src/aio/aio.c

   1 #include <aio.h>
   2 #include <pthread.h>
   3 #include <semaphore.h>
   4 #include <limits.h>
   5 #include <errno.h>
   6 #include <unistd.h>
   7 #include <stdlib.h>
   8 #include <sys/auxv.h>
   9 #include "syscall.h"
  10 #include "atomic.h"
  11 #include "pthread_impl.h"
  12
  13 /* The following is a threads-based implementation of AIO with minimal
  14  * dependence on implementation details. Most synchronization is
  15  * performed with pthread primitives, but atomics and futex operations
  16  * are used for notification in a couple places where the pthread
  17  * primitives would be inefficient or impractical.
  18  *
  19  * For each fd with outstanding aio operations, an aio_queue structure
  20  * is maintained. These are reference-counted and destroyed by the last
  21  * aio worker thread to exit. Accessing any member of the aio_queue
  22  * structure requires a lock on the aio_queue. Adding and removing aio
  23  * queues themselves requires a write lock on the global map object,
  24  * a 4-level table mapping file descriptor numbers to aio queues. A
  25  * read lock on the map is used to obtain locks on existing queues by
  26  * excluding destruction of the queue by a different thread while it is
  27  * being locked.
  28  *
  29  * Each aio queue has a list of active threads/operations. Presently there
  30  * is a one to one relationship between threads and operations. The only
  31  * members of the aio_thread structure which are accessed by other threads
  32  * are the linked list pointers, op (which is immutable), running (which
  33  * is updated atomically), and err (which is synchronized via running),
  34  * so no locking is necessary. Most of the other other members are used
  35  * for sharing data between the main flow of execution and cancellation
  36  * cleanup handler.
  37  *
  38  * Taking any aio locks requires having all signals blocked. This is
  39  * necessary because aio_cancel is needed by close, and close is required
  40  * to be async-signal safe. All aio worker threads run with all signals
  41  * blocked permanently.
  42  */
  43
  44 struct aio_thread {
  45         pthread_t td;
  46         struct aiocb *cb;
  47         struct aio_thread *next, *prev;
  48         struct aio_queue *q;
  49         volatile int running;
  50         int err, op;
  51         ssize_t ret;
  52 };
  53
  54 struct aio_queue {
  55         int fd, seekable, append, ref, init;
  56         pthread_mutex_t lock;
  57         pthread_cond_t cond;
  58         struct aio_thread *head;
  59 };
  60
  61 struct aio_args {
  62         struct aiocb *cb;
  63         struct aio_queue *q;
  64         int op;
  65         sem_t sem;
  66 };
  67
  68 static pthread_rwlock_t maplock = PTHREAD_RWLOCK_INITIALIZER;
  69 static struct aio_queue *****map;
  70 static volatile int aio_fd_cnt;
  71 volatile int __aio_fut;
  72
  73 static struct aio_queue *__aio_get_queue(int fd, int need)
  74 {
  75         if (fd < 0) {
  76                 errno = EBADF;
  77                 return 0;
  78         }
  79         int a=fd>>24;
  80         unsigned char b=fd>>16, c=fd>>8, d=fd;
  81         struct aio_queue *q = 0;
  82         pthread_rwlock_rdlock(&maplock);
  83         if ((!map || !map[a] || !map[a][b] || !map[a][b][c] || !(q=map[a][b][c][d])) && need) {
  84                 pthread_rwlock_unlock(&maplock);
  85                 if (fcntl(fd, F_GETFD) < 0) return 0;
  86                 pthread_rwlock_wrlock(&maplock);
  87                 if (!map) map = calloc(sizeof *map, (-1U/2+1)>>24);
  88                 if (!map) goto out;
