1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/perf_event.h>
3 #include <linux/jump_label.h>
4 #include <linux/export.h>
5 #include <linux/types.h>
6 #include <linux/init.h>
7 #include <linux/slab.h>
8 #include <linux/delay.h>
9 #include <linux/jiffies.h>
10 #include <asm/apicdef.h>
11 #include <asm/apic.h>
12 #include <asm/nmi.h>
13
14 #include "../perf_event.h"
15
16 static DEFINE_PER_CPU(unsigned long, perf_nmi_tstamp);
17 static unsigned long perf_nmi_window;
18
19 /* AMD Event 0xFFF: Merge. Used with Large Increment per Cycle events */
20 #define AMD_MERGE_EVENT ((0xFULL << 32) | 0xFFULL)
21 #define AMD_MERGE_EVENT_ENABLE (AMD_MERGE_EVENT | ARCH_PERFMON_EVENTSEL_ENABLE)
22
23 /* PMC Enable and Overflow bits for PerfCntrGlobal* registers */
24 static u64 amd_pmu_global_cntr_mask __read_mostly;
25
26 static __initconst const u64 amd_hw_cache_event_ids
27 [PERF_COUNT_HW_CACHE_MAX]
28 [PERF_COUNT_HW_CACHE_OP_MAX]
29 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
30 {
31 [ C(L1D) ] = {
32 [ C(OP_READ) ] = {
33 [ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses */
34 [ C(RESULT_MISS) ] = 0x0141, /* Data Cache Misses */
35 },
36 [ C(OP_WRITE) ] = {
37 [ C(RESULT_ACCESS) ] = 0,
38 [ C(RESULT_MISS) ] = 0,
39 },
40 [ C(OP_PREFETCH) ] = {
41 [ C(RESULT_ACCESS) ] = 0x0267, /* Data Prefetcher :attempts */
42 [ C(RESULT_MISS) ] = 0x0167, /* Data Prefetcher :cancelled */
43 },
44 },
45 [ C(L1I ) ] = {
46 [ C(OP_READ) ] = {
47 [ C(RESULT_ACCESS) ] = 0x0080, /* Instruction cache fetches */
48 [ C(RESULT_MISS) ] = 0x0081, /* Instruction cache misses */
49 },
50 [ C(OP_WRITE) ] = {
51 [ C(RESULT_ACCESS) ] = -1,
52 [ C(RESULT_MISS) ] = -1,
53 },
54 [ C(OP_PREFETCH) ] = {
55 [ C(RESULT_ACCESS) ] = 0x014B, /* Prefetch Instructions :Load */
56 [ C(RESULT_MISS) ] = 0,
57 },
58 },
59 [ C(LL ) ] = {
60 [ C(OP_READ) ] = {
61 [ C(RESULT_ACCESS) ] = 0x037D, /* Requests to L2 Cache :IC+DC */
62 [ C(RESULT_MISS) ] = 0x037E, /* L2 Cache Misses : IC+DC */
63 },
64 [ C(OP_WRITE) ] = {
65 [ C(RESULT_ACCESS) ] = 0x017F, /* L2 Fill/Writeback */
66 [ C(RESULT_MISS) ] = 0,
67 },
68 [ C(OP_PREFETCH) ] = {
69 [ C(RESULT_ACCESS) ] = 0,
70 [ C(RESULT_MISS) ] = 0,
71 },
72 },
73 [ C(DTLB) ] = {
74 [ C(OP_READ) ] = {
75 [ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses */
76 [ C(RESULT_MISS) ] = 0x0746, /* L1_DTLB_AND_L2_DLTB_MISS.ALL */
77 },
78 [ C(OP_WRITE) ] = {
79 [ C(RESULT_ACCESS) ] = 0,
80 [ C(RESULT_MISS) ] = 0,
81 },
82 [ C(OP_PREFETCH) ] = {
83 [ C(RESULT_ACCESS) ] = 0,
84 [ C(RESULT_MISS) ] = 0,
85 },
86 },
87 [ C(ITLB) ] = {
88 [ C(OP_READ) ] = {
89 [ C(RESULT_ACCESS) ] = 0x0080, /* Instruction fecthes */
90 [ C(RESULT_MISS) ] = 0x0385, /* L1_ITLB_AND_L2_ITLB_MISS.ALL */
91 },
92 [ C(OP_WRITE) ] = {
93 [ C(RESULT_ACCESS) ] = -1,
94 [ C(RESULT_MISS) ] = -1,
95 },
96 [ C(OP_PREFETCH) ] = {
97 [ C(RESULT_ACCESS) ] = -1,
98 [ C(RESULT_MISS) ] = -1,
99 },
100 },
101 [ C(BPU ) ] = {
102 [ C(OP_READ) ] = {
103 [ C(RESULT_ACCESS) ] = 0x00c2, /* Retired Branch Instr. */
104 [ C(RESULT_MISS) ] = 0x00c3, /* Retired Mispredicted BI */
105 },
106 [ C(OP_WRITE) ] = {
107 [ C(RESULT_ACCESS) ] = -1,
108 [ C(RESULT_MISS) ] = -1,
109 },
110 [ C(OP_PREFETCH) ] = {
111 [ C(RESULT_ACCESS) ] = -1,
112 [ C(RESULT_MISS) ] = -1,
113 },
114 },
115 [ C(NODE) ] = {
116 [ C(OP_READ) ] = {
117 [ C(RESULT_ACCESS) ] = 0xb8e9, /* CPU Request to Memory, l+r */
118 [ C(RESULT_MISS) ] = 0x98e9, /* CPU Request to Memory, r */
119 },
120 [ C(OP_WRITE) ] = {
121 [ C(RESULT_ACCESS) ] = -1,
122 [ C(RESULT_MISS) ] = -1,
123 },
124 [ C(OP_PREFETCH) ] = {
125 [ C(RESULT_ACCESS) ] = -1,
126 [ C(RESULT_MISS) ] = -1,
127 },
128 },
129 };
130
131 static __initconst const u64 amd_hw_cache_event_ids_f17h
132 [PERF_COUNT_HW_CACHE_MAX]
133 [PERF_COUNT_HW_CACHE_OP_MAX]
134 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
135 [C(L1D)] = {
136 [C(OP_READ)] = {
137 [C(RESULT_ACCESS)] = 0x0040, /* Data Cache Accesses */
138 [C(RESULT_MISS)] = 0xc860, /* L2$ access from DC Miss */
139 },
140 [C(OP_WRITE)] = {
141 [C(RESULT_ACCESS)] = 0,
142 [C(RESULT_MISS)] = 0,
143 },
144 [C(OP_PREFETCH)] = {
145 [C(RESULT_ACCESS)] = 0xff5a, /* h/w prefetch DC Fills */
146 [C(RESULT_MISS)] = 0,
147 },
148 },
149 [C(L1I)] = {
150 [C(OP_READ)] = {
151 [C(RESULT_ACCESS)] = 0x0080, /* Instruction cache fetches */
152 [C(RESULT_MISS)] = 0x0081, /* Instruction cache misses */
153 },
154 [C(OP_WRITE)] = {
155 [C(RESULT_ACCESS)] = -1,
156 [C(RESULT_MISS)] = -1,
157 },
158 [C(OP_PREFETCH)] = {
159 [C(RESULT_ACCESS)] = 0,
160 [C(RESULT_MISS)] = 0,
161 },
162 },
163 [C(LL)] = {
164 [C(OP_READ)] = {
165 [C(RESULT_ACCESS)] = 0,
166 [C(RESULT_MISS)] = 0,
167 },
168 [C(OP_WRITE)] = {
169 [C(RESULT_ACCESS)] = 0,
170 [C(RESULT_MISS)] = 0,
171 },
172 [C(OP_PREFETCH)] = {
173 [C(RESULT_ACCESS)] = 0,
174 [C(RESULT_MISS)] = 0,
175 },
176 },
177 [C(DTLB)] = {
178 [C(OP_READ)] = {
179 [C(RESULT_ACCESS)] = 0xff45, /* All L2 DTLB accesses */
180 [C(RESULT_MISS)] = 0xf045, /* L2 DTLB misses (PT walks) */
181 },
182 [C(OP_WRITE)] = {
183 [C(RESULT_ACCESS)] = 0,
184 [C(RESULT_MISS)] = 0,
185 },
186 [C(OP_PREFETCH)] = {
187 [C(RESULT_ACCESS)] = 0,
188 [C(RESULT_MISS)] = 0,
189 },
190 },
191 [C(ITLB)] = {
192 [C(OP_READ)] = {
193 [C(RESULT_ACCESS)] = 0x0084, /* L1 ITLB misses, L2 ITLB hits */
194 [C(RESULT_MISS)] = 0xff85, /* L1 ITLB misses, L2 misses */
195 },
196 [C(OP_WRITE)] = {
197 [C(RESULT_ACCESS)] = -1,
198 [C(RESULT_MISS)] = -1,
199 },
200 [C(OP_PREFETCH)] = {
