xref: /openbmc/linux/arch/x86/events/amd/core.c (revision dc6a81c3)
1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/perf_event.h>
3 #include <linux/export.h>
4 #include <linux/types.h>
5 #include <linux/init.h>
6 #include <linux/slab.h>
7 #include <linux/delay.h>
8 #include <linux/jiffies.h>
9 #include <asm/apicdef.h>
10 #include <asm/nmi.h>
11 
12 #include "../perf_event.h"
13 
14 static DEFINE_PER_CPU(unsigned long, perf_nmi_tstamp);
15 static unsigned long perf_nmi_window;
16 
17 /* AMD Event 0xFFF: Merge.  Used with Large Increment per Cycle events */
18 #define AMD_MERGE_EVENT ((0xFULL << 32) | 0xFFULL)
19 #define AMD_MERGE_EVENT_ENABLE (AMD_MERGE_EVENT | ARCH_PERFMON_EVENTSEL_ENABLE)
20 
21 static __initconst const u64 amd_hw_cache_event_ids
22 				[PERF_COUNT_HW_CACHE_MAX]
23 				[PERF_COUNT_HW_CACHE_OP_MAX]
24 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
25 {
26  [ C(L1D) ] = {
27 	[ C(OP_READ) ] = {
28 		[ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses        */
29 		[ C(RESULT_MISS)   ] = 0x0141, /* Data Cache Misses          */
30 	},
31 	[ C(OP_WRITE) ] = {
32 		[ C(RESULT_ACCESS) ] = 0,
33 		[ C(RESULT_MISS)   ] = 0,
34 	},
35 	[ C(OP_PREFETCH) ] = {
36 		[ C(RESULT_ACCESS) ] = 0x0267, /* Data Prefetcher :attempts  */
37 		[ C(RESULT_MISS)   ] = 0x0167, /* Data Prefetcher :cancelled */
38 	},
39  },
40  [ C(L1I ) ] = {
41 	[ C(OP_READ) ] = {
42 		[ C(RESULT_ACCESS) ] = 0x0080, /* Instruction cache fetches  */
43 		[ C(RESULT_MISS)   ] = 0x0081, /* Instruction cache misses   */
44 	},
45 	[ C(OP_WRITE) ] = {
46 		[ C(RESULT_ACCESS) ] = -1,
47 		[ C(RESULT_MISS)   ] = -1,
48 	},
49 	[ C(OP_PREFETCH) ] = {
50 		[ C(RESULT_ACCESS) ] = 0x014B, /* Prefetch Instructions :Load */
51 		[ C(RESULT_MISS)   ] = 0,
52 	},
53  },
54  [ C(LL  ) ] = {
55 	[ C(OP_READ) ] = {
56 		[ C(RESULT_ACCESS) ] = 0x037D, /* Requests to L2 Cache :IC+DC */
57 		[ C(RESULT_MISS)   ] = 0x037E, /* L2 Cache Misses : IC+DC     */
58 	},
59 	[ C(OP_WRITE) ] = {
60 		[ C(RESULT_ACCESS) ] = 0x017F, /* L2 Fill/Writeback           */
61 		[ C(RESULT_MISS)   ] = 0,
62 	},
63 	[ C(OP_PREFETCH) ] = {
64 		[ C(RESULT_ACCESS) ] = 0,
65 		[ C(RESULT_MISS)   ] = 0,
66 	},
67  },
68  [ C(DTLB) ] = {
69 	[ C(OP_READ) ] = {
70 		[ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses        */
71 		[ C(RESULT_MISS)   ] = 0x0746, /* L1_DTLB_AND_L2_DLTB_MISS.ALL */
72 	},
73 	[ C(OP_WRITE) ] = {
74 		[ C(RESULT_ACCESS) ] = 0,
75 		[ C(RESULT_MISS)   ] = 0,
76 	},
77 	[ C(OP_PREFETCH) ] = {
78 		[ C(RESULT_ACCESS) ] = 0,
79 		[ C(RESULT_MISS)   ] = 0,
80 	},
81  },
82  [ C(ITLB) ] = {
83 	[ C(OP_READ) ] = {
84 		[ C(RESULT_ACCESS) ] = 0x0080, /* Instruction fecthes        */
85 		[ C(RESULT_MISS)   ] = 0x0385, /* L1_ITLB_AND_L2_ITLB_MISS.ALL */
86 	},
87 	[ C(OP_WRITE) ] = {
88 		[ C(RESULT_ACCESS) ] = -1,
89 		[ C(RESULT_MISS)   ] = -1,
90 	},
91 	[ C(OP_PREFETCH) ] = {
92 		[ C(RESULT_ACCESS) ] = -1,
93 		[ C(RESULT_MISS)   ] = -1,
94 	},
95  },
96  [ C(BPU ) ] = {
97 	[ C(OP_READ) ] = {
98 		[ C(RESULT_ACCESS) ] = 0x00c2, /* Retired Branch Instr.      */
99 		[ C(RESULT_MISS)   ] = 0x00c3, /* Retired Mispredicted BI    */
100 	},
101 	[ C(OP_WRITE) ] = {
102 		[ C(RESULT_ACCESS) ] = -1,
103 		[ C(RESULT_MISS)   ] = -1,
104 	},
105 	[ C(OP_PREFETCH) ] = {
106 		[ C(RESULT_ACCESS) ] = -1,
107 		[ C(RESULT_MISS)   ] = -1,
108 	},
109  },
110  [ C(NODE) ] = {
111 	[ C(OP_READ) ] = {
112 		[ C(RESULT_ACCESS) ] = 0xb8e9, /* CPU Request to Memory, l+r */
113 		[ C(RESULT_MISS)   ] = 0x98e9, /* CPU Request to Memory, r   */
114 	},
115 	[ C(OP_WRITE) ] = {
116 		[ C(RESULT_ACCESS) ] = -1,
117 		[ C(RESULT_MISS)   ] = -1,
118 	},
119 	[ C(OP_PREFETCH) ] = {
120 		[ C(RESULT_ACCESS) ] = -1,
121 		[ C(RESULT_MISS)   ] = -1,
122 	},
123  },
124 };
125 
126 static __initconst const u64 amd_hw_cache_event_ids_f17h
127 				[PERF_COUNT_HW_CACHE_MAX]
128 				[PERF_COUNT_HW_CACHE_OP_MAX]
129 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
