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> |
aacfbe6a | 15 | #include <linux/nmi.h> |
de0428a7 | 16 | |
3a632cb2 | 17 | #include <asm/cpufeature.h> |
de0428a7 KW |
18 | #include <asm/hardirq.h> |
19 | #include <asm/apic.h> | |
20 | ||
21 | #include "perf_event.h" | |
a7e3ed1e | 22 | |
f22f54f4 | 23 | /* |
b622d644 | 24 | * Intel PerfMon, used on Core and later. |
f22f54f4 | 25 | */ |
ec75a716 | 26 | static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly = |
f22f54f4 | 27 | { |
c3b7cdf1 PE |
28 | [PERF_COUNT_HW_CPU_CYCLES] = 0x003c, |
29 | [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0, | |
30 | [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e, | |
31 | [PERF_COUNT_HW_CACHE_MISSES] = 0x412e, | |
32 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4, | |
33 | [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5, | |
34 | [PERF_COUNT_HW_BUS_CYCLES] = 0x013c, | |
35 | [PERF_COUNT_HW_REF_CPU_CYCLES] = 0x0300, /* pseudo-encoding */ | |
f22f54f4 PZ |
36 | }; |
37 | ||
5c543e3c | 38 | static struct event_constraint intel_core_event_constraints[] __read_mostly = |
f22f54f4 PZ |
39 | { |
40 | INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ | |
41 | INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ | |
42 | INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */ | |
43 | INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */ | |
44 | INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */ | |
45 | INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */ | |
46 | EVENT_CONSTRAINT_END | |
47 | }; | |
48 | ||
5c543e3c | 49 | static struct event_constraint intel_core2_event_constraints[] __read_mostly = |
f22f54f4 | 50 | { |
b622d644 PZ |
51 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
52 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
cd09c0c4 | 53 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ |
f22f54f4 PZ |
54 | INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */ |
55 | INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ | |
56 | INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ | |
57 | INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */ | |
58 | INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */ | |
59 | INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */ | |
60 | INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */ | |
61 | INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */ | |
b622d644 | 62 | INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */ |
f22f54f4 PZ |
63 | INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */ |
64 | EVENT_CONSTRAINT_END | |
65 | }; | |
66 | ||
5c543e3c | 67 | static struct event_constraint intel_nehalem_event_constraints[] __read_mostly = |
f22f54f4 | 68 | { |
b622d644 PZ |
69 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
70 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
cd09c0c4 | 71 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ |
f22f54f4 PZ |
72 | INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */ |
73 | INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */ | |
74 | INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */ | |
75 | INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */ | |
76 | INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */ | |
77 | INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */ | |
78 | INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ | |
79 | INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ | |
80 | EVENT_CONSTRAINT_END | |
81 | }; | |
82 | ||
5c543e3c | 83 | static struct extra_reg intel_nehalem_extra_regs[] __read_mostly = |
a7e3ed1e | 84 | { |
53ad0447 YZ |
85 | /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ |
86 | INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), | |
f20093ee | 87 | INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b), |
a7e3ed1e AK |
88 | EVENT_EXTRA_END |
89 | }; | |
90 | ||
5c543e3c | 91 | static struct event_constraint intel_westmere_event_constraints[] __read_mostly = |
f22f54f4 | 92 | { |
b622d644 PZ |
93 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
94 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
cd09c0c4 | 95 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ |
f22f54f4 PZ |
96 | INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ |
97 | INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */ | |
98 | INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ | |
d1100770 | 99 | INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */ |
f22f54f4 PZ |
100 | EVENT_CONSTRAINT_END |
101 | }; | |
102 | ||
5c543e3c | 103 | static struct event_constraint intel_snb_event_constraints[] __read_mostly = |
b06b3d49 LM |
104 | { |
105 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | |
106 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
cd09c0c4 | 107 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ |
fd4a5aef SE |
108 | INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */ |
109 | INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */ | |
110 | INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ | |
111 | INTEL_UEVENT_CONSTRAINT(0x06a3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ | |
b06b3d49 | 112 | INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */ |
b06b3d49 LM |
113 | INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ |
114 | INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ | |
f8378f52 AK |
115 | INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */ |
116 | INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ | |
93fcf72c MD |
117 | |
118 | INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ | |
119 | INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ | |
120 | INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ | |
121 | INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ | |
122 | ||
b06b3d49 LM |
123 | EVENT_CONSTRAINT_END |
124 | }; | |
125 | ||
69943182 SE |
126 | static struct event_constraint intel_ivb_event_constraints[] __read_mostly = |
127 | { | |
128 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | |
129 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
130 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ | |
131 | INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */ | |
132 | INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMTPY */ | |
133 | INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */ | |
6113af14 | 134 | INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_LDM_PENDING */ |
69943182 SE |
135 | INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */ |
136 | INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */ | |
137 | INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */ | |
138 | INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ | |
139 | INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ | |
140 | INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ | |
93fcf72c MD |
141 | |
142 | INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ | |
143 | INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ | |
144 | INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ | |
145 | INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ | |
146 | ||
69943182 SE |
147 | EVENT_CONSTRAINT_END |
148 | }; | |
149 | ||
5c543e3c | 150 | static struct extra_reg intel_westmere_extra_regs[] __read_mostly = |
a7e3ed1e | 151 | { |
53ad0447 YZ |
152 | /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ |
153 | INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), | |
154 | INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1), | |
f20093ee | 155 | INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b), |
a7e3ed1e AK |
156 | EVENT_EXTRA_END |
157 | }; | |
158 | ||
0af3ac1f AK |
159 | static struct event_constraint intel_v1_event_constraints[] __read_mostly = |
160 | { | |
161 | EVENT_CONSTRAINT_END | |
162 | }; | |
163 | ||
5c543e3c | 164 | static struct event_constraint intel_gen_event_constraints[] __read_mostly = |
f22f54f4 | 165 | { |
b622d644 PZ |
166 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
167 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
cd09c0c4 | 168 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ |
f22f54f4 PZ |
169 | EVENT_CONSTRAINT_END |
170 | }; | |
171 | ||
1fa64180 YZ |
172 | static struct event_constraint intel_slm_event_constraints[] __read_mostly = |
173 | { | |
174 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | |
175 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
1fa64180 YZ |
176 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */ |
177 | EVENT_CONSTRAINT_END | |
178 | }; | |
179 | ||
9a92e16f AK |
180 | struct event_constraint intel_skl_event_constraints[] = { |
181 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | |
182 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
183 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ | |
184 | INTEL_UEVENT_CONSTRAINT(0x1c0, 0x2), /* INST_RETIRED.PREC_DIST */ | |
185 | EVENT_CONSTRAINT_END | |
186 | }; | |
187 | ||
ee89cbc2 | 188 | static struct extra_reg intel_snb_extra_regs[] __read_mostly = { |
53ad0447 YZ |
189 | /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ |
190 | INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0), | |
191 | INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1), | |
f20093ee | 192 | INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), |
f1923820 SE |
193 | EVENT_EXTRA_END |
194 | }; | |
195 | ||
196 | static struct extra_reg intel_snbep_extra_regs[] __read_mostly = { | |
53ad0447 YZ |
197 | /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ |
198 | INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0), | |
199 | INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1), | |
f1a52789 | 200 | INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), |
ee89cbc2 SE |
201 | EVENT_EXTRA_END |
202 | }; | |
203 | ||
9a92e16f AK |
204 | static struct extra_reg intel_skl_extra_regs[] __read_mostly = { |
205 | INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0), | |
206 | INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1), | |
207 | INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), | |
dfe1f3cb AK |
208 | /* |
209 | * Note the low 8 bits eventsel code is not a continuous field, containing | |
210 | * some #GPing bits. These are masked out. | |
211 | */ | |
212 | INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE), | |
9a92e16f AK |
213 | EVENT_EXTRA_END |
214 | }; | |
215 | ||
7f2ee91f IM |
216 | EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3"); |
217 | EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3"); | |
218 | EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2"); | |
f20093ee SE |
219 | |
220 | struct attribute *nhm_events_attrs[] = { | |
221 | EVENT_PTR(mem_ld_nhm), | |
222 | NULL, | |
223 | }; | |
224 | ||
225 | struct attribute *snb_events_attrs[] = { | |
226 | EVENT_PTR(mem_ld_snb), | |
9ad64c0f | 227 | EVENT_PTR(mem_st_snb), |
f20093ee SE |
228 | NULL, |
229 | }; | |
230 | ||
3a632cb2 AK |
231 | static struct event_constraint intel_hsw_event_constraints[] = { |
232 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | |
233 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
234 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ | |
235 | INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.* */ | |
236 | INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ | |
237 | INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ | |
238 | /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ | |
c420f19b | 239 | INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), |
3a632cb2 | 240 | /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ |
c420f19b | 241 | INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), |
3a632cb2 | 242 | /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */ |
c420f19b | 243 | INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), |
93fcf72c MD |
244 | |
245 | INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ | |
246 | INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ | |
247 | INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ | |
248 | INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ | |
249 | ||
3a632cb2 AK |
250 | EVENT_CONSTRAINT_END |
251 | }; | |
252 | ||
91f1b705 AK |
253 | struct event_constraint intel_bdw_event_constraints[] = { |
254 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | |
255 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | |
256 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ | |
257 | INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */ | |
5e176213 | 258 | INTEL_UEVENT_CONSTRAINT(0x8a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_MISS */ |
91f1b705 AK |
259 | EVENT_CONSTRAINT_END |
260 | }; | |
261 | ||
f22f54f4 PZ |
262 | static u64 intel_pmu_event_map(int hw_event) |
263 | { | |
264 | return intel_perfmon_event_map[hw_event]; | |
265 | } | |
266 | ||
9a92e16f AK |
267 | /* |
268 | * Notes on the events: | |
269 | * - data reads do not include code reads (comparable to earlier tables) | |
270 | * - data counts include speculative execution (except L1 write, dtlb, bpu) | |
271 | * - remote node access includes remote memory, remote cache, remote mmio. | |
272 | * - prefetches are not included in the counts. | |
273 | * - icache miss does not include decoded icache | |
274 | */ | |
275 | ||
276 | #define SKL_DEMAND_DATA_RD BIT_ULL(0) | |
277 | #define SKL_DEMAND_RFO BIT_ULL(1) | |
278 | #define SKL_ANY_RESPONSE BIT_ULL(16) | |
279 | #define SKL_SUPPLIER_NONE BIT_ULL(17) | |
280 | #define SKL_L3_MISS_LOCAL_DRAM BIT_ULL(26) | |
281 | #define SKL_L3_MISS_REMOTE_HOP0_DRAM BIT_ULL(27) | |
282 | #define SKL_L3_MISS_REMOTE_HOP1_DRAM BIT_ULL(28) | |
283 | #define SKL_L3_MISS_REMOTE_HOP2P_DRAM BIT_ULL(29) | |
284 | #define SKL_L3_MISS (SKL_L3_MISS_LOCAL_DRAM| \ | |
285 | SKL_L3_MISS_REMOTE_HOP0_DRAM| \ | |
286 | SKL_L3_MISS_REMOTE_HOP1_DRAM| \ | |
287 | SKL_L3_MISS_REMOTE_HOP2P_DRAM) | |
288 | #define SKL_SPL_HIT BIT_ULL(30) | |
289 | #define SKL_SNOOP_NONE BIT_ULL(31) | |
290 | #define SKL_SNOOP_NOT_NEEDED BIT_ULL(32) | |
291 | #define SKL_SNOOP_MISS BIT_ULL(33) | |
292 | #define SKL_SNOOP_HIT_NO_FWD BIT_ULL(34) | |
293 | #define SKL_SNOOP_HIT_WITH_FWD BIT_ULL(35) | |
294 | #define SKL_SNOOP_HITM BIT_ULL(36) | |
295 | #define SKL_SNOOP_NON_DRAM BIT_ULL(37) | |
296 | #define SKL_ANY_SNOOP (SKL_SPL_HIT|SKL_SNOOP_NONE| \ | |
297 | SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \ | |
298 | SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \ | |
299 | SKL_SNOOP_HITM|SKL_SNOOP_NON_DRAM) | |
300 | #define SKL_DEMAND_READ SKL_DEMAND_DATA_RD | |
301 | #define SKL_SNOOP_DRAM (SKL_SNOOP_NONE| \ | |
302 | SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \ | |
303 | SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \ | |
304 | SKL_SNOOP_HITM|SKL_SPL_HIT) | |
305 | #define SKL_DEMAND_WRITE SKL_DEMAND_RFO | |
306 | #define SKL_LLC_ACCESS SKL_ANY_RESPONSE | |
307 | #define SKL_L3_MISS_REMOTE (SKL_L3_MISS_REMOTE_HOP0_DRAM| \ | |
308 | SKL_L3_MISS_REMOTE_HOP1_DRAM| \ | |
309 | SKL_L3_MISS_REMOTE_HOP2P_DRAM) | |
310 | ||
311 | static __initconst const u64 skl_hw_cache_event_ids | |
312 | [PERF_COUNT_HW_CACHE_MAX] | |
313 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
314 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
315 | { | |
316 | [ C(L1D ) ] = { | |
317 | [ C(OP_READ) ] = { | |
318 | [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */ | |
319 | [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */ | |
320 | }, | |
321 | [ C(OP_WRITE) ] = { | |
322 | [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */ | |
323 | [ C(RESULT_MISS) ] = 0x0, | |
324 | }, | |
325 | [ C(OP_PREFETCH) ] = { | |
326 | [ C(RESULT_ACCESS) ] = 0x0, | |
327 | [ C(RESULT_MISS) ] = 0x0, | |
328 | }, | |
329 | }, | |
330 | [ C(L1I ) ] = { | |
331 | [ C(OP_READ) ] = { | |
332 | [ C(RESULT_ACCESS) ] = 0x0, | |
333 | [ C(RESULT_MISS) ] = 0x283, /* ICACHE_64B.MISS */ | |
334 | }, | |
335 | [ C(OP_WRITE) ] = { | |
336 | [ C(RESULT_ACCESS) ] = -1, | |
337 | [ C(RESULT_MISS) ] = -1, | |
338 | }, | |
339 | [ C(OP_PREFETCH) ] = { | |
340 | [ C(RESULT_ACCESS) ] = 0x0, | |
341 | [ C(RESULT_MISS) ] = 0x0, | |
342 | }, | |
343 | }, | |
344 | [ C(LL ) ] = { | |
345 | [ C(OP_READ) ] = { | |
346 | [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
347 | [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
348 | }, | |
349 | [ C(OP_WRITE) ] = { | |
350 | [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
351 | [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
352 | }, | |
353 | [ C(OP_PREFETCH) ] = { | |
354 | [ C(RESULT_ACCESS) ] = 0x0, | |
355 | [ C(RESULT_MISS) ] = 0x0, | |
356 | }, | |
357 | }, | |
358 | [ C(DTLB) ] = { | |
359 | [ C(OP_READ) ] = { | |
360 | [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */ | |
361 | [ C(RESULT_MISS) ] = 0x608, /* DTLB_LOAD_MISSES.WALK_COMPLETED */ | |
362 | }, | |
363 | [ C(OP_WRITE) ] = { | |
364 | [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */ | |
365 | [ C(RESULT_MISS) ] = 0x649, /* DTLB_STORE_MISSES.WALK_COMPLETED */ | |
366 | }, | |
367 | [ C(OP_PREFETCH) ] = { | |
368 | [ C(RESULT_ACCESS) ] = 0x0, | |
369 | [ C(RESULT_MISS) ] = 0x0, | |
370 | }, | |
371 | }, | |
372 | [ C(ITLB) ] = { | |
373 | [ C(OP_READ) ] = { | |
374 | [ C(RESULT_ACCESS) ] = 0x2085, /* ITLB_MISSES.STLB_HIT */ | |
375 | [ C(RESULT_MISS) ] = 0xe85, /* ITLB_MISSES.WALK_COMPLETED */ | |
376 | }, | |
377 | [ C(OP_WRITE) ] = { | |
378 | [ C(RESULT_ACCESS) ] = -1, | |
379 | [ C(RESULT_MISS) ] = -1, | |
380 | }, | |
381 | [ C(OP_PREFETCH) ] = { | |
382 | [ C(RESULT_ACCESS) ] = -1, | |
383 | [ C(RESULT_MISS) ] = -1, | |
384 | }, | |
385 | }, | |
386 | [ C(BPU ) ] = { | |
387 | [ C(OP_READ) ] = { | |
388 | [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
389 | [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */ | |
390 | }, | |
391 | [ C(OP_WRITE) ] = { | |
392 | [ C(RESULT_ACCESS) ] = -1, | |
393 | [ C(RESULT_MISS) ] = -1, | |
394 | }, | |
395 | [ C(OP_PREFETCH) ] = { | |
396 | [ C(RESULT_ACCESS) ] = -1, | |
397 | [ C(RESULT_MISS) ] = -1, | |
398 | }, | |
399 | }, | |
