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 "../lib/types.h"
  34 #include "../lib/roundup.h"
  35 #include "../lib/rand.h"
  36 #include "../minmax.h"
  37 #include "../os/linux/io_uring.h"
  38 #include "../engines/nvme.h"
  39
  40 struct io_sq_ring {
  41         unsigned *head;
  42         unsigned *tail;
  43         unsigned *ring_mask;
  44         unsigned *ring_entries;
  45         unsigned *flags;
  46         unsigned *array;
  47 };
  48
  49 struct io_cq_ring {
  50         unsigned *head;
  51         unsigned *tail;
  52         unsigned *ring_mask;
  53         unsigned *ring_entries;
  54         struct io_uring_cqe *cqes;
  55 };
  56
  57 #define DEPTH                   128
  58 #define BATCH_SUBMIT            32
  59 #define BATCH_COMPLETE          32
  60 #define BS                      4096
  61
  62 #define MAX_FDS                 16
  63
  64 static unsigned sq_ring_mask, cq_ring_mask;
  65
  66 struct file {
  67         unsigned long max_blocks;
  68         unsigned long max_size;
  69         unsigned long cur_off;
  70         unsigned pending_ios;
  71         unsigned int nsid;      /* nsid field required for nvme-passthrough */
  72         unsigned int lba_shift; /* lba_shift field required for nvme-passthrough */
  73         int real_fd;
  74         int fixed_fd;
  75         int fileno;
  76 };
  77
  78 #define PLAT_BITS               6
  79 #define PLAT_VAL                (1 << PLAT_BITS)
  80 #define PLAT_GROUP_NR           29
  81 #define PLAT_NR                 (PLAT_GROUP_NR * PLAT_VAL)
  82
  83 struct submitter {
  84         pthread_t thread;
  85         int ring_fd;
  86         int enter_ring_fd;
  87         int index;
  88         struct io_sq_ring sq_ring;
  89         struct io_uring_sqe *sqes;
  90         struct io_cq_ring cq_ring;
  91         int inflight;
  92         int tid;
  93         unsigned long reaps;
  94         unsigned long done;
  95         unsigned long calls;
  96         volatile int finish;
  97
  98         __s32 *fds;
  99
 100         struct taus258_state rand_state;
 101
 102         unsigned long *clock_batch;
 103         int clock_index;
 104         unsigned long *plat;
 105
 106 #ifdef CONFIG_LIBAIO
 107         io_context_t aio_ctx;
 108 #endif
 109
 110         int numa_node;
 111         const char *filename;
 112
 113         struct file files[MAX_FDS];
 114         unsigned nr_files;
 115         unsigned cur_file;
 116         struct iovec iovecs[];
 117 };
 118
 119 static struct submitter *submitter;
 120 static volatile int finish;
 121 static int stats_running;
 122 static unsigned long max_iops;
 123 static long t_io_uring_page_size;
 124
 125 static int depth = DEPTH;
 126 static int batch_submit = BATCH_SUBMIT;
 127 static int batch_complete = BATCH_COMPLETE;
 128 static int bs = BS;
 129 static int polled = 1;          /* use IO polling */
 130 static int fixedbufs = 1;       /* use fixed user buffers */
 131 static int dma_map;             /* pre-map DMA buffers */
 132 static int register_files = 1;  /* use fixed files */
 133 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
 134 static int sq_thread_poll = 0;  /* use kernel submission/poller thread */
 135 static int sq_thread_cpu = -1;  /* pin above thread to this CPU */
 136 static int do_nop = 0;          /* no-op SQ ring commands */
 137 static int nthreads = 1;
 138 static int stats = 0;           /* generate IO stats */
 139 static int aio = 0;             /* use libaio */
 140 static int runtime = 0;         /* runtime */
 141 static int random_io = 1;       /* random or sequential IO */
 142 static int register_ring = 1;   /* register ring */
 143 static int use_sync = 0;        /* use preadv2 */
 144 static int numa_placement = 0;  /* set to node of device */
 145 static int pt = 0;              /* passthrough I/O or not */
 146
 147 static unsigned long tsc_rate;
 148
 149 #define TSC_RATE_FILE   "tsc-rate"
 150
 151 static int vectored = 1;
 152
 153 static float plist[] = { 1.0, 5.0, 10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0,
 154                         80.0, 90.0, 95.0, 99.0, 99.5, 99.9, 99.95, 99.99 };
 155 static int plist_len = 17;
 156
 157 #ifndef IORING_REGISTER_MAP_BUFFERS
 158 #define IORING_REGISTER_MAP_BUFFERS     22
 159 struct io_uring_map_buffers {
 160         __s32   fd;
 161         __u32   buf_start;
 162         __u32   buf_end;
 163         __u32   flags;
 164         __u64   rsvd[2];
 165 };
 166 #endif
 167
 168 static int nvme_identify(int fd, __u32 nsid, enum nvme_identify_cns cns,
 169                          enum nvme_csi csi, void *data)
 170 {
 171         struct nvme_passthru_cmd cmd = {
 172                 .opcode         = nvme_admin_identify,
 173                 .nsid           = nsid,
 174                 .addr           = (__u64)(uintptr_t)data,
 175                 .data_len       = NVME_IDENTIFY_DATA_SIZE,
 176                 .cdw10          = cns,
 177                 .cdw11          = csi << NVME_IDENTIFY_CSI_SHIFT,
 178                 .timeout_ms     = NVME_DEFAULT_IOCTL_TIMEOUT,
 179         };
 180
 181         return ioctl(fd, NVME_IOCTL_ADMIN_CMD, &cmd);
 182 }
 183
 184 static int nvme_get_info(int fd, __u32 *nsid, __u32 *lba_sz, __u64 *nlba)
 185 {
 186         struct nvme_id_ns ns;
 187         int namespace_id;
 188         int err;
 189
 190         namespace_id = ioctl(fd, NVME_IOCTL_ID);
 191         if (namespace_id < 0) {
 192                 fprintf(stderr, "error failed to fetch namespace-id\n");
 193                 close(fd);
 194                 return -errno;
 195         }
 196
 197         /*
 198          * Identify namespace to get namespace-id, namespace size in LBA's
 199          * and LBA data size.
