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/msr.h>
13 #include <asm/nmi.h>
14
15 #include "../perf_event.h"
16
17 static DEFINE_PER_CPU(unsigned long, perf_nmi_tstamp);
18 static unsigned long perf_nmi_window;
19
20 /* AMD Event 0xFFF: Merge. Used with Large Increment per Cycle events */
21 #define AMD_MERGE_EVENT ((0xFULL << 32) | 0xFFULL)
22 #define AMD_MERGE_EVENT_ENABLE (AMD_MERGE_EVENT | ARCH_PERFMON_EVENTSEL_ENABLE)
23
24 /* PMC Enable and Overflow bits for PerfCntrGlobal* registers */
25 static u64 amd_pmu_global_cntr_mask __read_mostly;
26
27 static __initconst const u64 amd_hw_cache_event_ids
28 [PERF_COUNT_HW_CACHE_MAX]
29 [PERF_COUNT_HW_CACHE_OP_MAX]
30 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
31 {
32 [ C(L1D) ] = {
33 [ C(OP_READ) ] = {
34 [ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses */
35 [ C(RESULT_MISS) ] = 0x0141, /* Data Cache Misses */
36 },
37 [ C(OP_WRITE) ] = {
38 [ C(RESULT_ACCESS) ] = 0,
39 [ C(RESULT_MISS) ] = 0,
40 },
41 [ C(OP_PREFETCH) ] = {
42 [ C(RESULT_ACCESS) ] = 0x0267, /* Data Prefetcher :attempts */
43 [ C(RESULT_MISS) ] = 0x0167, /* Data Prefetcher :cancelled */
44 },
45 },
46 [ C(L1I ) ] = {
47 [ C(OP_READ) ] = {
48 [ C(RESULT_ACCESS) ] = 0x0080, /* Instruction cache fetches */
49 [ C(RESULT_MISS) ] = 0x0081, /* Instruction cache misses */
50 },
51 [ C(OP_WRITE) ] = {
52 [ C(RESULT_ACCESS) ] = -1,
53 [ C(RESULT_MISS) ] = -1,
54 },
55 [ C(OP_PREFETCH) ] = {
56 [ C(RESULT_ACCESS) ] = 0x014B, /* Prefetch Instructions :Load */
57 [ C(RESULT_MISS) ] = 0,
58 },
59 },
60 [ C(LL ) ] = {
61 [ C(OP_READ) ] = {
62 [ C(RESULT_ACCESS) ] = 0x037D, /* Requests to L2 Cache :IC+DC */
63 [ C(RESULT_MISS) ] = 0x037E, /* L2 Cache Misses : IC+DC */
64 },
65 [ C(OP_WRITE) ] = {
66 [ C(RESULT_ACCESS) ] = 0x017F, /* L2 Fill/Writeback */
67 [ C(RESULT_MISS) ] = 0,
68 },
69 [ C(OP_PREFETCH) ] = {
70 [ C(RESULT_ACCESS) ] = 0,
71 [ C(RESULT_MISS) ] = 0,
72 },
73 },
74 [ C(DTLB) ] = {
75 [ C(OP_READ) ] = {
76 [ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses */
77 [ C(RESULT_MISS) ] = 0x0746, /* L1_DTLB_AND_L2_DLTB_MISS.ALL */
78 },
79 [ C(OP_WRITE) ] = {
80 [ C(RESULT_ACCESS) ] = 0,
81 [ C(RESULT_MISS) ] = 0,
82 },
83 [ C(OP_PREFETCH) ] = {
84 [ C(RESULT_ACCESS) ] = 0,
85 [ C(RESULT_MISS) ] = 0,
86 },
87 },
88 [ C(ITLB) ] = {
89 [ C(OP_READ) ] = {
90 [ C(RESULT_ACCESS) ] = 0x0080, /* Instruction fecthes */
91 [ C(RESULT_MISS) ] = 0x0385, /* L1_ITLB_AND_L2_ITLB_MISS.ALL */
92 },
93 [ C(OP_WRITE) ] = {
94 [ C(RESULT_ACCESS) ] = -1,
95 [ C(RESULT_MISS) ] = -1,
96 },
97 [ C(OP_PREFETCH) ] = {
98 [ C(RESULT_ACCESS) ] = -1,
99 [ C(RESULT_MISS) ] = -1,
100 },
101 },
102 [ C(BPU ) ] = {
103 [ C(OP_READ) ] = {
104 [ C(RESULT_ACCESS) ] = 0x00c2, /* Retired Branch Instr. */
105 [ C(RESULT_MISS) ] = 0x00c3, /* Retired Mispredicted BI */
106 },
107 [ C(OP_WRITE) ] = {
108 [ C(RESULT_ACCESS) ] = -1,
109 [ C(RESULT_MISS) ] = -1,
110 },
111 [ C(OP_PREFETCH) ] = {
112 [ C(RESULT_ACCESS) ] = -1,
113 [ C(RESULT_MISS) ] = -1,
114 },
115 },
116 [ C(NODE) ] = {
117 [ C(OP_READ) ] = {
118 [ C(RESULT_ACCESS) ] = 0xb8e9, /* CPU Request to Memory, l+r */
119 [ C(RESULT_MISS) ] = 0x98e9, /* CPU Request to Memory, r */
120 },
121 [ C(OP_WRITE) ] = {
