1 // SPDX-License-Identifier: GPL-2.0-only
5 * Used to coordinate shared registers between HT threads or
6 * among events on a single PMU.
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/stddef.h>
12 #include <linux/types.h>
13 #include <linux/init.h>
14 #include <linux/slab.h>
15 #include <linux/export.h>
16 #include <linux/nmi.h>
18 #include <asm/cpufeature.h>
19 #include <asm/hardirq.h>
20 #include <asm/intel-family.h>
21 #include <asm/intel_pt.h>
23 #include <asm/cpu_device_id.h>
25 #include "../perf_event.h"
28 * Intel PerfMon, used on Core and later.
30 static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly =
32 [PERF_COUNT_HW_CPU_CYCLES] = 0x003c,
33 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
34 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e,
35 [PERF_COUNT_HW_CACHE_MISSES] = 0x412e,
36 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4,
37 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5,
38 [PERF_COUNT_HW_BUS_CYCLES] = 0x013c,
39 [PERF_COUNT_HW_REF_CPU_CYCLES] = 0x0300, /* pseudo-encoding */
42 static struct event_constraint intel_core_event_constraints[] __read_mostly =
44 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
45 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
46 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
47 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
48 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
49 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */
53 static struct event_constraint intel_core2_event_constraints[] __read_mostly =
55 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
56 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
57 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
58 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
59 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
60 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
61 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
62 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
63 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
64 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
65 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
66 INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */
67 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
71 static struct event_constraint intel_nehalem_event_constraints[] __read_mostly =
73 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
74 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
75 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
76 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
77 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
78 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
79 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
80 INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */
81 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
82 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
83 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
87 static struct extra_reg intel_nehalem_extra_regs[] __read_mostly =
89 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
90 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
91 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
95 static struct event_constraint intel_westmere_event_constraints[] __read_mostly =
97 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
98 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
99 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
100 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
101 INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */
102 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
103 INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */
107 static struct event_constraint intel_snb_event_constraints[] __read_mostly =
109 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
110 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
111 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
112 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
113 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
114 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
115 INTEL_UEVENT_CONSTRAINT(0x06a3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
116 INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */
117 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
118 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
119 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
120 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
123 * When HT is off these events can only run on the bottom 4 counters
124 * When HT is on, they are impacted by the HT bug and require EXCL access
126 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
127 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
128 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
129 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
134 static struct event_constraint intel_ivb_event_constraints[] __read_mostly =
136 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
137 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
138 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
139 INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */
140 INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMTPY */
141 INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */
142 INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_LDM_PENDING */
143 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
144 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
145 INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */
146 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
147 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
148 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
151 * When HT is off these events can only run on the bottom 4 counters
152 * When HT is on, they are impacted by the HT bug and require EXCL access
154 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
155 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
156 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
157 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
162 static struct extra_reg intel_westmere_extra_regs[] __read_mostly =
164 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
165 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
166 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1),
167 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
171 static struct event_constraint intel_v1_event_constraints[] __read_mostly =
176 static struct event_constraint intel_gen_event_constraints[] __read_mostly =
178 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
179 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
180 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
184 static struct event_constraint intel_slm_event_constraints[] __read_mostly =
186 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
187 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
188 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */
192 static struct event_constraint intel_skl_event_constraints[] = {
193 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
194 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
195 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
196 INTEL_UEVENT_CONSTRAINT(0x1c0, 0x2), /* INST_RETIRED.PREC_DIST */
199 * when HT is off, these can only run on the bottom 4 counters
201 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */
202 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */
203 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */
204 INTEL_EVENT_CONSTRAINT(0xcd, 0xf), /* MEM_TRANS_RETIRED.* */
205 INTEL_EVENT_CONSTRAINT(0xc6, 0xf), /* FRONTEND_RETIRED.* */
210 static struct extra_reg intel_knl_extra_regs[] __read_mostly = {
211 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x799ffbb6e7ull, RSP_0),
212 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x399ffbffe7ull, RSP_1),
216 static struct extra_reg intel_snb_extra_regs[] __read_mostly = {
217 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
218 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0),
219 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1),
220 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
224 static struct extra_reg intel_snbep_extra_regs[] __read_mostly = {
225 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
226 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
227 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
228 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
232 static struct extra_reg intel_skl_extra_regs[] __read_mostly = {
233 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
234 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
235 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
237 * Note the low 8 bits eventsel code is not a continuous field, containing
238 * some #GPing bits. These are masked out.
240 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE),
244 static struct event_constraint intel_icl_event_constraints[] = {
245 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
246 FIXED_EVENT_CONSTRAINT(0x01c0, 0), /* INST_RETIRED.PREC_DIST */
247 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
248 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
249 FIXED_EVENT_CONSTRAINT(0x0400, 3), /* SLOTS */
250 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_RETIRING, 0),
251 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BAD_SPEC, 1),
252 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FE_BOUND, 2),
253 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BE_BOUND, 3),
254 INTEL_EVENT_CONSTRAINT_RANGE(0x03, 0x0a, 0xf),
255 INTEL_EVENT_CONSTRAINT_RANGE(0x1f, 0x28, 0xf),
256 INTEL_EVENT_CONSTRAINT(0x32, 0xf), /* SW_PREFETCH_ACCESS.* */
257 INTEL_EVENT_CONSTRAINT_RANGE(0x48, 0x54, 0xf),
258 INTEL_EVENT_CONSTRAINT_RANGE(0x60, 0x8b, 0xf),
259 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xff), /* CYCLE_ACTIVITY.STALLS_TOTAL */
260 INTEL_UEVENT_CONSTRAINT(0x10a3, 0xff), /* CYCLE_ACTIVITY.CYCLES_MEM_ANY */
261 INTEL_UEVENT_CONSTRAINT(0x14a3, 0xff), /* CYCLE_ACTIVITY.STALLS_MEM_ANY */
262 INTEL_EVENT_CONSTRAINT(0xa3, 0xf), /* CYCLE_ACTIVITY.* */
263 INTEL_EVENT_CONSTRAINT_RANGE(0xa8, 0xb0, 0xf),
264 INTEL_EVENT_CONSTRAINT_RANGE(0xb7, 0xbd, 0xf),
265 INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xe6, 0xf),
266 INTEL_EVENT_CONSTRAINT_RANGE(0xf0, 0xf4, 0xf),
270 static struct extra_reg intel_icl_extra_regs[] __read_mostly = {
271 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffffbfffull, RSP_0),
272 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffffbfffull, RSP_1),
273 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
274 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE),
278 static struct extra_reg intel_spr_extra_regs[] __read_mostly = {
279 INTEL_UEVENT_EXTRA_REG(0x012a, MSR_OFFCORE_RSP_0, 0x3fffffffffull, RSP_0),
280 INTEL_UEVENT_EXTRA_REG(0x012b, MSR_OFFCORE_RSP_1, 0x3fffffffffull, RSP_1),
281 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
282 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE),
286 static struct event_constraint intel_spr_event_constraints[] = {
287 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
288 FIXED_EVENT_CONSTRAINT(0x01c0, 0), /* INST_RETIRED.PREC_DIST */
289 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
290 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
291 FIXED_EVENT_CONSTRAINT(0x0400, 3), /* SLOTS */
292 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_RETIRING, 0),
293 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BAD_SPEC, 1),
294 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FE_BOUND, 2),
295 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BE_BOUND, 3),
296 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_HEAVY_OPS, 4),
297 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BR_MISPREDICT, 5),
298 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FETCH_LAT, 6),
299 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_MEM_BOUND, 7),
301 INTEL_EVENT_CONSTRAINT(0x2e, 0xff),
302 INTEL_EVENT_CONSTRAINT(0x3c, 0xff),
304 * Generally event codes < 0x90 are restricted to counters 0-3.
305 * The 0x2E and 0x3C are exception, which has no restriction.
307 INTEL_EVENT_CONSTRAINT_RANGE(0x01, 0x8f, 0xf),
309 INTEL_UEVENT_CONSTRAINT(0x01a3, 0xf),
310 INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf),
311 INTEL_UEVENT_CONSTRAINT(0x08a3, 0xf),
312 INTEL_UEVENT_CONSTRAINT(0x04a4, 0x1),
313 INTEL_UEVENT_CONSTRAINT(0x08a4, 0x1),
314 INTEL_UEVENT_CONSTRAINT(0x02cd, 0x1),
315 INTEL_EVENT_CONSTRAINT(0xce, 0x1),
316 INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xdf, 0xf),
318 * Generally event codes >= 0x90 are likely to have no restrictions.
319 * The exception are defined as above.
321 INTEL_EVENT_CONSTRAINT_RANGE(0x90, 0xfe, 0xff),
327 EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3");
328 EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3");
329 EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2");
331 static struct attribute *nhm_mem_events_attrs[] = {
332 EVENT_PTR(mem_ld_nhm),
337 * topdown events for Intel Core CPUs.
339 * The events are all in slots, which is a free slot in a 4 wide
340 * pipeline. Some events are already reported in slots, for cycle
341 * events we multiply by the pipeline width (4).
343 * With Hyper Threading on, topdown metrics are either summed or averaged
344 * between the threads of a core: (count_t0 + count_t1).
346 * For the average case the metric is always scaled to pipeline width,
347 * so we use factor 2 ((count_t0 + count_t1) / 2 * 4)
350 EVENT_ATTR_STR_HT(topdown-total-slots, td_total_slots,
351 "event=0x3c,umask=0x0", /* cpu_clk_unhalted.thread */
352 "event=0x3c,umask=0x0,any=1"); /* cpu_clk_unhalted.thread_any */
353 EVENT_ATTR_STR_HT(topdown-total-slots.scale, td_total_slots_scale, "4", "2");
354 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued,
355 "event=0xe,umask=0x1"); /* uops_issued.any */
356 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired,
357 "event=0xc2,umask=0x2"); /* uops_retired.retire_slots */
358 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles,
359 "event=0x9c,umask=0x1"); /* idq_uops_not_delivered_core */
360 EVENT_ATTR_STR_HT(topdown-recovery-bubbles, td_recovery_bubbles,
361 "event=0xd,umask=0x3,cmask=1", /* int_misc.recovery_cycles */
362 "event=0xd,umask=0x3,cmask=1,any=1"); /* int_misc.recovery_cycles_any */
363 EVENT_ATTR_STR_HT(topdown-recovery-bubbles.scale, td_recovery_bubbles_scale,
366 EVENT_ATTR_STR(slots, slots, "event=0x00,umask=0x4");
367 EVENT_ATTR_STR(topdown-retiring, td_retiring, "event=0x00,umask=0x80");
368 EVENT_ATTR_STR(topdown-bad-spec, td_bad_spec, "event=0x00,umask=0x81");
369 EVENT_ATTR_STR(topdown-fe-bound, td_fe_bound, "event=0x00,umask=0x82");
370 EVENT_ATTR_STR(topdown-be-bound, td_be_bound, "event=0x00,umask=0x83");
371 EVENT_ATTR_STR(topdown-heavy-ops, td_heavy_ops, "event=0x00,umask=0x84");
372 EVENT_ATTR_STR(topdown-br-mispredict, td_br_mispredict, "event=0x00,umask=0x85");
373 EVENT_ATTR_STR(topdown-fetch-lat, td_fetch_lat, "event=0x00,umask=0x86");
374 EVENT_ATTR_STR(topdown-mem-bound, td_mem_bound, "event=0x00,umask=0x87");
376 static struct attribute *snb_events_attrs[] = {
377 EVENT_PTR(td_slots_issued),
378 EVENT_PTR(td_slots_retired),
379 EVENT_PTR(td_fetch_bubbles),
380 EVENT_PTR(td_total_slots),
381 EVENT_PTR(td_total_slots_scale),
382 EVENT_PTR(td_recovery_bubbles),
383 EVENT_PTR(td_recovery_bubbles_scale),
387 static struct attribute *snb_mem_events_attrs[] = {
388 EVENT_PTR(mem_ld_snb),
389 EVENT_PTR(mem_st_snb),
393 static struct event_constraint intel_hsw_event_constraints[] = {
394 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
395 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
396 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
397 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */
398 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
399 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
400 /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
401 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4),
402 /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
403 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4),
404 /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
405 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf),
408 * When HT is off these events can only run on the bottom 4 counters
409 * When HT is on, they are impacted by the HT bug and require EXCL access
411 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
412 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
413 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
414 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
419 static struct event_constraint intel_bdw_event_constraints[] = {
420 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
421 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
422 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
423 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */
424 INTEL_UBIT_EVENT_CONSTRAINT(0x8a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_MISS */
426 * when HT is off, these can only run on the bottom 4 counters
428 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */
429 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */
430 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */
431 INTEL_EVENT_CONSTRAINT(0xcd, 0xf), /* MEM_TRANS_RETIRED.* */
435 static u64 intel_pmu_event_map(int hw_event)
437 return intel_perfmon_event_map[hw_event];
440 static __initconst const u64 spr_hw_cache_event_ids
441 [PERF_COUNT_HW_CACHE_MAX]
442 [PERF_COUNT_HW_CACHE_OP_MAX]
443 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
447 [ C(RESULT_ACCESS) ] = 0x81d0,
448 [ C(RESULT_MISS) ] = 0xe124,
451 [ C(RESULT_ACCESS) ] = 0x82d0,
456 [ C(RESULT_MISS) ] = 0xe424,
459 [ C(RESULT_ACCESS) ] = -1,
460 [ C(RESULT_MISS) ] = -1,
465 [ C(RESULT_ACCESS) ] = 0x12a,
466 [ C(RESULT_MISS) ] = 0x12a,
469 [ C(RESULT_ACCESS) ] = 0x12a,
470 [ C(RESULT_MISS) ] = 0x12a,
475 [ C(RESULT_ACCESS) ] = 0x81d0,
476 [ C(RESULT_MISS) ] = 0xe12,
479 [ C(RESULT_ACCESS) ] = 0x82d0,
480 [ C(RESULT_MISS) ] = 0xe13,
485 [ C(RESULT_ACCESS) ] = -1,
486 [ C(RESULT_MISS) ] = 0xe11,
489 [ C(RESULT_ACCESS) ] = -1,
490 [ C(RESULT_MISS) ] = -1,
492 [ C(OP_PREFETCH) ] = {
493 [ C(RESULT_ACCESS) ] = -1,
494 [ C(RESULT_MISS) ] = -1,
499 [ C(RESULT_ACCESS) ] = 0x4c4,
500 [ C(RESULT_MISS) ] = 0x4c5,
503 [ C(RESULT_ACCESS) ] = -1,
504 [ C(RESULT_MISS) ] = -1,
506 [ C(OP_PREFETCH) ] = {
507 [ C(RESULT_ACCESS) ] = -1,
508 [ C(RESULT_MISS) ] = -1,
513 [ C(RESULT_ACCESS) ] = 0x12a,
514 [ C(RESULT_MISS) ] = 0x12a,
519 static __initconst const u64 spr_hw_cache_extra_regs
520 [PERF_COUNT_HW_CACHE_MAX]
521 [PERF_COUNT_HW_CACHE_OP_MAX]
522 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
526 [ C(RESULT_ACCESS) ] = 0x10001,
527 [ C(RESULT_MISS) ] = 0x3fbfc00001,
530 [ C(RESULT_ACCESS) ] = 0x3f3ffc0002,
531 [ C(RESULT_MISS) ] = 0x3f3fc00002,
536 [ C(RESULT_ACCESS) ] = 0x10c000001,
537 [ C(RESULT_MISS) ] = 0x3fb3000001,
543 * Notes on the events:
544 * - data reads do not include code reads (comparable to earlier tables)
545 * - data counts include speculative execution (except L1 write, dtlb, bpu)
546 * - remote node access includes remote memory, remote cache, remote mmio.
547 * - prefetches are not included in the counts.
548 * - icache miss does not include decoded icache
551 #define SKL_DEMAND_DATA_RD BIT_ULL(0)
552 #define SKL_DEMAND_RFO BIT_ULL(1)
553 #define SKL_ANY_RESPONSE BIT_ULL(16)
554 #define SKL_SUPPLIER_NONE BIT_ULL(17)
555 #define SKL_L3_MISS_LOCAL_DRAM BIT_ULL(26)
556 #define SKL_L3_MISS_REMOTE_HOP0_DRAM BIT_ULL(27)
557 #define SKL_L3_MISS_REMOTE_HOP1_DRAM BIT_ULL(28)
558 #define SKL_L3_MISS_REMOTE_HOP2P_DRAM BIT_ULL(29)
559 #define SKL_L3_MISS (SKL_L3_MISS_LOCAL_DRAM| \
560 SKL_L3_MISS_REMOTE_HOP0_DRAM| \
561 SKL_L3_MISS_REMOTE_HOP1_DRAM| \
562 SKL_L3_MISS_REMOTE_HOP2P_DRAM)
563 #define SKL_SPL_HIT BIT_ULL(30)
564 #define SKL_SNOOP_NONE BIT_ULL(31)
565 #define SKL_SNOOP_NOT_NEEDED BIT_ULL(32)
566 #define SKL_SNOOP_MISS BIT_ULL(33)
567 #define SKL_SNOOP_HIT_NO_FWD BIT_ULL(34)
568 #define SKL_SNOOP_HIT_WITH_FWD BIT_ULL(35)
569 #define SKL_SNOOP_HITM BIT_ULL(36)
570 #define SKL_SNOOP_NON_DRAM BIT_ULL(37)
571 #define SKL_ANY_SNOOP (SKL_SPL_HIT|SKL_SNOOP_NONE| \
572 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \
573 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \
574 SKL_SNOOP_HITM|SKL_SNOOP_NON_DRAM)
575 #define SKL_DEMAND_READ SKL_DEMAND_DATA_RD
576 #define SKL_SNOOP_DRAM (SKL_SNOOP_NONE| \
577 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \
578 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \
579 SKL_SNOOP_HITM|SKL_SPL_HIT)
580 #define SKL_DEMAND_WRITE SKL_DEMAND_RFO
581 #define SKL_LLC_ACCESS SKL_ANY_RESPONSE
582 #define SKL_L3_MISS_REMOTE (SKL_L3_MISS_REMOTE_HOP0_DRAM| \
583 SKL_L3_MISS_REMOTE_HOP1_DRAM| \
584 SKL_L3_MISS_REMOTE_HOP2P_DRAM)
586 static __initconst const u64 skl_hw_cache_event_ids
587 [PERF_COUNT_HW_CACHE_MAX]
588 [PERF_COUNT_HW_CACHE_OP_MAX]
589 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
593 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */
594 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */
597 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */
598 [ C(RESULT_MISS) ] = 0x0,
600 [ C(OP_PREFETCH) ] = {
601 [ C(RESULT_ACCESS) ] = 0x0,
602 [ C(RESULT_MISS) ] = 0x0,
607 [ C(RESULT_ACCESS) ] = 0x0,
608 [ C(RESULT_MISS) ] = 0x283, /* ICACHE_64B.MISS */
611 [ C(RESULT_ACCESS) ] = -1,
612 [ C(RESULT_MISS) ] = -1,
614 [ C(OP_PREFETCH) ] = {
615 [ C(RESULT_ACCESS) ] = 0x0,
616 [ C(RESULT_MISS) ] = 0x0,
621 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
622 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
625 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
626 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
628 [ C(OP_PREFETCH) ] = {
629 [ C(RESULT_ACCESS) ] = 0x0,
630 [ C(RESULT_MISS) ] = 0x0,
635 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */
636 [ C(RESULT_MISS) ] = 0xe08, /* DTLB_LOAD_MISSES.WALK_COMPLETED */
639 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */
640 [ C(RESULT_MISS) ] = 0xe49, /* DTLB_STORE_MISSES.WALK_COMPLETED */
642 [ C(OP_PREFETCH) ] = {
643 [ C(RESULT_ACCESS) ] = 0x0,
644 [ C(RESULT_MISS) ] = 0x0,
649 [ C(RESULT_ACCESS) ] = 0x2085, /* ITLB_MISSES.STLB_HIT */
650 [ C(RESULT_MISS) ] = 0xe85, /* ITLB_MISSES.WALK_COMPLETED */
653 [ C(RESULT_ACCESS) ] = -1,
654 [ C(RESULT_MISS) ] = -1,
656 [ C(OP_PREFETCH) ] = {
657 [ C(RESULT_ACCESS) ] = -1,
658 [ C(RESULT_MISS) ] = -1,
663 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */
664 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */
667 [ C(RESULT_ACCESS) ] = -1,
668 [ C(RESULT_MISS) ] = -1,
670 [ C(OP_PREFETCH) ] = {
671 [ C(RESULT_ACCESS) ] = -1,
672 [ C(RESULT_MISS) ] = -1,
677 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
678 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
681 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
682 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
684 [ C(OP_PREFETCH) ] = {
685 [ C(RESULT_ACCESS) ] = 0x0,
686 [ C(RESULT_MISS) ] = 0x0,
691 static __initconst const u64 skl_hw_cache_extra_regs
692 [PERF_COUNT_HW_CACHE_MAX]
693 [PERF_COUNT_HW_CACHE_OP_MAX]
694 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
698 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ|
699 SKL_LLC_ACCESS|SKL_ANY_SNOOP,
700 [ C(RESULT_MISS) ] = SKL_DEMAND_READ|
701 SKL_L3_MISS|SKL_ANY_SNOOP|
705 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE|
706 SKL_LLC_ACCESS|SKL_ANY_SNOOP,
707 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE|
708 SKL_L3_MISS|SKL_ANY_SNOOP|
711 [ C(OP_PREFETCH) ] = {
712 [ C(RESULT_ACCESS) ] = 0x0,
713 [ C(RESULT_MISS) ] = 0x0,
718 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ|
719 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM,
720 [ C(RESULT_MISS) ] = SKL_DEMAND_READ|
721 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM,
724 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE|
725 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM,
726 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE|
727 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM,
729 [ C(OP_PREFETCH) ] = {
730 [ C(RESULT_ACCESS) ] = 0x0,
731 [ C(RESULT_MISS) ] = 0x0,
736 #define SNB_DMND_DATA_RD (1ULL << 0)
737 #define SNB_DMND_RFO (1ULL << 1)
738 #define SNB_DMND_IFETCH (1ULL << 2)
739 #define SNB_DMND_WB (1ULL << 3)
740 #define SNB_PF_DATA_RD (1ULL << 4)
741 #define SNB_PF_RFO (1ULL << 5)
742 #define SNB_PF_IFETCH (1ULL << 6)
743 #define SNB_LLC_DATA_RD (1ULL << 7)
744 #define SNB_LLC_RFO (1ULL << 8)
745 #define SNB_LLC_IFETCH (1ULL << 9)
746 #define SNB_BUS_LOCKS (1ULL << 10)
747 #define SNB_STRM_ST (1ULL << 11)
748 #define SNB_OTHER (1ULL << 15)
749 #define SNB_RESP_ANY (1ULL << 16)
750 #define SNB_NO_SUPP (1ULL << 17)
751 #define SNB_LLC_HITM (1ULL << 18)
752 #define SNB_LLC_HITE (1ULL << 19)
753 #define SNB_LLC_HITS (1ULL << 20)
754 #define SNB_LLC_HITF (1ULL << 21)
755 #define SNB_LOCAL (1ULL << 22)
756 #define SNB_REMOTE (0xffULL << 23)
757 #define SNB_SNP_NONE (1ULL << 31)
758 #define SNB_SNP_NOT_NEEDED (1ULL << 32)
759 #define SNB_SNP_MISS (1ULL << 33)
760 #define SNB_NO_FWD (1ULL << 34)
761 #define SNB_SNP_FWD (1ULL << 35)
762 #define SNB_HITM (1ULL << 36)
763 #define SNB_NON_DRAM (1ULL << 37)
765 #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD)
766 #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO)
767 #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
769 #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \
770 SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \
773 #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY)
774 #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY)
776 #define SNB_L3_ACCESS SNB_RESP_ANY
777 #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM)
779 static __initconst const u64 snb_hw_cache_extra_regs
780 [PERF_COUNT_HW_CACHE_MAX]
781 [PERF_COUNT_HW_CACHE_OP_MAX]
782 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
786 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS,
787 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_L3_MISS,
790 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS,
791 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_L3_MISS,
793 [ C(OP_PREFETCH) ] = {
794 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS,
795 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_L3_MISS,
800 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY,
801 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_DRAM_REMOTE,
804 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY,
805 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE,
807 [ C(OP_PREFETCH) ] = {
808 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY,
809 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE,
814 static __initconst const u64 snb_hw_cache_event_ids
815 [PERF_COUNT_HW_CACHE_MAX]
816 [PERF_COUNT_HW_CACHE_OP_MAX]
817 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
821 [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */
822 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */
825 [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */
826 [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */
828 [ C(OP_PREFETCH) ] = {
829 [ C(RESULT_ACCESS) ] = 0x0,
830 [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */
835 [ C(RESULT_ACCESS) ] = 0x0,
836 [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */
839 [ C(RESULT_ACCESS) ] = -1,
840 [ C(RESULT_MISS) ] = -1,
842 [ C(OP_PREFETCH) ] = {
843 [ C(RESULT_ACCESS) ] = 0x0,
844 [ C(RESULT_MISS) ] = 0x0,
849 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
850 [ C(RESULT_ACCESS) ] = 0x01b7,
851 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
852 [ C(RESULT_MISS) ] = 0x01b7,
855 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
856 [ C(RESULT_ACCESS) ] = 0x01b7,
857 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
858 [ C(RESULT_MISS) ] = 0x01b7,
860 [ C(OP_PREFETCH) ] = {
861 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
862 [ C(RESULT_ACCESS) ] = 0x01b7,
863 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
864 [ C(RESULT_MISS) ] = 0x01b7,
869 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */
870 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */
873 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */
874 [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
876 [ C(OP_PREFETCH) ] = {
877 [ C(RESULT_ACCESS) ] = 0x0,
878 [ C(RESULT_MISS) ] = 0x0,
883 [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */
884 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */
887 [ C(RESULT_ACCESS) ] = -1,
888 [ C(RESULT_MISS) ] = -1,
890 [ C(OP_PREFETCH) ] = {
891 [ C(RESULT_ACCESS) ] = -1,
892 [ C(RESULT_MISS) ] = -1,
897 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
898 [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
901 [ C(RESULT_ACCESS) ] = -1,
902 [ C(RESULT_MISS) ] = -1,
904 [ C(OP_PREFETCH) ] = {
905 [ C(RESULT_ACCESS) ] = -1,
906 [ C(RESULT_MISS) ] = -1,
911 [ C(RESULT_ACCESS) ] = 0x01b7,
912 [ C(RESULT_MISS) ] = 0x01b7,
915 [ C(RESULT_ACCESS) ] = 0x01b7,
916 [ C(RESULT_MISS) ] = 0x01b7,
918 [ C(OP_PREFETCH) ] = {
919 [ C(RESULT_ACCESS) ] = 0x01b7,
920 [ C(RESULT_MISS) ] = 0x01b7,
927 * Notes on the events:
928 * - data reads do not include code reads (comparable to earlier tables)
929 * - data counts include speculative execution (except L1 write, dtlb, bpu)
930 * - remote node access includes remote memory, remote cache, remote mmio.
