]>
Commit | Line | Data |
---|---|---|
241771ef | 1 | /* |
cdd6c482 | 2 | * Performance events x86 architecture code |
241771ef | 3 | * |
98144511 IM |
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> | |
30dd568c | 9 | * Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com> |
1da53e02 | 10 | * Copyright (C) 2009 Google, Inc., Stephane Eranian |
241771ef IM |
11 | * |
12 | * For licencing details see kernel-base/COPYING | |
13 | */ | |
14 | ||
cdd6c482 | 15 | #include <linux/perf_event.h> |
241771ef IM |
16 | #include <linux/capability.h> |
17 | #include <linux/notifier.h> | |
18 | #include <linux/hardirq.h> | |
19 | #include <linux/kprobes.h> | |
4ac13294 | 20 | #include <linux/module.h> |
241771ef IM |
21 | #include <linux/kdebug.h> |
22 | #include <linux/sched.h> | |
d7d59fb3 | 23 | #include <linux/uaccess.h> |
5a0e3ad6 | 24 | #include <linux/slab.h> |
74193ef0 | 25 | #include <linux/highmem.h> |
30dd568c | 26 | #include <linux/cpu.h> |
272d30be | 27 | #include <linux/bitops.h> |
241771ef | 28 | |
241771ef | 29 | #include <asm/apic.h> |
d7d59fb3 | 30 | #include <asm/stacktrace.h> |
4e935e47 | 31 | #include <asm/nmi.h> |
257ef9d2 | 32 | #include <asm/compat.h> |
69092624 | 33 | #include <asm/smp.h> |
c8e5910e | 34 | #include <asm/alternative.h> |
241771ef | 35 | |
7645a24c PZ |
36 | #if 0 |
37 | #undef wrmsrl | |
38 | #define wrmsrl(msr, val) \ | |
39 | do { \ | |
40 | trace_printk("wrmsrl(%lx, %lx)\n", (unsigned long)(msr),\ | |
41 | (unsigned long)(val)); \ | |
42 | native_write_msr((msr), (u32)((u64)(val)), \ | |
43 | (u32)((u64)(val) >> 32)); \ | |
44 | } while (0) | |
45 | #endif | |
46 | ||
efc9f05d SE |
47 | /* |
48 | * | NHM/WSM | SNB | | |
49 | * register ------------------------------- | |
50 | * | HT | no HT | HT | no HT | | |
51 | *----------------------------------------- | |
52 | * offcore | core | core | cpu | core | | |
53 | * lbr_sel | core | core | cpu | core | | |
54 | * ld_lat | cpu | core | cpu | core | | |
55 | *----------------------------------------- | |
56 | * | |
57 | * Given that there is a small number of shared regs, | |
58 | * we can pre-allocate their slot in the per-cpu | |
59 | * per-core reg tables. | |
60 | */ | |
61 | enum extra_reg_type { | |
62 | EXTRA_REG_NONE = -1, /* not used */ | |
63 | ||
64 | EXTRA_REG_RSP_0 = 0, /* offcore_response_0 */ | |
65 | EXTRA_REG_RSP_1 = 1, /* offcore_response_1 */ | |
66 | ||
67 | EXTRA_REG_MAX /* number of entries needed */ | |
68 | }; | |
69 | ||
ef21f683 PZ |
70 | /* |
71 | * best effort, GUP based copy_from_user() that assumes IRQ or NMI context | |
72 | */ | |
73 | static unsigned long | |
74 | copy_from_user_nmi(void *to, const void __user *from, unsigned long n) | |
75 | { | |
76 | unsigned long offset, addr = (unsigned long)from; | |
ef21f683 PZ |
77 | unsigned long size, len = 0; |
78 | struct page *page; | |
79 | void *map; | |
80 | int ret; | |
81 | ||
82 | do { | |
83 | ret = __get_user_pages_fast(addr, 1, 0, &page); | |
84 | if (!ret) | |
85 | break; | |
86 | ||
87 | offset = addr & (PAGE_SIZE - 1); | |
88 | size = min(PAGE_SIZE - offset, n - len); | |
89 | ||
7a837d1b | 90 | map = kmap_atomic(page); |
ef21f683 | 91 | memcpy(to, map+offset, size); |
7a837d1b | 92 | kunmap_atomic(map); |
ef21f683 PZ |
93 | put_page(page); |
94 | ||
95 | len += size; | |
96 | to += size; | |
97 | addr += size; | |
98 | ||
99 | } while (len < n); | |
100 | ||
101 | return len; | |
102 | } | |
103 | ||
1da53e02 | 104 | struct event_constraint { |
c91e0f5d PZ |
105 | union { |
106 | unsigned long idxmsk[BITS_TO_LONGS(X86_PMC_IDX_MAX)]; | |
b622d644 | 107 | u64 idxmsk64; |
c91e0f5d | 108 | }; |
b622d644 PZ |
109 | u64 code; |
110 | u64 cmask; | |
272d30be | 111 | int weight; |
1da53e02 SE |
112 | }; |
113 | ||
38331f62 SE |
114 | struct amd_nb { |
115 | int nb_id; /* NorthBridge id */ | |
116 | int refcnt; /* reference count */ | |
117 | struct perf_event *owners[X86_PMC_IDX_MAX]; | |
118 | struct event_constraint event_constraints[X86_PMC_IDX_MAX]; | |
119 | }; | |
120 | ||
a7e3ed1e AK |
121 | struct intel_percore; |
122 | ||
caff2bef PZ |
123 | #define MAX_LBR_ENTRIES 16 |
124 | ||
cdd6c482 | 125 | struct cpu_hw_events { |
ca037701 PZ |
126 | /* |
127 | * Generic x86 PMC bits | |
128 | */ | |
1da53e02 | 129 | struct perf_event *events[X86_PMC_IDX_MAX]; /* in counter order */ |
43f6201a | 130 | unsigned long active_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)]; |
63e6be6d | 131 | unsigned long running[BITS_TO_LONGS(X86_PMC_IDX_MAX)]; |
b0f3f28e | 132 | int enabled; |
241771ef | 133 | |
1da53e02 SE |
134 | int n_events; |
135 | int n_added; | |
90151c35 | 136 | int n_txn; |
1da53e02 | 137 | int assign[X86_PMC_IDX_MAX]; /* event to counter assignment */ |
447a194b | 138 | u64 tags[X86_PMC_IDX_MAX]; |
1da53e02 | 139 | struct perf_event *event_list[X86_PMC_IDX_MAX]; /* in enabled order */ |
ca037701 | 140 | |
4d1c52b0 LM |
141 | unsigned int group_flag; |
142 | ||
ca037701 PZ |
143 | /* |
144 | * Intel DebugStore bits | |
145 | */ | |
146 | struct debug_store *ds; | |
147 | u64 pebs_enabled; | |
148 | ||
caff2bef PZ |
149 | /* |
150 | * Intel LBR bits | |
151 | */ | |
152 | int lbr_users; | |
153 | void *lbr_context; | |
154 | struct perf_branch_stack lbr_stack; | |
155 | struct perf_branch_entry lbr_entries[MAX_LBR_ENTRIES]; | |
156 | ||
a7e3ed1e | 157 | /* |
efc9f05d SE |
158 | * manage shared (per-core, per-cpu) registers |
159 | * used on Intel NHM/WSM/SNB | |
a7e3ed1e | 160 | */ |
efc9f05d | 161 | struct intel_shared_regs *shared_regs; |
a7e3ed1e | 162 | |
ca037701 PZ |
163 | /* |
164 | * AMD specific bits | |
165 | */ | |
38331f62 | 166 | struct amd_nb *amd_nb; |
b690081d SE |
167 | }; |
168 | ||
fce877e3 | 169 | #define __EVENT_CONSTRAINT(c, n, m, w) {\ |
b622d644 | 170 | { .idxmsk64 = (n) }, \ |
c91e0f5d PZ |
171 | .code = (c), \ |
172 | .cmask = (m), \ | |
fce877e3 | 173 | .weight = (w), \ |
c91e0f5d | 174 | } |
b690081d | 175 | |
fce877e3 PZ |
176 | #define EVENT_CONSTRAINT(c, n, m) \ |
177 | __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n)) | |
178 | ||
ca037701 PZ |
179 | /* |
180 | * Constraint on the Event code. | |
181 | */ | |
ed8777fc | 182 | #define INTEL_EVENT_CONSTRAINT(c, n) \ |
a098f448 | 183 | EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT) |
8433be11 | 184 | |
ca037701 PZ |
185 | /* |
186 | * Constraint on the Event code + UMask + fixed-mask | |
a098f448 RR |
187 | * |
188 | * filter mask to validate fixed counter events. | |
189 | * the following filters disqualify for fixed counters: | |
190 | * - inv | |
191 | * - edge | |
192 | * - cnt-mask | |
193 | * The other filters are supported by fixed counters. | |
194 | * The any-thread option is supported starting with v3. | |
ca037701 | 195 | */ |
ed8777fc | 196 | #define FIXED_EVENT_CONSTRAINT(c, n) \ |
a098f448 | 197 | EVENT_CONSTRAINT(c, (1ULL << (32+n)), X86_RAW_EVENT_MASK) |
8433be11 | 198 | |
ca037701 PZ |
199 | /* |
200 | * Constraint on the Event code + UMask | |
201 | */ | |
b06b3d49 | 202 | #define INTEL_UEVENT_CONSTRAINT(c, n) \ |
ca037701 PZ |
203 | EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK) |
204 | ||
ed8777fc PZ |
205 | #define EVENT_CONSTRAINT_END \ |
206 | EVENT_CONSTRAINT(0, 0, 0) | |
207 | ||
208 | #define for_each_event_constraint(e, c) \ | |
a1f2b70a | 209 | for ((e) = (c); (e)->weight; (e)++) |
b690081d | 210 | |
efc9f05d SE |
211 | /* |
212 | * Per register state. | |
213 | */ | |
214 | struct er_account { | |
215 | raw_spinlock_t lock; /* per-core: protect structure */ | |
216 | u64 config; /* extra MSR config */ | |
217 | u64 reg; /* extra MSR number */ | |
218 | atomic_t ref; /* reference count */ | |
219 | }; | |
220 | ||
a7e3ed1e AK |
221 | /* |
222 | * Extra registers for specific events. | |
efc9f05d | 223 | * |
a7e3ed1e | 224 | * Some events need large masks and require external MSRs. |
efc9f05d SE |
225 | * Those extra MSRs end up being shared for all events on |
226 | * a PMU and sometimes between PMU of sibling HT threads. | |
227 | * In either case, the kernel needs to handle conflicting | |
228 | * accesses to those extra, shared, regs. The data structure | |
229 | * to manage those registers is stored in cpu_hw_event. | |
a7e3ed1e AK |
230 | */ |
231 | struct extra_reg { | |
232 | unsigned int event; | |
233 | unsigned int msr; | |
234 | u64 config_mask; | |
235 | u64 valid_mask; | |
efc9f05d | 236 | int idx; /* per_xxx->regs[] reg index */ |
a7e3ed1e AK |
237 | }; |
238 | ||
efc9f05d | 239 | #define EVENT_EXTRA_REG(e, ms, m, vm, i) { \ |
a7e3ed1e AK |
240 | .event = (e), \ |
241 | .msr = (ms), \ | |
242 | .config_mask = (m), \ | |
243 | .valid_mask = (vm), \ | |
efc9f05d | 244 | .idx = EXTRA_REG_##i \ |
a7e3ed1e | 245 | } |
efc9f05d SE |
246 | |
247 | #define INTEL_EVENT_EXTRA_REG(event, msr, vm, idx) \ | |
248 | EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT, vm, idx) | |
249 | ||
250 | #define EVENT_EXTRA_END EVENT_EXTRA_REG(0, 0, 0, 0, RSP_0) | |
