Commit | Line | Data |
---|---|---|
f22f54f4 PZ |
1 | #ifdef CONFIG_CPU_SUP_INTEL |
2 | ||
a7e3ed1e | 3 | /* |
efc9f05d SE |
4 | * Per core/cpu state |
5 | * | |
6 | * Used to coordinate shared registers between HT threads or | |
7 | * among events on a single PMU. | |
a7e3ed1e | 8 | */ |
efc9f05d SE |
9 | struct intel_shared_regs { |
10 | struct er_account regs[EXTRA_REG_MAX]; | |
11 | int refcnt; /* per-core: #HT threads */ | |
12 | unsigned core_id; /* per-core: core id */ | |
a7e3ed1e AK |
13 | }; |
14 | ||
f22f54f4 | 15 | /* |
b622d644 | 16 | * Intel PerfMon, used on Core and later. |
f22f54f4 | 17 | */ |
ec75a716 | 18 | static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly = |
f22f54f4 PZ |
19 | { |
20 | [PERF_COUNT_HW_CPU_CYCLES] = 0x003c, | |
21 | [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0, | |
22 | [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e, | |
23 | [PERF_COUNT_HW_CACHE_MISSES] = 0x412e, | |
24 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4, | |
25 | [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5, | |
26 | [PERF_COUNT_HW_BUS_CYCLES] = 0x013c, | |
27 | }; | |
28 | ||
5c543e3c | 29 | static struct event_constraint intel_core_event_constraints[] __read_mostly = |
f22f54f4 PZ |
30 | { |
31 | INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ | |
32 | INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ | |
33 | INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */ | |
34 | INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */ | |
35 | INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */ | |
36 | INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */ | |
37 | EVENT_CONSTRAINT_END | |
38 | }; | |
39 | ||
5c543e3c | 40 | static struct event_constraint intel_core2_event_constraints[] __read_mostly = |
f22f54f4 | 41 | { |
b622d644 PZ |
42 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
43 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
44 | /* | |
45 | * Core2 has Fixed Counter 2 listed as CPU_CLK_UNHALTED.REF and event | |
46 | * 0x013c as CPU_CLK_UNHALTED.BUS and specifies there is a fixed | |
47 | * ratio between these counters. | |
48 | */ | |
49 | /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */ | |
f22f54f4 PZ |
50 | INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */ |
51 | INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ | |
52 | INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ | |
53 | INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */ | |
54 | INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */ | |
55 | INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */ | |
56 | INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */ | |
57 | INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */ | |
b622d644 | 58 | INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */ |
f22f54f4 PZ |
59 | INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */ |
60 | EVENT_CONSTRAINT_END | |
61 | }; | |
62 | ||
5c543e3c | 63 | static struct event_constraint intel_nehalem_event_constraints[] __read_mostly = |
f22f54f4 | 64 | { |
b622d644 PZ |
65 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
66 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
67 | /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */ | |
f22f54f4 PZ |
68 | INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */ |
69 | INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */ | |
70 | INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */ | |
71 | INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */ | |
72 | INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */ | |
73 | INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */ | |
74 | INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ | |
75 | INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ | |
76 | EVENT_CONSTRAINT_END | |
77 | }; | |
78 | ||
5c543e3c | 79 | static struct extra_reg intel_nehalem_extra_regs[] __read_mostly = |
a7e3ed1e | 80 | { |
efc9f05d | 81 | INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), |
a7e3ed1e AK |
82 | EVENT_EXTRA_END |
83 | }; | |
84 | ||
5c543e3c | 85 | static struct event_constraint intel_westmere_event_constraints[] __read_mostly = |
f22f54f4 | 86 | { |
b622d644 PZ |
87 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
88 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
89 | /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */ | |
f22f54f4 PZ |
90 | INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ |
91 | INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */ | |
92 | INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ | |
d1100770 | 93 | INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */ |
f22f54f4 PZ |
94 | EVENT_CONSTRAINT_END |
95 | }; | |
96 | ||
5c543e3c | 97 | static struct event_constraint intel_snb_event_constraints[] __read_mostly = |
b06b3d49 LM |
98 | { |
99 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | |
100 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
101 | /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */ | |
102 | INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */ | |
b06b3d49 LM |
103 | INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ |
104 | INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ | |
105 | EVENT_CONSTRAINT_END | |
106 | }; | |
107 | ||
5c543e3c | 108 | static struct extra_reg intel_westmere_extra_regs[] __read_mostly = |
a7e3ed1e | 109 | { |
efc9f05d SE |
110 | INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), |
111 | INTEL_EVENT_EXTRA_REG(0xbb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1), | |
a7e3ed1e AK |
112 | EVENT_EXTRA_END |
113 | }; | |
114 | ||
5c543e3c | 115 | static struct event_constraint intel_gen_event_constraints[] __read_mostly = |
f22f54f4 | 116 | { |
b622d644 PZ |
117 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
118 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
119 | /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */ | |
f22f54f4 PZ |
120 | EVENT_CONSTRAINT_END |
121 | }; | |
122 | ||
ee89cbc2 SE |
123 | static struct extra_reg intel_snb_extra_regs[] __read_mostly = { |
124 | INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0x3fffffffffull, RSP_0), | |
125 | INTEL_EVENT_EXTRA_REG(0xbb, MSR_OFFCORE_RSP_1, 0x3fffffffffull, RSP_1), | |
126 | EVENT_EXTRA_END | |
127 | }; | |
128 | ||
f22f54f4 PZ |
129 | static u64 intel_pmu_event_map(int hw_event) |
130 | { | |
131 | return intel_perfmon_event_map[hw_event]; | |
132 | } | |
133 | ||
b06b3d49 LM |
134 | static __initconst const u64 snb_hw_cache_event_ids |
135 | [PERF_COUNT_HW_CACHE_MAX] | |
136 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
137 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
138 | { | |
139 | [ C(L1D) ] = { | |
140 | [ C(OP_READ) ] = { | |
141 | [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */ | |
142 | [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */ | |
143 | }, | |
144 | [ C(OP_WRITE) ] = { | |
145 | [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */ | |
146 | [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */ | |
147 | }, | |
148 | [ C(OP_PREFETCH) ] = { | |
149 | [ C(RESULT_ACCESS) ] = 0x0, | |
150 | [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */ | |
151 | }, | |
152 | }, | |
153 | [ C(L1I ) ] = { | |
154 | [ C(OP_READ) ] = { | |
155 | [ C(RESULT_ACCESS) ] = 0x0, | |
156 | [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */ | |
157 | }, | |
158 | [ C(OP_WRITE) ] = { | |
159 | [ C(RESULT_ACCESS) ] = -1, | |
