xref: /linux/arch/x86/events/amd/core.c (revision ae22a94997b8a03dcb3c922857c203246711f9d4)
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_amd_node_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 
608 	if (!x86_pmu.amd_nb_constraints)
609 		return;
610 
611 	if (cpuhw->amd_nb) {
612 		struct amd_nb *nb = cpuhw->amd_nb;
613 
614 		if (nb->nb_id == -1 || --nb->refcnt == 0)
615 			kfree(nb);
616 
617 		cpuhw->amd_nb = NULL;
618 	}
619 }
620 
621 static inline void amd_pmu_set_global_ctl(u64 ctl)
622 {
623 	wrmsrl(MSR_AMD64_PERF_CNTR_GLOBAL_CTL, ctl);
624 }
625 
626 static inline u64 amd_pmu_get_global_status(void)
627 {
628 	u64 status;
629 
630 	/* PerfCntrGlobalStatus is read-only */
631 	rdmsrl(MSR_AMD64_PERF_CNTR_GLOBAL_STATUS, status);
632 
633 	return status;
634 }
635 
636 static inline void amd_pmu_ack_global_status(u64 status)
637 {
638 	/*
639 	 * PerfCntrGlobalStatus is read-only but an overflow acknowledgment
640 	 * mechanism exists; writing 1 to a bit in PerfCntrGlobalStatusClr
641 	 * clears the same bit in PerfCntrGlobalStatus
642 	 */
643 
644 	wrmsrl(MSR_AMD64_PERF_CNTR_GLOBAL_STATUS_CLR, status);
645 }
646 
647 static bool amd_pmu_test_overflow_topbit(int idx)
648 {
649 	u64 counter;
650 
651 	rdmsrl(x86_pmu_event_addr(idx), counter);
652 
653 	return !(counter & BIT_ULL(x86_pmu.cntval_bits - 1));
654 }
655 
656 static bool amd_pmu_test_overflow_status(int idx)
657 {
658 	return amd_pmu_get_global_status() & BIT_ULL(idx);
659 }
660 
661 DEFINE_STATIC_CALL(amd_pmu_test_overflow, amd_pmu_test_overflow_topbit);
662 
663 /*
664  * When a PMC counter overflows, an NMI is used to process the event and
665  * reset the counter. NMI latency can result in the counter being updated
666  * before the NMI can run, which can result in what appear to be spurious
667  * NMIs. This function is intended to wait for the NMI to run and reset
668  * the counter to avoid possible unhandled NMI messages.
669  */
670 #define OVERFLOW_WAIT_COUNT	50
671 
672 static void amd_pmu_wait_on_overflow(int idx)
673 {
674 	unsigned int i;
675 
676 	/*
677 	 * Wait for the counter to be reset if it has overflowed. This loop
678 	 * should exit very, very quickly, but just in case, don't wait
679 	 * forever...
680 	 */
681 	for (i = 0; i < OVERFLOW_WAIT_COUNT; i++) {
682 		if (!static_call(amd_pmu_test_overflow)(idx))
683 			break;
684 
685 		/* Might be in IRQ context, so can't sleep */
686 		udelay(1);
687 	}
688 }
689 
690 static void amd_pmu_check_overflow(void)
691 {
692 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
693 	int idx;
694 
695 	/*
696 	 * This shouldn't be called from NMI context, but add a safeguard here
697 	 * to return, since if we're in NMI context we can't wait for an NMI
698 	 * to reset an overflowed counter value.
699 	 */
700 	if (in_nmi())
701 		return;
702 
703 	/*
704 	 * Check each counter for overflow and wait for it to be reset by the
705 	 * NMI if it has overflowed. This relies on the fact that all active
706 	 * counters are always enabled when this function is called and
707 	 * ARCH_PERFMON_EVENTSEL_INT is always set.
708 	 */
709 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
710 		if (!test_bit(idx, cpuc->active_mask))
711 			continue;
712 
713 		amd_pmu_wait_on_overflow(idx);
714 	}
715 }
716 
717 static void amd_pmu_enable_event(struct perf_event *event)
718 {
719 	x86_pmu_enable_event(event);
720 }
721 
722 static void amd_pmu_enable_all(int added)
723 {
724 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
725 	int idx;
726 
727 	amd_brs_enable_all();
728 
729 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
730 		/* only activate events which are marked as active */
731 		if (!test_bit(idx, cpuc->active_mask))
732 			continue;
733 
734 		amd_pmu_enable_event(cpuc->events[idx]);
735 	}
736 }
737 
738 static void amd_pmu_v2_enable_event(struct perf_event *event)
739 {
740 	struct hw_perf_event *hwc = &event->hw;
741 
742 	/*
743 	 * Testing cpu_hw_events.enabled should be skipped in this case unlike
744 	 * in x86_pmu_enable_event().
