?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
x86/speculation/mds: Fix documentation typo
[linux/fpc-iii.git] / arch / x86 / events / intel / core.c
blob0b93f5519ddab2dfa594401f06411b2bc921826b
1 /*
2 * Per core/cpu state
3 *
4 * Used to coordinate shared registers between HT threads or
5 * among events on a single PMU.
6 */
7
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10 #include <linux/stddef.h>
11 #include <linux/types.h>
12 #include <linux/init.h>
13 #include <linux/slab.h>
14 #include <linux/export.h>
15 #include <linux/nmi.h>
16
17 #include <asm/cpufeature.h>
18 #include <asm/hardirq.h>
19 #include <asm/intel-family.h>
20 #include <asm/apic.h>
21
22 #include "../perf_event.h"
23
24 /*
25 * Intel PerfMon, used on Core and later.
26 */
27 static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly =
28 {
29 [PERF_COUNT_HW_CPU_CYCLES] = 0x003c,
30 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
31 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e,
32 [PERF_COUNT_HW_CACHE_MISSES] = 0x412e,
33 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4,
34 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5,
35 [PERF_COUNT_HW_BUS_CYCLES] = 0x013c,
36 [PERF_COUNT_HW_REF_CPU_CYCLES] = 0x0300, /* pseudo-encoding */
37 };
38
39 static struct event_constraint intel_core_event_constraints[] __read_mostly =
40 {
41 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
42 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
43 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
44 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
45 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
46 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */
47 EVENT_CONSTRAINT_END
48 };
49
50 static struct event_constraint intel_core2_event_constraints[] __read_mostly =
51 {
52 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
53 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
54 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
55 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
56 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
57 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
58 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
59 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
60 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
61 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
62 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
63 INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */
64 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
65 EVENT_CONSTRAINT_END
66 };
67
68 static struct event_constraint intel_nehalem_event_constraints[] __read_mostly =
69 {
70 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
71 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
72 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
73 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
74 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
75 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
76 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
77 INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */
78 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
79 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
80 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
81 EVENT_CONSTRAINT_END
82 };
83
84 static struct extra_reg intel_nehalem_extra_regs[] __read_mostly =
85 {
86 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
87 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
88 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
89 EVENT_EXTRA_END
90 };
91
92 static struct event_constraint intel_westmere_event_constraints[] __read_mostly =
93 {
94 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
95 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
96 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
97 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
98 INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */
99 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
100 INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */
101 EVENT_CONSTRAINT_END
102 };
103
104 static struct event_constraint intel_snb_event_constraints[] __read_mostly =
105 {
106 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
107 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
108 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
109 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
110 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
111 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
112 INTEL_UEVENT_CONSTRAINT(0x06a3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
113 INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */
114 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
115 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
116 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
117 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
118
119 /*
120 * When HT is off these events can only run on the bottom 4 counters
121 * When HT is on, they are impacted by the HT bug and require EXCL access
122 */
123 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
124 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
125 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
126 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
127
128 EVENT_CONSTRAINT_END
129 };
130
131 static struct event_constraint intel_ivb_event_constraints[] __read_mostly =
132 {
133 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
134 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
135 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
136 INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */
137 INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMTPY */
138 INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */
139 INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_LDM_PENDING */
140 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
141 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
142 INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */
143 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
144 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
145 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
146
147 /*
148 * When HT is off these events can only run on the bottom 4 counters
149 * When HT is on, they are impacted by the HT bug and require EXCL access
150 */
151 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
152 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
153 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
154 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
155
156 EVENT_CONSTRAINT_END
157 };
158
159 static struct extra_reg intel_westmere_extra_regs[] __read_mostly =
160 {
161 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
162 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
163 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1),
164 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
165 EVENT_EXTRA_END
166 };
167
168 static struct event_constraint intel_v1_event_constraints[] __read_mostly =
169 {
170 EVENT_CONSTRAINT_END
171 };
172
173 static struct event_constraint intel_gen_event_constraints[] __read_mostly =
174 {
175 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
176 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
177 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
178 EVENT_CONSTRAINT_END
179 };
180
181 static struct event_constraint intel_slm_event_constraints[] __read_mostly =
182 {
183 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
184 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
185 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */
186 EVENT_CONSTRAINT_END
187 };
188
189 static struct event_constraint intel_skl_event_constraints[] = {
190 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
191 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
192 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
193 INTEL_UEVENT_CONSTRAINT(0x1c0, 0x2), /* INST_RETIRED.PREC_DIST */
194
195 /*
196 * when HT is off, these can only run on the bottom 4 counters
197 */
198 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */
199 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */
200 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */
201 INTEL_EVENT_CONSTRAINT(0xcd, 0xf), /* MEM_TRANS_RETIRED.* */
202 INTEL_EVENT_CONSTRAINT(0xc6, 0xf), /* FRONTEND_RETIRED.* */
203
204 EVENT_CONSTRAINT_END
205 };
206
207 static struct extra_reg intel_knl_extra_regs[] __read_mostly = {
208 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x799ffbb6e7ull, RSP_0),
209 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x399ffbffe7ull, RSP_1),
210 EVENT_EXTRA_END
211 };
212
213 static struct extra_reg intel_snb_extra_regs[] __read_mostly = {
214 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
215 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0),
216 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1),
217 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
218 EVENT_EXTRA_END
219 };
220
221 static struct extra_reg intel_snbep_extra_regs[] __read_mostly = {
222 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
223 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
224 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
225 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
226 EVENT_EXTRA_END
227 };
228
229 static struct extra_reg intel_skl_extra_regs[] __read_mostly = {
230 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
231 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
232 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
233 /*
234 * Note the low 8 bits eventsel code is not a continuous field, containing
235 * some #GPing bits. These are masked out.
236 */
237 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE),
238 EVENT_EXTRA_END
239 };
240
241 EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3");
242 EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3");
243 EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2");
244
245 static struct attribute *nhm_events_attrs[] = {
246 EVENT_PTR(mem_ld_nhm),
247 NULL,
248 };
249
250 /*
251 * topdown events for Intel Core CPUs.
252 *
253 * The events are all in slots, which is a free slot in a 4 wide
254 * pipeline. Some events are already reported in slots, for cycle
255 * events we multiply by the pipeline width (4).
256 *
257 * With Hyper Threading on, topdown metrics are either summed or averaged
258 * between the threads of a core: (count_t0 + count_t1).
259 *
260 * For the average case the metric is always scaled to pipeline width,
261 * so we use factor 2 ((count_t0 + count_t1) / 2 * 4)
262 */
263
264 EVENT_ATTR_STR_HT(topdown-total-slots, td_total_slots,
265 "event=0x3c,umask=0x0", /* cpu_clk_unhalted.thread */
266 "event=0x3c,umask=0x0,any=1"); /* cpu_clk_unhalted.thread_any */
267 EVENT_ATTR_STR_HT(topdown-total-slots.scale, td_total_slots_scale, "4", "2");
268 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued,
269 "event=0xe,umask=0x1"); /* uops_issued.any */
270 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired,
271 "event=0xc2,umask=0x2"); /* uops_retired.retire_slots */
272 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles,
273 "event=0x9c,umask=0x1"); /* idq_uops_not_delivered_core */
274 EVENT_ATTR_STR_HT(topdown-recovery-bubbles, td_recovery_bubbles,
275 "event=0xd,umask=0x3,cmask=1", /* int_misc.recovery_cycles */
276 "event=0xd,umask=0x3,cmask=1,any=1"); /* int_misc.recovery_cycles_any */
277 EVENT_ATTR_STR_HT(topdown-recovery-bubbles.scale, td_recovery_bubbles_scale,
278 "4", "2");
279
280 static struct attribute *snb_events_attrs[] = {
281 EVENT_PTR(mem_ld_snb),
282 EVENT_PTR(mem_st_snb),
283 EVENT_PTR(td_slots_issued),
284 EVENT_PTR(td_slots_retired),
285 EVENT_PTR(td_fetch_bubbles),
286 EVENT_PTR(td_total_slots),
287 EVENT_PTR(td_total_slots_scale),
288 EVENT_PTR(td_recovery_bubbles),
289 EVENT_PTR(td_recovery_bubbles_scale),
290 NULL,
291 };
292
293 static struct event_constraint intel_hsw_event_constraints[] = {
294 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
295 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
296 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
297 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */
298 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
299 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
300 /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
301 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4),
302 /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
303 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4),
304 /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
305 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf),
306
307 /*
308 * When HT is off these events can only run on the bottom 4 counters
309 * When HT is on, they are impacted by the HT bug and require EXCL access
310 */
311 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
312 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
313 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
314 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
315
316 EVENT_CONSTRAINT_END
317 };
318
319 static struct event_constraint intel_bdw_event_constraints[] = {
320 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
321 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
322 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
323 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */
324 INTEL_UBIT_EVENT_CONSTRAINT(0x8a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_MISS */
325 /*
326 * when HT is off, these can only run on the bottom 4 counters
327 */
328 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */
329 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */
330 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */
331 INTEL_EVENT_CONSTRAINT(0xcd, 0xf), /* MEM_TRANS_RETIRED.* */
332 EVENT_CONSTRAINT_END
333 };
334
335 static u64 intel_pmu_event_map(int hw_event)
336 {
337 return intel_perfmon_event_map[hw_event];
338 }
339
340 /*
341 * Notes on the events:
342 * - data reads do not include code reads (comparable to earlier tables)
343 * - data counts include speculative execution (except L1 write, dtlb, bpu)
344 * - remote node access includes remote memory, remote cache, remote mmio.
345 * - prefetches are not included in the counts.
346 * - icache miss does not include decoded icache
347 */
348
349 #define SKL_DEMAND_DATA_RD BIT_ULL(0)
350 #define SKL_DEMAND_RFO BIT_ULL(1)
351 #define SKL_ANY_RESPONSE BIT_ULL(16)
352 #define SKL_SUPPLIER_NONE BIT_ULL(17)
353 #define SKL_L3_MISS_LOCAL_DRAM BIT_ULL(26)
354 #define SKL_L3_MISS_REMOTE_HOP0_DRAM BIT_ULL(27)
355 #define SKL_L3_MISS_REMOTE_HOP1_DRAM BIT_ULL(28)
356 #define SKL_L3_MISS_REMOTE_HOP2P_DRAM BIT_ULL(29)
357 #define SKL_L3_MISS (SKL_L3_MISS_LOCAL_DRAM| \
358 SKL_L3_MISS_REMOTE_HOP0_DRAM| \
359 SKL_L3_MISS_REMOTE_HOP1_DRAM| \
360 SKL_L3_MISS_REMOTE_HOP2P_DRAM)
361 #define SKL_SPL_HIT BIT_ULL(30)
362 #define SKL_SNOOP_NONE BIT_ULL(31)
363 #define SKL_SNOOP_NOT_NEEDED BIT_ULL(32)
364 #define SKL_SNOOP_MISS BIT_ULL(33)
365 #define SKL_SNOOP_HIT_NO_FWD BIT_ULL(34)
366 #define SKL_SNOOP_HIT_WITH_FWD BIT_ULL(35)
367 #define SKL_SNOOP_HITM BIT_ULL(36)
368 #define SKL_SNOOP_NON_DRAM BIT_ULL(37)
369 #define SKL_ANY_SNOOP (SKL_SPL_HIT|SKL_SNOOP_NONE| \
370 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \
371 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \
372 SKL_SNOOP_HITM|SKL_SNOOP_NON_DRAM)
373 #define SKL_DEMAND_READ SKL_DEMAND_DATA_RD
374 #define SKL_SNOOP_DRAM (SKL_SNOOP_NONE| \
375 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \
376 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \
377 SKL_SNOOP_HITM|SKL_SPL_HIT)
378 #define SKL_DEMAND_WRITE SKL_DEMAND_RFO
379 #define SKL_LLC_ACCESS SKL_ANY_RESPONSE
380 #define SKL_L3_MISS_REMOTE (SKL_L3_MISS_REMOTE_HOP0_DRAM| \
381 SKL_L3_MISS_REMOTE_HOP1_DRAM| \
382 SKL_L3_MISS_REMOTE_HOP2P_DRAM)
383
384 static __initconst const u64 skl_hw_cache_event_ids
385 [PERF_COUNT_HW_CACHE_MAX]
386 [PERF_COUNT_HW_CACHE_OP_MAX]
387 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
388 {
389 [ C(L1D ) ] = {
390 [ C(OP_READ) ] = {
391 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */
392 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */
393 },
394 [ C(OP_WRITE) ] = {
395 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */
396 [ C(RESULT_MISS) ] = 0x0,
397 },
398 [ C(OP_PREFETCH) ] = {
399 [ C(RESULT_ACCESS) ] = 0x0,
400 [ C(RESULT_MISS) ] = 0x0,
401 },
402 },
403 [ C(L1I ) ] = {
404 [ C(OP_READ) ] = {
405 [ C(RESULT_ACCESS) ] = 0x0,
406 [ C(RESULT_MISS) ] = 0x283, /* ICACHE_64B.MISS */
407 },
408 [ C(OP_WRITE) ] = {
409 [ C(RESULT_ACCESS) ] = -1,
410 [ C(RESULT_MISS) ] = -1,
411 },
412 [ C(OP_PREFETCH) ] = {
413 [ C(RESULT_ACCESS) ] = 0x0,
414 [ C(RESULT_MISS) ] = 0x0,
415 },
416 },
417 [ C(LL ) ] = {
418 [ C(OP_READ) ] = {
419 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
420 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
421 },
422 [ C(OP_WRITE) ] = {
423 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
424 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
425 },
426 [ C(OP_PREFETCH) ] = {
427 [ C(RESULT_ACCESS) ] = 0x0,
428 [ C(RESULT_MISS) ] = 0x0,
429 },
430 },
431 [ C(DTLB) ] = {
432 [ C(OP_READ) ] = {
433 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */
434 [ C(RESULT_MISS) ] = 0xe08, /* DTLB_LOAD_MISSES.WALK_COMPLETED */
435 },
436 [ C(OP_WRITE) ] = {
437 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */
438 [ C(RESULT_MISS) ] = 0xe49, /* DTLB_STORE_MISSES.WALK_COMPLETED */
439 },
440 [ C(OP_PREFETCH) ] = {
441 [ C(RESULT_ACCESS) ] = 0x0,
442 [ C(RESULT_MISS) ] = 0x0,
443 },
444 },
445 [ C(ITLB) ] = {
446 [ C(OP_READ) ] = {
447 [ C(RESULT_ACCESS) ] = 0x2085, /* ITLB_MISSES.STLB_HIT */
448 [ C(RESULT_MISS) ] = 0xe85, /* ITLB_MISSES.WALK_COMPLETED */
449 },
450 [ C(OP_WRITE) ] = {
451 [ C(RESULT_ACCESS) ] = -1,
452 [ C(RESULT_MISS) ] = -1,
453 },
454 [ C(OP_PREFETCH) ] = {
455 [ C(RESULT_ACCESS) ] = -1,
456 [ C(RESULT_MISS) ] = -1,
457 },
458 },
459 [ C(BPU ) ] = {
460 [ C(OP_READ) ] = {
461 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */
462 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */
463 },
464 [ C(OP_WRITE) ] = {
465 [ C(RESULT_ACCESS) ] = -1,
466 [ C(RESULT_MISS) ] = -1,
467 },
468 [ C(OP_PREFETCH) ] = {
469 [ C(RESULT_ACCESS) ] = -1,
470 [ C(RESULT_MISS) ] = -1,
471 },
472 },
473 [ C(NODE) ] = {
474 [ C(OP_READ) ] = {
475 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
476 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
477 },
478 [ C(OP_WRITE) ] = {
479 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
480 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
481 },
482 [ C(OP_PREFETCH) ] = {
483 [ C(RESULT_ACCESS) ] = 0x0,
484 [ C(RESULT_MISS) ] = 0x0,
485 },
486 },
487 };
488
489 static __initconst const u64 skl_hw_cache_extra_regs
490 [PERF_COUNT_HW_CACHE_MAX]
491 [PERF_COUNT_HW_CACHE_OP_MAX]
492 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
493 {
494 [ C(LL ) ] = {
495 [ C(OP_READ) ] = {
496 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ|
497 SKL_LLC_ACCESS|SKL_ANY_SNOOP,
498 [ C(RESULT_MISS) ] = SKL_DEMAND_READ|
499 SKL_L3_MISS|SKL_ANY_SNOOP|
500 SKL_SUPPLIER_NONE,
501 },
502 [ C(OP_WRITE) ] = {
503 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE|
504 SKL_LLC_ACCESS|SKL_ANY_SNOOP,
505 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE|
506 SKL_L3_MISS|SKL_ANY_SNOOP|
507 SKL_SUPPLIER_NONE,
508 },
509 [ C(OP_PREFETCH) ] = {
510 [ C(RESULT_ACCESS) ] = 0x0,
511 [ C(RESULT_MISS) ] = 0x0,
512 },
513 },
514 [ C(NODE) ] = {
515 [ C(OP_READ) ] = {
516 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ|
517 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM,
518 [ C(RESULT_MISS) ] = SKL_DEMAND_READ|
519 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM,
520 },
521 [ C(OP_WRITE) ] = {
522 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE|
523 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM,
524 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE|
525 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM,
526 },
527 [ C(OP_PREFETCH) ] = {
528 [ C(RESULT_ACCESS) ] = 0x0,
529 [ C(RESULT_MISS) ] = 0x0,
530 },
531 },
532 };
533
534 #define SNB_DMND_DATA_RD (1ULL << 0)
535 #define SNB_DMND_RFO (1ULL << 1)
536 #define SNB_DMND_IFETCH (1ULL << 2)
537 #define SNB_DMND_WB (1ULL << 3)
538 #define SNB_PF_DATA_RD (1ULL << 4)
539 #define SNB_PF_RFO (1ULL << 5)
540 #define SNB_PF_IFETCH (1ULL << 6)
541 #define SNB_LLC_DATA_RD (1ULL << 7)
542 #define SNB_LLC_RFO (1ULL << 8)
543 #define SNB_LLC_IFETCH (1ULL << 9)
544 #define SNB_BUS_LOCKS (1ULL << 10)
545 #define SNB_STRM_ST (1ULL << 11)
546 #define SNB_OTHER (1ULL << 15)
547 #define SNB_RESP_ANY (1ULL << 16)
548 #define SNB_NO_SUPP (1ULL << 17)
549 #define SNB_LLC_HITM (1ULL << 18)
550 #define SNB_LLC_HITE (1ULL << 19)
551 #define SNB_LLC_HITS (1ULL << 20)
552 #define SNB_LLC_HITF (1ULL << 21)
553 #define SNB_LOCAL (1ULL << 22)
554 #define SNB_REMOTE (0xffULL << 23)
555 #define SNB_SNP_NONE (1ULL << 31)
556 #define SNB_SNP_NOT_NEEDED (1ULL << 32)
557 #define SNB_SNP_MISS (1ULL << 33)
558 #define SNB_NO_FWD (1ULL << 34)
559 #define SNB_SNP_FWD (1ULL << 35)
560 #define SNB_HITM (1ULL << 36)
561 #define SNB_NON_DRAM (1ULL << 37)
562
563 #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD)
564 #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO)
565 #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
566
567 #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \
568 SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \
569 SNB_HITM)
570
571 #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY)
572 #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY)
573
574 #define SNB_L3_ACCESS SNB_RESP_ANY
575 #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM)
576
577 static __initconst const u64 snb_hw_cache_extra_regs
578 [PERF_COUNT_HW_CACHE_MAX]
579 [PERF_COUNT_HW_CACHE_OP_MAX]
580 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
581 {
582 [ C(LL ) ] = {
583 [ C(OP_READ) ] = {
584 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS,
585 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_L3_MISS,
586 },
587 [ C(OP_WRITE) ] = {
588 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS,
589 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_L3_MISS,
590 },
591 [ C(OP_PREFETCH) ] = {
592 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS,
593 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_L3_MISS,
594 },
595 },
596 [ C(NODE) ] = {
597 [ C(OP_READ) ] = {
598 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY,
599 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_DRAM_REMOTE,
600 },
601 [ C(OP_WRITE) ] = {
602 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY,
603 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE,
604 },
605 [ C(OP_PREFETCH) ] = {
606 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY,
607 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE,
608 },
609 },
610 };
611
612 static __initconst const u64 snb_hw_cache_event_ids
613 [PERF_COUNT_HW_CACHE_MAX]
614 [PERF_COUNT_HW_CACHE_OP_MAX]
615 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
616 {
617 [ C(L1D) ] = {
618 [ C(OP_READ) ] = {
619 [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */
620 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */
621 },
622 [ C(OP_WRITE) ] = {
623 [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */
624 [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */
625 },
626 [ C(OP_PREFETCH) ] = {
627 [ C(RESULT_ACCESS) ] = 0x0,
628 [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */
629 },
630 },
631 [ C(L1I ) ] = {
632 [ C(OP_READ) ] = {
633 [ C(RESULT_ACCESS) ] = 0x0,
634 [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */
635 },
636 [ C(OP_WRITE) ] = {
637 [ C(RESULT_ACCESS) ] = -1,
638 [ C(RESULT_MISS) ] = -1,
639 },
640 [ C(OP_PREFETCH) ] = {
641 [ C(RESULT_ACCESS) ] = 0x0,
642 [ C(RESULT_MISS) ] = 0x0,
643 },
644 },
645 [ C(LL ) ] = {
646 [ C(OP_READ) ] = {
647 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
648 [ C(RESULT_ACCESS) ] = 0x01b7,
649 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
650 [ C(RESULT_MISS) ] = 0x01b7,
651 },
652 [ C(OP_WRITE) ] = {
653 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
654 [ C(RESULT_ACCESS) ] = 0x01b7,
655 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
656 [ C(RESULT_MISS) ] = 0x01b7,
657 },
658 [ C(OP_PREFETCH) ] = {
659 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
660 [ C(RESULT_ACCESS) ] = 0x01b7,
661 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
662 [ C(RESULT_MISS) ] = 0x01b7,
663 },
664 },
665 [ C(DTLB) ] = {
666 [ C(OP_READ) ] = {
667 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */
668 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */
669 },
670 [ C(OP_WRITE) ] = {
671 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */
672 [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
673 },
674 [ C(OP_PREFETCH) ] = {
675 [ C(RESULT_ACCESS) ] = 0x0,
676 [ C(RESULT_MISS) ] = 0x0,
677 },
678 },
679 [ C(ITLB) ] = {
680 [ C(OP_READ) ] = {
681 [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */
682 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */
683 },
684 [ C(OP_WRITE) ] = {
685 [ C(RESULT_ACCESS) ] = -1,
686 [ C(RESULT_MISS) ] = -1,
687 },
688 [ C(OP_PREFETCH) ] = {
689 [ C(RESULT_ACCESS) ] = -1,
690 [ C(RESULT_MISS) ] = -1,
691 },
692 },
693 [ C(BPU ) ] = {
694 [ C(OP_READ) ] = {
695 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
696 [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
697 },
698 [ C(OP_WRITE) ] = {
699 [ C(RESULT_ACCESS) ] = -1,
700 [ C(RESULT_MISS) ] = -1,
701 },
702 [ C(OP_PREFETCH) ] = {
703 [ C(RESULT_ACCESS) ] = -1,
704 [ C(RESULT_MISS) ] = -1,
705 },
706 },
707 [ C(NODE) ] = {
708 [ C(OP_READ) ] = {
709 [ C(RESULT_ACCESS) ] = 0x01b7,
710 [ C(RESULT_MISS) ] = 0x01b7,
711 },
712 [ C(OP_WRITE) ] = {
713 [ C(RESULT_ACCESS) ] = 0x01b7,
714 [ C(RESULT_MISS) ] = 0x01b7,
715 },
716 [ C(OP_PREFETCH) ] = {
717 [ C(RESULT_ACCESS) ] = 0x01b7,
718 [ C(RESULT_MISS) ] = 0x01b7,
719 },
720 },
721
722 };
723
724 /*
725 * Notes on the events:
726 * - data reads do not include code reads (comparable to earlier tables)
727 * - data counts include speculative execution (except L1 write, dtlb, bpu)
728 * - remote node access includes remote memory, remote cache, remote mmio.
