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
)];
84 int nb_id
; /* NorthBridge id */
85 int refcnt
; /* reference count */
86 struct perf_event
*owners
[X86_PMC_IDX_MAX
];
87 struct event_constraint event_constraints
[X86_PMC_IDX_MAX
];
90 struct cpu_hw_events
{
91 struct perf_event
*events
[X86_PMC_IDX_MAX
]; /* in counter order */
92 unsigned long active_mask
[BITS_TO_LONGS(X86_PMC_IDX_MAX
)];
93 unsigned long interrupts
;
95 struct debug_store
*ds
;
99 int assign
[X86_PMC_IDX_MAX
]; /* event to counter assignment */
100 u64 tags
[X86_PMC_IDX_MAX
];
101 struct perf_event
*event_list
[X86_PMC_IDX_MAX
]; /* in enabled order */
102 struct amd_nb
*amd_nb
;
105 #define __EVENT_CONSTRAINT(c, n, m, w) {\
106 { .idxmsk64[0] = (n) }, \
112 #define EVENT_CONSTRAINT(c, n, m) \
113 __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n))
115 #define INTEL_EVENT_CONSTRAINT(c, n) \
116 EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVTSEL_MASK)
118 #define FIXED_EVENT_CONSTRAINT(c, n) \
119 EVENT_CONSTRAINT(c, n, INTEL_ARCH_FIXED_MASK)
121 #define EVENT_CONSTRAINT_END \
122 EVENT_CONSTRAINT(0, 0, 0)
124 #define for_each_event_constraint(e, c) \
125 for ((e) = (c); (e)->cmask; (e)++)
128 * struct x86_pmu - generic x86 pmu
133 int (*handle_irq
)(struct pt_regs
*);
134 void (*disable_all
)(void);
135 void (*enable_all
)(void);
136 void (*enable
)(struct hw_perf_event
*, int);
137 void (*disable
)(struct hw_perf_event
*, int);
140 u64 (*event_map
)(int);
141 u64 (*raw_event
)(u64
);
144 int num_events_fixed
;
150 void (*enable_bts
)(u64 config
);
151 void (*disable_bts
)(void);
153 struct event_constraint
*
154 (*get_event_constraints
)(struct cpu_hw_events
*cpuc
,
155 struct perf_event
*event
);
157 void (*put_event_constraints
)(struct cpu_hw_events
*cpuc
,
158 struct perf_event
*event
);
159 struct event_constraint
*event_constraints
;
162 static struct x86_pmu x86_pmu __read_mostly
;
164 static raw_spinlock_t amd_nb_lock
;
166 static DEFINE_PER_CPU(struct cpu_hw_events
, cpu_hw_events
) = {
170 static int x86_perf_event_set_period(struct perf_event
*event
,
171 struct hw_perf_event
*hwc
, int idx
);
174 * Not sure about some of these
176 static const u64 p6_perfmon_event_map
[] =
178 [PERF_COUNT_HW_CPU_CYCLES
] = 0x0079,
179 [PERF_COUNT_HW_INSTRUCTIONS
] = 0x00c0,
180 [PERF_COUNT_HW_CACHE_REFERENCES
] = 0x0f2e,
181 [PERF_COUNT_HW_CACHE_MISSES
] = 0x012e,
182 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS
] = 0x00c4,
183 [PERF_COUNT_HW_BRANCH_MISSES
] = 0x00c5,
184 [PERF_COUNT_HW_BUS_CYCLES
] = 0x0062,
187 static u64
p6_pmu_event_map(int hw_event
)
189 return p6_perfmon_event_map
[hw_event
];
193 * Event setting that is specified not to count anything.
194 * We use this to effectively disable a counter.
196 * L2_RQSTS with 0 MESI unit mask.
198 #define P6_NOP_EVENT 0x0000002EULL
200 static u64
p6_pmu_raw_event(u64 hw_event
)
202 #define P6_EVNTSEL_EVENT_MASK 0x000000FFULL
203 #define P6_EVNTSEL_UNIT_MASK 0x0000FF00ULL
204 #define P6_EVNTSEL_EDGE_MASK 0x00040000ULL
205 #define P6_EVNTSEL_INV_MASK 0x00800000ULL
206 #define P6_EVNTSEL_REG_MASK 0xFF000000ULL
208 #define P6_EVNTSEL_MASK \
209 (P6_EVNTSEL_EVENT_MASK | \
210 P6_EVNTSEL_UNIT_MASK | \
211 P6_EVNTSEL_EDGE_MASK | \
212 P6_EVNTSEL_INV_MASK | \
215 return hw_event
& P6_EVNTSEL_MASK
;
218 static struct event_constraint intel_p6_event_constraints
[] =
220 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FLOPS */
221 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
222 INTEL_EVENT_CONSTRAINT(0x11, 0x1), /* FP_ASSIST */
223 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
224 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
225 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
230 * Intel PerfMon v3. Used on Core2 and later.
232 static const u64 intel_perfmon_event_map
[] =
234 [PERF_COUNT_HW_CPU_CYCLES
] = 0x003c,
235 [PERF_COUNT_HW_INSTRUCTIONS
] = 0x00c0,
236 [PERF_COUNT_HW_CACHE_REFERENCES
] = 0x4f2e,
237 [PERF_COUNT_HW_CACHE_MISSES
] = 0x412e,
238 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS
] = 0x00c4,
239 [PERF_COUNT_HW_BRANCH_MISSES
] = 0x00c5,
240 [PERF_COUNT_HW_BUS_CYCLES
] = 0x013c,
243 static struct event_constraint intel_core_event_constraints
[] =
245 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
246 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
247 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
248 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
249 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
250 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */
254 static struct event_constraint intel_core2_event_constraints
[] =
256 FIXED_EVENT_CONSTRAINT(0xc0, (0x3|(1ULL<<32))), /* INSTRUCTIONS_RETIRED */
257 FIXED_EVENT_CONSTRAINT(0x3c, (0x3|(1ULL<<33))), /* UNHALTED_CORE_CYCLES */
258 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
259 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
260 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
261 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
262 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
263 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
264 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
265 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
266 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
270 static struct event_constraint intel_nehalem_event_constraints
[] =
272 FIXED_EVENT_CONSTRAINT(0xc0, (0xf|(1ULL<<32))), /* INSTRUCTIONS_RETIRED */
273 FIXED_EVENT_CONSTRAINT(0x3c, (0xf|(1ULL<<33))), /* UNHALTED_CORE_CYCLES */
274 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
275 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
276 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
277 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
278 INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */
279 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
280 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
281 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
285 static struct event_constraint intel_westmere_event_constraints
[] =
287 FIXED_EVENT_CONSTRAINT(0xc0, (0xf|(1ULL<<32))), /* INSTRUCTIONS_RETIRED */
288 FIXED_EVENT_CONSTRAINT(0x3c, (0xf|(1ULL<<33))), /* UNHALTED_CORE_CYCLES */
289 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
290 INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */
291 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
295 static struct event_constraint intel_gen_event_constraints
[] =
297 FIXED_EVENT_CONSTRAINT(0xc0, (0x3|(1ULL<<32))), /* INSTRUCTIONS_RETIRED */
298 FIXED_EVENT_CONSTRAINT(0x3c, (0x3|(1ULL<<33))), /* UNHALTED_CORE_CYCLES */
302 static u64
intel_pmu_event_map(int hw_event
)
304 return intel_perfmon_event_map
[hw_event
];
308 * Generalized hw caching related hw_event table, filled
309 * in on a per model basis. A value of 0 means
310 * 'not supported', -1 means 'hw_event makes no sense on
311 * this CPU', any other value means the raw hw_event
315 #define C(x) PERF_COUNT_HW_CACHE_##x
317 static u64 __read_mostly hw_cache_event_ids
318 [PERF_COUNT_HW_CACHE_MAX
]
319 [PERF_COUNT_HW_CACHE_OP_MAX
]
320 [PERF_COUNT_HW_CACHE_RESULT_MAX
];
322 static __initconst u64 westmere_hw_cache_event_ids
323 [PERF_COUNT_HW_CACHE_MAX
]
324 [PERF_COUNT_HW_CACHE_OP_MAX
]
325 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
329 [ C(RESULT_ACCESS
) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
330 [ C(RESULT_MISS
) ] = 0x0151, /* L1D.REPL */
333 [ C(RESULT_ACCESS
) ] = 0x020b, /* MEM_INST_RETURED.STORES */
334 [ C(RESULT_MISS
) ] = 0x0251, /* L1D.M_REPL */
336 [ C(OP_PREFETCH
) ] = {
337 [ C(RESULT_ACCESS
) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
338 [ C(RESULT_MISS
) ] = 0x024e, /* L1D_PREFETCH.MISS */
343 [ C(RESULT_ACCESS
) ] = 0x0380, /* L1I.READS */
344 [ C(RESULT_MISS
) ] = 0x0280, /* L1I.MISSES */
347 [ C(RESULT_ACCESS
) ] = -1,
348 [ C(RESULT_MISS
) ] = -1,
350 [ C(OP_PREFETCH
) ] = {
351 [ C(RESULT_ACCESS
) ] = 0x0,
352 [ C(RESULT_MISS
) ] = 0x0,
357 [ C(RESULT_ACCESS
) ] = 0x0324, /* L2_RQSTS.LOADS */
358 [ C(RESULT_MISS
) ] = 0x0224, /* L2_RQSTS.LD_MISS */
361 [ C(RESULT_ACCESS
) ] = 0x0c24, /* L2_RQSTS.RFOS */
362 [ C(RESULT_MISS
) ] = 0x0824, /* L2_RQSTS.RFO_MISS */
364 [ C(OP_PREFETCH
) ] = {
365 [ C(RESULT_ACCESS
) ] = 0x4f2e, /* LLC Reference */
366 [ C(RESULT_MISS
) ] = 0x412e, /* LLC Misses */
371 [ C(RESULT_ACCESS
) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
372 [ C(RESULT_MISS
) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
375 [ C(RESULT_ACCESS
) ] = 0x020b, /* MEM_INST_RETURED.STORES */
376 [ C(RESULT_MISS
) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
378 [ C(OP_PREFETCH
) ] = {
379 [ C(RESULT_ACCESS
) ] = 0x0,
380 [ C(RESULT_MISS
) ] = 0x0,
385 [ C(RESULT_ACCESS
) ] = 0x01c0, /* INST_RETIRED.ANY_P */
386 [ C(RESULT_MISS
) ] = 0x0185, /* ITLB_MISSES.ANY */
389 [ C(RESULT_ACCESS
) ] = -1,
390 [ C(RESULT_MISS
) ] = -1,
392 [ C(OP_PREFETCH
) ] = {
393 [ C(RESULT_ACCESS
) ] = -1,
394 [ C(RESULT_MISS
) ] = -1,
399 [ C(RESULT_ACCESS
) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
400 [ C(RESULT_MISS
) ] = 0x03e8, /* BPU_CLEARS.ANY */
403 [ C(RESULT_ACCESS
) ] = -1,
404 [ C(RESULT_MISS
) ] = -1,
406 [ C(OP_PREFETCH
) ] = {
407 [ C(RESULT_ACCESS
) ] = -1,
408 [ C(RESULT_MISS
) ] = -1,
413 static __initconst u64 nehalem_hw_cache_event_ids
414 [PERF_COUNT_HW_CACHE_MAX
]
415 [PERF_COUNT_HW_CACHE_OP_MAX
]
416 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
420 [ C(RESULT_ACCESS
) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
421 [ C(RESULT_MISS
) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
424 [ C(RESULT_ACCESS
) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
425 [ C(RESULT_MISS
) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
427 [ C(OP_PREFETCH
) ] = {
428 [ C(RESULT_ACCESS
) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
429 [ C(RESULT_MISS
) ] = 0x024e, /* L1D_PREFETCH.MISS */
434 [ C(RESULT_ACCESS
) ] = 0x0380, /* L1I.READS */
435 [ C(RESULT_MISS
) ] = 0x0280, /* L1I.MISSES */
438 [ C(RESULT_ACCESS
) ] = -1,
439 [ C(RESULT_MISS
) ] = -1,
441 [ C(OP_PREFETCH
) ] = {
442 [ C(RESULT_ACCESS
) ] = 0x0,
443 [ C(RESULT_MISS
) ] = 0x0,
448 [ C(RESULT_ACCESS
) ] = 0x0324, /* L2_RQSTS.LOADS */
449 [ C(RESULT_MISS
) ] = 0x0224, /* L2_RQSTS.LD_MISS */
452 [ C(RESULT_ACCESS
) ] = 0x0c24, /* L2_RQSTS.RFOS */
453 [ C(RESULT_MISS
) ] = 0x0824, /* L2_RQSTS.RFO_MISS */
455 [ C(OP_PREFETCH
) ] = {
456 [ C(RESULT_ACCESS
) ] = 0x4f2e, /* LLC Reference */
457 [ C(RESULT_MISS
) ] = 0x412e, /* LLC Misses */
462 [ C(RESULT_ACCESS
) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
463 [ C(RESULT_MISS
) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
466 [ C(RESULT_ACCESS
) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
467 [ C(RESULT_MISS
) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
469 [ C(OP_PREFETCH
) ] = {
470 [ C(RESULT_ACCESS
) ] = 0x0,
471 [ C(RESULT_MISS
) ] = 0x0,
476 [ C(RESULT_ACCESS
) ] = 0x01c0, /* INST_RETIRED.ANY_P */
477 [ C(RESULT_MISS
) ] = 0x20c8, /* ITLB_MISS_RETIRED */
480 [ C(RESULT_ACCESS
) ] = -1,
481 [ C(RESULT_MISS
) ] = -1,
483 [ C(OP_PREFETCH
) ] = {
484 [ C(RESULT_ACCESS
) ] = -1,
485 [ C(RESULT_MISS
) ] = -1,
490 [ C(RESULT_ACCESS
) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
491 [ C(RESULT_MISS
) ] = 0x03e8, /* BPU_CLEARS.ANY */
494 [ C(RESULT_ACCESS
) ] = -1,
495 [ C(RESULT_MISS
) ] = -1,
497 [ C(OP_PREFETCH
) ] = {
498 [ C(RESULT_ACCESS
) ] = -1,
499 [ C(RESULT_MISS
) ] = -1,
504 static __initconst u64 core2_hw_cache_event_ids
505 [PERF_COUNT_HW_CACHE_MAX
]
506 [PERF_COUNT_HW_CACHE_OP_MAX
]
507 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
511 [ C(RESULT_ACCESS
) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
512 [ C(RESULT_MISS
) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
515 [ C(RESULT_ACCESS
) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
516 [ C(RESULT_MISS
) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
518 [ C(OP_PREFETCH
) ] = {
519 [ C(RESULT_ACCESS
) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */
520 [ C(RESULT_MISS
) ] = 0,
525 [ C(RESULT_ACCESS
) ] = 0x0080, /* L1I.READS */
526 [ C(RESULT_MISS
) ] = 0x0081, /* L1I.MISSES */
529 [ C(RESULT_ACCESS
) ] = -1,
530 [ C(RESULT_MISS
) ] = -1,
532 [ C(OP_PREFETCH
) ] = {
533 [ C(RESULT_ACCESS
) ] = 0,
534 [ C(RESULT_MISS
) ] = 0,
539 [ C(RESULT_ACCESS
) ] = 0x4f29, /* L2_LD.MESI */
540 [ C(RESULT_MISS
) ] = 0x4129, /* L2_LD.ISTATE */
543 [ C(RESULT_ACCESS
) ] = 0x4f2A, /* L2_ST.MESI */
544 [ C(RESULT_MISS
) ] = 0x412A, /* L2_ST.ISTATE */
546 [ C(OP_PREFETCH
) ] = {
547 [ C(RESULT_ACCESS
) ] = 0,
548 [ C(RESULT_MISS
) ] = 0,
553 [ C(RESULT_ACCESS
) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
554 [ C(RESULT_MISS
) ] = 0x0208, /* DTLB_MISSES.MISS_LD */
557 [ C(RESULT_ACCESS
) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
558 [ C(RESULT_MISS
) ] = 0x0808, /* DTLB_MISSES.MISS_ST */
560 [ C(OP_PREFETCH
) ] = {
561 [ C(RESULT_ACCESS
) ] = 0,
562 [ C(RESULT_MISS
) ] = 0,
567 [ C(RESULT_ACCESS
) ] = 0x00c0, /* INST_RETIRED.ANY_P */
568 [ C(RESULT_MISS
) ] = 0x1282, /* ITLBMISSES */
571 [ C(RESULT_ACCESS
) ] = -1,
572 [ C(RESULT_MISS
) ] = -1,
574 [ C(OP_PREFETCH
) ] = {
575 [ C(RESULT_ACCESS
) ] = -1,
576 [ C(RESULT_MISS
) ] = -1,
581 [ C(RESULT_ACCESS
) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
582 [ C(RESULT_MISS
) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
585 [ C(RESULT_ACCESS
) ] = -1,
586 [ C(RESULT_MISS
) ] = -1,
588 [ C(OP_PREFETCH
) ] = {
589 [ C(RESULT_ACCESS
) ] = -1,
590 [ C(RESULT_MISS
) ] = -1,
595 static __initconst u64 atom_hw_cache_event_ids
596 [PERF_COUNT_HW_CACHE_MAX
]
597 [PERF_COUNT_HW_CACHE_OP_MAX
]
598 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
602 [ C(RESULT_ACCESS
) ] = 0x2140, /* L1D_CACHE.LD */
603 [ C(RESULT_MISS
) ] = 0,
606 [ C(RESULT_ACCESS
) ] = 0x2240, /* L1D_CACHE.ST */
607 [ C(RESULT_MISS
) ] = 0,
609 [ C(OP_PREFETCH
) ] = {
610 [ C(RESULT_ACCESS
) ] = 0x0,
611 [ C(RESULT_MISS
) ] = 0,
616 [ C(RESULT_ACCESS
) ] = 0x0380, /* L1I.READS */
617 [ C(RESULT_MISS
) ] = 0x0280, /* L1I.MISSES */
620 [ C(RESULT_ACCESS
) ] = -1,
621 [ C(RESULT_MISS
) ] = -1,
623 [ C(OP_PREFETCH
) ] = {
624 [ C(RESULT_ACCESS
) ] = 0,
625 [ C(RESULT_MISS
) ] = 0,
630 [ C(RESULT_ACCESS
) ] = 0x4f29, /* L2_LD.MESI */
631 [ C(RESULT_MISS
) ] = 0x4129, /* L2_LD.ISTATE */
634 [ C(RESULT_ACCESS
) ] = 0x4f2A, /* L2_ST.MESI */
635 [ C(RESULT_MISS
) ] = 0x412A, /* L2_ST.ISTATE */
637 [ C(OP_PREFETCH
) ] = {
638 [ C(RESULT_ACCESS
) ] = 0,
639 [ C(RESULT_MISS
) ] = 0,
644 [ C(RESULT_ACCESS
) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */
645 [ C(RESULT_MISS
) ] = 0x0508, /* DTLB_MISSES.MISS_LD */
648 [ C(RESULT_ACCESS
) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */
649 [ C(RESULT_MISS
) ] = 0x0608, /* DTLB_MISSES.MISS_ST */
651 [ C(OP_PREFETCH
) ] = {
652 [ C(RESULT_ACCESS
) ] = 0,
653 [ C(RESULT_MISS
) ] = 0,
658 [ C(RESULT_ACCESS
) ] = 0x00c0, /* INST_RETIRED.ANY_P */
659 [ C(RESULT_MISS
) ] = 0x0282, /* ITLB.MISSES */
662 [ C(RESULT_ACCESS
) ] = -1,
663 [ C(RESULT_MISS
) ] = -1,
665 [ C(OP_PREFETCH
) ] = {
666 [ C(RESULT_ACCESS
) ] = -1,
667 [ C(RESULT_MISS
) ] = -1,
672 [ C(RESULT_ACCESS
) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
673 [ C(RESULT_MISS
) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
676 [ C(RESULT_ACCESS
) ] = -1,
677 [ C(RESULT_MISS
) ] = -1,
679 [ C(OP_PREFETCH
) ] = {
680 [ C(RESULT_ACCESS
) ] = -1,
681 [ C(RESULT_MISS
) ] = -1,
686 static u64
intel_pmu_raw_event(u64 hw_event
)
688 #define CORE_EVNTSEL_EVENT_MASK 0x000000FFULL
689 #define CORE_EVNTSEL_UNIT_MASK 0x0000FF00ULL
690 #define CORE_EVNTSEL_EDGE_MASK 0x00040000ULL
691 #define CORE_EVNTSEL_INV_MASK 0x00800000ULL
692 #define CORE_EVNTSEL_REG_MASK 0xFF000000ULL
694 #define CORE_EVNTSEL_MASK \
695 (INTEL_ARCH_EVTSEL_MASK | \
696 INTEL_ARCH_UNIT_MASK | \
697 INTEL_ARCH_EDGE_MASK | \
698 INTEL_ARCH_INV_MASK | \
701 return hw_event
& CORE_EVNTSEL_MASK
;
704 static __initconst u64 amd_hw_cache_event_ids
705 [PERF_COUNT_HW_CACHE_MAX
]
706 [PERF_COUNT_HW_CACHE_OP_MAX
]
707 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
711 [ C(RESULT_ACCESS
) ] = 0x0040, /* Data Cache Accesses */
712 [ C(RESULT_MISS
) ] = 0x0041, /* Data Cache Misses */
715 [ C(RESULT_ACCESS
) ] = 0x0142, /* Data Cache Refills :system */
716 [ C(RESULT_MISS
) ] = 0,
718 [ C(OP_PREFETCH
) ] = {
719 [ C(RESULT_ACCESS
) ] = 0x0267, /* Data Prefetcher :attempts */
720 [ C(RESULT_MISS
) ] = 0x0167, /* Data Prefetcher :cancelled */
725 [ C(RESULT_ACCESS
) ] = 0x0080, /* Instruction cache fetches */
726 [ C(RESULT_MISS
) ] = 0x0081, /* Instruction cache misses */
729 [ C(RESULT_ACCESS
) ] = -1,
730 [ C(RESULT_MISS
) ] = -1,
732 [ C(OP_PREFETCH
) ] = {
733 [ C(RESULT_ACCESS
) ] = 0x014B, /* Prefetch Instructions :Load */
734 [ C(RESULT_MISS
) ] = 0,
739 [ C(RESULT_ACCESS
) ] = 0x037D, /* Requests to L2 Cache :IC+DC */
740 [ C(RESULT_MISS
) ] = 0x037E, /* L2 Cache Misses : IC+DC */
743 [ C(RESULT_ACCESS
) ] = 0x017F, /* L2 Fill/Writeback */
744 [ C(RESULT_MISS
) ] = 0,
746 [ C(OP_PREFETCH
) ] = {
747 [ C(RESULT_ACCESS
) ] = 0,
748 [ C(RESULT_MISS
) ] = 0,
753 [ C(RESULT_ACCESS
) ] = 0x0040, /* Data Cache Accesses */
754 [ C(RESULT_MISS
) ] = 0x0046, /* L1 DTLB and L2 DLTB Miss */
757 [ C(RESULT_ACCESS
) ] = 0,
758 [ C(RESULT_MISS
) ] = 0,
760 [ C(OP_PREFETCH
) ] = {
761 [ C(RESULT_ACCESS
) ] = 0,
762 [ C(RESULT_MISS
) ] = 0,
767 [ C(RESULT_ACCESS
) ] = 0x0080, /* Instruction fecthes */
768 [ C(RESULT_MISS
) ] = 0x0085, /* Instr. fetch ITLB misses */
771 [ C(RESULT_ACCESS
) ] = -1,
772 [ C(RESULT_MISS
) ] = -1,
774 [ C(OP_PREFETCH
) ] = {
775 [ C(RESULT_ACCESS
) ] = -1,
776 [ C(RESULT_MISS
) ] = -1,
781 [ C(RESULT_ACCESS
) ] = 0x00c2, /* Retired Branch Instr. */
782 [ C(RESULT_MISS
) ] = 0x00c3, /* Retired Mispredicted BI */
785 [ C(RESULT_ACCESS
) ] = -1,
786 [ C(RESULT_MISS
) ] = -1,
788 [ C(OP_PREFETCH
) ] = {
789 [ C(RESULT_ACCESS
) ] = -1,
790 [ C(RESULT_MISS
) ] = -1,
796 * AMD Performance Monitor K7 and later.
