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 #include <libgen.h>
  32
  33 #include "../arch/arch.h"
  34 #include "../os/os.h"
  35 #include "../lib/types.h"
  36 #include "../lib/roundup.h"
  37 #include "../lib/rand.h"
  38 #include "../minmax.h"
  39 #include "../os/linux/io_uring.h"
  40 #include "../engines/nvme.h"
  41
  42 struct io_sq_ring {
  43         unsigned *head;
  44         unsigned *tail;
  45         unsigned *ring_mask;
  46         unsigned *ring_entries;
  47         unsigned *flags;
  48         unsigned *array;
  49 };
  50
  51 struct io_cq_ring {
  52         unsigned *head;
  53         unsigned *tail;
  54         unsigned *ring_mask;
  55         unsigned *ring_entries;
  56         struct io_uring_cqe *cqes;
  57 };
  58
  59 #define DEPTH                   128
  60 #define BATCH_SUBMIT            32
  61 #define BATCH_COMPLETE          32
  62 #define BS                      4096
  63
  64 #define MAX_FDS                 16
  65
  66 static unsigned sq_ring_mask, cq_ring_mask;
  67
  68 struct file {
  69         unsigned long max_blocks;
  70         unsigned long max_size;
  71         unsigned long cur_off;
  72         unsigned pending_ios;
  73         unsigned int nsid;      /* nsid field required for nvme-passthrough */
  74         unsigned int lba_shift; /* lba_shift field required for nvme-passthrough */
  75         int real_fd;
  76         int fixed_fd;
  77         int fileno;
  78 };
  79
  80 #define PLAT_BITS               6
  81 #define PLAT_VAL                (1 << PLAT_BITS)
  82 #define PLAT_GROUP_NR           29
  83 #define PLAT_NR                 (PLAT_GROUP_NR * PLAT_VAL)
  84
  85 struct submitter {
  86         pthread_t thread;
  87         int ring_fd;
  88         int enter_ring_fd;
  89         int index;
  90         struct io_sq_ring sq_ring;
  91         struct io_uring_sqe *sqes;
  92         struct io_cq_ring cq_ring;
  93         int inflight;
  94         int tid;
  95         unsigned long reaps;
  96         unsigned long done;
  97         unsigned long calls;
  98         unsigned long io_errors;
  99         volatile int finish;
 100
 101         __s32 *fds;
 102
 103         struct taus258_state rand_state;
 104
 105         unsigned long *clock_batch;
 106         int clock_index;
 107         unsigned long *plat;
 108
 109 #ifdef CONFIG_LIBAIO
 110         io_context_t aio_ctx;
 111 #endif
 112
 113         int numa_node;
 114         int per_file_depth;
 115         const char *filename;
 116
 117         struct file files[MAX_FDS];
 118         unsigned nr_files;
 119         unsigned cur_file;
 120         struct iovec iovecs[];
 121 };
 122
 123 static struct submitter *submitter;
 124 static volatile int finish;
 125 static int stats_running;
 126 static unsigned long max_iops;
 127 static long t_io_uring_page_size;
 128
 129 static int depth = DEPTH;
 130 static int batch_submit = BATCH_SUBMIT;
 131 static int batch_complete = BATCH_COMPLETE;
 132 static int bs = BS;
 133 static int polled = 1;          /* use IO polling */
 134 static int fixedbufs = 1;       /* use fixed user buffers */
 135 static int register_files = 1;  /* use fixed files */
 136 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
 137 static int sq_thread_poll = 0;  /* use kernel submission/poller thread */
 138 static int sq_thread_cpu = -1;  /* pin above thread to this CPU */
 139 static int do_nop = 0;          /* no-op SQ ring commands */
 140 static int nthreads = 1;
 141 static int stats = 0;           /* generate IO stats */
 142 static int aio = 0;             /* use libaio */
 143 static int runtime = 0;         /* runtime */
 144 static int random_io = 1;       /* random or sequential IO */
 145 static int register_ring = 1;   /* register ring */
 146 static int use_sync = 0;        /* use preadv2 */
 147 static int numa_placement = 0;  /* set to node of device */
 148 static int pt = 0;              /* passthrough I/O or not */
 149
 150 static unsigned long tsc_rate;
 151
 152 #define TSC_RATE_FILE   "tsc-rate"
 153
 154 static int vectored = 1;
 155
 156 static float plist[] = { 1.0, 5.0, 10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0,
 157                         80.0, 90.0, 95.0, 99.0, 99.5, 99.9, 99.95, 99.99 };
 158 static int plist_len = 17;
 159
 160 static int nvme_identify(int fd, __u32 nsid, enum nvme_identify_cns cns,
 161                          enum nvme_csi csi, void *data)
 162 {
 163         struct nvme_passthru_cmd cmd = {
 164                 .opcode         = nvme_admin_identify,
 165                 .nsid           = nsid,
 166                 .addr           = (__u64)(uintptr_t)data,
 167                 .data_len       = NVME_IDENTIFY_DATA_SIZE,
 168                 .cdw10          = cns,
 169                 .cdw11          = csi << NVME_IDENTIFY_CSI_SHIFT,
 170                 .timeout_ms     = NVME_DEFAULT_IOCTL_TIMEOUT,
 171         };
 172
 173         return ioctl(fd, NVME_IOCTL_ADMIN_CMD, &cmd);
 174 }
 175
 176 static int nvme_get_info(int fd, __u32 *nsid, __u32 *lba_sz, __u64 *nlba)
 177 {
 178         struct nvme_id_ns ns;
 179         int namespace_id;
 180         int err;
 181
 182         namespace_id = ioctl(fd, NVME_IOCTL_ID);
 183         if (namespace_id < 0) {
 184                 fprintf(stderr, "error failed to fetch namespace-id\n");
 185                 close(fd);
 186                 return -errno;
 187         }
 188
 189         /*
 190          * Identify namespace to get namespace-id, namespace size in LBA's
 191          * and LBA data size.
