xref: /openbmc/linux/arch/x86/events/amd/core.c (revision b285d2ae)
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_CACHE_MISSES]		= 0x0964,
254 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x00c2,
255 	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x00c3,
256 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= 0x0287,
257 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= 0x0187,
258 };
259 
260 static u64 amd_pmu_event_map(int hw_event)
261 {
262 	if (boot_cpu_data.x86 >= 0x17)
263 		return amd_f17h_perfmon_event_map[hw_event];
264 
265 	return amd_perfmon_event_map[hw_event];
266 }
267 
268 /*
269  * Previously calculated offsets
270  */
271 static unsigned int event_offsets[X86_PMC_IDX_MAX] __read_mostly;
272 static unsigned int count_offsets[X86_PMC_IDX_MAX] __read_mostly;
273 
274 /*
275  * Legacy CPUs:
276  *   4 counters starting at 0xc0010000 each offset by 1
277  *
278  * CPUs with core performance counter extensions:
279  *   6 counters starting at 0xc0010200 each offset by 2
280  */
281 static inline int amd_pmu_addr_offset(int index, bool eventsel)
282 {
283 	int offset;
284 
285 	if (!index)
286 		return index;
287 
288 	if (eventsel)
289 		offset = event_offsets[index];
290 	else
291 		offset = count_offsets[index];
292 
293 	if (offset)
294 		return offset;
295 
296 	if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
297 		offset = index;
298 	else
299 		offset = index << 1;
300 
301 	if (eventsel)
302 		event_offsets[index] = offset;
303 	else
304 		count_offsets[index] = offset;
305 
306 	return offset;
307 }
308 
309 /*
310  * AMD64 events are detected based on their event codes.
311  */
312 static inline unsigned int amd_get_event_code(struct hw_perf_event *hwc)
313 {
314 	return ((hwc->config >> 24) & 0x0f00) | (hwc->config & 0x00ff);
315 }
316 
317 static inline bool amd_is_pair_event_code(struct hw_perf_event *hwc)
318 {
319 	if (!(x86_pmu.flags & PMU_FL_PAIR))
320 		return false;
321 
322 	switch (amd_get_event_code(hwc)) {
323 	case 0x003:	return true;	/* Retired SSE/AVX FLOPs */
324 	default:	return false;
325 	}
326 }
327 
328 static int amd_core_hw_config(struct perf_event *event)
329 {
330 	if (event->attr.exclude_host && event->attr.exclude_guest)
331 		/*
332 		 * When HO == GO == 1 the hardware treats that as GO == HO == 0
333 		 * and will count in both modes. We don't want to count in that
334 		 * case so we emulate no-counting by setting US = OS = 0.
335 		 */
336 		event->hw.config &= ~(ARCH_PERFMON_EVENTSEL_USR |
337 				      ARCH_PERFMON_EVENTSEL_OS);
338 	else if (event->attr.exclude_host)
339 		event->hw.config |= AMD64_EVENTSEL_GUESTONLY;
340 	else if (event->attr.exclude_guest)
341 		event->hw.config |= AMD64_EVENTSEL_HOSTONLY;
342 
343 	if ((x86_pmu.flags & PMU_FL_PAIR) && amd_is_pair_event_code(&event->hw))
344 		event->hw.flags |= PERF_X86_EVENT_PAIR;
345 
346 	return 0;
347 }
348 
349 static inline int amd_is_nb_event(struct hw_perf_event *hwc)
350 {
351 	return (hwc->config & 0xe0) == 0xe0;
352 }
353 
354 static inline int amd_has_nb(struct cpu_hw_events *cpuc)
355 {
356 	struct amd_nb *nb = cpuc->amd_nb;
357 
358 	return nb && nb->nb_id != -1;
359 }
360 
361 static int amd_pmu_hw_config(struct perf_event *event)
362 {
363 	int ret;
364 
365 	/* pass precise event sampling to ibs: */
366 	if (event->attr.precise_ip && get_ibs_caps())
367 		return -ENOENT;
368 
369 	if (has_branch_stack(event))
370 		return -EOPNOTSUPP;
371 
372 	ret = x86_pmu_hw_config(event);
373 	if (ret)
374 		return ret;
375 
376 	if (event->attr.type == PERF_TYPE_RAW)
377 		event->hw.config |= event->attr.config & AMD64_RAW_EVENT_MASK;
378 
379 	return amd_core_hw_config(event);
380 }
381 
382 static void __amd_put_nb_event_constraints(struct cpu_hw_events *cpuc,
383 					   struct perf_event *event)
384 {
385 	struct amd_nb *nb = cpuc->amd_nb;
386 	int i;
387 
388 	/*
389 	 * need to scan whole list because event may not have
390 	 * been assigned during scheduling
391 	 *
392 	 * no race condition possible because event can only
393 	 * be removed on one CPU at a time AND PMU is disabled
394 	 * when we come here
395 	 */
396 	for (i = 0; i < x86_pmu.num_counters; i++) {
397 		if (cmpxchg(nb->owners + i, event, NULL) == event)
398 			break;
399 	}
400 }
401 
402  /*
403   * AMD64 NorthBridge events need special treatment because
404   * counter access needs to be synchronized across all cores
405   * of a package. Refer to BKDG section 3.12
406   *
407   * NB events are events measuring L3 cache, Hypertransport
408   * traffic. They are identified by an event code >= 0xe00.
