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 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_SUBMIT      (1 << 0)
  68 #define IORING_FLAG_GETEVENTS   (1 << 1)
  69
  70 #define DEPTH                   32
  71 #define RING_SIZE               (DEPTH + 1)
  72
  73 #define BATCH_SUBMIT            8
  74 #define BATCH_COMPLETE          8
  75
  76 #define BS                      4096
  77
  78 struct submitter {
  79         pthread_t thread;
  80         unsigned long max_blocks;
  81         io_context_t ioc;
  82         struct drand48_data rand;
  83         struct aio_sq_ring *sq_ring;
  84         struct iocb *iocbs;
  85         struct aio_cq_ring *cq_ring;
  86         int inflight;
  87         unsigned long reaps;
  88         unsigned long done;
  89         unsigned long calls;
  90         volatile int finish;
  91         char filename[128];
  92 };
  93
  94 static struct submitter submitters[1];
  95 static volatile int finish;
  96
  97 static int polled = 1;          /* use IO polling */
  98 static int fixedbufs = 1;       /* use fixed user buffers */
  99 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
 100 static int sq_thread = 0;       /* use kernel submission thread */
 101 static int sq_thread_cpu = 0;   /* pin above thread to this CPU */
 102
 103 static int io_setup2(unsigned int nr_events, unsigned int flags,
 104                      struct aio_sq_ring *sq_ring, struct aio_cq_ring *cq_ring,
 105                      io_context_t *ctx_idp)
 106 {
 107         return syscall(335, nr_events, flags, sq_ring, cq_ring, ctx_idp);
 108 }
 109
 110 static int io_ring_enter(io_context_t ctx, unsigned int to_submit,
 111                          unsigned int min_complete, unsigned int flags)
 112 {
 113         return syscall(336, ctx, to_submit, min_complete, flags);
 114 }
 115
 116 static int gettid(void)
 117 {
 118         return syscall(__NR_gettid);
 119 }
 120
 121 static void init_io(struct submitter *s, int fd, struct iocb *iocb)
 122 {
 123         unsigned long offset;
 124         long r;
 125
 126         lrand48_r(&s->rand, &r);
 127         offset = (r % (s->max_blocks - 1)) * BS;
 128
 129         iocb->aio_fildes = fd;
 130         iocb->aio_lio_opcode = IO_CMD_PREAD;
 131         iocb->u.c.offset = offset;
 132         if (polled)
 133                 iocb->u.c.flags = IOCB_FLAG_HIPRI;
 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 tail, next_tail, prepped = 0;
 142
 143         next_tail = tail = ring->tail;
 144         do {
 145                 next_tail++;
 146                 if (next_tail == ring->nr_events)
 147                         next_tail = 0;
 148
 149                 barrier();
 150                 if (next_tail == ring->head)
 151                         break;
 152
 153                 init_io(s, fd, &s->iocbs[tail]);
 154                 s->sq_ring->array[tail] = tail;
 155                 prepped++;
 156                 tail = next_tail;
 157         } while (prepped < max_ios);
 158
 159         if (ring->tail != tail) {
 160                 /* order tail store with writes to iocbs above */
 161                 barrier();
 162                 ring->tail = tail;
 163                 barrier();
 164         }
 165         return prepped;
 166 }
 167
 168 static int get_file_size(int fd, unsigned long *blocks)
 169 {
 170         struct stat st;
 171
 172         if (fstat(fd, &st) < 0)
 173                 return -1;
 174         if (S_ISBLK(st.st_mode)) {
 175                 unsigned long long bytes;
 176
 177                 if (ioctl(fd, BLKGETSIZE64, &bytes) != 0)
 178                         return -1;
 179
 180                 *blocks = bytes / BS;
 181                 return 0;
 182         } else if (S_ISREG(st.st_mode)) {
 183                 *blocks = st.st_size / BS;
 184                 return 0;
 185         }
 186
 187         return -1;
 188 }
 189
 190 static int reap_events(struct submitter *s)
 191 {
 192         struct aio_cq_ring *ring = s->cq_ring;
 193         struct io_event *ev;
 194         u32 head, reaped = 0;
 195
 196         head = ring->head;
 197         do {
 198                 barrier();
 199                 if (head == ring->tail)
 200                         break;
 201                 ev = &ring->events[head];
 202                 if (ev->res != BS) {
 203                         struct iocb *iocb = ev->obj;
 204
 205                         printf("io: unexpected ret=%ld\n", ev->res);
 206                         printf("offset=%lu, size=%lu\n", (unsigned long) iocb->u.c.offset, (unsigned long) iocb->u.c.nbytes);
 207                         return -1;
 208                 }
 209                 reaped++;
 210                 head++;
 211                 if (head == ring->nr_events)
 212                         head = 0;
 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, flags, 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                 flags = IORING_FLAG_GETEVENTS;
 266                 if (to_submit)
 267                         flags |= IORING_FLAG_SUBMIT;
 268
 269                 ret = io_ring_enter(s->ioc, to_submit, to_wait, flags);
 270                 s->calls++;
 271
 272                 this_reap = reap_events(s);
 273                 if (this_reap == -1)
 274                         break;
 275                 s->reaps += this_reap;
 276
 277                 if (ret >= 0) {
 278                         if (!ret) {
 279                                 to_submit = 0;
 280                                 if (s->inflight)
 281                                         goto submit;
 282                                 continue;
 283                         } else if (ret < to_submit) {
 284                                 int diff = to_submit - ret;
