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 barrier()       __asm__ __volatile__("": : :"memory")
  28
  29 #define min(a, b)               ((a < b) ? (a) : (b))
  30
  31 struct io_sq_ring {
  32         unsigned *head;
  33         unsigned *tail;
  34         unsigned *ring_mask;
  35         unsigned *ring_entries;
  36         unsigned *flags;
  37         unsigned *array;
  38 };
  39
  40 struct io_cq_ring {
  41         unsigned *head;
  42         unsigned *tail;
  43         unsigned *ring_mask;
  44         unsigned *ring_entries;
  45         struct io_uring_cqe *cqes;
  46 };
  47
  48 #define DEPTH                   128
  49
  50 #define BATCH_SUBMIT            32
  51 #define BATCH_COMPLETE          32
  52
  53 #define BS                      4096
  54
  55 #define MAX_FDS                 16
  56
  57 static unsigned sq_ring_mask, cq_ring_mask;
  58
  59 struct file {
  60         unsigned long max_blocks;
  61         unsigned pending_ios;
  62         int real_fd;
  63         int fixed_fd;
  64 };
  65
  66 struct submitter {
  67         pthread_t thread;
  68         int ring_fd;
  69         struct drand48_data rand;
  70         struct io_sq_ring sq_ring;
  71         struct io_uring_sqe *sqes;
  72         struct iovec iovecs[DEPTH];
  73         struct io_cq_ring cq_ring;
  74         int inflight;
  75         unsigned long reaps;
  76         unsigned long done;
  77         unsigned long calls;
  78         unsigned long cachehit, cachemiss;
  79         volatile int finish;
  80
  81         __s32 *fds;
  82
  83         struct file files[MAX_FDS];
  84         unsigned nr_files;
  85         unsigned cur_file;
  86 };
  87
  88 static struct submitter submitters[1];
  89 static volatile int finish;
  90
  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         struct io_uring_register_buffers reg = {
 102                 .iovecs = s->iovecs,
 103                 .nr_iovecs = DEPTH
 104         };
 105
 106         if (do_nop)
 107                 return 0;
 108
 109         return syscall(__NR_sys_io_uring_register, s->ring_fd,
 110                         IORING_REGISTER_BUFFERS, &reg);
 111 }
 112
 113 static int io_uring_register_files(struct submitter *s)
 114 {
 115         struct io_uring_register_files reg;
 116         int i;
 117
 118         if (do_nop)
 119                 return 0;
 120
 121         s->fds = calloc(s->nr_files, sizeof(__s32));
 122         for (i = 0; i < s->nr_files; i++) {
 123                 s->fds[i] = s->files[i].real_fd;
 124                 s->files[i].fixed_fd = i;
 125         }
 126         reg.fds = s->fds;
 127         reg.nr_fds = s->nr_files;
 128
 129         return syscall(__NR_sys_io_uring_register, s->ring_fd,
 130                         IORING_REGISTER_FILES, &reg);
 131 }
 132
 133 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 134 {
 135         return syscall(__NR_sys_io_uring_setup, entries, p);
 136 }
 137
 138 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 139                           unsigned int min_complete, unsigned int flags)
 140 {
 141         return syscall(__NR_sys_io_uring_enter, s->ring_fd, to_submit,
 142                         min_complete, flags);
 143 }
 144
 145 static int gettid(void)
 146 {
 147         return syscall(__NR_gettid);
 148 }
 149
 150 static unsigned file_depth(struct submitter *s)
 151 {
 152         return (DEPTH + s->nr_files - 1) / s->nr_files;
 153 }
 154
 155 static void init_io(struct submitter *s, unsigned index)
 156 {
 157         struct io_uring_sqe *sqe = &s->sqes[index];
 158         unsigned long offset;
 159         struct file *f;
 160         long r;
 161
 162         if (do_nop) {
 163                 sqe->opcode = IORING_OP_NOP;
 164                 return;
 165         }
 166
 167         if (s->nr_files == 1) {
 168                 f = &s->files[0];
 169         } else {
 170                 f = &s->files[s->cur_file];
 171                 if (f->pending_ios >= file_depth(s)) {
 172                         s->cur_file++;
 173                         if (s->cur_file == s->nr_files)
 174                                 s->cur_file = 0;
 175                         f = &s->files[s->cur_file];
 176                 }
 177         }
 178         f->pending_ios++;
 179
 180         lrand48_r(&s->rand, &r);
 181         offset = (r % (f->max_blocks - 1)) * BS;
 182
 183         if (register_files) {
 184                 sqe->flags = IOSQE_FIXED_FILE;
 185                 sqe->fd = f->fixed_fd;
 186         } else {
 187                 sqe->flags = 0;
 188                 sqe->fd = f->real_fd;
 189         }
 190         if (fixedbufs) {
 191                 sqe->opcode = IORING_OP_READ_FIXED;
 192                 sqe->addr = s->iovecs[index].iov_base;
 193                 sqe->len = BS;
 194                 sqe->buf_index = index;
 195         } else {
 196                 sqe->opcode = IORING_OP_READV;
