xref: /linux/arch/x86/events/intel/pt.c (revision b0f84a84fff180718995b1269da2988e5b28be42)
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/slab.h>
17 #include <linux/device.h>
18 
19 #include <asm/perf_event.h>
20 #include <asm/insn.h>
21 #include <asm/io.h>
22 #include <asm/intel_pt.h>
23 #include <asm/intel-family.h>
24 
25 #include "../perf_event.h"
26 #include "pt.h"
27 
28 static DEFINE_PER_CPU(struct pt, pt_ctx);
29 
30 static struct pt_pmu pt_pmu;
31 
32 /*
33  * Capabilities of Intel PT hardware, such as number of address bits or
34  * supported output schemes, are cached and exported to userspace as "caps"
35  * attribute group of pt pmu device
36  * (/sys/bus/event_source/devices/intel_pt/caps/) so that userspace can store
37  * relevant bits together with intel_pt traces.
38  *
39  * These are necessary for both trace decoding (payloads_lip, contains address
40  * width encoded in IP-related packets), and event configuration (bitmasks with
41  * permitted values for certain bit fields).
42  */
43 #define PT_CAP(_n, _l, _r, _m)						\
44 	[PT_CAP_ ## _n] = { .name = __stringify(_n), .leaf = _l,	\
45 			    .reg = _r, .mask = _m }
46 
47 static struct pt_cap_desc {
48 	const char	*name;
49 	u32		leaf;
50 	u8		reg;
51 	u32		mask;
52 } pt_caps[] = {
53 	PT_CAP(max_subleaf,		0, CPUID_EAX, 0xffffffff),
54 	PT_CAP(cr3_filtering,		0, CPUID_EBX, BIT(0)),
55 	PT_CAP(psb_cyc,			0, CPUID_EBX, BIT(1)),
56 	PT_CAP(ip_filtering,		0, CPUID_EBX, BIT(2)),
57 	PT_CAP(mtc,			0, CPUID_EBX, BIT(3)),
58 	PT_CAP(ptwrite,			0, CPUID_EBX, BIT(4)),
59 	PT_CAP(power_event_trace,	0, CPUID_EBX, BIT(5)),
60 	PT_CAP(topa_output,		0, CPUID_ECX, BIT(0)),
61 	PT_CAP(topa_multiple_entries,	0, CPUID_ECX, BIT(1)),
62 	PT_CAP(single_range_output,	0, CPUID_ECX, BIT(2)),
63 	PT_CAP(output_subsys,		0, CPUID_ECX, BIT(3)),
64 	PT_CAP(payloads_lip,		0, CPUID_ECX, BIT(31)),
65 	PT_CAP(num_address_ranges,	1, CPUID_EAX, 0x3),
66 	PT_CAP(mtc_periods,		1, CPUID_EAX, 0xffff0000),
67 	PT_CAP(cycle_thresholds,	1, CPUID_EBX, 0xffff),
68 	PT_CAP(psb_periods,		1, CPUID_EBX, 0xffff0000),
69 };
70 
71 u32 intel_pt_validate_cap(u32 *caps, enum pt_capabilities capability)
72 {
73 	struct pt_cap_desc *cd = &pt_caps[capability];
74 	u32 c = caps[cd->leaf * PT_CPUID_REGS_NUM + cd->reg];
75 	unsigned int shift = __ffs(cd->mask);
76 
77 	return (c & cd->mask) >> shift;
78 }
79 EXPORT_SYMBOL_GPL(intel_pt_validate_cap);
80 
81 u32 intel_pt_validate_hw_cap(enum pt_capabilities cap)
82 {
83 	return intel_pt_validate_cap(pt_pmu.caps, cap);
84 }
85 EXPORT_SYMBOL_GPL(intel_pt_validate_hw_cap);
86 
87 static ssize_t pt_cap_show(struct device *cdev,
88 			   struct device_attribute *attr,
89 			   char *buf)
90 {
91 	struct dev_ext_attribute *ea =
92 		container_of(attr, struct dev_ext_attribute, attr);
93 	enum pt_capabilities cap = (long)ea->var;
94 
95 	return snprintf(buf, PAGE_SIZE, "%x\n", intel_pt_validate_hw_cap(cap));
96 }
97 
98 static struct attribute_group pt_cap_group __ro_after_init = {
99 	.name	= "caps",
100 };
101 
102 PMU_FORMAT_ATTR(pt,		"config:0"	);
103 PMU_FORMAT_ATTR(cyc,		"config:1"	);
104 PMU_FORMAT_ATTR(pwr_evt,	"config:4"	);
105 PMU_FORMAT_ATTR(fup_on_ptw,	"config:5"	);
106 PMU_FORMAT_ATTR(mtc,		"config:9"	);
107 PMU_FORMAT_ATTR(tsc,		"config:10"	);
108 PMU_FORMAT_ATTR(noretcomp,	"config:11"	);
109 PMU_FORMAT_ATTR(ptw,		"config:12"	);
110 PMU_FORMAT_ATTR(branch,		"config:13"	);
111 PMU_FORMAT_ATTR(mtc_period,	"config:14-17"	);
112 PMU_FORMAT_ATTR(cyc_thresh,	"config:19-22"	);
113 PMU_FORMAT_ATTR(psb_period,	"config:24-27"	);
114 
115 static struct attribute *pt_formats_attr[] = {
116 	&format_attr_pt.attr,
117 	&format_attr_cyc.attr,
118 	&format_attr_pwr_evt.attr,
119 	&format_attr_fup_on_ptw.attr,
120 	&format_attr_mtc.attr,
121 	&format_attr_tsc.attr,
122 	&format_attr_noretcomp.attr,
123 	&format_attr_ptw.attr,
124 	&format_attr_branch.attr,
125 	&format_attr_mtc_period.attr,
126 	&format_attr_cyc_thresh.attr,
127 	&format_attr_psb_period.attr,
128 	NULL,
129 };
130 
131 static struct attribute_group pt_format_group = {
132 	.name	= "format",
133 	.attrs	= pt_formats_attr,
134 };
135 
136 static ssize_t
137 pt_timing_attr_show(struct device *dev, struct device_attribute *attr,
138 		    char *page)
139 {
140 	struct perf_pmu_events_attr *pmu_attr =
141 		container_of(attr, struct perf_pmu_events_attr, attr);
142 
143 	switch (pmu_attr->id) {
144 	case 0:
145 		return sprintf(page, "%lu\n", pt_pmu.max_nonturbo_ratio);
146 	case 1:
147 		return sprintf(page, "%u:%u\n",
148 			       pt_pmu.tsc_art_num,
149 			       pt_pmu.tsc_art_den);
150 	default:
151 		break;
152 	}
153 
154 	return -EINVAL;
155 }
156 
157 PMU_EVENT_ATTR(max_nonturbo_ratio, timing_attr_max_nonturbo_ratio, 0,
158 	       pt_timing_attr_show);
159 PMU_EVENT_ATTR(tsc_art_ratio, timing_attr_tsc_art_ratio, 1,
