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 */ | |
93fcf72c MD |
116 | |
117 | INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ | |
118 | INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ | |
119 | INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ | |
120 | INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ | |
121 | ||
b06b3d49 LM |
122 | EVENT_CONSTRAINT_END |
123 | }; | |
124 | ||
69943182 SE |
125 | static struct event_constraint intel_ivb_event_constraints[] __read_mostly = |
126 | { | |
127 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | |
128 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
129 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ | |
130 | INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */ | |
131 | INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMTPY */ | |
132 | INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */ | |
6113af14 | 133 | INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_LDM_PENDING */ |
69943182 SE |
134 | INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */ |
135 | INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */ | |
136 | INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */ | |
137 | INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ | |
138 | INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ | |
139 | INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ | |
93fcf72c MD |
140 | |
141 | INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ | |
142 | INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ | |
143 | INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ | |
144 | INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ | |
145 | ||
69943182 SE |
146 | EVENT_CONSTRAINT_END |
147 | }; | |
148 | ||
5c543e3c | 149 | static struct extra_reg intel_westmere_extra_regs[] __read_mostly = |
a7e3ed1e | 150 | { |
53ad0447 YZ |
151 | /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ |
152 | INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), | |
153 | INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1), | |
f20093ee | 154 | INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b), |
a7e3ed1e AK |
155 | EVENT_EXTRA_END |
156 | }; | |
157 | ||
0af3ac1f AK |
158 | static struct event_constraint intel_v1_event_constraints[] __read_mostly = |
159 | { | |
160 | EVENT_CONSTRAINT_END | |
161 | }; | |
162 | ||
5c543e3c | 163 | static struct event_constraint intel_gen_event_constraints[] __read_mostly = |
f22f54f4 | 164 | { |
b622d644 PZ |
165 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
166 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
cd09c0c4 | 167 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ |
f22f54f4 PZ |
168 | EVENT_CONSTRAINT_END |
169 | }; | |
170 | ||
1fa64180 YZ |
171 | static struct event_constraint intel_slm_event_constraints[] __read_mostly = |
172 | { | |
173 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | |
174 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
1fa64180 YZ |
175 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */ |
176 | EVENT_CONSTRAINT_END | |
177 | }; | |
178 | ||
ee89cbc2 | 179 | static struct extra_reg intel_snb_extra_regs[] __read_mostly = { |
53ad0447 YZ |
180 | /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ |
181 | INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0), | |
182 | INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1), | |
f20093ee | 183 | INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), |
f1923820 SE |
184 | EVENT_EXTRA_END |
185 | }; | |
186 | ||
187 | static struct extra_reg intel_snbep_extra_regs[] __read_mostly = { | |
53ad0447 YZ |
188 | /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ |
189 | INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0), | |
190 | INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1), | |
f1a52789 | 191 | INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), |
ee89cbc2 SE |
192 | EVENT_EXTRA_END |
193 | }; | |
194 | ||
7f2ee91f IM |
195 | EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3"); |
196 | EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3"); | |
197 | EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2"); | |
f20093ee SE |
198 | |
199 | struct attribute *nhm_events_attrs[] = { | |
200 | EVENT_PTR(mem_ld_nhm), | |
201 | NULL, | |
202 | }; | |
203 | ||
204 | struct attribute *snb_events_attrs[] = { | |
205 | EVENT_PTR(mem_ld_snb), | |
9ad64c0f | 206 | EVENT_PTR(mem_st_snb), |
f20093ee SE |
207 | NULL, |
208 | }; | |
209 | ||
3a632cb2 AK |
210 | static struct event_constraint intel_hsw_event_constraints[] = { |
211 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | |
212 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
213 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ | |
214 | INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.* */ | |
215 | INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ | |
216 | INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ | |
217 | /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ | |
c420f19b | 218 | INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), |
3a632cb2 | 219 | /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ |
c420f19b | 220 | INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), |
3a632cb2 | 221 | /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */ |
c420f19b | 222 | INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), |
93fcf72c MD |
223 | |
224 | INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ | |
225 | INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ | |
226 | INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ | |
227 | INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ | |
228 | ||
3a632cb2 AK |
229 | EVENT_CONSTRAINT_END |
230 | }; | |
231 | ||
91f1b705 AK |
232 | struct event_constraint intel_bdw_event_constraints[] = { |
233 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | |
234 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
235 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ | |
236 | INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */ | |
237 | INTEL_EVENT_CONSTRAINT(0xa3, 0x4), /* CYCLE_ACTIVITY.* */ | |
238 | EVENT_CONSTRAINT_END | |
239 | }; | |
240 | ||
f22f54f4 PZ |
241 | static u64 intel_pmu_event_map(int hw_event) |
242 | { | |
243 | return intel_perfmon_event_map[hw_event]; | |
244 | } | |
245 | ||
74e6543f YZ |
246 | #define SNB_DMND_DATA_RD (1ULL << 0) |
247 | #define SNB_DMND_RFO (1ULL << 1) | |
248 | #define SNB_DMND_IFETCH (1ULL << 2) | |
249 | #define SNB_DMND_WB (1ULL << 3) | |
250 | #define SNB_PF_DATA_RD (1ULL << 4) | |
251 | #define SNB_PF_RFO (1ULL << 5) | |
252 | #define SNB_PF_IFETCH (1ULL << 6) | |
253 | #define SNB_LLC_DATA_RD (1ULL << 7) | |
254 | #define SNB_LLC_RFO (1ULL << 8) | |
255 | #define SNB_LLC_IFETCH (1ULL << 9) | |
256 | #define SNB_BUS_LOCKS (1ULL << 10) | |
257 | #define SNB_STRM_ST (1ULL << 11) | |
258 | #define SNB_OTHER (1ULL << 15) | |
259 | #define SNB_RESP_ANY (1ULL << 16) | |
260 | #define SNB_NO_SUPP (1ULL << 17) | |
261 | #define SNB_LLC_HITM (1ULL << 18) | |
262 | #define SNB_LLC_HITE (1ULL << 19) | |
263 | #define SNB_LLC_HITS (1ULL << 20) | |
264 | #define SNB_LLC_HITF (1ULL << 21) | |
265 | #define SNB_LOCAL (1ULL << 22) | |
266 | #define SNB_REMOTE (0xffULL << 23) | |
267 | #define SNB_SNP_NONE (1ULL << 31) | |
268 | #define SNB_SNP_NOT_NEEDED (1ULL << 32) | |
269 | #define SNB_SNP_MISS (1ULL << 33) | |
270 | #define SNB_NO_FWD (1ULL << 34) | |
271 | #define SNB_SNP_FWD (1ULL << 35) | |
272 | #define SNB_HITM (1ULL << 36) | |
273 | #define SNB_NON_DRAM (1ULL << 37) | |
274 | ||
275 | #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD) | |
276 | #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO) | |
277 | #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) | |
278 | ||
279 | #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \ | |
280 | SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \ | |
281 | SNB_HITM) | |
282 | ||
283 | #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY) | |
284 | #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY) | |
285 | ||
286 | #define SNB_L3_ACCESS SNB_RESP_ANY | |
287 | #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM) | |
288 | ||
289 | static __initconst const u64 snb_hw_cache_extra_regs | |
290 | [PERF_COUNT_HW_CACHE_MAX] | |
291 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
292 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
293 | { | |
294 | [ C(LL ) ] = { | |
295 | [ C(OP_READ) ] = { | |
296 | [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS, | |
297 | [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_L3_MISS, | |
298 | }, | |
299 | [ C(OP_WRITE) ] = { | |
300 | [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS, | |
301 | [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_L3_MISS, | |
302 | }, | |
303 | [ C(OP_PREFETCH) ] = { | |
304 | [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS, | |
305 | [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_L3_MISS, | |
306 | }, | |
307 | }, | |
308 | [ C(NODE) ] = { | |
309 | [ C(OP_READ) ] = { | |
310 | [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY, | |
311 | [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_DRAM_REMOTE, | |
312 | }, | |
313 | [ C(OP_WRITE) ] = { | |
314 | [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY, | |
315 | [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE, | |
316 | }, | |
317 | [ C(OP_PREFETCH) ] = { | |
318 | [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY, | |
319 | [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE, | |
320 | }, | |
321 | }, | |
322 | }; | |
323 | ||
b06b3d49 LM |
324 | static __initconst const u64 snb_hw_cache_event_ids |
325 | [PERF_COUNT_HW_CACHE_MAX] | |
326 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
327 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
328 | { | |
329 | [ C(L1D) ] = { | |
330 | [ C(OP_READ) ] = { | |
331 | [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */ | |
332 | [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */ | |
333 | }, | |
334 | [ C(OP_WRITE) ] = { | |
335 | [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */ | |
336 | [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */ | |
337 | }, | |
338 | [ C(OP_PREFETCH) ] = { | |
339 | [ C(RESULT_ACCESS) ] = 0x0, | |
340 | [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */ | |
341 | }, | |
342 | }, | |
343 | [ C(L1I ) ] = { | |
344 | [ C(OP_READ) ] = { | |
345 | [ C(RESULT_ACCESS) ] = 0x0, | |
346 | [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */ | |
347 | }, | |
348 | [ C(OP_WRITE) ] = { | |
349 | [ C(RESULT_ACCESS) ] = -1, | |
350 | [ C(RESULT_MISS) ] = -1, | |
351 | }, | |
352 | [ C(OP_PREFETCH) ] = { | |
353 | [ C(RESULT_ACCESS) ] = 0x0, | |
354 | [ C(RESULT_MISS) ] = 0x0, | |
355 | }, | |
356 | }, | |
357 | [ C(LL ) ] = { | |
b06b3d49 | 358 | [ C(OP_READ) ] = { |
63b6a675 | 359 | /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ |
b06b3d49 | 360 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
361 | /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ |
362 | [ C(RESULT_MISS) ] = 0x01b7, | |
b06b3d49 LM |
363 | }, |
364 | [ C(OP_WRITE) ] = { | |
63b6a675 | 365 | /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ |
b06b3d49 | 366 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
367 | /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ |
368 | [ C(RESULT_MISS) ] = 0x01b7, | |
b06b3d49 LM |
369 | }, |
370 | [ C(OP_PREFETCH) ] = { | |
63b6a675 | 371 | /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ |
b06b3d49 | 372 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
373 | /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ |
374 | [ C(RESULT_MISS) ] = 0x01b7, | |
b06b3d49 LM |
375 | }, |
376 | }, | |
377 | [ C(DTLB) ] = { | |
378 | [ C(OP_READ) ] = { | |
379 | [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */ | |
380 | [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */ | |
381 | }, | |
382 | [ C(OP_WRITE) ] = { | |
383 | [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */ | |
384 | [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */ | |
385 | }, | |
386 | [ C(OP_PREFETCH) ] = { | |
387 | [ C(RESULT_ACCESS) ] = 0x0, | |
388 | [ C(RESULT_MISS) ] = 0x0, | |
389 | }, | |
390 | }, | |
391 | [ C(ITLB) ] = { | |
392 | [ C(OP_READ) ] = { | |
393 | [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */ | |
394 | [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */ | |
395 | }, | |
396 | [ C(OP_WRITE) ] = { | |
397 | [ C(RESULT_ACCESS) ] = -1, | |
398 | [ C(RESULT_MISS) ] = -1, | |
399 | }, | |
400 | [ C(OP_PREFETCH) ] = { | |
401 | [ C(RESULT_ACCESS) ] = -1, | |
402 | [ C(RESULT_MISS) ] = -1, | |
403 | }, | |
404 | }, | |
405 | [ C(BPU ) ] = { | |
406 | [ C(OP_READ) ] = { | |
407 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
408 | [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ | |
409 | }, | |
410 | [ C(OP_WRITE) ] = { | |
411 | [ C(RESULT_ACCESS) ] = -1, | |
412 | [ C(RESULT_MISS) ] = -1, | |
413 | }, | |
414 | [ C(OP_PREFETCH) ] = { | |
415 | [ C(RESULT_ACCESS) ] = -1, | |
416 | [ C(RESULT_MISS) ] = -1, | |
417 | }, | |
418 | }, | |
89d6c0b5 PZ |
419 | [ C(NODE) ] = { |
420 | [ C(OP_READ) ] = { | |
74e6543f YZ |
421 | [ C(RESULT_ACCESS) ] = 0x01b7, |
422 | [ C(RESULT_MISS) ] = 0x01b7, | |
89d6c0b5 PZ |
423 | }, |
424 | [ C(OP_WRITE) ] = { | |
74e6543f YZ |
425 | [ C(RESULT_ACCESS) ] = 0x01b7, |
426 | [ C(RESULT_MISS) ] = 0x01b7, | |
89d6c0b5 PZ |
427 | }, |
428 | [ C(OP_PREFETCH) ] = { | |
74e6543f YZ |
429 | [ C(RESULT_ACCESS) ] = 0x01b7, |
430 | [ C(RESULT_MISS) ] = 0x01b7, | |
89d6c0b5 PZ |
431 | }, |
432 | }, | |
433 | ||
b06b3d49 LM |
434 | }; |
435 | ||
0f1b5ca2 AK |
436 | /* |
437 | * Notes on the events: | |
438 | * - data reads do not include code reads (comparable to earlier tables) | |
439 | * - data counts include speculative execution (except L1 write, dtlb, bpu) | |
440 | * - remote node access includes remote memory, remote cache, remote mmio. | |
441 | * - prefetches are not included in the counts because they are not | |
442 | * reliably counted. | |
443 | */ | |
444 | ||
445 | #define HSW_DEMAND_DATA_RD BIT_ULL(0) | |
446 | #define HSW_DEMAND_RFO BIT_ULL(1) | |
447 | #define HSW_ANY_RESPONSE BIT_ULL(16) | |
448 | #define HSW_SUPPLIER_NONE BIT_ULL(17) | |
449 | #define HSW_L3_MISS_LOCAL_DRAM BIT_ULL(22) | |
450 | #define HSW_L3_MISS_REMOTE_HOP0 BIT_ULL(27) | |
451 | #define HSW_L3_MISS_REMOTE_HOP1 BIT_ULL(28) | |
452 | #define HSW_L3_MISS_REMOTE_HOP2P BIT_ULL(29) | |
453 | #define HSW_L3_MISS (HSW_L3_MISS_LOCAL_DRAM| \ | |
454 | HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \ | |
455 | HSW_L3_MISS_REMOTE_HOP2P) | |
456 | #define HSW_SNOOP_NONE BIT_ULL(31) | |
457 | #define HSW_SNOOP_NOT_NEEDED BIT_ULL(32) | |
458 | #define HSW_SNOOP_MISS BIT_ULL(33) | |
459 | #define HSW_SNOOP_HIT_NO_FWD BIT_ULL(34) | |
460 | #define HSW_SNOOP_HIT_WITH_FWD BIT_ULL(35) | |
461 | #define HSW_SNOOP_HITM BIT_ULL(36) | |
462 | #define HSW_SNOOP_NON_DRAM BIT_ULL(37) | |
463 | #define HSW_ANY_SNOOP (HSW_SNOOP_NONE| \ | |
464 | HSW_SNOOP_NOT_NEEDED|HSW_SNOOP_MISS| \ | |
465 | HSW_SNOOP_HIT_NO_FWD|HSW_SNOOP_HIT_WITH_FWD| \ | |
466 | HSW_SNOOP_HITM|HSW_SNOOP_NON_DRAM) | |
467 | #define HSW_SNOOP_DRAM (HSW_ANY_SNOOP & ~HSW_SNOOP_NON_DRAM) | |
468 | #define HSW_DEMAND_READ HSW_DEMAND_DATA_RD | |
469 | #define HSW_DEMAND_WRITE HSW_DEMAND_RFO | |
470 | #define HSW_L3_MISS_REMOTE (HSW_L3_MISS_REMOTE_HOP0|\ | |
471 | HSW_L3_MISS_REMOTE_HOP1|HSW_L3_MISS_REMOTE_HOP2P) | |
