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_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_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_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_io_uring_enter, s->ring_fd, to_submit, min_complete,
 133                         flags, NULL, 0);
 134 }
 135
 136 #ifndef CONFIG_HAVE_GETTID
 137 static int gettid(void)
 138 {
 139         return syscall(__NR_gettid);
 140 }
 141 #endif
 142
 143 static unsigned file_depth(struct submitter *s)
 144 {
 145         return (depth + s->nr_files - 1) / s->nr_files;
 146 }
 147
 148 static void init_io(struct submitter *s, unsigned index)
 149 {
 150         struct io_uring_sqe *sqe = &s->sqes[index];
 151         unsigned long offset;
 152         struct file *f;
 153         long r;
 154
 155         if (do_nop) {
 156                 sqe->opcode = IORING_OP_NOP;
 157                 return;
 158         }
 159
 160         if (s->nr_files == 1) {
 161                 f = &s->files[0];
 162         } else {
 163                 f = &s->files[s->cur_file];
 164                 if (f->pending_ios >= file_depth(s)) {
 165                         s->cur_file++;
 166                         if (s->cur_file == s->nr_files)
 167                                 s->cur_file = 0;
 168                         f = &s->files[s->cur_file];
 169                 }
 170         }
 171         f->pending_ios++;
 172
 173         lrand48_r(&s->rand, &r);
 174         offset = (r % (f->max_blocks - 1)) * BS;
 175
 176         if (register_files) {
 177                 sqe->flags = IOSQE_FIXED_FILE;
 178                 sqe->fd = f->fixed_fd;
 179         } else {
 180                 sqe->flags = 0;
 181                 sqe->fd = f->real_fd;
 182         }
 183         if (fixedbufs) {
 184                 sqe->opcode = IORING_OP_READ_FIXED;
 185                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 186                 sqe->len = BS;
 187                 sqe->buf_index = index;
 188         } else {
 189                 sqe->opcode = IORING_OP_READV;
 190                 sqe->addr = (unsigned long) &s->iovecs[index];
 191                 sqe->len = 1;
 192                 sqe->buf_index = 0;
 193         }
 194         sqe->ioprio = 0;
 195         sqe->off = offset;
 196         sqe->user_data = (unsigned long) f;
 197 }
 198
 199 static int prep_more_ios(struct submitter *s, int max_ios)
 200 {
 201         struct io_sq_ring *ring = &s->sq_ring;
 202         unsigned index, tail, next_tail, prepped = 0;
 203
 204         next_tail = tail = *ring->tail;
 205         do {
 206                 next_tail++;
 207                 read_barrier();
 208                 if (next_tail == *ring->head)
 209                         break;
 210
 211                 index = tail & sq_ring_mask;
 212                 init_io(s, index);
 213                 ring->array[index] = index;
 214                 prepped++;
 215                 tail = next_tail;
 216         } while (prepped < max_ios);
 217
 218         if (*ring->tail != tail) {
 219                 /* order tail store with writes to sqes above */
 220                 write_barrier();
 221                 *ring->tail = tail;
 222                 write_barrier();
 223         }
 224         return prepped;
 225 }
 226
 227 static int get_file_size(struct file *f)
 228 {
 229         struct stat st;
 230
 231         if (fstat(f->real_fd, &st) < 0)
 232                 return -1;
 233         if (S_ISBLK(st.st_mode)) {
 234                 unsigned long long bytes;
 235
 236                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 237                         return -1;
 238
 239                 f->max_blocks = bytes / BS;
 240                 return 0;
 241         } else if (S_ISREG(st.st_mode)) {
 242                 f->max_blocks = st.st_size / BS;
 243                 return 0;
 244         }
 245
 246         return -1;
 247 }
 248
 249 static int reap_events(struct submitter *s)
 250 {
 251         struct io_cq_ring *ring = &s->cq_ring;
 252         struct io_uring_cqe *cqe;
 253         unsigned head, reaped = 0;
 254
 255         head = *ring->head;
 256         do {
 257                 struct file *f;
 258
 259                 read_barrier();
 260                 if (head == *ring->tail)
 261                         break;
 262                 cqe = &ring->cqes[head & cq_ring_mask];
 263                 if (!do_nop) {
 264                         f = (struct file *) (uintptr_t) cqe->user_data;
 265                         f->pending_ios--;
 266                         if (cqe->res != BS) {
