xref: /kernel/linux/linux-5.10/arch/x86/events/intel/pt.c (revision 8c2ecf20)
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Intel(R) Processor Trace PMU driver for perf
4 * Copyright (c) 2013-2014, Intel Corporation.
5 *
6 * Intel PT is specified in the Intel Architecture Instruction Set Extensions
7 * Programming Reference:
8 * http://software.intel.com/en-us/intel-isa-extensions
9 */
10
11#undef DEBUG
12
13#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14
15#include <linux/types.h>
16#include <linux/bits.h>
17#include <linux/limits.h>
18#include <linux/slab.h>
19#include <linux/device.h>
20
21#include <asm/perf_event.h>
22#include <asm/insn.h>
23#include <asm/io.h>
24#include <asm/intel_pt.h>
25#include <asm/intel-family.h>
26
27#include "../perf_event.h"
28#include "pt.h"
29
30static DEFINE_PER_CPU(struct pt, pt_ctx);
31
32static struct pt_pmu pt_pmu;
33
34/*
35 * Capabilities of Intel PT hardware, such as number of address bits or
36 * supported output schemes, are cached and exported to userspace as "caps"
37 * attribute group of pt pmu device
38 * (/sys/bus/event_source/devices/intel_pt/caps/) so that userspace can store
39 * relevant bits together with intel_pt traces.
40 *
41 * These are necessary for both trace decoding (payloads_lip, contains address
42 * width encoded in IP-related packets), and event configuration (bitmasks with
43 * permitted values for certain bit fields).
44 */
45#define PT_CAP(_n, _l, _r, _m)						\
46	[PT_CAP_ ## _n] = { .name = __stringify(_n), .leaf = _l,	\
47			    .reg = _r, .mask = _m }
48
49static struct pt_cap_desc {
50	const char	*name;
51	u32		leaf;
52	u8		reg;
53	u32		mask;
54} pt_caps[] = {
55	PT_CAP(max_subleaf,		0, CPUID_EAX, 0xffffffff),
56	PT_CAP(cr3_filtering,		0, CPUID_EBX, BIT(0)),
57	PT_CAP(psb_cyc,			0, CPUID_EBX, BIT(1)),
58	PT_CAP(ip_filtering,		0, CPUID_EBX, BIT(2)),
59	PT_CAP(mtc,			0, CPUID_EBX, BIT(3)),
60	PT_CAP(ptwrite,			0, CPUID_EBX, BIT(4)),
61	PT_CAP(power_event_trace,	0, CPUID_EBX, BIT(5)),
62	PT_CAP(topa_output,		0, CPUID_ECX, BIT(0)),
63	PT_CAP(topa_multiple_entries,	0, CPUID_ECX, BIT(1)),
64	PT_CAP(single_range_output,	0, CPUID_ECX, BIT(2)),
65	PT_CAP(output_subsys,		0, CPUID_ECX, BIT(3)),
66	PT_CAP(payloads_lip,		0, CPUID_ECX, BIT(31)),
67	PT_CAP(num_address_ranges,	1, CPUID_EAX, 0x7),
68	PT_CAP(mtc_periods,		1, CPUID_EAX, 0xffff0000),
69	PT_CAP(cycle_thresholds,	1, CPUID_EBX, 0xffff),
70	PT_CAP(psb_periods,		1, CPUID_EBX, 0xffff0000),
71};
72
73u32 intel_pt_validate_cap(u32 *caps, enum pt_capabilities capability)
74{
75	struct pt_cap_desc *cd = &pt_caps[capability];
76	u32 c = caps[cd->leaf * PT_CPUID_REGS_NUM + cd->reg];
77	unsigned int shift = __ffs(cd->mask);
78
79	return (c & cd->mask) >> shift;
80}
81EXPORT_SYMBOL_GPL(intel_pt_validate_cap);
82
83u32 intel_pt_validate_hw_cap(enum pt_capabilities cap)
84{
85	return intel_pt_validate_cap(pt_pmu.caps, cap);
86}
87EXPORT_SYMBOL_GPL(intel_pt_validate_hw_cap);
88
89static ssize_t pt_cap_show(struct device *cdev,
90			   struct device_attribute *attr,
91			   char *buf)
92{
93	struct dev_ext_attribute *ea =
94		container_of(attr, struct dev_ext_attribute, attr);
95	enum pt_capabilities cap = (long)ea->var;
96
97	return snprintf(buf, PAGE_SIZE, "%x\n", intel_pt_validate_hw_cap(cap));
98}
99
100static struct attribute_group pt_cap_group __ro_after_init = {
101	.name	= "caps",
102};
103
104PMU_FORMAT_ATTR(pt,		"config:0"	);
105PMU_FORMAT_ATTR(cyc,		"config:1"	);
106PMU_FORMAT_ATTR(pwr_evt,	"config:4"	);
107PMU_FORMAT_ATTR(fup_on_ptw,	"config:5"	);
108PMU_FORMAT_ATTR(mtc,		"config:9"	);
109PMU_FORMAT_ATTR(tsc,		"config:10"	);
110PMU_FORMAT_ATTR(noretcomp,	"config:11"	);
111PMU_FORMAT_ATTR(ptw,		"config:12"	);
112PMU_FORMAT_ATTR(branch,		"config:13"	);
113PMU_FORMAT_ATTR(mtc_period,	"config:14-17"	);
114PMU_FORMAT_ATTR(cyc_thresh,	"config:19-22"	);
115PMU_FORMAT_ATTR(psb_period,	"config:24-27"	);
116
117static struct attribute *pt_formats_attr[] = {
118	&format_attr_pt.attr,
119	&format_attr_cyc.attr,
120	&format_attr_pwr_evt.attr,
121	&format_attr_fup_on_ptw.attr,
122	&format_attr_mtc.attr,
123	&format_attr_tsc.attr,
124	&format_attr_noretcomp.attr,
125	&format_attr_ptw.attr,
126	&format_attr_branch.attr,
127	&format_attr_mtc_period.attr,
128	&format_attr_cyc_thresh.attr,
129	&format_attr_psb_period.attr,
130	NULL,
131};
132
133static struct attribute_group pt_format_group = {
134	.name	= "format",
135	.attrs	= pt_formats_attr,
136};
137
138static ssize_t
139pt_timing_attr_show(struct device *dev, struct device_attribute *attr,
140		    char *page)
141{
142	struct perf_pmu_events_attr *pmu_attr =
143		container_of(attr, struct perf_pmu_events_attr, attr);
144
145	switch (pmu_attr->id) {
146	case 0:
147		return sprintf(page, "%lu\n", pt_pmu.max_nonturbo_ratio);
148	case 1:
149		return sprintf(page, "%u:%u\n",
150			       pt_pmu.tsc_art_num,
151			       pt_pmu.tsc_art_den);
152	default:
153		break;
154	}
155
156	return -EINVAL;
157}
158
159PMU_EVENT_ATTR(max_nonturbo_ratio, timing_attr_max_nonturbo_ratio, 0,
160	       pt_timing_attr_show);
161PMU_EVENT_ATTR(tsc_art_ratio, timing_attr_tsc_art_ratio, 1,
162	       pt_timing_attr_show);
163
164static struct attribute *pt_timing_attr[] = {
165	&timing_attr_max_nonturbo_ratio.attr.attr,
166	&timing_attr_tsc_art_ratio.attr.attr,
167	NULL,
168};
169
170static struct attribute_group pt_timing_group = {
171	.attrs	= pt_timing_attr,
172};
173
174static const struct attribute_group *pt_attr_groups[] = {
175	&pt_cap_group,
176	&pt_format_group,
177	&pt_timing_group,
178	NULL,
179};
180
181static int __init pt_pmu_hw_init(void)
182{
183	struct dev_ext_attribute *de_attrs;
184	struct attribute **attrs;
185	size_t size;
186	u64 reg;
187	int ret;
188	long i;
189
190	rdmsrl(MSR_PLATFORM_INFO, reg);
191	pt_pmu.max_nonturbo_ratio = (reg & 0xff00) >> 8;
192
193	/*
194	 * if available, read in TSC to core crystal clock ratio,
195	 * otherwise, zero for numerator stands for "not enumerated"
196	 * as per SDM
197	 */
198	if (boot_cpu_data.cpuid_level >= CPUID_TSC_LEAF) {
199		u32 eax, ebx, ecx, edx;
200
201		cpuid(CPUID_TSC_LEAF, &eax, &ebx, &ecx, &edx);
202
203		pt_pmu.tsc_art_num = ebx;
204		pt_pmu.tsc_art_den = eax;
205	}
206
207	/* model-specific quirks */
208	switch (boot_cpu_data.x86_model) {
209	case INTEL_FAM6_BROADWELL:
210	case INTEL_FAM6_BROADWELL_D:
211	case INTEL_FAM6_BROADWELL_G:
212	case INTEL_FAM6_BROADWELL_X:
213		/* not setting BRANCH_EN will #GP, erratum BDM106 */
214		pt_pmu.branch_en_always_on = true;
215		break;
216	default:
217		break;
218	}
219
220	if (boot_cpu_has(X86_FEATURE_VMX)) {
221		/*
222		 * Intel SDM, 36.5 "Tracing post-VMXON" says that
223		 * "IA32_VMX_MISC[bit 14]" being 1 means PT can trace
224		 * post-VMXON.