729 * - prefetches are not included in the counts because they are not
730 * reliably counted.
731 */
732
733 #define HSW_DEMAND_DATA_RD BIT_ULL(0)
734 #define HSW_DEMAND_RFO BIT_ULL(1)
735 #define HSW_ANY_RESPONSE BIT_ULL(16)
736 #define HSW_SUPPLIER_NONE BIT_ULL(17)
737 #define HSW_L3_MISS_LOCAL_DRAM BIT_ULL(22)
738 #define HSW_L3_MISS_REMOTE_HOP0 BIT_ULL(27)
739 #define HSW_L3_MISS_REMOTE_HOP1 BIT_ULL(28)
740 #define HSW_L3_MISS_REMOTE_HOP2P BIT_ULL(29)
741 #define HSW_L3_MISS (HSW_L3_MISS_LOCAL_DRAM| \
742 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \
743 HSW_L3_MISS_REMOTE_HOP2P)
744 #define HSW_SNOOP_NONE BIT_ULL(31)
745 #define HSW_SNOOP_NOT_NEEDED BIT_ULL(32)
746 #define HSW_SNOOP_MISS BIT_ULL(33)
747 #define HSW_SNOOP_HIT_NO_FWD BIT_ULL(34)
748 #define HSW_SNOOP_HIT_WITH_FWD BIT_ULL(35)
749 #define HSW_SNOOP_HITM BIT_ULL(36)
750 #define HSW_SNOOP_NON_DRAM BIT_ULL(37)
751 #define HSW_ANY_SNOOP (HSW_SNOOP_NONE| \
752 HSW_SNOOP_NOT_NEEDED|HSW_SNOOP_MISS| \
753 HSW_SNOOP_HIT_NO_FWD|HSW_SNOOP_HIT_WITH_FWD| \
754 HSW_SNOOP_HITM|HSW_SNOOP_NON_DRAM)
755 #define HSW_SNOOP_DRAM (HSW_ANY_SNOOP & ~HSW_SNOOP_NON_DRAM)
756 #define HSW_DEMAND_READ HSW_DEMAND_DATA_RD
757 #define HSW_DEMAND_WRITE HSW_DEMAND_RFO
758 #define HSW_L3_MISS_REMOTE (HSW_L3_MISS_REMOTE_HOP0|\
759 HSW_L3_MISS_REMOTE_HOP1|HSW_L3_MISS_REMOTE_HOP2P)
760 #define HSW_LLC_ACCESS HSW_ANY_RESPONSE
761
762 #define BDW_L3_MISS_LOCAL BIT(26)
763 #define BDW_L3_MISS (BDW_L3_MISS_LOCAL| \
764 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \
765 HSW_L3_MISS_REMOTE_HOP2P)
766
767
768 static __initconst const u64 hsw_hw_cache_event_ids
769 [PERF_COUNT_HW_CACHE_MAX]
770 [PERF_COUNT_HW_CACHE_OP_MAX]
771 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
772 {
773 [ C(L1D ) ] = {
774 [ C(OP_READ) ] = {
775 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
776 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */
777 },
778 [ C(OP_WRITE) ] = {
779 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
780 [ C(RESULT_MISS) ] = 0x0,
781 },
782 [ C(OP_PREFETCH) ] = {
783 [ C(RESULT_ACCESS) ] = 0x0,
784 [ C(RESULT_MISS) ] = 0x0,
785 },
786 },
787 [ C(L1I ) ] = {
788 [ C(OP_READ) ] = {
789 [ C(RESULT_ACCESS) ] = 0x0,
790 [ C(RESULT_MISS) ] = 0x280, /* ICACHE.MISSES */
791 },
792 [ C(OP_WRITE) ] = {
793 [ C(RESULT_ACCESS) ] = -1,
794 [ C(RESULT_MISS) ] = -1,
795 },
796 [ C(OP_PREFETCH) ] = {
797 [ C(RESULT_ACCESS) ] = 0x0,
798 [ C(RESULT_MISS) ] = 0x0,
799 },
800 },
801 [ C(LL ) ] = {
802 [ C(OP_READ) ] = {
803 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
804 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
805 },
806 [ C(OP_WRITE) ] = {
807 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
808 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
809 },
810 [ C(OP_PREFETCH) ] = {
811 [ C(RESULT_ACCESS) ] = 0x0,
812 [ C(RESULT_MISS) ] = 0x0,
813 },
814 },
815 [ C(DTLB) ] = {
816 [ C(OP_READ) ] = {
817 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
818 [ C(RESULT_MISS) ] = 0x108, /* DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK */
819 },
820 [ C(OP_WRITE) ] = {
821 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
822 [ C(RESULT_MISS) ] = 0x149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
823 },
824 [ C(OP_PREFETCH) ] = {
825 [ C(RESULT_ACCESS) ] = 0x0,
826 [ C(RESULT_MISS) ] = 0x0,
827 },
828 },
829 [ C(ITLB) ] = {
830 [ C(OP_READ) ] = {
831 [ C(RESULT_ACCESS) ] = 0x6085, /* ITLB_MISSES.STLB_HIT */
832 [ C(RESULT_MISS) ] = 0x185, /* ITLB_MISSES.MISS_CAUSES_A_WALK */
833 },
834 [ C(OP_WRITE) ] = {
835 [ C(RESULT_ACCESS) ] = -1,
836 [ C(RESULT_MISS) ] = -1,
837 },
838 [ C(OP_PREFETCH) ] = {
839 [ C(RESULT_ACCESS) ] = -1,
840 [ C(RESULT_MISS) ] = -1,
841 },
842 },
843 [ C(BPU ) ] = {
844 [ C(OP_READ) ] = {
845 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */
846 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */
847 },
848 [ C(OP_WRITE) ] = {
849 [ C(RESULT_ACCESS) ] = -1,
850 [ C(RESULT_MISS) ] = -1,
851 },
852 [ C(OP_PREFETCH) ] = {
853 [ C(RESULT_ACCESS) ] = -1,
854 [ C(RESULT_MISS) ] = -1,
855 },
856 },
857 [ C(NODE) ] = {
858 [ C(OP_READ) ] = {
859 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
860 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
861 },
862 [ C(OP_WRITE) ] = {
863 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
864 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
865 },
866 [ C(OP_PREFETCH) ] = {
867 [ C(RESULT_ACCESS) ] = 0x0,
868 [ C(RESULT_MISS) ] = 0x0,
869 },
870 },
871 };
872
873 static __initconst const u64 hsw_hw_cache_extra_regs
874 [PERF_COUNT_HW_CACHE_MAX]
875 [PERF_COUNT_HW_CACHE_OP_MAX]
876 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
877 {
878 [ C(LL ) ] = {
879 [ C(OP_READ) ] = {
880 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ|
881 HSW_LLC_ACCESS,
882 [ C(RESULT_MISS) ] = HSW_DEMAND_READ|
883 HSW_L3_MISS|HSW_ANY_SNOOP,
884 },
885 [ C(OP_WRITE) ] = {
886 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE|
887 HSW_LLC_ACCESS,
888 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE|
889 HSW_L3_MISS|HSW_ANY_SNOOP,
890 },
891 [ C(OP_PREFETCH) ] = {
892 [ C(RESULT_ACCESS) ] = 0x0,
893 [ C(RESULT_MISS) ] = 0x0,
894 },
895 },
896 [ C(NODE) ] = {
897 [ C(OP_READ) ] = {
898 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ|
899 HSW_L3_MISS_LOCAL_DRAM|
900 HSW_SNOOP_DRAM,
901 [ C(RESULT_MISS) ] = HSW_DEMAND_READ|
902 HSW_L3_MISS_REMOTE|
903 HSW_SNOOP_DRAM,
904 },
905 [ C(OP_WRITE) ] = {
906 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE|
907 HSW_L3_MISS_LOCAL_DRAM|
908 HSW_SNOOP_DRAM,
909 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE|
910 HSW_L3_MISS_REMOTE|
911 HSW_SNOOP_DRAM,
912 },
913 [ C(OP_PREFETCH) ] = {
914 [ C(RESULT_ACCESS) ] = 0x0,
915 [ C(RESULT_MISS) ] = 0x0,
916 },
917 },
918 };
919
920 static __initconst const u64 westmere_hw_cache_event_ids
921 [PERF_COUNT_HW_CACHE_MAX]
922 [PERF_COUNT_HW_CACHE_OP_MAX]
923 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
924 {
925 [ C(L1D) ] = {
926 [ C(OP_READ) ] = {
927 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
928 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
929 },
930 [ C(OP_WRITE) ] = {
931 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
932 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
933 },
934 [ C(OP_PREFETCH) ] = {
935 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
936 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
937 },
938 },
939 [ C(L1I ) ] = {
940 [ C(OP_READ) ] = {
941 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
942 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
943 },
944 [ C(OP_WRITE) ] = {
945 [ C(RESULT_ACCESS) ] = -1,
946 [ C(RESULT_MISS) ] = -1,
947 },
948 [ C(OP_PREFETCH) ] = {
949 [ C(RESULT_ACCESS) ] = 0x0,
950 [ C(RESULT_MISS) ] = 0x0,
951 },
952 },
953 [ C(LL ) ] = {
954 [ C(OP_READ) ] = {
955 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
956 [ C(RESULT_ACCESS) ] = 0x01b7,
957 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
958 [ C(RESULT_MISS) ] = 0x01b7,
959 },
960 /*
961 * Use RFO, not WRITEBACK, because a write miss would typically occur
962 * on RFO.
963 */
964 [ C(OP_WRITE) ] = {
965 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
966 [ C(RESULT_ACCESS) ] = 0x01b7,
967 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
968 [ C(RESULT_MISS) ] = 0x01b7,
969 },
970 [ C(OP_PREFETCH) ] = {
971 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
972 [ C(RESULT_ACCESS) ] = 0x01b7,
973 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
974 [ C(RESULT_MISS) ] = 0x01b7,
975 },
976 },
977 [ C(DTLB) ] = {
978 [ C(OP_READ) ] = {
979 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
980 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
981 },
982 [ C(OP_WRITE) ] = {
983 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
984 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
985 },
986 [ C(OP_PREFETCH) ] = {
987 [ C(RESULT_ACCESS) ] = 0x0,
988 [ C(RESULT_MISS) ] = 0x0,
989 },
990 },
991 [ C(ITLB) ] = {
992 [ C(OP_READ) ] = {
993 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
994 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */
995 },
996 [ C(OP_WRITE) ] = {
997 [ C(RESULT_ACCESS) ] = -1,
998 [ C(RESULT_MISS) ] = -1,
999 },
1000 [ C(OP_PREFETCH) ] = {
1001 [ C(RESULT_ACCESS) ] = -1,
1002 [ C(RESULT_MISS) ] = -1,
1003 },
1004 },
1005 [ C(BPU ) ] = {
1006 [ C(OP_READ) ] = {
1007 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1008 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
1009 },
1010 [ C(OP_WRITE) ] = {
1011 [ C(RESULT_ACCESS) ] = -1,
1012 [ C(RESULT_MISS) ] = -1,
1013 },
1014 [ C(OP_PREFETCH) ] = {
1015 [ C(RESULT_ACCESS) ] = -1,
1016 [ C(RESULT_MISS) ] = -1,
1017 },
1018 },
1019 [ C(NODE) ] = {
1020 [ C(OP_READ) ] = {
1021 [ C(RESULT_ACCESS) ] = 0x01b7,
1022 [ C(RESULT_MISS) ] = 0x01b7,
1023 },
1024 [ C(OP_WRITE) ] = {
1025 [ C(RESULT_ACCESS) ] = 0x01b7,
1026 [ C(RESULT_MISS) ] = 0x01b7,
1027 },
1028 [ C(OP_PREFETCH) ] = {
1029 [ C(RESULT_ACCESS) ] = 0x01b7,
1030 [ C(RESULT_MISS) ] = 0x01b7,
1031 },
1032 },
1033 };
1034
1035 /*
1036 * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits;
1037 * See IA32 SDM Vol 3B 30.6.1.3
1038 */
1039
1040 #define NHM_DMND_DATA_RD (1 << 0)
1041 #define NHM_DMND_RFO (1 << 1)
1042 #define NHM_DMND_IFETCH (1 << 2)
1043 #define NHM_DMND_WB (1 << 3)
1044 #define NHM_PF_DATA_RD (1 << 4)
1045 #define NHM_PF_DATA_RFO (1 << 5)
1046 #define NHM_PF_IFETCH (1 << 6)
1047 #define NHM_OFFCORE_OTHER (1 << 7)
1048 #define NHM_UNCORE_HIT (1 << 8)
1049 #define NHM_OTHER_CORE_HIT_SNP (1 << 9)
1050 #define NHM_OTHER_CORE_HITM (1 << 10)
1051 /* reserved */
1052 #define NHM_REMOTE_CACHE_FWD (1 << 12)
1053 #define NHM_REMOTE_DRAM (1 << 13)
1054 #define NHM_LOCAL_DRAM (1 << 14)
1055 #define NHM_NON_DRAM (1 << 15)
1056
1057 #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD)
1058 #define NHM_REMOTE (NHM_REMOTE_DRAM)
1059
1060 #define NHM_DMND_READ (NHM_DMND_DATA_RD)
1061 #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB)
1062 #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO)
1063
1064 #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM)
1065 #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD)
1066 #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS)
1067
1068 static __initconst const u64 nehalem_hw_cache_extra_regs
1069 [PERF_COUNT_HW_CACHE_MAX]
1070 [PERF_COUNT_HW_CACHE_OP_MAX]
1071 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1072 {
1073 [ C(LL ) ] = {
1074 [ C(OP_READ) ] = {
1075 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS,
1076 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS,
1077 },
1078 [ C(OP_WRITE) ] = {
1079 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS,
1080 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS,
1081 },
1082 [ C(OP_PREFETCH) ] = {
1083 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS,
1084 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS,
1085 },
1086 },
1087 [ C(NODE) ] = {
1088 [ C(OP_READ) ] = {
1089 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE,
1090 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE,
1091 },
1092 [ C(OP_WRITE) ] = {
1093 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE,
1094 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE,
1095 },
1096 [ C(OP_PREFETCH) ] = {
1097 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE,
1098 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE,
1099 },
1100 },
1101 };
1102
1103 static __initconst const u64 nehalem_hw_cache_event_ids
1104 [PERF_COUNT_HW_CACHE_MAX]
1105 [PERF_COUNT_HW_CACHE_OP_MAX]
1106 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1107 {
1108 [ C(L1D) ] = {
1109 [ C(OP_READ) ] = {
1110 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
1111 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
1112 },
1113 [ C(OP_WRITE) ] = {
1114 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
1115 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
1116 },
1117 [ C(OP_PREFETCH) ] = {
1118 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
1119 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
1120 },
1121 },
1122 [ C(L1I ) ] = {
1123 [ C(OP_READ) ] = {
1124 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
1125 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
1126 },
1127 [ C(OP_WRITE) ] = {
1128 [ C(RESULT_ACCESS) ] = -1,
1129 [ C(RESULT_MISS) ] = -1,
1130 },
1131 [ C(OP_PREFETCH) ] = {
1132 [ C(RESULT_ACCESS) ] = 0x0,
1133 [ C(RESULT_MISS) ] = 0x0,
1134 },
1135 },
1136 [ C(LL ) ] = {
1137 [ C(OP_READ) ] = {
1138 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1139 [ C(RESULT_ACCESS) ] = 0x01b7,
1140 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
1141 [ C(RESULT_MISS) ] = 0x01b7,
1142 },
1143 /*
1144 * Use RFO, not WRITEBACK, because a write miss would typically occur
1145 * on RFO.
