1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Performance event support - powerpc architecture code
5 * Copyright 2008-2009 Paul Mackerras, IBM Corporation.
7 #include <linux/kernel.h>
8 #include <linux/sched.h>
9 #include <linux/sched/clock.h>
10 #include <linux/perf_event.h>
11 #include <linux/percpu.h>
12 #include <linux/hardirq.h>
13 #include <linux/uaccess.h>
16 #include <asm/machdep.h>
17 #include <asm/firmware.h>
18 #include <asm/ptrace.h>
19 #include <asm/code-patching.h>
25 #define BHRB_MAX_ENTRIES 32
26 #define BHRB_TARGET 0x0000000000000002
27 #define BHRB_PREDICTION 0x0000000000000001
28 #define BHRB_EA 0xFFFFFFFFFFFFFFFCUL
30 struct cpu_hw_events {
37 struct perf_event *event[MAX_HWEVENTS];
38 u64 events[MAX_HWEVENTS];
39 unsigned int flags[MAX_HWEVENTS];
40 struct mmcr_regs mmcr;
41 struct perf_event *limited_counter[MAX_LIMITED_HWCOUNTERS];
42 u8 limited_hwidx[MAX_LIMITED_HWCOUNTERS];
43 u64 alternatives[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
44 unsigned long amasks[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
45 unsigned long avalues[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
47 unsigned int txn_flags;
51 u64 bhrb_filter; /* BHRB HW branch filter */
52 unsigned int bhrb_users;
54 struct perf_branch_stack bhrb_stack;
55 struct perf_branch_entry bhrb_entries[BHRB_MAX_ENTRIES];
58 /* Store the PMC values */
59 unsigned long pmcs[MAX_HWEVENTS];
62 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
64 static struct power_pmu *ppmu;
67 * Normally, to ignore kernel events we set the FCS (freeze counters
68 * in supervisor mode) bit in MMCR0, but if the kernel runs with the
69 * hypervisor bit set in the MSR, or if we are running on a processor
70 * where the hypervisor bit is forced to 1 (as on Apple G5 processors),
71 * then we need to use the FCHV bit to ignore kernel events.
73 static unsigned int freeze_events_kernel = MMCR0_FCS;
76 * 32-bit doesn't have MMCRA but does have an MMCR2,
77 * and a few other names are different.
78 * Also 32-bit doesn't have MMCR3, SIER2 and SIER3.
79 * Define them as zero knowing that any code path accessing
80 * these registers (via mtspr/mfspr) are done under ppmu flag
81 * check for PPMU_ARCH_31 and we will not enter that code path
87 #define MMCR0_PMCjCE MMCR0_PMCnCE
93 #define MMCR0_PMCC_U6 0
95 #define SPRN_MMCRA SPRN_MMCR2
99 #define MMCRA_SAMPLE_ENABLE 0
100 #define MMCRA_BHRB_DISABLE 0
101 #define MMCR0_PMCCEXT 0
103 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
107 static inline void perf_get_data_addr(struct perf_event *event, struct pt_regs *regs, u64 *addrp) { }
108 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
112 static inline void perf_read_regs(struct pt_regs *regs)
117 static inline int siar_valid(struct pt_regs *regs)
122 static bool is_ebb_event(struct perf_event *event) { return false; }
123 static int ebb_event_check(struct perf_event *event) { return 0; }
124 static void ebb_event_add(struct perf_event *event) { }
125 static void ebb_switch_out(unsigned long mmcr0) { }
126 static unsigned long ebb_switch_in(bool ebb, struct cpu_hw_events *cpuhw)
128 return cpuhw->mmcr.mmcr0;
131 static inline void power_pmu_bhrb_enable(struct perf_event *event) {}
132 static inline void power_pmu_bhrb_disable(struct perf_event *event) {}
133 static void power_pmu_sched_task(struct perf_event_context *ctx, bool sched_in) {}
134 static inline void power_pmu_bhrb_read(struct perf_event *event, struct cpu_hw_events *cpuhw) {}
135 static void pmao_restore_workaround(bool ebb) { }
136 #endif /* CONFIG_PPC32 */
138 bool is_sier_available(void)
143 if (ppmu->flags & PPMU_HAS_SIER)
150 * Return PMC value corresponding to the
153 unsigned long get_pmcs_ext_regs(int idx)
155 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
157 return cpuhw->pmcs[idx];
160 static bool regs_use_siar(struct pt_regs *regs)
163 * When we take a performance monitor exception the regs are setup
164 * using perf_read_regs() which overloads some fields, in particular
165 * regs->result to tell us whether to use SIAR.
167 * However if the regs are from another exception, eg. a syscall, then
168 * they have not been setup using perf_read_regs() and so regs->result
169 * is something random.
171 return ((TRAP(regs) == 0xf00) && regs->result);
175 * Things that are specific to 64-bit implementations.
179 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
181 unsigned long mmcra = regs->dsisr;
183 if ((ppmu->flags & PPMU_HAS_SSLOT) && (mmcra & MMCRA_SAMPLE_ENABLE)) {
184 unsigned long slot = (mmcra & MMCRA_SLOT) >> MMCRA_SLOT_SHIFT;
186 return 4 * (slot - 1);
193 * The user wants a data address recorded.
194 * If we're not doing instruction sampling, give them the SDAR
195 * (sampled data address). If we are doing instruction sampling, then
196 * only give them the SDAR if it corresponds to the instruction
197 * pointed to by SIAR; this is indicated by the [POWER6_]MMCRA_SDSYNC, the
198 * [POWER7P_]MMCRA_SDAR_VALID bit in MMCRA, or the SDAR_VALID bit in SIER.
200 static inline void perf_get_data_addr(struct perf_event *event, struct pt_regs *regs, u64 *addrp)
202 unsigned long mmcra = regs->dsisr;
205 if (ppmu->flags & PPMU_HAS_SIER)
206 sdar_valid = regs->dar & SIER_SDAR_VALID;
208 unsigned long sdsync;
210 if (ppmu->flags & PPMU_SIAR_VALID)
211 sdsync = POWER7P_MMCRA_SDAR_VALID;
212 else if (ppmu->flags & PPMU_ALT_SIPR)
213 sdsync = POWER6_MMCRA_SDSYNC;
214 else if (ppmu->flags & PPMU_NO_SIAR)
215 sdsync = MMCRA_SAMPLE_ENABLE;
217 sdsync = MMCRA_SDSYNC;
219 sdar_valid = mmcra & sdsync;
222 if (!(mmcra & MMCRA_SAMPLE_ENABLE) || sdar_valid)
223 *addrp = mfspr(SPRN_SDAR);
225 if (is_kernel_addr(mfspr(SPRN_SDAR)) && perf_allow_kernel(&event->attr) != 0)
229 static bool regs_sihv(struct pt_regs *regs)
231 unsigned long sihv = MMCRA_SIHV;
233 if (ppmu->flags & PPMU_HAS_SIER)
234 return !!(regs->dar & SIER_SIHV);
236 if (ppmu->flags & PPMU_ALT_SIPR)
237 sihv = POWER6_MMCRA_SIHV;
239 return !!(regs->dsisr & sihv);
242 static bool regs_sipr(struct pt_regs *regs)
244 unsigned long sipr = MMCRA_SIPR;
246 if (ppmu->flags & PPMU_HAS_SIER)
247 return !!(regs->dar & SIER_SIPR);
249 if (ppmu->flags & PPMU_ALT_SIPR)
250 sipr = POWER6_MMCRA_SIPR;
252 return !!(regs->dsisr & sipr);
255 static inline u32 perf_flags_from_msr(struct pt_regs *regs)
257 if (regs->msr & MSR_PR)
258 return PERF_RECORD_MISC_USER;
259 if ((regs->msr & MSR_HV) && freeze_events_kernel != MMCR0_FCHV)
260 return PERF_RECORD_MISC_HYPERVISOR;
261 return PERF_RECORD_MISC_KERNEL;
264 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
266 bool use_siar = regs_use_siar(regs);
267 unsigned long mmcra = regs->dsisr;
268 int marked = mmcra & MMCRA_SAMPLE_ENABLE;
271 return perf_flags_from_msr(regs);
274 * Check the address in SIAR to identify the
275 * privilege levels since the SIER[MSR_HV, MSR_PR]
276 * bits are not set for marked events in power10
279 if (marked && (ppmu->flags & PPMU_P10_DD1)) {
280 unsigned long siar = mfspr(SPRN_SIAR);
282 if (is_kernel_addr(siar))
283 return PERF_RECORD_MISC_KERNEL;
284 return PERF_RECORD_MISC_USER;
286 if (is_kernel_addr(regs->nip))
287 return PERF_RECORD_MISC_KERNEL;
288 return PERF_RECORD_MISC_USER;
293 * If we don't have flags in MMCRA, rather than using
294 * the MSR, we intuit the flags from the address in
295 * SIAR which should give slightly more reliable
298 if (ppmu->flags & PPMU_NO_SIPR) {
299 unsigned long siar = mfspr(SPRN_SIAR);
300 if (is_kernel_addr(siar))
301 return PERF_RECORD_MISC_KERNEL;
302 return PERF_RECORD_MISC_USER;
305 /* PR has priority over HV, so order below is important */
307 return PERF_RECORD_MISC_USER;
309 if (regs_sihv(regs) && (freeze_events_kernel != MMCR0_FCHV))
310 return PERF_RECORD_MISC_HYPERVISOR;
312 return PERF_RECORD_MISC_KERNEL;
316 * Overload regs->dsisr to store MMCRA so we only need to read it once
318 * Overload regs->dar to store SIER if we have it.
