xref: /linux/arch/x86/events/intel/pt.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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/cpu_device_id.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_vfm) {
215 	case INTEL_BROADWELL:
216 	case INTEL_BROADWELL_D:
217 	case INTEL_BROADWELL_G:
218 	case INTEL_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.aux_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.aux_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.aux_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.aux_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.aux_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.aux_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  * @cpu:	CPU on which to allocate.
740  * @gfp:	Allocation flags.
741  *
742  * This initializes a list of ToPA tables with entries from
743  * the data_pages provided by rb_alloc_aux().
744  *
745  * Return:	0 on success or error code.
746  */
747 static int topa_insert_pages(struct pt_buffer *buf, int cpu, gfp_t gfp)
748 {
749 	struct topa *topa = buf->last;
750 	int order = 0;
751 	struct page *p;
752 
753 	p = virt_to_page(buf->data_pages[buf->nr_pages]);
754 	if (PagePrivate(p))
755 		order = page_private(p);
756 
757 	if (topa_table_full(topa)) {
758 		topa = topa_alloc(cpu, gfp);
759 		if (!topa)
760 			return -ENOMEM;
761 
762 		topa_insert_table(buf, topa);
763 	}
764 
765 	if (topa->z_count == topa->last - 1) {
766 		if (order == TOPA_ENTRY(topa, topa->last - 1)->size)
767 			topa->z_count++;
768 	}
769 
770 	TOPA_ENTRY(topa, -1)->base = page_to_phys(p) >> TOPA_SHIFT;
771 	TOPA_ENTRY(topa, -1)->size = order;
772 	if (!buf->snapshot &&
773 	    !intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries)) {
774 		TOPA_ENTRY(topa, -1)->intr = 1;
775 		TOPA_ENTRY(topa, -1)->stop = 1;
776 	}
777 
778 	topa->last++;
779 	topa->size += sizes(order);
780 
781 	buf->nr_pages += 1ul << order;
782 
783 	return 0;
784 }
785 
786 /**
787  * pt_topa_dump() - print ToPA tables and their entries
788  * @buf:	PT buffer.
789  */
790 static void pt_topa_dump(struct pt_buffer *buf)
791 {
792 	struct topa *topa;
793 
794 	list_for_each_entry(topa, &buf->tables, list) {
795 		struct topa_page *tp = topa_to_page(topa);
796 		int i;
797 
798 		pr_debug("# table @%p, off %llx size %zx\n", tp->table,
799 			 topa->offset, topa->size);
800 		for (i = 0; i < TENTS_PER_PAGE; i++) {
801 			pr_debug("# entry @%p (%lx sz %u %c%c%c) raw=%16llx\n",
802 				 &tp->table[i],
803 				 (unsigned long)tp->table[i].base << TOPA_SHIFT,
804 				 sizes(tp->table[i].size),
805 				 tp->table[i].end ?  'E' : ' ',
806 				 tp->table[i].intr ? 'I' : ' ',
807 				 tp->table[i].stop ? 'S' : ' ',
808 				 *(u64 *)&tp->table[i]);
809 			if ((intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) &&
810 			     tp->table[i].stop) ||
811 			    tp->table[i].end)
812 				break;
813 			if (!i && topa->z_count)
814 				i += topa->z_count;
815 		}
816 	}
817 }
818 
819 /**
820  * pt_buffer_advance() - advance to the next output region
821  * @buf:	PT buffer.
822  *
823  * Advance the current pointers in the buffer to the next ToPA entry.
824  */
825 static void pt_buffer_advance(struct pt_buffer *buf)
826 {
827 	buf->output_off = 0;
828 	buf->cur_idx++;
829 
830 	if (buf->cur_idx == buf->cur->last) {
831 		if (buf->cur == buf->last)
832 			buf->cur = buf->first;
833 		else
834 			buf->cur = list_entry(buf->cur->list.next, struct topa,
835 					      list);
836 		buf->cur_idx = 0;
837 	}
838 }
839 
840 /**
841  * pt_update_head() - calculate current offsets and sizes
842  * @pt:		Per-cpu pt context.
843  *
844  * Update buffer's current write pointer position and data size.
845  */
846 static void pt_update_head(struct pt *pt)
847 {
848 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
849 	u64 topa_idx, base, old;
850 
851 	if (buf->single) {
852 		local_set(&buf->data_size, buf->output_off);
853 		return;
854 	}
855 
856 	/* offset of the first region in this table from the beginning of buf */
857 	base = buf->cur->offset + buf->output_off;
858 
859 	/* offset of the current output region within this table */
860 	for (topa_idx = 0; topa_idx < buf->cur_idx; topa_idx++)
861 		base += TOPA_ENTRY_SIZE(buf->cur, topa_idx);
862 
863 	if (buf->snapshot) {
864 		local_set(&buf->data_size, base);
865 	} else {
866 		old = (local64_xchg(&buf->head, base) &
867 		       ((buf->nr_pages << PAGE_SHIFT) - 1));
868 		if (base < old)
869 			base += buf->nr_pages << PAGE_SHIFT;
870 
871 		local_add(base - old, &buf->data_size);
872 	}
873 }
874 
875 /**
876  * pt_buffer_region() - obtain current output region's address
877  * @buf:	PT buffer.
