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 dma_map;             /* pre-map DMA buffers */
 133 static int register_files = 1;  /* use fixed files */
 134 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
 135 static int sq_thread_poll = 0;  /* use kernel submission/poller thread */
 136 static int sq_thread_cpu = -1;  /* pin above thread to this CPU */
 137 static int do_nop = 0;          /* no-op SQ ring commands */
 138 static int nthreads = 1;
 139 static int stats = 0;           /* generate IO stats */
 140 static int aio = 0;             /* use libaio */
 141 static int runtime = 0;         /* runtime */
 142 static int random_io = 1;       /* random or sequential IO */
 143 static int register_ring = 1;   /* register ring */
 144 static int use_sync = 0;        /* use preadv2 */
 145 static int numa_placement = 0;  /* set to node of device */
 146 static int pt = 0;              /* passthrough I/O or not */
 147
 148 static unsigned long tsc_rate;
 149
 150 #define TSC_RATE_FILE   "tsc-rate"
 151
 152 static int vectored = 1;
 153
 154 static float plist[] = { 1.0, 5.0, 10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0,
 155                         80.0, 90.0, 95.0, 99.0, 99.5, 99.9, 99.95, 99.99 };
 156 static int plist_len = 17;
 157
 158 #ifndef IORING_REGISTER_MAP_BUFFERS
 159 #define IORING_REGISTER_MAP_BUFFERS     22
 160 struct io_uring_map_buffers {
 161         __s32   fd;
 162         __u32   buf_start;
 163         __u32   buf_end;
 164         __u32   flags;
 165         __u64   rsvd[2];
 166 };
 167 #endif
 168
 169 static int nvme_identify(int fd, __u32 nsid, enum nvme_identify_cns cns,
 170                          enum nvme_csi csi, void *data)
 171 {
 172         struct nvme_passthru_cmd cmd = {
 173                 .opcode         = nvme_admin_identify,
 174                 .nsid           = nsid,
 175                 .addr           = (__u64)(uintptr_t)data,
 176                 .data_len       = NVME_IDENTIFY_DATA_SIZE,
 177                 .cdw10          = cns,
 178                 .cdw11          = csi << NVME_IDENTIFY_CSI_SHIFT,
 179                 .timeout_ms     = NVME_DEFAULT_IOCTL_TIMEOUT,
 180         };
 181
 182         return ioctl(fd, NVME_IOCTL_ADMIN_CMD, &cmd);
 183 }
 184
 185 static int nvme_get_info(int fd, __u32 *nsid, __u32 *lba_sz, __u64 *nlba)
 186 {
 187         struct nvme_id_ns ns;
 188         int namespace_id;
 189         int err;
 190
 191         namespace_id = ioctl(fd, NVME_IOCTL_ID);
 192         if (namespace_id < 0) {
 193                 fprintf(stderr, "error failed to fetch namespace-id\n");
 194                 close(fd);
 195                 return -errno;
 196         }
 197
 198         /*
 199          * Identify namespace to get namespace-id, namespace size in LBA's
 200          * and LBA data size.
 201          */
 202         err = nvme_identify(fd, namespace_id, NVME_IDENTIFY_CNS_NS,
 203                                 NVME_CSI_NVM, &ns);
 204         if (err) {
 205                 fprintf(stderr, "error failed to fetch identify namespace\n");
 206                 close(fd);
 207                 return err;
 208         }
 209
 210         *nsid = namespace_id;
 211         *lba_sz = 1 << ns.lbaf[(ns.flbas & 0x0f)].ds;
 212         *nlba = ns.nsze;
 213
 214         return 0;
 215 }
 216
 217 static unsigned long cycles_to_nsec(unsigned long cycles)
 218 {
 219         uint64_t val;
 220
 221         if (!tsc_rate)
 222                 return cycles;
 223
 224         val = cycles * 1000000000ULL;
 225         return val / tsc_rate;
 226 }
 227
 228 static unsigned long plat_idx_to_val(unsigned int idx)
 229 {
 230         unsigned int error_bits;
 231         unsigned long k, base;
 232
 233         assert(idx < PLAT_NR);
 234
 235         /* MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 236          * all bits of the sample as index */
 237         if (idx < (PLAT_VAL << 1))
 238                 return cycles_to_nsec(idx);
 239
 240         /* Find the group and compute the minimum value of that group */
 241         error_bits = (idx >> PLAT_BITS) - 1;
 242         base = ((unsigned long) 1) << (error_bits + PLAT_BITS);
 243
 244         /* Find its bucket number of the group */
 245         k = idx % PLAT_VAL;
 246
 247         /* Return the mean of the range of the bucket */
 248         return cycles_to_nsec(base + ((k + 0.5) * (1 << error_bits)));
 249 }
 250
 251 unsigned int calculate_clat_percentiles(unsigned long *io_u_plat,
 252                 unsigned long nr, unsigned long **output,
 253                 unsigned long *maxv, unsigned long *minv)
 254 {
 255         unsigned long sum = 0;
 256         unsigned int len = plist_len, i, j = 0;
 257         unsigned long *ovals = NULL;
 258         bool is_last;
 259
 260         *minv = -1UL;
 261         *maxv = 0;
 262
 263         ovals = malloc(len * sizeof(*ovals));
 264         if (!ovals)
 265                 return 0;
 266
 267         /*
 268          * Calculate bucket values, note down max and min values
 269          */
 270         is_last = false;
 271         for (i = 0; i < PLAT_NR && !is_last; i++) {
 272                 sum += io_u_plat[i];
 273                 while (sum >= ((long double) plist[j] / 100.0 * nr)) {
 274                         assert(plist[j] <= 100.0);
 275
 276                         ovals[j] = plat_idx_to_val(i);
 277                         if (ovals[j] < *minv)
 278                                 *minv = ovals[j];
 279                         if (ovals[j] > *maxv)
 280                                 *maxv = ovals[j];
 281
 282                         is_last = (j == len - 1) != 0;
 283                         if (is_last)
 284                                 break;
 285
 286                         j++;
 287                 }
 288         }
 289
 290         if (!is_last)
 291                 fprintf(stderr, "error calculating latency percentiles\n");
 292
 293         *output = ovals;
 294         return len;
 295 }
 296
 297 static void show_clat_percentiles(unsigned long *io_u_plat, unsigned long nr,
 298                                   unsigned int precision)
 299 {
 300         unsigned int divisor, len, i, j = 0;
 301         unsigned long minv, maxv;
 302         unsigned long *ovals;
 303         int per_line, scale_down, time_width;
 304         bool is_last;
 305         char fmt[32];
 306
 307         len = calculate_clat_percentiles(io_u_plat, nr, &ovals, &maxv, &minv);
 308         if (!len || !ovals)
 309                 goto out;
 310
 311         if (!tsc_rate) {
 312                 scale_down = 0;
 313                 divisor = 1;
 314                 printf("    percentiles (tsc ticks):\n     |");
 315         } else if (minv > 2000 && maxv > 99999) {
 316                 scale_down = 1;
 317                 divisor = 1000;
 318                 printf("    percentiles (usec):\n     |");
 319         } else {
 320                 scale_down = 0;
 321                 divisor = 1;
 322                 printf("    percentiles (nsec):\n     |");
 323         }
 324
 325         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 326         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 327                         precision, time_width);
 328         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 329         per_line = (80 - 7) / (precision + 10 + time_width);
 330
 331         for (j = 0; j < len; j++) {
 332                 /* for formatting */
 333                 if (j != 0 && (j % per_line) == 0)
 334                         printf("     |");
 335
 336                 /* end of the list */
 337                 is_last = (j == len - 1) != 0;
 338
 339                 for (i = 0; i < scale_down; i++)
 340                         ovals[j] = (ovals[j] + 999) / 1000;
 341
 342                 printf(fmt, plist[j], ovals[j], is_last ? '\n' : ',');
 343
 344                 if (is_last)
 345                         break;
 346
 347                 if ((j % per_line) == per_line - 1)     /* for formatting */
 348                         printf("\n");
 349         }
 350
 351 out:
 352         free(ovals);
 353 }
 354
 355 #ifdef ARCH_HAVE_CPU_CLOCK
 356 static unsigned int plat_val_to_idx(unsigned long val)
 357 {
 358         unsigned int msb, error_bits, base, offset, idx;
 359
 360         /* Find MSB starting from bit 0 */
 361         if (val == 0)
 362                 msb = 0;
 363         else
 364                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
 365
 366         /*
 367          * MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 368          * all bits of the sample as index
 369          */
 370         if (msb <= PLAT_BITS)
 371                 return val;
 372
 373         /* Compute the number of error bits to discard*/
 374         error_bits = msb - PLAT_BITS;
 375
 376         /* Compute the number of buckets before the group */
 377         base = (error_bits + 1) << PLAT_BITS;
 378
 379         /*
 380          * Discard the error bits and apply the mask to find the
 381          * index for the buckets in the group
 382          */
 383         offset = (PLAT_VAL - 1) & (val >> error_bits);
 384
 385         /* Make sure the index does not exceed (array size - 1) */
 386         idx = (base + offset) < (PLAT_NR - 1) ?
