Commit | Line | Data |
---|---|---|
241771ef IM |
1 | /* |
2 | * Performance counter x86 architecture code | |
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> | |
241771ef IM |
9 | * |
10 | * For licencing details see kernel-base/COPYING | |
11 | */ | |
12 | ||
13 | #include <linux/perf_counter.h> | |
14 | #include <linux/capability.h> | |
15 | #include <linux/notifier.h> | |
16 | #include <linux/hardirq.h> | |
17 | #include <linux/kprobes.h> | |
4ac13294 | 18 | #include <linux/module.h> |
241771ef IM |
19 | #include <linux/kdebug.h> |
20 | #include <linux/sched.h> | |
d7d59fb3 | 21 | #include <linux/uaccess.h> |
241771ef | 22 | |
241771ef | 23 | #include <asm/apic.h> |
d7d59fb3 | 24 | #include <asm/stacktrace.h> |
4e935e47 | 25 | #include <asm/nmi.h> |
241771ef | 26 | |
862a1a5f | 27 | static u64 perf_counter_mask __read_mostly; |
703e937c | 28 | |
241771ef | 29 | struct cpu_hw_counters { |
862a1a5f | 30 | struct perf_counter *counters[X86_PMC_IDX_MAX]; |
43f6201a RR |
31 | unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)]; |
32 | unsigned long active_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)]; | |
4b39fd96 | 33 | unsigned long interrupts; |
b0f3f28e | 34 | int enabled; |
241771ef IM |
35 | }; |
36 | ||
37 | /* | |
5f4ec28f | 38 | * struct x86_pmu - generic x86 pmu |
241771ef | 39 | */ |
5f4ec28f | 40 | struct x86_pmu { |
faa28ae0 RR |
41 | const char *name; |
42 | int version; | |
39d81eab | 43 | int (*handle_irq)(struct pt_regs *, int); |
9e35ad38 PZ |
44 | void (*disable_all)(void); |
45 | void (*enable_all)(void); | |
7c90cc45 | 46 | void (*enable)(struct hw_perf_counter *, int); |
d4369891 | 47 | void (*disable)(struct hw_perf_counter *, int); |
169e41eb JSR |
48 | unsigned eventsel; |
49 | unsigned perfctr; | |
b0f3f28e PZ |
50 | u64 (*event_map)(int); |
51 | u64 (*raw_event)(u64); | |
169e41eb | 52 | int max_events; |
0933e5c6 RR |
53 | int num_counters; |
54 | int num_counters_fixed; | |
55 | int counter_bits; | |
56 | u64 counter_mask; | |
c619b8ff | 57 | u64 max_period; |
9e35ad38 | 58 | u64 intel_ctrl; |
b56a3802 JSR |
59 | }; |
60 | ||
4a06bd85 | 61 | static struct x86_pmu x86_pmu __read_mostly; |
b56a3802 | 62 | |
b0f3f28e PZ |
63 | static DEFINE_PER_CPU(struct cpu_hw_counters, cpu_hw_counters) = { |
64 | .enabled = 1, | |
65 | }; | |
241771ef | 66 | |
b56a3802 JSR |
67 | /* |
68 | * Intel PerfMon v3. Used on Core2 and later. | |
69 | */ | |
b0f3f28e | 70 | static const u64 intel_perfmon_event_map[] = |
241771ef | 71 | { |
f650a672 | 72 | [PERF_COUNT_CPU_CYCLES] = 0x003c, |
241771ef IM |
73 | [PERF_COUNT_INSTRUCTIONS] = 0x00c0, |
74 | [PERF_COUNT_CACHE_REFERENCES] = 0x4f2e, | |
75 | [PERF_COUNT_CACHE_MISSES] = 0x412e, | |
76 | [PERF_COUNT_BRANCH_INSTRUCTIONS] = 0x00c4, | |
77 | [PERF_COUNT_BRANCH_MISSES] = 0x00c5, | |
f650a672 | 78 | [PERF_COUNT_BUS_CYCLES] = 0x013c, |
241771ef IM |
79 | }; |
80 | ||
5f4ec28f | 81 | static u64 intel_pmu_event_map(int event) |
b56a3802 JSR |
82 | { |
83 | return intel_perfmon_event_map[event]; | |
84 | } | |
241771ef | 85 | |
5f4ec28f | 86 | static u64 intel_pmu_raw_event(u64 event) |
b0f3f28e | 87 | { |
82bae4f8 PZ |
88 | #define CORE_EVNTSEL_EVENT_MASK 0x000000FFULL |
89 | #define CORE_EVNTSEL_UNIT_MASK 0x0000FF00ULL | |
90 | #define CORE_EVNTSEL_COUNTER_MASK 0xFF000000ULL | |
b0f3f28e PZ |
91 | |
92 | #define CORE_EVNTSEL_MASK \ | |
93 | (CORE_EVNTSEL_EVENT_MASK | \ | |
94 | CORE_EVNTSEL_UNIT_MASK | \ | |
95 | CORE_EVNTSEL_COUNTER_MASK) | |
96 | ||
97 | return event & CORE_EVNTSEL_MASK; | |
98 | } | |
99 | ||
f87ad35d JSR |
100 | /* |
101 | * AMD Performance Monitor K7 and later. | |
102 | */ | |
b0f3f28e | 103 | static const u64 amd_perfmon_event_map[] = |
f87ad35d JSR |
104 | { |
105 | [PERF_COUNT_CPU_CYCLES] = 0x0076, | |
106 | [PERF_COUNT_INSTRUCTIONS] = 0x00c0, | |
107 | [PERF_COUNT_CACHE_REFERENCES] = 0x0080, | |
108 | [PERF_COUNT_CACHE_MISSES] = 0x0081, | |
109 | [PERF_COUNT_BRANCH_INSTRUCTIONS] = 0x00c4, | |
110 | [PERF_COUNT_BRANCH_MISSES] = 0x00c5, | |
111 | }; | |
112 | ||
5f4ec28f | 113 | static u64 amd_pmu_event_map(int event) |
f87ad35d JSR |
114 | { |
115 | return amd_perfmon_event_map[event]; | |
116 | } | |
117 | ||
5f4ec28f | 118 | static u64 amd_pmu_raw_event(u64 event) |
b0f3f28e | 119 | { |
82bae4f8 PZ |
120 | #define K7_EVNTSEL_EVENT_MASK 0x7000000FFULL |
121 | #define K7_EVNTSEL_UNIT_MASK 0x00000FF00ULL | |
122 | #define K7_EVNTSEL_COUNTER_MASK 0x0FF000000ULL | |
b0f3f28e PZ |
123 | |
124 | #define K7_EVNTSEL_MASK \ | |
125 | (K7_EVNTSEL_EVENT_MASK | \ | |
126 | K7_EVNTSEL_UNIT_MASK | \ | |
127 | K7_EVNTSEL_COUNTER_MASK) | |
128 | ||
129 | return event & K7_EVNTSEL_MASK; | |
130 | } | |
131 | ||
ee06094f IM |
132 | /* |
133 | * Propagate counter elapsed time into the generic counter. | |
134 | * Can only be executed on the CPU where the counter is active. | |
135 | * Returns the delta events processed. | |
136 | */ | |
4b7bfd0d | 137 | static u64 |
ee06094f IM |
138 | x86_perf_counter_update(struct perf_counter *counter, |
139 | struct hw_perf_counter *hwc, int idx) | |
140 | { | |
ec3232bd PZ |
141 | int shift = 64 - x86_pmu.counter_bits; |
142 | u64 prev_raw_count, new_raw_count; | |
143 | s64 delta; | |
ee06094f | 144 | |
ee06094f IM |
145 | /* |
146 | * Careful: an NMI might modify the previous counter value. | |
147 | * | |
148 | * Our tactic to handle this is to first atomically read and | |
149 | * exchange a new raw count - then add that new-prev delta | |
150 | * count to the generic counter atomically: | |
151 | */ | |
152 | again: | |
153 | prev_raw_count = atomic64_read(&hwc->prev_count); | |
