF00ULL
121 #define K7_EVNTSEL_COUNTER_MASK 0x0FF000000ULL
123 #define K7_EVNTSEL_MASK \
124 (K7_EVNTSEL_EVENT_MASK | \
125 K7_EVNTSEL_UNIT_MASK | \
126 K7_EVNTSEL_COUNTER_MASK)
128 return event & K7_EVNTSEL_MASK;
132 * Propagate counter elapsed time into the generic counter.
133 * Can only be executed on the CPU where the counter is active.
134 * Returns the delta events processed.
137 x86_perf_counter_update(struct perf_counter *counter,
138 struct hw_perf_counter *hwc, int idx)
140 u64 prev_raw_count, new_raw_count, delta;
143 * Careful: an NMI might modify the previous counter value.
145 * Our tactic to handle this is to first atomically read and
146 * exchange a new raw count - then add that new-prev delta
147 * count to the generic counter atomically:
150 prev_raw_count = atomic64_read(&hwc->prev_count);
151 rdmsrl(hwc->counter_base + idx, new_raw_count);
153 if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count,
154 new_raw_count) != prev_raw_count)
158 * Now we have the new raw value and have updated the prev
159 * timestamp already. We can now calculate the elapsed delta
160 * (counter-)time and add that to the generic counter.
162 * Careful, not all hw sign-extends above the physical width
163 * of the count, so we do that by clipping the delta to 32 bits:
165 delta = (u64)(u32)((s32)new_raw_count - (s32)prev_raw_count);
167 atomic64_add(delta, &counter->count);
168 atomic64_sub(delta, &hwc->period_left);
171 static atomic_t num_counters;
172 static DEFINE_MUTEX(pmc_reserve_mutex);
174 static bool reserve_pmc_hardware(void)
178 if (nmi_watchdog == NMI_LOCAL_APIC)
179 disable_lapic_nmi_watchdog();
181 for (i = 0; i < x86_pmu.num_counters; i++) {
182 if (!reserve_perfctr_nmi(x86_pmu.perfctr + i))
186 for (i = 0; i < x86_pmu.num_counters; i++) {
187 if (!reserve_evntsel_nmi(x86_pmu.eventsel + i))
194 for (i--; i >= 0; i--)
195 release_evntsel_nmi(x86_pmu.eventsel + i);
197 i = x86_pmu.num_counters;
200 for (i--; i >= 0; i--)
201 release_perfctr_nmi(x86_pmu.perfctr + i);
203 if (nmi_watchdog == NMI_LOCAL_APIC)
204 enable_lapic_nmi_watchdog();
209 static void release_pmc_hardware(void)
213 for (i = 0; i < x86_pmu.num_counters; i++) {
214 release_perfctr_nmi(x86_pmu.perfctr + i);
215 release_evntsel_nmi(x86_pmu.eventsel + i);
218 if (nmi_watchdog == NMI_LOCAL_APIC)
219 enable_lapic_nmi_watchdog();
222 static void hw_perf_counter_destroy(struct perf_counter *counter)
224 if (atomic_dec_and_mutex_lock(&num_counters, &pmc_reserve_mutex)) {
225 release_pmc_hardware();
226 mutex_unlock(&pmc_reserve_mutex);
231 * Setup the hardware configuration for a given hw_event_type
233 static int __hw_perf_counter_init(struct perf_counter *counter)
235 struct perf_counter_hw_event *hw_event = &counter->hw_event;
236 struct hw_perf_counter *hwc = &counter->hw;
239 /* disable temporarily */
240 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
243 if (unlikely(!perf_counters_initialized))
247 if (atomic_inc_not_zero(&num_counters)) {
248 mutex_lock(&pmc_reserve_mutex);
249 if (atomic_read(&num_counters) == 0 && !reserve_pmc_hardware())
252 atomic_inc(&num_counters);
253 mutex_unlock(&pmc_reserve_mutex);
260 * (keep 'enabled' bit clear for now)
262 hwc->config = ARCH_PERFMON_EVENTSEL_INT;
265 * Count user and OS events unless requested not to.