472 | #define HSW_LLC_ACCESS HSW_ANY_RESPONSE | |
473 | ||
91f1b705 AK |
474 | #define BDW_L3_MISS_LOCAL BIT(26) |
475 | #define BDW_L3_MISS (BDW_L3_MISS_LOCAL| \ | |
476 | HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \ | |
477 | HSW_L3_MISS_REMOTE_HOP2P) | |
478 | ||
479 | ||
0f1b5ca2 AK |
480 | static __initconst const u64 hsw_hw_cache_event_ids |
481 | [PERF_COUNT_HW_CACHE_MAX] | |
482 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
483 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
484 | { | |
485 | [ C(L1D ) ] = { | |
486 | [ C(OP_READ) ] = { | |
487 | [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ | |
488 | [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */ | |
489 | }, | |
490 | [ C(OP_WRITE) ] = { | |
491 | [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ | |
492 | [ C(RESULT_MISS) ] = 0x0, | |
493 | }, | |
494 | [ C(OP_PREFETCH) ] = { | |
495 | [ C(RESULT_ACCESS) ] = 0x0, | |
496 | [ C(RESULT_MISS) ] = 0x0, | |
497 | }, | |
498 | }, | |
499 | [ C(L1I ) ] = { | |
500 | [ C(OP_READ) ] = { | |
501 | [ C(RESULT_ACCESS) ] = 0x0, | |
502 | [ C(RESULT_MISS) ] = 0x280, /* ICACHE.MISSES */ | |
503 | }, | |
504 | [ C(OP_WRITE) ] = { | |
505 | [ C(RESULT_ACCESS) ] = -1, | |
506 | [ C(RESULT_MISS) ] = -1, | |
507 | }, | |
508 | [ C(OP_PREFETCH) ] = { | |
509 | [ C(RESULT_ACCESS) ] = 0x0, | |
510 | [ C(RESULT_MISS) ] = 0x0, | |
511 | }, | |
512 | }, | |
513 | [ C(LL ) ] = { | |
514 | [ C(OP_READ) ] = { | |
515 | [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
516 | [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
517 | }, | |
518 | [ C(OP_WRITE) ] = { | |
519 | [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
520 | [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
521 | }, | |
522 | [ C(OP_PREFETCH) ] = { | |
523 | [ C(RESULT_ACCESS) ] = 0x0, | |
524 | [ C(RESULT_MISS) ] = 0x0, | |
525 | }, | |
526 | }, | |
527 | [ C(DTLB) ] = { | |
528 | [ C(OP_READ) ] = { | |
529 | [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ | |
530 | [ C(RESULT_MISS) ] = 0x108, /* DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK */ | |
531 | }, | |
532 | [ C(OP_WRITE) ] = { | |
533 | [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ | |
534 | [ C(RESULT_MISS) ] = 0x149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */ | |
535 | }, | |
536 | [ C(OP_PREFETCH) ] = { | |
537 | [ C(RESULT_ACCESS) ] = 0x0, | |
538 | [ C(RESULT_MISS) ] = 0x0, | |
539 | }, | |
540 | }, | |
541 | [ C(ITLB) ] = { | |
542 | [ C(OP_READ) ] = { | |
543 | [ C(RESULT_ACCESS) ] = 0x6085, /* ITLB_MISSES.STLB_HIT */ | |
544 | [ C(RESULT_MISS) ] = 0x185, /* ITLB_MISSES.MISS_CAUSES_A_WALK */ | |
545 | }, | |
546 | [ C(OP_WRITE) ] = { | |
547 | [ C(RESULT_ACCESS) ] = -1, | |
548 | [ C(RESULT_MISS) ] = -1, | |
549 | }, | |
550 | [ C(OP_PREFETCH) ] = { | |
551 | [ C(RESULT_ACCESS) ] = -1, | |
552 | [ C(RESULT_MISS) ] = -1, | |
553 | }, | |
554 | }, | |
555 | [ C(BPU ) ] = { | |
556 | [ C(OP_READ) ] = { | |
557 | [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
558 | [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */ | |
559 | }, | |
560 | [ C(OP_WRITE) ] = { | |
561 | [ C(RESULT_ACCESS) ] = -1, | |
562 | [ C(RESULT_MISS) ] = -1, | |
563 | }, | |
564 | [ C(OP_PREFETCH) ] = { | |
565 | [ C(RESULT_ACCESS) ] = -1, | |
566 | [ C(RESULT_MISS) ] = -1, | |
567 | }, | |
568 | }, | |
569 | [ C(NODE) ] = { | |
570 | [ C(OP_READ) ] = { | |
571 | [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
572 | [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
573 | }, | |
574 | [ C(OP_WRITE) ] = { | |
575 | [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
576 | [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
577 | }, | |
578 | [ C(OP_PREFETCH) ] = { | |
579 | [ C(RESULT_ACCESS) ] = 0x0, | |
580 | [ C(RESULT_MISS) ] = 0x0, | |
581 | }, | |
582 | }, | |
583 | }; | |
584 | ||
585 | static __initconst const u64 hsw_hw_cache_extra_regs | |
586 | [PERF_COUNT_HW_CACHE_MAX] | |
587 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
588 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
589 | { | |
590 | [ C(LL ) ] = { | |
591 | [ C(OP_READ) ] = { | |
592 | [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ| | |
593 | HSW_LLC_ACCESS, | |
594 | [ C(RESULT_MISS) ] = HSW_DEMAND_READ| | |
595 | HSW_L3_MISS|HSW_ANY_SNOOP, | |
596 | }, | |
597 | [ C(OP_WRITE) ] = { | |
598 | [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE| | |
599 | HSW_LLC_ACCESS, | |
600 | [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE| | |
601 | HSW_L3_MISS|HSW_ANY_SNOOP, | |
602 | }, | |
603 | [ C(OP_PREFETCH) ] = { | |
604 | [ C(RESULT_ACCESS) ] = 0x0, | |
605 | [ C(RESULT_MISS) ] = 0x0, | |
606 | }, | |
607 | }, | |
608 | [ C(NODE) ] = { | |
609 | [ C(OP_READ) ] = { | |
610 | [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ| | |
611 | HSW_L3_MISS_LOCAL_DRAM| | |
612 | HSW_SNOOP_DRAM, | |
613 | [ C(RESULT_MISS) ] = HSW_DEMAND_READ| | |
614 | HSW_L3_MISS_REMOTE| | |
615 | HSW_SNOOP_DRAM, | |
616 | }, | |
617 | [ C(OP_WRITE) ] = { | |
618 | [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE| | |
619 | HSW_L3_MISS_LOCAL_DRAM| | |
620 | HSW_SNOOP_DRAM, | |
621 | [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE| | |
622 | HSW_L3_MISS_REMOTE| | |
623 | HSW_SNOOP_DRAM, | |
624 | }, | |
625 | [ C(OP_PREFETCH) ] = { | |
626 | [ C(RESULT_ACCESS) ] = 0x0, | |
627 | [ C(RESULT_MISS) ] = 0x0, | |
628 | }, | |
629 | }, | |
630 | }; | |
631 | ||
caaa8be3 | 632 | static __initconst const u64 westmere_hw_cache_event_ids |
f22f54f4 PZ |
633 | [PERF_COUNT_HW_CACHE_MAX] |
634 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
635 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
636 | { | |
637 | [ C(L1D) ] = { | |
638 | [ C(OP_READ) ] = { | |
639 | [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ | |
640 | [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ | |
641 | }, | |
642 | [ C(OP_WRITE) ] = { | |
643 | [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ | |
644 | [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ | |
645 | }, | |
646 | [ C(OP_PREFETCH) ] = { | |
647 | [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ | |
648 | [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ | |
649 | }, | |
650 | }, | |
651 | [ C(L1I ) ] = { | |
652 | [ C(OP_READ) ] = { | |
653 | [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ | |
654 | [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ | |
655 | }, | |
656 | [ C(OP_WRITE) ] = { | |
657 | [ C(RESULT_ACCESS) ] = -1, | |
658 | [ C(RESULT_MISS) ] = -1, | |
659 | }, | |
660 | [ C(OP_PREFETCH) ] = { | |
661 | [ C(RESULT_ACCESS) ] = 0x0, | |
662 | [ C(RESULT_MISS) ] = 0x0, | |
663 | }, | |
664 | }, | |
665 | [ C(LL ) ] = { | |
666 | [ C(OP_READ) ] = { | |
63b6a675 | 667 | /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ |
e994d7d2 | 668 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
669 | /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ |
670 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 | 671 | }, |
e994d7d2 AK |
672 | /* |
673 | * Use RFO, not WRITEBACK, because a write miss would typically occur | |
674 | * on RFO. | |
675 | */ | |
f22f54f4 | 676 | [ C(OP_WRITE) ] = { |
63b6a675 PZ |
677 | /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ |
678 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
679 | /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ | |
e994d7d2 | 680 | [ C(RESULT_MISS) ] = 0x01b7, |
f22f54f4 PZ |
681 | }, |
682 | [ C(OP_PREFETCH) ] = { | |
63b6a675 | 683 | /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ |
e994d7d2 | 684 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
685 | /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ |
686 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 PZ |
687 | }, |
688 | }, | |
689 | [ C(DTLB) ] = { | |
690 | [ C(OP_READ) ] = { | |
691 | [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ | |
692 | [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ | |
693 | }, | |
694 | [ C(OP_WRITE) ] = { | |
695 | [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ | |
696 | [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ | |
697 | }, | |
698 | [ C(OP_PREFETCH) ] = { | |
699 | [ C(RESULT_ACCESS) ] = 0x0, | |
700 | [ C(RESULT_MISS) ] = 0x0, | |
701 | }, | |
702 | }, | |
703 | [ C(ITLB) ] = { | |
704 | [ C(OP_READ) ] = { | |
705 | [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ | |
706 | [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */ | |
707 | }, | |
708 | [ C(OP_WRITE) ] = { | |
709 | [ C(RESULT_ACCESS) ] = -1, | |
710 | [ C(RESULT_MISS) ] = -1, | |
711 | }, | |
712 | [ C(OP_PREFETCH) ] = { | |
713 | [ C(RESULT_ACCESS) ] = -1, | |
714 | [ C(RESULT_MISS) ] = -1, | |
715 | }, | |
716 | }, | |
717 | [ C(BPU ) ] = { | |
718 | [ C(OP_READ) ] = { | |
719 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
720 | [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ | |
721 | }, | |
722 | [ C(OP_WRITE) ] = { | |
723 | [ C(RESULT_ACCESS) ] = -1, | |
724 | [ C(RESULT_MISS) ] = -1, | |
725 | }, | |
726 | [ C(OP_PREFETCH) ] = { | |
727 | [ C(RESULT_ACCESS) ] = -1, | |
728 | [ C(RESULT_MISS) ] = -1, | |
729 | }, | |
730 | }, | |
89d6c0b5 PZ |
731 | [ C(NODE) ] = { |
732 | [ C(OP_READ) ] = { | |
733 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
734 | [ C(RESULT_MISS) ] = 0x01b7, | |
735 | }, | |
736 | [ C(OP_WRITE) ] = { | |
737 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
738 | [ C(RESULT_MISS) ] = 0x01b7, | |
739 | }, | |
740 | [ C(OP_PREFETCH) ] = { | |
741 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
742 | [ C(RESULT_MISS) ] = 0x01b7, | |
743 | }, | |
744 | }, | |
f22f54f4 PZ |
745 | }; |
746 | ||
e994d7d2 | 747 | /* |
63b6a675 PZ |
748 | * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits; |
749 | * See IA32 SDM Vol 3B 30.6.1.3 | |
e994d7d2 AK |
750 | */ |
751 | ||
63b6a675 PZ |
752 | #define NHM_DMND_DATA_RD (1 << 0) |
753 | #define NHM_DMND_RFO (1 << 1) | |
754 | #define NHM_DMND_IFETCH (1 << 2) | |
755 | #define NHM_DMND_WB (1 << 3) | |
756 | #define NHM_PF_DATA_RD (1 << 4) | |
757 | #define NHM_PF_DATA_RFO (1 << 5) | |
758 | #define NHM_PF_IFETCH (1 << 6) | |
759 | #define NHM_OFFCORE_OTHER (1 << 7) | |
760 | #define NHM_UNCORE_HIT (1 << 8) | |
761 | #define NHM_OTHER_CORE_HIT_SNP (1 << 9) | |
762 | #define NHM_OTHER_CORE_HITM (1 << 10) | |
763 | /* reserved */ | |
764 | #define NHM_REMOTE_CACHE_FWD (1 << 12) | |
765 | #define NHM_REMOTE_DRAM (1 << 13) | |
766 | #define NHM_LOCAL_DRAM (1 << 14) | |
767 | #define NHM_NON_DRAM (1 << 15) | |
768 | ||
87e24f4b PZ |
769 | #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD) |
770 | #define NHM_REMOTE (NHM_REMOTE_DRAM) | |
63b6a675 PZ |
771 | |
772 | #define NHM_DMND_READ (NHM_DMND_DATA_RD) | |
773 | #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB) | |
774 | #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO) | |
775 | ||
776 | #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM) | |
87e24f4b | 777 | #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD) |
63b6a675 | 778 | #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS) |
e994d7d2 AK |
779 | |
780 | static __initconst const u64 nehalem_hw_cache_extra_regs | |
781 | [PERF_COUNT_HW_CACHE_MAX] | |
782 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
783 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
784 | { | |
785 | [ C(LL ) ] = { | |
786 | [ C(OP_READ) ] = { | |
63b6a675 PZ |
787 | [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS, |
788 | [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS, | |
e994d7d2 AK |
789 | }, |
790 | [ C(OP_WRITE) ] = { | |
63b6a675 PZ |
791 | [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS, |
792 | [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS, | |
e994d7d2 AK |
793 | }, |
794 | [ C(OP_PREFETCH) ] = { | |
63b6a675 PZ |
795 | [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS, |
796 | [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS, | |
e994d7d2 | 797 | }, |
89d6c0b5 PZ |
798 | }, |
799 | [ C(NODE) ] = { | |
800 | [ C(OP_READ) ] = { | |
87e24f4b PZ |
801 | [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE, |
802 | [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE, | |
89d6c0b5 PZ |
803 | }, |
804 | [ C(OP_WRITE) ] = { | |
87e24f4b PZ |
805 | [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE, |
806 | [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE, | |
89d6c0b5 PZ |
807 | }, |
808 | [ C(OP_PREFETCH) ] = { | |
87e24f4b PZ |
809 | [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE, |
810 | [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE, | |
89d6c0b5 PZ |
811 | }, |
812 | }, | |
e994d7d2 AK |
813 | }; |
814 | ||
caaa8be3 | 815 | static __initconst const u64 nehalem_hw_cache_event_ids |
f22f54f4 PZ |
816 | [PERF_COUNT_HW_CACHE_MAX] |
817 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
818 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
819 | { | |
820 | [ C(L1D) ] = { | |
821 | [ C(OP_READ) ] = { | |
f4929bd3 PZ |
822 | [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ |
823 | [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ | |
f22f54f4 PZ |
824 | }, |
825 | [ C(OP_WRITE) ] = { | |
f4929bd3 PZ |
826 | [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ |
827 | [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ | |
f22f54f4 PZ |
828 | }, |
829 | [ C(OP_PREFETCH) ] = { | |
830 | [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ | |
831 | [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ | |
832 | }, | |
833 | }, | |
834 | [ C(L1I ) ] = { | |
835 | [ C(OP_READ) ] = { | |
836 | [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ | |
837 | [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ | |
838 | }, | |
839 | [ C(OP_WRITE) ] = { | |
840 | [ C(RESULT_ACCESS) ] = -1, | |
841 | [ C(RESULT_MISS) ] = -1, | |
842 | }, | |
843 | [ C(OP_PREFETCH) ] = { | |
844 | [ C(RESULT_ACCESS) ] = 0x0, | |
845 | [ C(RESULT_MISS) ] = 0x0, | |
846 | }, | |
847 | }, | |
848 | [ C(LL ) ] = { | |
849 | [ C(OP_READ) ] = { | |
e994d7d2 AK |
850 | /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ |
851 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
852 | /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ | |
853 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 | 854 | }, |
e994d7d2 AK |
855 | /* |
856 | * Use RFO, not WRITEBACK, because a write miss would typically occur | |
857 | * on RFO. | |
858 | */ | |
f22f54f4 | 859 | [ C(OP_WRITE) ] = { |
e994d7d2 AK |
860 | /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ |
861 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
862 | /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ | |
863 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 PZ |
864 | }, |
865 | [ C(OP_PREFETCH) ] = { | |
e994d7d2 AK |
866 | /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ |
867 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
868 | /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ | |
869 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 PZ |
870 | }, |
871 | }, | |
872 | [ C(DTLB) ] = { | |
873 | [ C(OP_READ) ] = { | |
874 | [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ | |
875 | [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ | |
876 | }, | |
877 | [ C(OP_WRITE) ] = { | |
878 | [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ | |
879 | [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ | |
880 | }, | |
881 | [ C(OP_PREFETCH) ] = { | |
882 | [ C(RESULT_ACCESS) ] = 0x0, | |
883 | [ C(RESULT_MISS) ] = 0x0, | |