798 static const u64 amd_perfmon_event_map
[] =
800 [PERF_COUNT_HW_CPU_CYCLES
] = 0x0076,
801 [PERF_COUNT_HW_INSTRUCTIONS
] = 0x00c0,
802 [PERF_COUNT_HW_CACHE_REFERENCES
] = 0x0080,
803 [PERF_COUNT_HW_CACHE_MISSES
] = 0x0081,
804 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS
] = 0x00c4,
805 [PERF_COUNT_HW_BRANCH_MISSES
] = 0x00c5,
808 static u64
amd_pmu_event_map(int hw_event
)
810 return amd_perfmon_event_map
[hw_event
];
813 static u64
amd_pmu_raw_event(u64 hw_event
)
815 #define K7_EVNTSEL_EVENT_MASK 0xF000000FFULL
816 #define K7_EVNTSEL_UNIT_MASK 0x00000FF00ULL
817 #define K7_EVNTSEL_EDGE_MASK 0x000040000ULL
818 #define K7_EVNTSEL_INV_MASK 0x000800000ULL
819 #define K7_EVNTSEL_REG_MASK 0x0FF000000ULL
821 #define K7_EVNTSEL_MASK \
822 (K7_EVNTSEL_EVENT_MASK | \
823 K7_EVNTSEL_UNIT_MASK | \
824 K7_EVNTSEL_EDGE_MASK | \
825 K7_EVNTSEL_INV_MASK | \
828 return hw_event
& K7_EVNTSEL_MASK
;
832 * Propagate event elapsed time into the generic event.
833 * Can only be executed on the CPU where the event is active.
834 * Returns the delta events processed.
837 x86_perf_event_update(struct perf_event
*event
,
838 struct hw_perf_event
*hwc
, int idx
)
840 int shift
= 64 - x86_pmu
.event_bits
;
841 u64 prev_raw_count
, new_raw_count
;
844 if (idx
== X86_PMC_IDX_FIXED_BTS
)
848 * Careful: an NMI might modify the previous event value.
850 * Our tactic to handle this is to first atomically read and
851 * exchange a new raw count - then add that new-prev delta
852 * count to the generic event atomically:
855 prev_raw_count
= atomic64_read(&hwc
->prev_count
);
856 rdmsrl(hwc
->event_base
+ idx
, new_raw_count
);
858 if (atomic64_cmpxchg(&hwc
->prev_count
, prev_raw_count
,
859 new_raw_count
) != prev_raw_count
)
863 * Now we have the new raw value and have updated the prev
864 * timestamp already. We can now calculate the elapsed delta
865 * (event-)time and add that to the generic event.
867 * Careful, not all hw sign-extends above the physical width
870 delta
= (new_raw_count
<< shift
) - (prev_raw_count
<< shift
);
873 atomic64_add(delta
, &event
->count
);
874 atomic64_sub(delta
, &hwc
->period_left
);
876 return new_raw_count
;
879 static atomic_t active_events
;
880 static DEFINE_MUTEX(pmc_reserve_mutex
);
882 static bool reserve_pmc_hardware(void)
884 #ifdef CONFIG_X86_LOCAL_APIC
887 if (nmi_watchdog
== NMI_LOCAL_APIC
)
888 disable_lapic_nmi_watchdog();
890 for (i
= 0; i
< x86_pmu
.num_events
; i
++) {
891 if (!reserve_perfctr_nmi(x86_pmu
.perfctr
+ i
))
895 for (i
= 0; i
< x86_pmu
.num_events
; i
++) {
896 if (!reserve_evntsel_nmi(x86_pmu
.eventsel
+ i
))
903 #ifdef CONFIG_X86_LOCAL_APIC
905 for (i
--; i
>= 0; i
--)
906 release_evntsel_nmi(x86_pmu
.eventsel
+ i
);
908 i
= x86_pmu
.num_events
;
911 for (i
--; i
>= 0; i
--)
912 release_perfctr_nmi(x86_pmu
.perfctr
+ i
);
914 if (nmi_watchdog
== NMI_LOCAL_APIC
)
915 enable_lapic_nmi_watchdog();
921 static void release_pmc_hardware(void)
923 #ifdef CONFIG_X86_LOCAL_APIC
926 for (i
= 0; i
< x86_pmu
.num_events
; i
++) {
927 release_perfctr_nmi(x86_pmu
.perfctr
+ i
);
928 release_evntsel_nmi(x86_pmu
.eventsel
+ i
);
931 if (nmi_watchdog
== NMI_LOCAL_APIC
)
932 enable_lapic_nmi_watchdog();
936 static inline bool bts_available(void)
938 return x86_pmu
.enable_bts
!= NULL
;
941 static inline void init_debug_store_on_cpu(int cpu
)
943 struct debug_store
*ds
= per_cpu(cpu_hw_events
, cpu
).ds
;
948 wrmsr_on_cpu(cpu
, MSR_IA32_DS_AREA
,
949 (u32
)((u64
)(unsigned long)ds
),
950 (u32
)((u64
)(unsigned long)ds
>> 32));
953 static inline void fini_debug_store_on_cpu(int cpu
)
955 if (!per_cpu(cpu_hw_events
, cpu
).ds
)
958 wrmsr_on_cpu(cpu
, MSR_IA32_DS_AREA
, 0, 0);
961 static void release_bts_hardware(void)
965 if (!bts_available())
970 for_each_online_cpu(cpu
)
971 fini_debug_store_on_cpu(cpu
);
973 for_each_possible_cpu(cpu
) {
974 struct debug_store
*ds
= per_cpu(cpu_hw_events
, cpu
).ds
;
979 per_cpu(cpu_hw_events
, cpu
).ds
= NULL
;
981 kfree((void *)(unsigned long)ds
->bts_buffer_base
);
988 static int reserve_bts_hardware(void)
992 if (!bts_available())
997 for_each_possible_cpu(cpu
) {
998 struct debug_store
*ds
;
1002 buffer
= kzalloc(BTS_BUFFER_SIZE
, GFP_KERNEL
);
1003 if (unlikely(!buffer
))
1006 ds
= kzalloc(sizeof(*ds
), GFP_KERNEL
);
1007 if (unlikely(!ds
)) {
1012 ds
->bts_buffer_base
= (u64
)(unsigned long)buffer
;
1013 ds
->bts_index
= ds
->bts_buffer_base
;
1014 ds
->bts_absolute_maximum
=
1015 ds
->bts_buffer_base
+ BTS_BUFFER_SIZE
;
1016 ds
->bts_interrupt_threshold
=
1017 ds
->bts_absolute_maximum
- BTS_OVFL_TH
;
1019 per_cpu(cpu_hw_events
, cpu
).ds
= ds
;
1024 release_bts_hardware();
1026 for_each_online_cpu(cpu
)
1027 init_debug_store_on_cpu(cpu
);
1035 static void hw_perf_event_destroy(struct perf_event
*event
)
1037 if (atomic_dec_and_mutex_lock(&active_events
, &pmc_reserve_mutex
)) {
1038 release_pmc_hardware();
1039 release_bts_hardware();
1040 mutex_unlock(&pmc_reserve_mutex
);
1044 static inline int x86_pmu_initialized(void)
1046 return x86_pmu
.handle_irq
!= NULL
;
1050 set_ext_hw_attr(struct hw_perf_event
*hwc
, struct perf_event_attr
*attr
)
1052 unsigned int cache_type
, cache_op
, cache_result
;
1055 config
= attr
->config
;
1057 cache_type
= (config
>> 0) & 0xff;
1058 if (cache_type
>= PERF_COUNT_HW_CACHE_MAX
)
1061 cache_op
= (config
>> 8) & 0xff;
1062 if (cache_op
>= PERF_COUNT_HW_CACHE_OP_MAX
)
1065 cache_result
= (config
>> 16) & 0xff;
1066 if (cache_result
>= PERF_COUNT_HW_CACHE_RESULT_MAX
)
1069 val
= hw_cache_event_ids
[cache_type
][cache_op
][cache_result
];
1082 static void intel_pmu_enable_bts(u64 config
)
1084 unsigned long debugctlmsr
;
1086 debugctlmsr
= get_debugctlmsr();
1088 debugctlmsr
|= X86_DEBUGCTL_TR
;
1089 debugctlmsr
|= X86_DEBUGCTL_BTS
;
1090 debugctlmsr
|= X86_DEBUGCTL_BTINT
;
1092 if (!(config
& ARCH_PERFMON_EVENTSEL_OS
))
1093 debugctlmsr
|= X86_DEBUGCTL_BTS_OFF_OS
;
1095 if (!(config
& ARCH_PERFMON_EVENTSEL_USR
))
1096 debugctlmsr
|= X86_DEBUGCTL_BTS_OFF_USR
;
1098 update_debugctlmsr(debugctlmsr
);
1101 static void intel_pmu_disable_bts(void)
1103 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1104 unsigned long debugctlmsr
;
1109 debugctlmsr
= get_debugctlmsr();
1112 ~(X86_DEBUGCTL_TR
| X86_DEBUGCTL_BTS
| X86_DEBUGCTL_BTINT
|
1113 X86_DEBUGCTL_BTS_OFF_OS
| X86_DEBUGCTL_BTS_OFF_USR
);
1115 update_debugctlmsr(debugctlmsr
);
1119 * Setup the hardware configuration for a given attr_type
1121 static int __hw_perf_event_init(struct perf_event
*event
)
1123 struct perf_event_attr
*attr
= &event
->attr
;
1124 struct hw_perf_event
*hwc
= &event
->hw
;
1128 if (!x86_pmu_initialized())
1132 if (!atomic_inc_not_zero(&active_events
)) {
1133 mutex_lock(&pmc_reserve_mutex
);
1134 if (atomic_read(&active_events
) == 0) {
1135 if (!reserve_pmc_hardware())
1138 err
= reserve_bts_hardware();
1141 atomic_inc(&active_events
);
1142 mutex_unlock(&pmc_reserve_mutex
);
1147 event
->destroy
= hw_perf_event_destroy
;
1150 * Generate PMC IRQs:
1151 * (keep 'enabled' bit clear for now)
1153 hwc
->config
= ARCH_PERFMON_EVENTSEL_INT
;
1157 hwc
->last_tag
= ~0ULL;
1160 * Count user and OS events unless requested not to.