160 	       pt_timing_attr_show);
161 
162 static struct attribute *pt_timing_attr[] = {
163 	&timing_attr_max_nonturbo_ratio.attr.attr,
164 	&timing_attr_tsc_art_ratio.attr.attr,
165 	NULL,
166 };
167 
168 static struct attribute_group pt_timing_group = {
169 	.attrs	= pt_timing_attr,
170 };
171 
172 static const struct attribute_group *pt_attr_groups[] = {
173 	&pt_cap_group,
174 	&pt_format_group,
175 	&pt_timing_group,
176 	NULL,
177 };
178 
179 static int __init pt_pmu_hw_init(void)
180 {
181 	struct dev_ext_attribute *de_attrs;
182 	struct attribute **attrs;
183 	size_t size;
184 	u64 reg;
185 	int ret;
186 	long i;
187 
188 	rdmsrl(MSR_PLATFORM_INFO, reg);
189 	pt_pmu.max_nonturbo_ratio = (reg & 0xff00) >> 8;
190 
191 	/*
192 	 * if available, read in TSC to core crystal clock ratio,
193 	 * otherwise, zero for numerator stands for "not enumerated"
194 	 * as per SDM
195 	 */
196 	if (boot_cpu_data.cpuid_level >= CPUID_TSC_LEAF) {
197 		u32 eax, ebx, ecx, edx;
198 
199 		cpuid(CPUID_TSC_LEAF, &eax, &ebx, &ecx, &edx);
200 
201 		pt_pmu.tsc_art_num = ebx;
202 		pt_pmu.tsc_art_den = eax;
203 	}
204 
205 	/* model-specific quirks */
206 	switch (boot_cpu_data.x86_model) {
207 	case INTEL_FAM6_BROADWELL_CORE:
208 	case INTEL_FAM6_BROADWELL_XEON_D:
209 	case INTEL_FAM6_BROADWELL_GT3E:
210 	case INTEL_FAM6_BROADWELL_X:
211 		/* not setting BRANCH_EN will #GP, erratum BDM106 */
212 		pt_pmu.branch_en_always_on = true;
213 		break;
214 	default:
215 		break;
216 	}
217 
218 	if (boot_cpu_has(X86_FEATURE_VMX)) {
219 		/*
220 		 * Intel SDM, 36.5 "Tracing post-VMXON" says that
221 		 * "IA32_VMX_MISC[bit 14]" being 1 means PT can trace
222 		 * post-VMXON.
223 		 */
224 		rdmsrl(MSR_IA32_VMX_MISC, reg);
225 		if (reg & BIT(14))
226 			pt_pmu.vmx = true;
227 	}
228 
229 	attrs = NULL;
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 
268 fail:
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 
304 static 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 
400 /* Address ranges and their corresponding msr configuration registers */
401 static const struct pt_address_range {
402 	unsigned long	msr_a;
403 	unsigned long	msr_b;
404 	unsigned int	reg_off;
405 } pt_address_ranges[] = {
406 	{
407 		.msr_a	 = MSR_IA32_RTIT_ADDR0_A,
408 		.msr_b	 = MSR_IA32_RTIT_ADDR0_B,
409 		.reg_off = RTIT_CTL_ADDR0_OFFSET,
410 	},
411 	{
412 		.msr_a	 = MSR_IA32_RTIT_ADDR1_A,
413 		.msr_b	 = MSR_IA32_RTIT_ADDR1_B,
414 		.reg_off = RTIT_CTL_ADDR1_OFFSET,
415 	},
416 	{
417 		.msr_a	 = MSR_IA32_RTIT_ADDR2_A,
418 		.msr_b	 = MSR_IA32_RTIT_ADDR2_B,
419 		.reg_off = RTIT_CTL_ADDR2_OFFSET,
420 	},
421 	{
422 		.msr_a	 = MSR_IA32_RTIT_ADDR3_A,
423 		.msr_b	 = MSR_IA32_RTIT_ADDR3_B,
424 		.reg_off = RTIT_CTL_ADDR3_OFFSET,
425 	}
426 };
427 
428 static u64 pt_config_filters(struct perf_event *event)
429 {
430 	struct pt_filters *filters = event->hw.addr_filters;
431 	struct pt *pt = this_cpu_ptr(&pt_ctx);
432 	unsigned int range = 0;
433 	u64 rtit_ctl = 0;
434 
435 	if (!filters)
436 		return 0;
437 
438 	perf_event_addr_filters_sync(event);
439 
440 	for (range = 0; range < filters->nr_filters; range++) {
441 		struct pt_filter *filter = &filters->filter[range];
442 
443 		/*
444 		 * Note, if the range has zero start/end addresses due
445 		 * to its dynamic object not being loaded yet, we just
446 		 * go ahead and program zeroed range, which will simply
447 		 * produce no data. Note^2: if executable code at 0x0
448 		 * is a concern, we can set up an "invalid" configuration
449 		 * such as msr_b < msr_a.
450 		 */
451 
452 		/* avoid redundant msr writes */
453 		if (pt->filters.filter[range].msr_a != filter->msr_a) {
454 			wrmsrl(pt_address_ranges[range].msr_a, filter->msr_a);
455 			pt->filters.filter[range].msr_a = filter->msr_a;
456 		}
457 
458 		if (pt->filters.filter[range].msr_b != filter->msr_b) {
459 			wrmsrl(pt_address_ranges[range].msr_b, filter->msr_b);
460 			pt->filters.filter[range].msr_b = filter->msr_b;
461 		}
462 
463 		rtit_ctl |= filter->config << pt_address_ranges[range].reg_off;
464 	}
465 
466 	return rtit_ctl;
467 }
468 
469 static void pt_config(struct perf_event *event)
470 {
471 	struct pt *pt = this_cpu_ptr(&pt_ctx);
472 	u64 reg;
473 
474 	/* First round: clear STATUS, in particular the PSB byte counter. */
475 	if (!event->hw.config) {
476 		perf_event_itrace_started(event);
477 		wrmsrl(MSR_IA32_RTIT_STATUS, 0);
478 	}
479 
480 	reg = pt_config_filters(event);
481 	reg |= RTIT_CTL_TOPA | RTIT_CTL_TRACEEN;
482 
483 	/*
484 	 * Previously, we had BRANCH_EN on by default, but now that PT has
485 	 * grown features outside of branch tracing, it is useful to allow
486 	 * the user to disable it. Setting bit 0 in the event's attr.config
487 	 * allows BRANCH_EN to pass through instead of being always on. See
488 	 * also the comment in pt_event_valid().
