perf, x86: Fix accidentally ack'ing a second event on intel perf counter
[deliverable/linux.git] / arch / x86 / kernel / cpu / perf_event_intel.c
CommitLineData
f22f54f4
PZ
1#ifdef CONFIG_CPU_SUP_INTEL
2
3/*
b622d644 4 * Intel PerfMon, used on Core and later.
f22f54f4
PZ
5 */
6static const u64 intel_perfmon_event_map[] =
7{
8 [PERF_COUNT_HW_CPU_CYCLES] = 0x003c,
9 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
10 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e,
11 [PERF_COUNT_HW_CACHE_MISSES] = 0x412e,
12 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4,
13 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5,
14 [PERF_COUNT_HW_BUS_CYCLES] = 0x013c,
15};
16
17static struct event_constraint intel_core_event_constraints[] =
18{
19 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
20 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
21 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
22 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
23 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
24 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */
25 EVENT_CONSTRAINT_END
26};
27
28static struct event_constraint intel_core2_event_constraints[] =
29{
b622d644
PZ
30 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
31 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
32 /*
33 * Core2 has Fixed Counter 2 listed as CPU_CLK_UNHALTED.REF and event
34 * 0x013c as CPU_CLK_UNHALTED.BUS and specifies there is a fixed
35 * ratio between these counters.
36 */
37 /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
f22f54f4
PZ
38 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
39 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
40 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
41 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
42 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
43 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
44 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
45 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
b622d644 46 INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */
f22f54f4
PZ
47 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
48 EVENT_CONSTRAINT_END
49};
50
51static struct event_constraint intel_nehalem_event_constraints[] =
52{
b622d644
PZ
53 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
54 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
55 /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
f22f54f4
PZ
56 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
57 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
58 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
59 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
60 INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */
61 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
62 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
63 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
64 EVENT_CONSTRAINT_END
65};
66
67static struct event_constraint intel_westmere_event_constraints[] =
68{
b622d644
PZ
69 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
70 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
71 /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
f22f54f4
PZ
72 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
73 INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */
74 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
d1100770 75 INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */
f22f54f4
PZ
76 EVENT_CONSTRAINT_END
77};
78
79static struct event_constraint intel_gen_event_constraints[] =
80{
b622d644
PZ
81 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
82 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
83 /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
f22f54f4
PZ
84 EVENT_CONSTRAINT_END
85};
86
87static u64 intel_pmu_event_map(int hw_event)
88{
89 return intel_perfmon_event_map[hw_event];
90}
91
caaa8be3 92static __initconst const u64 westmere_hw_cache_event_ids
f22f54f4
PZ
93 [PERF_COUNT_HW_CACHE_MAX]
94 [PERF_COUNT_HW_CACHE_OP_MAX]
95 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
96{
97 [ C(L1D) ] = {
98 [ C(OP_READ) ] = {
