]> git.karo-electronics.de Git - karo-tx-linux.git/blob - arch/x86/kernel/cpu/perf_event_intel_ds.c
Merge branch 'x86-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[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         struct pebs_record_nhm nhm;
186         /*
187          * Real IP of the event. In the Intel documentation this
188          * is called eventingrip.
189          */
190         u64 real_ip;
191         /*
192          * TSX tuning information field: abort cycles and abort flags.
193          */
194         u64 tsx_tuning;
195 };
196
197 void init_debug_store_on_cpu(int cpu)
198 {
199         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
200
201         if (!ds)
202                 return;
203
204         wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
205                      (u32)((u64)(unsigned long)ds),
206                      (u32)((u64)(unsigned long)ds >> 32));
207 }
208
209 void fini_debug_store_on_cpu(int cpu)
210 {
211         if (!per_cpu(cpu_hw_events, cpu).ds)
212                 return;
213
214         wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
215 }
216
217 static int alloc_pebs_buffer(int cpu)
218 {
219         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
220         int node = cpu_to_node(cpu);
221         int max, thresh = 1; /* always use a single PEBS record */
222         void *buffer;
223
224         if (!x86_pmu.pebs)
225                 return 0;
226
227         buffer = kzalloc_node(PEBS_BUFFER_SIZE, GFP_KERNEL, node);
228         if (unlikely(!buffer))
229                 return -ENOMEM;
230
231         max = PEBS_BUFFER_SIZE / x86_pmu.pebs_record_size;
232
233         ds->pebs_buffer_base = (u64)(unsigned long)buffer;
234         ds->pebs_index = ds->pebs_buffer_base;
235         ds->pebs_absolute_maximum = ds->pebs_buffer_base +
236                 max * x86_pmu.pebs_record_size;
237
238         ds->pebs_interrupt_threshold = ds->pebs_buffer_base +
239                 thresh * x86_pmu.pebs_record_size;
240
241         return 0;
242 }
243
244 static void release_pebs_buffer(int cpu)
245 {
246         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
247
248         if (!ds || !x86_pmu.pebs)
249                 return;
250
251         kfree((void *)(unsigned long)ds->pebs_buffer_base);
252         ds->pebs_buffer_base = 0;
253 }
254
255 static int alloc_bts_buffer(int cpu)
256 {
257         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
258         int node = cpu_to_node(cpu);
259         int max, thresh;
260         void *buffer;
261
262         if (!x86_pmu.bts)
263                 return 0;
264
265         buffer = kzalloc_node(BTS_BUFFER_SIZE, GFP_KERNEL, node);
266         if (unlikely(!buffer))
267                 return -ENOMEM;
268
269         max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
270         thresh = max / 16;
271
272         ds->bts_buffer_base = (u64)(unsigned long)buffer;
273         ds->bts_index = ds->bts_buffer_base;
274         ds->bts_absolute_maximum = ds->bts_buffer_base +
275                 max * BTS_RECORD_SIZE;
276         ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
277                 thresh * BTS_RECORD_SIZE;
278
279         return 0;
280 }
281
282 static void release_bts_buffer(int cpu)
283 {
284         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
285
286         if (!ds || !x86_pmu.bts)
287                 return;
288
289         kfree((void *)(unsigned long)ds->bts_buffer_base);
290         ds->bts_buffer_base = 0;
291 }
292
293 static int alloc_ds_buffer(int cpu)
294 {
295         int node = cpu_to_node(cpu);
296         struct debug_store *ds;
297
298         ds = kzalloc_node(sizeof(*ds), GFP_KERNEL, node);
299         if (unlikely(!ds))
300                 return -ENOMEM;
301
302         per_cpu(cpu_hw_events, cpu).ds = ds;
303
304         return 0;
305 }
306
307 static void release_ds_buffer(int cpu)
308 {
309         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
310
311         if (!ds)
312                 return;
313
314         per_cpu(cpu_hw_events, cpu).ds = NULL;
315         kfree(ds);
316 }
317
318 void release_ds_buffers(void)
319 {
320         int cpu;
321
322         if (!x86_pmu.bts && !x86_pmu.pebs)
323                 return;
324
325         get_online_cpus();
326         for_each_online_cpu(cpu)
327                 fini_debug_store_on_cpu(cpu);
328
329         for_each_possible_cpu(cpu) {
