xref: /openbmc/linux/arch/x86/events/amd/core.c (revision 151f4e2b)
1 #include <linux/perf_event.h>
2 #include <linux/export.h>
3 #include <linux/types.h>
4 #include <linux/init.h>
5 #include <linux/slab.h>
6 #include <linux/delay.h>
7 #include <asm/apicdef.h>
8 #include <asm/nmi.h>
9 
10 #include "../perf_event.h"
11 
12 static DEFINE_PER_CPU(unsigned int, perf_nmi_counter);
13 
14 static __initconst const u64 amd_hw_cache_event_ids
15 				[PERF_COUNT_HW_CACHE_MAX]
16 				[PERF_COUNT_HW_CACHE_OP_MAX]
17 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
18 {
19  [ C(L1D) ] = {
20 	[ C(OP_READ) ] = {
21 		[ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses        */
22 		[ C(RESULT_MISS)   ] = 0x0141, /* Data Cache Misses          */
23 	},
24 	[ C(OP_WRITE) ] = {
25 		[ C(RESULT_ACCESS) ] = 0,
26 		[ C(RESULT_MISS)   ] = 0,
27 	},
28 	[ C(OP_PREFETCH) ] = {
29 		[ C(RESULT_ACCESS) ] = 0x0267, /* Data Prefetcher :attempts  */
30 		[ C(RESULT_MISS)   ] = 0x0167, /* Data Prefetcher :cancelled */
31 	},
32  },
33  [ C(L1I ) ] = {
34 	[ C(OP_READ) ] = {
35 		[ C(RESULT_ACCESS) ] = 0x0080, /* Instruction cache fetches  */
36 		[ C(RESULT_MISS)   ] = 0x0081, /* Instruction cache misses   */
37 	},
38 	[ C(OP_WRITE) ] = {
39 		[ C(RESULT_ACCESS) ] = -1,
40 		[ C(RESULT_MISS)   ] = -1,
41 	},
42 	[ C(OP_PREFETCH) ] = {
43 		[ C(RESULT_ACCESS) ] = 0x014B, /* Prefetch Instructions :Load */
44 		[ C(RESULT_MISS)   ] = 0,
45 	},
46  },
47  [ C(LL  ) ] = {
48 	[ C(OP_READ) ] = {
49 		[ C(RESULT_ACCESS) ] = 0x037D, /* Requests to L2 Cache :IC+DC */
50 		[ C(RESULT_MISS)   ] = 0x037E, /* L2 Cache Misses : IC+DC     */
51 	},
52 	[ C(OP_WRITE) ] = {
53 		[ C(RESULT_ACCESS) ] = 0x017F, /* L2 Fill/Writeback           */
54 		[ C(RESULT_MISS)   ] = 0,
55 	},
56 	[ C(OP_PREFETCH) ] = {
57 		[ C(RESULT_ACCESS) ] = 0,
58 		[ C(RESULT_MISS)   ] = 0,
59 	},
60  },
61  [ C(DTLB) ] = {
62 	[ C(OP_READ) ] = {
63 		[ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses        */
64 		[ C(RESULT_MISS)   ] = 0x0746, /* L1_DTLB_AND_L2_DLTB_MISS.ALL */
65 	},
66 	[ C(OP_WRITE) ] = {
67 		[ C(RESULT_ACCESS) ] = 0,
68 		[ C(RESULT_MISS)   ] = 0,
69 	},
70 	[ C(OP_PREFETCH) ] = {
71 		[ C(RESULT_ACCESS) ] = 0,
72 		[ C(RESULT_MISS)   ] = 0,
73 	},
74  },
75  [ C(ITLB) ] = {
76 	[ C(OP_READ) ] = {
77 		[ C(RESULT_ACCESS) ] = 0x0080, /* Instruction fecthes        */
78 		[ C(RESULT_MISS)   ] = 0x0385, /* L1_ITLB_AND_L2_ITLB_MISS.ALL */
79 	},
80 	[ C(OP_WRITE) ] = {
81 		[ C(RESULT_ACCESS) ] = -1,
82 		[ C(RESULT_MISS)   ] = -1,
83 	},
84 	[ C(OP_PREFETCH) ] = {
85 		[ C(RESULT_ACCESS) ] = -1,
86 		[ C(RESULT_MISS)   ] = -1,
87 	},
88  },
89  [ C(BPU ) ] = {
90 	[ C(OP_READ) ] = {
91 		[ C(RESULT_ACCESS) ] = 0x00c2, /* Retired Branch Instr.      */
92 		[ C(RESULT_MISS)   ] = 0x00c3, /* Retired Mispredicted BI    */
93 	},
94 	[ C(OP_WRITE) ] = {
95 		[ C(RESULT_ACCESS) ] = -1,
96 		[ C(RESULT_MISS)   ] = -1,
97 	},
98 	[ C(OP_PREFETCH) ] = {
99 		[ C(RESULT_ACCESS) ] = -1,
100 		[ C(RESULT_MISS)   ] = -1,
101 	},
102  },
103  [ C(NODE) ] = {
104 	[ C(OP_READ) ] = {
105 		[ C(RESULT_ACCESS) ] = 0xb8e9, /* CPU Request to Memory, l+r */
106 		[ C(RESULT_MISS)   ] = 0x98e9, /* CPU Request to Memory, r   */
107 	},
108 	[ C(OP_WRITE) ] = {
109 		[ C(RESULT_ACCESS) ] = -1,
110 		[ C(RESULT_MISS)   ] = -1,
111 	},
112 	[ C(OP_PREFETCH) ] = {
113 		[ C(RESULT_ACCESS) ] = -1,
114 		[ C(RESULT_MISS)   ] = -1,
115 	},
116  },
117 };
118 
119 static __initconst const u64 amd_hw_cache_event_ids_f17h
120 				[PERF_COUNT_HW_CACHE_MAX]
121 				[PERF_COUNT_HW_CACHE_OP_MAX]
122 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
123 [C(L1D)] = {
124 	[C(OP_READ)] = {