160 | [ C(RESULT_MISS) ] = -1, | |
161 | }, | |
162 | [ C(OP_PREFETCH) ] = { | |
163 | [ C(RESULT_ACCESS) ] = 0x0, | |
164 | [ C(RESULT_MISS) ] = 0x0, | |
165 | }, | |
166 | }, | |
167 | [ C(LL ) ] = { | |
b06b3d49 | 168 | [ C(OP_READ) ] = { |
63b6a675 | 169 | /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ |
b06b3d49 | 170 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
171 | /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ |
172 | [ C(RESULT_MISS) ] = 0x01b7, | |
b06b3d49 LM |
173 | }, |
174 | [ C(OP_WRITE) ] = { | |
63b6a675 | 175 | /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ |
b06b3d49 | 176 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
177 | /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ |
178 | [ C(RESULT_MISS) ] = 0x01b7, | |
b06b3d49 LM |
179 | }, |
180 | [ C(OP_PREFETCH) ] = { | |
63b6a675 | 181 | /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ |
b06b3d49 | 182 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
183 | /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ |
184 | [ C(RESULT_MISS) ] = 0x01b7, | |
b06b3d49 LM |
185 | }, |
186 | }, | |
187 | [ C(DTLB) ] = { | |
188 | [ C(OP_READ) ] = { | |
189 | [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */ | |
190 | [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */ | |
191 | }, | |
192 | [ C(OP_WRITE) ] = { | |
193 | [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */ | |
194 | [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */ | |
195 | }, | |
196 | [ C(OP_PREFETCH) ] = { | |
197 | [ C(RESULT_ACCESS) ] = 0x0, | |
198 | [ C(RESULT_MISS) ] = 0x0, | |
199 | }, | |
200 | }, | |
201 | [ C(ITLB) ] = { | |
202 | [ C(OP_READ) ] = { | |
203 | [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */ | |
204 | [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */ | |
205 | }, | |
206 | [ C(OP_WRITE) ] = { | |
207 | [ C(RESULT_ACCESS) ] = -1, | |
208 | [ C(RESULT_MISS) ] = -1, | |
209 | }, | |
210 | [ C(OP_PREFETCH) ] = { | |
211 | [ C(RESULT_ACCESS) ] = -1, | |
212 | [ C(RESULT_MISS) ] = -1, | |
213 | }, | |
214 | }, | |
215 | [ C(BPU ) ] = { | |
216 | [ C(OP_READ) ] = { | |
217 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
218 | [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ | |
219 | }, | |
220 | [ C(OP_WRITE) ] = { | |
221 | [ C(RESULT_ACCESS) ] = -1, | |
222 | [ C(RESULT_MISS) ] = -1, | |
223 | }, | |
224 | [ C(OP_PREFETCH) ] = { | |
225 | [ C(RESULT_ACCESS) ] = -1, | |
226 | [ C(RESULT_MISS) ] = -1, | |
227 | }, | |
228 | }, | |
229 | }; | |
230 | ||
caaa8be3 | 231 | static __initconst const u64 westmere_hw_cache_event_ids |
f22f54f4 PZ |
232 | [PERF_COUNT_HW_CACHE_MAX] |
233 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
234 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
235 | { | |
236 | [ C(L1D) ] = { | |
237 | [ C(OP_READ) ] = { | |
238 | [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ | |
239 | [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ | |
240 | }, | |
241 | [ C(OP_WRITE) ] = { | |
242 | [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ | |
243 | [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ | |
244 | }, | |
245 | [ C(OP_PREFETCH) ] = { | |
246 | [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ | |
247 | [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ | |
248 | }, | |
249 | }, | |
250 | [ C(L1I ) ] = { | |
251 | [ C(OP_READ) ] = { | |
252 | [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ | |
253 | [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ | |
254 | }, | |
255 | [ C(OP_WRITE) ] = { | |
256 | [ C(RESULT_ACCESS) ] = -1, | |
257 | [ C(RESULT_MISS) ] = -1, | |
258 | }, | |
259 | [ C(OP_PREFETCH) ] = { | |
260 | [ C(RESULT_ACCESS) ] = 0x0, | |
261 | [ C(RESULT_MISS) ] = 0x0, | |
262 | }, | |
263 | }, | |
264 | [ C(LL ) ] = { | |
265 | [ C(OP_READ) ] = { | |
63b6a675 | 266 | /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ |
e994d7d2 | 267 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
268 | /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ |
269 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 | 270 | }, |
e994d7d2 AK |
271 | /* |
272 | * Use RFO, not WRITEBACK, because a write miss would typically occur | |
273 | * on RFO. | |
274 | */ | |
f22f54f4 | 275 | [ C(OP_WRITE) ] = { |
63b6a675 PZ |
276 | /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ |
277 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
278 | /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ | |
e994d7d2 | 279 | [ C(RESULT_MISS) ] = 0x01b7, |
f22f54f4 PZ |
280 | }, |
281 | [ C(OP_PREFETCH) ] = { | |
63b6a675 | 282 | /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ |
e994d7d2 | 283 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
284 | /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ |
285 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 PZ |
286 | }, |
287 | }, | |
288 | [ C(DTLB) ] = { | |
289 | [ C(OP_READ) ] = { | |
290 | [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ | |
291 | [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ | |
292 | }, | |
293 | [ C(OP_WRITE) ] = { | |
294 | [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ | |
295 | [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ | |
296 | }, | |
297 | [ C(OP_PREFETCH) ] = { | |
298 | [ C(RESULT_ACCESS) ] = 0x0, | |
299 | [ C(RESULT_MISS) ] = 0x0, | |
300 | }, | |
301 | }, | |
302 | [ C(ITLB) ] = { | |
303 | [ C(OP_READ) ] = { | |
304 | [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ | |
305 | [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */ | |
306 | }, | |
307 | [ C(OP_WRITE) ] = { | |
308 | [ C(RESULT_ACCESS) ] = -1, | |
309 | [ C(RESULT_MISS) ] = -1, | |
310 | }, | |
311 | [ C(OP_PREFETCH) ] = { | |
312 | [ C(RESULT_ACCESS) ] = -1, | |
313 | [ C(RESULT_MISS) ] = -1, | |
314 | }, | |
315 | }, | |
316 | [ C(BPU ) ] = { | |
317 | [ C(OP_READ) ] = { | |
318 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
319 | [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ | |
320 | }, | |
321 | [ C(OP_WRITE) ] = { | |
322 | [ C(RESULT_ACCESS) ] = -1, | |
323 | [ C(RESULT_MISS) ] = -1, | |
324 | }, | |
325 | [ C(OP_PREFETCH) ] = { | |
326 | [ C(RESULT_ACCESS) ] = -1, | |
327 | [ C(RESULT_MISS) ] = -1, | |
328 | }, | |
329 | }, | |
330 | }; | |
331 | ||
e994d7d2 | 332 | /* |
63b6a675 PZ |
333 | * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits; |
334 | * See IA32 SDM Vol 3B 30.6.1.3 | |
e994d7d2 AK |
335 | */ |
336 | ||
63b6a675 PZ |
337 | #define NHM_DMND_DATA_RD (1 << 0) |
338 | #define NHM_DMND_RFO (1 << 1) | |
339 | #define NHM_DMND_IFETCH (1 << 2) | |
340 | #define NHM_DMND_WB (1 << 3) | |
341 | #define NHM_PF_DATA_RD (1 << 4) | |
342 | #define NHM_PF_DATA_RFO (1 << 5) | |
343 | #define NHM_PF_IFETCH (1 << 6) | |
344 | #define NHM_OFFCORE_OTHER (1 << 7) | |
345 | #define NHM_UNCORE_HIT (1 << 8) | |
346 | #define NHM_OTHER_CORE_HIT_SNP (1 << 9) | |
347 | #define NHM_OTHER_CORE_HITM (1 << 10) | |
348 | /* reserved */ | |
349 | #define NHM_REMOTE_CACHE_FWD (1 << 12) | |
350 | #define NHM_REMOTE_DRAM (1 << 13) | |
351 | #define NHM_LOCAL_DRAM (1 << 14) | |
352 | #define NHM_NON_DRAM (1 << 15) | |
353 | ||
354 | #define NHM_ALL_DRAM (NHM_REMOTE_DRAM|NHM_LOCAL_DRAM) | |
355 | ||
356 | #define NHM_DMND_READ (NHM_DMND_DATA_RD) | |
357 | #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB) | |
358 | #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO) | |
359 | ||
360 | #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM) | |
361 | #define NHM_L3_MISS (NHM_NON_DRAM|NHM_ALL_DRAM|NHM_REMOTE_CACHE_FWD) | |
