root / posix-aio-compat.c @ 5bc89ef6
History | View | Annotate | Download (14.5 kB)
1 | 3c529d93 | aliguori | /*
|
---|---|---|---|
2 | 3c529d93 | aliguori | * QEMU posix-aio emulation
|
3 | 3c529d93 | aliguori | *
|
4 | 3c529d93 | aliguori | * Copyright IBM, Corp. 2008
|
5 | 3c529d93 | aliguori | *
|
6 | 3c529d93 | aliguori | * Authors:
|
7 | 3c529d93 | aliguori | * Anthony Liguori <aliguori@us.ibm.com>
|
8 | 3c529d93 | aliguori | *
|
9 | 3c529d93 | aliguori | * This work is licensed under the terms of the GNU GPL, version 2. See
|
10 | 3c529d93 | aliguori | * the COPYING file in the top-level directory.
|
11 | 3c529d93 | aliguori | *
|
12 | 3c529d93 | aliguori | */
|
13 | 3c529d93 | aliguori | |
14 | 221f715d | aliguori | #include <sys/ioctl.h> |
15 | 9ef91a67 | Christoph Hellwig | #include <sys/types.h> |
16 | 3c529d93 | aliguori | #include <pthread.h> |
17 | 3c529d93 | aliguori | #include <unistd.h> |
18 | 3c529d93 | aliguori | #include <errno.h> |
19 | 30525aff | malc | #include <time.h> |
20 | 9ef91a67 | Christoph Hellwig | #include <signal.h> |
21 | 8653c015 | malc | #include <string.h> |
22 | 8653c015 | malc | #include <stdlib.h> |
23 | 8653c015 | malc | #include <stdio.h> |
24 | 9ef91a67 | Christoph Hellwig | |
25 | 9ef91a67 | Christoph Hellwig | #include "sys-queue.h" |
26 | 3c529d93 | aliguori | #include "osdep.h" |
27 | f141eafe | aliguori | #include "qemu-common.h" |
28 | 9ef91a67 | Christoph Hellwig | #include "block_int.h" |
29 | 9ef91a67 | Christoph Hellwig | |
30 | 9ef91a67 | Christoph Hellwig | #include "block/raw-posix-aio.h" |
31 | 9ef91a67 | Christoph Hellwig | |
32 | 9ef91a67 | Christoph Hellwig | |
33 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb {
|
34 | 9ef91a67 | Christoph Hellwig | BlockDriverAIOCB common; |
35 | 9ef91a67 | Christoph Hellwig | int aio_fildes;
|
36 | 9ef91a67 | Christoph Hellwig | union {
|
37 | 9ef91a67 | Christoph Hellwig | struct iovec *aio_iov;
|
38 | 9ef91a67 | Christoph Hellwig | void *aio_ioctl_buf;
|
39 | 9ef91a67 | Christoph Hellwig | }; |
40 | 9ef91a67 | Christoph Hellwig | int aio_niov;
|
41 | 9ef91a67 | Christoph Hellwig | size_t aio_nbytes; |
42 | 9ef91a67 | Christoph Hellwig | #define aio_ioctl_cmd aio_nbytes /* for QEMU_AIO_IOCTL */ |
43 | 9ef91a67 | Christoph Hellwig | int ev_signo;
|
44 | 9ef91a67 | Christoph Hellwig | off_t aio_offset; |
45 | 9ef91a67 | Christoph Hellwig | |
46 | 9ef91a67 | Christoph Hellwig | TAILQ_ENTRY(qemu_paiocb) node; |
47 | 9ef91a67 | Christoph Hellwig | int aio_type;
|
48 | 9ef91a67 | Christoph Hellwig | ssize_t ret; |
49 | 9ef91a67 | Christoph Hellwig | int active;
|
50 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *next;
|
51 | 9ef91a67 | Christoph Hellwig | }; |
52 | 9ef91a67 | Christoph Hellwig | |
53 | 9ef91a67 | Christoph Hellwig | typedef struct PosixAioState { |
54 | 9ef91a67 | Christoph Hellwig | int rfd, wfd;
|
55 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *first_aio;
|
56 | 9ef91a67 | Christoph Hellwig | } PosixAioState; |
57 | 3c529d93 | aliguori | |
58 | 3c529d93 | aliguori | |
59 | 3c529d93 | aliguori | static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
|
60 | 3c529d93 | aliguori | static pthread_cond_t cond = PTHREAD_COND_INITIALIZER;
|
61 | 3c529d93 | aliguori | static pthread_t thread_id;
|
62 | a8227a5a | malc | static pthread_attr_t attr;
|
63 | 3c529d93 | aliguori | static int max_threads = 64; |
64 | 3c529d93 | aliguori | static int cur_threads = 0; |
65 | 3c529d93 | aliguori | static int idle_threads = 0; |
66 | 3c529d93 | aliguori | static TAILQ_HEAD(, qemu_paiocb) request_list;
|
67 | 3c529d93 | aliguori | |
68 | 2341f9a1 | Juan Quintela | #ifdef CONFIG_PREADV
|
69 | ceb42de8 | aliguori | static int preadv_present = 1; |
70 | ceb42de8 | aliguori | #else
|
71 | ceb42de8 | aliguori | static int preadv_present = 0; |
72 | ceb42de8 | aliguori | #endif
|
73 | ceb42de8 | aliguori | |
74 | 8653c015 | malc | static void die2(int err, const char *what) |
75 | 8653c015 | malc | { |
76 | 8653c015 | malc | fprintf(stderr, "%s failed: %s\n", what, strerror(err));
|