a7e3ed1e | 251 | |
8db909a7 PZ |
252 | union perf_capabilities { |
253 | struct { | |
254 | u64 lbr_format : 6; | |
255 | u64 pebs_trap : 1; | |
256 | u64 pebs_arch_reg : 1; | |
257 | u64 pebs_format : 4; | |
258 | u64 smm_freeze : 1; | |
259 | }; | |
260 | u64 capabilities; | |
261 | }; | |
262 | ||
241771ef | 263 | /* |
5f4ec28f | 264 | * struct x86_pmu - generic x86 pmu |
241771ef | 265 | */ |
5f4ec28f | 266 | struct x86_pmu { |
ca037701 PZ |
267 | /* |
268 | * Generic x86 PMC bits | |
269 | */ | |
faa28ae0 RR |
270 | const char *name; |
271 | int version; | |
a3288106 | 272 | int (*handle_irq)(struct pt_regs *); |
9e35ad38 | 273 | void (*disable_all)(void); |
11164cd4 | 274 | void (*enable_all)(int added); |
aff3d91a PZ |
275 | void (*enable)(struct perf_event *); |
276 | void (*disable)(struct perf_event *); | |
1880c4ae | 277 | void (*hw_watchdog_set_attr)(struct perf_event_attr *attr); |
b4cdc5c2 | 278 | int (*hw_config)(struct perf_event *event); |
a072738e | 279 | int (*schedule_events)(struct cpu_hw_events *cpuc, int n, int *assign); |
169e41eb JSR |
280 | unsigned eventsel; |
281 | unsigned perfctr; | |
b0f3f28e | 282 | u64 (*event_map)(int); |
169e41eb | 283 | int max_events; |
948b1bb8 RR |
284 | int num_counters; |
285 | int num_counters_fixed; | |
286 | int cntval_bits; | |
287 | u64 cntval_mask; | |
04da8a43 | 288 | int apic; |
c619b8ff | 289 | u64 max_period; |
63b14649 PZ |
290 | struct event_constraint * |
291 | (*get_event_constraints)(struct cpu_hw_events *cpuc, | |
292 | struct perf_event *event); | |
293 | ||
c91e0f5d PZ |
294 | void (*put_event_constraints)(struct cpu_hw_events *cpuc, |
295 | struct perf_event *event); | |
63b14649 | 296 | struct event_constraint *event_constraints; |
3c44780b | 297 | void (*quirks)(void); |
68aa00ac | 298 | int perfctr_second_write; |
3f6da390 | 299 | |
b38b24ea | 300 | int (*cpu_prepare)(int cpu); |
3f6da390 PZ |
301 | void (*cpu_starting)(int cpu); |
302 | void (*cpu_dying)(int cpu); | |
303 | void (*cpu_dead)(int cpu); | |
ca037701 PZ |
304 | |
305 | /* | |
306 | * Intel Arch Perfmon v2+ | |
307 | */ | |
8db909a7 PZ |
308 | u64 intel_ctrl; |
309 | union perf_capabilities intel_cap; | |
ca037701 PZ |
310 | |
311 | /* | |
312 | * Intel DebugStore bits | |
313 | */ | |
314 | int bts, pebs; | |
6809b6ea | 315 | int bts_active, pebs_active; |
ca037701 PZ |
316 | int pebs_record_size; |
317 | void (*drain_pebs)(struct pt_regs *regs); | |
318 | struct event_constraint *pebs_constraints; | |
caff2bef PZ |
319 | |
320 | /* | |
321 | * Intel LBR | |
322 | */ | |
323 | unsigned long lbr_tos, lbr_from, lbr_to; /* MSR base regs */ | |
324 | int lbr_nr; /* hardware stack size */ | |
a7e3ed1e AK |
325 | |
326 | /* | |
327 | * Extra registers for events | |
328 | */ | |
329 | struct extra_reg *extra_regs; | |
b56a3802 JSR |
330 | }; |
331 | ||
4a06bd85 | 332 | static struct x86_pmu x86_pmu __read_mostly; |
b56a3802 | 333 | |
cdd6c482 | 334 | static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = { |
b0f3f28e PZ |
335 | .enabled = 1, |
336 | }; | |
241771ef | 337 | |
07088edb | 338 | static int x86_perf_event_set_period(struct perf_event *event); |
b690081d | 339 | |
8326f44d | 340 | /* |
dfc65094 | 341 | * Generalized hw caching related hw_event table, filled |
8326f44d | 342 | * in on a per model basis. A value of 0 means |
dfc65094 IM |
343 | * 'not supported', -1 means 'hw_event makes no sense on |
344 | * this CPU', any other value means the raw hw_event | |
8326f44d IM |
345 | * ID. |
346 | */ | |
347 | ||
348 | #define C(x) PERF_COUNT_HW_CACHE_##x | |
349 | ||
350 | static u64 __read_mostly hw_cache_event_ids | |
351 | [PERF_COUNT_HW_CACHE_MAX] | |
352 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
353 | [PERF_COUNT_HW_CACHE_RESULT_MAX]; | |
e994d7d2 AK |
354 | static u64 __read_mostly hw_cache_extra_regs |
355 | [PERF_COUNT_HW_CACHE_MAX] | |
356 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
357 | [PERF_COUNT_HW_CACHE_RESULT_MAX]; | |
8326f44d | 358 | |
1880c4ae CG |
359 | void hw_nmi_watchdog_set_attr(struct perf_event_attr *wd_attr) |
360 | { | |
361 | if (x86_pmu.hw_watchdog_set_attr) | |
362 | x86_pmu.hw_watchdog_set_attr(wd_attr); | |
363 | } | |
364 | ||
ee06094f | 365 | /* |
cdd6c482 IM |
366 | * Propagate event elapsed time into the generic event. |
367 | * Can only be executed on the CPU where the event is active. | |
ee06094f IM |
368 | * Returns the delta events processed. |
369 | */ | |
4b7bfd0d | 370 | static u64 |
cc2ad4ba | 371 | x86_perf_event_update(struct perf_event *event) |
ee06094f | 372 | { |
cc2ad4ba | 373 | struct hw_perf_event *hwc = &event->hw; |
948b1bb8 | 374 | int shift = 64 - x86_pmu.cntval_bits; |
ec3232bd | 375 | u64 prev_raw_count, new_raw_count; |
cc2ad4ba | 376 | int idx = hwc->idx; |
ec3232bd | 377 | s64 delta; |
ee06094f | 378 | |
30dd568c MM |
379 | if (idx == X86_PMC_IDX_FIXED_BTS) |
380 | return 0; | |
381 | ||
ee06094f | 382 | /* |
cdd6c482 | 383 | * Careful: an NMI might modify the previous event value. |
ee06094f IM |
384 | * |
385 | * Our tactic to handle this is to first atomically read and | |
386 | * exchange a new raw count - then add that new-prev delta | |
cdd6c482 | 387 | * count to the generic event atomically: |
ee06094f IM |
388 | */ |
389 | again: | |
e7850595 | 390 | prev_raw_count = local64_read(&hwc->prev_count); |
73d6e522 | 391 | rdmsrl(hwc->event_base, new_raw_count); |
ee06094f | 392 | |
e7850595 | 393 | if (local64_cmpxchg(&hwc->prev_count, prev_raw_count, |
ee06094f IM |
394 | new_raw_count) != prev_raw_count) |
395 | goto again; | |
396 | ||
397 | /* | |
398 | * Now we have the new raw value and have updated the prev | |
399 | * timestamp already. We can now calculate the elapsed delta | |
cdd6c482 | 400 | * (event-)time and add that to the generic event. |
ee06094f IM |
401 | * |
402 | * Careful, not all hw sign-extends above the physical width | |
ec3232bd | 403 | * of the count. |
ee06094f | 404 | */ |
ec3232bd PZ |
405 | delta = (new_raw_count << shift) - (prev_raw_count << shift); |
406 | delta >>= shift; | |
ee06094f | 407 | |
e7850595 PZ |
408 | local64_add(delta, &event->count); |
409 | local64_sub(delta, &hwc->period_left); | |
4b7bfd0d RR |
410 | |
411 | return new_raw_count; | |
ee06094f IM |
412 | } |
413 | ||
4979d272 RR |
414 | static inline int x86_pmu_addr_offset(int index) |
415 | { | |
c8e5910e RR |
416 | int offset; |
417 | ||
418 | /* offset = X86_FEATURE_PERFCTR_CORE ? index << 1 : index */ | |
419 | alternative_io(ASM_NOP2, | |
420 | "shll $1, %%eax", | |
421 | X86_FEATURE_PERFCTR_CORE, | |
422 | "=a" (offset), | |
423 | "a" (index)); | |
424 | ||
425 | return offset; | |
4979d272 RR |
426 | } |
427 | ||
41bf4989 RR |
428 | static inline unsigned int x86_pmu_config_addr(int index) |
429 | { | |
4979d272 | 430 | return x86_pmu.eventsel + x86_pmu_addr_offset(index); |
41bf4989 RR |
431 | } |
432 | ||
433 | static inline unsigned int x86_pmu_event_addr(int index) | |
434 | { | |
4979d272 | 435 | return x86_pmu.perfctr + x86_pmu_addr_offset(index); |
41bf4989 RR |
436 | } |
437 | ||
a7e3ed1e AK |
438 | /* |
439 | * Find and validate any extra registers to set up. | |
440 | */ | |
441 | static int x86_pmu_extra_regs(u64 config, struct perf_event *event) | |
442 | { | |
efc9f05d | 443 | struct hw_perf_event_extra *reg; |
a7e3ed1e AK |
444 | struct extra_reg *er; |
445 | ||
efc9f05d | 446 | reg = &event->hw.extra_reg; |
a7e3ed1e AK |
447 | |
448 | if (!x86_pmu.extra_regs) | |
449 | return 0; | |
450 | ||
451 | for (er = x86_pmu.extra_regs; er->msr; er++) { | |
452 | if (er->event != (config & er->config_mask)) | |
453 | continue; | |
454 | if (event->attr.config1 & ~er->valid_mask) | |
455 | return -EINVAL; | |
efc9f05d SE |
456 | |
457 | reg->idx = er->idx; | |
458 | reg->config = event->attr.config1; | |
459 | reg->reg = er->msr; | |
a7e3ed1e AK |
460 | break; |
461 | } | |
462 | return 0; | |
463 | } | |
464 | ||
cdd6c482 | 465 | static atomic_t active_events; |
4e935e47 PZ |
466 | static DEFINE_MUTEX(pmc_reserve_mutex); |
467 | ||
b27ea29c RR |
468 | #ifdef CONFIG_X86_LOCAL_APIC |
469 | ||
4e935e47 PZ |
470 | static bool reserve_pmc_hardware(void) |
471 | { | |
472 | int i; | |
473 | ||
948b1bb8 | 474 | for (i = 0; i < x86_pmu.num_counters; i++) { |
41bf4989 | 475 | if (!reserve_perfctr_nmi(x86_pmu_event_addr(i))) |
4e935e47 PZ |
476 | goto perfctr_fail; |
477 | } | |
478 | ||
948b1bb8 | 479 | for (i = 0; i < x86_pmu.num_counters; i++) { |
41bf4989 | 480 | if (!reserve_evntsel_nmi(x86_pmu_config_addr(i))) |
4e935e47 PZ |
481 | goto eventsel_fail; |
482 | } | |
483 | ||
484 | return true; | |
485 | ||
486 | eventsel_fail: | |
487 | for (i--; i >= 0; i--) | |
41bf4989 | 488 | release_evntsel_nmi(x86_pmu_config_addr(i)); |
4e935e47 | 489 | |
948b1bb8 | 490 | i = x86_pmu.num_counters; |
4e935e47 PZ |
491 | |
492 | perfctr_fail: | |
493 | for (i--; i >= 0; i--) | |
