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     26
 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                 unsigned long long block;
 534
 535                 r = __rand64(&s->rand_state);
 536                 block = r % f->max_blocks;
 537                 offset = block * (unsigned long long) bs;
 538         } else {
 539                 offset = f->cur_off;
 540                 f->cur_off += bs;
 541                 if (f->cur_off + bs > f->max_size)
 542                         f->cur_off = 0;
 543         }
 544
 545         return offset;
 546 }
 547
 548 static void init_io(struct submitter *s, unsigned index)
 549 {
 550         struct io_uring_sqe *sqe = &s->sqes[index];
 551         struct file *f;
 552
 553         if (do_nop) {
 554                 sqe->opcode = IORING_OP_NOP;
 555                 return;
 556         }
 557
 558         if (s->nr_files == 1) {
 559                 f = &s->files[0];
 560         } else {
 561                 f = &s->files[s->cur_file];
 562                 if (f->pending_ios >= file_depth(s)) {
 563                         s->cur_file++;
 564                         if (s->cur_file == s->nr_files)
 565                                 s->cur_file = 0;
 566                         f = &s->files[s->cur_file];
 567                 }
 568         }
 569         f->pending_ios++;
 570
 571         if (register_files) {
 572                 sqe->flags = IOSQE_FIXED_FILE;
 573                 sqe->fd = f->fixed_fd;
 574         } else {
 575                 sqe->flags = 0;
 576                 sqe->fd = f->real_fd;
 577         }
 578         if (fixedbufs) {
 579                 sqe->opcode = IORING_OP_READ_FIXED;
 580                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 581                 sqe->len = bs;
 582                 sqe->buf_index = index;
 583         } else if (!vectored) {
 584                 sqe->opcode = IORING_OP_READ;
 585                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 586                 sqe->len = bs;
 587                 sqe->buf_index = 0;
 588         } else {
 589                 sqe->opcode = IORING_OP_READV;
 590                 sqe->addr = (unsigned long) &s->iovecs[index];
 591                 sqe->len = 1;
 592                 sqe->buf_index = 0;
 593         }
 594         sqe->ioprio = 0;
 595         sqe->off = get_offset(s, f);
 596         sqe->user_data = (unsigned long) f->fileno;
 597         if (stats && stats_running)
 598                 sqe->user_data |= ((uint64_t)s->clock_index << 32);
 599 }
 600
 601 static void init_io_pt(struct submitter *s, unsigned index)
 602 {
 603         struct io_uring_sqe *sqe = &s->sqes[index << 1];
 604         unsigned long offset;
 605         struct file *f;
 606         struct nvme_uring_cmd *cmd;
 607         unsigned long long slba;
 608         unsigned long long nlb;
 609
 610         if (s->nr_files == 1) {
 611                 f = &s->files[0];
 612         } else {
 613                 f = &s->files[s->cur_file];
 614                 if (f->pending_ios >= file_depth(s)) {
 615                         s->cur_file++;
 616                         if (s->cur_file == s->nr_files)
 617                                 s->cur_file = 0;
 618                         f = &s->files[s->cur_file];
 619                 }
 620         }
 621         f->pending_ios++;
 622
 623         offset = get_offset(s, f);
 624
 625         if (register_files) {
 626                 sqe->fd = f->fixed_fd;
 627                 sqe->flags = IOSQE_FIXED_FILE;
 628         } else {
 629                 sqe->fd = f->real_fd;
 630                 sqe->flags = 0;
 631         }
 632         sqe->opcode = IORING_OP_URING_CMD;
 633         sqe->user_data = (unsigned long) f->fileno;
 634         if (stats)
 635                 sqe->user_data |= ((__u64) s->clock_index << 32ULL);
 636         sqe->cmd_op = NVME_URING_CMD_IO;
 637         slba = offset >> f->lba_shift;
 638         nlb = (bs >> f->lba_shift) - 1;
 639         cmd = (struct nvme_uring_cmd *)&sqe->cmd;
 640         /* cdw10 and cdw11 represent starting slba*/
 641         cmd->cdw10 = slba & 0xffffffff;
 642         cmd->cdw11 = slba >> 32;
 643         /* cdw12 represent number of lba to be read*/
 644         cmd->cdw12 = nlb;
 645         cmd->addr = (unsigned long) s->iovecs[index].iov_base;
 646         cmd->data_len = bs;
 647         if (fixedbufs) {
 648                 sqe->uring_cmd_flags = IORING_URING_CMD_FIXED;
 649                 sqe->buf_index = index;
 650         }
 651         cmd->nsid = f->nsid;
 652         cmd->opcode = 2;
 653 }
 654
 655 static int prep_more_ios_uring(struct submitter *s, int max_ios)
 656 {
 657         struct io_sq_ring *ring = &s->sq_ring;
 658         unsigned head, index, tail, next_tail, prepped = 0;
 659
 660         if (sq_thread_poll)
 661                 head = atomic_load_acquire(ring->head);
 662         else
 663                 head = *ring->head;
 664
 665         next_tail = tail = *ring->tail;
 666         do {
 667                 next_tail++;
 668                 if (next_tail == head)
 669                         break;
 670
 671                 index = tail & sq_ring_mask;
 672                 if (pt)
 673                         init_io_pt(s, index);
 674                 else
 675                         init_io(s, index);
 676                 prepped++;
 677                 tail = next_tail;
 678         } while (prepped < max_ios);
 679
 680         if (prepped)
 681                 atomic_store_release(ring->tail, tail);
 682         return prepped;
 683 }
 684
