xref: /openbmc/linux/tools/perf/arch/arm64/util/arm-spe.c (revision da6a5afd)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Arm Statistical Profiling Extensions (SPE) support
4  * Copyright (c) 2017-2018, Arm Ltd.
5  */
6 
7 #include <linux/kernel.h>
8 #include <linux/types.h>
9 #include <linux/bitops.h>
10 #include <linux/log2.h>
11 #include <linux/zalloc.h>
12 #include <time.h>
13 
14 #include "../../../util/cpumap.h"
15 #include "../../../util/event.h"
16 #include "../../../util/evsel.h"
17 #include "../../../util/evsel_config.h"
18 #include "../../../util/evlist.h"
19 #include "../../../util/session.h"
20 #include <internal/lib.h> // page_size
21 #include "../../../util/pmu.h"
22 #include "../../../util/debug.h"
23 #include "../../../util/auxtrace.h"
24 #include "../../../util/record.h"
25 #include "../../../util/arm-spe.h"
26 #include <tools/libc_compat.h> // reallocarray
27 
28 #define KiB(x) ((x) * 1024)
29 #define MiB(x) ((x) * 1024 * 1024)
30 
31 struct arm_spe_recording {
32 	struct auxtrace_record		itr;
33 	struct perf_pmu			*arm_spe_pmu;
34 	struct evlist		*evlist;
35 	int			wrapped_cnt;
36 	bool			*wrapped;
37 };
38 
39 static size_t
arm_spe_info_priv_size(struct auxtrace_record * itr __maybe_unused,struct evlist * evlist __maybe_unused)40 arm_spe_info_priv_size(struct auxtrace_record *itr __maybe_unused,
41 		       struct evlist *evlist __maybe_unused)
42 {
43 	return ARM_SPE_AUXTRACE_PRIV_SIZE;
44 }
45 
arm_spe_info_fill(struct auxtrace_record * itr,struct perf_session * session,struct perf_record_auxtrace_info * auxtrace_info,size_t priv_size)46 static int arm_spe_info_fill(struct auxtrace_record *itr,
47 			     struct perf_session *session,
48 			     struct perf_record_auxtrace_info *auxtrace_info,
49 			     size_t priv_size)
50 {
51 	struct arm_spe_recording *sper =
52 			container_of(itr, struct arm_spe_recording, itr);
53 	struct perf_pmu *arm_spe_pmu = sper->arm_spe_pmu;
54 
55 	if (priv_size != ARM_SPE_AUXTRACE_PRIV_SIZE)
56 		return -EINVAL;
57 
58 	if (!session->evlist->core.nr_mmaps)
59 		return -EINVAL;
60 
61 	auxtrace_info->type = PERF_AUXTRACE_ARM_SPE;
62 	auxtrace_info->priv[ARM_SPE_PMU_TYPE] = arm_spe_pmu->type;
63 
64 	return 0;
65 }
66 
67 static void
arm_spe_snapshot_resolve_auxtrace_defaults(struct record_opts * opts,bool privileged)68 arm_spe_snapshot_resolve_auxtrace_defaults(struct record_opts *opts,
69 					   bool privileged)
70 {
71 	/*
72 	 * The default snapshot size is the auxtrace mmap size. If neither auxtrace mmap size nor
73 	 * snapshot size is specified, then the default is 4MiB for privileged users, 128KiB for
74 	 * unprivileged users.
75 	 *
76 	 * The default auxtrace mmap size is 4MiB/page_size for privileged users, 128KiB for
77 	 * unprivileged users. If an unprivileged user does not specify mmap pages, the mmap pages
78 	 * will be reduced from the default 512KiB/page_size to 256KiB/page_size, otherwise the
79 	 * user is likely to get an error as they exceed their mlock limmit.
80 	 */
81 
82 	/*
83 	 * No size were given to '-S' or '-m,', so go with the default
84 	 */
85 	if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {
86 		if (privileged) {
87 			opts->auxtrace_mmap_pages = MiB(4) / page_size;
88 		} else {
89 			opts->auxtrace_mmap_pages = KiB(128) / page_size;
90 			if (opts->mmap_pages == UINT_MAX)
91 				opts->mmap_pages = KiB(256) / page_size;
92 		}
93 	} else if (!opts->auxtrace_mmap_pages && !privileged && opts->mmap_pages == UINT_MAX) {
94 		opts->mmap_pages = KiB(256) / page_size;
95 	}
96 
97 	/*
98 	 * '-m,xyz' was specified but no snapshot size, so make the snapshot size as big as the
99 	 * auxtrace mmap area.
