1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
|
#include <aio.h>
#include <pthread.h>
#include <semaphore.h>
#include <limits.h>
#include <errno.h>
#include <unistd.h>
#include <stdlib.h>
#include "syscall.h"
#include "atomic.h"
#include "libc.h"
#include "pthread_impl.h"
/* The following is a threads-based implementation of AIO with minimal
* dependence on implementation details. Most synchronization is
* performed with pthread primitives, but atomics and futex operations
* are used for notification in a couple places where the pthread
* primitives would be inefficient or impractical.
*
* For each fd with outstanding aio operations, an aio_queue structure
* is maintained. These are reference-counted and destroyed by the last
* aio worker thread to exit. Accessing any member of the aio_queue
* structure requires a lock on the aio_queue. Adding and removing aio
* queues themselves requires a write lock on the global map object,
* a 4-level table mapping file descriptor numbers to aio queues. A
* read lock on the map is used to obtain locks on existing queues by
* excluding destruction of the queue by a different thread while it is
* being locked.
*
* Each aio queue has a list of active threads/operations. Presently there
* is a one to one relationship between threads and operations. The only
* members of the aio_thread structure which are accessed by other threads
* are the linked list pointers, op (which is immutable), running (which
* is updated atomically), and err (which is synchronized via running),
* so no locking is necessary. Most of the other other members are used
* for sharing data between the main flow of execution and cancellation
* cleanup handler.
*
* Taking any aio locks requires having all signals blocked. This is
* necessary because aio_cancel is needed by close, and close is required
* to be async-signal safe. All aio worker threads run with all signals
* blocked permanently.
*/
struct aio_args {
struct aiocb *cb;
int op;
int err;
sem_t sem;
};
struct aio_thread {
pthread_t td;
struct aiocb *cb;
struct aio_thread *next, *prev;
struct aio_queue *q;
volatile int running;
int err, op;
ssize_t ret;
};
struct aio_queue {
int fd, seekable, append, ref, init;
pthread_mutex_t lock;
pthread_cond_t cond;
struct aio_thread *head;
};
static pthread_rwlock_t maplock = PTHREAD_RWLOCK_INITIALIZER;
static struct aio_queue *****map;
static volatile int aio_fd_cnt;
volatile int __aio_fut;
static struct aio_queue *__aio_get_queue(int fd, int need)
{
if (fd < 0) return 0;
int a=fd>>24;
unsigned char b=fd>>16, c=fd>>8, d=fd;
struct aio_queue *q = 0;
pthread_rwlock_rdlock(&maplock);
if ((!map || !map[a] || !map[a][b] || !map[a][b][c] || !(q=map[a][b][c][d])) && need) {
pthread_rwlock_unlock(&maplock);
pthread_rwlock_wrlock(&maplock);
if (!map) map = calloc(sizeof *map, (-1U/2+1)>>24);
if (!map) goto out;
if (!map[a]) map[a] = calloc(sizeof **map, 256);
if (!map[a]) goto out;
if (!map[a][b]) map[a][b] = calloc(sizeof ***map, 256);
if (!map[a][b]) goto out;
if (!map[a][b][c]) map[a][b][c] = calloc(sizeof ****map, 256);
if (!map[a][b][c]) goto out;
if (!(q = map[a][b][c][d])) {
map[a][b][c][d] = q = calloc(sizeof *****map, 1);
if (q) {
q->fd = fd;
pthread_mutex_init(&q->lock, 0);
pthread_cond_init(&q->cond, 0);
a_inc(&aio_fd_cnt);
}
}
}
if (q) pthread_mutex_lock(&q->lock);
out:
pthread_rwlock_unlock(&maplock);
return q;
}
static void __aio_unref_queue(struct aio_queue *q)
{
if (q->ref > 1) {
q->ref--;
pthread_mutex_unlock(&q->lock);
return;
}
/* This is potentially the last reference, but a new reference
* may arrive since we cannot free the queue object without first
* taking the maplock, which requires releasing the queue lock. */
pthread_mutex_unlock(&q->lock);
pthread_rwlock_wrlock(&maplock);
pthread_mutex_lock(&q->lock);
if (q->ref == 1) {
int fd=q->fd;
int a=fd>>24;
unsigned char b=fd>>16, c=fd>>8, d=fd;
map[a][b][c][d] = 0;
a_dec(&aio_fd_cnt);
pthread_rwlock_unlock(&maplock);
pthread_mutex_unlock(&q->lock);
free(q);
} else {
q->ref--;
pthread_rwlock_unlock(&maplock);
pthread_mutex_unlock(&q->lock);
}
}
static void cleanup(void *ctx)
{
struct aio_thread *at = ctx;
struct aio_queue *q = at->q;
struct aiocb *cb = at->cb;
struct sigevent sev = cb->aio_sigevent;
/* There are four potential types of waiters we could need to wake:
* 1. Callers of aio_cancel/close.
* 2. Callers of aio_suspend with a single aiocb.
* 3. Callers of aio_suspend with a list.
* 4. AIO worker threads waiting for sequenced operations.
