xref: /openbmc/linux/arch/x86/events/intel/pt.c (revision 1f327613)
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 #include <asm/intel-family.h>
32 
33 #include "../perf_event.h"
34 #include "pt.h"
35 
36 static DEFINE_PER_CPU(struct pt, pt_ctx);
37 
38 static struct pt_pmu pt_pmu;
39 
40 /*
41  * Capabilities of Intel PT hardware, such as number of address bits or
42  * supported output schemes, are cached and exported to userspace as "caps"
43  * attribute group of pt pmu device
44  * (/sys/bus/event_source/devices/intel_pt/caps/) so that userspace can store
45  * relevant bits together with intel_pt traces.
46  *
47  * These are necessary for both trace decoding (payloads_lip, contains address
48  * width encoded in IP-related packets), and event configuration (bitmasks with
49  * permitted values for certain bit fields).
50  */
51 #define PT_CAP(_n, _l, _r, _m)						\
52 	[PT_CAP_ ## _n] = { .name = __stringify(_n), .leaf = _l,	\
53 			    .reg = _r, .mask = _m }
54 
55 static struct pt_cap_desc {
56 	const char	*name;
57 	u32		leaf;
58 	u8		reg;
59 	u32		mask;
60 } pt_caps[] = {
61 	PT_CAP(max_subleaf,		0, CPUID_EAX, 0xffffffff),
62 	PT_CAP(cr3_filtering,		0, CPUID_EBX, BIT(0)),
63 	PT_CAP(psb_cyc,			0, CPUID_EBX, BIT(1)),
64 	PT_CAP(ip_filtering,		0, CPUID_EBX, BIT(2)),
65 	PT_CAP(mtc,			0, CPUID_EBX, BIT(3)),
66 	PT_CAP(ptwrite,			0, CPUID_EBX, BIT(4)),
67 	PT_CAP(power_event_trace,	0, CPUID_EBX, BIT(5)),
68 	PT_CAP(topa_output,		0, CPUID_ECX, BIT(0)),
69 	PT_CAP(topa_multiple_entries,	0, CPUID_ECX, BIT(1)),
70 	PT_CAP(single_range_output,	0, CPUID_ECX, BIT(2)),
71 	PT_CAP(output_subsys,		0, CPUID_ECX, BIT(3)),
72 	PT_CAP(payloads_lip,		0, CPUID_ECX, BIT(31)),
73 	PT_CAP(num_address_ranges,	1, CPUID_EAX, 0x3),
74 	PT_CAP(mtc_periods,		1, CPUID_EAX, 0xffff0000),
75 	PT_CAP(cycle_thresholds,	1, CPUID_EBX, 0xffff),
76 	PT_CAP(psb_periods,		1, CPUID_EBX, 0xffff0000),
77 };
78 
79 u32 intel_pt_validate_cap(u32 *caps, enum pt_capabilities capability)
80 {
81 	struct pt_cap_desc *cd = &pt_caps[capability];
82 	u32 c = caps[cd->leaf * PT_CPUID_REGS_NUM + cd->reg];
83 	unsigned int shift = __ffs(cd->mask);
84 
85 	return (c & cd->mask) >> shift;
86 }
87 EXPORT_SYMBOL_GPL(intel_pt_validate_cap);
88 
89 u32 intel_pt_validate_hw_cap(enum pt_capabilities cap)
90 {
91 	return intel_pt_validate_cap(pt_pmu.caps, cap);
92 }
93 EXPORT_SYMBOL_GPL(intel_pt_validate_hw_cap);
94 
95 static ssize_t pt_cap_show(struct device *cdev,
96 			   struct device_attribute *attr,
97 			   char *buf)
98 {
99 	struct dev_ext_attribute *ea =
100 		container_of(attr, struct dev_ext_attribute, attr);
101 	enum pt_capabilities cap = (long)ea->var;
102 
103 	return snprintf(buf, PAGE_SIZE, "%x\n", intel_pt_validate_hw_cap(cap));
104 }
105 
106 static struct attribute_group pt_cap_group __ro_after_init = {
107 	.name	= "caps",
108 };
109 
110 PMU_FORMAT_ATTR(pt,		"config:0"	);
111 PMU_FORMAT_ATTR(cyc,		"config:1"	);
112 PMU_FORMAT_ATTR(pwr_evt,	"config:4"	);
113 PMU_FORMAT_ATTR(fup_on_ptw,	"config:5"	);
114 PMU_FORMAT_ATTR(mtc,		"config:9"	);
115 PMU_FORMAT_ATTR(tsc,		"config:10"	);
116 PMU_FORMAT_ATTR(noretcomp,	"config:11"	);
117 PMU_FORMAT_ATTR(ptw,		"config:12"	);
118 PMU_FORMAT_ATTR(branch,		"config:13"	);
119 PMU_FORMAT_ATTR(mtc_period,	"config:14-17"	);
120 PMU_FORMAT_ATTR(cyc_thresh,	"config:19-22"	);
121 PMU_FORMAT_ATTR(psb_period,	"config:24-27"	);
122 
123 static struct attribute *pt_formats_attr[] = {
124 	&format_attr_pt.attr,
125 	&format_attr_cyc.attr,
126 	&format_attr_pwr_evt.attr,
127 	&format_attr_fup_on_ptw.attr,
128 	&format_attr_mtc.attr,
129 	&format_attr_tsc.attr,
130 	&format_attr_noretcomp.attr,
131 	&format_attr_ptw.attr,
132 	&format_attr_branch.attr,
133 	&format_attr_mtc_period.attr,
134 	&format_attr_cyc_thresh.attr,
135 	&format_attr_psb_period.attr,
136 	NULL,
137 };
138 
139 static struct attribute_group pt_format_group = {
140 	.name	= "format",
141 	.attrs	= pt_formats_attr,
142 };
143 
144 static ssize_t
145 pt_timing_attr_show(struct device *dev, struct device_attribute *attr,
146 		    char *page)
147 {
148 	struct perf_pmu_events_attr *pmu_attr =
149 		container_of(attr, struct perf_pmu_events_attr, attr);
150 
151 	switch (pmu_attr->id) {
152 	case 0:
153 		return sprintf(page, "%lu\n", pt_pmu.max_nonturbo_ratio);
154 	case 1:
155 		return sprintf(page, "%u:%u\n",
156 			       pt_pmu.tsc_art_num,
157 			       pt_pmu.tsc_art_den);
158 	default:
159 		break;
160 	}
161 
162 	return -EINVAL;
163 }
164 
165 PMU_EVENT_ATTR(max_nonturbo_ratio, timing_attr_max_nonturbo_ratio, 0,
166 	       pt_timing_attr_show);
167 PMU_EVENT_ATTR(tsc_art_ratio, timing_attr_tsc_art_ratio, 1,
168 	       pt_timing_attr_show);
169 
170 static struct attribute *pt_timing_attr[] = {
171 	&timing_attr_max_nonturbo_ratio.attr.attr,
172 	&timing_attr_tsc_art_ratio.attr.attr,
173 	NULL,
174 };
175 
176 static struct attribute_group pt_timing_group = {
177 	.attrs	= pt_timing_attr,
178 };
179 
180 static const struct attribute_group *pt_attr_groups[] = {
181 	&pt_cap_group,
182 	&pt_format_group,
183 	&pt_timing_group,
184 	NULL,
185 };
186 
187 static int __init pt_pmu_hw_init(void)
188 {
189 	struct dev_ext_attribute *de_attrs;
190 	struct attribute **attrs;
191 	size_t size;
192 	u64 reg;
193 	int ret;
194 	long i;
195 
196 	rdmsrl(MSR_PLATFORM_INFO, reg);
197 	pt_pmu.max_nonturbo_ratio = (reg & 0xff00) >> 8;
198 
199 	/*
200 	 * if available, read in TSC to core crystal clock ratio,
201 	 * otherwise, zero for numerator stands for "not enumerated"
202 	 * as per SDM
203 	 */
204 	if (boot_cpu_data.cpuid_level >= CPUID_TSC_LEAF) {
205 		u32 eax, ebx, ecx, edx;
206 
207 		cpuid(CPUID_TSC_LEAF, &eax, &ebx, &ecx, &edx);
208 
209 		pt_pmu.tsc_art_num = ebx;
210 		pt_pmu.tsc_art_den = eax;
211 	}
212 
213 	/* model-specific quirks */
214 	switch (boot_cpu_data.x86_model) {
215 	case INTEL_FAM6_BROADWELL_CORE:
216 	case INTEL_FAM6_BROADWELL_XEON_D:
217 	case INTEL_FAM6_BROADWELL_GT3E:
218 	case INTEL_FAM6_BROADWELL_X:
219 		/* not setting BRANCH_EN will #GP, erratum BDM106 */
220 		pt_pmu.branch_en_always_on = true;
221 		break;
222 	default:
223 		break;
224 	}
225 
226 	if (boot_cpu_has(X86_FEATURE_VMX)) {
227 		/*
228 		 * Intel SDM, 36.5 "Tracing post-VMXON" says that
229 		 * "IA32_VMX_MISC[bit 14]" being 1 means PT can trace
230 		 * post-VMXON.
