1/*
  2 * builtin-record.c
  3 *
  4 * Builtin record command: Record the profile of a workload
  5 * (or a CPU, or a PID) into the perf.data output file - for
  6 * later analysis via perf report.
  7 */
  8#define _FILE_OFFSET_BITS 64
  9
 10#include "builtin.h"
 11
 12#include "perf.h"
 13
 14#include "util/build-id.h"
 15#include "util/util.h"
 16#include "util/parse-options.h"
 17#include "util/parse-events.h"
 
 18
 
 
 19#include "util/header.h"
 20#include "util/event.h"
 21#include "util/evlist.h"
 22#include "util/evsel.h"
 23#include "util/debug.h"
 
 
 
 24#include "util/session.h"
 
 25#include "util/symbol.h"
 
 26#include "util/cpumap.h"
 27#include "util/thread_map.h"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 28
 
 
 
 
 
 29#include <unistd.h>
 
 
 
 30#include <sched.h>
 
 
 
 
 31#include <sys/mman.h>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 32
 33enum write_mode_t {
 34	WRITE_FORCE,
 35	WRITE_APPEND
 36};
 37
 38static u64			user_interval			= ULLONG_MAX;
 39static u64			default_interval		=      0;
 
 
 
 
 
 
 
 40
 41static unsigned int		page_size;
 42static unsigned int		mmap_pages			= UINT_MAX;
 43static unsigned int		user_freq 			= UINT_MAX;
 44static int			freq				=   1000;
 45static int			output;
 46static int			pipe_output			=      0;
 47static const char		*output_name			= NULL;
 48static bool			group				=  false;
 49static int			realtime_prio			=      0;
 50static bool			nodelay				=  false;
 51static bool			raw_samples			=  false;
 52static bool			sample_id_all_avail		=   true;
 53static bool			system_wide			=  false;
 54static pid_t			target_pid			=     -1;
 55static pid_t			target_tid			=     -1;
 56static pid_t			child_pid			=     -1;
 57static bool			no_inherit			=  false;
 58static enum write_mode_t	write_mode			= WRITE_FORCE;
 59static bool			call_graph			=  false;
 60static bool			inherit_stat			=  false;
 61static bool			no_samples			=  false;
 62static bool			sample_address			=  false;
 63static bool			sample_time			=  false;
 64static bool			no_buildid			=  false;
 65static bool			no_buildid_cache		=  false;
 66static struct perf_evlist	*evsel_list;
 67
 68static long			samples				=      0;
 69static u64			bytes_written			=      0;
 
 
 70
 71static int			file_new			=      1;
 72static off_t			post_processing_offset;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 73
 74static struct perf_session	*session;
 75static const char		*cpu_list;
 76
 77static void advance_output(size_t size)
 
 
 
 
 
 
 
 
 
 78{
 79	bytes_written += size;
 80}
 
 81
 82static void write_output(void *buf, size_t size)
 83{
 84	while (size) {
 85		int ret = write(output, buf, size);
 86
 87		if (ret < 0)
 88			die("failed to write");
 
 
 
 89
 90		size -= ret;
 91		buf += ret;
 
 
 
 
 92
 93		bytes_written += ret;
 94	}
 
 
 95}
 96
 97static int process_synthesized_event(union perf_event *event,
 98				     struct perf_sample *sample __used,
 99				     struct perf_session *self __used)
100{
101	write_output(event, event->header.size);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
102	return 0;
103}
104
105static void mmap_read(struct perf_mmap *md)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
106{
107	unsigned int head = perf_mmap__read_head(md);
108	unsigned int old = md->prev;
109	unsigned char *data = md->base + page_size;
110	unsigned long size;
111	void *buf;
 
 
 
 
 
 
 
 
 
 
 
 
 
112
113	if (old == head)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
114		return;
115
116	samples++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
117
118	size = head - old;
 
119
120	if ((old & md->mask) + size != (head & md->mask)) {
121		buf = &data[old & md->mask];
122		size = md->mask + 1 - (old & md->mask);
123		old += size;
124
125		write_output(buf, size);
 
 
 
 
 
 
126	}
127
128	buf = &data[old & md->mask];
129	size = head - old;
130	old += size;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
131
132	write_output(buf, size);
133
134	md->prev = old;
135	perf_mmap__write_tail(md, old);
 