41bf4989 | 494 | release_perfctr_nmi(x86_pmu_event_addr(i)); |
4e935e47 | 495 | |
4e935e47 PZ |
496 | return false; |
497 | } | |
498 | ||
499 | static void release_pmc_hardware(void) | |
500 | { | |
501 | int i; | |
502 | ||
948b1bb8 | 503 | for (i = 0; i < x86_pmu.num_counters; i++) { |
41bf4989 RR |
504 | release_perfctr_nmi(x86_pmu_event_addr(i)); |
505 | release_evntsel_nmi(x86_pmu_config_addr(i)); | |
4e935e47 | 506 | } |
4e935e47 PZ |
507 | } |
508 | ||
b27ea29c RR |
509 | #else |
510 | ||
511 | static bool reserve_pmc_hardware(void) { return true; } | |
512 | static void release_pmc_hardware(void) {} | |
513 | ||
514 | #endif | |
515 | ||
33c6d6a7 DZ |
516 | static bool check_hw_exists(void) |
517 | { | |
518 | u64 val, val_new = 0; | |
4407204c | 519 | int i, reg, ret = 0; |
33c6d6a7 | 520 | |
4407204c PZ |
521 | /* |
522 | * Check to see if the BIOS enabled any of the counters, if so | |
523 | * complain and bail. | |
524 | */ | |
525 | for (i = 0; i < x86_pmu.num_counters; i++) { | |
41bf4989 | 526 | reg = x86_pmu_config_addr(i); |
4407204c PZ |
527 | ret = rdmsrl_safe(reg, &val); |
528 | if (ret) | |
529 | goto msr_fail; | |
530 | if (val & ARCH_PERFMON_EVENTSEL_ENABLE) | |
531 | goto bios_fail; | |
532 | } | |
533 | ||
534 | if (x86_pmu.num_counters_fixed) { | |
535 | reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL; | |
536 | ret = rdmsrl_safe(reg, &val); | |
537 | if (ret) | |
538 | goto msr_fail; | |
539 | for (i = 0; i < x86_pmu.num_counters_fixed; i++) { | |
540 | if (val & (0x03 << i*4)) | |
541 | goto bios_fail; | |
542 | } | |
543 | } | |
544 | ||
545 | /* | |
546 | * Now write a value and read it back to see if it matches, | |
547 | * this is needed to detect certain hardware emulators (qemu/kvm) | |
548 | * that don't trap on the MSR access and always return 0s. | |
549 | */ | |
33c6d6a7 | 550 | val = 0xabcdUL; |
41bf4989 RR |
551 | ret = checking_wrmsrl(x86_pmu_event_addr(0), val); |
552 | ret |= rdmsrl_safe(x86_pmu_event_addr(0), &val_new); | |
33c6d6a7 | 553 | if (ret || val != val_new) |
4407204c | 554 | goto msr_fail; |
33c6d6a7 DZ |
555 | |
556 | return true; | |
4407204c PZ |
557 | |
558 | bios_fail: | |
45daae57 IM |
559 | /* |
560 | * We still allow the PMU driver to operate: | |
561 | */ | |
562 | printk(KERN_CONT "Broken BIOS detected, complain to your hardware vendor.\n"); | |
4407204c | 563 | printk(KERN_ERR FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n", reg, val); |
45daae57 IM |
564 | |
565 | return true; | |
4407204c PZ |
566 | |
567 | msr_fail: | |
568 | printk(KERN_CONT "Broken PMU hardware detected, using software events only.\n"); | |
45daae57 | 569 | |
4407204c | 570 | return false; |
33c6d6a7 DZ |
571 | } |
572 | ||
f80c9e30 | 573 | static void reserve_ds_buffers(void); |
ca037701 | 574 | static void release_ds_buffers(void); |
30dd568c | 575 | |
cdd6c482 | 576 | static void hw_perf_event_destroy(struct perf_event *event) |
4e935e47 | 577 | { |
cdd6c482 | 578 | if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) { |
4e935e47 | 579 | release_pmc_hardware(); |
ca037701 | 580 | release_ds_buffers(); |
4e935e47 PZ |
581 | mutex_unlock(&pmc_reserve_mutex); |
582 | } | |
583 | } | |
584 | ||
85cf9dba RR |
585 | static inline int x86_pmu_initialized(void) |
586 | { | |
587 | return x86_pmu.handle_irq != NULL; | |
588 | } | |
589 | ||
8326f44d | 590 | static inline int |
e994d7d2 | 591 | set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event) |
8326f44d | 592 | { |
e994d7d2 | 593 | struct perf_event_attr *attr = &event->attr; |
8326f44d IM |
594 | unsigned int cache_type, cache_op, cache_result; |
595 | u64 config, val; | |
596 | ||
597 | config = attr->config; | |
598 | ||
599 | cache_type = (config >> 0) & 0xff; | |
600 | if (cache_type >= PERF_COUNT_HW_CACHE_MAX) | |
601 | return -EINVAL; | |
602 | ||
603 | cache_op = (config >> 8) & 0xff; | |
604 | if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX) | |
605 | return -EINVAL; | |
606 | ||
607 | cache_result = (config >> 16) & 0xff; | |
608 | if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX) | |
609 | return -EINVAL; | |
610 | ||
611 | val = hw_cache_event_ids[cache_type][cache_op][cache_result]; | |
612 | ||
613 | if (val == 0) | |
614 | return -ENOENT; | |
615 | ||
616 | if (val == -1) | |
617 | return -EINVAL; | |
618 | ||
619 | hwc->config |= val; | |
e994d7d2 AK |
620 | attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result]; |
621 | return x86_pmu_extra_regs(val, event); | |
8326f44d IM |
622 | } |
623 | ||
c1726f34 RR |
624 | static int x86_setup_perfctr(struct perf_event *event) |
625 | { | |
626 | struct perf_event_attr *attr = &event->attr; | |
627 | struct hw_perf_event *hwc = &event->hw; | |
628 | u64 config; | |
629 | ||
6c7e550f | 630 | if (!is_sampling_event(event)) { |
c1726f34 RR |
631 | hwc->sample_period = x86_pmu.max_period; |
632 | hwc->last_period = hwc->sample_period; | |
e7850595 | 633 | local64_set(&hwc->period_left, hwc->sample_period); |
c1726f34 RR |
634 | } else { |
635 | /* | |
636 | * If we have a PMU initialized but no APIC | |
637 | * interrupts, we cannot sample hardware | |
638 | * events (user-space has to fall back and | |
639 | * sample via a hrtimer based software event): | |
640 | */ | |
641 | if (!x86_pmu.apic) | |
642 | return -EOPNOTSUPP; | |
643 | } | |
644 | ||
b52c55c6 IM |
645 | /* |
646 | * Do not allow config1 (extended registers) to propagate, | |
647 | * there's no sane user-space generalization yet: | |
648 | */ | |
c1726f34 | 649 | if (attr->type == PERF_TYPE_RAW) |
b52c55c6 | 650 | return 0; |
c1726f34 RR |
651 | |
652 | if (attr->type == PERF_TYPE_HW_CACHE) | |
e994d7d2 | 653 | return set_ext_hw_attr(hwc, event); |
c1726f34 RR |
654 | |
655 | if (attr->config >= x86_pmu.max_events) | |
656 | return -EINVAL; | |
657 | ||
658 | /* | |
659 | * The generic map: | |
660 | */ | |
661 | config = x86_pmu.event_map(attr->config); | |
662 | ||
663 | if (config == 0) | |
664 | return -ENOENT; | |
665 | ||
666 | if (config == -1LL) | |
667 | return -EINVAL; | |
668 | ||
669 | /* | |
670 | * Branch tracing: | |
671 | */ | |
18a073a3 PZ |
672 | if (attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS && |
673 | !attr->freq && hwc->sample_period == 1) { | |
c1726f34 | 674 | /* BTS is not supported by this architecture. */ |
6809b6ea | 675 | if (!x86_pmu.bts_active) |
c1726f34 RR |
676 | return -EOPNOTSUPP; |
677 | ||
678 | /* BTS is currently only allowed for user-mode. */ | |
679 | if (!attr->exclude_kernel) | |
680 | return -EOPNOTSUPP; | |
681 | } | |
682 | ||
683 | hwc->config |= config; | |
684 | ||
685 | return 0; | |
686 | } | |
4261e0e0 | 687 | |
b4cdc5c2 | 688 | static int x86_pmu_hw_config(struct perf_event *event) |
a072738e | 689 | { |
ab608344 PZ |
690 | if (event->attr.precise_ip) { |
691 | int precise = 0; | |
692 | ||
693 | /* Support for constant skid */ | |
6809b6ea | 694 | if (x86_pmu.pebs_active) { |
ab608344 PZ |
695 | precise++; |
696 | ||
5553be26 PZ |
697 | /* Support for IP fixup */ |
698 | if (x86_pmu.lbr_nr) | |
699 | precise++; | |
700 | } | |
ab608344 PZ |
701 | |
702 | if (event->attr.precise_ip > precise) | |
703 | return -EOPNOTSUPP; | |
704 | } | |
705 | ||
a072738e CG |
706 | /* |
707 | * Generate PMC IRQs: | |
708 | * (keep 'enabled' bit clear for now) | |
709 | */ | |
b4cdc5c2 | 710 | event->hw.config = ARCH_PERFMON_EVENTSEL_INT; |
a072738e CG |
711 | |
712 | /* | |
713 | * Count user and OS events unless requested not to | |
714 | */ | |
b4cdc5c2 PZ |
715 | if (!event->attr.exclude_user) |
716 | event->hw.config |= ARCH_PERFMON_EVENTSEL_USR; | |
717 | if (!event->attr.exclude_kernel) | |
718 | event->hw.config |= ARCH_PERFMON_EVENTSEL_OS; | |
a072738e | 719 | |
b4cdc5c2 PZ |
720 | if (event->attr.type == PERF_TYPE_RAW) |
721 | event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK; | |
a072738e | 722 | |
9d0fcba6 | 723 | return x86_setup_perfctr(event); |
a098f448 RR |
724 | } |
725 | ||
241771ef | 726 | /* |
0d48696f | 727 | * Setup the hardware configuration for a given attr_type |
241771ef | 728 | */ |
b0a873eb | 729 | static int __x86_pmu_event_init(struct perf_event *event) |
241771ef | 730 | { |
4e935e47 | 731 | int err; |
241771ef | 732 | |
85cf9dba RR |
733 | if (!x86_pmu_initialized()) |
734 | return -ENODEV; | |
241771ef | 735 | |
4e935e47 | 736 | err = 0; |
cdd6c482 | 737 | if (!atomic_inc_not_zero(&active_events)) { |
4e935e47 | 738 | mutex_lock(&pmc_reserve_mutex); |
cdd6c482 | 739 | if (atomic_read(&active_events) == 0) { |
30dd568c MM |
740 | if (!reserve_pmc_hardware()) |
741 | err = -EBUSY; | |
f80c9e30 PZ |
742 | else |
743 | reserve_ds_buffers(); | |
30dd568c MM |
744 | } |
745 | if (!err) | |
cdd6c482 | 746 | atomic_inc(&active_events); |
4e935e47 PZ |
747 | mutex_unlock(&pmc_reserve_mutex); |
748 | } | |
749 | if (err) | |
750 | return err; | |
751 | ||
cdd6c482 | 752 | event->destroy = hw_perf_event_destroy; |
a1792cda | 753 | |
4261e0e0 RR |