 192          */
 193         err = nvme_identify(fd, namespace_id, NVME_IDENTIFY_CNS_NS,
 194                                 NVME_CSI_NVM, &ns);
 195         if (err) {
 196                 fprintf(stderr, "error failed to fetch identify namespace\n");
 197                 close(fd);
 198                 return err;
 199         }
 200
 201         *nsid = namespace_id;
 202         *lba_sz = 1 << ns.lbaf[(ns.flbas & 0x0f)].ds;
 203         *nlba = ns.nsze;
 204
 205         return 0;
 206 }
 207
 208 static unsigned long cycles_to_nsec(unsigned long cycles)
 209 {
 210         uint64_t val;
 211
 212         if (!tsc_rate)
 213                 return cycles;
 214
 215         val = cycles * 1000000000ULL;
 216         return val / tsc_rate;
 217 }
 218
 219 static unsigned long plat_idx_to_val(unsigned int idx)
 220 {
 221         unsigned int error_bits;
 222         unsigned long k, base;
 223
 224         assert(idx < PLAT_NR);
 225
 226         /* MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 227          * all bits of the sample as index */
 228         if (idx < (PLAT_VAL << 1))
 229                 return cycles_to_nsec(idx);
 230
 231         /* Find the group and compute the minimum value of that group */
 232         error_bits = (idx >> PLAT_BITS) - 1;
 233         base = ((unsigned long) 1) << (error_bits + PLAT_BITS);
 234
 235         /* Find its bucket number of the group */
 236         k = idx % PLAT_VAL;
 237
 238         /* Return the mean of the range of the bucket */
 239         return cycles_to_nsec(base + ((k + 0.5) * (1 << error_bits)));
 240 }
 241
 242 unsigned int calculate_clat_percentiles(unsigned long *io_u_plat,
 243                 unsigned long nr, unsigned long **output,
 244                 unsigned long *maxv, unsigned long *minv)
 245 {
 246         unsigned long sum = 0;
 247         unsigned int len = plist_len, i, j = 0;
 248         unsigned long *ovals = NULL;
 249         bool is_last;
 250
 251         *minv = -1UL;
 252         *maxv = 0;
 253
 254         ovals = malloc(len * sizeof(*ovals));
 255         if (!ovals)
 256                 return 0;
 257
 258         /*
 259          * Calculate bucket values, note down max and min values
 260          */
 261         is_last = false;
 262         for (i = 0; i < PLAT_NR && !is_last; i++) {
 263                 sum += io_u_plat[i];
 264                 while (sum >= ((long double) plist[j] / 100.0 * nr)) {
 265                         assert(plist[j] <= 100.0);
 266
 267                         ovals[j] = plat_idx_to_val(i);
 268                         if (ovals[j] < *minv)
 269                                 *minv = ovals[j];
 270                         if (ovals[j] > *maxv)
 271                                 *maxv = ovals[j];
 272
 273                         is_last = (j == len - 1) != 0;
 274                         if (is_last)
 275                                 break;
 276
 277                         j++;
 278                 }
 279         }
 280
 281         if (!is_last)
 282                 fprintf(stderr, "error calculating latency percentiles\n");
 283
 284         *output = ovals;
 285         return len;
 286 }
 287
 288 static void show_clat_percentiles(unsigned long *io_u_plat, unsigned long nr,
 289                                   unsigned int precision)
 290 {
 291         unsigned int divisor, len, i, j = 0;
 292         unsigned long minv, maxv;
 293         unsigned long *ovals;
 294         int per_line, scale_down, time_width;
 295         bool is_last;
 296         char fmt[32];
 297
 298         len = calculate_clat_percentiles(io_u_plat, nr, &ovals, &maxv, &minv);
 299         if (!len || !ovals)
 300                 goto out;
 301
 302         if (!tsc_rate) {
 303                 scale_down = 0;
 304                 divisor = 1;
 305                 printf("    percentiles (tsc ticks):\n     |");
 306         } else if (minv > 2000 && maxv > 99999) {
 307                 scale_down = 1;
 308                 divisor = 1000;
 309                 printf("    percentiles (usec):\n     |");
 310         } else {
 311                 scale_down = 0;
 312                 divisor = 1;
 313                 printf("    percentiles (nsec):\n     |");
 314         }
 315
 316         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 317         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 318                         precision, time_width);
 319         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 320         per_line = (80 - 7) / (precision + 10 + time_width);
 321
 322         for (j = 0; j < len; j++) {
 323                 /* for formatting */
 324                 if (j != 0 && (j % per_line) == 0)
 325                         printf("     |");
 326
 327                 /* end of the list */
 328                 is_last = (j == len - 1) != 0;
 329
 330                 for (i = 0; i < scale_down; i++)
 331                         ovals[j] = (ovals[j] + 999) / 1000;
 332
 333                 printf(fmt, plist[j], ovals[j], is_last ? '\n' : ',');
 334
 335                 if (is_last)
 336                         break;
 337
 338                 if ((j % per_line) == per_line - 1)     /* for formatting */
 339                         printf("\n");
 340         }
 341
 342 out:
 343         free(ovals);
 344 }
 345
 346 #ifdef ARCH_HAVE_CPU_CLOCK
 347 static unsigned int plat_val_to_idx(unsigned long val)
 348 {
 349         unsigned int msb, error_bits, base, offset, idx;
 350
 351         /* Find MSB starting from bit 0 */
 352         if (val == 0)
 353                 msb = 0;
 354         else
 355                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
 356
 357         /*
 358          * MSB <= (PLAT_BITS-1), cannot be rounded off. Use
 359          * all bits of the sample as index
 360          */
 361         if (msb <= PLAT_BITS)
 362                 return val;
 363
 364         /* Compute the number of error bits to discard*/
 365         error_bits = msb - PLAT_BITS;
 366
 367         /* Compute the number of buckets before the group */
 368         base = (error_bits + 1) << PLAT_BITS;
 369
 370         /*
 371          * Discard the error bits and apply the mask to find the
 372          * index for the buckets in the group
 373          */
 374         offset = (PLAT_VAL - 1) & (val >> error_bits);
 375
 376         /* Make sure the index does not exceed (array size - 1) */
 377         idx = (base + offset) < (PLAT_NR - 1) ?
 378                 (base + offset) : (PLAT_NR - 1);
 379
 380         return idx;
 381 }
 382 #endif
 383
 384 static void add_stat(struct submitter *s, int clock_index, int nr)
 385 {
 386 #ifdef ARCH_HAVE_CPU_CLOCK
 387         unsigned long cycles;
 388         unsigned int pidx;
 389
 390         if (!s->finish && clock_index) {
 391                 cycles = get_cpu_clock();
 392                 cycles -= s->clock_batch[clock_index];
 393                 pidx = plat_val_to_idx(cycles);
 394                 s->plat[pidx] += nr;
 395         }
 396 #endif
 397 }
 398
 399 static int io_uring_register_buffers(struct submitter *s)
 400 {
 401         if (do_nop)
 402                 return 0;
 403
 404         return syscall(__NR_io_uring_register, s->ring_fd,
 405                         IORING_REGISTER_BUFFERS, s->iovecs, roundup_pow2(depth));
 406 }
 407
 408 static int io_uring_register_files(struct submitter *s)
 409 {
 410         int i;
 411
 412         if (do_nop)
 413                 return 0;
 414
 415         s->fds = calloc(s->nr_files, sizeof(__s32));
 416         for (i = 0; i < s->nr_files; i++) {
 417                 s->fds[i] = s->files[i].real_fd;
 418                 s->files[i].fixed_fd = i;
 419         }
 420
 421         return syscall(__NR_io_uring_register, s->ring_fd,
 422                         IORING_REGISTER_FILES, s->fds, s->nr_files);
 423 }
 424
 425 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 426 {
 427         int ret;
 428
 429         /*
 430          * Clamp CQ ring size at our SQ ring size, we don't need more entries
 431          * than that.
