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 #define BS                      4096
  50
  51 #define MAX_FDS                 16
  52
  53 static unsigned sq_ring_mask, cq_ring_mask;
  54
  55 struct file {
  56         unsigned long max_blocks;
  57         unsigned pending_ios;
  58         int real_fd;
  59         int fixed_fd;
  60 };
  61
  62 struct submitter {
  63         pthread_t thread;
  64         int ring_fd;
  65         struct io_sq_ring sq_ring;
  66         struct io_uring_sqe *sqes;
  67         struct io_cq_ring cq_ring;
  68         int inflight;
  69         unsigned long reaps;
  70         unsigned long done;
  71         unsigned long calls;
  72         volatile int finish;
  73
  74         __s32 *fds;
  75
  76         struct file files[MAX_FDS];
  77         unsigned nr_files;
  78         unsigned cur_file;
  79         struct iovec iovecs[];
  80 };
  81
  82 static struct submitter *submitter;
  83 static volatile int finish;
  84
  85 static int depth = DEPTH;
  86 static int batch_submit = BATCH_SUBMIT;
  87 static int batch_complete = BATCH_COMPLETE;
  88 static int bs = BS;
  89 static int polled = 1;          /* use IO polling */
  90 static int fixedbufs = 1;       /* use fixed user buffers */
  91 static int register_files = 1;  /* use fixed files */
  92 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
  93 static int sq_thread_poll = 0;  /* use kernel submission/poller thread */
  94 static int sq_thread_cpu = -1;  /* pin above thread to this CPU */
  95 static int do_nop = 0;          /* no-op SQ ring commands */
  96
  97 static int vectored = 1;
  98
  99 static int io_uring_register_buffers(struct submitter *s)
 100 {
 101         if (do_nop)
 102                 return 0;
 103
 104         return syscall(__NR_io_uring_register, s->ring_fd,
 105                         IORING_REGISTER_BUFFERS, s->iovecs, depth);
 106 }
 107
 108 static int io_uring_register_files(struct submitter *s)
 109 {
 110         int i;
 111
 112         if (do_nop)
 113                 return 0;
 114
 115         s->fds = calloc(s->nr_files, sizeof(__s32));
 116         for (i = 0; i < s->nr_files; i++) {
 117                 s->fds[i] = s->files[i].real_fd;
 118                 s->files[i].fixed_fd = i;
 119         }
 120
 121         return syscall(__NR_io_uring_register, s->ring_fd,
 122                         IORING_REGISTER_FILES, s->fds, s->nr_files);
 123 }
 124
 125 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 126 {
 127         return syscall(__NR_io_uring_setup, entries, p);
 128 }
 129
 130 static void io_uring_probe(int fd)
 131 {
 132         struct io_uring_probe *p;
 133         int ret;
 134
 135         p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 136         if (!p)
 137                 return;
 138
 139         memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 140         ret = syscall(__NR_io_uring_register, fd, IORING_REGISTER_PROBE, p, 256);
 141         if (ret < 0)
 142                 goto out;
 143
 144         if (IORING_OP_READ > p->ops_len)
 145                 goto out;
 146
 147         if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED))
 148                 vectored = 0;
 149 out:
 150         free(p);
 151 }
 152
 153 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 154                           unsigned int min_complete, unsigned int flags)
 155 {
 156         return syscall(__NR_io_uring_enter, s->ring_fd, to_submit, min_complete,
 157                         flags, NULL, 0);
 158 }
 159
 160 #ifndef CONFIG_HAVE_GETTID
 161 static int gettid(void)
 162 {
 163         return syscall(__NR_gettid);
 164 }
 165 #endif
 166
 167 static unsigned file_depth(struct submitter *s)
 168 {
 169         return (depth + s->nr_files - 1) / s->nr_files;
 170 }
 171
 172 static void init_io(struct submitter *s, unsigned index)
 173 {
 174         struct io_uring_sqe *sqe = &s->sqes[index];
 175         unsigned long offset;
 176         struct file *f;
 177         long r;
 178
 179         if (do_nop) {
 180                 sqe->opcode = IORING_OP_NOP;
 181                 return;
 182         }
 183
 184         if (s->nr_files == 1) {
 185                 f = &s->files[0];
 186         } else {
 187                 f = &s->files[s->cur_file];
 188                 if (f->pending_ios >= file_depth(s)) {
 189                         s->cur_file++;
 190                         if (s->cur_file == s->nr_files)
