xref: /linux/arch/x86/events/amd/core.c (revision 7fbcd195e2b8cc952e4aeaeb50867b798040314c)
1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/perf_event.h>
3 #include <linux/jump_label.h>
4 #include <linux/export.h>
5 #include <linux/types.h>
6 #include <linux/init.h>
7 #include <linux/slab.h>
8 #include <linux/delay.h>
9 #include <linux/jiffies.h>
10 #include <asm/apicdef.h>
11 #include <asm/apic.h>
12 #include <asm/nmi.h>
13 
14 #include "../perf_event.h"
15 
16 static DEFINE_PER_CPU(unsigned long, perf_nmi_tstamp);
17 static unsigned long perf_nmi_window;
18 
19 /* AMD Event 0xFFF: Merge.  Used with Large Increment per Cycle events */
20 #define AMD_MERGE_EVENT ((0xFULL << 32) | 0xFFULL)
21 #define AMD_MERGE_EVENT_ENABLE (AMD_MERGE_EVENT | ARCH_PERFMON_EVENTSEL_ENABLE)
22 
23 /* PMC Enable and Overflow bits for PerfCntrGlobal* registers */
24 static u64 amd_pmu_global_cntr_mask __read_mostly;
25 
26 static __initconst const u64 amd_hw_cache_event_ids
27 				[PERF_COUNT_HW_CACHE_MAX]
28 				[PERF_COUNT_HW_CACHE_OP_MAX]
29 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
30 {
31  [ C(L1D) ] = {
32 	[ C(OP_READ) ] = {
33 		[ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses        */
34 		[ C(RESULT_MISS)   ] = 0x0141, /* Data Cache Misses          */
35 	},
36 	[ C(OP_WRITE) ] = {
37 		[ C(RESULT_ACCESS) ] = 0,
38 		[ C(RESULT_MISS)   ] = 0,
39 	},
40 	[ C(OP_PREFETCH) ] = {
41 		[ C(RESULT_ACCESS) ] = 0x0267, /* Data Prefetcher :attempts  */
42 		[ C(RESULT_MISS)   ] = 0x0167, /* Data Prefetcher :cancelled */
43 	},
44  },
45  [ C(L1I ) ] = {
46 	[ C(OP_READ) ] = {
47 		[ C(RESULT_ACCESS) ] = 0x0080, /* Instruction cache fetches  */
48 		[ C(RESULT_MISS)   ] = 0x0081, /* Instruction cache misses   */
49 	},
50 	[ C(OP_WRITE) ] = {
51 		[ C(RESULT_ACCESS) ] = -1,
52 		[ C(RESULT_MISS)   ] = -1,
53 	},
54 	[ C(OP_PREFETCH) ] = {
55 		[ C(RESULT_ACCESS) ] = 0x014B, /* Prefetch Instructions :Load */
56 		[ C(RESULT_MISS)   ] = 0,
57 	},
58  },
59  [ C(LL  ) ] = {
60 	[ C(OP_READ) ] = {
61 		[ C(RESULT_ACCESS) ] = 0x037D, /* Requests to L2 Cache :IC+DC */
62 		[ C(RESULT_MISS)   ] = 0x037E, /* L2 Cache Misses : IC+DC     */
63 	},
64 	[ C(OP_WRITE) ] = {
65 		[ C(RESULT_ACCESS) ] = 0x017F, /* L2 Fill/Writeback           */
66 		[ C(RESULT_MISS)   ] = 0,
67 	},
68 	[ C(OP_PREFETCH) ] = {
69 		[ C(RESULT_ACCESS) ] = 0,
70 		[ C(RESULT_MISS)   ] = 0,
71 	},
72  },
73  [ C(DTLB) ] = {
74 	[ C(OP_READ) ] = {
75 		[ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses        */
76 		[ C(RESULT_MISS)   ] = 0x0746, /* L1_DTLB_AND_L2_DLTB_MISS.ALL */
77 	},
78 	[ C(OP_WRITE) ] = {
79 		[ C(RESULT_ACCESS) ] = 0,
80 		[ C(RESULT_MISS)   ] = 0,
81 	},
82 	[ C(OP_PREFETCH) ] = {
83 		[ C(RESULT_ACCESS) ] = 0,
84 		[ C(RESULT_MISS)   ] = 0,
85 	},
86  },
87  [ C(ITLB) ] = {
88 	[ C(OP_READ) ] = {
89 		[ C(RESULT_ACCESS) ] = 0x0080, /* Instruction fecthes        */
90 		[ C(RESULT_MISS)   ] = 0x0385, /* L1_ITLB_AND_L2_ITLB_MISS.ALL */
91 	},
92 	[ C(OP_WRITE) ] = {
93 		[ C(RESULT_ACCESS) ] = -1,
94 		[ C(RESULT_MISS)   ] = -1,
95 	},
96 	[ C(OP_PREFETCH) ] = {
97 		[ C(RESULT_ACCESS) ] = -1,
98 		[ C(RESULT_MISS)   ] = -1,
99 	},
100  },
101  [ C(BPU ) ] = {
102 	[ C(OP_READ) ] = {
103 		[ C(RESULT_ACCESS) ] = 0x00c2, /* Retired Branch Instr.      */
104 		[ C(RESULT_MISS)   ] = 0x00c3, /* Retired Mispredicted BI    */
105 	},
106 	[ C(OP_WRITE) ] = {
107 		[ C(RESULT_ACCESS) ] = -1,
108 		[ C(RESULT_MISS)   ] = -1,
109 	},
110 	[ C(OP_PREFETCH) ] = {
111 		[ C(RESULT_ACCESS) ] = -1,
112 		[ C(RESULT_MISS)   ] = -1,
113 	},
114  },
115  [ C(NODE) ] = {
116 	[ C(OP_READ) ] = {
117 		[ C(RESULT_ACCESS) ] = 0xb8e9, /* CPU Request to Memory, l+r */
118 		[ C(RESULT_MISS)   ] = 0x98e9, /* CPU Request to Memory, r   */
119 	},
120 	[ C(OP_WRITE) ] = {
121 		[ C(RESULT_ACCESS) ] = -1,
122 		[ C(RESULT_MISS)   ] = -1,
123 	},
124 	[ C(OP_PREFETCH) ] = {
125 		[ C(RESULT_ACCESS) ] = -1,
126 		[ C(RESULT_MISS)   ] = -1,
127 	},
128  },
129 };
130 
131 static __initconst const u64 amd_hw_cache_event_ids_f17h
132 				[PERF_COUNT_HW_CACHE_MAX]
133 				[PERF_COUNT_HW_CACHE_OP_MAX]
134 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
135 [C(L1D)] = {
136 	[C(OP_READ)] = {
137 		[C(RESULT_ACCESS)] = 0x0040, /* Data Cache Accesses */
138 		[C(RESULT_MISS)]   = 0xc860, /* L2$ access from DC Miss */
139 	},
140 	[C(OP_WRITE)] = {
141 		[C(RESULT_ACCESS)] = 0,
142 		[C(RESULT_MISS)]   = 0,
143 	},
144 	[C(OP_PREFETCH)] = {
145 		[C(RESULT_ACCESS)] = 0xff5a, /* h/w prefetch DC Fills */
146 		[C(RESULT_MISS)]   = 0,
147 	},
148 },
149 [C(L1I)] = {
150 	[C(OP_READ)] = {
151 		[C(RESULT_ACCESS)] = 0x0080, /* Instruction cache fetches  */
152 		[C(RESULT_MISS)]   = 0x0081, /* Instruction cache misses   */
153 	},
154 	[C(OP_WRITE)] = {
155 		[C(RESULT_ACCESS)] = -1,
156 		[C(RESULT_MISS)]   = -1,
157 	},
158 	[C(OP_PREFETCH)] = {
159 		[C(RESULT_ACCESS)] = 0,
160 		[C(RESULT_MISS)]   = 0,
161 	},
162 },
163 [C(LL)] = {
164 	[C(OP_READ)] = {
165 		[C(RESULT_ACCESS)] = 0,
