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
   9 #include <sys/types.h>
  10 #include <sys/stat.h>
  11 #include <sys/ioctl.h>
  12 #include <sys/syscall.h>
  13 #include <sys/resource.h>
  14 #include <sys/mman.h>
  15 #include <sys/uio.h>
  16 #include <linux/fs.h>
  17 #include <fcntl.h>
  18 #include <unistd.h>
  19 #include <string.h>
  20 #include <pthread.h>
  21 #include <sched.h>
  22
  23 #include "../arch/arch.h"
  24 #include "../lib/types.h"
  25 #include "../os/linux/io_uring.h"
  26
  27 #define min(a, b)               ((a < b) ? (a) : (b))
  28
  29 struct io_sq_ring {
  30         unsigned *head;
  31         unsigned *tail;
  32         unsigned *ring_mask;
  33         unsigned *ring_entries;
  34         unsigned *flags;
  35         unsigned *array;
  36 };
  37
  38 struct io_cq_ring {
  39         unsigned *head;
  40         unsigned *tail;
  41         unsigned *ring_mask;
  42         unsigned *ring_entries;
  43         struct io_uring_cqe *cqes;
  44 };
  45
  46 #define DEPTH                   128
  47 #define BATCH_SUBMIT            32
  48 #define BATCH_COMPLETE          32
  49
  50 #define BS                      4096
  51
  52 #define MAX_FDS                 16
  53
  54 static unsigned sq_ring_mask, cq_ring_mask;
  55
  56 struct file {
  57         unsigned long max_blocks;
  58         unsigned pending_ios;
  59         int real_fd;
  60         int fixed_fd;
  61 };
  62
  63 struct submitter {
  64         pthread_t thread;
  65         int ring_fd;
  66         struct drand48_data rand;
  67         struct io_sq_ring sq_ring;
  68         struct io_uring_sqe *sqes;
  69         struct io_cq_ring cq_ring;
  70         int inflight;
  71         unsigned long reaps;
  72         unsigned long done;
  73         unsigned long calls;
  74         volatile int finish;
  75
  76         __s32 *fds;
  77
  78         struct file files[MAX_FDS];
  79         unsigned nr_files;
  80         unsigned cur_file;
  81         struct iovec iovecs[];
  82 };
  83
  84 static struct submitter *submitter;
  85 static volatile int finish;
  86
  87 static int depth = DEPTH;
  88 static int batch_submit = BATCH_SUBMIT;
  89 static int batch_complete = BATCH_COMPLETE;
  90 static int polled = 1;          /* use IO polling */
  91 static int fixedbufs = 1;       /* use fixed user buffers */
  92 static int register_files = 1;  /* use fixed files */
  93 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
  94 static int sq_thread_poll = 0;  /* use kernel submission/poller thread */
  95 static int sq_thread_cpu = -1;  /* pin above thread to this CPU */
  96 static int do_nop = 0;          /* no-op SQ ring commands */
  97
  98 static int io_uring_register_buffers(struct submitter *s)
  99 {
 100         if (do_nop)
 101                 return 0;
 102
 103         return syscall(__NR_sys_io_uring_register, s->ring_fd,
 104                         IORING_REGISTER_BUFFERS, s->iovecs, depth);
 105 }
 106
 107 static int io_uring_register_files(struct submitter *s)
 108 {
 109         int i;
 110
 111         if (do_nop)
 112                 return 0;
 113
 114         s->fds = calloc(s->nr_files, sizeof(__s32));
 115         for (i = 0; i < s->nr_files; i++) {
 116                 s->fds[i] = s->files[i].real_fd;
 117                 s->files[i].fixed_fd = i;
 118         }
 119
 120         return syscall(__NR_sys_io_uring_register, s->ring_fd,
 121                         IORING_REGISTER_FILES, s->fds, s->nr_files);
 122 }
 123
 124 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 125 {
 126         return syscall(__NR_sys_io_uring_setup, entries, p);
 127 }
 128
 129 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 130                           unsigned int min_complete, unsigned int flags)
 131 {
 132         return syscall(__NR_sys_io_uring_enter, s->ring_fd, to_submit,
 133                         min_complete, flags, NULL, 0);
 134 }
 135
 136 static int gettid(void)
 137 {
 138         return syscall(__NR_gettid);
 139 }
 140
 141 static unsigned file_depth(struct submitter *s)
 142 {
 143         return (depth + s->nr_files - 1) / s->nr_files;
 144 }
 145
 146 static void init_io(struct submitter *s, unsigned index)
 147 {
 148         struct io_uring_sqe *sqe = &s->sqes[index];
 149         unsigned long offset;
 150         struct file *f;
 151         long r;
 152
 153         if (do_nop) {