77 | 8653c015 | malc | abort(); |
78 | 8653c015 | malc | } |
79 | 8653c015 | malc | |
80 | 8653c015 | malc | static void die(const char *what) |
81 | 8653c015 | malc | { |
82 | 8653c015 | malc | die2(errno, what); |
83 | 8653c015 | malc | } |
84 | 8653c015 | malc | |
85 | 8653c015 | malc | static void mutex_lock(pthread_mutex_t *mutex) |
86 | 8653c015 | malc | { |
87 | 8653c015 | malc | int ret = pthread_mutex_lock(mutex);
|
88 | 8653c015 | malc | if (ret) die2(ret, "pthread_mutex_lock"); |
89 | 8653c015 | malc | } |
90 | 8653c015 | malc | |
91 | 8653c015 | malc | static void mutex_unlock(pthread_mutex_t *mutex) |
92 | 8653c015 | malc | { |
93 | 8653c015 | malc | int ret = pthread_mutex_unlock(mutex);
|
94 | 8653c015 | malc | if (ret) die2(ret, "pthread_mutex_unlock"); |
95 | 8653c015 | malc | } |
96 | 8653c015 | malc | |
97 | 8653c015 | malc | static int cond_timedwait(pthread_cond_t *cond, pthread_mutex_t *mutex, |
98 | 8653c015 | malc | struct timespec *ts)
|
99 | 8653c015 | malc | { |
100 | 8653c015 | malc | int ret = pthread_cond_timedwait(cond, mutex, ts);
|
101 | 8653c015 | malc | if (ret && ret != ETIMEDOUT) die2(ret, "pthread_cond_timedwait"); |
102 | 8653c015 | malc | return ret;
|
103 | 8653c015 | malc | } |
104 | 8653c015 | malc | |
105 | 5d47e372 | malc | static void cond_signal(pthread_cond_t *cond) |
106 | 8653c015 | malc | { |
107 | 5d47e372 | malc | int ret = pthread_cond_signal(cond);
|
108 | 5d47e372 | malc | if (ret) die2(ret, "pthread_cond_signal"); |
109 | 8653c015 | malc | } |
110 | 8653c015 | malc | |
111 | 8653c015 | malc | static void thread_create(pthread_t *thread, pthread_attr_t *attr, |
112 | 8653c015 | malc | void *(*start_routine)(void*), void *arg) |
113 | 8653c015 | malc | { |
114 | 8653c015 | malc | int ret = pthread_create(thread, attr, start_routine, arg);
|
115 | 8653c015 | malc | if (ret) die2(ret, "pthread_create"); |
116 | 8653c015 | malc | } |
117 | 8653c015 | malc | |
118 | f141eafe | aliguori | static size_t handle_aiocb_ioctl(struct qemu_paiocb *aiocb) |
119 | f141eafe | aliguori | { |
120 | f141eafe | aliguori | int ret;
|
121 | f141eafe | aliguori | |
122 | f141eafe | aliguori | ret = ioctl(aiocb->aio_fildes, aiocb->aio_ioctl_cmd, aiocb->aio_ioctl_buf); |
123 | f141eafe | aliguori | if (ret == -1) |
124 | f141eafe | aliguori | return -errno;
|
125 | e7d54ae8 | Christoph Hellwig | |
126 | e7d54ae8 | Christoph Hellwig | /*
|
127 | e7d54ae8 | Christoph Hellwig | * This looks weird, but the aio code only consideres a request
|
128 | e7d54ae8 | Christoph Hellwig | * successfull if it has written the number full number of bytes.
|
129 | e7d54ae8 | Christoph Hellwig | *
|
130 | e7d54ae8 | Christoph Hellwig | * Now we overload aio_nbytes as aio_ioctl_cmd for the ioctl command,
|
131 | e7d54ae8 | Christoph Hellwig | * so in fact we return the ioctl command here to make posix_aio_read()
|
132 | e7d54ae8 | Christoph Hellwig | * happy..
|
133 | e7d54ae8 | Christoph Hellwig | */
|
134 | e7d54ae8 | Christoph Hellwig | return aiocb->aio_nbytes;
|
135 | f141eafe | aliguori | } |
136 | f141eafe | aliguori | |
137 | 2341f9a1 | Juan Quintela | #ifdef CONFIG_PREADV
|
138 | ceb42de8 | aliguori | |
139 | ceb42de8 | aliguori | static ssize_t
|
140 | ceb42de8 | aliguori | qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset) |
141 | ceb42de8 | aliguori | { |
142 | ceb42de8 | aliguori | return preadv(fd, iov, nr_iov, offset);
|
143 | ceb42de8 | aliguori | } |
144 | ceb42de8 | aliguori | |
145 | ceb42de8 | aliguori | static ssize_t
|
146 | ceb42de8 | aliguori | qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset) |
147 | ceb42de8 | aliguori | { |
148 | ceb42de8 | aliguori | return pwritev(fd, iov, nr_iov, offset);
|
149 | ceb42de8 | aliguori | } |
150 | ceb42de8 | aliguori | |
151 | ceb42de8 | aliguori | #else
|
152 | ceb42de8 | aliguori | |
153 | ceb42de8 | aliguori | static ssize_t
|
154 | ceb42de8 | aliguori | qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset) |
155 | ceb42de8 | aliguori | { |
156 | ceb42de8 | aliguori | return -ENOSYS;
|
157 | ceb42de8 | aliguori | } |
