?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Input: wacom - revert "Cintiq 21UX2 does not have menu strips"
[linux-btrfs-devel.git] / arch / x86 / kernel / cpu / perf_event_intel.c
blob41178c826c48507cac37d32f234b0f66a4c99708
1 #ifdef CONFIG_CPU_SUP_INTEL
2
3 #define MAX_EXTRA_REGS 2
4
5 /*
6 * Per register state.
7 */
8 struct 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 */
18 struct 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
25 /*
26 * Intel PerfMon, used on Core and later.
27 */
28 static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly =
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
39 static struct event_constraint intel_core_event_constraints[] __read_mostly =
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
50 static struct event_constraint intel_core2_event_constraints[] __read_mostly =
51 {
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 */
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 */
68 INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */
69 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
70 EVENT_CONSTRAINT_END
71 };
72
73 static struct event_constraint intel_nehalem_event_constraints[] __read_mostly =
74 {
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 */
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
89 static struct extra_reg intel_nehalem_extra_regs[] __read_mostly =
90 {
91 INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0xffff),
92 EVENT_EXTRA_END
93 };
94
95 static struct event_constraint intel_nehalem_percore_constraints[] __read_mostly =
96 {
97 INTEL_EVENT_CONSTRAINT(0xb7, 0),
98 EVENT_CONSTRAINT_END
99 };
100
101 static struct event_constraint intel_westmere_event_constraints[] __read_mostly =
102 {
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 */
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 */
109 INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */
110 EVENT_CONSTRAINT_END
111 };
112
113 static struct event_constraint intel_snb_event_constraints[] __read_mostly =
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
126 static struct extra_reg intel_westmere_extra_regs[] __read_mostly =
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
133 static struct event_constraint intel_westmere_percore_constraints[] __read_mostly =
134 {
135 INTEL_EVENT_CONSTRAINT(0xb7, 0),
136 INTEL_EVENT_CONSTRAINT(0xbb, 0),
137 EVENT_CONSTRAINT_END
138 };
139
140 static struct event_constraint intel_gen_event_constraints[] __read_mostly =
141 {
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 */
145 EVENT_CONSTRAINT_END
146 };
147
148 static u64 intel_pmu_event_map(int hw_event)
149 {
150 return intel_perfmon_event_map[hw_event];
151 }
152
153 static __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 [ C(OP_READ) ] = {
188 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
189 [ C(RESULT_ACCESS) ] = 0x01b7,
190 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
191 [ C(RESULT_MISS) ] = 0x01b7,
192 },
193 [ C(OP_WRITE) ] = {
194 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
195 [ C(RESULT_ACCESS) ] = 0x01b7,
196 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
197 [ C(RESULT_MISS) ] = 0x01b7,
198 },
199 [ C(OP_PREFETCH) ] = {
200 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
201 [ C(RESULT_ACCESS) ] = 0x01b7,
202 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
203 [ C(RESULT_MISS) ] = 0x01b7,
204 },
205 },
206 [ C(DTLB) ] = {
207 [ C(OP_READ) ] = {
208 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */
209 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */
210 },
211 [ C(OP_WRITE) ] = {
212 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */
213 [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
214 },
215 [ C(OP_PREFETCH) ] = {
216 [ C(RESULT_ACCESS) ] = 0x0,
217 [ C(RESULT_MISS) ] = 0x0,
218 },
219 },
220 [ C(ITLB) ] = {
221 [ C(OP_READ) ] = {
222 [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */
223 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */
224 },
225 [ C(OP_WRITE) ] = {
226 [ C(RESULT_ACCESS) ] = -1,
227 [ C(RESULT_MISS) ] = -1,
228 },
229 [ C(OP_PREFETCH) ] = {
230 [ C(RESULT_ACCESS) ] = -1,
231 [ C(RESULT_MISS) ] = -1,
232 },
233 },
234 [ C(BPU ) ] = {
235 [ C(OP_READ) ] = {
236 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
237 [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
238 },
239 [ C(OP_WRITE) ] = {
240 [ C(RESULT_ACCESS) ] = -1,
241 [ C(RESULT_MISS) ] = -1,
242 },
243 [ C(OP_PREFETCH) ] = {
244 [ C(RESULT_ACCESS) ] = -1,
245 [ C(RESULT_MISS) ] = -1,
246 },
247 },
248 };
249
250 static __initconst const u64 westmere_hw_cache_event_ids
251 [PERF_COUNT_HW_CACHE_MAX]
252 [PERF_COUNT_HW_CACHE_OP_MAX]
253 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
254 {
255 [ C(L1D) ] = {
256 [ C(OP_READ) ] = {
257 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
258 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
259 },
260 [ C(OP_WRITE) ] = {
261 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
262 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
263 },
264 [ C(OP_PREFETCH) ] = {
265 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
266 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
267 },
268 },
269 [ C(L1I ) ] = {
270 [ C(OP_READ) ] = {
271 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
272 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
273 },
274 [ C(OP_WRITE) ] = {
275 [ C(RESULT_ACCESS) ] = -1,
276 [ C(RESULT_MISS) ] = -1,
277 },
278 [ C(OP_PREFETCH) ] = {
279 [ C(RESULT_ACCESS) ] = 0x0,
280 [ C(RESULT_MISS) ] = 0x0,
281 },
282 },
283 [ C(LL ) ] = {
284 [ C(OP_READ) ] = {
285 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
286 [ C(RESULT_ACCESS) ] = 0x01b7,
287 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
288 [ C(RESULT_MISS) ] = 0x01b7,
289 },
290 /*
291 * Use RFO, not WRITEBACK, because a write miss would typically occur
292 * on RFO.
