xref: /openbmc/linux/tools/perf/builtin-mem.c (revision a5d46d9a)
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 <linux/err.h>
24 
25 #define MEM_OPERATION_LOAD	0x1
26 #define MEM_OPERATION_STORE	0x2
27 
28 struct perf_mem {
29 	struct perf_tool	tool;
30 	char const		*input_name;
31 	bool			hide_unresolved;
32 	bool			dump_raw;
33 	bool			force;
34 	bool			phys_addr;
35 	bool			data_page_size;
36 	int			operation;
37 	const char		*cpu_list;
38 	DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
39 };
40 
41 static int parse_record_events(const struct option *opt,
42 			       const char *str, int unset __maybe_unused)
43 {
44 	struct perf_mem *mem = *(struct perf_mem **)opt->value;
45 
46 	if (!strcmp(str, "list")) {
47 		perf_mem_events__list();
48 		exit(0);
49 	}
50 	if (perf_mem_events__parse(str))
51 		exit(-1);
52 
53 	mem->operation = 0;
54 	return 0;
55 }
56 
57 static const char * const __usage[] = {
58 	"perf mem record [<options>] [<command>]",
59 	"perf mem record [<options>] -- <command> [<options>]",
60 	NULL
61 };
62 
63 static const char * const *record_mem_usage = __usage;
64 
65 static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
66 {
67 	int rec_argc, i = 0, j, tmp_nr = 0;
68 	int start, end;
69 	const char **rec_argv;
70 	char **rec_tmp;
71 	int ret;
72 	bool all_user = false, all_kernel = false;
73 	struct perf_mem_event *e;
74 	struct option options[] = {
75 	OPT_CALLBACK('e', "event", &mem, "event",
76 		     "event selector. use 'perf mem record -e list' to list available events",
77 		     parse_record_events),
78 	OPT_UINTEGER(0, "ldlat", &perf_mem_events__loads_ldlat, "mem-loads latency"),
79 	OPT_INCR('v', "verbose", &verbose,
80 		 "be more verbose (show counter open errors, etc)"),
81 	OPT_BOOLEAN('U', "all-user", &all_user, "collect only user level data"),
82 	OPT_BOOLEAN('K', "all-kernel", &all_kernel, "collect only kernel level data"),
83 	OPT_END()
84 	};
85 
86 	if (perf_mem_events__init()) {
87 		pr_err("failed: memory events not supported\n");
88 		return -1;
89 	}
90 
91 	argc = parse_options(argc, argv, options, record_mem_usage,
92 			     PARSE_OPT_KEEP_UNKNOWN);
93 
94 	if (!perf_pmu__has_hybrid())
95 		rec_argc = argc + 9; /* max number of arguments */
96 	else
97 		rec_argc = argc + 9 * perf_pmu__hybrid_pmu_num();
98 
99 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
100 	if (!rec_argv)
101 		return -1;
102 
103 	/*
104 	 * Save the allocated event name strings.
105 	 */
106 	rec_tmp = calloc(rec_argc + 1, sizeof(char *));
107 	if (!rec_tmp) {
108 		free(rec_argv);
109 		return -1;
110 	}
111 
112 	rec_argv[i++] = "record";
113 
114 	e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD_STORE);
115 
116 	/*
117 	 * The load and store operations are required, use the event
118 	 * PERF_MEM_EVENTS__LOAD_STORE if it is supported.
