t/io_uring.c

   1 #include <stdio.h>
   2 #include <errno.h>
   3 #include <assert.h>
   4 #include <stdlib.h>
   5 #include <stddef.h>
   6 #include <signal.h>
   7 #include <inttypes.h>
   8 #include <math.h>
   9
  10 #ifdef CONFIG_LIBAIO
  11 #include <libaio.h>
  12 #endif
  13
  14 #include <sys/types.h>
  15 #include <sys/stat.h>
  16 #include <sys/ioctl.h>
  17 #include <sys/syscall.h>
  18 #include <sys/resource.h>
  19 #include <sys/mman.h>
  20 #include <sys/uio.h>
  21 #include <linux/fs.h>
  22 #include <fcntl.h>
  23 #include <unistd.h>
  24 #include <string.h>
  25 #include <pthread.h>
  26 #include <sched.h>
  27
  28 #include "../arch/arch.h"
  29 #include "../lib/types.h"
  30 #include "../lib/roundup.h"
  31 #include "../lib/rand.h"
  32 #include "../minmax.h"
  33 #include "../os/linux/io_uring.h"
  34
  35 struct io_sq_ring {
  36         unsigned *head;
  37         unsigned *tail;
  38         unsigned *ring_mask;
  39         unsigned *ring_entries;
  40         unsigned *flags;
  41         unsigned *array;
  42 };
  43
  44 struct io_cq_ring {
  45         unsigned *head;
  46         unsigned *tail;
  47         unsigned *ring_mask;
  48         unsigned *ring_entries;
  49         struct io_uring_cqe *cqes;
  50 };
  51
  52 #define DEPTH                   128
  53 #define BATCH_SUBMIT            32
  54 #define BATCH_COMPLETE          32
  55 #define BS                      4096
  56
  57 #define MAX_FDS                 16
  58
  59 static unsigned sq_ring_mask, cq_ring_mask;
  60
  61 struct file {
  62         unsigned long max_blocks;
  63         unsigned long max_size;
  64         unsigned long cur_off;
  65         unsigned pending_ios;
  66         int real_fd;
  67         int fixed_fd;
  68         int fileno;
  69 };
  70
  71 #define PLAT_BITS               6
  72 #define PLAT_VAL                (1 << PLAT_BITS)
  73 #define PLAT_GROUP_NR           29
  74 #define PLAT_NR                 (PLAT_GROUP_NR * PLAT_VAL)
  75
  76 struct submitter {
  77         pthread_t thread;
  78         int ring_fd;
  79         int index;
  80         struct io_sq_ring sq_ring;
  81         struct io_uring_sqe *sqes;
  82         struct io_cq_ring cq_ring;
  83         int inflight;
  84         int tid;
  85         unsigned long reaps;
  86         unsigned long done;
  87         unsigned long calls;
  88         volatile int finish;
  89
  90         __s32 *fds;
  91
  92         struct taus258_state rand_state;
  93
  94         unsigned long *clock_batch;
  95         int clock_index;
  96         unsigned long *plat;
  97
  98 #ifdef CONFIG_LIBAIO
  99         io_context_t aio_ctx;
 100 #endif
 101
 102         struct file files[MAX_FDS];
 103         unsigned nr_files;
 104         unsigned cur_file;
 105         struct iovec iovecs[];
 106 };
 107
 108 static struct submitter *submitter;
 109 static volatile int finish;
 110 static int stats_running;
 111 static unsigned long max_iops;
 112
 113 static int depth = DEPTH;
 114 static int batch_submit = BATCH_SUBMIT;
 115 static int batch_complete = BATCH_COMPLETE;
 116 static int bs = BS;
 117 static int polled = 1;          /* use IO polling */
 118 static int fixedbufs = 1;       /* use fixed user buffers */
 119 static int dma_map;             /* pre-map DMA buffers */
 120 static int register_files = 1;  /* use fixed files */
 121 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
 122 static int sq_thread_poll = 0;  /* use kernel submission/poller thread */
 123 static int sq_thread_cpu = -1;  /* pin above thread to this CPU */
 124 static int do_nop = 0;          /* no-op SQ ring commands */
 125 static int nthreads = 1;
 126 static int stats = 0;           /* generate IO stats */
 127 static int aio = 0;             /* use libaio */
 128 static int runtime = 0;         /* runtime */
 129 static int random_io = 1;       /* random or sequential IO */
 130
 131 static unsigned long tsc_rate;
 132
 133 #define TSC_RATE_FILE   "tsc-rate"
 134
 135 static int vectored = 1;
 136
 137 static float plist[] = { 1.0, 5.0, 10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0,
 138                         80.0, 90.0, 95.0, 99.0, 99.5, 99.9, 99.95, 99.99 };
 139 static int plist_len = 17;
 140
 141 #ifndef IORING_REGISTER_MAP_BUFFERS
 142 #define IORING_REGISTER_MAP_BUFFERS     20
 143 struct io_uring_map_buffers {
 144         __s32   fd;
 145         __u32   buf_start;
 146         __u32   buf_end;
 147         __u32   flags;
 148         __u64   rsvd[2];
 149 };
 150 #endif
 151
 152 static unsigned long cycles_to_nsec(unsigned long cycles)
 153 {
 154         uint64_t val;
 155
 156         if (!tsc_rate)
 157                 return cycles;
 158
 159         val = cycles * 1000000000ULL;
 160         return val / tsc_rate;
 161 }
 162
 163 static unsigned long plat_idx_to_val(unsigned int idx)
 164 {
 165         unsigned int error_bits;
 166         unsigned long k, base;
 167
 168         assert(idx < PLAT_NR);
 169
 170         /* MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 171          * all bits of the sample as index */
 172         if (idx < (PLAT_VAL << 1))
 173                 return cycles_to_nsec(idx);
 174
 175         /* Find the group and compute the minimum value of that group */
