Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
de0428a7 KW |
2 | #include <linux/bitops.h> |
3 | #include <linux/types.h> | |
4 | #include <linux/slab.h> | |
ca037701 | 5 | |
c1961a46 | 6 | #include <asm/cpu_entry_area.h> |
de0428a7 | 7 | #include <asm/perf_event.h> |
42f3bdc5 | 8 | #include <asm/tlbflush.h> |
3e702ff6 | 9 | #include <asm/insn.h> |
59e9f587 | 10 | #include <asm/io.h> |
de0428a7 | 11 | |
27f6d22b | 12 | #include "../perf_event.h" |
ca037701 | 13 | |
10043e02 TG |
14 | /* Waste a full page so it can be mapped into the cpu_entry_area */ |
15 | DEFINE_PER_CPU_PAGE_ALIGNED(struct debug_store, cpu_debug_store); | |
16 | ||
ca037701 PZ |
17 | /* The size of a BTS record in bytes: */ |
18 | #define BTS_RECORD_SIZE 24 | |
19 | ||
9536c8d2 | 20 | #define PEBS_FIXUP_SIZE PAGE_SIZE |
ca037701 PZ |
21 | |
22 | /* | |
23 | * pebs_record_32 for p4 and core not supported | |
24 | ||
25 | struct pebs_record_32 { | |
26 | u32 flags, ip; | |
27 | u32 ax, bc, cx, dx; | |
28 | u32 si, di, bp, sp; | |
29 | }; | |
30 | ||
31 | */ | |
32 | ||
f20093ee SE |
33 | union intel_x86_pebs_dse { |
34 | u64 val; | |
35 | struct { | |
36 | unsigned int ld_dse:4; | |
37 | unsigned int ld_stlb_miss:1; | |
38 | unsigned int ld_locked:1; | |
61b985e3 KL |
39 | unsigned int ld_data_blk:1; |
40 | unsigned int ld_addr_blk:1; | |
41 | unsigned int ld_reserved:24; | |
f20093ee SE |
42 | }; |
43 | struct { | |
44 | unsigned int st_l1d_hit:1; | |
45 | unsigned int st_reserved1:3; | |
46 | unsigned int st_stlb_miss:1; | |
47 | unsigned int st_locked:1; | |
48 | unsigned int st_reserved2:26; | |
49 | }; | |
61b985e3 KL |
50 | struct { |
51 | unsigned int st_lat_dse:4; | |
52 | unsigned int st_lat_stlb_miss:1; | |
53 | unsigned int st_lat_locked:1; | |
54 | unsigned int ld_reserved3:26; | |
55 | }; | |
38aaf921 KL |
56 | struct { |
57 | unsigned int mtl_dse:5; | |
58 | unsigned int mtl_locked:1; | |
59 | unsigned int mtl_stlb_miss:1; | |
60 | unsigned int mtl_fwd_blk:1; | |
61 | unsigned int ld_reserved4:24; | |
62 | }; | |
f20093ee SE |
63 | }; |
64 | ||
65 | ||
66 | /* | |
67 | * Map PEBS Load Latency Data Source encodings to generic | |
68 | * memory data source information | |
69 | */ | |
70 | #define P(a, b) PERF_MEM_S(a, b) | |
71 | #define OP_LH (P(OP, LOAD) | P(LVL, HIT)) | |
6ae5fa61 AK |
72 | #define LEVEL(x) P(LVLNUM, x) |
73 | #define REM P(REMOTE, REMOTE) | |
f20093ee SE |
74 | #define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS)) |
75 | ||
e17dc653 AK |
76 | /* Version for Sandy Bridge and later */ |
77 | static u64 pebs_data_source[] = { | |
6ae5fa61 AK |
78 | P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),/* 0x00:ukn L3 */ |
79 | OP_LH | P(LVL, L1) | LEVEL(L1) | P(SNOOP, NONE), /* 0x01: L1 local */ | |
80 | OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE), /* 0x02: LFB hit */ | |
81 | OP_LH | P(LVL, L2) | LEVEL(L2) | P(SNOOP, NONE), /* 0x03: L2 hit */ | |
82 | OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, NONE), /* 0x04: L3 hit */ | |
83 | OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, MISS), /* 0x05: L3 hit, snoop miss */ | |
84 | OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT), /* 0x06: L3 hit, snoop hit */ | |
85 | OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM), /* 0x07: L3 hit, snoop hitm */ | |
86 | OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HIT), /* 0x08: L3 miss snoop hit */ | |
87 | OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/ | |
88 | OP_LH | P(LVL, LOC_RAM) | LEVEL(RAM) | P(SNOOP, HIT), /* 0x0a: L3 miss, shared */ | |
89 | OP_LH | P(LVL, REM_RAM1) | REM | LEVEL(L3) | P(SNOOP, HIT), /* 0x0b: L3 miss, shared */ | |
90 | OP_LH | P(LVL, LOC_RAM) | LEVEL(RAM) | SNOOP_NONE_MISS, /* 0x0c: L3 miss, excl */ | |
91 | OP_LH | P(LVL, REM_RAM1) | LEVEL(RAM) | REM | SNOOP_NONE_MISS, /* 0x0d: L3 miss, excl */ | |
92 | OP_LH | P(LVL, IO) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0e: I/O */ | |
93 | OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0f: uncached */ | |
f20093ee SE |
94 | }; |
95 | ||
e17dc653 AK |
96 | /* Patch up minor differences in the bits */ |
97 | void __init intel_pmu_pebs_data_source_nhm(void) | |
98 | { | |
6ae5fa61 AK |
99 | pebs_data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT); |
100 | pebs_data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM); | |
101 | pebs_data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM); | |
102 | } | |
103 | ||
ccf170e9 | 104 | static void __init __intel_pmu_pebs_data_source_skl(bool pmem, u64 *data_source) |
6ae5fa61 AK |
105 | { |
106 | u64 pmem_or_l4 = pmem ? LEVEL(PMEM) : LEVEL(L4); | |
107 | ||
ccf170e9 KL |
108 | data_source[0x08] = OP_LH | pmem_or_l4 | P(SNOOP, HIT); |
109 | data_source[0x09] = OP_LH | pmem_or_l4 | REM | P(SNOOP, HIT); | |
110 | data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE); | |
111 | data_source[0x0c] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOPX, FWD); | |
112 | data_source[0x0d] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOP, HITM); | |
113 | } | |
114 | ||
115 | void __init intel_pmu_pebs_data_source_skl(bool pmem) | |
116 | { | |
117 | __intel_pmu_pebs_data_source_skl(pmem, pebs_data_source); | |
118 | } | |
119 | ||
24919fde | 120 | static void __init __intel_pmu_pebs_data_source_grt(u64 *data_source) |
ccf170e9 KL |
121 | { |
122 | data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT); | |
123 | data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM); | |
124 | data_source[0x08] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOPX, FWD); | |
125 | } | |
126 | ||
24919fde KL |
127 | void __init intel_pmu_pebs_data_source_grt(void) |
128 | { | |
129 | __intel_pmu_pebs_data_source_grt(pebs_data_source); | |
130 | } | |
131 | ||
ccf170e9 KL |
132 | void __init intel_pmu_pebs_data_source_adl(void) |
133 | { | |
134 | u64 *data_source; | |
135 | ||
136 | data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX].pebs_data_source; | |
137 | memcpy(data_source, pebs_data_source, sizeof(pebs_data_source)); | |
138 | __intel_pmu_pebs_data_source_skl(false, data_source); | |
139 | ||
140 | data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX].pebs_data_source; | |
141 | memcpy(data_source, pebs_data_source, sizeof(pebs_data_source)); | |
24919fde | 142 | __intel_pmu_pebs_data_source_grt(data_source); |
e17dc653 AK |
143 | } |
144 | ||
38aaf921 KL |
145 | static void __init intel_pmu_pebs_data_source_cmt(u64 *data_source) |
146 | { | |
147 | data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOPX, FWD); | |
148 | data_source[0x08] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM); | |
149 | data_source[0x0a] = OP_LH | P(LVL, LOC_RAM) | LEVEL(RAM) | P(SNOOP, NONE); | |
150 | data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE); | |
151 | data_source[0x0c] = OP_LH | LEVEL(RAM) | REM | P(SNOOPX, FWD); | |
152 | data_source[0x0d] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, HITM); | |
153 | } | |
154 | ||
155 | void __init intel_pmu_pebs_data_source_mtl(void) | |
156 | { | |
157 | u64 *data_source; | |
158 | ||
159 | data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX].pebs_data_source; | |
160 | memcpy(data_source, pebs_data_source, sizeof(pebs_data_source)); | |
161 | __intel_pmu_pebs_data_source_skl(false, data_source); | |
162 | ||
163 | data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX].pebs_data_source; | |
164 | memcpy(data_source, pebs_data_source, sizeof(pebs_data_source)); | |
165 | intel_pmu_pebs_data_source_cmt(data_source); | |
166 | } | |
167 | ||
9ad64c0f SE |
168 | static u64 precise_store_data(u64 status) |
169 | { | |
170 | union intel_x86_pebs_dse dse; | |
171 | u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2); | |
172 | ||
173 | dse.val = status; | |
174 | ||
175 | /* | |
176 | * bit 4: TLB access | |
177 | * 1 = stored missed 2nd level TLB | |
178 | * | |
179 | * so it either hit the walker or the OS | |
180 | * otherwise hit 2nd level TLB | |
181 | */ | |
182 | if (dse.st_stlb_miss) | |
183 | val |= P(TLB, MISS); | |
184 | else | |
185 | val |= P(TLB, HIT); | |
186 | ||
187 | /* | |
188 | * bit 0: hit L1 data cache | |
189 | * if not set, then all we know is that | |
190 | * it missed L1D | |
191 | */ | |
192 | if (dse.st_l1d_hit) | |
193 | val |= P(LVL, HIT); | |
194 | else | |
195 | val |= P(LVL, MISS); | |
196 | ||
197 | /* | |
198 | * bit 5: Locked prefix | |
199 | */ | |
200 | if (dse.st_locked) | |
201 | val |= P(LOCK, LOCKED); | |
202 | ||
203 | return val; | |
204 | } | |
205 | ||
c8aab2e0 | 206 | static u64 precise_datala_hsw(struct perf_event *event, u64 status) |
f9134f36 AK |
207 | { |
208 | union perf_mem_data_src dse; | |
209 | ||
770eee1f SE |
210 | dse.val = PERF_MEM_NA; |
211 | ||
212 | if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) | |
213 | dse.mem_op = PERF_MEM_OP_STORE; | |
214 | else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW) | |
215 | dse.mem_op = PERF_MEM_OP_LOAD; | |
722e76e6 SE |
216 | |
217 | /* | |
218 | * L1 info only valid for following events: | |
219 | * | |
220 | * MEM_UOPS_RETIRED.STLB_MISS_STORES | |
221 | * MEM_UOPS_RETIRED.LOCK_STORES | |
222 | * MEM_UOPS_RETIRED.SPLIT_STORES | |
223 | * MEM_UOPS_RETIRED.ALL_STORES | |
224 | */ | |
c8aab2e0 SE |
225 | if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) { |
226 | if (status & 1) | |
227 | dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT; | |
228 | else | |
229 | dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS; | |
230 | } | |
f9134f36 AK |
231 | return dse.val; |
232 | } | |
233 | ||
39a41278 KL |
234 | static inline void pebs_set_tlb_lock(u64 *val, bool tlb, bool lock) |
235 | { | |
236 | /* | |
237 | * TLB access | |
238 | * 0 = did not miss 2nd level TLB | |
239 | * 1 = missed 2nd level TLB | |
240 | */ | |
241 | if (tlb) | |
242 | *val |= P(TLB, MISS) | P(TLB, L2); | |
243 | else | |
244 | *val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2); | |
245 | ||
246 | /* locked prefix */ | |
247 | if (lock) | |
248 | *val |= P(LOCK, LOCKED); | |
249 | } | |
250 | ||
251 | /* Retrieve the latency data for e-core of ADL */ | |
38aaf921 KL |
252 | static u64 __adl_latency_data_small(struct perf_event *event, u64 status, |
253 | u8 dse, bool tlb, bool lock, bool blk) | |
39a41278 | 254 | { |
39a41278 KL |
255 | u64 val; |
256 | ||
257 | WARN_ON_ONCE(hybrid_pmu(event->pmu)->cpu_type == hybrid_big); | |
258 | ||
38aaf921 KL |
259 | dse &= PERF_PEBS_DATA_SOURCE_MASK; |
260 | val = hybrid_var(event->pmu, pebs_data_source)[dse]; | |
39a41278 | 261 | |
38aaf921 | 262 | pebs_set_tlb_lock(&val, tlb, lock); |
39a41278 | 263 | |
38aaf921 | 264 | if (blk) |
39a41278 KL |
265 | val |= P(BLK, DATA); |
266 | else | |
267 | val |= P(BLK, NA); | |
268 | ||
269 | return val; | |
270 | } | |
271 | ||
38aaf921 KL |
272 | u64 adl_latency_data_small(struct perf_event *event, u64 status) |
273 | { | |
274 | union intel_x86_pebs_dse dse; | |
275 | ||
276 | dse.val = status; | |
277 | ||
278 | return __adl_latency_data_small(event, status, dse.ld_dse, | |
279 | dse.ld_locked, dse.ld_stlb_miss, | |
280 | dse.ld_data_blk); | |
281 | } | |
282 | ||
283 | /* Retrieve the latency data for e-core of MTL */ | |
284 | u64 mtl_latency_data_small(struct perf_event *event, u64 status) | |
285 | { | |
286 | union intel_x86_pebs_dse dse; | |
287 | ||
288 | dse.val = status; | |
289 | ||
290 | return __adl_latency_data_small(event, status, dse.mtl_dse, | |
291 | dse.mtl_stlb_miss, dse.mtl_locked, | |
292 | dse.mtl_fwd_blk); | |
293 | } | |
294 | ||
ccf170e9 | 295 | static u64 load_latency_data(struct perf_event *event, u64 status) |
f20093ee SE |
296 | { |
297 | union intel_x86_pebs_dse dse; | |
298 | u64 val; | |
f20093ee SE |
299 | |
300 | dse.val = status; | |
301 | ||
302 | /* | |
303 | * use the mapping table for bit 0-3 | |
304 | */ | |
ccf170e9 | 305 | val = hybrid_var(event->pmu, pebs_data_source)[dse.ld_dse]; |
f20093ee SE |
306 | |
307 | /* | |
308 | * Nehalem models do not support TLB, Lock infos | |
309 | */ | |
95298355 | 310 | if (x86_pmu.pebs_no_tlb) { |
f20093ee SE |
311 | val |= P(TLB, NA) | P(LOCK, NA); |
312 | return val; | |
313 | } | |
f20093ee | 314 | |
