kvm tools, setup: Create private directory
[linux-2.6/next.git] / arch / powerpc / kernel / perf_event.c
blob822f63008ae11642b570986c7faf8ca61478ae10
1 /*
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 <asm/reg.h>
17 #include <asm/pmc.h>
18 #include <asm/machdep.h>
19 #include <asm/firmware.h>
20 #include <asm/ptrace.h>
22 struct cpu_hw_events {
23 int n_events;
24 int n_percpu;
25 int disabled;
26 int n_added;
27 int n_limited;
28 u8 pmcs_enabled;
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;
40 int n_txn_start;
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.
59 #ifdef CONFIG_PPC32
61 #define MMCR0_FCHV 0
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)
69 return 0;
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)
74 return 0;
76 static inline void perf_read_regs(struct pt_regs *regs) { }
77 static inline int perf_intr_is_nmi(struct pt_regs *regs)
79 return 0;
82 #endif /* CONFIG_PPC32 */
85 * Things that are specific to 64-bit implementations.
87 #ifdef CONFIG_PPC64
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;
95 if (slot > 1)
96 return 4 * (slot - 1);
98 return 0;
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
107 * bit in MMCRA.
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 */
134 if (mmcra & sipr)
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
143 * on each interrupt.
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
152 * it as an NMI.
154 static inline int perf_intr_is_nmi(struct pt_regs *regs)
156 return !regs->softe;
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)
172 unsigned long val;
174 switch (idx) {
175 case 1:
176 val = mfspr(SPRN_PMC1);
177 break;
178 case 2:
179 val = mfspr(SPRN_PMC2);
180 break;
181 case 3:
182 val = mfspr(SPRN_PMC3);
183 break;
184 case 4:
185 val = mfspr(SPRN_PMC4);
186 break;
187 case 5:
188 val = mfspr(SPRN_PMC5);
189 break;
190 case 6:
191 val = mfspr(SPRN_PMC6);
192 break;
193 #ifdef CONFIG_PPC64
194 case 7:
195 val = mfspr(SPRN_PMC7);
196 break;
197 case 8:
198 val = mfspr(SPRN_PMC8);
199 break;
200 #endif /* CONFIG_PPC64 */
201 default:
202 printk(KERN_ERR "oops trying to read PMC%d\n", idx);
203 val = 0;
205 return val;
209 * Write one PMC.
211 static void write_pmc(int idx, unsigned long val)
213 switch (idx) {
214 case 1:
215 mtspr(SPRN_PMC1, val);
216 break;
217 case 2:
218 mtspr(SPRN_PMC2, val);
219 break;
220 case 3:
221 mtspr(SPRN_PMC3, val);
222 break;
223 case 4:
224 mtspr(SPRN_PMC4, val);
225 break;
226 case 5:
227 mtspr(SPRN_PMC5, val);
228 break;
229 case 6:
230 mtspr(SPRN_PMC6, val);
231 break;
232 #ifdef CONFIG_PPC64
233 case 7:
234 mtspr(SPRN_PMC7, val);
235 break;
236 case 8:
237 mtspr(SPRN_PMC8, val);
238 break;
239 #endif /* CONFIG_PPC64 */
240 default:
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[],
253 int n_ev)
255 unsigned long mask, value, nv;
256 unsigned long smasks[MAX_HWEVENTS], svalues[MAX_HWEVENTS];
257 int n_alt[MAX_HWEVENTS], choice[MAX_HWEVENTS];
258 int i, j;
259 unsigned long addf = ppmu->add_fields;
260 unsigned long tadd = ppmu->test_adder;
262 if (n_ev > ppmu->n_counter)
263 return -1;
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]))
275 return -1;
277 value = mask = 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)
284 break;
285 value = nv;
286 mask |= cpuhw->amasks[i][0];
288 if (i == n_ev)
289 return 0; /* all OK */
291 /* doesn't work, gather alternatives... */
292 if (!ppmu->get_alternatives)
293 return -1;
294 for (i = 0; i < n_ev; ++i) {
295 choice[i] = 0;
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 */
305 i = 0;
306 j = -1;
307 value = mask = nv = 0;
308 while (i < n_ev) {
309 if (j >= 0) {
310 /* we're backtracking, restore context */
311 value = svalues[i];
312 mask = smasks[i];
313 j = choice[i];
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)
325 break;
327 if (j >= n_alt[i]) {
329 * No feasible alternative, backtrack
330 * to event_id i-1 and continue enumerating its
331 * alternatives from where we got up to.
333 if (--i < 0)
334 return -1;
335 } else {
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.