158 | ceb42de8 | aliguori | |
159 | ceb42de8 | aliguori | static ssize_t
|
160 | ceb42de8 | aliguori | qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset) |
161 | ceb42de8 | aliguori | { |
162 | ceb42de8 | aliguori | return -ENOSYS;
|
163 | ceb42de8 | aliguori | } |
164 | ceb42de8 | aliguori | |
165 | ceb42de8 | aliguori | #endif
|
166 | ceb42de8 | aliguori | |
167 | ceb42de8 | aliguori | static size_t handle_aiocb_rw_vector(struct qemu_paiocb *aiocb) |
168 | ceb42de8 | aliguori | { |
169 | ceb42de8 | aliguori | size_t offset = 0;
|
170 | ceb42de8 | aliguori | ssize_t len; |
171 | ceb42de8 | aliguori | |
172 | ceb42de8 | aliguori | do {
|
173 | 9ef91a67 | Christoph Hellwig | if (aiocb->aio_type & QEMU_AIO_WRITE)
|
174 | ceb42de8 | aliguori | len = qemu_pwritev(aiocb->aio_fildes, |
175 | ceb42de8 | aliguori | aiocb->aio_iov, |
176 | ceb42de8 | aliguori | aiocb->aio_niov, |
177 | ceb42de8 | aliguori | aiocb->aio_offset + offset); |
178 | ceb42de8 | aliguori | else
|
179 | ceb42de8 | aliguori | len = qemu_preadv(aiocb->aio_fildes, |
180 | ceb42de8 | aliguori | aiocb->aio_iov, |
181 | ceb42de8 | aliguori | aiocb->aio_niov, |
182 | ceb42de8 | aliguori | aiocb->aio_offset + offset); |
183 | ceb42de8 | aliguori | } while (len == -1 && errno == EINTR); |
184 | ceb42de8 | aliguori | |
185 | ceb42de8 | aliguori | if (len == -1) |
186 | ceb42de8 | aliguori | return -errno;
|
187 | ceb42de8 | aliguori | return len;
|
188 | ceb42de8 | aliguori | } |
189 | ceb42de8 | aliguori | |
190 | f141eafe | aliguori | static size_t handle_aiocb_rw_linear(struct qemu_paiocb *aiocb, char *buf) |
191 | 221f715d | aliguori | { |
192 | 221f715d | aliguori | size_t offset = 0;
|
193 | f141eafe | aliguori | size_t len; |
194 | 221f715d | aliguori | |
195 | 221f715d | aliguori | while (offset < aiocb->aio_nbytes) {
|
196 | 9ef91a67 | Christoph Hellwig | if (aiocb->aio_type & QEMU_AIO_WRITE)
|
197 | f141eafe | aliguori | len = pwrite(aiocb->aio_fildes, |
198 | f141eafe | aliguori | (const char *)buf + offset, |
199 | f141eafe | aliguori | aiocb->aio_nbytes - offset, |
200 | f141eafe | aliguori | aiocb->aio_offset + offset); |
201 | f141eafe | aliguori | else
|
202 | f141eafe | aliguori | len = pread(aiocb->aio_fildes, |
203 | f141eafe | aliguori | buf + offset, |
204 | 221f715d | aliguori | aiocb->aio_nbytes - offset, |
205 | 221f715d | aliguori | aiocb->aio_offset + offset); |
206 | 221f715d | aliguori | |
207 | f141eafe | aliguori | if (len == -1 && errno == EINTR) |
208 | f141eafe | aliguori | continue;
|
209 | f141eafe | aliguori | else if (len == -1) { |
210 | f141eafe | aliguori | offset = -errno; |
211 | f141eafe | aliguori | break;
|
212 | f141eafe | aliguori | } else if (len == 0) |
213 | f141eafe | aliguori | break;
|
214 | f141eafe | aliguori | |
215 | f141eafe | aliguori | offset += len; |
216 | 221f715d | aliguori | } |
217 | 221f715d | aliguori | |
218 | 221f715d | aliguori | return offset;
|
219 | 221f715d | aliguori | } |
220 | 221f715d | aliguori | |
221 | f141eafe | aliguori | static size_t handle_aiocb_rw(struct qemu_paiocb *aiocb) |
222 | 221f715d | aliguori | { |
223 | f141eafe | aliguori | size_t nbytes; |
224 | f141eafe | aliguori | char *buf;
|
225 | f141eafe | aliguori | |
226 | 9ef91a67 | Christoph Hellwig | if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
|
227 | f141eafe | aliguori | /*
|
228 | f141eafe | aliguori | * If there is just a single buffer, and it is properly aligned
|
229 | f141eafe | aliguori | * we can just use plain pread/pwrite without any problems.
|
230 | f141eafe | aliguori | */
|
231 | ceb42de8 | aliguori | if (aiocb->aio_niov == 1) |
232 | ceb42de8 | aliguori | return handle_aiocb_rw_linear(aiocb, aiocb->aio_iov->iov_base);
|
233 | ceb42de8 | aliguori | |
234 | ceb42de8 | aliguori | /*
|
235 | ceb42de8 | aliguori | * We have more than one iovec, and all are properly aligned.
|
236 | ceb42de8 | aliguori | *
|
237 | ceb42de8 | aliguori | * Try preadv/pwritev first and fall back to linearizing the
|
238 | ceb42de8 | aliguori | * buffer if it's not supported.