 
 
 
 
 
 
 
 
136}
137
138static volatile int done = 0;
139static volatile int signr = -1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
140
141static void sig_handler(int sig)
142{
 
 
 
 
 
143	done = 1;
144	signr = sig;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
145}
146
147static void sig_atexit(void)
148{
149	if (child_pid > 0)
150		kill(child_pid, SIGTERM);
 
151
152	if (signr == -1 || signr == SIGUSR1)
 
 
153		return;
154
155	signal(signr, SIG_DFL);
156	kill(getpid(), signr);
157}
158
159static void config_attr(struct perf_evsel *evsel, struct perf_evlist *evlist)
 
 
 
 
 
160{
161	struct perf_event_attr *attr = &evsel->attr;
162	int track = !evsel->idx; /* only the first counter needs these */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
163
164	attr->disabled		= 1;
165	attr->inherit		= !no_inherit;
166	attr->read_format	= PERF_FORMAT_TOTAL_TIME_ENABLED |
167				  PERF_FORMAT_TOTAL_TIME_RUNNING |
168				  PERF_FORMAT_ID;
 
 
 
 
 
169
170	attr->sample_type	|= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 
171
172	if (evlist->nr_entries > 1)
173		attr->sample_type |= PERF_SAMPLE_ID;
 
 
174
175	/*
176	 * We default some events to a 1 default interval. But keep
177	 * it a weak assumption overridable by the user.
178	 */
179	if (!attr->sample_period || (user_freq != UINT_MAX &&
180				     user_interval != ULLONG_MAX)) {
181		if (freq) {
182			attr->sample_type	|= PERF_SAMPLE_PERIOD;
183			attr->freq		= 1;
184			attr->sample_freq	= freq;
185		} else {
186			attr->sample_period = default_interval;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
187		}
188	}
 
 
 
189
190	if (no_samples)
191		attr->sample_freq = 0;
 
 
 
 
 
 
 
 
 
 
192
193	if (inherit_stat)
194		attr->inherit_stat = 1;
 
 
195
196	if (sample_address) {
197		attr->sample_type	|= PERF_SAMPLE_ADDR;
198		attr->mmap_data = track;
199	}
200
201	if (call_graph)
202		attr->sample_type	|= PERF_SAMPLE_CALLCHAIN;
 
203
204	if (system_wide)
205		attr->sample_type	|= PERF_SAMPLE_CPU;
206
207	if (sample_id_all_avail &&
208	    (sample_time || system_wide || !no_inherit || cpu_list))
209		attr->sample_type	|= PERF_SAMPLE_TIME;
210
211	if (raw_samples) {
212		attr->sample_type	|= PERF_SAMPLE_TIME;
213		attr->sample_type	|= PERF_SAMPLE_RAW;
214		attr->sample_type	|= PERF_SAMPLE_CPU;
215	}
216
217	if (nodelay) {
218		attr->watermark = 0;
219		attr->wakeup_events = 1;
 
220	}
221
222	attr->mmap		= track;
223	attr->comm		= track;
 
 
 
 
 
 
 
224
225	if (target_pid == -1 && target_tid == -1 && !system_wide) {
226		attr->disabled = 1;
227		attr->enable_on_exec = 1;
228	}
229}
230
231static bool perf_evlist__equal(struct perf_evlist *evlist,
232			       struct perf_evlist *other)
 
 
 
233{
234	struct perf_evsel *pos, *pair;
 
235
236	if (evlist->nr_entries != other->nr_entries)
237		return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
238
239	pair = list_entry(other->entries.next, struct perf_evsel, node);
240
241	list_for_each_entry(pos, &evlist->entries, node) {
242		if (memcmp(&pos->attr, &pair->attr, sizeof(pos->attr) != 0))
243			return false;
244		pair = list_entry(pair->node.next, struct perf_evsel, node);
 
 
 
 
245	}
246
247	return true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
248}
249
250static void open_counters(struct perf_evlist *evlist)
251{
252	struct perf_evsel *pos;
 