342 choice[i] = j;
343 svalues[i] = value;
344 smasks[i] = mask;
345 value = nv;
346 mask |= cpuhw->amasks[i][j];
347 ++i;
348 j = -1;
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]];
355 return 0;
359 * Check if newly-added events have consistent settings for
360 * exclude_{user,kernel,hv} with each other and any previously
361 * added events.
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;
367 int i, n, first;
368 struct perf_event *event;
370 n = n_prev + n_new;
371 if (n <= 1)
372 return 0;
374 first = 1;
375 for (i = 0; i < n; ++i) {
376 if (cflags[i] & PPMU_LIMITED_PMC_OK) {
377 cflags[i] &= ~PPMU_LIMITED_PMC_REQD;
378 continue;
380 event = ctrs[i];
381 if (first) {
382 eu = event->attr.exclude_user;
383 ek = event->attr.exclude_kernel;
384 eh = event->attr.exclude_hv;
385 first = 0;
386 } else if (event->attr.exclude_user != eu ||
387 event->attr.exclude_kernel != ek ||
388 event->attr.exclude_hv != eh) {
389 return -EAGAIN;
393 if (eu || ek || eh)
394 for (i = 0; i < n; ++i)
395 if (cflags[i] & PPMU_LIMITED_PMC_OK)
396 cflags[i] |= PPMU_LIMITED_PMC_REQD;
398 return 0;
401 static u64 check_and_compute_delta(u64 prev, u64 val)
403 u64 delta = (val - prev) & 0xfffffffful;
406 * POWER7 can roll back counter values, if the new value is smaller
407 * than the previous value it will cause the delta and the counter to
408 * have bogus values unless we rolled a counter over. If a coutner is
409 * rolled back, it will be smaller, but within 256, which is the maximum
410 * number of events to rollback at once. If we dectect a rollback
411 * return 0. This can lead to a small lack of precision in the
412 * counters.
414 if (prev > val && (prev - val) < 256)
415 delta = 0;
417 return delta;
420 static void power_pmu_read(struct perf_event *event)
422 s64 val, delta, prev;
424 if (event->hw.state & PERF_HES_STOPPED)
425 return;
427 if (!event->hw.idx)
428 return;
430 * Performance monitor interrupts come even when interrupts
431 * are soft-disabled, as long as interrupts are hard-enabled.
432 * Therefore we treat them like NMIs.
434 do {
435 prev = local64_read(&event->hw.prev_count);
436 barrier();
437 val = read_pmc(event->hw.idx);
438 delta = check_and_compute_delta(prev, val);
439 if (!delta)
440 return;
441 } while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
443 local64_add(delta, &event->count);
444 local64_sub(delta, &event->hw.period_left);
448 * On some machines, PMC5 and PMC6 can't be written, don't respect
449 * the freeze conditions, and don't generate interrupts. This tells
450 * us if `event' is using such a PMC.
452 static int is_limited_pmc(int pmcnum)
454 return (ppmu->flags & PPMU_LIMITED_PMC5_6)
455 && (pmcnum == 5 || pmcnum == 6);
458 static void freeze_limited_counters(struct cpu_hw_events *cpuhw,
459 unsigned long pmc5, unsigned long pmc6)
461 struct perf_event *event;
462 u64 val, prev, delta;
463 int i;
465 for (i = 0; i < cpuhw->n_limited; ++i) {
466 event = cpuhw->limited_counter[i];
467 if (!event->hw.idx)
468 continue;
469 val = (event->hw.idx == 5) ? pmc5 : pmc6;
470 prev = local64_read(&event->hw.prev_count);
471 event->hw.idx = 0;
472 delta = check_and_compute_delta(prev, val);
473 if (delta)
474 local64_add(delta, &event->count);
478 static void thaw_limited_counters(struct cpu_hw_events *cpuhw,
479 unsigned long pmc5, unsigned long pmc6)
481 struct perf_event *event;
482 u64 val, prev;
483 int i;
485 for (i = 0; i < cpuhw->n_limited; ++i) {
486 event = cpuhw->limited_counter[i];
487 event->hw.idx = cpuhw->limited_hwidx[i];
488 val = (event->hw.idx == 5) ? pmc5 : pmc6;
489 prev = local64_read(&event->hw.prev_count);
490 if (check_and_compute_delta(prev, val))
491 local64_set(&event->hw.prev_count, val);
492 perf_event_update_userpage(event);
497 * Since limited events don't respect the freeze conditions, we
498 * have to read them immediately after freezing or unfreezing the
499 * other events. We try to keep the values from the limited
500 * events as consistent as possible by keeping the delay (in
501 * cycles and instructions) between freezing/unfreezing and reading
502 * the limited events as small and consistent as possible.
503 * Therefore, if any limited events are in use, we read them
504 * both, and always in the same order, to minimize variability,
505 * and do it inside the same asm that writes MMCR0.
