xref: /openbmc/linux/arch/x86/events/intel/pt.c (revision 9a29f5fc)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Intel(R) Processor Trace PMU driver for perf
4  * Copyright (c) 2013-2014, Intel Corporation.
5  *
6  * Intel PT is specified in the Intel Architecture Instruction Set Extensions
7  * Programming Reference:
8  * http://software.intel.com/en-us/intel-isa-extensions
9  */
10 
11 #undef DEBUG
12 
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14 
15 #include <linux/types.h>
16 #include <linux/bits.h>
17 #include <linux/limits.h>
18 #include <linux/slab.h>
19 #include <linux/device.h>
20 
21 #include <asm/perf_event.h>
22 #include <asm/insn.h>
23 #include <asm/io.h>
24 #include <asm/intel_pt.h>
25 #include <asm/intel-family.h>
26 
27 #include "../perf_event.h"
28 #include "pt.h"
29 
30 static DEFINE_PER_CPU(struct pt, pt_ctx);
31 
32 static struct pt_pmu pt_pmu;
33 
34 /*
35  * Capabilities of Intel PT hardware, such as number of address bits or
36  * supported output schemes, are cached and exported to userspace as "caps"
37  * attribute group of pt pmu device
38  * (/sys/bus/event_source/devices/intel_pt/caps/) so that userspace can store
39  * relevant bits together with intel_pt traces.
40  *
41  * These are necessary for both trace decoding (payloads_lip, contains address
42  * width encoded in IP-related packets), and event configuration (bitmasks with
43  * permitted values for certain bit fields).
44  */
45 #define PT_CAP(_n, _l, _r, _m)						\
46 	[PT_CAP_ ## _n] = { .name = __stringify(_n), .leaf = _l,	\
47 			    .reg = _r, .mask = _m }
48 
49 static struct pt_cap_desc {
50 	const char	*name;
51 	u32		leaf;
52 	u8		reg;
53 	u32		mask;
54 } pt_caps[] = {
55 	PT_CAP(max_subleaf,		0, CPUID_EAX, 0xffffffff),
56 	PT_CAP(cr3_filtering,		0, CPUID_EBX, BIT(0)),
57 	PT_CAP(psb_cyc,			0, CPUID_EBX, BIT(1)),
58 	PT_CAP(ip_filtering,		0, CPUID_EBX, BIT(2)),
59 	PT_CAP(mtc,			0, CPUID_EBX, BIT(3)),
60 	PT_CAP(ptwrite,			0, CPUID_EBX, BIT(4)),
61 	PT_CAP(power_event_trace,	0, CPUID_EBX, BIT(5)),
62 	PT_CAP(event_trace,		0, CPUID_EBX, BIT(7)),
63 	PT_CAP(tnt_disable,		0, CPUID_EBX, BIT(8)),
64 	PT_CAP(topa_output,		0, CPUID_ECX, BIT(0)),
65 	PT_CAP(topa_multiple_entries,	0, CPUID_ECX, BIT(1)),
66 	PT_CAP(single_range_output,	0, CPUID_ECX, BIT(2)),
67 	PT_CAP(output_subsys,		0, CPUID_ECX, BIT(3)),
68 	PT_CAP(payloads_lip,		0, CPUID_ECX, BIT(31)),
69 	PT_CAP(num_address_ranges,	1, CPUID_EAX, 0x7),
70 	PT_CAP(mtc_periods,		1, CPUID_EAX, 0xffff0000),
71 	PT_CAP(cycle_thresholds,	1, CPUID_EBX, 0xffff),
72 	PT_CAP(psb_periods,		1, CPUID_EBX, 0xffff0000),
73 };
74 
75 u32 intel_pt_validate_cap(u32 *caps, enum pt_capabilities capability)
76 {
77 	struct pt_cap_desc *cd = &pt_caps[capability];
78 	u32 c = caps[cd->leaf * PT_CPUID_REGS_NUM + cd->reg];
79 	unsigned int shift = __ffs(cd->mask);
80 
81 	return (c & cd->mask) >> shift;
82 }
83 EXPORT_SYMBOL_GPL(intel_pt_validate_cap);
84 
85 u32 intel_pt_validate_hw_cap(enum pt_capabilities cap)
86 {
87 	return intel_pt_validate_cap(pt_pmu.caps, cap);
88 }
89 EXPORT_SYMBOL_GPL(intel_pt_validate_hw_cap);
90 
91 static ssize_t pt_cap_show(struct device *cdev,
92 			   struct device_attribute *attr,
93 			   char *buf)
94 {
95 	struct dev_ext_attribute *ea =
96 		container_of(attr, struct dev_ext_attribute, attr);
97 	enum pt_capabilities cap = (long)ea->var;
98 
99 	return snprintf(buf, PAGE_SIZE, "%x\n", intel_pt_validate_hw_cap(cap));
100 }
101 
102 static struct attribute_group pt_cap_group __ro_after_init = {
103 	.name	= "caps",
104 };
105 
106 PMU_FORMAT_ATTR(pt,		"config:0"	);
107 PMU_FORMAT_ATTR(cyc,		"config:1"	);
108 PMU_FORMAT_ATTR(pwr_evt,	"config:4"	);
109 PMU_FORMAT_ATTR(fup_on_ptw,	"config:5"	);
110 PMU_FORMAT_ATTR(mtc,		"config:9"	);
111 PMU_FORMAT_ATTR(tsc,		"config:10"	);
112 PMU_FORMAT_ATTR(noretcomp,	"config:11"	);
113 PMU_FORMAT_ATTR(ptw,		"config:12"	);
114 PMU_FORMAT_ATTR(branch,		"config:13"	);
115 PMU_FORMAT_ATTR(event,		"config:31"	);
116 PMU_FORMAT_ATTR(notnt,		"config:55"	);
117 PMU_FORMAT_ATTR(mtc_period,	"config:14-17"	);
118 PMU_FORMAT_ATTR(cyc_thresh,	"config:19-22"	);
119 PMU_FORMAT_ATTR(psb_period,	"config:24-27"	);
120 
121 static struct attribute *pt_formats_attr[] = {
122 	&format_attr_pt.attr,
123 	&format_attr_cyc.attr,
124 	&format_attr_pwr_evt.attr,
125 	&format_attr_event.attr,
126 	&format_attr_notnt.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:
216 	case INTEL_FAM6_BROADWELL_D:
217 	case INTEL_FAM6_BROADWELL_G:
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 	for (i = 0; i < PT_CPUID_LEAVES; i++) {
238 		cpuid_count(20, i,
239 			    &pt_pmu.caps[CPUID_EAX + i*PT_CPUID_REGS_NUM],
240 			    &pt_pmu.caps[CPUID_EBX + i*PT_CPUID_REGS_NUM],
241 			    &pt_pmu.caps[CPUID_ECX + i*PT_CPUID_REGS_NUM],
242 			    &pt_pmu.caps[CPUID_EDX + i*PT_CPUID_REGS_NUM]);
243 	}
244 
245 	ret = -ENOMEM;
246 	size = sizeof(struct attribute *) * (ARRAY_SIZE(pt_caps)+1);
247 	attrs = kzalloc(size, GFP_KERNEL);
248 	if (!attrs)
249 		goto fail;
250 
251 	size = sizeof(struct dev_ext_attribute) * (ARRAY_SIZE(pt_caps)+1);
252 	de_attrs = kzalloc(size, GFP_KERNEL);
253 	if (!de_attrs)
254 		goto fail;
255 
256 	for (i = 0; i < ARRAY_SIZE(pt_caps); i++) {
257 		struct dev_ext_attribute *de_attr = de_attrs + i;
258 
259 		de_attr->attr.attr.name = pt_caps[i].name;
260 
261 		sysfs_attr_init(&de_attr->attr.attr);
262 
263 		de_attr->attr.attr.mode		= S_IRUGO;
264 		de_attr->attr.show		= pt_cap_show;
265 		de_attr->var			= (void *)i;
266 
267 		attrs[i] = &de_attr->attr.attr;
268 	}
269 
270 	pt_cap_group.attrs = attrs;
271 
272 	return 0;
273 
274 fail:
275 	kfree(attrs);
276 
277 	return ret;
278 }
279 
280 #define RTIT_CTL_CYC_PSB (RTIT_CTL_CYCLEACC	| \
281 			  RTIT_CTL_CYC_THRESH	| \
282 			  RTIT_CTL_PSB_FREQ)
283 
284 #define RTIT_CTL_MTC	(RTIT_CTL_MTC_EN	| \
285 			 RTIT_CTL_MTC_RANGE)
286 
287 #define RTIT_CTL_PTW	(RTIT_CTL_PTW_EN	| \
288 			 RTIT_CTL_FUP_ON_PTW)
289 
290 /*
291  * Bit 0 (TraceEn) in the attr.config is meaningless as the
292  * corresponding bit in the RTIT_CTL can only be controlled
293  * by the driver; therefore, repurpose it to mean: pass
294  * through the bit that was previously assumed to be always
295  * on for PT, thereby allowing the user to *not* set it if
296  * they so wish. See also pt_event_valid() and pt_config().