 154                 sqe->opcode = IORING_OP_NOP;
 155                 return;
 156         }
 157
 158         if (s->nr_files == 1) {
 159                 f = &s->files[0];
 160         } else {
 161                 f = &s->files[s->cur_file];
 162                 if (f->pending_ios >= file_depth(s)) {
 163                         s->cur_file++;
 164                         if (s->cur_file == s->nr_files)
 165                                 s->cur_file = 0;
 166                         f = &s->files[s->cur_file];
 167                 }
 168         }
 169         f->pending_ios++;
 170
 171         lrand48_r(&s->rand, &r);
 172         offset = (r % (f->max_blocks - 1)) * BS;
 173
 174         if (register_files) {
 175                 sqe->flags = IOSQE_FIXED_FILE;
 176                 sqe->fd = f->fixed_fd;
 177         } else {
 178                 sqe->flags = 0;
 179                 sqe->fd = f->real_fd;
 180         }
 181         if (fixedbufs) {
 182                 sqe->opcode = IORING_OP_READ_FIXED;
 183                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 184                 sqe->len = BS;
 185                 sqe->buf_index = index;
 186         } else {
 187                 sqe->opcode = IORING_OP_READV;
 188                 sqe->addr = (unsigned long) &s->iovecs[index];
 189                 sqe->len = 1;
 190                 sqe->buf_index = 0;
 191         }
 192         sqe->ioprio = 0;
 193         sqe->off = offset;
 194         sqe->user_data = (unsigned long) f;
 195 }
 196
 197 static int prep_more_ios(struct submitter *s, int max_ios)
 198 {
 199         struct io_sq_ring *ring = &s->sq_ring;
 200         unsigned index, tail, next_tail, prepped = 0;
 201
 202         next_tail = tail = *ring->tail;
 203         do {
 204                 next_tail++;
 205                 read_barrier();
 206                 if (next_tail == *ring->head)
 207                         break;
 208
 209                 index = tail & sq_ring_mask;
 210                 init_io(s, index);
 211                 ring->array[index] = index;
 212                 prepped++;
 213                 tail = next_tail;
 214         } while (prepped < max_ios);
 215
 216         if (*ring->tail != tail) {
 217                 /* order tail store with writes to sqes above */
 218                 write_barrier();
 219                 *ring->tail = tail;
 220                 write_barrier();
 221         }
 222         return prepped;
 223 }
 224
 225 static int get_file_size(struct file *f)
 226 {
 227         struct stat st;
 228
 229         if (fstat(f->real_fd, &st) < 0)
 230                 return -1;
 231         if (S_ISBLK(st.st_mode)) {
 232                 unsigned long long bytes;
 233
 234                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 235                         return -1;
 236
 237                 f->max_blocks = bytes / BS;
 238                 return 0;
 239         } else if (S_ISREG(st.st_mode)) {
 240                 f->max_blocks = st.st_size / BS;
 241                 return 0;
 242         }
 243
 244         return -1;
 245 }
 246
 247 static int reap_events(struct submitter *s)
 248 {
 249         struct io_cq_ring *ring = &s->cq_ring;
 250         struct io_uring_cqe *cqe;
 251         unsigned head, reaped = 0;
 252
 253         head = *ring->head;
 254         do {
 255                 struct file *f;
 256
 257                 read_barrier();
 258                 if (head == *ring->tail)
 259                         break;
 260                 cqe = &ring->cqes[head & cq_ring_mask];
 261                 if (!do_nop) {
 262                         f = (struct file *) (uintptr_t) cqe->user_data;
 263                         f->pending_ios--;
 264                         if (cqe->res != BS) {
 265                                 printf("io: unexpected ret=%d\n", cqe->res);
 266                                 if (polled && cqe->res == -EOPNOTSUPP)
 267                                         printf("Your filesystem/device doesn't "
 268                                                 "support polled IO\n");
 269                                 return -1;
 270                         }
 271                 }
 272                 reaped++;
 273                 head++;
 274         } while (1);
 275
 276         s->inflight -= reaped;
 277         *ring->head = head;
 278         write_barrier();
