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         unsigned long cachehit, cachemiss;
  75         volatile int finish;
  76
  77         __s32 *fds;
  78
  79         struct file files[MAX_FDS];
  80         unsigned nr_files;
  81         unsigned cur_file;
  82         struct iovec iovecs[];
  83 };
  84
  85 static struct submitter *submitter;
  86 static volatile int finish;
  87
  88 static int depth = DEPTH;
  89 static int batch_submit = BATCH_SUBMIT;
  90 static int batch_complete = BATCH_COMPLETE;
  91 static int polled = 1;          /* use IO polling */
  92 static int fixedbufs = 1;       /* use fixed user buffers */
  93 static int register_files = 1;  /* use fixed files */
  94 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
  95 static int sq_thread_poll = 0;  /* use kernel submission/poller thread */
  96 static int sq_thread_cpu = -1;  /* pin above thread to this CPU */
  97 static int do_nop = 0;          /* no-op SQ ring commands */
  98
  99 static int io_uring_register_buffers(struct submitter *s)
 100 {
 101         if (do_nop)
 102                 return 0;
 103
 104         return syscall(__NR_sys_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_sys_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_sys_io_uring_setup, entries, p);
 128 }
 129
 130 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 131                           unsigned int min_complete, unsigned int flags)
 132 {
 133         return syscall(__NR_sys_io_uring_enter, s->ring_fd, to_submit,
 134                         min_complete, flags, NULL, 0);
 135 }
 136
 137 static int gettid(void)
 138 {
 139         return syscall(__NR_gettid);
 140 }
 141
 142 static unsigned file_depth(struct submitter *s)
 143 {
 144         return (depth + s->nr_files - 1) / s->nr_files;
 145 }
 146
 147 static void init_io(struct submitter *s, unsigned index)
 148 {
 149         struct io_uring_sqe *sqe = &s->sqes[index];
 150         unsigned long offset;
 151         struct file *f;
 152         long r;
 153
 154         if (do_nop) {
 155                 sqe->opcode = IORING_OP_NOP;
 156                 return;
 157         }
 158
 159         if (s->nr_files == 1) {
 160                 f = &s->files[0];
 161         } else {
 162                 f = &s->files[s->cur_file];
 163                 if (f->pending_ios >= file_depth(s)) {
 164                         s->cur_file++;
 165                         if (s->cur_file == s->nr_files)
 166                                 s->cur_file = 0;
 167                         f = &s->files[s->cur_file];
 168                 }
 169         }
 170         f->pending_ios++;
 171
 172         lrand48_r(&s->rand, &r);
 173         offset = (r % (f->max_blocks - 1)) * BS;
 174
 175         if (register_files) {
 176                 sqe->flags = IOSQE_FIXED_FILE;
 177                 sqe->fd = f->fixed_fd;
 178         } else {
 179                 sqe->flags = 0;
 180                 sqe->fd = f->real_fd;
 181         }
 182         if (fixedbufs) {
 183                 sqe->opcode = IORING_OP_READ_FIXED;
 184                 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
 185                 sqe->len = BS;
 186                 sqe->buf_index = index;
 187         } else {
 188                 sqe->opcode = IORING_OP_READV;
 189                 sqe->addr = (unsigned long) &s->iovecs[index];
 190                 sqe->len = 1;
 191                 sqe->buf_index = 0;
 192         }
 193         sqe->ioprio = 0;
 194         sqe->off = offset;
 195         sqe->user_data = (unsigned long) f;
 196 }
 197
 198 static int prep_more_ios(struct submitter *s, int max_ios)
 199 {
 200         struct io_sq_ring *ring = &s->sq_ring;
 201         unsigned index, tail, next_tail, prepped = 0;
 202
 203         next_tail = tail = *ring->tail;
 204         do {
 205                 next_tail++;
 206                 read_barrier();
 207                 if (next_tail == *ring->head)
 208                         break;
 209
 210                 index = tail & sq_ring_mask;
 211                 init_io(s, index);
 212                 ring->array[index] = index;
 213                 prepped++;
 214                 tail = next_tail;
 215         } while (prepped < max_ios);
 216
 217         if (*ring->tail != tail) {
