t/aio-ring.c

   1 /*
   2  * gcc -D_GNU_SOURCE -Wall -O2 -o aio-ring aio-ring.c  -lpthread -laio
   3  */
   4 #include <stdio.h>
   5 #include <errno.h>
   6 #include <assert.h>
   7 #include <stdlib.h>
   8 #include <stddef.h>
   9 #include <signal.h>
  10 #include <inttypes.h>
  11
  12 #include <sys/types.h>
  13 #include <sys/stat.h>
  14 #include <sys/ioctl.h>
  15 #include <sys/syscall.h>
  16 #include <sys/resource.h>
  17 #include <linux/fs.h>
  18 #include <fcntl.h>
  19 #include <unistd.h>
  20 #include <libaio.h>
  21 #include <string.h>
  22 #include <pthread.h>
  23 #include <sched.h>
  24
  25 #define IOCB_FLAG_HIPRI         (1 << 2)
  26
  27 #define IOCTX_FLAG_IOPOLL       (1 << 0)
  28 #define IOCTX_FLAG_SCQRING      (1 << 1)        /* Use SQ/CQ rings */
  29 #define IOCTX_FLAG_FIXEDBUFS    (1 << 2)
  30 #define IOCTX_FLAG_SQTHREAD     (1 << 3)        /* Use SQ thread */
  31 #define IOCTX_FLAG_SQWQ         (1 << 4)        /* Use SQ wq */
  32
  33 #define IOEV_RES2_CACHEHIT      (1 << 0)
  34
  35 #define barrier()       __asm__ __volatile__("": : :"memory")
  36
  37 #define min(a, b)               ((a < b) ? (a) : (b))
  38
  39 typedef uint64_t u64;
  40 typedef uint32_t u32;
  41 typedef uint16_t u16;
  42
  43 struct aio_sq_ring {
  44         union {
  45                 struct {
  46                         u32 head;
  47                         u32 tail;
  48                         u32 nr_events;
  49                         u16 sq_thread_cpu;
  50                         u64 iocbs;
  51                 };
  52                 u32 pad[16];
  53         };
  54         u32 array[0];
  55 };
  56
  57 struct aio_cq_ring {
  58         union {
  59                 struct {
  60                         u32 head;
  61                         u32 tail;
  62                         u32 nr_events;
  63                 };
  64                 struct io_event pad;
  65         };
  66         struct io_event events[0];
  67 };
  68
  69 #define IORING_FLAG_SUBMIT      (1 << 0)
  70 #define IORING_FLAG_GETEVENTS   (1 << 1)
  71
  72 #define DEPTH                   32
  73 #define RING_SIZE               (DEPTH + 1)
  74
  75 #define BATCH_SUBMIT            8
  76 #define BATCH_COMPLETE          8
  77
  78 #define BS                      4096
  79
  80 struct submitter {
  81         pthread_t thread;
  82         unsigned long max_blocks;
  83         io_context_t ioc;
  84         struct drand48_data rand;
  85         struct aio_sq_ring *sq_ring;
  86         struct iocb *iocbs;
  87         struct aio_cq_ring *cq_ring;
  88         int inflight;
  89         unsigned long reaps;
  90         unsigned long done;
  91         unsigned long calls;
  92         unsigned long cachehit, cachemiss;
  93         volatile int finish;
  94         char filename[128];
  95 };
  96
  97 static struct submitter submitters[1];
  98 static volatile int finish;
  99
 100 static int polled = 1;          /* use IO polling */
 101 static int fixedbufs = 1;       /* use fixed user buffers */
 102 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
 103 static int sq_thread = 0;       /* use kernel submission thread */
 104 static int sq_thread_cpu = 0;   /* pin above thread to this CPU */
 105
 106 static int io_setup2(unsigned int nr_events, unsigned int flags,
 107                      struct aio_sq_ring *sq_ring, struct aio_cq_ring *cq_ring,
 108                      io_context_t *ctx_idp)
 109 {
 110         return syscall(335, nr_events, flags, sq_ring, cq_ring, ctx_idp);
 111 }
 112
 113 static int io_ring_enter(io_context_t ctx, unsigned int to_submit,
 114                          unsigned int min_complete, unsigned int flags)
 115 {
 116         return syscall(336, ctx, to_submit, min_complete, flags);
 117 }
 118
 119 static int gettid(void)
 120 {
 121         return syscall(__NR_gettid);
 122 }
 123
 124 static void init_io(struct submitter *s, int fd, struct iocb *iocb)
 125 {
 126         unsigned long offset;
 127         long r;
 128