884 | }, | |
885 | }, | |
886 | [ C(ITLB) ] = { | |
887 | [ C(OP_READ) ] = { | |
888 | [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ | |
889 | [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */ | |
890 | }, | |
891 | [ C(OP_WRITE) ] = { | |
892 | [ C(RESULT_ACCESS) ] = -1, | |
893 | [ C(RESULT_MISS) ] = -1, | |
894 | }, | |
895 | [ C(OP_PREFETCH) ] = { | |
896 | [ C(RESULT_ACCESS) ] = -1, | |
897 | [ C(RESULT_MISS) ] = -1, | |
898 | }, | |
899 | }, | |
900 | [ C(BPU ) ] = { | |
901 | [ C(OP_READ) ] = { | |
902 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
903 | [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ | |
904 | }, | |
905 | [ C(OP_WRITE) ] = { | |
906 | [ C(RESULT_ACCESS) ] = -1, | |
907 | [ C(RESULT_MISS) ] = -1, | |
908 | }, | |
909 | [ C(OP_PREFETCH) ] = { | |
910 | [ C(RESULT_ACCESS) ] = -1, | |
911 | [ C(RESULT_MISS) ] = -1, | |
912 | }, | |
913 | }, | |
89d6c0b5 PZ |
914 | [ C(NODE) ] = { |
915 | [ C(OP_READ) ] = { | |
916 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
917 | [ C(RESULT_MISS) ] = 0x01b7, | |
918 | }, | |
919 | [ C(OP_WRITE) ] = { | |
920 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
921 | [ C(RESULT_MISS) ] = 0x01b7, | |
922 | }, | |
923 | [ C(OP_PREFETCH) ] = { | |
924 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
925 | [ C(RESULT_MISS) ] = 0x01b7, | |
926 | }, | |
927 | }, | |
f22f54f4 PZ |
928 | }; |
929 | ||
caaa8be3 | 930 | static __initconst const u64 core2_hw_cache_event_ids |
f22f54f4 PZ |
931 | [PERF_COUNT_HW_CACHE_MAX] |
932 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
933 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
934 | { | |
935 | [ C(L1D) ] = { | |
936 | [ C(OP_READ) ] = { | |
937 | [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */ | |
938 | [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */ | |
939 | }, | |
940 | [ C(OP_WRITE) ] = { | |
941 | [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */ | |
942 | [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */ | |
943 | }, | |
944 | [ C(OP_PREFETCH) ] = { | |
945 | [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */ | |
946 | [ C(RESULT_MISS) ] = 0, | |
947 | }, | |
948 | }, | |
949 | [ C(L1I ) ] = { | |
950 | [ C(OP_READ) ] = { | |
951 | [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */ | |
952 | [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */ | |
953 | }, | |
954 | [ C(OP_WRITE) ] = { | |
955 | [ C(RESULT_ACCESS) ] = -1, | |
956 | [ C(RESULT_MISS) ] = -1, | |
957 | }, | |
958 | [ C(OP_PREFETCH) ] = { | |
959 | [ C(RESULT_ACCESS) ] = 0, | |
960 | [ C(RESULT_MISS) ] = 0, | |
961 | }, | |
962 | }, | |
963 | [ C(LL ) ] = { | |
964 | [ C(OP_READ) ] = { | |
965 | [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ | |
966 | [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ | |
967 | }, | |
968 | [ C(OP_WRITE) ] = { | |
969 | [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ | |
970 | [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ | |
971 | }, | |
972 | [ C(OP_PREFETCH) ] = { | |
973 | [ C(RESULT_ACCESS) ] = 0, | |
974 | [ C(RESULT_MISS) ] = 0, | |
975 | }, | |
976 | }, | |
977 | [ C(DTLB) ] = { | |
978 | [ C(OP_READ) ] = { | |
979 | [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ | |
980 | [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */ | |
981 | }, | |
982 | [ C(OP_WRITE) ] = { | |
983 | [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ | |
984 | [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */ | |
985 | }, | |
986 | [ C(OP_PREFETCH) ] = { | |
987 | [ C(RESULT_ACCESS) ] = 0, | |
988 | [ C(RESULT_MISS) ] = 0, | |
989 | }, | |
990 | }, | |
991 | [ C(ITLB) ] = { | |
992 | [ C(OP_READ) ] = { | |
993 | [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ | |
994 | [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */ | |
995 | }, | |
996 | [ C(OP_WRITE) ] = { | |
997 | [ C(RESULT_ACCESS) ] = -1, | |
998 | [ C(RESULT_MISS) ] = -1, | |
999 | }, | |
1000 | [ C(OP_PREFETCH) ] = { | |
1001 | [ C(RESULT_ACCESS) ] = -1, | |
1002 | [ C(RESULT_MISS) ] = -1, | |
1003 | }, | |
1004 | }, | |
1005 | [ C(BPU ) ] = { | |
1006 | [ C(OP_READ) ] = { | |
1007 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ | |
1008 | [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ | |
1009 | }, | |
1010 | [ C(OP_WRITE) ] = { | |
1011 | [ C(RESULT_ACCESS) ] = -1, | |
1012 | [ C(RESULT_MISS) ] = -1, | |
1013 | }, | |
1014 | [ C(OP_PREFETCH) ] = { | |
1015 | [ C(RESULT_ACCESS) ] = -1, | |
1016 | [ C(RESULT_MISS) ] = -1, | |
1017 | }, | |
1018 | }, | |
1019 | }; | |
1020 | ||
caaa8be3 | 1021 | static __initconst const u64 atom_hw_cache_event_ids |
f22f54f4 PZ |
1022 | [PERF_COUNT_HW_CACHE_MAX] |
1023 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
1024 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
1025 | { | |
1026 | [ C(L1D) ] = { | |
1027 | [ C(OP_READ) ] = { | |
1028 | [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */ | |
1029 | [ C(RESULT_MISS) ] = 0, | |
1030 | }, | |
1031 | [ C(OP_WRITE) ] = { | |
1032 | [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */ | |
1033 | [ C(RESULT_MISS) ] = 0, | |
1034 | }, | |
1035 | [ C(OP_PREFETCH) ] = { | |
1036 | [ C(RESULT_ACCESS) ] = 0x0, | |
1037 | [ C(RESULT_MISS) ] = 0, | |
1038 | }, | |
1039 | }, | |
1040 | [ C(L1I ) ] = { | |
1041 | [ C(OP_READ) ] = { | |
1042 | [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ | |
1043 | [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ | |
1044 | }, | |
1045 | [ C(OP_WRITE) ] = { | |
1046 | [ C(RESULT_ACCESS) ] = -1, | |
1047 | [ C(RESULT_MISS) ] = -1, | |
1048 | }, | |
1049 | [ C(OP_PREFETCH) ] = { | |
1050 | [ C(RESULT_ACCESS) ] = 0, | |
1051 | [ C(RESULT_MISS) ] = 0, | |
1052 | }, | |
1053 | }, | |
1054 | [ C(LL ) ] = { | |
1055 | [ C(OP_READ) ] = { | |
1056 | [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ | |
1057 | [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ | |
1058 | }, | |
1059 | [ C(OP_WRITE) ] = { | |
1060 | [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ | |
1061 | [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ | |
1062 | }, | |
1063 | [ C(OP_PREFETCH) ] = { | |
1064 | [ C(RESULT_ACCESS) ] = 0, | |
1065 | [ C(RESULT_MISS) ] = 0, | |
1066 | }, | |
1067 | }, | |
1068 | [ C(DTLB) ] = { | |
1069 | [ C(OP_READ) ] = { | |
1070 | [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */ | |
1071 | [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */ | |
1072 | }, | |
1073 | [ C(OP_WRITE) ] = { | |
1074 | [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */ | |
1075 | [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */ | |
1076 | }, | |
1077 | [ C(OP_PREFETCH) ] = { | |
1078 | [ C(RESULT_ACCESS) ] = 0, | |
1079 | [ C(RESULT_MISS) ] = 0, | |
1080 | }, | |
1081 | }, | |
1082 | [ C(ITLB) ] = { | |
1083 | [ C(OP_READ) ] = { | |
1084 | [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ | |
1085 | [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */ | |
1086 | }, | |
1087 | [ C(OP_WRITE) ] = { | |
1088 | [ C(RESULT_ACCESS) ] = -1, | |
1089 | [ C(RESULT_MISS) ] = -1, | |
1090 | }, | |
1091 | [ C(OP_PREFETCH) ] = { | |
1092 | [ C(RESULT_ACCESS) ] = -1, | |
1093 | [ C(RESULT_MISS) ] = -1, | |
1094 | }, | |
1095 | }, | |
1096 | [ C(BPU ) ] = { | |
1097 | [ C(OP_READ) ] = { | |
1098 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ | |
1099 | [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ | |
1100 | }, | |
1101 | [ C(OP_WRITE) ] = { | |
1102 | [ C(RESULT_ACCESS) ] = -1, | |
1103 | [ C(RESULT_MISS) ] = -1, | |
1104 | }, | |
1105 | [ C(OP_PREFETCH) ] = { | |
1106 | [ C(RESULT_ACCESS) ] = -1, | |
1107 | [ C(RESULT_MISS) ] = -1, | |
1108 | }, | |
1109 | }, | |
1110 | }; | |
1111 | ||
1fa64180 YZ |
1112 | static struct extra_reg intel_slm_extra_regs[] __read_mostly = |
1113 | { | |
1114 | /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ | |
06c939c1 PZ |
1115 | INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x768005ffffull, RSP_0), |
1116 | INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x768005ffffull, RSP_1), | |
1fa64180 YZ |
1117 | EVENT_EXTRA_END |
1118 | }; | |
1119 | ||
1120 | #define SLM_DMND_READ SNB_DMND_DATA_RD | |
1121 | #define SLM_DMND_WRITE SNB_DMND_RFO | |
1122 | #define SLM_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) | |
1123 | ||
1124 | #define SLM_SNP_ANY (SNB_SNP_NONE|SNB_SNP_MISS|SNB_NO_FWD|SNB_HITM) | |
1125 | #define SLM_LLC_ACCESS SNB_RESP_ANY | |
1126 | #define SLM_LLC_MISS (SLM_SNP_ANY|SNB_NON_DRAM) | |
1127 | ||
1128 | static __initconst const u64 slm_hw_cache_extra_regs | |
1129 | [PERF_COUNT_HW_CACHE_MAX] | |
1130 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
1131 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
1132 | { | |
1133 | [ C(LL ) ] = { | |
1134 | [ C(OP_READ) ] = { | |
1135 | [ C(RESULT_ACCESS) ] = SLM_DMND_READ|SLM_LLC_ACCESS, | |
1136 | [ C(RESULT_MISS) ] = SLM_DMND_READ|SLM_LLC_MISS, | |
1137 | }, | |
1138 | [ C(OP_WRITE) ] = { | |
1139 | [ C(RESULT_ACCESS) ] = SLM_DMND_WRITE|SLM_LLC_ACCESS, | |
1140 | [ C(RESULT_MISS) ] = SLM_DMND_WRITE|SLM_LLC_MISS, | |
1141 | }, | |
1142 | [ C(OP_PREFETCH) ] = { | |
1143 | [ C(RESULT_ACCESS) ] = SLM_DMND_PREFETCH|SLM_LLC_ACCESS, | |
1144 | [ C(RESULT_MISS) ] = SLM_DMND_PREFETCH|SLM_LLC_MISS, | |
1145 | }, | |
1146 | }, | |
1147 | }; | |
1148 | ||
1149 | static __initconst const u64 slm_hw_cache_event_ids | |
1150 | [PERF_COUNT_HW_CACHE_MAX] | |
1151 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
1152 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
1153 | { | |
1154 | [ C(L1D) ] = { | |
1155 | [ C(OP_READ) ] = { | |
1156 | [ C(RESULT_ACCESS) ] = 0, | |
1157 | [ C(RESULT_MISS) ] = 0x0104, /* LD_DCU_MISS */ | |
1158 | }, | |
1159 | [ C(OP_WRITE) ] = { | |
1160 | [ C(RESULT_ACCESS) ] = 0, | |
1161 | [ C(RESULT_MISS) ] = 0, | |
1162 | }, | |
1163 | [ C(OP_PREFETCH) ] = { | |
1164 | [ C(RESULT_ACCESS) ] = 0, | |
1165 | [ C(RESULT_MISS) ] = 0, | |
1166 | }, | |
1167 | }, | |
1168 | [ C(L1I ) ] = { | |
1169 | [ C(OP_READ) ] = { | |
1170 | [ C(RESULT_ACCESS) ] = 0x0380, /* ICACHE.ACCESSES */ | |
1171 | [ C(RESULT_MISS) ] = 0x0280, /* ICACGE.MISSES */ | |
1172 | }, | |
1173 | [ C(OP_WRITE) ] = { | |
1174 | [ C(RESULT_ACCESS) ] = -1, | |
1175 | [ C(RESULT_MISS) ] = -1, | |
1176 | }, | |
1177 | [ C(OP_PREFETCH) ] = { | |
1178 | [ C(RESULT_ACCESS) ] = 0, | |
1179 | [ C(RESULT_MISS) ] = 0, | |
1180 | }, | |
1181 | }, | |
1182 | [ C(LL ) ] = { | |
1183 | [ C(OP_READ) ] = { | |
1184 | /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ | |
1185 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
1186 | /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ | |
1187 | [ C(RESULT_MISS) ] = 0x01b7, | |
1188 | }, | |
1189 | [ C(OP_WRITE) ] = { | |
1190 | /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ | |
1191 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
1192 | /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ | |
1193 | [ C(RESULT_MISS) ] = 0x01b7, | |
1194 | }, | |
1195 | [ C(OP_PREFETCH) ] = { | |
1196 | /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ | |
1197 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
1198 | /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ | |
1199 | [ C(RESULT_MISS) ] = 0x01b7, | |
1200 | }, | |
1201 | }, | |
1202 | [ C(DTLB) ] = { | |
1203 | [ C(OP_READ) ] = { | |
1204 | [ C(RESULT_ACCESS) ] = 0, | |
1205 | [ C(RESULT_MISS) ] = 0x0804, /* LD_DTLB_MISS */ | |
1206 | }, | |
1207 | [ C(OP_WRITE) ] = { | |
1208 | [ C(RESULT_ACCESS) ] = 0, | |
1209 | [ C(RESULT_MISS) ] = 0, | |
1210 | }, | |
1211 | [ C(OP_PREFETCH) ] = { | |
1212 | [ C(RESULT_ACCESS) ] = 0, | |
1213 | [ C(RESULT_MISS) ] = 0, | |
1214 | }, | |
1215 | }, | |
1216 | [ C(ITLB) ] = { | |
1217 | [ C(OP_READ) ] = { | |
1218 | [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ | |
1219 | [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */ | |
1220 | }, | |
1221 | [ C(OP_WRITE) ] = { | |
1222 | [ C(RESULT_ACCESS) ] = -1, | |
1223 | [ C(RESULT_MISS) ] = -1, | |
1224 | }, | |
1225 | [ C(OP_PREFETCH) ] = { | |
1226 | [ C(RESULT_ACCESS) ] = -1, | |
1227 | [ C(RESULT_MISS) ] = -1, | |
1228 | }, | |
1229 | }, | |
1230 | [ C(BPU ) ] = { | |
1231 | [ C(OP_READ) ] = { | |
1232 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ | |
1233 | [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ | |
1234 | }, | |
1235 | [ C(OP_WRITE) ] = { | |
1236 | [ C(RESULT_ACCESS) ] = -1, | |
1237 | [ C(RESULT_MISS) ] = -1, | |
1238 | }, | |
1239 | [ C(OP_PREFETCH) ] = { | |
1240 | [ C(RESULT_ACCESS) ] = -1, | |
1241 | [ C(RESULT_MISS) ] = -1, | |
1242 | }, | |
1243 | }, | |
1244 | }; | |
1245 | ||
f22f54f4 PZ |
1246 | static void intel_pmu_disable_all(void) |
1247 | { | |
89cbc767 | 1248 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
f22f54f4 PZ |
1249 | |
1250 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); | |
1251 | ||
15c7ad51 | 1252 | if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) |
f22f54f4 | 1253 | intel_pmu_disable_bts(); |
8062382c AS |
1254 | else |
1255 | intel_bts_disable_local(); | |
ca037701 PZ |
1256 | |
1257 | intel_pmu_pebs_disable_all(); | |
caff2bef | 1258 | intel_pmu_lbr_disable_all(); |
f22f54f4 PZ |
1259 | } |
1260 | ||
11164cd4 | 1261 | static void intel_pmu_enable_all(int added) |
f22f54f4 | 1262 | { |
89cbc767 | 1263 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
f22f54f4 | 1264 | |
d329527e PZ |
1265 | intel_pmu_pebs_enable_all(); |
1266 | intel_pmu_lbr_enable_all(); | |
144d31e6 GN |
1267 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, |
1268 | x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask); | |
f22f54f4 | 1269 | |
15c7ad51 | 1270 | if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) { |
f22f54f4 | 1271 | struct perf_event *event = |
15c7ad51 | 1272 | cpuc->events[INTEL_PMC_IDX_FIXED_BTS]; |
f22f54f4 PZ |
1273 | |
1274 | if (WARN_ON_ONCE(!event)) | |
1275 | return; | |
1276 | ||
1277 | intel_pmu_enable_bts(event->hw.config); | |
8062382c AS |
1278 | } else |
1279 | intel_bts_enable_local(); | |
f22f54f4 PZ |
1280 | } |
1281 | ||
11164cd4 PZ |
1282 | /* |
1283 | * Workaround for: | |
1284 | * Intel Errata AAK100 (model 26) | |
1285 | * Intel Errata AAP53 (model 30) | |
40b91cd1 | 1286 | * Intel Errata BD53 (model 44) |
11164cd4 | 1287 | * |
351af072 ZY |
1288 | * The official story: |
1289 | * These chips need to be 'reset' when adding counters by programming the | |
1290 | * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either | |
1291 | * in sequence on the same PMC or on different PMCs. | |
1292 | * | |
1293 | * In practise it appears some of these events do in fact count, and | |
1294 | * we need to programm all 4 events. | |
11164cd4 | 1295 | */ |
351af072 | 1296 | static void intel_pmu_nhm_workaround(void) |
11164cd4 | 1297 | { |
89cbc767 | 1298 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
351af072 ZY |
1299 | static const unsigned long nhm_magic[4] = { |
1300 | 0x4300B5, | |
1301 | 0x4300D2, | |
1302 | 0x4300B1, | |
1303 | 0x4300B1 | |
1304 | }; | |
1305 | struct perf_event *event; | |
1306 | int i; | |
11164cd4 | 1307 | |
351af072 ZY |
1308 | /* |
1309 | * The Errata requires below steps: | |
1310 | * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL; | |
1311 | * 2) Configure 4 PERFEVTSELx with the magic events and clear | |
1312 | * the corresponding PMCx; | |
1313 | * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL; | |
1314 | * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL; | |
1315 | * 5) Clear 4 pairs of ERFEVTSELx and PMCx; | |
1316 | */ | |
11164cd4 | 1317 | |
351af072 ZY |
