4 * Used to coordinate shared registers between HT threads or
5 * among events on a single PMU.
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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>
16 #include <asm/cpufeature.h>
17 #include <asm/hardirq.h>
20 #include "perf_event.h"
23 * Intel PerfMon, used on Core and later.
25 static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly =
27 [PERF_COUNT_HW_CPU_CYCLES] = 0x003c,
28 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
29 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e,
30 [PERF_COUNT_HW_CACHE_MISSES] = 0x412e,
31 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4,
32 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5,
33 [PERF_COUNT_HW_BUS_CYCLES] = 0x013c,
34 [PERF_COUNT_HW_REF_CPU_CYCLES] = 0x0300, /* pseudo-encoding */
37 static struct event_constraint intel_core_event_constraints[] __read_mostly =
39 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
40 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
41 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
42 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
43 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
44 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */
48 static struct event_constraint intel_core2_event_constraints[] __read_mostly =
50 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
51 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
52 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
53 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
54 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
55 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
56 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
57 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
58 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
59 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
60 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
61 INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */
62 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
66 static struct event_constraint intel_nehalem_event_constraints[] __read_mostly =
68 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
69 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
70 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
71 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
72 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
73 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
74 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
75 INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */
76 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
77 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
78 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
82 static struct extra_reg intel_nehalem_extra_regs[] __read_mostly =
84 INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
85 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
89 static struct event_constraint intel_westmere_event_constraints[] __read_mostly =
91 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
92 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
93 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
94 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
95 INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */
96 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
97 INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */
101 static struct event_constraint intel_snb_event_constraints[] __read_mostly =
103 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
104 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
105 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
106 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
107 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
108 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
109 INTEL_UEVENT_CONSTRAINT(0x06a3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
110 INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */
111 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
112 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
113 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
114 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
118 static struct event_constraint intel_ivb_event_constraints[] __read_mostly =
120 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
121 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
122 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
123 INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */
124 INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMTPY */
125 INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */
126 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
127 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
128 INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */
129 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
130 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
131 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
133 * Errata BV98 -- MEM_*_RETIRED events can leak between counters of SMT
134 * siblings; disable these events because they can corrupt unrelated
137 INTEL_EVENT_CONSTRAINT(0xd0, 0x0), /* MEM_UOPS_RETIRED.* */
138 INTEL_EVENT_CONSTRAINT(0xd1, 0x0), /* MEM_LOAD_UOPS_RETIRED.* */
139 INTEL_EVENT_CONSTRAINT(0xd2, 0x0), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
140 INTEL_EVENT_CONSTRAINT(0xd3, 0x0), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
144 static struct extra_reg intel_westmere_extra_regs[] __read_mostly =
146 INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
147 INTEL_EVENT_EXTRA_REG(0xbb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1),
148 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
152 static struct event_constraint intel_v1_event_constraints[] __read_mostly =
157 static struct event_constraint intel_gen_event_constraints[] __read_mostly =
159 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
160 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
161 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
165 static struct extra_reg intel_snb_extra_regs[] __read_mostly = {
166 INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0),
167 INTEL_EVENT_EXTRA_REG(0xbb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1),
168 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
172 static struct extra_reg intel_snbep_extra_regs[] __read_mostly = {
173 INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
174 INTEL_EVENT_EXTRA_REG(0xbb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
175 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
179 EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3");
180 EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3");
181 EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2");
183 struct attribute *nhm_events_attrs[] = {
184 EVENT_PTR(mem_ld_nhm),
188 struct attribute *snb_events_attrs[] = {
189 EVENT_PTR(mem_ld_snb),
190 EVENT_PTR(mem_st_snb),
194 static struct event_constraint intel_hsw_event_constraints[] = {
195 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
196 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
197 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
198 INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.* */
199 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
200 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
201 /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
202 INTEL_EVENT_CONSTRAINT(0x08a3, 0x4),
203 /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
204 INTEL_EVENT_CONSTRAINT(0x0ca3, 0x4),
205 /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
206 INTEL_EVENT_CONSTRAINT(0x04a3, 0xf),
210 static u64 intel_pmu_event_map(int hw_event)
212 return intel_perfmon_event_map[hw_event];
215 #define SNB_DMND_DATA_RD (1ULL << 0)
216 #define SNB_DMND_RFO (1ULL << 1)
217 #define SNB_DMND_IFETCH (1ULL << 2)
218 #define SNB_DMND_WB (1ULL << 3)
219 #define SNB_PF_DATA_RD (1ULL << 4)
220 #define SNB_PF_RFO (1ULL << 5)
221 #define SNB_PF_IFETCH (1ULL << 6)
222 #define SNB_LLC_DATA_RD (1ULL << 7)
223 #define SNB_LLC_RFO (1ULL << 8)
224 #define SNB_LLC_IFETCH (1ULL << 9)
225 #define SNB_BUS_LOCKS (1ULL << 10)
226 #define SNB_STRM_ST (1ULL << 11)
227 #define SNB_OTHER (1ULL << 15)
228 #define SNB_RESP_ANY (1ULL << 16)
229 #define SNB_NO_SUPP (1ULL << 17)
230 #define SNB_LLC_HITM (1ULL << 18)
231 #define SNB_LLC_HITE (1ULL << 19)
232 #define SNB_LLC_HITS (1ULL << 20)
233 #define SNB_LLC_HITF (1ULL << 21)
234 #define SNB_LOCAL (1ULL << 22)
235 #define SNB_REMOTE (0xffULL << 23)
236 #define SNB_SNP_NONE (1ULL << 31)
237 #define SNB_SNP_NOT_NEEDED (1ULL << 32)
238 #define SNB_SNP_MISS (1ULL << 33)