297  */
298 #define RTIT_CTL_PASSTHROUGH RTIT_CTL_TRACEEN
299 
300 #define PT_CONFIG_MASK (RTIT_CTL_TRACEEN	| \
301 			RTIT_CTL_TSC_EN		| \
302 			RTIT_CTL_DISRETC	| \
303 			RTIT_CTL_BRANCH_EN	| \
304 			RTIT_CTL_CYC_PSB	| \
305 			RTIT_CTL_MTC		| \
306 			RTIT_CTL_PWR_EVT_EN	| \
307 			RTIT_CTL_EVENT_EN	| \
308 			RTIT_CTL_NOTNT		| \
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_EVENT_EN &&
364 	    !intel_pt_validate_hw_cap(PT_CAP_event_trace))
365 		return false;
366 
367 	if (config & RTIT_CTL_NOTNT &&
368 	    !intel_pt_validate_hw_cap(PT_CAP_tnt_disable))
369 		return false;
370 
371 	if (config & RTIT_CTL_PTW) {
372 		if (!intel_pt_validate_hw_cap(PT_CAP_ptwrite))
373 			return false;
374 
375 		/* FUPonPTW without PTW doesn't make sense */
376 		if ((config & RTIT_CTL_FUP_ON_PTW) &&
377 		    !(config & RTIT_CTL_PTW_EN))
378 			return false;
379 	}
380 
381 	/*
382 	 * Setting bit 0 (TraceEn in RTIT_CTL MSR) in the attr.config
383 	 * clears the assumption that BranchEn must always be enabled,
384 	 * as was the case with the first implementation of PT.
385 	 * If this bit is not set, the legacy behavior is preserved
386 	 * for compatibility with the older userspace.
387 	 *
388 	 * Re-using bit 0 for this purpose is fine because it is never
389 	 * directly set by the user; previous attempts at setting it in
390 	 * the attr.config resulted in -EINVAL.
391 	 */
392 	if (config & RTIT_CTL_PASSTHROUGH) {
393 		/*
394 		 * Disallow not setting BRANCH_EN where BRANCH_EN is
395 		 * always required.
396 		 */
397 		if (pt_pmu.branch_en_always_on &&
398 		    !(config & RTIT_CTL_BRANCH_EN))
399 			return false;
400 	} else {
401 		/*
402 		 * Disallow BRANCH_EN without the PASSTHROUGH.
403 		 */
404 		if (config & RTIT_CTL_BRANCH_EN)
405 			return false;
406 	}
407 
408 	return true;
409 }
410 
411 /*
412  * PT configuration helpers
413  * These all are cpu affine and operate on a local PT
414  */
415 
416 static void pt_config_start(struct perf_event *event)
417 {
418 	struct pt *pt = this_cpu_ptr(&pt_ctx);
419 	u64 ctl = event->hw.config;
420 
421 	ctl |= RTIT_CTL_TRACEEN;
422 	if (READ_ONCE(pt->vmx_on))
423 		perf_aux_output_flag(&pt->handle, PERF_AUX_FLAG_PARTIAL);
424 	else
425 		wrmsrl(MSR_IA32_RTIT_CTL, ctl);
426 
427 	WRITE_ONCE(event->hw.config, ctl);
428 }
429 
430 /* Address ranges and their corresponding msr configuration registers */
431 static const struct pt_address_range {
432 	unsigned long	msr_a;
433 	unsigned long	msr_b;
434 	unsigned int	reg_off;
435 } pt_address_ranges[] = {
436 	{
437 		.msr_a	 = MSR_IA32_RTIT_ADDR0_A,
438 		.msr_b	 = MSR_IA32_RTIT_ADDR0_B,
439 		.reg_off = RTIT_CTL_ADDR0_OFFSET,
440 	},
441 	{
442 		.msr_a	 = MSR_IA32_RTIT_ADDR1_A,
443 		.msr_b	 = MSR_IA32_RTIT_ADDR1_B,
444 		.reg_off = RTIT_CTL_ADDR1_OFFSET,
445 	},
446 	{
447 		.msr_a	 = MSR_IA32_RTIT_ADDR2_A,
448 		.msr_b	 = MSR_IA32_RTIT_ADDR2_B,
449 		.reg_off = RTIT_CTL_ADDR2_OFFSET,
450 	},
451 	{
452 		.msr_a	 = MSR_IA32_RTIT_ADDR3_A,
453 		.msr_b	 = MSR_IA32_RTIT_ADDR3_B,
454 		.reg_off = RTIT_CTL_ADDR3_OFFSET,
455 	}
456 };
457 
458 static u64 pt_config_filters(struct perf_event *event)
459 {
460 	struct pt_filters *filters = event->hw.addr_filters;
461 	struct pt *pt = this_cpu_ptr(&pt_ctx);
462 	unsigned int range = 0;
463 	u64 rtit_ctl = 0;
464 
465 	if (!filters)
466 		return 0;
467 
468 	perf_event_addr_filters_sync(event);
469 
470 	for (range = 0; range < filters->nr_filters; range++) {
471 		struct pt_filter *filter = &filters->filter[range];
472 
473 		/*
474 		 * Note, if the range has zero start/end addresses due
475 		 * to its dynamic object not being loaded yet, we just
476 		 * go ahead and program zeroed range, which will simply
477 		 * produce no data. Note^2: if executable code at 0x0
478 		 * is a concern, we can set up an "invalid" configuration
479 		 * such as msr_b < msr_a.
480 		 */
481 
482 		/* avoid redundant msr writes */
483 		if (pt->filters.filter[range].msr_a != filter->msr_a) {
484 			wrmsrl(pt_address_ranges[range].msr_a, filter->msr_a);
485 			pt->filters.filter[range].msr_a = filter->msr_a;
486 		}
487 
488 		if (pt->filters.filter[range].msr_b != filter->msr_b) {
489 			wrmsrl(pt_address_ranges[range].msr_b, filter->msr_b);
490 			pt->filters.filter[range].msr_b = filter->msr_b;
491 		}
492 
493 		rtit_ctl |= (u64)filter->config << pt_address_ranges[range].reg_off;
494 	}
495 
496 	return rtit_ctl;
497 }
498 
499 static void pt_config(struct perf_event *event)
500 {
501 	struct pt *pt = this_cpu_ptr(&pt_ctx);
502 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
503 	u64 reg;
504 
505 	/* First round: clear STATUS, in particular the PSB byte counter. */
506 	if (!event->hw.config) {
507 		perf_event_itrace_started(event);
508 		wrmsrl(MSR_IA32_RTIT_STATUS, 0);
509 	}
510 
511 	reg = pt_config_filters(event);
512 	reg |= RTIT_CTL_TRACEEN;
513 	if (!buf->single)
514 		reg |= RTIT_CTL_TOPA;
515 
516 	/*
517 	 * Previously, we had BRANCH_EN on by default, but now that PT has
518 	 * grown features outside of branch tracing, it is useful to allow
519 	 * the user to disable it. Setting bit 0 in the event's attr.config
520 	 * allows BRANCH_EN to pass through instead of being always on. See
521 	 * also the comment in pt_event_valid().