 218                 /* order tail store with writes to sqes above */
 219                 write_barrier();
 220                 *ring->tail = tail;
 221                 write_barrier();
 222         }
 223         return prepped;
 224 }
 225
 226 static int get_file_size(struct file *f)
 227 {
 228         struct stat st;
 229
 230         if (fstat(f->real_fd, &st) < 0)
 231                 return -1;
 232         if (S_ISBLK(st.st_mode)) {
 233                 unsigned long long bytes;
 234
 235                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 236                         return -1;
 237
 238                 f->max_blocks = bytes / BS;
 239                 return 0;
 240         } else if (S_ISREG(st.st_mode)) {
 241                 f->max_blocks = st.st_size / BS;
 242                 return 0;
 243         }
 244
 245         return -1;
 246 }
 247
 248 static int reap_events(struct submitter *s)
 249 {
 250         struct io_cq_ring *ring = &s->cq_ring;
 251         struct io_uring_cqe *cqe;
 252         unsigned head, reaped = 0;
 253
 254         head = *ring->head;
 255         do {
 256                 struct file *f;
 257
 258                 read_barrier();
 259                 if (head == *ring->tail)
 260                         break;
 261                 cqe = &ring->cqes[head & cq_ring_mask];
 262                 if (!do_nop) {
 263                         f = (struct file *) (uintptr_t) cqe->user_data;
 264                         f->pending_ios--;
 265                         if (cqe->res != BS) {
 266                                 printf("io: unexpected ret=%d\n", cqe->res);
 267                                 if (polled && cqe->res == -EOPNOTSUPP)
 268                                         printf("Your filesystem doesn't support poll\n");
 269                                 return -1;
 270                         }
 271                 }
 272                 if (cqe->flags & IOCQE_FLAG_CACHEHIT)
 273                         s->cachehit++;
 274                 else
 275                         s->cachemiss++;
 276                 reaped++;
 277                 head++;
 278         } while (1);
 279
 280         s->inflight -= reaped;
 281         *ring->head = head;
 282         write_barrier();
 283         return reaped;
 284 }
 285
 286 static void *submitter_fn(void *data)
 287 {
 288         struct submitter *s = data;
 289         struct io_sq_ring *ring = &s->sq_ring;
 290         int ret, prepped;
 291
 292         printf("submitter=%d\n", gettid());
 293
 294         srand48_r(pthread_self(), &s->rand);
 295
 296         prepped = 0;
 297         do {
 298                 int to_wait, to_submit, this_reap, to_prep;
 299
 300                 if (!prepped && s->inflight < depth) {
 301                         to_prep = min(depth - s->inflight, batch_submit);
 302                         prepped = prep_more_ios(s, to_prep);
 303                 }
 304                 s->inflight += prepped;
 305 submit_more:
 306                 to_submit = prepped;
 307 submit:
 308                 if (to_submit && (s->inflight + to_submit <= depth))
 309                         to_wait = 0;
 310                 else
 311                         to_wait = min(s->inflight + to_submit, batch_complete);
 312
 313                 /*
 314                  * Only need to call io_uring_enter if we're not using SQ thread
 315                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
 316                  */
 317                 if (!sq_thread_poll || (*ring->flags & IORING_SQ_NEED_WAKEUP)) {
 318                         unsigned flags = 0;
 319
 320                         if (to_wait)
 321                                 flags = IORING_ENTER_GETEVENTS;
 322                         if ((*ring->flags & IORING_SQ_NEED_WAKEUP))
 323                                 flags |= IORING_ENTER_SQ_WAKEUP;
 324                         ret = io_uring_enter(s, to_submit, to_wait, flags);
 325                         s->calls++;
 326                 }
 327
 328                 /*
 329                  * For non SQ thread poll, we already got the events we needed
 330                  * through the io_uring_enter() above. For SQ thread poll, we
 331                  * need to loop here until we find enough events.
