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 #include <stdio.h>
  24 #include <stdlib.h>
  25 #include <unistd.h>
  26 #include <fcntl.h>
  27 #include <sys/time.h>
  28 #include <sys/types.h>
  29 #include <sys/stat.h>
  30 #include <inttypes.h>
  31 #include <string.h>
  32 #include <pthread.h>
  33 #include <errno.h>
  34 #include <assert.h>
  35
  36 #include "../flist.h"
  37
  38 static int bs = 4096;
  39 static int max_us = 10000;
  40 static char *file;
  41 static int separate_writer = 1;
  42
  43 #define PLAT_BITS       8
  44 #define PLAT_VAL        (1 << PLAT_BITS)
  45 #define PLAT_GROUP_NR   19
  46 #define PLAT_NR         (PLAT_GROUP_NR * PLAT_VAL)
  47 #define PLAT_LIST_MAX   20
  48
  49 struct stats {
  50         unsigned int plat[PLAT_NR];
  51         unsigned int nr_samples;
  52         unsigned int max;
  53         unsigned int min;
  54         unsigned int over;
  55 };
  56
  57 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, };
  58
  59 struct thread_data {
  60         int exit;
  61         int done;
  62         pthread_mutex_t lock;
  63         pthread_cond_t cond;
  64         pthread_mutex_t done_lock;
  65         pthread_cond_t done_cond;
  66         pthread_t thread;
  67 };
  68
  69 struct writer_thread {
  70         struct flist_head list;
  71         struct flist_head done_list;
  72         struct stats s;
  73         struct thread_data thread;
  74 };
  75
  76 struct reader_thread {
  77         struct flist_head list;
  78         struct flist_head done_list;
  79         int started;
  80         int busy;
  81         int write_seq;
  82         struct stats s;
  83         struct thread_data thread;
  84 };
  85
  86 struct work_item {
  87         struct flist_head list;
  88         void *buf;
  89         size_t buf_size;
  90         off_t off;
  91         int fd;
  92         int seq;
  93         struct writer_thread *writer;
  94         struct reader_thread *reader;
  95         pthread_mutex_t lock;
  96         pthread_cond_t cond;
  97         pthread_t thread;
  98 };
  99
 100 static struct reader_thread reader_thread;
 101 static struct writer_thread writer_thread;
 102
 103 uint64_t utime_since(const struct timeval *s, const struct timeval *e)
 104 {
 105         long sec, usec;
 106         uint64_t ret;
 107
 108         sec = e->tv_sec - s->tv_sec;
 109         usec = e->tv_usec - s->tv_usec;
 110         if (sec > 0 && usec < 0) {
 111                 sec--;
 112                 usec += 1000000;
 113         }
 114
 115         if (sec < 0 || (sec == 0 && usec < 0))
 116                 return 0;
 117
 118         ret = sec * 1000000ULL + usec;
 119
 120         return ret;
 121 }
 122
 123 static struct work_item *find_seq(struct writer_thread *w, unsigned int seq)
 124 {
 125         struct work_item *work;
 126         struct flist_head *entry;
 127
 128         if (flist_empty(&w->list))
 129                 return NULL;
 130
 131         flist_for_each(entry, &w->list) {
 132                 work = flist_entry(entry, struct work_item, list);
 133                 if (work->seq == seq)
 134                         return work;
 135         }
 136
 137         return NULL;
 138 }
 139
 140 static unsigned int plat_val_to_idx(unsigned int val)
 141 {
 142         unsigned int msb, error_bits, base, offset;
 143
 144         /* Find MSB starting from bit 0 */
 145         if (val == 0)
 146                 msb = 0;
 147         else
 148                 msb = sizeof(val)*8 - __builtin_clz(val) - 1;
 149
 150         /*
 151          * MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 152          * all bits of the sample as index
 153          */
 154         if (msb <= PLAT_BITS)
 155                 return val;
 156
 157         /* Compute the number of error bits to discard*/
 158         error_bits = msb - PLAT_BITS;
 159
 160         /* Compute the number of buckets before the group */
 161         base = (error_bits + 1) << PLAT_BITS;
 162
 163         /*
 164          * Discard the error bits and apply the mask to find the
 165          * index for the buckets in the group
 166          */
 167         offset = (PLAT_VAL - 1) & (val >> error_bits);
 168
 169         /* Make sure the index does not exceed (array size - 1) */
 170         return (base + offset) < (PLAT_NR - 1) ?