330                 release_pebs_buffer(cpu);
331                 release_bts_buffer(cpu);
332                 release_ds_buffer(cpu);
333         }
334         put_online_cpus();
335 }
336
337 void reserve_ds_buffers(void)
338 {
339         int bts_err = 0, pebs_err = 0;
340         int cpu;
341
342         x86_pmu.bts_active = 0;
343         x86_pmu.pebs_active = 0;
344
345         if (!x86_pmu.bts && !x86_pmu.pebs)
346                 return;
347
348         if (!x86_pmu.bts)
349                 bts_err = 1;
350
351         if (!x86_pmu.pebs)
352                 pebs_err = 1;
353
354         get_online_cpus();
355
356         for_each_possible_cpu(cpu) {
357                 if (alloc_ds_buffer(cpu)) {
358                         bts_err = 1;
359                         pebs_err = 1;
360                 }
361
362                 if (!bts_err && alloc_bts_buffer(cpu))
363                         bts_err = 1;
364
365                 if (!pebs_err && alloc_pebs_buffer(cpu))
366                         pebs_err = 1;
367
368                 if (bts_err && pebs_err)
369                         break;
370         }
371
372         if (bts_err) {
373                 for_each_possible_cpu(cpu)
374                         release_bts_buffer(cpu);
375         }
376
377         if (pebs_err) {
378                 for_each_possible_cpu(cpu)
379                         release_pebs_buffer(cpu);
380         }
381
382         if (bts_err && pebs_err) {
383                 for_each_possible_cpu(cpu)
384                         release_ds_buffer(cpu);
385         } else {
386                 if (x86_pmu.bts && !bts_err)
387                         x86_pmu.bts_active = 1;
388
389                 if (x86_pmu.pebs && !pebs_err)
390                         x86_pmu.pebs_active = 1;
391
392                 for_each_online_cpu(cpu)
393                         init_debug_store_on_cpu(cpu);
394         }
395
396         put_online_cpus();
397 }
398
399 /*
400  * BTS
401  */
402
403 struct event_constraint bts_constraint =
404         EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
405
406 void intel_pmu_enable_bts(u64 config)
407 {
408         unsigned long debugctlmsr;
409
410         debugctlmsr = get_debugctlmsr();
411
412         debugctlmsr |= DEBUGCTLMSR_TR;
413         debugctlmsr |= DEBUGCTLMSR_BTS;
414         debugctlmsr |= DEBUGCTLMSR_BTINT;
415
416         if (!(config & ARCH_PERFMON_EVENTSEL_OS))
417                 debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
418
419         if (!(config & ARCH_PERFMON_EVENTSEL_USR))
420                 debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
421
422         update_debugctlmsr(debugctlmsr);
423 }
424
425 void intel_pmu_disable_bts(void)
426 {
427         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
428         unsigned long debugctlmsr;
429
430         if (!cpuc->ds)
431                 return;
432
433         debugctlmsr = get_debugctlmsr();
434
435         debugctlmsr &=
436                 ~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
437                   DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
438
439         update_debugctlmsr(debugctlmsr);
440 }
441
442 int intel_pmu_drain_bts_buffer(void)
443 {
444         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
445         struct debug_store *ds = cpuc->ds;
446         struct bts_record {
447                 u64     from;
448                 u64     to;
449                 u64     flags;
450         };
451         struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
452         struct bts_record *at, *top;
453         struct perf_output_handle handle;
454         struct perf_event_header header;
455         struct perf_sample_data data;
456         struct pt_regs regs;
457
458         if (!event)
459                 return 0;
460
461         if (!x86_pmu.bts_active)
462                 return 0;
463
464         at  = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
465         top = (struct bts_record *)(unsigned long)ds->bts_index;
466
467         if (top <= at)
468                 return 0;
469
470         memset(&regs, 0, sizeof(regs));
471
472         ds->bts_index = ds->bts_buffer_base;
473
474         perf_sample_data_init(&data, 0, event->hw.last_period);
475
476         /*
477          * Prepare a generic sample, i.e. fill in the invariant fields.
478          * We will overwrite the from and to address before we output
479          * the sample.