201 [C(RESULT_ACCESS)] = -1,
202 [C(RESULT_MISS)] = -1,
203 },
204 },
205 [C(BPU)] = {
206 [C(OP_READ)] = {
207 [C(RESULT_ACCESS)] = 0x00c2, /* Retired Branch Instr. */
208 [C(RESULT_MISS)] = 0x00c3, /* Retired Mispredicted BI */
209 },
210 [C(OP_WRITE)] = {
211 [C(RESULT_ACCESS)] = -1,
212 [C(RESULT_MISS)] = -1,
213 },
214 [C(OP_PREFETCH)] = {
215 [C(RESULT_ACCESS)] = -1,
216 [C(RESULT_MISS)] = -1,
217 },
218 },
219 [C(NODE)] = {
220 [C(OP_READ)] = {
221 [C(RESULT_ACCESS)] = 0,
222 [C(RESULT_MISS)] = 0,
223 },
224 [C(OP_WRITE)] = {
225 [C(RESULT_ACCESS)] = -1,
226 [C(RESULT_MISS)] = -1,
227 },
228 [C(OP_PREFETCH)] = {
229 [C(RESULT_ACCESS)] = -1,
230 [C(RESULT_MISS)] = -1,
231 },
232 },
233 };
234
235 /*
236 * AMD Performance Monitor K7 and later, up to and including Family 16h:
237 */
238 static const u64 amd_perfmon_event_map[PERF_COUNT_HW_MAX] =
239 {
240 [PERF_COUNT_HW_CPU_CYCLES] = 0x0076,
241 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
242 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x077d,
243 [PERF_COUNT_HW_CACHE_MISSES] = 0x077e,
244 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c2,
245 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c3,
246 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x00d0, /* "Decoder empty" event */
247 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x00d1, /* "Dispatch stalls" event */
248 };
249
250 /*
251 * AMD Performance Monitor Family 17h and later:
252 */
253 static const u64 amd_zen1_perfmon_event_map[PERF_COUNT_HW_MAX] =
254 {
255 [PERF_COUNT_HW_CPU_CYCLES] = 0x0076,
256 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
257 [PERF_COUNT_HW_CACHE_REFERENCES] = 0xff60,
258 [PERF_COUNT_HW_CACHE_MISSES] = 0x0964,
259 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c2,
260 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c3,
261 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x0287,
262 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x0187,
263 };
264
265 static const u64 amd_zen2_perfmon_event_map[PERF_COUNT_HW_MAX] =
266 {
267 [PERF_COUNT_HW_CPU_CYCLES] = 0x0076,
268 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
269 [PERF_COUNT_HW_CACHE_REFERENCES] = 0xff60,
270 [PERF_COUNT_HW_CACHE_MISSES] = 0x0964,
271 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c2,
272 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c3,
273 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x00a9,
274 };
275
276 static const u64 amd_zen4_perfmon_event_map[PERF_COUNT_HW_MAX] =
277 {
278 [PERF_COUNT_HW_CPU_CYCLES] = 0x0076,
279 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
280 [PERF_COUNT_HW_CACHE_REFERENCES] = 0xff60,
281 [PERF_COUNT_HW_CACHE_MISSES] = 0x0964,
282 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c2,
283 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c3,
284 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x00a9,
285 [PERF_COUNT_HW_REF_CPU_CYCLES] = 0x100000120,
286 };
287
amd_pmu_event_map(int hw_event)288 static u64 amd_pmu_event_map(int hw_event)
289 {
290 if (cpu_feature_enabled(X86_FEATURE_ZEN4) || boot_cpu_data.x86 >= 0x1a)
291 return amd_zen4_perfmon_event_map[hw_event];
292
293 if (cpu_feature_enabled(X86_FEATURE_ZEN2) || boot_cpu_data.x86 >= 0x19)
294 return amd_zen2_perfmon_event_map[hw_event];
295
296 if (cpu_feature_enabled(X86_FEATURE_ZEN1))
297 return amd_zen1_perfmon_event_map[hw_event];
298
299 return amd_perfmon_event_map[hw_event];
300 }
301
302 /*
303 * Previously calculated offsets
304 */
305 static unsigned int event_offsets[X86_PMC_IDX_MAX] __read_mostly;
306 static unsigned int count_offsets[X86_PMC_IDX_MAX] __read_mostly;
307
308 /*
309 * Legacy CPUs:
310 * 4 counters starting at 0xc0010000 each offset by 1
311 *
312 * CPUs with core performance counter extensions:
313 * 6 counters starting at 0xc0010200 each offset by 2
314 */
amd_pmu_addr_offset(int index,bool eventsel)315 static inline int amd_pmu_addr_offset(int index, bool eventsel)
316 {
317 int offset;
318
319 if (!index)
320 return index;
321
322 if (eventsel)
323 offset = event_offsets[index];
324 else
325 offset = count_offsets[index];
326
327 if (offset)
328 return offset;
329
330 if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
331 offset = index;
332 else
333 offset = index << 1;
334
335 if (eventsel)
336 event_offsets[index] = offset;
337 else
338 count_offsets[index] = offset;
339
340 return offset;
341 }
342
343 /*
344 * AMD64 events are detected based on their event codes.
345 */
amd_get_event_code(struct hw_perf_event * hwc)346 static inline unsigned int amd_get_event_code(struct hw_perf_event *hwc)
347 {
348 return ((hwc->config >> 24) & 0x0f00) | (hwc->config & 0x00ff);
349 }
350
amd_is_pair_event_code(struct hw_perf_event * hwc)351 static inline bool amd_is_pair_event_code(struct hw_perf_event *hwc)
352 {
353 if (!(x86_pmu.flags & PMU_FL_PAIR))
354 return false;
355
356 switch (amd_get_event_code(hwc)) {
357 case 0x003: return true; /* Retired SSE/AVX FLOPs */
358 default: return false;
359 }
360 }
361
362 DEFINE_STATIC_CALL_RET0(amd_pmu_branch_hw_config, *x86_pmu.hw_config);
363
amd_core_hw_config(struct perf_event * event)364 static int amd_core_hw_config(struct perf_event *event)
365 {
366 if (event->attr.exclude_host && event->attr.exclude_guest)
367 /*
368 * When HO == GO == 1 the hardware treats that as GO == HO == 0
369 * and will count in both modes. We don't want to count in that
370 * case so we emulate no-counting by setting US = OS = 0.