1162 if (!attr
->exclude_user
)
1163 hwc
->config
|= ARCH_PERFMON_EVENTSEL_USR
;
1164 if (!attr
->exclude_kernel
)
1165 hwc
->config
|= ARCH_PERFMON_EVENTSEL_OS
;
1167 if (!hwc
->sample_period
) {
1168 hwc
->sample_period
= x86_pmu
.max_period
;
1169 hwc
->last_period
= hwc
->sample_period
;
1170 atomic64_set(&hwc
->period_left
, hwc
->sample_period
);
1173 * If we have a PMU initialized but no APIC
1174 * interrupts, we cannot sample hardware
1175 * events (user-space has to fall back and
1176 * sample via a hrtimer based software event):
1183 * Raw hw_event type provide the config in the hw_event structure
1185 if (attr
->type
== PERF_TYPE_RAW
) {
1186 hwc
->config
|= x86_pmu
.raw_event(attr
->config
);
1190 if (attr
->type
== PERF_TYPE_HW_CACHE
)
1191 return set_ext_hw_attr(hwc
, attr
);
1193 if (attr
->config
>= x86_pmu
.max_events
)
1199 config
= x86_pmu
.event_map(attr
->config
);
1210 if ((attr
->config
== PERF_COUNT_HW_BRANCH_INSTRUCTIONS
) &&
1211 (hwc
->sample_period
== 1)) {
1212 /* BTS is not supported by this architecture. */
1213 if (!bts_available())
1216 /* BTS is currently only allowed for user-mode. */
1217 if (hwc
->config
& ARCH_PERFMON_EVENTSEL_OS
)
1221 hwc
->config
|= config
;
1226 static void p6_pmu_disable_all(void)
1230 /* p6 only has one enable register */
1231 rdmsrl(MSR_P6_EVNTSEL0
, val
);
1232 val
&= ~ARCH_PERFMON_EVENTSEL0_ENABLE
;
1233 wrmsrl(MSR_P6_EVNTSEL0
, val
);
1236 static void intel_pmu_disable_all(void)
1238 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1240 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, 0);
1242 if (test_bit(X86_PMC_IDX_FIXED_BTS
, cpuc
->active_mask
))
1243 intel_pmu_disable_bts();
1246 static void x86_pmu_disable_all(void)
1248 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1251 for (idx
= 0; idx
< x86_pmu
.num_events
; idx
++) {
1254 if (!test_bit(idx
, cpuc
->active_mask
))
1256 rdmsrl(x86_pmu
.eventsel
+ idx
, val
);
1257 if (!(val
& ARCH_PERFMON_EVENTSEL0_ENABLE
))
1259 val
&= ~ARCH_PERFMON_EVENTSEL0_ENABLE
;
1260 wrmsrl(x86_pmu
.eventsel
+ idx
, val
);
1264 void hw_perf_disable(void)
1266 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1268 if (!x86_pmu_initialized())
1278 x86_pmu
.disable_all();
1281 static void p6_pmu_enable_all(void)
1285 /* p6 only has one enable register */
1286 rdmsrl(MSR_P6_EVNTSEL0
, val
);
1287 val
|= ARCH_PERFMON_EVENTSEL0_ENABLE
;
1288 wrmsrl(MSR_P6_EVNTSEL0
, val
);
1291 static void intel_pmu_enable_all(void)
1293 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1295 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, x86_pmu
.intel_ctrl
);
1297 if (test_bit(X86_PMC_IDX_FIXED_BTS
, cpuc
->active_mask
)) {
1298 struct perf_event
*event
=
1299 cpuc
->events
[X86_PMC_IDX_FIXED_BTS
];
1301 if (WARN_ON_ONCE(!event
))
1304 intel_pmu_enable_bts(event
->hw
.config
);
1308 static void x86_pmu_enable_all(void)
1310 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1313 for (idx
= 0; idx
< x86_pmu
.num_events
; idx
++) {
1314 struct perf_event
*event
= cpuc
->events
[idx
];
1317 if (!test_bit(idx
, cpuc
->active_mask
))
1320 val
= event
->hw
.config
;
1321 val
|= ARCH_PERFMON_EVENTSEL0_ENABLE
;
1322 wrmsrl(x86_pmu
.eventsel
+ idx
, val
);
1326 static const struct pmu pmu
;
1328 static inline int is_x86_event(struct perf_event
*event
)
1330 return event
->pmu
== &pmu
;
1333 static int x86_schedule_events(struct cpu_hw_events
*cpuc
, int n
, int *assign
)
1335 struct event_constraint
*c
, *constraints
[X86_PMC_IDX_MAX
];
1336 unsigned long used_mask
[BITS_TO_LONGS(X86_PMC_IDX_MAX
)];
1337 int i
, j
, w
, wmax
, num
= 0;
1338 struct hw_perf_event
*hwc
;
1340 bitmap_zero(used_mask
, X86_PMC_IDX_MAX
);
1342 for (i
= 0; i
< n
; i
++) {
1344 x86_pmu
.get_event_constraints(cpuc
, cpuc
->event_list
[i
]);
1348 * fastpath, try to reuse previous register
1350 for (i
= 0; i
< n
; i
++) {
1351 hwc
= &cpuc
->event_list
[i
]->hw
;
1354 /* never assigned */
1358 /* constraint still honored */
1359 if (!test_bit(hwc
->idx
, c
->idxmsk
))
1362 /* not already used */
1363 if (test_bit(hwc
->idx
, used_mask
))
1366 set_bit(hwc
->idx
, used_mask
);
1368 assign
[i
] = hwc
->idx
;
1377 bitmap_zero(used_mask
, X86_PMC_IDX_MAX
);
1380 * weight = number of possible counters
1382 * 1 = most constrained, only works on one counter
1383 * wmax = least constrained, works on any counter
1385 * assign events to counters starting with most
1386 * constrained events.
1388 wmax
= x86_pmu
.num_events
;
1391 * when fixed event counters are present,
1392 * wmax is incremented by 1 to account
1393 * for one more choice
1395 if (x86_pmu
.num_events_fixed
)
1398 for (w
= 1, num
= n
; num
&& w
<= wmax
; w
++) {
1399 /* for each event */
1400 for (i
= 0; num
&& i
< n
; i
++) {
1402 hwc
= &cpuc
->event_list
[i
]->hw
;
1407 for_each_bit(j
, c
->idxmsk
, X86_PMC_IDX_MAX
) {
1408 if (!test_bit(j
, used_mask
))
1412 if (j
== X86_PMC_IDX_MAX
)
1415 set_bit(j
, used_mask
);
1424 * scheduling failed or is just a simulation,
1425 * free resources if necessary
1427 if (!assign
|| num
) {
1428 for (i
= 0; i
< n
; i
++) {
1429 if (x86_pmu
.put_event_constraints
)
1430 x86_pmu
.put_event_constraints(cpuc
, cpuc
->event_list
[i
]);
1433 return num
? -ENOSPC
: 0;
1437 * dogrp: true if must collect siblings events (group)
1438 * returns total number of events and error code
1440 static int collect_events(struct cpu_hw_events
*cpuc
, struct perf_event
*leader
, bool dogrp
)
1442 struct perf_event
*event
;
1445 max_count
= x86_pmu
.num_events
+ x86_pmu
.num_events_fixed
;
1447 /* current number of events already accepted */
1450 if (is_x86_event(leader
)) {
1453 cpuc
->event_list
[n
] = leader
;
1459 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
) {
1460 if (!is_x86_event(event
) ||
1461 event
->state
<= PERF_EVENT_STATE_OFF
)
1467 cpuc
->event_list
[n
] = event
;
1473 static inline void x86_assign_hw_event(struct perf_event
*event
,
1474 struct cpu_hw_events
*cpuc
, int i
)
1476 struct hw_perf_event
*hwc
= &event
->hw
;
1478 hwc
->idx
= cpuc
->assign
[i
];
1479 hwc
->last_cpu
= smp_processor_id();
1480 hwc
->last_tag
= ++cpuc
->tags
[i
];
1482 if (hwc
->idx
== X86_PMC_IDX_FIXED_BTS
) {
1483 hwc
->config_base
= 0;
1484 hwc
->event_base
= 0;
1485 } else if (hwc
->idx
>= X86_PMC_IDX_FIXED
) {
1486 hwc
->config_base
= MSR_ARCH_PERFMON_FIXED_CTR_CTRL
;
1488 * We set it so that event_base + idx in wrmsr/rdmsr maps to
1489 * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
1492 MSR_ARCH_PERFMON_FIXED_CTR0
- X86_PMC_IDX_FIXED
;
1494 hwc
->config_base
= x86_pmu
.eventsel
;
1495 hwc
->event_base
= x86_pmu
.perfctr
;
1499 static inline int match_prev_assignment(struct hw_perf_event
*hwc
,
1500 struct cpu_hw_events
*cpuc
,
1503 return hwc
->idx
== cpuc
->assign
[i
] &&
1504 hwc
->last_cpu
== smp_processor_id() &&
1505 hwc
->last_tag
== cpuc
->tags
[i
];
1508 static void x86_pmu_stop(struct perf_event
*event
);
1510 void hw_perf_enable(void)
1512 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1513 struct perf_event
*event
;
1514 struct hw_perf_event
*hwc
;
1517 if (!x86_pmu_initialized())
1523 if (cpuc
->n_added
) {
1525 * apply assignment obtained either from
1526 * hw_perf_group_sched_in() or x86_pmu_enable()
1528 * step1: save events moving to new counters
1529 * step2: reprogram moved events into new counters
1531 for (i
= 0; i
< cpuc
->n_events
; i
++) {
1533 event
= cpuc
->event_list
[i
];
1537 * we can avoid reprogramming counter if:
1538 * - assigned same counter as last time
1539 * - running on same CPU as last time
1540 * - no other event has used the counter since
1542 if (hwc
->idx
== -1 ||
1543 match_prev_assignment(hwc
, cpuc
, i
))
1546 x86_pmu_stop(event
);
1551 for (i
= 0; i
< cpuc
->n_events
; i
++) {
1553 event
= cpuc
->event_list
[i
];
1556 if (hwc
->idx
== -1) {
1557 x86_assign_hw_event(event
, cpuc
, i
);
1558 x86_perf_event_set_period(event
, hwc
, hwc
->idx
);
1561 * need to mark as active because x86_pmu_disable()
1562 * clear active_mask and events[] yet it preserves
1565 set_bit(hwc
->idx
, cpuc
->active_mask
);
1566 cpuc
->events
[hwc
->idx
] = event
;
1568 x86_pmu
.enable(hwc
, hwc
->idx
);
1569 perf_event_update_userpage(event
);
1572 perf_events_lapic_init();
1578 x86_pmu
.enable_all();
1581 static inline u64
intel_pmu_get_status(void)
1585 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS
, status
);
1590 static inline void intel_pmu_ack_status(u64 ack
)
1592 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL
, ack
);
1595 static inline void __x86_pmu_enable_event(struct hw_perf_event
*hwc
, int idx
)
1597 (void)checking_wrmsrl(hwc
->config_base
+ idx
,
1598 hwc
->config
| ARCH_PERFMON_EVENTSEL0_ENABLE
);
1601 static inline void x86_pmu_disable_event(struct hw_perf_event
*hwc
, int idx
)
1603 (void)checking_wrmsrl(hwc
->config_base
+ idx
, hwc
->config
);
1607 intel_pmu_disable_fixed(struct hw_perf_event
*hwc
, int __idx
)
1609 int idx
= __idx
- X86_PMC_IDX_FIXED
;
1612 mask
= 0xfULL
<< (idx
* 4);
1614 rdmsrl(hwc
->config_base
, ctrl_val
);
1616 (void)checking_wrmsrl(hwc
->config_base
, ctrl_val
);
1620 p6_pmu_disable_event(struct hw_perf_event
*hwc
, int idx
)
1622 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1623 u64 val
= P6_NOP_EVENT
;
1626 val
|= ARCH_PERFMON_EVENTSEL0_ENABLE
;
1628 (void)checking_wrmsrl(hwc
->config_base
+ idx
, val
);
1632 intel_pmu_disable_event(struct hw_perf_event
*hwc
, int idx
)
1634 if (unlikely(idx
== X86_PMC_IDX_FIXED_BTS
)) {
1635 intel_pmu_disable_bts();
1639 if (unlikely(hwc
->config_base
== MSR_ARCH_PERFMON_FIXED_CTR_CTRL
)) {
1640 intel_pmu_disable_fixed(hwc
, idx
);
1644 x86_pmu_disable_event(hwc
, idx
);
1647 static DEFINE_PER_CPU(u64
[X86_PMC_IDX_MAX
], pmc_prev_left
);
1650 * Set the next IRQ period, based on the hwc->period_left value.
