t/read-to-pipe-async.c

   1 /*
   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.
   8  *
   9  * One test case:
  10  *
  11  * $ time (./read-to-pipe-async -f randfile.gz | gzip -dc > outfile; sync)
  12  *
  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
  16  * end of the run
  17  *
  18  * gcc -Wall -g -O2 -o read-to-pipe-async read-to-pipe-async.c -lpthread
  19  *
  20  * Copyright (C) 2016 Jens Axboe
  21  *
  22  */
  23
  24 #include <stdio.h>
  25 #include <stdlib.h>
  26 #include <unistd.h>
  27 #include <fcntl.h>
  28 #include <sys/time.h>
  29 #include <sys/types.h>
  30 #include <sys/stat.h>
  31 #include <inttypes.h>
  32 #include <string.h>
  33 #include <pthread.h>
  34 #include <errno.h>
  35 #include <assert.h>
  36
  37 #include "../flist.h"
  38
  39 static int bs = 4096;
  40 static int max_us = 10000;
  41 static char *file;
  42 static int separate_writer = 1;
  43
  44 #define PLAT_BITS       8
  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
  49
  50 struct stats {
  51         unsigned int plat[PLAT_NR];
  52         unsigned int nr_samples;
  53         unsigned int max;
  54         unsigned int min;
  55         unsigned int over;
  56 };
  57
  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, };
  59
  60 struct thread_data {
  61         int exit;
  62         int done;
  63         pthread_mutex_t lock;
  64         pthread_cond_t cond;
  65         pthread_mutex_t done_lock;
  66         pthread_cond_t done_cond;
  67         pthread_t thread;
  68 };
  69
  70 struct writer_thread {
  71         struct flist_head list;
  72         struct flist_head done_list;
  73         struct stats s;
  74         struct thread_data thread;
  75 };
  76
  77 struct reader_thread {
  78         struct flist_head list;
  79         struct flist_head done_list;
  80         int started;
  81         int busy;
  82         int write_seq;
  83         struct stats s;
  84         struct thread_data thread;
  85 };
  86
  87 struct work_item {
  88         struct flist_head list;
  89         void *buf;
  90         size_t buf_size;
  91         off_t off;
  92         int fd;
  93         int seq;
  94         struct writer_thread *writer;
  95         struct reader_thread *reader;
  96         pthread_mutex_t lock;
  97         pthread_cond_t cond;
  98         pthread_t thread;
  99 };
 100
 101 static struct reader_thread reader_thread;
 102 static struct writer_thread writer_thread;
 103
 104 uint64_t utime_since(const struct timespec *s, const struct timespec *e)
 105 {
 106         long sec, usec;
 107         uint64_t ret;
 108
 109         sec = e->tv_sec - s->tv_sec;
 110         usec = (e->tv_nsec - s->tv_nsec) / 1000;
 111         if (sec > 0 && usec < 0) {
 112                 sec--;
 113                 usec += 1000000;
 114         }
 115
 116         if (sec < 0 || (sec == 0 && usec < 0))
 117                 return 0;
 118
 119         ret = sec * 1000000ULL + usec;
 120
 121         return ret;
 122 }
 123
 124 static struct work_item *find_seq(struct writer_thread *w, unsigned int seq)
 125 {
 126         struct work_item *work;
 127         struct flist_head *entry;
 128
 129         if (flist_empty(&w->list))
 130                 return NULL;
 131
 132         flist_for_each(entry, &w->list) {
 133                 work = flist_entry(entry, struct work_item, list);
 134                 if (work->seq == seq)
 135                         return work;
 136         }
 137
 138         return NULL;
 139 }
 140
 141 static unsigned int plat_val_to_idx(unsigned int val)
 142 {
 143         unsigned int msb, error_bits, base, offset;
 144
 145         /* Find MSB starting from bit 0 */
 146         if (val == 0)
 147                 msb = 0;
 148         else
 149                 msb = sizeof(val)*8 - __builtin_clz(val) - 1;
 150
 151         /*
 152          * MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 153          * all bits of the sample as index
 154          */
 155         if (msb <= PLAT_BITS)
 156                 return val;
 157
 158         /* Compute the number of error bits to discard*/
 159         error_bits = msb - PLAT_BITS;
 160
 161         /* Compute the number of buckets before the group */
 162         base = (error_bits + 1) << PLAT_BITS;
 163
 164         /*
 165          * Discard the error bits and apply the mask to find the
 166          * index for the buckets in the group
 167          */
 168         offset = (PLAT_VAL - 1) & (val >> error_bits);
 169
 170         /* Make sure the index does not exceed (array size - 1) */
 171         return (base + offset) < (PLAT_NR - 1) ?
