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 */
92 static bool regs_use_siar(struct pt_regs *regs)
94 return !!(regs->result & 1);
98 * Things that are specific to 64-bit implementations.
102 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
104 unsigned long mmcra = regs->dsisr;
106 if ((ppmu->flags & PPMU_HAS_SSLOT) && (mmcra & MMCRA_SAMPLE_ENABLE)) {
107 unsigned long slot = (mmcra & MMCRA_SLOT) >> MMCRA_SLOT_SHIFT;
109 return 4 * (slot - 1);
116 * The user wants a data address recorded.
117 * If we're not doing instruction sampling, give them the SDAR
118 * (sampled data address). If we are doing instruction sampling, then
119 * only give them the SDAR if it corresponds to the instruction
120 * pointed to by SIAR; this is indicated by the [POWER6_]MMCRA_SDSYNC or
121 * the [POWER7P_]MMCRA_SDAR_VALID bit in MMCRA.
123 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp)
125 unsigned long mmcra = regs->dsisr;
126 unsigned long sdsync;
128 if (ppmu->flags & PPMU_SIAR_VALID)
129 sdsync = POWER7P_MMCRA_SDAR_VALID;
130 else if (ppmu->flags & PPMU_ALT_SIPR)
131 sdsync = POWER6_MMCRA_SDSYNC;
133 sdsync = MMCRA_SDSYNC;
135 if (!(mmcra & MMCRA_SAMPLE_ENABLE) || (mmcra & sdsync))
136 *addrp = mfspr(SPRN_SDAR);
139 static bool regs_sihv(struct pt_regs *regs)
141 unsigned long sihv = MMCRA_SIHV;
143 if (ppmu->flags & PPMU_ALT_SIPR)
144 sihv = POWER6_MMCRA_SIHV;
146 return !!(regs->dsisr & sihv);
149 static bool regs_sipr(struct pt_regs *regs)
151 unsigned long sipr = MMCRA_SIPR;
153 if (ppmu->flags & PPMU_ALT_SIPR)
154 sipr = POWER6_MMCRA_SIPR;
156 return !!(regs->dsisr & sipr);
159 static inline u32 perf_flags_from_msr(struct pt_regs *regs)
161 if (regs->msr & MSR_PR)
162 return PERF_RECORD_MISC_USER;
163 if ((regs->msr & MSR_HV) && freeze_events_kernel != MMCR0_FCHV)
164 return PERF_RECORD_MISC_HYPERVISOR;
165 return PERF_RECORD_MISC_KERNEL;
168 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
170 bool use_siar = regs_use_siar(regs);
173 return perf_flags_from_msr(regs);
176 * If we don't have flags in MMCRA, rather than using
177 * the MSR, we intuit the flags from the address in
178 * SIAR which should give slightly more reliable
181 if (ppmu->flags & PPMU_NO_SIPR) {
182 unsigned long siar = mfspr(SPRN_SIAR);
183 if (siar >= PAGE_OFFSET)
184 return PERF_RECORD_MISC_KERNEL;
185 return PERF_RECORD_MISC_USER;
188 /* PR has priority over HV, so order below is important */
190 return PERF_RECORD_MISC_USER;
192 if (regs_sihv(regs) && (freeze_events_kernel != MMCR0_FCHV))
193 return PERF_RECORD_MISC_HYPERVISOR;
195 return PERF_RECORD_MISC_KERNEL;
199 * Overload regs->dsisr to store MMCRA so we only need to read it once
201 * Overload regs->result to specify whether we should use the MSR (result
202 * is zero) or the SIAR (result is non zero).
204 static inline void perf_read_regs(struct pt_regs *regs)
206 unsigned long mmcra = mfspr(SPRN_MMCRA);
207 int marked = mmcra & MMCRA_SAMPLE_ENABLE;
213 * If this isn't a PMU exception (eg a software event) the SIAR is
214 * not valid. Use pt_regs.
216 * If it is a marked event use the SIAR.
218 * If the PMU doesn't update the SIAR for non marked events use
221 * If the PMU has HV/PR flags then check to see if they
222 * place the exception in userspace. If so, use pt_regs. In
223 * continuous sampling mode the SIAR and the PMU exception are
224 * not synchronised, so they may be many instructions apart.
225 * This can result in confusing backtraces. We still want
226 * hypervisor samples as well as samples in the kernel with
227 * interrupts off hence the userspace check.
229 if (TRAP(regs) != 0xf00)
233 else if ((ppmu->flags & PPMU_NO_CONT_SAMPLING))
235 else if (!(ppmu->flags & PPMU_NO_SIPR) && regs_sipr(regs))
240 regs->result = use_siar;
244 * If interrupts were soft-disabled when a PMU interrupt occurs, treat
247 static inline int perf_intr_is_nmi(struct pt_regs *regs)
253 * On processors like P7+ that have the SIAR-Valid bit, marked instructions
254 * must be sampled only if the SIAR-valid bit is set.
