Commit | Line | Data |
---|---|---|
a7e3ed1e | 1 | /* |
efc9f05d SE |
2 | * Per core/cpu state |
3 | * | |
4 | * Used to coordinate shared registers between HT threads or | |
5 | * among events on a single PMU. | |
a7e3ed1e | 6 | */ |
de0428a7 | 7 | |
c767a54b JP |
8 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
9 | ||
de0428a7 KW |
10 | #include <linux/stddef.h> |
11 | #include <linux/types.h> | |
12 | #include <linux/init.h> | |
13 | #include <linux/slab.h> | |
69c60c88 | 14 | #include <linux/export.h> |
de0428a7 | 15 | |
3a632cb2 | 16 | #include <asm/cpufeature.h> |
de0428a7 KW |
17 | #include <asm/hardirq.h> |
18 | #include <asm/apic.h> | |
19 | ||
20 | #include "perf_event.h" | |
a7e3ed1e | 21 | |
f22f54f4 | 22 | /* |
b622d644 | 23 | * Intel PerfMon, used on Core and later. |
f22f54f4 | 24 | */ |
ec75a716 | 25 | static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly = |
f22f54f4 | 26 | { |
c3b7cdf1 PE |
27 | [PERF_COUNT_HW_CPU_CYCLES] = 0x003c, |
28 | [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0, | |
29 | [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e, | |
30 | [PERF_COUNT_HW_CACHE_MISSES] = 0x412e, | |
31 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4, | |
32 | [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5, | |
33 | [PERF_COUNT_HW_BUS_CYCLES] = 0x013c, | |
34 | [PERF_COUNT_HW_REF_CPU_CYCLES] = 0x0300, /* pseudo-encoding */ | |
f22f54f4 PZ |
35 | }; |
36 | ||
5c543e3c | 37 | static struct event_constraint intel_core_event_constraints[] __read_mostly = |
f22f54f4 PZ |
38 | { |
39 | INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ | |
40 | INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ | |
41 | INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */ | |
42 | INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */ | |
43 | INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */ | |
44 | INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */ | |
45 | EVENT_CONSTRAINT_END | |
46 | }; | |
47 | ||
5c543e3c | 48 | static struct event_constraint intel_core2_event_constraints[] __read_mostly = |
f22f54f4 | 49 | { |
b622d644 PZ |
50 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
51 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
cd09c0c4 | 52 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ |
f22f54f4 PZ |
53 | INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */ |
54 | INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ | |
55 | INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ | |
56 | INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */ | |
57 | INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */ | |
58 | INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */ | |
59 | INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */ | |
60 | INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */ | |
b622d644 | 61 | INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */ |
f22f54f4 PZ |
62 | INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */ |
63 | EVENT_CONSTRAINT_END | |
64 | }; | |
65 | ||
5c543e3c | 66 | static struct event_constraint intel_nehalem_event_constraints[] __read_mostly = |
f22f54f4 | 67 | { |
b622d644 PZ |
68 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
69 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
cd09c0c4 | 70 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ |
f22f54f4 PZ |
71 | INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */ |
72 | INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */ | |
73 | INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */ | |
74 | INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */ | |
75 | INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */ | |
76 | INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */ | |
77 | INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ | |
78 | INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ | |
79 | EVENT_CONSTRAINT_END | |
80 | }; | |
81 | ||
5c543e3c | 82 | static struct extra_reg intel_nehalem_extra_regs[] __read_mostly = |
a7e3ed1e | 83 | { |
53ad0447 YZ |
84 | /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ |
85 | INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), | |
f20093ee | 86 | INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b), |
a7e3ed1e AK |
87 | EVENT_EXTRA_END |
88 | }; | |
89 | ||
5c543e3c | 90 | static struct event_constraint intel_westmere_event_constraints[] __read_mostly = |
f22f54f4 | 91 | { |
b622d644 PZ |
92 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
93 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
cd09c0c4 | 94 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ |
f22f54f4 PZ |
95 | INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ |
96 | INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */ | |
97 | INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ | |
d1100770 | 98 | INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */ |
f22f54f4 PZ |
99 | EVENT_CONSTRAINT_END |
100 | }; | |
101 | ||
5c543e3c | 102 | static struct event_constraint intel_snb_event_constraints[] __read_mostly = |
b06b3d49 LM |
103 | { |
104 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | |
105 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
cd09c0c4 | 106 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ |
fd4a5aef SE |
107 | INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */ |
108 | INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */ | |
109 | INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ | |
110 | INTEL_UEVENT_CONSTRAINT(0x06a3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ | |
b06b3d49 | 111 | INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */ |
b06b3d49 LM |
112 | INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ |
113 | INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ | |
f8378f52 AK |
114 | INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */ |
115 | INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ | |
b06b3d49 LM |
116 | EVENT_CONSTRAINT_END |
117 | }; | |
118 | ||
69943182 SE |
119 | static struct event_constraint intel_ivb_event_constraints[] __read_mostly = |
120 | { | |
121 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | |
122 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
123 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ | |
124 | INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */ | |
125 | INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMTPY */ | |
126 | INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */ | |
127 | INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */ | |
128 | INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */ | |
129 | INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */ | |
130 | INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ | |
131 | INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ | |
132 | INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ | |
741a698f PZ |
133 | /* |
134 | * Errata BV98 -- MEM_*_RETIRED events can leak between counters of SMT | |
135 | * siblings; disable these events because they can corrupt unrelated | |
136 | * counters. | |
137 | */ | |
138 | INTEL_EVENT_CONSTRAINT(0xd0, 0x0), /* MEM_UOPS_RETIRED.* */ | |
139 | INTEL_EVENT_CONSTRAINT(0xd1, 0x0), /* MEM_LOAD_UOPS_RETIRED.* */ | |
140 | INTEL_EVENT_CONSTRAINT(0xd2, 0x0), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ | |
141 | INTEL_EVENT_CONSTRAINT(0xd3, 0x0), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ | |
69943182 SE |
142 | EVENT_CONSTRAINT_END |
143 | }; | |
144 | ||
5c543e3c | 145 | static struct extra_reg intel_westmere_extra_regs[] __read_mostly = |
a7e3ed1e | 146 | { |
53ad0447 YZ |
147 | /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ |
148 | INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), | |
149 | INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1), | |
f20093ee | 150 | INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b), |
a7e3ed1e AK |
151 | EVENT_EXTRA_END |
152 | }; | |
153 | ||
0af3ac1f AK |
154 | static struct event_constraint intel_v1_event_constraints[] __read_mostly = |
155 | { | |
156 | EVENT_CONSTRAINT_END | |
157 | }; | |
158 | ||
5c543e3c | 159 | static struct event_constraint intel_gen_event_constraints[] __read_mostly = |
f22f54f4 | 160 | { |
b622d644 PZ |
161 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
162 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
cd09c0c4 | 163 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ |
f22f54f4 PZ |
164 | EVENT_CONSTRAINT_END |
165 | }; | |
166 | ||
ee89cbc2 | 167 | static struct extra_reg intel_snb_extra_regs[] __read_mostly = { |
53ad0447 YZ |
168 | /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ |
169 | INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0), | |
170 | INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1), | |
f20093ee | 171 | INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), |
f1923820 SE |
172 | EVENT_EXTRA_END |
173 | }; | |
174 | ||
175 | static struct extra_reg intel_snbep_extra_regs[] __read_mostly = { | |
53ad0447 YZ |
176 | /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ |
177 | INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0), | |
178 | INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1), | |
f1a52789 | 179 | INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), |
ee89cbc2 SE |
180 | EVENT_EXTRA_END |
181 | }; | |
182 | ||
f20093ee SE |
183 | EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3"); |
184 | EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3"); | |
9ad64c0f | 185 | EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2"); |
f20093ee SE |
186 | |
187 | struct attribute *nhm_events_attrs[] = { | |
188 | EVENT_PTR(mem_ld_nhm), | |
189 | NULL, | |
190 | }; | |
191 | ||
192 | struct attribute *snb_events_attrs[] = { | |
193 | EVENT_PTR(mem_ld_snb), | |
9ad64c0f | 194 | EVENT_PTR(mem_st_snb), |
f20093ee SE |
195 | NULL, |
196 | }; | |
197 | ||
3a632cb2 AK |
198 | static struct event_constraint intel_hsw_event_constraints[] = { |
199 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | |
200 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
201 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ | |
202 | INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.* */ | |
203 | INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ | |
204 | INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ | |
205 | /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ | |
206 | INTEL_EVENT_CONSTRAINT(0x08a3, 0x4), | |
207 | /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ | |
208 | INTEL_EVENT_CONSTRAINT(0x0ca3, 0x4), | |
209 | /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */ | |
210 | INTEL_EVENT_CONSTRAINT(0x04a3, 0xf), | |
211 | EVENT_CONSTRAINT_END | |
212 | }; | |
213 | ||
f22f54f4 PZ |
214 | static u64 intel_pmu_event_map(int hw_event) |
215 | { | |
216 | return intel_perfmon_event_map[hw_event]; | |
217 | } | |
218 | ||
74e6543f YZ |
219 | #define SNB_DMND_DATA_RD (1ULL << 0) |
220 | #define SNB_DMND_RFO (1ULL << 1) | |
221 | #define SNB_DMND_IFETCH (1ULL << 2) | |
222 | #define SNB_DMND_WB (1ULL << 3) | |
223 | #define SNB_PF_DATA_RD (1ULL << 4) | |
224 | #define SNB_PF_RFO (1ULL << 5) | |
225 | #define SNB_PF_IFETCH (1ULL << 6) | |
226 | #define SNB_LLC_DATA_RD (1ULL << 7) | |
227 | #define SNB_LLC_RFO (1ULL << 8) | |
228 | #define SNB_LLC_IFETCH (1ULL << 9) | |
229 | #define SNB_BUS_LOCKS (1ULL << 10) | |
230 | #define SNB_STRM_ST (1ULL << 11) | |
231 | #define SNB_OTHER (1ULL << 15) | |
232 | #define SNB_RESP_ANY (1ULL << 16) | |
233 | #define SNB_NO_SUPP (1ULL << 17) | |
234 | #define SNB_LLC_HITM (1ULL << 18) | |
