]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - arch/x86/kernel/cpu/perf_event_intel_ds.c
projects / mirror_ubuntu-bionic-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge tag 'drm-intel-next-2014-02-14' of ssh://git.freedesktop.org/git/drm-intel...
[mirror_ubuntu-bionic-kernel.git] / arch / x86 / kernel / cpu / perf_event_intel_ds.c
1 #include <linux/bitops.h>
2 #include <linux/types.h>
3 #include <linux/slab.h>
4
5 #include <asm/perf_event.h>
6 #include <asm/insn.h>
7
8 #include "perf_event.h"
9
10 /* The size of a BTS record in bytes: */
11 #define BTS_RECORD_SIZE 24
12
13 #define BTS_BUFFER_SIZE (PAGE_SIZE << 4)
14 #define PEBS_BUFFER_SIZE PAGE_SIZE
15 #define PEBS_FIXUP_SIZE PAGE_SIZE
16
17 /*
18 * pebs_record_32 for p4 and core not supported
19
20 struct pebs_record_32 {
21 u32 flags, ip;
22 u32 ax, bc, cx, dx;
23 u32 si, di, bp, sp;
24 };
25
26 */
27
28 union intel_x86_pebs_dse {
29 u64 val;
30 struct {
31 unsigned int ld_dse:4;
32 unsigned int ld_stlb_miss:1;
33 unsigned int ld_locked:1;
34 unsigned int ld_reserved:26;
35 };
36 struct {
37 unsigned int st_l1d_hit:1;
38 unsigned int st_reserved1:3;
39 unsigned int st_stlb_miss:1;
40 unsigned int st_locked:1;
41 unsigned int st_reserved2:26;
42 };
43 };
44
45
46 /*
47 * Map PEBS Load Latency Data Source encodings to generic
48 * memory data source information
49 */
50 #define P(a, b) PERF_MEM_S(a, b)
51 #define OP_LH (P(OP, LOAD) | P(LVL, HIT))
52 #define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
53
54 static const u64 pebs_data_source[] = {
55 P(OP, LOAD) | P(LVL, MISS) | P(LVL, L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
56 OP_LH | P(LVL, L1) | P(SNOOP, NONE), /* 0x01: L1 local */
57 OP_LH | P(LVL, LFB) | P(SNOOP, NONE), /* 0x02: LFB hit */
58 OP_LH | P(LVL, L2) | P(SNOOP, NONE), /* 0x03: L2 hit */
59 OP_LH | P(LVL, L3) | P(SNOOP, NONE), /* 0x04: L3 hit */
60 OP_LH | P(LVL, L3) | P(SNOOP, MISS), /* 0x05: L3 hit, snoop miss */
61 OP_LH | P(LVL, L3) | P(SNOOP, HIT), /* 0x06: L3 hit, snoop hit */
62 OP_LH | P(LVL, L3) | P(SNOOP, HITM), /* 0x07: L3 hit, snoop hitm */
63 OP_LH | P(LVL, REM_CCE1) | P(SNOOP, HIT), /* 0x08: L3 miss snoop hit */
64 OP_LH | P(LVL, REM_CCE1) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
65 OP_LH | P(LVL, LOC_RAM) | P(SNOOP, HIT), /* 0x0a: L3 miss, shared */
66 OP_LH | P(LVL, REM_RAM1) | P(SNOOP, HIT), /* 0x0b: L3 miss, shared */
67 OP_LH | P(LVL, LOC_RAM) | SNOOP_NONE_MISS,/* 0x0c: L3 miss, excl */
68 OP_LH | P(LVL, REM_RAM1) | SNOOP_NONE_MISS,/* 0x0d: L3 miss, excl */
69 OP_LH | P(LVL, IO) | P(SNOOP, NONE), /* 0x0e: I/O */
70 OP_LH | P(LVL, UNC) | P(SNOOP, NONE), /* 0x0f: uncached */
71 };
72
73 static u64 precise_store_data(u64 status)
74 {
75 union intel_x86_pebs_dse dse;
76 u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
77
78 dse.val = status;
79
80 /*
81 * bit 4: TLB access
82 * 1 = stored missed 2nd level TLB
83 *
84 * so it either hit the walker or the OS
85 * otherwise hit 2nd level TLB
86 */
87 if (dse.st_stlb_miss)
88 val |= P(TLB, MISS);
89 else
90 val |= P(TLB, HIT);
91
92 /*
93 * bit 0: hit L1 data cache
94 * if not set, then all we know is that
95 * it missed L1D
96 */
97 if (dse.st_l1d_hit)
98 val |= P(LVL, HIT);
99 else
100 val |= P(LVL, MISS);
101
102 /*