154 | rdmsrl(hwc->counter_base + idx, new_raw_count); | |
155 | ||
156 | if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count, | |
157 | new_raw_count) != prev_raw_count) | |
158 | goto again; | |
159 | ||
160 | /* | |
161 | * Now we have the new raw value and have updated the prev | |
162 | * timestamp already. We can now calculate the elapsed delta | |
163 | * (counter-)time and add that to the generic counter. | |
164 | * | |
165 | * Careful, not all hw sign-extends above the physical width | |
ec3232bd | 166 | * of the count. |
ee06094f | 167 | */ |
ec3232bd PZ |
168 | delta = (new_raw_count << shift) - (prev_raw_count << shift); |
169 | delta >>= shift; | |
ee06094f IM |
170 | |
171 | atomic64_add(delta, &counter->count); | |
172 | atomic64_sub(delta, &hwc->period_left); | |
4b7bfd0d RR |
173 | |
174 | return new_raw_count; | |
ee06094f IM |
175 | } |
176 | ||
ba77813a | 177 | static atomic_t active_counters; |
4e935e47 PZ |
178 | static DEFINE_MUTEX(pmc_reserve_mutex); |
179 | ||
180 | static bool reserve_pmc_hardware(void) | |
181 | { | |
182 | int i; | |
183 | ||
184 | if (nmi_watchdog == NMI_LOCAL_APIC) | |
185 | disable_lapic_nmi_watchdog(); | |
186 | ||
0933e5c6 | 187 | for (i = 0; i < x86_pmu.num_counters; i++) { |
4a06bd85 | 188 | if (!reserve_perfctr_nmi(x86_pmu.perfctr + i)) |
4e935e47 PZ |
189 | goto perfctr_fail; |
190 | } | |
191 | ||
0933e5c6 | 192 | for (i = 0; i < x86_pmu.num_counters; i++) { |
4a06bd85 | 193 | if (!reserve_evntsel_nmi(x86_pmu.eventsel + i)) |
4e935e47 PZ |
194 | goto eventsel_fail; |
195 | } | |
196 | ||
197 | return true; | |
198 | ||
199 | eventsel_fail: | |
200 | for (i--; i >= 0; i--) | |
4a06bd85 | 201 | release_evntsel_nmi(x86_pmu.eventsel + i); |
4e935e47 | 202 | |
0933e5c6 | 203 | i = x86_pmu.num_counters; |
4e935e47 PZ |
204 | |
205 | perfctr_fail: | |
206 | for (i--; i >= 0; i--) | |
4a06bd85 | 207 | release_perfctr_nmi(x86_pmu.perfctr + i); |
4e935e47 PZ |
208 | |
209 | if (nmi_watchdog == NMI_LOCAL_APIC) | |
210 | enable_lapic_nmi_watchdog(); | |
211 | ||
212 | return false; | |
213 | } | |
214 | ||
215 | static void release_pmc_hardware(void) | |
216 | { | |
217 | int i; | |
218 | ||
0933e5c6 | 219 | for (i = 0; i < x86_pmu.num_counters; i++) { |
4a06bd85 RR |
220 | release_perfctr_nmi(x86_pmu.perfctr + i); |
221 | release_evntsel_nmi(x86_pmu.eventsel + i); | |
4e935e47 PZ |
222 | } |
223 | ||
224 | if (nmi_watchdog == NMI_LOCAL_APIC) | |
225 | enable_lapic_nmi_watchdog(); | |
226 | } | |
227 | ||
228 | static void hw_perf_counter_destroy(struct perf_counter *counter) | |
229 | { | |
ba77813a | 230 | if (atomic_dec_and_mutex_lock(&active_counters, &pmc_reserve_mutex)) { |
4e935e47 PZ |
231 | release_pmc_hardware(); |
232 | mutex_unlock(&pmc_reserve_mutex); | |
233 | } | |
234 | } | |
235 | ||
85cf9dba RR |
236 | static inline int x86_pmu_initialized(void) |
237 | { | |
238 | return x86_pmu.handle_irq != NULL; | |
239 | } | |
240 | ||
241771ef IM |
241 | /* |
242 | * Setup the hardware configuration for a given hw_event_type | |
243 | */ | |
621a01ea | 244 | static int __hw_perf_counter_init(struct perf_counter *counter) |
241771ef | 245 | { |
9f66a381 | 246 | struct perf_counter_hw_event *hw_event = &counter->hw_event; |
241771ef | 247 | struct hw_perf_counter *hwc = &counter->hw; |
4e935e47 | 248 | int err; |
241771ef | 249 | |
85cf9dba RR |
250 | if (!x86_pmu_initialized()) |
251 | return -ENODEV; | |
241771ef | 252 | |
4e935e47 | 253 | err = 0; |
ba77813a | 254 | if (!atomic_inc_not_zero(&active_counters)) { |
4e935e47 | 255 | mutex_lock(&pmc_reserve_mutex); |
ba77813a | 256 | if (atomic_read(&active_counters) == 0 && !reserve_pmc_hardware()) |
4e935e47 PZ |
257 | err = -EBUSY; |
258 | else | |
ba77813a | 259 | atomic_inc(&active_counters); |
4e935e47 PZ |
260 | mutex_unlock(&pmc_reserve_mutex); |
261 | } | |
262 | if (err) | |
263 | return err; | |
264 | ||
241771ef | 265 | /* |
0475f9ea | 266 | * Generate PMC IRQs: |
241771ef IM |
267 | * (keep 'enabled' bit clear for now) |
268 | */ | |
0475f9ea | 269 | hwc->config = ARCH_PERFMON_EVENTSEL_INT; |
241771ef IM |
270 | |
271 | /* | |
0475f9ea | 272 | * Count user and OS events unless requested not to. |
241771ef | 273 | */ |
0475f9ea PM |
274 | if (!hw_event->exclude_user) |
275 | hwc->config |= ARCH_PERFMON_EVENTSEL_USR; | |
276 | if (!hw_event->exclude_kernel) | |
241771ef | 277 | hwc->config |= ARCH_PERFMON_EVENTSEL_OS; |
0475f9ea PM |
278 | |
279 | /* | |
280 | * If privileged enough, allow NMI events: | |
281 | */ | |
282 | hwc->nmi = 0; | |
a026dfec PZ |
283 | if (hw_event->nmi) { |
284 | if (sysctl_perf_counter_priv && !capable(CAP_SYS_ADMIN)) | |
285 | return -EACCES; | |
0475f9ea | 286 | hwc->nmi = 1; |
a026dfec | 287 | } |
b68f1d2e | 288 | perf_counters_lapic_init(hwc->nmi); |
241771ef | 289 | |
d2517a49 IM |
290 | if (!hwc->irq_period) |
291 | hwc->irq_period = x86_pmu.max_period; | |
292 | ||
60db5e09 PZ |
293 | atomic64_set(&hwc->period_left, |
294 | min(x86_pmu.max_period, hwc->irq_period)); | |
241771ef IM |
295 | |
296 | /* | |
dfa7c899 | 297 | * Raw event type provide the config in the event structure |
241771ef | 298 | */ |
f4a2deb4 | 299 | if (perf_event_raw(hw_event)) { |
4a06bd85 | 300 | hwc->config |= x86_pmu.raw_event(perf_event_config(hw_event)); |
241771ef | 301 | } else { |
4a06bd85 | 302 | if (perf_event_id(hw_event) >= x86_pmu.max_events) |
241771ef IM |
303 | return -EINVAL; |
304 | /* | |
305 | * The generic map: | |
306 | */ | |
4a06bd85 | 307 | hwc->config |= x86_pmu.event_map(perf_event_id(hw_event)); |
241771ef | 308 | } |
241771ef | 309 | |
4e935e47 PZ |
310 | counter->destroy = hw_perf_counter_destroy; |