235 | #define SNB_LLC_HITE (1ULL << 19) | |
236 | #define SNB_LLC_HITS (1ULL << 20) | |
237 | #define SNB_LLC_HITF (1ULL << 21) | |
238 | #define SNB_LOCAL (1ULL << 22) | |
239 | #define SNB_REMOTE (0xffULL << 23) | |
240 | #define SNB_SNP_NONE (1ULL << 31) | |
241 | #define SNB_SNP_NOT_NEEDED (1ULL << 32) | |
242 | #define SNB_SNP_MISS (1ULL << 33) | |
243 | #define SNB_NO_FWD (1ULL << 34) | |
244 | #define SNB_SNP_FWD (1ULL << 35) | |
245 | #define SNB_HITM (1ULL << 36) | |
246 | #define SNB_NON_DRAM (1ULL << 37) | |
247 | ||
248 | #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD) | |
249 | #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO) | |
250 | #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) | |
251 | ||
252 | #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \ | |
253 | SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \ | |
254 | SNB_HITM) | |
255 | ||
256 | #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY) | |
257 | #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY) | |
258 | ||
259 | #define SNB_L3_ACCESS SNB_RESP_ANY | |
260 | #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM) | |
261 | ||
262 | static __initconst const u64 snb_hw_cache_extra_regs | |
263 | [PERF_COUNT_HW_CACHE_MAX] | |
264 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
265 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
266 | { | |
267 | [ C(LL ) ] = { | |
268 | [ C(OP_READ) ] = { | |
269 | [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS, | |
270 | [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_L3_MISS, | |
271 | }, | |
272 | [ C(OP_WRITE) ] = { | |
273 | [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS, | |
274 | [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_L3_MISS, | |
275 | }, | |
276 | [ C(OP_PREFETCH) ] = { | |
277 | [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS, | |
278 | [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_L3_MISS, | |
279 | }, | |
280 | }, | |
281 | [ C(NODE) ] = { | |
282 | [ C(OP_READ) ] = { | |
283 | [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY, | |
284 | [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_DRAM_REMOTE, | |
285 | }, | |
286 | [ C(OP_WRITE) ] = { | |
287 | [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY, | |
288 | [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE, | |
289 | }, | |
290 | [ C(OP_PREFETCH) ] = { | |
291 | [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY, | |
292 | [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE, | |
293 | }, | |
294 | }, | |
295 | }; | |
296 | ||
b06b3d49 LM |
297 | static __initconst const u64 snb_hw_cache_event_ids |
298 | [PERF_COUNT_HW_CACHE_MAX] | |
299 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
300 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
301 | { | |
302 | [ C(L1D) ] = { | |
303 | [ C(OP_READ) ] = { | |
304 | [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */ | |
305 | [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */ | |
306 | }, | |
307 | [ C(OP_WRITE) ] = { | |
308 | [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */ | |
309 | [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */ | |
310 | }, | |
311 | [ C(OP_PREFETCH) ] = { | |
312 | [ C(RESULT_ACCESS) ] = 0x0, | |
313 | [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */ | |
314 | }, | |
315 | }, | |
316 | [ C(L1I ) ] = { | |
317 | [ C(OP_READ) ] = { | |
318 | [ C(RESULT_ACCESS) ] = 0x0, | |
319 | [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */ | |
320 | }, | |
321 | [ C(OP_WRITE) ] = { | |
322 | [ C(RESULT_ACCESS) ] = -1, | |
323 | [ C(RESULT_MISS) ] = -1, | |
324 | }, | |
325 | [ C(OP_PREFETCH) ] = { | |
326 | [ C(RESULT_ACCESS) ] = 0x0, | |
327 | [ C(RESULT_MISS) ] = 0x0, | |
328 | }, | |
329 | }, | |
330 | [ C(LL ) ] = { | |
b06b3d49 | 331 | [ C(OP_READ) ] = { |
63b6a675 | 332 | /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ |
b06b3d49 | 333 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
334 | /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ |
335 | [ C(RESULT_MISS) ] = 0x01b7, | |
b06b3d49 LM |
336 | }, |
337 | [ C(OP_WRITE) ] = { | |
63b6a675 | 338 | /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ |
b06b3d49 | 339 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
340 | /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ |
341 | [ C(RESULT_MISS) ] = 0x01b7, | |
b06b3d49 LM |
342 | }, |
343 | [ C(OP_PREFETCH) ] = { | |
63b6a675 | 344 | /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ |
b06b3d49 | 345 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
346 | /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ |
347 | [ C(RESULT_MISS) ] = 0x01b7, | |
b06b3d49 LM |
348 | }, |
349 | }, | |
350 | [ C(DTLB) ] = { | |
351 | [ C(OP_READ) ] = { | |
352 | [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */ | |
353 | [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */ | |
354 | }, | |
355 | [ C(OP_WRITE) ] = { | |
356 | [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */ | |
357 | [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */ | |
358 | }, | |
359 | [ C(OP_PREFETCH) ] = { | |
360 | [ C(RESULT_ACCESS) ] = 0x0, | |
361 | [ C(RESULT_MISS) ] = 0x0, | |
362 | }, | |
363 | }, | |
364 | [ C(ITLB) ] = { | |
365 | [ C(OP_READ) ] = { | |
366 | [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */ | |
367 | [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */ | |
368 | }, | |
369 | [ C(OP_WRITE) ] = { | |
370 | [ C(RESULT_ACCESS) ] = -1, | |
371 | [ C(RESULT_MISS) ] = -1, | |
372 | }, | |
373 | [ C(OP_PREFETCH) ] = { | |
374 | [ C(RESULT_ACCESS) ] = -1, | |
375 | [ C(RESULT_MISS) ] = -1, | |
376 | }, | |
377 | }, | |
378 | [ C(BPU ) ] = { | |
379 | [ C(OP_READ) ] = { | |
380 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
381 | [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ | |
382 | }, | |
383 | [ C(OP_WRITE) ] = { | |
384 | [ C(RESULT_ACCESS) ] = -1, | |
385 | [ C(RESULT_MISS) ] = -1, | |
386 | }, | |
387 | [ C(OP_PREFETCH) ] = { | |
388 | [ C(RESULT_ACCESS) ] = -1, | |
389 | [ C(RESULT_MISS) ] = -1, | |
390 | }, | |
391 | }, | |
89d6c0b5 PZ |
392 | [ C(NODE) ] = { |
393 | [ C(OP_READ) ] = { | |
74e6543f YZ |
394 | [ C(RESULT_ACCESS) ] = 0x01b7, |
395 | [ C(RESULT_MISS) ] = 0x01b7, | |
89d6c0b5 PZ |
396 | }, |
397 | [ C(OP_WRITE) ] = { | |
74e6543f YZ |
398 | [ C(RESULT_ACCESS) ] = 0x01b7, |
399 | [ C(RESULT_MISS) ] = 0x01b7, | |
89d6c0b5 PZ |
400 | }, |
401 | [ C(OP_PREFETCH) ] = { | |
74e6543f YZ |
402 | [ C(RESULT_ACCESS) ] = 0x01b7, |
403 | [ C(RESULT_MISS) ] = 0x01b7, | |
89d6c0b5 PZ |
404 | }, |
405 | }, | |
406 | ||
b06b3d49 LM |
407 | }; |
408 | ||
caaa8be3 | 409 | static __initconst const u64 westmere_hw_cache_event_ids |
f22f54f4 PZ |
410 | [PERF_COUNT_HW_CACHE_MAX] |
411 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
412 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
413 | { | |
414 | [ C(L1D) ] = { | |
415 | [ C(OP_READ) ] = { | |
416 | [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ | |
417 | [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ | |
418 | }, | |
419 | [ C(OP_WRITE) ] = { | |
420 | [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ | |
421 | [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ | |
422 | }, | |
423 | [ C(OP_PREFETCH) ] = { | |
424 | [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ | |
425 | [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ | |
426 | }, | |
427 | }, | |
428 | [ C(L1I ) ] = { | |
429 | [ C(OP_READ) ] = { | |
430 | [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ | |
431 | [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ | |
432 | }, | |
433 | [ C(OP_WRITE) ] = { | |
434 | [ C(RESULT_ACCESS) ] = -1, | |
435 | [ C(RESULT_MISS) ] = -1, | |
436 | }, | |
437 | [ C(OP_PREFETCH) ] = { | |
438 | [ C(RESULT_ACCESS) ] = 0x0, | |
439 | [ C(RESULT_MISS) ] = 0x0, | |
440 | }, | |
441 | }, | |
442 | [ C(LL ) ] = { | |
443 | [ C(OP_READ) ] = { | |
63b6a675 | 444 | /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ |
e994d7d2 | 445 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
446 | /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ |
447 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 | 448 | }, |
e994d7d2 AK |
449 | /* |
450 | * Use RFO, not WRITEBACK, because a write miss would typically occur | |
451 | * on RFO. | |
452 | */ | |
f22f54f4 | 453 | [ C(OP_WRITE) ] = { |
63b6a675 PZ |
454 | /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ |
455 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
456 | /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ | |
e994d7d2 | 457 | [ C(RESULT_MISS) ] = 0x01b7, |
f22f54f4 PZ |
458 | }, |
459 | [ C(OP_PREFETCH) ] = { | |
63b6a675 | 460 | /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ |
e994d7d2 | 461 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
462 | /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ |
463 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 PZ |
464 | }, |
465 | }, | |
466 | [ C(DTLB) ] = { | |
467 | [ C(OP_READ) ] = { | |
468 | [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ | |
469 | [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ | |
470 | }, | |
471 | [ C(OP_WRITE) ] = { | |
472 | [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ | |
473 | [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ | |
474 | }, | |
475 | [ C(OP_PREFETCH) ] = { | |
476 | [ C(RESULT_ACCESS) ] = 0x0, | |
477 | [ C(RESULT_MISS) ] = 0x0, | |
478 | }, | |
479 | }, | |
480 | [ C(ITLB) ] = { | |
481 | [ C(OP_READ) ] = { | |
482 | [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ | |
483 | [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */ | |
484 | }, | |
485 | [ C(OP_WRITE) ] = { | |
486 | [ C(RESULT_ACCESS) ] = -1, | |
487 | [ C(RESULT_MISS) ] = -1, | |
488 | }, | |
489 | [ C(OP_PREFETCH) ] = { | |
490 | [ C(RESULT_ACCESS) ] = -1, | |
491 | [ C(RESULT_MISS) ] = -1, | |
492 | }, | |
493 | }, | |
494 | [ C(BPU ) ] = { | |
495 | [ C(OP_READ) ] = { | |
496 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
497 | [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ | |
498 | }, | |
499 | [ C(OP_WRITE) ] = { | |
500 | [ C(RESULT_ACCESS) ] = -1, | |
501 | [ C(RESULT_MISS) ] = -1, | |
502 | }, | |
503 | [ C(OP_PREFETCH) ] = { | |
504 | [ C(RESULT_ACCESS) ] = -1, | |
505 | [ C(RESULT_MISS) ] = -1, | |
506 | }, | |
507 | }, | |
89d6c0b5 PZ |
508 | [ C(NODE) ] = { |
509 | [ C(OP_READ) ] = { | |
510 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
511 | [ C(RESULT_MISS) ] = 0x01b7, | |
512 | }, | |
513 | [ C(OP_WRITE) ] = { | |
514 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
515 | [ C(RESULT_MISS) ] = 0x01b7, | |
516 | }, | |
517 | [ C(OP_PREFETCH) ] = { | |
518 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
519 | [ C(RESULT_MISS) ] = 0x01b7, | |
520 | }, | |
521 | }, | |
f22f54f4 PZ |
522 | }; |
523 | ||
e994d7d2 | 524 | /* |
63b6a675 PZ |
525 | * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits; |
526 | * See IA32 SDM Vol 3B 30.6.1.3 | |
e994d7d2 AK |
527 | */ |
528 | ||
63b6a675 PZ |
529 | #define NHM_DMND_DATA_RD (1 << 0) |
530 | #define NHM_DMND_RFO (1 << 1) | |
531 | #define NHM_DMND_IFETCH (1 << 2) | |
532 | #define NHM_DMND_WB (1 << 3) | |
533 | #define NHM_PF_DATA_RD (1 << 4) | |
534 | #define NHM_PF_DATA_RFO (1 << 5) | |
535 | #define NHM_PF_IFETCH (1 << 6) | |
536 | #define NHM_OFFCORE_OTHER (1 << 7) | |
537 | #define NHM_UNCORE_HIT (1 << 8) | |
538 | #define NHM_OTHER_CORE_HIT_SNP (1 << 9) | |
539 | #define NHM_OTHER_CORE_HITM (1 << 10) | |
540 | /* reserved */ | |
541 | #define NHM_REMOTE_CACHE_FWD (1 << 12) | |
542 | #define NHM_REMOTE_DRAM (1 << 13) | |
543 | #define NHM_LOCAL_DRAM (1 << 14) | |