 191                                 s->cur_file = 0;
 192                         f = &s->files[s->cur_file];
 193                 }
 194         }
 195         f->pending_ios++;
 196
 197         r = lrand48();
 198         offset = (r % (f->max_blocks - 1)) * bs;
 199
 200         if (register_files) {
 201                 sqe->flags = IOSQE_FIXED_FILE;
 202                 sqe->fd = f->fixed_fd;
 203         } else {
 204                 sqe->flags = 0;
 205                 sqe->fd = f->real_fd;
 206         }
 207         if (fixedbufs) {
 208                 sqe->opcode = IORING_OP_READ_FIXED;
 209                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 210                 sqe->len = bs;
 211                 sqe->buf_index = index;
 212         } else if (!vectored) {
 213                 sqe->opcode = IORING_OP_READ;
 214                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 215                 sqe->len = bs;
 216                 sqe->buf_index = 0;
 217         } else {
 218                 sqe->opcode = IORING_OP_READV;
 219                 sqe->addr = (unsigned long) &s->iovecs[index];
 220                 sqe->len = 1;
 221                 sqe->buf_index = 0;
 222         }
 223         sqe->ioprio = 0;
 224         sqe->off = offset;
 225         sqe->user_data = (unsigned long) f;
 226 }
 227
 228 static int prep_more_ios(struct submitter *s, int max_ios)
 229 {
 230         struct io_sq_ring *ring = &s->sq_ring;
 231         unsigned index, tail, next_tail, prepped = 0;
 232
 233         next_tail = tail = *ring->tail;
 234         do {
 235                 next_tail++;
 236                 read_barrier();
 237                 if (next_tail == *ring->head)
 238                         break;
 239
 240                 index = tail & sq_ring_mask;
 241                 init_io(s, index);
 242                 ring->array[index] = index;
 243                 prepped++;
 244                 tail = next_tail;
 245         } while (prepped < max_ios);
 246
 247         if (*ring->tail != tail) {
 248                 *ring->tail = tail;
 249                 write_barrier();
 250         }
 251         return prepped;
 252 }
 253
 254 static int get_file_size(struct file *f)
 255 {
 256         struct stat st;
 257
 258         if (fstat(f->real_fd, &st) < 0)
 259                 return -1;
 260         if (S_ISBLK(st.st_mode)) {
 261                 unsigned long long bytes;
 262
 263                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 264                         return -1;
 265
 266                 f->max_blocks = bytes / bs;
 267                 return 0;
 268         } else if (S_ISREG(st.st_mode)) {
 269                 f->max_blocks = st.st_size / bs;
 270                 return 0;
 271         }
 272
 273         return -1;
 274 }
 275
 276 static int reap_events(struct submitter *s)
 277 {
 278         struct io_cq_ring *ring = &s->cq_ring;
 279         struct io_uring_cqe *cqe;
 280         unsigned head, reaped = 0;
 281
 282         head = *ring->head;
 283         do {
 284                 struct file *f;
 285
 286                 read_barrier();
 287                 if (head == *ring->tail)
 288                         break;
 289                 cqe = &ring->cqes[head & cq_ring_mask];
 290                 if (!do_nop) {
 291                         f = (struct file *) (uintptr_t) cqe->user_data;
 292                         f->pending_ios--;
 293                         if (cqe->res != bs) {
 294                                 printf("io: unexpected ret=%d\n", cqe->res);
 295                                 if (polled && cqe->res == -EOPNOTSUPP)
 296                                         printf("Your filesystem/driver/kernel doesn't support polled IO\n");
 297                                 return -1;
 298                         }
 299                 }
 300                 reaped++;
 301                 head++;
 302         } while (1);
 303
 304         s->inflight -= reaped;
 305         *ring->head = head;
 306         write_barrier();
 307         return reaped;
 308 }
 309
 310 static void *submitter_fn(void *data)
 311 {
 312         struct submitter *s = data;
 313         struct io_sq_ring *ring = &s->sq_ring;
 314         int ret, prepped;
 315
 316         printf("submitter=%d\n", gettid());
 317
 318         srand48(pthread_self());
 319
 320         prepped = 0;
 321         do {
 322                 int to_wait, to_submit, this_reap, to_prep;
 323