239 #define SNB_NO_FWD (1ULL << 34)
240 #define SNB_SNP_FWD (1ULL << 35)
241 #define SNB_HITM (1ULL << 36)
242 #define SNB_NON_DRAM (1ULL << 37)
244 #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD)
245 #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO)
246 #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
248 #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \
249 SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \
252 #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY)
253 #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY)
255 #define SNB_L3_ACCESS SNB_RESP_ANY
256 #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM)
258 static __initconst const u64 snb_hw_cache_extra_regs
259 [PERF_COUNT_HW_CACHE_MAX]
260 [PERF_COUNT_HW_CACHE_OP_MAX]
261 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
265 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS,
266 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_L3_MISS,
269 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS,
270 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_L3_MISS,
272 [ C(OP_PREFETCH) ] = {
273 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS,
274 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_L3_MISS,
279 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY,
280 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_DRAM_REMOTE,
283 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY,
284 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE,
286 [ C(OP_PREFETCH) ] = {
287 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY,
288 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE,
293 static __initconst const u64 snb_hw_cache_event_ids
294 [PERF_COUNT_HW_CACHE_MAX]
295 [PERF_COUNT_HW_CACHE_OP_MAX]
296 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
300 [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */
301 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */
304 [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */
305 [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */
307 [ C(OP_PREFETCH) ] = {
308 [ C(RESULT_ACCESS) ] = 0x0,
309 [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */
314 [ C(RESULT_ACCESS) ] = 0x0,
315 [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */
318 [ C(RESULT_ACCESS) ] = -1,
319 [ C(RESULT_MISS) ] = -1,
321 [ C(OP_PREFETCH) ] = {
322 [ C(RESULT_ACCESS) ] = 0x0,
323 [ C(RESULT_MISS) ] = 0x0,
328 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
329 [ C(RESULT_ACCESS) ] = 0x01b7,
330 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
331 [ C(RESULT_MISS) ] = 0x01b7,
334 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
335 [ C(RESULT_ACCESS) ] = 0x01b7,
336 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
337 [ C(RESULT_MISS) ] = 0x01b7,
339 [ C(OP_PREFETCH) ] = {
340 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
341 [ C(RESULT_ACCESS) ] = 0x01b7,
342 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
343 [ C(RESULT_MISS) ] = 0x01b7,
348 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */
349 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */
352 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */
353 [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
355 [ C(OP_PREFETCH) ] = {
356 [ C(RESULT_ACCESS) ] = 0x0,
357 [ C(RESULT_MISS) ] = 0x0,
362 [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */
363 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */
366 [ C(RESULT_ACCESS) ] = -1,
367 [ C(RESULT_MISS) ] = -1,
369 [ C(OP_PREFETCH) ] = {
370 [ C(RESULT_ACCESS) ] = -1,
371 [ C(RESULT_MISS) ] = -1,
376 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
377 [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
380 [ C(RESULT_ACCESS) ] = -1,
381 [ C(RESULT_MISS) ] = -1,
383 [ C(OP_PREFETCH) ] = {
384 [ C(RESULT_ACCESS) ] = -1,
385 [ C(RESULT_MISS) ] = -1,
390 [ C(RESULT_ACCESS) ] = 0x01b7,
391 [ C(RESULT_MISS) ] = 0x01b7,
394 [ C(RESULT_ACCESS) ] = 0x01b7,
395 [ C(RESULT_MISS) ] = 0x01b7,
397 [ C(OP_PREFETCH) ] = {
398 [ C(RESULT_ACCESS) ] = 0x01b7,
399 [ C(RESULT_MISS) ] = 0x01b7,
405 static __initconst const u64 westmere_hw_cache_event_ids
406 [PERF_COUNT_HW_CACHE_MAX]
407 [PERF_COUNT_HW_CACHE_OP_MAX]
408 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
412 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
413 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
416 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
417 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
419 [ C(OP_PREFETCH) ] = {
420 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
421 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
426 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
427 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
430 [ C(RESULT_ACCESS) ] = -1,
431 [ C(RESULT_MISS) ] = -1,
433 [ C(OP_PREFETCH) ] = {
434 [ C(RESULT_ACCESS) ] = 0x0,
435 [ C(RESULT_MISS) ] = 0x0,
440 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
441 [ C(RESULT_ACCESS) ] = 0x01b7,
442 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
443 [ C(RESULT_MISS) ] = 0x01b7,
446 * Use RFO, not WRITEBACK, because a write miss would typically occur
450 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
451 [ C(RESULT_ACCESS) ] = 0x01b7,
452 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
453 [ C(RESULT_MISS) ] = 0x01b7,
455 [ C(OP_PREFETCH) ] = {
456 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
457 [ C(RESULT_ACCESS) ] = 0x01b7,
458 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
459 [ C(RESULT_MISS) ] = 0x01b7,
464 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
465 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
468 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
469 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
471 [ C(OP_PREFETCH) ] = {
472 [ C(RESULT_ACCESS) ] = 0x0,
473 [ C(RESULT_MISS) ] = 0x0,
478 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
479 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */
482 [ C(RESULT_ACCESS) ] = -1,
483 [ C(RESULT_MISS) ] = -1,
485 [ C(OP_PREFETCH) ] = {
486 [ C(RESULT_ACCESS) ] = -1,
487 [ C(RESULT_MISS) ] = -1,
492 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
493 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
496 [ C(RESULT_ACCESS) ] = -1,
497 [ C(RESULT_MISS) ] = -1,
499 [ C(OP_PREFETCH) ] = {
500 [ C(RESULT_ACCESS) ] = -1,
501 [ C(RESULT_MISS) ] = -1,
506 [ C(RESULT_ACCESS) ] = 0x01b7,
507 [ C(RESULT_MISS) ] = 0x01b7,
510 [ C(RESULT_ACCESS) ] = 0x01b7,
511 [ C(RESULT_MISS) ] = 0x01b7,
513 [ C(OP_PREFETCH) ] = {
514 [ C(RESULT_ACCESS) ] = 0x01b7,
515 [ C(RESULT_MISS) ] = 0x01b7,
521 * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits;
522 * See IA32 SDM Vol 3B 30.6.1.3
525 #define NHM_DMND_DATA_RD (1 << 0)
526 #define NHM_DMND_RFO (1 << 1)
527 #define NHM_DMND_IFETCH (1 << 2)
528 #define NHM_DMND_WB (1 << 3)
529 #define NHM_PF_DATA_RD (1 << 4)
530 #define NHM_PF_DATA_RFO (1 << 5)
531 #define NHM_PF_IFETCH (1 << 6)
532 #define NHM_OFFCORE_OTHER (1 << 7)
533 #define NHM_UNCORE_HIT (1 << 8)
534 #define NHM_OTHER_CORE_HIT_SNP (1 << 9)
535 #define NHM_OTHER_CORE_HITM (1 << 10)
537 #define NHM_REMOTE_CACHE_FWD (1 << 12)
538 #define NHM_REMOTE_DRAM (1 << 13)
539 #define NHM_LOCAL_DRAM (1 << 14)
540 #define NHM_NON_DRAM (1 << 15)
542 #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD)
543 #define NHM_REMOTE (NHM_REMOTE_DRAM)
545 #define NHM_DMND_READ (NHM_DMND_DATA_RD)
546 #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB)
547 #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO)
549 #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM)
550 #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD)
551 #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS)
553 static __initconst const u64 nehalem_hw_cache_extra_regs
554 [PERF_COUNT_HW_CACHE_MAX]
555 [PERF_COUNT_HW_CACHE_OP_MAX]
556 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
560 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS,
561 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS,
564 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS,
565 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS,
567 [ C(OP_PREFETCH) ] = {
568 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS,
569 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS,
574 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE,
575 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE,
578 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE,
579 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE,
581 [ C(OP_PREFETCH) ] = {
582 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE,
583 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE,
588 static __initconst const u64 nehalem_hw_cache_event_ids
589 [PERF_COUNT_HW_CACHE_MAX]
