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 #ifdef CONFIG_LIBNUMA
  15 #include <numa.h>
  16 #endif
  17
  18 #include <sys/types.h>
  19 #include <sys/stat.h>
  20 #include <sys/ioctl.h>
  21 #include <sys/syscall.h>
  22 #include <sys/resource.h>
  23 #include <sys/mman.h>
  24 #include <sys/uio.h>
  25 #include <linux/fs.h>
  26 #include <fcntl.h>
  27 #include <unistd.h>
  28 #include <string.h>
  29 #include <pthread.h>
  30 #include <sched.h>
  31
  32 #include "../arch/arch.h"
  33 #include "../os/os.h"
  34 #include "../lib/types.h"
  35 #include "../lib/roundup.h"
  36 #include "../lib/rand.h"
  37 #include "../minmax.h"
  38 #include "../os/linux/io_uring.h"
  39 #include "../engines/nvme.h"
  40
  41 struct io_sq_ring {
  42         unsigned *head;
  43         unsigned *tail;
  44         unsigned *ring_mask;
  45         unsigned *ring_entries;
  46         unsigned *flags;
  47         unsigned *array;
  48 };
  49
  50 struct io_cq_ring {
  51         unsigned *head;
  52         unsigned *tail;
  53         unsigned *ring_mask;
  54         unsigned *ring_entries;
  55         struct io_uring_cqe *cqes;
  56 };
  57
  58 #define DEPTH                   128
  59 #define BATCH_SUBMIT            32
  60 #define BATCH_COMPLETE          32
  61 #define BS                      4096
  62
  63 #define MAX_FDS                 16
  64
  65 static unsigned sq_ring_mask, cq_ring_mask;
  66
  67 struct file {
  68         unsigned long max_blocks;
  69         unsigned long max_size;
  70         unsigned long cur_off;
  71         unsigned pending_ios;
  72         unsigned int nsid;      /* nsid field required for nvme-passthrough */
  73         unsigned int lba_shift; /* lba_shift field required for nvme-passthrough */
  74         int real_fd;
  75         int fixed_fd;
  76         int fileno;
  77 };
  78
  79 #define PLAT_BITS               6
  80 #define PLAT_VAL                (1 << PLAT_BITS)
  81 #define PLAT_GROUP_NR           29
  82 #define PLAT_NR                 (PLAT_GROUP_NR * PLAT_VAL)
  83
  84 struct submitter {
  85         pthread_t thread;
  86         int ring_fd;
  87         int enter_ring_fd;
  88         int index;
  89         struct io_sq_ring sq_ring;
  90         struct io_uring_sqe *sqes;
  91         struct io_cq_ring cq_ring;
  92         int inflight;
  93         int tid;
  94         unsigned long reaps;
  95         unsigned long done;
  96         unsigned long calls;
  97         volatile int finish;
  98
  99         __s32 *fds;
 100
 101         struct taus258_state rand_state;
 102
 103         unsigned long *clock_batch;
 104         int clock_index;
 105         unsigned long *plat;
 106
 107 #ifdef CONFIG_LIBAIO
 108         io_context_t aio_ctx;
 109 #endif
 110
 111         int numa_node;
 112         const char *filename;
 113
 114         struct file files[MAX_FDS];
 115         unsigned nr_files;
 116         unsigned cur_file;
 117         struct iovec iovecs[];
 118 };
 119
 120 static struct submitter *submitter;
 121 static volatile int finish;
 122 static int stats_running;
 123 static unsigned long max_iops;
 124 static long t_io_uring_page_size;
 125
 126 static int depth = DEPTH;
 127 static int batch_submit = BATCH_SUBMIT;
 128 static int batch_complete = BATCH_COMPLETE;
 129 static int bs = BS;
 130 static int polled = 1;          /* use IO polling */
 131 static int fixedbufs = 1;       /* use fixed user buffers */
 132 static int register_files = 1;  /* use fixed files */
 133 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
 134 static int sq_thread_poll = 0;  /* use kernel submission/poller thread */
 135 static int sq_thread_cpu = -1;  /* pin above thread to this CPU */
 136 static int do_nop = 0;          /* no-op SQ ring commands */
 137 static int nthreads = 1;
 138 static int stats = 0;           /* generate IO stats */
 139 static int aio = 0;             /* use libaio */
 140 static int runtime = 0;         /* runtime */
 141 static int random_io = 1;       /* random or sequential IO */
 142 static int register_ring = 1;   /* register ring */
 143 static int use_sync = 0;        /* use preadv2 */
 144 static int numa_placement = 0;  /* set to node of device */
 145 static int pt = 0;              /* passthrough I/O or not */
 146
 147 static unsigned long tsc_rate;
 148
 149 #define TSC_RATE_FILE   "tsc-rate"
 150
 151 static int vectored = 1;
 152
 153 static float plist[] = { 1.0, 5.0, 10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0,
 154                         80.0, 90.0, 95.0, 99.0, 99.5, 99.9, 99.95, 99.99 };
 155 static int plist_len = 17;
 156
 157 static int nvme_identify(int fd, __u32 nsid, enum nvme_identify_cns cns,
 158                          enum nvme_csi csi, void *data)
 159 {
 160         struct nvme_passthru_cmd cmd = {
 161                 .opcode         = nvme_admin_identify,
 162                 .nsid           = nsid,
 163                 .addr           = (__u64)(uintptr_t)data,
 164                 .data_len       = NVME_IDENTIFY_DATA_SIZE,
 165                 .cdw10          = cns,
 166                 .cdw11          = csi << NVME_IDENTIFY_CSI_SHIFT,
 167                 .timeout_ms     = NVME_DEFAULT_IOCTL_TIMEOUT,
 168         };
 169
 170         return ioctl(fd, NVME_IOCTL_ADMIN_CMD, &cmd);
 171 }
 172
 173 static int nvme_get_info(int fd, __u32 *nsid, __u32 *lba_sz, __u64 *nlba)
 174 {
 175         struct nvme_id_ns ns;
 176         int namespace_id;
 177         int err;
 178
 179         namespace_id = ioctl(fd, NVME_IOCTL_ID);
 180         if (namespace_id < 0) {
 181                 fprintf(stderr, "error failed to fetch namespace-id\n");
 182                 close(fd);
 183                 return -errno;
 184         }
 185
 186         /*
 187          * Identify namespace to get namespace-id, namespace size in LBA's
 188          * and LBA data size.
