projects / deliverable / linux.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
perf, x86: Restrict the ANY flag
[deliverable/linux.git] / arch / x86 / kernel / cpu / perf_event_intel.c
CommitLineData
f22f54f4
PZ		 1#ifdef CONFIG_CPU_SUP_INTEL
2
3/*
4 * Intel PerfMon v3. Used on Core2 and later.
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{
30 FIXED_EVENT_CONSTRAINT(0xc0, (0x3|(1ULL<<32))), /* INSTRUCTIONS_RETIRED */
31 FIXED_EVENT_CONSTRAINT(0x3c, (0x3|(1ULL<<33))), /* UNHALTED_CORE_CYCLES */
32 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
33 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
34 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
35 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
36 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
37 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
38 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
39 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
40 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
41 EVENT_CONSTRAINT_END
42};
43
44static struct event_constraint intel_nehalem_event_constraints[] =
45{
46 FIXED_EVENT_CONSTRAINT(0xc0, (0xf|(1ULL<<32))), /* INSTRUCTIONS_RETIRED */
47 FIXED_EVENT_CONSTRAINT(0x3c, (0xf|(1ULL<<33))), /* UNHALTED_CORE_CYCLES */
48 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
49 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
50 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
51 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
52 INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */
53 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
54 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
55 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
56 EVENT_CONSTRAINT_END
57};
58
59static struct event_constraint intel_westmere_event_constraints[] =
60{
61 FIXED_EVENT_CONSTRAINT(0xc0, (0xf|(1ULL<<32))), /* INSTRUCTIONS_RETIRED */
62 FIXED_EVENT_CONSTRAINT(0x3c, (0xf|(1ULL<<33))), /* UNHALTED_CORE_CYCLES */
63 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
64 INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */
65 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
66 EVENT_CONSTRAINT_END
67};
68
69static struct event_constraint intel_gen_event_constraints[] =
70{
71 FIXED_EVENT_CONSTRAINT(0xc0, (0x3|(1ULL<<32))), /* INSTRUCTIONS_RETIRED */
72 FIXED_EVENT_CONSTRAINT(0x3c, (0x3|(1ULL<<33))), /* UNHALTED_CORE_CYCLES */
73 EVENT_CONSTRAINT_END
74};
75
76static u64 intel_pmu_event_map(int hw_event)
77{
78 return intel_perfmon_event_map[hw_event];
79}
80
81static __initconst u64 westmere_hw_cache_event_ids
82 [PERF_COUNT_HW_CACHE_MAX]
83 [PERF_COUNT_HW_CACHE_OP_MAX]
84 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
85{
86 [ C(L1D) ] = {
87 [ C(OP_READ) ] = {
88 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
89 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
90 },
91 [ C(OP_WRITE) ] = {
92 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
93 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
94 },
95 [ C(OP_PREFETCH) ] = {
96 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
97 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
98 },
99 },
100 [ C(L1I ) ] = {
101 [ C(OP_READ) ] = {
102 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
103 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
104 },
105 [ C(OP_WRITE) ] = {
106 [ C(RESULT_ACCESS) ] = -1,
107 [ C(RESULT_MISS) ] = -1,
108 },
109 [ C(OP_PREFETCH) ] = {
110 [ C(RESULT_ACCESS) ] = 0x0,
111 [ C(RESULT_MISS) ] = 0x0,
112 },
113 },
114 [ C(LL ) ] = {
115 [ C(OP_READ) ] = {
116 [ C(RESULT_ACCESS) ] = 0x0324, /* L2_RQSTS.LOADS */
117 [ C(RESULT_MISS) ] = 0x0224, /* L2_RQSTS.LD_MISS */
118 },
119 [ C(OP_WRITE) ] = {
120 [ C(RESULT_ACCESS) ] = 0x0c24, /* L2_RQSTS.RFOS */
