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         int index;
  66         struct io_sq_ring sq_ring;
  67         struct io_uring_sqe *sqes;
  68         struct io_cq_ring cq_ring;
  69         int inflight;
  70         unsigned long reaps;
  71         unsigned long done;
  72         unsigned long calls;
  73         volatile int finish;
  74
  75         __s32 *fds;
  76
  77         struct file files[MAX_FDS];
  78         unsigned nr_files;
  79         unsigned cur_file;
  80         struct iovec iovecs[];
  81 };
  82
  83 static struct submitter *submitter;
  84 static volatile int finish;
  85
  86 static int depth = DEPTH;
  87 static int batch_submit = BATCH_SUBMIT;
  88 static int batch_complete = BATCH_COMPLETE;
  89 static int bs = BS;
  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 static int nthreads = 1;
  98
  99 static int vectored = 1;
 100
 101 static int io_uring_register_buffers(struct submitter *s)
 102 {
 103         if (do_nop)
 104                 return 0;
 105
 106         return syscall(__NR_io_uring_register, s->ring_fd,
 107                         IORING_REGISTER_BUFFERS, s->iovecs, depth);
 108 }
 109
 110 static int io_uring_register_files(struct submitter *s)
 111 {
 112         int i;
 113
 114         if (do_nop)
 115                 return 0;
 116
 117         s->fds = calloc(s->nr_files, sizeof(__s32));
 118         for (i = 0; i < s->nr_files; i++) {
 119                 s->fds[i] = s->files[i].real_fd;
 120                 s->files[i].fixed_fd = i;
 121         }
 122
 123         return syscall(__NR_io_uring_register, s->ring_fd,
 124                         IORING_REGISTER_FILES, s->fds, s->nr_files);
 125 }
 126
 127 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 128 {
 129         return syscall(__NR_io_uring_setup, entries, p);
 130 }
 131
 132 static void io_uring_probe(int fd)
 133 {
 134         struct io_uring_probe *p;
 135         int ret;
 136
 137         p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 138         if (!p)
 139                 return;
 140
 141         memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
 142         ret = syscall(__NR_io_uring_register, fd, IORING_REGISTER_PROBE, p, 256);
 143         if (ret < 0)
 144                 goto out;
 145
 146         if (IORING_OP_READ > p->ops_len)
 147                 goto out;
 148
 149         if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED))
 150                 vectored = 0;
 151 out:
 152         free(p);
 153 }
 154
 155 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 156                           unsigned int min_complete, unsigned int flags)
 157 {
 158         return syscall(__NR_io_uring_enter, s->ring_fd, to_submit, min_complete,
 159                         flags, NULL, 0);
 160 }
 161
 162 #ifndef CONFIG_HAVE_GETTID
 163 static int gettid(void)
 164 {
 165         return syscall(__NR_gettid);
 166 }
 167 #endif
 168
 169 static unsigned file_depth(struct submitter *s)
 170 {
 171         return (depth + s->nr_files - 1) / s->nr_files;
 172 }
 173
 174 static void init_io(struct submitter *s, unsigned index)
 175 {
 176         struct io_uring_sqe *sqe = &s->sqes[index];
 177         unsigned long offset;
 178         struct file *f;
 179         long r;
 180
 181         if (do_nop) {
 182                 sqe->opcode = IORING_OP_NOP;
 183                 return;
 184         }
 185
 186         if (s->nr_files == 1) {
 187                 f = &s->files[0];
 188         } else {
 189                 f = &s->files[s->cur_file];
 190                 if (f->pending_ios >= file_depth(s)) {
 191                         s->cur_file++;
 192                         if (s->cur_file == s->nr_files)
 193                                 s->cur_file = 0;
 194                         f = &s->files[s->cur_file];
 195                 }
 196         }
 197         f->pending_ios++;
 198
 199         r = lrand48();
 200         offset = (r % (f->max_blocks - 1)) * bs;
 201
 202         if (register_files) {
 203                 sqe->flags = IOSQE_FIXED_FILE;
 204                 sqe->fd = f->fixed_fd;
 205         } else {
 206                 sqe->flags = 0;
 207                 sqe->fd = f->real_fd;
 208         }
 209         if (fixedbufs) {
 210                 sqe->opcode = IORING_OP_READ_FIXED;
 211                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 212                 sqe->len = bs;
 213                 sqe->buf_index = index;
 214         } else if (!vectored) {
 215                 sqe->opcode = IORING_OP_READ;
 216                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 217                 sqe->len = bs;
 218                 sqe->buf_index = 0;
 219         } else {
 220                 sqe->opcode = IORING_OP_READV;
 221                 sqe->addr = (unsigned long) &s->iovecs[index];
