2 * Performance events x86 architecture code
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2009 Jaswinder Singh Rajput
7 * Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
8 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
9 * Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com>
10 * Copyright (C) 2009 Google, Inc., Stephane Eranian
12 * For licencing details see kernel-base/COPYING
15 #include <linux/perf_event.h>
16 #include <linux/capability.h>
17 #include <linux/notifier.h>
18 #include <linux/hardirq.h>
19 #include <linux/kprobes.h>
20 #include <linux/module.h>
21 #include <linux/kdebug.h>
22 #include <linux/sched.h>
23 #include <linux/uaccess.h>
24 #include <linux/highmem.h>
25 #include <linux/cpu.h>
26 #include <linux/bitops.h>
29 #include <asm/stacktrace.h>
32 static u64 perf_event_mask __read_mostly;
34 /* The maximal number of PEBS events: */
35 #define MAX_PEBS_EVENTS 4
37 /* The size of a BTS record in bytes: */
38 #define BTS_RECORD_SIZE 24
40 /* The size of a per-cpu BTS buffer in bytes: */
41 #define BTS_BUFFER_SIZE (BTS_RECORD_SIZE * 2048)
43 /* The BTS overflow threshold in bytes from the end of the buffer: */
44 #define BTS_OVFL_TH (BTS_RECORD_SIZE * 128)
48 * Bits in the debugctlmsr controlling branch tracing.
50 #define X86_DEBUGCTL_TR (1 << 6)
51 #define X86_DEBUGCTL_BTS (1 << 7)
52 #define X86_DEBUGCTL_BTINT (1 << 8)
53 #define X86_DEBUGCTL_BTS_OFF_OS (1 << 9)
54 #define X86_DEBUGCTL_BTS_OFF_USR (1 << 10)
57 * A debug store configuration.
59 * We only support architectures that use 64bit fields.
64 u64 bts_absolute_maximum;
65 u64 bts_interrupt_threshold;
68 u64 pebs_absolute_maximum;
69 u64 pebs_interrupt_threshold;
70 u64 pebs_event_reset[MAX_PEBS_EVENTS];
73 struct event_constraint {
75 unsigned long idxmsk[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
83 struct cpu_hw_events {
84 struct perf_event *events[X86_PMC_IDX_MAX]; /* in counter order */
85 unsigned long active_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
86 unsigned long interrupts;
88 struct debug_store *ds;
92 int assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
93 struct perf_event *event_list[X86_PMC_IDX_MAX]; /* in enabled order */
96 #define EVENT_CONSTRAINT(c, n, m) { \
97 { .idxmsk64[0] = (n) }, \
100 .weight = HWEIGHT64((u64)(n)), \
103 #define INTEL_EVENT_CONSTRAINT(c, n) EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK)
104 #define FIXED_EVENT_CONSTRAINT(c, n) EVENT_CONSTRAINT(c, n, INTEL_ARCH_FIXED_MASK)
106 #define EVENT_CONSTRAINT_END EVENT_CONSTRAINT(0, 0, 0)
108 #define for_each_event_constraint(e, c) for ((e) = (c); (e)->cmask; (e)++)
111 * struct x86_pmu - generic x86 pmu
116 int (*handle_irq)(struct pt_regs *);
117 void (*disable_all)(void);
118 void (*enable_all)(void);
119 void (*enable)(struct hw_perf_event *, int);
120 void (*disable)(struct hw_perf_event *, int);
123 u64 (*event_map)(int);
124 u64 (*raw_event)(u64);
127 int num_events_fixed;
133 void (*enable_bts)(u64 config);
134 void (*disable_bts)(void);
136 struct event_constraint *
137 (*get_event_constraints)(struct cpu_hw_events *cpuc,
138 struct perf_event *event);
140 void (*put_event_constraints)(struct cpu_hw_events *cpuc,
141 struct perf_event *event);
142 struct event_constraint *event_constraints;
145 static struct x86_pmu x86_pmu __read_mostly;
147 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
151 static int x86_perf_event_set_period(struct perf_event *event,
152 struct hw_perf_event *hwc, int idx);
155 * Not sure about some of these
157 static const u64 p6_perfmon_event_map[] =
159 [PERF_COUNT_HW_CPU_CYCLES] = 0x0079,
160 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
161 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x0f2e,
162 [PERF_COUNT_HW_CACHE_MISSES] = 0x012e,
163 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4,
164 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5,
165 [PERF_COUNT_HW_BUS_CYCLES] = 0x0062,
168 static u64 p6_pmu_event_map(int hw_event)
170 return p6_perfmon_event_map[hw_event];
174 * Event setting that is specified not to count anything.
175 * We use this to effectively disable a counter.
177 * L2_RQSTS with 0 MESI unit mask.
179 #define P6_NOP_EVENT 0x0000002EULL
181 static u64 p6_pmu_raw_event(u64 hw_event)
183 #define P6_EVNTSEL_EVENT_MASK 0x000000FFULL
184 #define P6_EVNTSEL_UNIT_MASK 0x0000FF00ULL
185 #define P6_EVNTSEL_EDGE_MASK 0x00040000ULL
186 #define P6_EVNTSEL_INV_MASK 0x00800000ULL
187 #define P6_EVNTSEL_REG_MASK 0xFF000000ULL
189 #define P6_EVNTSEL_MASK \
190 (P6_EVNTSEL_EVENT_MASK | \
191 P6_EVNTSEL_UNIT_MASK | \
192 P6_EVNTSEL_EDGE_MASK | \
193 P6_EVNTSEL_INV_MASK | \
196 return hw_event & P6_EVNTSEL_MASK;
199 static struct event_constraint intel_p6_event_constraints[] =
201 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FLOPS */
202 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
203 INTEL_EVENT_CONSTRAINT(0x11, 0x1), /* FP_ASSIST */
204 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
205 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
206 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
211 * Intel PerfMon v3. Used on Core2 and later.