 200          */
 201         err = nvme_identify(fd, namespace_id, NVME_IDENTIFY_CNS_NS,
 202                                 NVME_CSI_NVM, &ns);
 203         if (err) {
 204                 fprintf(stderr, "error failed to fetch identify namespace\n");
 205                 close(fd);
 206                 return err;
 207         }
 208
 209         *nsid = namespace_id;
 210         *lba_sz = 1 << ns.lbaf[(ns.flbas & 0x0f)].ds;
 211         *nlba = ns.nsze;
 212
 213         return 0;
 214 }
 215
 216 static unsigned long cycles_to_nsec(unsigned long cycles)
 217 {
 218         uint64_t val;
 219
 220         if (!tsc_rate)
 221                 return cycles;
 222
 223         val = cycles * 1000000000ULL;
 224         return val / tsc_rate;
 225 }
 226
 227 static unsigned long plat_idx_to_val(unsigned int idx)
 228 {
 229         unsigned int error_bits;
 230         unsigned long k, base;
 231
 232         assert(idx < PLAT_NR);
 233
 234         /* MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 235          * all bits of the sample as index */
 236         if (idx < (PLAT_VAL << 1))
 237                 return cycles_to_nsec(idx);
 238
 239         /* Find the group and compute the minimum value of that group */
 240         error_bits = (idx >> PLAT_BITS) - 1;
 241         base = ((unsigned long) 1) << (error_bits + PLAT_BITS);
 242
 243         /* Find its bucket number of the group */
 244         k = idx % PLAT_VAL;
 245
 246         /* Return the mean of the range of the bucket */
 247         return cycles_to_nsec(base + ((k + 0.5) * (1 << error_bits)));
 248 }
 249
 250 unsigned int calculate_clat_percentiles(unsigned long *io_u_plat,
 251                 unsigned long nr, unsigned long **output,
 252                 unsigned long *maxv, unsigned long *minv)
 253 {
 254         unsigned long sum = 0;
 255         unsigned int len = plist_len, i, j = 0;
 256         unsigned long *ovals = NULL;
 257         bool is_last;
 258
 259         *minv = -1UL;
 260         *maxv = 0;
 261
 262         ovals = malloc(len * sizeof(*ovals));
 263         if (!ovals)
 264                 return 0;
 265
 266         /*
 267          * Calculate bucket values, note down max and min values
 268          */
 269         is_last = false;
 270         for (i = 0; i < PLAT_NR && !is_last; i++) {
 271                 sum += io_u_plat[i];
 272                 while (sum >= ((long double) plist[j] / 100.0 * nr)) {
 273                         assert(plist[j] <= 100.0);
 274
 275                         ovals[j] = plat_idx_to_val(i);
 276                         if (ovals[j] < *minv)
 277                                 *minv = ovals[j];
 278                         if (ovals[j] > *maxv)
 279                                 *maxv = ovals[j];
 280
 281                         is_last = (j == len - 1) != 0;
 282                         if (is_last)
 283                                 break;
 284
 285                         j++;
 286                 }
 287         }
 288
 289         if (!is_last)
 290                 fprintf(stderr, "error calculating latency percentiles\n");
 291
 292         *output = ovals;
 293         return len;
 294 }
 295
 296 static void show_clat_percentiles(unsigned long *io_u_plat, unsigned long nr,
 297                                   unsigned int precision)
 298 {
 299         unsigned int divisor, len, i, j = 0;
 300         unsigned long minv, maxv;
 301         unsigned long *ovals;
 302         int per_line, scale_down, time_width;
 303         bool is_last;
 304         char fmt[32];
 305
 306         len = calculate_clat_percentiles(io_u_plat, nr, &ovals, &maxv, &minv);
 307         if (!len || !ovals)
 308                 goto out;
 309
 310         if (!tsc_rate) {
 311                 scale_down = 0;
 312                 divisor = 1;
 313                 printf("    percentiles (tsc ticks):\n     |");
 314         } else if (minv > 2000 && maxv > 99999) {
 315                 scale_down = 1;
 316                 divisor = 1000;
 317                 printf("    percentiles (usec):\n     |");
 318         } else {
 319                 scale_down = 0;
 320                 divisor = 1;
 321                 printf("    percentiles (nsec):\n     |");
 322         }
 323
 324         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 325         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 326                         precision, time_width);
 327         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 328         per_line = (80 - 7) / (precision + 10 + time_width);
 329
 330         for (j = 0; j < len; j++) {
 331                 /* for formatting */
 332                 if (j != 0 && (j % per_line) == 0)
 333                         printf("     |");
 334
 335                 /* end of the list */
 336                 is_last = (j == len - 1) != 0;
 337
 338                 for (i = 0; i < scale_down; i++)
 339                         ovals[j] = (ovals[j] + 999) / 1000;
 340
 341                 printf(fmt, plist[j], ovals[j], is_last ? '\n' : ',');
 342
 343                 if (is_last)
 344                         break;
 345
 346                 if ((j % per_line) == per_line - 1)     /* for formatting */
 347                         printf("\n");
 348         }
 349
 350 out:
 351         free(ovals);
 352 }
 353
 354 #ifdef ARCH_HAVE_CPU_CLOCK
 355 static unsigned int plat_val_to_idx(unsigned long val)
 356 {
 357         unsigned int msb, error_bits, base, offset, idx;
 358
 359         /* Find MSB starting from bit 0 */
 360         if (val == 0)
 361                 msb = 0;
 362         else
 363                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
 364
 365         /*
 366          * MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 367          * all bits of the sample as index
 368          */
 369         if (msb <= PLAT_BITS)
 370                 return val;
 371
 372         /* Compute the number of error bits to discard*/
 373         error_bits = msb - PLAT_BITS;
 374
 375         /* Compute the number of buckets before the group */
 376         base = (error_bits + 1) << PLAT_BITS;
 377
 378         /*
 379          * Discard the error bits and apply the mask to find the
 380          * index for the buckets in the group
 381          */
 382         offset = (PLAT_VAL - 1) & (val >> error_bits);
 383
 384         /* Make sure the index does not exceed (array size - 1) */
 385         idx = (base + offset) < (PLAT_NR - 1) ?
 386                 (base + offset) : (PLAT_NR - 1);
 387
 388         return idx;
 389 }
 390 #endif
 391
 392 static void add_stat(struct submitter *s, int clock_index, int nr)
 393 {
 394 #ifdef ARCH_HAVE_CPU_CLOCK
 395         unsigned long cycles;
 396         unsigned int pidx;
 397
 398         if (!s->finish && clock_index) {
 399                 cycles = get_cpu_clock();
 400                 cycles -= s->clock_batch[clock_index];
 401                 pidx = plat_val_to_idx(cycles);
 402                 s->plat[pidx] += nr;
 403         }
 404 #endif
 405 }
 406
 407 static int io_uring_map_buffers(struct submitter *s)
 408 {
 409         struct io_uring_map_buffers map = {
 410                 .fd             = s->files[0].real_fd,
 411                 .buf_end        = depth,
 412         };
 413
 414         if (do_nop)
 415                 return 0;
 416         if (s->nr_files > 1)
 417                 fprintf(stdout, "Mapping buffers may not work with multiple files\n");
 418
 419         return syscall(__NR_io_uring_register, s->ring_fd,
 420                         IORING_REGISTER_MAP_BUFFERS, &map, 1);
 421 }
 422
 423 static int io_uring_register_buffers(struct submitter *s)
 424 {
 425         if (do_nop)
 426                 return 0;
 427
 428         return syscall(__NR_io_uring_register, s->ring_fd,
 429                         IORING_REGISTER_BUFFERS, s->iovecs, roundup_pow2(depth));
 430 }
 431
 432 static int io_uring_register_files(struct submitter *s)
 433 {
 434         int i;
 435
 436         if (do_nop)
 437                 return 0;
 438
 439         s->fds = calloc(s->nr_files, sizeof(__s32));
 440         for (i = 0; i < s->nr_files; i++) {
 441                 s->fds[i] = s->files[i].real_fd;
 442                 s->files[i].fixed_fd = i;
 443         }
 444
 445         return syscall(__NR_io_uring_register, s->ring_fd,
 446                         IORING_REGISTER_FILES, s->fds, s->nr_files);
 447 }
 448
 449 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 450 {
 451         /*
 452          * Clamp CQ ring size at our SQ ring size, we don't need more entries
 453          * than that.
