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/intel_arch_perfmon.h>
22 static bool perf_counters_initialized __read_mostly;
25 * Number of (generic) HW counters:
27 static int nr_hw_counters __read_mostly;
28 static u32 perf_counter_mask __read_mostly;
30 /* No support for fixed function counters yet */
32 #define MAX_HW_COUNTERS 8
34 struct cpu_hw_counters {
35 struct perf_counter *counters[MAX_HW_COUNTERS];
36 unsigned long used[BITS_TO_LONGS(MAX_HW_COUNTERS)];
40 * Intel PerfMon v3. Used on Core2 and later.
42 static DEFINE_PER_CPU(struct cpu_hw_counters, cpu_hw_counters);
44 static const int intel_perfmon_event_map[] =
46 [PERF_COUNT_CYCLES] = 0x003c,
47 [PERF_COUNT_INSTRUCTIONS] = 0x00c0,
48 [PERF_COUNT_CACHE_REFERENCES] = 0x4f2e,
49 [PERF_COUNT_CACHE_MISSES] = 0x412e,
50 [PERF_COUNT_BRANCH_INSTRUCTIONS] = 0x00c4,
51 [PERF_COUNT_BRANCH_MISSES] = 0x00c5,
54 static const int max_intel_perfmon_events = ARRAY_SIZE(intel_perfmon_event_map);
57 * Propagate counter elapsed time into the generic counter.
58 * Can only be executed on the CPU where the counter is active.
59 * Returns the delta events processed.
62 x86_perf_counter_update(struct perf_counter *counter,
63 struct hw_perf_counter *hwc, int idx)
65 u64 prev_raw_count, new_raw_count, delta;
68 * Careful: an NMI might modify the previous counter value.
70 * Our tactic to handle this is to first atomically read and
71 * exchange a new raw count - then add that new-prev delta
72 * count to the generic counter atomically:
75 prev_raw_count = atomic64_read(&hwc->prev_count);
76 rdmsrl(hwc->counter_base + idx, new_raw_count);
78 if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count,
79 new_raw_count) != prev_raw_count)
83 * Now we have the new raw value and have updated the prev
84 * timestamp already. We can now calculate the elapsed delta
85 * (counter-)time and add that to the generic counter.
87 * Careful, not all hw sign-extends above the physical width
88 * of the count, so we do that by clipping the delta to 32 bits:
90 delta = (u64)(u32)((s32)new_raw_count - (s32)prev_raw_count);
92 atomic64_add(delta, &counter->count);
93 atomic64_sub(delta, &hwc->period_left);
97 * Setup the hardware configuration for a given hw_event_type
99 static int __hw_perf_counter_init(struct perf_counter *counter)
101 struct perf_counter_hw_event *hw_event = &counter->hw_event;
102 struct hw_perf_counter *hwc = &counter->hw;
104 if (unlikely(!perf_counters_initialized))
108 * Count user events, and generate PMC IRQs:
109 * (keep 'enabled' bit clear for now)
111 hwc->config = ARCH_PERFMON_EVENTSEL_USR | ARCH_PERFMON_EVENTSEL_INT;
114 * If privileged enough, count OS events too, and allow
115 * NMI events as well:
118 if (capable(CAP_SYS_ADMIN)) {
119 hwc->config |= ARCH_PERFMON_EVENTSEL_OS;
124 hwc->config_base = MSR_ARCH_PERFMON_EVENTSEL0;
125 hwc->counter_base = MSR_ARCH_PERFMON_PERFCTR0;
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 void hw_perf_enable_all(void)
158 if (unlikely(!perf_counters_initialized))
161 wrmsr(MSR_CORE_PERF_GLOBAL_CTRL, perf_counter_mask, 0);
164 u64 hw_perf_save_disable(void)
168 if (unlikely(!perf_counters_initialized))
171 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
172 wrmsr(MSR_CORE_PERF_GLOBAL_CTRL, 0, 0);
176 EXPORT_SYMBOL_GPL(hw_perf_save_disable);
178 void hw_perf_restore(u64 ctrl)
180 if (unlikely(!perf_counters_initialized))
183 wrmsr(MSR_CORE_PERF_GLOBAL_CTRL, ctrl, 0);
185 EXPORT_SYMBOL_GPL(hw_perf_restore);
188 __x86_perf_counter_disable(struct perf_counter *counter,
189 struct hw_perf_counter *hwc, unsigned int idx)
193 err = wrmsr_safe(hwc->config_base + idx, hwc->config, 0);
196 static DEFINE_PER_CPU(u64, prev_left[MAX_HW_COUNTERS]);
199 * Set the next IRQ period, based on the hwc->period_left value.