 432          */
 433         p->flags |= IORING_SETUP_CQSIZE;
 434         p->cq_entries = entries;
 435
 436         p->flags |= IORING_SETUP_COOP_TASKRUN;
 437         p->flags |= IORING_SETUP_SINGLE_ISSUER;
 438         p->flags |= IORING_SETUP_DEFER_TASKRUN;
 439 retry:
 440         ret = syscall(__NR_io_uring_setup, entries, p);
 441         if (!ret)
 442                 return 0;
 443
 444         if (errno == EINVAL && p->flags & IORING_SETUP_COOP_TASKRUN) {
 445                 p->flags &= ~IORING_SETUP_COOP_TASKRUN;
 446                 goto retry;
 447         }
 448         if (errno == EINVAL && p->flags & IORING_SETUP_SINGLE_ISSUER) {
 449                 p->flags &= ~IORING_SETUP_SINGLE_ISSUER;
 450                 goto retry;
 451         }
 452         if (errno == EINVAL && p->flags & IORING_SETUP_DEFER_TASKRUN) {
 453                 p->flags &= ~IORING_SETUP_DEFER_TASKRUN;
 454                 goto retry;
 455         }
 456
 457         return ret;
 458 }
 459
 460 static void io_uring_probe(int fd)
 461 {
 462         struct io_uring_probe *p;
 463         int ret;
 464
 465         p = calloc(1, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 466         if (!p)
 467                 return;
 468
 469         ret = syscall(__NR_io_uring_register, fd, IORING_REGISTER_PROBE, p, 256);
 470         if (ret < 0)
 471                 goto out;
 472
 473         if (IORING_OP_READ > p->ops_len)
 474                 goto out;
 475
 476         if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED))
 477                 vectored = 0;
 478 out:
 479         free(p);
 480 }
 481
 482 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 483                           unsigned int min_complete, unsigned int flags)
 484 {
 485         if (register_ring)
 486                 flags |= IORING_ENTER_REGISTERED_RING;
 487 #ifdef FIO_ARCH_HAS_SYSCALL
 488         return __do_syscall6(__NR_io_uring_enter, s->enter_ring_fd, to_submit,
 489                                 min_complete, flags, NULL, 0);
 490 #else
 491         return syscall(__NR_io_uring_enter, s->enter_ring_fd, to_submit,
 492                         min_complete, flags, NULL, 0);
 493 #endif
 494 }
 495
 496 static unsigned long long get_offset(struct submitter *s, struct file *f)
 497 {
 498         unsigned long long offset;
 499         long r;
 500
 501         if (random_io) {
 502                 unsigned long long block;
 503
 504                 r = __rand64(&s->rand_state);
 505                 block = r % f->max_blocks;
 506                 offset = block * (unsigned long long) bs;
 507         } else {
 508                 offset = f->cur_off;
 509                 f->cur_off += bs;
 510                 if (f->cur_off + bs > f->max_size)
 511                         f->cur_off = 0;
 512         }
 513
 514         return offset;
 515 }
 516
 517 static struct file *get_next_file(struct submitter *s)
 518 {
 519         struct file *f;
 520
 521         if (s->nr_files == 1) {
 522                 f = &s->files[0];
 523         } else {
 524                 f = &s->files[s->cur_file];
 525                 if (f->pending_ios >= s->per_file_depth) {
 526                         s->cur_file++;
 527                         if (s->cur_file == s->nr_files)
 528                                 s->cur_file = 0;
 529                         f = &s->files[s->cur_file];
 530                 }
 531         }
 532
 533         f->pending_ios++;
 534         return f;
 535 }
 536
 537 static void init_io(struct submitter *s, unsigned index)
 538 {
 539         struct io_uring_sqe *sqe = &s->sqes[index];
 540         struct file *f;
 541
 542         if (do_nop) {
 543                 sqe->opcode = IORING_OP_NOP;
 544                 return;
 545         }
 546
 547         f = get_next_file(s);
 548
 549         if (register_files) {
 550                 sqe->flags = IOSQE_FIXED_FILE;
 551                 sqe->fd = f->fixed_fd;
 552         } else {
 553                 sqe->flags = 0;
 554                 sqe->fd = f->real_fd;
 555         }
 556         if (fixedbufs) {
 557                 sqe->opcode = IORING_OP_READ_FIXED;
 558                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 559                 sqe->len = bs;
 560                 sqe->buf_index = index;
 561         } else if (!vectored) {
 562                 sqe->opcode = IORING_OP_READ;
 563                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 564                 sqe->len = bs;
 565                 sqe->buf_index = 0;
 566         } else {
 567                 sqe->opcode = IORING_OP_READV;
 568                 sqe->addr = (unsigned long) &s->iovecs[index];
 569                 sqe->len = 1;
 570                 sqe->buf_index = 0;
 571         }
 572         sqe->ioprio = 0;
 573         sqe->off = get_offset(s, f);
 574         sqe->user_data = (unsigned long) f->fileno;
 575         if (stats && stats_running)
 576                 sqe->user_data |= ((uint64_t)s->clock_index << 32);
 577 }
 578
 579 static void init_io_pt(struct submitter *s, unsigned index)
 580 {
 581         struct io_uring_sqe *sqe = &s->sqes[index << 1];
 582         unsigned long offset;
 583         struct file *f;
 584         struct nvme_uring_cmd *cmd;
 585         unsigned long long slba;
 586         unsigned long long nlb;
 587
 588         f = get_next_file(s);
 589
 590         offset = get_offset(s, f);
 591
 592         if (register_files) {
 593                 sqe->fd = f->fixed_fd;
 594                 sqe->flags = IOSQE_FIXED_FILE;
 595         } else {
 596                 sqe->fd = f->real_fd;
 597                 sqe->flags = 0;
 598         }
 599         sqe->opcode = IORING_OP_URING_CMD;
 600         sqe->user_data = (unsigned long) f->fileno;
 601         if (stats)
 602                 sqe->user_data |= ((__u64) s->clock_index << 32ULL);
 603         sqe->cmd_op = NVME_URING_CMD_IO;
 604         slba = offset >> f->lba_shift;
 605         nlb = (bs >> f->lba_shift) - 1;
 606         cmd = (struct nvme_uring_cmd *)&sqe->cmd;
 607         /* cdw10 and cdw11 represent starting slba*/
 608         cmd->cdw10 = slba & 0xffffffff;
 609         cmd->cdw11 = slba >> 32;
 610         /* cdw12 represent number of lba to be read*/
 611         cmd->cdw12 = nlb;
 612         cmd->addr = (unsigned long) s->iovecs[index].iov_base;
 613         cmd->data_len = bs;
 614         if (fixedbufs) {
 615                 sqe->uring_cmd_flags = IORING_URING_CMD_FIXED;
 616                 sqe->buf_index = index;
 617         }
 618         cmd->nsid = f->nsid;
 619         cmd->opcode = 2;
 620 }
 621
 622 static int prep_more_ios_uring(struct submitter *s, int max_ios)
 623 {
 624         struct io_sq_ring *ring = &s->sq_ring;
 625         unsigned head, index, tail, next_tail, prepped = 0;
 626
 627         if (sq_thread_poll)
 628                 head = atomic_load_acquire(ring->head);
 629         else
 630                 head = *ring->head;
 631
 632         next_tail = tail = *ring->tail;
 633         do {
 634                 next_tail++;
 635                 if (next_tail == head)
 636                         break;
 637
 638                 index = tail & sq_ring_mask;
 639                 if (pt)
 640                         init_io_pt(s, index);
 641                 else
 642                         init_io(s, index);
 643                 prepped++;
 644                 tail = next_tail;
 645         } while (prepped < max_ios);
 646
 647         if (prepped)
 648                 atomic_store_release(ring->tail, tail);
 649         return prepped;
 650 }
 651
 652 static int get_file_size(struct file *f)
 653 {
 654         struct stat st;
 655
 656         if (fstat(f->real_fd, &st) < 0)
 657                 return -1;
 658         if (pt) {
 659                 __u64 nlba;
 660                 __u32 lbs;
 661                 int ret;
 662
 663                 if (!S_ISCHR(st.st_mode)) {