99 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
100 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
101 },
102 [ C(OP_WRITE) ] = {
103 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
104 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
105 },
106 [ C(OP_PREFETCH) ] = {
107 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
108 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
109 },
110 },
111 [ C(L1I ) ] = {
112 [ C(OP_READ) ] = {
113 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
114 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
115 },
116 [ C(OP_WRITE) ] = {
117 [ C(RESULT_ACCESS) ] = -1,
118 [ C(RESULT_MISS) ] = -1,
119 },
120 [ C(OP_PREFETCH) ] = {
121 [ C(RESULT_ACCESS) ] = 0x0,
122 [ C(RESULT_MISS) ] = 0x0,
123 },
124 },
125 [ C(LL ) ] = {
126 [ C(OP_READ) ] = {
127 [ C(RESULT_ACCESS) ] = 0x0324, /* L2_RQSTS.LOADS */
128 [ C(RESULT_MISS) ] = 0x0224, /* L2_RQSTS.LD_MISS */
129 },
130 [ C(OP_WRITE) ] = {
131 [ C(RESULT_ACCESS) ] = 0x0c24, /* L2_RQSTS.RFOS */
132 [ C(RESULT_MISS) ] = 0x0824, /* L2_RQSTS.RFO_MISS */
133 },
134 [ C(OP_PREFETCH) ] = {
135 [ C(RESULT_ACCESS) ] = 0x4f2e, /* LLC Reference */
136 [ C(RESULT_MISS) ] = 0x412e, /* LLC Misses */
137 },
138 },
139 [ C(DTLB) ] = {
140 [ C(OP_READ) ] = {
141 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
142 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
143 },
144 [ C(OP_WRITE) ] = {
145 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
146 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
147 },
148 [ C(OP_PREFETCH) ] = {
149 [ C(RESULT_ACCESS) ] = 0x0,
150 [ C(RESULT_MISS) ] = 0x0,
151 },
152 },
153 [ C(ITLB) ] = {
154 [ C(OP_READ) ] = {
155 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
156 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */
157 },
158 [ C(OP_WRITE) ] = {
159 [ C(RESULT_ACCESS) ] = -1,
160 [ C(RESULT_MISS) ] = -1,
161 },
162 [ C(OP_PREFETCH) ] = {
163 [ C(RESULT_ACCESS) ] = -1,
164 [ C(RESULT_MISS) ] = -1,
165 },
166 },
167 [ C(BPU ) ] = {
168 [ C(OP_READ) ] = {
169 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
170 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
171 },
172 [ C(OP_WRITE) ] = {
173 [ C(RESULT_ACCESS) ] = -1,
174 [ C(RESULT_MISS) ] = -1,
175 },
176 [ C(OP_PREFETCH) ] = {
177 [ C(RESULT_ACCESS) ] = -1,
178 [ C(RESULT_MISS) ] = -1,
179 },
180 },
181};
182
caaa8be3 183static __initconst const u64 nehalem_hw_cache_event_ids
f22f54f4
PZ
184 [PERF_COUNT_HW_CACHE_MAX]
185 [PERF_COUNT_HW_CACHE_OP_MAX]
186 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
187{
188 [ C(L1D) ] = {
189 [ C(OP_READ) ] = {
190 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
191 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
192 },
193 [ C(OP_WRITE) ] = {
194 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
195 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
196 },
197 [ C(OP_PREFETCH) ] = {
198 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
199 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
200 },
201 },
202 [ C(L1I ) ] = {
203 [ C(OP_READ) ] = {
204 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
205 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
206 },
207 [ C(OP_WRITE) ] = {
208 [ C(RESULT_ACCESS) ] = -1,
209 [ C(RESULT_MISS) ] = -1,
210 },
211 [ C(OP_PREFETCH) ] = {
212 [ C(RESULT_ACCESS) ] = 0x0,
213 [ C(RESULT_MISS) ] = 0x0,
214 },
215 },
216 [ C(LL ) ] = {
217 [ C(OP_READ) ] = {
218 [ C(RESULT_ACCESS) ] = 0x0324, /* L2_RQSTS.LOADS */
219 [ C(RESULT_MISS) ] = 0x0224, /* L2_RQSTS.LD_MISS */
220 },
221 [ C(OP_WRITE) ] = {
222 [ C(RESULT_ACCESS) ] = 0x0c24, /* L2_RQSTS.RFOS */
223 [ C(RESULT_MISS) ] = 0x0824, /* L2_RQSTS.RFO_MISS */
224 },
225 [ C(OP_PREFETCH) ] = {
226 [ C(RESULT_ACCESS) ] = 0x4f2e, /* LLC Reference */
227 [ C(RESULT_MISS) ] = 0x412e, /* LLC Misses */
228 },
229 },
230 [ C(DTLB) ] = {
231 [ C(OP_READ) ] = {
232 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
233 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
234 },
235 [ C(OP_WRITE) ] = {
236 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
237 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
238 },
239 [ C(OP_PREFETCH) ] = {
