xref: /linux/arch/x86/events/amd/ibs.c (revision 8886640dade4ae2595fcdce511c8bcc716aa47d3)
1 /*
2  * Performance events - AMD IBS
3  *
4  *  Copyright (C) 2011 Advanced Micro Devices, Inc., Robert Richter
5  *
6  *  For licencing details see kernel-base/COPYING
7  */
8 
9 #include <linux/perf_event.h>
10 #include <linux/init.h>
11 #include <linux/export.h>
12 #include <linux/pci.h>
13 #include <linux/ptrace.h>
14 #include <linux/syscore_ops.h>
15 #include <linux/sched/clock.h>
16 
17 #include <asm/apic.h>
18 
19 #include "../perf_event.h"
20 
21 static u32 ibs_caps;
22 
23 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_AMD)
24 
25 #include <linux/kprobes.h>
26 #include <linux/hardirq.h>
27 
28 #include <asm/nmi.h>
29 #include <asm/amd-ibs.h>
30 
31 #define IBS_FETCH_CONFIG_MASK	(IBS_FETCH_RAND_EN | IBS_FETCH_MAX_CNT)
32 #define IBS_OP_CONFIG_MASK	IBS_OP_MAX_CNT
33 
34 
35 /*
36  * IBS states:
37  *
38  * ENABLED; tracks the pmu::add(), pmu::del() state, when set the counter is taken
39  * and any further add()s must fail.
40  *
41  * STARTED/STOPPING/STOPPED; deal with pmu::start(), pmu::stop() state but are
42  * complicated by the fact that the IBS hardware can send late NMIs (ie. after
43  * we've cleared the EN bit).
44  *
45  * In order to consume these late NMIs we have the STOPPED state, any NMI that
46  * happens after we've cleared the EN state will clear this bit and report the
47  * NMI handled (this is fundamentally racy in the face or multiple NMI sources,
48  * someone else can consume our BIT and our NMI will go unhandled).
49  *
50  * And since we cannot set/clear this separate bit together with the EN bit,
51  * there are races; if we cleared STARTED early, an NMI could land in
52  * between clearing STARTED and clearing the EN bit (in fact multiple NMIs
53  * could happen if the period is small enough), and consume our STOPPED bit
54  * and trigger streams of unhandled NMIs.
55  *
56  * If, however, we clear STARTED late, an NMI can hit between clearing the
57  * EN bit and clearing STARTED, still see STARTED set and process the event.
58  * If this event will have the VALID bit clear, we bail properly, but this
59  * is not a given. With VALID set we can end up calling pmu::stop() again
60  * (the throttle logic) and trigger the WARNs in there.
61  *
62  * So what we do is set STOPPING before clearing EN to avoid the pmu::stop()
63  * nesting, and clear STARTED late, so that we have a well defined state over
64  * the clearing of the EN bit.
65  *
66  * XXX: we could probably be using !atomic bitops for all this.
67  */
68 
69 enum ibs_states {
70 	IBS_ENABLED	= 0,
71 	IBS_STARTED	= 1,
72 	IBS_STOPPING	= 2,
73 	IBS_STOPPED	= 3,
74 
75 	IBS_MAX_STATES,
76 };
77 
78 struct cpu_perf_ibs {
79 	struct perf_event	*event;
80 	unsigned long		state[BITS_TO_LONGS(IBS_MAX_STATES)];
81 };
82 
83 struct perf_ibs {
84 	struct pmu			pmu;
85 	unsigned int			msr;
86 	u64				config_mask;
87 	u64				cnt_mask;
88 	u64				enable_mask;
89 	u64				valid_mask;
90 	u64				max_period;
91 	unsigned long			offset_mask[1];
92 	int				offset_max;
93 	unsigned int			fetch_count_reset_broken : 1;
94 	unsigned int			fetch_ignore_if_zero_rip : 1;
95 	struct cpu_perf_ibs __percpu	*pcpu;
96 
97 	u64				(*get_count)(u64 config);
98 };
99 
100 static int
101 perf_event_set_period(struct hw_perf_event *hwc, u64 min, u64 max, u64 *hw_period)
102 {
103 	s64 left = local64_read(&hwc->period_left);
104 	s64 period = hwc->sample_period;
105 	int overflow = 0;
106 
107 	/*
108 	 * If we are way outside a reasonable range then just skip forward:
109 	 */
110 	if (unlikely(left <= -period)) {
111 		left = period;
112 		local64_set(&hwc->period_left, left);
113 		hwc->last_period = period;
114 		overflow = 1;
115 	}
116 
117 	if (unlikely(left < (s64)min)) {
118 		left += period;
119 		local64_set(&hwc->period_left, left);
120 		hwc->last_period = period;
121 		overflow = 1;
122 	}
123 
124 	/*
125 	 * If the hw period that triggers the sw overflow is too short
126 	 * we might hit the irq handler. This biases the results.
127 	 * Thus we shorten the next-to-last period and set the last
128 	 * period to the max period.
129 	 */
130 	if (left > max) {
131 		left -= max;
132 		if (left > max)
133 			left = max;
134 		else if (left < min)
135 			left = min;
136 	}
137 
138 	*hw_period = (u64)left;
139 
140 	return overflow;
141 }
142 
143 static  int
144 perf_event_try_update(struct perf_event *event, u64 new_raw_count, int width)
145 {
146 	struct hw_perf_event *hwc = &event->hw;
147 	int shift = 64 - width;
148 	u64 prev_raw_count;
149 	u64 delta;
150 
151 	/*
152 	 * Careful: an NMI might modify the previous event value.
153 	 *
154 	 * Our tactic to handle this is to first atomically read and
155 	 * exchange a new raw count - then add that new-prev delta
156 	 * count to the generic event atomically:
157 	 */
158 	prev_raw_count = local64_read(&hwc->prev_count);
159 	if (!local64_try_cmpxchg(&hwc->prev_count,
160 				 &prev_raw_count, new_raw_count))
161 		return 0;
162 
163 	/*
164 	 * Now we have the new raw value and have updated the prev
165 	 * timestamp already. We can now calculate the elapsed delta
166 	 * (event-)time and add that to the generic event.
167 	 *
168 	 * Careful, not all hw sign-extends above the physical width
169 	 * of the count.
170 	 */
171 	delta = (new_raw_count << shift) - (prev_raw_count << shift);
172 	delta >>= shift;
173 
174 	local64_add(delta, &event->count);
175 	local64_sub(delta, &hwc->period_left);
176 
177 	return 1;
178 }
179 
180 static struct perf_ibs perf_ibs_fetch;
181 static struct perf_ibs perf_ibs_op;
182 
183 static struct perf_ibs *get_ibs_pmu(int type)
184 {
185 	if (perf_ibs_fetch.pmu.type == type)
186 		return &perf_ibs_fetch;
187 	if (perf_ibs_op.pmu.type == type)
188 		return &perf_ibs_op;
189 	return NULL;
190 }
191 
192 /*
193  * core pmu config -> IBS config
194  *
195  *  perf record -a -e cpu-cycles:p ...    # use ibs op counting cycle count
196  *  perf record -a -e r076:p ...          # same as -e cpu-cycles:p
197  *  perf record -a -e r0C1:p ...          # use ibs op counting micro-ops
198  *
199  * IbsOpCntCtl (bit 19) of IBS Execution Control Register (IbsOpCtl,
200  * MSRC001_1033) is used to select either cycle or micro-ops counting
201  * mode.
