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>
15 #include <linux/watchdog.h>
17 #include <asm/cpufeature.h>
18 #include <asm/hardirq.h>
21 #include "perf_event.h"
24 * Intel PerfMon, used on Core and later.
26 static u64 intel_perfmon_event_map
[PERF_COUNT_HW_MAX
] __read_mostly
=
28 [PERF_COUNT_HW_CPU_CYCLES
] = 0x003c,
29 [PERF_COUNT_HW_INSTRUCTIONS
] = 0x00c0,
30 [PERF_COUNT_HW_CACHE_REFERENCES
] = 0x4f2e,
31 [PERF_COUNT_HW_CACHE_MISSES
] = 0x412e,
32 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS
] = 0x00c4,
33 [PERF_COUNT_HW_BRANCH_MISSES
] = 0x00c5,
34 [PERF_COUNT_HW_BUS_CYCLES
] = 0x013c,
35 [PERF_COUNT_HW_REF_CPU_CYCLES
] = 0x0300, /* pseudo-encoding */
38 static struct event_constraint intel_core_event_constraints
[] __read_mostly
=
40 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
41 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
42 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
43 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
44 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
45 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */
49 static struct event_constraint intel_core2_event_constraints
[] __read_mostly
=
51 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
52 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
53 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
54 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
55 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
56 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
57 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
58 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
59 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
60 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
61 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
62 INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */
63 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
67 static struct event_constraint intel_nehalem_event_constraints
[] __read_mostly
=
69 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
70 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
71 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
72 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
73 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
74 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
75 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
76 INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */
77 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
78 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
79 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
83 static struct extra_reg intel_nehalem_extra_regs
[] __read_mostly
=
85 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
86 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0
, 0xffff, RSP_0
),
87 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
91 static struct event_constraint intel_westmere_event_constraints
[] __read_mostly
=
93 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
94 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
95 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
96 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
97 INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */
98 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
99 INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */
103 static struct event_constraint intel_snb_event_constraints
[] __read_mostly
=
105 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
106 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
107 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
108 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
109 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
110 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
111 INTEL_UEVENT_CONSTRAINT(0x06a3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
112 INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */
113 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
114 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
115 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
116 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
118 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
119 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
120 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
121 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
126 static struct event_constraint intel_ivb_event_constraints
[] __read_mostly
=
128 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
129 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
130 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
131 INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */
132 INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMTPY */
133 INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */
134 INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_LDM_PENDING */
135 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
136 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
137 INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */
138 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
139 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
140 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
142 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
143 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
144 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
145 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
150 static struct extra_reg intel_westmere_extra_regs
[] __read_mostly
=
152 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
153 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0
, 0xffff, RSP_0
),
154 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1
, 0xffff, RSP_1
),
155 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
159 static struct event_constraint intel_v1_event_constraints
[] __read_mostly
=
164 static struct event_constraint intel_gen_event_constraints
[] __read_mostly
=
166 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
167 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
168 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
172 static struct event_constraint intel_slm_event_constraints
[] __read_mostly
=
174 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
175 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
176 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */
180 static struct extra_reg intel_snb_extra_regs
[] __read_mostly
= {
181 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
182 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0
, 0x3f807f8fffull
, RSP_0
),
183 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1
, 0x3f807f8fffull
, RSP_1
),
184 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
188 static struct extra_reg intel_snbep_extra_regs
[] __read_mostly
= {
189 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
190 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0
, 0x3fffff8fffull
, RSP_0
),
191 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1
, 0x3fffff8fffull
, RSP_1
),
192 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
196 EVENT_ATTR_STR(mem
-loads
, mem_ld_nhm
, "event=0x0b,umask=0x10,ldlat=3");
197 EVENT_ATTR_STR(mem
-loads
, mem_ld_snb
, "event=0xcd,umask=0x1,ldlat=3");
198 EVENT_ATTR_STR(mem
-stores
, mem_st_snb
, "event=0xcd,umask=0x2");
200 struct attribute
*nhm_events_attrs
[] = {
201 EVENT_PTR(mem_ld_nhm
),
205 struct attribute
*snb_events_attrs
[] = {
206 EVENT_PTR(mem_ld_snb
),
207 EVENT_PTR(mem_st_snb
),
211 static struct event_constraint intel_hsw_event_constraints
[] = {
212 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
213 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
214 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
215 INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.* */
216 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
217 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
218 /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
219 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4),
220 /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
221 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4),
222 /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
223 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf),
225 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
226 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
227 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
228 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
233 struct event_constraint intel_bdw_event_constraints
[] = {
234 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
235 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
236 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
237 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */
238 INTEL_EVENT_CONSTRAINT(0xa3, 0x4), /* CYCLE_ACTIVITY.* */
242 static u64
intel_pmu_event_map(int hw_event
)
244 return intel_perfmon_event_map
[hw_event
];
247 #define SNB_DMND_DATA_RD (1ULL << 0)
248 #define SNB_DMND_RFO (1ULL << 1)
249 #define SNB_DMND_IFETCH (1ULL << 2)
250 #define SNB_DMND_WB (1ULL << 3)
251 #define SNB_PF_DATA_RD (1ULL << 4)
252 #define SNB_PF_RFO (1ULL << 5)
253 #define SNB_PF_IFETCH (1ULL << 6)
254 #define SNB_LLC_DATA_RD (1ULL << 7)
255 #define SNB_LLC_RFO (1ULL << 8)
256 #define SNB_LLC_IFETCH (1ULL << 9)
257 #define SNB_BUS_LOCKS (1ULL << 10)
258 #define SNB_STRM_ST (1ULL << 11)
259 #define SNB_OTHER (1ULL << 15)
260 #define SNB_RESP_ANY (1ULL << 16)
261 #define SNB_NO_SUPP (1ULL << 17)
262 #define SNB_LLC_HITM (1ULL << 18)
263 #define SNB_LLC_HITE (1ULL << 19)
264 #define SNB_LLC_HITS (1ULL << 20)
265 #define SNB_LLC_HITF (1ULL << 21)
266 #define SNB_LOCAL (1ULL << 22)
267 #define SNB_REMOTE (0xffULL << 23)
268 #define SNB_SNP_NONE (1ULL << 31)
269 #define SNB_SNP_NOT_NEEDED (1ULL << 32)
270 #define SNB_SNP_MISS (1ULL << 33)
271 #define SNB_NO_FWD (1ULL << 34)
272 #define SNB_SNP_FWD (1ULL << 35)
273 #define SNB_HITM (1ULL << 36)
274 #define SNB_NON_DRAM (1ULL << 37)
276 #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD)
277 #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO)
278 #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
280 #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \
281 SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \
284 #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY)
285 #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY)
287 #define SNB_L3_ACCESS SNB_RESP_ANY
288 #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM)
290 static __initconst
const u64 snb_hw_cache_extra_regs
291 [PERF_COUNT_HW_CACHE_MAX
]
292 [PERF_COUNT_HW_CACHE_OP_MAX
]
293 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
297 [ C(RESULT_ACCESS
) ] = SNB_DMND_READ
|SNB_L3_ACCESS
,
298 [ C(RESULT_MISS
) ] = SNB_DMND_READ
|SNB_L3_MISS
,
301 [ C(RESULT_ACCESS
) ] = SNB_DMND_WRITE
|SNB_L3_ACCESS
,
302 [ C(RESULT_MISS
) ] = SNB_DMND_WRITE
|SNB_L3_MISS
,
304 [ C(OP_PREFETCH
) ] = {
305 [ C(RESULT_ACCESS
) ] = SNB_DMND_PREFETCH
|SNB_L3_ACCESS
,
306 [ C(RESULT_MISS
) ] = SNB_DMND_PREFETCH
|SNB_L3_MISS
,
311 [ C(RESULT_ACCESS
) ] = SNB_DMND_READ
|SNB_DRAM_ANY
,
312 [ C(RESULT_MISS
) ] = SNB_DMND_READ
|SNB_DRAM_REMOTE
,
315 [ C(RESULT_ACCESS
) ] = SNB_DMND_WRITE
|SNB_DRAM_ANY
,
316 [ C(RESULT_MISS
) ] = SNB_DMND_WRITE
|SNB_DRAM_REMOTE
,
318 [ C(OP_PREFETCH
) ] = {
319 [ C(RESULT_ACCESS
) ] = SNB_DMND_PREFETCH
|SNB_DRAM_ANY
,
320 [ C(RESULT_MISS
) ] = SNB_DMND_PREFETCH
|SNB_DRAM_REMOTE
,
325 static __initconst
const u64 snb_hw_cache_event_ids
326 [PERF_COUNT_HW_CACHE_MAX
]
327 [PERF_COUNT_HW_CACHE_OP_MAX
]
328 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
332 [ C(RESULT_ACCESS
) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */
333 [ C(RESULT_MISS
) ] = 0x0151, /* L1D.REPLACEMENT */
336 [ C(RESULT_ACCESS
) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */
337 [ C(RESULT_MISS
) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */
339 [ C(OP_PREFETCH
) ] = {
340 [ C(RESULT_ACCESS
) ] = 0x0,
341 [ C(RESULT_MISS
) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */
346 [ C(RESULT_ACCESS
) ] = 0x0,
347 [ C(RESULT_MISS
) ] = 0x0280, /* ICACHE.MISSES */
350 [ C(RESULT_ACCESS
) ] = -1,
351 [ C(RESULT_MISS
) ] = -1,
353 [ C(OP_PREFETCH
) ] = {
354 [ C(RESULT_ACCESS
) ] = 0x0,
355 [ C(RESULT_MISS
) ] = 0x0,
360 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
361 [ C(RESULT_ACCESS
) ] = 0x01b7,
362 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
363 [ C(RESULT_MISS
) ] = 0x01b7,
366 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
367 [ C(RESULT_ACCESS
) ] = 0x01b7,
368 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
369 [ C(RESULT_MISS
) ] = 0x01b7,
371 [ C(OP_PREFETCH
) ] = {
372 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
373 [ C(RESULT_ACCESS
) ] = 0x01b7,
374 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
375 [ C(RESULT_MISS
) ] = 0x01b7,
380 [ C(RESULT_ACCESS
) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */
381 [ C(RESULT_MISS
) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */
384 [ C(RESULT_ACCESS
) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */
385 [ C(RESULT_MISS
) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
387 [ C(OP_PREFETCH
) ] = {
388 [ C(RESULT_ACCESS
) ] = 0x0,
389 [ C(RESULT_MISS
) ] = 0x0,
394 [ C(RESULT_ACCESS
) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */
395 [ C(RESULT_MISS
) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */
398 [ C(RESULT_ACCESS
) ] = -1,
399 [ C(RESULT_MISS
) ] = -1,
401 [ C(OP_PREFETCH
) ] = {
402 [ C(RESULT_ACCESS
) ] = -1,
403 [ C(RESULT_MISS
) ] = -1,
408 [ C(RESULT_ACCESS
) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
409 [ C(RESULT_MISS
) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
412 [ C(RESULT_ACCESS
) ] = -1,
413 [ C(RESULT_MISS
) ] = -1,
415 [ C(OP_PREFETCH
) ] = {
416 [ C(RESULT_ACCESS
) ] = -1,
417 [ C(RESULT_MISS
) ] = -1,
422 [ C(RESULT_ACCESS
) ] = 0x01b7,
423 [ C(RESULT_MISS
) ] = 0x01b7,
426 [ C(RESULT_ACCESS
) ] = 0x01b7,
427 [ C(RESULT_MISS
) ] = 0x01b7,
429 [ C(OP_PREFETCH
) ] = {
430 [ C(RESULT_ACCESS
) ] = 0x01b7,
431 [ C(RESULT_MISS
) ] = 0x01b7,
438 * Notes on the events:
439 * - data reads do not include code reads (comparable to earlier tables)
440 * - data counts include speculative execution (except L1 write, dtlb, bpu)
441 * - remote node access includes remote memory, remote cache, remote mmio.
