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 INTEL_EVENT_CONSTRAINT_RANGE(0x03, 0x0a, 0xf),
251 INTEL_EVENT_CONSTRAINT_RANGE(0x1f, 0x28, 0xf),
252 INTEL_EVENT_CONSTRAINT(0x32, 0xf), /* SW_PREFETCH_ACCESS.* */
253 INTEL_EVENT_CONSTRAINT_RANGE(0x48, 0x56, 0xf),
254 INTEL_EVENT_CONSTRAINT_RANGE(0x60, 0x8b, 0xf),
255 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xff), /* CYCLE_ACTIVITY.STALLS_TOTAL */
256 INTEL_UEVENT_CONSTRAINT(0x10a3, 0xff), /* CYCLE_ACTIVITY.CYCLES_MEM_ANY */
257 INTEL_UEVENT_CONSTRAINT(0x14a3, 0xff), /* CYCLE_ACTIVITY.STALLS_MEM_ANY */
258 INTEL_EVENT_CONSTRAINT(0xa3, 0xf), /* CYCLE_ACTIVITY.* */
259 INTEL_EVENT_CONSTRAINT_RANGE(0xa8, 0xb0, 0xf),
260 INTEL_EVENT_CONSTRAINT_RANGE(0xb7, 0xbd, 0xf),
261 INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xe6, 0xf),
262 INTEL_EVENT_CONSTRAINT(0xef, 0xf),
263 INTEL_EVENT_CONSTRAINT_RANGE(0xf0, 0xf4, 0xf),
267 static struct extra_reg intel_icl_extra_regs[] __read_mostly = {
268 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffffbfffull, RSP_0),
269 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffffbfffull, RSP_1),
270 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
271 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE),
275 EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3");
276 EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3");
277 EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2");
279 static struct attribute *nhm_mem_events_attrs[] = {
280 EVENT_PTR(mem_ld_nhm),
285 * topdown events for Intel Core CPUs.
287 * The events are all in slots, which is a free slot in a 4 wide
288 * pipeline. Some events are already reported in slots, for cycle
289 * events we multiply by the pipeline width (4).
291 * With Hyper Threading on, topdown metrics are either summed or averaged
292 * between the threads of a core: (count_t0 + count_t1).
294 * For the average case the metric is always scaled to pipeline width,
295 * so we use factor 2 ((count_t0 + count_t1) / 2 * 4)
298 EVENT_ATTR_STR_HT(topdown-total-slots, td_total_slots,
299 "event=0x3c,umask=0x0", /* cpu_clk_unhalted.thread */
300 "event=0x3c,umask=0x0,any=1"); /* cpu_clk_unhalted.thread_any */
301 EVENT_ATTR_STR_HT(topdown-total-slots.scale, td_total_slots_scale, "4", "2");
302 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued,
303 "event=0xe,umask=0x1"); /* uops_issued.any */
304 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired,
305 "event=0xc2,umask=0x2"); /* uops_retired.retire_slots */
306 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles,
307 "event=0x9c,umask=0x1"); /* idq_uops_not_delivered_core */
308 EVENT_ATTR_STR_HT(topdown-recovery-bubbles, td_recovery_bubbles,
309 "event=0xd,umask=0x3,cmask=1", /* int_misc.recovery_cycles */
310 "event=0xd,umask=0x3,cmask=1,any=1"); /* int_misc.recovery_cycles_any */
311 EVENT_ATTR_STR_HT(topdown-recovery-bubbles.scale, td_recovery_bubbles_scale,
314 static struct attribute *snb_events_attrs[] = {
315 EVENT_PTR(td_slots_issued),
316 EVENT_PTR(td_slots_retired),
317 EVENT_PTR(td_fetch_bubbles),
318 EVENT_PTR(td_total_slots),
319 EVENT_PTR(td_total_slots_scale),
320 EVENT_PTR(td_recovery_bubbles),
321 EVENT_PTR(td_recovery_bubbles_scale),
325 static struct attribute *snb_mem_events_attrs[] = {
326 EVENT_PTR(mem_ld_snb),
327 EVENT_PTR(mem_st_snb),
331 static struct event_constraint intel_hsw_event_constraints[] = {
332 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
333 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
334 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
335 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */
336 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
337 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
338 /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
339 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4),
340 /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
341 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4),
342 /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
343 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf),
346 * When HT is off these events can only run on the bottom 4 counters
347 * When HT is on, they are impacted by the HT bug and require EXCL access
349 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
350 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
351 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
352 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
357 static struct event_constraint intel_bdw_event_constraints[] = {
358 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
359 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
360 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
361 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */
362 INTEL_UBIT_EVENT_CONSTRAINT(0x8a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_MISS */
364 * when HT is off, these can only run on the bottom 4 counters
366 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */
367 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */
368 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */
369 INTEL_EVENT_CONSTRAINT(0xcd, 0xf), /* MEM_TRANS_RETIRED.* */
373 static u64 intel_pmu_event_map(int hw_event)
375 return intel_perfmon_event_map[hw_event];
379 * Notes on the events:
380 * - data reads do not include code reads (comparable to earlier tables)
381 * - data counts include speculative execution (except L1 write, dtlb, bpu)
382 * - remote node access includes remote memory, remote cache, remote mmio.
383 * - prefetches are not included in the counts.
384 * - icache miss does not include decoded icache
387 #define SKL_DEMAND_DATA_RD BIT_ULL(0)
388 #define SKL_DEMAND_RFO BIT_ULL(1)
389 #define SKL_ANY_RESPONSE BIT_ULL(16)
390 #define SKL_SUPPLIER_NONE BIT_ULL(17)
391 #define SKL_L3_MISS_LOCAL_DRAM BIT_ULL(26)
392 #define SKL_L3_MISS_REMOTE_HOP0_DRAM BIT_ULL(27)
393 #define SKL_L3_MISS_REMOTE_HOP1_DRAM BIT_ULL(28)
394 #define SKL_L3_MISS_REMOTE_HOP2P_DRAM BIT_ULL(29)
395 #define SKL_L3_MISS (SKL_L3_MISS_LOCAL_DRAM| \
396 SKL_L3_MISS_REMOTE_HOP0_DRAM| \
397 SKL_L3_MISS_REMOTE_HOP1_DRAM| \
398 SKL_L3_MISS_REMOTE_HOP2P_DRAM)
399 #define SKL_SPL_HIT BIT_ULL(30)
400 #define SKL_SNOOP_NONE BIT_ULL(31)
401 #define SKL_SNOOP_NOT_NEEDED BIT_ULL(32)
402 #define SKL_SNOOP_MISS BIT_ULL(33)
403 #define SKL_SNOOP_HIT_NO_FWD BIT_ULL(34)
404 #define SKL_SNOOP_HIT_WITH_FWD BIT_ULL(35)
405 #define SKL_SNOOP_HITM BIT_ULL(36)
406 #define SKL_SNOOP_NON_DRAM BIT_ULL(37)
407 #define SKL_ANY_SNOOP (SKL_SPL_HIT|SKL_SNOOP_NONE| \
408 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \
409 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \
410 SKL_SNOOP_HITM|SKL_SNOOP_NON_DRAM)
411 #define SKL_DEMAND_READ SKL_DEMAND_DATA_RD
412 #define SKL_SNOOP_DRAM (SKL_SNOOP_NONE| \
413 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \
414 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \
415 SKL_SNOOP_HITM|SKL_SPL_HIT)
416 #define SKL_DEMAND_WRITE SKL_DEMAND_RFO
417 #define SKL_LLC_ACCESS SKL_ANY_RESPONSE
418 #define SKL_L3_MISS_REMOTE (SKL_L3_MISS_REMOTE_HOP0_DRAM| \
419 SKL_L3_MISS_REMOTE_HOP1_DRAM| \
420 SKL_L3_MISS_REMOTE_HOP2P_DRAM)
422 static __initconst const u64 skl_hw_cache_event_ids
423 [PERF_COUNT_HW_CACHE_MAX]
424 [PERF_COUNT_HW_CACHE_OP_MAX]
425 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
429 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */
430 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */
433 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */
434 [ C(RESULT_MISS) ] = 0x0,
436 [ C(OP_PREFETCH) ] = {
437 [ C(RESULT_ACCESS) ] = 0x0,
438 [ C(RESULT_MISS) ] = 0x0,
443 [ C(RESULT_ACCESS) ] = 0x0,
444 [ C(RESULT_MISS) ] = 0x283, /* ICACHE_64B.MISS */
447 [ C(RESULT_ACCESS) ] = -1,
448 [ C(RESULT_MISS) ] = -1,
450 [ C(OP_PREFETCH) ] = {
451 [ C(RESULT_ACCESS) ] = 0x0,
452 [ C(RESULT_MISS) ] = 0x0,
457 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
458 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
461 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
462 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
464 [ C(OP_PREFETCH) ] = {
465 [ C(RESULT_ACCESS) ] = 0x0,
466 [ C(RESULT_MISS) ] = 0x0,
471 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */
472 [ C(RESULT_MISS) ] = 0xe08, /* DTLB_LOAD_MISSES.WALK_COMPLETED */
475 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */
476 [ C(RESULT_MISS) ] = 0xe49, /* DTLB_STORE_MISSES.WALK_COMPLETED */
478 [ C(OP_PREFETCH) ] = {
479 [ C(RESULT_ACCESS) ] = 0x0,
480 [ C(RESULT_MISS) ] = 0x0,
485 [ C(RESULT_ACCESS) ] = 0x2085, /* ITLB_MISSES.STLB_HIT */
486 [ C(RESULT_MISS) ] = 0xe85, /* ITLB_MISSES.WALK_COMPLETED */
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) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */
500 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */
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) ] = 0x1b7, /* OFFCORE_RESPONSE */
514 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
517 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
518 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
520 [ C(OP_PREFETCH) ] = {
521 [ C(RESULT_ACCESS) ] = 0x0,
522 [ C(RESULT_MISS) ] = 0x0,
527 static __initconst const u64 skl_hw_cache_extra_regs
528 [PERF_COUNT_HW_CACHE_MAX]
529 [PERF_COUNT_HW_CACHE_OP_MAX]
530 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
534 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ|
535 SKL_LLC_ACCESS|SKL_ANY_SNOOP,
536 [ C(RESULT_MISS) ] = SKL_DEMAND_READ|
537 SKL_L3_MISS|SKL_ANY_SNOOP|
541 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE|
542 SKL_LLC_ACCESS|SKL_ANY_SNOOP,
543 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE|
544 SKL_L3_MISS|SKL_ANY_SNOOP|
547 [ C(OP_PREFETCH) ] = {
548 [ C(RESULT_ACCESS) ] = 0x0,
549 [ C(RESULT_MISS) ] = 0x0,
554 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ|
555 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM,
556 [ C(RESULT_MISS) ] = SKL_DEMAND_READ|
557 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM,
560 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE|
561 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM,
562 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE|
563 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM,
565 [ C(OP_PREFETCH) ] = {
566 [ C(RESULT_ACCESS) ] = 0x0,
567 [ C(RESULT_MISS) ] = 0x0,
572 #define SNB_DMND_DATA_RD (1ULL << 0)
573 #define SNB_DMND_RFO (1ULL << 1)
574 #define SNB_DMND_IFETCH (1ULL << 2)
575 #define SNB_DMND_WB (1ULL << 3)
576 #define SNB_PF_DATA_RD (1ULL << 4)
577 #define SNB_PF_RFO (1ULL << 5)
578 #define SNB_PF_IFETCH (1ULL << 6)
579 #define SNB_LLC_DATA_RD (1ULL << 7)
580 #define SNB_LLC_RFO (1ULL << 8)
581 #define SNB_LLC_IFETCH (1ULL << 9)
582 #define SNB_BUS_LOCKS (1ULL << 10)
583 #define SNB_STRM_ST (1ULL << 11)
584 #define SNB_OTHER (1ULL << 15)
585 #define SNB_RESP_ANY (1ULL << 16)
586 #define SNB_NO_SUPP (1ULL << 17)
587 #define SNB_LLC_HITM (1ULL << 18)
588 #define SNB_LLC_HITE (1ULL << 19)
589 #define SNB_LLC_HITS (1ULL << 20)
590 #define SNB_LLC_HITF (1ULL << 21)
591 #define SNB_LOCAL (1ULL << 22)
592 #define SNB_REMOTE (0xffULL << 23)
593 #define SNB_SNP_NONE (1ULL << 31)
594 #define SNB_SNP_NOT_NEEDED (1ULL << 32)
595 #define SNB_SNP_MISS (1ULL << 33)
596 #define SNB_NO_FWD (1ULL << 34)
597 #define SNB_SNP_FWD (1ULL << 35)
598 #define SNB_HITM (1ULL << 36)
599 #define SNB_NON_DRAM (1ULL << 37)
601 #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD)
602 #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO)
603 #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
605 #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \
606 SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \
609 #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY)
610 #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY)
612 #define SNB_L3_ACCESS SNB_RESP_ANY
613 #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM)
615 static __initconst const u64 snb_hw_cache_extra_regs
616 [PERF_COUNT_HW_CACHE_MAX]
617 [PERF_COUNT_HW_CACHE_OP_MAX]
618 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
622 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS,
623 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_L3_MISS,
626 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS,
627 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_L3_MISS,
629 [ C(OP_PREFETCH) ] = {
630 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS,
631 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_L3_MISS,
636 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY,
637 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_DRAM_REMOTE,
640 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY,
641 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE,
643 [ C(OP_PREFETCH) ] = {
644 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY,
645 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE,
650 static __initconst const u64 snb_hw_cache_event_ids
651 [PERF_COUNT_HW_CACHE_MAX]
652 [PERF_COUNT_HW_CACHE_OP_MAX]
653 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
657 [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */
658 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */
661 [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */
662 [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */
664 [ C(OP_PREFETCH) ] = {
665 [ C(RESULT_ACCESS) ] = 0x0,
666 [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */
671 [ C(RESULT_ACCESS) ] = 0x0,
672 [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */
675 [ C(RESULT_ACCESS) ] = -1,
676 [ C(RESULT_MISS) ] = -1,
678 [ C(OP_PREFETCH) ] = {
679 [ C(RESULT_ACCESS) ] = 0x0,
680 [ C(RESULT_MISS) ] = 0x0,
685 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
686 [ C(RESULT_ACCESS) ] = 0x01b7,
687 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
688 [ C(RESULT_MISS) ] = 0x01b7,
691 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
692 [ C(RESULT_ACCESS) ] = 0x01b7,
693 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
694 [ C(RESULT_MISS) ] = 0x01b7,
696 [ C(OP_PREFETCH) ] = {
697 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
698 [ C(RESULT_ACCESS) ] = 0x01b7,
699 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
700 [ C(RESULT_MISS) ] = 0x01b7,
705 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */
706 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */
709 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */
710 [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
712 [ C(OP_PREFETCH) ] = {
713 [ C(RESULT_ACCESS) ] = 0x0,
714 [ C(RESULT_MISS) ] = 0x0,
719 [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */
720 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */
723 [ C(RESULT_ACCESS) ] = -1,
724 [ C(RESULT_MISS) ] = -1,
726 [ C(OP_PREFETCH) ] = {
727 [ C(RESULT_ACCESS) ] = -1,
728 [ C(RESULT_MISS) ] = -1,
733 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
734 [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
737 [ C(RESULT_ACCESS) ] = -1,
738 [ C(RESULT_MISS) ] = -1,
740 [ C(OP_PREFETCH) ] = {
741 [ C(RESULT_ACCESS) ] = -1,
742 [ C(RESULT_MISS) ] = -1,
747 [ C(RESULT_ACCESS) ] = 0x01b7,
748 [ C(RESULT_MISS) ] = 0x01b7,
751 [ C(RESULT_ACCESS) ] = 0x01b7,
752 [ C(RESULT_MISS) ] = 0x01b7,
754 [ C(OP_PREFETCH) ] = {
755 [ C(RESULT_ACCESS) ] = 0x01b7,
756 [ C(RESULT_MISS) ] = 0x01b7,
763 * Notes on the events:
764 * - data reads do not include code reads (comparable to earlier tables)
765 * - data counts include speculative execution (except L1 write, dtlb, bpu)
766 * - remote node access includes remote memory, remote cache, remote mmio.
