xref: /openbmc/linux/arch/x86/events/amd/core.c (revision 20e2fc42)
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 static __initconst const u64 amd_hw_cache_event_ids
18 				[PERF_COUNT_HW_CACHE_MAX]
19 				[PERF_COUNT_HW_CACHE_OP_MAX]
20 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
21 {
22  [ C(L1D) ] = {
23 	[ C(OP_READ) ] = {
24 		[ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses        */
25 		[ C(RESULT_MISS)   ] = 0x0141, /* Data Cache Misses          */
26 	},
27 	[ C(OP_WRITE) ] = {
28 		[ C(RESULT_ACCESS) ] = 0,
29 		[ C(RESULT_MISS)   ] = 0,
30 	},
31 	[ C(OP_PREFETCH) ] = {
32 		[ C(RESULT_ACCESS) ] = 0x0267, /* Data Prefetcher :attempts  */
33 		[ C(RESULT_MISS)   ] = 0x0167, /* Data Prefetcher :cancelled */
34 	},
35  },
36  [ C(L1I ) ] = {
37 	[ C(OP_READ) ] = {
38 		[ C(RESULT_ACCESS) ] = 0x0080, /* Instruction cache fetches  */
39 		[ C(RESULT_MISS)   ] = 0x0081, /* Instruction cache misses   */
40 	},
41 	[ C(OP_WRITE) ] = {
42 		[ C(RESULT_ACCESS) ] = -1,
43 		[ C(RESULT_MISS)   ] = -1,
44 	},
45 	[ C(OP_PREFETCH) ] = {
46 		[ C(RESULT_ACCESS) ] = 0x014B, /* Prefetch Instructions :Load */
47 		[ C(RESULT_MISS)   ] = 0,
48 	},
49  },
50  [ C(LL  ) ] = {
51 	[ C(OP_READ) ] = {
52 		[ C(RESULT_ACCESS) ] = 0x037D, /* Requests to L2 Cache :IC+DC */
53 		[ C(RESULT_MISS)   ] = 0x037E, /* L2 Cache Misses : IC+DC     */
54 	},
55 	[ C(OP_WRITE) ] = {
56 		[ C(RESULT_ACCESS) ] = 0x017F, /* L2 Fill/Writeback           */
57 		[ C(RESULT_MISS)   ] = 0,
58 	},
59 	[ C(OP_PREFETCH) ] = {
60 		[ C(RESULT_ACCESS) ] = 0,
61 		[ C(RESULT_MISS)   ] = 0,
62 	},
63  },
64  [ C(DTLB) ] = {
65 	[ C(OP_READ) ] = {
66 		[ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses        */
67 		[ C(RESULT_MISS)   ] = 0x0746, /* L1_DTLB_AND_L2_DLTB_MISS.ALL */
68 	},
69 	[ C(OP_WRITE) ] = {
70 		[ C(RESULT_ACCESS) ] = 0,
71 		[ C(RESULT_MISS)   ] = 0,
72 	},
73 	[ C(OP_PREFETCH) ] = {
74 		[ C(RESULT_ACCESS) ] = 0,
75 		[ C(RESULT_MISS)   ] = 0,
76 	},
77  },
78  [ C(ITLB) ] = {
79 	[ C(OP_READ) ] = {
80 		[ C(RESULT_ACCESS) ] = 0x0080, /* Instruction fecthes        */
81 		[ C(RESULT_MISS)   ] = 0x0385, /* L1_ITLB_AND_L2_ITLB_MISS.ALL */
82 	},
83 	[ C(OP_WRITE) ] = {
84 		[ C(RESULT_ACCESS) ] = -1,
85 		[ C(RESULT_MISS)   ] = -1,
86 	},
87 	[ C(OP_PREFETCH) ] = {
88 		[ C(RESULT_ACCESS) ] = -1,
89 		[ C(RESULT_MISS)   ] = -1,
90 	},
91  },
92  [ C(BPU ) ] = {
93 	[ C(OP_READ) ] = {
94 		[ C(RESULT_ACCESS) ] = 0x00c2, /* Retired Branch Instr.      */
95 		[ C(RESULT_MISS)   ] = 0x00c3, /* Retired Mispredicted BI    */
96 	},
97 	[ C(OP_WRITE) ] = {
98 		[ C(RESULT_ACCESS) ] = -1,
99 		[ C(RESULT_MISS)   ] = -1,
100 	},
101 	[ C(OP_PREFETCH) ] = {
102 		[ C(RESULT_ACCESS) ] = -1,
103 		[ C(RESULT_MISS)   ] = -1,
104 	},
105  },
106  [ C(NODE) ] = {
107 	[ C(OP_READ) ] = {
108 		[ C(RESULT_ACCESS) ] = 0xb8e9, /* CPU Request to Memory, l+r */
109 		[ C(RESULT_MISS)   ] = 0x98e9, /* CPU Request to Memory, r   */
110 	},
111 	[ C(OP_WRITE) ] = {
112 		[ C(RESULT_ACCESS) ] = -1,
113 		[ C(RESULT_MISS)   ] = -1,
114 	},
115 	[ C(OP_PREFETCH) ] = {
116 		[ C(RESULT_ACCESS) ] = -1,
117 		[ C(RESULT_MISS)   ] = -1,
118 	},
119  },
120 };
121 
122 static __initconst const u64 amd_hw_cache_event_ids_f17h
123 				[PERF_COUNT_HW_CACHE_MAX]
124 				[PERF_COUNT_HW_CACHE_OP_MAX]
125 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
126 [C(L1D)] = {
127 	[C(OP_READ)] = {
128 		[C(RESULT_ACCESS)] = 0x0040, /* Data Cache Accesses */
129 		[C(RESULT_MISS)]   = 0xc860, /* L2$ access from DC Miss */
130 	},
131 	[C(OP_WRITE)] = {
132 		[C(RESULT_ACCESS)] = 0,
133 		[C(RESULT_MISS)]   = 0,
