23 #include <sys/syscall.h>
24 #include <sys/ioctl.h>
26 #include <sys/prctl.h>
31 #include <linux/unistd.h>
32 #include <linux/types.h>
34 #include "../../include/linux/perf_counter.h"
38 static int nr_counters = 0;
39 static __u64 event_id[MAX_COUNTERS] = { };
40 static int default_interval = 100000;
41 static int event_count[MAX_COUNTERS];
42 static int fd[MAX_NR_CPUS][MAX_COUNTERS];
43 static int nr_cpus = 0;
44 static unsigned int page_size;
45 static unsigned int mmap_pages = 16;
47 static char *output_name = "output.perf";
49 static unsigned int realtime_prio = 0;
51 const unsigned int default_count[] = {
65 static struct event_symbol event_symbols[] = {
66 {EID(PERF_TYPE_HARDWARE, PERF_COUNT_CPU_CYCLES), "cpu-cycles", },
67 {EID(PERF_TYPE_HARDWARE, PERF_COUNT_CPU_CYCLES), "cycles", },
68 {EID(PERF_TYPE_HARDWARE, PERF_COUNT_INSTRUCTIONS), "instructions", },
69 {EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_REFERENCES), "cache-references", },
70 {EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_MISSES), "cache-misses", },
71 {EID(PERF_TYPE_HARDWARE, PERF_COUNT_BRANCH_INSTRUCTIONS), "branch-instructions", },
72 {EID(PERF_TYPE_HARDWARE, PERF_COUNT_BRANCH_INSTRUCTIONS), "branches", },
73 {EID(PERF_TYPE_HARDWARE, PERF_COUNT_BRANCH_MISSES), "branch-misses", },
74 {EID(PERF_TYPE_HARDWARE, PERF_COUNT_BUS_CYCLES), "bus-cycles", },
76 {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_CLOCK), "cpu-clock", },
77 {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK), "task-clock", },
78 {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS), "page-faults", },
79 {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS), "faults", },
80 {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS_MIN), "minor-faults", },
81 {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS_MAJ), "major-faults", },
82 {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CONTEXT_SWITCHES), "context-switches", },
83 {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CONTEXT_SWITCHES), "cs", },
84 {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_MIGRATIONS), "cpu-migrations", },
85 {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_MIGRATIONS), "migrations", },
89 * Each event can have multiple symbolic names.
90 * Symbolic names are (almost) exactly matched.
92 static __u64 match_event_symbols(char *str)
98 if (sscanf(str, "r%llx", &config) == 1)
99 return config | PERF_COUNTER_RAW_MASK;
101 if (sscanf(str, "%d:%llu", &type, &id) == 2)
102 return EID(type, id);
104 for (i = 0; i < ARRAY_SIZE(event_symbols); i++) {
105 if (!strncmp(str, event_symbols[i].symbol,
106 strlen(event_symbols[i].symbol)))
107 return event_symbols[i].event;
113 static int parse_events(char *str)
118 if (nr_counters == MAX_COUNTERS)
121 config = match_event_symbols(str);
125 event_id[nr_counters] = config;
128 str = strstr(str, ",");
137 #define __PERF_COUNTER_FIELD(config, name) \
138 ((config & PERF_COUNTER_##name##_MASK) >> PERF_COUNTER_##name##_SHIFT)
140 #define PERF_COUNTER_RAW(config) __PERF_COUNTER_FIELD(config, RAW)
141 #define PERF_COUNTER_CONFIG(config) __PERF_COUNTER_FIELD(config, CONFIG)
142 #define PERF_COUNTER_TYPE(config) __PERF_COUNTER_FIELD(config, TYPE)
143 #define PERF_COUNTER_ID(config) __PERF_COUNTER_FIELD(config, EVENT)
145 static void display_events_help(void)
151 " -e EVENT --event=EVENT # symbolic-name abbreviations");
153 for (i = 0; i < ARRAY_SIZE(event_symbols); i++) {
156 e = event_symbols[i].event;
157 type = PERF_COUNTER_TYPE(e);
158 id = PERF_COUNTER_ID(e);
160 printf("\n %d:%d: %-20s",
161 type, id, event_symbols[i].symbol);
165 " rNNN: raw PMU events (eventsel+umask)\n\n");
168 static void display_help(void)
171 "Usage: perf-record [<options>]\n"
172 "perf-record Options (up to %d event types can be specified at once):\n\n",
175 display_events_help();
178 " -c CNT --count=CNT # event period to sample\n"
179 " -m pages --mmap_pages=<pages> # number of mmap data pages\n"
180 " -o file --output=<file> # output file\n"
181 " -r prio --realtime=<prio> # use RT prio\n"
187 static void process_options(int argc, char *argv[])
189 int error = 0, counter;
192 int option_index = 0;
193 /** Options for getopt */
194 static struct option long_options[] = {
195 {"count", required_argument, NULL, 'c'},