225		 */
226		rdmsrl(MSR_IA32_VMX_MISC, reg);
227		if (reg & BIT(14))
228			pt_pmu.vmx = true;
229	}
230
231	for (i = 0; i < PT_CPUID_LEAVES; i++) {
232		cpuid_count(20, i,
233			    &pt_pmu.caps[CPUID_EAX + i*PT_CPUID_REGS_NUM],
234			    &pt_pmu.caps[CPUID_EBX + i*PT_CPUID_REGS_NUM],
235			    &pt_pmu.caps[CPUID_ECX + i*PT_CPUID_REGS_NUM],
236			    &pt_pmu.caps[CPUID_EDX + i*PT_CPUID_REGS_NUM]);
237	}
238
239	ret = -ENOMEM;
240	size = sizeof(struct attribute *) * (ARRAY_SIZE(pt_caps)+1);
241	attrs = kzalloc(size, GFP_KERNEL);
242	if (!attrs)
243		goto fail;
244
245	size = sizeof(struct dev_ext_attribute) * (ARRAY_SIZE(pt_caps)+1);
246	de_attrs = kzalloc(size, GFP_KERNEL);
247	if (!de_attrs)
248		goto fail;
249
250	for (i = 0; i < ARRAY_SIZE(pt_caps); i++) {
251		struct dev_ext_attribute *de_attr = de_attrs + i;
252
253		de_attr->attr.attr.name = pt_caps[i].name;
254
255		sysfs_attr_init(&de_attr->attr.attr);
256
257		de_attr->attr.attr.mode		= S_IRUGO;
258		de_attr->attr.show		= pt_cap_show;
259		de_attr->var			= (void *)i;
260
261		attrs[i] = &de_attr->attr.attr;
262	}
263
264	pt_cap_group.attrs = attrs;
265
266	return 0;
267
268fail:
269	kfree(attrs);
270
271	return ret;
272}
273
274#define RTIT_CTL_CYC_PSB (RTIT_CTL_CYCLEACC	| \
275			  RTIT_CTL_CYC_THRESH	| \
276			  RTIT_CTL_PSB_FREQ)
277
278#define RTIT_CTL_MTC	(RTIT_CTL_MTC_EN	| \
279			 RTIT_CTL_MTC_RANGE)
280
281#define RTIT_CTL_PTW	(RTIT_CTL_PTW_EN	| \
282			 RTIT_CTL_FUP_ON_PTW)
283
284/*
285 * Bit 0 (TraceEn) in the attr.config is meaningless as the
286 * corresponding bit in the RTIT_CTL can only be controlled
287 * by the driver; therefore, repurpose it to mean: pass
288 * through the bit that was previously assumed to be always
289 * on for PT, thereby allowing the user to *not* set it if
290 * they so wish. See also pt_event_valid() and pt_config().
291 */
292#define RTIT_CTL_PASSTHROUGH RTIT_CTL_TRACEEN
293
294#define PT_CONFIG_MASK (RTIT_CTL_TRACEEN	| \
295			RTIT_CTL_TSC_EN		| \
296			RTIT_CTL_DISRETC	| \
297			RTIT_CTL_BRANCH_EN	| \
298			RTIT_CTL_CYC_PSB	| \
299			RTIT_CTL_MTC		| \
300			RTIT_CTL_PWR_EVT_EN	| \
301			RTIT_CTL_FUP_ON_PTW	| \
302			RTIT_CTL_PTW_EN)
303
304static bool pt_event_valid(struct perf_event *event)
305{
306	u64 config = event->attr.config;
307	u64 allowed, requested;
308
309	if ((config & PT_CONFIG_MASK) != config)
310		return false;
311
312	if (config & RTIT_CTL_CYC_PSB) {
313		if (!intel_pt_validate_hw_cap(PT_CAP_psb_cyc))
314			return false;
315
316		allowed = intel_pt_validate_hw_cap(PT_CAP_psb_periods);
317		requested = (config & RTIT_CTL_PSB_FREQ) >>
318			RTIT_CTL_PSB_FREQ_OFFSET;
319		if (requested && (!(allowed & BIT(requested))))
320			return false;
321
322		allowed = intel_pt_validate_hw_cap(PT_CAP_cycle_thresholds);
323		requested = (config & RTIT_CTL_CYC_THRESH) >>
324			RTIT_CTL_CYC_THRESH_OFFSET;
325		if (requested && (!(allowed & BIT(requested))))
326			return false;
327	}
328
329	if (config & RTIT_CTL_MTC) {
330		/*
331		 * In the unlikely case that CPUID lists valid mtc periods,
332		 * but not the mtc capability, drop out here.
333		 *
334		 * Spec says that setting mtc period bits while mtc bit in
335		 * CPUID is 0 will #GP, so better safe than sorry.
336		 */
337		if (!intel_pt_validate_hw_cap(PT_CAP_mtc))
338			return false;
339
340		allowed = intel_pt_validate_hw_cap(PT_CAP_mtc_periods);
341		if (!allowed)
342			return false;
343
344		requested = (config & RTIT_CTL_MTC_RANGE) >>
345			RTIT_CTL_MTC_RANGE_OFFSET;
346
347		if (!(allowed & BIT(requested)))
348			return false;
349	}
350
351	if (config & RTIT_CTL_PWR_EVT_EN &&
352	    !intel_pt_validate_hw_cap(PT_CAP_power_event_trace))
353		return false;
354
355	if (config & RTIT_CTL_PTW) {
356		if (!intel_pt_validate_hw_cap(PT_CAP_ptwrite))
357			return false;
358
359		/* FUPonPTW without PTW doesn't make sense */
360		if ((config & RTIT_CTL_FUP_ON_PTW) &&
361		    !(config & RTIT_CTL_PTW_EN))
362			return false;
363	}
364
365	/*
366	 * Setting bit 0 (TraceEn in RTIT_CTL MSR) in the attr.config
367	 * clears the assomption that BranchEn must always be enabled,
368	 * as was the case with the first implementation of PT.
369	 * If this bit is not set, the legacy behavior is preserved
370	 * for compatibility with the older userspace.
371	 *
372	 * Re-using bit 0 for this purpose is fine because it is never
373	 * directly set by the user; previous attempts at setting it in
374	 * the attr.config resulted in -EINVAL.
375	 */
376	if (config & RTIT_CTL_PASSTHROUGH) {
377		/*
378		 * Disallow not setting BRANCH_EN where BRANCH_EN is
379		 * always required.
380		 */
381		if (pt_pmu.branch_en_always_on &&
382		    !(config & RTIT_CTL_BRANCH_EN))
383			return false;
384	} else {
385		/*
386		 * Disallow BRANCH_EN without the PASSTHROUGH.