319 * Overload regs->result to specify whether we should use the MSR (result
320 * is zero) or the SIAR (result is non zero).
322 static inline void perf_read_regs(struct pt_regs *regs)
324 unsigned long mmcra = mfspr(SPRN_MMCRA);
325 int marked = mmcra & MMCRA_SAMPLE_ENABLE;
330 if (ppmu->flags & PPMU_HAS_SIER)
331 regs->dar = mfspr(SPRN_SIER);
334 * If this isn't a PMU exception (eg a software event) the SIAR is
335 * not valid. Use pt_regs.
337 * If it is a marked event use the SIAR.
339 * If the PMU doesn't update the SIAR for non marked events use
342 * If the PMU has HV/PR flags then check to see if they
343 * place the exception in userspace. If so, use pt_regs. In
344 * continuous sampling mode the SIAR and the PMU exception are
345 * not synchronised, so they may be many instructions apart.
346 * This can result in confusing backtraces. We still want
347 * hypervisor samples as well as samples in the kernel with
348 * interrupts off hence the userspace check.
350 if (TRAP(regs) != 0xf00)
352 else if ((ppmu->flags & PPMU_NO_SIAR))
356 else if ((ppmu->flags & PPMU_NO_CONT_SAMPLING))
358 else if (!(ppmu->flags & PPMU_NO_SIPR) && regs_sipr(regs))
363 regs->result = use_siar;
367 * On processors like P7+ that have the SIAR-Valid bit, marked instructions
368 * must be sampled only if the SIAR-valid bit is set.
370 * For unmarked instructions and for processors that don't have the SIAR-Valid
371 * bit, assume that SIAR is valid.
373 static inline int siar_valid(struct pt_regs *regs)
375 unsigned long mmcra = regs->dsisr;
376 int marked = mmcra & MMCRA_SAMPLE_ENABLE;
380 * SIER[SIAR_VALID] is not set for some
381 * marked events on power10 DD1, so drop
382 * the check for SIER[SIAR_VALID] and return true.
384 if (ppmu->flags & PPMU_P10_DD1)
386 else if (ppmu->flags & PPMU_HAS_SIER)
387 return regs->dar & SIER_SIAR_VALID;
389 if (ppmu->flags & PPMU_SIAR_VALID)
390 return mmcra & POWER7P_MMCRA_SIAR_VALID;
397 /* Reset all possible BHRB entries */
398 static void power_pmu_bhrb_reset(void)
400 asm volatile(PPC_CLRBHRB);
403 static void power_pmu_bhrb_enable(struct perf_event *event)
405 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
410 /* Clear BHRB if we changed task context to avoid data leaks */
411 if (event->ctx->task && cpuhw->bhrb_context != event->ctx) {
412 power_pmu_bhrb_reset();
413 cpuhw->bhrb_context = event->ctx;
416 perf_sched_cb_inc(event->ctx->pmu);
419 static void power_pmu_bhrb_disable(struct perf_event *event)
421 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
426 WARN_ON_ONCE(!cpuhw->bhrb_users);
428 perf_sched_cb_dec(event->ctx->pmu);
430 if (!cpuhw->disabled && !cpuhw->bhrb_users) {
431 /* BHRB cannot be turned off when other
432 * events are active on the PMU.
435 /* avoid stale pointer */
436 cpuhw->bhrb_context = NULL;
440 /* Called from ctxsw to prevent one process's branch entries to
441 * mingle with the other process's entries during context switch.
443 static void power_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
449 power_pmu_bhrb_reset();
451 /* Calculate the to address for a branch */
452 static __u64 power_pmu_bhrb_to(u64 addr)
457 if (is_kernel_addr(addr)) {
458 if (copy_from_kernel_nofault(&instr, (void *)addr,
462 return branch_target((struct ppc_inst *)&instr);
465 /* Userspace: need copy instruction here then translate it */
466 if (copy_from_user_nofault(&instr, (unsigned int __user *)addr,
470 target = branch_target((struct ppc_inst *)&instr);
471 if ((!target) || (instr & BRANCH_ABSOLUTE))
474 /* Translate relative branch target from kernel to user address */
475 return target - (unsigned long)&instr + addr;
478 /* Processing BHRB entries */
479 static void power_pmu_bhrb_read(struct perf_event *event, struct cpu_hw_events *cpuhw)
483 int r_index, u_index, pred;
487 while (r_index < ppmu->bhrb_nr) {
488 /* Assembly read function */
489 val = read_bhrb(r_index++);
491 /* Terminal marker: End of valid BHRB entries */
494 addr = val & BHRB_EA;
495 pred = val & BHRB_PREDICTION;
502 * BHRB rolling buffer could very much contain the kernel
503 * addresses at this point. Check the privileges before
504 * exporting it to userspace (avoid exposure of regions
505 * where we could have speculative execution)
506 * Incase of ISA v3.1, BHRB will capture only user-space
507 * addresses, hence include a check before filtering code
509 if (!(ppmu->flags & PPMU_ARCH_31) &&
510 is_kernel_addr(addr) && perf_allow_kernel(&event->attr) != 0)
513 /* Branches are read most recent first (ie. mfbhrb 0 is
514 * the most recent branch).
515 * There are two types of valid entries:
516 * 1) a target entry which is the to address of a
517 * computed goto like a blr,bctr,btar. The next
518 * entry read from the bhrb will be branch
519 * corresponding to this target (ie. the actual
520 * blr/bctr/btar instruction).
521 * 2) a from address which is an actual branch. If a
522 * target entry proceeds this, then this is the
523 * matching branch for that target. If this is not
524 * following a target entry, then this is a branch
525 * where the target is given as an immediate field
526 * in the instruction (ie. an i or b form branch).
527 * In this case we need to read the instruction from
528 * memory to determine the target/to address.
531 if (val & BHRB_TARGET) {
532 /* Target branches use two entries
533 * (ie. computed gotos/XL form)
535 cpuhw->bhrb_entries[u_index].to = addr;
536 cpuhw->bhrb_entries[u_index].mispred = pred;
537 cpuhw->bhrb_entries[u_index].predicted = ~pred;
539 /* Get from address in next entry */
540 val = read_bhrb(r_index++);
541 addr = val & BHRB_EA;
542 if (val & BHRB_TARGET) {
543 /* Shouldn't have two targets in a
544 row.. Reset index and try again */
548 cpuhw->bhrb_entries[u_index].from = addr;
550 /* Branches to immediate field
552 cpuhw->bhrb_entries[u_index].from = addr;
553 cpuhw->bhrb_entries[u_index].to =
554 power_pmu_bhrb_to(addr);
555 cpuhw->bhrb_entries[u_index].mispred = pred;
556 cpuhw->bhrb_entries[u_index].predicted = ~pred;
562 cpuhw->bhrb_stack.nr = u_index;
563 cpuhw->bhrb_stack.hw_idx = -1ULL;
567 static bool is_ebb_event(struct perf_event *event)
570 * This could be a per-PMU callback, but we'd rather avoid the cost. We
571 * check that the PMU supports EBB, meaning those that don't can still
572 * use bit 63 of the event code for something else if they wish.
574 return (ppmu->flags & PPMU_ARCH_207S) &&
575 ((event->attr.config >> PERF_EVENT_CONFIG_EBB_SHIFT) & 1);
578 static int ebb_event_check(struct perf_event *event)
580 struct perf_event *leader = event->group_leader;
582 /* Event and group leader must agree on EBB */
583 if (is_ebb_event(leader) != is_ebb_event(event))
586 if (is_ebb_event(event)) {
587 if (!(event->attach_state & PERF_ATTACH_TASK))
590 if (!leader->attr.pinned || !leader->attr.exclusive)
593 if (event->attr.freq ||
594 event->attr.inherit ||
595 event->attr.sample_type ||
596 event->attr.sample_period ||
597 event->attr.enable_on_exec)
604 static void ebb_event_add(struct perf_event *event)
606 if (!is_ebb_event(event) || current->thread.used_ebb)
610 * IFF this is the first time we've added an EBB event, set
611 * PMXE in the user MMCR0 so we can detect when it's cleared by
612 * userspace. We need this so that we can context switch while
613 * userspace is in the EBB handler (where PMXE is 0).
