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         /*
 388          * Clamp CQ ring size at our SQ ring size, we don't need more entries
 389          * than that.
 390          */
 391         p->flags |= IORING_SETUP_CQSIZE;
 392         p->cq_entries = entries;
 393
 394         return syscall(__NR_io_uring_setup, entries, p);
 395 }
 396
 397 static void io_uring_probe(int fd)
 398 {
 399         struct io_uring_probe *p;
 400         int ret;
 401
 402         p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 403         if (!p)
 404                 return;
 405
 406         memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 407         ret = syscall(__NR_io_uring_register, fd, IORING_REGISTER_PROBE, p, 256);
 408         if (ret < 0)
 409                 goto out;
 410
 411         if (IORING_OP_READ > p->ops_len)
 412                 goto out;
 413
 414         if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED))
 415                 vectored = 0;
 416 out:
 417         free(p);
 418 }
 419
 420 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 421                           unsigned int min_complete, unsigned int flags)
 422 {
 423         return syscall(__NR_io_uring_enter, s->ring_fd, to_submit, min_complete,
 424                         flags, NULL, 0);
 425 }
 426
 427 #ifndef CONFIG_HAVE_GETTID
 428 static int gettid(void)
 429 {
 430         return syscall(__NR_gettid);
 431 }
 432 #endif
 433
 434 static unsigned file_depth(struct submitter *s)
 435 {
 436         return (depth + s->nr_files - 1) / s->nr_files;
 437 }
 438
 439 static void init_io(struct submitter *s, unsigned index)
 440 {
 441         struct io_uring_sqe *sqe = &s->sqes[index];
 442         unsigned long offset;
 443         struct file *f;
 444         long r;
 445
 446         if (do_nop) {
 447                 sqe->opcode = IORING_OP_NOP;
 448                 return;
 449         }
 450
 451         if (s->nr_files == 1) {
 452                 f = &s->files[0];
 453         } else {
 454                 f = &s->files[s->cur_file];
 455                 if (f->pending_ios >= file_depth(s)) {
 456                         s->cur_file++;
 457                         if (s->cur_file == s->nr_files)
 458                                 s->cur_file = 0;
 459                         f = &s->files[s->cur_file];
 460                 }
 461         }
 462         f->pending_ios++;
 463
 464         if (random_io) {
 465                 r = __rand64(&s->rand_state);
 466                 offset = (r % (f->max_blocks - 1)) * bs;
 467         } else {
 468                 offset = f->cur_off;
 469                 f->cur_off += bs;
 470                 if (f->cur_off + bs > f->max_size)
 471                         f->cur_off = 0;
 472         }
 473
 474         if (register_files) {
 475                 sqe->flags = IOSQE_FIXED_FILE;
 476                 sqe->fd = f->fixed_fd;
 477         } else {
 478                 sqe->flags = 0;
 479                 sqe->fd = f->real_fd;
 480         }
 481         if (fixedbufs) {
 482                 sqe->opcode = IORING_OP_READ_FIXED;
 483                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 484                 sqe->len = bs;
 485                 sqe->buf_index = index;
 486         } else if (!vectored) {
 487                 sqe->opcode = IORING_OP_READ;
 488                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 489                 sqe->len = bs;
 490                 sqe->buf_index = 0;
 491         } else {
 492                 sqe->opcode = IORING_OP_READV;
 493                 sqe->addr = (unsigned long) &s->iovecs[index];
 494                 sqe->len = 1;
 495                 sqe->buf_index = 0;
 496         }
 497         sqe->ioprio = 0;
 498         sqe->off = offset;
 499         sqe->user_data = (unsigned long) f->fileno;
 500         if (stats && stats_running)
 501                 sqe->user_data |= ((unsigned long)s->clock_index << 32);
 502 }
 503
 504 static int prep_more_ios_uring(struct submitter *s, int max_ios)
 505 {
 506         struct io_sq_ring *ring = &s->sq_ring;
 507         unsigned index, tail, next_tail, prepped = 0;
 508
 509         next_tail = tail = *ring->tail;
 510         do {
 511                 next_tail++;
 512                 if (next_tail == atomic_load_acquire(ring->head))
 513                         break;
 514
 515                 index = tail & sq_ring_mask;
 516                 init_io(s, index);
 517                 ring->array[index] = index;
 518                 prepped++;
 519                 tail = next_tail;
 520         } while (prepped < max_ios);
 521
 522         if (prepped)
 523                 atomic_store_release(ring->tail, tail);
 524         return prepped;
 525 }
 526
 527 static int get_file_size(struct file *f)
 528 {
 529         struct stat st;
 530