 189          */
 190         err = nvme_identify(fd, namespace_id, NVME_IDENTIFY_CNS_NS,
 191                                 NVME_CSI_NVM, &ns);
 192         if (err) {
 193                 fprintf(stderr, "error failed to fetch identify namespace\n");
 194                 close(fd);
 195                 return err;
 196         }
 197
 198         *nsid = namespace_id;
 199         *lba_sz = 1 << ns.lbaf[(ns.flbas & 0x0f)].ds;
 200         *nlba = ns.nsze;
 201
 202         return 0;
 203 }
 204
 205 static unsigned long cycles_to_nsec(unsigned long cycles)
 206 {
 207         uint64_t val;
 208
 209         if (!tsc_rate)
 210                 return cycles;
 211
 212         val = cycles * 1000000000ULL;
 213         return val / tsc_rate;
 214 }
 215
 216 static unsigned long plat_idx_to_val(unsigned int idx)
 217 {
 218         unsigned int error_bits;
 219         unsigned long k, base;
 220
 221         assert(idx < PLAT_NR);
 222
 223         /* MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 224          * all bits of the sample as index */
 225         if (idx < (PLAT_VAL << 1))
 226                 return cycles_to_nsec(idx);
 227
 228         /* Find the group and compute the minimum value of that group */
 229         error_bits = (idx >> PLAT_BITS) - 1;
 230         base = ((unsigned long) 1) << (error_bits + PLAT_BITS);
 231
 232         /* Find its bucket number of the group */
 233         k = idx % PLAT_VAL;
 234
 235         /* Return the mean of the range of the bucket */
 236         return cycles_to_nsec(base + ((k + 0.5) * (1 << error_bits)));
 237 }
 238
 239 unsigned int calculate_clat_percentiles(unsigned long *io_u_plat,
 240                 unsigned long nr, unsigned long **output,
 241                 unsigned long *maxv, unsigned long *minv)
 242 {
 243         unsigned long sum = 0;
 244         unsigned int len = plist_len, i, j = 0;
 245         unsigned long *ovals = NULL;
 246         bool is_last;
 247
 248         *minv = -1UL;
 249         *maxv = 0;
 250
 251         ovals = malloc(len * sizeof(*ovals));
 252         if (!ovals)
 253                 return 0;
 254
 255         /*
 256          * Calculate bucket values, note down max and min values
 257          */
 258         is_last = false;
 259         for (i = 0; i < PLAT_NR && !is_last; i++) {
 260                 sum += io_u_plat[i];
 261                 while (sum >= ((long double) plist[j] / 100.0 * nr)) {
 262                         assert(plist[j] <= 100.0);
 263
 264                         ovals[j] = plat_idx_to_val(i);
 265                         if (ovals[j] < *minv)
 266                                 *minv = ovals[j];
 267                         if (ovals[j] > *maxv)
 268                                 *maxv = ovals[j];
 269
 270                         is_last = (j == len - 1) != 0;
 271                         if (is_last)
 272                                 break;
 273
 274                         j++;
 275                 }
 276         }
 277
 278         if (!is_last)
 279                 fprintf(stderr, "error calculating latency percentiles\n");
 280
 281         *output = ovals;
 282         return len;
 283 }
 284
 285 static void show_clat_percentiles(unsigned long *io_u_plat, unsigned long nr,
 286                                   unsigned int precision)
 287 {
 288         unsigned int divisor, len, i, j = 0;
 289         unsigned long minv, maxv;
 290         unsigned long *ovals;
 291         int per_line, scale_down, time_width;
 292         bool is_last;
 293         char fmt[32];
 294
 295         len = calculate_clat_percentiles(io_u_plat, nr, &ovals, &maxv, &minv);
 296         if (!len || !ovals)
 297                 goto out;
 298
 299         if (!tsc_rate) {
 300                 scale_down = 0;
 301                 divisor = 1;
 302                 printf("    percentiles (tsc ticks):\n     |");
 303         } else if (minv > 2000 && maxv > 99999) {
 304                 scale_down = 1;
 305                 divisor = 1000;
 306                 printf("    percentiles (usec):\n     |");
 307         } else {
 308                 scale_down = 0;
 309                 divisor = 1;
 310                 printf("    percentiles (nsec):\n     |");
 311         }
 312
 313         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 314         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 315                         precision, time_width);
 316         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 317         per_line = (80 - 7) / (precision + 10 + time_width);
 318
 319         for (j = 0; j < len; j++) {
 320                 /* for formatting */
 321                 if (j != 0 && (j % per_line) == 0)
 322                         printf("     |");
 323
 324                 /* end of the list */
 325                 is_last = (j == len - 1) != 0;
 326
 327                 for (i = 0; i < scale_down; i++)
 328                         ovals[j] = (ovals[j] + 999) / 1000;
 329
 330                 printf(fmt, plist[j], ovals[j], is_last ? '\n' : ',');
 331
 332                 if (is_last)
 333                         break;
 334
 335                 if ((j % per_line) == per_line - 1)     /* for formatting */
 336                         printf("\n");
 337         }
 338
 339 out:
 340         free(ovals);
 341 }
 342
 343 #ifdef ARCH_HAVE_CPU_CLOCK
 344 static unsigned int plat_val_to_idx(unsigned long val)
 345 {
 346         unsigned int msb, error_bits, base, offset, idx;
 347
 348         /* Find MSB starting from bit 0 */
 349         if (val == 0)
 350                 msb = 0;
 351         else
 352                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
 353
 354         /*
 355          * MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 356          * all bits of the sample as index
 357          */
 358         if (msb <= PLAT_BITS)
 359                 return val;
 360
 361         /* Compute the number of error bits to discard*/
 362         error_bits = msb - PLAT_BITS;
 363
 364         /* Compute the number of buckets before the group */
 365         base = (error_bits + 1) << PLAT_BITS;
 366
 367         /*
 368          * Discard the error bits and apply the mask to find the
 369          * index for the buckets in the group
 370          */
 371         offset = (PLAT_VAL - 1) & (val >> error_bits);
 372
 373         /* Make sure the index does not exceed (array size - 1) */
 374         idx = (base + offset) < (PLAT_NR - 1) ?
 375                 (base + offset) : (PLAT_NR - 1);
 376
 377         return idx;
 378 }
 379 #endif
 380
 381 static void add_stat(struct submitter *s, int clock_index, int nr)
 382 {
 383 #ifdef ARCH_HAVE_CPU_CLOCK
 384         unsigned long cycles;
 385         unsigned int pidx;
 386
 387         if (!s->finish && clock_index) {
 388                 cycles = get_cpu_clock();
 389                 cycles -= s->clock_batch[clock_index];
 390                 pidx = plat_val_to_idx(cycles);
 391                 s->plat[pidx] += nr;
 392         }
 393 #endif
 394 }
 395
 396 static int io_uring_register_buffers(struct submitter *s)
 397 {
 398         if (do_nop)
 399                 return 0;
 400
 401         return syscall(__NR_io_uring_register, s->ring_fd,
 402                         IORING_REGISTER_BUFFERS, s->iovecs, roundup_pow2(depth));
 403 }
 404
 405 static int io_uring_register_files(struct submitter *s)
 406 {
 407         int i;
 408
 409         if (do_nop)
 410                 return 0;
 411
 412         s->fds = calloc(s->nr_files, sizeof(__s32));
 413         for (i = 0; i < s->nr_files; i++) {
 414                 s->fds[i] = s->files[i].real_fd;
 415                 s->files[i].fixed_fd = i;
 416         }
 417
 418         return syscall(__NR_io_uring_register, s->ring_fd,
 419                         IORING_REGISTER_FILES, s->fds, s->nr_files);
 420 }
 421
 422 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 423 {
 424         int ret;
 425
 426         /*
 427          * Clamp CQ ring size at our SQ ring size, we don't need more entries
 428          * than that.