122 [ C(RESULT_ACCESS) ] = -1,
123 [ C(RESULT_MISS) ] = -1,
124 },
125 [ C(OP_PREFETCH) ] = {
126 [ C(RESULT_ACCESS) ] = -1,
127 [ C(RESULT_MISS) ] = -1,
128 },
129 },
130 };
131
132 static __initconst const u64 amd_hw_cache_event_ids_f17h
133 [PERF_COUNT_HW_CACHE_MAX]
134 [PERF_COUNT_HW_CACHE_OP_MAX]
135 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
136 [C(L1D)] = {
137 [C(OP_READ)] = {
138 [C(RESULT_ACCESS)] = 0x0040, /* Data Cache Accesses */
139 [C(RESULT_MISS)] = 0xc860, /* L2$ access from DC Miss */
140 },
141 [C(OP_WRITE)] = {
142 [C(RESULT_ACCESS)] = 0,
143 [C(RESULT_MISS)] = 0,
144 },
145 [C(OP_PREFETCH)] = {
146 [C(RESULT_ACCESS)] = 0xff5a, /* h/w prefetch DC Fills */
147 [C(RESULT_MISS)] = 0,
148 },
149 },
150 [C(L1I)] = {
151 [C(OP_READ)] = {
152 [C(RESULT_ACCESS)] = 0x0080, /* Instruction cache fetches */
153 [C(RESULT_MISS)] = 0x0081, /* Instruction cache misses */
154 },
155 [C(OP_WRITE)] = {
156 [C(RESULT_ACCESS)] = -1,
157 [C(RESULT_MISS)] = -1,
158 },
159 [C(OP_PREFETCH)] = {
160 [C(RESULT_ACCESS)] = 0,
161 [C(RESULT_MISS)] = 0,
162 },
163 },
164 [C(LL)] = {
165 [C(OP_READ)] = {
166 [C(RESULT_ACCESS)] = 0,
167 [C(RESULT_MISS)] = 0,
168 },
169 [C(OP_WRITE)] = {
170 [C(RESULT_ACCESS)] = 0,
171 [C(RESULT_MISS)] = 0,
172 },
173 [C(OP_PREFETCH)] = {
174 [C(RESULT_ACCESS)] = 0,
175 [C(RESULT_MISS)] = 0,
176 },
177 },
178 [C(DTLB)] = {
179 [C(OP_READ)] = {
180 [C(RESULT_ACCESS)] = 0xff45, /* All L2 DTLB accesses */
181 [C(RESULT_MISS)] = 0xf045, /* L2 DTLB misses (PT walks) */
182 },
183 [C(OP_WRITE)] = {
184 [C(RESULT_ACCESS)] = 0,
185 [C(RESULT_MISS)] = 0,
186 },
187 [C(OP_PREFETCH)] = {
188 [C(RESULT_ACCESS)] = 0,
189 [C(RESULT_MISS)] = 0,
190 },
191 },
192 [C(ITLB)] = {
193 [C(OP_READ)] = {
194 [C(RESULT_ACCESS)] = 0x0084, /* L1 ITLB misses, L2 ITLB hits */
195 [C(RESULT_MISS)] = 0xff85, /* L1 ITLB misses, L2 misses */
196 },
197 [C(OP_WRITE)] = {
198 [C(RESULT_ACCESS)] = -1,
199 [C(RESULT_MISS)] = -1,
200 },
201 [C(OP_PREFETCH)] = {
202 [C(RESULT_ACCESS)] = -1,
203 [C(RESULT_MISS)] = -1,
204 },
205 },
206 [C(BPU)] = {
207 [C(OP_READ)] = {
208 [C(RESULT_ACCESS)] = 0x00c2, /* Retired Branch Instr. */
209 [C(RESULT_MISS)] = 0x00c3, /* Retired Mispredicted BI */
210 },
211 [C(OP_WRITE)] = {
212 [C(RESULT_ACCESS)] = -1,
213 [C(RESULT_MISS)] = -1,
214 },
215 [C(OP_PREFETCH)] = {
216 [C(RESULT_ACCESS)] = -1,
217 [C(RESULT_MISS)] = -1,
218 },
219 },
220 [C(NODE)] = {
221 [C(OP_READ)] = {
222 [C(RESULT_ACCESS)] = 0,
223 [C(RESULT_MISS)] = 0,
224 },
225 [C(OP_WRITE)] = {
226 [C(RESULT_ACCESS)] = -1,
227 [C(RESULT_MISS)] = -1,
228 },
229 [C(OP_PREFETCH)] = {
230 [C(RESULT_ACCESS)] = -1,
231 [C(RESULT_MISS)] = -1,
232 },
233 },
234 };
235
236 /*
237 * AMD Performance Monitor K7 and later, up to and including Family 16h:
238 */
239 static const u64 amd_perfmon_event_map[PERF_COUNT_HW_MAX] =
240 {
241 [PERF_COUNT_HW_CPU_CYCLES] = 0x0076,
242 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
243 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x077d,
244 [PERF_COUNT_HW_CACHE_MISSES] = 0x077e,
245 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c2,
246 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c3,
247 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x00d0, /* "Decoder empty" event */
