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