xref: /kernel/linux/linux-5.10/arch/x86/events/amd/ibs.c (revision 8c2ecf20)
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
21static 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
30#define IBS_FETCH_CONFIG_MASK	(IBS_FETCH_RAND_EN | IBS_FETCH_MAX_CNT)
31#define IBS_OP_CONFIG_MASK	IBS_OP_MAX_CNT
32
33
34/*
35 * IBS states:
36 *
37 * ENABLED; tracks the pmu::add(), pmu::del() state, when set the counter is taken
38 * and any further add()s must fail.
39 *
40 * STARTED/STOPPING/STOPPED; deal with pmu::start(), pmu::stop() state but are
41 * complicated by the fact that the IBS hardware can send late NMIs (ie. after
42 * we've cleared the EN bit).
43 *
44 * In order to consume these late NMIs we have the STOPPED state, any NMI that
45 * happens after we've cleared the EN state will clear this bit and report the
46 * NMI handled (this is fundamentally racy in the face or multiple NMI sources,
47 * someone else can consume our BIT and our NMI will go unhandled).
48 *
49 * And since we cannot set/clear this separate bit together with the EN bit,
50 * there are races; if we cleared STARTED early, an NMI could land in
51 * between clearing STARTED and clearing the EN bit (in fact multiple NMIs
52 * could happen if the period is small enough), and consume our STOPPED bit
53 * and trigger streams of unhandled NMIs.
54 *
55 * If, however, we clear STARTED late, an NMI can hit between clearing the
56 * EN bit and clearing STARTED, still see STARTED set and process the event.
57 * If this event will have the VALID bit clear, we bail properly, but this
58 * is not a given. With VALID set we can end up calling pmu::stop() again
59 * (the throttle logic) and trigger the WARNs in there.
60 *
61 * So what we do is set STOPPING before clearing EN to avoid the pmu::stop()
62 * nesting, and clear STARTED late, so that we have a well defined state over
63 * the clearing of the EN bit.
64 *
65 * XXX: we could probably be using !atomic bitops for all this.
66 */
67
68enum ibs_states {
69	IBS_ENABLED	= 0,
70	IBS_STARTED	= 1,
71	IBS_STOPPING	= 2,
72	IBS_STOPPED	= 3,
73
74	IBS_MAX_STATES,
75};
76
77struct cpu_perf_ibs {
78	struct perf_event	*event;
79	unsigned long		state[BITS_TO_LONGS(IBS_MAX_STATES)];
80};
81
82struct perf_ibs {
83	struct pmu			pmu;
84	unsigned int			msr;
85	u64				config_mask;
86	u64				cnt_mask;
87	u64				enable_mask;
88	u64				valid_mask;
89	u64				max_period;
90	unsigned long			offset_mask[1];
91	int				offset_max;
92	unsigned int			fetch_count_reset_broken : 1;
93	unsigned int			fetch_ignore_if_zero_rip : 1;
94	struct cpu_perf_ibs __percpu	*pcpu;
95
96	struct attribute		**format_attrs;
97	struct attribute_group		format_group;
98	const struct attribute_group	*attr_groups[2];
99
100	u64				(*get_count)(u64 config);
101};
102
103struct perf_ibs_data {
104	u32		size;
105	union {
106		u32	data[0];	/* data buffer starts here */
107		u32	caps;
108	};
109	u64		regs[MSR_AMD64_IBS_REG_COUNT_MAX];
110};
111
112static int
113perf_event_set_period(struct hw_perf_event *hwc, u64 min, u64 max, u64 *hw_period)
114{
115	s64 left = local64_read(&hwc->period_left);
116	s64 period = hwc->sample_period;
117	int overflow = 0;
118
119	/*
120	 * If we are way outside a reasonable range then just skip forward:
121	 */
122	if (unlikely(left <= -period)) {
123		left = period;
124		local64_set(&hwc->period_left, left);
125		hwc->last_period = period;
126		overflow = 1;
127	}
128
129	if (unlikely(left < (s64)min)) {
130		left += period;
131		local64_set(&hwc->period_left, left);
132		hwc->last_period = period;
133		overflow = 1;
134	}
135
136	/*
137	 * If the hw period that triggers the sw overflow is too short
138	 * we might hit the irq handler. This biases the results.
139	 * Thus we shorten the next-to-last period and set the last
140	 * period to the max period.
141	 */
142	if (left > max) {
143		left -= max;
144		if (left > max)
145			left = max;
146		else if (left < min)
147			left = min;
148	}
149
150	*hw_period = (u64)left;
151
152	return overflow;
153}
154
155static  int
156perf_event_try_update(struct perf_event *event, u64 new_raw_count, int width)
157{
158	struct hw_perf_event *hwc = &event->hw;
159	int shift = 64 - width;
160	u64 prev_raw_count;
161	u64 delta;
162
163	/*
164	 * Careful: an NMI might modify the previous event value.