 429          */
 430         p->flags |= IORING_SETUP_CQSIZE;
 431         p->cq_entries = entries;
 432
 433         p->flags |= IORING_SETUP_COOP_TASKRUN;
 434         p->flags |= IORING_SETUP_SINGLE_ISSUER;
 435         p->flags |= IORING_SETUP_DEFER_TASKRUN;
 436 retry:
 437         ret = syscall(__NR_io_uring_setup, entries, p);
 438         if (!ret)
 439                 return 0;
 440
 441         if (errno == EINVAL && p->flags & IORING_SETUP_COOP_TASKRUN) {
 442                 p->flags &= ~IORING_SETUP_COOP_TASKRUN;
 443                 goto retry;
 444         }
 445         if (errno == EINVAL && p->flags & IORING_SETUP_SINGLE_ISSUER) {
 446                 p->flags &= ~IORING_SETUP_SINGLE_ISSUER;
 447                 goto retry;
 448         }
 449         if (errno == EINVAL && p->flags & IORING_SETUP_DEFER_TASKRUN) {
 450                 p->flags &= ~IORING_SETUP_DEFER_TASKRUN;
 451                 goto retry;
 452         }
 453
 454         return ret;
 455 }
 456
 457 static void io_uring_probe(int fd)
 458 {
 459         struct io_uring_probe *p;
 460         int ret;
 461
 462         p = calloc(1, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 463         if (!p)
 464                 return;
 465
 466         ret = syscall(__NR_io_uring_register, fd, IORING_REGISTER_PROBE, p, 256);
 467         if (ret < 0)
 468                 goto out;
 469
 470         if (IORING_OP_READ > p->ops_len)
 471                 goto out;
 472
 473         if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED))
 474                 vectored = 0;
 475 out:
 476         free(p);
 477 }
 478
 479 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 480                           unsigned int min_complete, unsigned int flags)
 481 {
 482         if (register_ring)
 483                 flags |= IORING_ENTER_REGISTERED_RING;
 484 #ifdef FIO_ARCH_HAS_SYSCALL
 485         return __do_syscall6(__NR_io_uring_enter, s->enter_ring_fd, to_submit,
 486                                 min_complete, flags, NULL, 0);
 487 #else
 488         return syscall(__NR_io_uring_enter, s->enter_ring_fd, to_submit,
 489                         min_complete, flags, NULL, 0);
 490 #endif
 491 }
 492
 493 static unsigned file_depth(struct submitter *s)
 494 {
 495         return (depth + s->nr_files - 1) / s->nr_files;
 496 }
 497
 498 static unsigned long long get_offset(struct submitter *s, struct file *f)
 499 {
 500         unsigned long long offset;
 501         long r;
 502
 503         if (random_io) {
 504                 unsigned long long block;
 505
 506                 r = __rand64(&s->rand_state);
 507                 block = r % f->max_blocks;
 508                 offset = block * (unsigned long long) bs;
 509         } else {
 510                 offset = f->cur_off;
 511                 f->cur_off += bs;
 512                 if (f->cur_off + bs > f->max_size)
 513                         f->cur_off = 0;
 514         }
 515
 516         return offset;
 517 }
 518
 519 static struct file *init_new_io(struct submitter *s)
 520 {
 521         struct file *f;
 522
 523         if (s->nr_files == 1) {
 524                 f = &s->files[0];
 525         } else {
 526                 f = &s->files[s->cur_file];
 527                 if (f->pending_ios >= file_depth(s)) {
 528                         s->cur_file++;
 529                         if (s->cur_file == s->nr_files)
 530                                 s->cur_file = 0;
 531                         f = &s->files[s->cur_file];
 532                 }
 533         }
 534
 535         f->pending_ios++;
 536         return f;
 537 }
 538
 539 static void init_io(struct submitter *s, unsigned index)
 540 {
 541         struct io_uring_sqe *sqe = &s->sqes[index];
 542         struct file *f;
 543
 544         if (do_nop) {
 545                 sqe->opcode = IORING_OP_NOP;
 546                 return;
 547         }
 548
 549         f = init_new_io(s);
 550
 551         if (register_files) {
 552                 sqe->flags = IOSQE_FIXED_FILE;
 553                 sqe->fd = f->fixed_fd;
 554         } else {
 555                 sqe->flags = 0;
 556                 sqe->fd = f->real_fd;
 557         }
 558         if (fixedbufs) {
 559                 sqe->opcode = IORING_OP_READ_FIXED;
 560                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 561                 sqe->len = bs;
 562                 sqe->buf_index = index;
 563         } else if (!vectored) {
 564                 sqe->opcode = IORING_OP_READ;
 565                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 566                 sqe->len = bs;
 567                 sqe->buf_index = 0;
 568         } else {
 569                 sqe->opcode = IORING_OP_READV;
 570                 sqe->addr = (unsigned long) &s->iovecs[index];
 571                 sqe->len = 1;
 572                 sqe->buf_index = 0;
 573         }
 574         sqe->ioprio = 0;
 575         sqe->off = get_offset(s, f);
 576         sqe->user_data = (unsigned long) f->fileno;
 577         if (stats && stats_running)
 578                 sqe->user_data |= ((uint64_t)s->clock_index << 32);
 579 }
 580
 581 static void init_io_pt(struct submitter *s, unsigned index)
 582 {
 583         struct io_uring_sqe *sqe = &s->sqes[index << 1];
 584         unsigned long offset;
 585         struct file *f;
 586         struct nvme_uring_cmd *cmd;
 587         unsigned long long slba;
 588         unsigned long long nlb;
 589
 590         f = init_new_io(s);
 591
 592         offset = get_offset(s, f);
 593
 594         if (register_files) {
 595                 sqe->fd = f->fixed_fd;
 596                 sqe->flags = IOSQE_FIXED_FILE;
 597         } else {
 598                 sqe->fd = f->real_fd;
 599                 sqe->flags = 0;
 600         }
 601         sqe->opcode = IORING_OP_URING_CMD;
 602         sqe->user_data = (unsigned long) f->fileno;
 603         if (stats)
 604                 sqe->user_data |= ((__u64) s->clock_index << 32ULL);
 605         sqe->cmd_op = NVME_URING_CMD_IO;
 606         slba = offset >> f->lba_shift;
 607         nlb = (bs >> f->lba_shift) - 1;
 608         cmd = (struct nvme_uring_cmd *)&sqe->cmd;
 609         /* cdw10 and cdw11 represent starting slba*/
 610         cmd->cdw10 = slba & 0xffffffff;
 611         cmd->cdw11 = slba >> 32;
 612         /* cdw12 represent number of lba to be read*/
 613         cmd->cdw12 = nlb;
 614         cmd->addr = (unsigned long) s->iovecs[index].iov_base;
 615         cmd->data_len = bs;
 616         if (fixedbufs) {
 617                 sqe->uring_cmd_flags = IORING_URING_CMD_FIXED;
 618                 sqe->buf_index = index;
 619         }
 620         cmd->nsid = f->nsid;
 621         cmd->opcode = 2;
 622 }
 623
 624 static int prep_more_ios_uring(struct submitter *s, int max_ios)
 625 {
 626         struct io_sq_ring *ring = &s->sq_ring;
 627         unsigned head, index, tail, next_tail, prepped = 0;
 628
 629         if (sq_thread_poll)
 630                 head = atomic_load_acquire(ring->head);
 631         else
 632                 head = *ring->head;
 633
 634         next_tail = tail = *ring->tail;
 635         do {
 636                 next_tail++;
 637                 if (next_tail == head)
 638                         break;
 639
 640                 index = tail & sq_ring_mask;
 641                 if (pt)
 642                         init_io_pt(s, index);
 643                 else
 644                         init_io(s, index);
 645                 prepped++;
 646                 tail = next_tail;
 647         } while (prepped < max_ios);
 648
 649         if (prepped)
 650                 atomic_store_release(ring->tail, tail);
 651         return prepped;
 652 }
 653
 654 static int get_file_size(struct file *f)
 655 {
 656         struct stat st;
 657
 658         if (fstat(f->real_fd, &st) < 0)
 659                 return -1;
 660         if (pt) {
 661                 __u64 nlba;
