diff options
author | Jonathan Herman <hermanjl@cs.unc.edu> | 2013-01-17 16:15:55 -0500 |
---|---|---|
committer | Jonathan Herman <hermanjl@cs.unc.edu> | 2013-01-17 16:15:55 -0500 |
commit | 8dea78da5cee153b8af9c07a2745f6c55057fe12 (patch) | |
tree | a8f4d49d63b1ecc92f2fddceba0655b2472c5bd9 /arch/x86/kernel/cpu/perf_event_intel.c | |
parent | 406089d01562f1e2bf9f089fd7637009ebaad589 (diff) |
Patched in Tegra support.
Diffstat (limited to 'arch/x86/kernel/cpu/perf_event_intel.c')
-rw-r--r-- | arch/x86/kernel/cpu/perf_event_intel.c | 826 |
1 files changed, 155 insertions, 671 deletions
diff --git a/arch/x86/kernel/cpu/perf_event_intel.c b/arch/x86/kernel/cpu/perf_event_intel.c index 93b9e1181f8..f88af2c2a56 100644 --- a/arch/x86/kernel/cpu/perf_event_intel.c +++ b/arch/x86/kernel/cpu/perf_event_intel.c | |||
@@ -1,36 +1,29 @@ | |||
1 | #ifdef CONFIG_CPU_SUP_INTEL | ||
2 | |||
1 | /* | 3 | /* |
2 | * Per core/cpu state | 4 | * Per core/cpu state |
3 | * | 5 | * |
4 | * Used to coordinate shared registers between HT threads or | 6 | * Used to coordinate shared registers between HT threads or |
5 | * among events on a single PMU. | 7 | * among events on a single PMU. |
6 | */ | 8 | */ |
7 | 9 | struct intel_shared_regs { | |
8 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt | 10 | struct er_account regs[EXTRA_REG_MAX]; |
9 | 11 | int refcnt; /* per-core: #HT threads */ | |
10 | #include <linux/stddef.h> | 12 | unsigned core_id; /* per-core: core id */ |
11 | #include <linux/types.h> | 13 | }; |
12 | #include <linux/init.h> | ||
13 | #include <linux/slab.h> | ||
14 | #include <linux/export.h> | ||
15 | |||
16 | #include <asm/hardirq.h> | ||
17 | #include <asm/apic.h> | ||
18 | |||
19 | #include "perf_event.h" | ||
20 | 14 | ||
21 | /* | 15 | /* |
22 | * Intel PerfMon, used on Core and later. | 16 | * Intel PerfMon, used on Core and later. |
23 | */ | 17 | */ |
24 | static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly = | 18 | static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly = |
25 | { | 19 | { |
26 | [PERF_COUNT_HW_CPU_CYCLES] = 0x003c, | 20 | [PERF_COUNT_HW_CPU_CYCLES] = 0x003c, |
27 | [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0, | 21 | [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0, |
28 | [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e, | 22 | [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e, |
29 | [PERF_COUNT_HW_CACHE_MISSES] = 0x412e, | 23 | [PERF_COUNT_HW_CACHE_MISSES] = 0x412e, |
30 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4, | 24 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4, |
31 | [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5, | 25 | [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5, |
32 | [PERF_COUNT_HW_BUS_CYCLES] = 0x013c, | 26 | [PERF_COUNT_HW_BUS_CYCLES] = 0x013c, |
33 | [PERF_COUNT_HW_REF_CPU_CYCLES] = 0x0300, /* pseudo-encoding */ | ||
34 | }; | 27 | }; |
35 | 28 | ||
36 | static struct event_constraint intel_core_event_constraints[] __read_mostly = | 29 | static struct event_constraint intel_core_event_constraints[] __read_mostly = |
@@ -48,7 +41,12 @@ static struct event_constraint intel_core2_event_constraints[] __read_mostly = | |||
48 | { | 41 | { |
49 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | 42 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
50 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | 43 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ |
51 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ | 44 | /* |
45 | * Core2 has Fixed Counter 2 listed as CPU_CLK_UNHALTED.REF and event | ||
46 | * 0x013c as CPU_CLK_UNHALTED.BUS and specifies there is a fixed | ||
47 | * ratio between these counters. | ||
48 | */ | ||
49 | /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */ | ||
52 | INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */ | 50 | INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */ |
53 | INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ | 51 | INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ |
54 | INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ | 52 | INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ |
@@ -66,7 +64,7 @@ static struct event_constraint intel_nehalem_event_constraints[] __read_mostly = | |||
66 | { | 64 | { |
67 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | 65 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
68 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | 66 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ |
69 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ | 67 | /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */ |
70 | INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */ | 68 | INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */ |
71 | INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */ | 69 | INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */ |
72 | INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */ | 70 | INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */ |
@@ -88,7 +86,7 @@ static struct event_constraint intel_westmere_event_constraints[] __read_mostly | |||
88 | { | 86 | { |
89 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | 87 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
90 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | 88 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ |
91 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ | 89 | /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */ |
92 | INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ | 90 | INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ |
93 | INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */ | 91 | INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */ |
94 | INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ | 92 | INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ |
@@ -100,7 +98,7 @@ static struct event_constraint intel_snb_event_constraints[] __read_mostly = | |||
100 | { | 98 | { |
101 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | 99 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
102 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | 100 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ |
103 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ | 101 | /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */ |
104 | INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */ | 102 | INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */ |
105 | INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ | 103 | INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ |
106 | INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ | 104 | INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ |
@@ -123,7 +121,7 @@ static struct event_constraint intel_gen_event_constraints[] __read_mostly = | |||
123 | { | 121 | { |
124 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ | 122 | FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ |
125 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ | 123 | FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ |
126 | FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ | 124 | /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */ |
127 | EVENT_CONSTRAINT_END | 125 | EVENT_CONSTRAINT_END |
128 | }; | 126 | }; |
129 | 127 | ||
@@ -138,84 +136,6 @@ static u64 intel_pmu_event_map(int hw_event) | |||
138 | return intel_perfmon_event_map[hw_event]; | 136 | return intel_perfmon_event_map[hw_event]; |
139 | } | 137 | } |
140 | 138 | ||
141 | #define SNB_DMND_DATA_RD (1ULL << 0) | ||
142 | #define SNB_DMND_RFO (1ULL << 1) | ||
143 | #define SNB_DMND_IFETCH (1ULL << 2) | ||
144 | #define SNB_DMND_WB (1ULL << 3) | ||
145 | #define SNB_PF_DATA_RD (1ULL << 4) | ||
146 | #define SNB_PF_RFO (1ULL << 5) | ||