165	 *
166	 * Our tactic to handle this is to first atomically read and
167	 * exchange a new raw count - then add that new-prev delta
168	 * count to the generic event atomically:
169	 */
170	prev_raw_count = local64_read(&hwc->prev_count);
171	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
172					new_raw_count) != prev_raw_count)
173		return 0;
174
175	/*
176	 * Now we have the new raw value and have updated the prev
177	 * timestamp already. We can now calculate the elapsed delta
178	 * (event-)time and add that to the generic event.
179	 *
180	 * Careful, not all hw sign-extends above the physical width
181	 * of the count.
182	 */
183	delta = (new_raw_count << shift) - (prev_raw_count << shift);
184	delta >>= shift;
185
186	local64_add(delta, &event->count);
187	local64_sub(delta, &hwc->period_left);
188
189	return 1;
190}
191
192static struct perf_ibs perf_ibs_fetch;
193static struct perf_ibs perf_ibs_op;
194
195static struct perf_ibs *get_ibs_pmu(int type)
196{
197	if (perf_ibs_fetch.pmu.type == type)
198		return &perf_ibs_fetch;
199	if (perf_ibs_op.pmu.type == type)
200		return &perf_ibs_op;
201	return NULL;
202}
203
204/*
205 * core pmu config -> IBS config
206 *
207 *  perf record -a -e cpu-cycles:p ...    # use ibs op counting cycle count
208 *  perf record -a -e r076:p ...          # same as -e cpu-cycles:p
209 *  perf record -a -e r0C1:p ...          # use ibs op counting micro-ops
210 *
211 * IbsOpCntCtl (bit 19) of IBS Execution Control Register (IbsOpCtl,
212 * MSRC001_1033) is used to select either cycle or micro-ops counting
213 * mode.
214 */
215static int core_pmu_ibs_config(struct perf_event *event, u64 *config)
216{
217	switch (event->attr.type) {
218	case PERF_TYPE_HARDWARE:
219		switch (event->attr.config) {
220		case PERF_COUNT_HW_CPU_CYCLES:
221			*config = 0;
222			return 0;
223		}
224		break;
225	case PERF_TYPE_RAW:
226		switch (event->attr.config) {
227		case 0x0076:
228			*config = 0;
229			return 0;
230		case 0x00C1:
231			*config = IBS_OP_CNT_CTL;
232			return 0;
233		}
234		break;
235	default:
236		return -ENOENT;
237	}
238
239	return -EOPNOTSUPP;
240}
241
242/*
243 * The rip of IBS samples has skid 0. Thus, IBS supports precise
244 * levels 1 and 2 and the PERF_EFLAGS_EXACT is set. In rare cases the
245 * rip is invalid when IBS was not able to record the rip correctly.
246 * We clear PERF_EFLAGS_EXACT and take the rip from pt_regs then.
247 */
248int forward_event_to_ibs(struct perf_event *event)
249{
250	u64 config = 0;
251
252	if (!event->attr.precise_ip || event->attr.precise_ip > 2)
253		return -EOPNOTSUPP;
254
255	if (!core_pmu_ibs_config(event, &config)) {
256		event->attr.type = perf_ibs_op.pmu.type;
257		event->attr.config = config;
258	}
259	return -ENOENT;
260}
261
262static int perf_ibs_init(struct perf_event *event)
263{
264	struct hw_perf_event *hwc = &event->hw;
265	struct perf_ibs *perf_ibs;
266	u64 max_cnt, config;
267
268	perf_ibs = get_ibs_pmu(event->attr.type);
269	if (!perf_ibs)
270		return -ENOENT;
271
272	config = event->attr.config;
273
274	if (event->pmu != &perf_ibs->pmu)
275		return -ENOENT;
276
277	if (config & ~perf_ibs->config_mask)
278		return -EINVAL;
279
280	if (hwc->sample_period) {
281		if (config & perf_ibs->cnt_mask)
282			/* raw max_cnt may not be set */
283			return -EINVAL;
284		if (!event->attr.sample_freq && hwc->sample_period & 0x0f)
285			/*
286			 * lower 4 bits can not be set in ibs max cnt,
287			 * but allowing it in case we adjust the
288			 * sample period to set a frequency.
289			 */
290			return -EINVAL;
291		hwc->sample_period &= ~0x0FULL;
292		if (!hwc->sample_period)
293			hwc->sample_period = 0x10;
294	} else {
295		max_cnt = config & perf_ibs->cnt_mask;
296		config &= ~perf_ibs->cnt_mask;
297		event->attr.sample_period = max_cnt << 4;
298		hwc->sample_period = event->attr.sample_period;
299	}
300
301	if (!hwc->sample_period)
302		return -EINVAL;
303
304	/*
305	 * If we modify hwc->sample_period, we also need to update
306	 * hwc->last_period and hwc->period_left.