119 	 */
120 	if (e->tag &&
121 	    (mem->operation & MEM_OPERATION_LOAD) &&
122 	    (mem->operation & MEM_OPERATION_STORE)) {
123 		e->record = true;
124 	} else {
125 		if (mem->operation & MEM_OPERATION_LOAD) {
126 			e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD);
127 			e->record = true;
128 		}
129 
130 		if (mem->operation & MEM_OPERATION_STORE) {
131 			e = perf_mem_events__ptr(PERF_MEM_EVENTS__STORE);
132 			e->record = true;
133 		}
134 	}
135 
136 	e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD);
137 	if (e->record)
138 		rec_argv[i++] = "-W";
139 
140 	rec_argv[i++] = "-d";
141 
142 	if (mem->phys_addr)
143 		rec_argv[i++] = "--phys-data";
144 
145 	if (mem->data_page_size)
146 		rec_argv[i++] = "--data-page-size";
147 
148 	start = i;
149 	ret = perf_mem_events__record_args(rec_argv, &i, rec_tmp, &tmp_nr);
150 	if (ret)
151 		goto out;
152 	end = i;
153 
154 	if (all_user)
155 		rec_argv[i++] = "--all-user";
156 
157 	if (all_kernel)
158 		rec_argv[i++] = "--all-kernel";
159 
160 	for (j = 0; j < argc; j++, i++)
161 		rec_argv[i] = argv[j];
162 
163 	if (verbose > 0) {
164 		pr_debug("calling: record ");
165 
166 		for (j = start; j < end; j++)
167 			pr_debug("%s ", rec_argv[j]);
168 
169 		pr_debug("\n");
170 	}
171 
172 	ret = cmd_record(i, rec_argv);
173 out:
174 	for (i = 0; i < tmp_nr; i++)
175 		free(rec_tmp[i]);
176 
177 	free(rec_tmp);
178 	free(rec_argv);
179 	return ret;
180 }
181 
182 static int
183 dump_raw_samples(struct perf_tool *tool,
184 		 union perf_event *event,
185 		 struct perf_sample *sample,
186 		 struct machine *machine)
187 {
188 	struct perf_mem *mem = container_of(tool, struct perf_mem, tool);
189 	struct addr_location al;
190 	const char *fmt, *field_sep;
191 	char str[PAGE_SIZE_NAME_LEN];
192 
193 	if (machine__resolve(machine, &al, sample) < 0) {
194 		fprintf(stderr, "problem processing %d event, skipping it.\n",
195 				event->header.type);
196 		return -1;
197 	}
198 
199 	if (al.filtered || (mem->hide_unresolved && al.sym == NULL))
200 		goto out_put;
201 
202 	if (al.map != NULL)
203 		al.map->dso->hit = 1;
204 
205 	field_sep = symbol_conf.field_sep;
206 	if (field_sep) {
207 		fmt = "%d%s%d%s0x%"PRIx64"%s0x%"PRIx64"%s";
208 	} else {
209 		fmt = "%5d%s%5d%s0x%016"PRIx64"%s0x016%"PRIx64"%s";
210 		symbol_conf.field_sep = " ";
211 	}
212 	printf(fmt,
213 		sample->pid,
214 		symbol_conf.field_sep,
215 		sample->tid,
216 		symbol_conf.field_sep,
217 		sample->ip,
218 		symbol_conf.field_sep,
219 		sample->addr,
220 		symbol_conf.field_sep);
221 
222 	if (mem->phys_addr) {
223 		printf("0x%016"PRIx64"%s",
224 			sample->phys_addr,
225 			symbol_conf.field_sep);
226 	}
227 
228 	if (mem->data_page_size) {
229 		printf("%s%s",
230 			get_page_size_name(sample->data_page_size, str),
231 			symbol_conf.field_sep);
232 	}
233 
234 	if (field_sep)
235 		fmt = "%"PRIu64"%s0x%"PRIx64"%s%s:%s\n";
236 	else
237 		fmt = "%5"PRIu64"%s0x%06"PRIx64"%s%s:%s\n";
238 
239 	printf(fmt,
240 		sample->weight,
241 		symbol_conf.field_sep,
242 		sample->data_src,
243 		symbol_conf.field_sep,
244 		al.map ? (al.map->dso ? al.map->dso->long_name : "???") : "???",
245 		al.sym ? al.sym->name : "???");
246 out_put:
247 	addr_location__put(&al);
248 	return 0;
249 }
250 
251 static int process_sample_event(struct perf_tool *tool,
252 				union perf_event *event,
253 				struct perf_sample *sample,
254 				struct evsel *evsel __maybe_unused,
255 				struct machine *machine)
256 {