745 	 *
746 	 * Since cpu_hw_events.enabled is set only after returning from
747 	 * x86_pmu_start(), the PMCs must be programmed and kept ready.
748 	 * Counting starts only after x86_pmu_enable_all() is called.
749 	 */
750 	__x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
751 }
752 
753 static __always_inline void amd_pmu_core_enable_all(void)
754 {
755 	amd_pmu_set_global_ctl(amd_pmu_global_cntr_mask);
756 }
757 
758 static void amd_pmu_v2_enable_all(int added)
759 {
760 	amd_pmu_lbr_enable_all();
761 	amd_pmu_core_enable_all();
762 }
763 
764 static void amd_pmu_disable_event(struct perf_event *event)
765 {
766 	x86_pmu_disable_event(event);
767 
768 	/*
769 	 * This can be called from NMI context (via x86_pmu_stop). The counter
770 	 * may have overflowed, but either way, we'll never see it get reset
771 	 * by the NMI if we're already in the NMI. And the NMI latency support
772 	 * below will take care of any pending NMI that might have been
773 	 * generated by the overflow.
774 	 */
775 	if (in_nmi())
776 		return;
777 
778 	amd_pmu_wait_on_overflow(event->hw.idx);
779 }
780 
781 static void amd_pmu_disable_all(void)
782 {
783 	amd_brs_disable_all();
784 	x86_pmu_disable_all();
785 	amd_pmu_check_overflow();
786 }
787 
788 static __always_inline void amd_pmu_core_disable_all(void)
789 {
790 	amd_pmu_set_global_ctl(0);
791 }
792 
793 static void amd_pmu_v2_disable_all(void)
794 {
795 	amd_pmu_core_disable_all();
796 	amd_pmu_lbr_disable_all();
797 	amd_pmu_check_overflow();
798 }
799 
800 DEFINE_STATIC_CALL_NULL(amd_pmu_branch_add, *x86_pmu.add);
801 
802 static void amd_pmu_add_event(struct perf_event *event)
803 {
804 	if (needs_branch_stack(event))
805 		static_call(amd_pmu_branch_add)(event);
806 }
807 
808 DEFINE_STATIC_CALL_NULL(amd_pmu_branch_del, *x86_pmu.del);
809 
810 static void amd_pmu_del_event(struct perf_event *event)
811 {
812 	if (needs_branch_stack(event))
813 		static_call(amd_pmu_branch_del)(event);
814 }
815 
816 /*
817  * Because of NMI latency, if multiple PMC counters are active or other sources
818  * of NMIs are received, the perf NMI handler can handle one or more overflowed
819  * PMC counters outside of the NMI associated with the PMC overflow. If the NMI
820  * doesn't arrive at the LAPIC in time to become a pending NMI, then the kernel
821  * back-to-back NMI support won't be active. This PMC handler needs to take into
822  * account that this can occur, otherwise this could result in unknown NMI
823  * messages being issued. Examples of this is PMC overflow while in the NMI
824  * handler when multiple PMCs are active or PMC overflow while handling some
825  * other source of an NMI.
826  *
827  * Attempt to mitigate this by creating an NMI window in which un-handled NMIs
828  * received during this window will be claimed. This prevents extending the
829  * window past when it is possible that latent NMIs should be received. The
830  * per-CPU perf_nmi_tstamp will be set to the window end time whenever perf has
831  * handled a counter. When an un-handled NMI is received, it will be claimed
832  * only if arriving within that window.
833  */
834 static inline int amd_pmu_adjust_nmi_window(int handled)
835 {
836 	/*
837 	 * If a counter was handled, record a timestamp such that un-handled
838 	 * NMIs will be claimed if arriving within that window.
839 	 */
840 	if (handled) {
841 		this_cpu_write(perf_nmi_tstamp, jiffies + perf_nmi_window);
842 
843 		return handled;
844 	}
845 
846 	if (time_after(jiffies, this_cpu_read(perf_nmi_tstamp)))
847 		return NMI_DONE;
848 
849 	return NMI_HANDLED;
850 }
851 
852 static int amd_pmu_handle_irq(struct pt_regs *regs)
853 {
854 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
855 	int handled;
856 	int pmu_enabled;
857 
858 	/*
859 	 * Save the PMU state.
860 	 * It needs to be restored when leaving the handler.