387		 */
388		if (config & RTIT_CTL_BRANCH_EN)
389			return false;
390	}
391
392	return true;
393}
394
395/*
396 * PT configuration helpers
397 * These all are cpu affine and operate on a local PT
398 */
399
400static void pt_config_start(struct perf_event *event)
401{
402	struct pt *pt = this_cpu_ptr(&pt_ctx);
403	u64 ctl = event->hw.config;
404
405	ctl |= RTIT_CTL_TRACEEN;
406	if (READ_ONCE(pt->vmx_on))
407		perf_aux_output_flag(&pt->handle, PERF_AUX_FLAG_PARTIAL);
408	else
409		wrmsrl(MSR_IA32_RTIT_CTL, ctl);
410
411	WRITE_ONCE(event->hw.config, ctl);
412}
413
414/* Address ranges and their corresponding msr configuration registers */
415static const struct pt_address_range {
416	unsigned long	msr_a;
417	unsigned long	msr_b;
418	unsigned int	reg_off;
419} pt_address_ranges[] = {
420	{
421		.msr_a	 = MSR_IA32_RTIT_ADDR0_A,
422		.msr_b	 = MSR_IA32_RTIT_ADDR0_B,
423		.reg_off = RTIT_CTL_ADDR0_OFFSET,
424	},
425	{
426		.msr_a	 = MSR_IA32_RTIT_ADDR1_A,
427		.msr_b	 = MSR_IA32_RTIT_ADDR1_B,
428		.reg_off = RTIT_CTL_ADDR1_OFFSET,
429	},
430	{
431		.msr_a	 = MSR_IA32_RTIT_ADDR2_A,
432		.msr_b	 = MSR_IA32_RTIT_ADDR2_B,
433		.reg_off = RTIT_CTL_ADDR2_OFFSET,
434	},
435	{
436		.msr_a	 = MSR_IA32_RTIT_ADDR3_A,
437		.msr_b	 = MSR_IA32_RTIT_ADDR3_B,
438		.reg_off = RTIT_CTL_ADDR3_OFFSET,
439	}
440};
441
442static u64 pt_config_filters(struct perf_event *event)
443{
444	struct pt_filters *filters = event->hw.addr_filters;
445	struct pt *pt = this_cpu_ptr(&pt_ctx);
446	unsigned int range = 0;
447	u64 rtit_ctl = 0;
448
449	if (!filters)
450		return 0;
451
452	perf_event_addr_filters_sync(event);
453
454	for (range = 0; range < filters->nr_filters; range++) {
455		struct pt_filter *filter = &filters->filter[range];
456
457		/*
458		 * Note, if the range has zero start/end addresses due
459		 * to its dynamic object not being loaded yet, we just
460		 * go ahead and program zeroed range, which will simply
461		 * produce no data. Note^2: if executable code at 0x0
462		 * is a concern, we can set up an "invalid" configuration
463		 * such as msr_b < msr_a.
464		 */
465
466		/* avoid redundant msr writes */
467		if (pt->filters.filter[range].msr_a != filter->msr_a) {
468			wrmsrl(pt_address_ranges[range].msr_a, filter->msr_a);
469			pt->filters.filter[range].msr_a = filter->msr_a;
470		}
471
472		if (pt->filters.filter[range].msr_b != filter->msr_b) {
473			wrmsrl(pt_address_ranges[range].msr_b, filter->msr_b);
474			pt->filters.filter[range].msr_b = filter->msr_b;
475		}
476
477		rtit_ctl |= (u64)filter->config << pt_address_ranges[range].reg_off;
478	}
479
480	return rtit_ctl;
481}
482
483static void pt_config(struct perf_event *event)
484{
485	struct pt *pt = this_cpu_ptr(&pt_ctx);
486	struct pt_buffer *buf = perf_get_aux(&pt->handle);
487	u64 reg;
488
489	/* First round: clear STATUS, in particular the PSB byte counter. */
490	if (!event->hw.config) {
491		perf_event_itrace_started(event);
492		wrmsrl(MSR_IA32_RTIT_STATUS, 0);
493	}
494
495	reg = pt_config_filters(event);
496	reg |= RTIT_CTL_TRACEEN;
497	if (!buf->single)
498		reg |= RTIT_CTL_TOPA;
499
500	/*
501	 * Previously, we had BRANCH_EN on by default, but now that PT has
502	 * grown features outside of branch tracing, it is useful to allow
503	 * the user to disable it. Setting bit 0 in the event's attr.config
504	 * allows BRANCH_EN to pass through instead of being always on. See
505	 * also the comment in pt_event_valid().
506	 */
507	if (event->attr.config & BIT(0)) {
508		reg |= event->attr.config & RTIT_CTL_BRANCH_EN;
509	} else {
510		reg |= RTIT_CTL_BRANCH_EN;
511	}
512
513	if (!event->attr.exclude_kernel)
514		reg |= RTIT_CTL_OS;
515	if (!event->attr.exclude_user)
516		reg |= RTIT_CTL_USR;
517
518	reg |= (event->attr.config & PT_CONFIG_MASK);
519
520	event->hw.config = reg;
521	pt_config_start(event);
522}
523
524static void pt_config_stop(struct perf_event *event)
525{
526	struct pt *pt = this_cpu_ptr(&pt_ctx);
527	u64 ctl = READ_ONCE(event->hw.config);
528
529	/* may be already stopped by a PMI */
530	if (!(ctl & RTIT_CTL_TRACEEN))
531		return;
532
533	ctl &= ~RTIT_CTL_TRACEEN;
534	if (!READ_ONCE(pt->vmx_on))
535		wrmsrl(MSR_IA32_RTIT_CTL, ctl);
536
537	WRITE_ONCE(event->hw.config, ctl);
538
539	/*
540	 * A wrmsr that disables trace generation serializes other PT
541	 * registers and causes all data packets to be written to memory,
542	 * but a fence is required for the data to become globally visible.
543	 *
544	 * The below WMB, separating data store and aux_head store matches
545	 * the consumer's RMB that separates aux_head load and data load.
546	 */
547	wmb();
548}
549
550/**
551 * struct topa - ToPA metadata
552 * @list:	linkage to struct pt_buffer's list of tables
553 * @offset:	offset of the first entry in this table in the buffer
554 * @size:	total size of all entries in this table
555 * @last:	index of the last initialized entry in this table
556 * @z_count:	how many times the first entry repeats
557 */
558struct topa {
559	struct list_head	list;
560	u64			offset;
561	size_t			size;
562	int			last;
563	unsigned int		z_count;
564};
565
566/*
567 * Keep ToPA table-related metadata on the same page as the actual table,
568 * taking up a few words from the top
569 */
570
571#define TENTS_PER_PAGE	\
572	((PAGE_SIZE - sizeof(struct topa)) / sizeof(struct topa_entry))
573
574/**
575 * struct topa_page - page-sized ToPA table with metadata at the top
576 * @table:	actual ToPA table entries, as understood by PT hardware
577 * @topa:	metadata
578 */
579struct topa_page {
580	struct topa_entry	table[TENTS_PER_PAGE];
581	struct topa		topa;
582};
583
584static inline struct topa_page *topa_to_page(struct topa *topa)
585{
586	return container_of(topa, struct topa_page, topa);
587}
588
589static inline struct topa_page *topa_entry_to_page(struct topa_entry *te)
590{
591	return (struct topa_page *)((unsigned long)te & PAGE_MASK);
592}
593
594static inline phys_addr_t topa_pfn(struct topa *topa)
595{
596	return PFN_DOWN(virt_to_phys(topa_to_page(topa)));
597}
598
599/* make -1 stand for the last table entry */
600#define TOPA_ENTRY(t, i)				\
601	((i) == -1					\
602		? &topa_to_page(t)->table[(t)->last]	\
603		: &topa_to_page(t)->table[(i)])
604#define TOPA_ENTRY_SIZE(t, i) (sizes(TOPA_ENTRY((t), (i))->size))
605#define TOPA_ENTRY_PAGES(t, i) (1 << TOPA_ENTRY((t), (i))->size)
606
607static void pt_config_buffer(struct pt_buffer *buf)
608{
609	struct pt *pt = this_cpu_ptr(&pt_ctx);
610	u64 reg, mask;
611	void *base;
612
613	if (buf->single) {
614		base = buf->data_pages[0];
615		mask = (buf->nr_pages * PAGE_SIZE - 1) >> 7;
616	} else {
617		base = topa_to_page(buf->cur)->table;
618		mask = (u64)buf->cur_idx;
619	}
620
621	reg = virt_to_phys(base);
622	if (pt->output_base != reg) {
623		pt->output_base = reg;
624		wrmsrl(MSR_IA32_RTIT_OUTPUT_BASE, reg);
625	}
626
627	reg = 0x7f | (mask << 7) | ((u64)buf->output_off << 32);
628	if (pt->output_mask != reg) {
629		pt->output_mask = reg;
630		wrmsrl(MSR_IA32_RTIT_OUTPUT_MASK, reg);
631	}
632}
633
634/**
635 * topa_alloc() - allocate page-sized ToPA table
636 * @cpu:	CPU on which to allocate.