480          */
481         perf_prepare_sample(&header, &data, event, &regs);
482
483         if (perf_output_begin(&handle, event, header.size * (top - at)))
484                 return 1;
485
486         for (; at < top; at++) {
487                 data.ip         = at->from;
488                 data.addr       = at->to;
489
490                 perf_output_sample(&handle, &header, &data, event);
491         }
492
493         perf_output_end(&handle);
494
495         /* There's new data available. */
496         event->hw.interrupts++;
497         event->pending_kill = POLL_IN;
498         return 1;
499 }
500
501 /*
502  * PEBS
503  */
504 struct event_constraint intel_core2_pebs_event_constraints[] = {
505         INTEL_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
506         INTEL_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
507         INTEL_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
508         INTEL_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
509         INTEL_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
510         EVENT_CONSTRAINT_END
511 };
512
513 struct event_constraint intel_atom_pebs_event_constraints[] = {
514         INTEL_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
515         INTEL_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
516         INTEL_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
517         EVENT_CONSTRAINT_END
518 };
519
520 struct event_constraint intel_slm_pebs_event_constraints[] = {
521         INTEL_UEVENT_CONSTRAINT(0x0103, 0x1), /* REHABQ.LD_BLOCK_ST_FORWARD_PS */
522         INTEL_UEVENT_CONSTRAINT(0x0803, 0x1), /* REHABQ.LD_SPLITS_PS */
523         INTEL_UEVENT_CONSTRAINT(0x0204, 0x1), /* MEM_UOPS_RETIRED.L2_HIT_LOADS_PS */
524         INTEL_UEVENT_CONSTRAINT(0x0404, 0x1), /* MEM_UOPS_RETIRED.L2_MISS_LOADS_PS */
525         INTEL_UEVENT_CONSTRAINT(0x0804, 0x1), /* MEM_UOPS_RETIRED.DTLB_MISS_LOADS_PS */
526         INTEL_UEVENT_CONSTRAINT(0x2004, 0x1), /* MEM_UOPS_RETIRED.HITM_PS */
527         INTEL_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY_PS */
528         INTEL_UEVENT_CONSTRAINT(0x00c4, 0x1), /* BR_INST_RETIRED.ALL_BRANCHES_PS */
529         INTEL_UEVENT_CONSTRAINT(0x7ec4, 0x1), /* BR_INST_RETIRED.JCC_PS */
530         INTEL_UEVENT_CONSTRAINT(0xbfc4, 0x1), /* BR_INST_RETIRED.FAR_BRANCH_PS */
531         INTEL_UEVENT_CONSTRAINT(0xebc4, 0x1), /* BR_INST_RETIRED.NON_RETURN_IND_PS */
532         INTEL_UEVENT_CONSTRAINT(0xf7c4, 0x1), /* BR_INST_RETIRED.RETURN_PS */
533         INTEL_UEVENT_CONSTRAINT(0xf9c4, 0x1), /* BR_INST_RETIRED.CALL_PS */
534         INTEL_UEVENT_CONSTRAINT(0xfbc4, 0x1), /* BR_INST_RETIRED.IND_CALL_PS */
535         INTEL_UEVENT_CONSTRAINT(0xfdc4, 0x1), /* BR_INST_RETIRED.REL_CALL_PS */
536         INTEL_UEVENT_CONSTRAINT(0xfec4, 0x1), /* BR_INST_RETIRED.TAKEN_JCC_PS */
537         INTEL_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_MISP_RETIRED.ALL_BRANCHES_PS */
538         INTEL_UEVENT_CONSTRAINT(0x7ec5, 0x1), /* BR_INST_MISP_RETIRED.JCC_PS */
539         INTEL_UEVENT_CONSTRAINT(0xebc5, 0x1), /* BR_INST_MISP_RETIRED.NON_RETURN_IND_PS */
540         INTEL_UEVENT_CONSTRAINT(0xf7c5, 0x1), /* BR_INST_MISP_RETIRED.RETURN_PS */
541         INTEL_UEVENT_CONSTRAINT(0xfbc5, 0x1), /* BR_INST_MISP_RETIRED.IND_CALL_PS */
542         INTEL_UEVENT_CONSTRAINT(0xfec5, 0x1), /* BR_INST_MISP_RETIRED.TAKEN_JCC_PS */
543         EVENT_CONSTRAINT_END
544 };
545
546 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
547         INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
548         INTEL_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
549         INTEL_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
550         INTEL_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
551         INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
552         INTEL_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
553         INTEL_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
554         INTEL_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
555         INTEL_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
556         INTEL_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
557         INTEL_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
558         EVENT_CONSTRAINT_END
559 };
560
561 struct event_constraint intel_westmere_pebs_event_constraints[] = {
562         INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
563         INTEL_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
564         INTEL_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
565         INTEL_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
566         INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
567         INTEL_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