400 | [ C(NODE) ] = { | |
401 | [ C(OP_READ) ] = { | |
402 | [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
403 | [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
404 | }, | |
405 | [ C(OP_WRITE) ] = { | |
406 | [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
407 | [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
408 | }, | |
409 | [ C(OP_PREFETCH) ] = { | |
410 | [ C(RESULT_ACCESS) ] = 0x0, | |
411 | [ C(RESULT_MISS) ] = 0x0, | |
412 | }, | |
413 | }, | |
414 | }; | |
415 | ||
416 | static __initconst const u64 skl_hw_cache_extra_regs | |
417 | [PERF_COUNT_HW_CACHE_MAX] | |
418 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
419 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
420 | { | |
421 | [ C(LL ) ] = { | |
422 | [ C(OP_READ) ] = { | |
423 | [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ| | |
424 | SKL_LLC_ACCESS|SKL_ANY_SNOOP, | |
425 | [ C(RESULT_MISS) ] = SKL_DEMAND_READ| | |
426 | SKL_L3_MISS|SKL_ANY_SNOOP| | |
427 | SKL_SUPPLIER_NONE, | |
428 | }, | |
429 | [ C(OP_WRITE) ] = { | |
430 | [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE| | |
431 | SKL_LLC_ACCESS|SKL_ANY_SNOOP, | |
432 | [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE| | |
433 | SKL_L3_MISS|SKL_ANY_SNOOP| | |
434 | SKL_SUPPLIER_NONE, | |
435 | }, | |
436 | [ C(OP_PREFETCH) ] = { | |
437 | [ C(RESULT_ACCESS) ] = 0x0, | |
438 | [ C(RESULT_MISS) ] = 0x0, | |
439 | }, | |
440 | }, | |
441 | [ C(NODE) ] = { | |
442 | [ C(OP_READ) ] = { | |
443 | [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ| | |
444 | SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM, | |
445 | [ C(RESULT_MISS) ] = SKL_DEMAND_READ| | |
446 | SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM, | |
447 | }, | |
448 | [ C(OP_WRITE) ] = { | |
449 | [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE| | |
450 | SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM, | |
451 | [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE| | |
452 | SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM, | |
453 | }, | |
454 | [ C(OP_PREFETCH) ] = { | |
455 | [ C(RESULT_ACCESS) ] = 0x0, | |
456 | [ C(RESULT_MISS) ] = 0x0, | |
457 | }, | |
458 | }, | |
459 | }; | |
460 | ||
74e6543f YZ |
461 | #define SNB_DMND_DATA_RD (1ULL << 0) |
462 | #define SNB_DMND_RFO (1ULL << 1) | |
463 | #define SNB_DMND_IFETCH (1ULL << 2) | |
464 | #define SNB_DMND_WB (1ULL << 3) | |
465 | #define SNB_PF_DATA_RD (1ULL << 4) | |
466 | #define SNB_PF_RFO (1ULL << 5) | |
467 | #define SNB_PF_IFETCH (1ULL << 6) | |
468 | #define SNB_LLC_DATA_RD (1ULL << 7) | |
469 | #define SNB_LLC_RFO (1ULL << 8) | |
470 | #define SNB_LLC_IFETCH (1ULL << 9) | |
471 | #define SNB_BUS_LOCKS (1ULL << 10) | |
472 | #define SNB_STRM_ST (1ULL << 11) | |
473 | #define SNB_OTHER (1ULL << 15) | |
474 | #define SNB_RESP_ANY (1ULL << 16) | |
475 | #define SNB_NO_SUPP (1ULL << 17) | |
476 | #define SNB_LLC_HITM (1ULL << 18) | |
477 | #define SNB_LLC_HITE (1ULL << 19) | |
478 | #define SNB_LLC_HITS (1ULL << 20) | |
479 | #define SNB_LLC_HITF (1ULL << 21) | |
480 | #define SNB_LOCAL (1ULL << 22) | |
481 | #define SNB_REMOTE (0xffULL << 23) | |
482 | #define SNB_SNP_NONE (1ULL << 31) | |
483 | #define SNB_SNP_NOT_NEEDED (1ULL << 32) | |
484 | #define SNB_SNP_MISS (1ULL << 33) | |
485 | #define SNB_NO_FWD (1ULL << 34) | |
486 | #define SNB_SNP_FWD (1ULL << 35) | |
487 | #define SNB_HITM (1ULL << 36) | |
488 | #define SNB_NON_DRAM (1ULL << 37) | |
489 | ||
490 | #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD) | |
491 | #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO) | |
492 | #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) | |
493 | ||
494 | #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \ | |
495 | SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \ | |
496 | SNB_HITM) | |
497 | ||
498 | #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY) | |
499 | #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY) | |
500 | ||
501 | #define SNB_L3_ACCESS SNB_RESP_ANY | |
502 | #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM) | |
503 | ||
504 | static __initconst const u64 snb_hw_cache_extra_regs | |
505 | [PERF_COUNT_HW_CACHE_MAX] | |
506 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
507 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
508 | { | |
509 | [ C(LL ) ] = { | |
510 | [ C(OP_READ) ] = { | |
511 | [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS, | |
512 | [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_L3_MISS, | |
513 | }, | |
514 | [ C(OP_WRITE) ] = { | |
515 | [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS, | |
516 | [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_L3_MISS, | |
517 | }, | |
518 | [ C(OP_PREFETCH) ] = { | |
519 | [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS, | |
520 | [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_L3_MISS, | |
521 | }, | |
522 | }, | |
523 | [ C(NODE) ] = { | |
524 | [ C(OP_READ) ] = { | |
525 | [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY, | |
526 | [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_DRAM_REMOTE, | |
527 | }, | |
528 | [ C(OP_WRITE) ] = { | |
529 | [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY, | |
530 | [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE, | |
531 | }, | |
532 | [ C(OP_PREFETCH) ] = { | |
533 | [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY, | |
534 | [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE, | |
535 | }, | |
536 | }, | |
537 | }; | |
538 | ||
b06b3d49 LM |
539 | static __initconst const u64 snb_hw_cache_event_ids |
540 | [PERF_COUNT_HW_CACHE_MAX] | |
541 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
542 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
543 | { | |
544 | [ C(L1D) ] = { | |
545 | [ C(OP_READ) ] = { | |
546 | [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */ | |
547 | [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */ | |
548 | }, | |
549 | [ C(OP_WRITE) ] = { | |
550 | [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */ | |
551 | [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */ | |
552 | }, | |
553 | [ C(OP_PREFETCH) ] = { | |
554 | [ C(RESULT_ACCESS) ] = 0x0, | |
555 | [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */ | |
556 | }, | |
557 | }, | |
558 | [ C(L1I ) ] = { | |
559 | [ C(OP_READ) ] = { | |
560 | [ C(RESULT_ACCESS) ] = 0x0, | |
561 | [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */ | |
562 | }, | |
563 | [ C(OP_WRITE) ] = { | |
564 | [ C(RESULT_ACCESS) ] = -1, | |
565 | [ C(RESULT_MISS) ] = -1, | |
566 | }, | |
567 | [ C(OP_PREFETCH) ] = { | |
568 | [ C(RESULT_ACCESS) ] = 0x0, | |
569 | [ C(RESULT_MISS) ] = 0x0, | |
570 | }, | |
571 | }, | |
572 | [ C(LL ) ] = { | |
b06b3d49 | 573 | [ C(OP_READ) ] = { |
63b6a675 | 574 | /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ |
b06b3d49 | 575 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
576 | /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ |
577 | [ C(RESULT_MISS) ] = 0x01b7, | |
b06b3d49 LM |
578 | }, |
579 | [ C(OP_WRITE) ] = { | |
63b6a675 | 580 | /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ |
b06b3d49 | 581 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
582 | /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ |
583 | [ C(RESULT_MISS) ] = 0x01b7, | |
b06b3d49 LM |
584 | }, |
585 | [ C(OP_PREFETCH) ] = { | |
63b6a675 | 586 | /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ |
b06b3d49 | 587 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
588 | /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ |
589 | [ C(RESULT_MISS) ] = 0x01b7, | |
b06b3d49 LM |
590 | }, |
591 | }, | |
592 | [ C(DTLB) ] = { | |
593 | [ C(OP_READ) ] = { | |
594 | [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */ | |
595 | [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */ | |
596 | }, | |
597 | [ C(OP_WRITE) ] = { | |
598 | [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */ | |
599 | [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */ | |
600 | }, | |
601 | [ C(OP_PREFETCH) ] = { | |
602 | [ C(RESULT_ACCESS) ] = 0x0, | |
603 | [ C(RESULT_MISS) ] = 0x0, | |
604 | }, | |
605 | }, | |
606 | [ C(ITLB) ] = { | |
607 | [ C(OP_READ) ] = { | |
608 | [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */ | |
609 | [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */ | |
610 | }, | |
611 | [ C(OP_WRITE) ] = { | |
612 | [ C(RESULT_ACCESS) ] = -1, | |
613 | [ C(RESULT_MISS) ] = -1, | |
614 | }, | |
615 | [ C(OP_PREFETCH) ] = { | |
616 | [ C(RESULT_ACCESS) ] = -1, | |
617 | [ C(RESULT_MISS) ] = -1, | |
618 | }, | |
619 | }, | |
620 | [ C(BPU ) ] = { | |
621 | [ C(OP_READ) ] = { | |
622 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
623 | [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ | |
624 | }, | |
625 | [ C(OP_WRITE) ] = { | |
626 | [ C(RESULT_ACCESS) ] = -1, | |
627 | [ C(RESULT_MISS) ] = -1, | |
628 | }, | |
629 | [ C(OP_PREFETCH) ] = { | |
630 | [ C(RESULT_ACCESS) ] = -1, | |
631 | [ C(RESULT_MISS) ] = -1, | |
632 | }, | |
633 | }, | |
89d6c0b5 PZ |
634 | [ C(NODE) ] = { |
635 | [ C(OP_READ) ] = { | |
74e6543f YZ |
636 | [ C(RESULT_ACCESS) ] = 0x01b7, |
637 | [ C(RESULT_MISS) ] = 0x01b7, | |
89d6c0b5 PZ |
638 | }, |
639 | [ C(OP_WRITE) ] = { | |
74e6543f YZ |
640 | [ C(RESULT_ACCESS) ] = 0x01b7, |
641 | [ C(RESULT_MISS) ] = 0x01b7, | |
89d6c0b5 PZ |
642 | }, |
643 | [ C(OP_PREFETCH) ] = { | |
74e6543f YZ |
644 | [ C(RESULT_ACCESS) ] = 0x01b7, |
645 | [ C(RESULT_MISS) ] = 0x01b7, | |
89d6c0b5 PZ |
646 | }, |
647 | }, | |
648 | ||
b06b3d49 LM |
649 | }; |
650 | ||
0f1b5ca2 AK |
651 | /* |
652 | * Notes on the events: | |
653 | * - data reads do not include code reads (comparable to earlier tables) | |
654 | * - data counts include speculative execution (except L1 write, dtlb, bpu) | |
655 | * - remote node access includes remote memory, remote cache, remote mmio. | |
656 | * - prefetches are not included in the counts because they are not | |
657 | * reliably counted. | |
658 | */ | |
659 | ||
660 | #define HSW_DEMAND_DATA_RD BIT_ULL(0) | |
661 | #define HSW_DEMAND_RFO BIT_ULL(1) | |
662 | #define HSW_ANY_RESPONSE BIT_ULL(16) | |
663 | #define HSW_SUPPLIER_NONE BIT_ULL(17) | |
664 | #define HSW_L3_MISS_LOCAL_DRAM BIT_ULL(22) | |
665 | #define HSW_L3_MISS_REMOTE_HOP0 BIT_ULL(27) | |
666 | #define HSW_L3_MISS_REMOTE_HOP1 BIT_ULL(28) | |
667 | #define HSW_L3_MISS_REMOTE_HOP2P BIT_ULL(29) | |
668 | #define HSW_L3_MISS (HSW_L3_MISS_LOCAL_DRAM| \ | |
669 | HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \ | |
670 | HSW_L3_MISS_REMOTE_HOP2P) | |
671 | #define HSW_SNOOP_NONE BIT_ULL(31) | |
672 | #define HSW_SNOOP_NOT_NEEDED BIT_ULL(32) | |
673 | #define HSW_SNOOP_MISS BIT_ULL(33) | |
674 | #define HSW_SNOOP_HIT_NO_FWD BIT_ULL(34) | |
675 | #define HSW_SNOOP_HIT_WITH_FWD BIT_ULL(35) | |
676 | #define HSW_SNOOP_HITM BIT_ULL(36) | |
677 | #define HSW_SNOOP_NON_DRAM BIT_ULL(37) | |
678 | #define HSW_ANY_SNOOP (HSW_SNOOP_NONE| \ | |
679 | HSW_SNOOP_NOT_NEEDED|HSW_SNOOP_MISS| \ | |
680 | HSW_SNOOP_HIT_NO_FWD|HSW_SNOOP_HIT_WITH_FWD| \ | |
681 | HSW_SNOOP_HITM|HSW_SNOOP_NON_DRAM) | |
682 | #define HSW_SNOOP_DRAM (HSW_ANY_SNOOP & ~HSW_SNOOP_NON_DRAM) | |
683 | #define HSW_DEMAND_READ HSW_DEMAND_DATA_RD | |
684 | #define HSW_DEMAND_WRITE HSW_DEMAND_RFO | |
685 | #define HSW_L3_MISS_REMOTE (HSW_L3_MISS_REMOTE_HOP0|\ | |
686 | HSW_L3_MISS_REMOTE_HOP1|HSW_L3_MISS_REMOTE_HOP2P) | |
687 | #define HSW_LLC_ACCESS HSW_ANY_RESPONSE | |
688 | ||
91f1b705 AK |
689 | #define BDW_L3_MISS_LOCAL BIT(26) |
690 | #define BDW_L3_MISS (BDW_L3_MISS_LOCAL| \ | |
691 | HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \ | |
692 | HSW_L3_MISS_REMOTE_HOP2P) | |
693 | ||
694 | ||
0f1b5ca2 AK |
695 | static __initconst const u64 hsw_hw_cache_event_ids |
696 | [PERF_COUNT_HW_CACHE_MAX] | |
697 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
698 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
699 | { | |
700 | [ C(L1D ) ] = { | |
701 | [ C(OP_READ) ] = { | |
702 | [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ | |
703 | [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */ | |
704 | }, | |
705 | [ C(OP_WRITE) ] = { | |
706 | [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ | |
707 | [ C(RESULT_MISS) ] = 0x0, | |
708 | }, | |
709 | [ C(OP_PREFETCH) ] = { | |
710 | [ C(RESULT_ACCESS) ] = 0x0, | |
711 | [ C(RESULT_MISS) ] = 0x0, | |
712 | }, | |
713 | }, | |
714 | [ C(L1I ) ] = { | |
715 | [ C(OP_READ) ] = { | |
716 | [ C(RESULT_ACCESS) ] = 0x0, | |
717 | [ C(RESULT_MISS) ] = 0x280, /* ICACHE.MISSES */ | |
718 | }, | |
719 | [ C(OP_WRITE) ] = { | |
720 | [ C(RESULT_ACCESS) ] = -1, | |
721 | [ C(RESULT_MISS) ] = -1, | |
722 | }, | |
723 | [ C(OP_PREFETCH) ] = { | |
724 | [ C(RESULT_ACCESS) ] = 0x0, | |
725 | [ C(RESULT_MISS) ] = 0x0, | |
726 | }, | |
727 | }, | |
728 | [ C(LL ) ] = { | |
729 | [ C(OP_READ) ] = { | |
730 | [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
731 | [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
732 | }, | |
733 | [ C(OP_WRITE) ] = { | |
734 | [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
735 | [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
736 | }, | |
737 | [ C(OP_PREFETCH) ] = { | |
738 | [ C(RESULT_ACCESS) ] = 0x0, | |
739 | [ C(RESULT_MISS) ] = 0x0, | |
740 | }, | |
741 | }, | |
742 | [ C(DTLB) ] = { | |
743 | [ C(OP_READ) ] = { | |
744 | [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ | |
745 | [ C(RESULT_MISS) ] = 0x108, /* DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK */ | |
746 | }, | |
747 | [ C(OP_WRITE) ] = { | |
748 | [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ | |
749 | [ C(RESULT_MISS) ] = 0x149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */ | |
750 | }, | |
751 | [ C(OP_PREFETCH) ] = { | |
752 | [ C(RESULT_ACCESS) ] = 0x0, | |
753 | [ C(RESULT_MISS) ] = 0x0, | |
754 | }, | |
755 | }, | |
756 | [ C(ITLB) ] = { | |
757 | [ C(OP_READ) ] = { | |
758 | [ C(RESULT_ACCESS) ] = 0x6085, /* ITLB_MISSES.STLB_HIT */ | |
759 | [ C(RESULT_MISS) ] = 0x185, /* ITLB_MISSES.MISS_CAUSES_A_WALK */ | |
760 | }, | |
761 | [ C(OP_WRITE) ] = { | |
762 | [ C(RESULT_ACCESS) ] = -1, | |
763 | [ C(RESULT_MISS) ] = -1, | |
764 | }, | |
765 | [ C(OP_PREFETCH) ] = { | |
766 | [ C(RESULT_ACCESS) ] = -1, | |
767 | [ C(RESULT_MISS) ] = -1, | |
768 | }, | |
769 | }, | |
770 | [ C(BPU ) ] = { | |
771 | [ C(OP_READ) ] = { | |
772 | [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
773 | [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */ | |
774 | }, | |
775 | [ C(OP_WRITE) ] = { | |
776 | [ C(RESULT_ACCESS) ] = -1, | |
777 | [ C(RESULT_MISS) ] = -1, | |
778 | }, | |
779 | [ C(OP_PREFETCH) ] = { | |
780 | [ C(RESULT_ACCESS) ] = -1, | |
781 | [ C(RESULT_MISS) ] = -1, | |
782 | }, | |
783 | }, | |
784 | [ C(NODE) ] = { | |
785 | [ C(OP_READ) ] = { | |
786 | [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
787 | [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
788 | }, | |
789 | [ C(OP_WRITE) ] = { | |
790 | [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
791 | [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ | |
792 | }, | |
793 | [ C(OP_PREFETCH) ] = { | |
794 | [ C(RESULT_ACCESS) ] = 0x0, | |
795 | [ C(RESULT_MISS) ] = 0x0, | |
796 | }, | |
797 | }, | |
798 | }; | |
799 | ||
800 | static __initconst const u64 hsw_hw_cache_extra_regs | |
801 | [PERF_COUNT_HW_CACHE_MAX] | |
802 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
803 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
804 | { | |
805 | [ C(LL ) ] = { | |
806 | [ C(OP_READ) ] = { | |
807 | [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ| | |
808 | HSW_LLC_ACCESS, | |
809 | [ C(RESULT_MISS) ] = HSW_DEMAND_READ| | |
810 | HSW_L3_MISS|HSW_ANY_SNOOP, | |
811 | }, | |
812 | [ C(OP_WRITE) ] = { | |
813 | [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE| | |
814 | HSW_LLC_ACCESS, | |
815 | [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE| | |
816 | HSW_L3_MISS|HSW_ANY_SNOOP, | |
817 | }, | |
818 | [ C(OP_PREFETCH) ] = { | |
819 | [ C(RESULT_ACCESS) ] = 0x0, | |
820 | [ C(RESULT_MISS) ] = 0x0, | |
821 | }, | |
822 | }, | |
823 | [ C(NODE) ] = { | |
824 | [ C(OP_READ) ] = { | |
825 | [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ| | |
826 | HSW_L3_MISS_LOCAL_DRAM| | |
827 | HSW_SNOOP_DRAM, | |
828 | [ C(RESULT_MISS) ] = HSW_DEMAND_READ| | |
829 | HSW_L3_MISS_REMOTE| | |
830 | HSW_SNOOP_DRAM, | |
831 | }, | |
832 | [ C(OP_WRITE) ] = { | |
833 | [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE| | |
834 | HSW_L3_MISS_LOCAL_DRAM| | |
835 | HSW_SNOOP_DRAM, | |
836 | [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE| | |
837 | HSW_L3_MISS_REMOTE| | |
838 | HSW_SNOOP_DRAM, | |
839 | }, | |
840 | [ C(OP_PREFETCH) ] = { | |
841 | [ C(RESULT_ACCESS) ] = 0x0, | |
842 | [ C(RESULT_MISS) ] = 0x0, | |
843 | }, | |
844 | }, | |
845 | }; | |
846 | ||
caaa8be3 | 847 | static __initconst const u64 westmere_hw_cache_event_ids |
f22f54f4 PZ |
848 | [PERF_COUNT_HW_CACHE_MAX] |
849 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
850 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
851 | { | |
852 | [ C(L1D) ] = { | |
853 | [ C(OP_READ) ] = { | |
854 | [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ | |
855 | [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ | |
856 | }, | |
857 | [ C(OP_WRITE) ] = { | |
858 | [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ | |
859 | [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ | |
860 | }, | |
861 | [ C(OP_PREFETCH) ] = { | |
862 | [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ | |
863 | [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ | |
864 | }, | |
865 | }, | |
866 | [ C(L1I ) ] = { | |
867 | [ C(OP_READ) ] = { | |
868 | [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ | |
869 | [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ | |
870 | }, | |
871 | [ C(OP_WRITE) ] = { | |
872 | [ C(RESULT_ACCESS) ] = -1, | |
873 | [ C(RESULT_MISS) ] = -1, | |
874 | }, | |
875 | [ C(OP_PREFETCH) ] = { | |
876 | [ C(RESULT_ACCESS) ] = 0x0, | |
877 | [ C(RESULT_MISS) ] = 0x0, | |
878 | }, | |
879 | }, | |
880 | [ C(LL ) ] = { | |
881 | [ C(OP_READ) ] = { | |