 664                         fprintf(stderr, "passthrough works with only nvme-ns "
 665                                         "generic devices (/dev/ngXnY)\n");
 666                         return -1;
 667                 }
 668                 ret = nvme_get_info(f->real_fd, &f->nsid, &lbs, &nlba);
 669                 if (ret)
 670                         return -1;
 671                 if ((bs % lbs) != 0) {
 672                         printf("error: bs:%d should be a multiple logical_block_size:%d\n",
 673                                         bs, lbs);
 674                         return -1;
 675                 }
 676                 f->max_blocks = nlba;
 677                 f->max_size = nlba;
 678                 f->lba_shift = ilog2(lbs);
 679                 return 0;
 680         } else if (S_ISBLK(st.st_mode)) {
 681                 unsigned long long bytes;
 682
 683                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 684                         return -1;
 685
 686                 f->max_blocks = bytes / bs;
 687                 f->max_size = bytes;
 688                 return 0;
 689         } else if (S_ISREG(st.st_mode)) {
 690                 f->max_blocks = st.st_size / bs;
 691                 f->max_size = st.st_size;
 692                 return 0;
 693         }
 694
 695         return -1;
 696 }
 697
 698 static int reap_events_uring(struct submitter *s)
 699 {
 700         struct io_cq_ring *ring = &s->cq_ring;
 701         struct io_uring_cqe *cqe;
 702         unsigned head, reaped = 0;
 703         int last_idx = -1, stat_nr = 0;
 704
 705         head = *ring->head;
 706         do {
 707                 struct file *f;
 708
 709                 if (head == atomic_load_acquire(ring->tail))
 710                         break;
 711                 cqe = &ring->cqes[head & cq_ring_mask];
 712                 if (!do_nop) {
 713                         int fileno = cqe->user_data & 0xffffffff;
 714
 715                         f = &s->files[fileno];
 716                         f->pending_ios--;
 717                         if (cqe->res != bs) {
 718                                 if (cqe->res == -ENODATA || cqe->res == -EIO) {
 719                                         s->io_errors++;
 720                                 } else {
 721                                         printf("io: unexpected ret=%d\n", cqe->res);
 722                                         if (polled && cqe->res == -EOPNOTSUPP)
 723                                                 printf("Your filesystem/driver/kernel doesn't support polled IO\n");
 724                                         return -1;
 725                                 }
 726                         }
 727                 }
 728                 if (stats) {
 729                         int clock_index = cqe->user_data >> 32;
 730
 731                         if (last_idx != clock_index) {
 732                                 if (last_idx != -1) {
 733                                         add_stat(s, last_idx, stat_nr);
 734                                         stat_nr = 0;
 735                                 }
 736                                 last_idx = clock_index;
 737                         }
 738                         stat_nr++;
 739                 }
 740                 reaped++;
 741                 head++;
 742         } while (1);
 743
 744         if (stat_nr)
 745                 add_stat(s, last_idx, stat_nr);
 746
 747         if (reaped) {
 748                 s->inflight -= reaped;
 749                 atomic_store_release(ring->head, head);
 750         }
 751         return reaped;
 752 }
 753
 754 static int reap_events_uring_pt(struct submitter *s)
 755 {
 756         struct io_cq_ring *ring = &s->cq_ring;
 757         struct io_uring_cqe *cqe;
 758         unsigned head, reaped = 0;
 759         int last_idx = -1, stat_nr = 0;
 760         unsigned index;
 761         int fileno;
 762
 763         head = *ring->head;
 764         do {
 765                 struct file *f;
 766
 767                 if (head == atomic_load_acquire(ring->tail))
 768                         break;
 769                 index = head & cq_ring_mask;
 770                 cqe = &ring->cqes[index << 1];
 771                 fileno = cqe->user_data & 0xffffffff;
 772                 f = &s->files[fileno];
 773                 f->pending_ios--;
 774
 775                 if (cqe->res != 0) {
 776                         printf("io: unexpected ret=%d\n", cqe->res);
 777                         if (polled && cqe->res == -EINVAL)
 778                                 printf("passthrough doesn't support polled IO\n");
 779                         return -1;
 780                 }
 781                 if (stats) {
 782                         int clock_index = cqe->user_data >> 32;
 783
 784                         if (last_idx != clock_index) {
 785                                 if (last_idx != -1) {
 786                                         add_stat(s, last_idx, stat_nr);
 787                                         stat_nr = 0;
 788                                 }
 789                                 last_idx = clock_index;
 790                         }
 791                         stat_nr++;
 792                 }
 793                 reaped++;
 794                 head++;
 795         } while (1);
 796
 797         if (stat_nr)
 798                 add_stat(s, last_idx, stat_nr);
 799
 800         if (reaped) {
 801                 s->inflight -= reaped;
 802                 atomic_store_release(ring->head, head);
 803         }
 804         return reaped;
 805 }
 806
 807 static void set_affinity(struct submitter *s)
 808 {
 809 #ifdef CONFIG_LIBNUMA
 810         struct bitmask *mask;
 811
 812         if (s->numa_node == -1)
 813                 return;
 814
 815         numa_set_preferred(s->numa_node);
 816
 817         mask = numa_allocate_cpumask();
 818         numa_node_to_cpus(s->numa_node, mask);
 819         numa_sched_setaffinity(s->tid, mask);
 820 #endif
 821 }
 822
 823 static int detect_node(struct submitter *s, char *name)
 824 {
 825 #ifdef CONFIG_LIBNUMA
 826         const char *base = basename(name);
 827         char str[128];
 828         int ret, fd, node;
 829
 830         if (pt)
 831                 sprintf(str, "/sys/class/nvme-generic/%s/device/numa_node", base);
 832         else
 833                 sprintf(str, "/sys/block/%s/device/numa_node", base);
 834         fd = open(str, O_RDONLY);
 835         if (fd < 0)
 836                 return -1;
 837
 838         ret = read(fd, str, sizeof(str));
 839         if (ret < 0) {
 840                 close(fd);
 841                 return -1;
 842         }
 843         node = atoi(str);
 844         s->numa_node = node;
 845         close(fd);
 846 #else
 847         s->numa_node = -1;
 848 #endif
 849         return 0;
 850 }
 851
 852 static int setup_aio(struct submitter *s)
 853 {
 854 #ifdef CONFIG_LIBAIO
 855         if (polled) {
 856                 fprintf(stderr, "aio does not support polled IO\n");
 857                 polled = 0;
 858         }
 859         if (sq_thread_poll) {
 860                 fprintf(stderr, "aio does not support SQPOLL IO\n");
 861                 sq_thread_poll = 0;
 862         }
 863         if (do_nop) {
 864                 fprintf(stderr, "aio does not support polled IO\n");
 865                 do_nop = 0;
 866         }
 867         if (fixedbufs || register_files) {
 868                 fprintf(stderr, "aio does not support registered files or buffers\n");
 869                 fixedbufs = register_files = 0;
 870         }
 871
 872         s->per_file_depth = (depth + s->nr_files - 1) / s->nr_files;
 873         return io_queue_init(roundup_pow2(depth), &s->aio_ctx);
 874 #else
 875         fprintf(stderr, "Legacy AIO not available on this system/build\n");
 876         errno = EINVAL;