 531         if (fstat(f->real_fd, &st) < 0)
 532                 return -1;
 533         if (S_ISBLK(st.st_mode)) {
 534                 unsigned long long bytes;
 535
 536                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 537                         return -1;
 538
 539                 f->max_blocks = bytes / bs;
 540                 f->max_size = bytes;
 541                 return 0;
 542         } else if (S_ISREG(st.st_mode)) {
 543                 f->max_blocks = st.st_size / bs;
 544                 f->max_size = st.st_size;
 545                 return 0;
 546         }
 547
 548         return -1;
 549 }
 550
 551 static int reap_events_uring(struct submitter *s)
 552 {
 553         struct io_cq_ring *ring = &s->cq_ring;
 554         struct io_uring_cqe *cqe;
 555         unsigned head, reaped = 0;
 556         int last_idx = -1, stat_nr = 0;
 557
 558         head = *ring->head;
 559         do {
 560                 struct file *f;
 561
 562                 read_barrier();
 563                 if (head == atomic_load_acquire(ring->tail))
 564                         break;
 565                 cqe = &ring->cqes[head & cq_ring_mask];
 566                 if (!do_nop) {
 567                         int fileno = cqe->user_data & 0xffffffff;
 568
 569                         f = &s->files[fileno];
 570                         f->pending_ios--;
 571                         if (cqe->res != bs) {
 572                                 printf("io: unexpected ret=%d\n", cqe->res);
 573                                 if (polled && cqe->res == -EOPNOTSUPP)
 574                                         printf("Your filesystem/driver/kernel doesn't support polled IO\n");
 575                                 return -1;
 576                         }
 577                 }
 578                 if (stats) {
 579                         int clock_index = cqe->user_data >> 32;
 580
 581                         if (last_idx != clock_index) {
 582                                 if (last_idx != -1) {
 583                                         add_stat(s, last_idx, stat_nr);
 584                                         stat_nr = 0;
 585                                 }
 586                                 last_idx = clock_index;
 587                         }
 588                         stat_nr++;
 589                 }
 590                 reaped++;
 591                 head++;
 592         } while (1);
 593
 594         if (stat_nr)
 595                 add_stat(s, last_idx, stat_nr);
 596
 597         if (reaped) {
 598                 s->inflight -= reaped;
 599                 atomic_store_release(ring->head, head);
 600         }
 601         return reaped;
 602 }
 603
 604 static int submitter_init(struct submitter *s)
 605 {
 606         int i, nr_batch;
 607
 608         s->tid = gettid();
 609         printf("submitter=%d, tid=%d\n", s->index, s->tid);
 610
 611         __init_rand64(&s->rand_state, pthread_self());
 612         srand48(pthread_self());
 613
 614         for (i = 0; i < MAX_FDS; i++)
 615                 s->files[i].fileno = i;
 616
 617         if (stats) {
 618                 nr_batch = roundup_pow2(depth / batch_submit);
 619                 if (nr_batch < 2)
 620                         nr_batch = 2;
 621                 s->clock_batch = calloc(nr_batch, sizeof(unsigned long));
 622                 s->clock_index = 1;
 623
 624                 s->plat = calloc(PLAT_NR, sizeof(unsigned long));
 625         } else {
 626                 s->clock_batch = NULL;
 627                 s->plat = NULL;
 628                 nr_batch = 0;
 629         }
 630
 631         return nr_batch;
 632 }
 633
 634 #ifdef CONFIG_LIBAIO
 635 static int prep_more_ios_aio(struct submitter *s, int max_ios, struct iocb *iocbs)
 636 {
 637         unsigned long offset, data;
 638         struct file *f;
 639         unsigned index;
 640         long r;
 641
 642         index = 0;
 643         while (index < max_ios) {
 644                 struct iocb *iocb = &iocbs[index];
 645
 646                 if (s->nr_files == 1) {
 647                         f = &s->files[0];
 648                 } else {
 649                         f = &s->files[s->cur_file];
 650                         if (f->pending_ios >= file_depth(s)) {
 651                                 s->cur_file++;
 652                                 if (s->cur_file == s->nr_files)
 653                                         s->cur_file = 0;
 654                                 f = &s->files[s->cur_file];
 655                         }
 656                 }
 657                 f->pending_ios++;
 658
 659                 r = lrand48();
 660                 offset = (r % (f->max_blocks - 1)) * bs;
 661                 io_prep_pread(iocb, f->real_fd, s->iovecs[index].iov_base,
 662                                 s->iovecs[index].iov_len, offset);
