1 /*
2 * Per core/cpu state
3 *
4 * Used to coordinate shared registers between HT threads or
5 * among events on a single PMU.
6 */
7
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10 #include <linux/stddef.h>
11 #include <linux/types.h>
12 #include <linux/init.h>
13 #include <linux/slab.h>
14 #include <linux/export.h>
15
16 #include <asm/hardirq.h>
17 #include <asm/apic.h>
18
19 #include "perf_event.h"
20
21 /*
22 * Intel PerfMon, used on Core and later.
23 */
24 static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly =
25 {
26 [PERF_COUNT_HW_CPU_CYCLES] = 0x003c,
27 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
28 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e,
29 [PERF_COUNT_HW_CACHE_MISSES] = 0x412e,
30 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4,
31 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5,
32 [PERF_COUNT_HW_BUS_CYCLES] = 0x013c,
33 [PERF_COUNT_HW_REF_CPU_CYCLES] = 0x0300, /* pseudo-encoding */
34 };
35
36 static struct event_constraint intel_core_event_constraints[] __read_mostly =
37 {
38 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
39 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
40 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
41 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
42 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
43 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */
44 EVENT_CONSTRAINT_END
45 };
46
47 static struct event_constraint intel_core2_event_constraints[] __read_mostly =
48 {
49 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
50 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
51 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
52 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
53 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
54 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
55 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
56 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
57 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
58 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
59 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
60 INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */
61 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
62 EVENT_CONSTRAINT_END
63 };
64
65 static struct event_constraint intel_nehalem_event_constraints[] __read_mostly =
66 {
67 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
68 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
69 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
70 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
71 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
72 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
73 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
74 INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */
75 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
76 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
77 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
78 EVENT_CONSTRAINT_END
79 };
80
81 static struct extra_reg intel_nehalem_extra_regs[] __read_mostly =
82 {
83 INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
84 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
85 EVENT_EXTRA_END
86 };
87
88 static struct event_constraint intel_westmere_event_constraints[] __read_mostly =
89 {
90 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
91 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
92 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
93 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
94 INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */
95 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
96 INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */
97 EVENT_CONSTRAINT_END
98 };
99
100 static struct event_constraint intel_snb_event_constraints[] __read_mostly =
101 {
102 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
103 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
104 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
105 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
106 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
107 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
108 INTEL_UEVENT_CONSTRAINT(0x06a3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
109 INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */
110 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
111 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
112 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
113 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
114 EVENT_CONSTRAINT_END
115 };
116
117 static struct event_constraint intel_ivb_event_constraints[] __read_mostly =
118 {
119 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
120 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
121 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
122 INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */
123 INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMTPY */
124 INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */
125 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
126 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
127 INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */
128 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
129 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
130 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
131 /*
132 * Errata BV98 -- MEM_*_RETIRED events can leak between counters of SMT
133 * siblings; disable these events because they can corrupt unrelated
134 * counters.
135 */
136 INTEL_EVENT_CONSTRAINT(0xd0, 0x0), /* MEM_UOPS_RETIRED.* */
137 INTEL_EVENT_CONSTRAINT(0xd1, 0x0), /* MEM_LOAD_UOPS_RETIRED.* */
138 INTEL_EVENT_CONSTRAINT(0xd2, 0x0), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
139 INTEL_EVENT_CONSTRAINT(0xd3, 0x0), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
140 EVENT_CONSTRAINT_END
141 };
142
143 static struct extra_reg intel_westmere_extra_regs[] __read_mostly =
144 {
145 INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
146 INTEL_EVENT_EXTRA_REG(0xbb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1),
147 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
148 EVENT_EXTRA_END
149 };
150
151 static struct event_constraint intel_v1_event_constraints[] __read_mostly =
152 {
153 EVENT_CONSTRAINT_END
154 };
155
156 static struct event_constraint intel_gen_event_constraints[] __read_mostly =
157 {
158 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
159 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
160 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
161 EVENT_CONSTRAINT_END
162 };
163
164 static struct extra_reg intel_snb_extra_regs[] __read_mostly = {
165 INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0),
166 INTEL_EVENT_EXTRA_REG(0xbb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1),
167 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
168 EVENT_EXTRA_END
169 };
170
171 static struct extra_reg intel_snbep_extra_regs[] __read_mostly = {
172 INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
173 INTEL_EVENT_EXTRA_REG(0xbb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
174 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
175 EVENT_EXTRA_END
176 };
177
178 EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3");
179 EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3");
180 EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2");
181
182 struct attribute *nhm_events_attrs[] = {
183 EVENT_PTR(mem_ld_nhm),
184 NULL,
185 };
186
187 struct attribute *snb_events_attrs[] = {
188 EVENT_PTR(mem_ld_snb),
189 EVENT_PTR(mem_st_snb),
190 NULL,
191 };
192
intel_pmu_event_map(int hw_event)193 static u64 intel_pmu_event_map(int hw_event)
194 {
195 return intel_perfmon_event_map[hw_event];
196 }
197
198 #define SNB_DMND_DATA_RD (1ULL << 0)
199 #define SNB_DMND_RFO (1ULL << 1)
200 #define SNB_DMND_IFETCH (1ULL << 2)
201 #define SNB_DMND_WB (1ULL << 3)
202 #define SNB_PF_DATA_RD (1ULL << 4)
203 #define SNB_PF_RFO (1ULL << 5)
204 #define SNB_PF_IFETCH (1ULL << 6)
205 #define SNB_LLC_DATA_RD (1ULL << 7)
206 #define SNB_LLC_RFO (1ULL << 8)
207 #define SNB_LLC_IFETCH (1ULL << 9)
208 #define SNB_BUS_LOCKS (1ULL << 10)
209 #define SNB_STRM_ST (1ULL << 11)
210 #define SNB_OTHER (1ULL << 15)
211 #define SNB_RESP_ANY (1ULL << 16)
212 #define SNB_NO_SUPP (1ULL << 17)
213 #define SNB_LLC_HITM (1ULL << 18)
214 #define SNB_LLC_HITE (1ULL << 19)
215 #define SNB_LLC_HITS (1ULL << 20)
216 #define SNB_LLC_HITF (1ULL << 21)
217 #define SNB_LOCAL (1ULL << 22)
218 #define SNB_REMOTE (0xffULL << 23)
219 #define SNB_SNP_NONE (1ULL << 31)
220 #define SNB_SNP_NOT_NEEDED (1ULL << 32)
221 #define SNB_SNP_MISS (1ULL << 33)
222 #define SNB_NO_FWD (1ULL << 34)
223 #define SNB_SNP_FWD (1ULL << 35)
224 #define SNB_HITM (1ULL << 36)
225 #define SNB_NON_DRAM (1ULL << 37)
226
227 #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD)
228 #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO)
229 #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
230
231 #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \
232 SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \
233 SNB_HITM)
234
235 #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY)
236 #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY)
237
238 #define SNB_L3_ACCESS SNB_RESP_ANY
239 #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM)
240
241 static __initconst const u64 snb_hw_cache_extra_regs
242 [PERF_COUNT_HW_CACHE_MAX]
243 [PERF_COUNT_HW_CACHE_OP_MAX]
244 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
245 {
246 [ C(LL ) ] = {
247 [ C(OP_READ) ] = {
248 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS,
249 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_L3_MISS,
250 },
251 [ C(OP_WRITE) ] = {
252 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS,
253 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_L3_MISS,
254 },
255 [ C(OP_PREFETCH) ] = {
256 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS,
257 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_L3_MISS,
258 },
259 },
260 [ C(NODE) ] = {
261 [ C(OP_READ) ] = {
262 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY,
263 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_DRAM_REMOTE,
264 },