147 | #define SNB_PF_IFETCH (1ULL << 6) | ||
148 | #define SNB_LLC_DATA_RD (1ULL << 7) | ||
149 | #define SNB_LLC_RFO (1ULL << 8) | ||
150 | #define SNB_LLC_IFETCH (1ULL << 9) | ||
151 | #define SNB_BUS_LOCKS (1ULL << 10) | ||
152 | #define SNB_STRM_ST (1ULL << 11) | ||
153 | #define SNB_OTHER (1ULL << 15) | ||
154 | #define SNB_RESP_ANY (1ULL << 16) | ||
155 | #define SNB_NO_SUPP (1ULL << 17) | ||
156 | #define SNB_LLC_HITM (1ULL << 18) | ||
157 | #define SNB_LLC_HITE (1ULL << 19) | ||
158 | #define SNB_LLC_HITS (1ULL << 20) | ||
159 | #define SNB_LLC_HITF (1ULL << 21) | ||
160 | #define SNB_LOCAL (1ULL << 22) | ||
161 | #define SNB_REMOTE (0xffULL << 23) | ||
162 | #define SNB_SNP_NONE (1ULL << 31) | ||
163 | #define SNB_SNP_NOT_NEEDED (1ULL << 32) | ||
164 | #define SNB_SNP_MISS (1ULL << 33) | ||
165 | #define SNB_NO_FWD (1ULL << 34) | ||
166 | #define SNB_SNP_FWD (1ULL << 35) | ||
167 | #define SNB_HITM (1ULL << 36) | ||
168 | #define SNB_NON_DRAM (1ULL << 37) | ||
169 | |||
170 | #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD) | ||
171 | #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO) | ||
172 | #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) | ||
173 | |||
174 | #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \ | ||
175 | SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \ | ||
176 | SNB_HITM) | ||
177 | |||
178 | #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY) | ||
179 | #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY) | ||
180 | |||
181 | #define SNB_L3_ACCESS SNB_RESP_ANY | ||
182 | #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM) | ||
183 | |||
184 | static __initconst const u64 snb_hw_cache_extra_regs | ||
185 | [PERF_COUNT_HW_CACHE_MAX] | ||
186 | [PERF_COUNT_HW_CACHE_OP_MAX] | ||
187 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = | ||
188 | { | ||
189 | [ C(LL ) ] = { | ||
190 | [ C(OP_READ) ] = { | ||
191 | [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS, | ||
192 | [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_L3_MISS, | ||
193 | }, | ||
194 | [ C(OP_WRITE) ] = { | ||
195 | [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS, | ||
196 | [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_L3_MISS, | ||
197 | }, | ||
198 | [ C(OP_PREFETCH) ] = { | ||
199 | [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS, | ||
200 | [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_L3_MISS, | ||
201 | }, | ||
202 | }, | ||
203 | [ C(NODE) ] = { | ||
204 | [ C(OP_READ) ] = { | ||
205 | [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY, | ||
206 | [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_DRAM_REMOTE, | ||
207 | }, | ||
208 | [ C(OP_WRITE) ] = { | ||
209 | [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY, | ||
210 | [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE, | ||
211 | }, | ||
212 | [ C(OP_PREFETCH) ] = { | ||
213 | [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY, | ||
214 | [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE, | ||
215 | }, | ||
216 | }, | ||
217 | }; | ||
218 | |||
219 | static __initconst const u64 snb_hw_cache_event_ids | 139 | static __initconst const u64 snb_hw_cache_event_ids |
220 | [PERF_COUNT_HW_CACHE_MAX] | 140 | [PERF_COUNT_HW_CACHE_MAX] |
221 | [PERF_COUNT_HW_CACHE_OP_MAX] | 141 | [PERF_COUNT_HW_CACHE_OP_MAX] |
@@ -313,16 +233,16 @@ static __initconst const u64 snb_hw_cache_event_ids | |||
313 | }, | 233 | }, |
314 | [ C(NODE) ] = { | 234 | [ C(NODE) ] = { |
315 | [ C(OP_READ) ] = { | 235 | [ C(OP_READ) ] = { |
316 | [ C(RESULT_ACCESS) ] = 0x01b7, | 236 | [ C(RESULT_ACCESS) ] = -1, |
317 | [ C(RESULT_MISS) ] = 0x01b7, | 237 | [ C(RESULT_MISS) ] = -1, |
318 | }, | 238 | }, |
319 | [ C(OP_WRITE) ] = { | 239 | [ C(OP_WRITE) ] = { |
320 | [ C(RESULT_ACCESS) ] = 0x01b7, | 240 | [ C(RESULT_ACCESS) ] = -1, |
321 | [ C(RESULT_MISS) ] = 0x01b7, | 241 | [ C(RESULT_MISS) ] = -1, |
322 | }, | 242 | }, |
323 | [ C(OP_PREFETCH) ] = { | 243 | [ C(OP_PREFETCH) ] = { |
324 | [ C(RESULT_ACCESS) ] = 0x01b7, | 244 | [ C(RESULT_ACCESS) ] = -1, |
325 | [ C(RESULT_MISS) ] = 0x01b7, | 245 | [ C(RESULT_MISS) ] = -1, |
326 | }, | 246 | }, |
327 | }, | 247 | }, |
328 | 248 | ||
@@ -465,15 +385,14 @@ static __initconst const u64 westmere_hw_cache_event_ids | |||
465 | #define NHM_LOCAL_DRAM (1 << 14) | 385 | #define NHM_LOCAL_DRAM (1 << 14) |
466 | #define NHM_NON_DRAM (1 << 15) | 386 | #define NHM_NON_DRAM (1 << 15) |
467 | 387 | ||
468 | #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD) | 388 | #define NHM_ALL_DRAM (NHM_REMOTE_DRAM|NHM_LOCAL_DRAM) |
469 | #define NHM_REMOTE (NHM_REMOTE_DRAM) | ||
470 | 389 | ||
471 | #define NHM_DMND_READ (NHM_DMND_DATA_RD) | 390 | #define NHM_DMND_READ (NHM_DMND_DATA_RD) |
472 | #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB) | 391 | #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB) |
473 | #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO) | 392 | #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO) |
474 | 393 | ||
475 | #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM) | 394 | #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM) |
476 | #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD) | 395 | #define NHM_L3_MISS (NHM_NON_DRAM|NHM_ALL_DRAM|NHM_REMOTE_CACHE_FWD) |
477 | #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS) | 396 | #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS) |
478 | 397 | ||
479 | static __initconst const u64 nehalem_hw_cache_extra_regs | 398 | static __initconst const u64 nehalem_hw_cache_extra_regs |
@@ -497,16 +416,16 @@ static __initconst const u64 nehalem_hw_cache_extra_regs | |||
497 | }, | 416 | }, |
498 | [ C(NODE) ] = { | 417 | [ C(NODE) ] = { |
499 | [ C(OP_READ) ] = { | 418 | [ C(OP_READ) ] = { |
500 | [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE, | 419 | [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_ALL_DRAM, |
501 | [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE, | 420 | [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE_DRAM, |
502 | }, | 421 | }, |
503 | [ C(OP_WRITE) ] = { | 422 | [ C(OP_WRITE) ] = { |
504 | [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE, | 423 | [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_ALL_DRAM, |
505 | [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE, | 424 | [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE_DRAM, |
506 | }, | 425 | }, |
507 | [ C(OP_PREFETCH) ] = { | 426 | [ C(OP_PREFETCH) ] = { |
508 | [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE, | 427 | [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_ALL_DRAM, |
509 | [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE, | 428 | [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE_DRAM, |
510 | }, | 429 | }, |
511 | }, | 430 | }, |
512 | }; | 431 | }; |
@@ -808,26 +727,13 @@ static __initconst const u64 atom_hw_cache_event_ids | |||
808 | }, | 727 | }, |
809 | }; | 728 | }; |
810 | 729 | ||
811 | static inline bool intel_pmu_needs_lbr_smpl(struct perf_event *event) | ||
812 | { | ||
813 | /* user explicitly requested branch sampling */ | ||
814 | if (has_branch_stack(event)) | ||
815 | return true; | ||
816 | |||
817 | /* implicit branch sampling to correct PEBS skid */ | ||
818 | if (x86_pmu.intel_cap.pebs_trap && event->attr.precise_ip > 1) | ||
819 | return true; | ||
820 | |||
821 | return false; | ||
822 | } | ||
823 | |||
824 | static void intel_pmu_disable_all(void) | 730 | static void intel_pmu_disable_all(void) |
825 | { | 731 | { |
826 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | 732 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); |
827 | 733 | ||
828 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); | 734 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); |
829 | 735 | ||
