xref: /openbmc/linux/arch/x86/events/amd/core.c (revision f519f0be)
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 <asm/apicdef.h>
9 #include <asm/nmi.h>
10 
11 #include "../perf_event.h"
12 
13 static DEFINE_PER_CPU(unsigned int, perf_nmi_counter);
14 
15 static __initconst const u64 amd_hw_cache_event_ids
16 				[PERF_COUNT_HW_CACHE_MAX]
17 				[PERF_COUNT_HW_CACHE_OP_MAX]
18 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
19 {
20  [ C(L1D) ] = {
21 	[ C(OP_READ) ] = {
22 		[ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses        */
23 		[ C(RESULT_MISS)   ] = 0x0141, /* Data Cache Misses          */
24 	},
25 	[ C(OP_WRITE) ] = {
26 		[ C(RESULT_ACCESS) ] = 0,
27 		[ C(RESULT_MISS)   ] = 0,
28 	},
29 	[ C(OP_PREFETCH) ] = {
30 		[ C(RESULT_ACCESS) ] = 0x0267, /* Data Prefetcher :attempts  */
31 		[ C(RESULT_MISS)   ] = 0x0167, /* Data Prefetcher :cancelled */
32 	},
33  },
34  [ C(L1I ) ] = {
35 	[ C(OP_READ) ] = {
36 		[ C(RESULT_ACCESS) ] = 0x0080, /* Instruction cache fetches  */
37 		[ C(RESULT_MISS)   ] = 0x0081, /* Instruction cache misses   */
38 	},
39 	[ C(OP_WRITE) ] = {
40 		[ C(RESULT_ACCESS) ] = -1,
41 		[ C(RESULT_MISS)   ] = -1,
42 	},
43 	[ C(OP_PREFETCH) ] = {
44 		[ C(RESULT_ACCESS) ] = 0x014B, /* Prefetch Instructions :Load */
45 		[ C(RESULT_MISS)   ] = 0,
46 	},
47  },
48  [ C(LL  ) ] = {
49 	[ C(OP_READ) ] = {
50 		[ C(RESULT_ACCESS) ] = 0x037D, /* Requests to L2 Cache :IC+DC */
51 		[ C(RESULT_MISS)   ] = 0x037E, /* L2 Cache Misses : IC+DC     */
52 	},
53 	[ C(OP_WRITE) ] = {
54 		[ C(RESULT_ACCESS) ] = 0x017F, /* L2 Fill/Writeback           */
55 		[ C(RESULT_MISS)   ] = 0,
56 	},
57 	[ C(OP_PREFETCH) ] = {
58 		[ C(RESULT_ACCESS) ] = 0,
59 		[ C(RESULT_MISS)   ] = 0,
60 	},
61  },
62  [ C(DTLB) ] = {
63 	[ C(OP_READ) ] = {
64 		[ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses        */
65 		[ C(RESULT_MISS)   ] = 0x0746, /* L1_DTLB_AND_L2_DLTB_MISS.ALL */
66 	},
67 	[ C(OP_WRITE) ] = {
68 		[ C(RESULT_ACCESS) ] = 0,
69 		[ C(RESULT_MISS)   ] = 0,
70 	},
71 	[ C(OP_PREFETCH) ] = {
72 		[ C(RESULT_ACCESS) ] = 0,
73 		[ C(RESULT_MISS)   ] = 0,
74 	},
75  },
76  [ C(ITLB) ] = {
77 	[ C(OP_READ) ] = {
78 		[ C(RESULT_ACCESS) ] = 0x0080, /* Instruction fecthes        */
79 		[ C(RESULT_MISS)   ] = 0x0385, /* L1_ITLB_AND_L2_ITLB_MISS.ALL */
80 	},
81 	[ C(OP_WRITE) ] = {
82 		[ C(RESULT_ACCESS) ] = -1,
83 		[ C(RESULT_MISS)   ] = -1,
84 	},
85 	[ C(OP_PREFETCH) ] = {
86 		[ C(RESULT_ACCESS) ] = -1,
87 		[ C(RESULT_MISS)   ] = -1,
88 	},
89  },
90  [ C(BPU ) ] = {
91 	[ C(OP_READ) ] = {
92 		[ C(RESULT_ACCESS) ] = 0x00c2, /* Retired Branch Instr.      */
93 		[ C(RESULT_MISS)   ] = 0x00c3, /* Retired Mispredicted BI    */
94 	},
95 	[ C(OP_WRITE) ] = {
96 		[ C(RESULT_ACCESS) ] = -1,
97 		[ C(RESULT_MISS)   ] = -1,
98 	},
99 	[ C(OP_PREFETCH) ] = {
100 		[ C(RESULT_ACCESS) ] = -1,
101 		[ C(RESULT_MISS)   ] = -1,
102 	},
103  },
104  [ C(NODE) ] = {
105 	[ C(OP_READ) ] = {
106 		[ C(RESULT_ACCESS) ] = 0xb8e9, /* CPU Request to Memory, l+r */
107 		[ C(RESULT_MISS)   ] = 0x98e9, /* CPU Request to Memory, r   */
108 	},
109 	[ C(OP_WRITE) ] = {
110 		[ C(RESULT_ACCESS) ] = -1,
111 		[ C(RESULT_MISS)   ] = -1,
112 	},
113 	[ C(OP_PREFETCH) ] = {
114 		[ C(RESULT_ACCESS) ] = -1,
115 		[ C(RESULT_MISS)   ] = -1,
116 	},
117  },
118 };
119 
120 static __initconst const u64 amd_hw_cache_event_ids_f17h
121 				[PERF_COUNT_HW_CACHE_MAX]
122 				[PERF_COUNT_HW_CACHE_OP_MAX]
123 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
124 [C(L1D)] = {
125 	[C(OP_READ)] = {
126 		[C(RESULT_ACCESS)] = 0x0040, /* Data Cache Accesses */
127 		[C(RESULT_MISS)]   = 0xc860, /* L2$ access from DC Miss */