267 if (!hw_event->exclude_user)
268 hwc->config |= ARCH_PERFMON_EVENTSEL_USR;
269 if (!hw_event->exclude_kernel)
270 hwc->config |= ARCH_PERFMON_EVENTSEL_OS;
273 * If privileged enough, allow NMI events:
276 if (capable(CAP_SYS_ADMIN) && hw_event->nmi)
279 hwc->irq_period = hw_event->irq_period;
281 * Intel PMCs cannot be accessed sanely above 32 bit width,
282 * so we install an artificial 1<<31 period regardless of
283 * the generic counter period:
285 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
286 if ((s64)hwc->irq_period <= 0 || hwc->irq_period > 0x7FFFFFFF)
287 hwc->irq_period = 0x7FFFFFFF;
289 atomic64_set(&hwc->period_left, hwc->irq_period);
292 * Raw event type provide the config in the event structure
294 if (perf_event_raw(hw_event)) {
295 hwc->config |= x86_pmu.raw_event(perf_event_config(hw_event));
297 if (perf_event_id(hw_event) >= x86_pmu.max_events)
302 hwc->config |= x86_pmu.event_map(perf_event_id(hw_event));
305 counter->destroy = hw_perf_counter_destroy;
310 static u64 intel_pmu_save_disable_all(void)
314 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
315 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
320 static u64 amd_pmu_save_disable_all(void)
322 struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
325 enabled = cpuc->enabled;
328 * ensure we write the disable before we start disabling the
329 * counters proper, so that amd_pmu_enable_counter() does the
334 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
337 if (!test_bit(idx, cpuc->active))
339 rdmsrl(MSR_K7_EVNTSEL0 + idx, val);
340 if (!(val & ARCH_PERFMON_EVENTSEL0_ENABLE))
342 val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
343 wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
349 u64 hw_perf_save_disable(void)
351 if (unlikely(!perf_counters_initialized))
354 return x86_pmu.save_disable_all();
357 * Exported because of ACPI idle
359 EXPORT_SYMBOL_GPL(hw_perf_save_disable);
361 static void intel_pmu_restore_all(u64 ctrl)
363 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
366 static void amd_pmu_restore_all(u64 ctrl)
368 struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
371 cpuc->enabled = ctrl;
376 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
379 if (!test_bit(idx, cpuc->active))
381 rdmsrl(MSR_K7_EVNTSEL0 + idx, val);
382 if (val & ARCH_PERFMON_EVENTSEL0_ENABLE)
384 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
385 wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
389 void hw_perf_restore(u64 ctrl)
391 if (unlikely(!perf_counters_initialized))
394 x86_pmu.restore_all(ctrl);
397 * Exported because of ACPI idle
399 EXPORT_SYMBOL_GPL(hw_perf_restore);
401 static inline u64 intel_pmu_get_status(u64 mask)
405 if (unlikely(!perf_counters_initialized))
407 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
412 static inline void intel_pmu_ack_status(u64 ack)
414 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
417 static inline void x86_pmu_enable_counter(struct hw_perf_counter *hwc, int idx)
421 if (unlikely(!perf_counters_initialized))
424 err = checking_wrmsrl(hwc->config_base + idx,
425 hwc->config | ARCH_PERFMON_EVENTSEL0_ENABLE);
428 static inline void x86_pmu_disable_counter(struct hw_perf_counter *hwc, int idx)
432 if (unlikely(!perf_counters_initialized))
435 err = checking_wrmsrl(hwc->config_base + idx,
440 intel_pmu_disable_fixed(struct hw_perf_counter *hwc, int __idx)
442 int idx = __idx - X86_PMC_IDX_FIXED;
446 mask = 0xfULL << (idx * 4);
448 rdmsrl(hwc->config_base, ctrl_val);
450 err = checking_wrmsrl(hwc->config_base, ctrl_val);
454 intel_pmu_disable_counter(struct hw_perf_counter *hwc, int idx)
456 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
457 intel_pmu_disable_fixed(hwc, idx);
461 x86_pmu_disable_counter(hwc, idx);
465 amd_pmu_disable_counter(struct hw_perf_counter *hwc, int idx)
467 x86_pmu_disable_counter(hwc, idx);
470 static DEFINE_PER_CPU(u64, prev_left[X86_PMC_IDX_MAX]);
473 * Set the next IRQ period, based on the hwc->period_left value.
