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         unsigned long io_errors;
  98         volatile int finish;
  99
 100         __s32 *fds;
 101
 102         struct taus258_state rand_state;
 103
 104         unsigned long *clock_batch;
 105         int clock_index;
 106         unsigned long *plat;
 107
 108 #ifdef CONFIG_LIBAIO
 109         io_context_t aio_ctx;
 110 #endif
 111
 112         int numa_node;
 113         int per_file_depth;
 114         const char *filename;
 115
 116         struct file files[MAX_FDS];
 117         unsigned nr_files;
 118         unsigned cur_file;
 119         struct iovec iovecs[];
 120 };
 121
 122 static struct submitter *submitter;
 123 static volatile int finish;
 124 static int stats_running;
 125 static unsigned long max_iops;
 126 static long t_io_uring_page_size;
 127
 128 static int depth = DEPTH;
 129 static int batch_submit = BATCH_SUBMIT;
 130 static int batch_complete = BATCH_COMPLETE;
 131 static int bs = BS;
 132 static int polled = 1;          /* use IO polling */
 133 static int fixedbufs = 1;       /* use fixed user buffers */
 134 static int register_files = 1;  /* use fixed files */
 135 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
 136 static int sq_thread_poll = 0;  /* use kernel submission/poller thread */
 137 static int sq_thread_cpu = -1;  /* pin above thread to this CPU */
 138 static int do_nop = 0;          /* no-op SQ ring commands */
 139 static int nthreads = 1;
 140 static int stats = 0;           /* generate IO stats */
 141 static int aio = 0;             /* use libaio */
 142 static int runtime = 0;         /* runtime */
 143 static int random_io = 1;       /* random or sequential IO */
 144 static int register_ring = 1;   /* register ring */
 145 static int use_sync = 0;        /* use preadv2 */
 146 static int numa_placement = 0;  /* set to node of device */
 147 static int pt = 0;              /* passthrough I/O or not */
 148
 149 static unsigned long tsc_rate;
 150
 151 #define TSC_RATE_FILE   "tsc-rate"
 152
 153 static int vectored = 1;
 154
 155 static float plist[] = { 1.0, 5.0, 10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0,
 156                         80.0, 90.0, 95.0, 99.0, 99.5, 99.9, 99.95, 99.99 };
 157 static int plist_len = 17;
 158
 159 static int nvme_identify(int fd, __u32 nsid, enum nvme_identify_cns cns,
 160                          enum nvme_csi csi, void *data)
 161 {
 162         struct nvme_passthru_cmd cmd = {
 163                 .opcode         = nvme_admin_identify,
 164                 .nsid           = nsid,
 165                 .addr           = (__u64)(uintptr_t)data,
 166                 .data_len       = NVME_IDENTIFY_DATA_SIZE,
 167                 .cdw10          = cns,
 168                 .cdw11          = csi << NVME_IDENTIFY_CSI_SHIFT,
 169                 .timeout_ms     = NVME_DEFAULT_IOCTL_TIMEOUT,
 170         };
 171
 172         return ioctl(fd, NVME_IOCTL_ADMIN_CMD, &cmd);
 173 }
 174
 175 static int nvme_get_info(int fd, __u32 *nsid, __u32 *lba_sz, __u64 *nlba)
 176 {
 177         struct nvme_id_ns ns;
 178         int namespace_id;
 179         int err;
 180
 181         namespace_id = ioctl(fd, NVME_IOCTL_ID);
 182         if (namespace_id < 0) {
 183                 fprintf(stderr, "error failed to fetch namespace-id\n");
 184                 close(fd);
 185                 return -errno;
 186         }
 187
 188         /*
 189          * Identify namespace to get namespace-id, namespace size in LBA's
 190          * and LBA data size.
 191          */
 192         err = nvme_identify(fd, namespace_id, NVME_IDENTIFY_CNS_NS,
 193                                 NVME_CSI_NVM, &ns);
 194         if (err) {
 195                 fprintf(stderr, "error failed to fetch identify namespace\n");
 196                 close(fd);
 197                 return err;
 198         }
 199
 200         *nsid = namespace_id;
 201         *lba_sz = 1 << ns.lbaf[(ns.flbas & 0x0f)].ds;
 202         *nlba = ns.nsze;
 203
 204         return 0;
 205 }
 206
 207 static unsigned long cycles_to_nsec(unsigned long cycles)
 208 {
 209         uint64_t val;
 210
 211         if (!tsc_rate)
 212                 return cycles;
 213
 214         val = cycles * 1000000000ULL;
 215         return val / tsc_rate;
 216 }
 217
 218 static unsigned long plat_idx_to_val(unsigned int idx)
 219 {
 220         unsigned int error_bits;
 221         unsigned long k, base;
 222
 223         assert(idx < PLAT_NR);
 224
 225         /* MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 226          * all bits of the sample as index */
 227         if (idx < (PLAT_VAL << 1))
 228                 return cycles_to_nsec(idx);
 229
 230         /* Find the group and compute the minimum value of that group */
 231         error_bits = (idx >> PLAT_BITS) - 1;
 232         base = ((unsigned long) 1) << (error_bits + PLAT_BITS);
 233
 234         /* Find its bucket number of the group */
 235         k = idx % PLAT_VAL;
 236
 237         /* Return the mean of the range of the bucket */
 238         return cycles_to_nsec(base + ((k + 0.5) * (1 << error_bits)));
 239 }
 240
 241 unsigned int calculate_clat_percentiles(unsigned long *io_u_plat,
 242                 unsigned long nr, unsigned long **output,
 243                 unsigned long *maxv, unsigned long *minv)
 244 {
 245         unsigned long sum = 0;
 246         unsigned int len = plist_len, i, j = 0;
 247         unsigned long *ovals = NULL;
 248         bool is_last;
 249
 250         *minv = -1UL;
 251         *maxv = 0;
 252
 253         ovals = malloc(len * sizeof(*ovals));
 254         if (!ovals)
 255                 return 0;
 256
 257         /*
 258          * Calculate bucket values, note down max and min values
 259          */
 260         is_last = false;
 261         for (i = 0; i < PLAT_NR && !is_last; i++) {
 262                 sum += io_u_plat[i];
 263                 while (sum >= ((long double) plist[j] / 100.0 * nr)) {
 264                         assert(plist[j] <= 100.0);
 265
 266                         ovals[j] = plat_idx_to_val(i);
 267                         if (ovals[j] < *minv)
 268                                 *minv = ovals[j];
 269                         if (ovals[j] > *maxv)
 270                                 *maxv = ovals[j];
 271
 272                         is_last = (j == len - 1) != 0;
 273                         if (is_last)
 274                                 break;
 275
 276                         j++;
 277                 }
 278         }
 279
 280         if (!is_last)
 281                 fprintf(stderr, "error calculating latency percentiles\n");
 282
 283         *output = ovals;
 284         return len;
 285 }
 286
 287 static void show_clat_percentiles(unsigned long *io_u_plat, unsigned long nr,
 288                                   unsigned int precision)
 289 {
 290         unsigned int divisor, len, i, j = 0;
 291         unsigned long minv, maxv;
 292         unsigned long *ovals;
 293         int per_line, scale_down, time_width;
 294         bool is_last;
 295         char fmt[32];
 296
 297         len = calculate_clat_percentiles(io_u_plat, nr, &ovals, &maxv, &minv);
 298         if (!len || !ovals)
 299                 goto out;
 300
 301         if (!tsc_rate) {
 302                 scale_down = 0;
 303                 divisor = 1;
 304                 printf("    percentiles (tsc ticks):\n     |");
 305         } else if (minv > 2000 && maxv > 99999) {
 306                 scale_down = 1;
 307                 divisor = 1000;
 308                 printf("    percentiles (usec):\n     |");
 309         } else {
 310                 scale_down = 0;
 311                 divisor = 1;
 312                 printf("    percentiles (nsec):\n     |");
 313         }
 314
 315         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 316         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 317                         precision, time_width);