240 [ C(RESULT_ACCESS) ] = 0x0,
241 [ C(RESULT_MISS) ] = 0x0,
242 },
243 },
244 [ C(ITLB) ] = {
245 [ C(OP_READ) ] = {
246 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
247 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */
248 },
249 [ C(OP_WRITE) ] = {
250 [ C(RESULT_ACCESS) ] = -1,
251 [ C(RESULT_MISS) ] = -1,
252 },
253 [ C(OP_PREFETCH) ] = {
254 [ C(RESULT_ACCESS) ] = -1,
255 [ C(RESULT_MISS) ] = -1,
256 },
257 },
258 [ C(BPU ) ] = {
259 [ C(OP_READ) ] = {
260 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
261 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
262 },
263 [ C(OP_WRITE) ] = {
264 [ C(RESULT_ACCESS) ] = -1,
265 [ C(RESULT_MISS) ] = -1,
266 },
267 [ C(OP_PREFETCH) ] = {
268 [ C(RESULT_ACCESS) ] = -1,
269 [ C(RESULT_MISS) ] = -1,
270 },
271 },
272};
273
caaa8be3 274static __initconst const u64 core2_hw_cache_event_ids
f22f54f4
PZ
275 [PERF_COUNT_HW_CACHE_MAX]
276 [PERF_COUNT_HW_CACHE_OP_MAX]
277 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
278{
279 [ C(L1D) ] = {
280 [ C(OP_READ) ] = {
281 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
282 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
283 },
284 [ C(OP_WRITE) ] = {
285 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
286 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
287 },
288 [ C(OP_PREFETCH) ] = {
289 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */
290 [ C(RESULT_MISS) ] = 0,
291 },
292 },
293 [ C(L1I ) ] = {
294 [ C(OP_READ) ] = {
295 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */
296 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */
297 },
298 [ C(OP_WRITE) ] = {
299 [ C(RESULT_ACCESS) ] = -1,
300 [ C(RESULT_MISS) ] = -1,
301 },
302 [ C(OP_PREFETCH) ] = {
303 [ C(RESULT_ACCESS) ] = 0,
304 [ C(RESULT_MISS) ] = 0,
305 },
306 },
307 [ C(LL ) ] = {
308 [ C(OP_READ) ] = {
309 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
310 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
311 },
312 [ C(OP_WRITE) ] = {
313 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
314 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
315 },
316 [ C(OP_PREFETCH) ] = {
317 [ C(RESULT_ACCESS) ] = 0,
318 [ C(RESULT_MISS) ] = 0,
319 },
320 },
321 [ C(DTLB) ] = {
322 [ C(OP_READ) ] = {
323 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
324 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */
325 },
326 [ C(OP_WRITE) ] = {
327 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
328 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */
329 },
330 [ C(OP_PREFETCH) ] = {
331 [ C(RESULT_ACCESS) ] = 0,
332 [ C(RESULT_MISS) ] = 0,
333 },
334 },
335 [ C(ITLB) ] = {
336 [ C(OP_READ) ] = {
337 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
338 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */
339 },
340 [ C(OP_WRITE) ] = {
341 [ C(RESULT_ACCESS) ] = -1,
342 [ C(RESULT_MISS) ] = -1,
343 },
344 [ C(OP_PREFETCH) ] = {
345 [ C(RESULT_ACCESS) ] = -1,
346 [ C(RESULT_MISS) ] = -1,
347 },
348 },
349 [ C(BPU ) ] = {
350 [ C(OP_READ) ] = {
351 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
352 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
353 },
354 [ C(OP_WRITE) ] = {
355 [ C(RESULT_ACCESS) ] = -1,
356 [ C(RESULT_MISS) ] = -1,
357 },
358 [ C(OP_PREFETCH) ] = {
359 [ C(RESULT_ACCESS) ] = -1,
360 [ C(RESULT_MISS) ] = -1,
361 },
362 },
363};
364
caaa8be3 365static __initconst const u64 atom_hw_cache_event_ids
f22f54f4
PZ
366 [PERF_COUNT_HW_CACHE_MAX]
367 [PERF_COUNT_HW_CACHE_OP_MAX]
368 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
369{
370 [ C(L1D) ] = {
371 [ C(OP_READ) ] = {
372 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */
373 [ C(RESULT_MISS) ] = 0,
374 },
375 [ C(OP_WRITE) ] = {
376 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */
377 [ C(RESULT_MISS) ] = 0,
378 },
379 [ C(OP_PREFETCH) ] = {
380 [ C(RESULT_ACCESS) ] = 0x0,
381 [ C(RESULT_MISS) ] = 0,
382 },
383 },
384 [ C(L1I ) ] = {
385 [ C(OP_READ) ] = {
386 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
387 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
388 },
389 [ C(OP_WRITE) ] = {
390 [ C(RESULT_ACCESS) ] = -1,
391 [ C(RESULT_MISS) ] = -1,
392 },
393 [ C(OP_PREFETCH) ] = {