507 static void write_mmcr0(struct cpu_hw_events *cpuhw, unsigned long mmcr0)
509 unsigned long pmc5, pmc6;
511 if (!cpuhw->n_limited) {
512 mtspr(SPRN_MMCR0, mmcr0);
513 return;
517 * Write MMCR0, then read PMC5 and PMC6 immediately.
518 * To ensure we don't get a performance monitor interrupt
519 * between writing MMCR0 and freezing/thawing the limited
520 * events, we first write MMCR0 with the event overflow
521 * interrupt enable bits turned off.
523 asm volatile("mtspr %3,%2; mfspr %0,%4; mfspr %1,%5"
524 : "=&r" (pmc5), "=&r" (pmc6)
525 : "r" (mmcr0 & ~(MMCR0_PMC1CE | MMCR0_PMCjCE)),
526 "i" (SPRN_MMCR0),
527 "i" (SPRN_PMC5), "i" (SPRN_PMC6));
529 if (mmcr0 & MMCR0_FC)
530 freeze_limited_counters(cpuhw, pmc5, pmc6);
531 else
532 thaw_limited_counters(cpuhw, pmc5, pmc6);
535 * Write the full MMCR0 including the event overflow interrupt
536 * enable bits, if necessary.
538 if (mmcr0 & (MMCR0_PMC1CE | MMCR0_PMCjCE))
539 mtspr(SPRN_MMCR0, mmcr0);
543 * Disable all events to prevent PMU interrupts and to allow
544 * events to be added or removed.
546 static void power_pmu_disable(struct pmu *pmu)
548 struct cpu_hw_events *cpuhw;
549 unsigned long flags;
551 if (!ppmu)
552 return;
553 local_irq_save(flags);
554 cpuhw = &__get_cpu_var(cpu_hw_events);
556 if (!cpuhw->disabled) {
557 cpuhw->disabled = 1;
558 cpuhw->n_added = 0;
561 * Check if we ever enabled the PMU on this cpu.
563 if (!cpuhw->pmcs_enabled) {
564 ppc_enable_pmcs();
565 cpuhw->pmcs_enabled = 1;
569 * Disable instruction sampling if it was enabled
571 if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
572 mtspr(SPRN_MMCRA,
573 cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
574 mb();
578 * Set the 'freeze counters' bit.
579 * The barrier is to make sure the mtspr has been
580 * executed and the PMU has frozen the events
581 * before we return.
583 write_mmcr0(cpuhw, mfspr(SPRN_MMCR0) | MMCR0_FC);
584 mb();
586 local_irq_restore(flags);
590 * Re-enable all events if disable == 0.
591 * If we were previously disabled and events were added, then
592 * put the new config on the PMU.
594 static void power_pmu_enable(struct pmu *pmu)
596 struct perf_event *event;
597 struct cpu_hw_events *cpuhw;
598 unsigned long flags;
599 long i;
600 unsigned long val;
601 s64 left;
602 unsigned int hwc_index[MAX_HWEVENTS];
603 int n_lim;
604 int idx;
606 if (!ppmu)
607 return;
608 local_irq_save(flags);
609 cpuhw = &__get_cpu_var(cpu_hw_events);
610 if (!cpuhw->disabled) {
611 local_irq_restore(flags);
612 return;
614 cpuhw->disabled = 0;
617 * If we didn't change anything, or only removed events,
618 * no need to recalculate MMCR* settings and reset the PMCs.
619 * Just reenable the PMU with the current MMCR* settings
620 * (possibly updated for removal of events).
622 if (!cpuhw->n_added) {
623 mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
624 mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
625 if (cpuhw->n_events == 0)
626 ppc_set_pmu_inuse(0);
627 goto out_enable;
631 * Compute MMCR* values for the new set of events
633 if (ppmu->compute_mmcr(cpuhw->events, cpuhw->n_events, hwc_index,
634 cpuhw->mmcr)) {
635 /* shouldn't ever get here */
636 printk(KERN_ERR "oops compute_mmcr failed\n");
637 goto out;
641 * Add in MMCR0 freeze bits corresponding to the
642 * attr.exclude_* bits for the first event.
643 * We have already checked that all events have the
644 * same values for these bits as the first event.
646 event = cpuhw->event[0];
647 if (event->attr.exclude_user)
648 cpuhw->mmcr[0] |= MMCR0_FCP;
649 if (event->attr.exclude_kernel)
650 cpuhw->mmcr[0] |= freeze_events_kernel;
651 if (event->attr.exclude_hv)
652 cpuhw->mmcr[0] |= MMCR0_FCHV;
655 * Write the new configuration to MMCR* with the freeze
656 * bit set and set the hardware events to their initial values.
657 * Then unfreeze the events.