 454          */
 455         p->flags |= IORING_SETUP_CQSIZE;
 456         p->cq_entries = entries;
 457
 458         return syscall(__NR_io_uring_setup, entries, p);
 459 }
 460
 461 static void io_uring_probe(int fd)
 462 {
 463         struct io_uring_probe *p;
 464         int ret;
 465
 466         p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 467         if (!p)
 468                 return;
 469
 470         memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 471         ret = syscall(__NR_io_uring_register, fd, IORING_REGISTER_PROBE, p, 256);
 472         if (ret < 0)
 473                 goto out;
 474
 475         if (IORING_OP_READ > p->ops_len)
 476                 goto out;
 477
 478         if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED))
 479                 vectored = 0;
 480 out:
 481         free(p);
 482 }
 483
 484 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 485                           unsigned int min_complete, unsigned int flags)
 486 {
 487         if (register_ring)
 488                 flags |= IORING_ENTER_REGISTERED_RING;
 489 #ifdef FIO_ARCH_HAS_SYSCALL
 490         return __do_syscall6(__NR_io_uring_enter, s->enter_ring_fd, to_submit,
 491                                 min_complete, flags, NULL, 0);
 492 #else
 493         return syscall(__NR_io_uring_enter, s->enter_ring_fd, to_submit,
 494                         min_complete, flags, NULL, 0);
 495 #endif
 496 }
 497
 498 #ifndef CONFIG_HAVE_GETTID
 499 static int gettid(void)
 500 {
 501         return syscall(__NR_gettid);
 502 }
 503 #endif
 504
 505 static unsigned file_depth(struct submitter *s)
 506 {
 507         return (depth + s->nr_files - 1) / s->nr_files;
 508 }
 509
 510 static void init_io(struct submitter *s, unsigned index)
 511 {
 512         struct io_uring_sqe *sqe = &s->sqes[index];
 513         unsigned long offset;
 514         struct file *f;
 515         long r;
 516
 517         if (do_nop) {
 518                 sqe->opcode = IORING_OP_NOP;
 519                 return;
 520         }
 521
 522         if (s->nr_files == 1) {
 523                 f = &s->files[0];
 524         } else {
 525                 f = &s->files[s->cur_file];
 526                 if (f->pending_ios >= file_depth(s)) {
 527                         s->cur_file++;
 528                         if (s->cur_file == s->nr_files)
 529                                 s->cur_file = 0;
 530                         f = &s->files[s->cur_file];
 531                 }
 532         }
 533         f->pending_ios++;
 534
 535         if (random_io) {
 536                 r = __rand64(&s->rand_state);
 537                 offset = (r % (f->max_blocks - 1)) * 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         if (register_files) {
 546                 sqe->flags = IOSQE_FIXED_FILE;
 547                 sqe->fd = f->fixed_fd;
 548         } else {
 549                 sqe->flags = 0;
 550                 sqe->fd = f->real_fd;
 551         }
 552         if (fixedbufs) {
 553                 sqe->opcode = IORING_OP_READ_FIXED;
 554                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 555                 sqe->len = bs;
 556                 sqe->buf_index = index;
 557         } else if (!vectored) {
 558                 sqe->opcode = IORING_OP_READ;
 559                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 560                 sqe->len = bs;
 561                 sqe->buf_index = 0;
 562         } else {
 563                 sqe->opcode = IORING_OP_READV;
 564                 sqe->addr = (unsigned long) &s->iovecs[index];
 565                 sqe->len = 1;
 566                 sqe->buf_index = 0;
 567         }
 568         sqe->ioprio = 0;
 569         sqe->off = offset;
 570         sqe->user_data = (unsigned long) f->fileno;
 571         if (stats && stats_running)
 572                 sqe->user_data |= ((uint64_t)s->clock_index << 32);
 573 }
 574
 575 static void init_io_pt(struct submitter *s, unsigned index)
 576 {
 577         struct io_uring_sqe *sqe = &s->sqes[index << 1];
 578         unsigned long offset;
 579         struct file *f;
 580         struct nvme_uring_cmd *cmd;
 581         unsigned long long slba;
 582         unsigned long long nlb;
 583         long r;
 584
 585         if (s->nr_files == 1) {
 586                 f = &s->files[0];
 587         } else {
 588                 f = &s->files[s->cur_file];
 589                 if (f->pending_ios >= file_depth(s)) {
 590                         s->cur_file++;
 591                         if (s->cur_file == s->nr_files)
 592                                 s->cur_file = 0;
 593                         f = &s->files[s->cur_file];
 594                 }
 595         }
 596         f->pending_ios++;
 597
 598         if (random_io) {
 599                 r = __rand64(&s->rand_state);
 600                 offset = (r % (f->max_blocks - 1)) * bs;
 601         } else {
 602                 offset = f->cur_off;
 603                 f->cur_off += bs;
 604                 if (f->cur_off + bs > f->max_size)
 605                         f->cur_off = 0;
 606         }
 607
 608         if (register_files) {
 609                 sqe->fd = f->fixed_fd;
 610                 sqe->flags = IOSQE_FIXED_FILE;
 611         } else {
 612                 sqe->fd = f->real_fd;
 613                 sqe->flags = 0;
 614         }
 615         sqe->opcode = IORING_OP_URING_CMD;
 616         sqe->user_data = (unsigned long) f->fileno;
 617         if (stats)
 618                 sqe->user_data |= ((unsigned long)s->clock_index << 32);
 619         sqe->cmd_op = NVME_URING_CMD_IO;
 620         slba = offset >> f->lba_shift;
 621         nlb = (bs >> f->lba_shift) - 1;
 622         cmd = (struct nvme_uring_cmd *)&sqe->cmd;
 623         /* cdw10 and cdw11 represent starting slba*/
 624         cmd->cdw10 = slba & 0xffffffff;
 625         cmd->cdw11 = slba >> 32;
 626         /* cdw12 represent number of lba to be read*/
 627         cmd->cdw12 = nlb;
 628         cmd->addr = (unsigned long) s->iovecs[index].iov_base;
 629         cmd->data_len = bs;
 630         cmd->nsid = f->nsid;
 631         cmd->opcode = 2;
 632 }
 633
 634 static int prep_more_ios_uring(struct submitter *s, int max_ios)
 635 {
 636         struct io_sq_ring *ring = &s->sq_ring;
 637         unsigned index, tail, next_tail, prepped = 0;
 638
 639         next_tail = tail = *ring->tail;
 640         do {
 641                 next_tail++;
 642                 if (next_tail == atomic_load_acquire(ring->head))
 643                         break;
 644
 645                 index = tail & sq_ring_mask;
 646                 if (pt)
 647                         init_io_pt(s, index);
 648                 else
 649                         init_io(s, index);
 650                 ring->array[index] = index;
 651                 prepped++;
 652                 tail = next_tail;
 653         } while (prepped < max_ios);
 654
 655         if (prepped)
 656                 atomic_store_release(ring->tail, tail);
 657         return prepped;
 658 }
 659
 660 static int get_file_size(struct file *f)
 661 {
 662         struct stat st;
 663
 664         if (fstat(f->real_fd, &st) < 0)
 665                 return -1;
 666         if (pt) {
 667                 __u64 nlba;
 668                 __u32 lbs;
 669                 int ret;
 670
 671                 if (!S_ISCHR(st.st_mode)) {
 672                         fprintf(stderr, "passthrough works with only nvme-ns "
 673                                         "generic devices (/dev/ngXnY)\n");
 674                         return -1;
 675                 }
 676                 ret = nvme_get_info(f->real_fd, &f->nsid, &lbs, &nlba);
 677                 if (ret)
 678                         return -1;
 679                 if ((bs % lbs) != 0) {
 680                         printf("error: bs:%d should be a multiple logical_block_size:%d\n",
 681                                         bs, lbs);
 682                         return -1;
 683                 }
 684                 f->max_blocks = nlba / bs;
 685                 f->max_size = nlba;
 686                 f->lba_shift = ilog2(lbs);
 687                 return 0;
 688         } else if (S_ISBLK(st.st_mode)) {
 689                 unsigned long long bytes;
 690
 691                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 692                         return -1;
 693
 694                 f->max_blocks = bytes / bs;
 695                 f->max_size = bytes;
 696                 return 0;
 697         } else if (S_ISREG(st.st_mode)) {
 698                 f->max_blocks = st.st_size / bs;
 699                 f->max_size = st.st_size;