103 * bit 5: Locked prefix
104 */
105 if (dse.st_locked)
106 val |= P(LOCK, LOCKED);
107
108 return val;
109 }
110
111 static u64 precise_store_data_hsw(u64 status)
112 {
113 union perf_mem_data_src dse;
114
115 dse.val = 0;
116 dse.mem_op = PERF_MEM_OP_STORE;
117 dse.mem_lvl = PERF_MEM_LVL_NA;
118 if (status & 1)
119 dse.mem_lvl = PERF_MEM_LVL_L1;
120 /* Nothing else supported. Sorry. */
121 return dse.val;
122 }
123
124 static u64 load_latency_data(u64 status)
125 {
126 union intel_x86_pebs_dse dse;
127 u64 val;
128 int model = boot_cpu_data.x86_model;
129 int fam = boot_cpu_data.x86;
130
131 dse.val = status;
132
133 /*
134 * use the mapping table for bit 0-3
135 */
136 val = pebs_data_source[dse.ld_dse];
137
138 /*
139 * Nehalem models do not support TLB, Lock infos
140 */
141 if (fam == 0x6 && (model == 26 || model == 30
142 || model == 31 || model == 46)) {
143 val |= P(TLB, NA) | P(LOCK, NA);
144 return val;
145 }
146 /*
147 * bit 4: TLB access
148 * 0 = did not miss 2nd level TLB
149 * 1 = missed 2nd level TLB
150 */
151 if (dse.ld_stlb_miss)
152 val |= P(TLB, MISS) | P(TLB, L2);
153 else
154 val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
155
156 /*
157 * bit 5: locked prefix
158 */
159 if (dse.ld_locked)
160 val |= P(LOCK, LOCKED);
161
162 return val;
163 }
164
165 struct pebs_record_core {
166 u64 flags, ip;
167 u64 ax, bx, cx, dx;
168 u64 si, di, bp, sp;
169 u64 r8, r9, r10, r11;
170 u64 r12, r13, r14, r15;
171 };
172
173 struct pebs_record_nhm {
174 u64 flags, ip;
175 u64 ax, bx, cx, dx;
176 u64 si, di, bp, sp;
177 u64 r8, r9, r10, r11;
178 u64 r12, r13, r14, r15;
179 u64 status, dla, dse, lat;
180 };
181
182 /*
183 * Same as pebs_record_nhm, with two additional fields.
184 */
185 struct pebs_record_hsw {
186 u64 flags, ip;
187 u64 ax, bx, cx, dx;
188 u64 si, di, bp, sp;
189 u64 r8, r9, r10, r11;
190 u64 r12, r13, r14, r15;
191 u64 status, dla, dse, lat;
192 u64 real_ip, tsx_tuning;
193 };
194
195 union hsw_tsx_tuning {
196 struct {
197 u32 cycles_last_block : 32,
198 hle_abort : 1,
199 rtm_abort : 1,
200 instruction_abort : 1,
201 non_instruction_abort : 1,
202 retry : 1,
203 data_conflict : 1,
204 capacity_writes : 1,
205 capacity_reads : 1;
206 };
207 u64 value;
208 };
209
210 #define PEBS_HSW_TSX_FLAGS 0xff00000000ULL
211
212 void init_debug_store_on_cpu(int cpu)
213 {
214 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
215
216 if (!ds)
217 return;
218
219 wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
220 (u32)((u64)(unsigned long)ds),
221 (u32)((u64)(unsigned long)ds >> 32));
222 }
223
224 void fini_debug_store_on_cpu(int cpu)
225 {
226 if (!per_cpu(cpu_hw_events, cpu).ds)
227 return;
228
229 wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
230 }
231
232 static DEFINE_PER_CPU(void *, insn_buffer);
233
234 static int alloc_pebs_buffer(int cpu)
235 {
236 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
237 int node = cpu_to_node(cpu);
238 int max, thresh = 1; /* always use a single PEBS record */
239 void *buffer, *ibuffer;
240
241 if (!x86_pmu.pebs)
242 return 0;
243
244 buffer = kzalloc_node(PEBS_BUFFER_SIZE, GFP_KERNEL, node);
245 if (unlikely(!buffer))
246 return -ENOMEM;
247
248 /*
249 * HSW+ already provides us the eventing ip; no need to allocate this
250 * buffer then.