409   * They measure events on the NorthBride which is shared
410   * by all cores on a package. NB events are counted on a
411   * shared set of counters. When a NB event is programmed
412   * in a counter, the data actually comes from a shared
413   * counter. Thus, access to those counters needs to be
414   * synchronized.
415   *
416   * We implement the synchronization such that no two cores
417   * can be measuring NB events using the same counters. Thus,
418   * we maintain a per-NB allocation table. The available slot
419   * is propagated using the event_constraint structure.
420   *
421   * We provide only one choice for each NB event based on
422   * the fact that only NB events have restrictions. Consequently,
423   * if a counter is available, there is a guarantee the NB event
424   * will be assigned to it. If no slot is available, an empty
425   * constraint is returned and scheduling will eventually fail
426   * for this event.
427   *
428   * Note that all cores attached the same NB compete for the same
429   * counters to host NB events, this is why we use atomic ops. Some
430   * multi-chip CPUs may have more than one NB.
431   *
432   * Given that resources are allocated (cmpxchg), they must be
433   * eventually freed for others to use. This is accomplished by
434   * calling __amd_put_nb_event_constraints()
435   *
436   * Non NB events are not impacted by this restriction.
437   */
438 static struct event_constraint *
439 __amd_get_nb_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
440 			       struct event_constraint *c)
441 {
442 	struct hw_perf_event *hwc = &event->hw;
443 	struct amd_nb *nb = cpuc->amd_nb;
444 	struct perf_event *old;
445 	int idx, new = -1;
446 
447 	if (!c)
448 		c = &unconstrained;
449 
450 	if (cpuc->is_fake)
451 		return c;
452 
453 	/*
454 	 * detect if already present, if so reuse
455 	 *
456 	 * cannot merge with actual allocation
457 	 * because of possible holes
458 	 *
459 	 * event can already be present yet not assigned (in hwc->idx)
460 	 * because of successive calls to x86_schedule_events() from
461 	 * hw_perf_group_sched_in() without hw_perf_enable()
462 	 */
463 	for_each_set_bit(idx, c->idxmsk, x86_pmu.num_counters) {
464 		if (new == -1 || hwc->idx == idx)
465 			/* assign free slot, prefer hwc->idx */
466 			old = cmpxchg(nb->owners + idx, NULL, event);
467 		else if (nb->owners[idx] == event)
468 			/* event already present */
469 			old = event;
470 		else
471 			continue;
472 
473 		if (old && old != event)
474 			continue;
475 
476 		/* reassign to this slot */
477 		if (new != -1)
478 			cmpxchg(nb->owners + new, event, NULL);
479 		new = idx;
480 
481 		/* already present, reuse */
482 		if (old == event)
483 			break;
484 	}
485 
486 	if (new == -1)
487 		return &emptyconstraint;
488 
489 	return &nb->event_constraints[new];
490 }
491 
492 static struct amd_nb *amd_alloc_nb(int cpu)
493 {
494 	struct amd_nb *nb;
495 	int i;
496 
497 	nb = kzalloc_node(sizeof(struct amd_nb), GFP_KERNEL, cpu_to_node(cpu));
498 	if (!nb)
499 		return NULL;
500 
501 	nb->nb_id = -1;
502 
503 	/*
504 	 * initialize all possible NB constraints
505 	 */
506 	for (i = 0; i < x86_pmu.num_counters; i++) {
507 		__set_bit(i, nb->event_constraints[i].idxmsk);
508 		nb->event_constraints[i].weight = 1;
509 	}
510 	return nb;
511 }
512 
513 static int amd_pmu_cpu_prepare(int cpu)
514 {
515 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
516 
517 	WARN_ON_ONCE(cpuc->amd_nb);
518 
519 	if (!x86_pmu.amd_nb_constraints)