 662                 __u32 lbs;
 663                 int ret;
 664
 665                 if (!S_ISCHR(st.st_mode)) {
 666                         fprintf(stderr, "passthrough works with only nvme-ns "
 667                                         "generic devices (/dev/ngXnY)\n");
 668                         return -1;
 669                 }
 670                 ret = nvme_get_info(f->real_fd, &f->nsid, &lbs, &nlba);
 671                 if (ret)
 672                         return -1;
 673                 if ((bs % lbs) != 0) {
 674                         printf("error: bs:%d should be a multiple logical_block_size:%d\n",
 675                                         bs, lbs);
 676                         return -1;
 677                 }
 678                 f->max_blocks = nlba;
 679                 f->max_size = nlba;
 680                 f->lba_shift = ilog2(lbs);
 681                 return 0;
 682         } else if (S_ISBLK(st.st_mode)) {
 683                 unsigned long long bytes;
 684
 685                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 686                         return -1;
 687
 688                 f->max_blocks = bytes / bs;
 689                 f->max_size = bytes;
 690                 return 0;
 691         } else if (S_ISREG(st.st_mode)) {
 692                 f->max_blocks = st.st_size / bs;
 693                 f->max_size = st.st_size;
 694                 return 0;
 695         }
 696
 697         return -1;
 698 }
 699
 700 static int reap_events_uring(struct submitter *s)
 701 {
 702         struct io_cq_ring *ring = &s->cq_ring;
 703         struct io_uring_cqe *cqe;
 704         unsigned head, reaped = 0;
 705         int last_idx = -1, stat_nr = 0;
 706
 707         head = *ring->head;
 708         do {
 709                 struct file *f;
 710
 711                 if (head == atomic_load_acquire(ring->tail))
 712                         break;
 713                 cqe = &ring->cqes[head & cq_ring_mask];
 714                 if (!do_nop) {
 715                         int fileno = cqe->user_data & 0xffffffff;
 716
 717                         f = &s->files[fileno];
 718                         f->pending_ios--;
 719                         if (cqe->res != bs) {
 720                                 printf("io: unexpected ret=%d\n", cqe->res);
 721                                 if (polled && cqe->res == -EOPNOTSUPP)
 722                                         printf("Your filesystem/driver/kernel doesn't support polled IO\n");
 723                                 return -1;
 724                         }
 725                 }
 726                 if (stats) {
 727                         int clock_index = cqe->user_data >> 32;
 728
 729                         if (last_idx != clock_index) {
 730                                 if (last_idx != -1) {
 731                                         add_stat(s, last_idx, stat_nr);
 732                                         stat_nr = 0;
 733                                 }
 734                                 last_idx = clock_index;
 735                         }
 736                         stat_nr++;
 737                 }
 738                 reaped++;
 739                 head++;
 740         } while (1);
 741
 742         if (stat_nr)
 743                 add_stat(s, last_idx, stat_nr);
 744
 745         if (reaped) {
 746                 s->inflight -= reaped;
 747                 atomic_store_release(ring->head, head);
 748         }
 749         return reaped;
 750 }
 751
 752 static int reap_events_uring_pt(struct submitter *s)
 753 {
 754         struct io_cq_ring *ring = &s->cq_ring;
 755         struct io_uring_cqe *cqe;
 756         unsigned head, reaped = 0;
 757         int last_idx = -1, stat_nr = 0;
 758         unsigned index;
 759         int fileno;
 760
 761         head = *ring->head;
 762         do {
 763                 struct file *f;
 764
 765                 if (head == atomic_load_acquire(ring->tail))
 766                         break;
 767                 index = head & cq_ring_mask;
 768                 cqe = &ring->cqes[index << 1];
 769                 fileno = cqe->user_data & 0xffffffff;
 770                 f = &s->files[fileno];
 771                 f->pending_ios--;
 772
 773                 if (cqe->res != 0) {
 774                         printf("io: unexpected ret=%d\n", cqe->res);
 775                         if (polled && cqe->res == -EINVAL)
 776                                 printf("passthrough doesn't support polled IO\n");
 777                         return -1;
 778                 }
 779                 if (stats) {
 780                         int clock_index = cqe->user_data >> 32;
 781
 782                         if (last_idx != clock_index) {
 783                                 if (last_idx != -1) {
 784                                         add_stat(s, last_idx, stat_nr);
 785                                         stat_nr = 0;
 786                                 }
 787                                 last_idx = clock_index;
 788                         }
 789                         stat_nr++;
 790                 }
 791                 reaped++;
 792                 head++;
 793         } while (1);
 794
 795         if (stat_nr)
 796                 add_stat(s, last_idx, stat_nr);
 797
 798         if (reaped) {
 799                 s->inflight -= reaped;
 800                 atomic_store_release(ring->head, head);
 801         }
 802         return reaped;
 803 }
 804
 805 static void set_affinity(struct submitter *s)
 806 {
 807 #ifdef CONFIG_LIBNUMA
 808         struct bitmask *mask;
 809
 810         if (s->numa_node == -1)
 811                 return;
 812
 813         numa_set_preferred(s->numa_node);
 814
 815         mask = numa_allocate_cpumask();
 816         numa_node_to_cpus(s->numa_node, mask);
 817         numa_sched_setaffinity(s->tid, mask);
 818 #endif
 819 }
 820
 821 static int detect_node(struct submitter *s, const char *name)
 822 {
 823 #ifdef CONFIG_LIBNUMA
 824         const char *base = basename(name);
 825         char str[128];
 826         int ret, fd, node;
 827
 828         if (pt)
 829                 sprintf(str, "/sys/class/nvme-generic/%s/device/numa_node", base);
 830         else
 831                 sprintf(str, "/sys/block/%s/device/numa_node", base);
 832         fd = open(str, O_RDONLY);
 833         if (fd < 0)
 834                 return -1;
 835
 836         ret = read(fd, str, sizeof(str));
 837         if (ret < 0) {
 838                 close(fd);
 839                 return -1;
 840         }
 841         node = atoi(str);
 842         s->numa_node = node;
 843         close(fd);
 844 #else
 845         s->numa_node = -1;
 846 #endif
 847         return 0;
 848 }
 849
 850 static int setup_aio(struct submitter *s)
 851 {
 852 #ifdef CONFIG_LIBAIO
 853         if (polled) {
 854                 fprintf(stderr, "aio does not support polled IO\n");
 855                 polled = 0;
 856         }
 857         if (sq_thread_poll) {
 858                 fprintf(stderr, "aio does not support SQPOLL IO\n");
 859                 sq_thread_poll = 0;
 860         }
 861         if (do_nop) {
 862                 fprintf(stderr, "aio does not support polled IO\n");
 863                 do_nop = 0;
 864         }
 865         if (fixedbufs || register_files) {
 866                 fprintf(stderr, "aio does not support registered files or buffers\n");
 867                 fixedbufs = register_files = 0;
 868         }
 869
 870         return io_queue_init(roundup_pow2(depth), &s->aio_ctx);
 871 #else
 872         fprintf(stderr, "Legacy AIO not available on this system/build\n");
 873         errno = EINVAL;
 874         return -1;
 875 #endif
 876 }
 877
 878 static int setup_ring(struct submitter *s)
 879 {
 880         struct io_sq_ring *sring = &s->sq_ring;
 881         struct io_cq_ring *cring = &s->cq_ring;
 882         struct io_uring_params p;
