root / posix-aio-compat.c @ 212b6008
History | View | Annotate | Download (15.8 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 | 72cf2d4f | Blue Swirl | #include "qemu-queue.h" |
26 | 3c529d93 | aliguori | #include "osdep.h" |
27 | f141eafe | aliguori | #include "qemu-common.h" |
28 | 6d519a5f | Stefan Hajnoczi | #include "trace.h" |
29 | 9ef91a67 | Christoph Hellwig | #include "block_int.h" |
30 | 9ef91a67 | Christoph Hellwig | |
31 | 9ef91a67 | Christoph Hellwig | #include "block/raw-posix-aio.h" |
32 | 9ef91a67 | Christoph Hellwig | |
33 | 9ef91a67 | Christoph Hellwig | |
34 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb {
|
35 | 9ef91a67 | Christoph Hellwig | BlockDriverAIOCB common; |
36 | 9ef91a67 | Christoph Hellwig | int aio_fildes;
|
37 | 9ef91a67 | Christoph Hellwig | union {
|
38 | 9ef91a67 | Christoph Hellwig | struct iovec *aio_iov;
|
39 | b587a52c | Stefan Hajnoczi | void *aio_ioctl_buf;
|
40 | 9ef91a67 | Christoph Hellwig | }; |
41 | 9ef91a67 | Christoph Hellwig | int aio_niov;
|
42 | 9ef91a67 | Christoph Hellwig | size_t aio_nbytes; |
43 | 9ef91a67 | Christoph Hellwig | #define aio_ioctl_cmd aio_nbytes /* for QEMU_AIO_IOCTL */ |
44 | 9ef91a67 | Christoph Hellwig | int ev_signo;
|
45 | 9ef91a67 | Christoph Hellwig | off_t aio_offset; |
46 | 9ef91a67 | Christoph Hellwig | |
47 | 72cf2d4f | Blue Swirl | QTAILQ_ENTRY(qemu_paiocb) node; |
48 | 9ef91a67 | Christoph Hellwig | int aio_type;
|
49 | 9ef91a67 | Christoph Hellwig | ssize_t ret; |
50 | 9ef91a67 | Christoph Hellwig | int active;
|
51 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *next;
|
52 | e5f37649 | Kevin Wolf | |
53 | e5f37649 | Kevin Wolf | int async_context_id;
|
54 | 9ef91a67 | Christoph Hellwig | }; |
55 | 9ef91a67 | Christoph Hellwig | |
56 | 9ef91a67 | Christoph Hellwig | typedef struct PosixAioState { |
57 | 9ef91a67 | Christoph Hellwig | int rfd, wfd;
|
58 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *first_aio;
|
59 | 9ef91a67 | Christoph Hellwig | } PosixAioState; |
60 | 3c529d93 | aliguori | |
61 | 3c529d93 | aliguori | |
62 | 3c529d93 | aliguori | static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
|
63 | 3c529d93 | aliguori | static pthread_cond_t cond = PTHREAD_COND_INITIALIZER;
|
64 | 3c529d93 | aliguori | static pthread_t thread_id;
|
65 | a8227a5a | malc | static pthread_attr_t attr;
|
66 | 3c529d93 | aliguori | static int max_threads = 64; |
67 | 3c529d93 | aliguori | static int cur_threads = 0; |
68 | 3c529d93 | aliguori | static int idle_threads = 0; |
69 | 72cf2d4f | Blue Swirl | static QTAILQ_HEAD(, qemu_paiocb) request_list;
|
70 | 3c529d93 | aliguori | |
71 | 2341f9a1 | Juan Quintela | #ifdef CONFIG_PREADV
|
72 | ceb42de8 | aliguori | static int preadv_present = 1; |
73 | ceb42de8 | aliguori | #else
|
74 | ceb42de8 | aliguori | static int preadv_present = 0; |
75 | ceb42de8 | aliguori | #endif