590 [PERF_COUNT_HW_CACHE_OP_MAX]
591 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
595 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
596 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
599 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
600 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
602 [ C(OP_PREFETCH) ] = {
603 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
604 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
609 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
610 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
613 [ C(RESULT_ACCESS) ] = -1,
614 [ C(RESULT_MISS) ] = -1,
616 [ C(OP_PREFETCH) ] = {
617 [ C(RESULT_ACCESS) ] = 0x0,
618 [ C(RESULT_MISS) ] = 0x0,
623 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
624 [ C(RESULT_ACCESS) ] = 0x01b7,
625 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
626 [ C(RESULT_MISS) ] = 0x01b7,
629 * Use RFO, not WRITEBACK, because a write miss would typically occur
633 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
634 [ C(RESULT_ACCESS) ] = 0x01b7,
635 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
636 [ C(RESULT_MISS) ] = 0x01b7,
638 [ C(OP_PREFETCH) ] = {
639 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
640 [ C(RESULT_ACCESS) ] = 0x01b7,
641 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
642 [ C(RESULT_MISS) ] = 0x01b7,
647 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
648 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
651 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
652 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
654 [ C(OP_PREFETCH) ] = {
655 [ C(RESULT_ACCESS) ] = 0x0,
656 [ C(RESULT_MISS) ] = 0x0,
661 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
662 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */
665 [ C(RESULT_ACCESS) ] = -1,
666 [ C(RESULT_MISS) ] = -1,
668 [ C(OP_PREFETCH) ] = {
669 [ C(RESULT_ACCESS) ] = -1,
670 [ C(RESULT_MISS) ] = -1,
675 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
676 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
679 [ C(RESULT_ACCESS) ] = -1,
680 [ C(RESULT_MISS) ] = -1,
682 [ C(OP_PREFETCH) ] = {
683 [ C(RESULT_ACCESS) ] = -1,
684 [ C(RESULT_MISS) ] = -1,
689 [ C(RESULT_ACCESS) ] = 0x01b7,
690 [ C(RESULT_MISS) ] = 0x01b7,
693 [ C(RESULT_ACCESS) ] = 0x01b7,
694 [ C(RESULT_MISS) ] = 0x01b7,
696 [ C(OP_PREFETCH) ] = {
697 [ C(RESULT_ACCESS) ] = 0x01b7,
698 [ C(RESULT_MISS) ] = 0x01b7,
703 static __initconst const u64 core2_hw_cache_event_ids
704 [PERF_COUNT_HW_CACHE_MAX]
705 [PERF_COUNT_HW_CACHE_OP_MAX]
706 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
710 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
711 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
714 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
715 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
717 [ C(OP_PREFETCH) ] = {
718 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */
719 [ C(RESULT_MISS) ] = 0,
724 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */
725 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */
728 [ C(RESULT_ACCESS) ] = -1,
729 [ C(RESULT_MISS) ] = -1,
731 [ C(OP_PREFETCH) ] = {
732 [ C(RESULT_ACCESS) ] = 0,
733 [ C(RESULT_MISS) ] = 0,
738 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
739 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
742 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
743 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
745 [ C(OP_PREFETCH) ] = {
746 [ C(RESULT_ACCESS) ] = 0,
747 [ C(RESULT_MISS) ] = 0,
752 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
753 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */
756 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
757 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */
759 [ C(OP_PREFETCH) ] = {
760 [ C(RESULT_ACCESS) ] = 0,
761 [ C(RESULT_MISS) ] = 0,
766 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
767 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */
770 [ C(RESULT_ACCESS) ] = -1,
771 [ C(RESULT_MISS) ] = -1,
773 [ C(OP_PREFETCH) ] = {
774 [ C(RESULT_ACCESS) ] = -1,
775 [ C(RESULT_MISS) ] = -1,
780 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
781 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
784 [ C(RESULT_ACCESS) ] = -1,
785 [ C(RESULT_MISS) ] = -1,
787 [ C(OP_PREFETCH) ] = {
788 [ C(RESULT_ACCESS) ] = -1,
789 [ C(RESULT_MISS) ] = -1,
794 static __initconst const u64 atom_hw_cache_event_ids
795 [PERF_COUNT_HW_CACHE_MAX]
796 [PERF_COUNT_HW_CACHE_OP_MAX]
797 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
801 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */
802 [ C(RESULT_MISS) ] = 0,
805 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */
806 [ C(RESULT_MISS) ] = 0,
808 [ C(OP_PREFETCH) ] = {
809 [ C(RESULT_ACCESS) ] = 0x0,
810 [ C(RESULT_MISS) ] = 0,
815 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
816 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
819 [ C(RESULT_ACCESS) ] = -1,
820 [ C(RESULT_MISS) ] = -1,
822 [ C(OP_PREFETCH) ] = {
823 [ C(RESULT_ACCESS) ] = 0,
824 [ C(RESULT_MISS) ] = 0,
829 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
830 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
833 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
834 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
836 [ C(OP_PREFETCH) ] = {
837 [ C(RESULT_ACCESS) ] = 0,
838 [ C(RESULT_MISS) ] = 0,
843 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */
844 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */
847 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */
848 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */
850 [ C(OP_PREFETCH) ] = {
851 [ C(RESULT_ACCESS) ] = 0,
852 [ C(RESULT_MISS) ] = 0,
857 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
858 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */
861 [ C(RESULT_ACCESS) ] = -1,
862 [ C(RESULT_MISS) ] = -1,
864 [ C(OP_PREFETCH) ] = {
865 [ C(RESULT_ACCESS) ] = -1,
866 [ C(RESULT_MISS) ] = -1,
871 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
872 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
875 [ C(RESULT_ACCESS) ] = -1,
876 [ C(RESULT_MISS) ] = -1,
878 [ C(OP_PREFETCH) ] = {
879 [ C(RESULT_ACCESS) ] = -1,
880 [ C(RESULT_MISS) ] = -1,
885 static inline bool intel_pmu_needs_lbr_smpl(struct perf_event *event)
887 /* user explicitly requested branch sampling */
888 if (has_branch_stack(event))
891 /* implicit branch sampling to correct PEBS skid */
892 if (x86_pmu.intel_cap.pebs_trap && event->attr.precise_ip > 1 &&
893 x86_pmu.intel_cap.pebs_format < 2)
899 static void intel_pmu_disable_all(void)
901 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
903 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
905 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask))
906 intel_pmu_disable_bts();
908 intel_pmu_pebs_disable_all();
909 intel_pmu_lbr_disable_all();
912 static void intel_pmu_enable_all(int added)
914 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
916 intel_pmu_pebs_enable_all();
917 intel_pmu_lbr_enable_all();
918 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL,
919 x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask);
921 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
922 struct perf_event *event =
923 cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
925 if (WARN_ON_ONCE(!event))
928 intel_pmu_enable_bts(event->hw.config);
934 * Intel Errata AAK100 (model 26)
935 * Intel Errata AAP53 (model 30)
936 * Intel Errata BD53 (model 44)
938 * The official story:
939 * These chips need to be 'reset' when adding counters by programming the
940 * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either
941 * in sequence on the same PMC or on different PMCs.
943 * In practise it appears some of these events do in fact count, and
944 * we need to programm all 4 events.
946 static void intel_pmu_nhm_workaround(void)
948 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
949 static const unsigned long nhm_magic[4] = {
955 struct perf_event *event;
959 * The Errata requires below steps:
960 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL;
961 * 2) Configure 4 PERFEVTSELx with the magic events and clear
962 * the corresponding PMCx;
963 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL;
964 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL;
965 * 5) Clear 4 pairs of ERFEVTSELx and PMCx;
969 * The real steps we choose are a little different from above.
970 * A) To reduce MSR operations, we don't run step 1) as they
971 * are already cleared before this function is called;
972 * B) Call x86_perf_event_update to save PMCx before configuring
973 * PERFEVTSELx with magic number;
974 * C) With step 5), we do clear only when the PERFEVTSELx is
975 * not used currently.