 883         int ret, fd, i;
 884         void *ptr;
 885         size_t len;
 886
 887         memset(&p, 0, sizeof(p));
 888
 889         if (polled && !do_nop)
 890                 p.flags |= IORING_SETUP_IOPOLL;
 891         if (sq_thread_poll) {
 892                 p.flags |= IORING_SETUP_SQPOLL;
 893                 if (sq_thread_cpu != -1) {
 894                         p.flags |= IORING_SETUP_SQ_AFF;
 895                         p.sq_thread_cpu = sq_thread_cpu;
 896                 }
 897         }
 898         if (pt) {
 899                 p.flags |= IORING_SETUP_SQE128;
 900                 p.flags |= IORING_SETUP_CQE32;
 901         }
 902
 903         fd = io_uring_setup(depth, &p);
 904         if (fd < 0) {
 905                 perror("io_uring_setup");
 906                 return 1;
 907         }
 908         s->ring_fd = s->enter_ring_fd = fd;
 909
 910         io_uring_probe(fd);
 911
 912         if (fixedbufs) {
 913                 struct rlimit rlim;
 914
 915                 rlim.rlim_cur = RLIM_INFINITY;
 916                 rlim.rlim_max = RLIM_INFINITY;
 917                 /* ignore potential error, not needed on newer kernels */
 918                 setrlimit(RLIMIT_MEMLOCK, &rlim);
 919
 920                 ret = io_uring_register_buffers(s);
 921                 if (ret < 0) {
 922                         perror("io_uring_register_buffers");
 923                         return 1;
 924                 }
 925         }
 926
 927         if (register_files) {
 928                 ret = io_uring_register_files(s);
 929                 if (ret < 0) {
 930                         perror("io_uring_register_files");
 931                         return 1;
 932                 }
 933         }
 934
 935         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 936                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 937                         IORING_OFF_SQ_RING);
 938         sring->head = ptr + p.sq_off.head;
 939         sring->tail = ptr + p.sq_off.tail;
 940         sring->ring_mask = ptr + p.sq_off.ring_mask;
 941         sring->ring_entries = ptr + p.sq_off.ring_entries;
 942         sring->flags = ptr + p.sq_off.flags;
 943         sring->array = ptr + p.sq_off.array;
 944         sq_ring_mask = *sring->ring_mask;
 945
 946         if (p.flags & IORING_SETUP_SQE128)
 947                 len = 2 * p.sq_entries * sizeof(struct io_uring_sqe);
 948         else
 949                 len = p.sq_entries * sizeof(struct io_uring_sqe);
 950         s->sqes = mmap(0, len,
 951                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 952                         IORING_OFF_SQES);
 953
 954         if (p.flags & IORING_SETUP_CQE32) {
 955                 len = p.cq_off.cqes +
 956                         2 * p.cq_entries * sizeof(struct io_uring_cqe);
 957         } else {
 958                 len = p.cq_off.cqes +
 959                         p.cq_entries * sizeof(struct io_uring_cqe);
 960         }
 961         ptr = mmap(0, len,
 962                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 963                         IORING_OFF_CQ_RING);
 964         cring->head = ptr + p.cq_off.head;
 965         cring->tail = ptr + p.cq_off.tail;
 966         cring->ring_mask = ptr + p.cq_off.ring_mask;
 967         cring->ring_entries = ptr + p.cq_off.ring_entries;
 968         cring->cqes = ptr + p.cq_off.cqes;
 969         cq_ring_mask = *cring->ring_mask;
 970
 971         for (i = 0; i < p.sq_entries; i++)
 972                 sring->array[i] = i;
 973
 974         return 0;
 975 }
 976
 977 static void *allocate_mem(struct submitter *s, int size)
 978 {
 979         void *buf;
 980
 981 #ifdef CONFIG_LIBNUMA
 982         if (s->numa_node != -1)
 983                 return numa_alloc_onnode(size, s->numa_node);
 984 #endif
 985
 986         if (posix_memalign(&buf, t_io_uring_page_size, bs)) {
 987                 printf("failed alloc\n");
 988                 return NULL;
 989         }
 990
 991         return buf;
 992 }
 993
 994 static int submitter_init(struct submitter *s)
 995 {
 996         int i, nr_batch, err;
 997         static int init_printed;
 998         char buf[80];
 999         s->tid = gettid();
1000         printf("submitter=%d, tid=%d, file=%s, node=%d\n", s->index, s->tid,
1001                                                         s->filename, s->numa_node);
1002
1003         set_affinity(s);
1004
1005         __init_rand64(&s->rand_state, s->tid);
1006         srand48(s->tid);
1007
1008         for (i = 0; i < MAX_FDS; i++)
1009                 s->files[i].fileno = i;
1010
1011         for (i = 0; i < roundup_pow2(depth); i++) {
1012                 void *buf;
1013
1014                 buf = allocate_mem(s, bs);
1015                 if (!buf)
1016                         return -1;
1017                 s->iovecs[i].iov_base = buf;
1018                 s->iovecs[i].iov_len = bs;
1019         }
1020
1021         if (use_sync) {
1022                 sprintf(buf, "Engine=preadv2\n");
1023                 err = 0;
1024         } else if (!aio) {
1025                 err = setup_ring(s);
1026                 if (!err)
1027                         sprintf(buf, "Engine=io_uring, sq_ring=%d, cq_ring=%d\n", *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
1028         } else {
1029                 sprintf(buf, "Engine=aio\n");
1030                 err = setup_aio(s);
1031         }
1032         if (err) {
1033                 printf("queue setup failed: %s, %d\n", strerror(errno), err);
1034                 return -1;
1035         }
1036
1037         if (!init_printed) {
1038                 printf("polled=%d, fixedbufs=%d, register_files=%d, buffered=%d, QD=%d\n", polled, fixedbufs, register_files, buffered, depth);
1039                 printf("%s", buf);
1040                 init_printed = 1;
1041         }
1042
1043         if (stats) {
1044                 nr_batch = roundup_pow2(depth / batch_submit);
1045                 if (nr_batch < 2)
1046                         nr_batch = 2;
1047                 s->clock_batch = calloc(nr_batch, sizeof(unsigned long));
1048                 s->clock_index = 1;
1049
1050                 s->plat = calloc(PLAT_NR, sizeof(unsigned long));
1051         } else {
1052                 s->clock_batch = NULL;
1053                 s->plat = NULL;
1054                 nr_batch = 0;
1055         }
1056         /* perform the expensive command initialization part for passthrough here
1057          * rather than in the fast path
1058          */
1059         if (pt) {
1060                 for (i = 0; i < roundup_pow2(depth); i++) {
1061                         struct io_uring_sqe *sqe = &s->sqes[i << 1];
1062
1063                         memset(&sqe->cmd, 0, sizeof(struct nvme_uring_cmd));
1064                 }
1065         }
1066         return nr_batch;
1067 }
1068
1069 #ifdef CONFIG_LIBAIO
1070 static int prep_more_ios_aio(struct submitter *s, int max_ios, struct iocb *iocbs)
1071 {
1072         uint64_t data;
1073         struct file *f;
1074         unsigned index;
1075
1076         index = 0;
1077         while (index < max_ios) {
1078                 struct iocb *iocb = &iocbs[index];
1079
1080                 f = init_new_io(s);
1081
1082                 io_prep_pread(iocb, f->real_fd, s->iovecs[index].iov_base,
1083                                 s->iovecs[index].iov_len, get_offset(s, f));
1084
1085                 data = f->fileno;
1086                 if (stats && stats_running)
1087                         data |= (((uint64_t) s->clock_index) << 32);
1088                 iocb->data = (void *) (uintptr_t) data;
1089                 index++;
1090         }
1091         return index;
1092 }
1093
1094 static int reap_events_aio(struct submitter *s, struct io_event *events, int evs)
1095 {
1096         int last_idx = -1, stat_nr = 0;
1097         int reaped = 0;
1098
1099         while (evs) {
1100                 uint64_t data = (uintptr_t) events[reaped].data;