 176         error_bits = (idx >> PLAT_BITS) - 1;
 177         base = ((unsigned long) 1) << (error_bits + PLAT_BITS);
 178
 179         /* Find its bucket number of the group */
 180         k = idx % PLAT_VAL;
 181
 182         /* Return the mean of the range of the bucket */
 183         return cycles_to_nsec(base + ((k + 0.5) * (1 << error_bits)));
 184 }
 185
 186 unsigned int calc_clat_percentiles(unsigned long *io_u_plat, unsigned long nr,
 187                                    unsigned long **output,
 188                                    unsigned long *maxv, unsigned long *minv)
 189 {
 190         unsigned long sum = 0;
 191         unsigned int len = plist_len, i, j = 0;
 192         unsigned long *ovals = NULL;
 193         bool is_last;
 194
 195         *minv = -1ULL;
 196         *maxv = 0;
 197
 198         ovals = malloc(len * sizeof(*ovals));
 199         if (!ovals)
 200                 return 0;
 201
 202         /*
 203          * Calculate bucket values, note down max and min values
 204          */
 205         is_last = false;
 206         for (i = 0; i < PLAT_NR && !is_last; i++) {
 207                 sum += io_u_plat[i];
 208                 while (sum >= ((long double) plist[j] / 100.0 * nr)) {
 209                         assert(plist[j] <= 100.0);
 210
 211                         ovals[j] = plat_idx_to_val(i);
 212                         if (ovals[j] < *minv)
 213                                 *minv = ovals[j];
 214                         if (ovals[j] > *maxv)
 215                                 *maxv = ovals[j];
 216
 217                         is_last = (j == len - 1) != 0;
 218                         if (is_last)
 219                                 break;
 220
 221                         j++;
 222                 }
 223         }
 224
 225         if (!is_last)
 226                 fprintf(stderr, "error calculating latency percentiles\n");
 227
 228         *output = ovals;
 229         return len;
 230 }
 231
 232 static void show_clat_percentiles(unsigned long *io_u_plat, unsigned long nr,
 233                                   unsigned int precision)
 234 {
 235         unsigned int divisor, len, i, j = 0;
 236         unsigned long minv, maxv;
 237         unsigned long *ovals;
 238         int per_line, scale_down, time_width;
 239         bool is_last;
 240         char fmt[32];
 241
 242         len = calc_clat_percentiles(io_u_plat, nr, &ovals, &maxv, &minv);
 243         if (!len || !ovals)
 244                 goto out;
 245
 246         if (!tsc_rate) {
 247                 scale_down = 0;
 248                 divisor = 1;
 249                 printf("    percentiles (tsc ticks):\n     |");
 250         } else if (minv > 2000 && maxv > 99999) {
 251                 scale_down = 1;
 252                 divisor = 1000;
 253                 printf("    percentiles (usec):\n     |");
 254         } else {
 255                 scale_down = 0;
 256                 divisor = 1;
 257                 printf("    percentiles (nsec):\n     |");
 258         }
 259
 260         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 261         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 262                         precision, time_width);
 263         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 264         per_line = (80 - 7) / (precision + 10 + time_width);
 265
 266         for (j = 0; j < len; j++) {
 267                 /* for formatting */
 268                 if (j != 0 && (j % per_line) == 0)
 269                         printf("     |");
 270
 271                 /* end of the list */
 272                 is_last = (j == len - 1) != 0;
 273
 274                 for (i = 0; i < scale_down; i++)
 275                         ovals[j] = (ovals[j] + 999) / 1000;
 276
 277                 printf(fmt, plist[j], ovals[j], is_last ? '\n' : ',');
 278
 279                 if (is_last)
 280                         break;
 281
 282                 if ((j % per_line) == per_line - 1)     /* for formatting */
 283                         printf("\n");
 284         }
 285
 286 out:
 287         free(ovals);
 288 }
 289
 290 static unsigned int plat_val_to_idx(unsigned long val)
 291 {
 292         unsigned int msb, error_bits, base, offset, idx;
 293
 294         /* Find MSB starting from bit 0 */
 295         if (val == 0)
 296                 msb = 0;
 297         else
 298                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
 299
 300         /*
 301          * MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 302          * all bits of the sample as index
 303          */
 304         if (msb <= PLAT_BITS)
 305                 return val;
 306
 307         /* Compute the number of error bits to discard*/
 308         error_bits = msb - PLAT_BITS;
 309
 310         /* Compute the number of buckets before the group */
 311         base = (error_bits + 1) << PLAT_BITS;
 312
 313         /*
 314          * Discard the error bits and apply the mask to find the
 315          * index for the buckets in the group
 316          */
 317         offset = (PLAT_VAL - 1) & (val >> error_bits);
 318
 319         /* Make sure the index does not exceed (array size - 1) */
 320         idx = (base + offset) < (PLAT_NR - 1) ?