767 * - prefetches are not included in the counts because they are not
771 #define HSW_DEMAND_DATA_RD BIT_ULL(0)
772 #define HSW_DEMAND_RFO BIT_ULL(1)
773 #define HSW_ANY_RESPONSE BIT_ULL(16)
774 #define HSW_SUPPLIER_NONE BIT_ULL(17)
775 #define HSW_L3_MISS_LOCAL_DRAM BIT_ULL(22)
776 #define HSW_L3_MISS_REMOTE_HOP0 BIT_ULL(27)
777 #define HSW_L3_MISS_REMOTE_HOP1 BIT_ULL(28)
778 #define HSW_L3_MISS_REMOTE_HOP2P BIT_ULL(29)
779 #define HSW_L3_MISS (HSW_L3_MISS_LOCAL_DRAM| \
780 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \
781 HSW_L3_MISS_REMOTE_HOP2P)
782 #define HSW_SNOOP_NONE BIT_ULL(31)
783 #define HSW_SNOOP_NOT_NEEDED BIT_ULL(32)
784 #define HSW_SNOOP_MISS BIT_ULL(33)
785 #define HSW_SNOOP_HIT_NO_FWD BIT_ULL(34)
786 #define HSW_SNOOP_HIT_WITH_FWD BIT_ULL(35)
787 #define HSW_SNOOP_HITM BIT_ULL(36)
788 #define HSW_SNOOP_NON_DRAM BIT_ULL(37)
789 #define HSW_ANY_SNOOP (HSW_SNOOP_NONE| \
790 HSW_SNOOP_NOT_NEEDED|HSW_SNOOP_MISS| \
791 HSW_SNOOP_HIT_NO_FWD|HSW_SNOOP_HIT_WITH_FWD| \
792 HSW_SNOOP_HITM|HSW_SNOOP_NON_DRAM)
793 #define HSW_SNOOP_DRAM (HSW_ANY_SNOOP & ~HSW_SNOOP_NON_DRAM)
794 #define HSW_DEMAND_READ HSW_DEMAND_DATA_RD
795 #define HSW_DEMAND_WRITE HSW_DEMAND_RFO
796 #define HSW_L3_MISS_REMOTE (HSW_L3_MISS_REMOTE_HOP0|\
797 HSW_L3_MISS_REMOTE_HOP1|HSW_L3_MISS_REMOTE_HOP2P)
798 #define HSW_LLC_ACCESS HSW_ANY_RESPONSE
800 #define BDW_L3_MISS_LOCAL BIT(26)
801 #define BDW_L3_MISS (BDW_L3_MISS_LOCAL| \
802 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \
803 HSW_L3_MISS_REMOTE_HOP2P)
806 static __initconst const u64 hsw_hw_cache_event_ids
807 [PERF_COUNT_HW_CACHE_MAX]
808 [PERF_COUNT_HW_CACHE_OP_MAX]
809 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
813 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
814 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */
817 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
818 [ C(RESULT_MISS) ] = 0x0,
820 [ C(OP_PREFETCH) ] = {
821 [ C(RESULT_ACCESS) ] = 0x0,
822 [ C(RESULT_MISS) ] = 0x0,
827 [ C(RESULT_ACCESS) ] = 0x0,
828 [ C(RESULT_MISS) ] = 0x280, /* ICACHE.MISSES */
831 [ C(RESULT_ACCESS) ] = -1,
832 [ C(RESULT_MISS) ] = -1,
834 [ C(OP_PREFETCH) ] = {
835 [ C(RESULT_ACCESS) ] = 0x0,
836 [ C(RESULT_MISS) ] = 0x0,
841 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
842 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
845 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
846 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
848 [ C(OP_PREFETCH) ] = {
849 [ C(RESULT_ACCESS) ] = 0x0,
850 [ C(RESULT_MISS) ] = 0x0,
855 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
856 [ C(RESULT_MISS) ] = 0x108, /* DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK */
859 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
860 [ C(RESULT_MISS) ] = 0x149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
862 [ C(OP_PREFETCH) ] = {
863 [ C(RESULT_ACCESS) ] = 0x0,
864 [ C(RESULT_MISS) ] = 0x0,
869 [ C(RESULT_ACCESS) ] = 0x6085, /* ITLB_MISSES.STLB_HIT */
870 [ C(RESULT_MISS) ] = 0x185, /* ITLB_MISSES.MISS_CAUSES_A_WALK */
873 [ C(RESULT_ACCESS) ] = -1,
874 [ C(RESULT_MISS) ] = -1,
876 [ C(OP_PREFETCH) ] = {
877 [ C(RESULT_ACCESS) ] = -1,
878 [ C(RESULT_MISS) ] = -1,
883 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */
884 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */
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) ] = 0x1b7, /* OFFCORE_RESPONSE */
898 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
901 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
902 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
904 [ C(OP_PREFETCH) ] = {
905 [ C(RESULT_ACCESS) ] = 0x0,
906 [ C(RESULT_MISS) ] = 0x0,
911 static __initconst const u64 hsw_hw_cache_extra_regs
912 [PERF_COUNT_HW_CACHE_MAX]
913 [PERF_COUNT_HW_CACHE_OP_MAX]
914 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
918 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ|
920 [ C(RESULT_MISS) ] = HSW_DEMAND_READ|
921 HSW_L3_MISS|HSW_ANY_SNOOP,
924 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE|
926 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE|
927 HSW_L3_MISS|HSW_ANY_SNOOP,
929 [ C(OP_PREFETCH) ] = {
930 [ C(RESULT_ACCESS) ] = 0x0,
931 [ C(RESULT_MISS) ] = 0x0,
936 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ|
937 HSW_L3_MISS_LOCAL_DRAM|
939 [ C(RESULT_MISS) ] = HSW_DEMAND_READ|
944 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE|
945 HSW_L3_MISS_LOCAL_DRAM|
947 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE|
951 [ C(OP_PREFETCH) ] = {
952 [ C(RESULT_ACCESS) ] = 0x0,
953 [ C(RESULT_MISS) ] = 0x0,
958 static __initconst const u64 westmere_hw_cache_event_ids
959 [PERF_COUNT_HW_CACHE_MAX]
960 [PERF_COUNT_HW_CACHE_OP_MAX]
961 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
965 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
966 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
969 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
970 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
972 [ C(OP_PREFETCH) ] = {
973 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
974 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
979 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
980 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
983 [ C(RESULT_ACCESS) ] = -1,
984 [ C(RESULT_MISS) ] = -1,
986 [ C(OP_PREFETCH) ] = {
987 [ C(RESULT_ACCESS) ] = 0x0,
988 [ C(RESULT_MISS) ] = 0x0,
993 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
994 [ C(RESULT_ACCESS) ] = 0x01b7,
995 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
996 [ C(RESULT_MISS) ] = 0x01b7,
999 * Use RFO, not WRITEBACK, because a write miss would typically occur
1003 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1004 [ C(RESULT_ACCESS) ] = 0x01b7,
1005 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1006 [ C(RESULT_MISS) ] = 0x01b7,
1008 [ C(OP_PREFETCH) ] = {
1009 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1010 [ C(RESULT_ACCESS) ] = 0x01b7,
1011 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1012 [ C(RESULT_MISS) ] = 0x01b7,
1017 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
1018 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
1021 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
1022 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
1024 [ C(OP_PREFETCH) ] = {
1025 [ C(RESULT_ACCESS) ] = 0x0,
1026 [ C(RESULT_MISS) ] = 0x0,
1031 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
1032 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */
1035 [ C(RESULT_ACCESS) ] = -1,
1036 [ C(RESULT_MISS) ] = -1,
1038 [ C(OP_PREFETCH) ] = {
1039 [ C(RESULT_ACCESS) ] = -1,
1040 [ C(RESULT_MISS) ] = -1,
1045 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1046 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
1049 [ C(RESULT_ACCESS) ] = -1,
1050 [ C(RESULT_MISS) ] = -1,
1052 [ C(OP_PREFETCH) ] = {
1053 [ C(RESULT_ACCESS) ] = -1,
1054 [ C(RESULT_MISS) ] = -1,
1059 [ C(RESULT_ACCESS) ] = 0x01b7,
1060 [ C(RESULT_MISS) ] = 0x01b7,
1063 [ C(RESULT_ACCESS) ] = 0x01b7,
1064 [ C(RESULT_MISS) ] = 0x01b7,
1066 [ C(OP_PREFETCH) ] = {
1067 [ C(RESULT_ACCESS) ] = 0x01b7,
1068 [ C(RESULT_MISS) ] = 0x01b7,
1074 * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits;
1075 * See IA32 SDM Vol 3B 30.6.1.3
1078 #define NHM_DMND_DATA_RD (1 << 0)
1079 #define NHM_DMND_RFO (1 << 1)
1080 #define NHM_DMND_IFETCH (1 << 2)
1081 #define NHM_DMND_WB (1 << 3)
1082 #define NHM_PF_DATA_RD (1 << 4)
1083 #define NHM_PF_DATA_RFO (1 << 5)
1084 #define NHM_PF_IFETCH (1 << 6)
1085 #define NHM_OFFCORE_OTHER (1 << 7)
1086 #define NHM_UNCORE_HIT (1 << 8)
1087 #define NHM_OTHER_CORE_HIT_SNP (1 << 9)
1088 #define NHM_OTHER_CORE_HITM (1 << 10)
1090 #define NHM_REMOTE_CACHE_FWD (1 << 12)
1091 #define NHM_REMOTE_DRAM (1 << 13)
1092 #define NHM_LOCAL_DRAM (1 << 14)
1093 #define NHM_NON_DRAM (1 << 15)
1095 #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD)
1096 #define NHM_REMOTE (NHM_REMOTE_DRAM)
1098 #define NHM_DMND_READ (NHM_DMND_DATA_RD)
1099 #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB)
1100 #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO)
1102 #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM)
1103 #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD)
1104 #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS)
1106 static __initconst const u64 nehalem_hw_cache_extra_regs
1107 [PERF_COUNT_HW_CACHE_MAX]
1108 [PERF_COUNT_HW_CACHE_OP_MAX]
1109 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1113 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS,
1114 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS,
1117 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS,
1118 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS,
1120 [ C(OP_PREFETCH) ] = {
1121 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS,
1122 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS,
1127 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE,
1128 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE,
1131 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE,
1132 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE,
1134 [ C(OP_PREFETCH) ] = {
1135 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE,
1136 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE,
1141 static __initconst const u64 nehalem_hw_cache_event_ids
1142 [PERF_COUNT_HW_CACHE_MAX]
1143 [PERF_COUNT_HW_CACHE_OP_MAX]
1144 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1148 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
1149 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
1152 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
1153 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
1155 [ C(OP_PREFETCH) ] = {
1156 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
1157 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
1162 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
1163 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
1166 [ C(RESULT_ACCESS) ] = -1,
1167 [ C(RESULT_MISS) ] = -1,
1169 [ C(OP_PREFETCH) ] = {
1170 [ C(RESULT_ACCESS) ] = 0x0,
1171 [ C(RESULT_MISS) ] = 0x0,
1176 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1177 [ C(RESULT_ACCESS) ] = 0x01b7,
1178 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
1179 [ C(RESULT_MISS) ] = 0x01b7,
1182 * Use RFO, not WRITEBACK, because a write miss would typically occur
1186 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1187 [ C(RESULT_ACCESS) ] = 0x01b7,
1188 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1189 [ C(RESULT_MISS) ] = 0x01b7,
1191 [ C(OP_PREFETCH) ] = {
1192 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1193 [ C(RESULT_ACCESS) ] = 0x01b7,
1194 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1195 [ C(RESULT_MISS) ] = 0x01b7,
1200 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
1201 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
1204 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
1205 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
1207 [ C(OP_PREFETCH) ] = {
1208 [ C(RESULT_ACCESS) ] = 0x0,
1209 [ C(RESULT_MISS) ] = 0x0,
1214 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
1215 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */
1218 [ C(RESULT_ACCESS) ] = -1,
1219 [ C(RESULT_MISS) ] = -1,
1221 [ C(OP_PREFETCH) ] = {
1222 [ C(RESULT_ACCESS) ] = -1,
1223 [ C(RESULT_MISS) ] = -1,
1228 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1229 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
1232 [ C(RESULT_ACCESS) ] = -1,
1233 [ C(RESULT_MISS) ] = -1,
1235 [ C(OP_PREFETCH) ] = {
1236 [ C(RESULT_ACCESS) ] = -1,
1237 [ C(RESULT_MISS) ] = -1,
1242 [ C(RESULT_ACCESS) ] = 0x01b7,
1243 [ C(RESULT_MISS) ] = 0x01b7,
1246 [ C(RESULT_ACCESS) ] = 0x01b7,
1247 [ C(RESULT_MISS) ] = 0x01b7,
1249 [ C(OP_PREFETCH) ] = {
1250 [ C(RESULT_ACCESS) ] = 0x01b7,
1251 [ C(RESULT_MISS) ] = 0x01b7,
1256 static __initconst const u64 core2_hw_cache_event_ids
1257 [PERF_COUNT_HW_CACHE_MAX]
1258 [PERF_COUNT_HW_CACHE_OP_MAX]
1259 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1263 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
1264 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
1267 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
1268 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
1270 [ C(OP_PREFETCH) ] = {
1271 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */
1272 [ C(RESULT_MISS) ] = 0,
1277 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */
1278 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */
1281 [ C(RESULT_ACCESS) ] = -1,
1282 [ C(RESULT_MISS) ] = -1,
1284 [ C(OP_PREFETCH) ] = {
1285 [ C(RESULT_ACCESS) ] = 0,
1286 [ C(RESULT_MISS) ] = 0,
1291 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
1292 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
1295 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
1296 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
1298 [ C(OP_PREFETCH) ] = {
1299 [ C(RESULT_ACCESS) ] = 0,
1300 [ C(RESULT_MISS) ] = 0,
1305 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
1306 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */
1309 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
1310 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */
1312 [ C(OP_PREFETCH) ] = {
1313 [ C(RESULT_ACCESS) ] = 0,
1314 [ C(RESULT_MISS) ] = 0,
1319 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1320 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */
1323 [ C(RESULT_ACCESS) ] = -1,
1324 [ C(RESULT_MISS) ] = -1,
1326 [ C(OP_PREFETCH) ] = {
1327 [ C(RESULT_ACCESS) ] = -1,
1328 [ C(RESULT_MISS) ] = -1,
1333 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1334 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1337 [ C(RESULT_ACCESS) ] = -1,
1338 [ C(RESULT_MISS) ] = -1,
1340 [ C(OP_PREFETCH) ] = {
1341 [ C(RESULT_ACCESS) ] = -1,
1342 [ C(RESULT_MISS) ] = -1,
1347 static __initconst const u64 atom_hw_cache_event_ids
1348 [PERF_COUNT_HW_CACHE_MAX]
1349 [PERF_COUNT_HW_CACHE_OP_MAX]
1350 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1354 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */
1355 [ C(RESULT_MISS) ] = 0,
1358 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */
1359 [ C(RESULT_MISS) ] = 0,
1361 [ C(OP_PREFETCH) ] = {
1362 [ C(RESULT_ACCESS) ] = 0x0,
1363 [ C(RESULT_MISS) ] = 0,
1368 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
1369 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
1372 [ C(RESULT_ACCESS) ] = -1,
1373 [ C(RESULT_MISS) ] = -1,
1375 [ C(OP_PREFETCH) ] = {
1376 [ C(RESULT_ACCESS) ] = 0,
1377 [ C(RESULT_MISS) ] = 0,
1382 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
1383 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
1386 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
1387 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
1389 [ C(OP_PREFETCH) ] = {
1390 [ C(RESULT_ACCESS) ] = 0,
1391 [ C(RESULT_MISS) ] = 0,
1396 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */
1397 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */
1400 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */
1401 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */
1403 [ C(OP_PREFETCH) ] = {
1404 [ C(RESULT_ACCESS) ] = 0,
1405 [ C(RESULT_MISS) ] = 0,
1410 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1411 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */
1414 [ C(RESULT_ACCESS) ] = -1,
1415 [ C(RESULT_MISS) ] = -1,
1417 [ C(OP_PREFETCH) ] = {
1418 [ C(RESULT_ACCESS) ] = -1,
1419 [ C(RESULT_MISS) ] = -1,
1424 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1425 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1428 [ C(RESULT_ACCESS) ] = -1,
1429 [ C(RESULT_MISS) ] = -1,
1431 [ C(OP_PREFETCH) ] = {
1432 [ C(RESULT_ACCESS) ] = -1,
1433 [ C(RESULT_MISS) ] = -1,
1438 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_slm, "event=0x3c");
1439 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_slm, "2");
1440 /* no_alloc_cycles.not_delivered */
1441 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_slm,
1442 "event=0xca,umask=0x50");
1443 EVENT_ATTR_STR(topdown-fetch-bubbles.scale, td_fetch_bubbles_scale_slm, "2");
1444 /* uops_retired.all */
1445 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_slm,
1446 "event=0xc2,umask=0x10");
1447 /* uops_retired.all */