166 		[C(RESULT_MISS)]   = 0,
167 	},
168 	[C(OP_WRITE)] = {
169 		[C(RESULT_ACCESS)] = 0,
170 		[C(RESULT_MISS)]   = 0,
171 	},
172 	[C(OP_PREFETCH)] = {
173 		[C(RESULT_ACCESS)] = 0,
174 		[C(RESULT_MISS)]   = 0,
175 	},
176 },
177 [C(DTLB)] = {
178 	[C(OP_READ)] = {
179 		[C(RESULT_ACCESS)] = 0xff45, /* All L2 DTLB accesses */
180 		[C(RESULT_MISS)]   = 0xf045, /* L2 DTLB misses (PT walks) */
181 	},
182 	[C(OP_WRITE)] = {
183 		[C(RESULT_ACCESS)] = 0,
184 		[C(RESULT_MISS)]   = 0,
185 	},
186 	[C(OP_PREFETCH)] = {
187 		[C(RESULT_ACCESS)] = 0,
188 		[C(RESULT_MISS)]   = 0,
189 	},
190 },
191 [C(ITLB)] = {
192 	[C(OP_READ)] = {
193 		[C(RESULT_ACCESS)] = 0x0084, /* L1 ITLB misses, L2 ITLB hits */
194 		[C(RESULT_MISS)]   = 0xff85, /* L1 ITLB misses, L2 misses */
195 	},
196 	[C(OP_WRITE)] = {
197 		[C(RESULT_ACCESS)] = -1,
198 		[C(RESULT_MISS)]   = -1,
199 	},
200 	[C(OP_PREFETCH)] = {
201 		[C(RESULT_ACCESS)] = -1,
202 		[C(RESULT_MISS)]   = -1,
203 	},
204 },
205 [C(BPU)] = {
206 	[C(OP_READ)] = {
207 		[C(RESULT_ACCESS)] = 0x00c2, /* Retired Branch Instr.      */
208 		[C(RESULT_MISS)]   = 0x00c3, /* Retired Mispredicted BI    */
209 	},
210 	[C(OP_WRITE)] = {
211 		[C(RESULT_ACCESS)] = -1,
212 		[C(RESULT_MISS)]   = -1,
213 	},
214 	[C(OP_PREFETCH)] = {
215 		[C(RESULT_ACCESS)] = -1,
216 		[C(RESULT_MISS)]   = -1,
217 	},
218 },
219 [C(NODE)] = {
220 	[C(OP_READ)] = {
221 		[C(RESULT_ACCESS)] = 0,
222 		[C(RESULT_MISS)]   = 0,
223 	},
224 	[C(OP_WRITE)] = {
225 		[C(RESULT_ACCESS)] = -1,
226 		[C(RESULT_MISS)]   = -1,
227 	},
228 	[C(OP_PREFETCH)] = {
229 		[C(RESULT_ACCESS)] = -1,
230 		[C(RESULT_MISS)]   = -1,
231 	},
232 },
233 };
234 
235 /*
236  * AMD Performance Monitor K7 and later, up to and including Family 16h:
237  */
238 static const u64 amd_perfmon_event_map[PERF_COUNT_HW_MAX] =
239 {
240 	[PERF_COUNT_HW_CPU_CYCLES]		= 0x0076,
241 	[PERF_COUNT_HW_INSTRUCTIONS]		= 0x00c0,
242 	[PERF_COUNT_HW_CACHE_REFERENCES]	= 0x077d,
243 	[PERF_COUNT_HW_CACHE_MISSES]		= 0x077e,
244 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x00c2,
245 	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x00c3,
246 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= 0x00d0, /* "Decoder empty" event */
247 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= 0x00d1, /* "Dispatch stalls" event */
248 };
249 
250 /*
251  * AMD Performance Monitor Family 17h and later:
252  */
253 static const u64 amd_f17h_perfmon_event_map[PERF_COUNT_HW_MAX] =
254 {
255 	[PERF_COUNT_HW_CPU_CYCLES]		= 0x0076,
256 	[PERF_COUNT_HW_INSTRUCTIONS]		= 0x00c0,
257 	[PERF_COUNT_HW_CACHE_REFERENCES]	= 0xff60,
258 	[PERF_COUNT_HW_CACHE_MISSES]		= 0x0964,
259 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x00c2,
260 	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x00c3,
261 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]	= 0x0287,
262 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND]	= 0x0187,
263 };
264 
265 static u64 amd_pmu_event_map(int hw_event)
266 {
267 	if (boot_cpu_data.x86 >= 0x17)
268 		return amd_f17h_perfmon_event_map[hw_event];
269 
270 	return amd_perfmon_event_map[hw_event];
271 }
272 
273 /*
274  * Previously calculated offsets
275  */
276 static unsigned int event_offsets[X86_PMC_IDX_MAX] __read_mostly;
277 static unsigned int count_offsets[X86_PMC_IDX_MAX] __read_mostly;
278 
279 /*
280  * Legacy CPUs:
281  *   4 counters starting at 0xc0010000 each offset by 1
282  *
283  * CPUs with core performance counter extensions:
284  *   6 counters starting at 0xc0010200 each offset by 2
285  */
286 static inline int amd_pmu_addr_offset(int index, bool eventsel)
287 {
288 	int offset;
289 
290 	if (!index)
291 		return index;
292 
293 	if (eventsel)
294 		offset = event_offsets[index];
295 	else
296 		offset = count_offsets[index];
297 
298 	if (offset)
299 		return offset;
300 
301 	if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
302 		offset = index;
303 	else
304 		offset = index << 1;
305 
306 	if (eventsel)
307 		event_offsets[index] = offset;
308 	else
309 		count_offsets[index] = offset;
310 
311 	return offset;
312 }
313 
314 /*
315  * AMD64 events are detected based on their event codes.
316  */
317 static inline unsigned int amd_get_event_code(struct hw_perf_event *hwc)
318 {
319 	return ((hwc->config >> 24) & 0x0f00) | (hwc->config & 0x00ff);
320 }
321 
322 static inline bool amd_is_pair_event_code(struct hw_perf_event *hwc)
323 {
324 	if (!(x86_pmu.flags & PMU_FL_PAIR))
325 		return false;
326 
327 	switch (amd_get_event_code(hwc)) {
328 	case 0x003:	return true;	/* Retired SSE/AVX FLOPs */
329 	default:	return false;
330 	}
331 }
332 
333 DEFINE_STATIC_CALL_RET0(amd_pmu_branch_hw_config, *x86_pmu.hw_config);
334 
335 static int amd_core_hw_config(struct perf_event *event)
336 {
337 	if (event->attr.exclude_host && event->attr.exclude_guest)
338 		/*
339 		 * When HO == GO == 1 the hardware treats that as GO == HO == 0
340 		 * and will count in both modes. We don't want to count in that
341 		 * case so we emulate no-counting by setting US = OS = 0.
