Linux Audio

Check our new training course

Loading...
Note: File does not exist in v4.17.
  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/module.h>
 11#include <linux/pci.h>
 12#include <linux/ptrace.h>
 13
 14#include <asm/apic.h>
 15
 16static u32 ibs_caps;
 17
 18#if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_AMD)
 19
 20#include <linux/kprobes.h>
 21#include <linux/hardirq.h>
 22
 23#include <asm/nmi.h>
 24
 25#define IBS_FETCH_CONFIG_MASK	(IBS_FETCH_RAND_EN | IBS_FETCH_MAX_CNT)
 26#define IBS_OP_CONFIG_MASK	IBS_OP_MAX_CNT
 27
 28enum ibs_states {
 29	IBS_ENABLED	= 0,
 30	IBS_STARTED	= 1,
 31	IBS_STOPPING	= 2,
 32
 33	IBS_MAX_STATES,
 34};
 35
 36struct cpu_perf_ibs {
 37	struct perf_event	*event;
 38	unsigned long		state[BITS_TO_LONGS(IBS_MAX_STATES)];
 39};
 40
 41struct perf_ibs {
 42	struct pmu	pmu;
 43	unsigned int	msr;
 44	u64		config_mask;
 45	u64		cnt_mask;
 46	u64		enable_mask;
 47	u64		valid_mask;
 48	u64		max_period;
 49	unsigned long	offset_mask[1];
 50	int		offset_max;
 51	struct cpu_perf_ibs __percpu *pcpu;
 52	u64		(*get_count)(u64 config);
 53};
 54
 55struct perf_ibs_data {
 56	u32		size;
 57	union {
 58		u32	data[0];	/* data buffer starts here */
 59		u32	caps;
 60	};
 61	u64		regs[MSR_AMD64_IBS_REG_COUNT_MAX];
 62};
 63
 64static int
 65perf_event_set_period(struct hw_perf_event *hwc, u64 min, u64 max, u64 *hw_period)
 66{
 67	s64 left = local64_read(&hwc->period_left);
 68	s64 period = hwc->sample_period;
 69	int overflow = 0;
 70
 71	/*
 72	 * If we are way outside a reasonable range then just skip forward:
 73	 */
 74	if (unlikely(left <= -period)) {
 75		left = period;
 76		local64_set(&hwc->period_left, left);
 77		hwc->last_period = period;
 78		overflow = 1;
 79	}
 80
 81	if (unlikely(left < (s64)min)) {
 82		left += period;
 83		local64_set(&hwc->period_left, left);
 84		hwc->last_period = period;
 85		overflow = 1;
 86	}
 87
 88	/*
 89	 * If the hw period that triggers the sw overflow is too short
 90	 * we might hit the irq handler. This biases the results.
 91	 * Thus we shorten the next-to-last period and set the last
 92	 * period to the max period.
 93	 */
 94	if (left > max) {
 95		left -= max;
 96		if (left > max)
 97			left = max;
 98		else if (left < min)
 99			left = min;
100	}
101
102	*hw_period = (u64)left;
103
104	return overflow;
105}
106
107static  int
108perf_event_try_update(struct perf_event *event, u64 new_raw_count, int width)
109{
110	struct hw_perf_event *hwc = &event->hw;
111	int shift = 64 - width;
112	u64 prev_raw_count;
113	u64 delta;
114
115	/*
116	 * Careful: an NMI might modify the previous event value.
117	 *
118	 * Our tactic to handle this is to first atomically read and
119	 * exchange a new raw count - then add that new-prev delta
120	 * count to the generic event atomically:
121	 */
122	prev_raw_count = local64_read(&hwc->prev_count);
123	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
124					new_raw_count) != prev_raw_count)
125		return 0;
126
127	/*
128	 * Now we have the new raw value and have updated the prev
129	 * timestamp already. We can now calculate the elapsed delta
130	 * (event-)time and add that to the generic event.
131	 *
132	 * Careful, not all hw sign-extends above the physical width
133	 * of the count.