 318         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 319         per_line = (80 - 7) / (precision + 10 + time_width);
 320
 321         for (j = 0; j < len; j++) {
 322                 /* for formatting */
 323                 if (j != 0 && (j % per_line) == 0)
 324                         printf("     |");
 325
 326                 /* end of the list */
 327                 is_last = (j == len - 1) != 0;
 328
 329                 for (i = 0; i < scale_down; i++)
 330                         ovals[j] = (ovals[j] + 999) / 1000;
 331
 332                 printf(fmt, plist[j], ovals[j], is_last ? '\n' : ',');
 333
 334                 if (is_last)
 335                         break;
 336
 337                 if ((j % per_line) == per_line - 1)     /* for formatting */
 338                         printf("\n");
 339         }
 340
 341 out:
 342         free(ovals);
 343 }
 344
 345 #ifdef ARCH_HAVE_CPU_CLOCK
 346 static unsigned int plat_val_to_idx(unsigned long val)
 347 {
 348         unsigned int msb, error_bits, base, offset, idx;
 349
 350         /* Find MSB starting from bit 0 */
 351         if (val == 0)
 352                 msb = 0;
 353         else
 354                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
 355
 356         /*
 357          * MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 358          * all bits of the sample as index
 359          */
 360         if (msb <= PLAT_BITS)
 361                 return val;
 362
 363         /* Compute the number of error bits to discard*/
 364         error_bits = msb - PLAT_BITS;
 365
 366         /* Compute the number of buckets before the group */
 367         base = (error_bits + 1) << PLAT_BITS;
 368
 369         /*
 370          * Discard the error bits and apply the mask to find the
 371          * index for the buckets in the group
 372          */
 373         offset = (PLAT_VAL - 1) & (val >> error_bits);
 374
 375         /* Make sure the index does not exceed (array size - 1) */
 376         idx = (base + offset) < (PLAT_NR - 1) ?
 377                 (base + offset) : (PLAT_NR - 1);
 378
 379         return idx;
 380 }
 381 #endif
 382
 383 static void add_stat(struct submitter *s, int clock_index, int nr)
 384 {
 385 #ifdef ARCH_HAVE_CPU_CLOCK
 386         unsigned long cycles;
 387         unsigned int pidx;
 388
 389         if (!s->finish && clock_index) {
 390                 cycles = get_cpu_clock();
 391                 cycles -= s->clock_batch[clock_index];
 392                 pidx = plat_val_to_idx(cycles);
 393                 s->plat[pidx] += nr;
 394         }
 395 #endif
 396 }
 397
 398 static int io_uring_register_buffers(struct submitter *s)
 399 {
 400         if (do_nop)
 401                 return 0;
 402
 403         return syscall(__NR_io_uring_register, s->ring_fd,
 404                         IORING_REGISTER_BUFFERS, s->iovecs, roundup_pow2(depth));
 405 }
 406
 407 static int io_uring_register_files(struct submitter *s)
 408 {
 409         int i;
 410
 411         if (do_nop)
 412                 return 0;
 413
 414         s->fds = calloc(s->nr_files, sizeof(__s32));
 415         for (i = 0; i < s->nr_files; i++) {
 416                 s->fds[i] = s->files[i].real_fd;
 417                 s->files[i].fixed_fd = i;
 418         }
 419
 420         return syscall(__NR_io_uring_register, s->ring_fd,
 421                         IORING_REGISTER_FILES, s->fds, s->nr_files);
 422 }
 423
 424 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 425 {
 426         int ret;
 427
 428         /*
 429          * Clamp CQ ring size at our SQ ring size, we don't need more entries
 430          * than that.
 431          */
 432         p->flags |= IORING_SETUP_CQSIZE;
 433         p->cq_entries = entries;
 434
 435         p->flags |= IORING_SETUP_COOP_TASKRUN;
 436         p->flags |= IORING_SETUP_SINGLE_ISSUER;
 437         p->flags |= IORING_SETUP_DEFER_TASKRUN;
 438 retry:
 439         ret = syscall(__NR_io_uring_setup, entries, p);
 440         if (!ret)
 441                 return 0;
 442
 443         if (errno == EINVAL && p->flags & IORING_SETUP_COOP_TASKRUN) {
 444                 p->flags &= ~IORING_SETUP_COOP_TASKRUN;
 445                 goto retry;
 446         }
 447         if (errno == EINVAL && p->flags & IORING_SETUP_SINGLE_ISSUER) {
 448                 p->flags &= ~IORING_SETUP_SINGLE_ISSUER;
 449                 goto retry;
 450         }
 451         if (errno == EINVAL && p->flags & IORING_SETUP_DEFER_TASKRUN) {
 452                 p->flags &= ~IORING_SETUP_DEFER_TASKRUN;
 453                 goto retry;
 454         }
 455
 456         return ret;
 457 }
 458
 459 static void io_uring_probe(int fd)
 460 {
 461         struct io_uring_probe *p;
 462         int ret;
 463
 464         p = calloc(1, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 465         if (!p)
 466                 return;
 467
 468         ret = syscall(__NR_io_uring_register, fd, IORING_REGISTER_PROBE, p, 256);
 469         if (ret < 0)
 470                 goto out;
 471
 472         if (IORING_OP_READ > p->ops_len)
 473                 goto out;
 474
 475         if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED))
 476                 vectored = 0;
 477 out:
 478         free(p);
 479 }
 480
 481 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 482                           unsigned int min_complete, unsigned int flags)
 483 {
 484         if (register_ring)
 485                 flags |= IORING_ENTER_REGISTERED_RING;
 486 #ifdef FIO_ARCH_HAS_SYSCALL
 487         return __do_syscall6(__NR_io_uring_enter, s->enter_ring_fd, to_submit,
 488                                 min_complete, flags, NULL, 0);
 489 #else
 490         return syscall(__NR_io_uring_enter, s->enter_ring_fd, to_submit,
 491                         min_complete, flags, NULL, 0);
 492 #endif
 493 }
 494
 495 static unsigned long long get_offset(struct submitter *s, struct file *f)
 496 {
 497         unsigned long long offset;
 498         long r;
 499
 500         if (random_io) {
 501                 unsigned long long block;
 502
 503                 r = __rand64(&s->rand_state);
 504                 block = r % f->max_blocks;
 505                 offset = block * (unsigned long long) bs;
 506         } else {
 507                 offset = f->cur_off;
 508                 f->cur_off += bs;
 509                 if (f->cur_off + bs > f->max_size)
 510                         f->cur_off = 0;
 511         }
 512
 513         return offset;
 514 }
 515
 516 static struct file *get_next_file(struct submitter *s)
 517 {
 518         struct file *f;
 519
 520         if (s->nr_files == 1) {
 521                 f = &s->files[0];
 522         } else {
 523                 f = &s->files[s->cur_file];
 524                 if (f->pending_ios >= s->per_file_depth) {
 525                         s->cur_file++;
 526                         if (s->cur_file == s->nr_files)
 527                                 s->cur_file = 0;
 528                         f = &s->files[s->cur_file];
 529                 }
 530         }
 531
 532         f->pending_ios++;
 533         return f;
 534 }
 535
 536 static void init_io(struct submitter *s, unsigned index)
 537 {
 538         struct io_uring_sqe *sqe = &s->sqes[index];
 539         struct file *f;
 540
 541         if (do_nop) {
 542                 sqe->opcode = IORING_OP_NOP;
 543                 return;
 544         }
 545
 546         f = get_next_file(s);
 547
 548         if (register_files) {
 549                 sqe->flags = IOSQE_FIXED_FILE;
 550                 sqe->fd = f->fixed_fd;
 551         } else {
 552                 sqe->flags = 0;
 553                 sqe->fd = f->real_fd;
 554         }
 555         if (fixedbufs) {
 556                 sqe->opcode = IORING_OP_READ_FIXED;
 557                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 558                 sqe->len = bs;