|
76 | ceb42de8 | aliguori | |
77 | 8653c015 | malc | static void die2(int err, const char *what) |
78 | 8653c015 | malc | { |
79 | 8653c015 | malc | fprintf(stderr, "%s failed: %s\n", what, strerror(err));
|
80 | 8653c015 | malc | abort(); |
81 | 8653c015 | malc | } |
82 | 8653c015 | malc | |
83 | 8653c015 | malc | static void die(const char *what) |
84 | 8653c015 | malc | { |
85 | 8653c015 | malc | die2(errno, what); |
86 | 8653c015 | malc | } |
87 | 8653c015 | malc | |
88 | 8653c015 | malc | static void mutex_lock(pthread_mutex_t *mutex) |
89 | 8653c015 | malc | { |
90 | 8653c015 | malc | int ret = pthread_mutex_lock(mutex);
|
91 | 8653c015 | malc | if (ret) die2(ret, "pthread_mutex_lock"); |
92 | 8653c015 | malc | } |
93 | 8653c015 | malc | |
94 | 8653c015 | malc | static void mutex_unlock(pthread_mutex_t *mutex) |
95 | 8653c015 | malc | { |
96 | 8653c015 | malc | int ret = pthread_mutex_unlock(mutex);
|
97 | 8653c015 | malc | if (ret) die2(ret, "pthread_mutex_unlock"); |
98 | 8653c015 | malc | } |
99 | 8653c015 | malc | |
100 | 8653c015 | malc | static int cond_timedwait(pthread_cond_t *cond, pthread_mutex_t *mutex, |
101 | 8653c015 | malc | struct timespec *ts)
|
102 | 8653c015 | malc | { |
103 | 8653c015 | malc | int ret = pthread_cond_timedwait(cond, mutex, ts);
|
104 | 8653c015 | malc | if (ret && ret != ETIMEDOUT) die2(ret, "pthread_cond_timedwait"); |
105 | 8653c015 | malc | return ret;
|
106 | 8653c015 | malc | } |
107 | 8653c015 | malc | |
108 | 5d47e372 | malc | static void cond_signal(pthread_cond_t *cond) |
109 | 8653c015 | malc | { |
110 | 5d47e372 | malc | int ret = pthread_cond_signal(cond);
|
111 | 5d47e372 | malc | if (ret) die2(ret, "pthread_cond_signal"); |
112 | 8653c015 | malc | } |
113 | 8653c015 | malc | |
114 | 8653c015 | malc | static void thread_create(pthread_t *thread, pthread_attr_t *attr, |
115 | 8653c015 | malc | void *(*start_routine)(void*), void *arg) |
116 | 8653c015 | malc | { |
117 | 8653c015 | malc | int ret = pthread_create(thread, attr, start_routine, arg);
|
118 | 8653c015 | malc | if (ret) die2(ret, "pthread_create"); |
119 | 8653c015 | malc | } |
120 | 8653c015 | malc | |
121 | 6769da29 | Kevin Wolf | static ssize_t handle_aiocb_ioctl(struct qemu_paiocb *aiocb) |
122 | f141eafe | aliguori | { |
123 | b587a52c | Stefan Hajnoczi | int ret;
|
124 | b587a52c | Stefan Hajnoczi | |
125 | b587a52c | Stefan Hajnoczi | ret = ioctl(aiocb->aio_fildes, aiocb->aio_ioctl_cmd, aiocb->aio_ioctl_buf); |
126 | b587a52c | Stefan Hajnoczi | if (ret == -1) |
127 | b587a52c | Stefan Hajnoczi | return -errno;
|
128 | b587a52c | Stefan Hajnoczi | |
129 | b587a52c | Stefan Hajnoczi | /*
|
130 | b587a52c | Stefan Hajnoczi | * This looks weird, but the aio code only consideres a request
|
131 | b587a52c | Stefan Hajnoczi | * successfull if it has written the number full number of bytes.