931 * - prefetches are not included in the counts because they are not
935 #define HSW_DEMAND_DATA_RD BIT_ULL(0)
936 #define HSW_DEMAND_RFO BIT_ULL(1)
937 #define HSW_ANY_RESPONSE BIT_ULL(16)
938 #define HSW_SUPPLIER_NONE BIT_ULL(17)
939 #define HSW_L3_MISS_LOCAL_DRAM BIT_ULL(22)
940 #define HSW_L3_MISS_REMOTE_HOP0 BIT_ULL(27)
941 #define HSW_L3_MISS_REMOTE_HOP1 BIT_ULL(28)
942 #define HSW_L3_MISS_REMOTE_HOP2P BIT_ULL(29)
943 #define HSW_L3_MISS (HSW_L3_MISS_LOCAL_DRAM| \
944 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \
945 HSW_L3_MISS_REMOTE_HOP2P)
946 #define HSW_SNOOP_NONE BIT_ULL(31)
947 #define HSW_SNOOP_NOT_NEEDED BIT_ULL(32)
948 #define HSW_SNOOP_MISS BIT_ULL(33)
949 #define HSW_SNOOP_HIT_NO_FWD BIT_ULL(34)
950 #define HSW_SNOOP_HIT_WITH_FWD BIT_ULL(35)
951 #define HSW_SNOOP_HITM BIT_ULL(36)
952 #define HSW_SNOOP_NON_DRAM BIT_ULL(37)
953 #define HSW_ANY_SNOOP (HSW_SNOOP_NONE| \
954 HSW_SNOOP_NOT_NEEDED|HSW_SNOOP_MISS| \
955 HSW_SNOOP_HIT_NO_FWD|HSW_SNOOP_HIT_WITH_FWD| \
956 HSW_SNOOP_HITM|HSW_SNOOP_NON_DRAM)
957 #define HSW_SNOOP_DRAM (HSW_ANY_SNOOP & ~HSW_SNOOP_NON_DRAM)
958 #define HSW_DEMAND_READ HSW_DEMAND_DATA_RD
959 #define HSW_DEMAND_WRITE HSW_DEMAND_RFO
960 #define HSW_L3_MISS_REMOTE (HSW_L3_MISS_REMOTE_HOP0|\
961 HSW_L3_MISS_REMOTE_HOP1|HSW_L3_MISS_REMOTE_HOP2P)
962 #define HSW_LLC_ACCESS HSW_ANY_RESPONSE
964 #define BDW_L3_MISS_LOCAL BIT(26)
965 #define BDW_L3_MISS (BDW_L3_MISS_LOCAL| \
966 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \
967 HSW_L3_MISS_REMOTE_HOP2P)
970 static __initconst const u64 hsw_hw_cache_event_ids
971 [PERF_COUNT_HW_CACHE_MAX]
972 [PERF_COUNT_HW_CACHE_OP_MAX]
973 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
977 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
978 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */
981 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
982 [ C(RESULT_MISS) ] = 0x0,
984 [ C(OP_PREFETCH) ] = {
985 [ C(RESULT_ACCESS) ] = 0x0,
986 [ C(RESULT_MISS) ] = 0x0,
991 [ C(RESULT_ACCESS) ] = 0x0,
992 [ C(RESULT_MISS) ] = 0x280, /* ICACHE.MISSES */
995 [ C(RESULT_ACCESS) ] = -1,
996 [ C(RESULT_MISS) ] = -1,
998 [ C(OP_PREFETCH) ] = {
999 [ C(RESULT_ACCESS) ] = 0x0,
1000 [ C(RESULT_MISS) ] = 0x0,
1005 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
1006 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
1009 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
1010 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
1012 [ C(OP_PREFETCH) ] = {
1013 [ C(RESULT_ACCESS) ] = 0x0,
1014 [ C(RESULT_MISS) ] = 0x0,
1019 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
1020 [ C(RESULT_MISS) ] = 0x108, /* DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK */
1023 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
1024 [ C(RESULT_MISS) ] = 0x149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
1026 [ C(OP_PREFETCH) ] = {
1027 [ C(RESULT_ACCESS) ] = 0x0,
1028 [ C(RESULT_MISS) ] = 0x0,
1033 [ C(RESULT_ACCESS) ] = 0x6085, /* ITLB_MISSES.STLB_HIT */
1034 [ C(RESULT_MISS) ] = 0x185, /* ITLB_MISSES.MISS_CAUSES_A_WALK */
1037 [ C(RESULT_ACCESS) ] = -1,
1038 [ C(RESULT_MISS) ] = -1,
1040 [ C(OP_PREFETCH) ] = {
1041 [ C(RESULT_ACCESS) ] = -1,
1042 [ C(RESULT_MISS) ] = -1,
1047 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */
1048 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */
1051 [ C(RESULT_ACCESS) ] = -1,
1052 [ C(RESULT_MISS) ] = -1,
1054 [ C(OP_PREFETCH) ] = {
1055 [ C(RESULT_ACCESS) ] = -1,
1056 [ C(RESULT_MISS) ] = -1,
1061 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
1062 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
1065 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
1066 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
1068 [ C(OP_PREFETCH) ] = {
1069 [ C(RESULT_ACCESS) ] = 0x0,
1070 [ C(RESULT_MISS) ] = 0x0,
1075 static __initconst const u64 hsw_hw_cache_extra_regs
1076 [PERF_COUNT_HW_CACHE_MAX]
1077 [PERF_COUNT_HW_CACHE_OP_MAX]
1078 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1082 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ|
1084 [ C(RESULT_MISS) ] = HSW_DEMAND_READ|
1085 HSW_L3_MISS|HSW_ANY_SNOOP,
1088 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE|
1090 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE|
1091 HSW_L3_MISS|HSW_ANY_SNOOP,
1093 [ C(OP_PREFETCH) ] = {
1094 [ C(RESULT_ACCESS) ] = 0x0,
1095 [ C(RESULT_MISS) ] = 0x0,
1100 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ|
1101 HSW_L3_MISS_LOCAL_DRAM|
1103 [ C(RESULT_MISS) ] = HSW_DEMAND_READ|
1108 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE|
1109 HSW_L3_MISS_LOCAL_DRAM|
1111 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE|
1115 [ C(OP_PREFETCH) ] = {
1116 [ C(RESULT_ACCESS) ] = 0x0,
1117 [ C(RESULT_MISS) ] = 0x0,
1122 static __initconst const u64 westmere_hw_cache_event_ids
1123 [PERF_COUNT_HW_CACHE_MAX]
1124 [PERF_COUNT_HW_CACHE_OP_MAX]
1125 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1129 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
1130 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
1133 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
1134 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
1136 [ C(OP_PREFETCH) ] = {
1137 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
1138 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
1143 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
1144 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
1147 [ C(RESULT_ACCESS) ] = -1,
1148 [ C(RESULT_MISS) ] = -1,
1150 [ C(OP_PREFETCH) ] = {
1151 [ C(RESULT_ACCESS) ] = 0x0,
1152 [ C(RESULT_MISS) ] = 0x0,
1157 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1158 [ C(RESULT_ACCESS) ] = 0x01b7,
1159 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
1160 [ C(RESULT_MISS) ] = 0x01b7,
1163 * Use RFO, not WRITEBACK, because a write miss would typically occur
1167 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1168 [ C(RESULT_ACCESS) ] = 0x01b7,
1169 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1170 [ C(RESULT_MISS) ] = 0x01b7,
1172 [ C(OP_PREFETCH) ] = {
1173 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1174 [ C(RESULT_ACCESS) ] = 0x01b7,
1175 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1176 [ C(RESULT_MISS) ] = 0x01b7,
1181 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
1182 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
1185 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
1186 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
1188 [ C(OP_PREFETCH) ] = {
1189 [ C(RESULT_ACCESS) ] = 0x0,
1190 [ C(RESULT_MISS) ] = 0x0,
1195 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
1196 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */
1199 [ C(RESULT_ACCESS) ] = -1,
1200 [ C(RESULT_MISS) ] = -1,
1202 [ C(OP_PREFETCH) ] = {
1203 [ C(RESULT_ACCESS) ] = -1,
1204 [ C(RESULT_MISS) ] = -1,
1209 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1210 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
1213 [ C(RESULT_ACCESS) ] = -1,
1214 [ C(RESULT_MISS) ] = -1,
1216 [ C(OP_PREFETCH) ] = {
1217 [ C(RESULT_ACCESS) ] = -1,
1218 [ C(RESULT_MISS) ] = -1,
1223 [ C(RESULT_ACCESS) ] = 0x01b7,
1224 [ C(RESULT_MISS) ] = 0x01b7,
1227 [ C(RESULT_ACCESS) ] = 0x01b7,
1228 [ C(RESULT_MISS) ] = 0x01b7,
1230 [ C(OP_PREFETCH) ] = {
1231 [ C(RESULT_ACCESS) ] = 0x01b7,
1232 [ C(RESULT_MISS) ] = 0x01b7,
1238 * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits;
1239 * See IA32 SDM Vol 3B 30.6.1.3
1242 #define NHM_DMND_DATA_RD (1 << 0)
1243 #define NHM_DMND_RFO (1 << 1)
1244 #define NHM_DMND_IFETCH (1 << 2)
1245 #define NHM_DMND_WB (1 << 3)
1246 #define NHM_PF_DATA_RD (1 << 4)
1247 #define NHM_PF_DATA_RFO (1 << 5)
1248 #define NHM_PF_IFETCH (1 << 6)
1249 #define NHM_OFFCORE_OTHER (1 << 7)
1250 #define NHM_UNCORE_HIT (1 << 8)
1251 #define NHM_OTHER_CORE_HIT_SNP (1 << 9)
1252 #define NHM_OTHER_CORE_HITM (1 << 10)
1254 #define NHM_REMOTE_CACHE_FWD (1 << 12)
1255 #define NHM_REMOTE_DRAM (1 << 13)
1256 #define NHM_LOCAL_DRAM (1 << 14)
1257 #define NHM_NON_DRAM (1 << 15)
1259 #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD)
1260 #define NHM_REMOTE (NHM_REMOTE_DRAM)
1262 #define NHM_DMND_READ (NHM_DMND_DATA_RD)
1263 #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB)
1264 #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO)
1266 #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM)
1267 #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD)
1268 #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS)
1270 static __initconst const u64 nehalem_hw_cache_extra_regs
1271 [PERF_COUNT_HW_CACHE_MAX]
1272 [PERF_COUNT_HW_CACHE_OP_MAX]
1273 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1277 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS,
1278 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS,
1281 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS,
1282 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS,
1284 [ C(OP_PREFETCH) ] = {
1285 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS,
1286 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS,
1291 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE,
1292 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE,
1295 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE,
1296 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE,
1298 [ C(OP_PREFETCH) ] = {
1299 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE,
1300 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE,
1305 static __initconst const u64 nehalem_hw_cache_event_ids
1306 [PERF_COUNT_HW_CACHE_MAX]
1307 [PERF_COUNT_HW_CACHE_OP_MAX]
1308 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1312 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
1313 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
1316 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
1317 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
1319 [ C(OP_PREFETCH) ] = {
1320 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
1321 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
1326 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
1327 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
1330 [ C(RESULT_ACCESS) ] = -1,
1331 [ C(RESULT_MISS) ] = -1,
1333 [ C(OP_PREFETCH) ] = {
1334 [ C(RESULT_ACCESS) ] = 0x0,
1335 [ C(RESULT_MISS) ] = 0x0,
1340 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1341 [ C(RESULT_ACCESS) ] = 0x01b7,
1342 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
1343 [ C(RESULT_MISS) ] = 0x01b7,
1346 * Use RFO, not WRITEBACK, because a write miss would typically occur
1350 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1351 [ C(RESULT_ACCESS) ] = 0x01b7,
1352 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1353 [ C(RESULT_MISS) ] = 0x01b7,
1355 [ C(OP_PREFETCH) ] = {
1356 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1357 [ C(RESULT_ACCESS) ] = 0x01b7,
1358 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1359 [ C(RESULT_MISS) ] = 0x01b7,
1364 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
1365 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
1368 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
1369 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
1371 [ C(OP_PREFETCH) ] = {
1372 [ C(RESULT_ACCESS) ] = 0x0,
1373 [ C(RESULT_MISS) ] = 0x0,
1378 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
1379 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */
1382 [ C(RESULT_ACCESS) ] = -1,
1383 [ C(RESULT_MISS) ] = -1,
1385 [ C(OP_PREFETCH) ] = {
1386 [ C(RESULT_ACCESS) ] = -1,
1387 [ C(RESULT_MISS) ] = -1,
1392 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1393 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
1396 [ C(RESULT_ACCESS) ] = -1,
1397 [ C(RESULT_MISS) ] = -1,
1399 [ C(OP_PREFETCH) ] = {
1400 [ C(RESULT_ACCESS) ] = -1,
1401 [ C(RESULT_MISS) ] = -1,
1406 [ C(RESULT_ACCESS) ] = 0x01b7,
1407 [ C(RESULT_MISS) ] = 0x01b7,
1410 [ C(RESULT_ACCESS) ] = 0x01b7,
1411 [ C(RESULT_MISS) ] = 0x01b7,
1413 [ C(OP_PREFETCH) ] = {
1414 [ C(RESULT_ACCESS) ] = 0x01b7,
1415 [ C(RESULT_MISS) ] = 0x01b7,
1420 static __initconst const u64 core2_hw_cache_event_ids
1421 [PERF_COUNT_HW_CACHE_MAX]
1422 [PERF_COUNT_HW_CACHE_OP_MAX]
1423 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1427 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
1428 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
1431 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
1432 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
1434 [ C(OP_PREFETCH) ] = {
1435 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */
1436 [ C(RESULT_MISS) ] = 0,
1441 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */
1442 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */
1445 [ C(RESULT_ACCESS) ] = -1,
1446 [ C(RESULT_MISS) ] = -1,
1448 [ C(OP_PREFETCH) ] = {
1449 [ C(RESULT_ACCESS) ] = 0,
1450 [ C(RESULT_MISS) ] = 0,
1455 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
1456 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
1459 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
1460 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
1462 [ C(OP_PREFETCH) ] = {
1463 [ C(RESULT_ACCESS) ] = 0,
1464 [ C(RESULT_MISS) ] = 0,
1469 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
1470 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */
1473 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
1474 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */
1476 [ C(OP_PREFETCH) ] = {
1477 [ C(RESULT_ACCESS) ] = 0,
1478 [ C(RESULT_MISS) ] = 0,
1483 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1484 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */
1487 [ C(RESULT_ACCESS) ] = -1,
1488 [ C(RESULT_MISS) ] = -1,
1490 [ C(OP_PREFETCH) ] = {
1491 [ C(RESULT_ACCESS) ] = -1,
1492 [ C(RESULT_MISS) ] = -1,
1497 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1498 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1501 [ C(RESULT_ACCESS) ] = -1,
1502 [ C(RESULT_MISS) ] = -1,
1504 [ C(OP_PREFETCH) ] = {
1505 [ C(RESULT_ACCESS) ] = -1,
1506 [ C(RESULT_MISS) ] = -1,
1511 static __initconst const u64 atom_hw_cache_event_ids
1512 [PERF_COUNT_HW_CACHE_MAX]
1513 [PERF_COUNT_HW_CACHE_OP_MAX]
1514 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1518 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */
1519 [ C(RESULT_MISS) ] = 0,
1522 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */
1523 [ C(RESULT_MISS) ] = 0,
1525 [ C(OP_PREFETCH) ] = {
1526 [ C(RESULT_ACCESS) ] = 0x0,
1527 [ C(RESULT_MISS) ] = 0,
1532 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
1533 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
1536 [ C(RESULT_ACCESS) ] = -1,
1537 [ C(RESULT_MISS) ] = -1,
1539 [ C(OP_PREFETCH) ] = {
1540 [ C(RESULT_ACCESS) ] = 0,
1541 [ C(RESULT_MISS) ] = 0,
1546 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
1547 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
1550 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
1551 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
1553 [ C(OP_PREFETCH) ] = {
1554 [ C(RESULT_ACCESS) ] = 0,
1555 [ C(RESULT_MISS) ] = 0,
1560 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */
1561 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */
1564 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */
1565 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */
1567 [ C(OP_PREFETCH) ] = {
1568 [ C(RESULT_ACCESS) ] = 0,
1569 [ C(RESULT_MISS) ] = 0,
1574 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1575 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */
1578 [ C(RESULT_ACCESS) ] = -1,
1579 [ C(RESULT_MISS) ] = -1,
1581 [ C(OP_PREFETCH) ] = {
1582 [ C(RESULT_ACCESS) ] = -1,
1583 [ C(RESULT_MISS) ] = -1,
1588 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1589 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1592 [ C(RESULT_ACCESS) ] = -1,
1593 [ C(RESULT_MISS) ] = -1,
1595 [ C(OP_PREFETCH) ] = {
1596 [ C(RESULT_ACCESS) ] = -1,
1597 [ C(RESULT_MISS) ] = -1,
1602 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_slm, "event=0x3c");
1603 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_slm, "2");
1604 /* no_alloc_cycles.not_delivered */
1605 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_slm,
1606 "event=0xca,umask=0x50");
