]> git.karo-electronics.de Git - mv-sheeva.git/blob - arch/x86/kernel/cpu/perf_event_intel_ds.c
perf, x86: Deal with multiple state bits for pebs-fmt1
[mv-sheeva.git] / arch / x86 / kernel / cpu / perf_event_intel_ds.c
1 #ifdef CONFIG_CPU_SUP_INTEL
2
3 /* The maximal number of PEBS events: */
4 #define MAX_PEBS_EVENTS         4
5
6 /* The size of a BTS record in bytes: */
7 #define BTS_RECORD_SIZE         24
8
9 #define BTS_BUFFER_SIZE         (PAGE_SIZE << 4)
10 #define PEBS_BUFFER_SIZE        PAGE_SIZE
11
12 /*
13  * pebs_record_32 for p4 and core not supported
14
15 struct pebs_record_32 {
16         u32 flags, ip;
17         u32 ax, bc, cx, dx;
18         u32 si, di, bp, sp;
19 };
20
21  */
22
23 struct pebs_record_core {
24         u64 flags, ip;
25         u64 ax, bx, cx, dx;
26         u64 si, di, bp, sp;
27         u64 r8,  r9,  r10, r11;
28         u64 r12, r13, r14, r15;
29 };
30
31 struct pebs_record_nhm {
32         u64 flags, ip;
33         u64 ax, bx, cx, dx;
34         u64 si, di, bp, sp;
35         u64 r8,  r9,  r10, r11;
36         u64 r12, r13, r14, r15;
37         u64 status, dla, dse, lat;
38 };
39
40 /*
41  * Bits in the debugctlmsr controlling branch tracing.
42  */
43 #define X86_DEBUGCTL_TR                 (1 << 6)
44 #define X86_DEBUGCTL_BTS                (1 << 7)
45 #define X86_DEBUGCTL_BTINT              (1 << 8)
46 #define X86_DEBUGCTL_BTS_OFF_OS         (1 << 9)
47 #define X86_DEBUGCTL_BTS_OFF_USR        (1 << 10)
48
49 /*
50  * A debug store configuration.
51  *
52  * We only support architectures that use 64bit fields.
53  */
54 struct debug_store {
55         u64     bts_buffer_base;
56         u64     bts_index;
57         u64     bts_absolute_maximum;
58         u64     bts_interrupt_threshold;
59         u64     pebs_buffer_base;
60         u64     pebs_index;
61         u64     pebs_absolute_maximum;
62         u64     pebs_interrupt_threshold;
63         u64     pebs_event_reset[MAX_PEBS_EVENTS];
64 };
65
66 static void init_debug_store_on_cpu(int cpu)
67 {
68         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
69
70         if (!ds)
71                 return;
72
73         wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
74                      (u32)((u64)(unsigned long)ds),
75                      (u32)((u64)(unsigned long)ds >> 32));
76 }
77
78 static void fini_debug_store_on_cpu(int cpu)
79 {
80         if (!per_cpu(cpu_hw_events, cpu).ds)
81                 return;
82
83         wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
84 }
85
86 static void release_ds_buffers(void)
87 {
88         int cpu;
89
90         if (!x86_pmu.bts && !x86_pmu.pebs)
91                 return;
92
93         get_online_cpus();
94
95         for_each_online_cpu(cpu)
96                 fini_debug_store_on_cpu(cpu);
97
98         for_each_possible_cpu(cpu) {
99                 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
100
101                 if (!ds)
102                         continue;
103
104                 per_cpu(cpu_hw_events, cpu).ds = NULL;
105
106                 kfree((void *)(unsigned long)ds->pebs_buffer_base);
107                 kfree((void *)(unsigned long)ds->bts_buffer_base);
108                 kfree(ds);
109         }
110
111         put_online_cpus();
112 }
113
114 static int reserve_ds_buffers(void)
115 {
116         int cpu, err = 0;
117
118         if (!x86_pmu.bts && !x86_pmu.pebs)
119                 return 0;
120
121         get_online_cpus();
122
123         for_each_possible_cpu(cpu) {
124                 struct debug_store *ds;
125                 void *buffer;
126                 int max, thresh;
127
128                 err = -ENOMEM;
129                 ds = kzalloc(sizeof(*ds), GFP_KERNEL);
130                 if (unlikely(!ds))
131                         break;
132                 per_cpu(cpu_hw_events, cpu).ds = ds;
133
134                 if (x86_pmu.bts) {
135                         buffer = kzalloc(BTS_BUFFER_SIZE, GFP_KERNEL);
136                         if (unlikely(!buffer))
137                                 break;
138
139                         max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
140                         thresh = max / 16;