861 	 */
862 	pmu_enabled = cpuc->enabled;
863 	cpuc->enabled = 0;
864 
865 	amd_brs_disable_all();
866 
867 	/* Drain BRS is in use (could be inactive) */
868 	if (cpuc->lbr_users)
869 		amd_brs_drain();
870 
871 	/* Process any counter overflows */
872 	handled = x86_pmu_handle_irq(regs);
873 
874 	cpuc->enabled = pmu_enabled;
875 	if (pmu_enabled)
876 		amd_brs_enable_all();
877 
878 	return amd_pmu_adjust_nmi_window(handled);
879 }
880 
881 static int amd_pmu_v2_handle_irq(struct pt_regs *regs)
882 {
883 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
884 	struct perf_sample_data data;
885 	struct hw_perf_event *hwc;
886 	struct perf_event *event;
887 	int handled = 0, idx;
888 	u64 reserved, status, mask;
889 	bool pmu_enabled;
890 
891 	/*
892 	 * Save the PMU state as it needs to be restored when leaving the
893 	 * handler
894 	 */
895 	pmu_enabled = cpuc->enabled;
896 	cpuc->enabled = 0;
897 
898 	/* Stop counting but do not disable LBR */
899 	amd_pmu_core_disable_all();
900 
901 	status = amd_pmu_get_global_status();
902 
903 	/* Check if any overflows are pending */
904 	if (!status)
905 		goto done;
906 
907 	/* Read branch records before unfreezing */
908 	if (status & GLOBAL_STATUS_LBRS_FROZEN) {
909 		amd_pmu_lbr_read();
910 		status &= ~GLOBAL_STATUS_LBRS_FROZEN;
911 	}
912 
913 	reserved = status & ~amd_pmu_global_cntr_mask;
914 	if (reserved)
915 		pr_warn_once("Reserved PerfCntrGlobalStatus bits are set (0x%llx), please consider updating microcode\n",
916 			     reserved);
917 
918 	/* Clear any reserved bits set by buggy microcode */
919 	status &= amd_pmu_global_cntr_mask;
920 
921 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
922 		if (!test_bit(idx, cpuc->active_mask))
923 			continue;
924 
925 		event = cpuc->events[idx];
926 		hwc = &event->hw;
927 		x86_perf_event_update(event);
928 		mask = BIT_ULL(idx);
929 
930 		if (!(status & mask))
931 			continue;
932 
933 		/* Event overflow */
934 		handled++;
935 		status &= ~mask;
936 		perf_sample_data_init(&data, 0, hwc->last_period);
937 
938 		if (!x86_perf_event_set_period(event))
939 			continue;
940 
941 		if (has_branch_stack(event))
942 			perf_sample_save_brstack(&data, event, &cpuc->lbr_stack, NULL);
943 
944 		if (perf_event_overflow(event, &data, regs))
945 			x86_pmu_stop(event, 0);
946 	}
947 
948 	/*
949 	 * It should never be the case that some overflows are not handled as
950 	 * the corresponding PMCs are expected to be inactive according to the
951 	 * active_mask
952 	 */
953 	WARN_ON(status > 0);
954 
955 	/* Clear overflow and freeze bits */
956 	amd_pmu_ack_global_status(~status);
957 
958 	/*
959 	 * Unmasking the LVTPC is not required as the Mask (M) bit of the LVT
960 	 * PMI entry is not set by the local APIC when a PMC overflow occurs
961 	 */
962 	inc_irq_stat(apic_perf_irqs);
963 
964 done:
965 	cpuc->enabled = pmu_enabled;
966 
967 	/* Resume counting only if PMU is active */
968 	if (pmu_enabled)
969 		amd_pmu_core_enable_all();
970 
971 	return amd_pmu_adjust_nmi_window(handled);
972 }
973 
974 static struct event_constraint *
975 amd_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
976 			  struct perf_event *event)
977 {
978 	/*
979 	 * if not NB event or no NB, then no constraints
980 	 */
981 	if (!(amd_has_nb(cpuc) && amd_is_nb_event(&event->hw)))