362 | #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS) | |
e994d7d2 AK |
363 | |
364 | static __initconst const u64 nehalem_hw_cache_extra_regs | |
365 | [PERF_COUNT_HW_CACHE_MAX] | |
366 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
367 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
368 | { | |
369 | [ C(LL ) ] = { | |
370 | [ C(OP_READ) ] = { | |
63b6a675 PZ |
371 | [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS, |
372 | [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS, | |
e994d7d2 AK |
373 | }, |
374 | [ C(OP_WRITE) ] = { | |
63b6a675 PZ |
375 | [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS, |
376 | [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS, | |
e994d7d2 AK |
377 | }, |
378 | [ C(OP_PREFETCH) ] = { | |
63b6a675 PZ |
379 | [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS, |
380 | [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS, | |
e994d7d2 AK |
381 | }, |
382 | } | |
383 | }; | |
384 | ||
caaa8be3 | 385 | static __initconst const u64 nehalem_hw_cache_event_ids |
f22f54f4 PZ |
386 | [PERF_COUNT_HW_CACHE_MAX] |
387 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
388 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
389 | { | |
390 | [ C(L1D) ] = { | |
391 | [ C(OP_READ) ] = { | |
f4929bd3 PZ |
392 | [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ |
393 | [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ | |
f22f54f4 PZ |
394 | }, |
395 | [ C(OP_WRITE) ] = { | |
f4929bd3 PZ |
396 | [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ |
397 | [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ | |
f22f54f4 PZ |
398 | }, |
399 | [ C(OP_PREFETCH) ] = { | |
400 | [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ | |
401 | [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ | |
402 | }, | |
403 | }, | |
404 | [ C(L1I ) ] = { | |
405 | [ C(OP_READ) ] = { | |
406 | [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ | |
407 | [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ | |
408 | }, | |
409 | [ C(OP_WRITE) ] = { | |
410 | [ C(RESULT_ACCESS) ] = -1, | |
411 | [ C(RESULT_MISS) ] = -1, | |
412 | }, | |
413 | [ C(OP_PREFETCH) ] = { | |
414 | [ C(RESULT_ACCESS) ] = 0x0, | |
415 | [ C(RESULT_MISS) ] = 0x0, | |
416 | }, | |
417 | }, | |
418 | [ C(LL ) ] = { | |
419 | [ C(OP_READ) ] = { | |
e994d7d2 AK |
420 | /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ |
421 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
422 | /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ | |
423 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 | 424 | }, |
e994d7d2 AK |
425 | /* |
426 | * Use RFO, not WRITEBACK, because a write miss would typically occur | |
427 | * on RFO. | |
428 | */ | |
f22f54f4 | 429 | [ C(OP_WRITE) ] = { |
e994d7d2 AK |
430 | /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ |
431 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
432 | /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ | |
433 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 PZ |
434 | }, |
435 | [ C(OP_PREFETCH) ] = { | |
e994d7d2 AK |
436 | /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ |
437 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
438 | /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ | |
439 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 PZ |
440 | }, |
441 | }, | |
442 | [ C(DTLB) ] = { | |
443 | [ C(OP_READ) ] = { | |
444 | [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ | |
445 | [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ | |
446 | }, | |
447 | [ C(OP_WRITE) ] = { | |
448 | [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ | |
449 | [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ | |
450 | }, | |
451 | [ C(OP_PREFETCH) ] = { | |
452 | [ C(RESULT_ACCESS) ] = 0x0, | |
453 | [ C(RESULT_MISS) ] = 0x0, | |
454 | }, | |
455 | }, | |
456 | [ C(ITLB) ] = { | |
457 | [ C(OP_READ) ] = { | |
458 | [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ | |
459 | [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */ | |
460 | }, | |
461 | [ C(OP_WRITE) ] = { | |
462 | [ C(RESULT_ACCESS) ] = -1, | |
463 | [ C(RESULT_MISS) ] = -1, | |
464 | }, | |
465 | [ C(OP_PREFETCH) ] = { | |
466 | [ C(RESULT_ACCESS) ] = -1, | |
467 | [ C(RESULT_MISS) ] = -1, | |
468 | }, | |
469 | }, | |
470 | [ C(BPU ) ] = { | |
471 | [ C(OP_READ) ] = { | |
472 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
473 | [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ | |
474 | }, | |
475 | [ C(OP_WRITE) ] = { | |
476 | [ C(RESULT_ACCESS) ] = -1, | |
477 | [ C(RESULT_MISS) ] = -1, | |
478 | }, | |
479 | [ C(OP_PREFETCH) ] = { | |
480 | [ C(RESULT_ACCESS) ] = -1, | |
481 | [ C(RESULT_MISS) ] = -1, | |
482 | }, | |
483 | }, | |
484 | }; | |
485 | ||
caaa8be3 | 486 | static __initconst const u64 core2_hw_cache_event_ids |
f22f54f4 PZ |
487 | [PERF_COUNT_HW_CACHE_MAX] |
488 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
489 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
490 | { | |
491 | [ C(L1D) ] = { | |
492 | [ C(OP_READ) ] = { | |
493 | [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */ | |
494 | [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */ | |
495 | }, | |
496 | [ C(OP_WRITE) ] = { | |
497 | [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */ | |
498 | [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */ | |
499 | }, | |
500 | [ C(OP_PREFETCH) ] = { | |
501 | [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */ | |
502 | [ C(RESULT_MISS) ] = 0, | |
503 | }, | |
504 | }, | |
505 | [ C(L1I ) ] = { | |
506 | [ C(OP_READ) ] = { | |
507 | [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */ | |
508 | [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */ | |
509 | }, | |
510 | [ C(OP_WRITE) ] = { | |
511 | [ C(RESULT_ACCESS) ] = -1, | |
512 | [ C(RESULT_MISS) ] = -1, | |
513 | }, | |
514 | [ C(OP_PREFETCH) ] = { | |
515 | [ C(RESULT_ACCESS) ] = 0, | |
516 | [ C(RESULT_MISS) ] = 0, | |
517 | }, | |
518 | }, | |
519 | [ C(LL ) ] = { | |
520 | [ C(OP_READ) ] = { | |
521 | [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ | |
522 | [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ | |
523 | }, | |
524 | [ C(OP_WRITE) ] = { | |
525 | [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ | |
526 | [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ | |
527 | }, | |
528 | [ C(OP_PREFETCH) ] = { | |
529 | [ C(RESULT_ACCESS) ] = 0, | |
530 | [ C(RESULT_MISS) ] = 0, | |
531 | }, | |
532 | }, | |
533 | [ C(DTLB) ] = { | |
534 | [ C(OP_READ) ] = { | |
535 | [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ | |
536 | [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */ | |
537 | }, | |
538 | [ C(OP_WRITE) ] = { | |
539 | [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ | |
540 | [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */ | |
541 | }, | |
542 | [ C(OP_PREFETCH) ] = { | |
543 | [ C(RESULT_ACCESS) ] = 0, | |
544 | [ C(RESULT_MISS) ] = 0, | |
545 | }, | |
546 | }, | |
547 | [ C(ITLB) ] = { | |
548 | [ C(OP_READ) ] = { | |
549 | [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ | |
550 | [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */ | |
551 | }, | |
552 | [ C(OP_WRITE) ] = { | |
553 | [ C(RESULT_ACCESS) ] = -1, | |
554 | [ C(RESULT_MISS) ] = -1, | |
555 | }, | |
556 | [ C(OP_PREFETCH) ] = { | |
557 | [ C(RESULT_ACCESS) ] = -1, | |
558 | [ C(RESULT_MISS) ] = -1, | |
559 | }, | |
560 | }, | |
561 | [ C(BPU ) ] = { | |
562 | [ C(OP_READ) ] = { | |