130 [C(L1D)] = {
131 	[C(OP_READ)] = {
132 		[C(RESULT_ACCESS)] = 0x0040, /* Data Cache Accesses */
133 		[C(RESULT_MISS)]   = 0xc860, /* L2$ access from DC Miss */
134 	},
135 	[C(OP_WRITE)] = {
136 		[C(RESULT_ACCESS)] = 0,
137 		[C(RESULT_MISS)]   = 0,
138 	},
139 	[C(OP_PREFETCH)] = {
140 		[C(RESULT_ACCESS)] = 0xff5a, /* h/w prefetch DC Fills */
141 		[C(RESULT_MISS)]   = 0,
142 	},
143 },
144 [C(L1I)] = {
145 	[C(OP_READ)] = {
146 		[C(RESULT_ACCESS)] = 0x0080, /* Instruction cache fetches  */
147 		[C(RESULT_MISS)]   = 0x0081, /* Instruction cache misses   */
148 	},
149 	[C(OP_WRITE)] = {
150 		[C(RESULT_ACCESS)] = -1,
151 		[C(RESULT_MISS)]   = -1,
152 	},
153 	[C(OP_PREFETCH)] = {
154 		[C(RESULT_ACCESS)] = 0,
155 		[C(RESULT_MISS)]   = 0,
156 	},
157 },
158 [C(LL)] = {
159 	[C(OP_READ)] = {
160 		[C(RESULT_ACCESS)] = 0,
161 		[C(RESULT_MISS)]   = 0,
162 	},
163 	[C(OP_WRITE)] = {
164 		[C(RESULT_ACCESS)] = 0,
165 		[C(RESULT_MISS)]   = 0,
166 	},
167 	[C(OP_PREFETCH)] = {
168 		[C(RESULT_ACCESS)] = 0,
169 		[C(RESULT_MISS)]   = 0,
170 	},
171 },
172 [C(DTLB)] = {
173 	[C(OP_READ)] = {
174 		[C(RESULT_ACCESS)] = 0xff45, /* All L2 DTLB accesses */
175 		[C(RESULT_MISS)]   = 0xf045, /* L2 DTLB misses (PT walks) */
176 	},
177 	[C(OP_WRITE)] = {
178 		[C(RESULT_ACCESS)] = 0,
179 		[C(RESULT_MISS)]   = 0,
180 	},
181 	[C(OP_PREFETCH)] = {
182 		[C(RESULT_ACCESS)] = 0,
183 		[C(RESULT_MISS)]   = 0,
184 	},
185 },
186 [C(ITLB)] = {
187 	[C(OP_READ)] = {
188 		[C(RESULT_ACCESS)] = 0x0084, /* L1 ITLB misses, L2 ITLB hits */
189 		[C(RESULT_MISS)]   = 0xff85, /* L1 ITLB misses, L2 misses */
190 	},
191 	[C(OP_WRITE)] = {
192 		[C(RESULT_ACCESS)] = -1,
193 		[C(RESULT_MISS)]   = -1,
194 	},
195 	[C(OP_PREFETCH)] = {
196 		[C(RESULT_ACCESS)] = -1,
197 		[C(RESULT_MISS)]   = -1,
198 	},
199 },
200 [C(BPU)] = {
201 	[C(OP_READ)] = {
202 		[C(RESULT_ACCESS)] = 0x00c2, /* Retired Branch Instr.      */
203 		[C(RESULT_MISS)]   = 0x00c3, /* Retired Mispredicted BI    */
204 	},
205 	[C(OP_WRITE)] = {
206 		[C(RESULT_ACCESS)] = -1,
207 		[C(RESULT_MISS)]   = -1,
208 	},
209 	[C(OP_PREFETCH)] = {
210 		[C(RESULT_ACCESS)] = -1,
211 		[C(RESULT_MISS)]   = -1,
212 	},
213 },
214 [C(NODE)] = {
215 	[C(OP_READ)] = {
216 		[C(RESULT_ACCESS)] = 0,
217 		[C(RESULT_MISS)]   = 0,
218 	},
219 	[C(OP_WRITE)] = {
220 		[C(RESULT_ACCESS)] = -1,
221 		[C(RESULT_MISS)]   = -1,
222 	},
223 	[C(OP_PREFETCH)] = {
224 		[C(RESULT_ACCESS)] = -1,
225 		[C(RESULT_MISS)]   = -1,
226 	},
227 },
228 };
229 
230 /*
231  * AMD Performance Monitor K7 and later, up to and including Family 16h:
232  */
233 static const u64 amd_perfmon_event_map[PERF_COUNT_HW_MAX] =
234 {
235 	[PERF_COUNT_HW_CPU_CYCLES]		= 0x0076,
236 	[PERF_COUNT_HW_INSTRUCTIONS]		= 0x00c0,
237 	[PERF_COUNT_HW_CACHE_REFERENCES]	= 0x077d,
238 	[PERF_COUNT_HW_CACHE_MISSES]		= 0x077e,
239 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x00c2,
240 	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x00c3,
241 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= 0x00d0, /* "Decoder empty" event */
242 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= 0x00d1, /* "Dispatch stalls" event */
243 };
244 
245 /*
246  * AMD Performance Monitor Family 17h and later:
247  */
248 static const u64 amd_f17h_perfmon_event_map[PERF_COUNT_HW_MAX] =
249 {
250 	[PERF_COUNT_HW_CPU_CYCLES]		= 0x0076,
251 	[PERF_COUNT_HW_INSTRUCTIONS]		= 0x00c0,
252 	[PERF_COUNT_HW_CACHE_REFERENCES]	= 0xff60,
253 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x00c2,
254 	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x00c3,
255 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= 0x0287,
256 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= 0x0187,
257 };
258 
259 static u64 amd_pmu_event_map(int hw_event)
260 {
261 	if (boot_cpu_data.x86 >= 0x17)
262 		return amd_f17h_perfmon_event_map[hw_event];
263 
264 	return amd_perfmon_event_map[hw_event];
265 }
266 
267 /*