544 | #define NHM_NON_DRAM (1 << 15) | |
545 | ||
87e24f4b PZ |
546 | #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD) |
547 | #define NHM_REMOTE (NHM_REMOTE_DRAM) | |
63b6a675 PZ |
548 | |
549 | #define NHM_DMND_READ (NHM_DMND_DATA_RD) | |
550 | #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB) | |
551 | #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO) | |
552 | ||
553 | #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM) | |
87e24f4b | 554 | #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD) |
63b6a675 | 555 | #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS) |
e994d7d2 AK |
556 | |
557 | static __initconst const u64 nehalem_hw_cache_extra_regs | |
558 | [PERF_COUNT_HW_CACHE_MAX] | |
559 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
560 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
561 | { | |
562 | [ C(LL ) ] = { | |
563 | [ C(OP_READ) ] = { | |
63b6a675 PZ |
564 | [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS, |
565 | [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS, | |
e994d7d2 AK |
566 | }, |
567 | [ C(OP_WRITE) ] = { | |
63b6a675 PZ |
568 | [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS, |
569 | [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS, | |
e994d7d2 AK |
570 | }, |
571 | [ C(OP_PREFETCH) ] = { | |
63b6a675 PZ |
572 | [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS, |
573 | [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS, | |
e994d7d2 | 574 | }, |
89d6c0b5 PZ |
575 | }, |
576 | [ C(NODE) ] = { | |
577 | [ C(OP_READ) ] = { | |
87e24f4b PZ |
578 | [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE, |
579 | [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE, | |
89d6c0b5 PZ |
580 | }, |
581 | [ C(OP_WRITE) ] = { | |
87e24f4b PZ |
582 | [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE, |
583 | [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE, | |
89d6c0b5 PZ |
584 | }, |
585 | [ C(OP_PREFETCH) ] = { | |
87e24f4b PZ |
586 | [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE, |
587 | [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE, | |
89d6c0b5 PZ |
588 | }, |
589 | }, | |
e994d7d2 AK |
590 | }; |
591 | ||
caaa8be3 | 592 | static __initconst const u64 nehalem_hw_cache_event_ids |
f22f54f4 PZ |
593 | [PERF_COUNT_HW_CACHE_MAX] |
594 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
595 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
596 | { | |
597 | [ C(L1D) ] = { | |
598 | [ C(OP_READ) ] = { | |
f4929bd3 PZ |
599 | [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ |
600 | [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ | |
f22f54f4 PZ |
601 | }, |
602 | [ C(OP_WRITE) ] = { | |
f4929bd3 PZ |
603 | [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ |
604 | [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ | |
f22f54f4 PZ |
605 | }, |
606 | [ C(OP_PREFETCH) ] = { | |
607 | [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ | |
608 | [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ | |
609 | }, | |
610 | }, | |
611 | [ C(L1I ) ] = { | |
612 | [ C(OP_READ) ] = { | |
613 | [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ | |
614 | [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ | |
615 | }, | |
616 | [ C(OP_WRITE) ] = { | |
617 | [ C(RESULT_ACCESS) ] = -1, | |
618 | [ C(RESULT_MISS) ] = -1, | |
619 | }, | |
620 | [ C(OP_PREFETCH) ] = { | |
621 | [ C(RESULT_ACCESS) ] = 0x0, | |
622 | [ C(RESULT_MISS) ] = 0x0, | |
623 | }, | |
624 | }, | |
625 | [ C(LL ) ] = { | |
626 | [ C(OP_READ) ] = { | |
e994d7d2 AK |
627 | /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ |
628 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
629 | /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ | |
630 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 | 631 | }, |
e994d7d2 AK |
632 | /* |
633 | * Use RFO, not WRITEBACK, because a write miss would typically occur | |
634 | * on RFO. | |
635 | */ | |
f22f54f4 | 636 | [ C(OP_WRITE) ] = { |
e994d7d2 AK |
637 | /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ |
638 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
639 | /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ | |
640 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 PZ |
641 | }, |
642 | [ C(OP_PREFETCH) ] = { | |
e994d7d2 AK |
643 | /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ |
644 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
645 | /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ | |
646 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 PZ |
647 | }, |
648 | }, | |
649 | [ C(DTLB) ] = { | |
650 | [ C(OP_READ) ] = { | |
651 | [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ | |
652 | [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ | |
653 | }, | |
654 | [ C(OP_WRITE) ] = { | |
655 | [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ | |
656 | [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ | |
657 | }, | |
658 | [ C(OP_PREFETCH) ] = { | |
659 | [ C(RESULT_ACCESS) ] = 0x0, | |
660 | [ C(RESULT_MISS) ] = 0x0, | |
661 | }, | |
662 | }, | |
663 | [ C(ITLB) ] = { | |
664 | [ C(OP_READ) ] = { | |
665 | [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ | |
666 | [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */ | |
667 | }, | |
668 | [ C(OP_WRITE) ] = { | |
669 | [ C(RESULT_ACCESS) ] = -1, | |
670 | [ C(RESULT_MISS) ] = -1, | |
671 | }, | |
672 | [ C(OP_PREFETCH) ] = { | |
673 | [ C(RESULT_ACCESS) ] = -1, | |
674 | [ C(RESULT_MISS) ] = -1, | |
675 | }, | |
676 | }, | |
677 | [ C(BPU ) ] = { | |
678 | [ C(OP_READ) ] = { | |
679 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
680 | [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ | |
681 | }, | |
682 | [ C(OP_WRITE) ] = { | |
683 | [ C(RESULT_ACCESS) ] = -1, | |
684 | [ C(RESULT_MISS) ] = -1, | |
685 | }, | |
686 | [ C(OP_PREFETCH) ] = { | |
687 | [ C(RESULT_ACCESS) ] = -1, | |
688 | [ C(RESULT_MISS) ] = -1, | |
689 | }, | |
690 | }, | |
89d6c0b5 PZ |
691 | [ C(NODE) ] = { |
692 | [ C(OP_READ) ] = { | |
693 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
694 | [ C(RESULT_MISS) ] = 0x01b7, | |
695 | }, | |
696 | [ C(OP_WRITE) ] = { | |
697 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
698 | [ C(RESULT_MISS) ] = 0x01b7, | |
699 | }, | |
700 | [ C(OP_PREFETCH) ] = { | |
701 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
702 | [ C(RESULT_MISS) ] = 0x01b7, | |
703 | }, | |
704 | }, | |
f22f54f4 PZ |
705 | }; |
706 | ||
caaa8be3 | 707 | static __initconst const u64 core2_hw_cache_event_ids |
f22f54f4 PZ |
708 | [PERF_COUNT_HW_CACHE_MAX] |
709 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
710 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
711 | { | |
712 | [ C(L1D) ] = { | |
713 | [ C(OP_READ) ] = { | |
714 | [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */ | |
715 | [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */ | |
716 | }, | |
717 | [ C(OP_WRITE) ] = { | |
718 | [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */ | |
719 | [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */ | |
720 | }, | |
721 | [ C(OP_PREFETCH) ] = { | |
722 | [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */ | |
723 | [ C(RESULT_MISS) ] = 0, | |
724 | }, | |
725 | }, | |
726 | [ C(L1I ) ] = { | |
727 | [ C(OP_READ) ] = { | |
728 | [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */ | |
729 | [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */ | |
730 | }, | |
731 | [ C(OP_WRITE) ] = { | |
732 | [ C(RESULT_ACCESS) ] = -1, | |
733 | [ C(RESULT_MISS) ] = -1, | |
734 | }, | |
735 | [ C(OP_PREFETCH) ] = { | |
736 | [ C(RESULT_ACCESS) ] = 0, | |
737 | [ C(RESULT_MISS) ] = 0, | |
738 | }, | |
739 | }, | |
740 | [ C(LL ) ] = { | |
741 | [ C(OP_READ) ] = { | |
742 | [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ | |
743 | [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ | |
744 | }, | |
745 | [ C(OP_WRITE) ] = { | |
746 | [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ | |
747 | [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ | |
748 | }, | |
749 | [ C(OP_PREFETCH) ] = { | |
750 | [ C(RESULT_ACCESS) ] = 0, | |
751 | [ C(RESULT_MISS) ] = 0, | |
752 | }, | |
753 | }, | |
754 | [ C(DTLB) ] = { | |
755 | [ C(OP_READ) ] = { | |
756 | [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ | |
757 | [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */ | |
758 | }, | |
759 | [ C(OP_WRITE) ] = { | |
760 | [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ | |
761 | [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */ | |
762 | }, | |
763 | [ C(OP_PREFETCH) ] = { | |
764 | [ C(RESULT_ACCESS) ] = 0, | |
765 | [ C(RESULT_MISS) ] = 0, | |
766 | }, | |
767 | }, | |
768 | [ C(ITLB) ] = { | |
769 | [ C(OP_READ) ] = { | |
770 | [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ | |
771 | [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */ | |
772 | }, | |
773 | [ C(OP_WRITE) ] = { | |
774 | [ C(RESULT_ACCESS) ] = -1, | |
775 | [ C(RESULT_MISS) ] = -1, | |
776 | }, | |
777 | [ C(OP_PREFETCH) ] = { | |
778 | [ C(RESULT_ACCESS) ] = -1, | |
779 | [ C(RESULT_MISS) ] = -1, | |
780 | }, | |
781 | }, | |
782 | [ C(BPU ) ] = { | |
783 | [ C(OP_READ) ] = { | |
784 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ | |
785 | [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ | |
786 | }, | |
787 | [ C(OP_WRITE) ] = { | |
788 | [ C(RESULT_ACCESS) ] = -1, | |
789 | [ C(RESULT_MISS) ] = -1, | |
790 | }, | |
791 | [ C(OP_PREFETCH) ] = { | |
792 | [ C(RESULT_ACCESS) ] = -1, | |
793 | [ C(RESULT_MISS) ] = -1, | |
794 | }, | |
795 | }, | |
796 | }; | |
797 | ||
caaa8be3 | 798 | static __initconst const u64 atom_hw_cache_event_ids |
f22f54f4 PZ |
799 | [PERF_COUNT_HW_CACHE_MAX] |
800 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
801 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
802 | { | |
803 | [ C(L1D) ] = { | |
804 | [ C(OP_READ) ] = { | |
805 | [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */ | |
806 | [ C(RESULT_MISS) ] = 0, | |
807 | }, | |
808 | [ C(OP_WRITE) ] = { | |
809 | [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */ | |
810 | [ C(RESULT_MISS) ] = 0, | |
811 | }, | |
812 | [ C(OP_PREFETCH) ] = { | |
813 | [ C(RESULT_ACCESS) ] = 0x0, | |
814 | [ C(RESULT_MISS) ] = 0, | |
815 | }, | |
816 | }, | |
817 | [ C(L1I ) ] = { | |
818 | [ C(OP_READ) ] = { | |
819 | [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ | |
820 | [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ | |
821 | }, | |
822 | [ C(OP_WRITE) ] = { | |
823 | [ C(RESULT_ACCESS) ] = -1, | |
824 | [ C(RESULT_MISS) ] = -1, | |
825 | }, | |
826 | [ C(OP_PREFETCH) ] = { | |
827 | [ C(RESULT_ACCESS) ] = 0, | |
828 | [ C(RESULT_MISS) ] = 0, | |
829 | }, | |
830 | }, | |
831 | [ C(LL ) ] = { | |
832 | [ C(OP_READ) ] = { | |
833 | [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ | |
834 | [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ | |
835 | }, | |
836 | [ C(OP_WRITE) ] = { | |
837 | [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ | |
838 | [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ | |
839 | }, | |
840 | [ C(OP_PREFETCH) ] = { | |
841 | [ C(RESULT_ACCESS) ] = 0, | |
842 | [ C(RESULT_MISS) ] = 0, | |
843 | }, | |
844 | }, | |
845 | [ C(DTLB) ] = { | |
846 | [ C(OP_READ) ] = { | |
847 | [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */ | |
848 | [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */ | |
849 | }, | |
850 | [ C(OP_WRITE) ] = { | |
851 | [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */ | |
852 | [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */ | |
853 | }, | |
854 | [ C(OP_PREFETCH) ] = { | |
855 | [ C(RESULT_ACCESS) ] = 0, | |
856 | [ C(RESULT_MISS) ] = 0, | |