563 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ | |
564 | [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ | |
565 | }, | |
566 | [ C(OP_WRITE) ] = { | |
567 | [ C(RESULT_ACCESS) ] = -1, | |
568 | [ C(RESULT_MISS) ] = -1, | |
569 | }, | |
570 | [ C(OP_PREFETCH) ] = { | |
571 | [ C(RESULT_ACCESS) ] = -1, | |
572 | [ C(RESULT_MISS) ] = -1, | |
573 | }, | |
574 | }, | |
575 | }; | |
576 | ||
caaa8be3 | 577 | static __initconst const u64 atom_hw_cache_event_ids |
f22f54f4 PZ |
578 | [PERF_COUNT_HW_CACHE_MAX] |
579 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
580 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
581 | { | |
582 | [ C(L1D) ] = { | |
583 | [ C(OP_READ) ] = { | |
584 | [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */ | |
585 | [ C(RESULT_MISS) ] = 0, | |
586 | }, | |
587 | [ C(OP_WRITE) ] = { | |
588 | [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */ | |
589 | [ C(RESULT_MISS) ] = 0, | |
590 | }, | |
591 | [ C(OP_PREFETCH) ] = { | |
592 | [ C(RESULT_ACCESS) ] = 0x0, | |
593 | [ C(RESULT_MISS) ] = 0, | |
594 | }, | |
595 | }, | |
596 | [ C(L1I ) ] = { | |
597 | [ C(OP_READ) ] = { | |
598 | [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ | |
599 | [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ | |
600 | }, | |
601 | [ C(OP_WRITE) ] = { | |
602 | [ C(RESULT_ACCESS) ] = -1, | |
603 | [ C(RESULT_MISS) ] = -1, | |
604 | }, | |
605 | [ C(OP_PREFETCH) ] = { | |
606 | [ C(RESULT_ACCESS) ] = 0, | |
607 | [ C(RESULT_MISS) ] = 0, | |
608 | }, | |
609 | }, | |
610 | [ C(LL ) ] = { | |
611 | [ C(OP_READ) ] = { | |
612 | [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ | |
613 | [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ | |
614 | }, | |
615 | [ C(OP_WRITE) ] = { | |
616 | [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ | |
617 | [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ | |
618 | }, | |
619 | [ C(OP_PREFETCH) ] = { | |
620 | [ C(RESULT_ACCESS) ] = 0, | |
621 | [ C(RESULT_MISS) ] = 0, | |
622 | }, | |
623 | }, | |
624 | [ C(DTLB) ] = { | |
625 | [ C(OP_READ) ] = { | |
626 | [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */ | |
627 | [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */ | |
628 | }, | |
629 | [ C(OP_WRITE) ] = { | |
630 | [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */ | |
631 | [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */ | |
632 | }, | |
633 | [ C(OP_PREFETCH) ] = { | |
634 | [ C(RESULT_ACCESS) ] = 0, | |
635 | [ C(RESULT_MISS) ] = 0, | |
636 | }, | |
637 | }, | |
638 | [ C(ITLB) ] = { | |
639 | [ C(OP_READ) ] = { | |
640 | [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ | |
641 | [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */ | |
642 | }, | |
643 | [ C(OP_WRITE) ] = { | |
644 | [ C(RESULT_ACCESS) ] = -1, | |
645 | [ C(RESULT_MISS) ] = -1, | |
646 | }, | |
647 | [ C(OP_PREFETCH) ] = { | |
648 | [ C(RESULT_ACCESS) ] = -1, | |
649 | [ C(RESULT_MISS) ] = -1, | |
650 | }, | |
651 | }, | |
652 | [ C(BPU ) ] = { | |
653 | [ C(OP_READ) ] = { | |
654 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ | |
655 | [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ | |
656 | }, | |
657 | [ C(OP_WRITE) ] = { | |
658 | [ C(RESULT_ACCESS) ] = -1, | |
659 | [ C(RESULT_MISS) ] = -1, | |
660 | }, | |
661 | [ C(OP_PREFETCH) ] = { | |
662 | [ C(RESULT_ACCESS) ] = -1, | |
663 | [ C(RESULT_MISS) ] = -1, | |
664 | }, | |
665 | }, | |
666 | }; | |
667 | ||
f22f54f4 PZ |
668 | static void intel_pmu_disable_all(void) |
669 | { | |
670 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | |
671 | ||
672 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); | |
673 | ||
674 | if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask)) | |
675 | intel_pmu_disable_bts(); | |
ca037701 PZ |
676 | |
677 | intel_pmu_pebs_disable_all(); | |
caff2bef | 678 | intel_pmu_lbr_disable_all(); |
f22f54f4 PZ |
679 | } |
680 | ||
11164cd4 | 681 | static void intel_pmu_enable_all(int added) |
f22f54f4 PZ |
682 | { |
683 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | |
684 | ||
d329527e PZ |
685 | intel_pmu_pebs_enable_all(); |
686 | intel_pmu_lbr_enable_all(); | |
f22f54f4 PZ |
687 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl); |
688 | ||
689 | if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask)) { | |
690 | struct perf_event *event = | |
691 | cpuc->events[X86_PMC_IDX_FIXED_BTS]; | |
692 | ||
693 | if (WARN_ON_ONCE(!event)) | |
694 | return; | |
695 | ||
696 | intel_pmu_enable_bts(event->hw.config); | |
697 | } | |
698 | } | |
699 | ||
11164cd4 PZ |
700 | /* |
701 | * Workaround for: | |
702 | * Intel Errata AAK100 (model 26) | |
703 | * Intel Errata AAP53 (model 30) | |
40b91cd1 | 704 | * Intel Errata BD53 (model 44) |
11164cd4 | 705 | * |
351af072 ZY |
706 | * The official story: |
707 | * These chips need to be 'reset' when adding counters by programming the | |
708 | * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either | |
709 | * in sequence on the same PMC or on different PMCs. | |
710 | * | |
711 | * In practise it appears some of these events do in fact count, and | |
712 | * we need to programm all 4 events. | |
11164cd4 | 713 | */ |
351af072 | 714 | static void intel_pmu_nhm_workaround(void) |
11164cd4 | 715 | { |
351af072 ZY |
716 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); |
717 | static const unsigned long nhm_magic[4] = { | |
718 | 0x4300B5, | |
719 | 0x4300D2, | |
720 | 0x4300B1, | |
721 | 0x4300B1 | |
722 | }; | |
723 | struct perf_event *event; | |
724 | int i; | |
11164cd4 | 725 | |
351af072 ZY |
726 | /* |
727 | * The Errata requires below steps: | |
728 | * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL; | |
729 | * 2) Configure 4 PERFEVTSELx with the magic events and clear | |
730 | * the corresponding PMCx; | |
731 | * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL; | |
732 | * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL; | |
733 | * 5) Clear 4 pairs of ERFEVTSELx and PMCx; | |
734 | */ | |
11164cd4 | 735 | |
351af072 ZY |
736 | /* |
737 | * The real steps we choose are a little different from above. | |
738 | * A) To reduce MSR operations, we don't run step 1) as they | |
739 | * are already cleared before this function is called; | |
740 | * B) Call x86_perf_event_update to save PMCx before configuring | |
741 | * PERFEVTSELx with magic number; | |
742 | * C) With step 5), we do clear only when the PERFEVTSELx is | |
743 | * not used currently. | |
744 | * D) Call x86_perf_event_set_period to restore PMCx; | |
745 | */ | |
11164cd4 | 746 | |
351af072 ZY |
747 | /* We always operate 4 pairs of PERF Counters */ |
748 | for (i = 0; i < 4; i++) { | |
749 | event = cpuc->events[i]; | |
750 | if (event) | |
751 | x86_perf_event_update(event); | |
752 | } | |
11164cd4 | 753 | |
351af072 ZY |
754 | for (i = 0; i < 4; i++) { |
755 | wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]); | |
756 | wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0); | |
757 | } | |
758 | ||
759 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf); | |
760 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0); | |
11164cd4 | 761 | |
351af072 ZY |
762 | for (i = 0; i < 4; i++) { |
763 | event = cpuc->events[i]; | |
764 | ||
765 | if (event) { | |
766 | x86_perf_event_set_period(event); | |
31fa58af | 767 | __x86_pmu_enable_event(&event->hw, |
351af072 ZY |
768 | ARCH_PERFMON_EVENTSEL_ENABLE); |
769 | } else | |
770 | wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0); | |
11164cd4 | 771 | } |
351af072 ZY |
772 | } |
773 | ||