982 		return &unconstrained;
983 
984 	return __amd_get_nb_event_constraints(cpuc, event, NULL);
985 }
986 
987 static void amd_put_event_constraints(struct cpu_hw_events *cpuc,
988 				      struct perf_event *event)
989 {
990 	if (amd_has_nb(cpuc) && amd_is_nb_event(&event->hw))
991 		__amd_put_nb_event_constraints(cpuc, event);
992 }
993 
994 PMU_FORMAT_ATTR(event,	"config:0-7,32-35");
995 PMU_FORMAT_ATTR(umask,	"config:8-15"	);
996 PMU_FORMAT_ATTR(edge,	"config:18"	);
997 PMU_FORMAT_ATTR(inv,	"config:23"	);
998 PMU_FORMAT_ATTR(cmask,	"config:24-31"	);
999 
1000 static struct attribute *amd_format_attr[] = {
1001 	&format_attr_event.attr,
1002 	&format_attr_umask.attr,
1003 	&format_attr_edge.attr,
1004 	&format_attr_inv.attr,
1005 	&format_attr_cmask.attr,
1006 	NULL,
1007 };
1008 
1009 /* AMD Family 15h */
1010 
1011 #define AMD_EVENT_TYPE_MASK	0x000000F0ULL
1012 
1013 #define AMD_EVENT_FP		0x00000000ULL ... 0x00000010ULL
1014 #define AMD_EVENT_LS		0x00000020ULL ... 0x00000030ULL
1015 #define AMD_EVENT_DC		0x00000040ULL ... 0x00000050ULL
1016 #define AMD_EVENT_CU		0x00000060ULL ... 0x00000070ULL
1017 #define AMD_EVENT_IC_DE		0x00000080ULL ... 0x00000090ULL
1018 #define AMD_EVENT_EX_LS		0x000000C0ULL
1019 #define AMD_EVENT_DE		0x000000D0ULL
1020 #define AMD_EVENT_NB		0x000000E0ULL ... 0x000000F0ULL
1021 
1022 /*
1023  * AMD family 15h event code/PMC mappings:
1024  *
1025  * type = event_code & 0x0F0:
1026  *
1027  * 0x000	FP	PERF_CTL[5:3]
1028  * 0x010	FP	PERF_CTL[5:3]
1029  * 0x020	LS	PERF_CTL[5:0]
1030  * 0x030	LS	PERF_CTL[5:0]
1031  * 0x040	DC	PERF_CTL[5:0]
1032  * 0x050	DC	PERF_CTL[5:0]
1033  * 0x060	CU	PERF_CTL[2:0]
1034  * 0x070	CU	PERF_CTL[2:0]
1035  * 0x080	IC/DE	PERF_CTL[2:0]
1036  * 0x090	IC/DE	PERF_CTL[2:0]
1037  * 0x0A0	---
1038  * 0x0B0	---
1039  * 0x0C0	EX/LS	PERF_CTL[5:0]
1040  * 0x0D0	DE	PERF_CTL[2:0]
1041  * 0x0E0	NB	NB_PERF_CTL[3:0]
1042  * 0x0F0	NB	NB_PERF_CTL[3:0]
1043  *
1044  * Exceptions:
1045  *
1046  * 0x000	FP	PERF_CTL[3], PERF_CTL[5:3] (*)
1047  * 0x003	FP	PERF_CTL[3]
1048  * 0x004	FP	PERF_CTL[3], PERF_CTL[5:3] (*)
1049  * 0x00B	FP	PERF_CTL[3]
1050  * 0x00D	FP	PERF_CTL[3]
1051  * 0x023	DE	PERF_CTL[2:0]
1052  * 0x02D	LS	PERF_CTL[3]
1053  * 0x02E	LS	PERF_CTL[3,0]
1054  * 0x031	LS	PERF_CTL[2:0] (**)
1055  * 0x043	CU	PERF_CTL[2:0]
1056  * 0x045	CU	PERF_CTL[2:0]
1057  * 0x046	CU	PERF_CTL[2:0]
1058  * 0x054	CU	PERF_CTL[2:0]
1059  * 0x055	CU	PERF_CTL[2:0]
1060  * 0x08F	IC	PERF_CTL[0]
1061  * 0x187	DE	PERF_CTL[0]
1062  * 0x188	DE	PERF_CTL[0]
1063  * 0x0DB	EX	PERF_CTL[5:0]
1064  * 0x0DC	LS	PERF_CTL[5:0]
1065  * 0x0DD	LS	PERF_CTL[5:0]
1066  * 0x0DE	LS	PERF_CTL[5:0]
1067  * 0x0DF	LS	PERF_CTL[5:0]
1068  * 0x1C0	EX	PERF_CTL[5:3]
1069  * 0x1D6	EX	PERF_CTL[5:0]
1070  * 0x1D8	EX	PERF_CTL[5:0]
1071  *
1072  * (*)  depending on the umask all FPU counters may be used
1073  * (**) only one unitmask enabled at a time
1074  */
1075 
1076 static struct event_constraint amd_f15_PMC0  = EVENT_CONSTRAINT(0, 0x01, 0);
1077 static struct event_constraint amd_f15_PMC20 = EVENT_CONSTRAINT(0, 0x07, 0);
1078 static struct event_constraint amd_f15_PMC3  = EVENT_CONSTRAINT(0, 0x08, 0);
1079 static struct event_constraint amd_f15_PMC30 = EVENT_CONSTRAINT_OVERLAP(0, 0x09, 0);
1080 static struct event_constraint amd_f15_PMC50 = EVENT_CONSTRAINT(0, 0x3F, 0);