1318 | /* |
1319 | * The real steps we choose are a little different from above. | |
1320 | * A) To reduce MSR operations, we don't run step 1) as they | |
1321 | * are already cleared before this function is called; | |
1322 | * B) Call x86_perf_event_update to save PMCx before configuring | |
1323 | * PERFEVTSELx with magic number; | |
1324 | * C) With step 5), we do clear only when the PERFEVTSELx is | |
1325 | * not used currently. | |
1326 | * D) Call x86_perf_event_set_period to restore PMCx; | |
1327 | */ | |
11164cd4 | 1328 | |
351af072 ZY |
1329 | /* We always operate 4 pairs of PERF Counters */ |
1330 | for (i = 0; i < 4; i++) { | |
1331 | event = cpuc->events[i]; | |
1332 | if (event) | |
1333 | x86_perf_event_update(event); | |
1334 | } | |
11164cd4 | 1335 | |
351af072 ZY |
1336 | for (i = 0; i < 4; i++) { |
1337 | wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]); | |
1338 | wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0); | |
1339 | } | |
1340 | ||
1341 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf); | |
1342 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0); | |
11164cd4 | 1343 | |
351af072 ZY |
1344 | for (i = 0; i < 4; i++) { |
1345 | event = cpuc->events[i]; | |
1346 | ||
1347 | if (event) { | |
1348 | x86_perf_event_set_period(event); | |
31fa58af | 1349 | __x86_pmu_enable_event(&event->hw, |
351af072 ZY |
1350 | ARCH_PERFMON_EVENTSEL_ENABLE); |
1351 | } else | |
1352 | wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0); | |
11164cd4 | 1353 | } |
351af072 ZY |
1354 | } |
1355 | ||
1356 | static void intel_pmu_nhm_enable_all(int added) | |
1357 | { | |
1358 | if (added) | |
1359 | intel_pmu_nhm_workaround(); | |
11164cd4 PZ |
1360 | intel_pmu_enable_all(added); |
1361 | } | |
1362 | ||
f22f54f4 PZ |
1363 | static inline u64 intel_pmu_get_status(void) |
1364 | { | |
1365 | u64 status; | |
1366 | ||
1367 | rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status); | |
1368 | ||
1369 | return status; | |
1370 | } | |
1371 | ||
1372 | static inline void intel_pmu_ack_status(u64 ack) | |
1373 | { | |
1374 | wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack); | |
1375 | } | |
1376 | ||
ca037701 | 1377 | static void intel_pmu_disable_fixed(struct hw_perf_event *hwc) |
f22f54f4 | 1378 | { |
15c7ad51 | 1379 | int idx = hwc->idx - INTEL_PMC_IDX_FIXED; |
f22f54f4 PZ |
1380 | u64 ctrl_val, mask; |
1381 | ||
1382 | mask = 0xfULL << (idx * 4); | |
1383 | ||
1384 | rdmsrl(hwc->config_base, ctrl_val); | |
1385 | ctrl_val &= ~mask; | |
7645a24c | 1386 | wrmsrl(hwc->config_base, ctrl_val); |
f22f54f4 PZ |
1387 | } |
1388 | ||
2b9e344d PZ |
1389 | static inline bool event_is_checkpointed(struct perf_event *event) |
1390 | { | |
1391 | return (event->hw.config & HSW_IN_TX_CHECKPOINTED) != 0; | |
1392 | } | |
1393 | ||
ca037701 | 1394 | static void intel_pmu_disable_event(struct perf_event *event) |
f22f54f4 | 1395 | { |
aff3d91a | 1396 | struct hw_perf_event *hwc = &event->hw; |
89cbc767 | 1397 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
aff3d91a | 1398 | |
15c7ad51 | 1399 | if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) { |
f22f54f4 PZ |
1400 | intel_pmu_disable_bts(); |
1401 | intel_pmu_drain_bts_buffer(); | |
1402 | return; | |
1403 | } | |
1404 | ||
144d31e6 GN |
1405 | cpuc->intel_ctrl_guest_mask &= ~(1ull << hwc->idx); |
1406 | cpuc->intel_ctrl_host_mask &= ~(1ull << hwc->idx); | |
2b9e344d | 1407 | cpuc->intel_cp_status &= ~(1ull << hwc->idx); |
144d31e6 | 1408 | |
60ce0fbd SE |
1409 | /* |
1410 | * must disable before any actual event | |
1411 | * because any event may be combined with LBR | |
1412 | */ | |
a46a2300 | 1413 | if (needs_branch_stack(event)) |
60ce0fbd SE |
1414 | intel_pmu_lbr_disable(event); |
1415 | ||
f22f54f4 | 1416 | if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { |
aff3d91a | 1417 | intel_pmu_disable_fixed(hwc); |
f22f54f4 PZ |
1418 | return; |
1419 | } | |
1420 | ||
aff3d91a | 1421 | x86_pmu_disable_event(event); |
ca037701 | 1422 | |
ab608344 | 1423 | if (unlikely(event->attr.precise_ip)) |
ef21f683 | 1424 | intel_pmu_pebs_disable(event); |
f22f54f4 PZ |
1425 | } |
1426 | ||
ca037701 | 1427 | static void intel_pmu_enable_fixed(struct hw_perf_event *hwc) |
f22f54f4 | 1428 | { |
15c7ad51 | 1429 | int idx = hwc->idx - INTEL_PMC_IDX_FIXED; |
f22f54f4 | 1430 | u64 ctrl_val, bits, mask; |
f22f54f4 PZ |
1431 | |
1432 | /* | |
1433 | * Enable IRQ generation (0x8), | |
1434 | * and enable ring-3 counting (0x2) and ring-0 counting (0x1) | |
1435 | * if requested: | |
1436 | */ | |
1437 | bits = 0x8ULL; | |
1438 | if (hwc->config & ARCH_PERFMON_EVENTSEL_USR) | |
1439 | bits |= 0x2; | |
1440 | if (hwc->config & ARCH_PERFMON_EVENTSEL_OS) | |
1441 | bits |= 0x1; | |
1442 | ||
1443 | /* | |
1444 | * ANY bit is supported in v3 and up | |
1445 | */ | |
1446 | if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY) | |
1447 | bits |= 0x4; | |
1448 | ||
1449 | bits <<= (idx * 4); | |
1450 | mask = 0xfULL << (idx * 4); | |
1451 | ||
1452 | rdmsrl(hwc->config_base, ctrl_val); | |
1453 | ctrl_val &= ~mask; | |
1454 | ctrl_val |= bits; | |
7645a24c | 1455 | wrmsrl(hwc->config_base, ctrl_val); |
f22f54f4 PZ |
1456 | } |
1457 | ||
aff3d91a | 1458 | static void intel_pmu_enable_event(struct perf_event *event) |
f22f54f4 | 1459 | { |
aff3d91a | 1460 | struct hw_perf_event *hwc = &event->hw; |
89cbc767 | 1461 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
aff3d91a | 1462 | |
15c7ad51 | 1463 | if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) { |
0a3aee0d | 1464 | if (!__this_cpu_read(cpu_hw_events.enabled)) |
f22f54f4 PZ |
1465 | return; |
1466 | ||
1467 | intel_pmu_enable_bts(hwc->config); | |
1468 | return; | |
1469 | } | |
60ce0fbd SE |
1470 | /* |
1471 | * must enabled before any actual event | |
1472 | * because any event may be combined with LBR | |
1473 | */ | |
a46a2300 | 1474 | if (needs_branch_stack(event)) |
60ce0fbd | 1475 | intel_pmu_lbr_enable(event); |
f22f54f4 | 1476 | |
144d31e6 GN |
1477 | if (event->attr.exclude_host) |
1478 | cpuc->intel_ctrl_guest_mask |= (1ull << hwc->idx); | |
1479 | if (event->attr.exclude_guest) | |
1480 | cpuc->intel_ctrl_host_mask |= (1ull << hwc->idx); | |
1481 | ||
2b9e344d PZ |
1482 | if (unlikely(event_is_checkpointed(event))) |
1483 | cpuc->intel_cp_status |= (1ull << hwc->idx); | |
1484 | ||
f22f54f4 | 1485 | if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { |
aff3d91a | 1486 | intel_pmu_enable_fixed(hwc); |
f22f54f4 PZ |
1487 | return; |
1488 | } | |
1489 | ||
ab608344 | 1490 | if (unlikely(event->attr.precise_ip)) |
ef21f683 | 1491 | intel_pmu_pebs_enable(event); |
ca037701 | 1492 | |
31fa58af | 1493 | __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); |
f22f54f4 PZ |
1494 | } |
1495 | ||
1496 | /* | |
1497 | * Save and restart an expired event. Called by NMI contexts, | |
1498 | * so it has to be careful about preempting normal event ops: | |
1499 | */ | |
de0428a7 | 1500 | int intel_pmu_save_and_restart(struct perf_event *event) |
f22f54f4 | 1501 | { |
cc2ad4ba | 1502 | x86_perf_event_update(event); |
2dbf0116 AK |
1503 | /* |
1504 | * For a checkpointed counter always reset back to 0. This | |
1505 | * avoids a situation where the counter overflows, aborts the | |
1506 | * transaction and is then set back to shortly before the | |
1507 | * overflow, and overflows and aborts again. | |
1508 | */ | |
1509 | if (unlikely(event_is_checkpointed(event))) { | |
1510 | /* No race with NMIs because the counter should not be armed */ | |
1511 | wrmsrl(event->hw.event_base, 0); | |
1512 | local64_set(&event->hw.prev_count, 0); | |
1513 | } | |
cc2ad4ba | 1514 | return x86_perf_event_set_period(event); |
f22f54f4 PZ |
1515 | } |
1516 | ||
1517 | static void intel_pmu_reset(void) | |
1518 | { | |
0a3aee0d | 1519 | struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds); |
f22f54f4 PZ |
1520 | unsigned long flags; |
1521 | int idx; | |
1522 | ||
948b1bb8 | 1523 | if (!x86_pmu.num_counters) |
f22f54f4 PZ |
1524 | return; |
1525 | ||
1526 | local_irq_save(flags); | |
1527 | ||
c767a54b | 1528 | pr_info("clearing PMU state on CPU#%d\n", smp_processor_id()); |
f22f54f4 | 1529 | |
948b1bb8 | 1530 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
715c85b1 PA |
1531 | wrmsrl_safe(x86_pmu_config_addr(idx), 0ull); |
1532 | wrmsrl_safe(x86_pmu_event_addr(idx), 0ull); | |
f22f54f4 | 1533 | } |
948b1bb8 | 1534 | for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) |
715c85b1 | 1535 | wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull); |
948b1bb8 | 1536 | |
f22f54f4 PZ |
1537 | if (ds) |
1538 | ds->bts_index = ds->bts_buffer_base; | |
1539 | ||
1540 | local_irq_restore(flags); | |
1541 | } | |
1542 | ||
1543 | /* | |
1544 | * This handler is triggered by the local APIC, so the APIC IRQ handling | |
1545 | * rules apply: | |
1546 | */ | |
1547 | static int intel_pmu_handle_irq(struct pt_regs *regs) | |
1548 | { | |
1549 | struct perf_sample_data data; | |
1550 | struct cpu_hw_events *cpuc; | |
1551 | int bit, loops; | |
2e556b5b | 1552 | u64 status; |
b0b2072d | 1553 | int handled; |
f22f54f4 | 1554 | |
89cbc767 | 1555 | cpuc = this_cpu_ptr(&cpu_hw_events); |
f22f54f4 | 1556 | |
2bce5dac | 1557 | /* |
72db5596 AK |
1558 | * No known reason to not always do late ACK, |
1559 | * but just in case do it opt-in. | |
2bce5dac | 1560 | */ |
72db5596 AK |
1561 | if (!x86_pmu.late_ack) |
1562 | apic_write(APIC_LVTPC, APIC_DM_NMI); | |
3fb2b8dd | 1563 | intel_pmu_disable_all(); |
b0b2072d | 1564 | handled = intel_pmu_drain_bts_buffer(); |
8062382c | 1565 | handled += intel_bts_interrupt(); |
f22f54f4 | 1566 | status = intel_pmu_get_status(); |
a3ef2229 MM |
1567 | if (!status) |
1568 | goto done; | |
f22f54f4 PZ |
1569 | |
1570 | loops = 0; | |
1571 | again: | |
2e556b5b | 1572 | intel_pmu_ack_status(status); |
f22f54f4 | 1573 | if (++loops > 100) { |
ae0def05 DH |
1574 | static bool warned = false; |
1575 | if (!warned) { | |
1576 | WARN(1, "perfevents: irq loop stuck!\n"); | |
1577 | perf_event_print_debug(); | |
1578 | warned = true; | |
1579 | } | |
f22f54f4 | 1580 | intel_pmu_reset(); |
3fb2b8dd | 1581 | goto done; |
f22f54f4 PZ |
1582 | } |
1583 | ||
1584 | inc_irq_stat(apic_perf_irqs); | |
ca037701 | 1585 | |
caff2bef PZ |
1586 | intel_pmu_lbr_read(); |
1587 | ||
b292d7a1 HD |
1588 | /* |
1589 | * CondChgd bit 63 doesn't mean any overflow status. Ignore | |
1590 | * and clear the bit. | |
1591 | */ | |
1592 | if (__test_and_clear_bit(63, (unsigned long *)&status)) { | |
1593 | if (!status) | |
1594 | goto done; | |
1595 | } | |
1596 | ||
ca037701 PZ |
1597 | /* |
1598 | * PEBS overflow sets bit 62 in the global status register | |
1599 | */ | |
de725dec PZ |
1600 | if (__test_and_clear_bit(62, (unsigned long *)&status)) { |
1601 | handled++; | |
ca037701 | 1602 | x86_pmu.drain_pebs(regs); |
de725dec | 1603 | } |
ca037701 | 1604 | |
52ca9ced AS |
1605 | /* |
1606 | * Intel PT | |
1607 | */ | |
1608 | if (__test_and_clear_bit(55, (unsigned long *)&status)) { | |
1609 | handled++; | |
1610 | intel_pt_interrupt(); | |
1611 | } | |
1612 | ||
2dbf0116 | 1613 | /* |
2b9e344d PZ |
1614 | * Checkpointed counters can lead to 'spurious' PMIs because the |
1615 | * rollback caused by the PMI will have cleared the overflow status | |
1616 | * bit. Therefore always force probe these counters. | |
2dbf0116 | 1617 | */ |
2b9e344d | 1618 | status |= cpuc->intel_cp_status; |
2dbf0116 | 1619 | |
984b3f57 | 1620 | for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) { |
f22f54f4 PZ |
1621 | struct perf_event *event = cpuc->events[bit]; |
1622 | ||
de725dec PZ |
1623 | handled++; |
1624 | ||
f22f54f4 PZ |
1625 | if (!test_bit(bit, cpuc->active_mask)) |
1626 | continue; | |
1627 | ||
1628 | if (!intel_pmu_save_and_restart(event)) | |
1629 | continue; | |
1630 | ||
fd0d000b | 1631 | perf_sample_data_init(&data, 0, event->hw.last_period); |
f22f54f4 | 1632 | |
60ce0fbd SE |
1633 | if (has_branch_stack(event)) |
1634 | data.br_stack = &cpuc->lbr_stack; | |
1635 | ||
a8b0ca17 | 1636 | if (perf_event_overflow(event, &data, regs)) |
a4eaf7f1 | 1637 | x86_pmu_stop(event, 0); |
f22f54f4 PZ |
1638 | } |
1639 | ||
f22f54f4 PZ |
1640 | /* |
1641 | * Repeat if there is more work to be done: | |
1642 | */ | |
1643 | status = intel_pmu_get_status(); | |
1644 | if (status) | |
1645 | goto again; | |
1646 | ||
3fb2b8dd | 1647 | done: |
11164cd4 | 1648 | intel_pmu_enable_all(0); |
72db5596 AK |
1649 | /* |
1650 | * Only unmask the NMI after the overflow counters | |
1651 | * have been reset. This avoids spurious NMIs on | |
1652 | * Haswell CPUs. | |
1653 | */ | |
1654 | if (x86_pmu.late_ack) | |
1655 | apic_write(APIC_LVTPC, APIC_DM_NMI); | |
de725dec | 1656 | return handled; |
f22f54f4 PZ |
1657 | } |
1658 | ||
f22f54f4 | 1659 | static struct event_constraint * |
ca037701 | 1660 | intel_bts_constraints(struct perf_event *event) |
f22f54f4 | 1661 | { |
ca037701 PZ |
1662 | struct hw_perf_event *hwc = &event->hw; |
1663 | unsigned int hw_event, bts_event; | |
f22f54f4 | 1664 | |
18a073a3 PZ |
1665 | if (event->attr.freq) |
1666 | return NULL; | |
1667 | ||
ca037701 PZ |
1668 | hw_event = hwc->config & INTEL_ARCH_EVENT_MASK; |
1669 | bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS); | |
f22f54f4 | 1670 | |
ca037701 | 1671 | if (unlikely(hw_event == bts_event && hwc->sample_period == 1)) |
f22f54f4 | 1672 | return &bts_constraint; |
ca037701 | 1673 | |
f22f54f4 PZ |
1674 | return NULL; |
1675 | } | |
1676 | ||
5a425294 | 1677 | static int intel_alt_er(int idx) |
b79e8941 | 1678 | { |
9a5e3fb5 | 1679 | if (!(x86_pmu.flags & PMU_FL_HAS_RSP_1)) |
5a425294 | 1680 | return idx; |
b79e8941 | 1681 | |
5a425294 PZ |
1682 | if (idx == EXTRA_REG_RSP_0) |
1683 | return EXTRA_REG_RSP_1; | |
1684 | ||
1685 | if (idx == EXTRA_REG_RSP_1) | |
1686 | return EXTRA_REG_RSP_0; | |
1687 | ||
1688 | return idx; | |
1689 | } | |
1690 | ||
1691 | static void intel_fixup_er(struct perf_event *event, int idx) | |
1692 | { | |
1693 | event->hw.extra_reg.idx = idx; | |
1694 | ||
1695 | if (idx == EXTRA_REG_RSP_0) { | |
b79e8941 | 1696 | event->hw.config &= ~INTEL_ARCH_EVENT_MASK; |
53ad0447 | 1697 | event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_0].event; |
b79e8941 | 1698 | event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0; |
5a425294 PZ |
1699 | } else if (idx == EXTRA_REG_RSP_1) { |
1700 | event->hw.config &= ~INTEL_ARCH_EVENT_MASK; | |
53ad0447 | 1701 | event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_1].event; |
5a425294 | 1702 | event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1; |
b79e8941 | 1703 | } |
b79e8941 PZ |
1704 | } |
1705 | ||
efc9f05d SE |
1706 | /* |
1707 | * manage allocation of shared extra msr for certain events | |
1708 | * | |
1709 | * sharing can be: | |
1710 | * per-cpu: to be shared between the various events on a single PMU | |
1711 | * per-core: per-cpu + shared by HT threads | |
1712 | */ | |
a7e3ed1e | 1713 | static struct event_constraint * |
efc9f05d | 1714 | __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc, |
b36817e8 SE |
1715 | struct perf_event *event, |
1716 | struct hw_perf_event_extra *reg) | |
a7e3ed1e | 1717 | { |
efc9f05d | 1718 | struct event_constraint *c = &emptyconstraint; |
a7e3ed1e | 1719 | struct er_account *era; |
cd8a38d3 | 1720 | unsigned long flags; |
5a425294 | 1721 | int idx = reg->idx; |
a7e3ed1e | 1722 | |
5a425294 PZ |
1723 | /* |
1724 | * reg->alloc can be set due to existing state, so for fake cpuc we | |
1725 | * need to ignore this, otherwise we might fail to allocate proper fake | |