 663
 664                 data = f->fileno;
 665                 if (stats && stats_running)
 666                         data |= ((unsigned long) s->clock_index << 32);
 667                 iocb->data = (void *) (uintptr_t) data;
 668                 index++;
 669         }
 670         return index;
 671 }
 672
 673 static int reap_events_aio(struct submitter *s, struct io_event *events, int evs)
 674 {
 675         int last_idx = -1, stat_nr = 0;
 676         int reaped = 0;
 677
 678         while (evs) {
 679                 unsigned long data = (uintptr_t) events[reaped].data;
 680                 struct file *f = &s->files[data & 0xffffffff];
 681
 682                 f->pending_ios--;
 683                 if (events[reaped].res != bs) {
 684                         printf("io: unexpected ret=%ld\n", events[reaped].res);
 685                         return -1;
 686                 }
 687                 if (stats) {
 688                         int clock_index = data >> 32;
 689
 690                         if (last_idx != clock_index) {
 691                                 if (last_idx != -1) {
 692                                         add_stat(s, last_idx, stat_nr);
 693                                         stat_nr = 0;
 694                                 }
 695                                 last_idx = clock_index;
 696                         }
 697                         stat_nr++;
 698                 }
 699                 reaped++;
 700                 evs--;
 701         }
 702
 703         if (stat_nr)
 704                 add_stat(s, last_idx, stat_nr);
 705
 706         s->inflight -= reaped;
 707         s->done += reaped;
 708         return reaped;
 709 }
 710
 711 static void *submitter_aio_fn(void *data)
 712 {
 713         struct submitter *s = data;
 714         int i, ret, prepped, nr_batch;
 715         struct iocb **iocbsptr;
 716         struct iocb *iocbs;
 717         struct io_event *events;
 718
 719         nr_batch = submitter_init(s);
 720
 721         iocbsptr = calloc(depth, sizeof(struct iocb *));
 722         iocbs = calloc(depth, sizeof(struct iocb));
 723         events = calloc(depth, sizeof(struct io_event));
 724
 725         for (i = 0; i < depth; i++)
 726                 iocbsptr[i] = &iocbs[i];
 727
 728         prepped = 0;
 729         do {
 730                 int to_wait, to_submit, to_prep;
 731
 732                 if (!prepped && s->inflight < depth) {
 733                         to_prep = min(depth - s->inflight, batch_submit);
 734                         prepped = prep_more_ios_aio(s, to_prep, iocbs);
 735 #ifdef ARCH_HAVE_CPU_CLOCK
 736                         if (prepped && stats) {
 737                                 s->clock_batch[s->clock_index] = get_cpu_clock();
 738                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
 739                         }
 740 #endif
 741                 }
 742                 s->inflight += prepped;
 743                 to_submit = prepped;
 744
 745                 if (to_submit && (s->inflight + to_submit <= depth))
 746                         to_wait = 0;
 747                 else
 748                         to_wait = min(s->inflight + to_submit, batch_complete);
 749
 750                 ret = io_submit(s->aio_ctx, to_submit, iocbsptr);
 751                 s->calls++;
 752                 if (ret < 0) {
 753                         perror("io_submit");
 754                         break;
 755                 } else if (ret != to_submit) {
 756                         printf("submitted %d, wanted %d\n", ret, to_submit);
 757                         break;
 758                 }
 759                 prepped = 0;
 760
 761                 while (to_wait) {
 762                         int r;
 763
 764                         s->calls++;
 765                         r = io_getevents(s->aio_ctx, to_wait, to_wait, events, NULL);
 766                         if (r < 0) {
 767                                 perror("io_getevents");
 768                                 break;
 769                         } else if (r != to_wait) {
 770                                 printf("r=%d, wait=%d\n", r, to_wait);
 771                                 break;
 772                         }
 773                         r = reap_events_aio(s, events, r);
 774                         s->reaps += r;
 775                         to_wait -= r;
 776                 }
 777         } while (!s->finish);
 778
 779         free(iocbsptr);
 780         free(iocbs);
 781         free(events);
 782         finish = 1;
 783         return NULL;
 784 }
 785 #endif
 786
 787 static void *submitter_uring_fn(void *data)
 788 {
 789         struct submitter *s = data;
 790         struct io_sq_ring *ring = &s->sq_ring;
 791         int ret, prepped, nr_batch;
 792