125 		[C(RESULT_ACCESS)] = 0x0040, /* Data Cache Accesses */
126 		[C(RESULT_MISS)]   = 0xc860, /* L2$ access from DC Miss */
127 	},
128 	[C(OP_WRITE)] = {
129 		[C(RESULT_ACCESS)] = 0,
130 		[C(RESULT_MISS)]   = 0,
131 	},
132 	[C(OP_PREFETCH)] = {
133 		[C(RESULT_ACCESS)] = 0xff5a, /* h/w prefetch DC Fills */
134 		[C(RESULT_MISS)]   = 0,
135 	},
136 },
137 [C(L1I)] = {
138 	[C(OP_READ)] = {
139 		[C(RESULT_ACCESS)] = 0x0080, /* Instruction cache fetches  */
140 		[C(RESULT_MISS)]   = 0x0081, /* Instruction cache misses   */
141 	},
142 	[C(OP_WRITE)] = {
143 		[C(RESULT_ACCESS)] = -1,
144 		[C(RESULT_MISS)]   = -1,
145 	},
146 	[C(OP_PREFETCH)] = {
147 		[C(RESULT_ACCESS)] = 0,
148 		[C(RESULT_MISS)]   = 0,
149 	},
150 },
151 [C(LL)] = {
152 	[C(OP_READ)] = {
153 		[C(RESULT_ACCESS)] = 0,
154 		[C(RESULT_MISS)]   = 0,
155 	},
156 	[C(OP_WRITE)] = {
157 		[C(RESULT_ACCESS)] = 0,
158 		[C(RESULT_MISS)]   = 0,
159 	},
160 	[C(OP_PREFETCH)] = {
161 		[C(RESULT_ACCESS)] = 0,
162 		[C(RESULT_MISS)]   = 0,
163 	},
164 },
165 [C(DTLB)] = {
166 	[C(OP_READ)] = {
167 		[C(RESULT_ACCESS)] = 0xff45, /* All L2 DTLB accesses */
168 		[C(RESULT_MISS)]   = 0xf045, /* L2 DTLB misses (PT walks) */
169 	},
170 	[C(OP_WRITE)] = {
171 		[C(RESULT_ACCESS)] = 0,
172 		[C(RESULT_MISS)]   = 0,
173 	},
174 	[C(OP_PREFETCH)] = {
175 		[C(RESULT_ACCESS)] = 0,
176 		[C(RESULT_MISS)]   = 0,
177 	},
178 },
179 [C(ITLB)] = {
180 	[C(OP_READ)] = {
181 		[C(RESULT_ACCESS)] = 0x0084, /* L1 ITLB misses, L2 ITLB hits */
182 		[C(RESULT_MISS)]   = 0xff85, /* L1 ITLB misses, L2 misses */
183 	},
184 	[C(OP_WRITE)] = {
185 		[C(RESULT_ACCESS)] = -1,
186 		[C(RESULT_MISS)]   = -1,
187 	},
188 	[C(OP_PREFETCH)] = {
189 		[C(RESULT_ACCESS)] = -1,
190 		[C(RESULT_MISS)]   = -1,
191 	},
192 },
193 [C(BPU)] = {
194 	[C(OP_READ)] = {
195 		[C(RESULT_ACCESS)] = 0x00c2, /* Retired Branch Instr.      */
196 		[C(RESULT_MISS)]   = 0x00c3, /* Retired Mispredicted BI    */
197 	},
198 	[C(OP_WRITE)] = {
199 		[C(RESULT_ACCESS)] = -1,
200 		[C(RESULT_MISS)]   = -1,
201 	},
202 	[C(OP_PREFETCH)] = {
203 		[C(RESULT_ACCESS)] = -1,
204 		[C(RESULT_MISS)]   = -1,
205 	},
206 },
207 [C(NODE)] = {
208 	[C(OP_READ)] = {
209 		[C(RESULT_ACCESS)] = 0,
210 		[C(RESULT_MISS)]   = 0,
211 	},
212 	[C(OP_WRITE)] = {
213 		[C(RESULT_ACCESS)] = -1,
214 		[C(RESULT_MISS)]   = -1,
215 	},
216 	[C(OP_PREFETCH)] = {
217 		[C(RESULT_ACCESS)] = -1,
218 		[C(RESULT_MISS)]   = -1,
219 	},
220 },
221 };
222 
223 /*
224  * AMD Performance Monitor K7 and later, up to and including Family 16h:
225  */
226 static const u64 amd_perfmon_event_map[PERF_COUNT_HW_MAX] =
227 {
228 	[PERF_COUNT_HW_CPU_CYCLES]		= 0x0076,
229 	[PERF_COUNT_HW_INSTRUCTIONS]		= 0x00c0,
230 	[PERF_COUNT_HW_CACHE_REFERENCES]	= 0x077d,
231 	[PERF_COUNT_HW_CACHE_MISSES]		= 0x077e,
232 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x00c2,
233 	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x00c3,
234 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= 0x00d0, /* "Decoder empty" event */
235 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= 0x00d1, /* "Dispatch stalls" event */
236 };
237 
238 /*
239  * AMD Performance Monitor Family 17h and later:
240  */
241 static const u64 amd_f17h_perfmon_event_map[PERF_COUNT_HW_MAX] =
242 {
243 	[PERF_COUNT_HW_CPU_CYCLES]		= 0x0076,
244 	[PERF_COUNT_HW_INSTRUCTIONS]		= 0x00c0,
245 	[PERF_COUNT_HW_CACHE_REFERENCES]	= 0xff60,
246 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x00c2,
247 	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x00c3,
248 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= 0x0287,
249 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= 0x0187,
250 };
251 
252 static u64 amd_pmu_event_map(int hw_event)