121 [ C(RESULT_MISS) ] = 0x0824, /* L2_RQSTS.RFO_MISS */
122 },
123 [ C(OP_PREFETCH) ] = {
124 [ C(RESULT_ACCESS) ] = 0x4f2e, /* LLC Reference */
125 [ C(RESULT_MISS) ] = 0x412e, /* LLC Misses */
126 },
127 },
128 [ C(DTLB) ] = {
129 [ C(OP_READ) ] = {
130 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
131 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
132 },
133 [ C(OP_WRITE) ] = {
134 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
135 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
136 },
137 [ C(OP_PREFETCH) ] = {
138 [ C(RESULT_ACCESS) ] = 0x0,
139 [ C(RESULT_MISS) ] = 0x0,
140 },
141 },
142 [ C(ITLB) ] = {
143 [ C(OP_READ) ] = {
144 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
145 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */
146 },
147 [ C(OP_WRITE) ] = {
148 [ C(RESULT_ACCESS) ] = -1,
149 [ C(RESULT_MISS) ] = -1,
150 },
151 [ C(OP_PREFETCH) ] = {
152 [ C(RESULT_ACCESS) ] = -1,
153 [ C(RESULT_MISS) ] = -1,
154 },
155 },
156 [ C(BPU ) ] = {
157 [ C(OP_READ) ] = {
158 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
159 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
160 },
161 [ C(OP_WRITE) ] = {
162 [ C(RESULT_ACCESS) ] = -1,
163 [ C(RESULT_MISS) ] = -1,
164 },
165 [ C(OP_PREFETCH) ] = {
166 [ C(RESULT_ACCESS) ] = -1,
167 [ C(RESULT_MISS) ] = -1,
168 },
169 },
170};
171
172static __initconst u64 nehalem_hw_cache_event_ids
173 [PERF_COUNT_HW_CACHE_MAX]
174 [PERF_COUNT_HW_CACHE_OP_MAX]
175 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
176{
177 [ C(L1D) ] = {
178 [ C(OP_READ) ] = {
179 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
180 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
181 },
182 [ C(OP_WRITE) ] = {
183 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
184 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
185 },
186 [ C(OP_PREFETCH) ] = {
187 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
188 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
189 },
190 },
191 [ C(L1I ) ] = {
192 [ C(OP_READ) ] = {
193 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
194 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
195 },
196 [ C(OP_WRITE) ] = {
197 [ C(RESULT_ACCESS) ] = -1,
198 [ C(RESULT_MISS) ] = -1,
199 },
200 [ C(OP_PREFETCH) ] = {
201 [ C(RESULT_ACCESS) ] = 0x0,
202 [ C(RESULT_MISS) ] = 0x0,
203 },
204 },
205 [ C(LL ) ] = {
206 [ C(OP_READ) ] = {
207 [ C(RESULT_ACCESS) ] = 0x0324, /* L2_RQSTS.LOADS */
208 [ C(RESULT_MISS) ] = 0x0224, /* L2_RQSTS.LD_MISS */
209 },
210 [ C(OP_WRITE) ] = {
211 [ C(RESULT_ACCESS) ] = 0x0c24, /* L2_RQSTS.RFOS */
212 [ C(RESULT_MISS) ] = 0x0824, /* L2_RQSTS.RFO_MISS */
213 },
214 [ C(OP_PREFETCH) ] = {
215 [ C(RESULT_ACCESS) ] = 0x4f2e, /* LLC Reference */
216 [ C(RESULT_MISS) ] = 0x412e, /* LLC Misses */
217 },
218 },
219 [ C(DTLB) ] = {
220 [ C(OP_READ) ] = {
221 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
222 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
223 },
224 [ C(OP_WRITE) ] = {
225 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
226 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
227 },
228 [ C(OP_PREFETCH) ] = {
229 [ C(RESULT_ACCESS) ] = 0x0,
230 [ C(RESULT_MISS) ] = 0x0,
231 },
232 },
233 [ C(ITLB) ] = {
234 [ C(OP_READ) ] = {
235 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
236 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */
237 },
238 [ C(OP_WRITE) ] = {
239 [ C(RESULT_ACCESS) ] = -1,
240 [ C(RESULT_MISS) ] = -1,
241 },
242 [ C(OP_PREFETCH) ] = {
243 [ C(RESULT_ACCESS) ] = -1,
244 [ C(RESULT_MISS) ] = -1,
245 },
246 },
247 [ C(BPU ) ] = {
248 [ C(OP_READ) ] = {
249 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
250 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
251 },
252 [ C(OP_WRITE) ] = {
253 [ C(RESULT_ACCESS) ] = -1,