307	 */
308	hwc->last_period = hwc->sample_period;
309	local64_set(&hwc->period_left, hwc->sample_period);
310
311	hwc->config_base = perf_ibs->msr;
312	hwc->config = config;
313
314	/*
315	 * rip recorded by IbsOpRip will not be consistent with rsp and rbp
316	 * recorded as part of interrupt regs. Thus we need to use rip from
317	 * interrupt regs while unwinding call stack. Setting _EARLY flag
318	 * makes sure we unwind call-stack before perf sample rip is set to
319	 * IbsOpRip.
320	 */
321	if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
322		event->attr.sample_type |= __PERF_SAMPLE_CALLCHAIN_EARLY;
323
324	return 0;
325}
326
327static int perf_ibs_set_period(struct perf_ibs *perf_ibs,
328			       struct hw_perf_event *hwc, u64 *period)
329{
330	int overflow;
331
332	/* ignore lower 4 bits in min count: */
333	overflow = perf_event_set_period(hwc, 1<<4, perf_ibs->max_period, period);
334	local64_set(&hwc->prev_count, 0);
335
336	return overflow;
337}
338
339static u64 get_ibs_fetch_count(u64 config)
340{
341	return (config & IBS_FETCH_CNT) >> 12;
342}
343
344static u64 get_ibs_op_count(u64 config)
345{
346	u64 count = 0;
347
348	/*
349	 * If the internal 27-bit counter rolled over, the count is MaxCnt
350	 * and the lower 7 bits of CurCnt are randomized.
351	 * Otherwise CurCnt has the full 27-bit current counter value.
352	 */
353	if (config & IBS_OP_VAL) {
354		count = (config & IBS_OP_MAX_CNT) << 4;
355		if (ibs_caps & IBS_CAPS_OPCNTEXT)
356			count += config & IBS_OP_MAX_CNT_EXT_MASK;
357	} else if (ibs_caps & IBS_CAPS_RDWROPCNT) {
358		count = (config & IBS_OP_CUR_CNT) >> 32;
359	}
360
361	return count;
362}
363
364static void
365perf_ibs_event_update(struct perf_ibs *perf_ibs, struct perf_event *event,
366		      u64 *config)
367{
368	u64 count = perf_ibs->get_count(*config);
369
370	/*
371	 * Set width to 64 since we do not overflow on max width but
372	 * instead on max count. In perf_ibs_set_period() we clear
373	 * prev count manually on overflow.
374	 */
375	while (!perf_event_try_update(event, count, 64)) {
376		rdmsrl(event->hw.config_base, *config);
377		count = perf_ibs->get_count(*config);
378	}
379}
380
381static inline void perf_ibs_enable_event(struct perf_ibs *perf_ibs,
382					 struct hw_perf_event *hwc, u64 config)
383{
384	u64 tmp = hwc->config | config;
385
386	if (perf_ibs->fetch_count_reset_broken)
387		wrmsrl(hwc->config_base, tmp & ~perf_ibs->enable_mask);
388
389	wrmsrl(hwc->config_base, tmp | perf_ibs->enable_mask);
390}
391
392/*
393 * Erratum #420 Instruction-Based Sampling Engine May Generate
394 * Interrupt that Cannot Be Cleared:
395 *
396 * Must clear counter mask first, then clear the enable bit. See
397 * Revision Guide for AMD Family 10h Processors, Publication #41322.
398 */
399static inline void perf_ibs_disable_event(struct perf_ibs *perf_ibs,
400					  struct hw_perf_event *hwc, u64 config)
401{
402	config &= ~perf_ibs->cnt_mask;
403	if (boot_cpu_data.x86 == 0x10)
404		wrmsrl(hwc->config_base, config);
405	config &= ~perf_ibs->enable_mask;
406	wrmsrl(hwc->config_base, config);
407}
408
409/*
410 * We cannot restore the ibs pmu state, so we always needs to update
411 * the event while stopping it and then reset the state when starting
412 * again. Thus, ignoring PERF_EF_RELOAD and PERF_EF_UPDATE flags in
413 * perf_ibs_start()/perf_ibs_stop() and instead always do it.
414 */
415static void perf_ibs_start(struct perf_event *event, int flags)
416{
417	struct hw_perf_event *hwc = &event->hw;
418	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
419	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
420	u64 period, config = 0;
421
422	if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
423		return;
424
425	WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
426	hwc->state = 0;
427
428	perf_ibs_set_period(perf_ibs, hwc, &period);
429	if (perf_ibs == &perf_ibs_op && (ibs_caps & IBS_CAPS_OPCNTEXT)) {
430		config |= period & IBS_OP_MAX_CNT_EXT_MASK;
431		period &= ~IBS_OP_MAX_CNT_EXT_MASK;
432	}
433	config |= period >> 4;
434
435	/*
436	 * Set STARTED before enabling the hardware, such that a subsequent NMI
437	 * must observe it.