134	 */
135	delta = (new_raw_count << shift) - (prev_raw_count << shift);
136	delta >>= shift;
137
138	local64_add(delta, &event->count);
139	local64_sub(delta, &hwc->period_left);
140
141	return 1;
142}
143
144static struct perf_ibs perf_ibs_fetch;
145static struct perf_ibs perf_ibs_op;
146
147static struct perf_ibs *get_ibs_pmu(int type)
148{
149	if (perf_ibs_fetch.pmu.type == type)
150		return &perf_ibs_fetch;
151	if (perf_ibs_op.pmu.type == type)
152		return &perf_ibs_op;
153	return NULL;
154}
155
156/*
157 * Use IBS for precise event sampling:
158 *
159 *  perf record -a -e cpu-cycles:p ...    # use ibs op counting cycle count
160 *  perf record -a -e r076:p ...          # same as -e cpu-cycles:p
161 *  perf record -a -e r0C1:p ...          # use ibs op counting micro-ops
162 *
163 * IbsOpCntCtl (bit 19) of IBS Execution Control Register (IbsOpCtl,
164 * MSRC001_1033) is used to select either cycle or micro-ops counting
165 * mode.
166 *
167 * The rip of IBS samples has skid 0. Thus, IBS supports precise
168 * levels 1 and 2 and the PERF_EFLAGS_EXACT is set. In rare cases the
169 * rip is invalid when IBS was not able to record the rip correctly.
170 * We clear PERF_EFLAGS_EXACT and take the rip from pt_regs then.
171 *
172 */
173static int perf_ibs_precise_event(struct perf_event *event, u64 *config)
174{
175	switch (event->attr.precise_ip) {
176	case 0:
177		return -ENOENT;
178	case 1:
179	case 2:
180		break;
181	default:
182		return -EOPNOTSUPP;
183	}
184
185	switch (event->attr.type) {
186	case PERF_TYPE_HARDWARE:
187		switch (event->attr.config) {
188		case PERF_COUNT_HW_CPU_CYCLES:
189			*config = 0;
190			return 0;
191		}
192		break;
193	case PERF_TYPE_RAW:
194		switch (event->attr.config) {
195		case 0x0076:
196			*config = 0;
197			return 0;
198		case 0x00C1:
199			*config = IBS_OP_CNT_CTL;
200			return 0;
201		}
202		break;
203	default:
204		return -ENOENT;
205	}
206
207	return -EOPNOTSUPP;
208}
209
210static const struct perf_event_attr ibs_notsupp = {
211	.exclude_user	= 1,
212	.exclude_kernel	= 1,
213	.exclude_hv	= 1,
214	.exclude_idle	= 1,
215	.exclude_host	= 1,
216	.exclude_guest	= 1,
217};
218
219static int perf_ibs_init(struct perf_event *event)
220{
221	struct hw_perf_event *hwc = &event->hw;
222	struct perf_ibs *perf_ibs;
223	u64 max_cnt, config;
224	int ret;
225
226	perf_ibs = get_ibs_pmu(event->attr.type);
227	if (perf_ibs) {
228		config = event->attr.config;
229	} else {
230		perf_ibs = &perf_ibs_op;
231		ret = perf_ibs_precise_event(event, &config);
232		if (ret)
233			return ret;
234	}
235
236	if (event->pmu != &perf_ibs->pmu)
237		return -ENOENT;
238
239	if (perf_flags(&event->attr) & perf_flags(&ibs_notsupp))
240		return -EINVAL;
241
242	if (config & ~perf_ibs->config_mask)
243		return -EINVAL;
244
245	if (hwc->sample_period) {
246		if (config & perf_ibs->cnt_mask)
247			/* raw max_cnt may not be set */
248			return -EINVAL;
249		if (!event->attr.sample_freq && hwc->sample_period & 0x0f)
250			/*
251			 * lower 4 bits can not be set in ibs max cnt,
252			 * but allowing it in case we adjust the
253			 * sample period to set a frequency.