265 [ C(OP_WRITE) ] = {
266 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY,
267 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE,
268 },
269 [ C(OP_PREFETCH) ] = {
270 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY,
271 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE,
272 },
273 },
274 };
275
276 static __initconst const u64 snb_hw_cache_event_ids
277 [PERF_COUNT_HW_CACHE_MAX]
278 [PERF_COUNT_HW_CACHE_OP_MAX]
279 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
280 {
281 [ C(L1D) ] = {
282 [ C(OP_READ) ] = {
283 [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */
284 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */
285 },
286 [ C(OP_WRITE) ] = {
287 [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */
288 [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */
289 },
290 [ C(OP_PREFETCH) ] = {
291 [ C(RESULT_ACCESS) ] = 0x0,
292 [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */
293 },
294 },
295 [ C(L1I ) ] = {
296 [ C(OP_READ) ] = {
297 [ C(RESULT_ACCESS) ] = 0x0,
298 [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */
299 },
300 [ C(OP_WRITE) ] = {
301 [ C(RESULT_ACCESS) ] = -1,
302 [ C(RESULT_MISS) ] = -1,
303 },
304 [ C(OP_PREFETCH) ] = {
305 [ C(RESULT_ACCESS) ] = 0x0,
306 [ C(RESULT_MISS) ] = 0x0,
307 },
308 },
309 [ C(LL ) ] = {
310 [ C(OP_READ) ] = {
311 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
312 [ C(RESULT_ACCESS) ] = 0x01b7,
313 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
314 [ C(RESULT_MISS) ] = 0x01b7,
315 },
316 [ C(OP_WRITE) ] = {
317 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
318 [ C(RESULT_ACCESS) ] = 0x01b7,
319 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
320 [ C(RESULT_MISS) ] = 0x01b7,
321 },
322 [ C(OP_PREFETCH) ] = {
323 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
324 [ C(RESULT_ACCESS) ] = 0x01b7,
325 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
326 [ C(RESULT_MISS) ] = 0x01b7,
327 },
328 },
329 [ C(DTLB) ] = {
330 [ C(OP_READ) ] = {
331 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */
332 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */
333 },
334 [ C(OP_WRITE) ] = {
335 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */
336 [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
337 },
338 [ C(OP_PREFETCH) ] = {
339 [ C(RESULT_ACCESS) ] = 0x0,
340 [ C(RESULT_MISS) ] = 0x0,
341 },
342 },
343 [ C(ITLB) ] = {
344 [ C(OP_READ) ] = {
345 [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */
346 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */
347 },
348 [ C(OP_WRITE) ] = {
349 [ C(RESULT_ACCESS) ] = -1,
350 [ C(RESULT_MISS) ] = -1,
351 },
352 [ C(OP_PREFETCH) ] = {
353 [ C(RESULT_ACCESS) ] = -1,
354 [ C(RESULT_MISS) ] = -1,
355 },
356 },
357 [ C(BPU ) ] = {
358 [ C(OP_READ) ] = {
359 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
360 [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
361 },
362 [ C(OP_WRITE) ] = {
363 [ C(RESULT_ACCESS) ] = -1,
364 [ C(RESULT_MISS) ] = -1,
365 },
366 [ C(OP_PREFETCH) ] = {
367 [ C(RESULT_ACCESS) ] = -1,
368 [ C(RESULT_MISS) ] = -1,
369 },
370 },
371 [ C(NODE) ] = {
372 [ C(OP_READ) ] = {
373 [ C(RESULT_ACCESS) ] = 0x01b7,
374 [ C(RESULT_MISS) ] = 0x01b7,
375 },
376 [ C(OP_WRITE) ] = {
377 [ C(RESULT_ACCESS) ] = 0x01b7,
378 [ C(RESULT_MISS) ] = 0x01b7,
379 },
380 [ C(OP_PREFETCH) ] = {
381 [ C(RESULT_ACCESS) ] = 0x01b7,
382 [ C(RESULT_MISS) ] = 0x01b7,
383 },
384 },
385
386 };
387
388 static __initconst const u64 westmere_hw_cache_event_ids
389 [PERF_COUNT_HW_CACHE_MAX]
390 [PERF_COUNT_HW_CACHE_OP_MAX]
391 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
392 {
393 [ C(L1D) ] = {
394 [ C(OP_READ) ] = {
395 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
396 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
397 },
398 [ C(OP_WRITE) ] = {
399 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
400 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
401 },
402 [ C(OP_PREFETCH) ] = {
403 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
404 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
405 },
406 },
407 [ C(L1I ) ] = {
408 [ C(OP_READ) ] = {
409 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
410 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
411 },
412 [ C(OP_WRITE) ] = {
413 [ C(RESULT_ACCESS) ] = -1,
414 [ C(RESULT_MISS) ] = -1,
415 },
416 [ C(OP_PREFETCH) ] = {
417 [ C(RESULT_ACCESS) ] = 0x0,
418 [ C(RESULT_MISS) ] = 0x0,
419 },
420 },
421 [ C(LL ) ] = {
422 [ C(OP_READ) ] = {
423 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
424 [ C(RESULT_ACCESS) ] = 0x01b7,
425 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
426 [ C(RESULT_MISS) ] = 0x01b7,
427 },
428 /*
429 * Use RFO, not WRITEBACK, because a write miss would typically occur
430 * on RFO.
431 */
432 [ C(OP_WRITE) ] = {
433 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
434 [ C(RESULT_ACCESS) ] = 0x01b7,
435 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
436 [ C(RESULT_MISS) ] = 0x01b7,
437 },
438 [ C(OP_PREFETCH) ] = {
439 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
440 [ C(RESULT_ACCESS) ] = 0x01b7,
441 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
442 [ C(RESULT_MISS) ] = 0x01b7,
443 },
444 },
445 [ C(DTLB) ] = {
446 [ C(OP_READ) ] = {
447 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
448 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
449 },
450 [ C(OP_WRITE) ] = {
451 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
452 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
453 },
454 [ C(OP_PREFETCH) ] = {
455 [ C(RESULT_ACCESS) ] = 0x0,
456 [ C(RESULT_MISS) ] = 0x0,
457 },
458 },
459 [ C(ITLB) ] = {
460 [ C(OP_READ) ] = {
461 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
462 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */
463 },
464 [ C(OP_WRITE) ] = {
465 [ C(RESULT_ACCESS) ] = -1,
466 [ C(RESULT_MISS) ] = -1,
467 },
468 [ C(OP_PREFETCH) ] = {
469 [ C(RESULT_ACCESS) ] = -1,
470 [ C(RESULT_MISS) ] = -1,
471 },
472 },
473 [ C(BPU ) ] = {
474 [ C(OP_READ) ] = {
475 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
476 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
477 },
478 [ C(OP_WRITE) ] = {
479 [ C(RESULT_ACCESS) ] = -1,
480 [ C(RESULT_MISS) ] = -1,
481 },
482 [ C(OP_PREFETCH) ] = {
483 [ C(RESULT_ACCESS) ] = -1,
484 [ C(RESULT_MISS) ] = -1,
485 },
486 },
487 [ C(NODE) ] = {
488 [ C(OP_READ) ] = {
489 [ C(RESULT_ACCESS) ] = 0x01b7,
490 [ C(RESULT_MISS) ] = 0x01b7,
491 },
492 [ C(OP_WRITE) ] = {
493 [ C(RESULT_ACCESS) ] = 0x01b7,
494 [ C(RESULT_MISS) ] = 0x01b7,
495 },
496 [ C(OP_PREFETCH) ] = {
497 [ C(RESULT_ACCESS) ] = 0x01b7,
498 [ C(RESULT_MISS) ] = 0x01b7,
499 },
500 },
501 };
502
503 /*
504 * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits;
505 * See IA32 SDM Vol 3B 30.6.1.3
506 */
507
508 #define NHM_DMND_DATA_RD (1 << 0)
509 #define NHM_DMND_RFO (1 << 1)
510 #define NHM_DMND_IFETCH (1 << 2)
511 #define NHM_DMND_WB (1 << 3)
512 #define NHM_PF_DATA_RD (1 << 4)
513 #define NHM_PF_DATA_RFO (1 << 5)
514 #define NHM_PF_IFETCH (1 << 6)
515 #define NHM_OFFCORE_OTHER (1 << 7)
516 #define NHM_UNCORE_HIT (1 << 8)
517 #define NHM_OTHER_CORE_HIT_SNP (1 << 9)
518 #define NHM_OTHER_CORE_HITM (1 << 10)
519 /* reserved */
520 #define NHM_REMOTE_CACHE_FWD (1 << 12)
521 #define NHM_REMOTE_DRAM (1 << 13)
522 #define NHM_LOCAL_DRAM (1 << 14)
523 #define NHM_NON_DRAM (1 << 15)
524
525 #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD)
526 #define NHM_REMOTE (NHM_REMOTE_DRAM)
527
528 #define NHM_DMND_READ (NHM_DMND_DATA_RD)
529 #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB)
530 #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO)
531
532 #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM)
533 #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD)
534 #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS)
535
536 static __initconst const u64 nehalem_hw_cache_extra_regs
537 [PERF_COUNT_HW_CACHE_MAX]
538 [PERF_COUNT_HW_CACHE_OP_MAX]
539 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
540 {
541 [ C(LL ) ] = {
542 [ C(OP_READ) ] = {
543 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS,
544 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS,
545 },
546 [ C(OP_WRITE) ] = {
547 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS,
548 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS,
549 },
550 [ C(OP_PREFETCH) ] = {
551 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS,
552 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS,
553 },
554 },
555 [ C(NODE) ] = {
556 [ C(OP_READ) ] = {
557 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE,
558 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE,
559 },
560 [ C(OP_WRITE) ] = {
561 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE,
562 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE,
563 },
564 [ C(OP_PREFETCH) ] = {
565 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE,
566 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE,
567 },
568 },
569 };
570
571 static __initconst const u64 nehalem_hw_cache_event_ids
572 [PERF_COUNT_HW_CACHE_MAX]
573 [PERF_COUNT_HW_CACHE_OP_MAX]
574 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
575 {
576 [ C(L1D) ] = {
577 [ C(OP_READ) ] = {
578 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
579 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
580 },
581 [ C(OP_WRITE) ] = {
582 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
583 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
584 },
585 [ C(OP_PREFETCH) ] = {
586 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
587 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
588 },
589 },
590 [ C(L1I ) ] = {
591 [ C(OP_READ) ] = {
592 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
593 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
594 },
595 [ C(OP_WRITE) ] = {
596 [ C(RESULT_ACCESS) ] = -1,
597 [ C(RESULT_MISS) ] = -1,
598 },
599 [ C(OP_PREFETCH) ] = {
600 [ C(RESULT_ACCESS) ] = 0x0,
601 [ C(RESULT_MISS) ] = 0x0,
602 },
603 },
604 [ C(LL ) ] = {
605 [ C(OP_READ) ] = {
606 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
607 [ C(RESULT_ACCESS) ] = 0x01b7,
608 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
609 [ C(RESULT_MISS) ] = 0x01b7,
610 },
611 /*
612 * Use RFO, not WRITEBACK, because a write miss would typically occur
613 * on RFO.