1726 | * state for this extra reg constraint. Also see the comment below. | |
1727 | */ | |
1728 | if (reg->alloc && !cpuc->is_fake) | |
b36817e8 | 1729 | return NULL; /* call x86_get_event_constraint() */ |
a7e3ed1e | 1730 | |
b79e8941 | 1731 | again: |
5a425294 | 1732 | era = &cpuc->shared_regs->regs[idx]; |
cd8a38d3 SE |
1733 | /* |
1734 | * we use spin_lock_irqsave() to avoid lockdep issues when | |
1735 | * passing a fake cpuc | |
1736 | */ | |
1737 | raw_spin_lock_irqsave(&era->lock, flags); | |
efc9f05d SE |
1738 | |
1739 | if (!atomic_read(&era->ref) || era->config == reg->config) { | |
1740 | ||
5a425294 PZ |
1741 | /* |
1742 | * If its a fake cpuc -- as per validate_{group,event}() we | |
1743 | * shouldn't touch event state and we can avoid doing so | |
1744 | * since both will only call get_event_constraints() once | |
1745 | * on each event, this avoids the need for reg->alloc. | |
1746 | * | |
1747 | * Not doing the ER fixup will only result in era->reg being | |
1748 | * wrong, but since we won't actually try and program hardware | |
1749 | * this isn't a problem either. | |
1750 | */ | |
1751 | if (!cpuc->is_fake) { | |
1752 | if (idx != reg->idx) | |
1753 | intel_fixup_er(event, idx); | |
1754 | ||
1755 | /* | |
1756 | * x86_schedule_events() can call get_event_constraints() | |
1757 | * multiple times on events in the case of incremental | |
1758 | * scheduling(). reg->alloc ensures we only do the ER | |
1759 | * allocation once. | |
1760 | */ | |
1761 | reg->alloc = 1; | |
1762 | } | |
1763 | ||
efc9f05d SE |
1764 | /* lock in msr value */ |
1765 | era->config = reg->config; | |
1766 | era->reg = reg->reg; | |
1767 | ||
1768 | /* one more user */ | |
1769 | atomic_inc(&era->ref); | |
1770 | ||
a7e3ed1e | 1771 | /* |
b36817e8 SE |
1772 | * need to call x86_get_event_constraint() |
1773 | * to check if associated event has constraints | |
a7e3ed1e | 1774 | */ |
b36817e8 | 1775 | c = NULL; |
5a425294 PZ |
1776 | } else { |
1777 | idx = intel_alt_er(idx); | |
1778 | if (idx != reg->idx) { | |
1779 | raw_spin_unlock_irqrestore(&era->lock, flags); | |
1780 | goto again; | |
1781 | } | |
a7e3ed1e | 1782 | } |
cd8a38d3 | 1783 | raw_spin_unlock_irqrestore(&era->lock, flags); |
a7e3ed1e | 1784 | |
efc9f05d SE |
1785 | return c; |
1786 | } | |
1787 | ||
1788 | static void | |
1789 | __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc, | |
1790 | struct hw_perf_event_extra *reg) | |
1791 | { | |
1792 | struct er_account *era; | |
1793 | ||
1794 | /* | |
5a425294 PZ |
1795 | * Only put constraint if extra reg was actually allocated. Also takes |
1796 | * care of event which do not use an extra shared reg. | |
1797 | * | |
1798 | * Also, if this is a fake cpuc we shouldn't touch any event state | |
1799 | * (reg->alloc) and we don't care about leaving inconsistent cpuc state | |
1800 | * either since it'll be thrown out. | |
efc9f05d | 1801 | */ |
5a425294 | 1802 | if (!reg->alloc || cpuc->is_fake) |
efc9f05d SE |
1803 | return; |
1804 | ||
1805 | era = &cpuc->shared_regs->regs[reg->idx]; | |
1806 | ||
1807 | /* one fewer user */ | |
1808 | atomic_dec(&era->ref); | |
1809 | ||
1810 | /* allocate again next time */ | |
1811 | reg->alloc = 0; | |
1812 | } | |
1813 | ||
1814 | static struct event_constraint * | |
1815 | intel_shared_regs_constraints(struct cpu_hw_events *cpuc, | |
1816 | struct perf_event *event) | |
1817 | { | |
b36817e8 SE |
1818 | struct event_constraint *c = NULL, *d; |
1819 | struct hw_perf_event_extra *xreg, *breg; | |
1820 | ||
1821 | xreg = &event->hw.extra_reg; | |
1822 | if (xreg->idx != EXTRA_REG_NONE) { | |
1823 | c = __intel_shared_reg_get_constraints(cpuc, event, xreg); | |
1824 | if (c == &emptyconstraint) | |
1825 | return c; | |
1826 | } | |
1827 | breg = &event->hw.branch_reg; | |
1828 | if (breg->idx != EXTRA_REG_NONE) { | |
1829 | d = __intel_shared_reg_get_constraints(cpuc, event, breg); | |
1830 | if (d == &emptyconstraint) { | |
1831 | __intel_shared_reg_put_constraints(cpuc, xreg); | |
1832 | c = d; | |
1833 | } | |
1834 | } | |
efc9f05d | 1835 | return c; |
a7e3ed1e AK |
1836 | } |
1837 | ||
de0428a7 | 1838 | struct event_constraint * |
79cba822 SE |
1839 | x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx, |
1840 | struct perf_event *event) | |
de0428a7 KW |
1841 | { |
1842 | struct event_constraint *c; | |
1843 | ||
1844 | if (x86_pmu.event_constraints) { | |
1845 | for_each_event_constraint(c, x86_pmu.event_constraints) { | |
9fac2cf3 | 1846 | if ((event->hw.config & c->cmask) == c->code) { |
9fac2cf3 | 1847 | event->hw.flags |= c->flags; |
de0428a7 | 1848 | return c; |
9fac2cf3 | 1849 | } |
de0428a7 KW |
1850 | } |
1851 | } | |
1852 | ||
1853 | return &unconstrained; | |
1854 | } | |
1855 | ||
f22f54f4 | 1856 | static struct event_constraint * |
e979121b | 1857 | __intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx, |
79cba822 | 1858 | struct perf_event *event) |
f22f54f4 PZ |
1859 | { |
1860 | struct event_constraint *c; | |
1861 | ||
ca037701 PZ |
1862 | c = intel_bts_constraints(event); |
1863 | if (c) | |
1864 | return c; | |
1865 | ||
687805e4 | 1866 | c = intel_shared_regs_constraints(cpuc, event); |
f22f54f4 PZ |
1867 | if (c) |
1868 | return c; | |
1869 | ||
687805e4 | 1870 | c = intel_pebs_constraints(event); |
a7e3ed1e AK |
1871 | if (c) |
1872 | return c; | |
1873 | ||
79cba822 | 1874 | return x86_get_event_constraints(cpuc, idx, event); |
f22f54f4 PZ |
1875 | } |
1876 | ||
e979121b MD |
1877 | static void |
1878 | intel_start_scheduling(struct cpu_hw_events *cpuc) | |
1879 | { | |
1880 | struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; | |
1881 | struct intel_excl_states *xl, *xlo; | |
1882 | int tid = cpuc->excl_thread_id; | |
1883 | int o_tid = 1 - tid; /* sibling thread */ | |
1884 | ||
1885 | /* | |
1886 | * nothing needed if in group validation mode | |
1887 | */ | |
1888 | if (cpuc->is_fake) | |
1889 | return; | |
1890 | /* | |
1891 | * no exclusion needed | |
1892 | */ | |
1893 | if (!excl_cntrs) | |
1894 | return; | |
1895 | ||
1896 | xlo = &excl_cntrs->states[o_tid]; | |
1897 | xl = &excl_cntrs->states[tid]; | |
1898 | ||
1899 | xl->sched_started = true; | |
1900 | ||
1901 | /* | |
1902 | * lock shared state until we are done scheduling | |
1903 | * in stop_event_scheduling() | |
1904 | * makes scheduling appear as a transaction | |
1905 | */ | |
1906 | WARN_ON_ONCE(!irqs_disabled()); | |
1907 | raw_spin_lock(&excl_cntrs->lock); | |
1908 | ||
1909 | /* | |
1910 | * save initial state of sibling thread | |
1911 | */ | |
1912 | memcpy(xlo->init_state, xlo->state, sizeof(xlo->init_state)); | |
1913 | } | |
1914 | ||
1915 | static void | |
1916 | intel_stop_scheduling(struct cpu_hw_events *cpuc) | |
1917 | { | |
1918 | struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; | |
1919 | struct intel_excl_states *xl, *xlo; | |
1920 | int tid = cpuc->excl_thread_id; | |
1921 | int o_tid = 1 - tid; /* sibling thread */ | |
1922 | ||
1923 | /* | |
1924 | * nothing needed if in group validation mode | |
1925 | */ | |
1926 | if (cpuc->is_fake) | |
1927 | return; | |
1928 | /* | |
1929 | * no exclusion needed | |
1930 | */ | |
1931 | if (!excl_cntrs) | |
1932 | return; | |
1933 | ||
1934 | xlo = &excl_cntrs->states[o_tid]; | |
1935 | xl = &excl_cntrs->states[tid]; | |
1936 | ||
1937 | /* | |
1938 | * make new sibling thread state visible | |
1939 | */ | |
1940 | memcpy(xlo->state, xlo->init_state, sizeof(xlo->state)); | |
1941 | ||
1942 | xl->sched_started = false; | |
1943 | /* | |
1944 | * release shared state lock (acquired in intel_start_scheduling()) | |
1945 | */ | |
1946 | raw_spin_unlock(&excl_cntrs->lock); | |
1947 | } | |
1948 | ||
1949 | static struct event_constraint * | |
1950 | intel_get_excl_constraints(struct cpu_hw_events *cpuc, struct perf_event *event, | |
1951 | int idx, struct event_constraint *c) | |
1952 | { | |
1953 | struct event_constraint *cx; | |
1954 | struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; | |
1955 | struct intel_excl_states *xl, *xlo; | |
1956 | int is_excl, i; | |
1957 | int tid = cpuc->excl_thread_id; | |
1958 | int o_tid = 1 - tid; /* alternate */ | |
1959 | ||
1960 | /* | |
1961 | * validating a group does not require | |
1962 | * enforcing cross-thread exclusion | |
1963 | */ | |
1964 | if (cpuc->is_fake) | |
1965 | return c; | |
1966 | ||
1967 | /* | |
1968 | * event requires exclusive counter access | |
1969 | * across HT threads | |
1970 | */ | |
1971 | is_excl = c->flags & PERF_X86_EVENT_EXCL; | |
1972 | ||
1973 | /* | |
1974 | * xl = state of current HT | |
1975 | * xlo = state of sibling HT | |
1976 | */ | |
1977 | xl = &excl_cntrs->states[tid]; | |
1978 | xlo = &excl_cntrs->states[o_tid]; | |
1979 | ||
1980 | cx = c; | |
1981 | ||
1982 | /* | |
1983 | * because we modify the constraint, we need | |
1984 | * to make a copy. Static constraints come | |
1985 | * from static const tables. | |
1986 | * | |
1987 | * only needed when constraint has not yet | |
1988 | * been cloned (marked dynamic) | |
1989 | */ | |
1990 | if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) { | |
1991 | ||
1992 | /* sanity check */ | |
1993 | if (idx < 0) | |
1994 | return &emptyconstraint; | |
1995 | ||
1996 | /* | |
1997 | * grab pre-allocated constraint entry | |
1998 | */ | |
1999 | cx = &cpuc->constraint_list[idx]; | |
2000 | ||
2001 | /* | |
2002 | * initialize dynamic constraint | |
2003 | * with static constraint | |
2004 | */ | |
2005 | memcpy(cx, c, sizeof(*cx)); | |
2006 | ||
2007 | /* | |
2008 | * mark constraint as dynamic, so we | |
2009 | * can free it later on | |
2010 | */ | |
2011 | cx->flags |= PERF_X86_EVENT_DYNAMIC; | |
2012 | } | |
2013 | ||
2014 | /* | |
2015 | * From here on, the constraint is dynamic. | |
2016 | * Either it was just allocated above, or it | |
2017 | * was allocated during a earlier invocation | |
2018 | * of this function | |
2019 | */ | |
2020 | ||
2021 | /* | |
2022 | * Modify static constraint with current dynamic | |
2023 | * state of thread | |
2024 | * | |
2025 | * EXCLUSIVE: sibling counter measuring exclusive event | |
2026 | * SHARED : sibling counter measuring non-exclusive event | |
2027 | * UNUSED : sibling counter unused | |
2028 | */ | |
2029 | for_each_set_bit(i, cx->idxmsk, X86_PMC_IDX_MAX) { | |
2030 | /* | |
2031 | * exclusive event in sibling counter | |
2032 | * our corresponding counter cannot be used | |
2033 | * regardless of our event | |
2034 | */ | |
2035 | if (xl->state[i] == INTEL_EXCL_EXCLUSIVE) | |
2036 | __clear_bit(i, cx->idxmsk); | |
2037 | /* | |
2038 | * if measuring an exclusive event, sibling | |
2039 | * measuring non-exclusive, then counter cannot | |
2040 | * be used | |
2041 | */ | |
2042 | if (is_excl && xl->state[i] == INTEL_EXCL_SHARED) | |
2043 | __clear_bit(i, cx->idxmsk); | |
2044 | } | |
2045 | ||
2046 | /* | |
2047 | * recompute actual bit weight for scheduling algorithm | |
2048 | */ | |
2049 | cx->weight = hweight64(cx->idxmsk64); | |
2050 | ||
2051 | /* | |
2052 | * if we return an empty mask, then switch | |
2053 | * back to static empty constraint to avoid | |
2054 | * the cost of freeing later on | |
2055 | */ | |
2056 | if (cx->weight == 0) | |
2057 | cx = &emptyconstraint; | |
2058 | ||
2059 | return cx; | |
2060 | } | |
2061 | ||
2062 | static struct event_constraint * | |
2063 | intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx, | |
2064 | struct perf_event *event) | |
2065 | { | |
2066 | struct event_constraint *c = event->hw.constraint; | |
2067 | ||
2068 | /* | |
2069 | * first time only | |
2070 | * - static constraint: no change across incremental scheduling calls | |
2071 | * - dynamic constraint: handled by intel_get_excl_constraints() | |
2072 | */ | |
2073 | if (!c) | |
2074 | c = __intel_get_event_constraints(cpuc, idx, event); | |
2075 | ||
2076 | if (cpuc->excl_cntrs) | |
2077 | return intel_get_excl_constraints(cpuc, event, idx, c); | |
2078 | ||
2079 | return c; | |
2080 | } | |
2081 | ||
2082 | static void intel_put_excl_constraints(struct cpu_hw_events *cpuc, | |
2083 | struct perf_event *event) | |
2084 | { | |
2085 | struct hw_perf_event *hwc = &event->hw; | |
2086 | struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; | |
2087 | struct intel_excl_states *xlo, *xl; | |
2088 | unsigned long flags = 0; /* keep compiler happy */ | |
2089 | int tid = cpuc->excl_thread_id; | |
2090 | int o_tid = 1 - tid; | |
2091 | ||
2092 | /* | |
2093 | * nothing needed if in group validation mode | |
2094 | */ | |
2095 | if (cpuc->is_fake) | |
2096 | return; | |
2097 | ||
2098 | WARN_ON_ONCE(!excl_cntrs); | |
2099 | ||
2100 | if (!excl_cntrs) | |
2101 | return; | |
2102 | ||
2103 | xl = &excl_cntrs->states[tid]; | |
2104 | xlo = &excl_cntrs->states[o_tid]; | |
2105 | ||
2106 | /* | |
2107 | * put_constraint may be called from x86_schedule_events() | |
2108 | * which already has the lock held so here make locking | |
2109 | * conditional | |
2110 | */ | |
2111 | if (!xl->sched_started) | |
2112 | raw_spin_lock_irqsave(&excl_cntrs->lock, flags); | |
2113 | ||
2114 | /* | |
2115 | * if event was actually assigned, then mark the | |
2116 | * counter state as unused now | |
2117 | */ | |
2118 | if (hwc->idx >= 0) | |
2119 | xlo->state[hwc->idx] = INTEL_EXCL_UNUSED; | |
2120 | ||
2121 | if (!xl->sched_started) | |
2122 | raw_spin_unlock_irqrestore(&excl_cntrs->lock, flags); | |
2123 | } | |
2124 | ||
efc9f05d SE |
2125 | static void |
2126 | intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc, | |
a7e3ed1e AK |
2127 | struct perf_event *event) |
2128 | { | |
efc9f05d | 2129 | struct hw_perf_event_extra *reg; |
a7e3ed1e | 2130 | |
efc9f05d SE |
2131 | reg = &event->hw.extra_reg; |
2132 | if (reg->idx != EXTRA_REG_NONE) | |
2133 | __intel_shared_reg_put_constraints(cpuc, reg); | |
b36817e8 SE |
2134 | |
2135 | reg = &event->hw.branch_reg; | |
2136 | if (reg->idx != EXTRA_REG_NONE) | |
2137 | __intel_shared_reg_put_constraints(cpuc, reg); | |
efc9f05d | 2138 | } |
a7e3ed1e | 2139 | |
efc9f05d SE |
2140 | static void intel_put_event_constraints(struct cpu_hw_events *cpuc, |
2141 | struct perf_event *event) | |
2142 | { | |
e979121b MD |
2143 | struct event_constraint *c = event->hw.constraint; |
2144 | ||
efc9f05d | 2145 | intel_put_shared_regs_event_constraints(cpuc, event); |
e979121b MD |
2146 | |
2147 | /* | |
2148 | * is PMU has exclusive counter restrictions, then | |
2149 | * all events are subject to and must call the | |
2150 | * put_excl_constraints() routine | |
2151 | */ | |
2152 | if (c && cpuc->excl_cntrs) | |
2153 | intel_put_excl_constraints(cpuc, event); | |
2154 | ||
2155 | /* cleanup dynamic constraint */ | |
2156 | if (c && (c->flags & PERF_X86_EVENT_DYNAMIC)) | |
2157 | event->hw.constraint = NULL; | |
2158 | } | |
2159 | ||
2160 | static void intel_commit_scheduling(struct cpu_hw_events *cpuc, | |
2161 | struct perf_event *event, int cntr) | |
2162 | { | |
2163 | struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; | |
2164 | struct event_constraint *c = event->hw.constraint; | |
2165 | struct intel_excl_states *xlo, *xl; | |
2166 | int tid = cpuc->excl_thread_id; | |
2167 | int o_tid = 1 - tid; | |
2168 | int is_excl; | |
2169 | ||
2170 | if (cpuc->is_fake || !c) | |
2171 | return; | |
2172 | ||
2173 | is_excl = c->flags & PERF_X86_EVENT_EXCL; | |
2174 | ||
2175 | if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) | |
2176 | return; | |
2177 | ||
2178 | WARN_ON_ONCE(!excl_cntrs); | |
2179 | ||
2180 | if (!excl_cntrs) | |
2181 | return; | |
2182 | ||
2183 | xl = &excl_cntrs->states[tid]; | |
2184 | xlo = &excl_cntrs->states[o_tid]; | |
2185 | ||
2186 | WARN_ON_ONCE(!raw_spin_is_locked(&excl_cntrs->lock)); | |