134 	},
135 	[C(OP_PREFETCH)] = {
136 		[C(RESULT_ACCESS)] = 0xff5a, /* h/w prefetch DC Fills */
137 		[C(RESULT_MISS)]   = 0,
138 	},
139 },
140 [C(L1I)] = {
141 	[C(OP_READ)] = {
142 		[C(RESULT_ACCESS)] = 0x0080, /* Instruction cache fetches  */
143 		[C(RESULT_MISS)]   = 0x0081, /* Instruction cache misses   */
144 	},
145 	[C(OP_WRITE)] = {
146 		[C(RESULT_ACCESS)] = -1,
147 		[C(RESULT_MISS)]   = -1,
148 	},
149 	[C(OP_PREFETCH)] = {
150 		[C(RESULT_ACCESS)] = 0,
151 		[C(RESULT_MISS)]   = 0,
152 	},
153 },
154 [C(LL)] = {
155 	[C(OP_READ)] = {
156 		[C(RESULT_ACCESS)] = 0,
157 		[C(RESULT_MISS)]   = 0,
158 	},
159 	[C(OP_WRITE)] = {
160 		[C(RESULT_ACCESS)] = 0,
161 		[C(RESULT_MISS)]   = 0,
162 	},
163 	[C(OP_PREFETCH)] = {
164 		[C(RESULT_ACCESS)] = 0,
165 		[C(RESULT_MISS)]   = 0,
166 	},
167 },
168 [C(DTLB)] = {
169 	[C(OP_READ)] = {
170 		[C(RESULT_ACCESS)] = 0xff45, /* All L2 DTLB accesses */
171 		[C(RESULT_MISS)]   = 0xf045, /* L2 DTLB misses (PT walks) */
172 	},
173 	[C(OP_WRITE)] = {
174 		[C(RESULT_ACCESS)] = 0,
175 		[C(RESULT_MISS)]   = 0,
176 	},
177 	[C(OP_PREFETCH)] = {
178 		[C(RESULT_ACCESS)] = 0,
179 		[C(RESULT_MISS)]   = 0,
180 	},
181 },
182 [C(ITLB)] = {
183 	[C(OP_READ)] = {
184 		[C(RESULT_ACCESS)] = 0x0084, /* L1 ITLB misses, L2 ITLB hits */
185 		[C(RESULT_MISS)]   = 0xff85, /* L1 ITLB misses, L2 misses */
186 	},
187 	[C(OP_WRITE)] = {
188 		[C(RESULT_ACCESS)] = -1,
189 		[C(RESULT_MISS)]   = -1,
190 	},
191 	[C(OP_PREFETCH)] = {
192 		[C(RESULT_ACCESS)] = -1,
193 		[C(RESULT_MISS)]   = -1,
194 	},
195 },
196 [C(BPU)] = {
197 	[C(OP_READ)] = {
198 		[C(RESULT_ACCESS)] = 0x00c2, /* Retired Branch Instr.      */
199 		[C(RESULT_MISS)]   = 0x00c3, /* Retired Mispredicted BI    */
200 	},
201 	[C(OP_WRITE)] = {
202 		[C(RESULT_ACCESS)] = -1,
203 		[C(RESULT_MISS)]   = -1,
204 	},
205 	[C(OP_PREFETCH)] = {
206 		[C(RESULT_ACCESS)] = -1,
207 		[C(RESULT_MISS)]   = -1,
208 	},
209 },
210 [C(NODE)] = {
211 	[C(OP_READ)] = {
212 		[C(RESULT_ACCESS)] = 0,
213 		[C(RESULT_MISS)]   = 0,
214 	},
215 	[C(OP_WRITE)] = {
216 		[C(RESULT_ACCESS)] = -1,
217 		[C(RESULT_MISS)]   = -1,
218 	},
219 	[C(OP_PREFETCH)] = {
220 		[C(RESULT_ACCESS)] = -1,
221 		[C(RESULT_MISS)]   = -1,
222 	},
223 },
224 };
225 
226 /*
227  * AMD Performance Monitor K7 and later, up to and including Family 16h:
228  */
229 static const u64 amd_perfmon_event_map[PERF_COUNT_HW_MAX] =
230 {
231 	[PERF_COUNT_HW_CPU_CYCLES]		= 0x0076,
232 	[PERF_COUNT_HW_INSTRUCTIONS]		= 0x00c0,
233 	[PERF_COUNT_HW_CACHE_REFERENCES]	= 0x077d,
234 	[PERF_COUNT_HW_CACHE_MISSES]		= 0x077e,
235 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x00c2,
236 	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x00c3,
237 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= 0x00d0, /* "Decoder empty" event */
238 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= 0x00d1, /* "Dispatch stalls" event */
239 };
240 
241 /*
242  * AMD Performance Monitor Family 17h and later:
243  */
244 static const u64 amd_f17h_perfmon_event_map[PERF_COUNT_HW_MAX] =
245 {
246 	[PERF_COUNT_HW_CPU_CYCLES]		= 0x0076,
247 	[PERF_COUNT_HW_INSTRUCTIONS]		= 0x00c0,
248 	[PERF_COUNT_HW_CACHE_REFERENCES]	= 0xff60,
249 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x00c2,
250 	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x00c3,
251 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= 0x0287,
252 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= 0x0187,
253 };
254 
255 static u64 amd_pmu_event_map(int hw_event)
256 {
257 	if (boot_cpu_data.x86 >= 0x17)
258 		return amd_f17h_perfmon_event_map[hw_event];