  89                 if (!map[a]) map[a] = calloc(sizeof **map, 256);
  90                 if (!map[a]) goto out;
  91                 if (!map[a][b]) map[a][b] = calloc(sizeof ***map, 256);
  92                 if (!map[a][b]) goto out;
  93                 if (!map[a][b][c]) map[a][b][c] = calloc(sizeof ****map, 256);
  94                 if (!map[a][b][c]) goto out;
  95                 if (!(q = map[a][b][c][d])) {
  96                         map[a][b][c][d] = q = calloc(sizeof *****map, 1);
  97                         if (q) {
  98                                 q->fd = fd;
  99                                 pthread_mutex_init(&q->lock, 0);
 100                                 pthread_cond_init(&q->cond, 0);
 101                                 a_inc(&aio_fd_cnt);
 102                         }
 103                 }
 104         }
 105         if (q) pthread_mutex_lock(&q->lock);
 106 out:
 107         pthread_rwlock_unlock(&maplock);
 108         return q;
 109 }
 110
 111 static void __aio_unref_queue(struct aio_queue *q)
 112 {
 113         if (q->ref > 1) {
 114                 q->ref--;
 115                 pthread_mutex_unlock(&q->lock);
 116                 return;
 117         }
 118
 119         /* This is potentially the last reference, but a new reference
 120          * may arrive since we cannot free the queue object without first
 121          * taking the maplock, which requires releasing the queue lock. */
 122         pthread_mutex_unlock(&q->lock);
 123         pthread_rwlock_wrlock(&maplock);
 124         pthread_mutex_lock(&q->lock);
 125         if (q->ref == 1) {
 126                 int fd=q->fd;
 127                 int a=fd>>24;
 128                 unsigned char b=fd>>16, c=fd>>8, d=fd;
 129                 map[a][b][c][d] = 0;
 130                 a_dec(&aio_fd_cnt);
 131                 pthread_rwlock_unlock(&maplock);
 132                 pthread_mutex_unlock(&q->lock);
 133                 free(q);
 134         } else {
 135                 q->ref--;
 136                 pthread_rwlock_unlock(&maplock);
 137                 pthread_mutex_unlock(&q->lock);
 138         }
 139 }
 140
 141 static void cleanup(void *ctx)
 142 {
 143         struct aio_thread *at = ctx;
 144         struct aio_queue *q = at->q;
 145         struct aiocb *cb = at->cb;
 146         struct sigevent sev = cb->aio_sigevent;
 147
 148         /* There are four potential types of waiters we could need to wake:
 149          *   1. Callers of aio_cancel/close.
 150          *   2. Callers of aio_suspend with a single aiocb.
 151          *   3. Callers of aio_suspend with a list.
 152          *   4. AIO worker threads waiting for sequenced operations.
 153          * Types 1-3 are notified via atomics/futexes, mainly for AS-safety
 154          * considerations. Type 4 is notified later via a cond var. */
 155
 156         cb->__ret = at->ret;
 157         if (a_swap(&at->running, 0) < 0)
 158                 __wake(&at->running, -1, 1);
 159         if (a_swap(&cb->__err, at->err) != EINPROGRESS)
 160                 __wake(&cb->__err, -1, 1);
 161         if (a_swap(&__aio_fut, 0))
 162                 __wake(&__aio_fut, -1, 1);
 163
 164         pthread_mutex_lock(&q->lock);
 165
 166         if (at->next) at->next->prev = at->prev;
 167         if (at->prev) at->prev->next = at->next;
 168         else q->head = at->next;
 169
 170         /* Signal aio worker threads waiting for sequenced operations. */
 171         pthread_cond_broadcast(&q->cond);
 172
 173         __aio_unref_queue(q);
 174
 175         if (sev.sigev_notify == SIGEV_SIGNAL) {
 176                 siginfo_t si = {