2187 | ||
2188 | if (cntr >= 0) { | |
2189 | if (is_excl) | |
2190 | xlo->init_state[cntr] = INTEL_EXCL_EXCLUSIVE; | |
2191 | else | |
2192 | xlo->init_state[cntr] = INTEL_EXCL_SHARED; | |
2193 | } | |
a7e3ed1e AK |
2194 | } |
2195 | ||
0780c927 | 2196 | static void intel_pebs_aliases_core2(struct perf_event *event) |
b4cdc5c2 | 2197 | { |
0780c927 | 2198 | if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { |
7639dae0 PZ |
2199 | /* |
2200 | * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P | |
2201 | * (0x003c) so that we can use it with PEBS. | |
2202 | * | |
2203 | * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't | |
2204 | * PEBS capable. However we can use INST_RETIRED.ANY_P | |
2205 | * (0x00c0), which is a PEBS capable event, to get the same | |
2206 | * count. | |
2207 | * | |
2208 | * INST_RETIRED.ANY_P counts the number of cycles that retires | |
2209 | * CNTMASK instructions. By setting CNTMASK to a value (16) | |
2210 | * larger than the maximum number of instructions that can be | |
2211 | * retired per cycle (4) and then inverting the condition, we | |
2212 | * count all cycles that retire 16 or less instructions, which | |
2213 | * is every cycle. | |
2214 | * | |
2215 | * Thereby we gain a PEBS capable cycle counter. | |
2216 | */ | |
f9b4eeb8 PZ |
2217 | u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16); |
2218 | ||
0780c927 PZ |
2219 | alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); |
2220 | event->hw.config = alt_config; | |
2221 | } | |
2222 | } | |
2223 | ||
2224 | static void intel_pebs_aliases_snb(struct perf_event *event) | |
2225 | { | |
2226 | if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { | |
2227 | /* | |
2228 | * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P | |
2229 | * (0x003c) so that we can use it with PEBS. | |
2230 | * | |
2231 | * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't | |
2232 | * PEBS capable. However we can use UOPS_RETIRED.ALL | |
2233 | * (0x01c2), which is a PEBS capable event, to get the same | |
2234 | * count. | |
2235 | * | |
2236 | * UOPS_RETIRED.ALL counts the number of cycles that retires | |
2237 | * CNTMASK micro-ops. By setting CNTMASK to a value (16) | |
2238 | * larger than the maximum number of micro-ops that can be | |
2239 | * retired per cycle (4) and then inverting the condition, we | |
2240 | * count all cycles that retire 16 or less micro-ops, which | |
2241 | * is every cycle. | |
2242 | * | |
2243 | * Thereby we gain a PEBS capable cycle counter. | |
2244 | */ | |
2245 | u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16); | |
7639dae0 PZ |
2246 | |
2247 | alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); | |
2248 | event->hw.config = alt_config; | |
2249 | } | |
0780c927 PZ |
2250 | } |
2251 | ||
2252 | static int intel_pmu_hw_config(struct perf_event *event) | |
2253 | { | |
2254 | int ret = x86_pmu_hw_config(event); | |
2255 | ||
2256 | if (ret) | |
2257 | return ret; | |
2258 | ||
2259 | if (event->attr.precise_ip && x86_pmu.pebs_aliases) | |
2260 | x86_pmu.pebs_aliases(event); | |
7639dae0 | 2261 | |
a46a2300 | 2262 | if (needs_branch_stack(event)) { |
60ce0fbd SE |
2263 | ret = intel_pmu_setup_lbr_filter(event); |
2264 | if (ret) | |
2265 | return ret; | |
48070342 AS |
2266 | |
2267 | /* | |
2268 | * BTS is set up earlier in this path, so don't account twice | |
2269 | */ | |
2270 | if (!intel_pmu_has_bts(event)) { | |
2271 | /* disallow lbr if conflicting events are present */ | |
2272 | if (x86_add_exclusive(x86_lbr_exclusive_lbr)) | |
2273 | return -EBUSY; | |
2274 | ||
2275 | event->destroy = hw_perf_lbr_event_destroy; | |
2276 | } | |
60ce0fbd SE |
2277 | } |
2278 | ||
b4cdc5c2 PZ |
2279 | if (event->attr.type != PERF_TYPE_RAW) |
2280 | return 0; | |
2281 | ||
2282 | if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY)) | |
2283 | return 0; | |
2284 | ||
2285 | if (x86_pmu.version < 3) | |
2286 | return -EINVAL; | |
2287 | ||
2288 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) | |
2289 | return -EACCES; | |
2290 | ||
2291 | event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY; | |
2292 | ||
2293 | return 0; | |
2294 | } | |
2295 | ||
144d31e6 GN |
2296 | struct perf_guest_switch_msr *perf_guest_get_msrs(int *nr) |
2297 | { | |
2298 | if (x86_pmu.guest_get_msrs) | |
2299 | return x86_pmu.guest_get_msrs(nr); | |
2300 | *nr = 0; | |
2301 | return NULL; | |
2302 | } | |
2303 | EXPORT_SYMBOL_GPL(perf_guest_get_msrs); | |
2304 | ||
2305 | static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr) | |
2306 | { | |
89cbc767 | 2307 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
144d31e6 GN |
2308 | struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; |
2309 | ||
2310 | arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL; | |
2311 | arr[0].host = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask; | |
2312 | arr[0].guest = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_host_mask; | |
26a4f3c0 GN |
2313 | /* |
2314 | * If PMU counter has PEBS enabled it is not enough to disable counter | |
2315 | * on a guest entry since PEBS memory write can overshoot guest entry | |
2316 | * and corrupt guest memory. Disabling PEBS solves the problem. | |
2317 | */ | |
2318 | arr[1].msr = MSR_IA32_PEBS_ENABLE; | |
2319 | arr[1].host = cpuc->pebs_enabled; | |
2320 | arr[1].guest = 0; | |
144d31e6 | 2321 | |
26a4f3c0 | 2322 | *nr = 2; |
144d31e6 GN |
2323 | return arr; |
2324 | } | |
2325 | ||
2326 | static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr) | |
2327 | { | |
89cbc767 | 2328 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
144d31e6 GN |
2329 | struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; |
2330 | int idx; | |
2331 | ||
2332 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { | |
2333 | struct perf_event *event = cpuc->events[idx]; | |
2334 | ||
2335 | arr[idx].msr = x86_pmu_config_addr(idx); | |
2336 | arr[idx].host = arr[idx].guest = 0; | |
2337 | ||
2338 | if (!test_bit(idx, cpuc->active_mask)) | |
2339 | continue; | |
2340 | ||
2341 | arr[idx].host = arr[idx].guest = | |
2342 | event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE; | |
2343 | ||
2344 | if (event->attr.exclude_host) | |
2345 | arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE; | |
2346 | else if (event->attr.exclude_guest) | |
2347 | arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE; | |
2348 | } | |
2349 | ||
2350 | *nr = x86_pmu.num_counters; | |
2351 | return arr; | |
2352 | } | |
2353 | ||
2354 | static void core_pmu_enable_event(struct perf_event *event) | |
2355 | { | |
2356 | if (!event->attr.exclude_host) | |
2357 | x86_pmu_enable_event(event); | |
2358 | } | |
2359 | ||
2360 | static void core_pmu_enable_all(int added) | |
2361 | { | |
89cbc767 | 2362 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
144d31e6 GN |
2363 | int idx; |
2364 | ||
2365 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { | |
2366 | struct hw_perf_event *hwc = &cpuc->events[idx]->hw; | |
2367 | ||
2368 | if (!test_bit(idx, cpuc->active_mask) || | |
2369 | cpuc->events[idx]->attr.exclude_host) | |
2370 | continue; | |
2371 | ||
2372 | __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); | |
2373 | } | |
2374 | } | |
2375 | ||
3a632cb2 AK |
2376 | static int hsw_hw_config(struct perf_event *event) |
2377 | { | |
2378 | int ret = intel_pmu_hw_config(event); | |
2379 | ||
2380 | if (ret) | |
2381 | return ret; | |
2382 | if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE)) | |
2383 | return 0; | |
2384 | event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED); | |
2385 | ||
2386 | /* | |
2387 | * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with | |
2388 | * PEBS or in ANY thread mode. Since the results are non-sensical forbid | |
2389 | * this combination. | |
2390 | */ | |
2391 | if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) && | |
2392 | ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) || | |
2393 | event->attr.precise_ip > 0)) | |
2394 | return -EOPNOTSUPP; | |
2395 | ||
2dbf0116 AK |
2396 | if (event_is_checkpointed(event)) { |
2397 | /* | |
2398 | * Sampling of checkpointed events can cause situations where | |
2399 | * the CPU constantly aborts because of a overflow, which is | |
2400 | * then checkpointed back and ignored. Forbid checkpointing | |
2401 | * for sampling. | |
2402 | * | |
2403 | * But still allow a long sampling period, so that perf stat | |
2404 | * from KVM works. | |
2405 | */ | |
2406 | if (event->attr.sample_period > 0 && | |
2407 | event->attr.sample_period < 0x7fffffff) | |
2408 | return -EOPNOTSUPP; | |
2409 | } | |
3a632cb2 AK |
2410 | return 0; |
2411 | } | |
2412 | ||
2413 | static struct event_constraint counter2_constraint = | |
2414 | EVENT_CONSTRAINT(0, 0x4, 0); | |
2415 | ||
2416 | static struct event_constraint * | |
79cba822 SE |
2417 | hsw_get_event_constraints(struct cpu_hw_events *cpuc, int idx, |
2418 | struct perf_event *event) | |
3a632cb2 | 2419 | { |
79cba822 SE |
2420 | struct event_constraint *c; |
2421 | ||
2422 | c = intel_get_event_constraints(cpuc, idx, event); | |
3a632cb2 AK |
2423 | |
2424 | /* Handle special quirk on in_tx_checkpointed only in counter 2 */ | |
2425 | if (event->hw.config & HSW_IN_TX_CHECKPOINTED) { | |
2426 | if (c->idxmsk64 & (1U << 2)) | |
2427 | return &counter2_constraint; | |
2428 | return &emptyconstraint; | |
2429 | } | |
2430 | ||
2431 | return c; | |
2432 | } | |
2433 | ||
294fe0f5 AK |
2434 | /* |
2435 | * Broadwell: | |
2436 | * | |
2437 | * The INST_RETIRED.ALL period always needs to have lowest 6 bits cleared | |
2438 | * (BDM55) and it must not use a period smaller than 100 (BDM11). We combine | |
2439 | * the two to enforce a minimum period of 128 (the smallest value that has bits | |
2440 | * 0-5 cleared and >= 100). | |
2441 | * | |
2442 | * Because of how the code in x86_perf_event_set_period() works, the truncation | |
2443 | * of the lower 6 bits is 'harmless' as we'll occasionally add a longer period | |
2444 | * to make up for the 'lost' events due to carrying the 'error' in period_left. | |
2445 | * | |
2446 | * Therefore the effective (average) period matches the requested period, | |
2447 | * despite coarser hardware granularity. | |
2448 | */ | |
2449 | static unsigned bdw_limit_period(struct perf_event *event, unsigned left) | |
2450 | { | |
2451 | if ((event->hw.config & INTEL_ARCH_EVENT_MASK) == | |
2452 | X86_CONFIG(.event=0xc0, .umask=0x01)) { | |
2453 | if (left < 128) | |
2454 | left = 128; | |
2455 | left &= ~0x3fu; | |
2456 | } | |
2457 | return left; | |
2458 | } | |
2459 | ||
641cc938 JO |
2460 | PMU_FORMAT_ATTR(event, "config:0-7" ); |
2461 | PMU_FORMAT_ATTR(umask, "config:8-15" ); | |
2462 | PMU_FORMAT_ATTR(edge, "config:18" ); | |
2463 | PMU_FORMAT_ATTR(pc, "config:19" ); | |
2464 | PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */ | |
2465 | PMU_FORMAT_ATTR(inv, "config:23" ); | |
2466 | PMU_FORMAT_ATTR(cmask, "config:24-31" ); | |
3a632cb2 AK |
2467 | PMU_FORMAT_ATTR(in_tx, "config:32"); |
2468 | PMU_FORMAT_ATTR(in_tx_cp, "config:33"); | |
641cc938 JO |
2469 | |
2470 | static struct attribute *intel_arch_formats_attr[] = { | |
2471 | &format_attr_event.attr, | |
2472 | &format_attr_umask.attr, | |
2473 | &format_attr_edge.attr, | |
2474 | &format_attr_pc.attr, | |
2475 | &format_attr_inv.attr, | |
2476 | &format_attr_cmask.attr, | |
2477 | NULL, | |
2478 | }; | |
2479 | ||
0bf79d44 JO |
2480 | ssize_t intel_event_sysfs_show(char *page, u64 config) |
2481 | { | |
2482 | u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT); | |
2483 | ||
2484 | return x86_event_sysfs_show(page, config, event); | |
2485 | } | |
2486 | ||
caaa8be3 | 2487 | static __initconst const struct x86_pmu core_pmu = { |
f22f54f4 PZ |
2488 | .name = "core", |
2489 | .handle_irq = x86_pmu_handle_irq, | |
2490 | .disable_all = x86_pmu_disable_all, | |
144d31e6 GN |
2491 | .enable_all = core_pmu_enable_all, |
2492 | .enable = core_pmu_enable_event, | |
f22f54f4 | 2493 | .disable = x86_pmu_disable_event, |
b4cdc5c2 | 2494 | .hw_config = x86_pmu_hw_config, |
a072738e | 2495 | .schedule_events = x86_schedule_events, |
f22f54f4 PZ |
2496 | .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, |
2497 | .perfctr = MSR_ARCH_PERFMON_PERFCTR0, | |
2498 | .event_map = intel_pmu_event_map, | |
f22f54f4 PZ |
2499 | .max_events = ARRAY_SIZE(intel_perfmon_event_map), |
2500 | .apic = 1, | |
2501 | /* | |
2502 | * Intel PMCs cannot be accessed sanely above 32 bit width, | |
2503 | * so we install an artificial 1<<31 period regardless of | |
2504 | * the generic event period: | |
2505 | */ | |
2506 | .max_period = (1ULL << 31) - 1, | |
2507 | .get_event_constraints = intel_get_event_constraints, | |
a7e3ed1e | 2508 | .put_event_constraints = intel_put_event_constraints, |
f22f54f4 | 2509 | .event_constraints = intel_core_event_constraints, |
144d31e6 | 2510 | .guest_get_msrs = core_guest_get_msrs, |
641cc938 | 2511 | .format_attrs = intel_arch_formats_attr, |
0bf79d44 | 2512 | .events_sysfs_show = intel_event_sysfs_show, |
f22f54f4 PZ |
2513 | }; |
2514 | ||
de0428a7 | 2515 | struct intel_shared_regs *allocate_shared_regs(int cpu) |
efc9f05d SE |
2516 | { |
2517 | struct intel_shared_regs *regs; | |
2518 | int i; | |
2519 | ||
2520 | regs = kzalloc_node(sizeof(struct intel_shared_regs), | |
2521 | GFP_KERNEL, cpu_to_node(cpu)); | |
2522 | if (regs) { | |
2523 | /* | |
2524 | * initialize the locks to keep lockdep happy | |
2525 | */ | |
2526 | for (i = 0; i < EXTRA_REG_MAX; i++) | |
2527 | raw_spin_lock_init(®s->regs[i].lock); | |
2528 | ||
2529 | regs->core_id = -1; | |
2530 | } | |
2531 | return regs; | |
2532 | } | |
2533 | ||
6f6539ca MD |
2534 | static struct intel_excl_cntrs *allocate_excl_cntrs(int cpu) |
2535 | { | |
2536 | struct intel_excl_cntrs *c; | |
2537 | int i; | |
2538 | ||
2539 | c = kzalloc_node(sizeof(struct intel_excl_cntrs), | |
2540 | GFP_KERNEL, cpu_to_node(cpu)); | |
2541 | if (c) { | |
2542 | raw_spin_lock_init(&c->lock); | |
2543 | for (i = 0; i < X86_PMC_IDX_MAX; i++) { | |
2544 | c->states[0].state[i] = INTEL_EXCL_UNUSED; | |
2545 | c->states[0].init_state[i] = INTEL_EXCL_UNUSED; | |
2546 | ||
2547 | c->states[1].state[i] = INTEL_EXCL_UNUSED; | |
2548 | c->states[1].init_state[i] = INTEL_EXCL_UNUSED; | |
2549 | } | |
2550 | c->core_id = -1; | |
2551 | } | |
2552 | return c; | |
2553 | } | |
2554 | ||
a7e3ed1e AK |
2555 | static int intel_pmu_cpu_prepare(int cpu) |
2556 | { | |
2557 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); | |
2558 | ||
6f6539ca MD |
2559 | if (x86_pmu.extra_regs || x86_pmu.lbr_sel_map) { |
2560 | cpuc->shared_regs = allocate_shared_regs(cpu); | |
2561 | if (!cpuc->shared_regs) | |
2562 | return NOTIFY_BAD; | |
2563 | } | |
69092624 | 2564 | |
6f6539ca MD |
2565 | if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) { |
2566 | size_t sz = X86_PMC_IDX_MAX * sizeof(struct event_constraint); | |
2567 | ||
2568 | cpuc->constraint_list = kzalloc(sz, GFP_KERNEL); | |
2569 | if (!cpuc->constraint_list) | |
2570 | return NOTIFY_BAD; | |
2571 | ||
2572 | cpuc->excl_cntrs = allocate_excl_cntrs(cpu); | |
2573 | if (!cpuc->excl_cntrs) { | |
2574 | kfree(cpuc->constraint_list); | |
2575 | kfree(cpuc->shared_regs); | |
2576 | return NOTIFY_BAD; | |
2577 | } | |
2578 | cpuc->excl_thread_id = 0; | |
2579 | } | |
a7e3ed1e | 2580 | |
a7e3ed1e AK |
2581 | return NOTIFY_OK; |
2582 | } | |
2583 | ||
74846d35 PZ |
2584 | static void intel_pmu_cpu_starting(int cpu) |
2585 | { | |
a7e3ed1e AK |
2586 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); |