213 static const u64 intel_perfmon_event_map[] =
215 [PERF_COUNT_HW_CPU_CYCLES] = 0x003c,
216 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
217 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e,
218 [PERF_COUNT_HW_CACHE_MISSES] = 0x412e,
219 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4,
220 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5,
221 [PERF_COUNT_HW_BUS_CYCLES] = 0x013c,
224 static struct event_constraint intel_core_event_constraints[] =
226 FIXED_EVENT_CONSTRAINT(0xc0, (0x3|(1ULL<<32))), /* INSTRUCTIONS_RETIRED */
227 FIXED_EVENT_CONSTRAINT(0x3c, (0x3|(1ULL<<33))), /* UNHALTED_CORE_CYCLES */
228 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
229 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
230 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
231 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
232 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
233 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
234 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
235 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
236 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
240 static struct event_constraint intel_nehalem_event_constraints[] =
242 FIXED_EVENT_CONSTRAINT(0xc0, (0x3|(1ULL<<32))), /* INSTRUCTIONS_RETIRED */
243 FIXED_EVENT_CONSTRAINT(0x3c, (0x3|(1ULL<<33))), /* UNHALTED_CORE_CYCLES */
244 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
245 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
246 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
247 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
248 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
249 INTEL_EVENT_CONSTRAINT(0x4c, 0x3), /* LOAD_HIT_PRE */
250 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
251 INTEL_EVENT_CONSTRAINT(0x52, 0x3), /* L1D_CACHE_PREFETCH_LOCK_FB_HIT */
252 INTEL_EVENT_CONSTRAINT(0x53, 0x3), /* L1D_CACHE_LOCK_FB_HIT */
253 INTEL_EVENT_CONSTRAINT(0xc5, 0x3), /* CACHE_LOCK_CYCLES */
257 static struct event_constraint intel_gen_event_constraints[] =
259 FIXED_EVENT_CONSTRAINT(0xc0, (0x3|(1ULL<<32))), /* INSTRUCTIONS_RETIRED */
260 FIXED_EVENT_CONSTRAINT(0x3c, (0x3|(1ULL<<33))), /* UNHALTED_CORE_CYCLES */
264 static u64 intel_pmu_event_map(int hw_event)
266 return intel_perfmon_event_map[hw_event];
270 * Generalized hw caching related hw_event table, filled
271 * in on a per model basis. A value of 0 means
272 * 'not supported', -1 means 'hw_event makes no sense on
273 * this CPU', any other value means the raw hw_event
277 #define C(x) PERF_COUNT_HW_CACHE_##x
279 static u64 __read_mostly hw_cache_event_ids
280 [PERF_COUNT_HW_CACHE_MAX]
281 [PERF_COUNT_HW_CACHE_OP_MAX]
282 [PERF_COUNT_HW_CACHE_RESULT_MAX];
284 static __initconst u64 nehalem_hw_cache_event_ids
285 [PERF_COUNT_HW_CACHE_MAX]
286 [PERF_COUNT_HW_CACHE_OP_MAX]
287 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
291 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
292 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
295 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
296 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
298 [ C(OP_PREFETCH) ] = {
299 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
300 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
305 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
306 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
309 [ C(RESULT_ACCESS) ] = -1,
310 [ C(RESULT_MISS) ] = -1,
312 [ C(OP_PREFETCH) ] = {
313 [ C(RESULT_ACCESS) ] = 0x0,
314 [ C(RESULT_MISS) ] = 0x0,
319 [ C(RESULT_ACCESS) ] = 0x0324, /* L2_RQSTS.LOADS */
320 [ C(RESULT_MISS) ] = 0x0224, /* L2_RQSTS.LD_MISS */
323 [ C(RESULT_ACCESS) ] = 0x0c24, /* L2_RQSTS.RFOS */
324 [ C(RESULT_MISS) ] = 0x0824, /* L2_RQSTS.RFO_MISS */
326 [ C(OP_PREFETCH) ] = {
327 [ C(RESULT_ACCESS) ] = 0x4f2e, /* LLC Reference */
328 [ C(RESULT_MISS) ] = 0x412e, /* LLC Misses */
333 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
334 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
337 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
338 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
340 [ C(OP_PREFETCH) ] = {
341 [ C(RESULT_ACCESS) ] = 0x0,
342 [ C(RESULT_MISS) ] = 0x0,
347 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
348 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */
351 [ C(RESULT_ACCESS) ] = -1,
352 [ C(RESULT_MISS) ] = -1,
354 [ C(OP_PREFETCH) ] = {
355 [ C(RESULT_ACCESS) ] = -1,
356 [ C(RESULT_MISS) ] = -1,
361 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
362 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
365 [ C(RESULT_ACCESS) ] = -1,
366 [ C(RESULT_MISS) ] = -1,
368 [ C(OP_PREFETCH) ] = {
369 [ C(RESULT_ACCESS) ] = -1,
370 [ C(RESULT_MISS) ] = -1,
375 static __initconst u64 core2_hw_cache_event_ids
376 [PERF_COUNT_HW_CACHE_MAX]
377 [PERF_COUNT_HW_CACHE_OP_MAX]
378 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
382 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
383 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
386 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
387 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
389 [ C(OP_PREFETCH) ] = {
390 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */
391 [ C(RESULT_MISS) ] = 0,
396 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */
397 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */
400 [ C(RESULT_ACCESS) ] = -1,
401 [ C(RESULT_MISS) ] = -1,
403 [ C(OP_PREFETCH) ] = {
404 [ C(RESULT_ACCESS) ] = 0,
405 [ C(RESULT_MISS) ] = 0,
410 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
411 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
414 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
415 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
417 [ C(OP_PREFETCH) ] = {
418 [ C(RESULT_ACCESS) ] = 0,
419 [ C(RESULT_MISS) ] = 0,
424 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
425 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */
428 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
429 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */
431 [ C(OP_PREFETCH) ] = {
432 [ C(RESULT_ACCESS) ] = 0,
433 [ C(RESULT_MISS) ] = 0,
438 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
439 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */
442 [ C(RESULT_ACCESS) ] = -1,
443 [ C(RESULT_MISS) ] = -1,
445 [ C(OP_PREFETCH) ] = {
446 [ C(RESULT_ACCESS) ] = -1,
447 [ C(RESULT_MISS) ] = -1,
452 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
453 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
456 [ C(RESULT_ACCESS) ] = -1,
457 [ C(RESULT_MISS) ] = -1,
459 [ C(OP_PREFETCH) ] = {
460 [ C(RESULT_ACCESS) ] = -1,
461 [ C(RESULT_MISS) ] = -1,
466 static __initconst u64 atom_hw_cache_event_ids
467 [PERF_COUNT_HW_CACHE_MAX]
468 [PERF_COUNT_HW_CACHE_OP_MAX]
469 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
473 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */
474 [ C(RESULT_MISS) ] = 0,
477 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */
478 [ C(RESULT_MISS) ] = 0,
480 [ C(OP_PREFETCH) ] = {
481 [ C(RESULT_ACCESS) ] = 0x0,
482 [ C(RESULT_MISS) ] = 0,
487 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
488 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
491 [ C(RESULT_ACCESS) ] = -1,
492 [ C(RESULT_MISS) ] = -1,
494 [ C(OP_PREFETCH) ] = {
495 [ C(RESULT_ACCESS) ] = 0,
496 [ C(RESULT_MISS) ] = 0,
501 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
502 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
505 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
506 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
508 [ C(OP_PREFETCH) ] = {
509 [ C(RESULT_ACCESS) ] = 0,
510 [ C(RESULT_MISS) ] = 0,
515 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */
516 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */
519 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */
520 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */
522 [ C(OP_PREFETCH) ] = {
523 [ C(RESULT_ACCESS) ] = 0,
524 [ C(RESULT_MISS) ] = 0,
529 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
530 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */
533 [ C(RESULT_ACCESS) ] = -1,
534 [ C(RESULT_MISS) ] = -1,
536 [ C(OP_PREFETCH) ] = {
537 [ C(RESULT_ACCESS) ] = -1,
538 [ C(RESULT_MISS) ] = -1,
543 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
544 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
547 [ C(RESULT_ACCESS) ] = -1,
548 [ C(RESULT_MISS) ] = -1,
550 [ C(OP_PREFETCH) ] = {
551 [ C(RESULT_ACCESS) ] = -1,
552 [ C(RESULT_MISS) ] = -1,
557 static u64 intel_pmu_raw_event(u64 hw_event)
559 #define CORE_EVNTSEL_EVENT_MASK 0x000000FFULL
560 #define CORE_EVNTSEL_UNIT_MASK 0x0000FF00ULL
561 #define CORE_EVNTSEL_EDGE_MASK 0x00040000ULL
562 #define CORE_EVNTSEL_INV_MASK 0x00800000ULL
563 #define CORE_EVNTSEL_REG_MASK 0xFF000000ULL
565 #define CORE_EVNTSEL_MASK \
566 (INTEL_ARCH_EVTSEL_MASK | \
567 INTEL_ARCH_UNIT_MASK | \
568 INTEL_ARCH_EDGE_MASK | \
569 INTEL_ARCH_INV_MASK | \
572 return hw_event & CORE_EVNTSEL_MASK;
575 static __initconst u64 amd_hw_cache_event_ids
576 [PERF_COUNT_HW_CACHE_MAX]
577 [PERF_COUNT_HW_CACHE_OP_MAX]
578 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
582 [ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses */
583 [ C(RESULT_MISS) ] = 0x0041, /* Data Cache Misses */
586 [ C(RESULT_ACCESS) ] = 0x0142, /* Data Cache Refills :system */
587 [ C(RESULT_MISS) ] = 0,
589 [ C(OP_PREFETCH) ] = {
590 [ C(RESULT_ACCESS) ] = 0x0267, /* Data Prefetcher :attempts */
591 [ C(RESULT_MISS) ] = 0x0167, /* Data Prefetcher :cancelled */
596 [ C(RESULT_ACCESS) ] = 0x0080, /* Instruction cache fetches */
597 [ C(RESULT_MISS) ] = 0x0081, /* Instruction cache misses */
600 [ C(RESULT_ACCESS) ] = -1,
601 [ C(RESULT_MISS) ] = -1,
603 [ C(OP_PREFETCH) ] = {
604 [ C(RESULT_ACCESS) ] = 0x014B, /* Prefetch Instructions :Load */
605 [ C(RESULT_MISS) ] = 0,
610 [ C(RESULT_ACCESS) ] = 0x037D, /* Requests to L2 Cache :IC+DC */
611 [ C(RESULT_MISS) ] = 0x037E, /* L2 Cache Misses : IC+DC */
614 [ C(RESULT_ACCESS) ] = 0x017F, /* L2 Fill/Writeback */
615 [ C(RESULT_MISS) ] = 0,
617 [ C(OP_PREFETCH) ] = {
618 [ C(RESULT_ACCESS) ] = 0,
619 [ C(RESULT_MISS) ] = 0,
624 [ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses */
625 [ C(RESULT_MISS) ] = 0x0046, /* L1 DTLB and L2 DLTB Miss */
628 [ C(RESULT_ACCESS) ] = 0,
629 [ C(RESULT_MISS) ] = 0,
631 [ C(OP_PREFETCH) ] = {
632 [ C(RESULT_ACCESS) ] = 0,
633 [ C(RESULT_MISS) ] = 0,
638 [ C(RESULT_ACCESS) ] = 0x0080, /* Instruction fecthes */
639 [ C(RESULT_MISS) ] = 0x0085, /* Instr. fetch ITLB misses */
642 [ C(RESULT_ACCESS) ] = -1,
643 [ C(RESULT_MISS) ] = -1,
645 [ C(OP_PREFETCH) ] = {
646 [ C(RESULT_ACCESS) ] = -1,
647 [ C(RESULT_MISS) ] = -1,
652 [ C(RESULT_ACCESS) ] = 0x00c2, /* Retired Branch Instr. */
653 [ C(RESULT_MISS) ] = 0x00c3, /* Retired Mispredicted BI */
656 [ C(RESULT_ACCESS) ] = -1,
657 [ C(RESULT_MISS) ] = -1,
659 [ C(OP_PREFETCH) ] = {
660 [ C(RESULT_ACCESS) ] = -1,
661 [ C(RESULT_MISS) ] = -1,
667 * AMD Performance Monitor K7 and later.