857 | }, | |
858 | }, | |
859 | [ C(ITLB) ] = { | |
860 | [ C(OP_READ) ] = { | |
861 | [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ | |
862 | [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */ | |
863 | }, | |
864 | [ C(OP_WRITE) ] = { | |
865 | [ C(RESULT_ACCESS) ] = -1, | |
866 | [ C(RESULT_MISS) ] = -1, | |
867 | }, | |
868 | [ C(OP_PREFETCH) ] = { | |
869 | [ C(RESULT_ACCESS) ] = -1, | |
870 | [ C(RESULT_MISS) ] = -1, | |
871 | }, | |
872 | }, | |
873 | [ C(BPU ) ] = { | |
874 | [ C(OP_READ) ] = { | |
875 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ | |
876 | [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ | |
877 | }, | |
878 | [ C(OP_WRITE) ] = { | |
879 | [ C(RESULT_ACCESS) ] = -1, | |
880 | [ C(RESULT_MISS) ] = -1, | |
881 | }, | |
882 | [ C(OP_PREFETCH) ] = { | |
883 | [ C(RESULT_ACCESS) ] = -1, | |
884 | [ C(RESULT_MISS) ] = -1, | |
885 | }, | |
886 | }, | |
887 | }; | |
888 | ||
60ce0fbd SE |
889 | static inline bool intel_pmu_needs_lbr_smpl(struct perf_event *event) |
890 | { | |
891 | /* user explicitly requested branch sampling */ | |
892 | if (has_branch_stack(event)) | |
893 | return true; | |
894 | ||
895 | /* implicit branch sampling to correct PEBS skid */ | |
3044318f AK |
896 | if (x86_pmu.intel_cap.pebs_trap && event->attr.precise_ip > 1 && |
897 | x86_pmu.intel_cap.pebs_format < 2) | |
60ce0fbd SE |
898 | return true; |
899 | ||
900 | return false; | |
901 | } | |
902 | ||
f22f54f4 PZ |
903 | static void intel_pmu_disable_all(void) |
904 | { | |
905 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | |
906 | ||
907 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); | |
908 | ||
15c7ad51 | 909 | if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) |
f22f54f4 | 910 | intel_pmu_disable_bts(); |
ca037701 PZ |
911 | |
912 | intel_pmu_pebs_disable_all(); | |
caff2bef | 913 | intel_pmu_lbr_disable_all(); |
f22f54f4 PZ |
914 | } |
915 | ||
11164cd4 | 916 | static void intel_pmu_enable_all(int added) |
f22f54f4 PZ |
917 | { |
918 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | |
919 | ||
d329527e PZ |
920 | intel_pmu_pebs_enable_all(); |
921 | intel_pmu_lbr_enable_all(); | |
144d31e6 GN |
922 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, |
923 | x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask); | |
f22f54f4 | 924 | |
15c7ad51 | 925 | if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) { |
f22f54f4 | 926 | struct perf_event *event = |
15c7ad51 | 927 | cpuc->events[INTEL_PMC_IDX_FIXED_BTS]; |
f22f54f4 PZ |
928 | |
929 | if (WARN_ON_ONCE(!event)) | |
930 | return; | |
931 | ||
932 | intel_pmu_enable_bts(event->hw.config); | |
933 | } | |
934 | } | |
935 | ||
11164cd4 PZ |
936 | /* |
937 | * Workaround for: | |
938 | * Intel Errata AAK100 (model 26) | |
939 | * Intel Errata AAP53 (model 30) | |
40b91cd1 | 940 | * Intel Errata BD53 (model 44) |
11164cd4 | 941 | * |
351af072 ZY |
942 | * The official story: |
943 | * These chips need to be 'reset' when adding counters by programming the | |
944 | * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either | |
945 | * in sequence on the same PMC or on different PMCs. | |
946 | * | |
947 | * In practise it appears some of these events do in fact count, and | |
948 | * we need to programm all 4 events. | |
11164cd4 | 949 | */ |
351af072 | 950 | static void intel_pmu_nhm_workaround(void) |
11164cd4 | 951 | { |
351af072 ZY |
952 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); |
953 | static const unsigned long nhm_magic[4] = { | |
954 | 0x4300B5, | |
955 | 0x4300D2, | |
956 | 0x4300B1, | |
957 | 0x4300B1 | |
958 | }; | |
959 | struct perf_event *event; | |
960 | int i; | |
11164cd4 | 961 | |
351af072 ZY |
962 | /* |
963 | * The Errata requires below steps: | |
964 | * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL; | |
965 | * 2) Configure 4 PERFEVTSELx with the magic events and clear | |
966 | * the corresponding PMCx; | |
967 | * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL; | |
968 | * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL; | |
969 | * 5) Clear 4 pairs of ERFEVTSELx and PMCx; | |
970 | */ | |
11164cd4 | 971 | |
351af072 ZY |
972 | /* |
973 | * The real steps we choose are a little different from above. | |
974 | * A) To reduce MSR operations, we don't run step 1) as they | |
975 | * are already cleared before this function is called; | |
976 | * B) Call x86_perf_event_update to save PMCx before configuring | |
977 | * PERFEVTSELx with magic number; | |
978 | * C) With step 5), we do clear only when the PERFEVTSELx is | |
979 | * not used currently. | |
980 | * D) Call x86_perf_event_set_period to restore PMCx; | |
981 | */ | |
11164cd4 | 982 | |
351af072 ZY |
983 | /* We always operate 4 pairs of PERF Counters */ |
984 | for (i = 0; i < 4; i++) { | |
985 | event = cpuc->events[i]; | |
986 | if (event) | |
987 | x86_perf_event_update(event); | |
988 | } | |
11164cd4 | 989 | |
351af072 ZY |
990 | for (i = 0; i < 4; i++) { |
991 | wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]); | |
992 | wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0); | |
993 | } | |
994 | ||
995 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf); | |
996 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0); | |
11164cd4 | 997 | |
351af072 ZY |
998 | for (i = 0; i < 4; i++) { |
999 | event = cpuc->events[i]; | |
1000 | ||
1001 | if (event) { | |
1002 | x86_perf_event_set_period(event); | |
31fa58af | 1003 | __x86_pmu_enable_event(&event->hw, |
351af072 ZY |
1004 | ARCH_PERFMON_EVENTSEL_ENABLE); |
1005 | } else | |
1006 | wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0); | |
11164cd4 | 1007 | } |
351af072 ZY |
1008 | } |
1009 | ||
1010 | static void intel_pmu_nhm_enable_all(int added) | |
1011 | { | |
1012 | if (added) | |
1013 | intel_pmu_nhm_workaround(); | |
11164cd4 PZ |
1014 | intel_pmu_enable_all(added); |
1015 | } | |
1016 | ||
f22f54f4 PZ |
1017 | static inline u64 intel_pmu_get_status(void) |
1018 | { | |
1019 | u64 status; | |
1020 | ||
1021 | rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status); | |
1022 | ||
1023 | return status; | |
1024 | } | |
1025 | ||
1026 | static inline void intel_pmu_ack_status(u64 ack) | |
1027 | { | |
1028 | wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack); | |
1029 | } | |
1030 | ||
ca037701 | 1031 | static void intel_pmu_disable_fixed(struct hw_perf_event *hwc) |
f22f54f4 | 1032 | { |
15c7ad51 | 1033 | int idx = hwc->idx - INTEL_PMC_IDX_FIXED; |
f22f54f4 PZ |
1034 | u64 ctrl_val, mask; |
1035 | ||
1036 | mask = 0xfULL << (idx * 4); | |
1037 | ||
1038 | rdmsrl(hwc->config_base, ctrl_val); | |
1039 | ctrl_val &= ~mask; | |
7645a24c | 1040 | wrmsrl(hwc->config_base, ctrl_val); |
f22f54f4 PZ |
1041 | } |
1042 | ||
ca037701 | 1043 | static void intel_pmu_disable_event(struct perf_event *event) |
f22f54f4 | 1044 | { |
aff3d91a | 1045 | struct hw_perf_event *hwc = &event->hw; |
144d31e6 | 1046 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); |
aff3d91a | 1047 | |
15c7ad51 | 1048 | if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) { |
f22f54f4 PZ |
1049 | intel_pmu_disable_bts(); |
1050 | intel_pmu_drain_bts_buffer(); | |
1051 | return; | |
1052 | } | |
1053 | ||
144d31e6 GN |
1054 | cpuc->intel_ctrl_guest_mask &= ~(1ull << hwc->idx); |
1055 | cpuc->intel_ctrl_host_mask &= ~(1ull << hwc->idx); | |
1056 | ||
60ce0fbd SE |
1057 | /* |
1058 | * must disable before any actual event | |
1059 | * because any event may be combined with LBR | |
1060 | */ | |
1061 | if (intel_pmu_needs_lbr_smpl(event)) | |
1062 | intel_pmu_lbr_disable(event); | |
1063 | ||
f22f54f4 | 1064 | if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { |
aff3d91a | 1065 | intel_pmu_disable_fixed(hwc); |
f22f54f4 PZ |
1066 | return; |
1067 | } | |
1068 | ||
aff3d91a | 1069 | x86_pmu_disable_event(event); |
ca037701 | 1070 | |
ab608344 | 1071 | if (unlikely(event->attr.precise_ip)) |
ef21f683 | 1072 | intel_pmu_pebs_disable(event); |
f22f54f4 PZ |
1073 | } |
1074 | ||
ca037701 | 1075 | static void intel_pmu_enable_fixed(struct hw_perf_event *hwc) |
f22f54f4 | 1076 | { |
15c7ad51 | 1077 | int idx = hwc->idx - INTEL_PMC_IDX_FIXED; |
f22f54f4 | 1078 | u64 ctrl_val, bits, mask; |
f22f54f4 PZ |
1079 | |
1080 | /* | |
1081 | * Enable IRQ generation (0x8), | |
1082 | * and enable ring-3 counting (0x2) and ring-0 counting (0x1) | |
1083 | * if requested: | |
1084 | */ | |
1085 | bits = 0x8ULL; | |
1086 | if (hwc->config & ARCH_PERFMON_EVENTSEL_USR) | |
1087 | bits |= 0x2; | |
1088 | if (hwc->config & ARCH_PERFMON_EVENTSEL_OS) | |
1089 | bits |= 0x1; | |
1090 | ||
1091 | /* | |
1092 | * ANY bit is supported in v3 and up | |
1093 | */ | |
1094 | if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY) | |
1095 | bits |= 0x4; | |
1096 | ||
1097 | bits <<= (idx * 4); | |
1098 | mask = 0xfULL << (idx * 4); | |
1099 | ||
1100 | rdmsrl(hwc->config_base, ctrl_val); | |
1101 | ctrl_val &= ~mask; | |
1102 | ctrl_val |= bits; | |
7645a24c | 1103 | wrmsrl(hwc->config_base, ctrl_val); |
f22f54f4 PZ |
1104 | } |
1105 | ||
aff3d91a | 1106 | static void intel_pmu_enable_event(struct perf_event *event) |
f22f54f4 | 1107 | { |
aff3d91a | 1108 | struct hw_perf_event *hwc = &event->hw; |
144d31e6 | 1109 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); |
aff3d91a | 1110 | |
15c7ad51 | 1111 | if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) { |
0a3aee0d | 1112 | if (!__this_cpu_read(cpu_hw_events.enabled)) |
f22f54f4 PZ |
1113 | return; |
1114 | ||
1115 | intel_pmu_enable_bts(hwc->config); | |
1116 | return; | |
1117 | } | |
60ce0fbd SE |
1118 | /* |
1119 | * must enabled before any actual event | |
1120 | * because any event may be combined with LBR | |
1121 | */ | |
1122 | if (intel_pmu_needs_lbr_smpl(event)) | |
1123 | intel_pmu_lbr_enable(event); | |
f22f54f4 | 1124 | |
144d31e6 GN |
1125 | if (event->attr.exclude_host) |
1126 | cpuc->intel_ctrl_guest_mask |= (1ull << hwc->idx); | |
1127 | if (event->attr.exclude_guest) | |
1128 | cpuc->intel_ctrl_host_mask |= (1ull << hwc->idx); | |
1129 | ||
f22f54f4 | 1130 | if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { |
aff3d91a | 1131 | intel_pmu_enable_fixed(hwc); |
f22f54f4 PZ |
1132 | return; |
1133 | } | |
1134 | ||
ab608344 | 1135 | if (unlikely(event->attr.precise_ip)) |
ef21f683 | 1136 | intel_pmu_pebs_enable(event); |
ca037701 | 1137 | |
31fa58af | 1138 | __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); |
f22f54f4 PZ |
1139 | } |
1140 | ||
1141 | /* | |
1142 | * Save and restart an expired event. Called by NMI contexts, | |
1143 | * so it has to be careful about preempting normal event ops: | |
1144 | */ | |
de0428a7 | 1145 | int intel_pmu_save_and_restart(struct perf_event *event) |
f22f54f4 | 1146 | { |
cc2ad4ba PZ |
1147 | x86_perf_event_update(event); |
1148 | return x86_perf_event_set_period(event); | |
f22f54f4 PZ |
1149 | } |
1150 | ||
1151 | static void intel_pmu_reset(void) | |
1152 | { | |
0a3aee0d | 1153 | struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds); |
f22f54f4 PZ |
1154 | unsigned long flags; |
1155 | int idx; | |
1156 | ||
948b1bb8 | 1157 | if (!x86_pmu.num_counters) |
f22f54f4 PZ |
1158 | return; |
1159 | ||
1160 | local_irq_save(flags); | |
1161 | ||
c767a54b | 1162 | pr_info("clearing PMU state on CPU#%d\n", smp_processor_id()); |
f22f54f4 | 1163 | |
948b1bb8 | 1164 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
715c85b1 PA |
1165 | wrmsrl_safe(x86_pmu_config_addr(idx), 0ull); |
1166 | wrmsrl_safe(x86_pmu_event_addr(idx), 0ull); | |
f22f54f4 | 1167 | } |
948b1bb8 | 1168 | for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) |
715c85b1 | 1169 | wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull); |