489 	 */
490 	if (event->attr.config & BIT(0)) {
491 		reg |= event->attr.config & RTIT_CTL_BRANCH_EN;
492 	} else {
493 		reg |= RTIT_CTL_BRANCH_EN;
494 	}
495 
496 	if (!event->attr.exclude_kernel)
497 		reg |= RTIT_CTL_OS;
498 	if (!event->attr.exclude_user)
499 		reg |= RTIT_CTL_USR;
500 
501 	reg |= (event->attr.config & PT_CONFIG_MASK);
502 
503 	event->hw.config = reg;
504 	if (READ_ONCE(pt->vmx_on))
505 		perf_aux_output_flag(&pt->handle, PERF_AUX_FLAG_PARTIAL);
506 	else
507 		wrmsrl(MSR_IA32_RTIT_CTL, reg);
508 }
509 
510 static void pt_config_stop(struct perf_event *event)
511 {
512 	struct pt *pt = this_cpu_ptr(&pt_ctx);
513 	u64 ctl = READ_ONCE(event->hw.config);
514 
515 	/* may be already stopped by a PMI */
516 	if (!(ctl & RTIT_CTL_TRACEEN))
517 		return;
518 
519 	ctl &= ~RTIT_CTL_TRACEEN;
520 	if (!READ_ONCE(pt->vmx_on))
521 		wrmsrl(MSR_IA32_RTIT_CTL, ctl);
522 
523 	WRITE_ONCE(event->hw.config, ctl);
524 
525 	/*
526 	 * A wrmsr that disables trace generation serializes other PT
527 	 * registers and causes all data packets to be written to memory,
528 	 * but a fence is required for the data to become globally visible.
529 	 *
530 	 * The below WMB, separating data store and aux_head store matches
531 	 * the consumer's RMB that separates aux_head load and data load.
532 	 */
533 	wmb();
534 }
535 
536 static void pt_config_buffer(void *buf, unsigned int topa_idx,
537 			     unsigned int output_off)
538 {
539 	u64 reg;
540 
541 	wrmsrl(MSR_IA32_RTIT_OUTPUT_BASE, virt_to_phys(buf));
542 
543 	reg = 0x7f | ((u64)topa_idx << 7) | ((u64)output_off << 32);
544 
545 	wrmsrl(MSR_IA32_RTIT_OUTPUT_MASK, reg);
546 }
547 
548 /*
549  * Keep ToPA table-related metadata on the same page as the actual table,
550  * taking up a few words from the top
551  */
552 
553 #define TENTS_PER_PAGE (((PAGE_SIZE - 40) / sizeof(struct topa_entry)) - 1)
554 
555 /**
556  * struct topa - page-sized ToPA table with metadata at the top
557  * @table:	actual ToPA table entries, as understood by PT hardware
558  * @list:	linkage to struct pt_buffer's list of tables
559  * @phys:	physical address of this page
560  * @offset:	offset of the first entry in this table in the buffer
561  * @size:	total size of all entries in this table
562  * @last:	index of the last initialized entry in this table
563  */
564 struct topa {
565 	struct topa_entry	table[TENTS_PER_PAGE];
566 	struct list_head	list;
567 	u64			phys;
568 	u64			offset;
569 	size_t			size;
570 	int			last;
571 };
572 
573 /* make -1 stand for the last table entry */
574 #define TOPA_ENTRY(t, i) ((i) == -1 ? &(t)->table[(t)->last] : &(t)->table[(i)])
575 
576 /**
577  * topa_alloc() - allocate page-sized ToPA table
578  * @cpu:	CPU on which to allocate.
579  * @gfp:	Allocation flags.
580  *
581  * Return:	On success, return the pointer to ToPA table page.
582  */
583 static struct topa *topa_alloc(int cpu, gfp_t gfp)
584 {
585 	int node = cpu_to_node(cpu);
586 	struct topa *topa;
587 	struct page *p;
588 
589 	p = alloc_pages_node(node, gfp | __GFP_ZERO, 0);
590 	if (!p)
591 		return NULL;
592 
593 	topa = page_address(p);
594 	topa->last = 0;
595 	topa->phys = page_to_phys(p);
596 
597 	/*
598 	 * In case of singe-entry ToPA, always put the self-referencing END
599 	 * link as the 2nd entry in the table
600 	 */
601 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries)) {
602 		TOPA_ENTRY(topa, 1)->base = topa->phys >> TOPA_SHIFT;
603 		TOPA_ENTRY(topa, 1)->end = 1;
604 	}
605 
606 	return topa;
607 }
608 
609 /**
610  * topa_free() - free a page-sized ToPA table
611  * @topa:	Table to deallocate.
612  */
613 static void topa_free(struct topa *topa)
614 {
615 	free_page((unsigned long)topa);
616 }
617 
618 /**
619  * topa_insert_table() - insert a ToPA table into a buffer
620  * @buf:	 PT buffer that's being extended.
621  * @topa:	 New topa table to be inserted.
622  *
623  * If it's the first table in this buffer, set up buffer's pointers
624  * accordingly; otherwise, add a END=1 link entry to @topa to the current
625  * "last" table and adjust the last table pointer to @topa.
626  */
627 static void topa_insert_table(struct pt_buffer *buf, struct topa *topa)
628 {
629 	struct topa *last = buf->last;
630 
631 	list_add_tail(&topa->list, &buf->tables);
632 
633 	if (!buf->first) {
634 		buf->first = buf->last = buf->cur = topa;
635 		return;
636 	}
637 
638 	topa->offset = last->offset + last->size;
639 	buf->last = topa;
640 
641 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
642 		return;
643 
644 	BUG_ON(last->last != TENTS_PER_PAGE - 1);
645 
646 	TOPA_ENTRY(last, -1)->base = topa->phys >> TOPA_SHIFT;
647 	TOPA_ENTRY(last, -1)->end = 1;
648 }
649 
650 /**
651  * topa_table_full() - check if a ToPA table is filled up
652  * @topa:	ToPA table.
