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 fd;
  63 };
  64
  65 struct submitter {
  66         pthread_t thread;
  67         int ring_fd;
  68         struct drand48_data rand;
  69         struct io_sq_ring sq_ring;
  70         struct io_uring_sqe *sqes;
  71         struct iovec iovecs[DEPTH];
  72         struct io_cq_ring cq_ring;
  73         int inflight;
  74         unsigned long reaps;
  75         unsigned long done;
  76         unsigned long calls;
  77         unsigned long cachehit, cachemiss;
  78         volatile int finish;
  79
  80         struct file files[MAX_FDS];
  81         unsigned nr_files;
  82         unsigned cur_file;
  83 };
  84
  85 static struct submitter submitters[1];
  86 static volatile int finish;
  87
  88 static int polled = 1;          /* use IO polling */
  89 static int fixedbufs = 1;       /* use fixed user buffers */
  90 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
  91 static int sq_thread_poll = 0;  /* use kernel submission/poller thread */
  92 static int sq_thread_cpu = -1;  /* pin above thread to this CPU */
  93
  94 static int io_uring_register_buffers(struct submitter *s)
  95 {
  96         struct io_uring_register_buffers reg = {
  97                 .iovecs = s->iovecs,
  98                 .nr_iovecs = DEPTH
  99         };
 100
 101         return syscall(__NR_sys_io_uring_register, s->ring_fd,
 102                         IORING_REGISTER_BUFFERS, &reg);
 103 }
 104
 105 static int io_uring_register_files(struct submitter *s)
 106 {
 107         struct io_uring_register_files reg;
 108         int i, ret;
 109
 110         reg.fds = calloc(s->nr_files, sizeof(int));
 111         for (i = 0; i < s->nr_files; i++)
 112                 reg.fds[i] = s->files[i].fd;
 113         reg.nr_fds = s->nr_files;
 114
 115         ret = syscall(__NR_sys_io_uring_register, s->ring_fd,
 116                         IORING_REGISTER_FILES, &reg);
 117         free(reg.fds);
 118         return ret;
 119 }
 120
 121 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 122 {
 123         return syscall(__NR_sys_io_uring_setup, entries, p);
 124 }
 125
 126 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 127                           unsigned int min_complete, unsigned int flags)
 128 {
 129         return syscall(__NR_sys_io_uring_enter, s->ring_fd, to_submit,
 130                         min_complete, flags);
 131 }
 132
 133 static int gettid(void)
 134 {
 135         return syscall(__NR_gettid);
 136 }
 137
 138 static unsigned file_depth(struct submitter *s)
 139 {
 140         return (DEPTH + s->nr_files - 1) / s->nr_files;
 141 }
 142
 143 static void init_io(struct submitter *s, unsigned index)
 144 {
 145         struct io_uring_sqe *sqe = &s->sqes[index];
 146         unsigned long offset;
 147         struct file *f;
 148         long r;
 149
 150         if (s->nr_files == 1) {
 151                 f = &s->files[0];
 152         } else {
 153                 f = &s->files[s->cur_file];
 154                 if (f->pending_ios >= file_depth(s)) {
 155                         s->cur_file++;
 156                         if (s->cur_file == s->nr_files)
 157                                 s->cur_file = 0;
 158                 }
 159         }
 160         f->pending_ios++;
 161
 162         lrand48_r(&s->rand, &r);
 163         offset = (r % (f->max_blocks - 1)) * BS;
 164
 165         sqe->flags = IOSQE_FIXED_FILE;
 166         sqe->opcode = IORING_OP_READV;
 167         if (fixedbufs) {
 168                 sqe->addr = s->iovecs[index].iov_base;
 169                 sqe->len = BS;
 170                 sqe->buf_index = index;
 171                 sqe->flags |= IOSQE_FIXED_BUFFER;
 172         } else {
 173                 sqe->addr = &s->iovecs[index];
 174                 sqe->len = 1;
 175                 sqe->buf_index = 0;
 176         }
 177         sqe->ioprio = 0;
 178         sqe->fd = f->fd;
 179         sqe->off = offset;
 180         sqe->data = (unsigned long) f;
 181 }
 182
 183 static int prep_more_ios(struct submitter *s, int max_ios)
 184 {
 185         struct io_sq_ring *ring = &s->sq_ring;
 186         unsigned index, tail, next_tail, prepped = 0;
 187