259 
260 	return amd_perfmon_event_map[hw_event];
261 }
262 
263 /*
264  * Previously calculated offsets
265  */
266 static unsigned int event_offsets[X86_PMC_IDX_MAX] __read_mostly;
267 static unsigned int count_offsets[X86_PMC_IDX_MAX] __read_mostly;
268 
269 /*
270  * Legacy CPUs:
271  *   4 counters starting at 0xc0010000 each offset by 1
272  *
273  * CPUs with core performance counter extensions:
274  *   6 counters starting at 0xc0010200 each offset by 2
275  */
276 static inline int amd_pmu_addr_offset(int index, bool eventsel)
277 {
278 	int offset;
279 
280 	if (!index)
281 		return index;
282 
283 	if (eventsel)
284 		offset = event_offsets[index];
285 	else
286 		offset = count_offsets[index];
287 
288 	if (offset)
289 		return offset;
290 
291 	if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
292 		offset = index;
293 	else
294 		offset = index << 1;
295 
296 	if (eventsel)
297 		event_offsets[index] = offset;
298 	else
299 		count_offsets[index] = offset;
300 
301 	return offset;
302 }
303 
304 static int amd_core_hw_config(struct perf_event *event)
305 {
306 	if (event->attr.exclude_host && event->attr.exclude_guest)
307 		/*
308 		 * When HO == GO == 1 the hardware treats that as GO == HO == 0
309 		 * and will count in both modes. We don't want to count in that
310 		 * case so we emulate no-counting by setting US = OS = 0.
311 		 */
312 		event->hw.config &= ~(ARCH_PERFMON_EVENTSEL_USR |
313 				      ARCH_PERFMON_EVENTSEL_OS);
314 	else if (event->attr.exclude_host)
315 		event->hw.config |= AMD64_EVENTSEL_GUESTONLY;
316 	else if (event->attr.exclude_guest)
317 		event->hw.config |= AMD64_EVENTSEL_HOSTONLY;
318 
319 	return 0;
320 }
321 
322 /*
323  * AMD64 events are detected based on their event codes.
324  */
325 static inline unsigned int amd_get_event_code(struct hw_perf_event *hwc)
326 {
327 	return ((hwc->config >> 24) & 0x0f00) | (hwc->config & 0x00ff);
328 }
329 
330 static inline int amd_is_nb_event(struct hw_perf_event *hwc)
331 {
332 	return (hwc->config & 0xe0) == 0xe0;
333 }
334 
335 static inline int amd_has_nb(struct cpu_hw_events *cpuc)
336 {
337 	struct amd_nb *nb = cpuc->amd_nb;
338 
339 	return nb && nb->nb_id != -1;
340 }
341 
342 static int amd_pmu_hw_config(struct perf_event *event)
343 {
344 	int ret;
345 
346 	/* pass precise event sampling to ibs: */
347 	if (event->attr.precise_ip && get_ibs_caps())
348 		return -ENOENT;
349 
350 	if (has_branch_stack(event))
351 		return -EOPNOTSUPP;
352 
353 	ret = x86_pmu_hw_config(event);
354 	if (ret)
355 		return ret;
356 
357 	if (event->attr.type == PERF_TYPE_RAW)
358 		event->hw.config |= event->attr.config & AMD64_RAW_EVENT_MASK;
359 
360 	return amd_core_hw_config(event);
361 }
362 
363 static void __amd_put_nb_event_constraints(struct cpu_hw_events *cpuc,
364 					   struct perf_event *event)
365 {
366 	struct amd_nb *nb = cpuc->amd_nb;
367 	int i;
368 
369 	/*
370 	 * need to scan whole list because event may not have
371 	 * been assigned during scheduling
372 	 *
373 	 * no race condition possible because event can only
374 	 * be removed on one CPU at a time AND PMU is disabled
375 	 * when we come here
376 	 */
377 	for (i = 0; i < x86_pmu.num_counters; i++) {
378 		if (cmpxchg(nb->owners + i, event, NULL) == event)
379 			break;
380 	}
381 }
382 
383  /*
384   * AMD64 NorthBridge events need special treatment because
385   * counter access needs to be synchronized across all cores
386   * of a package. Refer to BKDG section 3.12
387   *
388   * NB events are events measuring L3 cache, Hypertransport
389   * traffic. They are identified by an event code >= 0xe00.