 559                 sqe->buf_index = index;
 560         } else if (!vectored) {
 561                 sqe->opcode = IORING_OP_READ;
 562                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 563                 sqe->len = bs;
 564                 sqe->buf_index = 0;
 565         } else {
 566                 sqe->opcode = IORING_OP_READV;
 567                 sqe->addr = (unsigned long) &s->iovecs[index];
 568                 sqe->len = 1;
 569                 sqe->buf_index = 0;
 570         }
 571         sqe->ioprio = 0;
 572         sqe->off = get_offset(s, f);
 573         sqe->user_data = (unsigned long) f->fileno;
 574         if (stats && stats_running)
 575                 sqe->user_data |= ((uint64_t)s->clock_index << 32);
 576 }
 577
 578 static void init_io_pt(struct submitter *s, unsigned index)
 579 {
 580         struct io_uring_sqe *sqe = &s->sqes[index << 1];
 581         unsigned long offset;
 582         struct file *f;
 583         struct nvme_uring_cmd *cmd;
 584         unsigned long long slba;
 585         unsigned long long nlb;
 586
 587         f = get_next_file(s);
 588
 589         offset = get_offset(s, f);
 590
 591         if (register_files) {
 592                 sqe->fd = f->fixed_fd;
 593                 sqe->flags = IOSQE_FIXED_FILE;
 594         } else {
 595                 sqe->fd = f->real_fd;
 596                 sqe->flags = 0;
 597         }
 598         sqe->opcode = IORING_OP_URING_CMD;
 599         sqe->user_data = (unsigned long) f->fileno;
 600         if (stats)
 601                 sqe->user_data |= ((__u64) s->clock_index << 32ULL);
 602         sqe->cmd_op = NVME_URING_CMD_IO;
 603         slba = offset >> f->lba_shift;
 604         nlb = (bs >> f->lba_shift) - 1;
 605         cmd = (struct nvme_uring_cmd *)&sqe->cmd;
 606         /* cdw10 and cdw11 represent starting slba*/
 607         cmd->cdw10 = slba & 0xffffffff;
 608         cmd->cdw11 = slba >> 32;
 609         /* cdw12 represent number of lba to be read*/
 610         cmd->cdw12 = nlb;
 611         cmd->addr = (unsigned long) s->iovecs[index].iov_base;
 612         cmd->data_len = bs;
 613         if (fixedbufs) {
 614                 sqe->uring_cmd_flags = IORING_URING_CMD_FIXED;
 615                 sqe->buf_index = index;
 616         }
 617         cmd->nsid = f->nsid;
 618         cmd->opcode = 2;
 619 }
 620
 621 static int prep_more_ios_uring(struct submitter *s, int max_ios)
 622 {
 623         struct io_sq_ring *ring = &s->sq_ring;
 624         unsigned head, index, tail, next_tail, prepped = 0;
 625
 626         if (sq_thread_poll)
 627                 head = atomic_load_acquire(ring->head);
 628         else
 629                 head = *ring->head;
 630
 631         next_tail = tail = *ring->tail;
 632         do {
 633                 next_tail++;
 634                 if (next_tail == head)
 635                         break;
 636
 637                 index = tail & sq_ring_mask;
 638                 if (pt)
 639                         init_io_pt(s, index);
 640                 else
 641                         init_io(s, index);
 642                 prepped++;
 643                 tail = next_tail;
 644         } while (prepped < max_ios);
 645
 646         if (prepped)
 647                 atomic_store_release(ring->tail, tail);
 648         return prepped;
 649 }
 650
 651 static int get_file_size(struct file *f)
 652 {
 653         struct stat st;
 654
 655         if (fstat(f->real_fd, &st) < 0)
 656                 return -1;
 657         if (pt) {
 658                 __u64 nlba;
 659                 __u32 lbs;
 660                 int ret;
 661
 662                 if (!S_ISCHR(st.st_mode)) {
 663                         fprintf(stderr, "passthrough works with only nvme-ns "
 664                                         "generic devices (/dev/ngXnY)\n");
 665                         return -1;
 666                 }
 667                 ret = nvme_get_info(f->real_fd, &f->nsid, &lbs, &nlba);
 668                 if (ret)
 669                         return -1;
 670                 if ((bs % lbs) != 0) {
 671                         printf("error: bs:%d should be a multiple logical_block_size:%d\n",
 672                                         bs, lbs);
 673                         return -1;
 674                 }
 675                 f->max_blocks = nlba;
 676                 f->max_size = nlba;
 677                 f->lba_shift = ilog2(lbs);
 678                 return 0;
 679         } else if (S_ISBLK(st.st_mode)) {
 680                 unsigned long long bytes;
 681
 682                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 683                         return -1;
 684
 685                 f->max_blocks = bytes / bs;
 686                 f->max_size = bytes;
 687                 return 0;
 688         } else if (S_ISREG(st.st_mode)) {
 689                 f->max_blocks = st.st_size / bs;
 690                 f->max_size = st.st_size;
 691                 return 0;
 692         }
 693
 694         return -1;
 695 }
 696
 697 static int reap_events_uring(struct submitter *s)
 698 {
 699         struct io_cq_ring *ring = &s->cq_ring;
 700         struct io_uring_cqe *cqe;
 701         unsigned head, reaped = 0;
 702         int last_idx = -1, stat_nr = 0;
 703
 704         head = *ring->head;
 705         do {
 706                 struct file *f;
 707
 708                 if (head == atomic_load_acquire(ring->tail))
 709                         break;
 710                 cqe = &ring->cqes[head & cq_ring_mask];
 711                 if (!do_nop) {
 712                         int fileno = cqe->user_data & 0xffffffff;
 713
 714                         f = &s->files[fileno];
 715                         f->pending_ios--;
 716                         if (cqe->res != bs) {
 717                                 if (cqe->res == -ENODATA || cqe->res == -EIO) {
 718                                         s->io_errors++;
 719                                 } else {
 720                                         printf("io: unexpected ret=%d\n", cqe->res);
 721                                         if (polled && cqe->res == -EOPNOTSUPP)
 722                                                 printf("Your filesystem/driver/kernel doesn't support polled IO\n");
 723                                         return -1;
 724                                 }
 725                         }
 726                 }
 727                 if (stats) {
 728                         int clock_index = cqe->user_data >> 32;
 729
 730                         if (last_idx != clock_index) {
 731                                 if (last_idx != -1) {
 732                                         add_stat(s, last_idx, stat_nr);
 733                                         stat_nr = 0;
 734                                 }
 735                                 last_idx = clock_index;
 736                         }
 737                         stat_nr++;
 738                 }
 739                 reaped++;
 740                 head++;
 741         } while (1);
 742
 743         if (stat_nr)
 744                 add_stat(s, last_idx, stat_nr);
 745
 746         if (reaped) {
 747                 s->inflight -= reaped;
 748                 atomic_store_release(ring->head, head);
 749         }
 750         return reaped;
 751 }
 752
 753 static int reap_events_uring_pt(struct submitter *s)
 754 {
 755         struct io_cq_ring *ring = &s->cq_ring;
 756         struct io_uring_cqe *cqe;
 757         unsigned head, reaped = 0;
 758         int last_idx = -1, stat_nr = 0;
 759         unsigned index;
 760         int fileno;
 761
 762         head = *ring->head;
 763         do {
 764                 struct file *f;
 765
 766                 if (head == atomic_load_acquire(ring->tail))
 767                         break;
 768                 index = head & cq_ring_mask;
 769                 cqe = &ring->cqes[index << 1];
