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