251 */
252 if (x86_pmu.intel_cap.pebs_format < 2) {
253 ibuffer = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
254 if (!ibuffer) {
255 kfree(buffer);
256 return -ENOMEM;
257 }
258 per_cpu(insn_buffer, cpu) = ibuffer;
259 }
260
261 max = PEBS_BUFFER_SIZE / x86_pmu.pebs_record_size;
262
263 ds->pebs_buffer_base = (u64)(unsigned long)buffer;
264 ds->pebs_index = ds->pebs_buffer_base;
265 ds->pebs_absolute_maximum = ds->pebs_buffer_base +
266 max * x86_pmu.pebs_record_size;
267
268 ds->pebs_interrupt_threshold = ds->pebs_buffer_base +
269 thresh * x86_pmu.pebs_record_size;
270
271 return 0;
272 }
273
274 static void release_pebs_buffer(int cpu)
275 {
276 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
277
278 if (!ds || !x86_pmu.pebs)
279 return;
280
281 kfree(per_cpu(insn_buffer, cpu));
282 per_cpu(insn_buffer, cpu) = NULL;
283
284 kfree((void *)(unsigned long)ds->pebs_buffer_base);
285 ds->pebs_buffer_base = 0;
286 }
287
288 static int alloc_bts_buffer(int cpu)
289 {
290 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
291 int node = cpu_to_node(cpu);
292 int max, thresh;
293 void *buffer;
294
295 if (!x86_pmu.bts)
296 return 0;
297
298 buffer = kzalloc_node(BTS_BUFFER_SIZE, GFP_KERNEL, node);
299 if (unlikely(!buffer))
300 return -ENOMEM;
301
302 max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
303 thresh = max / 16;
304
305 ds->bts_buffer_base = (u64)(unsigned long)buffer;
306 ds->bts_index = ds->bts_buffer_base;
307 ds->bts_absolute_maximum = ds->bts_buffer_base +
308 max * BTS_RECORD_SIZE;
309 ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
310 thresh * BTS_RECORD_SIZE;
311
312 return 0;
313 }
314
315 static void release_bts_buffer(int cpu)
316 {
317 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
318
319 if (!ds || !x86_pmu.bts)
320 return;
321
322 kfree((void *)(unsigned long)ds->bts_buffer_base);
323 ds->bts_buffer_base = 0;
324 }
325
326 static int alloc_ds_buffer(int cpu)
327 {
328 int node = cpu_to_node(cpu);
329 struct debug_store *ds;
330
331 ds = kzalloc_node(sizeof(*ds), GFP_KERNEL, node);
332 if (unlikely(!ds))
333 return -ENOMEM;
334
335 per_cpu(cpu_hw_events, cpu).ds = ds;
336
337 return 0;
338 }
339
340 static void release_ds_buffer(int cpu)
341 {
342 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
343
344 if (!ds)
345 return;
346
347 per_cpu(cpu_hw_events, cpu).ds = NULL;
348 kfree(ds);
349 }
350
351 void release_ds_buffers(void)
352 {
353 int cpu;
354
355 if (!x86_pmu.bts && !x86_pmu.pebs)
356 return;
357
358 get_online_cpus();
359 for_each_online_cpu(cpu)
360 fini_debug_store_on_cpu(cpu);
361
362 for_each_possible_cpu(cpu) {
363 release_pebs_buffer(cpu);
364 release_bts_buffer(cpu);
365 release_ds_buffer(cpu);
366 }
367 put_online_cpus();
368 }
369
370 void reserve_ds_buffers(void)
371 {
372 int bts_err = 0, pebs_err = 0;
373 int cpu;
374
375 x86_pmu.bts_active = 0;
376 x86_pmu.pebs_active = 0;
377
378 if (!x86_pmu.bts && !x86_pmu.pebs)
379 return;
380
381 if (!x86_pmu.bts)
382 bts_err = 1;
383
384 if (!x86_pmu.pebs)
385 pebs_err = 1;
386
387 get_online_cpus();
388
389 for_each_possible_cpu(cpu) {
390 if (alloc_ds_buffer(cpu)) {
391 bts_err = 1;
392 pebs_err = 1;
393 }
394
395 if (!bts_err && alloc_bts_buffer(cpu))
396 bts_err = 1;
397
398 if (!pebs_err && alloc_pebs_buffer(cpu))
399 pebs_err = 1;
400
401 if (bts_err && pebs_err)
402 break;
403 }
404
405 if (bts_err) {
406 for_each_possible_cpu(cpu)
407 release_bts_buffer(cpu);
408 }
409
410 if (pebs_err) {
411 for_each_possible_cpu(cpu)
412 release_pebs_buffer(cpu);
413 }
414
415 if (bts_err && pebs_err) {
416 for_each_possible_cpu(cpu)
417 release_ds_buffer(cpu);
418 } else {
419 if (x86_pmu.bts && !bts_err)
420 x86_pmu.bts_active = 1;
421
422 if (x86_pmu.pebs && !pebs_err)
423 x86_pmu.pebs_active = 1;
424
425 for_each_online_cpu(cpu)
426 init_debug_store_on_cpu(cpu);
427 }
428
429 put_online_cpus();
430 }
431
432 /*
433 * BTS
434 */
435
436 struct event_constraint bts_constraint =
437 EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
438
439 void intel_pmu_enable_bts(u64 config)
440 {
441 unsigned long debugctlmsr;
442
443 debugctlmsr = get_debugctlmsr();
444
445 debugctlmsr |= DEBUGCTLMSR_TR;
446 debugctlmsr |= DEBUGCTLMSR_BTS;
447 debugctlmsr |= DEBUGCTLMSR_BTINT;
448
449 if (!(config & ARCH_PERFMON_EVENTSEL_OS))
450 debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
451
452 if (!(config & ARCH_PERFMON_EVENTSEL_USR))
453 debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
454
455 update_debugctlmsr(debugctlmsr);
456 }
457
458 void intel_pmu_disable_bts(void)
459 {
460 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
461 unsigned long debugctlmsr;
462
463 if (!cpuc->ds)
464 return;
465
466 debugctlmsr = get_debugctlmsr();
467
468 debugctlmsr &=
469 ~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
470 DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
471
472 update_debugctlmsr(debugctlmsr);
473 }
474
475 int intel_pmu_drain_bts_buffer(void)
476 {
477 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
478 struct debug_store *ds = cpuc->ds;
479 struct bts_record {
480 u64 from;
481 u64 to;
482 u64 flags;
483 };
484 struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
485 struct bts_record *at, *top;
486 struct perf_output_handle handle;
487 struct perf_event_header header;
488 struct perf_sample_data data;
489 struct pt_regs regs;
490
491 if (!event)
492 return 0;
493
494 if (!x86_pmu.bts_active)
495 return 0;
496
497 at = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
498 top = (struct bts_record *)(unsigned long)ds->bts_index;
499
500 if (top <= at)
501 return 0;
502
503 memset(&regs, 0, sizeof(regs));
504
505 ds->bts_index = ds->bts_buffer_base;
506
507 perf_sample_data_init(&data, 0, event->hw.last_period);
508
509 /*
510 * Prepare a generic sample, i.e. fill in the invariant fields.