257 	return dump_raw_samples(tool, event, sample, machine);
258 }
259 
260 static int report_raw_events(struct perf_mem *mem)
261 {
262 	struct itrace_synth_opts itrace_synth_opts = {
263 		.set = true,
264 		.mem = true,	/* Only enable memory event */
265 		.default_no_sample = true,
266 	};
267 
268 	struct perf_data data = {
269 		.path  = input_name,
270 		.mode  = PERF_DATA_MODE_READ,
271 		.force = mem->force,
272 	};
273 	int ret;
274 	struct perf_session *session = perf_session__new(&data, false,
275 							 &mem->tool);
276 
277 	if (IS_ERR(session))
278 		return PTR_ERR(session);
279 
280 	session->itrace_synth_opts = &itrace_synth_opts;
281 
282 	if (mem->cpu_list) {
283 		ret = perf_session__cpu_bitmap(session, mem->cpu_list,
284 					       mem->cpu_bitmap);
285 		if (ret < 0)
286 			goto out_delete;
287 	}
288 
289 	ret = symbol__init(&session->header.env);
290 	if (ret < 0)
291 		goto out_delete;
292 
293 	printf("# PID, TID, IP, ADDR, ");
294 
295 	if (mem->phys_addr)
296 		printf("PHYS ADDR, ");
297 
298 	if (mem->data_page_size)
299 		printf("DATA PAGE SIZE, ");
300 
301 	printf("LOCAL WEIGHT, DSRC, SYMBOL\n");
302 
303 	ret = perf_session__process_events(session);
304 
305 out_delete:
306 	perf_session__delete(session);
307 	return ret;
308 }
309 static char *get_sort_order(struct perf_mem *mem)
310 {
311 	bool has_extra_options = (mem->phys_addr | mem->data_page_size) ? true : false;
312 	char sort[128];
313 
314 	/*
315 	 * there is no weight (cost) associated with stores, so don't print
316 	 * the column
317 	 */
318 	if (!(mem->operation & MEM_OPERATION_LOAD)) {
319 		strcpy(sort, "--sort=mem,sym,dso,symbol_daddr,"
320 			     "dso_daddr,tlb,locked");
321 	} else if (has_extra_options) {
322 		strcpy(sort, "--sort=local_weight,mem,sym,dso,symbol_daddr,"
323 			     "dso_daddr,snoop,tlb,locked,blocked");
324 	} else
325 		return NULL;
326 
327 	if (mem->phys_addr)
328 		strcat(sort, ",phys_daddr");
329 
330 	if (mem->data_page_size)
331 		strcat(sort, ",data_page_size");
332 
333 	return strdup(sort);
334 }
335 
336 static int report_events(int argc, const char **argv, struct perf_mem *mem)
337 {
338 	const char **rep_argv;
339 	int ret, i = 0, j, rep_argc;
340 	char *new_sort_order;
341 
342 	if (mem->dump_raw)
343 		return report_raw_events(mem);
344 
345 	rep_argc = argc + 3;
346 	rep_argv = calloc(rep_argc + 1, sizeof(char *));
347 	if (!rep_argv)
348 		return -1;
349 
350 	rep_argv[i++] = "report";
351 	rep_argv[i++] = "--mem-mode";
352 	rep_argv[i++] = "-n"; /* display number of samples */
353 
354 	new_sort_order = get_sort_order(mem);
355 	if (new_sort_order)
356 		rep_argv[i++] = new_sort_order;
357 
358 	for (j = 1; j < argc; j++, i++)
359 		rep_argv[i] = argv[j];
360 
361 	ret = cmd_report(i, rep_argv);
362 	free(rep_argv);
363 	return ret;
364 }
365 
366 struct mem_mode {
367 	const char *name;
368 	int mode;
369 };
370 
371 #define MEM_OPT(n, m) \
372 	{ .name = n, .mode = (m) }
373 
374 #define MEM_END { .name = NULL }
375 
376 static const struct mem_mode mem_modes[]={
377 	MEM_OPT("load", MEM_OPERATION_LOAD),
378 	MEM_OPT("store", MEM_OPERATION_STORE),
379 	MEM_END
380 };
381 
382 static int
383 parse_mem_ops(const struct option *opt, const char *str, int unset)
384 {
385 	int *mode = (int *)opt->value;
386 	const struct mem_mode *m;
387 	char *s, *os = NULL, *p;
388 	int ret = -1;
389 
390 	if (unset)
391 		return 0;
392 
393 	/* str may be NULL in case no arg is passed to -t */
394 	if (str) {
395 		/* because str is read-only */
