2 * Performance counter x86 architecture code
4 * Copyright(C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2008 Red Hat, Inc., Ingo Molnar
7 * For licencing details see kernel-base/COPYING
10 #include <linux/perf_counter.h>
11 #include <linux/capability.h>
12 #include <linux/notifier.h>
13 #include <linux/hardirq.h>
14 #include <linux/kprobes.h>
15 #include <linux/module.h>
16 #include <linux/kdebug.h>
17 #include <linux/sched.h>
19 #include <asm/perf_counter.h>
22 static bool perf_counters_initialized __read_mostly;
25 * Number of (generic) HW counters:
27 static int nr_counters_generic __read_mostly;
28 static u64 perf_counter_mask __read_mostly;
29 static u64 counter_value_mask __read_mostly;
31 static int nr_counters_fixed __read_mostly;
33 struct cpu_hw_counters {
34 struct perf_counter *counters[X86_PMC_IDX_MAX];
35 unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
42 * Intel PerfMon v3. Used on Core2 and later.
44 static DEFINE_PER_CPU(struct cpu_hw_counters, cpu_hw_counters);
46 static const int intel_perfmon_event_map[] =
48 [PERF_COUNT_CPU_CYCLES] = 0x003c,
49 [PERF_COUNT_INSTRUCTIONS] = 0x00c0,
50 [PERF_COUNT_CACHE_REFERENCES] = 0x4f2e,
51 [PERF_COUNT_CACHE_MISSES] = 0x412e,
52 [PERF_COUNT_BRANCH_INSTRUCTIONS] = 0x00c4,
53 [PERF_COUNT_BRANCH_MISSES] = 0x00c5,
54 [PERF_COUNT_BUS_CYCLES] = 0x013c,
57 static const int max_intel_perfmon_events = ARRAY_SIZE(intel_perfmon_event_map);
60 * Propagate counter elapsed time into the generic counter.
61 * Can only be executed on the CPU where the counter is active.
62 * Returns the delta events processed.
65 x86_perf_counter_update(struct perf_counter *counter,
66 struct hw_perf_counter *hwc, int idx)
68 u64 prev_raw_count, new_raw_count, delta;
71 * Careful: an NMI might modify the previous counter value.
73 * Our tactic to handle this is to first atomically read and
74 * exchange a new raw count - then add that new-prev delta
75 * count to the generic counter atomically:
78 prev_raw_count = atomic64_read(&hwc->prev_count);
79 rdmsrl(hwc->counter_base + idx, new_raw_count);
81 if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count,
82 new_raw_count) != prev_raw_count)
86 * Now we have the new raw value and have updated the prev
87 * timestamp already. We can now calculate the elapsed delta
88 * (counter-)time and add that to the generic counter.
90 * Careful, not all hw sign-extends above the physical width
91 * of the count, so we do that by clipping the delta to 32 bits:
93 delta = (u64)(u32)((s32)new_raw_count - (s32)prev_raw_count);
95 atomic64_add(delta, &counter->count);
96 atomic64_sub(delta, &hwc->period_left);
100 * Setup the hardware configuration for a given hw_event_type
102 static int __hw_perf_counter_init(struct perf_counter *counter)
104 struct perf_counter_hw_event *hw_event = &counter->hw_event;
105 struct hw_perf_counter *hwc = &counter->hw;
107 if (unlikely(!perf_counters_initialized))
111 * Count user events, and generate PMC IRQs:
112 * (keep 'enabled' bit clear for now)
114 hwc->config = ARCH_PERFMON_EVENTSEL_USR | ARCH_PERFMON_EVENTSEL_INT;
117 * If privileged enough, count OS events too, and allow
118 * NMI events as well:
121 if (capable(CAP_SYS_ADMIN)) {
122 hwc->config |= ARCH_PERFMON_EVENTSEL_OS;
127 hwc->irq_period = hw_event->irq_period;
129 * Intel PMCs cannot be accessed sanely above 32 bit width,
130 * so we install an artificial 1<<31 period regardless of
131 * the generic counter period:
133 if ((s64)hwc->irq_period <= 0 || hwc->irq_period > 0x7FFFFFFF)
134 hwc->irq_period = 0x7FFFFFFF;
136 atomic64_set(&hwc->period_left, hwc->irq_period);
139 * Raw event type provide the config in the event structure
142 hwc->config |= hw_event->type;
144 if (hw_event->type >= max_intel_perfmon_events)
149 hwc->config |= intel_perfmon_event_map[hw_event->type];
151 counter->wakeup_pending = 0;
156 u64 hw_perf_save_disable(void)
160 if (unlikely(!perf_counters_initialized))
163 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
164 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
168 EXPORT_SYMBOL_GPL(hw_perf_save_disable);
170 void hw_perf_restore(u64 ctrl)
172 if (unlikely(!perf_counters_initialized))
175 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
177 EXPORT_SYMBOL_GPL(hw_perf_restore);
180 __pmc_fixed_disable(struct perf_counter *counter,
181 struct hw_perf_counter *hwc, unsigned int __idx)
183 int idx = __idx - X86_PMC_IDX_FIXED;
187 mask = 0xfULL << (idx * 4);
189 rdmsrl(hwc->config_base, ctrl_val);
191 err = checking_wrmsrl(hwc->config_base, ctrl_val);
195 __pmc_generic_disable(struct perf_counter *counter,
196 struct hw_perf_counter *hwc, unsigned int idx)
198 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL))
199 __pmc_fixed_disable(counter, hwc, idx);
201 wrmsr_safe(hwc->config_base + idx, hwc->config, 0);
204 static DEFINE_PER_CPU(u64, prev_left[X86_PMC_IDX_MAX]);
207 * Set the next IRQ period, based on the hwc->period_left value.
208 * To be called with the counter disabled in hw:
211 __hw_perf_counter_set_period(struct perf_counter *counter,
212 struct hw_perf_counter *hwc, int idx)
214 s64 left = atomic64_read(&hwc->period_left);
215 s32 period = hwc->irq_period;
219 * If we are way outside a reasoable range then just skip forward:
221 if (unlikely(left <= -period)) {
223 atomic64_set(&hwc->period_left, left);
226 if (unlikely(left <= 0)) {
228 atomic64_set(&hwc->period_left, left);
231 per_cpu(prev_left[idx], smp_processor_id()) = left;
234 * The hw counter starts counting from this counter offset,
235 * mark it to be able to extra future deltas:
237 atomic64_set(&hwc->prev_count, (u64)-left);
239 err = checking_wrmsrl(hwc->counter_base + idx,
240 (u64)(-left) & counter_value_mask);
244 __pmc_fixed_enable(struct perf_counter *counter,
245 struct hw_perf_counter *hwc, unsigned int __idx)
247 int idx = __idx - X86_PMC_IDX_FIXED;
248 u64 ctrl_val, bits, mask;
252 * Enable IRQ generation (0x8) and ring-3 counting (0x2),
253 * and enable ring-0 counting if allowed:
255 bits = 0x8ULL | 0x2ULL;
256 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
259 mask = 0xfULL << (idx * 4);
261 rdmsrl(hwc->config_base, ctrl_val);
264 err = checking_wrmsrl(hwc->config_base, ctrl_val);
268 __pmc_generic_enable(struct perf_counter *counter,
269 struct hw_perf_counter *hwc, int idx)
271 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL))
272 __pmc_fixed_enable(counter, hwc, idx);
274 wrmsr(hwc->config_base + idx,
275 hwc->config | ARCH_PERFMON_EVENTSEL0_ENABLE, 0);
279 fixed_mode_idx(struct perf_counter *counter, struct hw_perf_counter *hwc)
283 if (unlikely(hwc->nmi))
286 event = hwc->config & ARCH_PERFMON_EVENT_MASK;
288 if (unlikely(event == intel_perfmon_event_map[PERF_COUNT_INSTRUCTIONS]))
289 return X86_PMC_IDX_FIXED_INSTRUCTIONS;
290 if (unlikely(event == intel_perfmon_event_map[PERF_COUNT_CPU_CYCLES]))
291 return X86_PMC_IDX_FIXED_CPU_CYCLES;