|
239 | ceb42de8 | aliguori | */
|
240 | ceb42de8 | aliguori | if (preadv_present) {
|
241 | ceb42de8 | aliguori | nbytes = handle_aiocb_rw_vector(aiocb); |
242 | ceb42de8 | aliguori | if (nbytes == aiocb->aio_nbytes)
|
243 | ceb42de8 | aliguori | return nbytes;
|
244 | ceb42de8 | aliguori | if (nbytes < 0 && nbytes != -ENOSYS) |
245 | ceb42de8 | aliguori | return nbytes;
|
246 | ceb42de8 | aliguori | preadv_present = 0;
|
247 | ceb42de8 | aliguori | } |
248 | ceb42de8 | aliguori | |
249 | ceb42de8 | aliguori | /*
|
250 | ceb42de8 | aliguori | * XXX(hch): short read/write. no easy way to handle the reminder
|
251 | ceb42de8 | aliguori | * using these interfaces. For now retry using plain
|
252 | ceb42de8 | aliguori | * pread/pwrite?
|
253 | ceb42de8 | aliguori | */
|
254 | f141eafe | aliguori | } |
255 | 221f715d | aliguori | |
256 | f141eafe | aliguori | /*
|
257 | f141eafe | aliguori | * Ok, we have to do it the hard way, copy all segments into
|
258 | f141eafe | aliguori | * a single aligned buffer.
|
259 | f141eafe | aliguori | */
|
260 | f141eafe | aliguori | buf = qemu_memalign(512, aiocb->aio_nbytes);
|
261 | 9ef91a67 | Christoph Hellwig | if (aiocb->aio_type & QEMU_AIO_WRITE) {
|
262 | f141eafe | aliguori | char *p = buf;
|
263 | f141eafe | aliguori | int i;
|
264 | f141eafe | aliguori | |
265 | f141eafe | aliguori | for (i = 0; i < aiocb->aio_niov; ++i) { |
266 | f141eafe | aliguori | memcpy(p, aiocb->aio_iov[i].iov_base, aiocb->aio_iov[i].iov_len); |
267 | f141eafe | aliguori | p += aiocb->aio_iov[i].iov_len; |
268 | f141eafe | aliguori | } |
269 | f141eafe | aliguori | } |
270 | f141eafe | aliguori | |
271 | f141eafe | aliguori | nbytes = handle_aiocb_rw_linear(aiocb, buf); |
272 | 9ef91a67 | Christoph Hellwig | if (!(aiocb->aio_type & QEMU_AIO_WRITE)) {
|
273 | f141eafe | aliguori | char *p = buf;
|
274 | f141eafe | aliguori | size_t count = aiocb->aio_nbytes, copy; |
275 | f141eafe | aliguori | int i;
|
276 | f141eafe | aliguori | |
277 | f141eafe | aliguori | for (i = 0; i < aiocb->aio_niov && count; ++i) { |
278 | f141eafe | aliguori | copy = count; |
279 | f141eafe | aliguori | if (copy > aiocb->aio_iov[i].iov_len)
|
280 | f141eafe | aliguori | copy = aiocb->aio_iov[i].iov_len; |
281 | f141eafe | aliguori | memcpy(aiocb->aio_iov[i].iov_base, p, copy); |
282 | f141eafe | aliguori | p += copy; |
283 | f141eafe | aliguori | count -= copy; |
284 | f141eafe | aliguori | } |
285 | f141eafe | aliguori | } |
286 | f141eafe | aliguori | qemu_vfree(buf); |
287 | f141eafe | aliguori | |
288 | f141eafe | aliguori | return nbytes;
|
289 | 221f715d | aliguori | } |
290 | 221f715d | aliguori | |
291 | 3c529d93 | aliguori | static void *aio_thread(void *unused) |
292 | 3c529d93 | aliguori | { |
293 | a8227a5a | malc | pid_t pid; |
294 | 3c529d93 | aliguori | sigset_t set; |
295 | 3c529d93 | aliguori | |
296 | a8227a5a | malc | pid = getpid(); |
297 | a8227a5a | malc | |
298 | 3c529d93 | aliguori | /* block all signals */
|
299 | 8653c015 | malc | if (sigfillset(&set)) die("sigfillset"); |
300 | 8653c015 | malc | if (sigprocmask(SIG_BLOCK, &set, NULL)) die("sigprocmask"); |
301 | 3c529d93 | aliguori | |
302 | 3c529d93 | aliguori | while (1) { |
303 | 3c529d93 | aliguori | struct qemu_paiocb *aiocb;
|
304 | 221f715d | aliguori | size_t ret = 0;
|
305 | 30525aff | malc | qemu_timeval tv; |
306 | 30525aff | malc | struct timespec ts;
|
307 | 30525aff | malc | |
308 | 30525aff | malc | qemu_gettimeofday(&tv); |
309 | 30525aff | malc | ts.tv_sec = tv.tv_sec + 10;
|
310 | 30525aff | malc | ts.tv_nsec = 0;
|
311 | 3c529d93 | aliguori | |
312 | 8653c015 | malc | mutex_lock(&lock); |
313 | 3c529d93 | aliguori | |
314 | 3c529d93 | aliguori | while (TAILQ_EMPTY(&request_list) &&
|
315 | 3c529d93 | aliguori | !(ret == ETIMEDOUT)) { |
316 | 8653c015 | malc | ret = cond_timedwait(&cond, &lock, &ts); |
317 | 3c529d93 | aliguori | } |
318 | 3c529d93 | aliguori | |
319 | 514f7a27 | malc | if (TAILQ_EMPTY(&request_list))
|
320 | 3c529d93 | aliguori | break;
|
