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