976 * D) Call x86_perf_event_set_period to restore PMCx;
979 /* We always operate 4 pairs of PERF Counters */
980 for (i = 0; i < 4; i++) {
981 event = cpuc->events[i];
983 x86_perf_event_update(event);
986 for (i = 0; i < 4; i++) {
987 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]);
988 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0);
991 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf);
992 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
994 for (i = 0; i < 4; i++) {
995 event = cpuc->events[i];
998 x86_perf_event_set_period(event);
999 __x86_pmu_enable_event(&event->hw,
1000 ARCH_PERFMON_EVENTSEL_ENABLE);
1002 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0);
1006 static void intel_pmu_nhm_enable_all(int added)
1009 intel_pmu_nhm_workaround();
1010 intel_pmu_enable_all(added);
1013 static inline u64 intel_pmu_get_status(void)
1017 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1022 static inline void intel_pmu_ack_status(u64 ack)
1024 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
1027 static void intel_pmu_disable_fixed(struct hw_perf_event *hwc)
1029 int idx = hwc->idx - INTEL_PMC_IDX_FIXED;
1032 mask = 0xfULL << (idx * 4);
1034 rdmsrl(hwc->config_base, ctrl_val);
1036 wrmsrl(hwc->config_base, ctrl_val);
1039 static void intel_pmu_disable_event(struct perf_event *event)
1041 struct hw_perf_event *hwc = &event->hw;
1042 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1044 if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) {
1045 intel_pmu_disable_bts();
1046 intel_pmu_drain_bts_buffer();
1050 cpuc->intel_ctrl_guest_mask &= ~(1ull << hwc->idx);
1051 cpuc->intel_ctrl_host_mask &= ~(1ull << hwc->idx);
1054 * must disable before any actual event
1055 * because any event may be combined with LBR
1057 if (intel_pmu_needs_lbr_smpl(event))
1058 intel_pmu_lbr_disable(event);
1060 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
1061 intel_pmu_disable_fixed(hwc);
1065 x86_pmu_disable_event(event);
1067 if (unlikely(event->attr.precise_ip))
1068 intel_pmu_pebs_disable(event);
1071 static void intel_pmu_enable_fixed(struct hw_perf_event *hwc)
1073 int idx = hwc->idx - INTEL_PMC_IDX_FIXED;
1074 u64 ctrl_val, bits, mask;
1077 * Enable IRQ generation (0x8),
1078 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
1082 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
1084 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
1088 * ANY bit is supported in v3 and up
1090 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
1094 mask = 0xfULL << (idx * 4);
1096 rdmsrl(hwc->config_base, ctrl_val);
1099 wrmsrl(hwc->config_base, ctrl_val);
1102 static void intel_pmu_enable_event(struct perf_event *event)
1104 struct hw_perf_event *hwc = &event->hw;
1105 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1107 if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) {
1108 if (!__this_cpu_read(cpu_hw_events.enabled))
1111 intel_pmu_enable_bts(hwc->config);
1115 * must enabled before any actual event
1116 * because any event may be combined with LBR
1118 if (intel_pmu_needs_lbr_smpl(event))
1119 intel_pmu_lbr_enable(event);
1121 if (event->attr.exclude_host)
1122 cpuc->intel_ctrl_guest_mask |= (1ull << hwc->idx);
1123 if (event->attr.exclude_guest)
1124 cpuc->intel_ctrl_host_mask |= (1ull << hwc->idx);
1126 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
1127 intel_pmu_enable_fixed(hwc);
1131 if (unlikely(event->attr.precise_ip))
1132 intel_pmu_pebs_enable(event);
1134 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
1138 * Save and restart an expired event. Called by NMI contexts,
1139 * so it has to be careful about preempting normal event ops:
1141 int intel_pmu_save_and_restart(struct perf_event *event)
1143 x86_perf_event_update(event);
1144 return x86_perf_event_set_period(event);
1147 static void intel_pmu_reset(void)
1149 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
1150 unsigned long flags;
1153 if (!x86_pmu.num_counters)
1156 local_irq_save(flags);
1158 pr_info("clearing PMU state on CPU#%d\n", smp_processor_id());
1160 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1161 wrmsrl_safe(x86_pmu_config_addr(idx), 0ull);
1162 wrmsrl_safe(x86_pmu_event_addr(idx), 0ull);
1164 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++)
1165 wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
1168 ds->bts_index = ds->bts_buffer_base;
1170 local_irq_restore(flags);
1174 * This handler is triggered by the local APIC, so the APIC IRQ handling
1177 static int intel_pmu_handle_irq(struct pt_regs *regs)
1179 struct perf_sample_data data;
1180 struct cpu_hw_events *cpuc;
1185 cpuc = &__get_cpu_var(cpu_hw_events);
1188 * No known reason to not always do late ACK,
1189 * but just in case do it opt-in.
1191 if (!x86_pmu.late_ack)
1192 apic_write(APIC_LVTPC, APIC_DM_NMI);
1193 intel_pmu_disable_all();
1194 handled = intel_pmu_drain_bts_buffer();
1195 status = intel_pmu_get_status();
1197 intel_pmu_enable_all(0);
1203 intel_pmu_ack_status(status);
1204 if (++loops > 100) {
1205 static bool warned = false;
1207 WARN(1, "perfevents: irq loop stuck!\n");
1208 perf_event_print_debug();
1215 inc_irq_stat(apic_perf_irqs);
1217 intel_pmu_lbr_read();
1220 * PEBS overflow sets bit 62 in the global status register
1222 if (__test_and_clear_bit(62, (unsigned long *)&status)) {
1224 x86_pmu.drain_pebs(regs);
1227 for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
1228 struct perf_event *event = cpuc->events[bit];
1232 if (!test_bit(bit, cpuc->active_mask))
1235 if (!intel_pmu_save_and_restart(event))
1238 perf_sample_data_init(&data, 0, event->hw.last_period);
1240 if (has_branch_stack(event))
1241 data.br_stack = &cpuc->lbr_stack;
1243 if (perf_event_overflow(event, &data, regs))
1244 x86_pmu_stop(event, 0);
1248 * Repeat if there is more work to be done:
1250 status = intel_pmu_get_status();
1255 intel_pmu_enable_all(0);
1257 * Only unmask the NMI after the overflow counters
1258 * have been reset. This avoids spurious NMIs on
1261 if (x86_pmu.late_ack)
1262 apic_write(APIC_LVTPC, APIC_DM_NMI);
1266 static struct event_constraint *
1267 intel_bts_constraints(struct perf_event *event)
1269 struct hw_perf_event *hwc = &event->hw;
1270 unsigned int hw_event, bts_event;
1272 if (event->attr.freq)
1275 hw_event = hwc->config & INTEL_ARCH_EVENT_MASK;
1276 bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
1278 if (unlikely(hw_event == bts_event && hwc->sample_period == 1))
1279 return &bts_constraint;
1284 static int intel_alt_er(int idx)
1286 if (!(x86_pmu.er_flags & ERF_HAS_RSP_1))
1289 if (idx == EXTRA_REG_RSP_0)
1290 return EXTRA_REG_RSP_1;
1292 if (idx == EXTRA_REG_RSP_1)
1293 return EXTRA_REG_RSP_0;
1298 static void intel_fixup_er(struct perf_event *event, int idx)
1300 event->hw.extra_reg.idx = idx;
1302 if (idx == EXTRA_REG_RSP_0) {
1303 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
1304 event->hw.config |= 0x01b7;
1305 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0;
1306 } else if (idx == EXTRA_REG_RSP_1) {
1307 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
1308 event->hw.config |= 0x01bb;
1309 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1;
1314 * manage allocation of shared extra msr for certain events
1317 * per-cpu: to be shared between the various events on a single PMU
1318 * per-core: per-cpu + shared by HT threads
1320 static struct event_constraint *
1321 __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc,
1322 struct perf_event *event,
1323 struct hw_perf_event_extra *reg)
1325 struct event_constraint *c = &emptyconstraint;
1326 struct er_account *era;
1327 unsigned long flags;
1331 * reg->alloc can be set due to existing state, so for fake cpuc we
1332 * need to ignore this, otherwise we might fail to allocate proper fake
1333 * state for this extra reg constraint. Also see the comment below.
