]> git.karo-electronics.de Git - karo-tx-linux.git/blob - arch/x86/events/intel/ds.c
2ca4d2d7b5fc5a795490bfda7219e1000334ecc0
[karo-tx-linux.git] / arch / x86 / events / 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 << 4)
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 /* Version for Sandy Bridge and later */
55 static u64 pebs_data_source[] = {
56         P(OP, LOAD) | P(LVL, MISS) | P(LVL, L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
57         OP_LH | P(LVL, L1)  | P(SNOOP, NONE),   /* 0x01: L1 local */
58         OP_LH | P(LVL, LFB) | P(SNOOP, NONE),   /* 0x02: LFB hit */
59         OP_LH | P(LVL, L2)  | P(SNOOP, NONE),   /* 0x03: L2 hit */
60         OP_LH | P(LVL, L3)  | P(SNOOP, NONE),   /* 0x04: L3 hit */
61         OP_LH | P(LVL, L3)  | P(SNOOP, MISS),   /* 0x05: L3 hit, snoop miss */
62         OP_LH | P(LVL, L3)  | P(SNOOP, HIT),    /* 0x06: L3 hit, snoop hit */
63         OP_LH | P(LVL, L3)  | P(SNOOP, HITM),   /* 0x07: L3 hit, snoop hitm */
64         OP_LH | P(LVL, REM_CCE1) | P(SNOOP, HIT),  /* 0x08: L3 miss snoop hit */
65         OP_LH | P(LVL, REM_CCE1) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
66         OP_LH | P(LVL, LOC_RAM)  | P(SNOOP, HIT),  /* 0x0a: L3 miss, shared */
67         OP_LH | P(LVL, REM_RAM1) | P(SNOOP, HIT),  /* 0x0b: L3 miss, shared */
68         OP_LH | P(LVL, LOC_RAM)  | SNOOP_NONE_MISS,/* 0x0c: L3 miss, excl */
69         OP_LH | P(LVL, REM_RAM1) | SNOOP_NONE_MISS,/* 0x0d: L3 miss, excl */
70         OP_LH | P(LVL, IO)  | P(SNOOP, NONE), /* 0x0e: I/O */
71         OP_LH | P(LVL, UNC) | P(SNOOP, NONE), /* 0x0f: uncached */
72 };
73
74 /* Patch up minor differences in the bits */
75 void __init intel_pmu_pebs_data_source_nhm(void)
76 {
77         pebs_data_source[0x05] = OP_LH | P(LVL, L3)  | P(SNOOP, HIT);
78         pebs_data_source[0x06] = OP_LH | P(LVL, L3)  | P(SNOOP, HITM);
79         pebs_data_source[0x07] = OP_LH | P(LVL, L3)  | P(SNOOP, HITM);
80 }
81
82 static u64 precise_store_data(u64 status)
83 {
84         union intel_x86_pebs_dse dse;
85         u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
86
87         dse.val = status;
88
89         /*
90          * bit 4: TLB access
91          * 1 = stored missed 2nd level TLB
92          *
93          * so it either hit the walker or the OS
94          * otherwise hit 2nd level TLB
95          */
96         if (dse.st_stlb_miss)
97                 val |= P(TLB, MISS);
98         else
99                 val |= P(TLB, HIT);
100
101         /*
102          * bit 0: hit L1 data cache
103          * if not set, then all we know is that
104          * it missed L1D
105          */
106         if (dse.st_l1d_hit)
107                 val |= P(LVL, HIT);
108         else
109                 val |= P(LVL, MISS);
110
111         /*
112          * bit 5: Locked prefix
113          */
114         if (dse.st_locked)
115                 val |= P(LOCK, LOCKED);
116
117         return val;
118 }
119
120 static u64 precise_datala_hsw(struct perf_event *event, u64 status)
121 {
122         union perf_mem_data_src dse;
123
124         dse.val = PERF_MEM_NA;
125
126         if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
127                 dse.mem_op = PERF_MEM_OP_STORE;
128         else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
129                 dse.mem_op = PERF_MEM_OP_LOAD;
130
131         /*
132          * L1 info only valid for following events:
133          *
134          * MEM_UOPS_RETIRED.STLB_MISS_STORES
135          * MEM_UOPS_RETIRED.LOCK_STORES
136          * MEM_UOPS_RETIRED.SPLIT_STORES
137          * MEM_UOPS_RETIRED.ALL_STORES
138          */
139         if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
140                 if (status & 1)
141                         dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
142                 else
143                         dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
144         }
145         return dse.val;
146 }
147
148 static u64 load_latency_data(u64 status)
149 {
150         union intel_x86_pebs_dse dse;
151         u64 val;
152         int model = boot_cpu_data.x86_model;
153         int fam = boot_cpu_data.x86;
154
155         dse.val = status;
156
157         /*
158          * use the mapping table for bit 0-3
159          */
160         val = pebs_data_source[dse.ld_dse];
161
162         /*
163          * Nehalem models do not support TLB, Lock infos
164          */
165         if (fam == 0x6 && (model == 26 || model == 30
166             || model == 31 || model == 46)) {
167                 val |= P(TLB, NA) | P(LOCK, NA);
168                 return val;
169         }
170         /*
171          * bit 4: TLB access
172          * 0 = did not miss 2nd level TLB
173          * 1 = missed 2nd level TLB
174          */
175         if (dse.ld_stlb_miss)
176                 val |= P(TLB, MISS) | P(TLB, L2);
177         else
178                 val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
179
180         /*
181          * bit 5: locked prefix
182          */
183         if (dse.ld_locked)
184                 val |= P(LOCK, LOCKED);
185
186         return val;
187 }
188
189 struct pebs_record_core {
190         u64 flags, ip;
191         u64 ax, bx, cx, dx;
192         u64 si, di, bp, sp;
193         u64 r8,  r9,  r10, r11;
194         u64 r12, r13, r14, r15;
195 };
196
197 struct pebs_record_nhm {
198         u64 flags, ip;
199         u64 ax, bx, cx, dx;
200         u64 si, di, bp, sp;
201         u64 r8,  r9,  r10, r11;
202         u64 r12, r13, r14, r15;
203         u64 status, dla, dse, lat;
204 };
205
206 /*
207  * Same as pebs_record_nhm, with two additional fields.