63b6a675 | 882 | /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ |
e994d7d2 | 883 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
884 | /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ |
885 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 | 886 | }, |
e994d7d2 AK |
887 | /* |
888 | * Use RFO, not WRITEBACK, because a write miss would typically occur | |
889 | * on RFO. | |
890 | */ | |
f22f54f4 | 891 | [ C(OP_WRITE) ] = { |
63b6a675 PZ |
892 | /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ |
893 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
894 | /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ | |
e994d7d2 | 895 | [ C(RESULT_MISS) ] = 0x01b7, |
f22f54f4 PZ |
896 | }, |
897 | [ C(OP_PREFETCH) ] = { | |
63b6a675 | 898 | /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ |
e994d7d2 | 899 | [ C(RESULT_ACCESS) ] = 0x01b7, |
63b6a675 PZ |
900 | /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ |
901 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 PZ |
902 | }, |
903 | }, | |
904 | [ C(DTLB) ] = { | |
905 | [ C(OP_READ) ] = { | |
906 | [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ | |
907 | [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ | |
908 | }, | |
909 | [ C(OP_WRITE) ] = { | |
910 | [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ | |
911 | [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ | |
912 | }, | |
913 | [ C(OP_PREFETCH) ] = { | |
914 | [ C(RESULT_ACCESS) ] = 0x0, | |
915 | [ C(RESULT_MISS) ] = 0x0, | |
916 | }, | |
917 | }, | |
918 | [ C(ITLB) ] = { | |
919 | [ C(OP_READ) ] = { | |
920 | [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ | |
921 | [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */ | |
922 | }, | |
923 | [ C(OP_WRITE) ] = { | |
924 | [ C(RESULT_ACCESS) ] = -1, | |
925 | [ C(RESULT_MISS) ] = -1, | |
926 | }, | |
927 | [ C(OP_PREFETCH) ] = { | |
928 | [ C(RESULT_ACCESS) ] = -1, | |
929 | [ C(RESULT_MISS) ] = -1, | |
930 | }, | |
931 | }, | |
932 | [ C(BPU ) ] = { | |
933 | [ C(OP_READ) ] = { | |
934 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
935 | [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ | |
936 | }, | |
937 | [ C(OP_WRITE) ] = { | |
938 | [ C(RESULT_ACCESS) ] = -1, | |
939 | [ C(RESULT_MISS) ] = -1, | |
940 | }, | |
941 | [ C(OP_PREFETCH) ] = { | |
942 | [ C(RESULT_ACCESS) ] = -1, | |
943 | [ C(RESULT_MISS) ] = -1, | |
944 | }, | |
945 | }, | |
89d6c0b5 PZ |
946 | [ C(NODE) ] = { |
947 | [ C(OP_READ) ] = { | |
948 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
949 | [ C(RESULT_MISS) ] = 0x01b7, | |
950 | }, | |
951 | [ C(OP_WRITE) ] = { | |
952 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
953 | [ C(RESULT_MISS) ] = 0x01b7, | |
954 | }, | |
955 | [ C(OP_PREFETCH) ] = { | |
956 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
957 | [ C(RESULT_MISS) ] = 0x01b7, | |
958 | }, | |
959 | }, | |
f22f54f4 PZ |
960 | }; |
961 | ||
e994d7d2 | 962 | /* |
63b6a675 PZ |
963 | * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits; |
964 | * See IA32 SDM Vol 3B 30.6.1.3 | |
e994d7d2 AK |
965 | */ |
966 | ||
63b6a675 PZ |
967 | #define NHM_DMND_DATA_RD (1 << 0) |
968 | #define NHM_DMND_RFO (1 << 1) | |
969 | #define NHM_DMND_IFETCH (1 << 2) | |
970 | #define NHM_DMND_WB (1 << 3) | |
971 | #define NHM_PF_DATA_RD (1 << 4) | |
972 | #define NHM_PF_DATA_RFO (1 << 5) | |
973 | #define NHM_PF_IFETCH (1 << 6) | |
974 | #define NHM_OFFCORE_OTHER (1 << 7) | |
975 | #define NHM_UNCORE_HIT (1 << 8) | |
976 | #define NHM_OTHER_CORE_HIT_SNP (1 << 9) | |
977 | #define NHM_OTHER_CORE_HITM (1 << 10) | |
978 | /* reserved */ | |
979 | #define NHM_REMOTE_CACHE_FWD (1 << 12) | |
980 | #define NHM_REMOTE_DRAM (1 << 13) | |
981 | #define NHM_LOCAL_DRAM (1 << 14) | |
982 | #define NHM_NON_DRAM (1 << 15) | |
983 | ||
87e24f4b PZ |
984 | #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD) |
985 | #define NHM_REMOTE (NHM_REMOTE_DRAM) | |
63b6a675 PZ |
986 | |
987 | #define NHM_DMND_READ (NHM_DMND_DATA_RD) | |
988 | #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB) | |
989 | #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO) | |
990 | ||
991 | #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM) | |
87e24f4b | 992 | #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD) |
63b6a675 | 993 | #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS) |
e994d7d2 AK |
994 | |
995 | static __initconst const u64 nehalem_hw_cache_extra_regs | |
996 | [PERF_COUNT_HW_CACHE_MAX] | |
997 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
998 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
999 | { | |
1000 | [ C(LL ) ] = { | |
1001 | [ C(OP_READ) ] = { | |
63b6a675 PZ |
1002 | [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS, |
1003 | [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS, | |
e994d7d2 AK |
1004 | }, |
1005 | [ C(OP_WRITE) ] = { | |
63b6a675 PZ |
1006 | [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS, |
1007 | [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS, | |
e994d7d2 AK |
1008 | }, |
1009 | [ C(OP_PREFETCH) ] = { | |
63b6a675 PZ |
1010 | [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS, |
1011 | [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS, | |
e994d7d2 | 1012 | }, |
89d6c0b5 PZ |
1013 | }, |
1014 | [ C(NODE) ] = { | |
1015 | [ C(OP_READ) ] = { | |
87e24f4b PZ |
1016 | [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE, |
1017 | [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE, | |
89d6c0b5 PZ |
1018 | }, |
1019 | [ C(OP_WRITE) ] = { | |
87e24f4b PZ |
1020 | [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE, |
1021 | [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE, | |
89d6c0b5 PZ |
1022 | }, |
1023 | [ C(OP_PREFETCH) ] = { | |
87e24f4b PZ |
1024 | [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE, |
1025 | [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE, | |
89d6c0b5 PZ |
1026 | }, |
1027 | }, | |
e994d7d2 AK |
1028 | }; |
1029 | ||
caaa8be3 | 1030 | static __initconst const u64 nehalem_hw_cache_event_ids |
f22f54f4 PZ |
1031 | [PERF_COUNT_HW_CACHE_MAX] |
1032 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
1033 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
1034 | { | |
1035 | [ C(L1D) ] = { | |
1036 | [ C(OP_READ) ] = { | |
f4929bd3 PZ |
1037 | [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ |
1038 | [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ | |
f22f54f4 PZ |
1039 | }, |
1040 | [ C(OP_WRITE) ] = { | |
f4929bd3 PZ |
1041 | [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ |
1042 | [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ | |
f22f54f4 PZ |
1043 | }, |
1044 | [ C(OP_PREFETCH) ] = { | |
1045 | [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ | |
1046 | [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ | |
1047 | }, | |
1048 | }, | |
1049 | [ C(L1I ) ] = { | |
1050 | [ C(OP_READ) ] = { | |
1051 | [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ | |
1052 | [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ | |
1053 | }, | |
1054 | [ C(OP_WRITE) ] = { | |
1055 | [ C(RESULT_ACCESS) ] = -1, | |
1056 | [ C(RESULT_MISS) ] = -1, | |
1057 | }, | |
1058 | [ C(OP_PREFETCH) ] = { | |
1059 | [ C(RESULT_ACCESS) ] = 0x0, | |
1060 | [ C(RESULT_MISS) ] = 0x0, | |
1061 | }, | |
1062 | }, | |
1063 | [ C(LL ) ] = { | |
1064 | [ C(OP_READ) ] = { | |
e994d7d2 AK |
1065 | /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ |
1066 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
1067 | /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ | |
1068 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 | 1069 | }, |
e994d7d2 AK |
1070 | /* |
1071 | * Use RFO, not WRITEBACK, because a write miss would typically occur | |
1072 | * on RFO. | |
1073 | */ | |
f22f54f4 | 1074 | [ C(OP_WRITE) ] = { |
e994d7d2 AK |
1075 | /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ |
1076 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
1077 | /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ | |
1078 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 PZ |
1079 | }, |
1080 | [ C(OP_PREFETCH) ] = { | |
e994d7d2 AK |
1081 | /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ |
1082 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
1083 | /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ | |
1084 | [ C(RESULT_MISS) ] = 0x01b7, | |
f22f54f4 PZ |
1085 | }, |
1086 | }, | |
1087 | [ C(DTLB) ] = { | |
1088 | [ C(OP_READ) ] = { | |
1089 | [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ | |
1090 | [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ | |
1091 | }, | |
1092 | [ C(OP_WRITE) ] = { | |
1093 | [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ | |
1094 | [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ | |
1095 | }, | |
1096 | [ C(OP_PREFETCH) ] = { | |
1097 | [ C(RESULT_ACCESS) ] = 0x0, | |
1098 | [ C(RESULT_MISS) ] = 0x0, | |
1099 | }, | |
1100 | }, | |
1101 | [ C(ITLB) ] = { | |
1102 | [ C(OP_READ) ] = { | |
1103 | [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ | |
1104 | [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */ | |
1105 | }, | |
1106 | [ C(OP_WRITE) ] = { | |
1107 | [ C(RESULT_ACCESS) ] = -1, | |
1108 | [ C(RESULT_MISS) ] = -1, | |
1109 | }, | |
1110 | [ C(OP_PREFETCH) ] = { | |
1111 | [ C(RESULT_ACCESS) ] = -1, | |
1112 | [ C(RESULT_MISS) ] = -1, | |
1113 | }, | |
1114 | }, | |
1115 | [ C(BPU ) ] = { | |
1116 | [ C(OP_READ) ] = { | |
1117 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ | |
1118 | [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ | |
1119 | }, | |
1120 | [ C(OP_WRITE) ] = { | |
1121 | [ C(RESULT_ACCESS) ] = -1, | |
1122 | [ C(RESULT_MISS) ] = -1, | |
1123 | }, | |
1124 | [ C(OP_PREFETCH) ] = { | |
1125 | [ C(RESULT_ACCESS) ] = -1, | |
1126 | [ C(RESULT_MISS) ] = -1, | |
1127 | }, | |
1128 | }, | |
89d6c0b5 PZ |
1129 | [ C(NODE) ] = { |
1130 | [ C(OP_READ) ] = { | |
1131 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
1132 | [ C(RESULT_MISS) ] = 0x01b7, | |
1133 | }, | |
1134 | [ C(OP_WRITE) ] = { | |
1135 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
1136 | [ C(RESULT_MISS) ] = 0x01b7, | |
1137 | }, | |
1138 | [ C(OP_PREFETCH) ] = { | |
1139 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
1140 | [ C(RESULT_MISS) ] = 0x01b7, | |
1141 | }, | |
1142 | }, | |
f22f54f4 PZ |
1143 | }; |
1144 | ||
caaa8be3 | 1145 | static __initconst const u64 core2_hw_cache_event_ids |
f22f54f4 PZ |
1146 | [PERF_COUNT_HW_CACHE_MAX] |
1147 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
1148 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
1149 | { | |
1150 | [ C(L1D) ] = { | |
1151 | [ C(OP_READ) ] = { | |
1152 | [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */ | |
1153 | [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */ | |
1154 | }, | |
1155 | [ C(OP_WRITE) ] = { | |
1156 | [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */ | |
1157 | [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */ | |
1158 | }, | |
1159 | [ C(OP_PREFETCH) ] = { | |
1160 | [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */ | |
1161 | [ C(RESULT_MISS) ] = 0, | |
1162 | }, | |
1163 | }, | |
1164 | [ C(L1I ) ] = { | |
1165 | [ C(OP_READ) ] = { | |
1166 | [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */ | |
1167 | [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */ | |
1168 | }, | |
1169 | [ C(OP_WRITE) ] = { | |
1170 | [ C(RESULT_ACCESS) ] = -1, | |
1171 | [ C(RESULT_MISS) ] = -1, | |
1172 | }, | |
1173 | [ C(OP_PREFETCH) ] = { | |
1174 | [ C(RESULT_ACCESS) ] = 0, | |
1175 | [ C(RESULT_MISS) ] = 0, | |
1176 | }, | |
1177 | }, | |
1178 | [ C(LL ) ] = { | |
1179 | [ C(OP_READ) ] = { | |
1180 | [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ | |
1181 | [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ | |
1182 | }, | |
1183 | [ C(OP_WRITE) ] = { | |
1184 | [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ | |
1185 | [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ | |
1186 | }, | |
1187 | [ C(OP_PREFETCH) ] = { | |
1188 | [ C(RESULT_ACCESS) ] = 0, | |
1189 | [ C(RESULT_MISS) ] = 0, | |
1190 | }, | |
1191 | }, | |
1192 | [ C(DTLB) ] = { | |
1193 | [ C(OP_READ) ] = { | |
1194 | [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ | |
1195 | [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */ | |
1196 | }, | |
1197 | [ C(OP_WRITE) ] = { | |
1198 | [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ | |
1199 | [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */ | |
1200 | }, | |
1201 | [ C(OP_PREFETCH) ] = { | |
1202 | [ C(RESULT_ACCESS) ] = 0, | |
1203 | [ C(RESULT_MISS) ] = 0, | |
1204 | }, | |
1205 | }, | |
1206 | [ C(ITLB) ] = { | |
1207 | [ C(OP_READ) ] = { | |
1208 | [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ | |
1209 | [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */ | |
1210 | }, | |
1211 | [ C(OP_WRITE) ] = { | |
1212 | [ C(RESULT_ACCESS) ] = -1, | |
1213 | [ C(RESULT_MISS) ] = -1, | |
1214 | }, | |
1215 | [ C(OP_PREFETCH) ] = { | |
1216 | [ C(RESULT_ACCESS) ] = -1, | |
1217 | [ C(RESULT_MISS) ] = -1, | |
1218 | }, | |
1219 | }, | |
1220 | [ C(BPU ) ] = { | |
1221 | [ C(OP_READ) ] = { | |
1222 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ | |
1223 | [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ | |
1224 | }, | |
1225 | [ C(OP_WRITE) ] = { | |
1226 | [ C(RESULT_ACCESS) ] = -1, | |
1227 | [ C(RESULT_MISS) ] = -1, | |
1228 | }, | |
1229 | [ C(OP_PREFETCH) ] = { | |
1230 | [ C(RESULT_ACCESS) ] = -1, | |
1231 | [ C(RESULT_MISS) ] = -1, | |
1232 | }, | |
1233 | }, | |
1234 | }; | |
1235 | ||
caaa8be3 | 1236 | static __initconst const u64 atom_hw_cache_event_ids |
f22f54f4 PZ |
1237 | [PERF_COUNT_HW_CACHE_MAX] |
1238 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
1239 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
1240 | { | |
1241 | [ C(L1D) ] = { | |
1242 | [ C(OP_READ) ] = { | |
1243 | [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */ | |
1244 | [ C(RESULT_MISS) ] = 0, | |
1245 | }, | |
1246 | [ C(OP_WRITE) ] = { | |
1247 | [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */ | |
1248 | [ C(RESULT_MISS) ] = 0, | |
1249 | }, | |
1250 | [ C(OP_PREFETCH) ] = { | |
1251 | [ C(RESULT_ACCESS) ] = 0x0, | |
1252 | [ C(RESULT_MISS) ] = 0, | |
1253 | }, | |
1254 | }, | |
1255 | [ C(L1I ) ] = { | |
1256 | [ C(OP_READ) ] = { | |
1257 | [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ | |
1258 | [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ | |
1259 | }, | |
1260 | [ C(OP_WRITE) ] = { | |
1261 | [ C(RESULT_ACCESS) ] = -1, | |
1262 | [ C(RESULT_MISS) ] = -1, | |
1263 | }, | |
1264 | [ C(OP_PREFETCH) ] = { | |
1265 | [ C(RESULT_ACCESS) ] = 0, | |
1266 | [ C(RESULT_MISS) ] = 0, | |
1267 | }, | |
1268 | }, | |
1269 | [ C(LL ) ] = { | |
1270 | [ C(OP_READ) ] = { | |
1271 | [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ | |
1272 | [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ | |
1273 | }, | |
1274 | [ C(OP_WRITE) ] = { | |
1275 | [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ | |
1276 | [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ | |
1277 | }, | |
1278 | [ C(OP_PREFETCH) ] = { | |
1279 | [ C(RESULT_ACCESS) ] = 0, | |
1280 | [ C(RESULT_MISS) ] = 0, | |
1281 | }, | |
1282 | }, | |
1283 | [ C(DTLB) ] = { | |
1284 | [ C(OP_READ) ] = { | |
1285 | [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */ | |
1286 | [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */ | |
1287 | }, | |
1288 | [ C(OP_WRITE) ] = { | |
1289 | [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */ | |
1290 | [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */ | |
1291 | }, | |
1292 | [ C(OP_PREFETCH) ] = { | |
1293 | [ C(RESULT_ACCESS) ] = 0, | |
1294 | [ C(RESULT_MISS) ] = 0, | |
1295 | }, | |
1296 | }, | |
1297 | [ C(ITLB) ] = { | |
1298 | [ C(OP_READ) ] = { | |
1299 | [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ | |
1300 | [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */ | |
1301 | }, | |
1302 | [ C(OP_WRITE) ] = { | |
1303 | [ C(RESULT_ACCESS) ] = -1, | |
1304 | [ C(RESULT_MISS) ] = -1, | |
1305 | }, | |
1306 | [ C(OP_PREFETCH) ] = { | |
1307 | [ C(RESULT_ACCESS) ] = -1, | |
1308 | [ C(RESULT_MISS) ] = -1, | |
1309 | }, | |
1310 | }, | |
1311 | [ C(BPU ) ] = { | |
1312 | [ C(OP_READ) ] = { | |
1313 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ | |
1314 | [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ | |
1315 | }, | |
1316 | [ C(OP_WRITE) ] = { | |
1317 | [ C(RESULT_ACCESS) ] = -1, | |
1318 | [ C(RESULT_MISS) ] = -1, | |
1319 | }, | |
1320 | [ C(OP_PREFETCH) ] = { | |
1321 | [ C(RESULT_ACCESS) ] = -1, | |
1322 | [ C(RESULT_MISS) ] = -1, | |
1323 | }, | |
1324 | }, | |
1325 | }; | |
1326 | ||
1fa64180 YZ |
1327 | static struct extra_reg intel_slm_extra_regs[] __read_mostly = |
1328 | { | |
1329 | /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ | |
06c939c1 | 1330 | INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x768005ffffull, RSP_0), |
ae3f011f | 1331 | INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x368005ffffull, RSP_1), |
1fa64180 YZ |
1332 | EVENT_EXTRA_END |
1333 | }; | |
1334 | ||
1335 | #define SLM_DMND_READ SNB_DMND_DATA_RD | |
1336 | #define SLM_DMND_WRITE SNB_DMND_RFO | |
1337 | #define SLM_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) | |
1338 | ||
1339 | #define SLM_SNP_ANY (SNB_SNP_NONE|SNB_SNP_MISS|SNB_NO_FWD|SNB_HITM) | |
1340 | #define SLM_LLC_ACCESS SNB_RESP_ANY | |
1341 | #define SLM_LLC_MISS (SLM_SNP_ANY|SNB_NON_DRAM) | |
1342 | ||
1343 | static __initconst const u64 slm_hw_cache_extra_regs | |
1344 | [PERF_COUNT_HW_CACHE_MAX] | |
1345 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
1346 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
1347 | { | |
1348 | [ C(LL ) ] = { | |
1349 | [ C(OP_READ) ] = { | |
1350 | [ C(RESULT_ACCESS) ] = SLM_DMND_READ|SLM_LLC_ACCESS, | |
6d374056 | 1351 | [ C(RESULT_MISS) ] = 0, |
1fa64180 YZ |
1352 | }, |
1353 | [ C(OP_WRITE) ] = { | |
1354 | [ C(RESULT_ACCESS) ] = SLM_DMND_WRITE|SLM_LLC_ACCESS, | |
1355 | [ C(RESULT_MISS) ] = SLM_DMND_WRITE|SLM_LLC_MISS, | |
1356 | }, | |
1357 | [ C(OP_PREFETCH) ] = { | |
1358 | [ C(RESULT_ACCESS) ] = SLM_DMND_PREFETCH|SLM_LLC_ACCESS, | |
