]> git.proxmox.com Git - mirror_ubuntu-focal-kernel.git/blame - arch/x86/kernel/cpu/perf_event_intel_ds.c
perf, x86: Implement simple LBR support
[mirror_ubuntu-focal-kernel.git] / arch / x86 / kernel / cpu / perf_event_intel_ds.c
CommitLineData
ca037701
PZ
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
15struct pebs_record_32 {
16 u32 flags, ip;
17 u32 ax, bc, cx, dx;
18 u32 si, di, bp, sp;
19};
20
21 */
22
23struct 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
31struct 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 */
54struct 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
66static 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
78static 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
86static 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
114static 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 kfree(buffer);
132 break;
133 }
134 per_cpu(cpu_hw_events, cpu).ds = ds;
135
136 if (x86_pmu.bts) {
137 buffer = kzalloc(BTS_BUFFER_SIZE, GFP_KERNEL);
138 if (unlikely(!buffer))
139 break;
140
141 max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
142 thresh = max / 16;
143
144 ds->bts_buffer_base = (u64)(unsigned long)buffer;
145 ds->bts_index = ds->bts_buffer_base;
146 ds->bts_absolute_maximum = ds->bts_buffer_base +
147 max * BTS_RECORD_SIZE;
148 ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
149 thresh * BTS_RECORD_SIZE;
150 }
151
152 if (x86_pmu.pebs) {
153 buffer = kzalloc(PEBS_BUFFER_SIZE, GFP_KERNEL);
154 if (unlikely(!buffer))
155 break;
156
157 max = PEBS_BUFFER_SIZE / x86_pmu.pebs_record_size;
158
159 ds->pebs_buffer_base = (u64)(unsigned long)buffer;
160 ds->pebs_index = ds->pebs_buffer_base;
161 ds->pebs_absolute_maximum = ds->pebs_buffer_base +
162 max * x86_pmu.pebs_record_size;
163 /*
164 * Always use single record PEBS
165 */
166 ds->pebs_interrupt_threshold = ds->pebs_buffer_base +
167 x86_pmu.pebs_record_size;
168 }
169
170 err = 0;
171 }
172
173 if (err)
174 release_ds_buffers();
175 else {
176 for_each_online_cpu(cpu)
177 init_debug_store_on_cpu(cpu);
178 }
179
180 put_online_cpus();
181
182 return err;
183}
184
185/*
186 * BTS
187 */
188
189static struct event_constraint bts_constraint =
190 EVENT_CONSTRAINT(0, 1ULL << X86_PMC_IDX_FIXED_BTS, 0);
191
192static void intel_pmu_enable_bts(u64 config)
193{
194 unsigned long debugctlmsr;
195
196 debugctlmsr = get_debugctlmsr();
197
198 debugctlmsr |= X86_DEBUGCTL_TR;
199 debugctlmsr |= X86_DEBUGCTL_BTS;
200 debugctlmsr |= X86_DEBUGCTL_BTINT;
201
202 if (!(config & ARCH_PERFMON_EVENTSEL_OS))
203 debugctlmsr |= X86_DEBUGCTL_BTS_OFF_OS;
204
205 if (!(config & ARCH_PERFMON_EVENTSEL_USR))
206 debugctlmsr |= X86_DEBUGCTL_BTS_OFF_USR;
207
208 update_debugctlmsr(debugctlmsr);
209}
210
211static void intel_pmu_disable_bts(void)
212{
213 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
214 unsigned long debugctlmsr;
215
216 if (!cpuc->ds)
217 return;
218
219 debugctlmsr = get_debugctlmsr();
220
221 debugctlmsr &=
222 ~(X86_DEBUGCTL_TR | X86_DEBUGCTL_BTS | X86_DEBUGCTL_BTINT |
223 X86_DEBUGCTL_BTS_OFF_OS | X86_DEBUGCTL_BTS_OFF_USR);
224
225 update_debugctlmsr(debugctlmsr);
226}
227
228static void intel_pmu_drain_bts_buffer(void)
229{
230 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
231 struct debug_store *ds = cpuc->ds;
232 struct bts_record {
233 u64 from;
234 u64 to;
235 u64 flags;
236 };
237 struct perf_event *event = cpuc->events[X86_PMC_IDX_FIXED_BTS];
238 struct bts_record *at, *top;
239 struct perf_output_handle handle;
240 struct perf_event_header header;
241 struct perf_sample_data data;
242 struct pt_regs regs;
243
244 if (!event)
245 return;
246
247 if (!ds)
248 return;
249
250 at = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
251 top = (struct bts_record *)(unsigned long)ds->bts_index;
252
253 if (top <= at)
254 return;
255
256 ds->bts_index = ds->bts_buffer_base;
257
258 perf_sample_data_init(&data, 0);
259 data.period = event->hw.last_period;
260 regs.ip = 0;
261
262 /*
263 * Prepare a generic sample, i.e. fill in the invariant fields.
