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