520 		return 0;
521 
522 	cpuc->amd_nb = amd_alloc_nb(cpu);
523 	if (!cpuc->amd_nb)
524 		return -ENOMEM;
525 
526 	return 0;
527 }
528 
529 static void amd_pmu_cpu_starting(int cpu)
530 {
531 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
532 	void **onln = &cpuc->kfree_on_online[X86_PERF_KFREE_SHARED];
533 	struct amd_nb *nb;
534 	int i, nb_id;
535 
536 	cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
537 
538 	if (!x86_pmu.amd_nb_constraints)
539 		return;
540 
541 	nb_id = amd_get_nb_id(cpu);
542 	WARN_ON_ONCE(nb_id == BAD_APICID);
543 
544 	for_each_online_cpu(i) {
545 		nb = per_cpu(cpu_hw_events, i).amd_nb;
546 		if (WARN_ON_ONCE(!nb))
547 			continue;
548 
549 		if (nb->nb_id == nb_id) {
550 			*onln = cpuc->amd_nb;
551 			cpuc->amd_nb = nb;
552 			break;
553 		}
554 	}
555 
556 	cpuc->amd_nb->nb_id = nb_id;
557 	cpuc->amd_nb->refcnt++;
558 }
559 
560 static void amd_pmu_cpu_dead(int cpu)
561 {
562 	struct cpu_hw_events *cpuhw;
563 
564 	if (!x86_pmu.amd_nb_constraints)
565 		return;
566 
567 	cpuhw = &per_cpu(cpu_hw_events, cpu);
568 
569 	if (cpuhw->amd_nb) {
570 		struct amd_nb *nb = cpuhw->amd_nb;
571 
572 		if (nb->nb_id == -1 || --nb->refcnt == 0)
573 			kfree(nb);
574 
575 		cpuhw->amd_nb = NULL;
576 	}
577 }
578 
579 /*
580  * When a PMC counter overflows, an NMI is used to process the event and
581  * reset the counter. NMI latency can result in the counter being updated
582  * before the NMI can run, which can result in what appear to be spurious
583  * NMIs. This function is intended to wait for the NMI to run and reset
584  * the counter to avoid possible unhandled NMI messages.
585  */
586 #define OVERFLOW_WAIT_COUNT	50
587 
588 static void amd_pmu_wait_on_overflow(int idx)
589 {
590 	unsigned int i;
591 	u64 counter;
592 
593 	/*
594 	 * Wait for the counter to be reset if it has overflowed. This loop
595 	 * should exit very, very quickly, but just in case, don't wait
596 	 * forever...
597 	 */
598 	for (i = 0; i < OVERFLOW_WAIT_COUNT; i++) {
599 		rdmsrl(x86_pmu_event_addr(idx), counter);
600 		if (counter & (1ULL << (x86_pmu.cntval_bits - 1)))
601 			break;
602 
603 		/* Might be in IRQ context, so can't sleep */
604 		udelay(1);
605 	}
606 }
607 
608 static void amd_pmu_disable_all(void)
609 {
610 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
611 	int idx;
612 
613 	x86_pmu_disable_all();
614 
615 	/*
616 	 * This shouldn't be called from NMI context, but add a safeguard here
617 	 * to return, since if we're in NMI context we can't wait for an NMI
618 	 * to reset an overflowed counter value.
619 	 */
620 	if (in_nmi())
621 		return;
622 
623 	/*
624 	 * Check each counter for overflow and wait for it to be reset by the
625 	 * NMI if it has overflowed. This relies on the fact that all active
626 	 * counters are always enabled when this function is caled and
627 	 * ARCH_PERFMON_EVENTSEL_INT is always set.
628 	 */
629 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
630 		if (!test_bit(idx, cpuc->active_mask))
631 			continue;
632 
633 		amd_pmu_wait_on_overflow(idx);
634 	}
635 }
636 
637 static void amd_pmu_disable_event(struct perf_event *event)
638 {
639 	x86_pmu_disable_event(event);
640 
641 	/*
642 	 * This can be called from NMI context (via x86_pmu_stop). The counter
643 	 * may have overflowed, but either way, we'll never see it get reset
644 	 * by the NMI if we're already in the NMI. And the NMI latency support
645 	 * below will take care of any pending NMI that might have been
646 	 * generated by the overflow.