253
254	if (evlist->cpus->map[0] < 0)
255		no_inherit = true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
256
257	list_for_each_entry(pos, &evlist->entries, node) {
258		struct perf_event_attr *attr = &pos->attr;
259		/*
260		 * Check if parse_single_tracepoint_event has already asked for
261		 * PERF_SAMPLE_TIME.
262		 *
263		 * XXX this is kludgy but short term fix for problems introduced by
264		 * eac23d1c that broke 'perf script' by having different sample_types
265		 * when using multiple tracepoint events when we use a perf binary
266		 * that tries to use sample_id_all on an older kernel.
267		 *
268		 * We need to move counter creation to perf_session, support
269		 * different sample_types, etc.
270		 */
271		bool time_needed = attr->sample_type & PERF_SAMPLE_TIME;
 
272
273		config_attr(pos, evlist);
274retry_sample_id:
275		attr->sample_id_all = sample_id_all_avail ? 1 : 0;
276try_again:
277		if (perf_evsel__open(pos, evlist->cpus, evlist->threads, group) < 0) {
278			int err = errno;
279
280			if (err == EPERM || err == EACCES) {
281				ui__warning_paranoid();
282				exit(EXIT_FAILURE);
283			} else if (err ==  ENODEV && cpu_list) {
284				die("No such device - did you specify"
285					" an out-of-range profile CPU?\n");
286			} else if (err == EINVAL && sample_id_all_avail) {
287				/*
288				 * Old kernel, no attr->sample_id_type_all field
289				 */
290				sample_id_all_avail = false;
291				if (!sample_time && !raw_samples && !time_needed)
292					attr->sample_type &= ~PERF_SAMPLE_TIME;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
293
294				goto retry_sample_id;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
295			}
 
 
 
 
 
 
 
 
296
297			/*
298			 * If it's cycles then fall back to hrtimer
299			 * based cpu-clock-tick sw counter, which
300			 * is always available even if no PMU support:
301			 */
302			if (attr->type == PERF_TYPE_HARDWARE
303					&& attr->config == PERF_COUNT_HW_CPU_CYCLES) {
304
305				if (verbose)
306					ui__warning("The cycles event is not supported, "
307						    "trying to fall back to cpu-clock-ticks\n");
308				attr->type = PERF_TYPE_SOFTWARE;
309				attr->config = PERF_COUNT_SW_CPU_CLOCK;
310				goto try_again;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
311			}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
312
313			if (err == ENOENT) {
314				ui__warning("The %s event is not supported.\n",
315					    event_name(pos));
316				exit(EXIT_FAILURE);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
317			}
 
 
 
 
 
 
 
 
 
 
 
 
318
319			printf("\n");
320			error("sys_perf_event_open() syscall returned with %d (%s).  /bin/dmesg may provide additional information.\n",
321			      err, strerror(err));
322
323#if defined(__i386__) || defined(__x86_64__)
324			if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
325				die("No hardware sampling interrupt available."
326				    " No APIC? If so then you can boot the kernel"
327				    " with the \"lapic\" boot parameter to"
328				    " force-enable it.\n");
329#endif
330
331			die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
332		}
333	}
334
335	if (perf_evlist__set_filters(evlist)) {
336		error("failed to set filter with %d (%s)\n", errno,
337			strerror(errno));
338		exit(-1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
339	}
340
341	if (perf_evlist__mmap(evlist, mmap_pages, false) < 0)
342		die("failed to mmap with %d (%s)\n", errno, strerror(errno));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
343
344	if (file_new)
345		session->evlist = evlist;
346	else {
347		if (!perf_evlist__equal(session->evlist, evlist)) {
348			fprintf(stderr, "incompatible append\n");
349			exit(-1);
 
 
 
 
 
 
 
 
 
 
 
 
350		}
351 	}
352
353	perf_session__update_sample_type(session);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
354}
355
356static int process_buildids(void)
 
 
 
 
357{
358	u64 size = lseek(output, 0, SEEK_CUR);
359
360	if (size == 0)
 