 770                 fileno = cqe->user_data & 0xffffffff;
 771                 f = &s->files[fileno];
 772                 f->pending_ios--;
 773
 774                 if (cqe->res != 0) {
 775                         printf("io: unexpected ret=%d\n", cqe->res);
 776                         if (polled && cqe->res == -EINVAL)
 777                                 printf("passthrough doesn't support polled IO\n");
 778                         return -1;
 779                 }
 780                 if (stats) {
 781                         int clock_index = cqe->user_data >> 32;
 782
 783                         if (last_idx != clock_index) {
 784                                 if (last_idx != -1) {
 785                                         add_stat(s, last_idx, stat_nr);
 786                                         stat_nr = 0;
 787                                 }
 788                                 last_idx = clock_index;
 789                         }
 790                         stat_nr++;
 791                 }
 792                 reaped++;
 793                 head++;
 794         } while (1);
 795
 796         if (stat_nr)
 797                 add_stat(s, last_idx, stat_nr);
 798
 799         if (reaped) {
 800                 s->inflight -= reaped;
 801                 atomic_store_release(ring->head, head);
 802         }
 803         return reaped;
 804 }
 805
 806 static void set_affinity(struct submitter *s)
 807 {
 808 #ifdef CONFIG_LIBNUMA
 809         struct bitmask *mask;
 810
 811         if (s->numa_node == -1)
 812                 return;
 813
 814         numa_set_preferred(s->numa_node);
 815
 816         mask = numa_allocate_cpumask();
 817         numa_node_to_cpus(s->numa_node, mask);
 818         numa_sched_setaffinity(s->tid, mask);
 819 #endif
 820 }
 821
 822 static int detect_node(struct submitter *s, const char *name)
 823 {
 824 #ifdef CONFIG_LIBNUMA
 825         const char *base = basename(name);
 826         char str[128];
 827         int ret, fd, node;
 828
 829         if (pt)
 830                 sprintf(str, "/sys/class/nvme-generic/%s/device/numa_node", base);
 831         else
 832                 sprintf(str, "/sys/block/%s/device/numa_node", base);
 833         fd = open(str, O_RDONLY);
 834         if (fd < 0)
 835                 return -1;
 836
 837         ret = read(fd, str, sizeof(str));
 838         if (ret < 0) {
 839                 close(fd);
 840                 return -1;
 841         }
 842         node = atoi(str);
 843         s->numa_node = node;
 844         close(fd);
 845 #else
 846         s->numa_node = -1;
 847 #endif
 848         return 0;
 849 }
 850
 851 static int setup_aio(struct submitter *s)
 852 {
 853 #ifdef CONFIG_LIBAIO
 854         if (polled) {
 855                 fprintf(stderr, "aio does not support polled IO\n");
 856                 polled = 0;
 857         }
 858         if (sq_thread_poll) {
 859                 fprintf(stderr, "aio does not support SQPOLL IO\n");
 860                 sq_thread_poll = 0;
 861         }
 862         if (do_nop) {
 863                 fprintf(stderr, "aio does not support polled IO\n");
 864                 do_nop = 0;
 865         }
 866         if (fixedbufs || register_files) {
 867                 fprintf(stderr, "aio does not support registered files or buffers\n");
 868                 fixedbufs = register_files = 0;
 869         }
 870
 871         s->per_file_depth = (depth + s->nr_files - 1) / s->nr_files;
 872         return io_queue_init(roundup_pow2(depth), &s->aio_ctx);
 873 #else
 874         fprintf(stderr, "Legacy AIO not available on this system/build\n");
 875         errno = EINVAL;
 876         return -1;
 877 #endif
 878 }
 879
 880 static int setup_ring(struct submitter *s)
 881 {
 882         struct io_sq_ring *sring = &s->sq_ring;
 883         struct io_cq_ring *cring = &s->cq_ring;
 884         struct io_uring_params p;
 885         int ret, fd, i;
 886         void *ptr;
 887         size_t len;
 888
 889         memset(&p, 0, sizeof(p));
 890
 891         if (polled && !do_nop)
 892                 p.flags |= IORING_SETUP_IOPOLL;
 893         if (sq_thread_poll) {
 894                 p.flags |= IORING_SETUP_SQPOLL;
 895                 if (sq_thread_cpu != -1) {
 896                         p.flags |= IORING_SETUP_SQ_AFF;
 897                         p.sq_thread_cpu = sq_thread_cpu;
 898                 }
 899         }
 900         if (pt) {
 901                 p.flags |= IORING_SETUP_SQE128;
 902                 p.flags |= IORING_SETUP_CQE32;
 903         }
 904
 905         fd = io_uring_setup(depth, &p);
 906         if (fd < 0) {
 907                 perror("io_uring_setup");
 908                 return 1;
 909         }
 910         s->ring_fd = s->enter_ring_fd = fd;
 911
 912         io_uring_probe(fd);
 913
 914         if (fixedbufs) {
 915                 struct rlimit rlim;
 916
 917                 rlim.rlim_cur = RLIM_INFINITY;
 918                 rlim.rlim_max = RLIM_INFINITY;
 919                 /* ignore potential error, not needed on newer kernels */
 920                 setrlimit(RLIMIT_MEMLOCK, &rlim);
 921
 922                 ret = io_uring_register_buffers(s);
 923                 if (ret < 0) {
 924                         perror("io_uring_register_buffers");
 925                         return 1;
 926                 }
 927         }
 928
 929         if (register_files) {
 930                 ret = io_uring_register_files(s);
 931                 if (ret < 0) {
 932                         perror("io_uring_register_files");
 933                         return 1;
 934                 }
 935         }
 936
 937         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 938                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 939                         IORING_OFF_SQ_RING);
 940         sring->head = ptr + p.sq_off.head;
 941         sring->tail = ptr + p.sq_off.tail;
 942         sring->ring_mask = ptr + p.sq_off.ring_mask;
 943         sring->ring_entries = ptr + p.sq_off.ring_entries;
 944         sring->flags = ptr + p.sq_off.flags;
 945         sring->array = ptr + p.sq_off.array;
 946         sq_ring_mask = *sring->ring_mask;
 947
 948         if (p.flags & IORING_SETUP_SQE128)
 949                 len = 2 * p.sq_entries * sizeof(struct io_uring_sqe);
 950         else
 951                 len = p.sq_entries * sizeof(struct io_uring_sqe);
 952         s->sqes = mmap(0, len,
 953                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 954                         IORING_OFF_SQES);
 955
 956         if (p.flags & IORING_SETUP_CQE32) {
 957                 len = p.cq_off.cqes +
 958                         2 * p.cq_entries * sizeof(struct io_uring_cqe);
 959         } else {
 960                 len = p.cq_off.cqes +
 961                         p.cq_entries * sizeof(struct io_uring_cqe);
 962         }
 963         ptr = mmap(0, len,
 964                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 965                         IORING_OFF_CQ_RING);
 966         cring->head = ptr + p.cq_off.head;
 967         cring->tail = ptr + p.cq_off.tail;
 968         cring->ring_mask = ptr + p.cq_off.ring_mask;
 969         cring->ring_entries = ptr + p.cq_off.ring_entries;
 970         cring->cqes = ptr + p.cq_off.cqes;
 971         cq_ring_mask = *cring->ring_mask;
 972
 973         for (i = 0; i < p.sq_entries; i++)
 974                 sring->array[i] = i;
 975
 976         s->per_file_depth = (depth + s->nr_files - 1) / s->nr_files;
 977         return 0;
 978 }
 979
 980 static void *allocate_mem(struct submitter *s, int size)
 981 {
 982         void *buf;
 983
 984 #ifdef CONFIG_LIBNUMA
 985         if (s->numa_node != -1)
 986                 return numa_alloc_onnode(size, s->numa_node);
 987 #endif
 988
 989         if (posix_memalign(&buf, t_io_uring_page_size, bs)) {
 990                 printf("failed alloc\n");
 991                 return NULL;
 992         }