669 static const u64 amd_perfmon_event_map[] =
671 [PERF_COUNT_HW_CPU_CYCLES] = 0x0076,
672 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
673 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x0080,
674 [PERF_COUNT_HW_CACHE_MISSES] = 0x0081,
675 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4,
676 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5,
679 static u64 amd_pmu_event_map(int hw_event)
681 return amd_perfmon_event_map[hw_event];
684 static u64 amd_pmu_raw_event(u64 hw_event)
686 #define K7_EVNTSEL_EVENT_MASK 0x7000000FFULL
687 #define K7_EVNTSEL_UNIT_MASK 0x00000FF00ULL
688 #define K7_EVNTSEL_EDGE_MASK 0x000040000ULL
689 #define K7_EVNTSEL_INV_MASK 0x000800000ULL
690 #define K7_EVNTSEL_REG_MASK 0x0FF000000ULL
692 #define K7_EVNTSEL_MASK \
693 (K7_EVNTSEL_EVENT_MASK | \
694 K7_EVNTSEL_UNIT_MASK | \
695 K7_EVNTSEL_EDGE_MASK | \
696 K7_EVNTSEL_INV_MASK | \
699 return hw_event & K7_EVNTSEL_MASK;
703 * Propagate event elapsed time into the generic event.
704 * Can only be executed on the CPU where the event is active.
705 * Returns the delta events processed.
708 x86_perf_event_update(struct perf_event *event,
709 struct hw_perf_event *hwc, int idx)
711 int shift = 64 - x86_pmu.event_bits;
712 u64 prev_raw_count, new_raw_count;
715 if (idx == X86_PMC_IDX_FIXED_BTS)
719 * Careful: an NMI might modify the previous event value.
721 * Our tactic to handle this is to first atomically read and
722 * exchange a new raw count - then add that new-prev delta
723 * count to the generic event atomically:
726 prev_raw_count = atomic64_read(&hwc->prev_count);
727 rdmsrl(hwc->event_base + idx, new_raw_count);
729 if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count,
730 new_raw_count) != prev_raw_count)
734 * Now we have the new raw value and have updated the prev
735 * timestamp already. We can now calculate the elapsed delta
736 * (event-)time and add that to the generic event.
738 * Careful, not all hw sign-extends above the physical width
741 delta = (new_raw_count << shift) - (prev_raw_count << shift);
744 atomic64_add(delta, &event->count);
745 atomic64_sub(delta, &hwc->period_left);
747 return new_raw_count;
750 static atomic_t active_events;
751 static DEFINE_MUTEX(pmc_reserve_mutex);
753 static bool reserve_pmc_hardware(void)
755 #ifdef CONFIG_X86_LOCAL_APIC
758 if (nmi_watchdog == NMI_LOCAL_APIC)
759 disable_lapic_nmi_watchdog();
761 for (i = 0; i < x86_pmu.num_events; i++) {
762 if (!reserve_perfctr_nmi(x86_pmu.perfctr + i))
766 for (i = 0; i < x86_pmu.num_events; i++) {
767 if (!reserve_evntsel_nmi(x86_pmu.eventsel + i))
774 #ifdef CONFIG_X86_LOCAL_APIC
776 for (i--; i >= 0; i--)
777 release_evntsel_nmi(x86_pmu.eventsel + i);
779 i = x86_pmu.num_events;
782 for (i--; i >= 0; i--)
783 release_perfctr_nmi(x86_pmu.perfctr + i);
785 if (nmi_watchdog == NMI_LOCAL_APIC)
786 enable_lapic_nmi_watchdog();
792 static void release_pmc_hardware(void)
794 #ifdef CONFIG_X86_LOCAL_APIC
797 for (i = 0; i < x86_pmu.num_events; i++) {
798 release_perfctr_nmi(x86_pmu.perfctr + i);
799 release_evntsel_nmi(x86_pmu.eventsel + i);
802 if (nmi_watchdog == NMI_LOCAL_APIC)
803 enable_lapic_nmi_watchdog();
807 static inline bool bts_available(void)
809 return x86_pmu.enable_bts != NULL;
812 static inline void init_debug_store_on_cpu(int cpu)
814 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
819 wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
820 (u32)((u64)(unsigned long)ds),
821 (u32)((u64)(unsigned long)ds >> 32));
824 static inline void fini_debug_store_on_cpu(int cpu)
826 if (!per_cpu(cpu_hw_events, cpu).ds)
829 wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
832 static void release_bts_hardware(void)
836 if (!bts_available())
841 for_each_online_cpu(cpu)
842 fini_debug_store_on_cpu(cpu);
844 for_each_possible_cpu(cpu) {
845 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
850 per_cpu(cpu_hw_events, cpu).ds = NULL;
852 kfree((void *)(unsigned long)ds->bts_buffer_base);
859 static int reserve_bts_hardware(void)
863 if (!bts_available())
868 for_each_possible_cpu(cpu) {
869 struct debug_store *ds;
873 buffer = kzalloc(BTS_BUFFER_SIZE, GFP_KERNEL);
874 if (unlikely(!buffer))
877 ds = kzalloc(sizeof(*ds), GFP_KERNEL);
883 ds->bts_buffer_base = (u64)(unsigned long)buffer;
884 ds->bts_index = ds->bts_buffer_base;
885 ds->bts_absolute_maximum =
886 ds->bts_buffer_base + BTS_BUFFER_SIZE;
887 ds->bts_interrupt_threshold =
888 ds->bts_absolute_maximum - BTS_OVFL_TH;
890 per_cpu(cpu_hw_events, cpu).ds = ds;
895 release_bts_hardware();
897 for_each_online_cpu(cpu)
898 init_debug_store_on_cpu(cpu);
906 static void hw_perf_event_destroy(struct perf_event *event)
908 if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
909 release_pmc_hardware();
910 release_bts_hardware();
911 mutex_unlock(&pmc_reserve_mutex);
915 static inline int x86_pmu_initialized(void)
917 return x86_pmu.handle_irq != NULL;
921 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event_attr *attr)
923 unsigned int cache_type, cache_op, cache_result;
926 config = attr->config;
928 cache_type = (config >> 0) & 0xff;
929 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
932 cache_op = (config >> 8) & 0xff;
933 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
936 cache_result = (config >> 16) & 0xff;
937 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
940 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
953 static void intel_pmu_enable_bts(u64 config)
955 unsigned long debugctlmsr;
957 debugctlmsr = get_debugctlmsr();
959 debugctlmsr |= X86_DEBUGCTL_TR;
960 debugctlmsr |= X86_DEBUGCTL_BTS;
961 debugctlmsr |= X86_DEBUGCTL_BTINT;
963 if (!(config & ARCH_PERFMON_EVENTSEL_OS))
964 debugctlmsr |= X86_DEBUGCTL_BTS_OFF_OS;
966 if (!(config & ARCH_PERFMON_EVENTSEL_USR))
967 debugctlmsr |= X86_DEBUGCTL_BTS_OFF_USR;
969 update_debugctlmsr(debugctlmsr);
972 static void intel_pmu_disable_bts(void)
974 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
975 unsigned long debugctlmsr;
980 debugctlmsr = get_debugctlmsr();
983 ~(X86_DEBUGCTL_TR | X86_DEBUGCTL_BTS | X86_DEBUGCTL_BTINT |
984 X86_DEBUGCTL_BTS_OFF_OS | X86_DEBUGCTL_BTS_OFF_USR);
986 update_debugctlmsr(debugctlmsr);
990 * Setup the hardware configuration for a given attr_type
992 static int __hw_perf_event_init(struct perf_event *event)
994 struct perf_event_attr *attr = &event->attr;
995 struct hw_perf_event *hwc = &event->hw;
999 if (!x86_pmu_initialized())
1003 if (!atomic_inc_not_zero(&active_events)) {
1004 mutex_lock(&pmc_reserve_mutex);
1005 if (atomic_read(&active_events) == 0) {
1006 if (!reserve_pmc_hardware())
1009 err = reserve_bts_hardware();
1012 atomic_inc(&active_events);
1013 mutex_unlock(&pmc_reserve_mutex);
1018 event->destroy = hw_perf_event_destroy;
1021 * Generate PMC IRQs:
1022 * (keep 'enabled' bit clear for now)
1024 hwc->config = ARCH_PERFMON_EVENTSEL_INT;
1029 * Count user and OS events unless requested not to.