39a41278 | 315 | pebs_set_tlb_lock(&val, dse.ld_stlb_miss, dse.ld_locked); |
f20093ee | 316 | |
61b985e3 KL |
317 | /* |
318 | * Ice Lake and earlier models do not support block infos. | |
319 | */ | |
320 | if (!x86_pmu.pebs_block) { | |
321 | val |= P(BLK, NA); | |
322 | return val; | |
323 | } | |
324 | /* | |
325 | * bit 6: load was blocked since its data could not be forwarded | |
326 | * from a preceding store | |
327 | */ | |
328 | if (dse.ld_data_blk) | |
329 | val |= P(BLK, DATA); | |
330 | ||
331 | /* | |
332 | * bit 7: load was blocked due to potential address conflict with | |
333 | * a preceding store | |
334 | */ | |
335 | if (dse.ld_addr_blk) | |
336 | val |= P(BLK, ADDR); | |
337 | ||
338 | if (!dse.ld_data_blk && !dse.ld_addr_blk) | |
339 | val |= P(BLK, NA); | |
340 | ||
341 | return val; | |
342 | } | |
343 | ||
ccf170e9 | 344 | static u64 store_latency_data(struct perf_event *event, u64 status) |
61b985e3 KL |
345 | { |
346 | union intel_x86_pebs_dse dse; | |
d4bdb0be | 347 | union perf_mem_data_src src; |
61b985e3 KL |
348 | u64 val; |
349 | ||
350 | dse.val = status; | |
351 | ||
352 | /* | |
353 | * use the mapping table for bit 0-3 | |
354 | */ | |
ccf170e9 | 355 | val = hybrid_var(event->pmu, pebs_data_source)[dse.st_lat_dse]; |
61b985e3 | 356 | |
39a41278 | 357 | pebs_set_tlb_lock(&val, dse.st_lat_stlb_miss, dse.st_lat_locked); |
61b985e3 KL |
358 | |
359 | val |= P(BLK, NA); | |
360 | ||
d4bdb0be SE |
361 | /* |
362 | * the pebs_data_source table is only for loads | |
363 | * so override the mem_op to say STORE instead | |
364 | */ | |
365 | src.val = val; | |
366 | src.mem_op = P(OP,STORE); | |
367 | ||
368 | return src.val; | |
f20093ee SE |
369 | } |
370 | ||
ca037701 PZ |
371 | struct pebs_record_core { |
372 | u64 flags, ip; | |
373 | u64 ax, bx, cx, dx; | |
374 | u64 si, di, bp, sp; | |
375 | u64 r8, r9, r10, r11; | |
376 | u64 r12, r13, r14, r15; | |
377 | }; | |
378 | ||
379 | struct pebs_record_nhm { | |
380 | u64 flags, ip; | |
381 | u64 ax, bx, cx, dx; | |
382 | u64 si, di, bp, sp; | |
383 | u64 r8, r9, r10, r11; | |
384 | u64 r12, r13, r14, r15; | |
385 | u64 status, dla, dse, lat; | |
386 | }; | |
387 | ||
130768b8 AK |
388 | /* |
389 | * Same as pebs_record_nhm, with two additional fields. | |
390 | */ | |
391 | struct pebs_record_hsw { | |
748e86aa AK |
392 | u64 flags, ip; |
393 | u64 ax, bx, cx, dx; | |
394 | u64 si, di, bp, sp; | |
395 | u64 r8, r9, r10, r11; | |
396 | u64 r12, r13, r14, r15; | |
397 | u64 status, dla, dse, lat; | |
d2beea4a | 398 | u64 real_ip, tsx_tuning; |
748e86aa AK |
399 | }; |
400 | ||
401 | union hsw_tsx_tuning { | |
402 | struct { | |
403 | u32 cycles_last_block : 32, | |
404 | hle_abort : 1, | |
405 | rtm_abort : 1, | |
406 | instruction_abort : 1, | |
407 | non_instruction_abort : 1, | |
408 | retry : 1, | |
409 | data_conflict : 1, | |
410 | capacity_writes : 1, | |
411 | capacity_reads : 1; | |
412 | }; | |
413 | u64 value; | |
130768b8 AK |
414 | }; |
415 | ||
a405bad5 AK |
416 | #define PEBS_HSW_TSX_FLAGS 0xff00000000ULL |
417 | ||
2f7ebf2e AK |
418 | /* Same as HSW, plus TSC */ |
419 | ||
420 | struct pebs_record_skl { | |
421 | u64 flags, ip; | |
422 | u64 ax, bx, cx, dx; | |
423 | u64 si, di, bp, sp; | |
424 | u64 r8, r9, r10, r11; | |
425 | u64 r12, r13, r14, r15; | |
426 | u64 status, dla, dse, lat; | |
427 | u64 real_ip, tsx_tuning; | |
428 | u64 tsc; | |
429 | }; | |
430 | ||
de0428a7 | 431 | void init_debug_store_on_cpu(int cpu) |
ca037701 PZ |
432 | { |
433 | struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds; | |
434 | ||
435 | if (!ds) | |
436 | return; | |
437 | ||
438 | wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, | |
439 | (u32)((u64)(unsigned long)ds), | |
440 | (u32)((u64)(unsigned long)ds >> 32)); | |
441 | } | |
442 | ||
de0428a7 | 443 | void fini_debug_store_on_cpu(int cpu) |
ca037701 PZ |
444 | { |
445 | if (!per_cpu(cpu_hw_events, cpu).ds) | |
446 | return; | |
447 | ||
448 | wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0); | |
449 | } | |
450 | ||
9536c8d2 PZ |
451 | static DEFINE_PER_CPU(void *, insn_buffer); |
452 | ||
c1961a46 | 453 | static void ds_update_cea(void *cea, void *addr, size_t size, pgprot_t prot) |
5ee25c87 | 454 | { |
42f3bdc5 | 455 | unsigned long start = (unsigned long)cea; |
c1961a46 HD |
456 | phys_addr_t pa; |
457 | size_t msz = 0; | |
458 | ||
459 | pa = virt_to_phys(addr); | |
42f3bdc5 PZ |
460 | |
461 | preempt_disable(); | |
c1961a46 HD |
462 | for (; msz < size; msz += PAGE_SIZE, pa += PAGE_SIZE, cea += PAGE_SIZE) |
463 | cea_set_pte(cea, pa, prot); | |
42f3bdc5 PZ |
464 | |
465 | /* | |
466 | * This is a cross-CPU update of the cpu_entry_area, we must shoot down | |
467 | * all TLB entries for it. | |
468 | */ | |
469 | flush_tlb_kernel_range(start, start + size); | |
470 | preempt_enable(); | |
c1961a46 HD |
471 | } |
472 | ||
473 | static void ds_clear_cea(void *cea, size_t size) | |
474 | { | |
42f3bdc5 | 475 | unsigned long start = (unsigned long)cea; |
c1961a46 HD |
476 | size_t msz = 0; |
477 | ||
42f3bdc5 | 478 | preempt_disable(); |
c1961a46 HD |
479 | for (; msz < size; msz += PAGE_SIZE, cea += PAGE_SIZE) |
480 | cea_set_pte(cea, 0, PAGE_NONE); | |
42f3bdc5 PZ |
481 | |
482 | flush_tlb_kernel_range(start, start + size); | |
483 | preempt_enable(); | |
c1961a46 HD |
484 | } |
485 | ||
486 | static void *dsalloc_pages(size_t size, gfp_t flags, int cpu) | |
487 | { | |
488 | unsigned int order = get_order(size); | |
96681fc3 | 489 | int node = cpu_to_node(cpu); |
c1961a46 HD |
490 | struct page *page; |
491 | ||
492 | page = __alloc_pages_node(node, flags | __GFP_ZERO, order); | |
493 | return page ? page_address(page) : NULL; | |
494 | } | |
495 | ||
496 | static void dsfree_pages(const void *buffer, size_t size) | |
497 | { | |
498 | if (buffer) | |
499 | free_pages((unsigned long)buffer, get_order(size)); | |
500 | } | |
501 | ||
502 | static int alloc_pebs_buffer(int cpu) | |
503 | { | |
504 | struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu); | |
505 | struct debug_store *ds = hwev->ds; | |
506 | size_t bsiz = x86_pmu.pebs_buffer_size; | |
507 | int max, node = cpu_to_node(cpu); | |
1fc654cf | 508 | void *buffer, *insn_buff, *cea; |
5ee25c87 PZ |
509 | |
510 | if (!x86_pmu.pebs) | |
511 | return 0; | |
512 | ||
c1961a46 | 513 | buffer = dsalloc_pages(bsiz, GFP_KERNEL, cpu); |
5ee25c87 PZ |
514 | if (unlikely(!buffer)) |
515 | return -ENOMEM; | |
516 | ||
9536c8d2 PZ |
517 | /* |
518 | * HSW+ already provides us the eventing ip; no need to allocate this | |
519 | * buffer then. | |
520 | */ | |
521 | if (x86_pmu.intel_cap.pebs_format < 2) { | |
1fc654cf IM |
522 | insn_buff = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node); |
523 | if (!insn_buff) { | |
c1961a46 | 524 | dsfree_pages(buffer, bsiz); |
9536c8d2 PZ |
525 | return -ENOMEM; |
526 | } | |
1fc654cf | 527 | per_cpu(insn_buffer, cpu) = insn_buff; |
9536c8d2 | 528 | } |
c1961a46 HD |
529 | hwev->ds_pebs_vaddr = buffer; |
530 | /* Update the cpu entry area mapping */ | |
531 | cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer; | |
532 | ds->pebs_buffer_base = (unsigned long) cea; | |
533 | ds_update_cea(cea, buffer, bsiz, PAGE_KERNEL); | |
5ee25c87 | 534 | ds->pebs_index = ds->pebs_buffer_base; |
c1961a46 HD |
535 | max = x86_pmu.pebs_record_size * (bsiz / x86_pmu.pebs_record_size); |
536 | ds->pebs_absolute_maximum = ds->pebs_buffer_base + max; | |
5ee25c87 PZ |
537 | return 0; |
538 | } | |
539 | ||
b39f88ac PZ |
540 | static void release_pebs_buffer(int cpu) |
541 | { | |
c1961a46 | 542 | struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu); |
c1961a46 | 543 | void *cea; |
b39f88ac | 544 | |
efe951d3 | 545 | if (!x86_pmu.pebs) |
b39f88ac PZ |
546 | return; |
547 | ||
9536c8d2 PZ |
548 | kfree(per_cpu(insn_buffer, cpu)); |
549 | per_cpu(insn_buffer, cpu) = NULL; | |
550 | ||
c1961a46 HD |
551 | /* Clear the fixmap */ |
552 | cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer; | |
553 | ds_clear_cea(cea, x86_pmu.pebs_buffer_size); | |
c1961a46 HD |
554 | dsfree_pages(hwev->ds_pebs_vaddr, x86_pmu.pebs_buffer_size); |
555 | hwev->ds_pebs_vaddr = NULL; | |
b39f88ac PZ |
556 | } |
557 | ||
5ee25c87 PZ |
558 | static int alloc_bts_buffer(int cpu) |
559 | { | |
c1961a46 HD |
560 | struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu); |
561 | struct debug_store *ds = hwev->ds; | |
562 | void *buffer, *cea; | |
563 | int max; | |
5ee25c87 PZ |
564 | |
565 | if (!x86_pmu.bts) | |
566 | return 0; | |
567 | ||
c1961a46 | 568 | buffer = dsalloc_pages(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, cpu); |
44851541 DR |
569 | if (unlikely(!buffer)) { |
570 | WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__); | |
5ee25c87 | 571 | return -ENOMEM; |
44851541 | 572 | } |
c1961a46 HD |
573 | hwev->ds_bts_vaddr = buffer; |
574 | /* Update the fixmap */ | |
575 | cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer; | |
576 | ds->bts_buffer_base = (unsigned long) cea; | |
577 | ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL); | |
5ee25c87 | 578 | ds->bts_index = ds->bts_buffer_base; |
2c991e40 HD |
579 | max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE; |
580 | ds->bts_absolute_maximum = ds->bts_buffer_base + | |
581 | max * BTS_RECORD_SIZE; | |
582 | ds->bts_interrupt_threshold = ds->bts_absolute_maximum - | |
583 | (max / 16) * BTS_RECORD_SIZE; | |
5ee25c87 PZ |
584 | return 0; |
585 | } | |
586 | ||
b39f88ac PZ |
587 | static void release_bts_buffer(int cpu) |
588 | { | |
c1961a46 | 589 | struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu); |
c1961a46 | 590 | void *cea; |
b39f88ac | 591 | |
efe951d3 | 592 | if (!x86_pmu.bts) |
b39f88ac PZ |
593 | return; |
594 | ||
c1961a46 HD |
595 | /* Clear the fixmap */ |
596 | cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer; | |
597 | ds_clear_cea(cea, BTS_BUFFER_SIZE); | |
c1961a46 HD |
598 | dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE); |
599 | hwev->ds_bts_vaddr = NULL; | |
b39f88ac PZ |
600 | } |
601 | ||
65af94ba PZ |
602 | static int alloc_ds_buffer(int cpu) |
603 | { | |
c1961a46 | 604 | struct debug_store *ds = &get_cpu_entry_area(cpu)->cpu_debug_store; |
65af94ba | 605 | |
c1961a46 | 606 | memset(ds, 0, sizeof(*ds)); |
65af94ba | 607 | per_cpu(cpu_hw_events, cpu).ds = ds; |
65af94ba PZ |
608 | return 0; |
609 | } | |
610 | ||
611 | static void release_ds_buffer(int cpu) | |
612 | { | |
65af94ba | 613 | per_cpu(cpu_hw_events, cpu).ds = NULL; |
65af94ba PZ |
614 | } |
615 | ||
de0428a7 | 616 | void release_ds_buffers(void) |
ca037701 PZ |
617 | { |
618 | int cpu; | |
619 | ||
620 | if (!x86_pmu.bts && !x86_pmu.pebs) | |
621 | return; | |
622 | ||
efe951d3 PZ |
623 | for_each_possible_cpu(cpu) |
624 | release_ds_buffer(cpu); | |
625 | ||
626 | for_each_possible_cpu(cpu) { | |
627 | /* | |
628 | * Again, ignore errors from offline CPUs, they will no longer | |
629 | * observe cpu_hw_events.ds and not program the DS_AREA when | |
630 | * they come up. | |
631 | */ | |
ca037701 | 632 | fini_debug_store_on_cpu(cpu); |
efe951d3 | 633 | } |
ca037701 PZ |
634 | |
635 | for_each_possible_cpu(cpu) { | |
b39f88ac PZ |
636 | release_pebs_buffer(cpu); |
637 | release_bts_buffer(cpu); | |
ca037701 | 638 | } |
ca037701 PZ |
639 | } |
640 | ||
de0428a7 | 641 | void reserve_ds_buffers(void) |
ca037701 | 642 | { |
6809b6ea PZ |
643 | int bts_err = 0, pebs_err = 0; |
644 | int cpu; | |
645 | ||
646 | x86_pmu.bts_active = 0; | |
647 | x86_pmu.pebs_active = 0; | |
ca037701 PZ |
648 | |
649 | if (!x86_pmu.bts && !x86_pmu.pebs) | |
f80c9e30 | 650 | return; |
ca037701 | 651 | |
6809b6ea PZ |
652 | if (!x86_pmu.bts) |
653 | bts_err = 1; | |
654 | ||
655 | if (!x86_pmu.pebs) | |
656 | pebs_err = 1; | |
657 | ||
ca037701 | 658 | for_each_possible_cpu(cpu) { |
6809b6ea PZ |
659 | if (alloc_ds_buffer(cpu)) { |
660 | bts_err = 1; | |
661 | pebs_err = 1; | |
662 | } | |
ca037701 | 663 | |
6809b6ea PZ |
664 | if (!bts_err && alloc_bts_buffer(cpu)) |
665 | bts_err = 1; | |
666 | ||
667 | if (!pebs_err && alloc_pebs_buffer(cpu)) | |
668 | pebs_err = 1; | |
5ee25c87 | 669 | |
6809b6ea | 670 | if (bts_err && pebs_err) |