1081 static struct event_constraint amd_f15_PMC53 = EVENT_CONSTRAINT(0, 0x38, 0);
1082 
1083 static struct event_constraint *
1084 amd_get_event_constraints_f15h(struct cpu_hw_events *cpuc, int idx,
1085 			       struct perf_event *event)
1086 {
1087 	struct hw_perf_event *hwc = &event->hw;
1088 	unsigned int event_code = amd_get_event_code(hwc);
1089 
1090 	switch (event_code & AMD_EVENT_TYPE_MASK) {
1091 	case AMD_EVENT_FP:
1092 		switch (event_code) {
1093 		case 0x000:
1094 			if (!(hwc->config & 0x0000F000ULL))
1095 				break;
1096 			if (!(hwc->config & 0x00000F00ULL))
1097 				break;
1098 			return &amd_f15_PMC3;
1099 		case 0x004:
1100 			if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
1101 				break;
1102 			return &amd_f15_PMC3;
1103 		case 0x003:
1104 		case 0x00B:
1105 		case 0x00D:
1106 			return &amd_f15_PMC3;
1107 		}
1108 		return &amd_f15_PMC53;
1109 	case AMD_EVENT_LS:
1110 	case AMD_EVENT_DC:
1111 	case AMD_EVENT_EX_LS:
1112 		switch (event_code) {
1113 		case 0x023:
1114 		case 0x043:
1115 		case 0x045:
1116 		case 0x046:
1117 		case 0x054:
1118 		case 0x055:
1119 			return &amd_f15_PMC20;
1120 		case 0x02D:
1121 			return &amd_f15_PMC3;
1122 		case 0x02E:
1123 			return &amd_f15_PMC30;
1124 		case 0x031:
1125 			if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
1126 				return &amd_f15_PMC20;
1127 			return &emptyconstraint;
1128 		case 0x1C0:
1129 			return &amd_f15_PMC53;
1130 		default:
1131 			return &amd_f15_PMC50;
1132 		}
1133 	case AMD_EVENT_CU:
1134 	case AMD_EVENT_IC_DE:
1135 	case AMD_EVENT_DE:
1136 		switch (event_code) {
1137 		case 0x08F:
1138 		case 0x187:
1139 		case 0x188:
1140 			return &amd_f15_PMC0;
1141 		case 0x0DB ... 0x0DF:
1142 		case 0x1D6:
1143 		case 0x1D8:
1144 			return &amd_f15_PMC50;
1145 		default:
1146 			return &amd_f15_PMC20;
1147 		}
1148 	case AMD_EVENT_NB:
1149 		/* moved to uncore.c */
1150 		return &emptyconstraint;
1151 	default:
1152 		return &emptyconstraint;
1153 	}
1154 }
1155 
1156 static struct event_constraint pair_constraint;
1157 
1158 static struct event_constraint *
1159 amd_get_event_constraints_f17h(struct cpu_hw_events *cpuc, int idx,
1160 			       struct perf_event *event)
1161 {
1162 	struct hw_perf_event *hwc = &event->hw;
1163 
1164 	if (amd_is_pair_event_code(hwc))
1165 		return &pair_constraint;
1166 
1167 	return &unconstrained;
1168 }
1169 
1170 static void amd_put_event_constraints_f17h(struct cpu_hw_events *cpuc,
1171 					   struct perf_event *event)
1172 {
1173 	struct hw_perf_event *hwc = &event->hw;
1174 
1175 	if (is_counter_pair(hwc))
1176 		--cpuc->n_pair;
1177 }
1178 
1179 /*
1180  * Because of the way BRS operates with an inactive and active phases, and
1181  * the link to one counter, it is not possible to have two events using BRS
1182  * scheduled at the same time. There would be an issue with enforcing the
1183  * period of each one and given that the BRS saturates, it would not be possible
1184  * to guarantee correlated content for all events. Therefore, in situations
1185  * where multiple events want to use BRS, the kernel enforces mutual exclusion.
1186  * Exclusion is enforced by choosing only one counter for events using BRS.
1187  * The event scheduling logic will then automatically multiplex the
1188  * events and ensure that at most one event is actively using BRS.
