2 * Performance event support - powerpc architecture code
4 * Copyright 2008-2009 Paul Mackerras, IBM Corporation.
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
11 #include <linux/kernel.h>
12 #include <linux/sched.h>
13 #include <linux/perf_event.h>
14 #include <linux/percpu.h>
15 #include <linux/hardirq.h>
18 #include <asm/machdep.h>
19 #include <asm/firmware.h>
20 #include <asm/ptrace.h>
22 struct cpu_hw_events {
29 struct perf_event *event[MAX_HWEVENTS];
30 u64 events[MAX_HWEVENTS];
31 unsigned int flags[MAX_HWEVENTS];
32 unsigned long mmcr[3];
33 struct perf_event *limited_counter[MAX_LIMITED_HWCOUNTERS];
34 u8 limited_hwidx[MAX_LIMITED_HWCOUNTERS];
35 u64 alternatives[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
36 unsigned long amasks[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
37 unsigned long avalues[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
39 unsigned int group_flag;
42 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
44 struct power_pmu *ppmu;
47 * Normally, to ignore kernel events we set the FCS (freeze counters
48 * in supervisor mode) bit in MMCR0, but if the kernel runs with the
49 * hypervisor bit set in the MSR, or if we are running on a processor
50 * where the hypervisor bit is forced to 1 (as on Apple G5 processors),
51 * then we need to use the FCHV bit to ignore kernel events.
53 static unsigned int freeze_events_kernel = MMCR0_FCS;
56 * 32-bit doesn't have MMCRA but does have an MMCR2,
57 * and a few other names are different.
62 #define MMCR0_PMCjCE MMCR0_PMCnCE
64 #define SPRN_MMCRA SPRN_MMCR2
65 #define MMCRA_SAMPLE_ENABLE 0
67 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
71 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp) { }
72 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
76 static inline void perf_read_regs(struct pt_regs *regs)
80 static inline int perf_intr_is_nmi(struct pt_regs *regs)
85 static inline int siar_valid(struct pt_regs *regs)
90 #endif /* CONFIG_PPC32 */
93 * Things that are specific to 64-bit implementations.
97 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
99 unsigned long mmcra = regs->dsisr;
101 if ((mmcra & MMCRA_SAMPLE_ENABLE) && !(ppmu->flags & PPMU_ALT_SIPR)) {
102 unsigned long slot = (mmcra & MMCRA_SLOT) >> MMCRA_SLOT_SHIFT;
104 return 4 * (slot - 1);
110 * The user wants a data address recorded.
111 * If we're not doing instruction sampling, give them the SDAR
112 * (sampled data address). If we are doing instruction sampling, then
113 * only give them the SDAR if it corresponds to the instruction
114 * pointed to by SIAR; this is indicated by the [POWER6_]MMCRA_SDSYNC or
115 * the [POWER7P_]MMCRA_SDAR_VALID bit in MMCRA.
117 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp)
119 unsigned long mmcra = regs->dsisr;
120 unsigned long sdsync;
122 if (ppmu->flags & PPMU_SIAR_VALID)
123 sdsync = POWER7P_MMCRA_SDAR_VALID;
124 else if (ppmu->flags & PPMU_ALT_SIPR)
125 sdsync = POWER6_MMCRA_SDSYNC;
127 sdsync = MMCRA_SDSYNC;
129 if (!(mmcra & MMCRA_SAMPLE_ENABLE) || (mmcra & sdsync))
130 *addrp = mfspr(SPRN_SDAR);
133 static bool mmcra_sihv(unsigned long mmcra)
135 unsigned long sihv = MMCRA_SIHV;
137 if (ppmu->flags & PPMU_ALT_SIPR)
138 sihv = POWER6_MMCRA_SIHV;
140 return !!(mmcra & sihv);
143 static bool mmcra_sipr(unsigned long mmcra)
145 unsigned long sipr = MMCRA_SIPR;
147 if (ppmu->flags & PPMU_ALT_SIPR)
148 sipr = POWER6_MMCRA_SIPR;
150 return !!(mmcra & sipr);
153 static inline u32 perf_flags_from_msr(struct pt_regs *regs)
155 if (regs->msr & MSR_PR)
156 return PERF_RECORD_MISC_USER;
157 if ((regs->msr & MSR_HV) && freeze_events_kernel != MMCR0_FCHV)
158 return PERF_RECORD_MISC_HYPERVISOR;
159 return PERF_RECORD_MISC_KERNEL;
162 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
164 unsigned long mmcra = regs->dsisr;
165 unsigned long use_siar = regs->result;
168 return perf_flags_from_msr(regs);
171 * If we don't have flags in MMCRA, rather than using
172 * the MSR, we intuit the flags from the address in
173 * SIAR which should give slightly more reliable
176 if (ppmu->flags & PPMU_NO_SIPR) {
177 unsigned long siar = mfspr(SPRN_SIAR);
178 if (siar >= PAGE_OFFSET)
179 return PERF_RECORD_MISC_KERNEL;
180 return PERF_RECORD_MISC_USER;
183 /* PR has priority over HV, so order below is important */
184 if (mmcra_sipr(mmcra))
185 return PERF_RECORD_MISC_USER;
186 if (mmcra_sihv(mmcra) && (freeze_events_kernel != MMCR0_FCHV))
187 return PERF_RECORD_MISC_HYPERVISOR;
188 return PERF_RECORD_MISC_KERNEL;
192 * Overload regs->dsisr to store MMCRA so we only need to read it once
194 * Overload regs->result to specify whether we should use the MSR (result
195 * is zero) or the SIAR (result is non zero).