1448 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_slm,
1449 "event=0xc2,umask=0x10");
1451 static struct attribute *slm_events_attrs[] = {
1452 EVENT_PTR(td_total_slots_slm),
1453 EVENT_PTR(td_total_slots_scale_slm),
1454 EVENT_PTR(td_fetch_bubbles_slm),
1455 EVENT_PTR(td_fetch_bubbles_scale_slm),
1456 EVENT_PTR(td_slots_issued_slm),
1457 EVENT_PTR(td_slots_retired_slm),
1461 static struct extra_reg intel_slm_extra_regs[] __read_mostly =
1463 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
1464 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x768005ffffull, RSP_0),
1465 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x368005ffffull, RSP_1),
1469 #define SLM_DMND_READ SNB_DMND_DATA_RD
1470 #define SLM_DMND_WRITE SNB_DMND_RFO
1471 #define SLM_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
1473 #define SLM_SNP_ANY (SNB_SNP_NONE|SNB_SNP_MISS|SNB_NO_FWD|SNB_HITM)
1474 #define SLM_LLC_ACCESS SNB_RESP_ANY
1475 #define SLM_LLC_MISS (SLM_SNP_ANY|SNB_NON_DRAM)
1477 static __initconst const u64 slm_hw_cache_extra_regs
1478 [PERF_COUNT_HW_CACHE_MAX]
1479 [PERF_COUNT_HW_CACHE_OP_MAX]
1480 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1484 [ C(RESULT_ACCESS) ] = SLM_DMND_READ|SLM_LLC_ACCESS,
1485 [ C(RESULT_MISS) ] = 0,
1488 [ C(RESULT_ACCESS) ] = SLM_DMND_WRITE|SLM_LLC_ACCESS,
1489 [ C(RESULT_MISS) ] = SLM_DMND_WRITE|SLM_LLC_MISS,
1491 [ C(OP_PREFETCH) ] = {
1492 [ C(RESULT_ACCESS) ] = SLM_DMND_PREFETCH|SLM_LLC_ACCESS,
1493 [ C(RESULT_MISS) ] = SLM_DMND_PREFETCH|SLM_LLC_MISS,
1498 static __initconst const u64 slm_hw_cache_event_ids
1499 [PERF_COUNT_HW_CACHE_MAX]
1500 [PERF_COUNT_HW_CACHE_OP_MAX]
1501 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1505 [ C(RESULT_ACCESS) ] = 0,
1506 [ C(RESULT_MISS) ] = 0x0104, /* LD_DCU_MISS */
1509 [ C(RESULT_ACCESS) ] = 0,
1510 [ C(RESULT_MISS) ] = 0,
1512 [ C(OP_PREFETCH) ] = {
1513 [ C(RESULT_ACCESS) ] = 0,
1514 [ C(RESULT_MISS) ] = 0,
1519 [ C(RESULT_ACCESS) ] = 0x0380, /* ICACHE.ACCESSES */
1520 [ C(RESULT_MISS) ] = 0x0280, /* ICACGE.MISSES */
1523 [ C(RESULT_ACCESS) ] = -1,
1524 [ C(RESULT_MISS) ] = -1,
1526 [ C(OP_PREFETCH) ] = {
1527 [ C(RESULT_ACCESS) ] = 0,
1528 [ C(RESULT_MISS) ] = 0,
1533 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1534 [ C(RESULT_ACCESS) ] = 0x01b7,
1535 [ C(RESULT_MISS) ] = 0,
1538 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1539 [ C(RESULT_ACCESS) ] = 0x01b7,
1540 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1541 [ C(RESULT_MISS) ] = 0x01b7,
1543 [ C(OP_PREFETCH) ] = {
1544 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1545 [ C(RESULT_ACCESS) ] = 0x01b7,
1546 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1547 [ C(RESULT_MISS) ] = 0x01b7,
1552 [ C(RESULT_ACCESS) ] = 0,
1553 [ C(RESULT_MISS) ] = 0x0804, /* LD_DTLB_MISS */
1556 [ C(RESULT_ACCESS) ] = 0,
1557 [ C(RESULT_MISS) ] = 0,
1559 [ C(OP_PREFETCH) ] = {
1560 [ C(RESULT_ACCESS) ] = 0,
1561 [ C(RESULT_MISS) ] = 0,
1566 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1567 [ C(RESULT_MISS) ] = 0x40205, /* PAGE_WALKS.I_SIDE_WALKS */
1570 [ C(RESULT_ACCESS) ] = -1,
1571 [ C(RESULT_MISS) ] = -1,
1573 [ C(OP_PREFETCH) ] = {
1574 [ C(RESULT_ACCESS) ] = -1,
1575 [ C(RESULT_MISS) ] = -1,
1580 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1581 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1584 [ C(RESULT_ACCESS) ] = -1,
1585 [ C(RESULT_MISS) ] = -1,
1587 [ C(OP_PREFETCH) ] = {
1588 [ C(RESULT_ACCESS) ] = -1,
1589 [ C(RESULT_MISS) ] = -1,
1594 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_glm, "event=0x3c");
1595 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_glm, "3");
1596 /* UOPS_NOT_DELIVERED.ANY */
1597 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_glm, "event=0x9c");
1598 /* ISSUE_SLOTS_NOT_CONSUMED.RECOVERY */
1599 EVENT_ATTR_STR(topdown-recovery-bubbles, td_recovery_bubbles_glm, "event=0xca,umask=0x02");
1600 /* UOPS_RETIRED.ANY */
1601 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_glm, "event=0xc2");
1602 /* UOPS_ISSUED.ANY */
1603 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_glm, "event=0x0e");
1605 static struct attribute *glm_events_attrs[] = {
1606 EVENT_PTR(td_total_slots_glm),
1607 EVENT_PTR(td_total_slots_scale_glm),
1608 EVENT_PTR(td_fetch_bubbles_glm),
1609 EVENT_PTR(td_recovery_bubbles_glm),
1610 EVENT_PTR(td_slots_issued_glm),
1611 EVENT_PTR(td_slots_retired_glm),
1615 static struct extra_reg intel_glm_extra_regs[] __read_mostly = {
1616 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
1617 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x760005ffbfull, RSP_0),
1618 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x360005ffbfull, RSP_1),
1622 #define GLM_DEMAND_DATA_RD BIT_ULL(0)
1623 #define GLM_DEMAND_RFO BIT_ULL(1)
1624 #define GLM_ANY_RESPONSE BIT_ULL(16)
1625 #define GLM_SNP_NONE_OR_MISS BIT_ULL(33)
1626 #define GLM_DEMAND_READ GLM_DEMAND_DATA_RD
1627 #define GLM_DEMAND_WRITE GLM_DEMAND_RFO
1628 #define GLM_DEMAND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
1629 #define GLM_LLC_ACCESS GLM_ANY_RESPONSE
1630 #define GLM_SNP_ANY (GLM_SNP_NONE_OR_MISS|SNB_NO_FWD|SNB_HITM)
1631 #define GLM_LLC_MISS (GLM_SNP_ANY|SNB_NON_DRAM)
1633 static __initconst const u64 glm_hw_cache_event_ids
1634 [PERF_COUNT_HW_CACHE_MAX]
1635 [PERF_COUNT_HW_CACHE_OP_MAX]
1636 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1639 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
1640 [C(RESULT_MISS)] = 0x0,
1643 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
1644 [C(RESULT_MISS)] = 0x0,
1646 [C(OP_PREFETCH)] = {
1647 [C(RESULT_ACCESS)] = 0x0,
1648 [C(RESULT_MISS)] = 0x0,
1653 [C(RESULT_ACCESS)] = 0x0380, /* ICACHE.ACCESSES */
1654 [C(RESULT_MISS)] = 0x0280, /* ICACHE.MISSES */
1657 [C(RESULT_ACCESS)] = -1,
1658 [C(RESULT_MISS)] = -1,
1660 [C(OP_PREFETCH)] = {
1661 [C(RESULT_ACCESS)] = 0x0,
1662 [C(RESULT_MISS)] = 0x0,
1667 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1668 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1671 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1672 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1674 [C(OP_PREFETCH)] = {
1675 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1676 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1681 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
1682 [C(RESULT_MISS)] = 0x0,
1685 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
1686 [C(RESULT_MISS)] = 0x0,
1688 [C(OP_PREFETCH)] = {
1689 [C(RESULT_ACCESS)] = 0x0,
1690 [C(RESULT_MISS)] = 0x0,
1695 [C(RESULT_ACCESS)] = 0x00c0, /* INST_RETIRED.ANY_P */
1696 [C(RESULT_MISS)] = 0x0481, /* ITLB.MISS */
1699 [C(RESULT_ACCESS)] = -1,
1700 [C(RESULT_MISS)] = -1,
1702 [C(OP_PREFETCH)] = {
1703 [C(RESULT_ACCESS)] = -1,
1704 [C(RESULT_MISS)] = -1,
1709 [C(RESULT_ACCESS)] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1710 [C(RESULT_MISS)] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
1713 [C(RESULT_ACCESS)] = -1,
1714 [C(RESULT_MISS)] = -1,
1716 [C(OP_PREFETCH)] = {
1717 [C(RESULT_ACCESS)] = -1,
1718 [C(RESULT_MISS)] = -1,
1723 static __initconst const u64 glm_hw_cache_extra_regs
1724 [PERF_COUNT_HW_CACHE_MAX]
1725 [PERF_COUNT_HW_CACHE_OP_MAX]
1726 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1729 [C(RESULT_ACCESS)] = GLM_DEMAND_READ|
1731 [C(RESULT_MISS)] = GLM_DEMAND_READ|
1735 [C(RESULT_ACCESS)] = GLM_DEMAND_WRITE|
1737 [C(RESULT_MISS)] = GLM_DEMAND_WRITE|
1740 [C(OP_PREFETCH)] = {
1741 [C(RESULT_ACCESS)] = GLM_DEMAND_PREFETCH|
1743 [C(RESULT_MISS)] = GLM_DEMAND_PREFETCH|
1749 static __initconst const u64 glp_hw_cache_event_ids
1750 [PERF_COUNT_HW_CACHE_MAX]
1751 [PERF_COUNT_HW_CACHE_OP_MAX]
1752 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1755 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
1756 [C(RESULT_MISS)] = 0x0,
1759 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
1760 [C(RESULT_MISS)] = 0x0,
1762 [C(OP_PREFETCH)] = {
1763 [C(RESULT_ACCESS)] = 0x0,
1764 [C(RESULT_MISS)] = 0x0,
1769 [C(RESULT_ACCESS)] = 0x0380, /* ICACHE.ACCESSES */
1770 [C(RESULT_MISS)] = 0x0280, /* ICACHE.MISSES */
1773 [C(RESULT_ACCESS)] = -1,
1774 [C(RESULT_MISS)] = -1,
1776 [C(OP_PREFETCH)] = {
1777 [C(RESULT_ACCESS)] = 0x0,
1778 [C(RESULT_MISS)] = 0x0,
1783 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1784 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1787 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1788 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1790 [C(OP_PREFETCH)] = {
1791 [C(RESULT_ACCESS)] = 0x0,
1792 [C(RESULT_MISS)] = 0x0,
1797 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
1798 [C(RESULT_MISS)] = 0xe08, /* DTLB_LOAD_MISSES.WALK_COMPLETED */
1801 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
1802 [C(RESULT_MISS)] = 0xe49, /* DTLB_STORE_MISSES.WALK_COMPLETED */
1804 [C(OP_PREFETCH)] = {
1805 [C(RESULT_ACCESS)] = 0x0,
1806 [C(RESULT_MISS)] = 0x0,
1811 [C(RESULT_ACCESS)] = 0x00c0, /* INST_RETIRED.ANY_P */
1812 [C(RESULT_MISS)] = 0x0481, /* ITLB.MISS */
1815 [C(RESULT_ACCESS)] = -1,
1816 [C(RESULT_MISS)] = -1,
1818 [C(OP_PREFETCH)] = {
1819 [C(RESULT_ACCESS)] = -1,
1820 [C(RESULT_MISS)] = -1,
1825 [C(RESULT_ACCESS)] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1826 [C(RESULT_MISS)] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
1829 [C(RESULT_ACCESS)] = -1,
1830 [C(RESULT_MISS)] = -1,
1832 [C(OP_PREFETCH)] = {
1833 [C(RESULT_ACCESS)] = -1,
1834 [C(RESULT_MISS)] = -1,
1839 static __initconst const u64 glp_hw_cache_extra_regs
1840 [PERF_COUNT_HW_CACHE_MAX]
1841 [PERF_COUNT_HW_CACHE_OP_MAX]
1842 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1845 [C(RESULT_ACCESS)] = GLM_DEMAND_READ|
1847 [C(RESULT_MISS)] = GLM_DEMAND_READ|
1851 [C(RESULT_ACCESS)] = GLM_DEMAND_WRITE|
1853 [C(RESULT_MISS)] = GLM_DEMAND_WRITE|
1856 [C(OP_PREFETCH)] = {
1857 [C(RESULT_ACCESS)] = 0x0,
1858 [C(RESULT_MISS)] = 0x0,
1863 #define TNT_LOCAL_DRAM BIT_ULL(26)
1864 #define TNT_DEMAND_READ GLM_DEMAND_DATA_RD
1865 #define TNT_DEMAND_WRITE GLM_DEMAND_RFO
1866 #define TNT_LLC_ACCESS GLM_ANY_RESPONSE
1867 #define TNT_SNP_ANY (SNB_SNP_NOT_NEEDED|SNB_SNP_MISS| \
1868 SNB_NO_FWD|SNB_SNP_FWD|SNB_HITM)
1869 #define TNT_LLC_MISS (TNT_SNP_ANY|SNB_NON_DRAM|TNT_LOCAL_DRAM)
1871 static __initconst const u64 tnt_hw_cache_extra_regs
1872 [PERF_COUNT_HW_CACHE_MAX]
1873 [PERF_COUNT_HW_CACHE_OP_MAX]
1874 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1877 [C(RESULT_ACCESS)] = TNT_DEMAND_READ|
1879 [C(RESULT_MISS)] = TNT_DEMAND_READ|
1883 [C(RESULT_ACCESS)] = TNT_DEMAND_WRITE|
1885 [C(RESULT_MISS)] = TNT_DEMAND_WRITE|
1888 [C(OP_PREFETCH)] = {
1889 [C(RESULT_ACCESS)] = 0x0,
1890 [C(RESULT_MISS)] = 0x0,
1895 static struct extra_reg intel_tnt_extra_regs[] __read_mostly = {
1896 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
1897 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x800ff0ffffff9fffull, RSP_0),
1898 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0xff0ffffff9fffull, RSP_1),
1902 #define KNL_OT_L2_HITE BIT_ULL(19) /* Other Tile L2 Hit */
1903 #define KNL_OT_L2_HITF BIT_ULL(20) /* Other Tile L2 Hit */
1904 #define KNL_MCDRAM_LOCAL BIT_ULL(21)
1905 #define KNL_MCDRAM_FAR BIT_ULL(22)
1906 #define KNL_DDR_LOCAL BIT_ULL(23)
1907 #define KNL_DDR_FAR BIT_ULL(24)
1908 #define KNL_DRAM_ANY (KNL_MCDRAM_LOCAL | KNL_MCDRAM_FAR | \
1909 KNL_DDR_LOCAL | KNL_DDR_FAR)
1910 #define KNL_L2_READ SLM_DMND_READ
1911 #define KNL_L2_WRITE SLM_DMND_WRITE
1912 #define KNL_L2_PREFETCH SLM_DMND_PREFETCH
1913 #define KNL_L2_ACCESS SLM_LLC_ACCESS
1914 #define KNL_L2_MISS (KNL_OT_L2_HITE | KNL_OT_L2_HITF | \
1915 KNL_DRAM_ANY | SNB_SNP_ANY | \
1918 static __initconst const u64 knl_hw_cache_extra_regs
1919 [PERF_COUNT_HW_CACHE_MAX]
1920 [PERF_COUNT_HW_CACHE_OP_MAX]
1921 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1924 [C(RESULT_ACCESS)] = KNL_L2_READ | KNL_L2_ACCESS,
1925 [C(RESULT_MISS)] = 0,
1928 [C(RESULT_ACCESS)] = KNL_L2_WRITE | KNL_L2_ACCESS,
1929 [C(RESULT_MISS)] = KNL_L2_WRITE | KNL_L2_MISS,
1931 [C(OP_PREFETCH)] = {
1932 [C(RESULT_ACCESS)] = KNL_L2_PREFETCH | KNL_L2_ACCESS,
1933 [C(RESULT_MISS)] = KNL_L2_PREFETCH | KNL_L2_MISS,
1939 * Used from PMIs where the LBRs are already disabled.
1941 * This function could be called consecutively. It is required to remain in
1942 * disabled state if called consecutively.
1944 * During consecutive calls, the same disable value will be written to related
1945 * registers, so the PMU state remains unchanged.
1947 * intel_bts events don't coexist with intel PMU's BTS events because of
1948 * x86_add_exclusive(x86_lbr_exclusive_lbr); there's no need to keep them
1949 * disabled around intel PMU's event batching etc, only inside the PMI handler.
1951 static void __intel_pmu_disable_all(void)
1953 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1955 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
1957 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask))
1958 intel_pmu_disable_bts();
1960 intel_pmu_pebs_disable_all();
1963 static void intel_pmu_disable_all(void)
1965 __intel_pmu_disable_all();
1966 intel_pmu_lbr_disable_all();
1969 static void __intel_pmu_enable_all(int added, bool pmi)
1971 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1973 intel_pmu_pebs_enable_all();
1974 intel_pmu_lbr_enable_all(pmi);
1975 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL,
1976 x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask);
1978 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
1979 struct perf_event *event =
1980 cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
1982 if (WARN_ON_ONCE(!event))
1985 intel_pmu_enable_bts(event->hw.config);
1989 static void intel_pmu_enable_all(int added)
1991 __intel_pmu_enable_all(added, false);
1996 * Intel Errata AAK100 (model 26)
1997 * Intel Errata AAP53 (model 30)
1998 * Intel Errata BD53 (model 44)
2000 * The official story:
2001 * These chips need to be 'reset' when adding counters by programming the
2002 * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either
2003 * in sequence on the same PMC or on different PMCs.
2005 * In practise it appears some of these events do in fact count, and
2006 * we need to program all 4 events.
2008 static void intel_pmu_nhm_workaround(void)
2010 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2011 static const unsigned long nhm_magic[4] = {
2017 struct perf_event *event;
2021 * The Errata requires below steps:
2022 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL;
2023 * 2) Configure 4 PERFEVTSELx with the magic events and clear
2024 * the corresponding PMCx;
2025 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL;
2026 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL;
2027 * 5) Clear 4 pairs of ERFEVTSELx and PMCx;
2031 * The real steps we choose are a little different from above.
2032 * A) To reduce MSR operations, we don't run step 1) as they
2033 * are already cleared before this function is called;
2034 * B) Call x86_perf_event_update to save PMCx before configuring
2035 * PERFEVTSELx with magic number;
2036 * C) With step 5), we do clear only when the PERFEVTSELx is
2037 * not used currently.
2038 * D) Call x86_perf_event_set_period to restore PMCx;
2041 /* We always operate 4 pairs of PERF Counters */
2042 for (i = 0; i < 4; i++) {
2043 event = cpuc->events[i];
2045 x86_perf_event_update(event);
2048 for (i = 0; i < 4; i++) {
2049 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]);
2050 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0);
2053 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf);
2054 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
2056 for (i = 0; i < 4; i++) {
2057 event = cpuc->events[i];
2060 x86_perf_event_set_period(event);
2061 __x86_pmu_enable_event(&event->hw,
2062 ARCH_PERFMON_EVENTSEL_ENABLE);
2064 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0);
2068 static void intel_pmu_nhm_enable_all(int added)
2071 intel_pmu_nhm_workaround();
2072 intel_pmu_enable_all(added);
2075 static void intel_set_tfa(struct cpu_hw_events *cpuc, bool on)
2077 u64 val = on ? MSR_TFA_RTM_FORCE_ABORT : 0;
2079 if (cpuc->tfa_shadow != val) {
2080 cpuc->tfa_shadow = val;
2081 wrmsrl(MSR_TSX_FORCE_ABORT, val);
2085 static void intel_tfa_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr)
2088 * We're going to use PMC3, make sure TFA is set before we touch it.
