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#include <linux/syscore_ops.h>
14
15#include <asm/apic.h>
16
17#include "perf_event.h"
18
19static u32 ibs_caps;
20
21#if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_AMD)
22
23#include <linux/kprobes.h>
24#include <linux/hardirq.h>
25
26#include <asm/nmi.h>
27
28#define IBS_FETCH_CONFIG_MASK	(IBS_FETCH_RAND_EN | IBS_FETCH_MAX_CNT)
29#define IBS_OP_CONFIG_MASK	IBS_OP_MAX_CNT
30
31enum ibs_states {
32	IBS_ENABLED	= 0,
33	IBS_STARTED	= 1,
34	IBS_STOPPING	= 2,
35
36	IBS_MAX_STATES,
37};
38
39struct cpu_perf_ibs {
40	struct perf_event	*event;
41	unsigned long		state[BITS_TO_LONGS(IBS_MAX_STATES)];
42};
43
44struct perf_ibs {
45	struct pmu			pmu;
46	unsigned int			msr;
47	u64				config_mask;
48	u64				cnt_mask;
49	u64				enable_mask;
50	u64				valid_mask;
51	u64				max_period;
52	unsigned long			offset_mask[1];
53	int				offset_max;
54	struct cpu_perf_ibs __percpu	*pcpu;
55
56	struct attribute		**format_attrs;
57	struct attribute_group		format_group;
58	const struct attribute_group	*attr_groups[2];
59
60	u64				(*get_count)(u64 config);
61};
62
63struct perf_ibs_data {
64	u32		size;
65	union {
66		u32	data[0];	/* data buffer starts here */
67		u32	caps;
68	};
69	u64		regs[MSR_AMD64_IBS_REG_COUNT_MAX];
70};
71
72static int
73perf_event_set_period(struct hw_perf_event *hwc, u64 min, u64 max, u64 *hw_period)
74{
75	s64 left = local64_read(&hwc->period_left);
76	s64 period = hwc->sample_period;
77	int overflow = 0;
78
79	/*
80	 * If we are way outside a reasonable range then just skip forward:
81	 */
82	if (unlikely(left <= -period)) {
83		left = period;
84		local64_set(&hwc->period_left, left);
85		hwc->last_period = period;
86		overflow = 1;
87	}
88
89	if (unlikely(left < (s64)min)) {
90		left += period;
91		local64_set(&hwc->period_left, left);
92		hwc->last_period = period;
93		overflow = 1;
94	}
95
96	/*
97	 * If the hw period that triggers the sw overflow is too short
98	 * we might hit the irq handler. This biases the results.
99	 * Thus we shorten the next-to-last period and set the last
100	 * period to the max period.
101	 */
102	if (left > max) {
103		left -= max;
104		if (left > max)
105			left = max;
106		else if (left < min)
107			left = min;
108	}
109
110	*hw_period = (u64)left;
111
112	return overflow;
113}
114
115static  int
116perf_event_try_update(struct perf_event *event, u64 new_raw_count, int width)
117{
118	struct hw_perf_event *hwc = &event->hw;
119	int shift = 64 - width;
120	u64 prev_raw_count;
121	u64 delta;
122
123	/*
124	 * Careful: an NMI might modify the previous event value.
125	 *
126	 * Our tactic to handle this is to first atomically read and
127	 * exchange a new raw count - then add that new-prev delta
128	 * count to the generic event atomically:
129	 */
130	prev_raw_count = local64_read(&hwc->prev_count);
131	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
132					new_raw_count) != prev_raw_count)
133		return 0;
134
135	/*
136	 * Now we have the new raw value and have updated the prev
137	 * timestamp already. We can now calculate the elapsed delta
138	 * (event-)time and add that to the generic event.
139	 *
140	 * Careful, not all hw sign-extends above the physical width
141	 * of the count.
