• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
50 struct 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 
60 struct 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 
67 struct aio_args {
68 	struct aiocb *cb;
69 	struct aio_queue *q;
70 	int op;
71 	sem_t sem;
72 };
73 
74 static pthread_rwlock_t maplock = PTHREAD_RWLOCK_INITIALIZER;
75 static struct aio_queue *****map;
76 static volatile int aio_fd_cnt;
77 volatile int __aio_fut;
78 
79 static size_t io_thread_stack_size;
80 
81 #define MAX(a,b) ((a)>(b) ? (a) : (b))
82 
__aio_get_queue(int fd,int need)83 static struct aio_queue *__aio_get_queue(int fd, int need)
84 {
85 	if (fd < 0) {
86 		errno = EBADF;
87 		return 0;
88 	}
89 	int a=fd>>24;
90 	unsigned char b=fd>>16, c=fd>>8, d=fd;
91 	struct aio_queue *q = 0;
92 	pthread_rwlock_rdlock(&maplock);
93 	if ((!map || !map[a] || !map[a][b] || !map[a][b][c] || !(q=map[a][b][c][d])) && need) {
94 		pthread_rwlock_unlock(&maplock);
95 		if (fcntl(fd, F_GETFD) < 0) return 0;
96 		pthread_rwlock_wrlock(&maplock);
97 		if (!io_thread_stack_size) {
98 			unsigned long val = __getauxval(AT_MINSIGSTKSZ);
99 			io_thread_stack_size = MAX(MINSIGSTKSZ+2048, val+512);
100 		}
101 		if (!map) map = calloc(sizeof *map, (-1U/2+1)>>24);
102 		if (!map) goto out;
103 		if (!map[a]) map[a] = calloc(sizeof **map, 256);
104 		if (!map[a]) goto out;
105 		if (!map[a][b]) map[a][b] = calloc(sizeof ***map, 256);
106 		if (!map[a][b]) goto out;
107 		if (!map[a][b][c]) map[a][b][c] = calloc(sizeof ****map, 256);
108 		if (!map[a][b][c]) goto out;
109 		if (!(q = map[a][b][c][d])) {
110 			map[a][b][c][d] = q = calloc(sizeof *****map, 1);
111 			if (q) {
112 				q->fd = fd;
113 				pthread_mutex_init(&q->lock, 0);
114 				pthread_cond_init(&q->cond, 0);
115 				a_inc(&aio_fd_cnt);
116 			}
117 		}
118 	}
119 	if (q) pthread_mutex_lock(&q->lock);
120 out:
121 	pthread_rwlock_unlock(&maplock);
122 	return q;
123 }
124 
__aio_unref_queue(struct aio_queue * q)125 static void __aio_unref_queue(struct aio_queue *q)
126 {
127 	if (q->ref > 1) {
128 		q->ref--;
129 		pthread_mutex_unlock(&q->lock);
130 		return;
131 	}
132 
133 	/* This is potentially the last reference, but a new reference
134 	 * may arrive since we cannot free the queue object without first
135 	 * taking the maplock, which requires releasing the queue lock. */
136 	pthread_mutex_unlock(&q->lock);
137 	pthread_rwlock_wrlock(&maplock);
138 	pthread_mutex_lock(&q->lock);
139 	if (q->ref == 1) {
140 		int fd=q->fd;
141 		int a=fd>>24;
142 		unsigned char b=fd>>16, c=fd>>8, d=fd;
143 		map[a][b][c][d] = 0;
144 		a_dec(&aio_fd_cnt);
145 		pthread_rwlock_unlock(&maplock);
146 		pthread_mutex_unlock(&q->lock);
147 		free(q);
148 	} else {
149 		q->ref--;
150 		pthread_rwlock_unlock(&maplock);
151 		pthread_mutex_unlock(&q->lock);
152 	}
153 }
154 
cleanup(void * ctx)155 static void cleanup(void *ctx)
156 {
157 	struct aio_thread *at = ctx;
158 	struct aio_queue *q = at->q;
159 	struct aiocb *cb = at->cb;
160 	struct sigevent sev = cb->aio_sigevent;
161 
162 	/* There are four potential types of waiters we could need to wake:
163 	 *   1. Callers of aio_cancel/close.
164 	 *   2. Callers of aio_suspend with a single aiocb.
165 	 *   3. Callers of aio_suspend with a list.
166 	 *   4. AIO worker threads waiting for sequenced operations.