253 {
254 	if (boot_cpu_data.x86 >= 0x17)
255 		return amd_f17h_perfmon_event_map[hw_event];
256 
257 	return amd_perfmon_event_map[hw_event];
258 }
259 
260 /*
261  * Previously calculated offsets
262  */
263 static unsigned int event_offsets[X86_PMC_IDX_MAX] __read_mostly;
264 static unsigned int count_offsets[X86_PMC_IDX_MAX] __read_mostly;
265 
266 /*
267  * Legacy CPUs:
268  *   4 counters starting at 0xc0010000 each offset by 1
269  *
270  * CPUs with core performance counter extensions:
271  *   6 counters starting at 0xc0010200 each offset by 2
272  */
273 static inline int amd_pmu_addr_offset(int index, bool eventsel)
274 {
275 	int offset;
276 
277 	if (!index)
278 		return index;
279 
280 	if (eventsel)
281 		offset = event_offsets[index];
282 	else
283 		offset = count_offsets[index];
284 
285 	if (offset)
286 		return offset;
287 
288 	if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
289 		offset = index;
290 	else
291 		offset = index << 1;
292 
293 	if (eventsel)
294 		event_offsets[index] = offset;
295 	else
296 		count_offsets[index] = offset;
297 
298 	return offset;
299 }
300 
301 static int amd_core_hw_config(struct perf_event *event)
302 {
303 	if (event->attr.exclude_host && event->attr.exclude_guest)
304 		/*
305 		 * When HO == GO == 1 the hardware treats that as GO == HO == 0
306 		 * and will count in both modes. We don't want to count in that
307 		 * case so we emulate no-counting by setting US = OS = 0.
308 		 */
309 		event->hw.config &= ~(ARCH_PERFMON_EVENTSEL_USR |
310 				      ARCH_PERFMON_EVENTSEL_OS);
311 	else if (event->attr.exclude_host)
312 		event->hw.config |= AMD64_EVENTSEL_GUESTONLY;
313 	else if (event->attr.exclude_guest)
314 		event->hw.config |= AMD64_EVENTSEL_HOSTONLY;
315 
316 	return 0;
317 }
318 
319 /*
320  * AMD64 events are detected based on their event codes.
321  */
322 static inline unsigned int amd_get_event_code(struct hw_perf_event *hwc)
323 {
324 	return ((hwc->config >> 24) & 0x0f00) | (hwc->config & 0x00ff);
325 }
326 
327 static inline int amd_is_nb_event(struct hw_perf_event *hwc)
328 {
329 	return (hwc->config & 0xe0) == 0xe0;
330 }
331 
332 static inline int amd_has_nb(struct cpu_hw_events *cpuc)
333 {
334 	struct amd_nb *nb = cpuc->amd_nb;
335 
336 	return nb && nb->nb_id != -1;
337 }
338 
339 static int amd_pmu_hw_config(struct perf_event *event)
340 {
341 	int ret;
342 
343 	/* pass precise event sampling to ibs: */
344 	if (event->attr.precise_ip && get_ibs_caps())
345 		return -ENOENT;
346 
347 	if (has_branch_stack(event))
348 		return -EOPNOTSUPP;
349 
350 	ret = x86_pmu_hw_config(event);
351 	if (ret)
352 		return ret;
353 
354 	if (event->attr.type == PERF_TYPE_RAW)
355 		event->hw.config |= event->attr.config & AMD64_RAW_EVENT_MASK;
356 
357 	return amd_core_hw_config(event);
358 }
359 
360 static void __amd_put_nb_event_constraints(struct cpu_hw_events *cpuc,
361 					   struct perf_event *event)
362 {
363 	struct amd_nb *nb = cpuc->amd_nb;
364 	int i;
365 
366 	/*
367 	 * need to scan whole list because event may not have
368 	 * been assigned during scheduling
369 	 *
370 	 * no race condition possible because event can only
371 	 * be removed on one CPU at a time AND PMU is disabled
372 	 * when we come here
373 	 */
374 	for (i = 0; i < x86_pmu.num_counters; i++) {
375 		if (cmpxchg(nb->owners + i, event, NULL) == event)
376 			break;
377 	}
378 }
379 
380  /*
381   * AMD64 NorthBridge events need special treatment because
382   * counter access needs to be synchronized across all cores
383   * of a package. Refer to BKDG section 3.12
384   *
385   * NB events are events measuring L3 cache, Hypertransport
386   * traffic. They are identified by an event code >= 0xe00.