438	 */
439	set_bit(IBS_STARTED,    pcpu->state);
440	clear_bit(IBS_STOPPING, pcpu->state);
441	perf_ibs_enable_event(perf_ibs, hwc, config);
442
443	perf_event_update_userpage(event);
444}
445
446static void perf_ibs_stop(struct perf_event *event, int flags)
447{
448	struct hw_perf_event *hwc = &event->hw;
449	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
450	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
451	u64 config;
452	int stopping;
453
454	if (test_and_set_bit(IBS_STOPPING, pcpu->state))
455		return;
456
457	stopping = test_bit(IBS_STARTED, pcpu->state);
458
459	if (!stopping && (hwc->state & PERF_HES_UPTODATE))
460		return;
461
462	rdmsrl(hwc->config_base, config);
463
464	if (stopping) {
465		/*
466		 * Set STOPPED before disabling the hardware, such that it
467		 * must be visible to NMIs the moment we clear the EN bit,
468		 * at which point we can generate an !VALID sample which
469		 * we need to consume.
470		 */
471		set_bit(IBS_STOPPED, pcpu->state);
472		perf_ibs_disable_event(perf_ibs, hwc, config);
473		/*
474		 * Clear STARTED after disabling the hardware; if it were
475		 * cleared before an NMI hitting after the clear but before
476		 * clearing the EN bit might think it a spurious NMI and not
477		 * handle it.
478		 *
479		 * Clearing it after, however, creates the problem of the NMI
480		 * handler seeing STARTED but not having a valid sample.
481		 */
482		clear_bit(IBS_STARTED, pcpu->state);
483		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
484		hwc->state |= PERF_HES_STOPPED;
485	}
486
487	if (hwc->state & PERF_HES_UPTODATE)
488		return;
489
490	/*
491	 * Clear valid bit to not count rollovers on update, rollovers
492	 * are only updated in the irq handler.
493	 */
494	config &= ~perf_ibs->valid_mask;
495
496	perf_ibs_event_update(perf_ibs, event, &config);
497	hwc->state |= PERF_HES_UPTODATE;
498}
499
500static int perf_ibs_add(struct perf_event *event, int flags)
501{
502	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
503	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
504
505	if (test_and_set_bit(IBS_ENABLED, pcpu->state))
506		return -ENOSPC;
507
508	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
509
510	pcpu->event = event;
511
512	if (flags & PERF_EF_START)
513		perf_ibs_start(event, PERF_EF_RELOAD);
514
515	return 0;
516}
517
518static void perf_ibs_del(struct perf_event *event, int flags)
519{
520	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
521	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
522
523	if (!test_and_clear_bit(IBS_ENABLED, pcpu->state))
524		return;
525
526	perf_ibs_stop(event, PERF_EF_UPDATE);
527
528	pcpu->event = NULL;
529
530	perf_event_update_userpage(event);
531}
532
533static void perf_ibs_read(struct perf_event *event) { }
534
535PMU_FORMAT_ATTR(rand_en,	"config:57");
536PMU_FORMAT_ATTR(cnt_ctl,	"config:19");
537
538static struct attribute *ibs_fetch_format_attrs[] = {
539	&format_attr_rand_en.attr,
540	NULL,
541};
542
543static struct attribute *ibs_op_format_attrs[] = {
544	NULL,	/* &format_attr_cnt_ctl.attr if IBS_CAPS_OPCNT */
545	NULL,
546};
547
548static struct perf_ibs perf_ibs_fetch = {
549	.pmu = {
550		.task_ctx_nr	= perf_invalid_context,
551
552		.event_init	= perf_ibs_init,
553		.add		= perf_ibs_add,
554		.del		= perf_ibs_del,
555		.start		= perf_ibs_start,
556		.stop		= perf_ibs_stop,
557		.read		= perf_ibs_read,
558		.capabilities	= PERF_PMU_CAP_NO_EXCLUDE,
559	},
560	.msr			= MSR_AMD64_IBSFETCHCTL,
561	.config_mask		= IBS_FETCH_CONFIG_MASK,
562	.cnt_mask		= IBS_FETCH_MAX_CNT,
563	.enable_mask		= IBS_FETCH_ENABLE,
564	.valid_mask		= IBS_FETCH_VAL,
565	.max_period		= IBS_FETCH_MAX_CNT << 4,
566	.offset_mask		= { MSR_AMD64_IBSFETCH_REG_MASK },
567	.offset_max		= MSR_AMD64_IBSFETCH_REG_COUNT,
568	.format_attrs		= ibs_fetch_format_attrs,
569
570	.get_count		= get_ibs_fetch_count,
571};
572
573static struct perf_ibs perf_ibs_op = {
574	.pmu = {
575		.task_ctx_nr	= perf_invalid_context,