 877         return -1;
 878 #endif
 879 }
 880
 881 static int setup_ring(struct submitter *s)
 882 {
 883         struct io_sq_ring *sring = &s->sq_ring;
 884         struct io_cq_ring *cring = &s->cq_ring;
 885         struct io_uring_params p;
 886         int ret, fd, i;
 887         void *ptr;
 888         size_t len;
 889
 890         memset(&p, 0, sizeof(p));
 891
 892         if (polled && !do_nop)
 893                 p.flags |= IORING_SETUP_IOPOLL;
 894         if (sq_thread_poll) {
 895                 p.flags |= IORING_SETUP_SQPOLL;
 896                 if (sq_thread_cpu != -1) {
 897                         p.flags |= IORING_SETUP_SQ_AFF;
 898                         p.sq_thread_cpu = sq_thread_cpu;
 899                 }
 900         }
 901         if (pt) {
 902                 p.flags |= IORING_SETUP_SQE128;
 903                 p.flags |= IORING_SETUP_CQE32;
 904         }
 905
 906         fd = io_uring_setup(depth, &p);
 907         if (fd < 0) {
 908                 perror("io_uring_setup");
 909                 return 1;
 910         }
 911         s->ring_fd = s->enter_ring_fd = fd;
 912
 913         io_uring_probe(fd);
 914
 915         if (fixedbufs) {
 916                 struct rlimit rlim;
 917
 918                 rlim.rlim_cur = RLIM_INFINITY;
 919                 rlim.rlim_max = RLIM_INFINITY;
 920                 /* ignore potential error, not needed on newer kernels */
 921                 setrlimit(RLIMIT_MEMLOCK, &rlim);
 922
 923                 ret = io_uring_register_buffers(s);
 924                 if (ret < 0) {
 925                         perror("io_uring_register_buffers");
 926                         return 1;
 927                 }
 928         }
 929
 930         if (register_files) {
 931                 ret = io_uring_register_files(s);
 932                 if (ret < 0) {
 933                         perror("io_uring_register_files");
 934                         return 1;
 935                 }
 936         }
 937
 938         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 939                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 940                         IORING_OFF_SQ_RING);
 941         sring->head = ptr + p.sq_off.head;
 942         sring->tail = ptr + p.sq_off.tail;
 943         sring->ring_mask = ptr + p.sq_off.ring_mask;
 944         sring->ring_entries = ptr + p.sq_off.ring_entries;
 945         sring->flags = ptr + p.sq_off.flags;
 946         sring->array = ptr + p.sq_off.array;
 947         sq_ring_mask = *sring->ring_mask;
 948
 949         if (p.flags & IORING_SETUP_SQE128)
 950                 len = 2 * p.sq_entries * sizeof(struct io_uring_sqe);
 951         else
 952                 len = p.sq_entries * sizeof(struct io_uring_sqe);
 953         s->sqes = mmap(0, len,
 954                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 955                         IORING_OFF_SQES);
 956
 957         if (p.flags & IORING_SETUP_CQE32) {
 958                 len = p.cq_off.cqes +
 959                         2 * p.cq_entries * sizeof(struct io_uring_cqe);
 960         } else {
 961                 len = p.cq_off.cqes +
 962                         p.cq_entries * sizeof(struct io_uring_cqe);
 963         }
 964         ptr = mmap(0, len,
 965                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 966                         IORING_OFF_CQ_RING);
 967         cring->head = ptr + p.cq_off.head;
 968         cring->tail = ptr + p.cq_off.tail;
 969         cring->ring_mask = ptr + p.cq_off.ring_mask;
 970         cring->ring_entries = ptr + p.cq_off.ring_entries;
 971         cring->cqes = ptr + p.cq_off.cqes;
 972         cq_ring_mask = *cring->ring_mask;
 973
 974         for (i = 0; i < p.sq_entries; i++)
 975                 sring->array[i] = i;
 976
 977         s->per_file_depth = INT_MAX;
 978         if (s->nr_files)
 979                 s->per_file_depth = (depth + s->nr_files - 1) / s->nr_files;
 980         return 0;
 981 }
 982
 983 static void *allocate_mem(struct submitter *s, int size)
 984 {
 985         void *buf;
 986
 987 #ifdef CONFIG_LIBNUMA
 988         if (s->numa_node != -1)
 989                 return numa_alloc_onnode(size, s->numa_node);
 990 #endif
 991
 992         if (posix_memalign(&buf, t_io_uring_page_size, bs)) {
 993                 printf("failed alloc\n");
 994                 return NULL;
 995         }
 996
 997         return buf;
 998 }
 999
1000 static int submitter_init(struct submitter *s)
1001 {
1002         int i, nr_batch, err;
1003         static int init_printed;
1004         char buf[80];
1005         s->tid = gettid();
1006         printf("submitter=%d, tid=%d, file=%s, nfiles=%d, node=%d\n", s->index, s->tid,
1007                                                         s->filename, s->nr_files, s->numa_node);
1008
1009         set_affinity(s);
1010
1011         __init_rand64(&s->rand_state, s->tid);
1012         srand48(s->tid);
1013
1014         for (i = 0; i < MAX_FDS; i++)
1015                 s->files[i].fileno = i;
1016
1017         for (i = 0; i < roundup_pow2(depth); i++) {
1018                 void *buf;
1019
1020                 buf = allocate_mem(s, bs);
1021                 if (!buf)
1022                         return -1;
1023                 s->iovecs[i].iov_base = buf;
1024                 s->iovecs[i].iov_len = bs;
1025         }
1026
1027         if (use_sync) {
1028                 sprintf(buf, "Engine=preadv2\n");
1029                 err = 0;
1030         } else if (!aio) {
1031                 err = setup_ring(s);
1032                 if (!err)
1033                         sprintf(buf, "Engine=io_uring, sq_ring=%d, cq_ring=%d\n", *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
1034         } else {
1035                 sprintf(buf, "Engine=aio\n");
1036                 err = setup_aio(s);
1037         }
1038         if (err) {
1039                 printf("queue setup failed: %s, %d\n", strerror(errno), err);
1040                 return -1;
1041         }
1042
1043         if (!init_printed) {
1044                 printf("polled=%d, fixedbufs=%d, register_files=%d, buffered=%d, QD=%d\n", polled, fixedbufs, register_files, buffered, depth);
1045                 printf("%s", buf);
1046                 init_printed = 1;
1047         }
1048
1049         if (stats) {
1050                 nr_batch = roundup_pow2(depth / batch_submit);
1051                 if (nr_batch < 2)
1052                         nr_batch = 2;
1053                 s->clock_batch = calloc(nr_batch, sizeof(unsigned long));
1054                 s->clock_index = 1;
1055
1056                 s->plat = calloc(PLAT_NR, sizeof(unsigned long));
1057         } else {
1058                 s->clock_batch = NULL;
1059                 s->plat = NULL;
1060                 nr_batch = 0;
1061         }
1062         /* perform the expensive command initialization part for passthrough here
1063          * rather than in the fast path
1064          */
1065         if (pt) {
1066                 for (i = 0; i < roundup_pow2(depth); i++) {
1067                         struct io_uring_sqe *sqe = &s->sqes[i << 1];
1068
1069                         memset(&sqe->cmd, 0, sizeof(struct nvme_uring_cmd));
1070                 }
1071         }
1072         return nr_batch;
1073 }
1074
1075 #ifdef CONFIG_LIBAIO
1076 static int prep_more_ios_aio(struct submitter *s, int max_ios, struct iocb *iocbs)
1077 {
1078         uint64_t data;
1079         struct file *f;
1080         unsigned index;
1081
1082         index = 0;
1083         while (index < max_ios) {
1084                 struct iocb *iocb = &iocbs[index];
1085
1086                 f = get_next_file(s);
1087
1088                 io_prep_pread(iocb, f->real_fd, s->iovecs[index].iov_base,
1089                                 s->iovecs[index].iov_len, get_offset(s, f));
1090
1091                 data = f->fileno;
1092                 if (stats && stats_running)