 267                                 printf("io: unexpected ret=%d\n", cqe->res);
 268                                 if (polled && cqe->res == -EOPNOTSUPP)
 269                                         printf("Your filesystem/driver/kernel doesn't support polled IO\n");
 270                                 return -1;
 271                         }
 272                 }
 273                 reaped++;
 274                 head++;
 275         } while (1);
 276
 277         s->inflight -= reaped;
 278         *ring->head = head;
 279         write_barrier();
 280         return reaped;
 281 }
 282
 283 static void *submitter_fn(void *data)
 284 {
 285         struct submitter *s = data;
 286         struct io_sq_ring *ring = &s->sq_ring;
 287         int ret, prepped;
 288
 289         printf("submitter=%d\n", gettid());
 290
 291         srand48_r(pthread_self(), &s->rand);
 292
 293         prepped = 0;
 294         do {
 295                 int to_wait, to_submit, this_reap, to_prep;
 296
 297                 if (!prepped && s->inflight < depth) {
 298                         to_prep = min(depth - s->inflight, batch_submit);
 299                         prepped = prep_more_ios(s, to_prep);
 300                 }
 301                 s->inflight += prepped;
 302 submit_more:
 303                 to_submit = prepped;
 304 submit:
 305                 if (to_submit && (s->inflight + to_submit <= depth))
 306                         to_wait = 0;
 307                 else
 308                         to_wait = min(s->inflight + to_submit, batch_complete);
 309
 310                 /*
 311                  * Only need to call io_uring_enter if we're not using SQ thread
 312                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
 313                  */
 314                 if (!sq_thread_poll || (*ring->flags & IORING_SQ_NEED_WAKEUP)) {
 315                         unsigned flags = 0;
 316
 317                         if (to_wait)
 318                                 flags = IORING_ENTER_GETEVENTS;
 319                         if ((*ring->flags & IORING_SQ_NEED_WAKEUP))
 320                                 flags |= IORING_ENTER_SQ_WAKEUP;
 321                         ret = io_uring_enter(s, to_submit, to_wait, flags);
 322                         s->calls++;
 323                 }
 324
 325                 /*
 326                  * For non SQ thread poll, we already got the events we needed
 327                  * through the io_uring_enter() above. For SQ thread poll, we
 328                  * need to loop here until we find enough events.
 329                  */
 330                 this_reap = 0;
 331                 do {
 332                         int r;
 333                         r = reap_events(s);
 334                         if (r == -1) {
 335                                 s->finish = 1;
 336                                 break;
 337                         } else if (r > 0)
 338                                 this_reap += r;
 339                 } while (sq_thread_poll && this_reap < to_wait);
 340                 s->reaps += this_reap;
 341
 342                 if (ret >= 0) {
 343                         if (!ret) {
 344                                 to_submit = 0;
 345                                 if (s->inflight)
 346                                         goto submit;
 347                                 continue;
 348                         } else if (ret < to_submit) {
 349                                 int diff = to_submit - ret;
 350
 351                                 s->done += ret;
 352                                 prepped -= diff;
 353                                 goto submit_more;
 354                         }
 355                         s->done += ret;
 356                         prepped = 0;
 357                         continue;
 358                 } else if (ret < 0) {
 359                         if (errno == EAGAIN) {
 360                                 if (s->finish)
 361                                         break;
 362                                 if (this_reap)
 363                                         goto submit;
 364                                 to_submit = 0;
 365                                 goto submit;
 366                         }
 367                         printf("io_submit: %s\n", strerror(errno));
 368                         break;
 369                 }
 370         } while (!s->finish);
 371
 372         finish = 1;
 373         return NULL;
 374 }
 375
 376 static void sig_int(int sig)
 377 {