128 	},
129 	[C(OP_WRITE)] = {
130 		[C(RESULT_ACCESS)] = 0,
131 		[C(RESULT_MISS)]   = 0,
132 	},
133 	[C(OP_PREFETCH)] = {
134 		[C(RESULT_ACCESS)] = 0xff5a, /* h/w prefetch DC Fills */
135 		[C(RESULT_MISS)]   = 0,
136 	},
137 },
138 [C(L1I)] = {
139 	[C(OP_READ)] = {
140 		[C(RESULT_ACCESS)] = 0x0080, /* Instruction cache fetches  */
141 		[C(RESULT_MISS)]   = 0x0081, /* Instruction cache misses   */
142 	},
143 	[C(OP_WRITE)] = {
144 		[C(RESULT_ACCESS)] = -1,
145 		[C(RESULT_MISS)]   = -1,
146 	},
147 	[C(OP_PREFETCH)] = {
148 		[C(RESULT_ACCESS)] = 0,
149 		[C(RESULT_MISS)]   = 0,
150 	},
151 },
152 [C(LL)] = {
153 	[C(OP_READ)] = {
154 		[C(RESULT_ACCESS)] = 0,
155 		[C(RESULT_MISS)]   = 0,
156 	},
157 	[C(OP_WRITE)] = {
158 		[C(RESULT_ACCESS)] = 0,
159 		[C(RESULT_MISS)]   = 0,
160 	},
161 	[C(OP_PREFETCH)] = {
162 		[C(RESULT_ACCESS)] = 0,
163 		[C(RESULT_MISS)]   = 0,
164 	},
165 },
166 [C(DTLB)] = {
167 	[C(OP_READ)] = {
168 		[C(RESULT_ACCESS)] = 0xff45, /* All L2 DTLB accesses */
169 		[C(RESULT_MISS)]   = 0xf045, /* L2 DTLB misses (PT walks) */
170 	},
171 	[C(OP_WRITE)] = {
172 		[C(RESULT_ACCESS)] = 0,
173 		[C(RESULT_MISS)]   = 0,
174 	},
175 	[C(OP_PREFETCH)] = {
176 		[C(RESULT_ACCESS)] = 0,
177 		[C(RESULT_MISS)]   = 0,
178 	},
179 },
180 [C(ITLB)] = {
181 	[C(OP_READ)] = {
182 		[C(RESULT_ACCESS)] = 0x0084, /* L1 ITLB misses, L2 ITLB hits */
183 		[C(RESULT_MISS)]   = 0xff85, /* L1 ITLB misses, L2 misses */
184 	},
185 	[C(OP_WRITE)] = {
186 		[C(RESULT_ACCESS)] = -1,
187 		[C(RESULT_MISS)]   = -1,
188 	},
189 	[C(OP_PREFETCH)] = {
190 		[C(RESULT_ACCESS)] = -1,
191 		[C(RESULT_MISS)]   = -1,
192 	},
193 },
194 [C(BPU)] = {
195 	[C(OP_READ)] = {
196 		[C(RESULT_ACCESS)] = 0x00c2, /* Retired Branch Instr.      */
197 		[C(RESULT_MISS)]   = 0x00c3, /* Retired Mispredicted BI    */
198 	},
199 	[C(OP_WRITE)] = {
200 		[C(RESULT_ACCESS)] = -1,
201 		[C(RESULT_MISS)]   = -1,
202 	},
203 	[C(OP_PREFETCH)] = {
204 		[C(RESULT_ACCESS)] = -1,
205 		[C(RESULT_MISS)]   = -1,
206 	},
207 },
208 [C(NODE)] = {
209 	[C(OP_READ)] = {
210 		[C(RESULT_ACCESS)] = 0,
211 		[C(RESULT_MISS)]   = 0,
212 	},
213 	[C(OP_WRITE)] = {
214 		[C(RESULT_ACCESS)] = -1,
215 		[C(RESULT_MISS)]   = -1,
216 	},
217 	[C(OP_PREFETCH)] = {
218 		[C(RESULT_ACCESS)] = -1,
219 		[C(RESULT_MISS)]   = -1,
220 	},
221 },
222 };
223 
224 /*
225  * AMD Performance Monitor K7 and later, up to and including Family 16h:
226  */
227 static const u64 amd_perfmon_event_map[PERF_COUNT_HW_MAX] =
228 {
229 	[PERF_COUNT_HW_CPU_CYCLES]		= 0x0076,
230 	[PERF_COUNT_HW_INSTRUCTIONS]		= 0x00c0,
231 	[PERF_COUNT_HW_CACHE_REFERENCES]	= 0x077d,
232 	[PERF_COUNT_HW_CACHE_MISSES]		= 0x077e,
233 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x00c2,
234 	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x00c3,
235 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= 0x00d0, /* "Decoder empty" event */
236 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= 0x00d1, /* "Dispatch stalls" event */
237 };
238 
239 /*
240  * AMD Performance Monitor Family 17h and later:
241  */
242 static const u64 amd_f17h_perfmon_event_map[PERF_COUNT_HW_MAX] =
243 {
244 	[PERF_COUNT_HW_CPU_CYCLES]		= 0x0076,
245 	[PERF_COUNT_HW_INSTRUCTIONS]		= 0x00c0,
246 	[PERF_COUNT_HW_CACHE_REFERENCES]	= 0xff60,
247 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x00c2,
248 	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x00c3,
249 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= 0x0287,
250 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= 0x0187,
251 };
252 
253 static u64 amd_pmu_event_map(int hw_event)
254 {
255 	if (boot_cpu_data.x86 >= 0x17)