1146 */
1147 [ C(OP_WRITE) ] = {
1148 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1149 [ C(RESULT_ACCESS) ] = 0x01b7,
1150 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1151 [ C(RESULT_MISS) ] = 0x01b7,
1152 },
1153 [ C(OP_PREFETCH) ] = {
1154 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1155 [ C(RESULT_ACCESS) ] = 0x01b7,
1156 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1157 [ C(RESULT_MISS) ] = 0x01b7,
1158 },
1159 },
1160 [ C(DTLB) ] = {
1161 [ C(OP_READ) ] = {
1162 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
1163 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
1164 },
1165 [ C(OP_WRITE) ] = {
1166 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
1167 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
1168 },
1169 [ C(OP_PREFETCH) ] = {
1170 [ C(RESULT_ACCESS) ] = 0x0,
1171 [ C(RESULT_MISS) ] = 0x0,
1172 },
1173 },
1174 [ C(ITLB) ] = {
1175 [ C(OP_READ) ] = {
1176 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
1177 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */
1178 },
1179 [ C(OP_WRITE) ] = {
1180 [ C(RESULT_ACCESS) ] = -1,
1181 [ C(RESULT_MISS) ] = -1,
1182 },
1183 [ C(OP_PREFETCH) ] = {
1184 [ C(RESULT_ACCESS) ] = -1,
1185 [ C(RESULT_MISS) ] = -1,
1186 },
1187 },
1188 [ C(BPU ) ] = {
1189 [ C(OP_READ) ] = {
1190 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1191 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
1192 },
1193 [ C(OP_WRITE) ] = {
1194 [ C(RESULT_ACCESS) ] = -1,
1195 [ C(RESULT_MISS) ] = -1,
1196 },
1197 [ C(OP_PREFETCH) ] = {
1198 [ C(RESULT_ACCESS) ] = -1,
1199 [ C(RESULT_MISS) ] = -1,
1200 },
1201 },
1202 [ C(NODE) ] = {
1203 [ C(OP_READ) ] = {
1204 [ C(RESULT_ACCESS) ] = 0x01b7,
1205 [ C(RESULT_MISS) ] = 0x01b7,
1206 },
1207 [ C(OP_WRITE) ] = {
1208 [ C(RESULT_ACCESS) ] = 0x01b7,
1209 [ C(RESULT_MISS) ] = 0x01b7,
1210 },
1211 [ C(OP_PREFETCH) ] = {
1212 [ C(RESULT_ACCESS) ] = 0x01b7,
1213 [ C(RESULT_MISS) ] = 0x01b7,
1214 },
1215 },
1216 };
1217
1218 static __initconst const u64 core2_hw_cache_event_ids
1219 [PERF_COUNT_HW_CACHE_MAX]
1220 [PERF_COUNT_HW_CACHE_OP_MAX]
1221 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1222 {
1223 [ C(L1D) ] = {
1224 [ C(OP_READ) ] = {
1225 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
1226 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
1227 },
1228 [ C(OP_WRITE) ] = {
1229 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
1230 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
1231 },
1232 [ C(OP_PREFETCH) ] = {
1233 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */
1234 [ C(RESULT_MISS) ] = 0,
1235 },
1236 },
1237 [ C(L1I ) ] = {
1238 [ C(OP_READ) ] = {
1239 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */
1240 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */
1241 },
1242 [ C(OP_WRITE) ] = {
1243 [ C(RESULT_ACCESS) ] = -1,
1244 [ C(RESULT_MISS) ] = -1,
1245 },
1246 [ C(OP_PREFETCH) ] = {
1247 [ C(RESULT_ACCESS) ] = 0,
1248 [ C(RESULT_MISS) ] = 0,
1249 },
1250 },
1251 [ C(LL ) ] = {
1252 [ C(OP_READ) ] = {
1253 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
1254 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
1255 },
1256 [ C(OP_WRITE) ] = {
1257 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
1258 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
1259 },
1260 [ C(OP_PREFETCH) ] = {
1261 [ C(RESULT_ACCESS) ] = 0,
1262 [ C(RESULT_MISS) ] = 0,
1263 },
1264 },
1265 [ C(DTLB) ] = {
1266 [ C(OP_READ) ] = {
1267 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
1268 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */
1269 },
1270 [ C(OP_WRITE) ] = {
1271 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
1272 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */
1273 },
1274 [ C(OP_PREFETCH) ] = {
1275 [ C(RESULT_ACCESS) ] = 0,
1276 [ C(RESULT_MISS) ] = 0,
1277 },
1278 },
1279 [ C(ITLB) ] = {
1280 [ C(OP_READ) ] = {
1281 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1282 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */
1283 },
1284 [ C(OP_WRITE) ] = {
1285 [ C(RESULT_ACCESS) ] = -1,
1286 [ C(RESULT_MISS) ] = -1,
1287 },
1288 [ C(OP_PREFETCH) ] = {
1289 [ C(RESULT_ACCESS) ] = -1,
1290 [ C(RESULT_MISS) ] = -1,
1291 },
1292 },
1293 [ C(BPU ) ] = {
1294 [ C(OP_READ) ] = {
1295 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1296 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1297 },
1298 [ C(OP_WRITE) ] = {
1299 [ C(RESULT_ACCESS) ] = -1,
1300 [ C(RESULT_MISS) ] = -1,
1301 },
1302 [ C(OP_PREFETCH) ] = {
1303 [ C(RESULT_ACCESS) ] = -1,
1304 [ C(RESULT_MISS) ] = -1,
1305 },
1306 },
1307 };
1308
1309 static __initconst const u64 atom_hw_cache_event_ids
1310 [PERF_COUNT_HW_CACHE_MAX]
1311 [PERF_COUNT_HW_CACHE_OP_MAX]
1312 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1313 {
1314 [ C(L1D) ] = {
1315 [ C(OP_READ) ] = {
1316 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */
1317 [ C(RESULT_MISS) ] = 0,
1318 },
1319 [ C(OP_WRITE) ] = {
1320 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */
1321 [ C(RESULT_MISS) ] = 0,
1322 },
1323 [ C(OP_PREFETCH) ] = {
1324 [ C(RESULT_ACCESS) ] = 0x0,
1325 [ C(RESULT_MISS) ] = 0,
1326 },
1327 },
1328 [ C(L1I ) ] = {
1329 [ C(OP_READ) ] = {
1330 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
1331 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
1332 },
1333 [ C(OP_WRITE) ] = {
1334 [ C(RESULT_ACCESS) ] = -1,
1335 [ C(RESULT_MISS) ] = -1,
1336 },
1337 [ C(OP_PREFETCH) ] = {
1338 [ C(RESULT_ACCESS) ] = 0,
1339 [ C(RESULT_MISS) ] = 0,
1340 },
1341 },
1342 [ C(LL ) ] = {
1343 [ C(OP_READ) ] = {
1344 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
1345 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
1346 },
1347 [ C(OP_WRITE) ] = {
1348 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
1349 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
1350 },
1351 [ C(OP_PREFETCH) ] = {
1352 [ C(RESULT_ACCESS) ] = 0,
1353 [ C(RESULT_MISS) ] = 0,
1354 },
1355 },
1356 [ C(DTLB) ] = {
1357 [ C(OP_READ) ] = {
1358 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */
1359 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */
1360 },
1361 [ C(OP_WRITE) ] = {
1362 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */
1363 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */
1364 },
1365 [ C(OP_PREFETCH) ] = {
1366 [ C(RESULT_ACCESS) ] = 0,
1367 [ C(RESULT_MISS) ] = 0,
1368 },
1369 },
1370 [ C(ITLB) ] = {
1371 [ C(OP_READ) ] = {
1372 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1373 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */
1374 },
1375 [ C(OP_WRITE) ] = {
1376 [ C(RESULT_ACCESS) ] = -1,
1377 [ C(RESULT_MISS) ] = -1,
1378 },
1379 [ C(OP_PREFETCH) ] = {
1380 [ C(RESULT_ACCESS) ] = -1,
1381 [ C(RESULT_MISS) ] = -1,
1382 },
1383 },
1384 [ C(BPU ) ] = {
1385 [ C(OP_READ) ] = {
1386 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1387 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1388 },
1389 [ C(OP_WRITE) ] = {
1390 [ C(RESULT_ACCESS) ] = -1,
1391 [ C(RESULT_MISS) ] = -1,
1392 },
1393 [ C(OP_PREFETCH) ] = {
1394 [ C(RESULT_ACCESS) ] = -1,
1395 [ C(RESULT_MISS) ] = -1,
1396 },
1397 },
1398 };
1399
1400 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_slm, "event=0x3c");
1401 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_slm, "2");
1402 /* no_alloc_cycles.not_delivered */
1403 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_slm,
1404 "event=0xca,umask=0x50");
1405 EVENT_ATTR_STR(topdown-fetch-bubbles.scale, td_fetch_bubbles_scale_slm, "2");
1406 /* uops_retired.all */
1407 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_slm,
1408 "event=0xc2,umask=0x10");
1409 /* uops_retired.all */
1410 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_slm,
1411 "event=0xc2,umask=0x10");
1412
1413 static struct attribute *slm_events_attrs[] = {
1414 EVENT_PTR(td_total_slots_slm),
1415 EVENT_PTR(td_total_slots_scale_slm),
1416 EVENT_PTR(td_fetch_bubbles_slm),
1417 EVENT_PTR(td_fetch_bubbles_scale_slm),
1418 EVENT_PTR(td_slots_issued_slm),
1419 EVENT_PTR(td_slots_retired_slm),
1420 NULL
1421 };
1422
1423 static struct extra_reg intel_slm_extra_regs[] __read_mostly =
1424 {
1425 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
1426 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x768005ffffull, RSP_0),
1427 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x368005ffffull, RSP_1),
1428 EVENT_EXTRA_END
1429 };
1430
1431 #define SLM_DMND_READ SNB_DMND_DATA_RD
1432 #define SLM_DMND_WRITE SNB_DMND_RFO
1433 #define SLM_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
1434
1435 #define SLM_SNP_ANY (SNB_SNP_NONE|SNB_SNP_MISS|SNB_NO_FWD|SNB_HITM)
1436 #define SLM_LLC_ACCESS SNB_RESP_ANY
1437 #define SLM_LLC_MISS (SLM_SNP_ANY|SNB_NON_DRAM)
1438
1439 static __initconst const u64 slm_hw_cache_extra_regs
1440 [PERF_COUNT_HW_CACHE_MAX]
1441 [PERF_COUNT_HW_CACHE_OP_MAX]
1442 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1443 {
1444 [ C(LL ) ] = {
1445 [ C(OP_READ) ] = {
1446 [ C(RESULT_ACCESS) ] = SLM_DMND_READ|SLM_LLC_ACCESS,
1447 [ C(RESULT_MISS) ] = 0,
1448 },
1449 [ C(OP_WRITE) ] = {
1450 [ C(RESULT_ACCESS) ] = SLM_DMND_WRITE|SLM_LLC_ACCESS,
1451 [ C(RESULT_MISS) ] = SLM_DMND_WRITE|SLM_LLC_MISS,
1452 },
1453 [ C(OP_PREFETCH) ] = {
1454 [ C(RESULT_ACCESS) ] = SLM_DMND_PREFETCH|SLM_LLC_ACCESS,
1455 [ C(RESULT_MISS) ] = SLM_DMND_PREFETCH|SLM_LLC_MISS,
1456 },
1457 },
1458 };
1459
1460 static __initconst const u64 slm_hw_cache_event_ids
1461 [PERF_COUNT_HW_CACHE_MAX]
1462 [PERF_COUNT_HW_CACHE_OP_MAX]
1463 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1464 {
1465 [ C(L1D) ] = {
1466 [ C(OP_READ) ] = {
1467 [ C(RESULT_ACCESS) ] = 0,
1468 [ C(RESULT_MISS) ] = 0x0104, /* LD_DCU_MISS */
1469 },
1470 [ C(OP_WRITE) ] = {
1471 [ C(RESULT_ACCESS) ] = 0,
1472 [ C(RESULT_MISS) ] = 0,
1473 },
1474 [ C(OP_PREFETCH) ] = {
1475 [ C(RESULT_ACCESS) ] = 0,
1476 [ C(RESULT_MISS) ] = 0,
1477 },
1478 },
1479 [ C(L1I ) ] = {
1480 [ C(OP_READ) ] = {
1481 [ C(RESULT_ACCESS) ] = 0x0380, /* ICACHE.ACCESSES */
1482 [ C(RESULT_MISS) ] = 0x0280, /* ICACGE.MISSES */
1483 },
1484 [ C(OP_WRITE) ] = {
1485 [ C(RESULT_ACCESS) ] = -1,
1486 [ C(RESULT_MISS) ] = -1,
1487 },
1488 [ C(OP_PREFETCH) ] = {
1489 [ C(RESULT_ACCESS) ] = 0,
1490 [ C(RESULT_MISS) ] = 0,
1491 },
1492 },
1493 [ C(LL ) ] = {
1494 [ C(OP_READ) ] = {
1495 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1496 [ C(RESULT_ACCESS) ] = 0x01b7,
1497 [ C(RESULT_MISS) ] = 0,
1498 },
1499 [ C(OP_WRITE) ] = {
1500 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1501 [ C(RESULT_ACCESS) ] = 0x01b7,
1502 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1503 [ C(RESULT_MISS) ] = 0x01b7,
1504 },
1505 [ C(OP_PREFETCH) ] = {
1506 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1507 [ C(RESULT_ACCESS) ] = 0x01b7,
1508 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1509 [ C(RESULT_MISS) ] = 0x01b7,
1510 },
1511 },
1512 [ C(DTLB) ] = {
1513 [ C(OP_READ) ] = {
1514 [ C(RESULT_ACCESS) ] = 0,
1515 [ C(RESULT_MISS) ] = 0x0804, /* LD_DTLB_MISS */
1516 },
1517 [ C(OP_WRITE) ] = {
1518 [ C(RESULT_ACCESS) ] = 0,
1519 [ C(RESULT_MISS) ] = 0,
1520 },
1521 [ C(OP_PREFETCH) ] = {
1522 [ C(RESULT_ACCESS) ] = 0,
1523 [ C(RESULT_MISS) ] = 0,
1524 },
1525 },
1526 [ C(ITLB) ] = {
1527 [ C(OP_READ) ] = {
1528 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1529 [ C(RESULT_MISS) ] = 0x40205, /* PAGE_WALKS.I_SIDE_WALKS */
1530 },
1531 [ C(OP_WRITE) ] = {
1532 [ C(RESULT_ACCESS) ] = -1,
1533 [ C(RESULT_MISS) ] = -1,
1534 },
1535 [ C(OP_PREFETCH) ] = {
1536 [ C(RESULT_ACCESS) ] = -1,
1537 [ C(RESULT_MISS) ] = -1,
1538 },
1539 },
1540 [ C(BPU ) ] = {
1541 [ C(OP_READ) ] = {
1542 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1543 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1544 },
1545 [ C(OP_WRITE) ] = {
1546 [ C(RESULT_ACCESS) ] = -1,
1547 [ C(RESULT_MISS) ] = -1,
1548 },
1549 [ C(OP_PREFETCH) ] = {
1550 [ C(RESULT_ACCESS) ] = -1,
1551 [ C(RESULT_MISS) ] = -1,
1552 },
1553 },
1554 };
1555
1556 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_glm, "event=0x3c");
1557 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_glm, "3");
1558 /* UOPS_NOT_DELIVERED.ANY */
1559 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_glm, "event=0x9c");
1560 /* ISSUE_SLOTS_NOT_CONSUMED.RECOVERY */
1561 EVENT_ATTR_STR(topdown-recovery-bubbles, td_recovery_bubbles_glm, "event=0xca,umask=0x02");
1562 /* UOPS_RETIRED.ANY */
1563 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_glm, "event=0xc2");
1564 /* UOPS_ISSUED.ANY */
1565 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_glm, "event=0x0e");
1566
1567 static struct attribute *glm_events_attrs[] = {
1568 EVENT_PTR(td_total_slots_glm),
1569 EVENT_PTR(td_total_slots_scale_glm),
1570 EVENT_PTR(td_fetch_bubbles_glm),
1571 EVENT_PTR(td_recovery_bubbles_glm),
1572 EVENT_PTR(td_slots_issued_glm),
1573 EVENT_PTR(td_slots_retired_glm),
1574 NULL
1575 };
1576
1577 static struct extra_reg intel_glm_extra_regs[] __read_mostly = {
1578 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
1579 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x760005ffbfull, RSP_0),
1580 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x360005ffbfull, RSP_1),
1581 EVENT_EXTRA_END
1582 };
1583
1584 #define GLM_DEMAND_DATA_RD BIT_ULL(0)
1585 #define GLM_DEMAND_RFO BIT_ULL(1)
1586 #define GLM_ANY_RESPONSE BIT_ULL(16)
1587 #define GLM_SNP_NONE_OR_MISS BIT_ULL(33)
1588 #define GLM_DEMAND_READ GLM_DEMAND_DATA_RD
1589 #define GLM_DEMAND_WRITE GLM_DEMAND_RFO
1590 #define GLM_DEMAND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
1591 #define GLM_LLC_ACCESS GLM_ANY_RESPONSE
1592 #define GLM_SNP_ANY (GLM_SNP_NONE_OR_MISS|SNB_NO_FWD|SNB_HITM)
1593 #define GLM_LLC_MISS (GLM_SNP_ANY|SNB_NON_DRAM)
1594
1595 static __initconst const u64 glm_hw_cache_event_ids
1596 [PERF_COUNT_HW_CACHE_MAX]
1597 [PERF_COUNT_HW_CACHE_OP_MAX]
1598 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1599 [C(L1D)] = {
1600 [C(OP_READ)] = {
1601 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
1602 [C(RESULT_MISS)] = 0x0,
1603 },
1604 [C(OP_WRITE)] = {
1605 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
1606 [C(RESULT_MISS)] = 0x0,
1607 },
1608 [C(OP_PREFETCH)] = {
1609 [C(RESULT_ACCESS)] = 0x0,
1610 [C(RESULT_MISS)] = 0x0,
1611 },
1612 },
1613 [C(L1I)] = {
1614 [C(OP_READ)] = {
1615 [C(RESULT_ACCESS)] = 0x0380, /* ICACHE.ACCESSES */
1616 [C(RESULT_MISS)] = 0x0280, /* ICACHE.MISSES */
1617 },
1618 [C(OP_WRITE)] = {
1619 [C(RESULT_ACCESS)] = -1,
1620 [C(RESULT_MISS)] = -1,
1621 },
1622 [C(OP_PREFETCH)] = {
1623 [C(RESULT_ACCESS)] = 0x0,
1624 [C(RESULT_MISS)] = 0x0,
1625 },
1626 },
1627 [C(LL)] = {
1628 [C(OP_READ)] = {
1629 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1630 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1631 },
1632 [C(OP_WRITE)] = {
1633 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1634 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1635 },
1636 [C(OP_PREFETCH)] = {
1637 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1638 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1639 },
1640 },
1641 [C(DTLB)] = {
1642 [C(OP_READ)] = {
1643 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
1644 [C(RESULT_MISS)] = 0x0,
1645 },
1646 [C(OP_WRITE)] = {
1647 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
1648 [C(RESULT_MISS)] = 0x0,
1649 },
1650 [C(OP_PREFETCH)] = {
1651 [C(RESULT_ACCESS)] = 0x0,
1652 [C(RESULT_MISS)] = 0x0,
1653 },
1654 },
1655 [C(ITLB)] = {
1656 [C(OP_READ)] = {
1657 [C(RESULT_ACCESS)] = 0x00c0, /* INST_RETIRED.ANY_P */
1658 [C(RESULT_MISS)] = 0x0481, /* ITLB.MISS */
1659 },
1660 [C(OP_WRITE)] = {
1661 [C(RESULT_ACCESS)] = -1,
1662 [C(RESULT_MISS)] = -1,
1663 },
1664 [C(OP_PREFETCH)] = {
1665 [C(RESULT_ACCESS)] = -1,
1666 [C(RESULT_MISS)] = -1,
1667 },
1668 },
1669 [C(BPU)] = {
1670 [C(OP_READ)] = {
1671 [C(RESULT_ACCESS)] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1672 [C(RESULT_MISS)] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
1673 },
1674 [C(OP_WRITE)] = {
1675 [C(RESULT_ACCESS)] = -1,
1676 [C(RESULT_MISS)] = -1,
1677 },
1678 [C(OP_PREFETCH)] = {
1679 [C(RESULT_ACCESS)] = -1,
1680 [C(RESULT_MISS)] = -1,
1681 },
1682 },
1683 };
1684
1685 static __initconst const u64 glm_hw_cache_extra_regs
1686 [PERF_COUNT_HW_CACHE_MAX]
1687 [PERF_COUNT_HW_CACHE_OP_MAX]
1688 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1689 [C(LL)] = {
1690 [C(OP_READ)] = {
1691 [C(RESULT_ACCESS)] = GLM_DEMAND_READ|
1692 GLM_LLC_ACCESS,
1693 [C(RESULT_MISS)] = GLM_DEMAND_READ|
1694 GLM_LLC_MISS,
1695 },
1696 [C(OP_WRITE)] = {
1697 [C(RESULT_ACCESS)] = GLM_DEMAND_WRITE|
1698 GLM_LLC_ACCESS,
1699 [C(RESULT_MISS)] = GLM_DEMAND_WRITE|
1700 GLM_LLC_MISS,
1701 },
1702 [C(OP_PREFETCH)] = {