 324                 if (!prepped && s->inflight < depth) {
 325                         to_prep = min(depth - s->inflight, batch_submit);
 326                         prepped = prep_more_ios(s, to_prep);
 327                 }
 328                 s->inflight += prepped;
 329 submit_more:
 330                 to_submit = prepped;
 331 submit:
 332                 if (to_submit && (s->inflight + to_submit <= depth))
 333                         to_wait = 0;
 334                 else
 335                         to_wait = min(s->inflight + to_submit, batch_complete);
 336
 337                 /*
 338                  * Only need to call io_uring_enter if we're not using SQ thread
 339                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
 340                  */
 341                 if (!sq_thread_poll || (*ring->flags & IORING_SQ_NEED_WAKEUP)) {
 342                         unsigned flags = 0;
 343
 344                         if (to_wait)
 345                                 flags = IORING_ENTER_GETEVENTS;
 346                         if ((*ring->flags & IORING_SQ_NEED_WAKEUP))
 347                                 flags |= IORING_ENTER_SQ_WAKEUP;
 348                         ret = io_uring_enter(s, to_submit, to_wait, flags);
 349                         s->calls++;
 350                 }
 351
 352                 /*
 353                  * For non SQ thread poll, we already got the events we needed
 354                  * through the io_uring_enter() above. For SQ thread poll, we
 355                  * need to loop here until we find enough events.
 356                  */
 357                 this_reap = 0;
 358                 do {
 359                         int r;
 360                         r = reap_events(s);
 361                         if (r == -1) {
 362                                 s->finish = 1;
 363                                 break;
 364                         } else if (r > 0)
 365                                 this_reap += r;
 366                 } while (sq_thread_poll && this_reap < to_wait);
 367                 s->reaps += this_reap;
 368
 369                 if (ret >= 0) {
 370                         if (!ret) {
 371                                 to_submit = 0;
 372                                 if (s->inflight)
 373                                         goto submit;
 374                                 continue;
 375                         } else if (ret < to_submit) {
 376                                 int diff = to_submit - ret;
 377
 378                                 s->done += ret;
 379                                 prepped -= diff;
 380                                 goto submit_more;
 381                         }
 382                         s->done += ret;
 383                         prepped = 0;
 384                         continue;
 385                 } else if (ret < 0) {
 386                         if (errno == EAGAIN) {
 387                                 if (s->finish)
 388                                         break;
 389                                 if (this_reap)
 390                                         goto submit;
 391                                 to_submit = 0;
 392                                 goto submit;
 393                         }
 394                         printf("io_submit: %s\n", strerror(errno));
 395                         break;
 396                 }
 397         } while (!s->finish);
 398
 399         finish = 1;
 400         return NULL;
 401 }
 402
 403 static void sig_int(int sig)
 404 {
 405         printf("Exiting on signal %d\n", sig);
 406         submitter->finish = 1;
 407         finish = 1;
 408 }
 409
 410 static void arm_sig_int(void)
 411 {
 412         struct sigaction act;
 413
 414         memset(&act, 0, sizeof(act));
 415         act.sa_handler = sig_int;
 416         act.sa_flags = SA_RESTART;
 417         sigaction(SIGINT, &act, NULL);
 418 }
 419
 420 static int setup_ring(struct submitter *s)
 421 {
 422         struct io_sq_ring *sring = &s->sq_ring;
 423         struct io_cq_ring *cring = &s->cq_ring;
 424         struct io_uring_params p;
 425         int ret, fd;
 426         void *ptr;
 427
 428         memset(&p, 0, sizeof(p));
 429
 430         if (polled && !do_nop)
 431                 p.flags |= IORING_SETUP_IOPOLL;
 432         if (sq_thread_poll) {
 433                 p.flags |= IORING_SETUP_SQPOLL;
 434                 if (sq_thread_cpu != -1) {
 435                         p.flags |= IORING_SETUP_SQ_AFF;