948b1bb8 | 1170 | |
f22f54f4 PZ |
1171 | if (ds) |
1172 | ds->bts_index = ds->bts_buffer_base; | |
1173 | ||
1174 | local_irq_restore(flags); | |
1175 | } | |
1176 | ||
1177 | /* | |
1178 | * This handler is triggered by the local APIC, so the APIC IRQ handling | |
1179 | * rules apply: | |
1180 | */ | |
1181 | static int intel_pmu_handle_irq(struct pt_regs *regs) | |
1182 | { | |
1183 | struct perf_sample_data data; | |
1184 | struct cpu_hw_events *cpuc; | |
1185 | int bit, loops; | |
2e556b5b | 1186 | u64 status; |
b0b2072d | 1187 | int handled; |
f22f54f4 | 1188 | |
f22f54f4 PZ |
1189 | cpuc = &__get_cpu_var(cpu_hw_events); |
1190 | ||
2bce5dac | 1191 | /* |
72db5596 AK |
1192 | * No known reason to not always do late ACK, |
1193 | * but just in case do it opt-in. | |
2bce5dac | 1194 | */ |
72db5596 AK |
1195 | if (!x86_pmu.late_ack) |
1196 | apic_write(APIC_LVTPC, APIC_DM_NMI); | |
3fb2b8dd | 1197 | intel_pmu_disable_all(); |
b0b2072d | 1198 | handled = intel_pmu_drain_bts_buffer(); |
f22f54f4 PZ |
1199 | status = intel_pmu_get_status(); |
1200 | if (!status) { | |
11164cd4 | 1201 | intel_pmu_enable_all(0); |
b0b2072d | 1202 | return handled; |
f22f54f4 PZ |
1203 | } |
1204 | ||
1205 | loops = 0; | |
1206 | again: | |
2e556b5b | 1207 | intel_pmu_ack_status(status); |
f22f54f4 | 1208 | if (++loops > 100) { |
ae0def05 DH |
1209 | static bool warned = false; |
1210 | if (!warned) { | |
1211 | WARN(1, "perfevents: irq loop stuck!\n"); | |
1212 | perf_event_print_debug(); | |
1213 | warned = true; | |
1214 | } | |
f22f54f4 | 1215 | intel_pmu_reset(); |
3fb2b8dd | 1216 | goto done; |
f22f54f4 PZ |
1217 | } |
1218 | ||
1219 | inc_irq_stat(apic_perf_irqs); | |
ca037701 | 1220 | |
caff2bef PZ |
1221 | intel_pmu_lbr_read(); |
1222 | ||
ca037701 PZ |
1223 | /* |
1224 | * PEBS overflow sets bit 62 in the global status register | |
1225 | */ | |
de725dec PZ |
1226 | if (__test_and_clear_bit(62, (unsigned long *)&status)) { |
1227 | handled++; | |
ca037701 | 1228 | x86_pmu.drain_pebs(regs); |
de725dec | 1229 | } |
ca037701 | 1230 | |
984b3f57 | 1231 | for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) { |
f22f54f4 PZ |
1232 | struct perf_event *event = cpuc->events[bit]; |
1233 | ||
de725dec PZ |
1234 | handled++; |
1235 | ||
f22f54f4 PZ |
1236 | if (!test_bit(bit, cpuc->active_mask)) |
1237 | continue; | |
1238 | ||
1239 | if (!intel_pmu_save_and_restart(event)) | |
1240 | continue; | |
1241 | ||
fd0d000b | 1242 | perf_sample_data_init(&data, 0, event->hw.last_period); |
f22f54f4 | 1243 | |
60ce0fbd SE |
1244 | if (has_branch_stack(event)) |
1245 | data.br_stack = &cpuc->lbr_stack; | |
1246 | ||
a8b0ca17 | 1247 | if (perf_event_overflow(event, &data, regs)) |
a4eaf7f1 | 1248 | x86_pmu_stop(event, 0); |
f22f54f4 PZ |
1249 | } |
1250 | ||
f22f54f4 PZ |
1251 | /* |
1252 | * Repeat if there is more work to be done: | |
1253 | */ | |
1254 | status = intel_pmu_get_status(); | |
1255 | if (status) | |
1256 | goto again; | |
1257 | ||
3fb2b8dd | 1258 | done: |
11164cd4 | 1259 | intel_pmu_enable_all(0); |
72db5596 AK |
1260 | /* |
1261 | * Only unmask the NMI after the overflow counters | |
1262 | * have been reset. This avoids spurious NMIs on | |
1263 | * Haswell CPUs. | |
1264 | */ | |
1265 | if (x86_pmu.late_ack) | |
1266 | apic_write(APIC_LVTPC, APIC_DM_NMI); | |
de725dec | 1267 | return handled; |
f22f54f4 PZ |
1268 | } |
1269 | ||
f22f54f4 | 1270 | static struct event_constraint * |
ca037701 | 1271 | intel_bts_constraints(struct perf_event *event) |
f22f54f4 | 1272 | { |
ca037701 PZ |
1273 | struct hw_perf_event *hwc = &event->hw; |
1274 | unsigned int hw_event, bts_event; | |
f22f54f4 | 1275 | |
18a073a3 PZ |
1276 | if (event->attr.freq) |
1277 | return NULL; | |
1278 | ||
ca037701 PZ |
1279 | hw_event = hwc->config & INTEL_ARCH_EVENT_MASK; |
1280 | bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS); | |
f22f54f4 | 1281 | |
ca037701 | 1282 | if (unlikely(hw_event == bts_event && hwc->sample_period == 1)) |
f22f54f4 | 1283 | return &bts_constraint; |
ca037701 | 1284 | |
f22f54f4 PZ |
1285 | return NULL; |
1286 | } | |
1287 | ||
5a425294 | 1288 | static int intel_alt_er(int idx) |
b79e8941 PZ |
1289 | { |
1290 | if (!(x86_pmu.er_flags & ERF_HAS_RSP_1)) | |
5a425294 | 1291 | return idx; |
b79e8941 | 1292 | |
5a425294 PZ |
1293 | if (idx == EXTRA_REG_RSP_0) |
1294 | return EXTRA_REG_RSP_1; | |
1295 | ||
1296 | if (idx == EXTRA_REG_RSP_1) | |
1297 | return EXTRA_REG_RSP_0; | |
1298 | ||
1299 | return idx; | |
1300 | } | |
1301 | ||
1302 | static void intel_fixup_er(struct perf_event *event, int idx) | |
1303 | { | |
1304 | event->hw.extra_reg.idx = idx; | |
1305 | ||
1306 | if (idx == EXTRA_REG_RSP_0) { | |
b79e8941 | 1307 | event->hw.config &= ~INTEL_ARCH_EVENT_MASK; |
53ad0447 | 1308 | event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_0].event; |
b79e8941 | 1309 | event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0; |
5a425294 PZ |
1310 | } else if (idx == EXTRA_REG_RSP_1) { |
1311 | event->hw.config &= ~INTEL_ARCH_EVENT_MASK; | |
53ad0447 | 1312 | event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_1].event; |
5a425294 | 1313 | event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1; |
b79e8941 | 1314 | } |
b79e8941 PZ |
1315 | } |
1316 | ||
efc9f05d SE |
1317 | /* |
1318 | * manage allocation of shared extra msr for certain events | |
1319 | * | |
1320 | * sharing can be: | |
1321 | * per-cpu: to be shared between the various events on a single PMU | |
1322 | * per-core: per-cpu + shared by HT threads | |
1323 | */ | |
a7e3ed1e | 1324 | static struct event_constraint * |
efc9f05d | 1325 | __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc, |
b36817e8 SE |
1326 | struct perf_event *event, |
1327 | struct hw_perf_event_extra *reg) | |
a7e3ed1e | 1328 | { |
efc9f05d | 1329 | struct event_constraint *c = &emptyconstraint; |
a7e3ed1e | 1330 | struct er_account *era; |
cd8a38d3 | 1331 | unsigned long flags; |
5a425294 | 1332 | int idx = reg->idx; |
a7e3ed1e | 1333 | |
5a425294 PZ |
1334 | /* |
1335 | * reg->alloc can be set due to existing state, so for fake cpuc we | |
1336 | * need to ignore this, otherwise we might fail to allocate proper fake | |
1337 | * state for this extra reg constraint. Also see the comment below. | |
1338 | */ | |
1339 | if (reg->alloc && !cpuc->is_fake) | |
b36817e8 | 1340 | return NULL; /* call x86_get_event_constraint() */ |
a7e3ed1e | 1341 | |
b79e8941 | 1342 | again: |
5a425294 | 1343 | era = &cpuc->shared_regs->regs[idx]; |
cd8a38d3 SE |
1344 | /* |
1345 | * we use spin_lock_irqsave() to avoid lockdep issues when | |
1346 | * passing a fake cpuc | |
1347 | */ | |
1348 | raw_spin_lock_irqsave(&era->lock, flags); | |
efc9f05d SE |
1349 | |
1350 | if (!atomic_read(&era->ref) || era->config == reg->config) { | |
1351 | ||
5a425294 PZ |
1352 | /* |
1353 | * If its a fake cpuc -- as per validate_{group,event}() we | |
1354 | * shouldn't touch event state and we can avoid doing so | |
1355 | * since both will only call get_event_constraints() once | |
1356 | * on each event, this avoids the need for reg->alloc. | |
1357 | * | |
1358 | * Not doing the ER fixup will only result in era->reg being | |
1359 | * wrong, but since we won't actually try and program hardware | |
1360 | * this isn't a problem either. | |
1361 | */ | |
1362 | if (!cpuc->is_fake) { | |
1363 | if (idx != reg->idx) | |
1364 | intel_fixup_er(event, idx); | |
1365 | ||
1366 | /* | |
1367 | * x86_schedule_events() can call get_event_constraints() | |
1368 | * multiple times on events in the case of incremental | |
1369 | * scheduling(). reg->alloc ensures we only do the ER | |
1370 | * allocation once. | |
1371 | */ | |
1372 | reg->alloc = 1; | |
1373 | } | |
1374 | ||
efc9f05d SE |
1375 | /* lock in msr value */ |
1376 | era->config = reg->config; | |
1377 | era->reg = reg->reg; | |
1378 | ||
1379 | /* one more user */ | |
1380 | atomic_inc(&era->ref); | |
1381 | ||
a7e3ed1e | 1382 | /* |
b36817e8 SE |
1383 | * need to call x86_get_event_constraint() |
1384 | * to check if associated event has constraints | |
a7e3ed1e | 1385 | */ |
b36817e8 | 1386 | c = NULL; |
5a425294 PZ |
1387 | } else { |
1388 | idx = intel_alt_er(idx); | |
1389 | if (idx != reg->idx) { | |
1390 | raw_spin_unlock_irqrestore(&era->lock, flags); | |
1391 | goto again; | |
1392 | } | |
a7e3ed1e | 1393 | } |
cd8a38d3 | 1394 | raw_spin_unlock_irqrestore(&era->lock, flags); |
a7e3ed1e | 1395 | |
efc9f05d SE |
1396 | return c; |
1397 | } | |
1398 | ||
1399 | static void | |
1400 | __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc, | |
1401 | struct hw_perf_event_extra *reg) | |
1402 | { | |
1403 | struct er_account *era; | |
1404 | ||
1405 | /* | |
5a425294 PZ |
1406 | * Only put constraint if extra reg was actually allocated. Also takes |
1407 | * care of event which do not use an extra shared reg. | |
1408 | * | |
1409 | * Also, if this is a fake cpuc we shouldn't touch any event state | |
1410 | * (reg->alloc) and we don't care about leaving inconsistent cpuc state | |
1411 | * either since it'll be thrown out. | |
efc9f05d | 1412 | */ |
5a425294 | 1413 | if (!reg->alloc || cpuc->is_fake) |
efc9f05d SE |
1414 | return; |
1415 | ||
1416 | era = &cpuc->shared_regs->regs[reg->idx]; | |
1417 | ||
1418 | /* one fewer user */ | |
1419 | atomic_dec(&era->ref); | |
1420 | ||
1421 | /* allocate again next time */ | |
1422 | reg->alloc = 0; | |
1423 | } | |
1424 | ||
1425 | static struct event_constraint * | |
1426 | intel_shared_regs_constraints(struct cpu_hw_events *cpuc, | |
1427 | struct perf_event *event) | |
1428 | { | |
b36817e8 SE |
1429 | struct event_constraint *c = NULL, *d; |
1430 | struct hw_perf_event_extra *xreg, *breg; | |
1431 | ||
1432 | xreg = &event->hw.extra_reg; | |
1433 | if (xreg->idx != EXTRA_REG_NONE) { | |
1434 | c = __intel_shared_reg_get_constraints(cpuc, event, xreg); | |
1435 | if (c == &emptyconstraint) | |
1436 | return c; | |
1437 | } | |
1438 | breg = &event->hw.branch_reg; | |
1439 | if (breg->idx != EXTRA_REG_NONE) { | |
1440 | d = __intel_shared_reg_get_constraints(cpuc, event, breg); | |
1441 | if (d == &emptyconstraint) { | |
1442 | __intel_shared_reg_put_constraints(cpuc, xreg); | |
1443 | c = d; | |
1444 | } | |
1445 | } | |
efc9f05d | 1446 | return c; |
a7e3ed1e AK |
1447 | } |
1448 | ||
de0428a7 KW |
1449 | struct event_constraint * |
1450 | x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event) | |
1451 | { | |
1452 | struct event_constraint *c; | |
1453 | ||
1454 | if (x86_pmu.event_constraints) { | |
1455 | for_each_event_constraint(c, x86_pmu.event_constraints) { | |
9fac2cf3 | 1456 | if ((event->hw.config & c->cmask) == c->code) { |
9fac2cf3 | 1457 | event->hw.flags |= c->flags; |
de0428a7 | 1458 | return c; |
9fac2cf3 | 1459 | } |
de0428a7 KW |
1460 | } |
1461 | } | |
1462 | ||
1463 | return &unconstrained; | |
1464 | } | |
1465 | ||
f22f54f4 PZ |
1466 | static struct event_constraint * |
1467 | intel_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event) | |
1468 | { | |
1469 | struct event_constraint *c; | |
1470 | ||
ca037701 PZ |
1471 | c = intel_bts_constraints(event); |
1472 | if (c) | |
1473 | return c; | |
1474 | ||
1475 | c = intel_pebs_constraints(event); | |
f22f54f4 PZ |
1476 | if (c) |
1477 | return c; | |
1478 | ||
efc9f05d | 1479 | c = intel_shared_regs_constraints(cpuc, event); |
a7e3ed1e AK |
1480 | if (c) |
1481 | return c; | |
1482 | ||
f22f54f4 PZ |
1483 | return x86_get_event_constraints(cpuc, event); |
1484 | } | |
1485 | ||
efc9f05d SE |
1486 | static void |
1487 | intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc, | |
a7e3ed1e AK |
1488 | struct perf_event *event) |
1489 | { | |
efc9f05d | 1490 | struct hw_perf_event_extra *reg; |
a7e3ed1e | 1491 | |