614 */
615 [ C(OP_WRITE) ] = {
616 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
617 [ C(RESULT_ACCESS) ] = 0x01b7,
618 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
619 [ C(RESULT_MISS) ] = 0x01b7,
620 },
621 [ C(OP_PREFETCH) ] = {
622 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
623 [ C(RESULT_ACCESS) ] = 0x01b7,
624 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
625 [ C(RESULT_MISS) ] = 0x01b7,
626 },
627 },
628 [ C(DTLB) ] = {
629 [ C(OP_READ) ] = {
630 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
631 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
632 },
633 [ C(OP_WRITE) ] = {
634 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
635 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
636 },
637 [ C(OP_PREFETCH) ] = {
638 [ C(RESULT_ACCESS) ] = 0x0,
639 [ C(RESULT_MISS) ] = 0x0,
640 },
641 },
642 [ C(ITLB) ] = {
643 [ C(OP_READ) ] = {
644 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
645 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */
646 },
647 [ C(OP_WRITE) ] = {
648 [ C(RESULT_ACCESS) ] = -1,
649 [ C(RESULT_MISS) ] = -1,
650 },
651 [ C(OP_PREFETCH) ] = {
652 [ C(RESULT_ACCESS) ] = -1,
653 [ C(RESULT_MISS) ] = -1,
654 },
655 },
656 [ C(BPU ) ] = {
657 [ C(OP_READ) ] = {
658 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
659 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
660 },
661 [ C(OP_WRITE) ] = {
662 [ C(RESULT_ACCESS) ] = -1,
663 [ C(RESULT_MISS) ] = -1,
664 },
665 [ C(OP_PREFETCH) ] = {
666 [ C(RESULT_ACCESS) ] = -1,
667 [ C(RESULT_MISS) ] = -1,
668 },
669 },
670 [ C(NODE) ] = {
671 [ C(OP_READ) ] = {
672 [ C(RESULT_ACCESS) ] = 0x01b7,
673 [ C(RESULT_MISS) ] = 0x01b7,
674 },
675 [ C(OP_WRITE) ] = {
676 [ C(RESULT_ACCESS) ] = 0x01b7,
677 [ C(RESULT_MISS) ] = 0x01b7,
678 },
679 [ C(OP_PREFETCH) ] = {
680 [ C(RESULT_ACCESS) ] = 0x01b7,
681 [ C(RESULT_MISS) ] = 0x01b7,
682 },
683 },
684 };
685
686 static __initconst const u64 core2_hw_cache_event_ids
687 [PERF_COUNT_HW_CACHE_MAX]
688 [PERF_COUNT_HW_CACHE_OP_MAX]
689 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
690 {
691 [ C(L1D) ] = {
692 [ C(OP_READ) ] = {
693 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
694 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
695 },
696 [ C(OP_WRITE) ] = {
697 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
698 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
699 },
700 [ C(OP_PREFETCH) ] = {
701 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */
702 [ C(RESULT_MISS) ] = 0,
703 },
704 },
705 [ C(L1I ) ] = {
706 [ C(OP_READ) ] = {
707 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */
708 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */
709 },
710 [ C(OP_WRITE) ] = {
711 [ C(RESULT_ACCESS) ] = -1,
712 [ C(RESULT_MISS) ] = -1,
713 },
714 [ C(OP_PREFETCH) ] = {
715 [ C(RESULT_ACCESS) ] = 0,
716 [ C(RESULT_MISS) ] = 0,
717 },
718 },
719 [ C(LL ) ] = {
720 [ C(OP_READ) ] = {
721 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
722 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
723 },
724 [ C(OP_WRITE) ] = {
725 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
726 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
727 },
728 [ C(OP_PREFETCH) ] = {
729 [ C(RESULT_ACCESS) ] = 0,
730 [ C(RESULT_MISS) ] = 0,
731 },
732 },
733 [ C(DTLB) ] = {
734 [ C(OP_READ) ] = {
735 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
736 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */
737 },
738 [ C(OP_WRITE) ] = {
739 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
740 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */
741 },
742 [ C(OP_PREFETCH) ] = {
743 [ C(RESULT_ACCESS) ] = 0,
744 [ C(RESULT_MISS) ] = 0,
745 },
746 },
747 [ C(ITLB) ] = {
748 [ C(OP_READ) ] = {
749 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
750 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */
751 },
752 [ C(OP_WRITE) ] = {
753 [ C(RESULT_ACCESS) ] = -1,
754 [ C(RESULT_MISS) ] = -1,
755 },
756 [ C(OP_PREFETCH) ] = {
757 [ C(RESULT_ACCESS) ] = -1,
758 [ C(RESULT_MISS) ] = -1,
759 },
760 },
761 [ C(BPU ) ] = {
762 [ C(OP_READ) ] = {
763 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
764 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
765 },
766 [ C(OP_WRITE) ] = {
767 [ C(RESULT_ACCESS) ] = -1,
768 [ C(RESULT_MISS) ] = -1,
769 },
770 [ C(OP_PREFETCH) ] = {
771 [ C(RESULT_ACCESS) ] = -1,
772 [ C(RESULT_MISS) ] = -1,
773 },
774 },
775 };
776
777 static __initconst const u64 atom_hw_cache_event_ids
778 [PERF_COUNT_HW_CACHE_MAX]
779 [PERF_COUNT_HW_CACHE_OP_MAX]
780 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
781 {
782 [ C(L1D) ] = {
783 [ C(OP_READ) ] = {
784 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */
785 [ C(RESULT_MISS) ] = 0,
786 },
787 [ C(OP_WRITE) ] = {
788 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */
789 [ C(RESULT_MISS) ] = 0,
790 },
791 [ C(OP_PREFETCH) ] = {
792 [ C(RESULT_ACCESS) ] = 0x0,
793 [ C(RESULT_MISS) ] = 0,
794 },
795 },
796 [ C(L1I ) ] = {
797 [ C(OP_READ) ] = {
798 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
799 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
800 },
801 [ C(OP_WRITE) ] = {
802 [ C(RESULT_ACCESS) ] = -1,
803 [ C(RESULT_MISS) ] = -1,
804 },
805 [ C(OP_PREFETCH) ] = {
806 [ C(RESULT_ACCESS) ] = 0,
807 [ C(RESULT_MISS) ] = 0,
808 },
809 },
810 [ C(LL ) ] = {
811 [ C(OP_READ) ] = {
812 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
813 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
814 },
815 [ C(OP_WRITE) ] = {
816 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
817 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
818 },
819 [ C(OP_PREFETCH) ] = {
820 [ C(RESULT_ACCESS) ] = 0,
821 [ C(RESULT_MISS) ] = 0,
822 },
823 },
824 [ C(DTLB) ] = {
825 [ C(OP_READ) ] = {
826 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */
827 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */
828 },
829 [ C(OP_WRITE) ] = {
830 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */
831 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */
832 },
833 [ C(OP_PREFETCH) ] = {
834 [ C(RESULT_ACCESS) ] = 0,
835 [ C(RESULT_MISS) ] = 0,
836 },
837 },
838 [ C(ITLB) ] = {
839 [ C(OP_READ) ] = {
840 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
841 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */
842 },
843 [ C(OP_WRITE) ] = {
844 [ C(RESULT_ACCESS) ] = -1,
845 [ C(RESULT_MISS) ] = -1,
846 },
847 [ C(OP_PREFETCH) ] = {
848 [ C(RESULT_ACCESS) ] = -1,
849 [ C(RESULT_MISS) ] = -1,
850 },
851 },
852 [ C(BPU ) ] = {
853 [ C(OP_READ) ] = {
854 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
855 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
856 },
857 [ C(OP_WRITE) ] = {
858 [ C(RESULT_ACCESS) ] = -1,
859 [ C(RESULT_MISS) ] = -1,
860 },
861 [ C(OP_PREFETCH) ] = {
862 [ C(RESULT_ACCESS) ] = -1,
863 [ C(RESULT_MISS) ] = -1,
864 },
865 },
866 };
867
intel_pmu_needs_lbr_smpl(struct perf_event * event)868 static inline bool intel_pmu_needs_lbr_smpl(struct perf_event *event)
869 {
870 /* user explicitly requested branch sampling */
871 if (has_branch_stack(event))
872 return true;
873
874 /* implicit branch sampling to correct PEBS skid */
875 if (x86_pmu.intel_cap.pebs_trap && event->attr.precise_ip > 1)
876 return true;
877
878 return false;
879 }
880
intel_pmu_disable_all(void)881 static void intel_pmu_disable_all(void)
882 {
883 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
884
885 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
886
887 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask))
888 intel_pmu_disable_bts();
889
890 intel_pmu_pebs_disable_all();
891 intel_pmu_lbr_disable_all();
892 }
893
intel_pmu_enable_all(int added)894 static void intel_pmu_enable_all(int added)
895 {
896 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
897
898 intel_pmu_pebs_enable_all();
899 intel_pmu_lbr_enable_all();
900 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL,
901 x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask);
902
903 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
904 struct perf_event *event =
905 cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
906
907 if (WARN_ON_ONCE(!event))
908 return;
909
910 intel_pmu_enable_bts(event->hw.config);
911 }
912 }
913
914 /*
915 * Workaround for:
916 * Intel Errata AAK100 (model 26)
917 * Intel Errata AAP53 (model 30)
918 * Intel Errata BD53 (model 44)
919 *
920 * The official story:
921 * These chips need to be 'reset' when adding counters by programming the
922 * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either
923 * in sequence on the same PMC or on different PMCs.