830 | if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) | 736 | if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask)) |
831 | intel_pmu_disable_bts(); | 737 | intel_pmu_disable_bts(); |
832 | 738 | ||
833 | intel_pmu_pebs_disable_all(); | 739 | intel_pmu_pebs_disable_all(); |
@@ -840,12 +746,11 @@ static void intel_pmu_enable_all(int added) | |||
840 | 746 | ||
841 | intel_pmu_pebs_enable_all(); | 747 | intel_pmu_pebs_enable_all(); |
842 | intel_pmu_lbr_enable_all(); | 748 | intel_pmu_lbr_enable_all(); |
843 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, | 749 | wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl); |
844 | x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask); | ||
845 | 750 | ||
846 | if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) { | 751 | if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask)) { |
847 | struct perf_event *event = | 752 | struct perf_event *event = |
848 | cpuc->events[INTEL_PMC_IDX_FIXED_BTS]; | 753 | cpuc->events[X86_PMC_IDX_FIXED_BTS]; |
849 | 754 | ||
850 | if (WARN_ON_ONCE(!event)) | 755 | if (WARN_ON_ONCE(!event)) |
851 | return; | 756 | return; |
@@ -951,7 +856,7 @@ static inline void intel_pmu_ack_status(u64 ack) | |||
951 | 856 | ||
952 | static void intel_pmu_disable_fixed(struct hw_perf_event *hwc) | 857 | static void intel_pmu_disable_fixed(struct hw_perf_event *hwc) |
953 | { | 858 | { |
954 | int idx = hwc->idx - INTEL_PMC_IDX_FIXED; | 859 | int idx = hwc->idx - X86_PMC_IDX_FIXED; |
955 | u64 ctrl_val, mask; | 860 | u64 ctrl_val, mask; |
956 | 861 | ||
957 | mask = 0xfULL << (idx * 4); | 862 | mask = 0xfULL << (idx * 4); |
@@ -964,24 +869,13 @@ static void intel_pmu_disable_fixed(struct hw_perf_event *hwc) | |||
964 | static void intel_pmu_disable_event(struct perf_event *event) | 869 | static void intel_pmu_disable_event(struct perf_event *event) |
965 | { | 870 | { |
966 | struct hw_perf_event *hwc = &event->hw; | 871 | struct hw_perf_event *hwc = &event->hw; |
967 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | ||
968 | 872 | ||
969 | if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) { | 873 | if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) { |
970 | intel_pmu_disable_bts(); | 874 | intel_pmu_disable_bts(); |
971 | intel_pmu_drain_bts_buffer(); | 875 | intel_pmu_drain_bts_buffer(); |
972 | return; | 876 | return; |
973 | } | 877 | } |
974 | 878 | ||
975 | cpuc->intel_ctrl_guest_mask &= ~(1ull << hwc->idx); | ||
976 | cpuc->intel_ctrl_host_mask &= ~(1ull << hwc->idx); | ||
977 | |||
978 | /* | ||
979 | * must disable before any actual event | ||
980 | * because any event may be combined with LBR | ||
981 | */ | ||
982 | if (intel_pmu_needs_lbr_smpl(event)) | ||
983 | intel_pmu_lbr_disable(event); | ||
984 | |||
985 | if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { | 879 | if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { |
986 | intel_pmu_disable_fixed(hwc); | 880 | intel_pmu_disable_fixed(hwc); |
987 | return; | 881 | return; |
@@ -995,7 +889,7 @@ static void intel_pmu_disable_event(struct perf_event *event) | |||
995 | 889 | ||
996 | static void intel_pmu_enable_fixed(struct hw_perf_event *hwc) | 890 | static void intel_pmu_enable_fixed(struct hw_perf_event *hwc) |
997 | { | 891 | { |
998 | int idx = hwc->idx - INTEL_PMC_IDX_FIXED; | 892 | int idx = hwc->idx - X86_PMC_IDX_FIXED; |
999 | u64 ctrl_val, bits, mask; | 893 | u64 ctrl_val, bits, mask; |
1000 | 894 | ||
1001 | /* | 895 | /* |
@@ -1027,26 +921,14 @@ static void intel_pmu_enable_fixed(struct hw_perf_event *hwc) | |||
1027 | static void intel_pmu_enable_event(struct perf_event *event) | 921 | static void intel_pmu_enable_event(struct perf_event *event) |
1028 | { | 922 | { |
1029 | struct hw_perf_event *hwc = &event->hw; | 923 | struct hw_perf_event *hwc = &event->hw; |
1030 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | ||
1031 | 924 | ||
1032 | if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) { | 925 | if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) { |
1033 | if (!__this_cpu_read(cpu_hw_events.enabled)) | 926 | if (!__this_cpu_read(cpu_hw_events.enabled)) |
1034 | return; | 927 | return; |
1035 | 928 | ||
1036 | intel_pmu_enable_bts(hwc->config); | 929 | intel_pmu_enable_bts(hwc->config); |
1037 | return; | 930 | return; |
1038 | } | 931 | } |
1039 | /* | ||
1040 | * must enabled before any actual event | ||
1041 | * because any event may be combined with LBR | ||
1042 | */ | ||
1043 | if (intel_pmu_needs_lbr_smpl(event)) | ||
1044 | intel_pmu_lbr_enable(event); | ||
1045 | |||
1046 | if (event->attr.exclude_host) | ||
1047 | cpuc->intel_ctrl_guest_mask |= (1ull << hwc->idx); | ||
1048 | if (event->attr.exclude_guest) | ||
1049 | cpuc->intel_ctrl_host_mask |= (1ull << hwc->idx); | ||
1050 | 932 | ||
1051 | if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { | 933 | if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { |
1052 | intel_pmu_enable_fixed(hwc); | 934 | intel_pmu_enable_fixed(hwc); |
@@ -1063,7 +945,7 @@ static void intel_pmu_enable_event(struct perf_event *event) | |||
1063 | * Save and restart an expired event. Called by NMI contexts, | 945 | * Save and restart an expired event. Called by NMI contexts, |
1064 | * so it has to be careful about preempting normal event ops: | 946 | * so it has to be careful about preempting normal event ops: |
1065 | */ | 947 | */ |
1066 | int intel_pmu_save_and_restart(struct perf_event *event) | 948 | static int intel_pmu_save_and_restart(struct perf_event *event) |
1067 | { | 949 | { |
1068 | x86_perf_event_update(event); | 950 | x86_perf_event_update(event); |
1069 | return x86_perf_event_set_period(event); | 951 | return x86_perf_event_set_period(event); |
@@ -1080,14 +962,14 @@ static void intel_pmu_reset(void) | |||
1080 | 962 | ||
1081 | local_irq_save(flags); | 963 | local_irq_save(flags); |
1082 | 964 | ||
1083 | pr_info("clearing PMU state on CPU#%d\n", smp_processor_id()); | 965 | printk("clearing PMU state on CPU#%d\n", smp_processor_id()); |
1084 | 966 | ||
1085 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { | 967 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
1086 | wrmsrl_safe(x86_pmu_config_addr(idx), 0ull); | 968 | checking_wrmsrl(x86_pmu_config_addr(idx), 0ull); |
1087 | wrmsrl_safe(x86_pmu_event_addr(idx), 0ull); | 969 | checking_wrmsrl(x86_pmu_event_addr(idx), 0ull); |
1088 | } | 970 | } |
1089 | for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) | 971 | for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) |
1090 | wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull); | 972 | checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull); |
1091 | 973 | ||
1092 | if (ds) | 974 | if (ds) |
1093 | ds->bts_index = ds->bts_buffer_base; | 975 | ds->bts_index = ds->bts_buffer_base; |
@@ -1107,6 +989,8 @@ static int intel_pmu_handle_irq(struct pt_regs *regs) | |||
1107 | u64 status; | 989 | u64 status; |
1108 | int handled; | 990 | int handled; |
1109 | 991 | ||
992 | perf_sample_data_init(&data, 0); | ||
993 | |||
1110 | cpuc = &__get_cpu_var(cpu_hw_events); | 994 | cpuc = &__get_cpu_var(cpu_hw_events); |
1111 | 995 | ||
1112 | /* | 996 | /* |
@@ -1160,10 +1044,7 @@ again: | |||
1160 | if (!intel_pmu_save_and_restart(event)) | 1044 | if (!intel_pmu_save_and_restart(event)) |
1161 | continue; | 1045 | continue; |
1162 | 1046 | ||
1163 | perf_sample_data_init(&data, 0, event->hw.last_period); | 1047 | data.period = event->hw.last_period; |
1164 | |||
1165 | if (has_branch_stack(event)) | ||
1166 | data.br_stack = &cpuc->lbr_stack; | ||
1167 | 1048 | ||
1168 | if (perf_event_overflow(event, &data, regs)) | 1049 | if (perf_event_overflow(event, &data, regs)) |
1169 | x86_pmu_stop(event, 0); | 1050 | x86_pmu_stop(event, 0); |
@@ -1199,33 +1080,27 @@ intel_bts_constraints(struct perf_event *event) | |||