|
132 | b587a52c | Stefan Hajnoczi | *
|
133 | b587a52c | Stefan Hajnoczi | * Now we overload aio_nbytes as aio_ioctl_cmd for the ioctl command,
|
134 | b587a52c | Stefan Hajnoczi | * so in fact we return the ioctl command here to make posix_aio_read()
|
135 | b587a52c | Stefan Hajnoczi | * happy..
|
136 | b587a52c | Stefan Hajnoczi | */
|
137 | b587a52c | Stefan Hajnoczi | return aiocb->aio_nbytes;
|
138 | f141eafe | aliguori | } |
139 | f141eafe | aliguori | |
140 | 6769da29 | Kevin Wolf | static ssize_t handle_aiocb_flush(struct qemu_paiocb *aiocb) |
141 | b2e12bc6 | Christoph Hellwig | { |
142 | b2e12bc6 | Christoph Hellwig | int ret;
|
143 | b2e12bc6 | Christoph Hellwig | |
144 | 47faadc6 | Blue Swirl | ret = qemu_fdatasync(aiocb->aio_fildes); |
145 | b2e12bc6 | Christoph Hellwig | if (ret == -1) |
146 | b2e12bc6 | Christoph Hellwig | return -errno;
|
147 | b2e12bc6 | Christoph Hellwig | return 0; |
148 | b2e12bc6 | Christoph Hellwig | } |
149 | b2e12bc6 | Christoph Hellwig | |
150 | 2341f9a1 | Juan Quintela | #ifdef CONFIG_PREADV
|
151 | ceb42de8 | aliguori | |
152 | ceb42de8 | aliguori | static ssize_t
|
153 | ceb42de8 | aliguori | qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset) |
154 | ceb42de8 | aliguori | { |
155 | ceb42de8 | aliguori | return preadv(fd, iov, nr_iov, offset);
|
156 | ceb42de8 | aliguori | } |
157 | ceb42de8 | aliguori | |
158 | ceb42de8 | aliguori | static ssize_t
|
159 | ceb42de8 | aliguori | qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset) |
160 | ceb42de8 | aliguori | { |
161 | ceb42de8 | aliguori | return pwritev(fd, iov, nr_iov, offset);
|
162 | ceb42de8 | aliguori | } |
163 | ceb42de8 | aliguori | |
164 | ceb42de8 | aliguori | #else
|
165 | ceb42de8 | aliguori | |
166 | ceb42de8 | aliguori | static ssize_t
|
167 | ceb42de8 | aliguori | qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset) |
168 | ceb42de8 | aliguori | { |
169 | ceb42de8 | aliguori | return -ENOSYS;
|
170 | ceb42de8 | aliguori | } |
171 | ceb42de8 | aliguori | |
172 | ceb42de8 | aliguori | static ssize_t
|
173 | ceb42de8 | aliguori | qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset) |
174 | ceb42de8 | aliguori | { |
175 | ceb42de8 | aliguori | return -ENOSYS;
|
176 | ceb42de8 | aliguori | } |
177 | ceb42de8 | aliguori | |
178 | ceb42de8 | aliguori | #endif
|
179 | ceb42de8 | aliguori | |
180 | 6769da29 | Kevin Wolf | static ssize_t handle_aiocb_rw_vector(struct qemu_paiocb *aiocb) |
181 | ceb42de8 | aliguori | { |
182 | ceb42de8 | aliguori | size_t offset = 0;
|
183 | ceb42de8 | aliguori | ssize_t len; |
184 | ceb42de8 | aliguori | |
185 | ceb42de8 | aliguori | do {
|
186 | 9ef91a67 | Christoph Hellwig | if (aiocb->aio_type & QEMU_AIO_WRITE)
|
187 | ceb42de8 | aliguori | len = qemu_pwritev(aiocb->aio_fildes, |
188 | ceb42de8 | aliguori | aiocb->aio_iov, |
189 | ceb42de8 | aliguori | aiocb->aio_niov, |
190 | ceb42de8 | aliguori | aiocb->aio_offset + offset); |
