xref: /openbmc/linux/arch/x86/events/intel/pt.c (revision e4781421e883340b796da5a724bda7226817990b)
1 /*
2  * Intel(R) Processor Trace PMU driver for perf
3  * Copyright (c) 2013-2014, Intel Corporation.
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms and conditions of the GNU General Public License,
7  * version 2, as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  *
14  * Intel PT is specified in the Intel Architecture Instruction Set Extensions
15  * Programming Reference:
16  * http://software.intel.com/en-us/intel-isa-extensions
17  */
18 
19 #undef DEBUG
20 
21 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
22 
23 #include <linux/types.h>
24 #include <linux/slab.h>
25 #include <linux/device.h>
26 
27 #include <asm/perf_event.h>
28 #include <asm/insn.h>
29 #include <asm/io.h>
30 #include <asm/intel_pt.h>
31 
32 #include "../perf_event.h"
33 #include "pt.h"
34 
35 static DEFINE_PER_CPU(struct pt, pt_ctx);
36 
37 static struct pt_pmu pt_pmu;
38 
39 /*
40  * Capabilities of Intel PT hardware, such as number of address bits or
41  * supported output schemes, are cached and exported to userspace as "caps"
42  * attribute group of pt pmu device
43  * (/sys/bus/event_source/devices/intel_pt/caps/) so that userspace can store
44  * relevant bits together with intel_pt traces.
45  *
46  * These are necessary for both trace decoding (payloads_lip, contains address
47  * width encoded in IP-related packets), and event configuration (bitmasks with
48  * permitted values for certain bit fields).
49  */
50 #define PT_CAP(_n, _l, _r, _m)						\
51 	[PT_CAP_ ## _n] = { .name = __stringify(_n), .leaf = _l,	\
52 			    .reg = _r, .mask = _m }
53 
54 static struct pt_cap_desc {
55 	const char	*name;
56 	u32		leaf;
57 	u8		reg;
58 	u32		mask;
59 } pt_caps[] = {
60 	PT_CAP(max_subleaf,		0, CPUID_EAX, 0xffffffff),
61 	PT_CAP(cr3_filtering,		0, CPUID_EBX, BIT(0)),
62 	PT_CAP(psb_cyc,			0, CPUID_EBX, BIT(1)),
63 	PT_CAP(ip_filtering,		0, CPUID_EBX, BIT(2)),
64 	PT_CAP(mtc,			0, CPUID_EBX, BIT(3)),
65 	PT_CAP(ptwrite,			0, CPUID_EBX, BIT(4)),
66 	PT_CAP(power_event_trace,	0, CPUID_EBX, BIT(5)),
67 	PT_CAP(topa_output,		0, CPUID_ECX, BIT(0)),
68 	PT_CAP(topa_multiple_entries,	0, CPUID_ECX, BIT(1)),
69 	PT_CAP(single_range_output,	0, CPUID_ECX, BIT(2)),
70 	PT_CAP(payloads_lip,		0, CPUID_ECX, BIT(31)),
71 	PT_CAP(num_address_ranges,	1, CPUID_EAX, 0x3),
72 	PT_CAP(mtc_periods,		1, CPUID_EAX, 0xffff0000),
73 	PT_CAP(cycle_thresholds,	1, CPUID_EBX, 0xffff),
74 	PT_CAP(psb_periods,		1, CPUID_EBX, 0xffff0000),
75 };
76 
77 static u32 pt_cap_get(enum pt_capabilities cap)
78 {
79 	struct pt_cap_desc *cd = &pt_caps[cap];
80 	u32 c = pt_pmu.caps[cd->leaf * PT_CPUID_REGS_NUM + cd->reg];
81 	unsigned int shift = __ffs(cd->mask);
82 
83 	return (c & cd->mask) >> shift;
84 }
85 
86 static ssize_t pt_cap_show(struct device *cdev,
87 			   struct device_attribute *attr,
88 			   char *buf)
89 {
90 	struct dev_ext_attribute *ea =
91 		container_of(attr, struct dev_ext_attribute, attr);
92 	enum pt_capabilities cap = (long)ea->var;
93 
94 	return snprintf(buf, PAGE_SIZE, "%x\n", pt_cap_get(cap));
95 }
96 
97 static struct attribute_group pt_cap_group = {
98 	.name	= "caps",
99 };
100 
101 PMU_FORMAT_ATTR(cyc,		"config:1"	);
102 PMU_FORMAT_ATTR(mtc,		"config:9"	);
103 PMU_FORMAT_ATTR(tsc,		"config:10"	);
104 PMU_FORMAT_ATTR(noretcomp,	"config:11"	);
105 PMU_FORMAT_ATTR(mtc_period,	"config:14-17"	);
106 PMU_FORMAT_ATTR(cyc_thresh,	"config:19-22"	);
107 PMU_FORMAT_ATTR(psb_period,	"config:24-27"	);
108 
109 static struct attribute *pt_formats_attr[] = {
110 	&format_attr_cyc.attr,
111 	&format_attr_mtc.attr,
112 	&format_attr_tsc.attr,
113 	&format_attr_noretcomp.attr,
114 	&format_attr_mtc_period.attr,
115 	&format_attr_cyc_thresh.attr,
116 	&format_attr_psb_period.attr,
117 	NULL,
118 };
119 
120 static struct attribute_group pt_format_group = {
121 	.name	= "format",
122 	.attrs	= pt_formats_attr,
123 };
124 
125 static ssize_t
126 pt_timing_attr_show(struct device *dev, struct device_attribute *attr,
127 		    char *page)
128 {
129 	struct perf_pmu_events_attr *pmu_attr =
130 		container_of(attr, struct perf_pmu_events_attr, attr);
131 
132 	switch (pmu_attr->id) {
133 	case 0:
134 		return sprintf(page, "%lu\n", pt_pmu.max_nonturbo_ratio);
135 	case 1:
136 		return sprintf(page, "%u:%u\n",
137 			       pt_pmu.tsc_art_num,
138 			       pt_pmu.tsc_art_den);
139 	default:
140 		break;
141 	}
142 
143 	return -EINVAL;
144 }
145 
146 PMU_EVENT_ATTR(max_nonturbo_ratio, timing_attr_max_nonturbo_ratio, 0,
147 	       pt_timing_attr_show);
148 PMU_EVENT_ATTR(tsc_art_ratio, timing_attr_tsc_art_ratio, 1,
149 	       pt_timing_attr_show);
150 
151 static struct attribute *pt_timing_attr[] = {
152 	&timing_attr_max_nonturbo_ratio.attr.attr,
153 	&timing_attr_tsc_art_ratio.attr.attr,
154 	NULL,
155 };
156 
157 static struct attribute_group pt_timing_group = {
158 	.attrs	= pt_timing_attr,
159 };
160 
161 static const struct attribute_group *pt_attr_groups[] = {
162 	&pt_cap_group,
163 	&pt_format_group,
164 	&pt_timing_group,
165 	NULL,
166 };
167 
168 static int __init pt_pmu_hw_init(void)
169 {
170 	struct dev_ext_attribute *de_attrs;
171 	struct attribute **attrs;
172 	size_t size;
173 	u64 reg;
174 	int ret;
175 	long i;
176 
177 	rdmsrl(MSR_PLATFORM_INFO, reg);
178 	pt_pmu.max_nonturbo_ratio = (reg & 0xff00) >> 8;
179 
180 	/*
181 	 * if available, read in TSC to core crystal clock ratio,
182 	 * otherwise, zero for numerator stands for "not enumerated"
183 	 * as per SDM
184 	 */
185 	if (boot_cpu_data.cpuid_level >= CPUID_TSC_LEAF) {
186 		u32 eax, ebx, ecx, edx;
187 
188 		cpuid(CPUID_TSC_LEAF, &eax, &ebx, &ecx, &edx);
189 
190 		pt_pmu.tsc_art_num = ebx;
191 		pt_pmu.tsc_art_den = eax;
192 	}
193 
194 	if (boot_cpu_has(X86_FEATURE_VMX)) {
195 		/*
196 		 * Intel SDM, 36.5 "Tracing post-VMXON" says that
197 		 * "IA32_VMX_MISC[bit 14]" being 1 means PT can trace
198 		 * post-VMXON.