390   * They measure events on the NorthBride which is shared
391   * by all cores on a package. NB events are counted on a
392   * shared set of counters. When a NB event is programmed
393   * in a counter, the data actually comes from a shared
394   * counter. Thus, access to those counters needs to be
395   * synchronized.
396   *
397   * We implement the synchronization such that no two cores
398   * can be measuring NB events using the same counters. Thus,
399   * we maintain a per-NB allocation table. The available slot
400   * is propagated using the event_constraint structure.
401   *
402   * We provide only one choice for each NB event based on
403   * the fact that only NB events have restrictions. Consequently,
404   * if a counter is available, there is a guarantee the NB event
405   * will be assigned to it. If no slot is available, an empty
406   * constraint is returned and scheduling will eventually fail
407   * for this event.
408   *
409   * Note that all cores attached the same NB compete for the same
410   * counters to host NB events, this is why we use atomic ops. Some
411   * multi-chip CPUs may have more than one NB.
412   *
413   * Given that resources are allocated (cmpxchg), they must be
414   * eventually freed for others to use. This is accomplished by
415   * calling __amd_put_nb_event_constraints()
416   *
417   * Non NB events are not impacted by this restriction.
418   */
419 static struct event_constraint *
420 __amd_get_nb_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
421 			       struct event_constraint *c)
422 {
423 	struct hw_perf_event *hwc = &event->hw;
424 	struct amd_nb *nb = cpuc->amd_nb;
425 	struct perf_event *old;
426 	int idx, new = -1;
427 
428 	if (!c)
429 		c = &unconstrained;
430 
431 	if (cpuc->is_fake)
432 		return c;
433 
434 	/*
435 	 * detect if already present, if so reuse
436 	 *
437 	 * cannot merge with actual allocation
438 	 * because of possible holes
439 	 *
440 	 * event can already be present yet not assigned (in hwc->idx)
441 	 * because of successive calls to x86_schedule_events() from
442 	 * hw_perf_group_sched_in() without hw_perf_enable()
443 	 */
444 	for_each_set_bit(idx, c->idxmsk, x86_pmu.num_counters) {
445 		if (new == -1 || hwc->idx == idx)
446 			/* assign free slot, prefer hwc->idx */
447 			old = cmpxchg(nb->owners + idx, NULL, event);
448 		else if (nb->owners[idx] == event)
449 			/* event already present */
450 			old = event;
451 		else
452 			continue;
453 
454 		if (old && old != event)
455 			continue;
456 
457 		/* reassign to this slot */
458 		if (new != -1)
459 			cmpxchg(nb->owners + new, event, NULL);
460 		new = idx;
461 
462 		/* already present, reuse */
463 		if (old == event)
464 			break;
465 	}
466 
467 	if (new == -1)
468 		return &emptyconstraint;
469 
470 	return &nb->event_constraints[new];
471 }
472 
473 static struct amd_nb *amd_alloc_nb(int cpu)
474 {
475 	struct amd_nb *nb;
476 	int i;
477 
478 	nb = kzalloc_node(sizeof(struct amd_nb), GFP_KERNEL, cpu_to_node(cpu));
479 	if (!nb)
480 		return NULL;
481 
482 	nb->nb_id = -1;
483 
484 	/*
485 	 * initialize all possible NB constraints
486 	 */
487 	for (i = 0; i < x86_pmu.num_counters; i++) {
488 		__set_bit(i, nb->event_constraints[i].idxmsk);
489 		nb->event_constraints[i].weight = 1;
490 	}
491 	return nb;
492 }
493 
494 static int amd_pmu_cpu_prepare(int cpu)
495 {
496 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
497 
498 	WARN_ON_ONCE(cpuc->amd_nb);
499 
500 	if (!x86_pmu.amd_nb_constraints)
501 		return 0;
502 
503 	cpuc->amd_nb = amd_alloc_nb(cpu);
504 	if (!cpuc->amd_nb)
505 		return -ENOMEM;
506 
507 	return 0;
508 }
509 
510 static void amd_pmu_cpu_starting(int cpu)
511 {
512 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
513 	void **onln = &cpuc->kfree_on_online[X86_PERF_KFREE_SHARED];
514 	struct amd_nb *nb;
515 	int i, nb_id;
516 
517 	cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
518 
519 	if (!x86_pmu.amd_nb_constraints)
520 		return;
521 
522 	nb_id = amd_get_nb_id(cpu);
523 	WARN_ON_ONCE(nb_id == BAD_APICID);
524 
525 	for_each_online_cpu(i) {
526 		nb = per_cpu(cpu_hw_events, i).amd_nb;