1607 EVENT_ATTR_STR(topdown-fetch-bubbles.scale, td_fetch_bubbles_scale_slm, "2");
1608 /* uops_retired.all */
1609 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_slm,
1610 "event=0xc2,umask=0x10");
1611 /* uops_retired.all */
1612 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_slm,
1613 "event=0xc2,umask=0x10");
1615 static struct attribute *slm_events_attrs[] = {
1616 EVENT_PTR(td_total_slots_slm),
1617 EVENT_PTR(td_total_slots_scale_slm),
1618 EVENT_PTR(td_fetch_bubbles_slm),
1619 EVENT_PTR(td_fetch_bubbles_scale_slm),
1620 EVENT_PTR(td_slots_issued_slm),
1621 EVENT_PTR(td_slots_retired_slm),
1625 static struct extra_reg intel_slm_extra_regs[] __read_mostly =
1627 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
1628 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x768005ffffull, RSP_0),
1629 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x368005ffffull, RSP_1),
1633 #define SLM_DMND_READ SNB_DMND_DATA_RD
1634 #define SLM_DMND_WRITE SNB_DMND_RFO
1635 #define SLM_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
1637 #define SLM_SNP_ANY (SNB_SNP_NONE|SNB_SNP_MISS|SNB_NO_FWD|SNB_HITM)
1638 #define SLM_LLC_ACCESS SNB_RESP_ANY
1639 #define SLM_LLC_MISS (SLM_SNP_ANY|SNB_NON_DRAM)
1641 static __initconst const u64 slm_hw_cache_extra_regs
1642 [PERF_COUNT_HW_CACHE_MAX]
1643 [PERF_COUNT_HW_CACHE_OP_MAX]
1644 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1648 [ C(RESULT_ACCESS) ] = SLM_DMND_READ|SLM_LLC_ACCESS,
1649 [ C(RESULT_MISS) ] = 0,
1652 [ C(RESULT_ACCESS) ] = SLM_DMND_WRITE|SLM_LLC_ACCESS,
1653 [ C(RESULT_MISS) ] = SLM_DMND_WRITE|SLM_LLC_MISS,
1655 [ C(OP_PREFETCH) ] = {
1656 [ C(RESULT_ACCESS) ] = SLM_DMND_PREFETCH|SLM_LLC_ACCESS,
1657 [ C(RESULT_MISS) ] = SLM_DMND_PREFETCH|SLM_LLC_MISS,
1662 static __initconst const u64 slm_hw_cache_event_ids
1663 [PERF_COUNT_HW_CACHE_MAX]
1664 [PERF_COUNT_HW_CACHE_OP_MAX]
1665 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1669 [ C(RESULT_ACCESS) ] = 0,
1670 [ C(RESULT_MISS) ] = 0x0104, /* LD_DCU_MISS */
1673 [ C(RESULT_ACCESS) ] = 0,
1674 [ C(RESULT_MISS) ] = 0,
1676 [ C(OP_PREFETCH) ] = {
1677 [ C(RESULT_ACCESS) ] = 0,
1678 [ C(RESULT_MISS) ] = 0,
1683 [ C(RESULT_ACCESS) ] = 0x0380, /* ICACHE.ACCESSES */
1684 [ C(RESULT_MISS) ] = 0x0280, /* ICACGE.MISSES */
1687 [ C(RESULT_ACCESS) ] = -1,
1688 [ C(RESULT_MISS) ] = -1,
1690 [ C(OP_PREFETCH) ] = {
1691 [ C(RESULT_ACCESS) ] = 0,
1692 [ C(RESULT_MISS) ] = 0,
1697 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1698 [ C(RESULT_ACCESS) ] = 0x01b7,
1699 [ C(RESULT_MISS) ] = 0,
1702 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1703 [ C(RESULT_ACCESS) ] = 0x01b7,
1704 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1705 [ C(RESULT_MISS) ] = 0x01b7,
1707 [ C(OP_PREFETCH) ] = {
1708 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1709 [ C(RESULT_ACCESS) ] = 0x01b7,
1710 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1711 [ C(RESULT_MISS) ] = 0x01b7,
1716 [ C(RESULT_ACCESS) ] = 0,
1717 [ C(RESULT_MISS) ] = 0x0804, /* LD_DTLB_MISS */
1720 [ C(RESULT_ACCESS) ] = 0,
1721 [ C(RESULT_MISS) ] = 0,
1723 [ C(OP_PREFETCH) ] = {
1724 [ C(RESULT_ACCESS) ] = 0,
1725 [ C(RESULT_MISS) ] = 0,
1730 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1731 [ C(RESULT_MISS) ] = 0x40205, /* PAGE_WALKS.I_SIDE_WALKS */
1734 [ C(RESULT_ACCESS) ] = -1,
1735 [ C(RESULT_MISS) ] = -1,
1737 [ C(OP_PREFETCH) ] = {
1738 [ C(RESULT_ACCESS) ] = -1,
1739 [ C(RESULT_MISS) ] = -1,
1744 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1745 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1748 [ C(RESULT_ACCESS) ] = -1,
1749 [ C(RESULT_MISS) ] = -1,
1751 [ C(OP_PREFETCH) ] = {
1752 [ C(RESULT_ACCESS) ] = -1,
1753 [ C(RESULT_MISS) ] = -1,
1758 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_glm, "event=0x3c");
1759 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_glm, "3");
1760 /* UOPS_NOT_DELIVERED.ANY */
1761 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_glm, "event=0x9c");
1762 /* ISSUE_SLOTS_NOT_CONSUMED.RECOVERY */
1763 EVENT_ATTR_STR(topdown-recovery-bubbles, td_recovery_bubbles_glm, "event=0xca,umask=0x02");
1764 /* UOPS_RETIRED.ANY */
1765 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_glm, "event=0xc2");
1766 /* UOPS_ISSUED.ANY */
1767 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_glm, "event=0x0e");
1769 static struct attribute *glm_events_attrs[] = {
1770 EVENT_PTR(td_total_slots_glm),
1771 EVENT_PTR(td_total_slots_scale_glm),
1772 EVENT_PTR(td_fetch_bubbles_glm),
1773 EVENT_PTR(td_recovery_bubbles_glm),
1774 EVENT_PTR(td_slots_issued_glm),
1775 EVENT_PTR(td_slots_retired_glm),
1779 static struct extra_reg intel_glm_extra_regs[] __read_mostly = {
1780 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
1781 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x760005ffbfull, RSP_0),
1782 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x360005ffbfull, RSP_1),
1786 #define GLM_DEMAND_DATA_RD BIT_ULL(0)
1787 #define GLM_DEMAND_RFO BIT_ULL(1)
1788 #define GLM_ANY_RESPONSE BIT_ULL(16)
1789 #define GLM_SNP_NONE_OR_MISS BIT_ULL(33)
1790 #define GLM_DEMAND_READ GLM_DEMAND_DATA_RD
1791 #define GLM_DEMAND_WRITE GLM_DEMAND_RFO
1792 #define GLM_DEMAND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
1793 #define GLM_LLC_ACCESS GLM_ANY_RESPONSE
1794 #define GLM_SNP_ANY (GLM_SNP_NONE_OR_MISS|SNB_NO_FWD|SNB_HITM)
1795 #define GLM_LLC_MISS (GLM_SNP_ANY|SNB_NON_DRAM)
1797 static __initconst const u64 glm_hw_cache_event_ids
1798 [PERF_COUNT_HW_CACHE_MAX]
1799 [PERF_COUNT_HW_CACHE_OP_MAX]
1800 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1803 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
1804 [C(RESULT_MISS)] = 0x0,
1807 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
1808 [C(RESULT_MISS)] = 0x0,
1810 [C(OP_PREFETCH)] = {
1811 [C(RESULT_ACCESS)] = 0x0,
1812 [C(RESULT_MISS)] = 0x0,
1817 [C(RESULT_ACCESS)] = 0x0380, /* ICACHE.ACCESSES */
1818 [C(RESULT_MISS)] = 0x0280, /* ICACHE.MISSES */
1821 [C(RESULT_ACCESS)] = -1,
1822 [C(RESULT_MISS)] = -1,
1824 [C(OP_PREFETCH)] = {
1825 [C(RESULT_ACCESS)] = 0x0,
1826 [C(RESULT_MISS)] = 0x0,
1831 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1832 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1835 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1836 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1838 [C(OP_PREFETCH)] = {
1839 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1840 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1845 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
1846 [C(RESULT_MISS)] = 0x0,
1849 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
1850 [C(RESULT_MISS)] = 0x0,
1852 [C(OP_PREFETCH)] = {
1853 [C(RESULT_ACCESS)] = 0x0,
1854 [C(RESULT_MISS)] = 0x0,
1859 [C(RESULT_ACCESS)] = 0x00c0, /* INST_RETIRED.ANY_P */
1860 [C(RESULT_MISS)] = 0x0481, /* ITLB.MISS */
1863 [C(RESULT_ACCESS)] = -1,
1864 [C(RESULT_MISS)] = -1,
1866 [C(OP_PREFETCH)] = {
1867 [C(RESULT_ACCESS)] = -1,
1868 [C(RESULT_MISS)] = -1,
1873 [C(RESULT_ACCESS)] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1874 [C(RESULT_MISS)] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
1877 [C(RESULT_ACCESS)] = -1,
1878 [C(RESULT_MISS)] = -1,
1880 [C(OP_PREFETCH)] = {
1881 [C(RESULT_ACCESS)] = -1,
1882 [C(RESULT_MISS)] = -1,
1887 static __initconst const u64 glm_hw_cache_extra_regs
1888 [PERF_COUNT_HW_CACHE_MAX]
1889 [PERF_COUNT_HW_CACHE_OP_MAX]
1890 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1893 [C(RESULT_ACCESS)] = GLM_DEMAND_READ|
1895 [C(RESULT_MISS)] = GLM_DEMAND_READ|
1899 [C(RESULT_ACCESS)] = GLM_DEMAND_WRITE|
1901 [C(RESULT_MISS)] = GLM_DEMAND_WRITE|
1904 [C(OP_PREFETCH)] = {
1905 [C(RESULT_ACCESS)] = GLM_DEMAND_PREFETCH|
1907 [C(RESULT_MISS)] = GLM_DEMAND_PREFETCH|
1913 static __initconst const u64 glp_hw_cache_event_ids
1914 [PERF_COUNT_HW_CACHE_MAX]
1915 [PERF_COUNT_HW_CACHE_OP_MAX]
1916 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1919 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
1920 [C(RESULT_MISS)] = 0x0,
1923 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
1924 [C(RESULT_MISS)] = 0x0,
1926 [C(OP_PREFETCH)] = {
1927 [C(RESULT_ACCESS)] = 0x0,
1928 [C(RESULT_MISS)] = 0x0,
1933 [C(RESULT_ACCESS)] = 0x0380, /* ICACHE.ACCESSES */
1934 [C(RESULT_MISS)] = 0x0280, /* ICACHE.MISSES */
1937 [C(RESULT_ACCESS)] = -1,
1938 [C(RESULT_MISS)] = -1,
1940 [C(OP_PREFETCH)] = {
1941 [C(RESULT_ACCESS)] = 0x0,
1942 [C(RESULT_MISS)] = 0x0,
1947 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1948 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1951 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1952 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1954 [C(OP_PREFETCH)] = {
1955 [C(RESULT_ACCESS)] = 0x0,
1956 [C(RESULT_MISS)] = 0x0,
1961 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
1962 [C(RESULT_MISS)] = 0xe08, /* DTLB_LOAD_MISSES.WALK_COMPLETED */
1965 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
1966 [C(RESULT_MISS)] = 0xe49, /* DTLB_STORE_MISSES.WALK_COMPLETED */
1968 [C(OP_PREFETCH)] = {
1969 [C(RESULT_ACCESS)] = 0x0,
1970 [C(RESULT_MISS)] = 0x0,
1975 [C(RESULT_ACCESS)] = 0x00c0, /* INST_RETIRED.ANY_P */
1976 [C(RESULT_MISS)] = 0x0481, /* ITLB.MISS */
1979 [C(RESULT_ACCESS)] = -1,
1980 [C(RESULT_MISS)] = -1,
1982 [C(OP_PREFETCH)] = {
1983 [C(RESULT_ACCESS)] = -1,
1984 [C(RESULT_MISS)] = -1,
1989 [C(RESULT_ACCESS)] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1990 [C(RESULT_MISS)] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
1993 [C(RESULT_ACCESS)] = -1,
1994 [C(RESULT_MISS)] = -1,
1996 [C(OP_PREFETCH)] = {
1997 [C(RESULT_ACCESS)] = -1,
1998 [C(RESULT_MISS)] = -1,
2003 static __initconst const u64 glp_hw_cache_extra_regs
2004 [PERF_COUNT_HW_CACHE_MAX]
2005 [PERF_COUNT_HW_CACHE_OP_MAX]
2006 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
2009 [C(RESULT_ACCESS)] = GLM_DEMAND_READ|
2011 [C(RESULT_MISS)] = GLM_DEMAND_READ|
2015 [C(RESULT_ACCESS)] = GLM_DEMAND_WRITE|
2017 [C(RESULT_MISS)] = GLM_DEMAND_WRITE|
2020 [C(OP_PREFETCH)] = {
2021 [C(RESULT_ACCESS)] = 0x0,
2022 [C(RESULT_MISS)] = 0x0,
2027 #define TNT_LOCAL_DRAM BIT_ULL(26)
2028 #define TNT_DEMAND_READ GLM_DEMAND_DATA_RD
2029 #define TNT_DEMAND_WRITE GLM_DEMAND_RFO
2030 #define TNT_LLC_ACCESS GLM_ANY_RESPONSE
2031 #define TNT_SNP_ANY (SNB_SNP_NOT_NEEDED|SNB_SNP_MISS| \
2032 SNB_NO_FWD|SNB_SNP_FWD|SNB_HITM)
2033 #define TNT_LLC_MISS (TNT_SNP_ANY|SNB_NON_DRAM|TNT_LOCAL_DRAM)
2035 static __initconst const u64 tnt_hw_cache_extra_regs
2036 [PERF_COUNT_HW_CACHE_MAX]
2037 [PERF_COUNT_HW_CACHE_OP_MAX]
2038 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
2041 [C(RESULT_ACCESS)] = TNT_DEMAND_READ|
2043 [C(RESULT_MISS)] = TNT_DEMAND_READ|
2047 [C(RESULT_ACCESS)] = TNT_DEMAND_WRITE|
2049 [C(RESULT_MISS)] = TNT_DEMAND_WRITE|
2052 [C(OP_PREFETCH)] = {
2053 [C(RESULT_ACCESS)] = 0x0,
2054 [C(RESULT_MISS)] = 0x0,
2059 EVENT_ATTR_STR(topdown-fe-bound, td_fe_bound_tnt, "event=0x71,umask=0x0");
2060 EVENT_ATTR_STR(topdown-retiring, td_retiring_tnt, "event=0xc2,umask=0x0");
2061 EVENT_ATTR_STR(topdown-bad-spec, td_bad_spec_tnt, "event=0x73,umask=0x6");
2062 EVENT_ATTR_STR(topdown-be-bound, td_be_bound_tnt, "event=0x74,umask=0x0");
2064 static struct attribute *tnt_events_attrs[] = {
2065 EVENT_PTR(td_fe_bound_tnt),
2066 EVENT_PTR(td_retiring_tnt),
2067 EVENT_PTR(td_bad_spec_tnt),
2068 EVENT_PTR(td_be_bound_tnt),
2072 static struct extra_reg intel_tnt_extra_regs[] __read_mostly = {
2073 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
2074 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x800ff0ffffff9fffull, RSP_0),
2075 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0xff0ffffff9fffull, RSP_1),
2079 #define KNL_OT_L2_HITE BIT_ULL(19) /* Other Tile L2 Hit */
2080 #define KNL_OT_L2_HITF BIT_ULL(20) /* Other Tile L2 Hit */
2081 #define KNL_MCDRAM_LOCAL BIT_ULL(21)
2082 #define KNL_MCDRAM_FAR BIT_ULL(22)
2083 #define KNL_DDR_LOCAL BIT_ULL(23)
2084 #define KNL_DDR_FAR BIT_ULL(24)
2085 #define KNL_DRAM_ANY (KNL_MCDRAM_LOCAL | KNL_MCDRAM_FAR | \
2086 KNL_DDR_LOCAL | KNL_DDR_FAR)
2087 #define KNL_L2_READ SLM_DMND_READ
2088 #define KNL_L2_WRITE SLM_DMND_WRITE
2089 #define KNL_L2_PREFETCH SLM_DMND_PREFETCH
2090 #define KNL_L2_ACCESS SLM_LLC_ACCESS
2091 #define KNL_L2_MISS (KNL_OT_L2_HITE | KNL_OT_L2_HITF | \
2092 KNL_DRAM_ANY | SNB_SNP_ANY | \
2095 static __initconst const u64 knl_hw_cache_extra_regs
2096 [PERF_COUNT_HW_CACHE_MAX]
2097 [PERF_COUNT_HW_CACHE_OP_MAX]
2098 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
2101 [C(RESULT_ACCESS)] = KNL_L2_READ | KNL_L2_ACCESS,
2102 [C(RESULT_MISS)] = 0,
2105 [C(RESULT_ACCESS)] = KNL_L2_WRITE | KNL_L2_ACCESS,
2106 [C(RESULT_MISS)] = KNL_L2_WRITE | KNL_L2_MISS,
2108 [C(OP_PREFETCH)] = {
2109 [C(RESULT_ACCESS)] = KNL_L2_PREFETCH | KNL_L2_ACCESS,
2110 [C(RESULT_MISS)] = KNL_L2_PREFETCH | KNL_L2_MISS,
2116 * Used from PMIs where the LBRs are already disabled.
2118 * This function could be called consecutively. It is required to remain in
2119 * disabled state if called consecutively.
2121 * During consecutive calls, the same disable value will be written to related
2122 * registers, so the PMU state remains unchanged.
2124 * intel_bts events don't coexist with intel PMU's BTS events because of
2125 * x86_add_exclusive(x86_lbr_exclusive_lbr); there's no need to keep them
2126 * disabled around intel PMU's event batching etc, only inside the PMI handler.
2128 * Avoid PEBS_ENABLE MSR access in PMIs.
2129 * The GLOBAL_CTRL has been disabled. All the counters do not count anymore.
2130 * It doesn't matter if the PEBS is enabled or not.
2131 * Usually, the PEBS status are not changed in PMIs. It's unnecessary to
2132 * access PEBS_ENABLE MSR in disable_all()/enable_all().
2133 * However, there are some cases which may change PEBS status, e.g. PMI
2134 * throttle. The PEBS_ENABLE should be updated where the status changes.
2136 static void __intel_pmu_disable_all(void)
2138 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2140 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
2142 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask))
2143 intel_pmu_disable_bts();
2146 static void intel_pmu_disable_all(void)
2148 __intel_pmu_disable_all();
2149 intel_pmu_pebs_disable_all();
2150 intel_pmu_lbr_disable_all();
2153 static void __intel_pmu_enable_all(int added, bool pmi)
2155 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2156 u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl);
2158 intel_pmu_lbr_enable_all(pmi);
2159 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL,
2160 intel_ctrl & ~cpuc->intel_ctrl_guest_mask);
2162 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
2163 struct perf_event *event =
2164 cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
2166 if (WARN_ON_ONCE(!event))
2169 intel_pmu_enable_bts(event->hw.config);
2173 static void intel_pmu_enable_all(int added)
2175 intel_pmu_pebs_enable_all();
2176 __intel_pmu_enable_all(added, false);
2181 * Intel Errata AAK100 (model 26)
2182 * Intel Errata AAP53 (model 30)
2183 * Intel Errata BD53 (model 44)
2185 * The official story:
2186 * These chips need to be 'reset' when adding counters by programming the
2187 * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either
2188 * in sequence on the same PMC or on different PMCs.
2190 * In practise it appears some of these events do in fact count, and
2191 * we need to program all 4 events.
2193 static void intel_pmu_nhm_workaround(void)
2195 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2196 static const unsigned long nhm_magic[4] = {
2202 struct perf_event *event;
2206 * The Errata requires below steps:
2207 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL;
2208 * 2) Configure 4 PERFEVTSELx with the magic events and clear
2209 * the corresponding PMCx;
2210 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL;
2211 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL;
2212 * 5) Clear 4 pairs of ERFEVTSELx and PMCx;
2216 * The real steps we choose are a little different from above.
2217 * A) To reduce MSR operations, we don't run step 1) as they
2218 * are already cleared before this function is called;
2219 * B) Call x86_perf_event_update to save PMCx before configuring
2220 * PERFEVTSELx with magic number;
2221 * C) With step 5), we do clear only when the PERFEVTSELx is
2222 * not used currently.