511 * We will overwrite the from and to address before we output
512 * the sample.
513 */
514 perf_prepare_sample(&header, &data, event, &regs);
515
516 if (perf_output_begin(&handle, event, header.size * (top - at)))
517 return 1;
518
519 for (; at < top; at++) {
520 data.ip = at->from;
521 data.addr = at->to;
522
523 perf_output_sample(&handle, &header, &data, event);
524 }
525
526 perf_output_end(&handle);
527
528 /* There's new data available. */
529 event->hw.interrupts++;
530 event->pending_kill = POLL_IN;
531 return 1;
532 }
533
534 /*
535 * PEBS
536 */
537 struct event_constraint intel_core2_pebs_event_constraints[] = {
538 INTEL_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
539 INTEL_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
540 INTEL_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
541 INTEL_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
542 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED.* */
543 EVENT_CONSTRAINT_END
544 };
545
546 struct event_constraint intel_atom_pebs_event_constraints[] = {
547 INTEL_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
548 INTEL_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
549 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED.* */
550 EVENT_CONSTRAINT_END
551 };
552
553 struct event_constraint intel_slm_pebs_event_constraints[] = {
554 INTEL_UEVENT_CONSTRAINT(0x0103, 0x1), /* REHABQ.LD_BLOCK_ST_FORWARD_PS */
555 INTEL_UEVENT_CONSTRAINT(0x0803, 0x1), /* REHABQ.LD_SPLITS_PS */
556 INTEL_UEVENT_CONSTRAINT(0x0204, 0x1), /* MEM_UOPS_RETIRED.L2_HIT_LOADS_PS */
557 INTEL_UEVENT_CONSTRAINT(0x0404, 0x1), /* MEM_UOPS_RETIRED.L2_MISS_LOADS_PS */
558 INTEL_UEVENT_CONSTRAINT(0x0804, 0x1), /* MEM_UOPS_RETIRED.DTLB_MISS_LOADS_PS */
559 INTEL_UEVENT_CONSTRAINT(0x2004, 0x1), /* MEM_UOPS_RETIRED.HITM_PS */
560 INTEL_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY_PS */
561 INTEL_UEVENT_CONSTRAINT(0x00c4, 0x1), /* BR_INST_RETIRED.ALL_BRANCHES_PS */
562 INTEL_UEVENT_CONSTRAINT(0x7ec4, 0x1), /* BR_INST_RETIRED.JCC_PS */
563 INTEL_UEVENT_CONSTRAINT(0xbfc4, 0x1), /* BR_INST_RETIRED.FAR_BRANCH_PS */
564 INTEL_UEVENT_CONSTRAINT(0xebc4, 0x1), /* BR_INST_RETIRED.NON_RETURN_IND_PS */
565 INTEL_UEVENT_CONSTRAINT(0xf7c4, 0x1), /* BR_INST_RETIRED.RETURN_PS */
566 INTEL_UEVENT_CONSTRAINT(0xf9c4, 0x1), /* BR_INST_RETIRED.CALL_PS */
567 INTEL_UEVENT_CONSTRAINT(0xfbc4, 0x1), /* BR_INST_RETIRED.IND_CALL_PS */
568 INTEL_UEVENT_CONSTRAINT(0xfdc4, 0x1), /* BR_INST_RETIRED.REL_CALL_PS */
569 INTEL_UEVENT_CONSTRAINT(0xfec4, 0x1), /* BR_INST_RETIRED.TAKEN_JCC_PS */
570 INTEL_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_MISP_RETIRED.ALL_BRANCHES_PS */
571 INTEL_UEVENT_CONSTRAINT(0x7ec5, 0x1), /* BR_INST_MISP_RETIRED.JCC_PS */
572 INTEL_UEVENT_CONSTRAINT(0xebc5, 0x1), /* BR_INST_MISP_RETIRED.NON_RETURN_IND_PS */
573 INTEL_UEVENT_CONSTRAINT(0xf7c5, 0x1), /* BR_INST_MISP_RETIRED.RETURN_PS */
574 INTEL_UEVENT_CONSTRAINT(0xfbc5, 0x1), /* BR_INST_MISP_RETIRED.IND_CALL_PS */
575 INTEL_UEVENT_CONSTRAINT(0xfec5, 0x1), /* BR_INST_MISP_RETIRED.TAKEN_JCC_PS */
576 EVENT_CONSTRAINT_END
577 };
578
579 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
580 INTEL_PLD_CONSTRAINT(0x100b, 0xf), /* MEM_INST_RETIRED.* */
581 INTEL_EVENT_CONSTRAINT(0x0f, 0xf), /* MEM_UNCORE_RETIRED.* */
582 INTEL_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
583 INTEL_EVENT_CONSTRAINT(0xc0, 0xf), /* INST_RETIRED.ANY */
584 INTEL_EVENT_CONSTRAINT(0xc2, 0xf), /* UOPS_RETIRED.* */
585 INTEL_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
586 INTEL_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
587 INTEL_EVENT_CONSTRAINT(0xc7, 0xf), /* SSEX_UOPS_RETIRED.* */
588 INTEL_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
