projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
perf/x86: Add model numbers for Kabylake CPUs
[deliverable/linux.git] / arch / x86 / events / intel / core.c
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/apic.h>
20
21 #include "../perf_event.h"
22
23 /*
24 * Intel PerfMon, used on Core and later.
25 */
26 static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly =
27 {
28 [PERF_COUNT_HW_CPU_CYCLES] = 0x003c,
29 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
30 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e,
31 [PERF_COUNT_HW_CACHE_MISSES] = 0x412e,
32 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4,
33 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5,
34 [PERF_COUNT_HW_BUS_CYCLES] = 0x013c,
35 [PERF_COUNT_HW_REF_CPU_CYCLES] = 0x0300, /* pseudo-encoding */
36 };
37
38 static struct event_constraint intel_core_event_constraints[] __read_mostly =
39 {
40 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
41 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
42 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
43 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
44 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
45 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */
46 EVENT_CONSTRAINT_END
47 };
48
49 static struct event_constraint intel_core2_event_constraints[] __read_mostly =
50 {
51 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
52 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
53 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
54 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
55 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
56 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
57 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
58 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
59 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
60 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
61 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
62 INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */
63 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
64 EVENT_CONSTRAINT_END
65 };
66
67 static struct event_constraint intel_nehalem_event_constraints[] __read_mostly =
68 {
69 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
70 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
71 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
72 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
73 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
74 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
75 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
76 INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */
77 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
78 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
79 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
80 EVENT_CONSTRAINT_END
81 };
82
83 static struct extra_reg intel_nehalem_extra_regs[] __read_mostly =
84 {
85 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
86 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
87 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
88 EVENT_EXTRA_END
89 };
90
91 static struct event_constraint intel_westmere_event_constraints[] __read_mostly =
92 {
93 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
94 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
95 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
96 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
97 INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */
98 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
99 INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */
100 EVENT_CONSTRAINT_END
101 };
102
103 static struct event_constraint intel_snb_event_constraints[] __read_mostly =
104 {
105 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
106 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
107 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
108 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
109 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
110 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
111 INTEL_UEVENT_CONSTRAINT(0x06a3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
112 INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */
113 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
114 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
115 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
116 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
117
118 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
119 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
120 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
121 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
122
123 EVENT_CONSTRAINT_END
124 };
125
126 static struct event_constraint intel_ivb_event_constraints[] __read_mostly =
127 {
128 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
129 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
130 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
131 INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */
132 INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMTPY */
133 INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */
134 INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_LDM_PENDING */
135 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
136 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
137 INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */
138 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
139 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
140 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
141
142 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
143 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
144 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
145 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
146
147 EVENT_CONSTRAINT_END
148 };
149
150 static struct extra_reg intel_westmere_extra_regs[] __read_mostly =
151 {
152 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
153 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
154 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1),
155 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
156 EVENT_EXTRA_END
157 };
158
159 static struct event_constraint intel_v1_event_constraints[] __read_mostly =
160 {
161 EVENT_CONSTRAINT_END
162 };
163
164 static struct event_constraint intel_gen_event_constraints[] __read_mostly =
165 {
166 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
167 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
168 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
169 EVENT_CONSTRAINT_END
170 };
171
172 static struct event_constraint intel_slm_event_constraints[] __read_mostly =
173 {
174 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
175 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
176 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */
177 EVENT_CONSTRAINT_END
178 };
179
180 struct event_constraint intel_skl_event_constraints[] = {
181 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
182 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
183 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
184 INTEL_UEVENT_CONSTRAINT(0x1c0, 0x2), /* INST_RETIRED.PREC_DIST */
185 EVENT_CONSTRAINT_END
186 };
187
188 static struct extra_reg intel_knl_extra_regs[] __read_mostly = {
189 INTEL_UEVENT_EXTRA_REG(0x01b7,
190 MSR_OFFCORE_RSP_0, 0x7f9ffbffffull, RSP_0),
191 INTEL_UEVENT_EXTRA_REG(0x02b7,
192 MSR_OFFCORE_RSP_1, 0x3f9ffbffffull, RSP_1),
193 EVENT_EXTRA_END
194 };
195
196 static struct extra_reg intel_snb_extra_regs[] __read_mostly = {
197 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
198 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0),
199 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1),
200 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
201 EVENT_EXTRA_END
202 };
203
204 static struct extra_reg intel_snbep_extra_regs[] __read_mostly = {
205 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
206 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
207 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
208 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
209 EVENT_EXTRA_END
210 };
211
212 static struct extra_reg intel_skl_extra_regs[] __read_mostly = {
213 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
214 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
215 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
216 /*
217 * Note the low 8 bits eventsel code is not a continuous field, containing
218 * some #GPing bits. These are masked out.
219 */
220 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE),
221 EVENT_EXTRA_END
222 };
223
224 EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3");
225 EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3");
226 EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2");
227
228 struct attribute *nhm_events_attrs[] = {
229 EVENT_PTR(mem_ld_nhm),
230 NULL,
231 };
232
233 struct attribute *snb_events_attrs[] = {
234 EVENT_PTR(mem_ld_snb),
235 EVENT_PTR(mem_st_snb),
236 NULL,
237 };
238
239 static struct event_constraint intel_hsw_event_constraints[] = {
240 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
241 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
242 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
243 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */
244 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
245 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
246 /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
247 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4),
248 /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
249 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4),
250 /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
251 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf),
252
253 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
254 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
255 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
256 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
257
258 EVENT_CONSTRAINT_END
259 };
260
261 struct event_constraint intel_bdw_event_constraints[] = {
262 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
263 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
264 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
265 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */
266 INTEL_UBIT_EVENT_CONSTRAINT(0x8a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_MISS */
267 EVENT_CONSTRAINT_END
268 };
269
270 static u64 intel_pmu_event_map(int hw_event)
271 {
272 return intel_perfmon_event_map[hw_event];
273 }
274
275 /*
276 * Notes on the events:
277 * - data reads do not include code reads (comparable to earlier tables)
278 * - data counts include speculative execution (except L1 write, dtlb, bpu)
279 * - remote node access includes remote memory, remote cache, remote mmio.
280 * - prefetches are not included in the counts.
281 * - icache miss does not include decoded icache
282 */
283
284 #define SKL_DEMAND_DATA_RD BIT_ULL(0)
285 #define SKL_DEMAND_RFO BIT_ULL(1)
286 #define SKL_ANY_RESPONSE BIT_ULL(16)
287 #define SKL_SUPPLIER_NONE BIT_ULL(17)
288 #define SKL_L3_MISS_LOCAL_DRAM BIT_ULL(26)
289 #define SKL_L3_MISS_REMOTE_HOP0_DRAM BIT_ULL(27)
290 #define SKL_L3_MISS_REMOTE_HOP1_DRAM BIT_ULL(28)
291 #define SKL_L3_MISS_REMOTE_HOP2P_DRAM BIT_ULL(29)
292 #define SKL_L3_MISS (SKL_L3_MISS_LOCAL_DRAM| \
293 SKL_L3_MISS_REMOTE_HOP0_DRAM| \
294 SKL_L3_MISS_REMOTE_HOP1_DRAM| \
295 SKL_L3_MISS_REMOTE_HOP2P_DRAM)
296 #define SKL_SPL_HIT BIT_ULL(30)
297 #define SKL_SNOOP_NONE BIT_ULL(31)
298 #define SKL_SNOOP_NOT_NEEDED BIT_ULL(32)
299 #define SKL_SNOOP_MISS BIT_ULL(33)
300 #define SKL_SNOOP_HIT_NO_FWD BIT_ULL(34)
301 #define SKL_SNOOP_HIT_WITH_FWD BIT_ULL(35)
302 #define SKL_SNOOP_HITM BIT_ULL(36)
303 #define SKL_SNOOP_NON_DRAM BIT_ULL(37)
304 #define SKL_ANY_SNOOP (SKL_SPL_HIT|SKL_SNOOP_NONE| \
305 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \
306 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \
307 SKL_SNOOP_HITM|SKL_SNOOP_NON_DRAM)
308 #define SKL_DEMAND_READ SKL_DEMAND_DATA_RD
309 #define SKL_SNOOP_DRAM (SKL_SNOOP_NONE| \
310 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \
311 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \
312 SKL_SNOOP_HITM|SKL_SPL_HIT)
313 #define SKL_DEMAND_WRITE SKL_DEMAND_RFO
314 #define SKL_LLC_ACCESS SKL_ANY_RESPONSE
315 #define SKL_L3_MISS_REMOTE (SKL_L3_MISS_REMOTE_HOP0_DRAM| \
316 SKL_L3_MISS_REMOTE_HOP1_DRAM| \
317 SKL_L3_MISS_REMOTE_HOP2P_DRAM)
318
319 static __initconst const u64 skl_hw_cache_event_ids
320 [PERF_COUNT_HW_CACHE_MAX]
321 [PERF_COUNT_HW_CACHE_OP_MAX]
322 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
323 {
324 [ C(L1D ) ] = {
325 [ C(OP_READ) ] = {
326 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */
327 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */
328 },
329 [ C(OP_WRITE) ] = {
330 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */
331 [ C(RESULT_MISS) ] = 0x0,
332 },
333 [ C(OP_PREFETCH) ] = {
334 [ C(RESULT_ACCESS) ] = 0x0,
335 [ C(RESULT_MISS) ] = 0x0,
336 },
337 },
338 [ C(L1I ) ] = {
339 [ C(OP_READ) ] = {
340 [ C(RESULT_ACCESS) ] = 0x0,
341 [ C(RESULT_MISS) ] = 0x283, /* ICACHE_64B.MISS */
342 },
343 [ C(OP_WRITE) ] = {
344 [ C(RESULT_ACCESS) ] = -1,
345 [ C(RESULT_MISS) ] = -1,
346 },
347 [ C(OP_PREFETCH) ] = {
348 [ C(RESULT_ACCESS) ] = 0x0,
349 [ C(RESULT_MISS) ] = 0x0,
350 },
351 },
352 [ C(LL ) ] = {
353 [ C(OP_READ) ] = {
354 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
355 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
356 },
357 [ C(OP_WRITE) ] = {
358 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
359 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
360 },
361 [ C(OP_PREFETCH) ] = {
362 [ C(RESULT_ACCESS) ] = 0x0,
363 [ C(RESULT_MISS) ] = 0x0,
364 },
365 },
366 [ C(DTLB) ] = {
367 [ C(OP_READ) ] = {
368 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */
369 [ C(RESULT_MISS) ] = 0x608, /* DTLB_LOAD_MISSES.WALK_COMPLETED */
370 },
371 [ C(OP_WRITE) ] = {
372 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */
373 [ C(RESULT_MISS) ] = 0x649, /* DTLB_STORE_MISSES.WALK_COMPLETED */
374 },
375 [ C(OP_PREFETCH) ] = {
376 [ C(RESULT_ACCESS) ] = 0x0,
377 [ C(RESULT_MISS) ] = 0x0,
378 },
379 },
380 [ C(ITLB) ] = {
381 [ C(OP_READ) ] = {
382 [ C(RESULT_ACCESS) ] = 0x2085, /* ITLB_MISSES.STLB_HIT */
383 [ C(RESULT_MISS) ] = 0xe85, /* ITLB_MISSES.WALK_COMPLETED */
384 },
385 [ C(OP_WRITE) ] = {
386 [ C(RESULT_ACCESS) ] = -1,
387 [ C(RESULT_MISS) ] = -1,
388 },
389 [ C(OP_PREFETCH) ] = {
390 [ C(RESULT_ACCESS) ] = -1,
391 [ C(RESULT_MISS) ] = -1,
392 },
393 },
394 [ C(BPU ) ] = {
395 [ C(OP_READ) ] = {
396 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */
397 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */
398 },
399 [ C(OP_WRITE) ] = {
400 [ C(RESULT_ACCESS) ] = -1,
401 [ C(RESULT_MISS) ] = -1,
402 },
403 [ C(OP_PREFETCH) ] = {
404 [ C(RESULT_ACCESS) ] = -1,
405 [ C(RESULT_MISS) ] = -1,
406 },
407 },
408 [ C(NODE) ] = {
409 [ C(OP_READ) ] = {
410 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
411 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
412 },
413 [ C(OP_WRITE) ] = {
414 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
415 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
416 },
417 [ C(OP_PREFETCH) ] = {
418 [ C(RESULT_ACCESS) ] = 0x0,
419 [ C(RESULT_MISS) ] = 0x0,
420 },
421 },
422 };
423
424 static __initconst const u64 skl_hw_cache_extra_regs
425 [PERF_COUNT_HW_CACHE_MAX]
426 [PERF_COUNT_HW_CACHE_OP_MAX]
427 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
428 {
429 [ C(LL ) ] = {
430 [ C(OP_READ) ] = {
431 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ|
432 SKL_LLC_ACCESS|SKL_ANY_SNOOP,
433 [ C(RESULT_MISS) ] = SKL_DEMAND_READ|
434 SKL_L3_MISS|SKL_ANY_SNOOP|
435 SKL_SUPPLIER_NONE,
436 },
437 [ C(OP_WRITE) ] = {
438 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE|
439 SKL_LLC_ACCESS|SKL_ANY_SNOOP,
440 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE|
441 SKL_L3_MISS|SKL_ANY_SNOOP|
442 SKL_SUPPLIER_NONE,
443 },
444 [ C(OP_PREFETCH) ] = {
445 [ C(RESULT_ACCESS) ] = 0x0,
446 [ C(RESULT_MISS) ] = 0x0,
447 },
448 },
449 [ C(NODE) ] = {
450 [ C(OP_READ) ] = {
451 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ|
452 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM,
453 [ C(RESULT_MISS) ] = SKL_DEMAND_READ|
454 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM,
455 },
456 [ C(OP_WRITE) ] = {
457 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE|
458 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM,
459 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE|
460 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM,
461 },
462 [ C(OP_PREFETCH) ] = {
463 [ C(RESULT_ACCESS) ] = 0x0,
464 [ C(RESULT_MISS) ] = 0x0,
465 },
466 },
467 };
468
469 #define SNB_DMND_DATA_RD (1ULL << 0)
470 #define SNB_DMND_RFO (1ULL << 1)
471 #define SNB_DMND_IFETCH (1ULL << 2)
472 #define SNB_DMND_WB (1ULL << 3)
473 #define SNB_PF_DATA_RD (1ULL << 4)
474 #define SNB_PF_RFO (1ULL << 5)
475 #define SNB_PF_IFETCH (1ULL << 6)
476 #define SNB_LLC_DATA_RD (1ULL << 7)
477 #define SNB_LLC_RFO (1ULL << 8)
478 #define SNB_LLC_IFETCH (1ULL << 9)
479 #define SNB_BUS_LOCKS (1ULL << 10)
480 #define SNB_STRM_ST (1ULL << 11)
481 #define SNB_OTHER (1ULL << 15)
482 #define SNB_RESP_ANY (1ULL << 16)
483 #define SNB_NO_SUPP (1ULL << 17)
484 #define SNB_LLC_HITM (1ULL << 18)
485 #define SNB_LLC_HITE (1ULL << 19)
486 #define SNB_LLC_HITS (1ULL << 20)
487 #define SNB_LLC_HITF (1ULL << 21)
488 #define SNB_LOCAL (1ULL << 22)
489 #define SNB_REMOTE (0xffULL << 23)
490 #define SNB_SNP_NONE (1ULL << 31)
491 #define SNB_SNP_NOT_NEEDED (1ULL << 32)
492 #define SNB_SNP_MISS (1ULL << 33)
493 #define SNB_NO_FWD (1ULL << 34)
494 #define SNB_SNP_FWD (1ULL << 35)
495 #define SNB_HITM (1ULL << 36)
496 #define SNB_NON_DRAM (1ULL << 37)
497
498 #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD)
499 #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO)
500 #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
501
502 #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \
503 SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \
504 SNB_HITM)
505
506 #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY)
507 #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY)
508
509 #define SNB_L3_ACCESS SNB_RESP_ANY
510 #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM)
511
512 static __initconst const u64 snb_hw_cache_extra_regs
513 [PERF_COUNT_HW_CACHE_MAX]
514 [PERF_COUNT_HW_CACHE_OP_MAX]
515 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
516 {
517 [ C(LL ) ] = {
518 [ C(OP_READ) ] = {
519 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS,
520 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_L3_MISS,
521 },
522 [ C(OP_WRITE) ] = {
523 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS,
524 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_L3_MISS,
525 },
526 [ C(OP_PREFETCH) ] = {
527 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS,
528 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_L3_MISS,
529 },
530 },
531 [ C(NODE) ] = {
532 [ C(OP_READ) ] = {
533 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY,
534 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_DRAM_REMOTE,
535 },
536 [ C(OP_WRITE) ] = {
537 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY,
538 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE,
539 },
540 [ C(OP_PREFETCH) ] = {
541 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY,
542 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE,
543 },
544 },
545 };
546
547 static __initconst const u64 snb_hw_cache_event_ids
548 [PERF_COUNT_HW_CACHE_MAX]
549 [PERF_COUNT_HW_CACHE_OP_MAX]
550 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
551 {
552 [ C(L1D) ] = {
553 [ C(OP_READ) ] = {
554 [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */
555 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */
556 },
557 [ C(OP_WRITE) ] = {
558 [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */
559 [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */
560 },
561 [ C(OP_PREFETCH) ] = {
562 [ C(RESULT_ACCESS) ] = 0x0,
563 [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */
564 },
565 },
566 [ C(L1I ) ] = {
567 [ C(OP_READ) ] = {
568 [ C(RESULT_ACCESS) ] = 0x0,
569 [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */
570 },
571 [ C(OP_WRITE) ] = {
572 [ C(RESULT_ACCESS) ] = -1,
573 [ C(RESULT_MISS) ] = -1,
574 },
575 [ C(OP_PREFETCH) ] = {
576 [ C(RESULT_ACCESS) ] = 0x0,
577 [ C(RESULT_MISS) ] = 0x0,
578 },
579 },
580 [ C(LL ) ] = {
581 [ C(OP_READ) ] = {
582 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
583 [ C(RESULT_ACCESS) ] = 0x01b7,
584 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
585 [ C(RESULT_MISS) ] = 0x01b7,
586 },
587 [ C(OP_WRITE) ] = {
588 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
589 [ C(RESULT_ACCESS) ] = 0x01b7,
590 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
591 [ C(RESULT_MISS) ] = 0x01b7,
592 },
593 [ C(OP_PREFETCH) ] = {
594 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
595 [ C(RESULT_ACCESS) ] = 0x01b7,
596 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
597 [ C(RESULT_MISS) ] = 0x01b7,
598 },
599 },
600 [ C(DTLB) ] = {
601 [ C(OP_READ) ] = {
602 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */
603 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */
604 },
605 [ C(OP_WRITE) ] = {
606 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */
607 [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
608 },
609 [ C(OP_PREFETCH) ] = {
610 [ C(RESULT_ACCESS) ] = 0x0,
611 [ C(RESULT_MISS) ] = 0x0,
612 },
613 },
614 [ C(ITLB) ] = {
615 [ C(OP_READ) ] = {
616 [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */
617 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */
618 },
619 [ C(OP_WRITE) ] = {
620 [ C(RESULT_ACCESS) ] = -1,
621 [ C(RESULT_MISS) ] = -1,
622 },
623 [ C(OP_PREFETCH) ] = {
624 [ C(RESULT_ACCESS) ] = -1,
625 [ C(RESULT_MISS) ] = -1,
626 },
627 },
628 [ C(BPU ) ] = {
629 [ C(OP_READ) ] = {
630 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
631 [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
632 },
633 [ C(OP_WRITE) ] = {
634 [ C(RESULT_ACCESS) ] = -1,
635 [ C(RESULT_MISS) ] = -1,
636 },
637 [ C(OP_PREFETCH) ] = {
638 [ C(RESULT_ACCESS) ] = -1,
639 [ C(RESULT_MISS) ] = -1,
640 },
641 },
642 [ C(NODE) ] = {
643 [ C(OP_READ) ] = {
644 [ C(RESULT_ACCESS) ] = 0x01b7,
645 [ C(RESULT_MISS) ] = 0x01b7,
646 },
647 [ C(OP_WRITE) ] = {
648 [ C(RESULT_ACCESS) ] = 0x01b7,
649 [ C(RESULT_MISS) ] = 0x01b7,
650 },
651 [ C(OP_PREFETCH) ] = {
652 [ C(RESULT_ACCESS) ] = 0x01b7,
653 [ C(RESULT_MISS) ] = 0x01b7,
654 },
655 },
656
657 };
658
659 /*
660 * Notes on the events:
661 * - data reads do not include code reads (comparable to earlier tables)
662 * - data counts include speculative execution (except L1 write, dtlb, bpu)
663 * - remote node access includes remote memory, remote cache, remote mmio.