197 static inline void perf_read_regs(struct pt_regs *regs)
199 unsigned long mmcra = mfspr(SPRN_MMCRA);
200 int marked = mmcra & MMCRA_SAMPLE_ENABLE;
204 * If this isn't a PMU exception (eg a software event) the SIAR is
205 * not valid. Use pt_regs.
207 * If it is a marked event use the SIAR.
209 * If the PMU doesn't update the SIAR for non marked events use
212 * If the PMU has HV/PR flags then check to see if they
213 * place the exception in userspace. If so, use pt_regs. In
214 * continuous sampling mode the SIAR and the PMU exception are
215 * not synchronised, so they may be many instructions apart.
216 * This can result in confusing backtraces. We still want
217 * hypervisor samples as well as samples in the kernel with
218 * interrupts off hence the userspace check.
220 if (TRAP(regs) != 0xf00)
224 else if ((ppmu->flags & PPMU_NO_CONT_SAMPLING))
226 else if (!(ppmu->flags & PPMU_NO_SIPR) && mmcra_sipr(mmcra))
232 regs->result = use_siar;
236 * If interrupts were soft-disabled when a PMU interrupt occurs, treat
239 static inline int perf_intr_is_nmi(struct pt_regs *regs)
245 * On processors like P7+ that have the SIAR-Valid bit, marked instructions
246 * must be sampled only if the SIAR-valid bit is set.
248 * For unmarked instructions and for processors that don't have the SIAR-Valid
249 * bit, assume that SIAR is valid.
251 static inline int siar_valid(struct pt_regs *regs)
253 unsigned long mmcra = regs->dsisr;
254 int marked = mmcra & MMCRA_SAMPLE_ENABLE;
256 if ((ppmu->flags & PPMU_SIAR_VALID) && marked)
257 return mmcra & POWER7P_MMCRA_SIAR_VALID;
262 #endif /* CONFIG_PPC64 */
264 static void perf_event_interrupt(struct pt_regs *regs);
266 void perf_event_print_debug(void)
271 * Read one performance monitor counter (PMC).
273 static unsigned long read_pmc(int idx)
279 val = mfspr(SPRN_PMC1);
282 val = mfspr(SPRN_PMC2);
285 val = mfspr(SPRN_PMC3);
288 val = mfspr(SPRN_PMC4);
291 val = mfspr(SPRN_PMC5);
294 val = mfspr(SPRN_PMC6);
298 val = mfspr(SPRN_PMC7);
301 val = mfspr(SPRN_PMC8);
303 #endif /* CONFIG_PPC64 */
305 printk(KERN_ERR "oops trying to read PMC%d\n", idx);
314 static void write_pmc(int idx, unsigned long val)
318 mtspr(SPRN_PMC1, val);
321 mtspr(SPRN_PMC2, val);
324 mtspr(SPRN_PMC3, val);
327 mtspr(SPRN_PMC4, val);
330 mtspr(SPRN_PMC5, val);
333 mtspr(SPRN_PMC6, val);
337 mtspr(SPRN_PMC7, val);
340 mtspr(SPRN_PMC8, val);
342 #endif /* CONFIG_PPC64 */
344 printk(KERN_ERR "oops trying to write PMC%d\n", idx);
349 * Check if a set of events can all go on the PMU at once.
350 * If they can't, this will look at alternative codes for the events
351 * and see if any combination of alternative codes is feasible.
352 * The feasible set is returned in event_id[].
354 static int power_check_constraints(struct cpu_hw_events *cpuhw,
355 u64 event_id[], unsigned int cflags[],
358 unsigned long mask, value, nv;
359 unsigned long smasks[MAX_HWEVENTS], svalues[MAX_HWEVENTS];
360 int n_alt[MAX_HWEVENTS], choice[MAX_HWEVENTS];
362 unsigned long addf = ppmu->add_fields;
363 unsigned long tadd = ppmu->test_adder;
365 if (n_ev > ppmu->n_counter)
368 /* First see if the events will go on as-is */
369 for (i = 0; i < n_ev; ++i) {
370 if ((cflags[i] & PPMU_LIMITED_PMC_REQD)
371 && !ppmu->limited_pmc_event(event_id[i])) {
372 ppmu->get_alternatives(event_id[i], cflags[i],
373 cpuhw->alternatives[i]);
374 event_id[i] = cpuhw->alternatives[i][0];
376 if (ppmu->get_constraint(event_id[i], &cpuhw->amasks[i][0],
377 &cpuhw->avalues[i][0]))
381 for (i = 0; i < n_ev; ++i) {
382 nv = (value | cpuhw->avalues[i][0]) +
383 (value & cpuhw->avalues[i][0] & addf);
384 if ((((nv + tadd) ^ value) & mask) != 0 ||
385 (((nv + tadd) ^ cpuhw->avalues[i][0]) &
386 cpuhw->amasks[i][0]) != 0)
389 mask |= cpuhw->amasks[i][0];
392 return 0; /* all OK */
394 /* doesn't work, gather alternatives... */
395 if (!ppmu->get_alternatives)
397 for (i = 0; i < n_ev; ++i) {
399 n_alt[i] = ppmu->get_alternatives(event_id[i], cflags[i],
400 cpuhw->alternatives[i]);
401 for (j = 1; j < n_alt[i]; ++j)
402 ppmu->get_constraint(cpuhw->alternatives[i][j],
403 &cpuhw->amasks[i][j],
404 &cpuhw->avalues[i][j]);
407 /* enumerate all possibilities and see if any will work */
410 value = mask = nv = 0;
413 /* we're backtracking, restore context */
419 * See if any alternative k for event_id i,
420 * where k > j, will satisfy the constraints.