 197                 sqe->addr = &s->iovecs[index];
 198                 sqe->len = 1;
 199                 sqe->buf_index = 0;
 200         }
 201         sqe->ioprio = 0;
 202         sqe->off = offset;
 203         sqe->user_data = (unsigned long) f;
 204 }
 205
 206 static int prep_more_ios(struct submitter *s, int max_ios)
 207 {
 208         struct io_sq_ring *ring = &s->sq_ring;
 209         unsigned index, tail, next_tail, prepped = 0;
 210
 211         next_tail = tail = *ring->tail;
 212         do {
 213                 next_tail++;
 214                 barrier();
 215                 if (next_tail == *ring->head)
 216                         break;
 217
 218                 index = tail & sq_ring_mask;
 219                 init_io(s, index);
 220                 ring->array[index] = index;
 221                 prepped++;
 222                 tail = next_tail;
 223         } while (prepped < max_ios);
 224
 225         if (*ring->tail != tail) {
 226                 /* order tail store with writes to sqes above */
 227                 barrier();
 228                 *ring->tail = tail;
 229                 barrier();
 230         }
 231         return prepped;
 232 }
 233
 234 static int get_file_size(struct file *f)
 235 {
 236         struct stat st;
 237
 238         if (fstat(f->real_fd, &st) < 0)
 239                 return -1;
 240         if (S_ISBLK(st.st_mode)) {
 241                 unsigned long long bytes;
 242
 243                 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
 244                         return -1;
 245
 246                 f->max_blocks = bytes / BS;
 247                 return 0;
 248         } else if (S_ISREG(st.st_mode)) {
 249                 f->max_blocks = st.st_size / BS;
 250                 return 0;
 251         }
 252
 253         return -1;
 254 }
 255
 256 static int reap_events(struct submitter *s)
 257 {
 258         struct io_cq_ring *ring = &s->cq_ring;
 259         struct io_uring_cqe *cqe;
 260         unsigned head, reaped = 0;
 261
 262         head = *ring->head;
 263         do {
 264                 struct file *f;
 265
 266                 barrier();
 267                 if (head == *ring->tail)
 268                         break;
 269                 cqe = &ring->cqes[head & cq_ring_mask];
 270                 if (!do_nop) {
 271                         f = (struct file *) (uintptr_t) cqe->user_data;
 272                         f->pending_ios--;
 273                         if (cqe->res != BS) {
 274                                 printf("io: unexpected ret=%d\n", cqe->res);
 275                                 return -1;
 276                         }
 277                 }
 278                 if (cqe->flags & IOCQE_FLAG_CACHEHIT)
 279                         s->cachehit++;
 280                 else
 281                         s->cachemiss++;
 282                 reaped++;
 283                 head++;
 284         } while (1);
 285
 286         s->inflight -= reaped;
 287         *ring->head = head;
 288         barrier();
 289         return reaped;
 290 }
 291
 292 static void *submitter_fn(void *data)
 293 {
 294         struct submitter *s = data;
 295         struct io_sq_ring *ring = &s->sq_ring;
 296         int ret, prepped;
 297
 298         printf("submitter=%d\n", gettid());
 299
 300         srand48_r(pthread_self(), &s->rand);
 301
 302         prepped = 0;
 303         do {
 304                 int to_wait, to_submit, this_reap, to_prep;
 305
 306                 if (!prepped && s->inflight < DEPTH) {
 307                         to_prep = min(DEPTH - s->inflight, BATCH_SUBMIT);
 308                         prepped = prep_more_ios(s, to_prep);
 309                 }
 310                 s->inflight += prepped;
 311 submit_more:
 312                 to_submit = prepped;
 313 submit:
 314                 if (to_submit && (s->inflight + to_submit <= DEPTH))
 315                         to_wait = 0;
 316                 else
 317                         to_wait = min(s->inflight + to_submit, BATCH_COMPLETE);
 318
 319                 /*
 320                  * Only need to call io_uring_enter if we're not using SQ thread
 321                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
 322                  */
 323                 if (!sq_thread_poll || (*ring->flags & IORING_SQ_NEED_WAKEUP)) {
 324                         unsigned flags = 0;
 325
 326                         if (to_wait)
 327                                 flags = IORING_ENTER_GETEVENTS;
 328                         ret = io_uring_enter(s, to_submit, to_wait, flags);
 329                         s->calls++;
 330                 }
 331
 332                 /*
 333                  * For non SQ thread poll, we already got the events we needed
 334                  * through the io_uring_enter() above. For SQ thread poll, we
 335                  * need to loop here until we find enough events.