664 * - prefetches are not included in the counts because they are not
665 * reliably counted.
666 */
667
668 #define HSW_DEMAND_DATA_RD BIT_ULL(0)
669 #define HSW_DEMAND_RFO BIT_ULL(1)
670 #define HSW_ANY_RESPONSE BIT_ULL(16)
671 #define HSW_SUPPLIER_NONE BIT_ULL(17)
672 #define HSW_L3_MISS_LOCAL_DRAM BIT_ULL(22)
673 #define HSW_L3_MISS_REMOTE_HOP0 BIT_ULL(27)
674 #define HSW_L3_MISS_REMOTE_HOP1 BIT_ULL(28)
675 #define HSW_L3_MISS_REMOTE_HOP2P BIT_ULL(29)
676 #define HSW_L3_MISS (HSW_L3_MISS_LOCAL_DRAM| \
677 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \
678 HSW_L3_MISS_REMOTE_HOP2P)
679 #define HSW_SNOOP_NONE BIT_ULL(31)
680 #define HSW_SNOOP_NOT_NEEDED BIT_ULL(32)
681 #define HSW_SNOOP_MISS BIT_ULL(33)
682 #define HSW_SNOOP_HIT_NO_FWD BIT_ULL(34)
683 #define HSW_SNOOP_HIT_WITH_FWD BIT_ULL(35)
684 #define HSW_SNOOP_HITM BIT_ULL(36)
685 #define HSW_SNOOP_NON_DRAM BIT_ULL(37)
686 #define HSW_ANY_SNOOP (HSW_SNOOP_NONE| \
687 HSW_SNOOP_NOT_NEEDED|HSW_SNOOP_MISS| \
688 HSW_SNOOP_HIT_NO_FWD|HSW_SNOOP_HIT_WITH_FWD| \
689 HSW_SNOOP_HITM|HSW_SNOOP_NON_DRAM)
690 #define HSW_SNOOP_DRAM (HSW_ANY_SNOOP & ~HSW_SNOOP_NON_DRAM)
691 #define HSW_DEMAND_READ HSW_DEMAND_DATA_RD
692 #define HSW_DEMAND_WRITE HSW_DEMAND_RFO
693 #define HSW_L3_MISS_REMOTE (HSW_L3_MISS_REMOTE_HOP0|\
694 HSW_L3_MISS_REMOTE_HOP1|HSW_L3_MISS_REMOTE_HOP2P)
695 #define HSW_LLC_ACCESS HSW_ANY_RESPONSE
696
697 #define BDW_L3_MISS_LOCAL BIT(26)
698 #define BDW_L3_MISS (BDW_L3_MISS_LOCAL| \
699 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \
700 HSW_L3_MISS_REMOTE_HOP2P)
701
702
703 static __initconst const u64 hsw_hw_cache_event_ids
704 [PERF_COUNT_HW_CACHE_MAX]
705 [PERF_COUNT_HW_CACHE_OP_MAX]
706 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
707 {
708 [ C(L1D ) ] = {
709 [ C(OP_READ) ] = {
710 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
711 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */
712 },
713 [ C(OP_WRITE) ] = {
714 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
715 [ C(RESULT_MISS) ] = 0x0,
716 },
717 [ C(OP_PREFETCH) ] = {
718 [ C(RESULT_ACCESS) ] = 0x0,
719 [ C(RESULT_MISS) ] = 0x0,
720 },
721 },
722 [ C(L1I ) ] = {
723 [ C(OP_READ) ] = {
724 [ C(RESULT_ACCESS) ] = 0x0,
725 [ C(RESULT_MISS) ] = 0x280, /* ICACHE.MISSES */
726 },
727 [ C(OP_WRITE) ] = {
728 [ C(RESULT_ACCESS) ] = -1,
729 [ C(RESULT_MISS) ] = -1,
730 },
731 [ C(OP_PREFETCH) ] = {
732 [ C(RESULT_ACCESS) ] = 0x0,
733 [ C(RESULT_MISS) ] = 0x0,
734 },
735 },
736 [ C(LL ) ] = {
737 [ C(OP_READ) ] = {
738 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
739 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
740 },
741 [ C(OP_WRITE) ] = {
742 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
743 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
744 },
745 [ C(OP_PREFETCH) ] = {
746 [ C(RESULT_ACCESS) ] = 0x0,
747 [ C(RESULT_MISS) ] = 0x0,
748 },
749 },
750 [ C(DTLB) ] = {
751 [ C(OP_READ) ] = {
752 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */
753 [ C(RESULT_MISS) ] = 0x108, /* DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK */
754 },
755 [ C(OP_WRITE) ] = {
756 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */
757 [ C(RESULT_MISS) ] = 0x149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
758 },
759 [ C(OP_PREFETCH) ] = {
760 [ C(RESULT_ACCESS) ] = 0x0,
761 [ C(RESULT_MISS) ] = 0x0,
762 },
763 },
764 [ C(ITLB) ] = {
765 [ C(OP_READ) ] = {
766 [ C(RESULT_ACCESS) ] = 0x6085, /* ITLB_MISSES.STLB_HIT */
767 [ C(RESULT_MISS) ] = 0x185, /* ITLB_MISSES.MISS_CAUSES_A_WALK */
768 },
769 [ C(OP_WRITE) ] = {
770 [ C(RESULT_ACCESS) ] = -1,
771 [ C(RESULT_MISS) ] = -1,
772 },
773 [ C(OP_PREFETCH) ] = {
774 [ C(RESULT_ACCESS) ] = -1,
775 [ C(RESULT_MISS) ] = -1,
776 },
777 },
778 [ C(BPU ) ] = {
779 [ C(OP_READ) ] = {
780 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */
781 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */
782 },
783 [ C(OP_WRITE) ] = {
784 [ C(RESULT_ACCESS) ] = -1,
785 [ C(RESULT_MISS) ] = -1,
786 },
787 [ C(OP_PREFETCH) ] = {
788 [ C(RESULT_ACCESS) ] = -1,
789 [ C(RESULT_MISS) ] = -1,
790 },
791 },
792 [ C(NODE) ] = {
793 [ C(OP_READ) ] = {
794 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
795 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
796 },
797 [ C(OP_WRITE) ] = {
798 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */
799 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */
800 },
801 [ C(OP_PREFETCH) ] = {
802 [ C(RESULT_ACCESS) ] = 0x0,
803 [ C(RESULT_MISS) ] = 0x0,
804 },
805 },
806 };
807
808 static __initconst const u64 hsw_hw_cache_extra_regs
809 [PERF_COUNT_HW_CACHE_MAX]
810 [PERF_COUNT_HW_CACHE_OP_MAX]
811 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
812 {
813 [ C(LL ) ] = {
814 [ C(OP_READ) ] = {
815 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ|
816 HSW_LLC_ACCESS,
817 [ C(RESULT_MISS) ] = HSW_DEMAND_READ|
818 HSW_L3_MISS|HSW_ANY_SNOOP,
819 },
820 [ C(OP_WRITE) ] = {
821 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE|
822 HSW_LLC_ACCESS,
823 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE|
824 HSW_L3_MISS|HSW_ANY_SNOOP,
825 },
826 [ C(OP_PREFETCH) ] = {
827 [ C(RESULT_ACCESS) ] = 0x0,
828 [ C(RESULT_MISS) ] = 0x0,
829 },
830 },
831 [ C(NODE) ] = {
832 [ C(OP_READ) ] = {
833 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ|
834 HSW_L3_MISS_LOCAL_DRAM|
835 HSW_SNOOP_DRAM,
836 [ C(RESULT_MISS) ] = HSW_DEMAND_READ|
837 HSW_L3_MISS_REMOTE|
838 HSW_SNOOP_DRAM,
839 },
840 [ C(OP_WRITE) ] = {
841 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE|
842 HSW_L3_MISS_LOCAL_DRAM|
843 HSW_SNOOP_DRAM,
844 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE|
845 HSW_L3_MISS_REMOTE|
846 HSW_SNOOP_DRAM,
847 },
848 [ C(OP_PREFETCH) ] = {
849 [ C(RESULT_ACCESS) ] = 0x0,
850 [ C(RESULT_MISS) ] = 0x0,
851 },
852 },
853 };
854
855 static __initconst const u64 westmere_hw_cache_event_ids
856 [PERF_COUNT_HW_CACHE_MAX]
857 [PERF_COUNT_HW_CACHE_OP_MAX]
858 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
859 {
860 [ C(L1D) ] = {
861 [ C(OP_READ) ] = {
862 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
863 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
864 },
865 [ C(OP_WRITE) ] = {
866 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
867 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
868 },
869 [ C(OP_PREFETCH) ] = {
870 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
871 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
872 },
873 },
874 [ C(L1I ) ] = {
875 [ C(OP_READ) ] = {
876 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
877 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
878 },
879 [ C(OP_WRITE) ] = {
880 [ C(RESULT_ACCESS) ] = -1,
881 [ C(RESULT_MISS) ] = -1,
882 },
883 [ C(OP_PREFETCH) ] = {
884 [ C(RESULT_ACCESS) ] = 0x0,
885 [ C(RESULT_MISS) ] = 0x0,
886 },
887 },
888 [ C(LL ) ] = {
889 [ C(OP_READ) ] = {
890 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
891 [ C(RESULT_ACCESS) ] = 0x01b7,
892 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
893 [ C(RESULT_MISS) ] = 0x01b7,
894 },
895 /*
896 * Use RFO, not WRITEBACK, because a write miss would typically occur
897 * on RFO.