774 | static void intel_pmu_nhm_enable_all(int added) | |
775 | { | |
776 | if (added) | |
777 | intel_pmu_nhm_workaround(); | |
11164cd4 PZ |
778 | intel_pmu_enable_all(added); |
779 | } | |
780 | ||
f22f54f4 PZ |
781 | static inline u64 intel_pmu_get_status(void) |
782 | { | |
783 | u64 status; | |
784 | ||
785 | rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status); | |
786 | ||
787 | return status; | |
788 | } | |
789 | ||
790 | static inline void intel_pmu_ack_status(u64 ack) | |
791 | { | |
792 | wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack); | |
793 | } | |
794 | ||
ca037701 | 795 | static void intel_pmu_disable_fixed(struct hw_perf_event *hwc) |
f22f54f4 | 796 | { |
aff3d91a | 797 | int idx = hwc->idx - X86_PMC_IDX_FIXED; |
f22f54f4 PZ |
798 | u64 ctrl_val, mask; |
799 | ||
800 | mask = 0xfULL << (idx * 4); | |
801 | ||
802 | rdmsrl(hwc->config_base, ctrl_val); | |
803 | ctrl_val &= ~mask; | |
7645a24c | 804 | wrmsrl(hwc->config_base, ctrl_val); |
f22f54f4 PZ |
805 | } |
806 | ||
ca037701 | 807 | static void intel_pmu_disable_event(struct perf_event *event) |
f22f54f4 | 808 | { |
aff3d91a PZ |
809 | struct hw_perf_event *hwc = &event->hw; |
810 | ||
811 | if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) { | |
f22f54f4 PZ |
812 | intel_pmu_disable_bts(); |
813 | intel_pmu_drain_bts_buffer(); | |
814 | return; | |
815 | } | |
816 | ||
817 | if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { | |
aff3d91a | 818 | intel_pmu_disable_fixed(hwc); |
f22f54f4 PZ |
819 | return; |
820 | } | |
821 | ||
aff3d91a | 822 | x86_pmu_disable_event(event); |
ca037701 | 823 | |
ab608344 | 824 | if (unlikely(event->attr.precise_ip)) |
ef21f683 | 825 | intel_pmu_pebs_disable(event); |
f22f54f4 PZ |
826 | } |
827 | ||
ca037701 | 828 | static void intel_pmu_enable_fixed(struct hw_perf_event *hwc) |
f22f54f4 | 829 | { |
aff3d91a | 830 | int idx = hwc->idx - X86_PMC_IDX_FIXED; |
f22f54f4 | 831 | u64 ctrl_val, bits, mask; |
f22f54f4 PZ |
832 | |
833 | /* | |
834 | * Enable IRQ generation (0x8), | |
835 | * and enable ring-3 counting (0x2) and ring-0 counting (0x1) | |
836 | * if requested: | |
837 | */ | |
838 | bits = 0x8ULL; | |
839 | if (hwc->config & ARCH_PERFMON_EVENTSEL_USR) | |
840 | bits |= 0x2; | |
841 | if (hwc->config & ARCH_PERFMON_EVENTSEL_OS) | |
842 | bits |= 0x1; | |
843 | ||
844 | /* | |
845 | * ANY bit is supported in v3 and up | |
846 | */ | |
847 | if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY) | |
848 | bits |= 0x4; | |
849 | ||
850 | bits <<= (idx * 4); | |
851 | mask = 0xfULL << (idx * 4); | |
852 | ||
853 | rdmsrl(hwc->config_base, ctrl_val); | |
854 | ctrl_val &= ~mask; | |
855 | ctrl_val |= bits; | |
7645a24c | 856 | wrmsrl(hwc->config_base, ctrl_val); |
f22f54f4 PZ |
857 | } |
858 | ||
aff3d91a | 859 | static void intel_pmu_enable_event(struct perf_event *event) |
f22f54f4 | 860 | { |
aff3d91a PZ |
861 | struct hw_perf_event *hwc = &event->hw; |
862 | ||
863 | if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) { | |
0a3aee0d | 864 | if (!__this_cpu_read(cpu_hw_events.enabled)) |
f22f54f4 PZ |
865 | return; |
866 | ||
867 | intel_pmu_enable_bts(hwc->config); | |
868 | return; | |
869 | } | |
870 | ||
871 | if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { | |
aff3d91a | 872 | intel_pmu_enable_fixed(hwc); |
f22f54f4 PZ |
873 | return; |
874 | } | |
875 | ||
ab608344 | 876 | if (unlikely(event->attr.precise_ip)) |
ef21f683 | 877 | intel_pmu_pebs_enable(event); |
ca037701 | 878 | |
31fa58af | 879 | __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); |
f22f54f4 PZ |
880 | } |
881 | ||
882 | /* | |
883 | * Save and restart an expired event. Called by NMI contexts, | |
884 | * so it has to be careful about preempting normal event ops: | |
885 | */ | |
886 | static int intel_pmu_save_and_restart(struct perf_event *event) | |
887 | { | |
cc2ad4ba PZ |
888 | x86_perf_event_update(event); |
889 | return x86_perf_event_set_period(event); | |
f22f54f4 PZ |
890 | } |
891 | ||
892 | static void intel_pmu_reset(void) | |
893 | { | |
0a3aee0d | 894 | struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds); |
f22f54f4 PZ |
895 | unsigned long flags; |
896 | int idx; | |
897 | ||
948b1bb8 | 898 | if (!x86_pmu.num_counters) |
f22f54f4 PZ |
899 | return; |
900 | ||
901 | local_irq_save(flags); | |
902 | ||
903 | printk("clearing PMU state on CPU#%d\n", smp_processor_id()); | |
904 | ||
948b1bb8 | 905 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
41bf4989 RR |
906 | checking_wrmsrl(x86_pmu_config_addr(idx), 0ull); |
907 | checking_wrmsrl(x86_pmu_event_addr(idx), 0ull); | |
f22f54f4 | 908 | } |
948b1bb8 | 909 | for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) |
f22f54f4 | 910 | checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull); |
948b1bb8 | 911 | |
f22f54f4 PZ |
912 | if (ds) |
913 | ds->bts_index = ds->bts_buffer_base; | |
914 | ||
915 | local_irq_restore(flags); | |
916 | } | |
917 | ||
918 | /* | |
919 | * This handler is triggered by the local APIC, so the APIC IRQ handling | |
920 | * rules apply: | |
921 | */ | |
922 | static int intel_pmu_handle_irq(struct pt_regs *regs) | |
923 | { | |
924 | struct perf_sample_data data; | |
925 | struct cpu_hw_events *cpuc; | |
926 | int bit, loops; | |
2e556b5b | 927 | u64 status; |
b0b2072d | 928 | int handled; |
f22f54f4 | 929 | |
dc1d628a | 930 | perf_sample_data_init(&data, 0); |
f22f54f4 PZ |
931 | |
932 | cpuc = &__get_cpu_var(cpu_hw_events); | |
933 | ||
2bce5dac DZ |
934 | /* |
935 | * Some chipsets need to unmask the LVTPC in a particular spot | |
936 | * inside the nmi handler. As a result, the unmasking was pushed | |
937 | * into all the nmi handlers. | |
938 | * | |
939 | * This handler doesn't seem to have any issues with the unmasking | |
940 | * so it was left at the top. | |
941 | */ | |
942 | apic_write(APIC_LVTPC, APIC_DM_NMI); | |
943 | ||
3fb2b8dd | 944 | intel_pmu_disable_all(); |
b0b2072d | 945 | handled = intel_pmu_drain_bts_buffer(); |
f22f54f4 PZ |
946 | status = intel_pmu_get_status(); |
947 | if (!status) { | |
11164cd4 | 948 | intel_pmu_enable_all(0); |
b0b2072d | 949 | return handled; |
f22f54f4 PZ |
950 | } |
951 | ||
952 | loops = 0; | |
953 | again: | |
2e556b5b | 954 | intel_pmu_ack_status(status); |
f22f54f4 PZ |
955 | if (++loops > 100) { |
956 | WARN_ONCE(1, "perfevents: irq loop stuck!\n"); | |
957 | perf_event_print_debug(); | |
958 | intel_pmu_reset(); | |
3fb2b8dd | 959 | goto done; |
f22f54f4 PZ |
960 | } |
961 | ||
962 | inc_irq_stat(apic_perf_irqs); | |
ca037701 | 963 | |
caff2bef PZ |
964 | intel_pmu_lbr_read(); |
965 | ||
ca037701 PZ |
966 | /* |
967 | * PEBS overflow sets bit 62 in the global status register | |
968 | */ | |
de725dec PZ |
969 | if (__test_and_clear_bit(62, (unsigned long *)&status)) { |
970 | handled++; | |
ca037701 | 971 | x86_pmu.drain_pebs(regs); |
de725dec | 972 | } |
ca037701 | 973 | |
984b3f57 | 974 | for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) { |
f22f54f4 PZ |
975 | struct perf_event *event = cpuc->events[bit]; |
976 | ||
de725dec PZ |
977 | handled++; |
978 | ||
f22f54f4 PZ |
979 | if (!test_bit(bit, cpuc->active_mask)) |
980 | continue; | |
981 | ||
982 | if (!intel_pmu_save_and_restart(event)) | |
983 | continue; | |
984 | ||
985 | data.period = event->hw.last_period; | |
986 | ||
a8b0ca17 | 987 | if (perf_event_overflow(event, &data, regs)) |
a4eaf7f1 | 988 | x86_pmu_stop(event, 0); |
f22f54f4 PZ |
989 | } |
990 | ||
f22f54f4 PZ |
991 | /* |
992 | * Repeat if there is more work to be done: | |
993 | */ | |