2223 * D) Call x86_perf_event_set_period to restore PMCx;
2226 /* We always operate 4 pairs of PERF Counters */
2227 for (i = 0; i < 4; i++) {
2228 event = cpuc->events[i];
2230 x86_perf_event_update(event);
2233 for (i = 0; i < 4; i++) {
2234 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]);
2235 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0);
2238 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf);
2239 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
2241 for (i = 0; i < 4; i++) {
2242 event = cpuc->events[i];
2245 x86_perf_event_set_period(event);
2246 __x86_pmu_enable_event(&event->hw,
2247 ARCH_PERFMON_EVENTSEL_ENABLE);
2249 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0);
2253 static void intel_pmu_nhm_enable_all(int added)
2256 intel_pmu_nhm_workaround();
2257 intel_pmu_enable_all(added);
2260 static void intel_set_tfa(struct cpu_hw_events *cpuc, bool on)
2262 u64 val = on ? MSR_TFA_RTM_FORCE_ABORT : 0;
2264 if (cpuc->tfa_shadow != val) {
2265 cpuc->tfa_shadow = val;
2266 wrmsrl(MSR_TSX_FORCE_ABORT, val);
2270 static void intel_tfa_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr)
2273 * We're going to use PMC3, make sure TFA is set before we touch it.
2276 intel_set_tfa(cpuc, true);
2279 static void intel_tfa_pmu_enable_all(int added)
2281 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2284 * If we find PMC3 is no longer used when we enable the PMU, we can
2287 if (!test_bit(3, cpuc->active_mask))
2288 intel_set_tfa(cpuc, false);
2290 intel_pmu_enable_all(added);
2293 static inline u64 intel_pmu_get_status(void)
2297 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
2302 static inline void intel_pmu_ack_status(u64 ack)
2304 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
2307 static inline bool event_is_checkpointed(struct perf_event *event)
2309 return unlikely(event->hw.config & HSW_IN_TX_CHECKPOINTED) != 0;
2312 static inline void intel_set_masks(struct perf_event *event, int idx)
2314 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2316 if (event->attr.exclude_host)
2317 __set_bit(idx, (unsigned long *)&cpuc->intel_ctrl_guest_mask);
2318 if (event->attr.exclude_guest)
2319 __set_bit(idx, (unsigned long *)&cpuc->intel_ctrl_host_mask);
2320 if (event_is_checkpointed(event))
2321 __set_bit(idx, (unsigned long *)&cpuc->intel_cp_status);
2324 static inline void intel_clear_masks(struct perf_event *event, int idx)
2326 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2328 __clear_bit(idx, (unsigned long *)&cpuc->intel_ctrl_guest_mask);
2329 __clear_bit(idx, (unsigned long *)&cpuc->intel_ctrl_host_mask);
2330 __clear_bit(idx, (unsigned long *)&cpuc->intel_cp_status);
2333 static void intel_pmu_disable_fixed(struct perf_event *event)
2335 struct hw_perf_event *hwc = &event->hw;
2339 if (is_topdown_idx(idx)) {
2340 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2343 * When there are other active TopDown events,
2344 * don't disable the fixed counter 3.
2346 if (*(u64 *)cpuc->active_mask & INTEL_PMC_OTHER_TOPDOWN_BITS(idx))
2348 idx = INTEL_PMC_IDX_FIXED_SLOTS;
2351 intel_clear_masks(event, idx);
2353 mask = 0xfULL << ((idx - INTEL_PMC_IDX_FIXED) * 4);
2354 rdmsrl(hwc->config_base, ctrl_val);
2356 wrmsrl(hwc->config_base, ctrl_val);
2359 static void intel_pmu_disable_event(struct perf_event *event)
2361 struct hw_perf_event *hwc = &event->hw;
2365 case 0 ... INTEL_PMC_IDX_FIXED - 1:
2366 intel_clear_masks(event, idx);
2367 x86_pmu_disable_event(event);
2369 case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS - 1:
2370 case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END:
2371 intel_pmu_disable_fixed(event);
2373 case INTEL_PMC_IDX_FIXED_BTS:
2374 intel_pmu_disable_bts();
2375 intel_pmu_drain_bts_buffer();
2377 case INTEL_PMC_IDX_FIXED_VLBR:
2378 intel_clear_masks(event, idx);
2381 intel_clear_masks(event, idx);
2382 pr_warn("Failed to disable the event with invalid index %d\n",
2388 * Needs to be called after x86_pmu_disable_event,
2389 * so we don't trigger the event without PEBS bit set.
2391 if (unlikely(event->attr.precise_ip))
2392 intel_pmu_pebs_disable(event);
2395 static void intel_pmu_del_event(struct perf_event *event)
2397 if (needs_branch_stack(event))
2398 intel_pmu_lbr_del(event);
2399 if (event->attr.precise_ip)
2400 intel_pmu_pebs_del(event);
2403 static int icl_set_topdown_event_period(struct perf_event *event)
2405 struct hw_perf_event *hwc = &event->hw;
2406 s64 left = local64_read(&hwc->period_left);
2409 * The values in PERF_METRICS MSR are derived from fixed counter 3.
2410 * Software should start both registers, PERF_METRICS and fixed
2411 * counter 3, from zero.
2412 * Clear PERF_METRICS and Fixed counter 3 in initialization.
2413 * After that, both MSRs will be cleared for each read.
2414 * Don't need to clear them again.
2416 if (left == x86_pmu.max_period) {
2417 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, 0);
2418 wrmsrl(MSR_PERF_METRICS, 0);
2419 hwc->saved_slots = 0;
2420 hwc->saved_metric = 0;
2423 if ((hwc->saved_slots) && is_slots_event(event)) {
2424 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, hwc->saved_slots);
2425 wrmsrl(MSR_PERF_METRICS, hwc->saved_metric);
2428 perf_event_update_userpage(event);
2433 static inline u64 icl_get_metrics_event_value(u64 metric, u64 slots, int idx)
2438 * The metric is reported as an 8bit integer fraction
2439 * suming up to 0xff.
2440 * slots-in-metric = (Metric / 0xff) * slots
2442 val = (metric >> ((idx - INTEL_PMC_IDX_METRIC_BASE) * 8)) & 0xff;
2443 return mul_u64_u32_div(slots, val, 0xff);
2446 static u64 icl_get_topdown_value(struct perf_event *event,
2447 u64 slots, u64 metrics)
2449 int idx = event->hw.idx;
2452 if (is_metric_idx(idx))
2453 delta = icl_get_metrics_event_value(metrics, slots, idx);
2460 static void __icl_update_topdown_event(struct perf_event *event,
2461 u64 slots, u64 metrics,
2462 u64 last_slots, u64 last_metrics)
2464 u64 delta, last = 0;
2466 delta = icl_get_topdown_value(event, slots, metrics);
2468 last = icl_get_topdown_value(event, last_slots, last_metrics);
2471 * The 8bit integer fraction of metric may be not accurate,
2472 * especially when the changes is very small.
2473 * For example, if only a few bad_spec happens, the fraction
2474 * may be reduced from 1 to 0. If so, the bad_spec event value
2475 * will be 0 which is definitely less than the last value.
2476 * Avoid update event->count for this case.
2480 local64_add(delta, &event->count);
2484 static void update_saved_topdown_regs(struct perf_event *event, u64 slots,
2485 u64 metrics, int metric_end)
2487 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2488 struct perf_event *other;
2491 event->hw.saved_slots = slots;
2492 event->hw.saved_metric = metrics;
2494 for_each_set_bit(idx, cpuc->active_mask, metric_end + 1) {
2495 if (!is_topdown_idx(idx))
2497 other = cpuc->events[idx];
2498 other->hw.saved_slots = slots;
2499 other->hw.saved_metric = metrics;
2504 * Update all active Topdown events.
2506 * The PERF_METRICS and Fixed counter 3 are read separately. The values may be
2507 * modify by a NMI. PMU has to be disabled before calling this function.
2510 static u64 intel_update_topdown_event(struct perf_event *event, int metric_end)
2512 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2513 struct perf_event *other;
2518 /* read Fixed counter 3 */
2519 rdpmcl((3 | INTEL_PMC_FIXED_RDPMC_BASE), slots);
2523 /* read PERF_METRICS */
2524 rdpmcl(INTEL_PMC_FIXED_RDPMC_METRICS, metrics);
2526 for_each_set_bit(idx, cpuc->active_mask, metric_end + 1) {
2527 if (!is_topdown_idx(idx))
2529 other = cpuc->events[idx];
2530 __icl_update_topdown_event(other, slots, metrics,
2531 event ? event->hw.saved_slots : 0,
2532 event ? event->hw.saved_metric : 0);
2536 * Check and update this event, which may have been cleared
2537 * in active_mask e.g. x86_pmu_stop()
2539 if (event && !test_bit(event->hw.idx, cpuc->active_mask)) {
2540 __icl_update_topdown_event(event, slots, metrics,
2541 event->hw.saved_slots,
2542 event->hw.saved_metric);
2545 * In x86_pmu_stop(), the event is cleared in active_mask first,
2546 * then drain the delta, which indicates context switch for
2548 * Save metric and slots for context switch.
2549 * Don't need to reset the PERF_METRICS and Fixed counter 3.
2550 * Because the values will be restored in next schedule in.
2552 update_saved_topdown_regs(event, slots, metrics, metric_end);
2557 /* The fixed counter 3 has to be written before the PERF_METRICS. */
2558 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, 0);
2559 wrmsrl(MSR_PERF_METRICS, 0);
2561 update_saved_topdown_regs(event, 0, 0, metric_end);
2567 static u64 icl_update_topdown_event(struct perf_event *event)
2569 return intel_update_topdown_event(event, INTEL_PMC_IDX_METRIC_BASE +
2570 x86_pmu.num_topdown_events - 1);
2573 static void intel_pmu_read_topdown_event(struct perf_event *event)
2575 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2577 /* Only need to call update_topdown_event() once for group read. */
2578 if ((cpuc->txn_flags & PERF_PMU_TXN_READ) &&
2579 !is_slots_event(event))
2582 perf_pmu_disable(event->pmu);
2583 x86_pmu.update_topdown_event(event);
2584 perf_pmu_enable(event->pmu);
2587 static void intel_pmu_read_event(struct perf_event *event)
2589 if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
2590 intel_pmu_auto_reload_read(event);
2591 else if (is_topdown_count(event) && x86_pmu.update_topdown_event)
2592 intel_pmu_read_topdown_event(event);
2594 x86_perf_event_update(event);
2597 static void intel_pmu_enable_fixed(struct perf_event *event)
2599 struct hw_perf_event *hwc = &event->hw;
2600 u64 ctrl_val, mask, bits = 0;
2603 if (is_topdown_idx(idx)) {
2604 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2606 * When there are other active TopDown events,
2607 * don't enable the fixed counter 3 again.
2609 if (*(u64 *)cpuc->active_mask & INTEL_PMC_OTHER_TOPDOWN_BITS(idx))
2612 idx = INTEL_PMC_IDX_FIXED_SLOTS;
2615 intel_set_masks(event, idx);
2618 * Enable IRQ generation (0x8), if not PEBS,
2619 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
2622 if (!event->attr.precise_ip)
2624 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
2626 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
2630 * ANY bit is supported in v3 and up
2632 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
2635 idx -= INTEL_PMC_IDX_FIXED;
2637 mask = 0xfULL << (idx * 4);
2639 if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip) {
2640 bits |= ICL_FIXED_0_ADAPTIVE << (idx * 4);
2641 mask |= ICL_FIXED_0_ADAPTIVE << (idx * 4);
2644 rdmsrl(hwc->config_base, ctrl_val);
2647 wrmsrl(hwc->config_base, ctrl_val);
2650 static void intel_pmu_enable_event(struct perf_event *event)
2652 struct hw_perf_event *hwc = &event->hw;
2655 if (unlikely(event->attr.precise_ip))
2656 intel_pmu_pebs_enable(event);
2659 case 0 ... INTEL_PMC_IDX_FIXED - 1:
2660 intel_set_masks(event, idx);
2661 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
2663 case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS - 1:
2664 case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END:
2665 intel_pmu_enable_fixed(event);
2667 case INTEL_PMC_IDX_FIXED_BTS:
2668 if (!__this_cpu_read(cpu_hw_events.enabled))
2670 intel_pmu_enable_bts(hwc->config);
2672 case INTEL_PMC_IDX_FIXED_VLBR:
2673 intel_set_masks(event, idx);
2676 pr_warn("Failed to enable the event with invalid index %d\n",
2681 static void intel_pmu_add_event(struct perf_event *event)
2683 if (event->attr.precise_ip)
2684 intel_pmu_pebs_add(event);
2685 if (needs_branch_stack(event))
2686 intel_pmu_lbr_add(event);
2690 * Save and restart an expired event. Called by NMI contexts,
2691 * so it has to be careful about preempting normal event ops:
2693 int intel_pmu_save_and_restart(struct perf_event *event)
2695 x86_perf_event_update(event);
2697 * For a checkpointed counter always reset back to 0. This
2698 * avoids a situation where the counter overflows, aborts the
2699 * transaction and is then set back to shortly before the
2700 * overflow, and overflows and aborts again.
2702 if (unlikely(event_is_checkpointed(event))) {
2703 /* No race with NMIs because the counter should not be armed */
2704 wrmsrl(event->hw.event_base, 0);
2705 local64_set(&event->hw.prev_count, 0);
2707 return x86_perf_event_set_period(event);
2710 static void intel_pmu_reset(void)
2712 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
2713 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2714 int num_counters_fixed = hybrid(cpuc->pmu, num_counters_fixed);
2715 int num_counters = hybrid(cpuc->pmu, num_counters);
2716 unsigned long flags;
2722 local_irq_save(flags);
2724 pr_info("clearing PMU state on CPU#%d\n", smp_processor_id());
2726 for (idx = 0; idx < num_counters; idx++) {
2727 wrmsrl_safe(x86_pmu_config_addr(idx), 0ull);
2728 wrmsrl_safe(x86_pmu_event_addr(idx), 0ull);
2730 for (idx = 0; idx < num_counters_fixed; idx++) {
2731 if (fixed_counter_disabled(idx, cpuc->pmu))
2733 wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
2737 ds->bts_index = ds->bts_buffer_base;
2739 /* Ack all overflows and disable fixed counters */
2740 if (x86_pmu.version >= 2) {
2741 intel_pmu_ack_status(intel_pmu_get_status());
2742 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
2745 /* Reset LBRs and LBR freezing */
2746 if (x86_pmu.lbr_nr) {
2747 update_debugctlmsr(get_debugctlmsr() &
2748 ~(DEBUGCTLMSR_FREEZE_LBRS_ON_PMI|DEBUGCTLMSR_LBR));
2751 local_irq_restore(flags);
2754 static int handle_pmi_common(struct pt_regs *regs, u64 status)
2756 struct perf_sample_data data;
2757 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2760 u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl);
2762 inc_irq_stat(apic_perf_irqs);
2765 * Ignore a range of extra bits in status that do not indicate
2766 * overflow by themselves.
2768 status &= ~(GLOBAL_STATUS_COND_CHG |
2769 GLOBAL_STATUS_ASIF |
2770 GLOBAL_STATUS_LBRS_FROZEN);
2774 * In case multiple PEBS events are sampled at the same time,
2775 * it is possible to have GLOBAL_STATUS bit 62 set indicating
2776 * PEBS buffer overflow and also seeing at most 3 PEBS counters
2777 * having their bits set in the status register. This is a sign
2778 * that there was at least one PEBS record pending at the time
2779 * of the PMU interrupt. PEBS counters must only be processed
2780 * via the drain_pebs() calls and not via the regular sample
2781 * processing loop coming after that the function, otherwise
2782 * phony regular samples may be generated in the sampling buffer
2783 * not marked with the EXACT tag. Another possibility is to have
2784 * one PEBS event and at least one non-PEBS event whic hoverflows
2785 * while PEBS has armed. In this case, bit 62 of GLOBAL_STATUS will
2786 * not be set, yet the overflow status bit for the PEBS counter will
2789 * To avoid this problem, we systematically ignore the PEBS-enabled
2790 * counters from the GLOBAL_STATUS mask and we always process PEBS
2791 * events via drain_pebs().
2793 if (x86_pmu.flags & PMU_FL_PEBS_ALL)
2794 status &= ~cpuc->pebs_enabled;
2796 status &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK);
2799 * PEBS overflow sets bit 62 in the global status register
2801 if (__test_and_clear_bit(GLOBAL_STATUS_BUFFER_OVF_BIT, (unsigned long *)&status)) {
2802 u64 pebs_enabled = cpuc->pebs_enabled;
2805 x86_pmu.drain_pebs(regs, &data);
2806 status &= intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI;
2809 * PMI throttle may be triggered, which stops the PEBS event.
2810 * Although cpuc->pebs_enabled is updated accordingly, the
2811 * MSR_IA32_PEBS_ENABLE is not updated. Because the
2812 * cpuc->enabled has been forced to 0 in PMI.
2813 * Update the MSR if pebs_enabled is changed.
2815 if (pebs_enabled != cpuc->pebs_enabled)
2816 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
2822 if (__test_and_clear_bit(GLOBAL_STATUS_TRACE_TOPAPMI_BIT, (unsigned long *)&status)) {
2824 if (unlikely(perf_guest_cbs && perf_guest_cbs->is_in_guest() &&
2825 perf_guest_cbs->handle_intel_pt_intr))
2826 perf_guest_cbs->handle_intel_pt_intr();
2828 intel_pt_interrupt();
2832 * Intel Perf mertrics
2834 if (__test_and_clear_bit(GLOBAL_STATUS_PERF_METRICS_OVF_BIT, (unsigned long *)&status)) {
2836 if (x86_pmu.update_topdown_event)
2837 x86_pmu.update_topdown_event(NULL);
2841 * Checkpointed counters can lead to 'spurious' PMIs because the
2842 * rollback caused by the PMI will have cleared the overflow status
2843 * bit. Therefore always force probe these counters.
2845 status |= cpuc->intel_cp_status;
2847 for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
2848 struct perf_event *event = cpuc->events[bit];
2852 if (!test_bit(bit, cpuc->active_mask))
2855 if (!intel_pmu_save_and_restart(event))
2858 perf_sample_data_init(&data, 0, event->hw.last_period);
2860 if (has_branch_stack(event))
2861 data.br_stack = &cpuc->lbr_stack;
2863 if (perf_event_overflow(event, &data, regs))
2864 x86_pmu_stop(event, 0);
2871 * This handler is triggered by the local APIC, so the APIC IRQ handling
2874 static int intel_pmu_handle_irq(struct pt_regs *regs)
2876 struct cpu_hw_events *cpuc;
2882 cpuc = this_cpu_ptr(&cpu_hw_events);
2885 * Save the PMU state.
2886 * It needs to be restored when leaving the handler.
2888 pmu_enabled = cpuc->enabled;
2890 * No known reason to not always do late ACK,
2891 * but just in case do it opt-in.
2893 if (!x86_pmu.late_ack)
2894 apic_write(APIC_LVTPC, APIC_DM_NMI);
2895 intel_bts_disable_local();
2897 __intel_pmu_disable_all();
2898 handled = intel_pmu_drain_bts_buffer();
2899 handled += intel_bts_interrupt();
2900 status = intel_pmu_get_status();
2906 intel_pmu_lbr_read();
2907 intel_pmu_ack_status(status);
2908 if (++loops > 100) {
2912 WARN(1, "perfevents: irq loop stuck!\n");
2913 perf_event_print_debug();
2920 handled += handle_pmi_common(regs, status);
2923 * Repeat if there is more work to be done:
2925 status = intel_pmu_get_status();
2930 /* Only restore PMU state when it's active. See x86_pmu_disable(). */
2931 cpuc->enabled = pmu_enabled;
2933 __intel_pmu_enable_all(0, true);
2934 intel_bts_enable_local();
2937 * Only unmask the NMI after the overflow counters
2938 * have been reset. This avoids spurious NMIs on
2941 if (x86_pmu.late_ack)
2942 apic_write(APIC_LVTPC, APIC_DM_NMI);
2946 static struct event_constraint *
2947 intel_bts_constraints(struct perf_event *event)
2949 if (unlikely(intel_pmu_has_bts(event)))
2950 return &bts_constraint;
2956 * Note: matches a fake event, like Fixed2.
2958 static struct event_constraint *
2959 intel_vlbr_constraints(struct perf_event *event)
2961 struct event_constraint *c = &vlbr_constraint;
2963 if (unlikely(constraint_match(c, event->hw.config)))
2969 static int intel_alt_er(struct cpu_hw_events *cpuc,
2970 int idx, u64 config)
2972 struct extra_reg *extra_regs = hybrid(cpuc->pmu, extra_regs);
2975 if (!(x86_pmu.flags & PMU_FL_HAS_RSP_1))
2978 if (idx == EXTRA_REG_RSP_0)
2979 alt_idx = EXTRA_REG_RSP_1;
2981 if (idx == EXTRA_REG_RSP_1)
2982 alt_idx = EXTRA_REG_RSP_0;
2984 if (config & ~extra_regs[alt_idx].valid_mask)
2990 static void intel_fixup_er(struct perf_event *event, int idx)
2992 struct extra_reg *extra_regs = hybrid(event->pmu, extra_regs);
2993 event->hw.extra_reg.idx = idx;
2995 if (idx == EXTRA_REG_RSP_0) {
2996 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
2997 event->hw.config |= extra_regs[EXTRA_REG_RSP_0].event;
2998 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0;
2999 } else if (idx == EXTRA_REG_RSP_1) {
3000 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
3001 event->hw.config |= extra_regs[EXTRA_REG_RSP_1].event;
3002 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1;
3007 * manage allocation of shared extra msr for certain events
3010 * per-cpu: to be shared between the various events on a single PMU
3011 * per-core: per-cpu + shared by HT threads
3013 static struct event_constraint *
3014 __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc,
3015 struct perf_event *event,
3016 struct hw_perf_event_extra *reg)
3018 struct event_constraint *c = &emptyconstraint;
3019 struct er_account *era;
3020 unsigned long flags;
3024 * reg->alloc can be set due to existing state, so for fake cpuc we
3025 * need to ignore this, otherwise we might fail to allocate proper fake
3026 * state for this extra reg constraint. Also see the comment below.