 
361		return 0;
362
363	session->fd = output;
364	return __perf_session__process_events(session, post_processing_offset,
365					      size - post_processing_offset,
366					      size, &build_id__mark_dso_hit_ops);
367}
368
369static void atexit_header(void)
370{
371	if (!pipe_output) {
372		session->header.data_size += bytes_written;
373
374		if (!no_buildid)
375			process_buildids();
376		perf_session__write_header(session, evsel_list, output, true);
377		perf_session__delete(session);
378		perf_evlist__delete(evsel_list);
379		symbol__exit();
380	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
381}
382
383static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
384{
385	int err;
386	struct perf_session *psession = data;
387
388	if (machine__is_host(machine))
389		return;
390
391	/*
392	 *As for guest kernel when processing subcommand record&report,
393	 *we arrange module mmap prior to guest kernel mmap and trigger
394	 *a preload dso because default guest module symbols are loaded
395	 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
396	 *method is used to avoid symbol missing when the first addr is
397	 *in module instead of in guest kernel.
398	 */
399	err = perf_event__synthesize_modules(process_synthesized_event,
400					     psession, machine);
401	if (err < 0)
402		pr_err("Couldn't record guest kernel [%d]'s reference"
403		       " relocation symbol.\n", machine->pid);
404
405	/*
406	 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
407	 * have no _text sometimes.
408	 */
409	err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
410						 psession, machine, "_text");
411	if (err < 0)
412		err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
413							 psession, machine,
414							 "_stext");
415	if (err < 0)
416		pr_err("Couldn't record guest kernel [%d]'s reference"
417		       " relocation symbol.\n", machine->pid);
418}
419
420static struct perf_event_header finished_round_event = {
421	.size = sizeof(struct perf_event_header),
422	.type = PERF_RECORD_FINISHED_ROUND,
423};
424
425static void mmap_read_all(void)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
426{
427	int i;
 
 
 
 
 
428
429	for (i = 0; i < evsel_list->nr_mmaps; i++) {
430		if (evsel_list->mmap[i].base)
431			mmap_read(&evsel_list->mmap[i]);
 
 
432	}
433
434	if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
435		write_output(&finished_round_event, sizeof(finished_round_event));
 
 
 
 
 
 
 
436}
437
438static int __cmd_record(int argc, const char **argv)
439{
440	struct stat st;
441	int flags;
442	int err;
443	unsigned long waking = 0;
444	int child_ready_pipe[2], go_pipe[2];
445	const bool forks = argc > 0;
446	char buf;
447	struct machine *machine;
448
449	page_size = sysconf(_SC_PAGE_SIZE);
 
450
451	atexit(sig_atexit);
452	signal(SIGCHLD, sig_handler);
453	signal(SIGINT, sig_handler);
454	signal(SIGUSR1, sig_handler);
455
456	if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
457		perror("failed to create pipes");
458		exit(-1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
459	}
460
461	if (!output_name) {
462		if (!fstat(STDOUT_FILENO, &st) && S_ISFIFO(st.st_mode))
463			pipe_output = 1;
464		else
465			output_name = "perf.data";
 
466	}
467	if (output_name) {
468		if (!strcmp(output_name, "-"))
469			pipe_output = 1;
470		else if (!stat(output_name, &st) && st.st_size) {
471			if (write_mode == WRITE_FORCE) {
472				char oldname[PATH_MAX];
473				snprintf(oldname, sizeof(oldname), "%s.old",
474					 output_name);
475				unlink(oldname);
476				rename(output_name, oldname);
477			}
478		} else if (write_mode == WRITE_APPEND) {
479			write_mode = WRITE_FORCE;
 
480		}
 
 
 
481	}
482
483	flags = O_CREAT|O_RDWR;
484	if (write_mode == WRITE_APPEND)
485		file_new = 0;
486	else
487		flags |= O_TRUNC;
488
489	if (pipe_output)
490		output = STDOUT_FILENO;
491	else
492		output = open(output_name, flags, S_IRUSR | S_IWUSR);
493	if (output < 0) {
494		perror("failed to create output file");
495		exit(-1);
 
 
 
 
 
496	}
 
 
497
498	session = perf_session__new(output_name, O_WRONLY,
499				    write_mode == WRITE_FORCE, false, NULL);
500	if (session == NULL) {
501		pr_err("Not enough memory for reading perf file header\n");
502		return -1;
 
 
 
 
 
 
 
 
503	}
504
505	if (!no_buildid)
506		perf_header__set_feat(&session->header, HEADER_BUILD_ID);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
507
508	if (!file_new) {
509		err = perf_session__read_header(session, output);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
510		if (err < 0)
511			goto out_delete_session;
 