* Types 1-3 are notified via atomics/futexes, mainly for AS-safety
* considerations. Type 4 is notified later via a cond var. */
cb->__ret = at->ret;
if (a_swap(&at->running, 0) < 0)
__wake(&at->running, -1, 1);
if (a_swap(&cb->__err, at->err) != EINPROGRESS)
__wake(&cb->__err, -1, 1);
if (a_swap(&__aio_fut, 0))
__wake(&__aio_fut, -1, 1);
pthread_mutex_lock(&q->lock);
if (at->next) at->next->prev = at->prev;
if (at->prev) at->prev->next = at->next;
else q->head = at->next;
/* Signal aio worker threads waiting for sequenced operations. */
pthread_cond_broadcast(&q->cond);
__aio_unref_queue(q);
if (sev.sigev_notify == SIGEV_SIGNAL) {
siginfo_t si = {
.si_signo = sev.sigev_signo,
.si_value = sev.sigev_value,
.si_code = SI_ASYNCIO,
.si_pid = getpid(),
.si_uid = getuid()
};
__syscall(SYS_rt_sigqueueinfo, si.si_pid, si.si_signo, &si);
}
if (sev.sigev_notify == SIGEV_THREAD) {
a_store(&__pthread_self()->cancel, 0);
sev.sigev_notify_function(sev.sigev_value);
}
}
static void *io_thread_func(void *ctx)
{
struct aio_thread at, *p;
struct aio_args *args = ctx;
struct aiocb *cb = args->cb;
int fd = cb->aio_fildes;
int op = args->op;
void *buf = (void *)cb->aio_buf;
size_t len = cb->aio_nbytes;
off_t off = cb->aio_offset;
struct aio_queue *q = __aio_get_queue(fd, 1);
ssize_t ret;
args->err = q ? 0 : EAGAIN;
sem_post(&args->sem);
if (!q) return 0;
at.op = op;
at.running = 1;
at.ret = -1;
at.err = ECANCELED;
at.q = q;
at.td = __pthread_self();
at.cb = cb;
at.prev = 0;
if ((at.next = q->head)) at.next->prev = &at;
q->head = &at;
q->ref++;
if (!q->init) {
int seekable = lseek(fd, 0, SEEK_CUR) >= 0;
q->seekable = seekable;
q->append = !seekable || (fcntl(fd, F_GETFL) & O_APPEND);
q->init = 1;
}
pthread_cleanup_push(cleanup, &at);
/* Wait for sequenced operations. */
if (op!=LIO_READ && (op!=LIO_WRITE || q->append)) {
for (;;) {
for (p=at.next; p && p->op!=LIO_WRITE; p=p->next);
if (!p) break;
pthread_cond_wait(&q->cond, &q->lock);
}
}
pthread_mutex_unlock(&q->lock);
switch (op) {
case LIO_WRITE:
ret = q->append ? write(fd, buf, len) : pwrite(fd, buf, len, off);
break;
case LIO_READ:
ret = !q->seekable ? read(fd, buf, len) : pread(fd, buf, len, off);
break;
case O_SYNC:
ret = fsync(fd);
break;
case O_DSYNC:
ret = fdatasync(fd);
break;
}
at.ret = ret;
at.err = ret<0 ? errno : 0;
pthread_cleanup_pop(1);
return 0;
}
static int submit(struct aiocb *cb, int op)
{
int ret = 0;
pthread_attr_t a;
sigset_t allmask, origmask;
pthread_t td;
struct aio_args args = { .cb = cb, .op = op };
sem_init(&args.sem, 0, 0);
if (cb->aio_sigevent.sigev_notify == SIGEV_THREAD) {
if (cb->aio_sigevent.sigev_notify_attributes)
a = *cb->aio_sigevent.sigev_notify_attributes;
else
pthread_attr_init(&a);
} else {
pthread_attr_init(&a);
pthread_attr_setstacksize(&a, PTHREAD_STACK_MIN);
pthread_attr_setguardsize(&a, 0);
}
pthread_attr_setdetachstate(&a, PTHREAD_CREATE_DETACHED);
sigfillset(&allmask);
pthread_sigmask(SIG_BLOCK, &allmask, &origmask);
cb->__err = EINPROGRESS;
if (pthread_create(&td, &a, io_thread_func, &args)) {
errno = EAGAIN;
ret = -1;
}
pthread_sigmask(SIG_SETMASK, &origmask, 0);
if (!ret) {
while (sem_wait(&args.sem));
if (args.err) {
errno = args.err;
ret = -1;
}
}
return ret;
}
int aio_read(struct aiocb *cb)
{
return submit(cb, LIO_READ);
}
int aio_write(struct aiocb *cb)
{
return submit(cb, LIO_WRITE);
}
int aio_fsync(int op, struct aiocb *cb)
{
if (op != O_SYNC && op != O_DSYNC) {
errno = EINVAL;
return -1;
}
return submit(cb, op);
}
ssize_t aio_return(struct aiocb *cb)
{
return cb->__ret;
}
int aio_error(const struct aiocb *cb)
{
a_barrier();
return cb->__err & 0x7fffffff;
}
int aio_cancel(int fd, struct aiocb *cb)
{
sigset_t allmask, origmask;
int ret = AIO_ALLDONE;
struct aio_thread *p;
struct aio_queue *q;
/* Unspecified behavior case. Report an error. */
if (cb && fd != cb->aio_fildes) {
errno = EINVAL;
return -1;
}
sigfillset(&allmask);
pthread_sigmask(SIG_BLOCK, &allmask, &origmask);
if (!(q = __aio_get_queue(fd, 0))) {
if (fcntl(fd, F_GETFD) < 0) ret = -1;
goto done;
}
for (p = q->head; p; p = p->next) {
if (cb && cb != p->cb) continue;
/* Transition target from running to running-with-waiters */
if (a_cas(&p->running, 1, -1)) {
pthread_cancel(p->td);
__wait(&p->running, 0, -1, 1);
if (p->err == ECANCELED) ret = AIO_CANCELED;
}
}
pthread_mutex_unlock(&q->lock);
done:
pthread_sigmask(SIG_SETMASK, &origmask, 0);
return ret;
}
int __aio_close(int fd)
{
a_barrier();
if (aio_fd_cnt) aio_cancel(fd, 0);
return fd;
}
LFS64(aio_cancel);
LFS64(aio_error);
LFS64(aio_fsync);
LFS64(aio_read);
LFS64(aio_write);
LFS64(aio_return);
|