 793         nr_batch = submitter_init(s);
 794
 795         prepped = 0;
 796         do {
 797                 int to_wait, to_submit, this_reap, to_prep;
 798                 unsigned ring_flags = 0;
 799
 800                 if (!prepped && s->inflight < depth) {
 801                         to_prep = min(depth - s->inflight, batch_submit);
 802                         prepped = prep_more_ios_uring(s, to_prep);
 803 #ifdef ARCH_HAVE_CPU_CLOCK
 804                         if (prepped && stats) {
 805                                 s->clock_batch[s->clock_index] = get_cpu_clock();
 806                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
 807                         }
 808 #endif
 809                 }
 810                 s->inflight += prepped;
 811 submit_more:
 812                 to_submit = prepped;
 813 submit:
 814                 if (to_submit && (s->inflight + to_submit <= depth))
 815                         to_wait = 0;
 816                 else
 817                         to_wait = min(s->inflight + to_submit, batch_complete);
 818
 819                 /*
 820                  * Only need to call io_uring_enter if we're not using SQ thread
 821                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
 822                  */
 823                 if (sq_thread_poll)
 824                         ring_flags = atomic_load_acquire(ring->flags);
 825                 if (!sq_thread_poll || ring_flags & IORING_SQ_NEED_WAKEUP) {
 826                         unsigned flags = 0;
 827
 828                         if (to_wait)
 829                                 flags = IORING_ENTER_GETEVENTS;
 830                         if (ring_flags & IORING_SQ_NEED_WAKEUP)
 831                                 flags |= IORING_ENTER_SQ_WAKEUP;
 832                         ret = io_uring_enter(s, to_submit, to_wait, flags);
 833                         s->calls++;
 834                 } else {
 835                         /* for SQPOLL, we submitted it all effectively */
 836                         ret = to_submit;
 837                 }
 838
 839                 /*
 840                  * For non SQ thread poll, we already got the events we needed
 841                  * through the io_uring_enter() above. For SQ thread poll, we
 842                  * need to loop here until we find enough events.
 843                  */
 844                 this_reap = 0;
 845                 do {
 846                         int r;
 847
 848                         r = reap_events_uring(s);
 849                         if (r == -1) {
 850                                 s->finish = 1;
 851                                 break;
 852                         } else if (r > 0)
 853                                 this_reap += r;
 854                 } while (sq_thread_poll && this_reap < to_wait);
 855                 s->reaps += this_reap;
 856
 857                 if (ret >= 0) {
 858                         if (!ret) {
 859                                 to_submit = 0;
 860                                 if (s->inflight)
 861                                         goto submit;
 862                                 continue;
 863                         } else if (ret < to_submit) {
 864                                 int diff = to_submit - ret;
 865
 866                                 s->done += ret;
 867                                 prepped -= diff;
 868                                 goto submit_more;
 869                         }
 870                         s->done += ret;
 871                         prepped = 0;
 872                         continue;
 873                 } else if (ret < 0) {
 874                         if (errno == EAGAIN) {
 875                                 if (s->finish)
 876                                         break;
 877                                 if (this_reap)
 878                                         goto submit;
 879                                 to_submit = 0;
 880                                 goto submit;
 881                         }
 882                         printf("io_submit: %s\n", strerror(errno));
 883                         break;
 884                 }
 885         } while (!s->finish);
 886
 887         finish = 1;
 888         return NULL;
 889 }
 890
 891 static struct submitter *get_submitter(int offset)
 892 {
 893         void *ret;
 894
 895         ret = submitter;
 896         if (offset)
 897                 ret += offset * (sizeof(*submitter) + depth * sizeof(struct iovec));
 898         return ret;
 899 }
 900
 901 static void do_finish(const char *reason)
 902 {
 903         int j;
 904         printf("Exiting on %s\n", reason);
 905         for (j = 0; j < nthreads; j++) {
 906                 struct submitter *s = get_submitter(j);
 907                 s->finish = 1;
 908         }
 909         if (max_iops > 100000)
