xref: /linux/tools/perf/builtin-mem.c (revision cc8c418b4fc09ed58ddd27b8e90ec797e9ca1e67)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <inttypes.h>
3 #include <sys/types.h>
4 #include <sys/stat.h>
5 #include <unistd.h>
6 #include "builtin.h"
7 #include "perf.h"
8 
9 #include <subcmd/parse-options.h>
10 #include "util/auxtrace.h"
11 #include "util/trace-event.h"
12 #include "util/tool.h"
13 #include "util/session.h"
14 #include "util/data.h"
15 #include "util/map_symbol.h"
16 #include "util/mem-events.h"
17 #include "util/debug.h"
18 #include "util/dso.h"
19 #include "util/map.h"
20 #include "util/symbol.h"
21 #include "util/pmu.h"
22 #include "util/pmu-hybrid.h"
23 #include "util/string2.h"
24 #include <linux/err.h>
25 
26 #define MEM_OPERATION_LOAD	0x1
27 #define MEM_OPERATION_STORE	0x2
28 
29 struct perf_mem {
30 	struct perf_tool	tool;
31 	char const		*input_name;
32 	bool			hide_unresolved;
33 	bool			dump_raw;
34 	bool			force;
35 	bool			phys_addr;
36 	bool			data_page_size;
37 	int			operation;
38 	const char		*cpu_list;
39 	DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
40 };
41 
42 static int parse_record_events(const struct option *opt,
43 			       const char *str, int unset __maybe_unused)
44 {
45 	struct perf_mem *mem = *(struct perf_mem **)opt->value;
46 
47 	if (!strcmp(str, "list")) {
48 		perf_mem_events__list();
49 		exit(0);
50 	}
51 	if (perf_mem_events__parse(str))
52 		exit(-1);
53 
54 	mem->operation = 0;
55 	return 0;
56 }
57 
58 static const char * const __usage[] = {
59 	"perf mem record [<options>] [<command>]",
60 	"perf mem record [<options>] -- <command> [<options>]",
61 	NULL
62 };
63 
64 static const char * const *record_mem_usage = __usage;
65 
66 static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
67 {
68 	int rec_argc, i = 0, j, tmp_nr = 0;
69 	int start, end;
70 	const char **rec_argv;
71 	char **rec_tmp;
72 	int ret;
73 	bool all_user = false, all_kernel = false;
74 	struct perf_mem_event *e;
75 	struct option options[] = {
76 	OPT_CALLBACK('e', "event", &mem, "event",
77 		     "event selector. use 'perf mem record -e list' to list available events",
78 		     parse_record_events),
79 	OPT_UINTEGER(0, "ldlat", &perf_mem_events__loads_ldlat, "mem-loads latency"),
80 	OPT_INCR('v', "verbose", &verbose,
81 		 "be more verbose (show counter open errors, etc)"),
82 	OPT_BOOLEAN('U', "all-user", &all_user, "collect only user level data"),
83 	OPT_BOOLEAN('K', "all-kernel", &all_kernel, "collect only kernel level data"),
84 	OPT_END()
85 	};
86 
87 	if (perf_mem_events__init()) {
88 		pr_err("failed: memory events not supported\n");
89 		return -1;
90 	}
91 
92 	argc = parse_options(argc, argv, options, record_mem_usage,
93 			     PARSE_OPT_KEEP_UNKNOWN);
94 
95 	if (!perf_pmu__has_hybrid())
96 		rec_argc = argc + 9; /* max number of arguments */
97 	else
98 		rec_argc = argc + 9 * perf_pmu__hybrid_pmu_num();
99 
100 	if (mem->cpu_list)
101 		rec_argc += 2;
102 
103 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
104 	if (!rec_argv)
105 		return -1;
106 
107 	/*
108 	 * Save the allocated event name strings.
109 	 */
110 	rec_tmp = calloc(rec_argc + 1, sizeof(char *));
111 	if (!rec_tmp) {
112 		free(rec_argv);
113 		return -1;
114 	}
115 
116 	rec_argv[i++] = "record";
117 
118 	e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD_STORE);
119 
120 	/*
121 	 * The load and store operations are required, use the event
122 	 * PERF_MEM_EVENTS__LOAD_STORE if it is supported.