659 ppc_set_pmu_inuse(1);
660 mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
661 mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
662 mtspr(SPRN_MMCR0, (cpuhw->mmcr[0] & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
663 | MMCR0_FC);
666 * Read off any pre-existing events that need to move
667 * to another PMC.
669 for (i = 0; i < cpuhw->n_events; ++i) {
670 event = cpuhw->event[i];
671 if (event->hw.idx && event->hw.idx != hwc_index[i] + 1) {
672 power_pmu_read(event);
673 write_pmc(event->hw.idx, 0);
674 event->hw.idx = 0;
679 * Initialize the PMCs for all the new and moved events.
681 cpuhw->n_limited = n_lim = 0;
682 for (i = 0; i < cpuhw->n_events; ++i) {
683 event = cpuhw->event[i];
684 if (event->hw.idx)
685 continue;
686 idx = hwc_index[i] + 1;
687 if (is_limited_pmc(idx)) {
688 cpuhw->limited_counter[n_lim] = event;
689 cpuhw->limited_hwidx[n_lim] = idx;
690 ++n_lim;
691 continue;
693 val = 0;
694 if (event->hw.sample_period) {
695 left = local64_read(&event->hw.period_left);
696 if (left < 0x80000000L)
697 val = 0x80000000L - left;
699 local64_set(&event->hw.prev_count, val);
700 event->hw.idx = idx;
701 if (event->hw.state & PERF_HES_STOPPED)
702 val = 0;
703 write_pmc(idx, val);
704 perf_event_update_userpage(event);
706 cpuhw->n_limited = n_lim;
707 cpuhw->mmcr[0] |= MMCR0_PMXE | MMCR0_FCECE;
709 out_enable:
710 mb();
711 write_mmcr0(cpuhw, cpuhw->mmcr[0]);
714 * Enable instruction sampling if necessary
716 if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
717 mb();
718 mtspr(SPRN_MMCRA, cpuhw->mmcr[2]);
721 out:
722 local_irq_restore(flags);
725 static int collect_events(struct perf_event *group, int max_count,
726 struct perf_event *ctrs[], u64 *events,
727 unsigned int *flags)
729 int n = 0;
730 struct perf_event *event;
732 if (!is_software_event(group)) {
733 if (n >= max_count)
734 return -1;
735 ctrs[n] = group;
736 flags[n] = group->hw.event_base;
737 events[n++] = group->hw.config;
739 list_for_each_entry(event, &group->sibling_list, group_entry) {
740 if (!is_software_event(event) &&
741 event->state != PERF_EVENT_STATE_OFF) {
742 if (n >= max_count)
743 return -1;
744 ctrs[n] = event;
745 flags[n] = event->hw.event_base;
746 events[n++] = event->hw.config;
749 return n;
753 * Add a event to the PMU.
754 * If all events are not already frozen, then we disable and
755 * re-enable the PMU in order to get hw_perf_enable to do the
756 * actual work of reconfiguring the PMU.
758 static int power_pmu_add(struct perf_event *event, int ef_flags)
760 struct cpu_hw_events *cpuhw;
761 unsigned long flags;
762 int n0;
763 int ret = -EAGAIN;
765 local_irq_save(flags);
766 perf_pmu_disable(event->pmu);
769 * Add the event to the list (if there is room)
770 * and check whether the total set is still feasible.
772 cpuhw = &__get_cpu_var(cpu_hw_events);
773 n0 = cpuhw->n_events;
774 if (n0 >= ppmu->n_counter)
775 goto out;
776 cpuhw->event[n0] = event;
777 cpuhw->events[n0] = event->hw.config;
778 cpuhw->flags[n0] = event->hw.event_base;
780 if (!(ef_flags & PERF_EF_START))
781 event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
784 * If group events scheduling transaction was started,
785 * skip the schedulability test here, it will be performed
786 * at commit time(->commit_txn) as a whole
788 if (cpuhw->group_flag & PERF_EVENT_TXN)
789 goto nocheck;
791 if (check_excludes(cpuhw->event, cpuhw->flags, n0, 1))
792 goto out;
793 if (power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n0 + 1))
794 goto out;
795 event->hw.config = cpuhw->events[n0];
797 nocheck:
798 ++cpuhw->n_events;
799 ++cpuhw->n_added;
801 ret = 0;
802 out:
803 perf_pmu_enable(event->pmu);
804 local_irq_restore(flags);
805 return ret;
809 * Remove a event from the PMU.