 993
 994         return buf;
 995 }
 996
 997 static int submitter_init(struct submitter *s)
 998 {
 999         int i, nr_batch, err;
1000         static int init_printed;
1001         char buf[80];
1002         s->tid = gettid();
1003         printf("submitter=%d, tid=%d, file=%s, nfiles=%d, node=%d\n", s->index, s->tid,
1004                                                         s->filename, s->nr_files, s->numa_node);
1005
1006         set_affinity(s);
1007
1008         __init_rand64(&s->rand_state, s->tid);
1009         srand48(s->tid);
1010
1011         for (i = 0; i < MAX_FDS; i++)
1012                 s->files[i].fileno = i;
1013
1014         for (i = 0; i < roundup_pow2(depth); i++) {
1015                 void *buf;
1016
1017                 buf = allocate_mem(s, bs);
1018                 if (!buf)
1019                         return -1;
1020                 s->iovecs[i].iov_base = buf;
1021                 s->iovecs[i].iov_len = bs;
1022         }
1023
1024         if (use_sync) {
1025                 sprintf(buf, "Engine=preadv2\n");
1026                 err = 0;
1027         } else if (!aio) {
1028                 err = setup_ring(s);
1029                 if (!err)
1030                         sprintf(buf, "Engine=io_uring, sq_ring=%d, cq_ring=%d\n", *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
1031         } else {
1032                 sprintf(buf, "Engine=aio\n");
1033                 err = setup_aio(s);
1034         }
1035         if (err) {
1036                 printf("queue setup failed: %s, %d\n", strerror(errno), err);
1037                 return -1;
1038         }
1039
1040         if (!init_printed) {
1041                 printf("polled=%d, fixedbufs=%d, register_files=%d, buffered=%d, QD=%d\n", polled, fixedbufs, register_files, buffered, depth);
1042                 printf("%s", buf);
1043                 init_printed = 1;
1044         }
1045
1046         if (stats) {
1047                 nr_batch = roundup_pow2(depth / batch_submit);
1048                 if (nr_batch < 2)
1049                         nr_batch = 2;
1050                 s->clock_batch = calloc(nr_batch, sizeof(unsigned long));
1051                 s->clock_index = 1;
1052
1053                 s->plat = calloc(PLAT_NR, sizeof(unsigned long));
1054         } else {
1055                 s->clock_batch = NULL;
1056                 s->plat = NULL;
1057                 nr_batch = 0;
1058         }
1059         /* perform the expensive command initialization part for passthrough here
1060          * rather than in the fast path
1061          */
1062         if (pt) {
1063                 for (i = 0; i < roundup_pow2(depth); i++) {
1064                         struct io_uring_sqe *sqe = &s->sqes[i << 1];
1065
1066                         memset(&sqe->cmd, 0, sizeof(struct nvme_uring_cmd));
1067                 }
1068         }
1069         return nr_batch;
1070 }
1071
1072 #ifdef CONFIG_LIBAIO
1073 static int prep_more_ios_aio(struct submitter *s, int max_ios, struct iocb *iocbs)
1074 {
1075         uint64_t data;
1076         struct file *f;
1077         unsigned index;
1078
1079         index = 0;
1080         while (index < max_ios) {
1081                 struct iocb *iocb = &iocbs[index];
1082
1083                 f = get_next_file(s);
1084
1085                 io_prep_pread(iocb, f->real_fd, s->iovecs[index].iov_base,
1086                                 s->iovecs[index].iov_len, get_offset(s, f));
1087
1088                 data = f->fileno;
1089                 if (stats && stats_running)
1090                         data |= (((uint64_t) s->clock_index) << 32);
1091                 iocb->data = (void *) (uintptr_t) data;
1092                 index++;
1093         }
1094         return index;
1095 }
1096
1097 static int reap_events_aio(struct submitter *s, struct io_event *events, int evs)
1098 {
1099         int last_idx = -1, stat_nr = 0;
1100         int reaped = 0;
1101
1102         while (evs) {
1103                 uint64_t data = (uintptr_t) events[reaped].data;
1104                 struct file *f = &s->files[data & 0xffffffff];
1105
1106                 f->pending_ios--;
1107                 if (events[reaped].res != bs) {
1108                         if (events[reaped].res == -ENODATA ||
1109                             events[reaped].res == -EIO) {
1110                                 s->io_errors++;
1111                         } else {
1112                                 printf("io: unexpected ret=%ld\n", events[reaped].res);
1113                                 return -1;
1114                         }
1115                 } else if (stats) {
1116                         int clock_index = data >> 32;
1117
1118                         if (last_idx != clock_index) {
1119                                 if (last_idx != -1) {
1120                                         add_stat(s, last_idx, stat_nr);
1121                                         stat_nr = 0;
1122                                 }
1123                                 last_idx = clock_index;
1124                         }
1125                         stat_nr++;
1126                 }
1127                 reaped++;
1128                 evs--;
1129         }
1130
1131         if (stat_nr)
1132                 add_stat(s, last_idx, stat_nr);
1133
1134         s->inflight -= reaped;
1135         s->done += reaped;
1136         return reaped;
1137 }
1138
1139 static void *submitter_aio_fn(void *data)
1140 {
1141         struct submitter *s = data;
1142         int i, ret, prepped;
1143         struct iocb **iocbsptr;
1144         struct iocb *iocbs;
1145         struct io_event *events;
1146 #ifdef ARCH_HAVE_CPU_CLOCK
1147         int nr_batch;
1148 #endif
1149
1150         ret = submitter_init(s);
1151         if (ret < 0)
1152                 goto done;
1153
1154 #ifdef ARCH_HAVE_CPU_CLOCK
1155         nr_batch = ret;
1156 #endif
1157
1158         iocbsptr = calloc(depth, sizeof(struct iocb *));
1159         iocbs = calloc(depth, sizeof(struct iocb));
1160         events = calloc(depth, sizeof(struct io_event));
1161
1162         for (i = 0; i < depth; i++)
1163                 iocbsptr[i] = &iocbs[i];
1164
1165         prepped = 0;
1166         do {
1167                 int to_wait, to_submit, to_prep;
1168
1169                 if (!prepped && s->inflight < depth) {
1170                         to_prep = min(depth - s->inflight, batch_submit);
1171                         prepped = prep_more_ios_aio(s, to_prep, iocbs);
1172 #ifdef ARCH_HAVE_CPU_CLOCK
1173                         if (prepped && stats) {
1174                                 s->clock_batch[s->clock_index] = get_cpu_clock();
1175                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
1176                         }
1177 #endif
1178                 }
1179                 s->inflight += prepped;
1180                 to_submit = prepped;
1181
1182                 if (to_submit && (s->inflight + to_submit <= depth))
1183                         to_wait = 0;
1184                 else
1185                         to_wait = min(s->inflight + to_submit, batch_complete);
1186
1187                 ret = io_submit(s->aio_ctx, to_submit, iocbsptr);
1188                 s->calls++;
1189                 if (ret < 0) {
1190                         perror("io_submit");
1191                         break;
1192                 } else if (ret != to_submit) {
1193                         printf("submitted %d, wanted %d\n", ret, to_submit);
1194                         break;
1195                 }
1196                 prepped = 0;
1197
1198                 while (to_wait) {
1199                         int r;
1200
1201                         s->calls++;
1202                         r = io_getevents(s->aio_ctx, to_wait, to_wait, events, NULL);
1203                         if (r < 0) {
1204                                 perror("io_getevents");
1205                                 break;
1206                         } else if (r != to_wait) {