 700                 return 0;
 701         }
 702
 703         return -1;
 704 }
 705
 706 static int reap_events_uring(struct submitter *s)
 707 {
 708         struct io_cq_ring *ring = &s->cq_ring;
 709         struct io_uring_cqe *cqe;
 710         unsigned head, reaped = 0;
 711         int last_idx = -1, stat_nr = 0;
 712
 713         head = *ring->head;
 714         do {
 715                 struct file *f;
 716
 717                 read_barrier();
 718                 if (head == atomic_load_acquire(ring->tail))
 719                         break;
 720                 cqe = &ring->cqes[head & cq_ring_mask];
 721                 if (!do_nop) {
 722                         int fileno = cqe->user_data & 0xffffffff;
 723
 724                         f = &s->files[fileno];
 725                         f->pending_ios--;
 726                         if (cqe->res != bs) {
 727                                 printf("io: unexpected ret=%d\n", cqe->res);
 728                                 if (polled && cqe->res == -EOPNOTSUPP)
 729                                         printf("Your filesystem/driver/kernel doesn't support polled IO\n");
 730                                 return -1;
 731                         }
 732                 }
 733                 if (stats) {
 734                         int clock_index = cqe->user_data >> 32;
 735
 736                         if (last_idx != clock_index) {
 737                                 if (last_idx != -1) {
 738                                         add_stat(s, last_idx, stat_nr);
 739                                         stat_nr = 0;
 740                                 }
 741                                 last_idx = clock_index;
 742                         }
 743                         stat_nr++;
 744                 }
 745                 reaped++;
 746                 head++;
 747         } while (1);
 748
 749         if (stat_nr)
 750                 add_stat(s, last_idx, stat_nr);
 751
 752         if (reaped) {
 753                 s->inflight -= reaped;
 754                 atomic_store_release(ring->head, head);
 755         }
 756         return reaped;
 757 }
 758
 759 static int reap_events_uring_pt(struct submitter *s)
 760 {
 761         struct io_cq_ring *ring = &s->cq_ring;
 762         struct io_uring_cqe *cqe;
 763         unsigned head, reaped = 0;
 764         int last_idx = -1, stat_nr = 0;
 765         unsigned index;
 766         int fileno;
 767
 768         head = *ring->head;
 769         do {
 770                 struct file *f;
 771
 772                 read_barrier();
 773                 if (head == atomic_load_acquire(ring->tail))
 774                         break;
 775                 index = head & cq_ring_mask;
 776                 cqe = &ring->cqes[index << 1];
 777                 fileno = cqe->user_data & 0xffffffff;
 778                 f = &s->files[fileno];
 779                 f->pending_ios--;
 780
 781                 if (cqe->res != 0) {
 782                         printf("io: unexpected ret=%d\n", cqe->res);
 783                         if (polled && cqe->res == -EINVAL)
 784                                 printf("passthrough doesn't support polled IO\n");
 785                         return -1;
 786                 }
 787                 if (stats) {
 788                         int clock_index = cqe->user_data >> 32;
 789
 790                         if (last_idx != clock_index) {
 791                                 if (last_idx != -1) {
 792                                         add_stat(s, last_idx, stat_nr);
 793                                         stat_nr = 0;
 794                                 }
 795                                 last_idx = clock_index;
 796                         }
 797                         stat_nr++;
 798                 }
 799                 reaped++;
 800                 head++;
 801         } while (1);
 802
 803         if (stat_nr)
 804                 add_stat(s, last_idx, stat_nr);
 805
 806         if (reaped) {
 807                 s->inflight -= reaped;
 808                 atomic_store_release(ring->head, head);
 809         }
 810         return reaped;
 811 }
 812
 813 static void set_affinity(struct submitter *s)
 814 {
 815 #ifdef CONFIG_LIBNUMA
 816         struct bitmask *mask;
 817
 818         if (s->numa_node == -1)
 819                 return;
 820
 821         numa_set_preferred(s->numa_node);
 822
 823         mask = numa_allocate_cpumask();
 824         numa_node_to_cpus(s->numa_node, mask);
 825         numa_sched_setaffinity(s->tid, mask);
 826 #endif
 827 }
 828
 829 static int detect_node(struct submitter *s, const char *name)
 830 {
 831 #ifdef CONFIG_LIBNUMA
 832         const char *base = basename(name);
 833         char str[128];
 834         int ret, fd, node;
 835
 836         sprintf(str, "/sys/block/%s/device/numa_node", base);
 837         fd = open(str, O_RDONLY);
 838         if (fd < 0)
 839                 return -1;
 840
 841         ret = read(fd, str, sizeof(str));
 842         if (ret < 0) {
 843                 close(fd);
 844                 return -1;
 845         }
 846         node = atoi(str);
 847         s->numa_node = node;
 848         close(fd);
 849 #else
 850         s->numa_node = -1;
 851 #endif
 852         return 0;
 853 }
 854
 855 static int setup_aio(struct submitter *s)
 856 {
 857 #ifdef CONFIG_LIBAIO
 858         if (polled) {
 859                 fprintf(stderr, "aio does not support polled IO\n");
 860                 polled = 0;
 861         }
 862         if (sq_thread_poll) {
 863                 fprintf(stderr, "aio does not support SQPOLL IO\n");
 864                 sq_thread_poll = 0;
 865         }
 866         if (do_nop) {
 867                 fprintf(stderr, "aio does not support polled IO\n");
 868                 do_nop = 0;
 869         }
 870         if (fixedbufs || register_files) {
 871                 fprintf(stderr, "aio does not support registered files or buffers\n");
 872                 fixedbufs = register_files = 0;
 873         }
 874
 875         return io_queue_init(roundup_pow2(depth), &s->aio_ctx);
 876 #else
 877         fprintf(stderr, "Legacy AIO not available on this system/build\n");
 878         errno = EINVAL;
 879         return -1;
 880 #endif
 881 }
 882
 883 static int setup_ring(struct submitter *s)
 884 {
 885         struct io_sq_ring *sring = &s->sq_ring;
 886         struct io_cq_ring *cring = &s->cq_ring;
 887         struct io_uring_params p;
 888         int ret, fd;
 889         void *ptr;
 890         size_t len;
 891
 892         memset(&p, 0, sizeof(p));
 893
 894         if (polled && !do_nop)
 895                 p.flags |= IORING_SETUP_IOPOLL;
 896         if (sq_thread_poll) {
 897                 p.flags |= IORING_SETUP_SQPOLL;
 898                 if (sq_thread_cpu != -1) {
 899                         p.flags |= IORING_SETUP_SQ_AFF;
 900                         p.sq_thread_cpu = sq_thread_cpu;
 901                 }
 902         }
 903         if (pt) {
 904                 p.flags |= IORING_SETUP_SQE128;
 905                 p.flags |= IORING_SETUP_CQE32;
 906         }
 907
 908         fd = io_uring_setup(depth, &p);
 909         if (fd < 0) {
 910                 perror("io_uring_setup");
 911                 return 1;
 912         }
 913         s->ring_fd = s->enter_ring_fd = fd;
 914
 915         io_uring_probe(fd);
 916
 917         if (fixedbufs) {
 918                 struct rlimit rlim;
 919
 920                 rlim.rlim_cur = RLIM_INFINITY;
 921                 rlim.rlim_max = RLIM_INFINITY;
 922                 /* ignore potential error, not needed on newer kernels */
 923                 setrlimit(RLIMIT_MEMLOCK, &rlim);
 924
 925                 ret = io_uring_register_buffers(s);
 926                 if (ret < 0) {
 927                         perror("io_uring_register_buffers");
 928                         return 1;
 929                 }
 930
 931                 if (dma_map) {
 932                         ret = io_uring_map_buffers(s);
 933                         if (ret < 0) {
 934                                 perror("io_uring_map_buffers");
 935                                 return 1;
 936                         }
 937                 }
 938         }
 939
 940         if (register_files) {
 941                 ret = io_uring_register_files(s);
 942                 if (ret < 0) {
 943                         perror("io_uring_register_files");
 944                         return 1;
 945                 }
 946         }
 947
 948         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 949                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 950                         IORING_OFF_SQ_RING);