 
512	}
513
514	if (have_tracepoints(&evsel_list->entries))
515		perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
516
517	/* 512 kiB: default amount of unprivileged mlocked memory */
518	if (mmap_pages == UINT_MAX)
519		mmap_pages = (512 * 1024) / page_size;
 
 
 
 
 
 
 
 
 
 
520
521	if (forks) {
522		child_pid = fork();
523		if (child_pid < 0) {
524			perror("failed to fork");
525			exit(-1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
526		}
 
527
528		if (!child_pid) {
529			if (pipe_output)
530				dup2(2, 1);
531			close(child_ready_pipe[0]);
532			close(go_pipe[1]);
533			fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
534
535			/*
536			 * Do a dummy execvp to get the PLT entry resolved,
537			 * so we avoid the resolver overhead on the real
538			 * execvp call.
539			 */
540			execvp("", (char **)argv);
541
542			/*
543			 * Tell the parent we're ready to go
544			 */
545			close(child_ready_pipe[1]);
 
546
547			/*
548			 * Wait until the parent tells us to go.
549			 */
550			if (read(go_pipe[0], &buf, 1) == -1)
551				perror("unable to read pipe");
552
553			execvp(argv[0], (char **)argv);
 
 
554
555			perror(argv[0]);
556			kill(getppid(), SIGUSR1);
557			exit(-1);
558		}
 
 
559
560		if (!system_wide && target_tid == -1 && target_pid == -1)
561			evsel_list->threads->map[0] = child_pid;
 
562
563		close(child_ready_pipe[1]);
564		close(go_pipe[0]);
565		/*
566		 * wait for child to settle
 
 
567		 */
568		if (read(child_ready_pipe[0], &buf, 1) == -1) {
569			perror("unable to read pipe");
570			exit(-1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
571		}
572		close(child_ready_pipe[0]);
573	}
 
 
 
 
 
 
 
 
574
575	open_counters(evsel_list);
 
 
 
 
 
576
577	/*
578	 * perf_session__delete(session) will be called at atexit_header()
579	 */
580	atexit(atexit_header);
581
582	if (pipe_output) {
583		err = perf_header__write_pipe(output);
584		if (err < 0)
585			return err;
586	} else if (file_new) {
587		err = perf_session__write_header(session, evsel_list,
588						 output, false);
589		if (err < 0)
590			return err;
 
 
 
 
591	}
592
593	post_processing_offset = lseek(output, 0, SEEK_CUR);
 
594
595	if (pipe_output) {
596		err = perf_session__synthesize_attrs(session,
597						     process_synthesized_event);
598		if (err < 0) {
599			pr_err("Couldn't synthesize attrs.\n");
600			return err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
601		}
602
603		err = perf_event__synthesize_event_types(process_synthesized_event,
604							 session);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
605		if (err < 0) {
606			pr_err("Couldn't synthesize event_types.\n");
607			return err;
 
608		}
 
609
610		if (have_tracepoints(&evsel_list->entries)) {
611			/*
612			 * FIXME err <= 0 here actually means that
613			 * there were no tracepoints so its not really
614			 * an error, just that we don't need to
615			 * synthesize anything.  We really have to
616			 * return this more properly and also
617			 * propagate errors that now are calling die()
618			 */
619			err = perf_event__synthesize_tracing_data(output, evsel_list,
620								  process_synthesized_event,
621								  session);
622			if (err <= 0) {
623				pr_err("Couldn't record tracing data.\n");
624				return err;
625			}
626			advance_output(err);
 
 
 
 
 
 
 
 
 
627		}
628	}
629
630	machine = perf_session__find_host_machine(session);
631	if (!machine) {
632		pr_err("Couldn't find native kernel information.\n");
633		return -1;
 
 
 
634	}
635
636	err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
637						 session, machine, "_text");
638	if (err < 0)
639		err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
640							 session, machine, "_stext");
641	if (err < 0)
642		pr_err("Couldn't record kernel reference relocation symbol\n"
643		       "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
644		       "Check /proc/kallsyms permission or run as root.\n");
 
 
 
 
 
 
 
 
 
 
 
645
646	err = perf_event__synthesize_modules(process_synthesized_event,
647					     session, machine);
 
 
 