442 * - prefetches are not included in the counts because they are not
446 #define HSW_DEMAND_DATA_RD BIT_ULL(0)
447 #define HSW_DEMAND_RFO BIT_ULL(1)
448 #define HSW_ANY_RESPONSE BIT_ULL(16)
449 #define HSW_SUPPLIER_NONE BIT_ULL(17)
450 #define HSW_L3_MISS_LOCAL_DRAM BIT_ULL(22)
451 #define HSW_L3_MISS_REMOTE_HOP0 BIT_ULL(27)
452 #define HSW_L3_MISS_REMOTE_HOP1 BIT_ULL(28)
453 #define HSW_L3_MISS_REMOTE_HOP2P BIT_ULL(29)
454 #define HSW_L3_MISS (HSW_L3_MISS_LOCAL_DRAM| \
455 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \
456 HSW_L3_MISS_REMOTE_HOP2P)
457 #define HSW_SNOOP_NONE BIT_ULL(31)
458 #define HSW_SNOOP_NOT_NEEDED BIT_ULL(32)
459 #define HSW_SNOOP_MISS BIT_ULL(33)
460 #define HSW_SNOOP_HIT_NO_FWD BIT_ULL(34)
461 #define HSW_SNOOP_HIT_WITH_FWD BIT_ULL(35)
462 #define HSW_SNOOP_HITM BIT_ULL(36)
463 #define HSW_SNOOP_NON_DRAM BIT_ULL(37)
464 #define HSW_ANY_SNOOP (HSW_SNOOP_NONE| \
465 HSW_SNOOP_NOT_NEEDED|HSW_SNOOP_MISS| \
466 HSW_SNOOP_HIT_NO_FWD|HSW_SNOOP_HIT_WITH_FWD| \
467 HSW_SNOOP_HITM|HSW_SNOOP_NON_DRAM)
468 #define HSW_SNOOP_DRAM (HSW_ANY_SNOOP & ~HSW_SNOOP_NON_DRAM)
469 #define HSW_DEMAND_READ HSW_DEMAND_DATA_RD
470 #define HSW_DEMAND_WRITE HSW_DEMAND_RFO
471 #define HSW_L3_MISS_REMOTE (HSW_L3_MISS_REMOTE_HOP0|\
472 HSW_L3_MISS_REMOTE_HOP1|HSW_L3_MISS_REMOTE_HOP2P)
473 #define HSW_LLC_ACCESS HSW_ANY_RESPONSE
475 #define BDW_L3_MISS_LOCAL BIT(26)
476 #define BDW_L3_MISS (BDW_L3_MISS_LOCAL| \
477 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \
478 HSW_L3_MISS_REMOTE_HOP2P)
481 static __initconst
const u64 hsw_hw_cache_event_ids
482 [PERF_COUNT_HW_CACHE_MAX
]
483 [PERF_COUNT_HW_CACHE_OP_MAX
]
484 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
488 [ C(RESULT_ACCESS
) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
489 [ C(RESULT_MISS
) ] = 0x151, /* L1D.REPLACEMENT */
492 [ C(RESULT_ACCESS
) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
493 [ C(RESULT_MISS
) ] = 0x0,
495 [ C(OP_PREFETCH
) ] = {
496 [ C(RESULT_ACCESS
) ] = 0x0,
497 [ C(RESULT_MISS
) ] = 0x0,
502 [ C(RESULT_ACCESS
) ] = 0x0,
503 [ C(RESULT_MISS
) ] = 0x280, /* ICACHE.MISSES */
506 [ C(RESULT_ACCESS
) ] = -1,
507 [ C(RESULT_MISS
) ] = -1,
509 [ C(OP_PREFETCH
) ] = {
510 [ C(RESULT_ACCESS
) ] = 0x0,
511 [ C(RESULT_MISS
) ] = 0x0,
516 [ C(RESULT_ACCESS
) ] = 0x1b7, /* OFFCORE_RESPONSE */
517 [ C(RESULT_MISS
) ] = 0x1b7, /* OFFCORE_RESPONSE */
520 [ C(RESULT_ACCESS
) ] = 0x1b7, /* OFFCORE_RESPONSE */
521 [ C(RESULT_MISS
) ] = 0x1b7, /* OFFCORE_RESPONSE */
523 [ C(OP_PREFETCH
) ] = {
524 [ C(RESULT_ACCESS
) ] = 0x0,
525 [ C(RESULT_MISS
) ] = 0x0,
530 [ C(RESULT_ACCESS
) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
531 [ C(RESULT_MISS
) ] = 0x108, /* DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK */
534 [ C(RESULT_ACCESS
) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
535 [ C(RESULT_MISS
) ] = 0x149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
537 [ C(OP_PREFETCH
) ] = {
538 [ C(RESULT_ACCESS
) ] = 0x0,
539 [ C(RESULT_MISS
) ] = 0x0,
544 [ C(RESULT_ACCESS
) ] = 0x6085, /* ITLB_MISSES.STLB_HIT */
545 [ C(RESULT_MISS
) ] = 0x185, /* ITLB_MISSES.MISS_CAUSES_A_WALK */
548 [ C(RESULT_ACCESS
) ] = -1,
549 [ C(RESULT_MISS
) ] = -1,
551 [ C(OP_PREFETCH
) ] = {
552 [ C(RESULT_ACCESS
) ] = -1,
553 [ C(RESULT_MISS
) ] = -1,
558 [ C(RESULT_ACCESS
) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */
559 [ C(RESULT_MISS
) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */
562 [ C(RESULT_ACCESS
) ] = -1,
563 [ C(RESULT_MISS
) ] = -1,
565 [ C(OP_PREFETCH
) ] = {
566 [ C(RESULT_ACCESS
) ] = -1,
567 [ C(RESULT_MISS
) ] = -1,
572 [ C(RESULT_ACCESS
) ] = 0x1b7, /* OFFCORE_RESPONSE */
573 [ C(RESULT_MISS
) ] = 0x1b7, /* OFFCORE_RESPONSE */
576 [ C(RESULT_ACCESS
) ] = 0x1b7, /* OFFCORE_RESPONSE */
577 [ C(RESULT_MISS
) ] = 0x1b7, /* OFFCORE_RESPONSE */
579 [ C(OP_PREFETCH
) ] = {
580 [ C(RESULT_ACCESS
) ] = 0x0,
581 [ C(RESULT_MISS
) ] = 0x0,
586 static __initconst
const u64 hsw_hw_cache_extra_regs
587 [PERF_COUNT_HW_CACHE_MAX
]
588 [PERF_COUNT_HW_CACHE_OP_MAX
]
589 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
593 [ C(RESULT_ACCESS
) ] = HSW_DEMAND_READ
|
595 [ C(RESULT_MISS
) ] = HSW_DEMAND_READ
|
596 HSW_L3_MISS
|HSW_ANY_SNOOP
,
599 [ C(RESULT_ACCESS
) ] = HSW_DEMAND_WRITE
|
601 [ C(RESULT_MISS
) ] = HSW_DEMAND_WRITE
|
602 HSW_L3_MISS
|HSW_ANY_SNOOP
,
604 [ C(OP_PREFETCH
) ] = {
605 [ C(RESULT_ACCESS
) ] = 0x0,
606 [ C(RESULT_MISS
) ] = 0x0,
611 [ C(RESULT_ACCESS
) ] = HSW_DEMAND_READ
|
612 HSW_L3_MISS_LOCAL_DRAM
|
614 [ C(RESULT_MISS
) ] = HSW_DEMAND_READ
|
619 [ C(RESULT_ACCESS
) ] = HSW_DEMAND_WRITE
|
620 HSW_L3_MISS_LOCAL_DRAM
|
622 [ C(RESULT_MISS
) ] = HSW_DEMAND_WRITE
|
626 [ C(OP_PREFETCH
) ] = {
627 [ C(RESULT_ACCESS
) ] = 0x0,
628 [ C(RESULT_MISS
) ] = 0x0,
633 static __initconst
const u64 westmere_hw_cache_event_ids
634 [PERF_COUNT_HW_CACHE_MAX
]
635 [PERF_COUNT_HW_CACHE_OP_MAX
]
636 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
640 [ C(RESULT_ACCESS
) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
641 [ C(RESULT_MISS
) ] = 0x0151, /* L1D.REPL */
644 [ C(RESULT_ACCESS
) ] = 0x020b, /* MEM_INST_RETURED.STORES */
645 [ C(RESULT_MISS
) ] = 0x0251, /* L1D.M_REPL */
647 [ C(OP_PREFETCH
) ] = {
648 [ C(RESULT_ACCESS
) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
649 [ C(RESULT_MISS
) ] = 0x024e, /* L1D_PREFETCH.MISS */
654 [ C(RESULT_ACCESS
) ] = 0x0380, /* L1I.READS */
655 [ C(RESULT_MISS
) ] = 0x0280, /* L1I.MISSES */
658 [ C(RESULT_ACCESS
) ] = -1,
659 [ C(RESULT_MISS
) ] = -1,
661 [ C(OP_PREFETCH
) ] = {
662 [ C(RESULT_ACCESS
) ] = 0x0,
663 [ C(RESULT_MISS
) ] = 0x0,
668 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
669 [ C(RESULT_ACCESS
) ] = 0x01b7,
670 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
671 [ C(RESULT_MISS
) ] = 0x01b7,
674 * Use RFO, not WRITEBACK, because a write miss would typically occur
678 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
679 [ C(RESULT_ACCESS
) ] = 0x01b7,
680 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
681 [ C(RESULT_MISS
) ] = 0x01b7,
683 [ C(OP_PREFETCH
) ] = {
684 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
685 [ C(RESULT_ACCESS
) ] = 0x01b7,
686 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
687 [ C(RESULT_MISS
) ] = 0x01b7,
692 [ C(RESULT_ACCESS
) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
693 [ C(RESULT_MISS
) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
696 [ C(RESULT_ACCESS
) ] = 0x020b, /* MEM_INST_RETURED.STORES */
697 [ C(RESULT_MISS
) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
699 [ C(OP_PREFETCH
) ] = {
700 [ C(RESULT_ACCESS
) ] = 0x0,
701 [ C(RESULT_MISS
) ] = 0x0,
706 [ C(RESULT_ACCESS
) ] = 0x01c0, /* INST_RETIRED.ANY_P */
707 [ C(RESULT_MISS
) ] = 0x0185, /* ITLB_MISSES.ANY */
710 [ C(RESULT_ACCESS
) ] = -1,
711 [ C(RESULT_MISS
) ] = -1,
713 [ C(OP_PREFETCH
) ] = {
714 [ C(RESULT_ACCESS
) ] = -1,
715 [ C(RESULT_MISS
) ] = -1,
720 [ C(RESULT_ACCESS
) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
721 [ C(RESULT_MISS
) ] = 0x03e8, /* BPU_CLEARS.ANY */
724 [ C(RESULT_ACCESS
) ] = -1,
725 [ C(RESULT_MISS
) ] = -1,
727 [ C(OP_PREFETCH
) ] = {
728 [ C(RESULT_ACCESS
) ] = -1,
729 [ C(RESULT_MISS
) ] = -1,
734 [ C(RESULT_ACCESS
) ] = 0x01b7,
735 [ C(RESULT_MISS
) ] = 0x01b7,
738 [ C(RESULT_ACCESS
) ] = 0x01b7,
739 [ C(RESULT_MISS
) ] = 0x01b7,
741 [ C(OP_PREFETCH
) ] = {
742 [ C(RESULT_ACCESS
) ] = 0x01b7,
743 [ C(RESULT_MISS
) ] = 0x01b7,
749 * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits;
750 * See IA32 SDM Vol 3B 30.6.1.3
753 #define NHM_DMND_DATA_RD (1 << 0)
754 #define NHM_DMND_RFO (1 << 1)
755 #define NHM_DMND_IFETCH (1 << 2)
756 #define NHM_DMND_WB (1 << 3)
757 #define NHM_PF_DATA_RD (1 << 4)
758 #define NHM_PF_DATA_RFO (1 << 5)
759 #define NHM_PF_IFETCH (1 << 6)
760 #define NHM_OFFCORE_OTHER (1 << 7)
761 #define NHM_UNCORE_HIT (1 << 8)
762 #define NHM_OTHER_CORE_HIT_SNP (1 << 9)
763 #define NHM_OTHER_CORE_HITM (1 << 10)
765 #define NHM_REMOTE_CACHE_FWD (1 << 12)
766 #define NHM_REMOTE_DRAM (1 << 13)
767 #define NHM_LOCAL_DRAM (1 << 14)
768 #define NHM_NON_DRAM (1 << 15)
770 #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD)
771 #define NHM_REMOTE (NHM_REMOTE_DRAM)
773 #define NHM_DMND_READ (NHM_DMND_DATA_RD)
774 #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB)
775 #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO)
777 #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM)
778 #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD)
779 #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS)
781 static __initconst
const u64 nehalem_hw_cache_extra_regs
782 [PERF_COUNT_HW_CACHE_MAX
]
783 [PERF_COUNT_HW_CACHE_OP_MAX
]
784 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
788 [ C(RESULT_ACCESS
) ] = NHM_DMND_READ
|NHM_L3_ACCESS
,
789 [ C(RESULT_MISS
) ] = NHM_DMND_READ
|NHM_L3_MISS
,
792 [ C(RESULT_ACCESS
) ] = NHM_DMND_WRITE
|NHM_L3_ACCESS
,
793 [ C(RESULT_MISS
) ] = NHM_DMND_WRITE
|NHM_L3_MISS
,
795 [ C(OP_PREFETCH
) ] = {
796 [ C(RESULT_ACCESS
) ] = NHM_DMND_PREFETCH
|NHM_L3_ACCESS
,
797 [ C(RESULT_MISS
) ] = NHM_DMND_PREFETCH
|NHM_L3_MISS
,
802 [ C(RESULT_ACCESS
) ] = NHM_DMND_READ
|NHM_LOCAL
|NHM_REMOTE
,
803 [ C(RESULT_MISS
) ] = NHM_DMND_READ
|NHM_REMOTE
,
806 [ C(RESULT_ACCESS
) ] = NHM_DMND_WRITE
|NHM_LOCAL
|NHM_REMOTE
,
807 [ C(RESULT_MISS
) ] = NHM_DMND_WRITE
|NHM_REMOTE
,
809 [ C(OP_PREFETCH
) ] = {
810 [ C(RESULT_ACCESS
) ] = NHM_DMND_PREFETCH
|NHM_LOCAL
|NHM_REMOTE
,
811 [ C(RESULT_MISS
) ] = NHM_DMND_PREFETCH
|NHM_REMOTE
,
816 static __initconst
const u64 nehalem_hw_cache_event_ids
817 [PERF_COUNT_HW_CACHE_MAX
]
818 [PERF_COUNT_HW_CACHE_OP_MAX
]
819 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
823 [ C(RESULT_ACCESS
) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
824 [ C(RESULT_MISS
) ] = 0x0151, /* L1D.REPL */
827 [ C(RESULT_ACCESS
) ] = 0x020b, /* MEM_INST_RETURED.STORES */
828 [ C(RESULT_MISS
) ] = 0x0251, /* L1D.M_REPL */
830 [ C(OP_PREFETCH
) ] = {
831 [ C(RESULT_ACCESS
) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
832 [ C(RESULT_MISS
) ] = 0x024e, /* L1D_PREFETCH.MISS */
837 [ C(RESULT_ACCESS
) ] = 0x0380, /* L1I.READS */
838 [ C(RESULT_MISS
) ] = 0x0280, /* L1I.MISSES */
841 [ C(RESULT_ACCESS
) ] = -1,
842 [ C(RESULT_MISS
) ] = -1,
844 [ C(OP_PREFETCH
) ] = {
845 [ C(RESULT_ACCESS
) ] = 0x0,
846 [ C(RESULT_MISS
) ] = 0x0,
851 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
852 [ C(RESULT_ACCESS
) ] = 0x01b7,
853 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
854 [ C(RESULT_MISS
) ] = 0x01b7,
857 * Use RFO, not WRITEBACK, because a write miss would typically occur
861 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
862 [ C(RESULT_ACCESS
) ] = 0x01b7,
863 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
864 [ C(RESULT_MISS
) ] = 0x01b7,
866 [ C(OP_PREFETCH
) ] = {
867 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
868 [ C(RESULT_ACCESS
) ] = 0x01b7,
869 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
870 [ C(RESULT_MISS
) ] = 0x01b7,
875 [ C(RESULT_ACCESS
) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
876 [ C(RESULT_MISS
) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
879 [ C(RESULT_ACCESS
) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
880 [ C(RESULT_MISS
) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
882 [ C(OP_PREFETCH
) ] = {
883 [ C(RESULT_ACCESS
) ] = 0x0,
884 [ C(RESULT_MISS
) ] = 0x0,
889 [ C(RESULT_ACCESS
) ] = 0x01c0, /* INST_RETIRED.ANY_P */
890 [ C(RESULT_MISS
) ] = 0x20c8, /* ITLB_MISS_RETIRED */
893 [ C(RESULT_ACCESS
) ] = -1,
894 [ C(RESULT_MISS
) ] = -1,
896 [ C(OP_PREFETCH
) ] = {
897 [ C(RESULT_ACCESS
) ] = -1,
898 [ C(RESULT_MISS
) ] = -1,
903 [ C(RESULT_ACCESS
) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
904 [ C(RESULT_MISS
) ] = 0x03e8, /* BPU_CLEARS.ANY */
907 [ C(RESULT_ACCESS
) ] = -1,
908 [ C(RESULT_MISS
) ] = -1,
910 [ C(OP_PREFETCH
) ] = {
911 [ C(RESULT_ACCESS
) ] = -1,
912 [ C(RESULT_MISS
) ] = -1,
917 [ C(RESULT_ACCESS
) ] = 0x01b7,
918 [ C(RESULT_MISS
) ] = 0x01b7,
921 [ C(RESULT_ACCESS
) ] = 0x01b7,
922 [ C(RESULT_MISS
) ] = 0x01b7,
924 [ C(OP_PREFETCH
) ] = {
925 [ C(RESULT_ACCESS
) ] = 0x01b7,
926 [ C(RESULT_MISS
) ] = 0x01b7,
931 static __initconst
const u64 core2_hw_cache_event_ids
932 [PERF_COUNT_HW_CACHE_MAX
]
933 [PERF_COUNT_HW_CACHE_OP_MAX
]
934 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
938 [ C(RESULT_ACCESS
) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
939 [ C(RESULT_MISS
) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
942 [ C(RESULT_ACCESS
) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
943 [ C(RESULT_MISS
) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
945 [ C(OP_PREFETCH
) ] = {
946 [ C(RESULT_ACCESS
) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */
947 [ C(RESULT_MISS
) ] = 0,
952 [ C(RESULT_ACCESS
) ] = 0x0080, /* L1I.READS */
953 [ C(RESULT_MISS
) ] = 0x0081, /* L1I.MISSES */
956 [ C(RESULT_ACCESS
) ] = -1,
957 [ C(RESULT_MISS
) ] = -1,
959 [ C(OP_PREFETCH
) ] = {
960 [ C(RESULT_ACCESS
) ] = 0,
961 [ C(RESULT_MISS
) ] = 0,
966 [ C(RESULT_ACCESS
) ] = 0x4f29, /* L2_LD.MESI */
967 [ C(RESULT_MISS
) ] = 0x4129, /* L2_LD.ISTATE */
970 [ C(RESULT_ACCESS
) ] = 0x4f2A, /* L2_ST.MESI */
971 [ C(RESULT_MISS
) ] = 0x412A, /* L2_ST.ISTATE */
973 [ C(OP_PREFETCH
) ] = {
974 [ C(RESULT_ACCESS
) ] = 0,
975 [ C(RESULT_MISS
) ] = 0,
980 [ C(RESULT_ACCESS
) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
981 [ C(RESULT_MISS
) ] = 0x0208, /* DTLB_MISSES.MISS_LD */
984 [ C(RESULT_ACCESS
) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
985 [ C(RESULT_MISS
) ] = 0x0808, /* DTLB_MISSES.MISS_ST */
987 [ C(OP_PREFETCH
) ] = {
988 [ C(RESULT_ACCESS
) ] = 0,
989 [ C(RESULT_MISS
) ] = 0,
994 [ C(RESULT_ACCESS
) ] = 0x00c0, /* INST_RETIRED.ANY_P */
995 [ C(RESULT_MISS
) ] = 0x1282, /* ITLBMISSES */
998 [ C(RESULT_ACCESS
) ] = -1,
999 [ C(RESULT_MISS
) ] = -1,
1001 [ C(OP_PREFETCH
) ] = {
1002 [ C(RESULT_ACCESS
) ] = -1,
1003 [ C(RESULT_MISS
) ] = -1,
1008 [ C(RESULT_ACCESS
) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1009 [ C(RESULT_MISS
) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1012 [ C(RESULT_ACCESS
) ] = -1,
1013 [ C(RESULT_MISS
) ] = -1,
1015 [ C(OP_PREFETCH
) ] = {
1016 [ C(RESULT_ACCESS
) ] = -1,
1017 [ C(RESULT_MISS
) ] = -1,
1022 static __initconst
const u64 atom_hw_cache_event_ids
1023 [PERF_COUNT_HW_CACHE_MAX
]
1024 [PERF_COUNT_HW_CACHE_OP_MAX
]
1025 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
1029 [ C(RESULT_ACCESS
) ] = 0x2140, /* L1D_CACHE.LD */
1030 [ C(RESULT_MISS
) ] = 0,
1033 [ C(RESULT_ACCESS
) ] = 0x2240, /* L1D_CACHE.ST */
1034 [ C(RESULT_MISS
) ] = 0,
1036 [ C(OP_PREFETCH
) ] = {
1037 [ C(RESULT_ACCESS
) ] = 0x0,
1038 [ C(RESULT_MISS
) ] = 0,
1043 [ C(RESULT_ACCESS
) ] = 0x0380, /* L1I.READS */
1044 [ C(RESULT_MISS
) ] = 0x0280, /* L1I.MISSES */
1047 [ C(RESULT_ACCESS
) ] = -1,
1048 [ C(RESULT_MISS
) ] = -1,
1050 [ C(OP_PREFETCH
) ] = {
1051 [ C(RESULT_ACCESS
) ] = 0,
1052 [ C(RESULT_MISS
) ] = 0,
1057 [ C(RESULT_ACCESS
) ] = 0x4f29, /* L2_LD.MESI */
1058 [ C(RESULT_MISS
) ] = 0x4129, /* L2_LD.ISTATE */
1061 [ C(RESULT_ACCESS
) ] = 0x4f2A, /* L2_ST.MESI */
1062 [ C(RESULT_MISS
) ] = 0x412A, /* L2_ST.ISTATE */
1064 [ C(OP_PREFETCH
) ] = {
1065 [ C(RESULT_ACCESS
) ] = 0,
1066 [ C(RESULT_MISS
) ] = 0,
1071 [ C(RESULT_ACCESS
) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */
1072 [ C(RESULT_MISS
) ] = 0x0508, /* DTLB_MISSES.MISS_LD */
1075 [ C(RESULT_ACCESS
) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */
1076 [ C(RESULT_MISS
) ] = 0x0608, /* DTLB_MISSES.MISS_ST */
1078 [ C(OP_PREFETCH
) ] = {
1079 [ C(RESULT_ACCESS
) ] = 0,
1080 [ C(RESULT_MISS
) ] = 0,
1085 [ C(RESULT_ACCESS
) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1086 [ C(RESULT_MISS
) ] = 0x0282, /* ITLB.MISSES */
1089 [ C(RESULT_ACCESS
) ] = -1,
1090 [ C(RESULT_MISS
) ] = -1,
1092 [ C(OP_PREFETCH
) ] = {
1093 [ C(RESULT_ACCESS
) ] = -1,
1094 [ C(RESULT_MISS
) ] = -1,
1099 [ C(RESULT_ACCESS
) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1100 [ C(RESULT_MISS
) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1103 [ C(RESULT_ACCESS
) ] = -1,
1104 [ C(RESULT_MISS
) ] = -1,
1106 [ C(OP_PREFETCH
) ] = {
1107 [ C(RESULT_ACCESS
) ] = -1,
1108 [ C(RESULT_MISS
) ] = -1,
1113 static struct extra_reg intel_slm_extra_regs
[] __read_mostly
=
1115 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
1116 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0
, 0x768005ffffull
, RSP_0
),
1117 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1
, 0x368005ffffull
, RSP_1
),
1121 #define SLM_DMND_READ SNB_DMND_DATA_RD
1122 #define SLM_DMND_WRITE SNB_DMND_RFO
1123 #define SLM_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
1125 #define SLM_SNP_ANY (SNB_SNP_NONE|SNB_SNP_MISS|SNB_NO_FWD|SNB_HITM)
1126 #define SLM_LLC_ACCESS SNB_RESP_ANY
1127 #define SLM_LLC_MISS (SLM_SNP_ANY|SNB_NON_DRAM)
1129 static __initconst
const u64 slm_hw_cache_extra_regs
1130 [PERF_COUNT_HW_CACHE_MAX
]
1131 [PERF_COUNT_HW_CACHE_OP_MAX
]
1132 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
1136 [ C(RESULT_ACCESS
) ] = SLM_DMND_READ
|SLM_LLC_ACCESS
,
1137 [ C(RESULT_MISS
) ] = 0,
1140 [ C(RESULT_ACCESS
) ] = SLM_DMND_WRITE
|SLM_LLC_ACCESS
,
1141 [ C(RESULT_MISS
) ] = SLM_DMND_WRITE
|SLM_LLC_MISS
,
1143 [ C(OP_PREFETCH
) ] = {
1144 [ C(RESULT_ACCESS
) ] = SLM_DMND_PREFETCH
|SLM_LLC_ACCESS
,
1145 [ C(RESULT_MISS
) ] = SLM_DMND_PREFETCH
|SLM_LLC_MISS
,
1150 static __initconst
const u64 slm_hw_cache_event_ids
1151 [PERF_COUNT_HW_CACHE_MAX
]
1152 [PERF_COUNT_HW_CACHE_OP_MAX
]
1153 [PERF_COUNT_HW_CACHE_RESULT_MAX
] =
1157 [ C(RESULT_ACCESS
) ] = 0,
1158 [ C(RESULT_MISS
) ] = 0x0104, /* LD_DCU_MISS */
1161 [ C(RESULT_ACCESS
) ] = 0,
1162 [ C(RESULT_MISS
) ] = 0,
1164 [ C(OP_PREFETCH
) ] = {
1165 [ C(RESULT_ACCESS
) ] = 0,
1166 [ C(RESULT_MISS
) ] = 0,
1171 [ C(RESULT_ACCESS
) ] = 0x0380, /* ICACHE.ACCESSES */
1172 [ C(RESULT_MISS
) ] = 0x0280, /* ICACGE.MISSES */
1175 [ C(RESULT_ACCESS
) ] = -1,
1176 [ C(RESULT_MISS
) ] = -1,
1178 [ C(OP_PREFETCH
) ] = {
1179 [ C(RESULT_ACCESS
) ] = 0,
1180 [ C(RESULT_MISS
) ] = 0,
1185 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1186 [ C(RESULT_ACCESS
) ] = 0x01b7,
1187 [ C(RESULT_MISS
) ] = 0,
1190 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1191 [ C(RESULT_ACCESS
) ] = 0x01b7,
1192 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1193 [ C(RESULT_MISS
) ] = 0x01b7,
1195 [ C(OP_PREFETCH
) ] = {
1196 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1197 [ C(RESULT_ACCESS
) ] = 0x01b7,
1198 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1199 [ C(RESULT_MISS
) ] = 0x01b7,
1204 [ C(RESULT_ACCESS
) ] = 0,
1205 [ C(RESULT_MISS
) ] = 0x0804, /* LD_DTLB_MISS */
1208 [ C(RESULT_ACCESS
) ] = 0,
1209 [ C(RESULT_MISS
) ] = 0,
1211 [ C(OP_PREFETCH
) ] = {
1212 [ C(RESULT_ACCESS
) ] = 0,
1213 [ C(RESULT_MISS
) ] = 0,
1218 [ C(RESULT_ACCESS
) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1219 [ C(RESULT_MISS
) ] = 0x40205, /* PAGE_WALKS.I_SIDE_WALKS */
1222 [ C(RESULT_ACCESS
) ] = -1,
1223 [ C(RESULT_MISS
) ] = -1,
1225 [ C(OP_PREFETCH
) ] = {
1226 [ C(RESULT_ACCESS
) ] = -1,
1227 [ C(RESULT_MISS
) ] = -1,
1232 [ C(RESULT_ACCESS
) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1233 [ C(RESULT_MISS
) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1236 [ C(RESULT_ACCESS
) ] = -1,
1237 [ C(RESULT_MISS
) ] = -1,
1239 [ C(OP_PREFETCH
) ] = {
1240 [ C(RESULT_ACCESS
) ] = -1,
1241 [ C(RESULT_MISS
) ] = -1,
1247 * Use from PMIs where the LBRs are already disabled.
1249 static void __intel_pmu_disable_all(void)
1251 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
1253 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, 0);
1255 if (test_bit(INTEL_PMC_IDX_FIXED_BTS
, cpuc
->active_mask
))
1256 intel_pmu_disable_bts();
1258 intel_bts_disable_local();
1260 intel_pmu_pebs_disable_all();
1263 static void intel_pmu_disable_all(void)
1265 __intel_pmu_disable_all();
1266 intel_pmu_lbr_disable_all();
1269 static void __intel_pmu_enable_all(int added
, bool pmi
)
1271 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
1273 intel_pmu_pebs_enable_all();
1274 intel_pmu_lbr_enable_all(pmi
);
1275 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL
,
1276 x86_pmu
.intel_ctrl
& ~cpuc
->intel_ctrl_guest_mask
);
1278 if (test_bit(INTEL_PMC_IDX_FIXED_BTS
, cpuc
->active_mask
)) {
1279 struct perf_event
*event
=
1280 cpuc
->events
[INTEL_PMC_IDX_FIXED_BTS
];
1282 if (WARN_ON_ONCE(!event
))
1285 intel_pmu_enable_bts(event
->hw
.config
);
1287 intel_bts_enable_local();
1290 static void intel_pmu_enable_all(int added
)
1292 __intel_pmu_enable_all(added
, false);
1297 * Intel Errata AAK100 (model 26)
1298 * Intel Errata AAP53 (model 30)
1299 * Intel Errata BD53 (model 44)
1301 * The official story:
1302 * These chips need to be 'reset' when adding counters by programming the
1303 * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either
1304 * in sequence on the same PMC or on different PMCs.
1306 * In practise it appears some of these events do in fact count, and
1307 * we need to programm all 4 events.
1309 static void intel_pmu_nhm_workaround(void)
1311 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
1312 static const unsigned long nhm_magic
[4] = {
1318 struct perf_event
*event
;
1322 * The Errata requires below steps:
1323 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL;
1324 * 2) Configure 4 PERFEVTSELx with the magic events and clear
1325 * the corresponding PMCx;
1326 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL;
1327 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL;
1328 * 5) Clear 4 pairs of ERFEVTSELx and PMCx;
1332 * The real steps we choose are a little different from above.
1333 * A) To reduce MSR operations, we don't run step 1) as they
1334 * are already cleared before this function is called;
1335 * B) Call x86_perf_event_update to save PMCx before configuring
1336 * PERFEVTSELx with magic number;
1337 * C) With step 5), we do clear only when the PERFEVTSELx is
1338 * not used currently.