1703 [C(RESULT_ACCESS)] = GLM_DEMAND_PREFETCH|
1704 GLM_LLC_ACCESS,
1705 [C(RESULT_MISS)] = GLM_DEMAND_PREFETCH|
1706 GLM_LLC_MISS,
1707 },
1708 },
1709 };
1710
1711 static __initconst const u64 glp_hw_cache_event_ids
1712 [PERF_COUNT_HW_CACHE_MAX]
1713 [PERF_COUNT_HW_CACHE_OP_MAX]
1714 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1715 [C(L1D)] = {
1716 [C(OP_READ)] = {
1717 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
1718 [C(RESULT_MISS)] = 0x0,
1719 },
1720 [C(OP_WRITE)] = {
1721 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
1722 [C(RESULT_MISS)] = 0x0,
1723 },
1724 [C(OP_PREFETCH)] = {
1725 [C(RESULT_ACCESS)] = 0x0,
1726 [C(RESULT_MISS)] = 0x0,
1727 },
1728 },
1729 [C(L1I)] = {
1730 [C(OP_READ)] = {
1731 [C(RESULT_ACCESS)] = 0x0380, /* ICACHE.ACCESSES */
1732 [C(RESULT_MISS)] = 0x0280, /* ICACHE.MISSES */
1733 },
1734 [C(OP_WRITE)] = {
1735 [C(RESULT_ACCESS)] = -1,
1736 [C(RESULT_MISS)] = -1,
1737 },
1738 [C(OP_PREFETCH)] = {
1739 [C(RESULT_ACCESS)] = 0x0,
1740 [C(RESULT_MISS)] = 0x0,
1741 },
1742 },
1743 [C(LL)] = {
1744 [C(OP_READ)] = {
1745 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1746 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1747 },
1748 [C(OP_WRITE)] = {
1749 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */
1750 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */
1751 },
1752 [C(OP_PREFETCH)] = {
1753 [C(RESULT_ACCESS)] = 0x0,
1754 [C(RESULT_MISS)] = 0x0,
1755 },
1756 },
1757 [C(DTLB)] = {
1758 [C(OP_READ)] = {
1759 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
1760 [C(RESULT_MISS)] = 0xe08, /* DTLB_LOAD_MISSES.WALK_COMPLETED */
1761 },
1762 [C(OP_WRITE)] = {
1763 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
1764 [C(RESULT_MISS)] = 0xe49, /* DTLB_STORE_MISSES.WALK_COMPLETED */
1765 },
1766 [C(OP_PREFETCH)] = {
1767 [C(RESULT_ACCESS)] = 0x0,
1768 [C(RESULT_MISS)] = 0x0,
1769 },
1770 },
1771 [C(ITLB)] = {
1772 [C(OP_READ)] = {
1773 [C(RESULT_ACCESS)] = 0x00c0, /* INST_RETIRED.ANY_P */
1774 [C(RESULT_MISS)] = 0x0481, /* ITLB.MISS */
1775 },
1776 [C(OP_WRITE)] = {
1777 [C(RESULT_ACCESS)] = -1,
1778 [C(RESULT_MISS)] = -1,
1779 },
1780 [C(OP_PREFETCH)] = {
1781 [C(RESULT_ACCESS)] = -1,
1782 [C(RESULT_MISS)] = -1,
1783 },
1784 },
1785 [C(BPU)] = {
1786 [C(OP_READ)] = {
1787 [C(RESULT_ACCESS)] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1788 [C(RESULT_MISS)] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
1789 },
1790 [C(OP_WRITE)] = {
1791 [C(RESULT_ACCESS)] = -1,
1792 [C(RESULT_MISS)] = -1,
1793 },
1794 [C(OP_PREFETCH)] = {
1795 [C(RESULT_ACCESS)] = -1,
1796 [C(RESULT_MISS)] = -1,
1797 },
1798 },
1799 };
1800
1801 static __initconst const u64 glp_hw_cache_extra_regs
1802 [PERF_COUNT_HW_CACHE_MAX]
1803 [PERF_COUNT_HW_CACHE_OP_MAX]
1804 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1805 [C(LL)] = {
1806 [C(OP_READ)] = {
1807 [C(RESULT_ACCESS)] = GLM_DEMAND_READ|
1808 GLM_LLC_ACCESS,
1809 [C(RESULT_MISS)] = GLM_DEMAND_READ|
1810 GLM_LLC_MISS,
1811 },
1812 [C(OP_WRITE)] = {
1813 [C(RESULT_ACCESS)] = GLM_DEMAND_WRITE|
1814 GLM_LLC_ACCESS,
1815 [C(RESULT_MISS)] = GLM_DEMAND_WRITE|
1816 GLM_LLC_MISS,
1817 },
1818 [C(OP_PREFETCH)] = {
1819 [C(RESULT_ACCESS)] = 0x0,
1820 [C(RESULT_MISS)] = 0x0,
1821 },
1822 },
1823 };
1824
1825 #define KNL_OT_L2_HITE BIT_ULL(19) /* Other Tile L2 Hit */
1826 #define KNL_OT_L2_HITF BIT_ULL(20) /* Other Tile L2 Hit */
1827 #define KNL_MCDRAM_LOCAL BIT_ULL(21)
1828 #define KNL_MCDRAM_FAR BIT_ULL(22)
1829 #define KNL_DDR_LOCAL BIT_ULL(23)
1830 #define KNL_DDR_FAR BIT_ULL(24)
1831 #define KNL_DRAM_ANY (KNL_MCDRAM_LOCAL | KNL_MCDRAM_FAR | \
1832 KNL_DDR_LOCAL | KNL_DDR_FAR)
1833 #define KNL_L2_READ SLM_DMND_READ
1834 #define KNL_L2_WRITE SLM_DMND_WRITE
1835 #define KNL_L2_PREFETCH SLM_DMND_PREFETCH
1836 #define KNL_L2_ACCESS SLM_LLC_ACCESS
1837 #define KNL_L2_MISS (KNL_OT_L2_HITE | KNL_OT_L2_HITF | \
1838 KNL_DRAM_ANY | SNB_SNP_ANY | \
1839 SNB_NON_DRAM)
1840
1841 static __initconst const u64 knl_hw_cache_extra_regs
1842 [PERF_COUNT_HW_CACHE_MAX]
1843 [PERF_COUNT_HW_CACHE_OP_MAX]
1844 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1845 [C(LL)] = {
1846 [C(OP_READ)] = {
1847 [C(RESULT_ACCESS)] = KNL_L2_READ | KNL_L2_ACCESS,
1848 [C(RESULT_MISS)] = 0,
1849 },
1850 [C(OP_WRITE)] = {
1851 [C(RESULT_ACCESS)] = KNL_L2_WRITE | KNL_L2_ACCESS,
1852 [C(RESULT_MISS)] = KNL_L2_WRITE | KNL_L2_MISS,
1853 },
1854 [C(OP_PREFETCH)] = {
1855 [C(RESULT_ACCESS)] = KNL_L2_PREFETCH | KNL_L2_ACCESS,
1856 [C(RESULT_MISS)] = KNL_L2_PREFETCH | KNL_L2_MISS,
1857 },
1858 },
1859 };
1860
1861 /*
1862 * Used from PMIs where the LBRs are already disabled.
1863 *
1864 * This function could be called consecutively. It is required to remain in
1865 * disabled state if called consecutively.
1866 *
1867 * During consecutive calls, the same disable value will be written to related
1868 * registers, so the PMU state remains unchanged.
1869 *
1870 * intel_bts events don't coexist with intel PMU's BTS events because of
1871 * x86_add_exclusive(x86_lbr_exclusive_lbr); there's no need to keep them
1872 * disabled around intel PMU's event batching etc, only inside the PMI handler.
1873 */
1874 static void __intel_pmu_disable_all(void)
1875 {
1876 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1877
1878 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
1879
1880 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask))
1881 intel_pmu_disable_bts();
1882
1883 intel_pmu_pebs_disable_all();
1884 }
1885
1886 static void intel_pmu_disable_all(void)
1887 {
1888 __intel_pmu_disable_all();
1889 intel_pmu_lbr_disable_all();
1890 }
1891
1892 static void __intel_pmu_enable_all(int added, bool pmi)
1893 {
1894 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1895
1896 intel_pmu_pebs_enable_all();
1897 intel_pmu_lbr_enable_all(pmi);
1898 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL,
1899 x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask);
1900
1901 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
1902 struct perf_event *event =
1903 cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
1904
1905 if (WARN_ON_ONCE(!event))
1906 return;
1907
1908 intel_pmu_enable_bts(event->hw.config);
1909 }
1910 }
1911
1912 static void intel_pmu_enable_all(int added)
1913 {
1914 __intel_pmu_enable_all(added, false);
1915 }
1916
1917 /*
1918 * Workaround for:
1919 * Intel Errata AAK100 (model 26)
1920 * Intel Errata AAP53 (model 30)
1921 * Intel Errata BD53 (model 44)
1922 *
1923 * The official story:
1924 * These chips need to be 'reset' when adding counters by programming the
1925 * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either
1926 * in sequence on the same PMC or on different PMCs.
1927 *
1928 * In practise it appears some of these events do in fact count, and
1929 * we need to programm all 4 events.
1930 */
1931 static void intel_pmu_nhm_workaround(void)
1932 {
1933 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1934 static const unsigned long nhm_magic[4] = {
1935 0x4300B5,
1936 0x4300D2,
1937 0x4300B1,
1938 0x4300B1
1939 };
1940 struct perf_event *event;
1941 int i;
1942
1943 /*
1944 * The Errata requires below steps:
1945 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL;
1946 * 2) Configure 4 PERFEVTSELx with the magic events and clear
1947 * the corresponding PMCx;
1948 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL;
1949 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL;
1950 * 5) Clear 4 pairs of ERFEVTSELx and PMCx;
1951 */
1952
1953 /*
1954 * The real steps we choose are a little different from above.
1955 * A) To reduce MSR operations, we don't run step 1) as they
1956 * are already cleared before this function is called;
1957 * B) Call x86_perf_event_update to save PMCx before configuring
1958 * PERFEVTSELx with magic number;
1959 * C) With step 5), we do clear only when the PERFEVTSELx is
1960 * not used currently.
1961 * D) Call x86_perf_event_set_period to restore PMCx;
1962 */
1963
1964 /* We always operate 4 pairs of PERF Counters */
1965 for (i = 0; i < 4; i++) {
1966 event = cpuc->events[i];
1967 if (event)
1968 x86_perf_event_update(event);
1969 }
1970
1971 for (i = 0; i < 4; i++) {
1972 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]);
1973 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0);
1974 }
1975
1976 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf);
1977 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
1978
1979 for (i = 0; i < 4; i++) {
1980 event = cpuc->events[i];
1981
1982 if (event) {
1983 x86_perf_event_set_period(event);
1984 __x86_pmu_enable_event(&event->hw,
1985 ARCH_PERFMON_EVENTSEL_ENABLE);
1986 } else
1987 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0);
1988 }
1989 }
1990
1991 static void intel_pmu_nhm_enable_all(int added)
1992 {
1993 if (added)
1994 intel_pmu_nhm_workaround();
1995 intel_pmu_enable_all(added);
1996 }
1997
1998 static void intel_set_tfa(struct cpu_hw_events *cpuc, bool on)
1999 {
2000 u64 val = on ? MSR_TFA_RTM_FORCE_ABORT : 0;
2001
2002 if (cpuc->tfa_shadow != val) {
2003 cpuc->tfa_shadow = val;
2004 wrmsrl(MSR_TSX_FORCE_ABORT, val);
2005 }
2006 }
2007
2008 static void intel_tfa_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr)
2009 {
2010 /*
2011 * We're going to use PMC3, make sure TFA is set before we touch it.
2012 */
2013 if (cntr == 3 && !cpuc->is_fake)
2014 intel_set_tfa(cpuc, true);
2015 }
2016
2017 static void intel_tfa_pmu_enable_all(int added)
2018 {
2019 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2020
2021 /*
2022 * If we find PMC3 is no longer used when we enable the PMU, we can
2023 * clear TFA.
2024 */
2025 if (!test_bit(3, cpuc->active_mask))
2026 intel_set_tfa(cpuc, false);
2027
2028 intel_pmu_enable_all(added);
2029 }
2030
2031 static inline u64 intel_pmu_get_status(void)
2032 {
2033 u64 status;
2034
2035 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
2036
2037 return status;
2038 }
2039
2040 static inline void intel_pmu_ack_status(u64 ack)
2041 {
2042 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
2043 }
2044
2045 static void intel_pmu_disable_fixed(struct hw_perf_event *hwc)
2046 {
2047 int idx = hwc->idx - INTEL_PMC_IDX_FIXED;
2048 u64 ctrl_val, mask;
2049
2050 mask = 0xfULL << (idx * 4);
2051
2052 rdmsrl(hwc->config_base, ctrl_val);
2053 ctrl_val &= ~mask;
2054 wrmsrl(hwc->config_base, ctrl_val);
2055 }
2056
2057 static inline bool event_is_checkpointed(struct perf_event *event)
2058 {
2059 return (event->hw.config & HSW_IN_TX_CHECKPOINTED) != 0;
2060 }
2061
2062 static void intel_pmu_disable_event(struct perf_event *event)
2063 {
2064 struct hw_perf_event *hwc = &event->hw;
2065 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2066
2067 if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) {
2068 intel_pmu_disable_bts();
2069 intel_pmu_drain_bts_buffer();
2070 return;
2071 }
2072
2073 cpuc->intel_ctrl_guest_mask &= ~(1ull << hwc->idx);
2074 cpuc->intel_ctrl_host_mask &= ~(1ull << hwc->idx);
2075 cpuc->intel_cp_status &= ~(1ull << hwc->idx);
2076
2077 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
2078 intel_pmu_disable_fixed(hwc);
2079 return;
2080 }
2081
2082 x86_pmu_disable_event(event);
2083
2084 if (unlikely(event->attr.precise_ip))
2085 intel_pmu_pebs_disable(event);
2086 }
2087
2088 static void intel_pmu_del_event(struct perf_event *event)
2089 {
2090 if (needs_branch_stack(event))
2091 intel_pmu_lbr_del(event);
2092 if (event->attr.precise_ip)
2093 intel_pmu_pebs_del(event);
2094 }
2095
2096 static void intel_pmu_enable_fixed(struct hw_perf_event *hwc)
2097 {
2098 int idx = hwc->idx - INTEL_PMC_IDX_FIXED;
2099 u64 ctrl_val, bits, mask;
2100
2101 /*
2102 * Enable IRQ generation (0x8),
2103 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
2104 * if requested:
2105 */
2106 bits = 0x8ULL;
2107 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
2108 bits |= 0x2;
2109 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
2110 bits |= 0x1;
2111
2112 /*
2113 * ANY bit is supported in v3 and up
2114 */
2115 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
2116 bits |= 0x4;
2117
2118 bits <<= (idx * 4);
2119 mask = 0xfULL << (idx * 4);
2120
2121 rdmsrl(hwc->config_base, ctrl_val);
2122 ctrl_val &= ~mask;
2123 ctrl_val |= bits;
2124 wrmsrl(hwc->config_base, ctrl_val);
2125 }
2126
2127 static void intel_pmu_enable_event(struct perf_event *event)
2128 {
2129 struct hw_perf_event *hwc = &event->hw;
2130 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2131
2132 if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) {
2133 if (!__this_cpu_read(cpu_hw_events.enabled))
2134 return;
2135
2136 intel_pmu_enable_bts(hwc->config);
2137 return;
2138 }
2139
2140 if (event->attr.exclude_host)
2141 cpuc->intel_ctrl_guest_mask |= (1ull << hwc->idx);
2142 if (event->attr.exclude_guest)
2143 cpuc->intel_ctrl_host_mask |= (1ull << hwc->idx);
2144
2145 if (unlikely(event_is_checkpointed(event)))
2146 cpuc->intel_cp_status |= (1ull << hwc->idx);
2147
2148 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
2149 intel_pmu_enable_fixed(hwc);
2150 return;
2151 }
2152
2153 if (unlikely(event->attr.precise_ip))
2154 intel_pmu_pebs_enable(event);
2155
2156 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
2157 }
2158
2159 static void intel_pmu_add_event(struct perf_event *event)
2160 {
2161 if (event->attr.precise_ip)
2162 intel_pmu_pebs_add(event);
2163 if (needs_branch_stack(event))
2164 intel_pmu_lbr_add(event);
2165 }
2166
2167 /*
2168 * Save and restart an expired event. Called by NMI contexts,
2169 * so it has to be careful about preempting normal event ops:
2170 */
2171 int intel_pmu_save_and_restart(struct perf_event *event)
2172 {
2173 x86_perf_event_update(event);
2174 /*
2175 * For a checkpointed counter always reset back to 0. This
2176 * avoids a situation where the counter overflows, aborts the
2177 * transaction and is then set back to shortly before the
2178 * overflow, and overflows and aborts again.
2179 */
2180 if (unlikely(event_is_checkpointed(event))) {
2181 /* No race with NMIs because the counter should not be armed */
2182 wrmsrl(event->hw.event_base, 0);
2183 local64_set(&event->hw.prev_count, 0);
2184 }
2185 return x86_perf_event_set_period(event);
2186 }
2187
2188 static void intel_pmu_reset(void)
2189 {
2190 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
2191 unsigned long flags;
2192 int idx;
2193
2194 if (!x86_pmu.num_counters)
2195 return;
2196
2197 local_irq_save(flags);
2198
2199 pr_info("clearing PMU state on CPU#%d\n", smp_processor_id());
2200
2201 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
2202 wrmsrl_safe(x86_pmu_config_addr(idx), 0ull);
2203 wrmsrl_safe(x86_pmu_event_addr(idx), 0ull);
2204 }
2205 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++)
2206 wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
2207
2208 if (ds)
2209 ds->bts_index = ds->bts_buffer_base;
2210
2211 /* Ack all overflows and disable fixed counters */
2212 if (x86_pmu.version >= 2) {
2213 intel_pmu_ack_status(intel_pmu_get_status());
2214 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
2215 }
2216
2217 /* Reset LBRs and LBR freezing */
2218 if (x86_pmu.lbr_nr) {
2219 update_debugctlmsr(get_debugctlmsr() &
2220 ~(DEBUGCTLMSR_FREEZE_LBRS_ON_PMI|DEBUGCTLMSR_LBR));
2221 }
2222
2223 local_irq_restore(flags);
2224 }
2225
2226 /*
2227 * This handler is triggered by the local APIC, so the APIC IRQ handling
2228 * rules apply:
2229 */
2230 static int intel_pmu_handle_irq(struct pt_regs *regs)
2231 {
2232 struct perf_sample_data data;
2233 struct cpu_hw_events *cpuc;
2234 int bit, loops;
2235 u64 status;
2236 int handled;
2237 int pmu_enabled;
2238
2239 cpuc = this_cpu_ptr(&cpu_hw_events);
2240
2241 /*
2242 * Save the PMU state.
2243 * It needs to be restored when leaving the handler.
2244 */
2245 pmu_enabled = cpuc->enabled;
2246 /*
2247 * No known reason to not always do late ACK,
2248 * but just in case do it opt-in.
2249 */
2250 if (!x86_pmu.late_ack)
2251 apic_write(APIC_LVTPC, APIC_DM_NMI);
2252 intel_bts_disable_local();
2253 cpuc->enabled = 0;
2254 __intel_pmu_disable_all();
2255 handled = intel_pmu_drain_bts_buffer();
2256 handled += intel_bts_interrupt();
2257 status = intel_pmu_get_status();
2258 if (!status)
2259 goto done;
2260
2261 loops = 0;
2262 again:
2263 intel_pmu_lbr_read();
2264 intel_pmu_ack_status(status);
2265 if (++loops > 100) {
2266 static bool warned = false;
2267 if (!warned) {
2268 WARN(1, "perfevents: irq loop stuck!\n");
2269 perf_event_print_debug();
2270 warned = true;
2271 }
2272 intel_pmu_reset();
2273 goto done;
2274 }
2275
2276 inc_irq_stat(apic_perf_irqs);
2277
2278
2279 /*
2280 * Ignore a range of extra bits in status that do not indicate
2281 * overflow by themselves.