 951         sring->head = ptr + p.sq_off.head;
 952         sring->tail = ptr + p.sq_off.tail;
 953         sring->ring_mask = ptr + p.sq_off.ring_mask;
 954         sring->ring_entries = ptr + p.sq_off.ring_entries;
 955         sring->flags = ptr + p.sq_off.flags;
 956         sring->array = ptr + p.sq_off.array;
 957         sq_ring_mask = *sring->ring_mask;
 958
 959         if (p.flags & IORING_SETUP_SQE128)
 960                 len = 2 * p.sq_entries * sizeof(struct io_uring_sqe);
 961         else
 962                 len = p.sq_entries * sizeof(struct io_uring_sqe);
 963         s->sqes = mmap(0, len,
 964                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 965                         IORING_OFF_SQES);
 966
 967         if (p.flags & IORING_SETUP_CQE32) {
 968                 len = p.cq_off.cqes +
 969                         2 * p.cq_entries * sizeof(struct io_uring_cqe);
 970         } else {
 971                 len = p.cq_off.cqes +
 972                         p.cq_entries * sizeof(struct io_uring_cqe);
 973         }
 974         ptr = mmap(0, len,
 975                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 976                         IORING_OFF_CQ_RING);
 977         cring->head = ptr + p.cq_off.head;
 978         cring->tail = ptr + p.cq_off.tail;
 979         cring->ring_mask = ptr + p.cq_off.ring_mask;
 980         cring->ring_entries = ptr + p.cq_off.ring_entries;
 981         cring->cqes = ptr + p.cq_off.cqes;
 982         cq_ring_mask = *cring->ring_mask;
 983         return 0;
 984 }
 985
 986 static void *allocate_mem(struct submitter *s, int size)
 987 {
 988         void *buf;
 989
 990 #ifdef CONFIG_LIBNUMA
 991         if (s->numa_node != -1)
 992                 return numa_alloc_onnode(size, s->numa_node);
 993 #endif
 994
 995         if (posix_memalign(&buf, t_io_uring_page_size, bs)) {
 996                 printf("failed alloc\n");
 997                 return NULL;
 998         }
 999
1000         return buf;
1001 }
1002
1003 static int submitter_init(struct submitter *s)
1004 {
1005         int i, nr_batch, err;
1006         static int init_printed;
1007         char buf[80];
1008         s->tid = gettid();
1009         printf("submitter=%d, tid=%d, file=%s, node=%d\n", s->index, s->tid,
1010                                                         s->filename, s->numa_node);
1011
1012         set_affinity(s);
1013
1014         __init_rand64(&s->rand_state, s->tid);
1015         srand48(s->tid);
1016
1017         for (i = 0; i < MAX_FDS; i++)
1018                 s->files[i].fileno = i;
1019
1020         for (i = 0; i < roundup_pow2(depth); i++) {
1021                 void *buf;
1022
1023                 buf = allocate_mem(s, bs);
1024                 if (!buf)
1025                         return 1;
1026                 s->iovecs[i].iov_base = buf;
1027                 s->iovecs[i].iov_len = bs;
1028         }
1029
1030         if (use_sync) {
1031                 sprintf(buf, "Engine=preadv2\n");
1032                 err = 0;
1033         } else if (!aio) {
1034                 err = setup_ring(s);
1035                 sprintf(buf, "Engine=io_uring, sq_ring=%d, cq_ring=%d\n", *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
1036         } else {
1037                 sprintf(buf, "Engine=aio\n");
1038                 err = setup_aio(s);
1039         }
1040         if (err) {
1041                 printf("queue setup failed: %s, %d\n", strerror(errno), err);
1042                 return 1;
1043         }
1044
1045         if (!init_printed) {
1046                 printf("polled=%d, fixedbufs=%d/%d, register_files=%d, buffered=%d, QD=%d\n", polled, fixedbufs, dma_map, register_files, buffered, depth);
1047                 printf("%s", buf);
1048                 init_printed = 1;
1049         }
1050
1051         if (stats) {
1052                 nr_batch = roundup_pow2(depth / batch_submit);
1053                 if (nr_batch < 2)
1054                         nr_batch = 2;
1055                 s->clock_batch = calloc(nr_batch, sizeof(unsigned long));
1056                 s->clock_index = 1;
1057
1058                 s->plat = calloc(PLAT_NR, sizeof(unsigned long));
1059         } else {
1060                 s->clock_batch = NULL;
1061                 s->plat = NULL;
1062                 nr_batch = 0;
1063         }
1064         /* perform the expensive command initialization part for passthrough here
1065          * rather than in the fast path
1066          */
1067         if (pt) {
1068                 for (i = 0; i < roundup_pow2(depth); i++) {
1069                         struct io_uring_sqe *sqe = &s->sqes[i << 1];
1070
1071                         memset(&sqe->cmd, 0, sizeof(struct nvme_uring_cmd));
1072                 }
1073         }
1074         return nr_batch;
1075 }
1076
1077 #ifdef CONFIG_LIBAIO
1078 static int prep_more_ios_aio(struct submitter *s, int max_ios, struct iocb *iocbs)
1079 {
1080         uint64_t data;
1081         long long offset;
1082         struct file *f;
1083         unsigned index;
1084         long r;
1085
1086         index = 0;
1087         while (index < max_ios) {
1088                 struct iocb *iocb = &iocbs[index];
1089
1090                 if (s->nr_files == 1) {
1091                         f = &s->files[0];
1092                 } else {
1093                         f = &s->files[s->cur_file];
1094                         if (f->pending_ios >= file_depth(s)) {
1095                                 s->cur_file++;
1096                                 if (s->cur_file == s->nr_files)
1097                                         s->cur_file = 0;
1098                                 f = &s->files[s->cur_file];
1099                         }
1100                 }
1101                 f->pending_ios++;
1102
1103                 r = lrand48();
1104                 offset = (r % (f->max_blocks - 1)) * bs;
1105                 io_prep_pread(iocb, f->real_fd, s->iovecs[index].iov_base,
1106                                 s->iovecs[index].iov_len, offset);
1107
1108                 data = f->fileno;
1109                 if (stats && stats_running)
1110                         data |= (((uint64_t) s->clock_index) << 32);
1111                 iocb->data = (void *) (uintptr_t) data;
1112                 index++;
1113         }
1114         return index;
1115 }
1116
1117 static int reap_events_aio(struct submitter *s, struct io_event *events, int evs)
1118 {
1119         int last_idx = -1, stat_nr = 0;
1120         int reaped = 0;
1121
1122         while (evs) {
1123                 uint64_t data = (uintptr_t) events[reaped].data;
1124                 struct file *f = &s->files[data & 0xffffffff];
1125
1126                 f->pending_ios--;
1127                 if (events[reaped].res != bs) {
1128                         printf("io: unexpected ret=%ld\n", events[reaped].res);
1129                         return -1;
1130                 }
1131                 if (stats) {
1132                         int clock_index = data >> 32;
1133
1134                         if (last_idx != clock_index) {
1135                                 if (last_idx != -1) {
1136                                         add_stat(s, last_idx, stat_nr);
1137                                         stat_nr = 0;
1138                                 }
1139                                 last_idx = clock_index;
1140                         }
1141                         stat_nr++;
1142                 }
1143                 reaped++;
1144                 evs--;
1145         }
1146
1147         if (stat_nr)
1148                 add_stat(s, last_idx, stat_nr);
1149
1150         s->inflight -= reaped;
1151         s->done += reaped;
1152         return reaped;
1153 }
1154
1155 static void *submitter_aio_fn(void *data)
1156 {
1157         struct submitter *s = data;
1158         int i, ret, prepped;
1159         struct iocb **iocbsptr;
1160         struct iocb *iocbs;
1161         struct io_event *events;
1162 #ifdef ARCH_HAVE_CPU_CLOCK
1163         int nr_batch = submitter_init(s);
1164 #else
1165         submitter_init(s);
1166 #endif
1167
1168         iocbsptr = calloc(depth, sizeof(struct iocb *));
1169         iocbs = calloc(depth, sizeof(struct iocb));
1170         events = calloc(depth, sizeof(struct io_event));
1171
1172         for (i = 0; i < depth; i++)
1173                 iocbsptr[i] = &iocbs[i];
1174
1175         prepped = 0;
1176         do {
1177                 int to_wait, to_submit, to_prep;
1178
1179                 if (!prepped && s->inflight < depth) {
1180                         to_prep = min(depth - s->inflight, batch_submit);