1031 if (!attr->exclude_user)
1032 hwc->config |= ARCH_PERFMON_EVENTSEL_USR;
1033 if (!attr->exclude_kernel)
1034 hwc->config |= ARCH_PERFMON_EVENTSEL_OS;
1036 if (!hwc->sample_period) {
1037 hwc->sample_period = x86_pmu.max_period;
1038 hwc->last_period = hwc->sample_period;
1039 atomic64_set(&hwc->period_left, hwc->sample_period);
1042 * If we have a PMU initialized but no APIC
1043 * interrupts, we cannot sample hardware
1044 * events (user-space has to fall back and
1045 * sample via a hrtimer based software event):
1052 * Raw hw_event type provide the config in the hw_event structure
1054 if (attr->type == PERF_TYPE_RAW) {
1055 hwc->config |= x86_pmu.raw_event(attr->config);
1059 if (attr->type == PERF_TYPE_HW_CACHE)
1060 return set_ext_hw_attr(hwc, attr);
1062 if (attr->config >= x86_pmu.max_events)
1068 config = x86_pmu.event_map(attr->config);
1079 if ((attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS) &&
1080 (hwc->sample_period == 1)) {
1081 /* BTS is not supported by this architecture. */
1082 if (!bts_available())
1085 /* BTS is currently only allowed for user-mode. */
1086 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
1090 hwc->config |= config;
1095 static void p6_pmu_disable_all(void)
1097 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1106 /* p6 only has one enable register */
1107 rdmsrl(MSR_P6_EVNTSEL0, val);
1108 val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
1109 wrmsrl(MSR_P6_EVNTSEL0, val);
1112 static void intel_pmu_disable_all(void)
1114 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1122 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
1124 if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask))
1125 intel_pmu_disable_bts();
1128 static void amd_pmu_disable_all(void)
1130 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1138 * ensure we write the disable before we start disabling the
1139 * events proper, so that amd_pmu_enable_event() does the
1144 for (idx = 0; idx < x86_pmu.num_events; idx++) {
1147 if (!test_bit(idx, cpuc->active_mask))
1149 rdmsrl(MSR_K7_EVNTSEL0 + idx, val);
1150 if (!(val & ARCH_PERFMON_EVENTSEL0_ENABLE))
1152 val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
1153 wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
1157 void hw_perf_disable(void)
1159 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1161 if (!x86_pmu_initialized())
1167 x86_pmu.disable_all();
1170 static void p6_pmu_enable_all(void)
1172 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1181 /* p6 only has one enable register */
1182 rdmsrl(MSR_P6_EVNTSEL0, val);
1183 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
1184 wrmsrl(MSR_P6_EVNTSEL0, val);
1187 static void intel_pmu_enable_all(void)
1189 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1197 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl);
1199 if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
1200 struct perf_event *event =
1201 cpuc->events[X86_PMC_IDX_FIXED_BTS];
1203 if (WARN_ON_ONCE(!event))
1206 intel_pmu_enable_bts(event->hw.config);
1210 static void amd_pmu_enable_all(void)
1212 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1221 for (idx = 0; idx < x86_pmu.num_events; idx++) {
1222 struct perf_event *event = cpuc->events[idx];
1225 if (!test_bit(idx, cpuc->active_mask))
1228 val = event->hw.config;
1229 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
1230 wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
1234 static const struct pmu pmu;
1236 static inline int is_x86_event(struct perf_event *event)
1238 return event->pmu == &pmu;
1241 static int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
1243 struct event_constraint *c, *constraints[X86_PMC_IDX_MAX];
1244 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
1245 int i, j, w, wmax, num = 0;
1246 struct hw_perf_event *hwc;
1248 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
1250 for (i = 0; i < n; i++) {
1252 x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
1256 * fastpath, try to reuse previous register
1258 for (i = 0; i < n; i++) {
1259 hwc = &cpuc->event_list[i]->hw;
1262 /* never assigned */
1266 /* constraint still honored */
1267 if (!test_bit(hwc->idx, c->idxmsk))
1270 /* not already used */
1271 if (test_bit(hwc->idx, used_mask))
1274 set_bit(hwc->idx, used_mask);
1276 assign[i] = hwc->idx;
1285 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
1288 * weight = number of possible counters
1290 * 1 = most constrained, only works on one counter
1291 * wmax = least constrained, works on any counter
1293 * assign events to counters starting with most
1294 * constrained events.
