]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - arch/x86/events/intel/ds.c
Merge tag 'for-linus-20170812' of git://git.infradead.org/linux-mtd
[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 /*
610  * PEBS
611  */
612 struct event_constraint intel_core2_pebs_event_constraints[] = {
613         INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
614         INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
615         INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
616         INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
617         INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
618         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
619         INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
620         EVENT_CONSTRAINT_END
621 };
622
623 struct event_constraint intel_atom_pebs_event_constraints[] = {
624         INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
625         INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
626         INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
627         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
628         INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
629         /* Allow all events as PEBS with no flags */
630         INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
631         EVENT_CONSTRAINT_END
632 };
633
634 struct event_constraint intel_slm_pebs_event_constraints[] = {
635         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
636         INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x1),
637         /* Allow all events as PEBS with no flags */
638         INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
639         EVENT_CONSTRAINT_END
640 };
641
642 struct event_constraint intel_glm_pebs_event_constraints[] = {
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_glp_pebs_event_constraints[] = {
649         /* Allow all events as PEBS with no flags */
650         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
651         EVENT_CONSTRAINT_END
652 };
653
654 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
655         INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
656         INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
657         INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
658         INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
659         INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
660         INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
661         INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
662         INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
663         INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
664         INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
665         INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
666         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
667         INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
668         EVENT_CONSTRAINT_END
669 };
670
671 struct event_constraint intel_westmere_pebs_event_constraints[] = {
672         INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
673         INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
674         INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
675         INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
676         INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
677         INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
678         INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
679         INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
680         INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
681         INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
682         INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
683         /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
684         INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
685         EVENT_CONSTRAINT_END
686 };
687
688 struct event_constraint intel_snb_pebs_event_constraints[] = {
689         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
690         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
691         INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
692         /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
693         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
694         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
695         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
696         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
697         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
698         /* Allow all events as PEBS with no flags */
699         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
700         EVENT_CONSTRAINT_END
701 };
702
703 struct event_constraint intel_ivb_pebs_event_constraints[] = {
704         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
705         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
706         INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
707         /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
708         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
709         /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
710         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
711         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
712         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
713         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
714         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
715         /* Allow all events as PEBS with no flags */
716         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
717         EVENT_CONSTRAINT_END
718 };
719
720 struct event_constraint intel_hsw_pebs_event_constraints[] = {
721         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
722         INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
723         /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
724         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
725         /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
726         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
727         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
728         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
729         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
730         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
731         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
732         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
733         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
734         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
735         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
736         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
737         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
738         /* Allow all events as PEBS with no flags */
739         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
740         EVENT_CONSTRAINT_END
741 };
742
743 struct event_constraint intel_bdw_pebs_event_constraints[] = {
744         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
745         INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
746         /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
747         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
748         /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
749         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
750         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
751         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
752         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
753         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
754         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
755         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
756         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
757         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
758         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
759         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
760         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
761         /* Allow all events as PEBS with no flags */
762         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
763         EVENT_CONSTRAINT_END
764 };
765
766
767 struct event_constraint intel_skl_pebs_event_constraints[] = {
768         INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),      /* INST_RETIRED.PREC_DIST */
769         /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
770         INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
771         /* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
772         INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
773         INTEL_PLD_CONSTRAINT(0x1cd, 0xf),                     /* MEM_TRANS_RETIRED.* */
774         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
775         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
776         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
777         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
778         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
779         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
780         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
781         INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
782         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
783         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
784         INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
785         /* Allow all events as PEBS with no flags */
786         INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
787         EVENT_CONSTRAINT_END
788 };
789
790 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
791 {
792         struct event_constraint *c;
793
794         if (!event->attr.precise_ip)
795                 return NULL;
796
797         if (x86_pmu.pebs_constraints) {
798                 for_each_event_constraint(c, x86_pmu.pebs_constraints) {
799                         if ((event->hw.config & c->cmask) == c->code) {
800                                 event->hw.flags |= c->flags;
801                                 return c;
802                         }
803                 }
804         }
805
806         return &emptyconstraint;
807 }
808
809 /*
810  * We need the sched_task callback even for per-cpu events when we use
811  * the large interrupt threshold, such that we can provide PID and TID
812  * to PEBS samples.