1651 * To be called with the event disabled in hw:
1654 x86_perf_event_set_period(struct perf_event
*event
,
1655 struct hw_perf_event
*hwc
, int idx
)
1657 s64 left
= atomic64_read(&hwc
->period_left
);
1658 s64 period
= hwc
->sample_period
;
1661 if (idx
== X86_PMC_IDX_FIXED_BTS
)
1665 * If we are way outside a reasonable range then just skip forward:
1667 if (unlikely(left
<= -period
)) {
1669 atomic64_set(&hwc
->period_left
, left
);
1670 hwc
->last_period
= period
;
1674 if (unlikely(left
<= 0)) {
1676 atomic64_set(&hwc
->period_left
, left
);
1677 hwc
->last_period
= period
;
1681 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1683 if (unlikely(left
< 2))
1686 if (left
> x86_pmu
.max_period
)
1687 left
= x86_pmu
.max_period
;
1689 per_cpu(pmc_prev_left
[idx
], smp_processor_id()) = left
;
1692 * The hw event starts counting from this event offset,
1693 * mark it to be able to extra future deltas:
1695 atomic64_set(&hwc
->prev_count
, (u64
)-left
);
1697 err
= checking_wrmsrl(hwc
->event_base
+ idx
,
1698 (u64
)(-left
) & x86_pmu
.event_mask
);
1700 perf_event_update_userpage(event
);
1706 intel_pmu_enable_fixed(struct hw_perf_event
*hwc
, int __idx
)
1708 int idx
= __idx
- X86_PMC_IDX_FIXED
;
1709 u64 ctrl_val
, bits
, mask
;
1713 * Enable IRQ generation (0x8),
1714 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
1718 if (hwc
->config
& ARCH_PERFMON_EVENTSEL_USR
)
1720 if (hwc
->config
& ARCH_PERFMON_EVENTSEL_OS
)
1724 * ANY bit is supported in v3 and up
1726 if (x86_pmu
.version
> 2 && hwc
->config
& ARCH_PERFMON_EVENTSEL_ANY
)
1730 mask
= 0xfULL
<< (idx
* 4);
1732 rdmsrl(hwc
->config_base
, ctrl_val
);
1735 err
= checking_wrmsrl(hwc
->config_base
, ctrl_val
);
1738 static void p6_pmu_enable_event(struct hw_perf_event
*hwc
, int idx
)
1740 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1745 val
|= ARCH_PERFMON_EVENTSEL0_ENABLE
;
1747 (void)checking_wrmsrl(hwc
->config_base
+ idx
, val
);
1751 static void intel_pmu_enable_event(struct hw_perf_event
*hwc
, int idx
)
1753 if (unlikely(idx
== X86_PMC_IDX_FIXED_BTS
)) {
1754 if (!__get_cpu_var(cpu_hw_events
).enabled
)
1757 intel_pmu_enable_bts(hwc
->config
);
1761 if (unlikely(hwc
->config_base
== MSR_ARCH_PERFMON_FIXED_CTR_CTRL
)) {
1762 intel_pmu_enable_fixed(hwc
, idx
);
1766 __x86_pmu_enable_event(hwc
, idx
);
1769 static void x86_pmu_enable_event(struct hw_perf_event
*hwc
, int idx
)
1771 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1773 __x86_pmu_enable_event(hwc
, idx
);
1777 * activate a single event
1779 * The event is added to the group of enabled events
1780 * but only if it can be scehduled with existing events.
1782 * Called with PMU disabled. If successful and return value 1,
1783 * then guaranteed to call perf_enable() and hw_perf_enable()
1785 static int x86_pmu_enable(struct perf_event
*event
)
1787 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1788 struct hw_perf_event
*hwc
;
1789 int assign
[X86_PMC_IDX_MAX
];
1794 n0
= cpuc
->n_events
;
1795 n
= collect_events(cpuc
, event
, false);
1799 ret
= x86_schedule_events(cpuc
, n
, assign
);
1803 * copy new assignment, now we know it is possible
1804 * will be used by hw_perf_enable()
1806 memcpy(cpuc
->assign
, assign
, n
*sizeof(int));
1809 cpuc
->n_added
= n
- n0
;
1814 static int x86_pmu_start(struct perf_event
*event
)
1816 struct hw_perf_event
*hwc
= &event
->hw
;
1821 x86_perf_event_set_period(event
, hwc
, hwc
->idx
);
1822 x86_pmu
.enable(hwc
, hwc
->idx
);
1827 static void x86_pmu_unthrottle(struct perf_event
*event
)
1829 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1830 struct hw_perf_event
*hwc
= &event
->hw
;
1832 if (WARN_ON_ONCE(hwc
->idx
>= X86_PMC_IDX_MAX
||
1833 cpuc
->events
[hwc
->idx
] != event
))
1836 x86_pmu
.enable(hwc
, hwc
->idx
);
1839 void perf_event_print_debug(void)
1841 u64 ctrl
, status
, overflow
, pmc_ctrl
, pmc_count
, prev_left
, fixed
;
1842 struct cpu_hw_events
*cpuc
;
1843 unsigned long flags
;
1846 if (!x86_pmu
.num_events
)
1849 local_irq_save(flags
);
1851 cpu
= smp_processor_id();
1852 cpuc
= &per_cpu(cpu_hw_events
, cpu
);
1854 if (x86_pmu
.version
>= 2) {
1855 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, ctrl
);
1856 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS
, status
);
1857 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL
, overflow
);
1858 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL
, fixed
);
1861 pr_info("CPU#%d: ctrl: %016llx\n", cpu
, ctrl
);
1862 pr_info("CPU#%d: status: %016llx\n", cpu
, status
);
1863 pr_info("CPU#%d: overflow: %016llx\n", cpu
, overflow
);
1864 pr_info("CPU#%d: fixed: %016llx\n", cpu
, fixed
);
1866 pr_info("CPU#%d: active: %016llx\n", cpu
, *(u64
*)cpuc
->active_mask
);
1868 for (idx
= 0; idx
< x86_pmu
.num_events
; idx
++) {
1869 rdmsrl(x86_pmu
.eventsel
+ idx
, pmc_ctrl
);
1870 rdmsrl(x86_pmu
.perfctr
+ idx
, pmc_count
);
1872 prev_left
= per_cpu(pmc_prev_left
[idx
], cpu
);
1874 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1875 cpu
, idx
, pmc_ctrl
);
1876 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1877 cpu
, idx
, pmc_count
);
1878 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1879 cpu
, idx
, prev_left
);
1881 for (idx
= 0; idx
< x86_pmu
.num_events_fixed
; idx
++) {
1882 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0
+ idx
, pmc_count
);
1884 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1885 cpu
, idx
, pmc_count
);
1887 local_irq_restore(flags
);
1890 static void intel_pmu_drain_bts_buffer(struct cpu_hw_events
*cpuc
)
1892 struct debug_store
*ds
= cpuc
->ds
;
1898 struct perf_event
*event
= cpuc
->events
[X86_PMC_IDX_FIXED_BTS
];
1899 struct bts_record
*at
, *top
;
1900 struct perf_output_handle handle
;
1901 struct perf_event_header header
;
1902 struct perf_sample_data data
;
1903 struct pt_regs regs
;
1911 at
= (struct bts_record
*)(unsigned long)ds
->bts_buffer_base
;
1912 top
= (struct bts_record
*)(unsigned long)ds
->bts_index
;
1917 ds
->bts_index
= ds
->bts_buffer_base
;
1920 data
.period
= event
->hw
.last_period
;
1926 * Prepare a generic sample, i.e. fill in the invariant fields.
