2 * Read a file and write the contents to stdout. If a given read takes
3 * longer than 'max_us' time, then we schedule a new thread to handle
4 * the next read. This avoids the coordinated omission problem, where
5 * one request appears to take a long time, but in reality a lot of
6 * requests would have been slow, but we don't notice since new submissions
7 * are not being issued if just 1 is held up.
11 * $ time (./read-to-pipe-async -f randfile.gz | gzip -dc > outfile; sync)
13 * This will read randfile.gz and log the latencies of doing so, while
14 * piping the output to gzip to decompress it. Any latencies over max_us
15 * are logged when they happen, and latency buckets are displayed at the
18 * gcc -Wall -g -O2 -o read-to-pipe-async read-to-pipe-async.c -lpthread
20 * Copyright (C) 2016 Jens Axboe
28 #include <sys/types.h>
40 static int max_us = 10000;
42 static int separate_writer = 1;
45 #define PLAT_VAL (1 << PLAT_BITS)
46 #define PLAT_GROUP_NR 19
47 #define PLAT_NR (PLAT_GROUP_NR * PLAT_VAL)
48 #define PLAT_LIST_MAX 20
51 unsigned int plat[PLAT_NR];
52 unsigned int nr_samples;
58 static double plist[PLAT_LIST_MAX] = { 50.0, 75.0, 90.0, 95.0, 99.0, 99.5, 99.9, 99.99, 99.999, 99.9999, };
65 pthread_mutex_t done_lock;
66 pthread_cond_t done_cond;
70 struct writer_thread {
71 struct flist_head list;
72 struct flist_head done_list;
74 struct thread_data thread;
77 struct reader_thread {
78 struct flist_head list;
79 struct flist_head done_list;
84 struct thread_data thread;
88 struct flist_head list;
94 struct writer_thread *writer;
95 struct reader_thread *reader;
101 static struct reader_thread reader_thread;
102 static struct writer_thread writer_thread;
104 uint64_t utime_since(const struct timeval *s, const struct timeval *e)
109 sec = e->tv_sec - s->tv_sec;
110 usec = e->tv_usec - s->tv_usec;
111 if (sec > 0 && usec < 0) {
116 if (sec < 0 || (sec == 0 && usec < 0))
119 ret = sec * 1000000ULL + usec;
124 static struct work_item *find_seq(struct writer_thread *w, unsigned int seq)
126 struct work_item *work;
127 struct flist_head *entry;
129 if (flist_empty(&w->list))
132 flist_for_each(entry, &w->list) {
133 work = flist_entry(entry, struct work_item, list);
134 if (work->seq == seq)
141 static unsigned int plat_val_to_idx(unsigned int val)
143 unsigned int msb, error_bits, base, offset;
145 /* Find MSB starting from bit 0 */
149 msb = sizeof(val)*8 - __builtin_clz(val) - 1;
152 * MSB <= (PLAT_BITS-1), cannot be rounded off. Use
153 * all bits of the sample as index
155 if (msb <= PLAT_BITS)
158 /* Compute the number of error bits to discard*/
159 error_bits = msb - PLAT_BITS;
161 /* Compute the number of buckets before the group */
162 base = (error_bits + 1) << PLAT_BITS;
165 * Discard the error bits and apply the mask to find the
166 * index for the buckets in the group
168 offset = (PLAT_VAL - 1) & (val >> error_bits);
170 /* Make sure the index does not exceed (array size - 1) */
171 return (base + offset) < (PLAT_NR - 1) ?
