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
  48 #define BATCH_SUBMIT            32
  49 #define BATCH_COMPLETE          32
  50
  51 #define BS                      4096
  52
  53 #define MAX_FDS                 16
  54
  55 static unsigned sq_ring_mask, cq_ring_mask;
  56
  57 struct file {
  58         unsigned long max_blocks;
  59         unsigned pending_ios;
  60         int real_fd;
  61         int fixed_fd;
  62 };
  63
  64 struct submitter {
  65         pthread_t thread;
  66         int ring_fd;
  67         struct drand48_data rand;
  68         struct io_sq_ring sq_ring;
  69         struct io_uring_sqe *sqes;
  70         struct iovec iovecs[DEPTH];
  71         struct io_cq_ring cq_ring;
  72         int inflight;
  73         unsigned long reaps;
  74         unsigned long done;
  75         unsigned long calls;
  76         unsigned long cachehit, cachemiss;
  77         volatile int finish;
  78
  79         __s32 *fds;
  80
  81         struct file files[MAX_FDS];
  82         unsigned nr_files;
  83         unsigned cur_file;
  84 };
  85
  86 static struct submitter submitters[1];
  87 static volatile int finish;
  88
  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 io_uring_register_buffers(struct submitter *s)
  98 {
  99         struct io_uring_register_buffers reg = {
 100                 .iovecs = s->iovecs,
 101                 .nr_iovecs = DEPTH
 102         };
 103
 104         if (do_nop)
 105                 return 0;
 106
 107         return syscall(__NR_sys_io_uring_register, s->ring_fd,
 108                         IORING_REGISTER_BUFFERS, &reg);
 109 }
 110
 111 static int io_uring_register_files(struct submitter *s)
 112 {
 113         struct io_uring_register_files reg;
 114         int i;
 115
 116         if (do_nop)
 117                 return 0;
 118
 119         s->fds = calloc(s->nr_files, sizeof(__s32));
 120         for (i = 0; i < s->nr_files; i++) {
 121                 s->fds[i] = s->files[i].real_fd;
 122                 s->files[i].fixed_fd = i;
 123         }
 124         reg.fds = s->fds;
 125         reg.nr_fds = s->nr_files;
 126
 127         return syscall(__NR_sys_io_uring_register, s->ring_fd,
 128                         IORING_REGISTER_FILES, &reg);
 129 }
 130
 131 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 132 {
 133         return syscall(__NR_sys_io_uring_setup, entries, p);
 134 }
 135
 136 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 137                           unsigned int min_complete, unsigned int flags)
 138 {
 139         return syscall(__NR_sys_io_uring_enter, s->ring_fd, to_submit,
 140                         min_complete, flags);
 141 }
 142
 143 static int gettid(void)
 144 {
 145         return syscall(__NR_gettid);
 146 }
 147
 148 static unsigned file_depth(struct submitter *s)
 149 {
 150         return (DEPTH + s->nr_files - 1) / s->nr_files;
 151 }
 152
 153 static void init_io(struct submitter *s, unsigned index)
 154 {
 155         struct io_uring_sqe *sqe = &s->sqes[index];
 156         unsigned long offset;
 157         struct file *f;
 158         long r;
 159
 160         if (do_nop) {
 161                 sqe->opcode = IORING_OP_NOP;
 162                 return;
 163         }
 164
 165         if (s->nr_files == 1) {
 166                 f = &s->files[0];
 167         } else {
 168                 f = &s->files[s->cur_file];
 169                 if (f->pending_ios >= file_depth(s)) {
 170                         s->cur_file++;
 171                         if (s->cur_file == s->nr_files)
 172                                 s->cur_file = 0;
 173                         f = &s->files[s->cur_file];
 174                 }
 175         }
 176         f->pending_ios++;
 177
 178         lrand48_r(&s->rand, &r);
 179         offset = (r % (f->max_blocks - 1)) * BS;
 180
 181         if (register_files) {
 182                 sqe->flags = IOSQE_FIXED_FILE;
 183                 sqe->fd = f->fixed_fd;
 184         } else {
 185                 sqe->flags = 0;
 186                 sqe->fd = f->real_fd;
 187         }
 188         if (fixedbufs) {
 189                 sqe->opcode = IORING_OP_READ_FIXED;
 190                 sqe->addr = s->iovecs[index].iov_base;
 191                 sqe->len = BS;
 192                 sqe->buf_index = index;
 193         } else {
 194                 sqe->opcode = IORING_OP_READV;
 195                 sqe->addr = &s->iovecs[index];
 196                 sqe->len = 1;
 197                 sqe->buf_index = 0;
 198         }
 199         sqe->ioprio = 0;
 200         sqe->off = offset;
 201         sqe->user_data = (unsigned long) f;
 202 }
 203