 321                 (base + offset) : (PLAT_NR - 1);
 322
 323         return idx;
 324 }
 325
 326 static void add_stat(struct submitter *s, int clock_index, int nr)
 327 {
 328 #ifdef ARCH_HAVE_CPU_CLOCK
 329         unsigned long cycles;
 330         unsigned int pidx;
 331
 332         if (!s->finish && clock_index) {
 333                 cycles = get_cpu_clock();
 334                 cycles -= s->clock_batch[clock_index];
 335                 pidx = plat_val_to_idx(cycles);
 336                 s->plat[pidx] += nr;
 337         }
 338 #endif
 339 }
 340
 341 static int io_uring_map_buffers(struct submitter *s)
 342 {
 343         struct io_uring_map_buffers map = {
 344                 .fd             = s->files[0].real_fd,
 345                 .buf_end        = depth,
 346         };
 347
 348         if (do_nop)
 349                 return 0;
 350         if (s->nr_files > 1) {
 351                 fprintf(stderr, "Can't map buffers with multiple files\n");
 352                 return -1;
 353         }
 354
 355         return syscall(__NR_io_uring_register, s->ring_fd,
 356                         IORING_REGISTER_MAP_BUFFERS, &map, 1);
 357 }
 358
 359 static int io_uring_register_buffers(struct submitter *s)
 360 {
 361         if (do_nop)
 362                 return 0;
 363
 364         return syscall(__NR_io_uring_register, s->ring_fd,
 365                         IORING_REGISTER_BUFFERS, s->iovecs, depth);
 366 }
 367
 368 static int io_uring_register_files(struct submitter *s)
 369 {
 370         int i;
 371
 372         if (do_nop)
 373                 return 0;
 374
 375         s->fds = calloc(s->nr_files, sizeof(__s32));
 376         for (i = 0; i < s->nr_files; i++) {
 377                 s->fds[i] = s->files[i].real_fd;
 378                 s->files[i].fixed_fd = i;
 379         }
 380
 381         return syscall(__NR_io_uring_register, s->ring_fd,
 382                         IORING_REGISTER_FILES, s->fds, s->nr_files);
 383 }
 384
 385 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 386 {
 387         return syscall(__NR_io_uring_setup, entries, p);
 388 }
 389
 390 static void io_uring_probe(int fd)
 391 {
 392         struct io_uring_probe *p;
 393         int ret;
 394
 395         p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 396         if (!p)
 397                 return;
 398
 399         memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 400         ret = syscall(__NR_io_uring_register, fd, IORING_REGISTER_PROBE, p, 256);
 401         if (ret < 0)
 402                 goto out;
 403
 404         if (IORING_OP_READ > p->ops_len)
 405                 goto out;
 406
 407         if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED))
 408                 vectored = 0;
 409 out:
 410         free(p);
 411 }
 412
 413 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 414                           unsigned int min_complete, unsigned int flags)
 415 {
 416         return syscall(__NR_io_uring_enter, s->ring_fd, to_submit, min_complete,
 417                         flags, NULL, 0);
 418 }
 419
 420 #ifndef CONFIG_HAVE_GETTID
 421 static int gettid(void)
 422 {
 423         return syscall(__NR_gettid);
 424 }
 425 #endif
 426
 427 static unsigned file_depth(struct submitter *s)
 428 {
 429         return (depth + s->nr_files - 1) / s->nr_files;
 430 }
 431
 432 static void init_io(struct submitter *s, unsigned index)
 433 {
 434         struct io_uring_sqe *sqe = &s->sqes[index];
 435         unsigned long offset;
 436         struct file *f;
 437         long r;
 438
 439         if (do_nop) {
 440                 sqe->opcode = IORING_OP_NOP;
 441                 return;
 442         }
 443
 444         if (s->nr_files == 1) {
 445                 f = &s->files[0];
 446         } else {
 447                 f = &s->files[s->cur_file];
 448                 if (f->pending_ios >= file_depth(s)) {
 449                         s->cur_file++;
 450                         if (s->cur_file == s->nr_files)
 451                                 s->cur_file = 0;
 452                         f = &s->files[s->cur_file];
 453                 }
 454         }
 455         f->pending_ios++;
 456
 457         if (random_io) {
 458                 r = __rand64(&s->rand_state);
 459                 offset = (r % (f->max_blocks - 1)) * bs;
 460         } else {
 461                 offset = f->cur_off;
 462                 f->cur_off += bs;
 463                 if (f->cur_off + bs > f->max_size)
 464                         f->cur_off = 0;
 465         }
 466
 467         if (register_files) {
 468                 sqe->flags = IOSQE_FIXED_FILE;
 469                 sqe->fd = f->fixed_fd;
 470         } else {
 471                 sqe->flags = 0;
 472                 sqe->fd = f->real_fd;
 473         }
 474         if (fixedbufs) {
 475                 sqe->opcode = IORING_OP_READ_FIXED;
 476                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 477                 sqe->len = bs;
 478                 sqe->buf_index = index;
 479         } else if (!vectored) {
 480                 sqe->opcode = IORING_OP_READ;
 481                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 482                 sqe->len = bs;
 483                 sqe->buf_index = 0;
 484         } else {
 485                 sqe->opcode = IORING_OP_READV;
 486                 sqe->addr = (unsigned long) &s->iovecs[index];
 487                 sqe->len = 1;
 488                 sqe->buf_index = 0;
 489         }
 490         sqe->ioprio = 0;
 491         sqe->off = offset;
 492         sqe->user_data = (unsigned long) f->fileno;
 493         if (stats && stats_running)
 494                 sqe->user_data |= ((unsigned long)s->clock_index << 32);
 495 }
 496
 497 static int prep_more_ios_uring(struct submitter *s, int max_ios)
 498 {
 499         struct io_sq_ring *ring = &s->sq_ring;
 500         unsigned index, tail, next_tail, prepped = 0;
 501
 502         next_tail = tail = *ring->tail;
 503         do {
 504                 next_tail++;
 505                 if (next_tail == atomic_load_acquire(ring->head))
 506                         break;
 507
 508                 index = tail & sq_ring_mask;
 509                 init_io(s, index);
 510                 ring->array[index] = index;
 511                 prepped++;
 512                 tail = next_tail;
 513         } while (prepped < max_ios);
 514
 515         if (prepped)
 516                 atomic_store_release(ring->tail, tail);
 517         return prepped;
 518 }
 519
 520 static int get_file_size(struct file *f)
 521 {
 522         struct stat st;
 523
 524         if (fstat(f->real_fd, &st) < 0)
 525                 return -1;
 526         if (S_ISBLK(st.st_mode)) {
 527                 unsigned long long bytes;
 528