474 * To be called with the counter disabled in hw:
477 x86_perf_counter_set_period(struct perf_counter *counter,
478 struct hw_perf_counter *hwc, int idx)
480 s64 left = atomic64_read(&hwc->period_left);
481 s64 period = hwc->irq_period;
485 * If we are way outside a reasoable range then just skip forward:
487 if (unlikely(left <= -period)) {
489 atomic64_set(&hwc->period_left, left);
492 if (unlikely(left <= 0)) {
494 atomic64_set(&hwc->period_left, left);
497 per_cpu(prev_left[idx], smp_processor_id()) = left;
500 * The hw counter starts counting from this counter offset,
501 * mark it to be able to extra future deltas:
503 atomic64_set(&hwc->prev_count, (u64)-left);
505 err = checking_wrmsrl(hwc->counter_base + idx,
506 (u64)(-left) & x86_pmu.counter_mask);
510 intel_pmu_enable_fixed(struct hw_perf_counter *hwc, int __idx)
512 int idx = __idx - X86_PMC_IDX_FIXED;
513 u64 ctrl_val, bits, mask;
517 * Enable IRQ generation (0x8),
518 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
522 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
524 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
527 mask = 0xfULL << (idx * 4);
529 rdmsrl(hwc->config_base, ctrl_val);
532 err = checking_wrmsrl(hwc->config_base, ctrl_val);
535 static void intel_pmu_enable_counter(struct hw_perf_counter *hwc, int idx)
537 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
538 intel_pmu_enable_fixed(hwc, idx);
542 x86_pmu_enable_counter(hwc, idx);
545 static void amd_pmu_enable_counter(struct hw_perf_counter *hwc, int idx)
547 struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
550 x86_pmu_enable_counter(hwc, idx);
552 x86_pmu_disable_counter(hwc, idx);
556 fixed_mode_idx(struct perf_counter *counter, struct hw_perf_counter *hwc)
560 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
563 if (unlikely(hwc->nmi))
566 event = hwc->config & ARCH_PERFMON_EVENT_MASK;
568 if (unlikely(event == x86_pmu.event_map(PERF_COUNT_INSTRUCTIONS)))
569 return X86_PMC_IDX_FIXED_INSTRUCTIONS;
570 if (unlikely(event == x86_pmu.event_map(PERF_COUNT_CPU_CYCLES)))
571 return X86_PMC_IDX_FIXED_CPU_CYCLES;
572 if (unlikely(event == x86_pmu.event_map(PERF_COUNT_BUS_CYCLES)))
573 return X86_PMC_IDX_FIXED_BUS_CYCLES;
579 * Find a PMC slot for the freshly enabled / scheduled in counter:
581 static int x86_pmu_enable(struct perf_counter *counter)
583 struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
584 struct hw_perf_counter *hwc = &counter->hw;
587 idx = fixed_mode_idx(counter, hwc);
590 * Try to get the fixed counter, if that is already taken
591 * then try to get a generic counter:
593 if (test_and_set_bit(idx, cpuc->used))
596 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
598 * We set it so that counter_base + idx in wrmsr/rdmsr maps to
599 * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
602 MSR_ARCH_PERFMON_FIXED_CTR0 - X86_PMC_IDX_FIXED;
606 /* Try to get the previous generic counter again */
607 if (test_and_set_bit(idx, cpuc->used)) {
609 idx = find_first_zero_bit(cpuc->used,
610 x86_pmu.num_counters);
611 if (idx == x86_pmu.num_counters)
614 set_bit(idx, cpuc->used);
617 hwc->config_base = x86_pmu.eventsel;
618 hwc->counter_base = x86_pmu.perfctr;
621 perf_counters_lapic_init(hwc->nmi);
623 x86_pmu.disable(hwc, idx);
625 cpuc->counters[idx] = counter;
626 set_bit(idx, cpuc->active);
628 x86_perf_counter_set_period(counter, hwc, idx);
629 x86_pmu.enable(hwc, idx);
634 void perf_counter_print_debug(void)
636 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
637 struct cpu_hw_counters *cpuc;
640 if (!x86_pmu.num_counters)
645 cpu = smp_processor_id();
646 cpuc = &per_cpu(cpu_hw_counters, cpu);
648 if (x86_pmu.version >= 2) {