1335 if (reg->alloc && !cpuc->is_fake)
1336 return NULL; /* call x86_get_event_constraint() */
1339 era = &cpuc->shared_regs->regs[idx];
1341 * we use spin_lock_irqsave() to avoid lockdep issues when
1342 * passing a fake cpuc
1344 raw_spin_lock_irqsave(&era->lock, flags);
1346 if (!atomic_read(&era->ref) || era->config == reg->config) {
1349 * If its a fake cpuc -- as per validate_{group,event}() we
1350 * shouldn't touch event state and we can avoid doing so
1351 * since both will only call get_event_constraints() once
1352 * on each event, this avoids the need for reg->alloc.
1354 * Not doing the ER fixup will only result in era->reg being
1355 * wrong, but since we won't actually try and program hardware
1356 * this isn't a problem either.
1358 if (!cpuc->is_fake) {
1359 if (idx != reg->idx)
1360 intel_fixup_er(event, idx);
1363 * x86_schedule_events() can call get_event_constraints()
1364 * multiple times on events in the case of incremental
1365 * scheduling(). reg->alloc ensures we only do the ER
1371 /* lock in msr value */
1372 era->config = reg->config;
1373 era->reg = reg->reg;
1376 atomic_inc(&era->ref);
1379 * need to call x86_get_event_constraint()
1380 * to check if associated event has constraints
1384 idx = intel_alt_er(idx);
1385 if (idx != reg->idx) {
1386 raw_spin_unlock_irqrestore(&era->lock, flags);
1390 raw_spin_unlock_irqrestore(&era->lock, flags);
1396 __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc,
1397 struct hw_perf_event_extra *reg)
1399 struct er_account *era;
1402 * Only put constraint if extra reg was actually allocated. Also takes
1403 * care of event which do not use an extra shared reg.
1405 * Also, if this is a fake cpuc we shouldn't touch any event state
1406 * (reg->alloc) and we don't care about leaving inconsistent cpuc state
1407 * either since it'll be thrown out.
1409 if (!reg->alloc || cpuc->is_fake)
1412 era = &cpuc->shared_regs->regs[reg->idx];
1414 /* one fewer user */
1415 atomic_dec(&era->ref);
1417 /* allocate again next time */
1421 static struct event_constraint *
1422 intel_shared_regs_constraints(struct cpu_hw_events *cpuc,
1423 struct perf_event *event)
1425 struct event_constraint *c = NULL, *d;
1426 struct hw_perf_event_extra *xreg, *breg;
1428 xreg = &event->hw.extra_reg;
1429 if (xreg->idx != EXTRA_REG_NONE) {
1430 c = __intel_shared_reg_get_constraints(cpuc, event, xreg);
1431 if (c == &emptyconstraint)
1434 breg = &event->hw.branch_reg;
1435 if (breg->idx != EXTRA_REG_NONE) {
1436 d = __intel_shared_reg_get_constraints(cpuc, event, breg);
1437 if (d == &emptyconstraint) {
1438 __intel_shared_reg_put_constraints(cpuc, xreg);
1445 struct event_constraint *
1446 x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1448 struct event_constraint *c;
1450 if (x86_pmu.event_constraints) {
1451 for_each_event_constraint(c, x86_pmu.event_constraints) {
1452 if ((event->hw.config & c->cmask) == c->code) {
1453 /* hw.flags zeroed at initialization */
1454 event->hw.flags |= c->flags;
1460 return &unconstrained;
1463 static struct event_constraint *
1464 intel_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1466 struct event_constraint *c;
1468 c = intel_bts_constraints(event);
1472 c = intel_pebs_constraints(event);
1476 c = intel_shared_regs_constraints(cpuc, event);
1480 return x86_get_event_constraints(cpuc, event);
1484 intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc,
1485 struct perf_event *event)
1487 struct hw_perf_event_extra *reg;
1489 reg = &event->hw.extra_reg;
1490 if (reg->idx != EXTRA_REG_NONE)
1491 __intel_shared_reg_put_constraints(cpuc, reg);
1493 reg = &event->hw.branch_reg;
1494 if (reg->idx != EXTRA_REG_NONE)
1495 __intel_shared_reg_put_constraints(cpuc, reg);
1498 static void intel_put_event_constraints(struct cpu_hw_events *cpuc,
1499 struct perf_event *event)
1501 event->hw.flags = 0;
1502 intel_put_shared_regs_event_constraints(cpuc, event);
1505 static void intel_pebs_aliases_core2(struct perf_event *event)
1507 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
1509 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
1510 * (0x003c) so that we can use it with PEBS.
1512 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
1513 * PEBS capable. However we can use INST_RETIRED.ANY_P
1514 * (0x00c0), which is a PEBS capable event, to get the same
1517 * INST_RETIRED.ANY_P counts the number of cycles that retires
1518 * CNTMASK instructions. By setting CNTMASK to a value (16)
1519 * larger than the maximum number of instructions that can be
1520 * retired per cycle (4) and then inverting the condition, we
1521 * count all cycles that retire 16 or less instructions, which
1524 * Thereby we gain a PEBS capable cycle counter.
1526 u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16);
1528 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
1529 event->hw.config = alt_config;
1533 static void intel_pebs_aliases_snb(struct perf_event *event)
1535 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
1537 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
1538 * (0x003c) so that we can use it with PEBS.
1540 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
1541 * PEBS capable. However we can use UOPS_RETIRED.ALL
1542 * (0x01c2), which is a PEBS capable event, to get the same
1545 * UOPS_RETIRED.ALL counts the number of cycles that retires
1546 * CNTMASK micro-ops. By setting CNTMASK to a value (16)
1547 * larger than the maximum number of micro-ops that can be
1548 * retired per cycle (4) and then inverting the condition, we
1549 * count all cycles that retire 16 or less micro-ops, which
1552 * Thereby we gain a PEBS capable cycle counter.