231 		 */
232 		rdmsrl(MSR_IA32_VMX_MISC, reg);
233 		if (reg & BIT(14))
234 			pt_pmu.vmx = true;
235 	}
236 
237 	attrs = NULL;
238 
239 	for (i = 0; i < PT_CPUID_LEAVES; i++) {
240 		cpuid_count(20, i,
241 			    &pt_pmu.caps[CPUID_EAX + i*PT_CPUID_REGS_NUM],
242 			    &pt_pmu.caps[CPUID_EBX + i*PT_CPUID_REGS_NUM],
243 			    &pt_pmu.caps[CPUID_ECX + i*PT_CPUID_REGS_NUM],
244 			    &pt_pmu.caps[CPUID_EDX + i*PT_CPUID_REGS_NUM]);
245 	}
246 
247 	ret = -ENOMEM;
248 	size = sizeof(struct attribute *) * (ARRAY_SIZE(pt_caps)+1);
249 	attrs = kzalloc(size, GFP_KERNEL);
250 	if (!attrs)
251 		goto fail;
252 
253 	size = sizeof(struct dev_ext_attribute) * (ARRAY_SIZE(pt_caps)+1);
254 	de_attrs = kzalloc(size, GFP_KERNEL);
255 	if (!de_attrs)
256 		goto fail;
257 
258 	for (i = 0; i < ARRAY_SIZE(pt_caps); i++) {
259 		struct dev_ext_attribute *de_attr = de_attrs + i;
260 
261 		de_attr->attr.attr.name = pt_caps[i].name;
262 
263 		sysfs_attr_init(&de_attr->attr.attr);
264 
265 		de_attr->attr.attr.mode		= S_IRUGO;
266 		de_attr->attr.show		= pt_cap_show;
267 		de_attr->var			= (void *)i;
268 
269 		attrs[i] = &de_attr->attr.attr;
270 	}
271 
272 	pt_cap_group.attrs = attrs;
273 
274 	return 0;
275 
276 fail:
277 	kfree(attrs);
278 
279 	return ret;
280 }
281 
282 #define RTIT_CTL_CYC_PSB (RTIT_CTL_CYCLEACC	| \
283 			  RTIT_CTL_CYC_THRESH	| \
284 			  RTIT_CTL_PSB_FREQ)
285 
286 #define RTIT_CTL_MTC	(RTIT_CTL_MTC_EN	| \
287 			 RTIT_CTL_MTC_RANGE)
288 
289 #define RTIT_CTL_PTW	(RTIT_CTL_PTW_EN	| \
290 			 RTIT_CTL_FUP_ON_PTW)
291 
292 /*
293  * Bit 0 (TraceEn) in the attr.config is meaningless as the
294  * corresponding bit in the RTIT_CTL can only be controlled
295  * by the driver; therefore, repurpose it to mean: pass
296  * through the bit that was previously assumed to be always
297  * on for PT, thereby allowing the user to *not* set it if
298  * they so wish. See also pt_event_valid() and pt_config().
299  */
300 #define RTIT_CTL_PASSTHROUGH RTIT_CTL_TRACEEN
301 
302 #define PT_CONFIG_MASK (RTIT_CTL_TRACEEN	| \
303 			RTIT_CTL_TSC_EN		| \
304 			RTIT_CTL_DISRETC	| \
305 			RTIT_CTL_BRANCH_EN	| \
306 			RTIT_CTL_CYC_PSB	| \
307 			RTIT_CTL_MTC		| \
308 			RTIT_CTL_PWR_EVT_EN	| \
309 			RTIT_CTL_FUP_ON_PTW	| \
310 			RTIT_CTL_PTW_EN)
311 
312 static bool pt_event_valid(struct perf_event *event)
313 {
314 	u64 config = event->attr.config;
315 	u64 allowed, requested;
316 
317 	if ((config & PT_CONFIG_MASK) != config)
318 		return false;
319 
320 	if (config & RTIT_CTL_CYC_PSB) {
321 		if (!intel_pt_validate_hw_cap(PT_CAP_psb_cyc))
322 			return false;
323 
324 		allowed = intel_pt_validate_hw_cap(PT_CAP_psb_periods);
325 		requested = (config & RTIT_CTL_PSB_FREQ) >>
326 			RTIT_CTL_PSB_FREQ_OFFSET;
327 		if (requested && (!(allowed & BIT(requested))))
328 			return false;
329 
330 		allowed = intel_pt_validate_hw_cap(PT_CAP_cycle_thresholds);
331 		requested = (config & RTIT_CTL_CYC_THRESH) >>
332 			RTIT_CTL_CYC_THRESH_OFFSET;
333 		if (requested && (!(allowed & BIT(requested))))
334 			return false;
335 	}
336 
337 	if (config & RTIT_CTL_MTC) {
338 		/*
339 		 * In the unlikely case that CPUID lists valid mtc periods,
340 		 * but not the mtc capability, drop out here.
341 		 *
342 		 * Spec says that setting mtc period bits while mtc bit in
343 		 * CPUID is 0 will #GP, so better safe than sorry.