568         INTEL_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
569         INTEL_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
570         INTEL_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
571         INTEL_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
572         INTEL_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
573         EVENT_CONSTRAINT_END
574 };
575
576 struct event_constraint intel_snb_pebs_event_constraints[] = {
577         INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
578         INTEL_UEVENT_CONSTRAINT(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
579         INTEL_UEVENT_CONSTRAINT(0x02c2, 0xf), /* UOPS_RETIRED.RETIRE_SLOTS */
580         INTEL_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
581         INTEL_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
582         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
583         INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
584         INTEL_EVENT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
585         INTEL_EVENT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
586         INTEL_EVENT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
587         INTEL_EVENT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
588         INTEL_UEVENT_CONSTRAINT(0x02d4, 0xf), /* MEM_LOAD_UOPS_MISC_RETIRED.LLC_MISS */
589         EVENT_CONSTRAINT_END
590 };
591
592 struct event_constraint intel_ivb_pebs_event_constraints[] = {
593         INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
594         INTEL_UEVENT_CONSTRAINT(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
595         INTEL_UEVENT_CONSTRAINT(0x02c2, 0xf), /* UOPS_RETIRED.RETIRE_SLOTS */
596         INTEL_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
597         INTEL_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
598         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
599         INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
600         INTEL_EVENT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
601         INTEL_EVENT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
602         INTEL_EVENT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
603         INTEL_EVENT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
604         EVENT_CONSTRAINT_END
605 };
606
607 struct event_constraint intel_hsw_pebs_event_constraints[] = {
608         INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
609         INTEL_PST_HSW_CONSTRAINT(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
610         INTEL_UEVENT_CONSTRAINT(0x02c2, 0xf), /* UOPS_RETIRED.RETIRE_SLOTS */
611         INTEL_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
612         INTEL_UEVENT_CONSTRAINT(0x01c5, 0xf), /* BR_MISP_RETIRED.CONDITIONAL */
613         INTEL_UEVENT_CONSTRAINT(0x04c5, 0xf), /* BR_MISP_RETIRED.ALL_BRANCHES */
614         INTEL_UEVENT_CONSTRAINT(0x20c5, 0xf), /* BR_MISP_RETIRED.NEAR_TAKEN */
615         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.* */
616         /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
617         INTEL_UEVENT_CONSTRAINT(0x11d0, 0xf),
618         /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
619         INTEL_UEVENT_CONSTRAINT(0x12d0, 0xf),
620         INTEL_UEVENT_CONSTRAINT(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
621         INTEL_UEVENT_CONSTRAINT(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
622         /* MEM_UOPS_RETIRED.SPLIT_STORES */
623         INTEL_UEVENT_CONSTRAINT(0x42d0, 0xf),
624         INTEL_UEVENT_CONSTRAINT(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
625         INTEL_PST_HSW_CONSTRAINT(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
626         INTEL_UEVENT_CONSTRAINT(0x01d1, 0xf), /* MEM_LOAD_UOPS_RETIRED.L1_HIT */
627         INTEL_UEVENT_CONSTRAINT(0x02d1, 0xf), /* MEM_LOAD_UOPS_RETIRED.L2_HIT */
628         INTEL_UEVENT_CONSTRAINT(0x04d1, 0xf), /* MEM_LOAD_UOPS_RETIRED.L3_HIT */
629         /* MEM_LOAD_UOPS_RETIRED.HIT_LFB */
630         INTEL_UEVENT_CONSTRAINT(0x40d1, 0xf),
631         /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_MISS */
632         INTEL_UEVENT_CONSTRAINT(0x01d2, 0xf),
633         /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT */
634         INTEL_UEVENT_CONSTRAINT(0x02d2, 0xf),
635         /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.LOCAL_DRAM */
636         INTEL_UEVENT_CONSTRAINT(0x01d3, 0xf),
637         INTEL_UEVENT_CONSTRAINT(0x04c8, 0xf), /* HLE_RETIRED.Abort */
638         INTEL_UEVENT_CONSTRAINT(0x04c9, 0xf), /* RTM_RETIRED.Abort */
639
640         EVENT_CONSTRAINT_END
641 };
642
643 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
644 {
645         struct event_constraint *c;