311 | ||
241771ef IM |
312 | return 0; |
313 | } | |
314 | ||
9e35ad38 | 315 | static void intel_pmu_disable_all(void) |
4ac13294 | 316 | { |
862a1a5f | 317 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); |
241771ef | 318 | } |
b56a3802 | 319 | |
9e35ad38 | 320 | static void amd_pmu_disable_all(void) |
f87ad35d | 321 | { |
b0f3f28e | 322 | struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters); |
9e35ad38 PZ |
323 | int idx; |
324 | ||
325 | if (!cpuc->enabled) | |
326 | return; | |
b0f3f28e | 327 | |
b0f3f28e | 328 | cpuc->enabled = 0; |
60b3df9c PZ |
329 | /* |
330 | * ensure we write the disable before we start disabling the | |
5f4ec28f RR |
331 | * counters proper, so that amd_pmu_enable_counter() does the |
332 | * right thing. | |
60b3df9c | 333 | */ |
b0f3f28e | 334 | barrier(); |
f87ad35d | 335 | |
0933e5c6 | 336 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
b0f3f28e PZ |
337 | u64 val; |
338 | ||
43f6201a | 339 | if (!test_bit(idx, cpuc->active_mask)) |
4295ee62 | 340 | continue; |
f87ad35d | 341 | rdmsrl(MSR_K7_EVNTSEL0 + idx, val); |
4295ee62 RR |
342 | if (!(val & ARCH_PERFMON_EVENTSEL0_ENABLE)) |
343 | continue; | |
344 | val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE; | |
345 | wrmsrl(MSR_K7_EVNTSEL0 + idx, val); | |
f87ad35d | 346 | } |
f87ad35d JSR |
347 | } |
348 | ||
9e35ad38 | 349 | void hw_perf_disable(void) |
b56a3802 | 350 | { |
85cf9dba | 351 | if (!x86_pmu_initialized()) |
9e35ad38 PZ |
352 | return; |
353 | return x86_pmu.disable_all(); | |
b56a3802 | 354 | } |
241771ef | 355 | |
9e35ad38 | 356 | static void intel_pmu_enable_all(void) |
b56a3802 | 357 | { |
9e35ad38 | 358 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl); |
b56a3802 JSR |
359 | } |
360 | ||
9e35ad38 | 361 | static void amd_pmu_enable_all(void) |
f87ad35d | 362 | { |
b0f3f28e | 363 | struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters); |
f87ad35d JSR |
364 | int idx; |
365 | ||
9e35ad38 | 366 | if (cpuc->enabled) |
b0f3f28e PZ |
367 | return; |
368 | ||
9e35ad38 PZ |
369 | cpuc->enabled = 1; |
370 | barrier(); | |
371 | ||
0933e5c6 | 372 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
4295ee62 | 373 | u64 val; |
b0f3f28e | 374 | |
43f6201a | 375 | if (!test_bit(idx, cpuc->active_mask)) |
4295ee62 RR |
376 | continue; |
377 | rdmsrl(MSR_K7_EVNTSEL0 + idx, val); | |
378 | if (val & ARCH_PERFMON_EVENTSEL0_ENABLE) | |
379 | continue; | |
380 | val |= ARCH_PERFMON_EVENTSEL0_ENABLE; | |
381 | wrmsrl(MSR_K7_EVNTSEL0 + idx, val); | |
f87ad35d JSR |
382 | } |
383 | } | |
384 | ||
9e35ad38 | 385 | void hw_perf_enable(void) |
ee06094f | 386 | { |
85cf9dba | 387 | if (!x86_pmu_initialized()) |
2b9ff0db | 388 | return; |
9e35ad38 | 389 | x86_pmu.enable_all(); |
ee06094f | 390 | } |
ee06094f | 391 | |
19d84dab | 392 | static inline u64 intel_pmu_get_status(void) |
b0f3f28e PZ |
393 | { |
394 | u64 status; | |
395 | ||
b7f8859a | 396 | rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status); |
b0f3f28e | 397 | |
b7f8859a | 398 | return status; |
b0f3f28e PZ |
399 | } |
400 | ||
dee5d906 | 401 | static inline void intel_pmu_ack_status(u64 ack) |
b0f3f28e PZ |
402 | { |
403 | wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack); | |
404 | } | |
405 | ||
7c90cc45 | 406 | static inline void x86_pmu_enable_counter(struct hw_perf_counter *hwc, int idx) |
b0f3f28e | 407 | { |
7c90cc45 | 408 | int err; |
7c90cc45 RR |
409 | err = checking_wrmsrl(hwc->config_base + idx, |
410 | hwc->config | ARCH_PERFMON_EVENTSEL0_ENABLE); | |
b0f3f28e PZ |
411 | } |
412 | ||
d4369891 | 413 | static inline void x86_pmu_disable_counter(struct hw_perf_counter *hwc, int idx) |
b0f3f28e | 414 | { |
d4369891 | 415 | int err; |
d4369891 RR |
416 | err = checking_wrmsrl(hwc->config_base + idx, |
417 | hwc->config); | |
b0f3f28e PZ |
418 | } |
419 | ||
2f18d1e8 | 420 | static inline void |
d4369891 | 421 | intel_pmu_disable_fixed(struct hw_perf_counter *hwc, int __idx) |
2f18d1e8 IM |
422 | { |
423 | int idx = __idx - X86_PMC_IDX_FIXED; | |
424 | u64 ctrl_val, mask; | |
425 | int err; | |
426 | ||
427 | mask = 0xfULL << (idx * 4); | |
428 | ||
429 | rdmsrl(hwc->config_base, ctrl_val); | |
430 | ctrl_val &= ~mask; | |
431 | err = checking_wrmsrl(hwc->config_base, ctrl_val); | |
432 | } | |
433 | ||
7e2ae347 | 434 | static inline void |
d4369891 | 435 | intel_pmu_disable_counter(struct hw_perf_counter *hwc, int idx) |
7e2ae347 | 436 | { |
d4369891 RR |
437 | if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { |
438 | intel_pmu_disable_fixed(hwc, idx); | |
439 | return; | |
440 | } | |
441 | ||
442 | x86_pmu_disable_counter(hwc, idx); | |
443 | } | |
444 | ||
445 | static inline void | |
446 | amd_pmu_disable_counter(struct hw_perf_counter *hwc, int idx) | |
447 | { | |
448 | x86_pmu_disable_counter(hwc, idx); | |
7e2ae347 IM |
449 | } |
450 | ||
2f18d1e8 | 451 | static DEFINE_PER_CPU(u64, prev_left[X86_PMC_IDX_MAX]); |
241771ef | 452 | |
ee06094f IM |
453 | /* |
454 | * Set the next IRQ period, based on the hwc->period_left value. | |
455 | * To be called with the counter disabled in hw: | |
456 | */ | |
457 | static void | |
26816c28 | 458 | x86_perf_counter_set_period(struct perf_counter *counter, |
ee06094f | 459 | struct hw_perf_counter *hwc, int idx) |
241771ef | 460 | { |
2f18d1e8 | 461 | s64 left = atomic64_read(&hwc->period_left); |
60db5e09 | 462 | s64 period = min(x86_pmu.max_period, hwc->irq_period); |
2f18d1e8 | 463 | int err; |
ee06094f | 464 | |
ee06094f IM |
465 | /* |
466 | * If we are way outside a reasoable range then just skip forward: | |
467 | */ | |
468 | if (unlikely(left <= -period)) { | |
