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
  74 #define BATCH_SUBMIT            8
  75 #define BATCH_COMPLETE          8
  76
  77 #define BS                      4096
  78
  79 static unsigned sq_ring_mask = DEPTH - 1;
  80 static unsigned cq_ring_mask = (2 * DEPTH) - 1;
  81
  82 struct submitter {
  83         pthread_t thread;
  84         unsigned long max_blocks;
  85         io_context_t ioc;
  86         struct drand48_data rand;
  87         struct aio_sq_ring *sq_ring;
  88         struct iocb *iocbs;
  89         struct aio_cq_ring *cq_ring;
  90         int inflight;
  91         unsigned long reaps;
  92         unsigned long done;
  93         unsigned long calls;
  94         unsigned long cachehit, cachemiss;
  95         volatile int finish;
  96         char filename[128];
  97 };
  98
  99 static struct submitter submitters[1];
 100 static volatile int finish;
 101
 102 static int polled = 1;          /* use IO polling */
 103 static int fixedbufs = 1;       /* use fixed user buffers */
 104 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
 105 static int sq_thread = 0;       /* use kernel submission thread */
 106 static int sq_thread_cpu = 0;   /* pin above thread to this CPU */
 107
 108 static int io_setup2(unsigned int nr_events, unsigned int flags,
 109                      struct aio_sq_ring *sq_ring, struct aio_cq_ring *cq_ring,
 110                      io_context_t *ctx_idp)
 111 {
 112         return syscall(335, nr_events, flags, sq_ring, cq_ring, ctx_idp);
 113 }
 114
 115 static int io_ring_enter(io_context_t ctx, unsigned int to_submit,
 116                          unsigned int min_complete, unsigned int flags)
 117 {
 118         return syscall(336, ctx, to_submit, min_complete, flags);
 119 }
 120
 121 static int gettid(void)
 122 {
 123         return syscall(__NR_gettid);
 124 }
 125
 126 static void init_io(struct submitter *s, int fd, struct iocb *iocb)
 127 {
 128         unsigned long offset;
 129         long r;
 130
 131         lrand48_r(&s->rand, &r);
 132         offset = (r % (s->max_blocks - 1)) * BS;
 133
 134         iocb->aio_fildes = fd;
 135         iocb->aio_lio_opcode = IO_CMD_PREAD;
 136         iocb->u.c.offset = offset;
 137         if (polled)
 138                 iocb->u.c.flags = IOCB_FLAG_HIPRI;
 139         if (!fixedbufs)
 140                 iocb->u.c.nbytes = BS;
 141 }
 142
 143 static int prep_more_ios(struct submitter *s, int fd, int max_ios)
 144 {
 145         struct aio_sq_ring *ring = s->sq_ring;
 146         u32 index, tail, next_tail, prepped = 0;
 147
 148         next_tail = tail = ring->tail;
 149         do {
 150                 next_tail++;
 151                 barrier();
 152                 if (next_tail == ring->head)
 153                         break;
 154
 155                 index = tail & sq_ring_mask;
 156                 init_io(s, fd, &s->iocbs[index]);
 157                 s->sq_ring->array[index] = index;
 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 & cq_ring_mask];
 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         } while (1);
 219
 220         s->inflight -= reaped;
 221         ring->head = head;
 222         barrier();
 223         return reaped;
 224 }
 225
 226 static void *submitter_fn(void *data)
 227 {
 228         struct submitter *s = data;
 229         int fd, ret, prepped, flags;
 230
 231         printf("submitter=%d\n", gettid());
 232
 233         flags = O_RDONLY;
 234         if (!buffered)
 235                 flags |= O_DIRECT;
 236         fd = open(s->filename, flags);
 237         if (fd < 0) {
 238                 perror("open");
 239                 goto done;
 240         }
 241
 242         if (get_file_size(fd, &s->max_blocks)) {
 243                 printf("failed getting size of device/file\n");
 244                 goto err;
 245         }
 246         if (!s->max_blocks) {
 247                 printf("Zero file/device size?\n");
 248                 goto err;
 249         }
 250