100 	 */
101 	if (!opts->auxtrace_snapshot_size)
102 		opts->auxtrace_snapshot_size = opts->auxtrace_mmap_pages * (size_t)page_size;
103 
104 	/*
105 	 * '-Sxyz' was specified but no auxtrace mmap area, so make the auxtrace mmap area big
106 	 * enough to fit the requested snapshot size.
107 	 */
108 	if (!opts->auxtrace_mmap_pages) {
109 		size_t sz = opts->auxtrace_snapshot_size;
110 
111 		sz = round_up(sz, page_size) / page_size;
112 		opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
113 	}
114 }
115 
arm_spe_recording_options(struct auxtrace_record * itr,struct evlist * evlist,struct record_opts * opts)116 static int arm_spe_recording_options(struct auxtrace_record *itr,
117 				     struct evlist *evlist,
118 				     struct record_opts *opts)
119 {
120 	struct arm_spe_recording *sper =
121 			container_of(itr, struct arm_spe_recording, itr);
122 	struct perf_pmu *arm_spe_pmu = sper->arm_spe_pmu;
123 	struct evsel *evsel, *arm_spe_evsel = NULL;
124 	struct perf_cpu_map *cpus = evlist->core.user_requested_cpus;
125 	bool privileged = perf_event_paranoid_check(-1);
126 	struct evsel *tracking_evsel;
127 	int err;
128 	u64 bit;
129 
130 	sper->evlist = evlist;
131 
132 	evlist__for_each_entry(evlist, evsel) {
133 		if (evsel->core.attr.type == arm_spe_pmu->type) {
134 			if (arm_spe_evsel) {
135 				pr_err("There may be only one " ARM_SPE_PMU_NAME "x event\n");
136 				return -EINVAL;
137 			}
138 			evsel->core.attr.freq = 0;
139 			evsel->core.attr.sample_period = arm_spe_pmu->default_config->sample_period;
140 			evsel->needs_auxtrace_mmap = true;
141 			arm_spe_evsel = evsel;
142 			opts->full_auxtrace = true;
143 		}
144 	}
145 
146 	if (!opts->full_auxtrace)
147 		return 0;
148 
149 	/*
150 	 * we are in snapshot mode.
151 	 */
152 	if (opts->auxtrace_snapshot_mode) {
153 		/*
154 		 * Command arguments '-Sxyz' and/or '-m,xyz' are missing, so fill those in with
155 		 * default values.
156 		 */
157 		if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages)
158 			arm_spe_snapshot_resolve_auxtrace_defaults(opts, privileged);
159 
160 		/*
161 		 * Snapshot size can't be bigger than the auxtrace area.
162 		 */
163 		if (opts->auxtrace_snapshot_size > opts->auxtrace_mmap_pages * (size_t)page_size) {
164 			pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
165 			       opts->auxtrace_snapshot_size,
166 			       opts->auxtrace_mmap_pages * (size_t)page_size);
167 			return -EINVAL;
168 		}
169 
170 		/*
171 		 * Something went wrong somewhere - this shouldn't happen.
172 		 */
173 		if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {
174 			pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
175 			return -EINVAL;
176 		}
177 	}
178 
179 	/* We are in full trace mode but '-m,xyz' wasn't specified */
180 	if (!opts->auxtrace_mmap_pages) {
181 		if (privileged) {
182 			opts->auxtrace_mmap_pages = MiB(4) / page_size;
183 		} else {
184 			opts->auxtrace_mmap_pages = KiB(128) / page_size;
185 			if (opts->mmap_pages == UINT_MAX)
186 				opts->mmap_pages = KiB(256) / page_size;
187 		}
188 	}
189 
190 	/* Validate auxtrace_mmap_pages */
191 	if (opts->auxtrace_mmap_pages) {
192 		size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
193 		size_t min_sz = KiB(8);
194 
195 		if (sz < min_sz || !is_power_of_2(sz)) {
196 			pr_err("Invalid mmap size for ARM SPE: must be at least %zuKiB and a power of 2\n",
197 			       min_sz / 1024);
198 			return -EINVAL;
199 		}
200 	}
201 
202 	if (opts->auxtrace_snapshot_mode)
203 		pr_debug2("%sx snapshot size: %zu\n", ARM_SPE_PMU_NAME,
204 			  opts->auxtrace_snapshot_size);
205 
206 	/*
207 	 * To obtain the auxtrace buffer file descriptor, the auxtrace event
208 	 * must come first.