649 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
650 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
651 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
652 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
655 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
656 pr_info("CPU#%d: status: %016llx\n", cpu, status);
657 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
658 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
660 pr_info("CPU#%d: used: %016llx\n", cpu, *(u64 *)cpuc->used);
662 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
663 rdmsrl(x86_pmu.eventsel + idx, pmc_ctrl);
664 rdmsrl(x86_pmu.perfctr + idx, pmc_count);
666 prev_left = per_cpu(prev_left[idx], cpu);
668 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
670 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
671 cpu, idx, pmc_count);
672 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
673 cpu, idx, prev_left);
675 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
676 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
678 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
679 cpu, idx, pmc_count);
684 static void x86_pmu_disable(struct perf_counter *counter)
686 struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
687 struct hw_perf_counter *hwc = &counter->hw;
691 * Must be done before we disable, otherwise the nmi handler
692 * could reenable again:
694 clear_bit(idx, cpuc->active);
695 x86_pmu.disable(hwc, idx);
698 * Make sure the cleared pointer becomes visible before we
699 * (potentially) free the counter:
704 * Drain the remaining delta count out of a counter
705 * that we are disabling:
707 x86_perf_counter_update(counter, hwc, idx);
708 cpuc->counters[idx] = NULL;
709 clear_bit(idx, cpuc->used);
713 * Save and restart an expired counter. Called by NMI contexts,
714 * so it has to be careful about preempting normal counter ops:
716 static void intel_pmu_save_and_restart(struct perf_counter *counter)
718 struct hw_perf_counter *hwc = &counter->hw;
721 x86_perf_counter_update(counter, hwc, idx);
722 x86_perf_counter_set_period(counter, hwc, idx);
724 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
725 intel_pmu_enable_counter(hwc, idx);
729 * Maximum interrupt frequency of 100KHz per CPU
731 #define PERFMON_MAX_INTERRUPTS (100000/HZ)
734 * This handler is triggered by the local APIC, so the APIC IRQ handling
737 static int intel_pmu_handle_irq(struct pt_regs *regs, int nmi)
739 int bit, cpu = smp_processor_id();
741 struct cpu_hw_counters *cpuc = &per_cpu(cpu_hw_counters, cpu);
744 cpuc->throttle_ctrl = intel_pmu_save_disable_all();
746 status = intel_pmu_get_status(cpuc->throttle_ctrl);
752 inc_irq_stat(apic_perf_irqs);
754 for_each_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
755 struct perf_counter *counter = cpuc->counters[bit];
757 clear_bit(bit, (unsigned long *) &status);
758 if (!test_bit(bit, cpuc->active))
761 intel_pmu_save_and_restart(counter);
762 if (perf_counter_overflow(counter, nmi, regs, 0))
763 intel_pmu_disable_counter(&counter->hw, bit);
766 intel_pmu_ack_status(ack);
769 * Repeat if there is more work to be done:
771 status = intel_pmu_get_status(cpuc->throttle_ctrl);
776 * Restore - do not reenable when global enable is off or throttled:
778 if (++cpuc->interrupts < PERFMON_MAX_INTERRUPTS)
779 intel_pmu_restore_all(cpuc->throttle_ctrl);
784 static int amd_pmu_handle_irq(struct pt_regs *regs, int nmi) { return 0; }
786 void perf_counter_unthrottle(void)
788 struct cpu_hw_counters *cpuc;
790 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
793 if (unlikely(!perf_counters_initialized))
796 cpuc = &__get_cpu_var(cpu_hw_counters);
797 if (cpuc->interrupts >= PERFMON_MAX_INTERRUPTS) {