1927 * We will overwrite the from and to address before we output
1930 perf_prepare_sample(&header
, &data
, event
, ®s
);
1932 if (perf_output_begin(&handle
, event
,
1933 header
.size
* (top
- at
), 1, 1))
1936 for (; at
< top
; at
++) {
1940 perf_output_sample(&handle
, &header
, &data
, event
);
1943 perf_output_end(&handle
);
1945 /* There's new data available. */
1946 event
->hw
.interrupts
++;
1947 event
->pending_kill
= POLL_IN
;
1950 static void x86_pmu_stop(struct perf_event
*event
)
1952 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1953 struct hw_perf_event
*hwc
= &event
->hw
;
1957 * Must be done before we disable, otherwise the nmi handler
1958 * could reenable again:
1960 clear_bit(idx
, cpuc
->active_mask
);
1961 x86_pmu
.disable(hwc
, idx
);
1964 * Drain the remaining delta count out of a event
1965 * that we are disabling:
1967 x86_perf_event_update(event
, hwc
, idx
);
1969 /* Drain the remaining BTS records. */
1970 if (unlikely(idx
== X86_PMC_IDX_FIXED_BTS
))
1971 intel_pmu_drain_bts_buffer(cpuc
);
1973 cpuc
->events
[idx
] = NULL
;
1976 static void x86_pmu_disable(struct perf_event
*event
)
1978 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1981 x86_pmu_stop(event
);
1983 for (i
= 0; i
< cpuc
->n_events
; i
++) {
1984 if (event
== cpuc
->event_list
[i
]) {
1986 if (x86_pmu
.put_event_constraints
)
1987 x86_pmu
.put_event_constraints(cpuc
, event
);
1989 while (++i
< cpuc
->n_events
)
1990 cpuc
->event_list
[i
-1] = cpuc
->event_list
[i
];
1996 perf_event_update_userpage(event
);
2000 * Save and restart an expired event. Called by NMI contexts,
2001 * so it has to be careful about preempting normal event ops:
2003 static int intel_pmu_save_and_restart(struct perf_event
*event
)
2005 struct hw_perf_event
*hwc
= &event
->hw
;
2009 x86_perf_event_update(event
, hwc
, idx
);
2010 ret
= x86_perf_event_set_period(event
, hwc
, idx
);
2012 if (event
->state
== PERF_EVENT_STATE_ACTIVE
)
2013 intel_pmu_enable_event(hwc
, idx
);
2018 static void intel_pmu_reset(void)
2020 struct debug_store
*ds
= __get_cpu_var(cpu_hw_events
).ds
;
2021 unsigned long flags
;
2024 if (!x86_pmu
.num_events
)
2027 local_irq_save(flags
);
2029 printk("clearing PMU state on CPU#%d\n", smp_processor_id());
2031 for (idx
= 0; idx
< x86_pmu
.num_events
; idx
++) {
2032 checking_wrmsrl(x86_pmu
.eventsel
+ idx
, 0ull);
2033 checking_wrmsrl(x86_pmu
.perfctr
+ idx
, 0ull);
2035 for (idx
= 0; idx
< x86_pmu
.num_events_fixed
; idx
++) {
2036 checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0
+ idx
, 0ull);
2039 ds
->bts_index
= ds
->bts_buffer_base
;
2041 local_irq_restore(flags
);
2045 * This handler is triggered by the local APIC, so the APIC IRQ handling
2048 static int intel_pmu_handle_irq(struct pt_regs
*regs
)
2050 struct perf_sample_data data
;
2051 struct cpu_hw_events
*cpuc
;
2058 cpuc
= &__get_cpu_var(cpu_hw_events
);
2061 intel_pmu_drain_bts_buffer(cpuc
);
2062 status
= intel_pmu_get_status();
2070 if (++loops
> 100) {
2071 WARN_ONCE(1, "perfevents: irq loop stuck!\n");
2072 perf_event_print_debug();
2078 inc_irq_stat(apic_perf_irqs
);
2080 for_each_bit(bit
, (unsigned long *)&status
, X86_PMC_IDX_MAX
) {
2081 struct perf_event
*event
= cpuc
->events
[bit
];
2083 clear_bit(bit
, (unsigned long *) &status
);
2084 if (!test_bit(bit
, cpuc
->active_mask
))
2087 if (!intel_pmu_save_and_restart(event
))
2090 data
.period
= event
->hw
.last_period
;
2092 if (perf_event_overflow(event
, 1, &data
, regs
))
2093 intel_pmu_disable_event(&event
->hw
, bit
);
2096 intel_pmu_ack_status(ack
);
2099 * Repeat if there is more work to be done:
2101 status
= intel_pmu_get_status();
2110 static int x86_pmu_handle_irq(struct pt_regs
*regs
)
2112 struct perf_sample_data data
;
2113 struct cpu_hw_events
*cpuc
;
2114 struct perf_event
*event
;
2115 struct hw_perf_event
*hwc
;
2116 int idx
, handled
= 0;
2122 cpuc
= &__get_cpu_var(cpu_hw_events
);
2124 for (idx
= 0; idx
< x86_pmu
.num_events
; idx
++) {
2125 if (!test_bit(idx
, cpuc
->active_mask
))
2128 event
= cpuc
->events
[idx
];
2131 val
= x86_perf_event_update(event
, hwc
, idx
);
2132 if (val
& (1ULL << (x86_pmu
.event_bits
- 1)))
2139 data
.period
= event
->hw
.last_period
;
2141 if (!x86_perf_event_set_period(event
, hwc
, idx
))
2144 if (perf_event_overflow(event
, 1, &data
, regs
))
2145 x86_pmu
.disable(hwc
, idx
);
2149 inc_irq_stat(apic_perf_irqs
);
2154 void smp_perf_pending_interrupt(struct pt_regs
*regs
)
2158 inc_irq_stat(apic_pending_irqs
);
2159 perf_event_do_pending();
2163 void set_perf_event_pending(void)
2165 #ifdef CONFIG_X86_LOCAL_APIC
2166 if (!x86_pmu
.apic
|| !x86_pmu_initialized())
2169 apic
->send_IPI_self(LOCAL_PENDING_VECTOR
);
2173 void perf_events_lapic_init(void)
2175 #ifdef CONFIG_X86_LOCAL_APIC
2176 if (!x86_pmu
.apic
|| !x86_pmu_initialized())
2180 * Always use NMI for PMU
2182 apic_write(APIC_LVTPC
, APIC_DM_NMI
);
2186 static int __kprobes
2187 perf_event_nmi_handler(struct notifier_block
*self
,
2188 unsigned long cmd
, void *__args
)
2190 struct die_args
*args
= __args
;
2191 struct pt_regs
*regs
;
2193 if (!atomic_read(&active_events
))
2207 #ifdef CONFIG_X86_LOCAL_APIC
2208 apic_write(APIC_LVTPC
, APIC_DM_NMI
);
2211 * Can't rely on the handled return value to say it was our NMI, two
2212 * events could trigger 'simultaneously' raising two back-to-back NMIs.
2214 * If the first NMI handles both, the latter will be empty and daze
2217 x86_pmu
.handle_irq(regs
);
2222 static struct event_constraint unconstrained
;
2223 static struct event_constraint emptyconstraint
;
2225 static struct event_constraint bts_constraint
=
2226 EVENT_CONSTRAINT(0, 1ULL << X86_PMC_IDX_FIXED_BTS
, 0);
2228 static struct event_constraint
*
2229 intel_special_constraints(struct perf_event
*event
)
2231 unsigned int hw_event
;
2233 hw_event
= event
->hw
.config
& INTEL_ARCH_EVENT_MASK
;
2235 if (unlikely((hw_event
==
2236 x86_pmu
.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS
)) &&
2237 (event
->hw
.sample_period
== 1))) {
2239 return &bts_constraint
;
2244 static struct event_constraint
*
2245 intel_get_event_constraints(struct cpu_hw_events
*cpuc
, struct perf_event
*event
)
2247 struct event_constraint
*c
;
2249 c
= intel_special_constraints(event
);
2253 if (x86_pmu
.event_constraints
) {
2254 for_each_event_constraint(c
, x86_pmu
.event_constraints
) {
2255 if ((event
->hw
.config
& c
->cmask
) == c
->code
)
2260 return &unconstrained
;
2264 * AMD64 events are detected based on their event codes.
2266 static inline int amd_is_nb_event(struct hw_perf_event
*hwc
)
2268 return (hwc
->config
& 0xe0) == 0xe0;
2271 static void amd_put_event_constraints(struct cpu_hw_events
*cpuc
,
2272 struct perf_event
*event
)
2274 struct hw_perf_event
*hwc
= &event
->hw
;
2275 struct amd_nb
*nb
= cpuc
->amd_nb
;
2279 * only care about NB events
2281 if (!(nb
&& amd_is_nb_event(hwc
)))
2285 * need to scan whole list because event may not have
2286 * been assigned during scheduling
2288 * no race condition possible because event can only
2289 * be removed on one CPU at a time AND PMU is disabled
2292 for (i
= 0; i
< x86_pmu
.num_events
; i
++) {
2293 if (nb
->owners
[i
] == event
) {
2294 cmpxchg(nb
->owners
+i
, event
, NULL
);
2301 * AMD64 NorthBridge events need special treatment because
2302 * counter access needs to be synchronized across all cores
2303 * of a package. Refer to BKDG section 3.12
2305 * NB events are events measuring L3 cache, Hypertransport
2306 * traffic. They are identified by an event code >= 0xe00.
2307 * They measure events on the NorthBride which is shared
2308 * by all cores on a package. NB events are counted on a
2309 * shared set of counters. When a NB event is programmed
2310 * in a counter, the data actually comes from a shared
2311 * counter. Thus, access to those counters needs to be
2314 * We implement the synchronization such that no two cores
2315 * can be measuring NB events using the same counters. Thus,
2316 * we maintain a per-NB allocation table. The available slot
2317 * is propagated using the event_constraint structure.
2319 * We provide only one choice for each NB event based on
2320 * the fact that only NB events have restrictions. Consequently,
2321 * if a counter is available, there is a guarantee the NB event
2322 * will be assigned to it. If no slot is available, an empty
2323 * constraint is returned and scheduling will eventually fail
2326 * Note that all cores attached the same NB compete for the same
2327 * counters to host NB events, this is why we use atomic ops. Some
2328 * multi-chip CPUs may have more than one NB.
2330 * Given that resources are allocated (cmpxchg), they must be
2331 * eventually freed for others to use. This is accomplished by
2332 * calling amd_put_event_constraints().
2334 * Non NB events are not impacted by this restriction.
2336 static struct event_constraint
*
2337 amd_get_event_constraints(struct cpu_hw_events
*cpuc
, struct perf_event
*event
)
2339 struct hw_perf_event
*hwc
= &event
->hw
;
2340 struct amd_nb
*nb
= cpuc
->amd_nb
;
2341 struct perf_event
*old
= NULL
;
2342 int max
= x86_pmu
.num_events
;
2346 * if not NB event or no NB, then no constraints
2348 if (!(nb
&& amd_is_nb_event(hwc
)))
2349 return &unconstrained
;
2352 * detect if already present, if so reuse
2354 * cannot merge with actual allocation
2355 * because of possible holes
2357 * event can already be present yet not assigned (in hwc->idx)
2358 * because of successive calls to x86_schedule_events() from
2359 * hw_perf_group_sched_in() without hw_perf_enable()
2361 for (i
= 0; i
< max
; i
++) {
2363 * keep track of first free slot
2365 if (k
== -1 && !nb
->owners
[i
])
2368 /* already present, reuse */
2369 if (nb
->owners
[i
] == event
)
2373 * not present, so grab a new slot
2374 * starting either at:
2376 if (hwc
->idx
!= -1) {
2377 /* previous assignment */
2379 } else if (k
!= -1) {
2380 /* start from free slot found */
2384 * event not found, no slot found in
2385 * first pass, try again from the
2392 old
= cmpxchg(nb
->owners
+i
, NULL
, event
);
2400 return &nb
->event_constraints
[i
];
2402 return &emptyconstraint
;
2405 static int x86_event_sched_in(struct perf_event
*event
,
2406 struct perf_cpu_context
*cpuctx
)
2410 event
->state
= PERF_EVENT_STATE_ACTIVE
;
2411 event
->oncpu
= smp_processor_id();
2412 event
->tstamp_running
+= event
->ctx
->time
- event
->tstamp_stopped
;
2414 if (!is_x86_event(event
))
2415 ret
= event
->pmu
->enable(event
);
2417 if (!ret
&& !is_software_event(event
))
2418 cpuctx
->active_oncpu
++;
2420 if (!ret
&& event
->attr
.exclusive
)
2421 cpuctx
->exclusive
= 1;
2426 static void x86_event_sched_out(struct perf_event
*event
,
2427 struct perf_cpu_context
*cpuctx
)
2429 event
->state
= PERF_EVENT_STATE_INACTIVE
;
2432 if (!is_x86_event(event
))
2433 event
->pmu
->disable(event
);
2435 event
->tstamp_running
-= event
->ctx
->time
- event
->tstamp_stopped
;
2437 if (!is_software_event(event
))
2438 cpuctx
->active_oncpu
--;
2440 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
2441 cpuctx
->exclusive
= 0;
2445 * Called to enable a whole group of events.