342 		 */
343 		event->hw.config &= ~(ARCH_PERFMON_EVENTSEL_USR |
344 				      ARCH_PERFMON_EVENTSEL_OS);
345 	else if (event->attr.exclude_host)
346 		event->hw.config |= AMD64_EVENTSEL_GUESTONLY;
347 	else if (event->attr.exclude_guest)
348 		event->hw.config |= AMD64_EVENTSEL_HOSTONLY;
349 
350 	if ((x86_pmu.flags & PMU_FL_PAIR) && amd_is_pair_event_code(&event->hw))
351 		event->hw.flags |= PERF_X86_EVENT_PAIR;
352 
353 	if (has_branch_stack(event))
354 		return static_call(amd_pmu_branch_hw_config)(event);
355 
356 	return 0;
357 }
358 
359 static inline int amd_is_nb_event(struct hw_perf_event *hwc)
360 {
361 	return (hwc->config & 0xe0) == 0xe0;
362 }
363 
364 static inline int amd_has_nb(struct cpu_hw_events *cpuc)
365 {
366 	struct amd_nb *nb = cpuc->amd_nb;
367 
368 	return nb && nb->nb_id != -1;
369 }
370 
371 static int amd_pmu_hw_config(struct perf_event *event)
372 {
373 	int ret;
374 
375 	/* pass precise event sampling to ibs: */
376 	if (event->attr.precise_ip && get_ibs_caps())
377 		return forward_event_to_ibs(event);
378 
379 	if (has_branch_stack(event) && !x86_pmu.lbr_nr)
380 		return -EOPNOTSUPP;
381 
382 	ret = x86_pmu_hw_config(event);
383 	if (ret)
384 		return ret;
385 
386 	if (event->attr.type == PERF_TYPE_RAW)
387 		event->hw.config |= event->attr.config & AMD64_RAW_EVENT_MASK;
388 
389 	return amd_core_hw_config(event);
390 }
391 
392 static void __amd_put_nb_event_constraints(struct cpu_hw_events *cpuc,
393 					   struct perf_event *event)
394 {
395 	struct amd_nb *nb = cpuc->amd_nb;
396 	int i;
397 
398 	/*
399 	 * need to scan whole list because event may not have
400 	 * been assigned during scheduling
401 	 *
402 	 * no race condition possible because event can only
403 	 * be removed on one CPU at a time AND PMU is disabled
404 	 * when we come here
405 	 */
406 	for (i = 0; i < x86_pmu.num_counters; i++) {
407 		if (cmpxchg(nb->owners + i, event, NULL) == event)
408 			break;
409 	}
410 }
411 
412  /*
413   * AMD64 NorthBridge events need special treatment because
414   * counter access needs to be synchronized across all cores
415   * of a package. Refer to BKDG section 3.12
416   *
417   * NB events are events measuring L3 cache, Hypertransport
418   * traffic. They are identified by an event code >= 0xe00.
419   * They measure events on the NorthBride which is shared
420   * by all cores on a package. NB events are counted on a
421   * shared set of counters. When a NB event is programmed
422   * in a counter, the data actually comes from a shared
423   * counter. Thus, access to those counters needs to be
424   * synchronized.
425   *
426   * We implement the synchronization such that no two cores
427   * can be measuring NB events using the same counters. Thus,
428   * we maintain a per-NB allocation table. The available slot
429   * is propagated using the event_constraint structure.
430   *
431   * We provide only one choice for each NB event based on
432   * the fact that only NB events have restrictions. Consequently,
433   * if a counter is available, there is a guarantee the NB event
434   * will be assigned to it. If no slot is available, an empty
435   * constraint is returned and scheduling will eventually fail
436   * for this event.
437   *
438   * Note that all cores attached the same NB compete for the same
439   * counters to host NB events, this is why we use atomic ops. Some
440   * multi-chip CPUs may have more than one NB.
441   *
442   * Given that resources are allocated (cmpxchg), they must be
443   * eventually freed for others to use. This is accomplished by
444   * calling __amd_put_nb_event_constraints()
445   *
446   * Non NB events are not impacted by this restriction.
447   */
448 static struct event_constraint *
449 __amd_get_nb_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
450 			       struct event_constraint *c)
451 {
452 	struct hw_perf_event *hwc = &event->hw;
453 	struct amd_nb *nb = cpuc->amd_nb;
454 	struct perf_event *old;
455 	int idx, new = -1;
456 
457 	if (!c)
458 		c = &unconstrained;
459 
460 	if (cpuc->is_fake)
461 		return c;
462 
463 	/*
464 	 * detect if already present, if so reuse
465 	 *
466 	 * cannot merge with actual allocation
467 	 * because of possible holes
468 	 *
469 	 * event can already be present yet not assigned (in hwc->idx)
470 	 * because of successive calls to x86_schedule_events() from
471 	 * hw_perf_group_sched_in() without hw_perf_enable()
472 	 */
473 	for_each_set_bit(idx, c->idxmsk, x86_pmu.num_counters) {
474 		if (new == -1 || hwc->idx == idx)
475 			/* assign free slot, prefer hwc->idx */
476 			old = cmpxchg(nb->owners + idx, NULL, event);
477 		else if (nb->owners[idx] == event)
478 			/* event already present */
479 			old = event;
480 		else
481 			continue;
482 
483 		if (old && old != event)
484 			continue;
485 
486 		/* reassign to this slot */
487 		if (new != -1)
488 			cmpxchg(nb->owners + new, event, NULL);
489 		new = idx;
490 
491 		/* already present, reuse */
492 		if (old == event)
493 			break;
494 	}
495 
496 	if (new == -1)
497 		return &emptyconstraint;
498 
499 	return &nb->event_constraints[new];
500 }
501 
502 static struct amd_nb *amd_alloc_nb(int cpu)
503 {
504 	struct amd_nb *nb;
505 	int i;
506 
507 	nb = kzalloc_node(sizeof(struct amd_nb), GFP_KERNEL, cpu_to_node(cpu));
508 	if (!nb)
509 		return NULL;
510 
511 	nb->nb_id = -1;
512 
513 	/*
514 	 * initialize all possible NB constraints
515 	 */
516 	for (i = 0; i < x86_pmu.num_counters; i++) {
517 		__set_bit(i, nb->event_constraints[i].idxmsk);
518 		nb->event_constraints[i].weight = 1;
519 	}
520 	return nb;
521 }
522 
523 typedef void (amd_pmu_branch_reset_t)(void);
524 DEFINE_STATIC_CALL_NULL(amd_pmu_branch_reset, amd_pmu_branch_reset_t);
525 
526 static void amd_pmu_cpu_reset(int cpu)
527 {
528 	if (x86_pmu.lbr_nr)
529 		static_call(amd_pmu_branch_reset)();
530 
531 	if (x86_pmu.version < 2)
532 		return;
533 
534 	/* Clear enable bits i.e. PerfCntrGlobalCtl.PerfCntrEn */
535 	wrmsrl(MSR_AMD64_PERF_CNTR_GLOBAL_CTL, 0);
536 
537 	/*
538 	 * Clear freeze and overflow bits i.e. PerfCntrGLobalStatus.LbrFreeze
539 	 * and PerfCntrGLobalStatus.PerfCntrOvfl
540 	 */
541 	wrmsrl(MSR_AMD64_PERF_CNTR_GLOBAL_STATUS_CLR,
542 	       GLOBAL_STATUS_LBRS_FROZEN | amd_pmu_global_cntr_mask);
543 }
544 
545 static int amd_pmu_cpu_prepare(int cpu)