202  */
203 static int core_pmu_ibs_config(struct perf_event *event, u64 *config)
204 {
205 	switch (event->attr.type) {
206 	case PERF_TYPE_HARDWARE:
207 		switch (event->attr.config) {
208 		case PERF_COUNT_HW_CPU_CYCLES:
209 			*config = 0;
210 			return 0;
211 		}
212 		break;
213 	case PERF_TYPE_RAW:
214 		switch (event->attr.config) {
215 		case 0x0076:
216 			*config = 0;
217 			return 0;
218 		case 0x00C1:
219 			*config = IBS_OP_CNT_CTL;
220 			return 0;
221 		}
222 		break;
223 	default:
224 		return -ENOENT;
225 	}
226 
227 	return -EOPNOTSUPP;
228 }
229 
230 /*
231  * The rip of IBS samples has skid 0. Thus, IBS supports precise
232  * levels 1 and 2 and the PERF_EFLAGS_EXACT is set. In rare cases the
233  * rip is invalid when IBS was not able to record the rip correctly.
234  * We clear PERF_EFLAGS_EXACT and take the rip from pt_regs then.
235  */
236 int forward_event_to_ibs(struct perf_event *event)
237 {
238 	u64 config = 0;
239 
240 	if (!event->attr.precise_ip || event->attr.precise_ip > 2)
241 		return -EOPNOTSUPP;
242 
243 	if (!core_pmu_ibs_config(event, &config)) {
244 		event->attr.type = perf_ibs_op.pmu.type;
245 		event->attr.config = config;
246 	}
247 	return -ENOENT;
248 }
249 
250 /*
251  * Grouping of IBS events is not possible since IBS can have only
252  * one event active at any point in time.
253  */
254 static int validate_group(struct perf_event *event)
255 {
256 	struct perf_event *sibling;
257 
258 	if (event->group_leader == event)
259 		return 0;
260 
261 	if (event->group_leader->pmu == event->pmu)
262 		return -EINVAL;
263 
264 	for_each_sibling_event(sibling, event->group_leader) {
265 		if (sibling->pmu == event->pmu)
266 			return -EINVAL;
267 	}
268 	return 0;
269 }
270 
271 static int perf_ibs_init(struct perf_event *event)
272 {
273 	struct hw_perf_event *hwc = &event->hw;
274 	struct perf_ibs *perf_ibs;
275 	u64 max_cnt, config;
276 	int ret;
277 
278 	perf_ibs = get_ibs_pmu(event->attr.type);
279 	if (!perf_ibs)
280 		return -ENOENT;
281 
282 	config = event->attr.config;
283 
284 	if (event->pmu != &perf_ibs->pmu)
285 		return -ENOENT;
286 
287 	if (config & ~perf_ibs->config_mask)
288 		return -EINVAL;
289 
290 	if (has_branch_stack(event))
291 		return -EOPNOTSUPP;
292 
293 	ret = validate_group(event);
294 	if (ret)
295 		return ret;
296 
297 	if (hwc->sample_period) {
298 		if (config & perf_ibs->cnt_mask)
299 			/* raw max_cnt may not be set */
300 			return -EINVAL;
301 		if (!event->attr.sample_freq && hwc->sample_period & 0x0f)
302 			/*
303 			 * lower 4 bits can not be set in ibs max cnt,
304 			 * but allowing it in case we adjust the
305 			 * sample period to set a frequency.
306 			 */
307 			return -EINVAL;
308 		hwc->sample_period &= ~0x0FULL;
309 		if (!hwc->sample_period)
310 			hwc->sample_period = 0x10;
311 	} else {
312 		max_cnt = config & perf_ibs->cnt_mask;
313 		config &= ~perf_ibs->cnt_mask;
314 		event->attr.sample_period = max_cnt << 4;
315 		hwc->sample_period = event->attr.sample_period;
316 	}
317 
318 	if (!hwc->sample_period)
319 		return -EINVAL;
320 
321 	/*
322 	 * If we modify hwc->sample_period, we also need to update
323 	 * hwc->last_period and hwc->period_left.
324 	 */
325 	hwc->last_period = hwc->sample_period;
326 	local64_set(&hwc->period_left, hwc->sample_period);
327 
328 	hwc->config_base = perf_ibs->msr;
329 	hwc->config = config;
330 
331 	return 0;
332 }
333 
334 static int perf_ibs_set_period(struct perf_ibs *perf_ibs,
335 			       struct hw_perf_event *hwc, u64 *period)
336 {
337 	int overflow;
338 
339 	/* ignore lower 4 bits in min count: */
340 	overflow = perf_event_set_period(hwc, 1<<4, perf_ibs->max_period, period);
341 	local64_set(&hwc->prev_count, 0);
342 
343 	return overflow;
344 }
345 
346 static u64 get_ibs_fetch_count(u64 config)
347 {
348 	union ibs_fetch_ctl fetch_ctl = (union ibs_fetch_ctl)config;
349 
350 	return fetch_ctl.fetch_cnt << 4;
351 }
352 
353 static u64 get_ibs_op_count(u64 config)
354 {
355 	union ibs_op_ctl op_ctl = (union ibs_op_ctl)config;
356 	u64 count = 0;
357 
358 	/*
359 	 * If the internal 27-bit counter rolled over, the count is MaxCnt
360 	 * and the lower 7 bits of CurCnt are randomized.
361 	 * Otherwise CurCnt has the full 27-bit current counter value.
362 	 */
363 	if (op_ctl.op_val) {
364 		count = op_ctl.opmaxcnt << 4;
365 		if (ibs_caps & IBS_CAPS_OPCNTEXT)
366 			count += op_ctl.opmaxcnt_ext << 20;
367 	} else if (ibs_caps & IBS_CAPS_RDWROPCNT) {
368 		count = op_ctl.opcurcnt;
369 	}
370 
371 	return count;
372 }
373 
374 static void
375 perf_ibs_event_update(struct perf_ibs *perf_ibs, struct perf_event *event,
376 		      u64 *config)
377 {
378 	u64 count = perf_ibs->get_count(*config);
379 
380 	/*
381 	 * Set width to 64 since we do not overflow on max width but
382 	 * instead on max count. In perf_ibs_set_period() we clear
383 	 * prev count manually on overflow.
384 	 */
385 	while (!perf_event_try_update(event, count, 64)) {
386 		rdmsrl(event->hw.config_base, *config);
387 		count = perf_ibs->get_count(*config);
388 	}
389 }
390 
391 static inline void perf_ibs_enable_event(struct perf_ibs *perf_ibs,
392 					 struct hw_perf_event *hwc, u64 config)
393 {
394 	u64 tmp = hwc->config | config;
395 
396 	if (perf_ibs->fetch_count_reset_broken)
397 		wrmsrl(hwc->config_base, tmp & ~perf_ibs->enable_mask);
398 
399 	wrmsrl(hwc->config_base, tmp | perf_ibs->enable_mask);
400 }
401 
402 /*
403  * Erratum #420 Instruction-Based Sampling Engine May Generate
404  * Interrupt that Cannot Be Cleared:
405  *
406  * Must clear counter mask first, then clear the enable bit. See
407  * Revision Guide for AMD Family 10h Processors, Publication #41322.
408  */
409 static inline void perf_ibs_disable_event(struct perf_ibs *perf_ibs,
410 					  struct hw_perf_event *hwc, u64 config)
411 {
412 	config &= ~perf_ibs->cnt_mask;
413 	if (boot_cpu_data.x86 == 0x10)
414 		wrmsrl(hwc->config_base, config);
415 	config &= ~perf_ibs->enable_mask;
416 	wrmsrl(hwc->config_base, config);
417 }
418 
419 /*
420  * We cannot restore the ibs pmu state, so we always needs to update
421  * the event while stopping it and then reset the state when starting
422  * again. Thus, ignoring PERF_EF_RELOAD and PERF_EF_UPDATE flags in
423  * perf_ibs_start()/perf_ibs_stop() and instead always do it.