 529                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 530                         return -1;
 531
 532                 f->max_blocks = bytes / bs;
 533                 f->max_size = bytes;
 534                 return 0;
 535         } else if (S_ISREG(st.st_mode)) {
 536                 f->max_blocks = st.st_size / bs;
 537                 f->max_size = st.st_size;
 538                 return 0;
 539         }
 540
 541         return -1;
 542 }
 543
 544 static int reap_events_uring(struct submitter *s)
 545 {
 546         struct io_cq_ring *ring = &s->cq_ring;
 547         struct io_uring_cqe *cqe;
 548         unsigned head, reaped = 0;
 549         int last_idx = -1, stat_nr = 0;
 550
 551         head = *ring->head;
 552         do {
 553                 struct file *f;
 554
 555                 read_barrier();
 556                 if (head == atomic_load_acquire(ring->tail))
 557                         break;
 558                 cqe = &ring->cqes[head & cq_ring_mask];
 559                 if (!do_nop) {
 560                         int fileno = cqe->user_data & 0xffffffff;
 561
 562                         f = &s->files[fileno];
 563                         f->pending_ios--;
 564                         if (cqe->res != bs) {
 565                                 printf("io: unexpected ret=%d\n", cqe->res);
 566                                 if (polled && cqe->res == -EOPNOTSUPP)
 567                                         printf("Your filesystem/driver/kernel doesn't support polled IO\n");
 568                                 return -1;
 569                         }
 570                 }
 571                 if (stats) {
 572                         int clock_index = cqe->user_data >> 32;
 573
 574                         if (last_idx != clock_index) {
 575                                 if (last_idx != -1) {
 576                                         add_stat(s, last_idx, stat_nr);
 577                                         stat_nr = 0;
 578                                 }
 579                                 last_idx = clock_index;
 580                         }
 581                         stat_nr++;
 582                 }
 583                 reaped++;
 584                 head++;
 585         } while (1);
 586
 587         if (stat_nr)
 588                 add_stat(s, last_idx, stat_nr);
 589
 590         if (reaped) {
 591                 s->inflight -= reaped;
 592                 atomic_store_release(ring->head, head);
 593         }
 594         return reaped;
 595 }
 596
 597 static int submitter_init(struct submitter *s)
 598 {
 599         int i, nr_batch;
 600
 601         s->tid = gettid();
 602         printf("submitter=%d, tid=%d\n", s->index, s->tid);
 603
 604         __init_rand64(&s->rand_state, pthread_self());
 605         srand48(pthread_self());
 606
 607         for (i = 0; i < MAX_FDS; i++)
 608                 s->files[i].fileno = i;
 609
 610         if (stats) {
 611                 nr_batch = roundup_pow2(depth / batch_submit);
 612                 if (nr_batch < 2)
 613                         nr_batch = 2;
 614                 s->clock_batch = calloc(nr_batch, sizeof(unsigned long));
 615                 s->clock_index = 1;
 616
 617                 s->plat = calloc(PLAT_NR, sizeof(unsigned long));
 618         } else {
 619                 s->clock_batch = NULL;
 620                 s->plat = NULL;
 621                 nr_batch = 0;
 622         }
 623
 624         return nr_batch;
 625 }
 626
 627 #ifdef CONFIG_LIBAIO
 628 static int prep_more_ios_aio(struct submitter *s, int max_ios, struct iocb *iocbs)
 629 {
 630         unsigned long offset, data;
 631         struct file *f;
 632         unsigned index;
 633         long r;
 634
 635         index = 0;
 636         while (index < max_ios) {
 637                 struct iocb *iocb = &iocbs[index];
 638
 639                 if (s->nr_files == 1) {
 640                         f = &s->files[0];
 641                 } else {
 642                         f = &s->files[s->cur_file];
 643                         if (f->pending_ios >= file_depth(s)) {
 644                                 s->cur_file++;
 645                                 if (s->cur_file == s->nr_files)
 646                                         s->cur_file = 0;
 647                                 f = &s->files[s->cur_file];
 648                         }
 649                 }
 650                 f->pending_ios++;
 651
 652                 r = lrand48();
 653                 offset = (r % (f->max_blocks - 1)) * bs;
 654                 io_prep_pread(iocb, f->real_fd, s->iovecs[index].iov_base,
 655                                 s->iovecs[index].iov_len, offset);
 656
 657                 data = f->fileno;
 658                 if (stats && stats_running)
 659                         data |= ((unsigned long) s->clock_index << 32);
 660                 iocb->data = (void *) (uintptr_t) data;
 661                 index++;
 662         }
 663         return index;
 664 }
 665
 666 static int reap_events_aio(struct submitter *s, struct io_event *events, int evs)
 667 {
 668         int last_idx = -1, stat_nr = 0;
 669         int reaped = 0;
 670
 671         while (evs) {
 672                 unsigned long data = (uintptr_t) events[reaped].data;
 673                 struct file *f = &s->files[data & 0xffffffff];
 674
 675                 f->pending_ios--;
 676                 if (events[reaped].res != bs) {
 677                         printf("io: unexpected ret=%ld\n", events[reaped].res);
 678                         return -1;
 679                 }
 680                 if (stats) {
 681                         int clock_index = data >> 32;
 682
 683                         if (last_idx != clock_index) {
 684                                 if (last_idx != -1) {
 685                                         add_stat(s, last_idx, stat_nr);
 686                                         stat_nr = 0;
 687                                 }
 688                                 last_idx = clock_index;
 689                         }
 690                         stat_nr++;
 691                 }
 692                 reaped++;
 693                 evs--;
 694         }
 695
 696         if (stat_nr)
 697                 add_stat(s, last_idx, stat_nr);
 698
 699         s->inflight -= reaped;
 700         s->done += reaped;
 701         return reaped;
 702 }
 703
 704 static void *submitter_aio_fn(void *data)
 705 {
 706         struct submitter *s = data;
 707         int i, ret, prepped, nr_batch;
 708         struct iocb **iocbsptr;
 709         struct iocb *iocbs;
 710         struct io_event *events;
 711
 712         nr_batch = submitter_init(s);
 713
 714         iocbsptr = calloc(depth, sizeof(struct iocb *));
 715         iocbs = calloc(depth, sizeof(struct iocb));