321 | 3c529d93 | aliguori | |
322 | 3c529d93 | aliguori | aiocb = TAILQ_FIRST(&request_list); |
323 | 3c529d93 | aliguori | TAILQ_REMOVE(&request_list, aiocb, node); |
324 | 3c529d93 | aliguori | aiocb->active = 1;
|
325 | 3c529d93 | aliguori | idle_threads--; |
326 | 8653c015 | malc | mutex_unlock(&lock); |
327 | 3c529d93 | aliguori | |
328 | 9ef91a67 | Christoph Hellwig | switch (aiocb->aio_type & QEMU_AIO_TYPE_MASK) {
|
329 | 9ef91a67 | Christoph Hellwig | case QEMU_AIO_READ:
|
330 | 9ef91a67 | Christoph Hellwig | case QEMU_AIO_WRITE:
|
331 | f141eafe | aliguori | ret = handle_aiocb_rw(aiocb); |
332 | 221f715d | aliguori | break;
|
333 | 9ef91a67 | Christoph Hellwig | case QEMU_AIO_IOCTL:
|
334 | 221f715d | aliguori | ret = handle_aiocb_ioctl(aiocb); |
335 | 221f715d | aliguori | break;
|
336 | 221f715d | aliguori | default:
|
337 | 221f715d | aliguori | fprintf(stderr, "invalid aio request (0x%x)\n", aiocb->aio_type);
|
338 | 221f715d | aliguori | ret = -EINVAL; |
339 | 221f715d | aliguori | break;
|
340 | 221f715d | aliguori | } |
341 | 3c529d93 | aliguori | |
342 | 8653c015 | malc | mutex_lock(&lock); |
343 | 221f715d | aliguori | aiocb->ret = ret; |
344 | 3c529d93 | aliguori | idle_threads++; |
345 | 8653c015 | malc | mutex_unlock(&lock); |
346 | 3c529d93 | aliguori | |
347 | a8227a5a | malc | if (kill(pid, aiocb->ev_signo)) die("kill failed"); |
348 | 3c529d93 | aliguori | } |
349 | 3c529d93 | aliguori | |
350 | 3c529d93 | aliguori | idle_threads--; |
351 | 3c529d93 | aliguori | cur_threads--; |
352 | 8653c015 | malc | mutex_unlock(&lock); |
353 | 3c529d93 | aliguori | |
354 | 3c529d93 | aliguori | return NULL; |
355 | 3c529d93 | aliguori | } |
356 | 3c529d93 | aliguori | |
357 | 8653c015 | malc | static void spawn_thread(void) |
358 | 3c529d93 | aliguori | { |
359 | 3c529d93 | aliguori | cur_threads++; |
360 | 3c529d93 | aliguori | idle_threads++; |
361 | 8653c015 | malc | thread_create(&thread_id, &attr, aio_thread, NULL);
|
362 | 3c529d93 | aliguori | } |
363 | 3c529d93 | aliguori | |
364 | 9ef91a67 | Christoph Hellwig | static void qemu_paio_submit(struct qemu_paiocb *aiocb) |
365 | 3c529d93 | aliguori | { |
366 | 3c529d93 | aliguori | aiocb->ret = -EINPROGRESS; |
367 | 3c529d93 | aliguori | aiocb->active = 0;
|
368 | 8653c015 | malc | mutex_lock(&lock); |
369 | 3c529d93 | aliguori | if (idle_threads == 0 && cur_threads < max_threads) |
370 | 3c529d93 | aliguori | spawn_thread(); |
371 | 3c529d93 | aliguori | TAILQ_INSERT_TAIL(&request_list, aiocb, node); |
372 | 8653c015 | malc | mutex_unlock(&lock); |
373 | 5d47e372 | malc | cond_signal(&cond); |
374 | 3c529d93 | aliguori | } |
375 | 3c529d93 | aliguori | |
376 | 9ef91a67 | Christoph Hellwig | static ssize_t qemu_paio_return(struct qemu_paiocb *aiocb) |
377 | 3c529d93 | aliguori | { |
378 | 3c529d93 | aliguori | ssize_t ret; |
379 | 3c529d93 | aliguori | |
380 | 8653c015 | malc | mutex_lock(&lock); |
381 | 3c529d93 | aliguori | ret = aiocb->ret; |
382 | 8653c015 | malc | mutex_unlock(&lock); |
383 | 3c529d93 | aliguori | |
384 | 3c529d93 | aliguori | return ret;
|
385 | 3c529d93 | aliguori | } |
386 | 3c529d93 | aliguori | |
387 | 9ef91a67 | Christoph Hellwig | static int qemu_paio_error(struct qemu_paiocb *aiocb) |
388 | 3c529d93 | aliguori | { |
389 | 3c529d93 | aliguori | ssize_t ret = qemu_paio_return(aiocb); |
390 | 3c529d93 | aliguori | |
391 | 3c529d93 | aliguori | if (ret < 0) |
392 | 3c529d93 | aliguori | ret = -ret; |
393 | 3c529d93 | aliguori | else
|
394 | 3c529d93 | aliguori | ret = 0;
|
395 | 3c529d93 | aliguori | |
396 | 3c529d93 | aliguori | return ret;
|
397 | 3c529d93 | aliguori | } |
398 | 3c529d93 | aliguori | |
399 | 9ef91a67 | Christoph Hellwig | static void posix_aio_read(void *opaque) |
400 | 3c529d93 | aliguori | { |
401 | 9ef91a67 | Christoph Hellwig | PosixAioState *s = opaque; |
402 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *acb, **pacb;
|
403 | 3c529d93 | aliguori | int ret;
|
404 | 9ef91a67 | Christoph Hellwig | ssize_t len; |