208  */
209 struct pebs_record_hsw {
210         u64 flags, ip;
211         u64 ax, bx, cx, dx;
212         u64 si, di, bp, sp;
213         u64 r8,  r9,  r10, r11;
214         u64 r12, r13, r14, r15;
215         u64 status, dla, dse, lat;
216         u64 real_ip, tsx_tuning;
217 };
218
219 union hsw_tsx_tuning {
220         struct {
221                 u32 cycles_last_block     : 32,
222                     hle_abort             : 1,
223                     rtm_abort             : 1,
224                     instruction_abort     : 1,
225                     non_instruction_abort : 1,
226                     retry                 : 1,
227                     data_conflict         : 1,
228                     capacity_writes       : 1,
229                     capacity_reads        : 1;
230         };
231         u64         value;
232 };
233
234 #define PEBS_HSW_TSX_FLAGS      0xff00000000ULL
235
236 /* Same as HSW, plus TSC */
237
238 struct pebs_record_skl {
239         u64 flags, ip;
240         u64 ax, bx, cx, dx;
241         u64 si, di, bp, sp;
242         u64 r8,  r9,  r10, r11;
243         u64 r12, r13, r14, r15;
244         u64 status, dla, dse, lat;
245         u64 real_ip, tsx_tuning;
246         u64 tsc;
247 };
248
249 void init_debug_store_on_cpu(int cpu)
250 {
251         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
252
253         if (!ds)
254                 return;
255
256         wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
257                      (u32)((u64)(unsigned long)ds),
258                      (u32)((u64)(unsigned long)ds >> 32));
259 }
260
261 void fini_debug_store_on_cpu(int cpu)
262 {
263         if (!per_cpu(cpu_hw_events, cpu).ds)
264                 return;
265
266         wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
267 }
268
269 static DEFINE_PER_CPU(void *, insn_buffer);
270
271 static int alloc_pebs_buffer(int cpu)
272 {
273         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
274         int node = cpu_to_node(cpu);
275         int max;
276         void *buffer, *ibuffer;
277
278         if (!x86_pmu.pebs)
279                 return 0;
280
281         buffer = kzalloc_node(x86_pmu.pebs_buffer_size, GFP_KERNEL, node);
282         if (unlikely(!buffer))
283                 return -ENOMEM;
284
285         /*
286          * HSW+ already provides us the eventing ip; no need to allocate this
287          * buffer then.
288          */
289         if (x86_pmu.intel_cap.pebs_format < 2) {
290                 ibuffer = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
291                 if (!ibuffer) {
292                         kfree(buffer);
293                         return -ENOMEM;
294                 }
295                 per_cpu(insn_buffer, cpu) = ibuffer;
296         }
297
298         max = x86_pmu.pebs_buffer_size / x86_pmu.pebs_record_size;
299
300         ds->pebs_buffer_base = (u64)(unsigned long)buffer;
301         ds->pebs_index = ds->pebs_buffer_base;
302         ds->pebs_absolute_maximum = ds->pebs_buffer_base +
303                 max * x86_pmu.pebs_record_size;
304
305         return 0;
306 }
307
308 static void release_pebs_buffer(int cpu)
309 {
310         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
311
312         if (!ds || !x86_pmu.pebs)
313                 return;
314
315         kfree(per_cpu(insn_buffer, cpu));
316         per_cpu(insn_buffer, cpu) = NULL;
317
318         kfree((void *)(unsigned long)ds->pebs_buffer_base);
319         ds->pebs_buffer_base = 0;
320 }
321
322 static int alloc_bts_buffer(int cpu)
323 {
324         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
325         int node = cpu_to_node(cpu);
326         int max, thresh;
327         void *buffer;
328
329         if (!x86_pmu.bts)
330                 return 0;
331
332         buffer = kzalloc_node(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, node);
333         if (unlikely(!buffer)) {
334                 WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
335                 return -ENOMEM;
336         }
337
338         max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
339         thresh = max / 16;
340
341         ds->bts_buffer_base = (u64)(unsigned long)buffer;
342         ds->bts_index = ds->bts_buffer_base;
343         ds->bts_absolute_maximum = ds->bts_buffer_base +
344                 max * BTS_RECORD_SIZE;
345         ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
346                 thresh * BTS_RECORD_SIZE;
347
348         return 0;
349 }
350
351 static void release_bts_buffer(int cpu)
352 {
353         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
354
355         if (!ds || !x86_pmu.bts)
356                 return;
357
358         kfree((void *)(unsigned long)ds->bts_buffer_base);
359         ds->bts_buffer_base = 0;
360 }
361
362 static int alloc_ds_buffer(int cpu)
363 {
364         int node = cpu_to_node(cpu);
365         struct debug_store *ds;
366
367         ds = kzalloc_node(sizeof(*ds), GFP_KERNEL, node);
368         if (unlikely(!ds))
369                 return -ENOMEM;
370
371         per_cpu(cpu_hw_events, cpu).ds = ds;
372
373         return 0;
374 }
375
376 static void release_ds_buffer(int cpu)
377 {
378         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
379
380         if (!ds)
381                 return;
382