248 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x00d1, /* "Dispatch stalls" event */
249 };
250
251 /*
252 * AMD Performance Monitor Family 17h and later:
253 */
254 static const u64 amd_zen1_perfmon_event_map[PERF_COUNT_HW_MAX] =
255 {
256 [PERF_COUNT_HW_CPU_CYCLES] = 0x0076,
257 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
258 [PERF_COUNT_HW_CACHE_REFERENCES] = 0xff60,
259 [PERF_COUNT_HW_CACHE_MISSES] = 0x0964,
260 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c2,
261 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c3,
262 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x0287,
263 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x0187,
264 };
265
266 static const u64 amd_zen2_perfmon_event_map[PERF_COUNT_HW_MAX] =
267 {
268 [PERF_COUNT_HW_CPU_CYCLES] = 0x0076,
269 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
270 [PERF_COUNT_HW_CACHE_REFERENCES] = 0xff60,
271 [PERF_COUNT_HW_CACHE_MISSES] = 0x0964,
272 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c2,
273 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c3,
274 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x00a9,
275 };
276
277 static const u64 amd_zen4_perfmon_event_map[PERF_COUNT_HW_MAX] =
278 {
279 [PERF_COUNT_HW_CPU_CYCLES] = 0x0076,
280 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
281 [PERF_COUNT_HW_CACHE_REFERENCES] = 0xff60,
282 [PERF_COUNT_HW_CACHE_MISSES] = 0x0964,
283 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c2,
284 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c3,
285 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x00a9,
286 [PERF_COUNT_HW_REF_CPU_CYCLES] = 0x100000120,
287 };
288
amd_pmu_event_map(int hw_event)289 static u64 amd_pmu_event_map(int hw_event)
290 {
291 if (cpu_feature_enabled(X86_FEATURE_ZEN4) || boot_cpu_data.x86 >= 0x1a)
292 return amd_zen4_perfmon_event_map[hw_event];
293
294 if (cpu_feature_enabled(X86_FEATURE_ZEN2) || boot_cpu_data.x86 >= 0x19)
295 return amd_zen2_perfmon_event_map[hw_event];
296
297 if (cpu_feature_enabled(X86_FEATURE_ZEN1))
298 return amd_zen1_perfmon_event_map[hw_event];
299
300 return amd_perfmon_event_map[hw_event];
301 }
302
303 /*
304 * Previously calculated offsets
305 */
306 static unsigned int event_offsets[X86_PMC_IDX_MAX] __read_mostly;
307 static unsigned int count_offsets[X86_PMC_IDX_MAX] __read_mostly;
308
309 /*
310 * Legacy CPUs:
311 * 4 counters starting at 0xc0010000 each offset by 1
312 *
313 * CPUs with core performance counter extensions:
314 * 6 counters starting at 0xc0010200 each offset by 2
315 */
amd_pmu_addr_offset(int index,bool eventsel)316 static inline int amd_pmu_addr_offset(int index, bool eventsel)
317 {
318 int offset;
319
320 if (!index)
321 return index;
322
323 if (eventsel)
324 offset = event_offsets[index];
325 else
326 offset = count_offsets[index];
327
328 if (offset)
329 return offset;
330
331 if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
332 offset = index;
333 else
334 offset = index << 1;
335
336 if (eventsel)
337 event_offsets[index] = offset;
338 else
339 count_offsets[index] = offset;
340
341 return offset;
342 }
343
344 /*
345 * AMD64 events are detected based on their event codes.