2091 intel_set_tfa(cpuc, true);
2094 static void intel_tfa_pmu_enable_all(int added)
2096 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2099 * If we find PMC3 is no longer used when we enable the PMU, we can
2102 if (!test_bit(3, cpuc->active_mask))
2103 intel_set_tfa(cpuc, false);
2105 intel_pmu_enable_all(added);
2108 static void enable_counter_freeze(void)
2110 update_debugctlmsr(get_debugctlmsr() |
2111 DEBUGCTLMSR_FREEZE_PERFMON_ON_PMI);
2114 static void disable_counter_freeze(void)
2116 update_debugctlmsr(get_debugctlmsr() &
2117 ~DEBUGCTLMSR_FREEZE_PERFMON_ON_PMI);
2120 static inline u64 intel_pmu_get_status(void)
2124 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
2129 static inline void intel_pmu_ack_status(u64 ack)
2131 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
2134 static void intel_pmu_disable_fixed(struct hw_perf_event *hwc)
2136 int idx = hwc->idx - INTEL_PMC_IDX_FIXED;
2139 mask = 0xfULL << (idx * 4);
2141 rdmsrl(hwc->config_base, ctrl_val);
2143 wrmsrl(hwc->config_base, ctrl_val);
2146 static inline bool event_is_checkpointed(struct perf_event *event)
2148 return (event->hw.config & HSW_IN_TX_CHECKPOINTED) != 0;
2151 static void intel_pmu_disable_event(struct perf_event *event)
2153 struct hw_perf_event *hwc = &event->hw;
2154 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2156 if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) {
2157 intel_pmu_disable_bts();
2158 intel_pmu_drain_bts_buffer();
2162 cpuc->intel_ctrl_guest_mask &= ~(1ull << hwc->idx);
2163 cpuc->intel_ctrl_host_mask &= ~(1ull << hwc->idx);
2164 cpuc->intel_cp_status &= ~(1ull << hwc->idx);
2166 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL))
2167 intel_pmu_disable_fixed(hwc);
2169 x86_pmu_disable_event(event);
2172 * Needs to be called after x86_pmu_disable_event,
2173 * so we don't trigger the event without PEBS bit set.
2175 if (unlikely(event->attr.precise_ip))
2176 intel_pmu_pebs_disable(event);
2179 static void intel_pmu_del_event(struct perf_event *event)
2181 if (needs_branch_stack(event))
2182 intel_pmu_lbr_del(event);
2183 if (event->attr.precise_ip)
2184 intel_pmu_pebs_del(event);
2187 static void intel_pmu_read_event(struct perf_event *event)
2189 if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
2190 intel_pmu_auto_reload_read(event);
2192 x86_perf_event_update(event);
2195 static void intel_pmu_enable_fixed(struct perf_event *event)
2197 struct hw_perf_event *hwc = &event->hw;
2198 int idx = hwc->idx - INTEL_PMC_IDX_FIXED;
2199 u64 ctrl_val, mask, bits = 0;
2202 * Enable IRQ generation (0x8), if not PEBS,
2203 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
2206 if (!event->attr.precise_ip)
2208 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
2210 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
2214 * ANY bit is supported in v3 and up
2216 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
2220 mask = 0xfULL << (idx * 4);
2222 if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip) {
2223 bits |= ICL_FIXED_0_ADAPTIVE << (idx * 4);
2224 mask |= ICL_FIXED_0_ADAPTIVE << (idx * 4);
2227 rdmsrl(hwc->config_base, ctrl_val);
2230 wrmsrl(hwc->config_base, ctrl_val);
2233 static void intel_pmu_enable_event(struct perf_event *event)
2235 struct hw_perf_event *hwc = &event->hw;
2236 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2238 if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) {
2239 if (!__this_cpu_read(cpu_hw_events.enabled))
2242 intel_pmu_enable_bts(hwc->config);
2246 if (event->attr.exclude_host)
2247 cpuc->intel_ctrl_guest_mask |= (1ull << hwc->idx);
2248 if (event->attr.exclude_guest)
2249 cpuc->intel_ctrl_host_mask |= (1ull << hwc->idx);
2251 if (unlikely(event_is_checkpointed(event)))
2252 cpuc->intel_cp_status |= (1ull << hwc->idx);
2254 if (unlikely(event->attr.precise_ip))
2255 intel_pmu_pebs_enable(event);
2257 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
2258 intel_pmu_enable_fixed(event);
2262 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
2265 static void intel_pmu_add_event(struct perf_event *event)
2267 if (event->attr.precise_ip)
2268 intel_pmu_pebs_add(event);
2269 if (needs_branch_stack(event))
2270 intel_pmu_lbr_add(event);
2274 * Save and restart an expired event. Called by NMI contexts,
2275 * so it has to be careful about preempting normal event ops:
2277 int intel_pmu_save_and_restart(struct perf_event *event)
2279 x86_perf_event_update(event);
2281 * For a checkpointed counter always reset back to 0. This
2282 * avoids a situation where the counter overflows, aborts the
2283 * transaction and is then set back to shortly before the
2284 * overflow, and overflows and aborts again.
2286 if (unlikely(event_is_checkpointed(event))) {
2287 /* No race with NMIs because the counter should not be armed */
2288 wrmsrl(event->hw.event_base, 0);
2289 local64_set(&event->hw.prev_count, 0);
2291 return x86_perf_event_set_period(event);
2294 static void intel_pmu_reset(void)
2296 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
2297 unsigned long flags;
2300 if (!x86_pmu.num_counters)
2303 local_irq_save(flags);
2305 pr_info("clearing PMU state on CPU#%d\n", smp_processor_id());
2307 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
2308 wrmsrl_safe(x86_pmu_config_addr(idx), 0ull);
2309 wrmsrl_safe(x86_pmu_event_addr(idx), 0ull);
2311 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++)
2312 wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
2315 ds->bts_index = ds->bts_buffer_base;
2317 /* Ack all overflows and disable fixed counters */
2318 if (x86_pmu.version >= 2) {
2319 intel_pmu_ack_status(intel_pmu_get_status());
2320 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
2323 /* Reset LBRs and LBR freezing */
2324 if (x86_pmu.lbr_nr) {
2325 update_debugctlmsr(get_debugctlmsr() &
2326 ~(DEBUGCTLMSR_FREEZE_LBRS_ON_PMI|DEBUGCTLMSR_LBR));
2329 local_irq_restore(flags);
2332 static int handle_pmi_common(struct pt_regs *regs, u64 status)
2334 struct perf_sample_data data;
2335 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2336 struct perf_guest_info_callbacks *guest_cbs;
2340 inc_irq_stat(apic_perf_irqs);
2343 * Ignore a range of extra bits in status that do not indicate
2344 * overflow by themselves.
2346 status &= ~(GLOBAL_STATUS_COND_CHG |
2347 GLOBAL_STATUS_ASIF |
2348 GLOBAL_STATUS_LBRS_FROZEN);
2352 * In case multiple PEBS events are sampled at the same time,
2353 * it is possible to have GLOBAL_STATUS bit 62 set indicating
2354 * PEBS buffer overflow and also seeing at most 3 PEBS counters
2355 * having their bits set in the status register. This is a sign
2356 * that there was at least one PEBS record pending at the time
2357 * of the PMU interrupt. PEBS counters must only be processed
2358 * via the drain_pebs() calls and not via the regular sample
2359 * processing loop coming after that the function, otherwise
2360 * phony regular samples may be generated in the sampling buffer
2361 * not marked with the EXACT tag. Another possibility is to have
2362 * one PEBS event and at least one non-PEBS event whic hoverflows
2363 * while PEBS has armed. In this case, bit 62 of GLOBAL_STATUS will
2364 * not be set, yet the overflow status bit for the PEBS counter will
2367 * To avoid this problem, we systematically ignore the PEBS-enabled
2368 * counters from the GLOBAL_STATUS mask and we always process PEBS
2369 * events via drain_pebs().
2371 if (x86_pmu.flags & PMU_FL_PEBS_ALL)
2372 status &= ~cpuc->pebs_enabled;
2374 status &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK);
2377 * PEBS overflow sets bit 62 in the global status register
2379 if (__test_and_clear_bit(62, (unsigned long *)&status)) {
2381 x86_pmu.drain_pebs(regs);
2382 status &= x86_pmu.intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI;
2388 if (__test_and_clear_bit(55, (unsigned long *)&status)) {
2391 guest_cbs = perf_get_guest_cbs();
2392 if (unlikely(guest_cbs && guest_cbs->is_in_guest() &&
2393 guest_cbs->handle_intel_pt_intr))
2394 guest_cbs->handle_intel_pt_intr();
2396 intel_pt_interrupt();
2400 * Checkpointed counters can lead to 'spurious' PMIs because the
2401 * rollback caused by the PMI will have cleared the overflow status
2402 * bit. Therefore always force probe these counters.
2404 status |= cpuc->intel_cp_status;
2406 for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
2407 struct perf_event *event = cpuc->events[bit];
2411 if (!test_bit(bit, cpuc->active_mask))
2414 if (!intel_pmu_save_and_restart(event))
2417 perf_sample_data_init(&data, 0, event->hw.last_period);
2419 if (has_branch_stack(event))
2420 data.br_stack = &cpuc->lbr_stack;
2422 if (perf_event_overflow(event, &data, regs))
2423 x86_pmu_stop(event, 0);
2429 static bool disable_counter_freezing = true;
2430 static int __init intel_perf_counter_freezing_setup(char *s)
2434 if (kstrtobool(s, &res))
2437 disable_counter_freezing = !res;
2440 __setup("perf_v4_pmi=", intel_perf_counter_freezing_setup);
2443 * Simplified handler for Arch Perfmon v4:
2444 * - We rely on counter freezing/unfreezing to enable/disable the PMU.
2445 * This is done automatically on PMU ack.
2446 * - Ack the PMU only after the APIC.
2449 static int intel_pmu_handle_irq_v4(struct pt_regs *regs)
2451 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2455 int pmu_enabled = cpuc->enabled;
2458 /* PMU has been disabled because of counter freezing */
2460 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
2462 intel_bts_disable_local();
2463 handled = intel_pmu_drain_bts_buffer();
2464 handled += intel_bts_interrupt();
2466 status = intel_pmu_get_status();
2470 intel_pmu_lbr_read();
2471 if (++loops > 100) {
2475 WARN(1, "perfevents: irq loop stuck!\n");
2476 perf_event_print_debug();
2484 handled += handle_pmi_common(regs, status);
2486 /* Ack the PMI in the APIC */
2487 apic_write(APIC_LVTPC, APIC_DM_NMI);
2490 * The counters start counting immediately while ack the status.
2491 * Make it as close as possible to IRET. This avoids bogus
2492 * freezing on Skylake CPUs.
2495 intel_pmu_ack_status(status);
2498 * CPU may issues two PMIs very close to each other.
2499 * When the PMI handler services the first one, the
2500 * GLOBAL_STATUS is already updated to reflect both.
2501 * When it IRETs, the second PMI is immediately
2502 * handled and it sees clear status. At the meantime,
2503 * there may be a third PMI, because the freezing bit
2504 * isn't set since the ack in first PMI handlers.
2505 * Double check if there is more work to be done.
2507 status = intel_pmu_get_status();
2513 intel_bts_enable_local();
2514 cpuc->enabled = pmu_enabled;
2519 * This handler is triggered by the local APIC, so the APIC IRQ handling
2522 static int intel_pmu_handle_irq(struct pt_regs *regs)
2524 struct cpu_hw_events *cpuc;
2530 cpuc = this_cpu_ptr(&cpu_hw_events);
2533 * Save the PMU state.
2534 * It needs to be restored when leaving the handler.
2536 pmu_enabled = cpuc->enabled;
2538 * No known reason to not always do late ACK,
2539 * but just in case do it opt-in.
2541 if (!x86_pmu.late_ack)
2542 apic_write(APIC_LVTPC, APIC_DM_NMI);
2543 intel_bts_disable_local();
2545 __intel_pmu_disable_all();
2546 handled = intel_pmu_drain_bts_buffer();
2547 handled += intel_bts_interrupt();
2548 status = intel_pmu_get_status();
2554 intel_pmu_lbr_read();
2555 intel_pmu_ack_status(status);
2556 if (++loops > 100) {
2560 WARN(1, "perfevents: irq loop stuck!\n");
2561 perf_event_print_debug();
2568 handled += handle_pmi_common(regs, status);
2571 * Repeat if there is more work to be done:
2573 status = intel_pmu_get_status();
2578 /* Only restore PMU state when it's active. See x86_pmu_disable(). */
2579 cpuc->enabled = pmu_enabled;
2581 __intel_pmu_enable_all(0, true);
2582 intel_bts_enable_local();
2585 * Only unmask the NMI after the overflow counters
2586 * have been reset. This avoids spurious NMIs on
2589 if (x86_pmu.late_ack)
2590 apic_write(APIC_LVTPC, APIC_DM_NMI);
2594 static struct event_constraint *
2595 intel_bts_constraints(struct perf_event *event)
2597 if (unlikely(intel_pmu_has_bts(event)))
2598 return &bts_constraint;
2603 static int intel_alt_er(int idx, u64 config)
2607 if (!(x86_pmu.flags & PMU_FL_HAS_RSP_1))
2610 if (idx == EXTRA_REG_RSP_0)
2611 alt_idx = EXTRA_REG_RSP_1;
2613 if (idx == EXTRA_REG_RSP_1)
2614 alt_idx = EXTRA_REG_RSP_0;
2616 if (config & ~x86_pmu.extra_regs[alt_idx].valid_mask)
2622 static void intel_fixup_er(struct perf_event *event, int idx)
2624 event->hw.extra_reg.idx = idx;
2626 if (idx == EXTRA_REG_RSP_0) {
2627 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
2628 event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_0].event;
2629 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0;
2630 } else if (idx == EXTRA_REG_RSP_1) {
2631 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
2632 event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_1].event;
2633 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1;
2638 * manage allocation of shared extra msr for certain events
2641 * per-cpu: to be shared between the various events on a single PMU
2642 * per-core: per-cpu + shared by HT threads
2644 static struct event_constraint *
2645 __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc,
2646 struct perf_event *event,
2647 struct hw_perf_event_extra *reg)
2649 struct event_constraint *c = &emptyconstraint;
2650 struct er_account *era;
2651 unsigned long flags;
2655 * reg->alloc can be set due to existing state, so for fake cpuc we
2656 * need to ignore this, otherwise we might fail to allocate proper fake
2657 * state for this extra reg constraint. Also see the comment below.
2659 if (reg->alloc && !cpuc->is_fake)
2660 return NULL; /* call x86_get_event_constraint() */
2663 era = &cpuc->shared_regs->regs[idx];
2665 * we use spin_lock_irqsave() to avoid lockdep issues when
2666 * passing a fake cpuc
2668 raw_spin_lock_irqsave(&era->lock, flags);
2670 if (!atomic_read(&era->ref) || era->config == reg->config) {
2673 * If its a fake cpuc -- as per validate_{group,event}() we
2674 * shouldn't touch event state and we can avoid doing so
2675 * since both will only call get_event_constraints() once
2676 * on each event, this avoids the need for reg->alloc.
2678 * Not doing the ER fixup will only result in era->reg being
2679 * wrong, but since we won't actually try and program hardware
2680 * this isn't a problem either.
2682 if (!cpuc->is_fake) {
2683 if (idx != reg->idx)
2684 intel_fixup_er(event, idx);
2687 * x86_schedule_events() can call get_event_constraints()
2688 * multiple times on events in the case of incremental
2689 * scheduling(). reg->alloc ensures we only do the ER
2695 /* lock in msr value */
2696 era->config = reg->config;
2697 era->reg = reg->reg;
2700 atomic_inc(&era->ref);
2703 * need to call x86_get_event_constraint()
2704 * to check if associated event has constraints
2708 idx = intel_alt_er(idx, reg->config);
2709 if (idx != reg->idx) {
2710 raw_spin_unlock_irqrestore(&era->lock, flags);
2714 raw_spin_unlock_irqrestore(&era->lock, flags);
2720 __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc,
2721 struct hw_perf_event_extra *reg)
2723 struct er_account *era;
2726 * Only put constraint if extra reg was actually allocated. Also takes
2727 * care of event which do not use an extra shared reg.
2729 * Also, if this is a fake cpuc we shouldn't touch any event state
2730 * (reg->alloc) and we don't care about leaving inconsistent cpuc state
2731 * either since it'll be thrown out.