898 */
899 [ C(OP_WRITE) ] = {
900 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
901 [ C(RESULT_ACCESS) ] = 0x01b7,
902 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
903 [ C(RESULT_MISS) ] = 0x01b7,
904 },
905 [ C(OP_PREFETCH) ] = {
906 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
907 [ C(RESULT_ACCESS) ] = 0x01b7,
908 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
909 [ C(RESULT_MISS) ] = 0x01b7,
910 },
911 },
912 [ C(DTLB) ] = {
913 [ C(OP_READ) ] = {
914 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
915 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
916 },
917 [ C(OP_WRITE) ] = {
918 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
919 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
920 },
921 [ C(OP_PREFETCH) ] = {
922 [ C(RESULT_ACCESS) ] = 0x0,
923 [ C(RESULT_MISS) ] = 0x0,
924 },
925 },
926 [ C(ITLB) ] = {
927 [ C(OP_READ) ] = {
928 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
929 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */
930 },
931 [ C(OP_WRITE) ] = {
932 [ C(RESULT_ACCESS) ] = -1,
933 [ C(RESULT_MISS) ] = -1,
934 },
935 [ C(OP_PREFETCH) ] = {
936 [ C(RESULT_ACCESS) ] = -1,
937 [ C(RESULT_MISS) ] = -1,
938 },
939 },
940 [ C(BPU ) ] = {
941 [ C(OP_READ) ] = {
942 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
943 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
944 },
945 [ C(OP_WRITE) ] = {
946 [ C(RESULT_ACCESS) ] = -1,
947 [ C(RESULT_MISS) ] = -1,
948 },
949 [ C(OP_PREFETCH) ] = {
950 [ C(RESULT_ACCESS) ] = -1,
951 [ C(RESULT_MISS) ] = -1,
952 },
953 },
954 [ C(NODE) ] = {
955 [ C(OP_READ) ] = {
956 [ C(RESULT_ACCESS) ] = 0x01b7,
957 [ C(RESULT_MISS) ] = 0x01b7,
958 },
959 [ C(OP_WRITE) ] = {
960 [ C(RESULT_ACCESS) ] = 0x01b7,
961 [ C(RESULT_MISS) ] = 0x01b7,
962 },
963 [ C(OP_PREFETCH) ] = {
964 [ C(RESULT_ACCESS) ] = 0x01b7,
965 [ C(RESULT_MISS) ] = 0x01b7,
966 },
967 },
968 };
969
970 /*
971 * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits;
972 * See IA32 SDM Vol 3B 30.6.1.3
973 */
974
975 #define NHM_DMND_DATA_RD (1 << 0)
976 #define NHM_DMND_RFO (1 << 1)
977 #define NHM_DMND_IFETCH (1 << 2)
978 #define NHM_DMND_WB (1 << 3)
979 #define NHM_PF_DATA_RD (1 << 4)
980 #define NHM_PF_DATA_RFO (1 << 5)
981 #define NHM_PF_IFETCH (1 << 6)
982 #define NHM_OFFCORE_OTHER (1 << 7)
983 #define NHM_UNCORE_HIT (1 << 8)
984 #define NHM_OTHER_CORE_HIT_SNP (1 << 9)
985 #define NHM_OTHER_CORE_HITM (1 << 10)
986 /* reserved */
987 #define NHM_REMOTE_CACHE_FWD (1 << 12)
988 #define NHM_REMOTE_DRAM (1 << 13)
989 #define NHM_LOCAL_DRAM (1 << 14)
990 #define NHM_NON_DRAM (1 << 15)
991
992 #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD)
993 #define NHM_REMOTE (NHM_REMOTE_DRAM)
994
995 #define NHM_DMND_READ (NHM_DMND_DATA_RD)
996 #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB)
997 #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO)
998
999 #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM)
1000 #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD)
1001 #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS)
1002
1003 static __initconst const u64 nehalem_hw_cache_extra_regs
1004 [PERF_COUNT_HW_CACHE_MAX]
1005 [PERF_COUNT_HW_CACHE_OP_MAX]
1006 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1007 {
1008 [ C(LL ) ] = {
1009 [ C(OP_READ) ] = {
1010 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS,
1011 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS,
1012 },
1013 [ C(OP_WRITE) ] = {
1014 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS,
1015 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS,
1016 },
1017 [ C(OP_PREFETCH) ] = {
1018 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS,
1019 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS,
1020 },
1021 },
1022 [ C(NODE) ] = {
1023 [ C(OP_READ) ] = {
1024 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE,
1025 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE,
1026 },
1027 [ C(OP_WRITE) ] = {
1028 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE,
1029 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE,
1030 },
1031 [ C(OP_PREFETCH) ] = {
1032 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE,
1033 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE,
1034 },
1035 },
1036 };
1037
1038 static __initconst const u64 nehalem_hw_cache_event_ids
1039 [PERF_COUNT_HW_CACHE_MAX]
1040 [PERF_COUNT_HW_CACHE_OP_MAX]
1041 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1042 {
1043 [ C(L1D) ] = {
1044 [ C(OP_READ) ] = {
1045 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */
1046 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */
1047 },
1048 [ C(OP_WRITE) ] = {
1049 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */
1050 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */
1051 },
1052 [ C(OP_PREFETCH) ] = {
1053 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */
1054 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */
1055 },
1056 },
1057 [ C(L1I ) ] = {
1058 [ C(OP_READ) ] = {
1059 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
1060 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
1061 },
1062 [ C(OP_WRITE) ] = {
1063 [ C(RESULT_ACCESS) ] = -1,
1064 [ C(RESULT_MISS) ] = -1,
1065 },
1066 [ C(OP_PREFETCH) ] = {
1067 [ C(RESULT_ACCESS) ] = 0x0,
1068 [ C(RESULT_MISS) ] = 0x0,
1069 },
1070 },
1071 [ C(LL ) ] = {
1072 [ C(OP_READ) ] = {
1073 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1074 [ C(RESULT_ACCESS) ] = 0x01b7,
1075 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
1076 [ C(RESULT_MISS) ] = 0x01b7,
1077 },
1078 /*
1079 * Use RFO, not WRITEBACK, because a write miss would typically occur
1080 * on RFO.
1081 */
1082 [ C(OP_WRITE) ] = {
1083 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1084 [ C(RESULT_ACCESS) ] = 0x01b7,
1085 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1086 [ C(RESULT_MISS) ] = 0x01b7,
1087 },
1088 [ C(OP_PREFETCH) ] = {
1089 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1090 [ C(RESULT_ACCESS) ] = 0x01b7,
1091 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1092 [ C(RESULT_MISS) ] = 0x01b7,
1093 },
1094 },
1095 [ C(DTLB) ] = {
1096 [ C(OP_READ) ] = {
1097 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
1098 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */
1099 },
1100 [ C(OP_WRITE) ] = {
1101 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
1102 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */
1103 },
1104 [ C(OP_PREFETCH) ] = {
1105 [ C(RESULT_ACCESS) ] = 0x0,
1106 [ C(RESULT_MISS) ] = 0x0,
1107 },
1108 },
1109 [ C(ITLB) ] = {
1110 [ C(OP_READ) ] = {
1111 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */
1112 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */
1113 },
1114 [ C(OP_WRITE) ] = {
1115 [ C(RESULT_ACCESS) ] = -1,
1116 [ C(RESULT_MISS) ] = -1,
1117 },
1118 [ C(OP_PREFETCH) ] = {
1119 [ C(RESULT_ACCESS) ] = -1,
1120 [ C(RESULT_MISS) ] = -1,
1121 },
1122 },
1123 [ C(BPU ) ] = {
1124 [ C(OP_READ) ] = {
1125 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1126 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */
1127 },
1128 [ C(OP_WRITE) ] = {
1129 [ C(RESULT_ACCESS) ] = -1,
1130 [ C(RESULT_MISS) ] = -1,
1131 },
1132 [ C(OP_PREFETCH) ] = {
1133 [ C(RESULT_ACCESS) ] = -1,
1134 [ C(RESULT_MISS) ] = -1,
1135 },
1136 },
1137 [ C(NODE) ] = {
1138 [ C(OP_READ) ] = {
1139 [ C(RESULT_ACCESS) ] = 0x01b7,
1140 [ C(RESULT_MISS) ] = 0x01b7,
1141 },
1142 [ C(OP_WRITE) ] = {
1143 [ C(RESULT_ACCESS) ] = 0x01b7,
1144 [ C(RESULT_MISS) ] = 0x01b7,
1145 },
1146 [ C(OP_PREFETCH) ] = {
1147 [ C(RESULT_ACCESS) ] = 0x01b7,
1148 [ C(RESULT_MISS) ] = 0x01b7,
1149 },
1150 },
1151 };
1152
1153 static __initconst const u64 core2_hw_cache_event_ids
1154 [PERF_COUNT_HW_CACHE_MAX]
1155 [PERF_COUNT_HW_CACHE_OP_MAX]
1156 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1157 {
1158 [ C(L1D) ] = {
1159 [ C(OP_READ) ] = {
1160 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */
1161 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */
1162 },
1163 [ C(OP_WRITE) ] = {
1164 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */
1165 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */
1166 },
1167 [ C(OP_PREFETCH) ] = {
1168 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */
1169 [ C(RESULT_MISS) ] = 0,
1170 },
1171 },
1172 [ C(L1I ) ] = {
1173 [ C(OP_READ) ] = {
1174 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */
1175 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */
1176 },
1177 [ C(OP_WRITE) ] = {
1178 [ C(RESULT_ACCESS) ] = -1,
1179 [ C(RESULT_MISS) ] = -1,
1180 },
1181 [ C(OP_PREFETCH) ] = {
1182 [ C(RESULT_ACCESS) ] = 0,
1183 [ C(RESULT_MISS) ] = 0,
1184 },
1185 },
1186 [ C(LL ) ] = {
1187 [ C(OP_READ) ] = {
1188 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
1189 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
1190 },
1191 [ C(OP_WRITE) ] = {
1192 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
1193 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
1194 },
1195 [ C(OP_PREFETCH) ] = {
1196 [ C(RESULT_ACCESS) ] = 0,
1197 [ C(RESULT_MISS) ] = 0,
1198 },
1199 },
1200 [ C(DTLB) ] = {
1201 [ C(OP_READ) ] = {
1202 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */
1203 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */
1204 },
1205 [ C(OP_WRITE) ] = {
1206 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */
1207 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */
1208 },
1209 [ C(OP_PREFETCH) ] = {
1210 [ C(RESULT_ACCESS) ] = 0,
1211 [ C(RESULT_MISS) ] = 0,
1212 },
1213 },
1214 [ C(ITLB) ] = {
1215 [ C(OP_READ) ] = {
1216 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1217 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */
1218 },
1219 [ C(OP_WRITE) ] = {
1220 [ C(RESULT_ACCESS) ] = -1,
1221 [ C(RESULT_MISS) ] = -1,
1222 },
1223 [ C(OP_PREFETCH) ] = {
1224 [ C(RESULT_ACCESS) ] = -1,
1225 [ C(RESULT_MISS) ] = -1,
1226 },
1227 },
1228 [ C(BPU ) ] = {
1229 [ C(OP_READ) ] = {
1230 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1231 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1232 },
1233 [ C(OP_WRITE) ] = {
1234 [ C(RESULT_ACCESS) ] = -1,
1235 [ C(RESULT_MISS) ] = -1,
1236 },
1237 [ C(OP_PREFETCH) ] = {
1238 [ C(RESULT_ACCESS) ] = -1,
1239 [ C(RESULT_MISS) ] = -1,
1240 },
1241 },
1242 };
1243
1244 static __initconst const u64 atom_hw_cache_event_ids
1245 [PERF_COUNT_HW_CACHE_MAX]
1246 [PERF_COUNT_HW_CACHE_OP_MAX]
1247 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1248 {
1249 [ C(L1D) ] = {
1250 [ C(OP_READ) ] = {
1251 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */
1252 [ C(RESULT_MISS) ] = 0,
1253 },
1254 [ C(OP_WRITE) ] = {
1255 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */
1256 [ C(RESULT_MISS) ] = 0,
1257 },
1258 [ C(OP_PREFETCH) ] = {
1259 [ C(RESULT_ACCESS) ] = 0x0,
1260 [ C(RESULT_MISS) ] = 0,
1261 },
1262 },
1263 [ C(L1I ) ] = {
1264 [ C(OP_READ) ] = {
1265 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */
1266 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */
1267 },
1268 [ C(OP_WRITE) ] = {
1269 [ C(RESULT_ACCESS) ] = -1,
1270 [ C(RESULT_MISS) ] = -1,
1271 },
1272 [ C(OP_PREFETCH) ] = {
1273 [ C(RESULT_ACCESS) ] = 0,
1274 [ C(RESULT_MISS) ] = 0,
1275 },
1276 },
1277 [ C(LL ) ] = {
1278 [ C(OP_READ) ] = {
1279 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */
1280 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */
1281 },
1282 [ C(OP_WRITE) ] = {
1283 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */
1284 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */
1285 },
1286 [ C(OP_PREFETCH) ] = {
1287 [ C(RESULT_ACCESS) ] = 0,
1288 [ C(RESULT_MISS) ] = 0,
1289 },
1290 },
1291 [ C(DTLB) ] = {
1292 [ C(OP_READ) ] = {
1293 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */
1294 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */
1295 },
1296 [ C(OP_WRITE) ] = {
1297 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */
1298 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */
1299 },
1300 [ C(OP_PREFETCH) ] = {
1301 [ C(RESULT_ACCESS) ] = 0,
1302 [ C(RESULT_MISS) ] = 0,
1303 },
1304 },
1305 [ C(ITLB) ] = {
1306 [ C(OP_READ) ] = {
1307 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1308 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */
1309 },
1310 [ C(OP_WRITE) ] = {
1311 [ C(RESULT_ACCESS) ] = -1,
1312 [ C(RESULT_MISS) ] = -1,
1313 },
1314 [ C(OP_PREFETCH) ] = {
1315 [ C(RESULT_ACCESS) ] = -1,
1316 [ C(RESULT_MISS) ] = -1,
1317 },
1318 },
1319 [ C(BPU ) ] = {
1320 [ C(OP_READ) ] = {
1321 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1322 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1323 },
1324 [ C(OP_WRITE) ] = {
1325 [ C(RESULT_ACCESS) ] = -1,
1326 [ C(RESULT_MISS) ] = -1,
1327 },
1328 [ C(OP_PREFETCH) ] = {
1329 [ C(RESULT_ACCESS) ] = -1,
1330 [ C(RESULT_MISS) ] = -1,
1331 },
1332 },
1333 };
1334
1335 static struct extra_reg intel_slm_extra_regs[] __read_mostly =
1336 {
1337 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */
1338 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x768005ffffull, RSP_0),
1339 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x368005ffffull, RSP_1),
1340 EVENT_EXTRA_END
1341 };
1342
1343 #define SLM_DMND_READ SNB_DMND_DATA_RD
1344 #define SLM_DMND_WRITE SNB_DMND_RFO
1345 #define SLM_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO)
1346
1347 #define SLM_SNP_ANY (SNB_SNP_NONE|SNB_SNP_MISS|SNB_NO_FWD|SNB_HITM)
1348 #define SLM_LLC_ACCESS SNB_RESP_ANY
1349 #define SLM_LLC_MISS (SLM_SNP_ANY|SNB_NON_DRAM)
1350
1351 static __initconst const u64 slm_hw_cache_extra_regs
1352 [PERF_COUNT_HW_CACHE_MAX]
1353 [PERF_COUNT_HW_CACHE_OP_MAX]
1354 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1355 {
1356 [ C(LL ) ] = {
1357 [ C(OP_READ) ] = {
1358 [ C(RESULT_ACCESS) ] = SLM_DMND_READ|SLM_LLC_ACCESS,
1359 [ C(RESULT_MISS) ] = 0,
1360 },
1361 [ C(OP_WRITE) ] = {
1362 [ C(RESULT_ACCESS) ] = SLM_DMND_WRITE|SLM_LLC_ACCESS,
1363 [ C(RESULT_MISS) ] = SLM_DMND_WRITE|SLM_LLC_MISS,
1364 },
1365 [ C(OP_PREFETCH) ] = {
1366 [ C(RESULT_ACCESS) ] = SLM_DMND_PREFETCH|SLM_LLC_ACCESS,
1367 [ C(RESULT_MISS) ] = SLM_DMND_PREFETCH|SLM_LLC_MISS,
1368 },
1369 },
1370 };
1371
1372 static __initconst const u64 slm_hw_cache_event_ids
1373 [PERF_COUNT_HW_CACHE_MAX]
1374 [PERF_COUNT_HW_CACHE_OP_MAX]
1375 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
1376 {
1377 [ C(L1D) ] = {
1378 [ C(OP_READ) ] = {
1379 [ C(RESULT_ACCESS) ] = 0,
1380 [ C(RESULT_MISS) ] = 0x0104, /* LD_DCU_MISS */
1381 },
1382 [ C(OP_WRITE) ] = {
1383 [ C(RESULT_ACCESS) ] = 0,
1384 [ C(RESULT_MISS) ] = 0,
1385 },
1386 [ C(OP_PREFETCH) ] = {
1387 [ C(RESULT_ACCESS) ] = 0,
1388 [ C(RESULT_MISS) ] = 0,
1389 },
1390 },
1391 [ C(L1I ) ] = {
1392 [ C(OP_READ) ] = {
1393 [ C(RESULT_ACCESS) ] = 0x0380, /* ICACHE.ACCESSES */
1394 [ C(RESULT_MISS) ] = 0x0280, /* ICACGE.MISSES */
1395 },
1396 [ C(OP_WRITE) ] = {
1397 [ C(RESULT_ACCESS) ] = -1,
1398 [ C(RESULT_MISS) ] = -1,
1399 },
1400 [ C(OP_PREFETCH) ] = {
1401 [ C(RESULT_ACCESS) ] = 0,
1402 [ C(RESULT_MISS) ] = 0,
1403 },
1404 },
1405 [ C(LL ) ] = {
1406 [ C(OP_READ) ] = {
1407 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1408 [ C(RESULT_ACCESS) ] = 0x01b7,
1409 [ C(RESULT_MISS) ] = 0,
1410 },
1411 [ C(OP_WRITE) ] = {
1412 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1413 [ C(RESULT_ACCESS) ] = 0x01b7,
1414 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1415 [ C(RESULT_MISS) ] = 0x01b7,
1416 },
1417 [ C(OP_PREFETCH) ] = {
1418 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1419 [ C(RESULT_ACCESS) ] = 0x01b7,
1420 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1421 [ C(RESULT_MISS) ] = 0x01b7,
1422 },
1423 },
1424 [ C(DTLB) ] = {
1425 [ C(OP_READ) ] = {
1426 [ C(RESULT_ACCESS) ] = 0,
1427 [ C(RESULT_MISS) ] = 0x0804, /* LD_DTLB_MISS */
1428 },
1429 [ C(OP_WRITE) ] = {
1430 [ C(RESULT_ACCESS) ] = 0,
1431 [ C(RESULT_MISS) ] = 0,
1432 },
1433 [ C(OP_PREFETCH) ] = {
1434 [ C(RESULT_ACCESS) ] = 0,
1435 [ C(RESULT_MISS) ] = 0,
1436 },
1437 },
1438 [ C(ITLB) ] = {
1439 [ C(OP_READ) ] = {
1440 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1441 [ C(RESULT_MISS) ] = 0x40205, /* PAGE_WALKS.I_SIDE_WALKS */
1442 },
1443 [ C(OP_WRITE) ] = {
1444 [ C(RESULT_ACCESS) ] = -1,
1445 [ C(RESULT_MISS) ] = -1,
1446 },
1447 [ C(OP_PREFETCH) ] = {
1448 [ C(RESULT_ACCESS) ] = -1,
1449 [ C(RESULT_MISS) ] = -1,
1450 },
1451 },
1452 [ C(BPU ) ] = {
1453 [ C(OP_READ) ] = {
1454 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1455 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1456 },
1457 [ C(OP_WRITE) ] = {
1458 [ C(RESULT_ACCESS) ] = -1,
1459 [ C(RESULT_MISS) ] = -1,
1460 },
1461 [ C(OP_PREFETCH) ] = {
1462 [ C(RESULT_ACCESS) ] = -1,
1463 [ C(RESULT_MISS) ] = -1,
1464 },
1465 },
1466 };
1467
1468 #define KNL_OT_L2_HITE BIT_ULL(19) /* Other Tile L2 Hit */
1469 #define KNL_OT_L2_HITF BIT_ULL(20) /* Other Tile L2 Hit */
1470 #define KNL_MCDRAM_LOCAL BIT_ULL(21)
1471 #define KNL_MCDRAM_FAR BIT_ULL(22)
1472 #define KNL_DDR_LOCAL BIT_ULL(23)
1473 #define KNL_DDR_FAR BIT_ULL(24)
1474 #define KNL_DRAM_ANY (KNL_MCDRAM_LOCAL | KNL_MCDRAM_FAR | \
1475 KNL_DDR_LOCAL | KNL_DDR_FAR)
1476 #define KNL_L2_READ SLM_DMND_READ
1477 #define KNL_L2_WRITE SLM_DMND_WRITE
1478 #define KNL_L2_PREFETCH SLM_DMND_PREFETCH
1479 #define KNL_L2_ACCESS SLM_LLC_ACCESS
1480 #define KNL_L2_MISS (KNL_OT_L2_HITE | KNL_OT_L2_HITF | \
1481 KNL_DRAM_ANY | SNB_SNP_ANY | \
1482 SNB_NON_DRAM)
1483
1484 static __initconst const u64 knl_hw_cache_extra_regs
1485 [PERF_COUNT_HW_CACHE_MAX]
1486 [PERF_COUNT_HW_CACHE_OP_MAX]
1487 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1488 [C(LL)] = {
1489 [C(OP_READ)] = {
1490 [C(RESULT_ACCESS)] = KNL_L2_READ | KNL_L2_ACCESS,
1491 [C(RESULT_MISS)] = 0,
1492 },
1493 [C(OP_WRITE)] = {
1494 [C(RESULT_ACCESS)] = KNL_L2_WRITE | KNL_L2_ACCESS,
1495 [C(RESULT_MISS)] = KNL_L2_WRITE | KNL_L2_MISS,
1496 },
1497 [C(OP_PREFETCH)] = {
1498 [C(RESULT_ACCESS)] = KNL_L2_PREFETCH | KNL_L2_ACCESS,
1499 [C(RESULT_MISS)] = KNL_L2_PREFETCH | KNL_L2_MISS,
1500 },
1501 },
1502 };
1503
1504 /*
1505 * Used from PMIs where the LBRs are already disabled.