1199 | return NULL; | 1080 | return NULL; |
1200 | } | 1081 | } |
1201 | 1082 | ||
1202 | static int intel_alt_er(int idx) | 1083 | static bool intel_try_alt_er(struct perf_event *event, int orig_idx) |
1203 | { | 1084 | { |
1204 | if (!(x86_pmu.er_flags & ERF_HAS_RSP_1)) | 1085 | if (!(x86_pmu.er_flags & ERF_HAS_RSP_1)) |
1205 | return idx; | 1086 | return false; |
1206 | |||
1207 | if (idx == EXTRA_REG_RSP_0) | ||
1208 | return EXTRA_REG_RSP_1; | ||
1209 | 1087 | ||
1210 | if (idx == EXTRA_REG_RSP_1) | 1088 | if (event->hw.extra_reg.idx == EXTRA_REG_RSP_0) { |
1211 | return EXTRA_REG_RSP_0; | ||
1212 | |||
1213 | return idx; | ||
1214 | } | ||
1215 | |||
1216 | static void intel_fixup_er(struct perf_event *event, int idx) | ||
1217 | { | ||
1218 | event->hw.extra_reg.idx = idx; | ||
1219 | |||
1220 | if (idx == EXTRA_REG_RSP_0) { | ||
1221 | event->hw.config &= ~INTEL_ARCH_EVENT_MASK; | ||
1222 | event->hw.config |= 0x01b7; | ||
1223 | event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0; | ||
1224 | } else if (idx == EXTRA_REG_RSP_1) { | ||
1225 | event->hw.config &= ~INTEL_ARCH_EVENT_MASK; | 1089 | event->hw.config &= ~INTEL_ARCH_EVENT_MASK; |
1226 | event->hw.config |= 0x01bb; | 1090 | event->hw.config |= 0x01bb; |
1091 | event->hw.extra_reg.idx = EXTRA_REG_RSP_1; | ||
1227 | event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1; | 1092 | event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1; |
1093 | } else if (event->hw.extra_reg.idx == EXTRA_REG_RSP_1) { | ||
1094 | event->hw.config &= ~INTEL_ARCH_EVENT_MASK; | ||
1095 | event->hw.config |= 0x01b7; | ||
1096 | event->hw.extra_reg.idx = EXTRA_REG_RSP_0; | ||
1097 | event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0; | ||
1228 | } | 1098 | } |
1099 | |||
1100 | if (event->hw.extra_reg.idx == orig_idx) | ||
1101 | return false; | ||
1102 | |||
1103 | return true; | ||
1229 | } | 1104 | } |
1230 | 1105 | ||
1231 | /* | 1106 | /* |
@@ -1237,24 +1112,20 @@ static void intel_fixup_er(struct perf_event *event, int idx) | |||
1237 | */ | 1112 | */ |
1238 | static struct event_constraint * | 1113 | static struct event_constraint * |
1239 | __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc, | 1114 | __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc, |
1240 | struct perf_event *event, | 1115 | struct perf_event *event) |
1241 | struct hw_perf_event_extra *reg) | ||
1242 | { | 1116 | { |
1243 | struct event_constraint *c = &emptyconstraint; | 1117 | struct event_constraint *c = &emptyconstraint; |
1118 | struct hw_perf_event_extra *reg = &event->hw.extra_reg; | ||
1244 | struct er_account *era; | 1119 | struct er_account *era; |
1245 | unsigned long flags; | 1120 | unsigned long flags; |
1246 | int idx = reg->idx; | 1121 | int orig_idx = reg->idx; |
1247 | 1122 | ||
1248 | /* | 1123 | /* already allocated shared msr */ |
1249 | * reg->alloc can be set due to existing state, so for fake cpuc we | 1124 | if (reg->alloc) |
1250 | * need to ignore this, otherwise we might fail to allocate proper fake | 1125 | return &unconstrained; |
1251 | * state for this extra reg constraint. Also see the comment below. | ||
1252 | */ | ||
1253 | if (reg->alloc && !cpuc->is_fake) | ||
1254 | return NULL; /* call x86_get_event_constraint() */ | ||
1255 | 1126 | ||
1256 | again: | 1127 | again: |
1257 | era = &cpuc->shared_regs->regs[idx]; | 1128 | era = &cpuc->shared_regs->regs[reg->idx]; |
1258 | /* | 1129 | /* |
1259 | * we use spin_lock_irqsave() to avoid lockdep issues when | 1130 | * we use spin_lock_irqsave() to avoid lockdep issues when |
1260 | * passing a fake cpuc | 1131 | * passing a fake cpuc |
@@ -1263,29 +1134,6 @@ again: | |||
1263 | 1134 | ||
1264 | if (!atomic_read(&era->ref) || era->config == reg->config) { | 1135 | if (!atomic_read(&era->ref) || era->config == reg->config) { |
1265 | 1136 | ||
1266 | /* | ||
1267 | * If its a fake cpuc -- as per validate_{group,event}() we | ||
1268 | * shouldn't touch event state and we can avoid doing so | ||
1269 | * since both will only call get_event_constraints() once | ||
1270 | * on each event, this avoids the need for reg->alloc. | ||
1271 | * | ||
1272 | * Not doing the ER fixup will only result in era->reg being | ||
1273 | * wrong, but since we won't actually try and program hardware | ||
1274 | * this isn't a problem either. | ||
1275 | */ | ||
1276 | if (!cpuc->is_fake) { | ||
1277 | if (idx != reg->idx) | ||
1278 | intel_fixup_er(event, idx); | ||
1279 | |||
1280 | /* | ||
1281 | * x86_schedule_events() can call get_event_constraints() | ||
1282 | * multiple times on events in the case of incremental | ||
1283 | * scheduling(). reg->alloc ensures we only do the ER | ||
1284 | * allocation once. | ||
1285 | */ | ||
1286 | reg->alloc = 1; | ||
1287 | } | ||
1288 | |||
1289 | /* lock in msr value */ | 1137 | /* lock in msr value */ |
1290 | era->config = reg->config; | 1138 | era->config = reg->config; |
1291 | era->reg = reg->reg; | 1139 | era->reg = reg->reg; |
@@ -1293,17 +1141,21 @@ again: | |||
1293 | /* one more user */ | 1141 | /* one more user */ |
1294 | atomic_inc(&era->ref); | 1142 | atomic_inc(&era->ref); |
1295 | 1143 | ||
1144 | /* no need to reallocate during incremental event scheduling */ | ||
1145 | reg->alloc = 1; | ||
1146 | |||
1296 | /* | 1147 | /* |
1297 | * need to call x86_get_event_constraint() | 1148 | * All events using extra_reg are unconstrained. |
1298 | * to check if associated event has constraints | 1149 | * Avoids calling x86_get_event_constraints() |
1150 | * | ||
1151 | * Must revisit if extra_reg controlling events | ||
1152 | * ever have constraints. Worst case we go through | ||
1153 | * the regular event constraint table. | ||
1299 | */ | 1154 | */ |
1300 | c = NULL; | 1155 | c = &unconstrained; |
1301 | } else { | 1156 | } else if (intel_try_alt_er(event, orig_idx)) { |
1302 | idx = intel_alt_er(idx); | 1157 | raw_spin_unlock(&era->lock); |
1303 | if (idx != reg->idx) { | 1158 | goto again; |
1304 | raw_spin_unlock_irqrestore(&era->lock, flags); | ||
1305 | goto again; | ||
1306 | } | ||
1307 | } | 1159 | } |
1308 | raw_spin_unlock_irqrestore(&era->lock, flags); | 1160 | raw_spin_unlock_irqrestore(&era->lock, flags); |
1309 | 1161 | ||
@@ -1317,14 +1169,11 @@ __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc, | |||
1317 | struct er_account *era; | 1169 | struct er_account *era; |
1318 | 1170 | ||
1319 | /* | 1171 | /* |
1320 | * Only put constraint if extra reg was actually allocated. Also takes | 1172 | * only put constraint if extra reg was actually |
1321 | * care of event which do not use an extra shared reg. | 1173 | * allocated. Also takes care of event which do |
1322 | * | 1174 | * not use an extra shared reg |
1323 | * Also, if this is a fake cpuc we shouldn't touch any event state | ||
1324 | * (reg->alloc) and we don't care about leaving inconsistent cpuc state | ||
1325 | * either since it'll be thrown out. | ||
1326 | */ | 1175 | */ |
1327 | if (!reg->alloc || cpuc->is_fake) | 1176 | if (!reg->alloc) |
1328 | return; | 1177 | return; |
1329 | 1178 | ||
1330 | era = &cpuc->shared_regs->regs[reg->idx]; | 1179 | era = &cpuc->shared_regs->regs[reg->idx]; |
@@ -1340,39 +1189,12 @@ static struct event_constraint * | |||
1340 | intel_shared_regs_constraints(struct cpu_hw_events *cpuc, | 1189 | intel_shared_regs_constraints(struct cpu_hw_events *cpuc, |
1341 | struct perf_event *event) | 1190 | struct perf_event *event) |
1342 | { | 1191 | { |
1343 | struct event_constraint *c = NULL, *d; | 1192 | struct event_constraint *c = NULL; |
1344 | struct hw_perf_event_extra *xreg, *breg; | ||
1345 | |||
1346 | xreg = &event->hw.extra_reg; | ||
1347 | if (xreg->idx != EXTRA_REG_NONE) { | ||
1348 | c = __intel_shared_reg_get_constraints(cpuc, event, xreg); | ||
1349 | if (c == &emptyconstraint) | ||