200 * To be called with the counter disabled in hw:
203 __hw_perf_counter_set_period(struct perf_counter *counter,
204 struct hw_perf_counter *hwc, int idx)
206 s32 left = atomic64_read(&hwc->period_left);
207 s32 period = hwc->irq_period;
210 * If we are way outside a reasoable range then just skip forward:
212 if (unlikely(left <= -period)) {
214 atomic64_set(&hwc->period_left, left);
217 if (unlikely(left <= 0)) {
219 atomic64_set(&hwc->period_left, left);
222 per_cpu(prev_left[idx], smp_processor_id()) = left;
225 * The hw counter starts counting from this counter offset,
226 * mark it to be able to extra future deltas:
228 atomic64_set(&hwc->prev_count, (u64)(s64)-left);
230 wrmsr(hwc->counter_base + idx, -left, 0);
234 __x86_perf_counter_enable(struct perf_counter *counter,
235 struct hw_perf_counter *hwc, int idx)
237 wrmsr(hwc->config_base + idx,
238 hwc->config | ARCH_PERFMON_EVENTSEL0_ENABLE, 0);
242 * Find a PMC slot for the freshly enabled / scheduled in counter:
244 static void x86_perf_counter_enable(struct perf_counter *counter)
246 struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
247 struct hw_perf_counter *hwc = &counter->hw;
250 /* Try to get the previous counter again */
251 if (test_and_set_bit(idx, cpuc->used)) {
252 idx = find_first_zero_bit(cpuc->used, nr_hw_counters);
253 set_bit(idx, cpuc->used);
257 perf_counters_lapic_init(hwc->nmi);
259 __x86_perf_counter_disable(counter, hwc, idx);
261 cpuc->counters[idx] = counter;
263 __hw_perf_counter_set_period(counter, hwc, idx);
264 __x86_perf_counter_enable(counter, hwc, idx);
267 void perf_counter_print_debug(void)
269 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left;
277 cpu = smp_processor_id();
279 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
280 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
281 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
283 printk(KERN_INFO "\n");
284 printk(KERN_INFO "CPU#%d: ctrl: %016llx\n", cpu, ctrl);
285 printk(KERN_INFO "CPU#%d: status: %016llx\n", cpu, status);
286 printk(KERN_INFO "CPU#%d: overflow: %016llx\n", cpu, overflow);
288 for (idx = 0; idx < nr_hw_counters; idx++) {
289 rdmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + idx, pmc_ctrl);
290 rdmsrl(MSR_ARCH_PERFMON_PERFCTR0 + idx, pmc_count);
292 prev_left = per_cpu(prev_left[idx], cpu);
294 printk(KERN_INFO "CPU#%d: PMC%d ctrl: %016llx\n",
296 printk(KERN_INFO "CPU#%d: PMC%d count: %016llx\n",
297 cpu, idx, pmc_count);
298 printk(KERN_INFO "CPU#%d: PMC%d left: %016llx\n",
299 cpu, idx, prev_left);
304 static void x86_perf_counter_disable(struct perf_counter *counter)
306 struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
307 struct hw_perf_counter *hwc = &counter->hw;
308 unsigned int idx = hwc->idx;
310 __x86_perf_counter_disable(counter, hwc, idx);
312 clear_bit(idx, cpuc->used);
313 cpuc->counters[idx] = NULL;
316 * Drain the remaining delta count out of a counter
317 * that we are disabling:
319 x86_perf_counter_update(counter, hwc, idx);
322 static void perf_store_irq_data(struct perf_counter *counter, u64 data)
324 struct perf_data *irqdata = counter->irqdata;
326 if (irqdata->len > PERF_DATA_BUFLEN - sizeof(u64)) {
329 u64 *p = (u64 *) &irqdata->data[irqdata->len];
332 irqdata->len += sizeof(u64);
337 * Save and restart an expired counter. Called by NMI contexts,
338 * so it has to be careful about preempting normal counter ops:
340 static void perf_save_and_restart(struct perf_counter *counter)
342 struct hw_perf_counter *hwc = &counter->hw;
346 rdmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + idx, pmc_ctrl);
348 x86_perf_counter_update(counter, hwc, idx);
349 __hw_perf_counter_set_period(counter, hwc, idx);
351 if (pmc_ctrl & ARCH_PERFMON_EVENTSEL0_ENABLE)
352 __x86_perf_counter_enable(counter, hwc, idx);
356 perf_handle_group(struct perf_counter *sibling, u64 *status, u64 *overflown)
358 struct perf_counter *counter, *group_leader = sibling->group_leader;
361 * Store sibling timestamps (if any):
363 list_for_each_entry(counter, &group_leader->sibling_list, list_entry) {
364 x86_perf_counter_update(counter, &counter->hw, counter->hw.idx);
365 perf_store_irq_data(sibling, counter->hw_event.type);
366 perf_store_irq_data(sibling, atomic64_read(&counter->count));
371 * This handler is triggered by the local APIC, so the APIC IRQ handling
374 static void __smp_perf_counter_interrupt(struct pt_regs *regs, int nmi)
376 int bit, cpu = smp_processor_id();