 685 static int get_file_size(struct file *f)
 686 {
 687         struct stat st;
 688
 689         if (fstat(f->real_fd, &st) < 0)
 690                 return -1;
 691         if (pt) {
 692                 __u64 nlba;
 693                 __u32 lbs;
 694                 int ret;
 695
 696                 if (!S_ISCHR(st.st_mode)) {
 697                         fprintf(stderr, "passthrough works with only nvme-ns "
 698                                         "generic devices (/dev/ngXnY)\n");
 699                         return -1;
 700                 }
 701                 ret = nvme_get_info(f->real_fd, &f->nsid, &lbs, &nlba);
 702                 if (ret)
 703                         return -1;
 704                 if ((bs % lbs) != 0) {
 705                         printf("error: bs:%d should be a multiple logical_block_size:%d\n",
 706                                         bs, lbs);
 707                         return -1;
 708                 }
 709                 f->max_blocks = nlba / bs;
 710                 f->max_size = nlba;
 711                 f->lba_shift = ilog2(lbs);
 712                 return 0;
 713         } else if (S_ISBLK(st.st_mode)) {
 714                 unsigned long long bytes;
 715
 716                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 717                         return -1;
 718
 719                 f->max_blocks = bytes / bs;
 720                 f->max_size = bytes;
 721                 return 0;
 722         } else if (S_ISREG(st.st_mode)) {
 723                 f->max_blocks = st.st_size / bs;
 724                 f->max_size = st.st_size;
 725                 return 0;
 726         }
 727
 728         return -1;
 729 }
 730
 731 static int reap_events_uring(struct submitter *s)
 732 {
 733         struct io_cq_ring *ring = &s->cq_ring;
 734         struct io_uring_cqe *cqe;
 735         unsigned head, reaped = 0;
 736         int last_idx = -1, stat_nr = 0;
 737
 738         head = *ring->head;
 739         do {
 740                 struct file *f;
 741
 742                 if (head == atomic_load_acquire(ring->tail))
 743                         break;
 744                 cqe = &ring->cqes[head & cq_ring_mask];
 745                 if (!do_nop) {
 746                         int fileno = cqe->user_data & 0xffffffff;
 747
 748                         f = &s->files[fileno];
 749                         f->pending_ios--;
 750                         if (cqe->res != bs) {
 751                                 printf("io: unexpected ret=%d\n", cqe->res);
 752                                 if (polled && cqe->res == -EOPNOTSUPP)
 753                                         printf("Your filesystem/driver/kernel doesn't support polled IO\n");
 754                                 return -1;
 755                         }
 756                 }
 757                 if (stats) {
 758                         int clock_index = cqe->user_data >> 32;
 759
 760                         if (last_idx != clock_index) {
 761                                 if (last_idx != -1) {
 762                                         add_stat(s, last_idx, stat_nr);
 763                                         stat_nr = 0;
 764                                 }
 765                                 last_idx = clock_index;
 766                         }
 767                         stat_nr++;
 768                 }
 769                 reaped++;
 770                 head++;
 771         } while (1);
 772
 773         if (stat_nr)
 774                 add_stat(s, last_idx, stat_nr);
 775
 776         if (reaped) {
 777                 s->inflight -= reaped;
 778                 atomic_store_release(ring->head, head);
 779         }
 780         return reaped;
 781 }
 782
 783 static int reap_events_uring_pt(struct submitter *s)
 784 {
 785         struct io_cq_ring *ring = &s->cq_ring;
 786         struct io_uring_cqe *cqe;
 787         unsigned head, reaped = 0;
 788         int last_idx = -1, stat_nr = 0;
 789         unsigned index;
 790         int fileno;
 791
 792         head = *ring->head;
 793         do {
 794                 struct file *f;
 795
 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         if (pt)
 860                 sprintf(str, "/sys/class/nvme-generic/%s/device/numa_node", base);
 861         else
 862                 sprintf(str, "/sys/block/%s/device/numa_node", base);
 863         fd = open(str, O_RDONLY);
 864         if (fd < 0)
 865                 return -1;
 866
 867         ret = read(fd, str, sizeof(str));
 868         if (ret < 0) {
 869                 close(fd);
 870                 return -1;
 871         }
 872         node = atoi(str);
 873         s->numa_node = node;
 874         close(fd);
 875 #else
 876         s->numa_node = -1;
 877 #endif
 878         return 0;
 879 }
 880
 881 static int setup_aio(struct submitter *s)
 882 {
 883 #ifdef CONFIG_LIBAIO
 884         if (polled) {
 885                 fprintf(stderr, "aio does not support polled IO\n");
 886                 polled = 0;
 887         }
 888         if (sq_thread_poll) {
 889                 fprintf(stderr, "aio does not support SQPOLL IO\n");
 890                 sq_thread_poll = 0;
 891         }
 892         if (do_nop) {
 893                 fprintf(stderr, "aio does not support polled IO\n");
 894                 do_nop = 0;
 895         }
 896         if (fixedbufs || register_files) {
 897                 fprintf(stderr, "aio does not support registered files or buffers\n");
 898                 fixedbufs = register_files = 0;
 899         }
 900
 901         return io_queue_init(roundup_pow2(depth), &s->aio_ctx);