1207                                 printf("r=%d, wait=%d\n", r, to_wait);
1208                                 break;
1209                         }
1210                         r = reap_events_aio(s, events, r);
1211                         s->reaps += r;
1212                         to_wait -= r;
1213                 }
1214         } while (!s->finish);
1215
1216         free(iocbsptr);
1217         free(iocbs);
1218         free(events);
1219 done:
1220         finish = 1;
1221         return NULL;
1222 }
1223 #endif
1224
1225 static void io_uring_unregister_ring(struct submitter *s)
1226 {
1227         struct io_uring_rsrc_update up = {
1228                 .offset = s->enter_ring_fd,
1229         };
1230
1231         syscall(__NR_io_uring_register, s->ring_fd, IORING_UNREGISTER_RING_FDS,
1232                 &up, 1);
1233 }
1234
1235 static int io_uring_register_ring(struct submitter *s)
1236 {
1237         struct io_uring_rsrc_update up = {
1238                 .data   = s->ring_fd,
1239                 .offset = -1U,
1240         };
1241         int ret;
1242
1243         ret = syscall(__NR_io_uring_register, s->ring_fd,
1244                         IORING_REGISTER_RING_FDS, &up, 1);
1245         if (ret == 1) {
1246                 s->enter_ring_fd = up.offset;
1247                 return 0;
1248         }
1249         register_ring = 0;
1250         return -1;
1251 }
1252
1253 static void *submitter_uring_fn(void *data)
1254 {
1255         struct submitter *s = data;
1256         struct io_sq_ring *ring = &s->sq_ring;
1257         int ret, prepped;
1258 #ifdef ARCH_HAVE_CPU_CLOCK
1259         int nr_batch;
1260 #endif
1261
1262         ret = submitter_init(s);
1263         if (ret < 0)
1264                 goto done;
1265
1266 #ifdef ARCH_HAVE_CPU_CLOCK
1267         nr_batch = ret;
1268 #endif
1269
1270         if (register_ring)
1271                 io_uring_register_ring(s);
1272
1273         prepped = 0;
1274         do {
1275                 int to_wait, to_submit, this_reap, to_prep;
1276                 unsigned ring_flags = 0;
1277
1278                 if (!prepped && s->inflight < depth) {
1279                         to_prep = min(depth - s->inflight, batch_submit);
1280                         prepped = prep_more_ios_uring(s, to_prep);
1281 #ifdef ARCH_HAVE_CPU_CLOCK
1282                         if (prepped && stats) {
1283                                 s->clock_batch[s->clock_index] = get_cpu_clock();
1284                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
1285                         }
1286 #endif
1287                 }
1288                 s->inflight += prepped;
1289 submit_more:
1290                 to_submit = prepped;
1291 submit:
1292                 if (to_submit && (s->inflight + to_submit <= depth))
1293                         to_wait = 0;
1294                 else
1295                         to_wait = min(s->inflight + to_submit, batch_complete);
1296
1297                 /*
1298                  * Only need to call io_uring_enter if we're not using SQ thread
1299                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
1300                  */
1301                 if (sq_thread_poll)
1302                         ring_flags = atomic_load_acquire(ring->flags);
1303                 if (!sq_thread_poll || ring_flags & IORING_SQ_NEED_WAKEUP) {
1304                         unsigned flags = 0;
1305
1306                         if (to_wait)
1307                                 flags = IORING_ENTER_GETEVENTS;
1308                         if (ring_flags & IORING_SQ_NEED_WAKEUP)
1309                                 flags |= IORING_ENTER_SQ_WAKEUP;
1310                         ret = io_uring_enter(s, to_submit, to_wait, flags);
1311                         s->calls++;
1312                 } else {
1313                         /* for SQPOLL, we submitted it all effectively */
1314                         ret = to_submit;
1315                 }
1316
1317                 /*
1318                  * For non SQ thread poll, we already got the events we needed
1319                  * through the io_uring_enter() above. For SQ thread poll, we
1320                  * need to loop here until we find enough events.
1321                  */
1322                 this_reap = 0;
1323                 do {
1324                         int r;
1325
1326                         if (pt)
1327                                 r = reap_events_uring_pt(s);
1328                         else
1329                                 r = reap_events_uring(s);
1330                         if (r == -1) {
1331                                 s->finish = 1;
1332                                 break;
1333                         } else if (r > 0)
1334                                 this_reap += r;
1335                 } while (sq_thread_poll && this_reap < to_wait);
1336                 s->reaps += this_reap;
1337
1338                 if (ret >= 0) {
1339                         if (!ret) {
1340                                 to_submit = 0;
1341                                 if (s->inflight)
1342                                         goto submit;
1343                                 continue;
1344                         } else if (ret < to_submit) {
1345                                 int diff = to_submit - ret;
1346
1347                                 s->done += ret;
1348                                 prepped -= diff;
1349                                 goto submit_more;
1350                         }
1351                         s->done += ret;
1352                         prepped = 0;
1353                         continue;
1354                 } else if (ret < 0) {
1355                         if (errno == EAGAIN) {
1356                                 if (s->finish)
1357                                         break;
1358                                 if (this_reap)
1359                                         goto submit;
1360                                 to_submit = 0;
1361                                 goto submit;
1362                         }
1363                         printf("io_submit: %s\n", strerror(errno));
1364                         break;
1365                 }
1366         } while (!s->finish);
1367
1368         if (register_ring)
1369                 io_uring_unregister_ring(s);
1370
1371 done:
1372         finish = 1;
1373         return NULL;
1374 }
1375
1376 #ifdef CONFIG_PWRITEV2
1377 static void *submitter_sync_fn(void *data)
1378 {
1379         struct submitter *s = data;
1380         int ret;
1381
1382         if (submitter_init(s) < 0)
1383                 goto done;
1384
1385         do {
1386                 uint64_t offset;
1387                 struct file *f;
1388
1389                 f = get_next_file(s);
1390
1391 #ifdef ARCH_HAVE_CPU_CLOCK
1392                 if (stats)
1393                         s->clock_batch[s->clock_index] = get_cpu_clock();
1394 #endif
1395
1396                 s->inflight++;
1397                 s->calls++;
1398
1399                 offset = get_offset(s, f);
1400                 if (polled)
1401                         ret = preadv2(f->real_fd, &s->iovecs[0], 1, offset, RWF_HIPRI);
1402                 else
1403                         ret = preadv2(f->real_fd, &s->iovecs[0], 1, offset, 0);
1404
1405                 if (ret < 0) {
1406                         perror("preadv2");
1407                         break;
1408                 } else if (ret != bs) {
1409                         break;
1410                 }
1411
1412                 s->done++;
1413                 s->inflight--;
1414                 f->pending_ios--;
1415                 if (stats)
1416                         add_stat(s, s->clock_index, 1);
1417         } while (!s->finish);
1418
1419 done:
1420         finish = 1;
1421         return NULL;
1422 }
1423 #else
1424 static void *submitter_sync_fn(void *data)
1425 {
1426         finish = 1;
1427         return NULL;