1339 * D) Call x86_perf_event_set_period to restore PMCx;
1342 /* We always operate 4 pairs of PERF Counters */
1343 for (i
= 0; i
< 4; i
++) {
1344 event
= cpuc
->events
[i
];
1346 x86_perf_event_update(event
);
1349 for (i
= 0; i
< 4; i
++) {
1350 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0
+ i
, nhm_magic
[i
]);
1351 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0
+ i
, 0x0);
1354 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, 0xf);
1355 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, 0x0);
1357 for (i
= 0; i
< 4; i
++) {
1358 event
= cpuc
->events
[i
];
1361 x86_perf_event_set_period(event
);
1362 __x86_pmu_enable_event(&event
->hw
,
1363 ARCH_PERFMON_EVENTSEL_ENABLE
);
1365 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0
+ i
, 0x0);
1369 static void intel_pmu_nhm_enable_all(int added
)
1372 intel_pmu_nhm_workaround();
1373 intel_pmu_enable_all(added
);
1376 static inline u64
intel_pmu_get_status(void)
1380 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS
, status
);
1385 static inline void intel_pmu_ack_status(u64 ack
)
1387 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL
, ack
);
1390 static void intel_pmu_disable_fixed(struct hw_perf_event
*hwc
)
1392 int idx
= hwc
->idx
- INTEL_PMC_IDX_FIXED
;
1395 mask
= 0xfULL
<< (idx
* 4);
1397 rdmsrl(hwc
->config_base
, ctrl_val
);
1399 wrmsrl(hwc
->config_base
, ctrl_val
);
1402 static inline bool event_is_checkpointed(struct perf_event
*event
)
1404 return (event
->hw
.config
& HSW_IN_TX_CHECKPOINTED
) != 0;
1407 static void intel_pmu_disable_event(struct perf_event
*event
)
1409 struct hw_perf_event
*hwc
= &event
->hw
;
1410 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
1412 if (unlikely(hwc
->idx
== INTEL_PMC_IDX_FIXED_BTS
)) {
1413 intel_pmu_disable_bts();
1414 intel_pmu_drain_bts_buffer();
1418 cpuc
->intel_ctrl_guest_mask
&= ~(1ull << hwc
->idx
);
1419 cpuc
->intel_ctrl_host_mask
&= ~(1ull << hwc
->idx
);
1420 cpuc
->intel_cp_status
&= ~(1ull << hwc
->idx
);
1423 * must disable before any actual event
1424 * because any event may be combined with LBR
1426 if (needs_branch_stack(event
))
1427 intel_pmu_lbr_disable(event
);
1429 if (unlikely(hwc
->config_base
== MSR_ARCH_PERFMON_FIXED_CTR_CTRL
)) {
1430 intel_pmu_disable_fixed(hwc
);
1434 x86_pmu_disable_event(event
);
1436 if (unlikely(event
->attr
.precise_ip
))
1437 intel_pmu_pebs_disable(event
);
1440 static void intel_pmu_enable_fixed(struct hw_perf_event
*hwc
)
1442 int idx
= hwc
->idx
- INTEL_PMC_IDX_FIXED
;
1443 u64 ctrl_val
, bits
, mask
;
1446 * Enable IRQ generation (0x8),
1447 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
1451 if (hwc
->config
& ARCH_PERFMON_EVENTSEL_USR
)
1453 if (hwc
->config
& ARCH_PERFMON_EVENTSEL_OS
)
1457 * ANY bit is supported in v3 and up
1459 if (x86_pmu
.version
> 2 && hwc
->config
& ARCH_PERFMON_EVENTSEL_ANY
)
1463 mask
= 0xfULL
<< (idx
* 4);
1465 rdmsrl(hwc
->config_base
, ctrl_val
);
1468 wrmsrl(hwc
->config_base
, ctrl_val
);
1471 static void intel_pmu_enable_event(struct perf_event
*event
)
1473 struct hw_perf_event
*hwc
= &event
->hw
;
1474 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
1476 if (unlikely(hwc
->idx
== INTEL_PMC_IDX_FIXED_BTS
)) {
1477 if (!__this_cpu_read(cpu_hw_events
.enabled
))
1480 intel_pmu_enable_bts(hwc
->config
);
1484 * must enabled before any actual event
1485 * because any event may be combined with LBR
1487 if (needs_branch_stack(event
))
1488 intel_pmu_lbr_enable(event
);
1490 if (event
->attr
.exclude_host
)
1491 cpuc
->intel_ctrl_guest_mask
|= (1ull << hwc
->idx
);
1492 if (event
->attr
.exclude_guest
)
1493 cpuc
->intel_ctrl_host_mask
|= (1ull << hwc
->idx
);
1495 if (unlikely(event_is_checkpointed(event
)))
1496 cpuc
->intel_cp_status
|= (1ull << hwc
->idx
);
1498 if (unlikely(hwc
->config_base
== MSR_ARCH_PERFMON_FIXED_CTR_CTRL
)) {
1499 intel_pmu_enable_fixed(hwc
);
1503 if (unlikely(event
->attr
.precise_ip
))
1504 intel_pmu_pebs_enable(event
);
1506 __x86_pmu_enable_event(hwc
, ARCH_PERFMON_EVENTSEL_ENABLE
);
1510 * Save and restart an expired event. Called by NMI contexts,
1511 * so it has to be careful about preempting normal event ops:
1513 int intel_pmu_save_and_restart(struct perf_event
*event
)
1515 x86_perf_event_update(event
);
1517 * For a checkpointed counter always reset back to 0. This
1518 * avoids a situation where the counter overflows, aborts the
1519 * transaction and is then set back to shortly before the
1520 * overflow, and overflows and aborts again.
1522 if (unlikely(event_is_checkpointed(event
))) {
1523 /* No race with NMIs because the counter should not be armed */
1524 wrmsrl(event
->hw
.event_base
, 0);
1525 local64_set(&event
->hw
.prev_count
, 0);
1527 return x86_perf_event_set_period(event
);
1530 static void intel_pmu_reset(void)
1532 struct debug_store
*ds
= __this_cpu_read(cpu_hw_events
.ds
);
1533 unsigned long flags
;
1536 if (!x86_pmu
.num_counters
)
1539 local_irq_save(flags
);
1541 pr_info("clearing PMU state on CPU#%d\n", smp_processor_id());
1543 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
1544 wrmsrl_safe(x86_pmu_config_addr(idx
), 0ull);
1545 wrmsrl_safe(x86_pmu_event_addr(idx
), 0ull);
1547 for (idx
= 0; idx
< x86_pmu
.num_counters_fixed
; idx
++)
1548 wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0
+ idx
, 0ull);
1551 ds
->bts_index
= ds
->bts_buffer_base
;
1553 /* Ack all overflows and disable fixed counters */
1554 if (x86_pmu
.version
>= 2) {
1555 intel_pmu_ack_status(intel_pmu_get_status());
1556 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, 0);
1559 /* Reset LBRs and LBR freezing */
1560 if (x86_pmu
.lbr_nr
) {
1561 update_debugctlmsr(get_debugctlmsr() &
1562 ~(DEBUGCTLMSR_FREEZE_LBRS_ON_PMI
|DEBUGCTLMSR_LBR
));
1565 local_irq_restore(flags
);
1569 * This handler is triggered by the local APIC, so the APIC IRQ handling
1572 static int intel_pmu_handle_irq(struct pt_regs
*regs
)
1574 struct perf_sample_data data
;
1575 struct cpu_hw_events
*cpuc
;
1580 cpuc
= this_cpu_ptr(&cpu_hw_events
);
1583 * No known reason to not always do late ACK,
1584 * but just in case do it opt-in.
1586 if (!x86_pmu
.late_ack
)
1587 apic_write(APIC_LVTPC
, APIC_DM_NMI
);
1588 __intel_pmu_disable_all();
1589 handled
= intel_pmu_drain_bts_buffer();
1590 handled
+= intel_bts_interrupt();
1591 status
= intel_pmu_get_status();
1597 intel_pmu_ack_status(status
);
1598 if (++loops
> 100) {
1599 static bool warned
= false;
1601 WARN(1, "perfevents: irq loop stuck!\n");
1602 perf_event_print_debug();
1609 inc_irq_stat(apic_perf_irqs
);
1611 intel_pmu_lbr_read();
1614 * CondChgd bit 63 doesn't mean any overflow status. Ignore
1615 * and clear the bit.
1617 if (__test_and_clear_bit(63, (unsigned long *)&status
)) {
1623 * PEBS overflow sets bit 62 in the global status register
1625 if (__test_and_clear_bit(62, (unsigned long *)&status
)) {
1627 x86_pmu
.drain_pebs(regs
);
1633 if (__test_and_clear_bit(55, (unsigned long *)&status
)) {
1635 intel_pt_interrupt();
1639 * Checkpointed counters can lead to 'spurious' PMIs because the
1640 * rollback caused by the PMI will have cleared the overflow status
1641 * bit. Therefore always force probe these counters.
1643 status
|= cpuc
->intel_cp_status
;
1645 for_each_set_bit(bit
, (unsigned long *)&status
, X86_PMC_IDX_MAX
) {
1646 struct perf_event
*event
= cpuc
->events
[bit
];
1650 if (!test_bit(bit
, cpuc
->active_mask
))
1653 if (!intel_pmu_save_and_restart(event
))
1656 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
1658 if (has_branch_stack(event
))
1659 data
.br_stack
= &cpuc
->lbr_stack
;
1661 if (perf_event_overflow(event
, &data
, regs
))
1662 x86_pmu_stop(event
, 0);
1666 * Repeat if there is more work to be done:
1668 status
= intel_pmu_get_status();
1673 __intel_pmu_enable_all(0, true);
1675 * Only unmask the NMI after the overflow counters
1676 * have been reset. This avoids spurious NMIs on
1679 if (x86_pmu
.late_ack
)
1680 apic_write(APIC_LVTPC
, APIC_DM_NMI
);
1684 static struct event_constraint
*
1685 intel_bts_constraints(struct perf_event
*event
)
1687 struct hw_perf_event
*hwc
= &event
->hw
;
1688 unsigned int hw_event
, bts_event
;
1690 if (event
->attr
.freq
)
1693 hw_event
= hwc
->config
& INTEL_ARCH_EVENT_MASK
;
1694 bts_event
= x86_pmu
.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS
);
1696 if (unlikely(hw_event
== bts_event
&& hwc
->sample_period
== 1))
1697 return &bts_constraint
;
1702 static int intel_alt_er(int idx
, u64 config
)
1705 if (!(x86_pmu
.flags
& PMU_FL_HAS_RSP_1
))
1708 if (idx
== EXTRA_REG_RSP_0
)
1709 alt_idx
= EXTRA_REG_RSP_1
;
1711 if (idx
== EXTRA_REG_RSP_1
)
1712 alt_idx
= EXTRA_REG_RSP_0
;
1714 if (config
& ~x86_pmu
.extra_regs
[alt_idx
].valid_mask
)
1720 static void intel_fixup_er(struct perf_event
*event
, int idx
)
1722 event
->hw
.extra_reg
.idx
= idx
;
1724 if (idx
== EXTRA_REG_RSP_0
) {
1725 event
->hw
.config
&= ~INTEL_ARCH_EVENT_MASK
;
1726 event
->hw
.config
|= x86_pmu
.extra_regs
[EXTRA_REG_RSP_0
].event
;
1727 event
->hw
.extra_reg
.reg
= MSR_OFFCORE_RSP_0
;
1728 } else if (idx
== EXTRA_REG_RSP_1
) {
1729 event
->hw
.config
&= ~INTEL_ARCH_EVENT_MASK
;
1730 event
->hw
.config
|= x86_pmu
.extra_regs
[EXTRA_REG_RSP_1
].event
;
1731 event
->hw
.extra_reg
.reg
= MSR_OFFCORE_RSP_1
;
1736 * manage allocation of shared extra msr for certain events
1739 * per-cpu: to be shared between the various events on a single PMU
1740 * per-core: per-cpu + shared by HT threads
1742 static struct event_constraint
*
1743 __intel_shared_reg_get_constraints(struct cpu_hw_events
*cpuc
,
1744 struct perf_event
*event
,
1745 struct hw_perf_event_extra
*reg
)
1747 struct event_constraint
*c
= &emptyconstraint
;
1748 struct er_account
*era
;
1749 unsigned long flags
;
1753 * reg->alloc can be set due to existing state, so for fake cpuc we
1754 * need to ignore this, otherwise we might fail to allocate proper fake
1755 * state for this extra reg constraint. Also see the comment below.
1757 if (reg
->alloc
&& !cpuc
->is_fake
)
1758 return NULL
; /* call x86_get_event_constraint() */
1761 era
= &cpuc
->shared_regs
->regs
[idx
];
1763 * we use spin_lock_irqsave() to avoid lockdep issues when
1764 * passing a fake cpuc
1766 raw_spin_lock_irqsave(&era
->lock
, flags
);
1768 if (!atomic_read(&era
->ref
) || era
->config
== reg
->config
) {
1771 * If its a fake cpuc -- as per validate_{group,event}() we
1772 * shouldn't touch event state and we can avoid doing so
1773 * since both will only call get_event_constraints() once
1774 * on each event, this avoids the need for reg->alloc.
1776 * Not doing the ER fixup will only result in era->reg being
1777 * wrong, but since we won't actually try and program hardware
1778 * this isn't a problem either.
1780 if (!cpuc
->is_fake
) {
1781 if (idx
!= reg
->idx
)
1782 intel_fixup_er(event
, idx
);
1785 * x86_schedule_events() can call get_event_constraints()
1786 * multiple times on events in the case of incremental
1787 * scheduling(). reg->alloc ensures we only do the ER
1793 /* lock in msr value */
1794 era
->config
= reg
->config
;
1795 era
->reg
= reg
->reg
;
1798 atomic_inc(&era
->ref
);
1801 * need to call x86_get_event_constraint()
1802 * to check if associated event has constraints
1806 idx
= intel_alt_er(idx
, reg
->config
);
1807 if (idx
!= reg
->idx
) {
1808 raw_spin_unlock_irqrestore(&era
->lock
, flags
);
1812 raw_spin_unlock_irqrestore(&era
->lock
, flags
);
1818 __intel_shared_reg_put_constraints(struct cpu_hw_events
*cpuc
,
1819 struct hw_perf_event_extra
*reg
)
1821 struct er_account
*era
;
1824 * Only put constraint if extra reg was actually allocated. Also takes
1825 * care of event which do not use an extra shared reg.
1827 * Also, if this is a fake cpuc we shouldn't touch any event state
1828 * (reg->alloc) and we don't care about leaving inconsistent cpuc state
1829 * either since it'll be thrown out.
