projects / deliverable / linux.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
perf events: Add generic front-end and back-end stalled cycle event definitions
[deliverable/linux.git] / arch / x86 / kernel / cpu / perf_event_intel.c
CommitLineData
f22f54f4
PZ		 1#ifdef CONFIG_CPU_SUP_INTEL
2
a7e3ed1e
AK		 3#define MAX_EXTRA_REGS 2
4
5/*
6 * Per register state.
7 */
8struct er_account {
9 int ref; /* reference count */
10 unsigned int extra_reg; /* extra MSR number */
11 u64 extra_config; /* extra MSR config */
12};
13
14/*
15 * Per core state
16 * This used to coordinate shared registers for HT threads.
17 */
18struct intel_percore {
19 raw_spinlock_t lock; /* protect structure */
20 struct er_account regs[MAX_EXTRA_REGS];
21 int refcnt; /* number of threads */
22 unsigned core_id;
23};
24
f22f54f4 25/*
b622d644 26 * Intel PerfMon, used on Core and later.
f22f54f4 27 */
ec75a716 28static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly =
f22f54f4
PZ		 29{
30 [PERF_COUNT_HW_CPU_CYCLES] = 0x003c,
31 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
32 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e,
33 [PERF_COUNT_HW_CACHE_MISSES] = 0x412e,
34 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4,
35 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5,
36 [PERF_COUNT_HW_BUS_CYCLES] = 0x013c,
37};
38
5c543e3c 39static struct event_constraint intel_core_event_constraints[] __read_mostly =
f22f54f4
PZ		 40{
41 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
42 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
43 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
44 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
45 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
46 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */
47 EVENT_CONSTRAINT_END
48};
49
5c543e3c 50static struct event_constraint intel_core2_event_constraints[] __read_mostly =
f22f54f4 51{
b622d644
PZ		 52 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
53 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
54 /*
55 * Core2 has Fixed Counter 2 listed as CPU_CLK_UNHALTED.REF and event
56 * 0x013c as CPU_CLK_UNHALTED.BUS and specifies there is a fixed
57 * ratio between these counters.
58 */
59 /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
f22f54f4
PZ		 60 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
61 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
62 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
63 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
64 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
65 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
66 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
67 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
b622d644 68 INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */
f22f54f4
PZ		 69 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
70 EVENT_CONSTRAINT_END
71};
72
5c543e3c 73static struct event_constraint intel_nehalem_event_constraints[] __read_mostly =
f22f54f4 74{
b622d644
PZ		 75 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
76 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
77 /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
f22f54f4
PZ		 78 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
79 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
80 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
81 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
82 INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */
83 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
84 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
85 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
86 EVENT_CONSTRAINT_END
87};
88
5c543e3c 89static struct extra_reg intel_nehalem_extra_regs[] __read_mostly =
a7e3ed1e
AK		 90{
91 INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0xffff),
92 EVENT_EXTRA_END
93};
94
5c543e3c 95static struct event_constraint intel_nehalem_percore_constraints[] __read_mostly =
a7e3ed1e
AK		 96{
97 INTEL_EVENT_CONSTRAINT(0xb7, 0),
98 EVENT_CONSTRAINT_END
99};
100
5c543e3c 101static struct event_constraint intel_westmere_event_constraints[] __read_mostly =
f22f54f4 102{
b622d644
PZ		 103 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
104 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
105 /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
f22f54f4
PZ		 106 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
107 INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */
108 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
d1100770 109 INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */
f22f54f4
PZ		 110 EVENT_CONSTRAINT_END
111};
112
5c543e3c 113static struct event_constraint intel_snb_event_constraints[] __read_mostly =
b06b3d49
LM		 114{
115 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
116 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
117 /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
118 INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */
119 INTEL_EVENT_CONSTRAINT(0xb7, 0x1), /* OFF_CORE_RESPONSE_0 */
120 INTEL_EVENT_CONSTRAINT(0xbb, 0x8), /* OFF_CORE_RESPONSE_1 */
121 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
122 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
123 EVENT_CONSTRAINT_END
124};
125
5c543e3c 126static struct extra_reg intel_westmere_extra_regs[] __read_mostly =
a7e3ed1e
AK		 127{
128 INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0xffff),
129 INTEL_EVENT_EXTRA_REG(0xbb, MSR_OFFCORE_RSP_1, 0xffff),
130 EVENT_EXTRA_END
131};
132
5c543e3c 133static struct event_constraint intel_westmere_percore_constraints[] __read_mostly =
a7e3ed1e
AK		 134{
135 INTEL_EVENT_CONSTRAINT(0xb7, 0),
136 INTEL_EVENT_CONSTRAINT(0xbb, 0),
137 EVENT_CONSTRAINT_END
138};
139
5c543e3c 140static struct event_constraint intel_gen_event_constraints[] __read_mostly =