637 * @gfp:	Allocation flags.
638 *
639 * Return:	On success, return the pointer to ToPA table page.
640 */
641static struct topa *topa_alloc(int cpu, gfp_t gfp)
642{
643	int node = cpu_to_node(cpu);
644	struct topa_page *tp;
645	struct page *p;
646
647	p = alloc_pages_node(node, gfp | __GFP_ZERO, 0);
648	if (!p)
649		return NULL;
650
651	tp = page_address(p);
652	tp->topa.last = 0;
653
654	/*
655	 * In case of singe-entry ToPA, always put the self-referencing END
656	 * link as the 2nd entry in the table
657	 */
658	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries)) {
659		TOPA_ENTRY(&tp->topa, 1)->base = page_to_phys(p) >> TOPA_SHIFT;
660		TOPA_ENTRY(&tp->topa, 1)->end = 1;
661	}
662
663	return &tp->topa;
664}
665
666/**
667 * topa_free() - free a page-sized ToPA table
668 * @topa:	Table to deallocate.
669 */
670static void topa_free(struct topa *topa)
671{
672	free_page((unsigned long)topa);
673}
674
675/**
676 * topa_insert_table() - insert a ToPA table into a buffer
677 * @buf:	 PT buffer that's being extended.
678 * @topa:	 New topa table to be inserted.
679 *
680 * If it's the first table in this buffer, set up buffer's pointers
681 * accordingly; otherwise, add a END=1 link entry to @topa to the current
682 * "last" table and adjust the last table pointer to @topa.
683 */
684static void topa_insert_table(struct pt_buffer *buf, struct topa *topa)
685{
686	struct topa *last = buf->last;
687
688	list_add_tail(&topa->list, &buf->tables);
689
690	if (!buf->first) {
691		buf->first = buf->last = buf->cur = topa;
692		return;
693	}
694
695	topa->offset = last->offset + last->size;
696	buf->last = topa;
697
698	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
699		return;
700
701	BUG_ON(last->last != TENTS_PER_PAGE - 1);
702
703	TOPA_ENTRY(last, -1)->base = topa_pfn(topa);
704	TOPA_ENTRY(last, -1)->end = 1;
705}
706
707/**
708 * topa_table_full() - check if a ToPA table is filled up
709 * @topa:	ToPA table.
710 */
711static bool topa_table_full(struct topa *topa)
712{
713	/* single-entry ToPA is a special case */
714	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
715		return !!topa->last;
716
717	return topa->last == TENTS_PER_PAGE - 1;
718}
719
720/**
721 * topa_insert_pages() - create a list of ToPA tables
722 * @buf:	PT buffer being initialized.
723 * @gfp:	Allocation flags.
724 *
725 * This initializes a list of ToPA tables with entries from
726 * the data_pages provided by rb_alloc_aux().
727 *
728 * Return:	0 on success or error code.
729 */
730static int topa_insert_pages(struct pt_buffer *buf, int cpu, gfp_t gfp)
731{
732	struct topa *topa = buf->last;
733	int order = 0;
734	struct page *p;
735
736	p = virt_to_page(buf->data_pages[buf->nr_pages]);
737	if (PagePrivate(p))
738		order = page_private(p);
739
740	if (topa_table_full(topa)) {
741		topa = topa_alloc(cpu, gfp);
742		if (!topa)
743			return -ENOMEM;
744
745		topa_insert_table(buf, topa);
746	}
747
748	if (topa->z_count == topa->last - 1) {
749		if (order == TOPA_ENTRY(topa, topa->last - 1)->size)
750			topa->z_count++;
751	}
752
753	TOPA_ENTRY(topa, -1)->base = page_to_phys(p) >> TOPA_SHIFT;
754	TOPA_ENTRY(topa, -1)->size = order;
755	if (!buf->snapshot &&
756	    !intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries)) {
757		TOPA_ENTRY(topa, -1)->intr = 1;
758		TOPA_ENTRY(topa, -1)->stop = 1;
759	}
760
761	topa->last++;
762	topa->size += sizes(order);
763
764	buf->nr_pages += 1ul << order;
765
766	return 0;
767}
768
769/**
770 * pt_topa_dump() - print ToPA tables and their entries
771 * @buf:	PT buffer.
772 */
773static void pt_topa_dump(struct pt_buffer *buf)
774{
775	struct topa *topa;
776
777	list_for_each_entry(topa, &buf->tables, list) {
778		struct topa_page *tp = topa_to_page(topa);
779		int i;
780
781		pr_debug("# table @%p, off %llx size %zx\n", tp->table,
782			 topa->offset, topa->size);
783		for (i = 0; i < TENTS_PER_PAGE; i++) {
784			pr_debug("# entry @%p (%lx sz %u %c%c%c) raw=%16llx\n",
785				 &tp->table[i],
786				 (unsigned long)tp->table[i].base << TOPA_SHIFT,
787				 sizes(tp->table[i].size),
788				 tp->table[i].end ?  'E' : ' ',
789				 tp->table[i].intr ? 'I' : ' ',
790				 tp->table[i].stop ? 'S' : ' ',
791				 *(u64 *)&tp->table[i]);
792			if ((intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) &&
793			     tp->table[i].stop) ||
794			    tp->table[i].end)
795				break;
796			if (!i && topa->z_count)
797				i += topa->z_count;
798		}
799	}
800}
801
802/**
803 * pt_buffer_advance() - advance to the next output region
804 * @buf:	PT buffer.
805 *
806 * Advance the current pointers in the buffer to the next ToPA entry.
807 */
808static void pt_buffer_advance(struct pt_buffer *buf)
809{
810	buf->output_off = 0;
811	buf->cur_idx++;
812
813	if (buf->cur_idx == buf->cur->last) {
814		if (buf->cur == buf->last)
815			buf->cur = buf->first;
816		else
817			buf->cur = list_entry(buf->cur->list.next, struct topa,
818					      list);
819		buf->cur_idx = 0;
820	}
821}
822
823/**
824 * pt_update_head() - calculate current offsets and sizes
825 * @pt:		Per-cpu pt context.
826 *
827 * Update buffer's current write pointer position and data size.
828 */
829static void pt_update_head(struct pt *pt)
830{
831	struct pt_buffer *buf = perf_get_aux(&pt->handle);
832	u64 topa_idx, base, old;
833
834	if (buf->single) {
835		local_set(&buf->data_size, buf->output_off);
836		return;
837	}
838
839	/* offset of the first region in this table from the beginning of buf */
840	base = buf->cur->offset + buf->output_off;
841
842	/* offset of the current output region within this table */
843	for (topa_idx = 0; topa_idx < buf->cur_idx; topa_idx++)
844		base += TOPA_ENTRY_SIZE(buf->cur, topa_idx);
845
846	if (buf->snapshot) {
847		local_set(&buf->data_size, base);
848	} else {
849		old = (local64_xchg(&buf->head, base) &
850		       ((buf->nr_pages << PAGE_SHIFT) - 1));
851		if (base < old)
852			base += buf->nr_pages << PAGE_SHIFT;
853
854		local_add(base - old, &buf->data_size);
855	}
856}
857
858/**
859 * pt_buffer_region() - obtain current output region's address
860 * @buf:	PT buffer.
861 */
862static void *pt_buffer_region(struct pt_buffer *buf)
863{
864	return phys_to_virt(TOPA_ENTRY(buf->cur, buf->cur_idx)->base << TOPA_SHIFT);
865}
866
867/**
868 * pt_buffer_region_size() - obtain current output region's size
869 * @buf:	PT buffer.