256 		return amd_f17h_perfmon_event_map[hw_event];
257 
258 	return amd_perfmon_event_map[hw_event];
259 }
260 
261 /*
262  * Previously calculated offsets
263  */
264 static unsigned int event_offsets[X86_PMC_IDX_MAX] __read_mostly;
265 static unsigned int count_offsets[X86_PMC_IDX_MAX] __read_mostly;
266 
267 /*
268  * Legacy CPUs:
269  *   4 counters starting at 0xc0010000 each offset by 1
270  *
271  * CPUs with core performance counter extensions:
272  *   6 counters starting at 0xc0010200 each offset by 2
273  */
274 static inline int amd_pmu_addr_offset(int index, bool eventsel)
275 {
276 	int offset;
277 
278 	if (!index)
279 		return index;
280 
281 	if (eventsel)
282 		offset = event_offsets[index];
283 	else
284 		offset = count_offsets[index];
285 
286 	if (offset)
287 		return offset;
288 
289 	if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
290 		offset = index;
291 	else
292 		offset = index << 1;
293 
294 	if (eventsel)
295 		event_offsets[index] = offset;
296 	else
297 		count_offsets[index] = offset;
298 
299 	return offset;
300 }
301 
302 static int amd_core_hw_config(struct perf_event *event)
303 {
304 	if (event->attr.exclude_host && event->attr.exclude_guest)
305 		/*
306 		 * When HO == GO == 1 the hardware treats that as GO == HO == 0
307 		 * and will count in both modes. We don't want to count in that
308 		 * case so we emulate no-counting by setting US = OS = 0.
309 		 */
310 		event->hw.config &= ~(ARCH_PERFMON_EVENTSEL_USR |
311 				      ARCH_PERFMON_EVENTSEL_OS);
312 	else if (event->attr.exclude_host)
313 		event->hw.config |= AMD64_EVENTSEL_GUESTONLY;
314 	else if (event->attr.exclude_guest)
315 		event->hw.config |= AMD64_EVENTSEL_HOSTONLY;
316 
317 	return 0;
318 }
319 
320 /*
321  * AMD64 events are detected based on their event codes.
322  */
323 static inline unsigned int amd_get_event_code(struct hw_perf_event *hwc)
324 {
325 	return ((hwc->config >> 24) & 0x0f00) | (hwc->config & 0x00ff);
326 }
327 
328 static inline int amd_is_nb_event(struct hw_perf_event *hwc)
329 {
330 	return (hwc->config & 0xe0) == 0xe0;
331 }
332 
333 static inline int amd_has_nb(struct cpu_hw_events *cpuc)
334 {
335 	struct amd_nb *nb = cpuc->amd_nb;
336 
337 	return nb && nb->nb_id != -1;
338 }
339 
340 static int amd_pmu_hw_config(struct perf_event *event)
341 {
342 	int ret;
343 
344 	/* pass precise event sampling to ibs: */
345 	if (event->attr.precise_ip && get_ibs_caps())
346 		return -ENOENT;
347 
348 	if (has_branch_stack(event))
349 		return -EOPNOTSUPP;
350 
351 	ret = x86_pmu_hw_config(event);
352 	if (ret)
353 		return ret;
354 
355 	if (event->attr.type == PERF_TYPE_RAW)
356 		event->hw.config |= event->attr.config & AMD64_RAW_EVENT_MASK;
357 
358 	return amd_core_hw_config(event);
359 }
360 
361 static void __amd_put_nb_event_constraints(struct cpu_hw_events *cpuc,
362 					   struct perf_event *event)
363 {
364 	struct amd_nb *nb = cpuc->amd_nb;
365 	int i;
366 
367 	/*
368 	 * need to scan whole list because event may not have
369 	 * been assigned during scheduling
370 	 *
371 	 * no race condition possible because event can only
372 	 * be removed on one CPU at a time AND PMU is disabled
373 	 * when we come here
374 	 */
375 	for (i = 0; i < x86_pmu.num_counters; i++) {
376 		if (cmpxchg(nb->owners + i, event, NULL) == event)
377 			break;
378 	}
379 }
380 
381  /*
382   * AMD64 NorthBridge events need special treatment because
383   * counter access needs to be synchronized across all cores
384   * of a package. Refer to BKDG section 3.12
385   *
386   * NB events are events measuring L3 cache, Hypertransport
387   * traffic. They are identified by an event code >= 0xe00.