f22f54f4 141{
b622d644
PZ		 142 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
143 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
144 /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
f22f54f4
PZ		 145 EVENT_CONSTRAINT_END
146};
147
148static u64 intel_pmu_event_map(int hw_event)
149{
150 return intel_perfmon_event_map[hw_event];
151}
152
b06b3d49
LM		 153static __initconst const u64 snb_hw_cache_event_ids
154 [PERF_COUNT_HW_CACHE_MAX]
155 [PERF_COUNT_HW_CACHE_OP_MAX]
156 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
157{
158 [ C(L1D) ] = {
159 [ C(OP_READ) ] = {
160 [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */
161 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */
162 },
163 [ C(OP_WRITE) ] = {
164 [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */
165 [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */
166 },
167 [ C(OP_PREFETCH) ] = {
168 [ C(RESULT_ACCESS) ] = 0x0,
169 [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */
170 },
171 },
172 [ C(L1I ) ] = {
173 [ C(OP_READ) ] = {
174 [ C(RESULT_ACCESS) ] = 0x0,
175 [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */
176 },
177 [ C(OP_WRITE) ] = {
178 [ C(RESULT_ACCESS) ] = -1,
179 [ C(RESULT_MISS) ] = -1,
180 },
181 [ C(OP_PREFETCH) ] = {
182 [ C(RESULT_ACCESS) ] = 0x0,
183 [ C(RESULT_MISS) ] = 0x0,
184 },
185 },
186 [ C(LL ) ] = {
187 /*
188 * TBD: Need Off-core Response Performance Monitoring support
189 */
190 [ C(OP_READ) ] = {
191 /* OFFCORE_RESPONSE_0.ANY_DATA.LOCAL_CACHE */
192 [ C(RESULT_ACCESS) ] = 0x01b7,
193 /* OFFCORE_RESPONSE_1.ANY_DATA.ANY_LLC_MISS */
194 [ C(RESULT_MISS) ] = 0x01bb,
195 },
196 [ C(OP_WRITE) ] = {
197 /* OFFCORE_RESPONSE_0.ANY_RFO.LOCAL_CACHE */
198 [ C(RESULT_ACCESS) ] = 0x01b7,
199 /* OFFCORE_RESPONSE_1.ANY_RFO.ANY_LLC_MISS */
200 [ C(RESULT_MISS) ] = 0x01bb,
201 },
202 [ C(OP_PREFETCH) ] = {
203 /* OFFCORE_RESPONSE_0.PREFETCH.LOCAL_CACHE */
204 [ C(RESULT_ACCESS) ] = 0x01b7,
205 /* OFFCORE_RESPONSE_1.PREFETCH.ANY_LLC_MISS */
206 [ C(RESULT_MISS) ] = 0x01bb,
207 },
208 },
209 [ C(DTLB) ] = {
210 [ C(OP_READ) ] = {
211 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */
212 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */
213 },
214 [ C(OP_WRITE) ] = {
215 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */
216 [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
217 },
218 [ C(OP_PREFETCH) ] = {
219 [ C(RESULT_ACCESS) ] = 0x0,
220 [ C(RESULT_MISS) ] = 0x0,
221 },
222 },
223 [ C(ITLB) ] = {
224 [ C(OP_READ) ] = {
225 [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */
226 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */
227 },
228 [ C(OP_WRITE) ] = {
229 [ C(RESULT_ACCESS) ] = -1,
230 [ C(RESULT_MISS) ] = -1,
231 },
232 [ C(OP_PREFETCH) ] = {
233 [ C(RESULT_ACCESS) ] = -1,
234 [ C(RESULT_MISS) ] = -1,
235 },
236 },
237 [ C(BPU ) ] = {
238 [ C(OP_READ) ] = {
239 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
240 [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
241 },
242 [ C(OP_WRITE) ] = {
243 [ C(RESULT_ACCESS) ] = -1,
244 [ C(RESULT_MISS) ] = -1,
245 },
246 [ C(OP_PREFETCH) ] = {
247 [ C(RESULT_ACCESS) ] = -1,
248 [ C(RESULT_MISS) ] = -1,
249 },
250 },
251};
252
caaa8be3 253static __initconst const u64 westmere_hw_cache_event_ids
f22f54f4
PZ		 254 [PERF_COUNT_HW_CACHE_MAX]
255 [PERF_COUNT_HW_CACHE_OP_MAX]
256 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
257{
258 [ C(L1D) ] = {
259 [ C(OP_READ) ] = {
260 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
261 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
262 },
263 [ C(OP_WRITE) ] = {
264 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
265 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
266 },
267 [ C(OP_PREFETCH) ] = {
268 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
269 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
270 },
271 },
272 [ C(L1I ) ] = {
273 [ C(OP_READ) ] = {
274 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
275 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
276 },
277 [ C(OP_WRITE) ] = {
278 [ C(RESULT_ACCESS) ] = -1,
279 [ C(RESULT_MISS) ] = -1,
280 },
281 [ C(OP_PREFETCH) ] = {
282 [ C(RESULT_ACCESS) ] = 0x0,
283 [ C(RESULT_MISS) ] = 0x0,
284 },
285 },
286 [ C(LL ) ] = {
287 [ C(OP_READ) ] = {
e994d7d2
AK		 288 /* OFFCORE_RESPONSE_0.ANY_DATA.LOCAL_CACHE */
289 [ C(RESULT_ACCESS) ] = 0x01b7,
290 /* OFFCORE_RESPONSE_1.ANY_DATA.ANY_LLC_MISS */
291 [ C(RESULT_MISS) ] = 0x01bb,
f22f54f4 292 },
e994d7d2
AK		 293 /*
294 * Use RFO, not WRITEBACK, because a write miss would typically occur
295 * on RFO.
296 */
f22f54f4 297 [ C(OP_WRITE) ] = {
e994d7d2
AK		 298 /* OFFCORE_RESPONSE_1.ANY_RFO.LOCAL_CACHE */
299 [ C(RESULT_ACCESS) ] = 0x01bb,
300 /* OFFCORE_RESPONSE_0.ANY_RFO.ANY_LLC_MISS */
301 [ C(RESULT_MISS) ] = 0x01b7,
f22f54f4
PZ		 302 },
303 [ C(OP_PREFETCH) ] = {
e994d7d2
AK		 304 /* OFFCORE_RESPONSE_0.PREFETCH.LOCAL_CACHE */
305 [ C(RESULT_ACCESS) ] = 0x01b7,
306 /* OFFCORE_RESPONSE_1.PREFETCH.ANY_LLC_MISS */
307 [ C(RESULT_MISS) ] = 0x01bb,
f22f54f4
PZ		 308 },
309 },
310 [ C(DTLB) ] = {
311 [ C(OP_READ) ] = {
312 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
313 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
314 },
315 [ C(OP_WRITE) ] = {
316 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
317 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
318 },
319 [ C(OP_PREFETCH) ] = {
320 [ C(RESULT_ACCESS) ] = 0x0,
321 [ C(RESULT_MISS) ] = 0x0,
322 },
323 },
324 [ C(ITLB) ] = {
325 [ C(OP_READ) ] = {
326 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
327 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */
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.ALL_BRANCHES */
341 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
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
