xref: /linux/arch/x86/events/intel/pt.c (revision 110e6f26af80dfd90b6e5c645b1aed7228aa580d)
1 /*
2  * Intel(R) Processor Trace PMU driver for perf
3  * Copyright (c) 2013-2014, Intel Corporation.
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms and conditions of the GNU General Public License,
7  * version 2, as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  *
14  * Intel PT is specified in the Intel Architecture Instruction Set Extensions
15  * Programming Reference:
16  * http://software.intel.com/en-us/intel-isa-extensions
17  */
18 
19 #undef DEBUG
20 
21 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
22 
23 #include <linux/types.h>
24 #include <linux/slab.h>
25 #include <linux/device.h>
26 
27 #include <asm/perf_event.h>
28 #include <asm/insn.h>
29 #include <asm/io.h>
30 #include <asm/intel_pt.h>
31 
32 #include "../perf_event.h"
33 #include "pt.h"
34 
35 static DEFINE_PER_CPU(struct pt, pt_ctx);
36 
37 static struct pt_pmu pt_pmu;
38 
39 enum cpuid_regs {
40 	CR_EAX = 0,
41 	CR_ECX,
42 	CR_EDX,
43 	CR_EBX
44 };
45 
46 /*
47  * Capabilities of Intel PT hardware, such as number of address bits or
48  * supported output schemes, are cached and exported to userspace as "caps"
49  * attribute group of pt pmu device
50  * (/sys/bus/event_source/devices/intel_pt/caps/) so that userspace can store
51  * relevant bits together with intel_pt traces.
52  *
53  * These are necessary for both trace decoding (payloads_lip, contains address
54  * width encoded in IP-related packets), and event configuration (bitmasks with
55  * permitted values for certain bit fields).
56  */
57 #define PT_CAP(_n, _l, _r, _m)						\
58 	[PT_CAP_ ## _n] = { .name = __stringify(_n), .leaf = _l,	\
59 			    .reg = _r, .mask = _m }
60 
61 static struct pt_cap_desc {
62 	const char	*name;
63 	u32		leaf;
64 	u8		reg;
65 	u32		mask;
66 } pt_caps[] = {
67 	PT_CAP(max_subleaf,		0, CR_EAX, 0xffffffff),
68 	PT_CAP(cr3_filtering,		0, CR_EBX, BIT(0)),
69 	PT_CAP(psb_cyc,			0, CR_EBX, BIT(1)),
70 	PT_CAP(mtc,			0, CR_EBX, BIT(3)),
71 	PT_CAP(topa_output,		0, CR_ECX, BIT(0)),
72 	PT_CAP(topa_multiple_entries,	0, CR_ECX, BIT(1)),
73 	PT_CAP(single_range_output,	0, CR_ECX, BIT(2)),
74 	PT_CAP(payloads_lip,		0, CR_ECX, BIT(31)),
75 	PT_CAP(mtc_periods,		1, CR_EAX, 0xffff0000),
76 	PT_CAP(cycle_thresholds,	1, CR_EBX, 0xffff),
77 	PT_CAP(psb_periods,		1, CR_EBX, 0xffff0000),
78 };
79 
80 static u32 pt_cap_get(enum pt_capabilities cap)
81 {
82 	struct pt_cap_desc *cd = &pt_caps[cap];
83 	u32 c = pt_pmu.caps[cd->leaf * PT_CPUID_REGS_NUM + cd->reg];
84 	unsigned int shift = __ffs(cd->mask);
85 
86 	return (c & cd->mask) >> shift;
87 }
88 
89 static ssize_t pt_cap_show(struct device *cdev,
90 			   struct device_attribute *attr,
91 			   char *buf)
92 {
93 	struct dev_ext_attribute *ea =
94 		container_of(attr, struct dev_ext_attribute, attr);
95 	enum pt_capabilities cap = (long)ea->var;
96 
97 	return snprintf(buf, PAGE_SIZE, "%x\n", pt_cap_get(cap));
98 }
99 
100 static struct attribute_group pt_cap_group = {
101 	.name	= "caps",
102 };
103 
104 PMU_FORMAT_ATTR(cyc,		"config:1"	);
105 PMU_FORMAT_ATTR(mtc,		"config:9"	);
106 PMU_FORMAT_ATTR(tsc,		"config:10"	);
107 PMU_FORMAT_ATTR(noretcomp,	"config:11"	);
108 PMU_FORMAT_ATTR(mtc_period,	"config:14-17"	);
109 PMU_FORMAT_ATTR(cyc_thresh,	"config:19-22"	);
110 PMU_FORMAT_ATTR(psb_period,	"config:24-27"	);
111 
112 static struct attribute *pt_formats_attr[] = {
113 	&format_attr_cyc.attr,
114 	&format_attr_mtc.attr,
115 	&format_attr_tsc.attr,
116 	&format_attr_noretcomp.attr,
117 	&format_attr_mtc_period.attr,
118 	&format_attr_cyc_thresh.attr,
119 	&format_attr_psb_period.attr,
120 	NULL,
121 };
122 
123 static struct attribute_group pt_format_group = {
124 	.name	= "format",
125 	.attrs	= pt_formats_attr,
126 };
127 
128 static const struct attribute_group *pt_attr_groups[] = {
129 	&pt_cap_group,
130 	&pt_format_group,
131 	NULL,
132 };
133 
134 static int __init pt_pmu_hw_init(void)
135 {
136 	struct dev_ext_attribute *de_attrs;