 378         printf("Exiting on signal %d\n", sig);
 379         submitter->finish = 1;
 380         finish = 1;
 381 }
 382
 383 static void arm_sig_int(void)
 384 {
 385         struct sigaction act;
 386
 387         memset(&act, 0, sizeof(act));
 388         act.sa_handler = sig_int;
 389         act.sa_flags = SA_RESTART;
 390         sigaction(SIGINT, &act, NULL);
 391 }
 392
 393 static int setup_ring(struct submitter *s)
 394 {
 395         struct io_sq_ring *sring = &s->sq_ring;
 396         struct io_cq_ring *cring = &s->cq_ring;
 397         struct io_uring_params p;
 398         int ret, fd;
 399         void *ptr;
 400
 401         memset(&p, 0, sizeof(p));
 402
 403         if (polled && !do_nop)
 404                 p.flags |= IORING_SETUP_IOPOLL;
 405         if (sq_thread_poll) {
 406                 p.flags |= IORING_SETUP_SQPOLL;
 407                 if (sq_thread_cpu != -1) {
 408                         p.flags |= IORING_SETUP_SQ_AFF;
 409                         p.sq_thread_cpu = sq_thread_cpu;
 410                 }
 411         }
 412
 413         fd = io_uring_setup(depth, &p);
 414         if (fd < 0) {
 415                 perror("io_uring_setup");
 416                 return 1;
 417         }
 418         s->ring_fd = fd;
 419
 420         if (fixedbufs) {
 421                 ret = io_uring_register_buffers(s);
 422                 if (ret < 0) {
 423                         perror("io_uring_register_buffers");
 424                         return 1;
 425                 }
 426         }
 427
 428         if (register_files) {
 429                 ret = io_uring_register_files(s);
 430                 if (ret < 0) {
 431                         perror("io_uring_register_files");
 432                         return 1;
 433                 }
 434         }
 435
 436         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 437                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 438                         IORING_OFF_SQ_RING);
 439         printf("sq_ring ptr = 0x%p\n", ptr);
 440         sring->head = ptr + p.sq_off.head;
 441         sring->tail = ptr + p.sq_off.tail;
 442         sring->ring_mask = ptr + p.sq_off.ring_mask;
 443         sring->ring_entries = ptr + p.sq_off.ring_entries;
 444         sring->flags = ptr + p.sq_off.flags;
 445         sring->array = ptr + p.sq_off.array;
 446         sq_ring_mask = *sring->ring_mask;
 447
 448         s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
 449                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 450                         IORING_OFF_SQES);
 451         printf("sqes ptr    = 0x%p\n", s->sqes);
 452
 453         ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
 454                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 455                         IORING_OFF_CQ_RING);
 456         printf("cq_ring ptr = 0x%p\n", ptr);
 457         cring->head = ptr + p.cq_off.head;
 458         cring->tail = ptr + p.cq_off.tail;
 459         cring->ring_mask = ptr + p.cq_off.ring_mask;
 460         cring->ring_entries = ptr + p.cq_off.ring_entries;
 461         cring->cqes = ptr + p.cq_off.cqes;
 462         cq_ring_mask = *cring->ring_mask;
 463         return 0;
 464 }
 465
 466 static void file_depths(char *buf)
 467 {
 468         struct submitter *s = submitter;
 469         char *p;
 470         int i;
 471
 472         buf[0] = '\0';
 473         p = buf;
 474         for (i = 0; i < s->nr_files; i++) {
 475                 struct file *f = &s->files[i];
 476
 477                 if (i + 1 == s->nr_files)
 478                         p += sprintf(p, "%d", f->pending_ios);
 479                 else
 480                         p += sprintf(p, "%d, ", f->pending_ios);
 481         }
 482 }
 483
 484 static void usage(char *argv)
 485 {
 486         printf("%s [options] -- [filenames]\n"
 487                 " -d <int> : IO Depth, default %d\n"
 488                 " -s <int> : Batch submit, default %d\n"
 489                 " -c <int> : Batch complete, default %d\n",
 490                 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE);
 491         exit(0);
 492 }
 493
 494 int main(int argc, char *argv[])
 495 {
 496         struct submitter *s;
 497         unsigned long done, calls, reap;
 498         int err, i, flags, fd, opt;
 499         char *fdepths;
 500         void *ret;
 501
 502         if (!do_nop && argc < 2) {
 503                 printf("%s: filename [options]\n", argv[0]);
 504                 return 1;
 505         }
 506
 507         while ((opt = getopt(argc, argv, "d:s:c:h?")) != -1) {