878  */
879 static void *pt_buffer_region(struct pt_buffer *buf)
880 {
881 	return phys_to_virt((phys_addr_t)TOPA_ENTRY(buf->cur, buf->cur_idx)->base << TOPA_SHIFT);
882 }
883 
884 /**
885  * pt_buffer_region_size() - obtain current output region's size
886  * @buf:	PT buffer.
887  */
888 static size_t pt_buffer_region_size(struct pt_buffer *buf)
889 {
890 	return TOPA_ENTRY_SIZE(buf->cur, buf->cur_idx);
891 }
892 
893 /**
894  * pt_handle_status() - take care of possible status conditions
895  * @pt:		Per-cpu pt context.
896  */
897 static void pt_handle_status(struct pt *pt)
898 {
899 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
900 	int advance = 0;
901 	u64 status;
902 
903 	rdmsrl(MSR_IA32_RTIT_STATUS, status);
904 
905 	if (status & RTIT_STATUS_ERROR) {
906 		pr_err_ratelimited("ToPA ERROR encountered, trying to recover\n");
907 		pt_topa_dump(buf);
908 		status &= ~RTIT_STATUS_ERROR;
909 	}
910 
911 	if (status & RTIT_STATUS_STOPPED) {
912 		status &= ~RTIT_STATUS_STOPPED;
913 
914 		/*
915 		 * On systems that only do single-entry ToPA, hitting STOP
916 		 * means we are already losing data; need to let the decoder
917 		 * know.
918 		 */
919 		if (!buf->single &&
920 		    (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) ||
921 		     buf->output_off == pt_buffer_region_size(buf))) {
922 			perf_aux_output_flag(&pt->handle,
923 			                     PERF_AUX_FLAG_TRUNCATED);
924 			advance++;
925 		}
926 	}
927 
928 	/*
929 	 * Also on single-entry ToPA implementations, interrupt will come
930 	 * before the output reaches its output region's boundary.
931 	 */
932 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries) &&
933 	    !buf->snapshot &&
934 	    pt_buffer_region_size(buf) - buf->output_off <= TOPA_PMI_MARGIN) {
935 		void *head = pt_buffer_region(buf);
936 
937 		/* everything within this margin needs to be zeroed out */
938 		memset(head + buf->output_off, 0,
939 		       pt_buffer_region_size(buf) -
940 		       buf->output_off);
941 		advance++;
942 	}
943 
944 	if (advance)
945 		pt_buffer_advance(buf);
946 
947 	wrmsrl(MSR_IA32_RTIT_STATUS, status);
948 }
949 
950 /**
951  * pt_read_offset() - translate registers into buffer pointers
952  * @buf:	PT buffer.
953  *
954  * Set buffer's output pointers from MSR values.
955  */
956 static void pt_read_offset(struct pt_buffer *buf)
957 {
958 	struct pt *pt = this_cpu_ptr(&pt_ctx);
959 	struct topa_page *tp;
960 
961 	if (!buf->single) {
962 		rdmsrl(MSR_IA32_RTIT_OUTPUT_BASE, pt->output_base);
963 		tp = phys_to_virt(pt->output_base);
964 		buf->cur = &tp->topa;
965 	}
966 
967 	rdmsrl(MSR_IA32_RTIT_OUTPUT_MASK, pt->output_mask);
968 	/* offset within current output region */
969 	buf->output_off = pt->output_mask >> 32;
970 	/* index of current output region within this table */
971 	if (!buf->single)
972 		buf->cur_idx = (pt->output_mask & 0xffffff80) >> 7;
973 }
974 
975 static struct topa_entry *
976 pt_topa_entry_for_page(struct pt_buffer *buf, unsigned int pg)
977 {
978 	struct topa_page *tp;
979 	struct topa *topa;
980 	unsigned int idx, cur_pg = 0, z_pg = 0, start_idx = 0;
981 
982 	/*
983 	 * Indicates a bug in the caller.
984 	 */
985 	if (WARN_ON_ONCE(pg >= buf->nr_pages))
986 		return NULL;
987 
988 	/*
989 	 * First, find the ToPA table where @pg fits. With high
990 	 * order allocations, there shouldn't be many of these.
991 	 */
992 	list_for_each_entry(topa, &buf->tables, list) {
993 		if (topa->offset + topa->size > (unsigned long)pg << PAGE_SHIFT)
994 			goto found;
995 	}
996 
997 	/*
998 	 * Hitting this means we have a problem in the ToPA
999 	 * allocation code.
1000 	 */
1001 	WARN_ON_ONCE(1);
1002 
1003 	return NULL;
1004 
1005 found:
1006 	/*
1007 	 * Indicates a problem in the ToPA allocation code.