346 */
amd_get_event_code(struct hw_perf_event * hwc)347 static inline unsigned int amd_get_event_code(struct hw_perf_event *hwc)
348 {
349 return ((hwc->config >> 24) & 0x0f00) | (hwc->config & 0x00ff);
350 }
351
amd_is_pair_event_code(struct hw_perf_event * hwc)352 static inline bool amd_is_pair_event_code(struct hw_perf_event *hwc)
353 {
354 if (!(x86_pmu.flags & PMU_FL_PAIR))
355 return false;
356
357 switch (amd_get_event_code(hwc)) {
358 case 0x003: return true; /* Retired SSE/AVX FLOPs */
359 default: return false;
360 }
361 }
362
363 DEFINE_STATIC_CALL_RET0(amd_pmu_branch_hw_config, *x86_pmu.hw_config);
364
amd_core_hw_config(struct perf_event * event)365 static int amd_core_hw_config(struct perf_event *event)
366 {
367 if (event->attr.exclude_host && event->attr.exclude_guest)
368 /*
369 * When HO == GO == 1 the hardware treats that as GO == HO == 0
370 * and will count in both modes. We don't want to count in that
371 * case so we emulate no-counting by setting US = OS = 0.
372 */
373 event->hw.config &= ~(ARCH_PERFMON_EVENTSEL_USR |
374 ARCH_PERFMON_EVENTSEL_OS);
375 else if (event->attr.exclude_host)
376 event->hw.config |= AMD64_EVENTSEL_GUESTONLY;
377 else if (event->attr.exclude_guest)
378 event->hw.config |= AMD64_EVENTSEL_HOSTONLY;
379
380 if ((x86_pmu.flags & PMU_FL_PAIR) && amd_is_pair_event_code(&event->hw))
381 event->hw.flags |= PERF_X86_EVENT_PAIR;
382
383 if (has_branch_stack(event))
384 return static_call(amd_pmu_branch_hw_config)(event);
385
386 return 0;
387 }
388
amd_is_nb_event(struct hw_perf_event * hwc)389 static inline int amd_is_nb_event(struct hw_perf_event *hwc)
390 {
391 return (hwc->config & 0xe0) == 0xe0;
392 }
393
amd_has_nb(struct cpu_hw_events * cpuc)394 static inline int amd_has_nb(struct cpu_hw_events *cpuc)
395 {
396 struct amd_nb *nb = cpuc->amd_nb;
397
398 return nb && nb->nb_id != -1;
399 }
400
amd_pmu_hw_config(struct perf_event * event)401 static int amd_pmu_hw_config(struct perf_event *event)
402 {
403 int ret;
404
405 /* pass precise event sampling to ibs: */
406 if (event->attr.precise_ip && get_ibs_caps())
407 return forward_event_to_ibs(event);
408
409 if (has_branch_stack(event) && !x86_pmu.lbr_nr)
410 return -EOPNOTSUPP;
411
412 ret = x86_pmu_hw_config(event);
413 if (ret)
414 return ret;
415
416 if (event->attr.type == PERF_TYPE_RAW)
417 event->hw.config |= event->attr.config & AMD64_RAW_EVENT_MASK;
418
419 return amd_core_hw_config(event);
420 }
421
__amd_put_nb_event_constraints(struct cpu_hw_events * cpuc,struct perf_event * event)422 static void __amd_put_nb_event_constraints(struct cpu_hw_events *cpuc,
423 struct perf_event *event)
424 {
425 struct amd_nb *nb = cpuc->amd_nb;
426 int i;
427
428 /*
429 * need to scan whole list because event may not have
430 * been assigned during scheduling
431 *
432 * no race condition possible because event can only
433 * be removed on one CPU at a time AND PMU is disabled
434 * when we come here
435 */
436 for_each_set_bit(i, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
437 struct perf_event *tmp = event;
438
439 if (try_cmpxchg(nb->owners + i, &tmp, NULL))
440 break;
441 }
442 }
443
444 /*
445 * AMD64 NorthBridge events need special treatment because
446 * counter access needs to be synchronized across all cores
447 * of a package. Refer to BKDG section 3.12
448 *
449 * NB events are events measuring L3 cache, Hypertransport
450 * traffic. They are identified by an event code >= 0xe00.
451 * They measure events on the NorthBride which is shared
452 * by all cores on a package. NB events are counted on a
453 * shared set of counters. When a NB event is programmed
454 * in a counter, the data actually comes from a shared
455 * counter. Thus, access to those counters needs to be
456 * synchronized.
457 *
458 * We implement the synchronization such that no two cores
459 * can be measuring NB events using the same counters. Thus,
460 * we maintain a per-NB allocation table. The available slot
461 * is propagated using the event_constraint structure.
462 *
463 * We provide only one choice for each NB event based on
464 * the fact that only NB events have restrictions. Consequently,
465 * if a counter is available, there is a guarantee the NB event
466 * will be assigned to it. If no slot is available, an empty
467 * constraint is returned and scheduling will eventually fail
468 * for this event.
469 *
470 * Note that all cores attached the same NB compete for the same
471 * counters to host NB events, this is why we use atomic ops. Some
472 * multi-chip CPUs may have more than one NB.