 336                  */
 337                 this_reap = 0;
 338                 do {
 339                         int r;
 340                         r = reap_events(s);
 341                         if (r == -1) {
 342                                 s->finish = 1;
 343                                 break;
 344                         } else if (r > 0)
 345                                 this_reap += r;
 346                 } while (sq_thread_poll && this_reap < to_wait);
 347                 s->reaps += this_reap;
 348
 349                 if (ret >= 0) {
 350                         if (!ret) {
 351                                 to_submit = 0;
 352                                 if (s->inflight)
 353                                         goto submit;
 354                                 continue;
 355                         } else if (ret < to_submit) {
 356                                 int diff = to_submit - ret;
 357
 358                                 s->done += ret;
 359                                 prepped -= diff;
 360                                 goto submit_more;
 361                         }
 362                         s->done += ret;
 363                         prepped = 0;
 364                         continue;
 365                 } else if (ret < 0) {
 366                         if (errno == EAGAIN) {
 367                                 if (s->finish)
 368                                         break;
 369                                 if (this_reap)
 370                                         goto submit;
 371                                 to_submit = 0;
 372                                 goto submit;
 373                         }
 374                         printf("io_submit: %s\n", strerror(errno));
 375                         break;
 376                 }
 377         } while (!s->finish);
 378
 379         finish = 1;
 380         return NULL;
 381 }
 382
 383 static void sig_int(int sig)
 384 {
 385         printf("Exiting on signal %d\n", sig);
 386         submitters[0].finish = 1;
 387         finish = 1;
 388 }
 389
 390 static void arm_sig_int(void)
 391 {
 392         struct sigaction act;
 393
 394         memset(&act, 0, sizeof(act));
 395         act.sa_handler = sig_int;
 396         act.sa_flags = SA_RESTART;
 397         sigaction(SIGINT, &act, NULL);
 398 }
 399
 400 static int setup_ring(struct submitter *s)
 401 {
 402         struct io_sq_ring *sring = &s->sq_ring;
 403         struct io_cq_ring *cring = &s->cq_ring;
 404         struct io_uring_params p;
 405         int ret, fd;
 406         void *ptr;
 407
 408         memset(&p, 0, sizeof(p));
 409
 410         if (polled && !do_nop)
 411                 p.flags |= IORING_SETUP_IOPOLL;
 412         if (sq_thread_poll) {
 413                 p.flags |= IORING_SETUP_SQPOLL;
 414                 if (sq_thread_cpu != -1) {
 415                         p.flags |= IORING_SETUP_SQ_AFF;
 416                         p.sq_thread_cpu = sq_thread_cpu;
 417                 }
 418         }
 419
 420         fd = io_uring_setup(DEPTH, &p);
 421         if (fd < 0) {
 422                 perror("io_uring_setup");
 423                 return 1;
 424         }
 425         s->ring_fd = fd;
 426
 427         if (fixedbufs) {
 428                 ret = io_uring_register_buffers(s);
 429                 if (ret < 0) {
 430                         perror("io_uring_register_buffers");
 431                         return 1;
 432                 }
 433         }
 434
 435         if (register_files) {
 436                 ret = io_uring_register_files(s);
 437                 if (ret < 0) {
 438                         perror("io_uring_register_files");
 439                         return 1;
 440                 }
 441         }
 442
 443         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 444                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 445                         IORING_OFF_SQ_RING);
 446         printf("sq_ring ptr = 0x%p\n", ptr);
 447         sring->head = ptr + p.sq_off.head;
 448         sring->tail = ptr + p.sq_off.tail;
 449         sring->ring_mask = ptr + p.sq_off.ring_mask;
 450         sring->ring_entries = ptr + p.sq_off.ring_entries;
 451         sring->flags = ptr + p.sq_off.flags;
 452         sring->array = ptr + p.sq_off.array;
 453         sq_ring_mask = *sring->ring_mask;
 454
 455         s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
 456                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 457                         IORING_OFF_SQES);
 458         printf("sqes ptr    = 0x%p\n", s->sqes);
 459
 460         ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
 461                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 462                         IORING_OFF_CQ_RING);
 463         printf("cq_ring ptr = 0x%p\n", ptr);
 464         cring->head = ptr + p.cq_off.head;
 465         cring->tail = ptr + p.cq_off.tail;
 466         cring->ring_mask = ptr + p.cq_off.ring_mask;
 467         cring->ring_entries = ptr + p.cq_off.ring_entries;
 468         cring->cqes = ptr + p.cq_off.cqes;
 469         cq_ring_mask = *cring->ring_mask;
 470         return 0;
 471 }
 472
 473 static void file_depths(char *buf)