254			 */
255			return -EINVAL;
256		hwc->sample_period &= ~0x0FULL;
257		if (!hwc->sample_period)
258			hwc->sample_period = 0x10;
259	} else {
260		max_cnt = config & perf_ibs->cnt_mask;
261		config &= ~perf_ibs->cnt_mask;
262		event->attr.sample_period = max_cnt << 4;
263		hwc->sample_period = event->attr.sample_period;
264	}
265
266	if (!hwc->sample_period)
267		return -EINVAL;
268
269	/*
270	 * If we modify hwc->sample_period, we also need to update
271	 * hwc->last_period and hwc->period_left.
272	 */
273	hwc->last_period = hwc->sample_period;
274	local64_set(&hwc->period_left, hwc->sample_period);
275
276	hwc->config_base = perf_ibs->msr;
277	hwc->config = config;
278
279	return 0;
280}
281
282static int perf_ibs_set_period(struct perf_ibs *perf_ibs,
283			       struct hw_perf_event *hwc, u64 *period)
284{
285	int overflow;
286
287	/* ignore lower 4 bits in min count: */
288	overflow = perf_event_set_period(hwc, 1<<4, perf_ibs->max_period, period);
289	local64_set(&hwc->prev_count, 0);
290
291	return overflow;
292}
293
294static u64 get_ibs_fetch_count(u64 config)
295{
296	return (config & IBS_FETCH_CNT) >> 12;
297}
298
299static u64 get_ibs_op_count(u64 config)
300{
301	u64 count = 0;
302
303	if (config & IBS_OP_VAL)
304		count += (config & IBS_OP_MAX_CNT) << 4; /* cnt rolled over */
305
306	if (ibs_caps & IBS_CAPS_RDWROPCNT)
307		count += (config & IBS_OP_CUR_CNT) >> 32;
308
309	return count;
310}
311
312static void
313perf_ibs_event_update(struct perf_ibs *perf_ibs, struct perf_event *event,
314		      u64 *config)
315{
316	u64 count = perf_ibs->get_count(*config);
317
318	/*
319	 * Set width to 64 since we do not overflow on max width but
320	 * instead on max count. In perf_ibs_set_period() we clear
321	 * prev count manually on overflow.
322	 */
323	while (!perf_event_try_update(event, count, 64)) {
324		rdmsrl(event->hw.config_base, *config);
325		count = perf_ibs->get_count(*config);
326	}
327}
328
329static inline void perf_ibs_enable_event(struct perf_ibs *perf_ibs,
330					 struct hw_perf_event *hwc, u64 config)
331{
332	wrmsrl(hwc->config_base, hwc->config | config | perf_ibs->enable_mask);
333}
334
335/*
336 * Erratum #420 Instruction-Based Sampling Engine May Generate
337 * Interrupt that Cannot Be Cleared:
338 *
339 * Must clear counter mask first, then clear the enable bit. See
340 * Revision Guide for AMD Family 10h Processors, Publication #41322.
341 */
342static inline void perf_ibs_disable_event(struct perf_ibs *perf_ibs,
343					  struct hw_perf_event *hwc, u64 config)
344{
345	config &= ~perf_ibs->cnt_mask;
346	wrmsrl(hwc->config_base, config);
347	config &= ~perf_ibs->enable_mask;
348	wrmsrl(hwc->config_base, config);
349}
350
351/*
352 * We cannot restore the ibs pmu state, so we always needs to update
353 * the event while stopping it and then reset the state when starting
354 * again. Thus, ignoring PERF_EF_RELOAD and PERF_EF_UPDATE flags in
355 * perf_ibs_start()/perf_ibs_stop() and instead always do it.