 910                 printf("Maximum IOPS=%luK\n", max_iops / 1000);
 911         else if (max_iops)
 912                 printf("Maximum IOPS=%lu\n", max_iops);
 913         finish = 1;
 914 }
 915
 916 static void sig_int(int sig)
 917 {
 918         do_finish("signal");
 919 }
 920
 921 static void arm_sig_int(void)
 922 {
 923         struct sigaction act;
 924
 925         memset(&act, 0, sizeof(act));
 926         act.sa_handler = sig_int;
 927         act.sa_flags = SA_RESTART;
 928         sigaction(SIGINT, &act, NULL);
 929
 930         /* Windows uses SIGBREAK as a quit signal from other applications */
 931 #ifdef WIN32
 932         sigaction(SIGBREAK, &act, NULL);
 933 #endif
 934 }
 935
 936 static int setup_aio(struct submitter *s)
 937 {
 938 #ifdef CONFIG_LIBAIO
 939         if (polled) {
 940                 fprintf(stderr, "aio does not support polled IO\n");
 941                 polled = 0;
 942         }
 943         if (sq_thread_poll) {
 944                 fprintf(stderr, "aio does not support SQPOLL IO\n");
 945                 sq_thread_poll = 0;
 946         }
 947         if (do_nop) {
 948                 fprintf(stderr, "aio does not support polled IO\n");
 949                 do_nop = 0;
 950         }
 951         if (fixedbufs || register_files) {
 952                 fprintf(stderr, "aio does not support registered files or buffers\n");
 953                 fixedbufs = register_files = 0;
 954         }
 955
 956         return io_queue_init(depth, &s->aio_ctx);
 957 #else
 958         fprintf(stderr, "Legacy AIO not available on this system/build\n");
 959         errno = EINVAL;
 960         return -1;
 961 #endif
 962 }
 963
 964 static int setup_ring(struct submitter *s)
 965 {
 966         struct io_sq_ring *sring = &s->sq_ring;
 967         struct io_cq_ring *cring = &s->cq_ring;
 968         struct io_uring_params p;
 969         int ret, fd;
 970         void *ptr;
 971
 972         memset(&p, 0, sizeof(p));
 973
 974         if (polled && !do_nop)
 975                 p.flags |= IORING_SETUP_IOPOLL;
 976         if (sq_thread_poll) {
 977                 p.flags |= IORING_SETUP_SQPOLL;
 978                 if (sq_thread_cpu != -1) {
 979                         p.flags |= IORING_SETUP_SQ_AFF;
 980                         p.sq_thread_cpu = sq_thread_cpu;
 981                 }
 982         }
 983
 984         fd = io_uring_setup(depth, &p);
 985         if (fd < 0) {
 986                 perror("io_uring_setup");
 987                 return 1;
 988         }
 989         s->ring_fd = fd;
 990
 991         io_uring_probe(fd);
 992
 993         if (fixedbufs) {
 994                 struct rlimit rlim;
 995
 996                 rlim.rlim_cur = RLIM_INFINITY;
 997                 rlim.rlim_max = RLIM_INFINITY;
 998                 /* ignore potential error, not needed on newer kernels */
 999                 setrlimit(RLIMIT_MEMLOCK, &rlim);
1000
1001                 ret = io_uring_register_buffers(s);
1002                 if (ret < 0) {
1003                         perror("io_uring_register_buffers");
1004                         return 1;
1005                 }
1006
1007                 if (dma_map) {
1008                         ret = io_uring_map_buffers(s);
1009                         if (ret < 0) {
1010                                 perror("io_uring_map_buffers");
1011                                 return 1;
1012                         }
1013                 }
1014         }
1015
1016         if (register_files) {
1017                 ret = io_uring_register_files(s);
1018                 if (ret < 0) {
1019                         perror("io_uring_register_files");
1020                         return 1;
1021                 }
1022         }
1023
1024         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
1025                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
1026                         IORING_OFF_SQ_RING);
1027         sring->head = ptr + p.sq_off.head;
1028         sring->tail = ptr + p.sq_off.tail;
1029         sring->ring_mask = ptr + p.sq_off.ring_mask;
1030         sring->ring_entries = ptr + p.sq_off.ring_entries;
1031         sring->flags = ptr + p.sq_off.flags;
1032         sring->array = ptr + p.sq_off.array;
1033         sq_ring_mask = *sring->ring_mask;
1034
1035         s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
1036                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
1037                         IORING_OFF_SQES);
1038
1039         ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
1040                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
1041                         IORING_OFF_CQ_RING);
1042         cring->head = ptr + p.cq_off.head;
1043         cring->tail = ptr + p.cq_off.tail;
1044         cring->ring_mask = ptr + p.cq_off.ring_mask;