615 current->thread.used_ebb = 1;
616 current->thread.mmcr0 |= MMCR0_PMXE;
619 static void ebb_switch_out(unsigned long mmcr0)
621 if (!(mmcr0 & MMCR0_EBE))
624 current->thread.siar = mfspr(SPRN_SIAR);
625 current->thread.sier = mfspr(SPRN_SIER);
626 current->thread.sdar = mfspr(SPRN_SDAR);
627 current->thread.mmcr0 = mmcr0 & MMCR0_USER_MASK;
628 current->thread.mmcr2 = mfspr(SPRN_MMCR2) & MMCR2_USER_MASK;
629 if (ppmu->flags & PPMU_ARCH_31) {
630 current->thread.mmcr3 = mfspr(SPRN_MMCR3);
631 current->thread.sier2 = mfspr(SPRN_SIER2);
632 current->thread.sier3 = mfspr(SPRN_SIER3);
636 static unsigned long ebb_switch_in(bool ebb, struct cpu_hw_events *cpuhw)
638 unsigned long mmcr0 = cpuhw->mmcr.mmcr0;
643 /* Enable EBB and read/write to all 6 PMCs and BHRB for userspace */
644 mmcr0 |= MMCR0_EBE | MMCR0_BHRBA | MMCR0_PMCC_U6;
647 * Add any bits from the user MMCR0, FC or PMAO. This is compatible
648 * with pmao_restore_workaround() because we may add PMAO but we never
651 mmcr0 |= current->thread.mmcr0;
654 * Be careful not to set PMXE if userspace had it cleared. This is also
655 * compatible with pmao_restore_workaround() because it has already
656 * cleared PMXE and we leave PMAO alone.
658 if (!(current->thread.mmcr0 & MMCR0_PMXE))
659 mmcr0 &= ~MMCR0_PMXE;
661 mtspr(SPRN_SIAR, current->thread.siar);
662 mtspr(SPRN_SIER, current->thread.sier);
663 mtspr(SPRN_SDAR, current->thread.sdar);
666 * Merge the kernel & user values of MMCR2. The semantics we implement
667 * are that the user MMCR2 can set bits, ie. cause counters to freeze,
668 * but not clear bits. If a task wants to be able to clear bits, ie.
669 * unfreeze counters, it should not set exclude_xxx in its events and
670 * instead manage the MMCR2 entirely by itself.
672 mtspr(SPRN_MMCR2, cpuhw->mmcr.mmcr2 | current->thread.mmcr2);
674 if (ppmu->flags & PPMU_ARCH_31) {
675 mtspr(SPRN_MMCR3, current->thread.mmcr3);
676 mtspr(SPRN_SIER2, current->thread.sier2);
677 mtspr(SPRN_SIER3, current->thread.sier3);
683 static void pmao_restore_workaround(bool ebb)
687 if (!cpu_has_feature(CPU_FTR_PMAO_BUG))
691 * On POWER8E there is a hardware defect which affects the PMU context
692 * switch logic, ie. power_pmu_disable/enable().
694 * When a counter overflows PMXE is cleared and FC/PMAO is set in MMCR0
695 * by the hardware. Sometime later the actual PMU exception is
698 * If we context switch, or simply disable/enable, the PMU prior to the
699 * exception arriving, the exception will be lost when we clear PMAO.
701 * When we reenable the PMU, we will write the saved MMCR0 with PMAO
702 * set, and this _should_ generate an exception. However because of the
703 * defect no exception is generated when we write PMAO, and we get
704 * stuck with no counters counting but no exception delivered.
706 * The workaround is to detect this case and tweak the hardware to
707 * create another pending PMU exception.
709 * We do that by setting up PMC6 (cycles) for an imminent overflow and
710 * enabling the PMU. That causes a new exception to be generated in the
711 * chip, but we don't take it yet because we have interrupts hard
712 * disabled. We then write back the PMU state as we want it to be seen
713 * by the exception handler. When we reenable interrupts the exception
714 * handler will be called and see the correct state.
716 * The logic is the same for EBB, except that the exception is gated by
717 * us having interrupts hard disabled as well as the fact that we are
718 * not in userspace. The exception is finally delivered when we return
722 /* Only if PMAO is set and PMAO_SYNC is clear */
723 if ((current->thread.mmcr0 & (MMCR0_PMAO | MMCR0_PMAO_SYNC)) != MMCR0_PMAO)
726 /* If we're doing EBB, only if BESCR[GE] is set */
727 if (ebb && !(current->thread.bescr & BESCR_GE))
731 * We are already soft-disabled in power_pmu_enable(). We need to hard
732 * disable to actually prevent the PMU exception from firing.
737 * This is a bit gross, but we know we're on POWER8E and have 6 PMCs.
738 * Using read/write_pmc() in a for loop adds 12 function calls and
739 * almost doubles our code size.
741 pmcs[0] = mfspr(SPRN_PMC1);
742 pmcs[1] = mfspr(SPRN_PMC2);
743 pmcs[2] = mfspr(SPRN_PMC3);
744 pmcs[3] = mfspr(SPRN_PMC4);
745 pmcs[4] = mfspr(SPRN_PMC5);
746 pmcs[5] = mfspr(SPRN_PMC6);
748 /* Ensure all freeze bits are unset */
749 mtspr(SPRN_MMCR2, 0);
751 /* Set up PMC6 to overflow in one cycle */
752 mtspr(SPRN_PMC6, 0x7FFFFFFE);
754 /* Enable exceptions and unfreeze PMC6 */
755 mtspr(SPRN_MMCR0, MMCR0_PMXE | MMCR0_PMCjCE | MMCR0_PMAO);
757 /* Now we need to refreeze and restore the PMCs */
758 mtspr(SPRN_MMCR0, MMCR0_FC | MMCR0_PMAO);
760 mtspr(SPRN_PMC1, pmcs[0]);
761 mtspr(SPRN_PMC2, pmcs[1]);
762 mtspr(SPRN_PMC3, pmcs[2]);
763 mtspr(SPRN_PMC4, pmcs[3]);
764 mtspr(SPRN_PMC5, pmcs[4]);
765 mtspr(SPRN_PMC6, pmcs[5]);
768 #endif /* CONFIG_PPC64 */
770 static void perf_event_interrupt(struct pt_regs *regs);
773 * Read one performance monitor counter (PMC).
775 static unsigned long read_pmc(int idx)
781 val = mfspr(SPRN_PMC1);
784 val = mfspr(SPRN_PMC2);
787 val = mfspr(SPRN_PMC3);
790 val = mfspr(SPRN_PMC4);
793 val = mfspr(SPRN_PMC5);
796 val = mfspr(SPRN_PMC6);
800 val = mfspr(SPRN_PMC7);
803 val = mfspr(SPRN_PMC8);
805 #endif /* CONFIG_PPC64 */
807 printk(KERN_ERR "oops trying to read PMC%d\n", idx);
816 static void write_pmc(int idx, unsigned long val)
820 mtspr(SPRN_PMC1, val);
823 mtspr(SPRN_PMC2, val);
826 mtspr(SPRN_PMC3, val);
829 mtspr(SPRN_PMC4, val);
832 mtspr(SPRN_PMC5, val);
835 mtspr(SPRN_PMC6, val);
839 mtspr(SPRN_PMC7, val);
842 mtspr(SPRN_PMC8, val);
844 #endif /* CONFIG_PPC64 */
846 printk(KERN_ERR "oops trying to write PMC%d\n", idx);
850 /* Called from sysrq_handle_showregs() */
851 void perf_event_print_debug(void)
853 unsigned long sdar, sier, flags;
854 u32 pmcs[MAX_HWEVENTS];
858 pr_info("Performance monitor hardware not registered.\n");
862 if (!ppmu->n_counter)
865 local_irq_save(flags);
867 pr_info("CPU: %d PMU registers, ppmu = %s n_counters = %d",
868 smp_processor_id(), ppmu->name, ppmu->n_counter);
870 for (i = 0; i < ppmu->n_counter; i++)
871 pmcs[i] = read_pmc(i + 1);
873 for (; i < MAX_HWEVENTS; i++)
874 pmcs[i] = 0xdeadbeef;
876 pr_info("PMC1: %08x PMC2: %08x PMC3: %08x PMC4: %08x\n",
877 pmcs[0], pmcs[1], pmcs[2], pmcs[3]);
879 if (ppmu->n_counter > 4)
880 pr_info("PMC5: %08x PMC6: %08x PMC7: %08x PMC8: %08x\n",
881 pmcs[4], pmcs[5], pmcs[6], pmcs[7]);
883 pr_info("MMCR0: %016lx MMCR1: %016lx MMCRA: %016lx\n",
884 mfspr(SPRN_MMCR0), mfspr(SPRN_MMCR1), mfspr(SPRN_MMCRA));
888 sdar = mfspr(SPRN_SDAR);
890 if (ppmu->flags & PPMU_HAS_SIER)
891 sier = mfspr(SPRN_SIER);
893 if (ppmu->flags & PPMU_ARCH_207S) {
894 pr_info("MMCR2: %016lx EBBHR: %016lx\n",
895 mfspr(SPRN_MMCR2), mfspr(SPRN_EBBHR));
896 pr_info("EBBRR: %016lx BESCR: %016lx\n",
897 mfspr(SPRN_EBBRR), mfspr(SPRN_BESCR));
900 if (ppmu->flags & PPMU_ARCH_31) {
901 pr_info("MMCR3: %016lx SIER2: %016lx SIER3: %016lx\n",
902 mfspr(SPRN_MMCR3), mfspr(SPRN_SIER2), mfspr(SPRN_SIER3));
905 pr_info("SIAR: %016lx SDAR: %016lx SIER: %016lx\n",
906 mfspr(SPRN_SIAR), sdar, sier);
908 local_irq_restore(flags);
912 * Check if a set of events can all go on the PMU at once.