142	 */
143	delta = (new_raw_count << shift) - (prev_raw_count << shift);
144	delta >>= shift;
145
146	local64_add(delta, &event->count);
147	local64_sub(delta, &hwc->period_left);
148
149	return 1;
150}
151
152static struct perf_ibs perf_ibs_fetch;
153static struct perf_ibs perf_ibs_op;
154
155static struct perf_ibs *get_ibs_pmu(int type)
156{
157	if (perf_ibs_fetch.pmu.type == type)
158		return &perf_ibs_fetch;
159	if (perf_ibs_op.pmu.type == type)
160		return &perf_ibs_op;
161	return NULL;
162}
163
164/*
165 * Use IBS for precise event sampling:
166 *
167 *  perf record -a -e cpu-cycles:p ...    # use ibs op counting cycle count
168 *  perf record -a -e r076:p ...          # same as -e cpu-cycles:p
169 *  perf record -a -e r0C1:p ...          # use ibs op counting micro-ops
170 *
171 * IbsOpCntCtl (bit 19) of IBS Execution Control Register (IbsOpCtl,
172 * MSRC001_1033) is used to select either cycle or micro-ops counting
173 * mode.
174 *
175 * The rip of IBS samples has skid 0. Thus, IBS supports precise
176 * levels 1 and 2 and the PERF_EFLAGS_EXACT is set. In rare cases the
177 * rip is invalid when IBS was not able to record the rip correctly.
178 * We clear PERF_EFLAGS_EXACT and take the rip from pt_regs then.
179 *
180 */
181static int perf_ibs_precise_event(struct perf_event *event, u64 *config)
182{
183	switch (event->attr.precise_ip) {
184	case 0:
185		return -ENOENT;
186	case 1:
187	case 2:
188		break;
189	default:
190		return -EOPNOTSUPP;
191	}
192
193	switch (event->attr.type) {
194	case PERF_TYPE_HARDWARE:
195		switch (event->attr.config) {
196		case PERF_COUNT_HW_CPU_CYCLES:
197			*config = 0;
198			return 0;
199		}
200		break;
201	case PERF_TYPE_RAW:
202		switch (event->attr.config) {
203		case 0x0076:
204			*config = 0;
205			return 0;
206		case 0x00C1:
207			*config = IBS_OP_CNT_CTL;
208			return 0;
209		}
210		break;
211	default:
212		return -ENOENT;
213	}
214
215	return -EOPNOTSUPP;
216}
217
218static const struct perf_event_attr ibs_notsupp = {
219	.exclude_user	= 1,
220	.exclude_kernel	= 1,
221	.exclude_hv	= 1,
222	.exclude_idle	= 1,
223	.exclude_host	= 1,
224	.exclude_guest	= 1,
225};
226
227static int perf_ibs_init(struct perf_event *event)
228{
229	struct hw_perf_event *hwc = &event->hw;
230	struct perf_ibs *perf_ibs;
231	u64 max_cnt, config;
232	int ret;
233
234	perf_ibs = get_ibs_pmu(event->attr.type);
235	if (perf_ibs) {
236		config = event->attr.config;
237	} else {
238		perf_ibs = &perf_ibs_op;
239		ret = perf_ibs_precise_event(event, &config);
240		if (ret)
241			return ret;
242	}
243
244	if (event->pmu != &perf_ibs->pmu)
245		return -ENOENT;
246
247	if (perf_flags(&event->attr) & perf_flags(&ibs_notsupp))
248		return -EINVAL;
249
250	if (config & ~perf_ibs->config_mask)
251		return -EINVAL;
252
253	if (hwc->sample_period) {
254		if (config & perf_ibs->cnt_mask)
255			/* raw max_cnt may not be set */
256			return -EINVAL;
257		if (!event->attr.sample_freq && hwc->sample_period & 0x0f)
258			/*
259			 * lower 4 bits can not be set in ibs max cnt,
260			 * but allowing it in case we adjust the
261			 * sample period to set a frequency.