422 while (++j < n_alt[i]) {
423 nv = (value | cpuhw->avalues[i][j]) +
424 (value & cpuhw->avalues[i][j] & addf);
425 if ((((nv + tadd) ^ value) & mask) == 0 &&
426 (((nv + tadd) ^ cpuhw->avalues[i][j])
427 & cpuhw->amasks[i][j]) == 0)
432 * No feasible alternative, backtrack
433 * to event_id i-1 and continue enumerating its
434 * alternatives from where we got up to.
440 * Found a feasible alternative for event_id i,
441 * remember where we got up to with this event_id,
442 * go on to the next event_id, and start with
443 * the first alternative for it.
449 mask |= cpuhw->amasks[i][j];
455 /* OK, we have a feasible combination, tell the caller the solution */
456 for (i = 0; i < n_ev; ++i)
457 event_id[i] = cpuhw->alternatives[i][choice[i]];
462 * Check if newly-added events have consistent settings for
463 * exclude_{user,kernel,hv} with each other and any previously
466 static int check_excludes(struct perf_event **ctrs, unsigned int cflags[],
467 int n_prev, int n_new)
469 int eu = 0, ek = 0, eh = 0;
471 struct perf_event *event;
478 for (i = 0; i < n; ++i) {
479 if (cflags[i] & PPMU_LIMITED_PMC_OK) {
480 cflags[i] &= ~PPMU_LIMITED_PMC_REQD;
485 eu = event->attr.exclude_user;
486 ek = event->attr.exclude_kernel;
487 eh = event->attr.exclude_hv;
489 } else if (event->attr.exclude_user != eu ||
490 event->attr.exclude_kernel != ek ||
491 event->attr.exclude_hv != eh) {
497 for (i = 0; i < n; ++i)
498 if (cflags[i] & PPMU_LIMITED_PMC_OK)
499 cflags[i] |= PPMU_LIMITED_PMC_REQD;
504 static u64 check_and_compute_delta(u64 prev, u64 val)
506 u64 delta = (val - prev) & 0xfffffffful;
509 * POWER7 can roll back counter values, if the new value is smaller
510 * than the previous value it will cause the delta and the counter to
511 * have bogus values unless we rolled a counter over. If a coutner is
512 * rolled back, it will be smaller, but within 256, which is the maximum
513 * number of events to rollback at once. If we dectect a rollback
514 * return 0. This can lead to a small lack of precision in the
517 if (prev > val && (prev - val) < 256)
523 static void power_pmu_read(struct perf_event *event)
525 s64 val, delta, prev;
527 if (event->hw.state & PERF_HES_STOPPED)
533 * Performance monitor interrupts come even when interrupts
534 * are soft-disabled, as long as interrupts are hard-enabled.
535 * Therefore we treat them like NMIs.
538 prev = local64_read(&event->hw.prev_count);
540 val = read_pmc(event->hw.idx);
541 delta = check_and_compute_delta(prev, val);
544 } while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
546 local64_add(delta, &event->count);
547 local64_sub(delta, &event->hw.period_left);
551 * On some machines, PMC5 and PMC6 can't be written, don't respect
552 * the freeze conditions, and don't generate interrupts. This tells
553 * us if `event' is using such a PMC.
555 static int is_limited_pmc(int pmcnum)
557 return (ppmu->flags & PPMU_LIMITED_PMC5_6)
558 && (pmcnum == 5 || pmcnum == 6);
561 static void freeze_limited_counters(struct cpu_hw_events *cpuhw,
562 unsigned long pmc5, unsigned long pmc6)
564 struct perf_event *event;
565 u64 val, prev, delta;
568 for (i = 0; i < cpuhw->n_limited; ++i) {
569 event = cpuhw->limited_counter[i];
572 val = (event->hw.idx == 5) ? pmc5 : pmc6;
573 prev = local64_read(&event->hw.prev_count);
575 delta = check_and_compute_delta(prev, val);
577 local64_add(delta, &event->count);
581 static void thaw_limited_counters(struct cpu_hw_events *cpuhw,
582 unsigned long pmc5, unsigned long pmc6)
584 struct perf_event *event;
588 for (i = 0; i < cpuhw->n_limited; ++i) {
589 event = cpuhw->limited_counter[i];
590 event->hw.idx = cpuhw->limited_hwidx[i];
591 val = (event->hw.idx == 5) ? pmc5 : pmc6;
592 prev = local64_read(&event->hw.prev_count);
593 if (check_and_compute_delta(prev, val))
594 local64_set(&event->hw.prev_count, val);
595 perf_event_update_userpage(event);
600 * Since limited events don't respect the freeze conditions, we
601 * have to read them immediately after freezing or unfreezing the
602 * other events. We try to keep the values from the limited
603 * events as consistent as possible by keeping the delay (in
604 * cycles and instructions) between freezing/unfreezing and reading
605 * the limited events as small and consistent as possible.