646
647         if (!event->attr.precise_ip)
648                 return NULL;
649
650         if (x86_pmu.pebs_constraints) {
651                 for_each_event_constraint(c, x86_pmu.pebs_constraints) {
652                         if ((event->hw.config & c->cmask) == c->code) {
653                                 event->hw.flags |= c->flags;
654                                 return c;
655                         }
656                 }
657         }
658
659         return &emptyconstraint;
660 }
661
662 void intel_pmu_pebs_enable(struct perf_event *event)
663 {
664         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
665         struct hw_perf_event *hwc = &event->hw;
666
667         hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
668
669         cpuc->pebs_enabled |= 1ULL << hwc->idx;
670
671         if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
672                 cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
673         else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
674                 cpuc->pebs_enabled |= 1ULL << 63;
675 }
676
677 void intel_pmu_pebs_disable(struct perf_event *event)
678 {
679         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
680         struct hw_perf_event *hwc = &event->hw;
681
682         cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
683
684         if (event->hw.constraint->flags & PERF_X86_EVENT_PEBS_LDLAT)
685                 cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
686         else if (event->hw.constraint->flags & PERF_X86_EVENT_PEBS_ST)
687                 cpuc->pebs_enabled &= ~(1ULL << 63);
688
689         if (cpuc->enabled)
690                 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
691
692         hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
693 }
694
695 void intel_pmu_pebs_enable_all(void)
696 {
697         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
698
699         if (cpuc->pebs_enabled)
700                 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
701 }
702
703 void intel_pmu_pebs_disable_all(void)
704 {
705         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
706
707         if (cpuc->pebs_enabled)
708                 wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
709 }
710
711 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
712 {
713         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
714         unsigned long from = cpuc->lbr_entries[0].from;
715         unsigned long old_to, to = cpuc->lbr_entries[0].to;
716         unsigned long ip = regs->ip;
717         int is_64bit = 0;
718
719         /*
720          * We don't need to fixup if the PEBS assist is fault like
721          */
722         if (!x86_pmu.intel_cap.pebs_trap)
723                 return 1;
724
725         /*
726          * No LBR entry, no basic block, no rewinding
727          */
728         if (!cpuc->lbr_stack.nr || !from || !to)
729                 return 0;
730
731         /*
732          * Basic blocks should never cross user/kernel boundaries
733          */
734         if (kernel_ip(ip) != kernel_ip(to))
735                 return 0;
736
737         /*
738          * unsigned math, either ip is before the start (impossible) or
739          * the basic block is larger than 1 page (sanity)
740          */
741         if ((ip - to) > PAGE_SIZE)
742                 return 0;
743
744         /*
745          * We sampled a branch insn, rewind using the LBR stack
746          */
747         if (ip == to) {
748                 set_linear_ip(regs, from);
749                 return 1;
750         }
751
752         do {
753                 struct insn insn;
754                 u8 buf[MAX_INSN_SIZE];
755                 void *kaddr;
756
757                 old_to = to;
758                 if (!kernel_ip(ip)) {
759                         int bytes, size = MAX_INSN_SIZE;
760
761                         bytes = copy_from_user_nmi(buf, (void __user *)to, size);
762                         if (bytes != size)
763                                 return 0;
764
765                         kaddr = buf;
766                 } else
767                         kaddr = (void *)to;
768
769 #ifdef CONFIG_X86_64
770                 is_64bit = kernel_ip(to) || !test_thread_flag(TIF_IA32);
771 #endif
772                 insn_init(&insn, kaddr, is_64bit);
773                 insn_get_length(&insn);
774                 to += insn.length;
775         } while (to < ip);
776
777         if (to == ip) {
778                 set_linear_ip(regs, old_to);
779                 return 1;
780         }
781
782         /*
783          * Even though we decoded the basic block, the instruction stream
784          * never matched the given IP, either the TO or the IP got corrupted.