141
142                         ds->bts_buffer_base = (u64)(unsigned long)buffer;
143                         ds->bts_index = ds->bts_buffer_base;
144                         ds->bts_absolute_maximum = ds->bts_buffer_base +
145                                 max * BTS_RECORD_SIZE;
146                         ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
147                                 thresh * BTS_RECORD_SIZE;
148                 }
149
150                 if (x86_pmu.pebs) {
151                         buffer = kzalloc(PEBS_BUFFER_SIZE, GFP_KERNEL);
152                         if (unlikely(!buffer))
153                                 break;
154
155                         max = PEBS_BUFFER_SIZE / x86_pmu.pebs_record_size;
156
157                         ds->pebs_buffer_base = (u64)(unsigned long)buffer;
158                         ds->pebs_index = ds->pebs_buffer_base;
159                         ds->pebs_absolute_maximum = ds->pebs_buffer_base +
160                                 max * x86_pmu.pebs_record_size;
161                         /*
162                          * Always use single record PEBS
163                          */
164                         ds->pebs_interrupt_threshold = ds->pebs_buffer_base +
165                                 x86_pmu.pebs_record_size;
166                 }
167
168                 err = 0;
169         }
170
171         if (err)
172                 release_ds_buffers();
173         else {
174                 for_each_online_cpu(cpu)
175                         init_debug_store_on_cpu(cpu);
176         }
177
178         put_online_cpus();
179
180         return err;
181 }
182
183 /*
184  * BTS
185  */
186
187 static struct event_constraint bts_constraint =
188         EVENT_CONSTRAINT(0, 1ULL << X86_PMC_IDX_FIXED_BTS, 0);
189
190 static void intel_pmu_enable_bts(u64 config)
191 {
192         unsigned long debugctlmsr;
193
194         debugctlmsr = get_debugctlmsr();
195
196         debugctlmsr |= X86_DEBUGCTL_TR;
197         debugctlmsr |= X86_DEBUGCTL_BTS;
198         debugctlmsr |= X86_DEBUGCTL_BTINT;
199
200         if (!(config & ARCH_PERFMON_EVENTSEL_OS))
201                 debugctlmsr |= X86_DEBUGCTL_BTS_OFF_OS;
202
203         if (!(config & ARCH_PERFMON_EVENTSEL_USR))
204                 debugctlmsr |= X86_DEBUGCTL_BTS_OFF_USR;
205
206         update_debugctlmsr(debugctlmsr);
207 }
208
209 static void intel_pmu_disable_bts(void)
210 {
211         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
212         unsigned long debugctlmsr;
213
214         if (!cpuc->ds)
215                 return;
216
217         debugctlmsr = get_debugctlmsr();
218
219         debugctlmsr &=
220                 ~(X86_DEBUGCTL_TR | X86_DEBUGCTL_BTS | X86_DEBUGCTL_BTINT |
221                   X86_DEBUGCTL_BTS_OFF_OS | X86_DEBUGCTL_BTS_OFF_USR);
222
223         update_debugctlmsr(debugctlmsr);
224 }
225
226 static void intel_pmu_drain_bts_buffer(void)
227 {
228         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
229         struct debug_store *ds = cpuc->ds;
230         struct bts_record {
231                 u64     from;
232                 u64     to;
233                 u64     flags;
234         };
235         struct perf_event *event = cpuc->events[X86_PMC_IDX_FIXED_BTS];
236         struct bts_record *at, *top;
237         struct perf_output_handle handle;
238         struct perf_event_header header;
239         struct perf_sample_data data;
240         struct pt_regs regs;
241
242         if (!event)
243                 return;
244
245         if (!ds)
246                 return;
247
248         at  = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
249         top = (struct bts_record *)(unsigned long)ds->bts_index;
250
251         if (top <= at)
252                 return;
253
254         ds->bts_index = ds->bts_buffer_base;
255
256         perf_sample_data_init(&data, 0);
257         data.period = event->hw.last_period;
258         regs.ip     = 0;
259
260         /*
261          * Prepare a generic sample, i.e. fill in the invariant fields.
262          * We will overwrite the from and to address before we output
263          * the sample.