924 *
925 * In practise it appears some of these events do in fact count, and
926 * we need to programm all 4 events.
927 */
intel_pmu_nhm_workaround(void)928 static void intel_pmu_nhm_workaround(void)
929 {
930 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
931 static const unsigned long nhm_magic[4] = {
932 0x4300B5,
933 0x4300D2,
934 0x4300B1,
935 0x4300B1
936 };
937 struct perf_event *event;
938 int i;
939
940 /*
941 * The Errata requires below steps:
942 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL;
943 * 2) Configure 4 PERFEVTSELx with the magic events and clear
944 * the corresponding PMCx;
945 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL;
946 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL;
947 * 5) Clear 4 pairs of ERFEVTSELx and PMCx;
948 */
949
950 /*
951 * The real steps we choose are a little different from above.
952 * A) To reduce MSR operations, we don't run step 1) as they
953 * are already cleared before this function is called;
954 * B) Call x86_perf_event_update to save PMCx before configuring
955 * PERFEVTSELx with magic number;
956 * C) With step 5), we do clear only when the PERFEVTSELx is
957 * not used currently.
958 * D) Call x86_perf_event_set_period to restore PMCx;
959 */
960
961 /* We always operate 4 pairs of PERF Counters */
962 for (i = 0; i < 4; i++) {
963 event = cpuc->events[i];
964 if (event)
965 x86_perf_event_update(event);
966 }
967
968 for (i = 0; i < 4; i++) {
969 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]);
970 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0);
971 }
972
973 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf);
974 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
975
976 for (i = 0; i < 4; i++) {
977 event = cpuc->events[i];
978
979 if (event) {
980 x86_perf_event_set_period(event);
981 __x86_pmu_enable_event(&event->hw,
982 ARCH_PERFMON_EVENTSEL_ENABLE);
983 } else
984 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0);
985 }
986 }
987
intel_pmu_nhm_enable_all(int added)988 static void intel_pmu_nhm_enable_all(int added)
989 {
990 if (added)
991 intel_pmu_nhm_workaround();
992 intel_pmu_enable_all(added);
993 }
994
intel_pmu_get_status(void)995 static inline u64 intel_pmu_get_status(void)
996 {
997 u64 status;
998
999 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1000
1001 return status;
1002 }
1003
intel_pmu_ack_status(u64 ack)1004 static inline void intel_pmu_ack_status(u64 ack)
1005 {
1006 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
1007 }
1008
intel_pmu_disable_fixed(struct hw_perf_event * hwc)1009 static void intel_pmu_disable_fixed(struct hw_perf_event *hwc)
1010 {
1011 int idx = hwc->idx - INTEL_PMC_IDX_FIXED;
1012 u64 ctrl_val, mask;
1013
1014 mask = 0xfULL << (idx * 4);
1015
1016 rdmsrl(hwc->config_base, ctrl_val);
1017 ctrl_val &= ~mask;
1018 wrmsrl(hwc->config_base, ctrl_val);
1019 }
1020
intel_pmu_disable_event(struct perf_event * event)1021 static void intel_pmu_disable_event(struct perf_event *event)
1022 {
1023 struct hw_perf_event *hwc = &event->hw;
1024 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1025
1026 if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) {
1027 intel_pmu_disable_bts();
1028 intel_pmu_drain_bts_buffer();
1029 return;
1030 }
1031
1032 cpuc->intel_ctrl_guest_mask &= ~(1ull << hwc->idx);
1033 cpuc->intel_ctrl_host_mask &= ~(1ull << hwc->idx);
1034
1035 /*
1036 * must disable before any actual event
1037 * because any event may be combined with LBR
1038 */
1039 if (intel_pmu_needs_lbr_smpl(event))
1040 intel_pmu_lbr_disable(event);
1041
1042 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
1043 intel_pmu_disable_fixed(hwc);
1044 return;
1045 }
1046
1047 x86_pmu_disable_event(event);
1048
1049 if (unlikely(event->attr.precise_ip))
1050 intel_pmu_pebs_disable(event);
1051 }
1052
intel_pmu_enable_fixed(struct hw_perf_event * hwc)1053 static void intel_pmu_enable_fixed(struct hw_perf_event *hwc)
1054 {
1055 int idx = hwc->idx - INTEL_PMC_IDX_FIXED;
1056 u64 ctrl_val, bits, mask;
1057
1058 /*
1059 * Enable IRQ generation (0x8),
1060 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
1061 * if requested:
1062 */
1063 bits = 0x8ULL;
1064 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
1065 bits |= 0x2;
1066 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
1067 bits |= 0x1;
1068
1069 /*
1070 * ANY bit is supported in v3 and up
1071 */
1072 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
1073 bits |= 0x4;
1074
1075 bits <<= (idx * 4);
1076 mask = 0xfULL << (idx * 4);
1077
1078 rdmsrl(hwc->config_base, ctrl_val);
1079 ctrl_val &= ~mask;
1080 ctrl_val |= bits;
1081 wrmsrl(hwc->config_base, ctrl_val);
1082 }
1083
intel_pmu_enable_event(struct perf_event * event)1084 static void intel_pmu_enable_event(struct perf_event *event)
1085 {
1086 struct hw_perf_event *hwc = &event->hw;
1087 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1088
1089 if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) {
1090 if (!__this_cpu_read(cpu_hw_events.enabled))
1091 return;
1092
1093 intel_pmu_enable_bts(hwc->config);
1094 return;
1095 }
1096 /*
1097 * must enabled before any actual event
1098 * because any event may be combined with LBR
1099 */
1100 if (intel_pmu_needs_lbr_smpl(event))
1101 intel_pmu_lbr_enable(event);
1102
1103 if (event->attr.exclude_host)
1104 cpuc->intel_ctrl_guest_mask |= (1ull << hwc->idx);
1105 if (event->attr.exclude_guest)
1106 cpuc->intel_ctrl_host_mask |= (1ull << hwc->idx);
1107
1108 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
1109 intel_pmu_enable_fixed(hwc);
1110 return;
1111 }
1112
1113 if (unlikely(event->attr.precise_ip))
1114 intel_pmu_pebs_enable(event);
1115
1116 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
1117 }
1118
1119 /*
1120 * Save and restart an expired event. Called by NMI contexts,
1121 * so it has to be careful about preempting normal event ops:
1122 */
intel_pmu_save_and_restart(struct perf_event * event)1123 int intel_pmu_save_and_restart(struct perf_event *event)
1124 {
1125 x86_perf_event_update(event);
1126 return x86_perf_event_set_period(event);
1127 }
1128
intel_pmu_reset(void)1129 static void intel_pmu_reset(void)
1130 {
1131 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
1132 unsigned long flags;
1133 int idx;
1134
1135 if (!x86_pmu.num_counters)
1136 return;
1137
1138 local_irq_save(flags);
1139
1140 pr_info("clearing PMU state on CPU#%d\n", smp_processor_id());
1141
1142 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1143 wrmsrl_safe(x86_pmu_config_addr(idx), 0ull);
1144 wrmsrl_safe(x86_pmu_event_addr(idx), 0ull);
1145 }
1146 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++)
1147 wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
1148
1149 if (ds)
1150 ds->bts_index = ds->bts_buffer_base;
1151
1152 local_irq_restore(flags);
1153 }
1154
1155 /*
1156 * This handler is triggered by the local APIC, so the APIC IRQ handling
1157 * rules apply:
1158 */
intel_pmu_handle_irq(struct pt_regs * regs)1159 static int intel_pmu_handle_irq(struct pt_regs *regs)
1160 {
1161 struct perf_sample_data data;
1162 struct cpu_hw_events *cpuc;
1163 int bit, loops;
1164 u64 status;
1165 int handled;
1166
1167 cpuc = &__get_cpu_var(cpu_hw_events);
1168
1169 /*
1170 * Some chipsets need to unmask the LVTPC in a particular spot
1171 * inside the nmi handler. As a result, the unmasking was pushed
1172 * into all the nmi handlers.
1173 *
1174 * This handler doesn't seem to have any issues with the unmasking
1175 * so it was left at the top.