606 * Therefore, if any limited events are in use, we read them
607 * both, and always in the same order, to minimize variability,
608 * and do it inside the same asm that writes MMCR0.
610 static void write_mmcr0(struct cpu_hw_events *cpuhw, unsigned long mmcr0)
612 unsigned long pmc5, pmc6;
614 if (!cpuhw->n_limited) {
615 mtspr(SPRN_MMCR0, mmcr0);
620 * Write MMCR0, then read PMC5 and PMC6 immediately.
621 * To ensure we don't get a performance monitor interrupt
622 * between writing MMCR0 and freezing/thawing the limited
623 * events, we first write MMCR0 with the event overflow
624 * interrupt enable bits turned off.
626 asm volatile("mtspr %3,%2; mfspr %0,%4; mfspr %1,%5"
627 : "=&r" (pmc5), "=&r" (pmc6)
628 : "r" (mmcr0 & ~(MMCR0_PMC1CE | MMCR0_PMCjCE)),
630 "i" (SPRN_PMC5), "i" (SPRN_PMC6));
632 if (mmcr0 & MMCR0_FC)
633 freeze_limited_counters(cpuhw, pmc5, pmc6);
635 thaw_limited_counters(cpuhw, pmc5, pmc6);
638 * Write the full MMCR0 including the event overflow interrupt
639 * enable bits, if necessary.
641 if (mmcr0 & (MMCR0_PMC1CE | MMCR0_PMCjCE))
642 mtspr(SPRN_MMCR0, mmcr0);
646 * Disable all events to prevent PMU interrupts and to allow
647 * events to be added or removed.
649 static void power_pmu_disable(struct pmu *pmu)
651 struct cpu_hw_events *cpuhw;
656 local_irq_save(flags);
657 cpuhw = &__get_cpu_var(cpu_hw_events);
659 if (!cpuhw->disabled) {
664 * Check if we ever enabled the PMU on this cpu.
666 if (!cpuhw->pmcs_enabled) {
668 cpuhw->pmcs_enabled = 1;
672 * Disable instruction sampling if it was enabled
674 if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
676 cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
681 * Set the 'freeze counters' bit.
682 * The barrier is to make sure the mtspr has been
683 * executed and the PMU has frozen the events
686 write_mmcr0(cpuhw, mfspr(SPRN_MMCR0) | MMCR0_FC);
689 local_irq_restore(flags);
693 * Re-enable all events if disable == 0.
694 * If we were previously disabled and events were added, then
695 * put the new config on the PMU.
697 static void power_pmu_enable(struct pmu *pmu)
699 struct perf_event *event;
700 struct cpu_hw_events *cpuhw;
705 unsigned int hwc_index[MAX_HWEVENTS];
711 local_irq_save(flags);
712 cpuhw = &__get_cpu_var(cpu_hw_events);
713 if (!cpuhw->disabled) {
714 local_irq_restore(flags);
720 * If we didn't change anything, or only removed events,
721 * no need to recalculate MMCR* settings and reset the PMCs.
722 * Just reenable the PMU with the current MMCR* settings
723 * (possibly updated for removal of events).
725 if (!cpuhw->n_added) {
726 mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
727 mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
728 if (cpuhw->n_events == 0)
729 ppc_set_pmu_inuse(0);
734 * Compute MMCR* values for the new set of events
736 if (ppmu->compute_mmcr(cpuhw->events, cpuhw->n_events, hwc_index,
738 /* shouldn't ever get here */
739 printk(KERN_ERR "oops compute_mmcr failed\n");
744 * Add in MMCR0 freeze bits corresponding to the
745 * attr.exclude_* bits for the first event.
746 * We have already checked that all events have the
747 * same values for these bits as the first event.
749 event = cpuhw->event[0];
750 if (event->attr.exclude_user)
751 cpuhw->mmcr[0] |= MMCR0_FCP;
752 if (event->attr.exclude_kernel)
753 cpuhw->mmcr[0] |= freeze_events_kernel;
754 if (event->attr.exclude_hv)
755 cpuhw->mmcr[0] |= MMCR0_FCHV;
758 * Write the new configuration to MMCR* with the freeze
759 * bit set and set the hardware events to their initial values.
760 * Then unfreeze the events.
762 ppc_set_pmu_inuse(1);
763 mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
764 mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
765 mtspr(SPRN_MMCR0, (cpuhw->mmcr[0] & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
769 * Read off any pre-existing events that need to move
772 for (i = 0; i < cpuhw->n_events; ++i) {
773 event = cpuhw->event[i];
774 if (event->hw.idx && event->hw.idx != hwc_index[i] + 1) {
775 power_pmu_read(event);
776 write_pmc(event->hw.idx, 0);
782 * Initialize the PMCs for all the new and moved events.