647 	 */
648 	if (in_nmi())
649 		return;
650 
651 	amd_pmu_wait_on_overflow(event->hw.idx);
652 }
653 
654 /*
655  * Because of NMI latency, if multiple PMC counters are active or other sources
656  * of NMIs are received, the perf NMI handler can handle one or more overflowed
657  * PMC counters outside of the NMI associated with the PMC overflow. If the NMI
658  * doesn't arrive at the LAPIC in time to become a pending NMI, then the kernel
659  * back-to-back NMI support won't be active. This PMC handler needs to take into
660  * account that this can occur, otherwise this could result in unknown NMI
661  * messages being issued. Examples of this is PMC overflow while in the NMI
662  * handler when multiple PMCs are active or PMC overflow while handling some
663  * other source of an NMI.
664  *
665  * Attempt to mitigate this by creating an NMI window in which un-handled NMIs
666  * received during this window will be claimed. This prevents extending the
667  * window past when it is possible that latent NMIs should be received. The
668  * per-CPU perf_nmi_tstamp will be set to the window end time whenever perf has
669  * handled a counter. When an un-handled NMI is received, it will be claimed
670  * only if arriving within that window.
671  */
672 static int amd_pmu_handle_irq(struct pt_regs *regs)
673 {
674 	int handled;
675 
676 	/* Process any counter overflows */
677 	handled = x86_pmu_handle_irq(regs);
678 
679 	/*
680 	 * If a counter was handled, record a timestamp such that un-handled
681 	 * NMIs will be claimed if arriving within that window.
682 	 */
683 	if (handled) {
684 		this_cpu_write(perf_nmi_tstamp, jiffies + perf_nmi_window);
685 
686 		return handled;
687 	}
688 
689 	if (time_after(jiffies, this_cpu_read(perf_nmi_tstamp)))
690 		return NMI_DONE;
691 
692 	return NMI_HANDLED;
693 }
694 
695 static struct event_constraint *
696 amd_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
697 			  struct perf_event *event)
698 {
699 	/*
700 	 * if not NB event or no NB, then no constraints
701 	 */
702 	if (!(amd_has_nb(cpuc) && amd_is_nb_event(&event->hw)))
703 		return &unconstrained;
704 
705 	return __amd_get_nb_event_constraints(cpuc, event, NULL);
706 }
707 
708 static void amd_put_event_constraints(struct cpu_hw_events *cpuc,
709 				      struct perf_event *event)
710 {
711 	if (amd_has_nb(cpuc) && amd_is_nb_event(&event->hw))
712 		__amd_put_nb_event_constraints(cpuc, event);
713 }
714 
715 PMU_FORMAT_ATTR(event,	"config:0-7,32-35");
716 PMU_FORMAT_ATTR(umask,	"config:8-15"	);
717 PMU_FORMAT_ATTR(edge,	"config:18"	);
718 PMU_FORMAT_ATTR(inv,	"config:23"	);
719 PMU_FORMAT_ATTR(cmask,	"config:24-31"	);
720 
721 static struct attribute *amd_format_attr[] = {
722 	&format_attr_event.attr,
723 	&format_attr_umask.attr,
724 	&format_attr_edge.attr,
725 	&format_attr_inv.attr,
726 	&format_attr_cmask.attr,
727 	NULL,
728 };
729 
730 /* AMD Family 15h */
731 
732 #define AMD_EVENT_TYPE_MASK	0x000000F0ULL
733 
734 #define AMD_EVENT_FP		0x00000000ULL ... 0x00000010ULL
735 #define AMD_EVENT_LS		0x00000020ULL ... 0x00000030ULL
736 #define AMD_EVENT_DC		0x00000040ULL ... 0x00000050ULL
737 #define AMD_EVENT_CU		0x00000060ULL ... 0x00000070ULL
738 #define AMD_EVENT_IC_DE		0x00000080ULL ... 0x00000090ULL