5ee25c87 | 671 | break; |
6809b6ea PZ |
672 | } |
673 | ||
674 | if (bts_err) { | |
675 | for_each_possible_cpu(cpu) | |
676 | release_bts_buffer(cpu); | |
677 | } | |
ca037701 | 678 | |
6809b6ea PZ |
679 | if (pebs_err) { |
680 | for_each_possible_cpu(cpu) | |
681 | release_pebs_buffer(cpu); | |
ca037701 PZ |
682 | } |
683 | ||
6809b6ea PZ |
684 | if (bts_err && pebs_err) { |
685 | for_each_possible_cpu(cpu) | |
686 | release_ds_buffer(cpu); | |
687 | } else { | |
688 | if (x86_pmu.bts && !bts_err) | |
689 | x86_pmu.bts_active = 1; | |
690 | ||
691 | if (x86_pmu.pebs && !pebs_err) | |
692 | x86_pmu.pebs_active = 1; | |
693 | ||
efe951d3 PZ |
694 | for_each_possible_cpu(cpu) { |
695 | /* | |
696 | * Ignores wrmsr_on_cpu() errors for offline CPUs they | |
697 | * will get this call through intel_pmu_cpu_starting(). | |
698 | */ | |
ca037701 | 699 | init_debug_store_on_cpu(cpu); |
efe951d3 | 700 | } |
ca037701 | 701 | } |
ca037701 PZ |
702 | } |
703 | ||
704 | /* | |
705 | * BTS | |
706 | */ | |
707 | ||
de0428a7 | 708 | struct event_constraint bts_constraint = |
15c7ad51 | 709 | EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0); |
ca037701 | 710 | |
de0428a7 | 711 | void intel_pmu_enable_bts(u64 config) |
ca037701 PZ |
712 | { |
713 | unsigned long debugctlmsr; | |
714 | ||
715 | debugctlmsr = get_debugctlmsr(); | |
716 | ||
7c5ecaf7 PZ |
717 | debugctlmsr |= DEBUGCTLMSR_TR; |
718 | debugctlmsr |= DEBUGCTLMSR_BTS; | |
8062382c AS |
719 | if (config & ARCH_PERFMON_EVENTSEL_INT) |
720 | debugctlmsr |= DEBUGCTLMSR_BTINT; | |
ca037701 PZ |
721 | |
722 | if (!(config & ARCH_PERFMON_EVENTSEL_OS)) | |
7c5ecaf7 | 723 | debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS; |
ca037701 PZ |
724 | |
725 | if (!(config & ARCH_PERFMON_EVENTSEL_USR)) | |
7c5ecaf7 | 726 | debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR; |
ca037701 PZ |
727 | |
728 | update_debugctlmsr(debugctlmsr); | |
729 | } | |
730 | ||
de0428a7 | 731 | void intel_pmu_disable_bts(void) |
ca037701 | 732 | { |
89cbc767 | 733 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
ca037701 PZ |
734 | unsigned long debugctlmsr; |
735 | ||
736 | if (!cpuc->ds) | |
737 | return; | |
738 | ||
739 | debugctlmsr = get_debugctlmsr(); | |
740 | ||
741 | debugctlmsr &= | |
7c5ecaf7 PZ |
742 | ~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT | |
743 | DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR); | |
ca037701 PZ |
744 | |
745 | update_debugctlmsr(debugctlmsr); | |
746 | } | |
747 | ||
de0428a7 | 748 | int intel_pmu_drain_bts_buffer(void) |
ca037701 | 749 | { |
89cbc767 | 750 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
ca037701 PZ |
751 | struct debug_store *ds = cpuc->ds; |
752 | struct bts_record { | |
753 | u64 from; | |
754 | u64 to; | |
755 | u64 flags; | |
756 | }; | |
15c7ad51 | 757 | struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS]; |
a09d31f4 | 758 | struct bts_record *at, *base, *top; |
ca037701 PZ |
759 | struct perf_output_handle handle; |
760 | struct perf_event_header header; | |
761 | struct perf_sample_data data; | |
a09d31f4 | 762 | unsigned long skip = 0; |
ca037701 PZ |
763 | struct pt_regs regs; |
764 | ||
765 | if (!event) | |
b0b2072d | 766 | return 0; |
ca037701 | 767 | |
6809b6ea | 768 | if (!x86_pmu.bts_active) |
b0b2072d | 769 | return 0; |
ca037701 | 770 | |
a09d31f4 AS |
771 | base = (struct bts_record *)(unsigned long)ds->bts_buffer_base; |
772 | top = (struct bts_record *)(unsigned long)ds->bts_index; | |
ca037701 | 773 | |
a09d31f4 | 774 | if (top <= base) |
b0b2072d | 775 | return 0; |
ca037701 | 776 | |
0e48026a SE |
777 | memset(®s, 0, sizeof(regs)); |
778 | ||
ca037701 PZ |
779 | ds->bts_index = ds->bts_buffer_base; |
780 | ||
fd0d000b | 781 | perf_sample_data_init(&data, 0, event->hw.last_period); |
ca037701 | 782 | |
a09d31f4 AS |
783 | /* |
784 | * BTS leaks kernel addresses in branches across the cpl boundary, | |
785 | * such as traps or system calls, so unless the user is asking for | |
786 | * kernel tracing (and right now it's not possible), we'd need to | |
787 | * filter them out. But first we need to count how many of those we | |
788 | * have in the current batch. This is an extra O(n) pass, however, | |
789 | * it's much faster than the other one especially considering that | |
790 | * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the | |
791 | * alloc_bts_buffer()). | |
792 | */ | |
793 | for (at = base; at < top; at++) { | |
794 | /* | |
795 | * Note that right now *this* BTS code only works if | |
796 | * attr::exclude_kernel is set, but let's keep this extra | |
797 | * check here in case that changes. | |
798 | */ | |
799 | if (event->attr.exclude_kernel && | |
800 | (kernel_ip(at->from) || kernel_ip(at->to))) | |
801 | skip++; | |
802 | } | |
803 | ||
ca037701 PZ |
804 | /* |
805 | * Prepare a generic sample, i.e. fill in the invariant fields. | |
806 | * We will overwrite the from and to address before we output | |
807 | * the sample. | |
808 | */ | |
e8d8a90f | 809 | rcu_read_lock(); |
ca037701 PZ |
810 | perf_prepare_sample(&header, &data, event, ®s); |
811 | ||
267fb273 PZ |
812 | if (perf_output_begin(&handle, &data, event, |
813 | header.size * (top - base - skip))) | |
e8d8a90f | 814 | goto unlock; |
ca037701 | 815 | |
a09d31f4 AS |
816 | for (at = base; at < top; at++) { |
817 | /* Filter out any records that contain kernel addresses. */ | |
818 | if (event->attr.exclude_kernel && | |
819 | (kernel_ip(at->from) || kernel_ip(at->to))) | |
820 | continue; | |
821 | ||
ca037701 PZ |
822 | data.ip = at->from; |
823 | data.addr = at->to; | |
824 | ||
825 | perf_output_sample(&handle, &header, &data, event); | |
826 | } | |
827 | ||
828 | perf_output_end(&handle); | |
829 | ||
830 | /* There's new data available. */ | |
831 | event->hw.interrupts++; | |
832 | event->pending_kill = POLL_IN; | |
e8d8a90f PZ |
833 | unlock: |
834 | rcu_read_unlock(); | |
b0b2072d | 835 | return 1; |
ca037701 PZ |
836 | } |
837 | ||
9c964efa YZ |
838 | static inline void intel_pmu_drain_pebs_buffer(void) |
839 | { | |
9dfa9a5c PZ |
840 | struct perf_sample_data data; |
841 | ||
842 | x86_pmu.drain_pebs(NULL, &data); | |
9c964efa YZ |
843 | } |
844 | ||
ca037701 PZ |
845 | /* |
846 | * PEBS | |
847 | */ | |
de0428a7 | 848 | struct event_constraint intel_core2_pebs_event_constraints[] = { |
af4bdcf6 AK |
849 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */ |
850 | INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */ | |
851 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */ | |
852 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */ | |
853 | INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED.* */ | |
517e6341 | 854 | /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */ |
23e3983a | 855 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01), |
ca037701 PZ |
856 | EVENT_CONSTRAINT_END |
857 | }; | |
858 | ||
de0428a7 | 859 | struct event_constraint intel_atom_pebs_event_constraints[] = { |
af4bdcf6 AK |
860 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */ |
861 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */ | |
862 | INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED.* */ | |
517e6341 | 863 | /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */ |
23e3983a | 864 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01), |
673d188b SE |
865 | /* Allow all events as PEBS with no flags */ |
866 | INTEL_ALL_EVENT_CONSTRAINT(0, 0x1), | |
17e31629 SE |
867 | EVENT_CONSTRAINT_END |
868 | }; | |
869 | ||
1fa64180 | 870 | struct event_constraint intel_slm_pebs_event_constraints[] = { |
33636732 | 871 | /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */ |
23e3983a | 872 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x1), |
86a04461 AK |
873 | /* Allow all events as PEBS with no flags */ |
874 | INTEL_ALL_EVENT_CONSTRAINT(0, 0x1), | |
1fa64180 YZ |
875 | EVENT_CONSTRAINT_END |
876 | }; | |
877 | ||
8b92c3a7 KL |
878 | struct event_constraint intel_glm_pebs_event_constraints[] = { |
879 | /* Allow all events as PEBS with no flags */ | |
880 | INTEL_ALL_EVENT_CONSTRAINT(0, 0x1), | |
881 | EVENT_CONSTRAINT_END | |
882 | }; | |
883 | ||
f83d2f91 KL |
884 | struct event_constraint intel_grt_pebs_event_constraints[] = { |
885 | /* Allow all events as PEBS with no flags */ | |
cde643ff | 886 | INTEL_HYBRID_LAT_CONSTRAINT(0x5d0, 0x3), |
39a41278 | 887 | INTEL_HYBRID_LAT_CONSTRAINT(0x6d0, 0xf), |
f83d2f91 KL |
888 | EVENT_CONSTRAINT_END |
889 | }; | |
890 | ||
de0428a7 | 891 | struct event_constraint intel_nehalem_pebs_event_constraints[] = { |
f20093ee | 892 | INTEL_PLD_CONSTRAINT(0x100b, 0xf), /* MEM_INST_RETIRED.* */ |
af4bdcf6 AK |
893 | INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf), /* MEM_UNCORE_RETIRED.* */ |
894 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */ | |
895 | INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf), /* INST_RETIRED.ANY */ | |
7d5d02da | 896 | INTEL_EVENT_CONSTRAINT(0xc2, 0xf), /* UOPS_RETIRED.* */ |
af4bdcf6 AK |
897 | INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */ |
898 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */ | |
899 | INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf), /* SSEX_UOPS_RETIRED.* */ | |
900 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */ | |
901 | INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf), /* MEM_LOAD_RETIRED.* */ | |
902 | INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf), /* FP_ASSIST.* */ | |
517e6341 | 903 | /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */ |
23e3983a | 904 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f), |
17e31629 SE |
905 | EVENT_CONSTRAINT_END |
906 | }; | |
907 | ||
de0428a7 | 908 | struct event_constraint intel_westmere_pebs_event_constraints[] = { |
f20093ee | 909 | INTEL_PLD_CONSTRAINT(0x100b, 0xf), /* MEM_INST_RETIRED.* */ |
af4bdcf6 AK |
910 | INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf), /* MEM_UNCORE_RETIRED.* */ |
911 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */ | |
912 | INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf), /* INSTR_RETIRED.* */ | |
7d5d02da | 913 | INTEL_EVENT_CONSTRAINT(0xc2, 0xf), /* UOPS_RETIRED.* */ |
af4bdcf6 AK |
914 | INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */ |
915 | INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf), /* BR_MISP_RETIRED.* */ | |
916 | INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf), /* SSEX_UOPS_RETIRED.* */ | |
917 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */ | |
918 | INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf), /* MEM_LOAD_RETIRED.* */ | |
919 | INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf), /* FP_ASSIST.* */ | |
517e6341 | 920 | /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */ |
23e3983a | 921 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f), |
ca037701 PZ |
922 | EVENT_CONSTRAINT_END |
923 | }; | |
924 | ||
de0428a7 | 925 | struct event_constraint intel_snb_pebs_event_constraints[] = { |
0dbc9479 | 926 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */ |
f20093ee | 927 | INTEL_PLD_CONSTRAINT(0x01cd, 0x8), /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */ |
9ad64c0f | 928 | INTEL_PST_CONSTRAINT(0x02cd, 0x8), /* MEM_TRANS_RETIRED.PRECISE_STORES */ |
86a04461 | 929 | /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */ |
23e3983a | 930 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf), |
b63b4b45 MD |
931 | INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOP_RETIRED.* */ |
932 | INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ | |
933 | INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ | |
934 | INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ | |
86a04461 AK |
935 | /* Allow all events as PEBS with no flags */ |
936 | INTEL_ALL_EVENT_CONSTRAINT(0, 0xf), | |
b06b3d49 LM |
937 | EVENT_CONSTRAINT_END |
938 | }; | |
939 | ||
20a36e39 | 940 | struct event_constraint intel_ivb_pebs_event_constraints[] = { |
0dbc9479 | 941 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */ |
f20093ee | 942 | INTEL_PLD_CONSTRAINT(0x01cd, 0x8), /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */ |
9ad64c0f | 943 | INTEL_PST_CONSTRAINT(0x02cd, 0x8), /* MEM_TRANS_RETIRED.PRECISE_STORES */ |
86a04461 | 944 | /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */ |
23e3983a | 945 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf), |
72469764 | 946 | /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */ |
23e3983a | 947 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2), |
b63b4b45 MD |
948 | INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOP_RETIRED.* */ |
949 | INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ | |
950 | INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ | |
951 | INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ | |
86a04461 AK |
952 | /* Allow all events as PEBS with no flags */ |
953 | INTEL_ALL_EVENT_CONSTRAINT(0, 0xf), | |
20a36e39 SE |
954 | EVENT_CONSTRAINT_END |
955 | }; | |
956 | ||
3044318f | 957 | struct event_constraint intel_hsw_pebs_event_constraints[] = { |
0dbc9479 | 958 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */ |
86a04461 AK |
959 | INTEL_PLD_CONSTRAINT(0x01cd, 0xf), /* MEM_TRANS_RETIRED.* */ |
960 | /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */ | |
23e3983a | 961 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf), |
72469764 | 962 | /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */ |
23e3983a | 963 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2), |
86a04461 | 964 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */ |
b63b4b45 MD |
965 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */ |
966 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */ | |
967 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */ | |
968 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */ | |
969 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */ | |
970 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */ | |
971 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */ | |
972 | INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ | |
973 | INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf), /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */ | |
974 | INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf), /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */ | |
9a92e16f AK |
975 | /* Allow all events as PEBS with no flags */ |
976 | INTEL_ALL_EVENT_CONSTRAINT(0, 0xf), | |
977 | EVENT_CONSTRAINT_END | |
978 | }; | |
979 | ||
b3e62463 SE |
980 | struct event_constraint intel_bdw_pebs_event_constraints[] = { |
981 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */ | |
982 | INTEL_PLD_CONSTRAINT(0x01cd, 0xf), /* MEM_TRANS_RETIRED.* */ | |
983 | /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */ | |
23e3983a | 984 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf), |
b3e62463 | 985 | /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */ |
23e3983a | 986 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2), |
b3e62463 SE |
987 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */ |
988 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */ | |
989 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */ | |
990 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */ | |
991 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */ | |
992 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */ | |
993 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */ | |
994 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */ | |
995 | INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ | |
996 | INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf), /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */ | |
997 | INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf), /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */ | |
998 | /* Allow all events as PEBS with no flags */ | |
999 | INTEL_ALL_EVENT_CONSTRAINT(0, 0xf), | |
1000 | EVENT_CONSTRAINT_END | |
1001 | }; | |
1002 | ||
1003 | ||
9a92e16f AK |
1004 | struct event_constraint intel_skl_pebs_event_constraints[] = { |
1005 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2), /* INST_RETIRED.PREC_DIST */ | |
72469764 | 1006 | /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */ |
23e3983a | 1007 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2), |
442f5c74 | 1008 | /* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */ |
23e3983a | 1009 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f), |
9a92e16f AK |
1010 | INTEL_PLD_CONSTRAINT(0x1cd, 0xf), /* MEM_TRANS_RETIRED.* */ |
1011 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */ | |
1012 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */ | |
1013 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */ | |
1014 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */ | |
1015 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */ | |
1016 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */ | |
1017 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */ | |
1018 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */ | |
1019 | INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */ | |
1020 | INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */ | |
1021 | INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf), /* MEM_LOAD_L3_MISS_RETIRED.* */ | |
86a04461 AK |
1022 | /* Allow all events as PEBS with no flags */ |
1023 | INTEL_ALL_EVENT_CONSTRAINT(0, 0xf), | |
3044318f AK |
1024 | EVENT_CONSTRAINT_END |
1025 | }; | |
1026 | ||
60176089 | 1027 | struct event_constraint intel_icl_pebs_event_constraints[] = { |
2de71ee1 SE |
1028 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x100000000ULL), /* old INST_RETIRED.PREC_DIST */ |
1029 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x0100, 0x100000000ULL), /* INST_RETIRED.PREC_DIST */ | |
3d0c3953 | 1030 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL), /* SLOTS */ |
60176089 KL |
1031 | |
1032 | INTEL_PLD_CONSTRAINT(0x1cd, 0xff), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ | |
acc5568b KL |
1033 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */ |
1034 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */ | |
1035 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */ | |
1036 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */ | |
1037 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */ | |
1038 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */ | |
1039 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */ | |
60176089 KL |
1040 | |
1041 | INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf), /* MEM_LOAD_*_RETIRED.* */ | |
1042 | ||
1043 | INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */ | |
1044 | ||
1045 | /* | |
1046 | * Everything else is handled by PMU_FL_PEBS_ALL, because we | |
1047 | * need the full constraints from the main table. | |
1048 | */ | |
1049 | ||
1050 | EVENT_CONSTRAINT_END | |
1051 | }; | |
1052 | ||
61b985e3 | 1053 | struct event_constraint intel_spr_pebs_event_constraints[] = { |
2de71ee1 | 1054 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x100, 0x100000000ULL), /* INST_RETIRED.PREC_DIST */ |
61b985e3 KL |
1055 | INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL), |
1056 | ||
1057 | INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xfe), | |
1058 | INTEL_PLD_CONSTRAINT(0x1cd, 0xfe), | |
1059 | INTEL_PSD_CONSTRAINT(0x2cd, 0x1), | |
0916886b KL |
1060 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */ |
1061 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */ | |
1062 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */ | |
1063 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */ | |
1064 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */ | |
1065 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */ | |
1066 | INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */ | |
61b985e3 KL |
1067 | |
1068 | INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf), | |
1069 | ||
1070 | INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf), | |
1071 | ||
1072 | /* | |
1073 | * Everything else is handled by PMU_FL_PEBS_ALL, because we | |
1074 | * need the full constraints from the main table. | |
1075 | */ | |
1076 | ||
1077 | EVENT_CONSTRAINT_END | |
1078 | }; | |
1079 | ||
de0428a7 | 1080 | struct event_constraint *intel_pebs_constraints(struct perf_event *event) |
ca037701 | 1081 | { |
24ee38ff | 1082 | struct event_constraint *pebs_constraints = hybrid(event->pmu, pebs_constraints); |
ca037701 PZ |
1083 | struct event_constraint *c; |
1084 | ||
ab608344 | 1085 | if (!event->attr.precise_ip) |
ca037701 PZ |
1086 | return NULL; |
1087 | ||
24ee38ff KL |
1088 | if (pebs_constraints) { |
1089 | for_each_event_constraint(c, pebs_constraints) { | |
63b79f6e | 1090 | if (constraint_match(c, event->hw.config)) { |
9fac2cf3 | 1091 | event->hw.flags |= c->flags; |
ca037701 | 1092 | return c; |
9fac2cf3 | 1093 | } |
ca037701 PZ |
1094 | } |
1095 | } | |
1096 | ||
31962340 KL |
1097 | /* |
1098 | * Extended PEBS support | |
1099 | * Makes the PEBS code search the normal constraints. | |
1100 | */ | |
1101 | if (x86_pmu.flags & PMU_FL_PEBS_ALL) | |
1102 | return NULL; | |
1103 | ||
ca037701 PZ |
1104 | return &emptyconstraint; |
1105 | } | |
1106 | ||
09e61b4f PZ |
1107 | /* |
1108 | * We need the sched_task callback even for per-cpu events when we use | |
1109 | * the large interrupt threshold, such that we can provide PID and TID | |
1110 | * to PEBS samples. | |
1111 | */ | |
1112 | static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc) | |
1113 | { | |
42880f72 AS |
1114 | if (cpuc->n_pebs == cpuc->n_pebs_via_pt) |
1115 | return false; | |
1116 | ||
09e61b4f PZ |
1117 | return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs); |
1118 | } | |
1119 | ||
bd275681 | 1120 | void intel_pmu_pebs_sched_task(struct perf_event_pmu_context *pmu_ctx, bool sched_in) |
df6c3db8 JO |
1121 | { |
1122 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); | |
1123 | ||
1124 | if (!sched_in && pebs_needs_sched_cb(cpuc)) | |
1125 | intel_pmu_drain_pebs_buffer(); | |
1126 | } | |
1127 | ||
09e61b4f PZ |
1128 | static inline void pebs_update_threshold(struct cpu_hw_events *cpuc) |
1129 | { | |
1130 | struct debug_store *ds = cpuc->ds; | |
d4b294bf KL |
1131 | int max_pebs_events = hybrid(cpuc->pmu, max_pebs_events); |
1132 | int num_counters_fixed = hybrid(cpuc->pmu, num_counters_fixed); | |
09e61b4f | 1133 | u64 threshold; |
ec71a398 KL |
1134 | int reserved; |
1135 | ||
42880f72 AS |
1136 | if (cpuc->n_pebs_via_pt) |
1137 | return; | |
1138 | ||
ec71a398 | 1139 | if (x86_pmu.flags & PMU_FL_PEBS_ALL) |
d4b294bf | 1140 | reserved = max_pebs_events + num_counters_fixed; |
ec71a398 | 1141 | else |
d4b294bf | 1142 | reserved = max_pebs_events; |
09e61b4f PZ |
1143 | |
1144 | if (cpuc->n_pebs == cpuc->n_large_pebs) { | |
1145 | threshold = ds->pebs_absolute_maximum - | |
c22497f5 | 1146 | reserved * cpuc->pebs_record_size; |
09e61b4f | 1147 | } else { |
c22497f5 | 1148 | threshold = ds->pebs_buffer_base + cpuc->pebs_record_size; |
09e61b4f PZ |
1149 | } |
1150 | ||
1151 | ds->pebs_interrupt_threshold = threshold; | |
1152 | } | |
1153 | ||
c22497f5 KL |
1154 | static void adaptive_pebs_record_size_update(void) |
1155 | { | |
1156 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); | |
1157 | u64 pebs_data_cfg = cpuc->pebs_data_cfg; | |
1158 | int sz = sizeof(struct pebs_basic); | |
1159 | ||
1160 | if (pebs_data_cfg & PEBS_DATACFG_MEMINFO) | |
1161 | sz += sizeof(struct pebs_meminfo); | |
1162 | if (pebs_data_cfg & PEBS_DATACFG_GP) | |
1163 | sz += sizeof(struct pebs_gprs); | |
1164 | if (pebs_data_cfg & PEBS_DATACFG_XMMS) | |
1165 | sz += sizeof(struct pebs_xmm); | |
1166 | if (pebs_data_cfg & PEBS_DATACFG_LBRS) | |
5624986d | 1167 | sz += x86_pmu.lbr_nr * sizeof(struct lbr_entry); |
c22497f5 KL |
1168 | |
1169 | cpuc->pebs_record_size = sz; | |
1170 | } | |
1171 | ||
1172 | #define PERF_PEBS_MEMINFO_TYPE (PERF_SAMPLE_ADDR | PERF_SAMPLE_DATA_SRC | \ | |
2a6c6b7d KL |
1173 | PERF_SAMPLE_PHYS_ADDR | \ |
1174 | PERF_SAMPLE_WEIGHT_TYPE | \ | |
76a5433f KL |
1175 | PERF_SAMPLE_TRANSACTION | \ |
1176 | PERF_SAMPLE_DATA_PAGE_SIZE) | |
c22497f5 KL |
1177 | |
1178 | static u64 pebs_update_adaptive_cfg(struct perf_event *event) | |