522 	 */
523 	if (event->attr.config & BIT(0)) {
524 		reg |= event->attr.config & RTIT_CTL_BRANCH_EN;
525 	} else {
526 		reg |= RTIT_CTL_BRANCH_EN;
527 	}
528 
529 	if (!event->attr.exclude_kernel)
530 		reg |= RTIT_CTL_OS;
531 	if (!event->attr.exclude_user)
532 		reg |= RTIT_CTL_USR;
533 
534 	reg |= (event->attr.config & PT_CONFIG_MASK);
535 
536 	event->hw.config = reg;
537 	pt_config_start(event);
538 }
539 
540 static void pt_config_stop(struct perf_event *event)
541 {
542 	struct pt *pt = this_cpu_ptr(&pt_ctx);
543 	u64 ctl = READ_ONCE(event->hw.config);
544 
545 	/* may be already stopped by a PMI */
546 	if (!(ctl & RTIT_CTL_TRACEEN))
547 		return;
548 
549 	ctl &= ~RTIT_CTL_TRACEEN;
550 	if (!READ_ONCE(pt->vmx_on))
551 		wrmsrl(MSR_IA32_RTIT_CTL, ctl);
552 
553 	WRITE_ONCE(event->hw.config, ctl);
554 
555 	/*
556 	 * A wrmsr that disables trace generation serializes other PT
557 	 * registers and causes all data packets to be written to memory,
558 	 * but a fence is required for the data to become globally visible.
559 	 *
560 	 * The below WMB, separating data store and aux_head store matches
561 	 * the consumer's RMB that separates aux_head load and data load.
562 	 */
563 	wmb();
564 }
565 
566 /**
567  * struct topa - ToPA metadata
568  * @list:	linkage to struct pt_buffer's list of tables
569  * @offset:	offset of the first entry in this table in the buffer
570  * @size:	total size of all entries in this table
571  * @last:	index of the last initialized entry in this table
572  * @z_count:	how many times the first entry repeats
573  */
574 struct topa {
575 	struct list_head	list;
576 	u64			offset;
577 	size_t			size;
578 	int			last;
579 	unsigned int		z_count;
580 };
581 
582 /*
583  * Keep ToPA table-related metadata on the same page as the actual table,
584  * taking up a few words from the top
585  */
586 
587 #define TENTS_PER_PAGE	\
588 	((PAGE_SIZE - sizeof(struct topa)) / sizeof(struct topa_entry))
589 
590 /**
591  * struct topa_page - page-sized ToPA table with metadata at the top
592  * @table:	actual ToPA table entries, as understood by PT hardware
593  * @topa:	metadata
594  */
595 struct topa_page {
596 	struct topa_entry	table[TENTS_PER_PAGE];
597 	struct topa		topa;
598 };
599 
600 static inline struct topa_page *topa_to_page(struct topa *topa)
601 {
602 	return container_of(topa, struct topa_page, topa);
603 }
604 
605 static inline struct topa_page *topa_entry_to_page(struct topa_entry *te)
606 {
607 	return (struct topa_page *)((unsigned long)te & PAGE_MASK);
608 }
609 
610 static inline phys_addr_t topa_pfn(struct topa *topa)
611 {
612 	return PFN_DOWN(virt_to_phys(topa_to_page(topa)));
613 }
614 
615 /* make -1 stand for the last table entry */
616 #define TOPA_ENTRY(t, i)				\
617 	((i) == -1					\
618 		? &topa_to_page(t)->table[(t)->last]	\
619 		: &topa_to_page(t)->table[(i)])
620 #define TOPA_ENTRY_SIZE(t, i) (sizes(TOPA_ENTRY((t), (i))->size))
621 #define TOPA_ENTRY_PAGES(t, i) (1 << TOPA_ENTRY((t), (i))->size)
622 
623 static void pt_config_buffer(struct pt_buffer *buf)
624 {
625 	struct pt *pt = this_cpu_ptr(&pt_ctx);
626 	u64 reg, mask;
627 	void *base;
628 
629 	if (buf->single) {
630 		base = buf->data_pages[0];
631 		mask = (buf->nr_pages * PAGE_SIZE - 1) >> 7;
632 	} else {
633 		base = topa_to_page(buf->cur)->table;
634 		mask = (u64)buf->cur_idx;
635 	}
636 
637 	reg = virt_to_phys(base);
638 	if (pt->output_base != reg) {
639 		pt->output_base = reg;
640 		wrmsrl(MSR_IA32_RTIT_OUTPUT_BASE, reg);
641 	}
642 
643 	reg = 0x7f | (mask << 7) | ((u64)buf->output_off << 32);
644 	if (pt->output_mask != reg) {
645 		pt->output_mask = reg;
646 		wrmsrl(MSR_IA32_RTIT_OUTPUT_MASK, reg);
647 	}
648 }
649 
650 /**
651  * topa_alloc() - allocate page-sized ToPA table
652  * @cpu:	CPU on which to allocate.
653  * @gfp:	Allocation flags.
654  *
655  * Return:	On success, return the pointer to ToPA table page.
656  */
657 static struct topa *topa_alloc(int cpu, gfp_t gfp)
658 {
659 	int node = cpu_to_node(cpu);
660 	struct topa_page *tp;
661 	struct page *p;
662 
663 	p = alloc_pages_node(node, gfp | __GFP_ZERO, 0);
664 	if (!p)
665 		return NULL;
666 
667 	tp = page_address(p);
668 	tp->topa.last = 0;
669 
670 	/*
671 	 * In case of singe-entry ToPA, always put the self-referencing END
672 	 * link as the 2nd entry in the table
673 	 */
674 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries)) {
675 		TOPA_ENTRY(&tp->topa, 1)->base = page_to_phys(p) >> TOPA_SHIFT;
676 		TOPA_ENTRY(&tp->topa, 1)->end = 1;
677 	}
678 
679 	return &tp->topa;
680 }
681 
682 /**
683  * topa_free() - free a page-sized ToPA table
684  * @topa:	Table to deallocate.
685  */
686 static void topa_free(struct topa *topa)
687 {
688 	free_page((unsigned long)topa);
689 }
690 
691 /**
692  * topa_insert_table() - insert a ToPA table into a buffer
693  * @buf:	 PT buffer that's being extended.
694  * @topa:	 New topa table to be inserted.
695  *
696  * If it's the first table in this buffer, set up buffer's pointers
697  * accordingly; otherwise, add a END=1 link entry to @topa to the current
698  * "last" table and adjust the last table pointer to @topa.
699  */
700 static void topa_insert_table(struct pt_buffer *buf, struct topa *topa)
701 {
702 	struct topa *last = buf->last;
703 
704 	list_add_tail(&topa->list, &buf->tables);
705 
706 	if (!buf->first) {
707 		buf->first = buf->last = buf->cur = topa;
708 		return;
709 	}
710 
711 	topa->offset = last->offset + last->size;
712 	buf->last = topa;
713 
714 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
715 		return;
716 
717 	BUG_ON(last->last != TENTS_PER_PAGE - 1);
718 
719 	TOPA_ENTRY(last, -1)->base = topa_pfn(topa);
720 	TOPA_ENTRY(last, -1)->end = 1;
721 }
722 
723 /**
724  * topa_table_full() - check if a ToPA table is filled up
725  * @topa:	ToPA table.
