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                 if (next_tail == atomic_load_acquire(ring->head))
 237                         break;
 238
 239                 index = tail & sq_ring_mask;
 240                 init_io(s, index);
 241                 ring->array[index] = index;
 242                 prepped++;
 243                 tail = next_tail;
 244         } while (prepped < max_ios);
 245
 246         if (prepped)
 247                 atomic_store_release(ring->tail, tail);
 248         return prepped;
 249 }
 250
 251 static int get_file_size(struct file *f)
 252 {
 253         struct stat st;
 254
 255         if (fstat(f->real_fd, &st) < 0)
 256                 return -1;
 257         if (S_ISBLK(st.st_mode)) {
 258                 unsigned long long bytes;
 259
 260                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 261                         return -1;
 262
 263                 f->max_blocks = bytes / bs;
 264                 return 0;
 265         } else if (S_ISREG(st.st_mode)) {
 266                 f->max_blocks = st.st_size / bs;
 267                 return 0;
 268         }
 269
 270         return -1;
 271 }
 272
 273 static int reap_events(struct submitter *s)
 274 {
 275         struct io_cq_ring *ring = &s->cq_ring;
 276         struct io_uring_cqe *cqe;
 277         unsigned head, reaped = 0;
 278
 279         head = *ring->head;
 280         do {
 281                 struct file *f;
 282
 283                 read_barrier();
 284                 if (head == atomic_load_acquire(ring->tail))
 285                         break;
 286                 cqe = &ring->cqes[head & cq_ring_mask];
 287                 if (!do_nop) {
 288                         f = (struct file *) (uintptr_t) cqe->user_data;
 289                         f->pending_ios--;
 290                         if (cqe->res != bs) {
 291                                 printf("io: unexpected ret=%d\n", cqe->res);
 292                                 if (polled && cqe->res == -EOPNOTSUPP)
 293                                         printf("Your filesystem/driver/kernel doesn't support polled IO\n");
 294                                 return -1;
 295                         }
 296                 }
 297                 reaped++;
 298                 head++;
 299         } while (1);
 300
 301         if (reaped) {
 302                 s->inflight -= reaped;
 303                 atomic_store_release(ring->head, head);
 304         }
 305         return reaped;
 306 }
 307
 308 static void *submitter_fn(void *data)
 309 {
 310         struct submitter *s = data;
 311         struct io_sq_ring *ring = &s->sq_ring;
 312         int ret, prepped;
 313
 314         printf("submitter=%d\n", gettid());
 315
 316         srand48(pthread_self());
 317
 318         prepped = 0;
 319         do {
 320                 int to_wait, to_submit, this_reap, to_prep;
 321                 unsigned ring_flags = 0;
 322
 323                 if (!prepped && s->inflight < depth) {
 324                         to_prep = min(depth - s->inflight, batch_submit);
 325                         prepped = prep_more_ios(s, to_prep);
 326                 }
 327                 s->inflight += prepped;
 328 submit_more:
 329                 to_submit = prepped;
 330 submit:
 331                 if (to_submit && (s->inflight + to_submit <= depth))
 332                         to_wait = 0;
 333                 else
 334                         to_wait = min(s->inflight + to_submit, batch_complete);
 335
 336                 /*
 337                  * Only need to call io_uring_enter if we're not using SQ thread
 338                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
 339                  */
 340                 if (sq_thread_poll)
 341                         ring_flags = atomic_load_acquire(ring->flags);
 342                 if (!sq_thread_poll || ring_flags & IORING_SQ_NEED_WAKEUP) {
 343                         unsigned flags = 0;
 344
 345                         if (to_wait)
 346                                 flags = IORING_ENTER_GETEVENTS;
 347                         if (ring_flags & IORING_SQ_NEED_WAKEUP)
 348                                 flags |= IORING_ENTER_SQ_WAKEUP;
 349                         ret = io_uring_enter(s, to_submit, to_wait, flags);
 350                         s->calls++;
 351                 } else {
 352                         /* for SQPOLL, we submitted it all effectively */
 353                         ret = to_submit;
 354                 }
 355
 356                 /*
 357                  * For non SQ thread poll, we already got the events we needed
 358                  * through the io_uring_enter() above. For SQ thread poll, we
 359                  * need to loop here until we find enough events.