1831 if (!reg
->alloc
|| cpuc
->is_fake
)
1834 era
= &cpuc
->shared_regs
->regs
[reg
->idx
];
1836 /* one fewer user */
1837 atomic_dec(&era
->ref
);
1839 /* allocate again next time */
1843 static struct event_constraint
*
1844 intel_shared_regs_constraints(struct cpu_hw_events
*cpuc
,
1845 struct perf_event
*event
)
1847 struct event_constraint
*c
= NULL
, *d
;
1848 struct hw_perf_event_extra
*xreg
, *breg
;
1850 xreg
= &event
->hw
.extra_reg
;
1851 if (xreg
->idx
!= EXTRA_REG_NONE
) {
1852 c
= __intel_shared_reg_get_constraints(cpuc
, event
, xreg
);
1853 if (c
== &emptyconstraint
)
1856 breg
= &event
->hw
.branch_reg
;
1857 if (breg
->idx
!= EXTRA_REG_NONE
) {
1858 d
= __intel_shared_reg_get_constraints(cpuc
, event
, breg
);
1859 if (d
== &emptyconstraint
) {
1860 __intel_shared_reg_put_constraints(cpuc
, xreg
);
1867 struct event_constraint
*
1868 x86_get_event_constraints(struct cpu_hw_events
*cpuc
, int idx
,
1869 struct perf_event
*event
)
1871 struct event_constraint
*c
;
1873 if (x86_pmu
.event_constraints
) {
1874 for_each_event_constraint(c
, x86_pmu
.event_constraints
) {
1875 if ((event
->hw
.config
& c
->cmask
) == c
->code
) {
1876 event
->hw
.flags
|= c
->flags
;
1882 return &unconstrained
;
1885 static struct event_constraint
*
1886 __intel_get_event_constraints(struct cpu_hw_events
*cpuc
, int idx
,
1887 struct perf_event
*event
)
1889 struct event_constraint
*c
;
1891 c
= intel_bts_constraints(event
);
1895 c
= intel_shared_regs_constraints(cpuc
, event
);
1899 c
= intel_pebs_constraints(event
);
1903 return x86_get_event_constraints(cpuc
, idx
, event
);
1907 intel_start_scheduling(struct cpu_hw_events
*cpuc
)
1909 struct intel_excl_cntrs
*excl_cntrs
= cpuc
->excl_cntrs
;
1910 struct intel_excl_states
*xl
;
1911 int tid
= cpuc
->excl_thread_id
;
1914 * nothing needed if in group validation mode
1916 if (cpuc
->is_fake
|| !is_ht_workaround_enabled())
1920 * no exclusion needed
1922 if (WARN_ON_ONCE(!excl_cntrs
))
1925 xl
= &excl_cntrs
->states
[tid
];
1927 xl
->sched_started
= true;
1929 * lock shared state until we are done scheduling
1930 * in stop_event_scheduling()
1931 * makes scheduling appear as a transaction
1933 raw_spin_lock(&excl_cntrs
->lock
);
1936 static void intel_commit_scheduling(struct cpu_hw_events
*cpuc
, int idx
, int cntr
)
1938 struct intel_excl_cntrs
*excl_cntrs
= cpuc
->excl_cntrs
;
1939 struct event_constraint
*c
= cpuc
->event_constraint
[idx
];
1940 struct intel_excl_states
*xl
;
1941 int tid
= cpuc
->excl_thread_id
;
1943 if (cpuc
->is_fake
|| !is_ht_workaround_enabled())
1946 if (WARN_ON_ONCE(!excl_cntrs
))
1949 if (!(c
->flags
& PERF_X86_EVENT_DYNAMIC
))
1952 xl
= &excl_cntrs
->states
[tid
];
1954 lockdep_assert_held(&excl_cntrs
->lock
);
1956 if (c
->flags
& PERF_X86_EVENT_EXCL
)
1957 xl
->state
[cntr
] = INTEL_EXCL_EXCLUSIVE
;
1959 xl
->state
[cntr
] = INTEL_EXCL_SHARED
;
1963 intel_stop_scheduling(struct cpu_hw_events
*cpuc
)
1965 struct intel_excl_cntrs
*excl_cntrs
= cpuc
->excl_cntrs
;
1966 struct intel_excl_states
*xl
;
1967 int tid
= cpuc
->excl_thread_id
;
1970 * nothing needed if in group validation mode
1972 if (cpuc
->is_fake
|| !is_ht_workaround_enabled())
1975 * no exclusion needed
1977 if (WARN_ON_ONCE(!excl_cntrs
))
1980 xl
= &excl_cntrs
->states
[tid
];
1982 xl
->sched_started
= false;
1984 * release shared state lock (acquired in intel_start_scheduling())
1986 raw_spin_unlock(&excl_cntrs
->lock
);
1989 static struct event_constraint
*
1990 intel_get_excl_constraints(struct cpu_hw_events
*cpuc
, struct perf_event
*event
,
1991 int idx
, struct event_constraint
*c
)
1993 struct intel_excl_cntrs
*excl_cntrs
= cpuc
->excl_cntrs
;
1994 struct intel_excl_states
*xlo
;
1995 int tid
= cpuc
->excl_thread_id
;
1999 * validating a group does not require
2000 * enforcing cross-thread exclusion
2002 if (cpuc
->is_fake
|| !is_ht_workaround_enabled())
2006 * no exclusion needed
2008 if (WARN_ON_ONCE(!excl_cntrs
))
2012 * because we modify the constraint, we need
2013 * to make a copy. Static constraints come
2014 * from static const tables.
2016 * only needed when constraint has not yet
2017 * been cloned (marked dynamic)
2019 if (!(c
->flags
& PERF_X86_EVENT_DYNAMIC
)) {
2020 struct event_constraint
*cx
;
2023 * grab pre-allocated constraint entry
2025 cx
= &cpuc
->constraint_list
[idx
];
2028 * initialize dynamic constraint
2029 * with static constraint
2034 * mark constraint as dynamic, so we
2035 * can free it later on
2037 cx
->flags
|= PERF_X86_EVENT_DYNAMIC
;
2042 * From here on, the constraint is dynamic.
2043 * Either it was just allocated above, or it
2044 * was allocated during a earlier invocation
2049 * state of sibling HT
2051 xlo
= &excl_cntrs
->states
[tid
^ 1];
2054 * event requires exclusive counter access
2057 is_excl
= c
->flags
& PERF_X86_EVENT_EXCL
;
2058 if (is_excl
&& !(event
->hw
.flags
& PERF_X86_EVENT_EXCL_ACCT
)) {
2059 event
->hw
.flags
|= PERF_X86_EVENT_EXCL_ACCT
;
2060 if (!cpuc
->n_excl
++)
2061 WRITE_ONCE(excl_cntrs
->has_exclusive
[tid
], 1);
2065 * Modify static constraint with current dynamic
2068 * EXCLUSIVE: sibling counter measuring exclusive event
2069 * SHARED : sibling counter measuring non-exclusive event
2070 * UNUSED : sibling counter unused
2072 for_each_set_bit(i
, c
->idxmsk
, X86_PMC_IDX_MAX
) {
2074 * exclusive event in sibling counter
2075 * our corresponding counter cannot be used
2076 * regardless of our event
2078 if (xlo
->state
[i
] == INTEL_EXCL_EXCLUSIVE
)
2079 __clear_bit(i
, c
->idxmsk
);
2081 * if measuring an exclusive event, sibling
2082 * measuring non-exclusive, then counter cannot
2085 if (is_excl
&& xlo
->state
[i
] == INTEL_EXCL_SHARED
)
2086 __clear_bit(i
, c
->idxmsk
);
2090 * recompute actual bit weight for scheduling algorithm
2092 c
->weight
= hweight64(c
->idxmsk64
);
2095 * if we return an empty mask, then switch
2096 * back to static empty constraint to avoid
2097 * the cost of freeing later on
2100 c
= &emptyconstraint
;
2105 static struct event_constraint
*
2106 intel_get_event_constraints(struct cpu_hw_events
*cpuc
, int idx
,
2107 struct perf_event
*event
)
2109 struct event_constraint
*c1
= cpuc
->event_constraint
[idx
];
2110 struct event_constraint
*c2
;
2114 * - static constraint: no change across incremental scheduling calls
2115 * - dynamic constraint: handled by intel_get_excl_constraints()
2117 c2
= __intel_get_event_constraints(cpuc
, idx
, event
);
2118 if (c1
&& (c1
->flags
& PERF_X86_EVENT_DYNAMIC
)) {
2119 bitmap_copy(c1
->idxmsk
, c2
->idxmsk
, X86_PMC_IDX_MAX
);
2120 c1
->weight
= c2
->weight
;
2124 if (cpuc
->excl_cntrs
)
2125 return intel_get_excl_constraints(cpuc
, event
, idx
, c2
);
2130 static void intel_put_excl_constraints(struct cpu_hw_events
*cpuc
,
2131 struct perf_event
*event
)
2133 struct hw_perf_event
*hwc
= &event
->hw
;
2134 struct intel_excl_cntrs
*excl_cntrs
= cpuc
->excl_cntrs
;
2135 int tid
= cpuc
->excl_thread_id
;
2136 struct intel_excl_states
*xl
;
2139 * nothing needed if in group validation mode
2144 if (WARN_ON_ONCE(!excl_cntrs
))
2147 if (hwc
->flags
& PERF_X86_EVENT_EXCL_ACCT
) {
2148 hwc
->flags
&= ~PERF_X86_EVENT_EXCL_ACCT
;
2149 if (!--cpuc
->n_excl
)
2150 WRITE_ONCE(excl_cntrs
->has_exclusive
[tid
], 0);
2154 * If event was actually assigned, then mark the counter state as
2157 if (hwc
->idx
>= 0) {
2158 xl
= &excl_cntrs
->states
[tid
];
2161 * put_constraint may be called from x86_schedule_events()
2162 * which already has the lock held so here make locking
2165 if (!xl
->sched_started
)
2166 raw_spin_lock(&excl_cntrs
->lock
);
2168 xl
->state
[hwc
->idx
] = INTEL_EXCL_UNUSED
;
2170 if (!xl
->sched_started
)
2171 raw_spin_unlock(&excl_cntrs
->lock
);
2176 intel_put_shared_regs_event_constraints(struct cpu_hw_events
*cpuc
,
2177 struct perf_event
*event
)
2179 struct hw_perf_event_extra
*reg
;
2181 reg
= &event
->hw
.extra_reg
;
2182 if (reg
->idx
!= EXTRA_REG_NONE
)
2183 __intel_shared_reg_put_constraints(cpuc
, reg
);
2185 reg
= &event
->hw
.branch_reg
;
2186 if (reg
->idx
!= EXTRA_REG_NONE
)
2187 __intel_shared_reg_put_constraints(cpuc
, reg
);
2190 static void intel_put_event_constraints(struct cpu_hw_events
*cpuc
,
2191 struct perf_event
*event
)
2193 intel_put_shared_regs_event_constraints(cpuc
, event
);
2196 * is PMU has exclusive counter restrictions, then
2197 * all events are subject to and must call the
2198 * put_excl_constraints() routine
2200 if (cpuc
->excl_cntrs
)
2201 intel_put_excl_constraints(cpuc
, event
);
2204 static void intel_pebs_aliases_core2(struct perf_event
*event
)
2206 if ((event
->hw
.config
& X86_RAW_EVENT_MASK
) == 0x003c) {
2208 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
2209 * (0x003c) so that we can use it with PEBS.
2211 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
2212 * PEBS capable. However we can use INST_RETIRED.ANY_P
2213 * (0x00c0), which is a PEBS capable event, to get the same
2216 * INST_RETIRED.ANY_P counts the number of cycles that retires
2217 * CNTMASK instructions. By setting CNTMASK to a value (16)
2218 * larger than the maximum number of instructions that can be
2219 * retired per cycle (4) and then inverting the condition, we
2220 * count all cycles that retire 16 or less instructions, which
2223 * Thereby we gain a PEBS capable cycle counter.
2225 u64 alt_config
= X86_CONFIG(.event
=0xc0, .inv
=1, .cmask
=16);
2227 alt_config
|= (event
->hw
.config
& ~X86_RAW_EVENT_MASK
);
2228 event
->hw
.config
= alt_config
;
2232 static void intel_pebs_aliases_snb(struct perf_event
*event
)
2234 if ((event
->hw
.config
& X86_RAW_EVENT_MASK
) == 0x003c) {
2236 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
2237 * (0x003c) so that we can use it with PEBS.
2239 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
2240 * PEBS capable. However we can use UOPS_RETIRED.ALL
2241 * (0x01c2), which is a PEBS capable event, to get the same
2244 * UOPS_RETIRED.ALL counts the number of cycles that retires
2245 * CNTMASK micro-ops. By setting CNTMASK to a value (16)
2246 * larger than the maximum number of micro-ops that can be
2247 * retired per cycle (4) and then inverting the condition, we
2248 * count all cycles that retire 16 or less micro-ops, which
2251 * Thereby we gain a PEBS capable cycle counter.
2253 u64 alt_config
= X86_CONFIG(.event
=0xc2, .umask
=0x01, .inv
=1, .cmask
=16);
2255 alt_config
|= (event
->hw
.config
& ~X86_RAW_EVENT_MASK
);
2256 event
->hw
.config
= alt_config
;
2260 static int intel_pmu_hw_config(struct perf_event
*event
)
2262 int ret
= x86_pmu_hw_config(event
);
2267 if (event
->attr
.precise_ip
) {
2268 if (!event
->attr
.freq
) {
2269 event
->hw
.flags
|= PERF_X86_EVENT_AUTO_RELOAD
;
2270 if (!(event
->attr
.sample_type
& ~PEBS_FREERUNNING_FLAGS
))
2271 event
->hw
.flags
|= PERF_X86_EVENT_FREERUNNING
;
2273 if (x86_pmu
.pebs_aliases
)
2274 x86_pmu
.pebs_aliases(event
);
2277 if (needs_branch_stack(event
)) {
2278 ret
= intel_pmu_setup_lbr_filter(event
);
2283 * BTS is set up earlier in this path, so don't account twice
2285 if (!intel_pmu_has_bts(event
)) {
2286 /* disallow lbr if conflicting events are present */
2287 if (x86_add_exclusive(x86_lbr_exclusive_lbr
))
2290 event
->destroy
= hw_perf_lbr_event_destroy
;
2294 if (event
->attr
.type
!= PERF_TYPE_RAW
)
2297 if (!(event
->attr
.config
& ARCH_PERFMON_EVENTSEL_ANY
))
2300 if (x86_pmu
.version
< 3)
2303 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
2306 event
->hw
.config
|= ARCH_PERFMON_EVENTSEL_ANY
;
2311 struct perf_guest_switch_msr
*perf_guest_get_msrs(int *nr
)
2313 if (x86_pmu
.guest_get_msrs
)
2314 return x86_pmu
.guest_get_msrs(nr
);
2318 EXPORT_SYMBOL_GPL(perf_guest_get_msrs
);
2320 static struct perf_guest_switch_msr
*intel_guest_get_msrs(int *nr
)
2322 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
2323 struct perf_guest_switch_msr
*arr
= cpuc
->guest_switch_msrs
;
2325 arr
[0].msr
= MSR_CORE_PERF_GLOBAL_CTRL
;
2326 arr
[0].host
= x86_pmu
.intel_ctrl
& ~cpuc
->intel_ctrl_guest_mask
;
2327 arr
[0].guest
= x86_pmu
.intel_ctrl
& ~cpuc
->intel_ctrl_host_mask
;
2329 * If PMU counter has PEBS enabled it is not enough to disable counter
2330 * on a guest entry since PEBS memory write can overshoot guest entry
2331 * and corrupt guest memory. Disabling PEBS solves the problem.