1350 | return c; | ||
1351 | } | ||
1352 | breg = &event->hw.branch_reg; | ||
1353 | if (breg->idx != EXTRA_REG_NONE) { | ||
1354 | d = __intel_shared_reg_get_constraints(cpuc, event, breg); | ||
1355 | if (d == &emptyconstraint) { | ||
1356 | __intel_shared_reg_put_constraints(cpuc, xreg); | ||
1357 | c = d; | ||
1358 | } | ||
1359 | } | ||
1360 | return c; | ||
1361 | } | ||
1362 | 1193 | ||
1363 | struct event_constraint * | 1194 | if (event->hw.extra_reg.idx != EXTRA_REG_NONE) |
1364 | x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event) | 1195 | c = __intel_shared_reg_get_constraints(cpuc, event); |
1365 | { | ||
1366 | struct event_constraint *c; | ||
1367 | 1196 | ||
1368 | if (x86_pmu.event_constraints) { | 1197 | return c; |
1369 | for_each_event_constraint(c, x86_pmu.event_constraints) { | ||
1370 | if ((event->hw.config & c->cmask) == c->code) | ||
1371 | return c; | ||
1372 | } | ||
1373 | } | ||
1374 | |||
1375 | return &unconstrained; | ||
1376 | } | 1198 | } |
1377 | 1199 | ||
1378 | static struct event_constraint * | 1200 | static struct event_constraint * |
@@ -1404,10 +1226,6 @@ intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc, | |||
1404 | reg = &event->hw.extra_reg; | 1226 | reg = &event->hw.extra_reg; |
1405 | if (reg->idx != EXTRA_REG_NONE) | 1227 | if (reg->idx != EXTRA_REG_NONE) |
1406 | __intel_shared_reg_put_constraints(cpuc, reg); | 1228 | __intel_shared_reg_put_constraints(cpuc, reg); |
1407 | |||
1408 | reg = &event->hw.branch_reg; | ||
1409 | if (reg->idx != EXTRA_REG_NONE) | ||
1410 | __intel_shared_reg_put_constraints(cpuc, reg); | ||
1411 | } | 1229 | } |
1412 | 1230 | ||
1413 | static void intel_put_event_constraints(struct cpu_hw_events *cpuc, | 1231 | static void intel_put_event_constraints(struct cpu_hw_events *cpuc, |
@@ -1416,9 +1234,15 @@ static void intel_put_event_constraints(struct cpu_hw_events *cpuc, | |||
1416 | intel_put_shared_regs_event_constraints(cpuc, event); | 1234 | intel_put_shared_regs_event_constraints(cpuc, event); |
1417 | } | 1235 | } |
1418 | 1236 | ||
1419 | static void intel_pebs_aliases_core2(struct perf_event *event) | 1237 | static int intel_pmu_hw_config(struct perf_event *event) |
1420 | { | 1238 | { |
1421 | if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { | 1239 | int ret = x86_pmu_hw_config(event); |
1240 | |||
1241 | if (ret) | ||
1242 | return ret; | ||
1243 | |||
1244 | if (event->attr.precise_ip && | ||
1245 | (event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { | ||
1422 | /* | 1246 | /* |
1423 | * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P | 1247 | * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P |
1424 | * (0x003c) so that we can use it with PEBS. | 1248 | * (0x003c) so that we can use it with PEBS. |
@@ -1437,56 +1261,11 @@ static void intel_pebs_aliases_core2(struct perf_event *event) | |||
1437 | * | 1261 | * |
1438 | * Thereby we gain a PEBS capable cycle counter. | 1262 | * Thereby we gain a PEBS capable cycle counter. |
1439 | */ | 1263 | */ |
1440 | u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16); | 1264 | u64 alt_config = 0x108000c0; /* INST_RETIRED.TOTAL_CYCLES */ |
1441 | |||
1442 | alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); | ||
1443 | event->hw.config = alt_config; | ||
1444 | } | ||
1445 | } | ||
1446 | |||
1447 | static void intel_pebs_aliases_snb(struct perf_event *event) | ||
1448 | { | ||
1449 | if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { | ||
1450 | /* | ||
1451 | * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P | ||
1452 | * (0x003c) so that we can use it with PEBS. | ||
1453 | * | ||
1454 | * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't | ||
1455 | * PEBS capable. However we can use UOPS_RETIRED.ALL | ||
1456 | * (0x01c2), which is a PEBS capable event, to get the same | ||
1457 | * count. | ||
1458 | * | ||
1459 | * UOPS_RETIRED.ALL counts the number of cycles that retires | ||
1460 | * CNTMASK micro-ops. By setting CNTMASK to a value (16) | ||
1461 | * larger than the maximum number of micro-ops that can be | ||
1462 | * retired per cycle (4) and then inverting the condition, we | ||
1463 | * count all cycles that retire 16 or less micro-ops, which | ||
1464 | * is every cycle. | ||
1465 | * | ||
1466 | * Thereby we gain a PEBS capable cycle counter. | ||
1467 | */ | ||
1468 | u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16); | ||
1469 | 1265 | ||
1470 | alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); | 1266 | alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); |
1471 | event->hw.config = alt_config; | 1267 | event->hw.config = alt_config; |
1472 | } | 1268 | } |
1473 | } | ||
1474 | |||
1475 | static int intel_pmu_hw_config(struct perf_event *event) | ||
1476 | { | ||
1477 | int ret = x86_pmu_hw_config(event); | ||
1478 | |||
1479 | if (ret) | ||
1480 | return ret; | ||
1481 | |||
1482 | if (event->attr.precise_ip && x86_pmu.pebs_aliases) | ||
1483 | x86_pmu.pebs_aliases(event); | ||
1484 | |||
1485 | if (intel_pmu_needs_lbr_smpl(event)) { | ||
1486 | ret = intel_pmu_setup_lbr_filter(event); | ||
1487 | if (ret) | ||
1488 | return ret; | ||
1489 | } | ||
1490 | 1269 | ||
1491 | if (event->attr.type != PERF_TYPE_RAW) | 1270 | if (event->attr.type != PERF_TYPE_RAW) |
1492 | return 0; | 1271 | return 0; |
@@ -1505,117 +1284,12 @@ static int intel_pmu_hw_config(struct perf_event *event) | |||
1505 | return 0; | 1284 | return 0; |
1506 | } | 1285 | } |
1507 | 1286 | ||
1508 | struct perf_guest_switch_msr *perf_guest_get_msrs(int *nr) | ||
1509 | { | ||
1510 | if (x86_pmu.guest_get_msrs) | ||
1511 | return x86_pmu.guest_get_msrs(nr); | ||
1512 | *nr = 0; | ||
1513 | return NULL; | ||
1514 | } | ||
1515 | EXPORT_SYMBOL_GPL(perf_guest_get_msrs); | ||
1516 | |||
1517 | static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr) | ||
1518 | { | ||
1519 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | ||
1520 | struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; | ||
1521 | |||
1522 | arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL; | ||
1523 | arr[0].host = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask; | ||
1524 | arr[0].guest = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_host_mask; | ||
1525 | /* | ||
1526 | * If PMU counter has PEBS enabled it is not enough to disable counter | ||
1527 | * on a guest entry since PEBS memory write can overshoot guest entry | ||
1528 | * and corrupt guest memory. Disabling PEBS solves the problem. | ||
1529 | */ | ||
1530 | arr[1].msr = MSR_IA32_PEBS_ENABLE; | ||
1531 | arr[1].host = cpuc->pebs_enabled; | ||
1532 | arr[1].guest = 0; | ||
1533 | |||
1534 | *nr = 2; | ||
1535 | return arr; | ||
1536 | } | ||
1537 | |||
1538 | static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr) | ||
1539 | { | ||
1540 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | ||
1541 | struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; | ||
1542 | int idx; | ||
1543 | |||
1544 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { | ||
1545 | struct perf_event *event = cpuc->events[idx]; | ||
1546 | |||
1547 | arr[idx].msr = x86_pmu_config_addr(idx); | ||
1548 | arr[idx].host = arr[idx].guest = 0; | ||
1549 | |||
1550 | if (!test_bit(idx, cpuc->active_mask)) | ||
1551 | continue; | ||
1552 | |||
1553 | arr[idx].host = arr[idx].guest = | ||
1554 | event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE; | ||
1555 | |||
1556 | if (event->attr.exclude_host) | ||
1557 | arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE; | ||
1558 | else if (event->attr.exclude_guest) | ||
1559 | arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE; | ||
1560 | } | ||
1561 | |||
1562 | *nr = x86_pmu.num_counters; | ||
1563 | return arr; | ||
1564 | } | ||
1565 | |||
1566 | static void core_pmu_enable_event(struct perf_event *event) | ||