344 		 */
345 		if (!intel_pt_validate_hw_cap(PT_CAP_mtc))
346 			return false;
347 
348 		allowed = intel_pt_validate_hw_cap(PT_CAP_mtc_periods);
349 		if (!allowed)
350 			return false;
351 
352 		requested = (config & RTIT_CTL_MTC_RANGE) >>
353 			RTIT_CTL_MTC_RANGE_OFFSET;
354 
355 		if (!(allowed & BIT(requested)))
356 			return false;
357 	}
358 
359 	if (config & RTIT_CTL_PWR_EVT_EN &&
360 	    !intel_pt_validate_hw_cap(PT_CAP_power_event_trace))
361 		return false;
362 
363 	if (config & RTIT_CTL_PTW) {
364 		if (!intel_pt_validate_hw_cap(PT_CAP_ptwrite))
365 			return false;
366 
367 		/* FUPonPTW without PTW doesn't make sense */
368 		if ((config & RTIT_CTL_FUP_ON_PTW) &&
369 		    !(config & RTIT_CTL_PTW_EN))
370 			return false;
371 	}
372 
373 	/*
374 	 * Setting bit 0 (TraceEn in RTIT_CTL MSR) in the attr.config
375 	 * clears the assomption that BranchEn must always be enabled,
376 	 * as was the case with the first implementation of PT.
377 	 * If this bit is not set, the legacy behavior is preserved
378 	 * for compatibility with the older userspace.
379 	 *
380 	 * Re-using bit 0 for this purpose is fine because it is never
381 	 * directly set by the user; previous attempts at setting it in
382 	 * the attr.config resulted in -EINVAL.
383 	 */
384 	if (config & RTIT_CTL_PASSTHROUGH) {
385 		/*
386 		 * Disallow not setting BRANCH_EN where BRANCH_EN is
387 		 * always required.
388 		 */
389 		if (pt_pmu.branch_en_always_on &&
390 		    !(config & RTIT_CTL_BRANCH_EN))
391 			return false;
392 	} else {
393 		/*
394 		 * Disallow BRANCH_EN without the PASSTHROUGH.
395 		 */
396 		if (config & RTIT_CTL_BRANCH_EN)
397 			return false;
398 	}
399 
400 	return true;
401 }
402 
403 /*
404  * PT configuration helpers
405  * These all are cpu affine and operate on a local PT
406  */
407 
408 /* Address ranges and their corresponding msr configuration registers */
409 static const struct pt_address_range {
410 	unsigned long	msr_a;
411 	unsigned long	msr_b;
412 	unsigned int	reg_off;
413 } pt_address_ranges[] = {
414 	{
415 		.msr_a	 = MSR_IA32_RTIT_ADDR0_A,
416 		.msr_b	 = MSR_IA32_RTIT_ADDR0_B,
417 		.reg_off = RTIT_CTL_ADDR0_OFFSET,
418 	},
419 	{
420 		.msr_a	 = MSR_IA32_RTIT_ADDR1_A,
421 		.msr_b	 = MSR_IA32_RTIT_ADDR1_B,
422 		.reg_off = RTIT_CTL_ADDR1_OFFSET,
423 	},
424 	{
425 		.msr_a	 = MSR_IA32_RTIT_ADDR2_A,
426 		.msr_b	 = MSR_IA32_RTIT_ADDR2_B,
427 		.reg_off = RTIT_CTL_ADDR2_OFFSET,
428 	},
429 	{
430 		.msr_a	 = MSR_IA32_RTIT_ADDR3_A,
431 		.msr_b	 = MSR_IA32_RTIT_ADDR3_B,
432 		.reg_off = RTIT_CTL_ADDR3_OFFSET,
433 	}
434 };
435 
436 static u64 pt_config_filters(struct perf_event *event)
437 {
438 	struct pt_filters *filters = event->hw.addr_filters;
439 	struct pt *pt = this_cpu_ptr(&pt_ctx);
440 	unsigned int range = 0;
441 	u64 rtit_ctl = 0;
442 
443 	if (!filters)
444 		return 0;
445 
446 	perf_event_addr_filters_sync(event);
447 
448 	for (range = 0; range < filters->nr_filters; range++) {
449 		struct pt_filter *filter = &filters->filter[range];
450 
451 		/*
452 		 * Note, if the range has zero start/end addresses due
453 		 * to its dynamic object not being loaded yet, we just
454 		 * go ahead and program zeroed range, which will simply
455 		 * produce no data. Note^2: if executable code at 0x0
456 		 * is a concern, we can set up an "invalid" configuration
457 		 * such as msr_b < msr_a.
458 		 */
459 
460 		/* avoid redundant msr writes */
461 		if (pt->filters.filter[range].msr_a != filter->msr_a) {
462 			wrmsrl(pt_address_ranges[range].msr_a, filter->msr_a);
463 			pt->filters.filter[range].msr_a = filter->msr_a;
464 		}
465 
466 		if (pt->filters.filter[range].msr_b != filter->msr_b) {
467 			wrmsrl(pt_address_ranges[range].msr_b, filter->msr_b);
468 			pt->filters.filter[range].msr_b = filter->msr_b;
469 		}
470 
471 		rtit_ctl |= filter->config << pt_address_ranges[range].reg_off;
472 	}
473 
474 	return rtit_ctl;
475 }
476 
477 static void pt_config(struct perf_event *event)
478 {
479 	struct pt *pt = this_cpu_ptr(&pt_ctx);
480 	u64 reg;
481 
482 	/* First round: clear STATUS, in particular the PSB byte counter. */
483 	if (!event->hw.config) {
484 		perf_event_itrace_started(event);
485 		wrmsrl(MSR_IA32_RTIT_STATUS, 0);
486 	}
487 
488 	reg = pt_config_filters(event);
489 	reg |= RTIT_CTL_TOPA | RTIT_CTL_TRACEEN;
490 
491 	/*
492 	 * Previously, we had BRANCH_EN on by default, but now that PT has
493 	 * grown features outside of branch tracing, it is useful to allow
494 	 * the user to disable it. Setting bit 0 in the event's attr.config
495 	 * allows BRANCH_EN to pass through instead of being always on. See
496 	 * also the comment in pt_event_valid().
497 	 */
498 	if (event->attr.config & BIT(0)) {
499 		reg |= event->attr.config & RTIT_CTL_BRANCH_EN;
500 	} else {
501 		reg |= RTIT_CTL_BRANCH_EN;
502 	}
503 
504 	if (!event->attr.exclude_kernel)
505 		reg |= RTIT_CTL_OS;
506 	if (!event->attr.exclude_user)
507 		reg |= RTIT_CTL_USR;
508 
509 	reg |= (event->attr.config & PT_CONFIG_MASK);
510 
511 	event->hw.config = reg;
512 	if (READ_ONCE(pt->vmx_on))
513 		perf_aux_output_flag(&pt->handle, PERF_AUX_FLAG_PARTIAL);
514 	else
515 		wrmsrl(MSR_IA32_RTIT_CTL, reg);
516 }
517 
518 static void pt_config_stop(struct perf_event *event)
519 {
520 	struct pt *pt = this_cpu_ptr(&pt_ctx);
521 	u64 ctl = READ_ONCE(event->hw.config);
522 
523 	/* may be already stopped by a PMI */
524 	if (!(ctl & RTIT_CTL_TRACEEN))
525 		return;
526 
527 	ctl &= ~RTIT_CTL_TRACEEN;
528 	if (!READ_ONCE(pt->vmx_on))
529 		wrmsrl(MSR_IA32_RTIT_CTL, ctl);
530 
531 	WRITE_ONCE(event->hw.config, ctl);
532 
533 	/*
534 	 * A wrmsr that disables trace generation serializes other PT
535 	 * registers and causes all data packets to be written to memory,
536 	 * but a fence is required for the data to become globally visible.