2733 if (!reg->alloc || cpuc->is_fake)
2736 era = &cpuc->shared_regs->regs[reg->idx];
2738 /* one fewer user */
2739 atomic_dec(&era->ref);
2741 /* allocate again next time */
2745 static struct event_constraint *
2746 intel_shared_regs_constraints(struct cpu_hw_events *cpuc,
2747 struct perf_event *event)
2749 struct event_constraint *c = NULL, *d;
2750 struct hw_perf_event_extra *xreg, *breg;
2752 xreg = &event->hw.extra_reg;
2753 if (xreg->idx != EXTRA_REG_NONE) {
2754 c = __intel_shared_reg_get_constraints(cpuc, event, xreg);
2755 if (c == &emptyconstraint)
2758 breg = &event->hw.branch_reg;
2759 if (breg->idx != EXTRA_REG_NONE) {
2760 d = __intel_shared_reg_get_constraints(cpuc, event, breg);
2761 if (d == &emptyconstraint) {
2762 __intel_shared_reg_put_constraints(cpuc, xreg);
2769 struct event_constraint *
2770 x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
2771 struct perf_event *event)
2773 struct event_constraint *c;
2775 if (x86_pmu.event_constraints) {
2776 for_each_event_constraint(c, x86_pmu.event_constraints) {
2777 if (constraint_match(c, event->hw.config)) {
2778 event->hw.flags |= c->flags;
2784 return &unconstrained;
2787 static struct event_constraint *
2788 __intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
2789 struct perf_event *event)
2791 struct event_constraint *c;
2793 c = intel_bts_constraints(event);
2797 c = intel_shared_regs_constraints(cpuc, event);
2801 c = intel_pebs_constraints(event);
2805 return x86_get_event_constraints(cpuc, idx, event);
2809 intel_start_scheduling(struct cpu_hw_events *cpuc)
2811 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
2812 struct intel_excl_states *xl;
2813 int tid = cpuc->excl_thread_id;
2816 * nothing needed if in group validation mode
2818 if (cpuc->is_fake || !is_ht_workaround_enabled())
2822 * no exclusion needed
2824 if (WARN_ON_ONCE(!excl_cntrs))
2827 xl = &excl_cntrs->states[tid];
2829 xl->sched_started = true;
2831 * lock shared state until we are done scheduling
2832 * in stop_event_scheduling()
2833 * makes scheduling appear as a transaction
2835 raw_spin_lock(&excl_cntrs->lock);
2838 static void intel_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr)
2840 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
2841 struct event_constraint *c = cpuc->event_constraint[idx];
2842 struct intel_excl_states *xl;
2843 int tid = cpuc->excl_thread_id;
2845 if (cpuc->is_fake || !is_ht_workaround_enabled())
2848 if (WARN_ON_ONCE(!excl_cntrs))
2851 if (!(c->flags & PERF_X86_EVENT_DYNAMIC))
2854 xl = &excl_cntrs->states[tid];
2856 lockdep_assert_held(&excl_cntrs->lock);
2858 if (c->flags & PERF_X86_EVENT_EXCL)
2859 xl->state[cntr] = INTEL_EXCL_EXCLUSIVE;
2861 xl->state[cntr] = INTEL_EXCL_SHARED;
2865 intel_stop_scheduling(struct cpu_hw_events *cpuc)
2867 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
2868 struct intel_excl_states *xl;
2869 int tid = cpuc->excl_thread_id;
2872 * nothing needed if in group validation mode
2874 if (cpuc->is_fake || !is_ht_workaround_enabled())
2877 * no exclusion needed
2879 if (WARN_ON_ONCE(!excl_cntrs))
2882 xl = &excl_cntrs->states[tid];
2884 xl->sched_started = false;
2886 * release shared state lock (acquired in intel_start_scheduling())
2888 raw_spin_unlock(&excl_cntrs->lock);
2891 static struct event_constraint *
2892 dyn_constraint(struct cpu_hw_events *cpuc, struct event_constraint *c, int idx)
2894 WARN_ON_ONCE(!cpuc->constraint_list);
2896 if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) {
2897 struct event_constraint *cx;
2900 * grab pre-allocated constraint entry
2902 cx = &cpuc->constraint_list[idx];
2905 * initialize dynamic constraint
2906 * with static constraint
2911 * mark constraint as dynamic
2913 cx->flags |= PERF_X86_EVENT_DYNAMIC;
2920 static struct event_constraint *
2921 intel_get_excl_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
2922 int idx, struct event_constraint *c)
2924 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
2925 struct intel_excl_states *xlo;
2926 int tid = cpuc->excl_thread_id;
2930 * validating a group does not require
2931 * enforcing cross-thread exclusion
2933 if (cpuc->is_fake || !is_ht_workaround_enabled())
2937 * no exclusion needed
2939 if (WARN_ON_ONCE(!excl_cntrs))
2943 * because we modify the constraint, we need
2944 * to make a copy. Static constraints come
2945 * from static const tables.
2947 * only needed when constraint has not yet
2948 * been cloned (marked dynamic)
2950 c = dyn_constraint(cpuc, c, idx);
2953 * From here on, the constraint is dynamic.
2954 * Either it was just allocated above, or it
2955 * was allocated during a earlier invocation
2960 * state of sibling HT
2962 xlo = &excl_cntrs->states[tid ^ 1];
2965 * event requires exclusive counter access
2968 is_excl = c->flags & PERF_X86_EVENT_EXCL;
2969 if (is_excl && !(event->hw.flags & PERF_X86_EVENT_EXCL_ACCT)) {
2970 event->hw.flags |= PERF_X86_EVENT_EXCL_ACCT;
2971 if (!cpuc->n_excl++)
2972 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 1);
2976 * Modify static constraint with current dynamic
2979 * EXCLUSIVE: sibling counter measuring exclusive event
2980 * SHARED : sibling counter measuring non-exclusive event
2981 * UNUSED : sibling counter unused
2984 for_each_set_bit(i, c->idxmsk, X86_PMC_IDX_MAX) {
2986 * exclusive event in sibling counter
2987 * our corresponding counter cannot be used
2988 * regardless of our event
2990 if (xlo->state[i] == INTEL_EXCL_EXCLUSIVE) {
2991 __clear_bit(i, c->idxmsk);
2996 * if measuring an exclusive event, sibling
2997 * measuring non-exclusive, then counter cannot
3000 if (is_excl && xlo->state[i] == INTEL_EXCL_SHARED) {
3001 __clear_bit(i, c->idxmsk);
3008 * if we return an empty mask, then switch
3009 * back to static empty constraint to avoid
3010 * the cost of freeing later on
3013 c = &emptyconstraint;
3020 static struct event_constraint *
3021 intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3022 struct perf_event *event)
3024 struct event_constraint *c1, *c2;
3026 c1 = cpuc->event_constraint[idx];
3030 * - static constraint: no change across incremental scheduling calls
3031 * - dynamic constraint: handled by intel_get_excl_constraints()
3033 c2 = __intel_get_event_constraints(cpuc, idx, event);
3035 WARN_ON_ONCE(!(c1->flags & PERF_X86_EVENT_DYNAMIC));
3036 bitmap_copy(c1->idxmsk, c2->idxmsk, X86_PMC_IDX_MAX);
3037 c1->weight = c2->weight;
3041 if (cpuc->excl_cntrs)
3042 return intel_get_excl_constraints(cpuc, event, idx, c2);
3047 static void intel_put_excl_constraints(struct cpu_hw_events *cpuc,
3048 struct perf_event *event)
3050 struct hw_perf_event *hwc = &event->hw;
3051 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3052 int tid = cpuc->excl_thread_id;
3053 struct intel_excl_states *xl;
3056 * nothing needed if in group validation mode
3061 if (WARN_ON_ONCE(!excl_cntrs))
3064 if (hwc->flags & PERF_X86_EVENT_EXCL_ACCT) {
3065 hwc->flags &= ~PERF_X86_EVENT_EXCL_ACCT;
3066 if (!--cpuc->n_excl)
3067 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 0);
3071 * If event was actually assigned, then mark the counter state as
3074 if (hwc->idx >= 0) {
3075 xl = &excl_cntrs->states[tid];
3078 * put_constraint may be called from x86_schedule_events()
3079 * which already has the lock held so here make locking
3082 if (!xl->sched_started)
3083 raw_spin_lock(&excl_cntrs->lock);
3085 xl->state[hwc->idx] = INTEL_EXCL_UNUSED;
3087 if (!xl->sched_started)
3088 raw_spin_unlock(&excl_cntrs->lock);
3093 intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc,
3094 struct perf_event *event)
3096 struct hw_perf_event_extra *reg;
3098 reg = &event->hw.extra_reg;
3099 if (reg->idx != EXTRA_REG_NONE)
3100 __intel_shared_reg_put_constraints(cpuc, reg);
3102 reg = &event->hw.branch_reg;
3103 if (reg->idx != EXTRA_REG_NONE)
3104 __intel_shared_reg_put_constraints(cpuc, reg);
3107 static void intel_put_event_constraints(struct cpu_hw_events *cpuc,
3108 struct perf_event *event)
3110 intel_put_shared_regs_event_constraints(cpuc, event);
3113 * is PMU has exclusive counter restrictions, then
3114 * all events are subject to and must call the
3115 * put_excl_constraints() routine
3117 if (cpuc->excl_cntrs)
3118 intel_put_excl_constraints(cpuc, event);
3121 static void intel_pebs_aliases_core2(struct perf_event *event)
3123 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
3125 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
3126 * (0x003c) so that we can use it with PEBS.
3128 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
3129 * PEBS capable. However we can use INST_RETIRED.ANY_P
3130 * (0x00c0), which is a PEBS capable event, to get the same
3133 * INST_RETIRED.ANY_P counts the number of cycles that retires
3134 * CNTMASK instructions. By setting CNTMASK to a value (16)
3135 * larger than the maximum number of instructions that can be
3136 * retired per cycle (4) and then inverting the condition, we
3137 * count all cycles that retire 16 or less instructions, which
3140 * Thereby we gain a PEBS capable cycle counter.
3142 u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16);
3144 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
3145 event->hw.config = alt_config;
3149 static void intel_pebs_aliases_snb(struct perf_event *event)
3151 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
3153 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
3154 * (0x003c) so that we can use it with PEBS.
3156 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
3157 * PEBS capable. However we can use UOPS_RETIRED.ALL
3158 * (0x01c2), which is a PEBS capable event, to get the same
3161 * UOPS_RETIRED.ALL counts the number of cycles that retires
3162 * CNTMASK micro-ops. By setting CNTMASK to a value (16)
3163 * larger than the maximum number of micro-ops that can be
3164 * retired per cycle (4) and then inverting the condition, we
3165 * count all cycles that retire 16 or less micro-ops, which
3168 * Thereby we gain a PEBS capable cycle counter.
3170 u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16);
3172 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
3173 event->hw.config = alt_config;
3177 static void intel_pebs_aliases_precdist(struct perf_event *event)
3179 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
3181 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
3182 * (0x003c) so that we can use it with PEBS.
3184 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
3185 * PEBS capable. However we can use INST_RETIRED.PREC_DIST
3186 * (0x01c0), which is a PEBS capable event, to get the same
3189 * The PREC_DIST event has special support to minimize sample
3190 * shadowing effects. One drawback is that it can be
3191 * only programmed on counter 1, but that seems like an
3192 * acceptable trade off.
3194 u64 alt_config = X86_CONFIG(.event=0xc0, .umask=0x01, .inv=1, .cmask=16);
3196 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
3197 event->hw.config = alt_config;
3201 static void intel_pebs_aliases_ivb(struct perf_event *event)
3203 if (event->attr.precise_ip < 3)
3204 return intel_pebs_aliases_snb(event);
3205 return intel_pebs_aliases_precdist(event);
3208 static void intel_pebs_aliases_skl(struct perf_event *event)
3210 if (event->attr.precise_ip < 3)
3211 return intel_pebs_aliases_core2(event);
3212 return intel_pebs_aliases_precdist(event);
3215 static unsigned long intel_pmu_large_pebs_flags(struct perf_event *event)
3217 unsigned long flags = x86_pmu.large_pebs_flags;
3219 if (event->attr.use_clockid)
3220 flags &= ~PERF_SAMPLE_TIME;
3221 if (!event->attr.exclude_kernel)
3222 flags &= ~PERF_SAMPLE_REGS_USER;
3223 if (event->attr.sample_regs_user & ~PEBS_GP_REGS)
3224 flags &= ~(PERF_SAMPLE_REGS_USER | PERF_SAMPLE_REGS_INTR);
3228 static int intel_pmu_bts_config(struct perf_event *event)
3230 struct perf_event_attr *attr = &event->attr;
3232 if (unlikely(intel_pmu_has_bts(event))) {
3233 /* BTS is not supported by this architecture. */
3234 if (!x86_pmu.bts_active)
3237 /* BTS is currently only allowed for user-mode. */
3238 if (!attr->exclude_kernel)
3241 /* BTS is not allowed for precise events. */
3242 if (attr->precise_ip)
3245 /* disallow bts if conflicting events are present */
3246 if (x86_add_exclusive(x86_lbr_exclusive_lbr))
3249 event->destroy = hw_perf_lbr_event_destroy;
3255 static int core_pmu_hw_config(struct perf_event *event)
3257 int ret = x86_pmu_hw_config(event);
3262 return intel_pmu_bts_config(event);
3265 static int intel_pmu_hw_config(struct perf_event *event)
3267 int ret = x86_pmu_hw_config(event);
3272 ret = intel_pmu_bts_config(event);
3276 if (event->attr.precise_ip) {
3277 if (!(event->attr.freq || (event->attr.wakeup_events && !event->attr.watermark))) {
3278 event->hw.flags |= PERF_X86_EVENT_AUTO_RELOAD;
3279 if (!(event->attr.sample_type &
3280 ~intel_pmu_large_pebs_flags(event)))
3281 event->hw.flags |= PERF_X86_EVENT_LARGE_PEBS;
3283 if (x86_pmu.pebs_aliases)
3284 x86_pmu.pebs_aliases(event);
3286 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3287 event->attr.sample_type |= __PERF_SAMPLE_CALLCHAIN_EARLY;
3290 if (needs_branch_stack(event)) {
3291 ret = intel_pmu_setup_lbr_filter(event);
3296 * BTS is set up earlier in this path, so don't account twice
3298 if (!unlikely(intel_pmu_has_bts(event))) {
3299 /* disallow lbr if conflicting events are present */
3300 if (x86_add_exclusive(x86_lbr_exclusive_lbr))
3303 event->destroy = hw_perf_lbr_event_destroy;
3307 if (event->attr.aux_output) {
3308 if (!event->attr.precise_ip)
3311 event->hw.flags |= PERF_X86_EVENT_PEBS_VIA_PT;
3314 if (event->attr.type != PERF_TYPE_RAW)
3317 if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY))
3320 if (x86_pmu.version < 3)
3323 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3326 event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY;
3331 struct perf_guest_switch_msr *perf_guest_get_msrs(int *nr)
3333 if (x86_pmu.guest_get_msrs)
3334 return x86_pmu.guest_get_msrs(nr);
3338 EXPORT_SYMBOL_GPL(perf_guest_get_msrs);
3340 static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr)
3342 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3343 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
3345 arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL;
3346 arr[0].host = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask;
3347 arr[0].guest = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_host_mask;
3348 if (x86_pmu.flags & PMU_FL_PEBS_ALL)
3349 arr[0].guest &= ~cpuc->pebs_enabled;
3351 arr[0].guest &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK);
3354 if (x86_pmu.pebs && x86_pmu.pebs_no_isolation) {
3356 * If PMU counter has PEBS enabled it is not enough to
3357 * disable counter on a guest entry since PEBS memory
3358 * write can overshoot guest entry and corrupt guest
3359 * memory. Disabling PEBS solves the problem.
3361 * Don't do this if the CPU already enforces it.
3363 arr[1].msr = MSR_IA32_PEBS_ENABLE;
3364 arr[1].host = cpuc->pebs_enabled;
3372 static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr)
3374 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3375 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
3378 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
3379 struct perf_event *event = cpuc->events[idx];
3381 arr[idx].msr = x86_pmu_config_addr(idx);
3382 arr[idx].host = arr[idx].guest = 0;
3384 if (!test_bit(idx, cpuc->active_mask))
3387 arr[idx].host = arr[idx].guest =
3388 event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE;
3390 if (event->attr.exclude_host)
3391 arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
3392 else if (event->attr.exclude_guest)
3393 arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
3396 *nr = x86_pmu.num_counters;
3400 static void core_pmu_enable_event(struct perf_event *event)
3402 if (!event->attr.exclude_host)
3403 x86_pmu_enable_event(event);
3406 static void core_pmu_enable_all(int added)
3408 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3411 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
3412 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
3414 if (!test_bit(idx, cpuc->active_mask) ||
3415 cpuc->events[idx]->attr.exclude_host)
3418 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
3422 static int hsw_hw_config(struct perf_event *event)
3424 int ret = intel_pmu_hw_config(event);
3428 if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE))
3430 event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED);
3433 * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with
3434 * PEBS or in ANY thread mode. Since the results are non-sensical forbid
3437 if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) &&
3438 ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) ||
3439 event->attr.precise_ip > 0))
3442 if (event_is_checkpointed(event)) {
3444 * Sampling of checkpointed events can cause situations where
3445 * the CPU constantly aborts because of a overflow, which is
3446 * then checkpointed back and ignored. Forbid checkpointing
3449 * But still allow a long sampling period, so that perf stat
3452 if (event->attr.sample_period > 0 &&
3453 event->attr.sample_period < 0x7fffffff)
3459 static struct event_constraint counter0_constraint =
3460 INTEL_ALL_EVENT_CONSTRAINT(0, 0x1);
3462 static struct event_constraint counter2_constraint =
3463 EVENT_CONSTRAINT(0, 0x4, 0);
3465 static struct event_constraint fixed0_constraint =
3466 FIXED_EVENT_CONSTRAINT(0x00c0, 0);
3468 static struct event_constraint fixed0_counter0_constraint =
3469 INTEL_ALL_EVENT_CONSTRAINT(0, 0x100000001ULL);
3471 static struct event_constraint *
3472 hsw_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3473 struct perf_event *event)
3475 struct event_constraint *c;
3477 c = intel_get_event_constraints(cpuc, idx, event);
3479 /* Handle special quirk on in_tx_checkpointed only in counter 2 */
3480 if (event->hw.config & HSW_IN_TX_CHECKPOINTED) {
3481 if (c->idxmsk64 & (1U << 2))
3482 return &counter2_constraint;
3483 return &emptyconstraint;
3489 static struct event_constraint *
3490 icl_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3491 struct perf_event *event)
3494 * Fixed counter 0 has less skid.
3495 * Force instruction:ppp in Fixed counter 0
3497 if ((event->attr.precise_ip == 3) &&
3498 constraint_match(&fixed0_constraint, event->hw.config))
3499 return &fixed0_constraint;
3501 return hsw_get_event_constraints(cpuc, idx, event);
3504 static struct event_constraint *
3505 glp_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3506 struct perf_event *event)
3508 struct event_constraint *c;
3510 /* :ppp means to do reduced skid PEBS which is PMC0 only. */
3511 if (event->attr.precise_ip == 3)
3512 return &counter0_constraint;
3514 c = intel_get_event_constraints(cpuc, idx, event);
3519 static struct event_constraint *
3520 tnt_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3521 struct perf_event *event)
3523 struct event_constraint *c;
3526 * :ppp means to do reduced skid PEBS,
3527 * which is available on PMC0 and fixed counter 0.
3529 if (event->attr.precise_ip == 3) {
3530 /* Force instruction:ppp on PMC0 and Fixed counter 0 */
3531 if (constraint_match(&fixed0_constraint, event->hw.config))
3532 return &fixed0_counter0_constraint;
3534 return &counter0_constraint;
3537 c = intel_get_event_constraints(cpuc, idx, event);
3542 static bool allow_tsx_force_abort = true;
3544 static struct event_constraint *
3545 tfa_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3546 struct perf_event *event)
3548 struct event_constraint *c = hsw_get_event_constraints(cpuc, idx, event);
3551 * Without TFA we must not use PMC3.