1045         cring->ring_entries = ptr + p.cq_off.ring_entries;
1046         cring->cqes = ptr + p.cq_off.cqes;
1047         cq_ring_mask = *cring->ring_mask;
1048         return 0;
1049 }
1050
1051 static void file_depths(char *buf)
1052 {
1053         bool prev = false;
1054         char *p;
1055         int i, j;
1056
1057         buf[0] = '\0';
1058         p = buf;
1059         for (j = 0; j < nthreads; j++) {
1060                 struct submitter *s = get_submitter(j);
1061
1062                 for (i = 0; i < s->nr_files; i++) {
1063                         struct file *f = &s->files[i];
1064
1065                         if (prev)
1066                                 p += sprintf(p, " %d", f->pending_ios);
1067                         else
1068                                 p += sprintf(p, "%d", f->pending_ios);
1069                         prev = true;
1070                 }
1071         }
1072 }
1073
1074 static void usage(char *argv, int status)
1075 {
1076         char runtime_str[16];
1077         snprintf(runtime_str, sizeof(runtime_str), "%d", runtime);
1078         printf("%s [options] -- [filenames]\n"
1079                 " -d <int>  : IO Depth, default %d\n"
1080                 " -s <int>  : Batch submit, default %d\n"
1081                 " -c <int>  : Batch complete, default %d\n"
1082                 " -b <int>  : Block size, default %d\n"
1083                 " -p <bool> : Polled IO, default %d\n"
1084                 " -B <bool> : Fixed buffers, default %d\n"
1085                 " -D <bool> : DMA map fixed buffers, default %d\n"
1086                 " -F <bool> : Register files, default %d\n"
1087                 " -n <int>  : Number of threads, default %d\n"
1088                 " -O <bool> : Use O_DIRECT, default %d\n"
1089                 " -N <bool> : Perform just no-op requests, default %d\n"
1090                 " -t <bool> : Track IO latencies, default %d\n"
1091                 " -T <int>  : TSC rate in HZ\n"
1092                 " -r <int>  : Runtime in seconds, default %s\n"
1093                 " -R <bool> : Use random IO, default %d\n"
1094                 " -a <bool> : Use legacy aio, default %d\n",
1095                 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled,
1096                 fixedbufs, dma_map, register_files, nthreads, !buffered, do_nop,
1097                 stats, runtime == 0 ? "unlimited" : runtime_str, aio, random_io);
1098         exit(status);
1099 }
1100
1101 static void read_tsc_rate(void)
1102 {
1103         char buffer[32];
1104         int fd, ret;
1105
1106         if (tsc_rate)
1107                 return;
1108
1109         fd = open(TSC_RATE_FILE, O_RDONLY);
1110         if (fd < 0)
1111                 return;
1112
1113         ret = read(fd, buffer, sizeof(buffer));
1114         if (ret < 0) {
1115                 close(fd);
1116                 return;
1117         }
1118
1119         tsc_rate = strtoul(buffer, NULL, 10);
1120         printf("Using TSC rate %luHz\n", tsc_rate);
1121         close(fd);
1122 }
1123
1124 static void write_tsc_rate(void)
1125 {
1126         char buffer[32];
1127         struct stat sb;
1128         int fd, ret;
1129
1130         if (!stat(TSC_RATE_FILE, &sb))
1131                 return;
1132
1133         fd = open(TSC_RATE_FILE, O_WRONLY | O_CREAT, 0644);
1134         if (fd < 0)
1135                 return;
1136
1137         memset(buffer, 0, sizeof(buffer));
1138         sprintf(buffer, "%lu", tsc_rate);
1139         ret = write(fd, buffer, strlen(buffer));
1140         if (ret < 0)
1141                 perror("write");
1142         close(fd);
1143 }
1144
1145 int main(int argc, char *argv[])
1146 {
1147         struct submitter *s;
1148         unsigned long done, calls, reap;
1149         int err, i, j, flags, fd, opt, threads_per_f, threads_rem = 0, nfiles;
1150         struct file f;
1151         char *fdepths;
1152         void *ret;
1153
1154         if (!do_nop && argc < 2)
1155                 usage(argv[0], 1);
1156
1157         while ((opt = getopt(argc, argv, "d:s:c:b:p:B:F:n:N:O:t:T:a:r:D:R:h?")) != -1) {
1158                 switch (opt) {
1159                 case 'a':
1160                         aio = !!atoi(optarg);
1161                         break;
1162                 case 'd':
1163                         depth = atoi(optarg);
1164                         break;
1165                 case 's':
1166                         batch_submit = atoi(optarg);
1167                         if (!batch_submit)
1168                                 batch_submit = 1;
1169                         break;
1170                 case 'c':
1171                         batch_complete = atoi(optarg);
1172                         if (!batch_complete)
1173                                 batch_complete = 1;
1174                         break;
1175                 case 'b':
1176                         bs = atoi(optarg);