537 	 *
538 	 * The below WMB, separating data store and aux_head store matches
539 	 * the consumer's RMB that separates aux_head load and data load.
540 	 */
541 	wmb();
542 }
543 
544 static void pt_config_buffer(void *buf, unsigned int topa_idx,
545 			     unsigned int output_off)
546 {
547 	u64 reg;
548 
549 	wrmsrl(MSR_IA32_RTIT_OUTPUT_BASE, virt_to_phys(buf));
550 
551 	reg = 0x7f | ((u64)topa_idx << 7) | ((u64)output_off << 32);
552 
553 	wrmsrl(MSR_IA32_RTIT_OUTPUT_MASK, reg);
554 }
555 
556 /*
557  * Keep ToPA table-related metadata on the same page as the actual table,
558  * taking up a few words from the top
559  */
560 
561 #define TENTS_PER_PAGE (((PAGE_SIZE - 40) / sizeof(struct topa_entry)) - 1)
562 
563 /**
564  * struct topa - page-sized ToPA table with metadata at the top
565  * @table:	actual ToPA table entries, as understood by PT hardware
566  * @list:	linkage to struct pt_buffer's list of tables
567  * @phys:	physical address of this page
568  * @offset:	offset of the first entry in this table in the buffer
569  * @size:	total size of all entries in this table
570  * @last:	index of the last initialized entry in this table
571  */
572 struct topa {
573 	struct topa_entry	table[TENTS_PER_PAGE];
574 	struct list_head	list;
575 	u64			phys;
576 	u64			offset;
577 	size_t			size;
578 	int			last;
579 };
580 
581 /* make -1 stand for the last table entry */
582 #define TOPA_ENTRY(t, i) ((i) == -1 ? &(t)->table[(t)->last] : &(t)->table[(i)])
583 
584 /**
585  * topa_alloc() - allocate page-sized ToPA table
586  * @cpu:	CPU on which to allocate.
587  * @gfp:	Allocation flags.
588  *
589  * Return:	On success, return the pointer to ToPA table page.
590  */
591 static struct topa *topa_alloc(int cpu, gfp_t gfp)
592 {
593 	int node = cpu_to_node(cpu);
594 	struct topa *topa;
595 	struct page *p;
596 
597 	p = alloc_pages_node(node, gfp | __GFP_ZERO, 0);
598 	if (!p)
599 		return NULL;
600 
601 	topa = page_address(p);
602 	topa->last = 0;
603 	topa->phys = page_to_phys(p);
604 
605 	/*
606 	 * In case of singe-entry ToPA, always put the self-referencing END
607 	 * link as the 2nd entry in the table
608 	 */
609 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries)) {
610 		TOPA_ENTRY(topa, 1)->base = topa->phys >> TOPA_SHIFT;
611 		TOPA_ENTRY(topa, 1)->end = 1;
612 	}
613 
614 	return topa;
615 }
616 
617 /**
618  * topa_free() - free a page-sized ToPA table
619  * @topa:	Table to deallocate.
620  */
621 static void topa_free(struct topa *topa)
622 {
623 	free_page((unsigned long)topa);
624 }
625 
626 /**
627  * topa_insert_table() - insert a ToPA table into a buffer
628  * @buf:	 PT buffer that's being extended.
629  * @topa:	 New topa table to be inserted.
630  *
631  * If it's the first table in this buffer, set up buffer's pointers
632  * accordingly; otherwise, add a END=1 link entry to @topa to the current
633  * "last" table and adjust the last table pointer to @topa.
634  */
635 static void topa_insert_table(struct pt_buffer *buf, struct topa *topa)
636 {
637 	struct topa *last = buf->last;
638 
639 	list_add_tail(&topa->list, &buf->tables);
640 
641 	if (!buf->first) {
642 		buf->first = buf->last = buf->cur = topa;
643 		return;
644 	}
645 
646 	topa->offset = last->offset + last->size;
647 	buf->last = topa;
648 
649 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
650 		return;
651 
652 	BUG_ON(last->last != TENTS_PER_PAGE - 1);
653 
654 	TOPA_ENTRY(last, -1)->base = topa->phys >> TOPA_SHIFT;
655 	TOPA_ENTRY(last, -1)->end = 1;
656 }
657 
658 /**
659  * topa_table_full() - check if a ToPA table is filled up
660  * @topa:	ToPA table.
661  */
662 static bool topa_table_full(struct topa *topa)
663 {
664 	/* single-entry ToPA is a special case */
665 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
666 		return !!topa->last;
667 
668 	return topa->last == TENTS_PER_PAGE - 1;
669 }
670 
671 /**
672  * topa_insert_pages() - create a list of ToPA tables
673  * @buf:	PT buffer being initialized.
674  * @gfp:	Allocation flags.
675  *
676  * This initializes a list of ToPA tables with entries from
677  * the data_pages provided by rb_alloc_aux().
678  *
679  * Return:	0 on success or error code.
680  */
681 static int topa_insert_pages(struct pt_buffer *buf, gfp_t gfp)
682 {
683 	struct topa *topa = buf->last;
684 	int order = 0;
685 	struct page *p;
686 
687 	p = virt_to_page(buf->data_pages[buf->nr_pages]);
688 	if (PagePrivate(p))
689 		order = page_private(p);
690 
691 	if (topa_table_full(topa)) {
692 		topa = topa_alloc(buf->cpu, gfp);
693 		if (!topa)
694 			return -ENOMEM;
695 
696 		topa_insert_table(buf, topa);
697 	}
698 
699 	TOPA_ENTRY(topa, -1)->base = page_to_phys(p) >> TOPA_SHIFT;
700 	TOPA_ENTRY(topa, -1)->size = order;
701 	if (!buf->snapshot &&
702 	    !intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries)) {
703 		TOPA_ENTRY(topa, -1)->intr = 1;
704 		TOPA_ENTRY(topa, -1)->stop = 1;
705 	}
706 
707 	topa->last++;
708 	topa->size += sizes(order);
709 
710 	buf->nr_pages += 1ul << order;
711 
712 	return 0;
713 }
714 
715 /**
716  * pt_topa_dump() - print ToPA tables and their entries
717  * @buf:	PT buffer.
718  */
719 static void pt_topa_dump(struct pt_buffer *buf)
720 {
721 	struct topa *topa;
722 
723 	list_for_each_entry(topa, &buf->tables, list) {
724 		int i;
725 
726 		pr_debug("# table @%p (%016Lx), off %llx size %zx\n", topa->table,
727 			 topa->phys, topa->offset, topa->size);
728 		for (i = 0; i < TENTS_PER_PAGE; i++) {
729 			pr_debug("# entry @%p (%lx sz %u %c%c%c) raw=%16llx\n",
730 				 &topa->table[i],
731 				 (unsigned long)topa->table[i].base << TOPA_SHIFT,
732 				 sizes(topa->table[i].size),
733 				 topa->table[i].end ?  'E' : ' ',
734 				 topa->table[i].intr ? 'I' : ' ',
735 				 topa->table[i].stop ? 'S' : ' ',
736 				 *(u64 *)&topa->table[i]);
737 			if ((intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) &&
738 			     topa->table[i].stop) ||
739 			    topa->table[i].end)
740 				break;
741 		}
742 	}
743 }
744 
745 /**
746  * pt_buffer_advance() - advance to the next output region
747  * @buf:	PT buffer.
