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) { }
77 static inline int perf_intr_is_nmi(struct pt_regs *regs)
82 #endif /* CONFIG_PPC32 */
85 * Things that are specific to 64-bit implementations.
89 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
91 unsigned long mmcra = regs->dsisr;
93 if ((mmcra & MMCRA_SAMPLE_ENABLE) && !(ppmu->flags & PPMU_ALT_SIPR)) {
94 unsigned long slot = (mmcra & MMCRA_SLOT) >> MMCRA_SLOT_SHIFT;
96 return 4 * (slot - 1);
102 * The user wants a data address recorded.
103 * If we're not doing instruction sampling, give them the SDAR
104 * (sampled data address). If we are doing instruction sampling, then
105 * only give them the SDAR if it corresponds to the instruction
106 * pointed to by SIAR; this is indicated by the [POWER6_]MMCRA_SDSYNC
109 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp)
111 unsigned long mmcra = regs->dsisr;
112 unsigned long sdsync = (ppmu->flags & PPMU_ALT_SIPR) ?
113 POWER6_MMCRA_SDSYNC : MMCRA_SDSYNC;
115 if (!(mmcra & MMCRA_SAMPLE_ENABLE) || (mmcra & sdsync))
116 *addrp = mfspr(SPRN_SDAR);
119 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
121 unsigned long mmcra = regs->dsisr;
122 unsigned long sihv = MMCRA_SIHV;
123 unsigned long sipr = MMCRA_SIPR;
125 if (TRAP(regs) != 0xf00)
126 return 0; /* not a PMU interrupt */
128 if (ppmu->flags & PPMU_ALT_SIPR) {
129 sihv = POWER6_MMCRA_SIHV;
130 sipr = POWER6_MMCRA_SIPR;
133 /* PR has priority over HV, so order below is important */
135 return PERF_RECORD_MISC_USER;
136 if ((mmcra & sihv) && (freeze_events_kernel != MMCR0_FCHV))
137 return PERF_RECORD_MISC_HYPERVISOR;
138 return PERF_RECORD_MISC_KERNEL;
142 * Overload regs->dsisr to store MMCRA so we only need to read it once
145 static inline void perf_read_regs(struct pt_regs *regs)
147 regs->dsisr = mfspr(SPRN_MMCRA);
151 * If interrupts were soft-disabled when a PMU interrupt occurs, treat
154 static inline int perf_intr_is_nmi(struct pt_regs *regs)
159 #endif /* CONFIG_PPC64 */
161 static void perf_event_interrupt(struct pt_regs *regs);
163 void perf_event_print_debug(void)
168 * Read one performance monitor counter (PMC).
170 static unsigned long read_pmc(int idx)
176 val = mfspr(SPRN_PMC1);
179 val = mfspr(SPRN_PMC2);
182 val = mfspr(SPRN_PMC3);
185 val = mfspr(SPRN_PMC4);
188 val = mfspr(SPRN_PMC5);
191 val = mfspr(SPRN_PMC6);
195 val = mfspr(SPRN_PMC7);
198 val = mfspr(SPRN_PMC8);
200 #endif /* CONFIG_PPC64 */
202 printk(KERN_ERR "oops trying to read PMC%d\n", idx);
211 static void write_pmc(int idx, unsigned long val)
215 mtspr(SPRN_PMC1, val);
218 mtspr(SPRN_PMC2, val);
221 mtspr(SPRN_PMC3, val);
224 mtspr(SPRN_PMC4, val);
227 mtspr(SPRN_PMC5, val);
230 mtspr(SPRN_PMC6, val);
234 mtspr(SPRN_PMC7, val);
237 mtspr(SPRN_PMC8, val);
239 #endif /* CONFIG_PPC64 */
241 printk(KERN_ERR "oops trying to write PMC%d\n", idx);
246 * Check if a set of events can all go on the PMU at once.
247 * If they can't, this will look at alternative codes for the events
248 * and see if any combination of alternative codes is feasible.
249 * The feasible set is returned in event_id[].