137 	struct attribute **attrs;
138 	size_t size;
139 	int ret;
140 	long i;
141 
142 	attrs = NULL;
143 
144 	for (i = 0; i < PT_CPUID_LEAVES; i++) {
145 		cpuid_count(20, i,
146 			    &pt_pmu.caps[CR_EAX + i*PT_CPUID_REGS_NUM],
147 			    &pt_pmu.caps[CR_EBX + i*PT_CPUID_REGS_NUM],
148 			    &pt_pmu.caps[CR_ECX + i*PT_CPUID_REGS_NUM],
149 			    &pt_pmu.caps[CR_EDX + i*PT_CPUID_REGS_NUM]);
150 	}
151 
152 	ret = -ENOMEM;
153 	size = sizeof(struct attribute *) * (ARRAY_SIZE(pt_caps)+1);
154 	attrs = kzalloc(size, GFP_KERNEL);
155 	if (!attrs)
156 		goto fail;
157 
158 	size = sizeof(struct dev_ext_attribute) * (ARRAY_SIZE(pt_caps)+1);
159 	de_attrs = kzalloc(size, GFP_KERNEL);
160 	if (!de_attrs)
161 		goto fail;
162 
163 	for (i = 0; i < ARRAY_SIZE(pt_caps); i++) {
164 		struct dev_ext_attribute *de_attr = de_attrs + i;
165 
166 		de_attr->attr.attr.name = pt_caps[i].name;
167 
168 		sysfs_attr_init(&de_attr->attr.attr);
169 
170 		de_attr->attr.attr.mode		= S_IRUGO;
171 		de_attr->attr.show		= pt_cap_show;
172 		de_attr->var			= (void *)i;
173 
174 		attrs[i] = &de_attr->attr.attr;
175 	}
176 
177 	pt_cap_group.attrs = attrs;
178 
179 	return 0;
180 
181 fail:
182 	kfree(attrs);
183 
184 	return ret;
185 }
186 
187 #define RTIT_CTL_CYC_PSB (RTIT_CTL_CYCLEACC	| \
188 			  RTIT_CTL_CYC_THRESH	| \
189 			  RTIT_CTL_PSB_FREQ)
190 
191 #define RTIT_CTL_MTC	(RTIT_CTL_MTC_EN	| \
192 			 RTIT_CTL_MTC_RANGE)
193 
194 #define PT_CONFIG_MASK (RTIT_CTL_TSC_EN		| \
195 			RTIT_CTL_DISRETC	| \
196 			RTIT_CTL_CYC_PSB	| \
197 			RTIT_CTL_MTC)
198 
199 static bool pt_event_valid(struct perf_event *event)
200 {
201 	u64 config = event->attr.config;
202 	u64 allowed, requested;
203 
204 	if ((config & PT_CONFIG_MASK) != config)
205 		return false;
206 
207 	if (config & RTIT_CTL_CYC_PSB) {
208 		if (!pt_cap_get(PT_CAP_psb_cyc))
209 			return false;
210 
211 		allowed = pt_cap_get(PT_CAP_psb_periods);
212 		requested = (config & RTIT_CTL_PSB_FREQ) >>
213 			RTIT_CTL_PSB_FREQ_OFFSET;
214 		if (requested && (!(allowed & BIT(requested))))
215 			return false;
216 
217 		allowed = pt_cap_get(PT_CAP_cycle_thresholds);
218 		requested = (config & RTIT_CTL_CYC_THRESH) >>
219 			RTIT_CTL_CYC_THRESH_OFFSET;
220 		if (requested && (!(allowed & BIT(requested))))
221 			return false;
222 	}
223 
224 	if (config & RTIT_CTL_MTC) {
225 		/*
226 		 * In the unlikely case that CPUID lists valid mtc periods,
227 		 * but not the mtc capability, drop out here.
228 		 *
229 		 * Spec says that setting mtc period bits while mtc bit in
230 		 * CPUID is 0 will #GP, so better safe than sorry.
231 		 */
232 		if (!pt_cap_get(PT_CAP_mtc))
233 			return false;
234 
235 		allowed = pt_cap_get(PT_CAP_mtc_periods);
236 		if (!allowed)
237 			return false;
238 
239 		requested = (config & RTIT_CTL_MTC_RANGE) >>
240 			RTIT_CTL_MTC_RANGE_OFFSET;
241 
242 		if (!(allowed & BIT(requested)))
243 			return false;
244 	}
245 
246 	return true;
247 }
248 
249 /*
250  * PT configuration helpers
251  * These all are cpu affine and operate on a local PT
252  */
253 
254 static void pt_config(struct perf_event *event)
255 {
256 	u64 reg;
257 
258 	if (!event->hw.itrace_started) {
259 		event->hw.itrace_started = 1;
260 		wrmsrl(MSR_IA32_RTIT_STATUS, 0);
261 	}
262 
263 	reg = RTIT_CTL_TOPA | RTIT_CTL_BRANCH_EN | RTIT_CTL_TRACEEN;
264 
265 	if (!event->attr.exclude_kernel)
266 		reg |= RTIT_CTL_OS;
267 	if (!event->attr.exclude_user)
268 		reg |= RTIT_CTL_USR;
269 
270 	reg |= (event->attr.config & PT_CONFIG_MASK);
271 
272 	wrmsrl(MSR_IA32_RTIT_CTL, reg);
273 }
274 
275 static void pt_config_start(bool start)
276 {
277 	u64 ctl;
278 
279 	rdmsrl(MSR_IA32_RTIT_CTL, ctl);
280 	if (start)
281 		ctl |= RTIT_CTL_TRACEEN;
282 	else
283 		ctl &= ~RTIT_CTL_TRACEEN;
284 	wrmsrl(MSR_IA32_RTIT_CTL, ctl);
285 
286 	/*
287 	 * A wrmsr that disables trace generation serializes other PT
288 	 * registers and causes all data packets to be written to memory,
289 	 * but a fence is required for the data to become globally visible.
