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