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