 222                 sqe->len = 1;
 223                 sqe->buf_index = 0;
 224         }
 225         sqe->ioprio = 0;
 226         sqe->off = offset;
 227         sqe->user_data = (unsigned long) f;
 228 }
 229
 230 static int prep_more_ios(struct submitter *s, int max_ios)
 231 {
 232         struct io_sq_ring *ring = &s->sq_ring;
 233         unsigned index, tail, next_tail, prepped = 0;
 234
 235         next_tail = tail = *ring->tail;
 236         do {
 237                 next_tail++;
 238                 if (next_tail == atomic_load_acquire(ring->head))
 239                         break;
 240
 241                 index = tail & sq_ring_mask;
 242                 init_io(s, index);
 243                 ring->array[index] = index;
 244                 prepped++;
 245                 tail = next_tail;
 246         } while (prepped < max_ios);
 247
 248         if (prepped)
 249                 atomic_store_release(ring->tail, tail);
 250         return prepped;
 251 }
 252
 253 static int get_file_size(struct file *f)
 254 {
 255         struct stat st;
 256
 257         if (fstat(f->real_fd, &st) < 0)
 258                 return -1;
 259         if (S_ISBLK(st.st_mode)) {
 260                 unsigned long long bytes;
 261
 262                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 263                         return -1;
 264
 265                 f->max_blocks = bytes / bs;
 266                 return 0;
 267         } else if (S_ISREG(st.st_mode)) {
 268                 f->max_blocks = st.st_size / bs;
 269                 return 0;
 270         }
 271
 272         return -1;
 273 }
 274
 275 static int reap_events(struct submitter *s)
 276 {
 277         struct io_cq_ring *ring = &s->cq_ring;
 278         struct io_uring_cqe *cqe;
 279         unsigned head, reaped = 0;
 280
 281         head = *ring->head;
 282         do {
 283                 struct file *f;
 284
 285                 read_barrier();
 286                 if (head == atomic_load_acquire(ring->tail))
 287                         break;
 288                 cqe = &ring->cqes[head & cq_ring_mask];
 289                 if (!do_nop) {
 290                         f = (struct file *) (uintptr_t) cqe->user_data;
 291                         f->pending_ios--;
 292                         if (cqe->res != bs) {
 293                                 printf("io: unexpected ret=%d\n", cqe->res);
 294                                 if (polled && cqe->res == -EOPNOTSUPP)
 295                                         printf("Your filesystem/driver/kernel doesn't support polled IO\n");
 296                                 return -1;
 297                         }
 298                 }
 299                 reaped++;
 300                 head++;
 301         } while (1);
 302
 303         if (reaped) {
 304                 s->inflight -= reaped;
 305                 atomic_store_release(ring->head, head);
 306         }
 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                 unsigned ring_flags = 0;
 324
 325                 if (!prepped && s->inflight < depth) {
 326                         to_prep = min(depth - s->inflight, batch_submit);
 327                         prepped = prep_more_ios(s, to_prep);
 328                 }
 329                 s->inflight += prepped;
 330 submit_more:
 331                 to_submit = prepped;
 332 submit:
 333                 if (to_submit && (s->inflight + to_submit <= depth))
 334                         to_wait = 0;
 335                 else
 336                         to_wait = min(s->inflight + to_submit, batch_complete);
 337
 338                 /*
 339                  * Only need to call io_uring_enter if we're not using SQ thread
 340                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
 341                  */
 342                 if (sq_thread_poll)
 343                         ring_flags = atomic_load_acquire(ring->flags);
 344                 if (!sq_thread_poll || ring_flags & IORING_SQ_NEED_WAKEUP) {
 345                         unsigned flags = 0;
 346
 347                         if (to_wait)
 348                                 flags = IORING_ENTER_GETEVENTS;
 349                         if (ring_flags & IORING_SQ_NEED_WAKEUP)
 350                                 flags |= IORING_ENTER_SQ_WAKEUP;
 351                         ret = io_uring_enter(s, to_submit, to_wait, flags);
 352                         s->calls++;
 353                 } else {
 354                         /* for SQPOLL, we submitted it all effectively */
 355                         ret = to_submit;
 356                 }
 357
 358                 /*
 359                  * For non SQ thread poll, we already got the events we needed
 360                  * through the io_uring_enter() above. For SQ thread poll, we
 361                  * need to loop here until we find enough events.