576
577		.event_init	= perf_ibs_init,
578		.add		= perf_ibs_add,
579		.del		= perf_ibs_del,
580		.start		= perf_ibs_start,
581		.stop		= perf_ibs_stop,
582		.read		= perf_ibs_read,
583		.capabilities	= PERF_PMU_CAP_NO_EXCLUDE,
584	},
585	.msr			= MSR_AMD64_IBSOPCTL,
586	.config_mask		= IBS_OP_CONFIG_MASK,
587	.cnt_mask		= IBS_OP_MAX_CNT | IBS_OP_CUR_CNT |
588				  IBS_OP_CUR_CNT_RAND,
589	.enable_mask		= IBS_OP_ENABLE,
590	.valid_mask		= IBS_OP_VAL,
591	.max_period		= IBS_OP_MAX_CNT << 4,
592	.offset_mask		= { MSR_AMD64_IBSOP_REG_MASK },
593	.offset_max		= MSR_AMD64_IBSOP_REG_COUNT,
594	.format_attrs		= ibs_op_format_attrs,
595
596	.get_count		= get_ibs_op_count,
597};
598
599static int perf_ibs_handle_irq(struct perf_ibs *perf_ibs, struct pt_regs *iregs)
600{
601	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
602	struct perf_event *event = pcpu->event;
603	struct hw_perf_event *hwc;
604	struct perf_sample_data data;
605	struct perf_raw_record raw;
606	struct pt_regs regs;
607	struct perf_ibs_data ibs_data;
608	int offset, size, check_rip, offset_max, throttle = 0;
609	unsigned int msr;
610	u64 *buf, *config, period, new_config = 0;
611
612	if (!test_bit(IBS_STARTED, pcpu->state)) {
613fail:
614		/*
615		 * Catch spurious interrupts after stopping IBS: After
616		 * disabling IBS there could be still incoming NMIs
617		 * with samples that even have the valid bit cleared.
618		 * Mark all this NMIs as handled.
619		 */
620		if (test_and_clear_bit(IBS_STOPPED, pcpu->state))
621			return 1;
622
623		return 0;
624	}
625
626	if (WARN_ON_ONCE(!event))
627		goto fail;
628
629	hwc = &event->hw;
630	msr = hwc->config_base;
631	buf = ibs_data.regs;
632	rdmsrl(msr, *buf);
633	if (!(*buf++ & perf_ibs->valid_mask))
634		goto fail;
635
636	config = &ibs_data.regs[0];
637	perf_ibs_event_update(perf_ibs, event, config);
638	perf_sample_data_init(&data, 0, hwc->last_period);
639	if (!perf_ibs_set_period(perf_ibs, hwc, &period))
640		goto out;	/* no sw counter overflow */
641
642	ibs_data.caps = ibs_caps;
643	size = 1;
644	offset = 1;
645	check_rip = (perf_ibs == &perf_ibs_op && (ibs_caps & IBS_CAPS_RIPINVALIDCHK));
646	if (event->attr.sample_type & PERF_SAMPLE_RAW)
647		offset_max = perf_ibs->offset_max;
648	else if (check_rip)
649		offset_max = 3;
650	else
651		offset_max = 1;
652	do {
653		rdmsrl(msr + offset, *buf++);
654		size++;
655		offset = find_next_bit(perf_ibs->offset_mask,
656				       perf_ibs->offset_max,
657				       offset + 1);
658	} while (offset < offset_max);
659	/*
660	 * Read IbsBrTarget, IbsOpData4, and IbsExtdCtl separately
661	 * depending on their availability.
662	 * Can't add to offset_max as they are staggered
663	 */
664	if (event->attr.sample_type & PERF_SAMPLE_RAW) {
665		if (perf_ibs == &perf_ibs_op) {
666			if (ibs_caps & IBS_CAPS_BRNTRGT) {
667				rdmsrl(MSR_AMD64_IBSBRTARGET, *buf++);
668				size++;
669			}
670			if (ibs_caps & IBS_CAPS_OPDATA4) {
671				rdmsrl(MSR_AMD64_IBSOPDATA4, *buf++);
672				size++;
673			}
674		}
675		if (perf_ibs == &perf_ibs_fetch && (ibs_caps & IBS_CAPS_FETCHCTLEXTD)) {
676			rdmsrl(MSR_AMD64_ICIBSEXTDCTL, *buf++);
677			size++;
678		}
679	}
680	ibs_data.size = sizeof(u64) * size;
681
682	regs = *iregs;
683	if (check_rip && (ibs_data.regs[2] & IBS_RIP_INVALID)) {
684		regs.flags &= ~PERF_EFLAGS_EXACT;
685	} else {
686		/* Workaround for erratum #1197 */
687		if (perf_ibs->fetch_ignore_if_zero_rip && !(ibs_data.regs[1]))
688			goto out;
689
690		set_linear_ip(&regs, ibs_data.regs[1]);
691		regs.flags |= PERF_EFLAGS_EXACT;
692	}
693
694	if (event->attr.sample_type & PERF_SAMPLE_RAW) {
695		raw = (struct perf_raw_record){
696			.frag = {
697				.size = sizeof(u32) + ibs_data.size,
698				.data = ibs_data.data,
699			},
700		};
701		data.raw = &raw;
702	}
703
704	/*
705	 * rip recorded by IbsOpRip will not be consistent with rsp and rbp
706	 * recorded as part of interrupt regs. Thus we need to use rip from
707	 * interrupt regs while unwinding call stack.