264          */
265         perf_prepare_sample(&header, &data, event, &regs);
266
267         if (perf_output_begin(&handle, event, header.size * (top - at), 1, 1))
268                 return;
269
270         for (; at < top; at++) {
271                 data.ip         = at->from;
272                 data.addr       = at->to;
273
274                 perf_output_sample(&handle, &header, &data, event);
275         }
276
277         perf_output_end(&handle);
278
279         /* There's new data available. */
280         event->hw.interrupts++;
281         event->pending_kill = POLL_IN;
282 }
283
284 /*
285  * PEBS
286  */
287
288 static struct event_constraint intel_core_pebs_events[] = {
289         PEBS_EVENT_CONSTRAINT(0x00c0, 0x1), /* INSTR_RETIRED.ANY */
290         PEBS_EVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
291         PEBS_EVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
292         PEBS_EVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
293         PEBS_EVENT_CONSTRAINT(0x01cb, 0x1), /* MEM_LOAD_RETIRED.L1D_MISS */
294         PEBS_EVENT_CONSTRAINT(0x02cb, 0x1), /* MEM_LOAD_RETIRED.L1D_LINE_MISS */
295         PEBS_EVENT_CONSTRAINT(0x04cb, 0x1), /* MEM_LOAD_RETIRED.L2_MISS */
296         PEBS_EVENT_CONSTRAINT(0x08cb, 0x1), /* MEM_LOAD_RETIRED.L2_LINE_MISS */
297         PEBS_EVENT_CONSTRAINT(0x10cb, 0x1), /* MEM_LOAD_RETIRED.DTLB_MISS */
298         EVENT_CONSTRAINT_END
299 };
300
301 static struct event_constraint intel_nehalem_pebs_events[] = {
302         PEBS_EVENT_CONSTRAINT(0x00c0, 0xf), /* INSTR_RETIRED.ANY */
303         PEBS_EVENT_CONSTRAINT(0xfec1, 0xf), /* X87_OPS_RETIRED.ANY */
304         PEBS_EVENT_CONSTRAINT(0x00c5, 0xf), /* BR_INST_RETIRED.MISPRED */
305         PEBS_EVENT_CONSTRAINT(0x1fc7, 0xf), /* SIMD_INST_RETURED.ANY */
306         PEBS_EVENT_CONSTRAINT(0x01cb, 0xf), /* MEM_LOAD_RETIRED.L1D_MISS */
307         PEBS_EVENT_CONSTRAINT(0x02cb, 0xf), /* MEM_LOAD_RETIRED.L1D_LINE_MISS */
308         PEBS_EVENT_CONSTRAINT(0x04cb, 0xf), /* MEM_LOAD_RETIRED.L2_MISS */
309         PEBS_EVENT_CONSTRAINT(0x08cb, 0xf), /* MEM_LOAD_RETIRED.L2_LINE_MISS */
310         PEBS_EVENT_CONSTRAINT(0x10cb, 0xf), /* MEM_LOAD_RETIRED.DTLB_MISS */
311         EVENT_CONSTRAINT_END
312 };
313
314 static struct event_constraint *
315 intel_pebs_constraints(struct perf_event *event)
316 {
317         struct event_constraint *c;
318
319         if (!event->attr.precise)
320                 return NULL;
321
322         if (x86_pmu.pebs_constraints) {
323                 for_each_event_constraint(c, x86_pmu.pebs_constraints) {
324                         if ((event->hw.config & c->cmask) == c->code)
325                                 return c;
326                 }
327         }
328
329         return &emptyconstraint;
330 }
331
332 static void intel_pmu_pebs_enable(struct perf_event *event)
333 {
334         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
335         struct hw_perf_event *hwc = &event->hw;
336         u64 val = cpuc->pebs_enabled;
337
338         hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
339
340         val |= 1ULL << hwc->idx;
341         WARN_ON_ONCE(cpuc->enabled);
342
343         if (x86_pmu.intel_cap.pebs_trap)
344                 intel_pmu_lbr_enable(event);
345 }
346
347 static void intel_pmu_pebs_disable(struct perf_event *event)
348 {
349         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
350         struct hw_perf_event *hwc = &event->hw;
351         u64 val = cpuc->pebs_enabled;
352
353         val &= ~(1ULL << hwc->idx);
354         if (cpuc->enabled)
355                 wrmsrl(MSR_IA32_PEBS_ENABLE, val);
356
357         hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
358
359         if (x86_pmu.intel_cap.pebs_trap)
360                 intel_pmu_lbr_disable(event);
361 }
362
363 static void intel_pmu_pebs_enable_all(void)
364 {
365         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
366
367         if (cpuc->pebs_enabled)
368                 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
369 }
370
371 static void intel_pmu_pebs_disable_all(void)
372 {
373         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
374
375         if (cpuc->pebs_enabled)
376                 wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
377 }
378
379 #include <asm/insn.h>
380
381 static inline bool kernel_ip(unsigned long ip)