209 	 */
210 	evlist__to_front(evlist, arm_spe_evsel);
211 
212 	/*
213 	 * In the case of per-cpu mmaps, sample CPU for AUX event;
214 	 * also enable the timestamp tracing for samples correlation.
215 	 */
216 	if (!perf_cpu_map__empty(cpus)) {
217 		evsel__set_sample_bit(arm_spe_evsel, CPU);
218 		evsel__set_config_if_unset(arm_spe_pmu, arm_spe_evsel,
219 					   "ts_enable", 1);
220 	}
221 
222 	/*
223 	 * Set this only so that perf report knows that SPE generates memory info. It has no effect
224 	 * on the opening of the event or the SPE data produced.
225 	 */
226 	evsel__set_sample_bit(arm_spe_evsel, DATA_SRC);
227 
228 	/*
229 	 * The PHYS_ADDR flag does not affect the driver behaviour, it is used to
230 	 * inform that the resulting output's SPE samples contain physical addresses
231 	 * where applicable.
232 	 */
233 	bit = perf_pmu__format_bits(arm_spe_pmu, "pa_enable");
234 	if (arm_spe_evsel->core.attr.config & bit)
235 		evsel__set_sample_bit(arm_spe_evsel, PHYS_ADDR);
236 
237 	/* Add dummy event to keep tracking */
238 	err = parse_event(evlist, "dummy:u");
239 	if (err)
240 		return err;
241 
242 	tracking_evsel = evlist__last(evlist);
243 	evlist__set_tracking_event(evlist, tracking_evsel);
244 
245 	tracking_evsel->core.attr.freq = 0;
246 	tracking_evsel->core.attr.sample_period = 1;
247 
248 	/* In per-cpu case, always need the time of mmap events etc */
249 	if (!perf_cpu_map__empty(cpus)) {
250 		evsel__set_sample_bit(tracking_evsel, TIME);
251 		evsel__set_sample_bit(tracking_evsel, CPU);
252 
253 		/* also track task context switch */
254 		if (!record_opts__no_switch_events(opts))
255 			tracking_evsel->core.attr.context_switch = 1;
256 	}
257 
258 	return 0;
259 }
260 
arm_spe_parse_snapshot_options(struct auxtrace_record * itr __maybe_unused,struct record_opts * opts,const char * str)261 static int arm_spe_parse_snapshot_options(struct auxtrace_record *itr __maybe_unused,
262 					 struct record_opts *opts,
263 					 const char *str)
264 {
265 	unsigned long long snapshot_size = 0;
266 	char *endptr;
267 
268 	if (str) {
269 		snapshot_size = strtoull(str, &endptr, 0);
270 		if (*endptr || snapshot_size > SIZE_MAX)
271 			return -1;
272 	}
273 
274 	opts->auxtrace_snapshot_mode = true;
275 	opts->auxtrace_snapshot_size = snapshot_size;
276 
277 	return 0;
278 }
279 
arm_spe_snapshot_start(struct auxtrace_record * itr)280 static int arm_spe_snapshot_start(struct auxtrace_record *itr)
281 {
282 	struct arm_spe_recording *ptr =
283 			container_of(itr, struct arm_spe_recording, itr);
284 	struct evsel *evsel;
285 
286 	evlist__for_each_entry(ptr->evlist, evsel) {
287 		if (evsel->core.attr.type == ptr->arm_spe_pmu->type)
288 			return evsel__disable(evsel);
289 	}
290 	return -EINVAL;
291 }
292 
arm_spe_snapshot_finish(struct auxtrace_record * itr)293 static int arm_spe_snapshot_finish(struct auxtrace_record *itr)
294 {
295 	struct arm_spe_recording *ptr =
296 			container_of(itr, struct arm_spe_recording, itr);
297 	struct evsel *evsel;
298 
299 	evlist__for_each_entry(ptr->evlist, evsel) {
300 		if (evsel->core.attr.type == ptr->arm_spe_pmu->type)
301 			return evsel__enable(evsel);
302 	}
303 	return -EINVAL;
304 }
305 
arm_spe_alloc_wrapped_array(struct arm_spe_recording * ptr,int idx)306 static int arm_spe_alloc_wrapped_array(struct arm_spe_recording *ptr, int idx)
307 {
308 	bool *wrapped;
309 	int cnt = ptr->wrapped_cnt, new_cnt, i;
310 
311 	/*
312 	 * No need to allocate, so return early.