1181                         prepped = prep_more_ios_aio(s, to_prep, iocbs);
1182 #ifdef ARCH_HAVE_CPU_CLOCK
1183                         if (prepped && stats) {
1184                                 s->clock_batch[s->clock_index] = get_cpu_clock();
1185                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
1186                         }
1187 #endif
1188                 }
1189                 s->inflight += prepped;
1190                 to_submit = prepped;
1191
1192                 if (to_submit && (s->inflight + to_submit <= depth))
1193                         to_wait = 0;
1194                 else
1195                         to_wait = min(s->inflight + to_submit, batch_complete);
1196
1197                 ret = io_submit(s->aio_ctx, to_submit, iocbsptr);
1198                 s->calls++;
1199                 if (ret < 0) {
1200                         perror("io_submit");
1201                         break;
1202                 } else if (ret != to_submit) {
1203                         printf("submitted %d, wanted %d\n", ret, to_submit);
1204                         break;
1205                 }
1206                 prepped = 0;
1207
1208                 while (to_wait) {
1209                         int r;
1210
1211                         s->calls++;
1212                         r = io_getevents(s->aio_ctx, to_wait, to_wait, events, NULL);
1213                         if (r < 0) {
1214                                 perror("io_getevents");
1215                                 break;
1216                         } else if (r != to_wait) {
1217                                 printf("r=%d, wait=%d\n", r, to_wait);
1218                                 break;
1219                         }
1220                         r = reap_events_aio(s, events, r);
1221                         s->reaps += r;
1222                         to_wait -= r;
1223                 }
1224         } while (!s->finish);
1225
1226         free(iocbsptr);
1227         free(iocbs);
1228         free(events);
1229         finish = 1;
1230         return NULL;
1231 }
1232 #endif
1233
1234 static void io_uring_unregister_ring(struct submitter *s)
1235 {
1236         struct io_uring_rsrc_update up = {
1237                 .offset = s->enter_ring_fd,
1238         };
1239
1240         syscall(__NR_io_uring_register, s->ring_fd, IORING_UNREGISTER_RING_FDS,
1241                 &up, 1);
1242 }
1243
1244 static int io_uring_register_ring(struct submitter *s)
1245 {
1246         struct io_uring_rsrc_update up = {
1247                 .data   = s->ring_fd,
1248                 .offset = -1U,
1249         };
1250         int ret;
1251
1252         ret = syscall(__NR_io_uring_register, s->ring_fd,
1253                         IORING_REGISTER_RING_FDS, &up, 1);
1254         if (ret == 1) {
1255                 s->enter_ring_fd = up.offset;
1256                 return 0;
1257         }
1258         register_ring = 0;
1259         return -1;
1260 }
1261
1262 static void *submitter_uring_fn(void *data)
1263 {
1264         struct submitter *s = data;
1265         struct io_sq_ring *ring = &s->sq_ring;
1266         int ret, prepped;
1267 #ifdef ARCH_HAVE_CPU_CLOCK
1268         int nr_batch = submitter_init(s);
1269 #else
1270         submitter_init(s);
1271 #endif
1272
1273         if (register_ring)
1274                 io_uring_register_ring(s);
1275
1276         prepped = 0;
1277         do {
1278                 int to_wait, to_submit, this_reap, to_prep;
1279                 unsigned ring_flags = 0;
1280
1281                 if (!prepped && s->inflight < depth) {
1282                         to_prep = min(depth - s->inflight, batch_submit);
1283                         prepped = prep_more_ios_uring(s, to_prep);
1284 #ifdef ARCH_HAVE_CPU_CLOCK
1285                         if (prepped && stats) {
1286                                 s->clock_batch[s->clock_index] = get_cpu_clock();
1287                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
1288                         }
1289 #endif
1290                 }
1291                 s->inflight += prepped;
1292 submit_more:
1293                 to_submit = prepped;
1294 submit:
1295                 if (to_submit && (s->inflight + to_submit <= depth))
1296                         to_wait = 0;
1297                 else
1298                         to_wait = min(s->inflight + to_submit, batch_complete);
1299
1300                 /*
1301                  * Only need to call io_uring_enter if we're not using SQ thread
1302                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
1303                  */
1304                 if (sq_thread_poll)
1305                         ring_flags = atomic_load_acquire(ring->flags);
1306                 if (!sq_thread_poll || ring_flags & IORING_SQ_NEED_WAKEUP) {
1307                         unsigned flags = 0;
1308
1309                         if (to_wait)
1310                                 flags = IORING_ENTER_GETEVENTS;
1311                         if (ring_flags & IORING_SQ_NEED_WAKEUP)
1312                                 flags |= IORING_ENTER_SQ_WAKEUP;
1313                         ret = io_uring_enter(s, to_submit, to_wait, flags);
1314                         s->calls++;
1315                 } else {
1316                         /* for SQPOLL, we submitted it all effectively */
1317                         ret = to_submit;
1318                 }
1319
1320                 /*
1321                  * For non SQ thread poll, we already got the events we needed
1322                  * through the io_uring_enter() above. For SQ thread poll, we
1323                  * need to loop here until we find enough events.
1324                  */
1325                 this_reap = 0;
1326                 do {
1327                         int r;
1328
1329                         if (pt)
1330                                 r = reap_events_uring_pt(s);
1331                         else
1332                                 r = reap_events_uring(s);
1333                         if (r == -1) {
1334                                 s->finish = 1;
1335                                 break;
1336                         } else if (r > 0)
1337                                 this_reap += r;
1338                 } while (sq_thread_poll && this_reap < to_wait);
1339                 s->reaps += this_reap;
1340
1341                 if (ret >= 0) {
1342                         if (!ret) {
1343                                 to_submit = 0;
1344                                 if (s->inflight)
1345                                         goto submit;
1346                                 continue;
1347                         } else if (ret < to_submit) {
1348                                 int diff = to_submit - ret;
1349
1350                                 s->done += ret;
1351                                 prepped -= diff;
1352                                 goto submit_more;
1353                         }
1354                         s->done += ret;
1355                         prepped = 0;
1356                         continue;
1357                 } else if (ret < 0) {
1358                         if (errno == EAGAIN) {
1359                                 if (s->finish)
1360                                         break;
1361                                 if (this_reap)
1362                                         goto submit;
1363                                 to_submit = 0;
1364                                 goto submit;
1365                         }
1366                         printf("io_submit: %s\n", strerror(errno));
1367                         break;
1368                 }
1369         } while (!s->finish);
1370
1371         if (register_ring)
1372                 io_uring_unregister_ring(s);
1373
1374         finish = 1;
1375         return NULL;
1376 }
1377
1378 #ifdef CONFIG_PWRITEV2
1379 static void *submitter_sync_fn(void *data)
1380 {
1381         struct submitter *s = data;
1382         int ret;
1383
1384         submitter_init(s);
1385
1386         do {
1387                 uint64_t offset;
1388                 struct file *f;
1389                 long r;
1390
1391                 if (s->nr_files == 1) {
1392                         f = &s->files[0];
1393                 } else {
1394                         f = &s->files[s->cur_file];
1395                         if (f->pending_ios >= file_depth(s)) {
1396                                 s->cur_file++;
1397                                 if (s->cur_file == s->nr_files)
1398                                         s->cur_file = 0;
1399                                 f = &s->files[s->cur_file];
1400                         }
1401                 }