1008 	 */
1009 	if (WARN_ON_ONCE(topa->last == -1))
1010 		return NULL;
1011 
1012 	tp = topa_to_page(topa);
1013 	cur_pg = PFN_DOWN(topa->offset);
1014 	if (topa->z_count) {
1015 		z_pg = TOPA_ENTRY_PAGES(topa, 0) * (topa->z_count + 1);
1016 		start_idx = topa->z_count + 1;
1017 	}
1018 
1019 	/*
1020 	 * Multiple entries at the beginning of the table have the same size,
1021 	 * ideally all of them; if @pg falls there, the search is done.
1022 	 */
1023 	if (pg >= cur_pg && pg < cur_pg + z_pg) {
1024 		idx = (pg - cur_pg) / TOPA_ENTRY_PAGES(topa, 0);
1025 		return &tp->table[idx];
1026 	}
1027 
1028 	/*
1029 	 * Otherwise, slow path: iterate through the remaining entries.
1030 	 */
1031 	for (idx = start_idx, cur_pg += z_pg; idx < topa->last; idx++) {
1032 		if (cur_pg + TOPA_ENTRY_PAGES(topa, idx) > pg)
1033 			return &tp->table[idx];
1034 
1035 		cur_pg += TOPA_ENTRY_PAGES(topa, idx);
1036 	}
1037 
1038 	/*
1039 	 * Means we couldn't find a ToPA entry in the table that does match.
1040 	 */
1041 	WARN_ON_ONCE(1);
1042 
1043 	return NULL;
1044 }
1045 
1046 static struct topa_entry *
1047 pt_topa_prev_entry(struct pt_buffer *buf, struct topa_entry *te)
1048 {
1049 	unsigned long table = (unsigned long)te & ~(PAGE_SIZE - 1);
1050 	struct topa_page *tp;
1051 	struct topa *topa;
1052 
1053 	tp = (struct topa_page *)table;
1054 	if (tp->table != te)
1055 		return --te;
1056 
1057 	topa = &tp->topa;
1058 	if (topa == buf->first)
1059 		topa = buf->last;
1060 	else
1061 		topa = list_prev_entry(topa, list);
1062 
1063 	tp = topa_to_page(topa);
1064 
1065 	return &tp->table[topa->last - 1];
1066 }
1067 
1068 /**
1069  * pt_buffer_reset_markers() - place interrupt and stop bits in the buffer
1070  * @buf:	PT buffer.
1071  * @handle:	Current output handle.
1072  *
1073  * Place INT and STOP marks to prevent overwriting old data that the consumer
1074  * hasn't yet collected and waking up the consumer after a certain fraction of
1075  * the buffer has filled up. Only needed and sensible for non-snapshot counters.
1076  *
1077  * This obviously relies on buf::head to figure out buffer markers, so it has
1078  * to be called after pt_buffer_reset_offsets() and before the hardware tracing
1079  * is enabled.
1080  */
1081 static int pt_buffer_reset_markers(struct pt_buffer *buf,
1082 				   struct perf_output_handle *handle)
1083 
1084 {
1085 	unsigned long head = local64_read(&buf->head);
1086 	unsigned long idx, npages, wakeup;
1087 
1088 	if (buf->single)
1089 		return 0;
1090 
1091 	/* can't stop in the middle of an output region */
1092 	if (buf->output_off + handle->size + 1 < pt_buffer_region_size(buf)) {
1093 		perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED);
1094 		return -EINVAL;
1095 	}
1096 
1097 
1098 	/* single entry ToPA is handled by marking all regions STOP=1 INT=1 */
1099 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
1100 		return 0;
1101 
1102 	/* clear STOP and INT from current entry */
1103 	if (buf->stop_te) {
1104 		buf->stop_te->stop = 0;
1105 		buf->stop_te->intr = 0;
1106 	}
1107 
1108 	if (buf->intr_te)
1109 		buf->intr_te->intr = 0;
1110 
1111 	/* how many pages till the STOP marker */
1112 	npages = handle->size >> PAGE_SHIFT;
1113 
1114 	/* if it's on a page boundary, fill up one more page */
1115 	if (!offset_in_page(head + handle->size + 1))
1116 		npages++;
1117 
1118 	idx = (head >> PAGE_SHIFT) + npages;
1119 	idx &= buf->nr_pages - 1;
1120 
1121 	if (idx != buf->stop_pos) {
1122 		buf->stop_pos = idx;
1123 		buf->stop_te = pt_topa_entry_for_page(buf, idx);
1124 		buf->stop_te = pt_topa_prev_entry(buf, buf->stop_te);
1125 	}
1126 
1127 	wakeup = handle->wakeup >> PAGE_SHIFT;
1128 
1129 	/* in the worst case, wake up the consumer one page before hard stop */
1130 	idx = (head >> PAGE_SHIFT) + npages - 1;
1131 	if (idx > wakeup)
1132 		idx = wakeup;
1133 
1134 	idx &= buf->nr_pages - 1;
1135 	if (idx != buf->intr_pos) {
1136 		buf->intr_pos = idx;
1137 		buf->intr_te = pt_topa_entry_for_page(buf, idx);
1138 		buf->intr_te = pt_topa_prev_entry(buf, buf->intr_te);
1139 	}
1140 
1141 	buf->stop_te->stop = 1;
1142 	buf->stop_te->intr = 1;
1143 	buf->intr_te->intr = 1;
1144 
1145 	return 0;
1146 }
1147 
1148 /**
1149  * pt_buffer_reset_offsets() - adjust buffer's write pointers from aux_head
1150  * @buf:	PT buffer.