254 [ C(RESULT_MISS) ] = -1,
255 },
256 [ C(OP_PREFETCH) ] = {
257 [ C(RESULT_ACCESS) ] = -1,
258 [ C(RESULT_MISS) ] = -1,
259 },
260 },
261};
262
263static __initconst u64 core2_hw_cache_event_ids
264 [PERF_COUNT_HW_CACHE_MAX]
265 [PERF_COUNT_HW_CACHE_OP_MAX]
266 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
267{
268 [ C(L1D) ] = {
269 [ C(OP_READ) ] = {
270 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
271 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
272 },
273 [ C(OP_WRITE) ] = {
274 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
275 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
276 },
277 [ C(OP_PREFETCH) ] = {
278 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */
279 [ C(RESULT_MISS) ] = 0,
280 },
281 },
282 [ C(L1I ) ] = {
283 [ C(OP_READ) ] = {
284 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */
285 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */
286 },
287 [ C(OP_WRITE) ] = {
288 [ C(RESULT_ACCESS) ] = -1,
289 [ C(RESULT_MISS) ] = -1,
290 },
291 [ C(OP_PREFETCH) ] = {
292 [ C(RESULT_ACCESS) ] = 0,
293 [ C(RESULT_MISS) ] = 0,
294 },
295 },
296 [ C(LL ) ] = {
297 [ C(OP_READ) ] = {
298 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
299 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
300 },
301 [ C(OP_WRITE) ] = {
302 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
303 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
304 },
305 [ C(OP_PREFETCH) ] = {
306 [ C(RESULT_ACCESS) ] = 0,
307 [ C(RESULT_MISS) ] = 0,
308 },
309 },
310 [ C(DTLB) ] = {
311 [ C(OP_READ) ] = {
312 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
313 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */
314 },
315 [ C(OP_WRITE) ] = {
316 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
317 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */
318 },
319 [ C(OP_PREFETCH) ] = {
320 [ C(RESULT_ACCESS) ] = 0,
321 [ C(RESULT_MISS) ] = 0,
322 },
323 },
324 [ C(ITLB) ] = {
325 [ C(OP_READ) ] = {
326 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
327 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */
328 },
329 [ C(OP_WRITE) ] = {
330 [ C(RESULT_ACCESS) ] = -1,
331 [ C(RESULT_MISS) ] = -1,
332 },
333 [ C(OP_PREFETCH) ] = {
334 [ C(RESULT_ACCESS) ] = -1,
335 [ C(RESULT_MISS) ] = -1,
336 },
337 },
338 [ C(BPU ) ] = {
339 [ C(OP_READ) ] = {
340 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
341 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
342 },
343 [ C(OP_WRITE) ] = {
344 [ C(RESULT_ACCESS) ] = -1,
345 [ C(RESULT_MISS) ] = -1,
346 },
347 [ C(OP_PREFETCH) ] = {
348 [ C(RESULT_ACCESS) ] = -1,
349 [ C(RESULT_MISS) ] = -1,
350 },
351 },
352};
353
354static __initconst u64 atom_hw_cache_event_ids
355 [PERF_COUNT_HW_CACHE_MAX]
356 [PERF_COUNT_HW_CACHE_OP_MAX]
357 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
358{
359 [ C(L1D) ] = {
360 [ C(OP_READ) ] = {
361 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */
362 [ C(RESULT_MISS) ] = 0,
363 },
364 [ C(OP_WRITE) ] = {
365 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */
366 [ C(RESULT_MISS) ] = 0,
367 },
368 [ C(OP_PREFETCH) ] = {
369 [ C(RESULT_ACCESS) ] = 0x0,
370 [ C(RESULT_MISS) ] = 0,
371 },
372 },
373 [ C(L1I ) ] = {
374 [ C(OP_READ) ] = {
375 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
376 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
377 },
378 [ C(OP_WRITE) ] = {
379 [ C(RESULT_ACCESS) ] = -1,
380 [ C(RESULT_MISS) ] = -1,
381 },
382 [ C(OP_PREFETCH) ] = {
383 [ C(RESULT_ACCESS) ] = 0,
384 [ C(RESULT_MISS) ] = 0,
385 },
386 },
387 [ C(LL ) ] = {
388 [ C(OP_READ) ] = {
389 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
390 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
391 },
392 [ C(OP_WRITE) ] = {