 332                  */
 333                 this_reap = 0;
 334                 do {
 335                         int r;
 336                         r = reap_events(s);
 337                         if (r == -1) {
 338                                 s->finish = 1;
 339                                 break;
 340                         } else if (r > 0)
 341                                 this_reap += r;
 342                 } while (sq_thread_poll && this_reap < to_wait);
 343                 s->reaps += this_reap;
 344
 345                 if (ret >= 0) {
 346                         if (!ret) {
 347                                 to_submit = 0;
 348                                 if (s->inflight)
 349                                         goto submit;
 350                                 continue;
 351                         } else if (ret < to_submit) {
 352                                 int diff = to_submit - ret;
 353
 354                                 s->done += ret;
 355                                 prepped -= diff;
 356                                 goto submit_more;
 357                         }
 358                         s->done += ret;
 359                         prepped = 0;
 360                         continue;
 361                 } else if (ret < 0) {
 362                         if (errno == EAGAIN) {
 363                                 if (s->finish)
 364                                         break;
 365                                 if (this_reap)
 366                                         goto submit;
 367                                 to_submit = 0;
 368                                 goto submit;
 369                         }
 370                         printf("io_submit: %s\n", strerror(errno));
 371                         break;
 372                 }
 373         } while (!s->finish);
 374
 375         finish = 1;
 376         return NULL;
 377 }
 378
 379 static void sig_int(int sig)
 380 {
 381         printf("Exiting on signal %d\n", sig);
 382         submitter->finish = 1;
 383         finish = 1;
 384 }
 385
 386 static void arm_sig_int(void)
 387 {
 388         struct sigaction act;
 389
 390         memset(&act, 0, sizeof(act));
 391         act.sa_handler = sig_int;
 392         act.sa_flags = SA_RESTART;
 393         sigaction(SIGINT, &act, NULL);
 394 }
 395
 396 static int setup_ring(struct submitter *s)
 397 {
 398         struct io_sq_ring *sring = &s->sq_ring;
 399         struct io_cq_ring *cring = &s->cq_ring;
 400         struct io_uring_params p;
 401         int ret, fd;
 402         void *ptr;
 403
 404         memset(&p, 0, sizeof(p));
 405
 406         if (polled && !do_nop)
 407                 p.flags |= IORING_SETUP_IOPOLL;
 408         if (sq_thread_poll) {
 409                 p.flags |= IORING_SETUP_SQPOLL;
 410                 if (sq_thread_cpu != -1) {
 411                         p.flags |= IORING_SETUP_SQ_AFF;
 412                         p.sq_thread_cpu = sq_thread_cpu;
 413                 }
 414         }
 415
 416         fd = io_uring_setup(depth, &p);
 417         if (fd < 0) {
 418                 perror("io_uring_setup");
 419                 return 1;
 420         }
 421         s->ring_fd = fd;
 422
 423         if (fixedbufs) {
 424                 ret = io_uring_register_buffers(s);
 425                 if (ret < 0) {
 426                         perror("io_uring_register_buffers");
 427                         return 1;
 428                 }
 429         }
 430
 431         if (register_files) {
 432                 ret = io_uring_register_files(s);
 433                 if (ret < 0) {
 434                         perror("io_uring_register_files");
 435                         return 1;
 436                 }
 437         }
 438
 439         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 440                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 441                         IORING_OFF_SQ_RING);
 442         printf("sq_ring ptr = 0x%p\n", ptr);
 443         sring->head = ptr + p.sq_off.head;
 444         sring->tail = ptr + p.sq_off.tail;
 445         sring->ring_mask = ptr + p.sq_off.ring_mask;
 446         sring->ring_entries = ptr + p.sq_off.ring_entries;
 447         sring->flags = ptr + p.sq_off.flags;
 448         sring->array = ptr + p.sq_off.array;
 449         sq_ring_mask = *sring->ring_mask;
 450
 451         s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
 452                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 453                         IORING_OFF_SQES);
 454         printf("sqes ptr    = 0x%p\n", s->sqes);
 455
 456         ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
 457                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 458                         IORING_OFF_CQ_RING);
 459         printf("cq_ring ptr = 0x%p\n", ptr);
 460         cring->head = ptr + p.cq_off.head;
 461         cring->tail = ptr + p.cq_off.tail;
 462         cring->ring_mask = ptr + p.cq_off.ring_mask;
 463         cring->ring_entries = ptr + p.cq_off.ring_entries;
 464         cring->cqes = ptr + p.cq_off.cqes;
 465         cq_ring_mask = *cring->ring_mask;
 466         return 0;
 467 }
 468
 469 static void file_depths(char *buf)
 470 {
 471         struct submitter *s = submitter;
 472         char *p;
 473         int i;
 474
 475         buf[0] = '\0';
 476         p = buf;
 477         for (i = 0; i < s->nr_files; i++) {
 478                 struct file *f = &s->files[i];
 479
 480                 if (i + 1 == s->nr_files)
 481                         p += sprintf(p, "%d", f->pending_ios);
 482                 else
 483                         p += sprintf(p, "%d, ", f->pending_ios);
 484         }
 485 }
 486
 487 static void usage(char *argv)
 488 {
 489         printf("%s [options] -- [filenames]\n"
 490                 " -d <int> : IO Depth, default %d\n"
 491                 " -s <int> : Batch submit, default %d\n"