383         per_cpu(cpu_hw_events, cpu).ds = NULL;
384         kfree(ds);
385 }
386
387 void release_ds_buffers(void)
388 {
389         int cpu;
390
391         if (!x86_pmu.bts && !x86_pmu.pebs)
392                 return;
393
394         get_online_cpus();
395         for_each_online_cpu(cpu)
396                 fini_debug_store_on_cpu(cpu);
397
398         for_each_possible_cpu(cpu) {
399                 release_pebs_buffer(cpu);
400                 release_bts_buffer(cpu);
401                 release_ds_buffer(cpu);
402         }
403         put_online_cpus();
404 }
405
406 void reserve_ds_buffers(void)
407 {
408         int bts_err = 0, pebs_err = 0;
409         int cpu;
410
411         x86_pmu.bts_active = 0;
412         x86_pmu.pebs_active = 0;
413
414         if (!x86_pmu.bts && !x86_pmu.pebs)
415                 return;
416
417         if (!x86_pmu.bts)
418                 bts_err = 1;
419
420         if (!x86_pmu.pebs)
421                 pebs_err = 1;
422
423         get_online_cpus();
424
425         for_each_possible_cpu(cpu) {
426                 if (alloc_ds_buffer(cpu)) {
427                         bts_err = 1;
428                         pebs_err = 1;
429                 }
430
431                 if (!bts_err && alloc_bts_buffer(cpu))
432                         bts_err = 1;
433
434                 if (!pebs_err && alloc_pebs_buffer(cpu))
435                         pebs_err = 1;
436
437                 if (bts_err && pebs_err)
438                         break;
439         }
440
441         if (bts_err) {
442                 for_each_possible_cpu(cpu)
443                         release_bts_buffer(cpu);
444         }
445
446         if (pebs_err) {
447                 for_each_possible_cpu(cpu)
448                         release_pebs_buffer(cpu);
449         }
450
451         if (bts_err && pebs_err) {
452                 for_each_possible_cpu(cpu)
453                         release_ds_buffer(cpu);
454         } else {
455                 if (x86_pmu.bts && !bts_err)
456                         x86_pmu.bts_active = 1;
457
458                 if (x86_pmu.pebs && !pebs_err)
459                         x86_pmu.pebs_active = 1;
460
461                 for_each_online_cpu(cpu)
462                         init_debug_store_on_cpu(cpu);
463         }
464
465         put_online_cpus();
466 }
467
468 /*
469  * BTS
470  */
471
472 struct event_constraint bts_constraint =
473         EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
474
475 void intel_pmu_enable_bts(u64 config)
476 {
477         unsigned long debugctlmsr;
478
479         debugctlmsr = get_debugctlmsr();
480
481         debugctlmsr |= DEBUGCTLMSR_TR;
482         debugctlmsr |= DEBUGCTLMSR_BTS;
483         if (config & ARCH_PERFMON_EVENTSEL_INT)
484                 debugctlmsr |= DEBUGCTLMSR_BTINT;
485
486         if (!(config & ARCH_PERFMON_EVENTSEL_OS))
487                 debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
488
489         if (!(config & ARCH_PERFMON_EVENTSEL_USR))
490                 debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
491
492         update_debugctlmsr(debugctlmsr);
493 }
494
495 void intel_pmu_disable_bts(void)
496 {
497         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
498         unsigned long debugctlmsr;
499
500         if (!cpuc->ds)
501                 return;
502
503         debugctlmsr = get_debugctlmsr();
504
505         debugctlmsr &=
506                 ~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
507                   DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
508
509         update_debugctlmsr(debugctlmsr);
510 }
511
512 int intel_pmu_drain_bts_buffer(void)
513 {
514         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
515         struct debug_store *ds = cpuc->ds;
516         struct bts_record {
517                 u64     from;
518                 u64     to;
519                 u64     flags;
520         };
521         struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
522         struct bts_record *at, *base, *top;
523         struct perf_output_handle handle;
524         struct perf_event_header header;
525         struct perf_sample_data data;
526         unsigned long skip = 0;
527         struct pt_regs regs;
528
529         if (!event)
530                 return 0;
531
532         if (!x86_pmu.bts_active)
533                 return 0;
534
535         base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
536         top  = (struct bts_record *)(unsigned long)ds->bts_index;
537
538         if (top <= base)
539                 return 0;
540
541         memset(&regs, 0, sizeof(regs));
542
543         ds->bts_index = ds->bts_buffer_base;
544
545         perf_sample_data_init(&data, 0, event->hw.last_period);
546
547         /*
548          * BTS leaks kernel addresses in branches across the cpl boundary,
549          * such as traps or system calls, so unless the user is asking for
550          * kernel tracing (and right now it's not possible), we'd need to
551          * filter them out. But first we need to count how many of those we
552          * have in the current batch. This is an extra O(n) pass, however,
553          * it's much faster than the other one especially considering that
554          * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
555          * alloc_bts_buffer()).
