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>
16 #include <linux/uaccess.h>
19 #include <asm/machdep.h>
20 #include <asm/firmware.h>
21 #include <asm/ptrace.h>
22 #include <asm/code-patching.h>
24 #define BHRB_MAX_ENTRIES 32
25 #define BHRB_TARGET 0x0000000000000002
26 #define BHRB_PREDICTION 0x0000000000000001
27 #define BHRB_EA 0xFFFFFFFFFFFFFFFC
29 struct cpu_hw_events {
36 struct perf_event *event[MAX_HWEVENTS];
37 u64 events[MAX_HWEVENTS];
38 unsigned int flags[MAX_HWEVENTS];
39 unsigned long mmcr[3];
40 struct perf_event *limited_counter[MAX_LIMITED_HWCOUNTERS];
41 u8 limited_hwidx[MAX_LIMITED_HWCOUNTERS];
42 u64 alternatives[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
43 unsigned long amasks[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
44 unsigned long avalues[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
46 unsigned int group_flag;
50 u64 bhrb_filter; /* BHRB HW branch filter */
53 struct perf_branch_stack bhrb_stack;
54 struct perf_branch_entry bhrb_entries[BHRB_MAX_ENTRIES];
57 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
59 struct power_pmu *ppmu;
62 * Normally, to ignore kernel events we set the FCS (freeze counters
63 * in supervisor mode) bit in MMCR0, but if the kernel runs with the
64 * hypervisor bit set in the MSR, or if we are running on a processor
65 * where the hypervisor bit is forced to 1 (as on Apple G5 processors),
66 * then we need to use the FCHV bit to ignore kernel events.
68 static unsigned int freeze_events_kernel = MMCR0_FCS;
71 * 32-bit doesn't have MMCRA but does have an MMCR2,
72 * and a few other names are different.
77 #define MMCR0_PMCjCE MMCR0_PMCnCE
79 #define SPRN_MMCRA SPRN_MMCR2
80 #define MMCRA_SAMPLE_ENABLE 0
82 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
86 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp) { }
87 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
91 static inline void perf_read_regs(struct pt_regs *regs)
95 static inline int perf_intr_is_nmi(struct pt_regs *regs)
100 static inline int siar_valid(struct pt_regs *regs)
105 static inline void power_pmu_bhrb_enable(struct perf_event *event) {}
106 static inline void power_pmu_bhrb_disable(struct perf_event *event) {}
107 void power_pmu_flush_branch_stack(void) {}
108 static inline void power_pmu_bhrb_read(struct cpu_hw_events *cpuhw) {}
109 #endif /* CONFIG_PPC32 */
111 static bool regs_use_siar(struct pt_regs *regs)
113 return !!regs->result;
117 * Things that are specific to 64-bit implementations.
121 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
123 unsigned long mmcra = regs->dsisr;
125 if ((ppmu->flags & PPMU_HAS_SSLOT) && (mmcra & MMCRA_SAMPLE_ENABLE)) {
126 unsigned long slot = (mmcra & MMCRA_SLOT) >> MMCRA_SLOT_SHIFT;
128 return 4 * (slot - 1);
135 * The user wants a data address recorded.
136 * If we're not doing instruction sampling, give them the SDAR
137 * (sampled data address). If we are doing instruction sampling, then
138 * only give them the SDAR if it corresponds to the instruction
139 * pointed to by SIAR; this is indicated by the [POWER6_]MMCRA_SDSYNC or
140 * the [POWER7P_]MMCRA_SDAR_VALID bit in MMCRA.
142 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp)
144 unsigned long mmcra = regs->dsisr;
145 unsigned long sdsync;
147 if (ppmu->flags & PPMU_SIAR_VALID)
148 sdsync = POWER7P_MMCRA_SDAR_VALID;
149 else if (ppmu->flags & PPMU_ALT_SIPR)
150 sdsync = POWER6_MMCRA_SDSYNC;
152 sdsync = MMCRA_SDSYNC;
154 if (!(mmcra & MMCRA_SAMPLE_ENABLE) || (mmcra & sdsync))
155 *addrp = mfspr(SPRN_SDAR);
158 static bool regs_sihv(struct pt_regs *regs)
160 unsigned long sihv = MMCRA_SIHV;
162 if (ppmu->flags & PPMU_HAS_SIER)
163 return !!(regs->dar & SIER_SIHV);
165 if (ppmu->flags & PPMU_ALT_SIPR)
166 sihv = POWER6_MMCRA_SIHV;
168 return !!(regs->dsisr & sihv);
171 static bool regs_sipr(struct pt_regs *regs)
173 unsigned long sipr = MMCRA_SIPR;
175 if (ppmu->flags & PPMU_HAS_SIER)
176 return !!(regs->dar & SIER_SIPR);
178 if (ppmu->flags & PPMU_ALT_SIPR)
179 sipr = POWER6_MMCRA_SIPR;
181 return !!(regs->dsisr & sipr);
184 static inline u32 perf_flags_from_msr(struct pt_regs *regs)
186 if (regs->msr & MSR_PR)
187 return PERF_RECORD_MISC_USER;
188 if ((regs->msr & MSR_HV) && freeze_events_kernel != MMCR0_FCHV)
189 return PERF_RECORD_MISC_HYPERVISOR;
190 return PERF_RECORD_MISC_KERNEL;
193 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
195 bool use_siar = regs_use_siar(regs);
198 return perf_flags_from_msr(regs);
201 * If we don't have flags in MMCRA, rather than using
202 * the MSR, we intuit the flags from the address in
203 * SIAR which should give slightly more reliable
206 if (ppmu->flags & PPMU_NO_SIPR) {
207 unsigned long siar = mfspr(SPRN_SIAR);
208 if (siar >= PAGE_OFFSET)
209 return PERF_RECORD_MISC_KERNEL;
210 return PERF_RECORD_MISC_USER;
213 /* PR has priority over HV, so order below is important */
215 return PERF_RECORD_MISC_USER;
217 if (regs_sihv(regs) && (freeze_events_kernel != MMCR0_FCHV))
218 return PERF_RECORD_MISC_HYPERVISOR;
220 return PERF_RECORD_MISC_KERNEL;
224 * Overload regs->dsisr to store MMCRA so we only need to read it once
226 * Overload regs->dar to store SIER if we have it.