 285
 286                                 s->done += ret;
 287                                 prepped -= diff;
 288                                 goto submit_more;
 289                         }
 290                         s->done += ret;
 291                         prepped = 0;
 292                         continue;
 293                 } else if (ret < 0) {
 294                         if ((ret == -1 && errno == EAGAIN) || ret == -EAGAIN) {
 295                                 if (s->finish)
 296                                         break;
 297                                 if (this_reap)
 298                                         goto submit;
 299                                 to_submit = 0;
 300                                 goto submit;
 301                         }
 302                         if (ret == -1)
 303                                 printf("io_submit: %s\n", strerror(errno));
 304                         else
 305                                 printf("io_submit: %s\n", strerror(-ret));
 306                         break;
 307                 }
 308         } while (!s->finish);
 309 err:
 310         close(fd);
 311 done:
 312         finish = 1;
 313         return NULL;
 314 }
 315
 316 static void sig_int(int sig)
 317 {
 318         printf("Exiting on signal %d\n", sig);
 319         submitters[0].finish = 1;
 320         finish = 1;
 321 }
 322
 323 static void arm_sig_int(void)
 324 {
 325         struct sigaction act;
 326
 327         memset(&act, 0, sizeof(act));
 328         act.sa_handler = sig_int;
 329         act.sa_flags = SA_RESTART;
 330         sigaction(SIGINT, &act, NULL);
 331 }
 332
 333 int main(int argc, char *argv[])
 334 {
 335         struct submitter *s = &submitters[0];
 336         unsigned long done, calls, reap;
 337         int flags = 0, err;
 338         int j;
 339         size_t size;
 340         void *p, *ret;
 341         struct rlimit rlim;
 342
 343         if (argc < 2) {
 344                 printf("%s: filename\n", argv[0]);
 345                 return 1;
 346         }
 347
 348         rlim.rlim_cur = RLIM_INFINITY;
 349         rlim.rlim_max = RLIM_INFINITY;
 350         if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 351                 perror("setrlimit");
 352                 return 1;
 353         }
 354
 355         arm_sig_int();
 356
 357         size = sizeof(struct iocb) * RING_SIZE;
 358         if (posix_memalign(&p, 4096, size))
 359                 return 1;
 360         memset(p, 0, size);
 361         s->iocbs = p;
 362
 363         size = sizeof(struct aio_sq_ring) + RING_SIZE * sizeof(u32);
 364         if (posix_memalign(&p, 4096, size))
 365                 return 1;
 366         s->sq_ring = p;
 367         memset(p, 0, size);
 368         s->sq_ring->nr_events = RING_SIZE;
 369         s->sq_ring->iocbs = (u64) s->iocbs;
 370
 371         /* CQ ring must be twice as big */
 372         size = sizeof(struct aio_cq_ring) +
 373                         2 * RING_SIZE * sizeof(struct io_event);
 374         if (posix_memalign(&p, 4096, size))
 375                 return 1;
 376         s->cq_ring = p;
 377         memset(p, 0, size);
 378         s->cq_ring->nr_events = 2 * RING_SIZE;
 379
 380         for (j = 0; j < RING_SIZE; j++) {
 381                 struct iocb *iocb = &s->iocbs[j];
 382
 383                 if (posix_memalign(&iocb->u.c.buf, BS, BS)) {
 384                         printf("failed alloc\n");
 385                         return 1;
 386                 }
 387                 iocb->u.c.nbytes = BS;
 388         }
 389
 390         flags = IOCTX_FLAG_SCQRING;
 391         if (polled)
 392                 flags |= IOCTX_FLAG_IOPOLL;
 393         if (fixedbufs)
 394                 flags |= IOCTX_FLAG_FIXEDBUFS;
 395         if (buffered)
 396                 flags |= IOCTX_FLAG_SQWQ;
 397         else if (sq_thread) {
 398                 flags |= IOCTX_FLAG_SQTHREAD;
 399                 s->sq_ring->sq_thread_cpu = sq_thread_cpu;
 400         }
 401
 402         err = io_setup2(RING_SIZE, flags, s->sq_ring, s->cq_ring, &s->ioc);
 403         if (err) {
 404                 printf("ctx_init failed: %s, %d\n", strerror(errno), err);
 405                 return 1;
 406         }
 407         printf("polled=%d, fixedbufs=%d, buffered=%d\n", polled, fixedbufs, buffered);
 408         printf("  QD=%d, sq_ring=%d, cq_ring=%d\n", DEPTH, s->sq_ring->nr_events, s->cq_ring->nr_events);
 409         strcpy(s->filename, argv[1]);
 410
 411         pthread_create(&s->thread, NULL, submitter_fn, s);
 412
 413         reap = calls = done = 0;
 414         do {
 415                 unsigned long this_done = 0;
 416                 unsigned long this_reap = 0;
 417                 unsigned long this_call = 0;
 418                 unsigned long rpc = 0, ipc = 0;
 419
 420                 sleep(1);
 421                 this_done += s->done;
 422                 this_call += s->calls;
 423                 this_reap += s->reaps;
 424                 if (this_call - calls) {
 425                         rpc = (this_done - done) / (this_call - calls);
 426                         ipc = (this_reap - reap) / (this_call - calls);
 427                 }
 428                 printf("IOPS=%lu, IOS/call=%lu/%lu, inflight=%u (head=%d tail=%d), %lu, %lu\n",
 429                                 this_done - done, rpc, ipc, s->inflight,
 430                                 s->cq_ring->head, s->cq_ring->tail, s->reaps, s->done);
 431                 done = this_done;
 432                 calls = this_call;
 433                 reap = this_reap;
 434         } while (!finish);
 435
 436         pthread_join(s->thread, &ret);
 437         return 0;
 438 }