293 */
294 [ C(OP_WRITE) ] = {
295 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
296 [ C(RESULT_ACCESS) ] = 0x01b7,
297 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
298 [ C(RESULT_MISS) ] = 0x01b7,
299 },
300 [ C(OP_PREFETCH) ] = {
301 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
302 [ C(RESULT_ACCESS) ] = 0x01b7,
303 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
304 [ C(RESULT_MISS) ] = 0x01b7,
305 },
306 },
307 [ C(DTLB) ] = {
308 [ C(OP_READ) ] = {
309 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
310 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
311 },
312 [ C(OP_WRITE) ] = {
313 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
314 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
315 },
316 [ C(OP_PREFETCH) ] = {
317 [ C(RESULT_ACCESS) ] = 0x0,
318 [ C(RESULT_MISS) ] = 0x0,
319 },
320 },
321 [ C(ITLB) ] = {
322 [ C(OP_READ) ] = {
323 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
324 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */
325 },
326 [ C(OP_WRITE) ] = {
327 [ C(RESULT_ACCESS) ] = -1,
328 [ C(RESULT_MISS) ] = -1,
329 },
330 [ C(OP_PREFETCH) ] = {
331 [ C(RESULT_ACCESS) ] = -1,
332 [ C(RESULT_MISS) ] = -1,
333 },
334 },
335 [ C(BPU ) ] = {
336 [ C(OP_READ) ] = {
337 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
338 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
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 };
350
351 /*
352 * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits;
353 * See IA32 SDM Vol 3B 30.6.1.3
354 */
355
356 #define NHM_DMND_DATA_RD (1 << 0)
357 #define NHM_DMND_RFO (1 << 1)
358 #define NHM_DMND_IFETCH (1 << 2)
359 #define NHM_DMND_WB (1 << 3)
360 #define NHM_PF_DATA_RD (1 << 4)
361 #define NHM_PF_DATA_RFO (1 << 5)
362 #define NHM_PF_IFETCH (1 << 6)
363 #define NHM_OFFCORE_OTHER (1 << 7)
364 #define NHM_UNCORE_HIT (1 << 8)
365 #define NHM_OTHER_CORE_HIT_SNP (1 << 9)
366 #define NHM_OTHER_CORE_HITM (1 << 10)
367 /* reserved */
368 #define NHM_REMOTE_CACHE_FWD (1 << 12)
369 #define NHM_REMOTE_DRAM (1 << 13)
370 #define NHM_LOCAL_DRAM (1 << 14)
371 #define NHM_NON_DRAM (1 << 15)
372
373 #define NHM_ALL_DRAM (NHM_REMOTE_DRAM|NHM_LOCAL_DRAM)
374
375 #define NHM_DMND_READ (NHM_DMND_DATA_RD)
376 #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB)
377 #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO)
378
379 #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM)
380 #define NHM_L3_MISS (NHM_NON_DRAM|NHM_ALL_DRAM|NHM_REMOTE_CACHE_FWD)
381 #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS)
382
383 static __initconst const u64 nehalem_hw_cache_extra_regs
384 [PERF_COUNT_HW_CACHE_MAX]
385 [PERF_COUNT_HW_CACHE_OP_MAX]
386 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
387 {
388 [ C(LL ) ] = {
389 [ C(OP_READ) ] = {
390 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS,
391 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS,
392 },
393 [ C(OP_WRITE) ] = {
394 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS,
395 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS,
396 },
397 [ C(OP_PREFETCH) ] = {
398 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS,
399 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS,
400 },
401 }
402 };
403
404 static __initconst const u64 nehalem_hw_cache_event_ids
405 [PERF_COUNT_HW_CACHE_MAX]
406 [PERF_COUNT_HW_CACHE_OP_MAX]
407 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
408 {
409 [ C(L1D) ] = {
410 [ C(OP_READ) ] = {
411 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
412 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
413 },
414 [ C(OP_WRITE) ] = {
415 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
416 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
417 },
418 [ C(OP_PREFETCH) ] = {
419 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
420 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
421 },
422 },
423 [ C(L1I ) ] = {
424 [ C(OP_READ) ] = {
425 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
426 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
427 },
428 [ C(OP_WRITE) ] = {
429 [ C(RESULT_ACCESS) ] = -1,
430 [ C(RESULT_MISS) ] = -1,
431 },
432 [ C(OP_PREFETCH) ] = {
433 [ C(RESULT_ACCESS) ] = 0x0,
434 [ C(RESULT_MISS) ] = 0x0,
435 },
436 },
437 [ C(LL ) ] = {
438 [ C(OP_READ) ] = {
439 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
440 [ C(RESULT_ACCESS) ] = 0x01b7,
441 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
442 [ C(RESULT_MISS) ] = 0x01b7,
443 },
444 /*
445 * Use RFO, not WRITEBACK, because a write miss would typically occur
446 * on RFO.