290 	 *
291 	 * The below WMB, separating data store and aux_head store matches
292 	 * the consumer's RMB that separates aux_head load and data load.
293 	 */
294 	if (!start)
295 		wmb();
296 }
297 
298 static void pt_config_buffer(void *buf, unsigned int topa_idx,
299 			     unsigned int output_off)
300 {
301 	u64 reg;
302 
303 	wrmsrl(MSR_IA32_RTIT_OUTPUT_BASE, virt_to_phys(buf));
304 
305 	reg = 0x7f | ((u64)topa_idx << 7) | ((u64)output_off << 32);
306 
307 	wrmsrl(MSR_IA32_RTIT_OUTPUT_MASK, reg);
308 }
309 
310 /*
311  * Keep ToPA table-related metadata on the same page as the actual table,
312  * taking up a few words from the top
313  */
314 
315 #define TENTS_PER_PAGE (((PAGE_SIZE - 40) / sizeof(struct topa_entry)) - 1)
316 
317 /**
318  * struct topa - page-sized ToPA table with metadata at the top
319  * @table:	actual ToPA table entries, as understood by PT hardware
320  * @list:	linkage to struct pt_buffer's list of tables
321  * @phys:	physical address of this page
322  * @offset:	offset of the first entry in this table in the buffer
323  * @size:	total size of all entries in this table
324  * @last:	index of the last initialized entry in this table
325  */
326 struct topa {
327 	struct topa_entry	table[TENTS_PER_PAGE];
328 	struct list_head	list;
329 	u64			phys;
330 	u64			offset;
331 	size_t			size;
332 	int			last;
333 };
334 
335 /* make -1 stand for the last table entry */
336 #define TOPA_ENTRY(t, i) ((i) == -1 ? &(t)->table[(t)->last] : &(t)->table[(i)])
337 
338 /**
339  * topa_alloc() - allocate page-sized ToPA table
340  * @cpu:	CPU on which to allocate.
341  * @gfp:	Allocation flags.
342  *
343  * Return:	On success, return the pointer to ToPA table page.
344  */
345 static struct topa *topa_alloc(int cpu, gfp_t gfp)
346 {
347 	int node = cpu_to_node(cpu);
348 	struct topa *topa;
349 	struct page *p;
350 
351 	p = alloc_pages_node(node, gfp | __GFP_ZERO, 0);
352 	if (!p)
353 		return NULL;
354 
355 	topa = page_address(p);
356 	topa->last = 0;
357 	topa->phys = page_to_phys(p);
358 
359 	/*
360 	 * In case of singe-entry ToPA, always put the self-referencing END
361 	 * link as the 2nd entry in the table
362 	 */
363 	if (!pt_cap_get(PT_CAP_topa_multiple_entries)) {
364 		TOPA_ENTRY(topa, 1)->base = topa->phys >> TOPA_SHIFT;
365 		TOPA_ENTRY(topa, 1)->end = 1;
366 	}
367 
368 	return topa;
369 }
370 
371 /**
372  * topa_free() - free a page-sized ToPA table
373  * @topa:	Table to deallocate.
374  */
375 static void topa_free(struct topa *topa)
376 {
377 	free_page((unsigned long)topa);
378 }
379 
380 /**
381  * topa_insert_table() - insert a ToPA table into a buffer
382  * @buf:	 PT buffer that's being extended.
383  * @topa:	 New topa table to be inserted.
384  *
385  * If it's the first table in this buffer, set up buffer's pointers
386  * accordingly; otherwise, add a END=1 link entry to @topa to the current
387  * "last" table and adjust the last table pointer to @topa.
388  */
389 static void topa_insert_table(struct pt_buffer *buf, struct topa *topa)
390 {
391 	struct topa *last = buf->last;
392 
393 	list_add_tail(&topa->list, &buf->tables);
394 
395 	if (!buf->first) {
396 		buf->first = buf->last = buf->cur = topa;
397 		return;
398 	}
399 
400 	topa->offset = last->offset + last->size;
401 	buf->last = topa;
402 
403 	if (!pt_cap_get(PT_CAP_topa_multiple_entries))
404 		return;
405 
406 	BUG_ON(last->last != TENTS_PER_PAGE - 1);
407 
408 	TOPA_ENTRY(last, -1)->base = topa->phys >> TOPA_SHIFT;
409 	TOPA_ENTRY(last, -1)->end = 1;
410 }
411 
412 /**
413  * topa_table_full() - check if a ToPA table is filled up
414  * @topa:	ToPA table.
415  */
416 static bool topa_table_full(struct topa *topa)
417 {
418 	/* single-entry ToPA is a special case */
419 	if (!pt_cap_get(PT_CAP_topa_multiple_entries))
420 		return !!topa->last;
421 
422 	return topa->last == TENTS_PER_PAGE - 1;
423 }
424 
425 /**
426  * topa_insert_pages() - create a list of ToPA tables
427  * @buf:	PT buffer being initialized.
428  * @gfp:	Allocation flags.
429  *
430  * This initializes a list of ToPA tables with entries from
431  * the data_pages provided by rb_alloc_aux().
432  *
433  * Return:	0 on success or error code.