726  */
727 static bool topa_table_full(struct topa *topa)
728 {
729 	/* single-entry ToPA is a special case */
730 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
731 		return !!topa->last;
732 
733 	return topa->last == TENTS_PER_PAGE - 1;
734 }
735 
736 /**
737  * topa_insert_pages() - create a list of ToPA tables
738  * @buf:	PT buffer being initialized.
739  * @gfp:	Allocation flags.
740  *
741  * This initializes a list of ToPA tables with entries from
742  * the data_pages provided by rb_alloc_aux().
743  *
744  * Return:	0 on success or error code.
745  */
746 static int topa_insert_pages(struct pt_buffer *buf, int cpu, gfp_t gfp)
747 {
748 	struct topa *topa = buf->last;
749 	int order = 0;
750 	struct page *p;
751 
752 	p = virt_to_page(buf->data_pages[buf->nr_pages]);
753 	if (PagePrivate(p))
754 		order = page_private(p);
755 
756 	if (topa_table_full(topa)) {
757 		topa = topa_alloc(cpu, gfp);
758 		if (!topa)
759 			return -ENOMEM;
760 
761 		topa_insert_table(buf, topa);
762 	}
763 
764 	if (topa->z_count == topa->last - 1) {
765 		if (order == TOPA_ENTRY(topa, topa->last - 1)->size)
766 			topa->z_count++;
767 	}
768 
769 	TOPA_ENTRY(topa, -1)->base = page_to_phys(p) >> TOPA_SHIFT;
770 	TOPA_ENTRY(topa, -1)->size = order;
771 	if (!buf->snapshot &&
772 	    !intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries)) {
773 		TOPA_ENTRY(topa, -1)->intr = 1;
774 		TOPA_ENTRY(topa, -1)->stop = 1;
775 	}
776 
777 	topa->last++;
778 	topa->size += sizes(order);
779 
780 	buf->nr_pages += 1ul << order;
781 
782 	return 0;
783 }
784 
785 /**
786  * pt_topa_dump() - print ToPA tables and their entries
787  * @buf:	PT buffer.
788  */
789 static void pt_topa_dump(struct pt_buffer *buf)
790 {
791 	struct topa *topa;
792 
793 	list_for_each_entry(topa, &buf->tables, list) {
794 		struct topa_page *tp = topa_to_page(topa);
795 		int i;
796 
797 		pr_debug("# table @%p, off %llx size %zx\n", tp->table,
798 			 topa->offset, topa->size);
799 		for (i = 0; i < TENTS_PER_PAGE; i++) {
800 			pr_debug("# entry @%p (%lx sz %u %c%c%c) raw=%16llx\n",
801 				 &tp->table[i],
802 				 (unsigned long)tp->table[i].base << TOPA_SHIFT,
803 				 sizes(tp->table[i].size),
804 				 tp->table[i].end ?  'E' : ' ',
805 				 tp->table[i].intr ? 'I' : ' ',
806 				 tp->table[i].stop ? 'S' : ' ',
807 				 *(u64 *)&tp->table[i]);
808 			if ((intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) &&
809 			     tp->table[i].stop) ||
810 			    tp->table[i].end)
811 				break;
812 			if (!i && topa->z_count)
813 				i += topa->z_count;
814 		}
815 	}
816 }
817 
818 /**
819  * pt_buffer_advance() - advance to the next output region
820  * @buf:	PT buffer.
821  *
822  * Advance the current pointers in the buffer to the next ToPA entry.
823  */
824 static void pt_buffer_advance(struct pt_buffer *buf)
825 {
826 	buf->output_off = 0;
827 	buf->cur_idx++;
828 
829 	if (buf->cur_idx == buf->cur->last) {
830 		if (buf->cur == buf->last)
831 			buf->cur = buf->first;
832 		else
833 			buf->cur = list_entry(buf->cur->list.next, struct topa,
834 					      list);
835 		buf->cur_idx = 0;
836 	}
837 }
838 
839 /**
840  * pt_update_head() - calculate current offsets and sizes
841  * @pt:		Per-cpu pt context.
842  *
843  * Update buffer's current write pointer position and data size.
844  */
845 static void pt_update_head(struct pt *pt)
846 {
847 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
848 	u64 topa_idx, base, old;
849 
850 	if (buf->single) {
851 		local_set(&buf->data_size, buf->output_off);
852 		return;
853 	}
854 
855 	/* offset of the first region in this table from the beginning of buf */
856 	base = buf->cur->offset + buf->output_off;
857 
858 	/* offset of the current output region within this table */
859 	for (topa_idx = 0; topa_idx < buf->cur_idx; topa_idx++)
860 		base += TOPA_ENTRY_SIZE(buf->cur, topa_idx);
861 
862 	if (buf->snapshot) {
863 		local_set(&buf->data_size, base);
864 	} else {
865 		old = (local64_xchg(&buf->head, base) &
866 		       ((buf->nr_pages << PAGE_SHIFT) - 1));
867 		if (base < old)
868 			base += buf->nr_pages << PAGE_SHIFT;
869 
870 		local_add(base - old, &buf->data_size);
871 	}
872 }
873 
874 /**
875  * pt_buffer_region() - obtain current output region's address
876  * @buf:	PT buffer.
877  */
878 static void *pt_buffer_region(struct pt_buffer *buf)
879 {
880 	return phys_to_virt(TOPA_ENTRY(buf->cur, buf->cur_idx)->base << TOPA_SHIFT);
881 }
882 
883 /**
884  * pt_buffer_region_size() - obtain current output region's size
885  * @buf:	PT buffer.
886  */
887 static size_t pt_buffer_region_size(struct pt_buffer *buf)
888 {
889 	return TOPA_ENTRY_SIZE(buf->cur, buf->cur_idx);
890 }
891 
892 /**
893  * pt_handle_status() - take care of possible status conditions
894  * @pt:		Per-cpu pt context.
895  */
896 static void pt_handle_status(struct pt *pt)
897 {
898 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
899 	int advance = 0;
900 	u64 status;
901 
902 	rdmsrl(MSR_IA32_RTIT_STATUS, status);
903 
904 	if (status & RTIT_STATUS_ERROR) {
905 		pr_err_ratelimited("ToPA ERROR encountered, trying to recover\n");
906 		pt_topa_dump(buf);
907 		status &= ~RTIT_STATUS_ERROR;
908 	}
909 
910 	if (status & RTIT_STATUS_STOPPED) {
911 		status &= ~RTIT_STATUS_STOPPED;
912 
913 		/*
914 		 * On systems that only do single-entry ToPA, hitting STOP
915 		 * means we are already losing data; need to let the decoder
916 		 * know.
917 		 */
918 		if (!buf->single &&
919 		    (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) ||
920 		     buf->output_off == pt_buffer_region_size(buf))) {
921 			perf_aux_output_flag(&pt->handle,
922 			                     PERF_AUX_FLAG_TRUNCATED);
923 			advance++;
924 		}
925 	}
926 
927 	/*
928 	 * Also on single-entry ToPA implementations, interrupt will come
929 	 * before the output reaches its output region's boundary.
930 	 */
931 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) &&
932 	    !buf->snapshot &&
933 	    pt_buffer_region_size(buf) - buf->output_off <= TOPA_PMI_MARGIN) {
934 		void *head = pt_buffer_region(buf);
935 
936 		/* everything within this margin needs to be zeroed out */
937 		memset(head + buf->output_off, 0,
938 		       pt_buffer_region_size(buf) -
939 		       buf->output_off);
940 		advance++;
941 	}
942 
943 	if (advance)
944 		pt_buffer_advance(buf);
945 
946 	wrmsrl(MSR_IA32_RTIT_STATUS, status);
947 }
948 
949 /**
950  * pt_read_offset() - translate registers into buffer pointers
951  * @buf:	PT buffer.
952  *
953  * Set buffer's output pointers from MSR values.