589 INTEL_EVENT_CONSTRAINT(0xcb, 0xf), /* MEM_LOAD_RETIRED.* */
590 INTEL_EVENT_CONSTRAINT(0xf7, 0xf), /* FP_ASSIST.* */
591 EVENT_CONSTRAINT_END
592 };
593
594 struct event_constraint intel_westmere_pebs_event_constraints[] = {
595 INTEL_PLD_CONSTRAINT(0x100b, 0xf), /* MEM_INST_RETIRED.* */
596 INTEL_EVENT_CONSTRAINT(0x0f, 0xf), /* MEM_UNCORE_RETIRED.* */
597 INTEL_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
598 INTEL_EVENT_CONSTRAINT(0xc0, 0xf), /* INSTR_RETIRED.* */
599 INTEL_EVENT_CONSTRAINT(0xc2, 0xf), /* UOPS_RETIRED.* */
600 INTEL_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
601 INTEL_EVENT_CONSTRAINT(0xc5, 0xf), /* BR_MISP_RETIRED.* */
602 INTEL_EVENT_CONSTRAINT(0xc7, 0xf), /* SSEX_UOPS_RETIRED.* */
603 INTEL_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
604 INTEL_EVENT_CONSTRAINT(0xcb, 0xf), /* MEM_LOAD_RETIRED.* */
605 INTEL_EVENT_CONSTRAINT(0xf7, 0xf), /* FP_ASSIST.* */
606 EVENT_CONSTRAINT_END
607 };
608
609 struct event_constraint intel_snb_pebs_event_constraints[] = {
610 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
611 INTEL_UEVENT_CONSTRAINT(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
612 INTEL_UEVENT_CONSTRAINT(0x02c2, 0xf), /* UOPS_RETIRED.RETIRE_SLOTS */
613 INTEL_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
614 INTEL_EVENT_CONSTRAINT(0xc5, 0xf), /* BR_MISP_RETIRED.* */
615 INTEL_PLD_CONSTRAINT(0x01cd, 0x8), /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
616 INTEL_PST_CONSTRAINT(0x02cd, 0x8), /* MEM_TRANS_RETIRED.PRECISE_STORES */
617 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_UOP_RETIRED.* */
618 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
619 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
620 INTEL_EVENT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
621 INTEL_UEVENT_CONSTRAINT(0x02d4, 0xf), /* MEM_LOAD_UOPS_MISC_RETIRED.LLC_MISS */
622 EVENT_CONSTRAINT_END
623 };
624
625 struct event_constraint intel_ivb_pebs_event_constraints[] = {
626 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
627 INTEL_UEVENT_CONSTRAINT(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
628 INTEL_UEVENT_CONSTRAINT(0x02c2, 0xf), /* UOPS_RETIRED.RETIRE_SLOTS */
629 INTEL_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
630 INTEL_EVENT_CONSTRAINT(0xc5, 0xf), /* BR_MISP_RETIRED.* */
631 INTEL_PLD_CONSTRAINT(0x01cd, 0x8), /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
632 INTEL_PST_CONSTRAINT(0x02cd, 0x8), /* MEM_TRANS_RETIRED.PRECISE_STORES */
633 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_UOP_RETIRED.* */
634 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
635 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
636 INTEL_EVENT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
637 EVENT_CONSTRAINT_END
638 };
639
640 struct event_constraint intel_hsw_pebs_event_constraints[] = {
641 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
642 INTEL_PST_HSW_CONSTRAINT(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
643 INTEL_UEVENT_CONSTRAINT(0x02c2, 0xf), /* UOPS_RETIRED.RETIRE_SLOTS */
644 INTEL_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
645 INTEL_UEVENT_CONSTRAINT(0x01c5, 0xf), /* BR_MISP_RETIRED.CONDITIONAL */
646 INTEL_UEVENT_CONSTRAINT(0x04c5, 0xf), /* BR_MISP_RETIRED.ALL_BRANCHES */
647 INTEL_UEVENT_CONSTRAINT(0x20c5, 0xf), /* BR_MISP_RETIRED.NEAR_TAKEN */
648 INTEL_PLD_CONSTRAINT(0x01cd, 0x8), /* MEM_TRANS_RETIRED.* */
649 /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
650 INTEL_UEVENT_CONSTRAINT(0x11d0, 0xf),
651 /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
652 INTEL_UEVENT_CONSTRAINT(0x12d0, 0xf),
653 INTEL_UEVENT_CONSTRAINT(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