227 * Overload regs->result to specify whether we should use the MSR (result
228 * is zero) or the SIAR (result is non zero).
230 static inline void perf_read_regs(struct pt_regs *regs)
232 unsigned long mmcra = mfspr(SPRN_MMCRA);
233 int marked = mmcra & MMCRA_SAMPLE_ENABLE;
238 if (ppmu->flags & PPMU_HAS_SIER)
239 regs->dar = mfspr(SPRN_SIER);
242 * If this isn't a PMU exception (eg a software event) the SIAR is
243 * not valid. Use pt_regs.
245 * If it is a marked event use the SIAR.
247 * If the PMU doesn't update the SIAR for non marked events use
250 * If the PMU has HV/PR flags then check to see if they
251 * place the exception in userspace. If so, use pt_regs. In
252 * continuous sampling mode the SIAR and the PMU exception are
253 * not synchronised, so they may be many instructions apart.
254 * This can result in confusing backtraces. We still want
255 * hypervisor samples as well as samples in the kernel with
256 * interrupts off hence the userspace check.
258 if (TRAP(regs) != 0xf00)
262 else if ((ppmu->flags & PPMU_NO_CONT_SAMPLING))
264 else if (!(ppmu->flags & PPMU_NO_SIPR) && regs_sipr(regs))
269 regs->result = use_siar;
273 * If interrupts were soft-disabled when a PMU interrupt occurs, treat
276 static inline int perf_intr_is_nmi(struct pt_regs *regs)
282 * On processors like P7+ that have the SIAR-Valid bit, marked instructions
283 * must be sampled only if the SIAR-valid bit is set.
285 * For unmarked instructions and for processors that don't have the SIAR-Valid
286 * bit, assume that SIAR is valid.
288 static inline int siar_valid(struct pt_regs *regs)
290 unsigned long mmcra = regs->dsisr;
291 int marked = mmcra & MMCRA_SAMPLE_ENABLE;
293 if ((ppmu->flags & PPMU_SIAR_VALID) && marked)
294 return mmcra & POWER7P_MMCRA_SIAR_VALID;
300 /* Reset all possible BHRB entries */
301 static void power_pmu_bhrb_reset(void)
303 asm volatile(PPC_CLRBHRB);
306 static void power_pmu_bhrb_enable(struct perf_event *event)
308 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
313 /* Clear BHRB if we changed task context to avoid data leaks */
314 if (event->ctx->task && cpuhw->bhrb_context != event->ctx) {
315 power_pmu_bhrb_reset();
316 cpuhw->bhrb_context = event->ctx;
321 static void power_pmu_bhrb_disable(struct perf_event *event)
323 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
329 WARN_ON_ONCE(cpuhw->bhrb_users < 0);
331 if (!cpuhw->disabled && !cpuhw->bhrb_users) {
332 /* BHRB cannot be turned off when other
333 * events are active on the PMU.
336 /* avoid stale pointer */
337 cpuhw->bhrb_context = NULL;
341 /* Called from ctxsw to prevent one process's branch entries to
342 * mingle with the other process's entries during context switch.
344 void power_pmu_flush_branch_stack(void)
347 power_pmu_bhrb_reset();
349 /* Calculate the to address for a branch */
350 static __u64 power_pmu_bhrb_to(u64 addr)
356 if (is_kernel_addr(addr))
357 return branch_target((unsigned int *)addr);
359 /* Userspace: need copy instruction here then translate it */
361 ret = __get_user_inatomic(instr, (unsigned int __user *)addr);
368 target = branch_target(&instr);
369 if ((!target) || (instr & BRANCH_ABSOLUTE))
372 /* Translate relative branch target from kernel to user address */
373 return target - (unsigned long)&instr + addr;
376 /* Processing BHRB entries */
377 void power_pmu_bhrb_read(struct cpu_hw_events *cpuhw)
381 int r_index, u_index, pred;
385 while (r_index < ppmu->bhrb_nr) {
386 /* Assembly read function */
387 val = read_bhrb(r_index++);
389 /* Terminal marker: End of valid BHRB entries */
392 addr = val & BHRB_EA;
393 pred = val & BHRB_PREDICTION;
399 /* Branches are read most recent first (ie. mfbhrb 0 is
400 * the most recent branch).
401 * There are two types of valid entries:
402 * 1) a target entry which is the to address of a
403 * computed goto like a blr,bctr,btar. The next
404 * entry read from the bhrb will be branch
405 * corresponding to this target (ie. the actual
406 * blr/bctr/btar instruction).
407 * 2) a from address which is an actual branch. If a
408 * target entry proceeds this, then this is the
409 * matching branch for that target. If this is not
410 * following a target entry, then this is a branch
411 * where the target is given as an immediate field
412 * in the instruction (ie. an i or b form branch).
413 * In this case we need to read the instruction from
414 * memory to determine the target/to address.
417 if (val & BHRB_TARGET) {
418 /* Target branches use two entries
419 * (ie. computed gotos/XL form)
421 cpuhw->bhrb_entries[u_index].to = addr;
422 cpuhw->bhrb_entries[u_index].mispred = pred;
423 cpuhw->bhrb_entries[u_index].predicted = ~pred;
425 /* Get from address in next entry */
426 val = read_bhrb(r_index++);
427 addr = val & BHRB_EA;
428 if (val & BHRB_TARGET) {
429 /* Shouldn't have two targets in a
430 row.. Reset index and try again */
434 cpuhw->bhrb_entries[u_index].from = addr;
436 /* Branches to immediate field
438 cpuhw->bhrb_entries[u_index].from = addr;
439 cpuhw->bhrb_entries[u_index].to =
440 power_pmu_bhrb_to(addr);
441 cpuhw->bhrb_entries[u_index].mispred = pred;
442 cpuhw->bhrb_entries[u_index].predicted = ~pred;
448 cpuhw->bhrb_stack.nr = u_index;
452 #endif /* CONFIG_PPC64 */
454 static void perf_event_interrupt(struct pt_regs *regs);
456 void perf_event_print_debug(void)
461 * Read one performance monitor counter (PMC).