3028 if (reg->alloc && !cpuc->is_fake)
3029 return NULL; /* call x86_get_event_constraint() */
3032 era = &cpuc->shared_regs->regs[idx];
3034 * we use spin_lock_irqsave() to avoid lockdep issues when
3035 * passing a fake cpuc
3037 raw_spin_lock_irqsave(&era->lock, flags);
3039 if (!atomic_read(&era->ref) || era->config == reg->config) {
3042 * If its a fake cpuc -- as per validate_{group,event}() we
3043 * shouldn't touch event state and we can avoid doing so
3044 * since both will only call get_event_constraints() once
3045 * on each event, this avoids the need for reg->alloc.
3047 * Not doing the ER fixup will only result in era->reg being
3048 * wrong, but since we won't actually try and program hardware
3049 * this isn't a problem either.
3051 if (!cpuc->is_fake) {
3052 if (idx != reg->idx)
3053 intel_fixup_er(event, idx);
3056 * x86_schedule_events() can call get_event_constraints()
3057 * multiple times on events in the case of incremental
3058 * scheduling(). reg->alloc ensures we only do the ER
3064 /* lock in msr value */
3065 era->config = reg->config;
3066 era->reg = reg->reg;
3069 atomic_inc(&era->ref);
3072 * need to call x86_get_event_constraint()
3073 * to check if associated event has constraints
3077 idx = intel_alt_er(cpuc, idx, reg->config);
3078 if (idx != reg->idx) {
3079 raw_spin_unlock_irqrestore(&era->lock, flags);
3083 raw_spin_unlock_irqrestore(&era->lock, flags);
3089 __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc,
3090 struct hw_perf_event_extra *reg)
3092 struct er_account *era;
3095 * Only put constraint if extra reg was actually allocated. Also takes
3096 * care of event which do not use an extra shared reg.
3098 * Also, if this is a fake cpuc we shouldn't touch any event state
3099 * (reg->alloc) and we don't care about leaving inconsistent cpuc state
3100 * either since it'll be thrown out.
3102 if (!reg->alloc || cpuc->is_fake)
3105 era = &cpuc->shared_regs->regs[reg->idx];
3107 /* one fewer user */
3108 atomic_dec(&era->ref);
3110 /* allocate again next time */
3114 static struct event_constraint *
3115 intel_shared_regs_constraints(struct cpu_hw_events *cpuc,
3116 struct perf_event *event)
3118 struct event_constraint *c = NULL, *d;
3119 struct hw_perf_event_extra *xreg, *breg;
3121 xreg = &event->hw.extra_reg;
3122 if (xreg->idx != EXTRA_REG_NONE) {
3123 c = __intel_shared_reg_get_constraints(cpuc, event, xreg);
3124 if (c == &emptyconstraint)
3127 breg = &event->hw.branch_reg;
3128 if (breg->idx != EXTRA_REG_NONE) {
3129 d = __intel_shared_reg_get_constraints(cpuc, event, breg);
3130 if (d == &emptyconstraint) {
3131 __intel_shared_reg_put_constraints(cpuc, xreg);
3138 struct event_constraint *
3139 x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3140 struct perf_event *event)
3142 struct event_constraint *event_constraints = hybrid(cpuc->pmu, event_constraints);
3143 struct event_constraint *c;
3145 if (event_constraints) {
3146 for_each_event_constraint(c, event_constraints) {
3147 if (constraint_match(c, event->hw.config)) {
3148 event->hw.flags |= c->flags;
3154 return &hybrid_var(cpuc->pmu, unconstrained);
3157 static struct event_constraint *
3158 __intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3159 struct perf_event *event)
3161 struct event_constraint *c;
3163 c = intel_vlbr_constraints(event);
3167 c = intel_bts_constraints(event);
3171 c = intel_shared_regs_constraints(cpuc, event);
3175 c = intel_pebs_constraints(event);
3179 return x86_get_event_constraints(cpuc, idx, event);
3183 intel_start_scheduling(struct cpu_hw_events *cpuc)
3185 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3186 struct intel_excl_states *xl;
3187 int tid = cpuc->excl_thread_id;
3190 * nothing needed if in group validation mode
3192 if (cpuc->is_fake || !is_ht_workaround_enabled())
3196 * no exclusion needed
3198 if (WARN_ON_ONCE(!excl_cntrs))
3201 xl = &excl_cntrs->states[tid];
3203 xl->sched_started = true;
3205 * lock shared state until we are done scheduling
3206 * in stop_event_scheduling()
3207 * makes scheduling appear as a transaction
3209 raw_spin_lock(&excl_cntrs->lock);
3212 static void intel_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr)
3214 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3215 struct event_constraint *c = cpuc->event_constraint[idx];
3216 struct intel_excl_states *xl;
3217 int tid = cpuc->excl_thread_id;
3219 if (cpuc->is_fake || !is_ht_workaround_enabled())
3222 if (WARN_ON_ONCE(!excl_cntrs))
3225 if (!(c->flags & PERF_X86_EVENT_DYNAMIC))
3228 xl = &excl_cntrs->states[tid];
3230 lockdep_assert_held(&excl_cntrs->lock);
3232 if (c->flags & PERF_X86_EVENT_EXCL)
3233 xl->state[cntr] = INTEL_EXCL_EXCLUSIVE;
3235 xl->state[cntr] = INTEL_EXCL_SHARED;
3239 intel_stop_scheduling(struct cpu_hw_events *cpuc)
3241 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3242 struct intel_excl_states *xl;
3243 int tid = cpuc->excl_thread_id;
3246 * nothing needed if in group validation mode
3248 if (cpuc->is_fake || !is_ht_workaround_enabled())
3251 * no exclusion needed
3253 if (WARN_ON_ONCE(!excl_cntrs))
3256 xl = &excl_cntrs->states[tid];
3258 xl->sched_started = false;
3260 * release shared state lock (acquired in intel_start_scheduling())
3262 raw_spin_unlock(&excl_cntrs->lock);
3265 static struct event_constraint *
3266 dyn_constraint(struct cpu_hw_events *cpuc, struct event_constraint *c, int idx)
3268 WARN_ON_ONCE(!cpuc->constraint_list);
3270 if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) {
3271 struct event_constraint *cx;
3274 * grab pre-allocated constraint entry
3276 cx = &cpuc->constraint_list[idx];
3279 * initialize dynamic constraint
3280 * with static constraint
3285 * mark constraint as dynamic
3287 cx->flags |= PERF_X86_EVENT_DYNAMIC;
3294 static struct event_constraint *
3295 intel_get_excl_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
3296 int idx, struct event_constraint *c)
3298 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3299 struct intel_excl_states *xlo;
3300 int tid = cpuc->excl_thread_id;
3304 * validating a group does not require
3305 * enforcing cross-thread exclusion
3307 if (cpuc->is_fake || !is_ht_workaround_enabled())
3311 * no exclusion needed
3313 if (WARN_ON_ONCE(!excl_cntrs))
3317 * because we modify the constraint, we need
3318 * to make a copy. Static constraints come
3319 * from static const tables.
3321 * only needed when constraint has not yet
3322 * been cloned (marked dynamic)
3324 c = dyn_constraint(cpuc, c, idx);
3327 * From here on, the constraint is dynamic.
3328 * Either it was just allocated above, or it
3329 * was allocated during a earlier invocation
3334 * state of sibling HT
3336 xlo = &excl_cntrs->states[tid ^ 1];
3339 * event requires exclusive counter access
3342 is_excl = c->flags & PERF_X86_EVENT_EXCL;
3343 if (is_excl && !(event->hw.flags & PERF_X86_EVENT_EXCL_ACCT)) {
3344 event->hw.flags |= PERF_X86_EVENT_EXCL_ACCT;
3345 if (!cpuc->n_excl++)
3346 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 1);
3350 * Modify static constraint with current dynamic
3353 * EXCLUSIVE: sibling counter measuring exclusive event
3354 * SHARED : sibling counter measuring non-exclusive event
3355 * UNUSED : sibling counter unused
3358 for_each_set_bit(i, c->idxmsk, X86_PMC_IDX_MAX) {
3360 * exclusive event in sibling counter
3361 * our corresponding counter cannot be used
3362 * regardless of our event
3364 if (xlo->state[i] == INTEL_EXCL_EXCLUSIVE) {
3365 __clear_bit(i, c->idxmsk);
3370 * if measuring an exclusive event, sibling
3371 * measuring non-exclusive, then counter cannot
3374 if (is_excl && xlo->state[i] == INTEL_EXCL_SHARED) {
3375 __clear_bit(i, c->idxmsk);
3382 * if we return an empty mask, then switch
3383 * back to static empty constraint to avoid
3384 * the cost of freeing later on
3387 c = &emptyconstraint;
3394 static struct event_constraint *
3395 intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3396 struct perf_event *event)
3398 struct event_constraint *c1, *c2;
3400 c1 = cpuc->event_constraint[idx];
3404 * - static constraint: no change across incremental scheduling calls
3405 * - dynamic constraint: handled by intel_get_excl_constraints()
3407 c2 = __intel_get_event_constraints(cpuc, idx, event);
3409 WARN_ON_ONCE(!(c1->flags & PERF_X86_EVENT_DYNAMIC));
3410 bitmap_copy(c1->idxmsk, c2->idxmsk, X86_PMC_IDX_MAX);
3411 c1->weight = c2->weight;
3415 if (cpuc->excl_cntrs)
3416 return intel_get_excl_constraints(cpuc, event, idx, c2);
3421 static void intel_put_excl_constraints(struct cpu_hw_events *cpuc,
3422 struct perf_event *event)
3424 struct hw_perf_event *hwc = &event->hw;
3425 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3426 int tid = cpuc->excl_thread_id;
3427 struct intel_excl_states *xl;
3430 * nothing needed if in group validation mode
3435 if (WARN_ON_ONCE(!excl_cntrs))
3438 if (hwc->flags & PERF_X86_EVENT_EXCL_ACCT) {
3439 hwc->flags &= ~PERF_X86_EVENT_EXCL_ACCT;
3440 if (!--cpuc->n_excl)
3441 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 0);
3445 * If event was actually assigned, then mark the counter state as
3448 if (hwc->idx >= 0) {
3449 xl = &excl_cntrs->states[tid];
3452 * put_constraint may be called from x86_schedule_events()
3453 * which already has the lock held so here make locking
3456 if (!xl->sched_started)
3457 raw_spin_lock(&excl_cntrs->lock);
3459 xl->state[hwc->idx] = INTEL_EXCL_UNUSED;
3461 if (!xl->sched_started)
3462 raw_spin_unlock(&excl_cntrs->lock);
3467 intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc,
3468 struct perf_event *event)
3470 struct hw_perf_event_extra *reg;
3472 reg = &event->hw.extra_reg;
3473 if (reg->idx != EXTRA_REG_NONE)
3474 __intel_shared_reg_put_constraints(cpuc, reg);
3476 reg = &event->hw.branch_reg;
3477 if (reg->idx != EXTRA_REG_NONE)
3478 __intel_shared_reg_put_constraints(cpuc, reg);
3481 static void intel_put_event_constraints(struct cpu_hw_events *cpuc,
3482 struct perf_event *event)
3484 intel_put_shared_regs_event_constraints(cpuc, event);
3487 * is PMU has exclusive counter restrictions, then
3488 * all events are subject to and must call the
3489 * put_excl_constraints() routine
3491 if (cpuc->excl_cntrs)
3492 intel_put_excl_constraints(cpuc, event);
3495 static void intel_pebs_aliases_core2(struct perf_event *event)
3497 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
3499 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
3500 * (0x003c) so that we can use it with PEBS.
3502 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
3503 * PEBS capable. However we can use INST_RETIRED.ANY_P
3504 * (0x00c0), which is a PEBS capable event, to get the same
3507 * INST_RETIRED.ANY_P counts the number of cycles that retires
3508 * CNTMASK instructions. By setting CNTMASK to a value (16)
3509 * larger than the maximum number of instructions that can be
3510 * retired per cycle (4) and then inverting the condition, we
3511 * count all cycles that retire 16 or less instructions, which
3514 * Thereby we gain a PEBS capable cycle counter.
3516 u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16);
3518 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
3519 event->hw.config = alt_config;
3523 static void intel_pebs_aliases_snb(struct perf_event *event)
3525 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
3527 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
3528 * (0x003c) so that we can use it with PEBS.
3530 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
3531 * PEBS capable. However we can use UOPS_RETIRED.ALL
3532 * (0x01c2), which is a PEBS capable event, to get the same
3535 * UOPS_RETIRED.ALL counts the number of cycles that retires
3536 * CNTMASK micro-ops. By setting CNTMASK to a value (16)
3537 * larger than the maximum number of micro-ops that can be
3538 * retired per cycle (4) and then inverting the condition, we
3539 * count all cycles that retire 16 or less micro-ops, which
3542 * Thereby we gain a PEBS capable cycle counter.
3544 u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16);
3546 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
3547 event->hw.config = alt_config;
3551 static void intel_pebs_aliases_precdist(struct perf_event *event)
3553 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
3555 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
3556 * (0x003c) so that we can use it with PEBS.
3558 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
3559 * PEBS capable. However we can use INST_RETIRED.PREC_DIST
3560 * (0x01c0), which is a PEBS capable event, to get the same
3563 * The PREC_DIST event has special support to minimize sample
3564 * shadowing effects. One drawback is that it can be
3565 * only programmed on counter 1, but that seems like an
3566 * acceptable trade off.
3568 u64 alt_config = X86_CONFIG(.event=0xc0, .umask=0x01, .inv=1, .cmask=16);
3570 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
3571 event->hw.config = alt_config;
3575 static void intel_pebs_aliases_ivb(struct perf_event *event)
3577 if (event->attr.precise_ip < 3)
3578 return intel_pebs_aliases_snb(event);
3579 return intel_pebs_aliases_precdist(event);
3582 static void intel_pebs_aliases_skl(struct perf_event *event)
3584 if (event->attr.precise_ip < 3)
3585 return intel_pebs_aliases_core2(event);
3586 return intel_pebs_aliases_precdist(event);
3589 static unsigned long intel_pmu_large_pebs_flags(struct perf_event *event)
3591 unsigned long flags = x86_pmu.large_pebs_flags;
3593 if (event->attr.use_clockid)
3594 flags &= ~PERF_SAMPLE_TIME;
3595 if (!event->attr.exclude_kernel)
3596 flags &= ~PERF_SAMPLE_REGS_USER;
3597 if (event->attr.sample_regs_user & ~PEBS_GP_REGS)
3598 flags &= ~(PERF_SAMPLE_REGS_USER | PERF_SAMPLE_REGS_INTR);
3602 static int intel_pmu_bts_config(struct perf_event *event)
3604 struct perf_event_attr *attr = &event->attr;
3606 if (unlikely(intel_pmu_has_bts(event))) {
3607 /* BTS is not supported by this architecture. */
3608 if (!x86_pmu.bts_active)
3611 /* BTS is currently only allowed for user-mode. */
3612 if (!attr->exclude_kernel)
3615 /* BTS is not allowed for precise events. */
3616 if (attr->precise_ip)
3619 /* disallow bts if conflicting events are present */
3620 if (x86_add_exclusive(x86_lbr_exclusive_lbr))
3623 event->destroy = hw_perf_lbr_event_destroy;
3629 static int core_pmu_hw_config(struct perf_event *event)
3631 int ret = x86_pmu_hw_config(event);
3636 return intel_pmu_bts_config(event);
3639 #define INTEL_TD_METRIC_AVAILABLE_MAX (INTEL_TD_METRIC_RETIRING + \
3640 ((x86_pmu.num_topdown_events - 1) << 8))
3642 static bool is_available_metric_event(struct perf_event *event)
3644 return is_metric_event(event) &&
3645 event->attr.config <= INTEL_TD_METRIC_AVAILABLE_MAX;
3648 static inline bool is_mem_loads_event(struct perf_event *event)
3650 return (event->attr.config & INTEL_ARCH_EVENT_MASK) == X86_CONFIG(.event=0xcd, .umask=0x01);
3653 static inline bool is_mem_loads_aux_event(struct perf_event *event)
3655 return (event->attr.config & INTEL_ARCH_EVENT_MASK) == X86_CONFIG(.event=0x03, .umask=0x82);
3658 static inline bool intel_pmu_has_cap(struct perf_event *event, int idx)
3660 union perf_capabilities *intel_cap = &hybrid(event->pmu, intel_cap);
3662 return test_bit(idx, (unsigned long *)&intel_cap->capabilities);
3665 static int intel_pmu_hw_config(struct perf_event *event)
3667 int ret = x86_pmu_hw_config(event);
3672 ret = intel_pmu_bts_config(event);
3676 if (event->attr.precise_ip) {
3677 if (!(event->attr.freq || (event->attr.wakeup_events && !event->attr.watermark))) {
3678 event->hw.flags |= PERF_X86_EVENT_AUTO_RELOAD;
3679 if (!(event->attr.sample_type &
3680 ~intel_pmu_large_pebs_flags(event))) {
3681 event->hw.flags |= PERF_X86_EVENT_LARGE_PEBS;
3682 event->attach_state |= PERF_ATTACH_SCHED_CB;
3685 if (x86_pmu.pebs_aliases)
3686 x86_pmu.pebs_aliases(event);
3688 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3689 event->attr.sample_type |= __PERF_SAMPLE_CALLCHAIN_EARLY;
3692 if (needs_branch_stack(event)) {
3693 ret = intel_pmu_setup_lbr_filter(event);
3696 event->attach_state |= PERF_ATTACH_SCHED_CB;
3699 * BTS is set up earlier in this path, so don't account twice
3701 if (!unlikely(intel_pmu_has_bts(event))) {
3702 /* disallow lbr if conflicting events are present */
3703 if (x86_add_exclusive(x86_lbr_exclusive_lbr))
3706 event->destroy = hw_perf_lbr_event_destroy;
3710 if (event->attr.aux_output) {
3711 if (!event->attr.precise_ip)
3714 event->hw.flags |= PERF_X86_EVENT_PEBS_VIA_PT;
3717 if (event->attr.type != PERF_TYPE_RAW)
3721 * Config Topdown slots and metric events
3723 * The slots event on Fixed Counter 3 can support sampling,
3724 * which will be handled normally in x86_perf_event_update().
3726 * Metric events don't support sampling and require being paired
3727 * with a slots event as group leader. When the slots event
3728 * is used in a metrics group, it too cannot support sampling.
3730 if (intel_pmu_has_cap(event, PERF_CAP_METRICS_IDX) && is_topdown_event(event)) {
3731 if (event->attr.config1 || event->attr.config2)
3735 * The TopDown metrics events and slots event don't
3736 * support any filters.
3738 if (event->attr.config & X86_ALL_EVENT_FLAGS)
3741 if (is_available_metric_event(event)) {
3742 struct perf_event *leader = event->group_leader;
3744 /* The metric events don't support sampling. */
3745 if (is_sampling_event(event))
3748 /* The metric events require a slots group leader. */
3749 if (!is_slots_event(leader))
3753 * The leader/SLOTS must not be a sampling event for
3754 * metric use; hardware requires it starts at 0 when used
3755 * in conjunction with MSR_PERF_METRICS.
3757 if (is_sampling_event(leader))
3760 event->event_caps |= PERF_EV_CAP_SIBLING;
3762 * Only once we have a METRICs sibling do we
3763 * need TopDown magic.
3765 leader->hw.flags |= PERF_X86_EVENT_TOPDOWN;
3766 event->hw.flags |= PERF_X86_EVENT_TOPDOWN;
3771 * The load latency event X86_CONFIG(.event=0xcd, .umask=0x01) on SPR
3772 * doesn't function quite right. As a work-around it needs to always be
3773 * co-scheduled with a auxiliary event X86_CONFIG(.event=0x03, .umask=0x82).
3774 * The actual count of this second event is irrelevant it just needs
3775 * to be active to make the first event function correctly.
3777 * In a group, the auxiliary event must be in front of the load latency
3778 * event. The rule is to simplify the implementation of the check.
3779 * That's because perf cannot have a complete group at the moment.
3781 if (x86_pmu.flags & PMU_FL_MEM_LOADS_AUX &&
3782 (event->attr.sample_type & PERF_SAMPLE_DATA_SRC) &&
3783 is_mem_loads_event(event)) {
3784 struct perf_event *leader = event->group_leader;
3785 struct perf_event *sibling = NULL;
3787 if (!is_mem_loads_aux_event(leader)) {
3788 for_each_sibling_event(sibling, leader) {
3789 if (is_mem_loads_aux_event(sibling))
3792 if (list_entry_is_head(sibling, &leader->sibling_list, sibling_list))
3797 if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY))
3800 if (x86_pmu.version < 3)
3803 ret = perf_allow_cpu(&event->attr);
3807 event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY;
3812 static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr)
3814 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3815 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
3816 u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl);
3818 arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL;
3819 arr[0].host = intel_ctrl & ~cpuc->intel_ctrl_guest_mask;
3820 arr[0].guest = intel_ctrl & ~cpuc->intel_ctrl_host_mask;
3821 if (x86_pmu.flags & PMU_FL_PEBS_ALL)
3822 arr[0].guest &= ~cpuc->pebs_enabled;
3824 arr[0].guest &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK);
3827 if (x86_pmu.pebs && x86_pmu.pebs_no_isolation) {
3829 * If PMU counter has PEBS enabled it is not enough to
3830 * disable counter on a guest entry since PEBS memory
3831 * write can overshoot guest entry and corrupt guest
3832 * memory. Disabling PEBS solves the problem.