1402                 f->pending_ios++;
1403
1404                 if (random_io) {
1405                         r = __rand64(&s->rand_state);
1406                         offset = (r % (f->max_blocks - 1)) * bs;
1407                 } else {
1408                         offset = f->cur_off;
1409                         f->cur_off += bs;
1410                         if (f->cur_off + bs > f->max_size)
1411                                 f->cur_off = 0;
1412                 }
1413
1414 #ifdef ARCH_HAVE_CPU_CLOCK
1415                 if (stats)
1416                         s->clock_batch[s->clock_index] = get_cpu_clock();
1417 #endif
1418
1419                 s->inflight++;
1420                 s->calls++;
1421
1422                 if (polled)
1423                         ret = preadv2(f->real_fd, &s->iovecs[0], 1, offset, RWF_HIPRI);
1424                 else
1425                         ret = preadv2(f->real_fd, &s->iovecs[0], 1, offset, 0);
1426
1427                 if (ret < 0) {
1428                         perror("preadv2");
1429                         break;
1430                 } else if (ret != bs) {
1431                         break;
1432                 }
1433
1434                 s->done++;
1435                 s->inflight--;
1436                 f->pending_ios--;
1437                 if (stats)
1438                         add_stat(s, s->clock_index, 1);
1439         } while (!s->finish);
1440
1441         finish = 1;
1442         return NULL;
1443 }
1444 #else
1445 static void *submitter_sync_fn(void *data)
1446 {
1447         finish = 1;
1448         return NULL;
1449 }
1450 #endif
1451
1452 static struct submitter *get_submitter(int offset)
1453 {
1454         void *ret;
1455
1456         ret = submitter;
1457         if (offset)
1458                 ret += offset * (sizeof(*submitter) + depth * sizeof(struct iovec));
1459         return ret;
1460 }
1461
1462 static void do_finish(const char *reason)
1463 {
1464         int j;
1465
1466         printf("Exiting on %s\n", reason);
1467         for (j = 0; j < nthreads; j++) {
1468                 struct submitter *s = get_submitter(j);
1469                 s->finish = 1;
1470         }
1471         if (max_iops > 1000000) {
1472                 double miops = (double) max_iops / 1000000.0;
1473                 printf("Maximum IOPS=%.2fM\n", miops);
1474         } else if (max_iops > 100000) {
1475                 double kiops = (double) max_iops / 1000.0;
1476                 printf("Maximum IOPS=%.2fK\n", kiops);
1477         } else {
1478                 printf("Maximum IOPS=%lu\n", max_iops);
1479         }
1480         finish = 1;
1481 }
1482
1483 static void sig_int(int sig)
1484 {
1485         do_finish("signal");
1486 }
1487
1488 static void arm_sig_int(void)
1489 {
1490         struct sigaction act;
1491
1492         memset(&act, 0, sizeof(act));
1493         act.sa_handler = sig_int;
1494         act.sa_flags = SA_RESTART;
1495         sigaction(SIGINT, &act, NULL);
1496
1497         /* Windows uses SIGBREAK as a quit signal from other applications */
1498 #ifdef WIN32
1499         sigaction(SIGBREAK, &act, NULL);
1500 #endif
1501 }
1502
1503 static void usage(char *argv, int status)
1504 {
1505         char runtime_str[16];
1506         snprintf(runtime_str, sizeof(runtime_str), "%d", runtime);
1507         printf("%s [options] -- [filenames]\n"
1508                 " -d <int>  : IO Depth, default %d\n"
1509                 " -s <int>  : Batch submit, default %d\n"
1510                 " -c <int>  : Batch complete, default %d\n"
1511                 " -b <int>  : Block size, default %d\n"
1512                 " -p <bool> : Polled IO, default %d\n"
1513                 " -B <bool> : Fixed buffers, default %d\n"
1514                 " -D <bool> : DMA map fixed buffers, default %d\n"
1515                 " -F <bool> : Register files, default %d\n"
1516                 " -n <int>  : Number of threads, default %d\n"
1517                 " -O <bool> : Use O_DIRECT, default %d\n"
1518                 " -N <bool> : Perform just no-op requests, default %d\n"
1519                 " -t <bool> : Track IO latencies, default %d\n"
1520                 " -T <int>  : TSC rate in HZ\n"
1521                 " -r <int>  : Runtime in seconds, default %s\n"
1522                 " -R <bool> : Use random IO, default %d\n"
1523                 " -a <bool> : Use legacy aio, default %d\n"
1524                 " -S <bool> : Use sync IO (preadv2), default %d\n"
1525                 " -X <bool> : Use registered ring %d\n"
1526                 " -P <bool> : Automatically place on device home node %d\n"
1527                 " -u <bool> : Use nvme-passthrough I/O, default %d\n",
1528                 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled,
1529                 fixedbufs, dma_map, register_files, nthreads, !buffered, do_nop,
1530                 stats, runtime == 0 ? "unlimited" : runtime_str, random_io, aio,
1531                 use_sync, register_ring, numa_placement, pt);
1532         exit(status);
1533 }
1534
1535 static void read_tsc_rate(void)
1536 {
1537         char buffer[32];
1538         int fd, ret;
1539
1540         if (tsc_rate)
1541                 return;
1542
1543         fd = open(TSC_RATE_FILE, O_RDONLY);
1544         if (fd < 0)
1545                 return;
1546
1547         ret = read(fd, buffer, sizeof(buffer));
1548         if (ret < 0) {
1549                 close(fd);
1550                 return;
1551         }
1552
1553         tsc_rate = strtoul(buffer, NULL, 10);
1554         printf("Using TSC rate %luHz\n", tsc_rate);
1555         close(fd);
1556 }
1557
1558 static void write_tsc_rate(void)
1559 {
1560         char buffer[32];
1561         struct stat sb;
1562         int fd, ret;
1563
1564         if (!stat(TSC_RATE_FILE, &sb))
1565                 return;
1566
1567         fd = open(TSC_RATE_FILE, O_WRONLY | O_CREAT, 0644);
1568         if (fd < 0)
1569                 return;
1570
1571         memset(buffer, 0, sizeof(buffer));
1572         sprintf(buffer, "%lu", tsc_rate);
1573         ret = write(fd, buffer, strlen(buffer));
1574         if (ret < 0)
1575                 perror("write");
1576         close(fd);
1577 }
1578
1579 int main(int argc, char *argv[])
1580 {
1581         struct submitter *s;
1582         unsigned long done, calls, reap;
1583         int i, j, flags, fd, opt, threads_per_f, threads_rem = 0, nfiles;
1584         struct file f;
1585         void *ret;
1586
1587         if (!do_nop && argc < 2)
1588                 usage(argv[0], 1);
1589
1590         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) {
1591                 switch (opt) {
1592                 case 'a':
1593                         aio = !!atoi(optarg);
1594                         break;
1595                 case 'd':
1596                         depth = atoi(optarg);
1597                         break;
1598                 case 's':
1599                         batch_submit = atoi(optarg);
1600                         if (!batch_submit)
1601                                 batch_submit = 1;
1602                         break;
1603                 case 'c':
1604                         batch_complete = atoi(optarg);
1605                         if (!batch_complete)
1606                                 batch_complete = 1;
1607                         break;
1608                 case 'b':
1609                         bs = atoi(optarg);
1610                         break;
1611                 case 'p':
1612                         polled = !!atoi(optarg);
1613                         break;
1614                 case 'B':
1615                         fixedbufs = !!atoi(optarg);
1616                         break;
1617                 case 'F':
1618                         register_files = !!atoi(optarg);
1619                         break;
1620                 case 'n':
1621                         nthreads = atoi(optarg);
1622                         if (!nthreads) {
1623                                 printf("Threads must be non-zero\n");
1624                                 usage(argv[0], 1);
1625                         }
1626                         break;
1627                 case 'N':
1628                         do_nop = !!atoi(optarg);
1629                         break;
1630                 case 'O':
1631                         buffered = !atoi(optarg);
1632                         break;
1633                 case 't':
1634 #ifndef ARCH_HAVE_CPU_CLOCK
1635                         fprintf(stderr, "Stats not supported on this CPU\n");
1636                         return 1;
1637 #endif
1638                         stats = !!atoi(optarg);
1639                         break;
1640                 case 'T':