1506 *
1507 * This function could be called consecutively. It is required to remain in
1508 * disabled state if called consecutively.
1509 *
1510 * During consecutive calls, the same disable value will be written to related
1511 * registers, so the PMU state remains unchanged. hw.state in
1512 * intel_bts_disable_local will remain PERF_HES_STOPPED too in consecutive
1513 * calls.
1514 */
1515 static void __intel_pmu_disable_all(void)
1516 {
1517 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1518
1519 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
1520
1521 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask))
1522 intel_pmu_disable_bts();
1523 else
1524 intel_bts_disable_local();
1525
1526 intel_pmu_pebs_disable_all();
1527 }
1528
1529 static void intel_pmu_disable_all(void)
1530 {
1531 __intel_pmu_disable_all();
1532 intel_pmu_lbr_disable_all();
1533 }
1534
1535 static void __intel_pmu_enable_all(int added, bool pmi)
1536 {
1537 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1538
1539 intel_pmu_pebs_enable_all();
1540 intel_pmu_lbr_enable_all(pmi);
1541 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL,
1542 x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask);
1543
1544 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
1545 struct perf_event *event =
1546 cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
1547
1548 if (WARN_ON_ONCE(!event))
1549 return;
1550
1551 intel_pmu_enable_bts(event->hw.config);
1552 } else
1553 intel_bts_enable_local();
1554 }
1555
1556 static void intel_pmu_enable_all(int added)
1557 {
1558 __intel_pmu_enable_all(added, false);
1559 }
1560
1561 /*
1562 * Workaround for:
1563 * Intel Errata AAK100 (model 26)
1564 * Intel Errata AAP53 (model 30)
1565 * Intel Errata BD53 (model 44)
1566 *
1567 * The official story:
1568 * These chips need to be 'reset' when adding counters by programming the
1569 * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either
1570 * in sequence on the same PMC or on different PMCs.
1571 *
1572 * In practise it appears some of these events do in fact count, and
1573 * we need to programm all 4 events.
1574 */
1575 static void intel_pmu_nhm_workaround(void)
1576 {
1577 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1578 static const unsigned long nhm_magic[4] = {
1579 0x4300B5,
1580 0x4300D2,
1581 0x4300B1,
1582 0x4300B1
1583 };
1584 struct perf_event *event;
1585 int i;
1586
1587 /*
1588 * The Errata requires below steps:
1589 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL;
1590 * 2) Configure 4 PERFEVTSELx with the magic events and clear
1591 * the corresponding PMCx;
1592 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL;
1593 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL;
1594 * 5) Clear 4 pairs of ERFEVTSELx and PMCx;
1595 */
1596
1597 /*
1598 * The real steps we choose are a little different from above.
1599 * A) To reduce MSR operations, we don't run step 1) as they
1600 * are already cleared before this function is called;
1601 * B) Call x86_perf_event_update to save PMCx before configuring
1602 * PERFEVTSELx with magic number;
1603 * C) With step 5), we do clear only when the PERFEVTSELx is
1604 * not used currently.
1605 * D) Call x86_perf_event_set_period to restore PMCx;
1606 */
1607
1608 /* We always operate 4 pairs of PERF Counters */
1609 for (i = 0; i < 4; i++) {
1610 event = cpuc->events[i];
1611 if (event)
1612 x86_perf_event_update(event);
1613 }
1614
1615 for (i = 0; i < 4; i++) {
1616 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]);
1617 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0);
1618 }
1619
1620 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf);
1621 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
1622
1623 for (i = 0; i < 4; i++) {
1624 event = cpuc->events[i];
1625
1626 if (event) {
1627 x86_perf_event_set_period(event);
1628 __x86_pmu_enable_event(&event->hw,
1629 ARCH_PERFMON_EVENTSEL_ENABLE);
1630 } else
1631 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0);
1632 }
1633 }
1634
1635 static void intel_pmu_nhm_enable_all(int added)
1636 {
1637 if (added)
1638 intel_pmu_nhm_workaround();
1639 intel_pmu_enable_all(added);
1640 }
1641
1642 static inline u64 intel_pmu_get_status(void)
1643 {
1644 u64 status;
1645
1646 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1647
1648 return status;
1649 }
1650
1651 static inline void intel_pmu_ack_status(u64 ack)
1652 {
1653 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
1654 }
1655
1656 static void intel_pmu_disable_fixed(struct hw_perf_event *hwc)
1657 {
1658 int idx = hwc->idx - INTEL_PMC_IDX_FIXED;
1659 u64 ctrl_val, mask;
1660
1661 mask = 0xfULL << (idx * 4);
1662
1663 rdmsrl(hwc->config_base, ctrl_val);
1664 ctrl_val &= ~mask;
1665 wrmsrl(hwc->config_base, ctrl_val);
1666 }
1667
1668 static inline bool event_is_checkpointed(struct perf_event *event)
1669 {
1670 return (event->hw.config & HSW_IN_TX_CHECKPOINTED) != 0;
1671 }
1672
1673 static void intel_pmu_disable_event(struct perf_event *event)
1674 {
1675 struct hw_perf_event *hwc = &event->hw;
1676 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1677
1678 if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) {
1679 intel_pmu_disable_bts();
1680 intel_pmu_drain_bts_buffer();
1681 return;
1682 }
1683
1684 cpuc->intel_ctrl_guest_mask &= ~(1ull << hwc->idx);
1685 cpuc->intel_ctrl_host_mask &= ~(1ull << hwc->idx);
1686 cpuc->intel_cp_status &= ~(1ull << hwc->idx);
1687
1688 /*
1689 * must disable before any actual event
1690 * because any event may be combined with LBR
1691 */
1692 if (needs_branch_stack(event))
1693 intel_pmu_lbr_disable(event);
1694
1695 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
1696 intel_pmu_disable_fixed(hwc);
1697 return;
1698 }
1699
1700 x86_pmu_disable_event(event);
1701
1702 if (unlikely(event->attr.precise_ip))
1703 intel_pmu_pebs_disable(event);
1704 }
1705
1706 static void intel_pmu_enable_fixed(struct hw_perf_event *hwc)
1707 {
1708 int idx = hwc->idx - INTEL_PMC_IDX_FIXED;
1709 u64 ctrl_val, bits, mask;
1710
1711 /*
1712 * Enable IRQ generation (0x8),
1713 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
1714 * if requested:
1715 */
1716 bits = 0x8ULL;
1717 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
1718 bits |= 0x2;
1719 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
1720 bits |= 0x1;
1721
1722 /*
1723 * ANY bit is supported in v3 and up
1724 */
1725 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
1726 bits |= 0x4;
1727
1728 bits <<= (idx * 4);
1729 mask = 0xfULL << (idx * 4);
1730
1731 rdmsrl(hwc->config_base, ctrl_val);
1732 ctrl_val &= ~mask;
1733 ctrl_val |= bits;
1734 wrmsrl(hwc->config_base, ctrl_val);
1735 }
1736
1737 static void intel_pmu_enable_event(struct perf_event *event)
1738 {
1739 struct hw_perf_event *hwc = &event->hw;
1740 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1741
1742 if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) {
1743 if (!__this_cpu_read(cpu_hw_events.enabled))
1744 return;
1745
1746 intel_pmu_enable_bts(hwc->config);
1747 return;
1748 }
1749 /*
1750 * must enabled before any actual event
1751 * because any event may be combined with LBR
1752 */
1753 if (needs_branch_stack(event))
1754 intel_pmu_lbr_enable(event);
1755
1756 if (event->attr.exclude_host)
1757 cpuc->intel_ctrl_guest_mask |= (1ull << hwc->idx);
1758 if (event->attr.exclude_guest)
1759 cpuc->intel_ctrl_host_mask |= (1ull << hwc->idx);
1760
1761 if (unlikely(event_is_checkpointed(event)))
1762 cpuc->intel_cp_status |= (1ull << hwc->idx);
1763
1764 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
1765 intel_pmu_enable_fixed(hwc);
1766 return;
1767 }
1768
1769 if (unlikely(event->attr.precise_ip))
1770 intel_pmu_pebs_enable(event);
1771
1772 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
1773 }
1774
1775 /*
1776 * Save and restart an expired event. Called by NMI contexts,
1777 * so it has to be careful about preempting normal event ops:
1778 */
1779 int intel_pmu_save_and_restart(struct perf_event *event)
1780 {
1781 x86_perf_event_update(event);
1782 /*
1783 * For a checkpointed counter always reset back to 0. This
1784 * avoids a situation where the counter overflows, aborts the
1785 * transaction and is then set back to shortly before the
1786 * overflow, and overflows and aborts again.
1787 */
1788 if (unlikely(event_is_checkpointed(event))) {
1789 /* No race with NMIs because the counter should not be armed */
1790 wrmsrl(event->hw.event_base, 0);
1791 local64_set(&event->hw.prev_count, 0);
1792 }
1793 return x86_perf_event_set_period(event);
1794 }
1795
1796 static void intel_pmu_reset(void)
1797 {
1798 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
1799 unsigned long flags;
1800 int idx;
1801
1802 if (!x86_pmu.num_counters)
1803 return;
1804
1805 local_irq_save(flags);
1806
1807 pr_info("clearing PMU state on CPU#%d\n", smp_processor_id());
1808
1809 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1810 wrmsrl_safe(x86_pmu_config_addr(idx), 0ull);
1811 wrmsrl_safe(x86_pmu_event_addr(idx), 0ull);
1812 }
1813 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++)
1814 wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
1815
1816 if (ds)
1817 ds->bts_index = ds->bts_buffer_base;
1818
1819 /* Ack all overflows and disable fixed counters */
1820 if (x86_pmu.version >= 2) {
1821 intel_pmu_ack_status(intel_pmu_get_status());
1822 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
1823 }
1824
1825 /* Reset LBRs and LBR freezing */
1826 if (x86_pmu.lbr_nr) {
1827 update_debugctlmsr(get_debugctlmsr() &
1828 ~(DEBUGCTLMSR_FREEZE_LBRS_ON_PMI|DEBUGCTLMSR_LBR));
1829 }
1830
1831 local_irq_restore(flags);
1832 }
1833
1834 /*
1835 * This handler is triggered by the local APIC, so the APIC IRQ handling
1836 * rules apply:
1837 */
1838 static int intel_pmu_handle_irq(struct pt_regs *regs)
1839 {
1840 struct perf_sample_data data;
1841 struct cpu_hw_events *cpuc;
1842 int bit, loops;
1843 u64 status;
1844 int handled;
1845
1846 cpuc = this_cpu_ptr(&cpu_hw_events);
1847
1848 /*
1849 * No known reason to not always do late ACK,
1850 * but just in case do it opt-in.
1851 */
1852 if (!x86_pmu.late_ack)
1853 apic_write(APIC_LVTPC, APIC_DM_NMI);
1854 __intel_pmu_disable_all();
1855 handled = intel_pmu_drain_bts_buffer();
1856 handled += intel_bts_interrupt();
1857 status = intel_pmu_get_status();
1858 if (!status)
1859 goto done;
1860
1861 loops = 0;
1862 again:
1863 intel_pmu_lbr_read();
1864 intel_pmu_ack_status(status);
1865 if (++loops > 100) {
1866 static bool warned = false;
1867 if (!warned) {
1868 WARN(1, "perfevents: irq loop stuck!\n");
1869 perf_event_print_debug();
1870 warned = true;
1871 }
1872 intel_pmu_reset();
1873 goto done;
1874 }
1875
1876 inc_irq_stat(apic_perf_irqs);
1877
1878
1879 /*
1880 * Ignore a range of extra bits in status that do not indicate
1881 * overflow by themselves.
1882 */
1883 status &= ~(GLOBAL_STATUS_COND_CHG |
1884 GLOBAL_STATUS_ASIF |
1885 GLOBAL_STATUS_LBRS_FROZEN);
1886 if (!status)
1887 goto done;
1888
1889 /*
1890 * PEBS overflow sets bit 62 in the global status register
1891 */
1892 if (__test_and_clear_bit(62, (unsigned long *)&status)) {
1893 handled++;
1894 x86_pmu.drain_pebs(regs);
1895 /*
1896 * There are cases where, even though, the PEBS ovfl bit is set
1897 * in GLOBAL_OVF_STATUS, the PEBS events may also have their
1898 * overflow bits set for their counters. We must clear them
1899 * here because they have been processed as exact samples in
1900 * the drain_pebs() routine. They must not be processed again
1901 * in the for_each_bit_set() loop for regular samples below.