434  */
435 static int topa_insert_pages(struct pt_buffer *buf, gfp_t gfp)
436 {
437 	struct topa *topa = buf->last;
438 	int order = 0;
439 	struct page *p;
440 
441 	p = virt_to_page(buf->data_pages[buf->nr_pages]);
442 	if (PagePrivate(p))
443 		order = page_private(p);
444 
445 	if (topa_table_full(topa)) {
446 		topa = topa_alloc(buf->cpu, gfp);
447 		if (!topa)
448 			return -ENOMEM;
449 
450 		topa_insert_table(buf, topa);
451 	}
452 
453 	TOPA_ENTRY(topa, -1)->base = page_to_phys(p) >> TOPA_SHIFT;
454 	TOPA_ENTRY(topa, -1)->size = order;
455 	if (!buf->snapshot && !pt_cap_get(PT_CAP_topa_multiple_entries)) {
456 		TOPA_ENTRY(topa, -1)->intr = 1;
457 		TOPA_ENTRY(topa, -1)->stop = 1;
458 	}
459 
460 	topa->last++;
461 	topa->size += sizes(order);
462 
463 	buf->nr_pages += 1ul << order;
464 
465 	return 0;
466 }
467 
468 /**
469  * pt_topa_dump() - print ToPA tables and their entries
470  * @buf:	PT buffer.
471  */
472 static void pt_topa_dump(struct pt_buffer *buf)
473 {
474 	struct topa *topa;
475 
476 	list_for_each_entry(topa, &buf->tables, list) {
477 		int i;
478 
479 		pr_debug("# table @%p (%016Lx), off %llx size %zx\n", topa->table,
480 			 topa->phys, topa->offset, topa->size);
481 		for (i = 0; i < TENTS_PER_PAGE; i++) {
482 			pr_debug("# entry @%p (%lx sz %u %c%c%c) raw=%16llx\n",
483 				 &topa->table[i],
484 				 (unsigned long)topa->table[i].base << TOPA_SHIFT,
485 				 sizes(topa->table[i].size),
486 				 topa->table[i].end ?  'E' : ' ',
487 				 topa->table[i].intr ? 'I' : ' ',
488 				 topa->table[i].stop ? 'S' : ' ',
489 				 *(u64 *)&topa->table[i]);
490 			if ((pt_cap_get(PT_CAP_topa_multiple_entries) &&
491 			     topa->table[i].stop) ||
492 			    topa->table[i].end)
493 				break;
494 		}
495 	}
496 }
497 
498 /**
499  * pt_buffer_advance() - advance to the next output region
500  * @buf:	PT buffer.
501  *
502  * Advance the current pointers in the buffer to the next ToPA entry.
503  */
504 static void pt_buffer_advance(struct pt_buffer *buf)
505 {
506 	buf->output_off = 0;
507 	buf->cur_idx++;
508 
509 	if (buf->cur_idx == buf->cur->last) {
510 		if (buf->cur == buf->last)
511 			buf->cur = buf->first;
512 		else
513 			buf->cur = list_entry(buf->cur->list.next, struct topa,
514 					      list);
515 		buf->cur_idx = 0;
516 	}
517 }
518 
519 /**
520  * pt_update_head() - calculate current offsets and sizes
521  * @pt:		Per-cpu pt context.
522  *
523  * Update buffer's current write pointer position and data size.
524  */
525 static void pt_update_head(struct pt *pt)
526 {
527 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
528 	u64 topa_idx, base, old;
529 
530 	/* offset of the first region in this table from the beginning of buf */
531 	base = buf->cur->offset + buf->output_off;
532 
533 	/* offset of the current output region within this table */
534 	for (topa_idx = 0; topa_idx < buf->cur_idx; topa_idx++)
535 		base += sizes(buf->cur->table[topa_idx].size);
536 
537 	if (buf->snapshot) {
538 		local_set(&buf->data_size, base);
539 	} else {
540 		old = (local64_xchg(&buf->head, base) &
541 		       ((buf->nr_pages << PAGE_SHIFT) - 1));
542 		if (base < old)
543 			base += buf->nr_pages << PAGE_SHIFT;
544 
545 		local_add(base - old, &buf->data_size);
546 	}
547 }
548 
549 /**
550  * pt_buffer_region() - obtain current output region's address
551  * @buf:	PT buffer.
552  */
553 static void *pt_buffer_region(struct pt_buffer *buf)
554 {
555 	return phys_to_virt(buf->cur->table[buf->cur_idx].base << TOPA_SHIFT);
556 }
557 
558 /**
559  * pt_buffer_region_size() - obtain current output region's size
560  * @buf:	PT buffer.
561  */
562 static size_t pt_buffer_region_size(struct pt_buffer *buf)
563 {
564 	return sizes(buf->cur->table[buf->cur_idx].size);
565 }
566 
567 /**
568  * pt_handle_status() - take care of possible status conditions
569  * @pt:		Per-cpu pt context.
570  */
571 static void pt_handle_status(struct pt *pt)
572 {
573 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
574 	int advance = 0;
575 	u64 status;
576 
577 	rdmsrl(MSR_IA32_RTIT_STATUS, status);
578 
579 	if (status & RTIT_STATUS_ERROR) {
580 		pr_err_ratelimited("ToPA ERROR encountered, trying to recover\n");
581 		pt_topa_dump(buf);
582 		status &= ~RTIT_STATUS_ERROR;
583 	}
584 
585 	if (status & RTIT_STATUS_STOPPED) {
586 		status &= ~RTIT_STATUS_STOPPED;
587 
588 		/*
589 		 * On systems that only do single-entry ToPA, hitting STOP
590 		 * means we are already losing data; need to let the decoder
591 		 * know.