813  */
814 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
815 {
816         return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
817 }
818
819 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in)
820 {
821         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
822
823         if (!sched_in && pebs_needs_sched_cb(cpuc))
824                 intel_pmu_drain_pebs_buffer();
825 }
826
827 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
828 {
829         struct debug_store *ds = cpuc->ds;
830         u64 threshold;
831
832         if (cpuc->n_pebs == cpuc->n_large_pebs) {
833                 threshold = ds->pebs_absolute_maximum -
834                         x86_pmu.max_pebs_events * x86_pmu.pebs_record_size;
835         } else {
836                 threshold = ds->pebs_buffer_base + x86_pmu.pebs_record_size;
837         }
838
839         ds->pebs_interrupt_threshold = threshold;
840 }
841
842 static void
843 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc, struct pmu *pmu)
844 {
845         /*
846          * Make sure we get updated with the first PEBS
847          * event. It will trigger also during removal, but
848          * that does not hurt:
849          */
850         bool update = cpuc->n_pebs == 1;
851
852         if (needed_cb != pebs_needs_sched_cb(cpuc)) {
853                 if (!needed_cb)
854                         perf_sched_cb_inc(pmu);
855                 else
856                         perf_sched_cb_dec(pmu);
857
858                 update = true;
859         }
860
861         if (update)
862                 pebs_update_threshold(cpuc);
863 }
864
865 void intel_pmu_pebs_add(struct perf_event *event)
866 {
867         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
868         struct hw_perf_event *hwc = &event->hw;
869         bool needed_cb = pebs_needs_sched_cb(cpuc);
870
871         cpuc->n_pebs++;
872         if (hwc->flags & PERF_X86_EVENT_FREERUNNING)
873                 cpuc->n_large_pebs++;
874
875         pebs_update_state(needed_cb, cpuc, event->ctx->pmu);
876 }
877
878 void intel_pmu_pebs_enable(struct perf_event *event)
879 {
880         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
881         struct hw_perf_event *hwc = &event->hw;
882         struct debug_store *ds = cpuc->ds;
883
884         hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
885
886         cpuc->pebs_enabled |= 1ULL << hwc->idx;
887
888         if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
889                 cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
890         else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
891                 cpuc->pebs_enabled |= 1ULL << 63;
892
893         /*
894          * Use auto-reload if possible to save a MSR write in the PMI.
895          * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
896          */
897         if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
898                 ds->pebs_event_reset[hwc->idx] =
899                         (u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
900         } else {
901                 ds->pebs_event_reset[hwc->idx] = 0;
902         }
903 }
904
905 void intel_pmu_pebs_del(struct perf_event *event)
906 {
907         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
908         struct hw_perf_event *hwc = &event->hw;
909         bool needed_cb = pebs_needs_sched_cb(cpuc);
910
911         cpuc->n_pebs--;
912         if (hwc->flags & PERF_X86_EVENT_FREERUNNING)
913                 cpuc->n_large_pebs--;
914
915         pebs_update_state(needed_cb, cpuc, event->ctx->pmu);
916 }
917
918 void intel_pmu_pebs_disable(struct perf_event *event)
919 {
920         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
921         struct hw_perf_event *hwc = &event->hw;
922
923         if (cpuc->n_pebs == cpuc->n_large_pebs)
924                 intel_pmu_drain_pebs_buffer();
925
926         cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
927
928         if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
929                 cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
930         else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
931                 cpuc->pebs_enabled &= ~(1ULL << 63);
932
933         if (cpuc->enabled)
934                 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
935
936         hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
937 }
938
939 void intel_pmu_pebs_enable_all(void)
940 {
941         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
942
943         if (cpuc->pebs_enabled)
944                 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
945 }
946