798 if (printk_ratelimit())
799 printk(KERN_WARNING "PERFMON: max interrupts exceeded!\n");
800 hw_perf_restore(cpuc->throttle_ctrl);
802 cpuc->interrupts = 0;
805 void smp_perf_counter_interrupt(struct pt_regs *regs)
808 apic_write(APIC_LVTPC, LOCAL_PERF_VECTOR);
810 x86_pmu.handle_irq(regs, 0);
814 void smp_perf_pending_interrupt(struct pt_regs *regs)
818 inc_irq_stat(apic_pending_irqs);
819 perf_counter_do_pending();
823 void set_perf_counter_pending(void)
825 apic->send_IPI_self(LOCAL_PENDING_VECTOR);
828 void perf_counters_lapic_init(int nmi)
832 if (!perf_counters_initialized)
835 * Enable the performance counter vector in the APIC LVT:
837 apic_val = apic_read(APIC_LVTERR);
839 apic_write(APIC_LVTERR, apic_val | APIC_LVT_MASKED);
841 apic_write(APIC_LVTPC, APIC_DM_NMI);
843 apic_write(APIC_LVTPC, LOCAL_PERF_VECTOR);
844 apic_write(APIC_LVTERR, apic_val);
848 perf_counter_nmi_handler(struct notifier_block *self,
849 unsigned long cmd, void *__args)
851 struct die_args *args = __args;
852 struct pt_regs *regs;
866 apic_write(APIC_LVTPC, APIC_DM_NMI);
867 ret = x86_pmu.handle_irq(regs, 1);
869 return ret ? NOTIFY_STOP : NOTIFY_OK;
872 static __read_mostly struct notifier_block perf_counter_nmi_notifier = {
873 .notifier_call = perf_counter_nmi_handler,
878 static struct x86_pmu intel_pmu = {
880 .handle_irq = intel_pmu_handle_irq,
881 .save_disable_all = intel_pmu_save_disable_all,
882 .restore_all = intel_pmu_restore_all,
883 .enable = intel_pmu_enable_counter,
884 .disable = intel_pmu_disable_counter,
885 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
886 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
887 .event_map = intel_pmu_event_map,
888 .raw_event = intel_pmu_raw_event,
889 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
892 static struct x86_pmu amd_pmu = {
894 .handle_irq = amd_pmu_handle_irq,
895 .save_disable_all = amd_pmu_save_disable_all,
896 .restore_all = amd_pmu_restore_all,
897 .enable = amd_pmu_enable_counter,
898 .disable = amd_pmu_disable_counter,
899 .eventsel = MSR_K7_EVNTSEL0,
900 .perfctr = MSR_K7_PERFCTR0,
901 .event_map = amd_pmu_event_map,
902 .raw_event = amd_pmu_raw_event,
903 .max_events = ARRAY_SIZE(amd_perfmon_event_map),
906 .counter_mask = (1ULL << 48) - 1,
909 static int intel_pmu_init(void)
911 union cpuid10_edx edx;
912 union cpuid10_eax eax;
917 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
921 * Check whether the Architectural PerfMon supports
922 * Branch Misses Retired Event or not.
924 cpuid(10, &eax.full, &ebx, &unused, &edx.full);
925 if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
928 version = eax.split.version_id;
933 x86_pmu.version = version;
934 x86_pmu.num_counters = eax.split.num_counters;
935 x86_pmu.num_counters_fixed = edx.split.num_counters_fixed;
936 x86_pmu.counter_bits = eax.split.bit_width;
937 x86_pmu.counter_mask = (1ULL << eax.split.bit_width) - 1;
942 static int amd_pmu_init(void)
948 void __init init_hw_perf_counters(void)
952 switch (boot_cpu_data.x86_vendor) {
953 case X86_VENDOR_INTEL:
954 err = intel_pmu_init();
957 err = amd_pmu_init();
965 pr_info("%s Performance Monitoring support detected.\n", x86_pmu.name);
966 pr_info("... version: %d\n", x86_pmu.version);
967 pr_info("... bit width: %d\n", x86_pmu.counter_bits);
969 pr_info("... num counters: %d\n", x86_pmu.num_counters);
970 if (x86_pmu.num_counters > X86_PMC_MAX_GENERIC) {
971 x86_pmu.num_counters = X86_PMC_MAX_GENERIC;
972 WARN(1, KERN_ERR "hw perf counters %d > max(%d), clipping!",
973 x86_pmu.num_counters, X86_PMC_MAX_GENERIC);
975 perf_counter_mask = (1 << x86_pmu.num_counters) - 1;
976 perf_max_counters = x86_pmu.num_counters;