739 #define AMD_EVENT_EX_LS		0x000000C0ULL
740 #define AMD_EVENT_DE		0x000000D0ULL
741 #define AMD_EVENT_NB		0x000000E0ULL ... 0x000000F0ULL
742 
743 /*
744  * AMD family 15h event code/PMC mappings:
745  *
746  * type = event_code & 0x0F0:
747  *
748  * 0x000	FP	PERF_CTL[5:3]
749  * 0x010	FP	PERF_CTL[5:3]
750  * 0x020	LS	PERF_CTL[5:0]
751  * 0x030	LS	PERF_CTL[5:0]
752  * 0x040	DC	PERF_CTL[5:0]
753  * 0x050	DC	PERF_CTL[5:0]
754  * 0x060	CU	PERF_CTL[2:0]
755  * 0x070	CU	PERF_CTL[2:0]
756  * 0x080	IC/DE	PERF_CTL[2:0]
757  * 0x090	IC/DE	PERF_CTL[2:0]
758  * 0x0A0	---
759  * 0x0B0	---
760  * 0x0C0	EX/LS	PERF_CTL[5:0]
761  * 0x0D0	DE	PERF_CTL[2:0]
762  * 0x0E0	NB	NB_PERF_CTL[3:0]
763  * 0x0F0	NB	NB_PERF_CTL[3:0]
764  *
765  * Exceptions:
766  *
767  * 0x000	FP	PERF_CTL[3], PERF_CTL[5:3] (*)
768  * 0x003	FP	PERF_CTL[3]
769  * 0x004	FP	PERF_CTL[3], PERF_CTL[5:3] (*)
770  * 0x00B	FP	PERF_CTL[3]
771  * 0x00D	FP	PERF_CTL[3]
772  * 0x023	DE	PERF_CTL[2:0]
773  * 0x02D	LS	PERF_CTL[3]
774  * 0x02E	LS	PERF_CTL[3,0]
775  * 0x031	LS	PERF_CTL[2:0] (**)
776  * 0x043	CU	PERF_CTL[2:0]
777  * 0x045	CU	PERF_CTL[2:0]
778  * 0x046	CU	PERF_CTL[2:0]
779  * 0x054	CU	PERF_CTL[2:0]
780  * 0x055	CU	PERF_CTL[2:0]
781  * 0x08F	IC	PERF_CTL[0]
782  * 0x187	DE	PERF_CTL[0]
783  * 0x188	DE	PERF_CTL[0]
784  * 0x0DB	EX	PERF_CTL[5:0]
785  * 0x0DC	LS	PERF_CTL[5:0]
786  * 0x0DD	LS	PERF_CTL[5:0]
787  * 0x0DE	LS	PERF_CTL[5:0]
788  * 0x0DF	LS	PERF_CTL[5:0]
789  * 0x1C0	EX	PERF_CTL[5:3]
790  * 0x1D6	EX	PERF_CTL[5:0]
791  * 0x1D8	EX	PERF_CTL[5:0]
792  *
793  * (*)  depending on the umask all FPU counters may be used
794  * (**) only one unitmask enabled at a time
795  */
796 
797 static struct event_constraint amd_f15_PMC0  = EVENT_CONSTRAINT(0, 0x01, 0);
798 static struct event_constraint amd_f15_PMC20 = EVENT_CONSTRAINT(0, 0x07, 0);
799 static struct event_constraint amd_f15_PMC3  = EVENT_CONSTRAINT(0, 0x08, 0);
800 static struct event_constraint amd_f15_PMC30 = EVENT_CONSTRAINT_OVERLAP(0, 0x09, 0);
801 static struct event_constraint amd_f15_PMC50 = EVENT_CONSTRAINT(0, 0x3F, 0);
802 static struct event_constraint amd_f15_PMC53 = EVENT_CONSTRAINT(0, 0x38, 0);
803 
804 static struct event_constraint *
805 amd_get_event_constraints_f15h(struct cpu_hw_events *cpuc, int idx,
806 			       struct perf_event *event)
807 {
808 	struct hw_perf_event *hwc = &event->hw;
809 	unsigned int event_code = amd_get_event_code(hwc);
810 
811 	switch (event_code & AMD_EVENT_TYPE_MASK) {
812 	case AMD_EVENT_FP:
813 		switch (event_code) {
814 		case 0x000:
815 			if (!(hwc->config & 0x0000F000ULL))
816 				break;
817 			if (!(hwc->config & 0x00000F00ULL))
818 				break;
819 			return &amd_f15_PMC3;
820 		case 0x004:
821 			if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
822 				break;
823 			return &amd_f15_PMC3;
824 		case 0x003:
825 		case 0x00B:
826 		case 0x00D:
827 			return &amd_f15_PMC3;
828 		}
829 		return &amd_f15_PMC53;
830 	case AMD_EVENT_LS:
831 	case AMD_EVENT_DC:
832 	case AMD_EVENT_EX_LS:
833 		switch (event_code) {
834 		case 0x023:
835 		case 0x043:
836 		case 0x045:
837 		case 0x046:
838 		case 0x054:
839 		case 0x055:
840 			return &amd_f15_PMC20;
841 		case 0x02D:
842 			return &amd_f15_PMC3;
843 		case 0x02E:
844 			return &amd_f15_PMC30;
845 		case 0x031:
846 			if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
847 				return &amd_f15_PMC20;
848 			return &emptyconstraint;
849 		case 0x1C0:
850 			return &amd_f15_PMC53;
851 		default:
852 			return &amd_f15_PMC50;
853 		}
854 	case AMD_EVENT_CU:
855 	case AMD_EVENT_IC_DE:
856 	case AMD_EVENT_DE:
857 		switch (event_code) {
858 		case 0x08F:
859 		case 0x187:
860 		case 0x188:
861 			return &amd_f15_PMC0;
862 		case 0x0DB ... 0x0DF:
863 		case 0x1D6:
864 		case 0x1D8:
865 			return &amd_f15_PMC50;
866 		default:
867 			return &amd_f15_PMC20;
868 		}
869 	case AMD_EVENT_NB:
870 		/* moved to uncore.c */
871 		return &emptyconstraint;
872 	default:
873 		return &emptyconstraint;
874 	}
875 }
876 
877 static struct event_constraint pair_constraint;
878 
879 static struct event_constraint *
880 amd_get_event_constraints_f17h(struct cpu_hw_events *cpuc, int idx,
881 			       struct perf_event *event)
882 {
883 	struct hw_perf_event *hwc = &event->hw;
884 
885 	if (amd_is_pair_event_code(hwc))
886 		return &pair_constraint;
887 
888 	return &unconstrained;
889 }
890 
891 static void amd_put_event_constraints_f17h(struct cpu_hw_events *cpuc,
892 					   struct perf_event *event)
893 {
894 	struct hw_perf_event *hwc = &event->hw;
895 
896 	if (is_counter_pair(hwc))
897 		--cpuc->n_pair;
898 }
899 
900 static ssize_t amd_event_sysfs_show(char *page, u64 config)
901 {
902 	u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT) |
903 		    (config & AMD64_EVENTSEL_EVENT) >> 24;
904 
905 	return x86_event_sysfs_show(page, config, event);
906 }
907 
908 static __initconst const struct x86_pmu amd_pmu = {
909 	.name			= "AMD",
910 	.handle_irq		= amd_pmu_handle_irq,
911 	.disable_all		= amd_pmu_disable_all,
912 	.enable_all		= x86_pmu_enable_all,
913 	.enable			= x86_pmu_enable_event,
914 	.disable		= amd_pmu_disable_event,
915 	.hw_config		= amd_pmu_hw_config,
916 	.schedule_events	= x86_schedule_events,
917 	.eventsel		= MSR_K7_EVNTSEL0,
918 	.perfctr		= MSR_K7_PERFCTR0,
919 	.addr_offset            = amd_pmu_addr_offset,
920 	.event_map		= amd_pmu_event_map,
921 	.max_events		= ARRAY_SIZE(amd_perfmon_event_map),
922 	.num_counters		= AMD64_NUM_COUNTERS,
923 	.cntval_bits		= 48,
924 	.cntval_mask		= (1ULL << 48) - 1,
925 	.apic			= 1,
926 	/* use highest bit to detect overflow */
927 	.max_period		= (1ULL << 47) - 1,
928 	.get_event_constraints	= amd_get_event_constraints,
929 	.put_event_constraints	= amd_put_event_constraints,
930 
931 	.format_attrs		= amd_format_attr,
932 	.events_sysfs_show	= amd_event_sysfs_show,
933 
934 	.cpu_prepare		= amd_pmu_cpu_prepare,
935 	.cpu_starting		= amd_pmu_cpu_starting,
936 	.cpu_dead		= amd_pmu_cpu_dead,
937 
938 	.amd_nb_constraints	= 1,
939 };
940 
941 static int __init amd_core_pmu_init(void)
942 {
943 	u64 even_ctr_mask = 0ULL;
944 	int i;
945 
946 	if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
947 		return 0;
948 
949 	/* Avoid calculating the value each time in the NMI handler */
950 	perf_nmi_window = msecs_to_jiffies(100);
951 
952 	/*
953 	 * If core performance counter extensions exists, we must use
954 	 * MSR_F15H_PERF_CTL/MSR_F15H_PERF_CTR msrs. See also
955 	 * amd_pmu_addr_offset().