efc9f05d SE |
1492 | reg = &event->hw.extra_reg; |
1493 | if (reg->idx != EXTRA_REG_NONE) | |
1494 | __intel_shared_reg_put_constraints(cpuc, reg); | |
b36817e8 SE |
1495 | |
1496 | reg = &event->hw.branch_reg; | |
1497 | if (reg->idx != EXTRA_REG_NONE) | |
1498 | __intel_shared_reg_put_constraints(cpuc, reg); | |
efc9f05d | 1499 | } |
a7e3ed1e | 1500 | |
efc9f05d SE |
1501 | static void intel_put_event_constraints(struct cpu_hw_events *cpuc, |
1502 | struct perf_event *event) | |
1503 | { | |
1504 | intel_put_shared_regs_event_constraints(cpuc, event); | |
a7e3ed1e AK |
1505 | } |
1506 | ||
0780c927 | 1507 | static void intel_pebs_aliases_core2(struct perf_event *event) |
b4cdc5c2 | 1508 | { |
0780c927 | 1509 | if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { |
7639dae0 PZ |
1510 | /* |
1511 | * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P | |
1512 | * (0x003c) so that we can use it with PEBS. | |
1513 | * | |
1514 | * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't | |
1515 | * PEBS capable. However we can use INST_RETIRED.ANY_P | |
1516 | * (0x00c0), which is a PEBS capable event, to get the same | |
1517 | * count. | |
1518 | * | |
1519 | * INST_RETIRED.ANY_P counts the number of cycles that retires | |
1520 | * CNTMASK instructions. By setting CNTMASK to a value (16) | |
1521 | * larger than the maximum number of instructions that can be | |
1522 | * retired per cycle (4) and then inverting the condition, we | |
1523 | * count all cycles that retire 16 or less instructions, which | |
1524 | * is every cycle. | |
1525 | * | |
1526 | * Thereby we gain a PEBS capable cycle counter. | |
1527 | */ | |
f9b4eeb8 PZ |
1528 | u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16); |
1529 | ||
0780c927 PZ |
1530 | alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); |
1531 | event->hw.config = alt_config; | |
1532 | } | |
1533 | } | |
1534 | ||
1535 | static void intel_pebs_aliases_snb(struct perf_event *event) | |
1536 | { | |
1537 | if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { | |
1538 | /* | |
1539 | * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P | |
1540 | * (0x003c) so that we can use it with PEBS. | |
1541 | * | |
1542 | * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't | |
1543 | * PEBS capable. However we can use UOPS_RETIRED.ALL | |
1544 | * (0x01c2), which is a PEBS capable event, to get the same | |
1545 | * count. | |
1546 | * | |
1547 | * UOPS_RETIRED.ALL counts the number of cycles that retires | |
1548 | * CNTMASK micro-ops. By setting CNTMASK to a value (16) | |
1549 | * larger than the maximum number of micro-ops that can be | |
1550 | * retired per cycle (4) and then inverting the condition, we | |
1551 | * count all cycles that retire 16 or less micro-ops, which | |
1552 | * is every cycle. | |
1553 | * | |
1554 | * Thereby we gain a PEBS capable cycle counter. | |
1555 | */ | |
1556 | u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16); | |
7639dae0 PZ |
1557 | |
1558 | alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); | |
1559 | event->hw.config = alt_config; | |
1560 | } | |
0780c927 PZ |
1561 | } |
1562 | ||
1563 | static int intel_pmu_hw_config(struct perf_event *event) | |
1564 | { | |
1565 | int ret = x86_pmu_hw_config(event); | |
1566 | ||
1567 | if (ret) | |
1568 | return ret; | |
1569 | ||
1570 | if (event->attr.precise_ip && x86_pmu.pebs_aliases) | |
1571 | x86_pmu.pebs_aliases(event); | |
7639dae0 | 1572 | |
60ce0fbd SE |
1573 | if (intel_pmu_needs_lbr_smpl(event)) { |
1574 | ret = intel_pmu_setup_lbr_filter(event); | |
1575 | if (ret) | |
1576 | return ret; | |
1577 | } | |
1578 | ||
b4cdc5c2 PZ |
1579 | if (event->attr.type != PERF_TYPE_RAW) |
1580 | return 0; | |
1581 | ||
1582 | if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY)) | |
1583 | return 0; | |
1584 | ||
1585 | if (x86_pmu.version < 3) | |
1586 | return -EINVAL; | |
1587 | ||
1588 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) | |
1589 | return -EACCES; | |
1590 | ||
1591 | event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY; | |
1592 | ||
1593 | return 0; | |
1594 | } | |
1595 | ||
144d31e6 GN |
1596 | struct perf_guest_switch_msr *perf_guest_get_msrs(int *nr) |
1597 | { | |
1598 | if (x86_pmu.guest_get_msrs) | |
1599 | return x86_pmu.guest_get_msrs(nr); | |
1600 | *nr = 0; | |
1601 | return NULL; | |
1602 | } | |
1603 | EXPORT_SYMBOL_GPL(perf_guest_get_msrs); | |
1604 | ||
1605 | static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr) | |
1606 | { | |
1607 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | |
1608 | struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; | |
1609 | ||
1610 | arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL; | |
1611 | arr[0].host = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask; | |
1612 | arr[0].guest = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_host_mask; | |
26a4f3c0 GN |
1613 | /* |
1614 | * If PMU counter has PEBS enabled it is not enough to disable counter | |
1615 | * on a guest entry since PEBS memory write can overshoot guest entry | |
1616 | * and corrupt guest memory. Disabling PEBS solves the problem. | |
1617 | */ | |
1618 | arr[1].msr = MSR_IA32_PEBS_ENABLE; | |
1619 | arr[1].host = cpuc->pebs_enabled; | |
1620 | arr[1].guest = 0; | |
144d31e6 | 1621 | |
26a4f3c0 | 1622 | *nr = 2; |
144d31e6 GN |
1623 | return arr; |
1624 | } | |
1625 | ||
1626 | static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr) | |
1627 | { | |
1628 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | |
1629 | struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; | |
1630 | int idx; | |
1631 | ||
1632 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { | |
1633 | struct perf_event *event = cpuc->events[idx]; | |
1634 | ||
1635 | arr[idx].msr = x86_pmu_config_addr(idx); | |
1636 | arr[idx].host = arr[idx].guest = 0; | |
1637 | ||
1638 | if (!test_bit(idx, cpuc->active_mask)) | |
1639 | continue; | |
1640 | ||
1641 | arr[idx].host = arr[idx].guest = | |
1642 | event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE; | |
1643 | ||
1644 | if (event->attr.exclude_host) | |
1645 | arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE; | |
1646 | else if (event->attr.exclude_guest) | |
1647 | arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE; | |
1648 | } | |
1649 | ||
1650 | *nr = x86_pmu.num_counters; | |
1651 | return arr; | |
1652 | } | |
1653 | ||
1654 | static void core_pmu_enable_event(struct perf_event *event) | |
1655 | { | |
1656 | if (!event->attr.exclude_host) | |
1657 | x86_pmu_enable_event(event); | |
1658 | } | |
1659 | ||
1660 | static void core_pmu_enable_all(int added) | |
1661 | { | |
1662 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | |
1663 | int idx; | |
1664 | ||
1665 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { | |
1666 | struct hw_perf_event *hwc = &cpuc->events[idx]->hw; | |
1667 | ||
1668 | if (!test_bit(idx, cpuc->active_mask) || | |
1669 | cpuc->events[idx]->attr.exclude_host) | |
1670 | continue; | |
1671 | ||
1672 | __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); | |
1673 | } | |
1674 | } | |
1675 | ||
3a632cb2 AK |
1676 | static int hsw_hw_config(struct perf_event *event) |
1677 | { | |
1678 | int ret = intel_pmu_hw_config(event); | |
1679 | ||
1680 | if (ret) | |
1681 | return ret; | |
1682 | if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE)) | |
1683 | return 0; | |
1684 | event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED); | |
1685 | ||
1686 | /* | |
1687 | * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with | |
1688 | * PEBS or in ANY thread mode. Since the results are non-sensical forbid | |
1689 | * this combination. | |
1690 | */ | |
1691 | if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) && | |
1692 | ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) || | |
1693 | event->attr.precise_ip > 0)) | |
1694 | return -EOPNOTSUPP; | |
1695 | ||
1696 | return 0; | |
1697 | } | |
1698 | ||
1699 | static struct event_constraint counter2_constraint = | |
1700 | EVENT_CONSTRAINT(0, 0x4, 0); | |
1701 | ||
1702 | static struct event_constraint * | |
1703 | hsw_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event) | |
1704 | { | |
1705 | struct event_constraint *c = intel_get_event_constraints(cpuc, event); | |
1706 | ||
1707 | /* Handle special quirk on in_tx_checkpointed only in counter 2 */ | |
1708 | if (event->hw.config & HSW_IN_TX_CHECKPOINTED) { | |
1709 | if (c->idxmsk64 & (1U << 2)) | |
1710 | return &counter2_constraint; | |
1711 | return &emptyconstraint; | |
1712 | } | |
1713 | ||
1714 | return c; | |
1715 | } | |
1716 | ||
641cc938 JO |
1717 | PMU_FORMAT_ATTR(event, "config:0-7" ); |
1718 | PMU_FORMAT_ATTR(umask, "config:8-15" ); | |
1719 | PMU_FORMAT_ATTR(edge, "config:18" ); | |
1720 | PMU_FORMAT_ATTR(pc, "config:19" ); | |
1721 | PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */ | |
1722 | PMU_FORMAT_ATTR(inv, "config:23" ); | |
1723 | PMU_FORMAT_ATTR(cmask, "config:24-31" ); | |
3a632cb2 AK |
1724 | PMU_FORMAT_ATTR(in_tx, "config:32"); |
1725 | PMU_FORMAT_ATTR(in_tx_cp, "config:33"); | |
641cc938 JO |
1726 | |
1727 | static struct attribute *intel_arch_formats_attr[] = { | |
1728 | &format_attr_event.attr, | |
1729 | &format_attr_umask.attr, | |
1730 | &format_attr_edge.attr, | |
1731 | &format_attr_pc.attr, | |
1732 | &format_attr_inv.attr, | |
1733 | &format_attr_cmask.attr, | |
1734 | NULL, | |
1735 | }; | |
1736 | ||
0bf79d44 JO |
1737 | ssize_t intel_event_sysfs_show(char *page, u64 config) |
1738 | { | |
1739 | u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT); | |
1740 | ||
1741 | return x86_event_sysfs_show(page, config, event); | |
1742 | } | |
1743 | ||
caaa8be3 | 1744 | static __initconst const struct x86_pmu core_pmu = { |
f22f54f4 PZ |
1745 | .name = "core", |
1746 | .handle_irq = x86_pmu_handle_irq, | |
1747 | .disable_all = x86_pmu_disable_all, | |
144d31e6 GN |
1748 | .enable_all = core_pmu_enable_all, |
1749 | .enable = core_pmu_enable_event, | |
f22f54f4 | 1750 | .disable = x86_pmu_disable_event, |
b4cdc5c2 | 1751 | .hw_config = x86_pmu_hw_config, |
a072738e | 1752 | .schedule_events = x86_schedule_events, |
f22f54f4 PZ |
1753 | .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, |
1754 | .perfctr = MSR_ARCH_PERFMON_PERFCTR0, | |
1755 | .event_map = intel_pmu_event_map, | |
f22f54f4 PZ |
1756 | .max_events = ARRAY_SIZE(intel_perfmon_event_map), |
1757 | .apic = 1, | |
1758 | /* | |
1759 | * Intel PMCs cannot be accessed sanely above 32 bit width, | |
1760 | * so we install an artificial 1<<31 period regardless of | |
1761 | * the generic event period: | |
1762 | */ | |
1763 | .max_period = (1ULL << 31) - 1, | |
1764 | .get_event_constraints = intel_get_event_constraints, | |
a7e3ed1e | 1765 | .put_event_constraints = intel_put_event_constraints, |
f22f54f4 | 1766 | .event_constraints = intel_core_event_constraints, |
144d31e6 | 1767 | .guest_get_msrs = core_guest_get_msrs, |
641cc938 | 1768 | .format_attrs = intel_arch_formats_attr, |
0bf79d44 | 1769 | .events_sysfs_show = intel_event_sysfs_show, |
f22f54f4 PZ |
1770 | }; |
1771 | ||
de0428a7 | 1772 | struct intel_shared_regs *allocate_shared_regs(int cpu) |
efc9f05d SE |
1773 | { |
1774 | struct intel_shared_regs *regs; | |
1775 | int i; | |
1776 | ||
1777 | regs = kzalloc_node(sizeof(struct intel_shared_regs), | |
1778 | GFP_KERNEL, cpu_to_node(cpu)); | |
1779 | if (regs) { | |
1780 | /* | |
1781 | * initialize the locks to keep lockdep happy | |
1782 | */ | |
1783 | for (i = 0; i < EXTRA_REG_MAX; i++) | |
1784 | raw_spin_lock_init(®s->regs[i].lock); | |
1785 | ||
1786 | regs->core_id = -1; | |
1787 | } | |
1788 | return regs; | |
1789 | } | |
1790 | ||