199 		 */
200 		rdmsrl(MSR_IA32_VMX_MISC, reg);
201 		if (reg & BIT(14))
202 			pt_pmu.vmx = true;
203 	}
204 
205 	attrs = NULL;
206 
207 	for (i = 0; i < PT_CPUID_LEAVES; i++) {
208 		cpuid_count(20, i,
209 			    &pt_pmu.caps[CPUID_EAX + i*PT_CPUID_REGS_NUM],
210 			    &pt_pmu.caps[CPUID_EBX + i*PT_CPUID_REGS_NUM],
211 			    &pt_pmu.caps[CPUID_ECX + i*PT_CPUID_REGS_NUM],
212 			    &pt_pmu.caps[CPUID_EDX + i*PT_CPUID_REGS_NUM]);
213 	}
214 
215 	ret = -ENOMEM;
216 	size = sizeof(struct attribute *) * (ARRAY_SIZE(pt_caps)+1);
217 	attrs = kzalloc(size, GFP_KERNEL);
218 	if (!attrs)
219 		goto fail;
220 
221 	size = sizeof(struct dev_ext_attribute) * (ARRAY_SIZE(pt_caps)+1);
222 	de_attrs = kzalloc(size, GFP_KERNEL);
223 	if (!de_attrs)
224 		goto fail;
225 
226 	for (i = 0; i < ARRAY_SIZE(pt_caps); i++) {
227 		struct dev_ext_attribute *de_attr = de_attrs + i;
228 
229 		de_attr->attr.attr.name = pt_caps[i].name;
230 
231 		sysfs_attr_init(&de_attr->attr.attr);
232 
233 		de_attr->attr.attr.mode		= S_IRUGO;
234 		de_attr->attr.show		= pt_cap_show;
235 		de_attr->var			= (void *)i;
236 
237 		attrs[i] = &de_attr->attr.attr;
238 	}
239 
240 	pt_cap_group.attrs = attrs;
241 
242 	return 0;
243 
244 fail:
245 	kfree(attrs);
246 
247 	return ret;
248 }
249 
250 #define RTIT_CTL_CYC_PSB (RTIT_CTL_CYCLEACC	| \
251 			  RTIT_CTL_CYC_THRESH	| \
252 			  RTIT_CTL_PSB_FREQ)
253 
254 #define RTIT_CTL_MTC	(RTIT_CTL_MTC_EN	| \
255 			 RTIT_CTL_MTC_RANGE)
256 
257 #define RTIT_CTL_PTW	(RTIT_CTL_PTW_EN	| \
258 			 RTIT_CTL_FUP_ON_PTW)
259 
260 #define PT_CONFIG_MASK (RTIT_CTL_TSC_EN		| \
261 			RTIT_CTL_DISRETC	| \
262 			RTIT_CTL_CYC_PSB	| \
263 			RTIT_CTL_MTC		| \
264 			RTIT_CTL_PWR_EVT_EN	| \
265 			RTIT_CTL_FUP_ON_PTW	| \
266 			RTIT_CTL_PTW_EN)
267 
268 static bool pt_event_valid(struct perf_event *event)
269 {
270 	u64 config = event->attr.config;
271 	u64 allowed, requested;
272 
273 	if ((config & PT_CONFIG_MASK) != config)
274 		return false;
275 
276 	if (config & RTIT_CTL_CYC_PSB) {
277 		if (!pt_cap_get(PT_CAP_psb_cyc))
278 			return false;
279 
280 		allowed = pt_cap_get(PT_CAP_psb_periods);
281 		requested = (config & RTIT_CTL_PSB_FREQ) >>
282 			RTIT_CTL_PSB_FREQ_OFFSET;
283 		if (requested && (!(allowed & BIT(requested))))
284 			return false;
285 
286 		allowed = pt_cap_get(PT_CAP_cycle_thresholds);
287 		requested = (config & RTIT_CTL_CYC_THRESH) >>
288 			RTIT_CTL_CYC_THRESH_OFFSET;
289 		if (requested && (!(allowed & BIT(requested))))
290 			return false;
291 	}
292 
293 	if (config & RTIT_CTL_MTC) {
294 		/*
295 		 * In the unlikely case that CPUID lists valid mtc periods,
296 		 * but not the mtc capability, drop out here.
297 		 *
298 		 * Spec says that setting mtc period bits while mtc bit in
299 		 * CPUID is 0 will #GP, so better safe than sorry.