870 */
871static size_t pt_buffer_region_size(struct pt_buffer *buf)
872{
873	return TOPA_ENTRY_SIZE(buf->cur, buf->cur_idx);
874}
875
876/**
877 * pt_handle_status() - take care of possible status conditions
878 * @pt:		Per-cpu pt context.
879 */
880static void pt_handle_status(struct pt *pt)
881{
882	struct pt_buffer *buf = perf_get_aux(&pt->handle);
883	int advance = 0;
884	u64 status;
885
886	rdmsrl(MSR_IA32_RTIT_STATUS, status);
887
888	if (status & RTIT_STATUS_ERROR) {
889		pr_err_ratelimited("ToPA ERROR encountered, trying to recover\n");
890		pt_topa_dump(buf);
891		status &= ~RTIT_STATUS_ERROR;
892	}
893
894	if (status & RTIT_STATUS_STOPPED) {
895		status &= ~RTIT_STATUS_STOPPED;
896
897		/*
898		 * On systems that only do single-entry ToPA, hitting STOP
899		 * means we are already losing data; need to let the decoder
900		 * know.
901		 */
902		if (!buf->single &&
903		    (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) ||
904		     buf->output_off == pt_buffer_region_size(buf))) {
905			perf_aux_output_flag(&pt->handle,
906			                     PERF_AUX_FLAG_TRUNCATED);
907			advance++;
908		}
909	}
910
911	/*
912	 * Also on single-entry ToPA implementations, interrupt will come
913	 * before the output reaches its output region's boundary.
914	 */
915	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) &&
916	    !buf->snapshot &&
917	    pt_buffer_region_size(buf) - buf->output_off <= TOPA_PMI_MARGIN) {
918		void *head = pt_buffer_region(buf);
919
920		/* everything within this margin needs to be zeroed out */
921		memset(head + buf->output_off, 0,
922		       pt_buffer_region_size(buf) -
923		       buf->output_off);
924		advance++;
925	}
926
927	if (advance)
928		pt_buffer_advance(buf);
929
930	wrmsrl(MSR_IA32_RTIT_STATUS, status);
931}
932
933/**
934 * pt_read_offset() - translate registers into buffer pointers
935 * @buf:	PT buffer.
936 *
937 * Set buffer's output pointers from MSR values.
938 */
939static void pt_read_offset(struct pt_buffer *buf)
940{
941	struct pt *pt = this_cpu_ptr(&pt_ctx);
942	struct topa_page *tp;
943
944	if (!buf->single) {
945		rdmsrl(MSR_IA32_RTIT_OUTPUT_BASE, pt->output_base);
946		tp = phys_to_virt(pt->output_base);
947		buf->cur = &tp->topa;
948	}
949
950	rdmsrl(MSR_IA32_RTIT_OUTPUT_MASK, pt->output_mask);
951	/* offset within current output region */
952	buf->output_off = pt->output_mask >> 32;
953	/* index of current output region within this table */
954	if (!buf->single)
955		buf->cur_idx = (pt->output_mask & 0xffffff80) >> 7;
956}
957
958static struct topa_entry *
959pt_topa_entry_for_page(struct pt_buffer *buf, unsigned int pg)
960{
961	struct topa_page *tp;
962	struct topa *topa;
963	unsigned int idx, cur_pg = 0, z_pg = 0, start_idx = 0;
964
965	/*
966	 * Indicates a bug in the caller.
967	 */
968	if (WARN_ON_ONCE(pg >= buf->nr_pages))
969		return NULL;
970
971	/*
972	 * First, find the ToPA table where @pg fits. With high
973	 * order allocations, there shouldn't be many of these.
974	 */
975	list_for_each_entry(topa, &buf->tables, list) {
976		if (topa->offset + topa->size > pg << PAGE_SHIFT)
977			goto found;
978	}
979
980	/*
981	 * Hitting this means we have a problem in the ToPA
982	 * allocation code.
983	 */
984	WARN_ON_ONCE(1);
985
986	return NULL;
987
988found:
989	/*
990	 * Indicates a problem in the ToPA allocation code.
991	 */
992	if (WARN_ON_ONCE(topa->last == -1))
993		return NULL;
994
995	tp = topa_to_page(topa);
996	cur_pg = PFN_DOWN(topa->offset);
997	if (topa->z_count) {
998		z_pg = TOPA_ENTRY_PAGES(topa, 0) * (topa->z_count + 1);
999		start_idx = topa->z_count + 1;
1000	}
1001
1002	/*
1003	 * Multiple entries at the beginning of the table have the same size,
1004	 * ideally all of them; if @pg falls there, the search is done.
1005	 */
1006	if (pg >= cur_pg && pg < cur_pg + z_pg) {
1007		idx = (pg - cur_pg) / TOPA_ENTRY_PAGES(topa, 0);
1008		return &tp->table[idx];
1009	}
1010
1011	/*
1012	 * Otherwise, slow path: iterate through the remaining entries.
1013	 */
1014	for (idx = start_idx, cur_pg += z_pg; idx < topa->last; idx++) {
1015		if (cur_pg + TOPA_ENTRY_PAGES(topa, idx) > pg)
1016			return &tp->table[idx];
1017
1018		cur_pg += TOPA_ENTRY_PAGES(topa, idx);
1019	}
1020
1021	/*
1022	 * Means we couldn't find a ToPA entry in the table that does match.
1023	 */
1024	WARN_ON_ONCE(1);
1025
1026	return NULL;
1027}
1028
1029static struct topa_entry *
1030pt_topa_prev_entry(struct pt_buffer *buf, struct topa_entry *te)
1031{
1032	unsigned long table = (unsigned long)te & ~(PAGE_SIZE - 1);
1033	struct topa_page *tp;
1034	struct topa *topa;
1035
1036	tp = (struct topa_page *)table;
1037	if (tp->table != te)
1038		return --te;
1039
1040	topa = &tp->topa;
1041	if (topa == buf->first)
1042		topa = buf->last;
1043	else
1044		topa = list_prev_entry(topa, list);
1045
1046	tp = topa_to_page(topa);
1047
1048	return &tp->table[topa->last - 1];
1049}
1050
1051/**
1052 * pt_buffer_reset_markers() - place interrupt and stop bits in the buffer
1053 * @buf:	PT buffer.
1054 * @handle:	Current output handle.
1055 *
1056 * Place INT and STOP marks to prevent overwriting old data that the consumer
1057 * hasn't yet collected and waking up the consumer after a certain fraction of
1058 * the buffer has filled up. Only needed and sensible for non-snapshot counters.
1059 *
1060 * This obviously relies on buf::head to figure out buffer markers, so it has
1061 * to be called after pt_buffer_reset_offsets() and before the hardware tracing
1062 * is enabled.
1063 */
1064static int pt_buffer_reset_markers(struct pt_buffer *buf,
1065				   struct perf_output_handle *handle)
1066
1067{
1068	unsigned long head = local64_read(&buf->head);
1069	unsigned long idx, npages, wakeup;
1070
1071	if (buf->single)
1072		return 0;
1073
1074	/* can't stop in the middle of an output region */
1075	if (buf->output_off + handle->size + 1 < pt_buffer_region_size(buf)) {
1076		perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED);
1077		return -EINVAL;
1078	}
1079
1080
1081	/* single entry ToPA is handled by marking all regions STOP=1 INT=1 */
1082	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
1083		return 0;
1084
1085	/* clear STOP and INT from current entry */
1086	if (buf->stop_te) {
1087		buf->stop_te->stop = 0;
1088		buf->stop_te->intr = 0;
1089	}
1090
1091	if (buf->intr_te)
1092		buf->intr_te->intr = 0;
1093
1094	/* how many pages till the STOP marker */
1095	npages = handle->size >> PAGE_SHIFT;
1096
1097	/* if it's on a page boundary, fill up one more page */
1098	if (!offset_in_page(head + handle->size + 1))
1099		npages++;
1100
1101	idx = (head >> PAGE_SHIFT) + npages;
1102	idx &= buf->nr_pages - 1;
1103
1104	if (idx != buf->stop_pos) {
1105		buf->stop_pos = idx;
1106		buf->stop_te = pt_topa_entry_for_page(buf, idx);
1107		buf->stop_te = pt_topa_prev_entry(buf, buf->stop_te);
1108	}
1109
1110	wakeup = handle->wakeup >> PAGE_SHIFT;
1111
1112	/* in the worst case, wake up the consumer one page before hard stop */
1113	idx = (head >> PAGE_SHIFT) + npages - 1;
1114	if (idx > wakeup)
1115		idx = wakeup;
1116
1117	idx &= buf->nr_pages - 1;
1118	if (idx != buf->intr_pos) {
1119		buf->intr_pos = idx;
1120		buf->intr_te = pt_topa_entry_for_page(buf, idx);
1121		buf->intr_te = pt_topa_prev_entry(buf, buf->intr_te);
1122	}
1123
1124	buf->stop_te->stop = 1;
1125	buf->stop_te->intr = 1;
1126	buf->intr_te->intr = 1;
1127
1128	return 0;
1129}
1130
1131/**
1132 * pt_buffer_reset_offsets() - adjust buffer's write pointers from aux_head
1133 * @buf:	PT buffer.