388   * They measure events on the NorthBride which is shared
389   * by all cores on a package. NB events are counted on a
390   * shared set of counters. When a NB event is programmed
391   * in a counter, the data actually comes from a shared
392   * counter. Thus, access to those counters needs to be
393   * synchronized.
394   *
395   * We implement the synchronization such that no two cores
396   * can be measuring NB events using the same counters. Thus,
397   * we maintain a per-NB allocation table. The available slot
398   * is propagated using the event_constraint structure.
399   *
400   * We provide only one choice for each NB event based on
401   * the fact that only NB events have restrictions. Consequently,
402   * if a counter is available, there is a guarantee the NB event
403   * will be assigned to it. If no slot is available, an empty
404   * constraint is returned and scheduling will eventually fail
405   * for this event.
406   *
407   * Note that all cores attached the same NB compete for the same
408   * counters to host NB events, this is why we use atomic ops. Some
409   * multi-chip CPUs may have more than one NB.
410   *
411   * Given that resources are allocated (cmpxchg), they must be
412   * eventually freed for others to use. This is accomplished by
413   * calling __amd_put_nb_event_constraints()
414   *
415   * Non NB events are not impacted by this restriction.
416   */
417 static struct event_constraint *
418 __amd_get_nb_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
419 			       struct event_constraint *c)
420 {
421 	struct hw_perf_event *hwc = &event->hw;
422 	struct amd_nb *nb = cpuc->amd_nb;
423 	struct perf_event *old;
424 	int idx, new = -1;
425 
426 	if (!c)
427 		c = &unconstrained;
428 
429 	if (cpuc->is_fake)
430 		return c;
431 
432 	/*
433 	 * detect if already present, if so reuse
434 	 *
435 	 * cannot merge with actual allocation
436 	 * because of possible holes
437 	 *
438 	 * event can already be present yet not assigned (in hwc->idx)
439 	 * because of successive calls to x86_schedule_events() from
440 	 * hw_perf_group_sched_in() without hw_perf_enable()
441 	 */
442 	for_each_set_bit(idx, c->idxmsk, x86_pmu.num_counters) {
443 		if (new == -1 || hwc->idx == idx)
444 			/* assign free slot, prefer hwc->idx */
445 			old = cmpxchg(nb->owners + idx, NULL, event);
446 		else if (nb->owners[idx] == event)
447 			/* event already present */
448 			old = event;
449 		else
450 			continue;
451 
452 		if (old && old != event)
453 			continue;
454 
455 		/* reassign to this slot */
456 		if (new != -1)
457 			cmpxchg(nb->owners + new, event, NULL);
458 		new = idx;
459 
460 		/* already present, reuse */
461 		if (old == event)
462 			break;
463 	}
464 
465 	if (new == -1)
466 		return &emptyconstraint;
467 
468 	return &nb->event_constraints[new];
469 }
470 
471 static struct amd_nb *amd_alloc_nb(int cpu)
472 {
473 	struct amd_nb *nb;
474 	int i;
475 
476 	nb = kzalloc_node(sizeof(struct amd_nb), GFP_KERNEL, cpu_to_node(cpu));
477 	if (!nb)
478 		return NULL;
479 
480 	nb->nb_id = -1;
481 
482 	/*
483 	 * initialize all possible NB constraints
484 	 */
485 	for (i = 0; i < x86_pmu.num_counters; i++) {
486 		__set_bit(i, nb->event_constraints[i].idxmsk);
487 		nb->event_constraints[i].weight = 1;
488 	}
489 	return nb;
490 }
491 
492 static int amd_pmu_cpu_prepare(int cpu)
493 {
494 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
495 
496 	WARN_ON_ONCE(cpuc->amd_nb);
497 
498 	if (!x86_pmu.amd_nb_constraints)
499 		return 0;
500 
501 	cpuc->amd_nb = amd_alloc_nb(cpu);
502 	if (!cpuc->amd_nb)
503 		return -ENOMEM;
504 
505 	return 0;
506 }
507 
508 static void amd_pmu_cpu_starting(int cpu)
509 {
510 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
511 	void **onln = &cpuc->kfree_on_online[X86_PERF_KFREE_SHARED];
512 	struct amd_nb *nb;
513 	int i, nb_id;
514 
515 	cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
516 
517 	if (!x86_pmu.amd_nb_constraints)
518 		return;
519 
520 	nb_id = amd_get_nb_id(cpu);
521 	WARN_ON_ONCE(nb_id == BAD_APICID);
522 
523 	for_each_online_cpu(i) {