1296 wmax = x86_pmu.num_events;
1299 * when fixed event counters are present,
1300 * wmax is incremented by 1 to account
1301 * for one more choice
1303 if (x86_pmu.num_events_fixed)
1306 for (w = 1, num = n; num && w <= wmax; w++) {
1307 /* for each event */
1308 for (i = 0; num && i < n; i++) {
1310 hwc = &cpuc->event_list[i]->hw;
1315 for_each_bit(j, c->idxmsk, X86_PMC_IDX_MAX) {
1316 if (!test_bit(j, used_mask))
1320 if (j == X86_PMC_IDX_MAX)
1323 set_bit(j, used_mask);
1332 * scheduling failed or is just a simulation,
1333 * free resources if necessary
1335 if (!assign || num) {
1336 for (i = 0; i < n; i++) {
1337 if (x86_pmu.put_event_constraints)
1338 x86_pmu.put_event_constraints(cpuc, cpuc->event_list[i]);
1341 return num ? -ENOSPC : 0;
1345 * dogrp: true if must collect siblings events (group)
1346 * returns total number of events and error code
1348 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
1350 struct perf_event *event;
1353 max_count = x86_pmu.num_events + x86_pmu.num_events_fixed;
1355 /* current number of events already accepted */
1358 if (is_x86_event(leader)) {
1361 cpuc->event_list[n] = leader;
1367 list_for_each_entry(event, &leader->sibling_list, group_entry) {
1368 if (!is_x86_event(event) ||
1369 event->state <= PERF_EVENT_STATE_OFF)
1375 cpuc->event_list[n] = event;
1382 static inline void x86_assign_hw_event(struct perf_event *event,
1383 struct hw_perf_event *hwc, int idx)
1387 if (hwc->idx == X86_PMC_IDX_FIXED_BTS) {
1388 hwc->config_base = 0;
1389 hwc->event_base = 0;
1390 } else if (hwc->idx >= X86_PMC_IDX_FIXED) {
1391 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
1393 * We set it so that event_base + idx in wrmsr/rdmsr maps to
1394 * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
1397 MSR_ARCH_PERFMON_FIXED_CTR0 - X86_PMC_IDX_FIXED;
1399 hwc->config_base = x86_pmu.eventsel;
1400 hwc->event_base = x86_pmu.perfctr;
1404 void hw_perf_enable(void)
1406 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1407 struct perf_event *event;
1408 struct hw_perf_event *hwc;
1411 if (!x86_pmu_initialized())
1413 if (cpuc->n_added) {
1415 * apply assignment obtained either from
1416 * hw_perf_group_sched_in() or x86_pmu_enable()
1418 * step1: save events moving to new counters
1419 * step2: reprogram moved events into new counters
1421 for (i = 0; i < cpuc->n_events; i++) {
1423 event = cpuc->event_list[i];
1426 if (hwc->idx == -1 || hwc->idx == cpuc->assign[i])
1429 x86_pmu.disable(hwc, hwc->idx);
1431 clear_bit(hwc->idx, cpuc->active_mask);
1433 cpuc->events[hwc->idx] = NULL;
1435 x86_perf_event_update(event, hwc, hwc->idx);
1440 for (i = 0; i < cpuc->n_events; i++) {
1442 event = cpuc->event_list[i];
1445 if (hwc->idx == -1) {
1446 x86_assign_hw_event(event, hwc, cpuc->assign[i]);
1447 x86_perf_event_set_period(event, hwc, hwc->idx);
1450 * need to mark as active because x86_pmu_disable()
1451 * clear active_mask and eventsp[] yet it preserves
1454 set_bit(hwc->idx, cpuc->active_mask);
1455 cpuc->events[hwc->idx] = event;
1457 x86_pmu.enable(hwc, hwc->idx);
1458 perf_event_update_userpage(event);
1461 perf_events_lapic_init();
1463 x86_pmu.enable_all();
1466 static inline u64 intel_pmu_get_status(void)
1470 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1475 static inline void intel_pmu_ack_status(u64 ack)
1477 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
1480 static inline void x86_pmu_enable_event(struct hw_perf_event *hwc, int idx)
1482 (void)checking_wrmsrl(hwc->config_base + idx,
1483 hwc->config | ARCH_PERFMON_EVENTSEL0_ENABLE);
1486 static inline void x86_pmu_disable_event(struct hw_perf_event *hwc, int idx)
1488 (void)checking_wrmsrl(hwc->config_base + idx, hwc->config);
1492 intel_pmu_disable_fixed(struct hw_perf_event *hwc, int __idx)
1494 int idx = __idx - X86_PMC_IDX_FIXED;
1497 mask = 0xfULL << (idx * 4);
1499 rdmsrl(hwc->config_base, ctrl_val);
1501 (void)checking_wrmsrl(hwc->config_base, ctrl_val);
1505 p6_pmu_disable_event(struct hw_perf_event *hwc, int idx)
1507 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1508 u64 val = P6_NOP_EVENT;
1511 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
1513 (void)checking_wrmsrl(hwc->config_base + idx, val);
1517 intel_pmu_disable_event(struct hw_perf_event *hwc, int idx)
1519 if (unlikely(idx == X86_PMC_IDX_FIXED_BTS)) {
1520 intel_pmu_disable_bts();
1524 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
1525 intel_pmu_disable_fixed(hwc, idx);
1529 x86_pmu_disable_event(hwc, idx);
1533 amd_pmu_disable_event(struct hw_perf_event *hwc, int idx)
1535 x86_pmu_disable_event(hwc, idx);
1538 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1541 * Set the next IRQ period, based on the hwc->period_left value.
1542 * To be called with the event disabled in hw:
1545 x86_perf_event_set_period(struct perf_event *event,
1546 struct hw_perf_event *hwc, int idx)
1548 s64 left = atomic64_read(&hwc->period_left);
1549 s64 period = hwc->sample_period;
1552 if (idx == X86_PMC_IDX_FIXED_BTS)
1556 * If we are way outside a reasonable range then just skip forward:
1558 if (unlikely(left <= -period)) {
1560 atomic64_set(&hwc->period_left, left);
1561 hwc->last_period = period;
1565 if (unlikely(left <= 0)) {
1567 atomic64_set(&hwc->period_left, left);
1568 hwc->last_period = period;
1572 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1574 if (unlikely(left < 2))
1577 if (left > x86_pmu.max_period)
1578 left = x86_pmu.max_period;
1580 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1583 * The hw event starts counting from this event offset,
1584 * mark it to be able to extra future deltas:
1586 atomic64_set(&hwc->prev_count, (u64)-left);
1588 err = checking_wrmsrl(hwc->event_base + idx,
1589 (u64)(-left) & x86_pmu.event_mask);
1591 perf_event_update_userpage(event);
1597 intel_pmu_enable_fixed(struct hw_perf_event *hwc, int __idx)
1599 int idx = __idx - X86_PMC_IDX_FIXED;
1600 u64 ctrl_val, bits, mask;
1604 * Enable IRQ generation (0x8),
1605 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
1609 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
1611 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
1614 mask = 0xfULL << (idx * 4);
1616 rdmsrl(hwc->config_base, ctrl_val);
1619 err = checking_wrmsrl(hwc->config_base, ctrl_val);
1622 static void p6_pmu_enable_event(struct hw_perf_event *hwc, int idx)
1624 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1629 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
1631 (void)checking_wrmsrl(hwc->config_base + idx, val);
1635 static void intel_pmu_enable_event(struct hw_perf_event *hwc, int idx)
1637 if (unlikely(idx == X86_PMC_IDX_FIXED_BTS)) {
1638 if (!__get_cpu_var(cpu_hw_events).enabled)
1641 intel_pmu_enable_bts(hwc->config);
1645 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
1646 intel_pmu_enable_fixed(hwc, idx);
1650 x86_pmu_enable_event(hwc, idx);
1653 static void amd_pmu_enable_event(struct hw_perf_event *hwc, int idx)
1655 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1658 x86_pmu_enable_event(hwc, idx);
1662 * activate a single event
1664 * The event is added to the group of enabled events
1665 * but only if it can be scehduled with existing events.
1667 * Called with PMU disabled. If successful and return value 1,
1668 * then guaranteed to call perf_enable() and hw_perf_enable()
1670 static int x86_pmu_enable(struct perf_event *event)
1672 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1673 struct hw_perf_event *hwc;
1674 int assign[X86_PMC_IDX_MAX];
1679 n0 = cpuc->n_events;
1680 n = collect_events(cpuc, event, false);
1684 ret = x86_schedule_events(cpuc, n, assign);
1688 * copy new assignment, now we know it is possible
1689 * will be used by hw_perf_enable()
1691 memcpy(cpuc->assign, assign, n*sizeof(int));
1694 cpuc->n_added = n - n0;
1697 x86_perf_event_set_period(event, hwc, hwc->idx);
1702 static void x86_pmu_unthrottle(struct perf_event *event)
1704 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1705 struct hw_perf_event *hwc = &event->hw;
1707 if (WARN_ON_ONCE(hwc->idx >= X86_PMC_IDX_MAX ||
1708 cpuc->events[hwc->idx] != event))
1711 x86_pmu.enable(hwc, hwc->idx);
1714 void perf_event_print_debug(void)
1716 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1717 struct cpu_hw_events *cpuc;
1718 unsigned long flags;
1721 if (!x86_pmu.num_events)
1724 local_irq_save(flags);
1726 cpu = smp_processor_id();
1727 cpuc = &per_cpu(cpu_hw_events, cpu);
1729 if (x86_pmu.version >= 2) {
1730 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1731 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1732 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1733 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1736 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1737 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1738 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1739 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1741 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1743 for (idx = 0; idx < x86_pmu.num_events; idx++) {
1744 rdmsrl(x86_pmu.eventsel + idx, pmc_ctrl);
1745 rdmsrl(x86_pmu.perfctr + idx, pmc_count);
1747 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1749 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1750 cpu, idx, pmc_ctrl);
1751 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1752 cpu, idx, pmc_count);
1753 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1754 cpu, idx, prev_left);
1756 for (idx = 0; idx < x86_pmu.num_events_fixed; idx++) {
1757 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1759 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1760 cpu, idx, pmc_count);
1762 local_irq_restore(flags);
1765 static void intel_pmu_drain_bts_buffer(struct cpu_hw_events *cpuc)
1767 struct debug_store *ds = cpuc->ds;
1773 struct perf_event *event = cpuc->events[X86_PMC_IDX_FIXED_BTS];
1774 struct bts_record *at, *top;
1775 struct perf_output_handle handle;
1776 struct perf_event_header header;
1777 struct perf_sample_data data;
1778 struct pt_regs regs;
1786 at = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
1787 top = (struct bts_record *)(unsigned long)ds->bts_index;
1792 ds->bts_index = ds->bts_buffer_base;
1795 data.period = event->hw.last_period;
1801 * Prepare a generic sample, i.e. fill in the invariant fields.