356 */
357static void perf_ibs_start(struct perf_event *event, int flags)
358{
359	struct hw_perf_event *hwc = &event->hw;
360	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
361	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
362	u64 period;
363
364	if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
365		return;
366
367	WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
368	hwc->state = 0;
369
370	perf_ibs_set_period(perf_ibs, hwc, &period);
371	set_bit(IBS_STARTED, pcpu->state);
372	perf_ibs_enable_event(perf_ibs, hwc, period >> 4);
373
374	perf_event_update_userpage(event);
375}
376
377static void perf_ibs_stop(struct perf_event *event, int flags)
378{
379	struct hw_perf_event *hwc = &event->hw;
380	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
381	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
382	u64 config;
383	int stopping;
384
385	stopping = test_and_clear_bit(IBS_STARTED, pcpu->state);
386
387	if (!stopping && (hwc->state & PERF_HES_UPTODATE))
388		return;
389
390	rdmsrl(hwc->config_base, config);
391
392	if (stopping) {
393		set_bit(IBS_STOPPING, pcpu->state);
394		perf_ibs_disable_event(perf_ibs, hwc, config);
395		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
396		hwc->state |= PERF_HES_STOPPED;
397	}
398
399	if (hwc->state & PERF_HES_UPTODATE)
400		return;
401
402	/*
403	 * Clear valid bit to not count rollovers on update, rollovers
404	 * are only updated in the irq handler.
405	 */
406	config &= ~perf_ibs->valid_mask;
407
408	perf_ibs_event_update(perf_ibs, event, &config);
409	hwc->state |= PERF_HES_UPTODATE;
410}
411
412static int perf_ibs_add(struct perf_event *event, int flags)
413{
414	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
415	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
416
417	if (test_and_set_bit(IBS_ENABLED, pcpu->state))
418		return -ENOSPC;
419
420	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
421
422	pcpu->event = event;
423
424	if (flags & PERF_EF_START)
425		perf_ibs_start(event, PERF_EF_RELOAD);
426
427	return 0;
428}
429
430static void perf_ibs_del(struct perf_event *event, int flags)
431{
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
435	if (!test_and_clear_bit(IBS_ENABLED, pcpu->state))
436		return;
437
438	perf_ibs_stop(event, PERF_EF_UPDATE);
439
440	pcpu->event = NULL;
441
442	perf_event_update_userpage(event);
443}
444
445static void perf_ibs_read(struct perf_event *event) { }
446
447static struct perf_ibs perf_ibs_fetch = {
448	.pmu = {
449		.task_ctx_nr	= perf_invalid_context,
450
451		.event_init	= perf_ibs_init,
452		.add		= perf_ibs_add,
453		.del		= perf_ibs_del,
454		.start		= perf_ibs_start,
455		.stop		= perf_ibs_stop,
456		.read		= perf_ibs_read,
457	},
458	.msr			= MSR_AMD64_IBSFETCHCTL,
459	.config_mask		= IBS_FETCH_CONFIG_MASK,
460	.cnt_mask		= IBS_FETCH_MAX_CNT,
461	.enable_mask		= IBS_FETCH_ENABLE,
462	.valid_mask		= IBS_FETCH_VAL,
463	.max_period		= IBS_FETCH_MAX_CNT << 4,
464	.offset_mask		= { MSR_AMD64_IBSFETCH_REG_MASK },
465	.offset_max		= MSR_AMD64_IBSFETCH_REG_COUNT,
466
467	.get_count		= get_ibs_fetch_count,
468};
469
470static struct perf_ibs perf_ibs_op = {
471	.pmu = {
472		.task_ctx_nr	= perf_invalid_context,
473
474		.event_init	= perf_ibs_init,
475		.add		= perf_ibs_add,
476		.del		= perf_ibs_del,
477		.start		= perf_ibs_start,
478		.stop		= perf_ibs_stop,
479		.read		= perf_ibs_read,
480	},
481	.msr			= MSR_AMD64_IBSOPCTL,
482	.config_mask		= IBS_OP_CONFIG_MASK,
483	.cnt_mask		= IBS_OP_MAX_CNT,
484	.enable_mask		= IBS_OP_ENABLE,
485	.valid_mask		= IBS_OP_VAL,
486	.max_period		= IBS_OP_MAX_CNT << 4,
487	.offset_mask		= { MSR_AMD64_IBSOP_REG_MASK },
488	.offset_max		= MSR_AMD64_IBSOP_REG_COUNT,
489
490	.get_count		= get_ibs_op_count,
491};
492
493static int perf_ibs_handle_irq(struct perf_ibs *perf_ibs, struct pt_regs *iregs)
494{
495	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
496	struct perf_event *event = pcpu->event;
497	struct hw_perf_event *hwc = &event->hw;
498	struct perf_sample_data data;
499	struct perf_raw_record raw;
500	struct pt_regs regs;
501	struct perf_ibs_data ibs_data;
502	int offset, size, check_rip, offset_max, throttle = 0;
503	unsigned int msr;
504	u64 *buf, *config, period;
505
506	if (!test_bit(IBS_STARTED, pcpu->state)) {
507		/*
508		 * Catch spurious interrupts after stopping IBS: After
509		 * disabling IBS there could be still incomming NMIs
510		 * with samples that even have the valid bit cleared.
