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 << 1)
  28 #define IOCTX_FLAG_USERIOCB     (1 << 0)
  29 #define IOCTX_FLAG_FIXEDBUFS    (1 << 2)
  30 #define IOCTX_FLAG_SCQRING      (1 << 3)        /* Use SQ/CQ rings */
  31 #define IOCTX_FLAG_SQTHREAD     (1 << 4)        /* Use SQ thread */
  32 #define IOCTX_FLAG_SQWQ         (1 << 5)        /* Use SQ wq */
  33
  34 #define barrier()       __asm__ __volatile__("": : :"memory")
  35
  36 #define min(a, b)               ((a < b) ? (a) : (b))
  37
  38 typedef uint32_t u32;
  39
  40 struct aio_iocb_ring {
  41         union {
  42                 struct {
  43                         u32 head, tail;
  44                         u32 nr_events;
  45                         u32 sq_thread_cpu;
  46                 };
  47                 struct iocb pad_iocb;
  48         };
  49         struct iocb iocbs[0];
  50 };
  51
  52 struct aio_io_event_ring {
  53         union {
  54                 struct {
  55                         u32 head, tail;
  56                         u32 nr_events;
  57                 };
  58                 struct io_event pad_event;
  59         };
  60         struct io_event events[0];
  61 };
  62
  63 #define IORING_FLAG_SUBMIT      (1 << 0)
  64 #define IORING_FLAG_GETEVENTS   (1 << 1)
  65
  66 #define container_of(ptr, type, member)  ({                             \
  67         const __typeof__( ((type *)0)->member ) *__mptr = (ptr);        \
  68         (type *)( (char *)__mptr - offsetof(type,member) );})
  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_iocb_ring *sq_ring;
  84         struct aio_io_event_ring *cq_ring;
  85         int inflight;
  86         unsigned long reaps;
  87         unsigned long done;
  88         unsigned long calls;
  89         volatile int finish;
  90         char filename[128];
  91 };
  92
  93 static struct submitter submitters[1];
  94 static volatile int finish;
  95
  96 static int polled = 1;          /* use IO polling */
  97 static int fixedbufs = 1;       /* use fixed user buffers */
  98 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
  99 static int sq_thread = 0;
 100 static int sq_thread_cpu = 0;
 101
 102 static int io_setup2(unsigned int nr_events, unsigned int flags,
 103                      struct iocb *iocbs, struct aio_iocb_ring *sq_ring,
 104                      struct aio_io_event_ring *cq_ring, io_context_t *ctx_idp)
 105 {
 106         return syscall(335, nr_events, flags, iocbs, sq_ring, cq_ring, ctx_idp);
 107 }
 108
 109 static int io_ring_enter(io_context_t ctx, unsigned int to_submit,
 110                          unsigned int min_complete, unsigned int flags)
 111 {
 112         return syscall(336, ctx, to_submit, min_complete, flags);
 113 }
 114
 115 static int gettid(void)
 116 {
 117         return syscall(__NR_gettid);
 118 }
 119
 120 static void init_io(struct submitter *s, int fd, struct iocb *iocb)
 121 {
 122         unsigned long offset;
 123         long r;
 124
 125         lrand48_r(&s->rand, &r);
 126         offset = (r % (s->max_blocks - 1)) * BS;
 127
 128         iocb->aio_fildes = fd;
 129         iocb->aio_lio_opcode = IO_CMD_PREAD;
 130         iocb->u.c.offset = offset;
 131         if (polled)
 132                 iocb->u.c.flags = IOCB_FLAG_HIPRI;
 133         if (!fixedbufs)
 134                 iocb->u.c.nbytes = BS;
 135 }
 136
 137 static int prep_more_ios(struct submitter *s, int fd, int max_ios)
 138 {
 139         struct aio_iocb_ring *ring = s->sq_ring;
 140         struct iocb *iocb;
 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                 iocb = &s->sq_ring->iocbs[tail];
 154                 init_io(s, fd, iocb);
 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_io_event_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                         int index = (int) (uintptr_t) ev->obj;
 204                         struct iocb *iocb = &s->sq_ring->iocbs[index];
 205
 206                         printf("io: unexpected ret=%ld\n", ev->res);
 207                         printf("offset=%lu, size=%lu\n", (unsigned long) iocb->u.c.offset, (unsigned long) iocb->u.c.nbytes);
 208                         return -1;
 209                 }
 210                 reaped++;
 211                 head++;
 212                 if (head == ring->nr_events)
 213                         head = 0;
 214         } while (1);
 215
 216         s->inflight -= reaped;
 217         ring->head = head;
 218         barrier();
 219         return reaped;
 220 }
 221
 222 static void *submitter_fn(void *data)
 223 {
 224         struct submitter *s = data;
 225         int fd, ret, prepped, flags;
 226
 227         printf("submitter=%d\n", gettid());
 228
 229         flags = O_RDONLY;
 230         if (!buffered)
 231                 flags |= O_DIRECT;
 232         fd = open(s->filename, flags);
 233         if (fd < 0) {
 234                 perror("open");
 235                 goto done;
 236         }
 237
 238         if (get_file_size(fd, &s->max_blocks)) {
 239                 printf("failed getting size of device/file\n");
 240                 goto err;
 241         }
 242         if (!s->max_blocks) {
 243                 printf("Zero file/device size?\n");
 244                 goto err;
 245         }
 246
 247         s->max_blocks--;
 248
 249         srand48_r(pthread_self(), &s->rand);
 250
 251         prepped = 0;
 252         do {
 253                 int to_wait, flags, to_submit, this_reap;
 254
 255                 if (!prepped && s->inflight < DEPTH)