1189  *
1190  * The BRS counter could be any counter, but there is no constraint on Fam19h,
1191  * therefore all counters are equal and thus we pick the first one: PMC0
1192  */
1193 static struct event_constraint amd_fam19h_brs_cntr0_constraint =
1194 	EVENT_CONSTRAINT(0, 0x1, AMD64_RAW_EVENT_MASK);
1195 
1196 static struct event_constraint amd_fam19h_brs_pair_cntr0_constraint =
1197 	__EVENT_CONSTRAINT(0, 0x1, AMD64_RAW_EVENT_MASK, 1, 0, PERF_X86_EVENT_PAIR);
1198 
1199 static struct event_constraint *
1200 amd_get_event_constraints_f19h(struct cpu_hw_events *cpuc, int idx,
1201 			  struct perf_event *event)
1202 {
1203 	struct hw_perf_event *hwc = &event->hw;
1204 	bool has_brs = has_amd_brs(hwc);
1205 
1206 	/*
1207 	 * In case BRS is used with an event requiring a counter pair,
1208 	 * the kernel allows it but only on counter 0 & 1 to enforce
1209 	 * multiplexing requiring to protect BRS in case of multiple
1210 	 * BRS users
1211 	 */
1212 	if (amd_is_pair_event_code(hwc)) {
1213 		return has_brs ? &amd_fam19h_brs_pair_cntr0_constraint
1214 			       : &pair_constraint;
1215 	}
1216 
1217 	if (has_brs)
1218 		return &amd_fam19h_brs_cntr0_constraint;
1219 
1220 	return &unconstrained;
1221 }
1222 
1223 
1224 static ssize_t amd_event_sysfs_show(char *page, u64 config)
1225 {
1226 	u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT) |
1227 		    (config & AMD64_EVENTSEL_EVENT) >> 24;
1228 
1229 	return x86_event_sysfs_show(page, config, event);
1230 }
1231 
1232 static void amd_pmu_limit_period(struct perf_event *event, s64 *left)
1233 {
1234 	/*
1235 	 * Decrease period by the depth of the BRS feature to get the last N
1236 	 * taken branches and approximate the desired period
1237 	 */
1238 	if (has_branch_stack(event) && *left > x86_pmu.lbr_nr)
1239 		*left -= x86_pmu.lbr_nr;
1240 }
1241 
1242 static __initconst const struct x86_pmu amd_pmu = {
1243 	.name			= "AMD",
1244 	.handle_irq		= amd_pmu_handle_irq,
1245 	.disable_all		= amd_pmu_disable_all,
1246 	.enable_all		= amd_pmu_enable_all,
1247 	.enable			= amd_pmu_enable_event,
1248 	.disable		= amd_pmu_disable_event,
1249 	.hw_config		= amd_pmu_hw_config,
1250 	.schedule_events	= x86_schedule_events,
1251 	.eventsel		= MSR_K7_EVNTSEL0,
1252 	.perfctr		= MSR_K7_PERFCTR0,
1253 	.addr_offset            = amd_pmu_addr_offset,
1254 	.event_map		= amd_pmu_event_map,
1255 	.max_events		= ARRAY_SIZE(amd_perfmon_event_map),
1256 	.num_counters		= AMD64_NUM_COUNTERS,
1257 	.add			= amd_pmu_add_event,
1258 	.del			= amd_pmu_del_event,
1259 	.cntval_bits		= 48,
1260 	.cntval_mask		= (1ULL << 48) - 1,
1261 	.apic			= 1,
1262 	/* use highest bit to detect overflow */
1263 	.max_period		= (1ULL << 47) - 1,
1264 	.get_event_constraints	= amd_get_event_constraints,
1265 	.put_event_constraints	= amd_put_event_constraints,
1266 
1267 	.format_attrs		= amd_format_attr,
1268 	.events_sysfs_show	= amd_event_sysfs_show,
1269 
1270 	.cpu_prepare		= amd_pmu_cpu_prepare,
1271 	.cpu_starting		= amd_pmu_cpu_starting,
1272 	.cpu_dead		= amd_pmu_cpu_dead,
1273 
1274 	.amd_nb_constraints	= 1,
1275 };
1276 
1277 static ssize_t branches_show(struct device *cdev,
1278 			      struct device_attribute *attr,
1279 			      char *buf)
1280 {
1281 	return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu.lbr_nr);
1282 }
1283 
1284 static DEVICE_ATTR_RO(branches);
1285 
1286 static struct attribute *amd_pmu_branches_attrs[] = {
1287 	&dev_attr_branches.attr,
1288 	NULL,
1289 };
1290 
1291 static umode_t
1292 amd_branches_is_visible(struct kobject *kobj, struct attribute *attr, int i)
1293 {
1294 	return x86_pmu.lbr_nr ? attr->mode : 0;
1295 }
1296 
1297 static struct attribute_group group_caps_amd_branches = {
1298 	.name  = "caps",
1299 	.attrs = amd_pmu_branches_attrs,
1300 	.is_visible = amd_branches_is_visible,
1301 };
1302 
1303 #ifdef CONFIG_PERF_EVENTS_AMD_BRS
1304 
1305 EVENT_ATTR_STR(branch-brs, amd_branch_brs,
1306 	       "event=" __stringify(AMD_FAM19H_BRS_EVENT)"\n");
1307 
1308 static struct attribute *amd_brs_events_attrs[] = {
1309 	EVENT_PTR(amd_branch_brs),
1310 	NULL,
1311 };
1312 
1313 static umode_t
1314 amd_brs_is_visible(struct kobject *kobj, struct attribute *attr, int i)
1315 {
1316 	return static_cpu_has(X86_FEATURE_BRS) && x86_pmu.lbr_nr ?