447 */
448 [ C(OP_WRITE) ] = {
449 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
450 [ C(RESULT_ACCESS) ] = 0x01b7,
451 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
452 [ C(RESULT_MISS) ] = 0x01b7,
453 },
454 [ C(OP_PREFETCH) ] = {
455 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
456 [ C(RESULT_ACCESS) ] = 0x01b7,
457 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
458 [ C(RESULT_MISS) ] = 0x01b7,
459 },
460 },
461 [ C(DTLB) ] = {
462 [ C(OP_READ) ] = {
463 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
464 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
465 },
466 [ C(OP_WRITE) ] = {
467 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
468 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
469 },
470 [ C(OP_PREFETCH) ] = {
471 [ C(RESULT_ACCESS) ] = 0x0,
472 [ C(RESULT_MISS) ] = 0x0,
473 },
474 },
475 [ C(ITLB) ] = {
476 [ C(OP_READ) ] = {
477 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
478 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */
479 },
480 [ C(OP_WRITE) ] = {
481 [ C(RESULT_ACCESS) ] = -1,
482 [ C(RESULT_MISS) ] = -1,
483 },
484 [ C(OP_PREFETCH) ] = {
485 [ C(RESULT_ACCESS) ] = -1,
486 [ C(RESULT_MISS) ] = -1,
487 },
488 },
489 [ C(BPU ) ] = {
490 [ C(OP_READ) ] = {
491 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
492 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
493 },
494 [ C(OP_WRITE) ] = {
495 [ C(RESULT_ACCESS) ] = -1,
496 [ C(RESULT_MISS) ] = -1,
497 },
498 [ C(OP_PREFETCH) ] = {
499 [ C(RESULT_ACCESS) ] = -1,
500 [ C(RESULT_MISS) ] = -1,
501 },
502 },
503 };
504
505 static __initconst const u64 core2_hw_cache_event_ids
506 [PERF_COUNT_HW_CACHE_MAX]
507 [PERF_COUNT_HW_CACHE_OP_MAX]
508 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
509 {
510 [ C(L1D) ] = {
511 [ C(OP_READ) ] = {
512 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
513 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
514 },
515 [ C(OP_WRITE) ] = {
516 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
517 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
518 },
519 [ C(OP_PREFETCH) ] = {
520 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */
521 [ C(RESULT_MISS) ] = 0,
522 },
523 },
524 [ C(L1I ) ] = {
525 [ C(OP_READ) ] = {
526 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */
527 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */
528 },
529 [ C(OP_WRITE) ] = {
530 [ C(RESULT_ACCESS) ] = -1,
531 [ C(RESULT_MISS) ] = -1,
532 },
533 [ C(OP_PREFETCH) ] = {
534 [ C(RESULT_ACCESS) ] = 0,
535 [ C(RESULT_MISS) ] = 0,
536 },
537 },
538 [ C(LL ) ] = {
539 [ C(OP_READ) ] = {
540 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
541 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
542 },
543 [ C(OP_WRITE) ] = {
544 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
545 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
546 },
547 [ C(OP_PREFETCH) ] = {
548 [ C(RESULT_ACCESS) ] = 0,
549 [ C(RESULT_MISS) ] = 0,
550 },
551 },
552 [ C(DTLB) ] = {
553 [ C(OP_READ) ] = {
554 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
555 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */
556 },
557 [ C(OP_WRITE) ] = {
558 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
559 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */
560 },
561 [ C(OP_PREFETCH) ] = {
562 [ C(RESULT_ACCESS) ] = 0,
563 [ C(RESULT_MISS) ] = 0,
564 },
565 },
566 [ C(ITLB) ] = {
567 [ C(OP_READ) ] = {
568 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
569 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */
570 },
571 [ C(OP_WRITE) ] = {
572 [ C(RESULT_ACCESS) ] = -1,
573 [ C(RESULT_MISS) ] = -1,
574 },
575 [ C(OP_PREFETCH) ] = {
576 [ C(RESULT_ACCESS) ] = -1,
577 [ C(RESULT_MISS) ] = -1,
578 },
579 },
580 [ C(BPU ) ] = {
581 [ C(OP_READ) ] = {
582 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
583 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
584 },
585 [ C(OP_WRITE) ] = {
586 [ C(RESULT_ACCESS) ] = -1,
587 [ C(RESULT_MISS) ] = -1,
588 },
589 [ C(OP_PREFETCH) ] = {
590 [ C(RESULT_ACCESS) ] = -1,
591 [ C(RESULT_MISS) ] = -1,
592 },
593 },
594 };
595
596 static __initconst const u64 atom_hw_cache_event_ids
597 [PERF_COUNT_HW_CACHE_MAX]
598 [PERF_COUNT_HW_CACHE_OP_MAX]
599 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
600 {
601 [ C(L1D) ] = {
602 [ C(OP_READ) ] = {
603 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */
604 [ C(RESULT_MISS) ] = 0,
605 },
606 [ C(OP_WRITE) ] = {
607 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */
608 [ C(RESULT_MISS) ] = 0,
609 },
610 [ C(OP_PREFETCH) ] = {
611 [ C(RESULT_ACCESS) ] = 0x0,
612 [ C(RESULT_MISS) ] = 0,
613 },
614 },
615 [ C(L1I ) ] = {
616 [ C(OP_READ) ] = {
617 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
618 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
619 },
620 [ C(OP_WRITE) ] = {
621 [ C(RESULT_ACCESS) ] = -1,
622 [ C(RESULT_MISS) ] = -1,
623 },
624 [ C(OP_PREFETCH) ] = {
625 [ C(RESULT_ACCESS) ] = 0,
626 [ C(RESULT_MISS) ] = 0,
627 },
628 },
629 [ C(LL ) ] = {
630 [ C(OP_READ) ] = {
631 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
632 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
633 },
634 [ C(OP_WRITE) ] = {
635 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
636 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
637 },
638 [ C(OP_PREFETCH) ] = {
639 [ C(RESULT_ACCESS) ] = 0,
640 [ C(RESULT_MISS) ] = 0,
641 },
642 },
643 [ C(DTLB) ] = {
644 [ C(OP_READ) ] = {