947 void intel_pmu_pebs_disable_all(void)
948 {
949         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
950
951         if (cpuc->pebs_enabled)
952                 wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
953 }
954
955 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
956 {
957         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
958         unsigned long from = cpuc->lbr_entries[0].from;
959         unsigned long old_to, to = cpuc->lbr_entries[0].to;
960         unsigned long ip = regs->ip;
961         int is_64bit = 0;
962         void *kaddr;
963         int size;
964
965         /*
966          * We don't need to fixup if the PEBS assist is fault like
967          */
968         if (!x86_pmu.intel_cap.pebs_trap)
969                 return 1;
970
971         /*
972          * No LBR entry, no basic block, no rewinding
973          */
974         if (!cpuc->lbr_stack.nr || !from || !to)
975                 return 0;
976
977         /*
978          * Basic blocks should never cross user/kernel boundaries
979          */
980         if (kernel_ip(ip) != kernel_ip(to))
981                 return 0;
982
983         /*
984          * unsigned math, either ip is before the start (impossible) or
985          * the basic block is larger than 1 page (sanity)
986          */
987         if ((ip - to) > PEBS_FIXUP_SIZE)
988                 return 0;
989
990         /*
991          * We sampled a branch insn, rewind using the LBR stack
992          */
993         if (ip == to) {
994                 set_linear_ip(regs, from);
995                 return 1;
996         }
997
998         size = ip - to;
999         if (!kernel_ip(ip)) {
1000                 int bytes;
1001                 u8 *buf = this_cpu_read(insn_buffer);
1002
1003                 /* 'size' must fit our buffer, see above */
1004                 bytes = copy_from_user_nmi(buf, (void __user *)to, size);
1005                 if (bytes != 0)
1006                         return 0;
1007
1008                 kaddr = buf;
1009         } else {
1010                 kaddr = (void *)to;
1011         }
1012
1013         do {
1014                 struct insn insn;
1015
1016                 old_to = to;
1017
1018 #ifdef CONFIG_X86_64
1019                 is_64bit = kernel_ip(to) || !test_thread_flag(TIF_IA32);
1020 #endif
1021                 insn_init(&insn, kaddr, size, is_64bit);
1022                 insn_get_length(&insn);
1023                 /*
1024                  * Make sure there was not a problem decoding the
1025                  * instruction and getting the length.  This is
1026                  * doubly important because we have an infinite
1027                  * loop if insn.length=0.
1028                  */
1029                 if (!insn.length)
1030                         break;
1031
1032                 to += insn.length;
1033                 kaddr += insn.length;
1034                 size -= insn.length;
1035         } while (to < ip);
1036
1037         if (to == ip) {
1038                 set_linear_ip(regs, old_to);
1039                 return 1;
1040         }
1041
1042         /*
1043          * Even though we decoded the basic block, the instruction stream
1044          * never matched the given IP, either the TO or the IP got corrupted.
1045          */
1046         return 0;
1047 }
1048
1049 static inline u64 intel_hsw_weight(struct pebs_record_skl *pebs)
1050 {
1051         if (pebs->tsx_tuning) {
1052                 union hsw_tsx_tuning tsx = { .value = pebs->tsx_tuning };
1053                 return tsx.cycles_last_block;
1054         }
1055         return 0;
1056 }
1057
1058 static inline u64 intel_hsw_transaction(struct pebs_record_skl *pebs)
1059 {
1060         u64 txn = (pebs->tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1061
1062         /* For RTM XABORTs also log the abort code from AX */
1063         if ((txn & PERF_TXN_TRANSACTION) && (pebs->ax & 1))
1064                 txn |= ((pebs->ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1065         return txn;
1066 }
1067
1068 static void setup_pebs_sample_data(struct perf_event *event,
1069                                    struct pt_regs *iregs, void *__pebs,
1070                                    struct perf_sample_data *data,
1071                                    struct pt_regs *regs)
1072 {
1073 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1074                 (PERF_X86_EVENT_PEBS_ST_HSW | \
1075                  PERF_X86_EVENT_PEBS_LD_HSW | \
1076                  PERF_X86_EVENT_PEBS_NA_HSW)
1077         /*
1078          * We cast to the biggest pebs_record but are careful not to
1079          * unconditionally access the 'extra' entries.