268  * Previously calculated offsets
269  */
270 static unsigned int event_offsets[X86_PMC_IDX_MAX] __read_mostly;
271 static unsigned int count_offsets[X86_PMC_IDX_MAX] __read_mostly;
272 
273 /*
274  * Legacy CPUs:
275  *   4 counters starting at 0xc0010000 each offset by 1
276  *
277  * CPUs with core performance counter extensions:
278  *   6 counters starting at 0xc0010200 each offset by 2
279  */
280 static inline int amd_pmu_addr_offset(int index, bool eventsel)
281 {
282 	int offset;
283 
284 	if (!index)
285 		return index;
286 
287 	if (eventsel)
288 		offset = event_offsets[index];
289 	else
290 		offset = count_offsets[index];
291 
292 	if (offset)
293 		return offset;
294 
295 	if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
296 		offset = index;
297 	else
298 		offset = index << 1;
299 
300 	if (eventsel)
301 		event_offsets[index] = offset;
302 	else
303 		count_offsets[index] = offset;
304 
305 	return offset;
306 }
307 
308 /*
309  * AMD64 events are detected based on their event codes.
310  */
311 static inline unsigned int amd_get_event_code(struct hw_perf_event *hwc)
312 {
313 	return ((hwc->config >> 24) & 0x0f00) | (hwc->config & 0x00ff);
314 }
315 
316 static inline bool amd_is_pair_event_code(struct hw_perf_event *hwc)
317 {
318 	if (!(x86_pmu.flags & PMU_FL_PAIR))
319 		return false;
320 
321 	switch (amd_get_event_code(hwc)) {
322 	case 0x003:	return true;	/* Retired SSE/AVX FLOPs */
323 	default:	return false;
324 	}
325 }
326 
327 static int amd_core_hw_config(struct perf_event *event)
328 {
329 	if (event->attr.exclude_host && event->attr.exclude_guest)
330 		/*
331 		 * When HO == GO == 1 the hardware treats that as GO == HO == 0
332 		 * and will count in both modes. We don't want to count in that
333 		 * case so we emulate no-counting by setting US = OS = 0.
334 		 */
335 		event->hw.config &= ~(ARCH_PERFMON_EVENTSEL_USR |
336 				      ARCH_PERFMON_EVENTSEL_OS);
337 	else if (event->attr.exclude_host)
338 		event->hw.config |= AMD64_EVENTSEL_GUESTONLY;
339 	else if (event->attr.exclude_guest)
340 		event->hw.config |= AMD64_EVENTSEL_HOSTONLY;
341 
342 	if ((x86_pmu.flags & PMU_FL_PAIR) && amd_is_pair_event_code(&event->hw))
343 		event->hw.flags |= PERF_X86_EVENT_PAIR;
344 
345 	return 0;
346 }
347 
348 static inline int amd_is_nb_event(struct hw_perf_event *hwc)
349 {
350 	return (hwc->config & 0xe0) == 0xe0;
351 }
352 
353 static inline int amd_has_nb(struct cpu_hw_events *cpuc)
354 {
355 	struct amd_nb *nb = cpuc->amd_nb;
356 
357 	return nb && nb->nb_id != -1;
358 }
359 
360 static int amd_pmu_hw_config(struct perf_event *event)
361 {
362 	int ret;
363 
364 	/* pass precise event sampling to ibs: */
365 	if (event->attr.precise_ip && get_ibs_caps())
366 		return -ENOENT;
367 
368 	if (has_branch_stack(event))
369 		return -EOPNOTSUPP;
370 
371 	ret = x86_pmu_hw_config(event);
372 	if (ret)
373 		return ret;
374 
375 	if (event->attr.type == PERF_TYPE_RAW)
376 		event->hw.config |= event->attr.config & AMD64_RAW_EVENT_MASK;
377 
378 	return amd_core_hw_config(event);
379 }
380 
381 static void __amd_put_nb_event_constraints(struct cpu_hw_events *cpuc,
382 					   struct perf_event *event)
383 {
384 	struct amd_nb *nb = cpuc->amd_nb;
385 	int i;
386 
387 	/*
388 	 * need to scan whole list because event may not have
389 	 * been assigned during scheduling
390 	 *
391 	 * no race condition possible because event can only
392 	 * be removed on one CPU at a time AND PMU is disabled
393 	 * when we come here
394 	 */
395 	for (i = 0; i < x86_pmu.num_counters; i++) {
396 		if (cmpxchg(nb->owners + i, event, NULL) == event)
397 			break;
398 	}
399 }
400 
401  /*
402   * AMD64 NorthBridge events need special treatment because
403   * counter access needs to be synchronized across all cores
404   * of a package. Refer to BKDG section 3.12
405   *
406   * NB events are events measuring L3 cache, Hypertransport
407   * traffic. They are identified by an event code >= 0xe00.