811 static void power_pmu_del(struct perf_event *event, int ef_flags)
813 struct cpu_hw_events *cpuhw;
814 long i;
815 unsigned long flags;
817 local_irq_save(flags);
818 perf_pmu_disable(event->pmu);
820 power_pmu_read(event);
822 cpuhw = &__get_cpu_var(cpu_hw_events);
823 for (i = 0; i < cpuhw->n_events; ++i) {
824 if (event == cpuhw->event[i]) {
825 while (++i < cpuhw->n_events) {
826 cpuhw->event[i-1] = cpuhw->event[i];
827 cpuhw->events[i-1] = cpuhw->events[i];
828 cpuhw->flags[i-1] = cpuhw->flags[i];
830 --cpuhw->n_events;
831 ppmu->disable_pmc(event->hw.idx - 1, cpuhw->mmcr);
832 if (event->hw.idx) {
833 write_pmc(event->hw.idx, 0);
834 event->hw.idx = 0;
836 perf_event_update_userpage(event);
837 break;
840 for (i = 0; i < cpuhw->n_limited; ++i)
841 if (event == cpuhw->limited_counter[i])
842 break;
843 if (i < cpuhw->n_limited) {
844 while (++i < cpuhw->n_limited) {
845 cpuhw->limited_counter[i-1] = cpuhw->limited_counter[i];
846 cpuhw->limited_hwidx[i-1] = cpuhw->limited_hwidx[i];
848 --cpuhw->n_limited;
850 if (cpuhw->n_events == 0) {
851 /* disable exceptions if no events are running */
852 cpuhw->mmcr[0] &= ~(MMCR0_PMXE | MMCR0_FCECE);
855 perf_pmu_enable(event->pmu);
856 local_irq_restore(flags);
860 * POWER-PMU does not support disabling individual counters, hence
861 * program their cycle counter to their max value and ignore the interrupts.
864 static void power_pmu_start(struct perf_event *event, int ef_flags)
866 unsigned long flags;
867 s64 left;
869 if (!event->hw.idx || !event->hw.sample_period)
870 return;
872 if (!(event->hw.state & PERF_HES_STOPPED))
873 return;
875 if (ef_flags & PERF_EF_RELOAD)
876 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
878 local_irq_save(flags);
879 perf_pmu_disable(event->pmu);
881 event->hw.state = 0;
882 left = local64_read(&event->hw.period_left);
883 write_pmc(event->hw.idx, left);
885 perf_event_update_userpage(event);
886 perf_pmu_enable(event->pmu);
887 local_irq_restore(flags);
890 static void power_pmu_stop(struct perf_event *event, int ef_flags)
892 unsigned long flags;
894 if (!event->hw.idx || !event->hw.sample_period)
895 return;
897 if (event->hw.state & PERF_HES_STOPPED)
898 return;
900 local_irq_save(flags);
901 perf_pmu_disable(event->pmu);
903 power_pmu_read(event);
904 event->hw.state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
905 write_pmc(event->hw.idx, 0);
907 perf_event_update_userpage(event);
908 perf_pmu_enable(event->pmu);
909 local_irq_restore(flags);
913 * Start group events scheduling transaction
914 * Set the flag to make pmu::enable() not perform the
915 * schedulability test, it will be performed at commit time
917 void power_pmu_start_txn(struct pmu *pmu)
919 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
921 perf_pmu_disable(pmu);
922 cpuhw->group_flag |= PERF_EVENT_TXN;
923 cpuhw->n_txn_start = cpuhw->n_events;
927 * Stop group events scheduling transaction
928 * Clear the flag and pmu::enable() will perform the
929 * schedulability test.
931 void power_pmu_cancel_txn(struct pmu *pmu)
933 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
935 cpuhw->group_flag &= ~PERF_EVENT_TXN;
936 perf_pmu_enable(pmu);
940 * Commit group events scheduling transaction
941 * Perform the group schedulability test as a whole
942 * Return 0 if success
944 int power_pmu_commit_txn(struct pmu *pmu)
946 struct cpu_hw_events *cpuhw;
947 long i, n;
949 if (!ppmu)
950 return -EAGAIN;
951 cpuhw = &__get_cpu_var(cpu_hw_events);
952 n = cpuhw->n_events;
953 if (check_excludes(cpuhw->event, cpuhw->flags, 0, n))
954 return -EAGAIN;
955 i = power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n);
956 if (i < 0)
957 return -EAGAIN;
959 for (i = cpuhw->n_txn_start; i < n; ++i)
960 cpuhw->event[i]->hw.config = cpuhw->events[i];
962 cpuhw->group_flag &= ~PERF_EVENT_TXN;
963 perf_pmu_enable(pmu);
964 return 0;
968 * Return 1 if we might be able to put event on a limited PMC,
969 * or 0 if not.
970 * A event can only go on a limited PMC if it counts something
971 * that a limited PMC can count, doesn't require interrupts, and
972 * doesn't exclude any processor mode.