1151  * @head:	Write pointer (aux_head) from AUX buffer.
1152  *
1153  * Find the ToPA table and entry corresponding to given @head and set buffer's
1154  * "current" pointers accordingly. This is done after we have obtained the
1155  * current aux_head position from a successful call to perf_aux_output_begin()
1156  * to make sure the hardware is writing to the right place.
1157  *
1158  * This function modifies buf::{cur,cur_idx,output_off} that will be programmed
1159  * into PT msrs when the tracing is enabled and buf::head and buf::data_size,
1160  * which are used to determine INT and STOP markers' locations by a subsequent
1161  * call to pt_buffer_reset_markers().
1162  */
1163 static void pt_buffer_reset_offsets(struct pt_buffer *buf, unsigned long head)
1164 {
1165 	struct topa_page *cur_tp;
1166 	struct topa_entry *te;
1167 	int pg;
1168 
1169 	if (buf->snapshot)
1170 		head &= (buf->nr_pages << PAGE_SHIFT) - 1;
1171 
1172 	if (!buf->single) {
1173 		pg = (head >> PAGE_SHIFT) & (buf->nr_pages - 1);
1174 		te = pt_topa_entry_for_page(buf, pg);
1175 
1176 		cur_tp = topa_entry_to_page(te);
1177 		buf->cur = &cur_tp->topa;
1178 		buf->cur_idx = te - TOPA_ENTRY(buf->cur, 0);
1179 		buf->output_off = head & (pt_buffer_region_size(buf) - 1);
1180 	} else {
1181 		buf->output_off = head;
1182 	}
1183 
1184 	local64_set(&buf->head, head);
1185 	local_set(&buf->data_size, 0);
1186 }
1187 
1188 /**
1189  * pt_buffer_fini_topa() - deallocate ToPA structure of a buffer
1190  * @buf:	PT buffer.
1191  */
1192 static void pt_buffer_fini_topa(struct pt_buffer *buf)
1193 {
1194 	struct topa *topa, *iter;
1195 
1196 	if (buf->single)
1197 		return;
1198 
1199 	list_for_each_entry_safe(topa, iter, &buf->tables, list) {
1200 		/*
1201 		 * right now, this is in free_aux() path only, so
1202 		 * no need to unlink this table from the list
1203 		 */
1204 		topa_free(topa);
1205 	}
1206 }
1207 
1208 /**
1209  * pt_buffer_init_topa() - initialize ToPA table for pt buffer
1210  * @buf:	PT buffer.
1211  * @cpu:	CPU on which to allocate.
1212  * @nr_pages:	No. of pages to allocate.
1213  * @gfp:	Allocation flags.
1214  *
1215  * Return:	0 on success or error code.
1216  */
1217 static int pt_buffer_init_topa(struct pt_buffer *buf, int cpu,
1218 			       unsigned long nr_pages, gfp_t gfp)
1219 {
1220 	struct topa *topa;
1221 	int err;
1222 
1223 	topa = topa_alloc(cpu, gfp);
1224 	if (!topa)
1225 		return -ENOMEM;
1226 
1227 	topa_insert_table(buf, topa);
1228 
1229 	while (buf->nr_pages < nr_pages) {
1230 		err = topa_insert_pages(buf, cpu, gfp);
1231 		if (err) {
1232 			pt_buffer_fini_topa(buf);
1233 			return -ENOMEM;
1234 		}
1235 	}
1236 
1237 	/* link last table to the first one, unless we're double buffering */
1238 	if (intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries)) {
1239 		TOPA_ENTRY(buf->last, -1)->base = topa_pfn(buf->first);
1240 		TOPA_ENTRY(buf->last, -1)->end = 1;
1241 	}
1242 
1243 	pt_topa_dump(buf);
1244 	return 0;
1245 }
1246 
1247 static int pt_buffer_try_single(struct pt_buffer *buf, int nr_pages)
1248 {
1249 	struct page *p = virt_to_page(buf->data_pages[0]);
1250 	int ret = -ENOTSUPP, order = 0;
1251 
1252 	/*
1253 	 * We can use single range output mode
1254 	 * + in snapshot mode, where we don't need interrupts;
1255 	 * + if the hardware supports it;
1256 	 * + if the entire buffer is one contiguous allocation.