648	if (err < 0)
649		pr_err("Couldn't record kernel module information.\n"
650		       "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
651		       "Check /proc/modules permission or run as root.\n");
652
653	if (perf_guest)
654		perf_session__process_machines(session,
655					       perf_event__synthesize_guest_os);
656
657	if (!system_wide)
658		perf_event__synthesize_thread_map(evsel_list->threads,
659						  process_synthesized_event,
660						  session);
661	else
662		perf_event__synthesize_threads(process_synthesized_event,
663					       session);
664
665	if (realtime_prio) {
666		struct sched_param param;
667
668		param.sched_priority = realtime_prio;
669		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
670			pr_err("Could not set realtime priority.\n");
671			exit(-1);
 
672		}
673	}
674
675	perf_evlist__enable(evsel_list);
 
 
 
 
 
 
 
 
 
676
677	/*
678	 * Let the child rip
679	 */
680	if (forks)
681		close(go_pipe[1]);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
682
683	for (;;) {
684		int hits = samples;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
685
686		mmap_read_all();
 
 
 
 
 
 
 
 
 
 
 
 
687
688		if (hits == samples) {
689			if (done)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
690				break;
691			err = poll(evsel_list->pollfd, evsel_list->nr_fds, -1);
692			waking++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
693		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
694
695		if (done)
696			perf_evlist__disable(evsel_list);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
697	}
698
699	if (quiet || signr == SIGUSR1)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
700		return 0;
 
701
702	fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
703
704	/*
705	 * Approximate RIP event size: 24 bytes.
 
 
706	 */
707	fprintf(stderr,
708		"[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64 " samples) ]\n",
709		(double)bytes_written / 1024.0 / 1024.0,
710		output_name,
711		bytes_written / 24);
 
 
712
713	return 0;
 
714
715out_delete_session:
716	perf_session__delete(session);
717	return err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
718}
719
720static const char * const record_usage[] = {
721	"perf record [<options>] [<command>]",
722	"perf record [<options>] -- <command> [<options>]",
723	NULL
724};
 
725
726static bool force, append_file;
 
 
 
 
 
 
 
 
 
 
727
728const struct option record_options[] = {
729	OPT_CALLBACK('e', "event", &evsel_list, "event",
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
730		     "event selector. use 'perf list' to list available events",
731		     parse_events_option),
732	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
733		     "event filter", parse_filter),
734	OPT_INTEGER('p', "pid", &target_pid,
 
 
 
735		    "record events on existing process id"),
736	OPT_INTEGER('t', "tid", &target_tid,
737		    "record events on existing thread id"),
738	OPT_INTEGER('r', "realtime", &realtime_prio,
739		    "collect data with this RT SCHED_FIFO priority"),
740	OPT_BOOLEAN('D', "no-delay", &nodelay,
741		    "collect data without buffering"),
742	OPT_BOOLEAN('R', "raw-samples", &raw_samples,
743		    "collect raw sample records from all opened counters"),
744	OPT_BOOLEAN('a', "all-cpus", &system_wide,
745			    "system-wide collection from all CPUs"),
746	OPT_BOOLEAN('A', "append", &append_file,
747			    "append to the output file to do incremental profiling"),
748	OPT_STRING('C', "cpu", &cpu_list, "cpu",
749		    "list of cpus to monitor"),
750	OPT_BOOLEAN('f', "force", &force,
751			"overwrite existing data file (deprecated)"),
752	OPT_U64('c', "count", &user_interval, "event period to sample"),
753	OPT_STRING('o', "output", &output_name, "file",
754		    "output file name"),
755	OPT_BOOLEAN('i', "no-inherit", &no_inherit,
756		    "child tasks do not inherit counters"),
757	OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
758	OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
759	OPT_BOOLEAN(0, "group", &group,
760		    "put the counters into a counter group"),
761	OPT_BOOLEAN('g', "call-graph", &call_graph,
762		    "do call-graph (stack chain/backtrace) recording"),
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
763	OPT_INCR('v', "verbose", &verbose,
764		    "be more verbose (show counter open errors, etc)"),
765	OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
766	OPT_BOOLEAN('s', "stat", &inherit_stat,
767		    "per thread counts"),
768	OPT_BOOLEAN('d', "data", &sample_address,
769		    "Sample addresses"),
770	OPT_BOOLEAN('T', "timestamp", &sample_time, "Sample timestamps"),
771	OPT_BOOLEAN('n', "no-samples", &no_samples,
 