424  */
425 static void perf_ibs_start(struct perf_event *event, int flags)
426 {
427 	struct hw_perf_event *hwc = &event->hw;
428 	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
429 	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
430 	u64 period, config = 0;
431 
432 	if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
433 		return;
434 
435 	WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
436 	hwc->state = 0;
437 
438 	perf_ibs_set_period(perf_ibs, hwc, &period);
439 	if (perf_ibs == &perf_ibs_op && (ibs_caps & IBS_CAPS_OPCNTEXT)) {
440 		config |= period & IBS_OP_MAX_CNT_EXT_MASK;
441 		period &= ~IBS_OP_MAX_CNT_EXT_MASK;
442 	}
443 	config |= period >> 4;
444 
445 	/*
446 	 * Set STARTED before enabling the hardware, such that a subsequent NMI
447 	 * must observe it.
448 	 */
449 	set_bit(IBS_STARTED,    pcpu->state);
450 	clear_bit(IBS_STOPPING, pcpu->state);
451 	perf_ibs_enable_event(perf_ibs, hwc, config);
452 
453 	perf_event_update_userpage(event);
454 }
455 
456 static void perf_ibs_stop(struct perf_event *event, int flags)
457 {
458 	struct hw_perf_event *hwc = &event->hw;
459 	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
460 	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
461 	u64 config;
462 	int stopping;
463 
464 	if (test_and_set_bit(IBS_STOPPING, pcpu->state))
465 		return;
466 
467 	stopping = test_bit(IBS_STARTED, pcpu->state);
468 
469 	if (!stopping && (hwc->state & PERF_HES_UPTODATE))
470 		return;
471 
472 	rdmsrl(hwc->config_base, config);
473 
474 	if (stopping) {
475 		/*
476 		 * Set STOPPED before disabling the hardware, such that it
477 		 * must be visible to NMIs the moment we clear the EN bit,
478 		 * at which point we can generate an !VALID sample which
479 		 * we need to consume.
480 		 */
481 		set_bit(IBS_STOPPED, pcpu->state);
482 		perf_ibs_disable_event(perf_ibs, hwc, config);
483 		/*
484 		 * Clear STARTED after disabling the hardware; if it were
485 		 * cleared before an NMI hitting after the clear but before
486 		 * clearing the EN bit might think it a spurious NMI and not
487 		 * handle it.
488 		 *
489 		 * Clearing it after, however, creates the problem of the NMI
490 		 * handler seeing STARTED but not having a valid sample.
491 		 */
492 		clear_bit(IBS_STARTED, pcpu->state);
493 		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
494 		hwc->state |= PERF_HES_STOPPED;
495 	}
496 
497 	if (hwc->state & PERF_HES_UPTODATE)
498 		return;
499 
500 	/*
501 	 * Clear valid bit to not count rollovers on update, rollovers
502 	 * are only updated in the irq handler.
503 	 */
504 	config &= ~perf_ibs->valid_mask;
505 
506 	perf_ibs_event_update(perf_ibs, event, &config);
507 	hwc->state |= PERF_HES_UPTODATE;
508 }
509 
510 static int perf_ibs_add(struct perf_event *event, int flags)
511 {
512 	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
513 	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
514 
515 	if (test_and_set_bit(IBS_ENABLED, pcpu->state))
516 		return -ENOSPC;
517 
518 	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
519 
520 	pcpu->event = event;
521 
522 	if (flags & PERF_EF_START)
523 		perf_ibs_start(event, PERF_EF_RELOAD);
524 
525 	return 0;
526 }
527 
528 static void perf_ibs_del(struct perf_event *event, int flags)
529 {
530 	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
531 	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
532 
533 	if (!test_and_clear_bit(IBS_ENABLED, pcpu->state))
534 		return;
535 
536 	perf_ibs_stop(event, PERF_EF_UPDATE);
537 
538 	pcpu->event = NULL;
539 
540 	perf_event_update_userpage(event);
541 }
542 
543 static void perf_ibs_read(struct perf_event *event) { }
544 
545 /*
546  * We need to initialize with empty group if all attributes in the
547  * group are dynamic.
548  */
549 static struct attribute *attrs_empty[] = {
550 	NULL,
551 };
552 
553 static struct attribute_group empty_format_group = {
554 	.name = "format",
555 	.attrs = attrs_empty,
556 };
557 
558 static struct attribute_group empty_caps_group = {
559 	.name = "caps",
560 	.attrs = attrs_empty,
561 };
562 
563 static const struct attribute_group *empty_attr_groups[] = {
564 	&empty_format_group,
565 	&empty_caps_group,
566 	NULL,
567 };
568 
569 PMU_FORMAT_ATTR(rand_en,	"config:57");
570 PMU_FORMAT_ATTR(cnt_ctl,	"config:19");
571 PMU_EVENT_ATTR_STRING(l3missonly, fetch_l3missonly, "config:59");
572 PMU_EVENT_ATTR_STRING(l3missonly, op_l3missonly, "config:16");
573 PMU_EVENT_ATTR_STRING(zen4_ibs_extensions, zen4_ibs_extensions, "1");
574 
575 static umode_t
576 zen4_ibs_extensions_is_visible(struct kobject *kobj, struct attribute *attr, int i)
577 {
578 	return ibs_caps & IBS_CAPS_ZEN4 ? attr->mode : 0;
579 }
580 
581 static struct attribute *rand_en_attrs[] = {
582 	&format_attr_rand_en.attr,
583 	NULL,
584 };
585 
586 static struct attribute *fetch_l3missonly_attrs[] = {
587 	&fetch_l3missonly.attr.attr,
588 	NULL,
589 };
590 
591 static struct attribute *zen4_ibs_extensions_attrs[] = {
592 	&zen4_ibs_extensions.attr.attr,
593 	NULL,
594 };
595 
596 static struct attribute_group group_rand_en = {
597 	.name = "format",
598 	.attrs = rand_en_attrs,
599 };
600 
601 static struct attribute_group group_fetch_l3missonly = {
602 	.name = "format",
603 	.attrs = fetch_l3missonly_attrs,
604 	.is_visible = zen4_ibs_extensions_is_visible,
605 };
606 
607 static struct attribute_group group_zen4_ibs_extensions = {
608 	.name = "caps",
609 	.attrs = zen4_ibs_extensions_attrs,
610 	.is_visible = zen4_ibs_extensions_is_visible,
611 };
612 
613 static const struct attribute_group *fetch_attr_groups[] = {
614 	&group_rand_en,
615 	&empty_caps_group,