 172                 (base + offset) : (PLAT_NR - 1);
 173 }
 174
 175 /*
 176  * Convert the given index of the bucket array to the value
 177  * represented by the bucket
 178  */
 179 static unsigned int plat_idx_to_val(unsigned int idx)
 180 {
 181         unsigned int error_bits, k, base;
 182
 183         assert(idx < PLAT_NR);
 184
 185         /* MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 186          * all bits of the sample as index */
 187         if (idx < (PLAT_VAL << 1))
 188                 return idx;
 189
 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);
 193
 194         /* Find its bucket number of the group */
 195         k = idx % PLAT_VAL;
 196
 197         /* Return the mean of the range of the bucket */
 198         return base + ((k + 0.5) * (1 << error_bits));
 199 }
 200
 201 static void add_lat(struct stats *s, unsigned int us, const char *name)
 202 {
 203         int lat_index = 0;
 204
 205         if (us > s->max)
 206                 s->max = us;
 207         if (us < s->min)
 208                 s->min = us;
 209
 210         if (us > max_us) {
 211                 fprintf(stderr, "%s latency=%u usec\n", name, us);
 212                 s->over++;
 213         }
 214
 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);
 218 }
 219
 220 static int write_work(struct work_item *work)
 221 {
 222         struct timespec s, e;
 223         ssize_t ret;
 224
 225         clock_gettime(CLOCK_MONOTONIC, &s);
 226         ret = write(STDOUT_FILENO, work->buf, work->buf_size);
 227         clock_gettime(CLOCK_MONOTONIC, &e);
 228         assert(ret == work->buf_size);
 229
 230         add_lat(&work->writer->s, utime_since(&s, &e), "write");
 231         return work->seq + 1;
 232 }
 233
 234 static void thread_exiting(struct thread_data *thread)
 235 {
 236         __sync_fetch_and_add(&thread->done, 1);
 237         pthread_cond_signal(&thread->done_cond);
 238 }
 239
 240 static void *writer_fn(void *data)
 241 {
 242         struct writer_thread *wt = data;
 243         struct work_item *work;
 244         unsigned int seq = 1;
 245
 246         work = NULL;
 247         while (!wt->thread.exit || !flist_empty(&wt->list)) {
 248                 pthread_mutex_lock(&wt->thread.lock);
 249
 250                 if (work) {
 251                         flist_add_tail(&work->list, &wt->done_list);
 252                         work = NULL;
 253                 }
 254
 255                 work = find_seq(wt, seq);
 256                 if (work)
 257                         flist_del_init(&work->list);
 258                 else
 259                         pthread_cond_wait(&wt->thread.cond, &wt->thread.lock);
 260
 261                 pthread_mutex_unlock(&wt->thread.lock);
 262
 263                 if (work)
 264                         seq = write_work(work);
 265         }
 266
 267         thread_exiting(&wt->thread);
 268         return NULL;
 269 }
 270
 271 static void reader_work(struct work_item *work)
 272 {
 273         struct timespec s, e;
 274         ssize_t ret;
 275         size_t left;
 276         void *buf;
 277         off_t off;
 278
 279         clock_gettime(CLOCK_MONOTONIC, &s);
 280
 281         left = work->buf_size;
 282         buf = work->buf;
 283         off = work->off;
 284         while (left) {
 285                 ret = pread(work->fd, buf, left, off);
 286                 if (!ret) {
 287                         fprintf(stderr, "zero read\n");
 288                         break;
 289                 } else if (ret < 0) {
 290                         fprintf(stderr, "errno=%d\n", errno);
 291                         break;
 292                 }
 293                 left -= ret;
 294                 off += ret;
 295                 buf += ret;
 296         }
 297
 298         clock_gettime(CLOCK_MONOTONIC, &e);
 299
 300         add_lat(&work->reader->s, utime_since(&s, &e), "read");
 301
 302         pthread_cond_signal(&work->cond);
 303
 304         if (separate_writer) {
 305                 pthread_mutex_lock(&work->writer->thread.lock);
 306                 flist_add_tail(&work->list, &work->writer->list);
 307                 pthread_mutex_unlock(&work->writer->thread.lock);
 308                 pthread_cond_signal(&work->writer->thread.cond);
 309         } else {
 310                 struct reader_thread *rt = work->reader;
 311                 struct work_item *next = NULL;
 312                 struct flist_head *entry;
 313
 314                 /*
 315                  * Write current work if it matches in sequence.