511		 * Mark all this NMIs as handled.
512		 */
513		return test_and_clear_bit(IBS_STOPPING, pcpu->state) ? 1 : 0;
514	}
515
516	msr = hwc->config_base;
517	buf = ibs_data.regs;
518	rdmsrl(msr, *buf);
519	if (!(*buf++ & perf_ibs->valid_mask))
520		return 0;
521
522	config = &ibs_data.regs[0];
523	perf_ibs_event_update(perf_ibs, event, config);
524	perf_sample_data_init(&data, 0, hwc->last_period);
525	if (!perf_ibs_set_period(perf_ibs, hwc, &period))
526		goto out;	/* no sw counter overflow */
527
528	ibs_data.caps = ibs_caps;
529	size = 1;
530	offset = 1;
531	check_rip = (perf_ibs == &perf_ibs_op && (ibs_caps & IBS_CAPS_RIPINVALIDCHK));
532	if (event->attr.sample_type & PERF_SAMPLE_RAW)
533		offset_max = perf_ibs->offset_max;
534	else if (check_rip)
535		offset_max = 2;
536	else
537		offset_max = 1;
538	do {
539		rdmsrl(msr + offset, *buf++);
540		size++;
541		offset = find_next_bit(perf_ibs->offset_mask,
542				       perf_ibs->offset_max,
543				       offset + 1);
544	} while (offset < offset_max);
545	ibs_data.size = sizeof(u64) * size;
546
547	regs = *iregs;
548	if (check_rip && (ibs_data.regs[2] & IBS_RIP_INVALID)) {
549		regs.flags &= ~PERF_EFLAGS_EXACT;
550	} else {
551		instruction_pointer_set(&regs, ibs_data.regs[1]);
552		regs.flags |= PERF_EFLAGS_EXACT;
553	}
554
555	if (event->attr.sample_type & PERF_SAMPLE_RAW) {
556		raw.size = sizeof(u32) + ibs_data.size;
557		raw.data = ibs_data.data;
558		data.raw = &raw;
559	}
560
561	throttle = perf_event_overflow(event, &data, &regs);
562out:
563	if (throttle)
564		perf_ibs_disable_event(perf_ibs, hwc, *config);
565	else
566		perf_ibs_enable_event(perf_ibs, hwc, period >> 4);
567
568	perf_event_update_userpage(event);
569
570	return 1;
571}
572
573static int __kprobes
574perf_ibs_nmi_handler(unsigned int cmd, struct pt_regs *regs)
575{
576	int handled = 0;
577
578	handled += perf_ibs_handle_irq(&perf_ibs_fetch, regs);
579	handled += perf_ibs_handle_irq(&perf_ibs_op, regs);
580
581	if (handled)
582		inc_irq_stat(apic_perf_irqs);
583
584	return handled;
585}
586
587static __init int perf_ibs_pmu_init(struct perf_ibs *perf_ibs, char *name)
588{
589	struct cpu_perf_ibs __percpu *pcpu;
590	int ret;
591
592	pcpu = alloc_percpu(struct cpu_perf_ibs);
593	if (!pcpu)
594		return -ENOMEM;
595
596	perf_ibs->pcpu = pcpu;
597
598	ret = perf_pmu_register(&perf_ibs->pmu, name, -1);
599	if (ret) {
600		perf_ibs->pcpu = NULL;
601		free_percpu(pcpu);
602	}
603
604	return ret;
605}
606
607static __init int perf_event_ibs_init(void)
608{
609	if (!ibs_caps)
610		return -ENODEV;	/* ibs not supported by the cpu */
611
612	perf_ibs_pmu_init(&perf_ibs_fetch, "ibs_fetch");
613	if (ibs_caps & IBS_CAPS_OPCNT)
614		perf_ibs_op.config_mask |= IBS_OP_CNT_CTL;
615	perf_ibs_pmu_init(&perf_ibs_op, "ibs_op");
616	register_nmi_handler(NMI_LOCAL, perf_ibs_nmi_handler, 0, "perf_ibs");
617	printk(KERN_INFO "perf: AMD IBS detected (0x%08x)\n", ibs_caps);
618
619	return 0;
620}
621
622#else /* defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_AMD) */