913 * If they can't, this will look at alternative codes for the events
914 * and see if any combination of alternative codes is feasible.
915 * The feasible set is returned in event_id[].
917 static int power_check_constraints(struct cpu_hw_events *cpuhw,
918 u64 event_id[], unsigned int cflags[],
919 int n_ev, struct perf_event **event)
921 unsigned long mask, value, nv;
922 unsigned long smasks[MAX_HWEVENTS], svalues[MAX_HWEVENTS];
923 int n_alt[MAX_HWEVENTS], choice[MAX_HWEVENTS];
925 unsigned long addf = ppmu->add_fields;
926 unsigned long tadd = ppmu->test_adder;
927 unsigned long grp_mask = ppmu->group_constraint_mask;
928 unsigned long grp_val = ppmu->group_constraint_val;
930 if (n_ev > ppmu->n_counter)
933 /* First see if the events will go on as-is */
934 for (i = 0; i < n_ev; ++i) {
935 if ((cflags[i] & PPMU_LIMITED_PMC_REQD)
936 && !ppmu->limited_pmc_event(event_id[i])) {
937 ppmu->get_alternatives(event_id[i], cflags[i],
938 cpuhw->alternatives[i]);
939 event_id[i] = cpuhw->alternatives[i][0];
941 if (ppmu->get_constraint(event_id[i], &cpuhw->amasks[i][0],
942 &cpuhw->avalues[i][0], event[i]->attr.config1))
946 for (i = 0; i < n_ev; ++i) {
947 nv = (value | cpuhw->avalues[i][0]) +
948 (value & cpuhw->avalues[i][0] & addf);
950 if (((((nv + tadd) ^ value) & mask) & (~grp_mask)) != 0)
953 if (((((nv + tadd) ^ cpuhw->avalues[i][0]) & cpuhw->amasks[i][0])
958 mask |= cpuhw->amasks[i][0];
961 if ((value & mask & grp_mask) != (mask & grp_val))
964 return 0; /* all OK */
967 /* doesn't work, gather alternatives... */
968 if (!ppmu->get_alternatives)
970 for (i = 0; i < n_ev; ++i) {
972 n_alt[i] = ppmu->get_alternatives(event_id[i], cflags[i],
973 cpuhw->alternatives[i]);
974 for (j = 1; j < n_alt[i]; ++j)
975 ppmu->get_constraint(cpuhw->alternatives[i][j],
976 &cpuhw->amasks[i][j],
977 &cpuhw->avalues[i][j],
978 event[i]->attr.config1);
981 /* enumerate all possibilities and see if any will work */
984 value = mask = nv = 0;
987 /* we're backtracking, restore context */
993 * See if any alternative k for event_id i,
994 * where k > j, will satisfy the constraints.
996 while (++j < n_alt[i]) {
997 nv = (value | cpuhw->avalues[i][j]) +
998 (value & cpuhw->avalues[i][j] & addf);
999 if ((((nv + tadd) ^ value) & mask) == 0 &&
1000 (((nv + tadd) ^ cpuhw->avalues[i][j])
1001 & cpuhw->amasks[i][j]) == 0)
1004 if (j >= n_alt[i]) {
1006 * No feasible alternative, backtrack
1007 * to event_id i-1 and continue enumerating its
1008 * alternatives from where we got up to.
1014 * Found a feasible alternative for event_id i,
1015 * remember where we got up to with this event_id,
1016 * go on to the next event_id, and start with
1017 * the first alternative for it.
1023 mask |= cpuhw->amasks[i][j];
1029 /* OK, we have a feasible combination, tell the caller the solution */
1030 for (i = 0; i < n_ev; ++i)
1031 event_id[i] = cpuhw->alternatives[i][choice[i]];
1036 * Check if newly-added events have consistent settings for
1037 * exclude_{user,kernel,hv} with each other and any previously
1040 static int check_excludes(struct perf_event **ctrs, unsigned int cflags[],
1041 int n_prev, int n_new)
1043 int eu = 0, ek = 0, eh = 0;
1045 struct perf_event *event;
1048 * If the PMU we're on supports per event exclude settings then we
1049 * don't need to do any of this logic. NB. This assumes no PMU has both
1050 * per event exclude and limited PMCs.
1052 if (ppmu->flags & PPMU_ARCH_207S)
1060 for (i = 0; i < n; ++i) {
1061 if (cflags[i] & PPMU_LIMITED_PMC_OK) {
1062 cflags[i] &= ~PPMU_LIMITED_PMC_REQD;
1067 eu = event->attr.exclude_user;
1068 ek = event->attr.exclude_kernel;
1069 eh = event->attr.exclude_hv;
1071 } else if (event->attr.exclude_user != eu ||
1072 event->attr.exclude_kernel != ek ||
1073 event->attr.exclude_hv != eh) {
1079 for (i = 0; i < n; ++i)
1080 if (cflags[i] & PPMU_LIMITED_PMC_OK)
1081 cflags[i] |= PPMU_LIMITED_PMC_REQD;
1086 static u64 check_and_compute_delta(u64 prev, u64 val)
1088 u64 delta = (val - prev) & 0xfffffffful;
1091 * POWER7 can roll back counter values, if the new value is smaller
1092 * than the previous value it will cause the delta and the counter to
1093 * have bogus values unless we rolled a counter over. If a coutner is
1094 * rolled back, it will be smaller, but within 256, which is the maximum
1095 * number of events to rollback at once. If we detect a rollback
1096 * return 0. This can lead to a small lack of precision in the
1099 if (prev > val && (prev - val) < 256)
1105 static void power_pmu_read(struct perf_event *event)
1107 s64 val, delta, prev;
1109 if (event->hw.state & PERF_HES_STOPPED)
1115 if (is_ebb_event(event)) {
1116 val = read_pmc(event->hw.idx);
1117 local64_set(&event->hw.prev_count, val);
1122 * Performance monitor interrupts come even when interrupts
1123 * are soft-disabled, as long as interrupts are hard-enabled.
1124 * Therefore we treat them like NMIs.
1127 prev = local64_read(&event->hw.prev_count);
1129 val = read_pmc(event->hw.idx);
1130 delta = check_and_compute_delta(prev, val);
1133 } while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
1135 local64_add(delta, &event->count);
1138 * A number of places program the PMC with (0x80000000 - period_left).
1139 * We never want period_left to be less than 1 because we will program
1140 * the PMC with a value >= 0x800000000 and an edge detected PMC will
1141 * roll around to 0 before taking an exception. We have seen this
1144 * To fix this, clamp the minimum value of period_left to 1.
1147 prev = local64_read(&event->hw.period_left);
1151 } while (local64_cmpxchg(&event->hw.period_left, prev, val) != prev);
1155 * On some machines, PMC5 and PMC6 can't be written, don't respect
1156 * the freeze conditions, and don't generate interrupts. This tells
1157 * us if `event' is using such a PMC.
1159 static int is_limited_pmc(int pmcnum)
1161 return (ppmu->flags & PPMU_LIMITED_PMC5_6)
1162 && (pmcnum == 5 || pmcnum == 6);
1165 static void freeze_limited_counters(struct cpu_hw_events *cpuhw,
1166 unsigned long pmc5, unsigned long pmc6)
1168 struct perf_event *event;
1169 u64 val, prev, delta;
1172 for (i = 0; i < cpuhw->n_limited; ++i) {
1173 event = cpuhw->limited_counter[i];
1176 val = (event->hw.idx == 5) ? pmc5 : pmc6;
1177 prev = local64_read(&event->hw.prev_count);
1179 delta = check_and_compute_delta(prev, val);
1181 local64_add(delta, &event->count);
1185 static void thaw_limited_counters(struct cpu_hw_events *cpuhw,
1186 unsigned long pmc5, unsigned long pmc6)
1188 struct perf_event *event;
1192 for (i = 0; i < cpuhw->n_limited; ++i) {
1193 event = cpuhw->limited_counter[i];
1194 event->hw.idx = cpuhw->limited_hwidx[i];
1195 val = (event->hw.idx == 5) ? pmc5 : pmc6;
1196 prev = local64_read(&event->hw.prev_count);
1197 if (check_and_compute_delta(prev, val))
1198 local64_set(&event->hw.prev_count, val);
1199 perf_event_update_userpage(event);
1204 * Since limited events don't respect the freeze conditions, we
1205 * have to read them immediately after freezing or unfreezing the
1206 * other events. We try to keep the values from the limited
1207 * events as consistent as possible by keeping the delay (in
1208 * cycles and instructions) between freezing/unfreezing and reading
1209 * the limited events as small and consistent as possible.