371 */
372 event->hw.config &= ~(ARCH_PERFMON_EVENTSEL_USR |
373 ARCH_PERFMON_EVENTSEL_OS);
374 else if (event->attr.exclude_host)
375 event->hw.config |= AMD64_EVENTSEL_GUESTONLY;
376 else if (event->attr.exclude_guest)
377 event->hw.config |= AMD64_EVENTSEL_HOSTONLY;
378
379 if ((x86_pmu.flags & PMU_FL_PAIR) && amd_is_pair_event_code(&event->hw))
380 event->hw.flags |= PERF_X86_EVENT_PAIR;
381
382 if (has_branch_stack(event))
383 return static_call(amd_pmu_branch_hw_config)(event);
384
385 return 0;
386 }
387
amd_is_nb_event(struct hw_perf_event * hwc)388 static inline int amd_is_nb_event(struct hw_perf_event *hwc)
389 {
390 return (hwc->config & 0xe0) == 0xe0;
391 }
392
amd_has_nb(struct cpu_hw_events * cpuc)393 static inline int amd_has_nb(struct cpu_hw_events *cpuc)
394 {
395 struct amd_nb *nb = cpuc->amd_nb;
396
397 return nb && nb->nb_id != -1;
398 }
399
amd_pmu_hw_config(struct perf_event * event)400 static int amd_pmu_hw_config(struct perf_event *event)
401 {
402 int ret;
403
404 /* pass precise event sampling to ibs: */
405 if (event->attr.precise_ip && get_ibs_caps())
406 return forward_event_to_ibs(event);
407
408 if (has_branch_stack(event) && !x86_pmu.lbr_nr)
409 return -EOPNOTSUPP;
410
411 ret = x86_pmu_hw_config(event);
412 if (ret)
413 return ret;
414
415 if (event->attr.type == PERF_TYPE_RAW)
416 event->hw.config |= event->attr.config & AMD64_RAW_EVENT_MASK;
417
418 return amd_core_hw_config(event);
419 }
420
__amd_put_nb_event_constraints(struct cpu_hw_events * cpuc,struct perf_event * event)421 static void __amd_put_nb_event_constraints(struct cpu_hw_events *cpuc,
422 struct perf_event *event)
423 {
424 struct amd_nb *nb = cpuc->amd_nb;
425 int i;
426
427 /*
428 * need to scan whole list because event may not have
429 * been assigned during scheduling
430 *
431 * no race condition possible because event can only
432 * be removed on one CPU at a time AND PMU is disabled
433 * when we come here
434 */
435 for_each_set_bit(i, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
436 struct perf_event *tmp = event;
437
438 if (try_cmpxchg(nb->owners + i, &tmp, NULL))
439 break;
440 }
441 }
442
443 /*
444 * AMD64 NorthBridge events need special treatment because
445 * counter access needs to be synchronized across all cores
446 * of a package. Refer to BKDG section 3.12
447 *
448 * NB events are events measuring L3 cache, Hypertransport
449 * traffic. They are identified by an event code >= 0xe00.
450 * They measure events on the NorthBride which is shared
451 * by all cores on a package. NB events are counted on a
452 * shared set of counters. When a NB event is programmed
453 * in a counter, the data actually comes from a shared
454 * counter. Thus, access to those counters needs to be
455 * synchronized.
456 *
457 * We implement the synchronization such that no two cores
458 * can be measuring NB events using the same counters. Thus,
459 * we maintain a per-NB allocation table. The available slot
460 * is propagated using the event_constraint structure.
461 *
462 * We provide only one choice for each NB event based on
463 * the fact that only NB events have restrictions. Consequently,
464 * if a counter is available, there is a guarantee the NB event
465 * will be assigned to it. If no slot is available, an empty
466 * constraint is returned and scheduling will eventually fail
467 * for this event.
468 *
469 * Note that all cores attached the same NB compete for the same
470 * counters to host NB events, this is why we use atomic ops. Some
471 * multi-chip CPUs may have more than one NB.
472 *
473 * Given that resources are allocated (cmpxchg), they must be
474 * eventually freed for others to use. This is accomplished by
475 * calling __amd_put_nb_event_constraints()
476 *
477 * Non NB events are not impacted by this restriction.
478 */
479 static struct event_constraint *
__amd_get_nb_event_constraints(struct cpu_hw_events * cpuc,struct perf_event * event,struct event_constraint * c)480 __amd_get_nb_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
481 struct event_constraint *c)
482 {
483 struct hw_perf_event *hwc = &event->hw;
484 struct amd_nb *nb = cpuc->amd_nb;
485 struct perf_event *old;
486 int idx, new = -1;
487
488 if (!c)
489 c = &unconstrained;
490
491 if (cpuc->is_fake)
492 return c;
493
494 /*
495 * detect if already present, if so reuse
496 *
497 * cannot merge with actual allocation
498 * because of possible holes
499 *
500 * event can already be present yet not assigned (in hwc->idx)
501 * because of successive calls to x86_schedule_events() from
502 * hw_perf_group_sched_in() without hw_perf_enable()
503 */
504 for_each_set_bit(idx, c->idxmsk, x86_pmu_max_num_counters(NULL)) {
505 if (new == -1 || hwc->idx == idx)
506 /* assign free slot, prefer hwc->idx */
507 old = cmpxchg(nb->owners + idx, NULL, event);
508 else if (nb->owners[idx] == event)
509 /* event already present */
510 old = event;
511 else
512 continue;
513
514 if (old && old != event)
515 continue;
516
517 /* reassign to this slot */
518 if (new != -1)
519 cmpxchg(nb->owners + new, event, NULL);
520 new = idx;
521
522 /* already present, reuse */
523 if (old == event)
524 break;
525 }
526
527 if (new == -1)
528 return &emptyconstraint;
529
530 return &nb->event_constraints[new];
531 }
532
amd_alloc_nb(int cpu)533 static struct amd_nb *amd_alloc_nb(int cpu)
534 {
535 struct amd_nb *nb;
536 int i;
537
538 nb = kzalloc_node(sizeof(struct amd_nb), GFP_KERNEL, cpu_to_node(cpu));
539 if (!nb)
540 return NULL;
541
542 nb->nb_id = -1;
543
544 /*
545 * initialize all possible NB constraints
546 */
547 for_each_set_bit(i, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
548 __set_bit(i, nb->event_constraints[i].idxmsk);
549 nb->event_constraints[i].weight = 1;
550 }
551 return nb;
552 }
553
554 typedef void (amd_pmu_branch_reset_t)(void);
555 DEFINE_STATIC_CALL_NULL(amd_pmu_branch_reset, amd_pmu_branch_reset_t);
556
amd_pmu_cpu_reset(int cpu)557 static void amd_pmu_cpu_reset(int cpu)
558 {
559 if (x86_pmu.lbr_nr)
560 static_call(amd_pmu_branch_reset)();
561
562 if (x86_pmu.version < 2)
563 return;
564
565 /* Clear enable bits i.e. PerfCntrGlobalCtl.PerfCntrEn */
566 wrmsrl(MSR_AMD64_PERF_CNTR_GLOBAL_CTL, 0);
567
568 /*
569 * Clear freeze and overflow bits i.e. PerfCntrGLobalStatus.LbrFreeze
570 * and PerfCntrGLobalStatus.PerfCntrOvfl
571 */
572 wrmsrl(MSR_AMD64_PERF_CNTR_GLOBAL_STATUS_CLR,
573 GLOBAL_STATUS_LBRS_FROZEN | amd_pmu_global_cntr_mask);
574 }
575
amd_pmu_cpu_prepare(int cpu)576 static int amd_pmu_cpu_prepare(int cpu)
577 {
578 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