974 static int can_go_on_limited_pmc(struct perf_event *event, u64 ev,
975 unsigned int flags)
977 int n;
978 u64 alt[MAX_EVENT_ALTERNATIVES];
980 if (event->attr.exclude_user
981 || event->attr.exclude_kernel
982 || event->attr.exclude_hv
983 || event->attr.sample_period)
984 return 0;
986 if (ppmu->limited_pmc_event(ev))
987 return 1;
990 * The requested event_id isn't on a limited PMC already;
991 * see if any alternative code goes on a limited PMC.
993 if (!ppmu->get_alternatives)
994 return 0;
996 flags |= PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD;
997 n = ppmu->get_alternatives(ev, flags, alt);
999 return n > 0;
1003 * Find an alternative event_id that goes on a normal PMC, if possible,
1004 * and return the event_id code, or 0 if there is no such alternative.
1005 * (Note: event_id code 0 is "don't count" on all machines.)
1007 static u64 normal_pmc_alternative(u64 ev, unsigned long flags)
1009 u64 alt[MAX_EVENT_ALTERNATIVES];
1010 int n;
1012 flags &= ~(PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD);
1013 n = ppmu->get_alternatives(ev, flags, alt);
1014 if (!n)
1015 return 0;
1016 return alt[0];
1019 /* Number of perf_events counting hardware events */
1020 static atomic_t num_events;
1021 /* Used to avoid races in calling reserve/release_pmc_hardware */
1022 static DEFINE_MUTEX(pmc_reserve_mutex);
1025 * Release the PMU if this is the last perf_event.
1027 static void hw_perf_event_destroy(struct perf_event *event)
1029 if (!atomic_add_unless(&num_events, -1, 1)) {
1030 mutex_lock(&pmc_reserve_mutex);
1031 if (atomic_dec_return(&num_events) == 0)
1032 release_pmc_hardware();
1033 mutex_unlock(&pmc_reserve_mutex);
1038 * Translate a generic cache event_id config to a raw event_id code.
1040 static int hw_perf_cache_event(u64 config, u64 *eventp)
1042 unsigned long type, op, result;
1043 int ev;
1045 if (!ppmu->cache_events)
1046 return -EINVAL;
1048 /* unpack config */
1049 type = config & 0xff;
1050 op = (config >> 8) & 0xff;
1051 result = (config >> 16) & 0xff;
1053 if (type >= PERF_COUNT_HW_CACHE_MAX ||
1054 op >= PERF_COUNT_HW_CACHE_OP_MAX ||
1055 result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1056 return -EINVAL;
1058 ev = (*ppmu->cache_events)[type][op][result];
1059 if (ev == 0)
1060 return -EOPNOTSUPP;
1061 if (ev == -1)
1062 return -EINVAL;
1063 *eventp = ev;
1064 return 0;
1067 static int power_pmu_event_init(struct perf_event *event)
1069 u64 ev;
1070 unsigned long flags;
1071 struct perf_event *ctrs[MAX_HWEVENTS];
1072 u64 events[MAX_HWEVENTS];
1073 unsigned int cflags[MAX_HWEVENTS];
1074 int n;
1075 int err;
1076 struct cpu_hw_events *cpuhw;
1078 if (!ppmu)
1079 return -ENOENT;
1081 switch (event->attr.type) {
1082 case PERF_TYPE_HARDWARE:
1083 ev = event->attr.config;
1084 if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
1085 return -EOPNOTSUPP;
1086 ev = ppmu->generic_events[ev];
1087 break;
1088 case PERF_TYPE_HW_CACHE:
1089 err = hw_perf_cache_event(event->attr.config, &ev);
1090 if (err)
1091 return err;
1092 break;
1093 case PERF_TYPE_RAW:
1094 ev = event->attr.config;
1095 break;
1096 default:
1097 return -ENOENT;
1100 event->hw.config_base = ev;
1101 event->hw.idx = 0;
1104 * If we are not running on a hypervisor, force the
1105 * exclude_hv bit to 0 so that we don't care what
1106 * the user set it to.
1108 if (!firmware_has_feature(FW_FEATURE_LPAR))
1109 event->attr.exclude_hv = 0;
1112 * If this is a per-task event, then we can use
1113 * PM_RUN_* events interchangeably with their non RUN_*
1114 * equivalents, e.g. PM_RUN_CYC instead of PM_CYC.
1115 * XXX we should check if the task is an idle task.
1117 flags = 0;
1118 if (event->attach_state & PERF_ATTACH_TASK)
1119 flags |= PPMU_ONLY_COUNT_RUN;
1122 * If this machine has limited events, check whether this
1123 * event_id could go on a limited event.