405 | 9ef91a67 | Christoph Hellwig | |
406 | 9ef91a67 | Christoph Hellwig | /* read all bytes from signal pipe */
|
407 | 9ef91a67 | Christoph Hellwig | for (;;) {
|
408 | 9ef91a67 | Christoph Hellwig | char bytes[16]; |
409 | 9ef91a67 | Christoph Hellwig | |
410 | 9ef91a67 | Christoph Hellwig | len = read(s->rfd, bytes, sizeof(bytes));
|
411 | 9ef91a67 | Christoph Hellwig | if (len == -1 && errno == EINTR) |
412 | 9ef91a67 | Christoph Hellwig | continue; /* try again */ |
413 | 9ef91a67 | Christoph Hellwig | if (len == sizeof(bytes)) |
414 | 9ef91a67 | Christoph Hellwig | continue; /* more to read */ |
415 | 9ef91a67 | Christoph Hellwig | break;
|
416 | 9ef91a67 | Christoph Hellwig | } |
417 | 9ef91a67 | Christoph Hellwig | |
418 | 9ef91a67 | Christoph Hellwig | for(;;) {
|
419 | 9ef91a67 | Christoph Hellwig | pacb = &s->first_aio; |
420 | 9ef91a67 | Christoph Hellwig | for(;;) {
|
421 | 9ef91a67 | Christoph Hellwig | acb = *pacb; |
422 | 9ef91a67 | Christoph Hellwig | if (!acb)
|
423 | 9ef91a67 | Christoph Hellwig | goto the_end;
|
424 | 9ef91a67 | Christoph Hellwig | ret = qemu_paio_error(acb); |
425 | 9ef91a67 | Christoph Hellwig | if (ret == ECANCELED) {
|
426 | 9ef91a67 | Christoph Hellwig | /* remove the request */
|
427 | 9ef91a67 | Christoph Hellwig | *pacb = acb->next; |
428 | 9ef91a67 | Christoph Hellwig | qemu_aio_release(acb); |
429 | 9ef91a67 | Christoph Hellwig | } else if (ret != EINPROGRESS) { |
430 | 9ef91a67 | Christoph Hellwig | /* end of aio */
|
431 | 9ef91a67 | Christoph Hellwig | if (ret == 0) { |
432 | 9ef91a67 | Christoph Hellwig | ret = qemu_paio_return(acb); |
433 | 9ef91a67 | Christoph Hellwig | if (ret == acb->aio_nbytes)
|
434 | 9ef91a67 | Christoph Hellwig | ret = 0;
|
435 | 9ef91a67 | Christoph Hellwig | else
|
436 | 9ef91a67 | Christoph Hellwig | ret = -EINVAL; |
437 | 9ef91a67 | Christoph Hellwig | } else {
|
438 | 9ef91a67 | Christoph Hellwig | ret = -ret; |
439 | 9ef91a67 | Christoph Hellwig | } |
440 | 9ef91a67 | Christoph Hellwig | /* remove the request */
|
441 | 9ef91a67 | Christoph Hellwig | *pacb = acb->next; |
442 | 9ef91a67 | Christoph Hellwig | /* call the callback */
|
443 | 9ef91a67 | Christoph Hellwig | acb->common.cb(acb->common.opaque, ret); |
444 | 9ef91a67 | Christoph Hellwig | qemu_aio_release(acb); |
445 | 9ef91a67 | Christoph Hellwig | break;
|
446 | 9ef91a67 | Christoph Hellwig | } else {
|
447 | 9ef91a67 | Christoph Hellwig | pacb = &acb->next; |
448 | 9ef91a67 | Christoph Hellwig | } |
449 | 9ef91a67 | Christoph Hellwig | } |
450 | 9ef91a67 | Christoph Hellwig | } |
451 | 9ef91a67 | Christoph Hellwig | the_end: ;
|
452 | 9ef91a67 | Christoph Hellwig | } |
453 | 9ef91a67 | Christoph Hellwig | |
454 | 9ef91a67 | Christoph Hellwig | static int posix_aio_flush(void *opaque) |
455 | 9ef91a67 | Christoph Hellwig | { |
456 | 9ef91a67 | Christoph Hellwig | PosixAioState *s = opaque; |
457 | 9ef91a67 | Christoph Hellwig | return !!s->first_aio;
|
458 | 9ef91a67 | Christoph Hellwig | } |
459 | 9ef91a67 | Christoph Hellwig | |
460 | 9ef91a67 | Christoph Hellwig | static PosixAioState *posix_aio_state;
|
461 | 9ef91a67 | Christoph Hellwig | |
462 | 9ef91a67 | Christoph Hellwig | static void aio_signal_handler(int signum) |
463 | 9ef91a67 | Christoph Hellwig | { |
464 | 9ef91a67 | Christoph Hellwig | if (posix_aio_state) {
|
465 | 9ef91a67 | Christoph Hellwig | char byte = 0; |
466 | 9ef91a67 | Christoph Hellwig | |
467 | 9ef91a67 | Christoph Hellwig | write(posix_aio_state->wfd, &byte, sizeof(byte));
|
468 | 9ef91a67 | Christoph Hellwig | } |
469 | 9ef91a67 | Christoph Hellwig | |
470 | 9ef91a67 | Christoph Hellwig | qemu_service_io(); |
471 | 9ef91a67 | Christoph Hellwig | } |
472 | 9ef91a67 | Christoph Hellwig | |
473 | 9ef91a67 | Christoph Hellwig | static void paio_remove(struct qemu_paiocb *acb) |
474 | 9ef91a67 | Christoph Hellwig | { |
475 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb **pacb;
|
476 | 9ef91a67 | Christoph Hellwig | |
477 | 9ef91a67 | Christoph Hellwig | /* remove the callback from the queue */