653  */
654 static bool topa_table_full(struct topa *topa)
655 {
656 	/* single-entry ToPA is a special case */
657 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
658 		return !!topa->last;
659 
660 	return topa->last == TENTS_PER_PAGE - 1;
661 }
662 
663 /**
664  * topa_insert_pages() - create a list of ToPA tables
665  * @buf:	PT buffer being initialized.
666  * @gfp:	Allocation flags.
667  *
668  * This initializes a list of ToPA tables with entries from
669  * the data_pages provided by rb_alloc_aux().
670  *
671  * Return:	0 on success or error code.
672  */
673 static int topa_insert_pages(struct pt_buffer *buf, gfp_t gfp)
674 {
675 	struct topa *topa = buf->last;
676 	int order = 0;
677 	struct page *p;
678 
679 	p = virt_to_page(buf->data_pages[buf->nr_pages]);
680 	if (PagePrivate(p))
681 		order = page_private(p);
682 
683 	if (topa_table_full(topa)) {
684 		topa = topa_alloc(buf->cpu, gfp);
685 		if (!topa)
686 			return -ENOMEM;
687 
688 		topa_insert_table(buf, topa);
689 	}
690 
691 	TOPA_ENTRY(topa, -1)->base = page_to_phys(p) >> TOPA_SHIFT;
692 	TOPA_ENTRY(topa, -1)->size = order;
693 	if (!buf->snapshot &&
694 	    !intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries)) {
695 		TOPA_ENTRY(topa, -1)->intr = 1;
696 		TOPA_ENTRY(topa, -1)->stop = 1;
697 	}
698 
699 	topa->last++;
700 	topa->size += sizes(order);
701 
702 	buf->nr_pages += 1ul << order;
703 
704 	return 0;
705 }
706 
707 /**
708  * pt_topa_dump() - print ToPA tables and their entries
709  * @buf:	PT buffer.
710  */
711 static void pt_topa_dump(struct pt_buffer *buf)
712 {
713 	struct topa *topa;
714 
715 	list_for_each_entry(topa, &buf->tables, list) {
716 		int i;
717 
718 		pr_debug("# table @%p (%016Lx), off %llx size %zx\n", topa->table,
719 			 topa->phys, topa->offset, topa->size);
720 		for (i = 0; i < TENTS_PER_PAGE; i++) {
721 			pr_debug("# entry @%p (%lx sz %u %c%c%c) raw=%16llx\n",
722 				 &topa->table[i],
723 				 (unsigned long)topa->table[i].base << TOPA_SHIFT,
724 				 sizes(topa->table[i].size),
725 				 topa->table[i].end ?  'E' : ' ',
726 				 topa->table[i].intr ? 'I' : ' ',
727 				 topa->table[i].stop ? 'S' : ' ',
728 				 *(u64 *)&topa->table[i]);
729 			if ((intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) &&
730 			     topa->table[i].stop) ||
731 			    topa->table[i].end)
732 				break;
733 		}
734 	}
735 }
736 
737 /**
738  * pt_buffer_advance() - advance to the next output region
739  * @buf:	PT buffer.
740  *
741  * Advance the current pointers in the buffer to the next ToPA entry.
742  */
743 static void pt_buffer_advance(struct pt_buffer *buf)
744 {
745 	buf->output_off = 0;
746 	buf->cur_idx++;
747 
748 	if (buf->cur_idx == buf->cur->last) {
749 		if (buf->cur == buf->last)
750 			buf->cur = buf->first;
751 		else
752 			buf->cur = list_entry(buf->cur->list.next, struct topa,
753 					      list);
754 		buf->cur_idx = 0;
755 	}
756 }
757 
758 /**
759  * pt_update_head() - calculate current offsets and sizes
760  * @pt:		Per-cpu pt context.
761  *
762  * Update buffer's current write pointer position and data size.
763  */
764 static void pt_update_head(struct pt *pt)
765 {
766 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
767 	u64 topa_idx, base, old;
768 
769 	/* offset of the first region in this table from the beginning of buf */
770 	base = buf->cur->offset + buf->output_off;
771 
772 	/* offset of the current output region within this table */
773 	for (topa_idx = 0; topa_idx < buf->cur_idx; topa_idx++)
774 		base += sizes(buf->cur->table[topa_idx].size);
775 
776 	if (buf->snapshot) {
777 		local_set(&buf->data_size, base);
778 	} else {
779 		old = (local64_xchg(&buf->head, base) &
780 		       ((buf->nr_pages << PAGE_SHIFT) - 1));
781 		if (base < old)
782 			base += buf->nr_pages << PAGE_SHIFT;
783 
784 		local_add(base - old, &buf->data_size);
785 	}
786 }
787 
788 /**
789  * pt_buffer_region() - obtain current output region's address
790  * @buf:	PT buffer.
791  */
792 static void *pt_buffer_region(struct pt_buffer *buf)
793 {
794 	return phys_to_virt(buf->cur->table[buf->cur_idx].base << TOPA_SHIFT);
795 }
796 
797 /**
798  * pt_buffer_region_size() - obtain current output region's size
799  * @buf:	PT buffer.
800  */
801 static size_t pt_buffer_region_size(struct pt_buffer *buf)
802 {
803 	return sizes(buf->cur->table[buf->cur_idx].size);
804 }
805 
806 /**
807  * pt_handle_status() - take care of possible status conditions
808  * @pt:		Per-cpu pt context.
809  */
810 static void pt_handle_status(struct pt *pt)
811 {
812 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
813 	int advance = 0;
814 	u64 status;
815 
816 	rdmsrl(MSR_IA32_RTIT_STATUS, status);
817 
818 	if (status & RTIT_STATUS_ERROR) {
819 		pr_err_ratelimited("ToPA ERROR encountered, trying to recover\n");
820 		pt_topa_dump(buf);
821 		status &= ~RTIT_STATUS_ERROR;
822 	}
823 
824 	if (status & RTIT_STATUS_STOPPED) {
825 		status &= ~RTIT_STATUS_STOPPED;
826 
827 		/*
828 		 * On systems that only do single-entry ToPA, hitting STOP
829 		 * means we are already losing data; need to let the decoder
830 		 * know.
831 		 */
832 		if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) ||
833 		    buf->output_off == sizes(TOPA_ENTRY(buf->cur, buf->cur_idx)->size)) {
834 			perf_aux_output_flag(&pt->handle,
835 			                     PERF_AUX_FLAG_TRUNCATED);
836 			advance++;
837 		}
838 	}
839 
840 	/*
841 	 * Also on single-entry ToPA implementations, interrupt will come
842 	 * before the output reaches its output region's boundary.