251 static int power_check_constraints(struct cpu_hw_events *cpuhw,
252 u64 event_id[], unsigned int cflags[],
255 unsigned long mask, value, nv;
256 unsigned long smasks[MAX_HWEVENTS], svalues[MAX_HWEVENTS];
257 int n_alt[MAX_HWEVENTS], choice[MAX_HWEVENTS];
259 unsigned long addf = ppmu->add_fields;
260 unsigned long tadd = ppmu->test_adder;
262 if (n_ev > ppmu->n_counter)
265 /* First see if the events will go on as-is */
266 for (i = 0; i < n_ev; ++i) {
267 if ((cflags[i] & PPMU_LIMITED_PMC_REQD)
268 && !ppmu->limited_pmc_event(event_id[i])) {
269 ppmu->get_alternatives(event_id[i], cflags[i],
270 cpuhw->alternatives[i]);
271 event_id[i] = cpuhw->alternatives[i][0];
273 if (ppmu->get_constraint(event_id[i], &cpuhw->amasks[i][0],
274 &cpuhw->avalues[i][0]))
278 for (i = 0; i < n_ev; ++i) {
279 nv = (value | cpuhw->avalues[i][0]) +
280 (value & cpuhw->avalues[i][0] & addf);
281 if ((((nv + tadd) ^ value) & mask) != 0 ||
282 (((nv + tadd) ^ cpuhw->avalues[i][0]) &
283 cpuhw->amasks[i][0]) != 0)
286 mask |= cpuhw->amasks[i][0];
289 return 0; /* all OK */
291 /* doesn't work, gather alternatives... */
292 if (!ppmu->get_alternatives)
294 for (i = 0; i < n_ev; ++i) {
296 n_alt[i] = ppmu->get_alternatives(event_id[i], cflags[i],
297 cpuhw->alternatives[i]);
298 for (j = 1; j < n_alt[i]; ++j)
299 ppmu->get_constraint(cpuhw->alternatives[i][j],
300 &cpuhw->amasks[i][j],
301 &cpuhw->avalues[i][j]);
304 /* enumerate all possibilities and see if any will work */
307 value = mask = nv = 0;
310 /* we're backtracking, restore context */
316 * See if any alternative k for event_id i,
317 * where k > j, will satisfy the constraints.
319 while (++j < n_alt[i]) {
320 nv = (value | cpuhw->avalues[i][j]) +
321 (value & cpuhw->avalues[i][j] & addf);
322 if ((((nv + tadd) ^ value) & mask) == 0 &&
323 (((nv + tadd) ^ cpuhw->avalues[i][j])
324 & cpuhw->amasks[i][j]) == 0)
329 * No feasible alternative, backtrack
330 * to event_id i-1 and continue enumerating its
331 * alternatives from where we got up to.
337 * Found a feasible alternative for event_id i,
338 * remember where we got up to with this event_id,
339 * go on to the next event_id, and start with
340 * the first alternative for it.
346 mask |= cpuhw->amasks[i][j];
352 /* OK, we have a feasible combination, tell the caller the solution */
353 for (i = 0; i < n_ev; ++i)
354 event_id[i] = cpuhw->alternatives[i][choice[i]];
359 * Check if newly-added events have consistent settings for
360 * exclude_{user,kernel,hv} with each other and any previously
363 static int check_excludes(struct perf_event **ctrs, unsigned int cflags[],
364 int n_prev, int n_new)
366 int eu = 0, ek = 0, eh = 0;
368 struct perf_event *event;
375 for (i = 0; i < n; ++i) {
376 if (cflags[i] & PPMU_LIMITED_PMC_OK) {
377 cflags[i] &= ~PPMU_LIMITED_PMC_REQD;
382 eu = event->attr.exclude_user;
383 ek = event->attr.exclude_kernel;
384 eh = event->attr.exclude_hv;
386 } else if (event->attr.exclude_user != eu ||
387 event->attr.exclude_kernel != ek ||
388 event->attr.exclude_hv != eh) {
394 for (i = 0; i < n; ++i)
395 if (cflags[i] & PPMU_LIMITED_PMC_OK)
396 cflags[i] |= PPMU_LIMITED_PMC_REQD;
401 static void power_pmu_read(struct perf_event *event)
403 s64 val, delta, prev;
408 * Performance monitor interrupts come even when interrupts
409 * are soft-disabled, as long as interrupts are hard-enabled.