 171                 (base + offset) : (PLAT_NR - 1);
 172 }
 173
 174 /*
 175  * Convert the given index of the bucket array to the value
 176  * represented by the bucket
 177  */
 178 static unsigned int plat_idx_to_val(unsigned int idx)
 179 {
 180         unsigned int error_bits, k, base;
 181
 182         assert(idx < PLAT_NR);
 183
 184         /* MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 185          * all bits of the sample as index */
 186         if (idx < (PLAT_VAL << 1))
 187                 return idx;
 188
 189         /* Find the group and compute the minimum value of that group */
 190         error_bits = (idx >> PLAT_BITS) - 1;
 191         base = 1 << (error_bits + PLAT_BITS);
 192
 193         /* Find its bucket number of the group */
 194         k = idx % PLAT_VAL;
 195
 196         /* Return the mean of the range of the bucket */
 197         return base + ((k + 0.5) * (1 << error_bits));
 198 }
 199
 200 static void add_lat(struct stats *s, unsigned int us, const char *name)
 201 {
 202         int lat_index = 0;
 203
 204         if (us > s->max)
 205                 s->max = us;
 206         if (us < s->min)
 207                 s->min = us;
 208
 209         if (us > max_us) {
 210                 fprintf(stderr, "%s latency=%u usec\n", name, us);
 211                 s->over++;
 212         }
 213
 214         lat_index = plat_val_to_idx(us);
 215         __sync_fetch_and_add(&s->plat[lat_index], 1);
 216         __sync_fetch_and_add(&s->nr_samples, 1);
 217 }
 218
 219 static int write_work(struct work_item *work)
 220 {
 221         struct timeval s, e;
 222         ssize_t ret;
 223
 224         gettimeofday(&s, NULL);
 225         ret = write(STDOUT_FILENO, work->buf, work->buf_size);
 226         gettimeofday(&e, NULL);
 227         assert(ret == work->buf_size);
 228
 229         add_lat(&work->writer->s, utime_since(&s, &e), "write");
 230         return work->seq + 1;
 231 }
 232
 233 static void *writer_fn(void *data)
 234 {
 235         struct writer_thread *wt = data;
 236         struct work_item *work;
 237         unsigned int seq = 1;
 238
 239         work = NULL;
 240         while (!wt->thread.exit || !flist_empty(&wt->list)) {
 241                 pthread_mutex_lock(&wt->thread.lock);
 242
 243                 if (work) {
 244                         flist_add_tail(&work->list, &wt->done_list);
 245                         work = NULL;
 246                 }
 247
 248                 work = find_seq(wt, seq);
 249                 if (work)
 250                         flist_del_init(&work->list);
 251                 else
 252                         pthread_cond_wait(&wt->thread.cond, &wt->thread.lock);
 253
 254                 pthread_mutex_unlock(&wt->thread.lock);
 255
 256                 if (work)
 257                         seq = write_work(work);
 258         }
 259
 260         wt->thread.done = 1;
 261         pthread_cond_signal(&wt->thread.done_cond);
 262         return NULL;
 263 }
 264
 265 static void reader_work(struct work_item *work)
 266 {
 267         struct timeval s, e;
 268         ssize_t ret;
 269         size_t left;
 270         void *buf;
 271         off_t off;
 272
 273         gettimeofday(&s, NULL);
 274
 275         left = work->buf_size;
 276         buf = work->buf;
 277         off = work->off;
 278         while (left) {
 279                 ret = pread(work->fd, buf, left, off);
 280                 if (!ret) {
 281                         fprintf(stderr, "zero read\n");
 282                         break;
 283                 } else if (ret < 0) {
 284                         fprintf(stderr, "errno=%d\n", errno);
 285                         break;
 286                 }
 287                 left -= ret;
 288                 off += ret;
 289                 buf += ret;
 290         }
 291
 292         gettimeofday(&e, NULL);
 293
 294         add_lat(&work->reader->s, utime_since(&s, &e), "read");
 295
 296         pthread_cond_signal(&work->cond);
 297
 298         if (separate_writer) {
 299                 pthread_mutex_lock(&work->writer->thread.lock);
 300                 flist_add_tail(&work->list, &work->writer->list);
 301                 pthread_mutex_unlock(&work->writer->thread.lock);
 302                 pthread_cond_signal(&work->writer->thread.cond);
 303         } else {
 304                 struct reader_thread *rt = work->reader;
 305                 struct work_item *next = NULL;
 306                 struct flist_head *entry;
 307
 308                 /*
 309                  * Write current work if it matches in sequence.