843 	 */
844 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) &&
845 	    !buf->snapshot &&
846 	    pt_buffer_region_size(buf) - buf->output_off <= TOPA_PMI_MARGIN) {
847 		void *head = pt_buffer_region(buf);
848 
849 		/* everything within this margin needs to be zeroed out */
850 		memset(head + buf->output_off, 0,
851 		       pt_buffer_region_size(buf) -
852 		       buf->output_off);
853 		advance++;
854 	}
855 
856 	if (advance)
857 		pt_buffer_advance(buf);
858 
859 	wrmsrl(MSR_IA32_RTIT_STATUS, status);
860 }
861 
862 /**
863  * pt_read_offset() - translate registers into buffer pointers
864  * @buf:	PT buffer.
865  *
866  * Set buffer's output pointers from MSR values.
867  */
868 static void pt_read_offset(struct pt_buffer *buf)
869 {
870 	u64 offset, base_topa;
871 
872 	rdmsrl(MSR_IA32_RTIT_OUTPUT_BASE, base_topa);
873 	buf->cur = phys_to_virt(base_topa);
874 
875 	rdmsrl(MSR_IA32_RTIT_OUTPUT_MASK, offset);
876 	/* offset within current output region */
877 	buf->output_off = offset >> 32;
878 	/* index of current output region within this table */
879 	buf->cur_idx = (offset & 0xffffff80) >> 7;
880 }
881 
882 /**
883  * pt_topa_next_entry() - obtain index of the first page in the next ToPA entry
884  * @buf:	PT buffer.
885  * @pg:		Page offset in the buffer.
886  *
887  * When advancing to the next output region (ToPA entry), given a page offset
888  * into the buffer, we need to find the offset of the first page in the next
889  * region.
890  */
891 static unsigned int pt_topa_next_entry(struct pt_buffer *buf, unsigned int pg)
892 {
893 	struct topa_entry *te = buf->topa_index[pg];
894 
895 	/* one region */
896 	if (buf->first == buf->last && buf->first->last == 1)
897 		return pg;
898 
899 	do {
900 		pg++;
901 		pg &= buf->nr_pages - 1;
902 	} while (buf->topa_index[pg] == te);
903 
904 	return pg;
905 }
906 
907 /**
908  * pt_buffer_reset_markers() - place interrupt and stop bits in the buffer
909  * @buf:	PT buffer.
910  * @handle:	Current output handle.
911  *
912  * Place INT and STOP marks to prevent overwriting old data that the consumer
913  * hasn't yet collected and waking up the consumer after a certain fraction of
914  * the buffer has filled up. Only needed and sensible for non-snapshot counters.
915  *
916  * This obviously relies on buf::head to figure out buffer markers, so it has
917  * to be called after pt_buffer_reset_offsets() and before the hardware tracing
918  * is enabled.
919  */
920 static int pt_buffer_reset_markers(struct pt_buffer *buf,
921 				   struct perf_output_handle *handle)
922 
923 {
924 	unsigned long head = local64_read(&buf->head);
925 	unsigned long idx, npages, wakeup;
926 
927 	/* can't stop in the middle of an output region */
928 	if (buf->output_off + handle->size + 1 <
929 	    sizes(TOPA_ENTRY(buf->cur, buf->cur_idx)->size)) {
930 		perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED);
931 		return -EINVAL;
932 	}
933 
934 
935 	/* single entry ToPA is handled by marking all regions STOP=1 INT=1 */
936 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
937 		return 0;
938 
939 	/* clear STOP and INT from current entry */
940 	buf->topa_index[buf->stop_pos]->stop = 0;
941 	buf->topa_index[buf->stop_pos]->intr = 0;
942 	buf->topa_index[buf->intr_pos]->intr = 0;
943 
944 	/* how many pages till the STOP marker */
945 	npages = handle->size >> PAGE_SHIFT;
946 
947 	/* if it's on a page boundary, fill up one more page */
948 	if (!offset_in_page(head + handle->size + 1))
949 		npages++;
950 
951 	idx = (head >> PAGE_SHIFT) + npages;
952 	idx &= buf->nr_pages - 1;
953 	buf->stop_pos = idx;
954 
955 	wakeup = handle->wakeup >> PAGE_SHIFT;
956 
957 	/* in the worst case, wake up the consumer one page before hard stop */
958 	idx = (head >> PAGE_SHIFT) + npages - 1;
959 	if (idx > wakeup)
960 		idx = wakeup;
961 
962 	idx &= buf->nr_pages - 1;
963 	buf->intr_pos = idx;
964 
965 	buf->topa_index[buf->stop_pos]->stop = 1;
966 	buf->topa_index[buf->stop_pos]->intr = 1;
967 	buf->topa_index[buf->intr_pos]->intr = 1;
968 
969 	return 0;
970 }
971 
972 /**
973  * pt_buffer_setup_topa_index() - build topa_index[] table of regions
974  * @buf:	PT buffer.
975  *
976  * topa_index[] references output regions indexed by offset into the
977  * buffer for purposes of quick reverse lookup.
978  */
979 static void pt_buffer_setup_topa_index(struct pt_buffer *buf)
980 {
981 	struct topa *cur = buf->first, *prev = buf->last;
982 	struct topa_entry *te_cur = TOPA_ENTRY(cur, 0),
983 		*te_prev = TOPA_ENTRY(prev, prev->last - 1);
984 	int pg = 0, idx = 0;
985 
986 	while (pg < buf->nr_pages) {
987 		int tidx;
988 
989 		/* pages within one topa entry */
990 		for (tidx = 0; tidx < 1 << te_cur->size; tidx++, pg++)
991 			buf->topa_index[pg] = te_prev;
992 
993 		te_prev = te_cur;
994 
995 		if (idx == cur->last - 1) {
996 			/* advance to next topa table */
997 			idx = 0;
998 			cur = list_entry(cur->list.next, struct topa, list);
999 		} else {
1000 			idx++;
1001 		}
1002 		te_cur = TOPA_ENTRY(cur, idx);
1003 	}
1004 
1005 }
1006 
1007 /**
1008  * pt_buffer_reset_offsets() - adjust buffer's write pointers from aux_head
1009  * @buf:	PT buffer.
1010  * @head:	Write pointer (aux_head) from AUX buffer.