754 | event->hw.idx = -1; |
755 | event->hw.last_cpu = -1; | |
756 | event->hw.last_tag = ~0ULL; | |
b690081d | 757 | |
efc9f05d SE |
758 | /* mark unused */ |
759 | event->hw.extra_reg.idx = EXTRA_REG_NONE; | |
760 | ||
9d0fcba6 | 761 | return x86_pmu.hw_config(event); |
4261e0e0 RR |
762 | } |
763 | ||
8c48e444 | 764 | static void x86_pmu_disable_all(void) |
f87ad35d | 765 | { |
cdd6c482 | 766 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); |
9e35ad38 PZ |
767 | int idx; |
768 | ||
948b1bb8 | 769 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
b0f3f28e PZ |
770 | u64 val; |
771 | ||
43f6201a | 772 | if (!test_bit(idx, cpuc->active_mask)) |
4295ee62 | 773 | continue; |
41bf4989 | 774 | rdmsrl(x86_pmu_config_addr(idx), val); |
bb1165d6 | 775 | if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE)) |
4295ee62 | 776 | continue; |
bb1165d6 | 777 | val &= ~ARCH_PERFMON_EVENTSEL_ENABLE; |
41bf4989 | 778 | wrmsrl(x86_pmu_config_addr(idx), val); |
f87ad35d | 779 | } |
f87ad35d JSR |
780 | } |
781 | ||
a4eaf7f1 | 782 | static void x86_pmu_disable(struct pmu *pmu) |
b56a3802 | 783 | { |
1da53e02 SE |
784 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); |
785 | ||
85cf9dba | 786 | if (!x86_pmu_initialized()) |
9e35ad38 | 787 | return; |
1da53e02 | 788 | |
1a6e21f7 PZ |
789 | if (!cpuc->enabled) |
790 | return; | |
791 | ||
792 | cpuc->n_added = 0; | |
793 | cpuc->enabled = 0; | |
794 | barrier(); | |
1da53e02 SE |
795 | |
796 | x86_pmu.disable_all(); | |
b56a3802 | 797 | } |
241771ef | 798 | |
d45dd923 RR |
799 | static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc, |
800 | u64 enable_mask) | |
801 | { | |
efc9f05d SE |
802 | if (hwc->extra_reg.reg) |
803 | wrmsrl(hwc->extra_reg.reg, hwc->extra_reg.config); | |
73d6e522 | 804 | wrmsrl(hwc->config_base, hwc->config | enable_mask); |
d45dd923 RR |
805 | } |
806 | ||
11164cd4 | 807 | static void x86_pmu_enable_all(int added) |
f87ad35d | 808 | { |
cdd6c482 | 809 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); |
f87ad35d JSR |
810 | int idx; |
811 | ||
948b1bb8 | 812 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
d45dd923 | 813 | struct hw_perf_event *hwc = &cpuc->events[idx]->hw; |
b0f3f28e | 814 | |
43f6201a | 815 | if (!test_bit(idx, cpuc->active_mask)) |
4295ee62 | 816 | continue; |
984b838c | 817 | |
d45dd923 | 818 | __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); |
f87ad35d JSR |
819 | } |
820 | } | |
821 | ||
51b0fe39 | 822 | static struct pmu pmu; |
1da53e02 SE |
823 | |
824 | static inline int is_x86_event(struct perf_event *event) | |
825 | { | |
826 | return event->pmu == &pmu; | |
827 | } | |
828 | ||
829 | static int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign) | |
830 | { | |
63b14649 | 831 | struct event_constraint *c, *constraints[X86_PMC_IDX_MAX]; |
1da53e02 | 832 | unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)]; |
c933c1a6 | 833 | int i, j, w, wmax, num = 0; |
1da53e02 SE |
834 | struct hw_perf_event *hwc; |
835 | ||
836 | bitmap_zero(used_mask, X86_PMC_IDX_MAX); | |
837 | ||
838 | for (i = 0; i < n; i++) { | |
b622d644 PZ |
839 | c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]); |
840 | constraints[i] = c; | |
1da53e02 SE |
841 | } |
842 | ||
8113070d SE |
843 | /* |
844 | * fastpath, try to reuse previous register | |
845 | */ | |
c933c1a6 | 846 | for (i = 0; i < n; i++) { |
8113070d | 847 | hwc = &cpuc->event_list[i]->hw; |
81269a08 | 848 | c = constraints[i]; |
8113070d SE |
849 | |
850 | /* never assigned */ | |
851 | if (hwc->idx == -1) | |
852 | break; | |
853 | ||
854 | /* constraint still honored */ | |
63b14649 | 855 | if (!test_bit(hwc->idx, c->idxmsk)) |
8113070d SE |
856 | break; |
857 | ||
858 | /* not already used */ | |
859 | if (test_bit(hwc->idx, used_mask)) | |
860 | break; | |
861 | ||
34538ee7 | 862 | __set_bit(hwc->idx, used_mask); |
8113070d SE |
863 | if (assign) |
864 | assign[i] = hwc->idx; | |
865 | } | |
c933c1a6 | 866 | if (i == n) |
8113070d SE |
867 | goto done; |
868 | ||
869 | /* | |
870 | * begin slow path | |
871 | */ | |
872 | ||
873 | bitmap_zero(used_mask, X86_PMC_IDX_MAX); | |
874 | ||
1da53e02 SE |
875 | /* |
876 | * weight = number of possible counters | |
877 | * | |
878 | * 1 = most constrained, only works on one counter | |
879 | * wmax = least constrained, works on any counter | |
880 | * | |
881 | * assign events to counters starting with most | |
882 | * constrained events. | |
883 | */ | |
948b1bb8 | 884 | wmax = x86_pmu.num_counters; |
1da53e02 SE |
885 | |
886 | /* | |
887 | * when fixed event counters are present, | |
888 | * wmax is incremented by 1 to account | |
889 | * for one more choice | |
890 | */ | |
948b1bb8 | 891 | if (x86_pmu.num_counters_fixed) |
1da53e02 SE |
892 | wmax++; |
893 | ||
8113070d | 894 | for (w = 1, num = n; num && w <= wmax; w++) { |
1da53e02 | 895 | /* for each event */ |
8113070d | 896 | for (i = 0; num && i < n; i++) { |
81269a08 | 897 | c = constraints[i]; |
1da53e02 SE |
898 | hwc = &cpuc->event_list[i]->hw; |
899 | ||
272d30be | 900 | if (c->weight != w) |
1da53e02 SE |
901 | continue; |
902 | ||
984b3f57 | 903 | for_each_set_bit(j, c->idxmsk, X86_PMC_IDX_MAX) { |
1da53e02 SE |
904 | if (!test_bit(j, used_mask)) |
905 | break; | |
906 | } | |
907 | ||
908 | if (j == X86_PMC_IDX_MAX) | |
909 | break; | |
1da53e02 | 910 | |
34538ee7 | 911 | __set_bit(j, used_mask); |
8113070d | 912 | |
1da53e02 SE |
913 | if (assign) |
914 | assign[i] = j; | |
915 | num--; | |
916 | } | |
917 | } | |
8113070d | 918 | done: |
1da53e02 SE |
919 | /* |
920 | * scheduling failed or is just a simulation, | |
921 | * free resources if necessary | |
922 | */ | |
923 | if (!assign || num) { | |
924 | for (i = 0; i < n; i++) { | |
925 | if (x86_pmu.put_event_constraints) | |
926 | x86_pmu.put_event_constraints(cpuc, cpuc->event_list[i]); | |
927 | } | |
928 | } | |
929 | return num ? -ENOSPC : 0; | |
930 | } | |
931 | ||
932 | /* | |
933 | * dogrp: true if must collect siblings events (group) | |
934 | * returns total number of events and error code | |
935 | */ | |
936 | static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp) | |
937 | { | |
938 | struct perf_event *event; | |
939 | int n, max_count; | |
940 | ||
948b1bb8 | 941 | max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed; |
1da53e02 SE |
942 | |
943 | /* current number of events already accepted */ | |
944 | n = cpuc->n_events; | |
945 | ||
946 | if (is_x86_event(leader)) { | |
947 | if (n >= max_count) | |
948 | return -ENOSPC; | |
949 | cpuc->event_list[n] = leader; | |
950 | n++; | |
951 | } | |
952 | if (!dogrp) | |
953 | return n; | |
954 | ||
955 | list_for_each_entry(event, &leader->sibling_list, group_entry) { | |
956 | if (!is_x86_event(event) || | |
8113070d | 957 | event->state <= PERF_EVENT_STATE_OFF) |
1da53e02 SE |
958 | continue; |
959 | ||
960 | if (n >= max_count) | |
961 | return -ENOSPC; | |
962 | ||
963 | cpuc->event_list[n] = event; | |
964 | n++; | |
965 | } | |
966 | return n; | |
967 | } | |
968 | ||
1da53e02 | 969 | static inline void x86_assign_hw_event(struct perf_event *event, |
447a194b | 970 | struct cpu_hw_events *cpuc, int i) |
1da53e02 | 971 | { |
447a194b SE |
972 | struct hw_perf_event *hwc = &event->hw; |
973 | ||
974 | hwc->idx = cpuc->assign[i]; | |
975 | hwc->last_cpu = smp_processor_id(); | |
976 | hwc->last_tag = ++cpuc->tags[i]; | |
1da53e02 SE |
977 | |
978 | if (hwc->idx == X86_PMC_IDX_FIXED_BTS) { | |
979 | hwc->config_base = 0; | |
980 | hwc->event_base = 0; | |
981 | } else if (hwc->idx >= X86_PMC_IDX_FIXED) { | |
982 | hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL; | |
fc66c521 | 983 | hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - X86_PMC_IDX_FIXED); |
1da53e02 | 984 | } else { |
73d6e522 RR |
985 | hwc->config_base = x86_pmu_config_addr(hwc->idx); |
986 | hwc->event_base = x86_pmu_event_addr(hwc->idx); | |
1da53e02 SE |
987 | } |
988 | } | |
989 | ||
447a194b SE |
990 | static inline int match_prev_assignment(struct hw_perf_event *hwc, |
991 | struct cpu_hw_events *cpuc, | |
992 | int i) | |
993 | { | |
994 | return hwc->idx == cpuc->assign[i] && | |
995 | hwc->last_cpu == smp_processor_id() && | |
996 | hwc->last_tag == cpuc->tags[i]; | |
997 | } | |
998 | ||
a4eaf7f1 PZ |
999 | static void x86_pmu_start(struct perf_event *event, int flags); |
1000 | static void x86_pmu_stop(struct perf_event *event, int flags); | |
2e841873 | 1001 | |
a4eaf7f1 | 1002 | static void x86_pmu_enable(struct pmu *pmu) |
ee06094f | 1003 | { |
1da53e02 SE |
1004 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); |
1005 | struct perf_event *event; | |
1006 | struct hw_perf_event *hwc; | |
11164cd4 | 1007 | int i, added = cpuc->n_added; |
1da53e02 | 1008 | |
85cf9dba | 1009 | if (!x86_pmu_initialized()) |
2b9ff0db | 1010 | return; |
1a6e21f7 PZ |
1011 | |
1012 | if (cpuc->enabled) | |
1013 | return; | |
1014 | ||