256 * For unmarked instructions and for processors that don't have the SIAR-Valid
257 * bit, assume that SIAR is valid.
259 static inline int siar_valid(struct pt_regs *regs)
261 unsigned long mmcra = regs->dsisr;
262 int marked = mmcra & MMCRA_SAMPLE_ENABLE;
264 if ((ppmu->flags & PPMU_SIAR_VALID) && marked)
265 return mmcra & POWER7P_MMCRA_SIAR_VALID;
270 #endif /* CONFIG_PPC64 */
272 static void perf_event_interrupt(struct pt_regs *regs);
274 void perf_event_print_debug(void)
279 * Read one performance monitor counter (PMC).
281 static unsigned long read_pmc(int idx)
287 val = mfspr(SPRN_PMC1);
290 val = mfspr(SPRN_PMC2);
293 val = mfspr(SPRN_PMC3);
296 val = mfspr(SPRN_PMC4);
299 val = mfspr(SPRN_PMC5);
302 val = mfspr(SPRN_PMC6);
306 val = mfspr(SPRN_PMC7);
309 val = mfspr(SPRN_PMC8);
311 #endif /* CONFIG_PPC64 */
313 printk(KERN_ERR "oops trying to read PMC%d\n", idx);
322 static void write_pmc(int idx, unsigned long val)
326 mtspr(SPRN_PMC1, val);
329 mtspr(SPRN_PMC2, val);
332 mtspr(SPRN_PMC3, val);
335 mtspr(SPRN_PMC4, val);
338 mtspr(SPRN_PMC5, val);
341 mtspr(SPRN_PMC6, val);
345 mtspr(SPRN_PMC7, val);
348 mtspr(SPRN_PMC8, val);
350 #endif /* CONFIG_PPC64 */
352 printk(KERN_ERR "oops trying to write PMC%d\n", idx);
357 * Check if a set of events can all go on the PMU at once.
358 * If they can't, this will look at alternative codes for the events
359 * and see if any combination of alternative codes is feasible.
360 * The feasible set is returned in event_id[].
362 static int power_check_constraints(struct cpu_hw_events *cpuhw,
363 u64 event_id[], unsigned int cflags[],
366 unsigned long mask, value, nv;
367 unsigned long smasks[MAX_HWEVENTS], svalues[MAX_HWEVENTS];
368 int n_alt[MAX_HWEVENTS], choice[MAX_HWEVENTS];
370 unsigned long addf = ppmu->add_fields;
371 unsigned long tadd = ppmu->test_adder;
373 if (n_ev > ppmu->n_counter)
376 /* First see if the events will go on as-is */
377 for (i = 0; i < n_ev; ++i) {
378 if ((cflags[i] & PPMU_LIMITED_PMC_REQD)
379 && !ppmu->limited_pmc_event(event_id[i])) {
380 ppmu->get_alternatives(event_id[i], cflags[i],
381 cpuhw->alternatives[i]);
382 event_id[i] = cpuhw->alternatives[i][0];
384 if (ppmu->get_constraint(event_id[i], &cpuhw->amasks[i][0],
385 &cpuhw->avalues[i][0]))
389 for (i = 0; i < n_ev; ++i) {
390 nv = (value | cpuhw->avalues[i][0]) +
391 (value & cpuhw->avalues[i][0] & addf);
392 if ((((nv + tadd) ^ value) & mask) != 0 ||
393 (((nv + tadd) ^ cpuhw->avalues[i][0]) &
394 cpuhw->amasks[i][0]) != 0)
397 mask |= cpuhw->amasks[i][0];
400 return 0; /* all OK */
402 /* doesn't work, gather alternatives... */
403 if (!ppmu->get_alternatives)
405 for (i = 0; i < n_ev; ++i) {
407 n_alt[i] = ppmu->get_alternatives(event_id[i], cflags[i],
408 cpuhw->alternatives[i]);
409 for (j = 1; j < n_alt[i]; ++j)
410 ppmu->get_constraint(cpuhw->alternatives[i][j],
411 &cpuhw->amasks[i][j],
412 &cpuhw->avalues[i][j]);
415 /* enumerate all possibilities and see if any will work */
418 value = mask = nv = 0;
421 /* we're backtracking, restore context */
427 * See if any alternative k for event_id i,
428 * where k > j, will satisfy the constraints.
430 while (++j < n_alt[i]) {
431 nv = (value | cpuhw->avalues[i][j]) +
432 (value & cpuhw->avalues[i][j] & addf);
433 if ((((nv + tadd) ^ value) & mask) == 0 &&
434 (((nv + tadd) ^ cpuhw->avalues[i][j])
435 & cpuhw->amasks[i][j]) == 0)
440 * No feasible alternative, backtrack
441 * to event_id i-1 and continue enumerating its
442 * alternatives from where we got up to.
448 * Found a feasible alternative for event_id i,
449 * remember where we got up to with this event_id,
450 * go on to the next event_id, and start with
451 * the first alternative for it.