463 static unsigned long read_pmc(int idx)
469 val = mfspr(SPRN_PMC1);
472 val = mfspr(SPRN_PMC2);
475 val = mfspr(SPRN_PMC3);
478 val = mfspr(SPRN_PMC4);
481 val = mfspr(SPRN_PMC5);
484 val = mfspr(SPRN_PMC6);
488 val = mfspr(SPRN_PMC7);
491 val = mfspr(SPRN_PMC8);
493 #endif /* CONFIG_PPC64 */
495 printk(KERN_ERR "oops trying to read PMC%d\n", idx);
504 static void write_pmc(int idx, unsigned long val)
508 mtspr(SPRN_PMC1, val);
511 mtspr(SPRN_PMC2, val);
514 mtspr(SPRN_PMC3, val);
517 mtspr(SPRN_PMC4, val);
520 mtspr(SPRN_PMC5, val);
523 mtspr(SPRN_PMC6, val);
527 mtspr(SPRN_PMC7, val);
530 mtspr(SPRN_PMC8, val);
532 #endif /* CONFIG_PPC64 */
534 printk(KERN_ERR "oops trying to write PMC%d\n", idx);
539 * Check if a set of events can all go on the PMU at once.
540 * If they can't, this will look at alternative codes for the events
541 * and see if any combination of alternative codes is feasible.
542 * The feasible set is returned in event_id[].
544 static int power_check_constraints(struct cpu_hw_events *cpuhw,
545 u64 event_id[], unsigned int cflags[],
548 unsigned long mask, value, nv;
549 unsigned long smasks[MAX_HWEVENTS], svalues[MAX_HWEVENTS];
550 int n_alt[MAX_HWEVENTS], choice[MAX_HWEVENTS];
552 unsigned long addf = ppmu->add_fields;
553 unsigned long tadd = ppmu->test_adder;
555 if (n_ev > ppmu->n_counter)
558 /* First see if the events will go on as-is */
559 for (i = 0; i < n_ev; ++i) {
560 if ((cflags[i] & PPMU_LIMITED_PMC_REQD)
561 && !ppmu->limited_pmc_event(event_id[i])) {
562 ppmu->get_alternatives(event_id[i], cflags[i],
563 cpuhw->alternatives[i]);
564 event_id[i] = cpuhw->alternatives[i][0];
566 if (ppmu->get_constraint(event_id[i], &cpuhw->amasks[i][0],
567 &cpuhw->avalues[i][0]))
571 for (i = 0; i < n_ev; ++i) {
572 nv = (value | cpuhw->avalues[i][0]) +
573 (value & cpuhw->avalues[i][0] & addf);
574 if ((((nv + tadd) ^ value) & mask) != 0 ||
575 (((nv + tadd) ^ cpuhw->avalues[i][0]) &
576 cpuhw->amasks[i][0]) != 0)
579 mask |= cpuhw->amasks[i][0];
582 return 0; /* all OK */
584 /* doesn't work, gather alternatives... */
585 if (!ppmu->get_alternatives)
587 for (i = 0; i < n_ev; ++i) {
589 n_alt[i] = ppmu->get_alternatives(event_id[i], cflags[i],
590 cpuhw->alternatives[i]);
591 for (j = 1; j < n_alt[i]; ++j)
592 ppmu->get_constraint(cpuhw->alternatives[i][j],
593 &cpuhw->amasks[i][j],
594 &cpuhw->avalues[i][j]);
597 /* enumerate all possibilities and see if any will work */
600 value = mask = nv = 0;
603 /* we're backtracking, restore context */
609 * See if any alternative k for event_id i,
610 * where k > j, will satisfy the constraints.
612 while (++j < n_alt[i]) {
613 nv = (value | cpuhw->avalues[i][j]) +
614 (value & cpuhw->avalues[i][j] & addf);
615 if ((((nv + tadd) ^ value) & mask) == 0 &&
616 (((nv + tadd) ^ cpuhw->avalues[i][j])
617 & cpuhw->amasks[i][j]) == 0)
622 * No feasible alternative, backtrack
623 * to event_id i-1 and continue enumerating its
624 * alternatives from where we got up to.
630 * Found a feasible alternative for event_id i,
631 * remember where we got up to with this event_id,
632 * go on to the next event_id, and start with
633 * the first alternative for it.
639 mask |= cpuhw->amasks[i][j];
645 /* OK, we have a feasible combination, tell the caller the solution */
646 for (i = 0; i < n_ev; ++i)
647 event_id[i] = cpuhw->alternatives[i][choice[i]];
652 * Check if newly-added events have consistent settings for
653 * exclude_{user,kernel,hv} with each other and any previously
656 static int check_excludes(struct perf_event **ctrs, unsigned int cflags[],
657 int n_prev, int n_new)
659 int eu = 0, ek = 0, eh = 0;
661 struct perf_event *event;
668 for (i = 0; i < n; ++i) {
669 if (cflags[i] & PPMU_LIMITED_PMC_OK) {
670 cflags[i] &= ~PPMU_LIMITED_PMC_REQD;
675 eu = event->attr.exclude_user;
676 ek = event->attr.exclude_kernel;
677 eh = event->attr.exclude_hv;
679 } else if (event->attr.exclude_user != eu ||
680 event->attr.exclude_kernel != ek ||
681 event->attr.exclude_hv != eh) {
687 for (i = 0; i < n; ++i)
688 if (cflags[i] & PPMU_LIMITED_PMC_OK)
689 cflags[i] |= PPMU_LIMITED_PMC_REQD;
694 static u64 check_and_compute_delta(u64 prev, u64 val)
696 u64 delta = (val - prev) & 0xfffffffful;
699 * POWER7 can roll back counter values, if the new value is smaller
700 * than the previous value it will cause the delta and the counter to
701 * have bogus values unless we rolled a counter over. If a coutner is
702 * rolled back, it will be smaller, but within 256, which is the maximum
703 * number of events to rollback at once. If we dectect a rollback
704 * return 0. This can lead to a small lack of precision in the
707 if (prev > val && (prev - val) < 256)
713 static void power_pmu_read(struct perf_event *event)
715 s64 val, delta, prev;
717 if (event->hw.state & PERF_HES_STOPPED)
723 * Performance monitor interrupts come even when interrupts
724 * are soft-disabled, as long as interrupts are hard-enabled.