 492                 " -c <int> : Batch complete, default %d\n",
 493                 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE);
 494         exit(0);
 495 }
 496
 497 int main(int argc, char *argv[])
 498 {
 499         struct submitter *s;
 500         unsigned long done, calls, reap, cache_hit, cache_miss;
 501         int err, i, flags, fd, opt;
 502         char *fdepths;
 503         void *ret;
 504
 505         if (!do_nop && argc < 2) {
 506                 printf("%s: filename [options]\n", argv[0]);
 507                 return 1;
 508         }
 509
 510         while ((opt = getopt(argc, argv, "d:s:c:h?")) != -1) {
 511                 switch (opt) {
 512                 case 'd':
 513                         depth = atoi(optarg);
 514                         break;
 515                 case 's':
 516                         batch_submit = atoi(optarg);
 517                         break;
 518                 case 'c':
 519                         batch_complete = atoi(optarg);
 520                         break;
 521                 case 'h':
 522                 case '?':
 523                 default:
 524                         usage(argv[0]);
 525                         break;
 526                 }
 527         }
 528
 529         submitter = malloc(sizeof(*submitter) + depth * sizeof(struct iovec));
 530         memset(submitter, 0, sizeof(*submitter) + depth * sizeof(struct iovec));
 531         s = submitter;
 532
 533         flags = O_RDONLY | O_NOATIME;
 534         if (!buffered)
 535                 flags |= O_DIRECT;
 536
 537         i = optind;
 538         while (!do_nop && i < argc) {
 539                 struct file *f;
 540
 541                 if (s->nr_files == MAX_FDS) {
 542                         printf("Max number of files (%d) reached\n", MAX_FDS);
 543                         break;
 544                 }
 545                 fd = open(argv[i], flags);
 546                 if (fd < 0) {
 547                         perror("open");
 548                         return 1;
 549                 }
 550
 551                 f = &s->files[s->nr_files];
 552                 f->real_fd = fd;
 553                 if (get_file_size(f)) {
 554                         printf("failed getting size of device/file\n");
 555                         return 1;
 556                 }
 557                 if (f->max_blocks <= 1) {
 558                         printf("Zero file/device size?\n");
 559                         return 1;
 560                 }
 561                 f->max_blocks--;
 562
 563                 printf("Added file %s\n", argv[i]);
 564                 s->nr_files++;
 565                 i++;
 566         }
 567
 568         if (fixedbufs) {
 569                 struct rlimit rlim;
 570
 571                 rlim.rlim_cur = RLIM_INFINITY;
 572                 rlim.rlim_max = RLIM_INFINITY;
 573                 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 574                         perror("setrlimit");
 575                         return 1;
 576                 }
 577         }
 578
 579         arm_sig_int();
 580
 581         for (i = 0; i < depth; i++) {
 582                 void *buf;
 583
 584                 if (posix_memalign(&buf, BS, BS)) {
 585                         printf("failed alloc\n");
 586                         return 1;
 587                 }
 588                 s->iovecs[i].iov_base = buf;
 589                 s->iovecs[i].iov_len = BS;
 590         }
 591
 592         err = setup_ring(s);
 593         if (err) {
 594                 printf("ring setup failed: %s, %d\n", strerror(errno), err);
 595                 return 1;
 596         }
 597         printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered);
 598         printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", depth, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
 599
 600         pthread_create(&s->thread, NULL, submitter_fn, s);
 601
 602         fdepths = malloc(8 * s->nr_files);
 603         cache_hit = cache_miss = reap = calls = done = 0;
 604         do {
 605                 unsigned long this_done = 0;
 606                 unsigned long this_reap = 0;
 607                 unsigned long this_call = 0;
 608                 unsigned long this_cache_hit = 0;
 609                 unsigned long this_cache_miss = 0;
 610                 unsigned long rpc = 0, ipc = 0;
 611                 double hit = 0.0;
 612
 613                 sleep(1);
 614                 this_done += s->done;
 615                 this_call += s->calls;
 616                 this_reap += s->reaps;
 617                 this_cache_hit += s->cachehit;
 618                 this_cache_miss += s->cachemiss;
 619                 if (this_cache_hit && this_cache_miss) {
 620                         unsigned long hits, total;
 621
 622                         hits = this_cache_hit - cache_hit;
 623                         total = hits + this_cache_miss - cache_miss;
 624                         hit = (double) hits / (double) total;
 625                         hit *= 100.0;
 626                 }
 627                 if (this_call - calls) {
 628                         rpc = (this_done - done) / (this_call - calls);
 629                         ipc = (this_reap - reap) / (this_call - calls);
 630                 } else
 631                         rpc = ipc = -1;
 632                 file_depths(fdepths);
 633                 printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s), Cachehit=%0.2f%%\n",
 634                                 this_done - done, rpc, ipc, s->inflight,
 635                                 fdepths, hit);
 636                 done = this_done;
 637                 calls = this_call;
 638                 reap = this_reap;
 639                 cache_hit = s->cachehit;
 640                 cache_miss = s->cachemiss;
 641         } while (!finish);
 642
 643         pthread_join(s->thread, &ret);
 644         close(s->ring_fd);
 645         free(fdepths);
 646         return 0;
 647 }