1802 * We will overwrite the from and to address before we output
1805 perf_prepare_sample(&header, &data, event, ®s);
1807 if (perf_output_begin(&handle, event,
1808 header.size * (top - at), 1, 1))
1811 for (; at < top; at++) {
1815 perf_output_sample(&handle, &header, &data, event);
1818 perf_output_end(&handle);
1820 /* There's new data available. */
1821 event->hw.interrupts++;
1822 event->pending_kill = POLL_IN;
1825 static void x86_pmu_disable(struct perf_event *event)
1827 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1828 struct hw_perf_event *hwc = &event->hw;
1829 int i, idx = hwc->idx;
1832 * Must be done before we disable, otherwise the nmi handler
1833 * could reenable again:
1835 clear_bit(idx, cpuc->active_mask);
1836 x86_pmu.disable(hwc, idx);
1839 * Make sure the cleared pointer becomes visible before we
1840 * (potentially) free the event:
1845 * Drain the remaining delta count out of a event
1846 * that we are disabling:
1848 x86_perf_event_update(event, hwc, idx);
1850 /* Drain the remaining BTS records. */
1851 if (unlikely(idx == X86_PMC_IDX_FIXED_BTS))
1852 intel_pmu_drain_bts_buffer(cpuc);
1854 cpuc->events[idx] = NULL;
1856 for (i = 0; i < cpuc->n_events; i++) {
1857 if (event == cpuc->event_list[i]) {
1859 if (x86_pmu.put_event_constraints)
1860 x86_pmu.put_event_constraints(cpuc, event);
1862 while (++i < cpuc->n_events)
1863 cpuc->event_list[i-1] = cpuc->event_list[i];
1869 perf_event_update_userpage(event);
1873 * Save and restart an expired event. Called by NMI contexts,
1874 * so it has to be careful about preempting normal event ops:
1876 static int intel_pmu_save_and_restart(struct perf_event *event)
1878 struct hw_perf_event *hwc = &event->hw;
1882 x86_perf_event_update(event, hwc, idx);
1883 ret = x86_perf_event_set_period(event, hwc, idx);
1885 if (event->state == PERF_EVENT_STATE_ACTIVE)
1886 intel_pmu_enable_event(hwc, idx);
1891 static void intel_pmu_reset(void)
1893 struct debug_store *ds = __get_cpu_var(cpu_hw_events).ds;
1894 unsigned long flags;
1897 if (!x86_pmu.num_events)
1900 local_irq_save(flags);
1902 printk("clearing PMU state on CPU#%d\n", smp_processor_id());
1904 for (idx = 0; idx < x86_pmu.num_events; idx++) {
1905 checking_wrmsrl(x86_pmu.eventsel + idx, 0ull);
1906 checking_wrmsrl(x86_pmu.perfctr + idx, 0ull);
1908 for (idx = 0; idx < x86_pmu.num_events_fixed; idx++) {
1909 checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
1912 ds->bts_index = ds->bts_buffer_base;
1914 local_irq_restore(flags);
1917 static int p6_pmu_handle_irq(struct pt_regs *regs)
1919 struct perf_sample_data data;
1920 struct cpu_hw_events *cpuc;
1921 struct perf_event *event;
1922 struct hw_perf_event *hwc;
1923 int idx, handled = 0;
1929 cpuc = &__get_cpu_var(cpu_hw_events);
1931 for (idx = 0; idx < x86_pmu.num_events; idx++) {
1932 if (!test_bit(idx, cpuc->active_mask))
1935 event = cpuc->events[idx];
1938 val = x86_perf_event_update(event, hwc, idx);
1939 if (val & (1ULL << (x86_pmu.event_bits - 1)))
1946 data.period = event->hw.last_period;
1948 if (!x86_perf_event_set_period(event, hwc, idx))
1951 if (perf_event_overflow(event, 1, &data, regs))
1952 p6_pmu_disable_event(hwc, idx);
1956 inc_irq_stat(apic_perf_irqs);
1962 * This handler is triggered by the local APIC, so the APIC IRQ handling
1965 static int intel_pmu_handle_irq(struct pt_regs *regs)
1967 struct perf_sample_data data;
1968 struct cpu_hw_events *cpuc;
1975 cpuc = &__get_cpu_var(cpu_hw_events);
1978 intel_pmu_drain_bts_buffer(cpuc);
1979 status = intel_pmu_get_status();
1987 if (++loops > 100) {
1988 WARN_ONCE(1, "perfevents: irq loop stuck!\n");
1989 perf_event_print_debug();
1995 inc_irq_stat(apic_perf_irqs);
1997 for_each_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
1998 struct perf_event *event = cpuc->events[bit];
2000 clear_bit(bit, (unsigned long *) &status);
2001 if (!test_bit(bit, cpuc->active_mask))
2004 if (!intel_pmu_save_and_restart(event))
2007 data.period = event->hw.last_period;
2009 if (perf_event_overflow(event, 1, &data, regs))
2010 intel_pmu_disable_event(&event->hw, bit);
2013 intel_pmu_ack_status(ack);
2016 * Repeat if there is more work to be done:
2018 status = intel_pmu_get_status();
2027 static int amd_pmu_handle_irq(struct pt_regs *regs)
2029 struct perf_sample_data data;
2030 struct cpu_hw_events *cpuc;
2031 struct perf_event *event;
2032 struct hw_perf_event *hwc;
2033 int idx, handled = 0;
2039 cpuc = &__get_cpu_var(cpu_hw_events);
2041 for (idx = 0; idx < x86_pmu.num_events; idx++) {
2042 if (!test_bit(idx, cpuc->active_mask))
2045 event = cpuc->events[idx];
2048 val = x86_perf_event_update(event, hwc, idx);
2049 if (val & (1ULL << (x86_pmu.event_bits - 1)))
2056 data.period = event->hw.last_period;
2058 if (!x86_perf_event_set_period(event, hwc, idx))
2061 if (perf_event_overflow(event, 1, &data, regs))
2062 amd_pmu_disable_event(hwc, idx);
2066 inc_irq_stat(apic_perf_irqs);
2071 void smp_perf_pending_interrupt(struct pt_regs *regs)
2075 inc_irq_stat(apic_pending_irqs);
2076 perf_event_do_pending();
2080 void set_perf_event_pending(void)
2082 #ifdef CONFIG_X86_LOCAL_APIC
2083 if (!x86_pmu.apic || !x86_pmu_initialized())
2086 apic->send_IPI_self(LOCAL_PENDING_VECTOR);
2090 void perf_events_lapic_init(void)
2092 #ifdef CONFIG_X86_LOCAL_APIC
2093 if (!x86_pmu.apic || !x86_pmu_initialized())
2097 * Always use NMI for PMU
2099 apic_write(APIC_LVTPC, APIC_DM_NMI);
2103 static int __kprobes
2104 perf_event_nmi_handler(struct notifier_block *self,
2105 unsigned long cmd, void *__args)
2107 struct die_args *args = __args;
2108 struct pt_regs *regs;
2110 if (!atomic_read(&active_events))
2124 #ifdef CONFIG_X86_LOCAL_APIC
2125 apic_write(APIC_LVTPC, APIC_DM_NMI);
2128 * Can't rely on the handled return value to say it was our NMI, two
2129 * events could trigger 'simultaneously' raising two back-to-back NMIs.