1177                         break;
1178                 case 'p':
1179                         polled = !!atoi(optarg);
1180                         break;
1181                 case 'B':
1182                         fixedbufs = !!atoi(optarg);
1183                         break;
1184                 case 'F':
1185                         register_files = !!atoi(optarg);
1186                         break;
1187                 case 'n':
1188                         nthreads = atoi(optarg);
1189                         if (!nthreads) {
1190                                 printf("Threads must be non-zero\n");
1191                                 usage(argv[0], 1);
1192                         }
1193                         break;
1194                 case 'N':
1195                         do_nop = !!atoi(optarg);
1196                         break;
1197                 case 'O':
1198                         buffered = !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                         stats = !!atoi(optarg);
1206                         break;
1207                 case 'T':
1208 #ifndef ARCH_HAVE_CPU_CLOCK
1209                         fprintf(stderr, "Stats not supported on this CPU\n");
1210                         return 1;
1211 #endif
1212                         tsc_rate = strtoul(optarg, NULL, 10);
1213                         write_tsc_rate();
1214                         break;
1215                 case 'r':
1216                         runtime = atoi(optarg);
1217                         break;
1218                 case 'D':
1219                         dma_map = !!atoi(optarg);
1220                         break;
1221                 case 'R':
1222                         random_io = !!atoi(optarg);
1223                         break;
1224                 case 'h':
1225                 case '?':
1226                 default:
1227                         usage(argv[0], 0);
1228                         break;
1229                 }
1230         }
1231
1232         if (stats)
1233                 read_tsc_rate();
1234
1235         if (batch_complete > depth)
1236                 batch_complete = depth;
1237         if (batch_submit > depth)
1238                 batch_submit = depth;
1239         if (!fixedbufs && dma_map)
1240                 dma_map = 0;
1241
1242         submitter = calloc(nthreads, sizeof(*submitter) +
1243                                 depth * sizeof(struct iovec));
1244         for (j = 0; j < nthreads; j++) {
1245                 s = get_submitter(j);
1246                 s->index = j;
1247                 s->done = s->calls = s->reaps = 0;
1248         }
1249
1250         flags = O_RDONLY | O_NOATIME;
1251         if (!buffered)
1252                 flags |= O_DIRECT;
1253
1254         j = 0;
1255         i = optind;
1256         nfiles = argc - i;
1257         if (!do_nop) {
1258                 if (!nfiles) {
1259                         printf("No files specified\n");
1260                         usage(argv[0], 1);
1261                 }
1262                 threads_per_f = nthreads / nfiles;
1263                 /* make sure each thread gets assigned files */
1264                 if (threads_per_f == 0) {
1265                         threads_per_f = 1;
1266                 } else {
1267                         threads_rem = nthreads - threads_per_f * nfiles;
1268                 }
1269         }
1270         while (!do_nop && i < argc) {
1271                 int k, limit;
1272
1273                 memset(&f, 0, sizeof(f));
1274
1275                 fd = open(argv[i], flags);
1276                 if (fd < 0) {
1277                         perror("open");
1278                         return 1;
1279                 }
1280                 f.real_fd = fd;
1281                 if (get_file_size(&f)) {
1282                         printf("failed getting size of device/file\n");
1283                         return 1;
1284                 }
1285                 if (f.max_blocks <= 1) {
1286                         printf("Zero file/device size?\n");
1287                         return 1;
1288                 }
1289                 f.max_blocks--;
1290
1291                 limit = threads_per_f;
1292                 limit += threads_rem > 0 ? 1 : 0;
1293                 for (k = 0; k < limit; k++) {
1294                         s = get_submitter((j + k) % nthreads);
1295
1296                         if (s->nr_files == MAX_FDS) {
1297                                 printf("Max number of files (%d) reached\n", MAX_FDS);
1298                                 break;
1299                         }
1300
1301                         memcpy(&s->files[s->nr_files], &f, sizeof(f));
1302
1303                         printf("Added file %s (submitter %d)\n", argv[i], s->index);
1304                         s->nr_files++;
1305                 }