623
624static __init int perf_event_ibs_init(void) { return 0; }
625
626#endif
627
628/* IBS - apic initialization, for perf and oprofile */
629
630static __init u32 __get_ibs_caps(void)
631{
632	u32 caps;
633	unsigned int max_level;
634
635	if (!boot_cpu_has(X86_FEATURE_IBS))
636		return 0;
637
638	/* check IBS cpuid feature flags */
639	max_level = cpuid_eax(0x80000000);
640	if (max_level < IBS_CPUID_FEATURES)
641		return IBS_CAPS_DEFAULT;
642
643	caps = cpuid_eax(IBS_CPUID_FEATURES);
644	if (!(caps & IBS_CAPS_AVAIL))
645		/* cpuid flags not valid */
646		return IBS_CAPS_DEFAULT;
647
648	return caps;
649}
650
651u32 get_ibs_caps(void)
652{
653	return ibs_caps;
654}
655
656EXPORT_SYMBOL(get_ibs_caps);
657
658static inline int get_eilvt(int offset)
659{
660	return !setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_NMI, 1);
661}
662
663static inline int put_eilvt(int offset)
664{
665	return !setup_APIC_eilvt(offset, 0, 0, 1);
666}
667
668/*
669 * Check and reserve APIC extended interrupt LVT offset for IBS if available.
670 */
671static inline int ibs_eilvt_valid(void)
672{
673	int offset;
674	u64 val;
675	int valid = 0;
676
677	preempt_disable();
678
679	rdmsrl(MSR_AMD64_IBSCTL, val);
680	offset = val & IBSCTL_LVT_OFFSET_MASK;
681
682	if (!(val & IBSCTL_LVT_OFFSET_VALID)) {
683		pr_err(FW_BUG "cpu %d, invalid IBS interrupt offset %d (MSR%08X=0x%016llx)\n",
684		       smp_processor_id(), offset, MSR_AMD64_IBSCTL, val);
685		goto out;
686	}
687
688	if (!get_eilvt(offset)) {
689		pr_err(FW_BUG "cpu %d, IBS interrupt offset %d not available (MSR%08X=0x%016llx)\n",
690		       smp_processor_id(), offset, MSR_AMD64_IBSCTL, val);
691		goto out;
692	}
693
694	valid = 1;
695out:
696	preempt_enable();
697
698	return valid;
699}
700
701static int setup_ibs_ctl(int ibs_eilvt_off)
702{
703	struct pci_dev *cpu_cfg;
704	int nodes;
705	u32 value = 0;
706
707	nodes = 0;
708	cpu_cfg = NULL;
709	do {
710		cpu_cfg = pci_get_device(PCI_VENDOR_ID_AMD,
711					 PCI_DEVICE_ID_AMD_10H_NB_MISC,
712					 cpu_cfg);
713		if (!cpu_cfg)
714			break;
715		++nodes;
716		pci_write_config_dword(cpu_cfg, IBSCTL, ibs_eilvt_off
717				       | IBSCTL_LVT_OFFSET_VALID);
718		pci_read_config_dword(cpu_cfg, IBSCTL, &value);
719		if (value != (ibs_eilvt_off | IBSCTL_LVT_OFFSET_VALID)) {
720			pci_dev_put(cpu_cfg);
721			printk(KERN_DEBUG "Failed to setup IBS LVT offset, "
722			       "IBSCTL = 0x%08x\n", value);
723			return -EINVAL;
724		}
725	} while (1);
726
727	if (!nodes) {
728		printk(KERN_DEBUG "No CPU node configured for IBS\n");
729		return -ENODEV;
730	}
731
732	return 0;
733}
734
735/*
736 * This runs only on the current cpu. We try to find an LVT offset and
737 * setup the local APIC. For this we must disable preemption. On
738 * success we initialize all nodes with this offset. This updates then
739 * the offset in the IBS_CTL per-node msr. The per-core APIC setup of
740 * the IBS interrupt vector is handled by perf_ibs_cpu_notifier that
741 * is using the new offset.