 387                 (base + offset) : (PLAT_NR - 1);
 388
 389         return idx;
 390 }
 391 #endif
 392
 393 static void add_stat(struct submitter *s, int clock_index, int nr)
 394 {
 395 #ifdef ARCH_HAVE_CPU_CLOCK
 396         unsigned long cycles;
 397         unsigned int pidx;
 398
 399         if (!s->finish && clock_index) {
 400                 cycles = get_cpu_clock();
 401                 cycles -= s->clock_batch[clock_index];
 402                 pidx = plat_val_to_idx(cycles);
 403                 s->plat[pidx] += nr;
 404         }
 405 #endif
 406 }
 407
 408 static int io_uring_map_buffers(struct submitter *s)
 409 {
 410         struct io_uring_map_buffers map = {
 411                 .fd             = s->files[0].real_fd,
 412                 .buf_end        = depth,
 413         };
 414
 415         if (do_nop)
 416                 return 0;
 417         if (s->nr_files > 1)
 418                 fprintf(stdout, "Mapping buffers may not work with multiple files\n");
 419
 420         return syscall(__NR_io_uring_register, s->ring_fd,
 421                         IORING_REGISTER_MAP_BUFFERS, &map, 1);
 422 }
 423
 424 static int io_uring_register_buffers(struct submitter *s)
 425 {
 426         if (do_nop)
 427                 return 0;
 428
 429         return syscall(__NR_io_uring_register, s->ring_fd,
 430                         IORING_REGISTER_BUFFERS, s->iovecs, roundup_pow2(depth));
 431 }
 432
 433 static int io_uring_register_files(struct submitter *s)
 434 {
 435         int i;
 436
 437         if (do_nop)
 438                 return 0;
 439
 440         s->fds = calloc(s->nr_files, sizeof(__s32));
 441         for (i = 0; i < s->nr_files; i++) {
 442                 s->fds[i] = s->files[i].real_fd;
 443                 s->files[i].fixed_fd = i;
 444         }
 445
 446         return syscall(__NR_io_uring_register, s->ring_fd,
 447                         IORING_REGISTER_FILES, s->fds, s->nr_files);
 448 }
 449
 450 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 451 {
 452         int ret;
 453
 454         /*
 455          * Clamp CQ ring size at our SQ ring size, we don't need more entries
 456          * than that.
 457          */
 458         p->flags |= IORING_SETUP_CQSIZE;
 459         p->cq_entries = entries;
 460
 461         p->flags |= IORING_SETUP_COOP_TASKRUN;
 462         p->flags |= IORING_SETUP_SINGLE_ISSUER;
 463         p->flags |= IORING_SETUP_DEFER_TASKRUN;
 464 retry:
 465         ret = syscall(__NR_io_uring_setup, entries, p);
 466         if (!ret)
 467                 return 0;
 468
 469         if (errno == EINVAL && p->flags & IORING_SETUP_COOP_TASKRUN) {
 470                 p->flags &= ~IORING_SETUP_COOP_TASKRUN;
 471                 goto retry;
 472         }
 473         if (errno == EINVAL && p->flags & IORING_SETUP_SINGLE_ISSUER) {
 474                 p->flags &= ~IORING_SETUP_SINGLE_ISSUER;
 475                 goto retry;
 476         }
 477         if (errno == EINVAL && p->flags & IORING_SETUP_DEFER_TASKRUN) {
 478                 p->flags &= ~IORING_SETUP_DEFER_TASKRUN;
 479                 goto retry;
 480         }
 481
 482         return ret;
 483 }
 484
 485 static void io_uring_probe(int fd)
 486 {
 487         struct io_uring_probe *p;
 488         int ret;
 489
 490         p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 491         if (!p)
 492                 return;
 493
 494         memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 495         ret = syscall(__NR_io_uring_register, fd, IORING_REGISTER_PROBE, p, 256);
 496         if (ret < 0)
 497                 goto out;
 498
 499         if (IORING_OP_READ > p->ops_len)
 500                 goto out;
 501
 502         if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED))
 503                 vectored = 0;
 504 out:
 505         free(p);
 506 }
 507
 508 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 509                           unsigned int min_complete, unsigned int flags)
 510 {
 511         if (register_ring)
 512                 flags |= IORING_ENTER_REGISTERED_RING;
 513 #ifdef FIO_ARCH_HAS_SYSCALL
 514         return __do_syscall6(__NR_io_uring_enter, s->enter_ring_fd, to_submit,
 515                                 min_complete, flags, NULL, 0);
 516 #else
 517         return syscall(__NR_io_uring_enter, s->enter_ring_fd, to_submit,
 518                         min_complete, flags, NULL, 0);
 519 #endif
 520 }
 521
 522 static unsigned file_depth(struct submitter *s)
 523 {
 524         return (depth + s->nr_files - 1) / s->nr_files;
 525 }
 526
 527 static unsigned long long get_offset(struct submitter *s, struct file *f)
 528 {
 529         unsigned long long offset;
 530         long r;
 531
 532         if (random_io) {
 533                 r = __rand64(&s->rand_state);
 534                 offset = (r % (f->max_blocks - 1)) * bs;
 535         } else {
 536                 offset = f->cur_off;
 537                 f->cur_off += bs;
 538                 if (f->cur_off + bs > f->max_size)
 539                         f->cur_off = 0;
 540         }
 541
 542         return offset;
 543 }
 544
 545 static void init_io(struct submitter *s, unsigned index)
 546 {
 547         struct io_uring_sqe *sqe = &s->sqes[index];
 548         struct file *f;
 549
 550         if (do_nop) {
 551                 sqe->opcode = IORING_OP_NOP;
 552                 return;
 553         }
 554
 555         if (s->nr_files == 1) {
 556                 f = &s->files[0];
 557         } else {
 558                 f = &s->files[s->cur_file];
 559                 if (f->pending_ios >= file_depth(s)) {
 560                         s->cur_file++;
 561                         if (s->cur_file == s->nr_files)
 562                                 s->cur_file = 0;
 563                         f = &s->files[s->cur_file];
 564                 }
 565         }
 566         f->pending_ios++;
 567
 568         if (register_files) {
 569                 sqe->flags = IOSQE_FIXED_FILE;
 570                 sqe->fd = f->fixed_fd;
 571         } else {
 572                 sqe->flags = 0;
 573                 sqe->fd = f->real_fd;
 574         }
 575         if (fixedbufs) {
 576                 sqe->opcode = IORING_OP_READ_FIXED;
 577                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 578                 sqe->len = bs;
 579                 sqe->buf_index = index;
 580         } else if (!vectored) {
 581                 sqe->opcode = IORING_OP_READ;
 582                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 583                 sqe->len = bs;
 584                 sqe->buf_index = 0;
 585         } else {
 586                 sqe->opcode = IORING_OP_READV;
 587                 sqe->addr = (unsigned long) &s->iovecs[index];
 588                 sqe->len = 1;
 589                 sqe->buf_index = 0;
 590         }
 591         sqe->ioprio = 0;
 592         sqe->off = get_offset(s, f);
 593         sqe->user_data = (unsigned long) f->fileno;
 594         if (stats && stats_running)
 595                 sqe->user_data |= ((uint64_t)s->clock_index << 32);
 596 }
 597
 598 static void init_io_pt(struct submitter *s, unsigned index)
 599 {
 600         struct io_uring_sqe *sqe = &s->sqes[index << 1];
 601         unsigned long offset;
 602         struct file *f;
 603         struct nvme_uring_cmd *cmd;
 604         unsigned long long slba;
 605         unsigned long long nlb;
 606         long r;
 607
 608         if (s->nr_files == 1) {
 609                 f = &s->files[0];
 610         } else {
 611                 f = &s->files[s->cur_file];
 612                 if (f->pending_ios >= file_depth(s)) {
 613                         s->cur_file++;
 614                         if (s->cur_file == s->nr_files)
 615                                 s->cur_file = 0;