387   * They measure events on the NorthBride which is shared
388   * by all cores on a package. NB events are counted on a
389   * shared set of counters. When a NB event is programmed
390   * in a counter, the data actually comes from a shared
391   * counter. Thus, access to those counters needs to be
392   * synchronized.
393   *
394   * We implement the synchronization such that no two cores
395   * can be measuring NB events using the same counters. Thus,
396   * we maintain a per-NB allocation table. The available slot
397   * is propagated using the event_constraint structure.
398   *
399   * We provide only one choice for each NB event based on
400   * the fact that only NB events have restrictions. Consequently,
401   * if a counter is available, there is a guarantee the NB event
402   * will be assigned to it. If no slot is available, an empty
403   * constraint is returned and scheduling will eventually fail
404   * for this event.
405   *
406   * Note that all cores attached the same NB compete for the same
407   * counters to host NB events, this is why we use atomic ops. Some
408   * multi-chip CPUs may have more than one NB.
409   *
410   * Given that resources are allocated (cmpxchg), they must be
411   * eventually freed for others to use. This is accomplished by
412   * calling __amd_put_nb_event_constraints()
413   *
414   * Non NB events are not impacted by this restriction.
415   */
416 static struct event_constraint *
417 __amd_get_nb_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
418 			       struct event_constraint *c)
419 {
420 	struct hw_perf_event *hwc = &event->hw;
421 	struct amd_nb *nb = cpuc->amd_nb;
422 	struct perf_event *old;
423 	int idx, new = -1;
424 
425 	if (!c)
426 		c = &unconstrained;
427 
428 	if (cpuc->is_fake)
429 		return c;
430 
431 	/*
432 	 * detect if already present, if so reuse
433 	 *
434 	 * cannot merge with actual allocation
435 	 * because of possible holes
436 	 *
437 	 * event can already be present yet not assigned (in hwc->idx)
438 	 * because of successive calls to x86_schedule_events() from
439 	 * hw_perf_group_sched_in() without hw_perf_enable()
440 	 */
441 	for_each_set_bit(idx, c->idxmsk, x86_pmu.num_counters) {
442 		if (new == -1 || hwc->idx == idx)
443 			/* assign free slot, prefer hwc->idx */
444 			old = cmpxchg(nb->owners + idx, NULL, event);
445 		else if (nb->owners[idx] == event)
446 			/* event already present */
447 			old = event;
448 		else
449 			continue;
450 
451 		if (old && old != event)
452 			continue;
453 
454 		/* reassign to this slot */
455 		if (new != -1)
456 			cmpxchg(nb->owners + new, event, NULL);
457 		new = idx;
458 
459 		/* already present, reuse */
460 		if (old == event)
461 			break;
462 	}
463 
464 	if (new == -1)
465 		return &emptyconstraint;
466 
467 	return &nb->event_constraints[new];
468 }
469 
470 static struct amd_nb *amd_alloc_nb(int cpu)
471 {
472 	struct amd_nb *nb;
473 	int i;
474 
475 	nb = kzalloc_node(sizeof(struct amd_nb), GFP_KERNEL, cpu_to_node(cpu));
476 	if (!nb)
477 		return NULL;
478 
479 	nb->nb_id = -1;
480 
481 	/*
482 	 * initialize all possible NB constraints
483 	 */
484 	for (i = 0; i < x86_pmu.num_counters; i++) {
485 		__set_bit(i, nb->event_constraints[i].idxmsk);
486 		nb->event_constraints[i].weight = 1;
487 	}
488 	return nb;
489 }
490 
491 static int amd_pmu_cpu_prepare(int cpu)
492 {
493 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
494 
495 	WARN_ON_ONCE(cpuc->amd_nb);
496 
497 	if (!x86_pmu.amd_nb_constraints)
498 		return 0;
499 
500 	cpuc->amd_nb = amd_alloc_nb(cpu);
501 	if (!cpuc->amd_nb)
502 		return -ENOMEM;
503 
504 	return 0;
505 }
506 
507 static void amd_pmu_cpu_starting(int cpu)
508 {
509 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
510 	void **onln = &cpuc->kfree_on_online[X86_PERF_KFREE_SHARED];
511 	struct amd_nb *nb;
512 	int i, nb_id;
513 
514 	cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