 436                         p.sq_thread_cpu = sq_thread_cpu;
 437                 }
 438         }
 439
 440         fd = io_uring_setup(depth, &p);
 441         if (fd < 0) {
 442                 perror("io_uring_setup");
 443                 return 1;
 444         }
 445         s->ring_fd = fd;
 446
 447         io_uring_probe(fd);
 448
 449         if (fixedbufs) {
 450                 ret = io_uring_register_buffers(s);
 451                 if (ret < 0) {
 452                         perror("io_uring_register_buffers");
 453                         return 1;
 454                 }
 455         }
 456
 457         if (register_files) {
 458                 ret = io_uring_register_files(s);
 459                 if (ret < 0) {
 460                         perror("io_uring_register_files");
 461                         return 1;
 462                 }
 463         }
 464
 465         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 466                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 467                         IORING_OFF_SQ_RING);
 468         printf("sq_ring ptr = 0x%p\n", ptr);
 469         sring->head = ptr + p.sq_off.head;
 470         sring->tail = ptr + p.sq_off.tail;
 471         sring->ring_mask = ptr + p.sq_off.ring_mask;
 472         sring->ring_entries = ptr + p.sq_off.ring_entries;
 473         sring->flags = ptr + p.sq_off.flags;
 474         sring->array = ptr + p.sq_off.array;
 475         sq_ring_mask = *sring->ring_mask;
 476
 477         s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
 478                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 479                         IORING_OFF_SQES);
 480         printf("sqes ptr    = 0x%p\n", s->sqes);
 481
 482         ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
 483                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 484                         IORING_OFF_CQ_RING);
 485         printf("cq_ring ptr = 0x%p\n", ptr);
 486         cring->head = ptr + p.cq_off.head;
 487         cring->tail = ptr + p.cq_off.tail;
 488         cring->ring_mask = ptr + p.cq_off.ring_mask;
 489         cring->ring_entries = ptr + p.cq_off.ring_entries;
 490         cring->cqes = ptr + p.cq_off.cqes;
 491         cq_ring_mask = *cring->ring_mask;
 492         return 0;
 493 }
 494
 495 static void file_depths(char *buf)
 496 {
 497         struct submitter *s = submitter;
 498         char *p;
 499         int i;
 500
 501         buf[0] = '\0';
 502         p = buf;
 503         for (i = 0; i < s->nr_files; i++) {
 504                 struct file *f = &s->files[i];
 505
 506                 if (i + 1 == s->nr_files)
 507                         p += sprintf(p, "%d", f->pending_ios);
 508                 else
 509                         p += sprintf(p, "%d, ", f->pending_ios);
 510         }
 511 }
 512
 513 static void usage(char *argv)
 514 {
 515         printf("%s [options] -- [filenames]\n"
 516                 " -d <int> : IO Depth, default %d\n"
 517                 " -s <int> : Batch submit, default %d\n"
 518                 " -c <int> : Batch complete, default %d\n"
 519                 " -b <int> : Block size, default %d\n"
 520                 " -p <bool> : Polled IO, default %d\n",
 521                 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled);
 522         exit(0);
 523 }
 524
 525 int main(int argc, char *argv[])
 526 {
 527         struct submitter *s;
 528         unsigned long done, calls, reap;
 529         int err, i, flags, fd, opt;
 530         char *fdepths;
 531         void *ret;
 532
 533         if (!do_nop && argc < 2) {
 534                 printf("%s: filename [options]\n", argv[0]);
 535                 return 1;
 536         }
 537
 538         while ((opt = getopt(argc, argv, "d:s:c:b:p:B:F:h?")) != -1) {
 539                 switch (opt) {
 540                 case 'd':
 541                         depth = atoi(optarg);
 542                         break;
 543                 case 's':
 544                         batch_submit = atoi(optarg);
 545                         break;
 546                 case 'c':
 547                         batch_complete = atoi(optarg);
 548                         break;
 549                 case 'b':
 550                         bs = atoi(optarg);
 551                         break;
 552                 case 'p':
 553                         polled = !!atoi(optarg);
 554                         break;