527 		if (WARN_ON_ONCE(!nb))
528 			continue;
529 
530 		if (nb->nb_id == nb_id) {
531 			*onln = cpuc->amd_nb;
532 			cpuc->amd_nb = nb;
533 			break;
534 		}
535 	}
536 
537 	cpuc->amd_nb->nb_id = nb_id;
538 	cpuc->amd_nb->refcnt++;
539 }
540 
541 static void amd_pmu_cpu_dead(int cpu)
542 {
543 	struct cpu_hw_events *cpuhw;
544 
545 	if (!x86_pmu.amd_nb_constraints)
546 		return;
547 
548 	cpuhw = &per_cpu(cpu_hw_events, cpu);
549 
550 	if (cpuhw->amd_nb) {
551 		struct amd_nb *nb = cpuhw->amd_nb;
552 
553 		if (nb->nb_id == -1 || --nb->refcnt == 0)
554 			kfree(nb);
555 
556 		cpuhw->amd_nb = NULL;
557 	}
558 }
559 
560 /*
561  * When a PMC counter overflows, an NMI is used to process the event and
562  * reset the counter. NMI latency can result in the counter being updated
563  * before the NMI can run, which can result in what appear to be spurious
564  * NMIs. This function is intended to wait for the NMI to run and reset
565  * the counter to avoid possible unhandled NMI messages.
566  */
567 #define OVERFLOW_WAIT_COUNT	50
568 
569 static void amd_pmu_wait_on_overflow(int idx)
570 {
571 	unsigned int i;
572 	u64 counter;
573 
574 	/*
575 	 * Wait for the counter to be reset if it has overflowed. This loop
576 	 * should exit very, very quickly, but just in case, don't wait
577 	 * forever...
578 	 */
579 	for (i = 0; i < OVERFLOW_WAIT_COUNT; i++) {
580 		rdmsrl(x86_pmu_event_addr(idx), counter);
581 		if (counter & (1ULL << (x86_pmu.cntval_bits - 1)))
582 			break;
583 
584 		/* Might be in IRQ context, so can't sleep */
585 		udelay(1);
586 	}
587 }
588 
589 static void amd_pmu_disable_all(void)
590 {
591 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
592 	int idx;
593 
594 	x86_pmu_disable_all();
595 
596 	/*
597 	 * This shouldn't be called from NMI context, but add a safeguard here
598 	 * to return, since if we're in NMI context we can't wait for an NMI
599 	 * to reset an overflowed counter value.
600 	 */
601 	if (in_nmi())
602 		return;
603 
604 	/*
605 	 * Check each counter for overflow and wait for it to be reset by the
606 	 * NMI if it has overflowed. This relies on the fact that all active
607 	 * counters are always enabled when this function is caled and
608 	 * ARCH_PERFMON_EVENTSEL_INT is always set.
609 	 */
610 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
611 		if (!test_bit(idx, cpuc->active_mask))
612 			continue;
613 
614 		amd_pmu_wait_on_overflow(idx);
615 	}
616 }
617 
618 static void amd_pmu_disable_event(struct perf_event *event)
619 {
620 	x86_pmu_disable_event(event);
621 
622 	/*
623 	 * This can be called from NMI context (via x86_pmu_stop). The counter
624 	 * may have overflowed, but either way, we'll never see it get reset
625 	 * by the NMI if we're already in the NMI. And the NMI latency support
626 	 * below will take care of any pending NMI that might have been
627 	 * generated by the overflow.
628 	 */
629 	if (in_nmi())
630 		return;
631 
632 	amd_pmu_wait_on_overflow(event->hw.idx);
633 }
634 
635 /*
636  * Because of NMI latency, if multiple PMC counters are active or other sources
637  * of NMIs are received, the perf NMI handler can handle one or more overflowed
638  * PMC counters outside of the NMI associated with the PMC overflow. If the NMI
639  * doesn't arrive at the LAPIC in time to become a pending NMI, then the kernel
640  * back-to-back NMI support won't be active. This PMC handler needs to take into
641  * account that this can occur, otherwise this could result in unknown NMI
642  * messages being issued. Examples of this is PMC overflow while in the NMI
643  * handler when multiple PMCs are active or PMC overflow while handling some
644  * other source of an NMI.
645  *
646  * Attempt to mitigate this by creating an NMI window in which un-handled NMIs
647  * received during this window will be claimed. This prevents extending the
648  * window past when it is possible that latent NMIs should be received. The
649  * per-CPU perf_nmi_tstamp will be set to the window end time whenever perf has
650  * handled a counter. When an un-handled NMI is received, it will be claimed
651  * only if arriving within that window.