515 
516 	if (!x86_pmu.amd_nb_constraints)
517 		return;
518 
519 	nb_id = amd_get_nb_id(cpu);
520 	WARN_ON_ONCE(nb_id == BAD_APICID);
521 
522 	for_each_online_cpu(i) {
523 		nb = per_cpu(cpu_hw_events, i).amd_nb;
524 		if (WARN_ON_ONCE(!nb))
525 			continue;
526 
527 		if (nb->nb_id == nb_id) {
528 			*onln = cpuc->amd_nb;
529 			cpuc->amd_nb = nb;
530 			break;
531 		}
532 	}
533 
534 	cpuc->amd_nb->nb_id = nb_id;
535 	cpuc->amd_nb->refcnt++;
536 }
537 
538 static void amd_pmu_cpu_dead(int cpu)
539 {
540 	struct cpu_hw_events *cpuhw;
541 
542 	if (!x86_pmu.amd_nb_constraints)
543 		return;
544 
545 	cpuhw = &per_cpu(cpu_hw_events, cpu);
546 
547 	if (cpuhw->amd_nb) {
548 		struct amd_nb *nb = cpuhw->amd_nb;
549 
550 		if (nb->nb_id == -1 || --nb->refcnt == 0)
551 			kfree(nb);
552 
553 		cpuhw->amd_nb = NULL;
554 	}
555 }
556 
557 /*
558  * When a PMC counter overflows, an NMI is used to process the event and
559  * reset the counter. NMI latency can result in the counter being updated
560  * before the NMI can run, which can result in what appear to be spurious
561  * NMIs. This function is intended to wait for the NMI to run and reset
562  * the counter to avoid possible unhandled NMI messages.
563  */
564 #define OVERFLOW_WAIT_COUNT	50
565 
566 static void amd_pmu_wait_on_overflow(int idx)
567 {
568 	unsigned int i;
569 	u64 counter;
570 
571 	/*
572 	 * Wait for the counter to be reset if it has overflowed. This loop
573 	 * should exit very, very quickly, but just in case, don't wait
574 	 * forever...
575 	 */
576 	for (i = 0; i < OVERFLOW_WAIT_COUNT; i++) {
577 		rdmsrl(x86_pmu_event_addr(idx), counter);
578 		if (counter & (1ULL << (x86_pmu.cntval_bits - 1)))
579 			break;
580 
581 		/* Might be in IRQ context, so can't sleep */
582 		udelay(1);
583 	}
584 }
585 
586 static void amd_pmu_disable_all(void)
587 {
588 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
589 	int idx;
590 
591 	x86_pmu_disable_all();
592 
593 	/*
594 	 * This shouldn't be called from NMI context, but add a safeguard here
595 	 * to return, since if we're in NMI context we can't wait for an NMI
596 	 * to reset an overflowed counter value.
597 	 */
598 	if (in_nmi())
599 		return;
600 
601 	/*
602 	 * Check each counter for overflow and wait for it to be reset by the
603 	 * NMI if it has overflowed. This relies on the fact that all active
604 	 * counters are always enabled when this function is caled and
605 	 * ARCH_PERFMON_EVENTSEL_INT is always set.
606 	 */
607 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
608 		if (!test_bit(idx, cpuc->active_mask))
609 			continue;
610 
611 		amd_pmu_wait_on_overflow(idx);
612 	}
613 }
614 
615 static void amd_pmu_disable_event(struct perf_event *event)
616 {
617 	x86_pmu_disable_event(event);
618 
619 	/*
620 	 * This can be called from NMI context (via x86_pmu_stop). The counter
621 	 * may have overflowed, but either way, we'll never see it get reset
622 	 * by the NMI if we're already in the NMI. And the NMI latency support
623 	 * below will take care of any pending NMI that might have been
624 	 * generated by the overflow.
625 	 */
626 	if (in_nmi())
627 		return;
628 
629 	amd_pmu_wait_on_overflow(event->hw.idx);
630 }
631 
632 /*
633  * Because of NMI latency, if multiple PMC counters are active or other sources
634  * of NMIs are received, the perf NMI handler can handle one or more overflowed
635  * PMC counters outside of the NMI associated with the PMC overflow. If the NMI
636  * doesn't arrive at the LAPIC in time to become a pending NMI, then the kernel
637  * back-to-back NMI support won't be active. This PMC handler needs to take into
638  * account that this can occur, otherwise this could result in unknown NMI
639  * messages being issued. Examples of this is PMC overflow while in the NMI
640  * handler when multiple PMCs are active or PMC overflow while handling some
641  * other source of an NMI.