1428 }
1429 #endif
1430
1431 static struct submitter *get_submitter(int offset)
1432 {
1433         void *ret;
1434
1435         ret = submitter;
1436         if (offset)
1437                 ret += offset * (sizeof(*submitter) + depth * sizeof(struct iovec));
1438         return ret;
1439 }
1440
1441 static void do_finish(const char *reason)
1442 {
1443         int j;
1444
1445         printf("Exiting on %s\n", reason);
1446         for (j = 0; j < nthreads; j++) {
1447                 struct submitter *s = get_submitter(j);
1448                 s->finish = 1;
1449         }
1450         if (max_iops > 1000000) {
1451                 double miops = (double) max_iops / 1000000.0;
1452                 printf("Maximum IOPS=%.2fM\n", miops);
1453         } else if (max_iops > 100000) {
1454                 double kiops = (double) max_iops / 1000.0;
1455                 printf("Maximum IOPS=%.2fK\n", kiops);
1456         } else {
1457                 printf("Maximum IOPS=%lu\n", max_iops);
1458         }
1459         finish = 1;
1460 }
1461
1462 static void sig_int(int sig)
1463 {
1464         do_finish("signal");
1465 }
1466
1467 static void arm_sig_int(void)
1468 {
1469         struct sigaction act;
1470
1471         memset(&act, 0, sizeof(act));
1472         act.sa_handler = sig_int;
1473         act.sa_flags = SA_RESTART;
1474         sigaction(SIGINT, &act, NULL);
1475
1476         /* Windows uses SIGBREAK as a quit signal from other applications */
1477 #ifdef WIN32
1478         sigaction(SIGBREAK, &act, NULL);
1479 #endif
1480 }
1481
1482 static void usage(char *argv, int status)
1483 {
1484         char runtime_str[16];
1485         snprintf(runtime_str, sizeof(runtime_str), "%d", runtime);
1486         printf("%s [options] -- [filenames]\n"
1487                 " -d <int>  : IO Depth, default %d\n"
1488                 " -s <int>  : Batch submit, default %d\n"
1489                 " -c <int>  : Batch complete, default %d\n"
1490                 " -b <int>  : Block size, default %d\n"
1491                 " -p <bool> : Polled IO, default %d\n"
1492                 " -B <bool> : Fixed buffers, default %d\n"
1493                 " -F <bool> : Register files, default %d\n"
1494                 " -n <int>  : Number of threads, default %d\n"
1495                 " -O <bool> : Use O_DIRECT, default %d\n"
1496                 " -N <bool> : Perform just no-op requests, default %d\n"
1497                 " -t <bool> : Track IO latencies, default %d\n"
1498                 " -T <int>  : TSC rate in HZ\n"
1499                 " -r <int>  : Runtime in seconds, default %s\n"
1500                 " -R <bool> : Use random IO, default %d\n"
1501                 " -a <bool> : Use legacy aio, default %d\n"
1502                 " -S <bool> : Use sync IO (preadv2), default %d\n"
1503                 " -X <bool> : Use registered ring %d\n"
1504                 " -P <bool> : Automatically place on device home node %d\n"
1505                 " -u <bool> : Use nvme-passthrough I/O, default %d\n",
1506                 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled,
1507                 fixedbufs, register_files, nthreads, !buffered, do_nop,
1508                 stats, runtime == 0 ? "unlimited" : runtime_str, random_io, aio,
1509                 use_sync, register_ring, numa_placement, pt);
1510         exit(status);
1511 }
1512
1513 static void read_tsc_rate(void)
1514 {
1515         char buffer[32];
1516         int fd, ret;
1517
1518         if (tsc_rate)
1519                 return;
1520
1521         fd = open(TSC_RATE_FILE, O_RDONLY);
1522         if (fd < 0)
1523                 return;
1524
1525         ret = read(fd, buffer, sizeof(buffer));
1526         if (ret < 0) {
1527                 close(fd);
1528                 return;
1529         }
1530
1531         tsc_rate = strtoul(buffer, NULL, 10);
1532         printf("Using TSC rate %luHz\n", tsc_rate);
1533         close(fd);
1534 }
1535
1536 static void write_tsc_rate(void)
1537 {
1538         char buffer[32];
1539         struct stat sb;
1540         int fd, ret;
1541
1542         if (!stat(TSC_RATE_FILE, &sb))
1543                 return;
1544
1545         fd = open(TSC_RATE_FILE, O_WRONLY | O_CREAT, 0644);
1546         if (fd < 0)
1547                 return;
1548
1549         memset(buffer, 0, sizeof(buffer));
1550         sprintf(buffer, "%lu", tsc_rate);
1551         ret = write(fd, buffer, strlen(buffer));
1552         if (ret < 0)
1553                 perror("write");
1554         close(fd);
1555 }
1556
1557 int main(int argc, char *argv[])
1558 {
1559         struct submitter *s;
1560         unsigned long done, calls, reap, io_errors;
1561         int i, j, flags, fd, opt, threads_per_f, threads_rem = 0, nfiles;
1562         struct file f;
1563         void *ret;
1564
1565         if (!do_nop && argc < 2)
1566                 usage(argv[0], 1);
1567
1568         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) {
1569                 switch (opt) {
1570                 case 'a':
1571                         aio = !!atoi(optarg);
1572                         break;
1573                 case 'd':
1574                         depth = atoi(optarg);
1575                         break;
1576                 case 's':
1577                         batch_submit = atoi(optarg);
1578                         if (!batch_submit)
1579                                 batch_submit = 1;
1580                         break;
1581                 case 'c':
1582                         batch_complete = atoi(optarg);
1583                         if (!batch_complete)
1584                                 batch_complete = 1;
1585                         break;
1586                 case 'b':
1587                         bs = atoi(optarg);
1588                         break;
1589                 case 'p':
1590                         polled = !!atoi(optarg);
1591                         break;
1592                 case 'B':
1593                         fixedbufs = !!atoi(optarg);
1594                         break;
1595                 case 'F':
1596                         register_files = !!atoi(optarg);
1597                         break;
1598                 case 'n':
1599                         nthreads = atoi(optarg);
1600                         if (!nthreads) {
1601                                 printf("Threads must be non-zero\n");
1602                                 usage(argv[0], 1);
1603                         }
1604                         break;
1605                 case 'N':
1606                         do_nop = !!atoi(optarg);
1607                         break;
1608                 case 'O':
1609                         buffered = !atoi(optarg);
1610                         break;
1611                 case 't':
1612 #ifndef ARCH_HAVE_CPU_CLOCK
1613                         fprintf(stderr, "Stats not supported on this CPU\n");
1614                         return 1;
1615 #endif
1616                         stats = !!atoi(optarg);
1617                         break;
1618                 case 'T':
1619 #ifndef ARCH_HAVE_CPU_CLOCK
1620                         fprintf(stderr, "Stats not supported on this CPU\n");
1621                         return 1;
1622 #endif
1623                         tsc_rate = strtoul(optarg, NULL, 10);
1624                         write_tsc_rate();
1625                         break;
1626                 case 'r':
1627                         runtime = atoi(optarg);
1628                         break;
1629                 case 'R':
1630                         random_io = !!atoi(optarg);
1631                         break;
1632                 case 'X':
1633                         register_ring = !!atoi(optarg);
1634                         break;
1635                 case 'S':
1636 #ifdef CONFIG_PWRITEV2
1637                         use_sync = !!atoi(optarg);
1638 #else
1639                         fprintf(stderr, "preadv2 not supported\n");