645 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */
646 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */
647 },
648 [ C(OP_WRITE) ] = {
649 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */
650 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */
651 },
652 [ C(OP_PREFETCH) ] = {
653 [ C(RESULT_ACCESS) ] = 0,
654 [ C(RESULT_MISS) ] = 0,
655 },
656 },
657 [ C(ITLB) ] = {
658 [ C(OP_READ) ] = {
659 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
660 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */
661 },
662 [ C(OP_WRITE) ] = {
663 [ C(RESULT_ACCESS) ] = -1,
664 [ C(RESULT_MISS) ] = -1,
665 },
666 [ C(OP_PREFETCH) ] = {
667 [ C(RESULT_ACCESS) ] = -1,
668 [ C(RESULT_MISS) ] = -1,
669 },
670 },
671 [ C(BPU ) ] = {
672 [ C(OP_READ) ] = {
673 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
674 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
675 },
676 [ C(OP_WRITE) ] = {
677 [ C(RESULT_ACCESS) ] = -1,
678 [ C(RESULT_MISS) ] = -1,
679 },
680 [ C(OP_PREFETCH) ] = {
681 [ C(RESULT_ACCESS) ] = -1,
682 [ C(RESULT_MISS) ] = -1,
683 },
684 },
685 };
686
687 static void intel_pmu_disable_all(void)
688 {
689 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
690
691 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
692
693 if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask))
694 intel_pmu_disable_bts();
695
696 intel_pmu_pebs_disable_all();
697 intel_pmu_lbr_disable_all();
698 }
699
700 static void intel_pmu_enable_all(int added)
701 {
702 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
703
704 intel_pmu_pebs_enable_all();
705 intel_pmu_lbr_enable_all();
706 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl);
707
708 if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
709 struct perf_event *event =
710 cpuc->events[X86_PMC_IDX_FIXED_BTS];
711
712 if (WARN_ON_ONCE(!event))
713 return;
714
715 intel_pmu_enable_bts(event->hw.config);
716 }
717 }
718
719 /*
720 * Workaround for:
721 * Intel Errata AAK100 (model 26)
722 * Intel Errata AAP53 (model 30)
723 * Intel Errata BD53 (model 44)
724 *
725 * The official story:
726 * These chips need to be 'reset' when adding counters by programming the
727 * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either
728 * in sequence on the same PMC or on different PMCs.
729 *
730 * In practise it appears some of these events do in fact count, and
731 * we need to programm all 4 events.
732 */
733 static void intel_pmu_nhm_workaround(void)
734 {
735 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
736 static const unsigned long nhm_magic[4] = {
737 0x4300B5,
738 0x4300D2,
739 0x4300B1,
740 0x4300B1
741 };
742 struct perf_event *event;
743 int i;
744
745 /*
746 * The Errata requires below steps:
747 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL;
748 * 2) Configure 4 PERFEVTSELx with the magic events and clear
749 * the corresponding PMCx;
750 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL;
751 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL;
752 * 5) Clear 4 pairs of ERFEVTSELx and PMCx;
753 */
754
755 /*
756 * The real steps we choose are a little different from above.
757 * A) To reduce MSR operations, we don't run step 1) as they
758 * are already cleared before this function is called;
759 * B) Call x86_perf_event_update to save PMCx before configuring
760 * PERFEVTSELx with magic number;
761 * C) With step 5), we do clear only when the PERFEVTSELx is
762 * not used currently.
763 * D) Call x86_perf_event_set_period to restore PMCx;
764 */
765
766 /* We always operate 4 pairs of PERF Counters */
767 for (i = 0; i < 4; i++) {
768 event = cpuc->events[i];
769 if (event)
770 x86_perf_event_update(event);
771 }
772
773 for (i = 0; i < 4; i++) {
774 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]);
775 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0);
776 }
777
778 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf);
779 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
780
781 for (i = 0; i < 4; i++) {
782 event = cpuc->events[i];
783
784 if (event) {
785 x86_perf_event_set_period(event);
786 __x86_pmu_enable_event(&event->hw,
787 ARCH_PERFMON_EVENTSEL_ENABLE);
788 } else
789 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0);
790 }
791 }
792
793 static void intel_pmu_nhm_enable_all(int added)
794 {
795 if (added)
796 intel_pmu_nhm_workaround();
797 intel_pmu_enable_all(added);
798 }
799
800 static inline u64 intel_pmu_get_status(void)
801 {
802 u64 status;
803
804 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
805
806 return status;
807 }
808
809 static inline void intel_pmu_ack_status(u64 ack)
810 {
811 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
812 }
813
814 static void intel_pmu_disable_fixed(struct hw_perf_event *hwc)
815 {
816 int idx = hwc->idx - X86_PMC_IDX_FIXED;
817 u64 ctrl_val, mask;
818
819 mask = 0xfULL << (idx * 4);
820
821 rdmsrl(hwc->config_base, ctrl_val);
822 ctrl_val &= ~mask;
823 wrmsrl(hwc->config_base, ctrl_val);
824 }
825
826 static void intel_pmu_disable_event(struct perf_event *event)
827 {
828 struct hw_perf_event *hwc = &event->hw;
829
830 if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) {
831 intel_pmu_disable_bts();
832 intel_pmu_drain_bts_buffer();
833 return;
834 }
835
836 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
837 intel_pmu_disable_fixed(hwc);
838 return;
839 }
840
841 x86_pmu_disable_event(event);
842
843 if (unlikely(event->attr.precise_ip))