784 cpuhw->n_limited = n_lim = 0;
785 for (i = 0; i < cpuhw->n_events; ++i) {
786 event = cpuhw->event[i];
789 idx = hwc_index[i] + 1;
790 if (is_limited_pmc(idx)) {
791 cpuhw->limited_counter[n_lim] = event;
792 cpuhw->limited_hwidx[n_lim] = idx;
797 if (event->hw.sample_period) {
798 left = local64_read(&event->hw.period_left);
799 if (left < 0x80000000L)
800 val = 0x80000000L - left;
802 local64_set(&event->hw.prev_count, val);
804 if (event->hw.state & PERF_HES_STOPPED)
807 perf_event_update_userpage(event);
809 cpuhw->n_limited = n_lim;
810 cpuhw->mmcr[0] |= MMCR0_PMXE | MMCR0_FCECE;
814 write_mmcr0(cpuhw, cpuhw->mmcr[0]);
817 * Enable instruction sampling if necessary
819 if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
821 mtspr(SPRN_MMCRA, cpuhw->mmcr[2]);
825 local_irq_restore(flags);
828 static int collect_events(struct perf_event *group, int max_count,
829 struct perf_event *ctrs[], u64 *events,
833 struct perf_event *event;
835 if (!is_software_event(group)) {
839 flags[n] = group->hw.event_base;
840 events[n++] = group->hw.config;
842 list_for_each_entry(event, &group->sibling_list, group_entry) {
843 if (!is_software_event(event) &&
844 event->state != PERF_EVENT_STATE_OFF) {
848 flags[n] = event->hw.event_base;
849 events[n++] = event->hw.config;
856 * Add a event to the PMU.
857 * If all events are not already frozen, then we disable and
858 * re-enable the PMU in order to get hw_perf_enable to do the
859 * actual work of reconfiguring the PMU.
861 static int power_pmu_add(struct perf_event *event, int ef_flags)
863 struct cpu_hw_events *cpuhw;
868 local_irq_save(flags);
869 perf_pmu_disable(event->pmu);
872 * Add the event to the list (if there is room)
873 * and check whether the total set is still feasible.
875 cpuhw = &__get_cpu_var(cpu_hw_events);
876 n0 = cpuhw->n_events;
877 if (n0 >= ppmu->n_counter)
879 cpuhw->event[n0] = event;
880 cpuhw->events[n0] = event->hw.config;
881 cpuhw->flags[n0] = event->hw.event_base;
884 * This event may have been disabled/stopped in record_and_restart()
885 * because we exceeded the ->event_limit. If re-starting the event,
886 * clear the ->hw.state (STOPPED and UPTODATE flags), so the user
887 * notification is re-enabled.
889 if (!(ef_flags & PERF_EF_START))
890 event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
895 * If group events scheduling transaction was started,
896 * skip the schedulability test here, it will be performed
897 * at commit time(->commit_txn) as a whole
899 if (cpuhw->group_flag & PERF_EVENT_TXN)
902 if (check_excludes(cpuhw->event, cpuhw->flags, n0, 1))
904 if (power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n0 + 1))
906 event->hw.config = cpuhw->events[n0];
914 perf_pmu_enable(event->pmu);
915 local_irq_restore(flags);
920 * Remove a event from the PMU.
922 static void power_pmu_del(struct perf_event *event, int ef_flags)
924 struct cpu_hw_events *cpuhw;
928 local_irq_save(flags);
929 perf_pmu_disable(event->pmu);
931 power_pmu_read(event);
933 cpuhw = &__get_cpu_var(cpu_hw_events);
934 for (i = 0; i < cpuhw->n_events; ++i) {
935 if (event == cpuhw->event[i]) {
936 while (++i < cpuhw->n_events) {
937 cpuhw->event[i-1] = cpuhw->event[i];
938 cpuhw->events[i-1] = cpuhw->events[i];
939 cpuhw->flags[i-1] = cpuhw->flags[i];
942 ppmu->disable_pmc(event->hw.idx - 1, cpuhw->mmcr);
944 write_pmc(event->hw.idx, 0);
947 perf_event_update_userpage(event);
951 for (i = 0; i < cpuhw->n_limited; ++i)
952 if (event == cpuhw->limited_counter[i])
954 if (i < cpuhw->n_limited) {
955 while (++i < cpuhw->n_limited) {
956 cpuhw->limited_counter[i-1] = cpuhw->limited_counter[i];
957 cpuhw->limited_hwidx[i-1] = cpuhw->limited_hwidx[i];
961 if (cpuhw->n_events == 0) {
962 /* disable exceptions if no events are running */
963 cpuhw->mmcr[0] &= ~(MMCR0_PMXE | MMCR0_FCECE);
966 perf_pmu_enable(event->pmu);
967 local_irq_restore(flags);
971 * POWER-PMU does not support disabling individual counters, hence
972 * program their cycle counter to their max value and ignore the interrupts.