579
580 cpuc->lbr_sel = kzalloc_node(sizeof(struct er_account), GFP_KERNEL,
581 cpu_to_node(cpu));
582 if (!cpuc->lbr_sel)
583 return -ENOMEM;
584
585 WARN_ON_ONCE(cpuc->amd_nb);
586
587 if (!x86_pmu.amd_nb_constraints)
588 return 0;
589
590 cpuc->amd_nb = amd_alloc_nb(cpu);
591 if (cpuc->amd_nb)
592 return 0;
593
594 kfree(cpuc->lbr_sel);
595 cpuc->lbr_sel = NULL;
596
597 return -ENOMEM;
598 }
599
amd_pmu_cpu_starting(int cpu)600 static void amd_pmu_cpu_starting(int cpu)
601 {
602 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
603 void **onln = &cpuc->kfree_on_online[X86_PERF_KFREE_SHARED];
604 struct amd_nb *nb;
605 int i, nb_id;
606
607 cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
608 amd_pmu_cpu_reset(cpu);
609
610 if (!x86_pmu.amd_nb_constraints)
611 return;
612
613 nb_id = topology_amd_node_id(cpu);
614 WARN_ON_ONCE(nb_id == BAD_APICID);
615
616 for_each_online_cpu(i) {
617 nb = per_cpu(cpu_hw_events, i).amd_nb;
618 if (WARN_ON_ONCE(!nb))
619 continue;
620
621 if (nb->nb_id == nb_id) {
622 *onln = cpuc->amd_nb;
623 cpuc->amd_nb = nb;
624 break;
625 }
626 }
627
628 cpuc->amd_nb->nb_id = nb_id;
629 cpuc->amd_nb->refcnt++;
630 }
631
amd_pmu_cpu_dead(int cpu)632 static void amd_pmu_cpu_dead(int cpu)
633 {
634 struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
635
636 kfree(cpuhw->lbr_sel);
637 cpuhw->lbr_sel = NULL;
638
639 if (!x86_pmu.amd_nb_constraints)
640 return;
641
642 if (cpuhw->amd_nb) {
643 struct amd_nb *nb = cpuhw->amd_nb;
644
645 if (nb->nb_id == -1 || --nb->refcnt == 0)
646 kfree(nb);
647
648 cpuhw->amd_nb = NULL;
649 }
650 }
651
amd_pmu_set_global_ctl(u64 ctl)652 static __always_inline void amd_pmu_set_global_ctl(u64 ctl)
653 {
654 wrmsrl(MSR_AMD64_PERF_CNTR_GLOBAL_CTL, ctl);
655 }
656
amd_pmu_get_global_status(void)657 static inline u64 amd_pmu_get_global_status(void)
658 {
659 u64 status;
660
661 /* PerfCntrGlobalStatus is read-only */
662 rdmsrl(MSR_AMD64_PERF_CNTR_GLOBAL_STATUS, status);
663
664 return status;
665 }
666
amd_pmu_ack_global_status(u64 status)667 static inline void amd_pmu_ack_global_status(u64 status)
668 {
669 /*
670 * PerfCntrGlobalStatus is read-only but an overflow acknowledgment
671 * mechanism exists; writing 1 to a bit in PerfCntrGlobalStatusClr
672 * clears the same bit in PerfCntrGlobalStatus
673 */
674
675 wrmsrl(MSR_AMD64_PERF_CNTR_GLOBAL_STATUS_CLR, status);
676 }
677
amd_pmu_test_overflow_topbit(int idx)678 static bool amd_pmu_test_overflow_topbit(int idx)
679 {
680 u64 counter;
681
682 rdmsrl(x86_pmu_event_addr(idx), counter);
683
684 return !(counter & BIT_ULL(x86_pmu.cntval_bits - 1));
685 }
686
amd_pmu_test_overflow_status(int idx)687 static bool amd_pmu_test_overflow_status(int idx)
688 {
689 return amd_pmu_get_global_status() & BIT_ULL(idx);
690 }
691
692 DEFINE_STATIC_CALL(amd_pmu_test_overflow, amd_pmu_test_overflow_topbit);
693
694 /*
695 * When a PMC counter overflows, an NMI is used to process the event and
696 * reset the counter. NMI latency can result in the counter being updated
697 * before the NMI can run, which can result in what appear to be spurious
698 * NMIs. This function is intended to wait for the NMI to run and reset
699 * the counter to avoid possible unhandled NMI messages.
700 */
701 #define OVERFLOW_WAIT_COUNT 50
702
amd_pmu_wait_on_overflow(int idx)703 static void amd_pmu_wait_on_overflow(int idx)
704 {
705 unsigned int i;
706
707 /*
708 * Wait for the counter to be reset if it has overflowed. This loop
709 * should exit very, very quickly, but just in case, don't wait
710 * forever...
711 */
712 for (i = 0; i < OVERFLOW_WAIT_COUNT; i++) {
713 if (!static_call(amd_pmu_test_overflow)(idx))
714 break;
715
716 /* Might be in IRQ context, so can't sleep */
717 udelay(1);
718 }
719 }
720
amd_pmu_check_overflow(void)721 static void amd_pmu_check_overflow(void)
722 {
723 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
724 int idx;
725
726 /*
727 * This shouldn't be called from NMI context, but add a safeguard here
728 * to return, since if we're in NMI context we can't wait for an NMI
729 * to reset an overflowed counter value.
730 */
731 if (in_nmi())
732 return;
733
734 /*
735 * Check each counter for overflow and wait for it to be reset by the
736 * NMI if it has overflowed. This relies on the fact that all active
737 * counters are always enabled when this function is called and
738 * ARCH_PERFMON_EVENTSEL_INT is always set.
739 */
740 for_each_set_bit(idx, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
741 if (!test_bit(idx, cpuc->active_mask))
742 continue;
743
744 amd_pmu_wait_on_overflow(idx);
745 }
746 }
747
amd_pmu_enable_event(struct perf_event * event)748 static void amd_pmu_enable_event(struct perf_event *event)
749 {
750 x86_pmu_enable_event(event);
751 }
752
amd_pmu_enable_all(int added)753 static void amd_pmu_enable_all(int added)
754 {
755 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
756 int idx;
757
758 amd_brs_enable_all();
759
760 for_each_set_bit(idx, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
761 /* only activate events which are marked as active */
762 if (!test_bit(idx, cpuc->active_mask))
763 continue;
764
765 amd_pmu_enable_event(cpuc->events[idx]);
766 }
767 }
768
amd_pmu_v2_enable_event(struct perf_event * event)769 static void amd_pmu_v2_enable_event(struct perf_event *event)
770 {
771 struct hw_perf_event *hwc = &event->hw;
772
773 /*
774 * Testing cpu_hw_events.enabled should be skipped in this case unlike
775 * in x86_pmu_enable_event().
776 *
777 * Since cpu_hw_events.enabled is set only after returning from
778 * x86_pmu_start(), the PMCs must be programmed and kept ready.
779 * Counting starts only after x86_pmu_enable_all() is called.
780 */
781 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
782 }
783
amd_pmu_core_enable_all(void)784 static __always_inline void amd_pmu_core_enable_all(void)
785 {
786 amd_pmu_set_global_ctl(amd_pmu_global_cntr_mask);
787 }
788
amd_pmu_v2_enable_all(int added)789 static void amd_pmu_v2_enable_all(int added)
790 {
791 amd_pmu_lbr_enable_all();
792 amd_pmu_core_enable_all();
793 }
794
amd_pmu_disable_event(struct perf_event * event)795 static void amd_pmu_disable_event(struct perf_event *event)
796 {
797 x86_pmu_disable_event(event);
798
799 /*
800 * This can be called from NMI context (via x86_pmu_stop). The counter
801 * may have overflowed, but either way, we'll never see it get reset
802 * by the NMI if we're already in the NMI. And the NMI latency support
803 * below will take care of any pending NMI that might have been
804 * generated by the overflow.