408   * They measure events on the NorthBride which is shared
409   * by all cores on a package. NB events are counted on a
410   * shared set of counters. When a NB event is programmed
411   * in a counter, the data actually comes from a shared
412   * counter. Thus, access to those counters needs to be
413   * synchronized.
414   *
415   * We implement the synchronization such that no two cores
416   * can be measuring NB events using the same counters. Thus,
417   * we maintain a per-NB allocation table. The available slot
418   * is propagated using the event_constraint structure.
419   *
420   * We provide only one choice for each NB event based on
421   * the fact that only NB events have restrictions. Consequently,
422   * if a counter is available, there is a guarantee the NB event
423   * will be assigned to it. If no slot is available, an empty
424   * constraint is returned and scheduling will eventually fail
425   * for this event.
426   *
427   * Note that all cores attached the same NB compete for the same
428   * counters to host NB events, this is why we use atomic ops. Some
429   * multi-chip CPUs may have more than one NB.
430   *
431   * Given that resources are allocated (cmpxchg), they must be
432   * eventually freed for others to use. This is accomplished by
433   * calling __amd_put_nb_event_constraints()
434   *
435   * Non NB events are not impacted by this restriction.
436   */
437 static struct event_constraint *
438 __amd_get_nb_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
439 			       struct event_constraint *c)
440 {
441 	struct hw_perf_event *hwc = &event->hw;
442 	struct amd_nb *nb = cpuc->amd_nb;
443 	struct perf_event *old;
444 	int idx, new = -1;
445 
446 	if (!c)
447 		c = &unconstrained;
448 
449 	if (cpuc->is_fake)
450 		return c;
451 
452 	/*
453 	 * detect if already present, if so reuse
454 	 *
455 	 * cannot merge with actual allocation
456 	 * because of possible holes
457 	 *
458 	 * event can already be present yet not assigned (in hwc->idx)
459 	 * because of successive calls to x86_schedule_events() from
460 	 * hw_perf_group_sched_in() without hw_perf_enable()
461 	 */
462 	for_each_set_bit(idx, c->idxmsk, x86_pmu.num_counters) {
463 		if (new == -1 || hwc->idx == idx)
464 			/* assign free slot, prefer hwc->idx */
465 			old = cmpxchg(nb->owners + idx, NULL, event);
466 		else if (nb->owners[idx] == event)
467 			/* event already present */
468 			old = event;
469 		else
470 			continue;
471 
472 		if (old && old != event)
473 			continue;
474 
475 		/* reassign to this slot */
476 		if (new != -1)
477 			cmpxchg(nb->owners + new, event, NULL);
478 		new = idx;
479 
480 		/* already present, reuse */
481 		if (old == event)
482 			break;
483 	}
484 
485 	if (new == -1)
486 		return &emptyconstraint;
487 
488 	return &nb->event_constraints[new];
489 }
490 
491 static struct amd_nb *amd_alloc_nb(int cpu)
492 {
493 	struct amd_nb *nb;
494 	int i;
495 
496 	nb = kzalloc_node(sizeof(struct amd_nb), GFP_KERNEL, cpu_to_node(cpu));
497 	if (!nb)
498 		return NULL;
499 
500 	nb->nb_id = -1;
501 
502 	/*
503 	 * initialize all possible NB constraints
504 	 */
505 	for (i = 0; i < x86_pmu.num_counters; i++) {
506 		__set_bit(i, nb->event_constraints[i].idxmsk);
507 		nb->event_constraints[i].weight = 1;
508 	}
509 	return nb;
510 }
511 
512 static int amd_pmu_cpu_prepare(int cpu)
513 {
514 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
515 
516 	WARN_ON_ONCE(cpuc->amd_nb);
517 
518 	if (!x86_pmu.amd_nb_constraints)
519 		return 0;
520 
521 	cpuc->amd_nb = amd_alloc_nb(cpu);
522 	if (!cpuc->amd_nb)
523 		return -ENOMEM;
524 
525 	return 0;
526 }
527 
528 static void amd_pmu_cpu_starting(int cpu)
529 {
530 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
531 	void **onln = &cpuc->kfree_on_online[X86_PERF_KFREE_SHARED];
532 	struct amd_nb *nb;
533 	int i, nb_id;
534 
535 	cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
536 
537 	if (!x86_pmu.amd_nb_constraints)
538 		return;
539 
540 	nb_id = amd_get_nb_id(cpu);
541 	WARN_ON_ONCE(nb_id == BAD_APICID);
542 
543 	for_each_online_cpu(i) {
544 		nb = per_cpu(cpu_hw_events, i).amd_nb;
545 		if (WARN_ON_ONCE(!nb))
546 			continue;
547 
548 		if (nb->nb_id == nb_id) {
549 			*onln = cpuc->amd_nb;
550 			cpuc->amd_nb = nb;
551 			break;
552 		}