1179 | { | |
1180 | struct perf_event_attr *attr = &event->attr; | |
1181 | u64 sample_type = attr->sample_type; | |
1182 | u64 pebs_data_cfg = 0; | |
1183 | bool gprs, tsx_weight; | |
1184 | ||
1185 | if (!(sample_type & ~(PERF_SAMPLE_IP|PERF_SAMPLE_TIME)) && | |
1186 | attr->precise_ip > 1) | |
1187 | return pebs_data_cfg; | |
1188 | ||
1189 | if (sample_type & PERF_PEBS_MEMINFO_TYPE) | |
1190 | pebs_data_cfg |= PEBS_DATACFG_MEMINFO; | |
1191 | ||
1192 | /* | |
1193 | * We need GPRs when: | |
1194 | * + user requested them | |
1195 | * + precise_ip < 2 for the non event IP | |
1196 | * + For RTM TSX weight we need GPRs for the abort code. | |
1197 | */ | |
1198 | gprs = (sample_type & PERF_SAMPLE_REGS_INTR) && | |
1199 | (attr->sample_regs_intr & PEBS_GP_REGS); | |
1200 | ||
2a6c6b7d | 1201 | tsx_weight = (sample_type & PERF_SAMPLE_WEIGHT_TYPE) && |
c22497f5 KL |
1202 | ((attr->config & INTEL_ARCH_EVENT_MASK) == |
1203 | x86_pmu.rtm_abort_event); | |
1204 | ||
1205 | if (gprs || (attr->precise_ip < 2) || tsx_weight) | |
1206 | pebs_data_cfg |= PEBS_DATACFG_GP; | |
1207 | ||
1208 | if ((sample_type & PERF_SAMPLE_REGS_INTR) && | |
dce86ac7 | 1209 | (attr->sample_regs_intr & PERF_REG_EXTENDED_MASK)) |
c22497f5 KL |
1210 | pebs_data_cfg |= PEBS_DATACFG_XMMS; |
1211 | ||
1212 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
1213 | /* | |
1214 | * For now always log all LBRs. Could configure this | |
1215 | * later. | |
1216 | */ | |
1217 | pebs_data_cfg |= PEBS_DATACFG_LBRS | | |
1218 | ((x86_pmu.lbr_nr-1) << PEBS_DATACFG_LBR_SHIFT); | |
1219 | } | |
1220 | ||
1221 | return pebs_data_cfg; | |
1222 | } | |
1223 | ||
09e61b4f | 1224 | static void |
c22497f5 KL |
1225 | pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc, |
1226 | struct perf_event *event, bool add) | |
09e61b4f | 1227 | { |
bd275681 | 1228 | struct pmu *pmu = event->pmu; |
b6a32f02 JO |
1229 | /* |
1230 | * Make sure we get updated with the first PEBS | |
1231 | * event. It will trigger also during removal, but | |
1232 | * that does not hurt: | |
1233 | */ | |
1234 | bool update = cpuc->n_pebs == 1; | |
1235 | ||
09e61b4f PZ |
1236 | if (needed_cb != pebs_needs_sched_cb(cpuc)) { |
1237 | if (!needed_cb) | |
1238 | perf_sched_cb_inc(pmu); | |
1239 | else | |
1240 | perf_sched_cb_dec(pmu); | |
1241 | ||
b6a32f02 | 1242 | update = true; |
09e61b4f | 1243 | } |
b6a32f02 | 1244 | |
c22497f5 KL |
1245 | /* |
1246 | * The PEBS record doesn't shrink on pmu::del(). Doing so would require | |
1247 | * iterating all remaining PEBS events to reconstruct the config. | |
1248 | */ | |
1249 | if (x86_pmu.intel_cap.pebs_baseline && add) { | |
1250 | u64 pebs_data_cfg; | |
1251 | ||
1252 | /* Clear pebs_data_cfg and pebs_record_size for first PEBS. */ | |
1253 | if (cpuc->n_pebs == 1) { | |
1254 | cpuc->pebs_data_cfg = 0; | |
1255 | cpuc->pebs_record_size = sizeof(struct pebs_basic); | |
1256 | } | |
1257 | ||
1258 | pebs_data_cfg = pebs_update_adaptive_cfg(event); | |
1259 | ||
1260 | /* Update pebs_record_size if new event requires more data. */ | |
1261 | if (pebs_data_cfg & ~cpuc->pebs_data_cfg) { | |
1262 | cpuc->pebs_data_cfg |= pebs_data_cfg; | |
1263 | adaptive_pebs_record_size_update(); | |
1264 | update = true; | |
1265 | } | |
1266 | } | |
1267 | ||
b6a32f02 JO |
1268 | if (update) |
1269 | pebs_update_threshold(cpuc); | |
09e61b4f PZ |
1270 | } |
1271 | ||
68f7082f | 1272 | void intel_pmu_pebs_add(struct perf_event *event) |
3569c0d7 | 1273 | { |
09e61b4f PZ |
1274 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
1275 | struct hw_perf_event *hwc = &event->hw; | |
1276 | bool needed_cb = pebs_needs_sched_cb(cpuc); | |
1277 | ||
1278 | cpuc->n_pebs++; | |
174afc3e | 1279 | if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS) |
09e61b4f | 1280 | cpuc->n_large_pebs++; |
42880f72 AS |
1281 | if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT) |
1282 | cpuc->n_pebs_via_pt++; | |
09e61b4f | 1283 | |
c22497f5 | 1284 | pebs_update_state(needed_cb, cpuc, event, true); |
3569c0d7 YZ |
1285 | } |
1286 | ||
42880f72 AS |
1287 | static void intel_pmu_pebs_via_pt_disable(struct perf_event *event) |
1288 | { | |
1289 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); | |
1290 | ||
1291 | if (!is_pebs_pt(event)) | |
1292 | return; | |
1293 | ||
1294 | if (!(cpuc->pebs_enabled & ~PEBS_VIA_PT_MASK)) | |
1295 | cpuc->pebs_enabled &= ~PEBS_VIA_PT_MASK; | |
1296 | } | |
1297 | ||
1298 | static void intel_pmu_pebs_via_pt_enable(struct perf_event *event) | |
1299 | { | |
1300 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); | |
1301 | struct hw_perf_event *hwc = &event->hw; | |
1302 | struct debug_store *ds = cpuc->ds; | |
4c58d922 LX |
1303 | u64 value = ds->pebs_event_reset[hwc->idx]; |
1304 | u32 base = MSR_RELOAD_PMC0; | |
1305 | unsigned int idx = hwc->idx; | |
42880f72 AS |
1306 | |
1307 | if (!is_pebs_pt(event)) | |
1308 | return; | |
1309 | ||
1310 | if (!(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS)) | |
1311 | cpuc->pebs_enabled |= PEBS_PMI_AFTER_EACH_RECORD; | |
1312 | ||
1313 | cpuc->pebs_enabled |= PEBS_OUTPUT_PT; | |
1314 | ||
4c58d922 LX |
1315 | if (hwc->idx >= INTEL_PMC_IDX_FIXED) { |
1316 | base = MSR_RELOAD_FIXED_CTR0; | |
1317 | idx = hwc->idx - INTEL_PMC_IDX_FIXED; | |
2145e77f KL |
1318 | if (x86_pmu.intel_cap.pebs_format < 5) |
1319 | value = ds->pebs_event_reset[MAX_PEBS_EVENTS_FMT4 + idx]; | |
1320 | else | |
1321 | value = ds->pebs_event_reset[MAX_PEBS_EVENTS + idx]; | |
4c58d922 LX |
1322 | } |
1323 | wrmsrl(base + idx, value); | |
42880f72 AS |
1324 | } |
1325 | ||
de0428a7 | 1326 | void intel_pmu_pebs_enable(struct perf_event *event) |
ca037701 | 1327 | { |
89cbc767 | 1328 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
ef21f683 | 1329 | struct hw_perf_event *hwc = &event->hw; |
851559e3 | 1330 | struct debug_store *ds = cpuc->ds; |
4c58d922 | 1331 | unsigned int idx = hwc->idx; |
09e61b4f | 1332 | |
ca037701 PZ |
1333 | hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT; |
1334 | ||
ad0e6cfe | 1335 | cpuc->pebs_enabled |= 1ULL << hwc->idx; |
f20093ee | 1336 | |
60176089 | 1337 | if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) && (x86_pmu.version < 5)) |
f20093ee | 1338 | cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32); |
9ad64c0f SE |
1339 | else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST) |
1340 | cpuc->pebs_enabled |= 1ULL << 63; | |
851559e3 | 1341 | |
c22497f5 KL |
1342 | if (x86_pmu.intel_cap.pebs_baseline) { |
1343 | hwc->config |= ICL_EVENTSEL_ADAPTIVE; | |
1344 | if (cpuc->pebs_data_cfg != cpuc->active_pebs_data_cfg) { | |
1345 | wrmsrl(MSR_PEBS_DATA_CFG, cpuc->pebs_data_cfg); | |
1346 | cpuc->active_pebs_data_cfg = cpuc->pebs_data_cfg; | |
1347 | } | |
1348 | } | |
1349 | ||
2145e77f KL |
1350 | if (idx >= INTEL_PMC_IDX_FIXED) { |
1351 | if (x86_pmu.intel_cap.pebs_format < 5) | |
1352 | idx = MAX_PEBS_EVENTS_FMT4 + (idx - INTEL_PMC_IDX_FIXED); | |
1353 | else | |
1354 | idx = MAX_PEBS_EVENTS + (idx - INTEL_PMC_IDX_FIXED); | |
1355 | } | |
4c58d922 | 1356 | |
3569c0d7 | 1357 | /* |
09e61b4f PZ |
1358 | * Use auto-reload if possible to save a MSR write in the PMI. |
1359 | * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD. | |
3569c0d7 | 1360 | */ |
851559e3 | 1361 | if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) { |
ec71a398 | 1362 | ds->pebs_event_reset[idx] = |
851559e3 | 1363 | (u64)(-hwc->sample_period) & x86_pmu.cntval_mask; |
dc853e26 | 1364 | } else { |
4c58d922 | 1365 | ds->pebs_event_reset[idx] = 0; |
851559e3 | 1366 | } |
42880f72 AS |
1367 | |
1368 | intel_pmu_pebs_via_pt_enable(event); | |
09e61b4f PZ |
1369 | } |
1370 | ||
68f7082f | 1371 | void intel_pmu_pebs_del(struct perf_event *event) |
09e61b4f PZ |
1372 | { |
1373 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); | |
1374 | struct hw_perf_event *hwc = &event->hw; | |
1375 | bool needed_cb = pebs_needs_sched_cb(cpuc); | |
1376 | ||
1377 | cpuc->n_pebs--; | |
174afc3e | 1378 | if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS) |
09e61b4f | 1379 | cpuc->n_large_pebs--; |
42880f72 AS |
1380 | if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT) |
1381 | cpuc->n_pebs_via_pt--; | |
3569c0d7 | 1382 | |
c22497f5 | 1383 | pebs_update_state(needed_cb, cpuc, event, false); |
ca037701 PZ |
1384 | } |
1385 | ||
de0428a7 | 1386 | void intel_pmu_pebs_disable(struct perf_event *event) |
ca037701 | 1387 | { |
89cbc767 | 1388 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
ef21f683 | 1389 | struct hw_perf_event *hwc = &event->hw; |
2a853e11 | 1390 | |
42880f72 AS |
1391 | if (cpuc->n_pebs == cpuc->n_large_pebs && |
1392 | cpuc->n_pebs != cpuc->n_pebs_via_pt) | |
2a853e11 | 1393 | intel_pmu_drain_pebs_buffer(); |
ca037701 | 1394 | |
ad0e6cfe | 1395 | cpuc->pebs_enabled &= ~(1ULL << hwc->idx); |
983433b5 | 1396 | |
60176089 KL |
1397 | if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) && |
1398 | (x86_pmu.version < 5)) | |
983433b5 | 1399 | cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32)); |
b371b594 | 1400 | else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST) |
983433b5 SE |
1401 | cpuc->pebs_enabled &= ~(1ULL << 63); |
1402 | ||
42880f72 AS |
1403 | intel_pmu_pebs_via_pt_disable(event); |
1404 | ||
4807e3d5 | 1405 | if (cpuc->enabled) |
ad0e6cfe | 1406 | wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled); |
ca037701 PZ |
1407 | |
1408 | hwc->config |= ARCH_PERFMON_EVENTSEL_INT; | |
1409 | } | |
1410 | ||
de0428a7 | 1411 | void intel_pmu_pebs_enable_all(void) |
ca037701 | 1412 | { |
89cbc767 | 1413 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
ca037701 PZ |
1414 | |
1415 | if (cpuc->pebs_enabled) | |
1416 | wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled); | |
1417 | } | |
1418 | ||
de0428a7 | 1419 | void intel_pmu_pebs_disable_all(void) |
ca037701 | 1420 | { |
89cbc767 | 1421 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
ca037701 PZ |
1422 | |
1423 | if (cpuc->pebs_enabled) | |
c22ac2a3 | 1424 | __intel_pmu_pebs_disable_all(); |
ca037701 PZ |
1425 | } |
1426 | ||
ef21f683 PZ |
1427 | static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs) |
1428 | { | |
89cbc767 | 1429 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
ef21f683 PZ |
1430 | unsigned long from = cpuc->lbr_entries[0].from; |
1431 | unsigned long old_to, to = cpuc->lbr_entries[0].to; | |
1432 | unsigned long ip = regs->ip; | |
57d1c0c0 | 1433 | int is_64bit = 0; |
9536c8d2 | 1434 | void *kaddr; |
6ba48ff4 | 1435 | int size; |
ef21f683 | 1436 | |
8db909a7 PZ |
1437 | /* |
1438 | * We don't need to fixup if the PEBS assist is fault like | |
1439 | */ | |
1440 | if (!x86_pmu.intel_cap.pebs_trap) | |
1441 | return 1; | |
1442 | ||
a562b187 PZ |
1443 | /* |
1444 | * No LBR entry, no basic block, no rewinding | |
1445 | */ | |
ef21f683 PZ |
1446 | if (!cpuc->lbr_stack.nr || !from || !to) |
1447 | return 0; | |
1448 | ||
a562b187 PZ |
1449 | /* |
1450 | * Basic blocks should never cross user/kernel boundaries | |
1451 | */ | |
1452 | if (kernel_ip(ip) != kernel_ip(to)) | |
1453 | return 0; | |
1454 | ||
1455 | /* | |
1456 | * unsigned math, either ip is before the start (impossible) or | |
1457 | * the basic block is larger than 1 page (sanity) | |
1458 | */ | |
9536c8d2 | 1459 | if ((ip - to) > PEBS_FIXUP_SIZE) |
ef21f683 PZ |
1460 | return 0; |
1461 | ||
1462 | /* | |
1463 | * We sampled a branch insn, rewind using the LBR stack | |
1464 | */ | |
1465 | if (ip == to) { | |
d07bdfd3 | 1466 | set_linear_ip(regs, from); |
ef21f683 PZ |
1467 | return 1; |
1468 | } | |
1469 | ||
6ba48ff4 | 1470 | size = ip - to; |
9536c8d2 | 1471 | if (!kernel_ip(ip)) { |
6ba48ff4 | 1472 | int bytes; |
9536c8d2 PZ |
1473 | u8 *buf = this_cpu_read(insn_buffer); |
1474 | ||
6ba48ff4 | 1475 | /* 'size' must fit our buffer, see above */ |
9536c8d2 | 1476 | bytes = copy_from_user_nmi(buf, (void __user *)to, size); |
0a196848 | 1477 | if (bytes != 0) |
9536c8d2 PZ |
1478 | return 0; |
1479 | ||
1480 | kaddr = buf; | |
1481 | } else { | |
1482 | kaddr = (void *)to; | |
1483 | } | |
1484 | ||
ef21f683 PZ |
1485 | do { |
1486 | struct insn insn; | |
ef21f683 PZ |
1487 | |
1488 | old_to = to; | |