725 * Therefore we treat them like NMIs.
728 prev = local64_read(&event->hw.prev_count);
730 val = read_pmc(event->hw.idx);
731 delta = check_and_compute_delta(prev, val);
734 } while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
736 local64_add(delta, &event->count);
737 local64_sub(delta, &event->hw.period_left);
741 * On some machines, PMC5 and PMC6 can't be written, don't respect
742 * the freeze conditions, and don't generate interrupts. This tells
743 * us if `event' is using such a PMC.
745 static int is_limited_pmc(int pmcnum)
747 return (ppmu->flags & PPMU_LIMITED_PMC5_6)
748 && (pmcnum == 5 || pmcnum == 6);
751 static void freeze_limited_counters(struct cpu_hw_events *cpuhw,
752 unsigned long pmc5, unsigned long pmc6)
754 struct perf_event *event;
755 u64 val, prev, delta;
758 for (i = 0; i < cpuhw->n_limited; ++i) {
759 event = cpuhw->limited_counter[i];
762 val = (event->hw.idx == 5) ? pmc5 : pmc6;
763 prev = local64_read(&event->hw.prev_count);
765 delta = check_and_compute_delta(prev, val);
767 local64_add(delta, &event->count);
771 static void thaw_limited_counters(struct cpu_hw_events *cpuhw,
772 unsigned long pmc5, unsigned long pmc6)
774 struct perf_event *event;
778 for (i = 0; i < cpuhw->n_limited; ++i) {
779 event = cpuhw->limited_counter[i];
780 event->hw.idx = cpuhw->limited_hwidx[i];
781 val = (event->hw.idx == 5) ? pmc5 : pmc6;
782 prev = local64_read(&event->hw.prev_count);
783 if (check_and_compute_delta(prev, val))
784 local64_set(&event->hw.prev_count, val);
785 perf_event_update_userpage(event);
790 * Since limited events don't respect the freeze conditions, we
791 * have to read them immediately after freezing or unfreezing the
792 * other events. We try to keep the values from the limited
793 * events as consistent as possible by keeping the delay (in
794 * cycles and instructions) between freezing/unfreezing and reading
795 * the limited events as small and consistent as possible.
796 * Therefore, if any limited events are in use, we read them
797 * both, and always in the same order, to minimize variability,
798 * and do it inside the same asm that writes MMCR0.
800 static void write_mmcr0(struct cpu_hw_events *cpuhw, unsigned long mmcr0)
802 unsigned long pmc5, pmc6;
804 if (!cpuhw->n_limited) {
805 mtspr(SPRN_MMCR0, mmcr0);
810 * Write MMCR0, then read PMC5 and PMC6 immediately.
811 * To ensure we don't get a performance monitor interrupt
812 * between writing MMCR0 and freezing/thawing the limited
813 * events, we first write MMCR0 with the event overflow
814 * interrupt enable bits turned off.
816 asm volatile("mtspr %3,%2; mfspr %0,%4; mfspr %1,%5"
817 : "=&r" (pmc5), "=&r" (pmc6)
818 : "r" (mmcr0 & ~(MMCR0_PMC1CE | MMCR0_PMCjCE)),
820 "i" (SPRN_PMC5), "i" (SPRN_PMC6));
822 if (mmcr0 & MMCR0_FC)
823 freeze_limited_counters(cpuhw, pmc5, pmc6);
825 thaw_limited_counters(cpuhw, pmc5, pmc6);
828 * Write the full MMCR0 including the event overflow interrupt
829 * enable bits, if necessary.
831 if (mmcr0 & (MMCR0_PMC1CE | MMCR0_PMCjCE))
832 mtspr(SPRN_MMCR0, mmcr0);
836 * Disable all events to prevent PMU interrupts and to allow
837 * events to be added or removed.
839 static void power_pmu_disable(struct pmu *pmu)
841 struct cpu_hw_events *cpuhw;
846 local_irq_save(flags);
847 cpuhw = &__get_cpu_var(cpu_hw_events);
849 if (!cpuhw->disabled) {
854 * Check if we ever enabled the PMU on this cpu.
856 if (!cpuhw->pmcs_enabled) {
858 cpuhw->pmcs_enabled = 1;
862 * Disable instruction sampling if it was enabled
864 if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
866 cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
871 * Set the 'freeze counters' bit.
872 * The barrier is to make sure the mtspr has been
873 * executed and the PMU has frozen the events
876 write_mmcr0(cpuhw, mfspr(SPRN_MMCR0) | MMCR0_FC);
879 local_irq_restore(flags);
883 * Re-enable all events if disable == 0.
884 * If we were previously disabled and events were added, then
885 * put the new config on the PMU.
887 static void power_pmu_enable(struct pmu *pmu)
889 struct perf_event *event;
890 struct cpu_hw_events *cpuhw;
895 unsigned int hwc_index[MAX_HWEVENTS];
901 local_irq_save(flags);
902 cpuhw = &__get_cpu_var(cpu_hw_events);
903 if (!cpuhw->disabled) {
904 local_irq_restore(flags);
910 * If we didn't change anything, or only removed events,
911 * no need to recalculate MMCR* settings and reset the PMCs.
912 * Just reenable the PMU with the current MMCR* settings
913 * (possibly updated for removal of events).