292 if (unlikely(event == intel_perfmon_event_map[PERF_COUNT_BUS_CYCLES]))
293 return X86_PMC_IDX_FIXED_BUS_CYCLES;
299 * Find a PMC slot for the freshly enabled / scheduled in counter:
301 static int pmc_generic_enable(struct perf_counter *counter)
303 struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
304 struct hw_perf_counter *hwc = &counter->hw;
307 idx = fixed_mode_idx(counter, hwc);
310 * Try to get the fixed counter, if that is already taken
311 * then try to get a generic counter:
313 if (test_and_set_bit(idx, cpuc->used))
316 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
318 * We set it so that counter_base + idx in wrmsr/rdmsr maps to
319 * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
322 MSR_ARCH_PERFMON_FIXED_CTR0 - X86_PMC_IDX_FIXED;
326 /* Try to get the previous generic counter again */
327 if (test_and_set_bit(idx, cpuc->used)) {
329 idx = find_first_zero_bit(cpuc->used, nr_counters_generic);
330 if (idx == nr_counters_generic)
333 set_bit(idx, cpuc->used);
336 hwc->config_base = MSR_ARCH_PERFMON_EVENTSEL0;
337 hwc->counter_base = MSR_ARCH_PERFMON_PERFCTR0;
340 perf_counters_lapic_init(hwc->nmi);
342 __pmc_generic_disable(counter, hwc, idx);
344 cpuc->counters[idx] = counter;
346 * Make it visible before enabling the hw:
350 __hw_perf_counter_set_period(counter, hwc, idx);
351 __pmc_generic_enable(counter, hwc, idx);
356 void perf_counter_print_debug(void)
358 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
359 struct cpu_hw_counters *cpuc;
362 if (!nr_counters_generic)
367 cpu = smp_processor_id();
368 cpuc = &per_cpu(cpu_hw_counters, cpu);
370 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
371 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
372 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
373 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
375 printk(KERN_INFO "\n");
376 printk(KERN_INFO "CPU#%d: ctrl: %016llx\n", cpu, ctrl);
377 printk(KERN_INFO "CPU#%d: status: %016llx\n", cpu, status);
378 printk(KERN_INFO "CPU#%d: overflow: %016llx\n", cpu, overflow);
379 printk(KERN_INFO "CPU#%d: fixed: %016llx\n", cpu, fixed);
380 printk(KERN_INFO "CPU#%d: used: %016llx\n", cpu, *(u64 *)cpuc->used);
382 for (idx = 0; idx < nr_counters_generic; idx++) {
383 rdmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + idx, pmc_ctrl);
384 rdmsrl(MSR_ARCH_PERFMON_PERFCTR0 + idx, pmc_count);
386 prev_left = per_cpu(prev_left[idx], cpu);
388 printk(KERN_INFO "CPU#%d: gen-PMC%d ctrl: %016llx\n",
390 printk(KERN_INFO "CPU#%d: gen-PMC%d count: %016llx\n",
391 cpu, idx, pmc_count);
392 printk(KERN_INFO "CPU#%d: gen-PMC%d left: %016llx\n",
393 cpu, idx, prev_left);
395 for (idx = 0; idx < nr_counters_fixed; idx++) {
396 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
398 printk(KERN_INFO "CPU#%d: fixed-PMC%d count: %016llx\n",
399 cpu, idx, pmc_count);
404 static void pmc_generic_disable(struct perf_counter *counter)
406 struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
407 struct hw_perf_counter *hwc = &counter->hw;
408 unsigned int idx = hwc->idx;
410 __pmc_generic_disable(counter, hwc, idx);
412 clear_bit(idx, cpuc->used);
413 cpuc->counters[idx] = NULL;
415 * Make sure the cleared pointer becomes visible before we
416 * (potentially) free the counter:
421 * Drain the remaining delta count out of a counter
422 * that we are disabling:
424 x86_perf_counter_update(counter, hwc, idx);
427 static void perf_store_irq_data(struct perf_counter *counter, u64 data)