3834 * Don't do this if the CPU already enforces it.
3836 arr[1].msr = MSR_IA32_PEBS_ENABLE;
3837 arr[1].host = cpuc->pebs_enabled;
3845 static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr)
3847 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3848 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
3851 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
3852 struct perf_event *event = cpuc->events[idx];
3854 arr[idx].msr = x86_pmu_config_addr(idx);
3855 arr[idx].host = arr[idx].guest = 0;
3857 if (!test_bit(idx, cpuc->active_mask))
3860 arr[idx].host = arr[idx].guest =
3861 event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE;
3863 if (event->attr.exclude_host)
3864 arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
3865 else if (event->attr.exclude_guest)
3866 arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
3869 *nr = x86_pmu.num_counters;
3873 static void core_pmu_enable_event(struct perf_event *event)
3875 if (!event->attr.exclude_host)
3876 x86_pmu_enable_event(event);
3879 static void core_pmu_enable_all(int added)
3881 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3884 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
3885 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
3887 if (!test_bit(idx, cpuc->active_mask) ||
3888 cpuc->events[idx]->attr.exclude_host)
3891 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
3895 static int hsw_hw_config(struct perf_event *event)
3897 int ret = intel_pmu_hw_config(event);
3901 if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE))
3903 event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED);
3906 * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with
3907 * PEBS or in ANY thread mode. Since the results are non-sensical forbid
3910 if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) &&
3911 ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) ||
3912 event->attr.precise_ip > 0))
3915 if (event_is_checkpointed(event)) {
3917 * Sampling of checkpointed events can cause situations where
3918 * the CPU constantly aborts because of a overflow, which is
3919 * then checkpointed back and ignored. Forbid checkpointing
3922 * But still allow a long sampling period, so that perf stat
3925 if (event->attr.sample_period > 0 &&
3926 event->attr.sample_period < 0x7fffffff)
3932 static struct event_constraint counter0_constraint =
3933 INTEL_ALL_EVENT_CONSTRAINT(0, 0x1);
3935 static struct event_constraint counter2_constraint =
3936 EVENT_CONSTRAINT(0, 0x4, 0);
3938 static struct event_constraint fixed0_constraint =
3939 FIXED_EVENT_CONSTRAINT(0x00c0, 0);
3941 static struct event_constraint fixed0_counter0_constraint =
3942 INTEL_ALL_EVENT_CONSTRAINT(0, 0x100000001ULL);
3944 static struct event_constraint *
3945 hsw_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3946 struct perf_event *event)
3948 struct event_constraint *c;
3950 c = intel_get_event_constraints(cpuc, idx, event);
3952 /* Handle special quirk on in_tx_checkpointed only in counter 2 */
3953 if (event->hw.config & HSW_IN_TX_CHECKPOINTED) {
3954 if (c->idxmsk64 & (1U << 2))
3955 return &counter2_constraint;
3956 return &emptyconstraint;
3962 static struct event_constraint *
3963 icl_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3964 struct perf_event *event)
3967 * Fixed counter 0 has less skid.
3968 * Force instruction:ppp in Fixed counter 0
3970 if ((event->attr.precise_ip == 3) &&
3971 constraint_match(&fixed0_constraint, event->hw.config))
3972 return &fixed0_constraint;
3974 return hsw_get_event_constraints(cpuc, idx, event);
3977 static struct event_constraint *
3978 spr_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3979 struct perf_event *event)
3981 struct event_constraint *c;
3983 c = icl_get_event_constraints(cpuc, idx, event);
3986 * The :ppp indicates the Precise Distribution (PDist) facility, which
3987 * is only supported on the GP counter 0. If a :ppp event which is not
3988 * available on the GP counter 0, error out.
3990 if (event->attr.precise_ip == 3) {
3991 if (c->idxmsk64 & BIT_ULL(0))
3992 return &counter0_constraint;
3994 return &emptyconstraint;
4000 static struct event_constraint *
4001 glp_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4002 struct perf_event *event)
4004 struct event_constraint *c;
4006 /* :ppp means to do reduced skid PEBS which is PMC0 only. */
4007 if (event->attr.precise_ip == 3)
4008 return &counter0_constraint;
4010 c = intel_get_event_constraints(cpuc, idx, event);
4015 static struct event_constraint *
4016 tnt_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4017 struct perf_event *event)
4019 struct event_constraint *c;
4022 * :ppp means to do reduced skid PEBS,
4023 * which is available on PMC0 and fixed counter 0.
4025 if (event->attr.precise_ip == 3) {
4026 /* Force instruction:ppp on PMC0 and Fixed counter 0 */
4027 if (constraint_match(&fixed0_constraint, event->hw.config))
4028 return &fixed0_counter0_constraint;
4030 return &counter0_constraint;
4033 c = intel_get_event_constraints(cpuc, idx, event);
4038 static bool allow_tsx_force_abort = true;
4040 static struct event_constraint *
4041 tfa_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4042 struct perf_event *event)
4044 struct event_constraint *c = hsw_get_event_constraints(cpuc, idx, event);
4047 * Without TFA we must not use PMC3.
4049 if (!allow_tsx_force_abort && test_bit(3, c->idxmsk)) {
4050 c = dyn_constraint(cpuc, c, idx);
4051 c->idxmsk64 &= ~(1ULL << 3);
4061 * The INST_RETIRED.ALL period always needs to have lowest 6 bits cleared
4062 * (BDM55) and it must not use a period smaller than 100 (BDM11). We combine
4063 * the two to enforce a minimum period of 128 (the smallest value that has bits
4064 * 0-5 cleared and >= 100).
4066 * Because of how the code in x86_perf_event_set_period() works, the truncation
4067 * of the lower 6 bits is 'harmless' as we'll occasionally add a longer period
4068 * to make up for the 'lost' events due to carrying the 'error' in period_left.
4070 * Therefore the effective (average) period matches the requested period,
4071 * despite coarser hardware granularity.
4073 static u64 bdw_limit_period(struct perf_event *event, u64 left)
4075 if ((event->hw.config & INTEL_ARCH_EVENT_MASK) ==
4076 X86_CONFIG(.event=0xc0, .umask=0x01)) {
4084 static u64 nhm_limit_period(struct perf_event *event, u64 left)
4086 return max(left, 32ULL);
4089 static u64 spr_limit_period(struct perf_event *event, u64 left)
4091 if (event->attr.precise_ip == 3)
4092 return max(left, 128ULL);
4097 PMU_FORMAT_ATTR(event, "config:0-7" );
4098 PMU_FORMAT_ATTR(umask, "config:8-15" );
4099 PMU_FORMAT_ATTR(edge, "config:18" );
4100 PMU_FORMAT_ATTR(pc, "config:19" );
4101 PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */
4102 PMU_FORMAT_ATTR(inv, "config:23" );
4103 PMU_FORMAT_ATTR(cmask, "config:24-31" );
4104 PMU_FORMAT_ATTR(in_tx, "config:32");
4105 PMU_FORMAT_ATTR(in_tx_cp, "config:33");
4107 static struct attribute *intel_arch_formats_attr[] = {
4108 &format_attr_event.attr,
4109 &format_attr_umask.attr,
4110 &format_attr_edge.attr,
4111 &format_attr_pc.attr,
4112 &format_attr_inv.attr,
4113 &format_attr_cmask.attr,
4117 ssize_t intel_event_sysfs_show(char *page, u64 config)
4119 u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT);
4121 return x86_event_sysfs_show(page, config, event);
4124 static struct intel_shared_regs *allocate_shared_regs(int cpu)
4126 struct intel_shared_regs *regs;
4129 regs = kzalloc_node(sizeof(struct intel_shared_regs),
4130 GFP_KERNEL, cpu_to_node(cpu));
4133 * initialize the locks to keep lockdep happy
4135 for (i = 0; i < EXTRA_REG_MAX; i++)
4136 raw_spin_lock_init(®s->regs[i].lock);
4143 static struct intel_excl_cntrs *allocate_excl_cntrs(int cpu)
4145 struct intel_excl_cntrs *c;
4147 c = kzalloc_node(sizeof(struct intel_excl_cntrs),
4148 GFP_KERNEL, cpu_to_node(cpu));
4150 raw_spin_lock_init(&c->lock);
4157 int intel_cpuc_prepare(struct cpu_hw_events *cpuc, int cpu)
4159 cpuc->pebs_record_size = x86_pmu.pebs_record_size;
4161 if (is_hybrid() || x86_pmu.extra_regs || x86_pmu.lbr_sel_map) {
4162 cpuc->shared_regs = allocate_shared_regs(cpu);
4163 if (!cpuc->shared_regs)
4167 if (x86_pmu.flags & (PMU_FL_EXCL_CNTRS | PMU_FL_TFA)) {
4168 size_t sz = X86_PMC_IDX_MAX * sizeof(struct event_constraint);
4170 cpuc->constraint_list = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));
4171 if (!cpuc->constraint_list)
4172 goto err_shared_regs;
4175 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) {
4176 cpuc->excl_cntrs = allocate_excl_cntrs(cpu);
4177 if (!cpuc->excl_cntrs)
4178 goto err_constraint_list;
4180 cpuc->excl_thread_id = 0;
4185 err_constraint_list:
4186 kfree(cpuc->constraint_list);
4187 cpuc->constraint_list = NULL;
4190 kfree(cpuc->shared_regs);
4191 cpuc->shared_regs = NULL;
4197 static int intel_pmu_cpu_prepare(int cpu)
4199 return intel_cpuc_prepare(&per_cpu(cpu_hw_events, cpu), cpu);
4202 static void flip_smm_bit(void *data)
4204 unsigned long set = *(unsigned long *)data;
4207 msr_set_bit(MSR_IA32_DEBUGCTLMSR,
4208 DEBUGCTLMSR_FREEZE_IN_SMM_BIT);
4210 msr_clear_bit(MSR_IA32_DEBUGCTLMSR,
4211 DEBUGCTLMSR_FREEZE_IN_SMM_BIT);
4215 static void intel_pmu_cpu_starting(int cpu)
4217 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
4218 int core_id = topology_core_id(cpu);
4221 init_debug_store_on_cpu(cpu);
4223 * Deal with CPUs that don't clear their LBRs on power-up.
4225 intel_pmu_lbr_reset();
4227 cpuc->lbr_sel = NULL;
4229 if (x86_pmu.flags & PMU_FL_TFA) {
4230 WARN_ON_ONCE(cpuc->tfa_shadow);
4231 cpuc->tfa_shadow = ~0ULL;
4232 intel_set_tfa(cpuc, false);
4235 if (x86_pmu.version > 1)
4236 flip_smm_bit(&x86_pmu.attr_freeze_on_smi);
4239 * Disable perf metrics if any added CPU doesn't support it.
4241 * Turn off the check for a hybrid architecture, because the
4242 * architecture MSR, MSR_IA32_PERF_CAPABILITIES, only indicate
4243 * the architecture features. The perf metrics is a model-specific
4244 * feature for now. The corresponding bit should always be 0 on
4245 * a hybrid platform, e.g., Alder Lake.
4247 if (!is_hybrid() && x86_pmu.intel_cap.perf_metrics) {
4248 union perf_capabilities perf_cap;
4250 rdmsrl(MSR_IA32_PERF_CAPABILITIES, perf_cap.capabilities);
4251 if (!perf_cap.perf_metrics) {
4252 x86_pmu.intel_cap.perf_metrics = 0;
4253 x86_pmu.intel_ctrl &= ~(1ULL << GLOBAL_CTRL_EN_PERF_METRICS);
4257 if (!cpuc->shared_regs)
4260 if (!(x86_pmu.flags & PMU_FL_NO_HT_SHARING)) {
4261 for_each_cpu(i, topology_sibling_cpumask(cpu)) {
4262 struct intel_shared_regs *pc;
4264 pc = per_cpu(cpu_hw_events, i).shared_regs;
4265 if (pc && pc->core_id == core_id) {
4266 cpuc->kfree_on_online[0] = cpuc->shared_regs;
4267 cpuc->shared_regs = pc;
4271 cpuc->shared_regs->core_id = core_id;
4272 cpuc->shared_regs->refcnt++;
4275 if (x86_pmu.lbr_sel_map)
4276 cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR];
4278 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) {
4279 for_each_cpu(i, topology_sibling_cpumask(cpu)) {
4280 struct cpu_hw_events *sibling;
4281 struct intel_excl_cntrs *c;
4283 sibling = &per_cpu(cpu_hw_events, i);
4284 c = sibling->excl_cntrs;
4285 if (c && c->core_id == core_id) {
4286 cpuc->kfree_on_online[1] = cpuc->excl_cntrs;
4287 cpuc->excl_cntrs = c;
4288 if (!sibling->excl_thread_id)
4289 cpuc->excl_thread_id = 1;
4293 cpuc->excl_cntrs->core_id = core_id;
4294 cpuc->excl_cntrs->refcnt++;
4298 static void free_excl_cntrs(struct cpu_hw_events *cpuc)
4300 struct intel_excl_cntrs *c;
4302 c = cpuc->excl_cntrs;
4304 if (c->core_id == -1 || --c->refcnt == 0)
4306 cpuc->excl_cntrs = NULL;
4309 kfree(cpuc->constraint_list);
4310 cpuc->constraint_list = NULL;
4313 static void intel_pmu_cpu_dying(int cpu)
4315 fini_debug_store_on_cpu(cpu);
4318 void intel_cpuc_finish(struct cpu_hw_events *cpuc)
4320 struct intel_shared_regs *pc;
4322 pc = cpuc->shared_regs;
4324 if (pc->core_id == -1 || --pc->refcnt == 0)
4326 cpuc->shared_regs = NULL;
4329 free_excl_cntrs(cpuc);
4332 static void intel_pmu_cpu_dead(int cpu)
4334 intel_cpuc_finish(&per_cpu(cpu_hw_events, cpu));
4337 static void intel_pmu_sched_task(struct perf_event_context *ctx,
4340 intel_pmu_pebs_sched_task(ctx, sched_in);
4341 intel_pmu_lbr_sched_task(ctx, sched_in);
4344 static void intel_pmu_swap_task_ctx(struct perf_event_context *prev,
4345 struct perf_event_context *next)
4347 intel_pmu_lbr_swap_task_ctx(prev, next);
4350 static int intel_pmu_check_period(struct perf_event *event, u64 value)
4352 return intel_pmu_has_bts_period(event, value) ? -EINVAL : 0;
4355 static int intel_pmu_aux_output_match(struct perf_event *event)
4357 if (!x86_pmu.intel_cap.pebs_output_pt_available)
4360 return is_intel_pt_event(event);
4363 PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63");
4365 PMU_FORMAT_ATTR(ldlat, "config1:0-15");
4367 PMU_FORMAT_ATTR(frontend, "config1:0-23");
4369 static struct attribute *intel_arch3_formats_attr[] = {
4370 &format_attr_event.attr,
4371 &format_attr_umask.attr,
4372 &format_attr_edge.attr,
4373 &format_attr_pc.attr,
4374 &format_attr_any.attr,
4375 &format_attr_inv.attr,
4376 &format_attr_cmask.attr,
4380 static struct attribute *hsw_format_attr[] = {
4381 &format_attr_in_tx.attr,
4382 &format_attr_in_tx_cp.attr,
4383 &format_attr_offcore_rsp.attr,
4384 &format_attr_ldlat.attr,
4388 static struct attribute *nhm_format_attr[] = {
4389 &format_attr_offcore_rsp.attr,
4390 &format_attr_ldlat.attr,
4394 static struct attribute *slm_format_attr[] = {
4395 &format_attr_offcore_rsp.attr,
4399 static struct attribute *skl_format_attr[] = {
4400 &format_attr_frontend.attr,
4404 static __initconst const struct x86_pmu core_pmu = {
4406 .handle_irq = x86_pmu_handle_irq,
4407 .disable_all = x86_pmu_disable_all,
4408 .enable_all = core_pmu_enable_all,
4409 .enable = core_pmu_enable_event,
4410 .disable = x86_pmu_disable_event,
4411 .hw_config = core_pmu_hw_config,
4412 .schedule_events = x86_schedule_events,
4413 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
4414 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
4415 .event_map = intel_pmu_event_map,
4416 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
4418 .large_pebs_flags = LARGE_PEBS_FLAGS,
4421 * Intel PMCs cannot be accessed sanely above 32-bit width,
4422 * so we install an artificial 1<<31 period regardless of
4423 * the generic event period:
4425 .max_period = (1ULL<<31) - 1,
4426 .get_event_constraints = intel_get_event_constraints,
4427 .put_event_constraints = intel_put_event_constraints,
4428 .event_constraints = intel_core_event_constraints,
4429 .guest_get_msrs = core_guest_get_msrs,
4430 .format_attrs = intel_arch_formats_attr,
4431 .events_sysfs_show = intel_event_sysfs_show,
4434 * Virtual (or funny metal) CPU can define x86_pmu.extra_regs
4435 * together with PMU version 1 and thus be using core_pmu with
4436 * shared_regs. We need following callbacks here to allocate
4439 .cpu_prepare = intel_pmu_cpu_prepare,
4440 .cpu_starting = intel_pmu_cpu_starting,
4441 .cpu_dying = intel_pmu_cpu_dying,
4442 .cpu_dead = intel_pmu_cpu_dead,
4444 .check_period = intel_pmu_check_period,
4446 .lbr_reset = intel_pmu_lbr_reset_64,
4447 .lbr_read = intel_pmu_lbr_read_64,
4448 .lbr_save = intel_pmu_lbr_save,
4449 .lbr_restore = intel_pmu_lbr_restore,
4452 static __initconst const struct x86_pmu intel_pmu = {
4454 .handle_irq = intel_pmu_handle_irq,
4455 .disable_all = intel_pmu_disable_all,
4456 .enable_all = intel_pmu_enable_all,
4457 .enable = intel_pmu_enable_event,
4458 .disable = intel_pmu_disable_event,
4459 .add = intel_pmu_add_event,
4460 .del = intel_pmu_del_event,
4461 .read = intel_pmu_read_event,
4462 .hw_config = intel_pmu_hw_config,
4463 .schedule_events = x86_schedule_events,
4464 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
4465 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
4466 .event_map = intel_pmu_event_map,
4467 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
4469 .large_pebs_flags = LARGE_PEBS_FLAGS,
4471 * Intel PMCs cannot be accessed sanely above 32 bit width,
4472 * so we install an artificial 1<<31 period regardless of
4473 * the generic event period:
4475 .max_period = (1ULL << 31) - 1,
4476 .get_event_constraints = intel_get_event_constraints,
4477 .put_event_constraints = intel_put_event_constraints,
4478 .pebs_aliases = intel_pebs_aliases_core2,
4480 .format_attrs = intel_arch3_formats_attr,
4481 .events_sysfs_show = intel_event_sysfs_show,
4483 .cpu_prepare = intel_pmu_cpu_prepare,
4484 .cpu_starting = intel_pmu_cpu_starting,
4485 .cpu_dying = intel_pmu_cpu_dying,
4486 .cpu_dead = intel_pmu_cpu_dead,
4488 .guest_get_msrs = intel_guest_get_msrs,
4489 .sched_task = intel_pmu_sched_task,
4490 .swap_task_ctx = intel_pmu_swap_task_ctx,
4492 .check_period = intel_pmu_check_period,
4494 .aux_output_match = intel_pmu_aux_output_match,
4496 .lbr_reset = intel_pmu_lbr_reset_64,
4497 .lbr_read = intel_pmu_lbr_read_64,
4498 .lbr_save = intel_pmu_lbr_save,
4499 .lbr_restore = intel_pmu_lbr_restore,
4502 static __init void intel_clovertown_quirk(void)
4505 * PEBS is unreliable due to:
4507 * AJ67 - PEBS may experience CPL leaks
4508 * AJ68 - PEBS PMI may be delayed by one event
4509 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
4510 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
4512 * AJ67 could be worked around by restricting the OS/USR flags.
4513 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
4515 * AJ106 could possibly be worked around by not allowing LBR
4516 * usage from PEBS, including the fixup.
4517 * AJ68 could possibly be worked around by always programming
4518 * a pebs_event_reset[0] value and coping with the lost events.
4520 * But taken together it might just make sense to not enable PEBS on
4523 pr_warn("PEBS disabled due to CPU errata\n");
4525 x86_pmu.pebs_constraints = NULL;
4528 static const struct x86_cpu_desc isolation_ucodes[] = {
4529 INTEL_CPU_DESC(INTEL_FAM6_HASWELL, 3, 0x0000001f),
4530 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_L, 1, 0x0000001e),
4531 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_G, 1, 0x00000015),
4532 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X, 2, 0x00000037),
4533 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X, 4, 0x0000000a),
4534 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL, 4, 0x00000023),
4535 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_G, 1, 0x00000014),
4536 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 2, 0x00000010),
4537 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 3, 0x07000009),
4538 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 4, 0x0f000009),
4539 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 5, 0x0e000002),
4540 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_X, 2, 0x0b000014),
4541 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 3, 0x00000021),
4542 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 4, 0x00000000),
4543 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 5, 0x00000000),
4544 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 6, 0x00000000),
4545 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 7, 0x00000000),
4546 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_L, 3, 0x0000007c),
4547 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE, 3, 0x0000007c),
4548 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 9, 0x0000004e),
4549 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 9, 0x0000004e),
4550 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 10, 0x0000004e),
4551 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 11, 0x0000004e),
4552 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 12, 0x0000004e),
4553 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 10, 0x0000004e),
4554 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 11, 0x0000004e),
4555 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 12, 0x0000004e),
4556 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 13, 0x0000004e),
4560 static void intel_check_pebs_isolation(void)
4562 x86_pmu.pebs_no_isolation = !x86_cpu_has_min_microcode_rev(isolation_ucodes);
4565 static __init void intel_pebs_isolation_quirk(void)
4567 WARN_ON_ONCE(x86_pmu.check_microcode);
4568 x86_pmu.check_microcode = intel_check_pebs_isolation;
4569 intel_check_pebs_isolation();
4572 static const struct x86_cpu_desc pebs_ucodes[] = {
4573 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE, 7, 0x00000028),
4574 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X, 6, 0x00000618),
4575 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X, 7, 0x0000070c),
4579 static bool intel_snb_pebs_broken(void)
4581 return !x86_cpu_has_min_microcode_rev(pebs_ucodes);
4584 static void intel_snb_check_microcode(void)
4586 if (intel_snb_pebs_broken() == x86_pmu.pebs_broken)
4590 * Serialized by the microcode lock..