524 		nb = per_cpu(cpu_hw_events, i).amd_nb;
525 		if (WARN_ON_ONCE(!nb))
526 			continue;
527 
528 		if (nb->nb_id == nb_id) {
529 			*onln = cpuc->amd_nb;
530 			cpuc->amd_nb = nb;
531 			break;
532 		}
533 	}
534 
535 	cpuc->amd_nb->nb_id = nb_id;
536 	cpuc->amd_nb->refcnt++;
537 }
538 
539 static void amd_pmu_cpu_dead(int cpu)
540 {
541 	struct cpu_hw_events *cpuhw;
542 
543 	if (!x86_pmu.amd_nb_constraints)
544 		return;
545 
546 	cpuhw = &per_cpu(cpu_hw_events, cpu);
547 
548 	if (cpuhw->amd_nb) {
549 		struct amd_nb *nb = cpuhw->amd_nb;
550 
551 		if (nb->nb_id == -1 || --nb->refcnt == 0)
552 			kfree(nb);
553 
554 		cpuhw->amd_nb = NULL;
555 	}
556 }
557 
558 /*
559  * When a PMC counter overflows, an NMI is used to process the event and
560  * reset the counter. NMI latency can result in the counter being updated
561  * before the NMI can run, which can result in what appear to be spurious
562  * NMIs. This function is intended to wait for the NMI to run and reset
563  * the counter to avoid possible unhandled NMI messages.
564  */
565 #define OVERFLOW_WAIT_COUNT	50
566 
567 static void amd_pmu_wait_on_overflow(int idx)
568 {
569 	unsigned int i;
570 	u64 counter;
571 
572 	/*
573 	 * Wait for the counter to be reset if it has overflowed. This loop
574 	 * should exit very, very quickly, but just in case, don't wait
575 	 * forever...
576 	 */
577 	for (i = 0; i < OVERFLOW_WAIT_COUNT; i++) {
578 		rdmsrl(x86_pmu_event_addr(idx), counter);
579 		if (counter & (1ULL << (x86_pmu.cntval_bits - 1)))
580 			break;
581 
582 		/* Might be in IRQ context, so can't sleep */
583 		udelay(1);
584 	}
585 }
586 
587 static void amd_pmu_disable_all(void)
588 {
589 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
590 	int idx;
591 
592 	x86_pmu_disable_all();
593 
594 	/*
595 	 * This shouldn't be called from NMI context, but add a safeguard here
596 	 * to return, since if we're in NMI context we can't wait for an NMI
597 	 * to reset an overflowed counter value.
598 	 */
599 	if (in_nmi())
600 		return;
601 
602 	/*
603 	 * Check each counter for overflow and wait for it to be reset by the
604 	 * NMI if it has overflowed. This relies on the fact that all active
605 	 * counters are always enabled when this function is caled and
606 	 * ARCH_PERFMON_EVENTSEL_INT is always set.
607 	 */
608 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
609 		if (!test_bit(idx, cpuc->active_mask))
610 			continue;
611 
612 		amd_pmu_wait_on_overflow(idx);
613 	}
614 }
615 
616 static void amd_pmu_disable_event(struct perf_event *event)
617 {
618 	x86_pmu_disable_event(event);
619 
620 	/*
621 	 * This can be called from NMI context (via x86_pmu_stop). The counter
622 	 * may have overflowed, but either way, we'll never see it get reset
623 	 * by the NMI if we're already in the NMI. And the NMI latency support
624 	 * below will take care of any pending NMI that might have been
625 	 * generated by the overflow.
626 	 */
627 	if (in_nmi())
628 		return;
629 
630 	amd_pmu_wait_on_overflow(event->hw.idx);
631 }
632 
633 /*
634  * Because of NMI latency, if multiple PMC counters are active or other sources
635  * of NMIs are received, the perf NMI handler can handle one or more overflowed
636  * PMC counters outside of the NMI associated with the PMC overflow. If the NMI
637  * doesn't arrive at the LAPIC in time to become a pending NMI, then the kernel
638  * back-to-back NMI support won't be active. This PMC handler needs to take into
639  * account that this can occur, otherwise this could result in unknown NMI
640  * messages being issued. Examples of this is PMC overflow while in the NMI
641  * handler when multiple PMCs are active or PMC overflow while handling some
642  * other source of an NMI.
643  *
644  * Attempt to mitigate this by using the number of active PMCs to determine
645  * whether to return NMI_HANDLED if the perf NMI handler did not handle/reset
646  * any PMCs. The per-CPU perf_nmi_counter variable is set to a minimum of the
647  * number of active PMCs or 2. The value of 2 is used in case an NMI does not
648  * arrive at the LAPIC in time to be collapsed into an already pending NMI.