 616                         f = &s->files[s->cur_file];
 617                 }
 618         }
 619         f->pending_ios++;
 620
 621         if (random_io) {
 622                 r = __rand64(&s->rand_state);
 623                 offset = (r % (f->max_blocks - 1)) * bs;
 624         } else {
 625                 offset = f->cur_off;
 626                 f->cur_off += bs;
 627                 if (f->cur_off + bs > f->max_size)
 628                         f->cur_off = 0;
 629         }
 630
 631         if (register_files) {
 632                 sqe->fd = f->fixed_fd;
 633                 sqe->flags = IOSQE_FIXED_FILE;
 634         } else {
 635                 sqe->fd = f->real_fd;
 636                 sqe->flags = 0;
 637         }
 638         sqe->opcode = IORING_OP_URING_CMD;
 639         sqe->user_data = (unsigned long) f->fileno;
 640         if (stats)
 641                 sqe->user_data |= ((__u64) s->clock_index << 32ULL);
 642         sqe->cmd_op = NVME_URING_CMD_IO;
 643         slba = offset >> f->lba_shift;
 644         nlb = (bs >> f->lba_shift) - 1;
 645         cmd = (struct nvme_uring_cmd *)&sqe->cmd;
 646         /* cdw10 and cdw11 represent starting slba*/
 647         cmd->cdw10 = slba & 0xffffffff;
 648         cmd->cdw11 = slba >> 32;
 649         /* cdw12 represent number of lba to be read*/
 650         cmd->cdw12 = nlb;
 651         cmd->addr = (unsigned long) s->iovecs[index].iov_base;
 652         cmd->data_len = bs;
 653         cmd->nsid = f->nsid;
 654         cmd->opcode = 2;
 655 }
 656
 657 static int prep_more_ios_uring(struct submitter *s, int max_ios)
 658 {
 659         struct io_sq_ring *ring = &s->sq_ring;
 660         unsigned index, tail, next_tail, prepped = 0;
 661
 662         next_tail = tail = *ring->tail;
 663         do {
 664                 next_tail++;
 665                 if (next_tail == atomic_load_acquire(ring->head))
 666                         break;
 667
 668                 index = tail & sq_ring_mask;
 669                 if (pt)
 670                         init_io_pt(s, index);
 671                 else
 672                         init_io(s, index);
 673                 ring->array[index] = index;
 674                 prepped++;
 675                 tail = next_tail;
 676         } while (prepped < max_ios);
 677
 678         if (prepped)
 679                 atomic_store_release(ring->tail, tail);
 680         return prepped;
 681 }
 682
 683 static int get_file_size(struct file *f)
 684 {
 685         struct stat st;
 686
 687         if (fstat(f->real_fd, &st) < 0)
 688                 return -1;
 689         if (pt) {
 690                 __u64 nlba;
 691                 __u32 lbs;
 692                 int ret;
 693
 694                 if (!S_ISCHR(st.st_mode)) {
 695                         fprintf(stderr, "passthrough works with only nvme-ns "
 696                                         "generic devices (/dev/ngXnY)\n");
 697                         return -1;
 698                 }
 699                 ret = nvme_get_info(f->real_fd, &f->nsid, &lbs, &nlba);
 700                 if (ret)
 701                         return -1;
 702                 if ((bs % lbs) != 0) {
 703                         printf("error: bs:%d should be a multiple logical_block_size:%d\n",
 704                                         bs, lbs);
 705                         return -1;
 706                 }
 707                 f->max_blocks = nlba / bs;
 708                 f->max_size = nlba;
 709                 f->lba_shift = ilog2(lbs);
 710                 return 0;
 711         } else if (S_ISBLK(st.st_mode)) {
 712                 unsigned long long bytes;
 713
 714                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 715                         return -1;
 716
 717                 f->max_blocks = bytes / bs;
 718                 f->max_size = bytes;
 719                 return 0;
 720         } else if (S_ISREG(st.st_mode)) {
 721                 f->max_blocks = st.st_size / bs;
 722                 f->max_size = st.st_size;
 723                 return 0;
 724         }
 725
 726         return -1;
 727 }
 728
 729 static int reap_events_uring(struct submitter *s)
 730 {
 731         struct io_cq_ring *ring = &s->cq_ring;
 732         struct io_uring_cqe *cqe;
 733         unsigned head, reaped = 0;
 734         int last_idx = -1, stat_nr = 0;
 735
 736         head = *ring->head;
 737         do {
 738                 struct file *f;
 739
 740                 read_barrier();
 741                 if (head == atomic_load_acquire(ring->tail))
 742                         break;
 743                 cqe = &ring->cqes[head & cq_ring_mask];
 744                 if (!do_nop) {
 745                         int fileno = cqe->user_data & 0xffffffff;
 746
 747                         f = &s->files[fileno];
 748                         f->pending_ios--;
 749                         if (cqe->res != bs) {
 750                                 printf("io: unexpected ret=%d\n", cqe->res);
 751                                 if (polled && cqe->res == -EOPNOTSUPP)
 752                                         printf("Your filesystem/driver/kernel doesn't support polled IO\n");
 753                                 return -1;
 754                         }
 755                 }
 756                 if (stats) {
 757                         int clock_index = cqe->user_data >> 32;
 758
 759                         if (last_idx != clock_index) {
 760                                 if (last_idx != -1) {
 761                                         add_stat(s, last_idx, stat_nr);
 762                                         stat_nr = 0;
 763                                 }
 764                                 last_idx = clock_index;
 765                         }
 766                         stat_nr++;
 767                 }
 768                 reaped++;
 769                 head++;
 770         } while (1);
 771
 772         if (stat_nr)
 773                 add_stat(s, last_idx, stat_nr);
 774
 775         if (reaped) {
 776                 s->inflight -= reaped;
 777                 atomic_store_release(ring->head, head);
 778         }
 779         return reaped;
 780 }
 781
 782 static int reap_events_uring_pt(struct submitter *s)
 783 {
 784         struct io_cq_ring *ring = &s->cq_ring;
 785         struct io_uring_cqe *cqe;
 786         unsigned head, reaped = 0;
 787         int last_idx = -1, stat_nr = 0;
 788         unsigned index;
 789         int fileno;
 790
 791         head = *ring->head;
 792         do {
 793                 struct file *f;
 794
 795                 read_barrier();
 796                 if (head == atomic_load_acquire(ring->tail))
 797                         break;
 798                 index = head & cq_ring_mask;
 799                 cqe = &ring->cqes[index << 1];
 800                 fileno = cqe->user_data & 0xffffffff;
 801                 f = &s->files[fileno];
 802                 f->pending_ios--;
 803
 804                 if (cqe->res != 0) {
 805                         printf("io: unexpected ret=%d\n", cqe->res);
 806                         if (polled && cqe->res == -EINVAL)
 807                                 printf("passthrough doesn't support polled IO\n");
 808                         return -1;
 809                 }
 810                 if (stats) {
 811                         int clock_index = cqe->user_data >> 32;
 812
 813                         if (last_idx != clock_index) {
 814                                 if (last_idx != -1) {
 815                                         add_stat(s, last_idx, stat_nr);
 816                                         stat_nr = 0;
 817                                 }
 818                                 last_idx = clock_index;
 819                         }