2333 arr
[1].msr
= MSR_IA32_PEBS_ENABLE
;
2334 arr
[1].host
= cpuc
->pebs_enabled
;
2341 static struct perf_guest_switch_msr
*core_guest_get_msrs(int *nr
)
2343 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
2344 struct perf_guest_switch_msr
*arr
= cpuc
->guest_switch_msrs
;
2347 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
2348 struct perf_event
*event
= cpuc
->events
[idx
];
2350 arr
[idx
].msr
= x86_pmu_config_addr(idx
);
2351 arr
[idx
].host
= arr
[idx
].guest
= 0;
2353 if (!test_bit(idx
, cpuc
->active_mask
))
2356 arr
[idx
].host
= arr
[idx
].guest
=
2357 event
->hw
.config
| ARCH_PERFMON_EVENTSEL_ENABLE
;
2359 if (event
->attr
.exclude_host
)
2360 arr
[idx
].host
&= ~ARCH_PERFMON_EVENTSEL_ENABLE
;
2361 else if (event
->attr
.exclude_guest
)
2362 arr
[idx
].guest
&= ~ARCH_PERFMON_EVENTSEL_ENABLE
;
2365 *nr
= x86_pmu
.num_counters
;
2369 static void core_pmu_enable_event(struct perf_event
*event
)
2371 if (!event
->attr
.exclude_host
)
2372 x86_pmu_enable_event(event
);
2375 static void core_pmu_enable_all(int added
)
2377 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
2380 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
2381 struct hw_perf_event
*hwc
= &cpuc
->events
[idx
]->hw
;
2383 if (!test_bit(idx
, cpuc
->active_mask
) ||
2384 cpuc
->events
[idx
]->attr
.exclude_host
)
2387 __x86_pmu_enable_event(hwc
, ARCH_PERFMON_EVENTSEL_ENABLE
);
2391 static int hsw_hw_config(struct perf_event
*event
)
2393 int ret
= intel_pmu_hw_config(event
);
2397 if (!boot_cpu_has(X86_FEATURE_RTM
) && !boot_cpu_has(X86_FEATURE_HLE
))
2399 event
->hw
.config
|= event
->attr
.config
& (HSW_IN_TX
|HSW_IN_TX_CHECKPOINTED
);
2402 * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with
2403 * PEBS or in ANY thread mode. Since the results are non-sensical forbid
2406 if ((event
->hw
.config
& (HSW_IN_TX
|HSW_IN_TX_CHECKPOINTED
)) &&
2407 ((event
->hw
.config
& ARCH_PERFMON_EVENTSEL_ANY
) ||
2408 event
->attr
.precise_ip
> 0))
2411 if (event_is_checkpointed(event
)) {
2413 * Sampling of checkpointed events can cause situations where
2414 * the CPU constantly aborts because of a overflow, which is
2415 * then checkpointed back and ignored. Forbid checkpointing
2418 * But still allow a long sampling period, so that perf stat
2421 if (event
->attr
.sample_period
> 0 &&
2422 event
->attr
.sample_period
< 0x7fffffff)
2428 static struct event_constraint counter2_constraint
=
2429 EVENT_CONSTRAINT(0, 0x4, 0);
2431 static struct event_constraint
*
2432 hsw_get_event_constraints(struct cpu_hw_events
*cpuc
, int idx
,
2433 struct perf_event
*event
)
2435 struct event_constraint
*c
;
2437 c
= intel_get_event_constraints(cpuc
, idx
, event
);
2439 /* Handle special quirk on in_tx_checkpointed only in counter 2 */
2440 if (event
->hw
.config
& HSW_IN_TX_CHECKPOINTED
) {
2441 if (c
->idxmsk64
& (1U << 2))
2442 return &counter2_constraint
;
2443 return &emptyconstraint
;
2452 * The INST_RETIRED.ALL period always needs to have lowest 6 bits cleared
2453 * (BDM55) and it must not use a period smaller than 100 (BDM11). We combine
2454 * the two to enforce a minimum period of 128 (the smallest value that has bits
2455 * 0-5 cleared and >= 100).
2457 * Because of how the code in x86_perf_event_set_period() works, the truncation
2458 * of the lower 6 bits is 'harmless' as we'll occasionally add a longer period
2459 * to make up for the 'lost' events due to carrying the 'error' in period_left.
2461 * Therefore the effective (average) period matches the requested period,
2462 * despite coarser hardware granularity.
2464 static unsigned bdw_limit_period(struct perf_event
*event
, unsigned left
)
2466 if ((event
->hw
.config
& INTEL_ARCH_EVENT_MASK
) ==
2467 X86_CONFIG(.event
=0xc0, .umask
=0x01)) {
2475 PMU_FORMAT_ATTR(event
, "config:0-7" );
2476 PMU_FORMAT_ATTR(umask
, "config:8-15" );
2477 PMU_FORMAT_ATTR(edge
, "config:18" );
2478 PMU_FORMAT_ATTR(pc
, "config:19" );
2479 PMU_FORMAT_ATTR(any
, "config:21" ); /* v3 + */
2480 PMU_FORMAT_ATTR(inv
, "config:23" );
2481 PMU_FORMAT_ATTR(cmask
, "config:24-31" );
2482 PMU_FORMAT_ATTR(in_tx
, "config:32");
2483 PMU_FORMAT_ATTR(in_tx_cp
, "config:33");
2485 static struct attribute
*intel_arch_formats_attr
[] = {
2486 &format_attr_event
.attr
,
2487 &format_attr_umask
.attr
,
2488 &format_attr_edge
.attr
,
2489 &format_attr_pc
.attr
,
2490 &format_attr_inv
.attr
,
2491 &format_attr_cmask
.attr
,
2495 ssize_t
intel_event_sysfs_show(char *page
, u64 config
)
2497 u64 event
= (config
& ARCH_PERFMON_EVENTSEL_EVENT
);
2499 return x86_event_sysfs_show(page
, config
, event
);
2502 struct intel_shared_regs
*allocate_shared_regs(int cpu
)
2504 struct intel_shared_regs
*regs
;
2507 regs
= kzalloc_node(sizeof(struct intel_shared_regs
),
2508 GFP_KERNEL
, cpu_to_node(cpu
));
2511 * initialize the locks to keep lockdep happy
2513 for (i
= 0; i
< EXTRA_REG_MAX
; i
++)
2514 raw_spin_lock_init(®s
->regs
[i
].lock
);
2521 static struct intel_excl_cntrs
*allocate_excl_cntrs(int cpu
)
2523 struct intel_excl_cntrs
*c
;
2525 c
= kzalloc_node(sizeof(struct intel_excl_cntrs
),
2526 GFP_KERNEL
, cpu_to_node(cpu
));
2528 raw_spin_lock_init(&c
->lock
);
2534 static int intel_pmu_cpu_prepare(int cpu
)
2536 struct cpu_hw_events
*cpuc
= &per_cpu(cpu_hw_events
, cpu
);
2538 if (x86_pmu
.extra_regs
|| x86_pmu
.lbr_sel_map
) {
2539 cpuc
->shared_regs
= allocate_shared_regs(cpu
);
2540 if (!cpuc
->shared_regs
)
2544 if (x86_pmu
.flags
& PMU_FL_EXCL_CNTRS
) {
2545 size_t sz
= X86_PMC_IDX_MAX
* sizeof(struct event_constraint
);
2547 cpuc
->constraint_list
= kzalloc(sz
, GFP_KERNEL
);
2548 if (!cpuc
->constraint_list
)
2551 cpuc
->excl_cntrs
= allocate_excl_cntrs(cpu
);
2552 if (!cpuc
->excl_cntrs
) {
2553 kfree(cpuc
->constraint_list
);
2554 kfree(cpuc
->shared_regs
);
2557 cpuc
->excl_thread_id
= 0;
2563 static void intel_pmu_cpu_starting(int cpu
)
2565 struct cpu_hw_events
*cpuc
= &per_cpu(cpu_hw_events
, cpu
);
2566 int core_id
= topology_core_id(cpu
);
2569 init_debug_store_on_cpu(cpu
);
2571 * Deal with CPUs that don't clear their LBRs on power-up.
2573 intel_pmu_lbr_reset();
2575 cpuc
->lbr_sel
= NULL
;
2577 if (!cpuc
->shared_regs
)
2580 if (!(x86_pmu
.flags
& PMU_FL_NO_HT_SHARING
)) {
2581 void **onln
= &cpuc
->kfree_on_online
[X86_PERF_KFREE_SHARED
];
2583 for_each_cpu(i
, topology_sibling_cpumask(cpu
)) {
2584 struct intel_shared_regs
*pc
;
2586 pc
= per_cpu(cpu_hw_events
, i
).shared_regs
;
2587 if (pc
&& pc
->core_id
== core_id
) {
2588 *onln
= cpuc
->shared_regs
;
2589 cpuc
->shared_regs
= pc
;
2593 cpuc
->shared_regs
->core_id
= core_id
;
2594 cpuc
->shared_regs
->refcnt
++;
2597 if (x86_pmu
.lbr_sel_map
)
2598 cpuc
->lbr_sel
= &cpuc
->shared_regs
->regs
[EXTRA_REG_LBR
];
2600 if (x86_pmu
.flags
& PMU_FL_EXCL_CNTRS
) {
2601 for_each_cpu(i
, topology_sibling_cpumask(cpu
)) {
2602 struct intel_excl_cntrs
*c
;
2604 c
= per_cpu(cpu_hw_events
, i
).excl_cntrs
;
2605 if (c
&& c
->core_id
== core_id
) {
2606 cpuc
->kfree_on_online
[1] = cpuc
->excl_cntrs
;
2607 cpuc
->excl_cntrs
= c
;
2608 cpuc
->excl_thread_id
= 1;
2612 cpuc
->excl_cntrs
->core_id
= core_id
;
2613 cpuc
->excl_cntrs
->refcnt
++;
2617 static void free_excl_cntrs(int cpu
)
2619 struct cpu_hw_events
*cpuc
= &per_cpu(cpu_hw_events
, cpu
);
2620 struct intel_excl_cntrs
*c
;
2622 c
= cpuc
->excl_cntrs
;
2624 if (c
->core_id
== -1 || --c
->refcnt
== 0)
2626 cpuc
->excl_cntrs
= NULL
;
2627 kfree(cpuc
->constraint_list
);
2628 cpuc
->constraint_list
= NULL
;
2632 static void intel_pmu_cpu_dying(int cpu
)
2634 struct cpu_hw_events
*cpuc
= &per_cpu(cpu_hw_events
, cpu
);
2635 struct intel_shared_regs
*pc
;
2637 pc
= cpuc
->shared_regs
;
2639 if (pc
->core_id
== -1 || --pc
->refcnt
== 0)
2641 cpuc
->shared_regs
= NULL
;
2644 free_excl_cntrs(cpu
);
2646 fini_debug_store_on_cpu(cpu
);
2649 static void intel_pmu_sched_task(struct perf_event_context
*ctx
,
2652 if (x86_pmu
.pebs_active
)
2653 intel_pmu_pebs_sched_task(ctx
, sched_in
);
2655 intel_pmu_lbr_sched_task(ctx
, sched_in
);
2658 PMU_FORMAT_ATTR(offcore_rsp
, "config1:0-63");
2660 PMU_FORMAT_ATTR(ldlat
, "config1:0-15");
2662 static struct attribute
*intel_arch3_formats_attr
[] = {
2663 &format_attr_event
.attr
,
2664 &format_attr_umask
.attr
,
2665 &format_attr_edge
.attr
,
2666 &format_attr_pc
.attr
,
2667 &format_attr_any
.attr
,
2668 &format_attr_inv
.attr
,
2669 &format_attr_cmask
.attr
,
2670 &format_attr_in_tx
.attr
,
2671 &format_attr_in_tx_cp
.attr
,
2673 &format_attr_offcore_rsp
.attr
, /* XXX do NHM/WSM + SNB breakout */
2674 &format_attr_ldlat
.attr
, /* PEBS load latency */
2678 static __initconst
const struct x86_pmu core_pmu
= {
2680 .handle_irq
= x86_pmu_handle_irq
,
2681 .disable_all
= x86_pmu_disable_all
,
2682 .enable_all
= core_pmu_enable_all
,
2683 .enable
= core_pmu_enable_event
,
2684 .disable
= x86_pmu_disable_event
,
2685 .hw_config
= x86_pmu_hw_config
,
2686 .schedule_events
= x86_schedule_events
,
2687 .eventsel
= MSR_ARCH_PERFMON_EVENTSEL0
,
2688 .perfctr
= MSR_ARCH_PERFMON_PERFCTR0
,
2689 .event_map
= intel_pmu_event_map
,
2690 .max_events
= ARRAY_SIZE(intel_perfmon_event_map
),
2693 * Intel PMCs cannot be accessed sanely above 32-bit width,
2694 * so we install an artificial 1<<31 period regardless of
2695 * the generic event period:
2697 .max_period
= (1ULL<<31) - 1,
2698 .get_event_constraints
= intel_get_event_constraints
,
2699 .put_event_constraints
= intel_put_event_constraints
,
2700 .event_constraints
= intel_core_event_constraints
,
2701 .guest_get_msrs
= core_guest_get_msrs
,
2702 .format_attrs
= intel_arch_formats_attr
,
2703 .events_sysfs_show
= intel_event_sysfs_show
,
2706 * Virtual (or funny metal) CPU can define x86_pmu.extra_regs
2707 * together with PMU version 1 and thus be using core_pmu with
2708 * shared_regs. We need following callbacks here to allocate
2711 .cpu_prepare
= intel_pmu_cpu_prepare
,
2712 .cpu_starting
= intel_pmu_cpu_starting
,
2713 .cpu_dying
= intel_pmu_cpu_dying
,
2716 static __initconst
const struct x86_pmu intel_pmu
= {
2718 .handle_irq
= intel_pmu_handle_irq
,
2719 .disable_all
= intel_pmu_disable_all
,
2720 .enable_all
= intel_pmu_enable_all
,
2721 .enable
= intel_pmu_enable_event
,
2722 .disable
= intel_pmu_disable_event
,
2723 .hw_config
= intel_pmu_hw_config
,
2724 .schedule_events
= x86_schedule_events
,
2725 .eventsel
= MSR_ARCH_PERFMON_EVENTSEL0
,
2726 .perfctr
= MSR_ARCH_PERFMON_PERFCTR0
,
2727 .event_map
= intel_pmu_event_map
,
2728 .max_events
= ARRAY_SIZE(intel_perfmon_event_map
),
2731 * Intel PMCs cannot be accessed sanely above 32 bit width,
2732 * so we install an artificial 1<<31 period regardless of
2733 * the generic event period:
2735 .max_period
= (1ULL << 31) - 1,
2736 .get_event_constraints
= intel_get_event_constraints
,
2737 .put_event_constraints
= intel_put_event_constraints
,
2738 .pebs_aliases
= intel_pebs_aliases_core2
,
2740 .format_attrs
= intel_arch3_formats_attr
,
2741 .events_sysfs_show
= intel_event_sysfs_show
,
2743 .cpu_prepare
= intel_pmu_cpu_prepare
,
2744 .cpu_starting
= intel_pmu_cpu_starting
,
2745 .cpu_dying
= intel_pmu_cpu_dying
,
2746 .guest_get_msrs
= intel_guest_get_msrs
,
2747 .sched_task
= intel_pmu_sched_task
,
2750 static __init
void intel_clovertown_quirk(void)
2753 * PEBS is unreliable due to:
2755 * AJ67 - PEBS may experience CPL leaks
2756 * AJ68 - PEBS PMI may be delayed by one event
2757 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
2758 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
2760 * AJ67 could be worked around by restricting the OS/USR flags.