 188         next_tail = tail = *ring->tail;
 189         do {
 190                 next_tail++;
 191                 barrier();
 192                 if (next_tail == *ring->head)
 193                         break;
 194
 195                 index = tail & sq_ring_mask;
 196                 init_io(s, index);
 197                 ring->array[index] = index;
 198                 prepped++;
 199                 tail = next_tail;
 200         } while (prepped < max_ios);
 201
 202         if (*ring->tail != tail) {
 203                 /* order tail store with writes to sqes above */
 204                 barrier();
 205                 *ring->tail = tail;
 206                 barrier();
 207         }
 208         return prepped;
 209 }
 210
 211 static int get_file_size(struct file *f)
 212 {
 213         struct stat st;
 214
 215         if (fstat(f->fd, &st) < 0)
 216                 return -1;
 217         if (S_ISBLK(st.st_mode)) {
 218                 unsigned long long bytes;
 219
 220                 if (ioctl(f->fd, BLKGETSIZE64, &bytes) != 0)
 221                         return -1;
 222
 223                 f->max_blocks = bytes / BS;
 224                 return 0;
 225         } else if (S_ISREG(st.st_mode)) {
 226                 f->max_blocks = st.st_size / BS;
 227                 return 0;
 228         }
 229
 230         return -1;
 231 }
 232
 233 static int reap_events(struct submitter *s)
 234 {
 235         struct io_cq_ring *ring = &s->cq_ring;
 236         struct io_uring_cqe *cqe;
 237         unsigned head, reaped = 0;
 238
 239         head = *ring->head;
 240         do {
 241                 struct file *f;
 242
 243                 barrier();
 244                 if (head == *ring->tail)
 245                         break;
 246                 cqe = &ring->cqes[head & cq_ring_mask];
 247                 f = (struct file *) cqe->data;
 248                 f->pending_ios--;
 249                 if (cqe->res != BS) {
 250                         printf("io: unexpected ret=%d\n", cqe->res);
 251                         return -1;
 252                 }
 253                 if (cqe->flags & IOCQE_FLAG_CACHEHIT)
 254                         s->cachehit++;
 255                 else
 256                         s->cachemiss++;
 257                 reaped++;
 258                 head++;
 259         } while (1);
 260
 261         s->inflight -= reaped;
 262         *ring->head = head;
 263         barrier();
 264         return reaped;
 265 }
 266
 267 static void *submitter_fn(void *data)
 268 {
 269         struct submitter *s = data;
 270         struct io_sq_ring *ring = &s->sq_ring;
 271         int ret, prepped;
 272
 273         printf("submitter=%d\n", gettid());
 274
 275         srand48_r(pthread_self(), &s->rand);
 276
 277         prepped = 0;
 278         do {
 279                 int to_wait, to_submit, this_reap, to_prep;
 280
 281                 if (!prepped && s->inflight < DEPTH) {
 282                         to_prep = min(DEPTH - s->inflight, BATCH_SUBMIT);
 283                         prepped = prep_more_ios(s, to_prep);
 284                 }
 285                 s->inflight += prepped;
 286 submit_more:
 287                 to_submit = prepped;
 288 submit:
 289                 if (s->inflight + BATCH_SUBMIT < DEPTH)
 290                         to_wait = 0;
 291                 else
 292                         to_wait = min(s->inflight + to_submit, BATCH_COMPLETE);
 293
 294                 /*
 295                  * Only need to call io_uring_enter if we're not using SQ thread
 296                  * poll, or if IORING_SQ_NEED_WAKEUP is set.
 297                  */
 298                 if (!sq_thread_poll || (*ring->flags & IORING_SQ_NEED_WAKEUP)) {
 299                         ret = io_uring_enter(s, to_submit, to_wait,
 300                                                 IORING_ENTER_GETEVENTS);
 301                         s->calls++;
 302                 }
 303
 304                 /*
 305                  * For non SQ thread poll, we already got the events we needed
 306                  * through the io_uring_enter() above. For SQ thread poll, we
 307                  * need to loop here until we find enough events.