262			 */
263			return -EINVAL;
264		hwc->sample_period &= ~0x0FULL;
265		if (!hwc->sample_period)
266			hwc->sample_period = 0x10;
267	} else {
268		max_cnt = config & perf_ibs->cnt_mask;
269		config &= ~perf_ibs->cnt_mask;
270		event->attr.sample_period = max_cnt << 4;
271		hwc->sample_period = event->attr.sample_period;
272	}
273
274	if (!hwc->sample_period)
275		return -EINVAL;
276
277	/*
278	 * If we modify hwc->sample_period, we also need to update
279	 * hwc->last_period and hwc->period_left.
280	 */
281	hwc->last_period = hwc->sample_period;
282	local64_set(&hwc->period_left, hwc->sample_period);
283
284	hwc->config_base = perf_ibs->msr;
285	hwc->config = config;
286
287	return 0;
288}
289
290static int perf_ibs_set_period(struct perf_ibs *perf_ibs,
291			       struct hw_perf_event *hwc, u64 *period)
292{
293	int overflow;
294
295	/* ignore lower 4 bits in min count: */
296	overflow = perf_event_set_period(hwc, 1<<4, perf_ibs->max_period, period);
297	local64_set(&hwc->prev_count, 0);
298
299	return overflow;
300}
301
302static u64 get_ibs_fetch_count(u64 config)
303{
304	return (config & IBS_FETCH_CNT) >> 12;
305}
306
307static u64 get_ibs_op_count(u64 config)
308{
309	u64 count = 0;
310
311	if (config & IBS_OP_VAL)
312		count += (config & IBS_OP_MAX_CNT) << 4; /* cnt rolled over */
313
314	if (ibs_caps & IBS_CAPS_RDWROPCNT)
315		count += (config & IBS_OP_CUR_CNT) >> 32;
316
317	return count;
318}
319
320static void
321perf_ibs_event_update(struct perf_ibs *perf_ibs, struct perf_event *event,
322		      u64 *config)
323{
324	u64 count = perf_ibs->get_count(*config);
325
326	/*
327	 * Set width to 64 since we do not overflow on max width but
328	 * instead on max count. In perf_ibs_set_period() we clear
329	 * prev count manually on overflow.
330	 */
331	while (!perf_event_try_update(event, count, 64)) {
332		rdmsrl(event->hw.config_base, *config);
333		count = perf_ibs->get_count(*config);
334	}
335}
336
337static inline void perf_ibs_enable_event(struct perf_ibs *perf_ibs,
338					 struct hw_perf_event *hwc, u64 config)
339{
340	wrmsrl(hwc->config_base, hwc->config | config | perf_ibs->enable_mask);
341}
342
343/*
344 * Erratum #420 Instruction-Based Sampling Engine May Generate
345 * Interrupt that Cannot Be Cleared:
346 *
347 * Must clear counter mask first, then clear the enable bit. See
348 * Revision Guide for AMD Family 10h Processors, Publication #41322.
349 */
350static inline void perf_ibs_disable_event(struct perf_ibs *perf_ibs,
351					  struct hw_perf_event *hwc, u64 config)
352{
353	config &= ~perf_ibs->cnt_mask;
354	wrmsrl(hwc->config_base, config);
355	config &= ~perf_ibs->enable_mask;
356	wrmsrl(hwc->config_base, config);
357}
358
359/*
360 * We cannot restore the ibs pmu state, so we always needs to update
361 * the event while stopping it and then reset the state when starting
362 * again. Thus, ignoring PERF_EF_RELOAD and PERF_EF_UPDATE flags in
363 * perf_ibs_start()/perf_ibs_stop() and instead always do it.