642  *
643  * Attempt to mitigate this by using the number of active PMCs to determine
644  * whether to return NMI_HANDLED if the perf NMI handler did not handle/reset
645  * any PMCs. The per-CPU perf_nmi_counter variable is set to a minimum of the
646  * number of active PMCs or 2. The value of 2 is used in case an NMI does not
647  * arrive at the LAPIC in time to be collapsed into an already pending NMI.
648  */
649 static int amd_pmu_handle_irq(struct pt_regs *regs)
650 {
651 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
652 	int active, handled;
653 
654 	/*
655 	 * Obtain the active count before calling x86_pmu_handle_irq() since
656 	 * it is possible that x86_pmu_handle_irq() may make a counter
657 	 * inactive (through x86_pmu_stop).
658 	 */
659 	active = __bitmap_weight(cpuc->active_mask, X86_PMC_IDX_MAX);
660 
661 	/* Process any counter overflows */
662 	handled = x86_pmu_handle_irq(regs);
663 
664 	/*
665 	 * If a counter was handled, record the number of possible remaining
666 	 * NMIs that can occur.
667 	 */
668 	if (handled) {
669 		this_cpu_write(perf_nmi_counter,
670 			       min_t(unsigned int, 2, active));
671 
672 		return handled;
673 	}
674 
675 	if (!this_cpu_read(perf_nmi_counter))
676 		return NMI_DONE;
677 
678 	this_cpu_dec(perf_nmi_counter);
679 
680 	return NMI_HANDLED;
681 }
682 
683 static struct event_constraint *
684 amd_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
685 			  struct perf_event *event)
686 {
687 	/*
688 	 * if not NB event or no NB, then no constraints
689 	 */
690 	if (!(amd_has_nb(cpuc) && amd_is_nb_event(&event->hw)))
691 		return &unconstrained;
692 
693 	return __amd_get_nb_event_constraints(cpuc, event, NULL);
694 }
695 
696 static void amd_put_event_constraints(struct cpu_hw_events *cpuc,
697 				      struct perf_event *event)
698 {
699 	if (amd_has_nb(cpuc) && amd_is_nb_event(&event->hw))
700 		__amd_put_nb_event_constraints(cpuc, event);
701 }
702 
703 PMU_FORMAT_ATTR(event,	"config:0-7,32-35");
704 PMU_FORMAT_ATTR(umask,	"config:8-15"	);
705 PMU_FORMAT_ATTR(edge,	"config:18"	);
706 PMU_FORMAT_ATTR(inv,	"config:23"	);
707 PMU_FORMAT_ATTR(cmask,	"config:24-31"	);
708 
709 static struct attribute *amd_format_attr[] = {
710 	&format_attr_event.attr,
711 	&format_attr_umask.attr,
712 	&format_attr_edge.attr,
713 	&format_attr_inv.attr,
714 	&format_attr_cmask.attr,
715 	NULL,
716 };
717 
718 /* AMD Family 15h */
719 
720 #define AMD_EVENT_TYPE_MASK	0x000000F0ULL
721 
722 #define AMD_EVENT_FP		0x00000000ULL ... 0x00000010ULL
723 #define AMD_EVENT_LS		0x00000020ULL ... 0x00000030ULL
724 #define AMD_EVENT_DC		0x00000040ULL ... 0x00000050ULL
725 #define AMD_EVENT_CU		0x00000060ULL ... 0x00000070ULL
726 #define AMD_EVENT_IC_DE		0x00000080ULL ... 0x00000090ULL
727 #define AMD_EVENT_EX_LS		0x000000C0ULL
728 #define AMD_EVENT_DE		0x000000D0ULL
729 #define AMD_EVENT_NB		0x000000E0ULL ... 0x000000F0ULL
730 
731 /*
732  * AMD family 15h event code/PMC mappings:
733  *
734  * type = event_code & 0x0F0:
735  *
736  * 0x000	FP	PERF_CTL[5:3]
737  * 0x010	FP	PERF_CTL[5:3]
738  * 0x020	LS	PERF_CTL[5:0]
739  * 0x030	LS	PERF_CTL[5:0]
740  * 0x040	DC	PERF_CTL[5:0]
741  * 0x050	DC	PERF_CTL[5:0]
742  * 0x060	CU	PERF_CTL[2:0]
743  * 0x070	CU	PERF_CTL[2:0]
744  * 0x080	IC/DE	PERF_CTL[2:0]
745  * 0x090	IC/DE	PERF_CTL[2:0]
746  * 0x0A0	---
747  * 0x0B0	---
748  * 0x0C0	EX/LS	PERF_CTL[5:0]
749  * 0x0D0	DE	PERF_CTL[2:0]
750  * 0x0E0	NB	NB_PERF_CTL[3:0]
751  * 0x0F0	NB	NB_PERF_CTL[3:0]
752  *
753  * Exceptions:
754  *
755  * 0x000	FP	PERF_CTL[3], PERF_CTL[5:3] (*)
756  * 0x003	FP	PERF_CTL[3]
757  * 0x004	FP	PERF_CTL[3], PERF_CTL[5:3] (*)
758  * 0x00B	FP	PERF_CTL[3]
759  * 0x00D	FP	PERF_CTL[3]