649  */
650 static int amd_pmu_handle_irq(struct pt_regs *regs)
651 {
652 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
653 	int active, handled;
654 
655 	/*
656 	 * Obtain the active count before calling x86_pmu_handle_irq() since
657 	 * it is possible that x86_pmu_handle_irq() may make a counter
658 	 * inactive (through x86_pmu_stop).
659 	 */
660 	active = __bitmap_weight(cpuc->active_mask, X86_PMC_IDX_MAX);
661 
662 	/* Process any counter overflows */
663 	handled = x86_pmu_handle_irq(regs);
664 
665 	/*
666 	 * If a counter was handled, record the number of possible remaining
667 	 * NMIs that can occur.
668 	 */
669 	if (handled) {
670 		this_cpu_write(perf_nmi_counter,
671 			       min_t(unsigned int, 2, active));
672 
673 		return handled;
674 	}
675 
676 	if (!this_cpu_read(perf_nmi_counter))
677 		return NMI_DONE;
678 
679 	this_cpu_dec(perf_nmi_counter);
680 
681 	return NMI_HANDLED;
682 }
683 
684 static struct event_constraint *
685 amd_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
686 			  struct perf_event *event)
687 {
688 	/*
689 	 * if not NB event or no NB, then no constraints
690 	 */
691 	if (!(amd_has_nb(cpuc) && amd_is_nb_event(&event->hw)))
692 		return &unconstrained;
693 
694 	return __amd_get_nb_event_constraints(cpuc, event, NULL);
695 }
696 
697 static void amd_put_event_constraints(struct cpu_hw_events *cpuc,
698 				      struct perf_event *event)
699 {
700 	if (amd_has_nb(cpuc) && amd_is_nb_event(&event->hw))
701 		__amd_put_nb_event_constraints(cpuc, event);
702 }
703 
704 PMU_FORMAT_ATTR(event,	"config:0-7,32-35");
705 PMU_FORMAT_ATTR(umask,	"config:8-15"	);
706 PMU_FORMAT_ATTR(edge,	"config:18"	);
707 PMU_FORMAT_ATTR(inv,	"config:23"	);
708 PMU_FORMAT_ATTR(cmask,	"config:24-31"	);
709 
710 static struct attribute *amd_format_attr[] = {
711 	&format_attr_event.attr,
712 	&format_attr_umask.attr,
713 	&format_attr_edge.attr,
714 	&format_attr_inv.attr,
715 	&format_attr_cmask.attr,
716 	NULL,
717 };
718 
719 /* AMD Family 15h */
720 
721 #define AMD_EVENT_TYPE_MASK	0x000000F0ULL
722 
723 #define AMD_EVENT_FP		0x00000000ULL ... 0x00000010ULL
724 #define AMD_EVENT_LS		0x00000020ULL ... 0x00000030ULL
725 #define AMD_EVENT_DC		0x00000040ULL ... 0x00000050ULL
726 #define AMD_EVENT_CU		0x00000060ULL ... 0x00000070ULL
727 #define AMD_EVENT_IC_DE		0x00000080ULL ... 0x00000090ULL
728 #define AMD_EVENT_EX_LS		0x000000C0ULL
729 #define AMD_EVENT_DE		0x000000D0ULL
730 #define AMD_EVENT_NB		0x000000E0ULL ... 0x000000F0ULL
731 
732 /*
733  * AMD family 15h event code/PMC mappings:
734  *
735  * type = event_code & 0x0F0:
736  *
737  * 0x000	FP	PERF_CTL[5:3]
738  * 0x010	FP	PERF_CTL[5:3]
739  * 0x020	LS	PERF_CTL[5:0]
740  * 0x030	LS	PERF_CTL[5:0]
741  * 0x040	DC	PERF_CTL[5:0]
742  * 0x050	DC	PERF_CTL[5:0]
743  * 0x060	CU	PERF_CTL[2:0]
744  * 0x070	CU	PERF_CTL[2:0]
745  * 0x080	IC/DE	PERF_CTL[2:0]
746  * 0x090	IC/DE	PERF_CTL[2:0]
747  * 0x0A0	---
748  * 0x0B0	---
749  * 0x0C0	EX/LS	PERF_CTL[5:0]
750  * 0x0D0	DE	PERF_CTL[2:0]
751  * 0x0E0	NB	NB_PERF_CTL[3:0]
752  * 0x0F0	NB	NB_PERF_CTL[3:0]
753  *
754  * Exceptions:
755  *
756  * 0x000	FP	PERF_CTL[3], PERF_CTL[5:3] (*)
757  * 0x003	FP	PERF_CTL[3]
758  * 0x004	FP	PERF_CTL[3], PERF_CTL[5:3] (*)
759  * 0x00B	FP	PERF_CTL[3]
760  * 0x00D	FP	PERF_CTL[3]
761  * 0x023	DE	PERF_CTL[2:0]
762  * 0x02D	LS	PERF_CTL[3]
763  * 0x02E	LS	PERF_CTL[3,0]
764  * 0x031	LS	PERF_CTL[2:0] (**)
765  * 0x043	CU	PERF_CTL[2:0]
766  * 0x045	CU	PERF_CTL[2:0]
767  * 0x046	CU	PERF_CTL[2:0]
768  * 0x054	CU	PERF_CTL[2:0]
769  * 0x055	CU	PERF_CTL[2:0]