 902 #else
 903         fprintf(stderr, "Legacy AIO not available on this system/build\n");
 904         errno = EINVAL;
 905         return -1;
 906 #endif
 907 }
 908
 909 static int setup_ring(struct submitter *s)
 910 {
 911         struct io_sq_ring *sring = &s->sq_ring;
 912         struct io_cq_ring *cring = &s->cq_ring;
 913         struct io_uring_params p;
 914         int ret, fd, i;
 915         void *ptr;
 916         size_t len;
 917
 918         memset(&p, 0, sizeof(p));
 919
 920         if (polled && !do_nop)
 921                 p.flags |= IORING_SETUP_IOPOLL;
 922         if (sq_thread_poll) {
 923                 p.flags |= IORING_SETUP_SQPOLL;
 924                 if (sq_thread_cpu != -1) {
 925                         p.flags |= IORING_SETUP_SQ_AFF;
 926                         p.sq_thread_cpu = sq_thread_cpu;
 927                 }
 928         }
 929         if (pt) {
 930                 p.flags |= IORING_SETUP_SQE128;
 931                 p.flags |= IORING_SETUP_CQE32;
 932         }
 933
 934         fd = io_uring_setup(depth, &p);
 935         if (fd < 0) {
 936                 perror("io_uring_setup");
 937                 return 1;
 938         }
 939         s->ring_fd = s->enter_ring_fd = fd;
 940
 941         io_uring_probe(fd);
 942
 943         if (fixedbufs) {
 944                 struct rlimit rlim;
 945
 946                 rlim.rlim_cur = RLIM_INFINITY;
 947                 rlim.rlim_max = RLIM_INFINITY;
 948                 /* ignore potential error, not needed on newer kernels */
 949                 setrlimit(RLIMIT_MEMLOCK, &rlim);
 950
 951                 ret = io_uring_register_buffers(s);
 952                 if (ret < 0) {
 953                         perror("io_uring_register_buffers");
 954                         return 1;
 955                 }
 956
 957                 if (dma_map) {
 958                         ret = io_uring_map_buffers(s);
 959                         if (ret < 0) {
 960                                 perror("io_uring_map_buffers");
 961                                 return 1;
 962                         }
 963                 }
 964         }
 965
 966         if (register_files) {
 967                 ret = io_uring_register_files(s);
 968                 if (ret < 0) {
 969                         perror("io_uring_register_files");
 970                         return 1;
 971                 }
 972         }
 973
 974         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 975                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 976                         IORING_OFF_SQ_RING);
 977         sring->head = ptr + p.sq_off.head;
 978         sring->tail = ptr + p.sq_off.tail;
 979         sring->ring_mask = ptr + p.sq_off.ring_mask;
 980         sring->ring_entries = ptr + p.sq_off.ring_entries;
 981         sring->flags = ptr + p.sq_off.flags;
 982         sring->array = ptr + p.sq_off.array;
 983         sq_ring_mask = *sring->ring_mask;
 984
 985         if (p.flags & IORING_SETUP_SQE128)
 986                 len = 2 * p.sq_entries * sizeof(struct io_uring_sqe);
 987         else
 988                 len = p.sq_entries * sizeof(struct io_uring_sqe);
 989         s->sqes = mmap(0, len,
 990                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 991                         IORING_OFF_SQES);
 992
 993         if (p.flags & IORING_SETUP_CQE32) {
 994                 len = p.cq_off.cqes +
 995                         2 * p.cq_entries * sizeof(struct io_uring_cqe);
 996         } else {
 997                 len = p.cq_off.cqes +
 998                         p.cq_entries * sizeof(struct io_uring_cqe);
 999         }
1000         ptr = mmap(0, len,
1001                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
1002                         IORING_OFF_CQ_RING);
1003         cring->head = ptr + p.cq_off.head;
1004         cring->tail = ptr + p.cq_off.tail;
1005         cring->ring_mask = ptr + p.cq_off.ring_mask;
1006         cring->ring_entries = ptr + p.cq_off.ring_entries;
1007         cring->cqes = ptr + p.cq_off.cqes;
1008         cq_ring_mask = *cring->ring_mask;
1009
1010         for (i = 0; i < p.sq_entries; i++)
1011                 sring->array[i] = i;
1012
1013         return 0;
1014 }
1015
1016 static void *allocate_mem(struct submitter *s, int size)
1017 {
1018         void *buf;
1019
1020 #ifdef CONFIG_LIBNUMA
1021         if (s->numa_node != -1)
1022                 return numa_alloc_onnode(size, s->numa_node);
1023 #endif
1024
1025         if (posix_memalign(&buf, t_io_uring_page_size, bs)) {
1026                 printf("failed alloc\n");
1027                 return NULL;
1028         }
1029
1030         return buf;
1031 }
1032
1033 static int submitter_init(struct submitter *s)
1034 {
1035         int i, nr_batch, err;
1036         static int init_printed;
1037         char buf[80];
1038         s->tid = gettid();
1039         printf("submitter=%d, tid=%d, file=%s, node=%d\n", s->index, s->tid,
1040                                                         s->filename, s->numa_node);
1041
1042         set_affinity(s);
1043
1044         __init_rand64(&s->rand_state, s->tid);
1045         srand48(s->tid);
1046
1047         for (i = 0; i < MAX_FDS; i++)
1048                 s->files[i].fileno = i;
1049
1050         for (i = 0; i < roundup_pow2(depth); i++) {
1051                 void *buf;
1052
1053                 buf = allocate_mem(s, bs);
1054                 if (!buf)
1055                         return 1;
1056                 s->iovecs[i].iov_base = buf;
1057                 s->iovecs[i].iov_len = bs;
1058         }