2761 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
2763 * AJ106 could possibly be worked around by not allowing LBR
2764 * usage from PEBS, including the fixup.
2765 * AJ68 could possibly be worked around by always programming
2766 * a pebs_event_reset[0] value and coping with the lost events.
2768 * But taken together it might just make sense to not enable PEBS on
2771 pr_warn("PEBS disabled due to CPU errata\n");
2773 x86_pmu
.pebs_constraints
= NULL
;
2776 static int intel_snb_pebs_broken(int cpu
)
2778 u32 rev
= UINT_MAX
; /* default to broken for unknown models */
2780 switch (cpu_data(cpu
).x86_model
) {
2785 case 45: /* SNB-EP */
2786 switch (cpu_data(cpu
).x86_mask
) {
2787 case 6: rev
= 0x618; break;
2788 case 7: rev
= 0x70c; break;
2792 return (cpu_data(cpu
).microcode
< rev
);
2795 static void intel_snb_check_microcode(void)
2797 int pebs_broken
= 0;
2801 for_each_online_cpu(cpu
) {
2802 if ((pebs_broken
= intel_snb_pebs_broken(cpu
)))
2807 if (pebs_broken
== x86_pmu
.pebs_broken
)
2811 * Serialized by the microcode lock..
2813 if (x86_pmu
.pebs_broken
) {
2814 pr_info("PEBS enabled due to microcode update\n");
2815 x86_pmu
.pebs_broken
= 0;
2817 pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n");
2818 x86_pmu
.pebs_broken
= 1;
2823 * Under certain circumstances, access certain MSR may cause #GP.
2824 * The function tests if the input MSR can be safely accessed.
2826 static bool check_msr(unsigned long msr
, u64 mask
)
2828 u64 val_old
, val_new
, val_tmp
;
2831 * Read the current value, change it and read it back to see if it
2832 * matches, this is needed to detect certain hardware emulators
2833 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
2835 if (rdmsrl_safe(msr
, &val_old
))
2839 * Only change the bits which can be updated by wrmsrl.
2841 val_tmp
= val_old
^ mask
;
2842 if (wrmsrl_safe(msr
, val_tmp
) ||
2843 rdmsrl_safe(msr
, &val_new
))
2846 if (val_new
!= val_tmp
)
2849 /* Here it's sure that the MSR can be safely accessed.
2850 * Restore the old value and return.
2852 wrmsrl(msr
, val_old
);
2857 static __init
void intel_sandybridge_quirk(void)
2859 x86_pmu
.check_microcode
= intel_snb_check_microcode
;
2860 intel_snb_check_microcode();
2863 static const struct { int id
; char *name
; } intel_arch_events_map
[] __initconst
= {
2864 { PERF_COUNT_HW_CPU_CYCLES
, "cpu cycles" },
2865 { PERF_COUNT_HW_INSTRUCTIONS
, "instructions" },
2866 { PERF_COUNT_HW_BUS_CYCLES
, "bus cycles" },
2867 { PERF_COUNT_HW_CACHE_REFERENCES
, "cache references" },
2868 { PERF_COUNT_HW_CACHE_MISSES
, "cache misses" },
2869 { PERF_COUNT_HW_BRANCH_INSTRUCTIONS
, "branch instructions" },
2870 { PERF_COUNT_HW_BRANCH_MISSES
, "branch misses" },
2873 static __init
void intel_arch_events_quirk(void)
2877 /* disable event that reported as not presend by cpuid */
2878 for_each_set_bit(bit
, x86_pmu
.events_mask
, ARRAY_SIZE(intel_arch_events_map
)) {
2879 intel_perfmon_event_map
[intel_arch_events_map
[bit
].id
] = 0;
2880 pr_warn("CPUID marked event: \'%s\' unavailable\n",
2881 intel_arch_events_map
[bit
].name
);
2885 static __init
void intel_nehalem_quirk(void)
2887 union cpuid10_ebx ebx
;
2889 ebx
.full
= x86_pmu
.events_maskl
;
2890 if (ebx
.split
.no_branch_misses_retired
) {
2892 * Erratum AAJ80 detected, we work it around by using
2893 * the BR_MISP_EXEC.ANY event. This will over-count
2894 * branch-misses, but it's still much better than the
2895 * architectural event which is often completely bogus:
2897 intel_perfmon_event_map
[PERF_COUNT_HW_BRANCH_MISSES
] = 0x7f89;
2898 ebx
.split
.no_branch_misses_retired
= 0;
2899 x86_pmu
.events_maskl
= ebx
.full
;
2900 pr_info("CPU erratum AAJ80 worked around\n");
2905 * enable software workaround for errata:
2910 * Only needed when HT is enabled. However detecting
2911 * if HT is enabled is difficult (model specific). So instead,
2912 * we enable the workaround in the early boot, and verify if
2913 * it is needed in a later initcall phase once we have valid
2914 * topology information to check if HT is actually enabled
2916 static __init
void intel_ht_bug(void)
2918 x86_pmu
.flags
|= PMU_FL_EXCL_CNTRS
| PMU_FL_EXCL_ENABLED
;
2920 x86_pmu
.start_scheduling
= intel_start_scheduling
;
2921 x86_pmu
.commit_scheduling
= intel_commit_scheduling
;
2922 x86_pmu
.stop_scheduling
= intel_stop_scheduling
;
2925 EVENT_ATTR_STR(mem
-loads
, mem_ld_hsw
, "event=0xcd,umask=0x1,ldlat=3");
2926 EVENT_ATTR_STR(mem
-stores
, mem_st_hsw
, "event=0xd0,umask=0x82")
2928 /* Haswell special events */
2929 EVENT_ATTR_STR(tx
-start
, tx_start
, "event=0xc9,umask=0x1");
2930 EVENT_ATTR_STR(tx
-commit
, tx_commit
, "event=0xc9,umask=0x2");
2931 EVENT_ATTR_STR(tx
-abort
, tx_abort
, "event=0xc9,umask=0x4");
2932 EVENT_ATTR_STR(tx
-capacity
, tx_capacity
, "event=0x54,umask=0x2");
2933 EVENT_ATTR_STR(tx
-conflict
, tx_conflict
, "event=0x54,umask=0x1");
2934 EVENT_ATTR_STR(el
-start
, el_start
, "event=0xc8,umask=0x1");
2935 EVENT_ATTR_STR(el
-commit
, el_commit
, "event=0xc8,umask=0x2");
2936 EVENT_ATTR_STR(el
-abort
, el_abort
, "event=0xc8,umask=0x4");
2937 EVENT_ATTR_STR(el
-capacity
, el_capacity
, "event=0x54,umask=0x2");
2938 EVENT_ATTR_STR(el
-conflict
, el_conflict
, "event=0x54,umask=0x1");
2939 EVENT_ATTR_STR(cycles
-t
, cycles_t
, "event=0x3c,in_tx=1");
2940 EVENT_ATTR_STR(cycles
-ct
, cycles_ct
, "event=0x3c,in_tx=1,in_tx_cp=1");
2942 static struct attribute
*hsw_events_attrs
[] = {
2943 EVENT_PTR(tx_start
),
2944 EVENT_PTR(tx_commit
),
2945 EVENT_PTR(tx_abort
),
2946 EVENT_PTR(tx_capacity
),
2947 EVENT_PTR(tx_conflict
),
2948 EVENT_PTR(el_start
),
2949 EVENT_PTR(el_commit
),
2950 EVENT_PTR(el_abort
),
2951 EVENT_PTR(el_capacity
),
2952 EVENT_PTR(el_conflict
),
2953 EVENT_PTR(cycles_t
),
2954 EVENT_PTR(cycles_ct
),
2955 EVENT_PTR(mem_ld_hsw
),
2956 EVENT_PTR(mem_st_hsw
),
2960 __init
int intel_pmu_init(void)
2962 union cpuid10_edx edx
;
2963 union cpuid10_eax eax
;
2964 union cpuid10_ebx ebx
;
2965 struct event_constraint
*c
;
2966 unsigned int unused
;
2967 struct extra_reg
*er
;
2970 if (!cpu_has(&boot_cpu_data
, X86_FEATURE_ARCH_PERFMON
)) {
2971 switch (boot_cpu_data
.x86
) {
2973 return p6_pmu_init();
2975 return knc_pmu_init();
2977 return p4_pmu_init();
2983 * Check whether the Architectural PerfMon supports
2984 * Branch Misses Retired hw_event or not.
2986 cpuid(10, &eax
.full
, &ebx
.full
, &unused
, &edx
.full
);
2987 if (eax
.split
.mask_length
< ARCH_PERFMON_EVENTS_COUNT
)
2990 version
= eax
.split
.version_id
;
2994 x86_pmu
= intel_pmu
;
2996 x86_pmu
.version
= version
;
2997 x86_pmu
.num_counters
= eax
.split
.num_counters
;
2998 x86_pmu
.cntval_bits
= eax
.split
.bit_width
;
2999 x86_pmu
.cntval_mask
= (1ULL << eax
.split
.bit_width
) - 1;
3001 x86_pmu
.events_maskl
= ebx
.full
;
3002 x86_pmu
.events_mask_len
= eax
.split
.mask_length
;
3004 x86_pmu
.max_pebs_events
= min_t(unsigned, MAX_PEBS_EVENTS
, x86_pmu
.num_counters
);
3007 * Quirk: v2 perfmon does not report fixed-purpose events, so
3008 * assume at least 3 events:
3011 x86_pmu
.num_counters_fixed
= max((int)edx
.split
.num_counters_fixed
, 3);
3013 if (boot_cpu_has(X86_FEATURE_PDCM
)) {
3016 rdmsrl(MSR_IA32_PERF_CAPABILITIES
, capabilities
);
3017 x86_pmu
.intel_cap
.capabilities
= capabilities
;
3022 x86_add_quirk(intel_arch_events_quirk
); /* Install first, so it runs last */
3025 * Install the hw-cache-events table:
3027 switch (boot_cpu_data
.x86_model
) {
3028 case 14: /* 65nm Core "Yonah" */
3029 pr_cont("Core events, ");
3032 case 15: /* 65nm Core2 "Merom" */
3033 x86_add_quirk(intel_clovertown_quirk
);
3034 case 22: /* 65nm Core2 "Merom-L" */
3035 case 23: /* 45nm Core2 "Penryn" */
3036 case 29: /* 45nm Core2 "Dunnington (MP) */
3037 memcpy(hw_cache_event_ids
, core2_hw_cache_event_ids
,
3038 sizeof(hw_cache_event_ids
));
3040 intel_pmu_lbr_init_core();
3042 x86_pmu
.event_constraints
= intel_core2_event_constraints
;
3043 x86_pmu
.pebs_constraints
= intel_core2_pebs_event_constraints
;
3044 pr_cont("Core2 events, ");
3047 case 30: /* 45nm Nehalem */
3048 case 26: /* 45nm Nehalem-EP */
3049 case 46: /* 45nm Nehalem-EX */
3050 memcpy(hw_cache_event_ids
, nehalem_hw_cache_event_ids
,
3051 sizeof(hw_cache_event_ids
));
3052 memcpy(hw_cache_extra_regs
, nehalem_hw_cache_extra_regs
,
3053 sizeof(hw_cache_extra_regs
));
3055 intel_pmu_lbr_init_nhm();
3057 x86_pmu
.event_constraints
= intel_nehalem_event_constraints
;
3058 x86_pmu
.pebs_constraints
= intel_nehalem_pebs_event_constraints
;
3059 x86_pmu
.enable_all
= intel_pmu_nhm_enable_all
;
3060 x86_pmu
.extra_regs
= intel_nehalem_extra_regs
;
3062 x86_pmu
.cpu_events
= nhm_events_attrs
;
3064 /* UOPS_ISSUED.STALLED_CYCLES */
3065 intel_perfmon_event_map
[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
] =
3066 X86_CONFIG(.event
=0x0e, .umask
=0x01, .inv
=1, .cmask
=1);
3067 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
3068 intel_perfmon_event_map
[PERF_COUNT_HW_STALLED_CYCLES_BACKEND
] =
3069 X86_CONFIG(.event
=0xb1, .umask
=0x3f, .inv
=1, .cmask
=1);
3071 x86_add_quirk(intel_nehalem_quirk
);
3073 pr_cont("Nehalem events, ");
3076 case 28: /* 45nm Atom "Pineview" */
3077 case 38: /* 45nm Atom "Lincroft" */
3078 case 39: /* 32nm Atom "Penwell" */
3079 case 53: /* 32nm Atom "Cloverview" */
3080 case 54: /* 32nm Atom "Cedarview" */
3081 memcpy(hw_cache_event_ids
, atom_hw_cache_event_ids
,
3082 sizeof(hw_cache_event_ids
));
3084 intel_pmu_lbr_init_atom();
3086 x86_pmu
.event_constraints
= intel_gen_event_constraints
;
3087 x86_pmu
.pebs_constraints
= intel_atom_pebs_event_constraints
;
3088 pr_cont("Atom events, ");
3091 case 55: /* 22nm Atom "Silvermont" */
3092 case 76: /* 14nm Atom "Airmont" */
3093 case 77: /* 22nm Atom "Silvermont Avoton/Rangely" */
3094 memcpy(hw_cache_event_ids
, slm_hw_cache_event_ids
,
3095 sizeof(hw_cache_event_ids
));
3096 memcpy(hw_cache_extra_regs
, slm_hw_cache_extra_regs
,
3097 sizeof(hw_cache_extra_regs
));
3099 intel_pmu_lbr_init_atom();
3101 x86_pmu
.event_constraints
= intel_slm_event_constraints