1080          */
1081         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1082         struct pebs_record_skl *pebs = __pebs;
1083         u64 sample_type;
1084         int fll, fst, dsrc;
1085         int fl = event->hw.flags;
1086
1087         if (pebs == NULL)
1088                 return;
1089
1090         sample_type = event->attr.sample_type;
1091         dsrc = sample_type & PERF_SAMPLE_DATA_SRC;
1092
1093         fll = fl & PERF_X86_EVENT_PEBS_LDLAT;
1094         fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1095
1096         perf_sample_data_init(data, 0, event->hw.last_period);
1097
1098         data->period = event->hw.last_period;
1099
1100         /*
1101          * Use latency for weight (only avail with PEBS-LL)
1102          */
1103         if (fll && (sample_type & PERF_SAMPLE_WEIGHT))
1104                 data->weight = pebs->lat;
1105
1106         /*
1107          * data.data_src encodes the data source
1108          */
1109         if (dsrc) {
1110                 u64 val = PERF_MEM_NA;
1111                 if (fll)
1112                         val = load_latency_data(pebs->dse);
1113                 else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1114                         val = precise_datala_hsw(event, pebs->dse);
1115                 else if (fst)
1116                         val = precise_store_data(pebs->dse);
1117                 data->data_src.val = val;
1118         }
1119
1120         /*
1121          * We use the interrupt regs as a base because the PEBS record does not
1122          * contain a full regs set, specifically it seems to lack segment
1123          * descriptors, which get used by things like user_mode().
1124          *
1125          * In the simple case fix up only the IP for PERF_SAMPLE_IP.
1126          *
1127          * We must however always use BP,SP from iregs for the unwinder to stay
1128          * sane; the record BP,SP can point into thin air when the record is
1129          * from a previous PMI context or an (I)RET happend between the record
1130          * and PMI.
1131          */
1132         *regs = *iregs;
1133         regs->flags = pebs->flags;
1134         set_linear_ip(regs, pebs->ip);
1135
1136         if (sample_type & PERF_SAMPLE_REGS_INTR) {
1137                 regs->ax = pebs->ax;
1138                 regs->bx = pebs->bx;
1139                 regs->cx = pebs->cx;
1140                 regs->dx = pebs->dx;
1141                 regs->si = pebs->si;
1142                 regs->di = pebs->di;
1143
1144                 /*
1145                  * Per the above; only set BP,SP if we don't need callchains.
1146                  *
1147                  * XXX: does this make sense?
1148                  */
1149                 if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
1150                         regs->bp = pebs->bp;
1151                         regs->sp = pebs->sp;
1152                 }
1153
1154                 /*
1155                  * Preserve PERF_EFLAGS_VM from set_linear_ip().
1156                  */
1157                 regs->flags = pebs->flags | (regs->flags & PERF_EFLAGS_VM);
1158 #ifndef CONFIG_X86_32
1159                 regs->r8 = pebs->r8;
1160                 regs->r9 = pebs->r9;
1161                 regs->r10 = pebs->r10;
1162                 regs->r11 = pebs->r11;
1163                 regs->r12 = pebs->r12;
1164                 regs->r13 = pebs->r13;
1165                 regs->r14 = pebs->r14;
1166                 regs->r15 = pebs->r15;
1167 #endif
1168         }
1169
1170         if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format >= 2) {
1171                 regs->ip = pebs->real_ip;
1172                 regs->flags |= PERF_EFLAGS_EXACT;
1173         } else if (event->attr.precise_ip > 1 && intel_pmu_pebs_fixup_ip(regs))
1174                 regs->flags |= PERF_EFLAGS_EXACT;
1175         else
1176                 regs->flags &= ~PERF_EFLAGS_EXACT;
1177
1178         if ((sample_type & PERF_SAMPLE_ADDR) &&
1179             x86_pmu.intel_cap.pebs_format >= 1)
1180                 data->addr = pebs->dla;
1181
1182         if (x86_pmu.intel_cap.pebs_format >= 2) {
1183                 /* Only set the TSX weight when no memory weight. */
1184                 if ((sample_type & PERF_SAMPLE_WEIGHT) && !fll)
1185                         data->weight = intel_hsw_weight(pebs);
1186
1187                 if (sample_type & PERF_SAMPLE_TRANSACTION)
1188                         data->txn = intel_hsw_transaction(pebs);
1189         }
1190
1191         /*
1192          * v3 supplies an accurate time stamp, so we use that
1193          * for the time stamp.
1194          *
1195          * We can only do this for the default trace clock.
