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