3553 if (!allow_tsx_force_abort && test_bit(3, c->idxmsk)) {
3554 c = dyn_constraint(cpuc, c, idx);
3555 c->idxmsk64 &= ~(1ULL << 3);
3565 * The INST_RETIRED.ALL period always needs to have lowest 6 bits cleared
3566 * (BDM55) and it must not use a period smaller than 100 (BDM11). We combine
3567 * the two to enforce a minimum period of 128 (the smallest value that has bits
3568 * 0-5 cleared and >= 100).
3570 * Because of how the code in x86_perf_event_set_period() works, the truncation
3571 * of the lower 6 bits is 'harmless' as we'll occasionally add a longer period
3572 * to make up for the 'lost' events due to carrying the 'error' in period_left.
3574 * Therefore the effective (average) period matches the requested period,
3575 * despite coarser hardware granularity.
3577 static u64 bdw_limit_period(struct perf_event *event, u64 left)
3579 if ((event->hw.config & INTEL_ARCH_EVENT_MASK) ==
3580 X86_CONFIG(.event=0xc0, .umask=0x01)) {
3588 static u64 nhm_limit_period(struct perf_event *event, u64 left)
3590 return max(left, 32ULL);
3593 PMU_FORMAT_ATTR(event, "config:0-7" );
3594 PMU_FORMAT_ATTR(umask, "config:8-15" );
3595 PMU_FORMAT_ATTR(edge, "config:18" );
3596 PMU_FORMAT_ATTR(pc, "config:19" );
3597 PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */
3598 PMU_FORMAT_ATTR(inv, "config:23" );
3599 PMU_FORMAT_ATTR(cmask, "config:24-31" );
3600 PMU_FORMAT_ATTR(in_tx, "config:32");
3601 PMU_FORMAT_ATTR(in_tx_cp, "config:33");
3603 static struct attribute *intel_arch_formats_attr[] = {
3604 &format_attr_event.attr,
3605 &format_attr_umask.attr,
3606 &format_attr_edge.attr,
3607 &format_attr_pc.attr,
3608 &format_attr_inv.attr,
3609 &format_attr_cmask.attr,
3613 ssize_t intel_event_sysfs_show(char *page, u64 config)
3615 u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT);
3617 return x86_event_sysfs_show(page, config, event);
3620 static struct intel_shared_regs *allocate_shared_regs(int cpu)
3622 struct intel_shared_regs *regs;
3625 regs = kzalloc_node(sizeof(struct intel_shared_regs),
3626 GFP_KERNEL, cpu_to_node(cpu));
3629 * initialize the locks to keep lockdep happy
3631 for (i = 0; i < EXTRA_REG_MAX; i++)
3632 raw_spin_lock_init(®s->regs[i].lock);
3639 static struct intel_excl_cntrs *allocate_excl_cntrs(int cpu)
3641 struct intel_excl_cntrs *c;
3643 c = kzalloc_node(sizeof(struct intel_excl_cntrs),
3644 GFP_KERNEL, cpu_to_node(cpu));
3646 raw_spin_lock_init(&c->lock);
3653 int intel_cpuc_prepare(struct cpu_hw_events *cpuc, int cpu)
3655 cpuc->pebs_record_size = x86_pmu.pebs_record_size;
3657 if (x86_pmu.extra_regs || x86_pmu.lbr_sel_map) {
3658 cpuc->shared_regs = allocate_shared_regs(cpu);
3659 if (!cpuc->shared_regs)
3663 if (x86_pmu.flags & (PMU_FL_EXCL_CNTRS | PMU_FL_TFA)) {
3664 size_t sz = X86_PMC_IDX_MAX * sizeof(struct event_constraint);
3666 cpuc->constraint_list = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));
3667 if (!cpuc->constraint_list)
3668 goto err_shared_regs;
3671 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) {
3672 cpuc->excl_cntrs = allocate_excl_cntrs(cpu);
3673 if (!cpuc->excl_cntrs)
3674 goto err_constraint_list;
3676 cpuc->excl_thread_id = 0;
3681 err_constraint_list:
3682 kfree(cpuc->constraint_list);
3683 cpuc->constraint_list = NULL;
3686 kfree(cpuc->shared_regs);
3687 cpuc->shared_regs = NULL;
3693 static int intel_pmu_cpu_prepare(int cpu)
3695 return intel_cpuc_prepare(&per_cpu(cpu_hw_events, cpu), cpu);
3698 static void flip_smm_bit(void *data)
3700 unsigned long set = *(unsigned long *)data;
3703 msr_set_bit(MSR_IA32_DEBUGCTLMSR,
3704 DEBUGCTLMSR_FREEZE_IN_SMM_BIT);
3706 msr_clear_bit(MSR_IA32_DEBUGCTLMSR,
3707 DEBUGCTLMSR_FREEZE_IN_SMM_BIT);
3711 static void intel_pmu_cpu_starting(int cpu)
3713 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
3714 int core_id = topology_core_id(cpu);
3717 init_debug_store_on_cpu(cpu);
3719 * Deal with CPUs that don't clear their LBRs on power-up.
3721 intel_pmu_lbr_reset();
3723 cpuc->lbr_sel = NULL;
3725 if (x86_pmu.flags & PMU_FL_TFA) {
3726 WARN_ON_ONCE(cpuc->tfa_shadow);
3727 cpuc->tfa_shadow = ~0ULL;
3728 intel_set_tfa(cpuc, false);
3731 if (x86_pmu.version > 1)
3732 flip_smm_bit(&x86_pmu.attr_freeze_on_smi);
3734 if (x86_pmu.counter_freezing)
3735 enable_counter_freeze();
3737 if (!cpuc->shared_regs)
3740 if (!(x86_pmu.flags & PMU_FL_NO_HT_SHARING)) {
3741 for_each_cpu(i, topology_sibling_cpumask(cpu)) {
3742 struct intel_shared_regs *pc;
3744 pc = per_cpu(cpu_hw_events, i).shared_regs;
3745 if (pc && pc->core_id == core_id) {
3746 cpuc->kfree_on_online[0] = cpuc->shared_regs;
3747 cpuc->shared_regs = pc;
3751 cpuc->shared_regs->core_id = core_id;
3752 cpuc->shared_regs->refcnt++;
3755 if (x86_pmu.lbr_sel_map)
3756 cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR];
3758 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) {
3759 for_each_cpu(i, topology_sibling_cpumask(cpu)) {
3760 struct cpu_hw_events *sibling;
3761 struct intel_excl_cntrs *c;
3763 sibling = &per_cpu(cpu_hw_events, i);
3764 c = sibling->excl_cntrs;
3765 if (c && c->core_id == core_id) {
3766 cpuc->kfree_on_online[1] = cpuc->excl_cntrs;
3767 cpuc->excl_cntrs = c;
3768 if (!sibling->excl_thread_id)
3769 cpuc->excl_thread_id = 1;
3773 cpuc->excl_cntrs->core_id = core_id;
3774 cpuc->excl_cntrs->refcnt++;
3778 static void free_excl_cntrs(struct cpu_hw_events *cpuc)
3780 struct intel_excl_cntrs *c;
3782 c = cpuc->excl_cntrs;
3784 if (c->core_id == -1 || --c->refcnt == 0)
3786 cpuc->excl_cntrs = NULL;
3789 kfree(cpuc->constraint_list);
3790 cpuc->constraint_list = NULL;
3793 static void intel_pmu_cpu_dying(int cpu)
3795 fini_debug_store_on_cpu(cpu);
3797 if (x86_pmu.counter_freezing)
3798 disable_counter_freeze();
3801 void intel_cpuc_finish(struct cpu_hw_events *cpuc)
3803 struct intel_shared_regs *pc;
3805 pc = cpuc->shared_regs;
3807 if (pc->core_id == -1 || --pc->refcnt == 0)
3809 cpuc->shared_regs = NULL;
3812 free_excl_cntrs(cpuc);
3815 static void intel_pmu_cpu_dead(int cpu)
3817 intel_cpuc_finish(&per_cpu(cpu_hw_events, cpu));
3820 static void intel_pmu_sched_task(struct perf_event_context *ctx,
3823 intel_pmu_pebs_sched_task(ctx, sched_in);
3824 intel_pmu_lbr_sched_task(ctx, sched_in);
3827 static int intel_pmu_check_period(struct perf_event *event, u64 value)
3829 return intel_pmu_has_bts_period(event, value) ? -EINVAL : 0;
3832 static int intel_pmu_aux_output_match(struct perf_event *event)
3834 if (!x86_pmu.intel_cap.pebs_output_pt_available)
3837 return is_intel_pt_event(event);
3840 PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63");
3842 PMU_FORMAT_ATTR(ldlat, "config1:0-15");
3844 PMU_FORMAT_ATTR(frontend, "config1:0-23");
3846 static struct attribute *intel_arch3_formats_attr[] = {
3847 &format_attr_event.attr,
3848 &format_attr_umask.attr,
3849 &format_attr_edge.attr,
3850 &format_attr_pc.attr,
3851 &format_attr_any.attr,
3852 &format_attr_inv.attr,
3853 &format_attr_cmask.attr,
3857 static struct attribute *hsw_format_attr[] = {
3858 &format_attr_in_tx.attr,
3859 &format_attr_in_tx_cp.attr,
3860 &format_attr_offcore_rsp.attr,
3861 &format_attr_ldlat.attr,
3865 static struct attribute *nhm_format_attr[] = {
3866 &format_attr_offcore_rsp.attr,
3867 &format_attr_ldlat.attr,
3871 static struct attribute *slm_format_attr[] = {
3872 &format_attr_offcore_rsp.attr,
3876 static struct attribute *skl_format_attr[] = {
3877 &format_attr_frontend.attr,
3881 static __initconst const struct x86_pmu core_pmu = {
3883 .handle_irq = x86_pmu_handle_irq,
3884 .disable_all = x86_pmu_disable_all,
3885 .enable_all = core_pmu_enable_all,
3886 .enable = core_pmu_enable_event,
3887 .disable = x86_pmu_disable_event,
3888 .hw_config = core_pmu_hw_config,
3889 .schedule_events = x86_schedule_events,
3890 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
3891 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
3892 .event_map = intel_pmu_event_map,
3893 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
3895 .large_pebs_flags = LARGE_PEBS_FLAGS,
3898 * Intel PMCs cannot be accessed sanely above 32-bit width,
3899 * so we install an artificial 1<<31 period regardless of
3900 * the generic event period:
3902 .max_period = (1ULL<<31) - 1,
3903 .get_event_constraints = intel_get_event_constraints,
3904 .put_event_constraints = intel_put_event_constraints,
3905 .event_constraints = intel_core_event_constraints,
3906 .guest_get_msrs = core_guest_get_msrs,
3907 .format_attrs = intel_arch_formats_attr,
3908 .events_sysfs_show = intel_event_sysfs_show,
3911 * Virtual (or funny metal) CPU can define x86_pmu.extra_regs
3912 * together with PMU version 1 and thus be using core_pmu with
3913 * shared_regs. We need following callbacks here to allocate
3916 .cpu_prepare = intel_pmu_cpu_prepare,
3917 .cpu_starting = intel_pmu_cpu_starting,
3918 .cpu_dying = intel_pmu_cpu_dying,
3919 .cpu_dead = intel_pmu_cpu_dead,
3921 .check_period = intel_pmu_check_period,
3924 static __initconst const struct x86_pmu intel_pmu = {
3926 .handle_irq = intel_pmu_handle_irq,
3927 .disable_all = intel_pmu_disable_all,
3928 .enable_all = intel_pmu_enable_all,
3929 .enable = intel_pmu_enable_event,
3930 .disable = intel_pmu_disable_event,
3931 .add = intel_pmu_add_event,
3932 .del = intel_pmu_del_event,
3933 .read = intel_pmu_read_event,
3934 .hw_config = intel_pmu_hw_config,
3935 .schedule_events = x86_schedule_events,
3936 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
3937 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
3938 .event_map = intel_pmu_event_map,
3939 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
3941 .large_pebs_flags = LARGE_PEBS_FLAGS,
3943 * Intel PMCs cannot be accessed sanely above 32 bit width,
3944 * so we install an artificial 1<<31 period regardless of
3945 * the generic event period:
3947 .max_period = (1ULL << 31) - 1,
3948 .get_event_constraints = intel_get_event_constraints,
3949 .put_event_constraints = intel_put_event_constraints,
3950 .pebs_aliases = intel_pebs_aliases_core2,
3952 .format_attrs = intel_arch3_formats_attr,
3953 .events_sysfs_show = intel_event_sysfs_show,
3955 .cpu_prepare = intel_pmu_cpu_prepare,
3956 .cpu_starting = intel_pmu_cpu_starting,
3957 .cpu_dying = intel_pmu_cpu_dying,
3958 .cpu_dead = intel_pmu_cpu_dead,
3960 .guest_get_msrs = intel_guest_get_msrs,
3961 .sched_task = intel_pmu_sched_task,
3963 .check_period = intel_pmu_check_period,
3965 .aux_output_match = intel_pmu_aux_output_match,
3968 static __init void intel_clovertown_quirk(void)
3971 * PEBS is unreliable due to:
3973 * AJ67 - PEBS may experience CPL leaks
3974 * AJ68 - PEBS PMI may be delayed by one event
3975 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
3976 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
3978 * AJ67 could be worked around by restricting the OS/USR flags.
3979 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
3981 * AJ106 could possibly be worked around by not allowing LBR
3982 * usage from PEBS, including the fixup.
3983 * AJ68 could possibly be worked around by always programming
3984 * a pebs_event_reset[0] value and coping with the lost events.
3986 * But taken together it might just make sense to not enable PEBS on
3989 pr_warn("PEBS disabled due to CPU errata\n");
3991 x86_pmu.pebs_constraints = NULL;
3994 static const struct x86_cpu_desc isolation_ucodes[] = {
3995 INTEL_CPU_DESC(INTEL_FAM6_HASWELL, 3, 0x0000001f),
3996 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_L, 1, 0x0000001e),
3997 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_G, 1, 0x00000015),
3998 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X, 2, 0x00000037),
3999 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X, 4, 0x0000000a),
4000 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL, 4, 0x00000023),
4001 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_G, 1, 0x00000014),
4002 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 2, 0x00000010),
4003 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 3, 0x07000009),
4004 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 4, 0x0f000009),
4005 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 5, 0x0e000002),
4006 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_X, 1, 0x0b000014),
4007 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 3, 0x00000021),
4008 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 4, 0x00000000),
4009 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 5, 0x00000000),
4010 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 6, 0x00000000),
4011 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 7, 0x00000000),
4012 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 11, 0x00000000),
4013 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_L, 3, 0x0000007c),
4014 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE, 3, 0x0000007c),
4015 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 9, 0x0000004e),
4016 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 9, 0x0000004e),
4017 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 10, 0x0000004e),
4018 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 11, 0x0000004e),
4019 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 12, 0x0000004e),
4020 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 10, 0x0000004e),
4021 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 11, 0x0000004e),
4022 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 12, 0x0000004e),
4023 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 13, 0x0000004e),
4027 static void intel_check_pebs_isolation(void)
4029 x86_pmu.pebs_no_isolation = !x86_cpu_has_min_microcode_rev(isolation_ucodes);
4032 static __init void intel_pebs_isolation_quirk(void)
4034 WARN_ON_ONCE(x86_pmu.check_microcode);
4035 x86_pmu.check_microcode = intel_check_pebs_isolation;
4036 intel_check_pebs_isolation();
4039 static const struct x86_cpu_desc pebs_ucodes[] = {
4040 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE, 7, 0x00000028),
4041 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X, 6, 0x00000618),
4042 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X, 7, 0x0000070c),
4046 static bool intel_snb_pebs_broken(void)
4048 return !x86_cpu_has_min_microcode_rev(pebs_ucodes);
4051 static void intel_snb_check_microcode(void)
4053 if (intel_snb_pebs_broken() == x86_pmu.pebs_broken)
4057 * Serialized by the microcode lock..
4059 if (x86_pmu.pebs_broken) {
4060 pr_info("PEBS enabled due to microcode update\n");
4061 x86_pmu.pebs_broken = 0;
4063 pr_info("PEBS disabled due to CPU errata/*(DEBLOBBED)*/\n");
4064 x86_pmu.pebs_broken = 1;
4068 static bool is_lbr_from(unsigned long msr)
4070 unsigned long lbr_from_nr = x86_pmu.lbr_from + x86_pmu.lbr_nr;
4072 return x86_pmu.lbr_from <= msr && msr < lbr_from_nr;
4076 * Under certain circumstances, access certain MSR may cause #GP.
4077 * The function tests if the input MSR can be safely accessed.
4079 static bool check_msr(unsigned long msr, u64 mask)
4081 u64 val_old, val_new, val_tmp;
4084 * Disable the check for real HW, so we don't
4085 * mess with potentionaly enabled registers:
4087 if (!boot_cpu_has(X86_FEATURE_HYPERVISOR))
4091 * Read the current value, change it and read it back to see if it
4092 * matches, this is needed to detect certain hardware emulators
4093 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
4095 if (rdmsrl_safe(msr, &val_old))
4099 * Only change the bits which can be updated by wrmsrl.
4101 val_tmp = val_old ^ mask;
4103 if (is_lbr_from(msr))
4104 val_tmp = lbr_from_signext_quirk_wr(val_tmp);
4106 if (wrmsrl_safe(msr, val_tmp) ||
4107 rdmsrl_safe(msr, &val_new))
4111 * Quirk only affects validation in wrmsr(), so wrmsrl()'s value
4112 * should equal rdmsrl()'s even with the quirk.
4114 if (val_new != val_tmp)
4117 if (is_lbr_from(msr))
4118 val_old = lbr_from_signext_quirk_wr(val_old);
4120 /* Here it's sure that the MSR can be safely accessed.
4121 * Restore the old value and return.