1306                 threads_rem--;
1307                 i++;
1308                 j += limit;
1309         }
1310
1311         arm_sig_int();
1312
1313         for (j = 0; j < nthreads; j++) {
1314                 s = get_submitter(j);
1315                 for (i = 0; i < depth; i++) {
1316                         void *buf;
1317
1318                         if (posix_memalign(&buf, bs, bs)) {
1319                                 printf("failed alloc\n");
1320                                 return 1;
1321                         }
1322                         s->iovecs[i].iov_base = buf;
1323                         s->iovecs[i].iov_len = bs;
1324                 }
1325         }
1326
1327         for (j = 0; j < nthreads; j++) {
1328                 s = get_submitter(j);
1329
1330                 if (!aio)
1331                         err = setup_ring(s);
1332                 else
1333                         err = setup_aio(s);
1334                 if (err) {
1335                         printf("ring setup failed: %s, %d\n", strerror(errno), err);
1336                         return 1;
1337                 }
1338         }
1339         s = get_submitter(0);
1340         printf("polled=%d, fixedbufs=%d/%d, register_files=%d, buffered=%d, QD=%d\n", polled, fixedbufs, dma_map, register_files, buffered, depth);
1341         if (!aio)
1342                 printf("Engine=io_uring, sq_ring=%d, cq_ring=%d\n", *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
1343         else
1344                 printf("Engine=aio\n");
1345
1346         for (j = 0; j < nthreads; j++) {
1347                 s = get_submitter(j);
1348                 if (!aio)
1349                         pthread_create(&s->thread, NULL, submitter_uring_fn, s);
1350 #ifdef CONFIG_LIBAIO
1351                 else
1352                         pthread_create(&s->thread, NULL, submitter_aio_fn, s);
1353 #endif
1354         }
1355
1356         fdepths = malloc(8 * s->nr_files * nthreads);
1357         reap = calls = done = 0;
1358         do {
1359                 unsigned long this_done = 0;
1360                 unsigned long this_reap = 0;
1361                 unsigned long this_call = 0;
1362                 unsigned long rpc = 0, ipc = 0;
1363                 unsigned long iops, bw;
1364
1365                 sleep(1);
1366                 if (runtime && !--runtime)
1367                         do_finish("timeout");
1368
1369                 /* don't print partial run, if interrupted by signal */
1370                 if (finish)
1371                         break;
1372
1373                 /* one second in to the run, enable stats */
1374                 if (stats)
1375                         stats_running = 1;
1376
1377                 for (j = 0; j < nthreads; j++) {
1378                         s = get_submitter(j);
1379                         this_done += s->done;
1380                         this_call += s->calls;
1381                         this_reap += s->reaps;
1382                 }
1383                 if (this_call - calls) {
1384                         rpc = (this_done - done) / (this_call - calls);
1385                         ipc = (this_reap - reap) / (this_call - calls);
1386                 } else
1387                         rpc = ipc = -1;
1388                 file_depths(fdepths);
1389                 iops = this_done - done;
1390                 if (bs > 1048576)
1391                         bw = iops * (bs / 1048576);
1392                 else
1393                         bw = iops / (1048576 / bs);
1394                 if (iops > 100000)
1395                         printf("IOPS=%luK, ", iops / 1000);
1396                 else
1397                         printf("IOPS=%lu, ", iops);
1398                 max_iops = max(max_iops, iops);
1399                 if (!do_nop)
1400                         printf("BW=%luMiB/s, ", bw);
1401                 printf("IOS/call=%ld/%ld, inflight=(%s)\n", rpc, ipc, fdepths);
1402                 done = this_done;
1403                 calls = this_call;
1404                 reap = this_reap;
1405         } while (!finish);
1406
1407         for (j = 0; j < nthreads; j++) {
1408                 s = get_submitter(j);
1409                 pthread_join(s->thread, &ret);
1410                 close(s->ring_fd);
1411
1412                 if (stats) {
1413                         unsigned long nr;
1414
1415                         printf("%d: Latency percentiles:\n", s->tid);
1416                         for (i = 0, nr = 0; i < PLAT_NR; i++)
1417                                 nr += s->plat[i];
1418                         show_clat_percentiles(s->plat, nr, 4);
1419                         free(s->clock_batch);
1420                         free(s->plat);
1421                 }
1422         }
1423
1424         free(fdepths);
1425         free(submitter);
1426         return 0;
1427 }