 360                  */
 361                 this_reap = 0;
 362                 do {
 363                         int r;
 364                         r = reap_events(s);
 365                         if (r == -1) {
 366                                 s->finish = 1;
 367                                 break;
 368                         } else if (r > 0)
 369                                 this_reap += r;
 370                 } while (sq_thread_poll && this_reap < to_wait);
 371                 s->reaps += this_reap;
 372
 373                 if (ret >= 0) {
 374                         if (!ret) {
 375                                 to_submit = 0;
 376                                 if (s->inflight)
 377                                         goto submit;
 378                                 continue;
 379                         } else if (ret < to_submit) {
 380                                 int diff = to_submit - ret;
 381
 382                                 s->done += ret;
 383                                 prepped -= diff;
 384                                 goto submit_more;
 385                         }
 386                         s->done += ret;
 387                         prepped = 0;
 388                         continue;
 389                 } else if (ret < 0) {
 390                         if (errno == EAGAIN) {
 391                                 if (s->finish)
 392                                         break;
 393                                 if (this_reap)
 394                                         goto submit;
 395                                 to_submit = 0;
 396                                 goto submit;
 397                         }
 398                         printf("io_submit: %s\n", strerror(errno));
 399                         break;
 400                 }
 401         } while (!s->finish);
 402
 403         finish = 1;
 404         return NULL;
 405 }
 406
 407 static void sig_int(int sig)
 408 {
 409         printf("Exiting on signal %d\n", sig);
 410         submitter->finish = 1;
 411         finish = 1;
 412 }
 413
 414 static void arm_sig_int(void)
 415 {
 416         struct sigaction act;
 417
 418         memset(&act, 0, sizeof(act));
 419         act.sa_handler = sig_int;
 420         act.sa_flags = SA_RESTART;
 421         sigaction(SIGINT, &act, NULL);
 422 }
 423
 424 static int setup_ring(struct submitter *s)
 425 {
 426         struct io_sq_ring *sring = &s->sq_ring;
 427         struct io_cq_ring *cring = &s->cq_ring;
 428         struct io_uring_params p;
 429         int ret, fd;
 430         void *ptr;
 431
 432         memset(&p, 0, sizeof(p));
 433
 434         if (polled && !do_nop)
 435                 p.flags |= IORING_SETUP_IOPOLL;
 436         if (sq_thread_poll) {
 437                 p.flags |= IORING_SETUP_SQPOLL;
 438                 if (sq_thread_cpu != -1) {
 439                         p.flags |= IORING_SETUP_SQ_AFF;
 440                         p.sq_thread_cpu = sq_thread_cpu;
 441                 }
 442         }
 443
 444         fd = io_uring_setup(depth, &p);
 445         if (fd < 0) {
 446                 perror("io_uring_setup");
 447                 return 1;
 448         }
 449         s->ring_fd = fd;
 450
 451         io_uring_probe(fd);
 452
 453         if (fixedbufs) {
 454                 ret = io_uring_register_buffers(s);
 455                 if (ret < 0) {
 456                         perror("io_uring_register_buffers");
 457                         return 1;
 458                 }
 459         }
 460
 461         if (register_files) {
 462                 ret = io_uring_register_files(s);
 463                 if (ret < 0) {
 464                         perror("io_uring_register_files");
 465                         return 1;
 466                 }
 467         }
 468
 469         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 470                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 471                         IORING_OFF_SQ_RING);
 472         printf("sq_ring ptr = 0x%p\n", ptr);
 473         sring->head = ptr + p.sq_off.head;
 474         sring->tail = ptr + p.sq_off.tail;
 475         sring->ring_mask = ptr + p.sq_off.ring_mask;
 476         sring->ring_entries = ptr + p.sq_off.ring_entries;
 477         sring->flags = ptr + p.sq_off.flags;
 478         sring->array = ptr + p.sq_off.array;
 479         sq_ring_mask = *sring->ring_mask;
 480
 481         s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
 482                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 483                         IORING_OFF_SQES);
 484         printf("sqes ptr    = 0x%p\n", s->sqes);
 485
 486         ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
 487                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 488                         IORING_OFF_CQ_RING);
 489         printf("cq_ring ptr = 0x%p\n", ptr);
 490         cring->head = ptr + p.cq_off.head;
 491         cring->tail = ptr + p.cq_off.tail;
 492         cring->ring_mask = ptr + p.cq_off.ring_mask;
 493         cring->ring_entries = ptr + p.cq_off.ring_entries;
 494         cring->cqes = ptr + p.cq_off.cqes;
 495         cq_ring_mask = *cring->ring_mask;
 496         return 0;
 497 }
 498
 499 static void file_depths(char *buf)
 500 {
 501         struct submitter *s = submitter;
 502         char *p;
 503         int i;
 504
 505         buf[0] = '\0';
 506         p = buf;
 507         for (i = 0; i < s->nr_files; i++) {
 508                 struct file *f = &s->files[i];
 509
 510                 if (i + 1 == s->nr_files)
 511                         p += sprintf(p, "%d", f->pending_ios);
 512                 else
 513                         p += sprintf(p, "%d, ", f->pending_ios);
 514         }
 515 }
 516
 517 static void usage(char *argv)
 518 {
 519         printf("%s [options] -- [filenames]\n"
 520                 " -d <int> : IO Depth, default %d\n"