592 		 */
593 		if (!pt_cap_get(PT_CAP_topa_multiple_entries) ||
594 		    buf->output_off == sizes(TOPA_ENTRY(buf->cur, buf->cur_idx)->size)) {
595 			local_inc(&buf->lost);
596 			advance++;
597 		}
598 	}
599 
600 	/*
601 	 * Also on single-entry ToPA implementations, interrupt will come
602 	 * before the output reaches its output region's boundary.
603 	 */
604 	if (!pt_cap_get(PT_CAP_topa_multiple_entries) && !buf->snapshot &&
605 	    pt_buffer_region_size(buf) - buf->output_off <= TOPA_PMI_MARGIN) {
606 		void *head = pt_buffer_region(buf);
607 
608 		/* everything within this margin needs to be zeroed out */
609 		memset(head + buf->output_off, 0,
610 		       pt_buffer_region_size(buf) -
611 		       buf->output_off);
612 		advance++;
613 	}
614 
615 	if (advance)
616 		pt_buffer_advance(buf);
617 
618 	wrmsrl(MSR_IA32_RTIT_STATUS, status);
619 }
620 
621 /**
622  * pt_read_offset() - translate registers into buffer pointers
623  * @buf:	PT buffer.
624  *
625  * Set buffer's output pointers from MSR values.
626  */
627 static void pt_read_offset(struct pt_buffer *buf)
628 {
629 	u64 offset, base_topa;
630 
631 	rdmsrl(MSR_IA32_RTIT_OUTPUT_BASE, base_topa);
632 	buf->cur = phys_to_virt(base_topa);
633 
634 	rdmsrl(MSR_IA32_RTIT_OUTPUT_MASK, offset);
635 	/* offset within current output region */
636 	buf->output_off = offset >> 32;
637 	/* index of current output region within this table */
638 	buf->cur_idx = (offset & 0xffffff80) >> 7;
639 }
640 
641 /**
642  * pt_topa_next_entry() - obtain index of the first page in the next ToPA entry
643  * @buf:	PT buffer.
644  * @pg:		Page offset in the buffer.
645  *
646  * When advancing to the next output region (ToPA entry), given a page offset
647  * into the buffer, we need to find the offset of the first page in the next
648  * region.
649  */
650 static unsigned int pt_topa_next_entry(struct pt_buffer *buf, unsigned int pg)
651 {
652 	struct topa_entry *te = buf->topa_index[pg];
653 
654 	/* one region */
655 	if (buf->first == buf->last && buf->first->last == 1)
656 		return pg;
657 
658 	do {
659 		pg++;
660 		pg &= buf->nr_pages - 1;
661 	} while (buf->topa_index[pg] == te);
662 
663 	return pg;
664 }
665 
666 /**
667  * pt_buffer_reset_markers() - place interrupt and stop bits in the buffer
668  * @buf:	PT buffer.
669  * @handle:	Current output handle.
670  *
671  * Place INT and STOP marks to prevent overwriting old data that the consumer
672  * hasn't yet collected and waking up the consumer after a certain fraction of
673  * the buffer has filled up. Only needed and sensible for non-snapshot counters.
674  *
675  * This obviously relies on buf::head to figure out buffer markers, so it has
676  * to be called after pt_buffer_reset_offsets() and before the hardware tracing
677  * is enabled.
678  */
679 static int pt_buffer_reset_markers(struct pt_buffer *buf,
680 				   struct perf_output_handle *handle)
681 
682 {
683 	unsigned long head = local64_read(&buf->head);
684 	unsigned long idx, npages, wakeup;
685 
686 	/* can't stop in the middle of an output region */
687 	if (buf->output_off + handle->size + 1 <
688 	    sizes(TOPA_ENTRY(buf->cur, buf->cur_idx)->size))
689 		return -EINVAL;
690 
691 
692 	/* single entry ToPA is handled by marking all regions STOP=1 INT=1 */
693 	if (!pt_cap_get(PT_CAP_topa_multiple_entries))
694 		return 0;
695 
696 	/* clear STOP and INT from current entry */
697 	buf->topa_index[buf->stop_pos]->stop = 0;
698 	buf->topa_index[buf->intr_pos]->intr = 0;
699 
700 	/* how many pages till the STOP marker */
701 	npages = handle->size >> PAGE_SHIFT;
702 
703 	/* if it's on a page boundary, fill up one more page */
704 	if (!offset_in_page(head + handle->size + 1))
705 		npages++;
706 
707 	idx = (head >> PAGE_SHIFT) + npages;
708 	idx &= buf->nr_pages - 1;
709 	buf->stop_pos = idx;
710 
711 	wakeup = handle->wakeup >> PAGE_SHIFT;
712 
713 	/* in the worst case, wake up the consumer one page before hard stop */
714 	idx = (head >> PAGE_SHIFT) + npages - 1;
715 	if (idx > wakeup)
716 		idx = wakeup;
717 
718 	idx &= buf->nr_pages - 1;
719 	buf->intr_pos = idx;
720 
721 	buf->topa_index[buf->stop_pos]->stop = 1;
722 	buf->topa_index[buf->intr_pos]->intr = 1;
723 
724 	return 0;
725 }
726 
727 /**
728  * pt_buffer_setup_topa_index() - build topa_index[] table of regions
729  * @buf:	PT buffer.