654 INTEL_UEVENT_CONSTRAINT(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
655 /* MEM_UOPS_RETIRED.SPLIT_STORES */
656 INTEL_UEVENT_CONSTRAINT(0x42d0, 0xf),
657 INTEL_UEVENT_CONSTRAINT(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
658 INTEL_PST_HSW_CONSTRAINT(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
659 INTEL_UEVENT_CONSTRAINT(0x01d1, 0xf), /* MEM_LOAD_UOPS_RETIRED.L1_HIT */
660 INTEL_UEVENT_CONSTRAINT(0x02d1, 0xf), /* MEM_LOAD_UOPS_RETIRED.L2_HIT */
661 INTEL_UEVENT_CONSTRAINT(0x04d1, 0xf), /* MEM_LOAD_UOPS_RETIRED.L3_HIT */
662 /* MEM_LOAD_UOPS_RETIRED.HIT_LFB */
663 INTEL_UEVENT_CONSTRAINT(0x40d1, 0xf),
664 /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_MISS */
665 INTEL_UEVENT_CONSTRAINT(0x01d2, 0xf),
666 /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT */
667 INTEL_UEVENT_CONSTRAINT(0x02d2, 0xf),
668 /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.LOCAL_DRAM */
669 INTEL_UEVENT_CONSTRAINT(0x01d3, 0xf),
670 INTEL_UEVENT_CONSTRAINT(0x04c8, 0xf), /* HLE_RETIRED.Abort */
671 INTEL_UEVENT_CONSTRAINT(0x04c9, 0xf), /* RTM_RETIRED.Abort */
672
673 EVENT_CONSTRAINT_END
674 };
675
676 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
677 {
678 struct event_constraint *c;
679
680 if (!event->attr.precise_ip)
681 return NULL;
682
683 if (x86_pmu.pebs_constraints) {
684 for_each_event_constraint(c, x86_pmu.pebs_constraints) {
685 if ((event->hw.config & c->cmask) == c->code) {
686 event->hw.flags |= c->flags;
687 return c;
688 }
689 }
690 }
691
692 return &emptyconstraint;
693 }
694
695 void intel_pmu_pebs_enable(struct perf_event *event)
696 {
697 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
698 struct hw_perf_event *hwc = &event->hw;
699
700 hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
701
702 cpuc->pebs_enabled |= 1ULL << hwc->idx;
703
704 if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
705 cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
706 else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
707 cpuc->pebs_enabled |= 1ULL << 63;
708 }
709
710 void intel_pmu_pebs_disable(struct perf_event *event)
711 {
712 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
713 struct hw_perf_event *hwc = &event->hw;
714
715 cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
716
717 if (event->hw.constraint->flags & PERF_X86_EVENT_PEBS_LDLAT)
718 cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
719 else if (event->hw.constraint->flags & PERF_X86_EVENT_PEBS_ST)
720 cpuc->pebs_enabled &= ~(1ULL << 63);
721
722 if (cpuc->enabled)
723 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
724
725 hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
726 }
727
728 void intel_pmu_pebs_enable_all(void)
729 {
730 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
731
732 if (cpuc->pebs_enabled)
733 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
734 }
735
736 void intel_pmu_pebs_disable_all(void)
737 {
738 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
739
740 if (cpuc->pebs_enabled)
741 wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
742 }
743
744 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
745 {
746 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
747 unsigned long from = cpuc->lbr_entries[0].from;
748 unsigned long old_to, to = cpuc->lbr_entries[0].to;
749 unsigned long ip = regs->ip;
750 int is_64bit = 0;
751 void *kaddr;
752
753 /*
754 * We don't need to fixup if the PEBS assist is fault like
755 */
756 if (!x86_pmu.intel_cap.pebs_trap)
757 return 1;
758
759 /*
760 * No LBR entry, no basic block, no rewinding
761 */
762 if (!cpuc->lbr_stack.nr || !from || !to)
763 return 0;
764
765 /*
766 * Basic blocks should never cross user/kernel boundaries