844 intel_pmu_pebs_disable(event);
845 }
846
847 static void intel_pmu_enable_fixed(struct hw_perf_event *hwc)
848 {
849 int idx = hwc->idx - X86_PMC_IDX_FIXED;
850 u64 ctrl_val, bits, mask;
851
852 /*
853 * Enable IRQ generation (0x8),
854 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
855 * if requested:
856 */
857 bits = 0x8ULL;
858 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
859 bits |= 0x2;
860 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
861 bits |= 0x1;
862
863 /*
864 * ANY bit is supported in v3 and up
865 */
866 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
867 bits |= 0x4;
868
869 bits <<= (idx * 4);
870 mask = 0xfULL << (idx * 4);
871
872 rdmsrl(hwc->config_base, ctrl_val);
873 ctrl_val &= ~mask;
874 ctrl_val |= bits;
875 wrmsrl(hwc->config_base, ctrl_val);
876 }
877
878 static void intel_pmu_enable_event(struct perf_event *event)
879 {
880 struct hw_perf_event *hwc = &event->hw;
881
882 if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) {
883 if (!__this_cpu_read(cpu_hw_events.enabled))
884 return;
885
886 intel_pmu_enable_bts(hwc->config);
887 return;
888 }
889
890 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
891 intel_pmu_enable_fixed(hwc);
892 return;
893 }
894
895 if (unlikely(event->attr.precise_ip))
896 intel_pmu_pebs_enable(event);
897
898 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
899 }
900
901 /*
902 * Save and restart an expired event. Called by NMI contexts,
903 * so it has to be careful about preempting normal event ops:
904 */
905 static int intel_pmu_save_and_restart(struct perf_event *event)
906 {
907 x86_perf_event_update(event);
908 return x86_perf_event_set_period(event);
909 }
910
911 static void intel_pmu_reset(void)
912 {
913 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
914 unsigned long flags;
915 int idx;
916
917 if (!x86_pmu.num_counters)
918 return;
919
920 local_irq_save(flags);
921
922 printk("clearing PMU state on CPU#%d\n", smp_processor_id());
923
924 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
925 checking_wrmsrl(x86_pmu_config_addr(idx), 0ull);
926 checking_wrmsrl(x86_pmu_event_addr(idx), 0ull);
927 }
928 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++)
929 checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
930
931 if (ds)
932 ds->bts_index = ds->bts_buffer_base;
933
934 local_irq_restore(flags);
935 }
936
937 /*
938 * This handler is triggered by the local APIC, so the APIC IRQ handling
939 * rules apply:
940 */
941 static int intel_pmu_handle_irq(struct pt_regs *regs)
942 {
943 struct perf_sample_data data;
944 struct cpu_hw_events *cpuc;
945 int bit, loops;
946 u64 status;
947 int handled;
948
949 perf_sample_data_init(&data, 0);
950
951 cpuc = &__get_cpu_var(cpu_hw_events);
952
953 /*
954 * Some chipsets need to unmask the LVTPC in a particular spot
955 * inside the nmi handler. As a result, the unmasking was pushed
956 * into all the nmi handlers.
957 *
958 * This handler doesn't seem to have any issues with the unmasking
959 * so it was left at the top.
960 */
961 apic_write(APIC_LVTPC, APIC_DM_NMI);
962
963 intel_pmu_disable_all();
964 handled = intel_pmu_drain_bts_buffer();
965 status = intel_pmu_get_status();
966 if (!status) {
967 intel_pmu_enable_all(0);
968 return handled;
969 }
970
971 loops = 0;
972 again:
973 intel_pmu_ack_status(status);
974 if (++loops > 100) {
975 WARN_ONCE(1, "perfevents: irq loop stuck!\n");
976 perf_event_print_debug();
977 intel_pmu_reset();
978 goto done;
979 }
980
981 inc_irq_stat(apic_perf_irqs);
982
983 intel_pmu_lbr_read();
984
985 /*
986 * PEBS overflow sets bit 62 in the global status register
987 */
988 if (__test_and_clear_bit(62, (unsigned long *)&status)) {
989 handled++;
990 x86_pmu.drain_pebs(regs);
991 }
992
993 for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
994 struct perf_event *event = cpuc->events[bit];
995
996 handled++;
997
998 if (!test_bit(bit, cpuc->active_mask))
999 continue;
1000
1001 if (!intel_pmu_save_and_restart(event))
1002 continue;
1003
1004 data.period = event->hw.last_period;
1005
1006 if (perf_event_overflow(event, 1, &data, regs))
1007 x86_pmu_stop(event, 0);
1008 }
1009
1010 /*
1011 * Repeat if there is more work to be done:
1012 */
1013 status = intel_pmu_get_status();
1014 if (status)
1015 goto again;
1016
1017 done:
1018 intel_pmu_enable_all(0);
1019 return handled;
1020 }
1021
1022 static struct event_constraint *
1023 intel_bts_constraints(struct perf_event *event)
1024 {
1025 struct hw_perf_event *hwc = &event->hw;
1026 unsigned int hw_event, bts_event;
1027
1028 if (event->attr.freq)
1029 return NULL;
1030
1031 hw_event = hwc->config & INTEL_ARCH_EVENT_MASK;
1032 bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
1033
1034 if (unlikely(hw_event == bts_event && hwc->sample_period == 1))
1035 return &bts_constraint;
1036
1037 return NULL;
1038 }
1039
1040 static struct event_constraint *
1041 intel_percore_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1042 {
1043 struct hw_perf_event *hwc = &event->hw;
1044 unsigned int e = hwc->config & ARCH_PERFMON_EVENTSEL_EVENT;
1045 struct event_constraint *c;
1046 struct intel_percore *pc;
1047 struct er_account *era;
1048 int i;
1049 int free_slot;
1050 int found;
1051
1052 if (!x86_pmu.percore_constraints || hwc->extra_alloc)
1053 return NULL;
1054
1055 for (c = x86_pmu.percore_constraints; c->cmask; c++) {
1056 if (e != c->code)
1057 continue;
1058
1059 /*
1060 * Allocate resource per core.
