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