1011  *
1012  * Find the ToPA table and entry corresponding to given @head and set buffer's
1013  * "current" pointers accordingly. This is done after we have obtained the
1014  * current aux_head position from a successful call to perf_aux_output_begin()
1015  * to make sure the hardware is writing to the right place.
1016  *
1017  * This function modifies buf::{cur,cur_idx,output_off} that will be programmed
1018  * into PT msrs when the tracing is enabled and buf::head and buf::data_size,
1019  * which are used to determine INT and STOP markers' locations by a subsequent
1020  * call to pt_buffer_reset_markers().
1021  */
1022 static void pt_buffer_reset_offsets(struct pt_buffer *buf, unsigned long head)
1023 {
1024 	int pg;
1025 
1026 	if (buf->snapshot)
1027 		head &= (buf->nr_pages << PAGE_SHIFT) - 1;
1028 
1029 	pg = (head >> PAGE_SHIFT) & (buf->nr_pages - 1);
1030 	pg = pt_topa_next_entry(buf, pg);
1031 
1032 	buf->cur = (struct topa *)((unsigned long)buf->topa_index[pg] & PAGE_MASK);
1033 	buf->cur_idx = ((unsigned long)buf->topa_index[pg] -
1034 			(unsigned long)buf->cur) / sizeof(struct topa_entry);
1035 	buf->output_off = head & (sizes(buf->cur->table[buf->cur_idx].size) - 1);
1036 
1037 	local64_set(&buf->head, head);
1038 	local_set(&buf->data_size, 0);
1039 }
1040 
1041 /**
1042  * pt_buffer_fini_topa() - deallocate ToPA structure of a buffer
1043  * @buf:	PT buffer.
1044  */
1045 static void pt_buffer_fini_topa(struct pt_buffer *buf)
1046 {
1047 	struct topa *topa, *iter;
1048 
1049 	list_for_each_entry_safe(topa, iter, &buf->tables, list) {
1050 		/*
1051 		 * right now, this is in free_aux() path only, so
1052 		 * no need to unlink this table from the list
1053 		 */
1054 		topa_free(topa);
1055 	}
1056 }
1057 
1058 /**
1059  * pt_buffer_init_topa() - initialize ToPA table for pt buffer
1060  * @buf:	PT buffer.
1061  * @size:	Total size of all regions within this ToPA.
1062  * @gfp:	Allocation flags.
1063  */
1064 static int pt_buffer_init_topa(struct pt_buffer *buf, unsigned long nr_pages,
1065 			       gfp_t gfp)
1066 {
1067 	struct topa *topa;
1068 	int err;
1069 
1070 	topa = topa_alloc(buf->cpu, gfp);
1071 	if (!topa)
1072 		return -ENOMEM;
1073 
1074 	topa_insert_table(buf, topa);
1075 
1076 	while (buf->nr_pages < nr_pages) {
1077 		err = topa_insert_pages(buf, gfp);
1078 		if (err) {
1079 			pt_buffer_fini_topa(buf);
1080 			return -ENOMEM;
1081 		}
1082 	}
1083 
1084 	pt_buffer_setup_topa_index(buf);
1085 
1086 	/* link last table to the first one, unless we're double buffering */
1087 	if (intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries)) {
1088 		TOPA_ENTRY(buf->last, -1)->base = buf->first->phys >> TOPA_SHIFT;
1089 		TOPA_ENTRY(buf->last, -1)->end = 1;
1090 	}
1091 
1092 	pt_topa_dump(buf);
1093 	return 0;
1094 }
1095 
1096 /**
1097  * pt_buffer_setup_aux() - set up topa tables for a PT buffer
1098  * @cpu:	Cpu on which to allocate, -1 means current.
1099  * @pages:	Array of pointers to buffer pages passed from perf core.
1100  * @nr_pages:	Number of pages in the buffer.
1101  * @snapshot:	If this is a snapshot/overwrite counter.
1102  *
1103  * This is a pmu::setup_aux callback that sets up ToPA tables and all the
1104  * bookkeeping for an AUX buffer.
1105  *
1106  * Return:	Our private PT buffer structure.
1107  */
1108 static void *
1109 pt_buffer_setup_aux(struct perf_event *event, void **pages,
1110 		    int nr_pages, bool snapshot)
1111 {
1112 	struct pt_buffer *buf;
1113 	int node, ret, cpu = event->cpu;
1114 
1115 	if (!nr_pages)
1116 		return NULL;
1117 
1118 	if (cpu == -1)
1119 		cpu = raw_smp_processor_id();
1120 	node = cpu_to_node(cpu);
1121 
1122 	buf = kzalloc_node(offsetof(struct pt_buffer, topa_index[nr_pages]),
1123 			   GFP_KERNEL, node);
1124 	if (!buf)
1125 		return NULL;
1126 
1127 	buf->cpu = cpu;
1128 	buf->snapshot = snapshot;
1129 	buf->data_pages = pages;
1130 
1131 	INIT_LIST_HEAD(&buf->tables);
1132 
1133 	ret = pt_buffer_init_topa(buf, nr_pages, GFP_KERNEL);
1134 	if (ret) {
1135 		kfree(buf);
1136 		return NULL;
1137 	}
1138 
1139 	return buf;
1140 }
1141 
1142 /**
1143  * pt_buffer_free_aux() - perf AUX deallocation path callback
1144  * @data:	PT buffer.
1145  */
1146 static void pt_buffer_free_aux(void *data)
1147 {
1148 	struct pt_buffer *buf = data;
1149 
1150 	pt_buffer_fini_topa(buf);
1151 	kfree(buf);
1152 }
1153 
1154 static int pt_addr_filters_init(struct perf_event *event)
1155 {
1156 	struct pt_filters *filters;
1157 	int node = event->cpu == -1 ? -1 : cpu_to_node(event->cpu);
1158 
1159 	if (!intel_pt_validate_hw_cap(PT_CAP_num_address_ranges))
1160 		return 0;
1161 
1162 	filters = kzalloc_node(sizeof(struct pt_filters), GFP_KERNEL, node);
1163 	if (!filters)
1164 		return -ENOMEM;
1165 
1166 	if (event->parent)
1167 		memcpy(filters, event->parent->hw.addr_filters,
1168 		       sizeof(*filters));
1169 
1170 	event->hw.addr_filters = filters;
1171 
1172 	return 0;
1173 }
1174 
1175 static void pt_addr_filters_fini(struct perf_event *event)
1176 {
1177 	kfree(event->hw.addr_filters);
1178 	event->hw.addr_filters = NULL;
1179 }
1180 
1181 static inline bool valid_kernel_ip(unsigned long ip)
1182 {
1183 	return virt_addr_valid(ip) && kernel_ip(ip);
1184 }
1185 
1186 static int pt_event_addr_filters_validate(struct list_head *filters)
1187 {
1188 	struct perf_addr_filter *filter;
1189 	int range = 0;
1190 
1191 	list_for_each_entry(filter, filters, entry) {
1192 		/*
1193 		 * PT doesn't support single address triggers and
1194 		 * 'start' filters.