1640                         exit(1);
1641 #endif
1642                         break;
1643                 case 'P':
1644                         numa_placement = !!atoi(optarg);
1645                         break;
1646                 case 'u':
1647                         pt = !!atoi(optarg);
1648                         break;
1649                 case 'h':
1650                 case '?':
1651                 default:
1652                         usage(argv[0], 0);
1653                         break;
1654                 }
1655         }
1656
1657         if (stats)
1658                 read_tsc_rate();
1659
1660         if (batch_complete > depth)
1661                 batch_complete = depth;
1662         if (batch_submit > depth)
1663                 batch_submit = depth;
1664
1665         submitter = calloc(nthreads, sizeof(*submitter) +
1666                                 roundup_pow2(depth) * sizeof(struct iovec));
1667         for (j = 0; j < nthreads; j++) {
1668                 s = get_submitter(j);
1669                 s->numa_node = -1;
1670                 s->index = j;
1671                 s->done = s->calls = s->reaps = s->io_errors = 0;
1672         }
1673
1674         flags = O_RDONLY | O_NOATIME;
1675         if (!buffered)
1676                 flags |= O_DIRECT;
1677
1678         j = 0;
1679         i = optind;
1680         nfiles = argc - i;
1681         if (!do_nop) {
1682                 if (!nfiles) {
1683                         printf("No files specified\n");
1684                         usage(argv[0], 1);
1685                 }
1686                 threads_per_f = nthreads / nfiles;
1687                 /* make sure each thread gets assigned files */
1688                 if (threads_per_f == 0) {
1689                         threads_per_f = 1;
1690                 } else {
1691                         threads_rem = nthreads - threads_per_f * nfiles;
1692                 }
1693         }
1694         while (!do_nop && i < argc) {
1695                 int k, limit;
1696
1697                 memset(&f, 0, sizeof(f));
1698
1699                 fd = open(argv[i], flags);
1700                 if (fd < 0) {
1701                         perror("open");
1702                         return 1;
1703                 }
1704                 f.real_fd = fd;
1705                 if (get_file_size(&f)) {
1706                         printf("failed getting size of device/file\n");
1707                         return 1;
1708                 }
1709                 if (f.max_blocks <= 1) {
1710                         printf("Zero file/device size?\n");
1711                         return 1;
1712                 }
1713                 f.max_blocks--;
1714
1715                 limit = threads_per_f;
1716                 limit += threads_rem > 0 ? 1 : 0;
1717                 for (k = 0; k < limit; k++) {
1718                         s = get_submitter((j + k) % nthreads);
1719
1720                         if (s->nr_files == MAX_FDS) {
1721                                 printf("Max number of files (%d) reached\n", MAX_FDS);
1722                                 break;
1723                         }
1724
1725                         memcpy(&s->files[s->nr_files], &f, sizeof(f));
1726
1727                         if (numa_placement)
1728                                 detect_node(s, argv[i]);
1729
1730                         s->filename = argv[i];
1731                         s->nr_files++;
1732                 }
1733                 threads_rem--;
1734                 i++;
1735                 j += limit;
1736         }
1737
1738         arm_sig_int();
1739
1740         t_io_uring_page_size = sysconf(_SC_PAGESIZE);
1741         if (t_io_uring_page_size < 0)
1742                 t_io_uring_page_size = 4096;
1743
1744         for (j = 0; j < nthreads; j++) {
1745                 s = get_submitter(j);
1746                 if (use_sync)
1747                         pthread_create(&s->thread, NULL, submitter_sync_fn, s);
1748                 else if (!aio)
1749                         pthread_create(&s->thread, NULL, submitter_uring_fn, s);
1750 #ifdef CONFIG_LIBAIO
1751                 else
1752                         pthread_create(&s->thread, NULL, submitter_aio_fn, s);
1753 #endif
1754         }
1755
1756         reap = calls = done = io_errors = 0;
1757         do {
1758                 unsigned long this_done = 0;
1759                 unsigned long this_reap = 0;
1760                 unsigned long this_call = 0;
1761                 unsigned long this_io_errors = 0;
1762                 unsigned long rpc = 0, ipc = 0;
1763                 unsigned long iops, bw;
1764
1765                 sleep(1);
1766                 if (runtime && !--runtime)
1767                         do_finish("timeout");
1768
1769                 /* don't print partial run, if interrupted by signal */
1770                 if (finish)
1771                         break;
1772
1773                 /* one second in to the run, enable stats */
1774                 if (stats)
1775                         stats_running = 1;
1776
1777                 for (j = 0; j < nthreads; j++) {
1778                         s = get_submitter(j);
1779                         this_done += s->done;
1780                         this_call += s->calls;
1781                         this_reap += s->reaps;
1782                         this_io_errors += s->io_errors;
1783                 }
1784                 if (this_call - calls) {
1785                         rpc = (this_done - done) / (this_call - calls);
1786                         ipc = (this_reap - reap) / (this_call - calls);
1787                 } else
1788                         rpc = ipc = -1;
1789                 iops = this_done - done;
1790                 iops -= this_io_errors - io_errors;
1791                 if (bs > 1048576)
1792                         bw = iops * (bs / 1048576);
1793                 else
1794                         bw = iops / (1048576 / bs);
1795                 if (iops > 1000000) {
1796                         double miops = (double) iops / 1000000.0;
1797                         printf("IOPS=%.2fM, ", miops);
1798                 } else if (iops > 100000) {
1799                         double kiops = (double) iops / 1000.0;
1800                         printf("IOPS=%.2fK, ", kiops);
1801                 } else {
1802                         printf("IOPS=%lu, ", iops);
1803                 }
1804                 max_iops = max(max_iops, iops);
1805                 if (!do_nop) {
1806                         if (bw > 2000) {
1807                                 double bw_g = (double) bw / 1000.0;
1808
1809                                 printf("BW=%.2fGiB/s, ", bw_g);
1810                         } else {
1811                                 printf("BW=%luMiB/s, ", bw);
1812                         }
1813                 }
1814                 printf("IOS/call=%ld/%ld\n", rpc, ipc);
1815                 done = this_done;
1816                 calls = this_call;
1817                 reap = this_reap;
1818                 io_errors = this_io_errors;
1819         } while (!finish);
1820
1821         for (j = 0; j < nthreads; j++) {
1822                 s = get_submitter(j);
1823                 pthread_join(s->thread, &ret);
1824                 close(s->ring_fd);
1825
1826                 if (s->io_errors)
1827                         printf("%d: %lu IO errors\n", s->tid, s->io_errors);
1828
1829                 if (stats) {
1830                         unsigned long nr;
1831
1832                         printf("%d: Latency percentiles:\n", s->tid);
1833                         for (i = 0, nr = 0; i < PLAT_NR; i++)
1834                                 nr += s->plat[i];
1835                         show_clat_percentiles(s->plat, nr, 4);
1836                         free(s->clock_batch);
1837                         free(s->plat);
1838                 }
1839         }
1840
1841         free(submitter);
1842         return 0;
1843 }