 555                 case 'B':
 556                         fixedbufs = !!atoi(optarg);
 557                         break;
 558                 case 'F':
 559                         register_files = !!atoi(optarg);
 560                         break;
 561                 case 'h':
 562                 case '?':
 563                 default:
 564                         usage(argv[0]);
 565                         break;
 566                 }
 567         }
 568
 569         submitter = malloc(sizeof(*submitter) + depth * sizeof(struct iovec));
 570         memset(submitter, 0, sizeof(*submitter) + depth * sizeof(struct iovec));
 571         s = submitter;
 572
 573         flags = O_RDONLY | O_NOATIME;
 574         if (!buffered)
 575                 flags |= O_DIRECT;
 576
 577         i = optind;
 578         while (!do_nop && i < argc) {
 579                 struct file *f;
 580
 581                 if (s->nr_files == MAX_FDS) {
 582                         printf("Max number of files (%d) reached\n", MAX_FDS);
 583                         break;
 584                 }
 585                 fd = open(argv[i], flags);
 586                 if (fd < 0) {
 587                         perror("open");
 588                         return 1;
 589                 }
 590
 591                 f = &s->files[s->nr_files];
 592                 f->real_fd = fd;
 593                 if (get_file_size(f)) {
 594                         printf("failed getting size of device/file\n");
 595                         return 1;
 596                 }
 597                 if (f->max_blocks <= 1) {
 598                         printf("Zero file/device size?\n");
 599                         return 1;
 600                 }
 601                 f->max_blocks--;
 602
 603                 printf("Added file %s\n", argv[i]);
 604                 s->nr_files++;
 605                 i++;
 606         }
 607
 608         if (fixedbufs) {
 609                 struct rlimit rlim;
 610
 611                 rlim.rlim_cur = RLIM_INFINITY;
 612                 rlim.rlim_max = RLIM_INFINITY;
 613                 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 614                         perror("setrlimit");
 615                         return 1;
 616                 }
 617         }
 618
 619         arm_sig_int();
 620
 621         for (i = 0; i < depth; i++) {
 622                 void *buf;
 623
 624                 if (posix_memalign(&buf, bs, bs)) {
 625                         printf("failed alloc\n");
 626                         return 1;
 627                 }
 628                 s->iovecs[i].iov_base = buf;
 629                 s->iovecs[i].iov_len = bs;
 630         }
 631
 632         err = setup_ring(s);
 633         if (err) {
 634                 printf("ring setup failed: %s, %d\n", strerror(errno), err);
 635                 return 1;
 636         }
 637         printf("polled=%d, fixedbufs=%d, register_files=%d, buffered=%d", polled, fixedbufs, register_files, buffered);
 638         printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", depth, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
 639
 640         pthread_create(&s->thread, NULL, submitter_fn, s);
 641
 642         fdepths = malloc(8 * s->nr_files);
 643         reap = calls = done = 0;
 644         do {
 645                 unsigned long this_done = 0;
 646                 unsigned long this_reap = 0;
 647                 unsigned long this_call = 0;
 648                 unsigned long rpc = 0, ipc = 0;
 649
 650                 sleep(1);
 651                 this_done += s->done;
 652                 this_call += s->calls;
 653                 this_reap += s->reaps;
 654                 if (this_call - calls) {
 655                         rpc = (this_done - done) / (this_call - calls);
 656                         ipc = (this_reap - reap) / (this_call - calls);
 657                 } else
 658                         rpc = ipc = -1;
 659                 file_depths(fdepths);
 660                 printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s)\n",
 661                                 this_done - done, rpc, ipc, s->inflight,
 662                                 fdepths);
 663                 done = this_done;
 664                 calls = this_call;
 665                 reap = this_reap;
 666         } while (!finish);
 667
 668         pthread_join(s->thread, &ret);
 669         close(s->ring_fd);
 670         free(fdepths);
 671         return 0;
 672 }