1196          */
1197         if (x86_pmu.intel_cap.pebs_format >= 3 &&
1198                 event->attr.use_clockid == 0)
1199                 data->time = native_sched_clock_from_tsc(pebs->tsc);
1200
1201         if (has_branch_stack(event))
1202                 data->br_stack = &cpuc->lbr_stack;
1203 }
1204
1205 static inline void *
1206 get_next_pebs_record_by_bit(void *base, void *top, int bit)
1207 {
1208         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1209         void *at;
1210         u64 pebs_status;
1211
1212         /*
1213          * fmt0 does not have a status bitfield (does not use
1214          * perf_record_nhm format)
1215          */
1216         if (x86_pmu.intel_cap.pebs_format < 1)
1217                 return base;
1218
1219         if (base == NULL)
1220                 return NULL;
1221
1222         for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1223                 struct pebs_record_nhm *p = at;
1224
1225                 if (test_bit(bit, (unsigned long *)&p->status)) {
1226                         /* PEBS v3 has accurate status bits */
1227                         if (x86_pmu.intel_cap.pebs_format >= 3)
1228                                 return at;
1229
1230                         if (p->status == (1 << bit))
1231                                 return at;
1232
1233                         /* clear non-PEBS bit and re-check */
1234                         pebs_status = p->status & cpuc->pebs_enabled;
1235                         pebs_status &= PEBS_COUNTER_MASK;
1236                         if (pebs_status == (1 << bit))
1237                                 return at;
1238                 }
1239         }
1240         return NULL;
1241 }
1242
1243 static void __intel_pmu_pebs_event(struct perf_event *event,
1244                                    struct pt_regs *iregs,
1245                                    void *base, void *top,
1246                                    int bit, int count)
1247 {
1248         struct perf_sample_data data;
1249         struct pt_regs regs;
1250         void *at = get_next_pebs_record_by_bit(base, top, bit);
1251
1252         if (!intel_pmu_save_and_restart(event) &&
1253             !(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
1254                 return;
1255
1256         while (count > 1) {
1257                 setup_pebs_sample_data(event, iregs, at, &data, &regs);
1258                 perf_event_output(event, &data, &regs);
1259                 at += x86_pmu.pebs_record_size;
1260                 at = get_next_pebs_record_by_bit(at, top, bit);
1261                 count--;
1262         }
1263
1264         setup_pebs_sample_data(event, iregs, at, &data, &regs);
1265
1266         /*
1267          * All but the last records are processed.
1268          * The last one is left to be able to call the overflow handler.
1269          */
1270         if (perf_event_overflow(event, &data, &regs)) {
1271                 x86_pmu_stop(event, 0);
1272                 return;
1273         }
1274
1275 }
1276
1277 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
1278 {
1279         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1280         struct debug_store *ds = cpuc->ds;
1281         struct perf_event *event = cpuc->events[0]; /* PMC0 only */
1282         struct pebs_record_core *at, *top;
1283         int n;
1284
1285         if (!x86_pmu.pebs_active)
1286                 return;
1287
1288         at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
1289         top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
1290
1291         /*
1292          * Whatever else happens, drain the thing
1293          */
1294         ds->pebs_index = ds->pebs_buffer_base;
1295
1296         if (!test_bit(0, cpuc->active_mask))
1297                 return;
1298
1299         WARN_ON_ONCE(!event);
1300
1301         if (!event->attr.precise_ip)
1302                 return;
1303
1304         n = top - at;
1305         if (n <= 0)
1306                 return;
1307
1308         __intel_pmu_pebs_event(event, iregs, at, top, 0, n);
1309 }
1310
1311 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
1312 {
1313         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1314         struct debug_store *ds = cpuc->ds;
1315         struct perf_event *event;
1316         void *base, *at, *top;
1317         short counts[MAX_PEBS_EVENTS] = {};
1318         short error[MAX_PEBS_EVENTS] = {};
1319         int bit, i;
1320
1321         if (!x86_pmu.pebs_active)
1322                 return;
1323
1324         base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
1325         top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
1326
1327         ds->pebs_index = ds->pebs_buffer_base;
1328
1329         if (unlikely(base >= top))
1330                 return;
1331
1332         for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1333                 struct pebs_record_nhm *p = at;
1334                 u64 pebs_status;
1335
1336                 pebs_status = p->status & cpuc->pebs_enabled;
1337                 pebs_status &= (1ULL << x86_pmu.max_pebs_events) - 1;
1338
1339                 /* PEBS v3 has more accurate status bits */
1340                 if (x86_pmu.intel_cap.pebs_format >= 3) {
1341                         for_each_set_bit(bit, (unsigned long *)&pebs_status,
1342                                          x86_pmu.max_pebs_events)
1343                                 counts[bit]++;
1344
1345                         continue;
1346                 }
1347
1348                 /*
1349                  * On some CPUs the PEBS status can be zero when PEBS is
1350                  * racing with clearing of GLOBAL_STATUS.
