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