1059
1060         if (use_sync) {
1061                 sprintf(buf, "Engine=preadv2\n");
1062                 err = 0;
1063         } else if (!aio) {
1064                 err = setup_ring(s);
1065                 sprintf(buf, "Engine=io_uring, sq_ring=%d, cq_ring=%d\n", *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
1066         } else {
1067                 sprintf(buf, "Engine=aio\n");
1068                 err = setup_aio(s);
1069         }
1070         if (err) {
1071                 printf("queue setup failed: %s, %d\n", strerror(errno), err);
1072                 return 1;
1073         }
1074
1075         if (!init_printed) {
1076                 printf("polled=%d, fixedbufs=%d/%d, register_files=%d, buffered=%d, QD=%d\n", polled, fixedbufs, dma_map, register_files, buffered, depth);
1077                 printf("%s", buf);
1078                 init_printed = 1;
1079         }
1080
1081         if (stats) {
1082                 nr_batch = roundup_pow2(depth / batch_submit);
1083                 if (nr_batch < 2)
1084                         nr_batch = 2;
1085                 s->clock_batch = calloc(nr_batch, sizeof(unsigned long));
1086                 s->clock_index = 1;
1087
1088                 s->plat = calloc(PLAT_NR, sizeof(unsigned long));
1089         } else {
1090                 s->clock_batch = NULL;
1091                 s->plat = NULL;
1092                 nr_batch = 0;
1093         }
1094         /* perform the expensive command initialization part for passthrough here
1095          * rather than in the fast path
1096          */
1097         if (pt) {
1098                 for (i = 0; i < roundup_pow2(depth); i++) {
1099                         struct io_uring_sqe *sqe = &s->sqes[i << 1];
1100
1101                         memset(&sqe->cmd, 0, sizeof(struct nvme_uring_cmd));
1102                 }
1103         }
1104         return nr_batch;
1105 }
1106
1107 #ifdef CONFIG_LIBAIO
1108 static int prep_more_ios_aio(struct submitter *s, int max_ios, struct iocb *iocbs)
1109 {
1110         uint64_t data;
1111         struct file *f;
1112         unsigned index;
1113
1114         index = 0;
1115         while (index < max_ios) {
1116                 struct iocb *iocb = &iocbs[index];
1117
1118                 if (s->nr_files == 1) {
1119                         f = &s->files[0];
1120                 } else {
1121                         f = &s->files[s->cur_file];
1122                         if (f->pending_ios >= file_depth(s)) {
1123                                 s->cur_file++;
1124                                 if (s->cur_file == s->nr_files)
1125                                         s->cur_file = 0;
1126                                 f = &s->files[s->cur_file];
1127                         }
1128                 }
1129                 f->pending_ios++;
1130
1131                 io_prep_pread(iocb, f->real_fd, s->iovecs[index].iov_base,
1132                                 s->iovecs[index].iov_len, get_offset(s, f));
1133
1134                 data = f->fileno;
1135                 if (stats && stats_running)
1136                         data |= (((uint64_t) s->clock_index) << 32);
1137                 iocb->data = (void *) (uintptr_t) data;
1138                 index++;
1139         }
1140         return index;
1141 }
1142
1143 static int reap_events_aio(struct submitter *s, struct io_event *events, int evs)
1144 {
1145         int last_idx = -1, stat_nr = 0;
1146         int reaped = 0;
1147
1148         while (evs) {
1149                 uint64_t data = (uintptr_t) events[reaped].data;
1150                 struct file *f = &s->files[data & 0xffffffff];
1151
1152                 f->pending_ios--;
1153                 if (events[reaped].res != bs) {
1154                         printf("io: unexpected ret=%ld\n", events[reaped].res);
1155                         return -1;
1156                 }
1157                 if (stats) {
1158                         int clock_index = data >> 32;
1159
1160                         if (last_idx != clock_index) {
1161                                 if (last_idx != -1) {
1162                                         add_stat(s, last_idx, stat_nr);
1163                                         stat_nr = 0;
1164                                 }
1165                                 last_idx = clock_index;
1166                         }
1167                         stat_nr++;
1168                 }
1169                 reaped++;
1170                 evs--;
1171         }
1172
1173         if (stat_nr)
1174                 add_stat(s, last_idx, stat_nr);
1175
1176         s->inflight -= reaped;
1177         s->done += reaped;
1178         return reaped;
1179 }
1180
1181 static void *submitter_aio_fn(void *data)
1182 {
1183         struct submitter *s = data;
1184         int i, ret, prepped;
1185         struct iocb **iocbsptr;
1186         struct iocb *iocbs;
1187         struct io_event *events;
1188 #ifdef ARCH_HAVE_CPU_CLOCK
1189         int nr_batch = submitter_init(s);
1190 #else
1191         submitter_init(s);
1192 #endif
1193
1194         iocbsptr = calloc(depth, sizeof(struct iocb *));
1195         iocbs = calloc(depth, sizeof(struct iocb));
1196         events = calloc(depth, sizeof(struct io_event));
1197
1198         for (i = 0; i < depth; i++)
1199                 iocbsptr[i] = &iocbs[i];
1200
1201         prepped = 0;
1202         do {
1203                 int to_wait, to_submit, to_prep;
1204
1205                 if (!prepped && s->inflight < depth) {
1206                         to_prep = min(depth - s->inflight, batch_submit);
1207                         prepped = prep_more_ios_aio(s, to_prep, iocbs);