2587 | int core_id = topology_core_id(cpu); | |
2588 | int i; | |
2589 | ||
69092624 LM |
2590 | init_debug_store_on_cpu(cpu); |
2591 | /* | |
2592 | * Deal with CPUs that don't clear their LBRs on power-up. | |
2593 | */ | |
2594 | intel_pmu_lbr_reset(); | |
2595 | ||
b36817e8 SE |
2596 | cpuc->lbr_sel = NULL; |
2597 | ||
2598 | if (!cpuc->shared_regs) | |
69092624 LM |
2599 | return; |
2600 | ||
9a5e3fb5 | 2601 | if (!(x86_pmu.flags & PMU_FL_NO_HT_SHARING)) { |
90413464 SE |
2602 | void **onln = &cpuc->kfree_on_online[X86_PERF_KFREE_SHARED]; |
2603 | ||
b36817e8 SE |
2604 | for_each_cpu(i, topology_thread_cpumask(cpu)) { |
2605 | struct intel_shared_regs *pc; | |
a7e3ed1e | 2606 | |
b36817e8 SE |
2607 | pc = per_cpu(cpu_hw_events, i).shared_regs; |
2608 | if (pc && pc->core_id == core_id) { | |
90413464 | 2609 | *onln = cpuc->shared_regs; |
b36817e8 SE |
2610 | cpuc->shared_regs = pc; |
2611 | break; | |
2612 | } | |
a7e3ed1e | 2613 | } |
b36817e8 SE |
2614 | cpuc->shared_regs->core_id = core_id; |
2615 | cpuc->shared_regs->refcnt++; | |
a7e3ed1e AK |
2616 | } |
2617 | ||
b36817e8 SE |
2618 | if (x86_pmu.lbr_sel_map) |
2619 | cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR]; | |
6f6539ca MD |
2620 | |
2621 | if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) { | |
2622 | for_each_cpu(i, topology_thread_cpumask(cpu)) { | |
2623 | struct intel_excl_cntrs *c; | |
2624 | ||
2625 | c = per_cpu(cpu_hw_events, i).excl_cntrs; | |
2626 | if (c && c->core_id == core_id) { | |
2627 | cpuc->kfree_on_online[1] = cpuc->excl_cntrs; | |
2628 | cpuc->excl_cntrs = c; | |
2629 | cpuc->excl_thread_id = 1; | |
2630 | break; | |
2631 | } | |
2632 | } | |
2633 | cpuc->excl_cntrs->core_id = core_id; | |
2634 | cpuc->excl_cntrs->refcnt++; | |
2635 | } | |
74846d35 PZ |
2636 | } |
2637 | ||
2638 | static void intel_pmu_cpu_dying(int cpu) | |
2639 | { | |
a7e3ed1e | 2640 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); |
efc9f05d | 2641 | struct intel_shared_regs *pc; |
6f6539ca | 2642 | struct intel_excl_cntrs *c; |
a7e3ed1e | 2643 | |
efc9f05d | 2644 | pc = cpuc->shared_regs; |
a7e3ed1e AK |
2645 | if (pc) { |
2646 | if (pc->core_id == -1 || --pc->refcnt == 0) | |
2647 | kfree(pc); | |
efc9f05d | 2648 | cpuc->shared_regs = NULL; |
a7e3ed1e | 2649 | } |
6f6539ca MD |
2650 | c = cpuc->excl_cntrs; |
2651 | if (c) { | |
2652 | if (c->core_id == -1 || --c->refcnt == 0) | |
2653 | kfree(c); | |
2654 | cpuc->excl_cntrs = NULL; | |
2655 | kfree(cpuc->constraint_list); | |
2656 | cpuc->constraint_list = NULL; | |
2657 | } | |
a7e3ed1e | 2658 | |
e979121b MD |
2659 | c = cpuc->excl_cntrs; |
2660 | if (c) { | |
2661 | if (c->core_id == -1 || --c->refcnt == 0) | |
2662 | kfree(c); | |
2663 | cpuc->excl_cntrs = NULL; | |
2664 | } | |
2665 | ||
74846d35 PZ |
2666 | fini_debug_store_on_cpu(cpu); |
2667 | } | |
2668 | ||
641cc938 JO |
2669 | PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63"); |
2670 | ||
a63fcab4 SE |
2671 | PMU_FORMAT_ATTR(ldlat, "config1:0-15"); |
2672 | ||
641cc938 JO |
2673 | static struct attribute *intel_arch3_formats_attr[] = { |
2674 | &format_attr_event.attr, | |
2675 | &format_attr_umask.attr, | |
2676 | &format_attr_edge.attr, | |
2677 | &format_attr_pc.attr, | |
2678 | &format_attr_any.attr, | |
2679 | &format_attr_inv.attr, | |
2680 | &format_attr_cmask.attr, | |
3a632cb2 AK |
2681 | &format_attr_in_tx.attr, |
2682 | &format_attr_in_tx_cp.attr, | |
641cc938 JO |
2683 | |
2684 | &format_attr_offcore_rsp.attr, /* XXX do NHM/WSM + SNB breakout */ | |
a63fcab4 | 2685 | &format_attr_ldlat.attr, /* PEBS load latency */ |
641cc938 JO |
2686 | NULL, |
2687 | }; | |
2688 | ||
caaa8be3 | 2689 | static __initconst const struct x86_pmu intel_pmu = { |
f22f54f4 PZ |
2690 | .name = "Intel", |
2691 | .handle_irq = intel_pmu_handle_irq, | |
2692 | .disable_all = intel_pmu_disable_all, | |
2693 | .enable_all = intel_pmu_enable_all, | |
2694 | .enable = intel_pmu_enable_event, | |
2695 | .disable = intel_pmu_disable_event, | |
b4cdc5c2 | 2696 | .hw_config = intel_pmu_hw_config, |
a072738e | 2697 | .schedule_events = x86_schedule_events, |
f22f54f4 PZ |
2698 | .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, |
2699 | .perfctr = MSR_ARCH_PERFMON_PERFCTR0, | |
2700 | .event_map = intel_pmu_event_map, | |
f22f54f4 PZ |
2701 | .max_events = ARRAY_SIZE(intel_perfmon_event_map), |
2702 | .apic = 1, | |
2703 | /* | |
2704 | * Intel PMCs cannot be accessed sanely above 32 bit width, | |
2705 | * so we install an artificial 1<<31 period regardless of | |
2706 | * the generic event period: | |
2707 | */ | |
2708 | .max_period = (1ULL << 31) - 1, | |
3f6da390 | 2709 | .get_event_constraints = intel_get_event_constraints, |
a7e3ed1e | 2710 | .put_event_constraints = intel_put_event_constraints, |
0780c927 | 2711 | .pebs_aliases = intel_pebs_aliases_core2, |
3f6da390 | 2712 | |
641cc938 | 2713 | .format_attrs = intel_arch3_formats_attr, |
0bf79d44 | 2714 | .events_sysfs_show = intel_event_sysfs_show, |
641cc938 | 2715 | |
a7e3ed1e | 2716 | .cpu_prepare = intel_pmu_cpu_prepare, |
74846d35 PZ |
2717 | .cpu_starting = intel_pmu_cpu_starting, |
2718 | .cpu_dying = intel_pmu_cpu_dying, | |
144d31e6 | 2719 | .guest_get_msrs = intel_guest_get_msrs, |
2a0ad3b3 | 2720 | .sched_task = intel_pmu_lbr_sched_task, |
f22f54f4 PZ |
2721 | }; |
2722 | ||
c1d6f42f | 2723 | static __init void intel_clovertown_quirk(void) |
3c44780b PZ |
2724 | { |
2725 | /* | |
2726 | * PEBS is unreliable due to: | |
2727 | * | |
2728 | * AJ67 - PEBS may experience CPL leaks | |
2729 | * AJ68 - PEBS PMI may be delayed by one event | |
2730 | * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12] | |
2731 | * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS | |
2732 | * | |
2733 | * AJ67 could be worked around by restricting the OS/USR flags. | |
2734 | * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI. | |
2735 | * | |
2736 | * AJ106 could possibly be worked around by not allowing LBR | |
2737 | * usage from PEBS, including the fixup. | |
2738 | * AJ68 could possibly be worked around by always programming | |
ec75a716 | 2739 | * a pebs_event_reset[0] value and coping with the lost events. |
3c44780b PZ |
2740 | * |
2741 | * But taken together it might just make sense to not enable PEBS on | |
2742 | * these chips. | |
2743 | */ | |
c767a54b | 2744 | pr_warn("PEBS disabled due to CPU errata\n"); |
3c44780b PZ |
2745 | x86_pmu.pebs = 0; |
2746 | x86_pmu.pebs_constraints = NULL; | |
2747 | } | |
2748 | ||
c93dc84c PZ |
2749 | static int intel_snb_pebs_broken(int cpu) |
2750 | { | |
2751 | u32 rev = UINT_MAX; /* default to broken for unknown models */ | |
2752 | ||
2753 | switch (cpu_data(cpu).x86_model) { | |
2754 | case 42: /* SNB */ | |
2755 | rev = 0x28; | |
2756 | break; | |
2757 | ||
2758 | case 45: /* SNB-EP */ | |
2759 | switch (cpu_data(cpu).x86_mask) { | |
2760 | case 6: rev = 0x618; break; | |
2761 | case 7: rev = 0x70c; break; | |
2762 | } | |
2763 | } | |
2764 | ||
2765 | return (cpu_data(cpu).microcode < rev); | |
2766 | } | |
2767 | ||
2768 | static void intel_snb_check_microcode(void) | |
2769 | { | |
2770 | int pebs_broken = 0; | |
2771 | int cpu; | |
2772 | ||
2773 | get_online_cpus(); | |
2774 | for_each_online_cpu(cpu) { | |
2775 | if ((pebs_broken = intel_snb_pebs_broken(cpu))) | |
2776 | break; | |
2777 | } | |
2778 | put_online_cpus(); | |
2779 | ||
2780 | if (pebs_broken == x86_pmu.pebs_broken) | |
2781 | return; | |
2782 | ||
2783 | /* | |
2784 | * Serialized by the microcode lock.. | |
2785 | */ | |
2786 | if (x86_pmu.pebs_broken) { | |
2787 | pr_info("PEBS enabled due to microcode update\n"); | |
2788 | x86_pmu.pebs_broken = 0; | |
2789 | } else { | |
2790 | pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n"); | |
2791 | x86_pmu.pebs_broken = 1; | |
2792 | } | |
2793 | } | |
2794 | ||
338b522c KL |
2795 | /* |
2796 | * Under certain circumstances, access certain MSR may cause #GP. | |
2797 | * The function tests if the input MSR can be safely accessed. | |
2798 | */ | |
2799 | static bool check_msr(unsigned long msr, u64 mask) | |
2800 | { | |
2801 | u64 val_old, val_new, val_tmp; | |
2802 | ||
2803 | /* | |
2804 | * Read the current value, change it and read it back to see if it | |
2805 | * matches, this is needed to detect certain hardware emulators | |
2806 | * (qemu/kvm) that don't trap on the MSR access and always return 0s. | |
2807 | */ | |
2808 | if (rdmsrl_safe(msr, &val_old)) | |
2809 | return false; | |
2810 | ||
2811 | /* | |
2812 | * Only change the bits which can be updated by wrmsrl. | |
2813 | */ | |
2814 | val_tmp = val_old ^ mask; | |
2815 | if (wrmsrl_safe(msr, val_tmp) || | |
2816 | rdmsrl_safe(msr, &val_new)) | |
2817 | return false; | |
2818 | ||
2819 | if (val_new != val_tmp) | |
2820 | return false; | |
2821 | ||
2822 | /* Here it's sure that the MSR can be safely accessed. | |
2823 | * Restore the old value and return. | |
2824 | */ | |
2825 | wrmsrl(msr, val_old); | |
2826 | ||
2827 | return true; | |
2828 | } | |
2829 | ||
c1d6f42f | 2830 | static __init void intel_sandybridge_quirk(void) |
6a600a8b | 2831 | { |
c93dc84c PZ |
2832 | x86_pmu.check_microcode = intel_snb_check_microcode; |
2833 | intel_snb_check_microcode(); | |
6a600a8b PZ |
2834 | } |
2835 | ||
c1d6f42f PZ |
2836 | static const struct { int id; char *name; } intel_arch_events_map[] __initconst = { |
2837 | { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" }, | |
2838 | { PERF_COUNT_HW_INSTRUCTIONS, "instructions" }, | |
2839 | { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" }, | |
2840 | { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" }, | |
2841 | { PERF_COUNT_HW_CACHE_MISSES, "cache misses" }, | |
2842 | { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" }, | |
2843 | { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" }, | |
ffb871bc GN |
2844 | }; |
2845 | ||
c1d6f42f PZ |
2846 | static __init void intel_arch_events_quirk(void) |
2847 | { | |
2848 | int bit; | |
2849 | ||
2850 | /* disable event that reported as not presend by cpuid */ | |
2851 | for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) { | |
2852 | intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0; | |
c767a54b JP |
2853 | pr_warn("CPUID marked event: \'%s\' unavailable\n", |
2854 | intel_arch_events_map[bit].name); | |
c1d6f42f PZ |
2855 | } |
2856 | } | |
2857 | ||
2858 | static __init void intel_nehalem_quirk(void) | |
2859 | { | |
2860 | union cpuid10_ebx ebx; | |
2861 | ||
2862 | ebx.full = x86_pmu.events_maskl; | |
2863 | if (ebx.split.no_branch_misses_retired) { | |
2864 | /* | |
2865 | * Erratum AAJ80 detected, we work it around by using | |
2866 | * the BR_MISP_EXEC.ANY event. This will over-count | |
2867 | * branch-misses, but it's still much better than the | |
2868 | * architectural event which is often completely bogus: | |
2869 | */ | |
2870 | intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89; | |
2871 | ebx.split.no_branch_misses_retired = 0; | |
2872 | x86_pmu.events_maskl = ebx.full; | |
c767a54b | 2873 | pr_info("CPU erratum AAJ80 worked around\n"); |
c1d6f42f PZ |
2874 | } |
2875 | } | |
2876 | ||
93fcf72c MD |
2877 | /* |
2878 | * enable software workaround for errata: | |
2879 | * SNB: BJ122 | |
2880 | * IVB: BV98 | |
2881 | * HSW: HSD29 | |
2882 | * | |
2883 | * Only needed when HT is enabled. However detecting | |
2884 | * this is too difficult and model specific so we enable | |
2885 | * it even with HT off for now. | |
2886 | */ | |
2887 | static __init void intel_ht_bug(void) | |
2888 | { | |
2889 | x86_pmu.flags |= PMU_FL_EXCL_CNTRS; | |
2890 | ||
2891 | x86_pmu.commit_scheduling = intel_commit_scheduling; | |
2892 | x86_pmu.start_scheduling = intel_start_scheduling; | |
2893 | x86_pmu.stop_scheduling = intel_stop_scheduling; | |
2894 | ||
2895 | pr_info("CPU erratum BJ122, BV98, HSD29 worked around\n"); | |
2896 | } | |
2897 | ||
7f2ee91f IM |
2898 | EVENT_ATTR_STR(mem-loads, mem_ld_hsw, "event=0xcd,umask=0x1,ldlat=3"); |
2899 | EVENT_ATTR_STR(mem-stores, mem_st_hsw, "event=0xd0,umask=0x82") | |
f9134f36 | 2900 | |
4b2c4f1f | 2901 | /* Haswell special events */ |
7f2ee91f IM |
2902 | EVENT_ATTR_STR(tx-start, tx_start, "event=0xc9,umask=0x1"); |
2903 | EVENT_ATTR_STR(tx-commit, tx_commit, "event=0xc9,umask=0x2"); | |
2904 | EVENT_ATTR_STR(tx-abort, tx_abort, "event=0xc9,umask=0x4"); | |
2905 | EVENT_ATTR_STR(tx-capacity, tx_capacity, "event=0x54,umask=0x2"); | |
2906 | EVENT_ATTR_STR(tx-conflict, tx_conflict, "event=0x54,umask=0x1"); | |
2907 | EVENT_ATTR_STR(el-start, el_start, "event=0xc8,umask=0x1"); | |
2908 | EVENT_ATTR_STR(el-commit, el_commit, "event=0xc8,umask=0x2"); | |
2909 | EVENT_ATTR_STR(el-abort, el_abort, "event=0xc8,umask=0x4"); | |
2910 | EVENT_ATTR_STR(el-capacity, el_capacity, "event=0x54,umask=0x2"); | |
2911 | EVENT_ATTR_STR(el-conflict, el_conflict, "event=0x54,umask=0x1"); | |
2912 | EVENT_ATTR_STR(cycles-t, cycles_t, "event=0x3c,in_tx=1"); | |
2913 | EVENT_ATTR_STR(cycles-ct, cycles_ct, "event=0x3c,in_tx=1,in_tx_cp=1"); | |
4b2c4f1f | 2914 | |
f9134f36 | 2915 | static struct attribute *hsw_events_attrs[] = { |
4b2c4f1f AK |
2916 | EVENT_PTR(tx_start), |
2917 | EVENT_PTR(tx_commit), | |
2918 | EVENT_PTR(tx_abort), | |
2919 | EVENT_PTR(tx_capacity), | |
2920 | EVENT_PTR(tx_conflict), | |
2921 | EVENT_PTR(el_start), | |
2922 | EVENT_PTR(el_commit), | |
2923 | EVENT_PTR(el_abort), | |
2924 | EVENT_PTR(el_capacity), | |
2925 | EVENT_PTR(el_conflict), | |
2926 | EVENT_PTR(cycles_t), | |
2927 | EVENT_PTR(cycles_ct), | |
f9134f36 AK |
2928 | EVENT_PTR(mem_ld_hsw), |
2929 | EVENT_PTR(mem_st_hsw), | |
2930 | NULL | |
2931 | }; | |
2932 | ||
de0428a7 | 2933 | __init int intel_pmu_init(void) |
f22f54f4 PZ |
2934 | { |
2935 | union cpuid10_edx edx; | |
2936 | union cpuid10_eax eax; | |
ffb871bc | 2937 | union cpuid10_ebx ebx; |
a1eac7ac | 2938 | struct event_constraint *c; |
f22f54f4 | 2939 | unsigned int unused; |
338b522c KL |
2940 | struct extra_reg *er; |
2941 | int version, i; | |
f22f54f4 PZ |
2942 | |
2943 | if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) { | |
a072738e CG |
2944 | switch (boot_cpu_data.x86) { |
2945 | case 0x6: | |
2946 | return p6_pmu_init(); | |
e717bf4e VW |
2947 | case 0xb: |
2948 | return knc_pmu_init(); | |
a072738e CG |
2949 | case 0xf: |
2950 | return p4_pmu_init(); | |
2951 | } | |
f22f54f4 | 2952 | return -ENODEV; |
f22f54f4 PZ |
2953 | } |
2954 | ||
2955 | /* | |
2956 | * Check whether the Architectural PerfMon supports | |
2957 | * Branch Misses Retired hw_event or not. | |
2958 | */ | |
ffb871bc GN |
2959 | cpuid(10, &eax.full, &ebx.full, &unused, &edx.full); |
2960 | if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT) | |
f22f54f4 PZ |
2961 | return -ENODEV; |
2962 | ||
2963 | version = eax.split.version_id; | |
2964 | if (version < 2) | |
2965 | x86_pmu = core_pmu; | |
2966 | else | |
2967 | x86_pmu = intel_pmu; | |
2968 | ||
2969 | x86_pmu.version = version; | |
948b1bb8 RR |
2970 | x86_pmu.num_counters = eax.split.num_counters; |
2971 | x86_pmu.cntval_bits = eax.split.bit_width; | |
2972 | x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1; | |
f22f54f4 | 2973 | |
c1d6f42f PZ |
2974 | x86_pmu.events_maskl = ebx.full; |
2975 | x86_pmu.events_mask_len = eax.split.mask_length; | |
2976 | ||
70ab7003 AK |
2977 | x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters); |
2978 | ||
f22f54f4 PZ |
2979 | /* |
2980 | * Quirk: v2 perfmon does not report fixed-purpose events, so | |
2981 | * assume at least 3 events: | |
2982 | */ | |
2983 | if (version > 1) | |
948b1bb8 | 2984 | x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3); |