1176 */
1177 apic_write(APIC_LVTPC, APIC_DM_NMI);
1178
1179 intel_pmu_disable_all();
1180 handled = intel_pmu_drain_bts_buffer();
1181 status = intel_pmu_get_status();
1182 if (!status) {
1183 intel_pmu_enable_all(0);
1184 return handled;
1185 }
1186
1187 loops = 0;
1188 again:
1189 intel_pmu_ack_status(status);
1190 if (++loops > 100) {
1191 WARN_ONCE(1, "perfevents: irq loop stuck!\n");
1192 perf_event_print_debug();
1193 intel_pmu_reset();
1194 goto done;
1195 }
1196
1197 inc_irq_stat(apic_perf_irqs);
1198
1199 intel_pmu_lbr_read();
1200
1201 /*
1202 * PEBS overflow sets bit 62 in the global status register
1203 */
1204 if (__test_and_clear_bit(62, (unsigned long *)&status)) {
1205 handled++;
1206 x86_pmu.drain_pebs(regs);
1207 }
1208
1209 for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
1210 struct perf_event *event = cpuc->events[bit];
1211
1212 handled++;
1213
1214 if (!test_bit(bit, cpuc->active_mask))
1215 continue;
1216
1217 if (!intel_pmu_save_and_restart(event))
1218 continue;
1219
1220 perf_sample_data_init(&data, 0, event->hw.last_period);
1221
1222 if (has_branch_stack(event))
1223 data.br_stack = &cpuc->lbr_stack;
1224
1225 if (perf_event_overflow(event, &data, regs))
1226 x86_pmu_stop(event, 0);
1227 }
1228
1229 /*
1230 * Repeat if there is more work to be done:
1231 */
1232 status = intel_pmu_get_status();
1233 if (status)
1234 goto again;
1235
1236 done:
1237 intel_pmu_enable_all(0);
1238 return handled;
1239 }
1240
1241 static struct event_constraint *
intel_bts_constraints(struct perf_event * event)1242 intel_bts_constraints(struct perf_event *event)
1243 {
1244 struct hw_perf_event *hwc = &event->hw;
1245 unsigned int hw_event, bts_event;
1246
1247 if (event->attr.freq)
1248 return NULL;
1249
1250 hw_event = hwc->config & INTEL_ARCH_EVENT_MASK;
1251 bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
1252
1253 if (unlikely(hw_event == bts_event && hwc->sample_period == 1))
1254 return &bts_constraint;
1255
1256 return NULL;
1257 }
1258
intel_alt_er(int idx)1259 static int intel_alt_er(int idx)
1260 {
1261 if (!(x86_pmu.er_flags & ERF_HAS_RSP_1))
1262 return idx;
1263
1264 if (idx == EXTRA_REG_RSP_0)
1265 return EXTRA_REG_RSP_1;
1266
1267 if (idx == EXTRA_REG_RSP_1)
1268 return EXTRA_REG_RSP_0;
1269
1270 return idx;
1271 }
1272
intel_fixup_er(struct perf_event * event,int idx)1273 static void intel_fixup_er(struct perf_event *event, int idx)
1274 {
1275 event->hw.extra_reg.idx = idx;
1276
1277 if (idx == EXTRA_REG_RSP_0) {
1278 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
1279 event->hw.config |= 0x01b7;
1280 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0;
1281 } else if (idx == EXTRA_REG_RSP_1) {
1282 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
1283 event->hw.config |= 0x01bb;
1284 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1;
1285 }
1286 }
1287
1288 /*
1289 * manage allocation of shared extra msr for certain events
1290 *
1291 * sharing can be:
1292 * per-cpu: to be shared between the various events on a single PMU
1293 * per-core: per-cpu + shared by HT threads
1294 */
1295 static struct event_constraint *
__intel_shared_reg_get_constraints(struct cpu_hw_events * cpuc,struct perf_event * event,struct hw_perf_event_extra * reg)1296 __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc,
1297 struct perf_event *event,
1298 struct hw_perf_event_extra *reg)
1299 {
1300 struct event_constraint *c = &emptyconstraint;
1301 struct er_account *era;
1302 unsigned long flags;
1303 int idx = reg->idx;
1304
1305 /*
1306 * reg->alloc can be set due to existing state, so for fake cpuc we
1307 * need to ignore this, otherwise we might fail to allocate proper fake
1308 * state for this extra reg constraint. Also see the comment below.
1309 */
1310 if (reg->alloc && !cpuc->is_fake)
1311 return NULL; /* call x86_get_event_constraint() */
1312
1313 again:
1314 era = &cpuc->shared_regs->regs[idx];
1315 /*
1316 * we use spin_lock_irqsave() to avoid lockdep issues when
1317 * passing a fake cpuc
1318 */
1319 raw_spin_lock_irqsave(&era->lock, flags);
1320
1321 if (!atomic_read(&era->ref) || era->config == reg->config) {
1322
1323 /*
1324 * If its a fake cpuc -- as per validate_{group,event}() we
1325 * shouldn't touch event state and we can avoid doing so
1326 * since both will only call get_event_constraints() once
1327 * on each event, this avoids the need for reg->alloc.
1328 *
1329 * Not doing the ER fixup will only result in era->reg being
1330 * wrong, but since we won't actually try and program hardware
1331 * this isn't a problem either.
1332 */
1333 if (!cpuc->is_fake) {
1334 if (idx != reg->idx)
1335 intel_fixup_er(event, idx);
1336
1337 /*
1338 * x86_schedule_events() can call get_event_constraints()
1339 * multiple times on events in the case of incremental
1340 * scheduling(). reg->alloc ensures we only do the ER
1341 * allocation once.
1342 */
1343 reg->alloc = 1;
1344 }
1345
1346 /* lock in msr value */
1347 era->config = reg->config;
1348 era->reg = reg->reg;
1349
1350 /* one more user */
1351 atomic_inc(&era->ref);
1352
1353 /*
1354 * need to call x86_get_event_constraint()
1355 * to check if associated event has constraints
1356 */
1357 c = NULL;
1358 } else {
1359 idx = intel_alt_er(idx);
1360 if (idx != reg->idx) {
1361 raw_spin_unlock_irqrestore(&era->lock, flags);
1362 goto again;
1363 }
1364 }
1365 raw_spin_unlock_irqrestore(&era->lock, flags);
1366
1367 return c;
1368 }
1369
1370 static void
__intel_shared_reg_put_constraints(struct cpu_hw_events * cpuc,struct hw_perf_event_extra * reg)1371 __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc,
1372 struct hw_perf_event_extra *reg)
1373 {
1374 struct er_account *era;
1375
1376 /*
1377 * Only put constraint if extra reg was actually allocated. Also takes
1378 * care of event which do not use an extra shared reg.
1379 *
1380 * Also, if this is a fake cpuc we shouldn't touch any event state
1381 * (reg->alloc) and we don't care about leaving inconsistent cpuc state
1382 * either since it'll be thrown out.
1383 */
1384 if (!reg->alloc || cpuc->is_fake)
1385 return;
1386
1387 era = &cpuc->shared_regs->regs[reg->idx];
1388
1389 /* one fewer user */
1390 atomic_dec(&era->ref);
1391
1392 /* allocate again next time */
1393 reg->alloc = 0;
1394 }
1395
1396 static struct event_constraint *
intel_shared_regs_constraints(struct cpu_hw_events * cpuc,struct perf_event * event)1397 intel_shared_regs_constraints(struct cpu_hw_events *cpuc,
1398 struct perf_event *event)
1399 {
1400 struct event_constraint *c = NULL, *d;
1401 struct hw_perf_event_extra *xreg, *breg;
1402
1403 xreg = &event->hw.extra_reg;
1404 if (xreg->idx != EXTRA_REG_NONE) {
1405 c = __intel_shared_reg_get_constraints(cpuc, event, xreg);
1406 if (c == &emptyconstraint)
1407 return c;
1408 }
1409 breg = &event->hw.branch_reg;
1410 if (breg->idx != EXTRA_REG_NONE) {
1411 d = __intel_shared_reg_get_constraints(cpuc, event, breg);
1412 if (d == &emptyconstraint) {
1413 __intel_shared_reg_put_constraints(cpuc, xreg);
1414 c = d;
1415 }
1416 }
1417 return c;
1418 }
1419
1420 struct event_constraint *
x86_get_event_constraints(struct cpu_hw_events * cpuc,struct perf_event * event)1421 x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1422 {
1423 struct event_constraint *c;
1424
1425 if (x86_pmu.event_constraints) {
1426 for_each_event_constraint(c, x86_pmu.event_constraints) {
1427 if ((event->hw.config & c->cmask) == c->code) {
1428 /* hw.flags zeroed at initialization */
1429 event->hw.flags |= c->flags;
1430 return c;
1431 }
1432 }
1433 }
1434
1435 return &unconstrained;
1436 }
1437
1438 static struct event_constraint *
intel_get_event_constraints(struct cpu_hw_events * cpuc,struct perf_event * event)1439 intel_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1440 {
1441 struct event_constraint *c;
1442
1443 c = intel_bts_constraints(event);
1444 if (c)
1445 return c;
1446
1447 c = intel_pebs_constraints(event);
1448 if (c)
1449 return c;
1450
1451 c = intel_shared_regs_constraints(cpuc, event);
1452 if (c)
1453 return c;
1454
1455 return x86_get_event_constraints(cpuc, event);
1456 }
1457
1458 static void
intel_put_shared_regs_event_constraints(struct cpu_hw_events * cpuc,struct perf_event * event)1459 intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc,
1460 struct perf_event *event)
1461 {
1462 struct hw_perf_event_extra *reg;
1463
1464 reg = &event->hw.extra_reg;
1465 if (reg->idx != EXTRA_REG_NONE)
1466 __intel_shared_reg_put_constraints(cpuc, reg);
1467
1468 reg = &event->hw.branch_reg;
1469 if (reg->idx != EXTRA_REG_NONE)
1470 __intel_shared_reg_put_constraints(cpuc, reg);
1471 }
1472
intel_put_event_constraints(struct cpu_hw_events * cpuc,struct perf_event * event)1473 static void intel_put_event_constraints(struct cpu_hw_events *cpuc,
1474 struct perf_event *event)
1475 {
1476 event->hw.flags = 0;
1477 intel_put_shared_regs_event_constraints(cpuc, event);
1478 }
1479
intel_pebs_aliases_core2(struct perf_event * event)1480 static void intel_pebs_aliases_core2(struct perf_event *event)
1481 {
1482 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
1483 /*
1484 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
1485 * (0x003c) so that we can use it with PEBS.