1208 #ifdef ARCH_HAVE_CPU_CLOCK
1209                         if (prepped && stats) {
1210                                 s->clock_batch[s->clock_index] = get_cpu_clock();
1211                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
1212                         }
1213 #endif
1214                 }
1215                 s->inflight += prepped;
1216                 to_submit = prepped;
1217
1218                 if (to_submit && (s->inflight + to_submit <= depth))
1219                         to_wait = 0;
1220                 else
1221                         to_wait = min(s->inflight + to_submit, batch_complete);
1222
1223                 ret = io_submit(s->aio_ctx, to_submit, iocbsptr);
1224                 s->calls++;
1225                 if (ret < 0) {
1226                         perror("io_submit");
1227                         break;
1228                 } else if (ret != to_submit) {
1229                         printf("submitted %d, wanted %d\n", ret, to_submit);
1230                         break;
1231                 }
1232                 prepped = 0;
1233
1234                 while (to_wait) {
1235                         int r;
1236
1237                         s->calls++;
1238                         r = io_getevents(s->aio_ctx, to_wait, to_wait, events, NULL);
1239                         if (r < 0) {
1240                                 perror("io_getevents");
1241                                 break;
1242                         } else if (r != to_wait) {
1243                                 printf("r=%d, wait=%d\n", r, to_wait);
1244                                 break;
1245                         }
1246                         r = reap_events_aio(s, events, r);
1247                         s->reaps += r;
1248                         to_wait -= r;
1249                 }
1250         } while (!s->finish);
1251
1252         free(iocbsptr);
1253         free(iocbs);
1254         free(events);
1255         finish = 1;
1256         return NULL;
1257 }
1258 #endif
1259
1260 static void io_uring_unregister_ring(struct submitter *s)
1261 {
1262         struct io_uring_rsrc_update up = {
1263                 .offset = s->enter_ring_fd,
1264         };
1265
1266         syscall(__NR_io_uring_register, s->ring_fd, IORING_UNREGISTER_RING_FDS,
1267                 &up, 1);
1268 }
1269
1270 static int io_uring_register_ring(struct submitter *s)
1271 {
1272         struct io_uring_rsrc_update up = {
1273                 .data   = s->ring_fd,
1274                 .offset = -1U,
1275         };
1276         int ret;
1277
1278         ret = syscall(__NR_io_uring_register, s->ring_fd,
1279                         IORING_REGISTER_RING_FDS, &up, 1);
1280         if (ret == 1) {
1281                 s->enter_ring_fd = up.offset;
1282                 return 0;
1283         }
1284         register_ring = 0;
1285         return -1;
1286 }
1287
1288 static void *submitter_uring_fn(void *data)
1289 {
1290         struct submitter *s = data;
1291         struct io_sq_ring *ring = &s->sq_ring;
1292         int ret, prepped;
1293 #ifdef ARCH_HAVE_CPU_CLOCK
1294         int nr_batch = submitter_init(s);
1295 #else
1296         submitter_init(s);
1297 #endif
1298
1299         if (register_ring)
1300                 io_uring_register_ring(s);
1301
1302         prepped = 0;
1303         do {
1304                 int to_wait, to_submit, this_reap, to_prep;
1305                 unsigned ring_flags = 0;
1306
1307                 if (!prepped && s->inflight < depth) {
1308                         to_prep = min(depth - s->inflight, batch_submit);
1309                         prepped = prep_more_ios_uring(s, to_prep);
1310 #ifdef ARCH_HAVE_CPU_CLOCK
1311                         if (prepped && stats) {
1312                                 s->clock_batch[s->clock_index] = get_cpu_clock();
1313                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
1314                         }
1315 #endif
1316                 }
1317                 s->inflight += prepped;
1318 submit_more:
1319                 to_submit = prepped;
1320 submit:
1321                 if (to_submit && (s->inflight + to_submit <= depth))
1322                         to_wait = 0;
1323                 else
1324                         to_wait = min(s->inflight + to_submit, batch_complete);
1325
1326                 /*
1327                  * Only need to call io_uring_enter if we're not using SQ thread
1328                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
1329                  */
1330                 if (sq_thread_poll)
1331                         ring_flags = atomic_load_acquire(ring->flags);
1332                 if (!sq_thread_poll || ring_flags & IORING_SQ_NEED_WAKEUP) {
1333                         unsigned flags = 0;
1334
1335                         if (to_wait)
1336                                 flags = IORING_ENTER_GETEVENTS;
1337                         if (ring_flags & IORING_SQ_NEED_WAKEUP)
1338                                 flags |= IORING_ENTER_SQ_WAKEUP;
1339                         ret = io_uring_enter(s, to_submit, to_wait, flags);
1340                         s->calls++;
1341                 } else {
1342                         /* for SQPOLL, we submitted it all effectively */
1343                         ret = to_submit;
1344                 }
1345
1346                 /*
1347                  * For non SQ thread poll, we already got the events we needed
1348                  * through the io_uring_enter() above. For SQ thread poll, we
1349                  * need to loop here until we find enough events.