410 * Therefore we treat them like NMIs.
413 prev = local64_read(&event->hw.prev_count);
415 val = read_pmc(event->hw.idx);
416 } while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
418 /* The counters are only 32 bits wide */
419 delta = (val - prev) & 0xfffffffful;
420 local64_add(delta, &event->count);
421 local64_sub(delta, &event->hw.period_left);
425 * On some machines, PMC5 and PMC6 can't be written, don't respect
426 * the freeze conditions, and don't generate interrupts. This tells
427 * us if `event' is using such a PMC.
429 static int is_limited_pmc(int pmcnum)
431 return (ppmu->flags & PPMU_LIMITED_PMC5_6)
432 && (pmcnum == 5 || pmcnum == 6);
435 static void freeze_limited_counters(struct cpu_hw_events *cpuhw,
436 unsigned long pmc5, unsigned long pmc6)
438 struct perf_event *event;
439 u64 val, prev, delta;
442 for (i = 0; i < cpuhw->n_limited; ++i) {
443 event = cpuhw->limited_counter[i];
446 val = (event->hw.idx == 5) ? pmc5 : pmc6;
447 prev = local64_read(&event->hw.prev_count);
449 delta = (val - prev) & 0xfffffffful;
450 local64_add(delta, &event->count);
454 static void thaw_limited_counters(struct cpu_hw_events *cpuhw,
455 unsigned long pmc5, unsigned long pmc6)
457 struct perf_event *event;
461 for (i = 0; i < cpuhw->n_limited; ++i) {
462 event = cpuhw->limited_counter[i];
463 event->hw.idx = cpuhw->limited_hwidx[i];
464 val = (event->hw.idx == 5) ? pmc5 : pmc6;
465 local64_set(&event->hw.prev_count, val);
466 perf_event_update_userpage(event);
471 * Since limited events don't respect the freeze conditions, we
472 * have to read them immediately after freezing or unfreezing the
473 * other events. We try to keep the values from the limited
474 * events as consistent as possible by keeping the delay (in
475 * cycles and instructions) between freezing/unfreezing and reading
476 * the limited events as small and consistent as possible.
477 * Therefore, if any limited events are in use, we read them
478 * both, and always in the same order, to minimize variability,
479 * and do it inside the same asm that writes MMCR0.
481 static void write_mmcr0(struct cpu_hw_events *cpuhw, unsigned long mmcr0)
483 unsigned long pmc5, pmc6;
485 if (!cpuhw->n_limited) {
486 mtspr(SPRN_MMCR0, mmcr0);
491 * Write MMCR0, then read PMC5 and PMC6 immediately.
492 * To ensure we don't get a performance monitor interrupt
493 * between writing MMCR0 and freezing/thawing the limited
494 * events, we first write MMCR0 with the event overflow
495 * interrupt enable bits turned off.
497 asm volatile("mtspr %3,%2; mfspr %0,%4; mfspr %1,%5"
498 : "=&r" (pmc5), "=&r" (pmc6)
499 : "r" (mmcr0 & ~(MMCR0_PMC1CE | MMCR0_PMCjCE)),
501 "i" (SPRN_PMC5), "i" (SPRN_PMC6));
503 if (mmcr0 & MMCR0_FC)
504 freeze_limited_counters(cpuhw, pmc5, pmc6);
506 thaw_limited_counters(cpuhw, pmc5, pmc6);
509 * Write the full MMCR0 including the event overflow interrupt
510 * enable bits, if necessary.
512 if (mmcr0 & (MMCR0_PMC1CE | MMCR0_PMCjCE))
513 mtspr(SPRN_MMCR0, mmcr0);
517 * Disable all events to prevent PMU interrupts and to allow
518 * events to be added or removed.
520 void hw_perf_disable(void)
522 struct cpu_hw_events *cpuhw;
527 local_irq_save(flags);
528 cpuhw = &__get_cpu_var(cpu_hw_events);
530 if (!cpuhw->disabled) {
535 * Check if we ever enabled the PMU on this cpu.
