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