 474 {
 475         struct submitter *s = &submitters[0];
 476         char *p;
 477         int i;
 478
 479         buf[0] = '\0';
 480         p = buf;
 481         for (i = 0; i < s->nr_files; i++) {
 482                 struct file *f = &s->files[i];
 483
 484                 if (i + 1 == s->nr_files)
 485                         p += sprintf(p, "%d", f->pending_ios);
 486                 else
 487                         p += sprintf(p, "%d, ", f->pending_ios);
 488         }
 489 }
 490
 491 int main(int argc, char *argv[])
 492 {
 493         struct submitter *s = &submitters[0];
 494         unsigned long done, calls, reap, cache_hit, cache_miss;
 495         int err, i, flags, fd;
 496         char *fdepths;
 497         void *ret;
 498
 499         if (!do_nop && argc < 2) {
 500                 printf("%s: filename\n", argv[0]);
 501                 return 1;
 502         }
 503
 504         flags = O_RDONLY | O_NOATIME;
 505         if (!buffered)
 506                 flags |= O_DIRECT;
 507
 508         i = 1;
 509         while (!do_nop && i < argc) {
 510                 struct file *f = &s->files[s->nr_files];
 511
 512                 fd = open(argv[i], flags);
 513                 if (fd < 0) {
 514                         perror("open");
 515                         return 1;
 516                 }
 517                 f->real_fd = fd;
 518                 if (get_file_size(f)) {
 519                         printf("failed getting size of device/file\n");
 520                         return 1;
 521                 }
 522                 if (f->max_blocks <= 1) {
 523                         printf("Zero file/device size?\n");
 524                         return 1;
 525                 }
 526                 f->max_blocks--;
 527
 528                 printf("Added file %s\n", argv[i]);
 529                 s->nr_files++;
 530                 i++;
 531         }
 532
 533         if (fixedbufs) {
 534                 struct rlimit rlim;
 535
 536                 rlim.rlim_cur = RLIM_INFINITY;
 537                 rlim.rlim_max = RLIM_INFINITY;
 538                 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 539                         perror("setrlimit");
 540                         return 1;
 541                 }
 542         }
 543
 544         arm_sig_int();
 545
 546         for (i = 0; i < DEPTH; i++) {
 547                 void *buf;
 548
 549                 if (posix_memalign(&buf, BS, BS)) {
 550                         printf("failed alloc\n");
 551                         return 1;
 552                 }
 553                 s->iovecs[i].iov_base = buf;
 554                 s->iovecs[i].iov_len = BS;
 555         }
 556
 557         err = setup_ring(s);
 558         if (err) {
 559                 printf("ring setup failed: %s, %d\n", strerror(errno), err);
 560                 return 1;
 561         }
 562         printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered);
 563         printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", DEPTH, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
 564
 565         pthread_create(&s->thread, NULL, submitter_fn, s);
 566
 567         fdepths = malloc(8 * s->nr_files);
 568         cache_hit = cache_miss = reap = calls = done = 0;
 569         do {
 570                 unsigned long this_done = 0;
 571                 unsigned long this_reap = 0;
 572                 unsigned long this_call = 0;
 573                 unsigned long this_cache_hit = 0;
 574                 unsigned long this_cache_miss = 0;
 575                 unsigned long rpc = 0, ipc = 0;
 576                 double hit = 0.0;
 577
 578                 sleep(1);
 579                 this_done += s->done;
 580                 this_call += s->calls;
 581                 this_reap += s->reaps;
 582                 this_cache_hit += s->cachehit;
 583                 this_cache_miss += s->cachemiss;
 584                 if (this_cache_hit && this_cache_miss) {
 585                         unsigned long hits, total;
 586
 587                         hits = this_cache_hit - cache_hit;
 588                         total = hits + this_cache_miss - cache_miss;
 589                         hit = (double) hits / (double) total;
 590                         hit *= 100.0;
 591                 }
 592                 if (this_call - calls) {
 593                         rpc = (this_done - done) / (this_call - calls);
 594                         ipc = (this_reap - reap) / (this_call - calls);
 595                 } else
 596                         rpc = ipc = -1;
 597                 file_depths(fdepths);
 598                 printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s), Cachehit=%0.2f%%\n",
 599                                 this_done - done, rpc, ipc, s->inflight,
 600                                 fdepths, hit);
 601                 done = this_done;
 602                 calls = this_call;
 603                 reap = this_reap;
 604                 cache_hit = s->cachehit;
 605                 cache_miss = s->cachemiss;
 606         } while (!finish);
 607
 608         pthread_join(s->thread, &ret);
 609         close(s->ring_fd);
 610         free(fdepths);
 611         return 0;
 612 }