994 | status = intel_pmu_get_status(); | |
995 | if (status) | |
996 | goto again; | |
997 | ||
3fb2b8dd | 998 | done: |
11164cd4 | 999 | intel_pmu_enable_all(0); |
de725dec | 1000 | return handled; |
f22f54f4 PZ |
1001 | } |
1002 | ||
f22f54f4 | 1003 | static struct event_constraint * |
ca037701 | 1004 | intel_bts_constraints(struct perf_event *event) |
f22f54f4 | 1005 | { |
ca037701 PZ |
1006 | struct hw_perf_event *hwc = &event->hw; |
1007 | unsigned int hw_event, bts_event; | |
f22f54f4 | 1008 | |
18a073a3 PZ |
1009 | if (event->attr.freq) |
1010 | return NULL; | |
1011 | ||
ca037701 PZ |
1012 | hw_event = hwc->config & INTEL_ARCH_EVENT_MASK; |
1013 | bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS); | |
f22f54f4 | 1014 | |
ca037701 | 1015 | if (unlikely(hw_event == bts_event && hwc->sample_period == 1)) |
f22f54f4 | 1016 | return &bts_constraint; |
ca037701 | 1017 | |
f22f54f4 PZ |
1018 | return NULL; |
1019 | } | |
1020 | ||
b79e8941 PZ |
1021 | static bool intel_try_alt_er(struct perf_event *event, int orig_idx) |
1022 | { | |
1023 | if (!(x86_pmu.er_flags & ERF_HAS_RSP_1)) | |
1024 | return false; | |
1025 | ||
1026 | if (event->hw.extra_reg.idx == EXTRA_REG_RSP_0) { | |
1027 | event->hw.config &= ~INTEL_ARCH_EVENT_MASK; | |
1028 | event->hw.config |= 0x01bb; | |
1029 | event->hw.extra_reg.idx = EXTRA_REG_RSP_1; | |
1030 | event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1; | |
1031 | } else if (event->hw.extra_reg.idx == EXTRA_REG_RSP_1) { | |
1032 | event->hw.config &= ~INTEL_ARCH_EVENT_MASK; | |
1033 | event->hw.config |= 0x01b7; | |
1034 | event->hw.extra_reg.idx = EXTRA_REG_RSP_0; | |
1035 | event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0; | |
1036 | } | |
1037 | ||
1038 | if (event->hw.extra_reg.idx == orig_idx) | |
1039 | return false; | |
1040 | ||
1041 | return true; | |
1042 | } | |
1043 | ||
efc9f05d SE |
1044 | /* |
1045 | * manage allocation of shared extra msr for certain events | |
1046 | * | |
1047 | * sharing can be: | |
1048 | * per-cpu: to be shared between the various events on a single PMU | |
1049 | * per-core: per-cpu + shared by HT threads | |
1050 | */ | |
a7e3ed1e | 1051 | static struct event_constraint * |
efc9f05d | 1052 | __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc, |
b79e8941 | 1053 | struct perf_event *event) |
a7e3ed1e | 1054 | { |
efc9f05d | 1055 | struct event_constraint *c = &emptyconstraint; |
b79e8941 | 1056 | struct hw_perf_event_extra *reg = &event->hw.extra_reg; |
a7e3ed1e | 1057 | struct er_account *era; |
cd8a38d3 | 1058 | unsigned long flags; |
b79e8941 | 1059 | int orig_idx = reg->idx; |
a7e3ed1e | 1060 | |
efc9f05d | 1061 | /* already allocated shared msr */ |
cd8a38d3 | 1062 | if (reg->alloc) |
efc9f05d | 1063 | return &unconstrained; |
a7e3ed1e | 1064 | |
b79e8941 | 1065 | again: |
efc9f05d | 1066 | era = &cpuc->shared_regs->regs[reg->idx]; |
cd8a38d3 SE |
1067 | /* |
1068 | * we use spin_lock_irqsave() to avoid lockdep issues when | |
1069 | * passing a fake cpuc | |
1070 | */ | |
1071 | raw_spin_lock_irqsave(&era->lock, flags); | |
efc9f05d SE |
1072 | |
1073 | if (!atomic_read(&era->ref) || era->config == reg->config) { | |
1074 | ||
1075 | /* lock in msr value */ | |
1076 | era->config = reg->config; | |
1077 | era->reg = reg->reg; | |
1078 | ||
1079 | /* one more user */ | |
1080 | atomic_inc(&era->ref); | |
1081 | ||
1082 | /* no need to reallocate during incremental event scheduling */ | |
1083 | reg->alloc = 1; | |
a7e3ed1e AK |
1084 | |
1085 | /* | |
efc9f05d SE |
1086 | * All events using extra_reg are unconstrained. |
1087 | * Avoids calling x86_get_event_constraints() | |
1088 | * | |
1089 | * Must revisit if extra_reg controlling events | |
1090 | * ever have constraints. Worst case we go through | |
1091 | * the regular event constraint table. | |
a7e3ed1e | 1092 | */ |
efc9f05d | 1093 | c = &unconstrained; |
b79e8941 PZ |
1094 | } else if (intel_try_alt_er(event, orig_idx)) { |
1095 | raw_spin_unlock(&era->lock); | |
1096 | goto again; | |
a7e3ed1e | 1097 | } |
cd8a38d3 | 1098 | raw_spin_unlock_irqrestore(&era->lock, flags); |
a7e3ed1e | 1099 | |
efc9f05d SE |
1100 | return c; |
1101 | } | |
1102 | ||
1103 | static void | |
1104 | __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc, | |
1105 | struct hw_perf_event_extra *reg) | |
1106 | { | |
1107 | struct er_account *era; | |
1108 | ||
1109 | /* | |
1110 | * only put constraint if extra reg was actually | |
1111 | * allocated. Also takes care of event which do | |
1112 | * not use an extra shared reg | |
1113 | */ | |
1114 | if (!reg->alloc) | |
1115 | return; | |
1116 | ||
1117 | era = &cpuc->shared_regs->regs[reg->idx]; | |
1118 | ||
1119 | /* one fewer user */ | |
1120 | atomic_dec(&era->ref); | |
1121 | ||
1122 | /* allocate again next time */ | |
1123 | reg->alloc = 0; | |
1124 | } | |
1125 | ||
1126 | static struct event_constraint * | |
1127 | intel_shared_regs_constraints(struct cpu_hw_events *cpuc, | |
1128 | struct perf_event *event) | |
1129 | { | |
1130 | struct event_constraint *c = NULL; | |
efc9f05d | 1131 | |
b79e8941 PZ |
1132 | if (event->hw.extra_reg.idx != EXTRA_REG_NONE) |
1133 | c = __intel_shared_reg_get_constraints(cpuc, event); | |
1134 | ||
efc9f05d | 1135 | return c; |
a7e3ed1e AK |
1136 | } |
1137 | ||
f22f54f4 PZ |
1138 | static struct event_constraint * |
1139 | intel_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event) | |
1140 | { | |
1141 | struct event_constraint *c; | |
1142 | ||
ca037701 PZ |
1143 | c = intel_bts_constraints(event); |
1144 | if (c) | |
1145 | return c; | |
1146 | ||
1147 | c = intel_pebs_constraints(event); | |
f22f54f4 PZ |
1148 | if (c) |
1149 | return c; | |
1150 | ||
efc9f05d | 1151 | c = intel_shared_regs_constraints(cpuc, event); |
a7e3ed1e AK |
1152 | if (c) |
1153 | return c; | |
1154 | ||
f22f54f4 PZ |
1155 | return x86_get_event_constraints(cpuc, event); |
1156 | } | |
1157 | ||
efc9f05d SE |
1158 | static void |
1159 | intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc, | |
a7e3ed1e AK |
1160 | struct perf_event *event) |
1161 | { | |
efc9f05d | 1162 | struct hw_perf_event_extra *reg; |
a7e3ed1e | 1163 | |
efc9f05d SE |
1164 | reg = &event->hw.extra_reg; |
1165 | if (reg->idx != EXTRA_REG_NONE) | |
1166 | __intel_shared_reg_put_constraints(cpuc, reg); | |
1167 | } | |
a7e3ed1e | 1168 | |
efc9f05d SE |
1169 | static void intel_put_event_constraints(struct cpu_hw_events *cpuc, |
1170 | struct perf_event *event) | |
1171 | { | |
1172 | intel_put_shared_regs_event_constraints(cpuc, event); | |
a7e3ed1e AK |
1173 | } |
1174 | ||
b4cdc5c2 PZ |
1175 | static int intel_pmu_hw_config(struct perf_event *event) |
1176 | { | |
1177 | int ret = x86_pmu_hw_config(event); | |
1178 | ||
1179 | if (ret) | |
1180 | return ret; | |
1181 | ||
7639dae0 PZ |
1182 | if (event->attr.precise_ip && |
1183 | (event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { | |
1184 | /* | |
1185 | * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P | |
1186 | * (0x003c) so that we can use it with PEBS. | |
1187 | * | |
1188 | * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't | |
1189 | * PEBS capable. However we can use INST_RETIRED.ANY_P | |
1190 | * (0x00c0), which is a PEBS capable event, to get the same | |
1191 | * count. | |
1192 | * | |
1193 | * INST_RETIRED.ANY_P counts the number of cycles that retires | |
1194 | * CNTMASK instructions. By setting CNTMASK to a value (16) | |
1195 | * larger than the maximum number of instructions that can be | |
1196 | * retired per cycle (4) and then inverting the condition, we | |