e994d7d2
AK		 354/*
355 * OFFCORE_RESPONSE MSR bits (subset), See IA32 SDM Vol 3 30.6.1.3
356 */
357
358#define DMND_DATA_RD (1 << 0)
359#define DMND_RFO (1 << 1)
360#define DMND_WB (1 << 3)
361#define PF_DATA_RD (1 << 4)
362#define PF_DATA_RFO (1 << 5)
363#define RESP_UNCORE_HIT (1 << 8)
364#define RESP_MISS (0xf600) /* non uncore hit */
365
366static __initconst const u64 nehalem_hw_cache_extra_regs
367 [PERF_COUNT_HW_CACHE_MAX]
368 [PERF_COUNT_HW_CACHE_OP_MAX]
369 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
370{
371 [ C(LL ) ] = {
372 [ C(OP_READ) ] = {
373 [ C(RESULT_ACCESS) ] = DMND_DATA_RD|RESP_UNCORE_HIT,
374 [ C(RESULT_MISS) ] = DMND_DATA_RD|RESP_MISS,
375 },
376 [ C(OP_WRITE) ] = {
377 [ C(RESULT_ACCESS) ] = DMND_RFO|DMND_WB|RESP_UNCORE_HIT,
378 [ C(RESULT_MISS) ] = DMND_RFO|DMND_WB|RESP_MISS,
379 },
380 [ C(OP_PREFETCH) ] = {
381 [ C(RESULT_ACCESS) ] = PF_DATA_RD|PF_DATA_RFO|RESP_UNCORE_HIT,
382 [ C(RESULT_MISS) ] = PF_DATA_RD|PF_DATA_RFO|RESP_MISS,
383 },
384 }
385};
386
caaa8be3 387static __initconst const u64 nehalem_hw_cache_event_ids
f22f54f4
PZ		 388 [PERF_COUNT_HW_CACHE_MAX]
389 [PERF_COUNT_HW_CACHE_OP_MAX]
390 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
391{
392 [ C(L1D) ] = {
393 [ C(OP_READ) ] = {
f4929bd3
PZ		 394 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
395 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
f22f54f4
PZ		 396 },
397 [ C(OP_WRITE) ] = {
f4929bd3
PZ		 398 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
399 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
f22f54f4
PZ		 400 },
401 [ C(OP_PREFETCH) ] = {
402 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
403 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
404 },
405 },
406 [ C(L1I ) ] = {
407 [ C(OP_READ) ] = {
408 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
409 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
410 },
411 [ C(OP_WRITE) ] = {
412 [ C(RESULT_ACCESS) ] = -1,
413 [ C(RESULT_MISS) ] = -1,
414 },
415 [ C(OP_PREFETCH) ] = {
416 [ C(RESULT_ACCESS) ] = 0x0,
417 [ C(RESULT_MISS) ] = 0x0,
418 },
419 },
420 [ C(LL ) ] = {
421 [ C(OP_READ) ] = {
e994d7d2
AK		 422 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
423 [ C(RESULT_ACCESS) ] = 0x01b7,
424 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
425 [ C(RESULT_MISS) ] = 0x01b7,
f22f54f4 426 },
e994d7d2
AK		 427 /*
428 * Use RFO, not WRITEBACK, because a write miss would typically occur
429 * on RFO.
430 */
f22f54f4 431 [ C(OP_WRITE) ] = {
e994d7d2
AK		 432 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
433 [ C(RESULT_ACCESS) ] = 0x01b7,
434 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
435 [ C(RESULT_MISS) ] = 0x01b7,
f22f54f4
PZ		 436 },
437 [ C(OP_PREFETCH) ] = {
e994d7d2
AK		 438 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
439 [ C(RESULT_ACCESS) ] = 0x01b7,
440 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
441 [ C(RESULT_MISS) ] = 0x01b7,
f22f54f4
PZ		 442 },
443 },
444 [ C(DTLB) ] = {
445 [ C(OP_READ) ] = {
446 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
447 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
448 },
449 [ C(OP_WRITE) ] = {
450 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
451 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
452 },
453 [ C(OP_PREFETCH) ] = {
454 [ C(RESULT_ACCESS) ] = 0x0,
455 [ C(RESULT_MISS) ] = 0x0,
456 },
457 },
458 [ C(ITLB) ] = {
459 [ C(OP_READ) ] = {
460 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
461 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */
462 },
463 [ C(OP_WRITE) ] = {
464 [ C(RESULT_ACCESS) ] = -1,
465 [ C(RESULT_MISS) ] = -1,
466 },
467 [ C(OP_PREFETCH) ] = {
468 [ C(RESULT_ACCESS) ] = -1,
469 [ C(RESULT_MISS) ] = -1,
470 },
471 },
472 [ C(BPU ) ] = {
473 [ C(OP_READ) ] = {
474 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
475 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
476 },
477 [ C(OP_WRITE) ] = {
478 [ C(RESULT_ACCESS) ] = -1,
479 [ C(RESULT_MISS) ] = -1,
480 },
481 [ C(OP_PREFETCH) ] = {
482 [ C(RESULT_ACCESS) ] = -1,
483 [ C(RESULT_MISS) ] = -1,
484 },
485 },
486};
487
caaa8be3 488static __initconst const u64 core2_hw_cache_event_ids
f22f54f4
PZ		 489 [PERF_COUNT_HW_CACHE_MAX]
490 [PERF_COUNT_HW_CACHE_OP_MAX]
491 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
492{
493 [ C(L1D) ] = {
494 [ C(OP_READ) ] = {
495 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
496 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
497 },
498 [ C(OP_WRITE) ] = {
499 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
500 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
501 },
502 [ C(OP_PREFETCH) ] = {
503 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */
504 [ C(RESULT_MISS) ] = 0,
505 },
506 },
507 [ C(L1I ) ] = {
508 [ C(OP_READ) ] = {
509 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */
510 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */
511 },
512 [ C(OP_WRITE) ] = {
513 [ C(RESULT_ACCESS) ] = -1,
514 [ C(RESULT_MISS) ] = -1,
515 },
516 [ C(OP_PREFETCH) ] = {
517 [ C(RESULT_ACCESS) ] = 0,
518 [ C(RESULT_MISS) ] = 0,
519 },
520 },
521 [ C(LL ) ] = {
522 [ C(OP_READ) ] = {
523 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
524 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
525 },
526 [ C(OP_WRITE) ] = {
527 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
528 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
529 },
530 [ C(OP_PREFETCH) ] = {
531 [ C(RESULT_ACCESS) ] = 0,
532 [ C(RESULT_MISS) ] = 0,
533 },
534 },
535 [ C(DTLB) ] = {
536 [ C(OP_READ) ] = {
537 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
538 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */
539 },
540 [ C(OP_WRITE) ] = {
541 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
542 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */
543 },
544 [ C(OP_PREFETCH) ] = {