394 [ C(RESULT_ACCESS) ] = 0,
395 [ C(RESULT_MISS) ] = 0,
396 },
397 },
398 [ C(LL ) ] = {
399 [ C(OP_READ) ] = {
400 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
401 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
402 },
403 [ C(OP_WRITE) ] = {
404 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
405 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
406 },
407 [ C(OP_PREFETCH) ] = {
408 [ C(RESULT_ACCESS) ] = 0,
409 [ C(RESULT_MISS) ] = 0,
410 },
411 },
412 [ C(DTLB) ] = {
413 [ C(OP_READ) ] = {
414 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */
415 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */
416 },
417 [ C(OP_WRITE) ] = {
418 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */
419 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */
420 },
421 [ C(OP_PREFETCH) ] = {
422 [ C(RESULT_ACCESS) ] = 0,
423 [ C(RESULT_MISS) ] = 0,
424 },
425 },
426 [ C(ITLB) ] = {
427 [ C(OP_READ) ] = {
428 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
429 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */
430 },
431 [ C(OP_WRITE) ] = {
432 [ C(RESULT_ACCESS) ] = -1,
433 [ C(RESULT_MISS) ] = -1,
434 },
435 [ C(OP_PREFETCH) ] = {
436 [ C(RESULT_ACCESS) ] = -1,
437 [ C(RESULT_MISS) ] = -1,
438 },
439 },
440 [ C(BPU ) ] = {
441 [ C(OP_READ) ] = {
442 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
443 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
444 },
445 [ C(OP_WRITE) ] = {
446 [ C(RESULT_ACCESS) ] = -1,
447 [ C(RESULT_MISS) ] = -1,
448 },
449 [ C(OP_PREFETCH) ] = {
450 [ C(RESULT_ACCESS) ] = -1,
451 [ C(RESULT_MISS) ] = -1,
452 },
453 },
454};
455
f22f54f4
PZ
456static void intel_pmu_disable_all(void)
457{
458 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
459
460 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
461
462 if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask))
463 intel_pmu_disable_bts();
ca037701
PZ
464
465 intel_pmu_pebs_disable_all();
caff2bef 466 intel_pmu_lbr_disable_all();
f22f54f4
PZ
467}
468
11164cd4 469static void intel_pmu_enable_all(int added)
f22f54f4
PZ
470{
471 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
472
d329527e
PZ
473 intel_pmu_pebs_enable_all();
474 intel_pmu_lbr_enable_all();
f22f54f4
PZ
475 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl);
476
477 if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
478 struct perf_event *event =
479 cpuc->events[X86_PMC_IDX_FIXED_BTS];
480
481 if (WARN_ON_ONCE(!event))
482 return;
483
484 intel_pmu_enable_bts(event->hw.config);
485 }
486}
487
11164cd4
PZ
488/*
489 * Workaround for:
490 * Intel Errata AAK100 (model 26)
491 * Intel Errata AAP53 (model 30)
40b91cd1 492 * Intel Errata BD53 (model 44)
11164cd4 493 *
351af072
ZY
494 * The official story:
495 * These chips need to be 'reset' when adding counters by programming the
496 * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either
497 * in sequence on the same PMC or on different PMCs.
498 *
499 * In practise it appears some of these events do in fact count, and
500 * we need to programm all 4 events.
11164cd4 501 */
351af072 502static void intel_pmu_nhm_workaround(void)
11164cd4 503{
351af072
ZY
504 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
505 static const unsigned long nhm_magic[4] = {
506 0x4300B5,
507 0x4300D2,
508 0x4300B1,
509 0x4300B1
510 };
511 struct perf_event *event;
512 int i;
11164cd4 513
351af072
ZY
514 /*
515 * The Errata requires below steps:
516 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL;
517 * 2) Configure 4 PERFEVTSELx with the magic events and clear
518 * the corresponding PMCx;
519 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL;
520 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL;
521 * 5) Clear 4 pairs of ERFEVTSELx and PMCx;
522 */
11164cd4 523
351af072
ZY
524 /*
525 * The real steps we choose are a little different from above.
526 * A) To reduce MSR operations, we don't run step 1) as they
527 * are already cleared before this function is called;
528 * B) Call x86_perf_event_update to save PMCx before configuring
529 * PERFEVTSELx with magic number;
530 * C) With step 5), we do clear only when the PERFEVTSELx is
531 * not used currently.