708	 */
709	if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
710		data.callchain = perf_callchain(event, iregs);
711
712	throttle = perf_event_overflow(event, &data, &regs);
713out:
714	if (throttle) {
715		perf_ibs_stop(event, 0);
716	} else {
717		if (perf_ibs == &perf_ibs_op) {
718			if (ibs_caps & IBS_CAPS_OPCNTEXT) {
719				new_config = period & IBS_OP_MAX_CNT_EXT_MASK;
720				period &= ~IBS_OP_MAX_CNT_EXT_MASK;
721			}
722			if ((ibs_caps & IBS_CAPS_RDWROPCNT) && (*config & IBS_OP_CNT_CTL))
723				new_config |= *config & IBS_OP_CUR_CNT_RAND;
724		}
725		new_config |= period >> 4;
726
727		perf_ibs_enable_event(perf_ibs, hwc, new_config);
728	}
729
730	perf_event_update_userpage(event);
731
732	return 1;
733}
734
735static int
736perf_ibs_nmi_handler(unsigned int cmd, struct pt_regs *regs)
737{
738	u64 stamp = sched_clock();
739	int handled = 0;
740
741	handled += perf_ibs_handle_irq(&perf_ibs_fetch, regs);
742	handled += perf_ibs_handle_irq(&perf_ibs_op, regs);
743
744	if (handled)
745		inc_irq_stat(apic_perf_irqs);
746
747	perf_sample_event_took(sched_clock() - stamp);
748
749	return handled;
750}
751NOKPROBE_SYMBOL(perf_ibs_nmi_handler);
752
753static __init int perf_ibs_pmu_init(struct perf_ibs *perf_ibs, char *name)
754{
755	struct cpu_perf_ibs __percpu *pcpu;
756	int ret;
757
758	pcpu = alloc_percpu(struct cpu_perf_ibs);
759	if (!pcpu)
760		return -ENOMEM;
761
762	perf_ibs->pcpu = pcpu;
763
764	/* register attributes */
765	if (perf_ibs->format_attrs[0]) {
766		memset(&perf_ibs->format_group, 0, sizeof(perf_ibs->format_group));
767		perf_ibs->format_group.name	= "format";
768		perf_ibs->format_group.attrs	= perf_ibs->format_attrs;
769
770		memset(&perf_ibs->attr_groups, 0, sizeof(perf_ibs->attr_groups));
771		perf_ibs->attr_groups[0]	= &perf_ibs->format_group;
772		perf_ibs->pmu.attr_groups	= perf_ibs->attr_groups;
773	}
774
775	ret = perf_pmu_register(&perf_ibs->pmu, name, -1);
776	if (ret) {
777		perf_ibs->pcpu = NULL;
778		free_percpu(pcpu);
779	}
780
781	return ret;
782}
783
784static __init int perf_event_ibs_init(void)
785{
786	struct attribute **attr = ibs_op_format_attrs;
787	int ret;
788
789	/*
790	 * Some chips fail to reset the fetch count when it is written; instead
791	 * they need a 0-1 transition of IbsFetchEn.