616 	NULL,
617 };
618 
619 static const struct attribute_group *fetch_attr_update[] = {
620 	&group_fetch_l3missonly,
621 	&group_zen4_ibs_extensions,
622 	NULL,
623 };
624 
625 static umode_t
626 cnt_ctl_is_visible(struct kobject *kobj, struct attribute *attr, int i)
627 {
628 	return ibs_caps & IBS_CAPS_OPCNT ? attr->mode : 0;
629 }
630 
631 static struct attribute *cnt_ctl_attrs[] = {
632 	&format_attr_cnt_ctl.attr,
633 	NULL,
634 };
635 
636 static struct attribute *op_l3missonly_attrs[] = {
637 	&op_l3missonly.attr.attr,
638 	NULL,
639 };
640 
641 static struct attribute_group group_cnt_ctl = {
642 	.name = "format",
643 	.attrs = cnt_ctl_attrs,
644 	.is_visible = cnt_ctl_is_visible,
645 };
646 
647 static struct attribute_group group_op_l3missonly = {
648 	.name = "format",
649 	.attrs = op_l3missonly_attrs,
650 	.is_visible = zen4_ibs_extensions_is_visible,
651 };
652 
653 static const struct attribute_group *op_attr_update[] = {
654 	&group_cnt_ctl,
655 	&group_op_l3missonly,
656 	&group_zen4_ibs_extensions,
657 	NULL,
658 };
659 
660 static struct perf_ibs perf_ibs_fetch = {
661 	.pmu = {
662 		.task_ctx_nr	= perf_hw_context,
663 
664 		.event_init	= perf_ibs_init,
665 		.add		= perf_ibs_add,
666 		.del		= perf_ibs_del,
667 		.start		= perf_ibs_start,
668 		.stop		= perf_ibs_stop,
669 		.read		= perf_ibs_read,
670 		.capabilities	= PERF_PMU_CAP_NO_EXCLUDE,
671 	},
672 	.msr			= MSR_AMD64_IBSFETCHCTL,
673 	.config_mask		= IBS_FETCH_CONFIG_MASK,
674 	.cnt_mask		= IBS_FETCH_MAX_CNT,
675 	.enable_mask		= IBS_FETCH_ENABLE,
676 	.valid_mask		= IBS_FETCH_VAL,
677 	.max_period		= IBS_FETCH_MAX_CNT << 4,
678 	.offset_mask		= { MSR_AMD64_IBSFETCH_REG_MASK },
679 	.offset_max		= MSR_AMD64_IBSFETCH_REG_COUNT,
680 
681 	.get_count		= get_ibs_fetch_count,
682 };
683 
684 static struct perf_ibs perf_ibs_op = {
685 	.pmu = {
686 		.task_ctx_nr	= perf_hw_context,
687 
688 		.event_init	= perf_ibs_init,
689 		.add		= perf_ibs_add,
690 		.del		= perf_ibs_del,
691 		.start		= perf_ibs_start,
692 		.stop		= perf_ibs_stop,
693 		.read		= perf_ibs_read,
694 		.capabilities	= PERF_PMU_CAP_NO_EXCLUDE,
695 	},
696 	.msr			= MSR_AMD64_IBSOPCTL,
697 	.config_mask		= IBS_OP_CONFIG_MASK,
698 	.cnt_mask		= IBS_OP_MAX_CNT | IBS_OP_CUR_CNT |
699 				  IBS_OP_CUR_CNT_RAND,
700 	.enable_mask		= IBS_OP_ENABLE,
701 	.valid_mask		= IBS_OP_VAL,
702 	.max_period		= IBS_OP_MAX_CNT << 4,
703 	.offset_mask		= { MSR_AMD64_IBSOP_REG_MASK },
704 	.offset_max		= MSR_AMD64_IBSOP_REG_COUNT,
705 
706 	.get_count		= get_ibs_op_count,
707 };
708 
709 static void perf_ibs_get_mem_op(union ibs_op_data3 *op_data3,
710 				struct perf_sample_data *data)
711 {
712 	union perf_mem_data_src *data_src = &data->data_src;
713 
714 	data_src->mem_op = PERF_MEM_OP_NA;
715 
716 	if (op_data3->ld_op)
717 		data_src->mem_op = PERF_MEM_OP_LOAD;
718 	else if (op_data3->st_op)
719 		data_src->mem_op = PERF_MEM_OP_STORE;
720 }
721 
722 /*
723  * Processors having CPUID_Fn8000001B_EAX[11] aka IBS_CAPS_ZEN4 has
724  * more fine granular DataSrc encodings. Others have coarse.
725  */
726 static u8 perf_ibs_data_src(union ibs_op_data2 *op_data2)
727 {
728 	if (ibs_caps & IBS_CAPS_ZEN4)
729 		return (op_data2->data_src_hi << 3) | op_data2->data_src_lo;
730 
731 	return op_data2->data_src_lo;
732 }
733 
734 #define	L(x)		(PERF_MEM_S(LVL, x) | PERF_MEM_S(LVL, HIT))
735 #define	LN(x)		PERF_MEM_S(LVLNUM, x)
736 #define	REM		PERF_MEM_S(REMOTE, REMOTE)
737 #define	HOPS(x)		PERF_MEM_S(HOPS, x)
738 
739 static u64 g_data_src[8] = {
740 	[IBS_DATA_SRC_LOC_CACHE]	  = L(L3) | L(REM_CCE1) | LN(ANY_CACHE) | HOPS(0),
741 	[IBS_DATA_SRC_DRAM]		  = L(LOC_RAM) | LN(RAM),
742 	[IBS_DATA_SRC_REM_CACHE]	  = L(REM_CCE2) | LN(ANY_CACHE) | REM | HOPS(1),
743 	[IBS_DATA_SRC_IO]		  = L(IO) | LN(IO),
744 };
745 
746 #define RMT_NODE_BITS			(1 << IBS_DATA_SRC_DRAM)
747 #define RMT_NODE_APPLICABLE(x)		(RMT_NODE_BITS & (1 << x))
748 
749 static u64 g_zen4_data_src[32] = {
750 	[IBS_DATA_SRC_EXT_LOC_CACHE]	  = L(L3) | LN(L3),
751 	[IBS_DATA_SRC_EXT_NEAR_CCX_CACHE] = L(REM_CCE1) | LN(ANY_CACHE) | REM | HOPS(0),
752 	[IBS_DATA_SRC_EXT_DRAM]		  = L(LOC_RAM) | LN(RAM),
753 	[IBS_DATA_SRC_EXT_FAR_CCX_CACHE]  = L(REM_CCE2) | LN(ANY_CACHE) | REM | HOPS(1),
754 	[IBS_DATA_SRC_EXT_PMEM]		  = LN(PMEM),
755 	[IBS_DATA_SRC_EXT_IO]		  = L(IO) | LN(IO),
756 	[IBS_DATA_SRC_EXT_EXT_MEM]	  = LN(CXL),
757 };
758 
759 #define ZEN4_RMT_NODE_BITS		((1 << IBS_DATA_SRC_EXT_DRAM) | \
760 					 (1 << IBS_DATA_SRC_EXT_PMEM) | \
761 					 (1 << IBS_DATA_SRC_EXT_EXT_MEM))
762 #define ZEN4_RMT_NODE_APPLICABLE(x)	(ZEN4_RMT_NODE_BITS & (1 << x))
763 
764 static __u64 perf_ibs_get_mem_lvl(union ibs_op_data2 *op_data2,
765 				  union ibs_op_data3 *op_data3,
766 				  struct perf_sample_data *data)
767 {
768 	union perf_mem_data_src *data_src = &data->data_src;
769 	u8 ibs_data_src = perf_ibs_data_src(op_data2);
770 
771 	data_src->mem_lvl = 0;
772 	data_src->mem_lvl_num = 0;
773 
774 	/*
775 	 * DcMiss, L2Miss, DataSrc, DcMissLat etc. are all invalid for Uncached
776 	 * memory accesses. So, check DcUcMemAcc bit early.
777 	 */
778 	if (op_data3->dc_uc_mem_acc && ibs_data_src != IBS_DATA_SRC_EXT_IO)
779 		return L(UNC) | LN(UNC);
780 
781 	/* L1 Hit */
782 	if (op_data3->dc_miss == 0)
783 		return L(L1) | LN(L1);
784 
785 	/* L2 Hit */
786 	if (op_data3->l2_miss == 0) {
787 		/* Erratum #1293 */
788 		if (boot_cpu_data.x86 != 0x19 || boot_cpu_data.x86_model > 0xF ||
789 		    !(op_data3->sw_pf || op_data3->dc_miss_no_mab_alloc))
790 			return L(L2) | LN(L2);
791 	}
792 
793 	/*
794 	 * OP_DATA2 is valid only for load ops. Skip all checks which
795 	 * uses OP_DATA2[DataSrc].