 316                  */
 317                 if (work->seq == rt->write_seq)
 318                         goto write_it;
 319
 320                 pthread_mutex_lock(&rt->thread.lock);
 321
 322                 flist_add_tail(&work->list, &rt->done_list);
 323
 324                 /*
 325                  * See if the next work item is here, if so, write it
 326                  */
 327                 work = NULL;
 328                 flist_for_each(entry, &rt->done_list) {
 329                         next = flist_entry(entry, struct work_item, list);
 330                         if (next->seq == rt->write_seq) {
 331                                 work = next;
 332                                 flist_del(&work->list);
 333                                 break;
 334                         }
 335                 }
 336
 337                 pthread_mutex_unlock(&rt->thread.lock);
 338
 339                 if (work) {
 340 write_it:
 341                         write_work(work);
 342                         __sync_fetch_and_add(&rt->write_seq, 1);
 343                 }
 344         }
 345 }
 346
 347 static void *reader_one_off(void *data)
 348 {
 349         reader_work(data);
 350         return NULL;
 351 }
 352
 353 static void *reader_fn(void *data)
 354 {
 355         struct reader_thread *rt = data;
 356         struct work_item *work;
 357
 358         while (!rt->thread.exit || !flist_empty(&rt->list)) {
 359                 work = NULL;
 360                 pthread_mutex_lock(&rt->thread.lock);
 361                 if (!flist_empty(&rt->list)) {
 362                         work = flist_first_entry(&rt->list, struct work_item, list);
 363                         flist_del_init(&work->list);
 364                 } else
 365                         pthread_cond_wait(&rt->thread.cond, &rt->thread.lock);
 366                 pthread_mutex_unlock(&rt->thread.lock);
 367
 368                 if (work) {
 369                         __sync_fetch_and_add(&rt->busy, 1);
 370                         reader_work(work);
 371                         __sync_fetch_and_sub(&rt->busy, 1);
 372                 }
 373         }
 374
 375         thread_exiting(&rt->thread);
 376         return NULL;
 377 }
 378
 379 static void queue_work(struct reader_thread *rt, struct work_item *work)
 380 {
 381         if (!rt->started) {
 382                 pthread_mutex_lock(&rt->thread.lock);
 383                 flist_add_tail(&work->list, &rt->list);
 384                 pthread_mutex_unlock(&rt->thread.lock);
 385
 386                 rt->started = 1;
 387                 pthread_create(&rt->thread.thread, NULL, reader_fn, rt);
 388         } else if (!rt->busy && !pthread_mutex_trylock(&rt->thread.lock)) {
 389                 flist_add_tail(&work->list, &rt->list);
 390                 pthread_mutex_unlock(&rt->thread.lock);
 391
 392                 pthread_cond_signal(&rt->thread.cond);
 393         } else {
 394                 int ret = pthread_create(&work->thread, NULL, reader_one_off, work);
 395                 if (ret)
 396                         fprintf(stderr, "pthread_create=%d\n", ret);
 397                 else
 398                         pthread_detach(work->thread);
 399         }
 400 }
 401
 402 static unsigned int calc_percentiles(unsigned int *io_u_plat, unsigned long nr,
 403                                      unsigned int **output)
 404 {
 405         unsigned long sum = 0;
 406         unsigned int len, i, j = 0;
 407         unsigned int oval_len = 0;
 408         unsigned int *ovals = NULL;
 409         int is_last;
 410
 411         len = 0;
 412         while (len < PLAT_LIST_MAX && plist[len] != 0.0)
 413                 len++;
 414
 415         if (!len)
 416                 return 0;
 417
 418         /*
 419          * Calculate bucket values, note down max and min values
 420          */
 421         is_last = 0;
 422         for (i = 0; i < PLAT_NR && !is_last; i++) {
 423                 sum += io_u_plat[i];
 424                 while (sum >= (plist[j] / 100.0 * nr)) {
 425                         assert(plist[j] <= 100.0);
 426
 427                         if (j == oval_len) {
 428                                 oval_len += 100;
 429                                 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
 430                         }
 431
 432                         ovals[j] = plat_idx_to_val(i);