915 if (!cpuhw->n_added) {
916 mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
917 mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
918 if (cpuhw->n_events == 0)
919 ppc_set_pmu_inuse(0);
924 * Compute MMCR* values for the new set of events
926 if (ppmu->compute_mmcr(cpuhw->events, cpuhw->n_events, hwc_index,
928 /* shouldn't ever get here */
929 printk(KERN_ERR "oops compute_mmcr failed\n");
934 * Add in MMCR0 freeze bits corresponding to the
935 * attr.exclude_* bits for the first event.
936 * We have already checked that all events have the
937 * same values for these bits as the first event.
939 event = cpuhw->event[0];
940 if (event->attr.exclude_user)
941 cpuhw->mmcr[0] |= MMCR0_FCP;
942 if (event->attr.exclude_kernel)
943 cpuhw->mmcr[0] |= freeze_events_kernel;
944 if (event->attr.exclude_hv)
945 cpuhw->mmcr[0] |= MMCR0_FCHV;
948 * Write the new configuration to MMCR* with the freeze
949 * bit set and set the hardware events to their initial values.
950 * Then unfreeze the events.
952 ppc_set_pmu_inuse(1);
953 mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
954 mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
955 mtspr(SPRN_MMCR0, (cpuhw->mmcr[0] & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
959 * Read off any pre-existing events that need to move
962 for (i = 0; i < cpuhw->n_events; ++i) {
963 event = cpuhw->event[i];
964 if (event->hw.idx && event->hw.idx != hwc_index[i] + 1) {
965 power_pmu_read(event);
966 write_pmc(event->hw.idx, 0);
972 * Initialize the PMCs for all the new and moved events.
974 cpuhw->n_limited = n_lim = 0;
975 for (i = 0; i < cpuhw->n_events; ++i) {
976 event = cpuhw->event[i];
979 idx = hwc_index[i] + 1;
980 if (is_limited_pmc(idx)) {
981 cpuhw->limited_counter[n_lim] = event;
982 cpuhw->limited_hwidx[n_lim] = idx;
987 if (event->hw.sample_period) {
988 left = local64_read(&event->hw.period_left);
989 if (left < 0x80000000L)
990 val = 0x80000000L - left;
992 local64_set(&event->hw.prev_count, val);
994 if (event->hw.state & PERF_HES_STOPPED)
997 perf_event_update_userpage(event);
999 cpuhw->n_limited = n_lim;
1000 cpuhw->mmcr[0] |= MMCR0_PMXE | MMCR0_FCECE;
1004 write_mmcr0(cpuhw, cpuhw->mmcr[0]);
1007 * Enable instruction sampling if necessary
1009 if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
1011 mtspr(SPRN_MMCRA, cpuhw->mmcr[2]);
1015 if (cpuhw->bhrb_users)
1016 ppmu->config_bhrb(cpuhw->bhrb_filter);
1018 local_irq_restore(flags);
1021 static int collect_events(struct perf_event *group, int max_count,
1022 struct perf_event *ctrs[], u64 *events,
1023 unsigned int *flags)
1026 struct perf_event *event;
1028 if (!is_software_event(group)) {
1032 flags[n] = group->hw.event_base;
1033 events[n++] = group->hw.config;
1035 list_for_each_entry(event, &group->sibling_list, group_entry) {
1036 if (!is_software_event(event) &&
1037 event->state != PERF_EVENT_STATE_OFF) {
1041 flags[n] = event->hw.event_base;
1042 events[n++] = event->hw.config;
1049 * Add a event to the PMU.
1050 * If all events are not already frozen, then we disable and
1051 * re-enable the PMU in order to get hw_perf_enable to do the
1052 * actual work of reconfiguring the PMU.
1054 static int power_pmu_add(struct perf_event *event, int ef_flags)
1056 struct cpu_hw_events *cpuhw;
1057 unsigned long flags;
1061 local_irq_save(flags);
1062 perf_pmu_disable(event->pmu);
1065 * Add the event to the list (if there is room)
1066 * and check whether the total set is still feasible.
1068 cpuhw = &__get_cpu_var(cpu_hw_events);
1069 n0 = cpuhw->n_events;
1070 if (n0 >= ppmu->n_counter)
1072 cpuhw->event[n0] = event;
1073 cpuhw->events[n0] = event->hw.config;
1074 cpuhw->flags[n0] = event->hw.event_base;
1077 * This event may have been disabled/stopped in record_and_restart()
1078 * because we exceeded the ->event_limit. If re-starting the event,
1079 * clear the ->hw.state (STOPPED and UPTODATE flags), so the user
1080 * notification is re-enabled.
1082 if (!(ef_flags & PERF_EF_START))
1083 event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
1085 event->hw.state = 0;
1088 * If group events scheduling transaction was started,
1089 * skip the schedulability test here, it will be performed
1090 * at commit time(->commit_txn) as a whole
1092 if (cpuhw->group_flag & PERF_EVENT_TXN)
1095 if (check_excludes(cpuhw->event, cpuhw->flags, n0, 1))
1097 if (power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n0 + 1))
1099 event->hw.config = cpuhw->events[n0];
1107 if (has_branch_stack(event))
1108 power_pmu_bhrb_enable(event);
1110 perf_pmu_enable(event->pmu);
1111 local_irq_restore(flags);
1116 * Remove a event from the PMU.