796 	 */
797 	if (data_src->mem_op != PERF_MEM_OP_LOAD)
798 		goto check_mab;
799 
800 	if (ibs_caps & IBS_CAPS_ZEN4) {
801 		u64 val = g_zen4_data_src[ibs_data_src];
802 
803 		if (!val)
804 			goto check_mab;
805 
806 		/* HOPS_1 because IBS doesn't provide remote socket detail */
807 		if (op_data2->rmt_node && ZEN4_RMT_NODE_APPLICABLE(ibs_data_src)) {
808 			if (ibs_data_src == IBS_DATA_SRC_EXT_DRAM)
809 				val = L(REM_RAM1) | LN(RAM) | REM | HOPS(1);
810 			else
811 				val |= REM | HOPS(1);
812 		}
813 
814 		return val;
815 	} else {
816 		u64 val = g_data_src[ibs_data_src];
817 
818 		if (!val)
819 			goto check_mab;
820 
821 		/* HOPS_1 because IBS doesn't provide remote socket detail */
822 		if (op_data2->rmt_node && RMT_NODE_APPLICABLE(ibs_data_src)) {
823 			if (ibs_data_src == IBS_DATA_SRC_DRAM)
824 				val = L(REM_RAM1) | LN(RAM) | REM | HOPS(1);
825 			else
826 				val |= REM | HOPS(1);
827 		}
828 
829 		return val;
830 	}
831 
832 check_mab:
833 	/*
834 	 * MAB (Miss Address Buffer) Hit. MAB keeps track of outstanding
835 	 * DC misses. However, such data may come from any level in mem
836 	 * hierarchy. IBS provides detail about both MAB as well as actual
837 	 * DataSrc simultaneously. Prioritize DataSrc over MAB, i.e. set
838 	 * MAB only when IBS fails to provide DataSrc.
839 	 */
840 	if (op_data3->dc_miss_no_mab_alloc)
841 		return L(LFB) | LN(LFB);
842 
843 	/* Don't set HIT with NA */
844 	return PERF_MEM_S(LVL, NA) | LN(NA);
845 }
846 
847 static bool perf_ibs_cache_hit_st_valid(void)
848 {
849 	/* 0: Uninitialized, 1: Valid, -1: Invalid */
850 	static int cache_hit_st_valid;
851 
852 	if (unlikely(!cache_hit_st_valid)) {
853 		if (boot_cpu_data.x86 == 0x19 &&
854 		    (boot_cpu_data.x86_model <= 0xF ||
855 		    (boot_cpu_data.x86_model >= 0x20 &&
856 		     boot_cpu_data.x86_model <= 0x5F))) {
857 			cache_hit_st_valid = -1;
858 		} else {
859 			cache_hit_st_valid = 1;
860 		}
861 	}
862 
863 	return cache_hit_st_valid == 1;
864 }
865 
866 static void perf_ibs_get_mem_snoop(union ibs_op_data2 *op_data2,
867 				   struct perf_sample_data *data)
868 {
869 	union perf_mem_data_src *data_src = &data->data_src;
870 	u8 ibs_data_src;
871 
872 	data_src->mem_snoop = PERF_MEM_SNOOP_NA;
873 
874 	if (!perf_ibs_cache_hit_st_valid() ||
875 	    data_src->mem_op != PERF_MEM_OP_LOAD ||
876 	    data_src->mem_lvl & PERF_MEM_LVL_L1 ||
877 	    data_src->mem_lvl & PERF_MEM_LVL_L2 ||
878 	    op_data2->cache_hit_st)
879 		return;
880 
881 	ibs_data_src = perf_ibs_data_src(op_data2);
882 
883 	if (ibs_caps & IBS_CAPS_ZEN4) {
884 		if (ibs_data_src == IBS_DATA_SRC_EXT_LOC_CACHE ||
885 		    ibs_data_src == IBS_DATA_SRC_EXT_NEAR_CCX_CACHE ||
886 		    ibs_data_src == IBS_DATA_SRC_EXT_FAR_CCX_CACHE)
887 			data_src->mem_snoop = PERF_MEM_SNOOP_HITM;
888 	} else if (ibs_data_src == IBS_DATA_SRC_LOC_CACHE) {
889 		data_src->mem_snoop = PERF_MEM_SNOOP_HITM;
890 	}
891 }
892 
893 static void perf_ibs_get_tlb_lvl(union ibs_op_data3 *op_data3,
894 				 struct perf_sample_data *data)
895 {
896 	union perf_mem_data_src *data_src = &data->data_src;
897 
898 	data_src->mem_dtlb = PERF_MEM_TLB_NA;
899 
900 	if (!op_data3->dc_lin_addr_valid)
901 		return;
902 
903 	if (!op_data3->dc_l1tlb_miss) {
904 		data_src->mem_dtlb = PERF_MEM_TLB_L1 | PERF_MEM_TLB_HIT;
905 		return;
906 	}
907 
908 	if (!op_data3->dc_l2tlb_miss) {
909 		data_src->mem_dtlb = PERF_MEM_TLB_L2 | PERF_MEM_TLB_HIT;
910 		return;
911 	}
912 
913 	data_src->mem_dtlb = PERF_MEM_TLB_L2 | PERF_MEM_TLB_MISS;
914 }
915 
916 static void perf_ibs_get_mem_lock(union ibs_op_data3 *op_data3,
917 				  struct perf_sample_data *data)
918 {
919 	union perf_mem_data_src *data_src = &data->data_src;
920 
921 	data_src->mem_lock = PERF_MEM_LOCK_NA;
922 
923 	if (op_data3->dc_locked_op)
924 		data_src->mem_lock = PERF_MEM_LOCK_LOCKED;
925 }
926 
927 #define ibs_op_msr_idx(msr)	(msr - MSR_AMD64_IBSOPCTL)
928 
929 static void perf_ibs_get_data_src(struct perf_ibs_data *ibs_data,
930 				  struct perf_sample_data *data,
931 				  union ibs_op_data2 *op_data2,
932 				  union ibs_op_data3 *op_data3)
933 {
934 	union perf_mem_data_src *data_src = &data->data_src;
935 
936 	data_src->val |= perf_ibs_get_mem_lvl(op_data2, op_data3, data);
937 	perf_ibs_get_mem_snoop(op_data2, data);
938 	perf_ibs_get_tlb_lvl(op_data3, data);
939 	perf_ibs_get_mem_lock(op_data3, data);
940 }
941 
942 static __u64 perf_ibs_get_op_data2(struct perf_ibs_data *ibs_data,
943 				   union ibs_op_data3 *op_data3)
944 {
945 	__u64 val = ibs_data->regs[ibs_op_msr_idx(MSR_AMD64_IBSOPDATA2)];
946 
947 	/* Erratum #1293 */
948 	if (boot_cpu_data.x86 == 0x19 && boot_cpu_data.x86_model <= 0xF &&
949 	    (op_data3->sw_pf || op_data3->dc_miss_no_mab_alloc)) {
950 		/*
951 		 * OP_DATA2 has only two fields on Zen3: DataSrc and RmtNode.
952 		 * DataSrc=0 is 'No valid status' and RmtNode is invalid when
953 		 * DataSrc=0.