;
3102 x86_pmu
.pebs_constraints
= intel_slm_pebs_event_constraints
;
3103 x86_pmu
.extra_regs
= intel_slm_extra_regs
;
3104 x86_pmu
.flags
|= PMU_FL_HAS_RSP_1
;
3105 pr_cont("Silvermont events, ");
3108 case 37: /* 32nm Westmere */
3109 case 44: /* 32nm Westmere-EP */
3110 case 47: /* 32nm Westmere-EX */
3111 memcpy(hw_cache_event_ids
, westmere_hw_cache_event_ids
,
3112 sizeof(hw_cache_event_ids
));
3113 memcpy(hw_cache_extra_regs
, nehalem_hw_cache_extra_regs
,
3114 sizeof(hw_cache_extra_regs
));
3116 intel_pmu_lbr_init_nhm();
3118 x86_pmu
.event_constraints
= intel_westmere_event_constraints
;
3119 x86_pmu
.enable_all
= intel_pmu_nhm_enable_all
;
3120 x86_pmu
.pebs_constraints
= intel_westmere_pebs_event_constraints
;
3121 x86_pmu
.extra_regs
= intel_westmere_extra_regs
;
3122 x86_pmu
.flags
|= PMU_FL_HAS_RSP_1
;
3124 x86_pmu
.cpu_events
= nhm_events_attrs
;
3126 /* UOPS_ISSUED.STALLED_CYCLES */
3127 intel_perfmon_event_map
[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
] =
3128 X86_CONFIG(.event
=0x0e, .umask
=0x01, .inv
=1, .cmask
=1);
3129 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
3130 intel_perfmon_event_map
[PERF_COUNT_HW_STALLED_CYCLES_BACKEND
] =
3131 X86_CONFIG(.event
=0xb1, .umask
=0x3f, .inv
=1, .cmask
=1);
3133 pr_cont("Westmere events, ");
3136 case 42: /* 32nm SandyBridge */
3137 case 45: /* 32nm SandyBridge-E/EN/EP */
3138 x86_add_quirk(intel_sandybridge_quirk
);
3139 x86_add_quirk(intel_ht_bug
);
3140 memcpy(hw_cache_event_ids
, snb_hw_cache_event_ids
,
3141 sizeof(hw_cache_event_ids
));
3142 memcpy(hw_cache_extra_regs
, snb_hw_cache_extra_regs
,
3143 sizeof(hw_cache_extra_regs
));
3145 intel_pmu_lbr_init_snb();
3147 x86_pmu
.event_constraints
= intel_snb_event_constraints
;
3148 x86_pmu
.pebs_constraints
= intel_snb_pebs_event_constraints
;
3149 x86_pmu
.pebs_aliases
= intel_pebs_aliases_snb
;
3150 if (boot_cpu_data
.x86_model
== 45)
3151 x86_pmu
.extra_regs
= intel_snbep_extra_regs
;
3153 x86_pmu
.extra_regs
= intel_snb_extra_regs
;
3156 /* all extra regs are per-cpu when HT is on */
3157 x86_pmu
.flags
|= PMU_FL_HAS_RSP_1
;
3158 x86_pmu
.flags
|= PMU_FL_NO_HT_SHARING
;
3160 x86_pmu
.cpu_events
= snb_events_attrs
;
3162 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
3163 intel_perfmon_event_map
[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
] =
3164 X86_CONFIG(.event
=0x0e, .umask
=0x01, .inv
=1, .cmask
=1);
3165 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
3166 intel_perfmon_event_map
[PERF_COUNT_HW_STALLED_CYCLES_BACKEND
] =
3167 X86_CONFIG(.event
=0xb1, .umask
=0x01, .inv
=1, .cmask
=1);
3169 pr_cont("SandyBridge events, ");
3172 case 58: /* 22nm IvyBridge */
3173 case 62: /* 22nm IvyBridge-EP/EX */
3174 x86_add_quirk(intel_ht_bug
);
3175 memcpy(hw_cache_event_ids
, snb_hw_cache_event_ids
,
3176 sizeof(hw_cache_event_ids
));
3177 /* dTLB-load-misses on IVB is different than SNB */
3178 hw_cache_event_ids
[C(DTLB
)][C(OP_READ
)][C(RESULT_MISS
)] = 0x8108; /* DTLB_LOAD_MISSES.DEMAND_LD_MISS_CAUSES_A_WALK */
3180 memcpy(hw_cache_extra_regs
, snb_hw_cache_extra_regs
,
3181 sizeof(hw_cache_extra_regs
));
3183 intel_pmu_lbr_init_snb();
3185 x86_pmu
.event_constraints
= intel_ivb_event_constraints
;
3186 x86_pmu
.pebs_constraints
= intel_ivb_pebs_event_constraints
;
3187 x86_pmu
.pebs_aliases
= intel_pebs_aliases_snb
;
3188 if (boot_cpu_data
.x86_model
== 62)
3189 x86_pmu
.extra_regs
= intel_snbep_extra_regs
;
3191 x86_pmu
.extra_regs
= intel_snb_extra_regs
;
3192 /* all extra regs are per-cpu when HT is on */
3193 x86_pmu
.flags
|= PMU_FL_HAS_RSP_1
;
3194 x86_pmu
.flags
|= PMU_FL_NO_HT_SHARING
;
3196 x86_pmu
.cpu_events
= snb_events_attrs
;
3198 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
3199 intel_perfmon_event_map
[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
] =
3200 X86_CONFIG(.event
=0x0e, .umask
=0x01, .inv
=1, .cmask
=1);
3202 pr_cont("IvyBridge events, ");
3206 case 60: /* 22nm Haswell Core */
3207 case 63: /* 22nm Haswell Server */
3208 case 69: /* 22nm Haswell ULT */
3209 case 70: /* 22nm Haswell + GT3e (Intel Iris Pro graphics) */
3210 x86_add_quirk(intel_ht_bug
);
3211 x86_pmu
.late_ack
= true;
3212 memcpy(hw_cache_event_ids
, hsw_hw_cache_event_ids
, sizeof(hw_cache_event_ids
));
3213 memcpy(hw_cache_extra_regs
, hsw_hw_cache_extra_regs
, sizeof(hw_cache_extra_regs
));
3215 intel_pmu_lbr_init_hsw();
3217 x86_pmu
.event_constraints
= intel_hsw_event_constraints
;
3218 x86_pmu
.pebs_constraints
= intel_hsw_pebs_event_constraints
;
3219 x86_pmu
.extra_regs
= intel_snbep_extra_regs
;
3220 x86_pmu
.pebs_aliases
= intel_pebs_aliases_snb
;
3221 /* all extra regs are per-cpu when HT is on */
3222 x86_pmu
.flags
|= PMU_FL_HAS_RSP_1
;
3223 x86_pmu
.flags
|= PMU_FL_NO_HT_SHARING
;
3225 x86_pmu
.hw_config
= hsw_hw_config
;
3226 x86_pmu
.get_event_constraints
= hsw_get_event_constraints
;
3227 x86_pmu
.cpu_events
= hsw_events_attrs
;
3228 x86_pmu
.lbr_double_abort
= true;
3229 pr_cont("Haswell events, ");
3232 case 61: /* 14nm Broadwell Core-M */
3233 case 86: /* 14nm Broadwell Xeon D */
3234 case 71: /* 14nm Broadwell + GT3e (Intel Iris Pro graphics) */
3235 case 79: /* 14nm Broadwell Server */
3236 x86_pmu
.late_ack
= true;
3237 memcpy(hw_cache_event_ids
, hsw_hw_cache_event_ids
, sizeof(hw_cache_event_ids
));
3238 memcpy(hw_cache_extra_regs
, hsw_hw_cache_extra_regs
, sizeof(hw_cache_extra_regs
));
3240 /* L3_MISS_LOCAL_DRAM is BIT(26) in Broadwell */
3241 hw_cache_extra_regs
[C(LL
)][C(OP_READ
)][C(RESULT_MISS
)] = HSW_DEMAND_READ
|
3242 BDW_L3_MISS
|HSW_SNOOP_DRAM
;
3243 hw_cache_extra_regs
[C(LL
)][C(OP_WRITE
)][C(RESULT_MISS
)] = HSW_DEMAND_WRITE
|BDW_L3_MISS
|
3245 hw_cache_extra_regs
[C(NODE
)][C(OP_READ
)][C(RESULT_ACCESS
)] = HSW_DEMAND_READ
|
3246 BDW_L3_MISS_LOCAL
|HSW_SNOOP_DRAM
;
3247 hw_cache_extra_regs
[C(NODE
)][C(OP_WRITE
)][C(RESULT_ACCESS
)] = HSW_DEMAND_WRITE
|
3248 BDW_L3_MISS_LOCAL
|HSW_SNOOP_DRAM
;
3250 intel_pmu_lbr_init_hsw();
3252 x86_pmu
.event_constraints
= intel_bdw_event_constraints
;
3253 x86_pmu
.pebs_constraints
= intel_hsw_pebs_event_constraints
;
3254 x86_pmu
.extra_regs
= intel_snbep_extra_regs
;
3255 x86_pmu
.pebs_aliases
= intel_pebs_aliases_snb
;
3256 /* all extra regs are per-cpu when HT is on */
3257 x86_pmu
.flags
|= PMU_FL_HAS_RSP_1
;
3258 x86_pmu
.flags
|= PMU_FL_NO_HT_SHARING
;
3260 x86_pmu
.hw_config
= hsw_hw_config
;
3261 x86_pmu
.get_event_constraints
= hsw_get_event_constraints
;
3262 x86_pmu
.cpu_events
= hsw_events_attrs
;
3263 x86_pmu
.limit_period
= bdw_limit_period
;
3264 pr_cont("Broadwell events, ");
3268 switch (x86_pmu
.version
) {
3270 x86_pmu
.event_constraints
= intel_v1_event_constraints
;
3271 pr_cont("generic architected perfmon v1, ");
3275 * default constraints for v2 and up
3277 x86_pmu
.event_constraints
= intel_gen_event_constraints
;
3278 pr_cont("generic architected perfmon, ");
3283 if (x86_pmu
.num_counters
> INTEL_PMC_MAX_GENERIC
) {
3284 WARN(1, KERN_ERR
"hw perf events %d > max(%d), clipping!",
3285 x86_pmu
.num_counters
, INTEL_PMC_MAX_GENERIC
);
3286 x86_pmu
.num_counters
= INTEL_PMC_MAX_GENERIC
;
3288 x86_pmu
.intel_ctrl
= (1 << x86_pmu
.num_counters
) - 1;
3290 if (x86_pmu
.num_counters_fixed
> INTEL_PMC_MAX_FIXED
) {
3291 WARN(1, KERN_ERR
"hw perf events fixed %d > max(%d), clipping!",
3292 x86_pmu
.num_counters_fixed
, INTEL_PMC_MAX_FIXED
);
3293 x86_pmu
.num_counters_fixed
= INTEL_PMC_MAX_FIXED
;
3296 x86_pmu
.intel_ctrl
|=
3297 ((1LL << x86_pmu
.num_counters_fixed
)-1) << INTEL_PMC_IDX_FIXED
;
3299 if (x86_pmu
.event_constraints
) {
3301 * event on fixed counter2 (REF_CYCLES) only works on this
3302 * counter, so do not extend mask to generic counters
3304 for_each_event_constraint(c
, x86_pmu
.event_constraints
) {
3305 if (c
->cmask
== FIXED_EVENT_FLAGS
3306 && c
->idxmsk64
!= INTEL_PMC_MSK_FIXED_REF_CYCLES
) {
3307 c
->idxmsk64
|= (1ULL << x86_pmu
.num_counters
) - 1;
3310 ~(~0UL << (INTEL_PMC_IDX_FIXED
+ x86_pmu
.num_counters_fixed
));
3311 c
->weight
= hweight64(c
->idxmsk64
);
3316 * Access LBR MSR may cause #GP under certain circumstances.
3317 * E.g. KVM doesn't support LBR MSR
3318 * Check all LBT MSR here.
3319 * Disable LBR access if any LBR MSRs can not be accessed.
3321 if (x86_pmu
.lbr_nr
&& !check_msr(x86_pmu
.lbr_tos
, 0x3UL
))
3323 for (i
= 0; i
< x86_pmu
.lbr_nr
; i
++) {
3324 if (!(check_msr(x86_pmu
.lbr_from
+ i
, 0xffffUL
) &&
3325 check_msr(x86_pmu
.lbr_to
+ i
, 0xffffUL
)))
3330 * Access extra MSR may cause #GP under certain circumstances.
3331 * E.g. KVM doesn't support offcore event
3332 * Check all extra_regs here.
3334 if (x86_pmu
.extra_regs
) {
3335 for (er
= x86_pmu
.extra_regs
; er
->msr
; er
++) {
3336 er
->extra_msr_access
= check_msr(er
->msr
, 0x1ffUL
);
3337 /* Disable LBR select mapping */
3338 if ((er
->idx
== EXTRA_REG_LBR
) && !er
->extra_msr_access
)
3339 x86_pmu
.lbr_sel_map
= NULL
;
3343 /* Support full width counters using alternative MSR range */
3344 if (x86_pmu
.intel_cap
.full_width_write
) {
3345 x86_pmu
.max_period
= x86_pmu
.cntval_mask
;
3346 x86_pmu
.perfctr
= MSR_IA32_PMC0
;
3347 pr_cont("full-width counters, ");
3354 * HT bug: phase 2 init
3355 * Called once we have valid topology information to check
3356 * whether or not HT is enabled
3357 * If HT is off, then we disable the workaround
3359 static __init
int fixup_ht_bug(void)
3361 int cpu
= smp_processor_id();
3364 * problem not present on this CPU model, nothing to do
3366 if (!(x86_pmu
.flags
& PMU_FL_EXCL_ENABLED
))
3369 w
= cpumask_weight(topology_sibling_cpumask(cpu
));
3371 pr_info("PMU erratum BJ122, BV98, HSD29 worked around, HT is on\n");
3375 watchdog_nmi_disable_all();
3377 x86_pmu
.flags
&= ~(PMU_FL_EXCL_CNTRS
| PMU_FL_EXCL_ENABLED
);
3379 x86_pmu
.start_scheduling
= NULL
;
3380 x86_pmu
.commit_scheduling
= NULL
;
3381 x86_pmu
.stop_scheduling
= NULL
;
3383 watchdog_nmi_enable_all();
3387 for_each_online_cpu(c
) {
3392 pr_info("PMU erratum BJ122, BV98, HSD29 workaround disabled, HT off\n");
3395 subsys_initcall(fixup_ht_bug
)