 508                 switch (opt) {
 509                 case 'd':
 510                         depth = atoi(optarg);
 511                         break;
 512                 case 's':
 513                         batch_submit = atoi(optarg);
 514                         break;
 515                 case 'c':
 516                         batch_complete = atoi(optarg);
 517                         break;
 518                 case 'h':
 519                 case '?':
 520                 default:
 521                         usage(argv[0]);
 522                         break;
 523                 }
 524         }
 525
 526         submitter = malloc(sizeof(*submitter) + depth * sizeof(struct iovec));
 527         memset(submitter, 0, sizeof(*submitter) + depth * sizeof(struct iovec));
 528         s = submitter;
 529
 530         flags = O_RDONLY | O_NOATIME;
 531         if (!buffered)
 532                 flags |= O_DIRECT;
 533
 534         i = optind;
 535         while (!do_nop && i < argc) {
 536                 struct file *f;
 537
 538                 if (s->nr_files == MAX_FDS) {
 539                         printf("Max number of files (%d) reached\n", MAX_FDS);
 540                         break;
 541                 }
 542                 fd = open(argv[i], flags);
 543                 if (fd < 0) {
 544                         perror("open");
 545                         return 1;
 546                 }
 547
 548                 f = &s->files[s->nr_files];
 549                 f->real_fd = fd;
 550                 if (get_file_size(f)) {
 551                         printf("failed getting size of device/file\n");
 552                         return 1;
 553                 }
 554                 if (f->max_blocks <= 1) {
 555                         printf("Zero file/device size?\n");
 556                         return 1;
 557                 }
 558                 f->max_blocks--;
 559
 560                 printf("Added file %s\n", argv[i]);
 561                 s->nr_files++;
 562                 i++;
 563         }
 564
 565         if (fixedbufs) {
 566                 struct rlimit rlim;
 567
 568                 rlim.rlim_cur = RLIM_INFINITY;
 569                 rlim.rlim_max = RLIM_INFINITY;
 570                 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 571                         perror("setrlimit");
 572                         return 1;
 573                 }
 574         }
 575
 576         arm_sig_int();
 577
 578         for (i = 0; i < depth; i++) {
 579                 void *buf;
 580
 581                 if (posix_memalign(&buf, BS, BS)) {
 582                         printf("failed alloc\n");
 583                         return 1;
 584                 }
 585                 s->iovecs[i].iov_base = buf;
 586                 s->iovecs[i].iov_len = BS;
 587         }
 588
 589         err = setup_ring(s);
 590         if (err) {
 591                 printf("ring setup failed: %s, %d\n", strerror(errno), err);
 592                 return 1;
 593         }
 594         printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered);
 595         printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", depth, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
 596
 597         pthread_create(&s->thread, NULL, submitter_fn, s);
 598
 599         fdepths = malloc(8 * s->nr_files);
 600         reap = calls = done = 0;
 601         do {
 602                 unsigned long this_done = 0;
 603                 unsigned long this_reap = 0;
 604                 unsigned long this_call = 0;
 605                 unsigned long rpc = 0, ipc = 0;
 606
 607                 sleep(1);
 608                 this_done += s->done;
 609                 this_call += s->calls;
 610                 this_reap += s->reaps;
 611                 if (this_call - calls) {
 612                         rpc = (this_done - done) / (this_call - calls);
 613                         ipc = (this_reap - reap) / (this_call - calls);
 614                 } else
 615                         rpc = ipc = -1;
 616                 file_depths(fdepths);
 617                 printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s)\n",
 618                                 this_done - done, rpc, ipc, s->inflight,
 619                                 fdepths);
 620                 done = this_done;
 621                 calls = this_call;
 622                 reap = this_reap;
 623         } while (!finish);
 624
 625         pthread_join(s->thread, &ret);
 626         close(s->ring_fd);
 627         free(fdepths);
 628         return 0;
 629 }