313 	 */
314 	if (idx < cnt)
315 		return 0;
316 
317 	/*
318 	 * Make ptr->wrapped as big as idx.
319 	 */
320 	new_cnt = idx + 1;
321 
322 	/*
323 	 * Free'ed in arm_spe_recording_free().
324 	 */
325 	wrapped = reallocarray(ptr->wrapped, new_cnt, sizeof(bool));
326 	if (!wrapped)
327 		return -ENOMEM;
328 
329 	/*
330 	 * init new allocated values.
331 	 */
332 	for (i = cnt; i < new_cnt; i++)
333 		wrapped[i] = false;
334 
335 	ptr->wrapped_cnt = new_cnt;
336 	ptr->wrapped = wrapped;
337 
338 	return 0;
339 }
340 
arm_spe_buffer_has_wrapped(unsigned char * buffer,size_t buffer_size,u64 head)341 static bool arm_spe_buffer_has_wrapped(unsigned char *buffer,
342 				      size_t buffer_size, u64 head)
343 {
344 	u64 i, watermark;
345 	u64 *buf = (u64 *)buffer;
346 	size_t buf_size = buffer_size;
347 
348 	/*
349 	 * Defensively handle the case where head might be continually increasing - if its value is
350 	 * equal or greater than the size of the ring buffer, then we can safely determine it has
351 	 * wrapped around. Otherwise, continue to detect if head might have wrapped.
352 	 */
353 	if (head >= buffer_size)
354 		return true;
355 
356 	/*
357 	 * We want to look the very last 512 byte (chosen arbitrarily) in the ring buffer.
358 	 */
359 	watermark = buf_size - 512;
360 
361 	/*
362 	 * The value of head is somewhere within the size of the ring buffer. This can be that there
363 	 * hasn't been enough data to fill the ring buffer yet or the trace time was so long that
364 	 * head has numerically wrapped around.  To find we need to check if we have data at the
365 	 * very end of the ring buffer.  We can reliably do this because mmap'ed pages are zeroed
366 	 * out and there is a fresh mapping with every new session.
367 	 */
368 
369 	/*
370 	 * head is less than 512 byte from the end of the ring buffer.
371 	 */
372 	if (head > watermark)
373 		watermark = head;
374 
375 	/*
376 	 * Speed things up by using 64 bit transactions (see "u64 *buf" above)
377 	 */
378 	watermark /= sizeof(u64);
379 	buf_size /= sizeof(u64);
380 
381 	/*
382 	 * If we find trace data at the end of the ring buffer, head has been there and has
383 	 * numerically wrapped around at least once.
384 	 */
385 	for (i = watermark; i < buf_size; i++)
386 		if (buf[i])
387 			return true;
388 
389 	return false;
390 }
391 
arm_spe_find_snapshot(struct auxtrace_record * itr,int idx,struct auxtrace_mmap * mm,unsigned char * data,u64 * head,u64 * old)392 static int arm_spe_find_snapshot(struct auxtrace_record *itr, int idx,
393 				  struct auxtrace_mmap *mm, unsigned char *data,
394 				  u64 *head, u64 *old)
395 {
396 	int err;
397 	bool wrapped;
398 	struct arm_spe_recording *ptr =
399 			container_of(itr, struct arm_spe_recording, itr);
400 
401 	/*
402 	 * Allocate memory to keep track of wrapping if this is the first
403 	 * time we deal with this *mm.
404 	 */
405 	if (idx >= ptr->wrapped_cnt) {
406 		err = arm_spe_alloc_wrapped_array(ptr, idx);
407 		if (err)
408 			return err;
409 	}
410 
411 	/*
412 	 * Check to see if *head has wrapped around.  If it hasn't only the
413 	 * amount of data between *head and *old is snapshot'ed to avoid
414 	 * bloating the perf.data file with zeros.  But as soon as *head has
415 	 * wrapped around the entire size of the AUX ring buffer it taken.