553 	}
554 
555 	cpuc->amd_nb->nb_id = nb_id;
556 	cpuc->amd_nb->refcnt++;
557 }
558 
559 static void amd_pmu_cpu_dead(int cpu)
560 {
561 	struct cpu_hw_events *cpuhw;
562 
563 	if (!x86_pmu.amd_nb_constraints)
564 		return;
565 
566 	cpuhw = &per_cpu(cpu_hw_events, cpu);
567 
568 	if (cpuhw->amd_nb) {
569 		struct amd_nb *nb = cpuhw->amd_nb;
570 
571 		if (nb->nb_id == -1 || --nb->refcnt == 0)
572 			kfree(nb);
573 
574 		cpuhw->amd_nb = NULL;
575 	}
576 }
577 
578 /*
579  * When a PMC counter overflows, an NMI is used to process the event and
580  * reset the counter. NMI latency can result in the counter being updated
581  * before the NMI can run, which can result in what appear to be spurious
582  * NMIs. This function is intended to wait for the NMI to run and reset
583  * the counter to avoid possible unhandled NMI messages.
584  */
585 #define OVERFLOW_WAIT_COUNT	50
586 
587 static void amd_pmu_wait_on_overflow(int idx)
588 {
589 	unsigned int i;
590 	u64 counter;
591 
592 	/*
593 	 * Wait for the counter to be reset if it has overflowed. This loop
594 	 * should exit very, very quickly, but just in case, don't wait
595 	 * forever...
596 	 */
597 	for (i = 0; i < OVERFLOW_WAIT_COUNT; i++) {
598 		rdmsrl(x86_pmu_event_addr(idx), counter);
599 		if (counter & (1ULL << (x86_pmu.cntval_bits - 1)))
600 			break;
601 
602 		/* Might be in IRQ context, so can't sleep */
603 		udelay(1);
604 	}
605 }
606 
607 static void amd_pmu_disable_all(void)
608 {
609 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
610 	int idx;
611 
612 	x86_pmu_disable_all();
613 
614 	/*
615 	 * This shouldn't be called from NMI context, but add a safeguard here
616 	 * to return, since if we're in NMI context we can't wait for an NMI
617 	 * to reset an overflowed counter value.
618 	 */
619 	if (in_nmi())
620 		return;
621 
622 	/*
623 	 * Check each counter for overflow and wait for it to be reset by the
624 	 * NMI if it has overflowed. This relies on the fact that all active
625 	 * counters are always enabled when this function is caled and
626 	 * ARCH_PERFMON_EVENTSEL_INT is always set.
627 	 */
628 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
629 		if (!test_bit(idx, cpuc->active_mask))
630 			continue;
631 
632 		amd_pmu_wait_on_overflow(idx);
633 	}
634 }
635 
636 static void amd_pmu_disable_event(struct perf_event *event)
637 {
638 	x86_pmu_disable_event(event);
639 
640 	/*
641 	 * This can be called from NMI context (via x86_pmu_stop). The counter
642 	 * may have overflowed, but either way, we'll never see it get reset
643 	 * by the NMI if we're already in the NMI. And the NMI latency support
644 	 * below will take care of any pending NMI that might have been
645 	 * generated by the overflow.
646 	 */
647 	if (in_nmi())
648 		return;
649 
650 	amd_pmu_wait_on_overflow(event->hw.idx);
651 }
652 
653 /*
654  * Because of NMI latency, if multiple PMC counters are active or other sources
655  * of NMIs are received, the perf NMI handler can handle one or more overflowed
656  * PMC counters outside of the NMI associated with the PMC overflow. If the NMI
657  * doesn't arrive at the LAPIC in time to become a pending NMI, then the kernel
658  * back-to-back NMI support won't be active. This PMC handler needs to take into
659  * account that this can occur, otherwise this could result in unknown NMI
660  * messages being issued. Examples of this is PMC overflow while in the NMI
661  * handler when multiple PMCs are active or PMC overflow while handling some
662  * other source of an NMI.
663  *
664  * Attempt to mitigate this by creating an NMI window in which un-handled NMIs
665  * received during this window will be claimed. This prevents extending the
666  * window past when it is possible that latent NMIs should be received. The
667  * per-CPU perf_nmi_tstamp will be set to the window end time whenever perf has
668  * handled a counter. When an un-handled NMI is received, it will be claimed
669  * only if arriving within that window.
670  */
671 static int amd_pmu_handle_irq(struct pt_regs *regs)
672 {
673 	int handled;
674 
675 	/* Process any counter overflows */
676 	handled = x86_pmu_handle_irq(regs);
677 
678 	/*
679 	 * If a counter was handled, record a timestamp such that un-handled
680 	 * NMIs will be claimed if arriving within that window.