1061 */
1062 pc = cpuc->per_core;
1063 if (!pc)
1064 break;
1065 c = &emptyconstraint;
1066 raw_spin_lock(&pc->lock);
1067 free_slot = -1;
1068 found = 0;
1069 for (i = 0; i < MAX_EXTRA_REGS; i++) {
1070 era = &pc->regs[i];
1071 if (era->ref > 0 && hwc->extra_reg == era->extra_reg) {
1072 /* Allow sharing same config */
1073 if (hwc->extra_config == era->extra_config) {
1074 era->ref++;
1075 cpuc->percore_used = 1;
1076 hwc->extra_alloc = 1;
1077 c = NULL;
1078 }
1079 /* else conflict */
1080 found = 1;
1081 break;
1082 } else if (era->ref == 0 && free_slot == -1)
1083 free_slot = i;
1084 }
1085 if (!found && free_slot != -1) {
1086 era = &pc->regs[free_slot];
1087 era->ref = 1;
1088 era->extra_reg = hwc->extra_reg;
1089 era->extra_config = hwc->extra_config;
1090 cpuc->percore_used = 1;
1091 hwc->extra_alloc = 1;
1092 c = NULL;
1093 }
1094 raw_spin_unlock(&pc->lock);
1095 return c;
1096 }
1097
1098 return NULL;
1099 }
1100
1101 static struct event_constraint *
1102 intel_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1103 {
1104 struct event_constraint *c;
1105
1106 c = intel_bts_constraints(event);
1107 if (c)
1108 return c;
1109
1110 c = intel_pebs_constraints(event);
1111 if (c)
1112 return c;
1113
1114 c = intel_percore_constraints(cpuc, event);
1115 if (c)
1116 return c;
1117
1118 return x86_get_event_constraints(cpuc, event);
1119 }
1120
1121 static void intel_put_event_constraints(struct cpu_hw_events *cpuc,
1122 struct perf_event *event)
1123 {
1124 struct extra_reg *er;
1125 struct intel_percore *pc;
1126 struct er_account *era;
1127 struct hw_perf_event *hwc = &event->hw;
1128 int i, allref;
1129
1130 if (!cpuc->percore_used)
1131 return;
1132
1133 for (er = x86_pmu.extra_regs; er->msr; er++) {
1134 if (er->event != (hwc->config & er->config_mask))
1135 continue;
1136
1137 pc = cpuc->per_core;
1138 raw_spin_lock(&pc->lock);
1139 for (i = 0; i < MAX_EXTRA_REGS; i++) {
1140 era = &pc->regs[i];
1141 if (era->ref > 0 &&
1142 era->extra_config == hwc->extra_config &&
1143 era->extra_reg == er->msr) {
1144 era->ref--;
1145 hwc->extra_alloc = 0;
1146 break;
1147 }
1148 }
1149 allref = 0;
1150 for (i = 0; i < MAX_EXTRA_REGS; i++)
1151 allref += pc->regs[i].ref;
1152 if (allref == 0)
1153 cpuc->percore_used = 0;
1154 raw_spin_unlock(&pc->lock);
1155 break;
1156 }
1157 }
1158
1159 static int intel_pmu_hw_config(struct perf_event *event)
1160 {
1161 int ret = x86_pmu_hw_config(event);
1162
1163 if (ret)
1164 return ret;
1165
1166 if (event->attr.precise_ip &&
1167 (event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
1168 /*
1169 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
1170 * (0x003c) so that we can use it with PEBS.
1171 *
1172 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
1173 * PEBS capable. However we can use INST_RETIRED.ANY_P
1174 * (0x00c0), which is a PEBS capable event, to get the same
1175 * count.
1176 *
1177 * INST_RETIRED.ANY_P counts the number of cycles that retires
1178 * CNTMASK instructions. By setting CNTMASK to a value (16)
1179 * larger than the maximum number of instructions that can be
1180 * retired per cycle (4) and then inverting the condition, we
1181 * count all cycles that retire 16 or less instructions, which
1182 * is every cycle.
1183 *
1184 * Thereby we gain a PEBS capable cycle counter.