f22f54f4 | 2985 | |
c9b08884 | 2986 | if (boot_cpu_has(X86_FEATURE_PDCM)) { |
8db909a7 PZ |
2987 | u64 capabilities; |
2988 | ||
2989 | rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities); | |
2990 | x86_pmu.intel_cap.capabilities = capabilities; | |
2991 | } | |
2992 | ||
ca037701 PZ |
2993 | intel_ds_init(); |
2994 | ||
c1d6f42f PZ |
2995 | x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */ |
2996 | ||
f22f54f4 PZ |
2997 | /* |
2998 | * Install the hw-cache-events table: | |
2999 | */ | |
3000 | switch (boot_cpu_data.x86_model) { | |
0f7c29ce | 3001 | case 14: /* 65nm Core "Yonah" */ |
f22f54f4 PZ |
3002 | pr_cont("Core events, "); |
3003 | break; | |
3004 | ||
0f7c29ce | 3005 | case 15: /* 65nm Core2 "Merom" */ |
c1d6f42f | 3006 | x86_add_quirk(intel_clovertown_quirk); |
0f7c29ce PZ |
3007 | case 22: /* 65nm Core2 "Merom-L" */ |
3008 | case 23: /* 45nm Core2 "Penryn" */ | |
3009 | case 29: /* 45nm Core2 "Dunnington (MP) */ | |
f22f54f4 PZ |
3010 | memcpy(hw_cache_event_ids, core2_hw_cache_event_ids, |
3011 | sizeof(hw_cache_event_ids)); | |
3012 | ||
caff2bef PZ |
3013 | intel_pmu_lbr_init_core(); |
3014 | ||
f22f54f4 | 3015 | x86_pmu.event_constraints = intel_core2_event_constraints; |
17e31629 | 3016 | x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints; |
f22f54f4 PZ |
3017 | pr_cont("Core2 events, "); |
3018 | break; | |
3019 | ||
0f7c29ce PZ |
3020 | case 30: /* 45nm Nehalem */ |
3021 | case 26: /* 45nm Nehalem-EP */ | |
3022 | case 46: /* 45nm Nehalem-EX */ | |
f22f54f4 PZ |
3023 | memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids, |
3024 | sizeof(hw_cache_event_ids)); | |
e994d7d2 AK |
3025 | memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, |
3026 | sizeof(hw_cache_extra_regs)); | |
f22f54f4 | 3027 | |
caff2bef PZ |
3028 | intel_pmu_lbr_init_nhm(); |
3029 | ||
f22f54f4 | 3030 | x86_pmu.event_constraints = intel_nehalem_event_constraints; |
17e31629 | 3031 | x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints; |
11164cd4 | 3032 | x86_pmu.enable_all = intel_pmu_nhm_enable_all; |
a7e3ed1e | 3033 | x86_pmu.extra_regs = intel_nehalem_extra_regs; |
ec75a716 | 3034 | |
f20093ee SE |
3035 | x86_pmu.cpu_events = nhm_events_attrs; |
3036 | ||
91fc4cc0 | 3037 | /* UOPS_ISSUED.STALLED_CYCLES */ |
f9b4eeb8 PZ |
3038 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = |
3039 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | |
91fc4cc0 | 3040 | /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ |
f9b4eeb8 PZ |
3041 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = |
3042 | X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); | |
94403f88 | 3043 | |
c1d6f42f | 3044 | x86_add_quirk(intel_nehalem_quirk); |
ec75a716 | 3045 | |
11164cd4 | 3046 | pr_cont("Nehalem events, "); |
f22f54f4 | 3047 | break; |
caff2bef | 3048 | |
0f7c29ce PZ |
3049 | case 28: /* 45nm Atom "Pineview" */ |
3050 | case 38: /* 45nm Atom "Lincroft" */ | |
3051 | case 39: /* 32nm Atom "Penwell" */ | |
3052 | case 53: /* 32nm Atom "Cloverview" */ | |
3053 | case 54: /* 32nm Atom "Cedarview" */ | |
f22f54f4 PZ |
3054 | memcpy(hw_cache_event_ids, atom_hw_cache_event_ids, |
3055 | sizeof(hw_cache_event_ids)); | |
3056 | ||
caff2bef PZ |
3057 | intel_pmu_lbr_init_atom(); |
3058 | ||
f22f54f4 | 3059 | x86_pmu.event_constraints = intel_gen_event_constraints; |
17e31629 | 3060 | x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints; |
f22f54f4 PZ |
3061 | pr_cont("Atom events, "); |
3062 | break; | |
3063 | ||
0f7c29ce | 3064 | case 55: /* 22nm Atom "Silvermont" */ |
ef454cae | 3065 | case 76: /* 14nm Atom "Airmont" */ |
0f7c29ce | 3066 | case 77: /* 22nm Atom "Silvermont Avoton/Rangely" */ |
1fa64180 YZ |
3067 | memcpy(hw_cache_event_ids, slm_hw_cache_event_ids, |
3068 | sizeof(hw_cache_event_ids)); | |
3069 | memcpy(hw_cache_extra_regs, slm_hw_cache_extra_regs, | |
3070 | sizeof(hw_cache_extra_regs)); | |
3071 | ||
3072 | intel_pmu_lbr_init_atom(); | |
3073 | ||
3074 | x86_pmu.event_constraints = intel_slm_event_constraints; | |
3075 | x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints; | |
3076 | x86_pmu.extra_regs = intel_slm_extra_regs; | |
9a5e3fb5 | 3077 | x86_pmu.flags |= PMU_FL_HAS_RSP_1; |
1fa64180 YZ |
3078 | pr_cont("Silvermont events, "); |
3079 | break; | |
3080 | ||
0f7c29ce PZ |
3081 | case 37: /* 32nm Westmere */ |
3082 | case 44: /* 32nm Westmere-EP */ | |
3083 | case 47: /* 32nm Westmere-EX */ | |
f22f54f4 PZ |
3084 | memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids, |
3085 | sizeof(hw_cache_event_ids)); | |
e994d7d2 AK |
3086 | memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, |
3087 | sizeof(hw_cache_extra_regs)); | |
f22f54f4 | 3088 | |
caff2bef PZ |
3089 | intel_pmu_lbr_init_nhm(); |
3090 | ||
f22f54f4 | 3091 | x86_pmu.event_constraints = intel_westmere_event_constraints; |
40b91cd1 | 3092 | x86_pmu.enable_all = intel_pmu_nhm_enable_all; |
17e31629 | 3093 | x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints; |
a7e3ed1e | 3094 | x86_pmu.extra_regs = intel_westmere_extra_regs; |
9a5e3fb5 | 3095 | x86_pmu.flags |= PMU_FL_HAS_RSP_1; |
30112039 | 3096 | |
f20093ee SE |
3097 | x86_pmu.cpu_events = nhm_events_attrs; |
3098 | ||
30112039 | 3099 | /* UOPS_ISSUED.STALLED_CYCLES */ |
f9b4eeb8 PZ |
3100 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = |
3101 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | |
30112039 | 3102 | /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ |
f9b4eeb8 PZ |
3103 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = |
3104 | X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); | |
30112039 | 3105 | |
f22f54f4 PZ |
3106 | pr_cont("Westmere events, "); |
3107 | break; | |
b622d644 | 3108 | |
0f7c29ce PZ |
3109 | case 42: /* 32nm SandyBridge */ |
3110 | case 45: /* 32nm SandyBridge-E/EN/EP */ | |
47a8863d | 3111 | x86_add_quirk(intel_sandybridge_quirk); |
93fcf72c | 3112 | x86_add_quirk(intel_ht_bug); |
b06b3d49 LM |
3113 | memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, |
3114 | sizeof(hw_cache_event_ids)); | |
74e6543f YZ |
3115 | memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, |
3116 | sizeof(hw_cache_extra_regs)); | |
b06b3d49 | 3117 | |
c5cc2cd9 | 3118 | intel_pmu_lbr_init_snb(); |
b06b3d49 LM |
3119 | |
3120 | x86_pmu.event_constraints = intel_snb_event_constraints; | |
de0428a7 | 3121 | x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints; |
0780c927 | 3122 | x86_pmu.pebs_aliases = intel_pebs_aliases_snb; |
f1923820 SE |
3123 | if (boot_cpu_data.x86_model == 45) |
3124 | x86_pmu.extra_regs = intel_snbep_extra_regs; | |
3125 | else | |
3126 | x86_pmu.extra_regs = intel_snb_extra_regs; | |
93fcf72c MD |
3127 | |
3128 | ||
ee89cbc2 | 3129 | /* all extra regs are per-cpu when HT is on */ |
9a5e3fb5 SE |
3130 | x86_pmu.flags |= PMU_FL_HAS_RSP_1; |
3131 | x86_pmu.flags |= PMU_FL_NO_HT_SHARING; | |
e04d1b23 | 3132 | |
f20093ee SE |
3133 | x86_pmu.cpu_events = snb_events_attrs; |
3134 | ||
e04d1b23 | 3135 | /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ |
f9b4eeb8 PZ |
3136 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = |
3137 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | |
e04d1b23 | 3138 | /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/ |
f9b4eeb8 PZ |
3139 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = |
3140 | X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1); | |
e04d1b23 | 3141 | |
b06b3d49 LM |
3142 | pr_cont("SandyBridge events, "); |
3143 | break; | |
0f7c29ce PZ |
3144 | |
3145 | case 58: /* 22nm IvyBridge */ | |
3146 | case 62: /* 22nm IvyBridge-EP/EX */ | |
93fcf72c | 3147 | x86_add_quirk(intel_ht_bug); |
20a36e39 SE |
3148 | memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, |
3149 | sizeof(hw_cache_event_ids)); | |
1996388e VW |
3150 | /* dTLB-load-misses on IVB is different than SNB */ |
3151 | hw_cache_event_ids[C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = 0x8108; /* DTLB_LOAD_MISSES.DEMAND_LD_MISS_CAUSES_A_WALK */ | |
3152 | ||
20a36e39 SE |
3153 | memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, |
3154 | sizeof(hw_cache_extra_regs)); | |
3155 | ||
3156 | intel_pmu_lbr_init_snb(); | |
3157 | ||
69943182 | 3158 | x86_pmu.event_constraints = intel_ivb_event_constraints; |
20a36e39 SE |
3159 | x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints; |
3160 | x86_pmu.pebs_aliases = intel_pebs_aliases_snb; | |
f1923820 SE |
3161 | if (boot_cpu_data.x86_model == 62) |
3162 | x86_pmu.extra_regs = intel_snbep_extra_regs; | |
3163 | else | |
3164 | x86_pmu.extra_regs = intel_snb_extra_regs; | |
20a36e39 | 3165 | /* all extra regs are per-cpu when HT is on */ |
9a5e3fb5 SE |
3166 | x86_pmu.flags |= PMU_FL_HAS_RSP_1; |
3167 | x86_pmu.flags |= PMU_FL_NO_HT_SHARING; | |
20a36e39 | 3168 | |
f20093ee SE |
3169 | x86_pmu.cpu_events = snb_events_attrs; |
3170 | ||
20a36e39 SE |
3171 | /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ |
3172 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = | |
3173 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | |
3174 | ||
3175 | pr_cont("IvyBridge events, "); | |
3176 | break; | |
3177 | ||
b06b3d49 | 3178 | |
d86c8eaf AK |
3179 | case 60: /* 22nm Haswell Core */ |
3180 | case 63: /* 22nm Haswell Server */ | |
3181 | case 69: /* 22nm Haswell ULT */ | |
3182 | case 70: /* 22nm Haswell + GT3e (Intel Iris Pro graphics) */ | |
93fcf72c | 3183 | x86_add_quirk(intel_ht_bug); |
72db5596 | 3184 | x86_pmu.late_ack = true; |
0f1b5ca2 AK |
3185 | memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids)); |
3186 | memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); | |
3a632cb2 | 3187 | |
e9d7f7cd | 3188 | intel_pmu_lbr_init_hsw(); |
3a632cb2 AK |
3189 | |
3190 | x86_pmu.event_constraints = intel_hsw_event_constraints; | |
3044318f | 3191 | x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints; |
36bbb2f2 | 3192 | x86_pmu.extra_regs = intel_snbep_extra_regs; |
3044318f | 3193 | x86_pmu.pebs_aliases = intel_pebs_aliases_snb; |
3a632cb2 | 3194 | /* all extra regs are per-cpu when HT is on */ |
9a5e3fb5 SE |
3195 | x86_pmu.flags |= PMU_FL_HAS_RSP_1; |
3196 | x86_pmu.flags |= PMU_FL_NO_HT_SHARING; | |
3a632cb2 AK |
3197 | |
3198 | x86_pmu.hw_config = hsw_hw_config; | |
3199 | x86_pmu.get_event_constraints = hsw_get_event_constraints; | |
f9134f36 | 3200 | x86_pmu.cpu_events = hsw_events_attrs; |
b7af41a1 | 3201 | x86_pmu.lbr_double_abort = true; |
3a632cb2 AK |
3202 | pr_cont("Haswell events, "); |
3203 | break; | |
3204 | ||
91f1b705 AK |
3205 | case 61: /* 14nm Broadwell Core-M */ |
3206 | case 86: /* 14nm Broadwell Xeon D */ | |
3207 | x86_pmu.late_ack = true; | |
3208 | memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids)); | |
3209 | memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); | |
3210 | ||
3211 | /* L3_MISS_LOCAL_DRAM is BIT(26) in Broadwell */ | |
3212 | hw_cache_extra_regs[C(LL)][C(OP_READ)][C(RESULT_MISS)] = HSW_DEMAND_READ | | |
3213 | BDW_L3_MISS|HSW_SNOOP_DRAM; | |
3214 | hw_cache_extra_regs[C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = HSW_DEMAND_WRITE|BDW_L3_MISS| | |
3215 | HSW_SNOOP_DRAM; | |
3216 | hw_cache_extra_regs[C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = HSW_DEMAND_READ| | |
3217 | BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM; | |
3218 | hw_cache_extra_regs[C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = HSW_DEMAND_WRITE| | |
3219 | BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM; | |
3220 | ||
3221 | intel_pmu_lbr_init_snb(); | |
3222 | ||
3223 | x86_pmu.event_constraints = intel_bdw_event_constraints; | |
3224 | x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints; | |
3225 | x86_pmu.extra_regs = intel_snbep_extra_regs; | |
3226 | x86_pmu.pebs_aliases = intel_pebs_aliases_snb; | |
3227 | /* all extra regs are per-cpu when HT is on */ | |
9a5e3fb5 SE |
3228 | x86_pmu.flags |= PMU_FL_HAS_RSP_1; |
3229 | x86_pmu.flags |= PMU_FL_NO_HT_SHARING; | |
91f1b705 AK |
3230 | |
3231 | x86_pmu.hw_config = hsw_hw_config; | |
3232 | x86_pmu.get_event_constraints = hsw_get_event_constraints; | |
3233 | x86_pmu.cpu_events = hsw_events_attrs; | |
294fe0f5 | 3234 | x86_pmu.limit_period = bdw_limit_period; |
91f1b705 AK |
3235 | pr_cont("Broadwell events, "); |
3236 | break; | |
3237 | ||
f22f54f4 | 3238 | default: |
0af3ac1f AK |
3239 | switch (x86_pmu.version) { |
3240 | case 1: | |
3241 | x86_pmu.event_constraints = intel_v1_event_constraints; | |
3242 | pr_cont("generic architected perfmon v1, "); | |
3243 | break; | |
3244 | default: | |
3245 | /* | |
3246 | * default constraints for v2 and up | |
3247 | */ | |
3248 | x86_pmu.event_constraints = intel_gen_event_constraints; | |
3249 | pr_cont("generic architected perfmon, "); | |
3250 | break; | |
3251 | } | |
f22f54f4 | 3252 | } |
ffb871bc | 3253 | |
a1eac7ac RR |
3254 | if (x86_pmu.num_counters > INTEL_PMC_MAX_GENERIC) { |
3255 | WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!", | |
3256 | x86_pmu.num_counters, INTEL_PMC_MAX_GENERIC); | |
3257 | x86_pmu.num_counters = INTEL_PMC_MAX_GENERIC; | |
3258 | } | |
3259 | x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1; | |
3260 | ||
3261 | if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) { | |
3262 | WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!", | |
3263 | x86_pmu.num_counters_fixed, INTEL_PMC_MAX_FIXED); | |
3264 | x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED; | |
3265 | } | |
3266 | ||
3267 | x86_pmu.intel_ctrl |= | |
3268 | ((1LL << x86_pmu.num_counters_fixed)-1) << INTEL_PMC_IDX_FIXED; | |
3269 | ||
3270 | if (x86_pmu.event_constraints) { | |
3271 | /* | |
3272 | * event on fixed counter2 (REF_CYCLES) only works on this | |
3273 | * counter, so do not extend mask to generic counters | |
3274 | */ | |
3275 | for_each_event_constraint(c, x86_pmu.event_constraints) { | |
3a632cb2 | 3276 | if (c->cmask != FIXED_EVENT_FLAGS |
a1eac7ac RR |
3277 | || c->idxmsk64 == INTEL_PMC_MSK_FIXED_REF_CYCLES) { |
3278 | continue; | |
3279 | } | |
3280 | ||
3281 | c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1; | |
3282 | c->weight += x86_pmu.num_counters; | |
3283 | } | |
3284 | } | |
3285 | ||
338b522c KL |
3286 | /* |
3287 | * Access LBR MSR may cause #GP under certain circumstances. | |
3288 | * E.g. KVM doesn't support LBR MSR | |
3289 | * Check all LBT MSR here. | |
3290 | * Disable LBR access if any LBR MSRs can not be accessed. | |
3291 | */ | |
3292 | if (x86_pmu.lbr_nr && !check_msr(x86_pmu.lbr_tos, 0x3UL)) | |
3293 | x86_pmu.lbr_nr = 0; | |
3294 | for (i = 0; i < x86_pmu.lbr_nr; i++) { | |
3295 | if (!(check_msr(x86_pmu.lbr_from + i, 0xffffUL) && | |
3296 | check_msr(x86_pmu.lbr_to + i, 0xffffUL))) | |
3297 | x86_pmu.lbr_nr = 0; | |
3298 | } | |
3299 | ||
3300 | /* | |
3301 | * Access extra MSR may cause #GP under certain circumstances. | |
3302 | * E.g. KVM doesn't support offcore event | |
3303 | * Check all extra_regs here. | |
3304 | */ | |
3305 | if (x86_pmu.extra_regs) { | |
3306 | for (er = x86_pmu.extra_regs; er->msr; er++) { | |
3307 | er->extra_msr_access = check_msr(er->msr, 0x1ffUL); | |
3308 | /* Disable LBR select mapping */ | |
3309 | if ((er->idx == EXTRA_REG_LBR) && !er->extra_msr_access) | |
3310 | x86_pmu.lbr_sel_map = NULL; | |
3311 | } | |
3312 | } | |
3313 | ||
069e0c3c AK |
3314 | /* Support full width counters using alternative MSR range */ |
3315 | if (x86_pmu.intel_cap.full_width_write) { | |
3316 | x86_pmu.max_period = x86_pmu.cntval_mask; | |
3317 | x86_pmu.perfctr = MSR_IA32_PMC0; | |
3318 | pr_cont("full-width counters, "); | |
3319 | } | |
3320 | ||
f22f54f4 PZ |
3321 | return 0; |
3322 | } |