1da53e02 | 1015 | if (cpuc->n_added) { |
19925ce7 | 1016 | int n_running = cpuc->n_events - cpuc->n_added; |
1da53e02 SE |
1017 | /* |
1018 | * apply assignment obtained either from | |
1019 | * hw_perf_group_sched_in() or x86_pmu_enable() | |
1020 | * | |
1021 | * step1: save events moving to new counters | |
1022 | * step2: reprogram moved events into new counters | |
1023 | */ | |
19925ce7 | 1024 | for (i = 0; i < n_running; i++) { |
1da53e02 SE |
1025 | event = cpuc->event_list[i]; |
1026 | hwc = &event->hw; | |
1027 | ||
447a194b SE |
1028 | /* |
1029 | * we can avoid reprogramming counter if: | |
1030 | * - assigned same counter as last time | |
1031 | * - running on same CPU as last time | |
1032 | * - no other event has used the counter since | |
1033 | */ | |
1034 | if (hwc->idx == -1 || | |
1035 | match_prev_assignment(hwc, cpuc, i)) | |
1da53e02 SE |
1036 | continue; |
1037 | ||
a4eaf7f1 PZ |
1038 | /* |
1039 | * Ensure we don't accidentally enable a stopped | |
1040 | * counter simply because we rescheduled. | |
1041 | */ | |
1042 | if (hwc->state & PERF_HES_STOPPED) | |
1043 | hwc->state |= PERF_HES_ARCH; | |
1044 | ||
1045 | x86_pmu_stop(event, PERF_EF_UPDATE); | |
1da53e02 SE |
1046 | } |
1047 | ||
1048 | for (i = 0; i < cpuc->n_events; i++) { | |
1da53e02 SE |
1049 | event = cpuc->event_list[i]; |
1050 | hwc = &event->hw; | |
1051 | ||
45e16a68 | 1052 | if (!match_prev_assignment(hwc, cpuc, i)) |
447a194b | 1053 | x86_assign_hw_event(event, cpuc, i); |
45e16a68 PZ |
1054 | else if (i < n_running) |
1055 | continue; | |
1da53e02 | 1056 | |
a4eaf7f1 PZ |
1057 | if (hwc->state & PERF_HES_ARCH) |
1058 | continue; | |
1059 | ||
1060 | x86_pmu_start(event, PERF_EF_RELOAD); | |
1da53e02 SE |
1061 | } |
1062 | cpuc->n_added = 0; | |
1063 | perf_events_lapic_init(); | |
1064 | } | |
1a6e21f7 PZ |
1065 | |
1066 | cpuc->enabled = 1; | |
1067 | barrier(); | |
1068 | ||
11164cd4 | 1069 | x86_pmu.enable_all(added); |
ee06094f | 1070 | } |
ee06094f | 1071 | |
aff3d91a | 1072 | static inline void x86_pmu_disable_event(struct perf_event *event) |
b0f3f28e | 1073 | { |
aff3d91a | 1074 | struct hw_perf_event *hwc = &event->hw; |
7645a24c | 1075 | |
73d6e522 | 1076 | wrmsrl(hwc->config_base, hwc->config); |
b0f3f28e PZ |
1077 | } |
1078 | ||
245b2e70 | 1079 | static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left); |
241771ef | 1080 | |
ee06094f IM |
1081 | /* |
1082 | * Set the next IRQ period, based on the hwc->period_left value. | |
cdd6c482 | 1083 | * To be called with the event disabled in hw: |
ee06094f | 1084 | */ |
e4abb5d4 | 1085 | static int |
07088edb | 1086 | x86_perf_event_set_period(struct perf_event *event) |
241771ef | 1087 | { |
07088edb | 1088 | struct hw_perf_event *hwc = &event->hw; |
e7850595 | 1089 | s64 left = local64_read(&hwc->period_left); |
e4abb5d4 | 1090 | s64 period = hwc->sample_period; |
7645a24c | 1091 | int ret = 0, idx = hwc->idx; |
ee06094f | 1092 | |
30dd568c MM |
1093 | if (idx == X86_PMC_IDX_FIXED_BTS) |
1094 | return 0; | |
1095 | ||
ee06094f | 1096 | /* |
af901ca1 | 1097 | * If we are way outside a reasonable range then just skip forward: |
ee06094f IM |
1098 | */ |
1099 | if (unlikely(left <= -period)) { | |
1100 | left = period; | |
e7850595 | 1101 | local64_set(&hwc->period_left, left); |
9e350de3 | 1102 | hwc->last_period = period; |
e4abb5d4 | 1103 | ret = 1; |
ee06094f IM |
1104 | } |
1105 | ||
1106 | if (unlikely(left <= 0)) { | |
1107 | left += period; | |
e7850595 | 1108 | local64_set(&hwc->period_left, left); |
9e350de3 | 1109 | hwc->last_period = period; |
e4abb5d4 | 1110 | ret = 1; |
ee06094f | 1111 | } |
1c80f4b5 | 1112 | /* |
dfc65094 | 1113 | * Quirk: certain CPUs dont like it if just 1 hw_event is left: |
1c80f4b5 IM |
1114 | */ |
1115 | if (unlikely(left < 2)) | |
1116 | left = 2; | |
241771ef | 1117 | |
e4abb5d4 PZ |
1118 | if (left > x86_pmu.max_period) |
1119 | left = x86_pmu.max_period; | |
1120 | ||
245b2e70 | 1121 | per_cpu(pmc_prev_left[idx], smp_processor_id()) = left; |
ee06094f IM |
1122 | |
1123 | /* | |
cdd6c482 | 1124 | * The hw event starts counting from this event offset, |
ee06094f IM |
1125 | * mark it to be able to extra future deltas: |
1126 | */ | |
e7850595 | 1127 | local64_set(&hwc->prev_count, (u64)-left); |
ee06094f | 1128 | |
73d6e522 | 1129 | wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask); |
68aa00ac CG |
1130 | |
1131 | /* | |
1132 | * Due to erratum on certan cpu we need | |
1133 | * a second write to be sure the register | |
1134 | * is updated properly | |
1135 | */ | |
1136 | if (x86_pmu.perfctr_second_write) { | |
73d6e522 | 1137 | wrmsrl(hwc->event_base, |
948b1bb8 | 1138 | (u64)(-left) & x86_pmu.cntval_mask); |
68aa00ac | 1139 | } |
e4abb5d4 | 1140 | |
cdd6c482 | 1141 | perf_event_update_userpage(event); |
194002b2 | 1142 | |
e4abb5d4 | 1143 | return ret; |
2f18d1e8 IM |
1144 | } |
1145 | ||
aff3d91a | 1146 | static void x86_pmu_enable_event(struct perf_event *event) |
7c90cc45 | 1147 | { |
0a3aee0d | 1148 | if (__this_cpu_read(cpu_hw_events.enabled)) |
31fa58af RR |
1149 | __x86_pmu_enable_event(&event->hw, |
1150 | ARCH_PERFMON_EVENTSEL_ENABLE); | |
241771ef IM |
1151 | } |
1152 | ||
b690081d | 1153 | /* |
a4eaf7f1 | 1154 | * Add a single event to the PMU. |
1da53e02 SE |
1155 | * |
1156 | * The event is added to the group of enabled events | |
1157 | * but only if it can be scehduled with existing events. | |
fe9081cc | 1158 | */ |
a4eaf7f1 | 1159 | static int x86_pmu_add(struct perf_event *event, int flags) |
fe9081cc PZ |
1160 | { |
1161 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | |
1da53e02 SE |
1162 | struct hw_perf_event *hwc; |
1163 | int assign[X86_PMC_IDX_MAX]; | |
1164 | int n, n0, ret; | |
fe9081cc | 1165 | |
1da53e02 | 1166 | hwc = &event->hw; |
fe9081cc | 1167 | |
33696fc0 | 1168 | perf_pmu_disable(event->pmu); |
1da53e02 | 1169 | n0 = cpuc->n_events; |
24cd7f54 PZ |
1170 | ret = n = collect_events(cpuc, event, false); |
1171 | if (ret < 0) | |
1172 | goto out; | |
53b441a5 | 1173 | |
a4eaf7f1 PZ |
1174 | hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED; |
1175 | if (!(flags & PERF_EF_START)) | |
1176 | hwc->state |= PERF_HES_ARCH; | |
1177 | ||
4d1c52b0 LM |
1178 | /* |
1179 | * If group events scheduling transaction was started, | |
0d2eb44f | 1180 | * skip the schedulability test here, it will be performed |
a4eaf7f1 | 1181 | * at commit time (->commit_txn) as a whole |
4d1c52b0 | 1182 | */ |
8d2cacbb | 1183 | if (cpuc->group_flag & PERF_EVENT_TXN) |
24cd7f54 | 1184 | goto done_collect; |
4d1c52b0 | 1185 | |
a072738e | 1186 | ret = x86_pmu.schedule_events(cpuc, n, assign); |
1da53e02 | 1187 | if (ret) |
24cd7f54 | 1188 | goto out; |
1da53e02 SE |
1189 | /* |
1190 | * copy new assignment, now we know it is possible | |
1191 | * will be used by hw_perf_enable() | |
1192 | */ | |
1193 | memcpy(cpuc->assign, assign, n*sizeof(int)); | |
7e2ae347 | 1194 | |
24cd7f54 | 1195 | done_collect: |
1da53e02 | 1196 | cpuc->n_events = n; |
356e1f2e | 1197 | cpuc->n_added += n - n0; |
90151c35 | 1198 | cpuc->n_txn += n - n0; |
95cdd2e7 | 1199 | |
24cd7f54 PZ |
1200 | ret = 0; |
1201 | out: | |
33696fc0 | 1202 | perf_pmu_enable(event->pmu); |
24cd7f54 | 1203 | return ret; |
241771ef IM |
1204 | } |
1205 | ||
a4eaf7f1 | 1206 | static void x86_pmu_start(struct perf_event *event, int flags) |
d76a0812 | 1207 | { |
c08053e6 PZ |
1208 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); |
1209 | int idx = event->hw.idx; | |
1210 | ||
a4eaf7f1 PZ |
1211 | if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED))) |
1212 | return; | |
1213 | ||
1214 | if (WARN_ON_ONCE(idx == -1)) | |
1215 | return; | |
1216 | ||
1217 | if (flags & PERF_EF_RELOAD) { | |
1218 | WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE)); | |
1219 | x86_perf_event_set_period(event); | |
1220 | } | |
1221 | ||
1222 | event->hw.state = 0; | |
d76a0812 | 1223 | |
c08053e6 PZ |
1224 | cpuc->events[idx] = event; |
1225 | __set_bit(idx, cpuc->active_mask); | |
63e6be6d | 1226 | __set_bit(idx, cpuc->running); |
aff3d91a | 1227 | x86_pmu.enable(event); |
c08053e6 | 1228 | perf_event_update_userpage(event); |
a78ac325 PZ |
1229 | } |
1230 | ||
cdd6c482 | 1231 | void perf_event_print_debug(void) |
241771ef | 1232 | { |
2f18d1e8 | 1233 | u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed; |
ca037701 | 1234 | u64 pebs; |
cdd6c482 | 1235 | struct cpu_hw_events *cpuc; |
5bb9efe3 | 1236 | unsigned long flags; |
1e125676 IM |
1237 | int cpu, idx; |
1238 | ||
948b1bb8 | 1239 | if (!x86_pmu.num_counters) |
1e125676 | 1240 | return; |
241771ef | 1241 | |
5bb9efe3 | 1242 | local_irq_save(flags); |
241771ef IM |
1243 | |
1244 | cpu = smp_processor_id(); | |
cdd6c482 | 1245 | cpuc = &per_cpu(cpu_hw_events, cpu); |
241771ef | 1246 | |
faa28ae0 | 1247 | if (x86_pmu.version >= 2) { |
a1ef58f4 JSR |
1248 | rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl); |