545 [ C(RESULT_ACCESS) ] = 0,
546 [ C(RESULT_MISS) ] = 0,
547 },
548 },
549 [ C(ITLB) ] = {
550 [ C(OP_READ) ] = {
551 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
552 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */
553 },
554 [ C(OP_WRITE) ] = {
555 [ C(RESULT_ACCESS) ] = -1,
556 [ C(RESULT_MISS) ] = -1,
557 },
558 [ C(OP_PREFETCH) ] = {
559 [ C(RESULT_ACCESS) ] = -1,
560 [ C(RESULT_MISS) ] = -1,
561 },
562 },
563 [ C(BPU ) ] = {
564 [ C(OP_READ) ] = {
565 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
566 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
567 },
568 [ C(OP_WRITE) ] = {
569 [ C(RESULT_ACCESS) ] = -1,
570 [ C(RESULT_MISS) ] = -1,
571 },
572 [ C(OP_PREFETCH) ] = {
573 [ C(RESULT_ACCESS) ] = -1,
574 [ C(RESULT_MISS) ] = -1,
575 },
576 },
577};
578
caaa8be3 579static __initconst const u64 atom_hw_cache_event_ids
f22f54f4
PZ		 580 [PERF_COUNT_HW_CACHE_MAX]
581 [PERF_COUNT_HW_CACHE_OP_MAX]
582 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
583{
584 [ C(L1D) ] = {
585 [ C(OP_READ) ] = {
586 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */
587 [ C(RESULT_MISS) ] = 0,
588 },
589 [ C(OP_WRITE) ] = {
590 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */
591 [ C(RESULT_MISS) ] = 0,
592 },
593 [ C(OP_PREFETCH) ] = {
594 [ C(RESULT_ACCESS) ] = 0x0,
595 [ C(RESULT_MISS) ] = 0,
596 },
597 },
598 [ C(L1I ) ] = {
599 [ C(OP_READ) ] = {
600 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
601 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
602 },
603 [ C(OP_WRITE) ] = {
604 [ C(RESULT_ACCESS) ] = -1,
605 [ C(RESULT_MISS) ] = -1,
606 },
607 [ C(OP_PREFETCH) ] = {
608 [ C(RESULT_ACCESS) ] = 0,
609 [ C(RESULT_MISS) ] = 0,
610 },
611 },
612 [ C(LL ) ] = {
613 [ C(OP_READ) ] = {
614 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
615 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
616 },
617 [ C(OP_WRITE) ] = {
618 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
619 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
620 },
621 [ C(OP_PREFETCH) ] = {
622 [ C(RESULT_ACCESS) ] = 0,
623 [ C(RESULT_MISS) ] = 0,
624 },
625 },
626 [ C(DTLB) ] = {
627 [ C(OP_READ) ] = {
628 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */
629 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */
630 },
631 [ C(OP_WRITE) ] = {
632 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */
633 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */
634 },
635 [ C(OP_PREFETCH) ] = {
636 [ C(RESULT_ACCESS) ] = 0,
637 [ C(RESULT_MISS) ] = 0,
638 },
639 },
640 [ C(ITLB) ] = {
641 [ C(OP_READ) ] = {
642 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
643 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */
644 },
645 [ C(OP_WRITE) ] = {
646 [ C(RESULT_ACCESS) ] = -1,
647 [ C(RESULT_MISS) ] = -1,
648 },
649 [ C(OP_PREFETCH) ] = {
650 [ C(RESULT_ACCESS) ] = -1,
651 [ C(RESULT_MISS) ] = -1,
652 },
653 },
654 [ C(BPU ) ] = {
655 [ C(OP_READ) ] = {
656 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
657 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
658 },
659 [ C(OP_WRITE) ] = {
660 [ C(RESULT_ACCESS) ] = -1,
661 [ C(RESULT_MISS) ] = -1,
662 },
663 [ C(OP_PREFETCH) ] = {
664 [ C(RESULT_ACCESS) ] = -1,
665 [ C(RESULT_MISS) ] = -1,
666 },
667 },
668};
669
f22f54f4
PZ		 670static void intel_pmu_disable_all(void)
671{
672 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
673
674 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
675
676 if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask))
677 intel_pmu_disable_bts();
ca037701
PZ		 678
679 intel_pmu_pebs_disable_all();
caff2bef 680 intel_pmu_lbr_disable_all();
f22f54f4
PZ		 681}
682
11164cd4 683static void intel_pmu_enable_all(int added)
f22f54f4
PZ		 684{
685 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
686
d329527e
PZ		 687 intel_pmu_pebs_enable_all();
688 intel_pmu_lbr_enable_all();
f22f54f4
PZ		 689 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl);
690
691 if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
692 struct perf_event *event =
693 cpuc->events[X86_PMC_IDX_FIXED_BTS];
694
695 if (WARN_ON_ONCE(!event))
696 return;
697
698 intel_pmu_enable_bts(event->hw.config);
699 }
700}
701
11164cd4
PZ		 702/*
703 * Workaround for:
704 * Intel Errata AAK100 (model 26)
705 * Intel Errata AAP53 (model 30)
40b91cd1 706 * Intel Errata BD53 (model 44)
11164cd4 707 *
351af072
ZY		 708 * The official story:
709 * These chips need to be 'reset' when adding counters by programming the
710 * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either
711 * in sequence on the same PMC or on different PMCs.
712 *
713 * In practise it appears some of these events do in fact count, and
714 * we need to programm all 4 events.
11164cd4 715 */
351af072 716static void intel_pmu_nhm_workaround(void)
11164cd4 717{
351af072
ZY		 718 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
719 static const unsigned long nhm_magic[4] = {
720 0x4300B5,
721 0x4300D2,
722 0x4300B1,
723 0x4300B1
724 };
725 struct perf_event *event;
726 int i;
11164cd4 727
351af072
ZY		 728 /*
729 * The Errata requires below steps:
730 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL;
731 * 2) Configure 4 PERFEVTSELx with the magic events and clear
732 * the corresponding PMCx;
733 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL;
734 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL;
735 * 5) Clear 4 pairs of ERFEVTSELx and PMCx;
736 */
11164cd4 737
351af072
ZY		 738 /*
739 * The real steps we choose are a little different from above.
740 * A) To reduce MSR operations, we don't run step 1) as they
741 * are already cleared before this function is called;
742 * B) Call x86_perf_event_update to save PMCx before configuring
743 * PERFEVTSELx with magic number;
744 * C) With step 5), we do clear only when the PERFEVTSELx is
745 * not used currently.