 820                         stat_nr++;
 821                 }
 822                 reaped++;
 823                 head++;
 824         } while (1);
 825
 826         if (stat_nr)
 827                 add_stat(s, last_idx, stat_nr);
 828
 829         if (reaped) {
 830                 s->inflight -= reaped;
 831                 atomic_store_release(ring->head, head);
 832         }
 833         return reaped;
 834 }
 835
 836 static void set_affinity(struct submitter *s)
 837 {
 838 #ifdef CONFIG_LIBNUMA
 839         struct bitmask *mask;
 840
 841         if (s->numa_node == -1)
 842                 return;
 843
 844         numa_set_preferred(s->numa_node);
 845
 846         mask = numa_allocate_cpumask();
 847         numa_node_to_cpus(s->numa_node, mask);
 848         numa_sched_setaffinity(s->tid, mask);
 849 #endif
 850 }
 851
 852 static int detect_node(struct submitter *s, const char *name)
 853 {
 854 #ifdef CONFIG_LIBNUMA
 855         const char *base = basename(name);
 856         char str[128];
 857         int ret, fd, node;
 858
 859         sprintf(str, "/sys/block/%s/device/numa_node", base);
 860         fd = open(str, O_RDONLY);
 861         if (fd < 0)
 862                 return -1;
 863
 864         ret = read(fd, str, sizeof(str));
 865         if (ret < 0) {
 866                 close(fd);
 867                 return -1;
 868         }
 869         node = atoi(str);
 870         s->numa_node = node;
 871         close(fd);
 872 #else
 873         s->numa_node = -1;
 874 #endif
 875         return 0;
 876 }
 877
 878 static int setup_aio(struct submitter *s)
 879 {
 880 #ifdef CONFIG_LIBAIO
 881         if (polled) {
 882                 fprintf(stderr, "aio does not support polled IO\n");
 883                 polled = 0;
 884         }
 885         if (sq_thread_poll) {
 886                 fprintf(stderr, "aio does not support SQPOLL IO\n");
 887                 sq_thread_poll = 0;
 888         }
 889         if (do_nop) {
 890                 fprintf(stderr, "aio does not support polled IO\n");
 891                 do_nop = 0;
 892         }
 893         if (fixedbufs || register_files) {
 894                 fprintf(stderr, "aio does not support registered files or buffers\n");
 895                 fixedbufs = register_files = 0;
 896         }
 897
 898         return io_queue_init(roundup_pow2(depth), &s->aio_ctx);
 899 #else
 900         fprintf(stderr, "Legacy AIO not available on this system/build\n");
 901         errno = EINVAL;
 902         return -1;
 903 #endif
 904 }
 905
 906 static int setup_ring(struct submitter *s)
 907 {
 908         struct io_sq_ring *sring = &s->sq_ring;
 909         struct io_cq_ring *cring = &s->cq_ring;
 910         struct io_uring_params p;
 911         int ret, fd;
 912         void *ptr;
 913         size_t len;
 914
 915         memset(&p, 0, sizeof(p));
 916
 917         if (polled && !do_nop)
 918                 p.flags |= IORING_SETUP_IOPOLL;
 919         if (sq_thread_poll) {
 920                 p.flags |= IORING_SETUP_SQPOLL;
 921                 if (sq_thread_cpu != -1) {
 922                         p.flags |= IORING_SETUP_SQ_AFF;
 923                         p.sq_thread_cpu = sq_thread_cpu;
 924                 }
 925         }
 926         if (pt) {
 927                 p.flags |= IORING_SETUP_SQE128;
 928                 p.flags |= IORING_SETUP_CQE32;
 929         }
 930
 931         fd = io_uring_setup(depth, &p);
 932         if (fd < 0) {
 933                 perror("io_uring_setup");
 934                 return 1;
 935         }
 936         s->ring_fd = s->enter_ring_fd = fd;
 937
 938         io_uring_probe(fd);
 939
 940         if (fixedbufs) {
 941                 struct rlimit rlim;
 942
 943                 rlim.rlim_cur = RLIM_INFINITY;
 944                 rlim.rlim_max = RLIM_INFINITY;
 945                 /* ignore potential error, not needed on newer kernels */
 946                 setrlimit(RLIMIT_MEMLOCK, &rlim);
 947
 948                 ret = io_uring_register_buffers(s);
 949                 if (ret < 0) {
 950                         perror("io_uring_register_buffers");
 951                         return 1;
 952                 }
 953
 954                 if (dma_map) {
 955                         ret = io_uring_map_buffers(s);
 956                         if (ret < 0) {
 957                                 perror("io_uring_map_buffers");
 958                                 return 1;
 959                         }
 960                 }
 961         }
 962
 963         if (register_files) {
 964                 ret = io_uring_register_files(s);
 965                 if (ret < 0) {
 966                         perror("io_uring_register_files");
 967                         return 1;
 968                 }
 969         }
 970
 971         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 972                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 973                         IORING_OFF_SQ_RING);
 974         sring->head = ptr + p.sq_off.head;
 975         sring->tail = ptr + p.sq_off.tail;
 976         sring->ring_mask = ptr + p.sq_off.ring_mask;
 977         sring->ring_entries = ptr + p.sq_off.ring_entries;
 978         sring->flags = ptr + p.sq_off.flags;
 979         sring->array = ptr + p.sq_off.array;
 980         sq_ring_mask = *sring->ring_mask;
 981
 982         if (p.flags & IORING_SETUP_SQE128)
 983                 len = 2 * p.sq_entries * sizeof(struct io_uring_sqe);
 984         else
 985                 len = p.sq_entries * sizeof(struct io_uring_sqe);
 986         s->sqes = mmap(0, len,
 987                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 988                         IORING_OFF_SQES);
 989
 990         if (p.flags & IORING_SETUP_CQE32) {
 991                 len = p.cq_off.cqes +
 992                         2 * p.cq_entries * sizeof(struct io_uring_cqe);
 993         } else {
 994                 len = p.cq_off.cqes +
 995                         p.cq_entries * sizeof(struct io_uring_cqe);
 996         }
 997         ptr = mmap(0, len,
 998                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 999                         IORING_OFF_CQ_RING);
1000         cring->head = ptr + p.cq_off.head;
1001         cring->tail = ptr + p.cq_off.tail;
1002         cring->ring_mask = ptr + p.cq_off.ring_mask;
1003         cring->ring_entries = ptr + p.cq_off.ring_entries;
1004         cring->cqes = ptr + p.cq_off.cqes;
1005         cq_ring_mask = *cring->ring_mask;
1006         return 0;
1007 }
1008
1009 static void *allocate_mem(struct submitter *s, int size)
1010 {
1011         void *buf;
1012
1013 #ifdef CONFIG_LIBNUMA
1014         if (s->numa_node != -1)
1015                 return numa_alloc_onnode(size, s->numa_node);
1016 #endif
1017
1018         if (posix_memalign(&buf, t_io_uring_page_size, bs)) {
1019                 printf("failed alloc\n");
1020                 return NULL;
1021         }
1022
1023         return buf;
1024 }
1025
1026 static int submitter_init(struct submitter *s)
1027 {
1028         int i, nr_batch, err;
1029         static int init_printed;
1030         char buf[80];
1031         s->tid = gettid();
1032         printf("submitter=%d, tid=%d, file=%s, node=%d\n", s->index, s->tid,
1033                                                         s->filename, s->numa_node);
1034
1035         set_affinity(s);
1036
1037         __init_rand64(&s->rand_state, s->tid);
1038         srand48(s->tid);
1039
1040         for (i = 0; i < MAX_FDS; i++)
1041                 s->files[i].fileno = i;
1042
1043         for (i = 0; i < roundup_pow2(depth); i++) {
1044                 void *buf;
1045
1046                 buf = allocate_mem(s, bs);
1047                 if (!buf)
1048                         return 1;
1049                 s->iovecs[i].iov_base = buf;