1093                         data |= (((uint64_t) s->clock_index) << 32);
1094                 iocb->data = (void *) (uintptr_t) data;
1095                 index++;
1096         }
1097         return index;
1098 }
1099
1100 static int reap_events_aio(struct submitter *s, struct io_event *events, int evs)
1101 {
1102         int last_idx = -1, stat_nr = 0;
1103         int reaped = 0;
1104
1105         while (evs) {
1106                 uint64_t data = (uintptr_t) events[reaped].data;
1107                 struct file *f = &s->files[data & 0xffffffff];
1108
1109                 f->pending_ios--;
1110                 if (events[reaped].res != bs) {
1111                         if (events[reaped].res == -ENODATA ||
1112                             events[reaped].res == -EIO) {
1113                                 s->io_errors++;
1114                         } else {
1115                                 printf("io: unexpected ret=%ld\n", events[reaped].res);
1116                                 return -1;
1117                         }
1118                 } else if (stats) {
1119                         int clock_index = data >> 32;
1120
1121                         if (last_idx != clock_index) {
1122                                 if (last_idx != -1) {
1123                                         add_stat(s, last_idx, stat_nr);
1124                                         stat_nr = 0;
1125                                 }
1126                                 last_idx = clock_index;
1127                         }
1128                         stat_nr++;
1129                 }
1130                 reaped++;
1131                 evs--;
1132         }
1133
1134         if (stat_nr)
1135                 add_stat(s, last_idx, stat_nr);
1136
1137         s->inflight -= reaped;
1138         s->done += reaped;
1139         return reaped;
1140 }
1141
1142 static void *submitter_aio_fn(void *data)
1143 {
1144         struct submitter *s = data;
1145         int i, ret, prepped;
1146         struct iocb **iocbsptr;
1147         struct iocb *iocbs;
1148         struct io_event *events;
1149 #ifdef ARCH_HAVE_CPU_CLOCK
1150         int nr_batch;
1151 #endif
1152
1153         ret = submitter_init(s);
1154         if (ret < 0)
1155                 goto done;
1156
1157 #ifdef ARCH_HAVE_CPU_CLOCK
1158         nr_batch = ret;
1159 #endif
1160
1161         iocbsptr = calloc(depth, sizeof(struct iocb *));
1162         iocbs = calloc(depth, sizeof(struct iocb));
1163         events = calloc(depth, sizeof(struct io_event));
1164
1165         for (i = 0; i < depth; i++)
1166                 iocbsptr[i] = &iocbs[i];
1167
1168         prepped = 0;
1169         do {
1170                 int to_wait, to_submit, to_prep;
1171
1172                 if (!prepped && s->inflight < depth) {
1173                         to_prep = min(depth - s->inflight, batch_submit);
1174                         prepped = prep_more_ios_aio(s, to_prep, iocbs);
1175 #ifdef ARCH_HAVE_CPU_CLOCK
1176                         if (prepped && stats) {
1177                                 s->clock_batch[s->clock_index] = get_cpu_clock();
1178                                 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
1179                         }
1180 #endif
1181                 }
1182                 s->inflight += prepped;
1183                 to_submit = prepped;
1184
1185                 if (to_submit && (s->inflight + to_submit <= depth))
1186                         to_wait = 0;
1187                 else
1188                         to_wait = min(s->inflight + to_submit, batch_complete);
1189
1190                 ret = io_submit(s->aio_ctx, to_submit, iocbsptr);
1191                 s->calls++;
1192                 if (ret < 0) {
1193                         perror("io_submit");
1194                         break;
1195                 } else if (ret != to_submit) {
1196                         printf("submitted %d, wanted %d\n", ret, to_submit);
1197                         break;
1198                 }
1199                 prepped = 0;
1200
1201                 while (to_wait) {
1202                         int r;
1203
1204                         s->calls++;
1205                         r = io_getevents(s->aio_ctx, to_wait, to_wait, events, NULL);
1206                         if (r < 0) {
1207                                 perror("io_getevents");
1208                                 break;
1209                         } else if (r != to_wait) {
1210                                 printf("r=%d, wait=%d\n", r, to_wait);
1211                                 break;
1212                         }
1213                         r = reap_events_aio(s, events, r);
1214                         s->reaps += r;
1215                         to_wait -= r;
1216                 }
1217         } while (!s->finish);
1218
1219         free(iocbsptr);
1220         free(iocbs);
1221         free(events);
1222 done:
1223         finish = 1;
1224         return NULL;
1225 }
1226 #endif
1227
1228 static void io_uring_unregister_ring(struct submitter *s)
1229 {
1230         struct io_uring_rsrc_update up = {
1231                 .offset = s->enter_ring_fd,
1232         };
1233
1234         syscall(__NR_io_uring_register, s->ring_fd, IORING_UNREGISTER_RING_FDS,
1235                 &up, 1);
1236 }
1237
1238 static int io_uring_register_ring(struct submitter *s)
1239 {
1240         struct io_uring_rsrc_update up = {
1241                 .data   = s->ring_fd,
1242                 .offset = -1U,
1243         };
1244         int ret;
1245
1246         ret = syscall(__NR_io_uring_register, s->ring_fd,
1247                         IORING_REGISTER_RING_FDS, &up, 1);
1248         if (ret == 1) {
1249                 s->enter_ring_fd = up.offset;
1250                 return 0;
1251         }
1252         register_ring = 0;
1253         return -1;
1254 }
1255
1256 static void *submitter_uring_fn(void *data)
1257 {
1258         struct submitter *s = data;
1259         struct io_sq_ring *ring = &s->sq_ring;
1260         int ret, prepped;
1261 #ifdef ARCH_HAVE_CPU_CLOCK
1262         int nr_batch;
1263 #endif
1264
1265         ret = submitter_init(s);
1266         if (ret < 0)
1267                 goto done;
1268
1269 #ifdef ARCH_HAVE_CPU_CLOCK
1270         nr_batch = ret;
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 done:
1375         finish = 1;
1376         return NULL;
1377 }
1378
1379 #ifdef CONFIG_PWRITEV2
1380 static void *submitter_sync_fn(void *data)
1381 {
1382         struct submitter *s = data;
1383         int ret;
1384
1385         if (submitter_init(s) < 0)
1386                 goto done;
1387
1388         do {
1389                 uint64_t offset;
1390                 struct file *f;
1391
1392                 f = get_next_file(s);
1393
1394 #ifdef ARCH_HAVE_CPU_CLOCK
1395                 if (stats)
1396                         s->clock_batch[s->clock_index] = get_cpu_clock();
1397 #endif
1398
1399                 s->inflight++;
1400                 s->calls++;
1401
1402                 offset = get_offset(s, f);
1403                 if (polled)
1404                         ret = preadv2(f->real_fd, &s->iovecs[0], 1, offset, RWF_HIPRI);
1405                 else
1406                         ret = preadv2(f->real_fd, &s->iovecs[0], 1, offset, 0);
1407
1408                 if (ret < 0) {
1409                         perror("preadv2");
1410                         break;
1411                 } else if (ret != bs) {
1412                         break;
1413                 }
1414
1415                 s->done++;
1416                 s->inflight--;
1417                 f->pending_ios--;
1418                 if (stats)
1419                         add_stat(s, s->clock_index, 1);
1420         } while (!s->finish);
1421
1422 done:
1423         finish = 1;
1424         return NULL;
1425 }
1426 #else
1427 static void *submitter_sync_fn(void *data)
1428 {
1429         finish = 1;
1430         return NULL;
1431 }
1432 #endif
1433