393 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
394 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
395 },
396 [ C(OP_PREFETCH) ] = {
397 [ C(RESULT_ACCESS) ] = 0,
398 [ C(RESULT_MISS) ] = 0,
399 },
400 },
401 [ C(DTLB) ] = {
402 [ C(OP_READ) ] = {
403 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */
404 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */
405 },
406 [ C(OP_WRITE) ] = {
407 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */
408 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */
409 },
410 [ C(OP_PREFETCH) ] = {
411 [ C(RESULT_ACCESS) ] = 0,
412 [ C(RESULT_MISS) ] = 0,
413 },
414 },
415 [ C(ITLB) ] = {
416 [ C(OP_READ) ] = {
417 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
418 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */
419 },
420 [ C(OP_WRITE) ] = {
421 [ C(RESULT_ACCESS) ] = -1,
422 [ C(RESULT_MISS) ] = -1,
423 },
424 [ C(OP_PREFETCH) ] = {
425 [ C(RESULT_ACCESS) ] = -1,
426 [ C(RESULT_MISS) ] = -1,
427 },
428 },
429 [ C(BPU ) ] = {
430 [ C(OP_READ) ] = {
431 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
432 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
433 },
434 [ C(OP_WRITE) ] = {
435 [ C(RESULT_ACCESS) ] = -1,
436 [ C(RESULT_MISS) ] = -1,
437 },
438 [ C(OP_PREFETCH) ] = {
439 [ C(RESULT_ACCESS) ] = -1,
440 [ C(RESULT_MISS) ] = -1,
441 },
442 },
443};
444
445static u64 intel_pmu_raw_event(u64 hw_event)
446{
447#define CORE_EVNTSEL_EVENT_MASK 0x000000FFULL
448#define CORE_EVNTSEL_UNIT_MASK 0x0000FF00ULL
449#define CORE_EVNTSEL_EDGE_MASK 0x00040000ULL
450#define CORE_EVNTSEL_INV_MASK 0x00800000ULL
451#define CORE_EVNTSEL_REG_MASK 0xFF000000ULL
452
453#define CORE_EVNTSEL_MASK \
454 (INTEL_ARCH_EVTSEL_MASK | \
455 INTEL_ARCH_UNIT_MASK | \
456 INTEL_ARCH_EDGE_MASK | \
457 INTEL_ARCH_INV_MASK | \
458 INTEL_ARCH_CNT_MASK)
459
460 return hw_event & CORE_EVNTSEL_MASK;
461}
462
463static void intel_pmu_enable_bts(u64 config)
464{
465 unsigned long debugctlmsr;
466
467 debugctlmsr = get_debugctlmsr();
468
469 debugctlmsr |= X86_DEBUGCTL_TR;
470 debugctlmsr |= X86_DEBUGCTL_BTS;
471 debugctlmsr |= X86_DEBUGCTL_BTINT;
472
473 if (!(config & ARCH_PERFMON_EVENTSEL_OS))
474 debugctlmsr |= X86_DEBUGCTL_BTS_OFF_OS;
475
476 if (!(config & ARCH_PERFMON_EVENTSEL_USR))
477 debugctlmsr |= X86_DEBUGCTL_BTS_OFF_USR;
478
479 update_debugctlmsr(debugctlmsr);
480}
481
482static void intel_pmu_disable_bts(void)
483{
484 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
485 unsigned long debugctlmsr;
486
487 if (!cpuc->ds)
488 return;
489
490 debugctlmsr = get_debugctlmsr();
491
492 debugctlmsr &=
493 ~(X86_DEBUGCTL_TR | X86_DEBUGCTL_BTS | X86_DEBUGCTL_BTINT |
494 X86_DEBUGCTL_BTS_OFF_OS | X86_DEBUGCTL_BTS_OFF_USR);
495
496 update_debugctlmsr(debugctlmsr);
497}
498
499static void intel_pmu_disable_all(void)
500{
501 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
502
503 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
504
505 if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask))
506 intel_pmu_disable_bts();
507}
508
509static void intel_pmu_enable_all(void)
510{
511 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
512
513 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl);
514
515 if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
516 struct perf_event *event =
517 cpuc->events[X86_PMC_IDX_FIXED_BTS];
518
519 if (WARN_ON_ONCE(!event))
520 return;
521
522 intel_pmu_enable_bts(event->hw.config);
523 }
524}
525
526static inline u64 intel_pmu_get_status(void)
527{
528 u64 status;
529
530 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
531
532 return status;
533}
534
535static inline void intel_pmu_ack_status(u64 ack)
536{
537 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
538}
539
540static inline void
541intel_pmu_disable_fixed(struct hw_perf_event *hwc, int __idx)