469 | left = period; | |
470 | atomic64_set(&hwc->period_left, left); | |
471 | } | |
472 | ||
473 | if (unlikely(left <= 0)) { | |
474 | left += period; | |
475 | atomic64_set(&hwc->period_left, left); | |
476 | } | |
1c80f4b5 IM |
477 | /* |
478 | * Quirk: certain CPUs dont like it if just 1 event is left: | |
479 | */ | |
480 | if (unlikely(left < 2)) | |
481 | left = 2; | |
241771ef | 482 | |
ee06094f IM |
483 | per_cpu(prev_left[idx], smp_processor_id()) = left; |
484 | ||
485 | /* | |
486 | * The hw counter starts counting from this counter offset, | |
487 | * mark it to be able to extra future deltas: | |
488 | */ | |
2f18d1e8 | 489 | atomic64_set(&hwc->prev_count, (u64)-left); |
ee06094f | 490 | |
2f18d1e8 | 491 | err = checking_wrmsrl(hwc->counter_base + idx, |
0933e5c6 | 492 | (u64)(-left) & x86_pmu.counter_mask); |
2f18d1e8 IM |
493 | } |
494 | ||
495 | static inline void | |
7c90cc45 | 496 | intel_pmu_enable_fixed(struct hw_perf_counter *hwc, int __idx) |
2f18d1e8 IM |
497 | { |
498 | int idx = __idx - X86_PMC_IDX_FIXED; | |
499 | u64 ctrl_val, bits, mask; | |
500 | int err; | |
501 | ||
502 | /* | |
0475f9ea PM |
503 | * Enable IRQ generation (0x8), |
504 | * and enable ring-3 counting (0x2) and ring-0 counting (0x1) | |
505 | * if requested: | |
2f18d1e8 | 506 | */ |
0475f9ea PM |
507 | bits = 0x8ULL; |
508 | if (hwc->config & ARCH_PERFMON_EVENTSEL_USR) | |
509 | bits |= 0x2; | |
2f18d1e8 IM |
510 | if (hwc->config & ARCH_PERFMON_EVENTSEL_OS) |
511 | bits |= 0x1; | |
512 | bits <<= (idx * 4); | |
513 | mask = 0xfULL << (idx * 4); | |
514 | ||
515 | rdmsrl(hwc->config_base, ctrl_val); | |
516 | ctrl_val &= ~mask; | |
517 | ctrl_val |= bits; | |
518 | err = checking_wrmsrl(hwc->config_base, ctrl_val); | |
7e2ae347 IM |
519 | } |
520 | ||
7c90cc45 | 521 | static void intel_pmu_enable_counter(struct hw_perf_counter *hwc, int idx) |
7e2ae347 | 522 | { |
7c90cc45 RR |
523 | if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { |
524 | intel_pmu_enable_fixed(hwc, idx); | |
525 | return; | |
526 | } | |
527 | ||
528 | x86_pmu_enable_counter(hwc, idx); | |
529 | } | |
530 | ||
531 | static void amd_pmu_enable_counter(struct hw_perf_counter *hwc, int idx) | |
532 | { | |
533 | struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters); | |
534 | ||
535 | if (cpuc->enabled) | |
536 | x86_pmu_enable_counter(hwc, idx); | |
2b583d8b | 537 | else |
d4369891 | 538 | x86_pmu_disable_counter(hwc, idx); |
241771ef IM |
539 | } |
540 | ||
2f18d1e8 IM |
541 | static int |
542 | fixed_mode_idx(struct perf_counter *counter, struct hw_perf_counter *hwc) | |
862a1a5f | 543 | { |
2f18d1e8 IM |
544 | unsigned int event; |
545 | ||
ef7b3e09 | 546 | if (!x86_pmu.num_counters_fixed) |
f87ad35d JSR |
547 | return -1; |
548 | ||
2f18d1e8 IM |
549 | if (unlikely(hwc->nmi)) |
550 | return -1; | |
551 | ||
552 | event = hwc->config & ARCH_PERFMON_EVENT_MASK; | |
553 | ||
4a06bd85 | 554 | if (unlikely(event == x86_pmu.event_map(PERF_COUNT_INSTRUCTIONS))) |
2f18d1e8 | 555 | return X86_PMC_IDX_FIXED_INSTRUCTIONS; |
4a06bd85 | 556 | if (unlikely(event == x86_pmu.event_map(PERF_COUNT_CPU_CYCLES))) |
2f18d1e8 | 557 | return X86_PMC_IDX_FIXED_CPU_CYCLES; |
4a06bd85 | 558 | if (unlikely(event == x86_pmu.event_map(PERF_COUNT_BUS_CYCLES))) |
2f18d1e8 IM |
559 | return X86_PMC_IDX_FIXED_BUS_CYCLES; |
560 | ||
862a1a5f IM |
561 | return -1; |
562 | } | |
563 | ||
ee06094f IM |
564 | /* |
565 | * Find a PMC slot for the freshly enabled / scheduled in counter: | |
566 | */ | |
4aeb0b42 | 567 | static int x86_pmu_enable(struct perf_counter *counter) |
241771ef IM |
568 | { |
569 | struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters); | |
570 | struct hw_perf_counter *hwc = &counter->hw; | |
2f18d1e8 | 571 | int idx; |
241771ef | 572 | |
2f18d1e8 IM |
573 | idx = fixed_mode_idx(counter, hwc); |
574 | if (idx >= 0) { | |
575 | /* | |
576 | * Try to get the fixed counter, if that is already taken | |
577 | * then try to get a generic counter: | |
578 | */ | |
43f6201a | 579 | if (test_and_set_bit(idx, cpuc->used_mask)) |
2f18d1e8 | 580 | goto try_generic; |
0dff86aa | 581 | |
2f18d1e8 IM |
582 | hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL; |
583 | /* | |
584 | * We set it so that counter_base + idx in wrmsr/rdmsr maps to | |
585 | * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2: | |
586 | */ | |
587 | hwc->counter_base = | |
588 | MSR_ARCH_PERFMON_FIXED_CTR0 - X86_PMC_IDX_FIXED; | |
241771ef | 589 | hwc->idx = idx; |
2f18d1e8 IM |
590 | } else { |
591 | idx = hwc->idx; | |
592 | /* Try to get the previous generic counter again */ | |
43f6201a | 593 | if (test_and_set_bit(idx, cpuc->used_mask)) { |
2f18d1e8 | 594 | try_generic: |
43f6201a | 595 | idx = find_first_zero_bit(cpuc->used_mask, |
0933e5c6 RR |
596 | x86_pmu.num_counters); |
597 | if (idx == x86_pmu.num_counters) | |
2f18d1e8 IM |
598 | return -EAGAIN; |
599 | ||
43f6201a | 600 | set_bit(idx, cpuc->used_mask); |
2f18d1e8 IM |
601 | hwc->idx = idx; |
602 | } | |
4a06bd85 RR |
603 | hwc->config_base = x86_pmu.eventsel; |
604 | hwc->counter_base = x86_pmu.perfctr; | |
241771ef IM |
605 | } |
606 | ||
d4369891 | 607 | x86_pmu.disable(hwc, idx); |
241771ef | 608 | |
862a1a5f | 609 | cpuc->counters[idx] = counter; |
43f6201a | 610 | set_bit(idx, cpuc->active_mask); |
7e2ae347 | 611 | |
26816c28 | 612 | x86_perf_counter_set_period(counter, hwc, idx); |
7c90cc45 | 613 | x86_pmu.enable(hwc, idx); |
95cdd2e7 IM |
614 | |
615 | return 0; | |
241771ef IM |
616 | } |
617 | ||
618 | void perf_counter_print_debug(void) | |
619 | { | |
2f18d1e8 | 620 | u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed; |