 716         events = calloc(depth, sizeof(struct io_event));
 717
 718         for (i = 0; i < depth; i++)
 719                 iocbsptr[i] = &iocbs[i];
 720
 721         prepped = 0;
 722         do {
 723                 int to_wait, to_submit, to_prep;
 724
 725                 if (!prepped && s->inflight < depth) {
 726                         to_prep = min(depth - s->inflight, batch_submit);
 727                         prepped = prep_more_ios_aio(s, to_prep, iocbs);
 728 #ifdef ARCH_HAVE_CPU_CLOCK
 729                         if (prepped && stats) {
 730                                 s->clock_batch[s->clock_index] = get_cpu_clock();
 731                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
 732                         }
 733 #endif
 734                 }
 735                 s->inflight += prepped;
 736                 to_submit = prepped;
 737
 738                 if (to_submit && (s->inflight + to_submit <= depth))
 739                         to_wait = 0;
 740                 else
 741                         to_wait = min(s->inflight + to_submit, batch_complete);
 742
 743                 ret = io_submit(s->aio_ctx, to_submit, iocbsptr);
 744                 s->calls++;
 745                 if (ret < 0) {
 746                         perror("io_submit");
 747                         break;
 748                 } else if (ret != to_submit) {
 749                         printf("submitted %d, wanted %d\n", ret, to_submit);
 750                         break;
 751                 }
 752                 prepped = 0;
 753
 754                 while (to_wait) {
 755                         int r;
 756
 757                         s->calls++;
 758                         r = io_getevents(s->aio_ctx, to_wait, to_wait, events, NULL);
 759                         if (r < 0) {
 760                                 perror("io_getevents");
 761                                 break;
 762                         } else if (r != to_wait) {
 763                                 printf("r=%d, wait=%d\n", r, to_wait);
 764                                 break;
 765                         }
 766                         r = reap_events_aio(s, events, r);
 767                         s->reaps += r;
 768                         to_wait -= r;
 769                 }
 770         } while (!s->finish);
 771
 772         free(iocbsptr);
 773         free(iocbs);
 774         free(events);
 775         finish = 1;
 776         return NULL;
 777 }
 778 #endif
 779
 780 static void *submitter_uring_fn(void *data)
 781 {
 782         struct submitter *s = data;
 783         struct io_sq_ring *ring = &s->sq_ring;
 784         int ret, prepped, nr_batch;
 785
 786         nr_batch = submitter_init(s);
 787
 788         prepped = 0;
 789         do {
 790                 int to_wait, to_submit, this_reap, to_prep;
 791                 unsigned ring_flags = 0;
 792
 793                 if (!prepped && s->inflight < depth) {
 794                         to_prep = min(depth - s->inflight, batch_submit);
 795                         prepped = prep_more_ios_uring(s, to_prep);
 796 #ifdef ARCH_HAVE_CPU_CLOCK
 797                         if (prepped && stats) {
 798                                 s->clock_batch[s->clock_index] = get_cpu_clock();
 799                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
 800                         }
 801 #endif
 802                 }
 803                 s->inflight += prepped;
 804 submit_more:
 805                 to_submit = prepped;
 806 submit:
 807                 if (to_submit && (s->inflight + to_submit <= depth))
 808                         to_wait = 0;
 809                 else
 810                         to_wait = min(s->inflight + to_submit, batch_complete);
 811
 812                 /*
 813                  * Only need to call io_uring_enter if we're not using SQ thread
 814                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
 815                  */
 816                 if (sq_thread_poll)
 817                         ring_flags = atomic_load_acquire(ring->flags);
 818                 if (!sq_thread_poll || ring_flags & IORING_SQ_NEED_WAKEUP) {
 819                         unsigned flags = 0;
 820
 821                         if (to_wait)
 822                                 flags = IORING_ENTER_GETEVENTS;
 823                         if (ring_flags & IORING_SQ_NEED_WAKEUP)
 824                                 flags |= IORING_ENTER_SQ_WAKEUP;
 825                         ret = io_uring_enter(s, to_submit, to_wait, flags);
 826                         s->calls++;
 827                 } else {
 828                         /* for SQPOLL, we submitted it all effectively */
 829                         ret = to_submit;
 830                 }
 831
 832                 /*
 833                  * For non SQ thread poll, we already got the events we needed
 834                  * through the io_uring_enter() above. For SQ thread poll, we
 835                  * need to loop here until we find enough events.
 836                  */
 837                 this_reap = 0;
 838                 do {
 839                         int r;
 840
 841                         r = reap_events_uring(s);
 842                         if (r == -1) {
 843                                 s->finish = 1;
 844                                 break;
 845                         } else if (r > 0)
 846                                 this_reap += r;
 847                 } while (sq_thread_poll && this_reap < to_wait);
 848                 s->reaps += this_reap;
 849
 850                 if (ret >= 0) {
 851                         if (!ret) {
 852                                 to_submit = 0;
 853                                 if (s->inflight)
 854                                         goto submit;
 855                                 continue;
 856                         } else if (ret < to_submit) {
 857                                 int diff = to_submit - ret;
 858
 859                                 s->done += ret;
 860                                 prepped -= diff;
 861                                 goto submit_more;
 862                         }
 863                         s->done += ret;
 864                         prepped = 0;
 865                         continue;
 866                 } else if (ret < 0) {
 867                         if (errno == EAGAIN) {
 868                                 if (s->finish)
 869                                         break;
 870                                 if (this_reap)
 871                                         goto submit;
 872                                 to_submit = 0;
 873                                 goto submit;
 874                         }