364 */
365static void perf_ibs_start(struct perf_event *event, int flags)
366{
367	struct hw_perf_event *hwc = &event->hw;
368	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
369	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
370	u64 period;
371
372	if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
373		return;
374
375	WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
376	hwc->state = 0;
377
378	perf_ibs_set_period(perf_ibs, hwc, &period);
379	set_bit(IBS_STARTED, pcpu->state);
380	perf_ibs_enable_event(perf_ibs, hwc, period >> 4);
381
382	perf_event_update_userpage(event);
383}
384
385static void perf_ibs_stop(struct perf_event *event, int flags)
386{
387	struct hw_perf_event *hwc = &event->hw;
388	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
389	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
390	u64 config;
391	int stopping;
392
393	stopping = test_and_clear_bit(IBS_STARTED, pcpu->state);
394
395	if (!stopping && (hwc->state & PERF_HES_UPTODATE))
396		return;
397
398	rdmsrl(hwc->config_base, config);
399
400	if (stopping) {
401		set_bit(IBS_STOPPING, pcpu->state);
402		perf_ibs_disable_event(perf_ibs, hwc, config);
403		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
404		hwc->state |= PERF_HES_STOPPED;
405	}
406
407	if (hwc->state & PERF_HES_UPTODATE)
408		return;
409
410	/*
411	 * Clear valid bit to not count rollovers on update, rollovers
412	 * are only updated in the irq handler.
413	 */
414	config &= ~perf_ibs->valid_mask;
415
416	perf_ibs_event_update(perf_ibs, event, &config);
417	hwc->state |= PERF_HES_UPTODATE;
418}
419
420static int perf_ibs_add(struct perf_event *event, int flags)
421{
422	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
423	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
424
425	if (test_and_set_bit(IBS_ENABLED, pcpu->state))
426		return -ENOSPC;
427
428	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
429
430	pcpu->event = event;
431
432	if (flags & PERF_EF_START)
433		perf_ibs_start(event, PERF_EF_RELOAD);
434
435	return 0;
436}
437
438static void perf_ibs_del(struct perf_event *event, int flags)
439{
440	struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
441	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
442
443	if (!test_and_clear_bit(IBS_ENABLED, pcpu->state))
444		return;
445
446	perf_ibs_stop(event, PERF_EF_UPDATE);
447
448	pcpu->event = NULL;
449
450	perf_event_update_userpage(event);
451}
452
453static void perf_ibs_read(struct perf_event *event) { }
454
455PMU_FORMAT_ATTR(rand_en,	"config:57");
456PMU_FORMAT_ATTR(cnt_ctl,	"config:19");
457
458static struct attribute *ibs_fetch_format_attrs[] = {
459	&format_attr_rand_en.attr,
460	NULL,
461};
462
463static struct attribute *ibs_op_format_attrs[] = {
464	NULL,	/* &format_attr_cnt_ctl.attr if IBS_CAPS_OPCNT */
465	NULL,
466};
467
468static struct perf_ibs perf_ibs_fetch = {
469	.pmu = {
470		.task_ctx_nr	= perf_invalid_context,
471
472		.event_init	= perf_ibs_init,
473		.add		= perf_ibs_add,
474		.del		= perf_ibs_del,
475		.start		= perf_ibs_start,