457 mask |= cpuhw->amasks[i][j];
463 /* OK, we have a feasible combination, tell the caller the solution */
464 for (i = 0; i < n_ev; ++i)
465 event_id[i] = cpuhw->alternatives[i][choice[i]];
470 * Check if newly-added events have consistent settings for
471 * exclude_{user,kernel,hv} with each other and any previously
474 static int check_excludes(struct perf_event **ctrs, unsigned int cflags[],
475 int n_prev, int n_new)
477 int eu = 0, ek = 0, eh = 0;
479 struct perf_event *event;
486 for (i = 0; i < n; ++i) {
487 if (cflags[i] & PPMU_LIMITED_PMC_OK) {
488 cflags[i] &= ~PPMU_LIMITED_PMC_REQD;
493 eu = event->attr.exclude_user;
494 ek = event->attr.exclude_kernel;
495 eh = event->attr.exclude_hv;
497 } else if (event->attr.exclude_user != eu ||
498 event->attr.exclude_kernel != ek ||
499 event->attr.exclude_hv != eh) {
505 for (i = 0; i < n; ++i)
506 if (cflags[i] & PPMU_LIMITED_PMC_OK)
507 cflags[i] |= PPMU_LIMITED_PMC_REQD;
512 static u64 check_and_compute_delta(u64 prev, u64 val)
514 u64 delta = (val - prev) & 0xfffffffful;
517 * POWER7 can roll back counter values, if the new value is smaller
518 * than the previous value it will cause the delta and the counter to
519 * have bogus values unless we rolled a counter over. If a coutner is
520 * rolled back, it will be smaller, but within 256, which is the maximum
521 * number of events to rollback at once. If we dectect a rollback
522 * return 0. This can lead to a small lack of precision in the
525 if (prev > val && (prev - val) < 256)
531 static void power_pmu_read(struct perf_event *event)
533 s64 val, delta, prev;
535 if (event->hw.state & PERF_HES_STOPPED)
541 * Performance monitor interrupts come even when interrupts
542 * are soft-disabled, as long as interrupts are hard-enabled.
543 * Therefore we treat them like NMIs.
546 prev = local64_read(&event->hw.prev_count);
548 val = read_pmc(event->hw.idx);
549 delta = check_and_compute_delta(prev, val);
552 } while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
554 local64_add(delta, &event->count);
555 local64_sub(delta, &event->hw.period_left);
559 * On some machines, PMC5 and PMC6 can't be written, don't respect
560 * the freeze conditions, and don't generate interrupts. This tells
561 * us if `event' is using such a PMC.
563 static int is_limited_pmc(int pmcnum)
565 return (ppmu->flags & PPMU_LIMITED_PMC5_6)
566 && (pmcnum == 5 || pmcnum == 6);
569 static void freeze_limited_counters(struct cpu_hw_events *cpuhw,
570 unsigned long pmc5, unsigned long pmc6)
572 struct perf_event *event;
573 u64 val, prev, delta;
576 for (i = 0; i < cpuhw->n_limited; ++i) {
577 event = cpuhw->limited_counter[i];
580 val = (event->hw.idx == 5) ? pmc5 : pmc6;
581 prev = local64_read(&event->hw.prev_count);
583 delta = check_and_compute_delta(prev, val);
585 local64_add(delta, &event->count);
589 static void thaw_limited_counters(struct cpu_hw_events *cpuhw,
590 unsigned long pmc5, unsigned long pmc6)
592 struct perf_event *event;
596 for (i = 0; i < cpuhw->n_limited; ++i) {
597 event = cpuhw->limited_counter[i];
598 event->hw.idx = cpuhw->limited_hwidx[i];
599 val = (event->hw.idx == 5) ? pmc5 : pmc6;
600 prev = local64_read(&event->hw.prev_count);
601 if (check_and_compute_delta(prev, val))
602 local64_set(&event->hw.prev_count, val);
603 perf_event_update_userpage(event);
608 * Since limited events don't respect the freeze conditions, we
609 * have to read them immediately after freezing or unfreezing the
610 * other events. We try to keep the values from the limited
611 * events as consistent as possible by keeping the delay (in
612 * cycles and instructions) between freezing/unfreezing and reading
613 * the limited events as small and consistent as possible.
614 * Therefore, if any limited events are in use, we read them
615 * both, and always in the same order, to minimize variability,
616 * and do it inside the same asm that writes MMCR0.
618 static void write_mmcr0(struct cpu_hw_events *cpuhw, unsigned long mmcr0)
620 unsigned long pmc5, pmc6;
622 if (!cpuhw->n_limited) {
623 mtspr(SPRN_MMCR0, mmcr0);
628 * Write MMCR0, then read PMC5 and PMC6 immediately.
629 * To ensure we don't get a performance monitor interrupt
630 * between writing MMCR0 and freezing/thawing the limited
631 * events, we first write MMCR0 with the event overflow
632 * interrupt enable bits turned off.