748  *
749  * Advance the current pointers in the buffer to the next ToPA entry.
750  */
751 static void pt_buffer_advance(struct pt_buffer *buf)
752 {
753 	buf->output_off = 0;
754 	buf->cur_idx++;
755 
756 	if (buf->cur_idx == buf->cur->last) {
757 		if (buf->cur == buf->last)
758 			buf->cur = buf->first;
759 		else
760 			buf->cur = list_entry(buf->cur->list.next, struct topa,
761 					      list);
762 		buf->cur_idx = 0;
763 	}
764 }
765 
766 /**
767  * pt_update_head() - calculate current offsets and sizes
768  * @pt:		Per-cpu pt context.
769  *
770  * Update buffer's current write pointer position and data size.
771  */
772 static void pt_update_head(struct pt *pt)
773 {
774 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
775 	u64 topa_idx, base, old;
776 
777 	/* offset of the first region in this table from the beginning of buf */
778 	base = buf->cur->offset + buf->output_off;
779 
780 	/* offset of the current output region within this table */
781 	for (topa_idx = 0; topa_idx < buf->cur_idx; topa_idx++)
782 		base += sizes(buf->cur->table[topa_idx].size);
783 
784 	if (buf->snapshot) {
785 		local_set(&buf->data_size, base);
786 	} else {
787 		old = (local64_xchg(&buf->head, base) &
788 		       ((buf->nr_pages << PAGE_SHIFT) - 1));
789 		if (base < old)
790 			base += buf->nr_pages << PAGE_SHIFT;
791 
792 		local_add(base - old, &buf->data_size);
793 	}
794 }
795 
796 /**
797  * pt_buffer_region() - obtain current output region's address
798  * @buf:	PT buffer.
799  */
800 static void *pt_buffer_region(struct pt_buffer *buf)
801 {
802 	return phys_to_virt(buf->cur->table[buf->cur_idx].base << TOPA_SHIFT);
803 }
804 
805 /**
806  * pt_buffer_region_size() - obtain current output region's size
807  * @buf:	PT buffer.
808  */
809 static size_t pt_buffer_region_size(struct pt_buffer *buf)
810 {
811 	return sizes(buf->cur->table[buf->cur_idx].size);
812 }
813 
814 /**
815  * pt_handle_status() - take care of possible status conditions
816  * @pt:		Per-cpu pt context.
817  */
818 static void pt_handle_status(struct pt *pt)
819 {
820 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
821 	int advance = 0;
822 	u64 status;
823 
824 	rdmsrl(MSR_IA32_RTIT_STATUS, status);
825 
826 	if (status & RTIT_STATUS_ERROR) {
827 		pr_err_ratelimited("ToPA ERROR encountered, trying to recover\n");
828 		pt_topa_dump(buf);
829 		status &= ~RTIT_STATUS_ERROR;
830 	}
831 
832 	if (status & RTIT_STATUS_STOPPED) {
833 		status &= ~RTIT_STATUS_STOPPED;
834 
835 		/*
836 		 * On systems that only do single-entry ToPA, hitting STOP
837 		 * means we are already losing data; need to let the decoder
838 		 * know.
839 		 */
840 		if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) ||
841 		    buf->output_off == sizes(TOPA_ENTRY(buf->cur, buf->cur_idx)->size)) {
842 			perf_aux_output_flag(&pt->handle,
843 			                     PERF_AUX_FLAG_TRUNCATED);
844 			advance++;
845 		}
846 	}
847 
848 	/*
849 	 * Also on single-entry ToPA implementations, interrupt will come
850 	 * before the output reaches its output region's boundary.
851 	 */
852 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) &&
853 	    !buf->snapshot &&
854 	    pt_buffer_region_size(buf) - buf->output_off <= TOPA_PMI_MARGIN) {
855 		void *head = pt_buffer_region(buf);
856 
857 		/* everything within this margin needs to be zeroed out */
858 		memset(head + buf->output_off, 0,
859 		       pt_buffer_region_size(buf) -
860 		       buf->output_off);
861 		advance++;
862 	}
863 
864 	if (advance)
865 		pt_buffer_advance(buf);
866 
867 	wrmsrl(MSR_IA32_RTIT_STATUS, status);
868 }
869 
870 /**
871  * pt_read_offset() - translate registers into buffer pointers
872  * @buf:	PT buffer.
873  *
874  * Set buffer's output pointers from MSR values.
875  */
876 static void pt_read_offset(struct pt_buffer *buf)
877 {
878 	u64 offset, base_topa;
879 
880 	rdmsrl(MSR_IA32_RTIT_OUTPUT_BASE, base_topa);
881 	buf->cur = phys_to_virt(base_topa);
882 
883 	rdmsrl(MSR_IA32_RTIT_OUTPUT_MASK, offset);
884 	/* offset within current output region */
885 	buf->output_off = offset >> 32;
886 	/* index of current output region within this table */
887 	buf->cur_idx = (offset & 0xffffff80) >> 7;
888 }
889 
890 /**
891  * pt_topa_next_entry() - obtain index of the first page in the next ToPA entry
892  * @buf:	PT buffer.
893  * @pg:		Page offset in the buffer.
894  *
895  * When advancing to the next output region (ToPA entry), given a page offset
896  * into the buffer, we need to find the offset of the first page in the next
897  * region.
898  */
899 static unsigned int pt_topa_next_entry(struct pt_buffer *buf, unsigned int pg)
900 {
901 	struct topa_entry *te = buf->topa_index[pg];
902 
903 	/* one region */
904 	if (buf->first == buf->last && buf->first->last == 1)
905 		return pg;
906 
907 	do {
908 		pg++;
909 		pg &= buf->nr_pages - 1;
910 	} while (buf->topa_index[pg] == te);
911 
912 	return pg;
913 }
914 
915 /**
916  * pt_buffer_reset_markers() - place interrupt and stop bits in the buffer
917  * @buf:	PT buffer.
918  * @handle:	Current output handle.
919  *
920  * Place INT and STOP marks to prevent overwriting old data that the consumer
921  * hasn't yet collected and waking up the consumer after a certain fraction of
922  * the buffer has filled up. Only needed and sensible for non-snapshot counters.
923  *
924  * This obviously relies on buf::head to figure out buffer markers, so it has
925  * to be called after pt_buffer_reset_offsets() and before the hardware tracing
926  * is enabled.