1257 	 */
1258 	if (!buf->snapshot)
1259 		goto out;
1260 
1261 	if (!intel_pt_validate_hw_cap(PT_CAP_single_range_output))
1262 		goto out;
1263 
1264 	if (PagePrivate(p))
1265 		order = page_private(p);
1266 
1267 	if (1 << order != nr_pages)
1268 		goto out;
1269 
1270 	/*
1271 	 * Some processors cannot always support single range for more than
1272 	 * 4KB - refer errata TGL052, ADL037 and RPL017. Future processors might
1273 	 * also be affected, so for now rather than trying to keep track of
1274 	 * which ones, just disable it for all.
1275 	 */
1276 	if (nr_pages > 1)
1277 		goto out;
1278 
1279 	buf->single = true;
1280 	buf->nr_pages = nr_pages;
1281 	ret = 0;
1282 out:
1283 	return ret;
1284 }
1285 
1286 /**
1287  * pt_buffer_setup_aux() - set up topa tables for a PT buffer
1288  * @event:	Performance event
1289  * @pages:	Array of pointers to buffer pages passed from perf core.
1290  * @nr_pages:	Number of pages in the buffer.
1291  * @snapshot:	If this is a snapshot/overwrite counter.
1292  *
1293  * This is a pmu::setup_aux callback that sets up ToPA tables and all the
1294  * bookkeeping for an AUX buffer.
1295  *
1296  * Return:	Our private PT buffer structure.
1297  */
1298 static void *
1299 pt_buffer_setup_aux(struct perf_event *event, void **pages,
1300 		    int nr_pages, bool snapshot)
1301 {
1302 	struct pt_buffer *buf;
1303 	int node, ret, cpu = event->cpu;
1304 
1305 	if (!nr_pages)
1306 		return NULL;
1307 
1308 	/*
1309 	 * Only support AUX sampling in snapshot mode, where we don't
1310 	 * generate NMIs.
1311 	 */
1312 	if (event->attr.aux_sample_size && !snapshot)
1313 		return NULL;
1314 
1315 	if (cpu == -1)
1316 		cpu = raw_smp_processor_id();
1317 	node = cpu_to_node(cpu);
1318 
1319 	buf = kzalloc_node(sizeof(struct pt_buffer), GFP_KERNEL, node);
1320 	if (!buf)
1321 		return NULL;
1322 
1323 	buf->snapshot = snapshot;
1324 	buf->data_pages = pages;
1325 	buf->stop_pos = -1;
1326 	buf->intr_pos = -1;
1327 
1328 	INIT_LIST_HEAD(&buf->tables);
1329 
1330 	ret = pt_buffer_try_single(buf, nr_pages);
1331 	if (!ret)
1332 		return buf;
1333 
1334 	ret = pt_buffer_init_topa(buf, cpu, nr_pages, GFP_KERNEL);
1335 	if (ret) {
1336 		kfree(buf);
1337 		return NULL;
1338 	}
1339 
1340 	return buf;
1341 }
1342 
1343 /**
1344  * pt_buffer_free_aux() - perf AUX deallocation path callback
1345  * @data:	PT buffer.
1346  */
1347 static void pt_buffer_free_aux(void *data)
1348 {
1349 	struct pt_buffer *buf = data;
1350 
1351 	pt_buffer_fini_topa(buf);
1352 	kfree(buf);
1353 }
1354 
1355 static int pt_addr_filters_init(struct perf_event *event)
1356 {
1357 	struct pt_filters *filters;
1358 	int node = event->cpu == -1 ? -1 : cpu_to_node(event->cpu);
1359 
1360 	if (!intel_pt_validate_hw_cap(PT_CAP_num_address_ranges))
1361 		return 0;
1362 
1363 	filters = kzalloc_node(sizeof(struct pt_filters), GFP_KERNEL, node);
1364 	if (!filters)
1365 		return -ENOMEM;
1366 
1367 	if (event->parent)
1368 		memcpy(filters, event->parent->hw.addr_filters,
1369 		       sizeof(*filters));
1370 
1371 	event->hw.addr_filters = filters;
1372 
1373 	return 0;
1374 }
1375 
1376 static void pt_addr_filters_fini(struct perf_event *event)
1377 {
1378 	kfree(event->hw.addr_filters);
1379 	event->hw.addr_filters = NULL;
1380 }
1381 
1382 #ifdef CONFIG_X86_64
1383 /* Clamp to a canonical address greater-than-or-equal-to the address given */
1384 static u64 clamp_to_ge_canonical_addr(u64 vaddr, u8 vaddr_bits)
1385 {
1386 	return __is_canonical_address(vaddr, vaddr_bits) ?
1387 	       vaddr :
1388 	       -BIT_ULL(vaddr_bits - 1);
1389 }
1390 
1391 /* Clamp to a canonical address less-than-or-equal-to the address given */
1392 static u64 clamp_to_le_canonical_addr(u64 vaddr, u8 vaddr_bits)
1393 {
1394 	return __is_canonical_address(vaddr, vaddr_bits) ?