 204 static int prep_more_ios(struct submitter *s, int max_ios)
 205 {
 206         struct io_sq_ring *ring = &s->sq_ring;
 207         unsigned index, tail, next_tail, prepped = 0;
 208
 209         next_tail = tail = *ring->tail;
 210         do {
 211                 next_tail++;
 212                 read_barrier();
 213                 if (next_tail == *ring->head)
 214                         break;
 215
 216                 index = tail & sq_ring_mask;
 217                 init_io(s, index);
 218                 ring->array[index] = index;
 219                 prepped++;
 220                 tail = next_tail;
 221         } while (prepped < max_ios);
 222
 223         if (*ring->tail != tail) {
 224                 /* order tail store with writes to sqes above */
 225                 write_barrier();
 226                 *ring->tail = tail;
 227                 write_barrier();
 228         }
 229         return prepped;
 230 }
 231
 232 static int get_file_size(struct file *f)
 233 {
 234         struct stat st;
 235
 236         if (fstat(f->real_fd, &st) < 0)
 237                 return -1;
 238         if (S_ISBLK(st.st_mode)) {
 239                 unsigned long long bytes;
 240
 241                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 242                         return -1;
 243
 244                 f->max_blocks = bytes / BS;
 245                 return 0;
 246         } else if (S_ISREG(st.st_mode)) {
 247                 f->max_blocks = st.st_size / BS;
 248                 return 0;
 249         }
 250
 251         return -1;
 252 }
 253
 254 static int reap_events(struct submitter *s)
 255 {
 256         struct io_cq_ring *ring = &s->cq_ring;
 257         struct io_uring_cqe *cqe;
 258         unsigned head, reaped = 0;
 259
 260         head = *ring->head;
 261         do {
 262                 struct file *f;
 263
 264                 read_barrier();
 265                 if (head == *ring->tail)
 266                         break;
 267                 cqe = &ring->cqes[head & cq_ring_mask];
 268                 if (!do_nop) {
 269                         f = (struct file *) (uintptr_t) cqe->user_data;
 270                         f->pending_ios--;
 271                         if (cqe->res != BS) {
 272                                 printf("io: unexpected ret=%d\n", cqe->res);
 273                                 return -1;
 274                         }
 275                 }
 276                 if (cqe->flags & IOCQE_FLAG_CACHEHIT)
 277                         s->cachehit++;
 278                 else
 279                         s->cachemiss++;
 280                 reaped++;
 281                 head++;
 282         } while (1);
 283
 284         s->inflight -= reaped;
 285         *ring->head = head;
 286         write_barrier();
 287         return reaped;
 288 }
 289
 290 static void *submitter_fn(void *data)
 291 {
 292         struct submitter *s = data;
 293         struct io_sq_ring *ring = &s->sq_ring;
 294         int ret, prepped;
 295
 296         printf("submitter=%d\n", gettid());
 297
 298         srand48_r(pthread_self(), &s->rand);
 299
 300         prepped = 0;
 301         do {
 302                 int to_wait, to_submit, this_reap, to_prep;
 303
 304                 if (!prepped && s->inflight < DEPTH) {
 305                         to_prep = min(DEPTH - s->inflight, BATCH_SUBMIT);
 306                         prepped = prep_more_ios(s, to_prep);
 307                 }
 308                 s->inflight += prepped;
 309 submit_more:
 310                 to_submit = prepped;
 311 submit:
 312                 if (to_submit && (s->inflight + to_submit <= DEPTH))
 313                         to_wait = 0;
 314                 else
 315                         to_wait = min(s->inflight + to_submit, BATCH_COMPLETE);
 316
 317                 /*
 318                  * Only need to call io_uring_enter if we're not using SQ thread
 319                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
 320                  */
 321                 if (!sq_thread_poll || (*ring->flags & IORING_SQ_NEED_WAKEUP)) {
 322                         unsigned flags = 0;
 323
 324                         if (to_wait)
 325                                 flags = IORING_ENTER_GETEVENTS;
 326                         ret = io_uring_enter(s, to_submit, to_wait, flags);
 327                         s->calls++;
 328                 }
 329
 330                 /*
 331                  * For non SQ thread poll, we already got the events we needed
 332                  * through the io_uring_enter() above. For SQ thread poll, we
 333                  * need to loop here until we find enough events.