1486 *
1487 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
1488 * PEBS capable. However we can use INST_RETIRED.ANY_P
1489 * (0x00c0), which is a PEBS capable event, to get the same
1490 * count.
1491 *
1492 * INST_RETIRED.ANY_P counts the number of cycles that retires
1493 * CNTMASK instructions. By setting CNTMASK to a value (16)
1494 * larger than the maximum number of instructions that can be
1495 * retired per cycle (4) and then inverting the condition, we
1496 * count all cycles that retire 16 or less instructions, which
1497 * is every cycle.
1498 *
1499 * Thereby we gain a PEBS capable cycle counter.
1500 */
1501 u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16);
1502
1503 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
1504 event->hw.config = alt_config;
1505 }
1506 }
1507
intel_pebs_aliases_snb(struct perf_event * event)1508 static void intel_pebs_aliases_snb(struct perf_event *event)
1509 {
1510 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
1511 /*
1512 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
1513 * (0x003c) so that we can use it with PEBS.
1514 *
1515 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
1516 * PEBS capable. However we can use UOPS_RETIRED.ALL
1517 * (0x01c2), which is a PEBS capable event, to get the same
1518 * count.
1519 *
1520 * UOPS_RETIRED.ALL counts the number of cycles that retires
1521 * CNTMASK micro-ops. By setting CNTMASK to a value (16)
1522 * larger than the maximum number of micro-ops that can be
1523 * retired per cycle (4) and then inverting the condition, we
1524 * count all cycles that retire 16 or less micro-ops, which
1525 * is every cycle.
1526 *
1527 * Thereby we gain a PEBS capable cycle counter.
1528 */
1529 u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16);
1530
1531 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
1532 event->hw.config = alt_config;
1533 }
1534 }
1535
intel_pmu_hw_config(struct perf_event * event)1536 static int intel_pmu_hw_config(struct perf_event *event)
1537 {
1538 int ret = x86_pmu_hw_config(event);
1539
1540 if (ret)
1541 return ret;
1542
1543 if (event->attr.precise_ip && x86_pmu.pebs_aliases)
1544 x86_pmu.pebs_aliases(event);
1545
1546 if (intel_pmu_needs_lbr_smpl(event)) {
1547 ret = intel_pmu_setup_lbr_filter(event);
1548 if (ret)
1549 return ret;
1550 }
1551
1552 if (event->attr.type != PERF_TYPE_RAW)
1553 return 0;
1554
1555 if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY))
1556 return 0;
1557
1558 if (x86_pmu.version < 3)
1559 return -EINVAL;
1560
1561 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1562 return -EACCES;
1563
1564 event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY;
1565
1566 return 0;
1567 }
1568
perf_guest_get_msrs(int * nr)1569 struct perf_guest_switch_msr *perf_guest_get_msrs(int *nr)
1570 {
1571 if (x86_pmu.guest_get_msrs)
1572 return x86_pmu.guest_get_msrs(nr);
1573 *nr = 0;
1574 return NULL;
1575 }
1576 EXPORT_SYMBOL_GPL(perf_guest_get_msrs);
1577
intel_guest_get_msrs(int * nr)1578 static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr)
1579 {
1580 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1581 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
1582
1583 arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL;
1584 arr[0].host = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask;
1585 arr[0].guest = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_host_mask;
1586 /*
1587 * If PMU counter has PEBS enabled it is not enough to disable counter
1588 * on a guest entry since PEBS memory write can overshoot guest entry
1589 * and corrupt guest memory. Disabling PEBS solves the problem.
1590 */
1591 arr[1].msr = MSR_IA32_PEBS_ENABLE;
1592 arr[1].host = cpuc->pebs_enabled;
1593 arr[1].guest = 0;
1594
1595 *nr = 2;
1596 return arr;
1597 }
1598
core_guest_get_msrs(int * nr)1599 static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr)
1600 {
1601 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1602 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
1603 int idx;
1604
1605 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1606 struct perf_event *event = cpuc->events[idx];
1607
1608 arr[idx].msr = x86_pmu_config_addr(idx);
1609 arr[idx].host = arr[idx].guest = 0;
1610
1611 if (!test_bit(idx, cpuc->active_mask))
1612 continue;
1613
1614 arr[idx].host = arr[idx].guest =
1615 event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE;
1616
1617 if (event->attr.exclude_host)
1618 arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
1619 else if (event->attr.exclude_guest)
1620 arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
1621 }
1622
1623 *nr = x86_pmu.num_counters;
1624 return arr;
1625 }
1626
core_pmu_enable_event(struct perf_event * event)1627 static void core_pmu_enable_event(struct perf_event *event)
1628 {
1629 if (!event->attr.exclude_host)
1630 x86_pmu_enable_event(event);
1631 }
1632
core_pmu_enable_all(int added)1633 static void core_pmu_enable_all(int added)
1634 {
1635 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1636 int idx;
1637
1638 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1639 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
1640
1641 if (!test_bit(idx, cpuc->active_mask) ||
1642 cpuc->events[idx]->attr.exclude_host)
1643 continue;
1644
1645 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
1646 }
1647 }
1648
1649 PMU_FORMAT_ATTR(event, "config:0-7" );
1650 PMU_FORMAT_ATTR(umask, "config:8-15" );
1651 PMU_FORMAT_ATTR(edge, "config:18" );
1652 PMU_FORMAT_ATTR(pc, "config:19" );
1653 PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */
1654 PMU_FORMAT_ATTR(inv, "config:23" );
1655 PMU_FORMAT_ATTR(cmask, "config:24-31" );
1656
1657 static struct attribute *intel_arch_formats_attr[] = {
1658 &format_attr_event.attr,
1659 &format_attr_umask.attr,
1660 &format_attr_edge.attr,
1661 &format_attr_pc.attr,
1662 &format_attr_inv.attr,
1663 &format_attr_cmask.attr,
1664 NULL,
1665 };
1666
intel_event_sysfs_show(char * page,u64 config)1667 ssize_t intel_event_sysfs_show(char *page, u64 config)
1668 {
1669 u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT);
1670
1671 return x86_event_sysfs_show(page, config, event);
1672 }
1673
1674 static __initconst const struct x86_pmu core_pmu = {
1675 .name = "core",
1676 .handle_irq = x86_pmu_handle_irq,
1677 .disable_all = x86_pmu_disable_all,
1678 .enable_all = core_pmu_enable_all,
1679 .enable = core_pmu_enable_event,
1680 .disable = x86_pmu_disable_event,
1681 .hw_config = x86_pmu_hw_config,
1682 .schedule_events = x86_schedule_events,
1683 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
1684 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
1685 .event_map = intel_pmu_event_map,
1686 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
1687 .apic = 1,
1688 /*
1689 * Intel PMCs cannot be accessed sanely above 32 bit width,
1690 * so we install an artificial 1<<31 period regardless of
1691 * the generic event period:
1692 */
1693 .max_period = (1ULL << 31) - 1,
1694 .get_event_constraints = intel_get_event_constraints,
1695 .put_event_constraints = intel_put_event_constraints,
1696 .event_constraints = intel_core_event_constraints,
1697 .guest_get_msrs = core_guest_get_msrs,
1698 .format_attrs = intel_arch_formats_attr,
1699 .events_sysfs_show = intel_event_sysfs_show,
1700 };
1701
allocate_shared_regs(int cpu)1702 struct intel_shared_regs *allocate_shared_regs(int cpu)
1703 {
1704 struct intel_shared_regs *regs;
1705 int i;
1706
1707 regs = kzalloc_node(sizeof(struct intel_shared_regs),
1708 GFP_KERNEL, cpu_to_node(cpu));
1709 if (regs) {
1710 /*
1711 * initialize the locks to keep lockdep happy
1712 */
1713 for (i = 0; i < EXTRA_REG_MAX; i++)
1714 raw_spin_lock_init(®s->regs[i].lock);
1715
1716 regs->core_id = -1;
1717 }
1718 return regs;
1719 }
1720
intel_pmu_cpu_prepare(int cpu)1721 static int intel_pmu_cpu_prepare(int cpu)
1722 {
1723 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1724
1725 if (!(x86_pmu.extra_regs || x86_pmu.lbr_sel_map))
1726 return NOTIFY_OK;
1727
1728 cpuc->shared_regs = allocate_shared_regs(cpu);
1729 if (!cpuc->shared_regs)
1730 return NOTIFY_BAD;
1731
1732 return NOTIFY_OK;
1733 }
1734
intel_pmu_cpu_starting(int cpu)1735 static void intel_pmu_cpu_starting(int cpu)
1736 {
1737 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1738 int core_id = topology_core_id(cpu);
1739 int i;
1740
1741 init_debug_store_on_cpu(cpu);
1742 /*
1743 * Deal with CPUs that don't clear their LBRs on power-up.