1359 | [ C(RESULT_MISS) ] = SLM_DMND_PREFETCH|SLM_LLC_MISS, | |
1360 | }, | |
1361 | }, | |
1362 | }; | |
1363 | ||
1364 | static __initconst const u64 slm_hw_cache_event_ids | |
1365 | [PERF_COUNT_HW_CACHE_MAX] | |
1366 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
1367 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | |
1368 | { | |
1369 | [ C(L1D) ] = { | |
1370 | [ C(OP_READ) ] = { | |
1371 | [ C(RESULT_ACCESS) ] = 0, | |
1372 | [ C(RESULT_MISS) ] = 0x0104, /* LD_DCU_MISS */ | |
1373 | }, | |
1374 | [ C(OP_WRITE) ] = { | |
1375 | [ C(RESULT_ACCESS) ] = 0, | |
1376 | [ C(RESULT_MISS) ] = 0, | |
1377 | }, | |
1378 | [ C(OP_PREFETCH) ] = { | |
1379 | [ C(RESULT_ACCESS) ] = 0, | |
1380 | [ C(RESULT_MISS) ] = 0, | |
1381 | }, | |
1382 | }, | |
1383 | [ C(L1I ) ] = { | |
1384 | [ C(OP_READ) ] = { | |
1385 | [ C(RESULT_ACCESS) ] = 0x0380, /* ICACHE.ACCESSES */ | |
1386 | [ C(RESULT_MISS) ] = 0x0280, /* ICACGE.MISSES */ | |
1387 | }, | |
1388 | [ C(OP_WRITE) ] = { | |
1389 | [ C(RESULT_ACCESS) ] = -1, | |
1390 | [ C(RESULT_MISS) ] = -1, | |
1391 | }, | |
1392 | [ C(OP_PREFETCH) ] = { | |
1393 | [ C(RESULT_ACCESS) ] = 0, | |
1394 | [ C(RESULT_MISS) ] = 0, | |
1395 | }, | |
1396 | }, | |
1397 | [ C(LL ) ] = { | |
1398 | [ C(OP_READ) ] = { | |
1399 | /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ | |
1400 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
6d374056 | 1401 | [ C(RESULT_MISS) ] = 0, |
1fa64180 YZ |
1402 | }, |
1403 | [ C(OP_WRITE) ] = { | |
1404 | /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ | |
1405 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
1406 | /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ | |
1407 | [ C(RESULT_MISS) ] = 0x01b7, | |
1408 | }, | |
1409 | [ C(OP_PREFETCH) ] = { | |
1410 | /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ | |
1411 | [ C(RESULT_ACCESS) ] = 0x01b7, | |
1412 | /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ | |
1413 | [ C(RESULT_MISS) ] = 0x01b7, | |
1414 | }, | |
1415 | }, | |
1416 | [ C(DTLB) ] = { | |
1417 | [ C(OP_READ) ] = { | |
1418 | [ C(RESULT_ACCESS) ] = 0, | |
1419 | [ C(RESULT_MISS) ] = 0x0804, /* LD_DTLB_MISS */ | |
1420 | }, | |
1421 | [ C(OP_WRITE) ] = { | |
1422 | [ C(RESULT_ACCESS) ] = 0, | |
1423 | [ C(RESULT_MISS) ] = 0, | |
1424 | }, | |
1425 | [ C(OP_PREFETCH) ] = { | |
1426 | [ C(RESULT_ACCESS) ] = 0, | |
1427 | [ C(RESULT_MISS) ] = 0, | |
1428 | }, | |
1429 | }, | |
1430 | [ C(ITLB) ] = { | |
1431 | [ C(OP_READ) ] = { | |
1432 | [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ | |
6d374056 | 1433 | [ C(RESULT_MISS) ] = 0x40205, /* PAGE_WALKS.I_SIDE_WALKS */ |
1fa64180 YZ |
1434 | }, |
1435 | [ C(OP_WRITE) ] = { | |
1436 | [ C(RESULT_ACCESS) ] = -1, | |
1437 | [ C(RESULT_MISS) ] = -1, | |
1438 | }, | |
1439 | [ C(OP_PREFETCH) ] = { | |
1440 | [ C(RESULT_ACCESS) ] = -1, | |
1441 | [ C(RESULT_MISS) ] = -1, | |
1442 | }, | |
1443 | }, | |
1444 | [ C(BPU ) ] = { | |
1445 | [ C(OP_READ) ] = { | |
1446 | [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ | |
1447 | [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ | |
1448 | }, | |
1449 | [ C(OP_WRITE) ] = { | |
1450 | [ C(RESULT_ACCESS) ] = -1, | |
1451 | [ C(RESULT_MISS) ] = -1, | |
1452 | }, | |
1453 | [ C(OP_PREFETCH) ] = { | |
1454 | [ C(RESULT_ACCESS) ] = -1, | |
1455 | [ C(RESULT_MISS) ] = -1, | |
1456 | }, | |
1457 | }, | |
1458 | }; | |
1459 | ||
1a78d937 AK |
1460 | /* |
1461 | * Use from PMIs where the LBRs are already disabled. | |
1462 | */ | |
1463 | static void __intel_pmu_disable_all(void) | |
f22f54f4 | 1464 | { |
89cbc767 | 1465 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
f22f54f4 PZ |
1466 | |
1467 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); | |
1468 | ||
15c7ad51 | 1469 | if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) |
f22f54f4 | 1470 | intel_pmu_disable_bts(); |
8062382c AS |
1471 | else |
1472 | intel_bts_disable_local(); | |
ca037701 PZ |
1473 | |
1474 | intel_pmu_pebs_disable_all(); | |
1a78d937 AK |
1475 | } |
1476 | ||
1477 | static void intel_pmu_disable_all(void) | |
1478 | { | |
1479 | __intel_pmu_disable_all(); | |
caff2bef | 1480 | intel_pmu_lbr_disable_all(); |
f22f54f4 PZ |
1481 | } |
1482 | ||
1a78d937 | 1483 | static void __intel_pmu_enable_all(int added, bool pmi) |
f22f54f4 | 1484 | { |
89cbc767 | 1485 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
f22f54f4 | 1486 | |
d329527e | 1487 | intel_pmu_pebs_enable_all(); |
1a78d937 | 1488 | intel_pmu_lbr_enable_all(pmi); |
144d31e6 GN |
1489 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, |
1490 | x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask); | |
f22f54f4 | 1491 | |
15c7ad51 | 1492 | if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) { |
f22f54f4 | 1493 | struct perf_event *event = |
15c7ad51 | 1494 | cpuc->events[INTEL_PMC_IDX_FIXED_BTS]; |
f22f54f4 PZ |
1495 | |
1496 | if (WARN_ON_ONCE(!event)) | |
1497 | return; | |
1498 | ||
1499 | intel_pmu_enable_bts(event->hw.config); | |
8062382c AS |
1500 | } else |
1501 | intel_bts_enable_local(); | |
f22f54f4 PZ |
1502 | } |
1503 | ||
1a78d937 AK |
1504 | static void intel_pmu_enable_all(int added) |
1505 | { | |
1506 | __intel_pmu_enable_all(added, false); | |
1507 | } | |
1508 | ||
11164cd4 PZ |
1509 | /* |
1510 | * Workaround for: | |
1511 | * Intel Errata AAK100 (model 26) | |
1512 | * Intel Errata AAP53 (model 30) | |
40b91cd1 | 1513 | * Intel Errata BD53 (model 44) |
11164cd4 | 1514 | * |
351af072 ZY |
1515 | * The official story: |
1516 | * These chips need to be 'reset' when adding counters by programming the | |
1517 | * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either | |
1518 | * in sequence on the same PMC or on different PMCs. | |
1519 | * | |
1520 | * In practise it appears some of these events do in fact count, and | |
1521 | * we need to programm all 4 events. | |
11164cd4 | 1522 | */ |
351af072 | 1523 | static void intel_pmu_nhm_workaround(void) |
11164cd4 | 1524 | { |
89cbc767 | 1525 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
351af072 ZY |
1526 | static const unsigned long nhm_magic[4] = { |
1527 | 0x4300B5, | |
1528 | 0x4300D2, | |
1529 | 0x4300B1, | |
1530 | 0x4300B1 | |
1531 | }; | |
1532 | struct perf_event *event; | |
1533 | int i; | |
11164cd4 | 1534 | |
351af072 ZY |
1535 | /* |
1536 | * The Errata requires below steps: | |
1537 | * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL; | |
1538 | * 2) Configure 4 PERFEVTSELx with the magic events and clear | |
1539 | * the corresponding PMCx; | |
1540 | * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL; | |
1541 | * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL; | |
1542 | * 5) Clear 4 pairs of ERFEVTSELx and PMCx; | |
1543 | */ | |
11164cd4 | 1544 | |
351af072 ZY |
1545 | /* |
1546 | * The real steps we choose are a little different from above. | |
1547 | * A) To reduce MSR operations, we don't run step 1) as they | |
1548 | * are already cleared before this function is called; | |
1549 | * B) Call x86_perf_event_update to save PMCx before configuring | |
1550 | * PERFEVTSELx with magic number; | |
1551 | * C) With step 5), we do clear only when the PERFEVTSELx is | |
1552 | * not used currently. | |
1553 | * D) Call x86_perf_event_set_period to restore PMCx; | |
1554 | */ | |
11164cd4 | 1555 | |
351af072 ZY |
1556 | /* We always operate 4 pairs of PERF Counters */ |
1557 | for (i = 0; i < 4; i++) { | |
1558 | event = cpuc->events[i]; | |
1559 | if (event) | |
1560 | x86_perf_event_update(event); | |
1561 | } | |
11164cd4 | 1562 | |
351af072 ZY |
1563 | for (i = 0; i < 4; i++) { |
1564 | wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]); | |
1565 | wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0); | |
1566 | } | |
1567 | ||
1568 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf); | |
1569 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0); | |
11164cd4 | 1570 | |
351af072 ZY |
1571 | for (i = 0; i < 4; i++) { |
1572 | event = cpuc->events[i]; | |
1573 | ||
1574 | if (event) { | |
1575 | x86_perf_event_set_period(event); | |
31fa58af | 1576 | __x86_pmu_enable_event(&event->hw, |
351af072 ZY |
1577 | ARCH_PERFMON_EVENTSEL_ENABLE); |
1578 | } else | |
1579 | wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0); | |
11164cd4 | 1580 | } |
351af072 ZY |
1581 | } |
1582 | ||
1583 | static void intel_pmu_nhm_enable_all(int added) | |
1584 | { | |
1585 | if (added) | |
1586 | intel_pmu_nhm_workaround(); | |
11164cd4 PZ |
1587 | intel_pmu_enable_all(added); |
1588 | } | |
1589 | ||
f22f54f4 PZ |
1590 | static inline u64 intel_pmu_get_status(void) |
1591 | { | |
1592 | u64 status; | |
1593 | ||
1594 | rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status); | |
1595 | ||
1596 | return status; | |
1597 | } | |
1598 | ||
1599 | static inline void intel_pmu_ack_status(u64 ack) | |
1600 | { | |
1601 | wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack); | |
1602 | } | |
1603 | ||
ca037701 | 1604 | static void intel_pmu_disable_fixed(struct hw_perf_event *hwc) |
f22f54f4 | 1605 | { |
15c7ad51 | 1606 | int idx = hwc->idx - INTEL_PMC_IDX_FIXED; |
f22f54f4 PZ |
1607 | u64 ctrl_val, mask; |
1608 | ||
1609 | mask = 0xfULL << (idx * 4); | |
1610 | ||
1611 | rdmsrl(hwc->config_base, ctrl_val); | |
1612 | ctrl_val &= ~mask; | |
7645a24c | 1613 | wrmsrl(hwc->config_base, ctrl_val); |
f22f54f4 PZ |
1614 | } |
1615 | ||
2b9e344d PZ |
1616 | static inline bool event_is_checkpointed(struct perf_event *event) |
1617 | { | |
1618 | return (event->hw.config & HSW_IN_TX_CHECKPOINTED) != 0; | |
1619 | } | |
1620 | ||
ca037701 | 1621 | static void intel_pmu_disable_event(struct perf_event *event) |
f22f54f4 | 1622 | { |
aff3d91a | 1623 | struct hw_perf_event *hwc = &event->hw; |
89cbc767 | 1624 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
aff3d91a | 1625 | |
15c7ad51 | 1626 | if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) { |
f22f54f4 PZ |
1627 | intel_pmu_disable_bts(); |
1628 | intel_pmu_drain_bts_buffer(); | |
1629 | return; | |
1630 | } | |
1631 | ||
144d31e6 GN |
1632 | cpuc->intel_ctrl_guest_mask &= ~(1ull << hwc->idx); |
1633 | cpuc->intel_ctrl_host_mask &= ~(1ull << hwc->idx); | |
2b9e344d | 1634 | cpuc->intel_cp_status &= ~(1ull << hwc->idx); |
144d31e6 | 1635 | |
60ce0fbd SE |
1636 | /* |
1637 | * must disable before any actual event | |
1638 | * because any event may be combined with LBR | |
1639 | */ | |
a46a2300 | 1640 | if (needs_branch_stack(event)) |
60ce0fbd SE |
1641 | intel_pmu_lbr_disable(event); |
1642 | ||
f22f54f4 | 1643 | if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { |
aff3d91a | 1644 | intel_pmu_disable_fixed(hwc); |
f22f54f4 PZ |
1645 | return; |
1646 | } | |
1647 | ||
aff3d91a | 1648 | x86_pmu_disable_event(event); |
ca037701 | 1649 | |
ab608344 | 1650 | if (unlikely(event->attr.precise_ip)) |
ef21f683 | 1651 | intel_pmu_pebs_disable(event); |
f22f54f4 PZ |
1652 | } |
1653 | ||
ca037701 | 1654 | static void intel_pmu_enable_fixed(struct hw_perf_event *hwc) |
f22f54f4 | 1655 | { |
15c7ad51 | 1656 | int idx = hwc->idx - INTEL_PMC_IDX_FIXED; |
f22f54f4 | 1657 | u64 ctrl_val, bits, mask; |
f22f54f4 PZ |
1658 | |
1659 | /* | |
1660 | * Enable IRQ generation (0x8), | |
1661 | * and enable ring-3 counting (0x2) and ring-0 counting (0x1) | |
1662 | * if requested: | |
1663 | */ | |
1664 | bits = 0x8ULL; | |
1665 | if (hwc->config & ARCH_PERFMON_EVENTSEL_USR) | |
1666 | bits |= 0x2; | |
1667 | if (hwc->config & ARCH_PERFMON_EVENTSEL_OS) | |
1668 | bits |= 0x1; | |
1669 | ||
1670 | /* | |
1671 | * ANY bit is supported in v3 and up | |
1672 | */ | |
1673 | if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY) | |
1674 | bits |= 0x4; | |
1675 | ||
1676 | bits <<= (idx * 4); | |
1677 | mask = 0xfULL << (idx * 4); | |
1678 | ||
1679 | rdmsrl(hwc->config_base, ctrl_val); | |
1680 | ctrl_val &= ~mask; | |
1681 | ctrl_val |= bits; | |
7645a24c | 1682 | wrmsrl(hwc->config_base, ctrl_val); |
f22f54f4 PZ |
1683 | } |
1684 | ||
aff3d91a | 1685 | static void intel_pmu_enable_event(struct perf_event *event) |
f22f54f4 | 1686 | { |
aff3d91a | 1687 | struct hw_perf_event *hwc = &event->hw; |
89cbc767 | 1688 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
aff3d91a | 1689 | |
15c7ad51 | 1690 | if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) { |
0a3aee0d | 1691 | if (!__this_cpu_read(cpu_hw_events.enabled)) |
f22f54f4 PZ |
1692 | return; |
1693 | ||
1694 | intel_pmu_enable_bts(hwc->config); | |
1695 | return; | |
1696 | } | |
60ce0fbd SE |
1697 | /* |
1698 | * must enabled before any actual event | |
1699 | * because any event may be combined with LBR | |
1700 | */ | |
a46a2300 | 1701 | if (needs_branch_stack(event)) |
60ce0fbd | 1702 | intel_pmu_lbr_enable(event); |
f22f54f4 | 1703 | |
144d31e6 GN |
1704 | if (event->attr.exclude_host) |
1705 | cpuc->intel_ctrl_guest_mask |= (1ull << hwc->idx); | |
1706 | if (event->attr.exclude_guest) | |
1707 | cpuc->intel_ctrl_host_mask |= (1ull << hwc->idx); | |
1708 | ||
2b9e344d PZ |
1709 | if (unlikely(event_is_checkpointed(event))) |
1710 | cpuc->intel_cp_status |= (1ull << hwc->idx); | |
1711 | ||
f22f54f4 | 1712 | if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { |
aff3d91a | 1713 | intel_pmu_enable_fixed(hwc); |
f22f54f4 PZ |
1714 | return; |
1715 | } | |
1716 | ||
ab608344 | 1717 | if (unlikely(event->attr.precise_ip)) |
ef21f683 | 1718 | intel_pmu_pebs_enable(event); |
ca037701 | 1719 | |
31fa58af | 1720 | __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); |
f22f54f4 PZ |
1721 | } |
1722 | ||
1723 | /* | |
1724 | * Save and restart an expired event. Called by NMI contexts, | |
1725 | * so it has to be careful about preempting normal event ops: | |
1726 | */ | |
de0428a7 | 1727 | int intel_pmu_save_and_restart(struct perf_event *event) |
f22f54f4 | 1728 | { |
cc2ad4ba | 1729 | x86_perf_event_update(event); |
2dbf0116 AK |
1730 | /* |
1731 | * For a checkpointed counter always reset back to 0. This | |
1732 | * avoids a situation where the counter overflows, aborts the | |
1733 | * transaction and is then set back to shortly before the | |
1734 | * overflow, and overflows and aborts again. | |
1735 | */ | |
1736 | if (unlikely(event_is_checkpointed(event))) { | |
1737 | /* No race with NMIs because the counter should not be armed */ | |
1738 | wrmsrl(event->hw.event_base, 0); | |
1739 | local64_set(&event->hw.prev_count, 0); | |
1740 | } | |
cc2ad4ba | 1741 | return x86_perf_event_set_period(event); |
f22f54f4 PZ |
1742 | } |
1743 | ||
1744 | static void intel_pmu_reset(void) | |
1745 | { | |
0a3aee0d | 1746 | struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds); |
f22f54f4 PZ |
1747 | unsigned long flags; |
1748 | int idx; | |
1749 | ||
948b1bb8 | 1750 | if (!x86_pmu.num_counters) |
f22f54f4 PZ |
1751 | return; |
1752 | ||
1753 | local_irq_save(flags); | |
1754 | ||
c767a54b | 1755 | pr_info("clearing PMU state on CPU#%d\n", smp_processor_id()); |
f22f54f4 | 1756 | |
948b1bb8 | 1757 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
715c85b1 PA |
1758 | wrmsrl_safe(x86_pmu_config_addr(idx), 0ull); |
1759 | wrmsrl_safe(x86_pmu_event_addr(idx), 0ull); | |
f22f54f4 | 1760 | } |
948b1bb8 | 1761 | for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) |
715c85b1 | 1762 | wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull); |
948b1bb8 | 1763 | |
f22f54f4 PZ |
1764 | if (ds) |
1765 | ds->bts_index = ds->bts_buffer_base; | |
1766 | ||
8882edf7 AK |
1767 | /* Ack all overflows and disable fixed counters */ |
1768 | if (x86_pmu.version >= 2) { | |
1769 | intel_pmu_ack_status(intel_pmu_get_status()); | |
1770 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); | |
1771 | } | |
1772 | ||
1773 | /* Reset LBRs and LBR freezing */ | |
1774 | if (x86_pmu.lbr_nr) { | |
1775 | update_debugctlmsr(get_debugctlmsr() & | |
1776 | ~(DEBUGCTLMSR_FREEZE_LBRS_ON_PMI|DEBUGCTLMSR_LBR)); | |
1777 | } | |
1778 | ||
f22f54f4 PZ |
1779 | local_irq_restore(flags); |
1780 | } | |
1781 | ||
1782 | /* | |
1783 | * This handler is triggered by the local APIC, so the APIC IRQ handling | |
1784 | * rules apply: | |
1785 | */ | |
1786 | static int intel_pmu_handle_irq(struct pt_regs *regs) | |
1787 | { | |
1788 | struct perf_sample_data data; | |
1789 | struct cpu_hw_events *cpuc; | |
1790 | int bit, loops; | |
2e556b5b | 1791 | u64 status; |
b0b2072d | 1792 | int handled; |
f22f54f4 | 1793 | |
89cbc767 | 1794 | cpuc = this_cpu_ptr(&cpu_hw_events); |
f22f54f4 | 1795 | |
2bce5dac | 1796 | /* |
72db5596 AK |
1797 | * No known reason to not always do late ACK, |
1798 | * but just in case do it opt-in. | |
2bce5dac | 1799 | */ |
72db5596 AK |
1800 | if (!x86_pmu.late_ack) |
1801 | apic_write(APIC_LVTPC, APIC_DM_NMI); | |
1a78d937 | 1802 | __intel_pmu_disable_all(); |
b0b2072d | 1803 | handled = intel_pmu_drain_bts_buffer(); |
8062382c | 1804 | handled += intel_bts_interrupt(); |
f22f54f4 | 1805 | status = intel_pmu_get_status(); |
a3ef2229 MM |
1806 | if (!status) |
1807 | goto done; | |
f22f54f4 PZ |
1808 | |
1809 | loops = 0; | |
1810 | again: | |
0f29e573 | 1811 | intel_pmu_lbr_read(); |
2e556b5b | 1812 | intel_pmu_ack_status(status); |
f22f54f4 | 1813 | if (++loops > 100) { |
ae0def05 DH |
1814 | static bool warned = false; |
1815 | if (!warned) { | |
1816 | WARN(1, "perfevents: irq loop stuck!\n"); | |
1817 | perf_event_print_debug(); | |
1818 | warned = true; | |
1819 | } | |
f22f54f4 | 1820 | intel_pmu_reset(); |
3fb2b8dd | 1821 | goto done; |
f22f54f4 PZ |
1822 | } |
1823 | ||
1824 | inc_irq_stat(apic_perf_irqs); | |
ca037701 | 1825 | |
caff2bef | 1826 | |
b292d7a1 | 1827 | /* |
d8020bee AK |
1828 | * Ignore a range of extra bits in status that do not indicate |
1829 | * overflow by themselves. | |
b292d7a1 | 1830 | */ |
d8020bee AK |
1831 | status &= ~(GLOBAL_STATUS_COND_CHG | |
1832 | GLOBAL_STATUS_ASIF | | |
1833 | GLOBAL_STATUS_LBRS_FROZEN); | |
1834 | if (!status) | |
1835 | goto done; | |
b292d7a1 | 1836 | |
ca037701 PZ |
1837 | /* |
1838 | * PEBS overflow sets bit 62 in the global status register | |
1839 | */ | |
de725dec PZ |
1840 | if (__test_and_clear_bit(62, (unsigned long *)&status)) { |
1841 | handled++; | |
ca037701 | 1842 | x86_pmu.drain_pebs(regs); |
de725dec | 1843 | } |
ca037701 | 1844 | |