556          */
557         for (at = base; at < top; at++) {
558                 /*
559                  * Note that right now *this* BTS code only works if
560                  * attr::exclude_kernel is set, but let's keep this extra
561                  * check here in case that changes.
562                  */
563                 if (event->attr.exclude_kernel &&
564                     (kernel_ip(at->from) || kernel_ip(at->to)))
565                         skip++;
566         }
567
568         /*
569          * Prepare a generic sample, i.e. fill in the invariant fields.
570          * We will overwrite the from and to address before we output
571          * the sample.
572          */
573         rcu_read_lock();
574         perf_prepare_sample(&header, &data, event, &regs);
575
576         if (perf_output_begin(&handle, event, header.size *
577                               (top - base - skip)))
578                 goto unlock;
579
580         for (at = base; at < top; at++) {
581                 /* Filter out any records that contain kernel addresses. */
582                 if (event->attr.exclude_kernel &&
583                     (kernel_ip(at->from) || kernel_ip(at->to)))
584                         continue;
585
586                 data.ip         = at->from;
587                 data.addr       = at->to;
588
589                 perf_output_sample(&handle, &header, &data, event);
590         }
591
592         perf_output_end(&handle);
593
594         /* There's new data available. */
595         event->hw.interrupts++;
596         event->pending_kill = POLL_IN;
597 unlock:
598         rcu_read_unlock();
599         return 1;
600 }
601
602 static inline void intel_pmu_drain_pebs_buffer(void)
603 {
604         struct pt_regs regs;
605
606         x86_pmu.drain_pebs(&regs);
607 }
608
609 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in)
610 {
611         if (!sched_in)
612                 intel_pmu_drain_pebs_buffer();
613 }
614
615 /*
616  * PEBS
617  */
618 struct event_constraint intel_core2_pebs_event_constraints[] = {
619         INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
620         INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
621         INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
622         INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
623         INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
624         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
625         INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
626         EVENT_CONSTRAINT_END
627 };
628
629 struct event_constraint intel_atom_pebs_event_constraints[] = {
630         INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
631         INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
632         INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
633         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
634         INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
635         /* Allow all events as PEBS with no flags */
636         INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
637         EVENT_CONSTRAINT_END
638 };
639
640 struct event_constraint intel_slm_pebs_event_constraints[] = {
641         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
642         INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x1),
643         /* Allow all events as PEBS with no flags */
644         INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
645         EVENT_CONSTRAINT_END
646 };
647
648 struct event_constraint intel_glm_pebs_event_constraints[] = {
649         /* Allow all events as PEBS with no flags */
650         INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
651         EVENT_CONSTRAINT_END
652 };
653
654 struct event_constraint intel_glp_pebs_event_constraints[] = {
655         /* Allow all events as PEBS with no flags */
656         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
657         EVENT_CONSTRAINT_END
658 };
659
660 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
661         INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
662         INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
663         INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
664         INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
665         INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
666         INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
667         INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
668         INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
669         INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
670         INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
671         INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
672         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
673         INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
674         EVENT_CONSTRAINT_END
675 };
676
677 struct event_constraint intel_westmere_pebs_event_constraints[] = {
678         INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
679         INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
680         INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
681         INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
682         INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
683         INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
684         INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
685         INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
686         INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
687         INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
688         INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
689         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
690         INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
691         EVENT_CONSTRAINT_END
692 };
693
694 struct event_constraint intel_snb_pebs_event_constraints[] = {
695         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
696         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
697         INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
698         /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
699         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
700         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
701         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
702         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
703         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
704         /* Allow all events as PEBS with no flags */
705         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
706         EVENT_CONSTRAINT_END
707 };
708
709 struct event_constraint intel_ivb_pebs_event_constraints[] = {
710         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
711         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
712         INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
713         /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
714         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
715         /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
716         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
717         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
718         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
719         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
720         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
721         /* Allow all events as PEBS with no flags */
722         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
723         EVENT_CONSTRAINT_END
724 };
725
726 struct event_constraint intel_hsw_pebs_event_constraints[] = {
727         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
728         INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
729         /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
730         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
731         /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
732         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
733         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
734         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
735         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
736         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
737         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
738         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
739         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
740         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
741         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
742         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
743         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
744         /* Allow all events as PEBS with no flags */
745         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
746         EVENT_CONSTRAINT_END
747 };
748
749 struct event_constraint intel_bdw_pebs_event_constraints[] = {
750         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
751         INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
752         /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
753         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
754         /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
755         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
756         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
757         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
758         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
759         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
760         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
761         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
762         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
763         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
764         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
765         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
766         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
767         /* Allow all events as PEBS with no flags */
768         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
769         EVENT_CONSTRAINT_END
770 };
771
772
773 struct event_constraint intel_skl_pebs_event_constraints[] = {
774         INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),      /* INST_RETIRED.PREC_DIST */
775         /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
776         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
777         /* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
778         INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
779         INTEL_PLD_CONSTRAINT(0x1cd, 0xf),                     /* MEM_TRANS_RETIRED.* */
780         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
781         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
782         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
783         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
784         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
785         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
786         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
787         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
788         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
789         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
790         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
791         /* Allow all events as PEBS with no flags */
792         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
793         EVENT_CONSTRAINT_END
794 };
795
796 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
797 {
798         struct event_constraint *c;
799
800         if (!event->attr.precise_ip)
801                 return NULL;
802
803         if (x86_pmu.pebs_constraints) {
804                 for_each_event_constraint(c, x86_pmu.pebs_constraints) {
805                         if ((event->hw.config & c->cmask) == c->code) {
806                                 event->hw.flags |= c->flags;
807                                 return c;
808                         }
809                 }
810         }
811
812         return &emptyconstraint;
813 }
814
815 /*
816  * We need the sched_task callback even for per-cpu events when we use
817  * the large interrupt threshold, such that we can provide PID and TID
818  * to PEBS samples.