1185 */
1186 u64 alt_config = 0x108000c0; /* INST_RETIRED.TOTAL_CYCLES */
1187
1188 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
1189 event->hw.config = alt_config;
1190 }
1191
1192 if (event->attr.type != PERF_TYPE_RAW)
1193 return 0;
1194
1195 if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY))
1196 return 0;
1197
1198 if (x86_pmu.version < 3)
1199 return -EINVAL;
1200
1201 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1202 return -EACCES;
1203
1204 event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY;
1205
1206 return 0;
1207 }
1208
1209 static __initconst const struct x86_pmu core_pmu = {
1210 .name = "core",
1211 .handle_irq = x86_pmu_handle_irq,
1212 .disable_all = x86_pmu_disable_all,
1213 .enable_all = x86_pmu_enable_all,
1214 .enable = x86_pmu_enable_event,
1215 .disable = x86_pmu_disable_event,
1216 .hw_config = x86_pmu_hw_config,
1217 .schedule_events = x86_schedule_events,
1218 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
1219 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
1220 .event_map = intel_pmu_event_map,
1221 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
1222 .apic = 1,
1223 /*
1224 * Intel PMCs cannot be accessed sanely above 32 bit width,
1225 * so we install an artificial 1<<31 period regardless of
1226 * the generic event period:
1227 */
1228 .max_period = (1ULL << 31) - 1,
1229 .get_event_constraints = intel_get_event_constraints,
1230 .put_event_constraints = intel_put_event_constraints,
1231 .event_constraints = intel_core_event_constraints,
1232 };
1233
1234 static int intel_pmu_cpu_prepare(int cpu)
1235 {
1236 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1237
1238 if (!cpu_has_ht_siblings())
1239 return NOTIFY_OK;
1240
1241 cpuc->per_core = kzalloc_node(sizeof(struct intel_percore),
1242 GFP_KERNEL, cpu_to_node(cpu));
1243 if (!cpuc->per_core)
1244 return NOTIFY_BAD;
1245
1246 raw_spin_lock_init(&cpuc->per_core->lock);
1247 cpuc->per_core->core_id = -1;
1248 return NOTIFY_OK;
1249 }
1250
1251 static void intel_pmu_cpu_starting(int cpu)
1252 {
1253 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1254 int core_id = topology_core_id(cpu);
1255 int i;
1256
1257 init_debug_store_on_cpu(cpu);
1258 /*
1259 * Deal with CPUs that don't clear their LBRs on power-up.
1260 */
1261 intel_pmu_lbr_reset();
1262
1263 if (!cpu_has_ht_siblings())
1264 return;
1265
1266 for_each_cpu(i, topology_thread_cpumask(cpu)) {
1267 struct intel_percore *pc = per_cpu(cpu_hw_events, i).per_core;
1268
1269 if (pc && pc->core_id == core_id) {
1270 kfree(cpuc->per_core);
1271 cpuc->per_core = pc;
1272 break;
1273 }
1274 }
1275
1276 cpuc->per_core->core_id = core_id;
1277 cpuc->per_core->refcnt++;
1278 }
1279
1280 static void intel_pmu_cpu_dying(int cpu)
1281 {
1282 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1283 struct intel_percore *pc = cpuc->per_core;
1284
1285 if (pc) {
1286 if (pc->core_id == -1 || --pc->refcnt == 0)
1287 kfree(pc);
1288 cpuc->per_core = NULL;
1289 }
1290
1291 fini_debug_store_on_cpu(cpu);
1292 }
1293
1294 static __initconst const struct x86_pmu intel_pmu = {
1295 .name = "Intel",
1296 .handle_irq = intel_pmu_handle_irq,
1297 .disable_all = intel_pmu_disable_all,
1298 .enable_all = intel_pmu_enable_all,
1299 .enable = intel_pmu_enable_event,
1300 .disable = intel_pmu_disable_event,
1301 .hw_config = intel_pmu_hw_config,
1302 .schedule_events = x86_schedule_events,
1303 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
1304 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
1305 .event_map = intel_pmu_event_map,
1306 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
1307 .apic = 1,
1308 /*
1309 * Intel PMCs cannot be accessed sanely above 32 bit width,
1310 * so we install an artificial 1<<31 period regardless of
1311 * the generic event period:
1312 */
1313 .max_period = (1ULL << 31) - 1,
1314 .get_event_constraints = intel_get_event_constraints,
1315 .put_event_constraints = intel_put_event_constraints,
1316
1317 .cpu_prepare = intel_pmu_cpu_prepare,
1318 .cpu_starting = intel_pmu_cpu_starting,
1319 .cpu_dying = intel_pmu_cpu_dying,
1320 };
1321
1322 static void intel_clovertown_quirks(void)
1323 {
1324 /*
1325 * PEBS is unreliable due to:
1326 *
1327 * AJ67 - PEBS may experience CPL leaks
1328 * AJ68 - PEBS PMI may be delayed by one event
1329 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
1330 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
1331 *
1332 * AJ67 could be worked around by restricting the OS/USR flags.
1333 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
1334 *
1335 * AJ106 could possibly be worked around by not allowing LBR
1336 * usage from PEBS, including the fixup.
1337 * AJ68 could possibly be worked around by always programming
1338 * a pebs_event_reset[0] value and coping with the lost events.
1339 *
1340 * But taken together it might just make sense to not enable PEBS on
1341 * these chips.
1342 */
1343 printk(KERN_WARNING "PEBS disabled due to CPU errata.\n");
1344 x86_pmu.pebs = 0;
1345 x86_pmu.pebs_constraints = NULL;
1346 }
1347
1348 static __init int intel_pmu_init(void)
1349 {
1350 union cpuid10_edx edx;
1351 union cpuid10_eax eax;
1352 unsigned int unused;
1353 unsigned int ebx;
1354 int version;
1355
1356 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
1357 switch (boot_cpu_data.x86) {
1358 case 0x6:
1359 return p6_pmu_init();
1360 case 0xf:
1361 return p4_pmu_init();
1362 }
1363 return -ENODEV;
1364 }
1365
1366 /*
1367 * Check whether the Architectural PerfMon supports
1368 * Branch Misses Retired hw_event or not.