382 {
383 #ifdef CONFIG_X86_32
384         return ip > PAGE_OFFSET;
385 #else
386         return (long)ip < 0;
387 #endif
388 }
389
390 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
391 {
392         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
393         unsigned long from = cpuc->lbr_entries[0].from;
394         unsigned long old_to, to = cpuc->lbr_entries[0].to;
395         unsigned long ip = regs->ip;
396
397         /*
398          * We don't need to fixup if the PEBS assist is fault like
399          */
400         if (!x86_pmu.intel_cap.pebs_trap)
401                 return 1;
402
403         /*
404          * No LBR entry, no basic block, no rewinding
405          */
406         if (!cpuc->lbr_stack.nr || !from || !to)
407                 return 0;
408
409         /*
410          * Basic blocks should never cross user/kernel boundaries
411          */
412         if (kernel_ip(ip) != kernel_ip(to))
413                 return 0;
414
415         /*
416          * unsigned math, either ip is before the start (impossible) or
417          * the basic block is larger than 1 page (sanity)
418          */
419         if ((ip - to) > PAGE_SIZE)
420                 return 0;
421
422         /*
423          * We sampled a branch insn, rewind using the LBR stack
424          */
425         if (ip == to) {
426                 regs->ip = from;
427                 return 1;
428         }
429
430         do {
431                 struct insn insn;
432                 u8 buf[MAX_INSN_SIZE];
433                 void *kaddr;
434
435                 old_to = to;
436                 if (!kernel_ip(ip)) {
437                         int bytes, size = MAX_INSN_SIZE;
438
439                         bytes = copy_from_user_nmi(buf, (void __user *)to, size);
440                         if (bytes != size)
441                                 return 0;
442
443                         kaddr = buf;
444                 } else
445                         kaddr = (void *)to;
446
447                 kernel_insn_init(&insn, kaddr);
448                 insn_get_length(&insn);
449                 to += insn.length;
450         } while (to < ip);
451
452         if (to == ip) {
453                 regs->ip = old_to;
454                 return 1;
455         }
456
457         /*
458          * Even though we decoded the basic block, the instruction stream
459          * never matched the given IP, either the TO or the IP got corrupted.
460          */
461         return 0;
462 }
463
464 static int intel_pmu_save_and_restart(struct perf_event *event);
465
466 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
467 {
468         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
469         struct debug_store *ds = cpuc->ds;
470         struct perf_event *event = cpuc->events[0]; /* PMC0 only */
471         struct pebs_record_core *at, *top;
472         struct perf_sample_data data;
473         struct perf_raw_record raw;
474         struct pt_regs regs;
475         int n;
476
477         if (!event || !ds || !x86_pmu.pebs)
478                 return;
479
480         at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
481         top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
482
483         if (top <= at)
484                 return;
485
486         ds->pebs_index = ds->pebs_buffer_base;
487
488         if (!intel_pmu_save_and_restart(event))
489                 return;
490
491         perf_sample_data_init(&data, 0);
492         data.period = event->hw.last_period;
493
494         if (event->attr.sample_type & PERF_SAMPLE_RAW) {
495                 raw.size = x86_pmu.pebs_record_size;
496                 raw.data = at;
497                 data.raw = &raw;
498         }
499
500         n = top - at;
501
502         /*
503          * Should not happen, we program the threshold at 1 and do not
504          * set a reset value.
505          */
506         WARN_ON_ONCE(n > 1);
507
508         /*
509          * We use the interrupt regs as a base because the PEBS record
510          * does not contain a full regs set, specifically it seems to
511          * lack segment descriptors, which get used by things like
512          * user_mode().
513          *
514          * In the simple case fix up only the IP and BP,SP regs, for
515          * PERF_SAMPLE_IP and PERF_SAMPLE_CALLCHAIN to function properly.