1902 */
1903 status &= ~cpuc->pebs_enabled;
1904 status &= x86_pmu.intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI;
1905 }
1906
1907 /*
1908 * Intel PT
1909 */
1910 if (__test_and_clear_bit(55, (unsigned long *)&status)) {
1911 handled++;
1912 intel_pt_interrupt();
1913 }
1914
1915 /*
1916 * Checkpointed counters can lead to 'spurious' PMIs because the
1917 * rollback caused by the PMI will have cleared the overflow status
1918 * bit. Therefore always force probe these counters.
1919 */
1920 status |= cpuc->intel_cp_status;
1921
1922 for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
1923 struct perf_event *event = cpuc->events[bit];
1924
1925 handled++;
1926
1927 if (!test_bit(bit, cpuc->active_mask))
1928 continue;
1929
1930 if (!intel_pmu_save_and_restart(event))
1931 continue;
1932
1933 perf_sample_data_init(&data, 0, event->hw.last_period);
1934
1935 if (has_branch_stack(event))
1936 data.br_stack = &cpuc->lbr_stack;
1937
1938 if (perf_event_overflow(event, &data, regs))
1939 x86_pmu_stop(event, 0);
1940 }
1941
1942 /*
1943 * Repeat if there is more work to be done:
1944 */
1945 status = intel_pmu_get_status();
1946 if (status)
1947 goto again;
1948
1949 done:
1950 /* Only restore PMU state when it's active. See x86_pmu_disable(). */
1951 if (cpuc->enabled)
1952 __intel_pmu_enable_all(0, true);
1953
1954 /*
1955 * Only unmask the NMI after the overflow counters
1956 * have been reset. This avoids spurious NMIs on
1957 * Haswell CPUs.
1958 */
1959 if (x86_pmu.late_ack)
1960 apic_write(APIC_LVTPC, APIC_DM_NMI);
1961 return handled;
1962 }
1963
1964 static struct event_constraint *
1965 intel_bts_constraints(struct perf_event *event)
1966 {
1967 struct hw_perf_event *hwc = &event->hw;
1968 unsigned int hw_event, bts_event;
1969
1970 if (event->attr.freq)
1971 return NULL;
1972
1973 hw_event = hwc->config & INTEL_ARCH_EVENT_MASK;
1974 bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
1975
1976 if (unlikely(hw_event == bts_event && hwc->sample_period == 1))
1977 return &bts_constraint;
1978
1979 return NULL;
1980 }
1981
1982 static int intel_alt_er(int idx, u64 config)
1983 {
1984 int alt_idx = idx;
1985
1986 if (!(x86_pmu.flags & PMU_FL_HAS_RSP_1))
1987 return idx;
1988
1989 if (idx == EXTRA_REG_RSP_0)
1990 alt_idx = EXTRA_REG_RSP_1;
1991
1992 if (idx == EXTRA_REG_RSP_1)
1993 alt_idx = EXTRA_REG_RSP_0;
1994
1995 if (config & ~x86_pmu.extra_regs[alt_idx].valid_mask)
1996 return idx;
1997
1998 return alt_idx;
1999 }
2000
2001 static void intel_fixup_er(struct perf_event *event, int idx)
2002 {
2003 event->hw.extra_reg.idx = idx;
2004
2005 if (idx == EXTRA_REG_RSP_0) {
2006 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
2007 event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_0].event;
2008 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0;
2009 } else if (idx == EXTRA_REG_RSP_1) {
2010 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
2011 event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_1].event;
2012 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1;
2013 }
2014 }
2015
2016 /*
2017 * manage allocation of shared extra msr for certain events
2018 *
2019 * sharing can be:
2020 * per-cpu: to be shared between the various events on a single PMU
2021 * per-core: per-cpu + shared by HT threads
2022 */
2023 static struct event_constraint *
2024 __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc,
2025 struct perf_event *event,
2026 struct hw_perf_event_extra *reg)
2027 {
2028 struct event_constraint *c = &emptyconstraint;
2029 struct er_account *era;
2030 unsigned long flags;
2031 int idx = reg->idx;
2032
2033 /*
2034 * reg->alloc can be set due to existing state, so for fake cpuc we
2035 * need to ignore this, otherwise we might fail to allocate proper fake
2036 * state for this extra reg constraint. Also see the comment below.
2037 */
2038 if (reg->alloc && !cpuc->is_fake)
2039 return NULL; /* call x86_get_event_constraint() */
2040
2041 again:
2042 era = &cpuc->shared_regs->regs[idx];
2043 /*
2044 * we use spin_lock_irqsave() to avoid lockdep issues when
2045 * passing a fake cpuc
2046 */
2047 raw_spin_lock_irqsave(&era->lock, flags);
2048
2049 if (!atomic_read(&era->ref) || era->config == reg->config) {
2050
2051 /*
2052 * If its a fake cpuc -- as per validate_{group,event}() we
2053 * shouldn't touch event state and we can avoid doing so
2054 * since both will only call get_event_constraints() once
2055 * on each event, this avoids the need for reg->alloc.
2056 *
2057 * Not doing the ER fixup will only result in era->reg being
2058 * wrong, but since we won't actually try and program hardware
2059 * this isn't a problem either.
2060 */
2061 if (!cpuc->is_fake) {
2062 if (idx != reg->idx)
2063 intel_fixup_er(event, idx);
2064
2065 /*
2066 * x86_schedule_events() can call get_event_constraints()
2067 * multiple times on events in the case of incremental
2068 * scheduling(). reg->alloc ensures we only do the ER
2069 * allocation once.
2070 */
2071 reg->alloc = 1;
2072 }
2073
2074 /* lock in msr value */
2075 era->config = reg->config;
2076 era->reg = reg->reg;
2077
2078 /* one more user */
2079 atomic_inc(&era->ref);
2080
2081 /*
2082 * need to call x86_get_event_constraint()
2083 * to check if associated event has constraints
2084 */
2085 c = NULL;
2086 } else {
2087 idx = intel_alt_er(idx, reg->config);
2088 if (idx != reg->idx) {
2089 raw_spin_unlock_irqrestore(&era->lock, flags);
2090 goto again;
2091 }
2092 }
2093 raw_spin_unlock_irqrestore(&era->lock, flags);
2094
2095 return c;
2096 }
2097
2098 static void
2099 __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc,
2100 struct hw_perf_event_extra *reg)
2101 {
2102 struct er_account *era;
2103
2104 /*
2105 * Only put constraint if extra reg was actually allocated. Also takes
2106 * care of event which do not use an extra shared reg.
2107 *
2108 * Also, if this is a fake cpuc we shouldn't touch any event state
2109 * (reg->alloc) and we don't care about leaving inconsistent cpuc state
2110 * either since it'll be thrown out.
2111 */
2112 if (!reg->alloc || cpuc->is_fake)
2113 return;
2114
2115 era = &cpuc->shared_regs->regs[reg->idx];
2116
2117 /* one fewer user */
2118 atomic_dec(&era->ref);
2119
2120 /* allocate again next time */
2121 reg->alloc = 0;
2122 }
2123
2124 static struct event_constraint *
2125 intel_shared_regs_constraints(struct cpu_hw_events *cpuc,
2126 struct perf_event *event)
2127 {
2128 struct event_constraint *c = NULL, *d;
2129 struct hw_perf_event_extra *xreg, *breg;
2130
2131 xreg = &event->hw.extra_reg;
2132 if (xreg->idx != EXTRA_REG_NONE) {
2133 c = __intel_shared_reg_get_constraints(cpuc, event, xreg);
2134 if (c == &emptyconstraint)
2135 return c;
2136 }
2137 breg = &event->hw.branch_reg;
2138 if (breg->idx != EXTRA_REG_NONE) {
2139 d = __intel_shared_reg_get_constraints(cpuc, event, breg);
2140 if (d == &emptyconstraint) {
2141 __intel_shared_reg_put_constraints(cpuc, xreg);
2142 c = d;
2143 }
2144 }
2145 return c;
2146 }
2147
2148 struct event_constraint *
2149 x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
2150 struct perf_event *event)
2151 {
2152 struct event_constraint *c;
2153
2154 if (x86_pmu.event_constraints) {
2155 for_each_event_constraint(c, x86_pmu.event_constraints) {
2156 if ((event->hw.config & c->cmask) == c->code) {
2157 event->hw.flags |= c->flags;
2158 return c;
2159 }
2160 }
2161 }
2162
2163 return &unconstrained;
2164 }
2165
2166 static struct event_constraint *
2167 __intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
2168 struct perf_event *event)
2169 {
2170 struct event_constraint *c;
2171
2172 c = intel_bts_constraints(event);
2173 if (c)
2174 return c;
2175
2176 c = intel_shared_regs_constraints(cpuc, event);
2177 if (c)
2178 return c;
2179
2180 c = intel_pebs_constraints(event);
2181 if (c)
2182 return c;
2183
2184 return x86_get_event_constraints(cpuc, idx, event);
2185 }
2186
2187 static void
2188 intel_start_scheduling(struct cpu_hw_events *cpuc)
2189 {
2190 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
2191 struct intel_excl_states *xl;
2192 int tid = cpuc->excl_thread_id;
2193
2194 /*
2195 * nothing needed if in group validation mode
2196 */
2197 if (cpuc->is_fake || !is_ht_workaround_enabled())
2198 return;
2199
2200 /*
2201 * no exclusion needed
2202 */
2203 if (WARN_ON_ONCE(!excl_cntrs))
2204 return;
2205
2206 xl = &excl_cntrs->states[tid];
2207
2208 xl->sched_started = true;
2209 /*
2210 * lock shared state until we are done scheduling
2211 * in stop_event_scheduling()
2212 * makes scheduling appear as a transaction
2213 */
2214 raw_spin_lock(&excl_cntrs->lock);
2215 }
2216
2217 static void intel_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr)
2218 {
2219 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
2220 struct event_constraint *c = cpuc->event_constraint[idx];
2221 struct intel_excl_states *xl;
2222 int tid = cpuc->excl_thread_id;
2223
2224 if (cpuc->is_fake || !is_ht_workaround_enabled())
2225 return;
2226
2227 if (WARN_ON_ONCE(!excl_cntrs))
2228 return;
2229
2230 if (!(c->flags & PERF_X86_EVENT_DYNAMIC))
2231 return;
2232
2233 xl = &excl_cntrs->states[tid];
2234
2235 lockdep_assert_held(&excl_cntrs->lock);
2236
2237 if (c->flags & PERF_X86_EVENT_EXCL)
2238 xl->state[cntr] = INTEL_EXCL_EXCLUSIVE;
2239 else
2240 xl->state[cntr] = INTEL_EXCL_SHARED;
2241 }
2242
2243 static void
2244 intel_stop_scheduling(struct cpu_hw_events *cpuc)
2245 {
2246 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
2247 struct intel_excl_states *xl;
2248 int tid = cpuc->excl_thread_id;
2249
2250 /*
2251 * nothing needed if in group validation mode
2252 */
2253 if (cpuc->is_fake || !is_ht_workaround_enabled())
2254 return;
2255 /*
2256 * no exclusion needed
2257 */
2258 if (WARN_ON_ONCE(!excl_cntrs))
2259 return;
2260
2261 xl = &excl_cntrs->states[tid];
2262
2263 xl->sched_started = false;
2264 /*
2265 * release shared state lock (acquired in intel_start_scheduling())
2266 */
2267 raw_spin_unlock(&excl_cntrs->lock);
2268 }
2269
2270 static struct event_constraint *
2271 intel_get_excl_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
2272 int idx, struct event_constraint *c)
2273 {
2274 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
2275 struct intel_excl_states *xlo;
2276 int tid = cpuc->excl_thread_id;
2277 int is_excl, i;
2278
2279 /*
2280 * validating a group does not require
2281 * enforcing cross-thread exclusion
2282 */
2283 if (cpuc->is_fake || !is_ht_workaround_enabled())
2284 return c;
2285
2286 /*
2287 * no exclusion needed
2288 */
2289 if (WARN_ON_ONCE(!excl_cntrs))
2290 return c;
2291
2292 /*
2293 * because we modify the constraint, we need
2294 * to make a copy. Static constraints come
2295 * from static const tables.
2296 *
2297 * only needed when constraint has not yet
2298 * been cloned (marked dynamic)
2299 */
2300 if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) {
2301 struct event_constraint *cx;
2302
2303 /*
2304 * grab pre-allocated constraint entry
2305 */
2306 cx = &cpuc->constraint_list[idx];
2307
2308 /*
2309 * initialize dynamic constraint
2310 * with static constraint
2311 */
2312 *cx = *c;
2313
2314 /*
2315 * mark constraint as dynamic, so we
2316 * can free it later on
2317 */
2318 cx->flags |= PERF_X86_EVENT_DYNAMIC;
2319 c = cx;
2320 }
2321
2322 /*
2323 * From here on, the constraint is dynamic.
2324 * Either it was just allocated above, or it
2325 * was allocated during a earlier invocation
2326 * of this function
2327 */
2328
2329 /*
2330 * state of sibling HT
2331 */
2332 xlo = &excl_cntrs->states[tid ^ 1];
2333
2334 /*
2335 * event requires exclusive counter access
2336 * across HT threads
2337 */
2338 is_excl = c->flags & PERF_X86_EVENT_EXCL;
2339 if (is_excl && !(event->hw.flags & PERF_X86_EVENT_EXCL_ACCT)) {
2340 event->hw.flags |= PERF_X86_EVENT_EXCL_ACCT;
2341 if (!cpuc->n_excl++)
2342 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 1);
2343 }
2344
2345 /*
2346 * Modify static constraint with current dynamic
2347 * state of thread
2348 *
2349 * EXCLUSIVE: sibling counter measuring exclusive event
2350 * SHARED : sibling counter measuring non-exclusive event
2351 * UNUSED : sibling counter unused
2352 */
2353 for_each_set_bit(i, c->idxmsk, X86_PMC_IDX_MAX) {
2354 /*
2355 * exclusive event in sibling counter
2356 * our corresponding counter cannot be used
2357 * regardless of our event
2358 */
2359 if (xlo->state[i] == INTEL_EXCL_EXCLUSIVE)
2360 __clear_bit(i, c->idxmsk);
2361 /*
2362 * if measuring an exclusive event, sibling
2363 * measuring non-exclusive, then counter cannot
2364 * be used
2365 */
2366 if (is_excl && xlo->state[i] == INTEL_EXCL_SHARED)
2367 __clear_bit(i, c->idxmsk);
2368 }
2369
2370 /*
2371 * recompute actual bit weight for scheduling algorithm
2372 */
2373 c->weight = hweight64(c->idxmsk64);
2374
2375 /*
2376 * if we return an empty mask, then switch
2377 * back to static empty constraint to avoid
2378 * the cost of freeing later on
2379 */
2380 if (c->weight == 0)
2381 c = &emptyconstraint;
2382
2383 return c;
2384 }
2385
2386 static struct event_constraint *
2387 intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
2388 struct perf_event *event)
2389 {
2390 struct event_constraint *c1 = NULL;
2391 struct event_constraint *c2;
2392
2393 if (idx >= 0) /* fake does < 0 */
2394 c1 = cpuc->event_constraint[idx];
2395
2396 /*
2397 * first time only
2398 * - static constraint: no change across incremental scheduling calls
2399 * - dynamic constraint: handled by intel_get_excl_constraints()
2400 */
2401 c2 = __intel_get_event_constraints(cpuc, idx, event);
2402 if (c1 && (c1->flags & PERF_X86_EVENT_DYNAMIC)) {
2403 bitmap_copy(c1->idxmsk, c2->idxmsk, X86_PMC_IDX_MAX);
2404 c1->weight = c2->weight;
2405 c2 = c1;
2406 }
2407
2408 if (cpuc->excl_cntrs)
2409 return intel_get_excl_constraints(cpuc, event, idx, c2);
2410
2411 return c2;
2412 }
2413
2414 static void intel_put_excl_constraints(struct cpu_hw_events *cpuc,
2415 struct perf_event *event)
2416 {
2417 struct hw_perf_event *hwc = &event->hw;
2418 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
2419 int tid = cpuc->excl_thread_id;
2420 struct intel_excl_states *xl;
2421
2422 /*
2423 * nothing needed if in group validation mode
2424 */
2425 if (cpuc->is_fake)
2426 return;
2427
2428 if (WARN_ON_ONCE(!excl_cntrs))
2429 return;
2430
2431 if (hwc->flags & PERF_X86_EVENT_EXCL_ACCT) {
2432 hwc->flags &= ~PERF_X86_EVENT_EXCL_ACCT;
2433 if (!--cpuc->n_excl)
2434 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 0);
2435 }
2436
2437 /*
2438 * If event was actually assigned, then mark the counter state as
2439 * unused now.