1554 u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16);
1556 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
1557 event->hw.config = alt_config;
1561 static int intel_pmu_hw_config(struct perf_event *event)
1563 int ret = x86_pmu_hw_config(event);
1568 if (event->attr.precise_ip && x86_pmu.pebs_aliases)
1569 x86_pmu.pebs_aliases(event);
1571 if (intel_pmu_needs_lbr_smpl(event)) {
1572 ret = intel_pmu_setup_lbr_filter(event);
1577 if (event->attr.type != PERF_TYPE_RAW)
1580 if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY))
1583 if (x86_pmu.version < 3)
1586 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1589 event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY;
1594 struct perf_guest_switch_msr *perf_guest_get_msrs(int *nr)
1596 if (x86_pmu.guest_get_msrs)
1597 return x86_pmu.guest_get_msrs(nr);
1601 EXPORT_SYMBOL_GPL(perf_guest_get_msrs);
1603 static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr)
1605 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1606 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
1608 arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL;
1609 arr[0].host = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask;
1610 arr[0].guest = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_host_mask;
1612 * If PMU counter has PEBS enabled it is not enough to disable counter
1613 * on a guest entry since PEBS memory write can overshoot guest entry
1614 * and corrupt guest memory. Disabling PEBS solves the problem.
1616 arr[1].msr = MSR_IA32_PEBS_ENABLE;
1617 arr[1].host = cpuc->pebs_enabled;
1624 static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr)
1626 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1627 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
1630 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1631 struct perf_event *event = cpuc->events[idx];
1633 arr[idx].msr = x86_pmu_config_addr(idx);
1634 arr[idx].host = arr[idx].guest = 0;
1636 if (!test_bit(idx, cpuc->active_mask))
1639 arr[idx].host = arr[idx].guest =
1640 event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE;
1642 if (event->attr.exclude_host)
1643 arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
1644 else if (event->attr.exclude_guest)
1645 arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
1648 *nr = x86_pmu.num_counters;
1652 static void core_pmu_enable_event(struct perf_event *event)
1654 if (!event->attr.exclude_host)
1655 x86_pmu_enable_event(event);
1658 static void core_pmu_enable_all(int added)
1660 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1663 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1664 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
1666 if (!test_bit(idx, cpuc->active_mask) ||
1667 cpuc->events[idx]->attr.exclude_host)
1670 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
1674 static int hsw_hw_config(struct perf_event *event)
1676 int ret = intel_pmu_hw_config(event);
1680 if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE))
1682 event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED);
1685 * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with
1686 * PEBS or in ANY thread mode. Since the results are non-sensical forbid
1689 if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) &&
1690 ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) ||
1691 event->attr.precise_ip > 0))
1697 static struct event_constraint counter2_constraint =
1698 EVENT_CONSTRAINT(0, 0x4, 0);
1700 static struct event_constraint *
1701 hsw_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1703 struct event_constraint *c = intel_get_event_constraints(cpuc, event);
1705 /* Handle special quirk on in_tx_checkpointed only in counter 2 */
1706 if (event->hw.config & HSW_IN_TX_CHECKPOINTED) {
1707 if (c->idxmsk64 & (1U << 2))
1708 return &counter2_constraint;
1709 return &emptyconstraint;
1715 PMU_FORMAT_ATTR(event, "config:0-7" );
1716 PMU_FORMAT_ATTR(umask, "config:8-15" );
1717 PMU_FORMAT_ATTR(edge, "config:18" );
1718 PMU_FORMAT_ATTR(pc, "config:19" );
1719 PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */
1720 PMU_FORMAT_ATTR(inv, "config:23" );
1721 PMU_FORMAT_ATTR(cmask, "config:24-31" );
1722 PMU_FORMAT_ATTR(in_tx, "config:32");
1723 PMU_FORMAT_ATTR(in_tx_cp, "config:33");
1725 static struct attribute *intel_arch_formats_attr[] = {
1726 &format_attr_event.attr,
1727 &format_attr_umask.attr,
1728 &format_attr_edge.attr,
1729 &format_attr_pc.attr,
1730 &format_attr_inv.attr,
1731 &format_attr_cmask.attr,
1735 ssize_t intel_event_sysfs_show(char *page, u64 config)
1737 u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT);
1739 return x86_event_sysfs_show(page, config, event);
1742 static __initconst const struct x86_pmu core_pmu = {
1744 .handle_irq = x86_pmu_handle_irq,
1745 .disable_all = x86_pmu_disable_all,
1746 .enable_all = core_pmu_enable_all,
1747 .enable = core_pmu_enable_event,
1748 .disable = x86_pmu_disable_event,
1749 .hw_config = x86_pmu_hw_config,
1750 .schedule_events = x86_schedule_events,
1751 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
1752 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
1753 .event_map = intel_pmu_event_map,
1754 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
1757 * Intel PMCs cannot be accessed sanely above 32 bit width,
1758 * so we install an artificial 1<<31 period regardless of
1759 * the generic event period:
1761 .max_period = (1ULL << 31) - 1,
1762 .get_event_constraints = intel_get_event_constraints,
1763 .put_event_constraints = intel_put_event_constraints,
1764 .event_constraints = intel_core_event_constraints,
1765 .guest_get_msrs = core_guest_get_msrs,
1766 .format_attrs = intel_arch_formats_attr,
1767 .events_sysfs_show = intel_event_sysfs_show,
1770 struct intel_shared_regs *allocate_shared_regs(int cpu)
1772 struct intel_shared_regs *regs;
1775 regs = kzalloc_node(sizeof(struct intel_shared_regs),
1776 GFP_KERNEL, cpu_to_node(cpu));
1779 * initialize the locks to keep lockdep happy
1781 for (i = 0; i < EXTRA_REG_MAX; i++)
1782 raw_spin_lock_init(®s->regs[i].lock);
1789 static int intel_pmu_cpu_prepare(int cpu)
1791 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1793 if (!(x86_pmu.extra_regs || x86_pmu.lbr_sel_map))
1796 cpuc->shared_regs = allocate_shared_regs(cpu);
1797 if (!cpuc->shared_regs)
1803 static void intel_pmu_cpu_starting(int cpu)
1805 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1806 int core_id = topology_core_id(cpu);
1809 init_debug_store_on_cpu(cpu);
1811 * Deal with CPUs that don't clear their LBRs on power-up.