742 */
743static int force_ibs_eilvt_setup(void)
744{
745	int offset;
746	int ret;
747
748	preempt_disable();
749	/* find the next free available EILVT entry, skip offset 0 */
750	for (offset = 1; offset < APIC_EILVT_NR_MAX; offset++) {
751		if (get_eilvt(offset))
752			break;
753	}
754	preempt_enable();
755
756	if (offset == APIC_EILVT_NR_MAX) {
757		printk(KERN_DEBUG "No EILVT entry available\n");
758		return -EBUSY;
759	}
760
761	ret = setup_ibs_ctl(offset);
762	if (ret)
763		goto out;
764
765	if (!ibs_eilvt_valid()) {
766		ret = -EFAULT;
767		goto out;
768	}
769
770	pr_info("IBS: LVT offset %d assigned\n", offset);
771
772	return 0;
773out:
774	preempt_disable();
775	put_eilvt(offset);
776	preempt_enable();
777	return ret;
778}
779
780static inline int get_ibs_lvt_offset(void)
781{
782	u64 val;
783
784	rdmsrl(MSR_AMD64_IBSCTL, val);
785	if (!(val & IBSCTL_LVT_OFFSET_VALID))
786		return -EINVAL;
787
788	return val & IBSCTL_LVT_OFFSET_MASK;
789}
790
791static void setup_APIC_ibs(void *dummy)
792{
793	int offset;
794
795	offset = get_ibs_lvt_offset();
796	if (offset < 0)
797		goto failed;
798
799	if (!setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_NMI, 0))
800		return;
801failed:
802	pr_warn("perf: IBS APIC setup failed on cpu #%d\n",
803		smp_processor_id());
804}
805
806static void clear_APIC_ibs(void *dummy)
807{
808	int offset;
809
810	offset = get_ibs_lvt_offset();
811	if (offset >= 0)
812		setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_FIX, 1);
813}
814
815static int __cpuinit
816perf_ibs_cpu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
817{
818	switch (action & ~CPU_TASKS_FROZEN) {
819	case CPU_STARTING:
820		setup_APIC_ibs(NULL);
821		break;
822	case CPU_DYING:
823		clear_APIC_ibs(NULL);
824		break;
825	default:
826		break;
827	}
828
829	return NOTIFY_OK;
830}
831
832static __init int amd_ibs_init(void)
833{
834	u32 caps;
835	int ret = -EINVAL;
836
837	caps = __get_ibs_caps();
838	if (!caps)
839		return -ENODEV;	/* ibs not supported by the cpu */
840
841	/*
842	 * Force LVT offset assignment for family 10h: The offsets are
843	 * not assigned by the BIOS for this family, so the OS is
844	 * responsible for doing it. If the OS assignment fails, fall
845	 * back to BIOS settings and try to setup this.
846	 */
847	if (boot_cpu_data.x86 == 0x10)
848		force_ibs_eilvt_setup();
849
850	if (!ibs_eilvt_valid())
851		goto out;
852
853	get_online_cpus();
854	ibs_caps = caps;
855	/* make ibs_caps visible to other cpus: */
856	smp_mb();
857	perf_cpu_notifier(perf_ibs_cpu_notifier);
858	smp_call_function(setup_APIC_ibs, NULL, 1);
859	put_online_cpus();
860
861	ret = perf_event_ibs_init();
862out:
863	if (ret)
864		pr_err("Failed to setup IBS, %d\n", ret);
865	return ret;
866}
867
868/* Since we need the pci subsystem to init ibs we can't do this earlier: */
869device_initcall(amd_ibs_init);