52ca9ced AS |
1845 | /* |
1846 | * Intel PT | |
1847 | */ | |
1848 | if (__test_and_clear_bit(55, (unsigned long *)&status)) { | |
1849 | handled++; | |
1850 | intel_pt_interrupt(); | |
1851 | } | |
1852 | ||
2dbf0116 | 1853 | /* |
2b9e344d PZ |
1854 | * Checkpointed counters can lead to 'spurious' PMIs because the |
1855 | * rollback caused by the PMI will have cleared the overflow status | |
1856 | * bit. Therefore always force probe these counters. | |
2dbf0116 | 1857 | */ |
2b9e344d | 1858 | status |= cpuc->intel_cp_status; |
2dbf0116 | 1859 | |
984b3f57 | 1860 | for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) { |
f22f54f4 PZ |
1861 | struct perf_event *event = cpuc->events[bit]; |
1862 | ||
de725dec PZ |
1863 | handled++; |
1864 | ||
f22f54f4 PZ |
1865 | if (!test_bit(bit, cpuc->active_mask)) |
1866 | continue; | |
1867 | ||
1868 | if (!intel_pmu_save_and_restart(event)) | |
1869 | continue; | |
1870 | ||
fd0d000b | 1871 | perf_sample_data_init(&data, 0, event->hw.last_period); |
f22f54f4 | 1872 | |
60ce0fbd SE |
1873 | if (has_branch_stack(event)) |
1874 | data.br_stack = &cpuc->lbr_stack; | |
1875 | ||
a8b0ca17 | 1876 | if (perf_event_overflow(event, &data, regs)) |
a4eaf7f1 | 1877 | x86_pmu_stop(event, 0); |
f22f54f4 PZ |
1878 | } |
1879 | ||
f22f54f4 PZ |
1880 | /* |
1881 | * Repeat if there is more work to be done: | |
1882 | */ | |
1883 | status = intel_pmu_get_status(); | |
1884 | if (status) | |
1885 | goto again; | |
1886 | ||
3fb2b8dd | 1887 | done: |
1a78d937 | 1888 | __intel_pmu_enable_all(0, true); |
72db5596 AK |
1889 | /* |
1890 | * Only unmask the NMI after the overflow counters | |
1891 | * have been reset. This avoids spurious NMIs on | |
1892 | * Haswell CPUs. | |
1893 | */ | |
1894 | if (x86_pmu.late_ack) | |
1895 | apic_write(APIC_LVTPC, APIC_DM_NMI); | |
de725dec | 1896 | return handled; |
f22f54f4 PZ |
1897 | } |
1898 | ||
f22f54f4 | 1899 | static struct event_constraint * |
ca037701 | 1900 | intel_bts_constraints(struct perf_event *event) |
f22f54f4 | 1901 | { |
ca037701 PZ |
1902 | struct hw_perf_event *hwc = &event->hw; |
1903 | unsigned int hw_event, bts_event; | |
f22f54f4 | 1904 | |
18a073a3 PZ |
1905 | if (event->attr.freq) |
1906 | return NULL; | |
1907 | ||
ca037701 PZ |
1908 | hw_event = hwc->config & INTEL_ARCH_EVENT_MASK; |
1909 | bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS); | |
f22f54f4 | 1910 | |
ca037701 | 1911 | if (unlikely(hw_event == bts_event && hwc->sample_period == 1)) |
f22f54f4 | 1912 | return &bts_constraint; |
ca037701 | 1913 | |
f22f54f4 PZ |
1914 | return NULL; |
1915 | } | |
1916 | ||
ae3f011f | 1917 | static int intel_alt_er(int idx, u64 config) |
b79e8941 | 1918 | { |
ae3f011f | 1919 | int alt_idx; |
9a5e3fb5 | 1920 | if (!(x86_pmu.flags & PMU_FL_HAS_RSP_1)) |
5a425294 | 1921 | return idx; |
b79e8941 | 1922 | |
5a425294 | 1923 | if (idx == EXTRA_REG_RSP_0) |
ae3f011f | 1924 | alt_idx = EXTRA_REG_RSP_1; |
5a425294 PZ |
1925 | |
1926 | if (idx == EXTRA_REG_RSP_1) | |
ae3f011f | 1927 | alt_idx = EXTRA_REG_RSP_0; |
5a425294 | 1928 | |
ae3f011f KL |
1929 | if (config & ~x86_pmu.extra_regs[alt_idx].valid_mask) |
1930 | return idx; | |
1931 | ||
1932 | return alt_idx; | |
5a425294 PZ |
1933 | } |
1934 | ||
1935 | static void intel_fixup_er(struct perf_event *event, int idx) | |
1936 | { | |
1937 | event->hw.extra_reg.idx = idx; | |
1938 | ||
1939 | if (idx == EXTRA_REG_RSP_0) { | |
b79e8941 | 1940 | event->hw.config &= ~INTEL_ARCH_EVENT_MASK; |
53ad0447 | 1941 | event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_0].event; |
b79e8941 | 1942 | event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0; |
5a425294 PZ |
1943 | } else if (idx == EXTRA_REG_RSP_1) { |
1944 | event->hw.config &= ~INTEL_ARCH_EVENT_MASK; | |
53ad0447 | 1945 | event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_1].event; |
5a425294 | 1946 | event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1; |
b79e8941 | 1947 | } |
b79e8941 PZ |
1948 | } |
1949 | ||
efc9f05d SE |
1950 | /* |
1951 | * manage allocation of shared extra msr for certain events | |
1952 | * | |
1953 | * sharing can be: | |
1954 | * per-cpu: to be shared between the various events on a single PMU | |
1955 | * per-core: per-cpu + shared by HT threads | |
1956 | */ | |
a7e3ed1e | 1957 | static struct event_constraint * |
efc9f05d | 1958 | __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc, |
b36817e8 SE |
1959 | struct perf_event *event, |
1960 | struct hw_perf_event_extra *reg) | |
a7e3ed1e | 1961 | { |
efc9f05d | 1962 | struct event_constraint *c = &emptyconstraint; |
a7e3ed1e | 1963 | struct er_account *era; |
cd8a38d3 | 1964 | unsigned long flags; |
5a425294 | 1965 | int idx = reg->idx; |
a7e3ed1e | 1966 | |
5a425294 PZ |
1967 | /* |
1968 | * reg->alloc can be set due to existing state, so for fake cpuc we | |
1969 | * need to ignore this, otherwise we might fail to allocate proper fake | |
1970 | * state for this extra reg constraint. Also see the comment below. | |
1971 | */ | |
1972 | if (reg->alloc && !cpuc->is_fake) | |
b36817e8 | 1973 | return NULL; /* call x86_get_event_constraint() */ |
a7e3ed1e | 1974 | |
b79e8941 | 1975 | again: |
5a425294 | 1976 | era = &cpuc->shared_regs->regs[idx]; |
cd8a38d3 SE |
1977 | /* |
1978 | * we use spin_lock_irqsave() to avoid lockdep issues when | |
1979 | * passing a fake cpuc | |
1980 | */ | |
1981 | raw_spin_lock_irqsave(&era->lock, flags); | |
efc9f05d SE |
1982 | |
1983 | if (!atomic_read(&era->ref) || era->config == reg->config) { | |
1984 | ||
5a425294 PZ |
1985 | /* |
1986 | * If its a fake cpuc -- as per validate_{group,event}() we | |
1987 | * shouldn't touch event state and we can avoid doing so | |
1988 | * since both will only call get_event_constraints() once | |
1989 | * on each event, this avoids the need for reg->alloc. | |
1990 | * | |
1991 | * Not doing the ER fixup will only result in era->reg being | |
1992 | * wrong, but since we won't actually try and program hardware | |
1993 | * this isn't a problem either. | |
1994 | */ | |
1995 | if (!cpuc->is_fake) { | |
1996 | if (idx != reg->idx) | |
1997 | intel_fixup_er(event, idx); | |
1998 | ||
1999 | /* | |
2000 | * x86_schedule_events() can call get_event_constraints() | |
2001 | * multiple times on events in the case of incremental | |
2002 | * scheduling(). reg->alloc ensures we only do the ER | |
2003 | * allocation once. | |
2004 | */ | |
2005 | reg->alloc = 1; | |
2006 | } | |
2007 | ||
efc9f05d SE |
2008 | /* lock in msr value */ |
2009 | era->config = reg->config; | |
2010 | era->reg = reg->reg; | |
2011 | ||
2012 | /* one more user */ | |
2013 | atomic_inc(&era->ref); | |
2014 | ||
a7e3ed1e | 2015 | /* |
b36817e8 SE |
2016 | * need to call x86_get_event_constraint() |
2017 | * to check if associated event has constraints | |
a7e3ed1e | 2018 | */ |
b36817e8 | 2019 | c = NULL; |
5a425294 | 2020 | } else { |
ae3f011f | 2021 | idx = intel_alt_er(idx, reg->config); |
5a425294 PZ |
2022 | if (idx != reg->idx) { |
2023 | raw_spin_unlock_irqrestore(&era->lock, flags); | |
2024 | goto again; | |
2025 | } | |
a7e3ed1e | 2026 | } |
cd8a38d3 | 2027 | raw_spin_unlock_irqrestore(&era->lock, flags); |
a7e3ed1e | 2028 | |
efc9f05d SE |
2029 | return c; |
2030 | } | |
2031 | ||
2032 | static void | |
2033 | __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc, | |
2034 | struct hw_perf_event_extra *reg) | |
2035 | { | |
2036 | struct er_account *era; | |
2037 | ||
2038 | /* | |
5a425294 PZ |
2039 | * Only put constraint if extra reg was actually allocated. Also takes |
2040 | * care of event which do not use an extra shared reg. | |
2041 | * | |
2042 | * Also, if this is a fake cpuc we shouldn't touch any event state | |
2043 | * (reg->alloc) and we don't care about leaving inconsistent cpuc state | |
2044 | * either since it'll be thrown out. | |
efc9f05d | 2045 | */ |
5a425294 | 2046 | if (!reg->alloc || cpuc->is_fake) |
efc9f05d SE |
2047 | return; |
2048 | ||
2049 | era = &cpuc->shared_regs->regs[reg->idx]; | |
2050 | ||
2051 | /* one fewer user */ | |
2052 | atomic_dec(&era->ref); | |
2053 | ||
2054 | /* allocate again next time */ | |
2055 | reg->alloc = 0; | |
2056 | } | |
2057 | ||
2058 | static struct event_constraint * | |
2059 | intel_shared_regs_constraints(struct cpu_hw_events *cpuc, | |
2060 | struct perf_event *event) | |
2061 | { | |
b36817e8 SE |
2062 | struct event_constraint *c = NULL, *d; |
2063 | struct hw_perf_event_extra *xreg, *breg; | |
2064 | ||
2065 | xreg = &event->hw.extra_reg; | |
2066 | if (xreg->idx != EXTRA_REG_NONE) { | |
2067 | c = __intel_shared_reg_get_constraints(cpuc, event, xreg); | |
2068 | if (c == &emptyconstraint) | |
2069 | return c; | |
2070 | } | |
2071 | breg = &event->hw.branch_reg; | |
2072 | if (breg->idx != EXTRA_REG_NONE) { | |
2073 | d = __intel_shared_reg_get_constraints(cpuc, event, breg); | |
2074 | if (d == &emptyconstraint) { | |
2075 | __intel_shared_reg_put_constraints(cpuc, xreg); | |
2076 | c = d; | |
2077 | } | |
2078 | } | |
efc9f05d | 2079 | return c; |
a7e3ed1e AK |
2080 | } |
2081 | ||
de0428a7 | 2082 | struct event_constraint * |
79cba822 SE |
2083 | x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx, |
2084 | struct perf_event *event) | |
de0428a7 KW |
2085 | { |
2086 | struct event_constraint *c; | |
2087 | ||
2088 | if (x86_pmu.event_constraints) { | |
2089 | for_each_event_constraint(c, x86_pmu.event_constraints) { | |
9fac2cf3 | 2090 | if ((event->hw.config & c->cmask) == c->code) { |
9fac2cf3 | 2091 | event->hw.flags |= c->flags; |
de0428a7 | 2092 | return c; |
9fac2cf3 | 2093 | } |
de0428a7 KW |
2094 | } |
2095 | } | |
2096 | ||
2097 | return &unconstrained; | |
2098 | } | |
2099 | ||
f22f54f4 | 2100 | static struct event_constraint * |
e979121b | 2101 | __intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx, |
79cba822 | 2102 | struct perf_event *event) |
f22f54f4 PZ |
2103 | { |
2104 | struct event_constraint *c; | |
2105 | ||
ca037701 PZ |
2106 | c = intel_bts_constraints(event); |
2107 | if (c) | |
2108 | return c; | |
2109 | ||
687805e4 | 2110 | c = intel_shared_regs_constraints(cpuc, event); |
f22f54f4 PZ |
2111 | if (c) |
2112 | return c; | |
2113 | ||
687805e4 | 2114 | c = intel_pebs_constraints(event); |
a7e3ed1e AK |
2115 | if (c) |
2116 | return c; | |
2117 | ||
79cba822 | 2118 | return x86_get_event_constraints(cpuc, idx, event); |
f22f54f4 PZ |
2119 | } |
2120 | ||
e979121b MD |
2121 | static void |
2122 | intel_start_scheduling(struct cpu_hw_events *cpuc) | |
2123 | { | |
2124 | struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; | |
1c565833 | 2125 | struct intel_excl_states *xl; |
e979121b | 2126 | int tid = cpuc->excl_thread_id; |
e979121b MD |
2127 | |
2128 | /* | |
2129 | * nothing needed if in group validation mode | |
2130 | */ | |
b37609c3 | 2131 | if (cpuc->is_fake || !is_ht_workaround_enabled()) |
e979121b | 2132 | return; |
b37609c3 | 2133 | |
e979121b MD |
2134 | /* |
2135 | * no exclusion needed | |
2136 | */ | |
17186ccd | 2137 | if (WARN_ON_ONCE(!excl_cntrs)) |
e979121b MD |
2138 | return; |
2139 | ||
e979121b MD |
2140 | xl = &excl_cntrs->states[tid]; |
2141 | ||
2142 | xl->sched_started = true; | |
e979121b MD |
2143 | /* |
2144 | * lock shared state until we are done scheduling | |
2145 | * in stop_event_scheduling() | |
2146 | * makes scheduling appear as a transaction | |
2147 | */ | |
e979121b | 2148 | raw_spin_lock(&excl_cntrs->lock); |
e979121b MD |
2149 | } |
2150 | ||
0c41e756 PZ |
2151 | static void intel_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr) |
2152 | { | |
2153 | struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; | |
2154 | struct event_constraint *c = cpuc->event_constraint[idx]; | |
2155 | struct intel_excl_states *xl; | |
2156 | int tid = cpuc->excl_thread_id; | |
2157 | ||
2158 | if (cpuc->is_fake || !is_ht_workaround_enabled()) | |
2159 | return; | |
2160 | ||
2161 | if (WARN_ON_ONCE(!excl_cntrs)) | |
2162 | return; | |
2163 | ||
2164 | if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) | |
2165 | return; | |
2166 | ||
2167 | xl = &excl_cntrs->states[tid]; | |
2168 | ||
2169 | lockdep_assert_held(&excl_cntrs->lock); | |
2170 | ||
1fe684e3 | 2171 | if (c->flags & PERF_X86_EVENT_EXCL) |
43ef205b | 2172 | xl->state[cntr] = INTEL_EXCL_EXCLUSIVE; |
1fe684e3 | 2173 | else |
43ef205b | 2174 | xl->state[cntr] = INTEL_EXCL_SHARED; |
0c41e756 PZ |
2175 | } |
2176 | ||
e979121b MD |
2177 | static void |
2178 | intel_stop_scheduling(struct cpu_hw_events *cpuc) | |
2179 | { | |
2180 | struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; | |
1c565833 | 2181 | struct intel_excl_states *xl; |
e979121b | 2182 | int tid = cpuc->excl_thread_id; |
e979121b MD |
2183 | |
2184 | /* | |
2185 | * nothing needed if in group validation mode | |
2186 | */ | |
b37609c3 | 2187 | if (cpuc->is_fake || !is_ht_workaround_enabled()) |
e979121b MD |
2188 | return; |
2189 | /* | |
2190 | * no exclusion needed | |
2191 | */ | |
17186ccd | 2192 | if (WARN_ON_ONCE(!excl_cntrs)) |
e979121b MD |
2193 | return; |
2194 | ||
e979121b MD |
2195 | xl = &excl_cntrs->states[tid]; |
2196 | ||
e979121b MD |
2197 | xl->sched_started = false; |
2198 | /* | |
2199 | * release shared state lock (acquired in intel_start_scheduling()) | |
2200 | */ | |
2201 | raw_spin_unlock(&excl_cntrs->lock); | |
2202 | } | |
2203 | ||
2204 | static struct event_constraint * | |
2205 | intel_get_excl_constraints(struct cpu_hw_events *cpuc, struct perf_event *event, | |
2206 | int idx, struct event_constraint *c) | |
2207 | { | |
e979121b | 2208 | struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; |
1c565833 | 2209 | struct intel_excl_states *xlo; |
e979121b | 2210 | int tid = cpuc->excl_thread_id; |
1c565833 | 2211 | int is_excl, i; |
e979121b MD |
2212 | |
2213 | /* | |
2214 | * validating a group does not require | |
2215 | * enforcing cross-thread exclusion | |
2216 | */ | |
b37609c3 SE |
2217 | if (cpuc->is_fake || !is_ht_workaround_enabled()) |
2218 | return c; | |
2219 | ||
2220 | /* | |
2221 | * no exclusion needed | |
2222 | */ | |
17186ccd | 2223 | if (WARN_ON_ONCE(!excl_cntrs)) |
e979121b | 2224 | return c; |
e979121b | 2225 | |
e979121b MD |
2226 | /* |
2227 | * because we modify the constraint, we need | |
2228 | * to make a copy. Static constraints come | |
2229 | * from static const tables. | |
2230 | * | |
2231 | * only needed when constraint has not yet | |
2232 | * been cloned (marked dynamic) | |
2233 | */ | |
2234 | if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) { | |
aaf932e8 | 2235 | struct event_constraint *cx; |
e979121b | 2236 | |
e979121b MD |
2237 | /* |
2238 | * grab pre-allocated constraint entry | |
2239 | */ | |
2240 | cx = &cpuc->constraint_list[idx]; | |
2241 | ||
2242 | /* | |
2243 | * initialize dynamic constraint | |
2244 | * with static constraint | |
2245 | */ | |
aaf932e8 | 2246 | *cx = *c; |
e979121b MD |
2247 | |
2248 | /* | |
2249 | * mark constraint as dynamic, so we | |
2250 | * can free it later on | |
2251 | */ | |
2252 | cx->flags |= PERF_X86_EVENT_DYNAMIC; | |
aaf932e8 | 2253 | c = cx; |
e979121b MD |
2254 | } |
2255 | ||
2256 | /* | |
2257 | * From here on, the constraint is dynamic. | |
2258 | * Either it was just allocated above, or it | |
2259 | * was allocated during a earlier invocation | |
2260 | * of this function | |
2261 | */ | |
2262 | ||
1c565833 PZ |
2263 | /* |
2264 | * state of sibling HT | |
2265 | */ | |
2266 | xlo = &excl_cntrs->states[tid ^ 1]; | |
2267 | ||
2268 | /* | |
2269 | * event requires exclusive counter access | |
2270 | * across HT threads | |
2271 | */ | |
2272 | is_excl = c->flags & PERF_X86_EVENT_EXCL; | |
2273 | if (is_excl && !(event->hw.flags & PERF_X86_EVENT_EXCL_ACCT)) { | |
2274 | event->hw.flags |= PERF_X86_EVENT_EXCL_ACCT; | |
2275 | if (!cpuc->n_excl++) | |
2276 | WRITE_ONCE(excl_cntrs->has_exclusive[tid], 1); | |
2277 | } | |
2278 | ||
e979121b MD |
2279 | /* |
2280 | * Modify static constraint with current dynamic | |
2281 | * state of thread | |
2282 | * | |
2283 | * EXCLUSIVE: sibling counter measuring exclusive event | |
2284 | * SHARED : sibling counter measuring non-exclusive event | |
2285 | * UNUSED : sibling counter unused | |
2286 | */ | |
aaf932e8 | 2287 | for_each_set_bit(i, c->idxmsk, X86_PMC_IDX_MAX) { |
e979121b MD |
2288 | /* |
2289 | * exclusive event in sibling counter | |
2290 | * our corresponding counter cannot be used | |
2291 | * regardless of our event | |
2292 | */ | |
1c565833 | 2293 | if (xlo->state[i] == INTEL_EXCL_EXCLUSIVE) |
aaf932e8 | 2294 | __clear_bit(i, c->idxmsk); |
e979121b MD |
2295 | /* |
2296 | * if measuring an exclusive event, sibling | |
2297 | * measuring non-exclusive, then counter cannot | |
2298 | * be used | |
2299 | */ | |
1c565833 | 2300 | if (is_excl && xlo->state[i] == INTEL_EXCL_SHARED) |
aaf932e8 | 2301 | __clear_bit(i, c->idxmsk); |
e979121b MD |
2302 | } |
2303 | ||
2304 | /* | |
2305 | * recompute actual bit weight for scheduling algorithm | |
2306 | */ | |
aaf932e8 | 2307 | c->weight = hweight64(c->idxmsk64); |
e979121b MD |
2308 | |
2309 | /* | |
2310 | * if we return an empty mask, then switch | |
2311 | * back to static empty constraint to avoid | |
2312 | * the cost of freeing later on | |
2313 | */ | |
aaf932e8 PZ |
2314 | if (c->weight == 0) |
2315 | c = &emptyconstraint; | |
e979121b | 2316 | |
aaf932e8 | 2317 | return c; |
e979121b MD |
2318 | } |
2319 | ||
2320 | static struct event_constraint * | |
2321 | intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx, | |
2322 | struct perf_event *event) | |
2323 | { | |
ebfb4988 | 2324 | struct event_constraint *c1 = NULL; |
a90738c2 | 2325 | struct event_constraint *c2; |
e979121b | 2326 | |
ebfb4988 PZ |
2327 | if (idx >= 0) /* fake does < 0 */ |
2328 | c1 = cpuc->event_constraint[idx]; | |
2329 | ||
e979121b MD |
2330 | /* |
2331 | * first time only | |
2332 | * - static constraint: no change across incremental scheduling calls | |
2333 | * - dynamic constraint: handled by intel_get_excl_constraints() | |
2334 | */ | |
a90738c2 SE |
2335 | c2 = __intel_get_event_constraints(cpuc, idx, event); |
2336 | if (c1 && (c1->flags & PERF_X86_EVENT_DYNAMIC)) { | |
2337 | bitmap_copy(c1->idxmsk, c2->idxmsk, X86_PMC_IDX_MAX); | |