2131 * If the first NMI handles both, the latter will be empty and daze
2134 x86_pmu.handle_irq(regs);
2139 static struct event_constraint unconstrained;
2141 static struct event_constraint bts_constraint =
2142 EVENT_CONSTRAINT(0, 1ULL << X86_PMC_IDX_FIXED_BTS, 0);
2144 static struct event_constraint *
2145 intel_special_constraints(struct perf_event *event)
2147 unsigned int hw_event;
2149 hw_event = event->hw.config & INTEL_ARCH_EVENT_MASK;
2151 if (unlikely((hw_event ==
2152 x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS)) &&
2153 (event->hw.sample_period == 1))) {
2155 return &bts_constraint;
2160 static struct event_constraint *
2161 intel_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
2163 struct event_constraint *c;
2165 c = intel_special_constraints(event);
2169 if (x86_pmu.event_constraints) {
2170 for_each_event_constraint(c, x86_pmu.event_constraints) {
2171 if ((event->hw.config & c->cmask) == c->code)
2176 return &unconstrained;
2179 static struct event_constraint *
2180 amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
2182 return &unconstrained;
2185 static int x86_event_sched_in(struct perf_event *event,
2186 struct perf_cpu_context *cpuctx, int cpu)
2190 event->state = PERF_EVENT_STATE_ACTIVE;
2192 event->tstamp_running += event->ctx->time - event->tstamp_stopped;
2194 if (!is_x86_event(event))
2195 ret = event->pmu->enable(event);
2197 if (!ret && !is_software_event(event))
2198 cpuctx->active_oncpu++;
2200 if (!ret && event->attr.exclusive)
2201 cpuctx->exclusive = 1;
2206 static void x86_event_sched_out(struct perf_event *event,
2207 struct perf_cpu_context *cpuctx, int cpu)
2209 event->state = PERF_EVENT_STATE_INACTIVE;
2212 if (!is_x86_event(event))
2213 event->pmu->disable(event);
2215 event->tstamp_running -= event->ctx->time - event->tstamp_stopped;
2217 if (!is_software_event(event))
2218 cpuctx->active_oncpu--;
2220 if (event->attr.exclusive || !cpuctx->active_oncpu)
2221 cpuctx->exclusive = 0;
2225 * Called to enable a whole group of events.
2226 * Returns 1 if the group was enabled, or -EAGAIN if it could not be.
2227 * Assumes the caller has disabled interrupts and has
2228 * frozen the PMU with hw_perf_save_disable.
2230 * called with PMU disabled. If successful and return value 1,
2231 * then guaranteed to call perf_enable() and hw_perf_enable()
2233 int hw_perf_group_sched_in(struct perf_event *leader,
2234 struct perf_cpu_context *cpuctx,
2235 struct perf_event_context *ctx, int cpu)
2237 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
2238 struct perf_event *sub;
2239 int assign[X86_PMC_IDX_MAX];
2242 /* n0 = total number of events */
2243 n0 = collect_events(cpuc, leader, true);
2247 ret = x86_schedule_events(cpuc, n0, assign);
2251 ret = x86_event_sched_in(leader, cpuctx, cpu);
2256 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2257 if (sub->state > PERF_EVENT_STATE_OFF) {
2258 ret = x86_event_sched_in(sub, cpuctx, cpu);
2265 * copy new assignment, now we know it is possible
2266 * will be used by hw_perf_enable()
2268 memcpy(cpuc->assign, assign, n0*sizeof(int));
2270 cpuc->n_events = n0;
2272 ctx->nr_active += n1;
2275 * 1 means successful and events are active
2276 * This is not quite true because we defer
2277 * actual activation until hw_perf_enable() but
2278 * this way we* ensure caller won't try to enable
2283 x86_event_sched_out(leader, cpuctx, cpu);
2285 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2286 if (sub->state == PERF_EVENT_STATE_ACTIVE) {
2287 x86_event_sched_out(sub, cpuctx, cpu);
2295 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
2296 .notifier_call = perf_event_nmi_handler,
2301 static __initconst struct x86_pmu p6_pmu = {
2303 .handle_irq = p6_pmu_handle_irq,
2304 .disable_all = p6_pmu_disable_all,
2305 .enable_all = p6_pmu_enable_all,
2306 .enable = p6_pmu_enable_event,
2307 .disable = p6_pmu_disable_event,
2308 .eventsel = MSR_P6_EVNTSEL0,
2309 .perfctr = MSR_P6_PERFCTR0,
2310 .event_map = p6_pmu_event_map,
2311 .raw_event = p6_pmu_raw_event,
2312 .max_events = ARRAY_SIZE(p6_perfmon_event_map),
2314 .max_period = (1ULL << 31) - 1,
2318 * Events have 40 bits implemented. However they are designed such
2319 * that bits [32-39] are sign extensions of bit 31. As such the
2320 * effective width of a event for P6-like PMU is 32 bits only.
2322 * See IA-32 Intel Architecture Software developer manual Vol 3B
2325 .event_mask = (1ULL << 32) - 1,
2326 .get_event_constraints = intel_get_event_constraints,
2327 .event_constraints = intel_p6_event_constraints
2330 static __initconst struct x86_pmu intel_pmu = {
2332 .handle_irq = intel_pmu_handle_irq,
2333 .disable_all = intel_pmu_disable_all,
2334 .enable_all = intel_pmu_enable_all,
2335 .enable = intel_pmu_enable_event,
2336 .disable = intel_pmu_disable_event,
2337 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
2338 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
2339 .event_map = intel_pmu_event_map,
2340 .raw_event = intel_pmu_raw_event,
2341 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
2344 * Intel PMCs cannot be accessed sanely above 32 bit width,
2345 * so we install an artificial 1<<31 period regardless of
2346 * the generic event period:
2348 .max_period = (1ULL << 31) - 1,
2349 .enable_bts = intel_pmu_enable_bts,
2350 .disable_bts = intel_pmu_disable_bts,
2351 .get_event_constraints = intel_get_event_constraints
2354 static __initconst struct x86_pmu amd_pmu = {
2356 .handle_irq = amd_pmu_handle_irq,
2357 .disable_all = amd_pmu_disable_all,
2358 .enable_all = amd_pmu_enable_all,
2359 .enable = amd_pmu_enable_event,
2360 .disable = amd_pmu_disable_event,
2361 .eventsel = MSR_K7_EVNTSEL0,
2362 .perfctr = MSR_K7_PERFCTR0,
2363 .event_map = amd_pmu_event_map,
2364 .raw_event = amd_pmu_raw_event,
2365 .max_events = ARRAY_SIZE(amd_perfmon_event_map),
2368 .event_mask = (1ULL << 48) - 1,
2370 /* use highest bit to detect overflow */
2371 .max_period = (1ULL << 47) - 1,
2372 .get_event_constraints = amd_get_event_constraints
2375 static __init int p6_pmu_init(void)
2377 switch (boot_cpu_data.x86_model) {
2379 case 3: /* Pentium Pro */
2381 case 6: /* Pentium II */
2384 case 11: /* Pentium III */
2390 pr_cont("unsupported p6 CPU model %d ",
2391 boot_cpu_data.x86_model);
2400 static __init int intel_pmu_init(void)
2402 union cpuid10_edx edx;
2403 union cpuid10_eax eax;
2404 unsigned int unused;
2408 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
2409 /* check for P6 processor family */
2410 if (boot_cpu_data.x86 == 6) {
2411 return p6_pmu_init();
2418 * Check whether the Architectural PerfMon supports
2419 * Branch Misses Retired hw_event or not.