2282 */
2283 status &= ~(GLOBAL_STATUS_COND_CHG |
2284 GLOBAL_STATUS_ASIF |
2285 GLOBAL_STATUS_LBRS_FROZEN);
2286 if (!status)
2287 goto done;
2288 /*
2289 * In case multiple PEBS events are sampled at the same time,
2290 * it is possible to have GLOBAL_STATUS bit 62 set indicating
2291 * PEBS buffer overflow and also seeing at most 3 PEBS counters
2292 * having their bits set in the status register. This is a sign
2293 * that there was at least one PEBS record pending at the time
2294 * of the PMU interrupt. PEBS counters must only be processed
2295 * via the drain_pebs() calls and not via the regular sample
2296 * processing loop coming after that the function, otherwise
2297 * phony regular samples may be generated in the sampling buffer
2298 * not marked with the EXACT tag. Another possibility is to have
2299 * one PEBS event and at least one non-PEBS event whic hoverflows
2300 * while PEBS has armed. In this case, bit 62 of GLOBAL_STATUS will
2301 * not be set, yet the overflow status bit for the PEBS counter will
2302 * be on Skylake.
2303 *
2304 * To avoid this problem, we systematically ignore the PEBS-enabled
2305 * counters from the GLOBAL_STATUS mask and we always process PEBS
2306 * events via drain_pebs().
2307 */
2308 status &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK);
2309
2310 /*
2311 * PEBS overflow sets bit 62 in the global status register
2312 */
2313 if (__test_and_clear_bit(62, (unsigned long *)&status)) {
2314 handled++;
2315 x86_pmu.drain_pebs(regs);
2316 status &= x86_pmu.intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI;
2317 }
2318
2319 /*
2320 * Intel PT
2321 */
2322 if (__test_and_clear_bit(55, (unsigned long *)&status)) {
2323 handled++;
2324 intel_pt_interrupt();
2325 }
2326
2327 /*
2328 * Checkpointed counters can lead to 'spurious' PMIs because the
2329 * rollback caused by the PMI will have cleared the overflow status
2330 * bit. Therefore always force probe these counters.
2331 */
2332 status |= cpuc->intel_cp_status;
2333
2334 for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
2335 struct perf_event *event = cpuc->events[bit];
2336
2337 handled++;
2338
2339 if (!test_bit(bit, cpuc->active_mask))
2340 continue;
2341
2342 if (!intel_pmu_save_and_restart(event))
2343 continue;
2344
2345 perf_sample_data_init(&data, 0, event->hw.last_period);
2346
2347 if (has_branch_stack(event))
2348 data.br_stack = &cpuc->lbr_stack;
2349
2350 if (perf_event_overflow(event, &data, regs))
2351 x86_pmu_stop(event, 0);
2352 }
2353
2354 /*
2355 * Repeat if there is more work to be done:
2356 */
2357 status = intel_pmu_get_status();
2358 if (status)
2359 goto again;
2360
2361 done:
2362 /* Only restore PMU state when it's active. See x86_pmu_disable(). */
2363 cpuc->enabled = pmu_enabled;
2364 if (pmu_enabled)
2365 __intel_pmu_enable_all(0, true);
2366 intel_bts_enable_local();
2367
2368 /*
2369 * Only unmask the NMI after the overflow counters
2370 * have been reset. This avoids spurious NMIs on
2371 * Haswell CPUs.
2372 */
2373 if (x86_pmu.late_ack)
2374 apic_write(APIC_LVTPC, APIC_DM_NMI);
2375 return handled;
2376 }
2377
2378 static struct event_constraint *
2379 intel_bts_constraints(struct perf_event *event)
2380 {
2381 if (unlikely(intel_pmu_has_bts(event)))
2382 return &bts_constraint;
2383
2384 return NULL;
2385 }
2386
2387 static int intel_alt_er(int idx, u64 config)
2388 {
2389 int alt_idx = idx;
2390
2391 if (!(x86_pmu.flags & PMU_FL_HAS_RSP_1))
2392 return idx;
2393
2394 if (idx == EXTRA_REG_RSP_0)
2395 alt_idx = EXTRA_REG_RSP_1;
2396
2397 if (idx == EXTRA_REG_RSP_1)
2398 alt_idx = EXTRA_REG_RSP_0;
2399
2400 if (config & ~x86_pmu.extra_regs[alt_idx].valid_mask)
2401 return idx;
2402
2403 return alt_idx;
2404 }
2405
2406 static void intel_fixup_er(struct perf_event *event, int idx)
2407 {
2408 event->hw.extra_reg.idx = idx;
2409
2410 if (idx == EXTRA_REG_RSP_0) {
2411 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
2412 event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_0].event;
2413 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0;
2414 } else if (idx == EXTRA_REG_RSP_1) {
2415 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
2416 event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_1].event;
2417 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1;
2418 }
2419 }
2420
2421 /*
2422 * manage allocation of shared extra msr for certain events
2423 *
2424 * sharing can be:
2425 * per-cpu: to be shared between the various events on a single PMU
2426 * per-core: per-cpu + shared by HT threads
2427 */
2428 static struct event_constraint *
2429 __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc,
2430 struct perf_event *event,
2431 struct hw_perf_event_extra *reg)
2432 {
2433 struct event_constraint *c = &emptyconstraint;
2434 struct er_account *era;
2435 unsigned long flags;
2436 int idx = reg->idx;
2437
2438 /*
2439 * reg->alloc can be set due to existing state, so for fake cpuc we
2440 * need to ignore this, otherwise we might fail to allocate proper fake
2441 * state for this extra reg constraint. Also see the comment below.
2442 */
2443 if (reg->alloc && !cpuc->is_fake)
2444 return NULL; /* call x86_get_event_constraint() */
2445
2446 again:
2447 era = &cpuc->shared_regs->regs[idx];
2448 /*
2449 * we use spin_lock_irqsave() to avoid lockdep issues when
2450 * passing a fake cpuc
2451 */
2452 raw_spin_lock_irqsave(&era->lock, flags);
2453
2454 if (!atomic_read(&era->ref) || era->config == reg->config) {
2455
2456 /*
2457 * If its a fake cpuc -- as per validate_{group,event}() we
2458 * shouldn't touch event state and we can avoid doing so
2459 * since both will only call get_event_constraints() once
2460 * on each event, this avoids the need for reg->alloc.
2461 *
2462 * Not doing the ER fixup will only result in era->reg being
2463 * wrong, but since we won't actually try and program hardware
2464 * this isn't a problem either.
2465 */
2466 if (!cpuc->is_fake) {
2467 if (idx != reg->idx)
2468 intel_fixup_er(event, idx);
2469
2470 /*
2471 * x86_schedule_events() can call get_event_constraints()
2472 * multiple times on events in the case of incremental
2473 * scheduling(). reg->alloc ensures we only do the ER
2474 * allocation once.
2475 */
2476 reg->alloc = 1;
2477 }
2478
2479 /* lock in msr value */
2480 era->config = reg->config;
2481 era->reg = reg->reg;
2482
2483 /* one more user */
2484 atomic_inc(&era->ref);
2485
2486 /*
2487 * need to call x86_get_event_constraint()
2488 * to check if associated event has constraints
2489 */
2490 c = NULL;
2491 } else {
2492 idx = intel_alt_er(idx, reg->config);
2493 if (idx != reg->idx) {
2494 raw_spin_unlock_irqrestore(&era->lock, flags);
2495 goto again;
2496 }
2497 }
2498 raw_spin_unlock_irqrestore(&era->lock, flags);
2499
2500 return c;
2501 }
2502
2503 static void
2504 __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc,
2505 struct hw_perf_event_extra *reg)
2506 {
2507 struct er_account *era;
2508
2509 /*
2510 * Only put constraint if extra reg was actually allocated. Also takes
2511 * care of event which do not use an extra shared reg.
2512 *
2513 * Also, if this is a fake cpuc we shouldn't touch any event state
2514 * (reg->alloc) and we don't care about leaving inconsistent cpuc state
2515 * either since it'll be thrown out.
2516 */
2517 if (!reg->alloc || cpuc->is_fake)
2518 return;
2519
2520 era = &cpuc->shared_regs->regs[reg->idx];
2521
2522 /* one fewer user */
2523 atomic_dec(&era->ref);
2524
2525 /* allocate again next time */
2526 reg->alloc = 0;
2527 }
2528
2529 static struct event_constraint *
2530 intel_shared_regs_constraints(struct cpu_hw_events *cpuc,
2531 struct perf_event *event)
2532 {
2533 struct event_constraint *c = NULL, *d;
2534 struct hw_perf_event_extra *xreg, *breg;
2535
2536 xreg = &event->hw.extra_reg;
2537 if (xreg->idx != EXTRA_REG_NONE) {
2538 c = __intel_shared_reg_get_constraints(cpuc, event, xreg);
2539 if (c == &emptyconstraint)
2540 return c;
2541 }
2542 breg = &event->hw.branch_reg;
2543 if (breg->idx != EXTRA_REG_NONE) {
2544 d = __intel_shared_reg_get_constraints(cpuc, event, breg);
2545 if (d == &emptyconstraint) {
2546 __intel_shared_reg_put_constraints(cpuc, xreg);
2547 c = d;
2548 }
2549 }
2550 return c;
2551 }
2552
2553 struct event_constraint *
2554 x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
2555 struct perf_event *event)
2556 {
2557 struct event_constraint *c;
2558
2559 if (x86_pmu.event_constraints) {
2560 for_each_event_constraint(c, x86_pmu.event_constraints) {
2561 if ((event->hw.config & c->cmask) == c->code) {
2562 event->hw.flags |= c->flags;
2563 return c;
2564 }
2565 }
2566 }
2567
2568 return &unconstrained;
2569 }
2570
2571 static struct event_constraint *
2572 __intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
2573 struct perf_event *event)
2574 {
2575 struct event_constraint *c;
2576
2577 c = intel_bts_constraints(event);
2578 if (c)
2579 return c;
2580
2581 c = intel_shared_regs_constraints(cpuc, event);
2582 if (c)
2583 return c;
2584
2585 c = intel_pebs_constraints(event);
2586 if (c)
2587 return c;
2588
2589 return x86_get_event_constraints(cpuc, idx, event);
2590 }
2591
2592 static void
2593 intel_start_scheduling(struct cpu_hw_events *cpuc)
2594 {
2595 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
2596 struct intel_excl_states *xl;
2597 int tid = cpuc->excl_thread_id;
2598
2599 /*
2600 * nothing needed if in group validation mode
2601 */
2602 if (cpuc->is_fake || !is_ht_workaround_enabled())
2603 return;
2604
2605 /*
2606 * no exclusion needed
2607 */
2608 if (WARN_ON_ONCE(!excl_cntrs))
2609 return;
2610
2611 xl = &excl_cntrs->states[tid];
2612
2613 xl->sched_started = true;
2614 /*
2615 * lock shared state until we are done scheduling
2616 * in stop_event_scheduling()
2617 * makes scheduling appear as a transaction
2618 */
2619 raw_spin_lock(&excl_cntrs->lock);
2620 }
2621
2622 static void intel_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr)
2623 {
2624 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
2625 struct event_constraint *c = cpuc->event_constraint[idx];
2626 struct intel_excl_states *xl;
2627 int tid = cpuc->excl_thread_id;
2628
2629 if (cpuc->is_fake || !is_ht_workaround_enabled())
2630 return;
2631
2632 if (WARN_ON_ONCE(!excl_cntrs))
2633 return;
2634
2635 if (!(c->flags & PERF_X86_EVENT_DYNAMIC))
2636 return;
2637
2638 xl = &excl_cntrs->states[tid];
2639
2640 lockdep_assert_held(&excl_cntrs->lock);
2641
2642 if (c->flags & PERF_X86_EVENT_EXCL)
2643 xl->state[cntr] = INTEL_EXCL_EXCLUSIVE;
2644 else
2645 xl->state[cntr] = INTEL_EXCL_SHARED;
2646 }
2647
2648 static void
2649 intel_stop_scheduling(struct cpu_hw_events *cpuc)
2650 {
2651 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
2652 struct intel_excl_states *xl;
2653 int tid = cpuc->excl_thread_id;
2654
2655 /*
2656 * nothing needed if in group validation mode
2657 */
2658 if (cpuc->is_fake || !is_ht_workaround_enabled())
2659 return;
2660 /*
2661 * no exclusion needed
2662 */
2663 if (WARN_ON_ONCE(!excl_cntrs))
2664 return;
2665
2666 xl = &excl_cntrs->states[tid];
2667
2668 xl->sched_started = false;
2669 /*
2670 * release shared state lock (acquired in intel_start_scheduling())
2671 */
2672 raw_spin_unlock(&excl_cntrs->lock);
2673 }
2674
2675 static struct event_constraint *
2676 dyn_constraint(struct cpu_hw_events *cpuc, struct event_constraint *c, int idx)
2677 {
2678 WARN_ON_ONCE(!cpuc->constraint_list);
2679
2680 if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) {
2681 struct event_constraint *cx;
2682
2683 /*
2684 * grab pre-allocated constraint entry
2685 */
2686 cx = &cpuc->constraint_list[idx];
2687
2688 /*
2689 * initialize dynamic constraint
2690 * with static constraint
2691 */
2692 *cx = *c;
2693
2694 /*
2695 * mark constraint as dynamic
2696 */
2697 cx->flags |= PERF_X86_EVENT_DYNAMIC;
2698 c = cx;
2699 }
2700
2701 return c;
2702 }
2703
2704 static struct event_constraint *
2705 intel_get_excl_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
2706 int idx, struct event_constraint *c)
2707 {
2708 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
2709 struct intel_excl_states *xlo;
2710 int tid = cpuc->excl_thread_id;
2711 int is_excl, i;
2712
2713 /*
2714 * validating a group does not require
2715 * enforcing cross-thread exclusion
2716 */
2717 if (cpuc->is_fake || !is_ht_workaround_enabled())
2718 return c;
2719
2720 /*
2721 * no exclusion needed
2722 */
2723 if (WARN_ON_ONCE(!excl_cntrs))
2724 return c;
2725
2726 /*
2727 * because we modify the constraint, we need
2728 * to make a copy. Static constraints come
2729 * from static const tables.
2730 *
2731 * only needed when constraint has not yet
2732 * been cloned (marked dynamic)
2733 */
2734 c = dyn_constraint(cpuc, c, idx);
2735
2736 /*
2737 * From here on, the constraint is dynamic.
2738 * Either it was just allocated above, or it
2739 * was allocated during a earlier invocation
2740 * of this function
2741 */
2742
2743 /*
2744 * state of sibling HT
2745 */
2746 xlo = &excl_cntrs->states[tid ^ 1];
2747
2748 /*
2749 * event requires exclusive counter access
2750 * across HT threads
2751 */
2752 is_excl = c->flags & PERF_X86_EVENT_EXCL;
2753 if (is_excl && !(event->hw.flags & PERF_X86_EVENT_EXCL_ACCT)) {
2754 event->hw.flags |= PERF_X86_EVENT_EXCL_ACCT;
2755 if (!cpuc->n_excl++)
2756 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 1);
2757 }
2758
2759 /*
2760 * Modify static constraint with current dynamic
2761 * state of thread
2762 *
2763 * EXCLUSIVE: sibling counter measuring exclusive event
2764 * SHARED : sibling counter measuring non-exclusive event
2765 * UNUSED : sibling counter unused
2766 */
2767 for_each_set_bit(i, c->idxmsk, X86_PMC_IDX_MAX) {
2768 /*
2769 * exclusive event in sibling counter
2770 * our corresponding counter cannot be used
2771 * regardless of our event
2772 */
2773 if (xlo->state[i] == INTEL_EXCL_EXCLUSIVE)
2774 __clear_bit(i, c->idxmsk);
2775 /*
2776 * if measuring an exclusive event, sibling
2777 * measuring non-exclusive, then counter cannot
2778 * be used
2779 */
2780 if (is_excl && xlo->state[i] == INTEL_EXCL_SHARED)
2781 __clear_bit(i, c->idxmsk);
2782 }
2783
2784 /*
2785 * recompute actual bit weight for scheduling algorithm
2786 */
2787 c->weight = hweight64(c->idxmsk64);
2788
2789 /*
2790 * if we return an empty mask, then switch
2791 * back to static empty constraint to avoid
2792 * the cost of freeing later on
2793 */
2794 if (c->weight == 0)
2795 c = &emptyconstraint;
2796
2797 return c;
2798 }
2799
2800 static struct event_constraint *
2801 intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
2802 struct perf_event *event)
2803 {
2804 struct event_constraint *c1 = NULL;
2805 struct event_constraint *c2;
2806
2807 if (idx >= 0) /* fake does < 0 */
2808 c1 = cpuc->event_constraint[idx];
2809
2810 /*
2811 * first time only
2812 * - static constraint: no change across incremental scheduling calls
2813 * - dynamic constraint: handled by intel_get_excl_constraints()
2814 */
2815 c2 = __intel_get_event_constraints(cpuc, idx, event);
2816 if (c1 && (c1->flags & PERF_X86_EVENT_DYNAMIC)) {
2817 bitmap_copy(c1->idxmsk, c2->idxmsk, X86_PMC_IDX_MAX);
2818 c1->weight = c2->weight;
2819 c2 = c1;
2820 }
2821
2822 if (cpuc->excl_cntrs)
2823 return intel_get_excl_constraints(cpuc, event, idx, c2);
2824
2825 return c2;
2826 }
2827
2828 static void intel_put_excl_constraints(struct cpu_hw_events *cpuc,
2829 struct perf_event *event)
2830 {
2831 struct hw_perf_event *hwc = &event->hw;
2832 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
2833 int tid = cpuc->excl_thread_id;
2834 struct intel_excl_states *xl;
2835
2836 /*
2837 * nothing needed if in group validation mode
2838 */
2839 if (cpuc->is_fake)
2840 return;
2841
2842 if (WARN_ON_ONCE(!excl_cntrs))
2843 return;
2844
2845 if (hwc->flags & PERF_X86_EVENT_EXCL_ACCT) {
2846 hwc->flags &= ~PERF_X86_EVENT_EXCL_ACCT;
2847 if (!--cpuc->n_excl)
2848 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 0);
2849 }
2850
2851 /*
2852 * If event was actually assigned, then mark the counter state as
2853 * unused now.
2854 */
2855 if (hwc->idx >= 0) {
2856 xl = &excl_cntrs->states[tid];
2857
2858 /*
2859 * put_constraint may be called from x86_schedule_events()
2860 * which already has the lock held so here make locking
2861 * conditional.
2862 */
2863 if (!xl->sched_started)
2864 raw_spin_lock(&excl_cntrs->lock);
2865
2866 xl->state[hwc->idx] = INTEL_EXCL_UNUSED;
2867
2868 if (!xl->sched_started)
2869 raw_spin_unlock(&excl_cntrs->lock);
2870 }
2871 }
2872
2873 static void
2874 intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc,
2875 struct perf_event *event)
2876 {
2877 struct hw_perf_event_extra *reg;
2878
2879 reg = &event->hw.extra_reg;
2880 if (reg->idx != EXTRA_REG_NONE)
2881 __intel_shared_reg_put_constraints(cpuc, reg);
2882
2883 reg = &event->hw.branch_reg;
2884 if (reg->idx != EXTRA_REG_NONE)
2885 __intel_shared_reg_put_constraints(cpuc, reg);
2886 }
2887
2888 static void intel_put_event_constraints(struct cpu_hw_events *cpuc,
2889 struct perf_event *event)
2890 {
2891 intel_put_shared_regs_event_constraints(cpuc, event);
2892
2893 /*
2894 * is PMU has exclusive counter restrictions, then
2895 * all events are subject to and must call the
2896 * put_excl_constraints() routine
2897 */
2898 if (cpuc->excl_cntrs)
2899 intel_put_excl_constraints(cpuc, event);
2900 }
2901
2902 static void intel_pebs_aliases_core2(struct perf_event *event)
2903 {
2904 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
2905 /*
2906 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
2907 * (0x003c) so that we can use it with PEBS.
2908 *
2909 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
2910 * PEBS capable. However we can use INST_RETIRED.ANY_P
2911 * (0x00c0), which is a PEBS capable event, to get the same
2912 * count.
2913 *
2914 * INST_RETIRED.ANY_P counts the number of cycles that retires
2915 * CNTMASK instructions. By setting CNTMASK to a value (16)
2916 * larger than the maximum number of instructions that can be
2917 * retired per cycle (4) and then inverting the condition, we
2918 * count all cycles that retire 16 or less instructions, which
2919 * is every cycle.
2920 *
2921 * Thereby we gain a PEBS capable cycle counter.