2338 | c1->weight = c2->weight; | |
2339 | c2 = c1; | |
2340 | } | |
e979121b MD |
2341 | |
2342 | if (cpuc->excl_cntrs) | |
a90738c2 | 2343 | return intel_get_excl_constraints(cpuc, event, idx, c2); |
e979121b | 2344 | |
a90738c2 | 2345 | return c2; |
e979121b MD |
2346 | } |
2347 | ||
2348 | static void intel_put_excl_constraints(struct cpu_hw_events *cpuc, | |
2349 | struct perf_event *event) | |
2350 | { | |
2351 | struct hw_perf_event *hwc = &event->hw; | |
2352 | struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; | |
e979121b | 2353 | int tid = cpuc->excl_thread_id; |
1c565833 | 2354 | struct intel_excl_states *xl; |
e979121b MD |
2355 | |
2356 | /* | |
2357 | * nothing needed if in group validation mode | |
2358 | */ | |
2359 | if (cpuc->is_fake) | |
2360 | return; | |
2361 | ||
17186ccd | 2362 | if (WARN_ON_ONCE(!excl_cntrs)) |
e979121b MD |
2363 | return; |
2364 | ||
cc1790cf PZ |
2365 | if (hwc->flags & PERF_X86_EVENT_EXCL_ACCT) { |
2366 | hwc->flags &= ~PERF_X86_EVENT_EXCL_ACCT; | |
2367 | if (!--cpuc->n_excl) | |
2368 | WRITE_ONCE(excl_cntrs->has_exclusive[tid], 0); | |
2369 | } | |
e979121b MD |
2370 | |
2371 | /* | |
ba040653 PZ |
2372 | * If event was actually assigned, then mark the counter state as |
2373 | * unused now. | |
e979121b | 2374 | */ |
ba040653 PZ |
2375 | if (hwc->idx >= 0) { |
2376 | xl = &excl_cntrs->states[tid]; | |
2377 | ||
2378 | /* | |
2379 | * put_constraint may be called from x86_schedule_events() | |
2380 | * which already has the lock held so here make locking | |
2381 | * conditional. | |
2382 | */ | |
2383 | if (!xl->sched_started) | |
2384 | raw_spin_lock(&excl_cntrs->lock); | |
e979121b | 2385 | |
1c565833 | 2386 | xl->state[hwc->idx] = INTEL_EXCL_UNUSED; |
e979121b | 2387 | |
ba040653 PZ |
2388 | if (!xl->sched_started) |
2389 | raw_spin_unlock(&excl_cntrs->lock); | |
2390 | } | |
e979121b MD |
2391 | } |
2392 | ||
efc9f05d SE |
2393 | static void |
2394 | intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc, | |
a7e3ed1e AK |
2395 | struct perf_event *event) |
2396 | { | |
efc9f05d | 2397 | struct hw_perf_event_extra *reg; |
a7e3ed1e | 2398 | |
efc9f05d SE |
2399 | reg = &event->hw.extra_reg; |
2400 | if (reg->idx != EXTRA_REG_NONE) | |
2401 | __intel_shared_reg_put_constraints(cpuc, reg); | |
b36817e8 SE |
2402 | |
2403 | reg = &event->hw.branch_reg; | |
2404 | if (reg->idx != EXTRA_REG_NONE) | |
2405 | __intel_shared_reg_put_constraints(cpuc, reg); | |
efc9f05d | 2406 | } |
a7e3ed1e | 2407 | |
efc9f05d SE |
2408 | static void intel_put_event_constraints(struct cpu_hw_events *cpuc, |
2409 | struct perf_event *event) | |
2410 | { | |
2411 | intel_put_shared_regs_event_constraints(cpuc, event); | |
e979121b MD |
2412 | |
2413 | /* | |
2414 | * is PMU has exclusive counter restrictions, then | |
2415 | * all events are subject to and must call the | |
2416 | * put_excl_constraints() routine | |
2417 | */ | |
b371b594 | 2418 | if (cpuc->excl_cntrs) |
e979121b | 2419 | intel_put_excl_constraints(cpuc, event); |
e979121b MD |
2420 | } |
2421 | ||
0780c927 | 2422 | static void intel_pebs_aliases_core2(struct perf_event *event) |
b4cdc5c2 | 2423 | { |
0780c927 | 2424 | if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { |
7639dae0 PZ |
2425 | /* |
2426 | * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P | |
2427 | * (0x003c) so that we can use it with PEBS. | |
2428 | * | |
2429 | * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't | |
2430 | * PEBS capable. However we can use INST_RETIRED.ANY_P | |
2431 | * (0x00c0), which is a PEBS capable event, to get the same | |
2432 | * count. | |
2433 | * | |
2434 | * INST_RETIRED.ANY_P counts the number of cycles that retires | |
2435 | * CNTMASK instructions. By setting CNTMASK to a value (16) | |
2436 | * larger than the maximum number of instructions that can be | |
2437 | * retired per cycle (4) and then inverting the condition, we | |
2438 | * count all cycles that retire 16 or less instructions, which | |
2439 | * is every cycle. | |
2440 | * | |
2441 | * Thereby we gain a PEBS capable cycle counter. | |
2442 | */ | |
f9b4eeb8 PZ |
2443 | u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16); |
2444 | ||
0780c927 PZ |
2445 | alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); |
2446 | event->hw.config = alt_config; | |
2447 | } | |
2448 | } | |
2449 | ||
2450 | static void intel_pebs_aliases_snb(struct perf_event *event) | |
2451 | { | |
2452 | if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { | |
2453 | /* | |
2454 | * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P | |
2455 | * (0x003c) so that we can use it with PEBS. | |
2456 | * | |
2457 | * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't | |
2458 | * PEBS capable. However we can use UOPS_RETIRED.ALL | |
2459 | * (0x01c2), which is a PEBS capable event, to get the same | |
2460 | * count. | |
2461 | * | |
2462 | * UOPS_RETIRED.ALL counts the number of cycles that retires | |
2463 | * CNTMASK micro-ops. By setting CNTMASK to a value (16) | |
2464 | * larger than the maximum number of micro-ops that can be | |
2465 | * retired per cycle (4) and then inverting the condition, we | |
2466 | * count all cycles that retire 16 or less micro-ops, which | |
2467 | * is every cycle. | |
2468 | * | |
2469 | * Thereby we gain a PEBS capable cycle counter. | |
2470 | */ | |
2471 | u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16); | |
7639dae0 PZ |
2472 | |
2473 | alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); | |
2474 | event->hw.config = alt_config; | |
2475 | } | |
0780c927 PZ |
2476 | } |
2477 | ||
a7b58d21 AK |
2478 | static unsigned long intel_pmu_free_running_flags(struct perf_event *event) |
2479 | { | |
2480 | unsigned long flags = x86_pmu.free_running_flags; | |
2481 | ||
2482 | if (event->attr.use_clockid) | |
2483 | flags &= ~PERF_SAMPLE_TIME; | |
2484 | return flags; | |
2485 | } | |
2486 | ||
0780c927 PZ |
2487 | static int intel_pmu_hw_config(struct perf_event *event) |
2488 | { | |
2489 | int ret = x86_pmu_hw_config(event); | |
2490 | ||
2491 | if (ret) | |
2492 | return ret; | |
2493 | ||
851559e3 | 2494 | if (event->attr.precise_ip) { |
3569c0d7 | 2495 | if (!event->attr.freq) { |
851559e3 | 2496 | event->hw.flags |= PERF_X86_EVENT_AUTO_RELOAD; |
a7b58d21 AK |
2497 | if (!(event->attr.sample_type & |
2498 | ~intel_pmu_free_running_flags(event))) | |
3569c0d7 YZ |
2499 | event->hw.flags |= PERF_X86_EVENT_FREERUNNING; |
2500 | } | |
851559e3 YZ |
2501 | if (x86_pmu.pebs_aliases) |
2502 | x86_pmu.pebs_aliases(event); | |
2503 | } | |
7639dae0 | 2504 | |
a46a2300 | 2505 | if (needs_branch_stack(event)) { |
60ce0fbd SE |
2506 | ret = intel_pmu_setup_lbr_filter(event); |
2507 | if (ret) | |
2508 | return ret; | |
48070342 AS |
2509 | |
2510 | /* | |
2511 | * BTS is set up earlier in this path, so don't account twice | |
2512 | */ | |
2513 | if (!intel_pmu_has_bts(event)) { | |
2514 | /* disallow lbr if conflicting events are present */ | |
2515 | if (x86_add_exclusive(x86_lbr_exclusive_lbr)) | |
2516 | return -EBUSY; | |
2517 | ||
2518 | event->destroy = hw_perf_lbr_event_destroy; | |
2519 | } | |
60ce0fbd SE |
2520 | } |
2521 | ||
b4cdc5c2 PZ |
2522 | if (event->attr.type != PERF_TYPE_RAW) |
2523 | return 0; | |
2524 | ||
2525 | if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY)) | |
2526 | return 0; | |
2527 | ||
2528 | if (x86_pmu.version < 3) | |
2529 | return -EINVAL; | |
2530 | ||
2531 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) | |
2532 | return -EACCES; | |
2533 | ||
2534 | event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY; | |
2535 | ||
2536 | return 0; | |
2537 | } | |
2538 | ||
144d31e6 GN |
2539 | struct perf_guest_switch_msr *perf_guest_get_msrs(int *nr) |
2540 | { | |
2541 | if (x86_pmu.guest_get_msrs) | |
2542 | return x86_pmu.guest_get_msrs(nr); | |
2543 | *nr = 0; | |
2544 | return NULL; | |
2545 | } | |
2546 | EXPORT_SYMBOL_GPL(perf_guest_get_msrs); | |
2547 | ||
2548 | static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr) | |
2549 | { | |
89cbc767 | 2550 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
144d31e6 GN |
2551 | struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; |
2552 | ||
2553 | arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL; | |
2554 | arr[0].host = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask; | |
2555 | arr[0].guest = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_host_mask; | |
26a4f3c0 GN |
2556 | /* |
2557 | * If PMU counter has PEBS enabled it is not enough to disable counter | |
2558 | * on a guest entry since PEBS memory write can overshoot guest entry | |
2559 | * and corrupt guest memory. Disabling PEBS solves the problem. | |
2560 | */ | |
2561 | arr[1].msr = MSR_IA32_PEBS_ENABLE; | |
2562 | arr[1].host = cpuc->pebs_enabled; | |
2563 | arr[1].guest = 0; | |
144d31e6 | 2564 | |
26a4f3c0 | 2565 | *nr = 2; |
144d31e6 GN |
2566 | return arr; |
2567 | } | |
2568 | ||
2569 | static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr) | |
2570 | { | |
89cbc767 | 2571 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
144d31e6 GN |
2572 | struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; |
2573 | int idx; | |
2574 | ||
2575 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { | |
2576 | struct perf_event *event = cpuc->events[idx]; | |
2577 | ||
2578 | arr[idx].msr = x86_pmu_config_addr(idx); | |
2579 | arr[idx].host = arr[idx].guest = 0; | |
2580 | ||
2581 | if (!test_bit(idx, cpuc->active_mask)) | |
2582 | continue; | |
2583 | ||
2584 | arr[idx].host = arr[idx].guest = | |
2585 | event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE; | |
2586 | ||
2587 | if (event->attr.exclude_host) | |
2588 | arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE; | |
2589 | else if (event->attr.exclude_guest) | |
2590 | arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE; | |
2591 | } | |
2592 | ||
2593 | *nr = x86_pmu.num_counters; | |
2594 | return arr; | |
2595 | } | |
2596 | ||
2597 | static void core_pmu_enable_event(struct perf_event *event) | |
2598 | { | |
2599 | if (!event->attr.exclude_host) | |
2600 | x86_pmu_enable_event(event); | |
2601 | } | |
2602 | ||
2603 | static void core_pmu_enable_all(int added) | |
2604 | { | |
89cbc767 | 2605 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
144d31e6 GN |
2606 | int idx; |
2607 | ||
2608 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { | |
2609 | struct hw_perf_event *hwc = &cpuc->events[idx]->hw; | |
2610 | ||
2611 | if (!test_bit(idx, cpuc->active_mask) || | |
2612 | cpuc->events[idx]->attr.exclude_host) | |
2613 | continue; | |
2614 | ||
2615 | __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); | |
2616 | } | |
2617 | } | |
2618 | ||
3a632cb2 AK |
2619 | static int hsw_hw_config(struct perf_event *event) |
2620 | { | |
2621 | int ret = intel_pmu_hw_config(event); | |
2622 | ||
2623 | if (ret) | |
2624 | return ret; | |
2625 | if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE)) | |
2626 | return 0; | |
2627 | event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED); | |
2628 | ||
2629 | /* | |
2630 | * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with | |
2631 | * PEBS or in ANY thread mode. Since the results are non-sensical forbid | |
2632 | * this combination. | |
2633 | */ | |
2634 | if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) && | |
2635 | ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) || | |
2636 | event->attr.precise_ip > 0)) | |
2637 | return -EOPNOTSUPP; | |
2638 | ||
2dbf0116 AK |
2639 | if (event_is_checkpointed(event)) { |
2640 | /* | |
2641 | * Sampling of checkpointed events can cause situations where | |
2642 | * the CPU constantly aborts because of a overflow, which is | |
2643 | * then checkpointed back and ignored. Forbid checkpointing | |
2644 | * for sampling. | |
2645 | * | |
2646 | * But still allow a long sampling period, so that perf stat | |
2647 | * from KVM works. | |
2648 | */ | |
2649 | if (event->attr.sample_period > 0 && | |
2650 | event->attr.sample_period < 0x7fffffff) | |
2651 | return -EOPNOTSUPP; | |
2652 | } | |
3a632cb2 AK |
2653 | return 0; |
2654 | } | |
2655 | ||
2656 | static struct event_constraint counter2_constraint = | |
2657 | EVENT_CONSTRAINT(0, 0x4, 0); | |
2658 | ||
2659 | static struct event_constraint * | |
79cba822 SE |
2660 | hsw_get_event_constraints(struct cpu_hw_events *cpuc, int idx, |
2661 | struct perf_event *event) | |
3a632cb2 | 2662 | { |
79cba822 SE |
2663 | struct event_constraint *c; |
2664 | ||
2665 | c = intel_get_event_constraints(cpuc, idx, event); | |
3a632cb2 AK |
2666 | |
2667 | /* Handle special quirk on in_tx_checkpointed only in counter 2 */ | |
2668 | if (event->hw.config & HSW_IN_TX_CHECKPOINTED) { | |
2669 | if (c->idxmsk64 & (1U << 2)) | |
2670 | return &counter2_constraint; | |
2671 | return &emptyconstraint; | |
2672 | } | |
2673 | ||
2674 | return c; | |
2675 | } | |
2676 | ||
294fe0f5 AK |
2677 | /* |
2678 | * Broadwell: | |
2679 | * | |
2680 | * The INST_RETIRED.ALL period always needs to have lowest 6 bits cleared | |
2681 | * (BDM55) and it must not use a period smaller than 100 (BDM11). We combine | |
2682 | * the two to enforce a minimum period of 128 (the smallest value that has bits | |
2683 | * 0-5 cleared and >= 100). | |
2684 | * | |
2685 | * Because of how the code in x86_perf_event_set_period() works, the truncation | |
2686 | * of the lower 6 bits is 'harmless' as we'll occasionally add a longer period | |
2687 | * to make up for the 'lost' events due to carrying the 'error' in period_left. | |
2688 | * | |
2689 | * Therefore the effective (average) period matches the requested period, | |
2690 | * despite coarser hardware granularity. | |
2691 | */ | |
2692 | static unsigned bdw_limit_period(struct perf_event *event, unsigned left) | |
2693 | { | |
2694 | if ((event->hw.config & INTEL_ARCH_EVENT_MASK) == | |
2695 | X86_CONFIG(.event=0xc0, .umask=0x01)) { | |
2696 | if (left < 128) | |
2697 | left = 128; | |
2698 | left &= ~0x3fu; | |
2699 | } | |
2700 | return left; | |
2701 | } | |
2702 | ||
641cc938 JO |
2703 | PMU_FORMAT_ATTR(event, "config:0-7" ); |
2704 | PMU_FORMAT_ATTR(umask, "config:8-15" ); | |
2705 | PMU_FORMAT_ATTR(edge, "config:18" ); | |
2706 | PMU_FORMAT_ATTR(pc, "config:19" ); | |
2707 | PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */ | |
2708 | PMU_FORMAT_ATTR(inv, "config:23" ); | |
2709 | PMU_FORMAT_ATTR(cmask, "config:24-31" ); | |
3a632cb2 AK |
2710 | PMU_FORMAT_ATTR(in_tx, "config:32"); |
2711 | PMU_FORMAT_ATTR(in_tx_cp, "config:33"); | |
641cc938 JO |
2712 | |
2713 | static struct attribute *intel_arch_formats_attr[] = { | |
2714 | &format_attr_event.attr, | |
2715 | &format_attr_umask.attr, | |
2716 | &format_attr_edge.attr, | |
2717 | &format_attr_pc.attr, | |
2718 | &format_attr_inv.attr, | |
2719 | &format_attr_cmask.attr, | |
2720 | NULL, | |
2721 | }; | |
2722 | ||
0bf79d44 JO |
2723 | ssize_t intel_event_sysfs_show(char *page, u64 config) |
2724 | { | |
2725 | u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT); | |
2726 | ||
2727 | return x86_event_sysfs_show(page, config, event); | |
2728 | } | |
2729 | ||
de0428a7 | 2730 | struct intel_shared_regs *allocate_shared_regs(int cpu) |
efc9f05d SE |
2731 | { |
2732 | struct intel_shared_regs *regs; | |
2733 | int i; | |
2734 | ||
2735 | regs = kzalloc_node(sizeof(struct intel_shared_regs), | |
2736 | GFP_KERNEL, cpu_to_node(cpu)); | |
2737 | if (regs) { | |
2738 | /* | |
2739 | * initialize the locks to keep lockdep happy | |
2740 | */ | |
2741 | for (i = 0; i < EXTRA_REG_MAX; i++) | |
2742 | raw_spin_lock_init(®s->regs[i].lock); | |
2743 | ||
2744 | regs->core_id = -1; | |
2745 | } | |
2746 | return regs; | |
2747 | } | |
2748 | ||
6f6539ca MD |
2749 | static struct intel_excl_cntrs *allocate_excl_cntrs(int cpu) |
2750 | { | |
2751 | struct intel_excl_cntrs *c; | |
6f6539ca MD |
2752 | |
2753 | c = kzalloc_node(sizeof(struct intel_excl_cntrs), | |
2754 | GFP_KERNEL, cpu_to_node(cpu)); | |
2755 | if (c) { | |
2756 | raw_spin_lock_init(&c->lock); | |
6f6539ca MD |
2757 | c->core_id = -1; |
2758 | } | |
2759 | return c; | |
2760 | } | |
2761 | ||
a7e3ed1e AK |
2762 | static int intel_pmu_cpu_prepare(int cpu) |
2763 | { | |
2764 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); | |
2765 | ||
6f6539ca MD |
2766 | if (x86_pmu.extra_regs || x86_pmu.lbr_sel_map) { |
2767 | cpuc->shared_regs = allocate_shared_regs(cpu); | |
2768 | if (!cpuc->shared_regs) | |
dbc72b7a | 2769 | goto err; |
6f6539ca | 2770 | } |
69092624 | 2771 | |
6f6539ca MD |
2772 | if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) { |
2773 | size_t sz = X86_PMC_IDX_MAX * sizeof(struct event_constraint); | |
2774 | ||
2775 | cpuc->constraint_list = kzalloc(sz, GFP_KERNEL); | |
2776 | if (!cpuc->constraint_list) | |
dbc72b7a | 2777 | goto err_shared_regs; |
6f6539ca MD |
2778 | |
2779 | cpuc->excl_cntrs = allocate_excl_cntrs(cpu); | |
dbc72b7a PZ |
2780 | if (!cpuc->excl_cntrs) |
2781 | goto err_constraint_list; | |
2782 | ||
6f6539ca MD |
2783 | cpuc->excl_thread_id = 0; |
2784 | } | |
a7e3ed1e | 2785 | |
a7e3ed1e | 2786 | return NOTIFY_OK; |
dbc72b7a PZ |
2787 | |
2788 | err_constraint_list: | |
2789 | kfree(cpuc->constraint_list); | |
2790 | cpuc->constraint_list = NULL; | |
2791 | ||
2792 | err_shared_regs: | |
2793 | kfree(cpuc->shared_regs); | |
2794 | cpuc->shared_regs = NULL; | |
2795 | ||
2796 | err: | |
2797 | return NOTIFY_BAD; | |
a7e3ed1e AK |
2798 | } |
2799 | ||
74846d35 PZ |
2800 | static void intel_pmu_cpu_starting(int cpu) |
2801 | { | |
a7e3ed1e AK |
2802 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); |
2803 | int core_id = topology_core_id(cpu); | |
2804 | int i; | |
2805 | ||
69092624 LM |
2806 | init_debug_store_on_cpu(cpu); |
2807 | /* | |
2808 | * Deal with CPUs that don't clear their LBRs on power-up. | |
2809 | */ | |
2810 | intel_pmu_lbr_reset(); | |
2811 | ||
b36817e8 SE |
2812 | cpuc->lbr_sel = NULL; |
2813 | ||
2814 | if (!cpuc->shared_regs) | |
69092624 LM |
2815 | return; |
2816 | ||
9a5e3fb5 | 2817 | if (!(x86_pmu.flags & PMU_FL_NO_HT_SHARING)) { |
90413464 SE |
2818 | void **onln = &cpuc->kfree_on_online[X86_PERF_KFREE_SHARED]; |
2819 | ||