746 * D) Call x86_perf_event_set_period to restore PMCx;
747 */
11164cd4 748
351af072
ZY		 749 /* We always operate 4 pairs of PERF Counters */
750 for (i = 0; i < 4; i++) {
751 event = cpuc->events[i];
752 if (event)
753 x86_perf_event_update(event);
754 }
11164cd4 755
351af072
ZY		 756 for (i = 0; i < 4; i++) {
757 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]);
758 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0);
759 }
760
761 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf);
762 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
11164cd4 763
351af072
ZY		 764 for (i = 0; i < 4; i++) {
765 event = cpuc->events[i];
766
767 if (event) {
768 x86_perf_event_set_period(event);
31fa58af 769 __x86_pmu_enable_event(&event->hw,
351af072
ZY		 770 ARCH_PERFMON_EVENTSEL_ENABLE);
771 } else
772 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0);
11164cd4 773 }
351af072
ZY		 774}
775
776static void intel_pmu_nhm_enable_all(int added)
777{
778 if (added)
779 intel_pmu_nhm_workaround();
11164cd4
PZ		 780 intel_pmu_enable_all(added);
781}
782
f22f54f4
PZ		 783static inline u64 intel_pmu_get_status(void)
784{
785 u64 status;
786
787 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
788
789 return status;
790}
791
792static inline void intel_pmu_ack_status(u64 ack)
793{
794 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
795}
796
ca037701 797static void intel_pmu_disable_fixed(struct hw_perf_event *hwc)
f22f54f4 798{
aff3d91a 799 int idx = hwc->idx - X86_PMC_IDX_FIXED;
f22f54f4
PZ		 800 u64 ctrl_val, mask;
801
802 mask = 0xfULL << (idx * 4);
803
804 rdmsrl(hwc->config_base, ctrl_val);
805 ctrl_val &= ~mask;
7645a24c 806 wrmsrl(hwc->config_base, ctrl_val);
f22f54f4
PZ		 807}
808
ca037701 809static void intel_pmu_disable_event(struct perf_event *event)
f22f54f4 810{
aff3d91a
PZ		 811 struct hw_perf_event *hwc = &event->hw;
812
813 if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) {
f22f54f4
PZ		 814 intel_pmu_disable_bts();
815 intel_pmu_drain_bts_buffer();
816 return;
817 }
818
819 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
aff3d91a 820 intel_pmu_disable_fixed(hwc);
f22f54f4
PZ		 821 return;
822 }
823
aff3d91a 824 x86_pmu_disable_event(event);
ca037701 825
ab608344 826 if (unlikely(event->attr.precise_ip))
ef21f683 827 intel_pmu_pebs_disable(event);
f22f54f4
PZ		 828}
829
ca037701 830static void intel_pmu_enable_fixed(struct hw_perf_event *hwc)
f22f54f4 831{
aff3d91a 832 int idx = hwc->idx - X86_PMC_IDX_FIXED;
f22f54f4 833 u64 ctrl_val, bits, mask;
f22f54f4
PZ		 834
835 /*
836 * Enable IRQ generation (0x8),
837 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
838 * if requested:
839 */
840 bits = 0x8ULL;
841 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
842 bits |= 0x2;
843 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
844 bits |= 0x1;
845
846 /*
847 * ANY bit is supported in v3 and up
848 */
849 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
850 bits |= 0x4;
851
852 bits <<= (idx * 4);
853 mask = 0xfULL << (idx * 4);
854
855 rdmsrl(hwc->config_base, ctrl_val);
856 ctrl_val &= ~mask;
857 ctrl_val |= bits;
7645a24c 858 wrmsrl(hwc->config_base, ctrl_val);
f22f54f4
PZ		 859}
860
aff3d91a 861static void intel_pmu_enable_event(struct perf_event *event)
f22f54f4 862{
aff3d91a
PZ		 863 struct hw_perf_event *hwc = &event->hw;
864
865 if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) {
0a3aee0d 866 if (!__this_cpu_read(cpu_hw_events.enabled))
f22f54f4
PZ		 867 return;
868
869 intel_pmu_enable_bts(hwc->config);
870 return;
871 }
872
873 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
aff3d91a 874 intel_pmu_enable_fixed(hwc);
f22f54f4
PZ		 875 return;
876 }
877
ab608344 878 if (unlikely(event->attr.precise_ip))
ef21f683 879 intel_pmu_pebs_enable(event);
ca037701 880
31fa58af 881 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
f22f54f4
PZ		 882}
883
884/*
885 * Save and restart an expired event. Called by NMI contexts,
886 * so it has to be careful about preempting normal event ops:
887 */
888static int intel_pmu_save_and_restart(struct perf_event *event)
889{
cc2ad4ba
PZ		 890 x86_perf_event_update(event);
891 return x86_perf_event_set_period(event);
f22f54f4
PZ		 892}
893
894static void intel_pmu_reset(void)
895{
0a3aee0d 896 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
f22f54f4
PZ		 897 unsigned long flags;
898 int idx;
899
948b1bb8 900 if (!x86_pmu.num_counters)
f22f54f4
PZ		 901 return;
902
903 local_irq_save(flags);
904
905 printk("clearing PMU state on CPU#%d\n", smp_processor_id());
906
948b1bb8 907 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
41bf4989
RR		 908 checking_wrmsrl(x86_pmu_config_addr(idx), 0ull);
909 checking_wrmsrl(x86_pmu_event_addr(idx), 0ull);
f22f54f4 910 }
948b1bb8 911 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++)
f22f54f4 912 checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
948b1bb8 913
f22f54f4
PZ		 914 if (ds)
915 ds->bts_index = ds->bts_buffer_base;
916
917 local_irq_restore(flags);
918}
919
920/*
921 * This handler is triggered by the local APIC, so the APIC IRQ handling
922 * rules apply:
923 */
924static int intel_pmu_handle_irq(struct pt_regs *regs)
925{
926 struct perf_sample_data data;
927 struct cpu_hw_events *cpuc;
928 int bit, loops;
2e556b5b 929 u64 status;
b0b2072d 930 int handled;
f22f54f4 931
dc1d628a 932 perf_sample_data_init(&data, 0);
f22f54f4
PZ		 933
934 cpuc = &__get_cpu_var(cpu_hw_events);
935
3fb2b8dd 936 intel_pmu_disable_all();
b0b2072d 937 handled = intel_pmu_drain_bts_buffer();
f22f54f4
PZ		 938 status = intel_pmu_get_status();
939 if (!status) {
11164cd4 940 intel_pmu_enable_all(0);
b0b2072d 941 return handled;
f22f54f4
PZ		 942 }
943
944 loops = 0;
945again:
2e556b5b 946 intel_pmu_ack_status(status);
f22f54f4