a7e3ed1e AK |
1791 | static int intel_pmu_cpu_prepare(int cpu) |
1792 | { | |
1793 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); | |
1794 | ||
b36817e8 | 1795 | if (!(x86_pmu.extra_regs || x86_pmu.lbr_sel_map)) |
69092624 LM |
1796 | return NOTIFY_OK; |
1797 | ||
efc9f05d SE |
1798 | cpuc->shared_regs = allocate_shared_regs(cpu); |
1799 | if (!cpuc->shared_regs) | |
a7e3ed1e AK |
1800 | return NOTIFY_BAD; |
1801 | ||
a7e3ed1e AK |
1802 | return NOTIFY_OK; |
1803 | } | |
1804 | ||
74846d35 PZ |
1805 | static void intel_pmu_cpu_starting(int cpu) |
1806 | { | |
a7e3ed1e AK |
1807 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); |
1808 | int core_id = topology_core_id(cpu); | |
1809 | int i; | |
1810 | ||
69092624 LM |
1811 | init_debug_store_on_cpu(cpu); |
1812 | /* | |
1813 | * Deal with CPUs that don't clear their LBRs on power-up. | |
1814 | */ | |
1815 | intel_pmu_lbr_reset(); | |
1816 | ||
b36817e8 SE |
1817 | cpuc->lbr_sel = NULL; |
1818 | ||
1819 | if (!cpuc->shared_regs) | |
69092624 LM |
1820 | return; |
1821 | ||
b36817e8 SE |
1822 | if (!(x86_pmu.er_flags & ERF_NO_HT_SHARING)) { |
1823 | for_each_cpu(i, topology_thread_cpumask(cpu)) { | |
1824 | struct intel_shared_regs *pc; | |
a7e3ed1e | 1825 | |
b36817e8 SE |
1826 | pc = per_cpu(cpu_hw_events, i).shared_regs; |
1827 | if (pc && pc->core_id == core_id) { | |
1828 | cpuc->kfree_on_online = cpuc->shared_regs; | |
1829 | cpuc->shared_regs = pc; | |
1830 | break; | |
1831 | } | |
a7e3ed1e | 1832 | } |
b36817e8 SE |
1833 | cpuc->shared_regs->core_id = core_id; |
1834 | cpuc->shared_regs->refcnt++; | |
a7e3ed1e AK |
1835 | } |
1836 | ||
b36817e8 SE |
1837 | if (x86_pmu.lbr_sel_map) |
1838 | cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR]; | |
74846d35 PZ |
1839 | } |
1840 | ||
1841 | static void intel_pmu_cpu_dying(int cpu) | |
1842 | { | |
a7e3ed1e | 1843 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); |
efc9f05d | 1844 | struct intel_shared_regs *pc; |
a7e3ed1e | 1845 | |
efc9f05d | 1846 | pc = cpuc->shared_regs; |
a7e3ed1e AK |
1847 | if (pc) { |
1848 | if (pc->core_id == -1 || --pc->refcnt == 0) | |
1849 | kfree(pc); | |
efc9f05d | 1850 | cpuc->shared_regs = NULL; |
a7e3ed1e AK |
1851 | } |
1852 | ||
74846d35 PZ |
1853 | fini_debug_store_on_cpu(cpu); |
1854 | } | |
1855 | ||
d010b332 SE |
1856 | static void intel_pmu_flush_branch_stack(void) |
1857 | { | |
1858 | /* | |
1859 | * Intel LBR does not tag entries with the | |
1860 | * PID of the current task, then we need to | |
1861 | * flush it on ctxsw | |
1862 | * For now, we simply reset it | |
1863 | */ | |
1864 | if (x86_pmu.lbr_nr) | |
1865 | intel_pmu_lbr_reset(); | |
1866 | } | |
1867 | ||
641cc938 JO |
1868 | PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63"); |
1869 | ||
a63fcab4 SE |
1870 | PMU_FORMAT_ATTR(ldlat, "config1:0-15"); |
1871 | ||
641cc938 JO |
1872 | static struct attribute *intel_arch3_formats_attr[] = { |
1873 | &format_attr_event.attr, | |
1874 | &format_attr_umask.attr, | |
1875 | &format_attr_edge.attr, | |
1876 | &format_attr_pc.attr, | |
1877 | &format_attr_any.attr, | |
1878 | &format_attr_inv.attr, | |
1879 | &format_attr_cmask.attr, | |
3a632cb2 AK |
1880 | &format_attr_in_tx.attr, |
1881 | &format_attr_in_tx_cp.attr, | |
641cc938 JO |
1882 | |
1883 | &format_attr_offcore_rsp.attr, /* XXX do NHM/WSM + SNB breakout */ | |
a63fcab4 | 1884 | &format_attr_ldlat.attr, /* PEBS load latency */ |
641cc938 JO |
1885 | NULL, |
1886 | }; | |
1887 | ||
caaa8be3 | 1888 | static __initconst const struct x86_pmu intel_pmu = { |
f22f54f4 PZ |
1889 | .name = "Intel", |
1890 | .handle_irq = intel_pmu_handle_irq, | |
1891 | .disable_all = intel_pmu_disable_all, | |
1892 | .enable_all = intel_pmu_enable_all, | |
1893 | .enable = intel_pmu_enable_event, | |
1894 | .disable = intel_pmu_disable_event, | |
b4cdc5c2 | 1895 | .hw_config = intel_pmu_hw_config, |
a072738e | 1896 | .schedule_events = x86_schedule_events, |
f22f54f4 PZ |
1897 | .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, |
1898 | .perfctr = MSR_ARCH_PERFMON_PERFCTR0, | |
1899 | .event_map = intel_pmu_event_map, | |
f22f54f4 PZ |
1900 | .max_events = ARRAY_SIZE(intel_perfmon_event_map), |
1901 | .apic = 1, | |
1902 | /* | |
1903 | * Intel PMCs cannot be accessed sanely above 32 bit width, | |
1904 | * so we install an artificial 1<<31 period regardless of | |
1905 | * the generic event period: | |
1906 | */ | |
1907 | .max_period = (1ULL << 31) - 1, | |
3f6da390 | 1908 | .get_event_constraints = intel_get_event_constraints, |
a7e3ed1e | 1909 | .put_event_constraints = intel_put_event_constraints, |
0780c927 | 1910 | .pebs_aliases = intel_pebs_aliases_core2, |
3f6da390 | 1911 | |
641cc938 | 1912 | .format_attrs = intel_arch3_formats_attr, |
0bf79d44 | 1913 | .events_sysfs_show = intel_event_sysfs_show, |
641cc938 | 1914 | |
a7e3ed1e | 1915 | .cpu_prepare = intel_pmu_cpu_prepare, |
74846d35 PZ |
1916 | .cpu_starting = intel_pmu_cpu_starting, |
1917 | .cpu_dying = intel_pmu_cpu_dying, | |
144d31e6 | 1918 | .guest_get_msrs = intel_guest_get_msrs, |
d010b332 | 1919 | .flush_branch_stack = intel_pmu_flush_branch_stack, |
f22f54f4 PZ |
1920 | }; |
1921 | ||
c1d6f42f | 1922 | static __init void intel_clovertown_quirk(void) |
3c44780b PZ |
1923 | { |
1924 | /* | |
1925 | * PEBS is unreliable due to: | |
1926 | * | |
1927 | * AJ67 - PEBS may experience CPL leaks | |
1928 | * AJ68 - PEBS PMI may be delayed by one event | |
1929 | * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12] | |
1930 | * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS | |
1931 | * | |
1932 | * AJ67 could be worked around by restricting the OS/USR flags. | |
1933 | * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI. | |
1934 | * | |
1935 | * AJ106 could possibly be worked around by not allowing LBR | |
1936 | * usage from PEBS, including the fixup. | |
1937 | * AJ68 could possibly be worked around by always programming | |
ec75a716 | 1938 | * a pebs_event_reset[0] value and coping with the lost events. |
3c44780b PZ |
1939 | * |
1940 | * But taken together it might just make sense to not enable PEBS on | |
1941 | * these chips. | |
1942 | */ | |
c767a54b | 1943 | pr_warn("PEBS disabled due to CPU errata\n"); |
3c44780b PZ |
1944 | x86_pmu.pebs = 0; |
1945 | x86_pmu.pebs_constraints = NULL; | |
1946 | } | |
1947 | ||
c93dc84c PZ |
1948 | static int intel_snb_pebs_broken(int cpu) |
1949 | { | |
1950 | u32 rev = UINT_MAX; /* default to broken for unknown models */ | |
1951 | ||
1952 | switch (cpu_data(cpu).x86_model) { | |
1953 | case 42: /* SNB */ | |
1954 | rev = 0x28; | |
1955 | break; | |
1956 | ||
1957 | case 45: /* SNB-EP */ | |
1958 | switch (cpu_data(cpu).x86_mask) { | |
1959 | case 6: rev = 0x618; break; | |
1960 | case 7: rev = 0x70c; break; | |
1961 | } | |
1962 | } | |
1963 | ||
1964 | return (cpu_data(cpu).microcode < rev); | |
1965 | } | |
1966 | ||
1967 | static void intel_snb_check_microcode(void) | |
1968 | { | |
1969 | int pebs_broken = 0; | |
1970 | int cpu; | |
1971 | ||
1972 | get_online_cpus(); | |
1973 | for_each_online_cpu(cpu) { | |
1974 | if ((pebs_broken = intel_snb_pebs_broken(cpu))) | |
1975 | break; | |
1976 | } | |
1977 | put_online_cpus(); | |
1978 | ||
1979 | if (pebs_broken == x86_pmu.pebs_broken) | |
1980 | return; | |
1981 | ||
1982 | /* | |
1983 | * Serialized by the microcode lock.. | |
1984 | */ | |
1985 | if (x86_pmu.pebs_broken) { | |
1986 | pr_info("PEBS enabled due to microcode update\n"); | |
1987 | x86_pmu.pebs_broken = 0; | |
1988 | } else { | |
1989 | pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n"); | |
1990 | x86_pmu.pebs_broken = 1; | |
1991 | } | |
1992 | } | |
1993 | ||
c1d6f42f | 1994 | static __init void intel_sandybridge_quirk(void) |
6a600a8b | 1995 | { |
c93dc84c PZ |
1996 | x86_pmu.check_microcode = intel_snb_check_microcode; |
1997 | intel_snb_check_microcode(); | |
6a600a8b PZ |
1998 | } |
1999 | ||
c1d6f42f PZ |
2000 | static const struct { int id; char *name; } intel_arch_events_map[] __initconst = { |
2001 | { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" }, | |
2002 | { PERF_COUNT_HW_INSTRUCTIONS, "instructions" }, | |
2003 | { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" }, | |
2004 | { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" }, | |
2005 | { PERF_COUNT_HW_CACHE_MISSES, "cache misses" }, | |
2006 | { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" }, | |
2007 | { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" }, | |
ffb871bc GN |
2008 | }; |
2009 | ||
c1d6f42f PZ |
2010 | static __init void intel_arch_events_quirk(void) |
2011 | { | |
2012 | int bit; | |
2013 | ||
2014 | /* disable event that reported as not presend by cpuid */ | |
2015 | for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) { | |
2016 | intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0; | |
c767a54b JP |
2017 | pr_warn("CPUID marked event: \'%s\' unavailable\n", |
2018 | intel_arch_events_map[bit].name); | |
c1d6f42f PZ |
2019 | } |
2020 | } | |
2021 | ||
2022 | static __init void intel_nehalem_quirk(void) | |
2023 | { | |
2024 | union cpuid10_ebx ebx; | |
2025 | ||
2026 | ebx.full = x86_pmu.events_maskl; | |
2027 | if (ebx.split.no_branch_misses_retired) { | |
2028 | /* | |
2029 | * Erratum AAJ80 detected, we work it around by using | |
2030 | * the BR_MISP_EXEC.ANY event. This will over-count | |
2031 | * branch-misses, but it's still much better than the | |
2032 | * architectural event which is often completely bogus: | |
2033 | */ | |
2034 | intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89; | |
2035 | ebx.split.no_branch_misses_retired = 0; | |
2036 | x86_pmu.events_maskl = ebx.full; | |
c767a54b | 2037 | pr_info("CPU erratum AAJ80 worked around\n"); |
c1d6f42f PZ |
2038 | } |
2039 | } | |
2040 | ||
f9134f36 AK |
2041 | EVENT_ATTR_STR(mem-loads, mem_ld_hsw, "event=0xcd,umask=0x1,ldlat=3"); |
2042 | EVENT_ATTR_STR(mem-stores, mem_st_hsw, "event=0xd0,umask=0x82") | |
2043 | ||
2044 | static struct attribute *hsw_events_attrs[] = { | |
2045 | EVENT_PTR(mem_ld_hsw), | |
2046 | EVENT_PTR(mem_st_hsw), | |
2047 | NULL | |
2048 | }; | |
2049 | ||
de0428a7 | 2050 | __init int intel_pmu_init(void) |
f22f54f4 PZ |
2051 | { |
2052 | union cpuid10_edx edx; | |
2053 | union cpuid10_eax eax; | |
ffb871bc | 2054 | union cpuid10_ebx ebx; |
a1eac7ac | 2055 | struct event_constraint *c; |
f22f54f4 | 2056 | unsigned int unused; |
f22f54f4 PZ |
2057 | int version; |
2058 | ||
2059 | if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) { | |
a072738e CG |
2060 | switch (boot_cpu_data.x86) { |
2061 | case 0x6: | |
2062 | return p6_pmu_init(); | |
e717bf4e VW |
2063 | case 0xb: |
2064 | return knc_pmu_init(); | |
a072738e CG |
2065 | case 0xf: |
2066 | return p4_pmu_init(); | |
2067 | } | |
f22f54f4 | 2068 | return -ENODEV; |
f22f54f4 PZ |
2069 | } |
2070 | ||
2071 | /* | |
2072 | * Check whether the Architectural PerfMon supports | |
2073 | * Branch Misses Retired hw_event or not. | |
2074 | */ | |
ffb871bc GN |
2075 | cpuid(10, &eax.full, &ebx.full, &unused, &edx.full); |
2076 | if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT) | |
f22f54f4 PZ |
2077 | return -ENODEV; |
2078 | ||
2079 | version = eax.split.version_id; | |
2080 | if (version < 2) | |
2081 | x86_pmu = core_pmu; | |
2082 | else | |
2083 | x86_pmu = intel_pmu; | |
2084 | ||
2085 | x86_pmu.version = version; | |
948b1bb8 RR |
2086 | x86_pmu.num_counters = eax.split.num_counters; |
2087 | x86_pmu.cntval_bits = eax.split.bit_width; | |
2088 | x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1; | |
f22f54f4 | 2089 | |
c1d6f42f PZ |
2090 | x86_pmu.events_maskl = ebx.full; |
2091 | x86_pmu.events_mask_len = eax.split.mask_length; | |
2092 | ||
70ab7003 AK |