300 		 */
301 		if (!pt_cap_get(PT_CAP_mtc))
302 			return false;
303 
304 		allowed = pt_cap_get(PT_CAP_mtc_periods);
305 		if (!allowed)
306 			return false;
307 
308 		requested = (config & RTIT_CTL_MTC_RANGE) >>
309 			RTIT_CTL_MTC_RANGE_OFFSET;
310 
311 		if (!(allowed & BIT(requested)))
312 			return false;
313 	}
314 
315 	if (config & RTIT_CTL_PWR_EVT_EN &&
316 	    !pt_cap_get(PT_CAP_power_event_trace))
317 		return false;
318 
319 	if (config & RTIT_CTL_PTW) {
320 		if (!pt_cap_get(PT_CAP_ptwrite))
321 			return false;
322 
323 		/* FUPonPTW without PTW doesn't make sense */
324 		if ((config & RTIT_CTL_FUP_ON_PTW) &&
325 		    !(config & RTIT_CTL_PTW_EN))
326 			return false;
327 	}
328 
329 	return true;
330 }
331 
332 /*
333  * PT configuration helpers
334  * These all are cpu affine and operate on a local PT
335  */
336 
337 /* Address ranges and their corresponding msr configuration registers */
338 static const struct pt_address_range {
339 	unsigned long	msr_a;
340 	unsigned long	msr_b;
341 	unsigned int	reg_off;
342 } pt_address_ranges[] = {
343 	{
344 		.msr_a	 = MSR_IA32_RTIT_ADDR0_A,
345 		.msr_b	 = MSR_IA32_RTIT_ADDR0_B,
346 		.reg_off = RTIT_CTL_ADDR0_OFFSET,
347 	},
348 	{
349 		.msr_a	 = MSR_IA32_RTIT_ADDR1_A,
350 		.msr_b	 = MSR_IA32_RTIT_ADDR1_B,
351 		.reg_off = RTIT_CTL_ADDR1_OFFSET,
352 	},
353 	{
354 		.msr_a	 = MSR_IA32_RTIT_ADDR2_A,
355 		.msr_b	 = MSR_IA32_RTIT_ADDR2_B,
356 		.reg_off = RTIT_CTL_ADDR2_OFFSET,
357 	},
358 	{
359 		.msr_a	 = MSR_IA32_RTIT_ADDR3_A,
360 		.msr_b	 = MSR_IA32_RTIT_ADDR3_B,
361 		.reg_off = RTIT_CTL_ADDR3_OFFSET,
362 	}
363 };
364 
365 static u64 pt_config_filters(struct perf_event *event)
366 {
367 	struct pt_filters *filters = event->hw.addr_filters;
368 	struct pt *pt = this_cpu_ptr(&pt_ctx);
369 	unsigned int range = 0;
370 	u64 rtit_ctl = 0;
371 
372 	if (!filters)
373 		return 0;
374 
375 	perf_event_addr_filters_sync(event);
376 
377 	for (range = 0; range < filters->nr_filters; range++) {
378 		struct pt_filter *filter = &filters->filter[range];
379 
380 		/*
381 		 * Note, if the range has zero start/end addresses due
382 		 * to its dynamic object not being loaded yet, we just
383 		 * go ahead and program zeroed range, which will simply
384 		 * produce no data. Note^2: if executable code at 0x0
385 		 * is a concern, we can set up an "invalid" configuration
386 		 * such as msr_b < msr_a.
387 		 */
388 
389 		/* avoid redundant msr writes */
390 		if (pt->filters.filter[range].msr_a != filter->msr_a) {
391 			wrmsrl(pt_address_ranges[range].msr_a, filter->msr_a);
392 			pt->filters.filter[range].msr_a = filter->msr_a;
393 		}
394 
395 		if (pt->filters.filter[range].msr_b != filter->msr_b) {
396 			wrmsrl(pt_address_ranges[range].msr_b, filter->msr_b);
397 			pt->filters.filter[range].msr_b = filter->msr_b;
398 		}
399 
400 		rtit_ctl |= filter->config << pt_address_ranges[range].reg_off;
401 	}
402 
403 	return rtit_ctl;
404 }
405 
406 static void pt_config(struct perf_event *event)
407 {
408 	u64 reg;
409 
410 	if (!event->hw.itrace_started) {
411 		event->hw.itrace_started = 1;
412 		wrmsrl(MSR_IA32_RTIT_STATUS, 0);
413 	}
414 
415 	reg = pt_config_filters(event);
416 	reg |= RTIT_CTL_TOPA | RTIT_CTL_BRANCH_EN | RTIT_CTL_TRACEEN;
417 
418 	if (!event->attr.exclude_kernel)
419 		reg |= RTIT_CTL_OS;
420 	if (!event->attr.exclude_user)
421 		reg |= RTIT_CTL_USR;
422 
423 	reg |= (event->attr.config & PT_CONFIG_MASK);
424 
425 	event->hw.config = reg;
426 	wrmsrl(MSR_IA32_RTIT_CTL, reg);
427 }
428 
429 static void pt_config_stop(struct perf_event *event)
430 {
431 	u64 ctl = READ_ONCE(event->hw.config);
432 
433 	/* may be already stopped by a PMI */
434 	if (!(ctl & RTIT_CTL_TRACEEN))
435 		return;
436 
437 	ctl &= ~RTIT_CTL_TRACEEN;
438 	wrmsrl(MSR_IA32_RTIT_CTL, ctl);
439 
440 	WRITE_ONCE(event->hw.config, ctl);
441 
442 	/*
443 	 * A wrmsr that disables trace generation serializes other PT
444 	 * registers and causes all data packets to be written to memory,
445 	 * but a fence is required for the data to become globally visible.
446 	 *
447 	 * The below WMB, separating data store and aux_head store matches
448 	 * the consumer's RMB that separates aux_head load and data load.
449 	 */
450 	wmb();
451 }
452 
453 static void pt_config_buffer(void *buf, unsigned int topa_idx,
454 			     unsigned int output_off)
455 {
456 	u64 reg;
457 
458 	wrmsrl(MSR_IA32_RTIT_OUTPUT_BASE, virt_to_phys(buf));
459 
460 	reg = 0x7f | ((u64)topa_idx << 7) | ((u64)output_off << 32);
461 
462 	wrmsrl(MSR_IA32_RTIT_OUTPUT_MASK, reg);
463 }
464 
465 /*
466  * Keep ToPA table-related metadata on the same page as the actual table,
467  * taking up a few words from the top
468  */
469 
470 #define TENTS_PER_PAGE (((PAGE_SIZE - 40) / sizeof(struct topa_entry)) - 1)
471 
472 /**
473  * struct topa - page-sized ToPA table with metadata at the top
474  * @table:	actual ToPA table entries, as understood by PT hardware
475  * @list:	linkage to struct pt_buffer's list of tables
476  * @phys:	physical address of this page
477  * @offset:	offset of the first entry in this table in the buffer
478  * @size:	total size of all entries in this table
479  * @last:	index of the last initialized entry in this table
480  */
481 struct topa {
482 	struct topa_entry	table[TENTS_PER_PAGE];
483 	struct list_head	list;
484 	u64			phys;
485 	u64			offset;
486 	size_t			size;
487 	int			last;
488 };
489 
490 /* make -1 stand for the last table entry */
491 #define TOPA_ENTRY(t, i) ((i) == -1 ? &(t)->table[(t)->last] : &(t)->table[(i)])
492 
493 /**
494  * topa_alloc() - allocate page-sized ToPA table
495  * @cpu:	CPU on which to allocate.
496  * @gfp:	Allocation flags.
497  *
498  * Return:	On success, return the pointer to ToPA table page.