785          */
786         return 0;
787 }
788
789 static void __intel_pmu_pebs_event(struct perf_event *event,
790                                    struct pt_regs *iregs, void *__pebs)
791 {
792         /*
793          * We cast to pebs_record_nhm to get the load latency data
794          * if extra_reg MSR_PEBS_LD_LAT_THRESHOLD used
795          */
796         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
797         struct pebs_record_nhm *pebs = __pebs;
798         struct pebs_record_hsw *pebs_hsw = __pebs;
799         struct perf_sample_data data;
800         struct pt_regs regs;
801         u64 sample_type;
802         int fll, fst;
803
804         if (!intel_pmu_save_and_restart(event))
805                 return;
806
807         fll = event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT;
808         fst = event->hw.flags & (PERF_X86_EVENT_PEBS_ST |
809                                  PERF_X86_EVENT_PEBS_ST_HSW);
810
811         perf_sample_data_init(&data, 0, event->hw.last_period);
812
813         data.period = event->hw.last_period;
814         sample_type = event->attr.sample_type;
815
816         /*
817          * if PEBS-LL or PreciseStore
818          */
819         if (fll || fst) {
820                 /*
821                  * Use latency for weight (only avail with PEBS-LL)
822                  */
823                 if (fll && (sample_type & PERF_SAMPLE_WEIGHT))
824                         data.weight = pebs->lat;
825
826                 /*
827                  * data.data_src encodes the data source
828                  */
829                 if (sample_type & PERF_SAMPLE_DATA_SRC) {
830                         if (fll)
831                                 data.data_src.val = load_latency_data(pebs->dse);
832                         else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
833                                 data.data_src.val =
834                                         precise_store_data_hsw(pebs->dse);
835                         else
836                                 data.data_src.val = precise_store_data(pebs->dse);
837                 }
838         }
839
840         /*
841          * We use the interrupt regs as a base because the PEBS record
842          * does not contain a full regs set, specifically it seems to
843          * lack segment descriptors, which get used by things like
844          * user_mode().
845          *
846          * In the simple case fix up only the IP and BP,SP regs, for
847          * PERF_SAMPLE_IP and PERF_SAMPLE_CALLCHAIN to function properly.
848          * A possible PERF_SAMPLE_REGS will have to transfer all regs.
849          */
850         regs = *iregs;
851         regs.flags = pebs->flags;
852         set_linear_ip(&regs, pebs->ip);
853         regs.bp = pebs->bp;
854         regs.sp = pebs->sp;
855
856         if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format >= 2) {
857                 regs.ip = pebs_hsw->real_ip;
858                 regs.flags |= PERF_EFLAGS_EXACT;
859         } else if (event->attr.precise_ip > 1 && intel_pmu_pebs_fixup_ip(&regs))
860                 regs.flags |= PERF_EFLAGS_EXACT;
861         else
862                 regs.flags &= ~PERF_EFLAGS_EXACT;
863
864         if ((event->attr.sample_type & PERF_SAMPLE_ADDR) &&
865                 x86_pmu.intel_cap.pebs_format >= 1)
866                 data.addr = pebs->dla;
867
868         if (has_branch_stack(event))
869                 data.br_stack = &cpuc->lbr_stack;
870
871         if (perf_event_overflow(event, &data, &regs))
872                 x86_pmu_stop(event, 0);
873 }
874
875 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
876 {
877         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
878         struct debug_store *ds = cpuc->ds;
879         struct perf_event *event = cpuc->events[0]; /* PMC0 only */
880         struct pebs_record_core *at, *top;
881         int n;
882
883         if (!x86_pmu.pebs_active)
884                 return;
885
886         at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
887         top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
888
889         /*
890          * Whatever else happens, drain the thing
891          */
892         ds->pebs_index = ds->pebs_buffer_base;
893
894         if (!test_bit(0, cpuc->active_mask))
895                 return;
896
897         WARN_ON_ONCE(!event);
898
899         if (!event->attr.precise_ip)
900                 return;
901
902         n = top - at;
903         if (n <= 0)
904                 return;
905
906         /*
907          * Should not happen, we program the threshold at 1 and do not
908          * set a reset value.