 433                         is_last = (j == len - 1);
 434                         if (is_last)
 435                                 break;
 436
 437                         j++;
 438                 }
 439         }
 440
 441         *output = ovals;
 442         return len;
 443 }
 444
 445 static void show_latencies(struct stats *s, const char *msg)
 446 {
 447         unsigned int *ovals = NULL;
 448         unsigned int len, i;
 449
 450         len = calc_percentiles(s->plat, s->nr_samples, &ovals);
 451         if (len) {
 452                 fprintf(stderr, "Latency percentiles (usec) (%s)\n", msg);
 453                 for (i = 0; i < len; i++)
 454                         fprintf(stderr, "\t%2.4fth: %u\n", plist[i], ovals[i]);
 455         }
 456
 457         if (ovals)
 458                 free(ovals);
 459
 460         fprintf(stderr, "\tOver=%u, min=%u, max=%u\n", s->over, s->min, s->max);
 461 }
 462
 463 static void init_thread(struct thread_data *thread)
 464 {
 465         pthread_condattr_t cattr;
 466         int ret;
 467
 468         ret = pthread_condattr_init(&cattr);
 469         assert(ret == 0);
 470 #ifdef CONFIG_PTHREAD_CONDATTR_SETCLOCK
 471         ret = pthread_condattr_setclock(&cattr, CLOCK_MONOTONIC);
 472         assert(ret == 0);
 473 #endif
 474         pthread_cond_init(&thread->cond, &cattr);
 475         pthread_cond_init(&thread->done_cond, &cattr);
 476         pthread_mutex_init(&thread->lock, NULL);
 477         pthread_mutex_init(&thread->done_lock, NULL);
 478         thread->exit = 0;
 479 }
 480
 481 static void exit_thread(struct thread_data *thread,
 482                         void fn(struct writer_thread *),
 483                         struct writer_thread *wt)
 484 {
 485         __sync_fetch_and_add(&thread->exit, 1);
 486         pthread_cond_signal(&thread->cond);
 487
 488         while (!thread->done) {
 489                 pthread_mutex_lock(&thread->done_lock);
 490
 491                 if (fn) {
 492                         struct timespec ts;
 493
 494 #ifdef CONFIG_PTHREAD_CONDATTR_SETCLOCK
 495                         clock_gettime(CLOCK_MONOTONIC, &ts);
 496 #else
 497                         clock_gettime(CLOCK_REALTIME, &ts);
 498 #endif
 499                         ts.tv_sec++;
 500
 501                         pthread_cond_timedwait(&thread->done_cond, &thread->done_lock, &ts);
 502                         fn(wt);
 503                 } else
 504                         pthread_cond_wait(&thread->done_cond, &thread->done_lock);
 505
 506                 pthread_mutex_unlock(&thread->done_lock);
 507         }
 508 }
 509
 510 static int usage(char *argv[])
 511 {
 512         fprintf(stderr, "%s: [-b blocksize] [-t max usec] [-w separate writer] -f file\n", argv[0]);
 513         return 1;
 514 }
 515
 516 static int parse_options(int argc, char *argv[])
 517 {
 518         int c;
 519
 520         while ((c = getopt(argc, argv, "f:b:t:w:")) != -1) {
 521                 switch (c) {
 522                 case 'f':
 523                         if (file)
 524                                 return usage(argv);
 525                         file = strdup(optarg);
 526                         break;
 527                 case 'b':
 528                         bs = atoi(optarg);
 529                         break;
 530                 case 't':
 531                         max_us = atoi(optarg);
 532                         break;
 533                 case 'w':
 534                         separate_writer = atoi(optarg);
 535                         if (!separate_writer)
 536                                 fprintf(stderr, "inline writing is broken\n");
 537                         break;
 538                 case '?':
 539                 default:
 540                         return usage(argv);
 541                 }
 542         }
 543
 544         if (!file)
 545                 return usage(argv);
 546
 547         return 0;
 548 }
 549
 550 static void prune_done_entries(struct writer_thread *wt)
 551 {
 552         FLIST_HEAD(list);
 553
 554         if (flist_empty(&wt->done_list))
 555                 return;
 556
 557         if (pthread_mutex_trylock(&wt->thread.lock))
 558                 return;
 559
 560         if (!flist_empty(&wt->done_list))
 561                 flist_splice_init(&wt->done_list, &list);
 562         pthread_mutex_unlock(&wt->thread.lock);