|
478 | 9ef91a67 | Christoph Hellwig | pacb = &posix_aio_state->first_aio; |
479 | 9ef91a67 | Christoph Hellwig | for(;;) {
|
480 | 9ef91a67 | Christoph Hellwig | if (*pacb == NULL) { |
481 | 9ef91a67 | Christoph Hellwig | fprintf(stderr, "paio_remove: aio request not found!\n");
|
482 | 9ef91a67 | Christoph Hellwig | break;
|
483 | 9ef91a67 | Christoph Hellwig | } else if (*pacb == acb) { |
484 | 9ef91a67 | Christoph Hellwig | *pacb = acb->next; |
485 | 9ef91a67 | Christoph Hellwig | qemu_aio_release(acb); |
486 | 9ef91a67 | Christoph Hellwig | break;
|
487 | 9ef91a67 | Christoph Hellwig | } |
488 | 9ef91a67 | Christoph Hellwig | pacb = &(*pacb)->next; |
489 | 9ef91a67 | Christoph Hellwig | } |
490 | 9ef91a67 | Christoph Hellwig | } |
491 | 9ef91a67 | Christoph Hellwig | |
492 | 9ef91a67 | Christoph Hellwig | static void paio_cancel(BlockDriverAIOCB *blockacb) |
493 | 9ef91a67 | Christoph Hellwig | { |
494 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *acb = (struct qemu_paiocb *)blockacb; |
495 | 9ef91a67 | Christoph Hellwig | int active = 0; |
496 | 3c529d93 | aliguori | |
497 | 8653c015 | malc | mutex_lock(&lock); |
498 | 9ef91a67 | Christoph Hellwig | if (!acb->active) {
|
499 | 9ef91a67 | Christoph Hellwig | TAILQ_REMOVE(&request_list, acb, node); |
500 | 9ef91a67 | Christoph Hellwig | acb->ret = -ECANCELED; |
501 | 9ef91a67 | Christoph Hellwig | } else if (acb->ret == -EINPROGRESS) { |
502 | 9ef91a67 | Christoph Hellwig | active = 1;
|
503 | 9ef91a67 | Christoph Hellwig | } |
504 | 8653c015 | malc | mutex_unlock(&lock); |
505 | 3c529d93 | aliguori | |
506 | 9ef91a67 | Christoph Hellwig | if (active) {
|
507 | 9ef91a67 | Christoph Hellwig | /* fail safe: if the aio could not be canceled, we wait for
|
508 | 9ef91a67 | Christoph Hellwig | it */
|
509 | 9ef91a67 | Christoph Hellwig | while (qemu_paio_error(acb) == EINPROGRESS)
|
510 | 9ef91a67 | Christoph Hellwig | ; |
511 | 9ef91a67 | Christoph Hellwig | } |
512 | 9ef91a67 | Christoph Hellwig | |
513 | 9ef91a67 | Christoph Hellwig | paio_remove(acb); |
514 | 9ef91a67 | Christoph Hellwig | } |
515 | 9ef91a67 | Christoph Hellwig | |
516 | 9ef91a67 | Christoph Hellwig | static AIOPool raw_aio_pool = {
|
517 | 9ef91a67 | Christoph Hellwig | .aiocb_size = sizeof(struct qemu_paiocb), |
518 | 9ef91a67 | Christoph Hellwig | .cancel = paio_cancel, |
519 | 9ef91a67 | Christoph Hellwig | }; |
520 | 9ef91a67 | Christoph Hellwig | |
521 | 9ef91a67 | Christoph Hellwig | BlockDriverAIOCB *paio_submit(BlockDriverState *bs, void *aio_ctx, int fd, |
522 | 9ef91a67 | Christoph Hellwig | int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
|
523 | 9ef91a67 | Christoph Hellwig | BlockDriverCompletionFunc *cb, void *opaque, int type) |
524 | 9ef91a67 | Christoph Hellwig | { |
525 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *acb;
|
526 | 9ef91a67 | Christoph Hellwig | |
527 | 9ef91a67 | Christoph Hellwig | acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque); |
528 | 9ef91a67 | Christoph Hellwig | if (!acb)
|
529 | 9ef91a67 | Christoph Hellwig | return NULL; |
530 | 9ef91a67 | Christoph Hellwig | acb->aio_type = type; |
531 | 9ef91a67 | Christoph Hellwig | acb->aio_fildes = fd; |
532 | 9ef91a67 | Christoph Hellwig | acb->ev_signo = SIGUSR2; |
533 | 9ef91a67 | Christoph Hellwig | acb->aio_iov = qiov->iov; |
534 | 9ef91a67 | Christoph Hellwig | acb->aio_niov = qiov->niov; |
535 | 9ef91a67 | Christoph Hellwig | acb->aio_nbytes = nb_sectors * 512;
|
536 | 9ef91a67 | Christoph Hellwig | acb->aio_offset = sector_num * 512;
|
537 | 9ef91a67 | Christoph Hellwig | |
538 | 9ef91a67 | Christoph Hellwig | acb->next = posix_aio_state->first_aio; |
539 | 9ef91a67 | Christoph Hellwig | posix_aio_state->first_aio = acb; |
540 | 9ef91a67 | Christoph Hellwig | |
541 | 9ef91a67 | Christoph Hellwig | qemu_paio_submit(acb); |
542 | 9ef91a67 | Christoph Hellwig | return &acb->common;
|
543 | 9ef91a67 | Christoph Hellwig | } |
544 | 9ef91a67 | Christoph Hellwig | |