1351                  *
1352                  * Normally we would drop that record, but in the
1353                  * case when there is only a single active PEBS event
1354                  * we can assume it's for that event.
1355                  */
1356                 if (!pebs_status && cpuc->pebs_enabled &&
1357                         !(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
1358                         pebs_status = cpuc->pebs_enabled;
1359
1360                 bit = find_first_bit((unsigned long *)&pebs_status,
1361                                         x86_pmu.max_pebs_events);
1362                 if (bit >= x86_pmu.max_pebs_events)
1363                         continue;
1364
1365                 /*
1366                  * The PEBS hardware does not deal well with the situation
1367                  * when events happen near to each other and multiple bits
1368                  * are set. But it should happen rarely.
1369                  *
1370                  * If these events include one PEBS and multiple non-PEBS
1371                  * events, it doesn't impact PEBS record. The record will
1372                  * be handled normally. (slow path)
1373                  *
1374                  * If these events include two or more PEBS events, the
1375                  * records for the events can be collapsed into a single
1376                  * one, and it's not possible to reconstruct all events
1377                  * that caused the PEBS record. It's called collision.
1378                  * If collision happened, the record will be dropped.
1379                  */
1380                 if (p->status != (1ULL << bit)) {
1381                         for_each_set_bit(i, (unsigned long *)&pebs_status,
1382                                          x86_pmu.max_pebs_events)
1383                                 error[i]++;
1384                         continue;
1385                 }
1386
1387                 counts[bit]++;
1388         }
1389
1390         for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) {
1391                 if ((counts[bit] == 0) && (error[bit] == 0))
1392                         continue;
1393
1394                 event = cpuc->events[bit];
1395                 if (WARN_ON_ONCE(!event))
1396                         continue;
1397
1398                 if (WARN_ON_ONCE(!event->attr.precise_ip))
1399                         continue;
1400
1401                 /* log dropped samples number */
1402                 if (error[bit]) {
1403                         perf_log_lost_samples(event, error[bit]);
1404
1405                         if (perf_event_account_interrupt(event))
1406                                 x86_pmu_stop(event, 0);
1407                 }
1408
1409                 if (counts[bit]) {
1410                         __intel_pmu_pebs_event(event, iregs, base,
1411                                                top, bit, counts[bit]);
1412                 }
1413         }
1414 }
1415
1416 /*
1417  * BTS, PEBS probe and setup
1418  */
1419
1420 void __init intel_ds_init(void)
1421 {
1422         /*
1423          * No support for 32bit formats
1424          */
1425         if (!boot_cpu_has(X86_FEATURE_DTES64))
1426                 return;
1427
1428         x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
1429         x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
1430         x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
1431         if (x86_pmu.pebs) {
1432                 char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
1433                 int format = x86_pmu.intel_cap.pebs_format;
1434
1435                 switch (format) {
1436                 case 0:
1437                         pr_cont("PEBS fmt0%c, ", pebs_type);
1438                         x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
1439                         /*
1440                          * Using >PAGE_SIZE buffers makes the WRMSR to
1441                          * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
1442                          * mysteriously hang on Core2.
1443                          *
1444                          * As a workaround, we don't do this.
1445                          */
1446                         x86_pmu.pebs_buffer_size = PAGE_SIZE;
1447                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
1448                         break;
1449
1450                 case 1:
1451                         pr_cont("PEBS fmt1%c, ", pebs_type);
1452                         x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
1453                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1454                         break;
1455
1456                 case 2:
1457                         pr_cont("PEBS fmt2%c, ", pebs_type);
1458                         x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
1459                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1460                         break;
1461
1462                 case 3:
1463                         pr_cont("PEBS fmt3%c, ", pebs_type);
1464                         x86_pmu.pebs_record_size =
1465                                                 sizeof(struct pebs_record_skl);
1466                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1467                         x86_pmu.free_running_flags |= PERF_SAMPLE_TIME;
1468                         break;
1469
1470                 default:
1471                         pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
1472                         x86_pmu.pebs = 0;
1473                 }
1474         }
1475 }
1476
1477 void perf_restore_debug_store(void)
1478 {
1479         struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
1480
1481         if (!x86_pmu.bts && !x86_pmu.pebs)
1482                 return;
1483
1484         wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
1485 }