909          */
910         WARN_ONCE(n > 1, "bad leftover pebs %d\n", n);
911         at += n - 1;
912
913         __intel_pmu_pebs_event(event, iregs, at);
914 }
915
916 static void __intel_pmu_drain_pebs_nhm(struct pt_regs *iregs, void *at,
917                                         void *top)
918 {
919         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
920         struct debug_store *ds = cpuc->ds;
921         struct perf_event *event = NULL;
922         u64 status = 0;
923         int bit;
924
925         ds->pebs_index = ds->pebs_buffer_base;
926
927         for (; at < top; at += x86_pmu.pebs_record_size) {
928                 struct pebs_record_nhm *p = at;
929
930                 for_each_set_bit(bit, (unsigned long *)&p->status,
931                                  x86_pmu.max_pebs_events) {
932                         event = cpuc->events[bit];
933                         if (!test_bit(bit, cpuc->active_mask))
934                                 continue;
935
936                         WARN_ON_ONCE(!event);
937
938                         if (!event->attr.precise_ip)
939                                 continue;
940
941                         if (__test_and_set_bit(bit, (unsigned long *)&status))
942                                 continue;
943
944                         break;
945                 }
946
947                 if (!event || bit >= x86_pmu.max_pebs_events)
948                         continue;
949
950                 __intel_pmu_pebs_event(event, iregs, at);
951         }
952 }
953
954 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
955 {
956         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
957         struct debug_store *ds = cpuc->ds;
958         struct pebs_record_nhm *at, *top;
959         int n;
960
961         if (!x86_pmu.pebs_active)
962                 return;
963
964         at  = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
965         top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
966
967         ds->pebs_index = ds->pebs_buffer_base;
968
969         n = top - at;
970         if (n <= 0)
971                 return;
972
973         /*
974          * Should not happen, we program the threshold at 1 and do not
975          * set a reset value.
976          */
977         WARN_ONCE(n > x86_pmu.max_pebs_events,
978                   "Unexpected number of pebs records %d\n", n);
979
980         return __intel_pmu_drain_pebs_nhm(iregs, at, top);
981 }
982
983 static void intel_pmu_drain_pebs_hsw(struct pt_regs *iregs)
984 {
985         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
986         struct debug_store *ds = cpuc->ds;
987         struct pebs_record_hsw *at, *top;
988         int n;
989
990         if (!x86_pmu.pebs_active)
991                 return;
992
993         at  = (struct pebs_record_hsw *)(unsigned long)ds->pebs_buffer_base;
994         top = (struct pebs_record_hsw *)(unsigned long)ds->pebs_index;
995
996         n = top - at;
997         if (n <= 0)
998                 return;
999         /*
1000          * Should not happen, we program the threshold at 1 and do not
1001          * set a reset value.
1002          */
1003         WARN_ONCE(n > x86_pmu.max_pebs_events,
1004                   "Unexpected number of pebs records %d\n", n);
1005
1006         return __intel_pmu_drain_pebs_nhm(iregs, at, top);
1007 }
1008
1009 /*
1010  * BTS, PEBS probe and setup
1011  */
1012
1013 void intel_ds_init(void)
1014 {
1015         /*
1016          * No support for 32bit formats
1017          */
1018         if (!boot_cpu_has(X86_FEATURE_DTES64))
1019                 return;
1020
1021         x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
1022         x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
1023         if (x86_pmu.pebs) {
1024                 char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
1025                 int format = x86_pmu.intel_cap.pebs_format;
1026
1027                 switch (format) {
1028                 case 0:
1029                         printk(KERN_CONT "PEBS fmt0%c, ", pebs_type);
1030                         x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
1031                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
1032                         break;
1033
1034                 case 1:
1035                         printk(KERN_CONT "PEBS fmt1%c, ", pebs_type);
1036                         x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
1037                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1038                         break;
1039
1040                 case 2:
1041                         pr_cont("PEBS fmt2%c, ", pebs_type);
1042                         x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
1043                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_hsw;
1044                         break;
1045
1046                 default:
1047                         printk(KERN_CONT "no PEBS fmt%d%c, ", format, pebs_type);
1048                         x86_pmu.pebs = 0;
1049                 }
1050         }
1051 }
1052
1053 void perf_restore_debug_store(void)
1054 {
1055         struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
1056
1057         if (!x86_pmu.bts && !x86_pmu.pebs)
1058                 return;
1059
1060         wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
1061 }