634 asm volatile("mtspr %3,%2; mfspr %0,%4; mfspr %1,%5"
635 : "=&r" (pmc5), "=&r" (pmc6)
636 : "r" (mmcr0 & ~(MMCR0_PMC1CE | MMCR0_PMCjCE)),
638 "i" (SPRN_PMC5), "i" (SPRN_PMC6));
640 if (mmcr0 & MMCR0_FC)
641 freeze_limited_counters(cpuhw, pmc5, pmc6);
643 thaw_limited_counters(cpuhw, pmc5, pmc6);
646 * Write the full MMCR0 including the event overflow interrupt
647 * enable bits, if necessary.
649 if (mmcr0 & (MMCR0_PMC1CE | MMCR0_PMCjCE))
650 mtspr(SPRN_MMCR0, mmcr0);
654 * Disable all events to prevent PMU interrupts and to allow
655 * events to be added or removed.
657 static void power_pmu_disable(struct pmu *pmu)
659 struct cpu_hw_events *cpuhw;
664 local_irq_save(flags);
665 cpuhw = &__get_cpu_var(cpu_hw_events);
667 if (!cpuhw->disabled) {
672 * Check if we ever enabled the PMU on this cpu.
674 if (!cpuhw->pmcs_enabled) {
676 cpuhw->pmcs_enabled = 1;
680 * Disable instruction sampling if it was enabled
682 if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
684 cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
689 * Set the 'freeze counters' bit.
690 * The barrier is to make sure the mtspr has been
691 * executed and the PMU has frozen the events
694 write_mmcr0(cpuhw, mfspr(SPRN_MMCR0) | MMCR0_FC);
697 local_irq_restore(flags);
701 * Re-enable all events if disable == 0.
702 * If we were previously disabled and events were added, then
703 * put the new config on the PMU.
705 static void power_pmu_enable(struct pmu *pmu)
707 struct perf_event *event;
708 struct cpu_hw_events *cpuhw;
713 unsigned int hwc_index[MAX_HWEVENTS];
719 local_irq_save(flags);
720 cpuhw = &__get_cpu_var(cpu_hw_events);
721 if (!cpuhw->disabled) {
722 local_irq_restore(flags);
728 * If we didn't change anything, or only removed events,
729 * no need to recalculate MMCR* settings and reset the PMCs.
730 * Just reenable the PMU with the current MMCR* settings
731 * (possibly updated for removal of events).
733 if (!cpuhw->n_added) {
734 mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
735 mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
736 if (cpuhw->n_events == 0)
737 ppc_set_pmu_inuse(0);
742 * Compute MMCR* values for the new set of events
744 if (ppmu->compute_mmcr(cpuhw->events, cpuhw->n_events, hwc_index,
746 /* shouldn't ever get here */
747 printk(KERN_ERR "oops compute_mmcr failed\n");
752 * Add in MMCR0 freeze bits corresponding to the
753 * attr.exclude_* bits for the first event.
754 * We have already checked that all events have the
755 * same values for these bits as the first event.
757 event = cpuhw->event[0];
758 if (event->attr.exclude_user)
759 cpuhw->mmcr[0] |= MMCR0_FCP;
760 if (event->attr.exclude_kernel)
761 cpuhw->mmcr[0] |= freeze_events_kernel;
762 if (event->attr.exclude_hv)
763 cpuhw->mmcr[0] |= MMCR0_FCHV;
766 * Write the new configuration to MMCR* with the freeze
767 * bit set and set the hardware events to their initial values.
768 * Then unfreeze the events.
770 ppc_set_pmu_inuse(1);
771 mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
772 mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
773 mtspr(SPRN_MMCR0, (cpuhw->mmcr[0] & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
777 * Read off any pre-existing events that need to move
780 for (i = 0; i < cpuhw->n_events; ++i) {
781 event = cpuhw->event[i];
782 if (event->hw.idx && event->hw.idx != hwc_index[i] + 1) {
783 power_pmu_read(event);
784 write_pmc(event->hw.idx, 0);
790 * Initialize the PMCs for all the new and moved events.
792 cpuhw->n_limited = n_lim = 0;
793 for (i = 0; i < cpuhw->n_events; ++i) {
794 event = cpuhw->event[i];
797 idx = hwc_index[i] + 1;
798 if (is_limited_pmc(idx)) {
799 cpuhw->limited_counter[n_lim] = event;
800 cpuhw->limited_hwidx[n_lim] = idx;
805 if (event->hw.sample_period) {
806 left = local64_read(&event->hw.period_left);
807 if (left < 0x80000000L)
808 val = 0x80000000L - left;
810 local64_set(&event->hw.prev_count, val);
812 if (event->hw.state & PERF_HES_STOPPED)
815 perf_event_update_userpage(event);
817 cpuhw->n_limited = n_lim;
818 cpuhw->mmcr[0] |= MMCR0_PMXE | MMCR0_FCECE;
822 write_mmcr0(cpuhw, cpuhw->mmcr[0]);
825 * Enable instruction sampling if necessary
827 if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
829 mtspr(SPRN_MMCRA, cpuhw->mmcr[2]);
833 local_irq_restore(flags);
836 static int collect_events(struct perf_event *group, int max_count,
837 struct perf_event *ctrs[], u64 *events,
841 struct perf_event *event;
843 if (!is_software_event(group)) {
847 flags[n] = group->hw.event_base;
848 events[n++] = group->hw.config;
850 list_for_each_entry(event, &group->sibling_list, group_entry) {
851 if (!is_software_event(event) &&
852 event->state != PERF_EVENT_STATE_OFF) {
856 flags[n] = event->hw.event_base;
857 events[n++] = event->hw.config;
864 * Add a event to the PMU.