542{
543 int idx = __idx - X86_PMC_IDX_FIXED;
544 u64 ctrl_val, mask;
545
546 mask = 0xfULL << (idx * 4);
547
548 rdmsrl(hwc->config_base, ctrl_val);
549 ctrl_val &= ~mask;
550 (void)checking_wrmsrl(hwc->config_base, ctrl_val);
551}
552
553static void intel_pmu_drain_bts_buffer(void)
554{
555 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
556 struct debug_store *ds = cpuc->ds;
557 struct bts_record {
558 u64 from;
559 u64 to;
560 u64 flags;
561 };
562 struct perf_event *event = cpuc->events[X86_PMC_IDX_FIXED_BTS];
563 struct bts_record *at, *top;
564 struct perf_output_handle handle;
565 struct perf_event_header header;
566 struct perf_sample_data data;
567 struct pt_regs regs;
568
569 if (!event)
570 return;
571
572 if (!ds)
573 return;
574
575 at = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
576 top = (struct bts_record *)(unsigned long)ds->bts_index;
577
578 if (top <= at)
579 return;
580
581 ds->bts_index = ds->bts_buffer_base;
582
583
584 data.period = event->hw.last_period;
585 data.addr = 0;
586 data.raw = NULL;
587 regs.ip = 0;
588
589 /*
590 * Prepare a generic sample, i.e. fill in the invariant fields.
591 * We will overwrite the from and to address before we output
592 * the sample.
593 */
594 perf_prepare_sample(&header, &data, event, &regs);
595
596 if (perf_output_begin(&handle, event,
597 header.size * (top - at), 1, 1))
598 return;
599
600 for (; at < top; at++) {
601 data.ip = at->from;
602 data.addr = at->to;
603
604 perf_output_sample(&handle, &header, &data, event);
605 }
606
607 perf_output_end(&handle);
608
609 /* There's new data available. */
610 event->hw.interrupts++;
611 event->pending_kill = POLL_IN;
612}
613
614static inline void
615intel_pmu_disable_event(struct hw_perf_event *hwc, int idx)
616{
617 if (unlikely(idx == X86_PMC_IDX_FIXED_BTS)) {
618 intel_pmu_disable_bts();
619 intel_pmu_drain_bts_buffer();
620 return;
621 }
622
623 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
624 intel_pmu_disable_fixed(hwc, idx);
625 return;
626 }
627
628 x86_pmu_disable_event(hwc, idx);
629}
630
631static inline void
632intel_pmu_enable_fixed(struct hw_perf_event *hwc, int __idx)
633{
634 int idx = __idx - X86_PMC_IDX_FIXED;
635 u64 ctrl_val, bits, mask;
636 int err;
637
638 /*
639 * Enable IRQ generation (0x8),
640 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
641 * if requested:
642 */
643 bits = 0x8ULL;
644 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
645 bits |= 0x2;
646 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
647 bits |= 0x1;
648
649 /*
650 * ANY bit is supported in v3 and up
651 */
652 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
653 bits |= 0x4;
654
655 bits <<= (idx * 4);
656 mask = 0xfULL << (idx * 4);
657
658 rdmsrl(hwc->config_base, ctrl_val);
659 ctrl_val &= ~mask;
660 ctrl_val |= bits;
661 err = checking_wrmsrl(hwc->config_base, ctrl_val);
662}
663
664static void intel_pmu_enable_event(struct hw_perf_event *hwc, int idx)
665{
666 if (unlikely(idx == X86_PMC_IDX_FIXED_BTS)) {
667 if (!__get_cpu_var(cpu_hw_events).enabled)
668 return;
669
670 intel_pmu_enable_bts(hwc->config);
671 return;
672 }
673
674 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
675 intel_pmu_enable_fixed(hwc, idx);
676 return;
677 }
678
679 __x86_pmu_enable_event(hwc, idx);
680}
681
682/*
683 * Save and restart an expired event. Called by NMI contexts,
684 * so it has to be careful about preempting normal event ops:
685 */
686static int intel_pmu_save_and_restart(struct perf_event *event)
687{
688 struct hw_perf_event *hwc = &event->hw;
689 int idx = hwc->idx;
690 int ret;
691
692 x86_perf_event_update(event, hwc, idx);
693 ret = x86_perf_event_set_period(event, hwc, idx);
694
695 return ret;
696}
697
698static void intel_pmu_reset(void)
699{