1813 intel_pmu_lbr_reset();
1815 cpuc->lbr_sel = NULL;
1817 if (!cpuc->shared_regs)
1820 if (!(x86_pmu.er_flags & ERF_NO_HT_SHARING)) {
1821 for_each_cpu(i, topology_thread_cpumask(cpu)) {
1822 struct intel_shared_regs *pc;
1824 pc = per_cpu(cpu_hw_events, i).shared_regs;
1825 if (pc && pc->core_id == core_id) {
1826 cpuc->kfree_on_online = cpuc->shared_regs;
1827 cpuc->shared_regs = pc;
1831 cpuc->shared_regs->core_id = core_id;
1832 cpuc->shared_regs->refcnt++;
1835 if (x86_pmu.lbr_sel_map)
1836 cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR];
1839 static void intel_pmu_cpu_dying(int cpu)
1841 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1842 struct intel_shared_regs *pc;
1844 pc = cpuc->shared_regs;
1846 if (pc->core_id == -1 || --pc->refcnt == 0)
1848 cpuc->shared_regs = NULL;
1851 fini_debug_store_on_cpu(cpu);
1854 static void intel_pmu_flush_branch_stack(void)
1857 * Intel LBR does not tag entries with the
1858 * PID of the current task, then we need to
1860 * For now, we simply reset it
1863 intel_pmu_lbr_reset();
1866 PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63");
1868 PMU_FORMAT_ATTR(ldlat, "config1:0-15");
1870 static struct attribute *intel_arch3_formats_attr[] = {
1871 &format_attr_event.attr,
1872 &format_attr_umask.attr,
1873 &format_attr_edge.attr,
1874 &format_attr_pc.attr,
1875 &format_attr_any.attr,
1876 &format_attr_inv.attr,
1877 &format_attr_cmask.attr,
1878 &format_attr_in_tx.attr,
1879 &format_attr_in_tx_cp.attr,
1881 &format_attr_offcore_rsp.attr, /* XXX do NHM/WSM + SNB breakout */
1882 &format_attr_ldlat.attr, /* PEBS load latency */
1886 static __initconst const struct x86_pmu intel_pmu = {
1888 .handle_irq = intel_pmu_handle_irq,
1889 .disable_all = intel_pmu_disable_all,
1890 .enable_all = intel_pmu_enable_all,
1891 .enable = intel_pmu_enable_event,
1892 .disable = intel_pmu_disable_event,
1893 .hw_config = intel_pmu_hw_config,
1894 .schedule_events = x86_schedule_events,
1895 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
1896 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
1897 .event_map = intel_pmu_event_map,
1898 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
1901 * Intel PMCs cannot be accessed sanely above 32 bit width,
1902 * so we install an artificial 1<<31 period regardless of
1903 * the generic event period:
1905 .max_period = (1ULL << 31) - 1,
1906 .get_event_constraints = intel_get_event_constraints,
1907 .put_event_constraints = intel_put_event_constraints,
1908 .pebs_aliases = intel_pebs_aliases_core2,
1910 .format_attrs = intel_arch3_formats_attr,
1911 .events_sysfs_show = intel_event_sysfs_show,
1913 .cpu_prepare = intel_pmu_cpu_prepare,
1914 .cpu_starting = intel_pmu_cpu_starting,
1915 .cpu_dying = intel_pmu_cpu_dying,
1916 .guest_get_msrs = intel_guest_get_msrs,
1917 .flush_branch_stack = intel_pmu_flush_branch_stack,
1920 static __init void intel_clovertown_quirk(void)
1923 * PEBS is unreliable due to:
1925 * AJ67 - PEBS may experience CPL leaks
1926 * AJ68 - PEBS PMI may be delayed by one event
1927 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
1928 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
1930 * AJ67 could be worked around by restricting the OS/USR flags.
1931 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
1933 * AJ106 could possibly be worked around by not allowing LBR
1934 * usage from PEBS, including the fixup.
1935 * AJ68 could possibly be worked around by always programming
1936 * a pebs_event_reset[0] value and coping with the lost events.
1938 * But taken together it might just make sense to not enable PEBS on
1941 pr_warn("PEBS disabled due to CPU errata\n");
1943 x86_pmu.pebs_constraints = NULL;
1946 static int intel_snb_pebs_broken(int cpu)
1948 u32 rev = UINT_MAX; /* default to broken for unknown models */
1950 switch (cpu_data(cpu).x86_model) {
1955 case 45: /* SNB-EP */
1956 switch (cpu_data(cpu).x86_mask) {
1957 case 6: rev = 0x618; break;
1958 case 7: rev = 0x70c; break;
1962 return (cpu_data(cpu).microcode < rev);
1965 static void intel_snb_check_microcode(void)
1967 int pebs_broken = 0;
1971 for_each_online_cpu(cpu) {
1972 if ((pebs_broken = intel_snb_pebs_broken(cpu)))
1977 if (pebs_broken == x86_pmu.pebs_broken)
1981 * Serialized by the microcode lock..
1983 if (x86_pmu.pebs_broken) {
1984 pr_info("PEBS enabled due to microcode update\n");
1985 x86_pmu.pebs_broken = 0;
1987 pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n");
1988 x86_pmu.pebs_broken = 1;
1992 static __init void intel_sandybridge_quirk(void)
1994 x86_pmu.check_microcode = intel_snb_check_microcode;
1995 intel_snb_check_microcode();
1998 static const struct { int id; char *name; } intel_arch_events_map[] __initconst = {
1999 { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" },
2000 { PERF_COUNT_HW_INSTRUCTIONS, "instructions" },
2001 { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" },
2002 { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" },
2003 { PERF_COUNT_HW_CACHE_MISSES, "cache misses" },
2004 { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" },
2005 { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" },
2008 static __init void intel_arch_events_quirk(void)
2012 /* disable event that reported as not presend by cpuid */
2013 for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) {
2014 intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0;
2015 pr_warn("CPUID marked event: \'%s\' unavailable\n",
2016 intel_arch_events_map[bit].name);
2020 static __init void intel_nehalem_quirk(void)
2022 union cpuid10_ebx ebx;
2024 ebx.full = x86_pmu.events_maskl;
2025 if (ebx.split.no_branch_misses_retired) {
2027 * Erratum AAJ80 detected, we work it around by using
2028 * the BR_MISP_EXEC.ANY event. This will over-count
2029 * branch-misses, but it's still much better than the
2030 * architectural event which is often completely bogus:
2032 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89;
2033 ebx.split.no_branch_misses_retired = 0;
2034 x86_pmu.events_maskl = ebx.full;
2035 pr_info("CPU erratum AAJ80 worked around\n");
2039 EVENT_ATTR_STR(mem-loads, mem_ld_hsw, "event=0xcd,umask=0x1,ldlat=3");
2040 EVENT_ATTR_STR(mem-stores, mem_st_hsw, "event=0xd0,umask=0x82")
2042 static struct attribute *hsw_events_attrs[] = {
2043 EVENT_PTR(mem_ld_hsw),
2044 EVENT_PTR(mem_st_hsw),
2048 __init int intel_pmu_init(void)
2050 union cpuid10_edx edx;
2051 union cpuid10_eax eax;
2052 union cpuid10_ebx ebx;
2053 struct event_constraint *c;
2054 unsigned int unused;
2057 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
2058 switch (boot_cpu_data.x86) {
2060 return p6_pmu_init();
2062 return knc_pmu_init();
2064 return p4_pmu_init();
2070 * Check whether the Architectural PerfMon supports
2071 * Branch Misses Retired hw_event or not.
