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