516          * A possible PERF_SAMPLE_REGS will have to transfer all regs.
517          */
518         regs = *iregs;
519         regs.ip = at->ip;
520         regs.bp = at->bp;
521         regs.sp = at->sp;
522
523         if (intel_pmu_pebs_fixup_ip(&regs))
524                 regs.flags |= PERF_EFLAGS_EXACT;
525         else
526                 regs.flags &= ~PERF_EFLAGS_EXACT;
527
528         if (perf_event_overflow(event, 1, &data, &regs))
529                 x86_pmu_stop(event);
530 }
531
532 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
533 {
534         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
535         struct debug_store *ds = cpuc->ds;
536         struct pebs_record_nhm *at, *top;
537         struct perf_sample_data data;
538         struct perf_event *event = NULL;
539         struct perf_raw_record raw;
540         struct pt_regs regs;
541         u64 status = 0;
542         int bit, n;
543
544         if (!ds || !x86_pmu.pebs)
545                 return;
546
547         at  = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
548         top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
549
550         if (top <= at)
551                 return;
552
553         ds->pebs_index = ds->pebs_buffer_base;
554
555         n = top - at;
556
557         /*
558          * Should not happen, we program the threshold at 1 and do not
559          * set a reset value.
560          */
561         WARN_ON_ONCE(n > MAX_PEBS_EVENTS);
562
563         for ( ; at < top; at++) {
564                 for_each_bit(bit, (unsigned long *)&at->status, MAX_PEBS_EVENTS) {
565                         event = cpuc->events[bit];
566                         if (!test_bit(bit, cpuc->active_mask))
567                                 continue;
568
569                         WARN_ON_ONCE(!event);
570
571                         if (!event->attr.precise)
572                                 continue;
573
574                         if (__test_and_set_bit(bit, (unsigned long *)&status))
575                                 continue;
576
577                         break;
578                 }
579
580                 if (!event || bit >= MAX_PEBS_EVENTS)
581                         continue;
582
583                 if (!intel_pmu_save_and_restart(event))
584                         continue;
585
586                 perf_sample_data_init(&data, 0);
587                 data.period = event->hw.last_period;
588
589                 if (event->attr.sample_type & PERF_SAMPLE_RAW) {
590                         raw.size = x86_pmu.pebs_record_size;
591                         raw.data = at;
592                         data.raw = &raw;
593                 }
594
595                 /*
596                  * See the comment in intel_pmu_drain_pebs_core()
597                  */
598                 regs = *iregs;
599                 regs.ip = at->ip;
600                 regs.bp = at->bp;
601                 regs.sp = at->sp;
602
603                 if (intel_pmu_pebs_fixup_ip(&regs))
604                         regs.flags |= PERF_EFLAGS_EXACT;
605                 else
606                         regs.flags &= ~PERF_EFLAGS_EXACT;
607
608                 if (perf_event_overflow(event, 1, &data, &regs))
609                         x86_pmu_stop(event);
610         }
611 }
612
613 /*
614  * BTS, PEBS probe and setup
615  */
616
617 static void intel_ds_init(void)
618 {
619         /*
620          * No support for 32bit formats
621          */
622         if (!boot_cpu_has(X86_FEATURE_DTES64))
623                 return;
624
625         x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
626         x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
627         if (x86_pmu.pebs) {
628                 char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
629                 int format = x86_pmu.intel_cap.pebs_format;
630
631                 switch (format) {
632                 case 0:
633                         printk(KERN_CONT "PEBS fmt0%c, ", pebs_type);
634                         x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
635                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
636                         x86_pmu.pebs_constraints = intel_core_pebs_events;
637                         break;
638
639                 case 1:
640                         printk(KERN_CONT "PEBS fmt1%c, ", pebs_type);
641                         x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
642                         x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
643                         x86_pmu.pebs_constraints = intel_nehalem_pebs_events;
644                         break;
645
646                 default:
647                         printk(KERN_CONT "no PEBS fmt%d%c, ", format, pebs_type);
648                         x86_pmu.pebs = 0;
649                         break;
650                 }
651         }
652 }
653
654 #else /* CONFIG_CPU_SUP_INTEL */
655
656 static int reseve_ds_buffers(void)
657 {
658         return 0;
659 }
660
661 static void release_ds_buffers(void)
662 {
663 }
664
665 #endif /* CONFIG_CPU_SUP_INTEL */