730  *
731  * topa_index[] references output regions indexed by offset into the
732  * buffer for purposes of quick reverse lookup.
733  */
734 static void pt_buffer_setup_topa_index(struct pt_buffer *buf)
735 {
736 	struct topa *cur = buf->first, *prev = buf->last;
737 	struct topa_entry *te_cur = TOPA_ENTRY(cur, 0),
738 		*te_prev = TOPA_ENTRY(prev, prev->last - 1);
739 	int pg = 0, idx = 0;
740 
741 	while (pg < buf->nr_pages) {
742 		int tidx;
743 
744 		/* pages within one topa entry */
745 		for (tidx = 0; tidx < 1 << te_cur->size; tidx++, pg++)
746 			buf->topa_index[pg] = te_prev;
747 
748 		te_prev = te_cur;
749 
750 		if (idx == cur->last - 1) {
751 			/* advance to next topa table */
752 			idx = 0;
753 			cur = list_entry(cur->list.next, struct topa, list);
754 		} else {
755 			idx++;
756 		}
757 		te_cur = TOPA_ENTRY(cur, idx);
758 	}
759 
760 }
761 
762 /**
763  * pt_buffer_reset_offsets() - adjust buffer's write pointers from aux_head
764  * @buf:	PT buffer.
765  * @head:	Write pointer (aux_head) from AUX buffer.
766  *
767  * Find the ToPA table and entry corresponding to given @head and set buffer's
768  * "current" pointers accordingly. This is done after we have obtained the
769  * current aux_head position from a successful call to perf_aux_output_begin()
770  * to make sure the hardware is writing to the right place.
771  *
772  * This function modifies buf::{cur,cur_idx,output_off} that will be programmed
773  * into PT msrs when the tracing is enabled and buf::head and buf::data_size,
774  * which are used to determine INT and STOP markers' locations by a subsequent
775  * call to pt_buffer_reset_markers().
776  */
777 static void pt_buffer_reset_offsets(struct pt_buffer *buf, unsigned long head)
778 {
779 	int pg;
780 
781 	if (buf->snapshot)
782 		head &= (buf->nr_pages << PAGE_SHIFT) - 1;
783 
784 	pg = (head >> PAGE_SHIFT) & (buf->nr_pages - 1);
785 	pg = pt_topa_next_entry(buf, pg);
786 
787 	buf->cur = (struct topa *)((unsigned long)buf->topa_index[pg] & PAGE_MASK);
788 	buf->cur_idx = ((unsigned long)buf->topa_index[pg] -
789 			(unsigned long)buf->cur) / sizeof(struct topa_entry);
790 	buf->output_off = head & (sizes(buf->cur->table[buf->cur_idx].size) - 1);
791 
792 	local64_set(&buf->head, head);
793 	local_set(&buf->data_size, 0);
794 }
795 
796 /**
797  * pt_buffer_fini_topa() - deallocate ToPA structure of a buffer
798  * @buf:	PT buffer.
799  */
800 static void pt_buffer_fini_topa(struct pt_buffer *buf)
801 {
802 	struct topa *topa, *iter;
803 
804 	list_for_each_entry_safe(topa, iter, &buf->tables, list) {
805 		/*
806 		 * right now, this is in free_aux() path only, so
807 		 * no need to unlink this table from the list
808 		 */
809 		topa_free(topa);
810 	}
811 }
812 
813 /**
814  * pt_buffer_init_topa() - initialize ToPA table for pt buffer
815  * @buf:	PT buffer.
816  * @size:	Total size of all regions within this ToPA.
817  * @gfp:	Allocation flags.
818  */
819 static int pt_buffer_init_topa(struct pt_buffer *buf, unsigned long nr_pages,
820 			       gfp_t gfp)
821 {
822 	struct topa *topa;
823 	int err;
824 
825 	topa = topa_alloc(buf->cpu, gfp);
826 	if (!topa)
827 		return -ENOMEM;
828 
829 	topa_insert_table(buf, topa);
830 
831 	while (buf->nr_pages < nr_pages) {
832 		err = topa_insert_pages(buf, gfp);
833 		if (err) {
834 			pt_buffer_fini_topa(buf);
835 			return -ENOMEM;
836 		}
837 	}
838 
839 	pt_buffer_setup_topa_index(buf);
840 
841 	/* link last table to the first one, unless we're double buffering */
842 	if (pt_cap_get(PT_CAP_topa_multiple_entries)) {
843 		TOPA_ENTRY(buf->last, -1)->base = buf->first->phys >> TOPA_SHIFT;
844 		TOPA_ENTRY(buf->last, -1)->end = 1;
845 	}
846 
847 	pt_topa_dump(buf);
848 	return 0;
849 }
850 
851 /**
852  * pt_buffer_setup_aux() - set up topa tables for a PT buffer
853  * @cpu:	Cpu on which to allocate, -1 means current.
854  * @pages:	Array of pointers to buffer pages passed from perf core.
855  * @nr_pages:	Number of pages in the buffer.
856  * @snapshot:	If this is a snapshot/overwrite counter.
857  *
858  * This is a pmu::setup_aux callback that sets up ToPA tables and all the
859  * bookkeeping for an AUX buffer.
860  *
861  * Return:	Our private PT buffer structure.