767 */
768 if (kernel_ip(ip) != kernel_ip(to))
769 return 0;
770
771 /*
772 * unsigned math, either ip is before the start (impossible) or
773 * the basic block is larger than 1 page (sanity)
774 */
775 if ((ip - to) > PEBS_FIXUP_SIZE)
776 return 0;
777
778 /*
779 * We sampled a branch insn, rewind using the LBR stack
780 */
781 if (ip == to) {
782 set_linear_ip(regs, from);
783 return 1;
784 }
785
786 if (!kernel_ip(ip)) {
787 int size, bytes;
788 u8 *buf = this_cpu_read(insn_buffer);
789
790 size = ip - to; /* Must fit our buffer, see above */
791 bytes = copy_from_user_nmi(buf, (void __user *)to, size);
792 if (bytes != 0)
793 return 0;
794
795 kaddr = buf;
796 } else {
797 kaddr = (void *)to;
798 }
799
800 do {
801 struct insn insn;
802
803 old_to = to;
804
805 #ifdef CONFIG_X86_64
806 is_64bit = kernel_ip(to) || !test_thread_flag(TIF_IA32);
807 #endif
808 insn_init(&insn, kaddr, is_64bit);
809 insn_get_length(&insn);
810
811 to += insn.length;
812 kaddr += insn.length;
813 } while (to < ip);
814
815 if (to == ip) {
816 set_linear_ip(regs, old_to);
817 return 1;
818 }
819
820 /*
821 * Even though we decoded the basic block, the instruction stream
822 * never matched the given IP, either the TO or the IP got corrupted.
823 */
824 return 0;
825 }
826
827 static inline u64 intel_hsw_weight(struct pebs_record_hsw *pebs)
828 {
829 if (pebs->tsx_tuning) {
830 union hsw_tsx_tuning tsx = { .value = pebs->tsx_tuning };
831 return tsx.cycles_last_block;
832 }
833 return 0;
834 }
835
836 static inline u64 intel_hsw_transaction(struct pebs_record_hsw *pebs)
837 {
838 u64 txn = (pebs->tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
839
840 /* For RTM XABORTs also log the abort code from AX */
841 if ((txn & PERF_TXN_TRANSACTION) && (pebs->ax & 1))
842 txn |= ((pebs->ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
843 return txn;
844 }
845
846 static void __intel_pmu_pebs_event(struct perf_event *event,
847 struct pt_regs *iregs, void *__pebs)
848 {
849 /*
850 * We cast to the biggest pebs_record but are careful not to
851 * unconditionally access the 'extra' entries.
852 */
853 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
854 struct pebs_record_hsw *pebs = __pebs;
855 struct perf_sample_data data;
856 struct pt_regs regs;
857 u64 sample_type;
858 int fll, fst;
859
860 if (!intel_pmu_save_and_restart(event))
861 return;
862
863 fll = event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT;
864 fst = event->hw.flags & (PERF_X86_EVENT_PEBS_ST |
865 PERF_X86_EVENT_PEBS_ST_HSW);
866
867 perf_sample_data_init(&data, 0, event->hw.last_period);
868
869 data.period = event->hw.last_period;
870 sample_type = event->attr.sample_type;
871
872 /*
873 * if PEBS-LL or PreciseStore
874 */
875 if (fll || fst) {
876 /*
877 * Use latency for weight (only avail with PEBS-LL)
878 */
879 if (fll && (sample_type & PERF_SAMPLE_WEIGHT))
880 data.weight = pebs->lat;
881
882 /*
883 * data.data_src encodes the data source
884 */
885 if (sample_type & PERF_SAMPLE_DATA_SRC) {
886 if (fll)
887 data.data_src.val = load_latency_data(pebs->dse);
888 else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
889 data.data_src.val =
890 precise_store_data_hsw(pebs->dse);
891 else
892 data.data_src.val = precise_store_data(pebs->dse);
893 }
894 }
895
896 /*
897 * We use the interrupt regs as a base because the PEBS record
898 * does not contain a full regs set, specifically it seems to
899 * lack segment descriptors, which get used by things like
900 * user_mode().
901 *
902 * In the simple case fix up only the IP and BP,SP regs, for
903 * PERF_SAMPLE_IP and PERF_SAMPLE_CALLCHAIN to function properly.
904 * A possible PERF_SAMPLE_REGS will have to transfer all regs.