700 struct debug_store *ds = __get_cpu_var(cpu_hw_events).ds;
701 unsigned long flags;
702 int idx;
703
704 if (!x86_pmu.num_events)
705 return;
706
707 local_irq_save(flags);
708
709 printk("clearing PMU state on CPU#%d\n", smp_processor_id());
710
711 for (idx = 0; idx < x86_pmu.num_events; idx++) {
712 checking_wrmsrl(x86_pmu.eventsel + idx, 0ull);
713 checking_wrmsrl(x86_pmu.perfctr + idx, 0ull);
714 }
715 for (idx = 0; idx < x86_pmu.num_events_fixed; idx++) {
716 checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
717 }
718 if (ds)
719 ds->bts_index = ds->bts_buffer_base;
720
721 local_irq_restore(flags);
722}
723
724/*
725 * This handler is triggered by the local APIC, so the APIC IRQ handling
726 * rules apply:
727 */
728static int intel_pmu_handle_irq(struct pt_regs *regs)
729{
730 struct perf_sample_data data;
731 struct cpu_hw_events *cpuc;
732 int bit, loops;
733 u64 ack, status;
734
735 data.addr = 0;
736 data.raw = NULL;
737
738 cpuc = &__get_cpu_var(cpu_hw_events);
739
740 perf_disable();
741 intel_pmu_drain_bts_buffer();
742 status = intel_pmu_get_status();
743 if (!status) {
744 perf_enable();
745 return 0;
746 }
747
748 loops = 0;
749again:
750 if (++loops > 100) {
751 WARN_ONCE(1, "perfevents: irq loop stuck!\n");
752 perf_event_print_debug();
753 intel_pmu_reset();
754 perf_enable();
755 return 1;
756 }
757
758 inc_irq_stat(apic_perf_irqs);
759 ack = status;
760 for_each_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
761 struct perf_event *event = cpuc->events[bit];
762
763 clear_bit(bit, (unsigned long *) &status);
764 if (!test_bit(bit, cpuc->active_mask))
765 continue;
766
767 if (!intel_pmu_save_and_restart(event))
768 continue;
769
770 data.period = event->hw.last_period;
771
772 if (perf_event_overflow(event, 1, &data, regs))
773 intel_pmu_disable_event(&event->hw, bit);
774 }
775
776 intel_pmu_ack_status(ack);
777
778 /*
779 * Repeat if there is more work to be done:
780 */
781 status = intel_pmu_get_status();
782 if (status)
783 goto again;
784
785 perf_enable();
786
787 return 1;
788}
789
790static struct event_constraint bts_constraint =
791 EVENT_CONSTRAINT(0, 1ULL << X86_PMC_IDX_FIXED_BTS, 0);
792
793static struct event_constraint *
794intel_special_constraints(struct perf_event *event)
795{
796 unsigned int hw_event;
797
798 hw_event = event->hw.config & INTEL_ARCH_EVENT_MASK;
799
800 if (unlikely((hw_event ==
801 x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS)) &&
802 (event->hw.sample_period == 1))) {
803
804 return &bts_constraint;
805 }
806 return NULL;
807}
808
809static struct event_constraint *
810intel_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
811{
812 struct event_constraint *c;
813
814 c = intel_special_constraints(event);
815 if (c)
816 return c;
817
818 return x86_get_event_constraints(cpuc, event);
819}
820
821static __initconst struct x86_pmu core_pmu = {
822 .name = "core",
823 .handle_irq = x86_pmu_handle_irq,
824 .disable_all = x86_pmu_disable_all,
825 .enable_all = x86_pmu_enable_all,
826 .enable = x86_pmu_enable_event,
827 .disable = x86_pmu_disable_event,
828 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
829 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
830 .event_map = intel_pmu_event_map,
831 .raw_event = intel_pmu_raw_event,
832 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
833 .apic = 1,
834 /*
835 * Intel PMCs cannot be accessed sanely above 32 bit width,
836 * so we install an artificial 1<<31 period regardless of
837 * the generic event period:
838 */
839 .max_period = (1ULL << 31) - 1,
840 .get_event_constraints = intel_get_event_constraints,
841 .event_constraints = intel_core_event_constraints,
842};
843
844static __initconst struct x86_pmu intel_pmu = {
845 .name = "Intel",
846 .handle_irq = intel_pmu_handle_irq,