 362                  */
 363                 this_reap = 0;
 364                 do {
 365                         int r;
 366                         r = reap_events(s);
 367                         if (r == -1) {
 368                                 s->finish = 1;
 369                                 break;
 370                         } else if (r > 0)
 371                                 this_reap += r;
 372                 } while (sq_thread_poll && this_reap < to_wait);
 373                 s->reaps += this_reap;
 374
 375                 if (ret >= 0) {
 376                         if (!ret) {
 377                                 to_submit = 0;
 378                                 if (s->inflight)
 379                                         goto submit;
 380                                 continue;
 381                         } else if (ret < to_submit) {
 382                                 int diff = to_submit - ret;
 383
 384                                 s->done += ret;
 385                                 prepped -= diff;
 386                                 goto submit_more;
 387                         }
 388                         s->done += ret;
 389                         prepped = 0;
 390                         continue;
 391                 } else if (ret < 0) {
 392                         if (errno == EAGAIN) {
 393                                 if (s->finish)
 394                                         break;
 395                                 if (this_reap)
 396                                         goto submit;
 397                                 to_submit = 0;
 398                                 goto submit;
 399                         }
 400                         printf("io_submit: %s\n", strerror(errno));
 401                         break;
 402                 }
 403         } while (!s->finish);
 404
 405         finish = 1;
 406         return NULL;
 407 }
 408
 409 static struct submitter *get_submitter(int offset)
 410 {
 411         void *ret;
 412
 413         ret = submitter;
 414         if (offset)
 415                 ret += offset * (sizeof(*submitter) + depth * sizeof(struct iovec));
 416         return ret;
 417 }
 418
 419 static void sig_int(int sig)
 420 {
 421         int j;
 422
 423         printf("Exiting on signal %d\n", sig);
 424         for (j = 0; j < nthreads; j++) {
 425                 struct submitter *s = get_submitter(j);
 426                 s->finish = 1;
 427         }
 428         finish = 1;
 429 }
 430
 431 static void arm_sig_int(void)
 432 {
 433         struct sigaction act;
 434
 435         memset(&act, 0, sizeof(act));
 436         act.sa_handler = sig_int;
 437         act.sa_flags = SA_RESTART;
 438         sigaction(SIGINT, &act, NULL);
 439 }
 440
 441 static int setup_ring(struct submitter *s)
 442 {
 443         struct io_sq_ring *sring = &s->sq_ring;
 444         struct io_cq_ring *cring = &s->cq_ring;
 445         struct io_uring_params p;
 446         int ret, fd;
 447         void *ptr;
 448
 449         memset(&p, 0, sizeof(p));
 450
 451         if (polled && !do_nop)
 452                 p.flags |= IORING_SETUP_IOPOLL;
 453         if (sq_thread_poll) {
 454                 p.flags |= IORING_SETUP_SQPOLL;
 455                 if (sq_thread_cpu != -1) {
 456                         p.flags |= IORING_SETUP_SQ_AFF;
 457                         p.sq_thread_cpu = sq_thread_cpu;
 458                 }
 459         }
 460
 461         fd = io_uring_setup(depth, &p);
 462         if (fd < 0) {
 463                 perror("io_uring_setup");
 464                 return 1;
 465         }
 466         s->ring_fd = fd;
 467
 468         io_uring_probe(fd);
 469
 470         if (fixedbufs) {
 471                 ret = io_uring_register_buffers(s);
 472                 if (ret < 0) {
 473                         perror("io_uring_register_buffers");
 474                         return 1;
 475                 }
 476         }
 477
 478         if (register_files) {
 479                 ret = io_uring_register_files(s);
 480                 if (ret < 0) {
 481                         perror("io_uring_register_files");
 482                         return 1;
 483                 }
 484         }
 485
 486         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 487                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 488                         IORING_OFF_SQ_RING);
 489         printf("sq_ring ptr = 0x%p\n", ptr);
 490         sring->head = ptr + p.sq_off.head;