06931e62 | 2820 | for_each_cpu(i, topology_sibling_cpumask(cpu)) { |
b36817e8 | 2821 | struct intel_shared_regs *pc; |
a7e3ed1e | 2822 | |
b36817e8 SE |
2823 | pc = per_cpu(cpu_hw_events, i).shared_regs; |
2824 | if (pc && pc->core_id == core_id) { | |
90413464 | 2825 | *onln = cpuc->shared_regs; |
b36817e8 SE |
2826 | cpuc->shared_regs = pc; |
2827 | break; | |
2828 | } | |
a7e3ed1e | 2829 | } |
b36817e8 SE |
2830 | cpuc->shared_regs->core_id = core_id; |
2831 | cpuc->shared_regs->refcnt++; | |
a7e3ed1e AK |
2832 | } |
2833 | ||
b36817e8 SE |
2834 | if (x86_pmu.lbr_sel_map) |
2835 | cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR]; | |
6f6539ca MD |
2836 | |
2837 | if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) { | |
06931e62 | 2838 | for_each_cpu(i, topology_sibling_cpumask(cpu)) { |
6f6539ca MD |
2839 | struct intel_excl_cntrs *c; |
2840 | ||
2841 | c = per_cpu(cpu_hw_events, i).excl_cntrs; | |
2842 | if (c && c->core_id == core_id) { | |
2843 | cpuc->kfree_on_online[1] = cpuc->excl_cntrs; | |
2844 | cpuc->excl_cntrs = c; | |
2845 | cpuc->excl_thread_id = 1; | |
2846 | break; | |
2847 | } | |
2848 | } | |
2849 | cpuc->excl_cntrs->core_id = core_id; | |
2850 | cpuc->excl_cntrs->refcnt++; | |
2851 | } | |
74846d35 PZ |
2852 | } |
2853 | ||
b37609c3 | 2854 | static void free_excl_cntrs(int cpu) |
74846d35 | 2855 | { |
a7e3ed1e | 2856 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); |
6f6539ca | 2857 | struct intel_excl_cntrs *c; |
a7e3ed1e | 2858 | |
6f6539ca MD |
2859 | c = cpuc->excl_cntrs; |
2860 | if (c) { | |
2861 | if (c->core_id == -1 || --c->refcnt == 0) | |
2862 | kfree(c); | |
2863 | cpuc->excl_cntrs = NULL; | |
2864 | kfree(cpuc->constraint_list); | |
2865 | cpuc->constraint_list = NULL; | |
2866 | } | |
b37609c3 | 2867 | } |
a7e3ed1e | 2868 | |
b37609c3 SE |
2869 | static void intel_pmu_cpu_dying(int cpu) |
2870 | { | |
2871 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); | |
2872 | struct intel_shared_regs *pc; | |
2873 | ||
2874 | pc = cpuc->shared_regs; | |
2875 | if (pc) { | |
2876 | if (pc->core_id == -1 || --pc->refcnt == 0) | |
2877 | kfree(pc); | |
2878 | cpuc->shared_regs = NULL; | |
e979121b MD |
2879 | } |
2880 | ||
b37609c3 SE |
2881 | free_excl_cntrs(cpu); |
2882 | ||
74846d35 PZ |
2883 | fini_debug_store_on_cpu(cpu); |
2884 | } | |
2885 | ||
9c964efa YZ |
2886 | static void intel_pmu_sched_task(struct perf_event_context *ctx, |
2887 | bool sched_in) | |
2888 | { | |
2889 | if (x86_pmu.pebs_active) | |
2890 | intel_pmu_pebs_sched_task(ctx, sched_in); | |
2891 | if (x86_pmu.lbr_nr) | |
2892 | intel_pmu_lbr_sched_task(ctx, sched_in); | |
2893 | } | |
2894 | ||
641cc938 JO |
2895 | PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63"); |
2896 | ||
a63fcab4 SE |
2897 | PMU_FORMAT_ATTR(ldlat, "config1:0-15"); |
2898 | ||
d0dc8494 AK |
2899 | PMU_FORMAT_ATTR(frontend, "config1:0-23"); |
2900 | ||
641cc938 JO |
2901 | static struct attribute *intel_arch3_formats_attr[] = { |
2902 | &format_attr_event.attr, | |
2903 | &format_attr_umask.attr, | |
2904 | &format_attr_edge.attr, | |
2905 | &format_attr_pc.attr, | |
2906 | &format_attr_any.attr, | |
2907 | &format_attr_inv.attr, | |
2908 | &format_attr_cmask.attr, | |
3a632cb2 AK |
2909 | &format_attr_in_tx.attr, |
2910 | &format_attr_in_tx_cp.attr, | |
641cc938 JO |
2911 | |
2912 | &format_attr_offcore_rsp.attr, /* XXX do NHM/WSM + SNB breakout */ | |
a63fcab4 | 2913 | &format_attr_ldlat.attr, /* PEBS load latency */ |
641cc938 JO |
2914 | NULL, |
2915 | }; | |
2916 | ||
d0dc8494 AK |
2917 | static struct attribute *skl_format_attr[] = { |
2918 | &format_attr_frontend.attr, | |
2919 | NULL, | |
2920 | }; | |
2921 | ||
3b6e0421 JO |
2922 | static __initconst const struct x86_pmu core_pmu = { |
2923 | .name = "core", | |
2924 | .handle_irq = x86_pmu_handle_irq, | |
2925 | .disable_all = x86_pmu_disable_all, | |
2926 | .enable_all = core_pmu_enable_all, | |
2927 | .enable = core_pmu_enable_event, | |
2928 | .disable = x86_pmu_disable_event, | |
2929 | .hw_config = x86_pmu_hw_config, | |
2930 | .schedule_events = x86_schedule_events, | |
2931 | .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, | |
2932 | .perfctr = MSR_ARCH_PERFMON_PERFCTR0, | |
2933 | .event_map = intel_pmu_event_map, | |
2934 | .max_events = ARRAY_SIZE(intel_perfmon_event_map), | |
2935 | .apic = 1, | |
a7b58d21 AK |
2936 | .free_running_flags = PEBS_FREERUNNING_FLAGS, |
2937 | ||
3b6e0421 JO |
2938 | /* |
2939 | * Intel PMCs cannot be accessed sanely above 32-bit width, | |
2940 | * so we install an artificial 1<<31 period regardless of | |
2941 | * the generic event period: | |
2942 | */ | |
2943 | .max_period = (1ULL<<31) - 1, | |
2944 | .get_event_constraints = intel_get_event_constraints, | |
2945 | .put_event_constraints = intel_put_event_constraints, | |
2946 | .event_constraints = intel_core_event_constraints, | |
2947 | .guest_get_msrs = core_guest_get_msrs, | |
2948 | .format_attrs = intel_arch_formats_attr, | |
2949 | .events_sysfs_show = intel_event_sysfs_show, | |
2950 | ||
2951 | /* | |
2952 | * Virtual (or funny metal) CPU can define x86_pmu.extra_regs | |
2953 | * together with PMU version 1 and thus be using core_pmu with | |
2954 | * shared_regs. We need following callbacks here to allocate | |
2955 | * it properly. | |
2956 | */ | |
2957 | .cpu_prepare = intel_pmu_cpu_prepare, | |
2958 | .cpu_starting = intel_pmu_cpu_starting, | |
2959 | .cpu_dying = intel_pmu_cpu_dying, | |
2960 | }; | |
2961 | ||
caaa8be3 | 2962 | static __initconst const struct x86_pmu intel_pmu = { |
f22f54f4 PZ |
2963 | .name = "Intel", |
2964 | .handle_irq = intel_pmu_handle_irq, | |
2965 | .disable_all = intel_pmu_disable_all, | |
2966 | .enable_all = intel_pmu_enable_all, | |
2967 | .enable = intel_pmu_enable_event, | |
2968 | .disable = intel_pmu_disable_event, | |
b4cdc5c2 | 2969 | .hw_config = intel_pmu_hw_config, |
a072738e | 2970 | .schedule_events = x86_schedule_events, |
f22f54f4 PZ |
2971 | .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, |
2972 | .perfctr = MSR_ARCH_PERFMON_PERFCTR0, | |
2973 | .event_map = intel_pmu_event_map, | |
f22f54f4 PZ |
2974 | .max_events = ARRAY_SIZE(intel_perfmon_event_map), |
2975 | .apic = 1, | |
a7b58d21 | 2976 | .free_running_flags = PEBS_FREERUNNING_FLAGS, |
f22f54f4 PZ |
2977 | /* |
2978 | * Intel PMCs cannot be accessed sanely above 32 bit width, | |
2979 | * so we install an artificial 1<<31 period regardless of | |
2980 | * the generic event period: | |
2981 | */ | |
2982 | .max_period = (1ULL << 31) - 1, | |
3f6da390 | 2983 | .get_event_constraints = intel_get_event_constraints, |
a7e3ed1e | 2984 | .put_event_constraints = intel_put_event_constraints, |
0780c927 | 2985 | .pebs_aliases = intel_pebs_aliases_core2, |
3f6da390 | 2986 | |
641cc938 | 2987 | .format_attrs = intel_arch3_formats_attr, |
0bf79d44 | 2988 | .events_sysfs_show = intel_event_sysfs_show, |
641cc938 | 2989 | |
a7e3ed1e | 2990 | .cpu_prepare = intel_pmu_cpu_prepare, |
74846d35 PZ |
2991 | .cpu_starting = intel_pmu_cpu_starting, |
2992 | .cpu_dying = intel_pmu_cpu_dying, | |
144d31e6 | 2993 | .guest_get_msrs = intel_guest_get_msrs, |
9c964efa | 2994 | .sched_task = intel_pmu_sched_task, |
f22f54f4 PZ |
2995 | }; |
2996 | ||
c1d6f42f | 2997 | static __init void intel_clovertown_quirk(void) |
3c44780b PZ |
2998 | { |
2999 | /* | |
3000 | * PEBS is unreliable due to: | |
3001 | * | |
3002 | * AJ67 - PEBS may experience CPL leaks | |
3003 | * AJ68 - PEBS PMI may be delayed by one event | |
3004 | * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12] | |
3005 | * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS | |
3006 | * | |
3007 | * AJ67 could be worked around by restricting the OS/USR flags. | |
3008 | * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI. | |
3009 | * | |
3010 | * AJ106 could possibly be worked around by not allowing LBR | |
3011 | * usage from PEBS, including the fixup. | |
3012 | * AJ68 could possibly be worked around by always programming | |
ec75a716 | 3013 | * a pebs_event_reset[0] value and coping with the lost events. |
3c44780b PZ |
3014 | * |
3015 | * But taken together it might just make sense to not enable PEBS on | |
3016 | * these chips. | |
3017 | */ | |
c767a54b | 3018 | pr_warn("PEBS disabled due to CPU errata\n"); |
3c44780b PZ |
3019 | x86_pmu.pebs = 0; |
3020 | x86_pmu.pebs_constraints = NULL; | |
3021 | } | |
3022 | ||
c93dc84c PZ |
3023 | static int intel_snb_pebs_broken(int cpu) |
3024 | { | |
3025 | u32 rev = UINT_MAX; /* default to broken for unknown models */ | |
3026 | ||
3027 | switch (cpu_data(cpu).x86_model) { | |
3028 | case 42: /* SNB */ | |
3029 | rev = 0x28; | |
3030 | break; | |
3031 | ||
3032 | case 45: /* SNB-EP */ | |
3033 | switch (cpu_data(cpu).x86_mask) { | |
3034 | case 6: rev = 0x618; break; | |
3035 | case 7: rev = 0x70c; break; | |
3036 | } | |
3037 | } | |
3038 | ||
3039 | return (cpu_data(cpu).microcode < rev); | |
3040 | } | |
3041 | ||
3042 | static void intel_snb_check_microcode(void) | |
3043 | { | |
3044 | int pebs_broken = 0; | |
3045 | int cpu; | |
3046 | ||
3047 | get_online_cpus(); | |
3048 | for_each_online_cpu(cpu) { | |
3049 | if ((pebs_broken = intel_snb_pebs_broken(cpu))) | |
3050 | break; | |
3051 | } | |
3052 | put_online_cpus(); | |
3053 | ||
3054 | if (pebs_broken == x86_pmu.pebs_broken) | |
3055 | return; | |
3056 | ||
3057 | /* | |
3058 | * Serialized by the microcode lock.. | |
3059 | */ | |
3060 | if (x86_pmu.pebs_broken) { | |
3061 | pr_info("PEBS enabled due to microcode update\n"); | |
3062 | x86_pmu.pebs_broken = 0; | |
3063 | } else { | |
3064 | pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n"); | |
3065 | x86_pmu.pebs_broken = 1; | |
3066 | } | |
3067 | } | |
3068 | ||
338b522c KL |
3069 | /* |
3070 | * Under certain circumstances, access certain MSR may cause #GP. | |
3071 | * The function tests if the input MSR can be safely accessed. | |
3072 | */ | |
3073 | static bool check_msr(unsigned long msr, u64 mask) | |
3074 | { | |
3075 | u64 val_old, val_new, val_tmp; | |
3076 | ||
3077 | /* | |
3078 | * Read the current value, change it and read it back to see if it | |
3079 | * matches, this is needed to detect certain hardware emulators | |
3080 | * (qemu/kvm) that don't trap on the MSR access and always return 0s. | |
3081 | */ | |
3082 | if (rdmsrl_safe(msr, &val_old)) | |
3083 | return false; | |
3084 | ||
3085 | /* | |
3086 | * Only change the bits which can be updated by wrmsrl. | |
3087 | */ | |
3088 | val_tmp = val_old ^ mask; | |
3089 | if (wrmsrl_safe(msr, val_tmp) || | |
3090 | rdmsrl_safe(msr, &val_new)) | |
3091 | return false; | |
3092 | ||
3093 | if (val_new != val_tmp) | |
3094 | return false; | |
3095 | ||
3096 | /* Here it's sure that the MSR can be safely accessed. | |
3097 | * Restore the old value and return. | |
3098 | */ | |
3099 | wrmsrl(msr, val_old); | |
3100 | ||
3101 | return true; | |
3102 | } | |
3103 | ||
c1d6f42f | 3104 | static __init void intel_sandybridge_quirk(void) |
6a600a8b | 3105 | { |
c93dc84c PZ |
3106 | x86_pmu.check_microcode = intel_snb_check_microcode; |
3107 | intel_snb_check_microcode(); | |
6a600a8b PZ |
3108 | } |
3109 | ||
c1d6f42f PZ |
3110 | static const struct { int id; char *name; } intel_arch_events_map[] __initconst = { |
3111 | { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" }, | |
3112 | { PERF_COUNT_HW_INSTRUCTIONS, "instructions" }, | |
3113 | { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" }, | |
3114 | { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" }, | |
3115 | { PERF_COUNT_HW_CACHE_MISSES, "cache misses" }, | |
3116 | { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" }, | |
3117 | { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" }, | |
ffb871bc GN |
3118 | }; |
3119 | ||
c1d6f42f PZ |
3120 | static __init void intel_arch_events_quirk(void) |
3121 | { | |
3122 | int bit; | |
3123 | ||
3124 | /* disable event that reported as not presend by cpuid */ | |
3125 | for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) { | |
3126 | intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0; | |
c767a54b JP |
3127 | pr_warn("CPUID marked event: \'%s\' unavailable\n", |
3128 | intel_arch_events_map[bit].name); | |
c1d6f42f PZ |
3129 | } |
3130 | } | |
3131 | ||
3132 | static __init void intel_nehalem_quirk(void) | |
3133 | { | |
3134 | union cpuid10_ebx ebx; | |
3135 | ||
3136 | ebx.full = x86_pmu.events_maskl; | |
3137 | if (ebx.split.no_branch_misses_retired) { | |
3138 | /* | |
3139 | * Erratum AAJ80 detected, we work it around by using | |
3140 | * the BR_MISP_EXEC.ANY event. This will over-count | |
3141 | * branch-misses, but it's still much better than the | |
3142 | * architectural event which is often completely bogus: | |
3143 | */ | |
3144 | intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89; | |
3145 | ebx.split.no_branch_misses_retired = 0; | |
3146 | x86_pmu.events_maskl = ebx.full; | |
c767a54b | 3147 | pr_info("CPU erratum AAJ80 worked around\n"); |
c1d6f42f PZ |
3148 | } |
3149 | } | |
3150 | ||
93fcf72c MD |
3151 | /* |
3152 | * enable software workaround for errata: | |
3153 | * SNB: BJ122 | |
3154 | * IVB: BV98 | |
3155 | * HSW: HSD29 | |
3156 | * | |
3157 | * Only needed when HT is enabled. However detecting | |
b37609c3 SE |
3158 | * if HT is enabled is difficult (model specific). So instead, |
3159 | * we enable the workaround in the early boot, and verify if | |
3160 | * it is needed in a later initcall phase once we have valid | |
3161 | * topology information to check if HT is actually enabled | |
93fcf72c MD |
3162 | */ |
3163 | static __init void intel_ht_bug(void) | |
3164 | { | |
b37609c3 | 3165 | x86_pmu.flags |= PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED; |
93fcf72c | 3166 | |
93fcf72c | 3167 | x86_pmu.start_scheduling = intel_start_scheduling; |
0c41e756 | 3168 | x86_pmu.commit_scheduling = intel_commit_scheduling; |
93fcf72c | 3169 | x86_pmu.stop_scheduling = intel_stop_scheduling; |
93fcf72c MD |
3170 | } |
3171 | ||
7f2ee91f IM |
3172 | EVENT_ATTR_STR(mem-loads, mem_ld_hsw, "event=0xcd,umask=0x1,ldlat=3"); |
3173 | EVENT_ATTR_STR(mem-stores, mem_st_hsw, "event=0xd0,umask=0x82") | |
f9134f36 | 3174 | |
4b2c4f1f | 3175 | /* Haswell special events */ |
7f2ee91f IM |
3176 | EVENT_ATTR_STR(tx-start, tx_start, "event=0xc9,umask=0x1"); |
3177 | EVENT_ATTR_STR(tx-commit, tx_commit, "event=0xc9,umask=0x2"); | |
3178 | EVENT_ATTR_STR(tx-abort, tx_abort, "event=0xc9,umask=0x4"); | |
3179 | EVENT_ATTR_STR(tx-capacity, tx_capacity, "event=0x54,umask=0x2"); | |
3180 | EVENT_ATTR_STR(tx-conflict, tx_conflict, "event=0x54,umask=0x1"); | |
3181 | EVENT_ATTR_STR(el-start, el_start, "event=0xc8,umask=0x1"); | |
3182 | EVENT_ATTR_STR(el-commit, el_commit, "event=0xc8,umask=0x2"); | |
3183 | EVENT_ATTR_STR(el-abort, el_abort, "event=0xc8,umask=0x4"); | |
3184 | EVENT_ATTR_STR(el-capacity, el_capacity, "event=0x54,umask=0x2"); | |
3185 | EVENT_ATTR_STR(el-conflict, el_conflict, "event=0x54,umask=0x1"); | |
3186 | EVENT_ATTR_STR(cycles-t, cycles_t, "event=0x3c,in_tx=1"); | |
3187 | EVENT_ATTR_STR(cycles-ct, cycles_ct, "event=0x3c,in_tx=1,in_tx_cp=1"); | |
4b2c4f1f | 3188 | |
f9134f36 | 3189 | static struct attribute *hsw_events_attrs[] = { |
4b2c4f1f AK |
3190 | EVENT_PTR(tx_start), |
3191 | EVENT_PTR(tx_commit), | |
3192 | EVENT_PTR(tx_abort), | |
3193 | EVENT_PTR(tx_capacity), | |
3194 | EVENT_PTR(tx_conflict), | |
3195 | EVENT_PTR(el_start), | |
3196 | EVENT_PTR(el_commit), | |
3197 | EVENT_PTR(el_abort), | |
3198 | EVENT_PTR(el_capacity), | |
3199 | EVENT_PTR(el_conflict), | |
3200 | EVENT_PTR(cycles_t), | |
3201 | EVENT_PTR(cycles_ct), | |
f9134f36 AK |
3202 | EVENT_PTR(mem_ld_hsw), |
3203 | EVENT_PTR(mem_st_hsw), | |
3204 | NULL | |
3205 | }; | |
3206 | ||
de0428a7 | 3207 | __init int intel_pmu_init(void) |
f22f54f4 PZ |
3208 | { |
3209 | union cpuid10_edx edx; | |
3210 | union cpuid10_eax eax; | |
ffb871bc | 3211 | union cpuid10_ebx ebx; |
a1eac7ac | 3212 | struct event_constraint *c; |
f22f54f4 | 3213 | unsigned int unused; |
338b522c KL |
3214 | struct extra_reg *er; |
3215 | int version, i; | |
f22f54f4 PZ |
3216 | |
3217 | if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) { | |
a072738e CG |
3218 | switch (boot_cpu_data.x86) { |
3219 | case 0x6: | |
3220 | return p6_pmu_init(); | |
e717bf4e VW |
3221 | case 0xb: |
3222 | return knc_pmu_init(); | |
a072738e CG |
3223 | case 0xf: |
3224 | return p4_pmu_init(); | |
3225 | } | |
f22f54f4 | 3226 | return -ENODEV; |
f22f54f4 PZ |
3227 | } |
3228 | ||
3229 | /* | |
3230 | * Check whether the Architectural PerfMon supports | |
3231 | * Branch Misses Retired hw_event or not. | |
3232 | */ | |
ffb871bc GN |
3233 | cpuid(10, &eax.full, &ebx.full, &unused, &edx.full); |
3234 | if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT) | |
f22f54f4 PZ |
3235 | return -ENODEV; |
3236 | ||
3237 | version = eax.split.version_id; | |
3238 | if (version < 2) | |
3239 | x86_pmu = core_pmu; | |
3240 | else | |
3241 | x86_pmu = intel_pmu; | |
3242 | ||
3243 | x86_pmu.version = version; | |
948b1bb8 RR |
3244 | x86_pmu.num_counters = eax.split.num_counters; |
3245 | x86_pmu.cntval_bits = eax.split.bit_width; | |
3246 | x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1; | |
f22f54f4 | 3247 | |
c1d6f42f PZ |
3248 | x86_pmu.events_maskl = ebx.full; |
3249 | x86_pmu.events_mask_len = eax.split.mask_length; | |
3250 | ||
70ab7003 AK |
3251 | x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters); |
3252 | ||
f22f54f4 PZ |
3253 | /* |
3254 | * Quirk: v2 perfmon does not report fixed-purpose events, so | |
3255 | * assume at least 3 events: | |
3256 | */ | |
3257 | if (version > 1) | |
948b1bb8 | 3258 | x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3); |
f22f54f4 | 3259 | |
c9b08884 | 3260 | if (boot_cpu_has(X86_FEATURE_PDCM)) { |
8db909a7 PZ |
3261 | u64 capabilities; |
3262 | ||
3263 | rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities); | |
3264 | x86_pmu.intel_cap.capabilities = capabilities; | |
3265 | } | |
3266 | ||
ca037701 PZ |
3267 | intel_ds_init(); |
3268 | ||
c1d6f42f PZ |
3269 | x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */ |
3270 | ||
f22f54f4 PZ |
3271 | /* |
3272 | * Install the hw-cache-events table: | |