681 	 */
682 	if (handled) {
683 		this_cpu_write(perf_nmi_tstamp, jiffies + perf_nmi_window);
684 
685 		return handled;
686 	}
687 
688 	if (time_after(jiffies, this_cpu_read(perf_nmi_tstamp)))
689 		return NMI_DONE;
690 
691 	return NMI_HANDLED;
692 }
693 
694 static struct event_constraint *
695 amd_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
696 			  struct perf_event *event)
697 {
698 	/*
699 	 * if not NB event or no NB, then no constraints
700 	 */
701 	if (!(amd_has_nb(cpuc) && amd_is_nb_event(&event->hw)))
702 		return &unconstrained;
703 
704 	return __amd_get_nb_event_constraints(cpuc, event, NULL);
705 }
706 
707 static void amd_put_event_constraints(struct cpu_hw_events *cpuc,
708 				      struct perf_event *event)
709 {
710 	if (amd_has_nb(cpuc) && amd_is_nb_event(&event->hw))
711 		__amd_put_nb_event_constraints(cpuc, event);
712 }
713 
714 PMU_FORMAT_ATTR(event,	"config:0-7,32-35");
715 PMU_FORMAT_ATTR(umask,	"config:8-15"	);
716 PMU_FORMAT_ATTR(edge,	"config:18"	);
717 PMU_FORMAT_ATTR(inv,	"config:23"	);
718 PMU_FORMAT_ATTR(cmask,	"config:24-31"	);
719 
720 static struct attribute *amd_format_attr[] = {
721 	&format_attr_event.attr,
722 	&format_attr_umask.attr,
723 	&format_attr_edge.attr,
724 	&format_attr_inv.attr,
725 	&format_attr_cmask.attr,
726 	NULL,
727 };
728 
729 /* AMD Family 15h */
730 
731 #define AMD_EVENT_TYPE_MASK	0x000000F0ULL
732 
733 #define AMD_EVENT_FP		0x00000000ULL ... 0x00000010ULL
734 #define AMD_EVENT_LS		0x00000020ULL ... 0x00000030ULL
735 #define AMD_EVENT_DC		0x00000040ULL ... 0x00000050ULL
736 #define AMD_EVENT_CU		0x00000060ULL ... 0x00000070ULL
737 #define AMD_EVENT_IC_DE		0x00000080ULL ... 0x00000090ULL
738 #define AMD_EVENT_EX_LS		0x000000C0ULL
739 #define AMD_EVENT_DE		0x000000D0ULL
740 #define AMD_EVENT_NB		0x000000E0ULL ... 0x000000F0ULL
741 
742 /*
743  * AMD family 15h event code/PMC mappings:
744  *
745  * type = event_code & 0x0F0:
746  *
747  * 0x000	FP	PERF_CTL[5:3]
748  * 0x010	FP	PERF_CTL[5:3]
749  * 0x020	LS	PERF_CTL[5:0]
750  * 0x030	LS	PERF_CTL[5:0]
751  * 0x040	DC	PERF_CTL[5:0]
752  * 0x050	DC	PERF_CTL[5:0]
753  * 0x060	CU	PERF_CTL[2:0]
754  * 0x070	CU	PERF_CTL[2:0]
755  * 0x080	IC/DE	PERF_CTL[2:0]
756  * 0x090	IC/DE	PERF_CTL[2:0]
757  * 0x0A0	---
758  * 0x0B0	---
759  * 0x0C0	EX/LS	PERF_CTL[5:0]
760  * 0x0D0	DE	PERF_CTL[2:0]
761  * 0x0E0	NB	NB_PERF_CTL[3:0]
762  * 0x0F0	NB	NB_PERF_CTL[3:0]
763  *
764  * Exceptions:
765  *
766  * 0x000	FP	PERF_CTL[3], PERF_CTL[5:3] (*)
767  * 0x003	FP	PERF_CTL[3]
768  * 0x004	FP	PERF_CTL[3], PERF_CTL[5:3] (*)
769  * 0x00B	FP	PERF_CTL[3]
770  * 0x00D	FP	PERF_CTL[3]
771  * 0x023	DE	PERF_CTL[2:0]
772  * 0x02D	LS	PERF_CTL[3]
773  * 0x02E	LS	PERF_CTL[3,0]
774  * 0x031	LS	PERF_CTL[2:0] (**)
775  * 0x043	CU	PERF_CTL[2:0]
776  * 0x045	CU	PERF_CTL[2:0]
777  * 0x046	CU	PERF_CTL[2:0]
778  * 0x054	CU	PERF_CTL[2:0]
779  * 0x055	CU	PERF_CTL[2:0]
780  * 0x08F	IC	PERF_CTL[0]
781  * 0x187	DE	PERF_CTL[0]
782  * 0x188	DE	PERF_CTL[0]
783  * 0x0DB	EX	PERF_CTL[5:0]
784  * 0x0DC	LS	PERF_CTL[5:0]
785  * 0x0DD	LS	PERF_CTL[5:0]
786  * 0x0DE	LS	PERF_CTL[5:0]
787  * 0x0DF	LS	PERF_CTL[5:0]
788  * 0x1C0	EX	PERF_CTL[5:3]
789  * 0x1D6	EX	PERF_CTL[5:0]
790  * 0x1D8	EX	PERF_CTL[5:0]
791  *
792  * (*)  depending on the umask all FPU counters may be used
793  * (**) only one unitmask enabled at a time
794  */
795 
796 static struct event_constraint amd_f15_PMC0  = EVENT_CONSTRAINT(0, 0x01, 0);
797 static struct event_constraint amd_f15_PMC20 = EVENT_CONSTRAINT(0, 0x07, 0);
798 static struct event_constraint amd_f15_PMC3  = EVENT_CONSTRAINT(0, 0x08, 0);
799 static struct event_constraint amd_f15_PMC30 = EVENT_CONSTRAINT_OVERLAP(0, 0x09, 0);
800 static struct event_constraint amd_f15_PMC50 = EVENT_CONSTRAINT(0, 0x3F, 0);
801 static struct event_constraint amd_f15_PMC53 = EVENT_CONSTRAINT(0, 0x38, 0);
802 
803 static struct event_constraint *
804 amd_get_event_constraints_f15h(struct cpu_hw_events *cpuc, int idx,
805 			       struct perf_event *event)
806 {
807 	struct hw_perf_event *hwc = &event->hw;
808 	unsigned int event_code = amd_get_event_code(hwc);