 177                         .si_signo = sev.sigev_signo,
 178                         .si_value = sev.sigev_value,
 179                         .si_code = SI_ASYNCIO,
 180                         .si_pid = getpid(),
 181                         .si_uid = getuid()
 182                 };
 183                 __syscall(SYS_rt_sigqueueinfo, si.si_pid, si.si_signo, &si);
 184         }
 185         if (sev.sigev_notify == SIGEV_THREAD) {
 186                 a_store(&__pthread_self()->cancel, 0);
 187                 sev.sigev_notify_function(sev.sigev_value);
 188         }
 189 }
 190
 191 static void *io_thread_func(void *ctx)
 192 {
 193         struct aio_thread at, *p;
 194
 195         struct aio_args *args = ctx;
 196         struct aiocb *cb = args->cb;
 197         int fd = cb->aio_fildes;
 198         int op = args->op;
 199         void *buf = (void *)cb->aio_buf;
 200         size_t len = cb->aio_nbytes;
 201         off_t off = cb->aio_offset;
 202
 203         struct aio_queue *q = args->q;
 204         ssize_t ret;
 205
 206         pthread_mutex_lock(&q->lock);
 207         sem_post(&args->sem);
 208
 209         at.op = op;
 210         at.running = 1;
 211         at.ret = -1;
 212         at.err = ECANCELED;
 213         at.q = q;
 214         at.td = __pthread_self();
 215         at.cb = cb;
 216         at.prev = 0;
 217         if ((at.next = q->head)) at.next->prev = &at;
 218         q->head = &at;
 219
 220         if (!q->init) {
 221                 int seekable = lseek(fd, 0, SEEK_CUR) >= 0;
 222                 q->seekable = seekable;
 223                 q->append = !seekable || (fcntl(fd, F_GETFL) & O_APPEND);
 224                 q->init = 1;
 225         }
 226
 227         pthread_cleanup_push(cleanup, &at);
 228
 229         /* Wait for sequenced operations. */
 230         if (op!=LIO_READ && (op!=LIO_WRITE || q->append)) {
 231                 for (;;) {
 232                         for (p=at.next; p && p->op!=LIO_WRITE; p=p->next);
 233                         if (!p) break;
 234                         pthread_cond_wait(&q->cond, &q->lock);
 235                 }
 236         }
 237
 238         pthread_mutex_unlock(&q->lock);
 239
 240         switch (op) {
 241         case LIO_WRITE:
 242                 ret = q->append ? write(fd, buf, len) : pwrite(fd, buf, len, off);
 243                 break;
 244         case LIO_READ:
 245                 ret = !q->seekable ? read(fd, buf, len) : pread(fd, buf, len, off);
 246                 break;
 247         case O_SYNC:
 248                 ret = fsync(fd);
 249                 break;
 250         case O_DSYNC:
 251                 ret = fdatasync(fd);
 252                 break;
 253         }
 254         at.ret = ret;
 255         at.err = ret<0 ? errno : 0;
 256
 257         pthread_cleanup_pop(1);
 258
 259         return 0;
 260 }
 261
 262 static size_t io_thread_stack_size = MINSIGSTKSZ+2048;
 263 static pthread_once_t init_stack_size_once;
 264
 265 static void init_stack_size()
 266 {
 267         unsigned long val = __getauxval(AT_MINSIGSTKSZ);
 268         if (val > MINSIGSTKSZ) io_thread_stack_size = val + 512;
 269 }
 270
 271 static int submit(struct aiocb *cb, int op)
 272 {
 273         int ret = 0;
 274         pthread_attr_t a;
 275         sigset_t allmask, origmask;
 276         pthread_t td;
 277         struct aio_queue *q = __aio_get_queue(cb->aio_fildes, 1);
 278         struct aio_args args = { .cb = cb, .op = op, .q = q };
 279         sem_init(&args.sem, 0, 0);
 280
 281         if (!q) {
 282                 if (errno != EBADF) errno = EAGAIN;
 283                 return -1;