1350                  */
1351                 this_reap = 0;
1352                 do {
1353                         int r;
1354
1355                         if (pt)
1356                                 r = reap_events_uring_pt(s);
1357                         else
1358                                 r = reap_events_uring(s);
1359                         if (r == -1) {
1360                                 s->finish = 1;
1361                                 break;
1362                         } else if (r > 0)
1363                                 this_reap += r;
1364                 } while (sq_thread_poll && this_reap < to_wait);
1365                 s->reaps += this_reap;
1366
1367                 if (ret >= 0) {
1368                         if (!ret) {
1369                                 to_submit = 0;
1370                                 if (s->inflight)
1371                                         goto submit;
1372                                 continue;
1373                         } else if (ret < to_submit) {
1374                                 int diff = to_submit - ret;
1375
1376                                 s->done += ret;
1377                                 prepped -= diff;
1378                                 goto submit_more;
1379                         }
1380                         s->done += ret;
1381                         prepped = 0;
1382                         continue;
1383                 } else if (ret < 0) {
1384                         if (errno == EAGAIN) {
1385                                 if (s->finish)
1386                                         break;
1387                                 if (this_reap)
1388                                         goto submit;
1389                                 to_submit = 0;
1390                                 goto submit;
1391                         }
1392                         printf("io_submit: %s\n", strerror(errno));
1393                         break;
1394                 }
1395         } while (!s->finish);
1396
1397         if (register_ring)
1398                 io_uring_unregister_ring(s);
1399
1400         finish = 1;
1401         return NULL;
1402 }
1403
1404 #ifdef CONFIG_PWRITEV2
1405 static void *submitter_sync_fn(void *data)
1406 {
1407         struct submitter *s = data;
1408         int ret;
1409
1410         submitter_init(s);
1411
1412         do {
1413                 uint64_t offset;
1414                 struct file *f;
1415
1416                 if (s->nr_files == 1) {
1417                         f = &s->files[0];
1418                 } else {
1419                         f = &s->files[s->cur_file];
1420                         if (f->pending_ios >= file_depth(s)) {
1421                                 s->cur_file++;
1422                                 if (s->cur_file == s->nr_files)
1423                                         s->cur_file = 0;
1424                                 f = &s->files[s->cur_file];
1425                         }
1426                 }
1427                 f->pending_ios++;
1428
1429 #ifdef ARCH_HAVE_CPU_CLOCK
1430                 if (stats)
1431                         s->clock_batch[s->clock_index] = get_cpu_clock();
1432 #endif
1433
1434                 s->inflight++;
1435                 s->calls++;
1436
1437                 offset = get_offset(s, f);
1438                 if (polled)
1439                         ret = preadv2(f->real_fd, &s->iovecs[0], 1, offset, RWF_HIPRI);
1440                 else
1441                         ret = preadv2(f->real_fd, &s->iovecs[0], 1, offset, 0);
1442
1443                 if (ret < 0) {
1444                         perror("preadv2");
1445                         break;
1446                 } else if (ret != bs) {
1447                         break;
1448                 }
1449
1450                 s->done++;
1451                 s->inflight--;
1452                 f->pending_ios--;
1453                 if (stats)
1454                         add_stat(s, s->clock_index, 1);
1455         } while (!s->finish);
1456
1457         finish = 1;
1458         return NULL;
1459 }
1460 #else
1461 static void *submitter_sync_fn(void *data)
1462 {
1463         finish = 1;
1464         return NULL;
1465 }
1466 #endif
1467
1468 static struct submitter *get_submitter(int offset)
1469 {
1470         void *ret;
1471
1472         ret = submitter;
1473         if (offset)
1474                 ret += offset * (sizeof(*submitter) + depth * sizeof(struct iovec));
1475         return ret;
1476 }
1477
1478 static void do_finish(const char *reason)
1479 {
1480         int j;
1481
1482         printf("Exiting on %s\n", reason);
1483         for (j = 0; j < nthreads; j++) {
1484                 struct submitter *s = get_submitter(j);
1485                 s->finish = 1;
1486         }
1487         if (max_iops > 1000000) {
1488                 double miops = (double) max_iops / 1000000.0;
1489                 printf("Maximum IOPS=%.2fM\n", miops);
1490         } else if (max_iops > 100000) {
1491                 double kiops = (double) max_iops / 1000.0;
1492                 printf("Maximum IOPS=%.2fK\n", kiops);
1493         } else {
1494                 printf("Maximum IOPS=%lu\n", max_iops);
1495         }
1496         finish = 1;
1497 }
1498
1499 static void sig_int(int sig)
1500 {
1501         do_finish("signal");
1502 }
1503
1504 static void arm_sig_int(void)
1505 {
1506         struct sigaction act;
1507
1508         memset(&act, 0, sizeof(act));
1509         act.sa_handler = sig_int;
1510         act.sa_flags = SA_RESTART;
1511         sigaction(SIGINT, &act, NULL);
1512
1513         /* Windows uses SIGBREAK as a quit signal from other applications */
1514 #ifdef WIN32
1515         sigaction(SIGBREAK, &act, NULL);
1516 #endif
1517 }
1518
1519 static void usage(char *argv, int status)
1520 {
1521         char runtime_str[16];
1522         snprintf(runtime_str, sizeof(runtime_str), "%d", runtime);
1523         printf("%s [options] -- [filenames]\n"
1524                 " -d <int>  : IO Depth, default %d\n"
1525                 " -s <int>  : Batch submit, default %d\n"
1526                 " -c <int>  : Batch complete, default %d\n"
1527                 " -b <int>  : Block size, default %d\n"
1528                 " -p <bool> : Polled IO, default %d\n"
1529                 " -B <bool> : Fixed buffers, default %d\n"
1530                 " -D <bool> : DMA map fixed buffers, default %d\n"
1531                 " -F <bool> : Register files, default %d\n"
1532                 " -n <int>  : Number of threads, default %d\n"
1533                 " -O <bool> : Use O_DIRECT, default %d\n"