 256                         prepped = prep_more_ios(s, fd, min(DEPTH - s->inflight, BATCH_SUBMIT));
 257                 s->inflight += prepped;
 258 submit_more:
 259                 to_submit = prepped;
 260 submit:
 261                 if (s->inflight + BATCH_SUBMIT < DEPTH)
 262                         to_wait = 0;
 263                 else
 264                         to_wait = min(s->inflight + to_submit, BATCH_COMPLETE);
 265
 266                 flags = IORING_FLAG_GETEVENTS;
 267                 if (to_submit)
 268                         flags |= IORING_FLAG_SUBMIT;
 269
 270                 ret = io_ring_enter(s->ioc, to_submit, to_wait, flags);
 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                                 printf("EAGAIN reap=%d\n", reap_events(s));
 301                                 printf("%d/%d, %d/%d\n", s->sq_ring->head, s->sq_ring->tail, s->cq_ring->head, s->cq_ring->tail);
 302                                 to_submit = 0;
 303                                 goto submit;
 304                         }
 305                         if (ret == -1)
 306                                 printf("io_submit: %s\n", strerror(errno));
 307                         else
 308                                 printf("io_submit: %s\n", strerror(-ret));
 309                         break;
 310                 }
 311         } while (!s->finish);
 312 err:
 313         close(fd);
 314 done:
 315         finish = 1;
 316         return NULL;
 317 }
 318
 319 static void sig_int(int sig)
 320 {
 321         printf("Exiting on signal %d\n", sig);
 322         submitters[0].finish = 1;
 323         finish = 1;
 324 }
 325
 326 static void arm_sig_int(void)
 327 {
 328         struct sigaction act;
 329
 330         memset(&act, 0, sizeof(act));
 331         act.sa_handler = sig_int;
 332         act.sa_flags = SA_RESTART;
 333         sigaction(SIGINT, &act, NULL);
 334 }
 335
 336 int main(int argc, char *argv[])
 337 {
 338         struct submitter *s = &submitters[0];
 339         unsigned long done, calls, reap;
 340         int flags = 0, err;
 341         int j;
 342         size_t size;
 343         void *p, *ret;
 344         struct rlimit rlim;
 345
 346         if (argc < 2) {
 347                 printf("%s: filename\n", argv[0]);
 348                 return 1;
 349         }
 350
 351         rlim.rlim_cur = RLIM_INFINITY;
 352         rlim.rlim_max = RLIM_INFINITY;
 353         if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 354                 perror("setrlimit");
 355                 return 1;
 356         }
 357
 358         arm_sig_int();
 359
 360         size = sizeof(struct aio_iocb_ring) + RING_SIZE * sizeof(struct iocb);
 361         if (posix_memalign(&p, 4096, size))
 362                 return 1;
 363         s->sq_ring = p;
 364         memset(p, 0, size);
 365
 366         size = sizeof(struct aio_io_event_ring) + RING_SIZE * sizeof(struct io_event);
 367         if (posix_memalign(&p, 4096, size))
 368                 return 1;
 369         s->cq_ring = p;
 370         memset(p, 0, size);
 371
 372         for (j = 0; j < RING_SIZE; j++) {
 373                 struct iocb *iocb = &s->sq_ring->iocbs[j];
 374
 375                 if (posix_memalign(&iocb->u.c.buf, BS, BS)) {
 376                         printf("failed alloc\n");
 377                         return 1;
 378                 }
 379                 iocb->u.c.nbytes = BS;
 380         }
 381
 382         flags = IOCTX_FLAG_SCQRING;
 383         if (polled)
 384                 flags |= IOCTX_FLAG_IOPOLL;
 385         if (fixedbufs)
 386                 flags |= IOCTX_FLAG_FIXEDBUFS;
 387         if (buffered)
 388                 flags |= IOCTX_FLAG_SQWQ;
 389         else if (sq_thread) {
 390                 flags |= IOCTX_FLAG_SQTHREAD;
 391                 s->sq_ring->sq_thread_cpu = sq_thread_cpu;
 392         }
 393
 394         err = io_setup2(RING_SIZE, flags, s->sq_ring->iocbs, s->sq_ring, s->cq_ring, &s->ioc);
 395         if (err) {
 396                 printf("ctx_init failed: %s, %d\n", strerror(errno), err);
 397                 return 1;
 398         }
 399         printf("polled=%d, fixedbufs=%d, buffered=%d\n", polled, fixedbufs, buffered);
 400         printf("  QD=%d, sq_ring=%d, cq_ring=%d\n", DEPTH, s->sq_ring->nr_events, s->cq_ring->nr_events);
 401         strcpy(s->filename, argv[1]);
 402
 403         pthread_create(&s->thread, NULL, submitter_fn, s);
 404
 405         reap = calls = done = 0;
 406         do {
 407                 unsigned long this_done = 0;
 408                 unsigned long this_reap = 0;
 409                 unsigned long this_call = 0;
 410                 unsigned long rpc = 0, ipc = 0;
 411
 412                 sleep(1);
 413                 this_done += s->done;
 414                 this_call += s->calls;
 415                 this_reap += s->reaps;
 416                 if (this_call - calls) {
 417                         rpc = (this_done - done) / (this_call - calls);
 418                         ipc = (this_reap - reap) / (this_call - calls);
 419                 }
 420                 printf("IOPS=%lu, IOS/call=%lu/%lu, inflight=%u (head=%d tail=%d), %lu, %lu\n",
 421                                 this_done - done, rpc, ipc, s->inflight,
 422                                 s->cq_ring->head, s->cq_ring->tail, s->reaps, s->done);
 423                 done = this_done;
 424                 calls = this_call;
 425                 reap = this_reap;
 426         } while (!finish);
 427
 428         pthread_join(s->thread, &ret);
 429         return 0;
 430 }