1118 static void power_pmu_del(struct perf_event *event, int ef_flags)
1120 struct cpu_hw_events *cpuhw;
1122 unsigned long flags;
1124 local_irq_save(flags);
1125 perf_pmu_disable(event->pmu);
1127 power_pmu_read(event);
1129 cpuhw = &__get_cpu_var(cpu_hw_events);
1130 for (i = 0; i < cpuhw->n_events; ++i) {
1131 if (event == cpuhw->event[i]) {
1132 while (++i < cpuhw->n_events) {
1133 cpuhw->event[i-1] = cpuhw->event[i];
1134 cpuhw->events[i-1] = cpuhw->events[i];
1135 cpuhw->flags[i-1] = cpuhw->flags[i];
1138 ppmu->disable_pmc(event->hw.idx - 1, cpuhw->mmcr);
1139 if (event->hw.idx) {
1140 write_pmc(event->hw.idx, 0);
1143 perf_event_update_userpage(event);
1147 for (i = 0; i < cpuhw->n_limited; ++i)
1148 if (event == cpuhw->limited_counter[i])
1150 if (i < cpuhw->n_limited) {
1151 while (++i < cpuhw->n_limited) {
1152 cpuhw->limited_counter[i-1] = cpuhw->limited_counter[i];
1153 cpuhw->limited_hwidx[i-1] = cpuhw->limited_hwidx[i];
1157 if (cpuhw->n_events == 0) {
1158 /* disable exceptions if no events are running */
1159 cpuhw->mmcr[0] &= ~(MMCR0_PMXE | MMCR0_FCECE);
1162 if (has_branch_stack(event))
1163 power_pmu_bhrb_disable(event);
1165 perf_pmu_enable(event->pmu);
1166 local_irq_restore(flags);
1170 * POWER-PMU does not support disabling individual counters, hence
1171 * program their cycle counter to their max value and ignore the interrupts.
1174 static void power_pmu_start(struct perf_event *event, int ef_flags)
1176 unsigned long flags;
1180 if (!event->hw.idx || !event->hw.sample_period)
1183 if (!(event->hw.state & PERF_HES_STOPPED))
1186 if (ef_flags & PERF_EF_RELOAD)
1187 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1189 local_irq_save(flags);
1190 perf_pmu_disable(event->pmu);
1192 event->hw.state = 0;
1193 left = local64_read(&event->hw.period_left);
1196 if (left < 0x80000000L)
1197 val = 0x80000000L - left;
1199 write_pmc(event->hw.idx, val);
1201 perf_event_update_userpage(event);
1202 perf_pmu_enable(event->pmu);
1203 local_irq_restore(flags);
1206 static void power_pmu_stop(struct perf_event *event, int ef_flags)
1208 unsigned long flags;
1210 if (!event->hw.idx || !event->hw.sample_period)
1213 if (event->hw.state & PERF_HES_STOPPED)
1216 local_irq_save(flags);
1217 perf_pmu_disable(event->pmu);
1219 power_pmu_read(event);
1220 event->hw.state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1221 write_pmc(event->hw.idx, 0);
1223 perf_event_update_userpage(event);
1224 perf_pmu_enable(event->pmu);
1225 local_irq_restore(flags);
1229 * Start group events scheduling transaction
1230 * Set the flag to make pmu::enable() not perform the
1231 * schedulability test, it will be performed at commit time
1233 void power_pmu_start_txn(struct pmu *pmu)
1235 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1237 perf_pmu_disable(pmu);
1238 cpuhw->group_flag |= PERF_EVENT_TXN;
1239 cpuhw->n_txn_start = cpuhw->n_events;
1243 * Stop group events scheduling transaction
1244 * Clear the flag and pmu::enable() will perform the
1245 * schedulability test.
1247 void power_pmu_cancel_txn(struct pmu *pmu)
1249 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1251 cpuhw->group_flag &= ~PERF_EVENT_TXN;
1252 perf_pmu_enable(pmu);
1256 * Commit group events scheduling transaction
1257 * Perform the group schedulability test as a whole
1258 * Return 0 if success
1260 int power_pmu_commit_txn(struct pmu *pmu)
1262 struct cpu_hw_events *cpuhw;
1267 cpuhw = &__get_cpu_var(cpu_hw_events);
1268 n = cpuhw->n_events;
1269 if (check_excludes(cpuhw->event, cpuhw->flags, 0, n))
1271 i = power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n);
1275 for (i = cpuhw->n_txn_start; i < n; ++i)
1276 cpuhw->event[i]->hw.config = cpuhw->events[i];
1278 cpuhw->group_flag &= ~PERF_EVENT_TXN;
1279 perf_pmu_enable(pmu);
1284 * Return 1 if we might be able to put event on a limited PMC,
1286 * A event can only go on a limited PMC if it counts something
1287 * that a limited PMC can count, doesn't require interrupts, and
1288 * doesn't exclude any processor mode.
1290 static int can_go_on_limited_pmc(struct perf_event *event, u64 ev,
1294 u64 alt[MAX_EVENT_ALTERNATIVES];
1296 if (event->attr.exclude_user
1297 || event->attr.exclude_kernel
1298 || event->attr.exclude_hv
1299 || event->attr.sample_period)
1302 if (ppmu->limited_pmc_event(ev))
1306 * The requested event_id isn't on a limited PMC already;
1307 * see if any alternative code goes on a limited PMC.
1309 if (!ppmu->get_alternatives)
1312 flags |= PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD;
1313 n = ppmu->get_alternatives(ev, flags, alt);
1319 * Find an alternative event_id that goes on a normal PMC, if possible,
1320 * and return the event_id code, or 0 if there is no such alternative.
1321 * (Note: event_id code 0 is "don't count" on all machines.)
1323 static u64 normal_pmc_alternative(u64 ev, unsigned long flags)
1325 u64 alt[MAX_EVENT_ALTERNATIVES];
1328 flags &= ~(PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD);
1329 n = ppmu->get_alternatives(ev, flags, alt);
1335 /* Number of perf_events counting hardware events */
1336 static atomic_t num_events;
1337 /* Used to avoid races in calling reserve/release_pmc_hardware */
1338 static DEFINE_MUTEX(pmc_reserve_mutex);
1341 * Release the PMU if this is the last perf_event.
1343 static void hw_perf_event_destroy(struct perf_event *event)
1345 if (!atomic_add_unless(&num_events, -1, 1)) {
1346 mutex_lock(&pmc_reserve_mutex);
1347 if (atomic_dec_return(&num_events) == 0)
1348 release_pmc_hardware();
1349 mutex_unlock(&pmc_reserve_mutex);
1354 * Translate a generic cache event_id config to a raw event_id code.