805 */
806 if (in_nmi())
807 return;
808
809 amd_pmu_wait_on_overflow(event->hw.idx);
810 }
811
amd_pmu_disable_all(void)812 static void amd_pmu_disable_all(void)
813 {
814 amd_brs_disable_all();
815 x86_pmu_disable_all();
816 amd_pmu_check_overflow();
817 }
818
amd_pmu_core_disable_all(void)819 static __always_inline void amd_pmu_core_disable_all(void)
820 {
821 amd_pmu_set_global_ctl(0);
822 }
823
amd_pmu_v2_disable_all(void)824 static void amd_pmu_v2_disable_all(void)
825 {
826 amd_pmu_core_disable_all();
827 amd_pmu_lbr_disable_all();
828 amd_pmu_check_overflow();
829 }
830
831 DEFINE_STATIC_CALL_NULL(amd_pmu_branch_add, *x86_pmu.add);
832
amd_pmu_add_event(struct perf_event * event)833 static void amd_pmu_add_event(struct perf_event *event)
834 {
835 if (needs_branch_stack(event))
836 static_call(amd_pmu_branch_add)(event);
837 }
838
839 DEFINE_STATIC_CALL_NULL(amd_pmu_branch_del, *x86_pmu.del);
840
amd_pmu_del_event(struct perf_event * event)841 static void amd_pmu_del_event(struct perf_event *event)
842 {
843 if (needs_branch_stack(event))
844 static_call(amd_pmu_branch_del)(event);
845 }
846
847 /*
848 * Because of NMI latency, if multiple PMC counters are active or other sources
849 * of NMIs are received, the perf NMI handler can handle one or more overflowed
850 * PMC counters outside of the NMI associated with the PMC overflow. If the NMI
851 * doesn't arrive at the LAPIC in time to become a pending NMI, then the kernel
852 * back-to-back NMI support won't be active. This PMC handler needs to take into
853 * account that this can occur, otherwise this could result in unknown NMI
854 * messages being issued. Examples of this is PMC overflow while in the NMI
855 * handler when multiple PMCs are active or PMC overflow while handling some
856 * other source of an NMI.
857 *
858 * Attempt to mitigate this by creating an NMI window in which un-handled NMIs
859 * received during this window will be claimed. This prevents extending the
860 * window past when it is possible that latent NMIs should be received. The
861 * per-CPU perf_nmi_tstamp will be set to the window end time whenever perf has
862 * handled a counter. When an un-handled NMI is received, it will be claimed
863 * only if arriving within that window.
864 */
amd_pmu_adjust_nmi_window(int handled)865 static inline int amd_pmu_adjust_nmi_window(int handled)
866 {
867 /*
868 * If a counter was handled, record a timestamp such that un-handled
869 * NMIs will be claimed if arriving within that window.
870 */
871 if (handled) {
872 this_cpu_write(perf_nmi_tstamp, jiffies + perf_nmi_window);
873
874 return handled;
875 }
876
877 if (time_after(jiffies, this_cpu_read(perf_nmi_tstamp)))
878 return NMI_DONE;
879
880 return NMI_HANDLED;
881 }
882
amd_pmu_handle_irq(struct pt_regs * regs)883 static int amd_pmu_handle_irq(struct pt_regs *regs)
884 {
885 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
886 int handled;
887 int pmu_enabled;
888
889 /*
890 * Save the PMU state.
891 * It needs to be restored when leaving the handler.
892 */
893 pmu_enabled = cpuc->enabled;
894 cpuc->enabled = 0;
895
896 amd_brs_disable_all();
897
898 /* Drain BRS is in use (could be inactive) */
899 if (cpuc->lbr_users)
900 amd_brs_drain();
901
902 /* Process any counter overflows */
903 handled = x86_pmu_handle_irq(regs);
904
905 cpuc->enabled = pmu_enabled;
906 if (pmu_enabled)
907 amd_brs_enable_all();
908
909 return amd_pmu_adjust_nmi_window(handled);
910 }
911
912 /*
913 * AMD-specific callback invoked through perf_snapshot_branch_stack static
914 * call, defined in include/linux/perf_event.h. See its definition for API
915 * details. It's up to caller to provide enough space in *entries* to fit all
916 * LBR records, otherwise returned result will be truncated to *cnt* entries.
917 */
amd_pmu_v2_snapshot_branch_stack(struct perf_branch_entry * entries,unsigned int cnt)918 static int amd_pmu_v2_snapshot_branch_stack(struct perf_branch_entry *entries, unsigned int cnt)
919 {
920 struct cpu_hw_events *cpuc;
921 unsigned long flags;
922
923 /*
924 * The sequence of steps to freeze LBR should be completely inlined
925 * and contain no branches to minimize contamination of LBR snapshot
926 */
927 local_irq_save(flags);
928 amd_pmu_core_disable_all();
929 __amd_pmu_lbr_disable();
930
931 cpuc = this_cpu_ptr(&cpu_hw_events);
932
933 amd_pmu_lbr_read();
934 cnt = min(cnt, x86_pmu.lbr_nr);
935 memcpy(entries, cpuc->lbr_entries, sizeof(struct perf_branch_entry) * cnt);
936
937 amd_pmu_v2_enable_all(0);
938 local_irq_restore(flags);
939
940 return cnt;
941 }
942
amd_pmu_v2_handle_irq(struct pt_regs * regs)943 static int amd_pmu_v2_handle_irq(struct pt_regs *regs)
944 {
945 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
946 static atomic64_t status_warned = ATOMIC64_INIT(0);
947 u64 reserved, status, mask, new_bits, prev_bits;
948 struct perf_sample_data data;
949 struct hw_perf_event *hwc;
950 struct perf_event *event;
951 int handled = 0, idx;
952 bool pmu_enabled;
953
954 /*
955 * Save the PMU state as it needs to be restored when leaving the
956 * handler
957 */
958 pmu_enabled = cpuc->enabled;
959 cpuc->enabled = 0;
960
961 /* Stop counting but do not disable LBR */
962 amd_pmu_core_disable_all();
963
964 status = amd_pmu_get_global_status();
965
966 /* Check if any overflows are pending */
967 if (!status)
968 goto done;
969
970 /* Read branch records */
971 if (x86_pmu.lbr_nr) {
972 amd_pmu_lbr_read();
973 status &= ~GLOBAL_STATUS_LBRS_FROZEN;
974 }
975
976 reserved = status & ~amd_pmu_global_cntr_mask;
977 if (reserved)
978 pr_warn_once("Reserved PerfCntrGlobalStatus bits are set (0x%llx), please consider updating microcode\n",
979 reserved);
980
981 /* Clear any reserved bits set by buggy microcode */
982 status &= amd_pmu_global_cntr_mask;
983
984 for_each_set_bit(idx, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
985 if (!test_bit(idx, cpuc->active_mask))
986 continue;
987
988 event = cpuc->events[idx];
989 hwc = &event->hw;
990 x86_perf_event_update(event);
991 mask = BIT_ULL(idx);
992
993 if (!(status & mask))
994 continue;
995
996 /* Event overflow */
997 handled++;
998 status &= ~mask;
999 perf_sample_data_init(&data, 0, hwc->last_period);
1000
1001 if (!x86_perf_event_set_period(event))
1002 continue;
1003
1004 perf_sample_save_brstack(&data, event, &cpuc->lbr_stack, NULL);
1005
1006 if (perf_event_overflow(event, &data, regs))
1007 x86_pmu_stop(event, 0);
1008 }
1009
1010 /*
1011 * It should never be the case that some overflows are not handled as
1012 * the corresponding PMCs are expected to be inactive according to the
1013 * active_mask
1014 */
1015 if (status > 0) {
1016 prev_bits = atomic64_fetch_or(status, &status_warned);
1017 // A new bit was set for the very first time.