1210 * Therefore, if any limited events are in use, we read them
1211 * both, and always in the same order, to minimize variability,
1212 * and do it inside the same asm that writes MMCR0.
1214 static void write_mmcr0(struct cpu_hw_events *cpuhw, unsigned long mmcr0)
1216 unsigned long pmc5, pmc6;
1218 if (!cpuhw->n_limited) {
1219 mtspr(SPRN_MMCR0, mmcr0);
1224 * Write MMCR0, then read PMC5 and PMC6 immediately.
1225 * To ensure we don't get a performance monitor interrupt
1226 * between writing MMCR0 and freezing/thawing the limited
1227 * events, we first write MMCR0 with the event overflow
1228 * interrupt enable bits turned off.
1230 asm volatile("mtspr %3,%2; mfspr %0,%4; mfspr %1,%5"
1231 : "=&r" (pmc5), "=&r" (pmc6)
1232 : "r" (mmcr0 & ~(MMCR0_PMC1CE | MMCR0_PMCjCE)),
1234 "i" (SPRN_PMC5), "i" (SPRN_PMC6));
1236 if (mmcr0 & MMCR0_FC)
1237 freeze_limited_counters(cpuhw, pmc5, pmc6);
1239 thaw_limited_counters(cpuhw, pmc5, pmc6);
1242 * Write the full MMCR0 including the event overflow interrupt
1243 * enable bits, if necessary.
1245 if (mmcr0 & (MMCR0_PMC1CE | MMCR0_PMCjCE))
1246 mtspr(SPRN_MMCR0, mmcr0);
1250 * Disable all events to prevent PMU interrupts and to allow
1251 * events to be added or removed.
1253 static void power_pmu_disable(struct pmu *pmu)
1255 struct cpu_hw_events *cpuhw;
1256 unsigned long flags, mmcr0, val, mmcra;
1260 local_irq_save(flags);
1261 cpuhw = this_cpu_ptr(&cpu_hw_events);
1263 if (!cpuhw->disabled) {
1265 * Check if we ever enabled the PMU on this cpu.
1267 if (!cpuhw->pmcs_enabled) {
1269 cpuhw->pmcs_enabled = 1;
1273 * Set the 'freeze counters' bit, clear EBE/BHRBA/PMCC/PMAO/FC56
1275 val = mmcr0 = mfspr(SPRN_MMCR0);
1277 val &= ~(MMCR0_EBE | MMCR0_BHRBA | MMCR0_PMCC | MMCR0_PMAO |
1279 /* Set mmcr0 PMCCEXT for p10 */
1280 if (ppmu->flags & PPMU_ARCH_31)
1281 val |= MMCR0_PMCCEXT;
1284 * The barrier is to make sure the mtspr has been
1285 * executed and the PMU has frozen the events etc.
1288 write_mmcr0(cpuhw, val);
1292 val = mmcra = cpuhw->mmcr.mmcra;
1295 * Disable instruction sampling if it was enabled
1297 if (cpuhw->mmcr.mmcra & MMCRA_SAMPLE_ENABLE)
1298 val &= ~MMCRA_SAMPLE_ENABLE;
1300 /* Disable BHRB via mmcra (BHRBRD) for p10 */
1301 if (ppmu->flags & PPMU_ARCH_31)
1302 val |= MMCRA_BHRB_DISABLE;
1305 * Write SPRN_MMCRA if mmcra has either disabled
1306 * instruction sampling or BHRB.
1309 mtspr(SPRN_MMCRA, mmcra);
1314 cpuhw->disabled = 1;
1317 ebb_switch_out(mmcr0);
1321 * These are readable by userspace, may contain kernel
1322 * addresses and are not switched by context switch, so clear
1323 * them now to avoid leaking anything to userspace in general
1324 * including to another process.
1326 if (ppmu->flags & PPMU_ARCH_207S) {
1327 mtspr(SPRN_SDAR, 0);
1328 mtspr(SPRN_SIAR, 0);
1333 local_irq_restore(flags);
1337 * Re-enable all events if disable == 0.
1338 * If we were previously disabled and events were added, then
1339 * put the new config on the PMU.
1341 static void power_pmu_enable(struct pmu *pmu)
1343 struct perf_event *event;
1344 struct cpu_hw_events *cpuhw;
1345 unsigned long flags;
1347 unsigned long val, mmcr0;
1349 unsigned int hwc_index[MAX_HWEVENTS];
1356 local_irq_save(flags);
1358 cpuhw = this_cpu_ptr(&cpu_hw_events);
1359 if (!cpuhw->disabled)
1362 if (cpuhw->n_events == 0) {
1363 ppc_set_pmu_inuse(0);
1367 cpuhw->disabled = 0;
1370 * EBB requires an exclusive group and all events must have the EBB
1371 * flag set, or not set, so we can just check a single event. Also we
1372 * know we have at least one event.
1374 ebb = is_ebb_event(cpuhw->event[0]);
1377 * If we didn't change anything, or only removed events,
1378 * no need to recalculate MMCR* settings and reset the PMCs.
1379 * Just reenable the PMU with the current MMCR* settings
1380 * (possibly updated for removal of events).
1382 if (!cpuhw->n_added) {
1383 mtspr(SPRN_MMCRA, cpuhw->mmcr.mmcra & ~MMCRA_SAMPLE_ENABLE);
1384 mtspr(SPRN_MMCR1, cpuhw->mmcr.mmcr1);
1385 if (ppmu->flags & PPMU_ARCH_31)
1386 mtspr(SPRN_MMCR3, cpuhw->mmcr.mmcr3);
1391 * Clear all MMCR settings and recompute them for the new set of events.
1393 memset(&cpuhw->mmcr, 0, sizeof(cpuhw->mmcr));
1395 if (ppmu->compute_mmcr(cpuhw->events, cpuhw->n_events, hwc_index,
1396 &cpuhw->mmcr, cpuhw->event, ppmu->flags)) {
1397 /* shouldn't ever get here */
1398 printk(KERN_ERR "oops compute_mmcr failed\n");
1402 if (!(ppmu->flags & PPMU_ARCH_207S)) {
1404 * Add in MMCR0 freeze bits corresponding to the attr.exclude_*
1405 * bits for the first event. We have already checked that all
1406 * events have the same value for these bits as the first event.
1408 event = cpuhw->event[0];
1409 if (event->attr.exclude_user)
1410 cpuhw->mmcr.mmcr0 |= MMCR0_FCP;
1411 if (event->attr.exclude_kernel)
1412 cpuhw->mmcr.mmcr0 |= freeze_events_kernel;
1413 if (event->attr.exclude_hv)
1414 cpuhw->mmcr.mmcr0 |= MMCR0_FCHV;
1418 * Write the new configuration to MMCR* with the freeze
1419 * bit set and set the hardware events to their initial values.
1420 * Then unfreeze the events.
1422 ppc_set_pmu_inuse(1);
1423 mtspr(SPRN_MMCRA, cpuhw->mmcr.mmcra & ~MMCRA_SAMPLE_ENABLE);
1424 mtspr(SPRN_MMCR1, cpuhw->mmcr.mmcr1);
1425 mtspr(SPRN_MMCR0, (cpuhw->mmcr.mmcr0 & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
1427 if (ppmu->flags & PPMU_ARCH_207S)
1428 mtspr(SPRN_MMCR2, cpuhw->mmcr.mmcr2);
1430 if (ppmu->flags & PPMU_ARCH_31)
1431 mtspr(SPRN_MMCR3, cpuhw->mmcr.mmcr3);
1434 * Read off any pre-existing events that need to move
1437 for (i = 0; i < cpuhw->n_events; ++i) {
1438 event = cpuhw->event[i];
1439 if (event->hw.idx && event->hw.idx != hwc_index[i] + 1) {
1440 power_pmu_read(event);
1441 write_pmc(event->hw.idx, 0);
1447 * Initialize the PMCs for all the new and moved events.