473 *
474 * Given that resources are allocated (cmpxchg), they must be
475 * eventually freed for others to use. This is accomplished by
476 * calling __amd_put_nb_event_constraints()
477 *
478 * Non NB events are not impacted by this restriction.
479 */
480 static struct event_constraint *
__amd_get_nb_event_constraints(struct cpu_hw_events * cpuc,struct perf_event * event,struct event_constraint * c)481 __amd_get_nb_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
482 struct event_constraint *c)
483 {
484 struct hw_perf_event *hwc = &event->hw;
485 struct amd_nb *nb = cpuc->amd_nb;
486 struct perf_event *old;
487 int idx, new = -1;
488
489 if (!c)
490 c = &unconstrained;
491
492 if (cpuc->is_fake)
493 return c;
494
495 /*
496 * detect if already present, if so reuse
497 *
498 * cannot merge with actual allocation
499 * because of possible holes
500 *
501 * event can already be present yet not assigned (in hwc->idx)
502 * because of successive calls to x86_schedule_events() from
503 * hw_perf_group_sched_in() without hw_perf_enable()
504 */
505 for_each_set_bit(idx, c->idxmsk, x86_pmu_max_num_counters(NULL)) {
506 if (new == -1 || hwc->idx == idx)
507 /* assign free slot, prefer hwc->idx */
508 old = cmpxchg(nb->owners + idx, NULL, event);
509 else if (nb->owners[idx] == event)
510 /* event already present */
511 old = event;
512 else
513 continue;
514
515 if (old && old != event)
516 continue;
517
518 /* reassign to this slot */
519 if (new != -1)
520 cmpxchg(nb->owners + new, event, NULL);
521 new = idx;
522
523 /* already present, reuse */
524 if (old == event)
525 break;
526 }
527
528 if (new == -1)
529 return &emptyconstraint;
530
531 return &nb->event_constraints[new];
532 }
533
amd_alloc_nb(int cpu)534 static struct amd_nb *amd_alloc_nb(int cpu)
535 {
536 struct amd_nb *nb;
537 int i;
538
539 nb = kzalloc_node(sizeof(struct amd_nb), GFP_KERNEL, cpu_to_node(cpu));
540 if (!nb)
541 return NULL;
542
543 nb->nb_id = -1;
544
545 /*
546 * initialize all possible NB constraints
547 */
548 for_each_set_bit(i, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
549 __set_bit(i, nb->event_constraints[i].idxmsk);
550 nb->event_constraints[i].weight = 1;
551 }
552 return nb;
553 }
554
555 typedef void (amd_pmu_branch_reset_t)(void);
556 DEFINE_STATIC_CALL_NULL(amd_pmu_branch_reset, amd_pmu_branch_reset_t);
557
amd_pmu_cpu_reset(int cpu)558 static void amd_pmu_cpu_reset(int cpu)
559 {
560 if (x86_pmu.lbr_nr)
561 static_call(amd_pmu_branch_reset)();
562
563 if (x86_pmu.version < 2)
564 return;
565
566 /* Clear enable bits i.e. PerfCntrGlobalCtl.PerfCntrEn */
567 wrmsrq(MSR_AMD64_PERF_CNTR_GLOBAL_CTL, 0);
568
569 /*
570 * Clear freeze and overflow bits i.e. PerfCntrGLobalStatus.LbrFreeze
571 * and PerfCntrGLobalStatus.PerfCntrOvfl
572 */
573 wrmsrq(MSR_AMD64_PERF_CNTR_GLOBAL_STATUS_CLR,
574 GLOBAL_STATUS_LBRS_FROZEN | amd_pmu_global_cntr_mask);
575 }
576
amd_pmu_cpu_prepare(int cpu)577 static int amd_pmu_cpu_prepare(int cpu)
578 {
579 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
580
581 cpuc->lbr_sel = kzalloc_node(sizeof(struct er_account), GFP_KERNEL,
582 cpu_to_node(cpu));
583 if (!cpuc->lbr_sel)
584 return -ENOMEM;
585
586 WARN_ON_ONCE(cpuc->amd_nb);
587
588 if (!x86_pmu.amd_nb_constraints)
589 return 0;
590
591 cpuc->amd_nb = amd_alloc_nb(cpu);
592 if (cpuc->amd_nb)
593 return 0;
594
595 kfree(cpuc->lbr_sel);
596 cpuc->lbr_sel = NULL;
597
598 return -ENOMEM;
599 }
600
amd_pmu_cpu_starting(int cpu)601 static void amd_pmu_cpu_starting(int cpu)
602 {
603 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
604 void **onln = &cpuc->kfree_on_online[X86_PERF_KFREE_SHARED];
605 struct amd_nb *nb;
606 int i, nb_id;
607
608 cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