 310                  */
 311                 if (work->seq == rt->write_seq)
 312                         goto write_it;
 313
 314                 pthread_mutex_lock(&rt->thread.lock);
 315
 316                 flist_add_tail(&work->list, &rt->done_list);
 317
 318                 /*
 319                  * See if the next work item is here, if so, write it
 320                  */
 321                 work = NULL;
 322                 flist_for_each(entry, &rt->done_list) {
 323                         next = flist_entry(entry, struct work_item, list);
 324                         if (next->seq == rt->write_seq) {
 325                                 work = next;
 326                                 flist_del(&work->list);
 327                                 break;
 328                         }
 329                 }
 330
 331                 pthread_mutex_unlock(&rt->thread.lock);
 332
 333                 if (work) {
 334 write_it:
 335                         write_work(work);
 336                         __sync_fetch_and_add(&rt->write_seq, 1);
 337                 }
 338         }
 339 }
 340
 341 static void *reader_one_off(void *data)
 342 {
 343         reader_work(data);
 344         return NULL;
 345 }
 346
 347 static void *reader_fn(void *data)
 348 {
 349         struct reader_thread *rt = data;
 350         struct work_item *work;
 351
 352         while (!rt->thread.exit || !flist_empty(&rt->list)) {
 353                 work = NULL;
 354                 pthread_mutex_lock(&rt->thread.lock);
 355                 if (!flist_empty(&rt->list)) {
 356                         work = flist_first_entry(&rt->list, struct work_item, list);
 357                         flist_del_init(&work->list);
 358                 } else
 359                         pthread_cond_wait(&rt->thread.cond, &rt->thread.lock);
 360                 pthread_mutex_unlock(&rt->thread.lock);
 361
 362                 if (work) {
 363                         rt->busy = 1;
 364                         reader_work(work);
 365                         rt->busy = 0;
 366                 }
 367         }
 368
 369         rt->thread.done = 1;
 370         pthread_cond_signal(&rt->thread.done_cond);
 371         return NULL;
 372 }
 373
 374 static void queue_work(struct reader_thread *rt, struct work_item *work)
 375 {
 376         if (!rt->started) {
 377                 pthread_mutex_lock(&rt->thread.lock);
 378                 flist_add_tail(&work->list, &rt->list);
 379                 pthread_mutex_unlock(&rt->thread.lock);
 380
 381                 rt->started = 1;
 382                 pthread_create(&rt->thread.thread, NULL, reader_fn, rt);
 383         } else if (!rt->busy && !pthread_mutex_trylock(&rt->thread.lock)) {
 384                 flist_add_tail(&work->list, &rt->list);
 385                 pthread_mutex_unlock(&rt->thread.lock);
 386
 387                 pthread_cond_signal(&rt->thread.cond);
 388         } else {
 389                 int ret = pthread_create(&work->thread, NULL, reader_one_off, work);
 390                 if (ret)
 391                         fprintf(stderr, "pthread_create=%d\n", ret);
 392                 else
 393                         pthread_detach(work->thread);
 394         }
 395 }
 396
 397 static unsigned int calc_percentiles(unsigned int *io_u_plat, unsigned long nr,
 398                                      unsigned int **output)
 399 {
 400         unsigned long sum = 0;
 401         unsigned int len, i, j = 0;
 402         unsigned int oval_len = 0;
 403         unsigned int *ovals = NULL;
 404         int is_last;
 405
 406         len = 0;
 407         while (len < PLAT_LIST_MAX && plist[len] != 0.0)
 408                 len++;
 409
 410         if (!len)
 411                 return 0;
 412
 413         /*
 414          * Calculate bucket values, note down max and min values
 415          */
 416         is_last = 0;
 417         for (i = 0; i < PLAT_NR && !is_last; i++) {
 418                 sum += io_u_plat[i];
 419                 while (sum >= (plist[j] / 100.0 * nr)) {