770  * 0x08F	IC	PERF_CTL[0]
771  * 0x187	DE	PERF_CTL[0]
772  * 0x188	DE	PERF_CTL[0]
773  * 0x0DB	EX	PERF_CTL[5:0]
774  * 0x0DC	LS	PERF_CTL[5:0]
775  * 0x0DD	LS	PERF_CTL[5:0]
776  * 0x0DE	LS	PERF_CTL[5:0]
777  * 0x0DF	LS	PERF_CTL[5:0]
778  * 0x1C0	EX	PERF_CTL[5:3]
779  * 0x1D6	EX	PERF_CTL[5:0]
780  * 0x1D8	EX	PERF_CTL[5:0]
781  *
782  * (*)  depending on the umask all FPU counters may be used
783  * (**) only one unitmask enabled at a time
784  */
785 
786 static struct event_constraint amd_f15_PMC0  = EVENT_CONSTRAINT(0, 0x01, 0);
787 static struct event_constraint amd_f15_PMC20 = EVENT_CONSTRAINT(0, 0x07, 0);
788 static struct event_constraint amd_f15_PMC3  = EVENT_CONSTRAINT(0, 0x08, 0);
789 static struct event_constraint amd_f15_PMC30 = EVENT_CONSTRAINT_OVERLAP(0, 0x09, 0);
790 static struct event_constraint amd_f15_PMC50 = EVENT_CONSTRAINT(0, 0x3F, 0);
791 static struct event_constraint amd_f15_PMC53 = EVENT_CONSTRAINT(0, 0x38, 0);
792 
793 static struct event_constraint *
794 amd_get_event_constraints_f15h(struct cpu_hw_events *cpuc, int idx,
795 			       struct perf_event *event)
796 {
797 	struct hw_perf_event *hwc = &event->hw;
798 	unsigned int event_code = amd_get_event_code(hwc);
799 
800 	switch (event_code & AMD_EVENT_TYPE_MASK) {
801 	case AMD_EVENT_FP:
802 		switch (event_code) {
803 		case 0x000:
804 			if (!(hwc->config & 0x0000F000ULL))
805 				break;
806 			if (!(hwc->config & 0x00000F00ULL))
807 				break;
808 			return &amd_f15_PMC3;
809 		case 0x004:
810 			if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
811 				break;
812 			return &amd_f15_PMC3;
813 		case 0x003:
814 		case 0x00B:
815 		case 0x00D:
816 			return &amd_f15_PMC3;
817 		}
818 		return &amd_f15_PMC53;
819 	case AMD_EVENT_LS:
820 	case AMD_EVENT_DC:
821 	case AMD_EVENT_EX_LS:
822 		switch (event_code) {
823 		case 0x023:
824 		case 0x043:
825 		case 0x045:
826 		case 0x046:
827 		case 0x054:
828 		case 0x055:
829 			return &amd_f15_PMC20;
830 		case 0x02D:
831 			return &amd_f15_PMC3;
832 		case 0x02E:
833 			return &amd_f15_PMC30;
834 		case 0x031:
835 			if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
836 				return &amd_f15_PMC20;
837 			return &emptyconstraint;
838 		case 0x1C0:
839 			return &amd_f15_PMC53;
840 		default:
841 			return &amd_f15_PMC50;
842 		}
843 	case AMD_EVENT_CU:
844 	case AMD_EVENT_IC_DE:
845 	case AMD_EVENT_DE:
846 		switch (event_code) {
847 		case 0x08F:
848 		case 0x187:
849 		case 0x188:
850 			return &amd_f15_PMC0;
851 		case 0x0DB ... 0x0DF:
852 		case 0x1D6:
853 		case 0x1D8:
854 			return &amd_f15_PMC50;
855 		default:
856 			return &amd_f15_PMC20;
857 		}
858 	case AMD_EVENT_NB:
859 		/* moved to uncore.c */
860 		return &emptyconstraint;
861 	default:
862 		return &emptyconstraint;
863 	}
864 }
865 
866 static ssize_t amd_event_sysfs_show(char *page, u64 config)
867 {
868 	u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT) |
869 		    (config & AMD64_EVENTSEL_EVENT) >> 24;
870 
871 	return x86_event_sysfs_show(page, config, event);
872 }
873 
874 static __initconst const struct x86_pmu amd_pmu = {
875 	.name			= "AMD",
876 	.handle_irq		= amd_pmu_handle_irq,
877 	.disable_all		= amd_pmu_disable_all,
878 	.enable_all		= x86_pmu_enable_all,
879 	.enable			= x86_pmu_enable_event,
880 	.disable		= amd_pmu_disable_event,
881 	.hw_config		= amd_pmu_hw_config,
882 	.schedule_events	= x86_schedule_events,
883 	.eventsel		= MSR_K7_EVNTSEL0,
884 	.perfctr		= MSR_K7_PERFCTR0,
885 	.addr_offset            = amd_pmu_addr_offset,
886 	.event_map		= amd_pmu_event_map,