0dff86aa | 621 | struct cpu_hw_counters *cpuc; |
5bb9efe3 | 622 | unsigned long flags; |
1e125676 IM |
623 | int cpu, idx; |
624 | ||
0933e5c6 | 625 | if (!x86_pmu.num_counters) |
1e125676 | 626 | return; |
241771ef | 627 | |
5bb9efe3 | 628 | local_irq_save(flags); |
241771ef IM |
629 | |
630 | cpu = smp_processor_id(); | |
0dff86aa | 631 | cpuc = &per_cpu(cpu_hw_counters, cpu); |
241771ef | 632 | |
faa28ae0 | 633 | if (x86_pmu.version >= 2) { |
a1ef58f4 JSR |
634 | rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl); |
635 | rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status); | |
636 | rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow); | |
637 | rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed); | |
638 | ||
639 | pr_info("\n"); | |
640 | pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl); | |
641 | pr_info("CPU#%d: status: %016llx\n", cpu, status); | |
642 | pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow); | |
643 | pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed); | |
f87ad35d | 644 | } |
43f6201a | 645 | pr_info("CPU#%d: used: %016llx\n", cpu, *(u64 *)cpuc->used_mask); |
241771ef | 646 | |
0933e5c6 | 647 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
4a06bd85 RR |
648 | rdmsrl(x86_pmu.eventsel + idx, pmc_ctrl); |
649 | rdmsrl(x86_pmu.perfctr + idx, pmc_count); | |
241771ef | 650 | |
ee06094f | 651 | prev_left = per_cpu(prev_left[idx], cpu); |
241771ef | 652 | |
a1ef58f4 | 653 | pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n", |
241771ef | 654 | cpu, idx, pmc_ctrl); |
a1ef58f4 | 655 | pr_info("CPU#%d: gen-PMC%d count: %016llx\n", |
241771ef | 656 | cpu, idx, pmc_count); |
a1ef58f4 | 657 | pr_info("CPU#%d: gen-PMC%d left: %016llx\n", |
ee06094f | 658 | cpu, idx, prev_left); |
241771ef | 659 | } |
0933e5c6 | 660 | for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) { |
2f18d1e8 IM |
661 | rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count); |
662 | ||
a1ef58f4 | 663 | pr_info("CPU#%d: fixed-PMC%d count: %016llx\n", |
2f18d1e8 IM |
664 | cpu, idx, pmc_count); |
665 | } | |
5bb9efe3 | 666 | local_irq_restore(flags); |
241771ef IM |
667 | } |
668 | ||
4aeb0b42 | 669 | static void x86_pmu_disable(struct perf_counter *counter) |
241771ef IM |
670 | { |
671 | struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters); | |
672 | struct hw_perf_counter *hwc = &counter->hw; | |
6f00cada | 673 | int idx = hwc->idx; |
241771ef | 674 | |
09534238 RR |
675 | /* |
676 | * Must be done before we disable, otherwise the nmi handler | |
677 | * could reenable again: | |
678 | */ | |
43f6201a | 679 | clear_bit(idx, cpuc->active_mask); |
d4369891 | 680 | x86_pmu.disable(hwc, idx); |
241771ef | 681 | |
2f18d1e8 IM |
682 | /* |
683 | * Make sure the cleared pointer becomes visible before we | |
684 | * (potentially) free the counter: | |
685 | */ | |
527e26af | 686 | barrier(); |
241771ef | 687 | |
ee06094f IM |
688 | /* |
689 | * Drain the remaining delta count out of a counter | |
690 | * that we are disabling: | |
691 | */ | |
692 | x86_perf_counter_update(counter, hwc, idx); | |
09534238 | 693 | cpuc->counters[idx] = NULL; |
43f6201a | 694 | clear_bit(idx, cpuc->used_mask); |
241771ef IM |
695 | } |
696 | ||
7e2ae347 | 697 | /* |
ee06094f IM |
698 | * Save and restart an expired counter. Called by NMI contexts, |
699 | * so it has to be careful about preempting normal counter ops: | |
7e2ae347 | 700 | */ |
55de0f2e | 701 | static void intel_pmu_save_and_restart(struct perf_counter *counter) |
241771ef IM |
702 | { |
703 | struct hw_perf_counter *hwc = &counter->hw; | |
704 | int idx = hwc->idx; | |
241771ef | 705 | |
ee06094f | 706 | x86_perf_counter_update(counter, hwc, idx); |
26816c28 | 707 | x86_perf_counter_set_period(counter, hwc, idx); |
7e2ae347 | 708 | |
2f18d1e8 | 709 | if (counter->state == PERF_COUNTER_STATE_ACTIVE) |
7c90cc45 | 710 | intel_pmu_enable_counter(hwc, idx); |
241771ef IM |
711 | } |
712 | ||
4b39fd96 MG |
713 | /* |
714 | * Maximum interrupt frequency of 100KHz per CPU | |
715 | */ | |
169e41eb | 716 | #define PERFMON_MAX_INTERRUPTS (100000/HZ) |
4b39fd96 | 717 | |
241771ef IM |
718 | /* |
719 | * This handler is triggered by the local APIC, so the APIC IRQ handling | |
720 | * rules apply: | |
721 | */ | |
39d81eab | 722 | static int intel_pmu_handle_irq(struct pt_regs *regs, int nmi) |
241771ef | 723 | { |
9029a5e3 IM |
724 | struct cpu_hw_counters *cpuc; |
725 | struct cpu_hw_counters; | |
726 | int bit, cpu, loops; | |
4b39fd96 | 727 | u64 ack, status; |
9029a5e3 IM |
728 | |
729 | cpu = smp_processor_id(); | |
730 | cpuc = &per_cpu(cpu_hw_counters, cpu); | |
241771ef | 731 | |
9e35ad38 | 732 | perf_disable(); |
19d84dab | 733 | status = intel_pmu_get_status(); |
9e35ad38 PZ |
734 | if (!status) { |
735 | perf_enable(); | |
736 | return 0; | |
737 | } | |
87b9cf46 | 738 | |
9029a5e3 | 739 | loops = 0; |
241771ef | 740 | again: |
9029a5e3 IM |
741 | if (++loops > 100) { |
742 | WARN_ONCE(1, "perfcounters: irq loop stuck!\n"); | |
743 | return 1; | |
744 | } | |
745 | ||
d278c484 | 746 | inc_irq_stat(apic_perf_irqs); |
241771ef | 747 | ack = status; |
2f18d1e8 | 748 | for_each_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) { |
862a1a5f | 749 | struct perf_counter *counter = cpuc->counters[bit]; |
241771ef IM |
750 | |
751 | clear_bit(bit, (unsigned long *) &status); | |
43f6201a | 752 | if (!test_bit(bit, cpuc->active_mask)) |
241771ef IM |
753 | continue; |
754 | ||
55de0f2e | 755 | intel_pmu_save_and_restart(counter); |
78f13e95 | 756 | if (perf_counter_overflow(counter, nmi, regs, 0)) |