532 * D) Call x86_perf_event_set_period to restore PMCx;
533 */
11164cd4 534
351af072
ZY
535 /* We always operate 4 pairs of PERF Counters */
536 for (i = 0; i < 4; i++) {
537 event = cpuc->events[i];
538 if (event)
539 x86_perf_event_update(event);
540 }
11164cd4 541
351af072
ZY
542 for (i = 0; i < 4; i++) {
543 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]);
544 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0);
545 }
546
547 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf);
548 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
11164cd4 549
351af072
ZY
550 for (i = 0; i < 4; i++) {
551 event = cpuc->events[i];
552
553 if (event) {
554 x86_perf_event_set_period(event);
31fa58af 555 __x86_pmu_enable_event(&event->hw,
351af072
ZY
556 ARCH_PERFMON_EVENTSEL_ENABLE);
557 } else
558 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0);
11164cd4 559 }
351af072
ZY
560}
561
562static void intel_pmu_nhm_enable_all(int added)
563{
564 if (added)
565 intel_pmu_nhm_workaround();
11164cd4
PZ
566 intel_pmu_enable_all(added);
567}
568
f22f54f4
PZ
569static inline u64 intel_pmu_get_status(void)
570{
571 u64 status;
572
573 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
574
575 return status;
576}
577
578static inline void intel_pmu_ack_status(u64 ack)
579{
580 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
581}
582
ca037701 583static void intel_pmu_disable_fixed(struct hw_perf_event *hwc)
f22f54f4 584{
aff3d91a 585 int idx = hwc->idx - X86_PMC_IDX_FIXED;
f22f54f4
PZ
586 u64 ctrl_val, mask;
587
588 mask = 0xfULL << (idx * 4);
589
590 rdmsrl(hwc->config_base, ctrl_val);
591 ctrl_val &= ~mask;
7645a24c 592 wrmsrl(hwc->config_base, ctrl_val);
f22f54f4
PZ
593}
594
ca037701 595static void intel_pmu_disable_event(struct perf_event *event)
f22f54f4 596{
aff3d91a
PZ
597 struct hw_perf_event *hwc = &event->hw;
598
599 if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) {
f22f54f4
PZ
600 intel_pmu_disable_bts();
601 intel_pmu_drain_bts_buffer();
602 return;
603 }
604
605 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
aff3d91a 606 intel_pmu_disable_fixed(hwc);
f22f54f4
PZ
607 return;
608 }
609
aff3d91a 610 x86_pmu_disable_event(event);
ca037701 611
ab608344 612 if (unlikely(event->attr.precise_ip))
ef21f683 613 intel_pmu_pebs_disable(event);
f22f54f4
PZ
614}
615
ca037701 616static void intel_pmu_enable_fixed(struct hw_perf_event *hwc)
f22f54f4 617{
aff3d91a 618 int idx = hwc->idx - X86_PMC_IDX_FIXED;
f22f54f4 619 u64 ctrl_val, bits, mask;
f22f54f4
PZ
620
621 /*
622 * Enable IRQ generation (0x8),
623 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
624 * if requested:
625 */
626 bits = 0x8ULL;
627 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
628 bits |= 0x2;
629 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
630 bits |= 0x1;
631
632 /*
633 * ANY bit is supported in v3 and up
634 */
635 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
636 bits |= 0x4;
637
638 bits <<= (idx * 4);
639 mask = 0xfULL << (idx * 4);
640
641 rdmsrl(hwc->config_base, ctrl_val);
642 ctrl_val &= ~mask;
643 ctrl_val |= bits;
7645a24c 644 wrmsrl(hwc->config_base, ctrl_val);
f22f54f4
PZ
645}
646
aff3d91a 647static void intel_pmu_enable_event(struct perf_event *event)
f22f54f4 648{
aff3d91a
PZ
649 struct hw_perf_event *hwc = &event->hw;
650
651 if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) {
f22f54f4
PZ
652 if (!__get_cpu_var(cpu_hw_events).enabled)
653 return;
654
655 intel_pmu_enable_bts(hwc->config);
656 return;
657 }
658
659 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
aff3d91a 660 intel_pmu_enable_fixed(hwc);
f22f54f4
PZ
661 return;
662 }
663
ab608344 664 if (unlikely(event->attr.precise_ip))
ef21f683 665 intel_pmu_pebs_enable(event);
ca037701 666
31fa58af 667 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
f22f54f4
PZ
668}
669
670/*
671 * Save and restart an expired event. Called by NMI contexts,
672 * so it has to be careful about preempting normal event ops:
673 */
674static int intel_pmu_save_and_restart(struct perf_event *event)
675{
cc2ad4ba
PZ
676 x86_perf_event_update(event);
677 return x86_perf_event_set_period(event);
f22f54f4
PZ
678}
679
680static void intel_pmu_reset(void)
681{
682 struct debug_store *ds = __get_cpu_var(cpu_hw_events).ds;
683 unsigned long flags;
684 int idx;
685
948b1bb8 686 if (!x86_pmu.num_counters)
f22f54f4
PZ
687 return;
688
689 local_irq_save(flags);
690
691 printk("clearing PMU state on CPU#%d\n", smp_processor_id());
692
948b1bb8 693 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
f22f54f4
PZ
694 checking_wrmsrl(x86_pmu.eventsel + idx, 0ull);
695 checking_wrmsrl(x86_pmu.perfctr + idx, 0ull);
696 }
948b1bb8 697 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++)
f22f54f4 698 checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
948b1bb8 699
f22f54f4
PZ
700 if (ds)
701 ds->bts_index = ds->bts_buffer_base;
702
703 local_irq_restore(flags);
704}
705
706/*
707 * This handler is triggered by the local APIC, so the APIC IRQ handling
708 * rules apply:
709 */
710static int intel_pmu_handle_irq(struct pt_regs *regs)
711{
712 struct perf_sample_data data;
713 struct cpu_hw_events *cpuc;
714 int bit, loops;
2e556b5b 715 u64 status;
f22f54f4 716
dc1d628a 717 perf_sample_data_init(&data, 0);
f22f54f4
PZ
718
719 cpuc = &__get_cpu_var(cpu_hw_events);
720
3fb2b8dd 721 intel_pmu_disable_all();
f22f54f4
PZ
722 intel_pmu_drain_bts_buffer();
723 status = intel_pmu_get_status();
724 if (!status) {
11164cd4 725 intel_pmu_enable_all(0);
f22f54f4
PZ
726 return 0;
727 }
728
729 loops = 0;
730again:
2e556b5b 731 intel_pmu_ack_status(status);
f22f54f4
PZ
732 if (++loops > 100) {
733 WARN_ONCE(1, "perfevents: irq loop stuck!\n");
734 perf_event_print_debug();
735 intel_pmu_reset();
3fb2b8dd 736 goto done;
f22f54f4
PZ
737 }
738
739 inc_irq_stat(apic_perf_irqs);
ca037701 740
caff2bef
PZ
741 intel_pmu_lbr_read();
742
ca037701
PZ
743 /*
744 * PEBS overflow sets bit 62 in the global status register
745 */
746 if (__test_and_clear_bit(62, (unsigned long *)&status))
747 x86_pmu.drain_pebs(regs);
748
984b3f57 749 for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
f22f54f4
PZ
750 struct perf_event *event = cpuc->events[bit];
751
f22f54f4
PZ
752 if (!test_bit(bit, cpuc->active_mask))
753 continue;
754
755 if (!intel_pmu_save_and_restart(event))
756 continue;
757
758 data.period = event->hw.last_period;
759
760 if (perf_event_overflow(event, 1, &data, regs))
71e2d282 761 x86_pmu_stop(event);
f22f54f4
PZ
762 }
763
f22f54f4
PZ
764 /*
765 * Repeat if there is more work to be done:
766 */
767 status = intel_pmu_get_status();
768 if (status)
769 goto again;
770
3fb2b8dd 771done:
11164cd4 772 intel_pmu_enable_all(0);
f22f54f4
PZ
773 return 1;
774}
775
f22f54f4 776static struct event_constraint *
ca037701 777intel_bts_constraints(struct perf_event *event)
f22f54f4 778{
ca037701
PZ
779 struct hw_perf_event *hwc = &event->hw;
780 unsigned int hw_event, bts_event;
f22f54f4 781
ca037701
PZ
782 hw_event = hwc->config & INTEL_ARCH_EVENT_MASK;
783 bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
f22f54f4 784
ca037701 785 if (unlikely(hw_event == bts_event && hwc->sample_period == 1))
f22f54f4 786 return &bts_constraint;
ca037701 787
f22f54f4
PZ
788 return NULL;
789}
790
791static struct event_constraint *
792intel_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
793{
794 struct event_constraint *c;
795
ca037701
PZ
796 c = intel_bts_constraints(event);
797 if (c)
798 return c;
799
800 c = intel_pebs_constraints(event);
f22f54f4
PZ
801 if (c)
802 return c;
803
804 return x86_get_event_constraints(cpuc, event);
805}
806
b4cdc5c2
PZ
807static int intel_pmu_hw_config(struct perf_event *event)
808{
809 int ret = x86_pmu_hw_config(event);
810
811 if (ret)
812 return ret;
813
814 if (event->attr.type != PERF_TYPE_RAW)
815 return 0;
816
817 if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY))
818 return 0;
819
820 if (x86_pmu.version < 3)