2440 */
2441 if (hwc->idx >= 0) {
2442 xl = &excl_cntrs->states[tid];
2443
2444 /*
2445 * put_constraint may be called from x86_schedule_events()
2446 * which already has the lock held so here make locking
2447 * conditional.
2448 */
2449 if (!xl->sched_started)
2450 raw_spin_lock(&excl_cntrs->lock);
2451
2452 xl->state[hwc->idx] = INTEL_EXCL_UNUSED;
2453
2454 if (!xl->sched_started)
2455 raw_spin_unlock(&excl_cntrs->lock);
2456 }
2457 }
2458
2459 static void
2460 intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc,
2461 struct perf_event *event)
2462 {
2463 struct hw_perf_event_extra *reg;
2464
2465 reg = &event->hw.extra_reg;
2466 if (reg->idx != EXTRA_REG_NONE)
2467 __intel_shared_reg_put_constraints(cpuc, reg);
2468
2469 reg = &event->hw.branch_reg;
2470 if (reg->idx != EXTRA_REG_NONE)
2471 __intel_shared_reg_put_constraints(cpuc, reg);
2472 }
2473
2474 static void intel_put_event_constraints(struct cpu_hw_events *cpuc,
2475 struct perf_event *event)
2476 {
2477 intel_put_shared_regs_event_constraints(cpuc, event);
2478
2479 /*
2480 * is PMU has exclusive counter restrictions, then
2481 * all events are subject to and must call the
2482 * put_excl_constraints() routine
2483 */
2484 if (cpuc->excl_cntrs)
2485 intel_put_excl_constraints(cpuc, event);
2486 }
2487
2488 static void intel_pebs_aliases_core2(struct perf_event *event)
2489 {
2490 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
2491 /*
2492 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
2493 * (0x003c) so that we can use it with PEBS.
2494 *
2495 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
2496 * PEBS capable. However we can use INST_RETIRED.ANY_P
2497 * (0x00c0), which is a PEBS capable event, to get the same
2498 * count.
2499 *
2500 * INST_RETIRED.ANY_P counts the number of cycles that retires
2501 * CNTMASK instructions. By setting CNTMASK to a value (16)
2502 * larger than the maximum number of instructions that can be
2503 * retired per cycle (4) and then inverting the condition, we
2504 * count all cycles that retire 16 or less instructions, which
2505 * is every cycle.
2506 *
2507 * Thereby we gain a PEBS capable cycle counter.
2508 */
2509 u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16);
2510
2511 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
2512 event->hw.config = alt_config;
2513 }
2514 }
2515
2516 static void intel_pebs_aliases_snb(struct perf_event *event)
2517 {
2518 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
2519 /*
2520 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
2521 * (0x003c) so that we can use it with PEBS.
2522 *
2523 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
2524 * PEBS capable. However we can use UOPS_RETIRED.ALL
2525 * (0x01c2), which is a PEBS capable event, to get the same
2526 * count.
2527 *
2528 * UOPS_RETIRED.ALL counts the number of cycles that retires
2529 * CNTMASK micro-ops. By setting CNTMASK to a value (16)
2530 * larger than the maximum number of micro-ops that can be
2531 * retired per cycle (4) and then inverting the condition, we
2532 * count all cycles that retire 16 or less micro-ops, which
2533 * is every cycle.
2534 *
2535 * Thereby we gain a PEBS capable cycle counter.
2536 */
2537 u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16);
2538
2539 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
2540 event->hw.config = alt_config;
2541 }
2542 }
2543
2544 static void intel_pebs_aliases_precdist(struct perf_event *event)
2545 {
2546 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
2547 /*
2548 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
2549 * (0x003c) so that we can use it with PEBS.
2550 *
2551 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
2552 * PEBS capable. However we can use INST_RETIRED.PREC_DIST
2553 * (0x01c0), which is a PEBS capable event, to get the same
2554 * count.
2555 *
2556 * The PREC_DIST event has special support to minimize sample
2557 * shadowing effects. One drawback is that it can be
2558 * only programmed on counter 1, but that seems like an
2559 * acceptable trade off.
2560 */
2561 u64 alt_config = X86_CONFIG(.event=0xc0, .umask=0x01, .inv=1, .cmask=16);
2562
2563 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
2564 event->hw.config = alt_config;
2565 }
2566 }
2567
2568 static void intel_pebs_aliases_ivb(struct perf_event *event)
2569 {
2570 if (event->attr.precise_ip < 3)
2571 return intel_pebs_aliases_snb(event);
2572 return intel_pebs_aliases_precdist(event);
2573 }
2574
2575 static void intel_pebs_aliases_skl(struct perf_event *event)
2576 {
2577 if (event->attr.precise_ip < 3)
2578 return intel_pebs_aliases_core2(event);
2579 return intel_pebs_aliases_precdist(event);
2580 }
2581
2582 static unsigned long intel_pmu_free_running_flags(struct perf_event *event)
2583 {
2584 unsigned long flags = x86_pmu.free_running_flags;
2585
2586 if (event->attr.use_clockid)
2587 flags &= ~PERF_SAMPLE_TIME;
2588 return flags;
2589 }
2590
2591 static int intel_pmu_hw_config(struct perf_event *event)
2592 {
2593 int ret = x86_pmu_hw_config(event);
2594
2595 if (ret)
2596 return ret;
2597
2598 if (event->attr.precise_ip) {
2599 if (!event->attr.freq) {
2600 event->hw.flags |= PERF_X86_EVENT_AUTO_RELOAD;
2601 if (!(event->attr.sample_type &
2602 ~intel_pmu_free_running_flags(event)))
2603 event->hw.flags |= PERF_X86_EVENT_FREERUNNING;
2604 }
2605 if (x86_pmu.pebs_aliases)
2606 x86_pmu.pebs_aliases(event);
2607 }
2608
2609 if (needs_branch_stack(event)) {
2610 ret = intel_pmu_setup_lbr_filter(event);
2611 if (ret)
2612 return ret;
2613
2614 /*
2615 * BTS is set up earlier in this path, so don't account twice
2616 */
2617 if (!intel_pmu_has_bts(event)) {
2618 /* disallow lbr if conflicting events are present */
2619 if (x86_add_exclusive(x86_lbr_exclusive_lbr))
2620 return -EBUSY;
2621
2622 event->destroy = hw_perf_lbr_event_destroy;
2623 }
2624 }
2625
2626 if (event->attr.type != PERF_TYPE_RAW)
2627 return 0;
2628
2629 if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY))
2630 return 0;
2631
2632 if (x86_pmu.version < 3)
2633 return -EINVAL;
2634
2635 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2636 return -EACCES;
2637
2638 event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY;
2639
2640 return 0;
2641 }
2642
2643 struct perf_guest_switch_msr *perf_guest_get_msrs(int *nr)
2644 {
2645 if (x86_pmu.guest_get_msrs)
2646 return x86_pmu.guest_get_msrs(nr);
2647 *nr = 0;
2648 return NULL;
2649 }
2650 EXPORT_SYMBOL_GPL(perf_guest_get_msrs);
2651
2652 static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr)
2653 {
2654 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2655 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
2656
2657 arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL;
2658 arr[0].host = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask;
2659 arr[0].guest = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_host_mask;
2660 /*
2661 * If PMU counter has PEBS enabled it is not enough to disable counter
2662 * on a guest entry since PEBS memory write can overshoot guest entry
2663 * and corrupt guest memory. Disabling PEBS solves the problem.
2664 */
2665 arr[1].msr = MSR_IA32_PEBS_ENABLE;
2666 arr[1].host = cpuc->pebs_enabled;
2667 arr[1].guest = 0;
2668
2669 *nr = 2;
2670 return arr;
2671 }
2672
2673 static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr)
2674 {
2675 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2676 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
2677 int idx;
2678
2679 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
2680 struct perf_event *event = cpuc->events[idx];
2681
2682 arr[idx].msr = x86_pmu_config_addr(idx);
2683 arr[idx].host = arr[idx].guest = 0;
2684
2685 if (!test_bit(idx, cpuc->active_mask))
2686 continue;
2687
2688 arr[idx].host = arr[idx].guest =
2689 event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE;
2690
2691 if (event->attr.exclude_host)
2692 arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
2693 else if (event->attr.exclude_guest)
2694 arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
2695 }
2696
2697 *nr = x86_pmu.num_counters;
2698 return arr;
2699 }
2700
2701 static void core_pmu_enable_event(struct perf_event *event)
2702 {
2703 if (!event->attr.exclude_host)
2704 x86_pmu_enable_event(event);
2705 }
2706
2707 static void core_pmu_enable_all(int added)
2708 {
2709 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2710 int idx;
2711
2712 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
2713 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
2714
2715 if (!test_bit(idx, cpuc->active_mask) ||
2716 cpuc->events[idx]->attr.exclude_host)
2717 continue;
2718
2719 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
2720 }
2721 }
2722
2723 static int hsw_hw_config(struct perf_event *event)
2724 {
2725 int ret = intel_pmu_hw_config(event);
2726
2727 if (ret)
2728 return ret;
2729 if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE))
2730 return 0;
2731 event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED);
2732
2733 /*
2734 * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with
2735 * PEBS or in ANY thread mode. Since the results are non-sensical forbid
2736 * this combination.
2737 */
2738 if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) &&
2739 ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) ||
2740 event->attr.precise_ip > 0))
2741 return -EOPNOTSUPP;
2742
2743 if (event_is_checkpointed(event)) {
2744 /*
2745 * Sampling of checkpointed events can cause situations where
2746 * the CPU constantly aborts because of a overflow, which is
2747 * then checkpointed back and ignored. Forbid checkpointing
2748 * for sampling.
2749 *
2750 * But still allow a long sampling period, so that perf stat
2751 * from KVM works.
2752 */
2753 if (event->attr.sample_period > 0 &&
2754 event->attr.sample_period < 0x7fffffff)
2755 return -EOPNOTSUPP;
2756 }
2757 return 0;
2758 }
2759
2760 static struct event_constraint counter2_constraint =
2761 EVENT_CONSTRAINT(0, 0x4, 0);
2762
2763 static struct event_constraint *
2764 hsw_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
2765 struct perf_event *event)
2766 {
2767 struct event_constraint *c;
2768
2769 c = intel_get_event_constraints(cpuc, idx, event);
2770
2771 /* Handle special quirk on in_tx_checkpointed only in counter 2 */
2772 if (event->hw.config & HSW_IN_TX_CHECKPOINTED) {
2773 if (c->idxmsk64 & (1U << 2))
2774 return &counter2_constraint;
2775 return &emptyconstraint;
2776 }
2777
2778 return c;
2779 }
2780
2781 /*
2782 * Broadwell:
2783 *
2784 * The INST_RETIRED.ALL period always needs to have lowest 6 bits cleared
2785 * (BDM55) and it must not use a period smaller than 100 (BDM11). We combine
2786 * the two to enforce a minimum period of 128 (the smallest value that has bits
2787 * 0-5 cleared and >= 100).
2788 *
2789 * Because of how the code in x86_perf_event_set_period() works, the truncation
2790 * of the lower 6 bits is 'harmless' as we'll occasionally add a longer period
2791 * to make up for the 'lost' events due to carrying the 'error' in period_left.
2792 *
2793 * Therefore the effective (average) period matches the requested period,
2794 * despite coarser hardware granularity.
2795 */
2796 static unsigned bdw_limit_period(struct perf_event *event, unsigned left)
2797 {
2798 if ((event->hw.config & INTEL_ARCH_EVENT_MASK) ==
2799 X86_CONFIG(.event=0xc0, .umask=0x01)) {
2800 if (left < 128)
2801 left = 128;
2802 left &= ~0x3fu;
2803 }
2804 return left;
2805 }
2806
2807 PMU_FORMAT_ATTR(event, "config:0-7" );
2808 PMU_FORMAT_ATTR(umask, "config:8-15" );
2809 PMU_FORMAT_ATTR(edge, "config:18" );
2810 PMU_FORMAT_ATTR(pc, "config:19" );
2811 PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */
2812 PMU_FORMAT_ATTR(inv, "config:23" );
2813 PMU_FORMAT_ATTR(cmask, "config:24-31" );
2814 PMU_FORMAT_ATTR(in_tx, "config:32");
2815 PMU_FORMAT_ATTR(in_tx_cp, "config:33");
2816
2817 static struct attribute *intel_arch_formats_attr[] = {
2818 &format_attr_event.attr,
2819 &format_attr_umask.attr,
2820 &format_attr_edge.attr,
2821 &format_attr_pc.attr,
2822 &format_attr_inv.attr,
2823 &format_attr_cmask.attr,
2824 NULL,
2825 };
2826
2827 ssize_t intel_event_sysfs_show(char *page, u64 config)
2828 {
2829 u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT);
2830
2831 return x86_event_sysfs_show(page, config, event);
2832 }
2833
2834 struct intel_shared_regs *allocate_shared_regs(int cpu)
2835 {
2836 struct intel_shared_regs *regs;
2837 int i;
2838
2839 regs = kzalloc_node(sizeof(struct intel_shared_regs),
2840 GFP_KERNEL, cpu_to_node(cpu));
2841 if (regs) {
2842 /*
2843 * initialize the locks to keep lockdep happy
2844 */
2845 for (i = 0; i < EXTRA_REG_MAX; i++)
2846 raw_spin_lock_init(&regs->regs[i].lock);
2847
2848 regs->core_id = -1;
2849 }
2850 return regs;
2851 }
2852
2853 static struct intel_excl_cntrs *allocate_excl_cntrs(int cpu)
2854 {
2855 struct intel_excl_cntrs *c;
2856
2857 c = kzalloc_node(sizeof(struct intel_excl_cntrs),
2858 GFP_KERNEL, cpu_to_node(cpu));
2859 if (c) {
2860 raw_spin_lock_init(&c->lock);
2861 c->core_id = -1;
2862 }
2863 return c;
2864 }
2865
2866 static int intel_pmu_cpu_prepare(int cpu)
2867 {
2868 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
2869
2870 if (x86_pmu.extra_regs || x86_pmu.lbr_sel_map) {
2871 cpuc->shared_regs = allocate_shared_regs(cpu);
2872 if (!cpuc->shared_regs)
2873 goto err;
2874 }
2875
2876 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) {
2877 size_t sz = X86_PMC_IDX_MAX * sizeof(struct event_constraint);
2878
2879 cpuc->constraint_list = kzalloc(sz, GFP_KERNEL);
2880 if (!cpuc->constraint_list)
2881 goto err_shared_regs;
2882
2883 cpuc->excl_cntrs = allocate_excl_cntrs(cpu);
2884 if (!cpuc->excl_cntrs)
2885 goto err_constraint_list;
2886
2887 cpuc->excl_thread_id = 0;
2888 }
2889
2890 return NOTIFY_OK;
2891
2892 err_constraint_list:
2893 kfree(cpuc->constraint_list);
2894 cpuc->constraint_list = NULL;
2895
2896 err_shared_regs:
2897 kfree(cpuc->shared_regs);
2898 cpuc->shared_regs = NULL;
2899
2900 err:
2901 return NOTIFY_BAD;
2902 }
2903
2904 static void intel_pmu_cpu_starting(int cpu)
2905 {
2906 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
2907 int core_id = topology_core_id(cpu);
2908 int i;
2909
2910 init_debug_store_on_cpu(cpu);
2911 /*
2912 * Deal with CPUs that don't clear their LBRs on power-up.