954 		 */
955 		val = 0;
956 	}
957 	return val;
958 }
959 
960 static void perf_ibs_parse_ld_st_data(__u64 sample_type,
961 				      struct perf_ibs_data *ibs_data,
962 				      struct perf_sample_data *data)
963 {
964 	union ibs_op_data3 op_data3;
965 	union ibs_op_data2 op_data2;
966 	union ibs_op_data op_data;
967 
968 	data->data_src.val = PERF_MEM_NA;
969 	op_data3.val = ibs_data->regs[ibs_op_msr_idx(MSR_AMD64_IBSOPDATA3)];
970 
971 	perf_ibs_get_mem_op(&op_data3, data);
972 	if (data->data_src.mem_op != PERF_MEM_OP_LOAD &&
973 	    data->data_src.mem_op != PERF_MEM_OP_STORE)
974 		return;
975 
976 	op_data2.val = perf_ibs_get_op_data2(ibs_data, &op_data3);
977 
978 	if (sample_type & PERF_SAMPLE_DATA_SRC) {
979 		perf_ibs_get_data_src(ibs_data, data, &op_data2, &op_data3);
980 		data->sample_flags |= PERF_SAMPLE_DATA_SRC;
981 	}
982 
983 	if (sample_type & PERF_SAMPLE_WEIGHT_TYPE && op_data3.dc_miss &&
984 	    data->data_src.mem_op == PERF_MEM_OP_LOAD) {
985 		op_data.val = ibs_data->regs[ibs_op_msr_idx(MSR_AMD64_IBSOPDATA)];
986 
987 		if (sample_type & PERF_SAMPLE_WEIGHT_STRUCT) {
988 			data->weight.var1_dw = op_data3.dc_miss_lat;
989 			data->weight.var2_w = op_data.tag_to_ret_ctr;
990 		} else if (sample_type & PERF_SAMPLE_WEIGHT) {
991 			data->weight.full = op_data3.dc_miss_lat;
992 		}
993 		data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
994 	}
995 
996 	if (sample_type & PERF_SAMPLE_ADDR && op_data3.dc_lin_addr_valid) {
997 		data->addr = ibs_data->regs[ibs_op_msr_idx(MSR_AMD64_IBSDCLINAD)];
998 		data->sample_flags |= PERF_SAMPLE_ADDR;
999 	}
1000 
1001 	if (sample_type & PERF_SAMPLE_PHYS_ADDR && op_data3.dc_phy_addr_valid) {
1002 		data->phys_addr = ibs_data->regs[ibs_op_msr_idx(MSR_AMD64_IBSDCPHYSAD)];
1003 		data->sample_flags |= PERF_SAMPLE_PHYS_ADDR;
1004 	}
1005 }
1006 
1007 static int perf_ibs_get_offset_max(struct perf_ibs *perf_ibs, u64 sample_type,
1008 				   int check_rip)
1009 {
1010 	if (sample_type & PERF_SAMPLE_RAW ||
1011 	    (perf_ibs == &perf_ibs_op &&
1012 	     (sample_type & PERF_SAMPLE_DATA_SRC ||
1013 	      sample_type & PERF_SAMPLE_WEIGHT_TYPE ||
1014 	      sample_type & PERF_SAMPLE_ADDR ||
1015 	      sample_type & PERF_SAMPLE_PHYS_ADDR)))
1016 		return perf_ibs->offset_max;
1017 	else if (check_rip)
1018 		return 3;
1019 	return 1;
1020 }
1021 
1022 static int perf_ibs_handle_irq(struct perf_ibs *perf_ibs, struct pt_regs *iregs)
1023 {
1024 	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
1025 	struct perf_event *event = pcpu->event;
1026 	struct hw_perf_event *hwc;
1027 	struct perf_sample_data data;
1028 	struct perf_raw_record raw;
1029 	struct pt_regs regs;
1030 	struct perf_ibs_data ibs_data;
1031 	int offset, size, check_rip, offset_max, throttle = 0;
1032 	unsigned int msr;
1033 	u64 *buf, *config, period, new_config = 0;
1034 
1035 	if (!test_bit(IBS_STARTED, pcpu->state)) {
1036 fail:
1037 		/*
1038 		 * Catch spurious interrupts after stopping IBS: After
1039 		 * disabling IBS there could be still incoming NMIs
1040 		 * with samples that even have the valid bit cleared.
1041 		 * Mark all this NMIs as handled.
1042 		 */
1043 		if (test_and_clear_bit(IBS_STOPPED, pcpu->state))
1044 			return 1;
1045 
1046 		return 0;
1047 	}
1048 
1049 	if (WARN_ON_ONCE(!event))
1050 		goto fail;
1051 
1052 	hwc = &event->hw;
1053 	msr = hwc->config_base;
1054 	buf = ibs_data.regs;
1055 	rdmsrl(msr, *buf);
1056 	if (!(*buf++ & perf_ibs->valid_mask))
1057 		goto fail;
1058 
1059 	config = &ibs_data.regs[0];
1060 	perf_ibs_event_update(perf_ibs, event, config);
1061 	perf_sample_data_init(&data, 0, hwc->last_period);
1062 	if (!perf_ibs_set_period(perf_ibs, hwc, &period))
1063 		goto out;	/* no sw counter overflow */
1064 
1065 	ibs_data.caps = ibs_caps;
1066 	size = 1;
1067 	offset = 1;
1068 	check_rip = (perf_ibs == &perf_ibs_op && (ibs_caps & IBS_CAPS_RIPINVALIDCHK));
1069 
1070 	offset_max = perf_ibs_get_offset_max(perf_ibs, event->attr.sample_type, check_rip);
1071 
1072 	do {
1073 		rdmsrl(msr + offset, *buf++);
1074 		size++;
1075 		offset = find_next_bit(perf_ibs->offset_mask,
1076 				       perf_ibs->offset_max,
1077 				       offset + 1);
1078 	} while (offset < offset_max);
1079 	/*
1080 	 * Read IbsBrTarget, IbsOpData4, and IbsExtdCtl separately
1081 	 * depending on their availability.
1082 	 * Can't add to offset_max as they are staggered
1083 	 */
1084 	if (event->attr.sample_type & PERF_SAMPLE_RAW) {
1085 		if (perf_ibs == &perf_ibs_op) {
1086 			if (ibs_caps & IBS_CAPS_BRNTRGT) {
1087 				rdmsrl(MSR_AMD64_IBSBRTARGET, *buf++);
1088 				size++;
1089 			}
1090 			if (ibs_caps & IBS_CAPS_OPDATA4) {
1091 				rdmsrl(MSR_AMD64_IBSOPDATA4, *buf++);
1092 				size++;
1093 			}
1094 		}
1095 		if (perf_ibs == &perf_ibs_fetch && (ibs_caps & IBS_CAPS_FETCHCTLEXTD)) {
1096 			rdmsrl(MSR_AMD64_ICIBSEXTDCTL, *buf++);
1097 			size++;
1098 		}
1099 	}
1100 	ibs_data.size = sizeof(u64) * size;
1101 
1102 	regs = *iregs;
1103 	if (check_rip && (ibs_data.regs[2] & IBS_RIP_INVALID)) {
1104 		regs.flags &= ~PERF_EFLAGS_EXACT;
1105 	} else {
1106 		/* Workaround for erratum #1197 */
1107 		if (perf_ibs->fetch_ignore_if_zero_rip && !(ibs_data.regs[1]))
1108 			goto out;
1109 
1110 		set_linear_ip(&regs, ibs_data.regs[1]);
1111 		regs.flags |= PERF_EFLAGS_EXACT;
1112 	}
1113 
1114 	if (event->attr.sample_type & PERF_SAMPLE_RAW) {
1115 		raw = (struct perf_raw_record){
1116 			.frag = {
1117 				.size = sizeof(u32) + ibs_data.size,
1118 				.data = ibs_data.data,
1119 			},
1120 		};
1121 		perf_sample_save_raw_data(&data, &raw);
1122 	}
1123 
1124 	if (perf_ibs == &perf_ibs_op)
1125 		perf_ibs_parse_ld_st_data(event->attr.sample_type, &ibs_data, &data);
1126 
1127 	/*
1128 	 * rip recorded by IbsOpRip will not be consistent with rsp and rbp
1129 	 * recorded as part of interrupt regs. Thus we need to use rip from
1130 	 * interrupt regs while unwinding call stack.