 521                 " -s <int> : Batch submit, default %d\n"
 522                 " -c <int> : Batch complete, default %d\n"
 523                 " -b <int> : Block size, default %d\n"
 524                 " -p <bool> : Polled IO, default %d\n",
 525                 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled);
 526         exit(0);
 527 }
 528
 529 int main(int argc, char *argv[])
 530 {
 531         struct submitter *s;
 532         unsigned long done, calls, reap;
 533         int err, i, flags, fd, opt;
 534         char *fdepths;
 535         void *ret;
 536
 537         if (!do_nop && argc < 2) {
 538                 printf("%s: filename [options]\n", argv[0]);
 539                 return 1;
 540         }
 541
 542         while ((opt = getopt(argc, argv, "d:s:c:b:p:B:F:h?")) != -1) {
 543                 switch (opt) {
 544                 case 'd':
 545                         depth = atoi(optarg);
 546                         break;
 547                 case 's':
 548                         batch_submit = atoi(optarg);
 549                         break;
 550                 case 'c':
 551                         batch_complete = atoi(optarg);
 552                         break;
 553                 case 'b':
 554                         bs = atoi(optarg);
 555                         break;
 556                 case 'p':
 557                         polled = !!atoi(optarg);
 558                         break;
 559                 case 'B':
 560                         fixedbufs = !!atoi(optarg);
 561                         break;
 562                 case 'F':
 563                         register_files = !!atoi(optarg);
 564                         break;
 565                 case 'h':
 566                 case '?':
 567                 default:
 568                         usage(argv[0]);
 569                         break;
 570                 }
 571         }
 572
 573         submitter = malloc(sizeof(*submitter) + depth * sizeof(struct iovec));
 574         memset(submitter, 0, sizeof(*submitter) + depth * sizeof(struct iovec));
 575         s = submitter;
 576
 577         flags = O_RDONLY | O_NOATIME;
 578         if (!buffered)
 579                 flags |= O_DIRECT;
 580
 581         i = optind;
 582         while (!do_nop && i < argc) {
 583                 struct file *f;
 584
 585                 if (s->nr_files == MAX_FDS) {
 586                         printf("Max number of files (%d) reached\n", MAX_FDS);
 587                         break;
 588                 }
 589                 fd = open(argv[i], flags);
 590                 if (fd < 0) {
 591                         perror("open");
 592                         return 1;
 593                 }
 594
 595                 f = &s->files[s->nr_files];
 596                 f->real_fd = fd;
 597                 if (get_file_size(f)) {
 598                         printf("failed getting size of device/file\n");
 599                         return 1;
 600                 }
 601                 if (f->max_blocks <= 1) {
 602                         printf("Zero file/device size?\n");
 603                         return 1;
 604                 }
 605                 f->max_blocks--;
 606
 607                 printf("Added file %s\n", argv[i]);
 608                 s->nr_files++;
 609                 i++;
 610         }
 611
 612         if (fixedbufs) {
 613                 struct rlimit rlim;
 614
 615                 rlim.rlim_cur = RLIM_INFINITY;
 616                 rlim.rlim_max = RLIM_INFINITY;
 617                 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 618                         perror("setrlimit");
 619                         return 1;
 620                 }
 621         }
 622
 623         arm_sig_int();
 624
 625         for (i = 0; i < depth; i++) {
 626                 void *buf;
 627
 628                 if (posix_memalign(&buf, bs, bs)) {
 629                         printf("failed alloc\n");
 630                         return 1;
 631                 }
 632                 s->iovecs[i].iov_base = buf;
 633                 s->iovecs[i].iov_len = bs;
 634         }
 635
 636         err = setup_ring(s);
 637         if (err) {
 638                 printf("ring setup failed: %s, %d\n", strerror(errno), err);
 639                 return 1;
 640         }
 641         printf("polled=%d, fixedbufs=%d, register_files=%d, buffered=%d", polled, fixedbufs, register_files, buffered);
 642         printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", depth, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
 643
 644         pthread_create(&s->thread, NULL, submitter_fn, s);
 645
 646         fdepths = malloc(8 * s->nr_files);
 647         reap = calls = done = 0;
 648         do {
 649                 unsigned long this_done = 0;
 650                 unsigned long this_reap = 0;
 651                 unsigned long this_call = 0;
 652                 unsigned long rpc = 0, ipc = 0;
 653
 654                 sleep(1);
 655                 this_done += s->done;
 656                 this_call += s->calls;
 657                 this_reap += s->reaps;
 658                 if (this_call - calls) {
 659                         rpc = (this_done - done) / (this_call - calls);
 660                         ipc = (this_reap - reap) / (this_call - calls);
 661                 } else
 662                         rpc = ipc = -1;
 663                 file_depths(fdepths);
 664                 printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s)\n",
 665                                 this_done - done, rpc, ipc, s->inflight,
 666                                 fdepths);
 667                 done = this_done;
 668                 calls = this_call;
 669                 reap = this_reap;
 670         } while (!finish);
 671
 672         pthread_join(s->thread, &ret);
 673         close(s->ring_fd);
 674         free(fdepths);
 675         return 0;
 676 }