d4369891 | 757 | intel_pmu_disable_counter(&counter->hw, bit); |
241771ef IM |
758 | } |
759 | ||
dee5d906 | 760 | intel_pmu_ack_status(ack); |
241771ef IM |
761 | |
762 | /* | |
763 | * Repeat if there is more work to be done: | |
764 | */ | |
19d84dab | 765 | status = intel_pmu_get_status(); |
241771ef IM |
766 | if (status) |
767 | goto again; | |
b0f3f28e | 768 | |
9e35ad38 PZ |
769 | if (++cpuc->interrupts != PERFMON_MAX_INTERRUPTS) |
770 | perf_enable(); | |
771 | ||
772 | return 1; | |
1b023a96 MG |
773 | } |
774 | ||
a29aa8a7 RR |
775 | static int amd_pmu_handle_irq(struct pt_regs *regs, int nmi) |
776 | { | |
9029a5e3 IM |
777 | int cpu, idx, throttle = 0, handled = 0; |
778 | struct cpu_hw_counters *cpuc; | |
a29aa8a7 RR |
779 | struct perf_counter *counter; |
780 | struct hw_perf_counter *hwc; | |
9029a5e3 IM |
781 | u64 val; |
782 | ||
783 | cpu = smp_processor_id(); | |
784 | cpuc = &per_cpu(cpu_hw_counters, cpu); | |
962bf7a6 | 785 | |
9e35ad38 PZ |
786 | if (++cpuc->interrupts == PERFMON_MAX_INTERRUPTS) { |
787 | throttle = 1; | |
788 | __perf_disable(); | |
789 | cpuc->enabled = 0; | |
790 | barrier(); | |
962bf7a6 | 791 | } |
a29aa8a7 | 792 | |
a29aa8a7 | 793 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
962bf7a6 PZ |
794 | int disable = 0; |
795 | ||
43f6201a | 796 | if (!test_bit(idx, cpuc->active_mask)) |
a29aa8a7 | 797 | continue; |
962bf7a6 | 798 | |
a29aa8a7 RR |
799 | counter = cpuc->counters[idx]; |
800 | hwc = &counter->hw; | |
a4016a79 PZ |
801 | |
802 | if (counter->hw_event.nmi != nmi) | |
803 | goto next; | |
804 | ||
4b7bfd0d | 805 | val = x86_perf_counter_update(counter, hwc, idx); |
a29aa8a7 | 806 | if (val & (1ULL << (x86_pmu.counter_bits - 1))) |
962bf7a6 PZ |
807 | goto next; |
808 | ||
a29aa8a7 RR |
809 | /* counter overflow */ |
810 | x86_perf_counter_set_period(counter, hwc, idx); | |
811 | handled = 1; | |
812 | inc_irq_stat(apic_perf_irqs); | |
962bf7a6 PZ |
813 | disable = perf_counter_overflow(counter, nmi, regs, 0); |
814 | ||
815 | next: | |
816 | if (disable || throttle) | |
a29aa8a7 | 817 | amd_pmu_disable_counter(hwc, idx); |
a29aa8a7 | 818 | } |
962bf7a6 | 819 | |
a29aa8a7 RR |
820 | return handled; |
821 | } | |
39d81eab | 822 | |
1b023a96 MG |
823 | void perf_counter_unthrottle(void) |
824 | { | |
825 | struct cpu_hw_counters *cpuc; | |
826 | ||
85cf9dba | 827 | if (!x86_pmu_initialized()) |
1b023a96 MG |
828 | return; |
829 | ||
b0f3f28e | 830 | cpuc = &__get_cpu_var(cpu_hw_counters); |
4b39fd96 | 831 | if (cpuc->interrupts >= PERFMON_MAX_INTERRUPTS) { |
f5a5a2f6 IM |
832 | /* |
833 | * Clear them before re-enabling irqs/NMIs again: | |
834 | */ | |
835 | cpuc->interrupts = 0; | |
9e35ad38 | 836 | perf_enable(); |
f5a5a2f6 IM |
837 | } else { |
838 | cpuc->interrupts = 0; | |
1b023a96 | 839 | } |
241771ef IM |
840 | } |
841 | ||
842 | void smp_perf_counter_interrupt(struct pt_regs *regs) | |
843 | { | |
844 | irq_enter(); | |
241771ef | 845 | apic_write(APIC_LVTPC, LOCAL_PERF_VECTOR); |
b0f3f28e | 846 | ack_APIC_irq(); |
4a06bd85 | 847 | x86_pmu.handle_irq(regs, 0); |
241771ef IM |
848 | irq_exit(); |
849 | } | |
850 | ||
b6276f35 PZ |
851 | void smp_perf_pending_interrupt(struct pt_regs *regs) |
852 | { | |
853 | irq_enter(); | |
854 | ack_APIC_irq(); | |
855 | inc_irq_stat(apic_pending_irqs); | |
856 | perf_counter_do_pending(); | |
857 | irq_exit(); | |
858 | } | |
859 | ||
860 | void set_perf_counter_pending(void) | |
861 | { | |
862 | apic->send_IPI_self(LOCAL_PENDING_VECTOR); | |
863 | } | |
864 | ||
3415dd91 | 865 | void perf_counters_lapic_init(int nmi) |
241771ef IM |
866 | { |
867 | u32 apic_val; | |
868 | ||
85cf9dba | 869 | if (!x86_pmu_initialized()) |
241771ef | 870 | return; |
85cf9dba | 871 | |
241771ef IM |
872 | /* |
873 | * Enable the performance counter vector in the APIC LVT: | |
874 | */ | |
875 | apic_val = apic_read(APIC_LVTERR); | |
876 | ||
877 | apic_write(APIC_LVTERR, apic_val | APIC_LVT_MASKED); | |
878 | if (nmi) | |
879 | apic_write(APIC_LVTPC, APIC_DM_NMI); | |
880 | else | |
881 | apic_write(APIC_LVTPC, LOCAL_PERF_VECTOR); | |
882 | apic_write(APIC_LVTERR, apic_val); | |
883 | } | |
884 | ||
885 | static int __kprobes | |
886 | perf_counter_nmi_handler(struct notifier_block *self, | |
887 | unsigned long cmd, void *__args) | |
888 | { | |
889 | struct die_args *args = __args; | |
890 | struct pt_regs *regs; | |
b0f3f28e | 891 | |
ba77813a | 892 | if (!atomic_read(&active_counters)) |
63a809a2 PZ |
893 | return NOTIFY_DONE; |
894 | ||
b0f3f28e PZ |
895 | switch (cmd) { |
896 | case DIE_NMI: | |
897 | case DIE_NMI_IPI: | |
898 | break; | |
241771ef | 899 | |
b0f3f28e | 900 | default: |
241771ef | 901 | return NOTIFY_DONE; |
b0f3f28e | 902 | } |
241771ef IM |
903 | |
904 | regs = args->regs; | |
905 | ||
906 | apic_write(APIC_LVTPC, APIC_DM_NMI); | |
a4016a79 PZ |
907 | /* |
908 | * Can't rely on the handled return value to say it was our NMI, two | |
909 | * counters could trigger 'simultaneously' raising two back-to-back NMIs. | |
910 | * | |
911 | * If the first NMI handles both, the latter will be empty and daze | |
912 | * the CPU. | |
913 | */ | |
914 | x86_pmu.handle_irq(regs, 1); | |
241771ef | 915 | |
a4016a79 | 916 | return NOTIFY_STOP; |
241771ef IM |
917 | } |
918 | ||
919 | static __read_mostly struct notifier_block perf_counter_nmi_notifier = { | |
5b75af0a MG |
920 | .notifier_call = perf_counter_nmi_handler, |
921 | .next = NULL, | |
922 | .priority = 1 | |
241771ef IM |
923 | }; |
924 | ||
5f4ec28f | 925 | static struct x86_pmu intel_pmu = { |
faa28ae0 | 926 | .name = "Intel", |
39d81eab | 927 | .handle_irq = intel_pmu_handle_irq, |