476		.stop		= perf_ibs_stop,
477		.read		= perf_ibs_read,
478	},
479	.msr			= MSR_AMD64_IBSFETCHCTL,
480	.config_mask		= IBS_FETCH_CONFIG_MASK,
481	.cnt_mask		= IBS_FETCH_MAX_CNT,
482	.enable_mask		= IBS_FETCH_ENABLE,
483	.valid_mask		= IBS_FETCH_VAL,
484	.max_period		= IBS_FETCH_MAX_CNT << 4,
485	.offset_mask		= { MSR_AMD64_IBSFETCH_REG_MASK },
486	.offset_max		= MSR_AMD64_IBSFETCH_REG_COUNT,
487	.format_attrs		= ibs_fetch_format_attrs,
488
489	.get_count		= get_ibs_fetch_count,
490};
491
492static struct perf_ibs perf_ibs_op = {
493	.pmu = {
494		.task_ctx_nr	= perf_invalid_context,
495
496		.event_init	= perf_ibs_init,
497		.add		= perf_ibs_add,
498		.del		= perf_ibs_del,
499		.start		= perf_ibs_start,
500		.stop		= perf_ibs_stop,
501		.read		= perf_ibs_read,
502	},
503	.msr			= MSR_AMD64_IBSOPCTL,
504	.config_mask		= IBS_OP_CONFIG_MASK,
505	.cnt_mask		= IBS_OP_MAX_CNT,
506	.enable_mask		= IBS_OP_ENABLE,
507	.valid_mask		= IBS_OP_VAL,
508	.max_period		= IBS_OP_MAX_CNT << 4,
509	.offset_mask		= { MSR_AMD64_IBSOP_REG_MASK },
510	.offset_max		= MSR_AMD64_IBSOP_REG_COUNT,
511	.format_attrs		= ibs_op_format_attrs,
512
513	.get_count		= get_ibs_op_count,
514};
515
516static int perf_ibs_handle_irq(struct perf_ibs *perf_ibs, struct pt_regs *iregs)
517{
518	struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
519	struct perf_event *event = pcpu->event;
520	struct hw_perf_event *hwc = &event->hw;
521	struct perf_sample_data data;
522	struct perf_raw_record raw;
523	struct pt_regs regs;
524	struct perf_ibs_data ibs_data;
525	int offset, size, check_rip, offset_max, throttle = 0;
526	unsigned int msr;
527	u64 *buf, *config, period;
528
529	if (!test_bit(IBS_STARTED, pcpu->state)) {
530		/*
531		 * Catch spurious interrupts after stopping IBS: After
532		 * disabling IBS there could be still incoming NMIs
533		 * with samples that even have the valid bit cleared.
534		 * Mark all this NMIs as handled.
535		 */
536		return test_and_clear_bit(IBS_STOPPING, pcpu->state) ? 1 : 0;
537	}
538
539	msr = hwc->config_base;
540	buf = ibs_data.regs;
541	rdmsrl(msr, *buf);
542	if (!(*buf++ & perf_ibs->valid_mask))
543		return 0;
544
545	config = &ibs_data.regs[0];
546	perf_ibs_event_update(perf_ibs, event, config);
547	perf_sample_data_init(&data, 0, hwc->last_period);
548	if (!perf_ibs_set_period(perf_ibs, hwc, &period))
549		goto out;	/* no sw counter overflow */
550
551	ibs_data.caps = ibs_caps;
552	size = 1;
553	offset = 1;
554	check_rip = (perf_ibs == &perf_ibs_op && (ibs_caps & IBS_CAPS_RIPINVALIDCHK));
555	if (event->attr.sample_type & PERF_SAMPLE_RAW)
556		offset_max = perf_ibs->offset_max;
557	else if (check_rip)
558		offset_max = 2;
559	else
560		offset_max = 1;
561	do {
562		rdmsrl(msr + offset, *buf++);
563		size++;
564		offset = find_next_bit(perf_ibs->offset_mask,
565				       perf_ibs->offset_max,
566				       offset + 1);
567	} while (offset < offset_max);
568	if (event->attr.sample_type & PERF_SAMPLE_RAW) {
569		/*
570		 * Read IbsBrTarget and IbsOpData4 separately
571		 * depending on their availability.