1018 new_bits = status & ~prev_bits;
1019 WARN(new_bits, "New overflows for inactive PMCs: %llx\n", new_bits);
1020 }
1021
1022 /* Clear overflow and freeze bits */
1023 amd_pmu_ack_global_status(~status);
1024
1025 /*
1026 * Unmasking the LVTPC is not required as the Mask (M) bit of the LVT
1027 * PMI entry is not set by the local APIC when a PMC overflow occurs
1028 */
1029 inc_irq_stat(apic_perf_irqs);
1030
1031 done:
1032 cpuc->enabled = pmu_enabled;
1033
1034 /* Resume counting only if PMU is active */
1035 if (pmu_enabled)
1036 amd_pmu_core_enable_all();
1037
1038 return amd_pmu_adjust_nmi_window(handled);
1039 }
1040
1041 static struct event_constraint *
amd_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)1042 amd_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
1043 struct perf_event *event)
1044 {
1045 /*
1046 * if not NB event or no NB, then no constraints
1047 */
1048 if (!(amd_has_nb(cpuc) && amd_is_nb_event(&event->hw)))
1049 return &unconstrained;
1050
1051 return __amd_get_nb_event_constraints(cpuc, event, NULL);
1052 }
1053
amd_put_event_constraints(struct cpu_hw_events * cpuc,struct perf_event * event)1054 static void amd_put_event_constraints(struct cpu_hw_events *cpuc,
1055 struct perf_event *event)
1056 {
1057 if (amd_has_nb(cpuc) && amd_is_nb_event(&event->hw))
1058 __amd_put_nb_event_constraints(cpuc, event);
1059 }
1060
1061 PMU_FORMAT_ATTR(event, "config:0-7,32-35");
1062 PMU_FORMAT_ATTR(umask, "config:8-15" );
1063 PMU_FORMAT_ATTR(edge, "config:18" );
1064 PMU_FORMAT_ATTR(inv, "config:23" );
1065 PMU_FORMAT_ATTR(cmask, "config:24-31" );
1066
1067 static struct attribute *amd_format_attr[] = {
1068 &format_attr_event.attr,
1069 &format_attr_umask.attr,
1070 &format_attr_edge.attr,
1071 &format_attr_inv.attr,
1072 &format_attr_cmask.attr,
1073 NULL,
1074 };
1075
1076 /* AMD Family 15h */
1077
1078 #define AMD_EVENT_TYPE_MASK 0x000000F0ULL
1079
1080 #define AMD_EVENT_FP 0x00000000ULL ... 0x00000010ULL
1081 #define AMD_EVENT_LS 0x00000020ULL ... 0x00000030ULL
1082 #define AMD_EVENT_DC 0x00000040ULL ... 0x00000050ULL
1083 #define AMD_EVENT_CU 0x00000060ULL ... 0x00000070ULL
1084 #define AMD_EVENT_IC_DE 0x00000080ULL ... 0x00000090ULL
1085 #define AMD_EVENT_EX_LS 0x000000C0ULL
1086 #define AMD_EVENT_DE 0x000000D0ULL
1087 #define AMD_EVENT_NB 0x000000E0ULL ... 0x000000F0ULL
1088
1089 /*
1090 * AMD family 15h event code/PMC mappings:
1091 *
1092 * type = event_code & 0x0F0:
1093 *
1094 * 0x000 FP PERF_CTL[5:3]
1095 * 0x010 FP PERF_CTL[5:3]
1096 * 0x020 LS PERF_CTL[5:0]
1097 * 0x030 LS PERF_CTL[5:0]
1098 * 0x040 DC PERF_CTL[5:0]
1099 * 0x050 DC PERF_CTL[5:0]
1100 * 0x060 CU PERF_CTL[2:0]
1101 * 0x070 CU PERF_CTL[2:0]
1102 * 0x080 IC/DE PERF_CTL[2:0]
1103 * 0x090 IC/DE PERF_CTL[2:0]
1104 * 0x0A0 ---
1105 * 0x0B0 ---
1106 * 0x0C0 EX/LS PERF_CTL[5:0]
1107 * 0x0D0 DE PERF_CTL[2:0]
1108 * 0x0E0 NB NB_PERF_CTL[3:0]
1109 * 0x0F0 NB NB_PERF_CTL[3:0]
1110 *
1111 * Exceptions:
1112 *
1113 * 0x000 FP PERF_CTL[3], PERF_CTL[5:3] (*)
1114 * 0x003 FP PERF_CTL[3]
1115 * 0x004 FP PERF_CTL[3], PERF_CTL[5:3] (*)
1116 * 0x00B FP PERF_CTL[3]
1117 * 0x00D FP PERF_CTL[3]
1118 * 0x023 DE PERF_CTL[2:0]
1119 * 0x02D LS PERF_CTL[3]
1120 * 0x02E LS PERF_CTL[3,0]
1121 * 0x031 LS PERF_CTL[2:0] (**)
1122 * 0x043 CU PERF_CTL[2:0]
1123 * 0x045 CU PERF_CTL[2:0]
1124 * 0x046 CU PERF_CTL[2:0]
1125 * 0x054 CU PERF_CTL[2:0]
1126 * 0x055 CU PERF_CTL[2:0]
1127 * 0x08F IC PERF_CTL[0]
1128 * 0x187 DE PERF_CTL[0]
1129 * 0x188 DE PERF_CTL[0]
1130 * 0x0DB EX PERF_CTL[5:0]
1131 * 0x0DC LS PERF_CTL[5:0]
1132 * 0x0DD LS PERF_CTL[5:0]
1133 * 0x0DE LS PERF_CTL[5:0]
1134 * 0x0DF LS PERF_CTL[5:0]
1135 * 0x1C0 EX PERF_CTL[5:3]
1136 * 0x1D6 EX PERF_CTL[5:0]
1137 * 0x1D8 EX PERF_CTL[5:0]
1138 *
1139 * (*) depending on the umask all FPU counters may be used
1140 * (**) only one unitmask enabled at a time
1141 */
1142
1143 static struct event_constraint amd_f15_PMC0 = EVENT_CONSTRAINT(0, 0x01, 0);
1144 static struct event_constraint amd_f15_PMC20 = EVENT_CONSTRAINT(0, 0x07, 0);
1145 static struct event_constraint amd_f15_PMC3 = EVENT_CONSTRAINT(0, 0x08, 0);
1146 static struct event_constraint amd_f15_PMC30 = EVENT_CONSTRAINT_OVERLAP(0, 0x09, 0);
1147 static struct event_constraint amd_f15_PMC50 = EVENT_CONSTRAINT(0, 0x3F, 0);
1148 static struct event_constraint amd_f15_PMC53 = EVENT_CONSTRAINT(0, 0x38, 0);
1149
1150 static struct event_constraint *
amd_get_event_constraints_f15h(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)1151 amd_get_event_constraints_f15h(struct cpu_hw_events *cpuc, int idx,
1152 struct perf_event *event)
1153 {
1154 struct hw_perf_event *hwc = &event->hw;
1155 unsigned int event_code = amd_get_event_code(hwc);
1156
1157 switch (event_code & AMD_EVENT_TYPE_MASK) {
1158 case AMD_EVENT_FP:
1159 switch (event_code) {
1160 case 0x000:
1161 if (!(hwc->config & 0x0000F000ULL))
1162 break;
1163 if (!(hwc->config & 0x00000F00ULL))
1164 break;
1165 return &amd_f15_PMC3;
1166 case 0x004:
1167 if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
1168 break;
1169 return &amd_f15_PMC3;
1170 case 0x003:
1171 case 0x00B:
1172 case 0x00D:
1173 return &amd_f15_PMC3;
1174 }
1175 return &amd_f15_PMC53;
1176 case AMD_EVENT_LS:
1177 case AMD_EVENT_DC:
1178 case AMD_EVENT_EX_LS:
1179 switch (event_code) {
1180 case 0x023:
1181 case 0x043:
1182 case 0x045:
1183 case 0x046:
1184 case 0x054:
1185 case 0x055:
1186 return &amd_f15_PMC20;
1187 case 0x02D:
1188 return &amd_f15_PMC3;
1189 case 0x02E:
1190 return &amd_f15_PMC30;
1191 case 0x031:
1192 if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
1193 return &amd_f15_PMC20;
1194 return &emptyconstraint;
1195 case 0x1C0:
1196 return &amd_f15_PMC53;
1197 default:
1198 return &amd_f15_PMC50;
1199 }
1200 case AMD_EVENT_CU:
1201 case AMD_EVENT_IC_DE:
1202 case AMD_EVENT_DE:
1203 switch (event_code) {
1204 case 0x08F:
1205 case 0x187:
1206 case 0x188:
1207 return &amd_f15_PMC0;
1208 case 0x0DB ... 0x0DF:
1209 case 0x1D6:
1210 case 0x1D8:
1211 return &amd_f15_PMC50;
1212 default:
1213 return &amd_f15_PMC20;
1214 }
1215 case AMD_EVENT_NB:
1216 /* moved to uncore.c */
1217 return &emptyconstraint;
1218 default:
1219 return &emptyconstraint;
1220 }
1221 }
1222
1223 static struct event_constraint pair_constraint;
1224
1225 static struct event_constraint *
amd_get_event_constraints_f17h(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)1226 amd_get_event_constraints_f17h(struct cpu_hw_events *cpuc, int idx,
1227 struct perf_event *event)
1228 {
1229 struct hw_perf_event *hwc = &event->hw;
1230
1231 if (amd_is_pair_event_code(hwc))
1232 return &pair_constraint;
1233
1234 return &unconstrained;
1235 }
1236
amd_put_event_constraints_f17h(struct cpu_hw_events * cpuc,struct perf_event * event)1237 static void amd_put_event_constraints_f17h(struct cpu_hw_events *cpuc,
1238 struct perf_event *event)
1239 {
1240 struct hw_perf_event *hwc = &event->hw;
1241
1242 if (is_counter_pair(hwc))
1243 --cpuc->n_pair;
1244 }
1245
1246 /*
1247 * Because of the way BRS operates with an inactive and active phases, and
1248 * the link to one counter, it is not possible to have two events using BRS
1249 * scheduled at the same time. There would be an issue with enforcing the
1250 * period of each one and given that the BRS saturates, it would not be possible
1251 * to guarantee correlated content for all events. Therefore, in situations
1252 * where multiple events want to use BRS, the kernel enforces mutual exclusion.