546 {
547 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
548 
549 	cpuc->lbr_sel = kzalloc_node(sizeof(struct er_account), GFP_KERNEL,
550 				     cpu_to_node(cpu));
551 	if (!cpuc->lbr_sel)
552 		return -ENOMEM;
553 
554 	WARN_ON_ONCE(cpuc->amd_nb);
555 
556 	if (!x86_pmu.amd_nb_constraints)
557 		return 0;
558 
559 	cpuc->amd_nb = amd_alloc_nb(cpu);
560 	if (cpuc->amd_nb)
561 		return 0;
562 
563 	kfree(cpuc->lbr_sel);
564 	cpuc->lbr_sel = NULL;
565 
566 	return -ENOMEM;
567 }
568 
569 static void amd_pmu_cpu_starting(int cpu)
570 {
571 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
572 	void **onln = &cpuc->kfree_on_online[X86_PERF_KFREE_SHARED];
573 	struct amd_nb *nb;
574 	int i, nb_id;
575 
576 	cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
577 	amd_pmu_cpu_reset(cpu);
578 
579 	if (!x86_pmu.amd_nb_constraints)
580 		return;
581 
582 	nb_id = topology_die_id(cpu);
583 	WARN_ON_ONCE(nb_id == BAD_APICID);
584 
585 	for_each_online_cpu(i) {
586 		nb = per_cpu(cpu_hw_events, i).amd_nb;
587 		if (WARN_ON_ONCE(!nb))
588 			continue;
589 
590 		if (nb->nb_id == nb_id) {
591 			*onln = cpuc->amd_nb;
592 			cpuc->amd_nb = nb;
593 			break;
594 		}
595 	}
596 
597 	cpuc->amd_nb->nb_id = nb_id;
598 	cpuc->amd_nb->refcnt++;
599 }
600 
601 static void amd_pmu_cpu_dead(int cpu)
602 {
603 	struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
604 
605 	kfree(cpuhw->lbr_sel);
606 	cpuhw->lbr_sel = NULL;
607 	amd_pmu_cpu_reset(cpu);
608 
609 	if (!x86_pmu.amd_nb_constraints)
610 		return;
611 
612 	if (cpuhw->amd_nb) {
613 		struct amd_nb *nb = cpuhw->amd_nb;
614 
615 		if (nb->nb_id == -1 || --nb->refcnt == 0)
616 			kfree(nb);
617 
618 		cpuhw->amd_nb = NULL;
619 	}
620 }
621 
622 static inline void amd_pmu_set_global_ctl(u64 ctl)
623 {
624 	wrmsrl(MSR_AMD64_PERF_CNTR_GLOBAL_CTL, ctl);
625 }
626 
627 static inline u64 amd_pmu_get_global_status(void)
628 {
629 	u64 status;
630 
631 	/* PerfCntrGlobalStatus is read-only */
632 	rdmsrl(MSR_AMD64_PERF_CNTR_GLOBAL_STATUS, status);
633 
634 	return status;
635 }
636 
637 static inline void amd_pmu_ack_global_status(u64 status)
638 {
639 	/*
640 	 * PerfCntrGlobalStatus is read-only but an overflow acknowledgment
641 	 * mechanism exists; writing 1 to a bit in PerfCntrGlobalStatusClr
642 	 * clears the same bit in PerfCntrGlobalStatus
643 	 */
644 
645 	wrmsrl(MSR_AMD64_PERF_CNTR_GLOBAL_STATUS_CLR, status);
646 }
647 
648 static bool amd_pmu_test_overflow_topbit(int idx)
649 {
650 	u64 counter;
651 
652 	rdmsrl(x86_pmu_event_addr(idx), counter);
653 
654 	return !(counter & BIT_ULL(x86_pmu.cntval_bits - 1));
655 }
656 
657 static bool amd_pmu_test_overflow_status(int idx)
658 {
659 	return amd_pmu_get_global_status() & BIT_ULL(idx);
660 }
661 
662 DEFINE_STATIC_CALL(amd_pmu_test_overflow, amd_pmu_test_overflow_topbit);
663 
664 /*
665  * When a PMC counter overflows, an NMI is used to process the event and
666  * reset the counter. NMI latency can result in the counter being updated
667  * before the NMI can run, which can result in what appear to be spurious
668  * NMIs. This function is intended to wait for the NMI to run and reset
669  * the counter to avoid possible unhandled NMI messages.
670  */
671 #define OVERFLOW_WAIT_COUNT	50
672 
673 static void amd_pmu_wait_on_overflow(int idx)
674 {
675 	unsigned int i;
676 
677 	/*
678 	 * Wait for the counter to be reset if it has overflowed. This loop
679 	 * should exit very, very quickly, but just in case, don't wait
680 	 * forever...
681 	 */
682 	for (i = 0; i < OVERFLOW_WAIT_COUNT; i++) {
683 		if (!static_call(amd_pmu_test_overflow)(idx))
684 			break;
685 
686 		/* Might be in IRQ context, so can't sleep */
687 		udelay(1);
688 	}
689 }
690 
691 static void amd_pmu_check_overflow(void)
692 {
693 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
694 	int idx;
695 
696 	/*
697 	 * This shouldn't be called from NMI context, but add a safeguard here
698 	 * to return, since if we're in NMI context we can't wait for an NMI
699 	 * to reset an overflowed counter value.
700 	 */
701 	if (in_nmi())
702 		return;
703 
704 	/*
705 	 * Check each counter for overflow and wait for it to be reset by the
706 	 * NMI if it has overflowed. This relies on the fact that all active
707 	 * counters are always enabled when this function is called and
708 	 * ARCH_PERFMON_EVENTSEL_INT is always set.
709 	 */
710 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
711 		if (!test_bit(idx, cpuc->active_mask))
712 			continue;
713 
714 		amd_pmu_wait_on_overflow(idx);
715 	}
716 }
717 
718 static void amd_pmu_enable_event(struct perf_event *event)
719 {
720 	x86_pmu_enable_event(event);
721 }
722 
723 static void amd_pmu_enable_all(int added)
724 {
725 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
726 	int idx;
727 
728 	amd_brs_enable_all();
729 
730 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
731 		/* only activate events which are marked as active */
732 		if (!test_bit(idx, cpuc->active_mask))
733 			continue;
734 
735 		amd_pmu_enable_event(cpuc->events[idx]);
736 	}
737 }
738 
739 static void amd_pmu_v2_enable_event(struct perf_event *event)
740 {
741 	struct hw_perf_event *hwc = &event->hw;
742 
743 	/*
744 	 * Testing cpu_hw_events.enabled should be skipped in this case unlike
745 	 * in x86_pmu_enable_event().
746 	 *
747 	 * Since cpu_hw_events.enabled is set only after returning from
748 	 * x86_pmu_start(), the PMCs must be programmed and kept ready.
749 	 * Counting starts only after x86_pmu_enable_all() is called.
750 	 */
751 	__x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
752 }
753 
754 static __always_inline void amd_pmu_core_enable_all(void)
755 {
756 	amd_pmu_set_global_ctl(amd_pmu_global_cntr_mask);
757 }
758 
759 static void amd_pmu_v2_enable_all(int added)
760 {
761 	amd_pmu_lbr_enable_all();
762 	amd_pmu_core_enable_all();
763 }
764 
765 static void amd_pmu_disable_event(struct perf_event *event)
766 {
767 	x86_pmu_disable_event(event);
768 
769 	/*
770 	 * This can be called from NMI context (via x86_pmu_stop). The counter
771 	 * may have overflowed, but either way, we'll never see it get reset
772 	 * by the NMI if we're already in the NMI. And the NMI latency support
773 	 * below will take care of any pending NMI that might have been
774 	 * generated by the overflow.