ef21f683 | 1489 | |
57d1c0c0 | 1490 | #ifdef CONFIG_X86_64 |
375d4bfd | 1491 | is_64bit = kernel_ip(to) || any_64bit_mode(regs); |
57d1c0c0 | 1492 | #endif |
6ba48ff4 | 1493 | insn_init(&insn, kaddr, size, is_64bit); |
2ff49881 | 1494 | |
6ba48ff4 | 1495 | /* |
2ff49881 BP |
1496 | * Make sure there was not a problem decoding the instruction. |
1497 | * This is doubly important because we have an infinite loop if | |
1498 | * insn.length=0. | |
6ba48ff4 | 1499 | */ |
2ff49881 | 1500 | if (insn_get_length(&insn)) |
6ba48ff4 | 1501 | break; |
9536c8d2 | 1502 | |
ef21f683 | 1503 | to += insn.length; |
9536c8d2 | 1504 | kaddr += insn.length; |
6ba48ff4 | 1505 | size -= insn.length; |
ef21f683 PZ |
1506 | } while (to < ip); |
1507 | ||
1508 | if (to == ip) { | |
d07bdfd3 | 1509 | set_linear_ip(regs, old_to); |
ef21f683 PZ |
1510 | return 1; |
1511 | } | |
1512 | ||
a562b187 PZ |
1513 | /* |
1514 | * Even though we decoded the basic block, the instruction stream | |
1515 | * never matched the given IP, either the TO or the IP got corrupted. | |
1516 | */ | |
ef21f683 PZ |
1517 | return 0; |
1518 | } | |
1519 | ||
48f38aa4 | 1520 | static inline u64 intel_get_tsx_weight(u64 tsx_tuning) |
748e86aa | 1521 | { |
48f38aa4 AK |
1522 | if (tsx_tuning) { |
1523 | union hsw_tsx_tuning tsx = { .value = tsx_tuning }; | |
748e86aa AK |
1524 | return tsx.cycles_last_block; |
1525 | } | |
1526 | return 0; | |
1527 | } | |
1528 | ||
48f38aa4 | 1529 | static inline u64 intel_get_tsx_transaction(u64 tsx_tuning, u64 ax) |
a405bad5 | 1530 | { |
48f38aa4 | 1531 | u64 txn = (tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32; |
a405bad5 AK |
1532 | |
1533 | /* For RTM XABORTs also log the abort code from AX */ | |
48f38aa4 AK |
1534 | if ((txn & PERF_TXN_TRANSACTION) && (ax & 1)) |
1535 | txn |= ((ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT; | |
a405bad5 AK |
1536 | return txn; |
1537 | } | |
1538 | ||
c22497f5 KL |
1539 | static inline u64 get_pebs_status(void *n) |
1540 | { | |
1541 | if (x86_pmu.intel_cap.pebs_format < 4) | |
1542 | return ((struct pebs_record_nhm *)n)->status; | |
1543 | return ((struct pebs_basic *)n)->applicable_counters; | |
1544 | } | |
1545 | ||
48f38aa4 AK |
1546 | #define PERF_X86_EVENT_PEBS_HSW_PREC \ |
1547 | (PERF_X86_EVENT_PEBS_ST_HSW | \ | |
1548 | PERF_X86_EVENT_PEBS_LD_HSW | \ | |
1549 | PERF_X86_EVENT_PEBS_NA_HSW) | |
1550 | ||
1551 | static u64 get_data_src(struct perf_event *event, u64 aux) | |
1552 | { | |
1553 | u64 val = PERF_MEM_NA; | |
1554 | int fl = event->hw.flags; | |
1555 | bool fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC); | |
1556 | ||
1557 | if (fl & PERF_X86_EVENT_PEBS_LDLAT) | |
ccf170e9 | 1558 | val = load_latency_data(event, aux); |
61b985e3 | 1559 | else if (fl & PERF_X86_EVENT_PEBS_STLAT) |
ccf170e9 | 1560 | val = store_latency_data(event, aux); |
39a41278 KL |
1561 | else if (fl & PERF_X86_EVENT_PEBS_LAT_HYBRID) |
1562 | val = x86_pmu.pebs_latency_data(event, aux); | |
48f38aa4 AK |
1563 | else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC)) |
1564 | val = precise_datala_hsw(event, aux); | |
1565 | else if (fst) | |
1566 | val = precise_store_data(aux); | |
1567 | return val; | |
1568 | } | |
1569 | ||
76a5433f KL |
1570 | #define PERF_SAMPLE_ADDR_TYPE (PERF_SAMPLE_ADDR | \ |
1571 | PERF_SAMPLE_PHYS_ADDR | \ | |
1572 | PERF_SAMPLE_DATA_PAGE_SIZE) | |
1573 | ||
c22497f5 | 1574 | static void setup_pebs_fixed_sample_data(struct perf_event *event, |
43cf7631 YZ |
1575 | struct pt_regs *iregs, void *__pebs, |
1576 | struct perf_sample_data *data, | |
1577 | struct pt_regs *regs) | |
2b0b5c6f PZ |
1578 | { |
1579 | /* | |
d2beea4a PZ |
1580 | * We cast to the biggest pebs_record but are careful not to |
1581 | * unconditionally access the 'extra' entries. | |
2b0b5c6f | 1582 | */ |
89cbc767 | 1583 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
2f7ebf2e | 1584 | struct pebs_record_skl *pebs = __pebs; |
f20093ee | 1585 | u64 sample_type; |
48f38aa4 | 1586 | int fll; |
2b0b5c6f | 1587 | |
21509084 YZ |
1588 | if (pebs == NULL) |
1589 | return; | |
1590 | ||
c8aab2e0 | 1591 | sample_type = event->attr.sample_type; |
48f38aa4 | 1592 | fll = event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT; |
f20093ee | 1593 | |
43cf7631 | 1594 | perf_sample_data_init(data, 0, event->hw.last_period); |
2b0b5c6f | 1595 | |
43cf7631 | 1596 | data->period = event->hw.last_period; |
f20093ee SE |
1597 | |
1598 | /* | |
c8aab2e0 | 1599 | * Use latency for weight (only avail with PEBS-LL) |
f20093ee | 1600 | */ |
2abe681d | 1601 | if (fll && (sample_type & PERF_SAMPLE_WEIGHT_TYPE)) { |
2a6c6b7d | 1602 | data->weight.full = pebs->lat; |
2abe681d KL |
1603 | data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE; |
1604 | } | |
c8aab2e0 SE |
1605 | |
1606 | /* | |
1607 | * data.data_src encodes the data source | |
1608 | */ | |
e16fd7f2 | 1609 | if (sample_type & PERF_SAMPLE_DATA_SRC) { |
48f38aa4 | 1610 | data->data_src.val = get_data_src(event, pebs->dse); |
e16fd7f2 KL |
1611 | data->sample_flags |= PERF_SAMPLE_DATA_SRC; |
1612 | } | |
f20093ee | 1613 | |
6cbc304f PZ |
1614 | /* |
1615 | * We must however always use iregs for the unwinder to stay sane; the | |
1616 | * record BP,SP,IP can point into thin air when the record is from a | |
a97673a1 | 1617 | * previous PMI context or an (I)RET happened between the record and |
6cbc304f PZ |
1618 | * PMI. |
1619 | */ | |
31046500 NK |
1620 | if (sample_type & PERF_SAMPLE_CALLCHAIN) |
1621 | perf_sample_save_callchain(data, event, iregs); | |
6cbc304f | 1622 | |
2b0b5c6f | 1623 | /* |
b8000586 PZ |
1624 | * We use the interrupt regs as a base because the PEBS record does not |
1625 | * contain a full regs set, specifically it seems to lack segment | |
1626 | * descriptors, which get used by things like user_mode(). | |
2b0b5c6f | 1627 | * |
b8000586 | 1628 | * In the simple case fix up only the IP for PERF_SAMPLE_IP. |
2b0b5c6f | 1629 | */ |
43cf7631 | 1630 | *regs = *iregs; |
d1e7e602 SE |
1631 | |
1632 | /* | |
1633 | * Initialize regs_>flags from PEBS, | |
1634 | * Clear exact bit (which uses x86 EFLAGS Reserved bit 3), | |
1635 | * i.e., do not rely on it being zero: | |
1636 | */ | |
1637 | regs->flags = pebs->flags & ~PERF_EFLAGS_EXACT; | |
2b0b5c6f | 1638 | |
aea48559 | 1639 | if (sample_type & PERF_SAMPLE_REGS_INTR) { |
43cf7631 YZ |
1640 | regs->ax = pebs->ax; |
1641 | regs->bx = pebs->bx; | |
1642 | regs->cx = pebs->cx; | |
1643 | regs->dx = pebs->dx; | |
1644 | regs->si = pebs->si; | |
1645 | regs->di = pebs->di; | |
43cf7631 | 1646 | |
6cbc304f PZ |
1647 | regs->bp = pebs->bp; |
1648 | regs->sp = pebs->sp; | |
b8000586 | 1649 | |
aea48559 | 1650 | #ifndef CONFIG_X86_32 |
43cf7631 YZ |
1651 | regs->r8 = pebs->r8; |
1652 | regs->r9 = pebs->r9; | |
1653 | regs->r10 = pebs->r10; | |
1654 | regs->r11 = pebs->r11; | |
1655 | regs->r12 = pebs->r12; | |
1656 | regs->r13 = pebs->r13; | |
1657 | regs->r14 = pebs->r14; | |
1658 | regs->r15 = pebs->r15; | |
aea48559 SE |
1659 | #endif |
1660 | } | |
1661 | ||
71eb9ee9 | 1662 | if (event->attr.precise_ip > 1) { |
d1e7e602 SE |
1663 | /* |
1664 | * Haswell and later processors have an 'eventing IP' | |
1665 | * (real IP) which fixes the off-by-1 skid in hardware. | |
1666 | * Use it when precise_ip >= 2 : | |
1667 | */ | |
71eb9ee9 SE |
1668 | if (x86_pmu.intel_cap.pebs_format >= 2) { |
1669 | set_linear_ip(regs, pebs->real_ip); | |
1670 | regs->flags |= PERF_EFLAGS_EXACT; | |
1671 | } else { | |
d1e7e602 | 1672 | /* Otherwise, use PEBS off-by-1 IP: */ |
71eb9ee9 SE |
1673 | set_linear_ip(regs, pebs->ip); |
1674 | ||
d1e7e602 SE |
1675 | /* |
1676 | * With precise_ip >= 2, try to fix up the off-by-1 IP | |
1677 | * using the LBR. If successful, the fixup function | |
1678 | * corrects regs->ip and calls set_linear_ip() on regs: | |
1679 | */ | |
71eb9ee9 SE |
1680 | if (intel_pmu_pebs_fixup_ip(regs)) |
1681 | regs->flags |= PERF_EFLAGS_EXACT; | |
1682 | } | |
d1e7e602 SE |
1683 | } else { |
1684 | /* | |
1685 | * When precise_ip == 1, return the PEBS off-by-1 IP, | |
1686 | * no fixup attempted: | |
1687 | */ | |
71eb9ee9 | 1688 | set_linear_ip(regs, pebs->ip); |
d1e7e602 | 1689 | } |
71eb9ee9 | 1690 | |
2b0b5c6f | 1691 | |
76a5433f | 1692 | if ((sample_type & PERF_SAMPLE_ADDR_TYPE) && |
7b084630 | 1693 | x86_pmu.intel_cap.pebs_format >= 1) { |
43cf7631 | 1694 | data->addr = pebs->dla; |
7b084630 NK |
1695 | data->sample_flags |= PERF_SAMPLE_ADDR; |
1696 | } | |
f9134f36 | 1697 | |
a405bad5 AK |
1698 | if (x86_pmu.intel_cap.pebs_format >= 2) { |
1699 | /* Only set the TSX weight when no memory weight. */ | |
2abe681d | 1700 | if ((sample_type & PERF_SAMPLE_WEIGHT_TYPE) && !fll) { |
2a6c6b7d | 1701 | data->weight.full = intel_get_tsx_weight(pebs->tsx_tuning); |
2abe681d KL |
1702 | data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE; |
1703 | } | |
ee9db0e1 | 1704 | if (sample_type & PERF_SAMPLE_TRANSACTION) { |
48f38aa4 AK |
1705 | data->txn = intel_get_tsx_transaction(pebs->tsx_tuning, |
1706 | pebs->ax); | |
ee9db0e1 KL |
1707 | data->sample_flags |= PERF_SAMPLE_TRANSACTION; |
1708 | } | |
a405bad5 | 1709 | } |
748e86aa | 1710 | |
2f7ebf2e AK |
1711 | /* |
1712 | * v3 supplies an accurate time stamp, so we use that | |
1713 | * for the time stamp. | |
1714 | * | |
1715 | * We can only do this for the default trace clock. | |
1716 | */ | |
1717 | if (x86_pmu.intel_cap.pebs_format >= 3 && | |
47a3aeb3 | 1718 | event->attr.use_clockid == 0) { |
2f7ebf2e | 1719 | data->time = native_sched_clock_from_tsc(pebs->tsc); |
47a3aeb3 KL |
1720 | data->sample_flags |= PERF_SAMPLE_TIME; |
1721 | } | |
2f7ebf2e | 1722 | |
eb55b455 NK |
1723 | if (has_branch_stack(event)) |
1724 | perf_sample_save_brstack(data, event, &cpuc->lbr_stack); | |
43cf7631 YZ |
1725 | } |
1726 | ||
c22497f5 KL |
1727 | static void adaptive_pebs_save_regs(struct pt_regs *regs, |
1728 | struct pebs_gprs *gprs) | |
1729 | { | |
1730 | regs->ax = gprs->ax; | |
1731 | regs->bx = gprs->bx; | |
1732 | regs->cx = gprs->cx; | |
1733 | regs->dx = gprs->dx; | |
1734 | regs->si = gprs->si; | |
1735 | regs->di = gprs->di; | |
1736 | regs->bp = gprs->bp; | |
1737 | regs->sp = gprs->sp; | |
1738 | #ifndef CONFIG_X86_32 | |
1739 | regs->r8 = gprs->r8; | |
1740 | regs->r9 = gprs->r9; | |
1741 | regs->r10 = gprs->r10; | |
1742 | regs->r11 = gprs->r11; | |
1743 | regs->r12 = gprs->r12; | |
1744 | regs->r13 = gprs->r13; | |
1745 | regs->r14 = gprs->r14; | |
1746 | regs->r15 = gprs->r15; | |
1747 | #endif | |
1748 | } | |
1749 | ||
61b985e3 KL |
1750 | #define PEBS_LATENCY_MASK 0xffff |
1751 | #define PEBS_CACHE_LATENCY_OFFSET 32 | |
c87a3109 | 1752 | #define PEBS_RETIRE_LATENCY_OFFSET 32 |
61b985e3 | 1753 | |
c22497f5 KL |
1754 | /* |
1755 | * With adaptive PEBS the layout depends on what fields are configured. | |
1756 | */ | |
1757 | ||
1758 | static void setup_pebs_adaptive_sample_data(struct perf_event *event, | |
1759 | struct pt_regs *iregs, void *__pebs, | |
1760 | struct perf_sample_data *data, | |
1761 | struct pt_regs *regs) | |
1762 | { | |
1763 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); | |
1764 | struct pebs_basic *basic = __pebs; | |
1765 | void *next_record = basic + 1; | |
1766 | u64 sample_type; | |
1767 | u64 format_size; | |
1768 | struct pebs_meminfo *meminfo = NULL; | |
1769 | struct pebs_gprs *gprs = NULL; | |
1770 | struct x86_perf_regs *perf_regs; | |
1771 | ||
1772 | if (basic == NULL) | |
1773 | return; | |
1774 | ||
1775 | perf_regs = container_of(regs, struct x86_perf_regs, regs); | |
1776 | perf_regs->xmm_regs = NULL; | |
1777 | ||
1778 | sample_type = event->attr.sample_type; | |
1779 | format_size = basic->format_size; | |
1780 | perf_sample_data_init(data, 0, event->hw.last_period); | |
1781 | data->period = event->hw.last_period; | |
1782 | ||
47a3aeb3 | 1783 | if (event->attr.use_clockid == 0) { |
c22497f5 | 1784 | data->time = native_sched_clock_from_tsc(basic->tsc); |
47a3aeb3 KL |
1785 | data->sample_flags |= PERF_SAMPLE_TIME; |
1786 | } | |
c22497f5 KL |
1787 | |
1788 | /* | |
1789 | * We must however always use iregs for the unwinder to stay sane; the | |
1790 | * record BP,SP,IP can point into thin air when the record is from a | |