652  */
653 static int amd_pmu_handle_irq(struct pt_regs *regs)
654 {
655 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
656 	int active, handled;
657 
658 	/*
659 	 * Obtain the active count before calling x86_pmu_handle_irq() since
660 	 * it is possible that x86_pmu_handle_irq() may make a counter
661 	 * inactive (through x86_pmu_stop).
662 	 */
663 	active = __bitmap_weight(cpuc->active_mask, X86_PMC_IDX_MAX);
664 
665 	/* Process any counter overflows */
666 	handled = x86_pmu_handle_irq(regs);
667 
668 	/*
669 	 * If a counter was handled, record a timestamp such that un-handled
670 	 * NMIs will be claimed if arriving within that window.
671 	 */
672 	if (handled) {
673 		this_cpu_write(perf_nmi_tstamp,
674 			       jiffies + perf_nmi_window);
675 
676 		return handled;
677 	}
678 
679 	if (time_after(jiffies, this_cpu_read(perf_nmi_tstamp)))
680 		return NMI_DONE;
681 
682 	return NMI_HANDLED;
683 }
684 
685 static struct event_constraint *
686 amd_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
687 			  struct perf_event *event)
688 {
689 	/*
690 	 * if not NB event or no NB, then no constraints
691 	 */
692 	if (!(amd_has_nb(cpuc) && amd_is_nb_event(&event->hw)))
693 		return &unconstrained;
694 
695 	return __amd_get_nb_event_constraints(cpuc, event, NULL);
696 }
697 
698 static void amd_put_event_constraints(struct cpu_hw_events *cpuc,
699 				      struct perf_event *event)
700 {
701 	if (amd_has_nb(cpuc) && amd_is_nb_event(&event->hw))
702 		__amd_put_nb_event_constraints(cpuc, event);
703 }
704 
705 PMU_FORMAT_ATTR(event,	"config:0-7,32-35");
706 PMU_FORMAT_ATTR(umask,	"config:8-15"	);
707 PMU_FORMAT_ATTR(edge,	"config:18"	);
708 PMU_FORMAT_ATTR(inv,	"config:23"	);
709 PMU_FORMAT_ATTR(cmask,	"config:24-31"	);
710 
711 static struct attribute *amd_format_attr[] = {
712 	&format_attr_event.attr,
713 	&format_attr_umask.attr,
714 	&format_attr_edge.attr,
715 	&format_attr_inv.attr,
716 	&format_attr_cmask.attr,
717 	NULL,
718 };
719 
720 /* AMD Family 15h */
721 
722 #define AMD_EVENT_TYPE_MASK	0x000000F0ULL
723 
724 #define AMD_EVENT_FP		0x00000000ULL ... 0x00000010ULL
725 #define AMD_EVENT_LS		0x00000020ULL ... 0x00000030ULL
726 #define AMD_EVENT_DC		0x00000040ULL ... 0x00000050ULL
727 #define AMD_EVENT_CU		0x00000060ULL ... 0x00000070ULL
728 #define AMD_EVENT_IC_DE		0x00000080ULL ... 0x00000090ULL
729 #define AMD_EVENT_EX_LS		0x000000C0ULL
730 #define AMD_EVENT_DE		0x000000D0ULL
731 #define AMD_EVENT_NB		0x000000E0ULL ... 0x000000F0ULL
732 
733 /*
734  * AMD family 15h event code/PMC mappings:
735  *
736  * type = event_code & 0x0F0:
737  *
738  * 0x000	FP	PERF_CTL[5:3]
739  * 0x010	FP	PERF_CTL[5:3]
740  * 0x020	LS	PERF_CTL[5:0]
741  * 0x030	LS	PERF_CTL[5:0]
742  * 0x040	DC	PERF_CTL[5:0]
743  * 0x050	DC	PERF_CTL[5:0]
744  * 0x060	CU	PERF_CTL[2:0]
745  * 0x070	CU	PERF_CTL[2:0]
746  * 0x080	IC/DE	PERF_CTL[2:0]
747  * 0x090	IC/DE	PERF_CTL[2:0]
748  * 0x0A0	---
749  * 0x0B0	---
750  * 0x0C0	EX/LS	PERF_CTL[5:0]
751  * 0x0D0	DE	PERF_CTL[2:0]
752  * 0x0E0	NB	NB_PERF_CTL[3:0]
753  * 0x0F0	NB	NB_PERF_CTL[3:0]
754  *
755  * Exceptions:
756  *
757  * 0x000	FP	PERF_CTL[3], PERF_CTL[5:3] (*)
758  * 0x003	FP	PERF_CTL[3]
759  * 0x004	FP	PERF_CTL[3], PERF_CTL[5:3] (*)
760  * 0x00B	FP	PERF_CTL[3]
761  * 0x00D	FP	PERF_CTL[3]
762  * 0x023	DE	PERF_CTL[2:0]
763  * 0x02D	LS	PERF_CTL[3]
764  * 0x02E	LS	PERF_CTL[3,0]
765  * 0x031	LS	PERF_CTL[2:0] (**)
766  * 0x043	CU	PERF_CTL[2:0]
767  * 0x045	CU	PERF_CTL[2:0]
768  * 0x046	CU	PERF_CTL[2:0]
769  * 0x054	CU	PERF_CTL[2:0]
770  * 0x055	CU	PERF_CTL[2:0]
771  * 0x08F	IC	PERF_CTL[0]