545 | 9ef91a67 | Christoph Hellwig | BlockDriverAIOCB *paio_ioctl(BlockDriverState *bs, int fd,
|
546 | 9ef91a67 | Christoph Hellwig | unsigned long int req, void *buf, |
547 | 9ef91a67 | Christoph Hellwig | BlockDriverCompletionFunc *cb, void *opaque)
|
548 | 9ef91a67 | Christoph Hellwig | { |
549 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *acb;
|
550 | 9ef91a67 | Christoph Hellwig | |
551 | 9ef91a67 | Christoph Hellwig | acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque); |
552 | 9ef91a67 | Christoph Hellwig | if (!acb)
|
553 | 9ef91a67 | Christoph Hellwig | return NULL; |
554 | 9ef91a67 | Christoph Hellwig | acb->aio_type = QEMU_AIO_IOCTL; |
555 | 9ef91a67 | Christoph Hellwig | acb->aio_fildes = fd; |
556 | 9ef91a67 | Christoph Hellwig | acb->ev_signo = SIGUSR2; |
557 | 9ef91a67 | Christoph Hellwig | acb->aio_offset = 0;
|
558 | 9ef91a67 | Christoph Hellwig | acb->aio_ioctl_buf = buf; |
559 | 9ef91a67 | Christoph Hellwig | acb->aio_ioctl_cmd = req; |
560 | 9ef91a67 | Christoph Hellwig | |
561 | 9ef91a67 | Christoph Hellwig | acb->next = posix_aio_state->first_aio; |
562 | 9ef91a67 | Christoph Hellwig | posix_aio_state->first_aio = acb; |
563 | 9ef91a67 | Christoph Hellwig | |
564 | 9ef91a67 | Christoph Hellwig | qemu_paio_submit(acb); |
565 | 9ef91a67 | Christoph Hellwig | return &acb->common;
|
566 | 9ef91a67 | Christoph Hellwig | } |
567 | 9ef91a67 | Christoph Hellwig | |
568 | 9ef91a67 | Christoph Hellwig | void *paio_init(void) |
569 | 9ef91a67 | Christoph Hellwig | { |
570 | 9ef91a67 | Christoph Hellwig | struct sigaction act;
|
571 | 9ef91a67 | Christoph Hellwig | PosixAioState *s; |
572 | 9ef91a67 | Christoph Hellwig | int fds[2]; |
573 | 9ef91a67 | Christoph Hellwig | int ret;
|
574 | 9ef91a67 | Christoph Hellwig | |
575 | 9ef91a67 | Christoph Hellwig | if (posix_aio_state)
|
576 | 9ef91a67 | Christoph Hellwig | return posix_aio_state;
|
577 | 9ef91a67 | Christoph Hellwig | |
578 | 9ef91a67 | Christoph Hellwig | s = qemu_malloc(sizeof(PosixAioState));
|
579 | 9ef91a67 | Christoph Hellwig | |
580 | 9ef91a67 | Christoph Hellwig | sigfillset(&act.sa_mask); |
581 | 9ef91a67 | Christoph Hellwig | act.sa_flags = 0; /* do not restart syscalls to interrupt select() */ |
582 | 9ef91a67 | Christoph Hellwig | act.sa_handler = aio_signal_handler; |
583 | 9ef91a67 | Christoph Hellwig | sigaction(SIGUSR2, &act, NULL);
|
584 | 9ef91a67 | Christoph Hellwig | |
585 | 9ef91a67 | Christoph Hellwig | s->first_aio = NULL;
|
586 | 9ef91a67 | Christoph Hellwig | if (pipe(fds) == -1) { |
587 | 9ef91a67 | Christoph Hellwig | fprintf(stderr, "failed to create pipe\n");
|
588 | 9ef91a67 | Christoph Hellwig | return NULL; |
589 | 9ef91a67 | Christoph Hellwig | } |
590 | 9ef91a67 | Christoph Hellwig | |
591 | 9ef91a67 | Christoph Hellwig | s->rfd = fds[0];
|
592 | 9ef91a67 | Christoph Hellwig | s->wfd = fds[1];
|
593 | 9ef91a67 | Christoph Hellwig | |
594 | 9ef91a67 | Christoph Hellwig | fcntl(s->rfd, F_SETFL, O_NONBLOCK); |
595 | 9ef91a67 | Christoph Hellwig | fcntl(s->wfd, F_SETFL, O_NONBLOCK); |
596 | 9ef91a67 | Christoph Hellwig | |
597 | 9ef91a67 | Christoph Hellwig | qemu_aio_set_fd_handler(s->rfd, posix_aio_read, NULL, posix_aio_flush, s);
|
598 | 9ef91a67 | Christoph Hellwig | |
599 | 9ef91a67 | Christoph Hellwig | ret = pthread_attr_init(&attr); |
600 | 9ef91a67 | Christoph Hellwig | if (ret)
|
601 | 9ef91a67 | Christoph Hellwig | die2(ret, "pthread_attr_init");
|
602 | 9ef91a67 | Christoph Hellwig | |
603 | 9ef91a67 | Christoph Hellwig | ret = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED); |
604 | 9ef91a67 | Christoph Hellwig | if (ret)
|
605 | 9ef91a67 | Christoph Hellwig | die2(ret, "pthread_attr_setdetachstate");
|
606 | 9ef91a67 | Christoph Hellwig | |
607 | 9ef91a67 | Christoph Hellwig | TAILQ_INIT(&request_list); |
608 | 9ef91a67 | Christoph Hellwig | |
609 | 9ef91a67 | Christoph Hellwig | posix_aio_state = s; |
610 | 9ef91a67 | Christoph Hellwig | |
611 | 9ef91a67 | Christoph Hellwig | return posix_aio_state;
|
612 | 3c529d93 | aliguori | } |