1101                 struct file *f = &s->files[data & 0xffffffff];
1102
1103                 f->pending_ios--;
1104                 if (events[reaped].res != bs) {
1105                         printf("io: unexpected ret=%ld\n", events[reaped].res);
1106                         return -1;
1107                 }
1108                 if (stats) {
1109                         int clock_index = data >> 32;
1110
1111                         if (last_idx != clock_index) {
1112                                 if (last_idx != -1) {
1113                                         add_stat(s, last_idx, stat_nr);
1114                                         stat_nr = 0;
1115                                 }
1116                                 last_idx = clock_index;
1117                         }
1118                         stat_nr++;
1119                 }
1120                 reaped++;
1121                 evs--;
1122         }
1123
1124         if (stat_nr)
1125                 add_stat(s, last_idx, stat_nr);
1126
1127         s->inflight -= reaped;
1128         s->done += reaped;
1129         return reaped;
1130 }
1131
1132 static void *submitter_aio_fn(void *data)
1133 {
1134         struct submitter *s = data;
1135         int i, ret, prepped;
1136         struct iocb **iocbsptr;
1137         struct iocb *iocbs;
1138         struct io_event *events;
1139 #ifdef ARCH_HAVE_CPU_CLOCK
1140         int nr_batch;
1141 #endif
1142
1143         ret = submitter_init(s);
1144         if (ret < 0)
1145                 goto done;
1146
1147 #ifdef ARCH_HAVE_CPU_CLOCK
1148         nr_batch = ret;
1149 #endif
1150
1151         iocbsptr = calloc(depth, sizeof(struct iocb *));
1152         iocbs = calloc(depth, sizeof(struct iocb));
1153         events = calloc(depth, sizeof(struct io_event));
1154
1155         for (i = 0; i < depth; i++)
1156                 iocbsptr[i] = &iocbs[i];
1157
1158         prepped = 0;
1159         do {
1160                 int to_wait, to_submit, to_prep;
1161
1162                 if (!prepped && s->inflight < depth) {
1163                         to_prep = min(depth - s->inflight, batch_submit);
1164                         prepped = prep_more_ios_aio(s, to_prep, iocbs);
1165 #ifdef ARCH_HAVE_CPU_CLOCK
1166                         if (prepped && stats) {
1167                                 s->clock_batch[s->clock_index] = get_cpu_clock();
1168                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
1169                         }
1170 #endif
1171                 }
1172                 s->inflight += prepped;
1173                 to_submit = prepped;
1174
1175                 if (to_submit && (s->inflight + to_submit <= depth))
1176                         to_wait = 0;
1177                 else
1178                         to_wait = min(s->inflight + to_submit, batch_complete);
1179
1180                 ret = io_submit(s->aio_ctx, to_submit, iocbsptr);
1181                 s->calls++;
1182                 if (ret < 0) {
1183                         perror("io_submit");
1184                         break;
1185                 } else if (ret != to_submit) {
1186                         printf("submitted %d, wanted %d\n", ret, to_submit);
1187                         break;
1188                 }
1189                 prepped = 0;
1190
1191                 while (to_wait) {
1192                         int r;
1193
1194                         s->calls++;
1195                         r = io_getevents(s->aio_ctx, to_wait, to_wait, events, NULL);
1196                         if (r < 0) {
1197                                 perror("io_getevents");
1198                                 break;
1199                         } else if (r != to_wait) {
1200                                 printf("r=%d, wait=%d\n", r, to_wait);
1201                                 break;
1202                         }
1203                         r = reap_events_aio(s, events, r);
1204                         s->reaps += r;
1205                         to_wait -= r;
1206                 }
1207         } while (!s->finish);
1208
1209         free(iocbsptr);
1210         free(iocbs);
1211         free(events);
1212 done:
1213         finish = 1;
1214         return NULL;
1215 }
1216 #endif
1217
1218 static void io_uring_unregister_ring(struct submitter *s)
1219 {
1220         struct io_uring_rsrc_update up = {
1221                 .offset = s->enter_ring_fd,
1222         };
1223
1224         syscall(__NR_io_uring_register, s->ring_fd, IORING_UNREGISTER_RING_FDS,
1225                 &up, 1);
1226 }
1227
1228 static int io_uring_register_ring(struct submitter *s)
1229 {
1230         struct io_uring_rsrc_update up = {
1231                 .data   = s->ring_fd,
1232                 .offset = -1U,
1233         };
1234         int ret;
1235
1236         ret = syscall(__NR_io_uring_register, s->ring_fd,
1237                         IORING_REGISTER_RING_FDS, &up, 1);
1238         if (ret == 1) {
1239                 s->enter_ring_fd = up.offset;
1240                 return 0;
1241         }
1242         register_ring = 0;
1243         return -1;
1244 }
1245
1246 static void *submitter_uring_fn(void *data)
1247 {
1248         struct submitter *s = data;
1249         struct io_sq_ring *ring = &s->sq_ring;
1250         int ret, prepped;
1251 #ifdef ARCH_HAVE_CPU_CLOCK
1252         int nr_batch;
1253 #endif
1254
1255         ret = submitter_init(s);
1256         if (ret < 0)
1257                 goto done;
1258
1259 #ifdef ARCH_HAVE_CPU_CLOCK
1260         nr_batch = ret;
1261 #endif
1262
1263         if (register_ring)
1264                 io_uring_register_ring(s);
1265
1266         prepped = 0;
1267         do {
1268                 int to_wait, to_submit, this_reap, to_prep;
1269                 unsigned ring_flags = 0;
1270
1271                 if (!prepped && s->inflight < depth) {
1272                         to_prep = min(depth - s->inflight, batch_submit);
1273                         prepped = prep_more_ios_uring(s, to_prep);
1274 #ifdef ARCH_HAVE_CPU_CLOCK
1275                         if (prepped && stats) {
1276                                 s->clock_batch[s->clock_index] = get_cpu_clock();
1277                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
1278                         }
1279 #endif
1280                 }
1281                 s->inflight += prepped;
1282 submit_more:
1283                 to_submit = prepped;
1284 submit:
1285                 if (to_submit && (s->inflight + to_submit <= depth))
1286                         to_wait = 0;
1287                 else
1288                         to_wait = min(s->inflight + to_submit, batch_complete);
1289
1290                 /*
1291                  * Only need to call io_uring_enter if we're not using SQ thread
1292                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
1293                  */
1294                 if (sq_thread_poll)
1295                         ring_flags = atomic_load_acquire(ring->flags);
1296                 if (!sq_thread_poll || ring_flags & IORING_SQ_NEED_WAKEUP) {
1297                         unsigned flags = 0;
1298
1299                         if (to_wait)
1300                                 flags = IORING_ENTER_GETEVENTS;
1301                         if (ring_flags & IORING_SQ_NEED_WAKEUP)
1302                                 flags |= IORING_ENTER_SQ_WAKEUP;
1303                         ret = io_uring_enter(s, to_submit, to_wait, flags);
1304                         s->calls++;
1305                 } else {
1306                         /* for SQPOLL, we submitted it all effectively */
1307                         ret = to_submit;
1308                 }
1309
1310                 /*
1311                  * For non SQ thread poll, we already got the events we needed
1312                  * through the io_uring_enter() above. For SQ thread poll, we
1313                  * need to loop here until we find enough events.