2913 */
2914 intel_pmu_lbr_reset();
2915
2916 cpuc->lbr_sel = NULL;
2917
2918 if (!cpuc->shared_regs)
2919 return;
2920
2921 if (!(x86_pmu.flags & PMU_FL_NO_HT_SHARING)) {
2922 for_each_cpu(i, topology_sibling_cpumask(cpu)) {
2923 struct intel_shared_regs *pc;
2924
2925 pc = per_cpu(cpu_hw_events, i).shared_regs;
2926 if (pc && pc->core_id == core_id) {
2927 cpuc->kfree_on_online[0] = cpuc->shared_regs;
2928 cpuc->shared_regs = pc;
2929 break;
2930 }
2931 }
2932 cpuc->shared_regs->core_id = core_id;
2933 cpuc->shared_regs->refcnt++;
2934 }
2935
2936 if (x86_pmu.lbr_sel_map)
2937 cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR];
2938
2939 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) {
2940 for_each_cpu(i, topology_sibling_cpumask(cpu)) {
2941 struct intel_excl_cntrs *c;
2942
2943 c = per_cpu(cpu_hw_events, i).excl_cntrs;
2944 if (c && c->core_id == core_id) {
2945 cpuc->kfree_on_online[1] = cpuc->excl_cntrs;
2946 cpuc->excl_cntrs = c;
2947 cpuc->excl_thread_id = 1;
2948 break;
2949 }
2950 }
2951 cpuc->excl_cntrs->core_id = core_id;
2952 cpuc->excl_cntrs->refcnt++;
2953 }
2954 }
2955
2956 static void free_excl_cntrs(int cpu)
2957 {
2958 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
2959 struct intel_excl_cntrs *c;
2960
2961 c = cpuc->excl_cntrs;
2962 if (c) {
2963 if (c->core_id == -1 || --c->refcnt == 0)
2964 kfree(c);
2965 cpuc->excl_cntrs = NULL;
2966 kfree(cpuc->constraint_list);
2967 cpuc->constraint_list = NULL;
2968 }
2969 }
2970
2971 static void intel_pmu_cpu_dying(int cpu)
2972 {
2973 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
2974 struct intel_shared_regs *pc;
2975
2976 pc = cpuc->shared_regs;
2977 if (pc) {
2978 if (pc->core_id == -1 || --pc->refcnt == 0)
2979 kfree(pc);
2980 cpuc->shared_regs = NULL;
2981 }
2982
2983 free_excl_cntrs(cpu);
2984
2985 fini_debug_store_on_cpu(cpu);
2986 }
2987
2988 static void intel_pmu_sched_task(struct perf_event_context *ctx,
2989 bool sched_in)
2990 {
2991 if (x86_pmu.pebs_active)
2992 intel_pmu_pebs_sched_task(ctx, sched_in);
2993 if (x86_pmu.lbr_nr)
2994 intel_pmu_lbr_sched_task(ctx, sched_in);
2995 }
2996
2997 PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63");
2998
2999 PMU_FORMAT_ATTR(ldlat, "config1:0-15");
3000
3001 PMU_FORMAT_ATTR(frontend, "config1:0-23");
3002
3003 static struct attribute *intel_arch3_formats_attr[] = {
3004 &format_attr_event.attr,
3005 &format_attr_umask.attr,
3006 &format_attr_edge.attr,
3007 &format_attr_pc.attr,
3008 &format_attr_any.attr,
3009 &format_attr_inv.attr,
3010 &format_attr_cmask.attr,
3011 &format_attr_in_tx.attr,
3012 &format_attr_in_tx_cp.attr,
3013
3014 &format_attr_offcore_rsp.attr, /* XXX do NHM/WSM + SNB breakout */
3015 &format_attr_ldlat.attr, /* PEBS load latency */
3016 NULL,
3017 };
3018
3019 static struct attribute *skl_format_attr[] = {
3020 &format_attr_frontend.attr,
3021 NULL,
3022 };
3023
3024 static __initconst const struct x86_pmu core_pmu = {
3025 .name = "core",
3026 .handle_irq = x86_pmu_handle_irq,
3027 .disable_all = x86_pmu_disable_all,
3028 .enable_all = core_pmu_enable_all,
3029 .enable = core_pmu_enable_event,
3030 .disable = x86_pmu_disable_event,
3031 .hw_config = x86_pmu_hw_config,
3032 .schedule_events = x86_schedule_events,
3033 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
3034 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
3035 .event_map = intel_pmu_event_map,
3036 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
3037 .apic = 1,
3038 .free_running_flags = PEBS_FREERUNNING_FLAGS,
3039
3040 /*
3041 * Intel PMCs cannot be accessed sanely above 32-bit width,
3042 * so we install an artificial 1<<31 period regardless of
3043 * the generic event period:
3044 */
3045 .max_period = (1ULL<<31) - 1,
3046 .get_event_constraints = intel_get_event_constraints,
3047 .put_event_constraints = intel_put_event_constraints,
3048 .event_constraints = intel_core_event_constraints,
3049 .guest_get_msrs = core_guest_get_msrs,
3050 .format_attrs = intel_arch_formats_attr,
3051 .events_sysfs_show = intel_event_sysfs_show,
3052
3053 /*
3054 * Virtual (or funny metal) CPU can define x86_pmu.extra_regs
3055 * together with PMU version 1 and thus be using core_pmu with
3056 * shared_regs. We need following callbacks here to allocate
3057 * it properly.
3058 */
3059 .cpu_prepare = intel_pmu_cpu_prepare,
3060 .cpu_starting = intel_pmu_cpu_starting,
3061 .cpu_dying = intel_pmu_cpu_dying,
3062 };
3063
3064 static __initconst const struct x86_pmu intel_pmu = {
3065 .name = "Intel",
3066 .handle_irq = intel_pmu_handle_irq,
3067 .disable_all = intel_pmu_disable_all,
3068 .enable_all = intel_pmu_enable_all,
3069 .enable = intel_pmu_enable_event,
3070 .disable = intel_pmu_disable_event,
3071 .hw_config = intel_pmu_hw_config,
3072 .schedule_events = x86_schedule_events,
3073 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0,
3074 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
3075 .event_map = intel_pmu_event_map,
3076 .max_events = ARRAY_SIZE(intel_perfmon_event_map),
3077 .apic = 1,
3078 .free_running_flags = PEBS_FREERUNNING_FLAGS,
3079 /*
3080 * Intel PMCs cannot be accessed sanely above 32 bit width,
3081 * so we install an artificial 1<<31 period regardless of
3082 * the generic event period:
3083 */
3084 .max_period = (1ULL << 31) - 1,
3085 .get_event_constraints = intel_get_event_constraints,
3086 .put_event_constraints = intel_put_event_constraints,
3087 .pebs_aliases = intel_pebs_aliases_core2,
3088
3089 .format_attrs = intel_arch3_formats_attr,
3090 .events_sysfs_show = intel_event_sysfs_show,
3091
3092 .cpu_prepare = intel_pmu_cpu_prepare,
3093 .cpu_starting = intel_pmu_cpu_starting,
3094 .cpu_dying = intel_pmu_cpu_dying,
3095 .guest_get_msrs = intel_guest_get_msrs,
3096 .sched_task = intel_pmu_sched_task,
3097 };
3098
3099 static __init void intel_clovertown_quirk(void)
3100 {
3101 /*
3102 * PEBS is unreliable due to:
3103 *
3104 * AJ67 - PEBS may experience CPL leaks
3105 * AJ68 - PEBS PMI may be delayed by one event
3106 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
3107 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
3108 *
3109 * AJ67 could be worked around by restricting the OS/USR flags.
3110 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
3111 *
3112 * AJ106 could possibly be worked around by not allowing LBR
3113 * usage from PEBS, including the fixup.
3114 * AJ68 could possibly be worked around by always programming
3115 * a pebs_event_reset[0] value and coping with the lost events.
3116 *
3117 * But taken together it might just make sense to not enable PEBS on
3118 * these chips.
3119 */
3120 pr_warn("PEBS disabled due to CPU errata\n");
3121 x86_pmu.pebs = 0;
3122 x86_pmu.pebs_constraints = NULL;
3123 }
3124
3125 static int intel_snb_pebs_broken(int cpu)
3126 {
3127 u32 rev = UINT_MAX; /* default to broken for unknown models */
3128
3129 switch (cpu_data(cpu).x86_model) {
3130 case 42: /* SNB */
3131 rev = 0x28;
3132 break;
3133
3134 case 45: /* SNB-EP */
3135 switch (cpu_data(cpu).x86_mask) {
3136 case 6: rev = 0x618; break;
3137 case 7: rev = 0x70c; break;
3138 }
3139 }
3140
3141 return (cpu_data(cpu).microcode < rev);
3142 }
3143
3144 static void intel_snb_check_microcode(void)
3145 {
3146 int pebs_broken = 0;
3147 int cpu;
3148
3149 get_online_cpus();
3150 for_each_online_cpu(cpu) {
3151 if ((pebs_broken = intel_snb_pebs_broken(cpu)))
3152 break;
3153 }
3154 put_online_cpus();
3155
3156 if (pebs_broken == x86_pmu.pebs_broken)
3157 return;
3158
3159 /*
3160 * Serialized by the microcode lock..
3161 */
3162 if (x86_pmu.pebs_broken) {
3163 pr_info("PEBS enabled due to microcode update\n");
3164 x86_pmu.pebs_broken = 0;
3165 } else {
3166 pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n");
3167 x86_pmu.pebs_broken = 1;
3168 }
3169 }
3170
3171 /*
3172 * Under certain circumstances, access certain MSR may cause #GP.
3173 * The function tests if the input MSR can be safely accessed.
3174 */
3175 static bool check_msr(unsigned long msr, u64 mask)
3176 {
3177 u64 val_old, val_new, val_tmp;
3178
3179 /*
3180 * Read the current value, change it and read it back to see if it
3181 * matches, this is needed to detect certain hardware emulators
3182 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
3183 */
3184 if (rdmsrl_safe(msr, &val_old))
3185 return false;
3186
3187 /*
3188 * Only change the bits which can be updated by wrmsrl.
3189 */
3190 val_tmp = val_old ^ mask;
3191 if (wrmsrl_safe(msr, val_tmp) ||
3192 rdmsrl_safe(msr, &val_new))
3193 return false;
3194
3195 if (val_new != val_tmp)
3196 return false;
3197
3198 /* Here it's sure that the MSR can be safely accessed.
3199 * Restore the old value and return.
3200 */
3201 wrmsrl(msr, val_old);
3202
3203 return true;
3204 }
3205
3206 static __init void intel_sandybridge_quirk(void)
3207 {
3208 x86_pmu.check_microcode = intel_snb_check_microcode;
3209 intel_snb_check_microcode();
3210 }
3211
3212 static const struct { int id; char *name; } intel_arch_events_map[] __initconst = {
3213 { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" },
3214 { PERF_COUNT_HW_INSTRUCTIONS, "instructions" },
3215 { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" },
3216 { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" },
3217 { PERF_COUNT_HW_CACHE_MISSES, "cache misses" },
3218 { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" },
3219 { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" },
3220 };
3221
3222 static __init void intel_arch_events_quirk(void)
3223 {
3224 int bit;
3225
3226 /* disable event that reported as not presend by cpuid */
3227 for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) {
3228 intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0;
3229 pr_warn("CPUID marked event: \'%s\' unavailable\n",
3230 intel_arch_events_map[bit].name);
3231 }
3232 }
3233
3234 static __init void intel_nehalem_quirk(void)
3235 {
3236 union cpuid10_ebx ebx;
3237
3238 ebx.full = x86_pmu.events_maskl;
3239 if (ebx.split.no_branch_misses_retired) {
3240 /*
3241 * Erratum AAJ80 detected, we work it around by using
3242 * the BR_MISP_EXEC.ANY event. This will over-count
3243 * branch-misses, but it's still much better than the
3244 * architectural event which is often completely bogus:
3245 */
3246 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89;
3247 ebx.split.no_branch_misses_retired = 0;
3248 x86_pmu.events_maskl = ebx.full;
3249 pr_info("CPU erratum AAJ80 worked around\n");
3250 }
3251 }
3252
3253 /*
3254 * enable software workaround for errata:
3255 * SNB: BJ122
3256 * IVB: BV98
3257 * HSW: HSD29
3258 *
3259 * Only needed when HT is enabled. However detecting
3260 * if HT is enabled is difficult (model specific). So instead,
3261 * we enable the workaround in the early boot, and verify if
3262 * it is needed in a later initcall phase once we have valid
3263 * topology information to check if HT is actually enabled
3264 */
3265 static __init void intel_ht_bug(void)
3266 {
3267 x86_pmu.flags |= PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED;
3268
3269 x86_pmu.start_scheduling = intel_start_scheduling;
3270 x86_pmu.commit_scheduling = intel_commit_scheduling;
3271 x86_pmu.stop_scheduling = intel_stop_scheduling;
3272 }
3273
3274 EVENT_ATTR_STR(mem-loads, mem_ld_hsw, "event=0xcd,umask=0x1,ldlat=3");
3275 EVENT_ATTR_STR(mem-stores, mem_st_hsw, "event=0xd0,umask=0x82")
3276
3277 /* Haswell special events */
3278 EVENT_ATTR_STR(tx-start, tx_start, "event=0xc9,umask=0x1");
3279 EVENT_ATTR_STR(tx-commit, tx_commit, "event=0xc9,umask=0x2");
3280 EVENT_ATTR_STR(tx-abort, tx_abort, "event=0xc9,umask=0x4");
3281 EVENT_ATTR_STR(tx-capacity, tx_capacity, "event=0x54,umask=0x2");
3282 EVENT_ATTR_STR(tx-conflict, tx_conflict, "event=0x54,umask=0x1");
3283 EVENT_ATTR_STR(el-start, el_start, "event=0xc8,umask=0x1");
3284 EVENT_ATTR_STR(el-commit, el_commit, "event=0xc8,umask=0x2");
3285 EVENT_ATTR_STR(el-abort, el_abort, "event=0xc8,umask=0x4");
3286 EVENT_ATTR_STR(el-capacity, el_capacity, "event=0x54,umask=0x2");
3287 EVENT_ATTR_STR(el-conflict, el_conflict, "event=0x54,umask=0x1");
3288 EVENT_ATTR_STR(cycles-t, cycles_t, "event=0x3c,in_tx=1");
3289 EVENT_ATTR_STR(cycles-ct, cycles_ct, "event=0x3c,in_tx=1,in_tx_cp=1");
3290
3291 static struct attribute *hsw_events_attrs[] = {
3292 EVENT_PTR(tx_start),
3293 EVENT_PTR(tx_commit),
3294 EVENT_PTR(tx_abort),
3295 EVENT_PTR(tx_capacity),
3296 EVENT_PTR(tx_conflict),
3297 EVENT_PTR(el_start),
3298 EVENT_PTR(el_commit),
3299 EVENT_PTR(el_abort),
3300 EVENT_PTR(el_capacity),
3301 EVENT_PTR(el_conflict),
3302 EVENT_PTR(cycles_t),
3303 EVENT_PTR(cycles_ct),
3304 EVENT_PTR(mem_ld_hsw),
3305 EVENT_PTR(mem_st_hsw),
3306 NULL
3307 };
3308
3309 __init int intel_pmu_init(void)
3310 {
3311 union cpuid10_edx edx;
3312 union cpuid10_eax eax;
3313 union cpuid10_ebx ebx;
3314 struct event_constraint *c;
3315 unsigned int unused;
3316 struct extra_reg *er;
3317 int version, i;
3318
3319 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
3320 switch (boot_cpu_data.x86) {
3321 case 0x6:
3322 return p6_pmu_init();
3323 case 0xb:
3324 return knc_pmu_init();
3325 case 0xf:
3326 return p4_pmu_init();
3327 }
3328 return -ENODEV;
3329 }
3330
3331 /*
3332 * Check whether the Architectural PerfMon supports
3333 * Branch Misses Retired hw_event or not.