954  */
955 static void pt_read_offset(struct pt_buffer *buf)
956 {
957 	struct pt *pt = this_cpu_ptr(&pt_ctx);
958 	struct topa_page *tp;
959 
960 	if (!buf->single) {
961 		rdmsrl(MSR_IA32_RTIT_OUTPUT_BASE, pt->output_base);
962 		tp = phys_to_virt(pt->output_base);
963 		buf->cur = &tp->topa;
964 	}
965 
966 	rdmsrl(MSR_IA32_RTIT_OUTPUT_MASK, pt->output_mask);
967 	/* offset within current output region */
968 	buf->output_off = pt->output_mask >> 32;
969 	/* index of current output region within this table */
970 	if (!buf->single)
971 		buf->cur_idx = (pt->output_mask & 0xffffff80) >> 7;
972 }
973 
974 static struct topa_entry *
975 pt_topa_entry_for_page(struct pt_buffer *buf, unsigned int pg)
976 {
977 	struct topa_page *tp;
978 	struct topa *topa;
979 	unsigned int idx, cur_pg = 0, z_pg = 0, start_idx = 0;
980 
981 	/*
982 	 * Indicates a bug in the caller.
983 	 */
984 	if (WARN_ON_ONCE(pg >= buf->nr_pages))
985 		return NULL;
986 
987 	/*
988 	 * First, find the ToPA table where @pg fits. With high
989 	 * order allocations, there shouldn't be many of these.
990 	 */
991 	list_for_each_entry(topa, &buf->tables, list) {
992 		if (topa->offset + topa->size > pg << PAGE_SHIFT)
993 			goto found;
994 	}
995 
996 	/*
997 	 * Hitting this means we have a problem in the ToPA
998 	 * allocation code.
999 	 */
1000 	WARN_ON_ONCE(1);
1001 
1002 	return NULL;
1003 
1004 found:
1005 	/*
1006 	 * Indicates a problem in the ToPA allocation code.
1007 	 */
1008 	if (WARN_ON_ONCE(topa->last == -1))
1009 		return NULL;
1010 
1011 	tp = topa_to_page(topa);
1012 	cur_pg = PFN_DOWN(topa->offset);
1013 	if (topa->z_count) {
1014 		z_pg = TOPA_ENTRY_PAGES(topa, 0) * (topa->z_count + 1);
1015 		start_idx = topa->z_count + 1;
1016 	}
1017 
1018 	/*
1019 	 * Multiple entries at the beginning of the table have the same size,
1020 	 * ideally all of them; if @pg falls there, the search is done.
1021 	 */
1022 	if (pg >= cur_pg && pg < cur_pg + z_pg) {
1023 		idx = (pg - cur_pg) / TOPA_ENTRY_PAGES(topa, 0);
1024 		return &tp->table[idx];
1025 	}
1026 
1027 	/*
1028 	 * Otherwise, slow path: iterate through the remaining entries.
1029 	 */
1030 	for (idx = start_idx, cur_pg += z_pg; idx < topa->last; idx++) {
1031 		if (cur_pg + TOPA_ENTRY_PAGES(topa, idx) > pg)
1032 			return &tp->table[idx];
1033 
1034 		cur_pg += TOPA_ENTRY_PAGES(topa, idx);
1035 	}
1036 
1037 	/*
1038 	 * Means we couldn't find a ToPA entry in the table that does match.
1039 	 */
1040 	WARN_ON_ONCE(1);
1041 
1042 	return NULL;
1043 }
1044 
1045 static struct topa_entry *
1046 pt_topa_prev_entry(struct pt_buffer *buf, struct topa_entry *te)
1047 {
1048 	unsigned long table = (unsigned long)te & ~(PAGE_SIZE - 1);
1049 	struct topa_page *tp;
1050 	struct topa *topa;
1051 
1052 	tp = (struct topa_page *)table;
1053 	if (tp->table != te)
1054 		return --te;
1055 
1056 	topa = &tp->topa;
1057 	if (topa == buf->first)
1058 		topa = buf->last;
1059 	else
1060 		topa = list_prev_entry(topa, list);
1061 
1062 	tp = topa_to_page(topa);
1063 
1064 	return &tp->table[topa->last - 1];
1065 }
1066 
1067 /**
1068  * pt_buffer_reset_markers() - place interrupt and stop bits in the buffer
1069  * @buf:	PT buffer.
1070  * @handle:	Current output handle.
1071  *
1072  * Place INT and STOP marks to prevent overwriting old data that the consumer
1073  * hasn't yet collected and waking up the consumer after a certain fraction of
1074  * the buffer has filled up. Only needed and sensible for non-snapshot counters.
1075  *
1076  * This obviously relies on buf::head to figure out buffer markers, so it has
1077  * to be called after pt_buffer_reset_offsets() and before the hardware tracing
1078  * is enabled.
1079  */
1080 static int pt_buffer_reset_markers(struct pt_buffer *buf,
1081 				   struct perf_output_handle *handle)
1082 
1083 {
1084 	unsigned long head = local64_read(&buf->head);
1085 	unsigned long idx, npages, wakeup;
1086 
1087 	if (buf->single)
1088 		return 0;
1089 
1090 	/* can't stop in the middle of an output region */
1091 	if (buf->output_off + handle->size + 1 < pt_buffer_region_size(buf)) {
1092 		perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED);
1093 		return -EINVAL;
1094 	}
1095 
1096 
1097 	/* single entry ToPA is handled by marking all regions STOP=1 INT=1 */
1098 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
1099 		return 0;
1100 
1101 	/* clear STOP and INT from current entry */
1102 	if (buf->stop_te) {
1103 		buf->stop_te->stop = 0;
1104 		buf->stop_te->intr = 0;
1105 	}
1106 
1107 	if (buf->intr_te)
1108 		buf->intr_te->intr = 0;
1109 
1110 	/* how many pages till the STOP marker */
1111 	npages = handle->size >> PAGE_SHIFT;
1112 
1113 	/* if it's on a page boundary, fill up one more page */
1114 	if (!offset_in_page(head + handle->size + 1))
1115 		npages++;
1116 
1117 	idx = (head >> PAGE_SHIFT) + npages;
1118 	idx &= buf->nr_pages - 1;
1119 
1120 	if (idx != buf->stop_pos) {
1121 		buf->stop_pos = idx;
1122 		buf->stop_te = pt_topa_entry_for_page(buf, idx);
1123 		buf->stop_te = pt_topa_prev_entry(buf, buf->stop_te);
1124 	}
1125 
1126 	wakeup = handle->wakeup >> PAGE_SHIFT;
1127 
1128 	/* in the worst case, wake up the consumer one page before hard stop */
1129 	idx = (head >> PAGE_SHIFT) + npages - 1;
1130 	if (idx > wakeup)
1131 		idx = wakeup;
1132 
1133 	idx &= buf->nr_pages - 1;
1134 	if (idx != buf->intr_pos) {
1135 		buf->intr_pos = idx;
1136 		buf->intr_te = pt_topa_entry_for_page(buf, idx);
1137 		buf->intr_te = pt_topa_prev_entry(buf, buf->intr_te);
1138 	}
1139 
1140 	buf->stop_te->stop = 1;
1141 	buf->stop_te->intr = 1;
1142 	buf->intr_te->intr = 1;
1143 
1144 	return 0;
1145 }
1146 
1147 /**
1148  * pt_buffer_reset_offsets() - adjust buffer's write pointers from aux_head
1149  * @buf:	PT buffer.
1150  * @head:	Write pointer (aux_head) from AUX buffer.
1151  *
1152  * Find the ToPA table and entry corresponding to given @head and set buffer's
1153  * "current" pointers accordingly. This is done after we have obtained the
1154  * current aux_head position from a successful call to perf_aux_output_begin()
1155  * to make sure the hardware is writing to the right place.