760  * 0x023	DE	PERF_CTL[2:0]
761  * 0x02D	LS	PERF_CTL[3]
762  * 0x02E	LS	PERF_CTL[3,0]
763  * 0x031	LS	PERF_CTL[2:0] (**)
764  * 0x043	CU	PERF_CTL[2:0]
765  * 0x045	CU	PERF_CTL[2:0]
766  * 0x046	CU	PERF_CTL[2:0]
767  * 0x054	CU	PERF_CTL[2:0]
768  * 0x055	CU	PERF_CTL[2:0]
769  * 0x08F	IC	PERF_CTL[0]
770  * 0x187	DE	PERF_CTL[0]
771  * 0x188	DE	PERF_CTL[0]
772  * 0x0DB	EX	PERF_CTL[5:0]
773  * 0x0DC	LS	PERF_CTL[5:0]
774  * 0x0DD	LS	PERF_CTL[5:0]
775  * 0x0DE	LS	PERF_CTL[5:0]
776  * 0x0DF	LS	PERF_CTL[5:0]
777  * 0x1C0	EX	PERF_CTL[5:3]
778  * 0x1D6	EX	PERF_CTL[5:0]
779  * 0x1D8	EX	PERF_CTL[5:0]
780  *
781  * (*)  depending on the umask all FPU counters may be used
782  * (**) only one unitmask enabled at a time
783  */
784 
785 static struct event_constraint amd_f15_PMC0  = EVENT_CONSTRAINT(0, 0x01, 0);
786 static struct event_constraint amd_f15_PMC20 = EVENT_CONSTRAINT(0, 0x07, 0);
787 static struct event_constraint amd_f15_PMC3  = EVENT_CONSTRAINT(0, 0x08, 0);
788 static struct event_constraint amd_f15_PMC30 = EVENT_CONSTRAINT_OVERLAP(0, 0x09, 0);
789 static struct event_constraint amd_f15_PMC50 = EVENT_CONSTRAINT(0, 0x3F, 0);
790 static struct event_constraint amd_f15_PMC53 = EVENT_CONSTRAINT(0, 0x38, 0);
791 
792 static struct event_constraint *
793 amd_get_event_constraints_f15h(struct cpu_hw_events *cpuc, int idx,
794 			       struct perf_event *event)
795 {
796 	struct hw_perf_event *hwc = &event->hw;
797 	unsigned int event_code = amd_get_event_code(hwc);
798 
799 	switch (event_code & AMD_EVENT_TYPE_MASK) {
800 	case AMD_EVENT_FP:
801 		switch (event_code) {
802 		case 0x000:
803 			if (!(hwc->config & 0x0000F000ULL))
804 				break;
805 			if (!(hwc->config & 0x00000F00ULL))
806 				break;
807 			return &amd_f15_PMC3;
808 		case 0x004:
809 			if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
810 				break;
811 			return &amd_f15_PMC3;
812 		case 0x003:
813 		case 0x00B:
814 		case 0x00D:
815 			return &amd_f15_PMC3;
816 		}
817 		return &amd_f15_PMC53;
818 	case AMD_EVENT_LS:
819 	case AMD_EVENT_DC:
820 	case AMD_EVENT_EX_LS:
821 		switch (event_code) {
822 		case 0x023:
823 		case 0x043:
824 		case 0x045:
825 		case 0x046:
826 		case 0x054:
827 		case 0x055:
828 			return &amd_f15_PMC20;
829 		case 0x02D:
830 			return &amd_f15_PMC3;
831 		case 0x02E:
832 			return &amd_f15_PMC30;
833 		case 0x031:
834 			if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
835 				return &amd_f15_PMC20;
836 			return &emptyconstraint;
837 		case 0x1C0:
838 			return &amd_f15_PMC53;
839 		default:
840 			return &amd_f15_PMC50;
841 		}
842 	case AMD_EVENT_CU:
843 	case AMD_EVENT_IC_DE:
844 	case AMD_EVENT_DE:
845 		switch (event_code) {
846 		case 0x08F:
847 		case 0x187:
848 		case 0x188:
849 			return &amd_f15_PMC0;
850 		case 0x0DB ... 0x0DF:
851 		case 0x1D6:
852 		case 0x1D8:
853 			return &amd_f15_PMC50;
854 		default:
855 			return &amd_f15_PMC20;
856 		}
857 	case AMD_EVENT_NB:
858 		/* moved to uncore.c */
859 		return &emptyconstraint;
860 	default:
861 		return &emptyconstraint;
862 	}
863 }
864 
865 static ssize_t amd_event_sysfs_show(char *page, u64 config)
866 {
867 	u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT) |
868 		    (config & AMD64_EVENTSEL_EVENT) >> 24;
869 
870 	return x86_event_sysfs_show(page, config, event);
871 }
872 
873 static __initconst const struct x86_pmu amd_pmu = {
874 	.name			= "AMD",
875 	.handle_irq		= amd_pmu_handle_irq,
876 	.disable_all		= amd_pmu_disable_all,
877 	.enable_all		= x86_pmu_enable_all,
878 	.enable			= x86_pmu_enable_event,
879 	.disable		= amd_pmu_disable_event,
880 	.hw_config		= amd_pmu_hw_config,
881 	.schedule_events	= x86_schedule_events,