 129         lrand48_r(&s->rand, &r);
 130         offset = (r % (s->max_blocks - 1)) * BS;
 131
 132         iocb->aio_fildes = fd;
 133         iocb->aio_lio_opcode = IO_CMD_PREAD;
 134         iocb->u.c.offset = offset;
 135         if (polled)
 136                 iocb->u.c.flags = IOCB_FLAG_HIPRI;
 137         if (!fixedbufs)
 138                 iocb->u.c.nbytes = BS;
 139 }
 140
 141 static int prep_more_ios(struct submitter *s, int fd, int max_ios)
 142 {
 143         struct aio_sq_ring *ring = s->sq_ring;
 144         u32 tail, next_tail, prepped = 0;
 145
 146         next_tail = tail = ring->tail;
 147         do {
 148                 next_tail++;
 149                 if (next_tail == ring->nr_events)
 150                         next_tail = 0;
 151
 152                 barrier();
 153                 if (next_tail == ring->head)
 154                         break;
 155
 156                 init_io(s, fd, &s->iocbs[tail]);
 157                 s->sq_ring->array[tail] = tail;
 158                 prepped++;
 159                 tail = next_tail;
 160         } while (prepped < max_ios);
 161
 162         if (ring->tail != tail) {
 163                 /* order tail store with writes to iocbs above */
 164                 barrier();
 165                 ring->tail = tail;
 166                 barrier();
 167         }
 168         return prepped;
 169 }
 170
 171 static int get_file_size(int fd, unsigned long *blocks)
 172 {
 173         struct stat st;
 174
 175         if (fstat(fd, &st) < 0)
 176                 return -1;
 177         if (S_ISBLK(st.st_mode)) {
 178                 unsigned long long bytes;
 179
 180                 if (ioctl(fd, BLKGETSIZE64, &bytes) != 0)
 181                         return -1;
 182
 183                 *blocks = bytes / BS;
 184                 return 0;
 185         } else if (S_ISREG(st.st_mode)) {
 186                 *blocks = st.st_size / BS;
 187                 return 0;
 188         }
 189
 190         return -1;
 191 }
 192
 193 static int reap_events(struct submitter *s)
 194 {
 195         struct aio_cq_ring *ring = s->cq_ring;
 196         struct io_event *ev;
 197         u32 head, reaped = 0;
 198
 199         head = ring->head;
 200         do {
 201                 barrier();
 202                 if (head == ring->tail)
 203                         break;
 204                 ev = &ring->events[head];
 205                 if (ev->res != BS) {
 206                         struct iocb *iocb = ev->obj;
 207
 208                         printf("io: unexpected ret=%ld\n", ev->res);
 209                         printf("offset=%lu, size=%lu\n", (unsigned long) iocb->u.c.offset, (unsigned long) iocb->u.c.nbytes);
 210                         return -1;
 211                 }
 212                 if (ev->res2 & IOEV_RES2_CACHEHIT)
 213                         s->cachehit++;
 214                 else
 215                         s->cachemiss++;
 216                 reaped++;
 217                 head++;
 218                 if (head == ring->nr_events)
 219                         head = 0;
 220         } while (1);
 221
 222         s->inflight -= reaped;
 223         ring->head = head;
 224         barrier();
 225         return reaped;
 226 }
 227
 228 static void *submitter_fn(void *data)
 229 {
 230         struct submitter *s = data;
 231         int fd, ret, prepped, flags;
 232
 233         printf("submitter=%d\n", gettid());
 234
 235         flags = O_RDONLY;
 236         if (!buffered)
 237                 flags |= O_DIRECT;
 238         fd = open(s->filename, flags);
 239         if (fd < 0) {
 240                 perror("open");
 241                 goto done;
 242         }
 243
 244         if (get_file_size(fd, &s->max_blocks)) {
 245                 printf("failed getting size of device/file\n");
 246                 goto err;
 247         }
 248         if (!s->max_blocks) {
 249                 printf("Zero file/device size?\n");
 250                 goto err;
 251         }
 252
 253         s->max_blocks--;