 251         s->max_blocks--;
 252
 253         srand48_r(pthread_self(), &s->rand);
 254
 255         prepped = 0;
 256         do {
 257                 int to_wait, flags, to_submit, this_reap;
 258
 259                 if (!prepped && s->inflight < DEPTH)
 260                         prepped = prep_more_ios(s, fd, min(DEPTH - s->inflight, BATCH_SUBMIT));
 261                 s->inflight += prepped;
 262 submit_more:
 263                 to_submit = prepped;
 264 submit:
 265                 if (s->inflight + BATCH_SUBMIT < DEPTH)
 266                         to_wait = 0;
 267                 else
 268                         to_wait = min(s->inflight + to_submit, BATCH_COMPLETE);
 269
 270                 flags = IORING_FLAG_GETEVENTS;
 271                 if (to_submit)
 272                         flags |= IORING_FLAG_SUBMIT;
 273
 274                 ret = io_ring_enter(s->ioc, to_submit, to_wait, flags);
 275                 s->calls++;
 276
 277                 this_reap = reap_events(s);
 278                 if (this_reap == -1)
 279                         break;
 280                 s->reaps += this_reap;
 281
 282                 if (ret >= 0) {
 283                         if (!ret) {
 284                                 to_submit = 0;
 285                                 if (s->inflight)
 286                                         goto submit;
 287                                 continue;
 288                         } else if (ret < to_submit) {
 289                                 int diff = to_submit - ret;
 290
 291                                 s->done += ret;
 292                                 prepped -= diff;
 293                                 goto submit_more;
 294                         }
 295                         s->done += ret;
 296                         prepped = 0;
 297                         continue;
 298                 } else if (ret < 0) {
 299                         if ((ret == -1 && errno == EAGAIN) || ret == -EAGAIN) {
 300                                 if (s->finish)
 301                                         break;
 302                                 if (this_reap)
 303                                         goto submit;
 304                                 to_submit = 0;
 305                                 goto submit;
 306                         }
 307                         if (ret == -1)
 308                                 printf("io_submit: %s\n", strerror(errno));
 309                         else
 310                                 printf("io_submit: %s\n", strerror(-ret));
 311                         break;
 312                 }
 313         } while (!s->finish);
 314 err:
 315         close(fd);
 316 done:
 317         finish = 1;
 318         return NULL;
 319 }
 320
 321 static void sig_int(int sig)
 322 {
 323         printf("Exiting on signal %d\n", sig);
 324         submitters[0].finish = 1;
 325         finish = 1;
 326 }
 327
 328 static void arm_sig_int(void)
 329 {
 330         struct sigaction act;
 331
 332         memset(&act, 0, sizeof(act));
 333         act.sa_handler = sig_int;
 334         act.sa_flags = SA_RESTART;
 335         sigaction(SIGINT, &act, NULL);
 336 }
 337
 338 int main(int argc, char *argv[])
 339 {
 340         struct submitter *s = &submitters[0];
 341         unsigned long done, calls, reap, cache_hit, cache_miss;
 342         int flags = 0, err;
 343         int j;
 344         size_t size;
 345         void *p, *ret;
 346         struct rlimit rlim;
 347
 348         if (argc < 2) {
 349                 printf("%s: filename\n", argv[0]);
 350                 return 1;
 351         }
 352
 353         rlim.rlim_cur = RLIM_INFINITY;
 354         rlim.rlim_max = RLIM_INFINITY;
 355         if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 356                 perror("setrlimit");
 357                 return 1;
 358         }
 359
 360         arm_sig_int();
 361
 362         size = sizeof(struct iocb) * DEPTH;
 363         if (posix_memalign(&p, 4096, size))