956 	 */
957 	x86_pmu.eventsel	= MSR_F15H_PERF_CTL;
958 	x86_pmu.perfctr		= MSR_F15H_PERF_CTR;
959 	x86_pmu.num_counters	= AMD64_NUM_COUNTERS_CORE;
960 	/*
961 	 * AMD Core perfctr has separate MSRs for the NB events, see
962 	 * the amd/uncore.c driver.
963 	 */
964 	x86_pmu.amd_nb_constraints = 0;
965 
966 	if (boot_cpu_data.x86 == 0x15) {
967 		pr_cont("Fam15h ");
968 		x86_pmu.get_event_constraints = amd_get_event_constraints_f15h;
969 	}
970 	if (boot_cpu_data.x86 >= 0x17) {
971 		pr_cont("Fam17h+ ");
972 		/*
973 		 * Family 17h and compatibles have constraints for Large
974 		 * Increment per Cycle events: they may only be assigned an
975 		 * even numbered counter that has a consecutive adjacent odd
976 		 * numbered counter following it.
977 		 */
978 		for (i = 0; i < x86_pmu.num_counters - 1; i += 2)
979 			even_ctr_mask |= 1 << i;
980 
981 		pair_constraint = (struct event_constraint)
982 				    __EVENT_CONSTRAINT(0, even_ctr_mask, 0,
983 				    x86_pmu.num_counters / 2, 0,
984 				    PERF_X86_EVENT_PAIR);
985 
986 		x86_pmu.get_event_constraints = amd_get_event_constraints_f17h;
987 		x86_pmu.put_event_constraints = amd_put_event_constraints_f17h;
988 		x86_pmu.perf_ctr_pair_en = AMD_MERGE_EVENT_ENABLE;
989 		x86_pmu.flags |= PMU_FL_PAIR;
990 	}
991 
992 	pr_cont("core perfctr, ");
993 	return 0;
994 }
995 
996 __init int amd_pmu_init(void)
997 {
998 	int ret;
999 
1000 	/* Performance-monitoring supported from K7 and later: */
1001 	if (boot_cpu_data.x86 < 6)
1002 		return -ENODEV;
1003 
1004 	x86_pmu = amd_pmu;
1005 
1006 	ret = amd_core_pmu_init();
1007 	if (ret)
1008 		return ret;
1009 
1010 	if (num_possible_cpus() == 1) {
1011 		/*
1012 		 * No point in allocating data structures to serialize
1013 		 * against other CPUs, when there is only the one CPU.
1014 		 */
1015 		x86_pmu.amd_nb_constraints = 0;
1016 	}
1017 
1018 	if (boot_cpu_data.x86 >= 0x17)
1019 		memcpy(hw_cache_event_ids, amd_hw_cache_event_ids_f17h, sizeof(hw_cache_event_ids));
1020 	else
1021 		memcpy(hw_cache_event_ids, amd_hw_cache_event_ids, sizeof(hw_cache_event_ids));
1022 
1023 	return 0;
1024 }
1025 
1026 void amd_pmu_enable_virt(void)
1027 {
1028 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1029 
1030 	cpuc->perf_ctr_virt_mask = 0;
1031 
1032 	/* Reload all events */
1033 	amd_pmu_disable_all();
1034 	x86_pmu_enable_all(0);
1035 }
1036 EXPORT_SYMBOL_GPL(amd_pmu_enable_virt);
1037 
1038 void amd_pmu_disable_virt(void)
1039 {
1040 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1041 
1042 	/*
1043 	 * We only mask out the Host-only bit so that host-only counting works
1044 	 * when SVM is disabled. If someone sets up a guest-only counter when
1045 	 * SVM is disabled the Guest-only bits still gets set and the counter
1046 	 * will not count anything.
1047 	 */
1048 	cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
1049 
1050 	/* Reload all events */
1051 	amd_pmu_disable_all();
1052 	x86_pmu_enable_all(0);
1053 }
1054 EXPORT_SYMBOL_GPL(amd_pmu_disable_virt);
1055