537 if (!cpuhw->pmcs_enabled) {
539 cpuhw->pmcs_enabled = 1;
543 * Disable instruction sampling if it was enabled
545 if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
547 cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
552 * Set the 'freeze counters' bit.
553 * The barrier is to make sure the mtspr has been
554 * executed and the PMU has frozen the events
557 write_mmcr0(cpuhw, mfspr(SPRN_MMCR0) | MMCR0_FC);
560 local_irq_restore(flags);
564 * Re-enable all events if disable == 0.
565 * If we were previously disabled and events were added, then
566 * put the new config on the PMU.
568 void hw_perf_enable(void)
570 struct perf_event *event;
571 struct cpu_hw_events *cpuhw;
576 unsigned int hwc_index[MAX_HWEVENTS];
582 local_irq_save(flags);
583 cpuhw = &__get_cpu_var(cpu_hw_events);
584 if (!cpuhw->disabled) {
585 local_irq_restore(flags);
591 * If we didn't change anything, or only removed events,
592 * no need to recalculate MMCR* settings and reset the PMCs.
593 * Just reenable the PMU with the current MMCR* settings
594 * (possibly updated for removal of events).
596 if (!cpuhw->n_added) {
597 mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
598 mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
599 if (cpuhw->n_events == 0)
600 ppc_set_pmu_inuse(0);
605 * Compute MMCR* values for the new set of events
607 if (ppmu->compute_mmcr(cpuhw->events, cpuhw->n_events, hwc_index,
609 /* shouldn't ever get here */
610 printk(KERN_ERR "oops compute_mmcr failed\n");
615 * Add in MMCR0 freeze bits corresponding to the
616 * attr.exclude_* bits for the first event.
617 * We have already checked that all events have the
618 * same values for these bits as the first event.
620 event = cpuhw->event[0];
621 if (event->attr.exclude_user)
622 cpuhw->mmcr[0] |= MMCR0_FCP;
623 if (event->attr.exclude_kernel)
624 cpuhw->mmcr[0] |= freeze_events_kernel;
625 if (event->attr.exclude_hv)
626 cpuhw->mmcr[0] |= MMCR0_FCHV;
629 * Write the new configuration to MMCR* with the freeze
630 * bit set and set the hardware events to their initial values.
631 * Then unfreeze the events.
633 ppc_set_pmu_inuse(1);
634 mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
635 mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
636 mtspr(SPRN_MMCR0, (cpuhw->mmcr[0] & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
640 * Read off any pre-existing events that need to move
643 for (i = 0; i < cpuhw->n_events; ++i) {
644 event = cpuhw->event[i];
645 if (event->hw.idx && event->hw.idx != hwc_index[i] + 1) {
646 power_pmu_read(event);
647 write_pmc(event->hw.idx, 0);
653 * Initialize the PMCs for all the new and moved events.
655 cpuhw->n_limited = n_lim = 0;
656 for (i = 0; i < cpuhw->n_events; ++i) {
657 event = cpuhw->event[i];
660 idx = hwc_index[i] + 1;
661 if (is_limited_pmc(idx)) {
662 cpuhw->limited_counter[n_lim] = event;
663 cpuhw->limited_hwidx[n_lim] = idx;
668 if (event->hw.sample_period) {
669 left = local64_read(&event->hw.period_left);
670 if (left < 0x80000000L)
671 val = 0x80000000L - left;
673 local64_set(&event->hw.prev_count, val);
676 perf_event_update_userpage(event);
678 cpuhw->n_limited = n_lim;
679 cpuhw->mmcr[0] |= MMCR0_PMXE | MMCR0_FCECE;
683 write_mmcr0(cpuhw, cpuhw->mmcr[0]);
686 * Enable instruction sampling if necessary
688 if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
690 mtspr(SPRN_MMCRA, cpuhw->mmcr[2]);
694 local_irq_restore(flags);
697 static int collect_events(struct perf_event *group, int max_count,
698 struct perf_event *ctrs[], u64 *events,
702 struct perf_event *event;
704 if (!is_software_event(group)) {
708 flags[n] = group->hw.event_base;
709 events[n++] = group->hw.config;
711 list_for_each_entry(event, &group->sibling_list, group_entry) {
712 if (!is_software_event(event) &&
713 event->state != PERF_EVENT_STATE_OFF) {
717 flags[n] = event->hw.event_base;
718 events[n++] = event->hw.config;
725 * Add a event to the PMU.