609 amd_pmu_cpu_reset(cpu);
610
611 if (!x86_pmu.amd_nb_constraints)
612 return;
613
614 nb_id = topology_amd_node_id(cpu);
615 WARN_ON_ONCE(nb_id == BAD_APICID);
616
617 for_each_online_cpu(i) {
618 nb = per_cpu(cpu_hw_events, i).amd_nb;
619 if (WARN_ON_ONCE(!nb))
620 continue;
621
622 if (nb->nb_id == nb_id) {
623 *onln = cpuc->amd_nb;
624 cpuc->amd_nb = nb;
625 break;
626 }
627 }
628
629 cpuc->amd_nb->nb_id = nb_id;
630 cpuc->amd_nb->refcnt++;
631 }
632
amd_pmu_cpu_dead(int cpu)633 static void amd_pmu_cpu_dead(int cpu)
634 {
635 struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
636
637 kfree(cpuhw->lbr_sel);
638 cpuhw->lbr_sel = NULL;
639
640 if (!x86_pmu.amd_nb_constraints)
641 return;
642
643 if (cpuhw->amd_nb) {
644 struct amd_nb *nb = cpuhw->amd_nb;
645
646 if (nb->nb_id == -1 || --nb->refcnt == 0)
647 kfree(nb);
648
649 cpuhw->amd_nb = NULL;
650 }
651 }
652
amd_pmu_set_global_ctl(u64 ctl)653 static __always_inline void amd_pmu_set_global_ctl(u64 ctl)
654 {
655 wrmsrq(MSR_AMD64_PERF_CNTR_GLOBAL_CTL, ctl);
656 }
657
amd_pmu_get_global_status(void)658 static inline u64 amd_pmu_get_global_status(void)
659 {
660 u64 status;
661
662 /* PerfCntrGlobalStatus is read-only */
663 rdmsrq(MSR_AMD64_PERF_CNTR_GLOBAL_STATUS, status);
664
665 return status;
666 }
667
amd_pmu_ack_global_status(u64 status)668 static inline void amd_pmu_ack_global_status(u64 status)
669 {
670 /*
671 * PerfCntrGlobalStatus is read-only but an overflow acknowledgment
672 * mechanism exists; writing 1 to a bit in PerfCntrGlobalStatusClr
673 * clears the same bit in PerfCntrGlobalStatus
674 */
675
676 wrmsrq(MSR_AMD64_PERF_CNTR_GLOBAL_STATUS_CLR, status);
677 }
678
amd_pmu_test_overflow_topbit(int idx)679 static bool amd_pmu_test_overflow_topbit(int idx)
680 {
681 u64 counter;
682
683 rdmsrq(x86_pmu_event_addr(idx), counter);
684
685 return !(counter & BIT_ULL(x86_pmu.cntval_bits - 1));
686 }
687
amd_pmu_test_overflow_status(int idx)688 static bool amd_pmu_test_overflow_status(int idx)
689 {
690 return amd_pmu_get_global_status() & BIT_ULL(idx);
691 }
692
693 DEFINE_STATIC_CALL(amd_pmu_test_overflow, amd_pmu_test_overflow_topbit);
694
695 /*
696 * When a PMC counter overflows, an NMI is used to process the event and
697 * reset the counter. NMI latency can result in the counter being updated
698 * before the NMI can run, which can result in what appear to be spurious
699 * NMIs. This function is intended to wait for the NMI to run and reset
700 * the counter to avoid possible unhandled NMI messages.
701 */
702 #define OVERFLOW_WAIT_COUNT 50
703
amd_pmu_wait_on_overflow(int idx)704 static void amd_pmu_wait_on_overflow(int idx)
705 {
706 unsigned int i;
707
708 /*
709 * Wait for the counter to be reset if it has overflowed. This loop
710 * should exit very, very quickly, but just in case, don't wait
711 * forever...
712 */
713 for (i = 0; i < OVERFLOW_WAIT_COUNT; i++) {
714 if (!static_call(amd_pmu_test_overflow)(idx))
715 break;
716
717 /* Might be in IRQ context, so can't sleep */
718 udelay(1);
719 }
720 }
721
amd_pmu_check_overflow(void)722 static void amd_pmu_check_overflow(void)
723 {
724 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
725 int idx;
726
727 /*
728 * This shouldn't be called from NMI context, but add a safeguard here
729 * to return, since if we're in NMI context we can't wait for an NMI
730 * to reset an overflowed counter value.
731 */
732 if (in_nmi())
733 return;
734
735 /*
736 * Check each counter for overflow and wait for it to be reset by the
737 * NMI if it has overflowed. This relies on the fact that all active
738 * counters are always enabled when this function is called and
739 * ARCH_PERFMON_EVENTSEL_INT is always set.