PZ		 947 if (++loops > 100) {
948 WARN_ONCE(1, "perfevents: irq loop stuck!\n");
949 perf_event_print_debug();
950 intel_pmu_reset();
3fb2b8dd 951 goto done;
f22f54f4
PZ		 952 }
953
954 inc_irq_stat(apic_perf_irqs);
ca037701 955
caff2bef
PZ		 956 intel_pmu_lbr_read();
957
ca037701
PZ		 958 /*
959 * PEBS overflow sets bit 62 in the global status register
960 */
de725dec
PZ		 961 if (__test_and_clear_bit(62, (unsigned long *)&status)) {
962 handled++;
ca037701 963 x86_pmu.drain_pebs(regs);
de725dec 964 }
ca037701 965
984b3f57 966 for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
f22f54f4
PZ		 967 struct perf_event *event = cpuc->events[bit];
968
de725dec
PZ		 969 handled++;
970
f22f54f4
PZ		 971 if (!test_bit(bit, cpuc->active_mask))
972 continue;
973
974 if (!intel_pmu_save_and_restart(event))
975 continue;
976
977 data.period = event->hw.last_period;
978
979 if (perf_event_overflow(event, 1, &data, regs))
a4eaf7f1 980 x86_pmu_stop(event, 0);
f22f54f4
PZ		 981 }
982
f22f54f4
PZ		 983 /*
984 * Repeat if there is more work to be done:
985 */
986 status = intel_pmu_get_status();
987 if (status)
988 goto again;
989
3fb2b8dd 990done:
11164cd4 991 intel_pmu_enable_all(0);
de725dec 992 return handled;
f22f54f4
PZ		 993}
994
f22f54f4 995static struct event_constraint *
ca037701 996intel_bts_constraints(struct perf_event *event)
f22f54f4 997{
ca037701
PZ		 998 struct hw_perf_event *hwc = &event->hw;
999 unsigned int hw_event, bts_event;
f22f54f4 1000
18a073a3
PZ		 1001 if (event->attr.freq)
1002 return NULL;
1003
ca037701
PZ		 1004 hw_event = hwc->config & INTEL_ARCH_EVENT_MASK;
1005 bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
f22f54f4 1006
ca037701 1007 if (unlikely(hw_event == bts_event && hwc->sample_period == 1))
f22f54f4 1008 return &bts_constraint;
ca037701 1009
f22f54f4
PZ		 1010 return NULL;
1011}
1012
a7e3ed1e
AK		 1013static struct event_constraint *
1014intel_percore_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1015{
1016 struct hw_perf_event *hwc = &event->hw;
1017 unsigned int e = hwc->config & ARCH_PERFMON_EVENTSEL_EVENT;
1018 struct event_constraint *c;
1019 struct intel_percore *pc;
1020 struct er_account *era;
1021 int i;
1022 int free_slot;
1023 int found;
1024
1025 if (!x86_pmu.percore_constraints || hwc->extra_alloc)
1026 return NULL;
1027
1028 for (c = x86_pmu.percore_constraints; c->cmask; c++) {
1029 if (e != c->code)
1030 continue;
1031
1032 /*
1033 * Allocate resource per core.
1034 */
1035 pc = cpuc->per_core;
1036 if (!pc)
1037 break;
1038 c = &emptyconstraint;
1039 raw_spin_lock(&pc->lock);
1040 free_slot = -1;
1041 found = 0;
1042 for (i = 0; i < MAX_EXTRA_REGS; i++) {
1043 era = &pc->regs[i];
1044 if (era->ref > 0 && hwc->extra_reg == era->extra_reg) {
1045 /* Allow sharing same config */
1046 if (hwc->extra_config == era->extra_config) {
1047 era->ref++;
1048 cpuc->percore_used = 1;
1049 hwc->extra_alloc = 1;
1050 c = NULL;
1051 }
1052 /* else conflict */
1053 found = 1;
1054 break;
1055 } else if (era->ref == 0 && free_slot == -1)
1056 free_slot = i;
1057 }
1058 if (!found && free_slot != -1) {
1059 era = &pc->regs[free_slot];
1060 era->ref = 1;
1061 era->extra_reg = hwc->extra_reg;
1062 era->extra_config = hwc->extra_config;
1063 cpuc->percore_used = 1;
1064 hwc->extra_alloc = 1;
1065 c = NULL;
1066 }
1067 raw_spin_unlock(&pc->lock);
1068 return c;
1069 }
1070
1071 return NULL;
1072}
1073
f22f54f4
PZ		 1074static struct event_constraint *
1075intel_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1076{
1077 struct event_constraint *c;
1078
ca037701
PZ		 1079 c = intel_bts_constraints(event);
1080 if (c)
1081 return c;
1082
1083 c = intel_pebs_constraints(event);
f22f54f4
PZ		 1084 if (c)
1085 return c;
1086
a7e3ed1e
AK		 1087 c = intel_percore_constraints(cpuc, event);
1088 if (c)
1089 return c;
1090
f22f54f4
PZ		 1091 return x86_get_event_constraints(cpuc, event);
1092}
1093
a7e3ed1e
AK		 1094static void intel_put_event_constraints(struct cpu_hw_events *cpuc,
1095 struct perf_event *event)
1096{
1097 struct extra_reg *er;
1098 struct intel_percore *pc;
1099 struct er_account *era;
1100 struct hw_perf_event *hwc = &event->hw;
1101 int i, allref;
1102
1103 if (!cpuc->percore_used)
1104 return;
1105
1106 for (er = x86_pmu.extra_regs; er->msr; er++) {
1107 if (er->event != (hwc->config & er->config_mask))
1108 continue;
1109
1110 pc = cpuc->per_core;
1111 raw_spin_lock(&pc->lock);
1112 for (i = 0; i < MAX_EXTRA_REGS; i++) {
1113 era = &pc->regs[i];
1114 if (era->ref > 0 &&
1115 era->extra_config == hwc->extra_config &&
1116 era->extra_reg == er->msr) {
1117 era->ref--;
1118 hwc->extra_alloc = 0;
1119 break;
1120 }
1121 }
1122 allref = 0;
1123 for (i = 0; i < MAX_EXTRA_REGS; i++)
1124 allref += pc->regs[i].ref;
1125 if (allref == 0)
1126 cpuc->percore_used = 0;
1127 raw_spin_unlock(&pc->lock);
1128 break;
1129 }
1130}
1131
b4cdc5c2
PZ		 1132static int intel_pmu_hw_config(struct perf_event *event)
1133{
1134 int ret = x86_pmu_hw_config(event);
1135
1136 if (ret)
1137 return ret;
1138
7639dae0
PZ		 1139 if (event->attr.precise_ip &&
1140 (event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
1141 /*
1142 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
1143 * (0x003c) so that we can use it with PEBS.
1144 *
1145 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
1146 * PEBS capable. However we can use INST_RETIRED.ANY_P
1147 * (0x00c0), which is a PEBS capable event, to get the same
1148 * count.
1149 *
1150 * INST_RETIRED.ANY_P counts the number of cycles that retires
1151 * CNTMASK instructions. By setting CNTMASK to a value (16)
1152 * larger than the maximum number of instructions that can be
1153 * retired per cycle (4) and then inverting the condition, we
1154 * count all cycles that retire 16 or less instructions, which
1155 * is every cycle.