196 {"event", required_argument, NULL, 'e'},
197 {"mmap_pages", required_argument, NULL, 'm'},
198 {"output", required_argument, NULL, 'o'},
199 {"realtime", required_argument, NULL, 'r'},
202 int c = getopt_long(argc, argv, "+:c:e:m:o:r:",
203 long_options, &option_index);
208 case 'c': default_interval = atoi(optarg); break;
209 case 'e': error = parse_events(optarg); break;
210 case 'm': mmap_pages = atoi(optarg); break;
211 case 'o': output_name = strdup(optarg); break;
212 case 'r': realtime_prio = atoi(optarg); break;
213 default: error = 1; break;
224 for (counter = 0; counter < nr_counters; counter++) {
225 if (event_count[counter])
228 event_count[counter] = default_interval;
239 static unsigned int mmap_read_head(struct mmap_data *md)
241 struct perf_counter_mmap_page *pc = md->base;
244 head = pc->data_head;
251 static struct timeval last_read, this_read;
253 static void mmap_read(struct mmap_data *md)
255 unsigned int head = mmap_read_head(md);
256 unsigned int old = md->prev;
257 unsigned char *data = md->base + page_size;
262 gettimeofday(&this_read, NULL);
265 * If we're further behind than half the buffer, there's a chance
266 * the writer will bite our tail and screw up the events under us.
268 * If we somehow ended up ahead of the head, we got messed up.
270 * In either case, truncate and restart at head.
273 if (diff > md->mask / 2 || diff < 0) {
277 timersub(&this_read, &last_read, &iv);
278 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
280 fprintf(stderr, "WARNING: failed to keep up with mmap data."
281 " Last read %lu msecs ago.\n", msecs);
284 * head points to a known good entry, start there.
289 last_read = this_read;
296 if ((old & md->mask) + size != (head & md->mask)) {
297 buf = &data[old & md->mask];
298 size = md->mask + 1 - (old & md->mask);
301 int ret = write(output, buf, size);
303 perror("failed to write");
311 buf = &data[old & md->mask];
315 int ret = write(output, buf, size);
317 perror("failed to write");
327 static volatile int done = 0;
329 static void sigchld_handler(int sig)
335 int cmd_record(int argc, char **argv)
337 struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
338 struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
339 struct perf_counter_hw_event hw_event;
340 int i, counter, group_fd, nr_poll = 0;
344 page_size = sysconf(_SC_PAGE_SIZE);
346 process_options(argc, argv);
348 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
349 assert(nr_cpus <= MAX_NR_CPUS);
350 assert(nr_cpus >= 0);
352 output = open(output_name, O_CREAT|O_RDWR, S_IRWXU);
354 perror("failed to create output file");
361 for (i = 0; i < nr_cpus; i++) {
363 for (counter = 0; counter < nr_counters; counter++) {
365 memset(&hw_event, 0, sizeof(hw_event));
366 hw_event.config = event_id[counter];
367 hw_event.irq_period = event_count[counter];
368 hw_event.record_type = PERF_RECORD_IP | PERF_RECORD_TID;
373 fd[i][counter] = sys_perf_counter_open(&hw_event, -1, i, group_fd, 0);
374 if (fd[i][counter] < 0) {
376 printf("kerneltop error: syscall returned with %d (%s)\n",
377 fd[i][counter], strerror(err));
379 printf("Are you root?\n");
382 assert(fd[i][counter] >= 0);
383 fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
386 * First counter acts as the group leader:
388 if (group && group_fd == -1)
389 group_fd = fd[i][counter];
391 event_array[nr_poll].fd = fd[i][counter];
392 event_array[nr_poll].events = POLLIN;
395 mmap_array[i][counter].counter = counter;
396 mmap_array[i][counter].prev = 0;
397 mmap_array[i][counter].mask = mmap_pages*page_size - 1;
398 mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
399 PROT_READ, MAP_SHARED, fd[i][counter], 0);
400 if (mmap_array[i][counter].base == MAP_FAILED) {
401 printf("kerneltop error: failed to mmap with %d (%s)\n",
402 errno, strerror(errno));
408 signal(SIGCHLD, sigchld_handler);
412 perror("failed to fork");
415 if (execvp(argv[0], argv)) {
422 struct sched_param param;
424 param.sched_priority = realtime_prio;
425 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
426 printf("Could not set realtime priority.\n");
432 * TODO: store the current /proc/$/maps information somewhere
438 for (i = 0; i < nr_cpus; i++) {
439 for (counter = 0; counter < nr_counters; counter++)
440 mmap_read(&mmap_array[i][counter]);
444 ret = poll(event_array, nr_poll, 100);