1449 cpuhw->n_limited = n_lim = 0;
1450 for (i = 0; i < cpuhw->n_events; ++i) {
1451 event = cpuhw->event[i];
1454 idx = hwc_index[i] + 1;
1455 if (is_limited_pmc(idx)) {
1456 cpuhw->limited_counter[n_lim] = event;
1457 cpuhw->limited_hwidx[n_lim] = idx;
1463 val = local64_read(&event->hw.prev_count);
1466 if (event->hw.sample_period) {
1467 left = local64_read(&event->hw.period_left);
1468 if (left < 0x80000000L)
1469 val = 0x80000000L - left;
1471 local64_set(&event->hw.prev_count, val);
1474 event->hw.idx = idx;
1475 if (event->hw.state & PERF_HES_STOPPED)
1477 write_pmc(idx, val);
1479 perf_event_update_userpage(event);
1481 cpuhw->n_limited = n_lim;
1482 cpuhw->mmcr.mmcr0 |= MMCR0_PMXE | MMCR0_FCECE;
1485 pmao_restore_workaround(ebb);
1487 mmcr0 = ebb_switch_in(ebb, cpuhw);
1490 if (cpuhw->bhrb_users)
1491 ppmu->config_bhrb(cpuhw->bhrb_filter);
1493 write_mmcr0(cpuhw, mmcr0);
1496 * Enable instruction sampling if necessary
1498 if (cpuhw->mmcr.mmcra & MMCRA_SAMPLE_ENABLE) {
1500 mtspr(SPRN_MMCRA, cpuhw->mmcr.mmcra);
1505 local_irq_restore(flags);
1508 static int collect_events(struct perf_event *group, int max_count,
1509 struct perf_event *ctrs[], u64 *events,
1510 unsigned int *flags)
1513 struct perf_event *event;
1515 if (group->pmu->task_ctx_nr == perf_hw_context) {
1519 flags[n] = group->hw.event_base;
1520 events[n++] = group->hw.config;
1522 for_each_sibling_event(event, group) {
1523 if (event->pmu->task_ctx_nr == perf_hw_context &&
1524 event->state != PERF_EVENT_STATE_OFF) {
1528 flags[n] = event->hw.event_base;
1529 events[n++] = event->hw.config;
1536 * Add an event to the PMU.
1537 * If all events are not already frozen, then we disable and
1538 * re-enable the PMU in order to get hw_perf_enable to do the
1539 * actual work of reconfiguring the PMU.
1541 static int power_pmu_add(struct perf_event *event, int ef_flags)
1543 struct cpu_hw_events *cpuhw;
1544 unsigned long flags;
1548 local_irq_save(flags);
1549 perf_pmu_disable(event->pmu);
1552 * Add the event to the list (if there is room)
1553 * and check whether the total set is still feasible.
1555 cpuhw = this_cpu_ptr(&cpu_hw_events);
1556 n0 = cpuhw->n_events;
1557 if (n0 >= ppmu->n_counter)
1559 cpuhw->event[n0] = event;
1560 cpuhw->events[n0] = event->hw.config;
1561 cpuhw->flags[n0] = event->hw.event_base;
1564 * This event may have been disabled/stopped in record_and_restart()
1565 * because we exceeded the ->event_limit. If re-starting the event,
1566 * clear the ->hw.state (STOPPED and UPTODATE flags), so the user
1567 * notification is re-enabled.
1569 if (!(ef_flags & PERF_EF_START))
1570 event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
1572 event->hw.state = 0;
1575 * If group events scheduling transaction was started,
1576 * skip the schedulability test here, it will be performed
1577 * at commit time(->commit_txn) as a whole
1579 if (cpuhw->txn_flags & PERF_PMU_TXN_ADD)
1582 if (check_excludes(cpuhw->event, cpuhw->flags, n0, 1))
1584 if (power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n0 + 1, cpuhw->event))
1586 event->hw.config = cpuhw->events[n0];
1589 ebb_event_add(event);
1596 if (has_branch_stack(event)) {
1597 u64 bhrb_filter = -1;
1599 if (ppmu->bhrb_filter_map)
1600 bhrb_filter = ppmu->bhrb_filter_map(
1601 event->attr.branch_sample_type);
1603 if (bhrb_filter != -1) {
1604 cpuhw->bhrb_filter = bhrb_filter;
1605 power_pmu_bhrb_enable(event);
1609 perf_pmu_enable(event->pmu);
1610 local_irq_restore(flags);
1615 * Remove an event from the PMU.
1617 static void power_pmu_del(struct perf_event *event, int ef_flags)
1619 struct cpu_hw_events *cpuhw;
1621 unsigned long flags;
1623 local_irq_save(flags);
1624 perf_pmu_disable(event->pmu);
1626 power_pmu_read(event);
1628 cpuhw = this_cpu_ptr(&cpu_hw_events);
1629 for (i = 0; i < cpuhw->n_events; ++i) {
1630 if (event == cpuhw->event[i]) {
1631 while (++i < cpuhw->n_events) {
1632 cpuhw->event[i-1] = cpuhw->event[i];
1633 cpuhw->events[i-1] = cpuhw->events[i];
1634 cpuhw->flags[i-1] = cpuhw->flags[i];
1637 ppmu->disable_pmc(event->hw.idx - 1, &cpuhw->mmcr);
1638 if (event->hw.idx) {
1639 write_pmc(event->hw.idx, 0);
1642 perf_event_update_userpage(event);
1646 for (i = 0; i < cpuhw->n_limited; ++i)
1647 if (event == cpuhw->limited_counter[i])
1649 if (i < cpuhw->n_limited) {
1650 while (++i < cpuhw->n_limited) {
1651 cpuhw->limited_counter[i-1] = cpuhw->limited_counter[i];
1652 cpuhw->limited_hwidx[i-1] = cpuhw->limited_hwidx[i];
1656 if (cpuhw->n_events == 0) {
1657 /* disable exceptions if no events are running */
1658 cpuhw->mmcr.mmcr0 &= ~(MMCR0_PMXE | MMCR0_FCECE);
1661 if (has_branch_stack(event))
1662 power_pmu_bhrb_disable(event);
1664 perf_pmu_enable(event->pmu);
1665 local_irq_restore(flags);
1669 * POWER-PMU does not support disabling individual counters, hence
1670 * program their cycle counter to their max value and ignore the interrupts.
1673 static void power_pmu_start(struct perf_event *event, int ef_flags)
1675 unsigned long flags;
1679 if (!event->hw.idx || !event->hw.sample_period)
1682 if (!(event->hw.state & PERF_HES_STOPPED))
1685 if (ef_flags & PERF_EF_RELOAD)
1686 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1688 local_irq_save(flags);
1689 perf_pmu_disable(event->pmu);
1691 event->hw.state = 0;
1692 left = local64_read(&event->hw.period_left);
1695 if (left < 0x80000000L)
1696 val = 0x80000000L - left;
1698 write_pmc(event->hw.idx, val);
1700 perf_event_update_userpage(event);
1701 perf_pmu_enable(event->pmu);
1702 local_irq_restore(flags);
1705 static void power_pmu_stop(struct perf_event *event, int ef_flags)
1707 unsigned long flags;
1709 if (!event->hw.idx || !event->hw.sample_period)
1712 if (event->hw.state & PERF_HES_STOPPED)
1715 local_irq_save(flags);
1716 perf_pmu_disable(event->pmu);
1718 power_pmu_read(event);
1719 event->hw.state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1720 write_pmc(event->hw.idx, 0);
1722 perf_event_update_userpage(event);
1723 perf_pmu_enable(event->pmu);
1724 local_irq_restore(flags);
1728 * Start group events scheduling transaction
1729 * Set the flag to make pmu::enable() not perform the
1730 * schedulability test, it will be performed at commit time
1732 * We only support PERF_PMU_TXN_ADD transactions. Save the
1733 * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
1736 static void power_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1738 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1740 WARN_ON_ONCE(cpuhw->txn_flags); /* txn already in flight */
1742 cpuhw->txn_flags = txn_flags;
1743 if (txn_flags & ~PERF_PMU_TXN_ADD)
1746 perf_pmu_disable(pmu);
1747 cpuhw->n_txn_start = cpuhw->n_events;
1751 * Stop group events scheduling transaction
1752 * Clear the flag and pmu::enable() will perform the
1753 * schedulability test.
1755 static void power_pmu_cancel_txn(struct pmu *pmu)
1757 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1758 unsigned int txn_flags;
1760 WARN_ON_ONCE(!cpuhw->txn_flags); /* no txn in flight */
1762 txn_flags = cpuhw->txn_flags;
1763 cpuhw->txn_flags = 0;
1764 if (txn_flags & ~PERF_PMU_TXN_ADD)
1767 perf_pmu_enable(pmu);
1771 * Commit group events scheduling transaction
1772 * Perform the group schedulability test as a whole
1773 * Return 0 if success
1775 static int power_pmu_commit_txn(struct pmu *pmu)
1777 struct cpu_hw_events *cpuhw;
1783 cpuhw = this_cpu_ptr(&cpu_hw_events);
1784 WARN_ON_ONCE(!cpuhw->txn_flags); /* no txn in flight */
1786 if (cpuhw->txn_flags & ~PERF_PMU_TXN_ADD) {
1787 cpuhw->txn_flags = 0;
1791 n = cpuhw->n_events;
1792 if (check_excludes(cpuhw->event, cpuhw->flags, 0, n))
1794 i = power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n, cpuhw->event);
1798 for (i = cpuhw->n_txn_start; i < n; ++i)
1799 cpuhw->event[i]->hw.config = cpuhw->events[i];
1801 cpuhw->txn_flags = 0;
1802 perf_pmu_enable(pmu);
1807 * Return 1 if we might be able to put event on a limited PMC,
1809 * An event can only go on a limited PMC if it counts something
1810 * that a limited PMC can count, doesn't require interrupts, and
1811 * doesn't exclude any processor mode.
