master
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#include "aio_impl.h"
13
14#define malloc __libc_malloc
15#define calloc __libc_calloc
16#define realloc __libc_realloc
17#define free __libc_free
18
19/* The following is a threads-based implementation of AIO with minimal
20 * dependence on implementation details. Most synchronization is
21 * performed with pthread primitives, but atomics and futex operations
22 * are used for notification in a couple places where the pthread
23 * primitives would be inefficient or impractical.
24 *
25 * For each fd with outstanding aio operations, an aio_queue structure
26 * is maintained. These are reference-counted and destroyed by the last
27 * aio worker thread to exit. Accessing any member of the aio_queue
28 * structure requires a lock on the aio_queue. Adding and removing aio
29 * queues themselves requires a write lock on the global map object,
30 * a 4-level table mapping file descriptor numbers to aio queues. A
31 * read lock on the map is used to obtain locks on existing queues by
32 * excluding destruction of the queue by a different thread while it is
33 * being locked.
34 *
35 * Each aio queue has a list of active threads/operations. Presently there
36 * is a one to one relationship between threads and operations. The only
37 * members of the aio_thread structure which are accessed by other threads
38 * are the linked list pointers, op (which is immutable), running (which
39 * is updated atomically), and err (which is synchronized via running),
40 * so no locking is necessary. Most of the other other members are used
41 * for sharing data between the main flow of execution and cancellation
42 * cleanup handler.
43 *
44 * Taking any aio locks requires having all signals blocked. This is
45 * necessary because aio_cancel is needed by close, and close is required
46 * to be async-signal safe. All aio worker threads run with all signals
47 * blocked permanently.
48 */
49
50struct aio_thread {
51 pthread_t td;
52 struct aiocb *cb;
53 struct aio_thread *next, *prev;
54 struct aio_queue *q;
55 volatile int running;
56 int err, op;
57 ssize_t ret;
58};
59
60struct aio_queue {
61 int fd, seekable, append, ref, init;
62 pthread_mutex_t lock;
63 pthread_cond_t cond;
64 struct aio_thread *head;
65};
66
67struct aio_args {
68 struct aiocb *cb;
69 struct aio_queue *q;
70 int op;
71 sem_t sem;
72};
73
74static pthread_rwlock_t maplock = PTHREAD_RWLOCK_INITIALIZER;
75static struct aio_queue *****map;
76static volatile int aio_fd_cnt;
77volatile int __aio_fut;
78
79static size_t io_thread_stack_size;
80
81#define MAX(a,b) ((a)>(b) ? (a) : (b))
82
83static struct aio_queue *__aio_get_queue(int fd, int need)
84{
85 sigset_t allmask, origmask;
86 int masked = 0;
87 if (fd < 0) {
88 errno = EBADF;
89 return 0;
90 }
91 int a=fd>>24;
92 unsigned char b=fd>>16, c=fd>>8, d=fd;
93 struct aio_queue *q = 0;
94 pthread_rwlock_rdlock(&maplock);
95 if ((!map || !map[a] || !map[a][b] || !map[a][b][c] || !(q=map[a][b][c][d])) && need) {
96 pthread_rwlock_unlock(&maplock);
97 if (fcntl(fd, F_GETFD) < 0) return 0;
98 sigfillset(&allmask);
99 masked = 1;
100 pthread_sigmask(SIG_BLOCK, &allmask, &origmask);
101 pthread_rwlock_wrlock(&maplock);
102 if (!io_thread_stack_size) {
103 unsigned long val = __getauxval(AT_MINSIGSTKSZ);
104 io_thread_stack_size = MAX(MINSIGSTKSZ+2048, val+512);
105 }
106 if (!map) map = calloc(sizeof *map, (-1U/2+1)>>24);
107 if (!map) goto out;
108 if (!map[a]) map[a] = calloc(sizeof **map, 256);
109 if (!map[a]) goto out;
110 if (!map[a][b]) map[a][b] = calloc(sizeof ***map, 256);
111 if (!map[a][b]) goto out;
112 if (!map[a][b][c]) map[a][b][c] = calloc(sizeof ****map, 256);
113 if (!map[a][b][c]) goto out;
114 if (!(q = map[a][b][c][d])) {
115 map[a][b][c][d] = q = calloc(sizeof *****map, 1);
116 if (q) {
117 q->fd = fd;
118 pthread_mutex_init(&q->lock, 0);
119 pthread_cond_init(&q->cond, 0);
120 a_inc(&aio_fd_cnt);
121 }
122 }
123 }
124 if (q) pthread_mutex_lock(&q->lock);
125out:
126 pthread_rwlock_unlock(&maplock);
127 if (masked) pthread_sigmask(SIG_SETMASK, &origmask, 0);
128 return q;
129}
130
131static void __aio_unref_queue(struct aio_queue *q)
132{
133 if (q->ref > 1) {
134 q->ref--;
135 pthread_mutex_unlock(&q->lock);
136 return;
137 }
138
139 /* This is potentially the last reference, but a new reference
140 * may arrive since we cannot free the queue object without first
141 * taking the maplock, which requires releasing the queue lock. */
142 pthread_mutex_unlock(&q->lock);
143 pthread_rwlock_wrlock(&maplock);
144 pthread_mutex_lock(&q->lock);
145 if (q->ref == 1) {
146 int fd=q->fd;
147 int a=fd>>24;
148 unsigned char b=fd>>16, c=fd>>8, d=fd;
149 map[a][b][c][d] = 0;
150 a_dec(&aio_fd_cnt);
151 pthread_rwlock_unlock(&maplock);
152 pthread_mutex_unlock(&q->lock);
153 free(q);
154 } else {
155 q->ref--;
156 pthread_rwlock_unlock(&maplock);
157 pthread_mutex_unlock(&q->lock);
158 }
159}
160
161static void cleanup(void *ctx)
162{
163 struct aio_thread *at = ctx;
164 struct aio_queue *q = at->q;
165 struct aiocb *cb = at->cb;
166 struct sigevent sev = cb->aio_sigevent;
167
168 /* There are four potential types of waiters we could need to wake:
169 * 1. Callers of aio_cancel/close.