726 * If all events are not already frozen, then we disable and
727 * re-enable the PMU in order to get hw_perf_enable to do the
728 * actual work of reconfiguring the PMU.
730 static int power_pmu_enable(struct perf_event *event)
732 struct cpu_hw_events *cpuhw;
737 local_irq_save(flags);
741 * Add the event to the list (if there is room)
742 * and check whether the total set is still feasible.
744 cpuhw = &__get_cpu_var(cpu_hw_events);
745 n0 = cpuhw->n_events;
746 if (n0 >= ppmu->n_counter)
748 cpuhw->event[n0] = event;
749 cpuhw->events[n0] = event->hw.config;
750 cpuhw->flags[n0] = event->hw.event_base;
753 * If group events scheduling transaction was started,
754 * skip the schedulability test here, it will be peformed
755 * at commit time(->commit_txn) as a whole
757 if (cpuhw->group_flag & PERF_EVENT_TXN)
760 if (check_excludes(cpuhw->event, cpuhw->flags, n0, 1))
762 if (power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n0 + 1))
764 event->hw.config = cpuhw->events[n0];
773 local_irq_restore(flags);
778 * Remove a event from the PMU.
780 static void power_pmu_disable(struct perf_event *event)
782 struct cpu_hw_events *cpuhw;
786 local_irq_save(flags);
789 power_pmu_read(event);
791 cpuhw = &__get_cpu_var(cpu_hw_events);
792 for (i = 0; i < cpuhw->n_events; ++i) {
793 if (event == cpuhw->event[i]) {
794 while (++i < cpuhw->n_events) {
795 cpuhw->event[i-1] = cpuhw->event[i];
796 cpuhw->events[i-1] = cpuhw->events[i];
797 cpuhw->flags[i-1] = cpuhw->flags[i];
800 ppmu->disable_pmc(event->hw.idx - 1, cpuhw->mmcr);
802 write_pmc(event->hw.idx, 0);
805 perf_event_update_userpage(event);
809 for (i = 0; i < cpuhw->n_limited; ++i)
810 if (event == cpuhw->limited_counter[i])
812 if (i < cpuhw->n_limited) {
813 while (++i < cpuhw->n_limited) {
814 cpuhw->limited_counter[i-1] = cpuhw->limited_counter[i];
815 cpuhw->limited_hwidx[i-1] = cpuhw->limited_hwidx[i];
819 if (cpuhw->n_events == 0) {
820 /* disable exceptions if no events are running */
821 cpuhw->mmcr[0] &= ~(MMCR0_PMXE | MMCR0_FCECE);
825 local_irq_restore(flags);
829 * Re-enable interrupts on a event after they were throttled
830 * because they were coming too fast.
832 static void power_pmu_unthrottle(struct perf_event *event)
837 if (!event->hw.idx || !event->hw.sample_period)
839 local_irq_save(flags);
841 power_pmu_read(event);
842 left = event->hw.sample_period;
843 event->hw.last_period = left;
845 if (left < 0x80000000L)
846 val = 0x80000000L - left;
847 write_pmc(event->hw.idx, val);
848 local64_set(&event->hw.prev_count, val);
849 local64_set(&event->hw.period_left, left);
850 perf_event_update_userpage(event);
852 local_irq_restore(flags);
856 * Start group events scheduling transaction
857 * Set the flag to make pmu::enable() not perform the
858 * schedulability test, it will be performed at commit time
860 void power_pmu_start_txn(const struct pmu *pmu)
862 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
864 cpuhw->group_flag |= PERF_EVENT_TXN;
865 cpuhw->n_txn_start = cpuhw->n_events;
869 * Stop group events scheduling transaction
870 * Clear the flag and pmu::enable() will perform the
871 * schedulability test.