 334                  */
 335                 this_reap = 0;
 336                 do {
 337                         int r;
 338                         r = reap_events(s);
 339                         if (r == -1) {
 340                                 s->finish = 1;
 341                                 break;
 342                         } else if (r > 0)
 343                                 this_reap += r;
 344                 } while (sq_thread_poll && this_reap < to_wait);
 345                 s->reaps += this_reap;
 346
 347                 if (ret >= 0) {
 348                         if (!ret) {
 349                                 to_submit = 0;
 350                                 if (s->inflight)
 351                                         goto submit;
 352                                 continue;
 353                         } else if (ret < to_submit) {
 354                                 int diff = to_submit - ret;
 355
 356                                 s->done += ret;
 357                                 prepped -= diff;
 358                                 goto submit_more;
 359                         }
 360                         s->done += ret;
 361                         prepped = 0;
 362                         continue;
 363                 } else if (ret < 0) {
 364                         if (errno == EAGAIN) {
 365                                 if (s->finish)
 366                                         break;
 367                                 if (this_reap)
 368                                         goto submit;
 369                                 to_submit = 0;
 370                                 goto submit;
 371                         }
 372                         printf("io_submit: %s\n", strerror(errno));
 373                         break;
 374                 }
 375         } while (!s->finish);
 376
 377         finish = 1;
 378         return NULL;
 379 }
 380
 381 static void sig_int(int sig)
 382 {
 383         printf("Exiting on signal %d\n", sig);
 384         submitters[0].finish = 1;
 385         finish = 1;
 386 }
 387
 388 static void arm_sig_int(void)
 389 {
 390         struct sigaction act;
 391
 392         memset(&act, 0, sizeof(act));
 393         act.sa_handler = sig_int;
 394         act.sa_flags = SA_RESTART;
 395         sigaction(SIGINT, &act, NULL);
 396 }
 397
 398 static int setup_ring(struct submitter *s)
 399 {
 400         struct io_sq_ring *sring = &s->sq_ring;
 401         struct io_cq_ring *cring = &s->cq_ring;
 402         struct io_uring_params p;
 403         int ret, fd;
 404         void *ptr;
 405
 406         memset(&p, 0, sizeof(p));
 407
 408         if (polled && !do_nop)
 409                 p.flags |= IORING_SETUP_IOPOLL;
 410         if (sq_thread_poll) {
 411                 p.flags |= IORING_SETUP_SQPOLL;
 412                 if (sq_thread_cpu != -1) {
 413                         p.flags |= IORING_SETUP_SQ_AFF;
 414                         p.sq_thread_cpu = sq_thread_cpu;
 415                 }
 416         }
 417
 418         fd = io_uring_setup(DEPTH, &p);
 419         if (fd < 0) {
 420                 perror("io_uring_setup");
 421                 return 1;
 422         }
 423         s->ring_fd = fd;
 424
 425         if (fixedbufs) {
 426                 ret = io_uring_register_buffers(s);
 427                 if (ret < 0) {
 428                         perror("io_uring_register_buffers");
 429                         return 1;
 430                 }
 431         }
 432
 433         if (register_files) {
 434                 ret = io_uring_register_files(s);
 435                 if (ret < 0) {
 436                         perror("io_uring_register_files");
 437                         return 1;
 438                 }
 439         }
 440
 441         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 442                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 443                         IORING_OFF_SQ_RING);
 444         printf("sq_ring ptr = 0x%p\n", ptr);
 445         sring->head = ptr + p.sq_off.head;
 446         sring->tail = ptr + p.sq_off.tail;
 447         sring->ring_mask = ptr + p.sq_off.ring_mask;
 448         sring->ring_entries = ptr + p.sq_off.ring_entries;
 449         sring->flags = ptr + p.sq_off.flags;
 450         sring->array = ptr + p.sq_off.array;
 451         sq_ring_mask = *sring->ring_mask;
 452
 453         s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
 454                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 455                         IORING_OFF_SQES);
 456         printf("sqes ptr    = 0x%p\n", s->sqes);
 457
 458         ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
 459                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 460                         IORING_OFF_CQ_RING);
 461         printf("cq_ring ptr = 0x%p\n", ptr);
 462         cring->head = ptr + p.cq_off.head;
 463         cring->tail = ptr + p.cq_off.tail;
 464         cring->ring_mask = ptr + p.cq_off.ring_mask;
 465         cring->ring_entries = ptr + p.cq_off.ring_entries;
 466         cring->cqes = ptr + p.cq_off.cqes;
 467         cq_ring_mask = *cring->ring_mask;
 468         return 0;
 469 }
 470
 471 static void file_depths(char *buf)