4592 if (x86_pmu.pebs_broken) {
4593 pr_info("PEBS enabled due to microcode update\n");
4594 x86_pmu.pebs_broken = 0;
4596 pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n");
4597 x86_pmu.pebs_broken = 1;
4601 static bool is_lbr_from(unsigned long msr)
4603 unsigned long lbr_from_nr = x86_pmu.lbr_from + x86_pmu.lbr_nr;
4605 return x86_pmu.lbr_from <= msr && msr < lbr_from_nr;
4609 * Under certain circumstances, access certain MSR may cause #GP.
4610 * The function tests if the input MSR can be safely accessed.
4612 static bool check_msr(unsigned long msr, u64 mask)
4614 u64 val_old, val_new, val_tmp;
4617 * Disable the check for real HW, so we don't
4618 * mess with potentionaly enabled registers:
4620 if (!boot_cpu_has(X86_FEATURE_HYPERVISOR))
4624 * Read the current value, change it and read it back to see if it
4625 * matches, this is needed to detect certain hardware emulators
4626 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
4628 if (rdmsrl_safe(msr, &val_old))
4632 * Only change the bits which can be updated by wrmsrl.
4634 val_tmp = val_old ^ mask;
4636 if (is_lbr_from(msr))
4637 val_tmp = lbr_from_signext_quirk_wr(val_tmp);
4639 if (wrmsrl_safe(msr, val_tmp) ||
4640 rdmsrl_safe(msr, &val_new))
4644 * Quirk only affects validation in wrmsr(), so wrmsrl()'s value
4645 * should equal rdmsrl()'s even with the quirk.
4647 if (val_new != val_tmp)
4650 if (is_lbr_from(msr))
4651 val_old = lbr_from_signext_quirk_wr(val_old);
4653 /* Here it's sure that the MSR can be safely accessed.
4654 * Restore the old value and return.
4656 wrmsrl(msr, val_old);
4661 static __init void intel_sandybridge_quirk(void)
4663 x86_pmu.check_microcode = intel_snb_check_microcode;
4665 intel_snb_check_microcode();
4669 static const struct { int id; char *name; } intel_arch_events_map[] __initconst = {
4670 { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" },
4671 { PERF_COUNT_HW_INSTRUCTIONS, "instructions" },
4672 { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" },
4673 { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" },
4674 { PERF_COUNT_HW_CACHE_MISSES, "cache misses" },
4675 { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" },
4676 { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" },
4679 static __init void intel_arch_events_quirk(void)
4683 /* disable event that reported as not presend by cpuid */
4684 for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) {
4685 intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0;
4686 pr_warn("CPUID marked event: \'%s\' unavailable\n",
4687 intel_arch_events_map[bit].name);
4691 static __init void intel_nehalem_quirk(void)
4693 union cpuid10_ebx ebx;
4695 ebx.full = x86_pmu.events_maskl;
4696 if (ebx.split.no_branch_misses_retired) {
4698 * Erratum AAJ80 detected, we work it around by using
4699 * the BR_MISP_EXEC.ANY event. This will over-count
4700 * branch-misses, but it's still much better than the
4701 * architectural event which is often completely bogus:
4703 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89;
4704 ebx.split.no_branch_misses_retired = 0;
4705 x86_pmu.events_maskl = ebx.full;
4706 pr_info("CPU erratum AAJ80 worked around\n");
4711 * enable software workaround for errata:
4716 * Only needed when HT is enabled. However detecting
4717 * if HT is enabled is difficult (model specific). So instead,
4718 * we enable the workaround in the early boot, and verify if
4719 * it is needed in a later initcall phase once we have valid
4720 * topology information to check if HT is actually enabled
4722 static __init void intel_ht_bug(void)
4724 x86_pmu.flags |= PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED;
4726 x86_pmu.start_scheduling = intel_start_scheduling;
4727 x86_pmu.commit_scheduling = intel_commit_scheduling;
4728 x86_pmu.stop_scheduling = intel_stop_scheduling;
4731 EVENT_ATTR_STR(mem-loads, mem_ld_hsw, "event=0xcd,umask=0x1,ldlat=3");
4732 EVENT_ATTR_STR(mem-stores, mem_st_hsw, "event=0xd0,umask=0x82")
4734 /* Haswell special events */
4735 EVENT_ATTR_STR(tx-start, tx_start, "event=0xc9,umask=0x1");
4736 EVENT_ATTR_STR(tx-commit, tx_commit, "event=0xc9,umask=0x2");
4737 EVENT_ATTR_STR(tx-abort, tx_abort, "event=0xc9,umask=0x4");
4738 EVENT_ATTR_STR(tx-capacity, tx_capacity, "event=0x54,umask=0x2");
4739 EVENT_ATTR_STR(tx-conflict, tx_conflict, "event=0x54,umask=0x1");
4740 EVENT_ATTR_STR(el-start, el_start, "event=0xc8,umask=0x1");
4741 EVENT_ATTR_STR(el-commit, el_commit, "event=0xc8,umask=0x2");
4742 EVENT_ATTR_STR(el-abort, el_abort, "event=0xc8,umask=0x4");
4743 EVENT_ATTR_STR(el-capacity, el_capacity, "event=0x54,umask=0x2");
4744 EVENT_ATTR_STR(el-conflict, el_conflict, "event=0x54,umask=0x1");
4745 EVENT_ATTR_STR(cycles-t, cycles_t, "event=0x3c,in_tx=1");
4746 EVENT_ATTR_STR(cycles-ct, cycles_ct, "event=0x3c,in_tx=1,in_tx_cp=1");
4748 static struct attribute *hsw_events_attrs[] = {
4749 EVENT_PTR(td_slots_issued),
4750 EVENT_PTR(td_slots_retired),
4751 EVENT_PTR(td_fetch_bubbles),
4752 EVENT_PTR(td_total_slots),
4753 EVENT_PTR(td_total_slots_scale),
4754 EVENT_PTR(td_recovery_bubbles),
4755 EVENT_PTR(td_recovery_bubbles_scale),
4759 static struct attribute *hsw_mem_events_attrs[] = {
4760 EVENT_PTR(mem_ld_hsw),
4761 EVENT_PTR(mem_st_hsw),
4765 static struct attribute *hsw_tsx_events_attrs[] = {
4766 EVENT_PTR(tx_start),
4767 EVENT_PTR(tx_commit),
4768 EVENT_PTR(tx_abort),
4769 EVENT_PTR(tx_capacity),
4770 EVENT_PTR(tx_conflict),
4771 EVENT_PTR(el_start),
4772 EVENT_PTR(el_commit),
4773 EVENT_PTR(el_abort),
4774 EVENT_PTR(el_capacity),
4775 EVENT_PTR(el_conflict),
4776 EVENT_PTR(cycles_t),
4777 EVENT_PTR(cycles_ct),
4781 EVENT_ATTR_STR(tx-capacity-read, tx_capacity_read, "event=0x54,umask=0x80");
4782 EVENT_ATTR_STR(tx-capacity-write, tx_capacity_write, "event=0x54,umask=0x2");
4783 EVENT_ATTR_STR(el-capacity-read, el_capacity_read, "event=0x54,umask=0x80");
4784 EVENT_ATTR_STR(el-capacity-write, el_capacity_write, "event=0x54,umask=0x2");
4786 static struct attribute *icl_events_attrs[] = {
4787 EVENT_PTR(mem_ld_hsw),
4788 EVENT_PTR(mem_st_hsw),
4792 static struct attribute *icl_td_events_attrs[] = {
4794 EVENT_PTR(td_retiring),
4795 EVENT_PTR(td_bad_spec),
4796 EVENT_PTR(td_fe_bound),
4797 EVENT_PTR(td_be_bound),
4801 static struct attribute *icl_tsx_events_attrs[] = {
4802 EVENT_PTR(tx_start),
4803 EVENT_PTR(tx_abort),
4804 EVENT_PTR(tx_commit),
4805 EVENT_PTR(tx_capacity_read),
4806 EVENT_PTR(tx_capacity_write),
4807 EVENT_PTR(tx_conflict),
4808 EVENT_PTR(el_start),
4809 EVENT_PTR(el_abort),
4810 EVENT_PTR(el_commit),
4811 EVENT_PTR(el_capacity_read),
4812 EVENT_PTR(el_capacity_write),
4813 EVENT_PTR(el_conflict),
4814 EVENT_PTR(cycles_t),
4815 EVENT_PTR(cycles_ct),
4820 EVENT_ATTR_STR(mem-stores, mem_st_spr, "event=0xcd,umask=0x2");
4821 EVENT_ATTR_STR(mem-loads-aux, mem_ld_aux, "event=0x03,umask=0x82");
4823 static struct attribute *spr_events_attrs[] = {
4824 EVENT_PTR(mem_ld_hsw),
4825 EVENT_PTR(mem_st_spr),
4826 EVENT_PTR(mem_ld_aux),
4830 static struct attribute *spr_td_events_attrs[] = {
4832 EVENT_PTR(td_retiring),
4833 EVENT_PTR(td_bad_spec),
4834 EVENT_PTR(td_fe_bound),
4835 EVENT_PTR(td_be_bound),
4836 EVENT_PTR(td_heavy_ops),
4837 EVENT_PTR(td_br_mispredict),
4838 EVENT_PTR(td_fetch_lat),
4839 EVENT_PTR(td_mem_bound),
4843 static struct attribute *spr_tsx_events_attrs[] = {
4844 EVENT_PTR(tx_start),
4845 EVENT_PTR(tx_abort),
4846 EVENT_PTR(tx_commit),
4847 EVENT_PTR(tx_capacity_read),
4848 EVENT_PTR(tx_capacity_write),
4849 EVENT_PTR(tx_conflict),
4850 EVENT_PTR(cycles_t),
4851 EVENT_PTR(cycles_ct),
4855 static ssize_t freeze_on_smi_show(struct device *cdev,
4856 struct device_attribute *attr,
4859 return sprintf(buf, "%lu\n", x86_pmu.attr_freeze_on_smi);
4862 static DEFINE_MUTEX(freeze_on_smi_mutex);
4864 static ssize_t freeze_on_smi_store(struct device *cdev,
4865 struct device_attribute *attr,
4866 const char *buf, size_t count)
4871 ret = kstrtoul(buf, 0, &val);
4878 mutex_lock(&freeze_on_smi_mutex);
4880 if (x86_pmu.attr_freeze_on_smi == val)
4883 x86_pmu.attr_freeze_on_smi = val;
4886 on_each_cpu(flip_smm_bit, &val, 1);
4889 mutex_unlock(&freeze_on_smi_mutex);
4894 static void update_tfa_sched(void *ignored)
4896 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
4899 * check if PMC3 is used
4900 * and if so force schedule out for all event types all contexts
4902 if (test_bit(3, cpuc->active_mask))
4903 perf_pmu_resched(x86_get_pmu(smp_processor_id()));
4906 static ssize_t show_sysctl_tfa(struct device *cdev,
4907 struct device_attribute *attr,
4910 return snprintf(buf, 40, "%d\n", allow_tsx_force_abort);
4913 static ssize_t set_sysctl_tfa(struct device *cdev,
4914 struct device_attribute *attr,
4915 const char *buf, size_t count)
4920 ret = kstrtobool(buf, &val);
4925 if (val == allow_tsx_force_abort)
4928 allow_tsx_force_abort = val;
4931 on_each_cpu(update_tfa_sched, NULL, 1);
4938 static DEVICE_ATTR_RW(freeze_on_smi);
4940 static ssize_t branches_show(struct device *cdev,
4941 struct device_attribute *attr,
4944 return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu.lbr_nr);
4947 static DEVICE_ATTR_RO(branches);
4949 static struct attribute *lbr_attrs[] = {
4950 &dev_attr_branches.attr,
4954 static char pmu_name_str[30];
4956 static ssize_t pmu_name_show(struct device *cdev,
4957 struct device_attribute *attr,
4960 return snprintf(buf, PAGE_SIZE, "%s\n", pmu_name_str);
4963 static DEVICE_ATTR_RO(pmu_name);
4965 static struct attribute *intel_pmu_caps_attrs[] = {
4966 &dev_attr_pmu_name.attr,
4970 static DEVICE_ATTR(allow_tsx_force_abort, 0644,
4974 static struct attribute *intel_pmu_attrs[] = {
4975 &dev_attr_freeze_on_smi.attr,
4976 &dev_attr_allow_tsx_force_abort.attr,
4981 tsx_is_visible(struct kobject *kobj, struct attribute *attr, int i)
4983 return boot_cpu_has(X86_FEATURE_RTM) ? attr->mode : 0;
4987 pebs_is_visible(struct kobject *kobj, struct attribute *attr, int i)
4989 return x86_pmu.pebs ? attr->mode : 0;
4993 lbr_is_visible(struct kobject *kobj, struct attribute *attr, int i)
4995 return x86_pmu.lbr_nr ? attr->mode : 0;
4999 exra_is_visible(struct kobject *kobj, struct attribute *attr, int i)
5001 return x86_pmu.version >= 2 ? attr->mode : 0;
5005 default_is_visible(struct kobject *kobj, struct attribute *attr, int i)
5007 if (attr == &dev_attr_allow_tsx_force_abort.attr)
5008 return x86_pmu.flags & PMU_FL_TFA ? attr->mode : 0;
5013 static struct attribute_group group_events_td = {
5017 static struct attribute_group group_events_mem = {
5019 .is_visible = pebs_is_visible,
5022 static struct attribute_group group_events_tsx = {
5024 .is_visible = tsx_is_visible,
5027 static struct attribute_group group_caps_gen = {
5029 .attrs = intel_pmu_caps_attrs,
5032 static struct attribute_group group_caps_lbr = {
5035 .is_visible = lbr_is_visible,
5038 static struct attribute_group group_format_extra = {
5040 .is_visible = exra_is_visible,
5043 static struct attribute_group group_format_extra_skl = {
5045 .is_visible = exra_is_visible,
5048 static struct attribute_group group_default = {
5049 .attrs = intel_pmu_attrs,
5050 .is_visible = default_is_visible,
5053 static const struct attribute_group *attr_update[] = {
5059 &group_format_extra,
5060 &group_format_extra_skl,
5065 static struct attribute *empty_attrs;
5067 __init int intel_pmu_init(void)
5069 struct attribute **extra_skl_attr = &empty_attrs;
5070 struct attribute **extra_attr = &empty_attrs;
5071 struct attribute **td_attr = &empty_attrs;
5072 struct attribute **mem_attr = &empty_attrs;
5073 struct attribute **tsx_attr = &empty_attrs;
5074 union cpuid10_edx edx;
5075 union cpuid10_eax eax;
5076 union cpuid10_ebx ebx;
5077 struct event_constraint *c;
5078 unsigned int fixed_mask;
5079 struct extra_reg *er;
5084 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
5085 switch (boot_cpu_data.x86) {
5087 return p6_pmu_init();
5089 return knc_pmu_init();
5091 return p4_pmu_init();
5097 * Check whether the Architectural PerfMon supports
5098 * Branch Misses Retired hw_event or not.
5100 cpuid(10, &eax.full, &ebx.full, &fixed_mask, &edx.full);
5101 if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT)
5104 version = eax.split.version_id;
5108 x86_pmu = intel_pmu;
5110 x86_pmu.version = version;
5111 x86_pmu.num_counters = eax.split.num_counters;
5112 x86_pmu.cntval_bits = eax.split.bit_width;
5113 x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1;
5115 x86_pmu.events_maskl = ebx.full;
5116 x86_pmu.events_mask_len = eax.split.mask_length;
5118 x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters);
5121 * Quirk: v2 perfmon does not report fixed-purpose events, so
5122 * assume at least 3 events, when not running in a hypervisor:
5124 if (version > 1 && version < 5) {
5125 int assume = 3 * !boot_cpu_has(X86_FEATURE_HYPERVISOR);
5127 x86_pmu.num_counters_fixed =
5128 max((int)edx.split.num_counters_fixed, assume);
5130 fixed_mask = (1L << x86_pmu.num_counters_fixed) - 1;
5131 } else if (version >= 5)
5132 x86_pmu.num_counters_fixed = fls(fixed_mask);
5134 if (boot_cpu_has(X86_FEATURE_PDCM)) {
5137 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
5138 x86_pmu.intel_cap.capabilities = capabilities;
5141 if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32) {
5142 x86_pmu.lbr_reset = intel_pmu_lbr_reset_32;
5143 x86_pmu.lbr_read = intel_pmu_lbr_read_32;
5146 if (boot_cpu_has(X86_FEATURE_ARCH_LBR))
5147 intel_pmu_arch_lbr_init();
5151 x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */
5154 x86_pmu.intel_cap.anythread_deprecated = edx.split.anythread_deprecated;
5155 if (x86_pmu.intel_cap.anythread_deprecated)
5156 pr_cont(" AnyThread deprecated, ");
5160 * Install the hw-cache-events table:
5162 switch (boot_cpu_data.x86_model) {
5163 case INTEL_FAM6_CORE_YONAH:
5164 pr_cont("Core events, ");
5168 case INTEL_FAM6_CORE2_MEROM:
5169 x86_add_quirk(intel_clovertown_quirk);
5172 case INTEL_FAM6_CORE2_MEROM_L:
5173 case INTEL_FAM6_CORE2_PENRYN:
5174 case INTEL_FAM6_CORE2_DUNNINGTON:
5175 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
5176 sizeof(hw_cache_event_ids));
5178 intel_pmu_lbr_init_core();
5180 x86_pmu.event_constraints = intel_core2_event_constraints;
5181 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints;
5182 pr_cont("Core2 events, ");
5186 case INTEL_FAM6_NEHALEM:
5187 case INTEL_FAM6_NEHALEM_EP:
5188 case INTEL_FAM6_NEHALEM_EX:
5189 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
5190 sizeof(hw_cache_event_ids));
5191 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
5192 sizeof(hw_cache_extra_regs));
5194 intel_pmu_lbr_init_nhm();
5196 x86_pmu.event_constraints = intel_nehalem_event_constraints;
5197 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints;
5198 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
5199 x86_pmu.extra_regs = intel_nehalem_extra_regs;
5200 x86_pmu.limit_period = nhm_limit_period;
5202 mem_attr = nhm_mem_events_attrs;
5204 /* UOPS_ISSUED.STALLED_CYCLES */
5205 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
5206 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
5207 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
5208 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
5209 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
5211 intel_pmu_pebs_data_source_nhm();
5212 x86_add_quirk(intel_nehalem_quirk);
5213 x86_pmu.pebs_no_tlb = 1;
5214 extra_attr = nhm_format_attr;
5216 pr_cont("Nehalem events, ");
5220 case INTEL_FAM6_ATOM_BONNELL:
5221 case INTEL_FAM6_ATOM_BONNELL_MID:
5222 case INTEL_FAM6_ATOM_SALTWELL:
5223 case INTEL_FAM6_ATOM_SALTWELL_MID:
5224 case INTEL_FAM6_ATOM_SALTWELL_TABLET:
5225 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
5226 sizeof(hw_cache_event_ids));
5228 intel_pmu_lbr_init_atom();
5230 x86_pmu.event_constraints = intel_gen_event_constraints;
5231 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints;
5232 x86_pmu.pebs_aliases = intel_pebs_aliases_core2;
5233 pr_cont("Atom events, ");
5237 case INTEL_FAM6_ATOM_SILVERMONT:
5238 case INTEL_FAM6_ATOM_SILVERMONT_D:
5239 case INTEL_FAM6_ATOM_SILVERMONT_MID:
5240 case INTEL_FAM6_ATOM_AIRMONT:
5241 case INTEL_FAM6_ATOM_AIRMONT_MID:
5242 memcpy(hw_cache_event_ids, slm_hw_cache_event_ids,
5243 sizeof(hw_cache_event_ids));
5244 memcpy(hw_cache_extra_regs, slm_hw_cache_extra_regs,
5245 sizeof(hw_cache_extra_regs));
5247 intel_pmu_lbr_init_slm();
5249 x86_pmu.event_constraints = intel_slm_event_constraints;
5250 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints;
5251 x86_pmu.extra_regs = intel_slm_extra_regs;
5252 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5253 td_attr = slm_events_attrs;
5254 extra_attr = slm_format_attr;
5255 pr_cont("Silvermont events, ");
5256 name = "silvermont";
5259 case INTEL_FAM6_ATOM_GOLDMONT:
5260 case INTEL_FAM6_ATOM_GOLDMONT_D:
5261 memcpy(hw_cache_event_ids, glm_hw_cache_event_ids,
5262 sizeof(hw_cache_event_ids));
5263 memcpy(hw_cache_extra_regs, glm_hw_cache_extra_regs,
5264 sizeof(hw_cache_extra_regs));
5266 intel_pmu_lbr_init_skl();
5268 x86_pmu.event_constraints = intel_slm_event_constraints;
5269 x86_pmu.pebs_constraints = intel_glm_pebs_event_constraints;
5270 x86_pmu.extra_regs = intel_glm_extra_regs;
5272 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS
5273 * for precise cycles.