572		 * Can't add to offset_max as they are staggered
573		 */
574		if (ibs_caps & IBS_CAPS_BRNTRGT) {
575			rdmsrl(MSR_AMD64_IBSBRTARGET, *buf++);
576			size++;
577		}
578		if (ibs_caps & IBS_CAPS_OPDATA4) {
579			rdmsrl(MSR_AMD64_IBSOPDATA4, *buf++);
580			size++;
581		}
582	}
583	ibs_data.size = sizeof(u64) * size;
584
585	regs = *iregs;
586	if (check_rip && (ibs_data.regs[2] & IBS_RIP_INVALID)) {
587		regs.flags &= ~PERF_EFLAGS_EXACT;
588	} else {
589		set_linear_ip(&regs, ibs_data.regs[1]);
590		regs.flags |= PERF_EFLAGS_EXACT;
591	}
592
593	if (event->attr.sample_type & PERF_SAMPLE_RAW) {
594		raw.size = sizeof(u32) + ibs_data.size;
595		raw.data = ibs_data.data;
596		data.raw = &raw;
597	}
598
599	throttle = perf_event_overflow(event, &data, &regs);
600out:
601	if (throttle)
602		perf_ibs_disable_event(perf_ibs, hwc, *config);
603	else
604		perf_ibs_enable_event(perf_ibs, hwc, period >> 4);
605
606	perf_event_update_userpage(event);
607
608	return 1;
609}
610
611static int
612perf_ibs_nmi_handler(unsigned int cmd, struct pt_regs *regs)
613{
614	int handled = 0;
615
616	handled += perf_ibs_handle_irq(&perf_ibs_fetch, regs);
617	handled += perf_ibs_handle_irq(&perf_ibs_op, regs);
618
619	if (handled)
620		inc_irq_stat(apic_perf_irqs);
621
622	return handled;
623}
624NOKPROBE_SYMBOL(perf_ibs_nmi_handler);
625
626static __init int perf_ibs_pmu_init(struct perf_ibs *perf_ibs, char *name)
627{
628	struct cpu_perf_ibs __percpu *pcpu;
629	int ret;
630
631	pcpu = alloc_percpu(struct cpu_perf_ibs);
632	if (!pcpu)
633		return -ENOMEM;
634
635	perf_ibs->pcpu = pcpu;
636
637	/* register attributes */
638	if (perf_ibs->format_attrs[0]) {
639		memset(&perf_ibs->format_group, 0, sizeof(perf_ibs->format_group));
640		perf_ibs->format_group.name	= "format";
641		perf_ibs->format_group.attrs	= perf_ibs->format_attrs;
642
643		memset(&perf_ibs->attr_groups, 0, sizeof(perf_ibs->attr_groups));
644		perf_ibs->attr_groups[0]	= &perf_ibs->format_group;
645		perf_ibs->pmu.attr_groups	= perf_ibs->attr_groups;
646	}
647
648	ret = perf_pmu_register(&perf_ibs->pmu, name, -1);
649	if (ret) {
650		perf_ibs->pcpu = NULL;
651		free_percpu(pcpu);
652	}
653
654	return ret;
655}
656
657static __init int perf_event_ibs_init(void)
658{
659	struct attribute **attr = ibs_op_format_attrs;
660
661	if (!ibs_caps)
662		return -ENODEV;	/* ibs not supported by the cpu */
663
664	perf_ibs_pmu_init(&perf_ibs_fetch, "ibs_fetch");
665
666	if (ibs_caps & IBS_CAPS_OPCNT) {
667		perf_ibs_op.config_mask |= IBS_OP_CNT_CTL;
668		*attr++ = &format_attr_cnt_ctl.attr;
669	}
670	perf_ibs_pmu_init(&perf_ibs_op, "ibs_op");
671
672	register_nmi_handler(NMI_LOCAL, perf_ibs_nmi_handler, 0, "perf_ibs");
673	printk(KERN_INFO "perf: AMD IBS detected (0x%08x)\n", ibs_caps);
674
675	return 0;
676}
677
678#else /* defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_AMD) */
679
680static __init int perf_event_ibs_init(void) { return 0; }
681
682#endif
683
684/* IBS - apic initialization, for perf and oprofile */
685
686static __init u32 __get_ibs_caps(void)
687{
688	u32 caps;
689	unsigned int max_level;
690
691	if (!boot_cpu_has(X86_FEATURE_IBS))
692		return 0;
693
694	/* check IBS cpuid feature flags */
695	max_level = cpuid_eax(0x80000000);
696	if (max_level < IBS_CPUID_FEATURES)
697		return IBS_CAPS_DEFAULT;
698
699	caps = cpuid_eax(IBS_CPUID_FEATURES);
700	if (!(caps & IBS_CAPS_AVAIL))
701		/* cpuid flags not valid */
702		return IBS_CAPS_DEFAULT;
703
704	return caps;
705}
706
707u32 get_ibs_caps(void)
708{
709	return ibs_caps;
710}
711
712EXPORT_SYMBOL(get_ibs_caps);
713
714static inline int get_eilvt(int offset)
715{
716	return !setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_NMI, 1);
717}
718
719static inline int put_eilvt(int offset)
720{
721	return !setup_APIC_eilvt(offset, 0, 0, 1);
722}
723
724/*
725 * Check and reserve APIC extended interrupt LVT offset for IBS if available.