4123 wrmsrl(msr, val_old);
4128 static __init void intel_sandybridge_quirk(void)
4130 x86_pmu.check_microcode = intel_snb_check_microcode;
4132 intel_snb_check_microcode();
4136 static const struct { int id; char *name; } intel_arch_events_map[] __initconst = {
4137 { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" },
4138 { PERF_COUNT_HW_INSTRUCTIONS, "instructions" },
4139 { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" },
4140 { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" },
4141 { PERF_COUNT_HW_CACHE_MISSES, "cache misses" },
4142 { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" },
4143 { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" },
4146 static __init void intel_arch_events_quirk(void)
4150 /* disable event that reported as not presend by cpuid */
4151 for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) {
4152 intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0;
4153 pr_warn("CPUID marked event: \'%s\' unavailable\n",
4154 intel_arch_events_map[bit].name);
4158 static __init void intel_nehalem_quirk(void)
4160 union cpuid10_ebx ebx;
4162 ebx.full = x86_pmu.events_maskl;
4163 if (ebx.split.no_branch_misses_retired) {
4165 * Erratum AAJ80 detected, we work it around by using
4166 * the BR_MISP_EXEC.ANY event. This will over-count
4167 * branch-misses, but it's still much better than the
4168 * architectural event which is often completely bogus:
4170 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89;
4171 ebx.split.no_branch_misses_retired = 0;
4172 x86_pmu.events_maskl = ebx.full;
4173 pr_info("CPU erratum AAJ80 worked around\n");
4177 static const struct x86_cpu_desc counter_freezing_ucodes[] = {
4178 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT, 2, 0x0000000e),
4179 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT, 9, 0x0000002e),
4180 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT, 10, 0x00000008),
4181 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT_D, 1, 0x00000028),
4182 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT_PLUS, 1, 0x00000028),
4183 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT_PLUS, 8, 0x00000006),
4187 static bool intel_counter_freezing_broken(void)
4189 return !x86_cpu_has_min_microcode_rev(counter_freezing_ucodes);
4192 static __init void intel_counter_freezing_quirk(void)
4194 /* Check if it's already disabled */
4195 if (disable_counter_freezing)
4199 * If the system starts with the wrong ucode, leave the
4200 * counter-freezing feature permanently disabled.
4202 if (intel_counter_freezing_broken()) {
4203 pr_info("PMU counter freezing disabled due to CPU errata,"
4204 "please /*(DEBLOBBED)*/\n");
4205 x86_pmu.counter_freezing = false;
4206 x86_pmu.handle_irq = intel_pmu_handle_irq;
4211 * enable software workaround for errata:
4216 * Only needed when HT is enabled. However detecting
4217 * if HT is enabled is difficult (model specific). So instead,
4218 * we enable the workaround in the early boot, and verify if
4219 * it is needed in a later initcall phase once we have valid
4220 * topology information to check if HT is actually enabled
4222 static __init void intel_ht_bug(void)
4224 x86_pmu.flags |= PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED;
4226 x86_pmu.start_scheduling = intel_start_scheduling;
4227 x86_pmu.commit_scheduling = intel_commit_scheduling;
4228 x86_pmu.stop_scheduling = intel_stop_scheduling;
4231 EVENT_ATTR_STR(mem-loads, mem_ld_hsw, "event=0xcd,umask=0x1,ldlat=3");
4232 EVENT_ATTR_STR(mem-stores, mem_st_hsw, "event=0xd0,umask=0x82")
4234 /* Haswell special events */
4235 EVENT_ATTR_STR(tx-start, tx_start, "event=0xc9,umask=0x1");
4236 EVENT_ATTR_STR(tx-commit, tx_commit, "event=0xc9,umask=0x2");
4237 EVENT_ATTR_STR(tx-abort, tx_abort, "event=0xc9,umask=0x4");
4238 EVENT_ATTR_STR(tx-capacity, tx_capacity, "event=0x54,umask=0x2");
4239 EVENT_ATTR_STR(tx-conflict, tx_conflict, "event=0x54,umask=0x1");
4240 EVENT_ATTR_STR(el-start, el_start, "event=0xc8,umask=0x1");
4241 EVENT_ATTR_STR(el-commit, el_commit, "event=0xc8,umask=0x2");
4242 EVENT_ATTR_STR(el-abort, el_abort, "event=0xc8,umask=0x4");
4243 EVENT_ATTR_STR(el-capacity, el_capacity, "event=0x54,umask=0x2");
4244 EVENT_ATTR_STR(el-conflict, el_conflict, "event=0x54,umask=0x1");
4245 EVENT_ATTR_STR(cycles-t, cycles_t, "event=0x3c,in_tx=1");
4246 EVENT_ATTR_STR(cycles-ct, cycles_ct, "event=0x3c,in_tx=1,in_tx_cp=1");
4248 static struct attribute *hsw_events_attrs[] = {
4249 EVENT_PTR(td_slots_issued),
4250 EVENT_PTR(td_slots_retired),
4251 EVENT_PTR(td_fetch_bubbles),
4252 EVENT_PTR(td_total_slots),
4253 EVENT_PTR(td_total_slots_scale),
4254 EVENT_PTR(td_recovery_bubbles),
4255 EVENT_PTR(td_recovery_bubbles_scale),
4259 static struct attribute *hsw_mem_events_attrs[] = {
4260 EVENT_PTR(mem_ld_hsw),
4261 EVENT_PTR(mem_st_hsw),
4265 static struct attribute *hsw_tsx_events_attrs[] = {
4266 EVENT_PTR(tx_start),
4267 EVENT_PTR(tx_commit),
4268 EVENT_PTR(tx_abort),
4269 EVENT_PTR(tx_capacity),
4270 EVENT_PTR(tx_conflict),
4271 EVENT_PTR(el_start),
4272 EVENT_PTR(el_commit),
4273 EVENT_PTR(el_abort),
4274 EVENT_PTR(el_capacity),
4275 EVENT_PTR(el_conflict),
4276 EVENT_PTR(cycles_t),
4277 EVENT_PTR(cycles_ct),
4281 EVENT_ATTR_STR(tx-capacity-read, tx_capacity_read, "event=0x54,umask=0x80");
4282 EVENT_ATTR_STR(tx-capacity-write, tx_capacity_write, "event=0x54,umask=0x2");
4283 EVENT_ATTR_STR(el-capacity-read, el_capacity_read, "event=0x54,umask=0x80");
4284 EVENT_ATTR_STR(el-capacity-write, el_capacity_write, "event=0x54,umask=0x2");
4286 static struct attribute *icl_events_attrs[] = {
4287 EVENT_PTR(mem_ld_hsw),
4288 EVENT_PTR(mem_st_hsw),
4292 static struct attribute *icl_tsx_events_attrs[] = {
4293 EVENT_PTR(tx_start),
4294 EVENT_PTR(tx_abort),
4295 EVENT_PTR(tx_commit),
4296 EVENT_PTR(tx_capacity_read),
4297 EVENT_PTR(tx_capacity_write),
4298 EVENT_PTR(tx_conflict),
4299 EVENT_PTR(el_start),
4300 EVENT_PTR(el_abort),
4301 EVENT_PTR(el_commit),
4302 EVENT_PTR(el_capacity_read),
4303 EVENT_PTR(el_capacity_write),
4304 EVENT_PTR(el_conflict),
4305 EVENT_PTR(cycles_t),
4306 EVENT_PTR(cycles_ct),
4310 static ssize_t freeze_on_smi_show(struct device *cdev,
4311 struct device_attribute *attr,
4314 return sprintf(buf, "%lu\n", x86_pmu.attr_freeze_on_smi);
4317 static DEFINE_MUTEX(freeze_on_smi_mutex);
4319 static ssize_t freeze_on_smi_store(struct device *cdev,
4320 struct device_attribute *attr,
4321 const char *buf, size_t count)
4326 ret = kstrtoul(buf, 0, &val);
4333 mutex_lock(&freeze_on_smi_mutex);
4335 if (x86_pmu.attr_freeze_on_smi == val)
4338 x86_pmu.attr_freeze_on_smi = val;
4341 on_each_cpu(flip_smm_bit, &val, 1);
4344 mutex_unlock(&freeze_on_smi_mutex);
4349 static void update_tfa_sched(void *ignored)
4351 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
4354 * check if PMC3 is used
4355 * and if so force schedule out for all event types all contexts
4357 if (test_bit(3, cpuc->active_mask))
4358 perf_pmu_resched(x86_get_pmu());
4361 static ssize_t show_sysctl_tfa(struct device *cdev,
4362 struct device_attribute *attr,
4365 return snprintf(buf, 40, "%d\n", allow_tsx_force_abort);
4368 static ssize_t set_sysctl_tfa(struct device *cdev,
4369 struct device_attribute *attr,
4370 const char *buf, size_t count)
4375 ret = kstrtobool(buf, &val);
4380 if (val == allow_tsx_force_abort)
4383 allow_tsx_force_abort = val;
4386 on_each_cpu(update_tfa_sched, NULL, 1);
4393 static DEVICE_ATTR_RW(freeze_on_smi);
4395 static ssize_t branches_show(struct device *cdev,
4396 struct device_attribute *attr,
4399 return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu.lbr_nr);
4402 static DEVICE_ATTR_RO(branches);
4404 static struct attribute *lbr_attrs[] = {
4405 &dev_attr_branches.attr,
4409 static char pmu_name_str[30];
4411 static ssize_t pmu_name_show(struct device *cdev,
4412 struct device_attribute *attr,
4415 return snprintf(buf, PAGE_SIZE, "%s\n", pmu_name_str);
4418 static DEVICE_ATTR_RO(pmu_name);
4420 static struct attribute *intel_pmu_caps_attrs[] = {
4421 &dev_attr_pmu_name.attr,
4425 static DEVICE_ATTR(allow_tsx_force_abort, 0644,
4429 static struct attribute *intel_pmu_attrs[] = {
4430 &dev_attr_freeze_on_smi.attr,
4431 &dev_attr_allow_tsx_force_abort.attr,
4436 tsx_is_visible(struct kobject *kobj, struct attribute *attr, int i)
4438 return boot_cpu_has(X86_FEATURE_RTM) ? attr->mode : 0;
4442 pebs_is_visible(struct kobject *kobj, struct attribute *attr, int i)
4444 return x86_pmu.pebs ? attr->mode : 0;
4448 lbr_is_visible(struct kobject *kobj, struct attribute *attr, int i)
4450 return x86_pmu.lbr_nr ? attr->mode : 0;
4454 exra_is_visible(struct kobject *kobj, struct attribute *attr, int i)
4456 return x86_pmu.version >= 2 ? attr->mode : 0;
4460 default_is_visible(struct kobject *kobj, struct attribute *attr, int i)
4462 if (attr == &dev_attr_allow_tsx_force_abort.attr)
4463 return x86_pmu.flags & PMU_FL_TFA ? attr->mode : 0;
4468 static struct attribute_group group_events_td = {
4472 static struct attribute_group group_events_mem = {
4474 .is_visible = pebs_is_visible,
4477 static struct attribute_group group_events_tsx = {
4479 .is_visible = tsx_is_visible,
4482 static struct attribute_group group_caps_gen = {
4484 .attrs = intel_pmu_caps_attrs,
4487 static struct attribute_group group_caps_lbr = {
4490 .is_visible = lbr_is_visible,
4493 static struct attribute_group group_format_extra = {
4495 .is_visible = exra_is_visible,
4498 static struct attribute_group group_format_extra_skl = {
4500 .is_visible = exra_is_visible,
4503 static struct attribute_group group_default = {
4504 .attrs = intel_pmu_attrs,
4505 .is_visible = default_is_visible,
4508 static const struct attribute_group *attr_update[] = {
4514 &group_format_extra,
4515 &group_format_extra_skl,
4520 static struct attribute *empty_attrs;
4522 __init int intel_pmu_init(void)
4524 struct attribute **extra_skl_attr = &empty_attrs;
4525 struct attribute **extra_attr = &empty_attrs;
4526 struct attribute **td_attr = &empty_attrs;
4527 struct attribute **mem_attr = &empty_attrs;
4528 struct attribute **tsx_attr = &empty_attrs;
4529 union cpuid10_edx edx;
4530 union cpuid10_eax eax;
4531 union cpuid10_ebx ebx;
4532 struct event_constraint *c;
4533 unsigned int unused;
4534 struct extra_reg *er;
4539 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
4540 switch (boot_cpu_data.x86) {
4542 return p6_pmu_init();
4544 return knc_pmu_init();
4546 return p4_pmu_init();
4552 * Check whether the Architectural PerfMon supports
4553 * Branch Misses Retired hw_event or not.
4555 cpuid(10, &eax.full, &ebx.full, &unused, &edx.full);
4556 if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT)
4559 version = eax.split.version_id;
4563 x86_pmu = intel_pmu;
4565 x86_pmu.version = version;
4566 x86_pmu.num_counters = eax.split.num_counters;
4567 x86_pmu.cntval_bits = eax.split.bit_width;
4568 x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1;
4570 x86_pmu.events_maskl = ebx.full;
4571 x86_pmu.events_mask_len = eax.split.mask_length;
4573 x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters);
4576 * Quirk: v2 perfmon does not report fixed-purpose events, so
4577 * assume at least 3 events, when not running in a hypervisor:
4580 int assume = 3 * !boot_cpu_has(X86_FEATURE_HYPERVISOR);
4582 x86_pmu.num_counters_fixed =
4583 max((int)edx.split.num_counters_fixed, assume);
4587 x86_pmu.counter_freezing = !disable_counter_freezing;
4589 if (boot_cpu_has(X86_FEATURE_PDCM)) {
4592 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
4593 x86_pmu.intel_cap.capabilities = capabilities;
4598 x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */
4601 * Install the hw-cache-events table:
4603 switch (boot_cpu_data.x86_model) {
4604 case INTEL_FAM6_CORE_YONAH:
4605 pr_cont("Core events, ");
4609 case INTEL_FAM6_CORE2_MEROM:
4610 x86_add_quirk(intel_clovertown_quirk);
4613 case INTEL_FAM6_CORE2_MEROM_L:
4614 case INTEL_FAM6_CORE2_PENRYN:
4615 case INTEL_FAM6_CORE2_DUNNINGTON:
4616 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
4617 sizeof(hw_cache_event_ids));
4619 intel_pmu_lbr_init_core();
4621 x86_pmu.event_constraints = intel_core2_event_constraints;
4622 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints;
4623 pr_cont("Core2 events, ");
4627 case INTEL_FAM6_NEHALEM:
4628 case INTEL_FAM6_NEHALEM_EP:
4629 case INTEL_FAM6_NEHALEM_EX:
4630 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
4631 sizeof(hw_cache_event_ids));
4632 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
4633 sizeof(hw_cache_extra_regs));
4635 intel_pmu_lbr_init_nhm();
4637 x86_pmu.event_constraints = intel_nehalem_event_constraints;
4638 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints;
4639 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
4640 x86_pmu.extra_regs = intel_nehalem_extra_regs;
4641 x86_pmu.limit_period = nhm_limit_period;
4643 mem_attr = nhm_mem_events_attrs;
4645 /* UOPS_ISSUED.STALLED_CYCLES */
4646 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
4647 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
4648 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
4649 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
4650 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
4652 intel_pmu_pebs_data_source_nhm();
4653 x86_add_quirk(intel_nehalem_quirk);
4654 x86_pmu.pebs_no_tlb = 1;
4655 extra_attr = nhm_format_attr;
4657 pr_cont("Nehalem events, ");
4661 case INTEL_FAM6_ATOM_BONNELL:
4662 case INTEL_FAM6_ATOM_BONNELL_MID:
4663 case INTEL_FAM6_ATOM_SALTWELL:
4664 case INTEL_FAM6_ATOM_SALTWELL_MID:
4665 case INTEL_FAM6_ATOM_SALTWELL_TABLET:
4666 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
4667 sizeof(hw_cache_event_ids));
4669 intel_pmu_lbr_init_atom();
4671 x86_pmu.event_constraints = intel_gen_event_constraints;
4672 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints;
4673 x86_pmu.pebs_aliases = intel_pebs_aliases_core2;
4674 pr_cont("Atom events, ");
4678 case INTEL_FAM6_ATOM_SILVERMONT:
4679 case INTEL_FAM6_ATOM_SILVERMONT_D:
4680 case INTEL_FAM6_ATOM_SILVERMONT_MID:
4681 case INTEL_FAM6_ATOM_AIRMONT:
4682 case INTEL_FAM6_ATOM_AIRMONT_MID:
4683 memcpy(hw_cache_event_ids, slm_hw_cache_event_ids,
4684 sizeof(hw_cache_event_ids));
4685 memcpy(hw_cache_extra_regs, slm_hw_cache_extra_regs,
4686 sizeof(hw_cache_extra_regs));
4688 intel_pmu_lbr_init_slm();
4690 x86_pmu.event_constraints = intel_slm_event_constraints;
4691 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints;
4692 x86_pmu.extra_regs = intel_slm_extra_regs;
4693 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4694 td_attr = slm_events_attrs;
4695 extra_attr = slm_format_attr;
4696 pr_cont("Silvermont events, ");
4697 name = "silvermont";
4700 case INTEL_FAM6_ATOM_GOLDMONT:
4701 case INTEL_FAM6_ATOM_GOLDMONT_D:
4702 x86_add_quirk(intel_counter_freezing_quirk);
4703 memcpy(hw_cache_event_ids, glm_hw_cache_event_ids,
4704 sizeof(hw_cache_event_ids));
4705 memcpy(hw_cache_extra_regs, glm_hw_cache_extra_regs,
4706 sizeof(hw_cache_extra_regs));
4708 intel_pmu_lbr_init_skl();
4710 x86_pmu.event_constraints = intel_slm_event_constraints;
4711 x86_pmu.pebs_constraints = intel_glm_pebs_event_constraints;
4712 x86_pmu.extra_regs = intel_glm_extra_regs;
4714 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS
4715 * for precise cycles.