1134 * @head:	Write pointer (aux_head) from AUX buffer.
1135 *
1136 * Find the ToPA table and entry corresponding to given @head and set buffer's
1137 * "current" pointers accordingly. This is done after we have obtained the
1138 * current aux_head position from a successful call to perf_aux_output_begin()
1139 * to make sure the hardware is writing to the right place.
1140 *
1141 * This function modifies buf::{cur,cur_idx,output_off} that will be programmed
1142 * into PT msrs when the tracing is enabled and buf::head and buf::data_size,
1143 * which are used to determine INT and STOP markers' locations by a subsequent
1144 * call to pt_buffer_reset_markers().
1145 */
1146static void pt_buffer_reset_offsets(struct pt_buffer *buf, unsigned long head)
1147{
1148	struct topa_page *cur_tp;
1149	struct topa_entry *te;
1150	int pg;
1151
1152	if (buf->snapshot)
1153		head &= (buf->nr_pages << PAGE_SHIFT) - 1;
1154
1155	if (!buf->single) {
1156		pg = (head >> PAGE_SHIFT) & (buf->nr_pages - 1);
1157		te = pt_topa_entry_for_page(buf, pg);
1158
1159		cur_tp = topa_entry_to_page(te);
1160		buf->cur = &cur_tp->topa;
1161		buf->cur_idx = te - TOPA_ENTRY(buf->cur, 0);
1162		buf->output_off = head & (pt_buffer_region_size(buf) - 1);
1163	} else {
1164		buf->output_off = head;
1165	}
1166
1167	local64_set(&buf->head, head);
1168	local_set(&buf->data_size, 0);
1169}
1170
1171/**
1172 * pt_buffer_fini_topa() - deallocate ToPA structure of a buffer
1173 * @buf:	PT buffer.
1174 */
1175static void pt_buffer_fini_topa(struct pt_buffer *buf)
1176{
1177	struct topa *topa, *iter;
1178
1179	if (buf->single)
1180		return;
1181
1182	list_for_each_entry_safe(topa, iter, &buf->tables, list) {
1183		/*
1184		 * right now, this is in free_aux() path only, so
1185		 * no need to unlink this table from the list
1186		 */
1187		topa_free(topa);
1188	}
1189}
1190
1191/**
1192 * pt_buffer_init_topa() - initialize ToPA table for pt buffer
1193 * @buf:	PT buffer.
1194 * @size:	Total size of all regions within this ToPA.
1195 * @gfp:	Allocation flags.
1196 */
1197static int pt_buffer_init_topa(struct pt_buffer *buf, int cpu,
1198			       unsigned long nr_pages, gfp_t gfp)
1199{
1200	struct topa *topa;
1201	int err;
1202
1203	topa = topa_alloc(cpu, gfp);
1204	if (!topa)
1205		return -ENOMEM;
1206
1207	topa_insert_table(buf, topa);
1208
1209	while (buf->nr_pages < nr_pages) {
1210		err = topa_insert_pages(buf, cpu, gfp);
1211		if (err) {
1212			pt_buffer_fini_topa(buf);
1213			return -ENOMEM;
1214		}
1215	}
1216
1217	/* link last table to the first one, unless we're double buffering */
1218	if (intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries)) {
1219		TOPA_ENTRY(buf->last, -1)->base = topa_pfn(buf->first);
1220		TOPA_ENTRY(buf->last, -1)->end = 1;
1221	}
1222
1223	pt_topa_dump(buf);
1224	return 0;
1225}
1226
1227static int pt_buffer_try_single(struct pt_buffer *buf, int nr_pages)
1228{
1229	struct page *p = virt_to_page(buf->data_pages[0]);
1230	int ret = -ENOTSUPP, order = 0;
1231
1232	/*
1233	 * We can use single range output mode
1234	 * + in snapshot mode, where we don't need interrupts;
1235	 * + if the hardware supports it;
1236	 * + if the entire buffer is one contiguous allocation.
1237	 */
1238	if (!buf->snapshot)
1239		goto out;
1240
1241	if (!intel_pt_validate_hw_cap(PT_CAP_single_range_output))
1242		goto out;
1243
1244	if (PagePrivate(p))
1245		order = page_private(p);
1246
1247	if (1 << order != nr_pages)
1248		goto out;
1249
1250	/*
1251	 * Some processors cannot always support single range for more than
1252	 * 4KB - refer errata TGL052, ADL037 and RPL017. Future processors might
1253	 * also be affected, so for now rather than trying to keep track of
1254	 * which ones, just disable it for all.
1255	 */
1256	if (nr_pages > 1)
1257		goto out;
1258
1259	buf->single = true;
1260	buf->nr_pages = nr_pages;
1261	ret = 0;
1262out:
1263	return ret;
1264}
1265
1266/**
1267 * pt_buffer_setup_aux() - set up topa tables for a PT buffer
1268 * @cpu:	Cpu on which to allocate, -1 means current.
1269 * @pages:	Array of pointers to buffer pages passed from perf core.
1270 * @nr_pages:	Number of pages in the buffer.
1271 * @snapshot:	If this is a snapshot/overwrite counter.
1272 *
1273 * This is a pmu::setup_aux callback that sets up ToPA tables and all the
1274 * bookkeeping for an AUX buffer.
1275 *
1276 * Return:	Our private PT buffer structure.
1277 */
1278static void *
1279pt_buffer_setup_aux(struct perf_event *event, void **pages,
1280		    int nr_pages, bool snapshot)
1281{
1282	struct pt_buffer *buf;
1283	int node, ret, cpu = event->cpu;
1284
1285	if (!nr_pages)
1286		return NULL;
1287
1288	/*
1289	 * Only support AUX sampling in snapshot mode, where we don't
1290	 * generate NMIs.
1291	 */
1292	if (event->attr.aux_sample_size && !snapshot)
1293		return NULL;
1294
1295	if (cpu == -1)
1296		cpu = raw_smp_processor_id();
1297	node = cpu_to_node(cpu);
1298
1299	buf = kzalloc_node(sizeof(struct pt_buffer), GFP_KERNEL, node);
1300	if (!buf)
1301		return NULL;
1302
1303	buf->snapshot = snapshot;
1304	buf->data_pages = pages;
1305	buf->stop_pos = -1;
1306	buf->intr_pos = -1;
1307
1308	INIT_LIST_HEAD(&buf->tables);
1309
1310	ret = pt_buffer_try_single(buf, nr_pages);
1311	if (!ret)
1312		return buf;
1313
1314	ret = pt_buffer_init_topa(buf, cpu, nr_pages, GFP_KERNEL);
1315	if (ret) {
1316		kfree(buf);
1317		return NULL;
1318	}
1319
1320	return buf;
1321}
1322
1323/**
1324 * pt_buffer_free_aux() - perf AUX deallocation path callback
1325 * @data:	PT buffer.