873 void power_pmu_cancel_txn(const struct pmu *pmu)
875 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
877 cpuhw->group_flag &= ~PERF_EVENT_TXN;
881 * Commit group events scheduling transaction
882 * Perform the group schedulability test as a whole
883 * Return 0 if success
885 int power_pmu_commit_txn(const struct pmu *pmu)
887 struct cpu_hw_events *cpuhw;
892 cpuhw = &__get_cpu_var(cpu_hw_events);
894 if (check_excludes(cpuhw->event, cpuhw->flags, 0, n))
896 i = power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n);
900 for (i = cpuhw->n_txn_start; i < n; ++i)
901 cpuhw->event[i]->hw.config = cpuhw->events[i];
903 cpuhw->group_flag &= ~PERF_EVENT_TXN;
907 struct pmu power_pmu = {
908 .enable = power_pmu_enable,
909 .disable = power_pmu_disable,
910 .read = power_pmu_read,
911 .unthrottle = power_pmu_unthrottle,
912 .start_txn = power_pmu_start_txn,
913 .cancel_txn = power_pmu_cancel_txn,
914 .commit_txn = power_pmu_commit_txn,
918 * Return 1 if we might be able to put event on a limited PMC,
920 * A event can only go on a limited PMC if it counts something
921 * that a limited PMC can count, doesn't require interrupts, and
922 * doesn't exclude any processor mode.
924 static int can_go_on_limited_pmc(struct perf_event *event, u64 ev,
928 u64 alt[MAX_EVENT_ALTERNATIVES];
930 if (event->attr.exclude_user
931 || event->attr.exclude_kernel
932 || event->attr.exclude_hv
933 || event->attr.sample_period)
936 if (ppmu->limited_pmc_event(ev))
940 * The requested event_id isn't on a limited PMC already;
941 * see if any alternative code goes on a limited PMC.
943 if (!ppmu->get_alternatives)
946 flags |= PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD;
947 n = ppmu->get_alternatives(ev, flags, alt);
953 * Find an alternative event_id that goes on a normal PMC, if possible,
954 * and return the event_id code, or 0 if there is no such alternative.
955 * (Note: event_id code 0 is "don't count" on all machines.)
957 static u64 normal_pmc_alternative(u64 ev, unsigned long flags)
959 u64 alt[MAX_EVENT_ALTERNATIVES];
962 flags &= ~(PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD);
963 n = ppmu->get_alternatives(ev, flags, alt);
969 /* Number of perf_events counting hardware events */
970 static atomic_t num_events;
971 /* Used to avoid races in calling reserve/release_pmc_hardware */
972 static DEFINE_MUTEX(pmc_reserve_mutex);
975 * Release the PMU if this is the last perf_event.
977 static void hw_perf_event_destroy(struct perf_event *event)
979 if (!atomic_add_unless(&num_events, -1, 1)) {
980 mutex_lock(&pmc_reserve_mutex);
981 if (atomic_dec_return(&num_events) == 0)
982 release_pmc_hardware();
983 mutex_unlock(&pmc_reserve_mutex);
988 * Translate a generic cache event_id config to a raw event_id code.
990 static int hw_perf_cache_event(u64 config, u64 *eventp)
992 unsigned long type, op, result;
995 if (!ppmu->cache_events)
999 type = config & 0xff;
1000 op = (config >> 8) & 0xff;
1001 result = (config >> 16) & 0xff;
1003 if (type >= PERF_COUNT_HW_CACHE_MAX ||
1004 op >= PERF_COUNT_HW_CACHE_OP_MAX ||
1005 result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1008 ev = (*ppmu->cache_events)[type][op][result];
1017 const struct pmu *hw_perf_event_init(struct perf_event *event)
1020 unsigned long flags;
1021 struct perf_event *ctrs[MAX_HWEVENTS];
1022 u64 events[MAX_HWEVENTS];
1023 unsigned int cflags[MAX_HWEVENTS];
1026 struct cpu_hw_events *cpuhw;
1029 return ERR_PTR(-ENXIO);
1030 switch (event->attr.type) {
1031 case PERF_TYPE_HARDWARE:
1032 ev = event->attr.config;
1033 if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
1034 return ERR_PTR(-EOPNOTSUPP);
1035 ev = ppmu->generic_events[ev];
1037 case PERF_TYPE_HW_CACHE:
1038 err = hw_perf_cache_event(event->attr.config, &ev);
1040 return ERR_PTR(err);
1043 ev = event->attr.config;
1046 return ERR_PTR(-EINVAL);
1048 event->hw.config_base = ev;
1052 * If we are not running on a hypervisor, force the
1053 * exclude_hv bit to 0 so that we don't care what
1054 * the user set it to.