775 	 */
776 	if (in_nmi())
777 		return;
778 
779 	amd_pmu_wait_on_overflow(event->hw.idx);
780 }
781 
782 static void amd_pmu_disable_all(void)
783 {
784 	amd_brs_disable_all();
785 	x86_pmu_disable_all();
786 	amd_pmu_check_overflow();
787 }
788 
789 static __always_inline void amd_pmu_core_disable_all(void)
790 {
791 	amd_pmu_set_global_ctl(0);
792 }
793 
794 static void amd_pmu_v2_disable_all(void)
795 {
796 	amd_pmu_core_disable_all();
797 	amd_pmu_lbr_disable_all();
798 	amd_pmu_check_overflow();
799 }
800 
801 DEFINE_STATIC_CALL_NULL(amd_pmu_branch_add, *x86_pmu.add);
802 
803 static void amd_pmu_add_event(struct perf_event *event)
804 {
805 	if (needs_branch_stack(event))
806 		static_call(amd_pmu_branch_add)(event);
807 }
808 
809 DEFINE_STATIC_CALL_NULL(amd_pmu_branch_del, *x86_pmu.del);
810 
811 static void amd_pmu_del_event(struct perf_event *event)
812 {
813 	if (needs_branch_stack(event))
814 		static_call(amd_pmu_branch_del)(event);
815 }
816 
817 /*
818  * Because of NMI latency, if multiple PMC counters are active or other sources
819  * of NMIs are received, the perf NMI handler can handle one or more overflowed
820  * PMC counters outside of the NMI associated with the PMC overflow. If the NMI
821  * doesn't arrive at the LAPIC in time to become a pending NMI, then the kernel
822  * back-to-back NMI support won't be active. This PMC handler needs to take into
823  * account that this can occur, otherwise this could result in unknown NMI
824  * messages being issued. Examples of this is PMC overflow while in the NMI
825  * handler when multiple PMCs are active or PMC overflow while handling some
826  * other source of an NMI.
827  *
828  * Attempt to mitigate this by creating an NMI window in which un-handled NMIs
829  * received during this window will be claimed. This prevents extending the
830  * window past when it is possible that latent NMIs should be received. The
831  * per-CPU perf_nmi_tstamp will be set to the window end time whenever perf has
832  * handled a counter. When an un-handled NMI is received, it will be claimed
833  * only if arriving within that window.
834  */
835 static inline int amd_pmu_adjust_nmi_window(int handled)
836 {
837 	/*
838 	 * If a counter was handled, record a timestamp such that un-handled
839 	 * NMIs will be claimed if arriving within that window.
840 	 */
841 	if (handled) {
842 		this_cpu_write(perf_nmi_tstamp, jiffies + perf_nmi_window);
843 
844 		return handled;
845 	}
846 
847 	if (time_after(jiffies, this_cpu_read(perf_nmi_tstamp)))
848 		return NMI_DONE;
849 
850 	return NMI_HANDLED;
851 }
852 
853 static int amd_pmu_handle_irq(struct pt_regs *regs)
854 {
855 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
856 	int handled;
857 	int pmu_enabled;
858 
859 	/*
860 	 * Save the PMU state.
861 	 * It needs to be restored when leaving the handler.
862 	 */
863 	pmu_enabled = cpuc->enabled;
864 	cpuc->enabled = 0;
865 
866 	amd_brs_disable_all();
867 
868 	/* Drain BRS is in use (could be inactive) */
869 	if (cpuc->lbr_users)
870 		amd_brs_drain();
871 
872 	/* Process any counter overflows */
873 	handled = x86_pmu_handle_irq(regs);
874 
875 	cpuc->enabled = pmu_enabled;
876 	if (pmu_enabled)
877 		amd_brs_enable_all();
878 
879 	return amd_pmu_adjust_nmi_window(handled);
880 }
881 
882 static int amd_pmu_v2_handle_irq(struct pt_regs *regs)
883 {
884 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
885 	struct perf_sample_data data;
886 	struct hw_perf_event *hwc;
887 	struct perf_event *event;
888 	int handled = 0, idx;
889 	u64 reserved, status, mask;
890 	bool pmu_enabled;
891 
892 	/*
893 	 * Save the PMU state as it needs to be restored when leaving the
894 	 * handler
895 	 */
896 	pmu_enabled = cpuc->enabled;
897 	cpuc->enabled = 0;
898 
899 	/* Stop counting but do not disable LBR */
900 	amd_pmu_core_disable_all();
901 
902 	status = amd_pmu_get_global_status();
903 
904 	/* Check if any overflows are pending */
905 	if (!status)
906 		goto done;
907 
908 	/* Read branch records before unfreezing */
909 	if (status & GLOBAL_STATUS_LBRS_FROZEN) {
910 		amd_pmu_lbr_read();
911 		status &= ~GLOBAL_STATUS_LBRS_FROZEN;
912 	}
913 
914 	reserved = status & ~amd_pmu_global_cntr_mask;
915 	if (reserved)
916 		pr_warn_once("Reserved PerfCntrGlobalStatus bits are set (0x%llx), please consider updating microcode\n",
917 			     reserved);
918 
919 	/* Clear any reserved bits set by buggy microcode */
920 	status &= amd_pmu_global_cntr_mask;
921 
922 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
923 		if (!test_bit(idx, cpuc->active_mask))
924 			continue;
925 
926 		event = cpuc->events[idx];
927 		hwc = &event->hw;
928 		x86_perf_event_update(event);
929 		mask = BIT_ULL(idx);
930 
931 		if (!(status & mask))
932 			continue;
933 
934 		/* Event overflow */
935 		handled++;
936 		status &= ~mask;
937 		perf_sample_data_init(&data, 0, hwc->last_period);
938 
939 		if (!x86_perf_event_set_period(event))
940 			continue;
941 
942 		if (has_branch_stack(event))
943 			perf_sample_save_brstack(&data, event, &cpuc->lbr_stack);
944 
945 		if (perf_event_overflow(event, &data, regs))
946 			x86_pmu_stop(event, 0);
947 	}
948 
949 	/*
950 	 * It should never be the case that some overflows are not handled as
951 	 * the corresponding PMCs are expected to be inactive according to the
952 	 * active_mask
953 	 */
954 	WARN_ON(status > 0);
955 
956 	/* Clear overflow and freeze bits */
957 	amd_pmu_ack_global_status(~status);
958 
959 	/*
960 	 * Unmasking the LVTPC is not required as the Mask (M) bit of the LVT
961 	 * PMI entry is not set by the local APIC when a PMC overflow occurs
962 	 */
963 	inc_irq_stat(apic_perf_irqs);
964 
965 done:
966 	cpuc->enabled = pmu_enabled;
967 
968 	/* Resume counting only if PMU is active */
969 	if (pmu_enabled)
970 		amd_pmu_core_enable_all();
971 
972 	return amd_pmu_adjust_nmi_window(handled);
973 }
974 
975 static struct event_constraint *
976 amd_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
977 			  struct perf_event *event)
978 {
979 	/*
980 	 * if not NB event or no NB, then no constraints
981 	 */
982 	if (!(amd_has_nb(cpuc) && amd_is_nb_event(&event->hw)))
983 		return &unconstrained;
984 
985 	return __amd_get_nb_event_constraints(cpuc, event, NULL);
986 }
987 