1197 | * count all cycles that retire 16 or less instructions, which | |
1198 | * is every cycle. | |
1199 | * | |
1200 | * Thereby we gain a PEBS capable cycle counter. | |
1201 | */ | |
1202 | u64 alt_config = 0x108000c0; /* INST_RETIRED.TOTAL_CYCLES */ | |
1203 | ||
1204 | alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); | |
1205 | event->hw.config = alt_config; | |
1206 | } | |
1207 | ||
b4cdc5c2 PZ |
1208 | if (event->attr.type != PERF_TYPE_RAW) |
1209 | return 0; | |
1210 | ||
1211 | if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY)) | |
1212 | return 0; | |
1213 | ||
1214 | if (x86_pmu.version < 3) | |
1215 | return -EINVAL; | |
1216 | ||
1217 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) | |
1218 | return -EACCES; | |
1219 | ||
1220 | event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY; | |
1221 | ||
1222 | return 0; | |
1223 | } | |
1224 | ||
caaa8be3 | 1225 | static __initconst const struct x86_pmu core_pmu = { |
f22f54f4 PZ |
1226 | .name = "core", |
1227 | .handle_irq = x86_pmu_handle_irq, | |
1228 | .disable_all = x86_pmu_disable_all, | |
1229 | .enable_all = x86_pmu_enable_all, | |
1230 | .enable = x86_pmu_enable_event, | |
1231 | .disable = x86_pmu_disable_event, | |
b4cdc5c2 | 1232 | .hw_config = x86_pmu_hw_config, |
a072738e | 1233 | .schedule_events = x86_schedule_events, |
f22f54f4 PZ |
1234 | .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, |
1235 | .perfctr = MSR_ARCH_PERFMON_PERFCTR0, | |
1236 | .event_map = intel_pmu_event_map, | |
f22f54f4 PZ |
1237 | .max_events = ARRAY_SIZE(intel_perfmon_event_map), |
1238 | .apic = 1, | |
1239 | /* | |
1240 | * Intel PMCs cannot be accessed sanely above 32 bit width, | |
1241 | * so we install an artificial 1<<31 period regardless of | |
1242 | * the generic event period: | |
1243 | */ | |
1244 | .max_period = (1ULL << 31) - 1, | |
1245 | .get_event_constraints = intel_get_event_constraints, | |
a7e3ed1e | 1246 | .put_event_constraints = intel_put_event_constraints, |
f22f54f4 PZ |
1247 | .event_constraints = intel_core_event_constraints, |
1248 | }; | |
1249 | ||
efc9f05d SE |
1250 | static struct intel_shared_regs *allocate_shared_regs(int cpu) |
1251 | { | |
1252 | struct intel_shared_regs *regs; | |
1253 | int i; | |
1254 | ||
1255 | regs = kzalloc_node(sizeof(struct intel_shared_regs), | |
1256 | GFP_KERNEL, cpu_to_node(cpu)); | |
1257 | if (regs) { | |
1258 | /* | |
1259 | * initialize the locks to keep lockdep happy | |
1260 | */ | |
1261 | for (i = 0; i < EXTRA_REG_MAX; i++) | |
1262 | raw_spin_lock_init(®s->regs[i].lock); | |
1263 | ||
1264 | regs->core_id = -1; | |
1265 | } | |
1266 | return regs; | |
1267 | } | |
1268 | ||
a7e3ed1e AK |
1269 | static int intel_pmu_cpu_prepare(int cpu) |
1270 | { | |
1271 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); | |
1272 | ||
efc9f05d | 1273 | if (!x86_pmu.extra_regs) |
69092624 LM |
1274 | return NOTIFY_OK; |
1275 | ||
efc9f05d SE |
1276 | cpuc->shared_regs = allocate_shared_regs(cpu); |
1277 | if (!cpuc->shared_regs) | |
a7e3ed1e AK |
1278 | return NOTIFY_BAD; |
1279 | ||
a7e3ed1e AK |
1280 | return NOTIFY_OK; |
1281 | } | |
1282 | ||
74846d35 PZ |
1283 | static void intel_pmu_cpu_starting(int cpu) |
1284 | { | |
a7e3ed1e AK |
1285 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); |
1286 | int core_id = topology_core_id(cpu); | |
1287 | int i; | |
1288 | ||
69092624 LM |
1289 | init_debug_store_on_cpu(cpu); |
1290 | /* | |
1291 | * Deal with CPUs that don't clear their LBRs on power-up. | |
1292 | */ | |
1293 | intel_pmu_lbr_reset(); | |
1294 | ||
b79e8941 | 1295 | if (!cpuc->shared_regs || (x86_pmu.er_flags & ERF_NO_HT_SHARING)) |
69092624 LM |
1296 | return; |
1297 | ||
a7e3ed1e | 1298 | for_each_cpu(i, topology_thread_cpumask(cpu)) { |
efc9f05d | 1299 | struct intel_shared_regs *pc; |
a7e3ed1e | 1300 | |
efc9f05d | 1301 | pc = per_cpu(cpu_hw_events, i).shared_regs; |
a7e3ed1e | 1302 | if (pc && pc->core_id == core_id) { |
efc9f05d SE |
1303 | kfree(cpuc->shared_regs); |
1304 | cpuc->shared_regs = pc; | |
a7e3ed1e AK |
1305 | break; |
1306 | } | |
1307 | } | |
1308 | ||
efc9f05d SE |
1309 | cpuc->shared_regs->core_id = core_id; |
1310 | cpuc->shared_regs->refcnt++; | |
74846d35 PZ |
1311 | } |
1312 | ||
1313 | static void intel_pmu_cpu_dying(int cpu) | |
1314 | { | |
a7e3ed1e | 1315 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); |
efc9f05d | 1316 | struct intel_shared_regs *pc; |
a7e3ed1e | 1317 | |
efc9f05d | 1318 | pc = cpuc->shared_regs; |
a7e3ed1e AK |
1319 | if (pc) { |
1320 | if (pc->core_id == -1 || --pc->refcnt == 0) | |
1321 | kfree(pc); | |
efc9f05d | 1322 | cpuc->shared_regs = NULL; |
a7e3ed1e AK |
1323 | } |
1324 | ||
74846d35 PZ |
1325 | fini_debug_store_on_cpu(cpu); |
1326 | } | |
1327 | ||
caaa8be3 | 1328 | static __initconst const struct x86_pmu intel_pmu = { |
f22f54f4 PZ |
1329 | .name = "Intel", |
1330 | .handle_irq = intel_pmu_handle_irq, | |
1331 | .disable_all = intel_pmu_disable_all, | |
1332 | .enable_all = intel_pmu_enable_all, | |
1333 | .enable = intel_pmu_enable_event, | |
1334 | .disable = intel_pmu_disable_event, | |
b4cdc5c2 | 1335 | .hw_config = intel_pmu_hw_config, |
a072738e | 1336 | .schedule_events = x86_schedule_events, |
f22f54f4 PZ |
1337 | .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, |
1338 | .perfctr = MSR_ARCH_PERFMON_PERFCTR0, | |
1339 | .event_map = intel_pmu_event_map, | |
f22f54f4 PZ |
1340 | .max_events = ARRAY_SIZE(intel_perfmon_event_map), |
1341 | .apic = 1, | |
1342 | /* | |
1343 | * Intel PMCs cannot be accessed sanely above 32 bit width, | |
1344 | * so we install an artificial 1<<31 period regardless of | |
1345 | * the generic event period: | |
1346 | */ | |
1347 | .max_period = (1ULL << 31) - 1, | |
3f6da390 | 1348 | .get_event_constraints = intel_get_event_constraints, |
a7e3ed1e | 1349 | .put_event_constraints = intel_put_event_constraints, |
3f6da390 | 1350 | |
a7e3ed1e | 1351 | .cpu_prepare = intel_pmu_cpu_prepare, |
74846d35 PZ |
1352 | .cpu_starting = intel_pmu_cpu_starting, |
1353 | .cpu_dying = intel_pmu_cpu_dying, | |
f22f54f4 PZ |
1354 | }; |
1355 | ||
3c44780b PZ |
1356 | static void intel_clovertown_quirks(void) |
1357 | { | |
1358 | /* | |
1359 | * PEBS is unreliable due to: | |
1360 | * | |
1361 | * AJ67 - PEBS may experience CPL leaks | |
1362 | * AJ68 - PEBS PMI may be delayed by one event | |
1363 | * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12] | |
1364 | * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS | |
1365 | * | |
1366 | * AJ67 could be worked around by restricting the OS/USR flags. | |
1367 | * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI. | |
1368 | * | |
1369 | * AJ106 could possibly be worked around by not allowing LBR | |
1370 | * usage from PEBS, including the fixup. | |
1371 | * AJ68 could possibly be worked around by always programming | |
ec75a716 | 1372 | * a pebs_event_reset[0] value and coping with the lost events. |
3c44780b PZ |
1373 | * |
1374 | * But taken together it might just make sense to not enable PEBS on | |
1375 | * these chips. | |
1376 | */ | |
1377 | printk(KERN_WARNING "PEBS disabled due to CPU errata.\n"); | |
1378 | x86_pmu.pebs = 0; | |
1379 | x86_pmu.pebs_constraints = NULL; | |
1380 | } | |
1381 | ||
f22f54f4 PZ |
1382 | static __init int intel_pmu_init(void) |
1383 | { | |
1384 | union cpuid10_edx edx; | |
1385 | union cpuid10_eax eax; | |
1386 | unsigned int unused; | |
1387 | unsigned int ebx; | |
1388 | int version; | |
1389 | ||
1390 | if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) { | |