1813 static int can_go_on_limited_pmc(struct perf_event *event, u64 ev,
1817 u64 alt[MAX_EVENT_ALTERNATIVES];
1819 if (event->attr.exclude_user
1820 || event->attr.exclude_kernel
1821 || event->attr.exclude_hv
1822 || event->attr.sample_period)
1825 if (ppmu->limited_pmc_event(ev))
1829 * The requested event_id isn't on a limited PMC already;
1830 * see if any alternative code goes on a limited PMC.
1832 if (!ppmu->get_alternatives)
1835 flags |= PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD;
1836 n = ppmu->get_alternatives(ev, flags, alt);
1842 * Find an alternative event_id that goes on a normal PMC, if possible,
1843 * and return the event_id code, or 0 if there is no such alternative.
1844 * (Note: event_id code 0 is "don't count" on all machines.)
1846 static u64 normal_pmc_alternative(u64 ev, unsigned long flags)
1848 u64 alt[MAX_EVENT_ALTERNATIVES];
1851 flags &= ~(PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD);
1852 n = ppmu->get_alternatives(ev, flags, alt);
1858 /* Number of perf_events counting hardware events */
1859 static atomic_t num_events;
1860 /* Used to avoid races in calling reserve/release_pmc_hardware */
1861 static DEFINE_MUTEX(pmc_reserve_mutex);
1864 * Release the PMU if this is the last perf_event.
1866 static void hw_perf_event_destroy(struct perf_event *event)
1868 if (!atomic_add_unless(&num_events, -1, 1)) {
1869 mutex_lock(&pmc_reserve_mutex);
1870 if (atomic_dec_return(&num_events) == 0)
1871 release_pmc_hardware();
1872 mutex_unlock(&pmc_reserve_mutex);
1877 * Translate a generic cache event_id config to a raw event_id code.
1879 static int hw_perf_cache_event(u64 config, u64 *eventp)
1881 unsigned long type, op, result;
1884 if (!ppmu->cache_events)
1888 type = config & 0xff;
1889 op = (config >> 8) & 0xff;
1890 result = (config >> 16) & 0xff;
1892 if (type >= PERF_COUNT_HW_CACHE_MAX ||
1893 op >= PERF_COUNT_HW_CACHE_OP_MAX ||
1894 result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1897 ev = (*ppmu->cache_events)[type][op][result];
1906 static bool is_event_blacklisted(u64 ev)
1910 for (i=0; i < ppmu->n_blacklist_ev; i++) {
1911 if (ppmu->blacklist_ev[i] == ev)
1918 static int power_pmu_event_init(struct perf_event *event)
1921 unsigned long flags, irq_flags;
1922 struct perf_event *ctrs[MAX_HWEVENTS];
1923 u64 events[MAX_HWEVENTS];
1924 unsigned int cflags[MAX_HWEVENTS];
1927 struct cpu_hw_events *cpuhw;
1932 if (has_branch_stack(event)) {
1933 /* PMU has BHRB enabled */
1934 if (!(ppmu->flags & PPMU_ARCH_207S))
1938 switch (event->attr.type) {
1939 case PERF_TYPE_HARDWARE:
1940 ev = event->attr.config;
1941 if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
1944 if (ppmu->blacklist_ev && is_event_blacklisted(ev))
1946 ev = ppmu->generic_events[ev];
1948 case PERF_TYPE_HW_CACHE:
1949 err = hw_perf_cache_event(event->attr.config, &ev);
1953 if (ppmu->blacklist_ev && is_event_blacklisted(ev))
1957 ev = event->attr.config;
1959 if (ppmu->blacklist_ev && is_event_blacklisted(ev))
1966 event->hw.config_base = ev;
1970 * If we are not running on a hypervisor, force the
1971 * exclude_hv bit to 0 so that we don't care what
1972 * the user set it to.
1974 if (!firmware_has_feature(FW_FEATURE_LPAR))
1975 event->attr.exclude_hv = 0;
1978 * If this is a per-task event, then we can use
1979 * PM_RUN_* events interchangeably with their non RUN_*
1980 * equivalents, e.g. PM_RUN_CYC instead of PM_CYC.
1981 * XXX we should check if the task is an idle task.
1984 if (event->attach_state & PERF_ATTACH_TASK)
1985 flags |= PPMU_ONLY_COUNT_RUN;
1988 * If this machine has limited events, check whether this
1989 * event_id could go on a limited event.
1991 if (ppmu->flags & PPMU_LIMITED_PMC5_6) {
1992 if (can_go_on_limited_pmc(event, ev, flags)) {
1993 flags |= PPMU_LIMITED_PMC_OK;
1994 } else if (ppmu->limited_pmc_event(ev)) {
1996 * The requested event_id is on a limited PMC,
1997 * but we can't use a limited PMC; see if any
1998 * alternative goes on a normal PMC.
2000 ev = normal_pmc_alternative(ev, flags);
2006 /* Extra checks for EBB */
2007 err = ebb_event_check(event);
2012 * If this is in a group, check if it can go on with all the
2013 * other hardware events in the group. We assume the event
2014 * hasn't been linked into its leader's sibling list at this point.
2017 if (event->group_leader != event) {
2018 n = collect_events(event->group_leader, ppmu->n_counter - 1,
2019 ctrs, events, cflags);
2026 if (check_excludes(ctrs, cflags, n, 1))
2029 local_irq_save(irq_flags);
2030 cpuhw = this_cpu_ptr(&cpu_hw_events);
2032 err = power_check_constraints(cpuhw, events, cflags, n + 1, ctrs);
2034 if (has_branch_stack(event)) {
2035 u64 bhrb_filter = -1;
2037 if (ppmu->bhrb_filter_map)
2038 bhrb_filter = ppmu->bhrb_filter_map(
2039 event->attr.branch_sample_type);
2041 if (bhrb_filter == -1) {
2042 local_irq_restore(irq_flags);
2045 cpuhw->bhrb_filter = bhrb_filter;
2048 local_irq_restore(irq_flags);
2052 event->hw.config = events[n];
2053 event->hw.event_base = cflags[n];
2054 event->hw.last_period = event->hw.sample_period;
2055 local64_set(&event->hw.period_left, event->hw.last_period);
2058 * For EBB events we just context switch the PMC value, we don't do any
2059 * of the sample_period logic. We use hw.prev_count for this.
2061 if (is_ebb_event(event))
2062 local64_set(&event->hw.prev_count, 0);
2065 * See if we need to reserve the PMU.
2066 * If no events are currently in use, then we have to take a
2067 * mutex to ensure that we don't race with another task doing
2068 * reserve_pmc_hardware or release_pmc_hardware.
2071 if (!atomic_inc_not_zero(&num_events)) {
2072 mutex_lock(&pmc_reserve_mutex);
2073 if (atomic_read(&num_events) == 0 &&
2074 reserve_pmc_hardware(perf_event_interrupt))
2077 atomic_inc(&num_events);
2078 mutex_unlock(&pmc_reserve_mutex);
2080 event->destroy = hw_perf_event_destroy;
2085 static int power_pmu_event_idx(struct perf_event *event)
2087 return event->hw.idx;
2090 ssize_t power_events_sysfs_show(struct device *dev,
2091 struct device_attribute *attr, char *page)
2093 struct perf_pmu_events_attr *pmu_attr;
2095 pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
2097 return sprintf(page, "event=0x%02llx\n", pmu_attr->id);
2100 static struct pmu power_pmu = {
2101 .pmu_enable = power_pmu_enable,
2102 .pmu_disable = power_pmu_disable,
2103 .event_init = power_pmu_event_init,
2104 .add = power_pmu_add,
2105 .del = power_pmu_del,
2106 .start = power_pmu_start,
2107 .stop = power_pmu_stop,
2108 .read = power_pmu_read,
2109 .start_txn = power_pmu_start_txn,
2110 .cancel_txn = power_pmu_cancel_txn,
2111 .commit_txn = power_pmu_commit_txn,
2112 .event_idx = power_pmu_event_idx,
2113 .sched_task = power_pmu_sched_task,
2116 #define PERF_SAMPLE_ADDR_TYPE (PERF_SAMPLE_ADDR | \
2117 PERF_SAMPLE_PHYS_ADDR | \
2118 PERF_SAMPLE_DATA_PAGE_SIZE)
2120 * A counter has overflowed; update its count and record
2121 * things if requested. Note that interrupts are hard-disabled
2122 * here so there is no possibility of being interrupted.