988 static void amd_put_event_constraints(struct cpu_hw_events *cpuc,
989 				      struct perf_event *event)
990 {
991 	if (amd_has_nb(cpuc) && amd_is_nb_event(&event->hw))
992 		__amd_put_nb_event_constraints(cpuc, event);
993 }
994 
995 PMU_FORMAT_ATTR(event,	"config:0-7,32-35");
996 PMU_FORMAT_ATTR(umask,	"config:8-15"	);
997 PMU_FORMAT_ATTR(edge,	"config:18"	);
998 PMU_FORMAT_ATTR(inv,	"config:23"	);
999 PMU_FORMAT_ATTR(cmask,	"config:24-31"	);
1000 
1001 static struct attribute *amd_format_attr[] = {
1002 	&format_attr_event.attr,
1003 	&format_attr_umask.attr,
1004 	&format_attr_edge.attr,
1005 	&format_attr_inv.attr,
1006 	&format_attr_cmask.attr,
1007 	NULL,
1008 };
1009 
1010 /* AMD Family 15h */
1011 
1012 #define AMD_EVENT_TYPE_MASK	0x000000F0ULL
1013 
1014 #define AMD_EVENT_FP		0x00000000ULL ... 0x00000010ULL
1015 #define AMD_EVENT_LS		0x00000020ULL ... 0x00000030ULL
1016 #define AMD_EVENT_DC		0x00000040ULL ... 0x00000050ULL
1017 #define AMD_EVENT_CU		0x00000060ULL ... 0x00000070ULL
1018 #define AMD_EVENT_IC_DE		0x00000080ULL ... 0x00000090ULL
1019 #define AMD_EVENT_EX_LS		0x000000C0ULL
1020 #define AMD_EVENT_DE		0x000000D0ULL
1021 #define AMD_EVENT_NB		0x000000E0ULL ... 0x000000F0ULL
1022 
1023 /*
1024  * AMD family 15h event code/PMC mappings:
1025  *
1026  * type = event_code & 0x0F0:
1027  *
1028  * 0x000	FP	PERF_CTL[5:3]
1029  * 0x010	FP	PERF_CTL[5:3]
1030  * 0x020	LS	PERF_CTL[5:0]
1031  * 0x030	LS	PERF_CTL[5:0]
1032  * 0x040	DC	PERF_CTL[5:0]
1033  * 0x050	DC	PERF_CTL[5:0]
1034  * 0x060	CU	PERF_CTL[2:0]
1035  * 0x070	CU	PERF_CTL[2:0]
1036  * 0x080	IC/DE	PERF_CTL[2:0]
1037  * 0x090	IC/DE	PERF_CTL[2:0]
1038  * 0x0A0	---
1039  * 0x0B0	---
1040  * 0x0C0	EX/LS	PERF_CTL[5:0]
1041  * 0x0D0	DE	PERF_CTL[2:0]
1042  * 0x0E0	NB	NB_PERF_CTL[3:0]
1043  * 0x0F0	NB	NB_PERF_CTL[3:0]
1044  *
1045  * Exceptions:
1046  *
1047  * 0x000	FP	PERF_CTL[3], PERF_CTL[5:3] (*)
1048  * 0x003	FP	PERF_CTL[3]
1049  * 0x004	FP	PERF_CTL[3], PERF_CTL[5:3] (*)
1050  * 0x00B	FP	PERF_CTL[3]
1051  * 0x00D	FP	PERF_CTL[3]
1052  * 0x023	DE	PERF_CTL[2:0]
1053  * 0x02D	LS	PERF_CTL[3]
1054  * 0x02E	LS	PERF_CTL[3,0]
1055  * 0x031	LS	PERF_CTL[2:0] (**)
1056  * 0x043	CU	PERF_CTL[2:0]
1057  * 0x045	CU	PERF_CTL[2:0]
1058  * 0x046	CU	PERF_CTL[2:0]
1059  * 0x054	CU	PERF_CTL[2:0]
1060  * 0x055	CU	PERF_CTL[2:0]
1061  * 0x08F	IC	PERF_CTL[0]
1062  * 0x187	DE	PERF_CTL[0]
1063  * 0x188	DE	PERF_CTL[0]
1064  * 0x0DB	EX	PERF_CTL[5:0]
1065  * 0x0DC	LS	PERF_CTL[5:0]
1066  * 0x0DD	LS	PERF_CTL[5:0]
1067  * 0x0DE	LS	PERF_CTL[5:0]
1068  * 0x0DF	LS	PERF_CTL[5:0]
1069  * 0x1C0	EX	PERF_CTL[5:3]
1070  * 0x1D6	EX	PERF_CTL[5:0]
1071  * 0x1D8	EX	PERF_CTL[5:0]
1072  *
1073  * (*)  depending on the umask all FPU counters may be used
1074  * (**) only one unitmask enabled at a time
1075  */
1076 
1077 static struct event_constraint amd_f15_PMC0  = EVENT_CONSTRAINT(0, 0x01, 0);
1078 static struct event_constraint amd_f15_PMC20 = EVENT_CONSTRAINT(0, 0x07, 0);
1079 static struct event_constraint amd_f15_PMC3  = EVENT_CONSTRAINT(0, 0x08, 0);
1080 static struct event_constraint amd_f15_PMC30 = EVENT_CONSTRAINT_OVERLAP(0, 0x09, 0);
1081 static struct event_constraint amd_f15_PMC50 = EVENT_CONSTRAINT(0, 0x3F, 0);
1082 static struct event_constraint amd_f15_PMC53 = EVENT_CONSTRAINT(0, 0x38, 0);
1083 
1084 static struct event_constraint *
1085 amd_get_event_constraints_f15h(struct cpu_hw_events *cpuc, int idx,
1086 			       struct perf_event *event)
1087 {
1088 	struct hw_perf_event *hwc = &event->hw;
1089 	unsigned int event_code = amd_get_event_code(hwc);
1090 
1091 	switch (event_code & AMD_EVENT_TYPE_MASK) {
1092 	case AMD_EVENT_FP:
1093 		switch (event_code) {
1094 		case 0x000:
1095 			if (!(hwc->config & 0x0000F000ULL))
1096 				break;
1097 			if (!(hwc->config & 0x00000F00ULL))
1098 				break;
1099 			return &amd_f15_PMC3;
1100 		case 0x004:
1101 			if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
1102 				break;
1103 			return &amd_f15_PMC3;
1104 		case 0x003:
1105 		case 0x00B:
1106 		case 0x00D:
1107 			return &amd_f15_PMC3;
1108 		}
1109 		return &amd_f15_PMC53;
1110 	case AMD_EVENT_LS:
1111 	case AMD_EVENT_DC:
1112 	case AMD_EVENT_EX_LS:
1113 		switch (event_code) {
1114 		case 0x023:
1115 		case 0x043:
1116 		case 0x045:
1117 		case 0x046:
1118 		case 0x054:
1119 		case 0x055:
1120 			return &amd_f15_PMC20;
1121 		case 0x02D:
1122 			return &amd_f15_PMC3;
1123 		case 0x02E:
1124 			return &amd_f15_PMC30;
1125 		case 0x031:
1126 			if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
1127 				return &amd_f15_PMC20;
1128 			return &emptyconstraint;
1129 		case 0x1C0:
1130 			return &amd_f15_PMC53;
1131 		default:
1132 			return &amd_f15_PMC50;
1133 		}
1134 	case AMD_EVENT_CU:
1135 	case AMD_EVENT_IC_DE:
1136 	case AMD_EVENT_DE:
1137 		switch (event_code) {
1138 		case 0x08F:
1139 		case 0x187:
1140 		case 0x188:
1141 			return &amd_f15_PMC0;
1142 		case 0x0DB ... 0x0DF:
1143 		case 0x1D6:
1144 		case 0x1D8:
1145 			return &amd_f15_PMC50;
1146 		default:
1147 			return &amd_f15_PMC20;
1148 		}
1149 	case AMD_EVENT_NB:
1150 		/* moved to uncore.c */
1151 		return &emptyconstraint;
1152 	default:
1153 		return &emptyconstraint;
1154 	}
1155 }
1156 
1157 static struct event_constraint pair_constraint;
1158 
1159 static struct event_constraint *
1160 amd_get_event_constraints_f17h(struct cpu_hw_events *cpuc, int idx,
1161 			       struct perf_event *event)
1162 {
1163 	struct hw_perf_event *hwc = &event->hw;
1164 
1165 	if (amd_is_pair_event_code(hwc))
1166 		return &pair_constraint;
1167 
1168 	return &unconstrained;
1169 }
1170 
1171 static void amd_put_event_constraints_f17h(struct cpu_hw_events *cpuc,
1172 					   struct perf_event *event)
1173 {
1174 	struct hw_perf_event *hwc = &event->hw;
1175 
1176 	if (is_counter_pair(hwc))
1177 		--cpuc->n_pair;
1178 }
1179 
1180 /*
1181  * Because of the way BRS operates with an inactive and active phases, and
1182  * the link to one counter, it is not possible to have two events using BRS
1183  * scheduled at the same time. There would be an issue with enforcing the
1184  * period of each one and given that the BRS saturates, it would not be possible
1185  * to guarantee correlated content for all events. Therefore, in situations
1186  * where multiple events want to use BRS, the kernel enforces mutual exclusion.