821 return -EINVAL;
822
823 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
824 return -EACCES;
825
826 event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY;
827
828 return 0;
829}
830
caaa8be3 831static __initconst const struct x86_pmu core_pmu = {
f22f54f4
PZ
832 .name = "core",
833 .handle_irq = x86_pmu_handle_irq,
834 .disable_all = x86_pmu_disable_all,
835 .enable_all = x86_pmu_enable_all,
836 .enable = x86_pmu_enable_event,
837 .disable = x86_pmu_disable_event,
b4cdc5c2 838 .hw_config = x86_pmu_hw_config,
a072738e 839 .schedule_events = x86_schedule_events,
f22f54f4
PZ
840 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
841 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
842 .event_map = intel_pmu_event_map,
f22f54f4
PZ
843 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
844 .apic = 1,
845 /*
846 * Intel PMCs cannot be accessed sanely above 32 bit width,
847 * so we install an artificial 1<<31 period regardless of
848 * the generic event period:
849 */
850 .max_period = (1ULL << 31) - 1,
851 .get_event_constraints = intel_get_event_constraints,
852 .event_constraints = intel_core_event_constraints,
853};
854
74846d35
PZ
855static void intel_pmu_cpu_starting(int cpu)
856{
857 init_debug_store_on_cpu(cpu);
858 /*
859 * Deal with CPUs that don't clear their LBRs on power-up.
860 */
861 intel_pmu_lbr_reset();
862}
863
864static void intel_pmu_cpu_dying(int cpu)
865{
866 fini_debug_store_on_cpu(cpu);
867}
868
caaa8be3 869static __initconst const struct x86_pmu intel_pmu = {
f22f54f4
PZ
870 .name = "Intel",
871 .handle_irq = intel_pmu_handle_irq,
872 .disable_all = intel_pmu_disable_all,
873 .enable_all = intel_pmu_enable_all,
874 .enable = intel_pmu_enable_event,
875 .disable = intel_pmu_disable_event,
b4cdc5c2 876 .hw_config = intel_pmu_hw_config,
a072738e 877 .schedule_events = x86_schedule_events,
f22f54f4
PZ
878 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
879 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
880 .event_map = intel_pmu_event_map,
f22f54f4
PZ
881 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
882 .apic = 1,
883 /*
884 * Intel PMCs cannot be accessed sanely above 32 bit width,
885 * so we install an artificial 1<<31 period regardless of
886 * the generic event period:
887 */
888 .max_period = (1ULL << 31) - 1,
3f6da390
PZ
889 .get_event_constraints = intel_get_event_constraints,
890
74846d35
PZ
891 .cpu_starting = intel_pmu_cpu_starting,
892 .cpu_dying = intel_pmu_cpu_dying,
f22f54f4
PZ
893};
894
3c44780b
PZ
895static void intel_clovertown_quirks(void)
896{
897 /*
898 * PEBS is unreliable due to:
899 *
900 * AJ67 - PEBS may experience CPL leaks
901 * AJ68 - PEBS PMI may be delayed by one event
902 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
903 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
904 *
905 * AJ67 could be worked around by restricting the OS/USR flags.
906 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
907 *
908 * AJ106 could possibly be worked around by not allowing LBR
909 * usage from PEBS, including the fixup.
910 * AJ68 could possibly be worked around by always programming
911 * a pebs_event_reset[0] value and coping with the lost events.
912 *
913 * But taken together it might just make sense to not enable PEBS on
914 * these chips.
915 */
916 printk(KERN_WARNING "PEBS disabled due to CPU errata.\n");
917 x86_pmu.pebs = 0;
918 x86_pmu.pebs_constraints = NULL;
919}
920
f22f54f4
PZ
921static __init int intel_pmu_init(void)
922{
923 union cpuid10_edx edx;
924 union cpuid10_eax eax;
925 unsigned int unused;
926 unsigned int ebx;
927 int version;
928
929 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
a072738e
CG
930 switch (boot_cpu_data.x86) {
931 case 0x6:
932 return p6_pmu_init();
933 case 0xf:
934 return p4_pmu_init();
935 }
f22f54f4 936 return -ENODEV;
f22f54f4
PZ
937 }
938
939 /*
940 * Check whether the Architectural PerfMon supports
941 * Branch Misses Retired hw_event or not.