416 	 */
417 	wrapped = ptr->wrapped[idx];
418 	if (!wrapped && arm_spe_buffer_has_wrapped(data, mm->len, *head)) {
419 		wrapped = true;
420 		ptr->wrapped[idx] = true;
421 	}
422 
423 	pr_debug3("%s: mmap index %d old head %zu new head %zu size %zu\n",
424 		  __func__, idx, (size_t)*old, (size_t)*head, mm->len);
425 
426 	/*
427 	 * No wrap has occurred, we can just use *head and *old.
428 	 */
429 	if (!wrapped)
430 		return 0;
431 
432 	/*
433 	 * *head has wrapped around - adjust *head and *old to pickup the
434 	 * entire content of the AUX buffer.
435 	 */
436 	if (*head >= mm->len) {
437 		*old = *head - mm->len;
438 	} else {
439 		*head += mm->len;
440 		*old = *head - mm->len;
441 	}
442 
443 	return 0;
444 }
445 
arm_spe_reference(struct auxtrace_record * itr __maybe_unused)446 static u64 arm_spe_reference(struct auxtrace_record *itr __maybe_unused)
447 {
448 	struct timespec ts;
449 
450 	clock_gettime(CLOCK_MONOTONIC_RAW, &ts);
451 
452 	return ts.tv_sec ^ ts.tv_nsec;
453 }
454 
arm_spe_recording_free(struct auxtrace_record * itr)455 static void arm_spe_recording_free(struct auxtrace_record *itr)
456 {
457 	struct arm_spe_recording *sper =
458 			container_of(itr, struct arm_spe_recording, itr);
459 
460 	zfree(&sper->wrapped);
461 	free(sper);
462 }
463 
arm_spe_recording_init(int * err,struct perf_pmu * arm_spe_pmu)464 struct auxtrace_record *arm_spe_recording_init(int *err,
465 					       struct perf_pmu *arm_spe_pmu)
466 {
467 	struct arm_spe_recording *sper;
468 
469 	if (!arm_spe_pmu) {
470 		*err = -ENODEV;
471 		return NULL;
472 	}
473 
474 	sper = zalloc(sizeof(struct arm_spe_recording));
475 	if (!sper) {
476 		*err = -ENOMEM;
477 		return NULL;
478 	}
479 
480 	sper->arm_spe_pmu = arm_spe_pmu;
481 	sper->itr.pmu = arm_spe_pmu;
482 	sper->itr.snapshot_start = arm_spe_snapshot_start;
483 	sper->itr.snapshot_finish = arm_spe_snapshot_finish;
484 	sper->itr.find_snapshot = arm_spe_find_snapshot;
485 	sper->itr.parse_snapshot_options = arm_spe_parse_snapshot_options;
486 	sper->itr.recording_options = arm_spe_recording_options;
487 	sper->itr.info_priv_size = arm_spe_info_priv_size;
488 	sper->itr.info_fill = arm_spe_info_fill;
489 	sper->itr.free = arm_spe_recording_free;
490 	sper->itr.reference = arm_spe_reference;
491 	sper->itr.read_finish = auxtrace_record__read_finish;
492 	sper->itr.alignment = 0;
493 
494 	*err = 0;
495 	return &sper->itr;
496 }
497 
498 struct perf_event_attr
arm_spe_pmu_default_config(struct perf_pmu * arm_spe_pmu)499 *arm_spe_pmu_default_config(struct perf_pmu *arm_spe_pmu)
500 {
501 	struct perf_event_attr *attr;
502 
503 	attr = zalloc(sizeof(struct perf_event_attr));
504 	if (!attr) {
505 		pr_err("arm_spe default config cannot allocate a perf_event_attr\n");
506 		return NULL;
507 	}
508 
509 	/*
510 	 * If kernel driver doesn't advertise a minimum,
511 	 * use max allowable by PMSIDR_EL1.INTERVAL
512 	 */
513 	if (perf_pmu__scan_file(arm_spe_pmu, "caps/min_interval", "%llu",
514 				  &attr->sample_period) != 1) {
515 		pr_debug("arm_spe driver doesn't advertise a min. interval. Using 4096\n");
516 		attr->sample_period = 4096;
517 	}
518 
519 	arm_spe_pmu->selectable = true;
520 	arm_spe_pmu->is_uncore = false;
521 
522 	return attr;
523 }
524