1050                 s->iovecs[i].iov_len = bs;
1051         }
1052
1053         if (use_sync) {
1054                 sprintf(buf, "Engine=preadv2\n");
1055                 err = 0;
1056         } else if (!aio) {
1057                 err = setup_ring(s);
1058                 sprintf(buf, "Engine=io_uring, sq_ring=%d, cq_ring=%d\n", *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
1059         } else {
1060                 sprintf(buf, "Engine=aio\n");
1061                 err = setup_aio(s);
1062         }
1063         if (err) {
1064                 printf("queue setup failed: %s, %d\n", strerror(errno), err);
1065                 return 1;
1066         }
1067
1068         if (!init_printed) {
1069                 printf("polled=%d, fixedbufs=%d/%d, register_files=%d, buffered=%d, QD=%d\n", polled, fixedbufs, dma_map, register_files, buffered, depth);
1070                 printf("%s", buf);
1071                 init_printed = 1;
1072         }
1073
1074         if (stats) {
1075                 nr_batch = roundup_pow2(depth / batch_submit);
1076                 if (nr_batch < 2)
1077                         nr_batch = 2;
1078                 s->clock_batch = calloc(nr_batch, sizeof(unsigned long));
1079                 s->clock_index = 1;
1080
1081                 s->plat = calloc(PLAT_NR, sizeof(unsigned long));
1082         } else {
1083                 s->clock_batch = NULL;
1084                 s->plat = NULL;
1085                 nr_batch = 0;
1086         }
1087         /* perform the expensive command initialization part for passthrough here
1088          * rather than in the fast path
1089          */
1090         if (pt) {
1091                 for (i = 0; i < roundup_pow2(depth); i++) {
1092                         struct io_uring_sqe *sqe = &s->sqes[i << 1];
1093
1094                         memset(&sqe->cmd, 0, sizeof(struct nvme_uring_cmd));
1095                 }
1096         }
1097         return nr_batch;
1098 }
1099
1100 #ifdef CONFIG_LIBAIO
1101 static int prep_more_ios_aio(struct submitter *s, int max_ios, struct iocb *iocbs)
1102 {
1103         uint64_t data;
1104         struct file *f;
1105         unsigned index;
1106
1107         index = 0;
1108         while (index < max_ios) {
1109                 struct iocb *iocb = &iocbs[index];
1110
1111                 if (s->nr_files == 1) {
1112                         f = &s->files[0];
1113                 } else {
1114                         f = &s->files[s->cur_file];
1115                         if (f->pending_ios >= file_depth(s)) {
1116                                 s->cur_file++;
1117                                 if (s->cur_file == s->nr_files)
1118                                         s->cur_file = 0;
1119                                 f = &s->files[s->cur_file];
1120                         }
1121                 }
1122                 f->pending_ios++;
1123
1124                 io_prep_pread(iocb, f->real_fd, s->iovecs[index].iov_base,
1125                                 s->iovecs[index].iov_len, get_offset(s, f));
1126
1127                 data = f->fileno;
1128                 if (stats && stats_running)
1129                         data |= (((uint64_t) s->clock_index) << 32);
1130                 iocb->data = (void *) (uintptr_t) data;
1131                 index++;
1132         }
1133         return index;
1134 }
1135
1136 static int reap_events_aio(struct submitter *s, struct io_event *events, int evs)
1137 {
1138         int last_idx = -1, stat_nr = 0;
1139         int reaped = 0;
1140
1141         while (evs) {
1142                 uint64_t data = (uintptr_t) events[reaped].data;
1143                 struct file *f = &s->files[data & 0xffffffff];
1144
1145                 f->pending_ios--;
1146                 if (events[reaped].res != bs) {
1147                         printf("io: unexpected ret=%ld\n", events[reaped].res);
1148                         return -1;
1149                 }
1150                 if (stats) {
1151                         int clock_index = data >> 32;
1152
1153                         if (last_idx != clock_index) {
1154                                 if (last_idx != -1) {
1155                                         add_stat(s, last_idx, stat_nr);
1156                                         stat_nr = 0;
1157                                 }
1158                                 last_idx = clock_index;
1159                         }
1160                         stat_nr++;
1161                 }
1162                 reaped++;
1163                 evs--;
1164         }
1165
1166         if (stat_nr)
1167                 add_stat(s, last_idx, stat_nr);
1168
1169         s->inflight -= reaped;
1170         s->done += reaped;
1171         return reaped;
1172 }
1173
1174 static void *submitter_aio_fn(void *data)
1175 {
1176         struct submitter *s = data;
1177         int i, ret, prepped;
1178         struct iocb **iocbsptr;
1179         struct iocb *iocbs;
1180         struct io_event *events;
1181 #ifdef ARCH_HAVE_CPU_CLOCK
1182         int nr_batch = submitter_init(s);
1183 #else
1184         submitter_init(s);
1185 #endif
1186
1187         iocbsptr = calloc(depth, sizeof(struct iocb *));
1188         iocbs = calloc(depth, sizeof(struct iocb));
1189         events = calloc(depth, sizeof(struct io_event));
1190
1191         for (i = 0; i < depth; i++)
1192                 iocbsptr[i] = &iocbs[i];
1193
1194         prepped = 0;
1195         do {
1196                 int to_wait, to_submit, to_prep;
1197
1198                 if (!prepped && s->inflight < depth) {
1199                         to_prep = min(depth - s->inflight, batch_submit);
1200                         prepped = prep_more_ios_aio(s, to_prep, iocbs);
1201 #ifdef ARCH_HAVE_CPU_CLOCK
1202                         if (prepped && stats) {
1203                                 s->clock_batch[s->clock_index] = get_cpu_clock();
1204                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
1205                         }
1206 #endif
1207                 }
1208                 s->inflight += prepped;
1209                 to_submit = prepped;
1210
1211                 if (to_submit && (s->inflight + to_submit <= depth))
1212                         to_wait = 0;
1213                 else
1214                         to_wait = min(s->inflight + to_submit, batch_complete);
1215
1216                 ret = io_submit(s->aio_ctx, to_submit, iocbsptr);
1217                 s->calls++;
1218                 if (ret < 0) {
1219                         perror("io_submit");
1220                         break;
1221                 } else if (ret != to_submit) {
1222                         printf("submitted %d, wanted %d\n", ret, to_submit);
1223                         break;
1224                 }
1225                 prepped = 0;
1226
1227                 while (to_wait) {
1228                         int r;
1229
1230                         s->calls++;
1231                         r = io_getevents(s->aio_ctx, to_wait, to_wait, events, NULL);
1232                         if (r < 0) {
1233                                 perror("io_getevents");
1234                                 break;
1235                         } else if (r != to_wait) {
1236                                 printf("r=%d, wait=%d\n", r, to_wait);
1237                                 break;
1238                         }
1239                         r = reap_events_aio(s, events, r);
1240                         s->reaps += r;
1241                         to_wait -= r;
1242                 }
1243         } while (!s->finish);
1244
1245         free(iocbsptr);
1246         free(iocbs);
1247         free(events);
1248         finish = 1;
1249         return NULL;
1250 }
1251 #endif
1252
1253 static void io_uring_unregister_ring(struct submitter *s)
1254 {
1255         struct io_uring_rsrc_update up = {