862  */
863 static void *
864 pt_buffer_setup_aux(int cpu, void **pages, int nr_pages, bool snapshot)
865 {
866 	struct pt_buffer *buf;
867 	int node, ret;
868 
869 	if (!nr_pages)
870 		return NULL;
871 
872 	if (cpu == -1)
873 		cpu = raw_smp_processor_id();
874 	node = cpu_to_node(cpu);
875 
876 	buf = kzalloc_node(offsetof(struct pt_buffer, topa_index[nr_pages]),
877 			   GFP_KERNEL, node);
878 	if (!buf)
879 		return NULL;
880 
881 	buf->cpu = cpu;
882 	buf->snapshot = snapshot;
883 	buf->data_pages = pages;
884 
885 	INIT_LIST_HEAD(&buf->tables);
886 
887 	ret = pt_buffer_init_topa(buf, nr_pages, GFP_KERNEL);
888 	if (ret) {
889 		kfree(buf);
890 		return NULL;
891 	}
892 
893 	return buf;
894 }
895 
896 /**
897  * pt_buffer_free_aux() - perf AUX deallocation path callback
898  * @data:	PT buffer.
899  */
900 static void pt_buffer_free_aux(void *data)
901 {
902 	struct pt_buffer *buf = data;
903 
904 	pt_buffer_fini_topa(buf);
905 	kfree(buf);
906 }
907 
908 /**
909  * pt_buffer_is_full() - check if the buffer is full
910  * @buf:	PT buffer.
911  * @pt:		Per-cpu pt handle.
912  *
913  * If the user hasn't read data from the output region that aux_head
914  * points to, the buffer is considered full: the user needs to read at
915  * least this region and update aux_tail to point past it.
916  */
917 static bool pt_buffer_is_full(struct pt_buffer *buf, struct pt *pt)
918 {
919 	if (buf->snapshot)
920 		return false;
921 
922 	if (local_read(&buf->data_size) >= pt->handle.size)
923 		return true;
924 
925 	return false;
926 }
927 
928 /**
929  * intel_pt_interrupt() - PT PMI handler
930  */
931 void intel_pt_interrupt(void)
932 {
933 	struct pt *pt = this_cpu_ptr(&pt_ctx);
934 	struct pt_buffer *buf;
935 	struct perf_event *event = pt->handle.event;
936 
937 	/*
938 	 * There may be a dangling PT bit in the interrupt status register
939 	 * after PT has been disabled by pt_event_stop(). Make sure we don't
940 	 * do anything (particularly, re-enable) for this event here.
941 	 */
942 	if (!ACCESS_ONCE(pt->handle_nmi))
943 		return;
944 
945 	pt_config_start(false);
946 
947 	if (!event)
948 		return;
949 
950 	buf = perf_get_aux(&pt->handle);
951 	if (!buf)
952 		return;
953 
954 	pt_read_offset(buf);
955 
956 	pt_handle_status(pt);
957 
958 	pt_update_head(pt);
959 
960 	perf_aux_output_end(&pt->handle, local_xchg(&buf->data_size, 0),
961 			    local_xchg(&buf->lost, 0));
962 
963 	if (!event->hw.state) {
964 		int ret;
965 
966 		buf = perf_aux_output_begin(&pt->handle, event);
967 		if (!buf) {
968 			event->hw.state = PERF_HES_STOPPED;
969 			return;
970 		}
971 
972 		pt_buffer_reset_offsets(buf, pt->handle.head);
973 		/* snapshot counters don't use PMI, so it's safe */
974 		ret = pt_buffer_reset_markers(buf, &pt->handle);
975 		if (ret) {
976 			perf_aux_output_end(&pt->handle, 0, true);
977 			return;
978 		}
979 
980 		pt_config_buffer(buf->cur->table, buf->cur_idx,
981 				 buf->output_off);
982 		pt_config(event);
983 	}
984 }
985 
986 /*
987  * PMU callbacks
988  */
989 
990 static void pt_event_start(struct perf_event *event, int mode)
991 {
992 	struct pt *pt = this_cpu_ptr(&pt_ctx);
993 	struct pt_buffer *buf = perf_get_aux(&pt->handle);
994 
995 	if (!buf || pt_buffer_is_full(buf, pt)) {
996 		event->hw.state = PERF_HES_STOPPED;
997 		return;
998 	}
999 
1000 	ACCESS_ONCE(pt->handle_nmi) = 1;
1001 	event->hw.state = 0;
1002 
1003 	pt_config_buffer(buf->cur->table, buf->cur_idx,
1004 			 buf->output_off);
1005 	pt_config(event);
1006 }
1007 
1008 static void pt_event_stop(struct perf_event *event, int mode)
1009 {
1010 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1011 
1012 	/*
1013 	 * Protect against the PMI racing with disabling wrmsr,
1014 	 * see comment in intel_pt_interrupt().