 875                         printf("io_submit: %s\n", strerror(errno));
 876                         break;
 877                 }
 878         } while (!s->finish);
 879
 880         finish = 1;
 881         return NULL;
 882 }
 883
 884 static struct submitter *get_submitter(int offset)
 885 {
 886         void *ret;
 887
 888         ret = submitter;
 889         if (offset)
 890                 ret += offset * (sizeof(*submitter) + depth * sizeof(struct iovec));
 891         return ret;
 892 }
 893
 894 static void do_finish(const char *reason)
 895 {
 896         int j;
 897         printf("Exiting on %s\n", reason);
 898         for (j = 0; j < nthreads; j++) {
 899                 struct submitter *s = get_submitter(j);
 900                 s->finish = 1;
 901         }
 902         if (max_iops > 100000)
 903                 printf("Maximum IOPS=%luK\n", max_iops / 1000);
 904         else if (max_iops)
 905                 printf("Maximum IOPS=%lu\n", max_iops);
 906         finish = 1;
 907 }
 908
 909 static void sig_int(int sig)
 910 {
 911         do_finish("signal");
 912 }
 913
 914 static void arm_sig_int(void)
 915 {
 916         struct sigaction act;
 917
 918         memset(&act, 0, sizeof(act));
 919         act.sa_handler = sig_int;
 920         act.sa_flags = SA_RESTART;
 921         sigaction(SIGINT, &act, NULL);
 922
 923         /* Windows uses SIGBREAK as a quit signal from other applications */
 924 #ifdef WIN32
 925         sigaction(SIGBREAK, &act, NULL);
 926 #endif
 927 }
 928
 929 static int setup_aio(struct submitter *s)
 930 {
 931 #ifdef CONFIG_LIBAIO
 932         if (polled) {
 933                 fprintf(stderr, "aio does not support polled IO\n");
 934                 polled = 0;
 935         }
 936         if (sq_thread_poll) {
 937                 fprintf(stderr, "aio does not support SQPOLL IO\n");
 938                 sq_thread_poll = 0;
 939         }
 940         if (do_nop) {
 941                 fprintf(stderr, "aio does not support polled IO\n");
 942                 do_nop = 0;
 943         }
 944         if (fixedbufs || register_files) {
 945                 fprintf(stderr, "aio does not support registered files or buffers\n");
 946                 fixedbufs = register_files = 0;
 947         }
 948
 949         return io_queue_init(depth, &s->aio_ctx);
 950 #else
 951         fprintf(stderr, "Legacy AIO not available on this system/build\n");
 952         errno = EINVAL;
 953         return -1;
 954 #endif
 955 }
 956
 957 static int setup_ring(struct submitter *s)
 958 {
 959         struct io_sq_ring *sring = &s->sq_ring;
 960         struct io_cq_ring *cring = &s->cq_ring;
 961         struct io_uring_params p;
 962         int ret, fd;
 963         void *ptr;
 964
 965         memset(&p, 0, sizeof(p));
 966
 967         if (polled && !do_nop)
 968                 p.flags |= IORING_SETUP_IOPOLL;
 969         if (sq_thread_poll) {
 970                 p.flags |= IORING_SETUP_SQPOLL;
 971                 if (sq_thread_cpu != -1) {
 972                         p.flags |= IORING_SETUP_SQ_AFF;
 973                         p.sq_thread_cpu = sq_thread_cpu;
 974                 }
 975         }
 976
 977         fd = io_uring_setup(depth, &p);
 978         if (fd < 0) {
 979                 perror("io_uring_setup");
 980                 return 1;
 981         }
 982         s->ring_fd = fd;
 983
 984         io_uring_probe(fd);
 985
 986         if (fixedbufs) {
 987                 struct rlimit rlim;
 988
 989                 rlim.rlim_cur = RLIM_INFINITY;
 990                 rlim.rlim_max = RLIM_INFINITY;
 991                 /* ignore potential error, not needed on newer kernels */
 992                 setrlimit(RLIMIT_MEMLOCK, &rlim);
 993
 994                 ret = io_uring_register_buffers(s);
 995                 if (ret < 0) {
 996                         perror("io_uring_register_buffers");
 997                         return 1;
 998                 }
 999
1000                 if (dma_map) {
1001                         ret = io_uring_map_buffers(s);
1002                         if (ret < 0) {
1003                                 perror("io_uring_map_buffers");
1004                                 return 1;
1005                         }
1006                 }
1007         }
1008
1009         if (register_files) {
1010                 ret = io_uring_register_files(s);
1011                 if (ret < 0) {
1012                         perror("io_uring_register_files");
1013                         return 1;
1014                 }
1015         }
1016
1017         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
1018                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
1019                         IORING_OFF_SQ_RING);
1020         sring->head = ptr + p.sq_off.head;
1021         sring->tail = ptr + p.sq_off.tail;
1022         sring->ring_mask = ptr + p.sq_off.ring_mask;
1023         sring->ring_entries = ptr + p.sq_off.ring_entries;
1024         sring->flags = ptr + p.sq_off.flags;
1025         sring->array = ptr + p.sq_off.array;
1026         sq_ring_mask = *sring->ring_mask;
1027
1028         s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
1029                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
1030                         IORING_OFF_SQES);
1031
1032         ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
1033                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
1034                         IORING_OFF_CQ_RING);
1035         cring->head = ptr + p.cq_off.head;
1036         cring->tail = ptr + p.cq_off.tail;
1037         cring->ring_mask = ptr + p.cq_off.ring_mask;
1038         cring->ring_entries = ptr + p.cq_off.ring_entries;
1039         cring->cqes = ptr + p.cq_off.cqes;
1040         cq_ring_mask = *cring->ring_mask;
1041         return 0;
1042 }
1043
1044 static void file_depths(char *buf)
1045 {
1046         bool prev = false;
1047         char *p;
1048         int i, j;
1049
1050         buf[0] = '\0';
1051         p = buf;
1052         for (j = 0; j < nthreads; j++) {
1053                 struct submitter *s = get_submitter(j);
1054
1055                 for (i = 0; i < s->nr_files; i++) {
1056                         struct file *f = &s->files[i];
1057
1058                         if (prev)
1059                                 p += sprintf(p, " %d", f->pending_ios);
1060                         else
1061                                 p += sprintf(p, "%d", f->pending_ios);
1062                         prev = true;
1063                 }
1064         }
1065 }
1066
1067 static void usage(char *argv, int status)
1068 {
1069         char runtime_str[16];
1070         snprintf(runtime_str, sizeof(runtime_str), "%d", runtime);
1071         printf("%s [options] -- [filenames]\n"
1072                 " -d <int>  : IO Depth, default %d\n"