499  */
500 static struct topa *topa_alloc(int cpu, gfp_t gfp)
501 {
502 	int node = cpu_to_node(cpu);
503 	struct topa *topa;
504 	struct page *p;
505 
506 	p = alloc_pages_node(node, gfp | __GFP_ZERO, 0);
507 	if (!p)
508 		return NULL;
509 
510 	topa = page_address(p);
511 	topa->last = 0;
512 	topa->phys = page_to_phys(p);
513 
514 	/*
515 	 * In case of singe-entry ToPA, always put the self-referencing END
516 	 * link as the 2nd entry in the table
517 	 */
518 	if (!pt_cap_get(PT_CAP_topa_multiple_entries)) {
519 		TOPA_ENTRY(topa, 1)->base = topa->phys >> TOPA_SHIFT;
520 		TOPA_ENTRY(topa, 1)->end = 1;
521 	}
522 
523 	return topa;
524 }
525 
526 /**
527  * topa_free() - free a page-sized ToPA table
528  * @topa:	Table to deallocate.
529  */
530 static void topa_free(struct topa *topa)
531 {
532 	free_page((unsigned long)topa);
533 }
534 
535 /**
536  * topa_insert_table() - insert a ToPA table into a buffer
537  * @buf:	 PT buffer that's being extended.
538  * @topa:	 New topa table to be inserted.
539  *
540  * If it's the first table in this buffer, set up buffer's pointers
541  * accordingly; otherwise, add a END=1 link entry to @topa to the current
542  * "last" table and adjust the last table pointer to @topa.
543  */
544 static void topa_insert_table(struct pt_buffer *buf, struct topa *topa)
545 {
546 	struct topa *last = buf->last;
547 
548 	list_add_tail(&topa->list, &buf->tables);
549 
550 	if (!buf->first) {
551 		buf->first = buf->last = buf->cur = topa;
552 		return;
553 	}
554 
555 	topa->offset = last->offset + last->size;
556 	buf->last = topa;
557 
558 	if (!pt_cap_get(PT_CAP_topa_multiple_entries))
559 		return;
560 
561 	BUG_ON(last->last != TENTS_PER_PAGE - 1);
562 
563 	TOPA_ENTRY(last, -1)->base = topa->phys >> TOPA_SHIFT;
564 	TOPA_ENTRY(last, -1)->end = 1;
565 }
566 
567 /**
568  * topa_table_full() - check if a ToPA table is filled up
569  * @topa:	ToPA table.
570  */
571 static bool topa_table_full(struct topa *topa)
572 {
573 	/* single-entry ToPA is a special case */
574 	if (!pt_cap_get(PT_CAP_topa_multiple_entries))
575 		return !!topa->last;
576 
577 	return topa->last == TENTS_PER_PAGE - 1;
578 }
579 
580 /**
581  * topa_insert_pages() - create a list of ToPA tables
582  * @buf:	PT buffer being initialized.
583  * @gfp:	Allocation flags.
584  *
585  * This initializes a list of ToPA tables with entries from
586  * the data_pages provided by rb_alloc_aux().
587  *
588  * Return:	0 on success or error code.
589  */
590 static int topa_insert_pages(struct pt_buffer *buf, gfp_t gfp)
591 {
592 	struct topa *topa = buf->last;
593 	int order = 0;
594 	struct page *p;
595 
596 	p = virt_to_page(buf->data_pages[buf->nr_pages]);
597 	if (PagePrivate(p))
598 		order = page_private(p);
599 
600 	if (topa_table_full(topa)) {
601 		topa = topa_alloc(buf->cpu, gfp);
602 		if (!topa)
603 			return -ENOMEM;
604 
605 		topa_insert_table(buf, topa);
606 	}
607 
608 	TOPA_ENTRY(topa, -1)->base = page_to_phys(p) >> TOPA_SHIFT;
609 	TOPA_ENTRY(topa, -1)->size = order;
610 	if (!buf->snapshot && !pt_cap_get(PT_CAP_topa_multiple_entries)) {
611 		TOPA_ENTRY(topa, -1)->intr = 1;
612 		TOPA_ENTRY(topa, -1)->stop = 1;
613 	}
614 
615 	topa->last++;
616 	topa->size += sizes(order);
617 
618 	buf->nr_pages += 1ul << order;
619 
620 	return 0;
621 }
622 
623 /**
624  * pt_topa_dump() - print ToPA tables and their entries
625  * @buf:	PT buffer.
626  */
627 static void pt_topa_dump(struct pt_buffer *buf)
628 {
629 	struct topa *topa;
630 
631 	list_for_each_entry(topa, &buf->tables, list) {
632 		int i;
633 
634 		pr_debug("# table @%p (%016Lx), off %llx size %zx\n", topa->table,
635 			 topa->phys, topa->offset, topa->size);
636 		for (i = 0; i < TENTS_PER_PAGE; i++) {
637 			pr_debug("# entry @%p (%lx sz %u %c%c%c) raw=%16llx\n",
638 				 &topa->table[i],
639 				 (unsigned long)topa->table[i].base << TOPA_SHIFT,
640 				 sizes(topa->table[i].size),
641 				 topa->table[i].end ?  'E' : ' ',
642 				 topa->table[i].intr ? 'I' : ' ',
643 				 topa->table[i].stop ? 'S' : ' ',
644 				 *(u64 *)&topa->table[i]);
645 			if ((pt_cap_get(PT_CAP_topa_multiple_entries) &&
646 			     topa->table[i].stop) ||
647 			    topa->table[i].end)
648 				break;
649 		}
650 	}
651 }
652 
653 /**
654  * pt_buffer_advance() - advance to the next output region
655  * @buf:	PT buffer.
656  *
657  * Advance the current pointers in the buffer to the next ToPA entry.
658  */
659 static void pt_buffer_advance(struct pt_buffer *buf)
660 {
661 	buf->output_off = 0;
662 	buf->cur_idx++;
663 
664 	if (buf->cur_idx == buf->cur->last) {
665 		if (buf->cur == buf->last)
666 			buf->cur = buf->first;
667 		else
668 			buf->cur = list_entry(buf->cur->list.next, struct topa,
669 					      list);
670 		buf->cur_idx = 0;
671 	}
672 }
673 
674 /**
675  * pt_update_head() - calculate current offsets and sizes
676  * @pt:		Per-cpu pt context.
677  *
678  * Update buffer's current write pointer position and data size.
679  */
680 static void pt_update_head(struct pt *pt)
681 {
682 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
683 	u64 topa_idx, base, old;
684 
685 	/* offset of the first region in this table from the beginning of buf */
686 	base = buf->cur->offset + buf->output_off;
687 
688 	/* offset of the current output region within this table */
689 	for (topa_idx = 0; topa_idx < buf->cur_idx; topa_idx++)
690 		base += sizes(buf->cur->table[topa_idx].size);
691 
692 	if (buf->snapshot) {
693 		local_set(&buf->data_size, base);
694 	} else {
695 		old = (local64_xchg(&buf->head, base) &
696 		       ((buf->nr_pages << PAGE_SHIFT) - 1));
697 		if (base < old)
698 			base += buf->nr_pages << PAGE_SHIFT;
699 
700 		local_add(base - old, &buf->data_size);
701 	}
702 }
703 
704 /**
705  * pt_buffer_region() - obtain current output region's address
706  * @buf:	PT buffer.