1125 if (ppmu->flags & PPMU_LIMITED_PMC5_6) {
1126 if (can_go_on_limited_pmc(event, ev, flags)) {
1127 flags |= PPMU_LIMITED_PMC_OK;
1128 } else if (ppmu->limited_pmc_event(ev)) {
1130 * The requested event_id is on a limited PMC,
1131 * but we can't use a limited PMC; see if any
1132 * alternative goes on a normal PMC.
1134 ev = normal_pmc_alternative(ev, flags);
1135 if (!ev)
1136 return -EINVAL;
1141 * If this is in a group, check if it can go on with all the
1142 * other hardware events in the group. We assume the event
1143 * hasn't been linked into its leader's sibling list at this point.
1145 n = 0;
1146 if (event->group_leader != event) {
1147 n = collect_events(event->group_leader, ppmu->n_counter - 1,
1148 ctrs, events, cflags);
1149 if (n < 0)
1150 return -EINVAL;
1152 events[n] = ev;
1153 ctrs[n] = event;
1154 cflags[n] = flags;
1155 if (check_excludes(ctrs, cflags, n, 1))
1156 return -EINVAL;
1158 cpuhw = &get_cpu_var(cpu_hw_events);
1159 err = power_check_constraints(cpuhw, events, cflags, n + 1);
1160 put_cpu_var(cpu_hw_events);
1161 if (err)
1162 return -EINVAL;
1164 event->hw.config = events[n];
1165 event->hw.event_base = cflags[n];
1166 event->hw.last_period = event->hw.sample_period;
1167 local64_set(&event->hw.period_left, event->hw.last_period);
1170 * See if we need to reserve the PMU.
1171 * If no events are currently in use, then we have to take a
1172 * mutex to ensure that we don't race with another task doing
1173 * reserve_pmc_hardware or release_pmc_hardware.
1175 err = 0;
1176 if (!atomic_inc_not_zero(&num_events)) {
1177 mutex_lock(&pmc_reserve_mutex);
1178 if (atomic_read(&num_events) == 0 &&
1179 reserve_pmc_hardware(perf_event_interrupt))
1180 err = -EBUSY;
1181 else
1182 atomic_inc(&num_events);
1183 mutex_unlock(&pmc_reserve_mutex);
1185 event->destroy = hw_perf_event_destroy;
1187 return err;
1190 struct pmu power_pmu = {
1191 .pmu_enable = power_pmu_enable,
1192 .pmu_disable = power_pmu_disable,
1193 .event_init = power_pmu_event_init,
1194 .add = power_pmu_add,
1195 .del = power_pmu_del,
1196 .start = power_pmu_start,
1197 .stop = power_pmu_stop,
1198 .read = power_pmu_read,
1199 .start_txn = power_pmu_start_txn,
1200 .cancel_txn = power_pmu_cancel_txn,
1201 .commit_txn = power_pmu_commit_txn,
1205 * A counter has overflowed; update its count and record
1206 * things if requested. Note that interrupts are hard-disabled
1207 * here so there is no possibility of being interrupted.
1209 static void record_and_restart(struct perf_event *event, unsigned long val,
1210 struct pt_regs *regs, int nmi)
1212 u64 period = event->hw.sample_period;
1213 s64 prev, delta, left;
1214 int record = 0;
1216 if (event->hw.state & PERF_HES_STOPPED) {
1217 write_pmc(event->hw.idx, 0);
1218 return;
1221 /* we don't have to worry about interrupts here */
1222 prev = local64_read(&event->hw.prev_count);
1223 delta = check_and_compute_delta(prev, val);
1224 local64_add(delta, &event->count);
1227 * See if the total period for this event has expired,
1228 * and update for the next period.
1230 val = 0;
1231 left = local64_read(&event->hw.period_left) - delta;
1232 if (period) {
1233 if (left <= 0) {
1234 left += period;
1235 if (left <= 0)
1236 left = period;
1237 record = 1;
1238 event->hw.last_period = event->hw.sample_period;
1240 if (left < 0x80000000LL)
1241 val = 0x80000000LL - left;
1244 write_pmc(event->hw.idx, val);
1245 local64_set(&event->hw.prev_count, val);
1246 local64_set(&event->hw.period_left, left);
1247 perf_event_update_userpage(event);
1250 * Finally record data if requested.
1252 if (record) {
1253 struct perf_sample_data data;
1255 perf_sample_data_init(&data, ~0ULL);
1256 data.period = event->hw.last_period;
1258 if (event->attr.sample_type & PERF_SAMPLE_ADDR)
1259 perf_get_data_addr(regs, &data.addr);
1261 if (perf_event_overflow(event, nmi, &data, regs))
1262 power_pmu_stop(event, 0);
1267 * Called from generic code to get the misc flags (i.e. processor mode)
1268 * for an event_id.