1195 		 */
1196 		if (!filter->size ||
1197 		    filter->action == PERF_ADDR_FILTER_ACTION_START)
1198 			return -EOPNOTSUPP;
1199 
1200 		if (!filter->path.dentry) {
1201 			if (!valid_kernel_ip(filter->offset))
1202 				return -EINVAL;
1203 
1204 			if (!valid_kernel_ip(filter->offset + filter->size))
1205 				return -EINVAL;
1206 		}
1207 
1208 		if (++range > intel_pt_validate_hw_cap(PT_CAP_num_address_ranges))
1209 			return -EOPNOTSUPP;
1210 	}
1211 
1212 	return 0;
1213 }
1214 
1215 static void pt_event_addr_filters_sync(struct perf_event *event)
1216 {
1217 	struct perf_addr_filters_head *head = perf_event_addr_filters(event);
1218 	unsigned long msr_a, msr_b;
1219 	struct perf_addr_filter_range *fr = event->addr_filter_ranges;
1220 	struct pt_filters *filters = event->hw.addr_filters;
1221 	struct perf_addr_filter *filter;
1222 	int range = 0;
1223 
1224 	if (!filters)
1225 		return;
1226 
1227 	list_for_each_entry(filter, &head->list, entry) {
1228 		if (filter->path.dentry && !fr[range].start) {
1229 			msr_a = msr_b = 0;
1230 		} else {
1231 			/* apply the offset */
1232 			msr_a = fr[range].start;
1233 			msr_b = msr_a + fr[range].size - 1;
1234 		}
1235 
1236 		filters->filter[range].msr_a  = msr_a;
1237 		filters->filter[range].msr_b  = msr_b;
1238 		if (filter->action == PERF_ADDR_FILTER_ACTION_FILTER)
1239 			filters->filter[range].config = 1;
1240 		else
1241 			filters->filter[range].config = 2;
1242 		range++;
1243 	}
1244 
1245 	filters->nr_filters = range;
1246 }
1247 
1248 /**
1249  * intel_pt_interrupt() - PT PMI handler
1250  */
1251 void intel_pt_interrupt(void)
1252 {
1253 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1254 	struct pt_buffer *buf;
1255 	struct perf_event *event = pt->handle.event;
1256 
1257 	/*
1258 	 * There may be a dangling PT bit in the interrupt status register
1259 	 * after PT has been disabled by pt_event_stop(). Make sure we don't
1260 	 * do anything (particularly, re-enable) for this event here.
1261 	 */
1262 	if (!READ_ONCE(pt->handle_nmi))
1263 		return;
1264 
1265 	if (!event)
1266 		return;
1267 
1268 	pt_config_stop(event);
1269 
1270 	buf = perf_get_aux(&pt->handle);
1271 	if (!buf)
1272 		return;
1273 
1274 	pt_read_offset(buf);
1275 
1276 	pt_handle_status(pt);
1277 
1278 	pt_update_head(pt);
1279 
1280 	perf_aux_output_end(&pt->handle, local_xchg(&buf->data_size, 0));
1281 
1282 	if (!event->hw.state) {
1283 		int ret;
1284 
1285 		buf = perf_aux_output_begin(&pt->handle, event);
1286 		if (!buf) {
1287 			event->hw.state = PERF_HES_STOPPED;
1288 			return;
1289 		}
1290 
1291 		pt_buffer_reset_offsets(buf, pt->handle.head);
1292 		/* snapshot counters don't use PMI, so it's safe */
1293 		ret = pt_buffer_reset_markers(buf, &pt->handle);
1294 		if (ret) {
1295 			perf_aux_output_end(&pt->handle, 0);
1296 			return;
1297 		}
1298 
1299 		pt_config_buffer(buf->cur->table, buf->cur_idx,
1300 				 buf->output_off);
1301 		pt_config(event);
1302 	}
1303 }
1304 
1305 void intel_pt_handle_vmx(int on)
1306 {
1307 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1308 	struct perf_event *event;
1309 	unsigned long flags;
1310 
1311 	/* PT plays nice with VMX, do nothing */
1312 	if (pt_pmu.vmx)
1313 		return;
1314 
1315 	/*
1316 	 * VMXON will clear RTIT_CTL.TraceEn; we need to make
1317 	 * sure to not try to set it while VMX is on. Disable
1318 	 * interrupts to avoid racing with pmu callbacks;
1319 	 * concurrent PMI should be handled fine.
1320 	 */
1321 	local_irq_save(flags);
1322 	WRITE_ONCE(pt->vmx_on, on);
1323 
1324 	/*
1325 	 * If an AUX transaction is in progress, it will contain
1326 	 * gap(s), so flag it PARTIAL to inform the user.
1327 	 */
1328 	event = pt->handle.event;
1329 	if (event)
1330 		perf_aux_output_flag(&pt->handle,
1331 		                     PERF_AUX_FLAG_PARTIAL);
1332 
1333 	/* Turn PTs back on */
1334 	if (!on && event)
1335 		wrmsrl(MSR_IA32_RTIT_CTL, event->hw.config);
1336 
1337 	local_irq_restore(flags);
1338 }
1339 EXPORT_SYMBOL_GPL(intel_pt_handle_vmx);
1340 
1341 /*
1342  * PMU callbacks
1343  */
1344 
1345 static void pt_event_start(struct perf_event *event, int mode)
1346 {
1347 	struct hw_perf_event *hwc = &event->hw;
1348 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1349 	struct pt_buffer *buf;
1350 
1351 	buf = perf_aux_output_begin(&pt->handle, event);
1352 	if (!buf)
1353 		goto fail_stop;
1354 
1355 	pt_buffer_reset_offsets(buf, pt->handle.head);
1356 	if (!buf->snapshot) {
1357 		if (pt_buffer_reset_markers(buf, &pt->handle))
1358 			goto fail_end_stop;
1359 	}
1360 
1361 	WRITE_ONCE(pt->handle_nmi, 1);
1362 	hwc->state = 0;
1363 
1364 	pt_config_buffer(buf->cur->table, buf->cur_idx,
1365 			 buf->output_off);
1366 	pt_config(event);
1367 
1368 	return;
1369 
1370 fail_end_stop:
1371 	perf_aux_output_end(&pt->handle, 0);
1372 fail_stop:
1373 	hwc->state = PERF_HES_STOPPED;
1374 }
1375 
1376 static void pt_event_stop(struct perf_event *event, int mode)
1377 {
1378 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1379 
1380 	/*
1381 	 * Protect against the PMI racing with disabling wrmsr,
1382 	 * see comment in intel_pt_interrupt().