172 (base + offset) : (PLAT_NR - 1);
176 * Convert the given index of the bucket array to the value
177 * represented by the bucket
179 static unsigned int plat_idx_to_val(unsigned int idx)
181 unsigned int error_bits, k, base;
183 assert(idx < PLAT_NR);
185 /* MSB <= (PLAT_BITS-1), cannot be rounded off. Use
186 * all bits of the sample as index */
187 if (idx < (PLAT_VAL << 1))
190 /* Find the group and compute the minimum value of that group */
191 error_bits = (idx >> PLAT_BITS) - 1;
192 base = 1 << (error_bits + PLAT_BITS);
194 /* Find its bucket number of the group */
197 /* Return the mean of the range of the bucket */
198 return base + ((k + 0.5) * (1 << error_bits));
201 static void add_lat(struct stats *s, unsigned int us, const char *name)
211 fprintf(stderr, "%s latency=%u usec\n", name, us);
215 lat_index = plat_val_to_idx(us);
216 __sync_fetch_and_add(&s->plat[lat_index], 1);
217 __sync_fetch_and_add(&s->nr_samples, 1);
220 static int write_work(struct work_item *work)
225 gettimeofday(&s, NULL);
226 ret = write(STDOUT_FILENO, work->buf, work->buf_size);
227 gettimeofday(&e, NULL);
228 assert(ret == work->buf_size);
230 add_lat(&work->writer->s, utime_since(&s, &e), "write");
231 return work->seq + 1;
234 static void *writer_fn(void *data)
236 struct writer_thread *wt = data;
237 struct work_item *work;
238 unsigned int seq = 1;
241 while (!wt->thread.exit || !flist_empty(&wt->list)) {
242 pthread_mutex_lock(&wt->thread.lock);
245 flist_add_tail(&work->list, &wt->done_list);
249 work = find_seq(wt, seq);
251 flist_del_init(&work->list);
253 pthread_cond_wait(&wt->thread.cond, &wt->thread.lock);
255 pthread_mutex_unlock(&wt->thread.lock);
258 seq = write_work(work);
262 pthread_cond_signal(&wt->thread.done_cond);
266 static void reader_work(struct work_item *work)
274 gettimeofday(&s, NULL);
276 left = work->buf_size;
280 ret = pread(work->fd, buf, left, off);
282 fprintf(stderr, "zero read\n");
284 } else if (ret < 0) {
285 fprintf(stderr, "errno=%d\n", errno);
293 gettimeofday(&e, NULL);
295 add_lat(&work->reader->s, utime_since(&s, &e), "read");
297 pthread_cond_signal(&work->cond);
299 if (separate_writer) {
300 pthread_mutex_lock(&work->writer->thread.lock);
301 flist_add_tail(&work->list, &work->writer->list);
302 pthread_mutex_unlock(&work->writer->thread.lock);
303 pthread_cond_signal(&work->writer->thread.cond);
305 struct reader_thread *rt = work->reader;
306 struct work_item *next = NULL;
307 struct flist_head *entry;
310 * Write current work if it matches in sequence.
312 if (work->seq == rt->write_seq)
315 pthread_mutex_lock(&rt->thread.lock);
317 flist_add_tail(&work->list, &rt->done_list);
320 * See if the next work item is here, if so, write it
323 flist_for_each(entry, &rt->done_list) {
324 next = flist_entry(entry, struct work_item, list);
325 if (next->seq == rt->write_seq) {
327 flist_del(&work->list);
332 pthread_mutex_unlock(&rt->thread.lock);
337 __sync_fetch_and_add(&rt->write_seq, 1);
342 static void *reader_one_off(void *data)
348 static void *reader_fn(void *data)
350 struct reader_thread *rt = data;
351 struct work_item *work;
353 while (!rt->thread.exit || !flist_empty(&rt->list)) {
355 pthread_mutex_lock(&rt->thread.lock);
356 if (!flist_empty(&rt->list)) {
357 work = flist_first_entry(&rt->list, struct work_item, list);
358 flist_del_init(&work->list);
360 pthread_cond_wait(&rt->thread.cond, &rt->thread.lock);
361 pthread_mutex_unlock(&rt->thread.lock);
371 pthread_cond_signal(&rt->thread.done_cond);
375 static void queue_work(struct reader_thread *rt, struct work_item *work)
378 pthread_mutex_lock(&rt->thread.lock);
379 flist_add_tail(&work->list, &rt->list);
380 pthread_mutex_unlock(&rt->thread.lock);
383 pthread_create(&rt->thread.thread, NULL, reader_fn, rt);
384 } else if (!rt->busy && !pthread_mutex_trylock(&rt->thread.lock)) {
385 flist_add_tail(&work->list, &rt->list);
386 pthread_mutex_unlock(&rt->thread.lock);
388 pthread_cond_signal(&rt->thread.cond);
390 int ret = pthread_create(&work->thread, NULL, reader_one_off, work);
392 fprintf(stderr, "pthread_create=%d\n", ret);
394 pthread_detach(work->thread);