1434 static struct submitter *get_submitter(int offset)
1435 {
1436         void *ret;
1437
1438         ret = submitter;
1439         if (offset)
1440                 ret += offset * (sizeof(*submitter) + depth * sizeof(struct iovec));
1441         return ret;
1442 }
1443
1444 static void do_finish(const char *reason)
1445 {
1446         int j;
1447
1448         printf("Exiting on %s\n", reason);
1449         for (j = 0; j < nthreads; j++) {
1450                 struct submitter *s = get_submitter(j);
1451                 s->finish = 1;
1452         }
1453         if (max_iops > 1000000) {
1454                 double miops = (double) max_iops / 1000000.0;
1455                 printf("Maximum IOPS=%.2fM\n", miops);
1456         } else if (max_iops > 100000) {
1457                 double kiops = (double) max_iops / 1000.0;
1458                 printf("Maximum IOPS=%.2fK\n", kiops);
1459         } else {
1460                 printf("Maximum IOPS=%lu\n", max_iops);
1461         }
1462         finish = 1;
1463 }
1464
1465 static void sig_int(int sig)
1466 {
1467         do_finish("signal");
1468 }
1469
1470 static void arm_sig_int(void)
1471 {
1472         struct sigaction act;
1473
1474         memset(&act, 0, sizeof(act));
1475         act.sa_handler = sig_int;
1476         act.sa_flags = SA_RESTART;
1477         sigaction(SIGINT, &act, NULL);
1478
1479         /* Windows uses SIGBREAK as a quit signal from other applications */
1480 #ifdef WIN32
1481         sigaction(SIGBREAK, &act, NULL);
1482 #endif
1483 }
1484
1485 static void usage(char *argv, int status)
1486 {
1487         char runtime_str[16];
1488         snprintf(runtime_str, sizeof(runtime_str), "%d", runtime);
1489         printf("%s [options] -- [filenames]\n"
1490                 " -d <int>  : IO Depth, default %d\n"
1491                 " -s <int>  : Batch submit, default %d\n"
1492                 " -c <int>  : Batch complete, default %d\n"
1493                 " -b <int>  : Block size, default %d\n"
1494                 " -p <bool> : Polled IO, default %d\n"
1495                 " -B <bool> : Fixed buffers, default %d\n"
1496                 " -F <bool> : Register files, default %d\n"
1497                 " -n <int>  : Number of threads, default %d\n"
1498                 " -O <bool> : Use O_DIRECT, default %d\n"
1499                 " -N <bool> : Perform just no-op requests, default %d\n"
1500                 " -t <bool> : Track IO latencies, default %d\n"
1501                 " -T <int>  : TSC rate in HZ\n"
1502                 " -r <int>  : Runtime in seconds, default %s\n"
1503                 " -R <bool> : Use random IO, default %d\n"
1504                 " -a <bool> : Use legacy aio, default %d\n"
1505                 " -S <bool> : Use sync IO (preadv2), default %d\n"
1506                 " -X <bool> : Use registered ring %d\n"
1507                 " -P <bool> : Automatically place on device home node %d\n"
1508                 " -u <bool> : Use nvme-passthrough I/O, default %d\n",
1509                 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled,
1510                 fixedbufs, register_files, nthreads, !buffered, do_nop,
1511                 stats, runtime == 0 ? "unlimited" : runtime_str, random_io, aio,
1512                 use_sync, register_ring, numa_placement, pt);
1513         exit(status);
1514 }
1515
1516 static void read_tsc_rate(void)
1517 {
1518         char buffer[32];
1519         int fd, ret;
1520
1521         if (tsc_rate)
1522                 return;
1523
1524         fd = open(TSC_RATE_FILE, O_RDONLY);
1525         if (fd < 0)
1526                 return;
1527
1528         ret = read(fd, buffer, sizeof(buffer));
1529         if (ret < 0) {
1530                 close(fd);
1531                 return;
1532         }
1533
1534         tsc_rate = strtoul(buffer, NULL, 10);
1535         printf("Using TSC rate %luHz\n", tsc_rate);
1536         close(fd);
1537 }
1538
1539 static void write_tsc_rate(void)
1540 {
1541         char buffer[32];
1542         struct stat sb;
1543         int fd, ret;
1544
1545         if (!stat(TSC_RATE_FILE, &sb))
1546                 return;
1547
1548         fd = open(TSC_RATE_FILE, O_WRONLY | O_CREAT, 0644);
1549         if (fd < 0)
1550                 return;
1551
1552         memset(buffer, 0, sizeof(buffer));
1553         sprintf(buffer, "%lu", tsc_rate);
1554         ret = write(fd, buffer, strlen(buffer));
1555         if (ret < 0)
1556                 perror("write");
1557         close(fd);
1558 }
1559
1560 int main(int argc, char *argv[])
1561 {
1562         struct submitter *s;
1563         unsigned long done, calls, reap, io_errors;
1564         int i, j, flags, fd, opt, threads_per_f, threads_rem = 0, nfiles;
1565         struct file f;
1566         void *ret;
1567
1568         if (!do_nop && argc < 2)
1569                 usage(argv[0], 1);
1570
1571         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) {
1572                 switch (opt) {
1573                 case 'a':
1574                         aio = !!atoi(optarg);
1575                         break;
1576                 case 'd':
1577                         depth = atoi(optarg);
1578                         break;
1579                 case 's':
1580                         batch_submit = atoi(optarg);
1581                         if (!batch_submit)
1582                                 batch_submit = 1;
1583                         break;
1584                 case 'c':
1585                         batch_complete = atoi(optarg);
1586                         if (!batch_complete)
1587                                 batch_complete = 1;
1588                         break;
1589                 case 'b':
1590                         bs = atoi(optarg);
1591                         break;
1592                 case 'p':
1593                         polled = !!atoi(optarg);
1594                         break;
1595                 case 'B':
1596                         fixedbufs = !!atoi(optarg);
1597                         break;
1598                 case 'F':
1599                         register_files = !!atoi(optarg);
1600                         break;
1601                 case 'n':
1602                         nthreads = atoi(optarg);
1603                         if (!nthreads) {
1604                                 printf("Threads must be non-zero\n");
1605                                 usage(argv[0], 1);
1606                         }
1607                         break;
1608                 case 'N':
1609                         do_nop = !!atoi(optarg);
1610                         break;
1611                 case 'O':
1612                         buffered = !atoi(optarg);
1613                         break;
1614                 case 't':
1615 #ifndef ARCH_HAVE_CPU_CLOCK
1616                         fprintf(stderr, "Stats not supported on this CPU\n");
1617                         return 1;
1618 #endif
1619                         stats = !!atoi(optarg);
1620                         break;
1621                 case 'T':
1622 #ifndef ARCH_HAVE_CPU_CLOCK
1623                         fprintf(stderr, "Stats not supported on this CPU\n");
1624                         return 1;
1625 #endif
1626                         tsc_rate = strtoul(optarg, NULL, 10);
1627                         write_tsc_rate();
1628                         break;
1629                 case 'r':
1630                         runtime = atoi(optarg);
1631                         break;
1632                 case 'R':
1633                         random_io = !!atoi(optarg);
1634                         break;
1635                 case 'X':
1636                         register_ring = !!atoi(optarg);
1637                         break;
1638                 case 'S':
1639 #ifdef CONFIG_PWRITEV2
1640                         use_sync = !!atoi(optarg);
1641 #else
1642                         fprintf(stderr, "preadv2 not supported\n");
1643                         exit(1);