865 * If all events are not already frozen, then we disable and
866 * re-enable the PMU in order to get hw_perf_enable to do the
867 * actual work of reconfiguring the PMU.
869 static int power_pmu_add(struct perf_event *event, int ef_flags)
871 struct cpu_hw_events *cpuhw;
876 local_irq_save(flags);
877 perf_pmu_disable(event->pmu);
880 * Add the event to the list (if there is room)
881 * and check whether the total set is still feasible.
883 cpuhw = &__get_cpu_var(cpu_hw_events);
884 n0 = cpuhw->n_events;
885 if (n0 >= ppmu->n_counter)
887 cpuhw->event[n0] = event;
888 cpuhw->events[n0] = event->hw.config;
889 cpuhw->flags[n0] = event->hw.event_base;
892 * This event may have been disabled/stopped in record_and_restart()
893 * because we exceeded the ->event_limit. If re-starting the event,
894 * clear the ->hw.state (STOPPED and UPTODATE flags), so the user
895 * notification is re-enabled.
897 if (!(ef_flags & PERF_EF_START))
898 event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
903 * If group events scheduling transaction was started,
904 * skip the schedulability test here, it will be performed
905 * at commit time(->commit_txn) as a whole
907 if (cpuhw->group_flag & PERF_EVENT_TXN)
910 if (check_excludes(cpuhw->event, cpuhw->flags, n0, 1))
912 if (power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n0 + 1))
914 event->hw.config = cpuhw->events[n0];
922 perf_pmu_enable(event->pmu);
923 local_irq_restore(flags);
928 * Remove a event from the PMU.
930 static void power_pmu_del(struct perf_event *event, int ef_flags)
932 struct cpu_hw_events *cpuhw;
936 local_irq_save(flags);
937 perf_pmu_disable(event->pmu);
939 power_pmu_read(event);
941 cpuhw = &__get_cpu_var(cpu_hw_events);
942 for (i = 0; i < cpuhw->n_events; ++i) {
943 if (event == cpuhw->event[i]) {
944 while (++i < cpuhw->n_events) {
945 cpuhw->event[i-1] = cpuhw->event[i];
946 cpuhw->events[i-1] = cpuhw->events[i];
947 cpuhw->flags[i-1] = cpuhw->flags[i];
950 ppmu->disable_pmc(event->hw.idx - 1, cpuhw->mmcr);
952 write_pmc(event->hw.idx, 0);
955 perf_event_update_userpage(event);
959 for (i = 0; i < cpuhw->n_limited; ++i)
960 if (event == cpuhw->limited_counter[i])
962 if (i < cpuhw->n_limited) {
963 while (++i < cpuhw->n_limited) {
964 cpuhw->limited_counter[i-1] = cpuhw->limited_counter[i];
965 cpuhw->limited_hwidx[i-1] = cpuhw->limited_hwidx[i];
969 if (cpuhw->n_events == 0) {
970 /* disable exceptions if no events are running */
971 cpuhw->mmcr[0] &= ~(MMCR0_PMXE | MMCR0_FCECE);
974 perf_pmu_enable(event->pmu);
975 local_irq_restore(flags);
979 * POWER-PMU does not support disabling individual counters, hence
980 * program their cycle counter to their max value and ignore the interrupts.
983 static void power_pmu_start(struct perf_event *event, int ef_flags)
989 if (!event->hw.idx || !event->hw.sample_period)
992 if (!(event->hw.state & PERF_HES_STOPPED))
995 if (ef_flags & PERF_EF_RELOAD)
996 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
998 local_irq_save(flags);
999 perf_pmu_disable(event->pmu);
1001 event->hw.state = 0;
1002 left = local64_read(&event->hw.period_left);
1005 if (left < 0x80000000L)
1006 val = 0x80000000L - left;
1008 write_pmc(event->hw.idx, val);
1010 perf_event_update_userpage(event);
1011 perf_pmu_enable(event->pmu);
1012 local_irq_restore(flags);
1015 static void power_pmu_stop(struct perf_event *event, int ef_flags)
1017 unsigned long flags;
1019 if (!event->hw.idx || !event->hw.sample_period)
1022 if (event->hw.state & PERF_HES_STOPPED)
1025 local_irq_save(flags);
1026 perf_pmu_disable(event->pmu);
1028 power_pmu_read(event);
1029 event->hw.state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1030 write_pmc(event->hw.idx, 0);
1032 perf_event_update_userpage(event);
1033 perf_pmu_enable(event->pmu);
1034 local_irq_restore(flags);
1038 * Start group events scheduling transaction
1039 * Set the flag to make pmu::enable() not perform the
1040 * schedulability test, it will be performed at commit time
1042 void power_pmu_start_txn(struct pmu *pmu)
1044 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1046 perf_pmu_disable(pmu);
1047 cpuhw->group_flag |= PERF_EVENT_TXN;
1048 cpuhw->n_txn_start = cpuhw->n_events;
1052 * Stop group events scheduling transaction
1053 * Clear the flag and pmu::enable() will perform the
1054 * schedulability test.