167 	 * Types 1-3 are notified via atomics/futexes, mainly for AS-safety
168 	 * considerations. Type 4 is notified later via a cond var. */
169 
170 	cb->__ret = at->ret;
171 	if (a_swap(&at->running, 0) < 0)
172 		__wake(&at->running, -1, 1);
173 	if (a_swap(&cb->__err, at->err) != EINPROGRESS)
174 		__wake(&cb->__err, -1, 1);
175 	if (a_swap(&__aio_fut, 0))
176 		__wake(&__aio_fut, -1, 1);
177 
178 	pthread_mutex_lock(&q->lock);
179 
180 	if (at->next) at->next->prev = at->prev;
181 	if (at->prev) at->prev->next = at->next;
182 	else q->head = at->next;
183 
184 	/* Signal aio worker threads waiting for sequenced operations. */
185 	pthread_cond_broadcast(&q->cond);
186 
187 	__aio_unref_queue(q);
188 
189 	if (sev.sigev_notify == SIGEV_SIGNAL) {
190 		siginfo_t si = {
191 			.si_signo = sev.sigev_signo,
192 			.si_value = sev.sigev_value,
193 			.si_code = SI_ASYNCIO,
194 			.si_pid = getpid(),
195 			.si_uid = getuid()
196 		};
197 		__syscall(SYS_rt_sigqueueinfo, si.si_pid, si.si_signo, &si);
198 	}
199 	if (sev.sigev_notify == SIGEV_THREAD) {
200 		a_store(&__pthread_self()->cancel, 0);
201 		sev.sigev_notify_function(sev.sigev_value);
202 	}
203 }
204 
io_thread_func(void * ctx)205 static void *io_thread_func(void *ctx)
206 {
207 	struct aio_thread at, *p;
208 
209 	struct aio_args *args = ctx;
210 	struct aiocb *cb = args->cb;
211 	int fd = cb->aio_fildes;
212 	int op = args->op;
213 	void *buf = (void *)cb->aio_buf;
214 	size_t len = cb->aio_nbytes;
215 	off_t off = cb->aio_offset;
216 
217 	struct aio_queue *q = args->q;
218 	ssize_t ret;
219 
220 	pthread_mutex_lock(&q->lock);
221 	sem_post(&args->sem);
222 
223 	at.op = op;
224 	at.running = 1;
225 	at.ret = -1;
226 	at.err = ECANCELED;
227 	at.q = q;
228 	at.td = __pthread_self();
229 	at.cb = cb;
230 	at.prev = 0;
231 	if ((at.next = q->head)) at.next->prev = &at;
232 	q->head = &at;
233 
234 	if (!q->init) {
235 		int seekable = lseek(fd, 0, SEEK_CUR) >= 0;
236 		q->seekable = seekable;
237 		q->append = !seekable || (fcntl(fd, F_GETFL) & O_APPEND);
238 		q->init = 1;
239 	}
240 
241 	pthread_cleanup_push(cleanup, &at);
242 
243 	/* Wait for sequenced operations. */
244 	if (op!=LIO_READ && (op!=LIO_WRITE || q->append)) {
245 		for (;;) {
246 			for (p=at.next; p && p->op!=LIO_WRITE; p=p->next);
247 			if (!p) break;
248 			pthread_cond_wait(&q->cond, &q->lock);
249 		}
250 	}
251 
252 	pthread_mutex_unlock(&q->lock);
253 
254 	switch (op) {
255 	case LIO_WRITE:
256 		ret = q->append ? write(fd, buf, len) : pwrite(fd, buf, len, off);
257 		break;
258 	case LIO_READ:
259 		ret = !q->seekable ? read(fd, buf, len) : pread(fd, buf, len, off);
260 		break;
261 	case O_SYNC:
262 		ret = fsync(fd);
263 		break;
264 	case O_DSYNC:
265 		ret = fdatasync(fd);
266 		break;
267 	}
268 	at.ret = ret;
269 	at.err = ret<0 ? errno : 0;
270 
271 	pthread_cleanup_pop(1);
272 
273 	return 0;
274 }
275 
submit(struct aiocb * cb,int op)276 static int submit(struct aiocb *cb, int op)
277 {
278 	int ret = 0;
279 	pthread_attr_t a;
280 	sigset_t allmask, origmask;
281 	pthread_t td;
282 	struct aio_queue *q = __aio_get_queue(cb->aio_fildes, 1);
283 	struct aio_args args = { .cb = cb, .op = op, .q = q };
284 	sem_init(&args.sem, 0, 0);
285 
286 	if (!q) {
287 		if (errno != EBADF) errno = EAGAIN;
288 		cb->__ret = -1;
289 		cb->__err = errno;
290 		return -1;
291 	}
292 	q->ref++;
293 	pthread_mutex_unlock(&q->lock);