9e35ad38 PZ |
928 | .disable_all = intel_pmu_disable_all, |
929 | .enable_all = intel_pmu_enable_all, | |
5f4ec28f RR |
930 | .enable = intel_pmu_enable_counter, |
931 | .disable = intel_pmu_disable_counter, | |
b56a3802 JSR |
932 | .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, |
933 | .perfctr = MSR_ARCH_PERFMON_PERFCTR0, | |
5f4ec28f RR |
934 | .event_map = intel_pmu_event_map, |
935 | .raw_event = intel_pmu_raw_event, | |
b56a3802 | 936 | .max_events = ARRAY_SIZE(intel_perfmon_event_map), |
c619b8ff RR |
937 | /* |
938 | * Intel PMCs cannot be accessed sanely above 32 bit width, | |
939 | * so we install an artificial 1<<31 period regardless of | |
940 | * the generic counter period: | |
941 | */ | |
942 | .max_period = (1ULL << 31) - 1, | |
b56a3802 JSR |
943 | }; |
944 | ||
5f4ec28f | 945 | static struct x86_pmu amd_pmu = { |
faa28ae0 | 946 | .name = "AMD", |
39d81eab | 947 | .handle_irq = amd_pmu_handle_irq, |
9e35ad38 PZ |
948 | .disable_all = amd_pmu_disable_all, |
949 | .enable_all = amd_pmu_enable_all, | |
5f4ec28f RR |
950 | .enable = amd_pmu_enable_counter, |
951 | .disable = amd_pmu_disable_counter, | |
f87ad35d JSR |
952 | .eventsel = MSR_K7_EVNTSEL0, |
953 | .perfctr = MSR_K7_PERFCTR0, | |
5f4ec28f RR |
954 | .event_map = amd_pmu_event_map, |
955 | .raw_event = amd_pmu_raw_event, | |
f87ad35d | 956 | .max_events = ARRAY_SIZE(amd_perfmon_event_map), |
0933e5c6 RR |
957 | .num_counters = 4, |
958 | .counter_bits = 48, | |
959 | .counter_mask = (1ULL << 48) - 1, | |
c619b8ff RR |
960 | /* use highest bit to detect overflow */ |
961 | .max_period = (1ULL << 47) - 1, | |
f87ad35d JSR |
962 | }; |
963 | ||
72eae04d | 964 | static int intel_pmu_init(void) |
241771ef | 965 | { |
7bb497bd | 966 | union cpuid10_edx edx; |
241771ef | 967 | union cpuid10_eax eax; |
703e937c | 968 | unsigned int unused; |
7bb497bd | 969 | unsigned int ebx; |
faa28ae0 | 970 | int version; |
241771ef | 971 | |
da1a776b | 972 | if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) |
72eae04d | 973 | return -ENODEV; |
da1a776b | 974 | |
241771ef IM |
975 | /* |
976 | * Check whether the Architectural PerfMon supports | |
977 | * Branch Misses Retired Event or not. | |
978 | */ | |
703e937c | 979 | cpuid(10, &eax.full, &ebx, &unused, &edx.full); |
241771ef | 980 | if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED) |
72eae04d | 981 | return -ENODEV; |
241771ef | 982 | |
faa28ae0 RR |
983 | version = eax.split.version_id; |
984 | if (version < 2) | |
72eae04d | 985 | return -ENODEV; |
7bb497bd | 986 | |
4a06bd85 | 987 | x86_pmu = intel_pmu; |
faa28ae0 | 988 | x86_pmu.version = version; |
0933e5c6 | 989 | x86_pmu.num_counters = eax.split.num_counters; |
066d7dea IM |
990 | |
991 | /* | |
992 | * Quirk: v2 perfmon does not report fixed-purpose counters, so | |
993 | * assume at least 3 counters: | |
994 | */ | |
995 | x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3); | |
996 | ||
0933e5c6 RR |
997 | x86_pmu.counter_bits = eax.split.bit_width; |
998 | x86_pmu.counter_mask = (1ULL << eax.split.bit_width) - 1; | |
b56a3802 | 999 | |
9e35ad38 PZ |
1000 | rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl); |
1001 | ||
72eae04d | 1002 | return 0; |
b56a3802 JSR |
1003 | } |
1004 | ||
72eae04d | 1005 | static int amd_pmu_init(void) |
f87ad35d | 1006 | { |
4a06bd85 | 1007 | x86_pmu = amd_pmu; |
72eae04d | 1008 | return 0; |
f87ad35d JSR |
1009 | } |
1010 | ||
b56a3802 JSR |
1011 | void __init init_hw_perf_counters(void) |
1012 | { | |
72eae04d RR |
1013 | int err; |
1014 | ||
b56a3802 JSR |
1015 | switch (boot_cpu_data.x86_vendor) { |
1016 | case X86_VENDOR_INTEL: | |
72eae04d | 1017 | err = intel_pmu_init(); |
b56a3802 | 1018 | break; |
f87ad35d | 1019 | case X86_VENDOR_AMD: |
72eae04d | 1020 | err = amd_pmu_init(); |
f87ad35d | 1021 | break; |
4138960a RR |
1022 | default: |
1023 | return; | |
b56a3802 | 1024 | } |
72eae04d | 1025 | if (err != 0) |
b56a3802 JSR |
1026 | return; |
1027 | ||
faa28ae0 RR |
1028 | pr_info("%s Performance Monitoring support detected.\n", x86_pmu.name); |
1029 | pr_info("... version: %d\n", x86_pmu.version); | |
1030 | pr_info("... bit width: %d\n", x86_pmu.counter_bits); | |
1031 | ||
0933e5c6 RR |
1032 | pr_info("... num counters: %d\n", x86_pmu.num_counters); |
1033 | if (x86_pmu.num_counters > X86_PMC_MAX_GENERIC) { | |
1034 | x86_pmu.num_counters = X86_PMC_MAX_GENERIC; | |
241771ef | 1035 | WARN(1, KERN_ERR "hw perf counters %d > max(%d), clipping!", |
0933e5c6 | 1036 | x86_pmu.num_counters, X86_PMC_MAX_GENERIC); |
241771ef | 1037 | } |
0933e5c6 RR |
1038 | perf_counter_mask = (1 << x86_pmu.num_counters) - 1; |
1039 | perf_max_counters = x86_pmu.num_counters; | |
241771ef | 1040 | |
0933e5c6 | 1041 | pr_info("... value mask: %016Lx\n", x86_pmu.counter_mask); |
c619b8ff | 1042 | pr_info("... max period: %016Lx\n", x86_pmu.max_period); |
2f18d1e8 | 1043 | |
0933e5c6 RR |
1044 | if (x86_pmu.num_counters_fixed > X86_PMC_MAX_FIXED) { |
1045 | x86_pmu.num_counters_fixed = X86_PMC_MAX_FIXED; | |
703e937c | 1046 | WARN(1, KERN_ERR "hw perf counters fixed %d > max(%d), clipping!", |
0933e5c6 | 1047 | x86_pmu.num_counters_fixed, X86_PMC_MAX_FIXED); |
703e937c | 1048 | } |
0933e5c6 | 1049 | pr_info("... fixed counters: %d\n", x86_pmu.num_counters_fixed); |
862a1a5f | 1050 | |
0933e5c6 RR |
1051 | perf_counter_mask |= |
1052 | ((1LL << x86_pmu.num_counters_fixed)-1) << X86_PMC_IDX_FIXED; | |
241771ef | 1053 | |
a1ef58f4 | 1054 | pr_info("... counter mask: %016Lx\n", perf_counter_mask); |
75f224cf | 1055 | |
b68f1d2e | 1056 | perf_counters_lapic_init(1); |