1383 	 */
1384 	WRITE_ONCE(pt->handle_nmi, 0);
1385 
1386 	pt_config_stop(event);
1387 
1388 	if (event->hw.state == PERF_HES_STOPPED)
1389 		return;
1390 
1391 	event->hw.state = PERF_HES_STOPPED;
1392 
1393 	if (mode & PERF_EF_UPDATE) {
1394 		struct pt_buffer *buf = perf_get_aux(&pt->handle);
1395 
1396 		if (!buf)
1397 			return;
1398 
1399 		if (WARN_ON_ONCE(pt->handle.event != event))
1400 			return;
1401 
1402 		pt_read_offset(buf);
1403 
1404 		pt_handle_status(pt);
1405 
1406 		pt_update_head(pt);
1407 
1408 		if (buf->snapshot)
1409 			pt->handle.head =
1410 				local_xchg(&buf->data_size,
1411 					   buf->nr_pages << PAGE_SHIFT);
1412 		perf_aux_output_end(&pt->handle, local_xchg(&buf->data_size, 0));
1413 	}
1414 }
1415 
1416 static void pt_event_del(struct perf_event *event, int mode)
1417 {
1418 	pt_event_stop(event, PERF_EF_UPDATE);
1419 }
1420 
1421 static int pt_event_add(struct perf_event *event, int mode)
1422 {
1423 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1424 	struct hw_perf_event *hwc = &event->hw;
1425 	int ret = -EBUSY;
1426 
1427 	if (pt->handle.event)
1428 		goto fail;
1429 
1430 	if (mode & PERF_EF_START) {
1431 		pt_event_start(event, 0);
1432 		ret = -EINVAL;
1433 		if (hwc->state == PERF_HES_STOPPED)
1434 			goto fail;
1435 	} else {
1436 		hwc->state = PERF_HES_STOPPED;
1437 	}
1438 
1439 	ret = 0;
1440 fail:
1441 
1442 	return ret;
1443 }
1444 
1445 static void pt_event_read(struct perf_event *event)
1446 {
1447 }
1448 
1449 static void pt_event_destroy(struct perf_event *event)
1450 {
1451 	pt_addr_filters_fini(event);
1452 	x86_del_exclusive(x86_lbr_exclusive_pt);
1453 }
1454 
1455 static int pt_event_init(struct perf_event *event)
1456 {
1457 	if (event->attr.type != pt_pmu.pmu.type)
1458 		return -ENOENT;
1459 
1460 	if (!pt_event_valid(event))
1461 		return -EINVAL;
1462 
1463 	if (x86_add_exclusive(x86_lbr_exclusive_pt))
1464 		return -EBUSY;
1465 
1466 	if (pt_addr_filters_init(event)) {
1467 		x86_del_exclusive(x86_lbr_exclusive_pt);
1468 		return -ENOMEM;
1469 	}
1470 
1471 	event->destroy = pt_event_destroy;
1472 
1473 	return 0;
1474 }
1475 
1476 void cpu_emergency_stop_pt(void)
1477 {
1478 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1479 
1480 	if (pt->handle.event)
1481 		pt_event_stop(pt->handle.event, PERF_EF_UPDATE);
1482 }
1483 
1484 static __init int pt_init(void)
1485 {
1486 	int ret, cpu, prior_warn = 0;
1487 
1488 	BUILD_BUG_ON(sizeof(struct topa) > PAGE_SIZE);
1489 
1490 	if (!boot_cpu_has(X86_FEATURE_INTEL_PT))
1491 		return -ENODEV;
1492 
1493 	get_online_cpus();
1494 	for_each_online_cpu(cpu) {
1495 		u64 ctl;
1496 
1497 		ret = rdmsrl_safe_on_cpu(cpu, MSR_IA32_RTIT_CTL, &ctl);
1498 		if (!ret && (ctl & RTIT_CTL_TRACEEN))
1499 			prior_warn++;
1500 	}
1501 	put_online_cpus();
1502 
1503 	if (prior_warn) {
1504 		x86_add_exclusive(x86_lbr_exclusive_pt);
1505 		pr_warn("PT is enabled at boot time, doing nothing\n");
1506 
1507 		return -EBUSY;
1508 	}
1509 
1510 	ret = pt_pmu_hw_init();
1511 	if (ret)
1512 		return ret;
1513 
1514 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_output)) {
1515 		pr_warn("ToPA output is not supported on this CPU\n");
1516 		return -ENODEV;
1517 	}
1518 
1519 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
1520 		pt_pmu.pmu.capabilities = PERF_PMU_CAP_AUX_NO_SG;
1521 
1522 	pt_pmu.pmu.capabilities	|= PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE;
1523 	pt_pmu.pmu.attr_groups		 = pt_attr_groups;
1524 	pt_pmu.pmu.task_ctx_nr		 = perf_sw_context;
1525 	pt_pmu.pmu.event_init		 = pt_event_init;
1526 	pt_pmu.pmu.add			 = pt_event_add;
1527 	pt_pmu.pmu.del			 = pt_event_del;
1528 	pt_pmu.pmu.start		 = pt_event_start;
1529 	pt_pmu.pmu.stop			 = pt_event_stop;
1530 	pt_pmu.pmu.read			 = pt_event_read;
1531 	pt_pmu.pmu.setup_aux		 = pt_buffer_setup_aux;
1532 	pt_pmu.pmu.free_aux		 = pt_buffer_free_aux;
1533 	pt_pmu.pmu.addr_filters_sync     = pt_event_addr_filters_sync;
1534 	pt_pmu.pmu.addr_filters_validate = pt_event_addr_filters_validate;
1535 	pt_pmu.pmu.nr_addr_filters       =
1536 		intel_pt_validate_hw_cap(PT_CAP_num_address_ranges);
1537 
1538 	ret = perf_pmu_register(&pt_pmu.pmu, "intel_pt", -1);
1539 
1540 	return ret;
1541 }
1542 arch_initcall(pt_init);
1543