1314                  */
1315                 this_reap = 0;
1316                 do {
1317                         int r;
1318
1319                         if (pt)
1320                                 r = reap_events_uring_pt(s);
1321                         else
1322                                 r = reap_events_uring(s);
1323                         if (r == -1) {
1324                                 s->finish = 1;
1325                                 break;
1326                         } else if (r > 0)
1327                                 this_reap += r;
1328                 } while (sq_thread_poll && this_reap < to_wait);
1329                 s->reaps += this_reap;
1330
1331                 if (ret >= 0) {
1332                         if (!ret) {
1333                                 to_submit = 0;
1334                                 if (s->inflight)
1335                                         goto submit;
1336                                 continue;
1337                         } else if (ret < to_submit) {
1338                                 int diff = to_submit - ret;
1339
1340                                 s->done += ret;
1341                                 prepped -= diff;
1342                                 goto submit_more;
1343                         }
1344                         s->done += ret;
1345                         prepped = 0;
1346                         continue;
1347                 } else if (ret < 0) {
1348                         if (errno == EAGAIN) {
1349                                 if (s->finish)
1350                                         break;
1351                                 if (this_reap)
1352                                         goto submit;
1353                                 to_submit = 0;
1354                                 goto submit;
1355                         }
1356                         printf("io_submit: %s\n", strerror(errno));
1357                         break;
1358                 }
1359         } while (!s->finish);
1360
1361         if (register_ring)
1362                 io_uring_unregister_ring(s);
1363
1364 done:
1365         finish = 1;
1366         return NULL;
1367 }
1368
1369 #ifdef CONFIG_PWRITEV2
1370 static void *submitter_sync_fn(void *data)
1371 {
1372         struct submitter *s = data;
1373         int ret;
1374
1375         if (submitter_init(s) < 0)
1376                 goto done;
1377
1378         do {
1379                 uint64_t offset;
1380                 struct file *f;
1381
1382                 f = init_new_io(s);
1383
1384 #ifdef ARCH_HAVE_CPU_CLOCK
1385                 if (stats)
1386                         s->clock_batch[s->clock_index] = get_cpu_clock();
1387 #endif
1388
1389                 s->inflight++;
1390                 s->calls++;
1391
1392                 offset = get_offset(s, f);
1393                 if (polled)
1394                         ret = preadv2(f->real_fd, &s->iovecs[0], 1, offset, RWF_HIPRI);
1395                 else
1396                         ret = preadv2(f->real_fd, &s->iovecs[0], 1, offset, 0);
1397
1398                 if (ret < 0) {
1399                         perror("preadv2");
1400                         break;
1401                 } else if (ret != bs) {
1402                         break;
1403                 }
1404
1405                 s->done++;
1406                 s->inflight--;
1407                 f->pending_ios--;
1408                 if (stats)
1409                         add_stat(s, s->clock_index, 1);
1410         } while (!s->finish);
1411
1412 done:
1413         finish = 1;
1414         return NULL;
1415 }
1416 #else
1417 static void *submitter_sync_fn(void *data)
1418 {
1419         finish = 1;
1420         return NULL;
1421 }
1422 #endif
1423
1424 static struct submitter *get_submitter(int offset)
1425 {
1426         void *ret;
1427
1428         ret = submitter;
1429         if (offset)
1430                 ret += offset * (sizeof(*submitter) + depth * sizeof(struct iovec));
1431         return ret;
1432 }
1433
1434 static void do_finish(const char *reason)
1435 {
1436         int j;
1437
1438         printf("Exiting on %s\n", reason);
1439         for (j = 0; j < nthreads; j++) {
1440                 struct submitter *s = get_submitter(j);
1441                 s->finish = 1;
1442         }
1443         if (max_iops > 1000000) {
1444                 double miops = (double) max_iops / 1000000.0;
1445                 printf("Maximum IOPS=%.2fM\n", miops);
1446         } else if (max_iops > 100000) {
1447                 double kiops = (double) max_iops / 1000.0;
1448                 printf("Maximum IOPS=%.2fK\n", kiops);
1449         } else {
1450                 printf("Maximum IOPS=%lu\n", max_iops);
1451         }
1452         finish = 1;
1453 }
1454
1455 static void sig_int(int sig)
1456 {
1457         do_finish("signal");
1458 }
1459
1460 static void arm_sig_int(void)
1461 {
1462         struct sigaction act;
1463
1464         memset(&act, 0, sizeof(act));
1465         act.sa_handler = sig_int;
1466         act.sa_flags = SA_RESTART;
1467         sigaction(SIGINT, &act, NULL);
1468
1469         /* Windows uses SIGBREAK as a quit signal from other applications */
1470 #ifdef WIN32
1471         sigaction(SIGBREAK, &act, NULL);
1472 #endif
1473 }
1474
1475 static void usage(char *argv, int status)
1476 {
1477         char runtime_str[16];
1478         snprintf(runtime_str, sizeof(runtime_str), "%d", runtime);
1479         printf("%s [options] -- [filenames]\n"
1480                 " -d <int>  : IO Depth, default %d\n"
1481                 " -s <int>  : Batch submit, default %d\n"
1482                 " -c <int>  : Batch complete, default %d\n"
1483                 " -b <int>  : Block size, default %d\n"
1484                 " -p <bool> : Polled IO, default %d\n"
1485                 " -B <bool> : Fixed buffers, default %d\n"
1486                 " -F <bool> : Register files, default %d\n"
1487                 " -n <int>  : Number of threads, default %d\n"
1488                 " -O <bool> : Use O_DIRECT, default %d\n"
1489                 " -N <bool> : Perform just no-op requests, default %d\n"
1490                 " -t <bool> : Track IO latencies, default %d\n"
1491                 " -T <int>  : TSC rate in HZ\n"
1492                 " -r <int>  : Runtime in seconds, default %s\n"
1493                 " -R <bool> : Use random IO, default %d\n"
1494                 " -a <bool> : Use legacy aio, default %d\n"
1495                 " -S <bool> : Use sync IO (preadv2), default %d\n"
1496                 " -X <bool> : Use registered ring %d\n"
1497                 " -P <bool> : Automatically place on device home node %d\n"
1498                 " -u <bool> : Use nvme-passthrough I/O, default %d\n",
1499                 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled,
1500                 fixedbufs, register_files, nthreads, !buffered, do_nop,
1501                 stats, runtime == 0 ? "unlimited" : runtime_str, random_io, aio,
1502                 use_sync, register_ring, numa_placement, pt);
1503         exit(status);
1504 }
1505
1506 static void read_tsc_rate(void)
1507 {
1508         char buffer[32];
1509         int fd, ret;
1510
1511         if (tsc_rate)
1512                 return;
1513
1514         fd = open(TSC_RATE_FILE, O_RDONLY);
1515         if (fd < 0)
1516                 return;
1517
1518         ret = read(fd, buffer, sizeof(buffer));
1519         if (ret < 0) {
1520                 close(fd);
1521                 return;
1522         }
1523
1524         tsc_rate = strtoul(buffer, NULL, 10);
1525         printf("Using TSC rate %luHz\n", tsc_rate);
1526         close(fd);
1527 }
1528
1529 static void write_tsc_rate(void)
1530 {
1531         char buffer[32];
1532         struct stat sb;
1533         int fd, ret;
1534
1535         if (!stat(TSC_RATE_FILE, &sb))
1536                 return;
1537
1538         fd = open(TSC_RATE_FILE, O_WRONLY | O_CREAT, 0644);
1539         if (fd < 0)
1540                 return;
1541
1542         memset(buffer, 0, sizeof(buffer));
1543         sprintf(buffer, "%lu", tsc_rate);
1544         ret = write(fd, buffer, strlen(buffer));
1545         if (ret < 0)
1546                 perror("write");
1547         close(fd);
1548 }
1549
1550 int main(int argc, char *argv[])
1551 {
1552         struct submitter *s;
1553         unsigned long done, calls, reap;
1554         int i, j, flags, fd, opt, threads_per_f, threads_rem = 0, nfiles;
1555         struct file f;
1556         void *ret;
1557
1558         if (!do_nop && argc < 2)
1559                 usage(argv[0], 1);
1560
1561         while ((opt = getopt(argc, argv, "d:s:c:b:p:B:F:n:N:O:t:T:a:r:D:R:X:S:P:u:h?")) != -1) {
1562                 switch (opt) {
1563                 case 'a':
1564                         aio = !!atoi(optarg);
1565                         break;
1566                 case 'd':
1567                         depth = atoi(optarg);
1568                         break;
1569                 case 's':
1570                         batch_submit = atoi(optarg);
1571                         if (!batch_submit)
1572                                 batch_submit = 1;
1573                         break;
1574                 case 'c':
1575                         batch_complete = atoi(optarg);
1576                         if (!batch_complete)
1577                                 batch_complete = 1;
1578                         break;
1579                 case 'b':
1580                         bs = atoi(optarg);
1581                         break;
1582                 case 'p':
1583                         polled = !!atoi(optarg);