772  * 0x187	DE	PERF_CTL[0]
773  * 0x188	DE	PERF_CTL[0]
774  * 0x0DB	EX	PERF_CTL[5:0]
775  * 0x0DC	LS	PERF_CTL[5:0]
776  * 0x0DD	LS	PERF_CTL[5:0]
777  * 0x0DE	LS	PERF_CTL[5:0]
778  * 0x0DF	LS	PERF_CTL[5:0]
779  * 0x1C0	EX	PERF_CTL[5:3]
780  * 0x1D6	EX	PERF_CTL[5:0]
781  * 0x1D8	EX	PERF_CTL[5:0]
782  *
783  * (*)  depending on the umask all FPU counters may be used
784  * (**) only one unitmask enabled at a time
785  */
786 
787 static struct event_constraint amd_f15_PMC0  = EVENT_CONSTRAINT(0, 0x01, 0);
788 static struct event_constraint amd_f15_PMC20 = EVENT_CONSTRAINT(0, 0x07, 0);
789 static struct event_constraint amd_f15_PMC3  = EVENT_CONSTRAINT(0, 0x08, 0);
790 static struct event_constraint amd_f15_PMC30 = EVENT_CONSTRAINT_OVERLAP(0, 0x09, 0);
791 static struct event_constraint amd_f15_PMC50 = EVENT_CONSTRAINT(0, 0x3F, 0);
792 static struct event_constraint amd_f15_PMC53 = EVENT_CONSTRAINT(0, 0x38, 0);
793 
794 static struct event_constraint *
795 amd_get_event_constraints_f15h(struct cpu_hw_events *cpuc, int idx,
796 			       struct perf_event *event)
797 {
798 	struct hw_perf_event *hwc = &event->hw;
799 	unsigned int event_code = amd_get_event_code(hwc);
800 
801 	switch (event_code & AMD_EVENT_TYPE_MASK) {
802 	case AMD_EVENT_FP:
803 		switch (event_code) {
804 		case 0x000:
805 			if (!(hwc->config & 0x0000F000ULL))
806 				break;
807 			if (!(hwc->config & 0x00000F00ULL))
808 				break;
809 			return &amd_f15_PMC3;
810 		case 0x004:
811 			if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
812 				break;
813 			return &amd_f15_PMC3;
814 		case 0x003:
815 		case 0x00B:
816 		case 0x00D:
817 			return &amd_f15_PMC3;
818 		}
819 		return &amd_f15_PMC53;
820 	case AMD_EVENT_LS:
821 	case AMD_EVENT_DC:
822 	case AMD_EVENT_EX_LS:
823 		switch (event_code) {
824 		case 0x023:
825 		case 0x043:
826 		case 0x045:
827 		case 0x046:
828 		case 0x054:
829 		case 0x055:
830 			return &amd_f15_PMC20;
831 		case 0x02D:
832 			return &amd_f15_PMC3;
833 		case 0x02E:
834 			return &amd_f15_PMC30;
835 		case 0x031:
836 			if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
837 				return &amd_f15_PMC20;
838 			return &emptyconstraint;
839 		case 0x1C0:
840 			return &amd_f15_PMC53;
841 		default:
842 			return &amd_f15_PMC50;
843 		}
844 	case AMD_EVENT_CU:
845 	case AMD_EVENT_IC_DE:
846 	case AMD_EVENT_DE:
847 		switch (event_code) {
848 		case 0x08F:
849 		case 0x187:
850 		case 0x188:
851 			return &amd_f15_PMC0;
852 		case 0x0DB ... 0x0DF:
853 		case 0x1D6:
854 		case 0x1D8:
855 			return &amd_f15_PMC50;
856 		default:
857 			return &amd_f15_PMC20;
858 		}
859 	case AMD_EVENT_NB:
860 		/* moved to uncore.c */
861 		return &emptyconstraint;
862 	default:
863 		return &emptyconstraint;
864 	}
865 }
866 
867 static ssize_t amd_event_sysfs_show(char *page, u64 config)
868 {
869 	u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT) |
870 		    (config & AMD64_EVENTSEL_EVENT) >> 24;
871 
872 	return x86_event_sysfs_show(page, config, event);
873 }
874 
875 static __initconst const struct x86_pmu amd_pmu = {
876 	.name			= "AMD",
877 	.handle_irq		= amd_pmu_handle_irq,
878 	.disable_all		= amd_pmu_disable_all,
879 	.enable_all		= x86_pmu_enable_all,
880 	.enable			= x86_pmu_enable_event,
881 	.disable		= amd_pmu_disable_event,
882 	.hw_config		= amd_pmu_hw_config,
883 	.schedule_events	= x86_schedule_events,
884 	.eventsel		= MSR_K7_EVNTSEL0,
885 	.perfctr		= MSR_K7_PERFCTR0,
886 	.addr_offset            = amd_pmu_addr_offset,
887 	.event_map		= amd_pmu_event_map,
888 	.max_events		= ARRAY_SIZE(amd_perfmon_event_map),