 279         return reaped;
 280 }
 281
 282 static void *submitter_fn(void *data)
 283 {
 284         struct submitter *s = data;
 285         struct io_sq_ring *ring = &s->sq_ring;
 286         int ret, prepped;
 287
 288         printf("submitter=%d\n", gettid());
 289
 290         srand48_r(pthread_self(), &s->rand);
 291
 292         prepped = 0;
 293         do {
 294                 int to_wait, to_submit, this_reap, to_prep;
 295
 296                 if (!prepped && s->inflight < depth) {
 297                         to_prep = min(depth - s->inflight, batch_submit);
 298                         prepped = prep_more_ios(s, to_prep);
 299                 }
 300                 s->inflight += prepped;
 301 submit_more:
 302                 to_submit = prepped;
 303 submit:
 304                 if (to_submit && (s->inflight + to_submit <= depth))
 305                         to_wait = 0;
 306                 else
 307                         to_wait = min(s->inflight + to_submit, batch_complete);
 308
 309                 /*
 310                  * Only need to call io_uring_enter if we're not using SQ thread
 311                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
 312                  */
 313                 if (!sq_thread_poll || (*ring->flags & IORING_SQ_NEED_WAKEUP)) {
 314                         unsigned flags = 0;
 315
 316                         if (to_wait)
 317                                 flags = IORING_ENTER_GETEVENTS;
 318                         if ((*ring->flags & IORING_SQ_NEED_WAKEUP))
 319                                 flags |= IORING_ENTER_SQ_WAKEUP;
 320                         ret = io_uring_enter(s, to_submit, to_wait, flags);
 321                         s->calls++;
 322                 }
 323
 324                 /*
 325                  * For non SQ thread poll, we already got the events we needed
 326                  * through the io_uring_enter() above. For SQ thread poll, we
 327                  * need to loop here until we find enough events.
 328                  */
 329                 this_reap = 0;
 330                 do {
 331                         int r;
 332                         r = reap_events(s);
 333                         if (r == -1) {
 334                                 s->finish = 1;
 335                                 break;
 336                         } else if (r > 0)
 337                                 this_reap += r;
 338                 } while (sq_thread_poll && this_reap < to_wait);
 339                 s->reaps += this_reap;
 340
 341                 if (ret >= 0) {
 342                         if (!ret) {
 343                                 to_submit = 0;
 344                                 if (s->inflight)
 345                                         goto submit;
 346                                 continue;
 347                         } else if (ret < to_submit) {
 348                                 int diff = to_submit - ret;
 349
 350                                 s->done += ret;
 351                                 prepped -= diff;
 352                                 goto submit_more;
 353                         }
 354                         s->done += ret;
 355                         prepped = 0;
 356                         continue;
 357                 } else if (ret < 0) {
 358                         if (errno == EAGAIN) {
 359                                 if (s->finish)
 360                                         break;
 361                                 if (this_reap)
 362                                         goto submit;
 363                                 to_submit = 0;
 364                                 goto submit;
 365                         }
 366                         printf("io_submit: %s\n", strerror(errno));
 367                         break;
 368                 }
 369         } while (!s->finish);
 370
 371         finish = 1;
 372         return NULL;
 373 }
 374
 375 static void sig_int(int sig)
 376 {
 377         printf("Exiting on signal %d\n", sig);
 378         submitter->finish = 1;
 379         finish = 1;
 380 }
 381
 382 static void arm_sig_int(void)
 383 {
 384         struct sigaction act;
 385
 386         memset(&act, 0, sizeof(act));
 387         act.sa_handler = sig_int;
 388         act.sa_flags = SA_RESTART;
 389         sigaction(SIGINT, &act, NULL);
 390 }
 391