241771ef | 1057 | register_die_notifier(&perf_counter_nmi_notifier); |
241771ef | 1058 | } |
621a01ea | 1059 | |
bb775fc2 | 1060 | static inline void x86_pmu_read(struct perf_counter *counter) |
ee06094f IM |
1061 | { |
1062 | x86_perf_counter_update(counter, &counter->hw, counter->hw.idx); | |
1063 | } | |
1064 | ||
4aeb0b42 RR |
1065 | static const struct pmu pmu = { |
1066 | .enable = x86_pmu_enable, | |
1067 | .disable = x86_pmu_disable, | |
1068 | .read = x86_pmu_read, | |
621a01ea IM |
1069 | }; |
1070 | ||
4aeb0b42 | 1071 | const struct pmu *hw_perf_counter_init(struct perf_counter *counter) |
621a01ea IM |
1072 | { |
1073 | int err; | |
1074 | ||
1075 | err = __hw_perf_counter_init(counter); | |
1076 | if (err) | |
9ea98e19 | 1077 | return ERR_PTR(err); |
621a01ea | 1078 | |
4aeb0b42 | 1079 | return &pmu; |
621a01ea | 1080 | } |
d7d59fb3 PZ |
1081 | |
1082 | /* | |
1083 | * callchain support | |
1084 | */ | |
1085 | ||
1086 | static inline | |
1087 | void callchain_store(struct perf_callchain_entry *entry, unsigned long ip) | |
1088 | { | |
1089 | if (entry->nr < MAX_STACK_DEPTH) | |
1090 | entry->ip[entry->nr++] = ip; | |
1091 | } | |
1092 | ||
1093 | static DEFINE_PER_CPU(struct perf_callchain_entry, irq_entry); | |
1094 | static DEFINE_PER_CPU(struct perf_callchain_entry, nmi_entry); | |
1095 | ||
1096 | ||
1097 | static void | |
1098 | backtrace_warning_symbol(void *data, char *msg, unsigned long symbol) | |
1099 | { | |
1100 | /* Ignore warnings */ | |
1101 | } | |
1102 | ||
1103 | static void backtrace_warning(void *data, char *msg) | |
1104 | { | |
1105 | /* Ignore warnings */ | |
1106 | } | |
1107 | ||
1108 | static int backtrace_stack(void *data, char *name) | |
1109 | { | |
1110 | /* Don't bother with IRQ stacks for now */ | |
1111 | return -1; | |
1112 | } | |
1113 | ||
1114 | static void backtrace_address(void *data, unsigned long addr, int reliable) | |
1115 | { | |
1116 | struct perf_callchain_entry *entry = data; | |
1117 | ||
1118 | if (reliable) | |
1119 | callchain_store(entry, addr); | |
1120 | } | |
1121 | ||
1122 | static const struct stacktrace_ops backtrace_ops = { | |
1123 | .warning = backtrace_warning, | |
1124 | .warning_symbol = backtrace_warning_symbol, | |
1125 | .stack = backtrace_stack, | |
1126 | .address = backtrace_address, | |
1127 | }; | |
1128 | ||
1129 | static void | |
1130 | perf_callchain_kernel(struct pt_regs *regs, struct perf_callchain_entry *entry) | |
1131 | { | |
1132 | unsigned long bp; | |
1133 | char *stack; | |
5872bdb8 | 1134 | int nr = entry->nr; |
d7d59fb3 PZ |
1135 | |
1136 | callchain_store(entry, instruction_pointer(regs)); | |
1137 | ||
1138 | stack = ((char *)regs + sizeof(struct pt_regs)); | |
1139 | #ifdef CONFIG_FRAME_POINTER | |
1140 | bp = frame_pointer(regs); | |
1141 | #else | |
1142 | bp = 0; | |
1143 | #endif | |
1144 | ||
1145 | dump_trace(NULL, regs, (void *)stack, bp, &backtrace_ops, entry); | |
5872bdb8 PZ |
1146 | |
1147 | entry->kernel = entry->nr - nr; | |
d7d59fb3 PZ |
1148 | } |
1149 | ||
1150 | ||
1151 | struct stack_frame { | |
1152 | const void __user *next_fp; | |
1153 | unsigned long return_address; | |
1154 | }; | |
1155 | ||
1156 | static int copy_stack_frame(const void __user *fp, struct stack_frame *frame) | |
1157 | { | |
1158 | int ret; | |
1159 | ||
1160 | if (!access_ok(VERIFY_READ, fp, sizeof(*frame))) | |
1161 | return 0; | |
1162 | ||
1163 | ret = 1; | |
1164 | pagefault_disable(); | |
1165 | if (__copy_from_user_inatomic(frame, fp, sizeof(*frame))) | |
1166 | ret = 0; | |
1167 | pagefault_enable(); | |
1168 | ||
1169 | return ret; | |
1170 | } | |
1171 | ||
1172 | static void | |
1173 | perf_callchain_user(struct pt_regs *regs, struct perf_callchain_entry *entry) | |
1174 | { | |
1175 | struct stack_frame frame; | |
1176 | const void __user *fp; | |
5872bdb8 | 1177 | int nr = entry->nr; |
d7d59fb3 PZ |
1178 | |
1179 | regs = (struct pt_regs *)current->thread.sp0 - 1; | |
1180 | fp = (void __user *)regs->bp; | |
1181 | ||
1182 | callchain_store(entry, regs->ip); | |
1183 | ||
1184 | while (entry->nr < MAX_STACK_DEPTH) { | |
1185 | frame.next_fp = NULL; | |
1186 | frame.return_address = 0; | |
1187 | ||
1188 | if (!copy_stack_frame(fp, &frame)) | |
1189 | break; | |
1190 | ||
1191 | if ((unsigned long)fp < user_stack_pointer(regs)) | |
1192 | break; | |
1193 | ||
1194 | callchain_store(entry, frame.return_address); | |
1195 | fp = frame.next_fp; | |
1196 | } | |
5872bdb8 PZ |
1197 | |
1198 | entry->user = entry->nr - nr; | |
d7d59fb3 PZ |
1199 | } |
1200 | ||
1201 | static void | |
1202 | perf_do_callchain(struct pt_regs *regs, struct perf_callchain_entry *entry) | |
1203 | { | |
1204 | int is_user; | |
1205 | ||
1206 | if (!regs) | |
1207 | return; | |
1208 | ||
1209 | is_user = user_mode(regs); | |
1210 | ||
1211 | if (!current || current->pid == 0) | |
1212 | return; | |
1213 | ||
1214 | if (is_user && current->state != TASK_RUNNING) | |
1215 | return; | |
1216 | ||
1217 | if (!is_user) | |
1218 | perf_callchain_kernel(regs, entry); | |
1219 | ||
1220 | if (current->mm) | |
1221 | perf_callchain_user(regs, entry); | |
1222 | } | |
1223 | ||
1224 | struct perf_callchain_entry *perf_callchain(struct pt_regs *regs) | |
1225 | { | |
1226 | struct perf_callchain_entry *entry; | |
1227 | ||
1228 | if (in_nmi()) | |
1229 | entry = &__get_cpu_var(nmi_entry); | |
1230 | else | |
1231 | entry = &__get_cpu_var(irq_entry); | |
1232 | ||
1233 | entry->nr = 0; | |
5872bdb8 PZ |
1234 | entry->hv = 0; |
1235 | entry->kernel = 0; | |
1236 | entry->user = 0; | |
d7d59fb3 PZ |
1237 | |
1238 | perf_do_callchain(regs, entry); | |
1239 | ||
1240 | return entry; | |
1241 | } |