4716 * :pp is identical to :ppp
4718 x86_pmu.pebs_aliases = NULL;
4719 x86_pmu.pebs_prec_dist = true;
4720 x86_pmu.lbr_pt_coexist = true;
4721 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4722 td_attr = glm_events_attrs;
4723 extra_attr = slm_format_attr;
4724 pr_cont("Goldmont events, ");
4728 case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
4729 x86_add_quirk(intel_counter_freezing_quirk);
4730 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids,
4731 sizeof(hw_cache_event_ids));
4732 memcpy(hw_cache_extra_regs, glp_hw_cache_extra_regs,
4733 sizeof(hw_cache_extra_regs));
4735 intel_pmu_lbr_init_skl();
4737 x86_pmu.event_constraints = intel_slm_event_constraints;
4738 x86_pmu.extra_regs = intel_glm_extra_regs;
4740 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS
4741 * for precise cycles.
4743 x86_pmu.pebs_aliases = NULL;
4744 x86_pmu.pebs_prec_dist = true;
4745 x86_pmu.lbr_pt_coexist = true;
4746 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4747 x86_pmu.flags |= PMU_FL_PEBS_ALL;
4748 x86_pmu.get_event_constraints = glp_get_event_constraints;
4749 td_attr = glm_events_attrs;
4750 /* Goldmont Plus has 4-wide pipeline */
4751 event_attr_td_total_slots_scale_glm.event_str = "4";
4752 extra_attr = slm_format_attr;
4753 pr_cont("Goldmont plus events, ");
4754 name = "goldmont_plus";
4757 case INTEL_FAM6_ATOM_TREMONT_D:
4758 case INTEL_FAM6_ATOM_TREMONT:
4759 x86_pmu.late_ack = true;
4760 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids,
4761 sizeof(hw_cache_event_ids));
4762 memcpy(hw_cache_extra_regs, tnt_hw_cache_extra_regs,
4763 sizeof(hw_cache_extra_regs));
4764 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1;
4766 intel_pmu_lbr_init_skl();
4768 x86_pmu.event_constraints = intel_slm_event_constraints;
4769 x86_pmu.extra_regs = intel_tnt_extra_regs;
4771 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS
4772 * for precise cycles.
4774 x86_pmu.pebs_aliases = NULL;
4775 x86_pmu.pebs_prec_dist = true;
4776 x86_pmu.lbr_pt_coexist = true;
4777 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4778 x86_pmu.get_event_constraints = tnt_get_event_constraints;
4779 extra_attr = slm_format_attr;
4780 pr_cont("Tremont events, ");
4784 case INTEL_FAM6_WESTMERE:
4785 case INTEL_FAM6_WESTMERE_EP:
4786 case INTEL_FAM6_WESTMERE_EX:
4787 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
4788 sizeof(hw_cache_event_ids));
4789 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
4790 sizeof(hw_cache_extra_regs));
4792 intel_pmu_lbr_init_nhm();
4794 x86_pmu.event_constraints = intel_westmere_event_constraints;
4795 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
4796 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints;
4797 x86_pmu.extra_regs = intel_westmere_extra_regs;
4798 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4800 mem_attr = nhm_mem_events_attrs;
4802 /* UOPS_ISSUED.STALLED_CYCLES */
4803 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
4804 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
4805 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
4806 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
4807 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
4809 intel_pmu_pebs_data_source_nhm();
4810 extra_attr = nhm_format_attr;
4811 pr_cont("Westmere events, ");
4815 case INTEL_FAM6_SANDYBRIDGE:
4816 case INTEL_FAM6_SANDYBRIDGE_X:
4817 x86_add_quirk(intel_sandybridge_quirk);
4818 x86_add_quirk(intel_ht_bug);
4819 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
4820 sizeof(hw_cache_event_ids));
4821 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
4822 sizeof(hw_cache_extra_regs));
4824 intel_pmu_lbr_init_snb();
4826 x86_pmu.event_constraints = intel_snb_event_constraints;
4827 x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints;
4828 x86_pmu.pebs_aliases = intel_pebs_aliases_snb;
4829 if (boot_cpu_data.x86_model == INTEL_FAM6_SANDYBRIDGE_X)
4830 x86_pmu.extra_regs = intel_snbep_extra_regs;
4832 x86_pmu.extra_regs = intel_snb_extra_regs;
4835 /* all extra regs are per-cpu when HT is on */
4836 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4837 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
4839 td_attr = snb_events_attrs;
4840 mem_attr = snb_mem_events_attrs;
4842 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
4843 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
4844 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
4845 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
4846 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
4847 X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1);
4849 extra_attr = nhm_format_attr;
4851 pr_cont("SandyBridge events, ");
4852 name = "sandybridge";
4855 case INTEL_FAM6_IVYBRIDGE:
4856 case INTEL_FAM6_IVYBRIDGE_X:
4857 x86_add_quirk(intel_ht_bug);
4858 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
4859 sizeof(hw_cache_event_ids));
4860 /* dTLB-load-misses on IVB is different than SNB */
4861 hw_cache_event_ids[C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = 0x8108; /* DTLB_LOAD_MISSES.DEMAND_LD_MISS_CAUSES_A_WALK */
4863 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
4864 sizeof(hw_cache_extra_regs));
4866 intel_pmu_lbr_init_snb();
4868 x86_pmu.event_constraints = intel_ivb_event_constraints;
4869 x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints;
4870 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
4871 x86_pmu.pebs_prec_dist = true;
4872 if (boot_cpu_data.x86_model == INTEL_FAM6_IVYBRIDGE_X)
4873 x86_pmu.extra_regs = intel_snbep_extra_regs;
4875 x86_pmu.extra_regs = intel_snb_extra_regs;
4876 /* all extra regs are per-cpu when HT is on */
4877 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4878 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
4880 td_attr = snb_events_attrs;
4881 mem_attr = snb_mem_events_attrs;
4883 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
4884 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
4885 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
4887 extra_attr = nhm_format_attr;
4889 pr_cont("IvyBridge events, ");
4894 case INTEL_FAM6_HASWELL:
4895 case INTEL_FAM6_HASWELL_X:
4896 case INTEL_FAM6_HASWELL_L:
4897 case INTEL_FAM6_HASWELL_G:
4898 x86_add_quirk(intel_ht_bug);
4899 x86_add_quirk(intel_pebs_isolation_quirk);
4900 x86_pmu.late_ack = true;
4901 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids));
4902 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
4904 intel_pmu_lbr_init_hsw();
4906 x86_pmu.event_constraints = intel_hsw_event_constraints;
4907 x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints;
4908 x86_pmu.extra_regs = intel_snbep_extra_regs;
4909 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
4910 x86_pmu.pebs_prec_dist = true;
4911 /* all extra regs are per-cpu when HT is on */
4912 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4913 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
4915 x86_pmu.hw_config = hsw_hw_config;
4916 x86_pmu.get_event_constraints = hsw_get_event_constraints;
4917 x86_pmu.lbr_double_abort = true;
4918 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
4919 hsw_format_attr : nhm_format_attr;
4920 td_attr = hsw_events_attrs;
4921 mem_attr = hsw_mem_events_attrs;
4922 tsx_attr = hsw_tsx_events_attrs;
4923 pr_cont("Haswell events, ");
4927 case INTEL_FAM6_BROADWELL:
4928 case INTEL_FAM6_BROADWELL_D:
4929 case INTEL_FAM6_BROADWELL_G:
4930 case INTEL_FAM6_BROADWELL_X:
4931 x86_add_quirk(intel_pebs_isolation_quirk);
4932 x86_pmu.late_ack = true;
4933 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids));
4934 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
4936 /* L3_MISS_LOCAL_DRAM is BIT(26) in Broadwell */
4937 hw_cache_extra_regs[C(LL)][C(OP_READ)][C(RESULT_MISS)] = HSW_DEMAND_READ |
4938 BDW_L3_MISS|HSW_SNOOP_DRAM;
4939 hw_cache_extra_regs[C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = HSW_DEMAND_WRITE|BDW_L3_MISS|
4941 hw_cache_extra_regs[C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = HSW_DEMAND_READ|
4942 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM;
4943 hw_cache_extra_regs[C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = HSW_DEMAND_WRITE|
4944 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM;
4946 intel_pmu_lbr_init_hsw();
4948 x86_pmu.event_constraints = intel_bdw_event_constraints;
4949 x86_pmu.pebs_constraints = intel_bdw_pebs_event_constraints;
4950 x86_pmu.extra_regs = intel_snbep_extra_regs;
4951 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
4952 x86_pmu.pebs_prec_dist = true;
4953 /* all extra regs are per-cpu when HT is on */
4954 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4955 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
4957 x86_pmu.hw_config = hsw_hw_config;
4958 x86_pmu.get_event_constraints = hsw_get_event_constraints;
4959 x86_pmu.limit_period = bdw_limit_period;
4960 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
4961 hsw_format_attr : nhm_format_attr;
4962 td_attr = hsw_events_attrs;
4963 mem_attr = hsw_mem_events_attrs;
4964 tsx_attr = hsw_tsx_events_attrs;
4965 pr_cont("Broadwell events, ");
4969 case INTEL_FAM6_XEON_PHI_KNL:
4970 case INTEL_FAM6_XEON_PHI_KNM:
4971 memcpy(hw_cache_event_ids,
4972 slm_hw_cache_event_ids, sizeof(hw_cache_event_ids));
4973 memcpy(hw_cache_extra_regs,
4974 knl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
4975 intel_pmu_lbr_init_knl();
4977 x86_pmu.event_constraints = intel_slm_event_constraints;
4978 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints;
4979 x86_pmu.extra_regs = intel_knl_extra_regs;
4981 /* all extra regs are per-cpu when HT is on */
4982 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4983 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
4984 extra_attr = slm_format_attr;
4985 pr_cont("Knights Landing/Mill events, ");
4986 name = "knights-landing";
4989 case INTEL_FAM6_SKYLAKE_X:
4992 case INTEL_FAM6_SKYLAKE_L:
4993 case INTEL_FAM6_SKYLAKE:
4994 case INTEL_FAM6_KABYLAKE_L:
4995 case INTEL_FAM6_KABYLAKE:
4996 case INTEL_FAM6_COMETLAKE_L:
4997 case INTEL_FAM6_COMETLAKE:
4998 x86_add_quirk(intel_pebs_isolation_quirk);
4999 x86_pmu.late_ack = true;
5000 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids));
5001 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
5002 intel_pmu_lbr_init_skl();
5004 /* INT_MISC.RECOVERY_CYCLES has umask 1 in Skylake */
5005 event_attr_td_recovery_bubbles.event_str_noht =
5006 "event=0xd,umask=0x1,cmask=1";
5007 event_attr_td_recovery_bubbles.event_str_ht =
5008 "event=0xd,umask=0x1,cmask=1,any=1";
5010 x86_pmu.event_constraints = intel_skl_event_constraints;
5011 x86_pmu.pebs_constraints = intel_skl_pebs_event_constraints;
5012 x86_pmu.extra_regs = intel_skl_extra_regs;
5013 x86_pmu.pebs_aliases = intel_pebs_aliases_skl;
5014 x86_pmu.pebs_prec_dist = true;
5015 /* all extra regs are per-cpu when HT is on */
5016 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5017 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5019 x86_pmu.hw_config = hsw_hw_config;
5020 x86_pmu.get_event_constraints = hsw_get_event_constraints;
5021 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
5022 hsw_format_attr : nhm_format_attr;
5023 extra_skl_attr = skl_format_attr;
5024 td_attr = hsw_events_attrs;
5025 mem_attr = hsw_mem_events_attrs;
5026 tsx_attr = hsw_tsx_events_attrs;
5027 intel_pmu_pebs_data_source_skl(pmem);
5029 if (boot_cpu_has(X86_FEATURE_TSX_FORCE_ABORT)) {
5030 x86_pmu.flags |= PMU_FL_TFA;
5031 x86_pmu.get_event_constraints = tfa_get_event_constraints;
5032 x86_pmu.enable_all = intel_tfa_pmu_enable_all;
5033 x86_pmu.commit_scheduling = intel_tfa_commit_scheduling;
5036 pr_cont("Skylake events, ");
5040 case INTEL_FAM6_ICELAKE_X:
5041 case INTEL_FAM6_ICELAKE_D:
5044 case INTEL_FAM6_ICELAKE_L:
5045 case INTEL_FAM6_ICELAKE:
5046 case INTEL_FAM6_TIGERLAKE_L:
5047 case INTEL_FAM6_TIGERLAKE:
5048 x86_pmu.late_ack = true;
5049 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids));
5050 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
5051 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1;
5052 intel_pmu_lbr_init_skl();
5054 x86_pmu.event_constraints = intel_icl_event_constraints;
5055 x86_pmu.pebs_constraints = intel_icl_pebs_event_constraints;
5056 x86_pmu.extra_regs = intel_icl_extra_regs;
5057 x86_pmu.pebs_aliases = NULL;
5058 x86_pmu.pebs_prec_dist = true;
5059 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
5060 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
5062 x86_pmu.hw_config = hsw_hw_config;
5063 x86_pmu.get_event_constraints = icl_get_event_constraints;
5064 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
5065 hsw_format_attr : nhm_format_attr;
5066 extra_skl_attr = skl_format_attr;
5067 mem_attr = icl_events_attrs;
5068 tsx_attr = icl_tsx_events_attrs;
5069 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04);
5070 x86_pmu.lbr_pt_coexist = true;
5071 intel_pmu_pebs_data_source_skl(pmem);
5072 pr_cont("Icelake events, ");
5077 switch (x86_pmu.version) {
5079 x86_pmu.event_constraints = intel_v1_event_constraints;
5080 pr_cont("generic architected perfmon v1, ");
5081 name = "generic_arch_v1";
5085 * default constraints for v2 and up
5087 x86_pmu.event_constraints = intel_gen_event_constraints;
5088 pr_cont("generic architected perfmon, ");
5089 name = "generic_arch_v2+";
5094 snprintf(pmu_name_str, sizeof(pmu_name_str), "%s", name);
5097 group_events_td.attrs = td_attr;
5098 group_events_mem.attrs = mem_attr;
5099 group_events_tsx.attrs = tsx_attr;
5100 group_format_extra.attrs = extra_attr;
5101 group_format_extra_skl.attrs = extra_skl_attr;
5103 x86_pmu.attr_update = attr_update;
5105 if (x86_pmu.num_counters > INTEL_PMC_MAX_GENERIC) {
5106 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
5107 x86_pmu.num_counters, INTEL_PMC_MAX_GENERIC);
5108 x86_pmu.num_counters = INTEL_PMC_MAX_GENERIC;
5110 x86_pmu.intel_ctrl = (1ULL << x86_pmu.num_counters) - 1;
5112 if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) {
5113 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
5114 x86_pmu.num_counters_fixed, INTEL_PMC_MAX_FIXED);
5115 x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED;
5118 x86_pmu.intel_ctrl |=
5119 ((1LL << x86_pmu.num_counters_fixed)-1) << INTEL_PMC_IDX_FIXED;
5121 if (x86_pmu.event_constraints) {
5123 * event on fixed counter2 (REF_CYCLES) only works on this
5124 * counter, so do not extend mask to generic counters
5126 for_each_event_constraint(c, x86_pmu.event_constraints) {
5127 if (c->cmask == FIXED_EVENT_FLAGS
5128 && c->idxmsk64 != INTEL_PMC_MSK_FIXED_REF_CYCLES) {
5129 c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1;
5132 ~(~0ULL << (INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed));
5133 c->weight = hweight64(c->idxmsk64);
5138 * Access LBR MSR may cause #GP under certain circumstances.
5139 * E.g. KVM doesn't support LBR MSR
5140 * Check all LBT MSR here.
5141 * Disable LBR access if any LBR MSRs can not be accessed.
5143 if (x86_pmu.lbr_nr && !check_msr(x86_pmu.lbr_tos, 0x3UL))
5145 for (i = 0; i < x86_pmu.lbr_nr; i++) {
5146 if (!(check_msr(x86_pmu.lbr_from + i, 0xffffUL) &&
5147 check_msr(x86_pmu.lbr_to + i, 0xffffUL)))
5152 pr_cont("%d-deep LBR, ", x86_pmu.lbr_nr);
5155 * Access extra MSR may cause #GP under certain circumstances.
5156 * E.g. KVM doesn't support offcore event
5157 * Check all extra_regs here.
5159 if (x86_pmu.extra_regs) {
5160 for (er = x86_pmu.extra_regs; er->msr; er++) {
5161 er->extra_msr_access = check_msr(er->msr, 0x11UL);
5162 /* Disable LBR select mapping */
5163 if ((er->idx == EXTRA_REG_LBR) && !er->extra_msr_access)
5164 x86_pmu.lbr_sel_map = NULL;
5168 /* Support full width counters using alternative MSR range */
5169 if (x86_pmu.intel_cap.full_width_write) {
5170 x86_pmu.max_period = x86_pmu.cntval_mask >> 1;
5171 x86_pmu.perfctr = MSR_IA32_PMC0;
5172 pr_cont("full-width counters, ");
5176 * For arch perfmon 4 use counter freezing to avoid
5177 * several MSR accesses in the PMI.
5179 if (x86_pmu.counter_freezing)
5180 x86_pmu.handle_irq = intel_pmu_handle_irq_v4;
5186 * HT bug: phase 2 init
5187 * Called once we have valid topology information to check
5188 * whether or not HT is enabled
5189 * If HT is off, then we disable the workaround
5191 static __init int fixup_ht_bug(void)
5195 * problem not present on this CPU model, nothing to do
5197 if (!(x86_pmu.flags & PMU_FL_EXCL_ENABLED))
5200 if (topology_max_smt_threads() > 1) {
5201 pr_info("PMU erratum BJ122, BV98, HSD29 worked around, HT is on\n");
5207 hardlockup_detector_perf_stop();
5209 x86_pmu.flags &= ~(PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED);
5211 x86_pmu.start_scheduling = NULL;
5212 x86_pmu.commit_scheduling = NULL;
5213 x86_pmu.stop_scheduling = NULL;
5215 hardlockup_detector_perf_restart();
5217 for_each_online_cpu(c)
5218 free_excl_cntrs(&per_cpu(cpu_hw_events, c));
5221 pr_info("PMU erratum BJ122, BV98, HSD29 workaround disabled, HT off\n");
5224 subsys_initcall(fixup_ht_bug)