2922 */
2923 u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16);
2924
2925 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
2926 event->hw.config = alt_config;
2927 }
2928 }
2929
2930 static void intel_pebs_aliases_snb(struct perf_event *event)
2931 {
2932 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
2933 /*
2934 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
2935 * (0x003c) so that we can use it with PEBS.
2936 *
2937 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
2938 * PEBS capable. However we can use UOPS_RETIRED.ALL
2939 * (0x01c2), which is a PEBS capable event, to get the same
2940 * count.
2941 *
2942 * UOPS_RETIRED.ALL counts the number of cycles that retires
2943 * CNTMASK micro-ops. By setting CNTMASK to a value (16)
2944 * larger than the maximum number of micro-ops that can be
2945 * retired per cycle (4) and then inverting the condition, we
2946 * count all cycles that retire 16 or less micro-ops, which
2947 * is every cycle.
2948 *
2949 * Thereby we gain a PEBS capable cycle counter.
2950 */
2951 u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16);
2952
2953 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
2954 event->hw.config = alt_config;
2955 }
2956 }
2957
2958 static void intel_pebs_aliases_precdist(struct perf_event *event)
2959 {
2960 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
2961 /*
2962 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
2963 * (0x003c) so that we can use it with PEBS.
2964 *
2965 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
2966 * PEBS capable. However we can use INST_RETIRED.PREC_DIST
2967 * (0x01c0), which is a PEBS capable event, to get the same
2968 * count.
2969 *
2970 * The PREC_DIST event has special support to minimize sample
2971 * shadowing effects. One drawback is that it can be
2972 * only programmed on counter 1, but that seems like an
2973 * acceptable trade off.
2974 */
2975 u64 alt_config = X86_CONFIG(.event=0xc0, .umask=0x01, .inv=1, .cmask=16);
2976
2977 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
2978 event->hw.config = alt_config;
2979 }
2980 }
2981
2982 static void intel_pebs_aliases_ivb(struct perf_event *event)
2983 {
2984 if (event->attr.precise_ip < 3)
2985 return intel_pebs_aliases_snb(event);
2986 return intel_pebs_aliases_precdist(event);
2987 }
2988
2989 static void intel_pebs_aliases_skl(struct perf_event *event)
2990 {
2991 if (event->attr.precise_ip < 3)
2992 return intel_pebs_aliases_core2(event);
2993 return intel_pebs_aliases_precdist(event);
2994 }
2995
2996 static unsigned long intel_pmu_free_running_flags(struct perf_event *event)
2997 {
2998 unsigned long flags = x86_pmu.free_running_flags;
2999
3000 if (event->attr.use_clockid)
3001 flags &= ~PERF_SAMPLE_TIME;
3002 if (!event->attr.exclude_kernel)
3003 flags &= ~PERF_SAMPLE_REGS_USER;
3004 if (event->attr.sample_regs_user & ~PEBS_GP_REGS)
3005 flags &= ~(PERF_SAMPLE_REGS_USER | PERF_SAMPLE_REGS_INTR);
3006 return flags;
3007 }
3008
3009 static int intel_pmu_bts_config(struct perf_event *event)
3010 {
3011 struct perf_event_attr *attr = &event->attr;
3012
3013 if (unlikely(intel_pmu_has_bts(event))) {
3014 /* BTS is not supported by this architecture. */
3015 if (!x86_pmu.bts_active)
3016 return -EOPNOTSUPP;
3017
3018 /* BTS is currently only allowed for user-mode. */
3019 if (!attr->exclude_kernel)
3020 return -EOPNOTSUPP;
3021
3022 /* disallow bts if conflicting events are present */
3023 if (x86_add_exclusive(x86_lbr_exclusive_lbr))
3024 return -EBUSY;
3025
3026 event->destroy = hw_perf_lbr_event_destroy;
3027 }
3028
3029 return 0;
3030 }
3031
3032 static int core_pmu_hw_config(struct perf_event *event)
3033 {
3034 int ret = x86_pmu_hw_config(event);
3035
3036 if (ret)
3037 return ret;
3038
3039 return intel_pmu_bts_config(event);
3040 }
3041
3042 static int intel_pmu_hw_config(struct perf_event *event)
3043 {
3044 int ret = x86_pmu_hw_config(event);
3045
3046 if (ret)
3047 return ret;
3048
3049 ret = intel_pmu_bts_config(event);
3050 if (ret)
3051 return ret;
3052
3053 if (event->attr.precise_ip) {
3054 if (!(event->attr.freq || event->attr.wakeup_events)) {
3055 event->hw.flags |= PERF_X86_EVENT_AUTO_RELOAD;
3056 if (!(event->attr.sample_type &
3057 ~intel_pmu_free_running_flags(event)))
3058 event->hw.flags |= PERF_X86_EVENT_FREERUNNING;
3059 }
3060 if (x86_pmu.pebs_aliases)
3061 x86_pmu.pebs_aliases(event);
3062 }
3063
3064 if (needs_branch_stack(event)) {
3065 ret = intel_pmu_setup_lbr_filter(event);
3066 if (ret)
3067 return ret;
3068
3069 /*
3070 * BTS is set up earlier in this path, so don't account twice
3071 */
3072 if (!unlikely(intel_pmu_has_bts(event))) {
3073 /* disallow lbr if conflicting events are present */
3074 if (x86_add_exclusive(x86_lbr_exclusive_lbr))
3075 return -EBUSY;
3076
3077 event->destroy = hw_perf_lbr_event_destroy;
3078 }
3079 }
3080
3081 if (event->attr.type != PERF_TYPE_RAW)
3082 return 0;
3083
3084 if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY))
3085 return 0;
3086
3087 if (x86_pmu.version < 3)
3088 return -EINVAL;
3089
3090 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3091 return -EACCES;
3092
3093 event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY;
3094
3095 return 0;
3096 }
3097
3098 struct perf_guest_switch_msr *perf_guest_get_msrs(int *nr)
3099 {
3100 if (x86_pmu.guest_get_msrs)
3101 return x86_pmu.guest_get_msrs(nr);
3102 *nr = 0;
3103 return NULL;
3104 }
3105 EXPORT_SYMBOL_GPL(perf_guest_get_msrs);
3106
3107 static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr)
3108 {
3109 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3110 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
3111
3112 arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL;
3113 arr[0].host = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask;
3114 arr[0].guest = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_host_mask;
3115 /*
3116 * If PMU counter has PEBS enabled it is not enough to disable counter
3117 * on a guest entry since PEBS memory write can overshoot guest entry
3118 * and corrupt guest memory. Disabling PEBS solves the problem.
3119 */
3120 arr[1].msr = MSR_IA32_PEBS_ENABLE;
3121 arr[1].host = cpuc->pebs_enabled;
3122 arr[1].guest = 0;
3123
3124 *nr = 2;
3125 return arr;
3126 }
3127
3128 static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr)
3129 {
3130 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3131 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
3132 int idx;
3133
3134 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
3135 struct perf_event *event = cpuc->events[idx];
3136
3137 arr[idx].msr = x86_pmu_config_addr(idx);
3138 arr[idx].host = arr[idx].guest = 0;
3139
3140 if (!test_bit(idx, cpuc->active_mask))
3141 continue;
3142
3143 arr[idx].host = arr[idx].guest =
3144 event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE;
3145
3146 if (event->attr.exclude_host)
3147 arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
3148 else if (event->attr.exclude_guest)
3149 arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
3150 }
3151
3152 *nr = x86_pmu.num_counters;
3153 return arr;
3154 }
3155
3156 static void core_pmu_enable_event(struct perf_event *event)
3157 {
3158 if (!event->attr.exclude_host)
3159 x86_pmu_enable_event(event);
3160 }
3161
3162 static void core_pmu_enable_all(int added)
3163 {
3164 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3165 int idx;
3166
3167 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
3168 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
3169
3170 if (!test_bit(idx, cpuc->active_mask) ||
3171 cpuc->events[idx]->attr.exclude_host)
3172 continue;
3173
3174 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
3175 }
3176 }
3177
3178 static int hsw_hw_config(struct perf_event *event)
3179 {
3180 int ret = intel_pmu_hw_config(event);
3181
3182 if (ret)
3183 return ret;
3184 if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE))
3185 return 0;
3186 event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED);
3187
3188 /*
3189 * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with
3190 * PEBS or in ANY thread mode. Since the results are non-sensical forbid
3191 * this combination.
3192 */
3193 if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) &&
3194 ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) ||
3195 event->attr.precise_ip > 0))
3196 return -EOPNOTSUPP;
3197
3198 if (event_is_checkpointed(event)) {
3199 /*
3200 * Sampling of checkpointed events can cause situations where
3201 * the CPU constantly aborts because of a overflow, which is
3202 * then checkpointed back and ignored. Forbid checkpointing
3203 * for sampling.
3204 *
3205 * But still allow a long sampling period, so that perf stat
3206 * from KVM works.
3207 */
3208 if (event->attr.sample_period > 0 &&
3209 event->attr.sample_period < 0x7fffffff)
3210 return -EOPNOTSUPP;
3211 }
3212 return 0;
3213 }
3214
3215 static struct event_constraint counter0_constraint =
3216 INTEL_ALL_EVENT_CONSTRAINT(0, 0x1);
3217
3218 static struct event_constraint counter2_constraint =
3219 EVENT_CONSTRAINT(0, 0x4, 0);
3220
3221 static struct event_constraint *
3222 hsw_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3223 struct perf_event *event)
3224 {
3225 struct event_constraint *c;
3226
3227 c = intel_get_event_constraints(cpuc, idx, event);
3228
3229 /* Handle special quirk on in_tx_checkpointed only in counter 2 */
3230 if (event->hw.config & HSW_IN_TX_CHECKPOINTED) {
3231 if (c->idxmsk64 & (1U << 2))
3232 return &counter2_constraint;
3233 return &emptyconstraint;
3234 }
3235
3236 return c;
3237 }
3238
3239 static struct event_constraint *
3240 glp_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3241 struct perf_event *event)
3242 {
3243 struct event_constraint *c;
3244
3245 /* :ppp means to do reduced skid PEBS which is PMC0 only. */
3246 if (event->attr.precise_ip == 3)
3247 return &counter0_constraint;
3248
3249 c = intel_get_event_constraints(cpuc, idx, event);
3250
3251 return c;
3252 }
3253
3254 static bool allow_tsx_force_abort = true;
3255
3256 static struct event_constraint *
3257 tfa_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3258 struct perf_event *event)
3259 {
3260 struct event_constraint *c = hsw_get_event_constraints(cpuc, idx, event);
3261
3262 /*
3263 * Without TFA we must not use PMC3.
3264 */
3265 if (!allow_tsx_force_abort && test_bit(3, c->idxmsk) && idx >= 0) {
3266 c = dyn_constraint(cpuc, c, idx);
3267 c->idxmsk64 &= ~(1ULL << 3);
3268 c->weight--;
3269 }
3270
3271 return c;
3272 }
3273
3274 /*
3275 * Broadwell:
3276 *
3277 * The INST_RETIRED.ALL period always needs to have lowest 6 bits cleared
3278 * (BDM55) and it must not use a period smaller than 100 (BDM11). We combine
3279 * the two to enforce a minimum period of 128 (the smallest value that has bits
3280 * 0-5 cleared and >= 100).
3281 *
3282 * Because of how the code in x86_perf_event_set_period() works, the truncation
3283 * of the lower 6 bits is 'harmless' as we'll occasionally add a longer period
3284 * to make up for the 'lost' events due to carrying the 'error' in period_left.
3285 *
3286 * Therefore the effective (average) period matches the requested period,
3287 * despite coarser hardware granularity.
3288 */
3289 static u64 bdw_limit_period(struct perf_event *event, u64 left)
3290 {
3291 if ((event->hw.config & INTEL_ARCH_EVENT_MASK) ==
3292 X86_CONFIG(.event=0xc0, .umask=0x01)) {
3293 if (left < 128)
3294 left = 128;
3295 left &= ~0x3fULL;
3296 }
3297 return left;
3298 }
3299
3300 PMU_FORMAT_ATTR(event, "config:0-7" );
3301 PMU_FORMAT_ATTR(umask, "config:8-15" );
3302 PMU_FORMAT_ATTR(edge, "config:18" );
3303 PMU_FORMAT_ATTR(pc, "config:19" );
3304 PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */
3305 PMU_FORMAT_ATTR(inv, "config:23" );
3306 PMU_FORMAT_ATTR(cmask, "config:24-31" );
3307 PMU_FORMAT_ATTR(in_tx, "config:32");
3308 PMU_FORMAT_ATTR(in_tx_cp, "config:33");
3309
3310 static struct attribute *intel_arch_formats_attr[] = {
3311 &format_attr_event.attr,
3312 &format_attr_umask.attr,
3313 &format_attr_edge.attr,
3314 &format_attr_pc.attr,
3315 &format_attr_inv.attr,
3316 &format_attr_cmask.attr,
3317 NULL,
3318 };
3319
3320 ssize_t intel_event_sysfs_show(char *page, u64 config)
3321 {
3322 u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT);
3323
3324 return x86_event_sysfs_show(page, config, event);
3325 }
3326
3327 static struct intel_shared_regs *allocate_shared_regs(int cpu)
3328 {
3329 struct intel_shared_regs *regs;
3330 int i;
3331
3332 regs = kzalloc_node(sizeof(struct intel_shared_regs),
3333 GFP_KERNEL, cpu_to_node(cpu));
3334 if (regs) {
3335 /*
3336 * initialize the locks to keep lockdep happy
3337 */
3338 for (i = 0; i < EXTRA_REG_MAX; i++)
3339 raw_spin_lock_init(&regs->regs[i].lock);
3340
3341 regs->core_id = -1;
3342 }
3343 return regs;
3344 }
3345
3346 static struct intel_excl_cntrs *allocate_excl_cntrs(int cpu)
3347 {
3348 struct intel_excl_cntrs *c;
3349
3350 c = kzalloc_node(sizeof(struct intel_excl_cntrs),
3351 GFP_KERNEL, cpu_to_node(cpu));
3352 if (c) {
3353 raw_spin_lock_init(&c->lock);
3354 c->core_id = -1;
3355 }
3356 return c;
3357 }
3358
3359
3360 int intel_cpuc_prepare(struct cpu_hw_events *cpuc, int cpu)
3361 {
3362 if (x86_pmu.extra_regs || x86_pmu.lbr_sel_map) {
3363 cpuc->shared_regs = allocate_shared_regs(cpu);
3364 if (!cpuc->shared_regs)
3365 goto err;
3366 }
3367
3368 if (x86_pmu.flags & (PMU_FL_EXCL_CNTRS | PMU_FL_TFA)) {
3369 size_t sz = X86_PMC_IDX_MAX * sizeof(struct event_constraint);
3370
3371 cpuc->constraint_list = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));
3372 if (!cpuc->constraint_list)
3373 goto err_shared_regs;
3374 }
3375
3376 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) {
3377 cpuc->excl_cntrs = allocate_excl_cntrs(cpu);
3378 if (!cpuc->excl_cntrs)
3379 goto err_constraint_list;
3380
3381 cpuc->excl_thread_id = 0;
3382 }
3383
3384 return 0;
3385
3386 err_constraint_list:
3387 kfree(cpuc->constraint_list);
3388 cpuc->constraint_list = NULL;
3389
3390 err_shared_regs:
3391 kfree(cpuc->shared_regs);
3392 cpuc->shared_regs = NULL;
3393
3394 err:
3395 return -ENOMEM;
3396 }
3397
3398 static int intel_pmu_cpu_prepare(int cpu)
3399 {
3400 return intel_cpuc_prepare(&per_cpu(cpu_hw_events, cpu), cpu);
3401 }
3402
3403 static void flip_smm_bit(void *data)
3404 {
3405 unsigned long set = *(unsigned long *)data;
3406
3407 if (set > 0) {
3408 msr_set_bit(MSR_IA32_DEBUGCTLMSR,
3409 DEBUGCTLMSR_FREEZE_IN_SMM_BIT);
3410 } else {
3411 msr_clear_bit(MSR_IA32_DEBUGCTLMSR,
3412 DEBUGCTLMSR_FREEZE_IN_SMM_BIT);
3413 }
3414 }
3415
3416 static void intel_pmu_cpu_starting(int cpu)
3417 {
3418 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
3419 int core_id = topology_core_id(cpu);
3420 int i;
3421
3422 init_debug_store_on_cpu(cpu);
3423 /*
3424 * Deal with CPUs that don't clear their LBRs on power-up.
3425 */
3426 intel_pmu_lbr_reset();
3427
3428 cpuc->lbr_sel = NULL;
3429
3430 if (x86_pmu.flags & PMU_FL_TFA) {
3431 WARN_ON_ONCE(cpuc->tfa_shadow);
3432 cpuc->tfa_shadow = ~0ULL;
3433 intel_set_tfa(cpuc, false);
3434 }
3435
3436 if (x86_pmu.version > 1)
3437 flip_smm_bit(&x86_pmu.attr_freeze_on_smi);
3438
3439 if (!cpuc->shared_regs)
3440 return;
3441
3442 if (!(x86_pmu.flags & PMU_FL_NO_HT_SHARING)) {
3443 for_each_cpu(i, topology_sibling_cpumask(cpu)) {
3444 struct intel_shared_regs *pc;
3445
3446 pc = per_cpu(cpu_hw_events, i).shared_regs;
3447 if (pc && pc->core_id == core_id) {
3448 cpuc->kfree_on_online[0] = cpuc->shared_regs;
3449 cpuc->shared_regs = pc;
3450 break;
3451 }
3452 }
3453 cpuc->shared_regs->core_id = core_id;
3454 cpuc->shared_regs->refcnt++;
3455 }
3456
3457 if (x86_pmu.lbr_sel_map)
3458 cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR];
3459
3460 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) {
3461 for_each_cpu(i, topology_sibling_cpumask(cpu)) {
3462 struct cpu_hw_events *sibling;
3463 struct intel_excl_cntrs *c;
3464
3465 sibling = &per_cpu(cpu_hw_events, i);
3466 c = sibling->excl_cntrs;
3467 if (c && c->core_id == core_id) {
3468 cpuc->kfree_on_online[1] = cpuc->excl_cntrs;
3469 cpuc->excl_cntrs = c;
3470 if (!sibling->excl_thread_id)
3471 cpuc->excl_thread_id = 1;
3472 break;
3473 }
3474 }
3475 cpuc->excl_cntrs->core_id = core_id;
3476 cpuc->excl_cntrs->refcnt++;
3477 }
3478 }
3479
3480 static void free_excl_cntrs(struct cpu_hw_events *cpuc)
3481 {
3482 struct intel_excl_cntrs *c;
3483
3484 c = cpuc->excl_cntrs;
3485 if (c) {
3486 if (c->core_id == -1 || --c->refcnt == 0)
3487 kfree(c);
3488 cpuc->excl_cntrs = NULL;
3489 }
3490
3491 kfree(cpuc->constraint_list);
3492 cpuc->constraint_list = NULL;
3493 }
3494
3495 static void intel_pmu_cpu_dying(int cpu)
3496 {
3497 fini_debug_store_on_cpu(cpu);
3498 }
3499
3500 void intel_cpuc_finish(struct cpu_hw_events *cpuc)
3501 {
3502 struct intel_shared_regs *pc;
3503
3504 pc = cpuc->shared_regs;
3505 if (pc) {
3506 if (pc->core_id == -1 || --pc->refcnt == 0)
3507 kfree(pc);
3508 cpuc->shared_regs = NULL;
3509 }
3510
3511 free_excl_cntrs(cpuc);
3512 }
3513
3514 static void intel_pmu_cpu_dead(int cpu)
3515 {
3516 intel_cpuc_finish(&per_cpu(cpu_hw_events, cpu));
3517 }
3518
3519 static void intel_pmu_sched_task(struct perf_event_context *ctx,
3520 bool sched_in)
3521 {
3522 intel_pmu_pebs_sched_task(ctx, sched_in);
3523 intel_pmu_lbr_sched_task(ctx, sched_in);
3524 }
3525
3526 static int intel_pmu_check_period(struct perf_event *event, u64 value)
3527 {
3528 return intel_pmu_has_bts_period(event, value) ? -EINVAL : 0;
3529 }
3530
3531 PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63");
3532
3533 PMU_FORMAT_ATTR(ldlat, "config1:0-15");
3534
3535 PMU_FORMAT_ATTR(frontend, "config1:0-23");
3536
3537 static struct attribute *intel_arch3_formats_attr[] = {
3538 &format_attr_event.attr,
3539 &format_attr_umask.attr,
3540 &format_attr_edge.attr,
3541 &format_attr_pc.attr,
3542 &format_attr_any.attr,
3543 &format_attr_inv.attr,
3544 &format_attr_cmask.attr,
3545 NULL,
3546 };
3547
3548 static struct attribute *hsw_format_attr[] = {
3549 &format_attr_in_tx.attr,
3550 &format_attr_in_tx_cp.attr,
3551 &format_attr_offcore_rsp.attr,
3552 &format_attr_ldlat.attr,
3553 NULL
3554 };
3555
3556 static struct attribute *nhm_format_attr[] = {
3557 &format_attr_offcore_rsp.attr,
3558 &format_attr_ldlat.attr,
3559 NULL
3560 };
3561
3562 static struct attribute *slm_format_attr[] = {
3563 &format_attr_offcore_rsp.attr,
3564 NULL
3565 };
3566
3567 static struct attribute *skl_format_attr[] = {
3568 &format_attr_frontend.attr,
3569 NULL,
3570 };
3571
3572 static __initconst const struct x86_pmu core_pmu = {
3573 .name = "core",
3574 .handle_irq = x86_pmu_handle_irq,
3575 .disable_all = x86_pmu_disable_all,
3576 .enable_all = core_pmu_enable_all,
3577 .enable = core_pmu_enable_event,
3578 .disable = x86_pmu_disable_event,
3579 .hw_config = core_pmu_hw_config,
3580 .schedule_events = x86_schedule_events,
3581 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
3582 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
3583 .event_map = intel_pmu_event_map,
3584 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
3585 .apic = 1,
3586 .free_running_flags = PEBS_FREERUNNING_FLAGS,
3587
3588 /*
3589 * Intel PMCs cannot be accessed sanely above 32-bit width,
3590 * so we install an artificial 1<<31 period regardless of
3591 * the generic event period:
3592 */
3593 .max_period = (1ULL<<31) - 1,
3594 .get_event_constraints = intel_get_event_constraints,
3595 .put_event_constraints = intel_put_event_constraints,
3596 .event_constraints = intel_core_event_constraints,
3597 .guest_get_msrs = core_guest_get_msrs,
3598 .format_attrs = intel_arch_formats_attr,
3599 .events_sysfs_show = intel_event_sysfs_show,
3600
3601 /*
3602 * Virtual (or funny metal) CPU can define x86_pmu.extra_regs
3603 * together with PMU version 1 and thus be using core_pmu with
3604 * shared_regs. We need following callbacks here to allocate
3605 * it properly.