398 static unsigned int calc_percentiles(unsigned int *io_u_plat, unsigned long nr,
399 unsigned int **output)
401 unsigned long sum = 0;
402 unsigned int len, i, j = 0;
403 unsigned int oval_len = 0;
404 unsigned int *ovals = NULL;
408 while (len < PLAT_LIST_MAX && plist[len] != 0.0)
415 * Calculate bucket values, note down max and min values
418 for (i = 0; i < PLAT_NR && !is_last; i++) {
420 while (sum >= (plist[j] / 100.0 * nr)) {
421 assert(plist[j] <= 100.0);
425 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
428 ovals[j] = plat_idx_to_val(i);
429 is_last = (j == len - 1);
441 static void show_latencies(struct stats *s, const char *msg)
443 unsigned int *ovals = NULL;
446 len = calc_percentiles(s->plat, s->nr_samples, &ovals);
448 fprintf(stderr, "Latency percentiles (usec) (%s)\n", msg);
449 for (i = 0; i < len; i++)
450 fprintf(stderr, "\t%2.4fth: %u\n", plist[i], ovals[i]);
456 fprintf(stderr, "\tOver=%u, min=%u, max=%u\n", s->over, s->min, s->max);
459 static void init_thread(struct thread_data *thread)
461 pthread_cond_init(&thread->cond, NULL);
462 pthread_cond_init(&thread->done_cond, NULL);
463 pthread_mutex_init(&thread->lock, NULL);
464 pthread_mutex_init(&thread->done_lock, NULL);
468 static void exit_thread(struct thread_data *thread,
469 void fn(struct writer_thread *),
470 struct writer_thread *wt)
473 pthread_cond_signal(&thread->cond);
475 while (!thread->done) {
476 pthread_mutex_lock(&thread->done_lock);
481 clock_gettime(CLOCK_REALTIME, &t);
485 pthread_cond_timedwait(&thread->done_cond, &thread->done_lock, &t);
488 pthread_cond_wait(&thread->done_cond, &thread->done_lock);
490 pthread_mutex_unlock(&thread->done_lock);
494 static int usage(char *argv[])
496 fprintf(stderr, "%s: [-b blocksize] [-t max usec] [-w separate writer] -f file\n", argv[0]);
500 static int parse_options(int argc, char *argv[])
504 while ((c = getopt(argc, argv, "f:b:t:w:")) != -1) {
507 file = strdup(optarg);
513 max_us = atoi(optarg);
516 separate_writer = atoi(optarg);
517 if (!separate_writer)
518 fprintf(stderr, "inline writing is broken\n");
532 static void prune_done_entries(struct writer_thread *wt)
536 if (flist_empty(&wt->done_list))
539 if (pthread_mutex_trylock(&wt->thread.lock))
542 if (!flist_empty(&wt->done_list))
543 flist_splice_init(&wt->done_list, &list);
544 pthread_mutex_unlock(&wt->thread.lock);
546 while (!flist_empty(&list)) {
547 struct work_item *work;
549 work = flist_first_entry(&list, struct work_item, list);
550 flist_del(&work->list);
552 pthread_cond_destroy(&work->cond);
553 pthread_mutex_destroy(&work->lock);
559 int main(int argc, char *argv[])
561 struct timeval s, re, we;
562 struct reader_thread *rt;
563 struct writer_thread *wt;
570 if (parse_options(argc, argv))
573 fd = open(file, O_RDONLY);
579 if (fstat(fd, &sb) < 0) {
585 init_thread(&wt->thread);
586 INIT_FLIST_HEAD(&wt->list);
587 INIT_FLIST_HEAD(&wt->done_list);
590 pthread_create(&wt->thread.thread, NULL, writer_fn, wt);
593 init_thread(&rt->thread);
594 INIT_FLIST_HEAD(&rt->list);
595 INIT_FLIST_HEAD(&rt->done_list);
604 gettimeofday(&s, NULL);
607 struct work_item *work;
611 prune_done_entries(wt);
613 this_len = sb.st_size;
617 work = calloc(1, sizeof(*work));
618 work->buf = malloc(this_len);
619 work->buf_size = this_len;
625 pthread_cond_init(&work->cond, NULL);
626 pthread_mutex_init(&work->lock, NULL);
628 queue_work(rt, work);
630 clock_gettime(CLOCK_REALTIME, &t);
631 t.tv_nsec += max_us * 1000ULL;
632 if (t.tv_nsec >= 1000000000ULL) {
633 t.tv_nsec -= 1000000000ULL;
637 pthread_mutex_lock(&work->lock);
638 pthread_cond_timedwait(&work->cond, &work->lock, &t);
639 pthread_mutex_unlock(&work->lock);
642 sb.st_size -= this_len;
646 exit_thread(&rt->thread, NULL, NULL);
647 gettimeofday(&re, NULL);
649 exit_thread(&wt->thread, prune_done_entries, wt);
650 gettimeofday(&we, NULL);
652 show_latencies(&rt->s, "READERS");
653 show_latencies(&wt->s, "WRITERS");
656 rate = (bytes * 1000UL * 1000UL) / utime_since(&s, &re);
657 fprintf(stderr, "Read rate (KB/sec) : %lu\n", rate);
658 rate = (bytes * 1000UL * 1000UL) / utime_since(&s, &we);
659 fprintf(stderr, "Write rate (KB/sec): %lu\n", rate);