1744 */
1745 intel_pmu_lbr_reset();
1746
1747 cpuc->lbr_sel = NULL;
1748
1749 if (!cpuc->shared_regs)
1750 return;
1751
1752 if (!(x86_pmu.er_flags & ERF_NO_HT_SHARING)) {
1753 for_each_cpu(i, topology_thread_cpumask(cpu)) {
1754 struct intel_shared_regs *pc;
1755
1756 pc = per_cpu(cpu_hw_events, i).shared_regs;
1757 if (pc && pc->core_id == core_id) {
1758 cpuc->kfree_on_online = cpuc->shared_regs;
1759 cpuc->shared_regs = pc;
1760 break;
1761 }
1762 }
1763 cpuc->shared_regs->core_id = core_id;
1764 cpuc->shared_regs->refcnt++;
1765 }
1766
1767 if (x86_pmu.lbr_sel_map)
1768 cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR];
1769 }
1770
intel_pmu_cpu_dying(int cpu)1771 static void intel_pmu_cpu_dying(int cpu)
1772 {
1773 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1774 struct intel_shared_regs *pc;
1775
1776 pc = cpuc->shared_regs;
1777 if (pc) {
1778 if (pc->core_id == -1 || --pc->refcnt == 0)
1779 kfree(pc);
1780 cpuc->shared_regs = NULL;
1781 }
1782
1783 fini_debug_store_on_cpu(cpu);
1784 }
1785
intel_pmu_flush_branch_stack(void)1786 static void intel_pmu_flush_branch_stack(void)
1787 {
1788 /*
1789 * Intel LBR does not tag entries with the
1790 * PID of the current task, then we need to
1791 * flush it on ctxsw
1792 * For now, we simply reset it
1793 */
1794 if (x86_pmu.lbr_nr)
1795 intel_pmu_lbr_reset();
1796 }
1797
1798 PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63");
1799
1800 PMU_FORMAT_ATTR(ldlat, "config1:0-15");
1801
1802 static struct attribute *intel_arch3_formats_attr[] = {
1803 &format_attr_event.attr,
1804 &format_attr_umask.attr,
1805 &format_attr_edge.attr,
1806 &format_attr_pc.attr,
1807 &format_attr_any.attr,
1808 &format_attr_inv.attr,
1809 &format_attr_cmask.attr,
1810
1811 &format_attr_offcore_rsp.attr, /* XXX do NHM/WSM + SNB breakout */
1812 &format_attr_ldlat.attr, /* PEBS load latency */
1813 NULL,
1814 };
1815
1816 static __initconst const struct x86_pmu intel_pmu = {
1817 .name = "Intel",
1818 .handle_irq = intel_pmu_handle_irq,
1819 .disable_all = intel_pmu_disable_all,
1820 .enable_all = intel_pmu_enable_all,
1821 .enable = intel_pmu_enable_event,
1822 .disable = intel_pmu_disable_event,
1823 .hw_config = intel_pmu_hw_config,
1824 .schedule_events = x86_schedule_events,
1825 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
1826 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
1827 .event_map = intel_pmu_event_map,
1828 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
1829 .apic = 1,
1830 /*
1831 * Intel PMCs cannot be accessed sanely above 32 bit width,
1832 * so we install an artificial 1<<31 period regardless of
1833 * the generic event period:
1834 */
1835 .max_period = (1ULL << 31) - 1,
1836 .get_event_constraints = intel_get_event_constraints,
1837 .put_event_constraints = intel_put_event_constraints,
1838 .pebs_aliases = intel_pebs_aliases_core2,
1839
1840 .format_attrs = intel_arch3_formats_attr,
1841 .events_sysfs_show = intel_event_sysfs_show,
1842
1843 .cpu_prepare = intel_pmu_cpu_prepare,
1844 .cpu_starting = intel_pmu_cpu_starting,
1845 .cpu_dying = intel_pmu_cpu_dying,
1846 .guest_get_msrs = intel_guest_get_msrs,
1847 .flush_branch_stack = intel_pmu_flush_branch_stack,
1848 };
1849
intel_clovertown_quirk(void)1850 static __init void intel_clovertown_quirk(void)
1851 {
1852 /*
1853 * PEBS is unreliable due to:
1854 *
1855 * AJ67 - PEBS may experience CPL leaks
1856 * AJ68 - PEBS PMI may be delayed by one event
1857 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
1858 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
1859 *
1860 * AJ67 could be worked around by restricting the OS/USR flags.
1861 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
1862 *
1863 * AJ106 could possibly be worked around by not allowing LBR
1864 * usage from PEBS, including the fixup.
1865 * AJ68 could possibly be worked around by always programming
1866 * a pebs_event_reset[0] value and coping with the lost events.
1867 *
1868 * But taken together it might just make sense to not enable PEBS on
1869 * these chips.
1870 */
1871 pr_warn("PEBS disabled due to CPU errata\n");
1872 x86_pmu.pebs = 0;
1873 x86_pmu.pebs_constraints = NULL;
1874 }
1875
intel_snb_pebs_broken(int cpu)1876 static int intel_snb_pebs_broken(int cpu)
1877 {
1878 u32 rev = UINT_MAX; /* default to broken for unknown models */
1879
1880 switch (cpu_data(cpu).x86_model) {
1881 case 42: /* SNB */
1882 rev = 0x28;
1883 break;
1884
1885 case 45: /* SNB-EP */
1886 switch (cpu_data(cpu).x86_mask) {
1887 case 6: rev = 0x618; break;
1888 case 7: rev = 0x70c; break;
1889 }
1890 }
1891
1892 return (cpu_data(cpu).microcode < rev);
1893 }
1894
intel_snb_check_microcode(void)1895 static void intel_snb_check_microcode(void)
1896 {
1897 int pebs_broken = 0;
1898 int cpu;
1899
1900 get_online_cpus();
1901 for_each_online_cpu(cpu) {
1902 if ((pebs_broken = intel_snb_pebs_broken(cpu)))
1903 break;
1904 }
1905 put_online_cpus();
1906
1907 if (pebs_broken == x86_pmu.pebs_broken)
1908 return;
1909
1910 /*
1911 * Serialized by the microcode lock..
1912 */
1913 if (x86_pmu.pebs_broken) {
1914 pr_info("PEBS enabled due to microcode update\n");
1915 x86_pmu.pebs_broken = 0;
1916 } else {
1917 pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n");
1918 x86_pmu.pebs_broken = 1;
1919 }
1920 }
1921
intel_sandybridge_quirk(void)1922 static __init void intel_sandybridge_quirk(void)
1923 {
1924 x86_pmu.check_microcode = intel_snb_check_microcode;
1925 intel_snb_check_microcode();
1926 }
1927
1928 static const struct { int id; char *name; } intel_arch_events_map[] __initconst = {
1929 { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" },
1930 { PERF_COUNT_HW_INSTRUCTIONS, "instructions" },
1931 { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" },
1932 { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" },
1933 { PERF_COUNT_HW_CACHE_MISSES, "cache misses" },
1934 { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" },
1935 { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" },
1936 };
1937
intel_arch_events_quirk(void)1938 static __init void intel_arch_events_quirk(void)
1939 {
1940 int bit;
1941
1942 /* disable event that reported as not presend by cpuid */
1943 for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) {
1944 intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0;
1945 pr_warn("CPUID marked event: \'%s\' unavailable\n",
1946 intel_arch_events_map[bit].name);
1947 }
1948 }
1949
intel_nehalem_quirk(void)1950 static __init void intel_nehalem_quirk(void)
1951 {
1952 union cpuid10_ebx ebx;
1953
1954 ebx.full = x86_pmu.events_maskl;
1955 if (ebx.split.no_branch_misses_retired) {
1956 /*
1957 * Erratum AAJ80 detected, we work it around by using
1958 * the BR_MISP_EXEC.ANY event. This will over-count
1959 * branch-misses, but it's still much better than the
1960 * architectural event which is often completely bogus:
1961 */
1962 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89;
1963 ebx.split.no_branch_misses_retired = 0;
1964 x86_pmu.events_maskl = ebx.full;
1965 pr_info("CPU erratum AAJ80 worked around\n");
1966 }
1967 }
1968
intel_pmu_init(void)1969 __init int intel_pmu_init(void)
1970 {
1971 union cpuid10_edx edx;
1972 union cpuid10_eax eax;
1973 union cpuid10_ebx ebx;
1974 struct event_constraint *c;
1975 unsigned int unused;
1976 int version;
1977
1978 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
1979 switch (boot_cpu_data.x86) {
1980 case 0x6:
1981 return p6_pmu_init();
1982 case 0xb:
1983 return knc_pmu_init();
1984 case 0xf:
1985 return p4_pmu_init();
1986 }
1987 return -ENODEV;
1988 }
1989
1990 /*
1991 * Check whether the Architectural PerfMon supports
1992 * Branch Misses Retired hw_event or not.