1131 	 */
1132 	if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
1133 		perf_sample_save_callchain(&data, event, iregs);
1134 
1135 	throttle = perf_event_overflow(event, &data, &regs);
1136 out:
1137 	if (throttle) {
1138 		perf_ibs_stop(event, 0);
1139 	} else {
1140 		if (perf_ibs == &perf_ibs_op) {
1141 			if (ibs_caps & IBS_CAPS_OPCNTEXT) {
1142 				new_config = period & IBS_OP_MAX_CNT_EXT_MASK;
1143 				period &= ~IBS_OP_MAX_CNT_EXT_MASK;
1144 			}
1145 			if ((ibs_caps & IBS_CAPS_RDWROPCNT) && (*config & IBS_OP_CNT_CTL))
1146 				new_config |= *config & IBS_OP_CUR_CNT_RAND;
1147 		}
1148 		new_config |= period >> 4;
1149 
1150 		perf_ibs_enable_event(perf_ibs, hwc, new_config);
1151 	}
1152 
1153 	perf_event_update_userpage(event);
1154 
1155 	return 1;
1156 }
1157 
1158 static int
1159 perf_ibs_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1160 {
1161 	u64 stamp = sched_clock();
1162 	int handled = 0;
1163 
1164 	handled += perf_ibs_handle_irq(&perf_ibs_fetch, regs);
1165 	handled += perf_ibs_handle_irq(&perf_ibs_op, regs);
1166 
1167 	if (handled)
1168 		inc_irq_stat(apic_perf_irqs);
1169 
1170 	perf_sample_event_took(sched_clock() - stamp);
1171 
1172 	return handled;
1173 }
1174 NOKPROBE_SYMBOL(perf_ibs_nmi_handler);
1175 
1176 static __init int perf_ibs_pmu_init(struct perf_ibs *perf_ibs, char *name)
1177 {
1178 	struct cpu_perf_ibs __percpu *pcpu;
1179 	int ret;
1180 
1181 	pcpu = alloc_percpu(struct cpu_perf_ibs);
1182 	if (!pcpu)
1183 		return -ENOMEM;
1184 
1185 	perf_ibs->pcpu = pcpu;
1186 
1187 	ret = perf_pmu_register(&perf_ibs->pmu, name, -1);
1188 	if (ret) {
1189 		perf_ibs->pcpu = NULL;
1190 		free_percpu(pcpu);
1191 	}
1192 
1193 	return ret;
1194 }
1195 
1196 static __init int perf_ibs_fetch_init(void)
1197 {
1198 	/*
1199 	 * Some chips fail to reset the fetch count when it is written; instead
1200 	 * they need a 0-1 transition of IbsFetchEn.
1201 	 */
1202 	if (boot_cpu_data.x86 >= 0x16 && boot_cpu_data.x86 <= 0x18)
1203 		perf_ibs_fetch.fetch_count_reset_broken = 1;
1204 
1205 	if (boot_cpu_data.x86 == 0x19 && boot_cpu_data.x86_model < 0x10)
1206 		perf_ibs_fetch.fetch_ignore_if_zero_rip = 1;
1207 
1208 	if (ibs_caps & IBS_CAPS_ZEN4)
1209 		perf_ibs_fetch.config_mask |= IBS_FETCH_L3MISSONLY;
1210 
1211 	perf_ibs_fetch.pmu.attr_groups = fetch_attr_groups;
1212 	perf_ibs_fetch.pmu.attr_update = fetch_attr_update;
1213 
1214 	return perf_ibs_pmu_init(&perf_ibs_fetch, "ibs_fetch");
1215 }
1216 
1217 static __init int perf_ibs_op_init(void)
1218 {
1219 	if (ibs_caps & IBS_CAPS_OPCNT)
1220 		perf_ibs_op.config_mask |= IBS_OP_CNT_CTL;
1221 
1222 	if (ibs_caps & IBS_CAPS_OPCNTEXT) {
1223 		perf_ibs_op.max_period  |= IBS_OP_MAX_CNT_EXT_MASK;
1224 		perf_ibs_op.config_mask	|= IBS_OP_MAX_CNT_EXT_MASK;
1225 		perf_ibs_op.cnt_mask    |= IBS_OP_MAX_CNT_EXT_MASK;
1226 	}
1227 
1228 	if (ibs_caps & IBS_CAPS_ZEN4)
1229 		perf_ibs_op.config_mask |= IBS_OP_L3MISSONLY;
1230 
1231 	perf_ibs_op.pmu.attr_groups = empty_attr_groups;
1232 	perf_ibs_op.pmu.attr_update = op_attr_update;
1233 
1234 	return perf_ibs_pmu_init(&perf_ibs_op, "ibs_op");
1235 }
1236 
1237 static __init int perf_event_ibs_init(void)
1238 {
1239 	int ret;
1240 
1241 	ret = perf_ibs_fetch_init();
1242 	if (ret)
1243 		return ret;
1244 
1245 	ret = perf_ibs_op_init();
1246 	if (ret)
1247 		goto err_op;
1248 
1249 	ret = register_nmi_handler(NMI_LOCAL, perf_ibs_nmi_handler, 0, "perf_ibs");
1250 	if (ret)
1251 		goto err_nmi;
1252 
1253 	pr_info("perf: AMD IBS detected (0x%08x)\n", ibs_caps);
1254 	return 0;
1255 
1256 err_nmi:
1257 	perf_pmu_unregister(&perf_ibs_op.pmu);
1258 	free_percpu(perf_ibs_op.pcpu);
1259 	perf_ibs_op.pcpu = NULL;
1260 err_op:
1261 	perf_pmu_unregister(&perf_ibs_fetch.pmu);
1262 	free_percpu(perf_ibs_fetch.pcpu);
1263 	perf_ibs_fetch.pcpu = NULL;
1264 
1265 	return ret;
1266 }
1267 
1268 #else /* defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_AMD) */
1269 
1270 static __init int perf_event_ibs_init(void)
1271 {
1272 	return 0;
1273 }
1274 
1275 #endif
1276 
1277 /* IBS - apic initialization, for perf and oprofile */
1278 
1279 static __init u32 __get_ibs_caps(void)
1280 {
1281 	u32 caps;
1282 	unsigned int max_level;
1283 
1284 	if (!boot_cpu_has(X86_FEATURE_IBS))
1285 		return 0;
1286 
1287 	/* check IBS cpuid feature flags */
1288 	max_level = cpuid_eax(0x80000000);
1289 	if (max_level < IBS_CPUID_FEATURES)
1290 		return IBS_CAPS_DEFAULT;
1291 
1292 	caps = cpuid_eax(IBS_CPUID_FEATURES);
1293 	if (!(caps & IBS_CAPS_AVAIL))
1294 		/* cpuid flags not valid */
1295 		return IBS_CAPS_DEFAULT;
1296 
1297 	return caps;
1298 }
1299 
1300 u32 get_ibs_caps(void)
1301 {
1302 	return ibs_caps;
1303 }
1304 
1305 EXPORT_SYMBOL(get_ibs_caps);
1306 
1307 static inline int get_eilvt(int offset)
1308 {
1309 	return !setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_NMI, 1);
1310 }
1311 
1312 static inline int put_eilvt(int offset)
1313 {
1314 	return !setup_APIC_eilvt(offset, 0, 0, 1);
1315 }
1316 
1317 /*
1318  * Check and reserve APIC extended interrupt LVT offset for IBS if available.