1356 static int hw_perf_cache_event(u64 config, u64 *eventp)
1358 unsigned long type, op, result;
1361 if (!ppmu->cache_events)
1365 type = config & 0xff;
1366 op = (config >> 8) & 0xff;
1367 result = (config >> 16) & 0xff;
1369 if (type >= PERF_COUNT_HW_CACHE_MAX ||
1370 op >= PERF_COUNT_HW_CACHE_OP_MAX ||
1371 result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1374 ev = (*ppmu->cache_events)[type][op][result];
1383 static int power_pmu_event_init(struct perf_event *event)
1386 unsigned long flags;
1387 struct perf_event *ctrs[MAX_HWEVENTS];
1388 u64 events[MAX_HWEVENTS];
1389 unsigned int cflags[MAX_HWEVENTS];
1392 struct cpu_hw_events *cpuhw;
1397 if (has_branch_stack(event)) {
1398 /* PMU has BHRB enabled */
1399 if (!(ppmu->flags & PPMU_BHRB))
1403 switch (event->attr.type) {
1404 case PERF_TYPE_HARDWARE:
1405 ev = event->attr.config;
1406 if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
1408 ev = ppmu->generic_events[ev];
1410 case PERF_TYPE_HW_CACHE:
1411 err = hw_perf_cache_event(event->attr.config, &ev);
1416 ev = event->attr.config;
1422 event->hw.config_base = ev;
1426 * If we are not running on a hypervisor, force the
1427 * exclude_hv bit to 0 so that we don't care what
1428 * the user set it to.
1430 if (!firmware_has_feature(FW_FEATURE_LPAR))
1431 event->attr.exclude_hv = 0;
1434 * If this is a per-task event, then we can use
1435 * PM_RUN_* events interchangeably with their non RUN_*
1436 * equivalents, e.g. PM_RUN_CYC instead of PM_CYC.
1437 * XXX we should check if the task is an idle task.
1440 if (event->attach_state & PERF_ATTACH_TASK)
1441 flags |= PPMU_ONLY_COUNT_RUN;
1444 * If this machine has limited events, check whether this
1445 * event_id could go on a limited event.
1447 if (ppmu->flags & PPMU_LIMITED_PMC5_6) {
1448 if (can_go_on_limited_pmc(event, ev, flags)) {
1449 flags |= PPMU_LIMITED_PMC_OK;
1450 } else if (ppmu->limited_pmc_event(ev)) {
1452 * The requested event_id is on a limited PMC,
1453 * but we can't use a limited PMC; see if any
1454 * alternative goes on a normal PMC.
1456 ev = normal_pmc_alternative(ev, flags);
1463 * If this is in a group, check if it can go on with all the
1464 * other hardware events in the group. We assume the event
1465 * hasn't been linked into its leader's sibling list at this point.
1468 if (event->group_leader != event) {
1469 n = collect_events(event->group_leader, ppmu->n_counter - 1,
1470 ctrs, events, cflags);
1477 if (check_excludes(ctrs, cflags, n, 1))
1480 cpuhw = &get_cpu_var(cpu_hw_events);
1481 err = power_check_constraints(cpuhw, events, cflags, n + 1);
1483 if (has_branch_stack(event)) {
1484 cpuhw->bhrb_filter = ppmu->bhrb_filter_map(
1485 event->attr.branch_sample_type);
1487 if(cpuhw->bhrb_filter == -1)
1491 put_cpu_var(cpu_hw_events);
1495 event->hw.config = events[n];
1496 event->hw.event_base = cflags[n];
1497 event->hw.last_period = event->hw.sample_period;
1498 local64_set(&event->hw.period_left, event->hw.last_period);
1501 * See if we need to reserve the PMU.
1502 * If no events are currently in use, then we have to take a
1503 * mutex to ensure that we don't race with another task doing
1504 * reserve_pmc_hardware or release_pmc_hardware.
1507 if (!atomic_inc_not_zero(&num_events)) {
1508 mutex_lock(&pmc_reserve_mutex);
1509 if (atomic_read(&num_events) == 0 &&
1510 reserve_pmc_hardware(perf_event_interrupt))
1513 atomic_inc(&num_events);
1514 mutex_unlock(&pmc_reserve_mutex);
1516 event->destroy = hw_perf_event_destroy;
1521 static int power_pmu_event_idx(struct perf_event *event)
1523 return event->hw.idx;
1526 ssize_t power_events_sysfs_show(struct device *dev,
1527 struct device_attribute *attr, char *page)
1529 struct perf_pmu_events_attr *pmu_attr;
1531 pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
1533 return sprintf(page, "event=0x%02llx\n", pmu_attr->id);
1536 struct pmu power_pmu = {
1537 .pmu_enable = power_pmu_enable,
1538 .pmu_disable = power_pmu_disable,
1539 .event_init = power_pmu_event_init,
1540 .add = power_pmu_add,
1541 .del = power_pmu_del,
1542 .start = power_pmu_start,
1543 .stop = power_pmu_stop,
1544 .read = power_pmu_read,
1545 .start_txn = power_pmu_start_txn,
1546 .cancel_txn = power_pmu_cancel_txn,
1547 .commit_txn = power_pmu_commit_txn,
1548 .event_idx = power_pmu_event_idx,
1549 .flush_branch_stack = power_pmu_flush_branch_stack,
1553 * A counter has overflowed; update its count and record
1554 * things if requested. Note that interrupts are hard-disabled
1555 * here so there is no possibility of being interrupted.
1557 static void record_and_restart(struct perf_event *event, unsigned long val,
1558 struct pt_regs *regs)
1560 u64 period = event->hw.sample_period;
1561 s64 prev, delta, left;
1564 if (event->hw.state & PERF_HES_STOPPED) {
1565 write_pmc(event->hw.idx, 0);
1569 /* we don't have to worry about interrupts here */
1570 prev = local64_read(&event->hw.prev_count);
1571 delta = check_and_compute_delta(prev, val);
1572 local64_add(delta, &event->count);
1575 * See if the total period for this event has expired,
1576 * and update for the next period.