726 */
727static inline int ibs_eilvt_valid(void)
728{
729	int offset;
730	u64 val;
731	int valid = 0;
732
733	preempt_disable();
734
735	rdmsrl(MSR_AMD64_IBSCTL, val);
736	offset = val & IBSCTL_LVT_OFFSET_MASK;
737
738	if (!(val & IBSCTL_LVT_OFFSET_VALID)) {
739		pr_err(FW_BUG "cpu %d, invalid IBS interrupt offset %d (MSR%08X=0x%016llx)\n",
740		       smp_processor_id(), offset, MSR_AMD64_IBSCTL, val);
741		goto out;
742	}
743
744	if (!get_eilvt(offset)) {
745		pr_err(FW_BUG "cpu %d, IBS interrupt offset %d not available (MSR%08X=0x%016llx)\n",
746		       smp_processor_id(), offset, MSR_AMD64_IBSCTL, val);
747		goto out;
748	}
749
750	valid = 1;
751out:
752	preempt_enable();
753
754	return valid;
755}
756
757static int setup_ibs_ctl(int ibs_eilvt_off)
758{
759	struct pci_dev *cpu_cfg;
760	int nodes;
761	u32 value = 0;
762
763	nodes = 0;
764	cpu_cfg = NULL;
765	do {
766		cpu_cfg = pci_get_device(PCI_VENDOR_ID_AMD,
767					 PCI_DEVICE_ID_AMD_10H_NB_MISC,
768					 cpu_cfg);
769		if (!cpu_cfg)
770			break;
771		++nodes;
772		pci_write_config_dword(cpu_cfg, IBSCTL, ibs_eilvt_off
773				       | IBSCTL_LVT_OFFSET_VALID);
774		pci_read_config_dword(cpu_cfg, IBSCTL, &value);
775		if (value != (ibs_eilvt_off | IBSCTL_LVT_OFFSET_VALID)) {
776			pci_dev_put(cpu_cfg);
777			printk(KERN_DEBUG "Failed to setup IBS LVT offset, "
778			       "IBSCTL = 0x%08x\n", value);
779			return -EINVAL;
780		}
781	} while (1);
782
783	if (!nodes) {
784		printk(KERN_DEBUG "No CPU node configured for IBS\n");
785		return -ENODEV;
786	}
787
788	return 0;
789}
790
791/*
792 * This runs only on the current cpu. We try to find an LVT offset and
793 * setup the local APIC. For this we must disable preemption. On
794 * success we initialize all nodes with this offset. This updates then
795 * the offset in the IBS_CTL per-node msr. The per-core APIC setup of
796 * the IBS interrupt vector is handled by perf_ibs_cpu_notifier that
797 * is using the new offset.