1256                 .offset = s->enter_ring_fd,
1257         };
1258
1259         syscall(__NR_io_uring_register, s->ring_fd, IORING_UNREGISTER_RING_FDS,
1260                 &up, 1);
1261 }
1262
1263 static int io_uring_register_ring(struct submitter *s)
1264 {
1265         struct io_uring_rsrc_update up = {
1266                 .data   = s->ring_fd,
1267                 .offset = -1U,
1268         };
1269         int ret;
1270
1271         ret = syscall(__NR_io_uring_register, s->ring_fd,
1272                         IORING_REGISTER_RING_FDS, &up, 1);
1273         if (ret == 1) {
1274                 s->enter_ring_fd = up.offset;
1275                 return 0;
1276         }
1277         register_ring = 0;
1278         return -1;
1279 }
1280
1281 static void *submitter_uring_fn(void *data)
1282 {
1283         struct submitter *s = data;
1284         struct io_sq_ring *ring = &s->sq_ring;
1285         int ret, prepped;
1286 #ifdef ARCH_HAVE_CPU_CLOCK
1287         int nr_batch = submitter_init(s);
1288 #else
1289         submitter_init(s);
1290 #endif
1291
1292         if (register_ring)
1293                 io_uring_register_ring(s);
1294
1295         prepped = 0;
1296         do {
1297                 int to_wait, to_submit, this_reap, to_prep;
1298                 unsigned ring_flags = 0;
1299
1300                 if (!prepped && s->inflight < depth) {
1301                         to_prep = min(depth - s->inflight, batch_submit);
1302                         prepped = prep_more_ios_uring(s, to_prep);
1303 #ifdef ARCH_HAVE_CPU_CLOCK
1304                         if (prepped && stats) {
1305                                 s->clock_batch[s->clock_index] = get_cpu_clock();
1306                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
1307                         }
1308 #endif
1309                 }
1310                 s->inflight += prepped;
1311 submit_more:
1312                 to_submit = prepped;
1313 submit:
1314                 if (to_submit && (s->inflight + to_submit <= depth))
1315                         to_wait = 0;
1316                 else
1317                         to_wait = min(s->inflight + to_submit, batch_complete);
1318
1319                 /*
1320                  * Only need to call io_uring_enter if we're not using SQ thread
1321                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
1322                  */
1323                 if (sq_thread_poll)
1324                         ring_flags = atomic_load_acquire(ring->flags);
1325                 if (!sq_thread_poll || ring_flags & IORING_SQ_NEED_WAKEUP) {
1326                         unsigned flags = 0;
1327
1328                         if (to_wait)
1329                                 flags = IORING_ENTER_GETEVENTS;
1330                         if (ring_flags & IORING_SQ_NEED_WAKEUP)
1331                                 flags |= IORING_ENTER_SQ_WAKEUP;
1332                         ret = io_uring_enter(s, to_submit, to_wait, flags);
1333                         s->calls++;
1334                 } else {
1335                         /* for SQPOLL, we submitted it all effectively */
1336                         ret = to_submit;
1337                 }
1338
1339                 /*
1340                  * For non SQ thread poll, we already got the events we needed
1341                  * through the io_uring_enter() above. For SQ thread poll, we
1342                  * need to loop here until we find enough events.
1343                  */
1344                 this_reap = 0;
1345                 do {
1346                         int r;
1347
1348                         if (pt)
1349                                 r = reap_events_uring_pt(s);
1350                         else
1351                                 r = reap_events_uring(s);
1352                         if (r == -1) {
1353                                 s->finish = 1;
1354                                 break;
1355                         } else if (r > 0)
1356                                 this_reap += r;
1357                 } while (sq_thread_poll && this_reap < to_wait);
1358                 s->reaps += this_reap;
1359
1360                 if (ret >= 0) {
1361                         if (!ret) {
1362                                 to_submit = 0;
1363                                 if (s->inflight)
1364                                         goto submit;
1365                                 continue;
1366                         } else if (ret < to_submit) {
1367                                 int diff = to_submit - ret;
1368
1369                                 s->done += ret;
1370                                 prepped -= diff;
1371                                 goto submit_more;
1372                         }
1373                         s->done += ret;
1374                         prepped = 0;
1375                         continue;
1376                 } else if (ret < 0) {
1377                         if (errno == EAGAIN) {
1378                                 if (s->finish)
1379                                         break;
1380                                 if (this_reap)
1381                                         goto submit;
1382                                 to_submit = 0;
1383                                 goto submit;
1384                         }
1385                         printf("io_submit: %s\n", strerror(errno));
1386                         break;
1387                 }
1388         } while (!s->finish);
1389
1390         if (register_ring)
1391                 io_uring_unregister_ring(s);
1392
1393         finish = 1;
1394         return NULL;
1395 }
1396
1397 #ifdef CONFIG_PWRITEV2
1398 static void *submitter_sync_fn(void *data)
1399 {
1400         struct submitter *s = data;
1401         int ret;
1402
1403         submitter_init(s);
1404
1405         do {
1406                 uint64_t offset;
1407                 struct file *f;
1408
1409                 if (s->nr_files == 1) {
1410                         f = &s->files[0];
1411                 } else {
1412                         f = &s->files[s->cur_file];
1413                         if (f->pending_ios >= file_depth(s)) {
1414                                 s->cur_file++;
1415                                 if (s->cur_file == s->nr_files)
1416                                         s->cur_file = 0;
1417                                 f = &s->files[s->cur_file];
1418                         }
1419                 }
1420                 f->pending_ios++;
1421
1422 #ifdef ARCH_HAVE_CPU_CLOCK
1423                 if (stats)
1424                         s->clock_batch[s->clock_index] = get_cpu_clock();
1425 #endif
1426
1427                 s->inflight++;
1428                 s->calls++;
1429
1430                 offset = get_offset(s, f);
1431                 if (polled)
1432                         ret = preadv2(f->real_fd, &s->iovecs[0], 1, offset, RWF_HIPRI);
1433                 else
1434                         ret = preadv2(f->real_fd, &s->iovecs[0], 1, offset, 0);
1435
1436                 if (ret < 0) {
1437                         perror("preadv2");
1438                         break;
1439                 } else if (ret != bs) {
1440                         break;
1441                 }
1442
1443                 s->done++;
1444                 s->inflight--;
1445                 f->pending_ios--;
1446                 if (stats)
1447                         add_stat(s, s->clock_index, 1);
1448         } while (!s->finish);
1449
1450         finish = 1;
1451         return NULL;
1452 }
1453 #else
1454 static void *submitter_sync_fn(void *data)
1455 {
1456         finish = 1;
1457         return NULL;
1458 }
1459 #endif
1460
1461 static struct submitter *get_submitter(int offset)
1462 {
1463         void *ret;
1464
1465         ret = submitter;