887 	.max_events		= ARRAY_SIZE(amd_perfmon_event_map),
888 	.num_counters		= AMD64_NUM_COUNTERS,
889 	.cntval_bits		= 48,
890 	.cntval_mask		= (1ULL << 48) - 1,
891 	.apic			= 1,
892 	/* use highest bit to detect overflow */
893 	.max_period		= (1ULL << 47) - 1,
894 	.get_event_constraints	= amd_get_event_constraints,
895 	.put_event_constraints	= amd_put_event_constraints,
896 
897 	.format_attrs		= amd_format_attr,
898 	.events_sysfs_show	= amd_event_sysfs_show,
899 
900 	.cpu_prepare		= amd_pmu_cpu_prepare,
901 	.cpu_starting		= amd_pmu_cpu_starting,
902 	.cpu_dead		= amd_pmu_cpu_dead,
903 
904 	.amd_nb_constraints	= 1,
905 };
906 
907 static int __init amd_core_pmu_init(void)
908 {
909 	if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
910 		return 0;
911 
912 	switch (boot_cpu_data.x86) {
913 	case 0x15:
914 		pr_cont("Fam15h ");
915 		x86_pmu.get_event_constraints = amd_get_event_constraints_f15h;
916 		break;
917 	case 0x17:
918 		pr_cont("Fam17h ");
919 		/*
920 		 * In family 17h, there are no event constraints in the PMC hardware.
921 		 * We fallback to using default amd_get_event_constraints.
922 		 */
923 		break;
924 	case 0x18:
925 		pr_cont("Fam18h ");
926 		/* Using default amd_get_event_constraints. */
927 		break;
928 	default:
929 		pr_err("core perfctr but no constraints; unknown hardware!\n");
930 		return -ENODEV;
931 	}
932 
933 	/*
934 	 * If core performance counter extensions exists, we must use
935 	 * MSR_F15H_PERF_CTL/MSR_F15H_PERF_CTR msrs. See also
936 	 * amd_pmu_addr_offset().
937 	 */
938 	x86_pmu.eventsel	= MSR_F15H_PERF_CTL;
939 	x86_pmu.perfctr		= MSR_F15H_PERF_CTR;
940 	x86_pmu.num_counters	= AMD64_NUM_COUNTERS_CORE;
941 	/*
942 	 * AMD Core perfctr has separate MSRs for the NB events, see
943 	 * the amd/uncore.c driver.
944 	 */
945 	x86_pmu.amd_nb_constraints = 0;
946 
947 	pr_cont("core perfctr, ");
948 	return 0;
949 }
950 
951 __init int amd_pmu_init(void)
952 {
953 	int ret;
954 
955 	/* Performance-monitoring supported from K7 and later: */
956 	if (boot_cpu_data.x86 < 6)
957 		return -ENODEV;
958 
959 	x86_pmu = amd_pmu;
960 
961 	ret = amd_core_pmu_init();
962 	if (ret)
963 		return ret;
964 
965 	if (num_possible_cpus() == 1) {
966 		/*
967 		 * No point in allocating data structures to serialize
968 		 * against other CPUs, when there is only the one CPU.
969 		 */
970 		x86_pmu.amd_nb_constraints = 0;
971 	}
972 
973 	if (boot_cpu_data.x86 >= 0x17)
974 		memcpy(hw_cache_event_ids, amd_hw_cache_event_ids_f17h, sizeof(hw_cache_event_ids));
975 	else
976 		memcpy(hw_cache_event_ids, amd_hw_cache_event_ids, sizeof(hw_cache_event_ids));
977 
978 	return 0;
979 }
980 
981 void amd_pmu_enable_virt(void)
982 {
983 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
984 
985 	cpuc->perf_ctr_virt_mask = 0;
986 
987 	/* Reload all events */
988 	amd_pmu_disable_all();
989 	x86_pmu_enable_all(0);
990 }
991 EXPORT_SYMBOL_GPL(amd_pmu_enable_virt);
992 
993 void amd_pmu_disable_virt(void)
994 {
995 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
996 
997 	/*
998 	 * We only mask out the Host-only bit so that host-only counting works
999 	 * when SVM is disabled. If someone sets up a guest-only counter when
1000 	 * SVM is disabled the Guest-only bits still gets set and the counter
1001 	 * will not count anything.
1002 	 */
1003 	cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
1004 
1005 	/* Reload all events */
1006 	amd_pmu_disable_all();
1007 	x86_pmu_enable_all(0);
1008 }
1009 EXPORT_SYMBOL_GPL(amd_pmu_disable_virt);
1010