396 		s = os = strdup(str);
397 		if (!s)
398 			return -1;
399 
400 		/* reset mode */
401 		*mode = 0;
402 
403 		for (;;) {
404 			p = strchr(s, ',');
405 			if (p)
406 				*p = '\0';
407 
408 			for (m = mem_modes; m->name; m++) {
409 				if (!strcasecmp(s, m->name))
410 					break;
411 			}
412 			if (!m->name) {
413 				fprintf(stderr, "unknown sampling op %s,"
414 					    " check man page\n", s);
415 				goto error;
416 			}
417 
418 			*mode |= m->mode;
419 
420 			if (!p)
421 				break;
422 
423 			s = p + 1;
424 		}
425 	}
426 	ret = 0;
427 
428 	if (*mode == 0)
429 		*mode = MEM_OPERATION_LOAD;
430 error:
431 	free(os);
432 	return ret;
433 }
434 
435 int cmd_mem(int argc, const char **argv)
436 {
437 	struct stat st;
438 	struct perf_mem mem = {
439 		.tool = {
440 			.sample		= process_sample_event,
441 			.mmap		= perf_event__process_mmap,
442 			.mmap2		= perf_event__process_mmap2,
443 			.comm		= perf_event__process_comm,
444 			.lost		= perf_event__process_lost,
445 			.fork		= perf_event__process_fork,
446 			.attr		= perf_event__process_attr,
447 			.build_id	= perf_event__process_build_id,
448 			.namespaces	= perf_event__process_namespaces,
449 			.auxtrace_info  = perf_event__process_auxtrace_info,
450 			.auxtrace       = perf_event__process_auxtrace,
451 			.auxtrace_error = perf_event__process_auxtrace_error,
452 			.ordered_events	= true,
453 		},
454 		.input_name		 = "perf.data",
455 		/*
456 		 * default to both load an store sampling
457 		 */
458 		.operation		 = MEM_OPERATION_LOAD | MEM_OPERATION_STORE,
459 	};
460 	const struct option mem_options[] = {
461 	OPT_CALLBACK('t', "type", &mem.operation,
462 		   "type", "memory operations(load,store) Default load,store",
463 		    parse_mem_ops),
464 	OPT_BOOLEAN('D', "dump-raw-samples", &mem.dump_raw,
465 		    "dump raw samples in ASCII"),
466 	OPT_BOOLEAN('U', "hide-unresolved", &mem.hide_unresolved,
467 		    "Only display entries resolved to a symbol"),
468 	OPT_STRING('i', "input", &input_name, "file",
469 		   "input file name"),
470 	OPT_STRING('C', "cpu", &mem.cpu_list, "cpu",
471 		   "list of cpus to profile"),
472 	OPT_STRING_NOEMPTY('x', "field-separator", &symbol_conf.field_sep,
473 		   "separator",
474 		   "separator for columns, no spaces will be added"
475 		   " between columns '.' is reserved."),
476 	OPT_BOOLEAN('f', "force", &mem.force, "don't complain, do it"),
477 	OPT_BOOLEAN('p', "phys-data", &mem.phys_addr, "Record/Report sample physical addresses"),
478 	OPT_BOOLEAN(0, "data-page-size", &mem.data_page_size, "Record/Report sample data address page size"),
479 	OPT_END()
480 	};
481 	const char *const mem_subcommands[] = { "record", "report", NULL };
482 	const char *mem_usage[] = {
483 		NULL,
484 		NULL
485 	};
486 
487 	argc = parse_options_subcommand(argc, argv, mem_options, mem_subcommands,
488 					mem_usage, PARSE_OPT_KEEP_UNKNOWN);
489 
490 	if (!argc || !(strncmp(argv[0], "rec", 3) || mem.operation))
491 		usage_with_options(mem_usage, mem_options);
492 
493 	if (!mem.input_name || !strlen(mem.input_name)) {
494 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
495 			mem.input_name = "-";
496 		else
497 			mem.input_name = "perf.data";
498 	}
499 
500 	if (!strncmp(argv[0], "rec", 3))
501 		return __cmd_record(argc, argv, &mem);
502 	else if (!strncmp(argv[0], "rep", 3))
503 		return report_events(argc, argv, &mem);
504 	else
505 		usage_with_options(mem_usage, mem_options);
506 
507 	return 0;
508 }
509