 254
 255         srand48_r(pthread_self(), &s->rand);
 256
 257         prepped = 0;
 258         do {
 259                 int to_wait, flags, to_submit, this_reap;
 260
 261                 if (!prepped && s->inflight < DEPTH)
 262                         prepped = prep_more_ios(s, fd, min(DEPTH - s->inflight, BATCH_SUBMIT));
 263                 s->inflight += prepped;
 264 submit_more:
 265                 to_submit = prepped;
 266 submit:
 267                 if (s->inflight + BATCH_SUBMIT < DEPTH)
 268                         to_wait = 0;
 269                 else
 270                         to_wait = min(s->inflight + to_submit, BATCH_COMPLETE);
 271
 272                 flags = IORING_FLAG_GETEVENTS;
 273                 if (to_submit)
 274                         flags |= IORING_FLAG_SUBMIT;
 275
 276                 ret = io_ring_enter(s->ioc, to_submit, to_wait, flags);
 277                 s->calls++;
 278
 279                 this_reap = reap_events(s);
 280                 if (this_reap == -1)
 281                         break;
 282                 s->reaps += this_reap;
 283
 284                 if (ret >= 0) {
 285                         if (!ret) {
 286                                 to_submit = 0;
 287                                 if (s->inflight)
 288                                         goto submit;
 289                                 continue;
 290                         } else if (ret < to_submit) {
 291                                 int diff = to_submit - ret;
 292
 293                                 s->done += ret;
 294                                 prepped -= diff;
 295                                 goto submit_more;
 296                         }
 297                         s->done += ret;
 298                         prepped = 0;
 299                         continue;
 300                 } else if (ret < 0) {
 301                         if ((ret == -1 && errno == EAGAIN) || ret == -EAGAIN) {
 302                                 if (s->finish)
 303                                         break;
 304                                 if (this_reap)
 305                                         goto submit;
 306                                 to_submit = 0;
 307                                 goto submit;
 308                         }
 309                         if (ret == -1)
 310                                 printf("io_submit: %s\n", strerror(errno));
 311                         else
 312                                 printf("io_submit: %s\n", strerror(-ret));
 313                         break;
 314                 }
 315         } while (!s->finish);
 316 err:
 317         close(fd);
 318 done:
 319         finish = 1;
 320         return NULL;
 321 }
 322
 323 static void sig_int(int sig)
 324 {
 325         printf("Exiting on signal %d\n", sig);
 326         submitters[0].finish = 1;
 327         finish = 1;
 328 }
 329
 330 static void arm_sig_int(void)
 331 {
 332         struct sigaction act;
 333
 334         memset(&act, 0, sizeof(act));
 335         act.sa_handler = sig_int;
 336         act.sa_flags = SA_RESTART;
 337         sigaction(SIGINT, &act, NULL);
 338 }
 339
 340 int main(int argc, char *argv[])
 341 {
 342         struct submitter *s = &submitters[0];
 343         unsigned long done, calls, reap, cache_hit, cache_miss;
 344         int flags = 0, err;
 345         int j;
 346         size_t size;
 347         void *p, *ret;
 348         struct rlimit rlim;
 349
 350         if (argc < 2) {
 351                 printf("%s: filename\n", argv[0]);
 352                 return 1;
 353         }
 354
 355         rlim.rlim_cur = RLIM_INFINITY;
 356         rlim.rlim_max = RLIM_INFINITY;
 357         if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 358                 perror("setrlimit");
 359                 return 1;
 360         }
 361
 362         arm_sig_int();
 363
 364         size = sizeof(struct iocb) * RING_SIZE;
 365         if (posix_memalign(&p, 4096, size))