1056 void power_pmu_cancel_txn(struct pmu *pmu)
1058 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1060 cpuhw->group_flag &= ~PERF_EVENT_TXN;
1061 perf_pmu_enable(pmu);
1065 * Commit group events scheduling transaction
1066 * Perform the group schedulability test as a whole
1067 * Return 0 if success
1069 int power_pmu_commit_txn(struct pmu *pmu)
1071 struct cpu_hw_events *cpuhw;
1076 cpuhw = &__get_cpu_var(cpu_hw_events);
1077 n = cpuhw->n_events;
1078 if (check_excludes(cpuhw->event, cpuhw->flags, 0, n))
1080 i = power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n);
1084 for (i = cpuhw->n_txn_start; i < n; ++i)
1085 cpuhw->event[i]->hw.config = cpuhw->events[i];
1087 cpuhw->group_flag &= ~PERF_EVENT_TXN;
1088 perf_pmu_enable(pmu);
1093 * Return 1 if we might be able to put event on a limited PMC,
1095 * A event can only go on a limited PMC if it counts something
1096 * that a limited PMC can count, doesn't require interrupts, and
1097 * doesn't exclude any processor mode.
1099 static int can_go_on_limited_pmc(struct perf_event *event, u64 ev,
1103 u64 alt[MAX_EVENT_ALTERNATIVES];
1105 if (event->attr.exclude_user
1106 || event->attr.exclude_kernel
1107 || event->attr.exclude_hv
1108 || event->attr.sample_period)
1111 if (ppmu->limited_pmc_event(ev))
1115 * The requested event_id isn't on a limited PMC already;
1116 * see if any alternative code goes on a limited PMC.
1118 if (!ppmu->get_alternatives)
1121 flags |= PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD;
1122 n = ppmu->get_alternatives(ev, flags, alt);
1128 * Find an alternative event_id that goes on a normal PMC, if possible,
1129 * and return the event_id code, or 0 if there is no such alternative.
1130 * (Note: event_id code 0 is "don't count" on all machines.)
1132 static u64 normal_pmc_alternative(u64 ev, unsigned long flags)
1134 u64 alt[MAX_EVENT_ALTERNATIVES];
1137 flags &= ~(PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD);
1138 n = ppmu->get_alternatives(ev, flags, alt);
1144 /* Number of perf_events counting hardware events */
1145 static atomic_t num_events;
1146 /* Used to avoid races in calling reserve/release_pmc_hardware */
1147 static DEFINE_MUTEX(pmc_reserve_mutex);
1150 * Release the PMU if this is the last perf_event.
1152 static void hw_perf_event_destroy(struct perf_event *event)
1154 if (!atomic_add_unless(&num_events, -1, 1)) {
1155 mutex_lock(&pmc_reserve_mutex);
1156 if (atomic_dec_return(&num_events) == 0)
1157 release_pmc_hardware();
1158 mutex_unlock(&pmc_reserve_mutex);
1163 * Translate a generic cache event_id config to a raw event_id code.
1165 static int hw_perf_cache_event(u64 config, u64 *eventp)
1167 unsigned long type, op, result;
1170 if (!ppmu->cache_events)
1174 type = config & 0xff;
1175 op = (config >> 8) & 0xff;
1176 result = (config >> 16) & 0xff;
1178 if (type >= PERF_COUNT_HW_CACHE_MAX ||
1179 op >= PERF_COUNT_HW_CACHE_OP_MAX ||
1180 result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1183 ev = (*ppmu->cache_events)[type][op][result];
1192 static int power_pmu_event_init(struct perf_event *event)
1195 unsigned long flags;
1196 struct perf_event *ctrs[MAX_HWEVENTS];
1197 u64 events[MAX_HWEVENTS];
1198 unsigned int cflags[MAX_HWEVENTS];
1201 struct cpu_hw_events *cpuhw;
1206 /* does not support taken branch sampling */
1207 if (has_branch_stack(event))
1210 switch (event->attr.type) {
1211 case PERF_TYPE_HARDWARE:
1212 ev = event->attr.config;
1213 if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
1215 ev = ppmu->generic_events[ev];
1217 case PERF_TYPE_HW_CACHE:
1218 err = hw_perf_cache_event(event->attr.config, &ev);
1223 ev = event->attr.config;
1229 event->hw.config_base = ev;
1233 * If we are not running on a hypervisor, force the
1234 * exclude_hv bit to 0 so that we don't care what
1235 * the user set it to.