975 static void power_pmu_start(struct perf_event *event, int ef_flags)
981 if (!event->hw.idx || !event->hw.sample_period)
984 if (!(event->hw.state & PERF_HES_STOPPED))
987 if (ef_flags & PERF_EF_RELOAD)
988 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
990 local_irq_save(flags);
991 perf_pmu_disable(event->pmu);
994 left = local64_read(&event->hw.period_left);
997 if (left < 0x80000000L)
998 val = 0x80000000L - left;
1000 write_pmc(event->hw.idx, val);
1002 perf_event_update_userpage(event);
1003 perf_pmu_enable(event->pmu);
1004 local_irq_restore(flags);
1007 static void power_pmu_stop(struct perf_event *event, int ef_flags)
1009 unsigned long flags;
1011 if (!event->hw.idx || !event->hw.sample_period)
1014 if (event->hw.state & PERF_HES_STOPPED)
1017 local_irq_save(flags);
1018 perf_pmu_disable(event->pmu);
1020 power_pmu_read(event);
1021 event->hw.state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1022 write_pmc(event->hw.idx, 0);
1024 perf_event_update_userpage(event);
1025 perf_pmu_enable(event->pmu);
1026 local_irq_restore(flags);
1030 * Start group events scheduling transaction
1031 * Set the flag to make pmu::enable() not perform the
1032 * schedulability test, it will be performed at commit time
1034 void power_pmu_start_txn(struct pmu *pmu)
1036 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1038 perf_pmu_disable(pmu);
1039 cpuhw->group_flag |= PERF_EVENT_TXN;
1040 cpuhw->n_txn_start = cpuhw->n_events;
1044 * Stop group events scheduling transaction
1045 * Clear the flag and pmu::enable() will perform the
1046 * schedulability test.
1048 void power_pmu_cancel_txn(struct pmu *pmu)
1050 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1052 cpuhw->group_flag &= ~PERF_EVENT_TXN;
1053 perf_pmu_enable(pmu);
1057 * Commit group events scheduling transaction
1058 * Perform the group schedulability test as a whole
1059 * Return 0 if success
1061 int power_pmu_commit_txn(struct pmu *pmu)
1063 struct cpu_hw_events *cpuhw;
1068 cpuhw = &__get_cpu_var(cpu_hw_events);
1069 n = cpuhw->n_events;
1070 if (check_excludes(cpuhw->event, cpuhw->flags, 0, n))
1072 i = power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n);
1076 for (i = cpuhw->n_txn_start; i < n; ++i)
1077 cpuhw->event[i]->hw.config = cpuhw->events[i];
1079 cpuhw->group_flag &= ~PERF_EVENT_TXN;
1080 perf_pmu_enable(pmu);
1085 * Return 1 if we might be able to put event on a limited PMC,
1087 * A event can only go on a limited PMC if it counts something
1088 * that a limited PMC can count, doesn't require interrupts, and
1089 * doesn't exclude any processor mode.
1091 static int can_go_on_limited_pmc(struct perf_event *event, u64 ev,
1095 u64 alt[MAX_EVENT_ALTERNATIVES];
1097 if (event->attr.exclude_user
1098 || event->attr.exclude_kernel
1099 || event->attr.exclude_hv
1100 || event->attr.sample_period)
1103 if (ppmu->limited_pmc_event(ev))
1107 * The requested event_id isn't on a limited PMC already;
1108 * see if any alternative code goes on a limited PMC.
1110 if (!ppmu->get_alternatives)
1113 flags |= PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD;
1114 n = ppmu->get_alternatives(ev, flags, alt);
1120 * Find an alternative event_id that goes on a normal PMC, if possible,
1121 * and return the event_id code, or 0 if there is no such alternative.
1122 * (Note: event_id code 0 is "don't count" on all machines.)
1124 static u64 normal_pmc_alternative(u64 ev, unsigned long flags)
1126 u64 alt[MAX_EVENT_ALTERNATIVES];
1129 flags &= ~(PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD);
1130 n = ppmu->get_alternatives(ev, flags, alt);
1136 /* Number of perf_events counting hardware events */
1137 static atomic_t num_events;
1138 /* Used to avoid races in calling reserve/release_pmc_hardware */
1139 static DEFINE_MUTEX(pmc_reserve_mutex);
1142 * Release the PMU if this is the last perf_event.
1144 static void hw_perf_event_destroy(struct perf_event *event)
1146 if (!atomic_add_unless(&num_events, -1, 1)) {
1147 mutex_lock(&pmc_reserve_mutex);
1148 if (atomic_dec_return(&num_events) == 0)
1149 release_pmc_hardware();
1150 mutex_unlock(&pmc_reserve_mutex);
1155 * Translate a generic cache event_id config to a raw event_id code.
1157 static int hw_perf_cache_event(u64 config, u64 *eventp)
1159 unsigned long type, op, result;
1162 if (!ppmu->cache_events)
1166 type = config & 0xff;
1167 op = (config >> 8) & 0xff;
1168 result = (config >> 16) & 0xff;
1170 if (type >= PERF_COUNT_HW_CACHE_MAX ||
1171 op >= PERF_COUNT_HW_CACHE_OP_MAX ||
1172 result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1175 ev = (*ppmu->cache_events)[type][op][result];
1184 static int power_pmu_event_init(struct perf_event *event)
1187 unsigned long flags;
1188 struct perf_event *ctrs[MAX_HWEVENTS];
1189 u64 events[MAX_HWEVENTS];
1190 unsigned int cflags[MAX_HWEVENTS];
1193 struct cpu_hw_events *cpuhw;
1198 /* does not support taken branch sampling */
1199 if (has_branch_stack(event))
1202 switch (event->attr.type) {
1203 case PERF_TYPE_HARDWARE:
1204 ev = event->attr.config;
1205 if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
1207 ev = ppmu->generic_events[ev];
1209 case PERF_TYPE_HW_CACHE:
1210 err = hw_perf_cache_event(event->attr.config, &ev);
1215 ev = event->attr.config;
1221 event->hw.config_base = ev;
1225 * If we are not running on a hypervisor, force the
1226 * exclude_hv bit to 0 so that we don't care what
1227 * the user set it to.