3606 */
3607 .cpu_prepare = intel_pmu_cpu_prepare,
3608 .cpu_starting = intel_pmu_cpu_starting,
3609 .cpu_dying = intel_pmu_cpu_dying,
3610 .cpu_dead = intel_pmu_cpu_dead,
3611
3612 .check_period = intel_pmu_check_period,
3613 };
3614
3615 static struct attribute *intel_pmu_attrs[];
3616
3617 static __initconst const struct x86_pmu intel_pmu = {
3618 .name = "Intel",
3619 .handle_irq = intel_pmu_handle_irq,
3620 .disable_all = intel_pmu_disable_all,
3621 .enable_all = intel_pmu_enable_all,
3622 .enable = intel_pmu_enable_event,
3623 .disable = intel_pmu_disable_event,
3624 .add = intel_pmu_add_event,
3625 .del = intel_pmu_del_event,
3626 .hw_config = intel_pmu_hw_config,
3627 .schedule_events = x86_schedule_events,
3628 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
3629 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
3630 .event_map = intel_pmu_event_map,
3631 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
3632 .apic = 1,
3633 .free_running_flags = PEBS_FREERUNNING_FLAGS,
3634 /*
3635 * Intel PMCs cannot be accessed sanely above 32 bit width,
3636 * so we install an artificial 1<<31 period regardless of
3637 * the generic event period:
3638 */
3639 .max_period = (1ULL << 31) - 1,
3640 .get_event_constraints = intel_get_event_constraints,
3641 .put_event_constraints = intel_put_event_constraints,
3642 .pebs_aliases = intel_pebs_aliases_core2,
3643
3644 .format_attrs = intel_arch3_formats_attr,
3645 .events_sysfs_show = intel_event_sysfs_show,
3646
3647 .attrs = intel_pmu_attrs,
3648
3649 .cpu_prepare = intel_pmu_cpu_prepare,
3650 .cpu_starting = intel_pmu_cpu_starting,
3651 .cpu_dying = intel_pmu_cpu_dying,
3652 .cpu_dead = intel_pmu_cpu_dead,
3653
3654 .guest_get_msrs = intel_guest_get_msrs,
3655 .sched_task = intel_pmu_sched_task,
3656
3657 .check_period = intel_pmu_check_period,
3658 };
3659
3660 static __init void intel_clovertown_quirk(void)
3661 {
3662 /*
3663 * PEBS is unreliable due to:
3664 *
3665 * AJ67 - PEBS may experience CPL leaks
3666 * AJ68 - PEBS PMI may be delayed by one event
3667 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
3668 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
3669 *
3670 * AJ67 could be worked around by restricting the OS/USR flags.
3671 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
3672 *
3673 * AJ106 could possibly be worked around by not allowing LBR
3674 * usage from PEBS, including the fixup.
3675 * AJ68 could possibly be worked around by always programming
3676 * a pebs_event_reset[0] value and coping with the lost events.
3677 *
3678 * But taken together it might just make sense to not enable PEBS on
3679 * these chips.
3680 */
3681 pr_warn("PEBS disabled due to CPU errata\n");
3682 x86_pmu.pebs = 0;
3683 x86_pmu.pebs_constraints = NULL;
3684 }
3685
3686 static int intel_snb_pebs_broken(int cpu)
3687 {
3688 u32 rev = UINT_MAX; /* default to broken for unknown models */
3689
3690 switch (cpu_data(cpu).x86_model) {
3691 case INTEL_FAM6_SANDYBRIDGE:
3692 rev = 0x28;
3693 break;
3694
3695 case INTEL_FAM6_SANDYBRIDGE_X:
3696 switch (cpu_data(cpu).x86_stepping) {
3697 case 6: rev = 0x618; break;
3698 case 7: rev = 0x70c; break;
3699 }
3700 }
3701
3702 return (cpu_data(cpu).microcode < rev);
3703 }
3704
3705 static void intel_snb_check_microcode(void)
3706 {
3707 int pebs_broken = 0;
3708 int cpu;
3709
3710 for_each_online_cpu(cpu) {
3711 if ((pebs_broken = intel_snb_pebs_broken(cpu)))
3712 break;
3713 }
3714
3715 if (pebs_broken == x86_pmu.pebs_broken)
3716 return;
3717
3718 /*
3719 * Serialized by the microcode lock..
3720 */
3721 if (x86_pmu.pebs_broken) {
3722 pr_info("PEBS enabled due to microcode update\n");
3723 x86_pmu.pebs_broken = 0;
3724 } else {
3725 pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n");
3726 x86_pmu.pebs_broken = 1;
3727 }
3728 }
3729
3730 static bool is_lbr_from(unsigned long msr)
3731 {
3732 unsigned long lbr_from_nr = x86_pmu.lbr_from + x86_pmu.lbr_nr;
3733
3734 return x86_pmu.lbr_from <= msr && msr < lbr_from_nr;
3735 }
3736
3737 /*
3738 * Under certain circumstances, access certain MSR may cause #GP.
3739 * The function tests if the input MSR can be safely accessed.
3740 */
3741 static bool check_msr(unsigned long msr, u64 mask)
3742 {
3743 u64 val_old, val_new, val_tmp;
3744
3745 /*
3746 * Read the current value, change it and read it back to see if it
3747 * matches, this is needed to detect certain hardware emulators
3748 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
3749 */
3750 if (rdmsrl_safe(msr, &val_old))
3751 return false;
3752
3753 /*
3754 * Only change the bits which can be updated by wrmsrl.
3755 */
3756 val_tmp = val_old ^ mask;
3757
3758 if (is_lbr_from(msr))
3759 val_tmp = lbr_from_signext_quirk_wr(val_tmp);
3760
3761 if (wrmsrl_safe(msr, val_tmp) ||
3762 rdmsrl_safe(msr, &val_new))
3763 return false;
3764
3765 /*
3766 * Quirk only affects validation in wrmsr(), so wrmsrl()'s value
3767 * should equal rdmsrl()'s even with the quirk.
3768 */
3769 if (val_new != val_tmp)
3770 return false;
3771
3772 if (is_lbr_from(msr))
3773 val_old = lbr_from_signext_quirk_wr(val_old);
3774
3775 /* Here it's sure that the MSR can be safely accessed.
3776 * Restore the old value and return.
3777 */
3778 wrmsrl(msr, val_old);
3779
3780 return true;
3781 }
3782
3783 static __init void intel_sandybridge_quirk(void)
3784 {
3785 x86_pmu.check_microcode = intel_snb_check_microcode;
3786 cpus_read_lock();
3787 intel_snb_check_microcode();
3788 cpus_read_unlock();
3789 }
3790
3791 static const struct { int id; char *name; } intel_arch_events_map[] __initconst = {
3792 { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" },
3793 { PERF_COUNT_HW_INSTRUCTIONS, "instructions" },
3794 { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" },
3795 { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" },
3796 { PERF_COUNT_HW_CACHE_MISSES, "cache misses" },
3797 { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" },
3798 { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" },
3799 };
3800
3801 static __init void intel_arch_events_quirk(void)
3802 {
3803 int bit;
3804
3805 /* disable event that reported as not presend by cpuid */
3806 for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) {
3807 intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0;
3808 pr_warn("CPUID marked event: \'%s\' unavailable\n",
3809 intel_arch_events_map[bit].name);
3810 }
3811 }
3812
3813 static __init void intel_nehalem_quirk(void)
3814 {
3815 union cpuid10_ebx ebx;
3816
3817 ebx.full = x86_pmu.events_maskl;
3818 if (ebx.split.no_branch_misses_retired) {
3819 /*
3820 * Erratum AAJ80 detected, we work it around by using
3821 * the BR_MISP_EXEC.ANY event. This will over-count
3822 * branch-misses, but it's still much better than the
3823 * architectural event which is often completely bogus:
3824 */
3825 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89;
3826 ebx.split.no_branch_misses_retired = 0;
3827 x86_pmu.events_maskl = ebx.full;
3828 pr_info("CPU erratum AAJ80 worked around\n");
3829 }
3830 }
3831
3832 /*
3833 * enable software workaround for errata:
3834 * SNB: BJ122
3835 * IVB: BV98
3836 * HSW: HSD29
3837 *
3838 * Only needed when HT is enabled. However detecting
3839 * if HT is enabled is difficult (model specific). So instead,
3840 * we enable the workaround in the early boot, and verify if
3841 * it is needed in a later initcall phase once we have valid
3842 * topology information to check if HT is actually enabled
3843 */
3844 static __init void intel_ht_bug(void)
3845 {
3846 x86_pmu.flags |= PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED;
3847
3848 x86_pmu.start_scheduling = intel_start_scheduling;
3849 x86_pmu.commit_scheduling = intel_commit_scheduling;
3850 x86_pmu.stop_scheduling = intel_stop_scheduling;
3851 }
3852
3853 EVENT_ATTR_STR(mem-loads, mem_ld_hsw, "event=0xcd,umask=0x1,ldlat=3");
3854 EVENT_ATTR_STR(mem-stores, mem_st_hsw, "event=0xd0,umask=0x82")
3855
3856 /* Haswell special events */
3857 EVENT_ATTR_STR(tx-start, tx_start, "event=0xc9,umask=0x1");
3858 EVENT_ATTR_STR(tx-commit, tx_commit, "event=0xc9,umask=0x2");
3859 EVENT_ATTR_STR(tx-abort, tx_abort, "event=0xc9,umask=0x4");
3860 EVENT_ATTR_STR(tx-capacity, tx_capacity, "event=0x54,umask=0x2");
3861 EVENT_ATTR_STR(tx-conflict, tx_conflict, "event=0x54,umask=0x1");
3862 EVENT_ATTR_STR(el-start, el_start, "event=0xc8,umask=0x1");
3863 EVENT_ATTR_STR(el-commit, el_commit, "event=0xc8,umask=0x2");
3864 EVENT_ATTR_STR(el-abort, el_abort, "event=0xc8,umask=0x4");
3865 EVENT_ATTR_STR(el-capacity, el_capacity, "event=0x54,umask=0x2");
3866 EVENT_ATTR_STR(el-conflict, el_conflict, "event=0x54,umask=0x1");
3867 EVENT_ATTR_STR(cycles-t, cycles_t, "event=0x3c,in_tx=1");
3868 EVENT_ATTR_STR(cycles-ct, cycles_ct, "event=0x3c,in_tx=1,in_tx_cp=1");
3869
3870 static struct attribute *hsw_events_attrs[] = {
3871 EVENT_PTR(mem_ld_hsw),
3872 EVENT_PTR(mem_st_hsw),
3873 EVENT_PTR(td_slots_issued),
3874 EVENT_PTR(td_slots_retired),
3875 EVENT_PTR(td_fetch_bubbles),
3876 EVENT_PTR(td_total_slots),
3877 EVENT_PTR(td_total_slots_scale),
3878 EVENT_PTR(td_recovery_bubbles),
3879 EVENT_PTR(td_recovery_bubbles_scale),
3880 NULL
3881 };
3882
3883 static struct attribute *hsw_tsx_events_attrs[] = {
3884 EVENT_PTR(tx_start),
3885 EVENT_PTR(tx_commit),
3886 EVENT_PTR(tx_abort),
3887 EVENT_PTR(tx_capacity),
3888 EVENT_PTR(tx_conflict),
3889 EVENT_PTR(el_start),
3890 EVENT_PTR(el_commit),
3891 EVENT_PTR(el_abort),
3892 EVENT_PTR(el_capacity),
3893 EVENT_PTR(el_conflict),
3894 EVENT_PTR(cycles_t),
3895 EVENT_PTR(cycles_ct),
3896 NULL
3897 };
3898
3899 static __init struct attribute **get_hsw_events_attrs(void)
3900 {
3901 return boot_cpu_has(X86_FEATURE_RTM) ?
3902 merge_attr(hsw_events_attrs, hsw_tsx_events_attrs) :
3903 hsw_events_attrs;
3904 }
3905
3906 static ssize_t freeze_on_smi_show(struct device *cdev,
3907 struct device_attribute *attr,
3908 char *buf)
3909 {
3910 return sprintf(buf, "%lu\n", x86_pmu.attr_freeze_on_smi);
3911 }
3912
3913 static DEFINE_MUTEX(freeze_on_smi_mutex);
3914
3915 static ssize_t freeze_on_smi_store(struct device *cdev,
3916 struct device_attribute *attr,
3917 const char *buf, size_t count)
3918 {
3919 unsigned long val;
3920 ssize_t ret;
3921
3922 ret = kstrtoul(buf, 0, &val);
3923 if (ret)
3924 return ret;
3925
3926 if (val > 1)
3927 return -EINVAL;
3928
3929 mutex_lock(&freeze_on_smi_mutex);
3930
3931 if (x86_pmu.attr_freeze_on_smi == val)
3932 goto done;
3933
3934 x86_pmu.attr_freeze_on_smi = val;
3935
3936 get_online_cpus();
3937 on_each_cpu(flip_smm_bit, &val, 1);
3938 put_online_cpus();
3939 done:
3940 mutex_unlock(&freeze_on_smi_mutex);
3941
3942 return count;
3943 }
3944
3945 static DEVICE_ATTR_RW(freeze_on_smi);
3946
3947 static ssize_t branches_show(struct device *cdev,
3948 struct device_attribute *attr,
3949 char *buf)
3950 {
3951 return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu.lbr_nr);
3952 }
3953
3954 static DEVICE_ATTR_RO(branches);
3955
3956 static struct attribute *lbr_attrs[] = {
3957 &dev_attr_branches.attr,
3958 NULL
3959 };
3960
3961 static char pmu_name_str[30];
3962
3963 static ssize_t pmu_name_show(struct device *cdev,
3964 struct device_attribute *attr,
3965 char *buf)
3966 {
3967 return snprintf(buf, PAGE_SIZE, "%s\n", pmu_name_str);
3968 }
3969
3970 static DEVICE_ATTR_RO(pmu_name);
3971
3972 static struct attribute *intel_pmu_caps_attrs[] = {
3973 &dev_attr_pmu_name.attr,
3974 NULL
3975 };
3976
3977 static DEVICE_BOOL_ATTR(allow_tsx_force_abort, 0644, allow_tsx_force_abort);
3978
3979 static struct attribute *intel_pmu_attrs[] = {
3980 &dev_attr_freeze_on_smi.attr,
3981 NULL, /* &dev_attr_allow_tsx_force_abort.attr.attr */
3982 NULL,
3983 };
3984
3985 __init int intel_pmu_init(void)
3986 {
3987 struct attribute **extra_attr = NULL;
3988 struct attribute **to_free = NULL;
3989 union cpuid10_edx edx;
3990 union cpuid10_eax eax;
3991 union cpuid10_ebx ebx;
3992 struct event_constraint *c;
3993 unsigned int unused;
3994 struct extra_reg *er;
3995 int version, i;
3996 char *name;
3997
3998 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
3999 switch (boot_cpu_data.x86) {
4000 case 0x6:
4001 return p6_pmu_init();
4002 case 0xb:
4003 return knc_pmu_init();
4004 case 0xf:
4005 return p4_pmu_init();
4006 }
4007 return -ENODEV;
4008 }
4009
4010 /*
4011 * Check whether the Architectural PerfMon supports
4012 * Branch Misses Retired hw_event or not.