1395 	       vaddr :
1396 	       BIT_ULL(vaddr_bits - 1) - 1;
1397 }
1398 #else
1399 #define clamp_to_ge_canonical_addr(x, y) (x)
1400 #define clamp_to_le_canonical_addr(x, y) (x)
1401 #endif
1402 
1403 static int pt_event_addr_filters_validate(struct list_head *filters)
1404 {
1405 	struct perf_addr_filter *filter;
1406 	int range = 0;
1407 
1408 	list_for_each_entry(filter, filters, entry) {
1409 		/*
1410 		 * PT doesn't support single address triggers and
1411 		 * 'start' filters.
1412 		 */
1413 		if (!filter->size ||
1414 		    filter->action == PERF_ADDR_FILTER_ACTION_START)
1415 			return -EOPNOTSUPP;
1416 
1417 		if (++range > intel_pt_validate_hw_cap(PT_CAP_num_address_ranges))
1418 			return -EOPNOTSUPP;
1419 	}
1420 
1421 	return 0;
1422 }
1423 
1424 static void pt_event_addr_filters_sync(struct perf_event *event)
1425 {
1426 	struct perf_addr_filters_head *head = perf_event_addr_filters(event);
1427 	unsigned long msr_a, msr_b;
1428 	struct perf_addr_filter_range *fr = event->addr_filter_ranges;
1429 	struct pt_filters *filters = event->hw.addr_filters;
1430 	struct perf_addr_filter *filter;
1431 	int range = 0;
1432 
1433 	if (!filters)
1434 		return;
1435 
1436 	list_for_each_entry(filter, &head->list, entry) {
1437 		if (filter->path.dentry && !fr[range].start) {
1438 			msr_a = msr_b = 0;
1439 		} else {
1440 			unsigned long n = fr[range].size - 1;
1441 			unsigned long a = fr[range].start;
1442 			unsigned long b;
1443 
1444 			if (a > ULONG_MAX - n)
1445 				b = ULONG_MAX;
1446 			else
1447 				b = a + n;
1448 			/*
1449 			 * Apply the offset. 64-bit addresses written to the
1450 			 * MSRs must be canonical, but the range can encompass
1451 			 * non-canonical addresses. Since software cannot
1452 			 * execute at non-canonical addresses, adjusting to
1453 			 * canonical addresses does not affect the result of the
1454 			 * address filter.
1455 			 */
1456 			msr_a = clamp_to_ge_canonical_addr(a, boot_cpu_data.x86_virt_bits);
1457 			msr_b = clamp_to_le_canonical_addr(b, boot_cpu_data.x86_virt_bits);
1458 			if (msr_b < msr_a)
1459 				msr_a = msr_b = 0;
1460 		}
1461 
1462 		filters->filter[range].msr_a  = msr_a;
1463 		filters->filter[range].msr_b  = msr_b;
1464 		if (filter->action == PERF_ADDR_FILTER_ACTION_FILTER)
1465 			filters->filter[range].config = 1;
1466 		else
1467 			filters->filter[range].config = 2;
1468 		range++;
1469 	}
1470 
1471 	filters->nr_filters = range;
1472 }
1473 
1474 /**
1475  * intel_pt_interrupt() - PT PMI handler
1476  */
1477 void intel_pt_interrupt(void)
1478 {
1479 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1480 	struct pt_buffer *buf;
1481 	struct perf_event *event = pt->handle.event;
1482 
1483 	/*
1484 	 * There may be a dangling PT bit in the interrupt status register
1485 	 * after PT has been disabled by pt_event_stop(). Make sure we don't
1486 	 * do anything (particularly, re-enable) for this event here.
1487 	 */
1488 	if (!READ_ONCE(pt->handle_nmi))
1489 		return;
1490 
1491 	if (!event)
1492 		return;
1493 
1494 	pt_config_stop(event);
1495 
1496 	buf = perf_get_aux(&pt->handle);
1497 	if (!buf)
1498 		return;
1499 
1500 	pt_read_offset(buf);
1501 
1502 	pt_handle_status(pt);
1503 
1504 	pt_update_head(pt);
1505 
1506 	perf_aux_output_end(&pt->handle, local_xchg(&buf->data_size, 0));
1507 
1508 	if (!event->hw.state) {
1509 		int ret;
1510 
1511 		buf = perf_aux_output_begin(&pt->handle, event);
1512 		if (!buf) {
1513 			event->hw.state = PERF_HES_STOPPED;
1514 			return;
1515 		}
1516 
1517 		pt_buffer_reset_offsets(buf, pt->handle.head);
1518 		/* snapshot counters don't use PMI, so it's safe */
1519 		ret = pt_buffer_reset_markers(buf, &pt->handle);
1520 		if (ret) {
1521 			perf_aux_output_end(&pt->handle, 0);
1522 			return;
1523 		}
1524 
1525 		pt_config_buffer(buf);
1526 		pt_config_start(event);
1527 	}
1528 }
1529 
1530 void intel_pt_handle_vmx(int on)
1531 {
1532 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1533 	struct perf_event *event;
1534 	unsigned long flags;
1535 
1536 	/* PT plays nice with VMX, do nothing */
1537 	if (pt_pmu.vmx)
1538 		return;
1539 
1540 	/*
1541 	 * VMXON will clear RTIT_CTL.TraceEn; we need to make
1542 	 * sure to not try to set it while VMX is on. Disable
1543 	 * interrupts to avoid racing with pmu callbacks;
1544 	 * concurrent PMI should be handled fine.