707  */
708 static void *pt_buffer_region(struct pt_buffer *buf)
709 {
710 	return phys_to_virt(buf->cur->table[buf->cur_idx].base << TOPA_SHIFT);
711 }
712 
713 /**
714  * pt_buffer_region_size() - obtain current output region's size
715  * @buf:	PT buffer.
716  */
717 static size_t pt_buffer_region_size(struct pt_buffer *buf)
718 {
719 	return sizes(buf->cur->table[buf->cur_idx].size);
720 }
721 
722 /**
723  * pt_handle_status() - take care of possible status conditions
724  * @pt:		Per-cpu pt context.
725  */
726 static void pt_handle_status(struct pt *pt)
727 {
728 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
729 	int advance = 0;
730 	u64 status;
731 
732 	rdmsrl(MSR_IA32_RTIT_STATUS, status);
733 
734 	if (status & RTIT_STATUS_ERROR) {
735 		pr_err_ratelimited("ToPA ERROR encountered, trying to recover\n");
736 		pt_topa_dump(buf);
737 		status &= ~RTIT_STATUS_ERROR;
738 	}
739 
740 	if (status & RTIT_STATUS_STOPPED) {
741 		status &= ~RTIT_STATUS_STOPPED;
742 
743 		/*
744 		 * On systems that only do single-entry ToPA, hitting STOP
745 		 * means we are already losing data; need to let the decoder
746 		 * know.
747 		 */
748 		if (!pt_cap_get(PT_CAP_topa_multiple_entries) ||
749 		    buf->output_off == sizes(TOPA_ENTRY(buf->cur, buf->cur_idx)->size)) {
750 			local_inc(&buf->lost);
751 			advance++;
752 		}
753 	}
754 
755 	/*
756 	 * Also on single-entry ToPA implementations, interrupt will come
757 	 * before the output reaches its output region's boundary.
758 	 */
759 	if (!pt_cap_get(PT_CAP_topa_multiple_entries) && !buf->snapshot &&
760 	    pt_buffer_region_size(buf) - buf->output_off <= TOPA_PMI_MARGIN) {
761 		void *head = pt_buffer_region(buf);
762 
763 		/* everything within this margin needs to be zeroed out */
764 		memset(head + buf->output_off, 0,
765 		       pt_buffer_region_size(buf) -
766 		       buf->output_off);
767 		advance++;
768 	}
769 
770 	if (advance)
771 		pt_buffer_advance(buf);
772 
773 	wrmsrl(MSR_IA32_RTIT_STATUS, status);
774 }
775 
776 /**
777  * pt_read_offset() - translate registers into buffer pointers
778  * @buf:	PT buffer.
779  *
780  * Set buffer's output pointers from MSR values.
781  */
782 static void pt_read_offset(struct pt_buffer *buf)
783 {
784 	u64 offset, base_topa;
785 
786 	rdmsrl(MSR_IA32_RTIT_OUTPUT_BASE, base_topa);
787 	buf->cur = phys_to_virt(base_topa);
788 
789 	rdmsrl(MSR_IA32_RTIT_OUTPUT_MASK, offset);
790 	/* offset within current output region */
791 	buf->output_off = offset >> 32;
792 	/* index of current output region within this table */
793 	buf->cur_idx = (offset & 0xffffff80) >> 7;
794 }
795 
796 /**
797  * pt_topa_next_entry() - obtain index of the first page in the next ToPA entry
798  * @buf:	PT buffer.
799  * @pg:		Page offset in the buffer.
800  *
801  * When advancing to the next output region (ToPA entry), given a page offset
802  * into the buffer, we need to find the offset of the first page in the next
803  * region.
804  */
805 static unsigned int pt_topa_next_entry(struct pt_buffer *buf, unsigned int pg)
806 {
807 	struct topa_entry *te = buf->topa_index[pg];
808 
809 	/* one region */
810 	if (buf->first == buf->last && buf->first->last == 1)
811 		return pg;
812 
813 	do {
814 		pg++;
815 		pg &= buf->nr_pages - 1;
816 	} while (buf->topa_index[pg] == te);
817 
818 	return pg;
819 }
820 
821 /**
822  * pt_buffer_reset_markers() - place interrupt and stop bits in the buffer
823  * @buf:	PT buffer.
824  * @handle:	Current output handle.
825  *
826  * Place INT and STOP marks to prevent overwriting old data that the consumer
827  * hasn't yet collected and waking up the consumer after a certain fraction of
828  * the buffer has filled up. Only needed and sensible for non-snapshot counters.
829  *
830  * This obviously relies on buf::head to figure out buffer markers, so it has
831  * to be called after pt_buffer_reset_offsets() and before the hardware tracing
832  * is enabled.
833  */
834 static int pt_buffer_reset_markers(struct pt_buffer *buf,
835 				   struct perf_output_handle *handle)
836 
837 {
838 	unsigned long head = local64_read(&buf->head);
839 	unsigned long idx, npages, wakeup;
840 
841 	/* can't stop in the middle of an output region */
842 	if (buf->output_off + handle->size + 1 <
843 	    sizes(TOPA_ENTRY(buf->cur, buf->cur_idx)->size))
844 		return -EINVAL;
845 
846 
847 	/* single entry ToPA is handled by marking all regions STOP=1 INT=1 */
848 	if (!pt_cap_get(PT_CAP_topa_multiple_entries))
849 		return 0;
850 
851 	/* clear STOP and INT from current entry */
852 	buf->topa_index[buf->stop_pos]->stop = 0;
853 	buf->topa_index[buf->stop_pos]->intr = 0;
854 	buf->topa_index[buf->intr_pos]->intr = 0;
855 
856 	/* how many pages till the STOP marker */
857 	npages = handle->size >> PAGE_SHIFT;
858 
859 	/* if it's on a page boundary, fill up one more page */
860 	if (!offset_in_page(head + handle->size + 1))
861 		npages++;
862 
863 	idx = (head >> PAGE_SHIFT) + npages;
864 	idx &= buf->nr_pages - 1;
865 	buf->stop_pos = idx;
866 
867 	wakeup = handle->wakeup >> PAGE_SHIFT;
868 
869 	/* in the worst case, wake up the consumer one page before hard stop */
870 	idx = (head >> PAGE_SHIFT) + npages - 1;
871 	if (idx > wakeup)
872 		idx = wakeup;
873 
874 	idx &= buf->nr_pages - 1;
875 	buf->intr_pos = idx;
876 
877 	buf->topa_index[buf->stop_pos]->stop = 1;
878 	buf->topa_index[buf->stop_pos]->intr = 1;
879 	buf->topa_index[buf->intr_pos]->intr = 1;
880 
881 	return 0;
882 }
883 
884 /**
885  * pt_buffer_setup_topa_index() - build topa_index[] table of regions
886  * @buf:	PT buffer.