1056 if (!firmware_has_feature(FW_FEATURE_LPAR))
1057 event->attr.exclude_hv = 0;
1060 * If this is a per-task event, then we can use
1061 * PM_RUN_* events interchangeably with their non RUN_*
1062 * equivalents, e.g. PM_RUN_CYC instead of PM_CYC.
1063 * XXX we should check if the task is an idle task.
1066 if (event->ctx->task)
1067 flags |= PPMU_ONLY_COUNT_RUN;
1070 * If this machine has limited events, check whether this
1071 * event_id could go on a limited event.
1073 if (ppmu->flags & PPMU_LIMITED_PMC5_6) {
1074 if (can_go_on_limited_pmc(event, ev, flags)) {
1075 flags |= PPMU_LIMITED_PMC_OK;
1076 } else if (ppmu->limited_pmc_event(ev)) {
1078 * The requested event_id is on a limited PMC,
1079 * but we can't use a limited PMC; see if any
1080 * alternative goes on a normal PMC.
1082 ev = normal_pmc_alternative(ev, flags);
1084 return ERR_PTR(-EINVAL);
1089 * If this is in a group, check if it can go on with all the
1090 * other hardware events in the group. We assume the event
1091 * hasn't been linked into its leader's sibling list at this point.
1094 if (event->group_leader != event) {
1095 n = collect_events(event->group_leader, ppmu->n_counter - 1,
1096 ctrs, events, cflags);
1098 return ERR_PTR(-EINVAL);
1103 if (check_excludes(ctrs, cflags, n, 1))
1104 return ERR_PTR(-EINVAL);
1106 cpuhw = &get_cpu_var(cpu_hw_events);
1107 err = power_check_constraints(cpuhw, events, cflags, n + 1);
1108 put_cpu_var(cpu_hw_events);
1110 return ERR_PTR(-EINVAL);
1112 event->hw.config = events[n];
1113 event->hw.event_base = cflags[n];
1114 event->hw.last_period = event->hw.sample_period;
1115 local64_set(&event->hw.period_left, event->hw.last_period);
1118 * See if we need to reserve the PMU.
1119 * If no events are currently in use, then we have to take a
1120 * mutex to ensure that we don't race with another task doing
1121 * reserve_pmc_hardware or release_pmc_hardware.
1124 if (!atomic_inc_not_zero(&num_events)) {
1125 mutex_lock(&pmc_reserve_mutex);
1126 if (atomic_read(&num_events) == 0 &&
1127 reserve_pmc_hardware(perf_event_interrupt))
1130 atomic_inc(&num_events);
1131 mutex_unlock(&pmc_reserve_mutex);
1133 event->destroy = hw_perf_event_destroy;
1136 return ERR_PTR(err);
1141 * A counter has overflowed; update its count and record
1142 * things if requested. Note that interrupts are hard-disabled
1143 * here so there is no possibility of being interrupted.
1145 static void record_and_restart(struct perf_event *event, unsigned long val,
1146 struct pt_regs *regs, int nmi)
1148 u64 period = event->hw.sample_period;
1149 s64 prev, delta, left;
1152 /* we don't have to worry about interrupts here */
1153 prev = local64_read(&event->hw.prev_count);
1154 delta = (val - prev) & 0xfffffffful;
1155 local64_add(delta, &event->count);
1158 * See if the total period for this event has expired,
1159 * and update for the next period.