123 	 */
124 	if (e->tag &&
125 	    (mem->operation & MEM_OPERATION_LOAD) &&
126 	    (mem->operation & MEM_OPERATION_STORE)) {
127 		e->record = true;
128 		rec_argv[i++] = "-W";
129 	} else {
130 		if (mem->operation & MEM_OPERATION_LOAD) {
131 			e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD);
132 			e->record = true;
133 		}
134 
135 		if (mem->operation & MEM_OPERATION_STORE) {
136 			e = perf_mem_events__ptr(PERF_MEM_EVENTS__STORE);
137 			e->record = true;
138 		}
139 	}
140 
141 	e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD);
142 	if (e->record)
143 		rec_argv[i++] = "-W";
144 
145 	rec_argv[i++] = "-d";
146 
147 	if (mem->phys_addr)
148 		rec_argv[i++] = "--phys-data";
149 
150 	if (mem->data_page_size)
151 		rec_argv[i++] = "--data-page-size";
152 
153 	start = i;
154 	ret = perf_mem_events__record_args(rec_argv, &i, rec_tmp, &tmp_nr);
155 	if (ret)
156 		goto out;
157 	end = i;
158 
159 	if (all_user)
160 		rec_argv[i++] = "--all-user";
161 
162 	if (all_kernel)
163 		rec_argv[i++] = "--all-kernel";
164 
165 	if (mem->cpu_list) {
166 		rec_argv[i++] = "-C";
167 		rec_argv[i++] = mem->cpu_list;
168 	}
169 
170 	for (j = 0; j < argc; j++, i++)
171 		rec_argv[i] = argv[j];
172 
173 	if (verbose > 0) {
174 		pr_debug("calling: record ");
175 
176 		for (j = start; j < end; j++)
177 			pr_debug("%s ", rec_argv[j]);
178 
179 		pr_debug("\n");
180 	}
181 
182 	ret = cmd_record(i, rec_argv);
183 out:
184 	for (i = 0; i < tmp_nr; i++)
185 		free(rec_tmp[i]);
186 
187 	free(rec_tmp);
188 	free(rec_argv);
189 	return ret;
190 }
191 
192 static int
193 dump_raw_samples(struct perf_tool *tool,
194 		 union perf_event *event,
195 		 struct perf_sample *sample,
196 		 struct machine *machine)
197 {
198 	struct perf_mem *mem = container_of(tool, struct perf_mem, tool);
199 	struct addr_location al;
200 	const char *fmt, *field_sep;
201 	char str[PAGE_SIZE_NAME_LEN];
202 
203 	if (machine__resolve(machine, &al, sample) < 0) {
204 		fprintf(stderr, "problem processing %d event, skipping it.\n",
205 				event->header.type);
206 		return -1;
207 	}
208 
209 	if (al.filtered || (mem->hide_unresolved && al.sym == NULL))
210 		goto out_put;
211 
212 	if (al.map != NULL)
213 		al.map->dso->hit = 1;
214 
215 	field_sep = symbol_conf.field_sep;
216 	if (field_sep) {
217 		fmt = "%d%s%d%s0x%"PRIx64"%s0x%"PRIx64"%s";
218 	} else {
219 		fmt = "%5d%s%5d%s0x%016"PRIx64"%s0x016%"PRIx64"%s";
220 		symbol_conf.field_sep = " ";
221 	}
222 	printf(fmt,
223 		sample->pid,
224 		symbol_conf.field_sep,
225 		sample->tid,
226 		symbol_conf.field_sep,
227 		sample->ip,
228 		symbol_conf.field_sep,
229 		sample->addr,
230 		symbol_conf.field_sep);
231 
232 	if (mem->phys_addr) {
233 		printf("0x%016"PRIx64"%s",
234 			sample->phys_addr,
235 			symbol_conf.field_sep);
236 	}
237 
238 	if (mem->data_page_size) {
239 		printf("%s%s",
240 			get_page_size_name(sample->data_page_size, str),
241 			symbol_conf.field_sep);
242 	}
243 
244 	if (field_sep)
245 		fmt = "%"PRIu64"%s0x%"PRIx64"%s%s:%s\n";
246 	else
247 		fmt = "%5"PRIu64"%s0x%06"PRIx64"%s%s:%s\n";
248 
249 	printf(fmt,
250 		sample->weight,
251 		symbol_conf.field_sep,
252 		sample->data_src,
253 		symbol_conf.field_sep,
254 		al.map ? (al.map->dso ? al.map->dso->long_name : "???") : "???",
255 		al.sym ? al.sym->name : "???");
256 out_put:
257 	addr_location__put(&al);
258 	return 0;
259 }
260 
261 static int process_sample_event(struct perf_tool *tool,
262 				union perf_event *event,
263 				struct perf_sample *sample,