 392 static int setup_ring(struct submitter *s)
 393 {
 394         struct io_sq_ring *sring = &s->sq_ring;
 395         struct io_cq_ring *cring = &s->cq_ring;
 396         struct io_uring_params p;
 397         int ret, fd;
 398         void *ptr;
 399
 400         memset(&p, 0, sizeof(p));
 401
 402         if (polled && !do_nop)
 403                 p.flags |= IORING_SETUP_IOPOLL;
 404         if (sq_thread_poll) {
 405                 p.flags |= IORING_SETUP_SQPOLL;
 406                 if (sq_thread_cpu != -1) {
 407                         p.flags |= IORING_SETUP_SQ_AFF;
 408                         p.sq_thread_cpu = sq_thread_cpu;
 409                 }
 410         }
 411
 412         fd = io_uring_setup(depth, &p);
 413         if (fd < 0) {
 414                 perror("io_uring_setup");
 415                 return 1;
 416         }
 417         s->ring_fd = fd;
 418
 419         if (fixedbufs) {
 420                 ret = io_uring_register_buffers(s);
 421                 if (ret < 0) {
 422                         perror("io_uring_register_buffers");
 423                         return 1;
 424                 }
 425         }
 426
 427         if (register_files) {
 428                 ret = io_uring_register_files(s);
 429                 if (ret < 0) {
 430                         perror("io_uring_register_files");
 431                         return 1;
 432                 }
 433         }
 434
 435         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 436                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 437                         IORING_OFF_SQ_RING);
 438         printf("sq_ring ptr = 0x%p\n", ptr);
 439         sring->head = ptr + p.sq_off.head;
 440         sring->tail = ptr + p.sq_off.tail;
 441         sring->ring_mask = ptr + p.sq_off.ring_mask;
 442         sring->ring_entries = ptr + p.sq_off.ring_entries;
 443         sring->flags = ptr + p.sq_off.flags;
 444         sring->array = ptr + p.sq_off.array;
 445         sq_ring_mask = *sring->ring_mask;
 446
 447         s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
 448                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 449                         IORING_OFF_SQES);
 450         printf("sqes ptr    = 0x%p\n", s->sqes);
 451
 452         ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
 453                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 454                         IORING_OFF_CQ_RING);
 455         printf("cq_ring ptr = 0x%p\n", ptr);
 456         cring->head = ptr + p.cq_off.head;
 457         cring->tail = ptr + p.cq_off.tail;
 458         cring->ring_mask = ptr + p.cq_off.ring_mask;
 459         cring->ring_entries = ptr + p.cq_off.ring_entries;
 460         cring->cqes = ptr + p.cq_off.cqes;
 461         cq_ring_mask = *cring->ring_mask;
 462         return 0;
 463 }
 464
 465 static void file_depths(char *buf)
 466 {
 467         struct submitter *s = submitter;
 468         char *p;
 469         int i;
 470
 471         buf[0] = '\0';
 472         p = buf;
 473         for (i = 0; i < s->nr_files; i++) {
 474                 struct file *f = &s->files[i];
 475
 476                 if (i + 1 == s->nr_files)
 477                         p += sprintf(p, "%d", f->pending_ios);
 478                 else
 479                         p += sprintf(p, "%d, ", f->pending_ios);
 480         }
 481 }
 482
 483 static void usage(char *argv)
 484 {
 485         printf("%s [options] -- [filenames]\n"
 486                 " -d <int> : IO Depth, default %d\n"
 487                 " -s <int> : Batch submit, default %d\n"
 488                 " -c <int> : Batch complete, default %d\n",
 489                 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE);
 490         exit(0);
 491 }
 492
 493 int main(int argc, char *argv[])
 494 {
 495         struct submitter *s;
 496         unsigned long done, calls, reap;
 497         int err, i, flags, fd, opt;
 498         char *fdepths;
 499         void *ret;
 500
 501         if (!do_nop && argc < 2) {
 502                 printf("%s: filename [options]\n", argv[0]);
 503                 return 1;
 504         }
 505
 506         while ((opt = getopt(argc, argv, "d:s:c:h?")) != -1) {
 507                 switch (opt) {
 508                 case 'd':
 509                         depth = atoi(optarg);
 510                         break;
 511                 case 's':