 366                 return 1;
 367         memset(p, 0, size);
 368         s->iocbs = p;
 369
 370         size = sizeof(struct aio_sq_ring) + RING_SIZE * sizeof(u32);
 371         if (posix_memalign(&p, 4096, size))
 372                 return 1;
 373         s->sq_ring = p;
 374         memset(p, 0, size);
 375         s->sq_ring->nr_events = RING_SIZE;
 376         s->sq_ring->iocbs = (u64) s->iocbs;
 377
 378         /* CQ ring must be twice as big */
 379         size = sizeof(struct aio_cq_ring) +
 380                         2 * RING_SIZE * sizeof(struct io_event);
 381         if (posix_memalign(&p, 4096, size))
 382                 return 1;
 383         s->cq_ring = p;
 384         memset(p, 0, size);
 385         s->cq_ring->nr_events = 2 * RING_SIZE;
 386
 387         for (j = 0; j < RING_SIZE; j++) {
 388                 struct iocb *iocb = &s->iocbs[j];
 389
 390                 if (posix_memalign(&iocb->u.c.buf, BS, BS)) {
 391                         printf("failed alloc\n");
 392                         return 1;
 393                 }
 394                 iocb->u.c.nbytes = BS;
 395         }
 396
 397         flags = IOCTX_FLAG_SCQRING;
 398         if (polled)
 399                 flags |= IOCTX_FLAG_IOPOLL;
 400         if (fixedbufs)
 401                 flags |= IOCTX_FLAG_FIXEDBUFS;
 402         if (buffered)
 403                 flags |= IOCTX_FLAG_SQWQ;
 404         else if (sq_thread) {
 405                 flags |= IOCTX_FLAG_SQTHREAD;
 406                 s->sq_ring->sq_thread_cpu = sq_thread_cpu;
 407         }
 408
 409         err = io_setup2(RING_SIZE, flags, s->sq_ring, s->cq_ring, &s->ioc);
 410         if (err) {
 411                 printf("ctx_init failed: %s, %d\n", strerror(errno), err);
 412                 return 1;
 413         }
 414         printf("polled=%d, fixedbufs=%d, buffered=%d\n", polled, fixedbufs, buffered);
 415         printf("  QD=%d, sq_ring=%d, cq_ring=%d\n", DEPTH, s->sq_ring->nr_events, s->cq_ring->nr_events);
 416         strcpy(s->filename, argv[1]);
 417
 418         pthread_create(&s->thread, NULL, submitter_fn, s);
 419
 420         cache_hit = cache_miss = reap = calls = done = 0;
 421         do {
 422                 unsigned long this_done = 0;
 423                 unsigned long this_reap = 0;
 424                 unsigned long this_call = 0;
 425                 unsigned long this_cache_hit = 0;
 426                 unsigned long this_cache_miss = 0;
 427                 unsigned long rpc = 0, ipc = 0;
 428                 double hit = 0.0;
 429
 430                 sleep(1);
 431                 this_done += s->done;
 432                 this_call += s->calls;
 433                 this_reap += s->reaps;
 434                 this_cache_hit += s->cachehit;
 435                 this_cache_miss += s->cachemiss;
 436                 if (this_cache_hit && this_cache_miss) {
 437                         unsigned long hits, total;
 438
 439                         hits = this_cache_hit - cache_hit;
 440                         total = hits + this_cache_miss - cache_miss;
 441                         hit = (double) hits / (double) total;
 442                         hit *= 100.0;
 443                 }
 444                 if (this_call - calls) {
 445                         rpc = (this_done - done) / (this_call - calls);
 446                         ipc = (this_reap - reap) / (this_call - calls);
 447                 }
 448                 printf("IOPS=%lu, IOS/call=%lu/%lu, inflight=%u (head=%d tail=%d), Cachehit=%0.2f%%\n",
 449                                 this_done - done, rpc, ipc, s->inflight,
 450                                 s->cq_ring->head, s->cq_ring->tail, hit);
 451                 done = this_done;
 452                 calls = this_call;
 453                 reap = this_reap;
 454                 cache_hit = s->cachehit;
 455                 cache_miss = s->cachemiss;
 456         } while (!finish);
 457
 458         pthread_join(s->thread, &ret);
 459         return 0;
 460 }