942 */
943 cpuid(10, &eax.full, &ebx, &unused, &edx.full);
944 if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
945 return -ENODEV;
946
947 version = eax.split.version_id;
948 if (version < 2)
949 x86_pmu = core_pmu;
950 else
951 x86_pmu = intel_pmu;
952
953 x86_pmu.version = version;
948b1bb8
RR
954 x86_pmu.num_counters = eax.split.num_counters;
955 x86_pmu.cntval_bits = eax.split.bit_width;
956 x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1;
f22f54f4
PZ
957
958 /*
959 * Quirk: v2 perfmon does not report fixed-purpose events, so
960 * assume at least 3 events:
961 */
962 if (version > 1)
948b1bb8 963 x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3);
f22f54f4 964
8db909a7
PZ
965 /*
966 * v2 and above have a perf capabilities MSR
967 */
968 if (version > 1) {
969 u64 capabilities;
970
971 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
972 x86_pmu.intel_cap.capabilities = capabilities;
973 }
974
ca037701
PZ
975 intel_ds_init();
976
f22f54f4
PZ
977 /*
978 * Install the hw-cache-events table:
979 */
980 switch (boot_cpu_data.x86_model) {
981 case 14: /* 65 nm core solo/duo, "Yonah" */
982 pr_cont("Core events, ");
983 break;
984
985 case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
3c44780b 986 x86_pmu.quirks = intel_clovertown_quirks;
f22f54f4
PZ
987 case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */
988 case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */
989 case 29: /* six-core 45 nm xeon "Dunnington" */
990 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
991 sizeof(hw_cache_event_ids));
992
caff2bef
PZ
993 intel_pmu_lbr_init_core();
994
f22f54f4
PZ
995 x86_pmu.event_constraints = intel_core2_event_constraints;
996 pr_cont("Core2 events, ");
997 break;
998
999 case 26: /* 45 nm nehalem, "Bloomfield" */
1000 case 30: /* 45 nm nehalem, "Lynnfield" */
134fbadf 1001 case 46: /* 45 nm nehalem-ex, "Beckton" */
f22f54f4
PZ
1002 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
1003 sizeof(hw_cache_event_ids));
1004
caff2bef
PZ
1005 intel_pmu_lbr_init_nhm();
1006
f22f54f4 1007 x86_pmu.event_constraints = intel_nehalem_event_constraints;
11164cd4
PZ
1008 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
1009 pr_cont("Nehalem events, ");
f22f54f4 1010 break;
caff2bef 1011
b622d644 1012 case 28: /* Atom */
f22f54f4
PZ
1013 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
1014 sizeof(hw_cache_event_ids));
1015
caff2bef
PZ
1016 intel_pmu_lbr_init_atom();
1017
f22f54f4
PZ
1018 x86_pmu.event_constraints = intel_gen_event_constraints;
1019 pr_cont("Atom events, ");
1020 break;
1021
1022 case 37: /* 32 nm nehalem, "Clarkdale" */
1023 case 44: /* 32 nm nehalem, "Gulftown" */
1024 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
1025 sizeof(hw_cache_event_ids));
1026
caff2bef
PZ
1027 intel_pmu_lbr_init_nhm();
1028
f22f54f4 1029 x86_pmu.event_constraints = intel_westmere_event_constraints;
40b91cd1 1030 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
f22f54f4
PZ
1031 pr_cont("Westmere events, ");
1032 break;
b622d644 1033
f22f54f4
PZ
1034 default:
1035 /*
1036 * default constraints for v2 and up
1037 */
1038 x86_pmu.event_constraints = intel_gen_event_constraints;
1039 pr_cont("generic architected perfmon, ");
1040 }
1041 return 0;
1042}
1043
1044#else /* CONFIG_CPU_SUP_INTEL */
1045
1046static int intel_pmu_init(void)
1047{
1048 return 0;
1049}
1050
1051#endif /* CONFIG_CPU_SUP_INTEL */
This page took 0.104518 seconds and 5 git commands to generate.