1187  * Exclusion is enforced by chosing only one counter for events using BRS.
1188  * The event scheduling logic will then automatically multiplex the
1189  * events and ensure that at most one event is actively using BRS.
1190  *
1191  * The BRS counter could be any counter, but there is no constraint on Fam19h,
1192  * therefore all counters are equal and thus we pick the first one: PMC0
1193  */
1194 static struct event_constraint amd_fam19h_brs_cntr0_constraint =
1195 	EVENT_CONSTRAINT(0, 0x1, AMD64_RAW_EVENT_MASK);
1196 
1197 static struct event_constraint amd_fam19h_brs_pair_cntr0_constraint =
1198 	__EVENT_CONSTRAINT(0, 0x1, AMD64_RAW_EVENT_MASK, 1, 0, PERF_X86_EVENT_PAIR);
1199 
1200 static struct event_constraint *
1201 amd_get_event_constraints_f19h(struct cpu_hw_events *cpuc, int idx,
1202 			  struct perf_event *event)
1203 {
1204 	struct hw_perf_event *hwc = &event->hw;
1205 	bool has_brs = has_amd_brs(hwc);
1206 
1207 	/*
1208 	 * In case BRS is used with an event requiring a counter pair,
1209 	 * the kernel allows it but only on counter 0 & 1 to enforce
1210 	 * multiplexing requiring to protect BRS in case of multiple
1211 	 * BRS users
1212 	 */
1213 	if (amd_is_pair_event_code(hwc)) {
1214 		return has_brs ? &amd_fam19h_brs_pair_cntr0_constraint
1215 			       : &pair_constraint;
1216 	}
1217 
1218 	if (has_brs)
1219 		return &amd_fam19h_brs_cntr0_constraint;
1220 
1221 	return &unconstrained;
1222 }
1223 
1224 
1225 static ssize_t amd_event_sysfs_show(char *page, u64 config)
1226 {
1227 	u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT) |
1228 		    (config & AMD64_EVENTSEL_EVENT) >> 24;
1229 
1230 	return x86_event_sysfs_show(page, config, event);
1231 }
1232 
1233 static void amd_pmu_limit_period(struct perf_event *event, s64 *left)
1234 {
1235 	/*
1236 	 * Decrease period by the depth of the BRS feature to get the last N
1237 	 * taken branches and approximate the desired period
1238 	 */
1239 	if (has_branch_stack(event) && *left > x86_pmu.lbr_nr)
1240 		*left -= x86_pmu.lbr_nr;
1241 }
1242 
1243 static __initconst const struct x86_pmu amd_pmu = {
1244 	.name			= "AMD",
1245 	.handle_irq		= amd_pmu_handle_irq,
1246 	.disable_all		= amd_pmu_disable_all,
1247 	.enable_all		= amd_pmu_enable_all,
1248 	.enable			= amd_pmu_enable_event,
1249 	.disable		= amd_pmu_disable_event,
1250 	.hw_config		= amd_pmu_hw_config,
1251 	.schedule_events	= x86_schedule_events,
1252 	.eventsel		= MSR_K7_EVNTSEL0,
1253 	.perfctr		= MSR_K7_PERFCTR0,
1254 	.addr_offset            = amd_pmu_addr_offset,
1255 	.event_map		= amd_pmu_event_map,
1256 	.max_events		= ARRAY_SIZE(amd_perfmon_event_map),
1257 	.num_counters		= AMD64_NUM_COUNTERS,
1258 	.add			= amd_pmu_add_event,
1259 	.del			= amd_pmu_del_event,
1260 	.cntval_bits		= 48,
1261 	.cntval_mask		= (1ULL << 48) - 1,
1262 	.apic			= 1,
1263 	/* use highest bit to detect overflow */
1264 	.max_period		= (1ULL << 47) - 1,
1265 	.get_event_constraints	= amd_get_event_constraints,
1266 	.put_event_constraints	= amd_put_event_constraints,
1267 
1268 	.format_attrs		= amd_format_attr,
1269 	.events_sysfs_show	= amd_event_sysfs_show,
1270 
1271 	.cpu_prepare		= amd_pmu_cpu_prepare,
1272 	.cpu_starting		= amd_pmu_cpu_starting,
1273 	.cpu_dead		= amd_pmu_cpu_dead,
1274 
1275 	.amd_nb_constraints	= 1,
1276 };
1277 
1278 static ssize_t branches_show(struct device *cdev,
1279 			      struct device_attribute *attr,
1280 			      char *buf)
1281 {
1282 	return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu.lbr_nr);
1283 }
1284 
1285 static DEVICE_ATTR_RO(branches);
1286 
1287 static struct attribute *amd_pmu_branches_attrs[] = {
1288 	&dev_attr_branches.attr,
1289 	NULL,
1290 };
1291 
1292 static umode_t
1293 amd_branches_is_visible(struct kobject *kobj, struct attribute *attr, int i)
1294 {
1295 	return x86_pmu.lbr_nr ? attr->mode : 0;
1296 }
1297 
1298 static struct attribute_group group_caps_amd_branches = {
1299 	.name  = "caps",
1300 	.attrs = amd_pmu_branches_attrs,
1301 	.is_visible = amd_branches_is_visible,
1302 };
1303 
1304 #ifdef CONFIG_PERF_EVENTS_AMD_BRS
1305 
1306 EVENT_ATTR_STR(branch-brs, amd_branch_brs,
1307 	       "event=" __stringify(AMD_FAM19H_BRS_EVENT)"\n");
1308 
1309 static struct attribute *amd_brs_events_attrs[] = {
1310 	EVENT_PTR(amd_branch_brs),
1311 	NULL,
1312 };
1313 
1314 static umode_t
1315 amd_brs_is_visible(struct kobject *kobj, struct attribute *attr, int i)
1316 {
1317 	return static_cpu_has(X86_FEATURE_BRS) && x86_pmu.lbr_nr ?
1318 	       attr->mode : 0;
1319 }
1320 
1321 static struct attribute_group group_events_amd_brs = {
1322 	.name       = "events",
1323 	.attrs      = amd_brs_events_attrs,
1324 	.is_visible = amd_brs_is_visible,
1325 };
1326 
1327 #endif	/* CONFIG_PERF_EVENTS_AMD_BRS */
1328 
1329 static const struct attribute_group *amd_attr_update[] = {
1330 	&group_caps_amd_branches,
1331 #ifdef CONFIG_PERF_EVENTS_AMD_BRS
1332 	&group_events_amd_brs,
1333 #endif
1334 	NULL,
1335 };
1336 
1337 static int __init amd_core_pmu_init(void)
1338 {
1339 	union cpuid_0x80000022_ebx ebx;
1340 	u64 even_ctr_mask = 0ULL;
1341 	int i;
1342 
1343 	if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
1344 		return 0;
1345 
1346 	/* Avoid calculating the value each time in the NMI handler */
1347 	perf_nmi_window = msecs_to_jiffies(100);
1348 
1349 	/*
1350 	 * If core performance counter extensions exists, we must use
1351 	 * MSR_F15H_PERF_CTL/MSR_F15H_PERF_CTR msrs. See also
1352 	 * amd_pmu_addr_offset().