2124 static void record_and_restart(struct perf_event *event, unsigned long val,
2125 struct pt_regs *regs)
2127 u64 period = event->hw.sample_period;
2128 s64 prev, delta, left;
2131 if (event->hw.state & PERF_HES_STOPPED) {
2132 write_pmc(event->hw.idx, 0);
2136 /* we don't have to worry about interrupts here */
2137 prev = local64_read(&event->hw.prev_count);
2138 delta = check_and_compute_delta(prev, val);
2139 local64_add(delta, &event->count);
2142 * See if the total period for this event has expired,
2143 * and update for the next period.
2146 left = local64_read(&event->hw.period_left) - delta;
2156 * If address is not requested in the sample via
2157 * PERF_SAMPLE_IP, just record that sample irrespective
2158 * of SIAR valid check.
2160 if (event->attr.sample_type & PERF_SAMPLE_IP)
2161 record = siar_valid(regs);
2165 event->hw.last_period = event->hw.sample_period;
2167 if (left < 0x80000000LL)
2168 val = 0x80000000LL - left;
2171 write_pmc(event->hw.idx, val);
2172 local64_set(&event->hw.prev_count, val);
2173 local64_set(&event->hw.period_left, left);
2174 perf_event_update_userpage(event);
2177 * Due to hardware limitation, sometimes SIAR could sample a kernel
2178 * address even when freeze on supervisor state (kernel) is set in
2179 * MMCR2. Check attr.exclude_kernel and address to drop the sample in
2182 if (event->attr.exclude_kernel &&
2183 (event->attr.sample_type & PERF_SAMPLE_IP) &&
2184 is_kernel_addr(mfspr(SPRN_SIAR)))
2188 * Finally record data if requested.
2191 struct perf_sample_data data;
2193 perf_sample_data_init(&data, ~0ULL, event->hw.last_period);
2195 if (event->attr.sample_type & PERF_SAMPLE_ADDR_TYPE)
2196 perf_get_data_addr(event, regs, &data.addr);
2198 if (event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK) {
2199 struct cpu_hw_events *cpuhw;
2200 cpuhw = this_cpu_ptr(&cpu_hw_events);
2201 power_pmu_bhrb_read(event, cpuhw);
2202 data.br_stack = &cpuhw->bhrb_stack;
2205 if (event->attr.sample_type & PERF_SAMPLE_DATA_SRC &&
2206 ppmu->get_mem_data_src)
2207 ppmu->get_mem_data_src(&data.data_src, ppmu->flags, regs);
2209 if (event->attr.sample_type & PERF_SAMPLE_WEIGHT &&
2210 ppmu->get_mem_weight)
2211 ppmu->get_mem_weight(&data.weight.full);
2213 if (perf_event_overflow(event, &data, regs))
2214 power_pmu_stop(event, 0);
2215 } else if (period) {
2216 /* Account for interrupt in case of invalid SIAR */
2217 if (perf_event_account_interrupt(event))
2218 power_pmu_stop(event, 0);
2223 * Called from generic code to get the misc flags (i.e. processor mode)
2226 unsigned long perf_misc_flags(struct pt_regs *regs)
2228 u32 flags = perf_get_misc_flags(regs);
2232 return user_mode(regs) ? PERF_RECORD_MISC_USER :
2233 PERF_RECORD_MISC_KERNEL;
2237 * Called from generic code to get the instruction pointer
2240 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2242 bool use_siar = regs_use_siar(regs);
2243 unsigned long siar = mfspr(SPRN_SIAR);
2245 if (ppmu->flags & PPMU_P10_DD1) {
2250 } else if (use_siar && siar_valid(regs))
2251 return mfspr(SPRN_SIAR) + perf_ip_adjust(regs);
2253 return 0; // no valid instruction pointer
2258 static bool pmc_overflow_power7(unsigned long val)
2261 * Events on POWER7 can roll back if a speculative event doesn't
2262 * eventually complete. Unfortunately in some rare cases they will
2263 * raise a performance monitor exception. We need to catch this to
2264 * ensure we reset the PMC. In all cases the PMC will be 256 or less
2265 * cycles from overflow.
2267 * We only do this if the first pass fails to find any overflowing
2268 * PMCs because a user might set a period of less than 256 and we
2269 * don't want to mistakenly reset them.
2271 if ((0x80000000 - val) <= 256)
2277 static bool pmc_overflow(unsigned long val)
2286 * Performance monitor interrupt stuff
2288 static void __perf_event_interrupt(struct pt_regs *regs)
2291 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
2292 struct perf_event *event;
2295 if (cpuhw->n_limited)
2296 freeze_limited_counters(cpuhw, mfspr(SPRN_PMC5),
2299 perf_read_regs(regs);
2301 /* Read all the PMCs since we'll need them a bunch of times */
2302 for (i = 0; i < ppmu->n_counter; ++i)
2303 cpuhw->pmcs[i] = read_pmc(i + 1);
2305 /* Try to find what caused the IRQ */
2307 for (i = 0; i < ppmu->n_counter; ++i) {
2308 if (!pmc_overflow(cpuhw->pmcs[i]))
2310 if (is_limited_pmc(i + 1))
2311 continue; /* these won't generate IRQs */
2313 * We've found one that's overflowed. For active
2314 * counters we need to log this. For inactive
2315 * counters, we need to reset it anyway
2319 for (j = 0; j < cpuhw->n_events; ++j) {
2320 event = cpuhw->event[j];
2321 if (event->hw.idx == (i + 1)) {
2323 record_and_restart(event, cpuhw->pmcs[i], regs);
2328 /* reset non active counters that have overflowed */
2329 write_pmc(i + 1, 0);
2331 if (!found && pvr_version_is(PVR_POWER7)) {
2332 /* check active counters for special buggy p7 overflow */
2333 for (i = 0; i < cpuhw->n_events; ++i) {
2334 event = cpuhw->event[i];
2335 if (!event->hw.idx || is_limited_pmc(event->hw.idx))
2337 if (pmc_overflow_power7(cpuhw->pmcs[event->hw.idx - 1])) {
2338 /* event has overflowed in a buggy way*/
2340 record_and_restart(event,
2341 cpuhw->pmcs[event->hw.idx - 1],
2346 if (unlikely(!found) && !arch_irq_disabled_regs(regs))
2347 printk_ratelimited(KERN_WARNING "Can't find PMC that caused IRQ\n");
2350 * Reset MMCR0 to its normal value. This will set PMXE and
2351 * clear FC (freeze counters) and PMAO (perf mon alert occurred)
2352 * and thus allow interrupts to occur again.
2353 * XXX might want to use MSR.PM to keep the events frozen until
2354 * we get back out of this interrupt.
2356 write_mmcr0(cpuhw, cpuhw->mmcr.mmcr0);
2358 /* Clear the cpuhw->pmcs */
2359 memset(&cpuhw->pmcs, 0, sizeof(cpuhw->pmcs));
2363 static void perf_event_interrupt(struct pt_regs *regs)
2365 u64 start_clock = sched_clock();
2367 __perf_event_interrupt(regs);
2368 perf_sample_event_took(sched_clock() - start_clock);
2371 static int power_pmu_prepare_cpu(unsigned int cpu)
2373 struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
2376 memset(cpuhw, 0, sizeof(*cpuhw));
2377 cpuhw->mmcr.mmcr0 = MMCR0_FC;
2382 int register_power_pmu(struct power_pmu *pmu)
2385 return -EBUSY; /* something's already registered */
2388 pr_info("%s performance monitor hardware support registered\n",
2391 power_pmu.attr_groups = ppmu->attr_groups;
2392 power_pmu.capabilities |= (ppmu->capabilities & PERF_PMU_CAP_EXTENDED_REGS);
2396 * Use FCHV to ignore kernel events if MSR.HV is set.
2398 if (mfmsr() & MSR_HV)
2399 freeze_events_kernel = MMCR0_FCHV;
2400 #endif /* CONFIG_PPC64 */
2402 perf_pmu_register(&power_pmu, "cpu", PERF_TYPE_RAW);
2403 cpuhp_setup_state(CPUHP_PERF_POWER, "perf/powerpc:prepare",
2404 power_pmu_prepare_cpu, NULL);
2409 static int __init init_ppc64_pmu(void)
2411 /* run through all the pmu drivers one at a time */
2412 if (!init_power5_pmu())
2414 else if (!init_power5p_pmu())
2416 else if (!init_power6_pmu())
2418 else if (!init_power7_pmu())
2420 else if (!init_power8_pmu())
2422 else if (!init_power9_pmu())
2424 else if (!init_power10_pmu())
2426 else if (!init_ppc970_pmu())
2429 return init_generic_compat_pmu();
2431 early_initcall(init_ppc64_pmu);