887  *
888  * topa_index[] references output regions indexed by offset into the
889  * buffer for purposes of quick reverse lookup.
890  */
891 static void pt_buffer_setup_topa_index(struct pt_buffer *buf)
892 {
893 	struct topa *cur = buf->first, *prev = buf->last;
894 	struct topa_entry *te_cur = TOPA_ENTRY(cur, 0),
895 		*te_prev = TOPA_ENTRY(prev, prev->last - 1);
896 	int pg = 0, idx = 0;
897 
898 	while (pg < buf->nr_pages) {
899 		int tidx;
900 
901 		/* pages within one topa entry */
902 		for (tidx = 0; tidx < 1 << te_cur->size; tidx++, pg++)
903 			buf->topa_index[pg] = te_prev;
904 
905 		te_prev = te_cur;
906 
907 		if (idx == cur->last - 1) {
908 			/* advance to next topa table */
909 			idx = 0;
910 			cur = list_entry(cur->list.next, struct topa, list);
911 		} else {
912 			idx++;
913 		}
914 		te_cur = TOPA_ENTRY(cur, idx);
915 	}
916 
917 }
918 
919 /**
920  * pt_buffer_reset_offsets() - adjust buffer's write pointers from aux_head
921  * @buf:	PT buffer.
922  * @head:	Write pointer (aux_head) from AUX buffer.
923  *
924  * Find the ToPA table and entry corresponding to given @head and set buffer's
925  * "current" pointers accordingly. This is done after we have obtained the
926  * current aux_head position from a successful call to perf_aux_output_begin()
927  * to make sure the hardware is writing to the right place.
928  *
929  * This function modifies buf::{cur,cur_idx,output_off} that will be programmed
930  * into PT msrs when the tracing is enabled and buf::head and buf::data_size,
931  * which are used to determine INT and STOP markers' locations by a subsequent
932  * call to pt_buffer_reset_markers().
933  */
934 static void pt_buffer_reset_offsets(struct pt_buffer *buf, unsigned long head)
935 {
936 	int pg;
937 
938 	if (buf->snapshot)
939 		head &= (buf->nr_pages << PAGE_SHIFT) - 1;
940 
941 	pg = (head >> PAGE_SHIFT) & (buf->nr_pages - 1);
942 	pg = pt_topa_next_entry(buf, pg);
943 
944 	buf->cur = (struct topa *)((unsigned long)buf->topa_index[pg] & PAGE_MASK);
945 	buf->cur_idx = ((unsigned long)buf->topa_index[pg] -
946 			(unsigned long)buf->cur) / sizeof(struct topa_entry);
947 	buf->output_off = head & (sizes(buf->cur->table[buf->cur_idx].size) - 1);
948 
949 	local64_set(&buf->head, head);
950 	local_set(&buf->data_size, 0);
951 }
952 
953 /**
954  * pt_buffer_fini_topa() - deallocate ToPA structure of a buffer
955  * @buf:	PT buffer.
956  */
957 static void pt_buffer_fini_topa(struct pt_buffer *buf)
958 {
959 	struct topa *topa, *iter;
960 
961 	list_for_each_entry_safe(topa, iter, &buf->tables, list) {
962 		/*
963 		 * right now, this is in free_aux() path only, so
964 		 * no need to unlink this table from the list
965 		 */
966 		topa_free(topa);
967 	}
968 }
969 
970 /**
971  * pt_buffer_init_topa() - initialize ToPA table for pt buffer
972  * @buf:	PT buffer.
973  * @size:	Total size of all regions within this ToPA.
974  * @gfp:	Allocation flags.
975  */
976 static int pt_buffer_init_topa(struct pt_buffer *buf, unsigned long nr_pages,
977 			       gfp_t gfp)
978 {
979 	struct topa *topa;
980 	int err;
981 
982 	topa = topa_alloc(buf->cpu, gfp);
983 	if (!topa)
984 		return -ENOMEM;
985 
986 	topa_insert_table(buf, topa);
987 
988 	while (buf->nr_pages < nr_pages) {
989 		err = topa_insert_pages(buf, gfp);
990 		if (err) {
991 			pt_buffer_fini_topa(buf);
992 			return -ENOMEM;
993 		}
994 	}
995 
996 	pt_buffer_setup_topa_index(buf);
997 
998 	/* link last table to the first one, unless we're double buffering */
999 	if (pt_cap_get(PT_CAP_topa_multiple_entries)) {
1000 		TOPA_ENTRY(buf->last, -1)->base = buf->first->phys >> TOPA_SHIFT;
1001 		TOPA_ENTRY(buf->last, -1)->end = 1;
1002 	}
1003 
1004 	pt_topa_dump(buf);
1005 	return 0;
1006 }
1007 
1008 /**
1009  * pt_buffer_setup_aux() - set up topa tables for a PT buffer
1010  * @cpu:	Cpu on which to allocate, -1 means current.
1011  * @pages:	Array of pointers to buffer pages passed from perf core.
1012  * @nr_pages:	Number of pages in the buffer.
1013  * @snapshot:	If this is a snapshot/overwrite counter.
1014  *
1015  * This is a pmu::setup_aux callback that sets up ToPA tables and all the
1016  * bookkeeping for an AUX buffer.
1017  *
1018  * Return:	Our private PT buffer structure.
1019  */
1020 static void *
1021 pt_buffer_setup_aux(int cpu, void **pages, int nr_pages, bool snapshot)
1022 {
1023 	struct pt_buffer *buf;
1024 	int node, ret;
1025 
1026 	if (!nr_pages)
1027 		return NULL;
1028 
1029 	if (cpu == -1)
1030 		cpu = raw_smp_processor_id();
1031 	node = cpu_to_node(cpu);
1032 
1033 	buf = kzalloc_node(offsetof(struct pt_buffer, topa_index[nr_pages]),
1034 			   GFP_KERNEL, node);
1035 	if (!buf)
1036 		return NULL;
1037 
1038 	buf->cpu = cpu;
1039 	buf->snapshot = snapshot;
1040 	buf->data_pages = pages;
1041 
1042 	INIT_LIST_HEAD(&buf->tables);
1043 
1044 	ret = pt_buffer_init_topa(buf, nr_pages, GFP_KERNEL);
1045 	if (ret) {
1046 		kfree(buf);
1047 		return NULL;
1048 	}
1049 
1050 	return buf;
1051 }
1052 
1053 /**
1054  * pt_buffer_free_aux() - perf AUX deallocation path callback
1055  * @data:	PT buffer.