1545 	 */
1546 	local_irq_save(flags);
1547 	WRITE_ONCE(pt->vmx_on, on);
1548 
1549 	/*
1550 	 * If an AUX transaction is in progress, it will contain
1551 	 * gap(s), so flag it PARTIAL to inform the user.
1552 	 */
1553 	event = pt->handle.event;
1554 	if (event)
1555 		perf_aux_output_flag(&pt->handle,
1556 		                     PERF_AUX_FLAG_PARTIAL);
1557 
1558 	/* Turn PTs back on */
1559 	if (!on && event)
1560 		wrmsrl(MSR_IA32_RTIT_CTL, event->hw.aux_config);
1561 
1562 	local_irq_restore(flags);
1563 }
1564 EXPORT_SYMBOL_GPL(intel_pt_handle_vmx);
1565 
1566 /*
1567  * PMU callbacks
1568  */
1569 
1570 static void pt_event_start(struct perf_event *event, int mode)
1571 {
1572 	struct hw_perf_event *hwc = &event->hw;
1573 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1574 	struct pt_buffer *buf;
1575 
1576 	buf = perf_aux_output_begin(&pt->handle, event);
1577 	if (!buf)
1578 		goto fail_stop;
1579 
1580 	pt_buffer_reset_offsets(buf, pt->handle.head);
1581 	if (!buf->snapshot) {
1582 		if (pt_buffer_reset_markers(buf, &pt->handle))
1583 			goto fail_end_stop;
1584 	}
1585 
1586 	WRITE_ONCE(pt->handle_nmi, 1);
1587 	hwc->state = 0;
1588 
1589 	pt_config_buffer(buf);
1590 	pt_config(event);
1591 
1592 	return;
1593 
1594 fail_end_stop:
1595 	perf_aux_output_end(&pt->handle, 0);
1596 fail_stop:
1597 	hwc->state = PERF_HES_STOPPED;
1598 }
1599 
1600 static void pt_event_stop(struct perf_event *event, int mode)
1601 {
1602 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1603 
1604 	/*
1605 	 * Protect against the PMI racing with disabling wrmsr,
1606 	 * see comment in intel_pt_interrupt().
1607 	 */
1608 	WRITE_ONCE(pt->handle_nmi, 0);
1609 	barrier();
1610 
1611 	pt_config_stop(event);
1612 
1613 	if (event->hw.state == PERF_HES_STOPPED)
1614 		return;
1615 
1616 	event->hw.state = PERF_HES_STOPPED;
1617 
1618 	if (mode & PERF_EF_UPDATE) {
1619 		struct pt_buffer *buf = perf_get_aux(&pt->handle);
1620 
1621 		if (!buf)
1622 			return;
1623 
1624 		if (WARN_ON_ONCE(pt->handle.event != event))
1625 			return;
1626 
1627 		pt_read_offset(buf);
1628 
1629 		pt_handle_status(pt);
1630 
1631 		pt_update_head(pt);
1632 
1633 		if (buf->snapshot)
1634 			pt->handle.head =
1635 				local_xchg(&buf->data_size,
1636 					   buf->nr_pages << PAGE_SHIFT);
1637 		perf_aux_output_end(&pt->handle, local_xchg(&buf->data_size, 0));
1638 	}
1639 }
1640 
1641 static long pt_event_snapshot_aux(struct perf_event *event,
1642 				  struct perf_output_handle *handle,
1643 				  unsigned long size)
1644 {
1645 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1646 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
1647 	unsigned long from = 0, to;
1648 	long ret;
1649 
1650 	if (WARN_ON_ONCE(!buf))
1651 		return 0;
1652 
1653 	/*
1654 	 * Sampling is only allowed on snapshot events;
1655 	 * see pt_buffer_setup_aux().
1656 	 */
1657 	if (WARN_ON_ONCE(!buf->snapshot))
1658 		return 0;
1659 
1660 	/*
1661 	 * There is no PT interrupt in this mode, so stop the trace and it will
1662 	 * remain stopped while the buffer is copied.
1663 	 */
1664 	pt_config_stop(event);
1665 	pt_read_offset(buf);
1666 	pt_update_head(pt);
1667 
1668 	to = local_read(&buf->data_size);
1669 	if (to < size)
1670 		from = buf->nr_pages << PAGE_SHIFT;
1671 	from += to - size;
1672 
1673 	ret = perf_output_copy_aux(&pt->handle, handle, from, to);
1674 
1675 	/*
1676 	 * Here, handle_nmi tells us if the tracing was on.
1677 	 * If the tracing was on, restart it.