927  */
928 static int pt_buffer_reset_markers(struct pt_buffer *buf,
929 				   struct perf_output_handle *handle)
930 
931 {
932 	unsigned long head = local64_read(&buf->head);
933 	unsigned long idx, npages, wakeup;
934 
935 	/* can't stop in the middle of an output region */
936 	if (buf->output_off + handle->size + 1 <
937 	    sizes(TOPA_ENTRY(buf->cur, buf->cur_idx)->size)) {
938 		perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED);
939 		return -EINVAL;
940 	}
941 
942 
943 	/* single entry ToPA is handled by marking all regions STOP=1 INT=1 */
944 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
945 		return 0;
946 
947 	/* clear STOP and INT from current entry */
948 	buf->topa_index[buf->stop_pos]->stop = 0;
949 	buf->topa_index[buf->stop_pos]->intr = 0;
950 	buf->topa_index[buf->intr_pos]->intr = 0;
951 
952 	/* how many pages till the STOP marker */
953 	npages = handle->size >> PAGE_SHIFT;
954 
955 	/* if it's on a page boundary, fill up one more page */
956 	if (!offset_in_page(head + handle->size + 1))
957 		npages++;
958 
959 	idx = (head >> PAGE_SHIFT) + npages;
960 	idx &= buf->nr_pages - 1;
961 	buf->stop_pos = idx;
962 
963 	wakeup = handle->wakeup >> PAGE_SHIFT;
964 
965 	/* in the worst case, wake up the consumer one page before hard stop */
966 	idx = (head >> PAGE_SHIFT) + npages - 1;
967 	if (idx > wakeup)
968 		idx = wakeup;
969 
970 	idx &= buf->nr_pages - 1;
971 	buf->intr_pos = idx;
972 
973 	buf->topa_index[buf->stop_pos]->stop = 1;
974 	buf->topa_index[buf->stop_pos]->intr = 1;
975 	buf->topa_index[buf->intr_pos]->intr = 1;
976 
977 	return 0;
978 }
979 
980 /**
981  * pt_buffer_setup_topa_index() - build topa_index[] table of regions
982  * @buf:	PT buffer.
983  *
984  * topa_index[] references output regions indexed by offset into the
985  * buffer for purposes of quick reverse lookup.
986  */
987 static void pt_buffer_setup_topa_index(struct pt_buffer *buf)
988 {
989 	struct topa *cur = buf->first, *prev = buf->last;
990 	struct topa_entry *te_cur = TOPA_ENTRY(cur, 0),
991 		*te_prev = TOPA_ENTRY(prev, prev->last - 1);
992 	int pg = 0, idx = 0;
993 
994 	while (pg < buf->nr_pages) {
995 		int tidx;
996 
997 		/* pages within one topa entry */
998 		for (tidx = 0; tidx < 1 << te_cur->size; tidx++, pg++)
999 			buf->topa_index[pg] = te_prev;
1000 
1001 		te_prev = te_cur;
1002 
1003 		if (idx == cur->last - 1) {
1004 			/* advance to next topa table */
1005 			idx = 0;
1006 			cur = list_entry(cur->list.next, struct topa, list);
1007 		} else {
1008 			idx++;
1009 		}
1010 		te_cur = TOPA_ENTRY(cur, idx);
1011 	}
1012 
1013 }
1014 
1015 /**
1016  * pt_buffer_reset_offsets() - adjust buffer's write pointers from aux_head
1017  * @buf:	PT buffer.
1018  * @head:	Write pointer (aux_head) from AUX buffer.
1019  *
1020  * Find the ToPA table and entry corresponding to given @head and set buffer's
1021  * "current" pointers accordingly. This is done after we have obtained the
1022  * current aux_head position from a successful call to perf_aux_output_begin()
1023  * to make sure the hardware is writing to the right place.
1024  *
1025  * This function modifies buf::{cur,cur_idx,output_off} that will be programmed
1026  * into PT msrs when the tracing is enabled and buf::head and buf::data_size,
1027  * which are used to determine INT and STOP markers' locations by a subsequent
1028  * call to pt_buffer_reset_markers().
1029  */
1030 static void pt_buffer_reset_offsets(struct pt_buffer *buf, unsigned long head)
1031 {
1032 	int pg;
1033 
1034 	if (buf->snapshot)
1035 		head &= (buf->nr_pages << PAGE_SHIFT) - 1;
1036 
1037 	pg = (head >> PAGE_SHIFT) & (buf->nr_pages - 1);
1038 	pg = pt_topa_next_entry(buf, pg);
1039 
1040 	buf->cur = (struct topa *)((unsigned long)buf->topa_index[pg] & PAGE_MASK);
1041 	buf->cur_idx = ((unsigned long)buf->topa_index[pg] -
1042 			(unsigned long)buf->cur) / sizeof(struct topa_entry);
1043 	buf->output_off = head & (sizes(buf->cur->table[buf->cur_idx].size) - 1);
1044 
1045 	local64_set(&buf->head, head);
1046 	local_set(&buf->data_size, 0);
1047 }
1048 
1049 /**
1050  * pt_buffer_fini_topa() - deallocate ToPA structure of a buffer
1051  * @buf:	PT buffer.
1052  */
1053 static void pt_buffer_fini_topa(struct pt_buffer *buf)
1054 {
1055 	struct topa *topa, *iter;
1056 
1057 	list_for_each_entry_safe(topa, iter, &buf->tables, list) {
1058 		/*
1059 		 * right now, this is in free_aux() path only, so
1060 		 * no need to unlink this table from the list
1061 		 */
1062 		topa_free(topa);
1063 	}
1064 }
1065 
1066 /**
1067  * pt_buffer_init_topa() - initialize ToPA table for pt buffer
1068  * @buf:	PT buffer.
1069  * @size:	Total size of all regions within this ToPA.
1070  * @gfp:	Allocation flags.
1071  */
1072 static int pt_buffer_init_topa(struct pt_buffer *buf, unsigned long nr_pages,
1073 			       gfp_t gfp)
1074 {
1075 	struct topa *topa;
1076 	int err;
1077 
1078 	topa = topa_alloc(buf->cpu, gfp);
1079 	if (!topa)
1080 		return -ENOMEM;
1081 
1082 	topa_insert_table(buf, topa);
1083 
1084 	while (buf->nr_pages < nr_pages) {
1085 		err = topa_insert_pages(buf, gfp);
1086 		if (err) {
1087 			pt_buffer_fini_topa(buf);
1088 			return -ENOMEM;
1089 		}
1090 	}
1091 
1092 	pt_buffer_setup_topa_index(buf);
1093 
1094 	/* link last table to the first one, unless we're double buffering */
1095 	if (intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries)) {
1096 		TOPA_ENTRY(buf->last, -1)->base = buf->first->phys >> TOPA_SHIFT;
1097 		TOPA_ENTRY(buf->last, -1)->end = 1;
1098 	}
1099 
1100 	pt_topa_dump(buf);
1101 	return 0;
1102 }
1103 
1104 /**
1105  * pt_buffer_setup_aux() - set up topa tables for a PT buffer
1106  * @cpu:	Cpu on which to allocate, -1 means current.
1107  * @pages:	Array of pointers to buffer pages passed from perf core.
1108  * @nr_pages:	Number of pages in the buffer.
1109  * @snapshot:	If this is a snapshot/overwrite counter.
1110  *
1111  * This is a pmu::setup_aux callback that sets up ToPA tables and all the
1112  * bookkeeping for an AUX buffer.
1113  *
1114  * Return:	Our private PT buffer structure.
1115  */
1116 static void *
1117 pt_buffer_setup_aux(struct perf_event *event, void **pages,
1118 		    int nr_pages, bool snapshot)
1119 {
1120 	struct pt_buffer *buf;
1121 	int node, ret, cpu = event->cpu;
1122 
1123 	if (!nr_pages)
1124 		return NULL;
1125 
1126 	if (cpu == -1)
1127 		cpu = raw_smp_processor_id();
1128 	node = cpu_to_node(cpu);
1129 
1130 	buf = kzalloc_node(offsetof(struct pt_buffer, topa_index[nr_pages]),
1131 			   GFP_KERNEL, node);
1132 	if (!buf)
1133 		return NULL;
1134 
1135 	buf->cpu = cpu;
1136 	buf->snapshot = snapshot;
1137 	buf->data_pages = pages;
1138 
1139 	INIT_LIST_HEAD(&buf->tables);
1140 
1141 	ret = pt_buffer_init_topa(buf, nr_pages, GFP_KERNEL);
1142 	if (ret) {
1143 		kfree(buf);
1144 		return NULL;
1145 	}
1146 
1147 	return buf;
1148 }
1149 
1150 /**
1151  * pt_buffer_free_aux() - perf AUX deallocation path callback
1152  * @data:	PT buffer.