1253 * Exclusion is enforced by choosing only one counter for events using BRS.
1254 * The event scheduling logic will then automatically multiplex the
1255 * events and ensure that at most one event is actively using BRS.
1256 *
1257 * The BRS counter could be any counter, but there is no constraint on Fam19h,
1258 * therefore all counters are equal and thus we pick the first one: PMC0
1259 */
1260 static struct event_constraint amd_fam19h_brs_cntr0_constraint =
1261 EVENT_CONSTRAINT(0, 0x1, AMD64_RAW_EVENT_MASK);
1262
1263 static struct event_constraint amd_fam19h_brs_pair_cntr0_constraint =
1264 __EVENT_CONSTRAINT(0, 0x1, AMD64_RAW_EVENT_MASK, 1, 0, PERF_X86_EVENT_PAIR);
1265
1266 static struct event_constraint *
amd_get_event_constraints_f19h(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)1267 amd_get_event_constraints_f19h(struct cpu_hw_events *cpuc, int idx,
1268 struct perf_event *event)
1269 {
1270 struct hw_perf_event *hwc = &event->hw;
1271 bool has_brs = has_amd_brs(hwc);
1272
1273 /*
1274 * In case BRS is used with an event requiring a counter pair,
1275 * the kernel allows it but only on counter 0 & 1 to enforce
1276 * multiplexing requiring to protect BRS in case of multiple
1277 * BRS users
1278 */
1279 if (amd_is_pair_event_code(hwc)) {
1280 return has_brs ? &amd_fam19h_brs_pair_cntr0_constraint
1281 : &pair_constraint;
1282 }
1283
1284 if (has_brs)
1285 return &amd_fam19h_brs_cntr0_constraint;
1286
1287 return &unconstrained;
1288 }
1289
1290
amd_event_sysfs_show(char * page,u64 config)1291 static ssize_t amd_event_sysfs_show(char *page, u64 config)
1292 {
1293 u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT) |
1294 (config & AMD64_EVENTSEL_EVENT) >> 24;
1295
1296 return x86_event_sysfs_show(page, config, event);
1297 }
1298
amd_pmu_limit_period(struct perf_event * event,s64 * left)1299 static void amd_pmu_limit_period(struct perf_event *event, s64 *left)
1300 {
1301 /*
1302 * Decrease period by the depth of the BRS feature to get the last N
1303 * taken branches and approximate the desired period
1304 */
1305 if (has_branch_stack(event) && *left > x86_pmu.lbr_nr)
1306 *left -= x86_pmu.lbr_nr;
1307 }
1308
1309 static __initconst const struct x86_pmu amd_pmu = {
1310 .name = "AMD",
1311 .handle_irq = amd_pmu_handle_irq,
1312 .disable_all = amd_pmu_disable_all,
1313 .enable_all = amd_pmu_enable_all,
1314 .enable = amd_pmu_enable_event,
1315 .disable = amd_pmu_disable_event,
1316 .hw_config = amd_pmu_hw_config,
1317 .schedule_events = x86_schedule_events,
1318 .eventsel = MSR_K7_EVNTSEL0,
1319 .perfctr = MSR_K7_PERFCTR0,
1320 .addr_offset = amd_pmu_addr_offset,
1321 .event_map = amd_pmu_event_map,
1322 .max_events = ARRAY_SIZE(amd_perfmon_event_map),
1323 .cntr_mask64 = GENMASK_ULL(AMD64_NUM_COUNTERS - 1, 0),
1324 .add = amd_pmu_add_event,
1325 .del = amd_pmu_del_event,
1326 .cntval_bits = 48,
1327 .cntval_mask = (1ULL << 48) - 1,
1328 .apic = 1,
1329 /* use highest bit to detect overflow */
1330 .max_period = (1ULL << 47) - 1,
1331 .get_event_constraints = amd_get_event_constraints,
1332 .put_event_constraints = amd_put_event_constraints,
1333
1334 .format_attrs = amd_format_attr,
1335 .events_sysfs_show = amd_event_sysfs_show,
1336
1337 .cpu_prepare = amd_pmu_cpu_prepare,
1338 .cpu_starting = amd_pmu_cpu_starting,
1339 .cpu_dead = amd_pmu_cpu_dead,
1340
1341 .amd_nb_constraints = 1,
1342 };
1343
branches_show(struct device * cdev,struct device_attribute * attr,char * buf)1344 static ssize_t branches_show(struct device *cdev,
1345 struct device_attribute *attr,
1346 char *buf)
1347 {
1348 return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu.lbr_nr);
1349 }
1350
1351 static DEVICE_ATTR_RO(branches);
1352
1353 static struct attribute *amd_pmu_branches_attrs[] = {
1354 &dev_attr_branches.attr,
1355 NULL,
1356 };
1357
1358 static umode_t
amd_branches_is_visible(struct kobject * kobj,struct attribute * attr,int i)1359 amd_branches_is_visible(struct kobject *kobj, struct attribute *attr, int i)
1360 {
1361 return x86_pmu.lbr_nr ? attr->mode : 0;
1362 }
1363
1364 static struct attribute_group group_caps_amd_branches = {
1365 .name = "caps",
1366 .attrs = amd_pmu_branches_attrs,
1367 .is_visible = amd_branches_is_visible,
1368 };
1369
1370 #ifdef CONFIG_PERF_EVENTS_AMD_BRS
1371
1372 EVENT_ATTR_STR(branch-brs, amd_branch_brs,
1373 "event=" __stringify(AMD_FAM19H_BRS_EVENT)"\n");
1374
1375 static struct attribute *amd_brs_events_attrs[] = {
1376 EVENT_PTR(amd_branch_brs),
1377 NULL,
1378 };
1379
1380 static umode_t
amd_brs_is_visible(struct kobject * kobj,struct attribute * attr,int i)1381 amd_brs_is_visible(struct kobject *kobj, struct attribute *attr, int i)
1382 {
1383 return static_cpu_has(X86_FEATURE_BRS) && x86_pmu.lbr_nr ?
1384 attr->mode : 0;
1385 }
1386
1387 static struct attribute_group group_events_amd_brs = {
1388 .name = "events",
1389 .attrs = amd_brs_events_attrs,
1390 .is_visible = amd_brs_is_visible,
1391 };
1392
1393 #endif /* CONFIG_PERF_EVENTS_AMD_BRS */
1394
1395 static const struct attribute_group *amd_attr_update[] = {
1396 &group_caps_amd_branches,
1397 #ifdef CONFIG_PERF_EVENTS_AMD_BRS
1398 &group_events_amd_brs,
1399 #endif
1400 NULL,
1401 };
1402
amd_core_pmu_init(void)1403 static int __init amd_core_pmu_init(void)
1404 {
1405 union cpuid_0x80000022_ebx ebx;
1406 u64 even_ctr_mask = 0ULL;
1407 int i;
1408
1409 if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
1410 return 0;
1411
1412 /* Avoid calculating the value each time in the NMI handler */
1413 perf_nmi_window = msecs_to_jiffies(100);
1414
1415 /*
1416 * If core performance counter extensions exists, we must use
1417 * MSR_F15H_PERF_CTL/MSR_F15H_PERF_CTR msrs. See also
1418 * amd_pmu_addr_offset().