1156 *
1157 * Thereby we gain a PEBS capable cycle counter.
1158 */
1159 u64 alt_config = 0x108000c0; /* INST_RETIRED.TOTAL_CYCLES */
1160
1161 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
1162 event->hw.config = alt_config;
1163 }
1164
b4cdc5c2
PZ		 1165 if (event->attr.type != PERF_TYPE_RAW)
1166 return 0;
1167
1168 if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY))
1169 return 0;
1170
1171 if (x86_pmu.version < 3)
1172 return -EINVAL;
1173
1174 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1175 return -EACCES;
1176
1177 event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY;
1178
1179 return 0;
1180}
1181
caaa8be3 1182static __initconst const struct x86_pmu core_pmu = {
f22f54f4
PZ		 1183 .name = "core",
1184 .handle_irq = x86_pmu_handle_irq,
1185 .disable_all = x86_pmu_disable_all,
1186 .enable_all = x86_pmu_enable_all,
1187 .enable = x86_pmu_enable_event,
1188 .disable = x86_pmu_disable_event,
b4cdc5c2 1189 .hw_config = x86_pmu_hw_config,
a072738e 1190 .schedule_events = x86_schedule_events,
f22f54f4
PZ		 1191 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
1192 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
1193 .event_map = intel_pmu_event_map,
f22f54f4
PZ		 1194 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
1195 .apic = 1,
1196 /*
1197 * Intel PMCs cannot be accessed sanely above 32 bit width,
1198 * so we install an artificial 1<<31 period regardless of
1199 * the generic event period:
1200 */
1201 .max_period = (1ULL << 31) - 1,
1202 .get_event_constraints = intel_get_event_constraints,
a7e3ed1e 1203 .put_event_constraints = intel_put_event_constraints,
f22f54f4
PZ		 1204 .event_constraints = intel_core_event_constraints,
1205};
1206
a7e3ed1e
AK		 1207static int intel_pmu_cpu_prepare(int cpu)
1208{
1209 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1210
69092624
LM		 1211 if (!cpu_has_ht_siblings())
1212 return NOTIFY_OK;
1213
a7e3ed1e
AK		 1214 cpuc->per_core = kzalloc_node(sizeof(struct intel_percore),
1215 GFP_KERNEL, cpu_to_node(cpu));
1216 if (!cpuc->per_core)
1217 return NOTIFY_BAD;
1218
1219 raw_spin_lock_init(&cpuc->per_core->lock);
1220 cpuc->per_core->core_id = -1;
1221 return NOTIFY_OK;
1222}
1223
74846d35
PZ		 1224static void intel_pmu_cpu_starting(int cpu)
1225{
a7e3ed1e
AK		 1226 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1227 int core_id = topology_core_id(cpu);
1228 int i;
1229
69092624
LM		 1230 init_debug_store_on_cpu(cpu);
1231 /*
1232 * Deal with CPUs that don't clear their LBRs on power-up.
1233 */
1234 intel_pmu_lbr_reset();
1235
1236 if (!cpu_has_ht_siblings())
1237 return;
1238
a7e3ed1e
AK		 1239 for_each_cpu(i, topology_thread_cpumask(cpu)) {
1240 struct intel_percore *pc = per_cpu(cpu_hw_events, i).per_core;
1241
1242 if (pc && pc->core_id == core_id) {
1243 kfree(cpuc->per_core);
1244 cpuc->per_core = pc;
1245 break;
1246 }
1247 }
1248
1249 cpuc->per_core->core_id = core_id;
1250 cpuc->per_core->refcnt++;
74846d35
PZ		 1251}
1252
1253static void intel_pmu_cpu_dying(int cpu)
1254{
a7e3ed1e
AK		 1255 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1256 struct intel_percore *pc = cpuc->per_core;
1257
1258 if (pc) {
1259 if (pc->core_id == -1 || --pc->refcnt == 0)
1260 kfree(pc);
1261 cpuc->per_core = NULL;
1262 }
1263
74846d35
PZ		 1264 fini_debug_store_on_cpu(cpu);
1265}
1266
caaa8be3 1267static __initconst const struct x86_pmu intel_pmu = {
f22f54f4
PZ		 1268 .name = "Intel",
1269 .handle_irq = intel_pmu_handle_irq,
1270 .disable_all = intel_pmu_disable_all,
1271 .enable_all = intel_pmu_enable_all,
1272 .enable = intel_pmu_enable_event,
1273 .disable = intel_pmu_disable_event,
b4cdc5c2 1274 .hw_config = intel_pmu_hw_config,
a072738e 1275 .schedule_events = x86_schedule_events,
f22f54f4
PZ		 1276 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
1277 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
1278 .event_map = intel_pmu_event_map,
f22f54f4
PZ		 1279 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
1280 .apic = 1,
1281 /*
1282 * Intel PMCs cannot be accessed sanely above 32 bit width,
1283 * so we install an artificial 1<<31 period regardless of
1284 * the generic event period:
1285 */
1286 .max_period = (1ULL << 31) - 1,
3f6da390 1287 .get_event_constraints = intel_get_event_constraints,
a7e3ed1e 1288 .put_event_constraints = intel_put_event_constraints,
3f6da390 1289
a7e3ed1e 1290 .cpu_prepare = intel_pmu_cpu_prepare,
74846d35
PZ		 1291 .cpu_starting = intel_pmu_cpu_starting,
1292 .cpu_dying = intel_pmu_cpu_dying,
f22f54f4
PZ		 1293};
1294
3c44780b
PZ		 1295static void intel_clovertown_quirks(void)
1296{
1297 /*
1298 * PEBS is unreliable due to:
1299 *
1300 * AJ67 - PEBS may experience CPL leaks
1301 * AJ68 - PEBS PMI may be delayed by one event
1302 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
1303 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
1304 *
1305 * AJ67 could be worked around by restricting the OS/USR flags.
1306 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
1307 *
1308 * AJ106 could possibly be worked around by not allowing LBR
1309 * usage from PEBS, including the fixup.
1310 * AJ68 could possibly be worked around by always programming
ec75a716 1311 * a pebs_event_reset[0] value and coping with the lost events.
3c44780b
PZ		 1312 *
1313 * But taken together it might just make sense to not enable PEBS on
1314 * these chips.