 491         sring->tail = ptr + p.sq_off.tail;
 492         sring->ring_mask = ptr + p.sq_off.ring_mask;
 493         sring->ring_entries = ptr + p.sq_off.ring_entries;
 494         sring->flags = ptr + p.sq_off.flags;
 495         sring->array = ptr + p.sq_off.array;
 496         sq_ring_mask = *sring->ring_mask;
 497
 498         s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
 499                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 500                         IORING_OFF_SQES);
 501         printf("sqes ptr    = 0x%p\n", s->sqes);
 502
 503         ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
 504                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 505                         IORING_OFF_CQ_RING);
 506         printf("cq_ring ptr = 0x%p\n", ptr);
 507         cring->head = ptr + p.cq_off.head;
 508         cring->tail = ptr + p.cq_off.tail;
 509         cring->ring_mask = ptr + p.cq_off.ring_mask;
 510         cring->ring_entries = ptr + p.cq_off.ring_entries;
 511         cring->cqes = ptr + p.cq_off.cqes;
 512         cq_ring_mask = *cring->ring_mask;
 513         return 0;
 514 }
 515
 516 static void file_depths(char *buf)
 517 {
 518         char *p;
 519         int i, j;
 520
 521         buf[0] = '\0';
 522         p = buf;
 523         for (j = 0; j < nthreads; j++) {
 524                 struct submitter *s = get_submitter(j);
 525
 526                 for (i = 0; i < s->nr_files; i++) {
 527                         struct file *f = &s->files[i];
 528
 529                         if (i + 1 == s->nr_files)
 530                                 p += sprintf(p, "%d", f->pending_ios);
 531                         else
 532                                 p += sprintf(p, "%d, ", f->pending_ios);
 533                 }
 534         }
 535 }
 536
 537 static void usage(char *argv)
 538 {
 539         printf("%s [options] -- [filenames]\n"
 540                 " -d <int> : IO Depth, default %d\n"
 541                 " -s <int> : Batch submit, default %d\n"
 542                 " -c <int> : Batch complete, default %d\n"
 543                 " -b <int> : Block size, default %d\n"
 544                 " -p <bool> : Polled IO, default %d\n",
 545                 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled);
 546         exit(0);
 547 }
 548
 549 int main(int argc, char *argv[])
 550 {
 551         struct submitter *s;
 552         unsigned long done, calls, reap;
 553         int err, i, j, flags, fd, opt;
 554         char *fdepths;
 555         void *ret;
 556
 557         if (!do_nop && argc < 2) {
 558                 printf("%s: filename [options]\n", argv[0]);
 559                 return 1;
 560         }
 561
 562         while ((opt = getopt(argc, argv, "d:s:c:b:p:B:F:n:h?")) != -1) {
 563                 switch (opt) {
 564                 case 'd':
 565                         depth = atoi(optarg);
 566                         break;
 567                 case 's':
 568                         batch_submit = atoi(optarg);
 569                         break;
 570                 case 'c':
 571                         batch_complete = atoi(optarg);
 572                         break;
 573                 case 'b':
 574                         bs = atoi(optarg);
 575                         break;
 576                 case 'p':
 577                         polled = !!atoi(optarg);
 578                         break;
 579                 case 'B':
 580                         fixedbufs = !!atoi(optarg);
 581                         break;
 582                 case 'F':
 583                         register_files = !!atoi(optarg);
 584                         break;
 585                 case 'n':
 586                         nthreads = atoi(optarg);
 587                         break;
 588                 case 'h':
 589                 case '?':
 590                 default:
 591                         usage(argv[0]);
 592                         break;
 593                 }
 594         }
 595
 596         submitter = calloc(nthreads, sizeof(*submitter) +
 597                                 depth * sizeof(struct iovec));
 598         for (j = 0; j < nthreads; j++) {
 599                 s = get_submitter(j);
 600                 s->index = j;
 601                 s->done = s->calls = s->reaps = 0;
 602         }
 603
 604         flags = O_RDONLY | O_NOATIME;
 605         if (!buffered)
 606                 flags |= O_DIRECT;
 607
 608         j = 0;
 609         i = optind;
 610         printf("i %d, argc %d\n", i, argc);