294 
295 	if (cb->aio_sigevent.sigev_notify == SIGEV_THREAD) {
296 		if (cb->aio_sigevent.sigev_notify_attributes)
297 			a = *cb->aio_sigevent.sigev_notify_attributes;
298 		else
299 			pthread_attr_init(&a);
300 	} else {
301 		pthread_attr_init(&a);
302 		pthread_attr_setstacksize(&a, io_thread_stack_size);
303 		pthread_attr_setguardsize(&a, 0);
304 	}
305 	pthread_attr_setdetachstate(&a, PTHREAD_CREATE_DETACHED);
306 	sigfillset(&allmask);
307 	pthread_sigmask(SIG_BLOCK, &allmask, &origmask);
308 	cb->__err = EINPROGRESS;
309 	if (pthread_create(&td, &a, io_thread_func, &args)) {
310 		pthread_mutex_lock(&q->lock);
311 		__aio_unref_queue(q);
312 		cb->__err = errno = EAGAIN;
313 		cb->__ret = ret = -1;
314 	}
315 	pthread_sigmask(SIG_SETMASK, &origmask, 0);
316 
317 	if (!ret) {
318 		while (sem_wait(&args.sem));
319 	}
320 
321 	return ret;
322 }
323 
aio_read(struct aiocb * cb)324 int aio_read(struct aiocb *cb)
325 {
326 	return submit(cb, LIO_READ);
327 }
328 
aio_write(struct aiocb * cb)329 int aio_write(struct aiocb *cb)
330 {
331 	return submit(cb, LIO_WRITE);
332 }
333 
aio_fsync(int op,struct aiocb * cb)334 int aio_fsync(int op, struct aiocb *cb)
335 {
336 	if (op != O_SYNC && op != O_DSYNC) {
337 		errno = EINVAL;
338 		return -1;
339 	}
340 	return submit(cb, op);
341 }
342 
aio_return(struct aiocb * cb)343 ssize_t aio_return(struct aiocb *cb)
344 {
345 	return cb->__ret;
346 }
347 
aio_error(const struct aiocb * cb)348 int aio_error(const struct aiocb *cb)
349 {
350 	a_barrier();
351 	return cb->__err & 0x7fffffff;
352 }
353 
aio_cancel(int fd,struct aiocb * cb)354 int aio_cancel(int fd, struct aiocb *cb)
355 {
356 	sigset_t allmask, origmask;
357 	int ret = AIO_ALLDONE;
358 	struct aio_thread *p;
359 	struct aio_queue *q;
360 
361 	/* Unspecified behavior case. Report an error. */
362 	if (cb && fd != cb->aio_fildes) {
363 		errno = EINVAL;
364 		return -1;
365 	}
366 
367 	sigfillset(&allmask);
368 	pthread_sigmask(SIG_BLOCK, &allmask, &origmask);
369 
370 	errno = ENOENT;
371 	if (!(q = __aio_get_queue(fd, 0))) {
372 		if (errno == EBADF) ret = -1;
373 		goto done;
374 	}
375 
376 	for (p = q->head; p; p = p->next) {
377 		if (cb && cb != p->cb) continue;
378 		/* Transition target from running to running-with-waiters */
379 		if (a_cas(&p->running, 1, -1)) {
380 			pthread_cancel(p->td);
381 			__wait(&p->running, 0, -1, 1);
382 			if (p->err == ECANCELED) ret = AIO_CANCELED;
383 		}
384 	}
385 
386 	pthread_mutex_unlock(&q->lock);
387 done:
388 	pthread_sigmask(SIG_SETMASK, &origmask, 0);
389 	return ret;
390 }
391 
__aio_close(int fd)392 int __aio_close(int fd)
393 {
394 	a_barrier();
395 	if (aio_fd_cnt) aio_cancel(fd, 0);
396 	return fd;
397 }
398 
__aio_atfork(int who)399 void __aio_atfork(int who)
400 {
401 	if (who<0) {
402 		pthread_rwlock_rdlock(&maplock);
403 		return;
404 	}
405 	if (who>0 && map) for (int a=0; a<(-1U/2+1)>>24; a++)
406 		if (map[a]) for (int b=0; b<256; b++)
407 			if (map[a][b]) for (int c=0; c<256; c++)
408 				if (map[a][b][c]) for (int d=0; d<256; d++)
409 					map[a][b][c][d] = 0;
410 	pthread_rwlock_unlock(&maplock);
411 }
412 
413 weak_alias(aio_cancel, aio_cancel64);
414 weak_alias(aio_error, aio_error64);
415 weak_alias(aio_fsync, aio_fsync64);
416 weak_alias(aio_read, aio_read64);
417 weak_alias(aio_write, aio_write64);
418 weak_alias(aio_return, aio_return64);
419