5274 * :pp is identical to :ppp
5276 x86_pmu.pebs_aliases = NULL;
5277 x86_pmu.pebs_prec_dist = true;
5278 x86_pmu.lbr_pt_coexist = true;
5279 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5280 td_attr = glm_events_attrs;
5281 extra_attr = slm_format_attr;
5282 pr_cont("Goldmont events, ");
5286 case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
5287 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids,
5288 sizeof(hw_cache_event_ids));
5289 memcpy(hw_cache_extra_regs, glp_hw_cache_extra_regs,
5290 sizeof(hw_cache_extra_regs));
5292 intel_pmu_lbr_init_skl();
5294 x86_pmu.event_constraints = intel_slm_event_constraints;
5295 x86_pmu.extra_regs = intel_glm_extra_regs;
5297 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS
5298 * for precise cycles.
5300 x86_pmu.pebs_aliases = NULL;
5301 x86_pmu.pebs_prec_dist = true;
5302 x86_pmu.lbr_pt_coexist = true;
5303 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5304 x86_pmu.flags |= PMU_FL_PEBS_ALL;
5305 x86_pmu.get_event_constraints = glp_get_event_constraints;
5306 td_attr = glm_events_attrs;
5307 /* Goldmont Plus has 4-wide pipeline */
5308 event_attr_td_total_slots_scale_glm.event_str = "4";
5309 extra_attr = slm_format_attr;
5310 pr_cont("Goldmont plus events, ");
5311 name = "goldmont_plus";
5314 case INTEL_FAM6_ATOM_TREMONT_D:
5315 case INTEL_FAM6_ATOM_TREMONT:
5316 case INTEL_FAM6_ATOM_TREMONT_L:
5317 x86_pmu.late_ack = true;
5318 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids,
5319 sizeof(hw_cache_event_ids));
5320 memcpy(hw_cache_extra_regs, tnt_hw_cache_extra_regs,
5321 sizeof(hw_cache_extra_regs));
5322 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1;
5324 intel_pmu_lbr_init_skl();
5326 x86_pmu.event_constraints = intel_slm_event_constraints;
5327 x86_pmu.extra_regs = intel_tnt_extra_regs;
5329 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS
5330 * for precise cycles.
5332 x86_pmu.pebs_aliases = NULL;
5333 x86_pmu.pebs_prec_dist = true;
5334 x86_pmu.lbr_pt_coexist = true;
5335 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5336 x86_pmu.get_event_constraints = tnt_get_event_constraints;
5337 td_attr = tnt_events_attrs;
5338 extra_attr = slm_format_attr;
5339 pr_cont("Tremont events, ");
5343 case INTEL_FAM6_WESTMERE:
5344 case INTEL_FAM6_WESTMERE_EP:
5345 case INTEL_FAM6_WESTMERE_EX:
5346 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
5347 sizeof(hw_cache_event_ids));
5348 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
5349 sizeof(hw_cache_extra_regs));
5351 intel_pmu_lbr_init_nhm();
5353 x86_pmu.event_constraints = intel_westmere_event_constraints;
5354 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
5355 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints;
5356 x86_pmu.extra_regs = intel_westmere_extra_regs;
5357 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5359 mem_attr = nhm_mem_events_attrs;
5361 /* UOPS_ISSUED.STALLED_CYCLES */
5362 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
5363 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
5364 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
5365 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
5366 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
5368 intel_pmu_pebs_data_source_nhm();
5369 extra_attr = nhm_format_attr;
5370 pr_cont("Westmere events, ");
5374 case INTEL_FAM6_SANDYBRIDGE:
5375 case INTEL_FAM6_SANDYBRIDGE_X:
5376 x86_add_quirk(intel_sandybridge_quirk);
5377 x86_add_quirk(intel_ht_bug);
5378 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
5379 sizeof(hw_cache_event_ids));
5380 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
5381 sizeof(hw_cache_extra_regs));
5383 intel_pmu_lbr_init_snb();
5385 x86_pmu.event_constraints = intel_snb_event_constraints;
5386 x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints;
5387 x86_pmu.pebs_aliases = intel_pebs_aliases_snb;
5388 if (boot_cpu_data.x86_model == INTEL_FAM6_SANDYBRIDGE_X)
5389 x86_pmu.extra_regs = intel_snbep_extra_regs;
5391 x86_pmu.extra_regs = intel_snb_extra_regs;
5394 /* all extra regs are per-cpu when HT is on */
5395 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5396 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5398 td_attr = snb_events_attrs;
5399 mem_attr = snb_mem_events_attrs;
5401 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
5402 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
5403 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
5404 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
5405 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
5406 X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1);
5408 extra_attr = nhm_format_attr;
5410 pr_cont("SandyBridge events, ");
5411 name = "sandybridge";
5414 case INTEL_FAM6_IVYBRIDGE:
5415 case INTEL_FAM6_IVYBRIDGE_X:
5416 x86_add_quirk(intel_ht_bug);
5417 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
5418 sizeof(hw_cache_event_ids));
5419 /* dTLB-load-misses on IVB is different than SNB */
5420 hw_cache_event_ids[C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = 0x8108; /* DTLB_LOAD_MISSES.DEMAND_LD_MISS_CAUSES_A_WALK */
5422 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
5423 sizeof(hw_cache_extra_regs));
5425 intel_pmu_lbr_init_snb();
5427 x86_pmu.event_constraints = intel_ivb_event_constraints;
5428 x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints;
5429 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
5430 x86_pmu.pebs_prec_dist = true;
5431 if (boot_cpu_data.x86_model == INTEL_FAM6_IVYBRIDGE_X)
5432 x86_pmu.extra_regs = intel_snbep_extra_regs;
5434 x86_pmu.extra_regs = intel_snb_extra_regs;
5435 /* all extra regs are per-cpu when HT is on */
5436 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5437 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5439 td_attr = snb_events_attrs;
5440 mem_attr = snb_mem_events_attrs;
5442 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
5443 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
5444 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
5446 extra_attr = nhm_format_attr;
5448 pr_cont("IvyBridge events, ");
5453 case INTEL_FAM6_HASWELL:
5454 case INTEL_FAM6_HASWELL_X:
5455 case INTEL_FAM6_HASWELL_L:
5456 case INTEL_FAM6_HASWELL_G:
5457 x86_add_quirk(intel_ht_bug);
5458 x86_add_quirk(intel_pebs_isolation_quirk);
5459 x86_pmu.late_ack = true;
5460 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids));
5461 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
5463 intel_pmu_lbr_init_hsw();
5465 x86_pmu.event_constraints = intel_hsw_event_constraints;
5466 x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints;
5467 x86_pmu.extra_regs = intel_snbep_extra_regs;
5468 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
5469 x86_pmu.pebs_prec_dist = true;
5470 /* all extra regs are per-cpu when HT is on */
5471 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5472 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5474 x86_pmu.hw_config = hsw_hw_config;
5475 x86_pmu.get_event_constraints = hsw_get_event_constraints;
5476 x86_pmu.lbr_double_abort = true;
5477 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
5478 hsw_format_attr : nhm_format_attr;
5479 td_attr = hsw_events_attrs;
5480 mem_attr = hsw_mem_events_attrs;
5481 tsx_attr = hsw_tsx_events_attrs;
5482 pr_cont("Haswell events, ");
5486 case INTEL_FAM6_BROADWELL:
5487 case INTEL_FAM6_BROADWELL_D:
5488 case INTEL_FAM6_BROADWELL_G:
5489 case INTEL_FAM6_BROADWELL_X:
5490 x86_add_quirk(intel_pebs_isolation_quirk);
5491 x86_pmu.late_ack = true;
5492 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids));
5493 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
5495 /* L3_MISS_LOCAL_DRAM is BIT(26) in Broadwell */
5496 hw_cache_extra_regs[C(LL)][C(OP_READ)][C(RESULT_MISS)] = HSW_DEMAND_READ |
5497 BDW_L3_MISS|HSW_SNOOP_DRAM;
5498 hw_cache_extra_regs[C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = HSW_DEMAND_WRITE|BDW_L3_MISS|
5500 hw_cache_extra_regs[C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = HSW_DEMAND_READ|
5501 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM;
5502 hw_cache_extra_regs[C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = HSW_DEMAND_WRITE|
5503 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM;
5505 intel_pmu_lbr_init_hsw();
5507 x86_pmu.event_constraints = intel_bdw_event_constraints;
5508 x86_pmu.pebs_constraints = intel_bdw_pebs_event_constraints;
5509 x86_pmu.extra_regs = intel_snbep_extra_regs;
5510 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
5511 x86_pmu.pebs_prec_dist = true;
5512 /* all extra regs are per-cpu when HT is on */
5513 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5514 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5516 x86_pmu.hw_config = hsw_hw_config;
5517 x86_pmu.get_event_constraints = hsw_get_event_constraints;
5518 x86_pmu.limit_period = bdw_limit_period;
5519 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
5520 hsw_format_attr : nhm_format_attr;
5521 td_attr = hsw_events_attrs;
5522 mem_attr = hsw_mem_events_attrs;
5523 tsx_attr = hsw_tsx_events_attrs;
5524 pr_cont("Broadwell events, ");
5528 case INTEL_FAM6_XEON_PHI_KNL:
5529 case INTEL_FAM6_XEON_PHI_KNM:
5530 memcpy(hw_cache_event_ids,
5531 slm_hw_cache_event_ids, sizeof(hw_cache_event_ids));
5532 memcpy(hw_cache_extra_regs,
5533 knl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
5534 intel_pmu_lbr_init_knl();
5536 x86_pmu.event_constraints = intel_slm_event_constraints;
5537 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints;
5538 x86_pmu.extra_regs = intel_knl_extra_regs;
5540 /* all extra regs are per-cpu when HT is on */
5541 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5542 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5543 extra_attr = slm_format_attr;
5544 pr_cont("Knights Landing/Mill events, ");
5545 name = "knights-landing";
5548 case INTEL_FAM6_SKYLAKE_X:
5551 case INTEL_FAM6_SKYLAKE_L:
5552 case INTEL_FAM6_SKYLAKE:
5553 case INTEL_FAM6_KABYLAKE_L:
5554 case INTEL_FAM6_KABYLAKE:
5555 case INTEL_FAM6_COMETLAKE_L:
5556 case INTEL_FAM6_COMETLAKE:
5557 x86_add_quirk(intel_pebs_isolation_quirk);
5558 x86_pmu.late_ack = true;
5559 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids));
5560 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
5561 intel_pmu_lbr_init_skl();
5563 /* INT_MISC.RECOVERY_CYCLES has umask 1 in Skylake */
5564 event_attr_td_recovery_bubbles.event_str_noht =
5565 "event=0xd,umask=0x1,cmask=1";
5566 event_attr_td_recovery_bubbles.event_str_ht =
5567 "event=0xd,umask=0x1,cmask=1,any=1";
5569 x86_pmu.event_constraints = intel_skl_event_constraints;
5570 x86_pmu.pebs_constraints = intel_skl_pebs_event_constraints;
5571 x86_pmu.extra_regs = intel_skl_extra_regs;
5572 x86_pmu.pebs_aliases = intel_pebs_aliases_skl;
5573 x86_pmu.pebs_prec_dist = true;
5574 /* all extra regs are per-cpu when HT is on */
5575 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5576 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5578 x86_pmu.hw_config = hsw_hw_config;
5579 x86_pmu.get_event_constraints = hsw_get_event_constraints;
5580 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
5581 hsw_format_attr : nhm_format_attr;
5582 extra_skl_attr = skl_format_attr;
5583 td_attr = hsw_events_attrs;
5584 mem_attr = hsw_mem_events_attrs;
5585 tsx_attr = hsw_tsx_events_attrs;
5586 intel_pmu_pebs_data_source_skl(pmem);
5588 if (boot_cpu_has(X86_FEATURE_TSX_FORCE_ABORT)) {
5589 x86_pmu.flags |= PMU_FL_TFA;
5590 x86_pmu.get_event_constraints = tfa_get_event_constraints;
5591 x86_pmu.enable_all = intel_tfa_pmu_enable_all;
5592 x86_pmu.commit_scheduling = intel_tfa_commit_scheduling;
5595 pr_cont("Skylake events, ");
5599 case INTEL_FAM6_ICELAKE_X:
5600 case INTEL_FAM6_ICELAKE_D:
5603 case INTEL_FAM6_ICELAKE_L:
5604 case INTEL_FAM6_ICELAKE:
5605 case INTEL_FAM6_TIGERLAKE_L:
5606 case INTEL_FAM6_TIGERLAKE:
5607 case INTEL_FAM6_ROCKETLAKE:
5608 x86_pmu.late_ack = true;
5609 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids));
5610 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
5611 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1;
5612 intel_pmu_lbr_init_skl();
5614 x86_pmu.event_constraints = intel_icl_event_constraints;
5615 x86_pmu.pebs_constraints = intel_icl_pebs_event_constraints;
5616 x86_pmu.extra_regs = intel_icl_extra_regs;
5617 x86_pmu.pebs_aliases = NULL;
5618 x86_pmu.pebs_prec_dist = true;
5619 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5620 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5622 x86_pmu.hw_config = hsw_hw_config;
5623 x86_pmu.get_event_constraints = icl_get_event_constraints;
5624 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
5625 hsw_format_attr : nhm_format_attr;
5626 extra_skl_attr = skl_format_attr;
5627 mem_attr = icl_events_attrs;
5628 td_attr = icl_td_events_attrs;
5629 tsx_attr = icl_tsx_events_attrs;
5630 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04);
5631 x86_pmu.lbr_pt_coexist = true;
5632 intel_pmu_pebs_data_source_skl(pmem);
5633 x86_pmu.num_topdown_events = 4;
5634 x86_pmu.update_topdown_event = icl_update_topdown_event;
5635 x86_pmu.set_topdown_event_period = icl_set_topdown_event_period;
5636 pr_cont("Icelake events, ");
5640 case INTEL_FAM6_SAPPHIRERAPIDS_X:
5642 x86_pmu.late_ack = true;
5643 memcpy(hw_cache_event_ids, spr_hw_cache_event_ids, sizeof(hw_cache_event_ids));
5644 memcpy(hw_cache_extra_regs, spr_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
5646 x86_pmu.event_constraints = intel_spr_event_constraints;
5647 x86_pmu.pebs_constraints = intel_spr_pebs_event_constraints;
5648 x86_pmu.extra_regs = intel_spr_extra_regs;
5649 x86_pmu.limit_period = spr_limit_period;
5650 x86_pmu.pebs_aliases = NULL;
5651 x86_pmu.pebs_prec_dist = true;
5652 x86_pmu.pebs_block = true;
5653 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5654 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5655 x86_pmu.flags |= PMU_FL_PEBS_ALL;
5656 x86_pmu.flags |= PMU_FL_INSTR_LATENCY;
5657 x86_pmu.flags |= PMU_FL_MEM_LOADS_AUX;
5659 x86_pmu.hw_config = hsw_hw_config;
5660 x86_pmu.get_event_constraints = spr_get_event_constraints;
5661 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
5662 hsw_format_attr : nhm_format_attr;
5663 extra_skl_attr = skl_format_attr;
5664 mem_attr = spr_events_attrs;
5665 td_attr = spr_td_events_attrs;
5666 tsx_attr = spr_tsx_events_attrs;
5667 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04);
5668 x86_pmu.lbr_pt_coexist = true;
5669 intel_pmu_pebs_data_source_skl(pmem);
5670 x86_pmu.num_topdown_events = 8;
5671 x86_pmu.update_topdown_event = icl_update_topdown_event;
5672 x86_pmu.set_topdown_event_period = icl_set_topdown_event_period;
5673 pr_cont("Sapphire Rapids events, ");
5674 name = "sapphire_rapids";
5678 switch (x86_pmu.version) {
5680 x86_pmu.event_constraints = intel_v1_event_constraints;
5681 pr_cont("generic architected perfmon v1, ");
5682 name = "generic_arch_v1";
5686 * default constraints for v2 and up
5688 x86_pmu.event_constraints = intel_gen_event_constraints;
5689 pr_cont("generic architected perfmon, ");
5690 name = "generic_arch_v2+";
5695 snprintf(pmu_name_str, sizeof(pmu_name_str), "%s", name);
5698 group_events_td.attrs = td_attr;
5699 group_events_mem.attrs = mem_attr;
5700 group_events_tsx.attrs = tsx_attr;
5701 group_format_extra.attrs = extra_attr;
5702 group_format_extra_skl.attrs = extra_skl_attr;
5704 x86_pmu.attr_update = attr_update;
5706 if (x86_pmu.num_counters > INTEL_PMC_MAX_GENERIC) {
5707 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
5708 x86_pmu.num_counters, INTEL_PMC_MAX_GENERIC);
5709 x86_pmu.num_counters = INTEL_PMC_MAX_GENERIC;
5711 x86_pmu.intel_ctrl = (1ULL << x86_pmu.num_counters) - 1;
5713 if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) {
5714 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
5715 x86_pmu.num_counters_fixed, INTEL_PMC_MAX_FIXED);
5716 x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED;
5719 x86_pmu.intel_ctrl |= (u64)fixed_mask << INTEL_PMC_IDX_FIXED;
5721 /* AnyThread may be deprecated on arch perfmon v5 or later */
5722 if (x86_pmu.intel_cap.anythread_deprecated)
5723 x86_pmu.format_attrs = intel_arch_formats_attr;
5725 if (x86_pmu.event_constraints) {
5727 * event on fixed counter2 (REF_CYCLES) only works on this
5728 * counter, so do not extend mask to generic counters
5730 for_each_event_constraint(c, x86_pmu.event_constraints) {
5732 * Don't extend the topdown slots and metrics
5733 * events to the generic counters.
5735 if (c->idxmsk64 & INTEL_PMC_MSK_TOPDOWN) {
5737 * Disable topdown slots and metrics events,
5738 * if slots event is not in CPUID.
5740 if (!(INTEL_PMC_MSK_FIXED_SLOTS & x86_pmu.intel_ctrl))
5742 c->weight = hweight64(c->idxmsk64);
5746 if (c->cmask == FIXED_EVENT_FLAGS) {
5747 /* Disabled fixed counters which are not in CPUID */
5748 c->idxmsk64 &= x86_pmu.intel_ctrl;
5750 if (c->idxmsk64 != INTEL_PMC_MSK_FIXED_REF_CYCLES)
5751 c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1;
5754 ~(~0ULL << (INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed));
5755 c->weight = hweight64(c->idxmsk64);
5760 * Access LBR MSR may cause #GP under certain circumstances.
5761 * E.g. KVM doesn't support LBR MSR
5762 * Check all LBT MSR here.
5763 * Disable LBR access if any LBR MSRs can not be accessed.
5765 if (x86_pmu.lbr_nr && !check_msr(x86_pmu.lbr_tos, 0x3UL))
5767 for (i = 0; i < x86_pmu.lbr_nr; i++) {
5768 if (!(check_msr(x86_pmu.lbr_from + i, 0xffffUL) &&
5769 check_msr(x86_pmu.lbr_to + i, 0xffffUL)))
5774 pr_cont("%d-deep LBR, ", x86_pmu.lbr_nr);
5777 * Access extra MSR may cause #GP under certain circumstances.
5778 * E.g. KVM doesn't support offcore event
5779 * Check all extra_regs here.
5781 if (x86_pmu.extra_regs) {
5782 for (er = x86_pmu.extra_regs; er->msr; er++) {
5783 er->extra_msr_access = check_msr(er->msr, 0x11UL);
5784 /* Disable LBR select mapping */
5785 if ((er->idx == EXTRA_REG_LBR) && !er->extra_msr_access)
5786 x86_pmu.lbr_sel_map = NULL;
5790 /* Support full width counters using alternative MSR range */
5791 if (x86_pmu.intel_cap.full_width_write) {
5792 x86_pmu.max_period = x86_pmu.cntval_mask >> 1;
5793 x86_pmu.perfctr = MSR_IA32_PMC0;
5794 pr_cont("full-width counters, ");
5797 if (!is_hybrid() && x86_pmu.intel_cap.perf_metrics)
5798 x86_pmu.intel_ctrl |= 1ULL << GLOBAL_CTRL_EN_PERF_METRICS;
5804 * HT bug: phase 2 init
5805 * Called once we have valid topology information to check
5806 * whether or not HT is enabled
5807 * If HT is off, then we disable the workaround
5809 static __init int fixup_ht_bug(void)
5813 * problem not present on this CPU model, nothing to do
5815 if (!(x86_pmu.flags & PMU_FL_EXCL_ENABLED))
5818 if (topology_max_smt_threads() > 1) {
5819 pr_info("PMU erratum BJ122, BV98, HSD29 worked around, HT is on\n");
5825 hardlockup_detector_perf_stop();
5827 x86_pmu.flags &= ~(PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED);
5829 x86_pmu.start_scheduling = NULL;
5830 x86_pmu.commit_scheduling = NULL;
5831 x86_pmu.stop_scheduling = NULL;
5833 hardlockup_detector_perf_restart();
5835 for_each_online_cpu(c)
5836 free_excl_cntrs(&per_cpu(cpu_hw_events, c));
5839 pr_info("PMU erratum BJ122, BV98, HSD29 workaround disabled, HT off\n");
5842 subsys_initcall(fixup_ht_bug)