191 | ceb42de8 | aliguori | else
|
192 | ceb42de8 | aliguori | len = qemu_preadv(aiocb->aio_fildes, |
193 | ceb42de8 | aliguori | aiocb->aio_iov, |
194 | ceb42de8 | aliguori | aiocb->aio_niov, |
195 | ceb42de8 | aliguori | aiocb->aio_offset + offset); |
196 | ceb42de8 | aliguori | } while (len == -1 && errno == EINTR); |
197 | ceb42de8 | aliguori | |
198 | ceb42de8 | aliguori | if (len == -1) |
199 | ceb42de8 | aliguori | return -errno;
|
200 | ceb42de8 | aliguori | return len;
|
201 | ceb42de8 | aliguori | } |
202 | ceb42de8 | aliguori | |
203 | 6769da29 | Kevin Wolf | static ssize_t handle_aiocb_rw_linear(struct qemu_paiocb *aiocb, char *buf) |
204 | 221f715d | aliguori | { |
205 | 6769da29 | Kevin Wolf | ssize_t offset = 0;
|
206 | 6769da29 | Kevin Wolf | ssize_t len; |
207 | 221f715d | aliguori | |
208 | 221f715d | aliguori | while (offset < aiocb->aio_nbytes) {
|
209 | 9ef91a67 | Christoph Hellwig | if (aiocb->aio_type & QEMU_AIO_WRITE)
|
210 | f141eafe | aliguori | len = pwrite(aiocb->aio_fildes, |
211 | f141eafe | aliguori | (const char *)buf + offset, |
212 | f141eafe | aliguori | aiocb->aio_nbytes - offset, |
213 | f141eafe | aliguori | aiocb->aio_offset + offset); |
214 | f141eafe | aliguori | else
|
215 | f141eafe | aliguori | len = pread(aiocb->aio_fildes, |
216 | f141eafe | aliguori | buf + offset, |
217 | 221f715d | aliguori | aiocb->aio_nbytes - offset, |
218 | 221f715d | aliguori | aiocb->aio_offset + offset); |
219 | 221f715d | aliguori | |
220 | f141eafe | aliguori | if (len == -1 && errno == EINTR) |
221 | f141eafe | aliguori | continue;
|
222 | f141eafe | aliguori | else if (len == -1) { |
223 | f141eafe | aliguori | offset = -errno; |
224 | f141eafe | aliguori | break;
|
225 | f141eafe | aliguori | } else if (len == 0) |
226 | f141eafe | aliguori | break;
|
227 | f141eafe | aliguori | |
228 | f141eafe | aliguori | offset += len; |
229 | 221f715d | aliguori | } |
230 | 221f715d | aliguori | |
231 | 221f715d | aliguori | return offset;
|
232 | 221f715d | aliguori | } |
233 | 221f715d | aliguori | |
234 | 6769da29 | Kevin Wolf | static ssize_t handle_aiocb_rw(struct qemu_paiocb *aiocb) |
235 | 221f715d | aliguori | { |
236 | 6769da29 | Kevin Wolf | ssize_t nbytes; |
237 | f141eafe | aliguori | char *buf;
|
238 | f141eafe | aliguori | |
239 | 9ef91a67 | Christoph Hellwig | if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
|
240 | f141eafe | aliguori | /*
|
241 | f141eafe | aliguori | * If there is just a single buffer, and it is properly aligned
|
242 | f141eafe | aliguori | * we can just use plain pread/pwrite without any problems.
|
243 | f141eafe | aliguori | */
|
244 | ceb42de8 | aliguori | if (aiocb->aio_niov == 1) |
245 | ceb42de8 | aliguori | return handle_aiocb_rw_linear(aiocb, aiocb->aio_iov->iov_base);
|
246 | ceb42de8 | aliguori | |
247 | ceb42de8 | aliguori | /*
|
248 | ceb42de8 | aliguori | * We have more than one iovec, and all are properly aligned.