2093 | x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters); |
2094 | ||
f22f54f4 PZ |
2095 | /* |
2096 | * Quirk: v2 perfmon does not report fixed-purpose events, so | |
2097 | * assume at least 3 events: | |
2098 | */ | |
2099 | if (version > 1) | |
948b1bb8 | 2100 | x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3); |
f22f54f4 | 2101 | |
8db909a7 PZ |
2102 | /* |
2103 | * v2 and above have a perf capabilities MSR | |
2104 | */ | |
2105 | if (version > 1) { | |
2106 | u64 capabilities; | |
2107 | ||
2108 | rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities); | |
2109 | x86_pmu.intel_cap.capabilities = capabilities; | |
2110 | } | |
2111 | ||
ca037701 PZ |
2112 | intel_ds_init(); |
2113 | ||
c1d6f42f PZ |
2114 | x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */ |
2115 | ||
f22f54f4 PZ |
2116 | /* |
2117 | * Install the hw-cache-events table: | |
2118 | */ | |
2119 | switch (boot_cpu_data.x86_model) { | |
2120 | case 14: /* 65 nm core solo/duo, "Yonah" */ | |
2121 | pr_cont("Core events, "); | |
2122 | break; | |
2123 | ||
2124 | case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */ | |
c1d6f42f | 2125 | x86_add_quirk(intel_clovertown_quirk); |
f22f54f4 PZ |
2126 | case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */ |
2127 | case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */ | |
2128 | case 29: /* six-core 45 nm xeon "Dunnington" */ | |
2129 | memcpy(hw_cache_event_ids, core2_hw_cache_event_ids, | |
2130 | sizeof(hw_cache_event_ids)); | |
2131 | ||
caff2bef PZ |
2132 | intel_pmu_lbr_init_core(); |
2133 | ||
f22f54f4 | 2134 | x86_pmu.event_constraints = intel_core2_event_constraints; |
17e31629 | 2135 | x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints; |
f22f54f4 PZ |
2136 | pr_cont("Core2 events, "); |
2137 | break; | |
2138 | ||
2139 | case 26: /* 45 nm nehalem, "Bloomfield" */ | |
2140 | case 30: /* 45 nm nehalem, "Lynnfield" */ | |
134fbadf | 2141 | case 46: /* 45 nm nehalem-ex, "Beckton" */ |
f22f54f4 PZ |
2142 | memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids, |
2143 | sizeof(hw_cache_event_ids)); | |
e994d7d2 AK |
2144 | memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, |
2145 | sizeof(hw_cache_extra_regs)); | |
f22f54f4 | 2146 | |
caff2bef PZ |
2147 | intel_pmu_lbr_init_nhm(); |
2148 | ||
f22f54f4 | 2149 | x86_pmu.event_constraints = intel_nehalem_event_constraints; |
17e31629 | 2150 | x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints; |
11164cd4 | 2151 | x86_pmu.enable_all = intel_pmu_nhm_enable_all; |
a7e3ed1e | 2152 | x86_pmu.extra_regs = intel_nehalem_extra_regs; |
ec75a716 | 2153 | |
f20093ee SE |
2154 | x86_pmu.cpu_events = nhm_events_attrs; |
2155 | ||
91fc4cc0 | 2156 | /* UOPS_ISSUED.STALLED_CYCLES */ |
f9b4eeb8 PZ |
2157 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = |
2158 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | |
91fc4cc0 | 2159 | /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ |
f9b4eeb8 PZ |
2160 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = |
2161 | X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); | |
94403f88 | 2162 | |
c1d6f42f | 2163 | x86_add_quirk(intel_nehalem_quirk); |
ec75a716 | 2164 | |
11164cd4 | 2165 | pr_cont("Nehalem events, "); |
f22f54f4 | 2166 | break; |
caff2bef | 2167 | |
b622d644 | 2168 | case 28: /* Atom */ |
0927b482 SL |
2169 | case 38: /* Lincroft */ |
2170 | case 39: /* Penwell */ | |
2171 | case 53: /* Cloverview */ | |
2172 | case 54: /* Cedarview */ | |
f22f54f4 PZ |
2173 | memcpy(hw_cache_event_ids, atom_hw_cache_event_ids, |
2174 | sizeof(hw_cache_event_ids)); | |
2175 | ||
caff2bef PZ |
2176 | intel_pmu_lbr_init_atom(); |
2177 | ||
f22f54f4 | 2178 | x86_pmu.event_constraints = intel_gen_event_constraints; |
17e31629 | 2179 | x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints; |
f22f54f4 PZ |
2180 | pr_cont("Atom events, "); |
2181 | break; | |
2182 | ||
2183 | case 37: /* 32 nm nehalem, "Clarkdale" */ | |
2184 | case 44: /* 32 nm nehalem, "Gulftown" */ | |
b2508e82 | 2185 | case 47: /* 32 nm Xeon E7 */ |
f22f54f4 PZ |
2186 | memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids, |
2187 | sizeof(hw_cache_event_ids)); | |
e994d7d2 AK |
2188 | memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, |
2189 | sizeof(hw_cache_extra_regs)); | |
f22f54f4 | 2190 | |
caff2bef PZ |
2191 | intel_pmu_lbr_init_nhm(); |
2192 | ||
f22f54f4 | 2193 | x86_pmu.event_constraints = intel_westmere_event_constraints; |
40b91cd1 | 2194 | x86_pmu.enable_all = intel_pmu_nhm_enable_all; |
17e31629 | 2195 | x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints; |
a7e3ed1e | 2196 | x86_pmu.extra_regs = intel_westmere_extra_regs; |
b79e8941 | 2197 | x86_pmu.er_flags |= ERF_HAS_RSP_1; |
30112039 | 2198 | |
f20093ee SE |
2199 | x86_pmu.cpu_events = nhm_events_attrs; |
2200 | ||
30112039 | 2201 | /* UOPS_ISSUED.STALLED_CYCLES */ |
f9b4eeb8 PZ |
2202 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = |
2203 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | |
30112039 | 2204 | /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ |
f9b4eeb8 PZ |
2205 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = |
2206 | X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); | |
30112039 | 2207 | |
f22f54f4 PZ |
2208 | pr_cont("Westmere events, "); |
2209 | break; | |
b622d644 | 2210 | |
b06b3d49 | 2211 | case 42: /* SandyBridge */ |
a34668f6 | 2212 | case 45: /* SandyBridge, "Romely-EP" */ |
47a8863d | 2213 | x86_add_quirk(intel_sandybridge_quirk); |
b06b3d49 LM |
2214 | memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, |
2215 | sizeof(hw_cache_event_ids)); | |
74e6543f YZ |
2216 | memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, |
2217 | sizeof(hw_cache_extra_regs)); | |
b06b3d49 | 2218 | |
c5cc2cd9 | 2219 | intel_pmu_lbr_init_snb(); |
b06b3d49 LM |
2220 | |
2221 | x86_pmu.event_constraints = intel_snb_event_constraints; | |
de0428a7 | 2222 | x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints; |
0780c927 | 2223 | x86_pmu.pebs_aliases = intel_pebs_aliases_snb; |
f1923820 SE |
2224 | if (boot_cpu_data.x86_model == 45) |
2225 | x86_pmu.extra_regs = intel_snbep_extra_regs; | |
2226 | else | |
2227 | x86_pmu.extra_regs = intel_snb_extra_regs; | |
ee89cbc2 | 2228 | /* all extra regs are per-cpu when HT is on */ |
b79e8941 PZ |
2229 | x86_pmu.er_flags |= ERF_HAS_RSP_1; |
2230 | x86_pmu.er_flags |= ERF_NO_HT_SHARING; | |
e04d1b23 | 2231 | |
f20093ee SE |
2232 | x86_pmu.cpu_events = snb_events_attrs; |
2233 | ||
e04d1b23 | 2234 | /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ |
f9b4eeb8 PZ |
2235 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = |
2236 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | |
e04d1b23 | 2237 | /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/ |
f9b4eeb8 PZ |
2238 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = |
2239 | X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1); | |
e04d1b23 | 2240 | |
b06b3d49 LM |
2241 | pr_cont("SandyBridge events, "); |
2242 | break; | |
20a36e39 | 2243 | case 58: /* IvyBridge */ |
923d8697 | 2244 | case 62: /* IvyBridge EP */ |
20a36e39 SE |
2245 | memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, |
2246 | sizeof(hw_cache_event_ids)); | |
2247 | memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, | |
2248 | sizeof(hw_cache_extra_regs)); | |
2249 | ||
2250 | intel_pmu_lbr_init_snb(); | |
2251 | ||
69943182 | 2252 | x86_pmu.event_constraints = intel_ivb_event_constraints; |
20a36e39 SE |
2253 | x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints; |
2254 | x86_pmu.pebs_aliases = intel_pebs_aliases_snb; | |
f1923820 SE |
2255 | if (boot_cpu_data.x86_model == 62) |
2256 | x86_pmu.extra_regs = intel_snbep_extra_regs; | |
2257 | else | |
2258 | x86_pmu.extra_regs = intel_snb_extra_regs; | |
20a36e39 SE |
2259 | /* all extra regs are per-cpu when HT is on */ |
2260 | x86_pmu.er_flags |= ERF_HAS_RSP_1; | |
2261 | x86_pmu.er_flags |= ERF_NO_HT_SHARING; | |
2262 | ||
f20093ee SE |
2263 | x86_pmu.cpu_events = snb_events_attrs; |
2264 | ||
20a36e39 SE |
2265 | /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ |
2266 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = | |
2267 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | |
2268 | ||
2269 | pr_cont("IvyBridge events, "); | |
2270 | break; | |
2271 | ||
b06b3d49 | 2272 | |
3a632cb2 AK |
2273 | case 60: /* Haswell Client */ |
2274 | case 70: | |
2275 | case 71: | |
2276 | case 63: | |
0499bd86 | 2277 | case 69: |
72db5596 | 2278 | x86_pmu.late_ack = true; |
3a632cb2 AK |
2279 | memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, sizeof(hw_cache_event_ids)); |
2280 | memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); | |
2281 | ||
2282 | intel_pmu_lbr_init_snb(); | |
2283 | ||
2284 | x86_pmu.event_constraints = intel_hsw_event_constraints; | |
3044318f | 2285 | x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints; |
3a632cb2 | 2286 | x86_pmu.extra_regs = intel_snb_extra_regs; |
3044318f | 2287 | x86_pmu.pebs_aliases = intel_pebs_aliases_snb; |
3a632cb2 AK |
2288 | /* all extra regs are per-cpu when HT is on */ |
2289 | x86_pmu.er_flags |= ERF_HAS_RSP_1; | |
2290 | x86_pmu.er_flags |= ERF_NO_HT_SHARING; | |
2291 | ||
2292 | x86_pmu.hw_config = hsw_hw_config; | |
2293 | x86_pmu.get_event_constraints = hsw_get_event_constraints; | |
f9134f36 | 2294 | x86_pmu.cpu_events = hsw_events_attrs; |
3a632cb2 AK |
2295 | pr_cont("Haswell events, "); |
2296 | break; | |
2297 | ||
f22f54f4 | 2298 | default: |
0af3ac1f AK |
2299 | switch (x86_pmu.version) { |
2300 | case 1: | |
2301 | x86_pmu.event_constraints = intel_v1_event_constraints; | |
2302 | pr_cont("generic architected perfmon v1, "); | |
2303 | break; | |
2304 | default: | |
2305 | /* | |
2306 | * default constraints for v2 and up | |
2307 | */ | |
2308 | x86_pmu.event_constraints = intel_gen_event_constraints; | |
2309 | pr_cont("generic architected perfmon, "); | |
2310 | break; | |
2311 | } | |
f22f54f4 | 2312 | } |
ffb871bc | 2313 | |
a1eac7ac RR |
2314 | if (x86_pmu.num_counters > INTEL_PMC_MAX_GENERIC) { |
2315 | WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!", | |
2316 | x86_pmu.num_counters, INTEL_PMC_MAX_GENERIC); | |
2317 | x86_pmu.num_counters = INTEL_PMC_MAX_GENERIC; | |
2318 | } | |
2319 | x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1; | |
2320 | ||
2321 | if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) { | |
2322 | WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!", | |
2323 | x86_pmu.num_counters_fixed, INTEL_PMC_MAX_FIXED); | |
2324 | x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED; | |
2325 | } | |
2326 | ||
2327 | x86_pmu.intel_ctrl |= | |
2328 | ((1LL << x86_pmu.num_counters_fixed)-1) << INTEL_PMC_IDX_FIXED; | |
2329 | ||
2330 | if (x86_pmu.event_constraints) { | |
2331 | /* | |
2332 | * event on fixed counter2 (REF_CYCLES) only works on this | |
2333 | * counter, so do not extend mask to generic counters | |
2334 | */ | |
2335 | for_each_event_constraint(c, x86_pmu.event_constraints) { | |
3a632cb2 | 2336 | if (c->cmask != FIXED_EVENT_FLAGS |
a1eac7ac RR |
2337 | || c->idxmsk64 == INTEL_PMC_MSK_FIXED_REF_CYCLES) { |
2338 | continue; | |
2339 | } | |
2340 | ||
2341 | c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1; | |
2342 | c->weight += x86_pmu.num_counters; | |
2343 | } | |
2344 | } | |
2345 | ||
069e0c3c AK |
2346 | /* Support full width counters using alternative MSR range */ |
2347 | if (x86_pmu.intel_cap.full_width_write) { | |
2348 | x86_pmu.max_period = x86_pmu.cntval_mask; | |
2349 | x86_pmu.perfctr = MSR_IA32_PMC0; | |
2350 | pr_cont("full-width counters, "); | |
2351 | } | |
2352 | ||
f22f54f4 PZ |
2353 | return 0; |
2354 | } |