1791 | * previous PMI context or an (I)RET happened between the record and | |
1792 | * PMI. | |
1793 | */ | |
31046500 NK |
1794 | if (sample_type & PERF_SAMPLE_CALLCHAIN) |
1795 | perf_sample_save_callchain(data, event, iregs); | |
c22497f5 KL |
1796 | |
1797 | *regs = *iregs; | |
1798 | /* The ip in basic is EventingIP */ | |
1799 | set_linear_ip(regs, basic->ip); | |
1800 | regs->flags = PERF_EFLAGS_EXACT; | |
1801 | ||
c87a3109 KL |
1802 | if ((sample_type & PERF_SAMPLE_WEIGHT_STRUCT) && (x86_pmu.flags & PMU_FL_RETIRE_LATENCY)) |
1803 | data->weight.var3_w = format_size >> PEBS_RETIRE_LATENCY_OFFSET & PEBS_LATENCY_MASK; | |
1804 | ||
c22497f5 KL |
1805 | /* |
1806 | * The record for MEMINFO is in front of GP | |
1807 | * But PERF_SAMPLE_TRANSACTION needs gprs->ax. | |
1808 | * Save the pointer here but process later. | |
1809 | */ | |
1810 | if (format_size & PEBS_DATACFG_MEMINFO) { | |
1811 | meminfo = next_record; | |
1812 | next_record = meminfo + 1; | |
1813 | } | |
1814 | ||
1815 | if (format_size & PEBS_DATACFG_GP) { | |
1816 | gprs = next_record; | |
1817 | next_record = gprs + 1; | |
1818 | ||
1819 | if (event->attr.precise_ip < 2) { | |
1820 | set_linear_ip(regs, gprs->ip); | |
1821 | regs->flags &= ~PERF_EFLAGS_EXACT; | |
1822 | } | |
1823 | ||
1824 | if (sample_type & PERF_SAMPLE_REGS_INTR) | |
1825 | adaptive_pebs_save_regs(regs, gprs); | |
1826 | } | |
1827 | ||
1828 | if (format_size & PEBS_DATACFG_MEMINFO) { | |
61b985e3 KL |
1829 | if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) { |
1830 | u64 weight = meminfo->latency; | |
1831 | ||
1832 | if (x86_pmu.flags & PMU_FL_INSTR_LATENCY) { | |
1833 | data->weight.var2_w = weight & PEBS_LATENCY_MASK; | |
1834 | weight >>= PEBS_CACHE_LATENCY_OFFSET; | |
1835 | } | |
1836 | ||
1837 | /* | |
1838 | * Although meminfo::latency is defined as a u64, | |
1839 | * only the lower 32 bits include the valid data | |
1840 | * in practice on Ice Lake and earlier platforms. | |
1841 | */ | |
1842 | if (sample_type & PERF_SAMPLE_WEIGHT) { | |
1843 | data->weight.full = weight ?: | |
1844 | intel_get_tsx_weight(meminfo->tsx_tuning); | |
1845 | } else { | |
1846 | data->weight.var1_dw = (u32)(weight & PEBS_LATENCY_MASK) ?: | |
1847 | intel_get_tsx_weight(meminfo->tsx_tuning); | |
1848 | } | |
2abe681d | 1849 | data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE; |
61b985e3 | 1850 | } |
c22497f5 | 1851 | |
e16fd7f2 | 1852 | if (sample_type & PERF_SAMPLE_DATA_SRC) { |
c22497f5 | 1853 | data->data_src.val = get_data_src(event, meminfo->aux); |
e16fd7f2 KL |
1854 | data->sample_flags |= PERF_SAMPLE_DATA_SRC; |
1855 | } | |
c22497f5 | 1856 | |
7b084630 | 1857 | if (sample_type & PERF_SAMPLE_ADDR_TYPE) { |
c22497f5 | 1858 | data->addr = meminfo->address; |
7b084630 NK |
1859 | data->sample_flags |= PERF_SAMPLE_ADDR; |
1860 | } | |
c22497f5 | 1861 | |
ee9db0e1 | 1862 | if (sample_type & PERF_SAMPLE_TRANSACTION) { |
c22497f5 KL |
1863 | data->txn = intel_get_tsx_transaction(meminfo->tsx_tuning, |
1864 | gprs ? gprs->ax : 0); | |
ee9db0e1 KL |
1865 | data->sample_flags |= PERF_SAMPLE_TRANSACTION; |
1866 | } | |
c22497f5 KL |
1867 | } |
1868 | ||
1869 | if (format_size & PEBS_DATACFG_XMMS) { | |
1870 | struct pebs_xmm *xmm = next_record; | |
1871 | ||
1872 | next_record = xmm + 1; | |
1873 | perf_regs->xmm_regs = xmm->xmm; | |
1874 | } | |
1875 | ||
1876 | if (format_size & PEBS_DATACFG_LBRS) { | |
5624986d | 1877 | struct lbr_entry *lbr = next_record; |
c22497f5 KL |
1878 | int num_lbr = ((format_size >> PEBS_DATACFG_LBR_SHIFT) |
1879 | & 0xff) + 1; | |
5624986d | 1880 | next_record = next_record + num_lbr * sizeof(struct lbr_entry); |
c22497f5 KL |
1881 | |
1882 | if (has_branch_stack(event)) { | |
1883 | intel_pmu_store_pebs_lbrs(lbr); | |
eb55b455 | 1884 | perf_sample_save_brstack(data, event, &cpuc->lbr_stack); |
c22497f5 KL |
1885 | } |
1886 | } | |
1887 | ||
1888 | WARN_ONCE(next_record != __pebs + (format_size >> 48), | |
1889 | "PEBS record size %llu, expected %llu, config %llx\n", | |
1890 | format_size >> 48, | |
1891 | (u64)(next_record - __pebs), | |
1892 | basic->format_size); | |
1893 | } | |
1894 | ||
21509084 YZ |
1895 | static inline void * |
1896 | get_next_pebs_record_by_bit(void *base, void *top, int bit) | |
1897 | { | |
1898 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); | |
1899 | void *at; | |
1900 | u64 pebs_status; | |
1901 | ||
1424a09a SE |
1902 | /* |
1903 | * fmt0 does not have a status bitfield (does not use | |
1904 | * perf_record_nhm format) | |
1905 | */ | |
1906 | if (x86_pmu.intel_cap.pebs_format < 1) | |
1907 | return base; | |
1908 | ||
21509084 YZ |
1909 | if (base == NULL) |
1910 | return NULL; | |
1911 | ||
c22497f5 KL |
1912 | for (at = base; at < top; at += cpuc->pebs_record_size) { |
1913 | unsigned long status = get_pebs_status(at); | |
21509084 | 1914 | |
c22497f5 | 1915 | if (test_bit(bit, (unsigned long *)&status)) { |
a3d86542 PZ |
1916 | /* PEBS v3 has accurate status bits */ |
1917 | if (x86_pmu.intel_cap.pebs_format >= 3) | |
1918 | return at; | |
21509084 | 1919 | |
c22497f5 | 1920 | if (status == (1 << bit)) |
21509084 YZ |
1921 | return at; |
1922 | ||
1923 | /* clear non-PEBS bit and re-check */ | |
c22497f5 | 1924 | pebs_status = status & cpuc->pebs_enabled; |
fd583ad1 | 1925 | pebs_status &= PEBS_COUNTER_MASK; |
21509084 YZ |
1926 | if (pebs_status == (1 << bit)) |
1927 | return at; | |
1928 | } | |
1929 | } | |
1930 | return NULL; | |
1931 | } | |
1932 | ||
5bee2cc6 KL |
1933 | void intel_pmu_auto_reload_read(struct perf_event *event) |
1934 | { | |
1935 | WARN_ON(!(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)); | |
1936 | ||
1937 | perf_pmu_disable(event->pmu); | |
1938 | intel_pmu_drain_pebs_buffer(); | |
1939 | perf_pmu_enable(event->pmu); | |
1940 | } | |
1941 | ||
d31fc13f KL |
1942 | /* |
1943 | * Special variant of intel_pmu_save_and_restart() for auto-reload. | |
1944 | */ | |
1945 | static int | |
1946 | intel_pmu_save_and_restart_reload(struct perf_event *event, int count) | |
1947 | { | |
1948 | struct hw_perf_event *hwc = &event->hw; | |
1949 | int shift = 64 - x86_pmu.cntval_bits; | |
1950 | u64 period = hwc->sample_period; | |
1951 | u64 prev_raw_count, new_raw_count; | |
1952 | s64 new, old; | |
1953 | ||
1954 | WARN_ON(!period); | |
1955 | ||
1956 | /* | |
1957 | * drain_pebs() only happens when the PMU is disabled. | |
1958 | */ | |
1959 | WARN_ON(this_cpu_read(cpu_hw_events.enabled)); | |
1960 | ||
1961 | prev_raw_count = local64_read(&hwc->prev_count); | |
1962 | rdpmcl(hwc->event_base_rdpmc, new_raw_count); | |
1963 | local64_set(&hwc->prev_count, new_raw_count); | |
1964 | ||
1965 | /* | |
1966 | * Since the counter increments a negative counter value and | |
1967 | * overflows on the sign switch, giving the interval: | |
1968 | * | |
1969 | * [-period, 0] | |
1970 | * | |
d9f6e12f | 1971 | * the difference between two consecutive reads is: |
d31fc13f KL |
1972 | * |
1973 | * A) value2 - value1; | |
1974 | * when no overflows have happened in between, | |
1975 | * | |
1976 | * B) (0 - value1) + (value2 - (-period)); | |
1977 | * when one overflow happened in between, | |
1978 | * | |
1979 | * C) (0 - value1) + (n - 1) * (period) + (value2 - (-period)); | |
1980 | * when @n overflows happened in between. | |
1981 | * | |
1982 | * Here A) is the obvious difference, B) is the extension to the | |
1983 | * discrete interval, where the first term is to the top of the | |
1984 | * interval and the second term is from the bottom of the next | |
1985 | * interval and C) the extension to multiple intervals, where the | |
1986 | * middle term is the whole intervals covered. | |
1987 | * | |
1988 | * An equivalent of C, by reduction, is: | |
1989 | * | |
1990 | * value2 - value1 + n * period | |
1991 | */ | |
1992 | new = ((s64)(new_raw_count << shift) >> shift); | |
1993 | old = ((s64)(prev_raw_count << shift) >> shift); | |
1994 | local64_add(new - old + count * period, &event->count); | |
1995 | ||
f861854e KL |
1996 | local64_set(&hwc->period_left, -new); |
1997 | ||
d31fc13f KL |
1998 | perf_event_update_userpage(event); |
1999 | ||
2000 | return 0; | |
2001 | } | |
2002 | ||
9dfa9a5c PZ |
2003 | static __always_inline void |
2004 | __intel_pmu_pebs_event(struct perf_event *event, | |
2005 | struct pt_regs *iregs, | |
2006 | struct perf_sample_data *data, | |
2007 | void *base, void *top, | |
2008 | int bit, int count, | |
2009 | void (*setup_sample)(struct perf_event *, | |
2010 | struct pt_regs *, | |
2011 | void *, | |
2012 | struct perf_sample_data *, | |
2013 | struct pt_regs *)) | |
43cf7631 | 2014 | { |
c22497f5 | 2015 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
d31fc13f | 2016 | struct hw_perf_event *hwc = &event->hw; |
c22497f5 KL |
2017 | struct x86_perf_regs perf_regs; |
2018 | struct pt_regs *regs = &perf_regs.regs; | |
21509084 | 2019 | void *at = get_next_pebs_record_by_bit(base, top, bit); |
e506d1da | 2020 | static struct pt_regs dummy_iregs; |
43cf7631 | 2021 | |
d31fc13f KL |
2022 | if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) { |
2023 | /* | |
2024 | * Now, auto-reload is only enabled in fixed period mode. | |
2025 | * The reload value is always hwc->sample_period. | |
2026 | * May need to change it, if auto-reload is enabled in | |
2027 | * freq mode later. | |
2028 | */ | |
2029 | intel_pmu_save_and_restart_reload(event, count); | |
2030 | } else if (!intel_pmu_save_and_restart(event)) | |
43cf7631 YZ |
2031 | return; |
2032 | ||
35d1ce6b KL |
2033 | if (!iregs) |
2034 | iregs = &dummy_iregs; | |
2035 | ||
a3d86542 | 2036 | while (count > 1) { |
9dfa9a5c PZ |
2037 | setup_sample(event, iregs, at, data, regs); |
2038 | perf_event_output(event, data, regs); | |
c22497f5 | 2039 | at += cpuc->pebs_record_size; |
a3d86542 PZ |
2040 | at = get_next_pebs_record_by_bit(at, top, bit); |
2041 | count--; | |
21509084 YZ |
2042 | } |
2043 | ||
9dfa9a5c | 2044 | setup_sample(event, iregs, at, data, regs); |
35d1ce6b KL |
2045 | if (iregs == &dummy_iregs) { |
2046 | /* | |
2047 | * The PEBS records may be drained in the non-overflow context, | |
2048 | * e.g., large PEBS + context switch. Perf should treat the | |
2049 | * last record the same as other PEBS records, and doesn't | |
2050 | * invoke the generic overflow handler. | |
2051 | */ | |
9dfa9a5c | 2052 | perf_event_output(event, data, regs); |
35d1ce6b KL |
2053 | } else { |
2054 | /* | |
2055 | * All but the last records are processed. | |
2056 | * The last one is left to be able to call the overflow handler. | |
2057 | */ | |
9dfa9a5c | 2058 | if (perf_event_overflow(event, data, regs)) |
35d1ce6b | 2059 | x86_pmu_stop(event, 0); |
21509084 | 2060 | } |
2b0b5c6f PZ |
2061 | } |
2062 | ||
9dfa9a5c | 2063 | static void intel_pmu_drain_pebs_core(struct pt_regs *iregs, struct perf_sample_data *data) |
ca037701 | 2064 | { |
89cbc767 | 2065 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
ca037701 PZ |
2066 | struct debug_store *ds = cpuc->ds; |
2067 | struct perf_event *event = cpuc->events[0]; /* PMC0 only */ | |
2068 | struct pebs_record_core *at, *top; | |
ca037701 PZ |
2069 | int n; |
2070 | ||
6809b6ea | 2071 | if (!x86_pmu.pebs_active) |
ca037701 PZ |
2072 | return; |
2073 | ||
ca037701 PZ |
2074 | at = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base; |
2075 | top = (struct pebs_record_core *)(unsigned long)ds->pebs_index; | |
2076 | ||
d80c7502 PZ |
2077 | /* |
2078 | * Whatever else happens, drain the thing | |
2079 | */ | |
2080 | ds->pebs_index = ds->pebs_buffer_base; | |
2081 | ||
2082 | if (!test_bit(0, cpuc->active_mask)) | |
8f4aebd2 | 2083 | return; |
ca037701 | 2084 | |
d80c7502 PZ |
2085 | WARN_ON_ONCE(!event); |
2086 | ||
ab608344 | 2087 | if (!event->attr.precise_ip) |
d80c7502 PZ |
2088 | return; |
2089 | ||
1424a09a | 2090 | n = top - at; |
d31fc13f KL |
2091 | if (n <= 0) { |
2092 | if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD) | |
2093 | intel_pmu_save_and_restart_reload(event, 0); | |
d80c7502 | 2094 | return; |
d31fc13f | 2095 | } |
ca037701 | 2096 | |
9dfa9a5c | 2097 | __intel_pmu_pebs_event(event, iregs, data, at, top, 0, n, |