 308                  */
 309                 this_reap = 0;
 310                 do {
 311                         int r;
 312                         r = reap_events(s);
 313                         if (r == -1)
 314                                 break;
 315                         else if (r > 0)
 316                                 this_reap += r;
 317                 } while (sq_thread_poll && this_reap < to_wait);
 318                 s->reaps += this_reap;
 319
 320                 if (ret >= 0) {
 321                         if (!ret) {
 322                                 to_submit = 0;
 323                                 if (s->inflight)
 324                                         goto submit;
 325                                 continue;
 326                         } else if (ret < to_submit) {
 327                                 int diff = to_submit - ret;
 328
 329                                 s->done += ret;
 330                                 prepped -= diff;
 331                                 goto submit_more;
 332                         }
 333                         s->done += ret;
 334                         prepped = 0;
 335                         continue;
 336                 } else if (ret < 0) {
 337                         if (errno == EAGAIN) {
 338                                 if (s->finish)
 339                                         break;
 340                                 if (this_reap)
 341                                         goto submit;
 342                                 to_submit = 0;
 343                                 goto submit;
 344                         }
 345                         printf("io_submit: %s\n", strerror(errno));
 346                         break;
 347                 }
 348         } while (!s->finish);
 349
 350         finish = 1;
 351         return NULL;
 352 }
 353
 354 static void sig_int(int sig)
 355 {
 356         printf("Exiting on signal %d\n", sig);
 357         submitters[0].finish = 1;
 358         finish = 1;
 359 }
 360
 361 static void arm_sig_int(void)
 362 {
 363         struct sigaction act;
 364
 365         memset(&act, 0, sizeof(act));
 366         act.sa_handler = sig_int;
 367         act.sa_flags = SA_RESTART;
 368         sigaction(SIGINT, &act, NULL);
 369 }
 370
 371 static int setup_ring(struct submitter *s)
 372 {
 373         struct io_sq_ring *sring = &s->sq_ring;
 374         struct io_cq_ring *cring = &s->cq_ring;
 375         struct io_uring_params p;
 376         int ret, fd;
 377         void *ptr;
 378
 379         memset(&p, 0, sizeof(p));
 380
 381         if (polled)
 382                 p.flags |= IORING_SETUP_IOPOLL;
 383         if (sq_thread_poll) {
 384                 p.flags |= IORING_SETUP_SQPOLL;
 385                 if (sq_thread_cpu != -1)
 386                         p.flags |= IORING_SETUP_SQ_AFF;
 387         }
 388
 389         fd = io_uring_setup(DEPTH, &p);
 390         if (fd < 0) {
 391                 perror("io_uring_setup");
 392                 return 1;
 393         }
 394         s->ring_fd = fd;
 395
 396         if (fixedbufs) {
 397                 ret = io_uring_register_buffers(s);
 398                 if (ret < 0) {
 399                         perror("io_uring_register_buffers");
 400                         return 1;
 401                 }
 402         }
 403
 404         ret = io_uring_register_files(s);
 405         if (ret < 0) {
 406                 perror("io_uring_register_files");
 407                 return 1;
 408         }
 409
 410         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 411                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 412                         IORING_OFF_SQ_RING);
 413         printf("sq_ring ptr = 0x%p\n", ptr);
 414         sring->head = ptr + p.sq_off.head;
 415         sring->tail = ptr + p.sq_off.tail;
 416         sring->ring_mask = ptr + p.sq_off.ring_mask;
 417         sring->ring_entries = ptr + p.sq_off.ring_entries;
 418         sring->flags = ptr + p.sq_off.flags;
 419         sring->array = ptr + p.sq_off.array;
 420         sq_ring_mask = *sring->ring_mask;
 421
 422         s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
 423                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 424                         IORING_OFF_SQES);
 425         printf("sqes ptr    = 0x%p\n", s->sqes);
 426
 427         ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
 428                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 429                         IORING_OFF_CQ_RING);
 430         printf("cq_ring ptr = 0x%p\n", ptr);
 431         cring->head = ptr + p.cq_off.head;
 432         cring->tail = ptr + p.cq_off.tail;