819  */
820 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
821 {
822         return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
823 }
824
825 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
826 {
827         struct debug_store *ds = cpuc->ds;
828         u64 threshold;
829
830         if (cpuc->n_pebs == cpuc->n_large_pebs) {
831                 threshold = ds->pebs_absolute_maximum -
832                         x86_pmu.max_pebs_events * x86_pmu.pebs_record_size;
833         } else {
834                 threshold = ds->pebs_buffer_base + x86_pmu.pebs_record_size;
835         }
836
837         ds->pebs_interrupt_threshold = threshold;
838 }
839
840 static void
841 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc, struct pmu *pmu)
842 {
843         /*
844          * Make sure we get updated with the first PEBS
845          * event. It will trigger also during removal, but
846          * that does not hurt:
847          */
848         bool update = cpuc->n_pebs == 1;
849
850         if (needed_cb != pebs_needs_sched_cb(cpuc)) {
851                 if (!needed_cb)
852                         perf_sched_cb_inc(pmu);
853                 else
854                         perf_sched_cb_dec(pmu);
855
856                 update = true;
857         }
858
859         if (update)
860                 pebs_update_threshold(cpuc);
861 }
862
863 void intel_pmu_pebs_add(struct perf_event *event)
864 {
865         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
866         struct hw_perf_event *hwc = &event->hw;
867         bool needed_cb = pebs_needs_sched_cb(cpuc);
868
869         cpuc->n_pebs++;
870         if (hwc->flags & PERF_X86_EVENT_FREERUNNING)
871                 cpuc->n_large_pebs++;
872
873         pebs_update_state(needed_cb, cpuc, event->ctx->pmu);
874 }
875
876 void intel_pmu_pebs_enable(struct perf_event *event)
877 {
878         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
879         struct hw_perf_event *hwc = &event->hw;
880         struct debug_store *ds = cpuc->ds;
881
882         hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
883
884         cpuc->pebs_enabled |= 1ULL << hwc->idx;
885
886         if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
887                 cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
888         else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
889                 cpuc->pebs_enabled |= 1ULL << 63;
890
891         /*
892          * Use auto-reload if possible to save a MSR write in the PMI.
893          * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
894          */
895         if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
896                 ds->pebs_event_reset[hwc->idx] =
897                         (u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
898         }
899 }
900
901 void intel_pmu_pebs_del(struct perf_event *event)
902 {
903         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
904         struct hw_perf_event *hwc = &event->hw;
905         bool needed_cb = pebs_needs_sched_cb(cpuc);
906
907         cpuc->n_pebs--;
908         if (hwc->flags & PERF_X86_EVENT_FREERUNNING)
909                 cpuc->n_large_pebs--;
910
911         pebs_update_state(needed_cb, cpuc, event->ctx->pmu);
912 }
913
914 void intel_pmu_pebs_disable(struct perf_event *event)
915 {
916         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
917         struct hw_perf_event *hwc = &event->hw;
918
919         if (cpuc->n_pebs == cpuc->n_large_pebs)
920                 intel_pmu_drain_pebs_buffer();
921
922         cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
923
924         if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
925                 cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
926         else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
927                 cpuc->pebs_enabled &= ~(1ULL << 63);
928
929         if (cpuc->enabled)
930                 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
931
932         hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
933 }
934
935 void intel_pmu_pebs_enable_all(void)
936 {
937         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
938
939         if (cpuc->pebs_enabled)
940                 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
941 }
942
943 void intel_pmu_pebs_disable_all(void)
944 {
945         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
946
947         if (cpuc->pebs_enabled)
948                 wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
949 }
950
951 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
952 {
953         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
954         unsigned long from = cpuc->lbr_entries[0].from;
955         unsigned long old_to, to = cpuc->lbr_entries[0].to;
956         unsigned long ip = regs->ip;
957         int is_64bit = 0;
958         void *kaddr;
959         int size;
960
961         /*
962          * We don't need to fixup if the PEBS assist is fault like
963          */
964         if (!x86_pmu.intel_cap.pebs_trap)
965                 return 1;
966
967         /*
968          * No LBR entry, no basic block, no rewinding
969          */
970         if (!cpuc->lbr_stack.nr || !from || !to)
971                 return 0;
972
973         /*
974          * Basic blocks should never cross user/kernel boundaries
975          */
976         if (kernel_ip(ip) != kernel_ip(to))
977                 return 0;
978
979         /*
980          * unsigned math, either ip is before the start (impossible) or
981          * the basic block is larger than 1 page (sanity)
982          */
983         if ((ip - to) > PEBS_FIXUP_SIZE)
984                 return 0;
985
986         /*
987          * We sampled a branch insn, rewind using the LBR stack
988          */
989         if (ip == to) {
990                 set_linear_ip(regs, from);
991                 return 1;
992         }
993
994         size = ip - to;
995         if (!kernel_ip(ip)) {
996                 int bytes;
997                 u8 *buf = this_cpu_read(insn_buffer);
998
999                 /* 'size' must fit our buffer, see above */
1000                 bytes = copy_from_user_nmi(buf, (void __user *)to, size);
1001                 if (bytes != 0)
1002                         return 0;
1003
1004                 kaddr = buf;
1005         } else {
1006                 kaddr = (void *)to;
1007         }
1008
1009         do {
1010                 struct insn insn;
1011
1012                 old_to = to;
1013
1014 #ifdef CONFIG_X86_64
1015                 is_64bit = kernel_ip(to) || !test_thread_flag(TIF_IA32);
1016 #endif
1017                 insn_init(&insn, kaddr, size, is_64bit);
1018                 insn_get_length(&insn);
1019                 /*
1020                  * Make sure there was not a problem decoding the
1021                  * instruction and getting the length.  This is
1022                  * doubly important because we have an infinite
1023                  * loop if insn.length=0.