1534                 " -N <bool> : Perform just no-op requests, default %d\n"
1535                 " -t <bool> : Track IO latencies, default %d\n"
1536                 " -T <int>  : TSC rate in HZ\n"
1537                 " -r <int>  : Runtime in seconds, default %s\n"
1538                 " -R <bool> : Use random IO, default %d\n"
1539                 " -a <bool> : Use legacy aio, default %d\n"
1540                 " -S <bool> : Use sync IO (preadv2), default %d\n"
1541                 " -X <bool> : Use registered ring %d\n"
1542                 " -P <bool> : Automatically place on device home node %d\n"
1543                 " -u <bool> : Use nvme-passthrough I/O, default %d\n",
1544                 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled,
1545                 fixedbufs, dma_map, register_files, nthreads, !buffered, do_nop,
1546                 stats, runtime == 0 ? "unlimited" : runtime_str, random_io, aio,
1547                 use_sync, register_ring, numa_placement, pt);
1548         exit(status);
1549 }
1550
1551 static void read_tsc_rate(void)
1552 {
1553         char buffer[32];
1554         int fd, ret;
1555
1556         if (tsc_rate)
1557                 return;
1558
1559         fd = open(TSC_RATE_FILE, O_RDONLY);
1560         if (fd < 0)
1561                 return;
1562
1563         ret = read(fd, buffer, sizeof(buffer));
1564         if (ret < 0) {
1565                 close(fd);
1566                 return;
1567         }
1568
1569         tsc_rate = strtoul(buffer, NULL, 10);
1570         printf("Using TSC rate %luHz\n", tsc_rate);
1571         close(fd);
1572 }
1573
1574 static void write_tsc_rate(void)
1575 {
1576         char buffer[32];
1577         struct stat sb;
1578         int fd, ret;
1579
1580         if (!stat(TSC_RATE_FILE, &sb))
1581                 return;
1582
1583         fd = open(TSC_RATE_FILE, O_WRONLY | O_CREAT, 0644);
1584         if (fd < 0)
1585                 return;
1586
1587         memset(buffer, 0, sizeof(buffer));
1588         sprintf(buffer, "%lu", tsc_rate);
1589         ret = write(fd, buffer, strlen(buffer));
1590         if (ret < 0)
1591                 perror("write");
1592         close(fd);
1593 }
1594
1595 int main(int argc, char *argv[])
1596 {
1597         struct submitter *s;
1598         unsigned long done, calls, reap;
1599         int i, j, flags, fd, opt, threads_per_f, threads_rem = 0, nfiles;
1600         struct file f;
1601         void *ret;
1602
1603         if (!do_nop && argc < 2)
1604                 usage(argv[0], 1);
1605
1606         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) {
1607                 switch (opt) {
1608                 case 'a':
1609                         aio = !!atoi(optarg);
1610                         break;
1611                 case 'd':
1612                         depth = atoi(optarg);
1613                         break;
1614                 case 's':
1615                         batch_submit = atoi(optarg);
1616                         if (!batch_submit)
1617                                 batch_submit = 1;
1618                         break;
1619                 case 'c':
1620                         batch_complete = atoi(optarg);
1621                         if (!batch_complete)
1622                                 batch_complete = 1;
1623                         break;
1624                 case 'b':
1625                         bs = atoi(optarg);
1626                         break;
1627                 case 'p':
1628                         polled = !!atoi(optarg);
1629                         break;
1630                 case 'B':
1631                         fixedbufs = !!atoi(optarg);
1632                         break;
1633                 case 'F':
1634                         register_files = !!atoi(optarg);
1635                         break;
1636                 case 'n':
1637                         nthreads = atoi(optarg);
1638                         if (!nthreads) {
1639                                 printf("Threads must be non-zero\n");
1640                                 usage(argv[0], 1);
1641                         }
1642                         break;
1643                 case 'N':
1644                         do_nop = !!atoi(optarg);
1645                         break;
1646                 case 'O':
1647                         buffered = !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                         stats = !!atoi(optarg);
1655                         break;
1656                 case 'T':
1657 #ifndef ARCH_HAVE_CPU_CLOCK
1658                         fprintf(stderr, "Stats not supported on this CPU\n");
1659                         return 1;
1660 #endif
1661                         tsc_rate = strtoul(optarg, NULL, 10);
1662                         write_tsc_rate();
1663                         break;
1664                 case 'r':
1665                         runtime = atoi(optarg);
1666                         break;
1667                 case 'D':
1668                         dma_map = !!atoi(optarg);
1669                         break;
1670                 case 'R':
1671                         random_io = !!atoi(optarg);
1672                         break;
1673                 case 'X':
1674                         register_ring = !!atoi(optarg);
1675                         break;
1676                 case 'S':
1677 #ifdef CONFIG_PWRITEV2
1678                         use_sync = !!atoi(optarg);
1679 #else
1680                         fprintf(stderr, "preadv2 not supported\n");
1681                         exit(1);
1682 #endif
1683                         break;
1684                 case 'P':
1685                         numa_placement = !!atoi(optarg);
1686                         break;
1687                 case 'u':
1688                         pt = !!atoi(optarg);
1689                         break;
1690                 case 'h':
1691                 case '?':
1692                 default:
1693                         usage(argv[0], 0);
1694                         break;
1695                 }
1696         }
1697
1698         if (stats)
1699                 read_tsc_rate();
1700
1701         if (batch_complete > depth)
1702                 batch_complete = depth;
1703         if (batch_submit > depth)
1704                 batch_submit = depth;
1705         if (!fixedbufs && dma_map)
1706                 dma_map = 0;
1707
1708         submitter = calloc(nthreads, sizeof(*submitter) +
1709                                 roundup_pow2(depth) * sizeof(struct iovec));
1710         for (j = 0; j < nthreads; j++) {
1711                 s = get_submitter(j);