 284         }
 285         q->ref++;
 286         pthread_mutex_unlock(&q->lock);
 287
 288         if (cb->aio_sigevent.sigev_notify == SIGEV_THREAD) {
 289                 if (cb->aio_sigevent.sigev_notify_attributes)
 290                         a = *cb->aio_sigevent.sigev_notify_attributes;
 291                 else
 292                         pthread_attr_init(&a);
 293         } else {
 294                 pthread_once(&init_stack_size_once, init_stack_size);
 295                 pthread_attr_init(&a);
 296                 pthread_attr_setstacksize(&a, io_thread_stack_size);
 297                 pthread_attr_setguardsize(&a, 0);
 298         }
 299         pthread_attr_setdetachstate(&a, PTHREAD_CREATE_DETACHED);
 300         sigfillset(&allmask);
 301         pthread_sigmask(SIG_BLOCK, &allmask, &origmask);
 302         cb->__err = EINPROGRESS;
 303         if (pthread_create(&td, &a, io_thread_func, &args)) {
 304                 pthread_mutex_lock(&q->lock);
 305                 __aio_unref_queue(q);
 306                 errno = EAGAIN;
 307                 ret = -1;
 308         }
 309         pthread_sigmask(SIG_SETMASK, &origmask, 0);
 310
 311         if (!ret) {
 312                 while (sem_wait(&args.sem));
 313         }
 314
 315         return ret;
 316 }
 317
 318 int aio_read(struct aiocb *cb)
 319 {
 320         return submit(cb, LIO_READ);
 321 }
 322
 323 int aio_write(struct aiocb *cb)
 324 {
 325         return submit(cb, LIO_WRITE);
 326 }
 327
 328 int aio_fsync(int op, struct aiocb *cb)
 329 {
 330         if (op != O_SYNC && op != O_DSYNC) {
 331                 errno = EINVAL;
 332                 return -1;
 333         }
 334         return submit(cb, op);
 335 }
 336
 337 ssize_t aio_return(struct aiocb *cb)
 338 {
 339         return cb->__ret;
 340 }
 341
 342 int aio_error(const struct aiocb *cb)
 343 {
 344         a_barrier();
 345         return cb->__err & 0x7fffffff;
 346 }
 347
 348 int aio_cancel(int fd, struct aiocb *cb)
 349 {
 350         sigset_t allmask, origmask;
 351         int ret = AIO_ALLDONE;
 352         struct aio_thread *p;
 353         struct aio_queue *q;
 354
 355         /* Unspecified behavior case. Report an error. */
 356         if (cb && fd != cb->aio_fildes) {
 357                 errno = EINVAL;
 358                 return -1;
 359         }
 360
 361         sigfillset(&allmask);
 362         pthread_sigmask(SIG_BLOCK, &allmask, &origmask);
 363
 364         errno = ENOENT;
 365         if (!(q = __aio_get_queue(fd, 0))) {
 366                 if (errno == EBADF) ret = -1;
 367                 goto done;
 368         }
 369
 370         for (p = q->head; p; p = p->next) {
 371                 if (cb && cb != p->cb) continue;
 372                 /* Transition target from running to running-with-waiters */
 373                 if (a_cas(&p->running, 1, -1)) {
 374                         pthread_cancel(p->td);
 375                         __wait(&p->running, 0, -1, 1);
 376                         if (p->err == ECANCELED) ret = AIO_CANCELED;
 377                 }
 378         }
 379
 380         pthread_mutex_unlock(&q->lock);
 381 done:
 382         pthread_sigmask(SIG_SETMASK, &origmask, 0);
 383         return ret;
 384 }
 385
 386 int __aio_close(int fd)
 387 {
 388         a_barrier();
 389         if (aio_fd_cnt) aio_cancel(fd, 0);
 390         return fd;
 391 }
 392
 393 weak_alias(aio_cancel, aio_cancel64);
 394 weak_alias(aio_error, aio_error64);
 395 weak_alias(aio_fsync, aio_fsync64);
 396 weak_alias(aio_read, aio_read64);
 397 weak_alias(aio_write, aio_write64);
 398 weak_alias(aio_return, aio_return64);