1249 | rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status); | |
1250 | rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow); | |
1251 | rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed); | |
ca037701 | 1252 | rdmsrl(MSR_IA32_PEBS_ENABLE, pebs); |
a1ef58f4 JSR |
1253 | |
1254 | pr_info("\n"); | |
1255 | pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl); | |
1256 | pr_info("CPU#%d: status: %016llx\n", cpu, status); | |
1257 | pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow); | |
1258 | pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed); | |
ca037701 | 1259 | pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs); |
f87ad35d | 1260 | } |
7645a24c | 1261 | pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask); |
241771ef | 1262 | |
948b1bb8 | 1263 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
41bf4989 RR |
1264 | rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl); |
1265 | rdmsrl(x86_pmu_event_addr(idx), pmc_count); | |
241771ef | 1266 | |
245b2e70 | 1267 | prev_left = per_cpu(pmc_prev_left[idx], cpu); |
241771ef | 1268 | |
a1ef58f4 | 1269 | pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n", |
241771ef | 1270 | cpu, idx, pmc_ctrl); |
a1ef58f4 | 1271 | pr_info("CPU#%d: gen-PMC%d count: %016llx\n", |
241771ef | 1272 | cpu, idx, pmc_count); |
a1ef58f4 | 1273 | pr_info("CPU#%d: gen-PMC%d left: %016llx\n", |
ee06094f | 1274 | cpu, idx, prev_left); |
241771ef | 1275 | } |
948b1bb8 | 1276 | for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) { |
2f18d1e8 IM |
1277 | rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count); |
1278 | ||
a1ef58f4 | 1279 | pr_info("CPU#%d: fixed-PMC%d count: %016llx\n", |
2f18d1e8 IM |
1280 | cpu, idx, pmc_count); |
1281 | } | |
5bb9efe3 | 1282 | local_irq_restore(flags); |
241771ef IM |
1283 | } |
1284 | ||
a4eaf7f1 | 1285 | static void x86_pmu_stop(struct perf_event *event, int flags) |
241771ef | 1286 | { |
d76a0812 | 1287 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); |
cdd6c482 | 1288 | struct hw_perf_event *hwc = &event->hw; |
241771ef | 1289 | |
a4eaf7f1 PZ |
1290 | if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) { |
1291 | x86_pmu.disable(event); | |
1292 | cpuc->events[hwc->idx] = NULL; | |
1293 | WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED); | |
1294 | hwc->state |= PERF_HES_STOPPED; | |
1295 | } | |
30dd568c | 1296 | |
a4eaf7f1 PZ |
1297 | if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) { |
1298 | /* | |
1299 | * Drain the remaining delta count out of a event | |
1300 | * that we are disabling: | |
1301 | */ | |
1302 | x86_perf_event_update(event); | |
1303 | hwc->state |= PERF_HES_UPTODATE; | |
1304 | } | |
2e841873 PZ |
1305 | } |
1306 | ||
a4eaf7f1 | 1307 | static void x86_pmu_del(struct perf_event *event, int flags) |
2e841873 PZ |
1308 | { |
1309 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | |
1310 | int i; | |
1311 | ||
90151c35 SE |
1312 | /* |
1313 | * If we're called during a txn, we don't need to do anything. | |
1314 | * The events never got scheduled and ->cancel_txn will truncate | |
1315 | * the event_list. | |
1316 | */ | |
8d2cacbb | 1317 | if (cpuc->group_flag & PERF_EVENT_TXN) |
90151c35 SE |
1318 | return; |
1319 | ||
a4eaf7f1 | 1320 | x86_pmu_stop(event, PERF_EF_UPDATE); |
194002b2 | 1321 | |
1da53e02 SE |
1322 | for (i = 0; i < cpuc->n_events; i++) { |
1323 | if (event == cpuc->event_list[i]) { | |
1324 | ||
1325 | if (x86_pmu.put_event_constraints) | |
1326 | x86_pmu.put_event_constraints(cpuc, event); | |
1327 | ||
1328 | while (++i < cpuc->n_events) | |
1329 | cpuc->event_list[i-1] = cpuc->event_list[i]; | |
1330 | ||
1331 | --cpuc->n_events; | |
6c9687ab | 1332 | break; |
1da53e02 SE |
1333 | } |
1334 | } | |
cdd6c482 | 1335 | perf_event_update_userpage(event); |
241771ef IM |
1336 | } |
1337 | ||
8c48e444 | 1338 | static int x86_pmu_handle_irq(struct pt_regs *regs) |
a29aa8a7 | 1339 | { |
df1a132b | 1340 | struct perf_sample_data data; |
cdd6c482 IM |
1341 | struct cpu_hw_events *cpuc; |
1342 | struct perf_event *event; | |
11d1578f | 1343 | int idx, handled = 0; |
9029a5e3 IM |
1344 | u64 val; |
1345 | ||
dc1d628a | 1346 | perf_sample_data_init(&data, 0); |
df1a132b | 1347 | |
cdd6c482 | 1348 | cpuc = &__get_cpu_var(cpu_hw_events); |
962bf7a6 | 1349 | |
2bce5dac DZ |
1350 | /* |
1351 | * Some chipsets need to unmask the LVTPC in a particular spot | |
1352 | * inside the nmi handler. As a result, the unmasking was pushed | |
1353 | * into all the nmi handlers. | |
1354 | * | |
1355 | * This generic handler doesn't seem to have any issues where the | |
1356 | * unmasking occurs so it was left at the top. | |
1357 | */ | |
1358 | apic_write(APIC_LVTPC, APIC_DM_NMI); | |
1359 | ||
948b1bb8 | 1360 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
63e6be6d RR |
1361 | if (!test_bit(idx, cpuc->active_mask)) { |
1362 | /* | |
1363 | * Though we deactivated the counter some cpus | |
1364 | * might still deliver spurious interrupts still | |
1365 | * in flight. Catch them: | |
1366 | */ | |
1367 | if (__test_and_clear_bit(idx, cpuc->running)) | |
1368 | handled++; | |
a29aa8a7 | 1369 | continue; |
63e6be6d | 1370 | } |
962bf7a6 | 1371 | |
cdd6c482 | 1372 | event = cpuc->events[idx]; |
a4016a79 | 1373 | |
cc2ad4ba | 1374 | val = x86_perf_event_update(event); |
948b1bb8 | 1375 | if (val & (1ULL << (x86_pmu.cntval_bits - 1))) |
48e22d56 | 1376 | continue; |
962bf7a6 | 1377 | |
9e350de3 | 1378 | /* |
cdd6c482 | 1379 | * event overflow |
9e350de3 | 1380 | */ |
4177c42a | 1381 | handled++; |
cdd6c482 | 1382 | data.period = event->hw.last_period; |
9e350de3 | 1383 | |
07088edb | 1384 | if (!x86_perf_event_set_period(event)) |
e4abb5d4 PZ |
1385 | continue; |
1386 | ||
a8b0ca17 | 1387 | if (perf_event_overflow(event, &data, regs)) |
a4eaf7f1 | 1388 | x86_pmu_stop(event, 0); |
a29aa8a7 | 1389 | } |
962bf7a6 | 1390 | |
9e350de3 PZ |
1391 | if (handled) |
1392 | inc_irq_stat(apic_perf_irqs); | |
1393 | ||
a29aa8a7 RR |
1394 | return handled; |
1395 | } | |
39d81eab | 1396 | |
cdd6c482 | 1397 | void perf_events_lapic_init(void) |
241771ef | 1398 | { |
04da8a43 | 1399 | if (!x86_pmu.apic || !x86_pmu_initialized()) |
241771ef | 1400 | return; |
85cf9dba | 1401 | |
241771ef | 1402 | /* |
c323d95f | 1403 | * Always use NMI for PMU |
241771ef | 1404 | */ |
c323d95f | 1405 | apic_write(APIC_LVTPC, APIC_DM_NMI); |
241771ef IM |
1406 | } |
1407 | ||
4177c42a RR |
1408 | struct pmu_nmi_state { |
1409 | unsigned int marked; | |
1410 | int handled; | |
1411 | }; | |
1412 | ||
1413 | static DEFINE_PER_CPU(struct pmu_nmi_state, pmu_nmi); | |
1414 | ||
241771ef | 1415 | static int __kprobes |
cdd6c482 | 1416 | perf_event_nmi_handler(struct notifier_block *self, |
241771ef IM |
1417 | unsigned long cmd, void *__args) |
1418 | { | |
1419 | struct die_args *args = __args; | |
4177c42a RR |
1420 | unsigned int this_nmi; |
1421 | int handled; | |
b0f3f28e | 1422 | |
cdd6c482 | 1423 | if (!atomic_read(&active_events)) |
63a809a2 PZ |
1424 | return NOTIFY_DONE; |
1425 | ||
b0f3f28e PZ |
1426 | switch (cmd) { |
1427 | case DIE_NMI: | |
b0f3f28e | 1428 | break; |
4177c42a RR |
1429 | case DIE_NMIUNKNOWN: |
1430 | this_nmi = percpu_read(irq_stat.__nmi_count); | |
0a3aee0d | 1431 | if (this_nmi != __this_cpu_read(pmu_nmi.marked)) |
4177c42a RR |
1432 | /* let the kernel handle the unknown nmi */ |
1433 | return NOTIFY_DONE; | |
1434 | /* | |
1435 | * This one is a PMU back-to-back nmi. Two events | |
1436 | * trigger 'simultaneously' raising two back-to-back | |
1437 | * NMIs. If the first NMI handles both, the latter | |
1438 | * will be empty and daze the CPU. So, we drop it to | |
1439 | * avoid false-positive 'unknown nmi' messages. | |
1440 | */ | |
1441 | return NOTIFY_STOP; | |
b0f3f28e | 1442 | default: |
241771ef | 1443 | return NOTIFY_DONE; |
b0f3f28e | 1444 | } |
241771ef | 1445 | |
4177c42a RR |
1446 | handled = x86_pmu.handle_irq(args->regs); |
1447 | if (!handled) | |
1448 | return NOTIFY_DONE; | |
1449 | ||
1450 | this_nmi = percpu_read(irq_stat.__nmi_count); | |
1451 | if ((handled > 1) || | |
1452 | /* the next nmi could be a back-to-back nmi */ | |
0a3aee0d TH |
1453 | ((__this_cpu_read(pmu_nmi.marked) == this_nmi) && |
1454 | (__this_cpu_read(pmu_nmi.handled) > 1))) { | |
4177c42a RR |
1455 | /* |
1456 | * We could have two subsequent back-to-back nmis: The | |
1457 | * first handles more than one counter, the 2nd | |
1458 | * handles only one counter and the 3rd handles no | |
1459 | * counter. | |
1460 | * | |
1461 | * This is the 2nd nmi because the previous was | |
1462 | * handling more than one counter. We will mark the | |
1463 | * next (3rd) and then drop it if unhandled. | |
1464 | */ | |
0a3aee0d TH |
1465 | __this_cpu_write(pmu_nmi.marked, this_nmi + 1); |
1466 | __this_cpu_write(pmu_nmi.handled, handled); | |
4177c42a | 1467 | } |
241771ef | 1468 | |
a4016a79 | 1469 | return NOTIFY_STOP; |
241771ef IM |
1470 | } |