1156  *
1157  * This function modifies buf::{cur,cur_idx,output_off} that will be programmed
1158  * into PT msrs when the tracing is enabled and buf::head and buf::data_size,
1159  * which are used to determine INT and STOP markers' locations by a subsequent
1160  * call to pt_buffer_reset_markers().
1161  */
1162 static void pt_buffer_reset_offsets(struct pt_buffer *buf, unsigned long head)
1163 {
1164 	struct topa_page *cur_tp;
1165 	struct topa_entry *te;
1166 	int pg;
1167 
1168 	if (buf->snapshot)
1169 		head &= (buf->nr_pages << PAGE_SHIFT) - 1;
1170 
1171 	if (!buf->single) {
1172 		pg = (head >> PAGE_SHIFT) & (buf->nr_pages - 1);
1173 		te = pt_topa_entry_for_page(buf, pg);
1174 
1175 		cur_tp = topa_entry_to_page(te);
1176 		buf->cur = &cur_tp->topa;
1177 		buf->cur_idx = te - TOPA_ENTRY(buf->cur, 0);
1178 		buf->output_off = head & (pt_buffer_region_size(buf) - 1);
1179 	} else {
1180 		buf->output_off = head;
1181 	}
1182 
1183 	local64_set(&buf->head, head);
1184 	local_set(&buf->data_size, 0);
1185 }
1186 
1187 /**
1188  * pt_buffer_fini_topa() - deallocate ToPA structure of a buffer
1189  * @buf:	PT buffer.
1190  */
1191 static void pt_buffer_fini_topa(struct pt_buffer *buf)
1192 {
1193 	struct topa *topa, *iter;
1194 
1195 	if (buf->single)
1196 		return;
1197 
1198 	list_for_each_entry_safe(topa, iter, &buf->tables, list) {
1199 		/*
1200 		 * right now, this is in free_aux() path only, so
1201 		 * no need to unlink this table from the list
1202 		 */
1203 		topa_free(topa);
1204 	}
1205 }
1206 
1207 /**
1208  * pt_buffer_init_topa() - initialize ToPA table for pt buffer
1209  * @buf:	PT buffer.
1210  * @size:	Total size of all regions within this ToPA.
1211  * @gfp:	Allocation flags.
1212  */
1213 static int pt_buffer_init_topa(struct pt_buffer *buf, int cpu,
1214 			       unsigned long nr_pages, gfp_t gfp)
1215 {
1216 	struct topa *topa;
1217 	int err;
1218 
1219 	topa = topa_alloc(cpu, gfp);
1220 	if (!topa)
1221 		return -ENOMEM;
1222 
1223 	topa_insert_table(buf, topa);
1224 
1225 	while (buf->nr_pages < nr_pages) {
1226 		err = topa_insert_pages(buf, cpu, gfp);
1227 		if (err) {
1228 			pt_buffer_fini_topa(buf);
1229 			return -ENOMEM;
1230 		}
1231 	}
1232 
1233 	/* link last table to the first one, unless we're double buffering */
1234 	if (intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries)) {
1235 		TOPA_ENTRY(buf->last, -1)->base = topa_pfn(buf->first);
1236 		TOPA_ENTRY(buf->last, -1)->end = 1;
1237 	}
1238 
1239 	pt_topa_dump(buf);
1240 	return 0;
1241 }
1242 
1243 static int pt_buffer_try_single(struct pt_buffer *buf, int nr_pages)
1244 {
1245 	struct page *p = virt_to_page(buf->data_pages[0]);
1246 	int ret = -ENOTSUPP, order = 0;
1247 
1248 	/*
1249 	 * We can use single range output mode
1250 	 * + in snapshot mode, where we don't need interrupts;
1251 	 * + if the hardware supports it;
1252 	 * + if the entire buffer is one contiguous allocation.
1253 	 */
1254 	if (!buf->snapshot)
1255 		goto out;
1256 
1257 	if (!intel_pt_validate_hw_cap(PT_CAP_single_range_output))
1258 		goto out;
1259 
1260 	if (PagePrivate(p))
1261 		order = page_private(p);
1262 
1263 	if (1 << order != nr_pages)
1264 		goto out;
1265 
1266 	buf->single = true;
1267 	buf->nr_pages = nr_pages;
1268 	ret = 0;
1269 out:
1270 	return ret;
1271 }
1272 
1273 /**
1274  * pt_buffer_setup_aux() - set up topa tables for a PT buffer
1275  * @cpu:	Cpu on which to allocate, -1 means current.
1276  * @pages:	Array of pointers to buffer pages passed from perf core.
1277  * @nr_pages:	Number of pages in the buffer.
1278  * @snapshot:	If this is a snapshot/overwrite counter.
1279  *
1280  * This is a pmu::setup_aux callback that sets up ToPA tables and all the
1281  * bookkeeping for an AUX buffer.
1282  *
1283  * Return:	Our private PT buffer structure.
1284  */
1285 static void *
1286 pt_buffer_setup_aux(struct perf_event *event, void **pages,
1287 		    int nr_pages, bool snapshot)
1288 {
1289 	struct pt_buffer *buf;
1290 	int node, ret, cpu = event->cpu;
1291 
1292 	if (!nr_pages)
1293 		return NULL;
1294 
1295 	/*
1296 	 * Only support AUX sampling in snapshot mode, where we don't
1297 	 * generate NMIs.
1298 	 */
1299 	if (event->attr.aux_sample_size && !snapshot)
1300 		return NULL;
1301 
1302 	if (cpu == -1)
1303 		cpu = raw_smp_processor_id();
1304 	node = cpu_to_node(cpu);
1305 
1306 	buf = kzalloc_node(sizeof(struct pt_buffer), GFP_KERNEL, node);
1307 	if (!buf)
1308 		return NULL;
1309 
1310 	buf->snapshot = snapshot;
1311 	buf->data_pages = pages;
1312 	buf->stop_pos = -1;
1313 	buf->intr_pos = -1;
1314 
1315 	INIT_LIST_HEAD(&buf->tables);
1316 
1317 	ret = pt_buffer_try_single(buf, nr_pages);
1318 	if (!ret)
1319 		return buf;
1320 
1321 	ret = pt_buffer_init_topa(buf, cpu, nr_pages, GFP_KERNEL);
1322 	if (ret) {
1323 		kfree(buf);
1324 		return NULL;
1325 	}
1326 
1327 	return buf;
1328 }
1329 
1330 /**
1331  * pt_buffer_free_aux() - perf AUX deallocation path callback
1332  * @data:	PT buffer.
1333  */
1334 static void pt_buffer_free_aux(void *data)
1335 {
1336 	struct pt_buffer *buf = data;
1337 
1338 	pt_buffer_fini_topa(buf);
1339 	kfree(buf);
1340 }
1341 
1342 static int pt_addr_filters_init(struct perf_event *event)
1343 {
1344 	struct pt_filters *filters;
1345 	int node = event->cpu == -1 ? -1 : cpu_to_node(event->cpu);
1346 
1347 	if (!intel_pt_validate_hw_cap(PT_CAP_num_address_ranges))
1348 		return 0;
1349 
1350 	filters = kzalloc_node(sizeof(struct pt_filters), GFP_KERNEL, node);
1351 	if (!filters)
1352 		return -ENOMEM;
1353 
1354 	if (event->parent)
1355 		memcpy(filters, event->parent->hw.addr_filters,
1356 		       sizeof(*filters));
1357 
1358 	event->hw.addr_filters = filters;
1359 
1360 	return 0;
1361 }
1362 
1363 static void pt_addr_filters_fini(struct perf_event *event)
1364 {
1365 	kfree(event->hw.addr_filters);
1366 	event->hw.addr_filters = NULL;
1367 }
1368 
1369 #ifdef CONFIG_X86_64
1370 /* Clamp to a canonical address greater-than-or-equal-to the address given */
1371 static u64 clamp_to_ge_canonical_addr(u64 vaddr, u8 vaddr_bits)
1372 {
1373 	return __is_canonical_address(vaddr, vaddr_bits) ?