1678 	 */
1679 	if (READ_ONCE(pt->handle_nmi))
1680 		pt_config_start(event);
1681 
1682 	return ret;
1683 }
1684 
1685 static void pt_event_del(struct perf_event *event, int mode)
1686 {
1687 	pt_event_stop(event, PERF_EF_UPDATE);
1688 }
1689 
1690 static int pt_event_add(struct perf_event *event, int mode)
1691 {
1692 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1693 	struct hw_perf_event *hwc = &event->hw;
1694 	int ret = -EBUSY;
1695 
1696 	if (pt->handle.event)
1697 		goto fail;
1698 
1699 	if (mode & PERF_EF_START) {
1700 		pt_event_start(event, 0);
1701 		ret = -EINVAL;
1702 		if (hwc->state == PERF_HES_STOPPED)
1703 			goto fail;
1704 	} else {
1705 		hwc->state = PERF_HES_STOPPED;
1706 	}
1707 
1708 	ret = 0;
1709 fail:
1710 
1711 	return ret;
1712 }
1713 
1714 static void pt_event_read(struct perf_event *event)
1715 {
1716 }
1717 
1718 static void pt_event_destroy(struct perf_event *event)
1719 {
1720 	pt_addr_filters_fini(event);
1721 	x86_del_exclusive(x86_lbr_exclusive_pt);
1722 }
1723 
1724 static int pt_event_init(struct perf_event *event)
1725 {
1726 	if (event->attr.type != pt_pmu.pmu.type)
1727 		return -ENOENT;
1728 
1729 	if (!pt_event_valid(event))
1730 		return -EINVAL;
1731 
1732 	if (x86_add_exclusive(x86_lbr_exclusive_pt))
1733 		return -EBUSY;
1734 
1735 	if (pt_addr_filters_init(event)) {
1736 		x86_del_exclusive(x86_lbr_exclusive_pt);
1737 		return -ENOMEM;
1738 	}
1739 
1740 	event->destroy = pt_event_destroy;
1741 
1742 	return 0;
1743 }
1744 
1745 void cpu_emergency_stop_pt(void)
1746 {
1747 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1748 
1749 	if (pt->handle.event)
1750 		pt_event_stop(pt->handle.event, PERF_EF_UPDATE);
1751 }
1752 
1753 int is_intel_pt_event(struct perf_event *event)
1754 {
1755 	return event->pmu == &pt_pmu.pmu;
1756 }
1757 
1758 static __init int pt_init(void)
1759 {
1760 	int ret, cpu, prior_warn = 0;
1761 
1762 	BUILD_BUG_ON(sizeof(struct topa) > PAGE_SIZE);
1763 
1764 	if (!boot_cpu_has(X86_FEATURE_INTEL_PT))
1765 		return -ENODEV;
1766 
1767 	cpus_read_lock();
1768 	for_each_online_cpu(cpu) {
1769 		u64 ctl;
1770 
1771 		ret = rdmsrl_safe_on_cpu(cpu, MSR_IA32_RTIT_CTL, &ctl);
1772 		if (!ret && (ctl & RTIT_CTL_TRACEEN))
1773 			prior_warn++;
1774 	}
1775 	cpus_read_unlock();
1776 
1777 	if (prior_warn) {
1778 		x86_add_exclusive(x86_lbr_exclusive_pt);
1779 		pr_warn("PT is enabled at boot time, doing nothing\n");
1780 
1781 		return -EBUSY;
1782 	}
1783 
1784 	ret = pt_pmu_hw_init();
1785 	if (ret)
1786 		return ret;
1787 
1788 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_output)) {
1789 		pr_warn("ToPA output is not supported on this CPU\n");
1790 		return -ENODEV;
1791 	}
1792 
1793 	if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
1794 		pt_pmu.pmu.capabilities = PERF_PMU_CAP_AUX_NO_SG;
1795 
1796 	pt_pmu.pmu.capabilities	|= PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE;
1797 	pt_pmu.pmu.attr_groups		 = pt_attr_groups;
1798 	pt_pmu.pmu.task_ctx_nr		 = perf_sw_context;
1799 	pt_pmu.pmu.event_init		 = pt_event_init;
1800 	pt_pmu.pmu.add			 = pt_event_add;
1801 	pt_pmu.pmu.del			 = pt_event_del;
1802 	pt_pmu.pmu.start		 = pt_event_start;
1803 	pt_pmu.pmu.stop			 = pt_event_stop;
1804 	pt_pmu.pmu.snapshot_aux		 = pt_event_snapshot_aux;
1805 	pt_pmu.pmu.read			 = pt_event_read;
1806 	pt_pmu.pmu.setup_aux		 = pt_buffer_setup_aux;
1807 	pt_pmu.pmu.free_aux		 = pt_buffer_free_aux;
1808 	pt_pmu.pmu.addr_filters_sync     = pt_event_addr_filters_sync;
1809 	pt_pmu.pmu.addr_filters_validate = pt_event_addr_filters_validate;
1810 	pt_pmu.pmu.nr_addr_filters       =
1811 		intel_pt_validate_hw_cap(PT_CAP_num_address_ranges);
1812 
1813 	ret = perf_pmu_register(&pt_pmu.pmu, "intel_pt", -1);
1814 
1815 	return ret;
1816 }
1817 arch_initcall(pt_init);
1818