1073                 " -s <int>  : Batch submit, default %d\n"
1074                 " -c <int>  : Batch complete, default %d\n"
1075                 " -b <int>  : Block size, default %d\n"
1076                 " -p <bool> : Polled IO, default %d\n"
1077                 " -B <bool> : Fixed buffers, default %d\n"
1078                 " -D <bool> : DMA map fixed buffers, default %d\n"
1079                 " -F <bool> : Register files, default %d\n"
1080                 " -n <int>  : Number of threads, default %d\n"
1081                 " -O <bool> : Use O_DIRECT, default %d\n"
1082                 " -N <bool> : Perform just no-op requests, default %d\n"
1083                 " -t <bool> : Track IO latencies, default %d\n"
1084                 " -T <int>  : TSC rate in HZ\n"
1085                 " -r <int>  : Runtime in seconds, default %s\n"
1086                 " -R <bool> : Use random IO, default %d\n"
1087                 " -a <bool> : Use legacy aio, default %d\n",
1088                 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled,
1089                 fixedbufs, dma_map, register_files, nthreads, !buffered, do_nop,
1090                 stats, runtime == 0 ? "unlimited" : runtime_str, aio, random_io);
1091         exit(status);
1092 }
1093
1094 static void read_tsc_rate(void)
1095 {
1096         char buffer[32];
1097         int fd, ret;
1098
1099         if (tsc_rate)
1100                 return;
1101
1102         fd = open(TSC_RATE_FILE, O_RDONLY);
1103         if (fd < 0)
1104                 return;
1105
1106         ret = read(fd, buffer, sizeof(buffer));
1107         if (ret < 0) {
1108                 close(fd);
1109                 return;
1110         }
1111
1112         tsc_rate = strtoul(buffer, NULL, 10);
1113         printf("Using TSC rate %luHz\n", tsc_rate);
1114         close(fd);
1115 }
1116
1117 static void write_tsc_rate(void)
1118 {
1119         char buffer[32];
1120         struct stat sb;
1121         int fd, ret;
1122
1123         if (!stat(TSC_RATE_FILE, &sb))
1124                 return;
1125
1126         fd = open(TSC_RATE_FILE, O_WRONLY | O_CREAT, 0644);
1127         if (fd < 0)
1128                 return;
1129
1130         memset(buffer, 0, sizeof(buffer));
1131         sprintf(buffer, "%lu", tsc_rate);
1132         ret = write(fd, buffer, strlen(buffer));
1133         if (ret < 0)
1134                 perror("write");
1135         close(fd);
1136 }
1137
1138 int main(int argc, char *argv[])
1139 {
1140         struct submitter *s;
1141         unsigned long done, calls, reap;
1142         int err, i, j, flags, fd, opt, threads_per_f, threads_rem = 0, nfiles;
1143         struct file f;
1144         char *fdepths;
1145         void *ret;
1146
1147         if (!do_nop && argc < 2)
1148                 usage(argv[0], 1);
1149
1150         while ((opt = getopt(argc, argv, "d:s:c:b:p:B:F:n:N:O:t:T:a:r:D:R:h?")) != -1) {
1151                 switch (opt) {
1152                 case 'a':
1153                         aio = !!atoi(optarg);
1154                         break;
1155                 case 'd':
1156                         depth = atoi(optarg);
1157                         break;
1158                 case 's':
1159                         batch_submit = atoi(optarg);
1160                         if (!batch_submit)
1161                                 batch_submit = 1;
1162                         break;
1163                 case 'c':
1164                         batch_complete = atoi(optarg);
1165                         if (!batch_complete)
1166                                 batch_complete = 1;
1167                         break;
1168                 case 'b':
1169                         bs = atoi(optarg);
1170                         break;
1171                 case 'p':
1172                         polled = !!atoi(optarg);
1173                         break;
1174                 case 'B':
1175                         fixedbufs = !!atoi(optarg);
1176                         break;
1177                 case 'F':
1178                         register_files = !!atoi(optarg);
1179                         break;
1180                 case 'n':
1181                         nthreads = atoi(optarg);
1182                         if (!nthreads) {
1183                                 printf("Threads must be non-zero\n");
1184                                 usage(argv[0], 1);
1185                         }
1186                         break;
1187                 case 'N':
1188                         do_nop = !!atoi(optarg);
1189                         break;
1190                 case 'O':
1191                         buffered = !atoi(optarg);
1192                         break;
1193                 case 't':
1194 #ifndef ARCH_HAVE_CPU_CLOCK
1195                         fprintf(stderr, "Stats not supported on this CPU\n");
1196                         return 1;
1197 #endif
1198                         stats = !!atoi(optarg);
1199                         break;
1200                 case 'T':
1201 #ifndef ARCH_HAVE_CPU_CLOCK
1202                         fprintf(stderr, "Stats not supported on this CPU\n");
1203                         return 1;
1204 #endif
1205                         tsc_rate = strtoul(optarg, NULL, 10);
1206                         write_tsc_rate();
1207                         break;
1208                 case 'r':
1209                         runtime = atoi(optarg);
1210                         break;
1211                 case 'D':
1212                         dma_map = !!atoi(optarg);
1213                         break;
1214                 case 'R':
1215                         random_io = !!atoi(optarg);
1216                         break;
1217                 case 'h':
1218                 case '?':
1219                 default:
1220                         usage(argv[0], 0);
1221                         break;
1222                 }
1223         }
1224
1225         if (stats)
1226                 read_tsc_rate();
1227
1228         if (batch_complete > depth)
1229                 batch_complete = depth;
1230         if (batch_submit > depth)
1231                 batch_submit = depth;
1232         if (!fixedbufs && dma_map)
1233                 dma_map = 0;
1234
1235         submitter = calloc(nthreads, sizeof(*submitter) +
1236                                 depth * sizeof(struct iovec));
1237         for (j = 0; j < nthreads; j++) {
1238                 s = get_submitter(j);
1239                 s->index = j;
1240                 s->done = s->calls = s->reaps = 0;
1241         }
1242
1243         flags = O_RDONLY | O_NOATIME;
1244         if (!buffered)
1245                 flags |= O_DIRECT;
1246
1247         j = 0;
1248         i = optind;
1249         nfiles = argc - i;
1250         if (!do_nop) {
1251                 if (!nfiles) {