1237 if (!firmware_has_feature(FW_FEATURE_LPAR))
1238 event->attr.exclude_hv = 0;
1241 * If this is a per-task event, then we can use
1242 * PM_RUN_* events interchangeably with their non RUN_*
1243 * equivalents, e.g. PM_RUN_CYC instead of PM_CYC.
1244 * XXX we should check if the task is an idle task.
1247 if (event->attach_state & PERF_ATTACH_TASK)
1248 flags |= PPMU_ONLY_COUNT_RUN;
1251 * If this machine has limited events, check whether this
1252 * event_id could go on a limited event.
1254 if (ppmu->flags & PPMU_LIMITED_PMC5_6) {
1255 if (can_go_on_limited_pmc(event, ev, flags)) {
1256 flags |= PPMU_LIMITED_PMC_OK;
1257 } else if (ppmu->limited_pmc_event(ev)) {
1259 * The requested event_id is on a limited PMC,
1260 * but we can't use a limited PMC; see if any
1261 * alternative goes on a normal PMC.
1263 ev = normal_pmc_alternative(ev, flags);
1270 * If this is in a group, check if it can go on with all the
1271 * other hardware events in the group. We assume the event
1272 * hasn't been linked into its leader's sibling list at this point.
1275 if (event->group_leader != event) {
1276 n = collect_events(event->group_leader, ppmu->n_counter - 1,
1277 ctrs, events, cflags);
1284 if (check_excludes(ctrs, cflags, n, 1))
1287 cpuhw = &get_cpu_var(cpu_hw_events);
1288 err = power_check_constraints(cpuhw, events, cflags, n + 1);
1289 put_cpu_var(cpu_hw_events);
1293 event->hw.config = events[n];
1294 event->hw.event_base = cflags[n];
1295 event->hw.last_period = event->hw.sample_period;
1296 local64_set(&event->hw.period_left, event->hw.last_period);
1299 * See if we need to reserve the PMU.
1300 * If no events are currently in use, then we have to take a
1301 * mutex to ensure that we don't race with another task doing
1302 * reserve_pmc_hardware or release_pmc_hardware.
1305 if (!atomic_inc_not_zero(&num_events)) {
1306 mutex_lock(&pmc_reserve_mutex);
1307 if (atomic_read(&num_events) == 0 &&
1308 reserve_pmc_hardware(perf_event_interrupt))
1311 atomic_inc(&num_events);
1312 mutex_unlock(&pmc_reserve_mutex);
1314 event->destroy = hw_perf_event_destroy;
1319 static int power_pmu_event_idx(struct perf_event *event)
1321 return event->hw.idx;
1324 ssize_t power_events_sysfs_show(struct device *dev,
1325 struct device_attribute *attr, char *page)
1327 struct perf_pmu_events_attr *pmu_attr;
1329 pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
1331 return sprintf(page, "event=0x%02llx\n", pmu_attr->id);
1334 struct pmu power_pmu = {
1335 .pmu_enable = power_pmu_enable,
1336 .pmu_disable = power_pmu_disable,
1337 .event_init = power_pmu_event_init,
1338 .add = power_pmu_add,
1339 .del = power_pmu_del,
1340 .start = power_pmu_start,
1341 .stop = power_pmu_stop,
1342 .read = power_pmu_read,
1343 .start_txn = power_pmu_start_txn,
1344 .cancel_txn = power_pmu_cancel_txn,
1345 .commit_txn = power_pmu_commit_txn,
1346 .event_idx = power_pmu_event_idx,
1351 * A counter has overflowed; update its count and record
1352 * things if requested. Note that interrupts are hard-disabled
1353 * here so there is no possibility of being interrupted.
1355 static void record_and_restart(struct perf_event *event, unsigned long val,
1356 struct pt_regs *regs)
1358 u64 period = event->hw.sample_period;
1359 s64 prev, delta, left;
1362 if (event->hw.state & PERF_HES_STOPPED) {
1363 write_pmc(event->hw.idx, 0);
1367 /* we don't have to worry about interrupts here */
1368 prev = local64_read(&event->hw.prev_count);
1369 delta = check_and_compute_delta(prev, val);
1370 local64_add(delta, &event->count);
1373 * See if the total period for this event has expired,
1374 * and update for the next period.
1377 left = local64_read(&event->hw.period_left) - delta;
1385 record = siar_valid(regs);
1386 event->hw.last_period = event->hw.sample_period;
1388 if (left < 0x80000000LL)
1389 val = 0x80000000LL - left;
1392 write_pmc(event->hw.idx, val);
1393 local64_set(&event->hw.prev_count, val);
1394 local64_set(&event->hw.period_left, left);
1395 perf_event_update_userpage(event);
1398 * Finally record data if requested.