1641 #ifndef ARCH_HAVE_CPU_CLOCK
1642                         fprintf(stderr, "Stats not supported on this CPU\n");
1643                         return 1;
1644 #endif
1645                         tsc_rate = strtoul(optarg, NULL, 10);
1646                         write_tsc_rate();
1647                         break;
1648                 case 'r':
1649                         runtime = atoi(optarg);
1650                         break;
1651                 case 'D':
1652                         dma_map = !!atoi(optarg);
1653                         break;
1654                 case 'R':
1655                         random_io = !!atoi(optarg);
1656                         break;
1657                 case 'X':
1658                         register_ring = !!atoi(optarg);
1659                         break;
1660                 case 'S':
1661 #ifdef CONFIG_PWRITEV2
1662                         use_sync = !!atoi(optarg);
1663 #else
1664                         fprintf(stderr, "preadv2 not supported\n");
1665                         exit(1);
1666 #endif
1667                         break;
1668                 case 'P':
1669                         numa_placement = !!atoi(optarg);
1670                         break;
1671                 case 'u':
1672                         pt = !!atoi(optarg);
1673                         break;
1674                 case 'h':
1675                 case '?':
1676                 default:
1677                         usage(argv[0], 0);
1678                         break;
1679                 }
1680         }
1681
1682         if (stats)
1683                 read_tsc_rate();
1684
1685         if (batch_complete > depth)
1686                 batch_complete = depth;
1687         if (batch_submit > depth)
1688                 batch_submit = depth;
1689         if (!fixedbufs && dma_map)
1690                 dma_map = 0;
1691
1692         submitter = calloc(nthreads, sizeof(*submitter) +
1693                                 roundup_pow2(depth) * sizeof(struct iovec));
1694         for (j = 0; j < nthreads; j++) {
1695                 s = get_submitter(j);
1696                 s->numa_node = -1;
1697                 s->index = j;
1698                 s->done = s->calls = s->reaps = 0;
1699         }
1700
1701         flags = O_RDONLY | O_NOATIME;
1702         if (!buffered)
1703                 flags |= O_DIRECT;
1704
1705         j = 0;
1706         i = optind;
1707         nfiles = argc - i;
1708         if (!do_nop) {
1709                 if (!nfiles) {
1710                         printf("No files specified\n");
1711                         usage(argv[0], 1);
1712                 }
1713                 threads_per_f = nthreads / nfiles;
1714                 /* make sure each thread gets assigned files */
1715                 if (threads_per_f == 0) {
1716                         threads_per_f = 1;
1717                 } else {
1718                         threads_rem = nthreads - threads_per_f * nfiles;
1719                 }
1720         }
1721         while (!do_nop && i < argc) {
1722                 int k, limit;
1723
1724                 memset(&f, 0, sizeof(f));
1725
1726                 fd = open(argv[i], flags);
1727                 if (fd < 0) {
1728                         perror("open");
1729                         return 1;
1730                 }
1731                 f.real_fd = fd;
1732                 if (get_file_size(&f)) {
1733                         printf("failed getting size of device/file\n");
1734                         return 1;
1735                 }
1736                 if (f.max_blocks <= 1) {
1737                         printf("Zero file/device size?\n");
1738                         return 1;
1739                 }
1740                 f.max_blocks--;
1741
1742                 limit = threads_per_f;
1743                 limit += threads_rem > 0 ? 1 : 0;
1744                 for (k = 0; k < limit; k++) {
1745                         s = get_submitter((j + k) % nthreads);
1746
1747                         if (s->nr_files == MAX_FDS) {
1748                                 printf("Max number of files (%d) reached\n", MAX_FDS);
1749                                 break;
1750                         }
1751
1752                         memcpy(&s->files[s->nr_files], &f, sizeof(f));
1753
1754                         if (numa_placement)
1755                                 detect_node(s, argv[i]);
1756
1757                         s->filename = argv[i];
1758                         s->nr_files++;
1759                 }
1760                 threads_rem--;
1761                 i++;
1762                 j += limit;
1763         }
1764
1765         arm_sig_int();
1766
1767         t_io_uring_page_size = sysconf(_SC_PAGESIZE);
1768         if (t_io_uring_page_size < 0)
1769                 t_io_uring_page_size = 4096;
1770
1771         for (j = 0; j < nthreads; j++) {
1772                 s = get_submitter(j);
1773                 if (use_sync)
1774                         pthread_create(&s->thread, NULL, submitter_sync_fn, s);
1775                 else if (!aio)
1776                         pthread_create(&s->thread, NULL, submitter_uring_fn, s);
1777 #ifdef CONFIG_LIBAIO
1778                 else
1779                         pthread_create(&s->thread, NULL, submitter_aio_fn, s);
1780 #endif
1781         }
1782
1783         reap = calls = done = 0;
1784         do {
1785                 unsigned long this_done = 0;
1786                 unsigned long this_reap = 0;
1787                 unsigned long this_call = 0;
1788                 unsigned long rpc = 0, ipc = 0;
1789                 unsigned long iops, bw;
1790
1791                 sleep(1);
1792                 if (runtime && !--runtime)
1793                         do_finish("timeout");
1794
1795                 /* don't print partial run, if interrupted by signal */
1796                 if (finish)
1797                         break;
1798
1799                 /* one second in to the run, enable stats */
1800                 if (stats)
1801                         stats_running = 1;
1802
1803                 for (j = 0; j < nthreads; j++) {
1804                         s = get_submitter(j);
1805                         this_done += s->done;
1806                         this_call += s->calls;
1807                         this_reap += s->reaps;
1808                 }
1809                 if (this_call - calls) {
1810                         rpc = (this_done - done) / (this_call - calls);
1811                         ipc = (this_reap - reap) / (this_call - calls);
1812                 } else
1813                         rpc = ipc = -1;
1814                 iops = this_done - done;
1815                 if (bs > 1048576)
1816                         bw = iops * (bs / 1048576);
1817                 else
1818                         bw = iops / (1048576 / bs);
1819                 if (iops > 1000000) {
1820                         double miops = (double) iops / 1000000.0;
1821                         printf("IOPS=%.2fM, ", miops);
1822                 } else if (iops > 100000) {
1823                         double kiops = (double) iops / 1000.0;
1824                         printf("IOPS=%.2fK, ", kiops);
1825                 } else {
1826                         printf("IOPS=%lu, ", iops);
1827                 }
1828                 max_iops = max(max_iops, iops);
1829                 if (!do_nop) {
1830                         if (bw > 2000) {
1831                                 double bw_g = (double) bw / 1000.0;
1832
1833                                 printf("BW=%.2fGiB/s, ", bw_g);
1834                         } else {
1835                                 printf("BW=%luMiB/s, ", bw);
1836                         }
1837                 }
1838                 printf("IOS/call=%ld/%ld\n", rpc, ipc);
1839                 done = this_done;
1840                 calls = this_call;
1841                 reap = this_reap;
1842         } while (!finish);
1843
1844         for (j = 0; j < nthreads; j++) {
1845                 s = get_submitter(j);
1846                 pthread_join(s->thread, &ret);
1847                 close(s->ring_fd);
1848
1849                 if (stats) {
1850                         unsigned long nr;
1851
1852                         printf("%d: Latency percentiles:\n", s->tid);
1853                         for (i = 0, nr = 0; i < PLAT_NR; i++)
1854                                 nr += s->plat[i];
1855                         show_clat_percentiles(s->plat, nr, 4);
1856                         free(s->clock_batch);
1857                         free(s->plat);
1858                 }
1859         }
1860
1861         free(submitter);
1862         return 0;
1863 }