429 struct perf_data *irqdata = counter->irqdata;
431 if (irqdata->len > PERF_DATA_BUFLEN - sizeof(u64)) {
434 u64 *p = (u64 *) &irqdata->data[irqdata->len];
437 irqdata->len += sizeof(u64);
442 * Save and restart an expired counter. Called by NMI contexts,
443 * so it has to be careful about preempting normal counter ops:
445 static void perf_save_and_restart(struct perf_counter *counter)
447 struct hw_perf_counter *hwc = &counter->hw;
450 x86_perf_counter_update(counter, hwc, idx);
451 __hw_perf_counter_set_period(counter, hwc, idx);
453 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
454 __pmc_generic_enable(counter, hwc, idx);
458 perf_handle_group(struct perf_counter *sibling, u64 *status, u64 *overflown)
460 struct perf_counter *counter, *group_leader = sibling->group_leader;
463 * Store sibling timestamps (if any):
465 list_for_each_entry(counter, &group_leader->sibling_list, list_entry) {
467 x86_perf_counter_update(counter, &counter->hw, counter->hw.idx);
468 perf_store_irq_data(sibling, counter->hw_event.type);
469 perf_store_irq_data(sibling, atomic64_read(&counter->count));
474 * This handler is triggered by the local APIC, so the APIC IRQ handling
477 static void __smp_perf_counter_interrupt(struct pt_regs *regs, int nmi)
479 int bit, cpu = smp_processor_id();
480 u64 ack, status, now;
481 struct cpu_hw_counters *cpuc = &per_cpu(cpu_hw_counters, cpu);
483 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, cpuc->global_enable);
485 /* Disable counters globally */
486 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
490 if (now - cpuc->last_interrupt < PERFMON_MIN_PERIOD_NS)
492 cpuc->last_interrupt = now;
494 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
500 for_each_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
501 struct perf_counter *counter = cpuc->counters[bit];
503 clear_bit(bit, (unsigned long *) &status);
507 perf_save_and_restart(counter);
509 switch (counter->hw_event.record_type) {
510 case PERF_RECORD_SIMPLE:
512 case PERF_RECORD_IRQ:
513 perf_store_irq_data(counter, instruction_pointer(regs));
515 case PERF_RECORD_GROUP:
516 perf_handle_group(counter, &status, &ack);
520 * From NMI context we cannot call into the scheduler to
521 * do a task wakeup - but we mark these generic as
522 * wakeup_pending and initate a wakeup callback:
525 counter->wakeup_pending = 1;
526 set_tsk_thread_flag(current, TIF_PERF_COUNTERS);
528 wake_up(&counter->waitq);
532 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
535 * Repeat if there is more work to be done:
537 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
542 * Restore - do not reenable when global enable is off or throttled:
544 if (!cpuc->throttled)
545 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, cpuc->global_enable);
548 void perf_counter_unthrottle(void)
550 struct cpu_hw_counters *cpuc;
552 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
555 if (unlikely(!perf_counters_initialized))
558 cpuc = &per_cpu(cpu_hw_counters, smp_processor_id());
559 if (cpuc->throttled) {
560 if (printk_ratelimit())
561 printk(KERN_WARNING "PERFMON: max event frequency exceeded!\n");
562 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, cpuc->global_enable);
567 void smp_perf_counter_interrupt(struct pt_regs *regs)
570 inc_irq_stat(apic_perf_irqs);
571 apic_write(APIC_LVTPC, LOCAL_PERF_VECTOR);
572 __smp_perf_counter_interrupt(regs, 0);
578 * This handler is triggered by NMI contexts:
580 void perf_counter_notify(struct pt_regs *regs)
582 struct cpu_hw_counters *cpuc;
586 local_irq_save(flags);
587 cpu = smp_processor_id();
588 cpuc = &per_cpu(cpu_hw_counters, cpu);
590 for_each_bit(bit, cpuc->used, X86_PMC_IDX_MAX) {
591 struct perf_counter *counter = cpuc->counters[bit];
596 if (counter->wakeup_pending) {
597 counter->wakeup_pending = 0;
598 wake_up(&counter->waitq);
602 local_irq_restore(flags);
605 void __cpuinit perf_counters_lapic_init(int nmi)
609 if (!perf_counters_initialized)
612 * Enable the performance counter vector in the APIC LVT:
614 apic_val = apic_read(APIC_LVTERR);
616 apic_write(APIC_LVTERR, apic_val | APIC_LVT_MASKED);
618 apic_write(APIC_LVTPC, APIC_DM_NMI);
620 apic_write(APIC_LVTPC, LOCAL_PERF_VECTOR);
621 apic_write(APIC_LVTERR, apic_val);
625 perf_counter_nmi_handler(struct notifier_block *self,
626 unsigned long cmd, void *__args)
628 struct die_args *args = __args;
629 struct pt_regs *regs;
631 if (likely(cmd != DIE_NMI_IPI))
636 apic_write(APIC_LVTPC, APIC_DM_NMI);
637 __smp_perf_counter_interrupt(regs, 1);
642 static __read_mostly struct notifier_block perf_counter_nmi_notifier = {
643 .notifier_call = perf_counter_nmi_handler
646 void __init init_hw_perf_counters(void)
648 union cpuid10_eax eax;
651 union cpuid10_edx edx;
653 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
657 * Check whether the Architectural PerfMon supports
658 * Branch Misses Retired Event or not.
660 cpuid(10, &eax.full, &ebx, &unused, &edx.full);
661 if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
664 printk(KERN_INFO "Intel Performance Monitoring support detected.\n");
666 printk(KERN_INFO "... version: %d\n", eax.split.version_id);
667 printk(KERN_INFO "... num counters: %d\n", eax.split.num_counters);
668 nr_counters_generic = eax.split.num_counters;
669 if (nr_counters_generic > X86_PMC_MAX_GENERIC) {
670 nr_counters_generic = X86_PMC_MAX_GENERIC;
671 WARN(1, KERN_ERR "hw perf counters %d > max(%d), clipping!",
672 nr_counters_generic, X86_PMC_MAX_GENERIC);
674 perf_counter_mask = (1 << nr_counters_generic) - 1;
675 perf_max_counters = nr_counters_generic;
677 printk(KERN_INFO "... bit width: %d\n", eax.split.bit_width);
678 counter_value_mask = (1ULL << eax.split.bit_width) - 1;
679 printk(KERN_INFO "... value mask: %016Lx\n", counter_value_mask);
681 printk(KERN_INFO "... mask length: %d\n", eax.split.mask_length);
683 nr_counters_fixed = edx.split.num_counters_fixed;
684 if (nr_counters_fixed > X86_PMC_MAX_FIXED) {
685 nr_counters_fixed = X86_PMC_MAX_FIXED;
686 WARN(1, KERN_ERR "hw perf counters fixed %d > max(%d), clipping!",
687 nr_counters_fixed, X86_PMC_MAX_FIXED);
689 printk(KERN_INFO "... fixed counters: %d\n", nr_counters_fixed);
691 perf_counter_mask |= ((1LL << nr_counters_fixed)-1) << X86_PMC_IDX_FIXED;
693 printk(KERN_INFO "... counter mask: %016Lx\n", perf_counter_mask);
694 perf_counters_initialized = true;
696 perf_counters_lapic_init(0);
697 register_die_notifier(&perf_counter_nmi_notifier);
700 static void pmc_generic_read(struct perf_counter *counter)
702 x86_perf_counter_update(counter, &counter->hw, counter->hw.idx);
705 static const struct hw_perf_counter_ops x86_perf_counter_ops = {
706 .enable = pmc_generic_enable,
707 .disable = pmc_generic_disable,
708 .read = pmc_generic_read,
711 const struct hw_perf_counter_ops *
712 hw_perf_counter_init(struct perf_counter *counter)
716 err = __hw_perf_counter_init(counter);
720 return &x86_perf_counter_ops;