 563
 564         while (!flist_empty(&list)) {
 565                 struct work_item *work;
 566
 567                 work = flist_first_entry(&list, struct work_item, list);
 568                 flist_del(&work->list);
 569
 570                 pthread_cond_destroy(&work->cond);
 571                 pthread_mutex_destroy(&work->lock);
 572                 free(work->buf);
 573                 free(work);
 574         }
 575 }
 576
 577 int main(int argc, char *argv[])
 578 {
 579         pthread_condattr_t cattr;
 580         struct timespec s, re, we;
 581         struct reader_thread *rt;
 582         struct writer_thread *wt;
 583         unsigned long rate;
 584         struct stat sb;
 585         size_t bytes;
 586         off_t off;
 587         int fd, seq;
 588         int ret;
 589
 590         if (parse_options(argc, argv))
 591                 return 1;
 592
 593         fd = open(file, O_RDONLY);
 594         if (fd < 0) {
 595                 perror("open");
 596                 return 2;
 597         }
 598
 599         if (fstat(fd, &sb) < 0) {
 600                 perror("stat");
 601                 return 3;
 602         }
 603
 604         wt = &writer_thread;
 605         init_thread(&wt->thread);
 606         INIT_FLIST_HEAD(&wt->list);
 607         INIT_FLIST_HEAD(&wt->done_list);
 608         wt->s.max = 0;
 609         wt->s.min = -1U;
 610         pthread_create(&wt->thread.thread, NULL, writer_fn, wt);
 611
 612         rt = &reader_thread;
 613         init_thread(&rt->thread);
 614         INIT_FLIST_HEAD(&rt->list);
 615         INIT_FLIST_HEAD(&rt->done_list);
 616         rt->s.max = 0;
 617         rt->s.min = -1U;
 618         rt->write_seq = 1;
 619
 620         off = 0;
 621         seq = 0;
 622         bytes = 0;
 623
 624         ret = pthread_condattr_init(&cattr);
 625         assert(ret == 0);
 626 #ifdef CONFIG_PTHREAD_CONDATTR_SETCLOCK
 627         ret = pthread_condattr_setclock(&cattr, CLOCK_MONOTONIC);
 628         assert(ret == 0);
 629 #endif
 630
 631         clock_gettime(CLOCK_MONOTONIC, &s);
 632
 633         while (sb.st_size) {
 634                 struct work_item *work;
 635                 size_t this_len;
 636                 struct timespec ts;
 637
 638                 prune_done_entries(wt);
 639
 640                 this_len = sb.st_size;
 641                 if (this_len > bs)
 642                         this_len = bs;
 643
 644                 work = calloc(1, sizeof(*work));
 645                 work->buf = malloc(this_len);
 646                 work->buf_size = this_len;
 647                 work->off = off;
 648                 work->fd = fd;
 649                 work->seq = ++seq;
 650                 work->writer = wt;
 651                 work->reader = rt;
 652                 pthread_cond_init(&work->cond, &cattr);
 653                 pthread_mutex_init(&work->lock, NULL);
 654
 655                 queue_work(rt, work);
 656
 657 #ifdef CONFIG_PTHREAD_CONDATTR_SETCLOCK
 658                 clock_gettime(CLOCK_MONOTONIC, &ts);
 659 #else
 660                 clock_gettime(CLOCK_REALTIME, &ts);
 661 #endif
 662                 ts.tv_nsec += max_us * 1000ULL;
 663                 if (ts.tv_nsec >= 1000000000ULL) {
 664                         ts.tv_nsec -= 1000000000ULL;
 665                         ts.tv_sec++;
 666                 }
 667
 668                 pthread_mutex_lock(&work->lock);
 669                 pthread_cond_timedwait(&work->cond, &work->lock, &ts);
 670                 pthread_mutex_unlock(&work->lock);
 671
 672                 off += this_len;
 673                 sb.st_size -= this_len;
 674                 bytes += this_len;
 675         }
 676
 677         exit_thread(&rt->thread, NULL, NULL);
 678         clock_gettime(CLOCK_MONOTONIC, &re);
 679
 680         exit_thread(&wt->thread, prune_done_entries, wt);
 681         clock_gettime(CLOCK_MONOTONIC, &we);
 682
 683         show_latencies(&rt->s, "READERS");
 684         show_latencies(&wt->s, "WRITERS");
 685
 686         bytes /= 1024;
 687         rate = (bytes * 1000UL * 1000UL) / utime_since(&s, &re);
 688         fprintf(stderr, "Read rate (KiB/sec) : %lu\n", rate);
 689         rate = (bytes * 1000UL * 1000UL) / utime_since(&s, &we);
 690         fprintf(stderr, "Write rate (KiB/sec): %lu\n", rate);
 691
 692         close(fd);
 693         return 0;
 694 }