1584                         break;
1585                 case 'B':
1586                         fixedbufs = !!atoi(optarg);
1587                         break;
1588                 case 'F':
1589                         register_files = !!atoi(optarg);
1590                         break;
1591                 case 'n':
1592                         nthreads = atoi(optarg);
1593                         if (!nthreads) {
1594                                 printf("Threads must be non-zero\n");
1595                                 usage(argv[0], 1);
1596                         }
1597                         break;
1598                 case 'N':
1599                         do_nop = !!atoi(optarg);
1600                         break;
1601                 case 'O':
1602                         buffered = !atoi(optarg);
1603                         break;
1604                 case 't':
1605 #ifndef ARCH_HAVE_CPU_CLOCK
1606                         fprintf(stderr, "Stats not supported on this CPU\n");
1607                         return 1;
1608 #endif
1609                         stats = !!atoi(optarg);
1610                         break;
1611                 case 'T':
1612 #ifndef ARCH_HAVE_CPU_CLOCK
1613                         fprintf(stderr, "Stats not supported on this CPU\n");
1614                         return 1;
1615 #endif
1616                         tsc_rate = strtoul(optarg, NULL, 10);
1617                         write_tsc_rate();
1618                         break;
1619                 case 'r':
1620                         runtime = atoi(optarg);
1621                         break;
1622                 case 'R':
1623                         random_io = !!atoi(optarg);
1624                         break;
1625                 case 'X':
1626                         register_ring = !!atoi(optarg);
1627                         break;
1628                 case 'S':
1629 #ifdef CONFIG_PWRITEV2
1630                         use_sync = !!atoi(optarg);
1631 #else
1632                         fprintf(stderr, "preadv2 not supported\n");
1633                         exit(1);
1634 #endif
1635                         break;
1636                 case 'P':
1637                         numa_placement = !!atoi(optarg);
1638                         break;
1639                 case 'u':
1640                         pt = !!atoi(optarg);
1641                         break;
1642                 case 'h':
1643                 case '?':
1644                 default:
1645                         usage(argv[0], 0);
1646                         break;
1647                 }
1648         }
1649
1650         if (stats)
1651                 read_tsc_rate();
1652
1653         if (batch_complete > depth)
1654                 batch_complete = depth;
1655         if (batch_submit > depth)
1656                 batch_submit = depth;
1657
1658         submitter = calloc(nthreads, sizeof(*submitter) +
1659                                 roundup_pow2(depth) * sizeof(struct iovec));
1660         for (j = 0; j < nthreads; j++) {
1661                 s = get_submitter(j);
1662                 s->numa_node = -1;
1663                 s->index = j;
1664                 s->done = s->calls = s->reaps = 0;
1665         }
1666
1667         flags = O_RDONLY | O_NOATIME;
1668         if (!buffered)
1669                 flags |= O_DIRECT;
1670
1671         j = 0;
1672         i = optind;
1673         nfiles = argc - i;
1674         if (!do_nop) {
1675                 if (!nfiles) {
1676                         printf("No files specified\n");
1677                         usage(argv[0], 1);
1678                 }
1679                 threads_per_f = nthreads / nfiles;
1680                 /* make sure each thread gets assigned files */
1681                 if (threads_per_f == 0) {
1682                         threads_per_f = 1;
1683                 } else {
1684                         threads_rem = nthreads - threads_per_f * nfiles;
1685                 }
1686         }
1687         while (!do_nop && i < argc) {
1688                 int k, limit;
1689
1690                 memset(&f, 0, sizeof(f));
1691
1692                 fd = open(argv[i], flags);
1693                 if (fd < 0) {
1694                         perror("open");
1695                         return 1;
1696                 }
1697                 f.real_fd = fd;
1698                 if (get_file_size(&f)) {
1699                         printf("failed getting size of device/file\n");
1700                         return 1;
1701                 }
1702                 if (f.max_blocks <= 1) {
1703                         printf("Zero file/device size?\n");
1704                         return 1;
1705                 }
1706                 f.max_blocks--;
1707
1708                 limit = threads_per_f;
1709                 limit += threads_rem > 0 ? 1 : 0;
1710                 for (k = 0; k < limit; k++) {
1711                         s = get_submitter((j + k) % nthreads);
1712
1713                         if (s->nr_files == MAX_FDS) {
1714                                 printf("Max number of files (%d) reached\n", MAX_FDS);
1715                                 break;
1716                         }
1717
1718                         memcpy(&s->files[s->nr_files], &f, sizeof(f));
1719
1720                         if (numa_placement)
1721                                 detect_node(s, argv[i]);
1722
1723                         s->filename = argv[i];
1724                         s->nr_files++;
1725                 }
1726                 threads_rem--;
1727                 i++;
1728                 j += limit;
1729         }
1730
1731         arm_sig_int();
1732
1733         t_io_uring_page_size = sysconf(_SC_PAGESIZE);
1734         if (t_io_uring_page_size < 0)
1735                 t_io_uring_page_size = 4096;
1736
1737         for (j = 0; j < nthreads; j++) {
1738                 s = get_submitter(j);
1739                 if (use_sync)
1740                         pthread_create(&s->thread, NULL, submitter_sync_fn, s);
1741                 else if (!aio)
1742                         pthread_create(&s->thread, NULL, submitter_uring_fn, s);
1743 #ifdef CONFIG_LIBAIO
1744                 else
1745                         pthread_create(&s->thread, NULL, submitter_aio_fn, s);
1746 #endif
1747         }
1748
1749         reap = calls = done = 0;
1750         do {
1751                 unsigned long this_done = 0;
1752                 unsigned long this_reap = 0;
1753                 unsigned long this_call = 0;
1754                 unsigned long rpc = 0, ipc = 0;
1755                 unsigned long iops, bw;
1756
1757                 sleep(1);
1758                 if (runtime && !--runtime)
1759                         do_finish("timeout");
1760
1761                 /* don't print partial run, if interrupted by signal */
1762                 if (finish)
1763                         break;
1764
1765                 /* one second in to the run, enable stats */
1766                 if (stats)
1767                         stats_running = 1;
1768
1769                 for (j = 0; j < nthreads; j++) {
1770                         s = get_submitter(j);
1771                         this_done += s->done;
1772                         this_call += s->calls;
1773                         this_reap += s->reaps;
1774                 }
1775                 if (this_call - calls) {
1776                         rpc = (this_done - done) / (this_call - calls);
1777                         ipc = (this_reap - reap) / (this_call - calls);
1778                 } else
1779                         rpc = ipc = -1;
1780                 iops = this_done - done;
1781                 if (bs > 1048576)
1782                         bw = iops * (bs / 1048576);
1783                 else
1784                         bw = iops / (1048576 / bs);
1785                 if (iops > 1000000) {
1786                         double miops = (double) iops / 1000000.0;
1787                         printf("IOPS=%.2fM, ", miops);
1788                 } else if (iops > 100000) {
1789                         double kiops = (double) iops / 1000.0;
1790                         printf("IOPS=%.2fK, ", kiops);
1791                 } else {
1792                         printf("IOPS=%lu, ", iops);
1793                 }
1794                 max_iops = max(max_iops, iops);
1795                 if (!do_nop) {
1796                         if (bw > 2000) {
1797                                 double bw_g = (double) bw / 1000.0;
1798
1799                                 printf("BW=%.2fGiB/s, ", bw_g);
1800                         } else {
1801                                 printf("BW=%luMiB/s, ", bw);
1802                         }
1803                 }
1804                 printf("IOS/call=%ld/%ld\n", rpc, ipc);
1805                 done = this_done;
1806                 calls = this_call;
1807                 reap = this_reap;
1808         } while (!finish);
1809
1810         for (j = 0; j < nthreads; j++) {
1811                 s = get_submitter(j);
1812                 pthread_join(s->thread, &ret);
1813                 close(s->ring_fd);
1814
1815                 if (stats) {
1816                         unsigned long nr;
1817
1818                         printf("%d: Latency percentiles:\n", s->tid);
1819                         for (i = 0, nr = 0; i < PLAT_NR; i++)
1820                                 nr += s->plat[i];
1821                         show_clat_percentiles(s->plat, nr, 4);
1822                         free(s->clock_batch);
1823                         free(s->plat);
1824                 }
1825         }
1826
1827         free(submitter);
1828         return 0;
1829 }