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