1024                  */
1025                 if (!insn.length)
1026                         break;
1027
1028                 to += insn.length;
1029                 kaddr += insn.length;
1030                 size -= insn.length;
1031         } while (to < ip);
1032
1033         if (to == ip) {
1034                 set_linear_ip(regs, old_to);
1035                 return 1;
1036         }
1037
1038         /*
1039          * Even though we decoded the basic block, the instruction stream
1040          * never matched the given IP, either the TO or the IP got corrupted.
1041          */
1042         return 0;
1043 }
1044
1045 static inline u64 intel_hsw_weight(struct pebs_record_skl *pebs)
1046 {
1047         if (pebs->tsx_tuning) {
1048                 union hsw_tsx_tuning tsx = { .value = pebs->tsx_tuning };
1049                 return tsx.cycles_last_block;
1050         }
1051         return 0;
1052 }
1053
1054 static inline u64 intel_hsw_transaction(struct pebs_record_skl *pebs)
1055 {
1056         u64 txn = (pebs->tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1057
1058         /* For RTM XABORTs also log the abort code from AX */
1059         if ((txn & PERF_TXN_TRANSACTION) && (pebs->ax & 1))
1060                 txn |= ((pebs->ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1061         return txn;
1062 }
1063
1064 static void setup_pebs_sample_data(struct perf_event *event,
1065                                    struct pt_regs *iregs, void *__pebs,
1066                                    struct perf_sample_data *data,
1067                                    struct pt_regs *regs)
1068 {
1069 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1070                 (PERF_X86_EVENT_PEBS_ST_HSW | \
1071                  PERF_X86_EVENT_PEBS_LD_HSW | \
1072                  PERF_X86_EVENT_PEBS_NA_HSW)
1073         /*
1074          * We cast to the biggest pebs_record but are careful not to
1075          * unconditionally access the 'extra' entries.
1076          */
1077         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1078         struct pebs_record_skl *pebs = __pebs;
1079         u64 sample_type;
1080         int fll, fst, dsrc;
1081         int fl = event->hw.flags;
1082
1083         if (pebs == NULL)
1084                 return;
1085
1086         sample_type = event->attr.sample_type;
1087         dsrc = sample_type & PERF_SAMPLE_DATA_SRC;
1088
1089         fll = fl & PERF_X86_EVENT_PEBS_LDLAT;
1090         fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1091
1092         perf_sample_data_init(data, 0, event->hw.last_period);
1093
1094         data->period = event->hw.last_period;
1095
1096         /*
1097          * Use latency for weight (only avail with PEBS-LL)
1098          */
1099         if (fll && (sample_type & PERF_SAMPLE_WEIGHT))
1100                 data->weight = pebs->lat;
1101
1102         /*
1103          * data.data_src encodes the data source
1104          */
1105         if (dsrc) {
1106                 u64 val = PERF_MEM_NA;
1107                 if (fll)
1108                         val = load_latency_data(pebs->dse);
1109                 else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1110                         val = precise_datala_hsw(event, pebs->dse);
1111                 else if (fst)
1112                         val = precise_store_data(pebs->dse);
1113                 data->data_src.val = val;
1114         }
1115
1116         /*
1117          * We use the interrupt regs as a base because the PEBS record does not
1118          * contain a full regs set, specifically it seems to lack segment
1119          * descriptors, which get used by things like user_mode().
1120          *
1121          * In the simple case fix up only the IP for PERF_SAMPLE_IP.
1122          *
1123          * We must however always use BP,SP from iregs for the unwinder to stay
1124          * sane; the record BP,SP can point into thin air when the record is
1125          * from a previous PMI context or an (I)RET happend between the record
1126          * and PMI.
1127          */
1128         *regs = *iregs;
1129         regs->flags = pebs->flags;
1130         set_linear_ip(regs, pebs->ip);
1131
1132         if (sample_type & PERF_SAMPLE_REGS_INTR) {
1133                 regs->ax = pebs->ax;
1134                 regs->bx = pebs->bx;
1135                 regs->cx = pebs->cx;
1136                 regs->dx = pebs->dx;
1137                 regs->si = pebs->si;
1138                 regs->di = pebs->di;
1139
1140                 /*
1141                  * Per the above; only set BP,SP if we don't need callchains.
1142                  *
1143                  * XXX: does this make sense?
1144                  */
1145                 if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
1146                         regs->bp = pebs->bp;
1147                         regs->sp = pebs->sp;
1148                 }
1149
1150                 /*
1151                  * Preserve PERF_EFLAGS_VM from set_linear_ip().