1644 #endif
1645                         break;
1646                 case 'P':
1647                         numa_placement = !!atoi(optarg);
1648                         break;
1649                 case 'u':
1650                         pt = !!atoi(optarg);
1651                         break;
1652                 case 'h':
1653                 case '?':
1654                 default:
1655                         usage(argv[0], 0);
1656                         break;
1657                 }
1658         }
1659
1660         if (stats)
1661                 read_tsc_rate();
1662
1663         if (batch_complete > depth)
1664                 batch_complete = depth;
1665         if (batch_submit > depth)
1666                 batch_submit = depth;
1667
1668         submitter = calloc(nthreads, sizeof(*submitter) +
1669                                 roundup_pow2(depth) * sizeof(struct iovec));
1670         for (j = 0; j < nthreads; j++) {
1671                 s = get_submitter(j);
1672                 s->numa_node = -1;
1673                 s->index = j;
1674                 s->done = s->calls = s->reaps = s->io_errors = 0;
1675         }
1676
1677         flags = O_RDONLY | O_NOATIME;
1678         if (!buffered)
1679                 flags |= O_DIRECT;
1680
1681         j = 0;
1682         i = optind;
1683         nfiles = argc - i;
1684         if (!do_nop) {
1685                 if (!nfiles) {
1686                         printf("No files specified\n");
1687                         usage(argv[0], 1);
1688                 }
1689                 threads_per_f = nthreads / nfiles;
1690                 /* make sure each thread gets assigned files */
1691                 if (threads_per_f == 0) {
1692                         threads_per_f = 1;
1693                 } else {
1694                         threads_rem = nthreads - threads_per_f * nfiles;
1695                 }
1696         }
1697         while (!do_nop && i < argc) {
1698                 int k, limit;
1699
1700                 memset(&f, 0, sizeof(f));
1701
1702                 fd = open(argv[i], flags);
1703                 if (fd < 0) {
1704                         perror("open");
1705                         return 1;
1706                 }
1707                 f.real_fd = fd;
1708                 if (get_file_size(&f)) {
1709                         printf("failed getting size of device/file\n");
1710                         return 1;
1711                 }
1712                 if (f.max_blocks <= 1) {
1713                         printf("Zero file/device size?\n");
1714                         return 1;
1715                 }
1716                 f.max_blocks--;
1717
1718                 limit = threads_per_f;
1719                 limit += threads_rem > 0 ? 1 : 0;
1720                 for (k = 0; k < limit; k++) {
1721                         s = get_submitter((j + k) % nthreads);
1722
1723                         if (s->nr_files == MAX_FDS) {
1724                                 printf("Max number of files (%d) reached\n", MAX_FDS);
1725                                 break;
1726                         }
1727
1728                         memcpy(&s->files[s->nr_files], &f, sizeof(f));
1729
1730                         if (numa_placement)
1731                                 detect_node(s, argv[i]);
1732
1733                         s->filename = argv[i];
1734                         s->nr_files++;
1735                 }
1736                 threads_rem--;
1737                 i++;
1738                 j += limit;
1739         }
1740
1741         arm_sig_int();
1742
1743         t_io_uring_page_size = sysconf(_SC_PAGESIZE);
1744         if (t_io_uring_page_size < 0)
1745                 t_io_uring_page_size = 4096;
1746
1747         for (j = 0; j < nthreads; j++) {
1748                 s = get_submitter(j);
1749                 if (use_sync)
1750                         pthread_create(&s->thread, NULL, submitter_sync_fn, s);
1751                 else if (!aio)
1752                         pthread_create(&s->thread, NULL, submitter_uring_fn, s);
1753 #ifdef CONFIG_LIBAIO
1754                 else
1755                         pthread_create(&s->thread, NULL, submitter_aio_fn, s);
1756 #endif
1757         }
1758
1759         reap = calls = done = io_errors = 0;
1760         do {
1761                 unsigned long this_done = 0;
1762                 unsigned long this_reap = 0;
1763                 unsigned long this_call = 0;
1764                 unsigned long this_io_errors = 0;
1765                 unsigned long rpc = 0, ipc = 0;
1766                 unsigned long iops, bw;
1767
1768                 sleep(1);
1769                 if (runtime && !--runtime)
1770                         do_finish("timeout");
1771
1772                 /* don't print partial run, if interrupted by signal */
1773                 if (finish)
1774                         break;
1775
1776                 /* one second in to the run, enable stats */
1777                 if (stats)
1778                         stats_running = 1;
1779
1780                 for (j = 0; j < nthreads; j++) {
1781                         s = get_submitter(j);
1782                         this_done += s->done;
1783                         this_call += s->calls;
1784                         this_reap += s->reaps;
1785                         this_io_errors += s->io_errors;
1786                 }
1787                 if (this_call - calls) {
1788                         rpc = (this_done - done) / (this_call - calls);
1789                         ipc = (this_reap - reap) / (this_call - calls);
1790                 } else
1791                         rpc = ipc = -1;
1792                 iops = this_done - done;
1793                 iops -= this_io_errors - io_errors;
1794                 if (bs > 1048576)
1795                         bw = iops * (bs / 1048576);
1796                 else
1797                         bw = iops / (1048576 / bs);
1798                 if (iops > 1000000) {
1799                         double miops = (double) iops / 1000000.0;
1800                         printf("IOPS=%.2fM, ", miops);
1801                 } else if (iops > 100000) {
1802                         double kiops = (double) iops / 1000.0;
1803                         printf("IOPS=%.2fK, ", kiops);
1804                 } else {
1805                         printf("IOPS=%lu, ", iops);
1806                 }
1807                 max_iops = max(max_iops, iops);
1808                 if (!do_nop) {
1809                         if (bw > 2000) {
1810                                 double bw_g = (double) bw / 1000.0;
1811
1812                                 printf("BW=%.2fGiB/s, ", bw_g);
1813                         } else {
1814                                 printf("BW=%luMiB/s, ", bw);
1815                         }
1816                 }
1817                 printf("IOS/call=%ld/%ld\n", rpc, ipc);
1818                 done = this_done;
1819                 calls = this_call;
1820                 reap = this_reap;
1821                 io_errors = this_io_errors;
1822         } while (!finish);
1823
1824         for (j = 0; j < nthreads; j++) {
1825                 s = get_submitter(j);
1826                 pthread_join(s->thread, &ret);
1827                 close(s->ring_fd);
1828
1829                 if (s->io_errors)
1830                         printf("%d: %lu IO errors\n", s->tid, s->io_errors);
1831
1832                 if (stats) {
1833                         unsigned long nr;
1834
1835                         printf("%d: Latency percentiles:\n", s->tid);
1836                         for (i = 0, nr = 0; i < PLAT_NR; i++)
1837                                 nr += s->plat[i];
1838                         show_clat_percentiles(s->plat, nr, 4);
1839                         free(s->clock_batch);
1840                         free(s->plat);
1841                 }
1842         }
1843
1844         free(submitter);
1845         return 0;
1846 }