809 
810 	switch (event_code & AMD_EVENT_TYPE_MASK) {
811 	case AMD_EVENT_FP:
812 		switch (event_code) {
813 		case 0x000:
814 			if (!(hwc->config & 0x0000F000ULL))
815 				break;
816 			if (!(hwc->config & 0x00000F00ULL))
817 				break;
818 			return &amd_f15_PMC3;
819 		case 0x004:
820 			if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
821 				break;
822 			return &amd_f15_PMC3;
823 		case 0x003:
824 		case 0x00B:
825 		case 0x00D:
826 			return &amd_f15_PMC3;
827 		}
828 		return &amd_f15_PMC53;
829 	case AMD_EVENT_LS:
830 	case AMD_EVENT_DC:
831 	case AMD_EVENT_EX_LS:
832 		switch (event_code) {
833 		case 0x023:
834 		case 0x043:
835 		case 0x045:
836 		case 0x046:
837 		case 0x054:
838 		case 0x055:
839 			return &amd_f15_PMC20;
840 		case 0x02D:
841 			return &amd_f15_PMC3;
842 		case 0x02E:
843 			return &amd_f15_PMC30;
844 		case 0x031:
845 			if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
846 				return &amd_f15_PMC20;
847 			return &emptyconstraint;
848 		case 0x1C0:
849 			return &amd_f15_PMC53;
850 		default:
851 			return &amd_f15_PMC50;
852 		}
853 	case AMD_EVENT_CU:
854 	case AMD_EVENT_IC_DE:
855 	case AMD_EVENT_DE:
856 		switch (event_code) {
857 		case 0x08F:
858 		case 0x187:
859 		case 0x188:
860 			return &amd_f15_PMC0;
861 		case 0x0DB ... 0x0DF:
862 		case 0x1D6:
863 		case 0x1D8:
864 			return &amd_f15_PMC50;
865 		default:
866 			return &amd_f15_PMC20;
867 		}
868 	case AMD_EVENT_NB:
869 		/* moved to uncore.c */
870 		return &emptyconstraint;
871 	default:
872 		return &emptyconstraint;
873 	}
874 }
875 
876 static struct event_constraint pair_constraint;
877 
878 static struct event_constraint *
879 amd_get_event_constraints_f17h(struct cpu_hw_events *cpuc, int idx,
880 			       struct perf_event *event)
881 {
882 	struct hw_perf_event *hwc = &event->hw;
883 
884 	if (amd_is_pair_event_code(hwc))
885 		return &pair_constraint;
886 
887 	return &unconstrained;
888 }
889 
890 static void amd_put_event_constraints_f17h(struct cpu_hw_events *cpuc,
891 					   struct perf_event *event)
892 {
893 	struct hw_perf_event *hwc = &event->hw;
894 
895 	if (is_counter_pair(hwc))
896 		--cpuc->n_pair;
897 }
898 
899 static ssize_t amd_event_sysfs_show(char *page, u64 config)
900 {
901 	u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT) |
902 		    (config & AMD64_EVENTSEL_EVENT) >> 24;
903 
904 	return x86_event_sysfs_show(page, config, event);
905 }
906 
907 static __initconst const struct x86_pmu amd_pmu = {
908 	.name			= "AMD",
909 	.handle_irq		= amd_pmu_handle_irq,
910 	.disable_all		= amd_pmu_disable_all,
911 	.enable_all		= x86_pmu_enable_all,
912 	.enable			= x86_pmu_enable_event,
913 	.disable		= amd_pmu_disable_event,
914 	.hw_config		= amd_pmu_hw_config,
915 	.schedule_events	= x86_schedule_events,
916 	.eventsel		= MSR_K7_EVNTSEL0,
917 	.perfctr		= MSR_K7_PERFCTR0,
918 	.addr_offset            = amd_pmu_addr_offset,
919 	.event_map		= amd_pmu_event_map,
920 	.max_events		= ARRAY_SIZE(amd_perfmon_event_map),
921 	.num_counters		= AMD64_NUM_COUNTERS,
922 	.cntval_bits		= 48,
923 	.cntval_mask		= (1ULL << 48) - 1,
924 	.apic			= 1,
925 	/* use highest bit to detect overflow */
926 	.max_period		= (1ULL << 47) - 1,
927 	.get_event_constraints	= amd_get_event_constraints,
928 	.put_event_constraints	= amd_put_event_constraints,
929 
930 	.format_attrs		= amd_format_attr,
931 	.events_sysfs_show	= amd_event_sysfs_show,
932 
933 	.cpu_prepare		= amd_pmu_cpu_prepare,
934 	.cpu_starting		= amd_pmu_cpu_starting,
935 	.cpu_dead		= amd_pmu_cpu_dead,
936 
937 	.amd_nb_constraints	= 1,
938 };
939 
940 static int __init amd_core_pmu_init(void)
941 {
942 	u64 even_ctr_mask = 0ULL;
943 	int i;
944 
945 	if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
946 		return 0;
947 
948 	/* Avoid calculating the value each time in the NMI handler */
949 	perf_nmi_window = msecs_to_jiffies(100);
950 
951 	/*
952 	 * If core performance counter extensions exists, we must use
953 	 * MSR_F15H_PERF_CTL/MSR_F15H_PERF_CTR msrs. See also
954 	 * amd_pmu_addr_offset().