264 				struct evsel *evsel __maybe_unused,
265 				struct machine *machine)
266 {
267 	return dump_raw_samples(tool, event, sample, machine);
268 }
269 
270 static int report_raw_events(struct perf_mem *mem)
271 {
272 	struct itrace_synth_opts itrace_synth_opts = {
273 		.set = true,
274 		.mem = true,	/* Only enable memory event */
275 		.default_no_sample = true,
276 	};
277 
278 	struct perf_data data = {
279 		.path  = input_name,
280 		.mode  = PERF_DATA_MODE_READ,
281 		.force = mem->force,
282 	};
283 	int ret;
284 	struct perf_session *session = perf_session__new(&data, &mem->tool);
285 
286 	if (IS_ERR(session))
287 		return PTR_ERR(session);
288 
289 	session->itrace_synth_opts = &itrace_synth_opts;
290 
291 	if (mem->cpu_list) {
292 		ret = perf_session__cpu_bitmap(session, mem->cpu_list,
293 					       mem->cpu_bitmap);
294 		if (ret < 0)
295 			goto out_delete;
296 	}
297 
298 	ret = symbol__init(&session->header.env);
299 	if (ret < 0)
300 		goto out_delete;
301 
302 	printf("# PID, TID, IP, ADDR, ");
303 
304 	if (mem->phys_addr)
305 		printf("PHYS ADDR, ");
306 
307 	if (mem->data_page_size)
308 		printf("DATA PAGE SIZE, ");
309 
310 	printf("LOCAL WEIGHT, DSRC, SYMBOL\n");
311 
312 	ret = perf_session__process_events(session);
313 
314 out_delete:
315 	perf_session__delete(session);
316 	return ret;
317 }
318 static char *get_sort_order(struct perf_mem *mem)
319 {
320 	bool has_extra_options = (mem->phys_addr | mem->data_page_size) ? true : false;
321 	char sort[128];
322 
323 	/*
324 	 * there is no weight (cost) associated with stores, so don't print
325 	 * the column
326 	 */
327 	if (!(mem->operation & MEM_OPERATION_LOAD)) {
328 		strcpy(sort, "--sort=mem,sym,dso,symbol_daddr,"
329 			     "dso_daddr,tlb,locked");
330 	} else if (has_extra_options) {
331 		strcpy(sort, "--sort=local_weight,mem,sym,dso,symbol_daddr,"
332 			     "dso_daddr,snoop,tlb,locked,blocked");
333 	} else
334 		return NULL;
335 
336 	if (mem->phys_addr)
337 		strcat(sort, ",phys_daddr");
338 
339 	if (mem->data_page_size)
340 		strcat(sort, ",data_page_size");
341 
342 	return strdup(sort);
343 }
344 
345 static int report_events(int argc, const char **argv, struct perf_mem *mem)
346 {
347 	const char **rep_argv;
348 	int ret, i = 0, j, rep_argc;
349 	char *new_sort_order;
350 
351 	if (mem->dump_raw)
352 		return report_raw_events(mem);
353 
354 	rep_argc = argc + 3;
355 	rep_argv = calloc(rep_argc + 1, sizeof(char *));
356 	if (!rep_argv)
357 		return -1;
358 
359 	rep_argv[i++] = "report";
360 	rep_argv[i++] = "--mem-mode";
361 	rep_argv[i++] = "-n"; /* display number of samples */
362 
363 	new_sort_order = get_sort_order(mem);
364 	if (new_sort_order)
365 		rep_argv[i++] = new_sort_order;
366 
367 	for (j = 1; j < argc; j++, i++)
368 		rep_argv[i] = argv[j];
369 
370 	ret = cmd_report(i, rep_argv);
371 	free(rep_argv);
372 	return ret;
373 }
374 
375 struct mem_mode {
376 	const char *name;
377 	int mode;
378 };
379 
380 #define MEM_OPT(n, m) \
381 	{ .name = n, .mode = (m) }
382 
383 #define MEM_END { .name = NULL }
384 
385 static const struct mem_mode mem_modes[]={
386 	MEM_OPT("load", MEM_OPERATION_LOAD),
387 	MEM_OPT("store", MEM_OPERATION_STORE),
388 	MEM_END
389 };
390 
391 static int
392 parse_mem_ops(const struct option *opt, const char *str, int unset)
393 {
394 	int *mode = (int *)opt->value;
395 	const struct mem_mode *m;
396 	char *s, *os = NULL, *p;
397 	int ret = -1;
398 
399 	if (unset)
400 		return 0;
401 
402 	/* str may be NULL in case no arg is passed to -t */
403 	if (str) {