1317 	       attr->mode : 0;
1318 }
1319 
1320 static struct attribute_group group_events_amd_brs = {
1321 	.name       = "events",
1322 	.attrs      = amd_brs_events_attrs,
1323 	.is_visible = amd_brs_is_visible,
1324 };
1325 
1326 #endif	/* CONFIG_PERF_EVENTS_AMD_BRS */
1327 
1328 static const struct attribute_group *amd_attr_update[] = {
1329 	&group_caps_amd_branches,
1330 #ifdef CONFIG_PERF_EVENTS_AMD_BRS
1331 	&group_events_amd_brs,
1332 #endif
1333 	NULL,
1334 };
1335 
1336 static int __init amd_core_pmu_init(void)
1337 {
1338 	union cpuid_0x80000022_ebx ebx;
1339 	u64 even_ctr_mask = 0ULL;
1340 	int i;
1341 
1342 	if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
1343 		return 0;
1344 
1345 	/* Avoid calculating the value each time in the NMI handler */
1346 	perf_nmi_window = msecs_to_jiffies(100);
1347 
1348 	/*
1349 	 * If core performance counter extensions exists, we must use
1350 	 * MSR_F15H_PERF_CTL/MSR_F15H_PERF_CTR msrs. See also
1351 	 * amd_pmu_addr_offset().
1352 	 */
1353 	x86_pmu.eventsel	= MSR_F15H_PERF_CTL;
1354 	x86_pmu.perfctr		= MSR_F15H_PERF_CTR;
1355 	x86_pmu.num_counters	= AMD64_NUM_COUNTERS_CORE;
1356 
1357 	/* Check for Performance Monitoring v2 support */
1358 	if (boot_cpu_has(X86_FEATURE_PERFMON_V2)) {
1359 		ebx.full = cpuid_ebx(EXT_PERFMON_DEBUG_FEATURES);
1360 
1361 		/* Update PMU version for later usage */
1362 		x86_pmu.version = 2;
1363 
1364 		/* Find the number of available Core PMCs */
1365 		x86_pmu.num_counters = ebx.split.num_core_pmc;
1366 
1367 		amd_pmu_global_cntr_mask = (1ULL << x86_pmu.num_counters) - 1;
1368 
1369 		/* Update PMC handling functions */
1370 		x86_pmu.enable_all = amd_pmu_v2_enable_all;
1371 		x86_pmu.disable_all = amd_pmu_v2_disable_all;
1372 		x86_pmu.enable = amd_pmu_v2_enable_event;
1373 		x86_pmu.handle_irq = amd_pmu_v2_handle_irq;
1374 		static_call_update(amd_pmu_test_overflow, amd_pmu_test_overflow_status);
1375 	}
1376 
1377 	/*
1378 	 * AMD Core perfctr has separate MSRs for the NB events, see
1379 	 * the amd/uncore.c driver.
1380 	 */
1381 	x86_pmu.amd_nb_constraints = 0;
1382 
1383 	if (boot_cpu_data.x86 == 0x15) {
1384 		pr_cont("Fam15h ");
1385 		x86_pmu.get_event_constraints = amd_get_event_constraints_f15h;
1386 	}
1387 	if (boot_cpu_data.x86 >= 0x17) {
1388 		pr_cont("Fam17h+ ");
1389 		/*
1390 		 * Family 17h and compatibles have constraints for Large
1391 		 * Increment per Cycle events: they may only be assigned an
1392 		 * even numbered counter that has a consecutive adjacent odd
1393 		 * numbered counter following it.