1993 */
1994 cpuid(10, &eax.full, &ebx.full, &unused, &edx.full);
1995 if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT)
1996 return -ENODEV;
1997
1998 version = eax.split.version_id;
1999 if (version < 2)
2000 x86_pmu = core_pmu;
2001 else
2002 x86_pmu = intel_pmu;
2003
2004 x86_pmu.version = version;
2005 x86_pmu.num_counters = eax.split.num_counters;
2006 x86_pmu.cntval_bits = eax.split.bit_width;
2007 x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1;
2008
2009 x86_pmu.events_maskl = ebx.full;
2010 x86_pmu.events_mask_len = eax.split.mask_length;
2011
2012 x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters);
2013
2014 /*
2015 * Quirk: v2 perfmon does not report fixed-purpose events, so
2016 * assume at least 3 events:
2017 */
2018 if (version > 1)
2019 x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3);
2020
2021 if (boot_cpu_has(X86_FEATURE_PDCM)) {
2022 u64 capabilities;
2023
2024 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
2025 x86_pmu.intel_cap.capabilities = capabilities;
2026 }
2027
2028 intel_ds_init();
2029
2030 x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */
2031
2032 /*
2033 * Install the hw-cache-events table:
2034 */
2035 switch (boot_cpu_data.x86_model) {
2036 case 14: /* 65 nm core solo/duo, "Yonah" */
2037 pr_cont("Core events, ");
2038 break;
2039
2040 case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
2041 x86_add_quirk(intel_clovertown_quirk);
2042 case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */
2043 case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */
2044 case 29: /* six-core 45 nm xeon "Dunnington" */
2045 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
2046 sizeof(hw_cache_event_ids));
2047
2048 intel_pmu_lbr_init_core();
2049
2050 x86_pmu.event_constraints = intel_core2_event_constraints;
2051 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints;
2052 pr_cont("Core2 events, ");
2053 break;
2054
2055 case 26: /* 45 nm nehalem, "Bloomfield" */
2056 case 30: /* 45 nm nehalem, "Lynnfield" */
2057 case 46: /* 45 nm nehalem-ex, "Beckton" */
2058 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
2059 sizeof(hw_cache_event_ids));
2060 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
2061 sizeof(hw_cache_extra_regs));
2062
2063 intel_pmu_lbr_init_nhm();
2064
2065 x86_pmu.event_constraints = intel_nehalem_event_constraints;
2066 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints;
2067 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
2068 x86_pmu.extra_regs = intel_nehalem_extra_regs;
2069
2070 x86_pmu.cpu_events = nhm_events_attrs;
2071
2072 /* UOPS_ISSUED.STALLED_CYCLES */
2073 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
2074 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
2075 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
2076 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
2077 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
2078
2079 x86_add_quirk(intel_nehalem_quirk);
2080
2081 pr_cont("Nehalem events, ");
2082 break;
2083
2084 case 28: /* Atom */
2085 case 38: /* Lincroft */
2086 case 39: /* Penwell */
2087 case 53: /* Cloverview */
2088 case 54: /* Cedarview */
2089 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
2090 sizeof(hw_cache_event_ids));
2091
2092 intel_pmu_lbr_init_atom();
2093
2094 x86_pmu.event_constraints = intel_gen_event_constraints;
2095 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints;
2096 pr_cont("Atom events, ");
2097 break;
2098
2099 case 37: /* 32 nm nehalem, "Clarkdale" */
2100 case 44: /* 32 nm nehalem, "Gulftown" */
2101 case 47: /* 32 nm Xeon E7 */
2102 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
2103 sizeof(hw_cache_event_ids));
2104 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
2105 sizeof(hw_cache_extra_regs));
2106
2107 intel_pmu_lbr_init_nhm();
2108
2109 x86_pmu.event_constraints = intel_westmere_event_constraints;
2110 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
2111 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints;
2112 x86_pmu.extra_regs = intel_westmere_extra_regs;
2113 x86_pmu.er_flags |= ERF_HAS_RSP_1;
2114
2115 x86_pmu.cpu_events = nhm_events_attrs;
2116
2117 /* UOPS_ISSUED.STALLED_CYCLES */
2118 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
2119 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
2120 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
2121 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
2122 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
2123
2124 pr_cont("Westmere events, ");
2125 break;
2126
2127 case 42: /* SandyBridge */
2128 case 45: /* SandyBridge, "Romely-EP" */
2129 x86_add_quirk(intel_sandybridge_quirk);
2130 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
2131 sizeof(hw_cache_event_ids));
2132 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
2133 sizeof(hw_cache_extra_regs));
2134
2135 intel_pmu_lbr_init_snb();
2136
2137 x86_pmu.event_constraints = intel_snb_event_constraints;
2138 x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints;
2139 x86_pmu.pebs_aliases = intel_pebs_aliases_snb;
2140 if (boot_cpu_data.x86_model == 45)
2141 x86_pmu.extra_regs = intel_snbep_extra_regs;
2142 else
2143 x86_pmu.extra_regs = intel_snb_extra_regs;
2144 /* all extra regs are per-cpu when HT is on */
2145 x86_pmu.er_flags |= ERF_HAS_RSP_1;
2146 x86_pmu.er_flags |= ERF_NO_HT_SHARING;
2147
2148 x86_pmu.cpu_events = snb_events_attrs;
2149
2150 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
2151 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
2152 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
2153 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
2154 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
2155 X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1);
2156
2157 pr_cont("SandyBridge events, ");
2158 break;
2159 case 58: /* IvyBridge */
2160 case 62: /* IvyBridge EP */
2161 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
2162 sizeof(hw_cache_event_ids));
2163 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
2164 sizeof(hw_cache_extra_regs));
2165
2166 intel_pmu_lbr_init_snb();
2167
2168 x86_pmu.event_constraints = intel_ivb_event_constraints;
2169 x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints;
2170 x86_pmu.pebs_aliases = intel_pebs_aliases_snb;
2171 if (boot_cpu_data.x86_model == 62)
2172 x86_pmu.extra_regs = intel_snbep_extra_regs;
2173 else
2174 x86_pmu.extra_regs = intel_snb_extra_regs;
2175 /* all extra regs are per-cpu when HT is on */
2176 x86_pmu.er_flags |= ERF_HAS_RSP_1;
2177 x86_pmu.er_flags |= ERF_NO_HT_SHARING;
2178
2179 x86_pmu.cpu_events = snb_events_attrs;
2180
2181 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
2182 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
2183 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
2184
2185 pr_cont("IvyBridge events, ");
2186 break;
2187
2188
2189 default:
2190 switch (x86_pmu.version) {
2191 case 1:
2192 x86_pmu.event_constraints = intel_v1_event_constraints;
2193 pr_cont("generic architected perfmon v1, ");
2194 break;
2195 default:
2196 /*
2197 * default constraints for v2 and up
2198 */
2199 x86_pmu.event_constraints = intel_gen_event_constraints;
2200 pr_cont("generic architected perfmon, ");
2201 break;
2202 }
2203 }
2204
2205 if (x86_pmu.num_counters > INTEL_PMC_MAX_GENERIC) {
2206 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
2207 x86_pmu.num_counters, INTEL_PMC_MAX_GENERIC);
2208 x86_pmu.num_counters = INTEL_PMC_MAX_GENERIC;
2209 }
2210 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
2211
2212 if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) {
2213 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
2214 x86_pmu.num_counters_fixed, INTEL_PMC_MAX_FIXED);
2215 x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED;
2216 }
2217
2218 x86_pmu.intel_ctrl |=
2219 ((1LL << x86_pmu.num_counters_fixed)-1) << INTEL_PMC_IDX_FIXED;
2220
2221 if (x86_pmu.event_constraints) {
2222 /*
2223 * event on fixed counter2 (REF_CYCLES) only works on this
2224 * counter, so do not extend mask to generic counters
2225 */
2226 for_each_event_constraint(c, x86_pmu.event_constraints) {
2227 if (c->cmask != X86_RAW_EVENT_MASK
2228 || c->idxmsk64 == INTEL_PMC_MSK_FIXED_REF_CYCLES) {
2229 continue;
2230 }
2231
2232 c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1;
2233 c->weight += x86_pmu.num_counters;
2234 }
2235 }
2236
2237 return 0;
2238 }
2239