1419 */
1420 x86_pmu.eventsel = MSR_F15H_PERF_CTL;
1421 x86_pmu.perfctr = MSR_F15H_PERF_CTR;
1422 x86_pmu.cntr_mask64 = GENMASK_ULL(AMD64_NUM_COUNTERS_CORE - 1, 0);
1423
1424 /* Check for Performance Monitoring v2 support */
1425 if (boot_cpu_has(X86_FEATURE_PERFMON_V2)) {
1426 ebx.full = cpuid_ebx(EXT_PERFMON_DEBUG_FEATURES);
1427
1428 /* Update PMU version for later usage */
1429 x86_pmu.version = 2;
1430
1431 /* Find the number of available Core PMCs */
1432 x86_pmu.cntr_mask64 = GENMASK_ULL(ebx.split.num_core_pmc - 1, 0);
1433
1434 amd_pmu_global_cntr_mask = x86_pmu.cntr_mask64;
1435
1436 /* Update PMC handling functions */
1437 x86_pmu.enable_all = amd_pmu_v2_enable_all;
1438 x86_pmu.disable_all = amd_pmu_v2_disable_all;
1439 x86_pmu.enable = amd_pmu_v2_enable_event;
1440 x86_pmu.handle_irq = amd_pmu_v2_handle_irq;
1441 static_call_update(amd_pmu_test_overflow, amd_pmu_test_overflow_status);
1442 }
1443
1444 /*
1445 * AMD Core perfctr has separate MSRs for the NB events, see
1446 * the amd/uncore.c driver.
1447 */
1448 x86_pmu.amd_nb_constraints = 0;
1449
1450 if (boot_cpu_data.x86 == 0x15) {
1451 pr_cont("Fam15h ");
1452 x86_pmu.get_event_constraints = amd_get_event_constraints_f15h;
1453 }
1454 if (boot_cpu_data.x86 >= 0x17) {
1455 pr_cont("Fam17h+ ");
1456 /*
1457 * Family 17h and compatibles have constraints for Large
1458 * Increment per Cycle events: they may only be assigned an
1459 * even numbered counter that has a consecutive adjacent odd
1460 * numbered counter following it.
1461 */
1462 for (i = 0; i < x86_pmu_max_num_counters(NULL) - 1; i += 2)
1463 even_ctr_mask |= BIT_ULL(i);
1464
1465 pair_constraint = (struct event_constraint)
1466 __EVENT_CONSTRAINT(0, even_ctr_mask, 0,
1467 x86_pmu_max_num_counters(NULL) / 2, 0,
1468 PERF_X86_EVENT_PAIR);
1469
1470 x86_pmu.get_event_constraints = amd_get_event_constraints_f17h;
1471 x86_pmu.put_event_constraints = amd_put_event_constraints_f17h;
1472 x86_pmu.perf_ctr_pair_en = AMD_MERGE_EVENT_ENABLE;
1473 x86_pmu.flags |= PMU_FL_PAIR;
1474 }
1475
1476 /* LBR and BRS are mutually exclusive features */
1477 if (!amd_pmu_lbr_init()) {
1478 /* LBR requires flushing on context switch */
1479 x86_pmu.sched_task = amd_pmu_lbr_sched_task;
1480 static_call_update(amd_pmu_branch_hw_config, amd_pmu_lbr_hw_config);
1481 static_call_update(amd_pmu_branch_reset, amd_pmu_lbr_reset);
1482 static_call_update(amd_pmu_branch_add, amd_pmu_lbr_add);
1483 static_call_update(amd_pmu_branch_del, amd_pmu_lbr_del);
1484
1485 /* Only support branch_stack snapshot on perfmon v2 */
1486 if (x86_pmu.handle_irq == amd_pmu_v2_handle_irq)
1487 static_call_update(perf_snapshot_branch_stack, amd_pmu_v2_snapshot_branch_stack);
1488 } else if (!amd_brs_init()) {
1489 /*
1490 * BRS requires special event constraints and flushing on ctxsw.
1491 */
1492 x86_pmu.get_event_constraints = amd_get_event_constraints_f19h;
1493 x86_pmu.sched_task = amd_pmu_brs_sched_task;
1494 x86_pmu.limit_period = amd_pmu_limit_period;
1495
1496 static_call_update(amd_pmu_branch_hw_config, amd_brs_hw_config);
1497 static_call_update(amd_pmu_branch_reset, amd_brs_reset);
1498 static_call_update(amd_pmu_branch_add, amd_pmu_brs_add);
1499 static_call_update(amd_pmu_branch_del, amd_pmu_brs_del);
1500
1501 /*
1502 * put_event_constraints callback same as Fam17h, set above
1503 */
1504
1505 /* branch sampling must be stopped when entering low power */
1506 amd_brs_lopwr_init();
1507 }
1508
1509 x86_pmu.attr_update = amd_attr_update;
1510
1511 pr_cont("core perfctr, ");
1512 return 0;
1513 }
1514
amd_pmu_init(void)1515 __init int amd_pmu_init(void)
1516 {
1517 int ret;
1518
1519 /* Performance-monitoring supported from K7 and later: */
1520 if (boot_cpu_data.x86 < 6)
1521 return -ENODEV;
1522
1523 x86_pmu = amd_pmu;
1524
1525 ret = amd_core_pmu_init();
1526 if (ret)
1527 return ret;
1528
1529 if (num_possible_cpus() == 1) {
1530 /*
1531 * No point in allocating data structures to serialize
1532 * against other CPUs, when there is only the one CPU.
1533 */
1534 x86_pmu.amd_nb_constraints = 0;
1535 }
1536
1537 if (boot_cpu_data.x86 >= 0x17)
1538 memcpy(hw_cache_event_ids, amd_hw_cache_event_ids_f17h, sizeof(hw_cache_event_ids));
1539 else
1540 memcpy(hw_cache_event_ids, amd_hw_cache_event_ids, sizeof(hw_cache_event_ids));
1541
1542 return 0;
1543 }
1544
amd_pmu_reload_virt(void)1545 static inline void amd_pmu_reload_virt(void)
1546 {
1547 if (x86_pmu.version >= 2) {
1548 /*
1549 * Clear global enable bits, reprogram the PERF_CTL
1550 * registers with updated perf_ctr_virt_mask and then
1551 * set global enable bits once again
1552 */
1553 amd_pmu_v2_disable_all();
1554 amd_pmu_enable_all(0);
1555 amd_pmu_v2_enable_all(0);
1556 return;
1557 }
1558
1559 amd_pmu_disable_all();
1560 amd_pmu_enable_all(0);
1561 }
1562
amd_pmu_enable_virt(void)1563 void amd_pmu_enable_virt(void)
1564 {
1565 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1566
1567 cpuc->perf_ctr_virt_mask = 0;
1568
1569 /* Reload all events */
1570 amd_pmu_reload_virt();
1571 }
1572 EXPORT_SYMBOL_GPL(amd_pmu_enable_virt);
1573
amd_pmu_disable_virt(void)1574 void amd_pmu_disable_virt(void)
1575 {
1576 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1577
1578 /*
1579 * We only mask out the Host-only bit so that host-only counting works
1580 * when SVM is disabled. If someone sets up a guest-only counter when
1581 * SVM is disabled the Guest-only bits still gets set and the counter
1582 * will not count anything.
1583 */
1584 cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
1585
1586 /* Reload all events */
1587 amd_pmu_reload_virt();
1588 }
1589 EXPORT_SYMBOL_GPL(amd_pmu_disable_virt);
1590