 611         while (!do_nop && i < argc) {
 612                 struct file *f;
 613
 614                 s = get_submitter(j);
 615                 if (s->nr_files == MAX_FDS) {
 616                         printf("Max number of files (%d) reached\n", MAX_FDS);
 617                         break;
 618                 }
 619                 fd = open(argv[i], flags);
 620                 if (fd < 0) {
 621                         perror("open");
 622                         return 1;
 623                 }
 624
 625                 f = &s->files[s->nr_files];
 626                 f->real_fd = fd;
 627                 if (get_file_size(f)) {
 628                         printf("failed getting size of device/file\n");
 629                         return 1;
 630                 }
 631                 if (f->max_blocks <= 1) {
 632                         printf("Zero file/device size?\n");
 633                         return 1;
 634                 }
 635                 f->max_blocks--;
 636
 637                 printf("Added file %s (submitter %d)\n", argv[i], s->index);
 638                 s->nr_files++;
 639                 i++;
 640                 if (++j >= nthreads)
 641                         j = 0;
 642         }
 643
 644         if (fixedbufs) {
 645                 struct rlimit rlim;
 646
 647                 rlim.rlim_cur = RLIM_INFINITY;
 648                 rlim.rlim_max = RLIM_INFINITY;
 649                 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 650                         perror("setrlimit");
 651                         return 1;
 652                 }
 653         }
 654
 655         arm_sig_int();
 656
 657         for (j = 0; j < nthreads; j++) {
 658                 s = get_submitter(j);
 659                 for (i = 0; i < depth; i++) {
 660                         void *buf;
 661
 662                         if (posix_memalign(&buf, bs, bs)) {
 663                                 printf("failed alloc\n");
 664                                 return 1;
 665                         }
 666                         s->iovecs[i].iov_base = buf;
 667                         s->iovecs[i].iov_len = bs;
 668                 }
 669         }
 670
 671         for (j = 0; j < nthreads; j++) {
 672                 s = get_submitter(j);
 673
 674                 err = setup_ring(s);
 675                 if (err) {
 676                         printf("ring setup failed: %s, %d\n", strerror(errno), err);
 677                         return 1;
 678                 }
 679         }
 680         s = get_submitter(0);
 681         printf("polled=%d, fixedbufs=%d, register_files=%d, buffered=%d", polled, fixedbufs, register_files, buffered);
 682         printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", depth, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
 683
 684         for (j = 0; j < nthreads; j++) {
 685                 s = get_submitter(j);
 686                 pthread_create(&s->thread, NULL, submitter_fn, s);
 687         }
 688
 689         fdepths = malloc(8 * s->nr_files * nthreads);
 690         reap = calls = done = 0;
 691         do {
 692                 unsigned long this_done = 0;
 693                 unsigned long this_reap = 0;
 694                 unsigned long this_call = 0;
 695                 unsigned long rpc = 0, ipc = 0;
 696
 697                 sleep(1);
 698                 for (j = 0; j < nthreads; j++) {
 699                         this_done += s->done;
 700                         this_call += s->calls;
 701                         this_reap += s->reaps;
 702                 }
 703                 if (this_call - calls) {
 704                         rpc = (this_done - done) / (this_call - calls);
 705                         ipc = (this_reap - reap) / (this_call - calls);
 706                 } else
 707                         rpc = ipc = -1;
 708                 file_depths(fdepths);
 709                 printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s)\n",
 710                                 this_done - done, rpc, ipc, s->inflight,
 711                                 fdepths);
 712                 done = this_done;
 713                 calls = this_call;
 714                 reap = this_reap;
 715         } while (!finish);
 716
 717         for (j = 0; j < nthreads; j++) {
 718                 s = get_submitter(j);
 719                 pthread_join(s->thread, &ret);
 720                 close(s->ring_fd);
 721         }
 722         free(fdepths);
 723         free(submitter);
 724         return 0;
 725 }