792	 */
793	if (boot_cpu_data.x86 >= 0x16 && boot_cpu_data.x86 <= 0x18)
794		perf_ibs_fetch.fetch_count_reset_broken = 1;
795
796	if (boot_cpu_data.x86 == 0x19 && boot_cpu_data.x86_model < 0x10)
797		perf_ibs_fetch.fetch_ignore_if_zero_rip = 1;
798
799	ret = perf_ibs_pmu_init(&perf_ibs_fetch, "ibs_fetch");
800	if (ret)
801		return ret;
802
803	if (ibs_caps & IBS_CAPS_OPCNT) {
804		perf_ibs_op.config_mask |= IBS_OP_CNT_CTL;
805		*attr++ = &format_attr_cnt_ctl.attr;
806	}
807
808	if (ibs_caps & IBS_CAPS_OPCNTEXT) {
809		perf_ibs_op.max_period  |= IBS_OP_MAX_CNT_EXT_MASK;
810		perf_ibs_op.config_mask	|= IBS_OP_MAX_CNT_EXT_MASK;
811		perf_ibs_op.cnt_mask    |= IBS_OP_MAX_CNT_EXT_MASK;
812	}
813
814	ret = perf_ibs_pmu_init(&perf_ibs_op, "ibs_op");
815	if (ret)
816		goto err_op;
817
818	ret = register_nmi_handler(NMI_LOCAL, perf_ibs_nmi_handler, 0, "perf_ibs");
819	if (ret)
820		goto err_nmi;
821
822	pr_info("perf: AMD IBS detected (0x%08x)\n", ibs_caps);
823	return 0;
824
825err_nmi:
826	perf_pmu_unregister(&perf_ibs_op.pmu);
827	free_percpu(perf_ibs_op.pcpu);
828	perf_ibs_op.pcpu = NULL;
829err_op:
830	perf_pmu_unregister(&perf_ibs_fetch.pmu);
831	free_percpu(perf_ibs_fetch.pcpu);
832	perf_ibs_fetch.pcpu = NULL;
833
834	return ret;
835}
836
837#else /* defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_AMD) */
838
839static __init int perf_event_ibs_init(void)
840{
841	return 0;
842}
843
844#endif
845
846/* IBS - apic initialization, for perf and oprofile */
847
848static __init u32 __get_ibs_caps(void)
849{
850	u32 caps;
851	unsigned int max_level;
852
853	if (!boot_cpu_has(X86_FEATURE_IBS))
854		return 0;
855
856	/* check IBS cpuid feature flags */
857	max_level = cpuid_eax(0x80000000);
858	if (max_level < IBS_CPUID_FEATURES)
859		return IBS_CAPS_DEFAULT;
860
861	caps = cpuid_eax(IBS_CPUID_FEATURES);
862	if (!(caps & IBS_CAPS_AVAIL))
863		/* cpuid flags not valid */
864		return IBS_CAPS_DEFAULT;
865
866	return caps;
867}
868
869u32 get_ibs_caps(void)
870{
871	return ibs_caps;
872}
873
874EXPORT_SYMBOL(get_ibs_caps);
875
876static inline int get_eilvt(int offset)
877{
878	return !setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_NMI, 1);
879}
880
881static inline int put_eilvt(int offset)
882{
883	return !setup_APIC_eilvt(offset, 0, 0, 1);
884}
885
886/*
887 * Check and reserve APIC extended interrupt LVT offset for IBS if available.
888 */
889static inline int ibs_eilvt_valid(void)
890{
891	int offset;
892	u64 val;
893	int valid = 0;
894
895	preempt_disable();
896
897	rdmsrl(MSR_AMD64_IBSCTL, val);
898	offset = val & IBSCTL_LVT_OFFSET_MASK;
899
900	if (!(val & IBSCTL_LVT_OFFSET_VALID)) {
901		pr_err(FW_BUG "cpu %d, invalid IBS interrupt offset %d (MSR%08X=0x%016llx)\n",
902		       smp_processor_id(), offset, MSR_AMD64_IBSCTL, val);
903		goto out;
904	}
905
906	if (!get_eilvt(offset)) {
907		pr_err(FW_BUG "cpu %d, IBS interrupt offset %d not available (MSR%08X=0x%016llx)\n",
908		       smp_processor_id(), offset, MSR_AMD64_IBSCTL, val);
909		goto out;
910	}
911
912	valid = 1;
913out:
914	preempt_enable();
915
916	return valid;
917}
918
919static int setup_ibs_ctl(int ibs_eilvt_off)
920{
921	struct pci_dev *cpu_cfg;
922	int nodes;
923	u32 value = 0;
924
925	nodes = 0;
926	cpu_cfg = NULL;
927	do {
928		cpu_cfg = pci_get_device(PCI_VENDOR_ID_AMD,
929					 PCI_DEVICE_ID_AMD_10H_NB_MISC,
930					 cpu_cfg);
931		if (!cpu_cfg)
932			break;
933		++nodes;
934		pci_write_config_dword(cpu_cfg, IBSCTL, ibs_eilvt_off
935				       | IBSCTL_LVT_OFFSET_VALID);
936		pci_read_config_dword(cpu_cfg, IBSCTL, &value);
937		if (value != (ibs_eilvt_off | IBSCTL_LVT_OFFSET_VALID)) {
938			pci_dev_put(cpu_cfg);
939			pr_debug("Failed to setup IBS LVT offset, IBSCTL = 0x%08x\n",
940				 value);
941			return -EINVAL;
942		}
943	} while (1);
944
945	if (!nodes) {
946		pr_debug("No CPU node configured for IBS\n");
947		return -ENODEV;
948	}
949
950	return 0;
951}
952
953/*
954 * This runs only on the current cpu. We try to find an LVT offset and
955 * setup the local APIC. For this we must disable preemption. On
956 * success we initialize all nodes with this offset. This updates then
957 * the offset in the IBS_CTL per-node msr. The per-core APIC setup of
958 * the IBS interrupt vector is handled by perf_ibs_cpu_notifier that
959 * is using the new offset.