3334 */
3335 cpuid(10, &eax.full, &ebx.full, &unused, &edx.full);
3336 if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT)
3337 return -ENODEV;
3338
3339 version = eax.split.version_id;
3340 if (version < 2)
3341 x86_pmu = core_pmu;
3342 else
3343 x86_pmu = intel_pmu;
3344
3345 x86_pmu.version = version;
3346 x86_pmu.num_counters = eax.split.num_counters;
3347 x86_pmu.cntval_bits = eax.split.bit_width;
3348 x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1;
3349
3350 x86_pmu.events_maskl = ebx.full;
3351 x86_pmu.events_mask_len = eax.split.mask_length;
3352
3353 x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters);
3354
3355 /*
3356 * Quirk: v2 perfmon does not report fixed-purpose events, so
3357 * assume at least 3 events:
3358 */
3359 if (version > 1)
3360 x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3);
3361
3362 if (boot_cpu_has(X86_FEATURE_PDCM)) {
3363 u64 capabilities;
3364
3365 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
3366 x86_pmu.intel_cap.capabilities = capabilities;
3367 }
3368
3369 intel_ds_init();
3370
3371 x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */
3372
3373 /*
3374 * Install the hw-cache-events table:
3375 */
3376 switch (boot_cpu_data.x86_model) {
3377 case 14: /* 65nm Core "Yonah" */
3378 pr_cont("Core events, ");
3379 break;
3380
3381 case 15: /* 65nm Core2 "Merom" */
3382 x86_add_quirk(intel_clovertown_quirk);
3383 case 22: /* 65nm Core2 "Merom-L" */
3384 case 23: /* 45nm Core2 "Penryn" */
3385 case 29: /* 45nm Core2 "Dunnington (MP) */
3386 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
3387 sizeof(hw_cache_event_ids));
3388
3389 intel_pmu_lbr_init_core();
3390
3391 x86_pmu.event_constraints = intel_core2_event_constraints;
3392 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints;
3393 pr_cont("Core2 events, ");
3394 break;
3395
3396 case 30: /* 45nm Nehalem */
3397 case 26: /* 45nm Nehalem-EP */
3398 case 46: /* 45nm Nehalem-EX */
3399 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
3400 sizeof(hw_cache_event_ids));
3401 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
3402 sizeof(hw_cache_extra_regs));
3403
3404 intel_pmu_lbr_init_nhm();
3405
3406 x86_pmu.event_constraints = intel_nehalem_event_constraints;
3407 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints;
3408 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
3409 x86_pmu.extra_regs = intel_nehalem_extra_regs;
3410
3411 x86_pmu.cpu_events = nhm_events_attrs;
3412
3413 /* UOPS_ISSUED.STALLED_CYCLES */
3414 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
3415 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
3416 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
3417 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
3418 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
3419
3420 intel_pmu_pebs_data_source_nhm();
3421 x86_add_quirk(intel_nehalem_quirk);
3422
3423 pr_cont("Nehalem events, ");
3424 break;
3425
3426 case 28: /* 45nm Atom "Pineview" */
3427 case 38: /* 45nm Atom "Lincroft" */
3428 case 39: /* 32nm Atom "Penwell" */
3429 case 53: /* 32nm Atom "Cloverview" */
3430 case 54: /* 32nm Atom "Cedarview" */
3431 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
3432 sizeof(hw_cache_event_ids));
3433
3434 intel_pmu_lbr_init_atom();
3435
3436 x86_pmu.event_constraints = intel_gen_event_constraints;
3437 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints;
3438 x86_pmu.pebs_aliases = intel_pebs_aliases_core2;
3439 pr_cont("Atom events, ");
3440 break;
3441
3442 case 55: /* 22nm Atom "Silvermont" */
3443 case 76: /* 14nm Atom "Airmont" */
3444 case 77: /* 22nm Atom "Silvermont Avoton/Rangely" */
3445 memcpy(hw_cache_event_ids, slm_hw_cache_event_ids,
3446 sizeof(hw_cache_event_ids));
3447 memcpy(hw_cache_extra_regs, slm_hw_cache_extra_regs,
3448 sizeof(hw_cache_extra_regs));
3449
3450 intel_pmu_lbr_init_atom();
3451
3452 x86_pmu.event_constraints = intel_slm_event_constraints;
3453 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints;
3454 x86_pmu.extra_regs = intel_slm_extra_regs;
3455 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
3456 pr_cont("Silvermont events, ");
3457 break;
3458
3459 case 37: /* 32nm Westmere */
3460 case 44: /* 32nm Westmere-EP */
3461 case 47: /* 32nm Westmere-EX */
3462 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
3463 sizeof(hw_cache_event_ids));
3464 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
3465 sizeof(hw_cache_extra_regs));
3466
3467 intel_pmu_lbr_init_nhm();
3468
3469 x86_pmu.event_constraints = intel_westmere_event_constraints;
3470 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
3471 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints;
3472 x86_pmu.extra_regs = intel_westmere_extra_regs;
3473 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
3474
3475 x86_pmu.cpu_events = nhm_events_attrs;
3476
3477 /* UOPS_ISSUED.STALLED_CYCLES */
3478 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
3479 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
3480 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
3481 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
3482 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
3483
3484 intel_pmu_pebs_data_source_nhm();
3485 pr_cont("Westmere events, ");
3486 break;
3487
3488 case 42: /* 32nm SandyBridge */
3489 case 45: /* 32nm SandyBridge-E/EN/EP */
3490 x86_add_quirk(intel_sandybridge_quirk);
3491 x86_add_quirk(intel_ht_bug);
3492 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
3493 sizeof(hw_cache_event_ids));
3494 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
3495 sizeof(hw_cache_extra_regs));
3496
3497 intel_pmu_lbr_init_snb();
3498
3499 x86_pmu.event_constraints = intel_snb_event_constraints;
3500 x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints;
3501 x86_pmu.pebs_aliases = intel_pebs_aliases_snb;
3502 if (boot_cpu_data.x86_model == 45)
3503 x86_pmu.extra_regs = intel_snbep_extra_regs;
3504 else
3505 x86_pmu.extra_regs = intel_snb_extra_regs;
3506
3507
3508 /* all extra regs are per-cpu when HT is on */
3509 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
3510 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
3511
3512 x86_pmu.cpu_events = snb_events_attrs;
3513
3514 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
3515 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
3516 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
3517 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
3518 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
3519 X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1);
3520
3521 pr_cont("SandyBridge events, ");
3522 break;
3523
3524 case 58: /* 22nm IvyBridge */
3525 case 62: /* 22nm IvyBridge-EP/EX */
3526 x86_add_quirk(intel_ht_bug);
3527 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
3528 sizeof(hw_cache_event_ids));
3529 /* dTLB-load-misses on IVB is different than SNB */
3530 hw_cache_event_ids[C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = 0x8108; /* DTLB_LOAD_MISSES.DEMAND_LD_MISS_CAUSES_A_WALK */
3531
3532 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
3533 sizeof(hw_cache_extra_regs));
3534
3535 intel_pmu_lbr_init_snb();
3536
3537 x86_pmu.event_constraints = intel_ivb_event_constraints;
3538 x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints;
3539 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
3540 x86_pmu.pebs_prec_dist = true;
3541 if (boot_cpu_data.x86_model == 62)
3542 x86_pmu.extra_regs = intel_snbep_extra_regs;
3543 else
3544 x86_pmu.extra_regs = intel_snb_extra_regs;
3545 /* all extra regs are per-cpu when HT is on */
3546 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
3547 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
3548
3549 x86_pmu.cpu_events = snb_events_attrs;
3550
3551 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
3552 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
3553 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
3554
3555 pr_cont("IvyBridge events, ");
3556 break;
3557
3558
3559 case 60: /* 22nm Haswell Core */
3560 case 63: /* 22nm Haswell Server */
3561 case 69: /* 22nm Haswell ULT */
3562 case 70: /* 22nm Haswell + GT3e (Intel Iris Pro graphics) */
3563 x86_add_quirk(intel_ht_bug);
3564 x86_pmu.late_ack = true;
3565 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids));
3566 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
3567
3568 intel_pmu_lbr_init_hsw();
3569
3570 x86_pmu.event_constraints = intel_hsw_event_constraints;
3571 x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints;
3572 x86_pmu.extra_regs = intel_snbep_extra_regs;
3573 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
3574 x86_pmu.pebs_prec_dist = true;
3575 /* all extra regs are per-cpu when HT is on */
3576 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
3577 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
3578
3579 x86_pmu.hw_config = hsw_hw_config;
3580 x86_pmu.get_event_constraints = hsw_get_event_constraints;
3581 x86_pmu.cpu_events = hsw_events_attrs;
3582 x86_pmu.lbr_double_abort = true;
3583 pr_cont("Haswell events, ");
3584 break;
3585
3586 case 61: /* 14nm Broadwell Core-M */
3587 case 86: /* 14nm Broadwell Xeon D */
3588 case 71: /* 14nm Broadwell + GT3e (Intel Iris Pro graphics) */
3589 case 79: /* 14nm Broadwell Server */
3590 x86_pmu.late_ack = true;
3591 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids));
3592 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
3593
3594 /* L3_MISS_LOCAL_DRAM is BIT(26) in Broadwell */
3595 hw_cache_extra_regs[C(LL)][C(OP_READ)][C(RESULT_MISS)] = HSW_DEMAND_READ |
3596 BDW_L3_MISS|HSW_SNOOP_DRAM;
3597 hw_cache_extra_regs[C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = HSW_DEMAND_WRITE|BDW_L3_MISS|
3598 HSW_SNOOP_DRAM;
3599 hw_cache_extra_regs[C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = HSW_DEMAND_READ|
3600 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM;
3601 hw_cache_extra_regs[C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = HSW_DEMAND_WRITE|
3602 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM;
3603
3604 intel_pmu_lbr_init_hsw();
3605
3606 x86_pmu.event_constraints = intel_bdw_event_constraints;
3607 x86_pmu.pebs_constraints = intel_bdw_pebs_event_constraints;
3608 x86_pmu.extra_regs = intel_snbep_extra_regs;
3609 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
3610 x86_pmu.pebs_prec_dist = true;
3611 /* all extra regs are per-cpu when HT is on */
3612 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
3613 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
3614
3615 x86_pmu.hw_config = hsw_hw_config;
3616 x86_pmu.get_event_constraints = hsw_get_event_constraints;
3617 x86_pmu.cpu_events = hsw_events_attrs;
3618 x86_pmu.limit_period = bdw_limit_period;
3619 pr_cont("Broadwell events, ");
3620 break;
3621
3622 case 87: /* Knights Landing Xeon Phi */
3623 memcpy(hw_cache_event_ids,
3624 slm_hw_cache_event_ids, sizeof(hw_cache_event_ids));
3625 memcpy(hw_cache_extra_regs,
3626 knl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
3627 intel_pmu_lbr_init_knl();
3628
3629 x86_pmu.event_constraints = intel_slm_event_constraints;
3630 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints;
3631 x86_pmu.extra_regs = intel_knl_extra_regs;
3632
3633 /* all extra regs are per-cpu when HT is on */
3634 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
3635 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
3636
3637 pr_cont("Knights Landing events, ");
3638 break;
3639
3640 case 142: /* 14nm Kabylake Mobile */
3641 case 158: /* 14nm Kabylake Desktop */
3642 case 78: /* 14nm Skylake Mobile */
3643 case 94: /* 14nm Skylake Desktop */
3644 case 85: /* 14nm Skylake Server */
3645 x86_pmu.late_ack = true;
3646 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids));
3647 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
3648 intel_pmu_lbr_init_skl();
3649
3650 x86_pmu.event_constraints = intel_skl_event_constraints;
3651 x86_pmu.pebs_constraints = intel_skl_pebs_event_constraints;
3652 x86_pmu.extra_regs = intel_skl_extra_regs;
3653 x86_pmu.pebs_aliases = intel_pebs_aliases_skl;
3654 x86_pmu.pebs_prec_dist = true;
3655 /* all extra regs are per-cpu when HT is on */
3656 x86_pmu.flags |= PMU_FL_HAS_RSP_1;
3657 x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
3658
3659 x86_pmu.hw_config = hsw_hw_config;
3660 x86_pmu.get_event_constraints = hsw_get_event_constraints;
3661 x86_pmu.format_attrs = merge_attr(intel_arch3_formats_attr,
3662 skl_format_attr);
3663 WARN_ON(!x86_pmu.format_attrs);
3664 x86_pmu.cpu_events = hsw_events_attrs;
3665 pr_cont("Skylake events, ");
3666 break;
3667
3668 default:
3669 switch (x86_pmu.version) {
3670 case 1:
3671 x86_pmu.event_constraints = intel_v1_event_constraints;
3672 pr_cont("generic architected perfmon v1, ");
3673 break;
3674 default:
3675 /*
3676 * default constraints for v2 and up
3677 */
3678 x86_pmu.event_constraints = intel_gen_event_constraints;
3679 pr_cont("generic architected perfmon, ");
3680 break;
3681 }
3682 }
3683
3684 if (x86_pmu.num_counters > INTEL_PMC_MAX_GENERIC) {
3685 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
3686 x86_pmu.num_counters, INTEL_PMC_MAX_GENERIC);
3687 x86_pmu.num_counters = INTEL_PMC_MAX_GENERIC;
3688 }
3689 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
3690
3691 if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) {
3692 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
3693 x86_pmu.num_counters_fixed, INTEL_PMC_MAX_FIXED);
3694 x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED;
3695 }
3696
3697 x86_pmu.intel_ctrl |=
3698 ((1LL << x86_pmu.num_counters_fixed)-1) << INTEL_PMC_IDX_FIXED;
3699
3700 if (x86_pmu.event_constraints) {
3701 /*
3702 * event on fixed counter2 (REF_CYCLES) only works on this
3703 * counter, so do not extend mask to generic counters
3704 */
3705 for_each_event_constraint(c, x86_pmu.event_constraints) {
3706 if (c->cmask == FIXED_EVENT_FLAGS
3707 && c->idxmsk64 != INTEL_PMC_MSK_FIXED_REF_CYCLES) {
3708 c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1;
3709 }
3710 c->idxmsk64 &=
3711 ~(~0UL << (INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed));
3712 c->weight = hweight64(c->idxmsk64);
3713 }
3714 }
3715
3716 /*
3717 * Access LBR MSR may cause #GP under certain circumstances.
3718 * E.g. KVM doesn't support LBR MSR
3719 * Check all LBT MSR here.
3720 * Disable LBR access if any LBR MSRs can not be accessed.
3721 */
3722 if (x86_pmu.lbr_nr && !check_msr(x86_pmu.lbr_tos, 0x3UL))
3723 x86_pmu.lbr_nr = 0;
3724 for (i = 0; i < x86_pmu.lbr_nr; i++) {
3725 if (!(check_msr(x86_pmu.lbr_from + i, 0xffffUL) &&
3726 check_msr(x86_pmu.lbr_to + i, 0xffffUL)))
3727 x86_pmu.lbr_nr = 0;
3728 }
3729
3730 /*
3731 * Access extra MSR may cause #GP under certain circumstances.
3732 * E.g. KVM doesn't support offcore event
3733 * Check all extra_regs here.
3734 */
3735 if (x86_pmu.extra_regs) {
3736 for (er = x86_pmu.extra_regs; er->msr; er++) {
3737 er->extra_msr_access = check_msr(er->msr, 0x11UL);
3738 /* Disable LBR select mapping */
3739 if ((er->idx == EXTRA_REG_LBR) && !er->extra_msr_access)
3740 x86_pmu.lbr_sel_map = NULL;
3741 }
3742 }
3743
3744 /* Support full width counters using alternative MSR range */
3745 if (x86_pmu.intel_cap.full_width_write) {
3746 x86_pmu.max_period = x86_pmu.cntval_mask;
3747 x86_pmu.perfctr = MSR_IA32_PMC0;
3748 pr_cont("full-width counters, ");
3749 }
3750
3751 return 0;
3752 }
3753
3754 /*
3755 * HT bug: phase 2 init
3756 * Called once we have valid topology information to check
3757 * whether or not HT is enabled
3758 * If HT is off, then we disable the workaround
3759 */
3760 static __init int fixup_ht_bug(void)
3761 {
3762 int cpu = smp_processor_id();
3763 int w, c;
3764 /*
3765 * problem not present on this CPU model, nothing to do
3766 */
3767 if (!(x86_pmu.flags & PMU_FL_EXCL_ENABLED))
3768 return 0;
3769
3770 w = cpumask_weight(topology_sibling_cpumask(cpu));
3771 if (w > 1) {
3772 pr_info("PMU erratum BJ122, BV98, HSD29 worked around, HT is on\n");
3773 return 0;
3774 }
3775
3776 if (lockup_detector_suspend() != 0) {
3777 pr_debug("failed to disable PMU erratum BJ122, BV98, HSD29 workaround\n");
3778 return 0;
3779 }
3780
3781 x86_pmu.flags &= ~(PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED);
3782
3783 x86_pmu.start_scheduling = NULL;
3784 x86_pmu.commit_scheduling = NULL;
3785 x86_pmu.stop_scheduling = NULL;
3786
3787 lockup_detector_resume();
3788
3789 get_online_cpus();
3790
3791 for_each_online_cpu(c) {
3792 free_excl_cntrs(c);
3793 }
3794
3795 put_online_cpus();
3796 pr_info("PMU erratum BJ122, BV98, HSD29 workaround disabled, HT off\n");
3797 return 0;
3798 }
3799 subsys_initcall(fixup_ht_bug)
Linux kernel - Deliverables
This page took 0.15665 seconds and 5 git commands to generate.