2421 cpuid(10, &eax.full, &ebx, &unused, &edx.full);
2422 if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
2425 version = eax.split.version_id;
2429 x86_pmu = intel_pmu;
2430 x86_pmu.version = version;
2431 x86_pmu.num_events = eax.split.num_events;
2432 x86_pmu.event_bits = eax.split.bit_width;
2433 x86_pmu.event_mask = (1ULL << eax.split.bit_width) - 1;
2436 * Quirk: v2 perfmon does not report fixed-purpose events, so
2437 * assume at least 3 events:
2439 x86_pmu.num_events_fixed = max((int)edx.split.num_events_fixed, 3);
2442 * Install the hw-cache-events table:
2444 switch (boot_cpu_data.x86_model) {
2445 case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
2446 case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */
2447 case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */
2448 case 29: /* six-core 45 nm xeon "Dunnington" */
2449 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
2450 sizeof(hw_cache_event_ids));
2452 x86_pmu.event_constraints = intel_core_event_constraints;
2453 pr_cont("Core2 events, ");
2456 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
2457 sizeof(hw_cache_event_ids));
2459 x86_pmu.event_constraints = intel_nehalem_event_constraints;
2460 pr_cont("Nehalem/Corei7 events, ");
2463 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
2464 sizeof(hw_cache_event_ids));
2466 x86_pmu.event_constraints = intel_gen_event_constraints;
2467 pr_cont("Atom events, ");
2471 * default constraints for v2 and up
2473 x86_pmu.event_constraints = intel_gen_event_constraints;
2474 pr_cont("generic architected perfmon, ");
2479 static __init int amd_pmu_init(void)
2481 /* Performance-monitoring supported from K7 and later: */
2482 if (boot_cpu_data.x86 < 6)
2487 /* Events are common for all AMDs */
2488 memcpy(hw_cache_event_ids, amd_hw_cache_event_ids,
2489 sizeof(hw_cache_event_ids));
2494 static void __init pmu_check_apic(void)
2500 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
2501 pr_info("no hardware sampling interrupt available.\n");
2504 void __init init_hw_perf_events(void)
2508 pr_info("Performance Events: ");
2510 switch (boot_cpu_data.x86_vendor) {
2511 case X86_VENDOR_INTEL:
2512 err = intel_pmu_init();
2514 case X86_VENDOR_AMD:
2515 err = amd_pmu_init();
2521 pr_cont("no PMU driver, software events only.\n");
2527 pr_cont("%s PMU driver.\n", x86_pmu.name);
2529 if (x86_pmu.num_events > X86_PMC_MAX_GENERIC) {
2530 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
2531 x86_pmu.num_events, X86_PMC_MAX_GENERIC);
2532 x86_pmu.num_events = X86_PMC_MAX_GENERIC;
2534 perf_event_mask = (1 << x86_pmu.num_events) - 1;
2535 perf_max_events = x86_pmu.num_events;
2537 if (x86_pmu.num_events_fixed > X86_PMC_MAX_FIXED) {
2538 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
2539 x86_pmu.num_events_fixed, X86_PMC_MAX_FIXED);
2540 x86_pmu.num_events_fixed = X86_PMC_MAX_FIXED;
2544 ((1LL << x86_pmu.num_events_fixed)-1) << X86_PMC_IDX_FIXED;
2545 x86_pmu.intel_ctrl = perf_event_mask;
2547 perf_events_lapic_init();
2548 register_die_notifier(&perf_event_nmi_notifier);
2550 unconstrained = (struct event_constraint)
2551 EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_events) - 1, 0);
2553 pr_info("... version: %d\n", x86_pmu.version);
2554 pr_info("... bit width: %d\n", x86_pmu.event_bits);
2555 pr_info("... generic registers: %d\n", x86_pmu.num_events);
2556 pr_info("... value mask: %016Lx\n", x86_pmu.event_mask);
2557 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
2558 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_events_fixed);
2559 pr_info("... event mask: %016Lx\n", perf_event_mask);
2562 static inline void x86_pmu_read(struct perf_event *event)
2564 x86_perf_event_update(event, &event->hw, event->hw.idx);
2567 static const struct pmu pmu = {
2568 .enable = x86_pmu_enable,
2569 .disable = x86_pmu_disable,
2570 .read = x86_pmu_read,
2571 .unthrottle = x86_pmu_unthrottle,
2575 * validate a single event group
2577 * validation include:
2578 * - check events are compatible which each other
2579 * - events do not compete for the same counter
2580 * - number of events <= number of counters
2582 * validation ensures the group can be loaded onto the
2583 * PMU if it was the only group available.
2585 static int validate_group(struct perf_event *event)
2587 struct perf_event *leader = event->group_leader;
2588 struct cpu_hw_events *fake_cpuc;
2592 fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
2597 * the event is not yet connected with its
2598 * siblings therefore we must first collect
2599 * existing siblings, then add the new event
2600 * before we can simulate the scheduling
2603 n = collect_events(fake_cpuc, leader, true);
2607 fake_cpuc->n_events = n;
2608 n = collect_events(fake_cpuc, event, false);
2612 fake_cpuc->n_events = n;
2614 ret = x86_schedule_events(fake_cpuc, n, NULL);
2622 const struct pmu *hw_perf_event_init(struct perf_event *event)
2624 const struct pmu *tmp;
2627 err = __hw_perf_event_init(event);
2630 * we temporarily connect event to its pmu
2631 * such that validate_group() can classify
2632 * it as an x86 event using is_x86_event()
2637 if (event->group_leader != event)
2638 err = validate_group(event);
2644 event->destroy(event);
2645 return ERR_PTR(err);
2656 void callchain_store(struct perf_callchain_entry *entry, u64 ip)
2658 if (entry->nr < PERF_MAX_STACK_DEPTH)
2659 entry->ip[entry->nr++] = ip;
2662 static DEFINE_PER_CPU(struct perf_callchain_entry, pmc_irq_entry);
2663 static DEFINE_PER_CPU(struct perf_callchain_entry, pmc_nmi_entry);
2667 backtrace_warning_symbol(void *data, char *msg, unsigned long symbol)
2669 /* Ignore warnings */
2672 static void backtrace_warning(void *data, char *msg)
2674 /* Ignore warnings */
2677 static int backtrace_stack(void *data, char *name)
2682 static void backtrace_address(void *data, unsigned long addr, int reliable)
2684 struct perf_callchain_entry *entry = data;
2687 callchain_store(entry, addr);
2690 static const struct stacktrace_ops backtrace_ops = {
2691 .warning = backtrace_warning,
2692 .warning_symbol = backtrace_warning_symbol,
2693 .stack = backtrace_stack,
2694 .address = backtrace_address,
2695 .walk_stack = print_context_stack_bp,
2698 #include "../dumpstack.h"
2701 perf_callchain_kernel(struct pt_regs *regs, struct perf_callchain_entry *entry)
2703 callchain_store(entry, PERF_CONTEXT_KERNEL);
2704 callchain_store(entry, regs->ip);
2706 dump_trace(NULL, regs, NULL, regs->bp, &backtrace_ops, entry);
2710 * best effort, GUP based copy_from_user() that assumes IRQ or NMI context
2712 static unsigned long
2713 copy_from_user_nmi(void *to, const void __user *from, unsigned long n)
2715 unsigned long offset, addr = (unsigned long)from;
2716 int type = in_nmi() ? KM_NMI : KM_IRQ0;
2717 unsigned long size, len = 0;
2723 ret = __get_user_pages_fast(addr, 1, 0, &page);
2727 offset = addr & (PAGE_SIZE - 1);
2728 size = min(PAGE_SIZE - offset, n - len);
2730 map = kmap_atomic(page, type);
2731 memcpy(to, map+offset, size);
2732 kunmap_atomic(map, type);
2744 static int copy_stack_frame(const void __user *fp, struct stack_frame *frame)
2746 unsigned long bytes;
2748 bytes = copy_from_user_nmi(frame, fp, sizeof(*frame));
2750 return bytes == sizeof(*frame);
2754 perf_callchain_user(struct pt_regs *regs, struct perf_callchain_entry *entry)
2756 struct stack_frame frame;
2757 const void __user *fp;
2759 if (!user_mode(regs))
2760 regs = task_pt_regs(current);
2762 fp = (void __user *)regs->bp;
2764 callchain_store(entry, PERF_CONTEXT_USER);
2765 callchain_store(entry, regs->ip);
2767 while (entry->nr < PERF_MAX_STACK_DEPTH) {
2768 frame.next_frame = NULL;
2769 frame.return_address = 0;
2771 if (!copy_stack_frame(fp, &frame))
2774 if ((unsigned long)fp < regs->sp)
2777 callchain_store(entry, frame.return_address);
2778 fp = frame.next_frame;
2783 perf_do_callchain(struct pt_regs *regs, struct perf_callchain_entry *entry)
2790 is_user = user_mode(regs);
2792 if (is_user && current->state != TASK_RUNNING)
2796 perf_callchain_kernel(regs, entry);
2799 perf_callchain_user(regs, entry);
2802 struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2804 struct perf_callchain_entry *entry;
2807 entry = &__get_cpu_var(pmc_nmi_entry);
2809 entry = &__get_cpu_var(pmc_irq_entry);
2813 perf_do_callchain(regs, entry);
2818 void hw_perf_event_setup_online(int cpu)
2820 init_debug_store_on_cpu(cpu);