1229 if (!firmware_has_feature(FW_FEATURE_LPAR))
1230 event->attr.exclude_hv = 0;
1233 * If this is a per-task event, then we can use
1234 * PM_RUN_* events interchangeably with their non RUN_*
1235 * equivalents, e.g. PM_RUN_CYC instead of PM_CYC.
1236 * XXX we should check if the task is an idle task.
1239 if (event->attach_state & PERF_ATTACH_TASK)
1240 flags |= PPMU_ONLY_COUNT_RUN;
1243 * If this machine has limited events, check whether this
1244 * event_id could go on a limited event.
1246 if (ppmu->flags & PPMU_LIMITED_PMC5_6) {
1247 if (can_go_on_limited_pmc(event, ev, flags)) {
1248 flags |= PPMU_LIMITED_PMC_OK;
1249 } else if (ppmu->limited_pmc_event(ev)) {
1251 * The requested event_id is on a limited PMC,
1252 * but we can't use a limited PMC; see if any
1253 * alternative goes on a normal PMC.
1255 ev = normal_pmc_alternative(ev, flags);
1262 * If this is in a group, check if it can go on with all the
1263 * other hardware events in the group. We assume the event
1264 * hasn't been linked into its leader's sibling list at this point.
1267 if (event->group_leader != event) {
1268 n = collect_events(event->group_leader, ppmu->n_counter - 1,
1269 ctrs, events, cflags);
1276 if (check_excludes(ctrs, cflags, n, 1))
1279 cpuhw = &get_cpu_var(cpu_hw_events);
1280 err = power_check_constraints(cpuhw, events, cflags, n + 1);
1281 put_cpu_var(cpu_hw_events);
1285 event->hw.config = events[n];
1286 event->hw.event_base = cflags[n];
1287 event->hw.last_period = event->hw.sample_period;
1288 local64_set(&event->hw.period_left, event->hw.last_period);
1291 * See if we need to reserve the PMU.
1292 * If no events are currently in use, then we have to take a
1293 * mutex to ensure that we don't race with another task doing
1294 * reserve_pmc_hardware or release_pmc_hardware.
1297 if (!atomic_inc_not_zero(&num_events)) {
1298 mutex_lock(&pmc_reserve_mutex);
1299 if (atomic_read(&num_events) == 0 &&
1300 reserve_pmc_hardware(perf_event_interrupt))
1303 atomic_inc(&num_events);
1304 mutex_unlock(&pmc_reserve_mutex);
1306 event->destroy = hw_perf_event_destroy;
1311 static int power_pmu_event_idx(struct perf_event *event)
1313 return event->hw.idx;
1316 ssize_t power_events_sysfs_show(struct device *dev,
1317 struct device_attribute *attr, char *page)
1319 struct perf_pmu_events_attr *pmu_attr;
1321 pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
1323 return sprintf(page, "event=0x%02llx\n", pmu_attr->id);
1326 struct pmu power_pmu = {
1327 .pmu_enable = power_pmu_enable,
1328 .pmu_disable = power_pmu_disable,
1329 .event_init = power_pmu_event_init,
1330 .add = power_pmu_add,
1331 .del = power_pmu_del,
1332 .start = power_pmu_start,
1333 .stop = power_pmu_stop,
1334 .read = power_pmu_read,
1335 .start_txn = power_pmu_start_txn,
1336 .cancel_txn = power_pmu_cancel_txn,
1337 .commit_txn = power_pmu_commit_txn,
1338 .event_idx = power_pmu_event_idx,
1343 * A counter has overflowed; update its count and record
1344 * things if requested. Note that interrupts are hard-disabled
1345 * here so there is no possibility of being interrupted.
1347 static void record_and_restart(struct perf_event *event, unsigned long val,
1348 struct pt_regs *regs)
1350 u64 period = event->hw.sample_period;
1351 s64 prev, delta, left;
1354 if (event->hw.state & PERF_HES_STOPPED) {
1355 write_pmc(event->hw.idx, 0);
1359 /* we don't have to worry about interrupts here */
1360 prev = local64_read(&event->hw.prev_count);
1361 delta = check_and_compute_delta(prev, val);
1362 local64_add(delta, &event->count);
1365 * See if the total period for this event has expired,
1366 * and update for the next period.
1369 left = local64_read(&event->hw.period_left) - delta;
1377 record = siar_valid(regs);
1378 event->hw.last_period = event->hw.sample_period;
1380 if (left < 0x80000000LL)
1381 val = 0x80000000LL - left;
1384 write_pmc(event->hw.idx, val);
1385 local64_set(&event->hw.prev_count, val);
1386 local64_set(&event->hw.period_left, left);
1387 perf_event_update_userpage(event);
1390 * Finally record data if requested.