1401 struct perf_sample_data data;
1403 perf_sample_data_init(&data, ~0ULL, event->hw.last_period);
1405 if (event->attr.sample_type & PERF_SAMPLE_ADDR)
1406 perf_get_data_addr(regs, &data.addr);
1408 if (perf_event_overflow(event, &data, regs))
1409 power_pmu_stop(event, 0);
1414 * Called from generic code to get the misc flags (i.e. processor mode)
1417 unsigned long perf_misc_flags(struct pt_regs *regs)
1419 u32 flags = perf_get_misc_flags(regs);
1423 return user_mode(regs) ? PERF_RECORD_MISC_USER :
1424 PERF_RECORD_MISC_KERNEL;
1428 * Called from generic code to get the instruction pointer
1431 unsigned long perf_instruction_pointer(struct pt_regs *regs)
1433 bool use_siar = regs_use_siar(regs);
1435 if (use_siar && siar_valid(regs))
1436 return mfspr(SPRN_SIAR) + perf_ip_adjust(regs);
1438 return 0; // no valid instruction pointer
1443 static bool pmc_overflow_power7(unsigned long val)
1446 * Events on POWER7 can roll back if a speculative event doesn't
1447 * eventually complete. Unfortunately in some rare cases they will
1448 * raise a performance monitor exception. We need to catch this to
1449 * ensure we reset the PMC. In all cases the PMC will be 256 or less
1450 * cycles from overflow.
1452 * We only do this if the first pass fails to find any overflowing
1453 * PMCs because a user might set a period of less than 256 and we
1454 * don't want to mistakenly reset them.
1456 if ((0x80000000 - val) <= 256)
1462 static bool pmc_overflow(unsigned long val)
1471 * Performance monitor interrupt stuff
1473 static void perf_event_interrupt(struct pt_regs *regs)
1476 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1477 struct perf_event *event;
1478 unsigned long val[8];
1482 if (cpuhw->n_limited)
1483 freeze_limited_counters(cpuhw, mfspr(SPRN_PMC5),
1486 perf_read_regs(regs);
1488 nmi = perf_intr_is_nmi(regs);
1494 /* Read all the PMCs since we'll need them a bunch of times */
1495 for (i = 0; i < ppmu->n_counter; ++i)
1496 val[i] = read_pmc(i + 1);
1498 /* Try to find what caused the IRQ */
1500 for (i = 0; i < ppmu->n_counter; ++i) {
1501 if (!pmc_overflow(val[i]))
1503 if (is_limited_pmc(i + 1))
1504 continue; /* these won't generate IRQs */
1506 * We've found one that's overflowed. For active
1507 * counters we need to log this. For inactive
1508 * counters, we need to reset it anyway
1512 for (j = 0; j < cpuhw->n_events; ++j) {
1513 event = cpuhw->event[j];
1514 if (event->hw.idx == (i + 1)) {
1516 record_and_restart(event, val[i], regs);
1521 /* reset non active counters that have overflowed */
1522 write_pmc(i + 1, 0);
1524 if (!found && pvr_version_is(PVR_POWER7)) {
1525 /* check active counters for special buggy p7 overflow */
1526 for (i = 0; i < cpuhw->n_events; ++i) {
1527 event = cpuhw->event[i];
1528 if (!event->hw.idx || is_limited_pmc(event->hw.idx))
1530 if (pmc_overflow_power7(val[event->hw.idx - 1])) {
1531 /* event has overflowed in a buggy way*/
1533 record_and_restart(event,
1534 val[event->hw.idx - 1],
1539 if ((!found) && printk_ratelimit())
1540 printk(KERN_WARNING "Can't find PMC that caused IRQ\n");
1543 * Reset MMCR0 to its normal value. This will set PMXE and
1544 * clear FC (freeze counters) and PMAO (perf mon alert occurred)
1545 * and thus allow interrupts to occur again.
1546 * XXX might want to use MSR.PM to keep the events frozen until
1547 * we get back out of this interrupt.
1549 write_mmcr0(cpuhw, cpuhw->mmcr[0]);
1557 static void power_pmu_setup(int cpu)
1559 struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
1563 memset(cpuhw, 0, sizeof(*cpuhw));
1564 cpuhw->mmcr[0] = MMCR0_FC;
1567 static int __cpuinit
1568 power_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1570 unsigned int cpu = (long)hcpu;
1572 switch (action & ~CPU_TASKS_FROZEN) {
1573 case CPU_UP_PREPARE:
1574 power_pmu_setup(cpu);
1584 int __cpuinit register_power_pmu(struct power_pmu *pmu)
1587 return -EBUSY; /* something's already registered */
1590 pr_info("%s performance monitor hardware support registered\n",
1593 power_pmu.attr_groups = ppmu->attr_groups;
1597 * Use FCHV to ignore kernel events if MSR.HV is set.
1599 if (mfmsr() & MSR_HV)
1600 freeze_events_kernel = MMCR0_FCHV;
1601 #endif /* CONFIG_PPC64 */
1603 perf_pmu_register(&power_pmu, "cpu", PERF_TYPE_RAW);
1604 perf_cpu_notifier(power_pmu_notifier);