|
249 | ceb42de8 | aliguori | *
|
250 | ceb42de8 | aliguori | * Try preadv/pwritev first and fall back to linearizing the
|
251 | ceb42de8 | aliguori | * buffer if it's not supported.
|
252 | ceb42de8 | aliguori | */
|
253 | b587a52c | Stefan Hajnoczi | if (preadv_present) {
|
254 | ceb42de8 | aliguori | nbytes = handle_aiocb_rw_vector(aiocb); |
255 | ceb42de8 | aliguori | if (nbytes == aiocb->aio_nbytes)
|
256 | b587a52c | Stefan Hajnoczi | return nbytes;
|
257 | ceb42de8 | aliguori | if (nbytes < 0 && nbytes != -ENOSYS) |
258 | ceb42de8 | aliguori | return nbytes;
|
259 | ceb42de8 | aliguori | preadv_present = 0;
|
260 | ceb42de8 | aliguori | } |
261 | ceb42de8 | aliguori | |
262 | ceb42de8 | aliguori | /*
|
263 | ceb42de8 | aliguori | * XXX(hch): short read/write. no easy way to handle the reminder
|
264 | ceb42de8 | aliguori | * using these interfaces. For now retry using plain
|
265 | ceb42de8 | aliguori | * pread/pwrite?
|
266 | ceb42de8 | aliguori | */
|
267 | f141eafe | aliguori | } |
268 | 221f715d | aliguori | |
269 | f141eafe | aliguori | /*
|
270 | f141eafe | aliguori | * Ok, we have to do it the hard way, copy all segments into
|
271 | f141eafe | aliguori | * a single aligned buffer.
|
272 | f141eafe | aliguori | */
|
273 | 72aef731 | Christoph Hellwig | buf = qemu_blockalign(aiocb->common.bs, aiocb->aio_nbytes); |
274 | 9ef91a67 | Christoph Hellwig | if (aiocb->aio_type & QEMU_AIO_WRITE) {
|
275 | f141eafe | aliguori | char *p = buf;
|
276 | f141eafe | aliguori | int i;
|
277 | f141eafe | aliguori | |
278 | f141eafe | aliguori | for (i = 0; i < aiocb->aio_niov; ++i) { |
279 | f141eafe | aliguori | memcpy(p, aiocb->aio_iov[i].iov_base, aiocb->aio_iov[i].iov_len); |
280 | f141eafe | aliguori | p += aiocb->aio_iov[i].iov_len; |
281 | f141eafe | aliguori | } |
282 | f141eafe | aliguori | } |
283 | f141eafe | aliguori | |
284 | f141eafe | aliguori | nbytes = handle_aiocb_rw_linear(aiocb, buf); |
285 | 9ef91a67 | Christoph Hellwig | if (!(aiocb->aio_type & QEMU_AIO_WRITE)) {
|
286 | f141eafe | aliguori | char *p = buf;
|
287 | f141eafe | aliguori | size_t count = aiocb->aio_nbytes, copy; |
288 | f141eafe | aliguori | int i;
|
289 | f141eafe | aliguori | |
290 | f141eafe | aliguori | for (i = 0; i < aiocb->aio_niov && count; ++i) { |
291 | f141eafe | aliguori | copy = count; |
292 | f141eafe | aliguori | if (copy > aiocb->aio_iov[i].iov_len)
|
293 | f141eafe | aliguori | copy = aiocb->aio_iov[i].iov_len; |
294 | f141eafe | aliguori | memcpy(aiocb->aio_iov[i].iov_base, p, copy); |
295 | f141eafe | aliguori | p += copy; |
296 | f141eafe | aliguori | count -= copy; |
297 | f141eafe | aliguori | } |
298 | f141eafe | aliguori | } |