1153  */
1154 static void pt_buffer_free_aux(void *data)
1155 {
1156 	struct pt_buffer *buf = data;
1157 
1158 	pt_buffer_fini_topa(buf);
1159 	kfree(buf);
1160 }
1161 
1162 static int pt_addr_filters_init(struct perf_event *event)
1163 {
1164 	struct pt_filters *filters;
1165 	int node = event->cpu == -1 ? -1 : cpu_to_node(event->cpu);
1166 
1167 	if (!intel_pt_validate_hw_cap(PT_CAP_num_address_ranges))
1168 		return 0;
1169 
1170 	filters = kzalloc_node(sizeof(struct pt_filters), GFP_KERNEL, node);
1171 	if (!filters)
1172 		return -ENOMEM;
1173 
1174 	if (event->parent)
1175 		memcpy(filters, event->parent->hw.addr_filters,
1176 		       sizeof(*filters));
1177 
1178 	event->hw.addr_filters = filters;
1179 
1180 	return 0;
1181 }
1182 
1183 static void pt_addr_filters_fini(struct perf_event *event)
1184 {
1185 	kfree(event->hw.addr_filters);
1186 	event->hw.addr_filters = NULL;
1187 }
1188 
1189 static inline bool valid_kernel_ip(unsigned long ip)
1190 {
1191 	return virt_addr_valid(ip) && kernel_ip(ip);
1192 }
1193 
1194 static int pt_event_addr_filters_validate(struct list_head *filters)
1195 {
1196 	struct perf_addr_filter *filter;
1197 	int range = 0;
1198 
1199 	list_for_each_entry(filter, filters, entry) {
1200 		/*
1201 		 * PT doesn't support single address triggers and
1202 		 * 'start' filters.
1203 		 */
1204 		if (!filter->size ||
1205 		    filter->action == PERF_ADDR_FILTER_ACTION_START)
1206 			return -EOPNOTSUPP;
1207 
1208 		if (!filter->path.dentry) {
1209 			if (!valid_kernel_ip(filter->offset))
1210 				return -EINVAL;
1211 
1212 			if (!valid_kernel_ip(filter->offset + filter->size))
1213 				return -EINVAL;
1214 		}
1215 
1216 		if (++range > intel_pt_validate_hw_cap(PT_CAP_num_address_ranges))
1217 			return -EOPNOTSUPP;
1218 	}
1219 
1220 	return 0;
1221 }
1222 
1223 static void pt_event_addr_filters_sync(struct perf_event *event)
1224 {
1225 	struct perf_addr_filters_head *head = perf_event_addr_filters(event);
1226 	unsigned long msr_a, msr_b;
1227 	struct perf_addr_filter_range *fr = event->addr_filter_ranges;
1228 	struct pt_filters *filters = event->hw.addr_filters;
1229 	struct perf_addr_filter *filter;
1230 	int range = 0;
1231 
1232 	if (!filters)
1233 		return;
1234 
1235 	list_for_each_entry(filter, &head->list, entry) {
1236 		if (filter->path.dentry && !fr[range].start) {
1237 			msr_a = msr_b = 0;
1238 		} else {
1239 			/* apply the offset */
1240 			msr_a = fr[range].start;
1241 			msr_b = msr_a + fr[range].size - 1;
1242 		}
1243 
1244 		filters->filter[range].msr_a  = msr_a;
1245 		filters->filter[range].msr_b  = msr_b;
1246 		if (filter->action == PERF_ADDR_FILTER_ACTION_FILTER)
1247 			filters->filter[range].config = 1;
1248 		else
1249 			filters->filter[range].config = 2;
1250 		range++;
1251 	}
1252 
1253 	filters->nr_filters = range;
1254 }
1255 
1256 /**
1257  * intel_pt_interrupt() - PT PMI handler
1258  */
1259 void intel_pt_interrupt(void)
1260 {
1261 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1262 	struct pt_buffer *buf;
1263 	struct perf_event *event = pt->handle.event;
1264 
1265 	/*
1266 	 * There may be a dangling PT bit in the interrupt status register
1267 	 * after PT has been disabled by pt_event_stop(). Make sure we don't
1268 	 * do anything (particularly, re-enable) for this event here.
1269 	 */
1270 	if (!READ_ONCE(pt->handle_nmi))
1271 		return;
1272 
1273 	if (!event)
1274 		return;
1275 
1276 	pt_config_stop(event);
1277 
1278 	buf = perf_get_aux(&pt->handle);
1279 	if (!buf)
1280 		return;
1281 
1282 	pt_read_offset(buf);
1283 
1284 	pt_handle_status(pt);
1285 
1286 	pt_update_head(pt);
1287 
1288 	perf_aux_output_end(&pt->handle, local_xchg(&buf->data_size, 0));
1289 
1290 	if (!event->hw.state) {
1291 		int ret;
1292 
1293 		buf = perf_aux_output_begin(&pt->handle, event);
1294 		if (!buf) {
1295 			event->hw.state = PERF_HES_STOPPED;
1296 			return;
1297 		}
1298 
1299 		pt_buffer_reset_offsets(buf, pt->handle.head);
1300 		/* snapshot counters don't use PMI, so it's safe */
1301 		ret = pt_buffer_reset_markers(buf, &pt->handle);
1302 		if (ret) {
1303 			perf_aux_output_end(&pt->handle, 0);
1304 			return;
1305 		}
1306 
1307 		pt_config_buffer(buf->cur->table, buf->cur_idx,
1308 				 buf->output_off);
1309 		pt_config(event);
1310 	}
1311 }
1312 
1313 void intel_pt_handle_vmx(int on)
1314 {
1315 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1316 	struct perf_event *event;
1317 	unsigned long flags;
1318 
1319 	/* PT plays nice with VMX, do nothing */
1320 	if (pt_pmu.vmx)
1321 		return;
1322 
1323 	/*
1324 	 * VMXON will clear RTIT_CTL.TraceEn; we need to make
1325 	 * sure to not try to set it while VMX is on. Disable
1326 	 * interrupts to avoid racing with pmu callbacks;
1327 	 * concurrent PMI should be handled fine.
1328 	 */
1329 	local_irq_save(flags);
1330 	WRITE_ONCE(pt->vmx_on, on);
1331 
1332 	/*
1333 	 * If an AUX transaction is in progress, it will contain
1334 	 * gap(s), so flag it PARTIAL to inform the user.