1567 | { | ||
1568 | if (!event->attr.exclude_host) | ||
1569 | x86_pmu_enable_event(event); | ||
1570 | } | ||
1571 | |||
1572 | static void core_pmu_enable_all(int added) | ||
1573 | { | ||
1574 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | ||
1575 | int idx; | ||
1576 | |||
1577 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { | ||
1578 | struct hw_perf_event *hwc = &cpuc->events[idx]->hw; | ||
1579 | |||
1580 | if (!test_bit(idx, cpuc->active_mask) || | ||
1581 | cpuc->events[idx]->attr.exclude_host) | ||
1582 | continue; | ||
1583 | |||
1584 | __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); | ||
1585 | } | ||
1586 | } | ||
1587 | |||
1588 | PMU_FORMAT_ATTR(event, "config:0-7" ); | ||
1589 | PMU_FORMAT_ATTR(umask, "config:8-15" ); | ||
1590 | PMU_FORMAT_ATTR(edge, "config:18" ); | ||
1591 | PMU_FORMAT_ATTR(pc, "config:19" ); | ||
1592 | PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */ | ||
1593 | PMU_FORMAT_ATTR(inv, "config:23" ); | ||
1594 | PMU_FORMAT_ATTR(cmask, "config:24-31" ); | ||
1595 | |||
1596 | static struct attribute *intel_arch_formats_attr[] = { | ||
1597 | &format_attr_event.attr, | ||
1598 | &format_attr_umask.attr, | ||
1599 | &format_attr_edge.attr, | ||
1600 | &format_attr_pc.attr, | ||
1601 | &format_attr_inv.attr, | ||
1602 | &format_attr_cmask.attr, | ||
1603 | NULL, | ||
1604 | }; | ||
1605 | |||
1606 | ssize_t intel_event_sysfs_show(char *page, u64 config) | ||
1607 | { | ||
1608 | u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT); | ||
1609 | |||
1610 | return x86_event_sysfs_show(page, config, event); | ||
1611 | } | ||
1612 | |||
1613 | static __initconst const struct x86_pmu core_pmu = { | 1287 | static __initconst const struct x86_pmu core_pmu = { |
1614 | .name = "core", | 1288 | .name = "core", |
1615 | .handle_irq = x86_pmu_handle_irq, | 1289 | .handle_irq = x86_pmu_handle_irq, |
1616 | .disable_all = x86_pmu_disable_all, | 1290 | .disable_all = x86_pmu_disable_all, |
1617 | .enable_all = core_pmu_enable_all, | 1291 | .enable_all = x86_pmu_enable_all, |
1618 | .enable = core_pmu_enable_event, | 1292 | .enable = x86_pmu_enable_event, |
1619 | .disable = x86_pmu_disable_event, | 1293 | .disable = x86_pmu_disable_event, |
1620 | .hw_config = x86_pmu_hw_config, | 1294 | .hw_config = x86_pmu_hw_config, |
1621 | .schedule_events = x86_schedule_events, | 1295 | .schedule_events = x86_schedule_events, |
@@ -1633,12 +1307,9 @@ static __initconst const struct x86_pmu core_pmu = { | |||
1633 | .get_event_constraints = intel_get_event_constraints, | 1307 | .get_event_constraints = intel_get_event_constraints, |
1634 | .put_event_constraints = intel_put_event_constraints, | 1308 | .put_event_constraints = intel_put_event_constraints, |
1635 | .event_constraints = intel_core_event_constraints, | 1309 | .event_constraints = intel_core_event_constraints, |
1636 | .guest_get_msrs = core_guest_get_msrs, | ||
1637 | .format_attrs = intel_arch_formats_attr, | ||
1638 | .events_sysfs_show = intel_event_sysfs_show, | ||
1639 | }; | 1310 | }; |
1640 | 1311 | ||
1641 | struct intel_shared_regs *allocate_shared_regs(int cpu) | 1312 | static struct intel_shared_regs *allocate_shared_regs(int cpu) |
1642 | { | 1313 | { |
1643 | struct intel_shared_regs *regs; | 1314 | struct intel_shared_regs *regs; |
1644 | int i; | 1315 | int i; |
@@ -1661,7 +1332,7 @@ static int intel_pmu_cpu_prepare(int cpu) | |||
1661 | { | 1332 | { |
1662 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); | 1333 | struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); |
1663 | 1334 | ||
1664 | if (!(x86_pmu.extra_regs || x86_pmu.lbr_sel_map)) | 1335 | if (!x86_pmu.extra_regs) |
1665 | return NOTIFY_OK; | 1336 | return NOTIFY_OK; |
1666 | 1337 | ||
1667 | cpuc->shared_regs = allocate_shared_regs(cpu); | 1338 | cpuc->shared_regs = allocate_shared_regs(cpu); |
@@ -1683,28 +1354,22 @@ static void intel_pmu_cpu_starting(int cpu) | |||
1683 | */ | 1354 | */ |
1684 | intel_pmu_lbr_reset(); | 1355 | intel_pmu_lbr_reset(); |
1685 | 1356 | ||
1686 | cpuc->lbr_sel = NULL; | 1357 | if (!cpuc->shared_regs || (x86_pmu.er_flags & ERF_NO_HT_SHARING)) |
1687 | |||
1688 | if (!cpuc->shared_regs) | ||
1689 | return; | 1358 | return; |
1690 | 1359 | ||
1691 | if (!(x86_pmu.er_flags & ERF_NO_HT_SHARING)) { | 1360 | for_each_cpu(i, topology_thread_cpumask(cpu)) { |
1692 | for_each_cpu(i, topology_thread_cpumask(cpu)) { | 1361 | struct intel_shared_regs *pc; |
1693 | struct intel_shared_regs *pc; | ||
1694 | 1362 | ||
1695 | pc = per_cpu(cpu_hw_events, i).shared_regs; | 1363 | pc = per_cpu(cpu_hw_events, i).shared_regs; |
1696 | if (pc && pc->core_id == core_id) { | 1364 | if (pc && pc->core_id == core_id) { |
1697 | cpuc->kfree_on_online = cpuc->shared_regs; | 1365 | kfree(cpuc->shared_regs); |
1698 | cpuc->shared_regs = pc; | 1366 | cpuc->shared_regs = pc; |
1699 | break; | 1367 | break; |
1700 | } | ||
1701 | } | 1368 | } |
1702 | cpuc->shared_regs->core_id = core_id; | ||
1703 | cpuc->shared_regs->refcnt++; | ||
1704 | } | 1369 | } |
1705 | 1370 | ||
1706 | if (x86_pmu.lbr_sel_map) | 1371 | cpuc->shared_regs->core_id = core_id; |
1707 | cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR]; | 1372 | cpuc->shared_regs->refcnt++; |
1708 | } | 1373 | } |
1709 | 1374 | ||
1710 | static void intel_pmu_cpu_dying(int cpu) | 1375 | static void intel_pmu_cpu_dying(int cpu) |
@@ -1722,33 +1387,6 @@ static void intel_pmu_cpu_dying(int cpu) | |||
1722 | fini_debug_store_on_cpu(cpu); | 1387 | fini_debug_store_on_cpu(cpu); |
1723 | } | 1388 | } |
1724 | 1389 | ||
1725 | static void intel_pmu_flush_branch_stack(void) | ||
1726 | { | ||
1727 | /* | ||
1728 | * Intel LBR does not tag entries with the | ||
1729 | * PID of the current task, then we need to | ||
1730 | * flush it on ctxsw | ||
1731 | * For now, we simply reset it | ||
1732 | */ | ||
1733 | if (x86_pmu.lbr_nr) | ||
1734 | intel_pmu_lbr_reset(); | ||
1735 | } | ||
1736 | |||
1737 | PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63"); | ||
1738 | |||
1739 | static struct attribute *intel_arch3_formats_attr[] = { | ||
1740 | &format_attr_event.attr, | ||
1741 | &format_attr_umask.attr, | ||
1742 | &format_attr_edge.attr, | ||
1743 | &format_attr_pc.attr, | ||
1744 | &format_attr_any.attr, | ||
1745 | &format_attr_inv.attr, | ||
1746 | &format_attr_cmask.attr, | ||
1747 | |||
1748 | &format_attr_offcore_rsp.attr, /* XXX do NHM/WSM + SNB breakout */ | ||
1749 | NULL, | ||
1750 | }; | ||
1751 | |||
1752 | static __initconst const struct x86_pmu intel_pmu = { | 1390 | static __initconst const struct x86_pmu intel_pmu = { |
1753 | .name = "Intel", | 1391 | .name = "Intel", |
1754 | .handle_irq = intel_pmu_handle_irq, | 1392 | .handle_irq = intel_pmu_handle_irq, |
@@ -1771,19 +1409,13 @@ static __initconst const struct x86_pmu intel_pmu = { | |||
1771 | .max_period = (1ULL << 31) - 1, | 1409 | .max_period = (1ULL << 31) - 1, |
1772 | .get_event_constraints = intel_get_event_constraints, | 1410 | .get_event_constraints = intel_get_event_constraints, |
1773 | .put_event_constraints = intel_put_event_constraints, | 1411 | .put_event_constraints = intel_put_event_constraints, |
1774 | .pebs_aliases = intel_pebs_aliases_core2, | ||
1775 | |||
1776 | .format_attrs = intel_arch3_formats_attr, | ||
1777 | .events_sysfs_show = intel_event_sysfs_show, | ||
1778 | 1412 | ||
1779 | .cpu_prepare = intel_pmu_cpu_prepare, | 1413 | .cpu_prepare = intel_pmu_cpu_prepare, |
1780 | .cpu_starting = intel_pmu_cpu_starting, | 1414 | .cpu_starting = intel_pmu_cpu_starting, |
1781 | .cpu_dying = intel_pmu_cpu_dying, | 1415 | .cpu_dying = intel_pmu_cpu_dying, |
1782 | .guest_get_msrs = intel_guest_get_msrs, | ||
1783 | .flush_branch_stack = intel_pmu_flush_branch_stack, | ||
1784 | }; | 1416 | }; |
1785 | 1417 | ||