299 | f141eafe | aliguori | qemu_vfree(buf); |
300 | f141eafe | aliguori | |
301 | f141eafe | aliguori | return nbytes;
|
302 | 221f715d | aliguori | } |
303 | 221f715d | aliguori | |
304 | 3c529d93 | aliguori | static void *aio_thread(void *unused) |
305 | 3c529d93 | aliguori | { |
306 | a8227a5a | malc | pid_t pid; |
307 | 3c529d93 | aliguori | |
308 | a8227a5a | malc | pid = getpid(); |
309 | a8227a5a | malc | |
310 | 3c529d93 | aliguori | while (1) { |
311 | 3c529d93 | aliguori | struct qemu_paiocb *aiocb;
|
312 | 6769da29 | Kevin Wolf | ssize_t ret = 0;
|
313 | 30525aff | malc | qemu_timeval tv; |
314 | 30525aff | malc | struct timespec ts;
|
315 | 30525aff | malc | |
316 | 30525aff | malc | qemu_gettimeofday(&tv); |
317 | 30525aff | malc | ts.tv_sec = tv.tv_sec + 10;
|
318 | 30525aff | malc | ts.tv_nsec = 0;
|
319 | 3c529d93 | aliguori | |
320 | 8653c015 | malc | mutex_lock(&lock); |
321 | 3c529d93 | aliguori | |
322 | 72cf2d4f | Blue Swirl | while (QTAILQ_EMPTY(&request_list) &&
|
323 | 3c529d93 | aliguori | !(ret == ETIMEDOUT)) { |
324 | 8653c015 | malc | ret = cond_timedwait(&cond, &lock, &ts); |
325 | 3c529d93 | aliguori | } |
326 | 3c529d93 | aliguori | |
327 | 72cf2d4f | Blue Swirl | if (QTAILQ_EMPTY(&request_list))
|
328 | 3c529d93 | aliguori | break;
|
329 | 3c529d93 | aliguori | |
330 | 72cf2d4f | Blue Swirl | aiocb = QTAILQ_FIRST(&request_list); |
331 | 72cf2d4f | Blue Swirl | QTAILQ_REMOVE(&request_list, aiocb, node); |
332 | 3c529d93 | aliguori | aiocb->active = 1;
|
333 | 3c529d93 | aliguori | idle_threads--; |
334 | 8653c015 | malc | mutex_unlock(&lock); |
335 | 3c529d93 | aliguori | |
336 | 9ef91a67 | Christoph Hellwig | switch (aiocb->aio_type & QEMU_AIO_TYPE_MASK) {
|
337 | 9ef91a67 | Christoph Hellwig | case QEMU_AIO_READ:
|
338 | 9ef91a67 | Christoph Hellwig | case QEMU_AIO_WRITE:
|
339 | b587a52c | Stefan Hajnoczi | ret = handle_aiocb_rw(aiocb); |
340 | b587a52c | Stefan Hajnoczi | break;
|
341 | b2e12bc6 | Christoph Hellwig | case QEMU_AIO_FLUSH:
|
342 | b587a52c | Stefan Hajnoczi | ret = handle_aiocb_flush(aiocb); |
343 | b587a52c | Stefan Hajnoczi | break;
|
344 | 9ef91a67 | Christoph Hellwig | case QEMU_AIO_IOCTL:
|
345 | b587a52c | Stefan Hajnoczi | ret = handle_aiocb_ioctl(aiocb); |
346 | b587a52c | Stefan Hajnoczi | break;
|
347 | b587a52c | Stefan Hajnoczi | default:
|
348 | b587a52c | Stefan Hajnoczi | fprintf(stderr, "invalid aio request (0x%x)\n", aiocb->aio_type);
|
349 | b587a52c | Stefan Hajnoczi | ret = -EINVAL; |
350 | b587a52c | Stefan Hajnoczi | break;
|
351 | b587a52c | Stefan Hajnoczi | } |
352 | 3c529d93 | aliguori | |
353 | 8653c015 | malc | mutex_lock(&lock); |
354 | 221f715d | aliguori | aiocb->ret = ret; |