955 	 */
956 	x86_pmu.eventsel	= MSR_F15H_PERF_CTL;
957 	x86_pmu.perfctr		= MSR_F15H_PERF_CTR;
958 	x86_pmu.num_counters	= AMD64_NUM_COUNTERS_CORE;
959 	/*
960 	 * AMD Core perfctr has separate MSRs for the NB events, see
961 	 * the amd/uncore.c driver.
962 	 */
963 	x86_pmu.amd_nb_constraints = 0;
964 
965 	if (boot_cpu_data.x86 == 0x15) {
966 		pr_cont("Fam15h ");
967 		x86_pmu.get_event_constraints = amd_get_event_constraints_f15h;
968 	}
969 	if (boot_cpu_data.x86 >= 0x17) {
970 		pr_cont("Fam17h+ ");
971 		/*
972 		 * Family 17h and compatibles have constraints for Large
973 		 * Increment per Cycle events: they may only be assigned an
974 		 * even numbered counter that has a consecutive adjacent odd
975 		 * numbered counter following it.
976 		 */
977 		for (i = 0; i < x86_pmu.num_counters - 1; i += 2)
978 			even_ctr_mask |= 1 << i;
979 
980 		pair_constraint = (struct event_constraint)
981 				    __EVENT_CONSTRAINT(0, even_ctr_mask, 0,
982 				    x86_pmu.num_counters / 2, 0,
983 				    PERF_X86_EVENT_PAIR);
984 
985 		x86_pmu.get_event_constraints = amd_get_event_constraints_f17h;
986 		x86_pmu.put_event_constraints = amd_put_event_constraints_f17h;
987 		x86_pmu.perf_ctr_pair_en = AMD_MERGE_EVENT_ENABLE;
988 		x86_pmu.flags |= PMU_FL_PAIR;
989 	}
990 
991 	pr_cont("core perfctr, ");
992 	return 0;
993 }
994 
995 __init int amd_pmu_init(void)
996 {
997 	int ret;
998 
999 	/* Performance-monitoring supported from K7 and later: */
1000 	if (boot_cpu_data.x86 < 6)
1001 		return -ENODEV;
1002 
1003 	x86_pmu = amd_pmu;
1004 
1005 	ret = amd_core_pmu_init();
1006 	if (ret)
1007 		return ret;
1008 
1009 	if (num_possible_cpus() == 1) {
1010 		/*
1011 		 * No point in allocating data structures to serialize
1012 		 * against other CPUs, when there is only the one CPU.
1013 		 */
1014 		x86_pmu.amd_nb_constraints = 0;
1015 	}
1016 
1017 	if (boot_cpu_data.x86 >= 0x17)
1018 		memcpy(hw_cache_event_ids, amd_hw_cache_event_ids_f17h, sizeof(hw_cache_event_ids));
1019 	else
1020 		memcpy(hw_cache_event_ids, amd_hw_cache_event_ids, sizeof(hw_cache_event_ids));
1021 
1022 	return 0;
1023 }
1024 
1025 void amd_pmu_enable_virt(void)
1026 {
1027 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1028 
1029 	cpuc->perf_ctr_virt_mask = 0;
1030 
1031 	/* Reload all events */
1032 	amd_pmu_disable_all();
1033 	x86_pmu_enable_all(0);
1034 }
1035 EXPORT_SYMBOL_GPL(amd_pmu_enable_virt);
1036 
1037 void amd_pmu_disable_virt(void)
1038 {
1039 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1040 
1041 	/*
1042 	 * We only mask out the Host-only bit so that host-only counting works
1043 	 * when SVM is disabled. If someone sets up a guest-only counter when
1044 	 * SVM is disabled the Guest-only bits still gets set and the counter
1045 	 * will not count anything.
1046 	 */
1047 	cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
1048 
1049 	/* Reload all events */
1050 	amd_pmu_disable_all();
1051 	x86_pmu_enable_all(0);
1052 }
1053 EXPORT_SYMBOL_GPL(amd_pmu_disable_virt);
1054