1326 */
1327static void pt_buffer_free_aux(void *data)
1328{
1329	struct pt_buffer *buf = data;
1330
1331	pt_buffer_fini_topa(buf);
1332	kfree(buf);
1333}
1334
1335static int pt_addr_filters_init(struct perf_event *event)
1336{
1337	struct pt_filters *filters;
1338	int node = event->cpu == -1 ? -1 : cpu_to_node(event->cpu);
1339
1340	if (!intel_pt_validate_hw_cap(PT_CAP_num_address_ranges))
1341		return 0;
1342
1343	filters = kzalloc_node(sizeof(struct pt_filters), GFP_KERNEL, node);
1344	if (!filters)
1345		return -ENOMEM;
1346
1347	if (event->parent)
1348		memcpy(filters, event->parent->hw.addr_filters,
1349		       sizeof(*filters));
1350
1351	event->hw.addr_filters = filters;
1352
1353	return 0;
1354}
1355
1356static void pt_addr_filters_fini(struct perf_event *event)
1357{
1358	kfree(event->hw.addr_filters);
1359	event->hw.addr_filters = NULL;
1360}
1361
1362#ifdef CONFIG_X86_64
1363/* Clamp to a canonical address greater-than-or-equal-to the address given */
1364static u64 clamp_to_ge_canonical_addr(u64 vaddr, u8 vaddr_bits)
1365{
1366	return __is_canonical_address(vaddr, vaddr_bits) ?
1367	       vaddr :
1368	       -BIT_ULL(vaddr_bits - 1);
1369}
1370
1371/* Clamp to a canonical address less-than-or-equal-to the address given */
1372static u64 clamp_to_le_canonical_addr(u64 vaddr, u8 vaddr_bits)
1373{
1374	return __is_canonical_address(vaddr, vaddr_bits) ?
1375	       vaddr :
1376	       BIT_ULL(vaddr_bits - 1) - 1;
1377}
1378#else
1379#define clamp_to_ge_canonical_addr(x, y) (x)
1380#define clamp_to_le_canonical_addr(x, y) (x)
1381#endif
1382
1383static int pt_event_addr_filters_validate(struct list_head *filters)
1384{
1385	struct perf_addr_filter *filter;
1386	int range = 0;
1387
1388	list_for_each_entry(filter, filters, entry) {
1389		/*
1390		 * PT doesn't support single address triggers and
1391		 * 'start' filters.
1392		 */
1393		if (!filter->size ||
1394		    filter->action == PERF_ADDR_FILTER_ACTION_START)
1395			return -EOPNOTSUPP;
1396
1397		if (++range > intel_pt_validate_hw_cap(PT_CAP_num_address_ranges))
1398			return -EOPNOTSUPP;
1399	}
1400
1401	return 0;
1402}
1403
1404static void pt_event_addr_filters_sync(struct perf_event *event)
1405{
1406	struct perf_addr_filters_head *head = perf_event_addr_filters(event);
1407	unsigned long msr_a, msr_b;
1408	struct perf_addr_filter_range *fr = event->addr_filter_ranges;
1409	struct pt_filters *filters = event->hw.addr_filters;
1410	struct perf_addr_filter *filter;
1411	int range = 0;
1412
1413	if (!filters)
1414		return;
1415
1416	list_for_each_entry(filter, &head->list, entry) {
1417		if (filter->path.dentry && !fr[range].start) {
1418			msr_a = msr_b = 0;
1419		} else {
1420			unsigned long n = fr[range].size - 1;
1421			unsigned long a = fr[range].start;
1422			unsigned long b;
1423
1424			if (a > ULONG_MAX - n)
1425				b = ULONG_MAX;
1426			else
1427				b = a + n;
1428			/*
1429			 * Apply the offset. 64-bit addresses written to the
1430			 * MSRs must be canonical, but the range can encompass
1431			 * non-canonical addresses. Since software cannot
1432			 * execute at non-canonical addresses, adjusting to
1433			 * canonical addresses does not affect the result of the
1434			 * address filter.
1435			 */
1436			msr_a = clamp_to_ge_canonical_addr(a, boot_cpu_data.x86_virt_bits);
1437			msr_b = clamp_to_le_canonical_addr(b, boot_cpu_data.x86_virt_bits);
1438			if (msr_b < msr_a)
1439				msr_a = msr_b = 0;
1440		}
1441
1442		filters->filter[range].msr_a  = msr_a;
1443		filters->filter[range].msr_b  = msr_b;
1444		if (filter->action == PERF_ADDR_FILTER_ACTION_FILTER)
1445			filters->filter[range].config = 1;
1446		else
1447			filters->filter[range].config = 2;
1448		range++;
1449	}
1450
1451	filters->nr_filters = range;
1452}
1453
1454/**
1455 * intel_pt_interrupt() - PT PMI handler
1456 */
1457void intel_pt_interrupt(void)
1458{
1459	struct pt *pt = this_cpu_ptr(&pt_ctx);
1460	struct pt_buffer *buf;
1461	struct perf_event *event = pt->handle.event;
1462
1463	/*
1464	 * There may be a dangling PT bit in the interrupt status register
1465	 * after PT has been disabled by pt_event_stop(). Make sure we don't
1466	 * do anything (particularly, re-enable) for this event here.
1467	 */
1468	if (!READ_ONCE(pt->handle_nmi))
1469		return;
1470
1471	if (!event)
1472		return;
1473
1474	pt_config_stop(event);
1475
1476	buf = perf_get_aux(&pt->handle);
1477	if (!buf)
1478		return;
1479
1480	pt_read_offset(buf);
1481
1482	pt_handle_status(pt);
1483
1484	pt_update_head(pt);
1485
1486	perf_aux_output_end(&pt->handle, local_xchg(&buf->data_size, 0));
1487
1488	if (!event->hw.state) {
1489		int ret;
1490
1491		buf = perf_aux_output_begin(&pt->handle, event);
1492		if (!buf) {
1493			event->hw.state = PERF_HES_STOPPED;
1494			return;
1495		}
1496
1497		pt_buffer_reset_offsets(buf, pt->handle.head);
1498		/* snapshot counters don't use PMI, so it's safe */
1499		ret = pt_buffer_reset_markers(buf, &pt->handle);
1500		if (ret) {
1501			perf_aux_output_end(&pt->handle, 0);
1502			return;
1503		}
1504
1505		pt_config_buffer(buf);
1506		pt_config_start(event);
1507	}
1508}
1509
1510void intel_pt_handle_vmx(int on)
1511{
1512	struct pt *pt = this_cpu_ptr(&pt_ctx);
1513	struct perf_event *event;
1514	unsigned long flags;
1515
1516	/* PT plays nice with VMX, do nothing */
1517	if (pt_pmu.vmx)
1518		return;
1519
1520	/*
1521	 * VMXON will clear RTIT_CTL.TraceEn; we need to make
1522	 * sure to not try to set it while VMX is on. Disable
1523	 * interrupts to avoid racing with pmu callbacks;
1524	 * concurrent PMI should be handled fine.
1525	 */
1526	local_irq_save(flags);
1527	WRITE_ONCE(pt->vmx_on, on);
1528
1529	/*
1530	 * If an AUX transaction is in progress, it will contain
1531	 * gap(s), so flag it PARTIAL to inform the user.
1532	 */
1533	event = pt->handle.event;
1534	if (event)
1535		perf_aux_output_flag(&pt->handle,
1536		                     PERF_AUX_FLAG_PARTIAL);
1537
1538	/* Turn PTs back on */
1539	if (!on && event)
1540		wrmsrl(MSR_IA32_RTIT_CTL, event->hw.config);
1541
1542	local_irq_restore(flags);
1543}
1544EXPORT_SYMBOL_GPL(intel_pt_handle_vmx);
1545
1546/*
1547 * PMU callbacks
1548 */
1549
1550static void pt_event_start(struct perf_event *event, int mode)
1551{
1552	struct hw_perf_event *hwc = &event->hw;
1553	struct pt *pt = this_cpu_ptr(&pt_ctx);
1554	struct pt_buffer *buf;
1555
1556	buf = perf_aux_output_begin(&pt->handle, event);
1557	if (!buf)
1558		goto fail_stop;
1559
1560	pt_buffer_reset_offsets(buf, pt->handle.head);
1561	if (!buf->snapshot) {
1562		if (pt_buffer_reset_markers(buf, &pt->handle))
1563			goto fail_end_stop;
1564	}
1565
1566	WRITE_ONCE(pt->handle_nmi, 1);
1567	hwc->state = 0;
1568
1569	pt_config_buffer(buf);
1570	pt_config(event);
1571
1572	return;
1573
1574fail_end_stop:
1575	perf_aux_output_end(&pt->handle, 0);
1576fail_stop:
1577	hwc->state = PERF_HES_STOPPED;
1578}
1579
1580static void pt_event_stop(struct perf_event *event, int mode)
1581{
1582	struct pt *pt = this_cpu_ptr(&pt_ctx);
1583
1584	/*
1585	 * Protect against the PMI racing with disabling wrmsr,
1586	 * see comment in intel_pt_interrupt().