1374 	       vaddr :
1375 	       -BIT_ULL(vaddr_bits - 1);
1376 }
1377 
1378 /* Clamp to a canonical address less-than-or-equal-to the address given */
1379 static u64 clamp_to_le_canonical_addr(u64 vaddr, u8 vaddr_bits)
1380 {
1381 	return __is_canonical_address(vaddr, vaddr_bits) ?
1382 	       vaddr :
1383 	       BIT_ULL(vaddr_bits - 1) - 1;
1384 }
1385 #else
1386 #define clamp_to_ge_canonical_addr(x, y) (x)
1387 #define clamp_to_le_canonical_addr(x, y) (x)
1388 #endif
1389 
1390 static int pt_event_addr_filters_validate(struct list_head *filters)
1391 {
1392 	struct perf_addr_filter *filter;
1393 	int range = 0;
1394 
1395 	list_for_each_entry(filter, filters, entry) {
1396 		/*
1397 		 * PT doesn't support single address triggers and
1398 		 * 'start' filters.
1399 		 */
1400 		if (!filter->size ||
1401 		    filter->action == PERF_ADDR_FILTER_ACTION_START)
1402 			return -EOPNOTSUPP;
1403 
1404 		if (++range > intel_pt_validate_hw_cap(PT_CAP_num_address_ranges))
1405 			return -EOPNOTSUPP;
1406 	}
1407 
1408 	return 0;
1409 }
1410 
1411 static void pt_event_addr_filters_sync(struct perf_event *event)
1412 {
1413 	struct perf_addr_filters_head *head = perf_event_addr_filters(event);
1414 	unsigned long msr_a, msr_b;
1415 	struct perf_addr_filter_range *fr = event->addr_filter_ranges;
1416 	struct pt_filters *filters = event->hw.addr_filters;
1417 	struct perf_addr_filter *filter;
1418 	int range = 0;
1419 
1420 	if (!filters)
1421 		return;
1422 
1423 	list_for_each_entry(filter, &head->list, entry) {
1424 		if (filter->path.dentry && !fr[range].start) {
1425 			msr_a = msr_b = 0;
1426 		} else {
1427 			unsigned long n = fr[range].size - 1;
1428 			unsigned long a = fr[range].start;
1429 			unsigned long b;
1430 
1431 			if (a > ULONG_MAX - n)
1432 				b = ULONG_MAX;
1433 			else
1434 				b = a + n;
1435 			/*
1436 			 * Apply the offset. 64-bit addresses written to the
1437 			 * MSRs must be canonical, but the range can encompass
1438 			 * non-canonical addresses. Since software cannot
1439 			 * execute at non-canonical addresses, adjusting to
1440 			 * canonical addresses does not affect the result of the
1441 			 * address filter.
1442 			 */
1443 			msr_a = clamp_to_ge_canonical_addr(a, boot_cpu_data.x86_virt_bits);
1444 			msr_b = clamp_to_le_canonical_addr(b, boot_cpu_data.x86_virt_bits);
1445 			if (msr_b < msr_a)
1446 				msr_a = msr_b = 0;
1447 		}
1448 
1449 		filters->filter[range].msr_a  = msr_a;
1450 		filters->filter[range].msr_b  = msr_b;
1451 		if (filter->action == PERF_ADDR_FILTER_ACTION_FILTER)
1452 			filters->filter[range].config = 1;
1453 		else
1454 			filters->filter[range].config = 2;
1455 		range++;
1456 	}
1457 
1458 	filters->nr_filters = range;
1459 }
1460 
1461 /**
1462  * intel_pt_interrupt() - PT PMI handler
1463  */
1464 void intel_pt_interrupt(void)
1465 {
1466 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1467 	struct pt_buffer *buf;
1468 	struct perf_event *event = pt->handle.event;
1469 
1470 	/*
1471 	 * There may be a dangling PT bit in the interrupt status register
1472 	 * after PT has been disabled by pt_event_stop(). Make sure we don't
1473 	 * do anything (particularly, re-enable) for this event here.
1474 	 */
1475 	if (!READ_ONCE(pt->handle_nmi))
1476 		return;
1477 
1478 	if (!event)
1479 		return;
1480 
1481 	pt_config_stop(event);
1482 
1483 	buf = perf_get_aux(&pt->handle);
1484 	if (!buf)
1485 		return;
1486 
1487 	pt_read_offset(buf);
1488 
1489 	pt_handle_status(pt);
1490 
1491 	pt_update_head(pt);
1492 
1493 	perf_aux_output_end(&pt->handle, local_xchg(&buf->data_size, 0));
1494 
1495 	if (!event->hw.state) {
1496 		int ret;
1497 
1498 		buf = perf_aux_output_begin(&pt->handle, event);
1499 		if (!buf) {
1500 			event->hw.state = PERF_HES_STOPPED;
1501 			return;
1502 		}
1503 
1504 		pt_buffer_reset_offsets(buf, pt->handle.head);
1505 		/* snapshot counters don't use PMI, so it's safe */
1506 		ret = pt_buffer_reset_markers(buf, &pt->handle);
1507 		if (ret) {
1508 			perf_aux_output_end(&pt->handle, 0);
1509 			return;
1510 		}
1511 
1512 		pt_config_buffer(buf);
1513 		pt_config_start(event);
1514 	}
1515 }
1516 
1517 void intel_pt_handle_vmx(int on)
1518 {
1519 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1520 	struct perf_event *event;
1521 	unsigned long flags;
1522 
1523 	/* PT plays nice with VMX, do nothing */
1524 	if (pt_pmu.vmx)
1525 		return;
1526 
1527 	/*
1528 	 * VMXON will clear RTIT_CTL.TraceEn; we need to make
1529 	 * sure to not try to set it while VMX is on. Disable
1530 	 * interrupts to avoid racing with pmu callbacks;
1531 	 * concurrent PMI should be handled fine.
1532 	 */
1533 	local_irq_save(flags);
1534 	WRITE_ONCE(pt->vmx_on, on);
1535 
1536 	/*
1537 	 * If an AUX transaction is in progress, it will contain
1538 	 * gap(s), so flag it PARTIAL to inform the user.
1539 	 */
1540 	event = pt->handle.event;
1541 	if (event)
1542 		perf_aux_output_flag(&pt->handle,
1543 		                     PERF_AUX_FLAG_PARTIAL);
1544 
1545 	/* Turn PTs back on */
1546 	if (!on && event)
1547 		wrmsrl(MSR_IA32_RTIT_CTL, event->hw.config);
1548 
1549 	local_irq_restore(flags);
1550 }
1551 EXPORT_SYMBOL_GPL(intel_pt_handle_vmx);
1552 
1553 /*
1554  * PMU callbacks
1555  */
1556 
1557 static void pt_event_start(struct perf_event *event, int mode)
1558 {
1559 	struct hw_perf_event *hwc = &event->hw;
1560 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1561 	struct pt_buffer *buf;
1562 
1563 	buf = perf_aux_output_begin(&pt->handle, event);
1564 	if (!buf)
1565 		goto fail_stop;
1566 
1567 	pt_buffer_reset_offsets(buf, pt->handle.head);
1568 	if (!buf->snapshot) {
1569 		if (pt_buffer_reset_markers(buf, &pt->handle))
1570 			goto fail_end_stop;
1571 	}
1572 
1573 	WRITE_ONCE(pt->handle_nmi, 1);
1574 	hwc->state = 0;
1575 
1576 	pt_config_buffer(buf);
1577 	pt_config(event);
1578 
1579 	return;
1580 
1581 fail_end_stop:
1582 	perf_aux_output_end(&pt->handle, 0);
1583 fail_stop:
1584 	hwc->state = PERF_HES_STOPPED;
1585 }
1586 
1587 static void pt_event_stop(struct perf_event *event, int mode)
1588 {
1589 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1590 
1591 	/*
1592 	 * Protect against the PMI racing with disabling wrmsr,
1593 	 * see comment in intel_pt_interrupt().