1315 */
1316 printk(KERN_WARNING "PEBS disabled due to CPU errata.\n");
1317 x86_pmu.pebs = 0;
1318 x86_pmu.pebs_constraints = NULL;
1319}
1320
f22f54f4
PZ		 1321static __init int intel_pmu_init(void)
1322{
1323 union cpuid10_edx edx;
1324 union cpuid10_eax eax;
1325 unsigned int unused;
1326 unsigned int ebx;
1327 int version;
1328
1329 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
a072738e
CG		 1330 switch (boot_cpu_data.x86) {
1331 case 0x6:
1332 return p6_pmu_init();
1333 case 0xf:
1334 return p4_pmu_init();
1335 }
f22f54f4 1336 return -ENODEV;
f22f54f4
PZ		 1337 }
1338
1339 /*
1340 * Check whether the Architectural PerfMon supports
1341 * Branch Misses Retired hw_event or not.
1342 */
1343 cpuid(10, &eax.full, &ebx, &unused, &edx.full);
1344 if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
1345 return -ENODEV;
1346
1347 version = eax.split.version_id;
1348 if (version < 2)
1349 x86_pmu = core_pmu;
1350 else
1351 x86_pmu = intel_pmu;
1352
1353 x86_pmu.version = version;
948b1bb8
RR		 1354 x86_pmu.num_counters = eax.split.num_counters;
1355 x86_pmu.cntval_bits = eax.split.bit_width;
1356 x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1;
f22f54f4
PZ		 1357
1358 /*
1359 * Quirk: v2 perfmon does not report fixed-purpose events, so
1360 * assume at least 3 events:
1361 */
1362 if (version > 1)
948b1bb8 1363 x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3);
f22f54f4 1364
8db909a7
PZ		 1365 /*
1366 * v2 and above have a perf capabilities MSR
1367 */
1368 if (version > 1) {
1369 u64 capabilities;
1370
1371 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
1372 x86_pmu.intel_cap.capabilities = capabilities;
1373 }
1374
ca037701
PZ		 1375 intel_ds_init();
1376
f22f54f4
PZ		 1377 /*
1378 * Install the hw-cache-events table:
1379 */
1380 switch (boot_cpu_data.x86_model) {
1381 case 14: /* 65 nm core solo/duo, "Yonah" */
1382 pr_cont("Core events, ");
1383 break;
1384
1385 case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
3c44780b 1386 x86_pmu.quirks = intel_clovertown_quirks;
f22f54f4
PZ		 1387 case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */
1388 case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */
1389 case 29: /* six-core 45 nm xeon "Dunnington" */
1390 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
1391 sizeof(hw_cache_event_ids));
1392
caff2bef
PZ		 1393 intel_pmu_lbr_init_core();
1394
f22f54f4 1395 x86_pmu.event_constraints = intel_core2_event_constraints;
17e31629 1396 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints;
f22f54f4
PZ		 1397 pr_cont("Core2 events, ");
1398 break;
1399
1400 case 26: /* 45 nm nehalem, "Bloomfield" */
1401 case 30: /* 45 nm nehalem, "Lynnfield" */
134fbadf 1402 case 46: /* 45 nm nehalem-ex, "Beckton" */
f22f54f4
PZ		 1403 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
1404 sizeof(hw_cache_event_ids));
e994d7d2
AK		 1405 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
1406 sizeof(hw_cache_extra_regs));
f22f54f4 1407
caff2bef
PZ		 1408 intel_pmu_lbr_init_nhm();
1409
f22f54f4 1410 x86_pmu.event_constraints = intel_nehalem_event_constraints;
17e31629 1411 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints;
a7e3ed1e 1412 x86_pmu.percore_constraints = intel_nehalem_percore_constraints;
11164cd4 1413 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
a7e3ed1e 1414 x86_pmu.extra_regs = intel_nehalem_extra_regs;
ec75a716 1415
8a850cad 1416 /* Install the stalled-cycles event: UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
8f622422 1417 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x1803fb1;
94403f88 1418
ec75a716
IM		 1419 if (ebx & 0x40) {
1420 /*
1421 * Erratum AAJ80 detected, we work it around by using
1422 * the BR_MISP_EXEC.ANY event. This will over-count
1423 * branch-misses, but it's still much better than the
1424 * architectural event which is often completely bogus:
1425 */
1426 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89;
1427
1428 pr_cont("erratum AAJ80 worked around, ");
1429 }
11164cd4 1430 pr_cont("Nehalem events, ");
f22f54f4 1431 break;
caff2bef 1432
b622d644 1433 case 28: /* Atom */
f22f54f4
PZ		 1434 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
1435 sizeof(hw_cache_event_ids));
1436
caff2bef
PZ		 1437 intel_pmu_lbr_init_atom();
1438
f22f54f4 1439 x86_pmu.event_constraints = intel_gen_event_constraints;
17e31629 1440 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints;
f22f54f4
PZ		 1441 pr_cont("Atom events, ");
1442 break;
1443
1444 case 37: /* 32 nm nehalem, "Clarkdale" */
1445 case 44: /* 32 nm nehalem, "Gulftown" */
b2508e82 1446 case 47: /* 32 nm Xeon E7 */
f22f54f4
PZ		 1447 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
1448 sizeof(hw_cache_event_ids));
e994d7d2
AK		 1449 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
1450 sizeof(hw_cache_extra_regs));
f22f54f4 1451
caff2bef
PZ		 1452 intel_pmu_lbr_init_nhm();
1453
f22f54f4 1454 x86_pmu.event_constraints = intel_westmere_event_constraints;
a7e3ed1e 1455 x86_pmu.percore_constraints = intel_westmere_percore_constraints;
40b91cd1 1456 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
17e31629 1457 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints;
a7e3ed1e 1458 x86_pmu.extra_regs = intel_westmere_extra_regs;
f22f54f4
PZ		 1459 pr_cont("Westmere events, ");
1460 break;
b622d644 1461
b06b3d49
LM		 1462 case 42: /* SandyBridge */
1463 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
1464 sizeof(hw_cache_event_ids));
1465
1466 intel_pmu_lbr_init_nhm();
1467
1468 x86_pmu.event_constraints = intel_snb_event_constraints;
1469 x86_pmu.pebs_constraints = intel_snb_pebs_events;
1470 pr_cont("SandyBridge events, ");
1471 break;
1472
f22f54f4
PZ		 1473 default:
1474 /*
1475 * default constraints for v2 and up
1476 */
1477 x86_pmu.event_constraints = intel_gen_event_constraints;
1478 pr_cont("generic architected perfmon, ");
1479 }
1480 return 0;
1481}
1482
1483#else /* CONFIG_CPU_SUP_INTEL */
1484
1485static int intel_pmu_init(void)
1486{
1487 return 0;
1488}
1489
1490#endif /* CONFIG_CPU_SUP_INTEL */
Linux kernel - Deliverables
This page took 0.154705 seconds and 5 git commands to generate.