1015 	 */
1016 	ACCESS_ONCE(pt->handle_nmi) = 0;
1017 	pt_config_start(false);
1018 
1019 	if (event->hw.state == PERF_HES_STOPPED)
1020 		return;
1021 
1022 	event->hw.state = PERF_HES_STOPPED;
1023 
1024 	if (mode & PERF_EF_UPDATE) {
1025 		struct pt_buffer *buf = perf_get_aux(&pt->handle);
1026 
1027 		if (!buf)
1028 			return;
1029 
1030 		if (WARN_ON_ONCE(pt->handle.event != event))
1031 			return;
1032 
1033 		pt_read_offset(buf);
1034 
1035 		pt_handle_status(pt);
1036 
1037 		pt_update_head(pt);
1038 	}
1039 }
1040 
1041 static void pt_event_del(struct perf_event *event, int mode)
1042 {
1043 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1044 	struct pt_buffer *buf;
1045 
1046 	pt_event_stop(event, PERF_EF_UPDATE);
1047 
1048 	buf = perf_get_aux(&pt->handle);
1049 
1050 	if (buf) {
1051 		if (buf->snapshot)
1052 			pt->handle.head =
1053 				local_xchg(&buf->data_size,
1054 					   buf->nr_pages << PAGE_SHIFT);
1055 		perf_aux_output_end(&pt->handle, local_xchg(&buf->data_size, 0),
1056 				    local_xchg(&buf->lost, 0));
1057 	}
1058 }
1059 
1060 static int pt_event_add(struct perf_event *event, int mode)
1061 {
1062 	struct pt_buffer *buf;
1063 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1064 	struct hw_perf_event *hwc = &event->hw;
1065 	int ret = -EBUSY;
1066 
1067 	if (pt->handle.event)
1068 		goto fail;
1069 
1070 	buf = perf_aux_output_begin(&pt->handle, event);
1071 	ret = -EINVAL;
1072 	if (!buf)
1073 		goto fail_stop;
1074 
1075 	pt_buffer_reset_offsets(buf, pt->handle.head);
1076 	if (!buf->snapshot) {
1077 		ret = pt_buffer_reset_markers(buf, &pt->handle);
1078 		if (ret)
1079 			goto fail_end_stop;
1080 	}
1081 
1082 	if (mode & PERF_EF_START) {
1083 		pt_event_start(event, 0);
1084 		ret = -EBUSY;
1085 		if (hwc->state == PERF_HES_STOPPED)
1086 			goto fail_end_stop;
1087 	} else {
1088 		hwc->state = PERF_HES_STOPPED;
1089 	}
1090 
1091 	return 0;
1092 
1093 fail_end_stop:
1094 	perf_aux_output_end(&pt->handle, 0, true);
1095 fail_stop:
1096 	hwc->state = PERF_HES_STOPPED;
1097 fail:
1098 	return ret;
1099 }
1100 
1101 static void pt_event_read(struct perf_event *event)
1102 {
1103 }
1104 
1105 static void pt_event_destroy(struct perf_event *event)
1106 {
1107 	x86_del_exclusive(x86_lbr_exclusive_pt);
1108 }
1109 
1110 static int pt_event_init(struct perf_event *event)
1111 {
1112 	if (event->attr.type != pt_pmu.pmu.type)
1113 		return -ENOENT;
1114 
1115 	if (!pt_event_valid(event))
1116 		return -EINVAL;
1117 
1118 	if (x86_add_exclusive(x86_lbr_exclusive_pt))
1119 		return -EBUSY;
1120 
1121 	event->destroy = pt_event_destroy;
1122 
1123 	return 0;
1124 }
1125 
1126 void cpu_emergency_stop_pt(void)
1127 {
1128 	struct pt *pt = this_cpu_ptr(&pt_ctx);
1129 
1130 	if (pt->handle.event)
1131 		pt_event_stop(pt->handle.event, PERF_EF_UPDATE);
1132 }
1133 
1134 static __init int pt_init(void)
1135 {
1136 	int ret, cpu, prior_warn = 0;
1137 
1138 	BUILD_BUG_ON(sizeof(struct topa) > PAGE_SIZE);
1139 
1140 	if (!test_cpu_cap(&boot_cpu_data, X86_FEATURE_INTEL_PT))
1141 		return -ENODEV;
1142 
1143 	get_online_cpus();
1144 	for_each_online_cpu(cpu) {
1145 		u64 ctl;
1146 
1147 		ret = rdmsrl_safe_on_cpu(cpu, MSR_IA32_RTIT_CTL, &ctl);
1148 		if (!ret && (ctl & RTIT_CTL_TRACEEN))
1149 			prior_warn++;
1150 	}
1151 	put_online_cpus();
1152 
1153 	if (prior_warn) {
1154 		x86_add_exclusive(x86_lbr_exclusive_pt);
1155 		pr_warn("PT is enabled at boot time, doing nothing\n");
1156 
1157 		return -EBUSY;
1158 	}
1159 
1160 	ret = pt_pmu_hw_init();
1161 	if (ret)
1162 		return ret;
1163 
1164 	if (!pt_cap_get(PT_CAP_topa_output)) {
1165 		pr_warn("ToPA output is not supported on this CPU\n");
1166 		return -ENODEV;
1167 	}
1168 
1169 	if (!pt_cap_get(PT_CAP_topa_multiple_entries))
1170 		pt_pmu.pmu.capabilities =
1171 			PERF_PMU_CAP_AUX_NO_SG | PERF_PMU_CAP_AUX_SW_DOUBLEBUF;
1172 
1173 	pt_pmu.pmu.capabilities	|= PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE;
1174 	pt_pmu.pmu.attr_groups	= pt_attr_groups;
1175 	pt_pmu.pmu.task_ctx_nr	= perf_sw_context;
1176 	pt_pmu.pmu.event_init	= pt_event_init;
1177 	pt_pmu.pmu.add		= pt_event_add;
1178 	pt_pmu.pmu.del		= pt_event_del;
1179 	pt_pmu.pmu.start	= pt_event_start;
1180 	pt_pmu.pmu.stop		= pt_event_stop;
1181 	pt_pmu.pmu.read		= pt_event_read;
1182 	pt_pmu.pmu.setup_aux	= pt_buffer_setup_aux;
1183 	pt_pmu.pmu.free_aux	= pt_buffer_free_aux;
1184 	ret = perf_pmu_register(&pt_pmu.pmu, "intel_pt", -1);
1185 
1186 	return ret;
1187 }
1188 arch_initcall(pt_init);
1189