 364                 return 1;
 365         memset(p, 0, size);
 366         s->iocbs = p;
 367
 368         size = sizeof(struct aio_sq_ring) + DEPTH * sizeof(u32);
 369         if (posix_memalign(&p, 4096, size))
 370                 return 1;
 371         s->sq_ring = p;
 372         memset(p, 0, size);
 373         s->sq_ring->nr_events = DEPTH;
 374         s->sq_ring->iocbs = (u64) s->iocbs;
 375
 376         /* CQ ring must be twice as big */
 377         size = sizeof(struct aio_cq_ring) +
 378                         2 * DEPTH * sizeof(struct io_event);
 379         if (posix_memalign(&p, 4096, size))
 380                 return 1;
 381         s->cq_ring = p;
 382         memset(p, 0, size);
 383         s->cq_ring->nr_events = 2 * DEPTH;
 384
 385         for (j = 0; j < DEPTH; j++) {
 386                 struct iocb *iocb = &s->iocbs[j];
 387
 388                 if (posix_memalign(&iocb->u.c.buf, BS, BS)) {
 389                         printf("failed alloc\n");
 390                         return 1;
 391                 }
 392                 iocb->u.c.nbytes = BS;
 393         }
 394
 395         flags = IOCTX_FLAG_SCQRING;
 396         if (polled)
 397                 flags |= IOCTX_FLAG_IOPOLL;
 398         if (fixedbufs)
 399                 flags |= IOCTX_FLAG_FIXEDBUFS;
 400         if (buffered)
 401                 flags |= IOCTX_FLAG_SQWQ;
 402         else if (sq_thread) {
 403                 flags |= IOCTX_FLAG_SQTHREAD;
 404                 s->sq_ring->sq_thread_cpu = sq_thread_cpu;
 405         }
 406
 407         err = io_setup2(DEPTH, flags, s->sq_ring, s->cq_ring, &s->ioc);
 408         if (err) {
 409                 printf("ctx_init failed: %s, %d\n", strerror(errno), err);
 410                 return 1;
 411         }
 412         printf("polled=%d, fixedbufs=%d, buffered=%d\n", polled, fixedbufs, buffered);
 413         printf("  QD=%d, sq_ring=%d, cq_ring=%d\n", DEPTH, s->sq_ring->nr_events, s->cq_ring->nr_events);
 414         strcpy(s->filename, argv[1]);
 415
 416         pthread_create(&s->thread, NULL, submitter_fn, s);
 417
 418         cache_hit = cache_miss = reap = calls = done = 0;
 419         do {
 420                 unsigned long this_done = 0;
 421                 unsigned long this_reap = 0;
 422                 unsigned long this_call = 0;
 423                 unsigned long this_cache_hit = 0;
 424                 unsigned long this_cache_miss = 0;
 425                 unsigned long rpc = 0, ipc = 0;
 426                 double hit = 0.0;
 427
 428                 sleep(1);
 429                 this_done += s->done;
 430                 this_call += s->calls;
 431                 this_reap += s->reaps;
 432                 this_cache_hit += s->cachehit;
 433                 this_cache_miss += s->cachemiss;
 434                 if (this_cache_hit && this_cache_miss) {
 435                         unsigned long hits, total;
 436
 437                         hits = this_cache_hit - cache_hit;
 438                         total = hits + this_cache_miss - cache_miss;
 439                         hit = (double) hits / (double) total;
 440                         hit *= 100.0;
 441                 }
 442                 if (this_call - calls) {
 443                         rpc = (this_done - done) / (this_call - calls);
 444                         ipc = (this_reap - reap) / (this_call - calls);
 445                 }
 446                 printf("IOPS=%lu, IOS/call=%lu/%lu, inflight=%u (head=%u tail=%u), Cachehit=%0.2f%%\n",
 447                                 this_done - done, rpc, ipc, s->inflight,
 448                                 s->cq_ring->head, s->cq_ring->tail, hit);
 449                 done = this_done;
 450                 calls = this_call;
 451                 reap = this_reap;
 452                 cache_hit = s->cachehit;
 453                 cache_miss = s->cachemiss;
 454         } while (!finish);
 455
 456         pthread_join(s->thread, &ret);
 457         return 0;
 458 }