905 */
906 regs = *iregs;
907 regs.flags = pebs->flags;
908 set_linear_ip(&regs, pebs->ip);
909 regs.bp = pebs->bp;
910 regs.sp = pebs->sp;
911
912 if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format >= 2) {
913 regs.ip = pebs->real_ip;
914 regs.flags |= PERF_EFLAGS_EXACT;
915 } else if (event->attr.precise_ip > 1 && intel_pmu_pebs_fixup_ip(&regs))
916 regs.flags |= PERF_EFLAGS_EXACT;
917 else
918 regs.flags &= ~PERF_EFLAGS_EXACT;
919
920 if ((event->attr.sample_type & PERF_SAMPLE_ADDR) &&
921 x86_pmu.intel_cap.pebs_format >= 1)
922 data.addr = pebs->dla;
923
924 if (x86_pmu.intel_cap.pebs_format >= 2) {
925 /* Only set the TSX weight when no memory weight. */
926 if ((event->attr.sample_type & PERF_SAMPLE_WEIGHT) && !fll)
927 data.weight = intel_hsw_weight(pebs);
928
929 if (event->attr.sample_type & PERF_SAMPLE_TRANSACTION)
930 data.txn = intel_hsw_transaction(pebs);
931 }
932
933 if (has_branch_stack(event))
934 data.br_stack = &cpuc->lbr_stack;
935
936 if (perf_event_overflow(event, &data, &regs))
937 x86_pmu_stop(event, 0);
938 }
939
940 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
941 {
942 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
943 struct debug_store *ds = cpuc->ds;
944 struct perf_event *event = cpuc->events[0]; /* PMC0 only */
945 struct pebs_record_core *at, *top;
946 int n;
947
948 if (!x86_pmu.pebs_active)
949 return;
950
951 at = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
952 top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
953
954 /*
955 * Whatever else happens, drain the thing
956 */
957 ds->pebs_index = ds->pebs_buffer_base;
958
959 if (!test_bit(0, cpuc->active_mask))
960 return;
961
962 WARN_ON_ONCE(!event);
963
964 if (!event->attr.precise_ip)
965 return;
966
967 n = top - at;
968 if (n <= 0)
969 return;
970
971 /*
972 * Should not happen, we program the threshold at 1 and do not
973 * set a reset value.
974 */
975 WARN_ONCE(n > 1, "bad leftover pebs %d\n", n);
976 at += n - 1;
977
978 __intel_pmu_pebs_event(event, iregs, at);
979 }
980
981 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
982 {
983 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
984 struct debug_store *ds = cpuc->ds;
985 struct perf_event *event = NULL;
986 void *at, *top;
987 u64 status = 0;
988 int bit;
989
990 if (!x86_pmu.pebs_active)
991 return;
992
993 at = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
994 top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
995
996 ds->pebs_index = ds->pebs_buffer_base;
997
998 if (unlikely(at > top))
999 return;
1000
1001 /*
1002 * Should not happen, we program the threshold at 1 and do not
1003 * set a reset value.
1004 */
1005 WARN_ONCE(top - at > x86_pmu.max_pebs_events * x86_pmu.pebs_record_size,
1006 "Unexpected number of pebs records %ld\n",
1007 (long)(top - at) / x86_pmu.pebs_record_size);
1008
1009 for (; at < top; at += x86_pmu.pebs_record_size) {
1010 struct pebs_record_nhm *p = at;
1011
1012 for_each_set_bit(bit, (unsigned long *)&p->status,
1013 x86_pmu.max_pebs_events) {
1014 event = cpuc->events[bit];
1015 if (!test_bit(bit, cpuc->active_mask))
1016 continue;
1017
1018 WARN_ON_ONCE(!event);
1019
1020 if (!event->attr.precise_ip)
1021 continue;
1022
1023 if (__test_and_set_bit(bit, (unsigned long *)&status))
1024 continue;
1025
1026 break;
1027 }
1028
1029 if (!event || bit >= x86_pmu.max_pebs_events)
1030 continue;
1031
1032 __intel_pmu_pebs_event(event, iregs, at);
1033 }
1034 }
1035
1036 /*
1037 * BTS, PEBS probe and setup
1038 */
1039
1040 void intel_ds_init(void)
1041 {
1042 /*
1043 * No support for 32bit formats
1044 */
1045 if (!boot_cpu_has(X86_FEATURE_DTES64))
1046 return;
1047
1048 x86_pmu.bts = boot_cpu_has(X86_FEATURE_BTS);
1049 x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
1050 if (x86_pmu.pebs) {
1051 char pebs_type = x86_pmu.intel_cap.pebs_trap ? '+' : '-';
1052 int format = x86_pmu.intel_cap.pebs_format;
1053
1054 switch (format) {
1055 case 0:
1056 printk(KERN_CONT "PEBS fmt0%c, ", pebs_type);
1057 x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
1058 x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
1059 break;
1060
1061 case 1:
1062 printk(KERN_CONT "PEBS fmt1%c, ", pebs_type);
1063 x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
1064 x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1065 break;
1066
1067 case 2:
1068 pr_cont("PEBS fmt2%c, ", pebs_type);
1069 x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
1070 x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1071 break;
1072
1073 default:
1074 printk(KERN_CONT "no PEBS fmt%d%c, ", format, pebs_type);
1075 x86_pmu.pebs = 0;
1076 }
1077 }
1078 }
1079
1080 void perf_restore_debug_store(void)
1081 {
1082 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
1083
1084 if (!x86_pmu.bts && !x86_pmu.pebs)
1085 return;
1086
1087 wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
1088 }
ubuntu-bionic-kernel mirror