1252                         printf("No files specified\n");
1253                         usage(argv[0], 1);
1254                 }
1255                 threads_per_f = nthreads / nfiles;
1256                 /* make sure each thread gets assigned files */
1257                 if (threads_per_f == 0) {
1258                         threads_per_f = 1;
1259                 } else {
1260                         threads_rem = nthreads - threads_per_f * nfiles;
1261                 }
1262         }
1263         while (!do_nop && i < argc) {
1264                 int k, limit;
1265
1266                 memset(&f, 0, sizeof(f));
1267
1268                 fd = open(argv[i], flags);
1269                 if (fd < 0) {
1270                         perror("open");
1271                         return 1;
1272                 }
1273                 f.real_fd = fd;
1274                 if (get_file_size(&f)) {
1275                         printf("failed getting size of device/file\n");
1276                         return 1;
1277                 }
1278                 if (f.max_blocks <= 1) {
1279                         printf("Zero file/device size?\n");
1280                         return 1;
1281                 }
1282                 f.max_blocks--;
1283
1284                 limit = threads_per_f;
1285                 limit += threads_rem > 0 ? 1 : 0;
1286                 for (k = 0; k < limit; k++) {
1287                         s = get_submitter((j + k) % nthreads);
1288
1289                         if (s->nr_files == MAX_FDS) {
1290                                 printf("Max number of files (%d) reached\n", MAX_FDS);
1291                                 break;
1292                         }
1293
1294                         memcpy(&s->files[s->nr_files], &f, sizeof(f));
1295
1296                         printf("Added file %s (submitter %d)\n", argv[i], s->index);
1297                         s->nr_files++;
1298                 }
1299                 threads_rem--;
1300                 i++;
1301                 j += limit;
1302         }
1303
1304         arm_sig_int();
1305
1306         for (j = 0; j < nthreads; j++) {
1307                 s = get_submitter(j);
1308                 for (i = 0; i < depth; i++) {
1309                         void *buf;
1310
1311                         if (posix_memalign(&buf, bs, bs)) {
1312                                 printf("failed alloc\n");
1313                                 return 1;
1314                         }
1315                         s->iovecs[i].iov_base = buf;
1316                         s->iovecs[i].iov_len = bs;
1317                 }
1318         }
1319
1320         for (j = 0; j < nthreads; j++) {
1321                 s = get_submitter(j);
1322
1323                 if (!aio)
1324                         err = setup_ring(s);
1325                 else
1326                         err = setup_aio(s);
1327                 if (err) {
1328                         printf("ring setup failed: %s, %d\n", strerror(errno), err);
1329                         return 1;
1330                 }
1331         }
1332         s = get_submitter(0);
1333         printf("polled=%d, fixedbufs=%d/%d, register_files=%d, buffered=%d, QD=%d\n", polled, fixedbufs, dma_map, register_files, buffered, depth);
1334         if (!aio)
1335                 printf("Engine=io_uring, sq_ring=%d, cq_ring=%d\n", *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
1336         else
1337                 printf("Engine=aio\n");
1338
1339         for (j = 0; j < nthreads; j++) {
1340                 s = get_submitter(j);
1341                 if (!aio)
1342                         pthread_create(&s->thread, NULL, submitter_uring_fn, s);
1343 #ifdef CONFIG_LIBAIO
1344                 else
1345                         pthread_create(&s->thread, NULL, submitter_aio_fn, s);
1346 #endif
1347         }
1348
1349         fdepths = malloc(8 * s->nr_files * nthreads);
1350         reap = calls = done = 0;
1351         do {
1352                 unsigned long this_done = 0;
1353                 unsigned long this_reap = 0;
1354                 unsigned long this_call = 0;
1355                 unsigned long rpc = 0, ipc = 0;
1356                 unsigned long iops, bw;
1357
1358                 sleep(1);
1359                 if (runtime && !--runtime)
1360                         do_finish("timeout");
1361
1362                 /* don't print partial run, if interrupted by signal */
1363                 if (finish)
1364                         break;
1365
1366                 /* one second in to the run, enable stats */
1367                 if (stats)
1368                         stats_running = 1;
1369
1370                 for (j = 0; j < nthreads; j++) {
1371                         s = get_submitter(j);
1372                         this_done += s->done;
1373                         this_call += s->calls;
1374                         this_reap += s->reaps;
1375                 }
1376                 if (this_call - calls) {
1377                         rpc = (this_done - done) / (this_call - calls);
1378                         ipc = (this_reap - reap) / (this_call - calls);
1379                 } else
1380                         rpc = ipc = -1;
1381                 file_depths(fdepths);
1382                 iops = this_done - done;
1383                 if (bs > 1048576)
1384                         bw = iops * (bs / 1048576);
1385                 else
1386                         bw = iops / (1048576 / bs);
1387                 if (iops > 100000)
1388                         printf("IOPS=%luK, ", iops / 1000);
1389                 else
1390                         printf("IOPS=%lu, ", iops);
1391                 max_iops = max(max_iops, iops);
1392                 if (!do_nop)
1393                         printf("BW=%luMiB/s, ", bw);
1394                 printf("IOS/call=%ld/%ld, inflight=(%s)\n", rpc, ipc, fdepths);
1395                 done = this_done;
1396                 calls = this_call;
1397                 reap = this_reap;
1398         } while (!finish);
1399
1400         for (j = 0; j < nthreads; j++) {
1401                 s = get_submitter(j);
1402                 pthread_join(s->thread, &ret);
1403                 close(s->ring_fd);
1404
1405                 if (stats) {
1406                         unsigned long nr;
1407
1408                         printf("%d: Latency percentiles:\n", s->tid);
1409                         for (i = 0, nr = 0; i < PLAT_NR; i++)
1410                                 nr += s->plat[i];
1411                         show_clat_percentiles(s->plat, nr, 4);
1412                         free(s->clock_batch);
1413                         free(s->plat);
1414                 }
1415         }
1416
1417         free(fdepths);
1418         free(submitter);
1419         return 0;
1420 }