740 */
741 for_each_set_bit(idx, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
742 if (!test_bit(idx, cpuc->active_mask))
743 continue;
744
745 amd_pmu_wait_on_overflow(idx);
746 }
747 }
748
amd_pmu_enable_event(struct perf_event * event)749 static void amd_pmu_enable_event(struct perf_event *event)
750 {
751 x86_pmu_enable_event(event);
752 }
753
amd_pmu_enable_all(int added)754 static void amd_pmu_enable_all(int added)
755 {
756 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
757 int idx;
758
759 amd_brs_enable_all();
760
761 for_each_set_bit(idx, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
762 /* only activate events which are marked as active */
763 if (!test_bit(idx, cpuc->active_mask))
764 continue;
765
766 amd_pmu_enable_event(cpuc->events[idx]);
767 }
768 }
769
amd_pmu_v2_enable_event(struct perf_event * event)770 static void amd_pmu_v2_enable_event(struct perf_event *event)
771 {
772 struct hw_perf_event *hwc = &event->hw;
773
774 /*
775 * Testing cpu_hw_events.enabled should be skipped in this case unlike
776 * in x86_pmu_enable_event().
777 *
778 * Since cpu_hw_events.enabled is set only after returning from
779 * x86_pmu_start(), the PMCs must be programmed and kept ready.
780 * Counting starts only after x86_pmu_enable_all() is called.
781 */
782 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
783 }
784
amd_pmu_core_enable_all(void)785 static __always_inline void amd_pmu_core_enable_all(void)
786 {
787 amd_pmu_set_global_ctl(amd_pmu_global_cntr_mask);
788 }
789
amd_pmu_v2_enable_all(int added)790 static void amd_pmu_v2_enable_all(int added)
791 {
792 amd_pmu_lbr_enable_all();
793 amd_pmu_core_enable_all();
794 }
795
amd_pmu_disable_event(struct perf_event * event)796 static void amd_pmu_disable_event(struct perf_event *event)
797 {
798 x86_pmu_disable_event(event);
799
800 /*
801 * This can be called from NMI context (via x86_pmu_stop). The counter
802 * may have overflowed, but either way, we'll never see it get reset
803 * by the NMI if we're already in the NMI. And the NMI latency support
804 * below will take care of any pending NMI that might have been
805 * generated by the overflow.
806 */
807 if (in_nmi())
808 return;
809
810 amd_pmu_wait_on_overflow(event->hw.idx);
811 }
812
amd_pmu_disable_all(void)813 static void amd_pmu_disable_all(void)
814 {
815 amd_brs_disable_all();
816 x86_pmu_disable_all();
817 amd_pmu_check_overflow();
818 }
819
amd_pmu_core_disable_all(void)820 static __always_inline void amd_pmu_core_disable_all(void)
821 {
822 amd_pmu_set_global_ctl(0);
823 }
824
amd_pmu_v2_disable_all(void)825 static void amd_pmu_v2_disable_all(void)
826 {
827 amd_pmu_core_disable_all();
828 amd_pmu_lbr_disable_all();
829 amd_pmu_check_overflow();
830 }
831
832 DEFINE_STATIC_CALL_NULL(amd_pmu_branch_add, *x86_pmu.add);
833
amd_pmu_add_event(struct perf_event * event)834 static void amd_pmu_add_event(struct perf_event *event)
835 {
836 if (needs_branch_stack(event))
837 static_call(amd_pmu_branch_add)(event);
838 }
839
840 DEFINE_STATIC_CALL_NULL(amd_pmu_branch_del, *x86_pmu.del);
841
amd_pmu_del_event(struct perf_event * event)842 static void amd_pmu_del_event(struct perf_event *event)
843 {
844 if (needs_branch_stack(event))
845 static_call(amd_pmu_branch_del)(event);
846 }
847
848 /*
849 * Because of NMI latency, if multiple PMC counters are active or other sources
850 * of NMIs are received, the perf NMI handler can handle one or more overflowed
851 * PMC counters outside of the NMI associated with the PMC overflow. If the NMI
852 * doesn't arrive at the LAPIC in time to become a pending NMI, then the kernel
853 * back-to-back NMI support won't be active. This PMC handler needs to take into
854 * account that this can occur, otherwise this could result in unknown NMI
855 * messages being issued. Examples of this is PMC overflow while in the NMI
856 * handler when multiple PMCs are active or PMC overflow while handling some
857 * other source of an NMI.
858 *
859 * Attempt to mitigate this by creating an NMI window in which un-handled NMIs
860 * received during this window will be claimed. This prevents extending the
861 * window past when it is possible that latent NMIs should be received. The
862 * per-CPU perf_nmi_tstamp will be set to the window end time whenever perf has
863 * handled a counter. When an un-handled NMI is received, it will be claimed
864 * only if arriving within that window.
865 */
amd_pmu_adjust_nmi_window(int handled)866 static inline int amd_pmu_adjust_nmi_window(int handled)
867 {
868 /*
869 * If a counter was handled, record a timestamp such that un-handled
870 * NMIs will be claimed if arriving within that window.