3273 | */ | |
3274 | switch (boot_cpu_data.x86_model) { | |
0f7c29ce | 3275 | case 14: /* 65nm Core "Yonah" */ |
f22f54f4 PZ |
3276 | pr_cont("Core events, "); |
3277 | break; | |
3278 | ||
0f7c29ce | 3279 | case 15: /* 65nm Core2 "Merom" */ |
c1d6f42f | 3280 | x86_add_quirk(intel_clovertown_quirk); |
0f7c29ce PZ |
3281 | case 22: /* 65nm Core2 "Merom-L" */ |
3282 | case 23: /* 45nm Core2 "Penryn" */ | |
3283 | case 29: /* 45nm Core2 "Dunnington (MP) */ | |
f22f54f4 PZ |
3284 | memcpy(hw_cache_event_ids, core2_hw_cache_event_ids, |
3285 | sizeof(hw_cache_event_ids)); | |
3286 | ||
caff2bef PZ |
3287 | intel_pmu_lbr_init_core(); |
3288 | ||
f22f54f4 | 3289 | x86_pmu.event_constraints = intel_core2_event_constraints; |
17e31629 | 3290 | x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints; |
f22f54f4 PZ |
3291 | pr_cont("Core2 events, "); |
3292 | break; | |
3293 | ||
0f7c29ce PZ |
3294 | case 30: /* 45nm Nehalem */ |
3295 | case 26: /* 45nm Nehalem-EP */ | |
3296 | case 46: /* 45nm Nehalem-EX */ | |
f22f54f4 PZ |
3297 | memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids, |
3298 | sizeof(hw_cache_event_ids)); | |
e994d7d2 AK |
3299 | memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, |
3300 | sizeof(hw_cache_extra_regs)); | |
f22f54f4 | 3301 | |
caff2bef PZ |
3302 | intel_pmu_lbr_init_nhm(); |
3303 | ||
f22f54f4 | 3304 | x86_pmu.event_constraints = intel_nehalem_event_constraints; |
17e31629 | 3305 | x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints; |
11164cd4 | 3306 | x86_pmu.enable_all = intel_pmu_nhm_enable_all; |
a7e3ed1e | 3307 | x86_pmu.extra_regs = intel_nehalem_extra_regs; |
ec75a716 | 3308 | |
f20093ee SE |
3309 | x86_pmu.cpu_events = nhm_events_attrs; |
3310 | ||
91fc4cc0 | 3311 | /* UOPS_ISSUED.STALLED_CYCLES */ |
f9b4eeb8 PZ |
3312 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = |
3313 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | |
91fc4cc0 | 3314 | /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ |
f9b4eeb8 PZ |
3315 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = |
3316 | X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); | |
94403f88 | 3317 | |
c1d6f42f | 3318 | x86_add_quirk(intel_nehalem_quirk); |
ec75a716 | 3319 | |
11164cd4 | 3320 | pr_cont("Nehalem events, "); |
f22f54f4 | 3321 | break; |
caff2bef | 3322 | |
0f7c29ce PZ |
3323 | case 28: /* 45nm Atom "Pineview" */ |
3324 | case 38: /* 45nm Atom "Lincroft" */ | |
3325 | case 39: /* 32nm Atom "Penwell" */ | |
3326 | case 53: /* 32nm Atom "Cloverview" */ | |
3327 | case 54: /* 32nm Atom "Cedarview" */ | |
f22f54f4 PZ |
3328 | memcpy(hw_cache_event_ids, atom_hw_cache_event_ids, |
3329 | sizeof(hw_cache_event_ids)); | |
3330 | ||
caff2bef PZ |
3331 | intel_pmu_lbr_init_atom(); |
3332 | ||
f22f54f4 | 3333 | x86_pmu.event_constraints = intel_gen_event_constraints; |
17e31629 | 3334 | x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints; |
f22f54f4 PZ |
3335 | pr_cont("Atom events, "); |
3336 | break; | |
3337 | ||
0f7c29ce | 3338 | case 55: /* 22nm Atom "Silvermont" */ |
ef454cae | 3339 | case 76: /* 14nm Atom "Airmont" */ |
0f7c29ce | 3340 | case 77: /* 22nm Atom "Silvermont Avoton/Rangely" */ |
1fa64180 YZ |
3341 | memcpy(hw_cache_event_ids, slm_hw_cache_event_ids, |
3342 | sizeof(hw_cache_event_ids)); | |
3343 | memcpy(hw_cache_extra_regs, slm_hw_cache_extra_regs, | |
3344 | sizeof(hw_cache_extra_regs)); | |
3345 | ||
3346 | intel_pmu_lbr_init_atom(); | |
3347 | ||
3348 | x86_pmu.event_constraints = intel_slm_event_constraints; | |
3349 | x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints; | |
3350 | x86_pmu.extra_regs = intel_slm_extra_regs; | |
9a5e3fb5 | 3351 | x86_pmu.flags |= PMU_FL_HAS_RSP_1; |
1fa64180 YZ |
3352 | pr_cont("Silvermont events, "); |
3353 | break; | |
3354 | ||
0f7c29ce PZ |
3355 | case 37: /* 32nm Westmere */ |
3356 | case 44: /* 32nm Westmere-EP */ | |
3357 | case 47: /* 32nm Westmere-EX */ | |
f22f54f4 PZ |
3358 | memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids, |
3359 | sizeof(hw_cache_event_ids)); | |
e994d7d2 AK |
3360 | memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, |
3361 | sizeof(hw_cache_extra_regs)); | |
f22f54f4 | 3362 | |
caff2bef PZ |
3363 | intel_pmu_lbr_init_nhm(); |
3364 | ||
f22f54f4 | 3365 | x86_pmu.event_constraints = intel_westmere_event_constraints; |
40b91cd1 | 3366 | x86_pmu.enable_all = intel_pmu_nhm_enable_all; |
17e31629 | 3367 | x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints; |
a7e3ed1e | 3368 | x86_pmu.extra_regs = intel_westmere_extra_regs; |
9a5e3fb5 | 3369 | x86_pmu.flags |= PMU_FL_HAS_RSP_1; |
30112039 | 3370 | |
f20093ee SE |
3371 | x86_pmu.cpu_events = nhm_events_attrs; |
3372 | ||
30112039 | 3373 | /* UOPS_ISSUED.STALLED_CYCLES */ |
f9b4eeb8 PZ |
3374 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = |
3375 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | |
30112039 | 3376 | /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ |
f9b4eeb8 PZ |
3377 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = |
3378 | X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); | |
30112039 | 3379 | |
f22f54f4 PZ |
3380 | pr_cont("Westmere events, "); |
3381 | break; | |
b622d644 | 3382 | |
0f7c29ce PZ |
3383 | case 42: /* 32nm SandyBridge */ |
3384 | case 45: /* 32nm SandyBridge-E/EN/EP */ | |
47a8863d | 3385 | x86_add_quirk(intel_sandybridge_quirk); |
93fcf72c | 3386 | x86_add_quirk(intel_ht_bug); |
b06b3d49 LM |
3387 | memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, |
3388 | sizeof(hw_cache_event_ids)); | |
74e6543f YZ |
3389 | memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, |
3390 | sizeof(hw_cache_extra_regs)); | |
b06b3d49 | 3391 | |
c5cc2cd9 | 3392 | intel_pmu_lbr_init_snb(); |
b06b3d49 LM |
3393 | |
3394 | x86_pmu.event_constraints = intel_snb_event_constraints; | |
de0428a7 | 3395 | x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints; |
0780c927 | 3396 | x86_pmu.pebs_aliases = intel_pebs_aliases_snb; |
f1923820 SE |
3397 | if (boot_cpu_data.x86_model == 45) |
3398 | x86_pmu.extra_regs = intel_snbep_extra_regs; | |
3399 | else | |
3400 | x86_pmu.extra_regs = intel_snb_extra_regs; | |
93fcf72c MD |
3401 | |
3402 | ||
ee89cbc2 | 3403 | /* all extra regs are per-cpu when HT is on */ |
9a5e3fb5 SE |
3404 | x86_pmu.flags |= PMU_FL_HAS_RSP_1; |
3405 | x86_pmu.flags |= PMU_FL_NO_HT_SHARING; | |
e04d1b23 | 3406 | |
f20093ee SE |
3407 | x86_pmu.cpu_events = snb_events_attrs; |
3408 | ||
e04d1b23 | 3409 | /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ |
f9b4eeb8 PZ |
3410 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = |
3411 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | |
e04d1b23 | 3412 | /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/ |
f9b4eeb8 PZ |
3413 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = |
3414 | X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1); | |
e04d1b23 | 3415 | |
b06b3d49 LM |
3416 | pr_cont("SandyBridge events, "); |
3417 | break; | |
0f7c29ce PZ |
3418 | |
3419 | case 58: /* 22nm IvyBridge */ | |
3420 | case 62: /* 22nm IvyBridge-EP/EX */ | |
93fcf72c | 3421 | x86_add_quirk(intel_ht_bug); |
20a36e39 SE |
3422 | memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, |
3423 | sizeof(hw_cache_event_ids)); | |
1996388e VW |
3424 | /* dTLB-load-misses on IVB is different than SNB */ |
3425 | hw_cache_event_ids[C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = 0x8108; /* DTLB_LOAD_MISSES.DEMAND_LD_MISS_CAUSES_A_WALK */ | |
3426 | ||
20a36e39 SE |
3427 | memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, |
3428 | sizeof(hw_cache_extra_regs)); | |
3429 | ||
3430 | intel_pmu_lbr_init_snb(); | |
3431 | ||
69943182 | 3432 | x86_pmu.event_constraints = intel_ivb_event_constraints; |
20a36e39 SE |
3433 | x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints; |
3434 | x86_pmu.pebs_aliases = intel_pebs_aliases_snb; | |
f1923820 SE |
3435 | if (boot_cpu_data.x86_model == 62) |
3436 | x86_pmu.extra_regs = intel_snbep_extra_regs; | |
3437 | else | |
3438 | x86_pmu.extra_regs = intel_snb_extra_regs; | |
20a36e39 | 3439 | /* all extra regs are per-cpu when HT is on */ |
9a5e3fb5 SE |
3440 | x86_pmu.flags |= PMU_FL_HAS_RSP_1; |
3441 | x86_pmu.flags |= PMU_FL_NO_HT_SHARING; | |
20a36e39 | 3442 | |
f20093ee SE |
3443 | x86_pmu.cpu_events = snb_events_attrs; |
3444 | ||
20a36e39 SE |
3445 | /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ |
3446 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = | |
3447 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | |
3448 | ||
3449 | pr_cont("IvyBridge events, "); | |
3450 | break; | |
3451 | ||
b06b3d49 | 3452 | |
d86c8eaf AK |
3453 | case 60: /* 22nm Haswell Core */ |
3454 | case 63: /* 22nm Haswell Server */ | |
3455 | case 69: /* 22nm Haswell ULT */ | |
3456 | case 70: /* 22nm Haswell + GT3e (Intel Iris Pro graphics) */ | |
93fcf72c | 3457 | x86_add_quirk(intel_ht_bug); |
72db5596 | 3458 | x86_pmu.late_ack = true; |
0f1b5ca2 AK |
3459 | memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids)); |
3460 | memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); | |
3a632cb2 | 3461 | |
e9d7f7cd | 3462 | intel_pmu_lbr_init_hsw(); |
3a632cb2 AK |
3463 | |
3464 | x86_pmu.event_constraints = intel_hsw_event_constraints; | |
3044318f | 3465 | x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints; |
36bbb2f2 | 3466 | x86_pmu.extra_regs = intel_snbep_extra_regs; |
3044318f | 3467 | x86_pmu.pebs_aliases = intel_pebs_aliases_snb; |
3a632cb2 | 3468 | /* all extra regs are per-cpu when HT is on */ |
9a5e3fb5 SE |
3469 | x86_pmu.flags |= PMU_FL_HAS_RSP_1; |
3470 | x86_pmu.flags |= PMU_FL_NO_HT_SHARING; | |
3a632cb2 AK |
3471 | |
3472 | x86_pmu.hw_config = hsw_hw_config; | |
3473 | x86_pmu.get_event_constraints = hsw_get_event_constraints; | |
f9134f36 | 3474 | x86_pmu.cpu_events = hsw_events_attrs; |
b7af41a1 | 3475 | x86_pmu.lbr_double_abort = true; |
3a632cb2 AK |
3476 | pr_cont("Haswell events, "); |
3477 | break; | |
3478 | ||
91f1b705 AK |
3479 | case 61: /* 14nm Broadwell Core-M */ |
3480 | case 86: /* 14nm Broadwell Xeon D */ | |
4b36f1a4 AK |
3481 | case 71: /* 14nm Broadwell + GT3e (Intel Iris Pro graphics) */ |
3482 | case 79: /* 14nm Broadwell Server */ | |
91f1b705 AK |
3483 | x86_pmu.late_ack = true; |
3484 | memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids)); | |
3485 | memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); | |
3486 | ||
3487 | /* L3_MISS_LOCAL_DRAM is BIT(26) in Broadwell */ | |
3488 | hw_cache_extra_regs[C(LL)][C(OP_READ)][C(RESULT_MISS)] = HSW_DEMAND_READ | | |
3489 | BDW_L3_MISS|HSW_SNOOP_DRAM; | |
3490 | hw_cache_extra_regs[C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = HSW_DEMAND_WRITE|BDW_L3_MISS| | |
3491 | HSW_SNOOP_DRAM; | |
3492 | hw_cache_extra_regs[C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = HSW_DEMAND_READ| | |
3493 | BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM; | |
3494 | hw_cache_extra_regs[C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = HSW_DEMAND_WRITE| | |
3495 | BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM; | |
3496 | ||
78d504bc | 3497 | intel_pmu_lbr_init_hsw(); |
91f1b705 AK |
3498 | |
3499 | x86_pmu.event_constraints = intel_bdw_event_constraints; | |
3500 | x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints; | |
3501 | x86_pmu.extra_regs = intel_snbep_extra_regs; | |
3502 | x86_pmu.pebs_aliases = intel_pebs_aliases_snb; | |
3503 | /* all extra regs are per-cpu when HT is on */ | |
9a5e3fb5 SE |
3504 | x86_pmu.flags |= PMU_FL_HAS_RSP_1; |
3505 | x86_pmu.flags |= PMU_FL_NO_HT_SHARING; | |
91f1b705 AK |
3506 | |
3507 | x86_pmu.hw_config = hsw_hw_config; | |
3508 | x86_pmu.get_event_constraints = hsw_get_event_constraints; | |
3509 | x86_pmu.cpu_events = hsw_events_attrs; | |
294fe0f5 | 3510 | x86_pmu.limit_period = bdw_limit_period; |
91f1b705 AK |
3511 | pr_cont("Broadwell events, "); |
3512 | break; | |
3513 | ||
9a92e16f AK |
3514 | case 78: /* 14nm Skylake Mobile */ |
3515 | case 94: /* 14nm Skylake Desktop */ | |
3516 | x86_pmu.late_ack = true; | |
3517 | memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids)); | |
3518 | memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); | |
3519 | intel_pmu_lbr_init_skl(); | |
3520 | ||
3521 | x86_pmu.event_constraints = intel_skl_event_constraints; | |
3522 | x86_pmu.pebs_constraints = intel_skl_pebs_event_constraints; | |
3523 | x86_pmu.extra_regs = intel_skl_extra_regs; | |
3524 | x86_pmu.pebs_aliases = intel_pebs_aliases_snb; | |
3525 | /* all extra regs are per-cpu when HT is on */ | |
3526 | x86_pmu.flags |= PMU_FL_HAS_RSP_1; | |
3527 | x86_pmu.flags |= PMU_FL_NO_HT_SHARING; | |
3528 | ||
3529 | x86_pmu.hw_config = hsw_hw_config; | |
3530 | x86_pmu.get_event_constraints = hsw_get_event_constraints; | |
d0dc8494 AK |
3531 | x86_pmu.format_attrs = merge_attr(intel_arch3_formats_attr, |
3532 | skl_format_attr); | |
9a92e16f AK |
3533 | WARN_ON(!x86_pmu.format_attrs); |
3534 | x86_pmu.cpu_events = hsw_events_attrs; | |
3535 | pr_cont("Skylake events, "); | |
3536 | break; | |
3537 | ||
f22f54f4 | 3538 | default: |
0af3ac1f AK |
3539 | switch (x86_pmu.version) { |
3540 | case 1: | |
3541 | x86_pmu.event_constraints = intel_v1_event_constraints; | |
3542 | pr_cont("generic architected perfmon v1, "); | |
3543 | break; | |
3544 | default: | |
3545 | /* | |
3546 | * default constraints for v2 and up | |
3547 | */ | |
3548 | x86_pmu.event_constraints = intel_gen_event_constraints; | |
3549 | pr_cont("generic architected perfmon, "); | |
3550 | break; | |
3551 | } | |
f22f54f4 | 3552 | } |
ffb871bc | 3553 | |
a1eac7ac RR |
3554 | if (x86_pmu.num_counters > INTEL_PMC_MAX_GENERIC) { |
3555 | WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!", | |
3556 | x86_pmu.num_counters, INTEL_PMC_MAX_GENERIC); | |
3557 | x86_pmu.num_counters = INTEL_PMC_MAX_GENERIC; | |
3558 | } | |
3559 | x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1; | |
3560 | ||
3561 | if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) { | |
3562 | WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!", | |
3563 | x86_pmu.num_counters_fixed, INTEL_PMC_MAX_FIXED); | |
3564 | x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED; | |
3565 | } | |
3566 | ||
3567 | x86_pmu.intel_ctrl |= | |
3568 | ((1LL << x86_pmu.num_counters_fixed)-1) << INTEL_PMC_IDX_FIXED; | |
3569 | ||
3570 | if (x86_pmu.event_constraints) { | |
3571 | /* | |
3572 | * event on fixed counter2 (REF_CYCLES) only works on this | |
3573 | * counter, so do not extend mask to generic counters | |
3574 | */ | |
3575 | for_each_event_constraint(c, x86_pmu.event_constraints) { | |
2c33645d PI |
3576 | if (c->cmask == FIXED_EVENT_FLAGS |
3577 | && c->idxmsk64 != INTEL_PMC_MSK_FIXED_REF_CYCLES) { | |
3578 | c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1; | |
a1eac7ac | 3579 | } |
2c33645d PI |
3580 | c->idxmsk64 &= |
3581 | ~(~0UL << (INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed)); | |
3582 | c->weight = hweight64(c->idxmsk64); | |
a1eac7ac RR |
3583 | } |
3584 | } | |
3585 | ||
338b522c KL |
3586 | /* |
3587 | * Access LBR MSR may cause #GP under certain circumstances. | |
3588 | * E.g. KVM doesn't support LBR MSR | |
3589 | * Check all LBT MSR here. | |
3590 | * Disable LBR access if any LBR MSRs can not be accessed. | |
3591 | */ | |
3592 | if (x86_pmu.lbr_nr && !check_msr(x86_pmu.lbr_tos, 0x3UL)) | |
3593 | x86_pmu.lbr_nr = 0; | |
3594 | for (i = 0; i < x86_pmu.lbr_nr; i++) { | |
3595 | if (!(check_msr(x86_pmu.lbr_from + i, 0xffffUL) && | |
3596 | check_msr(x86_pmu.lbr_to + i, 0xffffUL))) | |
3597 | x86_pmu.lbr_nr = 0; | |
3598 | } | |
3599 | ||
3600 | /* | |
3601 | * Access extra MSR may cause #GP under certain circumstances. | |
3602 | * E.g. KVM doesn't support offcore event | |
3603 | * Check all extra_regs here. | |
3604 | */ | |
3605 | if (x86_pmu.extra_regs) { | |
3606 | for (er = x86_pmu.extra_regs; er->msr; er++) { | |
8c4fe709 | 3607 | er->extra_msr_access = check_msr(er->msr, 0x11UL); |
338b522c KL |
3608 | /* Disable LBR select mapping */ |
3609 | if ((er->idx == EXTRA_REG_LBR) && !er->extra_msr_access) | |
3610 | x86_pmu.lbr_sel_map = NULL; | |
3611 | } | |
3612 | } | |
3613 | ||
069e0c3c AK |
3614 | /* Support full width counters using alternative MSR range */ |
3615 | if (x86_pmu.intel_cap.full_width_write) { | |
3616 | x86_pmu.max_period = x86_pmu.cntval_mask; | |
3617 | x86_pmu.perfctr = MSR_IA32_PMC0; | |
3618 | pr_cont("full-width counters, "); | |
3619 | } | |
3620 | ||
f22f54f4 PZ |
3621 | return 0; |
3622 | } | |
b37609c3 SE |
3623 | |
3624 | /* | |
3625 | * HT bug: phase 2 init | |
3626 | * Called once we have valid topology information to check | |
3627 | * whether or not HT is enabled | |
3628 | * If HT is off, then we disable the workaround | |
3629 | */ | |
3630 | static __init int fixup_ht_bug(void) | |
3631 | { | |
3632 | int cpu = smp_processor_id(); | |
3633 | int w, c; | |
3634 | /* | |
3635 | * problem not present on this CPU model, nothing to do | |
3636 | */ | |
3637 | if (!(x86_pmu.flags & PMU_FL_EXCL_ENABLED)) | |
3638 | return 0; | |
3639 | ||
06931e62 | 3640 | w = cpumask_weight(topology_sibling_cpumask(cpu)); |
b37609c3 SE |
3641 | if (w > 1) { |
3642 | pr_info("PMU erratum BJ122, BV98, HSD29 worked around, HT is on\n"); | |
3643 | return 0; | |
3644 | } | |
3645 | ||
ec6a9066 | 3646 | if (lockup_detector_suspend() != 0) { |
999bbe49 UO |
3647 | pr_debug("failed to disable PMU erratum BJ122, BV98, HSD29 workaround\n"); |
3648 | return 0; | |
3649 | } | |
b37609c3 SE |
3650 | |
3651 | x86_pmu.flags &= ~(PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED); | |
3652 | ||
b37609c3 | 3653 | x86_pmu.start_scheduling = NULL; |
0c41e756 | 3654 | x86_pmu.commit_scheduling = NULL; |
b37609c3 SE |
3655 | x86_pmu.stop_scheduling = NULL; |
3656 | ||
ec6a9066 | 3657 | lockup_detector_resume(); |
b37609c3 SE |
3658 | |
3659 | get_online_cpus(); | |
3660 | ||
3661 | for_each_online_cpu(c) { | |
3662 | free_excl_cntrs(c); | |
3663 | } | |
3664 | ||
3665 | put_online_cpus(); | |
3666 | pr_info("PMU erratum BJ122, BV98, HSD29 workaround disabled, HT off\n"); | |
3667 | return 0; | |
3668 | } | |
3669 | subsys_initcall(fixup_ht_bug) |