1056  */
1057 static void pt_buffer_free_aux(void *data)
1058 {
1059 	struct pt_buffer *buf = data;
1060 
1061 	pt_buffer_fini_topa(buf);
1062 	kfree(buf);
1063 }
1064 
1065 static int pt_addr_filters_init(struct perf_event *event)
1066 {
1067 	struct pt_filters *filters;
1068 	int node = event->cpu == -1 ? -1 : cpu_to_node(event->cpu);
1069 
1070 	if (!pt_cap_get(PT_CAP_num_address_ranges))
1071 		return 0;
1072 
1073 	filters = kzalloc_node(sizeof(struct pt_filters), GFP_KERNEL, node);
1074 	if (!filters)
1075 		return -ENOMEM;
1076 
1077 	if (event->parent)
1078 		memcpy(filters, event->parent->hw.addr_filters,
1079 		       sizeof(*filters));
1080 
1081 	event->hw.addr_filters = filters;
1082 
1083 	return 0;
1084 }
1085 
1086 static void pt_addr_filters_fini(struct perf_event *event)
1087 {
1088 	kfree(event->hw.addr_filters);
1089 	event->hw.addr_filters = NULL;
1090 }
1091 
1092 static inline bool valid_kernel_ip(unsigned long ip)
1093 {
1094 	return virt_addr_valid(ip) && kernel_ip(ip);
1095 }
1096 
1097 static int pt_event_addr_filters_validate(struct list_head *filters)
1098 {
1099 	struct perf_addr_filter *filter;
1100 	int range = 0;
1101 
1102 	list_for_each_entry(filter, filters, entry) {
1103 		/* PT doesn't support single address triggers */
1104 		if (!filter->range || !filter->size)
1105 			return -EOPNOTSUPP;
1106 
1107 		if (!filter->inode) {
1108 			if (!valid_kernel_ip(filter->offset))
1109 				return -EINVAL;
1110 
1111 			if (!valid_kernel_ip(filter->offset + filter->size))
1112 				return -EINVAL;
1113 		}
1114 
1115 		if (++range > pt_cap_get(PT_CAP_num_address_ranges))
1116 			return -EOPNOTSUPP;
1117 	}
1118 
1119 	return 0;
1120 }
1121 
1122 static void pt_event_addr_filters_sync(struct perf_event *event)
1123 {
1124 	struct perf_addr_filters_head *head = perf_event_addr_filters(event);
1125 	unsigned long msr_a, msr_b, *offs = event->addr_filters_offs;
1126 	struct pt_filters *filters = event->hw.addr_filters;
1127 	struct perf_addr_filter *filter;
1128 	int range = 0;
1129 
1130 	if (!filters)
1131 		return;
1132 
1133 	list_for_each_entry(filter, &head->list, entry) {
1134 		if (filter->inode && !offs[range]) {
1135 			msr_a = msr_b = 0;
1136 		} else {
1137 			/* apply the offset */
1138 			msr_a = filter->offset + offs[range];
1139 			msr_b = filter->size + msr_a - 1;
1140 		}
1141 
1142 		filters->filter[range].msr_a  = msr_a;
1143 		filters->filter[range].msr_b  = msr_b;
1144 		filters->filter[range].config = filter->filter ? 1 : 2;
1145 		range++;
1146 	}
1147 
1148 	filters->nr_filters = range;
1149 }
1150 
1151 /**
1152  * intel_pt_interrupt() - PT PMI handler
1153  */
1154 void intel_pt_interrupt(void)
1155 {
1156 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1157 	struct pt_buffer *buf;
1158 	struct perf_event *event = pt->handle.event;
1159 
1160 	/*
1161 	 * There may be a dangling PT bit in the interrupt status register
1162 	 * after PT has been disabled by pt_event_stop(). Make sure we don't
1163 	 * do anything (particularly, re-enable) for this event here.
1164 	 */
1165 	if (!READ_ONCE(pt->handle_nmi))
1166 		return;
1167 
1168 	/*
1169 	 * If VMX is on and PT does not support it, don't touch anything.
1170 	 */
1171 	if (READ_ONCE(pt->vmx_on))
1172 		return;
1173 
1174 	if (!event)
1175 		return;
1176 
1177 	pt_config_stop(event);
1178 
1179 	buf = perf_get_aux(&pt->handle);
1180 	if (!buf)
1181 		return;
1182 
1183 	pt_read_offset(buf);
1184 
1185 	pt_handle_status(pt);
1186 
1187 	pt_update_head(pt);
1188 
1189 	perf_aux_output_end(&pt->handle, local_xchg(&buf->data_size, 0),
1190 			    local_xchg(&buf->lost, 0));
1191 
1192 	if (!event->hw.state) {
1193 		int ret;
1194 
1195 		buf = perf_aux_output_begin(&pt->handle, event);
1196 		if (!buf) {
1197 			event->hw.state = PERF_HES_STOPPED;
1198 			return;
1199 		}
1200 
1201 		pt_buffer_reset_offsets(buf, pt->handle.head);
1202 		/* snapshot counters don't use PMI, so it's safe */
1203 		ret = pt_buffer_reset_markers(buf, &pt->handle);
1204 		if (ret) {
1205 			perf_aux_output_end(&pt->handle, 0, true);
1206 			return;
1207 		}
1208 
1209 		pt_config_buffer(buf->cur->table, buf->cur_idx,
1210 				 buf->output_off);
1211 		pt_config(event);
1212 	}
1213 }
1214 
1215 void intel_pt_handle_vmx(int on)
1216 {
1217 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1218 	struct perf_event *event;
1219 	unsigned long flags;
1220 
1221 	/* PT plays nice with VMX, do nothing */
1222 	if (pt_pmu.vmx)
1223 		return;
1224 
1225 	/*
1226 	 * VMXON will clear RTIT_CTL.TraceEn; we need to make
1227 	 * sure to not try to set it while VMX is on. Disable
1228 	 * interrupts to avoid racing with pmu callbacks;
1229 	 * concurrent PMI should be handled fine.
1230 	 */
1231 	local_irq_save(flags);
1232 	WRITE_ONCE(pt->vmx_on, on);
1233 
1234 	if (on) {
1235 		/* prevent pt_config_stop() from writing RTIT_CTL */
1236 		event = pt->handle.event;
1237 		if (event)
1238 			event->hw.config = 0;
1239 	}
1240 	local_irq_restore(flags);
1241 }
1242 EXPORT_SYMBOL_GPL(intel_pt_handle_vmx);
1243 
1244 /*
1245  * PMU callbacks
1246  */
1247 
1248 static void pt_event_start(struct perf_event *event, int mode)
1249 {
1250 	struct hw_perf_event *hwc = &event->hw;
1251 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1252 	struct pt_buffer *buf;
1253 
1254 	if (READ_ONCE(pt->vmx_on))
1255 		return;
1256 
1257 	buf = perf_aux_output_begin(&pt->handle, event);
1258 	if (!buf)
1259 		goto fail_stop;
1260 
1261 	pt_buffer_reset_offsets(buf, pt->handle.head);
1262 	if (!buf->snapshot) {
1263 		if (pt_buffer_reset_markers(buf, &pt->handle))
1264 			goto fail_end_stop;
1265 	}
1266 
1267 	WRITE_ONCE(pt->handle_nmi, 1);
1268 	hwc->state = 0;
1269 
1270 	pt_config_buffer(buf->cur->table, buf->cur_idx,
1271 			 buf->output_off);
1272 	pt_config(event);
1273 
1274 	return;
1275 
1276 fail_end_stop:
1277 	perf_aux_output_end(&pt->handle, 0, true);
1278 fail_stop:
1279 	hwc->state = PERF_HES_STOPPED;
1280 }
1281 
1282 static void pt_event_stop(struct perf_event *event, int mode)
1283 {
1284 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1285 
1286 	/*
1287 	 * Protect against the PMI racing with disabling wrmsr,
1288 	 * see comment in intel_pt_interrupt().