1353 	 */
1354 	x86_pmu.eventsel	= MSR_F15H_PERF_CTL;
1355 	x86_pmu.perfctr		= MSR_F15H_PERF_CTR;
1356 	x86_pmu.num_counters	= AMD64_NUM_COUNTERS_CORE;
1357 
1358 	/* Check for Performance Monitoring v2 support */
1359 	if (boot_cpu_has(X86_FEATURE_PERFMON_V2)) {
1360 		ebx.full = cpuid_ebx(EXT_PERFMON_DEBUG_FEATURES);
1361 
1362 		/* Update PMU version for later usage */
1363 		x86_pmu.version = 2;
1364 
1365 		/* Find the number of available Core PMCs */
1366 		x86_pmu.num_counters = ebx.split.num_core_pmc;
1367 
1368 		amd_pmu_global_cntr_mask = (1ULL << x86_pmu.num_counters) - 1;
1369 
1370 		/* Update PMC handling functions */
1371 		x86_pmu.enable_all = amd_pmu_v2_enable_all;
1372 		x86_pmu.disable_all = amd_pmu_v2_disable_all;
1373 		x86_pmu.enable = amd_pmu_v2_enable_event;
1374 		x86_pmu.handle_irq = amd_pmu_v2_handle_irq;
1375 		static_call_update(amd_pmu_test_overflow, amd_pmu_test_overflow_status);
1376 	}
1377 
1378 	/*
1379 	 * AMD Core perfctr has separate MSRs for the NB events, see
1380 	 * the amd/uncore.c driver.
1381 	 */
1382 	x86_pmu.amd_nb_constraints = 0;
1383 
1384 	if (boot_cpu_data.x86 == 0x15) {
1385 		pr_cont("Fam15h ");
1386 		x86_pmu.get_event_constraints = amd_get_event_constraints_f15h;
1387 	}
1388 	if (boot_cpu_data.x86 >= 0x17) {
1389 		pr_cont("Fam17h+ ");
1390 		/*
1391 		 * Family 17h and compatibles have constraints for Large
1392 		 * Increment per Cycle events: they may only be assigned an
1393 		 * even numbered counter that has a consecutive adjacent odd
1394 		 * numbered counter following it.
1395 		 */
1396 		for (i = 0; i < x86_pmu.num_counters - 1; i += 2)
1397 			even_ctr_mask |= BIT_ULL(i);
1398 
1399 		pair_constraint = (struct event_constraint)
1400 				    __EVENT_CONSTRAINT(0, even_ctr_mask, 0,
1401 				    x86_pmu.num_counters / 2, 0,
1402 				    PERF_X86_EVENT_PAIR);
1403 
1404 		x86_pmu.get_event_constraints = amd_get_event_constraints_f17h;
1405 		x86_pmu.put_event_constraints = amd_put_event_constraints_f17h;
1406 		x86_pmu.perf_ctr_pair_en = AMD_MERGE_EVENT_ENABLE;
1407 		x86_pmu.flags |= PMU_FL_PAIR;
1408 	}
1409 
1410 	/* LBR and BRS are mutually exclusive features */
1411 	if (!amd_pmu_lbr_init()) {
1412 		/* LBR requires flushing on context switch */
1413 		x86_pmu.sched_task = amd_pmu_lbr_sched_task;
1414 		static_call_update(amd_pmu_branch_hw_config, amd_pmu_lbr_hw_config);
1415 		static_call_update(amd_pmu_branch_reset, amd_pmu_lbr_reset);
1416 		static_call_update(amd_pmu_branch_add, amd_pmu_lbr_add);
1417 		static_call_update(amd_pmu_branch_del, amd_pmu_lbr_del);
1418 	} else if (!amd_brs_init()) {
1419 		/*
1420 		 * BRS requires special event constraints and flushing on ctxsw.
1421 		 */
1422 		x86_pmu.get_event_constraints = amd_get_event_constraints_f19h;
1423 		x86_pmu.sched_task = amd_pmu_brs_sched_task;
1424 		x86_pmu.limit_period = amd_pmu_limit_period;
1425 
1426 		static_call_update(amd_pmu_branch_hw_config, amd_brs_hw_config);
1427 		static_call_update(amd_pmu_branch_reset, amd_brs_reset);
1428 		static_call_update(amd_pmu_branch_add, amd_pmu_brs_add);
1429 		static_call_update(amd_pmu_branch_del, amd_pmu_brs_del);
1430 
1431 		/*
1432 		 * put_event_constraints callback same as Fam17h, set above
1433 		 */
1434 
1435 		/* branch sampling must be stopped when entering low power */
1436 		amd_brs_lopwr_init();
1437 	}
1438 
1439 	x86_pmu.attr_update = amd_attr_update;
1440 
1441 	pr_cont("core perfctr, ");
1442 	return 0;
1443 }
1444 
1445 __init int amd_pmu_init(void)
1446 {
1447 	int ret;
1448 
1449 	/* Performance-monitoring supported from K7 and later: */
1450 	if (boot_cpu_data.x86 < 6)
1451 		return -ENODEV;
1452 
1453 	x86_pmu = amd_pmu;
1454 
1455 	ret = amd_core_pmu_init();
1456 	if (ret)
1457 		return ret;
1458 
1459 	if (num_possible_cpus() == 1) {
1460 		/*
1461 		 * No point in allocating data structures to serialize
1462 		 * against other CPUs, when there is only the one CPU.
1463 		 */
1464 		x86_pmu.amd_nb_constraints = 0;
1465 	}
1466 
1467 	if (boot_cpu_data.x86 >= 0x17)
1468 		memcpy(hw_cache_event_ids, amd_hw_cache_event_ids_f17h, sizeof(hw_cache_event_ids));
1469 	else
1470 		memcpy(hw_cache_event_ids, amd_hw_cache_event_ids, sizeof(hw_cache_event_ids));
1471 
1472 	return 0;
1473 }
1474 
1475 static inline void amd_pmu_reload_virt(void)
1476 {
1477 	if (x86_pmu.version >= 2) {
1478 		/*
1479 		 * Clear global enable bits, reprogram the PERF_CTL
1480 		 * registers with updated perf_ctr_virt_mask and then
1481 		 * set global enable bits once again
1482 		 */
1483 		amd_pmu_v2_disable_all();
1484 		amd_pmu_enable_all(0);
1485 		amd_pmu_v2_enable_all(0);
1486 		return;
1487 	}
1488 
1489 	amd_pmu_disable_all();
1490 	amd_pmu_enable_all(0);
1491 }
1492 
1493 void amd_pmu_enable_virt(void)
1494 {
1495 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1496 
1497 	cpuc->perf_ctr_virt_mask = 0;
1498 
1499 	/* Reload all events */
1500 	amd_pmu_reload_virt();
1501 }
1502 EXPORT_SYMBOL_GPL(amd_pmu_enable_virt);
1503 
1504 void amd_pmu_disable_virt(void)
1505 {
1506 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1507 
1508 	/*
1509 	 * We only mask out the Host-only bit so that host-only counting works
1510 	 * when SVM is disabled. If someone sets up a guest-only counter when
1511 	 * SVM is disabled the Guest-only bits still gets set and the counter
1512 	 * will not count anything.
1513 	 */
1514 	cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
1515 
1516 	/* Reload all events */
1517 	amd_pmu_reload_virt();
1518 }
1519 EXPORT_SYMBOL_GPL(amd_pmu_disable_virt);
1520