 433         cring->ring_mask = ptr + p.cq_off.ring_mask;
 434         cring->ring_entries = ptr + p.cq_off.ring_entries;
 435         cring->cqes = ptr + p.cq_off.cqes;
 436         cq_ring_mask = *cring->ring_mask;
 437         return 0;
 438 }
 439
 440 int main(int argc, char *argv[])
 441 {
 442         struct submitter *s = &submitters[0];
 443         unsigned long done, calls, reap, cache_hit, cache_miss;
 444         int err, i, flags, fd;
 445         struct rlimit rlim;
 446         void *ret;
 447
 448         if (argc < 2) {
 449                 printf("%s: filename\n", argv[0]);
 450                 return 1;
 451         }
 452
 453         flags = O_RDONLY | O_NOATIME;
 454         if (!buffered)
 455                 flags |= O_DIRECT;
 456
 457         i = 1;
 458         while (i < argc) {
 459                 struct file *f = &s->files[s->nr_files];
 460
 461                 fd = open(argv[i], flags);
 462                 if (fd < 0) {
 463                         perror("open");
 464                         return 1;
 465                 }
 466                 f->fd = fd;
 467                 if (get_file_size(f)) {
 468                         printf("failed getting size of device/file\n");
 469                         return 1;
 470                 }
 471                 if (f->max_blocks <= 1) {
 472                         printf("Zero file/device size?\n");
 473                         return 1;
 474                 }
 475                 f->max_blocks--;
 476
 477                 printf("Added file %s\n", argv[i]);
 478                 s->nr_files++;
 479                 i++;
 480         }
 481
 482         rlim.rlim_cur = RLIM_INFINITY;
 483         rlim.rlim_max = RLIM_INFINITY;
 484         if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 485                 perror("setrlimit");
 486                 return 1;
 487         }
 488
 489         arm_sig_int();
 490
 491         for (i = 0; i < DEPTH; i++) {
 492                 void *buf;
 493
 494                 if (posix_memalign(&buf, BS, BS)) {
 495                         printf("failed alloc\n");
 496                         return 1;
 497                 }
 498                 s->iovecs[i].iov_base = buf;
 499                 s->iovecs[i].iov_len = BS;
 500         }
 501
 502         err = setup_ring(s);
 503         if (err) {
 504                 printf("ring setup failed: %s, %d\n", strerror(errno), err);
 505                 return 1;
 506         }
 507         printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered);
 508         printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", DEPTH, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
 509
 510         pthread_create(&s->thread, NULL, submitter_fn, s);
 511
 512         cache_hit = cache_miss = reap = calls = done = 0;
 513         do {
 514                 unsigned long this_done = 0;
 515                 unsigned long this_reap = 0;
 516                 unsigned long this_call = 0;
 517                 unsigned long this_cache_hit = 0;
 518                 unsigned long this_cache_miss = 0;
 519                 unsigned long rpc = 0, ipc = 0;
 520                 double hit = 0.0;
 521
 522                 sleep(1);
 523                 this_done += s->done;
 524                 this_call += s->calls;
 525                 this_reap += s->reaps;
 526                 this_cache_hit += s->cachehit;
 527                 this_cache_miss += s->cachemiss;
 528                 if (this_cache_hit && this_cache_miss) {
 529                         unsigned long hits, total;
 530
 531                         hits = this_cache_hit - cache_hit;
 532                         total = hits + this_cache_miss - cache_miss;
 533                         hit = (double) hits / (double) total;
 534                         hit *= 100.0;
 535                 }
 536                 if (this_call - calls) {
 537                         rpc = (this_done - done) / (this_call - calls);
 538                         ipc = (this_reap - reap) / (this_call - calls);
 539                 } else
 540                         rpc = ipc = -1;
 541                 printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (head=%u tail=%u), Cachehit=%0.2f%%\n",
 542                                 this_done - done, rpc, ipc, s->inflight,
 543                                 *s->cq_ring.head, *s->cq_ring.tail, hit);
 544                 done = this_done;
 545                 calls = this_call;
 546                 reap = this_reap;
 547                 cache_hit = s->cachehit;
 548                 cache_miss = s->cachemiss;
 549         } while (!finish);
 550
 551         pthread_join(s->thread, &ret);
 552         close(s->ring_fd);
 553         return 0;
 554 }