a072738e CG |
1391 | switch (boot_cpu_data.x86) { |
1392 | case 0x6: | |
1393 | return p6_pmu_init(); | |
1394 | case 0xf: | |
1395 | return p4_pmu_init(); | |
1396 | } | |
f22f54f4 | 1397 | return -ENODEV; |
f22f54f4 PZ |
1398 | } |
1399 | ||
1400 | /* | |
1401 | * Check whether the Architectural PerfMon supports | |
1402 | * Branch Misses Retired hw_event or not. | |
1403 | */ | |
1404 | cpuid(10, &eax.full, &ebx, &unused, &edx.full); | |
1405 | if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED) | |
1406 | return -ENODEV; | |
1407 | ||
1408 | version = eax.split.version_id; | |
1409 | if (version < 2) | |
1410 | x86_pmu = core_pmu; | |
1411 | else | |
1412 | x86_pmu = intel_pmu; | |
1413 | ||
1414 | x86_pmu.version = version; | |
948b1bb8 RR |
1415 | x86_pmu.num_counters = eax.split.num_counters; |
1416 | x86_pmu.cntval_bits = eax.split.bit_width; | |
1417 | x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1; | |
f22f54f4 PZ |
1418 | |
1419 | /* | |
1420 | * Quirk: v2 perfmon does not report fixed-purpose events, so | |
1421 | * assume at least 3 events: | |
1422 | */ | |
1423 | if (version > 1) | |
948b1bb8 | 1424 | x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3); |
f22f54f4 | 1425 | |
8db909a7 PZ |
1426 | /* |
1427 | * v2 and above have a perf capabilities MSR | |
1428 | */ | |
1429 | if (version > 1) { | |
1430 | u64 capabilities; | |
1431 | ||
1432 | rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities); | |
1433 | x86_pmu.intel_cap.capabilities = capabilities; | |
1434 | } | |
1435 | ||
ca037701 PZ |
1436 | intel_ds_init(); |
1437 | ||
f22f54f4 PZ |
1438 | /* |
1439 | * Install the hw-cache-events table: | |
1440 | */ | |
1441 | switch (boot_cpu_data.x86_model) { | |
1442 | case 14: /* 65 nm core solo/duo, "Yonah" */ | |
1443 | pr_cont("Core events, "); | |
1444 | break; | |
1445 | ||
1446 | case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */ | |
3c44780b | 1447 | x86_pmu.quirks = intel_clovertown_quirks; |
f22f54f4 PZ |
1448 | case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */ |
1449 | case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */ | |
1450 | case 29: /* six-core 45 nm xeon "Dunnington" */ | |
1451 | memcpy(hw_cache_event_ids, core2_hw_cache_event_ids, | |
1452 | sizeof(hw_cache_event_ids)); | |
1453 | ||
caff2bef PZ |
1454 | intel_pmu_lbr_init_core(); |
1455 | ||
f22f54f4 | 1456 | x86_pmu.event_constraints = intel_core2_event_constraints; |
17e31629 | 1457 | x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints; |
f22f54f4 PZ |
1458 | pr_cont("Core2 events, "); |
1459 | break; | |
1460 | ||
1461 | case 26: /* 45 nm nehalem, "Bloomfield" */ | |
1462 | case 30: /* 45 nm nehalem, "Lynnfield" */ | |
134fbadf | 1463 | case 46: /* 45 nm nehalem-ex, "Beckton" */ |
f22f54f4 PZ |
1464 | memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids, |
1465 | sizeof(hw_cache_event_ids)); | |
e994d7d2 AK |
1466 | memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, |
1467 | sizeof(hw_cache_extra_regs)); | |
f22f54f4 | 1468 | |
caff2bef PZ |
1469 | intel_pmu_lbr_init_nhm(); |
1470 | ||
f22f54f4 | 1471 | x86_pmu.event_constraints = intel_nehalem_event_constraints; |
17e31629 | 1472 | x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints; |
11164cd4 | 1473 | x86_pmu.enable_all = intel_pmu_nhm_enable_all; |
a7e3ed1e | 1474 | x86_pmu.extra_regs = intel_nehalem_extra_regs; |
ec75a716 | 1475 | |
91fc4cc0 IM |
1476 | /* UOPS_ISSUED.STALLED_CYCLES */ |
1477 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x180010e; | |
1478 | /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ | |
8f622422 | 1479 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x1803fb1; |
94403f88 | 1480 | |
ec75a716 IM |
1481 | if (ebx & 0x40) { |
1482 | /* | |
1483 | * Erratum AAJ80 detected, we work it around by using | |
1484 | * the BR_MISP_EXEC.ANY event. This will over-count | |
1485 | * branch-misses, but it's still much better than the | |
1486 | * architectural event which is often completely bogus: | |
1487 | */ | |
1488 | intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89; | |
1489 | ||
1490 | pr_cont("erratum AAJ80 worked around, "); | |
1491 | } | |
11164cd4 | 1492 | pr_cont("Nehalem events, "); |
f22f54f4 | 1493 | break; |
caff2bef | 1494 | |
b622d644 | 1495 | case 28: /* Atom */ |
f22f54f4 PZ |
1496 | memcpy(hw_cache_event_ids, atom_hw_cache_event_ids, |
1497 | sizeof(hw_cache_event_ids)); | |
1498 | ||
caff2bef PZ |
1499 | intel_pmu_lbr_init_atom(); |
1500 | ||
f22f54f4 | 1501 | x86_pmu.event_constraints = intel_gen_event_constraints; |
17e31629 | 1502 | x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints; |
f22f54f4 PZ |
1503 | pr_cont("Atom events, "); |
1504 | break; | |
1505 | ||
1506 | case 37: /* 32 nm nehalem, "Clarkdale" */ | |
1507 | case 44: /* 32 nm nehalem, "Gulftown" */ | |
b2508e82 | 1508 | case 47: /* 32 nm Xeon E7 */ |
f22f54f4 PZ |
1509 | memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids, |
1510 | sizeof(hw_cache_event_ids)); | |
e994d7d2 AK |
1511 | memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, |
1512 | sizeof(hw_cache_extra_regs)); | |
f22f54f4 | 1513 | |
caff2bef PZ |
1514 | intel_pmu_lbr_init_nhm(); |
1515 | ||
f22f54f4 | 1516 | x86_pmu.event_constraints = intel_westmere_event_constraints; |
40b91cd1 | 1517 | x86_pmu.enable_all = intel_pmu_nhm_enable_all; |
17e31629 | 1518 | x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints; |
a7e3ed1e | 1519 | x86_pmu.extra_regs = intel_westmere_extra_regs; |
b79e8941 | 1520 | x86_pmu.er_flags |= ERF_HAS_RSP_1; |
30112039 IM |
1521 | |
1522 | /* UOPS_ISSUED.STALLED_CYCLES */ | |
1523 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x180010e; | |
1524 | /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ | |
1525 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x1803fb1; | |
1526 | ||
f22f54f4 PZ |
1527 | pr_cont("Westmere events, "); |
1528 | break; | |
b622d644 | 1529 | |
b06b3d49 LM |
1530 | case 42: /* SandyBridge */ |
1531 | memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, | |
1532 | sizeof(hw_cache_event_ids)); | |
1533 | ||
1534 | intel_pmu_lbr_init_nhm(); | |
1535 | ||
1536 | x86_pmu.event_constraints = intel_snb_event_constraints; | |
1537 | x86_pmu.pebs_constraints = intel_snb_pebs_events; | |
ee89cbc2 SE |
1538 | x86_pmu.extra_regs = intel_snb_extra_regs; |
1539 | /* all extra regs are per-cpu when HT is on */ | |
b79e8941 PZ |
1540 | x86_pmu.er_flags |= ERF_HAS_RSP_1; |
1541 | x86_pmu.er_flags |= ERF_NO_HT_SHARING; | |
e04d1b23 LM |
1542 | |
1543 | /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ | |
1544 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x180010e; | |
1545 | /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/ | |
1546 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x18001b1; | |
1547 | ||
b06b3d49 LM |
1548 | pr_cont("SandyBridge events, "); |
1549 | break; | |
1550 | ||
f22f54f4 PZ |
1551 | default: |
1552 | /* | |
1553 | * default constraints for v2 and up | |
1554 | */ | |
1555 | x86_pmu.event_constraints = intel_gen_event_constraints; | |
1556 | pr_cont("generic architected perfmon, "); | |
1557 | } | |
1558 | return 0; | |
1559 | } | |
1560 | ||
1561 | #else /* CONFIG_CPU_SUP_INTEL */ | |
1562 | ||
1563 | static int intel_pmu_init(void) | |
1564 | { | |
1565 | return 0; | |
1566 | } | |
1567 | ||
cd8a38d3 SE |
1568 | static struct intel_shared_regs *allocate_shared_regs(int cpu) |
1569 | { | |
1570 | return NULL; | |
1571 | } | |
f22f54f4 | 1572 | #endif /* CONFIG_CPU_SUP_INTEL */ |