1270 unsigned long perf_misc_flags(struct pt_regs *regs)
1272 u32 flags = perf_get_misc_flags(regs);
1274 if (flags)
1275 return flags;
1276 return user_mode(regs) ? PERF_RECORD_MISC_USER :
1277 PERF_RECORD_MISC_KERNEL;
1281 * Called from generic code to get the instruction pointer
1282 * for an event_id.
1284 unsigned long perf_instruction_pointer(struct pt_regs *regs)
1286 unsigned long ip;
1288 if (TRAP(regs) != 0xf00)
1289 return regs->nip; /* not a PMU interrupt */
1291 ip = mfspr(SPRN_SIAR) + perf_ip_adjust(regs);
1292 return ip;
1295 static bool pmc_overflow(unsigned long val)
1297 if ((int)val < 0)
1298 return true;
1301 * Events on POWER7 can roll back if a speculative event doesn't
1302 * eventually complete. Unfortunately in some rare cases they will
1303 * raise a performance monitor exception. We need to catch this to
1304 * ensure we reset the PMC. In all cases the PMC will be 256 or less
1305 * cycles from overflow.
1307 * We only do this if the first pass fails to find any overflowing
1308 * PMCs because a user might set a period of less than 256 and we
1309 * don't want to mistakenly reset them.
1311 if (__is_processor(PV_POWER7) && ((0x80000000 - val) <= 256))
1312 return true;
1314 return false;
1318 * Performance monitor interrupt stuff
1320 static void perf_event_interrupt(struct pt_regs *regs)
1322 int i;
1323 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1324 struct perf_event *event;
1325 unsigned long val;
1326 int found = 0;
1327 int nmi;
1329 if (cpuhw->n_limited)
1330 freeze_limited_counters(cpuhw, mfspr(SPRN_PMC5),
1331 mfspr(SPRN_PMC6));
1333 perf_read_regs(regs);
1335 nmi = perf_intr_is_nmi(regs);
1336 if (nmi)
1337 nmi_enter();
1338 else
1339 irq_enter();
1341 for (i = 0; i < cpuhw->n_events; ++i) {
1342 event = cpuhw->event[i];
1343 if (!event->hw.idx || is_limited_pmc(event->hw.idx))
1344 continue;
1345 val = read_pmc(event->hw.idx);
1346 if ((int)val < 0) {
1347 /* event has overflowed */
1348 found = 1;
1349 record_and_restart(event, val, regs, nmi);
1354 * In case we didn't find and reset the event that caused
1355 * the interrupt, scan all events and reset any that are
1356 * negative, to avoid getting continual interrupts.
1357 * Any that we processed in the previous loop will not be negative.
1359 if (!found) {
1360 for (i = 0; i < ppmu->n_counter; ++i) {
1361 if (is_limited_pmc(i + 1))
1362 continue;
1363 val = read_pmc(i + 1);
1364 if (pmc_overflow(val))
1365 write_pmc(i + 1, 0);
1370 * Reset MMCR0 to its normal value. This will set PMXE and
1371 * clear FC (freeze counters) and PMAO (perf mon alert occurred)
1372 * and thus allow interrupts to occur again.
1373 * XXX might want to use MSR.PM to keep the events frozen until
1374 * we get back out of this interrupt.
1376 write_mmcr0(cpuhw, cpuhw->mmcr[0]);
1378 if (nmi)
1379 nmi_exit();
1380 else
1381 irq_exit();
1384 static void power_pmu_setup(int cpu)
1386 struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
1388 if (!ppmu)
1389 return;
1390 memset(cpuhw, 0, sizeof(*cpuhw));
1391 cpuhw->mmcr[0] = MMCR0_FC;
1394 static int __cpuinit
1395 power_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1397 unsigned int cpu = (long)hcpu;
1399 switch (action & ~CPU_TASKS_FROZEN) {
1400 case CPU_UP_PREPARE:
1401 power_pmu_setup(cpu);
1402 break;
1404 default:
1405 break;
1408 return NOTIFY_OK;
1411 int register_power_pmu(struct power_pmu *pmu)
1413 if (ppmu)
1414 return -EBUSY; /* something's already registered */
1416 ppmu = pmu;
1417 pr_info("%s performance monitor hardware support registered\n",
1418 pmu->name);
1420 #ifdef MSR_HV
1422 * Use FCHV to ignore kernel events if MSR.HV is set.
1424 if (mfmsr() & MSR_HV)
1425 freeze_events_kernel = MMCR0_FCHV;
1426 #endif /* CONFIG_PPC64 */
1428 perf_pmu_register(&power_pmu, "cpu", PERF_TYPE_RAW);
1429 perf_cpu_notifier(power_pmu_notifier);
1431 return 0;