1394 		 */
1395 		for (i = 0; i < x86_pmu.num_counters - 1; i += 2)
1396 			even_ctr_mask |= BIT_ULL(i);
1397 
1398 		pair_constraint = (struct event_constraint)
1399 				    __EVENT_CONSTRAINT(0, even_ctr_mask, 0,
1400 				    x86_pmu.num_counters / 2, 0,
1401 				    PERF_X86_EVENT_PAIR);
1402 
1403 		x86_pmu.get_event_constraints = amd_get_event_constraints_f17h;
1404 		x86_pmu.put_event_constraints = amd_put_event_constraints_f17h;
1405 		x86_pmu.perf_ctr_pair_en = AMD_MERGE_EVENT_ENABLE;
1406 		x86_pmu.flags |= PMU_FL_PAIR;
1407 	}
1408 
1409 	/* LBR and BRS are mutually exclusive features */
1410 	if (!amd_pmu_lbr_init()) {
1411 		/* LBR requires flushing on context switch */
1412 		x86_pmu.sched_task = amd_pmu_lbr_sched_task;
1413 		static_call_update(amd_pmu_branch_hw_config, amd_pmu_lbr_hw_config);
1414 		static_call_update(amd_pmu_branch_reset, amd_pmu_lbr_reset);
1415 		static_call_update(amd_pmu_branch_add, amd_pmu_lbr_add);
1416 		static_call_update(amd_pmu_branch_del, amd_pmu_lbr_del);
1417 	} else if (!amd_brs_init()) {
1418 		/*
1419 		 * BRS requires special event constraints and flushing on ctxsw.
1420 		 */
1421 		x86_pmu.get_event_constraints = amd_get_event_constraints_f19h;
1422 		x86_pmu.sched_task = amd_pmu_brs_sched_task;
1423 		x86_pmu.limit_period = amd_pmu_limit_period;
1424 
1425 		static_call_update(amd_pmu_branch_hw_config, amd_brs_hw_config);
1426 		static_call_update(amd_pmu_branch_reset, amd_brs_reset);
1427 		static_call_update(amd_pmu_branch_add, amd_pmu_brs_add);
1428 		static_call_update(amd_pmu_branch_del, amd_pmu_brs_del);
1429 
1430 		/*
1431 		 * put_event_constraints callback same as Fam17h, set above
1432 		 */
1433 
1434 		/* branch sampling must be stopped when entering low power */
1435 		amd_brs_lopwr_init();
1436 	}
1437 
1438 	x86_pmu.attr_update = amd_attr_update;
1439 
1440 	pr_cont("core perfctr, ");
1441 	return 0;
1442 }
1443 
1444 __init int amd_pmu_init(void)
1445 {
1446 	int ret;
1447 
1448 	/* Performance-monitoring supported from K7 and later: */
1449 	if (boot_cpu_data.x86 < 6)
1450 		return -ENODEV;
1451 
1452 	x86_pmu = amd_pmu;
1453 
1454 	ret = amd_core_pmu_init();
1455 	if (ret)
1456 		return ret;
1457 
1458 	if (num_possible_cpus() == 1) {
1459 		/*
1460 		 * No point in allocating data structures to serialize
1461 		 * against other CPUs, when there is only the one CPU.
1462 		 */
1463 		x86_pmu.amd_nb_constraints = 0;
1464 	}
1465 
1466 	if (boot_cpu_data.x86 >= 0x17)
1467 		memcpy(hw_cache_event_ids, amd_hw_cache_event_ids_f17h, sizeof(hw_cache_event_ids));
1468 	else
1469 		memcpy(hw_cache_event_ids, amd_hw_cache_event_ids, sizeof(hw_cache_event_ids));
1470 
1471 	return 0;
1472 }
1473 
1474 static inline void amd_pmu_reload_virt(void)
1475 {
1476 	if (x86_pmu.version >= 2) {
1477 		/*
1478 		 * Clear global enable bits, reprogram the PERF_CTL
1479 		 * registers with updated perf_ctr_virt_mask and then
1480 		 * set global enable bits once again
1481 		 */
1482 		amd_pmu_v2_disable_all();
1483 		amd_pmu_enable_all(0);
1484 		amd_pmu_v2_enable_all(0);
1485 		return;
1486 	}
1487 
1488 	amd_pmu_disable_all();
1489 	amd_pmu_enable_all(0);
1490 }
1491 
1492 void amd_pmu_enable_virt(void)
1493 {
1494 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1495 
1496 	cpuc->perf_ctr_virt_mask = 0;
1497 
1498 	/* Reload all events */
1499 	amd_pmu_reload_virt();
1500 }
1501 EXPORT_SYMBOL_GPL(amd_pmu_enable_virt);
1502 
1503 void amd_pmu_disable_virt(void)
1504 {
1505 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1506 
1507 	/*
1508 	 * We only mask out the Host-only bit so that host-only counting works
1509 	 * when SVM is disabled. If someone sets up a guest-only counter when
1510 	 * SVM is disabled the Guest-only bits still gets set and the counter
1511 	 * will not count anything.
1512 	 */
1513 	cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
1514 
1515 	/* Reload all events */
1516 	amd_pmu_reload_virt();
1517 }
1518 EXPORT_SYMBOL_GPL(amd_pmu_disable_virt);
1519