404 		/* because str is read-only */
405 		s = os = strdup(str);
406 		if (!s)
407 			return -1;
408 
409 		/* reset mode */
410 		*mode = 0;
411 
412 		for (;;) {
413 			p = strchr(s, ',');
414 			if (p)
415 				*p = '\0';
416 
417 			for (m = mem_modes; m->name; m++) {
418 				if (!strcasecmp(s, m->name))
419 					break;
420 			}
421 			if (!m->name) {
422 				fprintf(stderr, "unknown sampling op %s,"
423 					    " check man page\n", s);
424 				goto error;
425 			}
426 
427 			*mode |= m->mode;
428 
429 			if (!p)
430 				break;
431 
432 			s = p + 1;
433 		}
434 	}
435 	ret = 0;
436 
437 	if (*mode == 0)
438 		*mode = MEM_OPERATION_LOAD;
439 error:
440 	free(os);
441 	return ret;
442 }
443 
444 int cmd_mem(int argc, const char **argv)
445 {
446 	struct stat st;
447 	struct perf_mem mem = {
448 		.tool = {
449 			.sample		= process_sample_event,
450 			.mmap		= perf_event__process_mmap,
451 			.mmap2		= perf_event__process_mmap2,
452 			.comm		= perf_event__process_comm,
453 			.lost		= perf_event__process_lost,
454 			.fork		= perf_event__process_fork,
455 			.attr		= perf_event__process_attr,
456 			.build_id	= perf_event__process_build_id,
457 			.namespaces	= perf_event__process_namespaces,
458 			.auxtrace_info  = perf_event__process_auxtrace_info,
459 			.auxtrace       = perf_event__process_auxtrace,
460 			.auxtrace_error = perf_event__process_auxtrace_error,
461 			.ordered_events	= true,
462 		},
463 		.input_name		 = "perf.data",
464 		/*
465 		 * default to both load an store sampling
466 		 */
467 		.operation		 = MEM_OPERATION_LOAD | MEM_OPERATION_STORE,
468 	};
469 	const struct option mem_options[] = {
470 	OPT_CALLBACK('t', "type", &mem.operation,
471 		   "type", "memory operations(load,store) Default load,store",
472 		    parse_mem_ops),
473 	OPT_BOOLEAN('D', "dump-raw-samples", &mem.dump_raw,
474 		    "dump raw samples in ASCII"),
475 	OPT_BOOLEAN('U', "hide-unresolved", &mem.hide_unresolved,
476 		    "Only display entries resolved to a symbol"),
477 	OPT_STRING('i', "input", &input_name, "file",
478 		   "input file name"),
479 	OPT_STRING('C', "cpu", &mem.cpu_list, "cpu",
480 		   "list of cpus to profile"),
481 	OPT_STRING_NOEMPTY('x', "field-separator", &symbol_conf.field_sep,
482 		   "separator",
483 		   "separator for columns, no spaces will be added"
484 		   " between columns '.' is reserved."),
485 	OPT_BOOLEAN('f', "force", &mem.force, "don't complain, do it"),
486 	OPT_BOOLEAN('p', "phys-data", &mem.phys_addr, "Record/Report sample physical addresses"),
487 	OPT_BOOLEAN(0, "data-page-size", &mem.data_page_size, "Record/Report sample data address page size"),
488 	OPT_END()
489 	};
490 	const char *const mem_subcommands[] = { "record", "report", NULL };
491 	const char *mem_usage[] = {
492 		NULL,
493 		NULL
494 	};
495 
496 	argc = parse_options_subcommand(argc, argv, mem_options, mem_subcommands,
497 					mem_usage, PARSE_OPT_KEEP_UNKNOWN);
498 
499 	if (!argc || !(strncmp(argv[0], "rec", 3) || mem.operation))
500 		usage_with_options(mem_usage, mem_options);
501 
502 	if (!mem.input_name || !strlen(mem.input_name)) {
503 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
504 			mem.input_name = "-";
505 		else
506 			mem.input_name = "perf.data";
507 	}
508 
509 	if (strlen(argv[0]) > 2 && strstarts("record", argv[0]))
510 		return __cmd_record(argc, argv, &mem);
511 	else if (strlen(argv[0]) > 2 && strstarts("report", argv[0]))
512 		return report_events(argc, argv, &mem);
513 	else
514 		usage_with_options(mem_usage, mem_options);
515 
516 	return 0;
517 }
518