264 * We will overwrite the from and to address before we output
265 * the sample.
266 */
267 perf_prepare_sample(&header, &data, event, &regs);
268
269 if (perf_output_begin(&handle, event, header.size * (top - at), 1, 1))
270 return;
271
272 for (; at < top; at++) {
273 data.ip = at->from;
274 data.addr = at->to;
275
276 perf_output_sample(&handle, &header, &data, event);
277 }
278
279 perf_output_end(&handle);
280
281 /* There's new data available. */
282 event->hw.interrupts++;
283 event->pending_kill = POLL_IN;
284}
285
286/*
287 * PEBS
288 */
289
290static struct event_constraint intel_core_pebs_events[] = {
291 PEBS_EVENT_CONSTRAINT(0x00c0, 0x1), /* INSTR_RETIRED.ANY */
292 PEBS_EVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
293 PEBS_EVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
294 PEBS_EVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
295 PEBS_EVENT_CONSTRAINT(0x01cb, 0x1), /* MEM_LOAD_RETIRED.L1D_MISS */
296 PEBS_EVENT_CONSTRAINT(0x02cb, 0x1), /* MEM_LOAD_RETIRED.L1D_LINE_MISS */
297 PEBS_EVENT_CONSTRAINT(0x04cb, 0x1), /* MEM_LOAD_RETIRED.L2_MISS */
298 PEBS_EVENT_CONSTRAINT(0x08cb, 0x1), /* MEM_LOAD_RETIRED.L2_LINE_MISS */
299 PEBS_EVENT_CONSTRAINT(0x10cb, 0x1), /* MEM_LOAD_RETIRED.DTLB_MISS */
300 EVENT_CONSTRAINT_END
301};
302
303static struct event_constraint intel_nehalem_pebs_events[] = {
304 PEBS_EVENT_CONSTRAINT(0x00c0, 0xf), /* INSTR_RETIRED.ANY */
305 PEBS_EVENT_CONSTRAINT(0xfec1, 0xf), /* X87_OPS_RETIRED.ANY */
306 PEBS_EVENT_CONSTRAINT(0x00c5, 0xf), /* BR_INST_RETIRED.MISPRED */
307 PEBS_EVENT_CONSTRAINT(0x1fc7, 0xf), /* SIMD_INST_RETURED.ANY */
308 PEBS_EVENT_CONSTRAINT(0x01cb, 0xf), /* MEM_LOAD_RETIRED.L1D_MISS */
309 PEBS_EVENT_CONSTRAINT(0x02cb, 0xf), /* MEM_LOAD_RETIRED.L1D_LINE_MISS */
310 PEBS_EVENT_CONSTRAINT(0x04cb, 0xf), /* MEM_LOAD_RETIRED.L2_MISS */
311 PEBS_EVENT_CONSTRAINT(0x08cb, 0xf), /* MEM_LOAD_RETIRED.L2_LINE_MISS */
312 PEBS_EVENT_CONSTRAINT(0x10cb, 0xf), /* MEM_LOAD_RETIRED.DTLB_MISS */
313 EVENT_CONSTRAINT_END
314};
315
316static struct event_constraint *
317intel_pebs_constraints(struct perf_event *event)
318{
319 struct event_constraint *c;
320
321 if (!event->attr.precise)
322 return NULL;
323
324 if (x86_pmu.pebs_constraints) {
325 for_each_event_constraint(c, x86_pmu.pebs_constraints) {
326 if ((event->hw.config & c->cmask) == c->code)
327 return c;
328 }
329 }
330
331 return &emptyconstraint;
332}
333
334static void intel_pmu_pebs_enable(struct hw_perf_event *hwc)
335{
336 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
337 u64 val = cpuc->pebs_enabled;
338
339 hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
340
341 val |= 1ULL << hwc->idx;
342 wrmsrl(MSR_IA32_PEBS_ENABLE, val);
343}
344
345static void intel_pmu_pebs_disable(struct hw_perf_event *hwc)
346{
347 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
348 u64 val = cpuc->pebs_enabled;
349
350 val &= ~(1ULL << hwc->idx);
351 wrmsrl(MSR_IA32_PEBS_ENABLE, val);
352
353 hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
354}
355
356static void intel_pmu_pebs_enable_all(void)
357{
358 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
359
360 if (cpuc->pebs_enabled)
361 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
362}
363
364static void intel_pmu_pebs_disable_all(void)
365{
366 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
367
368 if (cpuc->pebs_enabled)
369 wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
370}
371
372static int intel_pmu_save_and_restart(struct perf_event *event);
373static void intel_pmu_disable_event(struct perf_event *event);
374
375static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
376{
377 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
378 struct debug_store *ds = cpuc->ds;
379 struct perf_event *event = cpuc->events[0]; /* PMC0 only */
380 struct pebs_record_core *at, *top;
381 struct perf_sample_data data;
382 struct pt_regs regs;
383 int n;
384
385 if (!event || !ds || !x86_pmu.pebs)
386 return;
387
388 intel_pmu_pebs_disable_all();
389
390 at = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
391 top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
392
393 if (top <= at)
394 goto out;
395
396 ds->pebs_index = ds->pebs_buffer_base;
397
398 if (!intel_pmu_save_and_restart(event))
399 goto out;
400
401 perf_sample_data_init(&data, 0);
402 data.period = event->hw.last_period;
403
404 n = top - at;
405
406 /*
407 * Should not happen, we program the threshold at 1 and do not
408 * set a reset value.