4013 */
4014 cpuid(10, &eax.full, &ebx.full, &unused, &edx.full);
4015 if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT)
4016 return -ENODEV;
4017
4018 version = eax.split.version_id;
4019 if (version < 2)
4020 x86_pmu = core_pmu;
4021 else
4022 x86_pmu = intel_pmu;
4023
4024 x86_pmu.version = version;
4025 x86_pmu.num_counters = eax.split.num_counters;
4026 x86_pmu.cntval_bits = eax.split.bit_width;
4027 x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1;
4028
4029 x86_pmu.events_maskl = ebx.full;
4030 x86_pmu.events_mask_len = eax.split.mask_length;
4031
4032 x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters);
4033
4034 /*
4035 * Quirk: v2 perfmon does not report fixed-purpose events, so
4036 * assume at least 3 events, when not running in a hypervisor:
4037 */
4038 if (version > 1) {
4039 int assume = 3 * !boot_cpu_has(X86_FEATURE_HYPERVISOR);
4040
4041 x86_pmu.num_counters_fixed =
4042 max((int)edx.split.num_counters_fixed, assume);
4043 }
4044
4045 if (boot_cpu_has(X86_FEATURE_PDCM)) {
4046 u64 capabilities;
4047
4048 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
4049 x86_pmu.intel_cap.capabilities = capabilities;
4050 }
4051
4052 intel_ds_init();
4053
4054 x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */
4055
4056 /*
4057 * Install the hw-cache-events table:
4058 */
4059 switch (boot_cpu_data.x86_model) {
4060 case INTEL_FAM6_CORE_YONAH:
4061 pr_cont("Core events, ");
4062 name = "core";
4063 break;
4064
4065 case INTEL_FAM6_CORE2_MEROM:
4066 x86_add_quirk(intel_clovertown_quirk);
4067 case INTEL_FAM6_CORE2_MEROM_L:
4068 case INTEL_FAM6_CORE2_PENRYN:
4069 case INTEL_FAM6_CORE2_DUNNINGTON:
4070 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
4071 sizeof(hw_cache_event_ids));
4072
4073 intel_pmu_lbr_init_core();
4074
4075 x86_pmu.event_constraints = intel_core2_event_constraints;
4076 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints;
4077 pr_cont("Core2 events, ");
4078 name = "core2";
4079 break;
4080
4081 case INTEL_FAM6_NEHALEM:
4082 case INTEL_FAM6_NEHALEM_EP:
4083 case INTEL_FAM6_NEHALEM_EX:
4084 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
4085 sizeof(hw_cache_event_ids));
4086 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
4087 sizeof(hw_cache_extra_regs));
4088
4089 intel_pmu_lbr_init_nhm();
4090
4091 x86_pmu.event_constraints = intel_nehalem_event_constraints;
4092 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints;
4093 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
4094 x86_pmu.extra_regs = intel_nehalem_extra_regs;
4095
4096 x86_pmu.cpu_events = nhm_events_attrs;
4097
4098 /* UOPS_ISSUED.STALLED_CYCLES */
4099 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
4100 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
4101 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
4102 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
4103 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
4104
4105 intel_pmu_pebs_data_source_nhm();
4106 x86_add_quirk(intel_nehalem_quirk);
4107 x86_pmu.pebs_no_tlb = 1;
4108 extra_attr = nhm_format_attr;
4109
4110 pr_cont("Nehalem events, ");
4111 name = "nehalem";
4112 break;
4113
4114 case INTEL_FAM6_ATOM_BONNELL:
4115 case INTEL_FAM6_ATOM_BONNELL_MID:
4116 case INTEL_FAM6_ATOM_SALTWELL:
4117 case INTEL_FAM6_ATOM_SALTWELL_MID:
4118 case INTEL_FAM6_ATOM_SALTWELL_TABLET:
4119 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
4120 sizeof(hw_cache_event_ids));
4121
4122 intel_pmu_lbr_init_atom();
4123
4124 x86_pmu.event_constraints = intel_gen_event_constraints;
4125 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints;
4126 x86_pmu.pebs_aliases = intel_pebs_aliases_core2;
4127 pr_cont("Atom events, ");
4128 name = "bonnell";
4129 break;
4130
4131 case INTEL_FAM6_ATOM_SILVERMONT:
4132 case INTEL_FAM6_ATOM_SILVERMONT_X:
4133 case INTEL_FAM6_ATOM_SILVERMONT_MID:
4134 case INTEL_FAM6_ATOM_AIRMONT:
4135 case INTEL_FAM6_ATOM_AIRMONT_MID:
4136 memcpy(hw_cache_event_ids, slm_hw_cache_event_ids,
4137 sizeof(hw_cache_event_ids));
4138 memcpy(hw_cache_extra_regs, slm_hw_cache_extra_regs,
4139 sizeof(hw_cache_extra_regs));
4140
4141 intel_pmu_lbr_init_slm();
4142
4143 x86_pmu.event_constraints = intel_slm_event_constraints;
4144 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints;
4145 x86_pmu.extra_regs = intel_slm_extra_regs;
4146 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4147 x86_pmu.cpu_events = slm_events_attrs;
4148 extra_attr = slm_format_attr;
4149 pr_cont("Silvermont events, ");
4150 name = "silvermont";
4151 break;
4152
4153 case INTEL_FAM6_ATOM_GOLDMONT:
4154 case INTEL_FAM6_ATOM_GOLDMONT_X:
4155 memcpy(hw_cache_event_ids, glm_hw_cache_event_ids,
4156 sizeof(hw_cache_event_ids));
4157 memcpy(hw_cache_extra_regs, glm_hw_cache_extra_regs,
4158 sizeof(hw_cache_extra_regs));
4159
4160 intel_pmu_lbr_init_skl();
4161
4162 x86_pmu.event_constraints = intel_slm_event_constraints;
4163 x86_pmu.pebs_constraints = intel_glm_pebs_event_constraints;
4164 x86_pmu.extra_regs = intel_glm_extra_regs;
4165 /*
4166 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS
4167 * for precise cycles.
4168 * :pp is identical to :ppp
4169 */
4170 x86_pmu.pebs_aliases = NULL;
4171 x86_pmu.pebs_prec_dist = true;
4172 x86_pmu.lbr_pt_coexist = true;
4173 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4174 x86_pmu.cpu_events = glm_events_attrs;
4175 extra_attr = slm_format_attr;
4176 pr_cont("Goldmont events, ");
4177 name = "goldmont";
4178 break;
4179
4180 case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
4181 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids,
4182 sizeof(hw_cache_event_ids));
4183 memcpy(hw_cache_extra_regs, glp_hw_cache_extra_regs,
4184 sizeof(hw_cache_extra_regs));
4185
4186 intel_pmu_lbr_init_skl();
4187
4188 x86_pmu.event_constraints = intel_slm_event_constraints;
4189 x86_pmu.pebs_constraints = intel_glp_pebs_event_constraints;
4190 x86_pmu.extra_regs = intel_glm_extra_regs;
4191 /*
4192 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS
4193 * for precise cycles.
4194 */
4195 x86_pmu.pebs_aliases = NULL;
4196 x86_pmu.pebs_prec_dist = true;
4197 x86_pmu.lbr_pt_coexist = true;
4198 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4199 x86_pmu.get_event_constraints = glp_get_event_constraints;
4200 x86_pmu.cpu_events = glm_events_attrs;
4201 /* Goldmont Plus has 4-wide pipeline */
4202 event_attr_td_total_slots_scale_glm.event_str = "4";
4203 extra_attr = slm_format_attr;
4204 pr_cont("Goldmont plus events, ");
4205 name = "goldmont_plus";
4206 break;
4207
4208 case INTEL_FAM6_WESTMERE:
4209 case INTEL_FAM6_WESTMERE_EP:
4210 case INTEL_FAM6_WESTMERE_EX:
4211 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
4212 sizeof(hw_cache_event_ids));
4213 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
4214 sizeof(hw_cache_extra_regs));
4215
4216 intel_pmu_lbr_init_nhm();
4217
4218 x86_pmu.event_constraints = intel_westmere_event_constraints;
4219 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
4220 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints;
4221 x86_pmu.extra_regs = intel_westmere_extra_regs;
4222 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4223
4224 x86_pmu.cpu_events = nhm_events_attrs;
4225
4226 /* UOPS_ISSUED.STALLED_CYCLES */
4227 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
4228 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
4229 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
4230 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
4231 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
4232
4233 intel_pmu_pebs_data_source_nhm();
4234 extra_attr = nhm_format_attr;
4235 pr_cont("Westmere events, ");
4236 name = "westmere";
4237 break;
4238
4239 case INTEL_FAM6_SANDYBRIDGE:
4240 case INTEL_FAM6_SANDYBRIDGE_X:
4241 x86_add_quirk(intel_sandybridge_quirk);
4242 x86_add_quirk(intel_ht_bug);
4243 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
4244 sizeof(hw_cache_event_ids));
4245 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
4246 sizeof(hw_cache_extra_regs));
4247
4248 intel_pmu_lbr_init_snb();
4249
4250 x86_pmu.event_constraints = intel_snb_event_constraints;
4251 x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints;
4252 x86_pmu.pebs_aliases = intel_pebs_aliases_snb;
4253 if (boot_cpu_data.x86_model == INTEL_FAM6_SANDYBRIDGE_X)
4254 x86_pmu.extra_regs = intel_snbep_extra_regs;
4255 else
4256 x86_pmu.extra_regs = intel_snb_extra_regs;
4257
4258
4259 /* all extra regs are per-cpu when HT is on */
4260 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4261 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
4262
4263 x86_pmu.cpu_events = snb_events_attrs;
4264
4265 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
4266 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
4267 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
4268 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
4269 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
4270 X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1);
4271
4272 extra_attr = nhm_format_attr;
4273
4274 pr_cont("SandyBridge events, ");
4275 name = "sandybridge";
4276 break;
4277
4278 case INTEL_FAM6_IVYBRIDGE:
4279 case INTEL_FAM6_IVYBRIDGE_X:
4280 x86_add_quirk(intel_ht_bug);
4281 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
4282 sizeof(hw_cache_event_ids));
4283 /* dTLB-load-misses on IVB is different than SNB */
4284 hw_cache_event_ids[C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = 0x8108; /* DTLB_LOAD_MISSES.DEMAND_LD_MISS_CAUSES_A_WALK */
4285
4286 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
4287 sizeof(hw_cache_extra_regs));
4288
4289 intel_pmu_lbr_init_snb();
4290
4291 x86_pmu.event_constraints = intel_ivb_event_constraints;
4292 x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints;
4293 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
4294 x86_pmu.pebs_prec_dist = true;
4295 if (boot_cpu_data.x86_model == INTEL_FAM6_IVYBRIDGE_X)
4296 x86_pmu.extra_regs = intel_snbep_extra_regs;
4297 else
4298 x86_pmu.extra_regs = intel_snb_extra_regs;
4299 /* all extra regs are per-cpu when HT is on */
4300 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4301 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
4302
4303 x86_pmu.cpu_events = snb_events_attrs;
4304
4305 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
4306 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
4307 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
4308
4309 extra_attr = nhm_format_attr;
4310
4311 pr_cont("IvyBridge events, ");
4312 name = "ivybridge";
4313 break;
4314
4315
4316 case INTEL_FAM6_HASWELL_CORE:
4317 case INTEL_FAM6_HASWELL_X:
4318 case INTEL_FAM6_HASWELL_ULT:
4319 case INTEL_FAM6_HASWELL_GT3E:
4320 x86_add_quirk(intel_ht_bug);
4321 x86_pmu.late_ack = true;
4322 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids));
4323 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
4324
4325 intel_pmu_lbr_init_hsw();
4326
4327 x86_pmu.event_constraints = intel_hsw_event_constraints;
4328 x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints;
4329 x86_pmu.extra_regs = intel_snbep_extra_regs;
4330 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
4331 x86_pmu.pebs_prec_dist = true;
4332 /* all extra regs are per-cpu when HT is on */
4333 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4334 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
4335
4336 x86_pmu.hw_config = hsw_hw_config;
4337 x86_pmu.get_event_constraints = hsw_get_event_constraints;
4338 x86_pmu.cpu_events = get_hsw_events_attrs();
4339 x86_pmu.lbr_double_abort = true;
4340 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
4341 hsw_format_attr : nhm_format_attr;
4342 pr_cont("Haswell events, ");
4343 name = "haswell";
4344 break;
4345
4346 case INTEL_FAM6_BROADWELL_CORE:
4347 case INTEL_FAM6_BROADWELL_XEON_D:
4348 case INTEL_FAM6_BROADWELL_GT3E:
4349 case INTEL_FAM6_BROADWELL_X:
4350 x86_pmu.late_ack = true;
4351 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids));
4352 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
4353
4354 /* L3_MISS_LOCAL_DRAM is BIT(26) in Broadwell */
4355 hw_cache_extra_regs[C(LL)][C(OP_READ)][C(RESULT_MISS)] = HSW_DEMAND_READ |
4356 BDW_L3_MISS|HSW_SNOOP_DRAM;
4357 hw_cache_extra_regs[C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = HSW_DEMAND_WRITE|BDW_L3_MISS|
4358 HSW_SNOOP_DRAM;
4359 hw_cache_extra_regs[C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = HSW_DEMAND_READ|
4360 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM;
4361 hw_cache_extra_regs[C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = HSW_DEMAND_WRITE|
4362 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM;
4363
4364 intel_pmu_lbr_init_hsw();
4365
4366 x86_pmu.event_constraints = intel_bdw_event_constraints;
4367 x86_pmu.pebs_constraints = intel_bdw_pebs_event_constraints;
4368 x86_pmu.extra_regs = intel_snbep_extra_regs;
4369 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
4370 x86_pmu.pebs_prec_dist = true;
4371 /* all extra regs are per-cpu when HT is on */
4372 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4373 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
4374
4375 x86_pmu.hw_config = hsw_hw_config;
4376 x86_pmu.get_event_constraints = hsw_get_event_constraints;
4377 x86_pmu.cpu_events = get_hsw_events_attrs();
4378 x86_pmu.limit_period = bdw_limit_period;
4379 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
4380 hsw_format_attr : nhm_format_attr;
4381 pr_cont("Broadwell events, ");
4382 name = "broadwell";
4383 break;
4384
4385 case INTEL_FAM6_XEON_PHI_KNL:
4386 case INTEL_FAM6_XEON_PHI_KNM:
4387 memcpy(hw_cache_event_ids,
4388 slm_hw_cache_event_ids, sizeof(hw_cache_event_ids));
4389 memcpy(hw_cache_extra_regs,
4390 knl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
4391 intel_pmu_lbr_init_knl();
4392
4393 x86_pmu.event_constraints = intel_slm_event_constraints;
4394 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints;
4395 x86_pmu.extra_regs = intel_knl_extra_regs;
4396
4397 /* all extra regs are per-cpu when HT is on */
4398 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4399 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
4400 extra_attr = slm_format_attr;
4401 pr_cont("Knights Landing/Mill events, ");
4402 name = "knights-landing";
4403 break;
4404
4405 case INTEL_FAM6_SKYLAKE_MOBILE:
4406 case INTEL_FAM6_SKYLAKE_DESKTOP:
4407 case INTEL_FAM6_SKYLAKE_X:
4408 case INTEL_FAM6_KABYLAKE_MOBILE:
4409 case INTEL_FAM6_KABYLAKE_DESKTOP:
4410 x86_pmu.late_ack = true;
4411 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids));
4412 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
4413 intel_pmu_lbr_init_skl();
4414
4415 /* INT_MISC.RECOVERY_CYCLES has umask 1 in Skylake */
4416 event_attr_td_recovery_bubbles.event_str_noht =
4417 "event=0xd,umask=0x1,cmask=1";
4418 event_attr_td_recovery_bubbles.event_str_ht =
4419 "event=0xd,umask=0x1,cmask=1,any=1";
4420
4421 x86_pmu.event_constraints = intel_skl_event_constraints;
4422 x86_pmu.pebs_constraints = intel_skl_pebs_event_constraints;
4423 x86_pmu.extra_regs = intel_skl_extra_regs;
4424 x86_pmu.pebs_aliases = intel_pebs_aliases_skl;
4425 x86_pmu.pebs_prec_dist = true;
4426 /* all extra regs are per-cpu when HT is on */
4427 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
4428 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
4429
4430 x86_pmu.hw_config = hsw_hw_config;
4431 x86_pmu.get_event_constraints = hsw_get_event_constraints;
4432 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
4433 hsw_format_attr : nhm_format_attr;
4434 extra_attr = merge_attr(extra_attr, skl_format_attr);
4435 to_free = extra_attr;
4436 x86_pmu.cpu_events = get_hsw_events_attrs();
4437 intel_pmu_pebs_data_source_skl(
4438 boot_cpu_data.x86_model == INTEL_FAM6_SKYLAKE_X);
4439
4440 if (boot_cpu_has(X86_FEATURE_TSX_FORCE_ABORT)) {
4441 x86_pmu.flags |= PMU_FL_TFA;
4442 x86_pmu.get_event_constraints = tfa_get_event_constraints;
4443 x86_pmu.enable_all = intel_tfa_pmu_enable_all;
4444 x86_pmu.commit_scheduling = intel_tfa_commit_scheduling;
4445 intel_pmu_attrs[1] = &dev_attr_allow_tsx_force_abort.attr.attr;
4446 }
4447
4448 pr_cont("Skylake events, ");
4449 name = "skylake";
4450 break;
4451
4452 default:
4453 switch (x86_pmu.version) {
4454 case 1:
4455 x86_pmu.event_constraints = intel_v1_event_constraints;
4456 pr_cont("generic architected perfmon v1, ");
4457 name = "generic_arch_v1";
4458 break;
4459 default:
4460 /*
4461 * default constraints for v2 and up
4462 */
4463 x86_pmu.event_constraints = intel_gen_event_constraints;
4464 pr_cont("generic architected perfmon, ");
4465 name = "generic_arch_v2+";
4466 break;
4467 }
4468 }
4469
4470 snprintf(pmu_name_str, sizeof pmu_name_str, "%s", name);
4471
4472 if (version >= 2 && extra_attr) {
4473 x86_pmu.format_attrs = merge_attr(intel_arch3_formats_attr,
4474 extra_attr);
4475 WARN_ON(!x86_pmu.format_attrs);
4476 }
4477
4478 if (x86_pmu.num_counters > INTEL_PMC_MAX_GENERIC) {
4479 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
4480 x86_pmu.num_counters, INTEL_PMC_MAX_GENERIC);
4481 x86_pmu.num_counters = INTEL_PMC_MAX_GENERIC;
4482 }
4483 x86_pmu.intel_ctrl = (1ULL << x86_pmu.num_counters) - 1;
4484
4485 if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) {
4486 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
4487 x86_pmu.num_counters_fixed, INTEL_PMC_MAX_FIXED);
4488 x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED;
4489 }
4490
4491 x86_pmu.intel_ctrl |=
4492 ((1LL << x86_pmu.num_counters_fixed)-1) << INTEL_PMC_IDX_FIXED;
4493
4494 if (x86_pmu.event_constraints) {
4495 /*
4496 * event on fixed counter2 (REF_CYCLES) only works on this
4497 * counter, so do not extend mask to generic counters
4498 */
4499 for_each_event_constraint(c, x86_pmu.event_constraints) {
4500 if (c->cmask == FIXED_EVENT_FLAGS
4501 && c->idxmsk64 != INTEL_PMC_MSK_FIXED_REF_CYCLES) {
4502 c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1;
4503 }
4504 c->idxmsk64 &=
4505 ~(~0ULL << (INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed));
4506 c->weight = hweight64(c->idxmsk64);
4507 }
4508 }
4509
4510 /*
4511 * Access LBR MSR may cause #GP under certain circumstances.
4512 * E.g. KVM doesn't support LBR MSR
4513 * Check all LBT MSR here.
4514 * Disable LBR access if any LBR MSRs can not be accessed.
4515 */
4516 if (x86_pmu.lbr_nr && !check_msr(x86_pmu.lbr_tos, 0x3UL))
4517 x86_pmu.lbr_nr = 0;
4518 for (i = 0; i < x86_pmu.lbr_nr; i++) {
4519 if (!(check_msr(x86_pmu.lbr_from + i, 0xffffUL) &&
4520 check_msr(x86_pmu.lbr_to + i, 0xffffUL)))
4521 x86_pmu.lbr_nr = 0;
4522 }
4523
4524 x86_pmu.caps_attrs = intel_pmu_caps_attrs;
4525
4526 if (x86_pmu.lbr_nr) {
4527 x86_pmu.caps_attrs = merge_attr(x86_pmu.caps_attrs, lbr_attrs);
4528 pr_cont("%d-deep LBR, ", x86_pmu.lbr_nr);
4529 }
4530
4531 /*
4532 * Access extra MSR may cause #GP under certain circumstances.
4533 * E.g. KVM doesn't support offcore event
4534 * Check all extra_regs here.
4535 */
4536 if (x86_pmu.extra_regs) {
4537 for (er = x86_pmu.extra_regs; er->msr; er++) {
4538 er->extra_msr_access = check_msr(er->msr, 0x11UL);
4539 /* Disable LBR select mapping */
4540 if ((er->idx == EXTRA_REG_LBR) && !er->extra_msr_access)
4541 x86_pmu.lbr_sel_map = NULL;
4542 }
4543 }
4544
4545 /* Support full width counters using alternative MSR range */
4546 if (x86_pmu.intel_cap.full_width_write) {
4547 x86_pmu.max_period = x86_pmu.cntval_mask >> 1;
4548 x86_pmu.perfctr = MSR_IA32_PMC0;
4549 pr_cont("full-width counters, ");
4550 }
4551
4552 kfree(to_free);
4553 return 0;
4554 }
4555
4556 /*
4557 * HT bug: phase 2 init
4558 * Called once we have valid topology information to check
4559 * whether or not HT is enabled
4560 * If HT is off, then we disable the workaround
4561 */
4562 static __init int fixup_ht_bug(void)
4563 {
4564 int c;
4565 /*
4566 * problem not present on this CPU model, nothing to do
4567 */
4568 if (!(x86_pmu.flags & PMU_FL_EXCL_ENABLED))
4569 return 0;
4570
4571 if (topology_max_smt_threads() > 1) {
4572 pr_info("PMU erratum BJ122, BV98, HSD29 worked around, HT is on\n");
4573 return 0;
4574 }
4575
4576 cpus_read_lock();
4577
4578 hardlockup_detector_perf_stop();
4579
4580 x86_pmu.flags &= ~(PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED);
4581
4582 x86_pmu.start_scheduling = NULL;
4583 x86_pmu.commit_scheduling = NULL;
4584 x86_pmu.stop_scheduling = NULL;
4585
4586 hardlockup_detector_perf_restart();
4587
4588 for_each_online_cpu(c)
4589 free_excl_cntrs(&per_cpu(cpu_hw_events, c));
4590
4591 cpus_read_unlock();
4592 pr_info("PMU erratum BJ122, BV98, HSD29 workaround disabled, HT off\n");
4593 return 0;
4594 }
4595 subsys_initcall(fixup_ht_bug)
Linux with added FPC-III support