978 pr_info("... value mask: %016Lx\n", x86_pmu.counter_mask);
980 if (x86_pmu.num_counters_fixed > X86_PMC_MAX_FIXED) {
981 x86_pmu.num_counters_fixed = X86_PMC_MAX_FIXED;
982 WARN(1, KERN_ERR "hw perf counters fixed %d > max(%d), clipping!",
983 x86_pmu.num_counters_fixed, X86_PMC_MAX_FIXED);
985 pr_info("... fixed counters: %d\n", x86_pmu.num_counters_fixed);
988 ((1LL << x86_pmu.num_counters_fixed)-1) << X86_PMC_IDX_FIXED;
990 pr_info("... counter mask: %016Lx\n", perf_counter_mask);
991 perf_counters_initialized = true;
993 perf_counters_lapic_init(0);
994 register_die_notifier(&perf_counter_nmi_notifier);
997 static inline void x86_pmu_read(struct perf_counter *counter)
999 x86_perf_counter_update(counter, &counter->hw, counter->hw.idx);
1002 static const struct pmu pmu = {
1003 .enable = x86_pmu_enable,
1004 .disable = x86_pmu_disable,
1005 .read = x86_pmu_read,
1008 const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
1012 err = __hw_perf_counter_init(counter);
1014 return ERR_PTR(err);
1024 void callchain_store(struct perf_callchain_entry *entry, unsigned long ip)
1026 if (entry->nr < MAX_STACK_DEPTH)
1027 entry->ip[entry->nr++] = ip;
1030 static DEFINE_PER_CPU(struct perf_callchain_entry, irq_entry);
1031 static DEFINE_PER_CPU(struct perf_callchain_entry, nmi_entry);
1035 backtrace_warning_symbol(void *data, char *msg, unsigned long symbol)
1037 /* Ignore warnings */
1040 static void backtrace_warning(void *data, char *msg)
1042 /* Ignore warnings */
1045 static int backtrace_stack(void *data, char *name)
1047 /* Don't bother with IRQ stacks for now */
1051 static void backtrace_address(void *data, unsigned long addr, int reliable)
1053 struct perf_callchain_entry *entry = data;
1056 callchain_store(entry, addr);
1059 static const struct stacktrace_ops backtrace_ops = {
1060 .warning = backtrace_warning,
1061 .warning_symbol = backtrace_warning_symbol,
1062 .stack = backtrace_stack,
1063 .address = backtrace_address,
1067 perf_callchain_kernel(struct pt_regs *regs, struct perf_callchain_entry *entry)
1073 callchain_store(entry, instruction_pointer(regs));
1075 stack = ((char *)regs + sizeof(struct pt_regs));
1076 #ifdef CONFIG_FRAME_POINTER
1077 bp = frame_pointer(regs);
1082 dump_trace(NULL, regs, (void *)stack, bp, &backtrace_ops, entry);
1084 entry->kernel = entry->nr - nr;
1088 struct stack_frame {
1089 const void __user *next_fp;
1090 unsigned long return_address;
1093 static int copy_stack_frame(const void __user *fp, struct stack_frame *frame)
1097 if (!access_ok(VERIFY_READ, fp, sizeof(*frame)))
1101 pagefault_disable();
1102 if (__copy_from_user_inatomic(frame, fp, sizeof(*frame)))
1110 perf_callchain_user(struct pt_regs *regs, struct perf_callchain_entry *entry)
1112 struct stack_frame frame;
1113 const void __user *fp;
1116 regs = (struct pt_regs *)current->thread.sp0 - 1;
1117 fp = (void __user *)regs->bp;
1119 callchain_store(entry, regs->ip);
1121 while (entry->nr < MAX_STACK_DEPTH) {
1122 frame.next_fp = NULL;
1123 frame.return_address = 0;
1125 if (!copy_stack_frame(fp, &frame))
1128 if ((unsigned long)fp < user_stack_pointer(regs))
1131 callchain_store(entry, frame.return_address);
1135 entry->user = entry->nr - nr;
1139 perf_do_callchain(struct pt_regs *regs, struct perf_callchain_entry *entry)
1146 is_user = user_mode(regs);
1148 if (!current || current->pid == 0)
1151 if (is_user && current->state != TASK_RUNNING)
1155 perf_callchain_kernel(regs, entry);
1158 perf_callchain_user(regs, entry);
1161 struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1163 struct perf_callchain_entry *entry;
1166 entry = &__get_cpu_var(nmi_entry);
1168 entry = &__get_cpu_var(irq_entry);
1175 perf_do_callchain(regs, entry);