960 */
961static void force_ibs_eilvt_setup(void)
962{
963	int offset;
964	int ret;
965
966	preempt_disable();
967	/* find the next free available EILVT entry, skip offset 0 */
968	for (offset = 1; offset < APIC_EILVT_NR_MAX; offset++) {
969		if (get_eilvt(offset))
970			break;
971	}
972	preempt_enable();
973
974	if (offset == APIC_EILVT_NR_MAX) {
975		pr_debug("No EILVT entry available\n");
976		return;
977	}
978
979	ret = setup_ibs_ctl(offset);
980	if (ret)
981		goto out;
982
983	if (!ibs_eilvt_valid())
984		goto out;
985
986	pr_info("LVT offset %d assigned\n", offset);
987
988	return;
989out:
990	preempt_disable();
991	put_eilvt(offset);
992	preempt_enable();
993	return;
994}
995
996static void ibs_eilvt_setup(void)
997{
998	/*
999	 * Force LVT offset assignment for family 10h: The offsets are
1000	 * not assigned by the BIOS for this family, so the OS is
1001	 * responsible for doing it. If the OS assignment fails, fall
1002	 * back to BIOS settings and try to setup this.
1003	 */
1004	if (boot_cpu_data.x86 == 0x10)
1005		force_ibs_eilvt_setup();
1006}
1007
1008static inline int get_ibs_lvt_offset(void)
1009{
1010	u64 val;
1011
1012	rdmsrl(MSR_AMD64_IBSCTL, val);
1013	if (!(val & IBSCTL_LVT_OFFSET_VALID))
1014		return -EINVAL;
1015
1016	return val & IBSCTL_LVT_OFFSET_MASK;
1017}
1018
1019static void setup_APIC_ibs(void)
1020{
1021	int offset;
1022
1023	offset = get_ibs_lvt_offset();
1024	if (offset < 0)
1025		goto failed;
1026
1027	if (!setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_NMI, 0))
1028		return;
1029failed:
1030	pr_warn("perf: IBS APIC setup failed on cpu #%d\n",
1031		smp_processor_id());
1032}
1033
1034static void clear_APIC_ibs(void)
1035{
1036	int offset;
1037
1038	offset = get_ibs_lvt_offset();
1039	if (offset >= 0)
1040		setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_FIX, 1);
1041}
1042
1043static int x86_pmu_amd_ibs_starting_cpu(unsigned int cpu)
1044{
1045	setup_APIC_ibs();
1046	return 0;
1047}
1048
1049#ifdef CONFIG_PM
1050
1051static int perf_ibs_suspend(void)
1052{
1053	clear_APIC_ibs();
1054	return 0;
1055}
1056
1057static void perf_ibs_resume(void)
1058{
1059	ibs_eilvt_setup();
1060	setup_APIC_ibs();
1061}
1062
1063static struct syscore_ops perf_ibs_syscore_ops = {
1064	.resume		= perf_ibs_resume,
1065	.suspend	= perf_ibs_suspend,
1066};
1067
1068static void perf_ibs_pm_init(void)
1069{
1070	register_syscore_ops(&perf_ibs_syscore_ops);
1071}
1072
1073#else
1074
1075static inline void perf_ibs_pm_init(void) { }
1076
1077#endif
1078
1079static int x86_pmu_amd_ibs_dying_cpu(unsigned int cpu)
1080{
1081	clear_APIC_ibs();
1082	return 0;
1083}
1084
1085static __init int amd_ibs_init(void)
1086{
1087	u32 caps;
1088
1089	caps = __get_ibs_caps();
1090	if (!caps)
1091		return -ENODEV;	/* ibs not supported by the cpu */
1092
1093	ibs_eilvt_setup();
1094
1095	if (!ibs_eilvt_valid())
1096		return -EINVAL;
1097
1098	perf_ibs_pm_init();
1099
1100	ibs_caps = caps;
1101	/* make ibs_caps visible to other cpus: */
1102	smp_mb();
1103	/*
1104	 * x86_pmu_amd_ibs_starting_cpu will be called from core on
1105	 * all online cpus.
1106	 */
1107	cpuhp_setup_state(CPUHP_AP_PERF_X86_AMD_IBS_STARTING,
1108			  "perf/x86/amd/ibs:starting",
1109			  x86_pmu_amd_ibs_starting_cpu,
1110			  x86_pmu_amd_ibs_dying_cpu);
1111
1112	return perf_event_ibs_init();
1113}
1114
1115/* Since we need the pci subsystem to init ibs we can't do this earlier: */
1116device_initcall(amd_ibs_init);
1117