1587	 */
1588	WRITE_ONCE(pt->handle_nmi, 0);
1589
1590	pt_config_stop(event);
1591
1592	if (event->hw.state == PERF_HES_STOPPED)
1593		return;
1594
1595	event->hw.state = PERF_HES_STOPPED;
1596
1597	if (mode & PERF_EF_UPDATE) {
1598		struct pt_buffer *buf = perf_get_aux(&pt->handle);
1599
1600		if (!buf)
1601			return;
1602
1603		if (WARN_ON_ONCE(pt->handle.event != event))
1604			return;
1605
1606		pt_read_offset(buf);
1607
1608		pt_handle_status(pt);
1609
1610		pt_update_head(pt);
1611
1612		if (buf->snapshot)
1613			pt->handle.head =
1614				local_xchg(&buf->data_size,
1615					   buf->nr_pages << PAGE_SHIFT);
1616		perf_aux_output_end(&pt->handle, local_xchg(&buf->data_size, 0));
1617	}
1618}
1619
1620static long pt_event_snapshot_aux(struct perf_event *event,
1621				  struct perf_output_handle *handle,
1622				  unsigned long size)
1623{
1624	struct pt *pt = this_cpu_ptr(&pt_ctx);
1625	struct pt_buffer *buf = perf_get_aux(&pt->handle);
1626	unsigned long from = 0, to;
1627	long ret;
1628
1629	if (WARN_ON_ONCE(!buf))
1630		return 0;
1631
1632	/*
1633	 * Sampling is only allowed on snapshot events;
1634	 * see pt_buffer_setup_aux().
1635	 */
1636	if (WARN_ON_ONCE(!buf->snapshot))
1637		return 0;
1638
1639	/*
1640	 * Here, handle_nmi tells us if the tracing is on
1641	 */
1642	if (READ_ONCE(pt->handle_nmi))
1643		pt_config_stop(event);
1644
1645	pt_read_offset(buf);
1646	pt_update_head(pt);
1647
1648	to = local_read(&buf->data_size);
1649	if (to < size)
1650		from = buf->nr_pages << PAGE_SHIFT;
1651	from += to - size;
1652
1653	ret = perf_output_copy_aux(&pt->handle, handle, from, to);
1654
1655	/*
1656	 * If the tracing was on when we turned up, restart it.
1657	 * Compiler barrier not needed as we couldn't have been
1658	 * preempted by anything that touches pt->handle_nmi.
1659	 */
1660	if (pt->handle_nmi)
1661		pt_config_start(event);
1662
1663	return ret;
1664}
1665
1666static void pt_event_del(struct perf_event *event, int mode)
1667{
1668	pt_event_stop(event, PERF_EF_UPDATE);
1669}
1670
1671static int pt_event_add(struct perf_event *event, int mode)
1672{
1673	struct pt *pt = this_cpu_ptr(&pt_ctx);
1674	struct hw_perf_event *hwc = &event->hw;
1675	int ret = -EBUSY;
1676
1677	if (pt->handle.event)
1678		goto fail;
1679
1680	if (mode & PERF_EF_START) {
1681		pt_event_start(event, 0);
1682		ret = -EINVAL;
1683		if (hwc->state == PERF_HES_STOPPED)
1684			goto fail;
1685	} else {
1686		hwc->state = PERF_HES_STOPPED;
1687	}
1688
1689	ret = 0;
1690fail:
1691
1692	return ret;
1693}
1694
1695static void pt_event_read(struct perf_event *event)
1696{
1697}
1698
1699static void pt_event_destroy(struct perf_event *event)
1700{
1701	pt_addr_filters_fini(event);
1702	x86_del_exclusive(x86_lbr_exclusive_pt);
1703}
1704
1705static int pt_event_init(struct perf_event *event)
1706{
1707	if (event->attr.type != pt_pmu.pmu.type)
1708		return -ENOENT;
1709
1710	if (!pt_event_valid(event))
1711		return -EINVAL;
1712
1713	if (x86_add_exclusive(x86_lbr_exclusive_pt))
1714		return -EBUSY;
1715
1716	if (pt_addr_filters_init(event)) {
1717		x86_del_exclusive(x86_lbr_exclusive_pt);
1718		return -ENOMEM;
1719	}
1720
1721	event->destroy = pt_event_destroy;
1722
1723	return 0;
1724}
1725
1726void cpu_emergency_stop_pt(void)
1727{
1728	struct pt *pt = this_cpu_ptr(&pt_ctx);
1729
1730	if (pt->handle.event)
1731		pt_event_stop(pt->handle.event, PERF_EF_UPDATE);
1732}
1733
1734int is_intel_pt_event(struct perf_event *event)
1735{
1736	return event->pmu == &pt_pmu.pmu;
1737}
1738
1739static __init int pt_init(void)
1740{
1741	int ret, cpu, prior_warn = 0;
1742
1743	BUILD_BUG_ON(sizeof(struct topa) > PAGE_SIZE);
1744
1745	if (!boot_cpu_has(X86_FEATURE_INTEL_PT))
1746		return -ENODEV;
1747
1748	get_online_cpus();
1749	for_each_online_cpu(cpu) {
1750		u64 ctl;
1751
1752		ret = rdmsrl_safe_on_cpu(cpu, MSR_IA32_RTIT_CTL, &ctl);
1753		if (!ret && (ctl & RTIT_CTL_TRACEEN))
1754			prior_warn++;
1755	}
1756	put_online_cpus();
1757
1758	if (prior_warn) {
1759		x86_add_exclusive(x86_lbr_exclusive_pt);
1760		pr_warn("PT is enabled at boot time, doing nothing\n");
1761
1762		return -EBUSY;
1763	}
1764
1765	ret = pt_pmu_hw_init();
1766	if (ret)
1767		return ret;
1768
1769	if (!intel_pt_validate_hw_cap(PT_CAP_topa_output)) {
1770		pr_warn("ToPA output is not supported on this CPU\n");
1771		return -ENODEV;
1772	}
1773
1774	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
1775		pt_pmu.pmu.capabilities = PERF_PMU_CAP_AUX_NO_SG;
1776
1777	pt_pmu.pmu.capabilities	|= PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE;
1778	pt_pmu.pmu.attr_groups		 = pt_attr_groups;
1779	pt_pmu.pmu.task_ctx_nr		 = perf_sw_context;
1780	pt_pmu.pmu.event_init		 = pt_event_init;
1781	pt_pmu.pmu.add			 = pt_event_add;
1782	pt_pmu.pmu.del			 = pt_event_del;
1783	pt_pmu.pmu.start		 = pt_event_start;
1784	pt_pmu.pmu.stop			 = pt_event_stop;
1785	pt_pmu.pmu.snapshot_aux		 = pt_event_snapshot_aux;
1786	pt_pmu.pmu.read			 = pt_event_read;
1787	pt_pmu.pmu.setup_aux		 = pt_buffer_setup_aux;
1788	pt_pmu.pmu.free_aux		 = pt_buffer_free_aux;
1789	pt_pmu.pmu.addr_filters_sync     = pt_event_addr_filters_sync;
1790	pt_pmu.pmu.addr_filters_validate = pt_event_addr_filters_validate;
1791	pt_pmu.pmu.nr_addr_filters       =
1792		intel_pt_validate_hw_cap(PT_CAP_num_address_ranges);
1793
1794	ret = perf_pmu_register(&pt_pmu.pmu, "intel_pt", -1);
1795
1796	return ret;
1797}
1798arch_initcall(pt_init);
1799