1471 | ||
f22f54f4 PZ |
1472 | static __read_mostly struct notifier_block perf_event_nmi_notifier = { |
1473 | .notifier_call = perf_event_nmi_handler, | |
1474 | .next = NULL, | |
166d7514 | 1475 | .priority = NMI_LOCAL_LOW_PRIOR, |
f22f54f4 PZ |
1476 | }; |
1477 | ||
63b14649 | 1478 | static struct event_constraint unconstrained; |
38331f62 | 1479 | static struct event_constraint emptyconstraint; |
63b14649 | 1480 | |
63b14649 | 1481 | static struct event_constraint * |
f22f54f4 | 1482 | x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event) |
1da53e02 | 1483 | { |
63b14649 | 1484 | struct event_constraint *c; |
1da53e02 | 1485 | |
1da53e02 SE |
1486 | if (x86_pmu.event_constraints) { |
1487 | for_each_event_constraint(c, x86_pmu.event_constraints) { | |
63b14649 PZ |
1488 | if ((event->hw.config & c->cmask) == c->code) |
1489 | return c; | |
1da53e02 SE |
1490 | } |
1491 | } | |
63b14649 PZ |
1492 | |
1493 | return &unconstrained; | |
1da53e02 SE |
1494 | } |
1495 | ||
f22f54f4 PZ |
1496 | #include "perf_event_amd.c" |
1497 | #include "perf_event_p6.c" | |
a072738e | 1498 | #include "perf_event_p4.c" |
caff2bef | 1499 | #include "perf_event_intel_lbr.c" |
ca037701 | 1500 | #include "perf_event_intel_ds.c" |
f22f54f4 | 1501 | #include "perf_event_intel.c" |
f87ad35d | 1502 | |
3f6da390 PZ |
1503 | static int __cpuinit |
1504 | x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu) | |
1505 | { | |
1506 | unsigned int cpu = (long)hcpu; | |
b38b24ea | 1507 | int ret = NOTIFY_OK; |
3f6da390 PZ |
1508 | |
1509 | switch (action & ~CPU_TASKS_FROZEN) { | |
1510 | case CPU_UP_PREPARE: | |
1511 | if (x86_pmu.cpu_prepare) | |
b38b24ea | 1512 | ret = x86_pmu.cpu_prepare(cpu); |
3f6da390 PZ |
1513 | break; |
1514 | ||
1515 | case CPU_STARTING: | |
1516 | if (x86_pmu.cpu_starting) | |
1517 | x86_pmu.cpu_starting(cpu); | |
1518 | break; | |
1519 | ||
1520 | case CPU_DYING: | |
1521 | if (x86_pmu.cpu_dying) | |
1522 | x86_pmu.cpu_dying(cpu); | |
1523 | break; | |
1524 | ||
b38b24ea | 1525 | case CPU_UP_CANCELED: |
3f6da390 PZ |
1526 | case CPU_DEAD: |
1527 | if (x86_pmu.cpu_dead) | |
1528 | x86_pmu.cpu_dead(cpu); | |
1529 | break; | |
1530 | ||
1531 | default: | |
1532 | break; | |
1533 | } | |
1534 | ||
b38b24ea | 1535 | return ret; |
3f6da390 PZ |
1536 | } |
1537 | ||
12558038 CG |
1538 | static void __init pmu_check_apic(void) |
1539 | { | |
1540 | if (cpu_has_apic) | |
1541 | return; | |
1542 | ||
1543 | x86_pmu.apic = 0; | |
1544 | pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n"); | |
1545 | pr_info("no hardware sampling interrupt available.\n"); | |
1546 | } | |
1547 | ||
dda99116 | 1548 | static int __init init_hw_perf_events(void) |
b56a3802 | 1549 | { |
b622d644 | 1550 | struct event_constraint *c; |
72eae04d RR |
1551 | int err; |
1552 | ||
cdd6c482 | 1553 | pr_info("Performance Events: "); |
1123e3ad | 1554 | |
b56a3802 JSR |
1555 | switch (boot_cpu_data.x86_vendor) { |
1556 | case X86_VENDOR_INTEL: | |
72eae04d | 1557 | err = intel_pmu_init(); |
b56a3802 | 1558 | break; |
f87ad35d | 1559 | case X86_VENDOR_AMD: |
72eae04d | 1560 | err = amd_pmu_init(); |
f87ad35d | 1561 | break; |
4138960a | 1562 | default: |
004417a6 | 1563 | return 0; |
b56a3802 | 1564 | } |
1123e3ad | 1565 | if (err != 0) { |
cdd6c482 | 1566 | pr_cont("no PMU driver, software events only.\n"); |
004417a6 | 1567 | return 0; |
1123e3ad | 1568 | } |
b56a3802 | 1569 | |
12558038 CG |
1570 | pmu_check_apic(); |
1571 | ||
33c6d6a7 | 1572 | /* sanity check that the hardware exists or is emulated */ |
4407204c | 1573 | if (!check_hw_exists()) |
004417a6 | 1574 | return 0; |
33c6d6a7 | 1575 | |
1123e3ad | 1576 | pr_cont("%s PMU driver.\n", x86_pmu.name); |
faa28ae0 | 1577 | |
3c44780b PZ |
1578 | if (x86_pmu.quirks) |
1579 | x86_pmu.quirks(); | |
1580 | ||
948b1bb8 | 1581 | if (x86_pmu.num_counters > X86_PMC_MAX_GENERIC) { |
cdd6c482 | 1582 | WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!", |
948b1bb8 RR |
1583 | x86_pmu.num_counters, X86_PMC_MAX_GENERIC); |
1584 | x86_pmu.num_counters = X86_PMC_MAX_GENERIC; | |
241771ef | 1585 | } |
948b1bb8 | 1586 | x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1; |
241771ef | 1587 | |
948b1bb8 | 1588 | if (x86_pmu.num_counters_fixed > X86_PMC_MAX_FIXED) { |
cdd6c482 | 1589 | WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!", |
948b1bb8 RR |
1590 | x86_pmu.num_counters_fixed, X86_PMC_MAX_FIXED); |
1591 | x86_pmu.num_counters_fixed = X86_PMC_MAX_FIXED; | |
703e937c | 1592 | } |
862a1a5f | 1593 | |
d6dc0b4e | 1594 | x86_pmu.intel_ctrl |= |
948b1bb8 | 1595 | ((1LL << x86_pmu.num_counters_fixed)-1) << X86_PMC_IDX_FIXED; |
241771ef | 1596 | |
cdd6c482 IM |
1597 | perf_events_lapic_init(); |
1598 | register_die_notifier(&perf_event_nmi_notifier); | |
1123e3ad | 1599 | |
63b14649 | 1600 | unconstrained = (struct event_constraint) |
948b1bb8 RR |
1601 | __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1, |
1602 | 0, x86_pmu.num_counters); | |
63b14649 | 1603 | |
b622d644 PZ |
1604 | if (x86_pmu.event_constraints) { |
1605 | for_each_event_constraint(c, x86_pmu.event_constraints) { | |
a098f448 | 1606 | if (c->cmask != X86_RAW_EVENT_MASK) |
b622d644 PZ |
1607 | continue; |
1608 | ||
948b1bb8 RR |
1609 | c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1; |
1610 | c->weight += x86_pmu.num_counters; | |
b622d644 PZ |
1611 | } |
1612 | } | |
1613 | ||
57c0c15b | 1614 | pr_info("... version: %d\n", x86_pmu.version); |
948b1bb8 RR |
1615 | pr_info("... bit width: %d\n", x86_pmu.cntval_bits); |
1616 | pr_info("... generic registers: %d\n", x86_pmu.num_counters); | |
1617 | pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask); | |
57c0c15b | 1618 | pr_info("... max period: %016Lx\n", x86_pmu.max_period); |
948b1bb8 | 1619 | pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed); |
d6dc0b4e | 1620 | pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl); |
3f6da390 | 1621 | |
2e80a82a | 1622 | perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW); |
3f6da390 | 1623 | perf_cpu_notifier(x86_pmu_notifier); |
004417a6 PZ |
1624 | |
1625 | return 0; | |
241771ef | 1626 | } |
004417a6 | 1627 | early_initcall(init_hw_perf_events); |
621a01ea | 1628 | |
cdd6c482 | 1629 | static inline void x86_pmu_read(struct perf_event *event) |
ee06094f | 1630 | { |
cc2ad4ba | 1631 | x86_perf_event_update(event); |
ee06094f IM |
1632 | } |
1633 | ||
4d1c52b0 LM |
1634 | /* |
1635 | * Start group events scheduling transaction | |
1636 | * Set the flag to make pmu::enable() not perform the | |
1637 | * schedulability test, it will be performed at commit time | |
1638 | */ | |
51b0fe39 | 1639 | static void x86_pmu_start_txn(struct pmu *pmu) |
4d1c52b0 | 1640 | { |
33696fc0 | 1641 | perf_pmu_disable(pmu); |
0a3aee0d TH |
1642 | __this_cpu_or(cpu_hw_events.group_flag, PERF_EVENT_TXN); |
1643 | __this_cpu_write(cpu_hw_events.n_txn, 0); | |
4d1c52b0 LM |
1644 | } |
1645 | ||
1646 | /* | |
1647 | * Stop group events scheduling transaction | |
1648 | * Clear the flag and pmu::enable() will perform the | |
1649 | * schedulability test. | |
1650 | */ | |
51b0fe39 | 1651 | static void x86_pmu_cancel_txn(struct pmu *pmu) |
4d1c52b0 | 1652 | { |
0a3aee0d | 1653 | __this_cpu_and(cpu_hw_events.group_flag, ~PERF_EVENT_TXN); |
90151c35 SE |
1654 | /* |
1655 | * Truncate the collected events. | |
1656 | */ | |
0a3aee0d TH |
1657 | __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn)); |
1658 | __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn)); | |
33696fc0 | 1659 | perf_pmu_enable(pmu); |
4d1c52b0 LM |
1660 | } |
1661 | ||
1662 | /* | |
1663 | * Commit group events scheduling transaction | |
1664 | * Perform the group schedulability test as a whole | |
1665 | * Return 0 if success | |
1666 | */ | |
51b0fe39 | 1667 | static int x86_pmu_commit_txn(struct pmu *pmu) |
4d1c52b0 LM |
1668 | { |
1669 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | |
1670 | int assign[X86_PMC_IDX_MAX]; | |
1671 | int n, ret; | |
1672 | ||
1673 | n = cpuc->n_events; | |
1674 | ||
1675 | if (!x86_pmu_initialized()) | |
1676 | return -EAGAIN; | |
1677 | ||
1678 | ret = x86_pmu.schedule_events(cpuc, n, assign); | |
1679 | if (ret) | |
1680 | return ret; | |
1681 | ||
1682 | /* | |
1683 | * copy new assignment, now we know it is possible | |
1684 | * will be used by hw_perf_enable() | |
1685 | */ | |
1686 | memcpy(cpuc->assign, assign, n*sizeof(int)); | |
1687 | ||
8d2cacbb | 1688 | cpuc->group_flag &= ~PERF_EVENT_TXN; |
33696fc0 | 1689 | perf_pmu_enable(pmu); |
4d1c52b0 LM |
1690 | return 0; |
1691 | } | |
1692 | ||
ca037701 PZ |
1693 | /* |
1694 | * validate that we can schedule this event | |
1695 | */ | |
1696 | static int validate_event(struct perf_event *event) | |
1697 | { | |
1698 | struct cpu_hw_events *fake_cpuc; | |
1699 | struct event_constraint *c; | |
1700 | int ret = 0; | |
1701 | ||
1702 | fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO); | |