 
 
 
 
 
 
 
 
 
 
 
772		    "don't sample"),
773	OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
774		    "do not update the buildid cache"),
775	OPT_BOOLEAN('B', "no-buildid", &no_buildid,
776		    "do not collect buildids in perf.data"),
777	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
 
 
778		     "monitor event in cgroup name only",
779		     parse_cgroups),
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
780	OPT_END()
781};
782
783int cmd_record(int argc, const char **argv, const char *prefix __used)
 
 
784{
785	int err = -ENOMEM;
786	struct perf_evsel *pos;
787
788	evsel_list = perf_evlist__new(NULL, NULL);
789	if (evsel_list == NULL)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
790		return -ENOMEM;
791
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
792	argc = parse_options(argc, argv, record_options, record_usage,
793			    PARSE_OPT_STOP_AT_NON_OPTION);
794	if (!argc && target_pid == -1 && target_tid == -1 &&
795		!system_wide && !cpu_list)
796		usage_with_options(record_usage, record_options);
 
 
 
 
 
 
 
 
 
 
 
 
 
797
798	if (force && append_file) {
799		fprintf(stderr, "Can't overwrite and append at the same time."
800				" You need to choose between -f and -A");
801		usage_with_options(record_usage, record_options);
802	} else if (append_file) {
803		write_mode = WRITE_APPEND;
804	} else {
805		write_mode = WRITE_FORCE;
806	}
807
808	if (nr_cgroups && !system_wide) {
809		fprintf(stderr, "cgroup monitoring only available in"
810			" system-wide mode\n");
811		usage_with_options(record_usage, record_options);
 
 
 
 
 
 
 
 
 
812	}
813
814	symbol__init();
 
 
 
 
815
816	if (symbol_conf.kptr_restrict)
817		pr_warning(
818"WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
819"check /proc/sys/kernel/kptr_restrict.\n\n"
820"Samples in kernel functions may not be resolved if a suitable vmlinux\n"
821"file is not found in the buildid cache or in the vmlinux path.\n\n"
822"Samples in kernel modules won't be resolved at all.\n\n"
823"If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
824"even with a suitable vmlinux or kallsyms file.\n\n");
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
825
826	if (no_buildid_cache || no_buildid)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
827		disable_buildid_cache();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
828
829	if (evsel_list->nr_entries == 0 &&
830	    perf_evlist__add_default(evsel_list) < 0) {
831		pr_err("Not enough memory for event selector list\n");
832		goto out_symbol_exit;
 
 
 
 
 
833	}
834
835	if (target_pid != -1)
836		target_tid = target_pid;
 
 
 
837
838	if (perf_evlist__create_maps(evsel_list, target_pid,
839				     target_tid, cpu_list) < 0)
840		usage_with_options(record_usage, record_options);
 
 
 
 
841
842	list_for_each_entry(pos, &evsel_list->entries, node) {
843		if (perf_evsel__alloc_fd(pos, evsel_list->cpus->nr,
844					 evsel_list->threads->nr) < 0)
845			goto out_free_fd;
846		if (perf_header__push_event(pos->attr.config, event_name(pos)))
847			goto out_free_fd;
848	}
849
850	if (perf_evlist__alloc_pollfd(evsel_list) < 0)
851		goto out_free_fd;
 
 
 
 
 
 
 
 
852
853	if (user_interval != ULLONG_MAX)
854		default_interval = user_interval;
855	if (user_freq != UINT_MAX)
856		freq = user_freq;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
857
858	/*
859	 * User specified count overrides default frequency.
 
 
860	 */
861	if (default_interval)
862		freq = 0;
863	else if (freq) {
864		default_interval = freq;
865	} else {
866		fprintf(stderr, "frequency and count are zero, aborting\n");
 
 
 
 
 
 
 
 
 
 
 
 
 
 
867		err = -EINVAL;
868		goto out_free_fd;
 
 
 
 
 
 
 
 
 
 
 
 
869	}
870
871	err = __cmd_record(argc, argv);
872out_free_fd:
873	perf_evlist__delete_maps(evsel_list);
874out_symbol_exit:
 
 
 
 
 
 
 
 
 
 
875	symbol__exit();
 
 
 
 
 
876	return err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
877}