1289 	 */
1290 	WRITE_ONCE(pt->handle_nmi, 0);
1291 
1292 	pt_config_stop(event);
1293 
1294 	if (event->hw.state == PERF_HES_STOPPED)
1295 		return;
1296 
1297 	event->hw.state = PERF_HES_STOPPED;
1298 
1299 	if (mode & PERF_EF_UPDATE) {
1300 		struct pt_buffer *buf = perf_get_aux(&pt->handle);
1301 
1302 		if (!buf)
1303 			return;
1304 
1305 		if (WARN_ON_ONCE(pt->handle.event != event))
1306 			return;
1307 
1308 		pt_read_offset(buf);
1309 
1310 		pt_handle_status(pt);
1311 
1312 		pt_update_head(pt);
1313 
1314 		if (buf->snapshot)
1315 			pt->handle.head =
1316 				local_xchg(&buf->data_size,
1317 					   buf->nr_pages << PAGE_SHIFT);
1318 		perf_aux_output_end(&pt->handle, local_xchg(&buf->data_size, 0),
1319 				    local_xchg(&buf->lost, 0));
1320 	}
1321 }
1322 
1323 static void pt_event_del(struct perf_event *event, int mode)
1324 {
1325 	pt_event_stop(event, PERF_EF_UPDATE);
1326 }
1327 
1328 static int pt_event_add(struct perf_event *event, int mode)
1329 {
1330 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1331 	struct hw_perf_event *hwc = &event->hw;
1332 	int ret = -EBUSY;
1333 
1334 	if (pt->handle.event)
1335 		goto fail;
1336 
1337 	if (mode & PERF_EF_START) {
1338 		pt_event_start(event, 0);
1339 		ret = -EINVAL;
1340 		if (hwc->state == PERF_HES_STOPPED)
1341 			goto fail;
1342 	} else {
1343 		hwc->state = PERF_HES_STOPPED;
1344 	}
1345 
1346 	ret = 0;
1347 fail:
1348 
1349 	return ret;
1350 }
1351 
1352 static void pt_event_read(struct perf_event *event)
1353 {
1354 }
1355 
1356 static void pt_event_destroy(struct perf_event *event)
1357 {
1358 	pt_addr_filters_fini(event);
1359 	x86_del_exclusive(x86_lbr_exclusive_pt);
1360 }
1361 
1362 static int pt_event_init(struct perf_event *event)
1363 {
1364 	if (event->attr.type != pt_pmu.pmu.type)
1365 		return -ENOENT;
1366 
1367 	if (!pt_event_valid(event))
1368 		return -EINVAL;
1369 
1370 	if (x86_add_exclusive(x86_lbr_exclusive_pt))
1371 		return -EBUSY;
1372 
1373 	if (pt_addr_filters_init(event)) {
1374 		x86_del_exclusive(x86_lbr_exclusive_pt);
1375 		return -ENOMEM;
1376 	}
1377 
1378 	event->destroy = pt_event_destroy;
1379 
1380 	return 0;
1381 }
1382 
1383 void cpu_emergency_stop_pt(void)
1384 {
1385 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1386 
1387 	if (pt->handle.event)
1388 		pt_event_stop(pt->handle.event, PERF_EF_UPDATE);
1389 }
1390 
1391 static __init int pt_init(void)
1392 {
1393 	int ret, cpu, prior_warn = 0;
1394 
1395 	BUILD_BUG_ON(sizeof(struct topa) > PAGE_SIZE);
1396 
1397 	if (!boot_cpu_has(X86_FEATURE_INTEL_PT))
1398 		return -ENODEV;
1399 
1400 	get_online_cpus();
1401 	for_each_online_cpu(cpu) {
1402 		u64 ctl;
1403 
1404 		ret = rdmsrl_safe_on_cpu(cpu, MSR_IA32_RTIT_CTL, &ctl);
1405 		if (!ret && (ctl & RTIT_CTL_TRACEEN))
1406 			prior_warn++;
1407 	}
1408 	put_online_cpus();
1409 
1410 	if (prior_warn) {
1411 		x86_add_exclusive(x86_lbr_exclusive_pt);
1412 		pr_warn("PT is enabled at boot time, doing nothing\n");
1413 
1414 		return -EBUSY;
1415 	}
1416 
1417 	ret = pt_pmu_hw_init();
1418 	if (ret)
1419 		return ret;
1420 
1421 	if (!pt_cap_get(PT_CAP_topa_output)) {
1422 		pr_warn("ToPA output is not supported on this CPU\n");
1423 		return -ENODEV;
1424 	}
1425 
1426 	if (!pt_cap_get(PT_CAP_topa_multiple_entries))
1427 		pt_pmu.pmu.capabilities =
1428 			PERF_PMU_CAP_AUX_NO_SG | PERF_PMU_CAP_AUX_SW_DOUBLEBUF;
1429 
1430 	pt_pmu.pmu.capabilities	|= PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE;
1431 	pt_pmu.pmu.attr_groups		 = pt_attr_groups;
1432 	pt_pmu.pmu.task_ctx_nr		 = perf_sw_context;
1433 	pt_pmu.pmu.event_init		 = pt_event_init;
1434 	pt_pmu.pmu.add			 = pt_event_add;
1435 	pt_pmu.pmu.del			 = pt_event_del;
1436 	pt_pmu.pmu.start		 = pt_event_start;
1437 	pt_pmu.pmu.stop			 = pt_event_stop;
1438 	pt_pmu.pmu.read			 = pt_event_read;
1439 	pt_pmu.pmu.setup_aux		 = pt_buffer_setup_aux;
1440 	pt_pmu.pmu.free_aux		 = pt_buffer_free_aux;
1441 	pt_pmu.pmu.addr_filters_sync     = pt_event_addr_filters_sync;
1442 	pt_pmu.pmu.addr_filters_validate = pt_event_addr_filters_validate;
1443 	pt_pmu.pmu.nr_addr_filters       =
1444 		pt_cap_get(PT_CAP_num_address_ranges);
1445 
1446 	ret = perf_pmu_register(&pt_pmu.pmu, "intel_pt", -1);
1447 
1448 	return ret;
1449 }
1450 arch_initcall(pt_init);
1451