1152                  */
1153                 regs->flags = pebs->flags | (regs->flags & PERF_EFLAGS_VM);
1154 #ifndef CONFIG_X86_32
1155                 regs->r8 = pebs->r8;
1156                 regs->r9 = pebs->r9;
1157                 regs->r10 = pebs->r10;
1158                 regs->r11 = pebs->r11;
1159                 regs->r12 = pebs->r12;
1160                 regs->r13 = pebs->r13;
1161                 regs->r14 = pebs->r14;
1162                 regs->r15 = pebs->r15;
1163 #endif
1164         }
1165
1166         if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format >= 2) {
1167                 regs->ip = pebs->real_ip;
1168                 regs->flags |= PERF_EFLAGS_EXACT;
1169         } else if (event->attr.precise_ip > 1 && intel_pmu_pebs_fixup_ip(regs))
1170                 regs->flags |= PERF_EFLAGS_EXACT;
1171         else
1172                 regs->flags &= ~PERF_EFLAGS_EXACT;
1173
1174         if ((sample_type & PERF_SAMPLE_ADDR) &&
1175             x86_pmu.intel_cap.pebs_format >= 1)
1176                 data->addr = pebs->dla;
1177
1178         if (x86_pmu.intel_cap.pebs_format >= 2) {
1179                 /* Only set the TSX weight when no memory weight. */
1180                 if ((sample_type & PERF_SAMPLE_WEIGHT) && !fll)
1181                         data->weight = intel_hsw_weight(pebs);
1182
1183                 if (sample_type & PERF_SAMPLE_TRANSACTION)
1184                         data->txn = intel_hsw_transaction(pebs);
1185         }
1186
1187         /*
1188          * v3 supplies an accurate time stamp, so we use that
1189          * for the time stamp.
1190          *
1191          * We can only do this for the default trace clock.
1192          */
1193         if (x86_pmu.intel_cap.pebs_format >= 3 &&
1194                 event->attr.use_clockid == 0)
1195                 data->time = native_sched_clock_from_tsc(pebs->tsc);
1196
1197         if (has_branch_stack(event))
1198                 data->br_stack = &cpuc->lbr_stack;
1199 }
1200
1201 static inline void *
1202 get_next_pebs_record_by_bit(void *base, void *top, int bit)
1203 {
1204         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1205         void *at;
1206         u64 pebs_status;
1207
1208         /*
1209          * fmt0 does not have a status bitfield (does not use
1210          * perf_record_nhm format)
1211          */
1212         if (x86_pmu.intel_cap.pebs_format < 1)
1213                 return base;
1214
1215         if (base == NULL)
1216                 return NULL;
1217
1218         for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1219                 struct pebs_record_nhm *p = at;
1220
1221                 if (test_bit(bit, (unsigned long *)&p->status)) {
1222                         /* PEBS v3 has accurate status bits */
1223                         if (x86_pmu.intel_cap.pebs_format >= 3)
1224                                 return at;
1225
1226                         if (p->status == (1 << bit))
1227                                 return at;
1228
1229                         /* clear non-PEBS bit and re-check */
1230                         pebs_status = p->status & cpuc->pebs_enabled;
1231                         pebs_status &= PEBS_COUNTER_MASK;
1232                         if (pebs_status == (1 << bit))
1233                                 return at;
1234                 }
1235         }
1236         return NULL;
1237 }
1238
1239 static void __intel_pmu_pebs_event(struct perf_event *event,
1240                                    struct pt_regs *iregs,
1241                                    void *base, void *top,
1242                                    int bit, int count)
1243 {
1244         struct perf_sample_data data;
1245         struct pt_regs regs;
1246         void *at = get_next_pebs_record_by_bit(base, top, bit);
1247
1248         if (!intel_pmu_save_and_restart(event) &&
1249             !(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
1250                 return;
1251
1252         while (count > 1) {
1253                 setup_pebs_sample_data(event, iregs, at, &data, &regs);
1254                 perf_event_output(event, &data, &regs);
1255                 at += x86_pmu.pebs_record_size;
1256                 at = get_next_pebs_record_by_bit(at, top, bit);
1257                 count--;
1258         }
1259
1260         setup_pebs_sample_data(event, iregs, at, &data, &regs);
1261
1262         /*
1263          * All but the last records are processed.
1264          * The last one is left to be able to call the overflow handler.