 420                         assert(plist[j] <= 100.0);
 421
 422                         if (j == oval_len) {
 423                                 oval_len += 100;
 424                                 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
 425                         }
 426
 427                         ovals[j] = plat_idx_to_val(i);
 428                         is_last = (j == len - 1);
 429                         if (is_last)
 430                                 break;
 431
 432                         j++;
 433                 }
 434         }
 435
 436         *output = ovals;
 437         return len;
 438 }
 439
 440 static void show_latencies(struct stats *s, const char *msg)
 441 {
 442         unsigned int *ovals = NULL;
 443         unsigned int len, i;
 444
 445         len = calc_percentiles(s->plat, s->nr_samples, &ovals);
 446         if (len) {
 447                 fprintf(stderr, "Latency percentiles (usec) (%s)\n", msg);
 448                 for (i = 0; i < len; i++)
 449                         fprintf(stderr, "\t%2.4fth: %u\n", plist[i], ovals[i]);
 450         }
 451
 452         if (ovals)
 453                 free(ovals);
 454
 455         fprintf(stderr, "\tOver=%u, min=%u, max=%u\n", s->over, s->min, s->max);
 456 }
 457
 458 static void init_thread(struct thread_data *thread)
 459 {
 460         pthread_cond_init(&thread->cond, NULL);
 461         pthread_cond_init(&thread->done_cond, NULL);
 462         pthread_mutex_init(&thread->lock, NULL);
 463         pthread_mutex_init(&thread->done_lock, NULL);
 464         thread->exit = 0;
 465 }
 466
 467 static void exit_thread(struct thread_data *thread,
 468                         void fn(struct writer_thread *),
 469                         struct writer_thread *wt)
 470 {
 471         thread->exit = 1;
 472         pthread_cond_signal(&thread->cond);
 473
 474         while (!thread->done) {
 475                 pthread_mutex_lock(&thread->done_lock);
 476
 477                 if (fn) {
 478                         struct timespec t;
 479
 480                         clock_gettime(CLOCK_REALTIME, &t);
 481                         t.tv_sec++;
 482
 483
 484                         pthread_cond_timedwait(&thread->done_cond, &thread->done_lock, &t);
 485                         fn(wt);
 486                 } else
 487                         pthread_cond_wait(&thread->done_cond, &thread->done_lock);
 488
 489                 pthread_mutex_unlock(&thread->done_lock);
 490         }
 491 }
 492
 493 static int usage(char *argv[])
 494 {
 495         fprintf(stderr, "%s: [-b blocksize] [-t max usec] [-w separate writer] -f file\n", argv[0]);
 496         return 1;
 497 }
 498
 499 static int parse_options(int argc, char *argv[])
 500 {
 501         int c;
 502
 503         while ((c = getopt(argc, argv, "f:b:t:w:")) != -1) {
 504                 switch (c) {
 505                 case 'f':
 506                         file = strdup(optarg);
 507                         break;
 508                 case 'b':
 509                         bs = atoi(optarg);
 510                         break;
 511                 case 't':
 512                         max_us = atoi(optarg);
 513                         break;
 514                 case 'w':
 515                         separate_writer = atoi(optarg);
 516                         if (!separate_writer)
 517                                 fprintf(stderr, "inline writing is broken\n");
 518                         break;
 519                 case '?':
 520                 default:
 521                         return usage(argv);
 522                 }
 523         }
 524
 525         if (!file)
 526                 return usage(argv);
 527
 528         return 0;
 529 }
 530
 531 static void prune_done_entries(struct writer_thread *wt)
 532 {
 533         FLIST_HEAD(list);
 534
 535         if (flist_empty(&wt->done_list))
 536                 return;
 537
 538         if (pthread_mutex_trylock(&wt->thread.lock))
 539                 return;
 540