1162 left = local64_read(&event->hw.period_left) - delta;
1170 if (left < 0x80000000LL)
1171 val = 0x80000000LL - left;
1175 * Finally record data if requested.
1178 struct perf_sample_data data;
1180 perf_sample_data_init(&data, ~0ULL);
1181 data.period = event->hw.last_period;
1183 if (event->attr.sample_type & PERF_SAMPLE_ADDR)
1184 perf_get_data_addr(regs, &data.addr);
1186 if (perf_event_overflow(event, nmi, &data, regs)) {
1188 * Interrupts are coming too fast - throttle them
1189 * by setting the event to 0, so it will be
1190 * at least 2^30 cycles until the next interrupt
1191 * (assuming each event counts at most 2 counts
1199 write_pmc(event->hw.idx, val);
1200 local64_set(&event->hw.prev_count, val);
1201 local64_set(&event->hw.period_left, left);
1202 perf_event_update_userpage(event);
1206 * Called from generic code to get the misc flags (i.e. processor mode)
1209 unsigned long perf_misc_flags(struct pt_regs *regs)
1211 u32 flags = perf_get_misc_flags(regs);
1215 return user_mode(regs) ? PERF_RECORD_MISC_USER :
1216 PERF_RECORD_MISC_KERNEL;
1220 * Called from generic code to get the instruction pointer
1223 unsigned long perf_instruction_pointer(struct pt_regs *regs)
1227 if (TRAP(regs) != 0xf00)
1228 return regs->nip; /* not a PMU interrupt */
1230 ip = mfspr(SPRN_SIAR) + perf_ip_adjust(regs);
1235 * Performance monitor interrupt stuff
1237 static void perf_event_interrupt(struct pt_regs *regs)
1240 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1241 struct perf_event *event;
1246 if (cpuhw->n_limited)
1247 freeze_limited_counters(cpuhw, mfspr(SPRN_PMC5),
1250 perf_read_regs(regs);
1252 nmi = perf_intr_is_nmi(regs);
1258 for (i = 0; i < cpuhw->n_events; ++i) {
1259 event = cpuhw->event[i];
1260 if (!event->hw.idx || is_limited_pmc(event->hw.idx))
1262 val = read_pmc(event->hw.idx);
1264 /* event has overflowed */
1266 record_and_restart(event, val, regs, nmi);
1271 * In case we didn't find and reset the event that caused
1272 * the interrupt, scan all events and reset any that are
1273 * negative, to avoid getting continual interrupts.
1274 * Any that we processed in the previous loop will not be negative.
1277 for (i = 0; i < ppmu->n_counter; ++i) {
1278 if (is_limited_pmc(i + 1))
1280 val = read_pmc(i + 1);
1282 write_pmc(i + 1, 0);
1287 * Reset MMCR0 to its normal value. This will set PMXE and
1288 * clear FC (freeze counters) and PMAO (perf mon alert occurred)
1289 * and thus allow interrupts to occur again.
1290 * XXX might want to use MSR.PM to keep the events frozen until
1291 * we get back out of this interrupt.
1293 write_mmcr0(cpuhw, cpuhw->mmcr[0]);
1301 static void power_pmu_setup(int cpu)
1303 struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
1307 memset(cpuhw, 0, sizeof(*cpuhw));
1308 cpuhw->mmcr[0] = MMCR0_FC;
1311 static int __cpuinit
1312 power_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1314 unsigned int cpu = (long)hcpu;
1316 switch (action & ~CPU_TASKS_FROZEN) {
1317 case CPU_UP_PREPARE:
1318 power_pmu_setup(cpu);
1328 int register_power_pmu(struct power_pmu *pmu)
1331 return -EBUSY; /* something's already registered */
1334 pr_info("%s performance monitor hardware support registered\n",
1339 * Use FCHV to ignore kernel events if MSR.HV is set.
1341 if (mfmsr() & MSR_HV)
1342 freeze_events_kernel = MMCR0_FCHV;
1343 #endif /* CONFIG_PPC64 */
1345 perf_cpu_notifier(power_pmu_notifier);