170 * 2. Callers of aio_suspend with a single aiocb.
171 * 3. Callers of aio_suspend with a list.
172 * 4. AIO worker threads waiting for sequenced operations.
173 * Types 1-3 are notified via atomics/futexes, mainly for AS-safety
174 * considerations. Type 4 is notified later via a cond var. */
175
176 cb->__ret = at->ret;
177 if (a_swap(&at->running, 0) < 0)
178 __wake(&at->running, -1, 1);
179 if (a_swap(&cb->__err, at->err) != EINPROGRESS)
180 __wake(&cb->__err, -1, 1);
181 if (a_swap(&__aio_fut, 0))
182 __wake(&__aio_fut, -1, 1);
183
184 pthread_mutex_lock(&q->lock);
185
186 if (at->next) at->next->prev = at->prev;
187 if (at->prev) at->prev->next = at->next;
188 else q->head = at->next;
189
190 /* Signal aio worker threads waiting for sequenced operations. */
191 pthread_cond_broadcast(&q->cond);
192
193 __aio_unref_queue(q);
194
195 if (sev.sigev_notify == SIGEV_SIGNAL) {
196 siginfo_t si = {
197 .si_signo = sev.sigev_signo,
198 .si_value = sev.sigev_value,
199 .si_code = SI_ASYNCIO,
200 .si_pid = getpid(),
201 .si_uid = getuid()
202 };
203 __syscall(SYS_rt_sigqueueinfo, si.si_pid, si.si_signo, &si);
204 }
205 if (sev.sigev_notify == SIGEV_THREAD) {
206 a_store(&__pthread_self()->cancel, 0);
207 sev.sigev_notify_function(sev.sigev_value);
208 }
209}
210
211static void *io_thread_func(void *ctx)
212{
213 struct aio_thread at, *p;
214
215 struct aio_args *args = ctx;
216 struct aiocb *cb = args->cb;
217 int fd = cb->aio_fildes;
218 int op = args->op;
219 void *buf = (void *)cb->aio_buf;
220 size_t len = cb->aio_nbytes;
221 off_t off = cb->aio_offset;
222
223 struct aio_queue *q = args->q;
224 ssize_t ret;
225
226 pthread_mutex_lock(&q->lock);
227 sem_post(&args->sem);
228
229 at.op = op;
230 at.running = 1;
231 at.ret = -1;
232 at.err = ECANCELED;
233 at.q = q;
234 at.td = __pthread_self();
235 at.cb = cb;
236 at.prev = 0;
237 if ((at.next = q->head)) at.next->prev = &at;
238 q->head = &at;
239
240 if (!q->init) {
241 int seekable = lseek(fd, 0, SEEK_CUR) >= 0;
242 q->seekable = seekable;
243 q->append = !seekable || (fcntl(fd, F_GETFL) & O_APPEND);
244 q->init = 1;
245 }
246
247 pthread_cleanup_push(cleanup, &at);
248
249 /* Wait for sequenced operations. */
250 if (op!=LIO_READ && (op!=LIO_WRITE || q->append)) {
251 for (;;) {
252 for (p=at.next; p && p->op!=LIO_WRITE; p=p->next);
253 if (!p) break;
254 pthread_cond_wait(&q->cond, &q->lock);
255 }
256 }
257
258 pthread_mutex_unlock(&q->lock);
259
260 switch (op) {
261 case LIO_WRITE:
262 ret = q->append ? write(fd, buf, len) : pwrite(fd, buf, len, off);
263 break;
264 case LIO_READ:
265 ret = !q->seekable ? read(fd, buf, len) : pread(fd, buf, len, off);
266 break;
267 case O_SYNC:
268 ret = fsync(fd);
269 break;
270 case O_DSYNC:
271 ret = fdatasync(fd);
272 break;
273 }
274 at.ret = ret;
275 at.err = ret<0 ? errno : 0;
276
277 pthread_cleanup_pop(1);
278
279 return 0;
280}
281
282static int submit(struct aiocb *cb, int op)
283{
284 int ret = 0;
285 pthread_attr_t a;
286 sigset_t allmask, origmask;
287 pthread_t td;
288 struct aio_queue *q = __aio_get_queue(cb->aio_fildes, 1);
289 struct aio_args args = { .cb = cb, .op = op, .q = q };
290 sem_init(&args.sem, 0, 0);
291
292 if (!q) {
293 if (errno != EBADF) errno = EAGAIN;
294 cb->__ret = -1;
295 cb->__err = errno;
296 return -1;
297 }
298 q->ref++;
299 pthread_mutex_unlock(&q->lock);