882 	.eventsel		= MSR_K7_EVNTSEL0,
883 	.perfctr		= MSR_K7_PERFCTR0,
884 	.addr_offset            = amd_pmu_addr_offset,
885 	.event_map		= amd_pmu_event_map,
886 	.max_events		= ARRAY_SIZE(amd_perfmon_event_map),
887 	.num_counters		= AMD64_NUM_COUNTERS,
888 	.cntval_bits		= 48,
889 	.cntval_mask		= (1ULL << 48) - 1,
890 	.apic			= 1,
891 	/* use highest bit to detect overflow */
892 	.max_period		= (1ULL << 47) - 1,
893 	.get_event_constraints	= amd_get_event_constraints,
894 	.put_event_constraints	= amd_put_event_constraints,
895 
896 	.format_attrs		= amd_format_attr,
897 	.events_sysfs_show	= amd_event_sysfs_show,
898 
899 	.cpu_prepare		= amd_pmu_cpu_prepare,
900 	.cpu_starting		= amd_pmu_cpu_starting,
901 	.cpu_dead		= amd_pmu_cpu_dead,
902 
903 	.amd_nb_constraints	= 1,
904 };
905 
906 static int __init amd_core_pmu_init(void)
907 {
908 	if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
909 		return 0;
910 
911 	switch (boot_cpu_data.x86) {
912 	case 0x15:
913 		pr_cont("Fam15h ");
914 		x86_pmu.get_event_constraints = amd_get_event_constraints_f15h;
915 		break;
916 	case 0x17:
917 		pr_cont("Fam17h ");
918 		/*
919 		 * In family 17h, there are no event constraints in the PMC hardware.
920 		 * We fallback to using default amd_get_event_constraints.
921 		 */
922 		break;
923 	case 0x18:
924 		pr_cont("Fam18h ");
925 		/* Using default amd_get_event_constraints. */
926 		break;
927 	default:
928 		pr_err("core perfctr but no constraints; unknown hardware!\n");
929 		return -ENODEV;
930 	}
931 
932 	/*
933 	 * If core performance counter extensions exists, we must use
934 	 * MSR_F15H_PERF_CTL/MSR_F15H_PERF_CTR msrs. See also
935 	 * amd_pmu_addr_offset().
936 	 */
937 	x86_pmu.eventsel	= MSR_F15H_PERF_CTL;
938 	x86_pmu.perfctr		= MSR_F15H_PERF_CTR;
939 	x86_pmu.num_counters	= AMD64_NUM_COUNTERS_CORE;
940 	/*
941 	 * AMD Core perfctr has separate MSRs for the NB events, see
942 	 * the amd/uncore.c driver.
943 	 */
944 	x86_pmu.amd_nb_constraints = 0;
945 
946 	pr_cont("core perfctr, ");
947 	return 0;
948 }
949 
950 __init int amd_pmu_init(void)
951 {
952 	int ret;
953 
954 	/* Performance-monitoring supported from K7 and later: */
955 	if (boot_cpu_data.x86 < 6)
956 		return -ENODEV;
957 
958 	x86_pmu = amd_pmu;
959 
960 	ret = amd_core_pmu_init();
961 	if (ret)
962 		return ret;
963 
964 	if (num_possible_cpus() == 1) {
965 		/*
966 		 * No point in allocating data structures to serialize
967 		 * against other CPUs, when there is only the one CPU.
968 		 */
969 		x86_pmu.amd_nb_constraints = 0;
970 	}
971 
972 	if (boot_cpu_data.x86 >= 0x17)
973 		memcpy(hw_cache_event_ids, amd_hw_cache_event_ids_f17h, sizeof(hw_cache_event_ids));
974 	else
975 		memcpy(hw_cache_event_ids, amd_hw_cache_event_ids, sizeof(hw_cache_event_ids));
976 
977 	return 0;
978 }
979 
980 void amd_pmu_enable_virt(void)
981 {
982 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
983 
984 	cpuc->perf_ctr_virt_mask = 0;
985 
986 	/* Reload all events */
987 	amd_pmu_disable_all();
988 	x86_pmu_enable_all(0);
989 }
990 EXPORT_SYMBOL_GPL(amd_pmu_enable_virt);
991 
992 void amd_pmu_disable_virt(void)
993 {
994 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
995 
996 	/*
997 	 * We only mask out the Host-only bit so that host-only counting works
998 	 * when SVM is disabled. If someone sets up a guest-only counter when
999 	 * SVM is disabled the Guest-only bits still gets set and the counter
1000 	 * will not count anything.
1001 	 */
1002 	cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
1003 
1004 	/* Reload all events */
1005 	amd_pmu_disable_all();
1006 	x86_pmu_enable_all(0);
1007 }
1008 EXPORT_SYMBOL_GPL(amd_pmu_disable_virt);
1009