1594 	 */
1595 	WRITE_ONCE(pt->handle_nmi, 0);
1596 
1597 	pt_config_stop(event);
1598 
1599 	if (event->hw.state == PERF_HES_STOPPED)
1600 		return;
1601 
1602 	event->hw.state = PERF_HES_STOPPED;
1603 
1604 	if (mode & PERF_EF_UPDATE) {
1605 		struct pt_buffer *buf = perf_get_aux(&pt->handle);
1606 
1607 		if (!buf)
1608 			return;
1609 
1610 		if (WARN_ON_ONCE(pt->handle.event != event))
1611 			return;
1612 
1613 		pt_read_offset(buf);
1614 
1615 		pt_handle_status(pt);
1616 
1617 		pt_update_head(pt);
1618 
1619 		if (buf->snapshot)
1620 			pt->handle.head =
1621 				local_xchg(&buf->data_size,
1622 					   buf->nr_pages << PAGE_SHIFT);
1623 		perf_aux_output_end(&pt->handle, local_xchg(&buf->data_size, 0));
1624 	}
1625 }
1626 
1627 static long pt_event_snapshot_aux(struct perf_event *event,
1628 				  struct perf_output_handle *handle,
1629 				  unsigned long size)
1630 {
1631 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1632 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
1633 	unsigned long from = 0, to;
1634 	long ret;
1635 
1636 	if (WARN_ON_ONCE(!buf))
1637 		return 0;
1638 
1639 	/*
1640 	 * Sampling is only allowed on snapshot events;
1641 	 * see pt_buffer_setup_aux().
1642 	 */
1643 	if (WARN_ON_ONCE(!buf->snapshot))
1644 		return 0;
1645 
1646 	/*
1647 	 * Here, handle_nmi tells us if the tracing is on
1648 	 */
1649 	if (READ_ONCE(pt->handle_nmi))
1650 		pt_config_stop(event);
1651 
1652 	pt_read_offset(buf);
1653 	pt_update_head(pt);
1654 
1655 	to = local_read(&buf->data_size);
1656 	if (to < size)
1657 		from = buf->nr_pages << PAGE_SHIFT;
1658 	from += to - size;
1659 
1660 	ret = perf_output_copy_aux(&pt->handle, handle, from, to);
1661 
1662 	/*
1663 	 * If the tracing was on when we turned up, restart it.
1664 	 * Compiler barrier not needed as we couldn't have been
1665 	 * preempted by anything that touches pt->handle_nmi.
1666 	 */
1667 	if (pt->handle_nmi)
1668 		pt_config_start(event);
1669 
1670 	return ret;
1671 }
1672 
1673 static void pt_event_del(struct perf_event *event, int mode)
1674 {
1675 	pt_event_stop(event, PERF_EF_UPDATE);
1676 }
1677 
1678 static int pt_event_add(struct perf_event *event, int mode)
1679 {
1680 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1681 	struct hw_perf_event *hwc = &event->hw;
1682 	int ret = -EBUSY;
1683 
1684 	if (pt->handle.event)
1685 		goto fail;
1686 
1687 	if (mode & PERF_EF_START) {
1688 		pt_event_start(event, 0);
1689 		ret = -EINVAL;
1690 		if (hwc->state == PERF_HES_STOPPED)
1691 			goto fail;
1692 	} else {
1693 		hwc->state = PERF_HES_STOPPED;
1694 	}
1695 
1696 	ret = 0;
1697 fail:
1698 
1699 	return ret;
1700 }
1701 
1702 static void pt_event_read(struct perf_event *event)
1703 {
1704 }
1705 
1706 static void pt_event_destroy(struct perf_event *event)
1707 {
1708 	pt_addr_filters_fini(event);
1709 	x86_del_exclusive(x86_lbr_exclusive_pt);
1710 }
1711 
1712 static int pt_event_init(struct perf_event *event)
1713 {
1714 	if (event->attr.type != pt_pmu.pmu.type)
1715 		return -ENOENT;
1716 
1717 	if (!pt_event_valid(event))
1718 		return -EINVAL;
1719 
1720 	if (x86_add_exclusive(x86_lbr_exclusive_pt))
1721 		return -EBUSY;
1722 
1723 	if (pt_addr_filters_init(event)) {
1724 		x86_del_exclusive(x86_lbr_exclusive_pt);
1725 		return -ENOMEM;
1726 	}
1727 
1728 	event->destroy = pt_event_destroy;
1729 
1730 	return 0;
1731 }
1732 
1733 void cpu_emergency_stop_pt(void)
1734 {
1735 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1736 
1737 	if (pt->handle.event)
1738 		pt_event_stop(pt->handle.event, PERF_EF_UPDATE);
1739 }
1740 
1741 int is_intel_pt_event(struct perf_event *event)
1742 {
1743 	return event->pmu == &pt_pmu.pmu;
1744 }
1745 
1746 static __init int pt_init(void)
1747 {
1748 	int ret, cpu, prior_warn = 0;
1749 
1750 	BUILD_BUG_ON(sizeof(struct topa) > PAGE_SIZE);
1751 
1752 	if (!boot_cpu_has(X86_FEATURE_INTEL_PT))
1753 		return -ENODEV;
1754 
1755 	cpus_read_lock();
1756 	for_each_online_cpu(cpu) {
1757 		u64 ctl;
1758 
1759 		ret = rdmsrl_safe_on_cpu(cpu, MSR_IA32_RTIT_CTL, &ctl);
1760 		if (!ret && (ctl & RTIT_CTL_TRACEEN))
1761 			prior_warn++;
1762 	}
1763 	cpus_read_unlock();
1764 
1765 	if (prior_warn) {
1766 		x86_add_exclusive(x86_lbr_exclusive_pt);
1767 		pr_warn("PT is enabled at boot time, doing nothing\n");
1768 
1769 		return -EBUSY;
1770 	}
1771 
1772 	ret = pt_pmu_hw_init();
1773 	if (ret)
1774 		return ret;
1775 
1776 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_output)) {
1777 		pr_warn("ToPA output is not supported on this CPU\n");
1778 		return -ENODEV;
1779 	}
1780 
1781 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
1782 		pt_pmu.pmu.capabilities = PERF_PMU_CAP_AUX_NO_SG;
1783 
1784 	pt_pmu.pmu.capabilities	|= PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE;
1785 	pt_pmu.pmu.attr_groups		 = pt_attr_groups;
1786 	pt_pmu.pmu.task_ctx_nr		 = perf_sw_context;
1787 	pt_pmu.pmu.event_init		 = pt_event_init;
1788 	pt_pmu.pmu.add			 = pt_event_add;
1789 	pt_pmu.pmu.del			 = pt_event_del;
1790 	pt_pmu.pmu.start		 = pt_event_start;
1791 	pt_pmu.pmu.stop			 = pt_event_stop;
1792 	pt_pmu.pmu.snapshot_aux		 = pt_event_snapshot_aux;
1793 	pt_pmu.pmu.read			 = pt_event_read;
1794 	pt_pmu.pmu.setup_aux		 = pt_buffer_setup_aux;
1795 	pt_pmu.pmu.free_aux		 = pt_buffer_free_aux;
1796 	pt_pmu.pmu.addr_filters_sync     = pt_event_addr_filters_sync;
1797 	pt_pmu.pmu.addr_filters_validate = pt_event_addr_filters_validate;
1798 	pt_pmu.pmu.nr_addr_filters       =
1799 		intel_pt_validate_hw_cap(PT_CAP_num_address_ranges);
1800 
1801 	ret = perf_pmu_register(&pt_pmu.pmu, "intel_pt", -1);
1802 
1803 	return ret;
1804 }
1805 arch_initcall(pt_init);
1806