1703 | if (!fake_cpuc) | |
1704 | return -ENOMEM; | |
1705 | ||
1706 | c = x86_pmu.get_event_constraints(fake_cpuc, event); | |
1707 | ||
1708 | if (!c || !c->weight) | |
1709 | ret = -ENOSPC; | |
1710 | ||
1711 | if (x86_pmu.put_event_constraints) | |
1712 | x86_pmu.put_event_constraints(fake_cpuc, event); | |
1713 | ||
1714 | kfree(fake_cpuc); | |
1715 | ||
1716 | return ret; | |
1717 | } | |
1718 | ||
1da53e02 SE |
1719 | /* |
1720 | * validate a single event group | |
1721 | * | |
1722 | * validation include: | |
184f412c IM |
1723 | * - check events are compatible which each other |
1724 | * - events do not compete for the same counter | |
1725 | * - number of events <= number of counters | |
1da53e02 SE |
1726 | * |
1727 | * validation ensures the group can be loaded onto the | |
1728 | * PMU if it was the only group available. | |
1729 | */ | |
fe9081cc PZ |
1730 | static int validate_group(struct perf_event *event) |
1731 | { | |
1da53e02 | 1732 | struct perf_event *leader = event->group_leader; |
502568d5 PZ |
1733 | struct cpu_hw_events *fake_cpuc; |
1734 | int ret, n; | |
fe9081cc | 1735 | |
502568d5 PZ |
1736 | ret = -ENOMEM; |
1737 | fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO); | |
1738 | if (!fake_cpuc) | |
1739 | goto out; | |
1da53e02 SE |
1740 | /* |
1741 | * the event is not yet connected with its | |
1742 | * siblings therefore we must first collect | |
1743 | * existing siblings, then add the new event | |
1744 | * before we can simulate the scheduling | |
1745 | */ | |
502568d5 PZ |
1746 | ret = -ENOSPC; |
1747 | n = collect_events(fake_cpuc, leader, true); | |
1da53e02 | 1748 | if (n < 0) |
502568d5 | 1749 | goto out_free; |
fe9081cc | 1750 | |
502568d5 PZ |
1751 | fake_cpuc->n_events = n; |
1752 | n = collect_events(fake_cpuc, event, false); | |
1da53e02 | 1753 | if (n < 0) |
502568d5 | 1754 | goto out_free; |
fe9081cc | 1755 | |
502568d5 | 1756 | fake_cpuc->n_events = n; |
1da53e02 | 1757 | |
a072738e | 1758 | ret = x86_pmu.schedule_events(fake_cpuc, n, NULL); |
502568d5 PZ |
1759 | |
1760 | out_free: | |
1761 | kfree(fake_cpuc); | |
1762 | out: | |
1763 | return ret; | |
fe9081cc PZ |
1764 | } |
1765 | ||
dda99116 | 1766 | static int x86_pmu_event_init(struct perf_event *event) |
621a01ea | 1767 | { |
51b0fe39 | 1768 | struct pmu *tmp; |
621a01ea IM |
1769 | int err; |
1770 | ||
b0a873eb PZ |
1771 | switch (event->attr.type) { |
1772 | case PERF_TYPE_RAW: | |
1773 | case PERF_TYPE_HARDWARE: | |
1774 | case PERF_TYPE_HW_CACHE: | |
1775 | break; | |
1776 | ||
1777 | default: | |
1778 | return -ENOENT; | |
1779 | } | |
1780 | ||
1781 | err = __x86_pmu_event_init(event); | |
fe9081cc | 1782 | if (!err) { |
8113070d SE |
1783 | /* |
1784 | * we temporarily connect event to its pmu | |
1785 | * such that validate_group() can classify | |
1786 | * it as an x86 event using is_x86_event() | |
1787 | */ | |
1788 | tmp = event->pmu; | |
1789 | event->pmu = &pmu; | |
1790 | ||
fe9081cc PZ |
1791 | if (event->group_leader != event) |
1792 | err = validate_group(event); | |
ca037701 PZ |
1793 | else |
1794 | err = validate_event(event); | |
8113070d SE |
1795 | |
1796 | event->pmu = tmp; | |
fe9081cc | 1797 | } |
a1792cda | 1798 | if (err) { |
cdd6c482 IM |
1799 | if (event->destroy) |
1800 | event->destroy(event); | |
a1792cda | 1801 | } |
621a01ea | 1802 | |
b0a873eb | 1803 | return err; |
621a01ea | 1804 | } |
d7d59fb3 | 1805 | |
b0a873eb | 1806 | static struct pmu pmu = { |
a4eaf7f1 PZ |
1807 | .pmu_enable = x86_pmu_enable, |
1808 | .pmu_disable = x86_pmu_disable, | |
1809 | ||
b0a873eb | 1810 | .event_init = x86_pmu_event_init, |
a4eaf7f1 PZ |
1811 | |
1812 | .add = x86_pmu_add, | |
1813 | .del = x86_pmu_del, | |
b0a873eb PZ |
1814 | .start = x86_pmu_start, |
1815 | .stop = x86_pmu_stop, | |
1816 | .read = x86_pmu_read, | |
a4eaf7f1 | 1817 | |
b0a873eb PZ |
1818 | .start_txn = x86_pmu_start_txn, |
1819 | .cancel_txn = x86_pmu_cancel_txn, | |
1820 | .commit_txn = x86_pmu_commit_txn, | |
1821 | }; | |
1822 | ||
d7d59fb3 PZ |
1823 | /* |
1824 | * callchain support | |
1825 | */ | |
1826 | ||
d7d59fb3 PZ |
1827 | static int backtrace_stack(void *data, char *name) |
1828 | { | |
038e836e | 1829 | return 0; |
d7d59fb3 PZ |
1830 | } |
1831 | ||
1832 | static void backtrace_address(void *data, unsigned long addr, int reliable) | |
1833 | { | |
1834 | struct perf_callchain_entry *entry = data; | |
1835 | ||
70791ce9 | 1836 | perf_callchain_store(entry, addr); |
d7d59fb3 PZ |
1837 | } |
1838 | ||
1839 | static const struct stacktrace_ops backtrace_ops = { | |
d7d59fb3 PZ |
1840 | .stack = backtrace_stack, |
1841 | .address = backtrace_address, | |
06d65bda | 1842 | .walk_stack = print_context_stack_bp, |
d7d59fb3 PZ |
1843 | }; |
1844 | ||
56962b44 FW |
1845 | void |
1846 | perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs) | |
d7d59fb3 | 1847 | { |
927c7a9e FW |
1848 | if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) { |
1849 | /* TODO: We don't support guest os callchain now */ | |
ed805261 | 1850 | return; |
927c7a9e FW |
1851 | } |
1852 | ||
70791ce9 | 1853 | perf_callchain_store(entry, regs->ip); |
d7d59fb3 | 1854 | |
e8e999cf | 1855 | dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry); |
d7d59fb3 PZ |
1856 | } |
1857 | ||
257ef9d2 TE |
1858 | #ifdef CONFIG_COMPAT |
1859 | static inline int | |
1860 | perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry) | |
74193ef0 | 1861 | { |
257ef9d2 TE |
1862 | /* 32-bit process in 64-bit kernel. */ |
1863 | struct stack_frame_ia32 frame; | |
1864 | const void __user *fp; | |
74193ef0 | 1865 | |
257ef9d2 TE |
1866 | if (!test_thread_flag(TIF_IA32)) |
1867 | return 0; | |
1868 | ||
1869 | fp = compat_ptr(regs->bp); | |
1870 | while (entry->nr < PERF_MAX_STACK_DEPTH) { | |
1871 | unsigned long bytes; | |
1872 | frame.next_frame = 0; | |
1873 | frame.return_address = 0; | |
1874 | ||
1875 | bytes = copy_from_user_nmi(&frame, fp, sizeof(frame)); | |
1876 | if (bytes != sizeof(frame)) | |
1877 | break; | |
74193ef0 | 1878 | |
257ef9d2 TE |
1879 | if (fp < compat_ptr(regs->sp)) |
1880 | break; | |
74193ef0 | 1881 | |
70791ce9 | 1882 | perf_callchain_store(entry, frame.return_address); |
257ef9d2 TE |
1883 | fp = compat_ptr(frame.next_frame); |
1884 | } | |
1885 | return 1; | |
d7d59fb3 | 1886 | } |
257ef9d2 TE |
1887 | #else |
1888 | static inline int | |
1889 | perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry) | |
1890 | { | |
1891 | return 0; | |
1892 | } | |
1893 | #endif | |
d7d59fb3 | 1894 | |
56962b44 FW |
1895 | void |
1896 | perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs) | |
d7d59fb3 PZ |
1897 | { |
1898 | struct stack_frame frame; | |
1899 | const void __user *fp; | |
1900 | ||
927c7a9e FW |
1901 | if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) { |
1902 | /* TODO: We don't support guest os callchain now */ | |
ed805261 | 1903 | return; |
927c7a9e | 1904 | } |
5a6cec3a | 1905 | |
74193ef0 | 1906 | fp = (void __user *)regs->bp; |
d7d59fb3 | 1907 | |
70791ce9 | 1908 | perf_callchain_store(entry, regs->ip); |
d7d59fb3 | 1909 | |
257ef9d2 TE |
1910 | if (perf_callchain_user32(regs, entry)) |
1911 | return; | |
1912 | ||
f9188e02 | 1913 | while (entry->nr < PERF_MAX_STACK_DEPTH) { |
257ef9d2 | 1914 | unsigned long bytes; |
038e836e | 1915 | frame.next_frame = NULL; |
d7d59fb3 PZ |
1916 | frame.return_address = 0; |
1917 | ||
257ef9d2 TE |
1918 | bytes = copy_from_user_nmi(&frame, fp, sizeof(frame)); |
1919 | if (bytes != sizeof(frame)) | |
d7d59fb3 PZ |
1920 | break; |
1921 | ||
5a6cec3a | 1922 | if ((unsigned long)fp < regs->sp) |
d7d59fb3 PZ |
1923 | break; |
1924 | ||
70791ce9 | 1925 | perf_callchain_store(entry, frame.return_address); |
038e836e | 1926 | fp = frame.next_frame; |
d7d59fb3 PZ |
1927 | } |
1928 | } | |
1929 | ||
39447b38 ZY |
1930 | unsigned long perf_instruction_pointer(struct pt_regs *regs) |
1931 | { | |
1932 | unsigned long ip; | |
dcf46b94 | 1933 | |
39447b38 ZY |
1934 | if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) |
1935 | ip = perf_guest_cbs->get_guest_ip(); | |
1936 | else | |
1937 | ip = instruction_pointer(regs); | |
dcf46b94 | 1938 | |
39447b38 ZY |
1939 | return ip; |
1940 | } | |
1941 | ||
1942 | unsigned long perf_misc_flags(struct pt_regs *regs) | |
1943 | { | |
1944 | int misc = 0; | |
dcf46b94 | 1945 | |
39447b38 | 1946 | if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) { |
dcf46b94 ZY |
1947 | if (perf_guest_cbs->is_user_mode()) |
1948 | misc |= PERF_RECORD_MISC_GUEST_USER; | |
1949 | else | |
1950 | misc |= PERF_RECORD_MISC_GUEST_KERNEL; | |
1951 | } else { | |
1952 | if (user_mode(regs)) | |
1953 | misc |= PERF_RECORD_MISC_USER; | |
1954 | else | |
1955 | misc |= PERF_RECORD_MISC_KERNEL; | |
1956 | } | |
1957 | ||
39447b38 | 1958 | if (regs->flags & PERF_EFLAGS_EXACT) |
ab608344 | 1959 | misc |= PERF_RECORD_MISC_EXACT_IP; |
39447b38 ZY |
1960 | |
1961 | return misc; | |
1962 | } |