2446 * Returns 1 if the group was enabled, or -EAGAIN if it could not be.
2447 * Assumes the caller has disabled interrupts and has
2448 * frozen the PMU with hw_perf_save_disable.
2450 * called with PMU disabled. If successful and return value 1,
2451 * then guaranteed to call perf_enable() and hw_perf_enable()
2453 int hw_perf_group_sched_in(struct perf_event
*leader
,
2454 struct perf_cpu_context
*cpuctx
,
2455 struct perf_event_context
*ctx
)
2457 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
2458 struct perf_event
*sub
;
2459 int assign
[X86_PMC_IDX_MAX
];
2462 /* n0 = total number of events */
2463 n0
= collect_events(cpuc
, leader
, true);
2467 ret
= x86_schedule_events(cpuc
, n0
, assign
);
2471 ret
= x86_event_sched_in(leader
, cpuctx
);
2476 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2477 if (sub
->state
> PERF_EVENT_STATE_OFF
) {
2478 ret
= x86_event_sched_in(sub
, cpuctx
);
2485 * copy new assignment, now we know it is possible
2486 * will be used by hw_perf_enable()
2488 memcpy(cpuc
->assign
, assign
, n0
*sizeof(int));
2490 cpuc
->n_events
= n0
;
2492 ctx
->nr_active
+= n1
;
2495 * 1 means successful and events are active
2496 * This is not quite true because we defer
2497 * actual activation until hw_perf_enable() but
2498 * this way we* ensure caller won't try to enable
2503 x86_event_sched_out(leader
, cpuctx
);
2505 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
2506 if (sub
->state
== PERF_EVENT_STATE_ACTIVE
) {
2507 x86_event_sched_out(sub
, cpuctx
);
2515 static __read_mostly
struct notifier_block perf_event_nmi_notifier
= {
2516 .notifier_call
= perf_event_nmi_handler
,
2521 static __initconst
struct x86_pmu p6_pmu
= {
2523 .handle_irq
= x86_pmu_handle_irq
,
2524 .disable_all
= p6_pmu_disable_all
,
2525 .enable_all
= p6_pmu_enable_all
,
2526 .enable
= p6_pmu_enable_event
,
2527 .disable
= p6_pmu_disable_event
,
2528 .eventsel
= MSR_P6_EVNTSEL0
,
2529 .perfctr
= MSR_P6_PERFCTR0
,
2530 .event_map
= p6_pmu_event_map
,
2531 .raw_event
= p6_pmu_raw_event
,
2532 .max_events
= ARRAY_SIZE(p6_perfmon_event_map
),
2534 .max_period
= (1ULL << 31) - 1,
2538 * Events have 40 bits implemented. However they are designed such
2539 * that bits [32-39] are sign extensions of bit 31. As such the
2540 * effective width of a event for P6-like PMU is 32 bits only.
2542 * See IA-32 Intel Architecture Software developer manual Vol 3B
2545 .event_mask
= (1ULL << 32) - 1,
2546 .get_event_constraints
= intel_get_event_constraints
,
2547 .event_constraints
= intel_p6_event_constraints
2550 static __initconst
struct x86_pmu core_pmu
= {
2552 .handle_irq
= x86_pmu_handle_irq
,
2553 .disable_all
= x86_pmu_disable_all
,
2554 .enable_all
= x86_pmu_enable_all
,
2555 .enable
= x86_pmu_enable_event
,
2556 .disable
= x86_pmu_disable_event
,
2557 .eventsel
= MSR_ARCH_PERFMON_EVENTSEL0
,
2558 .perfctr
= MSR_ARCH_PERFMON_PERFCTR0
,
2559 .event_map
= intel_pmu_event_map
,
2560 .raw_event
= intel_pmu_raw_event
,
2561 .max_events
= ARRAY_SIZE(intel_perfmon_event_map
),
2564 * Intel PMCs cannot be accessed sanely above 32 bit width,
2565 * so we install an artificial 1<<31 period regardless of
2566 * the generic event period:
2568 .max_period
= (1ULL << 31) - 1,
2569 .get_event_constraints
= intel_get_event_constraints
,
2570 .event_constraints
= intel_core_event_constraints
,
2573 static __initconst
struct x86_pmu intel_pmu
= {
2575 .handle_irq
= intel_pmu_handle_irq
,
2576 .disable_all
= intel_pmu_disable_all
,
2577 .enable_all
= intel_pmu_enable_all
,
2578 .enable
= intel_pmu_enable_event
,
2579 .disable
= intel_pmu_disable_event
,
2580 .eventsel
= MSR_ARCH_PERFMON_EVENTSEL0
,
2581 .perfctr
= MSR_ARCH_PERFMON_PERFCTR0
,
2582 .event_map
= intel_pmu_event_map
,
2583 .raw_event
= intel_pmu_raw_event
,
2584 .max_events
= ARRAY_SIZE(intel_perfmon_event_map
),
2587 * Intel PMCs cannot be accessed sanely above 32 bit width,
2588 * so we install an artificial 1<<31 period regardless of
2589 * the generic event period:
2591 .max_period
= (1ULL << 31) - 1,
2592 .enable_bts
= intel_pmu_enable_bts
,
2593 .disable_bts
= intel_pmu_disable_bts
,
2594 .get_event_constraints
= intel_get_event_constraints
2597 static __initconst
struct x86_pmu amd_pmu
= {
2599 .handle_irq
= x86_pmu_handle_irq
,
2600 .disable_all
= x86_pmu_disable_all
,
2601 .enable_all
= x86_pmu_enable_all
,
2602 .enable
= x86_pmu_enable_event
,
2603 .disable
= x86_pmu_disable_event
,
2604 .eventsel
= MSR_K7_EVNTSEL0
,
2605 .perfctr
= MSR_K7_PERFCTR0
,
2606 .event_map
= amd_pmu_event_map
,
2607 .raw_event
= amd_pmu_raw_event
,
2608 .max_events
= ARRAY_SIZE(amd_perfmon_event_map
),
2611 .event_mask
= (1ULL << 48) - 1,
2613 /* use highest bit to detect overflow */
2614 .max_period
= (1ULL << 47) - 1,
2615 .get_event_constraints
= amd_get_event_constraints
,
2616 .put_event_constraints
= amd_put_event_constraints
2619 static __init
int p6_pmu_init(void)
2621 switch (boot_cpu_data
.x86_model
) {
2623 case 3: /* Pentium Pro */
2625 case 6: /* Pentium II */
2628 case 11: /* Pentium III */
2634 pr_cont("unsupported p6 CPU model %d ",
2635 boot_cpu_data
.x86_model
);
2644 static __init
int intel_pmu_init(void)
2646 union cpuid10_edx edx
;
2647 union cpuid10_eax eax
;
2648 unsigned int unused
;
2652 if (!cpu_has(&boot_cpu_data
, X86_FEATURE_ARCH_PERFMON
)) {
2653 /* check for P6 processor family */
2654 if (boot_cpu_data
.x86
== 6) {
2655 return p6_pmu_init();
2662 * Check whether the Architectural PerfMon supports
2663 * Branch Misses Retired hw_event or not.