1466         if (offset)
1467                 ret += offset * (sizeof(*submitter) + depth * sizeof(struct iovec));
1468         return ret;
1469 }
1470
1471 static void do_finish(const char *reason)
1472 {
1473         int j;
1474
1475         printf("Exiting on %s\n", reason);
1476         for (j = 0; j < nthreads; j++) {
1477                 struct submitter *s = get_submitter(j);
1478                 s->finish = 1;
1479         }
1480         if (max_iops > 1000000) {
1481                 double miops = (double) max_iops / 1000000.0;
1482                 printf("Maximum IOPS=%.2fM\n", miops);
1483         } else if (max_iops > 100000) {
1484                 double kiops = (double) max_iops / 1000.0;
1485                 printf("Maximum IOPS=%.2fK\n", kiops);
1486         } else {
1487                 printf("Maximum IOPS=%lu\n", max_iops);
1488         }
1489         finish = 1;
1490 }
1491
1492 static void sig_int(int sig)
1493 {
1494         do_finish("signal");
1495 }
1496
1497 static void arm_sig_int(void)
1498 {
1499         struct sigaction act;
1500
1501         memset(&act, 0, sizeof(act));
1502         act.sa_handler = sig_int;
1503         act.sa_flags = SA_RESTART;
1504         sigaction(SIGINT, &act, NULL);
1505
1506         /* Windows uses SIGBREAK as a quit signal from other applications */
1507 #ifdef WIN32
1508         sigaction(SIGBREAK, &act, NULL);
1509 #endif
1510 }
1511
1512 static void usage(char *argv, int status)
1513 {
1514         char runtime_str[16];
1515         snprintf(runtime_str, sizeof(runtime_str), "%d", runtime);
1516         printf("%s [options] -- [filenames]\n"
1517                 " -d <int>  : IO Depth, default %d\n"
1518                 " -s <int>  : Batch submit, default %d\n"
1519                 " -c <int>  : Batch complete, default %d\n"
1520                 " -b <int>  : Block size, default %d\n"
1521                 " -p <bool> : Polled IO, default %d\n"
1522                 " -B <bool> : Fixed buffers, default %d\n"
1523                 " -D <bool> : DMA map fixed buffers, default %d\n"
1524                 " -F <bool> : Register files, default %d\n"
1525                 " -n <int>  : Number of threads, default %d\n"
1526                 " -O <bool> : Use O_DIRECT, default %d\n"
1527                 " -N <bool> : Perform just no-op requests, default %d\n"
1528                 " -t <bool> : Track IO latencies, default %d\n"
1529                 " -T <int>  : TSC rate in HZ\n"
1530                 " -r <int>  : Runtime in seconds, default %s\n"
1531                 " -R <bool> : Use random IO, default %d\n"
1532                 " -a <bool> : Use legacy aio, default %d\n"
1533                 " -S <bool> : Use sync IO (preadv2), default %d\n"
1534                 " -X <bool> : Use registered ring %d\n"
1535                 " -P <bool> : Automatically place on device home node %d\n"
1536                 " -u <bool> : Use nvme-passthrough I/O, default %d\n",
1537                 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled,
1538                 fixedbufs, dma_map, register_files, nthreads, !buffered, do_nop,
1539                 stats, runtime == 0 ? "unlimited" : runtime_str, random_io, aio,
1540                 use_sync, register_ring, numa_placement, pt);
1541         exit(status);
1542 }
1543
1544 static void read_tsc_rate(void)
1545 {
1546         char buffer[32];
1547         int fd, ret;
1548
1549         if (tsc_rate)
1550                 return;
1551
1552         fd = open(TSC_RATE_FILE, O_RDONLY);
1553         if (fd < 0)
1554                 return;
1555
1556         ret = read(fd, buffer, sizeof(buffer));
1557         if (ret < 0) {
1558                 close(fd);
1559                 return;
1560         }
1561
1562         tsc_rate = strtoul(buffer, NULL, 10);
1563         printf("Using TSC rate %luHz\n", tsc_rate);
1564         close(fd);
1565 }
1566
1567 static void write_tsc_rate(void)
1568 {
1569         char buffer[32];
1570         struct stat sb;
1571         int fd, ret;
1572
1573         if (!stat(TSC_RATE_FILE, &sb))
1574                 return;
1575
1576         fd = open(TSC_RATE_FILE, O_WRONLY | O_CREAT, 0644);
1577         if (fd < 0)
1578                 return;
1579
1580         memset(buffer, 0, sizeof(buffer));
1581         sprintf(buffer, "%lu", tsc_rate);
1582         ret = write(fd, buffer, strlen(buffer));
1583         if (ret < 0)
1584                 perror("write");
1585         close(fd);
1586 }
1587
1588 int main(int argc, char *argv[])
1589 {
1590         struct submitter *s;
1591         unsigned long done, calls, reap;
1592         int i, j, flags, fd, opt, threads_per_f, threads_rem = 0, nfiles;
1593         struct file f;
1594         void *ret;
1595
1596         if (!do_nop && argc < 2)
1597                 usage(argv[0], 1);
1598
1599         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) {
1600                 switch (opt) {
1601                 case 'a':
1602                         aio = !!atoi(optarg);
1603                         break;
1604                 case 'd':
1605                         depth = atoi(optarg);
1606                         break;
1607                 case 's':
1608                         batch_submit = atoi(optarg);
1609                         if (!batch_submit)
1610                                 batch_submit = 1;
1611                         break;
1612                 case 'c':
1613                         batch_complete = atoi(optarg);
1614                         if (!batch_complete)
1615                                 batch_complete = 1;
1616                         break;
1617                 case 'b':
1618                         bs = atoi(optarg);
1619                         break;
1620                 case 'p':
1621                         polled = !!atoi(optarg);
1622                         break;
1623                 case 'B':
1624                         fixedbufs = !!atoi(optarg);
1625                         break;
1626                 case 'F':
1627                         register_files = !!atoi(optarg);
1628                         break;
1629                 case 'n':
1630                         nthreads = atoi(optarg);
1631                         if (!nthreads) {
1632                                 printf("Threads must be non-zero\n");
1633                                 usage(argv[0], 1);
1634                         }
1635                         break;
1636                 case 'N':
1637                         do_nop = !!atoi(optarg);
1638                         break;
1639                 case 'O':
1640                         buffered = !atoi(optarg);
1641                         break;
1642                 case 't':
1643 #ifndef ARCH_HAVE_CPU_CLOCK
1644                         fprintf(stderr, "Stats not supported on this CPU\n");
1645                         return 1;
1646 #endif
1647                         stats = !!atoi(optarg);
1648                         break;
1649                 case 'T':
1650 #ifndef ARCH_HAVE_CPU_CLOCK
1651                         fprintf(stderr, "Stats not supported on this CPU\n");
1652                         return 1;
1653 #endif
1654                         tsc_rate = strtoul(optarg, NULL, 10);
1655                         write_tsc_rate();
1656                         break;
1657                 case 'r':
1658                         runtime = atoi(optarg);
1659                         break;
1660                 case 'D':
1661                         dma_map = !!atoi(optarg);
1662                         break;
1663                 case 'R':
1664                         random_io = !!atoi(optarg);
1665                         break;
1666                 case 'X':
1667                         register_ring = !!atoi(optarg);
1668                         break;
1669                 case 'S':
1670 #ifdef CONFIG_PWRITEV2