889 	.num_counters		= AMD64_NUM_COUNTERS,
890 	.cntval_bits		= 48,
891 	.cntval_mask		= (1ULL << 48) - 1,
892 	.apic			= 1,
893 	/* use highest bit to detect overflow */
894 	.max_period		= (1ULL << 47) - 1,
895 	.get_event_constraints	= amd_get_event_constraints,
896 	.put_event_constraints	= amd_put_event_constraints,
897 
898 	.format_attrs		= amd_format_attr,
899 	.events_sysfs_show	= amd_event_sysfs_show,
900 
901 	.cpu_prepare		= amd_pmu_cpu_prepare,
902 	.cpu_starting		= amd_pmu_cpu_starting,
903 	.cpu_dead		= amd_pmu_cpu_dead,
904 
905 	.amd_nb_constraints	= 1,
906 };
907 
908 static int __init amd_core_pmu_init(void)
909 {
910 	if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
911 		return 0;
912 
913 	/* Avoid calulating the value each time in the NMI handler */
914 	perf_nmi_window = msecs_to_jiffies(100);
915 
916 	switch (boot_cpu_data.x86) {
917 	case 0x15:
918 		pr_cont("Fam15h ");
919 		x86_pmu.get_event_constraints = amd_get_event_constraints_f15h;
920 		break;
921 	case 0x17:
922 		pr_cont("Fam17h ");
923 		/*
924 		 * In family 17h, there are no event constraints in the PMC hardware.
925 		 * We fallback to using default amd_get_event_constraints.
926 		 */
927 		break;
928 	case 0x18:
929 		pr_cont("Fam18h ");
930 		/* Using default amd_get_event_constraints. */
931 		break;
932 	default:
933 		pr_err("core perfctr but no constraints; unknown hardware!\n");
934 		return -ENODEV;
935 	}
936 
937 	/*
938 	 * If core performance counter extensions exists, we must use
939 	 * MSR_F15H_PERF_CTL/MSR_F15H_PERF_CTR msrs. See also
940 	 * amd_pmu_addr_offset().
941 	 */
942 	x86_pmu.eventsel	= MSR_F15H_PERF_CTL;
943 	x86_pmu.perfctr		= MSR_F15H_PERF_CTR;
944 	x86_pmu.num_counters	= AMD64_NUM_COUNTERS_CORE;
945 	/*
946 	 * AMD Core perfctr has separate MSRs for the NB events, see
947 	 * the amd/uncore.c driver.
948 	 */
949 	x86_pmu.amd_nb_constraints = 0;
950 
951 	pr_cont("core perfctr, ");
952 	return 0;
953 }
954 
955 __init int amd_pmu_init(void)
956 {
957 	int ret;
958 
959 	/* Performance-monitoring supported from K7 and later: */
960 	if (boot_cpu_data.x86 < 6)
961 		return -ENODEV;
962 
963 	x86_pmu = amd_pmu;
964 
965 	ret = amd_core_pmu_init();
966 	if (ret)
967 		return ret;
968 
969 	if (num_possible_cpus() == 1) {
970 		/*
971 		 * No point in allocating data structures to serialize
972 		 * against other CPUs, when there is only the one CPU.
973 		 */
974 		x86_pmu.amd_nb_constraints = 0;
975 	}
976 
977 	if (boot_cpu_data.x86 >= 0x17)
978 		memcpy(hw_cache_event_ids, amd_hw_cache_event_ids_f17h, sizeof(hw_cache_event_ids));
979 	else
980 		memcpy(hw_cache_event_ids, amd_hw_cache_event_ids, sizeof(hw_cache_event_ids));
981 
982 	return 0;
983 }
984 
985 void amd_pmu_enable_virt(void)
986 {
987 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
988 
989 	cpuc->perf_ctr_virt_mask = 0;
990 
991 	/* Reload all events */
992 	amd_pmu_disable_all();
993 	x86_pmu_enable_all(0);
994 }
995 EXPORT_SYMBOL_GPL(amd_pmu_enable_virt);
996 
997 void amd_pmu_disable_virt(void)
998 {
999 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1000 
1001 	/*
1002 	 * We only mask out the Host-only bit so that host-only counting works
1003 	 * when SVM is disabled. If someone sets up a guest-only counter when
1004 	 * SVM is disabled the Guest-only bits still gets set and the counter
1005 	 * will not count anything.
1006 	 */
1007 	cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
1008 
1009 	/* Reload all events */
1010 	amd_pmu_disable_all();
1011 	x86_pmu_enable_all(0);
1012 }
1013 EXPORT_SYMBOL_GPL(amd_pmu_disable_virt);
1014