1712                 s->numa_node = -1;
1713                 s->index = j;
1714                 s->done = s->calls = s->reaps = 0;
1715         }
1716
1717         flags = O_RDONLY | O_NOATIME;
1718         if (!buffered)
1719                 flags |= O_DIRECT;
1720
1721         j = 0;
1722         i = optind;
1723         nfiles = argc - i;
1724         if (!do_nop) {
1725                 if (!nfiles) {
1726                         printf("No files specified\n");
1727                         usage(argv[0], 1);
1728                 }
1729                 threads_per_f = nthreads / nfiles;
1730                 /* make sure each thread gets assigned files */
1731                 if (threads_per_f == 0) {
1732                         threads_per_f = 1;
1733                 } else {
1734                         threads_rem = nthreads - threads_per_f * nfiles;
1735                 }
1736         }
1737         while (!do_nop && i < argc) {
1738                 int k, limit;
1739
1740                 memset(&f, 0, sizeof(f));
1741
1742                 fd = open(argv[i], flags);
1743                 if (fd < 0) {
1744                         perror("open");
1745                         return 1;
1746                 }
1747                 f.real_fd = fd;
1748                 if (get_file_size(&f)) {
1749                         printf("failed getting size of device/file\n");
1750                         return 1;
1751                 }
1752                 if (f.max_blocks <= 1) {
1753                         printf("Zero file/device size?\n");
1754                         return 1;
1755                 }
1756                 f.max_blocks--;
1757
1758                 limit = threads_per_f;
1759                 limit += threads_rem > 0 ? 1 : 0;
1760                 for (k = 0; k < limit; k++) {
1761                         s = get_submitter((j + k) % nthreads);
1762
1763                         if (s->nr_files == MAX_FDS) {
1764                                 printf("Max number of files (%d) reached\n", MAX_FDS);
1765                                 break;
1766                         }
1767
1768                         memcpy(&s->files[s->nr_files], &f, sizeof(f));
1769
1770                         if (numa_placement)
1771                                 detect_node(s, argv[i]);
1772
1773                         s->filename = argv[i];
1774                         s->nr_files++;
1775                 }
1776                 threads_rem--;
1777                 i++;
1778                 j += limit;
1779         }
1780
1781         arm_sig_int();
1782
1783         t_io_uring_page_size = sysconf(_SC_PAGESIZE);
1784         if (t_io_uring_page_size < 0)
1785                 t_io_uring_page_size = 4096;
1786
1787         for (j = 0; j < nthreads; j++) {
1788                 s = get_submitter(j);
1789                 if (use_sync)
1790                         pthread_create(&s->thread, NULL, submitter_sync_fn, s);
1791                 else if (!aio)
1792                         pthread_create(&s->thread, NULL, submitter_uring_fn, s);
1793 #ifdef CONFIG_LIBAIO
1794                 else
1795                         pthread_create(&s->thread, NULL, submitter_aio_fn, s);
1796 #endif
1797         }
1798
1799         reap = calls = done = 0;
1800         do {
1801                 unsigned long this_done = 0;
1802                 unsigned long this_reap = 0;
1803                 unsigned long this_call = 0;
1804                 unsigned long rpc = 0, ipc = 0;
1805                 unsigned long iops, bw;
1806
1807                 sleep(1);
1808                 if (runtime && !--runtime)
1809                         do_finish("timeout");
1810
1811                 /* don't print partial run, if interrupted by signal */
1812                 if (finish)
1813                         break;
1814
1815                 /* one second in to the run, enable stats */
1816                 if (stats)
1817                         stats_running = 1;
1818
1819                 for (j = 0; j < nthreads; j++) {
1820                         s = get_submitter(j);
1821                         this_done += s->done;
1822                         this_call += s->calls;
1823                         this_reap += s->reaps;
1824                 }
1825                 if (this_call - calls) {
1826                         rpc = (this_done - done) / (this_call - calls);
1827                         ipc = (this_reap - reap) / (this_call - calls);
1828                 } else
1829                         rpc = ipc = -1;
1830                 iops = this_done - done;
1831                 if (bs > 1048576)
1832                         bw = iops * (bs / 1048576);
1833                 else
1834                         bw = iops / (1048576 / bs);
1835                 if (iops > 1000000) {
1836                         double miops = (double) iops / 1000000.0;
1837                         printf("IOPS=%.2fM, ", miops);
1838                 } else if (iops > 100000) {
1839                         double kiops = (double) iops / 1000.0;
1840                         printf("IOPS=%.2fK, ", kiops);
1841                 } else {
1842                         printf("IOPS=%lu, ", iops);
1843                 }
1844                 max_iops = max(max_iops, iops);
1845                 if (!do_nop) {
1846                         if (bw > 2000) {
1847                                 double bw_g = (double) bw / 1000.0;
1848
1849                                 printf("BW=%.2fGiB/s, ", bw_g);
1850                         } else {
1851                                 printf("BW=%luMiB/s, ", bw);
1852                         }
1853                 }
1854                 printf("IOS/call=%ld/%ld\n", rpc, ipc);
1855                 done = this_done;
1856                 calls = this_call;
1857                 reap = this_reap;
1858         } while (!finish);
1859
1860         for (j = 0; j < nthreads; j++) {
1861                 s = get_submitter(j);
1862                 pthread_join(s->thread, &ret);
1863                 close(s->ring_fd);
1864
1865                 if (stats) {
1866                         unsigned long nr;
1867
1868                         printf("%d: Latency percentiles:\n", s->tid);
1869                         for (i = 0, nr = 0; i < PLAT_NR; i++)
1870                                 nr += s->plat[i];
1871                         show_clat_percentiles(s->plat, nr, 4);
1872                         free(s->clock_batch);
1873                         free(s->plat);
1874                 }
1875         }
1876
1877         free(submitter);
1878         return 0;
1879 }