2073 cpuid(10, &eax.full, &ebx.full, &unused, &edx.full);
2074 if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT)
2077 version = eax.split.version_id;
2081 x86_pmu = intel_pmu;
2083 x86_pmu.version = version;
2084 x86_pmu.num_counters = eax.split.num_counters;
2085 x86_pmu.cntval_bits = eax.split.bit_width;
2086 x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1;
2088 x86_pmu.events_maskl = ebx.full;
2089 x86_pmu.events_mask_len = eax.split.mask_length;
2091 x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters);
2094 * Quirk: v2 perfmon does not report fixed-purpose events, so
2095 * assume at least 3 events:
2098 x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3);
2101 * v2 and above have a perf capabilities MSR
2106 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
2107 x86_pmu.intel_cap.capabilities = capabilities;
2112 x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */
2115 * Install the hw-cache-events table:
2117 switch (boot_cpu_data.x86_model) {
2118 case 14: /* 65 nm core solo/duo, "Yonah" */
2119 pr_cont("Core events, ");
2122 case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
2123 x86_add_quirk(intel_clovertown_quirk);
2124 case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */
2125 case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */
2126 case 29: /* six-core 45 nm xeon "Dunnington" */
2127 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
2128 sizeof(hw_cache_event_ids));
2130 intel_pmu_lbr_init_core();
2132 x86_pmu.event_constraints = intel_core2_event_constraints;
2133 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints;
2134 pr_cont("Core2 events, ");
2137 case 26: /* 45 nm nehalem, "Bloomfield" */
2138 case 30: /* 45 nm nehalem, "Lynnfield" */
2139 case 46: /* 45 nm nehalem-ex, "Beckton" */
2140 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
2141 sizeof(hw_cache_event_ids));
2142 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
2143 sizeof(hw_cache_extra_regs));
2145 intel_pmu_lbr_init_nhm();
2147 x86_pmu.event_constraints = intel_nehalem_event_constraints;
2148 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints;
2149 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
2150 x86_pmu.extra_regs = intel_nehalem_extra_regs;
2152 x86_pmu.cpu_events = nhm_events_attrs;
2154 /* UOPS_ISSUED.STALLED_CYCLES */
2155 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
2156 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
2157 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
2158 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
2159 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
2161 x86_add_quirk(intel_nehalem_quirk);
2163 pr_cont("Nehalem events, ");
2167 case 38: /* Lincroft */
2168 case 39: /* Penwell */
2169 case 53: /* Cloverview */
2170 case 54: /* Cedarview */
2171 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
2172 sizeof(hw_cache_event_ids));
2174 intel_pmu_lbr_init_atom();
2176 x86_pmu.event_constraints = intel_gen_event_constraints;
2177 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints;
2178 pr_cont("Atom events, ");
2181 case 37: /* 32 nm nehalem, "Clarkdale" */
2182 case 44: /* 32 nm nehalem, "Gulftown" */
2183 case 47: /* 32 nm Xeon E7 */
2184 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
2185 sizeof(hw_cache_event_ids));
2186 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
2187 sizeof(hw_cache_extra_regs));
2189 intel_pmu_lbr_init_nhm();
2191 x86_pmu.event_constraints = intel_westmere_event_constraints;
2192 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
2193 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints;
2194 x86_pmu.extra_regs = intel_westmere_extra_regs;
2195 x86_pmu.er_flags |= ERF_HAS_RSP_1;
2197 x86_pmu.cpu_events = nhm_events_attrs;
2199 /* UOPS_ISSUED.STALLED_CYCLES */
2200 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
2201 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
2202 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
2203 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
2204 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
2206 pr_cont("Westmere events, ");
2209 case 42: /* SandyBridge */
2210 case 45: /* SandyBridge, "Romely-EP" */
2211 x86_add_quirk(intel_sandybridge_quirk);
2212 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
2213 sizeof(hw_cache_event_ids));
2214 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
2215 sizeof(hw_cache_extra_regs));
2217 intel_pmu_lbr_init_snb();
2219 x86_pmu.event_constraints = intel_snb_event_constraints;
2220 x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints;
2221 x86_pmu.pebs_aliases = intel_pebs_aliases_snb;
2222 if (boot_cpu_data.x86_model == 45)
2223 x86_pmu.extra_regs = intel_snbep_extra_regs;
2225 x86_pmu.extra_regs = intel_snb_extra_regs;
2226 /* all extra regs are per-cpu when HT is on */
2227 x86_pmu.er_flags |= ERF_HAS_RSP_1;
2228 x86_pmu.er_flags |= ERF_NO_HT_SHARING;
2230 x86_pmu.cpu_events = snb_events_attrs;
2232 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
2233 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
2234 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
2235 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
2236 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
2237 X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1);
2239 pr_cont("SandyBridge events, ");
2241 case 58: /* IvyBridge */
2242 case 62: /* IvyBridge EP */
2243 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
2244 sizeof(hw_cache_event_ids));
2245 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
2246 sizeof(hw_cache_extra_regs));
2248 intel_pmu_lbr_init_snb();
2250 x86_pmu.event_constraints = intel_ivb_event_constraints;
2251 x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints;
2252 x86_pmu.pebs_aliases = intel_pebs_aliases_snb;
2253 if (boot_cpu_data.x86_model == 62)
2254 x86_pmu.extra_regs = intel_snbep_extra_regs;
2256 x86_pmu.extra_regs = intel_snb_extra_regs;
2257 /* all extra regs are per-cpu when HT is on */
2258 x86_pmu.er_flags |= ERF_HAS_RSP_1;
2259 x86_pmu.er_flags |= ERF_NO_HT_SHARING;
2261 x86_pmu.cpu_events = snb_events_attrs;
2263 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
2264 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
2265 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
2267 pr_cont("IvyBridge events, ");
2271 case 60: /* Haswell Client */
2275 x86_pmu.late_ack = true;
2276 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, sizeof(hw_cache_event_ids));
2277 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
2279 intel_pmu_lbr_init_snb();
2281 x86_pmu.event_constraints = intel_hsw_event_constraints;
2282 x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints;
2283 x86_pmu.extra_regs = intel_snb_extra_regs;
2284 x86_pmu.pebs_aliases = intel_pebs_aliases_snb;
2285 /* all extra regs are per-cpu when HT is on */
2286 x86_pmu.er_flags |= ERF_HAS_RSP_1;
2287 x86_pmu.er_flags |= ERF_NO_HT_SHARING;
2289 x86_pmu.hw_config = hsw_hw_config;
2290 x86_pmu.get_event_constraints = hsw_get_event_constraints;
2291 x86_pmu.cpu_events = hsw_events_attrs;
2292 pr_cont("Haswell events, ");
2296 switch (x86_pmu.version) {
2298 x86_pmu.event_constraints = intel_v1_event_constraints;
2299 pr_cont("generic architected perfmon v1, ");
2303 * default constraints for v2 and up
2305 x86_pmu.event_constraints = intel_gen_event_constraints;
2306 pr_cont("generic architected perfmon, ");
2311 if (x86_pmu.num_counters > INTEL_PMC_MAX_GENERIC) {
2312 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
2313 x86_pmu.num_counters, INTEL_PMC_MAX_GENERIC);
2314 x86_pmu.num_counters = INTEL_PMC_MAX_GENERIC;
2316 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
2318 if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) {
2319 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
2320 x86_pmu.num_counters_fixed, INTEL_PMC_MAX_FIXED);
2321 x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED;
2324 x86_pmu.intel_ctrl |=
2325 ((1LL << x86_pmu.num_counters_fixed)-1) << INTEL_PMC_IDX_FIXED;
2327 if (x86_pmu.event_constraints) {
2329 * event on fixed counter2 (REF_CYCLES) only works on this
2330 * counter, so do not extend mask to generic counters
2332 for_each_event_constraint(c, x86_pmu.event_constraints) {
2333 if (c->cmask != FIXED_EVENT_FLAGS
2334 || c->idxmsk64 == INTEL_PMC_MSK_FIXED_REF_CYCLES) {
2338 c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1;
2339 c->weight += x86_pmu.num_counters;