300
301 if (cb->aio_sigevent.sigev_notify == SIGEV_THREAD) {
302 if (cb->aio_sigevent.sigev_notify_attributes)
303 a = *cb->aio_sigevent.sigev_notify_attributes;
304 else
305 pthread_attr_init(&a);
306 } else {
307 pthread_attr_init(&a);
308 pthread_attr_setstacksize(&a, io_thread_stack_size);
309 pthread_attr_setguardsize(&a, 0);
310 }
311 pthread_attr_setdetachstate(&a, PTHREAD_CREATE_DETACHED);
312 sigfillset(&allmask);
313 pthread_sigmask(SIG_BLOCK, &allmask, &origmask);
314 cb->__err = EINPROGRESS;
315 if (pthread_create(&td, &a, io_thread_func, &args)) {
316 pthread_mutex_lock(&q->lock);
317 __aio_unref_queue(q);
318 cb->__err = errno = EAGAIN;
319 cb->__ret = ret = -1;
320 }
321 pthread_sigmask(SIG_SETMASK, &origmask, 0);
322
323 if (!ret) {
324 while (sem_wait(&args.sem));
325 }
326
327 return ret;
328}
329
330int aio_read(struct aiocb *cb)
331{
332 return submit(cb, LIO_READ);
333}
334
335int aio_write(struct aiocb *cb)
336{
337 return submit(cb, LIO_WRITE);
338}
339
340int aio_fsync(int op, struct aiocb *cb)
341{
342 if (op != O_SYNC && op != O_DSYNC) {
343 errno = EINVAL;
344 return -1;
345 }
346 return submit(cb, op);
347}
348
349ssize_t aio_return(struct aiocb *cb)
350{
351 return cb->__ret;
352}
353
354int aio_error(const struct aiocb *cb)
355{
356 a_barrier();
357 return cb->__err & 0x7fffffff;
358}
359
360int aio_cancel(int fd, struct aiocb *cb)
361{
362 sigset_t allmask, origmask;
363 int ret = AIO_ALLDONE;
364 struct aio_thread *p;
365 struct aio_queue *q;
366
367 /* Unspecified behavior case. Report an error. */
368 if (cb && fd != cb->aio_fildes) {
369 errno = EINVAL;
370 return -1;
371 }
372
373 sigfillset(&allmask);
374 pthread_sigmask(SIG_BLOCK, &allmask, &origmask);
375
376 errno = ENOENT;
377 if (!(q = __aio_get_queue(fd, 0))) {
378 if (errno == EBADF) ret = -1;
379 goto done;
380 }
381
382 for (p = q->head; p; p = p->next) {
383 if (cb && cb != p->cb) continue;
384 /* Transition target from running to running-with-waiters */
385 if (a_cas(&p->running, 1, -1)) {
386 pthread_cancel(p->td);
387 __wait(&p->running, 0, -1, 1);
388 if (p->err == ECANCELED) ret = AIO_CANCELED;
389 }
390 }
391
392 pthread_mutex_unlock(&q->lock);
393done:
394 pthread_sigmask(SIG_SETMASK, &origmask, 0);
395 return ret;
396}
397
398int __aio_close(int fd)
399{
400 a_barrier();
401 if (aio_fd_cnt) aio_cancel(fd, 0);
402 return fd;
403}
404
405void __aio_atfork(int who)
406{
407 if (who<0) {
408 pthread_rwlock_rdlock(&maplock);
409 return;
410 } else if (!who) {
411 pthread_rwlock_unlock(&maplock);
412 return;
413 }
414 aio_fd_cnt = 0;
415 if (pthread_rwlock_tryrdlock(&maplock)) {
416 /* Obtaining lock may fail if _Fork was called nor via
417 * fork. In this case, no further aio is possible from
418 * child and we can just null out map so __aio_close
419 * does not attempt to do anything. */
420 map = 0;
421 return;
422 }
423 if (map) for (int a=0; a<(-1U/2+1)>>24; a++)
424 if (map[a]) for (int b=0; b<256; b++)
425 if (map[a][b]) for (int c=0; c<256; c++)
426 if (map[a][b][c]) for (int d=0; d<256; d++)
427 map[a][b][c][d] = 0;
428 /* Re-initialize the rwlock rather than unlocking since there
429 * may have been more than one reference on it in the parent.
430 * We are not a lock holder anyway; the thread in the parent was. */
431 pthread_rwlock_init(&maplock, 0);
432}