 472 {
 473         struct submitter *s = &submitters[0];
 474         char *p;
 475         int i;
 476
 477         buf[0] = '\0';
 478         p = buf;
 479         for (i = 0; i < s->nr_files; i++) {
 480                 struct file *f = &s->files[i];
 481
 482                 if (i + 1 == s->nr_files)
 483                         p += sprintf(p, "%d", f->pending_ios);
 484                 else
 485                         p += sprintf(p, "%d, ", f->pending_ios);
 486         }
 487 }
 488
 489 int main(int argc, char *argv[])
 490 {
 491         struct submitter *s = &submitters[0];
 492         unsigned long done, calls, reap, cache_hit, cache_miss;
 493         int err, i, flags, fd;
 494         char *fdepths;
 495         void *ret;
 496
 497         if (!do_nop && argc < 2) {
 498                 printf("%s: filename\n", argv[0]);
 499                 return 1;
 500         }
 501
 502         flags = O_RDONLY | O_NOATIME;
 503         if (!buffered)
 504                 flags |= O_DIRECT;
 505
 506         i = 1;
 507         while (!do_nop && i < argc) {
 508                 struct file *f = &s->files[s->nr_files];
 509
 510                 fd = open(argv[i], flags);
 511                 if (fd < 0) {
 512                         perror("open");
 513                         return 1;
 514                 }
 515                 f->real_fd = fd;
 516                 if (get_file_size(f)) {
 517                         printf("failed getting size of device/file\n");
 518                         return 1;
 519                 }
 520                 if (f->max_blocks <= 1) {
 521                         printf("Zero file/device size?\n");
 522                         return 1;
 523                 }
 524                 f->max_blocks--;
 525
 526                 printf("Added file %s\n", argv[i]);
 527                 s->nr_files++;
 528                 i++;
 529         }
 530
 531         if (fixedbufs) {
 532                 struct rlimit rlim;
 533
 534                 rlim.rlim_cur = RLIM_INFINITY;
 535                 rlim.rlim_max = RLIM_INFINITY;
 536                 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 537                         perror("setrlimit");
 538                         return 1;
 539                 }
 540         }
 541
 542         arm_sig_int();
 543
 544         for (i = 0; i < DEPTH; i++) {
 545                 void *buf;
 546
 547                 if (posix_memalign(&buf, BS, BS)) {
 548                         printf("failed alloc\n");
 549                         return 1;
 550                 }
 551                 s->iovecs[i].iov_base = buf;
 552                 s->iovecs[i].iov_len = BS;
 553         }
 554
 555         err = setup_ring(s);
 556         if (err) {
 557                 printf("ring setup failed: %s, %d\n", strerror(errno), err);
 558                 return 1;
 559         }
 560         printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered);
 561         printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", DEPTH, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
 562
 563         pthread_create(&s->thread, NULL, submitter_fn, s);
 564
 565         fdepths = malloc(8 * s->nr_files);
 566         cache_hit = cache_miss = reap = calls = done = 0;
 567         do {
 568                 unsigned long this_done = 0;
 569                 unsigned long this_reap = 0;
 570                 unsigned long this_call = 0;
 571                 unsigned long this_cache_hit = 0;
 572                 unsigned long this_cache_miss = 0;
 573                 unsigned long rpc = 0, ipc = 0;
 574                 double hit = 0.0;
 575
 576                 sleep(1);
 577                 this_done += s->done;
 578                 this_call += s->calls;
 579                 this_reap += s->reaps;
 580                 this_cache_hit += s->cachehit;
 581                 this_cache_miss += s->cachemiss;
 582                 if (this_cache_hit && this_cache_miss) {
 583                         unsigned long hits, total;
 584
 585                         hits = this_cache_hit - cache_hit;
 586                         total = hits + this_cache_miss - cache_miss;
 587                         hit = (double) hits / (double) total;
 588                         hit *= 100.0;
 589                 }
 590                 if (this_call - calls) {
 591                         rpc = (this_done - done) / (this_call - calls);
 592                         ipc = (this_reap - reap) / (this_call - calls);
 593                 } else
 594                         rpc = ipc = -1;
 595                 file_depths(fdepths);
 596                 printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s), Cachehit=%0.2f%%\n",
 597                                 this_done - done, rpc, ipc, s->inflight,
 598                                 fdepths, hit);
 599                 done = this_done;
 600                 calls = this_call;
 601                 reap = this_reap;
 602                 cache_hit = s->cachehit;
 603                 cache_miss = s->cachemiss;
 604         } while (!finish);
 605
 606         pthread_join(s->thread, &ret);
 607         close(s->ring_fd);
 608         free(fdepths);
 609         return 0;
 610 }