1393 struct perf_sample_data data;
1395 perf_sample_data_init(&data, ~0ULL, event->hw.last_period);
1397 if (event->attr.sample_type & PERF_SAMPLE_ADDR)
1398 perf_get_data_addr(regs, &data.addr);
1400 if (perf_event_overflow(event, &data, regs))
1401 power_pmu_stop(event, 0);
1406 * Called from generic code to get the misc flags (i.e. processor mode)
1409 unsigned long perf_misc_flags(struct pt_regs *regs)
1411 u32 flags = perf_get_misc_flags(regs);
1415 return user_mode(regs) ? PERF_RECORD_MISC_USER :
1416 PERF_RECORD_MISC_KERNEL;
1420 * Called from generic code to get the instruction pointer
1423 unsigned long perf_instruction_pointer(struct pt_regs *regs)
1425 unsigned long use_siar = regs->result;
1427 if (use_siar && siar_valid(regs))
1428 return mfspr(SPRN_SIAR) + perf_ip_adjust(regs);
1430 return 0; // no valid instruction pointer
1435 static bool pmc_overflow_power7(unsigned long val)
1438 * Events on POWER7 can roll back if a speculative event doesn't
1439 * eventually complete. Unfortunately in some rare cases they will
1440 * raise a performance monitor exception. We need to catch this to
1441 * ensure we reset the PMC. In all cases the PMC will be 256 or less
1442 * cycles from overflow.
1444 * We only do this if the first pass fails to find any overflowing
1445 * PMCs because a user might set a period of less than 256 and we
1446 * don't want to mistakenly reset them.
1448 if ((0x80000000 - val) <= 256)
1454 static bool pmc_overflow(unsigned long val)
1463 * Performance monitor interrupt stuff
1465 static void perf_event_interrupt(struct pt_regs *regs)
1468 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1469 struct perf_event *event;
1470 unsigned long val[8];
1474 if (cpuhw->n_limited)
1475 freeze_limited_counters(cpuhw, mfspr(SPRN_PMC5),
1478 perf_read_regs(regs);
1480 nmi = perf_intr_is_nmi(regs);
1486 /* Read all the PMCs since we'll need them a bunch of times */
1487 for (i = 0; i < ppmu->n_counter; ++i)
1488 val[i] = read_pmc(i + 1);
1490 /* Try to find what caused the IRQ */
1492 for (i = 0; i < ppmu->n_counter; ++i) {
1493 if (!pmc_overflow(val[i]))
1495 if (is_limited_pmc(i + 1))
1496 continue; /* these won't generate IRQs */
1498 * We've found one that's overflowed. For active
1499 * counters we need to log this. For inactive
1500 * counters, we need to reset it anyway
1504 for (j = 0; j < cpuhw->n_events; ++j) {
1505 event = cpuhw->event[j];
1506 if (event->hw.idx == (i + 1)) {
1508 record_and_restart(event, val[i], regs);
1513 /* reset non active counters that have overflowed */
1514 write_pmc(i + 1, 0);
1516 if (!found && pvr_version_is(PVR_POWER7)) {
1517 /* check active counters for special buggy p7 overflow */
1518 for (i = 0; i < cpuhw->n_events; ++i) {
1519 event = cpuhw->event[i];
1520 if (!event->hw.idx || is_limited_pmc(event->hw.idx))
1522 if (pmc_overflow_power7(val[event->hw.idx - 1])) {
1523 /* event has overflowed in a buggy way*/
1525 record_and_restart(event,
1526 val[event->hw.idx - 1],
1531 if ((!found) && printk_ratelimit())
1532 printk(KERN_WARNING "Can't find PMC that caused IRQ\n");
1535 * Reset MMCR0 to its normal value. This will set PMXE and
1536 * clear FC (freeze counters) and PMAO (perf mon alert occurred)
1537 * and thus allow interrupts to occur again.
1538 * XXX might want to use MSR.PM to keep the events frozen until
1539 * we get back out of this interrupt.
1541 write_mmcr0(cpuhw, cpuhw->mmcr[0]);
1549 static void power_pmu_setup(int cpu)
1551 struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
1555 memset(cpuhw, 0, sizeof(*cpuhw));
1556 cpuhw->mmcr[0] = MMCR0_FC;
1559 static int __cpuinit
1560 power_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1562 unsigned int cpu = (long)hcpu;
1564 switch (action & ~CPU_TASKS_FROZEN) {
1565 case CPU_UP_PREPARE:
1566 power_pmu_setup(cpu);
1576 int __cpuinit register_power_pmu(struct power_pmu *pmu)
1579 return -EBUSY; /* something's already registered */
1582 pr_info("%s performance monitor hardware support registered\n",
1585 power_pmu.attr_groups = ppmu->attr_groups;
1589 * Use FCHV to ignore kernel events if MSR.HV is set.
1591 if (mfmsr() & MSR_HV)
1592 freeze_events_kernel = MMCR0_FCHV;
1593 #endif /* CONFIG_PPC64 */
1595 perf_pmu_register(&power_pmu, "cpu", PERF_TYPE_RAW);
1596 perf_cpu_notifier(power_pmu_notifier);