1369 */
1370 cpuid(10, &eax.full, &ebx, &unused, &edx.full);
1371 if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
1372 return -ENODEV;
1373
1374 version = eax.split.version_id;
1375 if (version < 2)
1376 x86_pmu = core_pmu;
1377 else
1378 x86_pmu = intel_pmu;
1379
1380 x86_pmu.version = version;
1381 x86_pmu.num_counters = eax.split.num_counters;
1382 x86_pmu.cntval_bits = eax.split.bit_width;
1383 x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1;
1384
1385 /*
1386 * Quirk: v2 perfmon does not report fixed-purpose events, so
1387 * assume at least 3 events:
1388 */
1389 if (version > 1)
1390 x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3);
1391
1392 /*
1393 * v2 and above have a perf capabilities MSR
1394 */
1395 if (version > 1) {
1396 u64 capabilities;
1397
1398 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
1399 x86_pmu.intel_cap.capabilities = capabilities;
1400 }
1401
1402 intel_ds_init();
1403
1404 /*
1405 * Install the hw-cache-events table:
1406 */
1407 switch (boot_cpu_data.x86_model) {
1408 case 14: /* 65 nm core solo/duo, "Yonah" */
1409 pr_cont("Core events, ");
1410 break;
1411
1412 case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
1413 x86_pmu.quirks = intel_clovertown_quirks;
1414 case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */
1415 case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */
1416 case 29: /* six-core 45 nm xeon "Dunnington" */
1417 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
1418 sizeof(hw_cache_event_ids));
1419
1420 intel_pmu_lbr_init_core();
1421
1422 x86_pmu.event_constraints = intel_core2_event_constraints;
1423 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints;
1424 pr_cont("Core2 events, ");
1425 break;
1426
1427 case 26: /* 45 nm nehalem, "Bloomfield" */
1428 case 30: /* 45 nm nehalem, "Lynnfield" */
1429 case 46: /* 45 nm nehalem-ex, "Beckton" */
1430 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
1431 sizeof(hw_cache_event_ids));
1432 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
1433 sizeof(hw_cache_extra_regs));
1434
1435 intel_pmu_lbr_init_nhm();
1436
1437 x86_pmu.event_constraints = intel_nehalem_event_constraints;
1438 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints;
1439 x86_pmu.percore_constraints = intel_nehalem_percore_constraints;
1440 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
1441 x86_pmu.extra_regs = intel_nehalem_extra_regs;
1442
1443 /* UOPS_ISSUED.STALLED_CYCLES */
1444 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x180010e;
1445 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
1446 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x1803fb1;
1447
1448 if (ebx & 0x40) {
1449 /*
1450 * Erratum AAJ80 detected, we work it around by using
1451 * the BR_MISP_EXEC.ANY event. This will over-count
1452 * branch-misses, but it's still much better than the
1453 * architectural event which is often completely bogus:
1454 */
1455 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89;
1456
1457 pr_cont("erratum AAJ80 worked around, ");
1458 }
1459 pr_cont("Nehalem events, ");
1460 break;
1461
1462 case 28: /* Atom */
1463 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
1464 sizeof(hw_cache_event_ids));
1465
1466 intel_pmu_lbr_init_atom();
1467
1468 x86_pmu.event_constraints = intel_gen_event_constraints;
1469 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints;
1470 pr_cont("Atom events, ");
1471 break;
1472
1473 case 37: /* 32 nm nehalem, "Clarkdale" */
1474 case 44: /* 32 nm nehalem, "Gulftown" */
1475 case 47: /* 32 nm Xeon E7 */
1476 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
1477 sizeof(hw_cache_event_ids));
1478 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
1479 sizeof(hw_cache_extra_regs));
1480
1481 intel_pmu_lbr_init_nhm();
1482
1483 x86_pmu.event_constraints = intel_westmere_event_constraints;
1484 x86_pmu.percore_constraints = intel_westmere_percore_constraints;
1485 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
1486 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints;
1487 x86_pmu.extra_regs = intel_westmere_extra_regs;
1488
1489 /* UOPS_ISSUED.STALLED_CYCLES */
1490 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x180010e;
1491 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
1492 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x1803fb1;
1493
1494 pr_cont("Westmere events, ");
1495 break;
1496
1497 case 42: /* SandyBridge */
1498 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
1499 sizeof(hw_cache_event_ids));
1500
1501 intel_pmu_lbr_init_nhm();
1502
1503 x86_pmu.event_constraints = intel_snb_event_constraints;
1504 x86_pmu.pebs_constraints = intel_snb_pebs_events;
1505
1506 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
1507 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x180010e;
1508 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
1509 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x18001b1;
1510
1511 pr_cont("SandyBridge events, ");
1512 break;
1513
1514 default:
1515 /*
1516 * default constraints for v2 and up
1517 */
1518 x86_pmu.event_constraints = intel_gen_event_constraints;
1519 pr_cont("generic architected perfmon, ");
1520 }
1521 return 0;
1522 }
1523
1524 #else /* CONFIG_CPU_SUP_INTEL */
1525
1526 static int intel_pmu_init(void)
1527 {
1528 return 0;
1529 }
1530
1531 #endif /* CONFIG_CPU_SUP_INTEL */
Linux Btrfs Development