1265          */
1266         if (perf_event_overflow(event, &data, &regs)) {
1267                 x86_pmu_stop(event, 0);
1268                 return;
1269         }
1270
1271 }
1272
1273 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
1274 {
1275         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1276         struct debug_store *ds = cpuc->ds;
1277         struct perf_event *event = cpuc->events[0]; /* PMC0 only */
1278         struct pebs_record_core *at, *top;
1279         int n;
1280
1281         if (!x86_pmu.pebs_active)
1282                 return;
1283
1284         at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
1285         top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
1286
1287         /*
1288          * Whatever else happens, drain the thing
1289          */
1290         ds->pebs_index = ds->pebs_buffer_base;
1291
1292         if (!test_bit(0, cpuc->active_mask))
1293                 return;
1294
1295         WARN_ON_ONCE(!event);
1296
1297         if (!event->attr.precise_ip)
1298                 return;
1299
1300         n = top - at;
1301         if (n <= 0)
1302                 return;
1303
1304         __intel_pmu_pebs_event(event, iregs, at, top, 0, n);
1305 }
1306
1307 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
1308 {
1309         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1310         struct debug_store *ds = cpuc->ds;
1311         struct perf_event *event;
1312         void *base, *at, *top;
1313         short counts[MAX_PEBS_EVENTS] = {};
1314         short error[MAX_PEBS_EVENTS] = {};
1315         int bit, i;
1316
1317         if (!x86_pmu.pebs_active)
1318                 return;
1319
1320         base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
1321         top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
1322
1323         ds->pebs_index = ds->pebs_buffer_base;
1324
1325         if (unlikely(base >= top))
1326                 return;
1327
1328         for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1329                 struct pebs_record_nhm *p = at;
1330                 u64 pebs_status;
1331
1332                 pebs_status = p->status & cpuc->pebs_enabled;
1333                 pebs_status &= (1ULL << x86_pmu.max_pebs_events) - 1;
1334
1335                 /* PEBS v3 has more accurate status bits */
1336                 if (x86_pmu.intel_cap.pebs_format >= 3) {
1337                         for_each_set_bit(bit, (unsigned long *)&pebs_status,
1338                                          x86_pmu.max_pebs_events)
1339                                 counts[bit]++;
1340
1341                         continue;
1342                 }
1343
1344                 /*
1345                  * On some CPUs the PEBS status can be zero when PEBS is
1346                  * racing with clearing of GLOBAL_STATUS.
1347                  *
1348                  * Normally we would drop that record, but in the
1349                  * case when there is only a single active PEBS event
1350                  * we can assume it's for that event.
1351                  */
1352                 if (!pebs_status && cpuc->pebs_enabled &&
1353                         !(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
1354                         pebs_status = cpuc->pebs_enabled;
1355
1356                 bit = find_first_bit((unsigned long *)&pebs_status,
1357                                         x86_pmu.max_pebs_events);
1358                 if (bit >= x86_pmu.max_pebs_events)
1359                         continue;
1360
1361                 /*
1362                  * The PEBS hardware does not deal well with the situation
1363                  * when events happen near to each other and multiple bits
1364                  * are set. But it should happen rarely.
1365                  *
1366                  * If these events include one PEBS and multiple non-PEBS
1367                  * events, it doesn't impact PEBS record. The record will
1368                  * be handled normally. (slow path)
1369                  *
1370                  * If these events include two or more PEBS events, the
1371                  * records for the events can be collapsed into a single
1372                  * one, and it's not possible to reconstruct all events
1373                  * that caused the PEBS record. It's called collision.
1374                  * If collision happened, the record will be dropped.
1375                  */
1376                 if (p->status != (1ULL << bit)) {
1377                         for_each_set_bit(i, (unsigned long *)&pebs_status,
1378                                          x86_pmu.max_pebs_events)
1379                                 error[i]++;
1380                         continue;
1381                 }
1382
1383                 counts[bit]++;
1384         }
1385
1386         for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) {
1387                 if ((counts[bit] == 0) && (error[bit] == 0))
1388                         continue;
1389
1390                 event = cpuc->events[bit];
1391                 if (WARN_ON_ONCE(!event))
1392                         continue;
1393
1394                 if (WARN_ON_ONCE(!event->attr.precise_ip))
1395                         continue;
1396
1397                 /* log dropped samples number */
1398                 if (error[bit]) {
1399                         perf_log_lost_samples(event, error[bit]);
1400
1401                         if (perf_event_account_interrupt(event))
1402                                 x86_pmu_stop(event, 0);
1403                 }
1404
1405                 if (counts[bit]) {
1406                         __intel_pmu_pebs_event(event, iregs, base,
1407                                                top, bit, counts[bit]);
1408                 }
1409         }
1410 }
1411
1412 /*
1413  * BTS, PEBS probe and setup
1414  */
1415
1416 void __init intel_ds_init(void)
1417 {
1418         /*
1419          * No support for 32bit formats
1420          */
1421         if (!boot_cpu_has(X86_FEATURE_DTES64))
1422                 return;
1423
1424         x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
1425         x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
1426         x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
1427         if (x86_pmu.pebs) {
1428                 char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
1429                 int format = x86_pmu.intel_cap.pebs_format;
1430
1431                 switch (format) {
1432                 case 0:
1433                         pr_cont("PEBS fmt0%c, ", pebs_type);
1434                         x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
1435                         /*
1436                          * Using >PAGE_SIZE buffers makes the WRMSR to
1437                          * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
1438                          * mysteriously hang on Core2.
1439                          *
1440                          * As a workaround, we don't do this.
1441                          */
1442                         x86_pmu.pebs_buffer_size = PAGE_SIZE;
1443                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
1444                         break;
1445
1446                 case 1:
1447                         pr_cont("PEBS fmt1%c, ", pebs_type);
1448                         x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
1449                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1450                         break;
1451
1452                 case 2:
1453                         pr_cont("PEBS fmt2%c, ", pebs_type);
1454                         x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
1455                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1456                         break;
1457
1458                 case 3:
1459                         pr_cont("PEBS fmt3%c, ", pebs_type);
1460                         x86_pmu.pebs_record_size =
1461                                                 sizeof(struct pebs_record_skl);
1462                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1463                         x86_pmu.free_running_flags |= PERF_SAMPLE_TIME;
1464                         break;
1465
1466                 default:
1467                         pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
1468                         x86_pmu.pebs = 0;
1469                 }
1470         }
1471 }
1472
1473 void perf_restore_debug_store(void)
1474 {
1475         struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
1476
1477         if (!x86_pmu.bts && !x86_pmu.pebs)
1478                 return;
1479
1480         wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
1481 }