 512                         batch_submit = atoi(optarg);
 513                         break;
 514                 case 'c':
 515                         batch_complete = atoi(optarg);
 516                         break;
 517                 case 'h':
 518                 case '?':
 519                 default:
 520                         usage(argv[0]);
 521                         break;
 522                 }
 523         }
 524
 525         submitter = malloc(sizeof(*submitter) + depth * sizeof(struct iovec));
 526         memset(submitter, 0, sizeof(*submitter) + depth * sizeof(struct iovec));
 527         s = submitter;
 528
 529         flags = O_RDONLY | O_NOATIME;
 530         if (!buffered)
 531                 flags |= O_DIRECT;
 532
 533         i = optind;
 534         while (!do_nop && i < argc) {
 535                 struct file *f;
 536
 537                 if (s->nr_files == MAX_FDS) {
 538                         printf("Max number of files (%d) reached\n", MAX_FDS);
 539                         break;
 540                 }
 541                 fd = open(argv[i], flags);
 542                 if (fd < 0) {
 543                         perror("open");
 544                         return 1;
 545                 }
 546
 547                 f = &s->files[s->nr_files];
 548                 f->real_fd = fd;
 549                 if (get_file_size(f)) {
 550                         printf("failed getting size of device/file\n");
 551                         return 1;
 552                 }
 553                 if (f->max_blocks <= 1) {
 554                         printf("Zero file/device size?\n");
 555                         return 1;
 556                 }
 557                 f->max_blocks--;
 558
 559                 printf("Added file %s\n", argv[i]);
 560                 s->nr_files++;
 561                 i++;
 562         }
 563
 564         if (fixedbufs) {
 565                 struct rlimit rlim;
 566
 567                 rlim.rlim_cur = RLIM_INFINITY;
 568                 rlim.rlim_max = RLIM_INFINITY;
 569                 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 570                         perror("setrlimit");
 571                         return 1;
 572                 }
 573         }
 574
 575         arm_sig_int();
 576
 577         for (i = 0; i < depth; i++) {
 578                 void *buf;
 579
 580                 if (posix_memalign(&buf, BS, BS)) {
 581                         printf("failed alloc\n");
 582                         return 1;
 583                 }
 584                 s->iovecs[i].iov_base = buf;
 585                 s->iovecs[i].iov_len = BS;
 586         }
 587
 588         err = setup_ring(s);
 589         if (err) {
 590                 printf("ring setup failed: %s, %d\n", strerror(errno), err);
 591                 return 1;
 592         }
 593         printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered);
 594         printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", depth, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
 595
 596         pthread_create(&s->thread, NULL, submitter_fn, s);
 597
 598         fdepths = malloc(8 * s->nr_files);
 599         reap = calls = done = 0;
 600         do {
 601                 unsigned long this_done = 0;
 602                 unsigned long this_reap = 0;
 603                 unsigned long this_call = 0;
 604                 unsigned long rpc = 0, ipc = 0;
 605
 606                 sleep(1);
 607                 this_done += s->done;
 608                 this_call += s->calls;
 609                 this_reap += s->reaps;
 610                 if (this_call - calls) {
 611                         rpc = (this_done - done) / (this_call - calls);
 612                         ipc = (this_reap - reap) / (this_call - calls);
 613                 } else
 614                         rpc = ipc = -1;
 615                 file_depths(fdepths);
 616                 printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s)\n",
 617                                 this_done - done, rpc, ipc, s->inflight,
 618                                 fdepths);
 619                 done = this_done;
 620                 calls = this_call;
 621                 reap = this_reap;
 622         } while (!finish);
 623
 624         pthread_join(s->thread, &ret);
 625         close(s->ring_fd);
 626         free(fdepths);
 627         return 0;
 628 }