798 */
799static void force_ibs_eilvt_setup(void)
800{
801	int offset;
802	int ret;
803
804	preempt_disable();
805	/* find the next free available EILVT entry, skip offset 0 */
806	for (offset = 1; offset < APIC_EILVT_NR_MAX; offset++) {
807		if (get_eilvt(offset))
808			break;
809	}
810	preempt_enable();
811
812	if (offset == APIC_EILVT_NR_MAX) {
813		printk(KERN_DEBUG "No EILVT entry available\n");
814		return;
815	}
816
817	ret = setup_ibs_ctl(offset);
818	if (ret)
819		goto out;
820
821	if (!ibs_eilvt_valid())
822		goto out;
823
824	pr_info("IBS: LVT offset %d assigned\n", offset);
825
826	return;
827out:
828	preempt_disable();
829	put_eilvt(offset);
830	preempt_enable();
831	return;
832}
833
834static void ibs_eilvt_setup(void)
835{
836	/*
837	 * Force LVT offset assignment for family 10h: The offsets are
838	 * not assigned by the BIOS for this family, so the OS is
839	 * responsible for doing it. If the OS assignment fails, fall
840	 * back to BIOS settings and try to setup this.
841	 */
842	if (boot_cpu_data.x86 == 0x10)
843		force_ibs_eilvt_setup();
844}
845
846static inline int get_ibs_lvt_offset(void)
847{
848	u64 val;
849
850	rdmsrl(MSR_AMD64_IBSCTL, val);
851	if (!(val & IBSCTL_LVT_OFFSET_VALID))
852		return -EINVAL;
853
854	return val & IBSCTL_LVT_OFFSET_MASK;
855}
856
857static void setup_APIC_ibs(void *dummy)
858{
859	int offset;
860
861	offset = get_ibs_lvt_offset();
862	if (offset < 0)
863		goto failed;
864
865	if (!setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_NMI, 0))
866		return;
867failed:
868	pr_warn("perf: IBS APIC setup failed on cpu #%d\n",
869		smp_processor_id());
870}
871
872static void clear_APIC_ibs(void *dummy)
873{
874	int offset;
875
876	offset = get_ibs_lvt_offset();
877	if (offset >= 0)
878		setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_FIX, 1);
879}
880
881#ifdef CONFIG_PM
882
883static int perf_ibs_suspend(void)
884{
885	clear_APIC_ibs(NULL);
886	return 0;
887}
888
889static void perf_ibs_resume(void)
890{
891	ibs_eilvt_setup();
892	setup_APIC_ibs(NULL);
893}
894
895static struct syscore_ops perf_ibs_syscore_ops = {
896	.resume		= perf_ibs_resume,
897	.suspend	= perf_ibs_suspend,
898};
899
900static void perf_ibs_pm_init(void)
901{
902	register_syscore_ops(&perf_ibs_syscore_ops);
903}
904
905#else
906
907static inline void perf_ibs_pm_init(void) { }
908
909#endif
910
911static int
912perf_ibs_cpu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
913{
914	switch (action & ~CPU_TASKS_FROZEN) {
915	case CPU_STARTING:
916		setup_APIC_ibs(NULL);
917		break;
918	case CPU_DYING:
919		clear_APIC_ibs(NULL);
920		break;
921	default:
922		break;
923	}
924
925	return NOTIFY_OK;
926}
927
928static __init int amd_ibs_init(void)
929{
930	u32 caps;
931	int ret = -EINVAL;
932
933	caps = __get_ibs_caps();
934	if (!caps)
935		return -ENODEV;	/* ibs not supported by the cpu */
936
937	ibs_eilvt_setup();
938
939	if (!ibs_eilvt_valid())
940		goto out;
941
942	perf_ibs_pm_init();
943	cpu_notifier_register_begin();
944	ibs_caps = caps;
945	/* make ibs_caps visible to other cpus: */
946	smp_mb();
947	smp_call_function(setup_APIC_ibs, NULL, 1);
948	__perf_cpu_notifier(perf_ibs_cpu_notifier);
949	cpu_notifier_register_done();
950
951	ret = perf_event_ibs_init();
952out:
953	if (ret)
954		pr_err("Failed to setup IBS, %d\n", ret);
955	return ret;
956}
957
958/* Since we need the pci subsystem to init ibs we can't do this earlier: */
959device_initcall(amd_ibs_init);
960