1786 | static __init void intel_clovertown_quirk(void) | 1418 | static void intel_clovertown_quirks(void) |
1787 | { | 1419 | { |
1788 | /* | 1420 | /* |
1789 | * PEBS is unreliable due to: | 1421 | * PEBS is unreliable due to: |
@@ -1804,119 +1436,23 @@ static __init void intel_clovertown_quirk(void) | |||
1804 | * But taken together it might just make sense to not enable PEBS on | 1436 | * But taken together it might just make sense to not enable PEBS on |
1805 | * these chips. | 1437 | * these chips. |
1806 | */ | 1438 | */ |
1807 | pr_warn("PEBS disabled due to CPU errata\n"); | 1439 | printk(KERN_WARNING "PEBS disabled due to CPU errata.\n"); |
1808 | x86_pmu.pebs = 0; | 1440 | x86_pmu.pebs = 0; |
1809 | x86_pmu.pebs_constraints = NULL; | 1441 | x86_pmu.pebs_constraints = NULL; |
1810 | } | 1442 | } |
1811 | 1443 | ||
1812 | static int intel_snb_pebs_broken(int cpu) | 1444 | static __init int intel_pmu_init(void) |
1813 | { | ||
1814 | u32 rev = UINT_MAX; /* default to broken for unknown models */ | ||
1815 | |||
1816 | switch (cpu_data(cpu).x86_model) { | ||
1817 | case 42: /* SNB */ | ||
1818 | rev = 0x28; | ||
1819 | break; | ||
1820 | |||
1821 | case 45: /* SNB-EP */ | ||
1822 | switch (cpu_data(cpu).x86_mask) { | ||
1823 | case 6: rev = 0x618; break; | ||
1824 | case 7: rev = 0x70c; break; | ||
1825 | } | ||
1826 | } | ||
1827 | |||
1828 | return (cpu_data(cpu).microcode < rev); | ||
1829 | } | ||
1830 | |||
1831 | static void intel_snb_check_microcode(void) | ||
1832 | { | ||
1833 | int pebs_broken = 0; | ||
1834 | int cpu; | ||
1835 | |||
1836 | get_online_cpus(); | ||
1837 | for_each_online_cpu(cpu) { | ||
1838 | if ((pebs_broken = intel_snb_pebs_broken(cpu))) | ||
1839 | break; | ||
1840 | } | ||
1841 | put_online_cpus(); | ||
1842 | |||
1843 | if (pebs_broken == x86_pmu.pebs_broken) | ||
1844 | return; | ||
1845 | |||
1846 | /* | ||
1847 | * Serialized by the microcode lock.. | ||
1848 | */ | ||
1849 | if (x86_pmu.pebs_broken) { | ||
1850 | pr_info("PEBS enabled due to microcode update\n"); | ||
1851 | x86_pmu.pebs_broken = 0; | ||
1852 | } else { | ||
1853 | pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n"); | ||
1854 | x86_pmu.pebs_broken = 1; | ||
1855 | } | ||
1856 | } | ||
1857 | |||
1858 | static __init void intel_sandybridge_quirk(void) | ||
1859 | { | ||
1860 | x86_pmu.check_microcode = intel_snb_check_microcode; | ||
1861 | intel_snb_check_microcode(); | ||
1862 | } | ||
1863 | |||
1864 | static const struct { int id; char *name; } intel_arch_events_map[] __initconst = { | ||
1865 | { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" }, | ||
1866 | { PERF_COUNT_HW_INSTRUCTIONS, "instructions" }, | ||
1867 | { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" }, | ||
1868 | { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" }, | ||
1869 | { PERF_COUNT_HW_CACHE_MISSES, "cache misses" }, | ||
1870 | { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" }, | ||
1871 | { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" }, | ||
1872 | }; | ||
1873 | |||
1874 | static __init void intel_arch_events_quirk(void) | ||
1875 | { | ||
1876 | int bit; | ||
1877 | |||
1878 | /* disable event that reported as not presend by cpuid */ | ||
1879 | for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) { | ||
1880 | intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0; | ||
1881 | pr_warn("CPUID marked event: \'%s\' unavailable\n", | ||
1882 | intel_arch_events_map[bit].name); | ||
1883 | } | ||
1884 | } | ||
1885 | |||
1886 | static __init void intel_nehalem_quirk(void) | ||
1887 | { | ||
1888 | union cpuid10_ebx ebx; | ||
1889 | |||
1890 | ebx.full = x86_pmu.events_maskl; | ||
1891 | if (ebx.split.no_branch_misses_retired) { | ||
1892 | /* | ||
1893 | * Erratum AAJ80 detected, we work it around by using | ||
1894 | * the BR_MISP_EXEC.ANY event. This will over-count | ||
1895 | * branch-misses, but it's still much better than the | ||
1896 | * architectural event which is often completely bogus: | ||
1897 | */ | ||
1898 | intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89; | ||
1899 | ebx.split.no_branch_misses_retired = 0; | ||
1900 | x86_pmu.events_maskl = ebx.full; | ||
1901 | pr_info("CPU erratum AAJ80 worked around\n"); | ||
1902 | } | ||
1903 | } | ||
1904 | |||
1905 | __init int intel_pmu_init(void) | ||
1906 | { | 1445 | { |
1907 | union cpuid10_edx edx; | 1446 | union cpuid10_edx edx; |
1908 | union cpuid10_eax eax; | 1447 | union cpuid10_eax eax; |
1909 | union cpuid10_ebx ebx; | ||
1910 | struct event_constraint *c; | ||
1911 | unsigned int unused; | 1448 | unsigned int unused; |
1449 | unsigned int ebx; | ||
1912 | int version; | 1450 | int version; |
1913 | 1451 | ||
1914 | if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) { | 1452 | if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) { |
1915 | switch (boot_cpu_data.x86) { | 1453 | switch (boot_cpu_data.x86) { |
1916 | case 0x6: | 1454 | case 0x6: |
1917 | return p6_pmu_init(); | 1455 | return p6_pmu_init(); |
1918 | case 0xb: | ||
1919 | return knc_pmu_init(); | ||
1920 | case 0xf: | 1456 | case 0xf: |
1921 | return p4_pmu_init(); | 1457 | return p4_pmu_init(); |
1922 | } | 1458 | } |
@@ -1927,8 +1463,8 @@ __init int intel_pmu_init(void) | |||
1927 | * Check whether the Architectural PerfMon supports | 1463 | * Check whether the Architectural PerfMon supports |
1928 | * Branch Misses Retired hw_event or not. | 1464 | * Branch Misses Retired hw_event or not. |
1929 | */ | 1465 | */ |
1930 | cpuid(10, &eax.full, &ebx.full, &unused, &edx.full); | 1466 | cpuid(10, &eax.full, &ebx, &unused, &edx.full); |
1931 | if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT) | 1467 | if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED) |
1932 | return -ENODEV; | 1468 | return -ENODEV; |
1933 | 1469 | ||
1934 | version = eax.split.version_id; | 1470 | version = eax.split.version_id; |
@@ -1942,11 +1478,6 @@ __init int intel_pmu_init(void) | |||
1942 | x86_pmu.cntval_bits = eax.split.bit_width; | 1478 | x86_pmu.cntval_bits = eax.split.bit_width; |
1943 | x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1; | 1479 | x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1; |
1944 | 1480 | ||
1945 | x86_pmu.events_maskl = ebx.full; | ||
1946 | x86_pmu.events_mask_len = eax.split.mask_length; | ||
1947 | |||
1948 | x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters); | ||
1949 | |||
1950 | /* | 1481 | /* |
1951 | * Quirk: v2 perfmon does not report fixed-purpose events, so | 1482 | * Quirk: v2 perfmon does not report fixed-purpose events, so |
1952 | * assume at least 3 events: | 1483 | * assume at least 3 events: |
@@ -1966,8 +1497,6 @@ __init int intel_pmu_init(void) | |||
1966 | 1497 | ||
1967 | intel_ds_init(); | 1498 | intel_ds_init(); |
1968 | 1499 | ||
1969 | x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */ | ||
1970 | |||
1971 | /* | 1500 | /* |
1972 | * Install the hw-cache-events table: | 1501 | * Install the hw-cache-events table: |
1973 | */ | 1502 | */ |
@@ -1977,7 +1506,7 @@ __init int intel_pmu_init(void) | |||
1977 | break; | 1506 | break; |
1978 | 1507 | ||
1979 | case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */ | 1508 | case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */ |
1980 | x86_add_quirk(intel_clovertown_quirk); | 1509 | x86_pmu.quirks = intel_clovertown_quirks; |
1981 | case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */ | 1510 | case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */ |
1982 | case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */ | 1511 | case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */ |
1983 | case 29: /* six-core 45 nm xeon "Dunnington" */ | 1512 | case 29: /* six-core 45 nm xeon "Dunnington" */ |
@@ -2007,19 +1536,25 @@ __init int intel_pmu_init(void) | |||