1335 	 */
1336 	event = pt->handle.event;
1337 	if (event)
1338 		perf_aux_output_flag(&pt->handle,
1339 		                     PERF_AUX_FLAG_PARTIAL);
1340 
1341 	/* Turn PTs back on */
1342 	if (!on && event)
1343 		wrmsrl(MSR_IA32_RTIT_CTL, event->hw.config);
1344 
1345 	local_irq_restore(flags);
1346 }
1347 EXPORT_SYMBOL_GPL(intel_pt_handle_vmx);
1348 
1349 /*
1350  * PMU callbacks
1351  */
1352 
1353 static void pt_event_start(struct perf_event *event, int mode)
1354 {
1355 	struct hw_perf_event *hwc = &event->hw;
1356 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1357 	struct pt_buffer *buf;
1358 
1359 	buf = perf_aux_output_begin(&pt->handle, event);
1360 	if (!buf)
1361 		goto fail_stop;
1362 
1363 	pt_buffer_reset_offsets(buf, pt->handle.head);
1364 	if (!buf->snapshot) {
1365 		if (pt_buffer_reset_markers(buf, &pt->handle))
1366 			goto fail_end_stop;
1367 	}
1368 
1369 	WRITE_ONCE(pt->handle_nmi, 1);
1370 	hwc->state = 0;
1371 
1372 	pt_config_buffer(buf->cur->table, buf->cur_idx,
1373 			 buf->output_off);
1374 	pt_config(event);
1375 
1376 	return;
1377 
1378 fail_end_stop:
1379 	perf_aux_output_end(&pt->handle, 0);
1380 fail_stop:
1381 	hwc->state = PERF_HES_STOPPED;
1382 }
1383 
1384 static void pt_event_stop(struct perf_event *event, int mode)
1385 {
1386 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1387 
1388 	/*
1389 	 * Protect against the PMI racing with disabling wrmsr,
1390 	 * see comment in intel_pt_interrupt().
1391 	 */
1392 	WRITE_ONCE(pt->handle_nmi, 0);
1393 
1394 	pt_config_stop(event);
1395 
1396 	if (event->hw.state == PERF_HES_STOPPED)
1397 		return;
1398 
1399 	event->hw.state = PERF_HES_STOPPED;
1400 
1401 	if (mode & PERF_EF_UPDATE) {
1402 		struct pt_buffer *buf = perf_get_aux(&pt->handle);
1403 
1404 		if (!buf)
1405 			return;
1406 
1407 		if (WARN_ON_ONCE(pt->handle.event != event))
1408 			return;
1409 
1410 		pt_read_offset(buf);
1411 
1412 		pt_handle_status(pt);
1413 
1414 		pt_update_head(pt);
1415 
1416 		if (buf->snapshot)
1417 			pt->handle.head =
1418 				local_xchg(&buf->data_size,
1419 					   buf->nr_pages << PAGE_SHIFT);
1420 		perf_aux_output_end(&pt->handle, local_xchg(&buf->data_size, 0));
1421 	}
1422 }
1423 
1424 static void pt_event_del(struct perf_event *event, int mode)
1425 {
1426 	pt_event_stop(event, PERF_EF_UPDATE);
1427 }
1428 
1429 static int pt_event_add(struct perf_event *event, int mode)
1430 {
1431 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1432 	struct hw_perf_event *hwc = &event->hw;
1433 	int ret = -EBUSY;
1434 
1435 	if (pt->handle.event)
1436 		goto fail;
1437 
1438 	if (mode & PERF_EF_START) {
1439 		pt_event_start(event, 0);
1440 		ret = -EINVAL;
1441 		if (hwc->state == PERF_HES_STOPPED)
1442 			goto fail;
1443 	} else {
1444 		hwc->state = PERF_HES_STOPPED;
1445 	}
1446 
1447 	ret = 0;
1448 fail:
1449 
1450 	return ret;
1451 }
1452 
1453 static void pt_event_read(struct perf_event *event)
1454 {
1455 }
1456 
1457 static void pt_event_destroy(struct perf_event *event)
1458 {
1459 	pt_addr_filters_fini(event);
1460 	x86_del_exclusive(x86_lbr_exclusive_pt);
1461 }
1462 
1463 static int pt_event_init(struct perf_event *event)
1464 {
1465 	if (event->attr.type != pt_pmu.pmu.type)
1466 		return -ENOENT;
1467 
1468 	if (!pt_event_valid(event))
1469 		return -EINVAL;
1470 
1471 	if (x86_add_exclusive(x86_lbr_exclusive_pt))
1472 		return -EBUSY;
1473 
1474 	if (pt_addr_filters_init(event)) {
1475 		x86_del_exclusive(x86_lbr_exclusive_pt);
1476 		return -ENOMEM;
1477 	}
1478 
1479 	event->destroy = pt_event_destroy;
1480 
1481 	return 0;
1482 }
1483 
1484 void cpu_emergency_stop_pt(void)
1485 {
1486 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1487 
1488 	if (pt->handle.event)
1489 		pt_event_stop(pt->handle.event, PERF_EF_UPDATE);
1490 }
1491 
1492 static __init int pt_init(void)
1493 {
1494 	int ret, cpu, prior_warn = 0;
1495 
1496 	BUILD_BUG_ON(sizeof(struct topa) > PAGE_SIZE);
1497 
1498 	if (!boot_cpu_has(X86_FEATURE_INTEL_PT))
1499 		return -ENODEV;
1500 
1501 	get_online_cpus();
1502 	for_each_online_cpu(cpu) {
1503 		u64 ctl;
1504 
1505 		ret = rdmsrl_safe_on_cpu(cpu, MSR_IA32_RTIT_CTL, &ctl);
1506 		if (!ret && (ctl & RTIT_CTL_TRACEEN))
1507 			prior_warn++;
1508 	}
1509 	put_online_cpus();
1510 
1511 	if (prior_warn) {
1512 		x86_add_exclusive(x86_lbr_exclusive_pt);
1513 		pr_warn("PT is enabled at boot time, doing nothing\n");
1514 
1515 		return -EBUSY;
1516 	}
1517 
1518 	ret = pt_pmu_hw_init();
1519 	if (ret)
1520 		return ret;
1521 
1522 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_output)) {
1523 		pr_warn("ToPA output is not supported on this CPU\n");
1524 		return -ENODEV;
1525 	}
1526 
1527 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
1528 		pt_pmu.pmu.capabilities = PERF_PMU_CAP_AUX_NO_SG;
1529 
1530 	pt_pmu.pmu.capabilities	|= PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE;
1531 	pt_pmu.pmu.attr_groups		 = pt_attr_groups;
1532 	pt_pmu.pmu.task_ctx_nr		 = perf_sw_context;
1533 	pt_pmu.pmu.event_init		 = pt_event_init;
1534 	pt_pmu.pmu.add			 = pt_event_add;
1535 	pt_pmu.pmu.del			 = pt_event_del;
1536 	pt_pmu.pmu.start		 = pt_event_start;
1537 	pt_pmu.pmu.stop			 = pt_event_stop;
1538 	pt_pmu.pmu.read			 = pt_event_read;
1539 	pt_pmu.pmu.setup_aux		 = pt_buffer_setup_aux;
1540 	pt_pmu.pmu.free_aux		 = pt_buffer_free_aux;
1541 	pt_pmu.pmu.addr_filters_sync     = pt_event_addr_filters_sync;
1542 	pt_pmu.pmu.addr_filters_validate = pt_event_addr_filters_validate;
1543 	pt_pmu.pmu.nr_addr_filters       =
1544 		intel_pt_validate_hw_cap(PT_CAP_num_address_ranges);
1545 
1546 	ret = perf_pmu_register(&pt_pmu.pmu, "intel_pt", -1);
1547 
1548 	return ret;
1549 }
1550 arch_initcall(pt_init);
1551