1579 left = local64_read(&event->hw.period_left) - delta;
1587 record = siar_valid(regs);
1588 event->hw.last_period = event->hw.sample_period;
1590 if (left < 0x80000000LL)
1591 val = 0x80000000LL - left;
1594 write_pmc(event->hw.idx, val);
1595 local64_set(&event->hw.prev_count, val);
1596 local64_set(&event->hw.period_left, left);
1597 perf_event_update_userpage(event);
1600 * Finally record data if requested.
1603 struct perf_sample_data data;
1605 perf_sample_data_init(&data, ~0ULL, event->hw.last_period);
1607 if (event->attr.sample_type & PERF_SAMPLE_ADDR)
1608 perf_get_data_addr(regs, &data.addr);
1610 if (event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK) {
1611 struct cpu_hw_events *cpuhw;
1612 cpuhw = &__get_cpu_var(cpu_hw_events);
1613 power_pmu_bhrb_read(cpuhw);
1614 data.br_stack = &cpuhw->bhrb_stack;
1617 if (perf_event_overflow(event, &data, regs))
1618 power_pmu_stop(event, 0);
1623 * Called from generic code to get the misc flags (i.e. processor mode)
1626 unsigned long perf_misc_flags(struct pt_regs *regs)
1628 u32 flags = perf_get_misc_flags(regs);
1632 return user_mode(regs) ? PERF_RECORD_MISC_USER :
1633 PERF_RECORD_MISC_KERNEL;
1637 * Called from generic code to get the instruction pointer
1640 unsigned long perf_instruction_pointer(struct pt_regs *regs)
1642 bool use_siar = regs_use_siar(regs);
1644 if (use_siar && siar_valid(regs))
1645 return mfspr(SPRN_SIAR) + perf_ip_adjust(regs);
1647 return 0; // no valid instruction pointer
1652 static bool pmc_overflow_power7(unsigned long val)
1655 * Events on POWER7 can roll back if a speculative event doesn't
1656 * eventually complete. Unfortunately in some rare cases they will
1657 * raise a performance monitor exception. We need to catch this to
1658 * ensure we reset the PMC. In all cases the PMC will be 256 or less
1659 * cycles from overflow.
1661 * We only do this if the first pass fails to find any overflowing
1662 * PMCs because a user might set a period of less than 256 and we
1663 * don't want to mistakenly reset them.
1665 if ((0x80000000 - val) <= 256)
1671 static bool pmc_overflow(unsigned long val)
1680 * Performance monitor interrupt stuff
1682 static void perf_event_interrupt(struct pt_regs *regs)
1685 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1686 struct perf_event *event;
1687 unsigned long val[8];
1691 if (cpuhw->n_limited)
1692 freeze_limited_counters(cpuhw, mfspr(SPRN_PMC5),
1695 perf_read_regs(regs);
1697 nmi = perf_intr_is_nmi(regs);
1703 /* Read all the PMCs since we'll need them a bunch of times */
1704 for (i = 0; i < ppmu->n_counter; ++i)
1705 val[i] = read_pmc(i + 1);
1707 /* Try to find what caused the IRQ */
1709 for (i = 0; i < ppmu->n_counter; ++i) {
1710 if (!pmc_overflow(val[i]))
1712 if (is_limited_pmc(i + 1))
1713 continue; /* these won't generate IRQs */
1715 * We've found one that's overflowed. For active
1716 * counters we need to log this. For inactive
1717 * counters, we need to reset it anyway
1721 for (j = 0; j < cpuhw->n_events; ++j) {
1722 event = cpuhw->event[j];
1723 if (event->hw.idx == (i + 1)) {
1725 record_and_restart(event, val[i], regs);
1730 /* reset non active counters that have overflowed */
1731 write_pmc(i + 1, 0);
1733 if (!found && pvr_version_is(PVR_POWER7)) {
1734 /* check active counters for special buggy p7 overflow */
1735 for (i = 0; i < cpuhw->n_events; ++i) {
1736 event = cpuhw->event[i];
1737 if (!event->hw.idx || is_limited_pmc(event->hw.idx))
1739 if (pmc_overflow_power7(val[event->hw.idx - 1])) {
1740 /* event has overflowed in a buggy way*/
1742 record_and_restart(event,
1743 val[event->hw.idx - 1],
1748 if ((!found) && printk_ratelimit())
1749 printk(KERN_WARNING "Can't find PMC that caused IRQ\n");
1752 * Reset MMCR0 to its normal value. This will set PMXE and
1753 * clear FC (freeze counters) and PMAO (perf mon alert occurred)
1754 * and thus allow interrupts to occur again.
1755 * XXX might want to use MSR.PM to keep the events frozen until
1756 * we get back out of this interrupt.
1758 write_mmcr0(cpuhw, cpuhw->mmcr[0]);
1766 static void power_pmu_setup(int cpu)
1768 struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
1772 memset(cpuhw, 0, sizeof(*cpuhw));
1773 cpuhw->mmcr[0] = MMCR0_FC;
1776 static int __cpuinit
1777 power_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1779 unsigned int cpu = (long)hcpu;
1781 switch (action & ~CPU_TASKS_FROZEN) {
1782 case CPU_UP_PREPARE:
1783 power_pmu_setup(cpu);
1793 int __cpuinit register_power_pmu(struct power_pmu *pmu)
1796 return -EBUSY; /* something's already registered */
1799 pr_info("%s performance monitor hardware support registered\n",
1802 power_pmu.attr_groups = ppmu->attr_groups;
1806 * Use FCHV to ignore kernel events if MSR.HV is set.
1808 if (mfmsr() & MSR_HV)
1809 freeze_events_kernel = MMCR0_FCHV;
1810 #endif /* CONFIG_PPC64 */
1812 perf_pmu_register(&power_pmu, "cpu", PERF_TYPE_RAW);
1813 perf_cpu_notifier(power_pmu_notifier);