1671                         use_sync = !!atoi(optarg);
1672 #else
1673                         fprintf(stderr, "preadv2 not supported\n");
1674                         exit(1);
1675 #endif
1676                         break;
1677                 case 'P':
1678                         numa_placement = !!atoi(optarg);
1679                         break;
1680                 case 'u':
1681                         pt = !!atoi(optarg);
1682                         break;
1683                 case 'h':
1684                 case '?':
1685                 default:
1686                         usage(argv[0], 0);
1687                         break;
1688                 }
1689         }
1690
1691         if (stats)
1692                 read_tsc_rate();
1693
1694         if (batch_complete > depth)
1695                 batch_complete = depth;
1696         if (batch_submit > depth)
1697                 batch_submit = depth;
1698         if (!fixedbufs && dma_map)
1699                 dma_map = 0;
1700
1701         submitter = calloc(nthreads, sizeof(*submitter) +
1702                                 roundup_pow2(depth) * sizeof(struct iovec));
1703         for (j = 0; j < nthreads; j++) {
1704                 s = get_submitter(j);
1705                 s->numa_node = -1;
1706                 s->index = j;
1707                 s->done = s->calls = s->reaps = 0;
1708         }
1709
1710         flags = O_RDONLY | O_NOATIME;
1711         if (!buffered)
1712                 flags |= O_DIRECT;
1713
1714         j = 0;
1715         i = optind;
1716         nfiles = argc - i;
1717         if (!do_nop) {
1718                 if (!nfiles) {
1719                         printf("No files specified\n");
1720                         usage(argv[0], 1);
1721                 }
1722                 threads_per_f = nthreads / nfiles;
1723                 /* make sure each thread gets assigned files */
1724                 if (threads_per_f == 0) {
1725                         threads_per_f = 1;
1726                 } else {
1727                         threads_rem = nthreads - threads_per_f * nfiles;
1728                 }
1729         }
1730         while (!do_nop && i < argc) {
1731                 int k, limit;
1732
1733                 memset(&f, 0, sizeof(f));
1734
1735                 fd = open(argv[i], flags);
1736                 if (fd < 0) {
1737                         perror("open");
1738                         return 1;
1739                 }
1740                 f.real_fd = fd;
1741                 if (get_file_size(&f)) {
1742                         printf("failed getting size of device/file\n");
1743                         return 1;
1744                 }
1745                 if (f.max_blocks <= 1) {
1746                         printf("Zero file/device size?\n");
1747                         return 1;
1748                 }
1749                 f.max_blocks--;
1750
1751                 limit = threads_per_f;
1752                 limit += threads_rem > 0 ? 1 : 0;
1753                 for (k = 0; k < limit; k++) {
1754                         s = get_submitter((j + k) % nthreads);
1755
1756                         if (s->nr_files == MAX_FDS) {
1757                                 printf("Max number of files (%d) reached\n", MAX_FDS);
1758                                 break;
1759                         }
1760
1761                         memcpy(&s->files[s->nr_files], &f, sizeof(f));
1762
1763                         if (numa_placement)
1764                                 detect_node(s, argv[i]);
1765
1766                         s->filename = argv[i];
1767                         s->nr_files++;
1768                 }
1769                 threads_rem--;
1770                 i++;
1771                 j += limit;
1772         }
1773
1774         arm_sig_int();
1775
1776         t_io_uring_page_size = sysconf(_SC_PAGESIZE);
1777         if (t_io_uring_page_size < 0)
1778                 t_io_uring_page_size = 4096;
1779
1780         for (j = 0; j < nthreads; j++) {
1781                 s = get_submitter(j);
1782                 if (use_sync)
1783                         pthread_create(&s->thread, NULL, submitter_sync_fn, s);
1784                 else if (!aio)
1785                         pthread_create(&s->thread, NULL, submitter_uring_fn, s);
1786 #ifdef CONFIG_LIBAIO
1787                 else
1788                         pthread_create(&s->thread, NULL, submitter_aio_fn, s);
1789 #endif
1790         }
1791
1792         reap = calls = done = 0;
1793         do {
1794                 unsigned long this_done = 0;
1795                 unsigned long this_reap = 0;
1796                 unsigned long this_call = 0;
1797                 unsigned long rpc = 0, ipc = 0;
1798                 unsigned long iops, bw;
1799
1800                 sleep(1);
1801                 if (runtime && !--runtime)
1802                         do_finish("timeout");
1803
1804                 /* don't print partial run, if interrupted by signal */
1805                 if (finish)
1806                         break;
1807
1808                 /* one second in to the run, enable stats */
1809                 if (stats)
1810                         stats_running = 1;
1811
1812                 for (j = 0; j < nthreads; j++) {
1813                         s = get_submitter(j);
1814                         this_done += s->done;
1815                         this_call += s->calls;
1816                         this_reap += s->reaps;
1817                 }
1818                 if (this_call - calls) {
1819                         rpc = (this_done - done) / (this_call - calls);
1820                         ipc = (this_reap - reap) / (this_call - calls);
1821                 } else
1822                         rpc = ipc = -1;
1823                 iops = this_done - done;
1824                 if (bs > 1048576)
1825                         bw = iops * (bs / 1048576);
1826                 else
1827                         bw = iops / (1048576 / bs);
1828                 if (iops > 1000000) {
1829                         double miops = (double) iops / 1000000.0;
1830                         printf("IOPS=%.2fM, ", miops);
1831                 } else if (iops > 100000) {
1832                         double kiops = (double) iops / 1000.0;
1833                         printf("IOPS=%.2fK, ", kiops);
1834                 } else {
1835                         printf("IOPS=%lu, ", iops);
1836                 }
1837                 max_iops = max(max_iops, iops);
1838                 if (!do_nop) {
1839                         if (bw > 2000) {
1840                                 double bw_g = (double) bw / 1000.0;
1841
1842                                 printf("BW=%.2fGiB/s, ", bw_g);
1843                         } else {
1844                                 printf("BW=%luMiB/s, ", bw);
1845                         }
1846                 }
1847                 printf("IOS/call=%ld/%ld\n", rpc, ipc);
1848                 done = this_done;
1849                 calls = this_call;
1850                 reap = this_reap;
1851         } while (!finish);
1852
1853         for (j = 0; j < nthreads; j++) {
1854                 s = get_submitter(j);
1855                 pthread_join(s->thread, &ret);
1856                 close(s->ring_fd);
1857
1858                 if (stats) {
1859                         unsigned long nr;
1860
1861                         printf("%d: Latency percentiles:\n", s->tid);
1862                         for (i = 0, nr = 0; i < PLAT_NR; i++)
1863                                 nr += s->plat[i];
1864                         show_clat_percentiles(s->plat, nr, 4);
1865                         free(s->clock_batch);
1866                         free(s->plat);
1867                 }
1868         }
1869
1870         free(submitter);
1871         return 0;
1872 }