1319  */
1320 static inline int ibs_eilvt_valid(void)
1321 {
1322 	int offset;
1323 	u64 val;
1324 	int valid = 0;
1325 
1326 	preempt_disable();
1327 
1328 	rdmsrl(MSR_AMD64_IBSCTL, val);
1329 	offset = val & IBSCTL_LVT_OFFSET_MASK;
1330 
1331 	if (!(val & IBSCTL_LVT_OFFSET_VALID)) {
1332 		pr_err(FW_BUG "cpu %d, invalid IBS interrupt offset %d (MSR%08X=0x%016llx)\n",
1333 		       smp_processor_id(), offset, MSR_AMD64_IBSCTL, val);
1334 		goto out;
1335 	}
1336 
1337 	if (!get_eilvt(offset)) {
1338 		pr_err(FW_BUG "cpu %d, IBS interrupt offset %d not available (MSR%08X=0x%016llx)\n",
1339 		       smp_processor_id(), offset, MSR_AMD64_IBSCTL, val);
1340 		goto out;
1341 	}
1342 
1343 	valid = 1;
1344 out:
1345 	preempt_enable();
1346 
1347 	return valid;
1348 }
1349 
1350 static int setup_ibs_ctl(int ibs_eilvt_off)
1351 {
1352 	struct pci_dev *cpu_cfg;
1353 	int nodes;
1354 	u32 value = 0;
1355 
1356 	nodes = 0;
1357 	cpu_cfg = NULL;
1358 	do {
1359 		cpu_cfg = pci_get_device(PCI_VENDOR_ID_AMD,
1360 					 PCI_DEVICE_ID_AMD_10H_NB_MISC,
1361 					 cpu_cfg);
1362 		if (!cpu_cfg)
1363 			break;
1364 		++nodes;
1365 		pci_write_config_dword(cpu_cfg, IBSCTL, ibs_eilvt_off
1366 				       | IBSCTL_LVT_OFFSET_VALID);
1367 		pci_read_config_dword(cpu_cfg, IBSCTL, &value);
1368 		if (value != (ibs_eilvt_off | IBSCTL_LVT_OFFSET_VALID)) {
1369 			pci_dev_put(cpu_cfg);
1370 			pr_debug("Failed to setup IBS LVT offset, IBSCTL = 0x%08x\n",
1371 				 value);
1372 			return -EINVAL;
1373 		}
1374 	} while (1);
1375 
1376 	if (!nodes) {
1377 		pr_debug("No CPU node configured for IBS\n");
1378 		return -ENODEV;
1379 	}
1380 
1381 	return 0;
1382 }
1383 
1384 /*
1385  * This runs only on the current cpu. We try to find an LVT offset and
1386  * setup the local APIC. For this we must disable preemption. On
1387  * success we initialize all nodes with this offset. This updates then
1388  * the offset in the IBS_CTL per-node msr. The per-core APIC setup of
1389  * the IBS interrupt vector is handled by perf_ibs_cpu_notifier that
1390  * is using the new offset.
1391  */
1392 static void force_ibs_eilvt_setup(void)
1393 {
1394 	int offset;
1395 	int ret;
1396 
1397 	preempt_disable();
1398 	/* find the next free available EILVT entry, skip offset 0 */
1399 	for (offset = 1; offset < APIC_EILVT_NR_MAX; offset++) {
1400 		if (get_eilvt(offset))
1401 			break;
1402 	}
1403 	preempt_enable();
1404 
1405 	if (offset == APIC_EILVT_NR_MAX) {
1406 		pr_debug("No EILVT entry available\n");
1407 		return;
1408 	}
1409 
1410 	ret = setup_ibs_ctl(offset);
1411 	if (ret)
1412 		goto out;
1413 
1414 	if (!ibs_eilvt_valid())
1415 		goto out;
1416 
1417 	pr_info("LVT offset %d assigned\n", offset);
1418 
1419 	return;
1420 out:
1421 	preempt_disable();
1422 	put_eilvt(offset);
1423 	preempt_enable();
1424 	return;
1425 }
1426 
1427 static void ibs_eilvt_setup(void)
1428 {
1429 	/*
1430 	 * Force LVT offset assignment for family 10h: The offsets are
1431 	 * not assigned by the BIOS for this family, so the OS is
1432 	 * responsible for doing it. If the OS assignment fails, fall
1433 	 * back to BIOS settings and try to setup this.
1434 	 */
1435 	if (boot_cpu_data.x86 == 0x10)
1436 		force_ibs_eilvt_setup();
1437 }
1438 
1439 static inline int get_ibs_lvt_offset(void)
1440 {
1441 	u64 val;
1442 
1443 	rdmsrl(MSR_AMD64_IBSCTL, val);
1444 	if (!(val & IBSCTL_LVT_OFFSET_VALID))
1445 		return -EINVAL;
1446 
1447 	return val & IBSCTL_LVT_OFFSET_MASK;
1448 }
1449 
1450 static void setup_APIC_ibs(void)
1451 {
1452 	int offset;
1453 
1454 	offset = get_ibs_lvt_offset();
1455 	if (offset < 0)
1456 		goto failed;
1457 
1458 	if (!setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_NMI, 0))
1459 		return;
1460 failed:
1461 	pr_warn("perf: IBS APIC setup failed on cpu #%d\n",
1462 		smp_processor_id());
1463 }
1464 
1465 static void clear_APIC_ibs(void)
1466 {
1467 	int offset;
1468 
1469 	offset = get_ibs_lvt_offset();
1470 	if (offset >= 0)
1471 		setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_FIX, 1);
1472 }
1473 
1474 static int x86_pmu_amd_ibs_starting_cpu(unsigned int cpu)
1475 {
1476 	setup_APIC_ibs();
1477 	return 0;
1478 }
1479 
1480 #ifdef CONFIG_PM
1481 
1482 static int perf_ibs_suspend(void)
1483 {
1484 	clear_APIC_ibs();
1485 	return 0;
1486 }
1487 
1488 static void perf_ibs_resume(void)
1489 {
1490 	ibs_eilvt_setup();
1491 	setup_APIC_ibs();
1492 }
1493 
1494 static struct syscore_ops perf_ibs_syscore_ops = {
1495 	.resume		= perf_ibs_resume,
1496 	.suspend	= perf_ibs_suspend,
1497 };
1498 
1499 static void perf_ibs_pm_init(void)
1500 {
1501 	register_syscore_ops(&perf_ibs_syscore_ops);
1502 }
1503 
1504 #else
1505 
1506 static inline void perf_ibs_pm_init(void) { }
1507 
1508 #endif
1509 
1510 static int x86_pmu_amd_ibs_dying_cpu(unsigned int cpu)
1511 {
1512 	clear_APIC_ibs();
1513 	return 0;
1514 }
1515 
1516 static __init int amd_ibs_init(void)
1517 {
1518 	u32 caps;
1519 
1520 	caps = __get_ibs_caps();
1521 	if (!caps)
1522 		return -ENODEV;	/* ibs not supported by the cpu */
1523 
1524 	ibs_eilvt_setup();
1525 
1526 	if (!ibs_eilvt_valid())
1527 		return -EINVAL;
1528 
1529 	perf_ibs_pm_init();
1530 
1531 	ibs_caps = caps;
1532 	/* make ibs_caps visible to other cpus: */
1533 	smp_mb();
1534 	/*
1535 	 * x86_pmu_amd_ibs_starting_cpu will be called from core on
1536 	 * all online cpus.
1537 	 */
1538 	cpuhp_setup_state(CPUHP_AP_PERF_X86_AMD_IBS_STARTING,
1539 			  "perf/x86/amd/ibs:starting",
1540 			  x86_pmu_amd_ibs_starting_cpu,
1541 			  x86_pmu_amd_ibs_dying_cpu);
1542 
1543 	return perf_event_ibs_init();
1544 }
1545 
1546 /* Since we need the pci subsystem to init ibs we can't do this earlier: */
1547 device_initcall(amd_ibs_init);
1548