c22497f5 | 2098 | setup_pebs_fixed_sample_data); |
ca037701 PZ |
2099 | } |
2100 | ||
477f00f9 KL |
2101 | static void intel_pmu_pebs_event_update_no_drain(struct cpu_hw_events *cpuc, int size) |
2102 | { | |
2103 | struct perf_event *event; | |
2104 | int bit; | |
2105 | ||
2106 | /* | |
2107 | * The drain_pebs() could be called twice in a short period | |
2108 | * for auto-reload event in pmu::read(). There are no | |
2109 | * overflows have happened in between. | |
2110 | * It needs to call intel_pmu_save_and_restart_reload() to | |
2111 | * update the event->count for this case. | |
2112 | */ | |
2113 | for_each_set_bit(bit, (unsigned long *)&cpuc->pebs_enabled, size) { | |
2114 | event = cpuc->events[bit]; | |
2115 | if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD) | |
2116 | intel_pmu_save_and_restart_reload(event, 0); | |
2117 | } | |
2118 | } | |
2119 | ||
9dfa9a5c | 2120 | static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs, struct perf_sample_data *data) |
ca037701 | 2121 | { |
89cbc767 | 2122 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
ca037701 | 2123 | struct debug_store *ds = cpuc->ds; |
21509084 YZ |
2124 | struct perf_event *event; |
2125 | void *base, *at, *top; | |
ec71a398 KL |
2126 | short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {}; |
2127 | short error[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {}; | |
2128 | int bit, i, size; | |
2129 | u64 mask; | |
d2beea4a PZ |
2130 | |
2131 | if (!x86_pmu.pebs_active) | |
2132 | return; | |
2133 | ||
21509084 | 2134 | base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base; |
d2beea4a | 2135 | top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index; |
ca037701 | 2136 | |
ca037701 PZ |
2137 | ds->pebs_index = ds->pebs_buffer_base; |
2138 | ||
ec71a398 KL |
2139 | mask = (1ULL << x86_pmu.max_pebs_events) - 1; |
2140 | size = x86_pmu.max_pebs_events; | |
2141 | if (x86_pmu.flags & PMU_FL_PEBS_ALL) { | |
2142 | mask |= ((1ULL << x86_pmu.num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED; | |
2143 | size = INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed; | |
2144 | } | |
2145 | ||
d31fc13f | 2146 | if (unlikely(base >= top)) { |
477f00f9 | 2147 | intel_pmu_pebs_event_update_no_drain(cpuc, size); |
d2beea4a | 2148 | return; |
d31fc13f | 2149 | } |
d2beea4a | 2150 | |
21509084 | 2151 | for (at = base; at < top; at += x86_pmu.pebs_record_size) { |
130768b8 | 2152 | struct pebs_record_nhm *p = at; |
75f80859 | 2153 | u64 pebs_status; |
ca037701 | 2154 | |
8ef9b845 | 2155 | pebs_status = p->status & cpuc->pebs_enabled; |
ec71a398 | 2156 | pebs_status &= mask; |
8ef9b845 PZ |
2157 | |
2158 | /* PEBS v3 has more accurate status bits */ | |
a3d86542 | 2159 | if (x86_pmu.intel_cap.pebs_format >= 3) { |
c22497f5 | 2160 | for_each_set_bit(bit, (unsigned long *)&pebs_status, size) |
a3d86542 PZ |
2161 | counts[bit]++; |
2162 | ||
2163 | continue; | |
2164 | } | |
2165 | ||
01330d72 AK |
2166 | /* |
2167 | * On some CPUs the PEBS status can be zero when PEBS is | |
2168 | * racing with clearing of GLOBAL_STATUS. | |
2169 | * | |
2170 | * Normally we would drop that record, but in the | |
2171 | * case when there is only a single active PEBS event | |
2172 | * we can assume it's for that event. | |
2173 | */ | |
2174 | if (!pebs_status && cpuc->pebs_enabled && | |
2175 | !(cpuc->pebs_enabled & (cpuc->pebs_enabled-1))) | |
d88d05a9 | 2176 | pebs_status = p->status = cpuc->pebs_enabled; |
01330d72 | 2177 | |
75f80859 | 2178 | bit = find_first_bit((unsigned long *)&pebs_status, |
21509084 | 2179 | x86_pmu.max_pebs_events); |
957ea1fd | 2180 | if (bit >= x86_pmu.max_pebs_events) |
21509084 | 2181 | continue; |
75f80859 | 2182 | |
21509084 YZ |
2183 | /* |
2184 | * The PEBS hardware does not deal well with the situation | |
2185 | * when events happen near to each other and multiple bits | |
2186 | * are set. But it should happen rarely. | |
2187 | * | |
2188 | * If these events include one PEBS and multiple non-PEBS | |
2189 | * events, it doesn't impact PEBS record. The record will | |
2190 | * be handled normally. (slow path) | |
2191 | * | |
2192 | * If these events include two or more PEBS events, the | |
2193 | * records for the events can be collapsed into a single | |
2194 | * one, and it's not possible to reconstruct all events | |
2195 | * that caused the PEBS record. It's called collision. | |
2196 | * If collision happened, the record will be dropped. | |
21509084 | 2197 | */ |
fc17db8a | 2198 | if (pebs_status != (1ULL << bit)) { |
c22497f5 | 2199 | for_each_set_bit(i, (unsigned long *)&pebs_status, size) |
75f80859 PZ |
2200 | error[i]++; |
2201 | continue; | |
ca037701 | 2202 | } |
75f80859 | 2203 | |
21509084 YZ |
2204 | counts[bit]++; |
2205 | } | |
ca037701 | 2206 | |
c22497f5 | 2207 | for_each_set_bit(bit, (unsigned long *)&mask, size) { |
f38b0dbb | 2208 | if ((counts[bit] == 0) && (error[bit] == 0)) |
ca037701 | 2209 | continue; |
75f80859 | 2210 | |
21509084 | 2211 | event = cpuc->events[bit]; |
8ef9b845 PZ |
2212 | if (WARN_ON_ONCE(!event)) |
2213 | continue; | |
2214 | ||
2215 | if (WARN_ON_ONCE(!event->attr.precise_ip)) | |
2216 | continue; | |
ca037701 | 2217 | |
f38b0dbb | 2218 | /* log dropped samples number */ |
475113d9 | 2219 | if (error[bit]) { |
f38b0dbb KL |
2220 | perf_log_lost_samples(event, error[bit]); |
2221 | ||
5debf021 | 2222 | if (iregs && perf_event_account_interrupt(event)) |
475113d9 JO |
2223 | x86_pmu_stop(event, 0); |
2224 | } | |
2225 | ||
f38b0dbb | 2226 | if (counts[bit]) { |
9dfa9a5c | 2227 | __intel_pmu_pebs_event(event, iregs, data, base, |
c22497f5 KL |
2228 | top, bit, counts[bit], |
2229 | setup_pebs_fixed_sample_data); | |
f38b0dbb | 2230 | } |
ca037701 | 2231 | } |
ca037701 PZ |
2232 | } |
2233 | ||
9dfa9a5c | 2234 | static void intel_pmu_drain_pebs_icl(struct pt_regs *iregs, struct perf_sample_data *data) |
c22497f5 KL |
2235 | { |
2236 | short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {}; | |
2237 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); | |
d4b294bf KL |
2238 | int max_pebs_events = hybrid(cpuc->pmu, max_pebs_events); |
2239 | int num_counters_fixed = hybrid(cpuc->pmu, num_counters_fixed); | |
c22497f5 KL |
2240 | struct debug_store *ds = cpuc->ds; |
2241 | struct perf_event *event; | |
2242 | void *base, *at, *top; | |
2243 | int bit, size; | |
2244 | u64 mask; | |
2245 | ||
2246 | if (!x86_pmu.pebs_active) | |
2247 | return; | |
2248 | ||
2249 | base = (struct pebs_basic *)(unsigned long)ds->pebs_buffer_base; | |
2250 | top = (struct pebs_basic *)(unsigned long)ds->pebs_index; | |
2251 | ||
2252 | ds->pebs_index = ds->pebs_buffer_base; | |
2253 | ||
d4b294bf KL |
2254 | mask = ((1ULL << max_pebs_events) - 1) | |
2255 | (((1ULL << num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED); | |
2256 | size = INTEL_PMC_IDX_FIXED + num_counters_fixed; | |
c22497f5 KL |
2257 | |
2258 | if (unlikely(base >= top)) { | |
2259 | intel_pmu_pebs_event_update_no_drain(cpuc, size); | |
2260 | return; | |
2261 | } | |
2262 | ||
2263 | for (at = base; at < top; at += cpuc->pebs_record_size) { | |
2264 | u64 pebs_status; | |
2265 | ||
2266 | pebs_status = get_pebs_status(at) & cpuc->pebs_enabled; | |
2267 | pebs_status &= mask; | |
2268 | ||
2269 | for_each_set_bit(bit, (unsigned long *)&pebs_status, size) | |
2270 | counts[bit]++; | |
2271 | } | |
2272 | ||
2273 | for_each_set_bit(bit, (unsigned long *)&mask, size) { | |
2274 | if (counts[bit] == 0) | |
2275 | continue; | |
2276 | ||
2277 | event = cpuc->events[bit]; | |
2278 | if (WARN_ON_ONCE(!event)) | |
2279 | continue; | |
2280 | ||
2281 | if (WARN_ON_ONCE(!event->attr.precise_ip)) | |
2282 | continue; | |
2283 | ||
9dfa9a5c | 2284 | __intel_pmu_pebs_event(event, iregs, data, base, |
c22497f5 KL |
2285 | top, bit, counts[bit], |
2286 | setup_pebs_adaptive_sample_data); | |
2287 | } | |
2288 | } | |
2289 | ||
ca037701 PZ |
2290 | /* |
2291 | * BTS, PEBS probe and setup | |
2292 | */ | |
2293 | ||
066ce64c | 2294 | void __init intel_ds_init(void) |
ca037701 PZ |
2295 | { |
2296 | /* | |
2297 | * No support for 32bit formats | |
2298 | */ | |
2299 | if (!boot_cpu_has(X86_FEATURE_DTES64)) | |
2300 | return; | |
2301 | ||
2302 | x86_pmu.bts = boot_cpu_has(X86_FEATURE_BTS); | |
2303 | x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS); | |
e72daf3f | 2304 | x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE; |
cd6b984f | 2305 | if (x86_pmu.version <= 4) |
9b545c04 | 2306 | x86_pmu.pebs_no_isolation = 1; |
cd6b984f | 2307 | |
ca037701 | 2308 | if (x86_pmu.pebs) { |
8db909a7 | 2309 | char pebs_type = x86_pmu.intel_cap.pebs_trap ? '+' : '-'; |
c22497f5 | 2310 | char *pebs_qual = ""; |
8db909a7 | 2311 | int format = x86_pmu.intel_cap.pebs_format; |
ca037701 | 2312 | |
c22497f5 KL |
2313 | if (format < 4) |
2314 | x86_pmu.intel_cap.pebs_baseline = 0; | |
2315 | ||
ca037701 PZ |
2316 | switch (format) { |
2317 | case 0: | |
1b74dde7 | 2318 | pr_cont("PEBS fmt0%c, ", pebs_type); |
ca037701 | 2319 | x86_pmu.pebs_record_size = sizeof(struct pebs_record_core); |
e72daf3f JO |
2320 | /* |
2321 | * Using >PAGE_SIZE buffers makes the WRMSR to | |
2322 | * PERF_GLOBAL_CTRL in intel_pmu_enable_all() | |
2323 | * mysteriously hang on Core2. | |
2324 | * | |
2325 | * As a workaround, we don't do this. | |
2326 | */ | |
2327 | x86_pmu.pebs_buffer_size = PAGE_SIZE; | |
ca037701 | 2328 | x86_pmu.drain_pebs = intel_pmu_drain_pebs_core; |
ca037701 PZ |
2329 | break; |
2330 | ||
2331 | case 1: | |
1b74dde7 | 2332 | pr_cont("PEBS fmt1%c, ", pebs_type); |
ca037701 PZ |
2333 | x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm); |
2334 | x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm; | |
ca037701 PZ |
2335 | break; |
2336 | ||
130768b8 AK |
2337 | case 2: |
2338 | pr_cont("PEBS fmt2%c, ", pebs_type); | |
2339 | x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw); | |
d2beea4a | 2340 | x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm; |
130768b8 AK |
2341 | break; |
2342 | ||
2f7ebf2e AK |
2343 | case 3: |
2344 | pr_cont("PEBS fmt3%c, ", pebs_type); | |
2345 | x86_pmu.pebs_record_size = | |
2346 | sizeof(struct pebs_record_skl); | |
2347 | x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm; | |
174afc3e | 2348 | x86_pmu.large_pebs_flags |= PERF_SAMPLE_TIME; |
2f7ebf2e AK |
2349 | break; |
2350 | ||
c22497f5 | 2351 | case 4: |
2145e77f | 2352 | case 5: |
c22497f5 KL |
2353 | x86_pmu.drain_pebs = intel_pmu_drain_pebs_icl; |
2354 | x86_pmu.pebs_record_size = sizeof(struct pebs_basic); | |
2355 | if (x86_pmu.intel_cap.pebs_baseline) { | |
2356 | x86_pmu.large_pebs_flags |= | |
2357 | PERF_SAMPLE_BRANCH_STACK | | |
2358 | PERF_SAMPLE_TIME; | |
2359 | x86_pmu.flags |= PMU_FL_PEBS_ALL; | |
7d359886 | 2360 | x86_pmu.pebs_capable = ~0ULL; |
c22497f5 | 2361 | pebs_qual = "-baseline"; |
61e76d53 | 2362 | x86_get_pmu(smp_processor_id())->capabilities |= PERF_PMU_CAP_EXTENDED_REGS; |
c22497f5 KL |
2363 | } else { |
2364 | /* Only basic record supported */ | |
c22497f5 KL |
2365 | x86_pmu.large_pebs_flags &= |
2366 | ~(PERF_SAMPLE_ADDR | | |
2367 | PERF_SAMPLE_TIME | | |
2368 | PERF_SAMPLE_DATA_SRC | | |
2369 | PERF_SAMPLE_TRANSACTION | | |
2370 | PERF_SAMPLE_REGS_USER | | |
2371 | PERF_SAMPLE_REGS_INTR); | |
2372 | } | |
2373 | pr_cont("PEBS fmt4%c%s, ", pebs_type, pebs_qual); | |
42880f72 | 2374 | |
d0946a88 | 2375 | if (!is_hybrid() && x86_pmu.intel_cap.pebs_output_pt_available) { |
42880f72 | 2376 | pr_cont("PEBS-via-PT, "); |
61e76d53 | 2377 | x86_get_pmu(smp_processor_id())->capabilities |= PERF_PMU_CAP_AUX_OUTPUT; |
42880f72 AS |
2378 | } |
2379 | ||
c22497f5 KL |
2380 | break; |
2381 | ||
ca037701 | 2382 | default: |
1b74dde7 | 2383 | pr_cont("no PEBS fmt%d%c, ", format, pebs_type); |
ca037701 | 2384 | x86_pmu.pebs = 0; |
ca037701 PZ |
2385 | } |
2386 | } | |
2387 | } | |
1d9d8639 SE |
2388 | |
2389 | void perf_restore_debug_store(void) | |
2390 | { | |
2a6e06b2 LT |
2391 | struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds); |
2392 | ||
1d9d8639 SE |
2393 | if (!x86_pmu.bts && !x86_pmu.pebs) |
2394 | return; | |
2395 | ||
2a6e06b2 | 2396 | wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds); |
1d9d8639 | 2397 | } |