847 .disable_all = intel_pmu_disable_all,
848 .enable_all = intel_pmu_enable_all,
849 .enable = intel_pmu_enable_event,
850 .disable = intel_pmu_disable_event,
851 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
852 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
853 .event_map = intel_pmu_event_map,
854 .raw_event = intel_pmu_raw_event,
855 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
856 .apic = 1,
857 /*
858 * Intel PMCs cannot be accessed sanely above 32 bit width,
859 * so we install an artificial 1<<31 period regardless of
860 * the generic event period:
861 */
862 .max_period = (1ULL << 31) - 1,
863 .enable_bts = intel_pmu_enable_bts,
864 .disable_bts = intel_pmu_disable_bts,
865 .get_event_constraints = intel_get_event_constraints
866};
867
868static __init int intel_pmu_init(void)
869{
870 union cpuid10_edx edx;
871 union cpuid10_eax eax;
872 unsigned int unused;
873 unsigned int ebx;
874 int version;
875
876 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
877 /* check for P6 processor family */
878 if (boot_cpu_data.x86 == 6) {
879 return p6_pmu_init();
880 } else {
881 return -ENODEV;
882 }
883 }
884
885 /*
886 * Check whether the Architectural PerfMon supports
887 * Branch Misses Retired hw_event or not.
888 */
889 cpuid(10, &eax.full, &ebx, &unused, &edx.full);
890 if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
891 return -ENODEV;
892
893 version = eax.split.version_id;
894 if (version < 2)
895 x86_pmu = core_pmu;
896 else
897 x86_pmu = intel_pmu;
898
899 x86_pmu.version = version;
900 x86_pmu.num_events = eax.split.num_events;
901 x86_pmu.event_bits = eax.split.bit_width;
902 x86_pmu.event_mask = (1ULL << eax.split.bit_width) - 1;
903
904 /*
905 * Quirk: v2 perfmon does not report fixed-purpose events, so
906 * assume at least 3 events:
907 */
908 if (version > 1)
909 x86_pmu.num_events_fixed = max((int)edx.split.num_events_fixed, 3);
910
911 /*
912 * Install the hw-cache-events table:
913 */
914 switch (boot_cpu_data.x86_model) {
915 case 14: /* 65 nm core solo/duo, "Yonah" */
916 pr_cont("Core events, ");
917 break;
918
919 case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
920 case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */
921 case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */
922 case 29: /* six-core 45 nm xeon "Dunnington" */
923 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
924 sizeof(hw_cache_event_ids));
925
926 x86_pmu.event_constraints = intel_core2_event_constraints;
927 pr_cont("Core2 events, ");
928 break;
929
930 case 26: /* 45 nm nehalem, "Bloomfield" */
931 case 30: /* 45 nm nehalem, "Lynnfield" */
932 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
933 sizeof(hw_cache_event_ids));
934
935 x86_pmu.event_constraints = intel_nehalem_event_constraints;
936 pr_cont("Nehalem/Corei7 events, ");
937 break;
938 case 28:
939 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
940 sizeof(hw_cache_event_ids));
941
942 x86_pmu.event_constraints = intel_gen_event_constraints;
943 pr_cont("Atom events, ");
944 break;
945
946 case 37: /* 32 nm nehalem, "Clarkdale" */
947 case 44: /* 32 nm nehalem, "Gulftown" */
948 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
949 sizeof(hw_cache_event_ids));
950
951 x86_pmu.event_constraints = intel_westmere_event_constraints;
952 pr_cont("Westmere events, ");
953 break;
954 default:
955 /*
956 * default constraints for v2 and up
957 */
958 x86_pmu.event_constraints = intel_gen_event_constraints;
959 pr_cont("generic architected perfmon, ");
960 }
961 return 0;
962}
963
964#else /* CONFIG_CPU_SUP_INTEL */
965
966static int intel_pmu_init(void)
967{
968 return 0;
969}
970
971#endif /* CONFIG_CPU_SUP_INTEL */
Linux kernel - Deliverables
This page took 0.091596 seconds and 5 git commands to generate.