871 */
872 if (handled) {
873 this_cpu_write(perf_nmi_tstamp, jiffies + perf_nmi_window);
874
875 return handled;
876 }
877
878 if (time_after(jiffies, this_cpu_read(perf_nmi_tstamp)))
879 return NMI_DONE;
880
881 return NMI_HANDLED;
882 }
883
amd_pmu_handle_irq(struct pt_regs * regs)884 static int amd_pmu_handle_irq(struct pt_regs *regs)
885 {
886 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
887 int handled;
888 int pmu_enabled;
889
890 /*
891 * Save the PMU state.
892 * It needs to be restored when leaving the handler.
893 */
894 pmu_enabled = cpuc->enabled;
895 cpuc->enabled = 0;
896
897 amd_brs_disable_all();
898
899 /* Drain BRS is in use (could be inactive) */
900 if (cpuc->lbr_users)
901 amd_brs_drain();
902
903 /* Process any counter overflows */
904 handled = x86_pmu_handle_irq(regs);
905
906 cpuc->enabled = pmu_enabled;
907 if (pmu_enabled)
908 amd_brs_enable_all();
909
910 return amd_pmu_adjust_nmi_window(handled);
911 }
912
913 /*
914 * AMD-specific callback invoked through perf_snapshot_branch_stack static
915 * call, defined in include/linux/perf_event.h. See its definition for API
916 * details. It's up to caller to provide enough space in *entries* to fit all
917 * LBR records, otherwise returned result will be truncated to *cnt* entries.
918 */
amd_pmu_v2_snapshot_branch_stack(struct perf_branch_entry * entries,unsigned int cnt)919 static int amd_pmu_v2_snapshot_branch_stack(struct perf_branch_entry *entries, unsigned int cnt)
920 {
921 struct cpu_hw_events *cpuc;
922 unsigned long flags;
923
924 /*
925 * The sequence of steps to freeze LBR should be completely inlined
926 * and contain no branches to minimize contamination of LBR snapshot
927 */
928 local_irq_save(flags);
929 amd_pmu_core_disable_all();
930 __amd_pmu_lbr_disable();
931
932 cpuc = this_cpu_ptr(&cpu_hw_events);
933
934 amd_pmu_lbr_read();
935 cnt = min(cnt, x86_pmu.lbr_nr);
936 memcpy(entries, cpuc->lbr_entries, sizeof(struct perf_branch_entry) * cnt);
937
938 amd_pmu_v2_enable_all(0);
939 local_irq_restore(flags);
940
941 return cnt;
942 }
943
amd_pmu_v2_handle_irq(struct pt_regs * regs)944 static int amd_pmu_v2_handle_irq(struct pt_regs *regs)
945 {
946 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
947 static atomic64_t status_warned = ATOMIC64_INIT(0);
948 u64 reserved, status, mask, new_bits, prev_bits;
949 struct perf_sample_data data;
950 struct hw_perf_event *hwc;
951 struct perf_event *event;
952 int handled = 0, idx;
953 bool pmu_enabled;
954
955 /*
956 * Save the PMU state as it needs to be restored when leaving the
957 * handler
958 */
959 pmu_enabled = cpuc->enabled;
960 cpuc->enabled = 0;
961
962 /* Stop counting but do not disable LBR */
963 amd_pmu_core_disable_all();
964
965 status = amd_pmu_get_global_status();
966
967 /* Check if any overflows are pending */
968 if (!status)
969 goto done;
970
971 /* Read branch records */
972 if (x86_pmu.lbr_nr) {
973 amd_pmu_lbr_read();
974 status &= ~GLOBAL_STATUS_LBRS_FROZEN;
975 }
976
977 reserved = status & ~amd_pmu_global_cntr_mask;
978 if (reserved)
979 pr_warn_once("Reserved PerfCntrGlobalStatus bits are set (0x%llx), please consider updating microcode\n",
980 reserved);
981
982 /* Clear any reserved bits set by buggy microcode */
983 status &= amd_pmu_global_cntr_mask;
984
985 for_each_set_bit(idx, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
986 if (!test_bit(idx, cpuc->active_mask))
987 continue;
988
989 event = cpuc->events[idx];
990 hwc = &event->hw;
991 x86_perf_event_update(event);
992 mask = BIT_ULL(idx);
993
994 if (!(status & mask))
995 continue;
996
997 /* Event overflow */
998 handled++;
999 status &= ~mask;
1000 perf_sample_data_init(&data, 0, hwc->last_period);
1001
1002 if (!x86_perf_event_set_period(event))
1003 continue;
1004
1005 perf_sample_save_brstack(&data, event, &cpuc->lbr_stack, NULL);
1006
1007 perf_event_overflow(event, &data, regs);
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