2665 cpuid(10, &eax
.full
, &ebx
, &unused
, &edx
.full
);
2666 if (eax
.split
.mask_length
<= ARCH_PERFMON_BRANCH_MISSES_RETIRED
)
2669 version
= eax
.split
.version_id
;
2673 x86_pmu
= intel_pmu
;
2675 x86_pmu
.version
= version
;
2676 x86_pmu
.num_events
= eax
.split
.num_events
;
2677 x86_pmu
.event_bits
= eax
.split
.bit_width
;
2678 x86_pmu
.event_mask
= (1ULL << eax
.split
.bit_width
) - 1;
2681 * Quirk: v2 perfmon does not report fixed-purpose events, so
2682 * assume at least 3 events:
2685 x86_pmu
.num_events_fixed
= max((int)edx
.split
.num_events_fixed
, 3);
2688 * Install the hw-cache-events table:
2690 switch (boot_cpu_data
.x86_model
) {
2691 case 14: /* 65 nm core solo/duo, "Yonah" */
2692 pr_cont("Core events, ");
2695 case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
2696 case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */
2697 case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */
2698 case 29: /* six-core 45 nm xeon "Dunnington" */
2699 memcpy(hw_cache_event_ids
, core2_hw_cache_event_ids
,
2700 sizeof(hw_cache_event_ids
));
2702 x86_pmu
.event_constraints
= intel_core2_event_constraints
;
2703 pr_cont("Core2 events, ");
2706 case 26: /* 45 nm nehalem, "Bloomfield" */
2707 case 30: /* 45 nm nehalem, "Lynnfield" */
2708 memcpy(hw_cache_event_ids
, nehalem_hw_cache_event_ids
,
2709 sizeof(hw_cache_event_ids
));
2711 x86_pmu
.event_constraints
= intel_nehalem_event_constraints
;
2712 pr_cont("Nehalem/Corei7 events, ");
2715 memcpy(hw_cache_event_ids
, atom_hw_cache_event_ids
,
2716 sizeof(hw_cache_event_ids
));
2718 x86_pmu
.event_constraints
= intel_gen_event_constraints
;
2719 pr_cont("Atom events, ");
2722 case 37: /* 32 nm nehalem, "Clarkdale" */
2723 case 44: /* 32 nm nehalem, "Gulftown" */
2724 memcpy(hw_cache_event_ids
, westmere_hw_cache_event_ids
,
2725 sizeof(hw_cache_event_ids
));
2727 x86_pmu
.event_constraints
= intel_westmere_event_constraints
;
2728 pr_cont("Westmere events, ");
2732 * default constraints for v2 and up
2734 x86_pmu
.event_constraints
= intel_gen_event_constraints
;
2735 pr_cont("generic architected perfmon, ");
2740 static struct amd_nb
*amd_alloc_nb(int cpu
, int nb_id
)
2745 nb
= kmalloc(sizeof(struct amd_nb
), GFP_KERNEL
);
2749 memset(nb
, 0, sizeof(*nb
));
2753 * initialize all possible NB constraints
2755 for (i
= 0; i
< x86_pmu
.num_events
; i
++) {
2756 set_bit(i
, nb
->event_constraints
[i
].idxmsk
);
2757 nb
->event_constraints
[i
].weight
= 1;
2762 static void amd_pmu_cpu_online(int cpu
)
2764 struct cpu_hw_events
*cpu1
, *cpu2
;
2765 struct amd_nb
*nb
= NULL
;
2768 if (boot_cpu_data
.x86_max_cores
< 2)
2772 * function may be called too early in the
2773 * boot process, in which case nb_id is bogus
2775 nb_id
= amd_get_nb_id(cpu
);
2776 if (nb_id
== BAD_APICID
)
2779 cpu1
= &per_cpu(cpu_hw_events
, cpu
);
2780 cpu1
->amd_nb
= NULL
;
2782 raw_spin_lock(&amd_nb_lock
);
2784 for_each_online_cpu(i
) {
2785 cpu2
= &per_cpu(cpu_hw_events
, i
);
2789 if (nb
->nb_id
== nb_id
)
2793 nb
= amd_alloc_nb(cpu
, nb_id
);
2795 pr_err("perf_events: failed NB allocation for CPU%d\n", cpu
);
2796 raw_spin_unlock(&amd_nb_lock
);
2803 raw_spin_unlock(&amd_nb_lock
);
2806 static void amd_pmu_cpu_offline(int cpu
)
2808 struct cpu_hw_events
*cpuhw
;
2810 if (boot_cpu_data
.x86_max_cores
< 2)
2813 cpuhw
= &per_cpu(cpu_hw_events
, cpu
);
2815 raw_spin_lock(&amd_nb_lock
);
2817 if (--cpuhw
->amd_nb
->refcnt
== 0)
2818 kfree(cpuhw
->amd_nb
);
2820 cpuhw
->amd_nb
= NULL
;
2822 raw_spin_unlock(&amd_nb_lock
);
2825 static __init
int amd_pmu_init(void)
2827 /* Performance-monitoring supported from K7 and later: */
2828 if (boot_cpu_data
.x86
< 6)
2833 /* Events are common for all AMDs */
2834 memcpy(hw_cache_event_ids
, amd_hw_cache_event_ids
,
2835 sizeof(hw_cache_event_ids
));
2838 * explicitly initialize the boot cpu, other cpus will get
2839 * the cpu hotplug callbacks from smp_init()
2841 amd_pmu_cpu_online(smp_processor_id());
2845 static void __init
pmu_check_apic(void)
2851 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
2852 pr_info("no hardware sampling interrupt available.\n");
2855 void __init
init_hw_perf_events(void)
2859 pr_info("Performance Events: ");
2861 switch (boot_cpu_data
.x86_vendor
) {
2862 case X86_VENDOR_INTEL
:
2863 err
= intel_pmu_init();
2865 case X86_VENDOR_AMD
:
2866 err
= amd_pmu_init();
2872 pr_cont("no PMU driver, software events only.\n");
2878 pr_cont("%s PMU driver.\n", x86_pmu
.name
);
2880 if (x86_pmu
.num_events
> X86_PMC_MAX_GENERIC
) {
2881 WARN(1, KERN_ERR
"hw perf events %d > max(%d), clipping!",
2882 x86_pmu
.num_events
, X86_PMC_MAX_GENERIC
);
2883 x86_pmu
.num_events
= X86_PMC_MAX_GENERIC
;
2885 perf_event_mask
= (1 << x86_pmu
.num_events
) - 1;
2886 perf_max_events
= x86_pmu
.num_events
;
2888 if (x86_pmu
.num_events_fixed
> X86_PMC_MAX_FIXED
) {
2889 WARN(1, KERN_ERR
"hw perf events fixed %d > max(%d), clipping!",
2890 x86_pmu
.num_events_fixed
, X86_PMC_MAX_FIXED
);
2891 x86_pmu
.num_events_fixed
= X86_PMC_MAX_FIXED
;
2895 ((1LL << x86_pmu
.num_events_fixed
)-1) << X86_PMC_IDX_FIXED
;
2896 x86_pmu
.intel_ctrl
= perf_event_mask
;
2898 perf_events_lapic_init();
2899 register_die_notifier(&perf_event_nmi_notifier
);
2901 unconstrained
= (struct event_constraint
)
2902 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu
.num_events
) - 1,
2903 0, x86_pmu
.num_events
);
2905 pr_info("... version: %d\n", x86_pmu
.version
);
2906 pr_info("... bit width: %d\n", x86_pmu
.event_bits
);
2907 pr_info("... generic registers: %d\n", x86_pmu
.num_events
);
2908 pr_info("... value mask: %016Lx\n", x86_pmu
.event_mask
);
2909 pr_info("... max period: %016Lx\n", x86_pmu
.max_period
);
2910 pr_info("... fixed-purpose events: %d\n", x86_pmu
.num_events_fixed
);
2911 pr_info("... event mask: %016Lx\n", perf_event_mask
);
2914 static inline void x86_pmu_read(struct perf_event
*event
)
2916 x86_perf_event_update(event
, &event
->hw
, event
->hw
.idx
);
2919 static const struct pmu pmu
= {
2920 .enable
= x86_pmu_enable
,
2921 .disable
= x86_pmu_disable
,
2922 .start
= x86_pmu_start
,
2923 .stop
= x86_pmu_stop
,
2924 .read
= x86_pmu_read
,
2925 .unthrottle
= x86_pmu_unthrottle
,
2929 * validate a single event group
2931 * validation include:
2932 * - check events are compatible which each other
2933 * - events do not compete for the same counter
2934 * - number of events <= number of counters
2936 * validation ensures the group can be loaded onto the
2937 * PMU if it was the only group available.
2939 static int validate_group(struct perf_event
*event
)
2941 struct perf_event
*leader
= event
->group_leader
;
2942 struct cpu_hw_events
*fake_cpuc
;
2946 fake_cpuc
= kmalloc(sizeof(*fake_cpuc
), GFP_KERNEL
| __GFP_ZERO
);
2951 * the event is not yet connected with its
2952 * siblings therefore we must first collect
2953 * existing siblings, then add the new event
2954 * before we can simulate the scheduling
2957 n
= collect_events(fake_cpuc
, leader
, true);
2961 fake_cpuc
->n_events
= n
;
2962 n
= collect_events(fake_cpuc
, event
, false);
2966 fake_cpuc
->n_events
= n
;
2968 ret
= x86_schedule_events(fake_cpuc
, n
, NULL
);
2976 const struct pmu
*hw_perf_event_init(struct perf_event
*event
)
2978 const struct pmu
*tmp
;
2981 err
= __hw_perf_event_init(event
);
2984 * we temporarily connect event to its pmu
2985 * such that validate_group() can classify
2986 * it as an x86 event using is_x86_event()
2991 if (event
->group_leader
!= event
)
2992 err
= validate_group(event
);
2998 event
->destroy(event
);
2999 return ERR_PTR(err
);
3010 void callchain_store(struct perf_callchain_entry
*entry
, u64 ip
)
3012 if (entry
->nr
< PERF_MAX_STACK_DEPTH
)
3013 entry
->ip
[entry
->nr
++] = ip
;
3016 static DEFINE_PER_CPU(struct perf_callchain_entry
, pmc_irq_entry
);
3017 static DEFINE_PER_CPU(struct perf_callchain_entry
, pmc_nmi_entry
);
3021 backtrace_warning_symbol(void *data
, char *msg
, unsigned long symbol
)
3023 /* Ignore warnings */
3026 static void backtrace_warning(void *data
, char *msg
)
3028 /* Ignore warnings */
3031 static int backtrace_stack(void *data
, char *name
)
3036 static void backtrace_address(void *data
, unsigned long addr
, int reliable
)
3038 struct perf_callchain_entry
*entry
= data
;
3041 callchain_store(entry
, addr
);
3044 static const struct stacktrace_ops backtrace_ops
= {
3045 .warning
= backtrace_warning
,
3046 .warning_symbol
= backtrace_warning_symbol
,
3047 .stack
= backtrace_stack
,
3048 .address
= backtrace_address
,
3049 .walk_stack
= print_context_stack_bp
,
3052 #include "../dumpstack.h"
3055 perf_callchain_kernel(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
3057 callchain_store(entry
, PERF_CONTEXT_KERNEL
);
3058 callchain_store(entry
, regs
->ip
);
3060 dump_trace(NULL
, regs
, NULL
, regs
->bp
, &backtrace_ops
, entry
);
3064 * best effort, GUP based copy_from_user() that assumes IRQ or NMI context
3066 static unsigned long
3067 copy_from_user_nmi(void *to
, const void __user
*from
, unsigned long n
)
3069 unsigned long offset
, addr
= (unsigned long)from
;
3070 int type
= in_nmi() ? KM_NMI
: KM_IRQ0
;
3071 unsigned long size
, len
= 0;
3077 ret
= __get_user_pages_fast(addr
, 1, 0, &page
);
3081 offset
= addr
& (PAGE_SIZE
- 1);
3082 size
= min(PAGE_SIZE
- offset
, n
- len
);
3084 map
= kmap_atomic(page
, type
);
3085 memcpy(to
, map
+offset
, size
);
3086 kunmap_atomic(map
, type
);
3098 static int copy_stack_frame(const void __user
*fp
, struct stack_frame
*frame
)
3100 unsigned long bytes
;
3102 bytes
= copy_from_user_nmi(frame
, fp
, sizeof(*frame
));
3104 return bytes
== sizeof(*frame
);
3108 perf_callchain_user(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
3110 struct stack_frame frame
;
3111 const void __user
*fp
;
3113 if (!user_mode(regs
))
3114 regs
= task_pt_regs(current
);
3116 fp
= (void __user
*)regs
->bp
;
3118 callchain_store(entry
, PERF_CONTEXT_USER
);
3119 callchain_store(entry
, regs
->ip
);
3121 while (entry
->nr
< PERF_MAX_STACK_DEPTH
) {
3122 frame
.next_frame
= NULL
;
3123 frame
.return_address
= 0;
3125 if (!copy_stack_frame(fp
, &frame
))
3128 if ((unsigned long)fp
< regs
->sp
)
3131 callchain_store(entry
, frame
.return_address
);
3132 fp
= frame
.next_frame
;
3137 perf_do_callchain(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
3144 is_user
= user_mode(regs
);
3146 if (is_user
&& current
->state
!= TASK_RUNNING
)
3150 perf_callchain_kernel(regs
, entry
);
3153 perf_callchain_user(regs
, entry
);
3156 struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
3158 struct perf_callchain_entry
*entry
;
3161 entry
= &__get_cpu_var(pmc_nmi_entry
);
3163 entry
= &__get_cpu_var(pmc_irq_entry
);
3167 perf_do_callchain(regs
, entry
);
3172 void hw_perf_event_setup_online(int cpu
)
3174 init_debug_store_on_cpu(cpu
);
3176 switch (boot_cpu_data
.x86_vendor
) {
3177 case X86_VENDOR_AMD
:
3178 amd_pmu_cpu_online(cpu
);
3185 void hw_perf_event_setup_offline(int cpu
)
3187 init_debug_store_on_cpu(cpu
);
3189 switch (boot_cpu_data
.x86_vendor
) {
3190 case X86_VENDOR_AMD
:
3191 amd_pmu_cpu_offline(cpu
);