 541         if (!flist_empty(&wt->done_list))
 542                 flist_splice_init(&wt->done_list, &list);
 543         pthread_mutex_unlock(&wt->thread.lock);
 544
 545         while (!flist_empty(&list)) {
 546                 struct work_item *work;
 547
 548                 work = flist_first_entry(&list, struct work_item, list);
 549                 flist_del(&work->list);
 550
 551                 pthread_cond_destroy(&work->cond);
 552                 pthread_mutex_destroy(&work->lock);
 553                 free(work->buf);
 554                 free(work);
 555         }
 556 }
 557
 558 int main(int argc, char *argv[])
 559 {
 560         struct timeval s, re, we;
 561         struct reader_thread *rt;
 562         struct writer_thread *wt;
 563         unsigned long rate;
 564         struct stat sb;
 565         size_t bytes;
 566         off_t off;
 567         int fd, seq;
 568
 569         if (parse_options(argc, argv))
 570                 return 1;
 571
 572         fd = open(file, O_RDONLY);
 573         if (fd < 0) {
 574                 perror("open");
 575                 return 2;
 576         }
 577
 578         if (fstat(fd, &sb) < 0) {
 579                 perror("stat");
 580                 return 3;
 581         }
 582
 583         wt = &writer_thread;
 584         init_thread(&wt->thread);
 585         INIT_FLIST_HEAD(&wt->list);
 586         INIT_FLIST_HEAD(&wt->done_list);
 587         wt->s.max = 0;
 588         wt->s.min = -1U;
 589         pthread_create(&wt->thread.thread, NULL, writer_fn, wt);
 590
 591         rt = &reader_thread;
 592         init_thread(&rt->thread);
 593         INIT_FLIST_HEAD(&rt->list);
 594         INIT_FLIST_HEAD(&rt->done_list);
 595         rt->s.max = 0;
 596         rt->s.min = -1U;
 597         rt->write_seq = 1;
 598
 599         off = 0;
 600         seq = 0;
 601         bytes = 0;
 602
 603         gettimeofday(&s, NULL);
 604
 605         while (sb.st_size) {
 606                 struct work_item *work;
 607                 size_t this_len;
 608                 struct timespec t;
 609
 610                 prune_done_entries(wt);
 611
 612                 this_len = sb.st_size;
 613                 if (this_len > bs)
 614                         this_len = bs;
 615
 616                 work = calloc(1, sizeof(*work));
 617                 work->buf = malloc(this_len);
 618                 work->buf_size = this_len;
 619                 work->off = off;
 620                 work->fd = fd;
 621                 work->seq = ++seq;
 622                 work->writer = wt;
 623                 work->reader = rt;
 624                 pthread_cond_init(&work->cond, NULL);
 625                 pthread_mutex_init(&work->lock, NULL);
 626
 627                 queue_work(rt, work);
 628
 629                 clock_gettime(CLOCK_REALTIME, &t);
 630                 t.tv_nsec += max_us * 1000ULL;
 631                 if (t.tv_nsec >= 1000000000ULL) {
 632                         t.tv_nsec -= 1000000000ULL;
 633                         t.tv_sec++;
 634                 }
 635
 636                 pthread_mutex_lock(&work->lock);
 637                 pthread_cond_timedwait(&work->cond, &work->lock, &t);
 638                 pthread_mutex_unlock(&work->lock);
 639
 640                 off += this_len;
 641                 sb.st_size -= this_len;
 642                 bytes += this_len;
 643         }
 644
 645         exit_thread(&rt->thread, NULL, NULL);
 646         gettimeofday(&re, NULL);
 647
 648         exit_thread(&wt->thread, prune_done_entries, wt);
 649         gettimeofday(&we, NULL);
 650
 651         show_latencies(&rt->s, "READERS");
 652         show_latencies(&wt->s, "WRITERS");
 653
 654         bytes /= 1024;
 655         rate = (bytes * 1000UL * 1000UL) / utime_since(&s, &re);
 656         fprintf(stderr, "Read rate (KB/sec) : %lu\n", rate);
 657         rate = (bytes * 1000UL * 1000UL) / utime_since(&s, &we);
 658         fprintf(stderr, "Write rate (KB/sec): %lu\n", rate);
 659
 660         close(fd);
 661         return 0;
 662 }