409 */
410 WARN_ON_ONCE(n > 1);
411
412 /*
413 * We use the interrupt regs as a base because the PEBS record
414 * does not contain a full regs set, specifically it seems to
415 * lack segment descriptors, which get used by things like
416 * user_mode().
417 *
418 * In the simple case fix up only the IP and BP,SP regs, for
419 * PERF_SAMPLE_IP and PERF_SAMPLE_CALLCHAIN to function properly.
420 * A possible PERF_SAMPLE_REGS will have to transfer all regs.
421 */
422 regs = *iregs;
423 regs.ip = at->ip;
424 regs.bp = at->bp;
425 regs.sp = at->sp;
426
427 if (perf_event_overflow(event, 1, &data, &regs))
428 intel_pmu_disable_event(event);
429
430out:
431 intel_pmu_pebs_enable_all();
432}
433
434static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
435{
436 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
437 struct debug_store *ds = cpuc->ds;
438 struct pebs_record_nhm *at, *top;
439 struct perf_sample_data data;
440 struct perf_event *event = NULL;
441 struct pt_regs regs;
442 int bit, n;
443
444 if (!ds || !x86_pmu.pebs)
445 return;
446
447 intel_pmu_pebs_disable_all();
448
449 at = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
450 top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
451
452 if (top <= at)
453 goto out;
454
455 ds->pebs_index = ds->pebs_buffer_base;
456
457 n = top - at;
458
459 /*
460 * Should not happen, we program the threshold at 1 and do not
461 * set a reset value.
462 */
463 WARN_ON_ONCE(n > MAX_PEBS_EVENTS);
464
465 for ( ; at < top; at++) {
466 for_each_bit(bit, (unsigned long *)&at->status, MAX_PEBS_EVENTS) {
467 if (!cpuc->events[bit]->attr.precise)
468 continue;
469
470 event = cpuc->events[bit];
471 }
472
473 if (!event)
474 continue;
475
476 if (!intel_pmu_save_and_restart(event))
477 continue;
478
479 perf_sample_data_init(&data, 0);
480 data.period = event->hw.last_period;
481
482 /*
483 * See the comment in intel_pmu_drain_pebs_core()
484 */
485 regs = *iregs;
486 regs.ip = at->ip;
487 regs.bp = at->bp;
488 regs.sp = at->sp;
489
490 if (perf_event_overflow(event, 1, &data, &regs))
491 intel_pmu_disable_event(event);
492 }
493out:
494 intel_pmu_pebs_enable_all();
495}
496
497/*
498 * BTS, PEBS probe and setup
499 */
500
501static void intel_ds_init(void)
502{
503 /*
504 * No support for 32bit formats
505 */
506 if (!boot_cpu_has(X86_FEATURE_DTES64))
507 return;
508
509 x86_pmu.bts = boot_cpu_has(X86_FEATURE_BTS);
510 x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
511 if (x86_pmu.pebs) {
512 int format = 0;
513
514 if (x86_pmu.version > 1) {
515 u64 capabilities;
516 /*
517 * v2+ has a PEBS format field
518 */
519 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
520 format = (capabilities >> 8) & 0xf;
521 }
522
523 switch (format) {
524 case 0:
525 printk(KERN_CONT "PEBS v0, ");
526 x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
527 x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
528 x86_pmu.pebs_constraints = intel_core_pebs_events;
529 break;
530
531 case 1:
532 printk(KERN_CONT "PEBS v1, ");
533 x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
534 x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
535 x86_pmu.pebs_constraints = intel_nehalem_pebs_events;
536 break;
537
538 default:
539 printk(KERN_CONT "PEBS unknown format: %d, ", format);
540 x86_pmu.pebs = 0;
541 break;
542 }
543 }
544}
545
546#else /* CONFIG_CPU_SUP_INTEL */
547
548static int reseve_ds_buffers(void)
549{
550 return 0;
551}
552
553static void release_ds_buffers(void)
554{
555}
556
557#endif /* CONFIG_CPU_SUP_INTEL */