377 u64 ack, status, saved_global;
378 struct cpu_hw_counters *cpuc;
380 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, saved_global);
382 /* Disable counters globally */
383 wrmsr(MSR_CORE_PERF_GLOBAL_CTRL, 0, 0);
386 cpuc = &per_cpu(cpu_hw_counters, cpu);
388 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
394 for_each_bit(bit, (unsigned long *) &status, nr_hw_counters) {
395 struct perf_counter *counter = cpuc->counters[bit];
397 clear_bit(bit, (unsigned long *) &status);
401 perf_save_and_restart(counter);
403 switch (counter->hw_event.record_type) {
404 case PERF_RECORD_SIMPLE:
406 case PERF_RECORD_IRQ:
407 perf_store_irq_data(counter, instruction_pointer(regs));
409 case PERF_RECORD_GROUP:
410 perf_handle_group(counter, &status, &ack);
414 * From NMI context we cannot call into the scheduler to
415 * do a task wakeup - but we mark these counters as
416 * wakeup_pending and initate a wakeup callback:
419 counter->wakeup_pending = 1;
420 set_tsk_thread_flag(current, TIF_PERF_COUNTERS);
422 wake_up(&counter->waitq);
426 wrmsr(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack, 0);
429 * Repeat if there is more work to be done:
431 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
436 * Restore - do not reenable when global enable is off:
438 wrmsr(MSR_CORE_PERF_GLOBAL_CTRL, saved_global, 0);
441 void smp_perf_counter_interrupt(struct pt_regs *regs)
444 inc_irq_stat(apic_perf_irqs);
445 apic_write(APIC_LVTPC, LOCAL_PERF_VECTOR);
446 __smp_perf_counter_interrupt(regs, 0);
452 * This handler is triggered by NMI contexts:
454 void perf_counter_notify(struct pt_regs *regs)
456 struct cpu_hw_counters *cpuc;
460 local_irq_save(flags);
461 cpu = smp_processor_id();
462 cpuc = &per_cpu(cpu_hw_counters, cpu);
464 for_each_bit(bit, cpuc->used, nr_hw_counters) {
465 struct perf_counter *counter = cpuc->counters[bit];
470 if (counter->wakeup_pending) {
471 counter->wakeup_pending = 0;
472 wake_up(&counter->waitq);
476 local_irq_restore(flags);
479 void __cpuinit perf_counters_lapic_init(int nmi)
483 if (!perf_counters_initialized)
486 * Enable the performance counter vector in the APIC LVT:
488 apic_val = apic_read(APIC_LVTERR);
490 apic_write(APIC_LVTERR, apic_val | APIC_LVT_MASKED);
492 apic_write(APIC_LVTPC, APIC_DM_NMI);
494 apic_write(APIC_LVTPC, LOCAL_PERF_VECTOR);
495 apic_write(APIC_LVTERR, apic_val);
499 perf_counter_nmi_handler(struct notifier_block *self,
500 unsigned long cmd, void *__args)
502 struct die_args *args = __args;
503 struct pt_regs *regs;
505 if (likely(cmd != DIE_NMI_IPI))
510 apic_write(APIC_LVTPC, APIC_DM_NMI);
511 __smp_perf_counter_interrupt(regs, 1);
516 static __read_mostly struct notifier_block perf_counter_nmi_notifier = {
517 .notifier_call = perf_counter_nmi_handler
520 void __init init_hw_perf_counters(void)
522 union cpuid10_eax eax;
526 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
530 * Check whether the Architectural PerfMon supports
531 * Branch Misses Retired Event or not.
533 cpuid(10, &(eax.full), &ebx, &unused, &unused);
534 if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
537 printk(KERN_INFO "Intel Performance Monitoring support detected.\n");
539 printk(KERN_INFO "... version: %d\n", eax.split.version_id);
540 printk(KERN_INFO "... num_counters: %d\n", eax.split.num_counters);
541 nr_hw_counters = eax.split.num_counters;
542 if (nr_hw_counters > MAX_HW_COUNTERS) {
543 nr_hw_counters = MAX_HW_COUNTERS;
544 WARN(1, KERN_ERR "hw perf counters %d > max(%d), clipping!",
545 nr_hw_counters, MAX_HW_COUNTERS);
547 perf_counter_mask = (1 << nr_hw_counters) - 1;
548 perf_max_counters = nr_hw_counters;
550 printk(KERN_INFO "... bit_width: %d\n", eax.split.bit_width);
551 printk(KERN_INFO "... mask_length: %d\n", eax.split.mask_length);
553 perf_counters_initialized = true;
555 perf_counters_lapic_init(0);
556 register_die_notifier(&perf_counter_nmi_notifier);
559 static void x86_perf_counter_read(struct perf_counter *counter)
561 x86_perf_counter_update(counter, &counter->hw, counter->hw.idx);
564 static const struct hw_perf_counter_ops x86_perf_counter_ops = {
565 .hw_perf_counter_enable = x86_perf_counter_enable,
566 .hw_perf_counter_disable = x86_perf_counter_disable,
567 .hw_perf_counter_read = x86_perf_counter_read,
570 const struct hw_perf_counter_ops *
571 hw_perf_counter_init(struct perf_counter *counter)
575 err = __hw_perf_counter_init(counter);
579 return &x86_perf_counter_ops;