2007 | x86_pmu.extra_regs = intel_nehalem_extra_regs; | 1536 | x86_pmu.extra_regs = intel_nehalem_extra_regs; |
2008 | 1537 | ||
2009 | /* UOPS_ISSUED.STALLED_CYCLES */ | 1538 | /* UOPS_ISSUED.STALLED_CYCLES */ |
2010 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = | 1539 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x180010e; |
2011 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | ||
2012 | /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ | 1540 | /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ |
2013 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = | 1541 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x1803fb1; |
2014 | X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); | ||
2015 | 1542 | ||
2016 | x86_add_quirk(intel_nehalem_quirk); | 1543 | if (ebx & 0x40) { |
1544 | /* | ||
1545 | * Erratum AAJ80 detected, we work it around by using | ||
1546 | * the BR_MISP_EXEC.ANY event. This will over-count | ||
1547 | * branch-misses, but it's still much better than the | ||
1548 | * architectural event which is often completely bogus: | ||
1549 | */ | ||
1550 | intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89; | ||
2017 | 1551 | ||
1552 | pr_cont("erratum AAJ80 worked around, "); | ||
1553 | } | ||
2018 | pr_cont("Nehalem events, "); | 1554 | pr_cont("Nehalem events, "); |
2019 | break; | 1555 | break; |
2020 | 1556 | ||
2021 | case 28: /* Atom */ | 1557 | case 28: /* Atom */ |
2022 | case 54: /* Cedariew */ | ||
2023 | memcpy(hw_cache_event_ids, atom_hw_cache_event_ids, | 1558 | memcpy(hw_cache_event_ids, atom_hw_cache_event_ids, |
2024 | sizeof(hw_cache_event_ids)); | 1559 | sizeof(hw_cache_event_ids)); |
2025 | 1560 | ||
@@ -2047,65 +1582,34 @@ __init int intel_pmu_init(void) | |||
2047 | x86_pmu.er_flags |= ERF_HAS_RSP_1; | 1582 | x86_pmu.er_flags |= ERF_HAS_RSP_1; |
2048 | 1583 | ||
2049 | /* UOPS_ISSUED.STALLED_CYCLES */ | 1584 | /* UOPS_ISSUED.STALLED_CYCLES */ |
2050 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = | 1585 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x180010e; |
2051 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | ||
2052 | /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ | 1586 | /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ |
2053 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = | 1587 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x1803fb1; |
2054 | X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); | ||
2055 | 1588 | ||
2056 | pr_cont("Westmere events, "); | 1589 | pr_cont("Westmere events, "); |
2057 | break; | 1590 | break; |
2058 | 1591 | ||
2059 | case 42: /* SandyBridge */ | 1592 | case 42: /* SandyBridge */ |
2060 | case 45: /* SandyBridge, "Romely-EP" */ | 1593 | case 45: /* SandyBridge, "Romely-EP" */ |
2061 | x86_add_quirk(intel_sandybridge_quirk); | ||
2062 | memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, | 1594 | memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, |
2063 | sizeof(hw_cache_event_ids)); | 1595 | sizeof(hw_cache_event_ids)); |
2064 | memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, | ||
2065 | sizeof(hw_cache_extra_regs)); | ||
2066 | 1596 | ||
2067 | intel_pmu_lbr_init_snb(); | 1597 | intel_pmu_lbr_init_nhm(); |
2068 | 1598 | ||
2069 | x86_pmu.event_constraints = intel_snb_event_constraints; | 1599 | x86_pmu.event_constraints = intel_snb_event_constraints; |
2070 | x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints; | 1600 | x86_pmu.pebs_constraints = intel_snb_pebs_events; |
2071 | x86_pmu.pebs_aliases = intel_pebs_aliases_snb; | ||
2072 | x86_pmu.extra_regs = intel_snb_extra_regs; | 1601 | x86_pmu.extra_regs = intel_snb_extra_regs; |
2073 | /* all extra regs are per-cpu when HT is on */ | 1602 | /* all extra regs are per-cpu when HT is on */ |
2074 | x86_pmu.er_flags |= ERF_HAS_RSP_1; | 1603 | x86_pmu.er_flags |= ERF_HAS_RSP_1; |
2075 | x86_pmu.er_flags |= ERF_NO_HT_SHARING; | 1604 | x86_pmu.er_flags |= ERF_NO_HT_SHARING; |
2076 | 1605 | ||
2077 | /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ | 1606 | /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ |
2078 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = | 1607 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x180010e; |
2079 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | ||
2080 | /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/ | 1608 | /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/ |
2081 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = | 1609 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x18001b1; |
2082 | X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1); | ||
2083 | 1610 | ||
2084 | pr_cont("SandyBridge events, "); | 1611 | pr_cont("SandyBridge events, "); |
2085 | break; | 1612 | break; |
2086 | case 58: /* IvyBridge */ | ||
2087 | memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, | ||
2088 | sizeof(hw_cache_event_ids)); | ||
2089 | memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, | ||
2090 | sizeof(hw_cache_extra_regs)); | ||
2091 | |||
2092 | intel_pmu_lbr_init_snb(); | ||
2093 | |||
2094 | x86_pmu.event_constraints = intel_snb_event_constraints; | ||
2095 | x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints; | ||
2096 | x86_pmu.pebs_aliases = intel_pebs_aliases_snb; | ||
2097 | x86_pmu.extra_regs = intel_snb_extra_regs; | ||
2098 | /* all extra regs are per-cpu when HT is on */ | ||
2099 | x86_pmu.er_flags |= ERF_HAS_RSP_1; | ||
2100 | x86_pmu.er_flags |= ERF_NO_HT_SHARING; | ||
2101 | |||
2102 | /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ | ||
2103 | intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = | ||
2104 | X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); | ||
2105 | |||
2106 | pr_cont("IvyBridge events, "); | ||
2107 | break; | ||
2108 | |||
2109 | 1613 | ||
2110 | default: | 1614 | default: |
2111 | switch (x86_pmu.version) { | 1615 | switch (x86_pmu.version) { |
@@ -2122,38 +1626,18 @@ __init int intel_pmu_init(void) | |||
2122 | break; | 1626 | break; |
2123 | } | 1627 | } |
2124 | } | 1628 | } |
1629 | return 0; | ||
1630 | } | ||
2125 | 1631 | ||
2126 | if (x86_pmu.num_counters > INTEL_PMC_MAX_GENERIC) { | 1632 | #else /* CONFIG_CPU_SUP_INTEL */ |
2127 | WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!", | ||
2128 | x86_pmu.num_counters, INTEL_PMC_MAX_GENERIC); | ||
2129 | x86_pmu.num_counters = INTEL_PMC_MAX_GENERIC; | ||
2130 | } | ||
2131 | x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1; | ||
2132 | |||
2133 | if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) { | ||
2134 | WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!", | ||
2135 | x86_pmu.num_counters_fixed, INTEL_PMC_MAX_FIXED); | ||
2136 | x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED; | ||
2137 | } | ||
2138 | |||
2139 | x86_pmu.intel_ctrl |= | ||
2140 | ((1LL << x86_pmu.num_counters_fixed)-1) << INTEL_PMC_IDX_FIXED; | ||
2141 | |||
2142 | if (x86_pmu.event_constraints) { | ||
2143 | /* | ||
2144 | * event on fixed counter2 (REF_CYCLES) only works on this | ||
2145 | * counter, so do not extend mask to generic counters | ||
2146 | */ | ||
2147 | for_each_event_constraint(c, x86_pmu.event_constraints) { | ||
2148 | if (c->cmask != X86_RAW_EVENT_MASK | ||
2149 | || c->idxmsk64 == INTEL_PMC_MSK_FIXED_REF_CYCLES) { | ||
2150 | continue; | ||
2151 | } | ||
2152 | |||
2153 | c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1; | ||
2154 | c->weight += x86_pmu.num_counters; | ||
2155 | } | ||
2156 | } | ||
2157 | 1633 | ||
1634 | static int intel_pmu_init(void) | ||
1635 | { | ||
2158 | return 0; | 1636 | return 0; |
2159 | } | 1637 | } |
1638 | |||
1639 | static struct intel_shared_regs *allocate_shared_regs(int cpu) | ||
1640 | { | ||
1641 | return NULL; | ||
1642 | } | ||
1643 | #endif /* CONFIG_CPU_SUP_INTEL */ | ||