355 | 3c529d93 | aliguori | idle_threads++; |
356 | 8653c015 | malc | mutex_unlock(&lock); |
357 | 3c529d93 | aliguori | |
358 | a8227a5a | malc | if (kill(pid, aiocb->ev_signo)) die("kill failed"); |
359 | 3c529d93 | aliguori | } |
360 | 3c529d93 | aliguori | |
361 | 3c529d93 | aliguori | idle_threads--; |
362 | 3c529d93 | aliguori | cur_threads--; |
363 | 8653c015 | malc | mutex_unlock(&lock); |
364 | 3c529d93 | aliguori | |
365 | 3c529d93 | aliguori | return NULL; |
366 | 3c529d93 | aliguori | } |
367 | 3c529d93 | aliguori | |
368 | 8653c015 | malc | static void spawn_thread(void) |
369 | 3c529d93 | aliguori | { |
370 | ee399306 | malc | sigset_t set, oldset; |
371 | ee399306 | malc | |
372 | 3c529d93 | aliguori | cur_threads++; |
373 | 3c529d93 | aliguori | idle_threads++; |
374 | ee399306 | malc | |
375 | ee399306 | malc | /* block all signals */
|
376 | ee399306 | malc | if (sigfillset(&set)) die("sigfillset"); |
377 | ee399306 | malc | if (sigprocmask(SIG_SETMASK, &set, &oldset)) die("sigprocmask"); |
378 | ee399306 | malc | |
379 | 8653c015 | malc | thread_create(&thread_id, &attr, aio_thread, NULL);
|
380 | ee399306 | malc | |
381 | ee399306 | malc | if (sigprocmask(SIG_SETMASK, &oldset, NULL)) die("sigprocmask restore"); |
382 | 3c529d93 | aliguori | } |
383 | 3c529d93 | aliguori | |
384 | 9ef91a67 | Christoph Hellwig | static void qemu_paio_submit(struct qemu_paiocb *aiocb) |
385 | 3c529d93 | aliguori | { |
386 | 3c529d93 | aliguori | aiocb->ret = -EINPROGRESS; |
387 | 3c529d93 | aliguori | aiocb->active = 0;
|
388 | 8653c015 | malc | mutex_lock(&lock); |
389 | 3c529d93 | aliguori | if (idle_threads == 0 && cur_threads < max_threads) |
390 | 3c529d93 | aliguori | spawn_thread(); |
391 | 72cf2d4f | Blue Swirl | QTAILQ_INSERT_TAIL(&request_list, aiocb, node); |
392 | 8653c015 | malc | mutex_unlock(&lock); |
393 | 5d47e372 | malc | cond_signal(&cond); |
394 | 3c529d93 | aliguori | } |
395 | 3c529d93 | aliguori | |
396 | 9ef91a67 | Christoph Hellwig | static ssize_t qemu_paio_return(struct qemu_paiocb *aiocb) |
397 | 3c529d93 | aliguori | { |
398 | 3c529d93 | aliguori | ssize_t ret; |
399 | 3c529d93 | aliguori | |
400 | 8653c015 | malc | mutex_lock(&lock); |
401 | 3c529d93 | aliguori | ret = aiocb->ret; |
402 | 8653c015 | malc | mutex_unlock(&lock); |
403 | 3c529d93 | aliguori | |
404 | 3c529d93 | aliguori | return ret;
|
405 | 3c529d93 | aliguori | } |
406 | 3c529d93 | aliguori | |
407 | 9ef91a67 | Christoph Hellwig | static int qemu_paio_error(struct qemu_paiocb *aiocb) |
408 | 3c529d93 | aliguori | { |
409 | 3c529d93 | aliguori | ssize_t ret = qemu_paio_return(aiocb); |
410 | 3c529d93 | aliguori | |
411 | 3c529d93 | aliguori | if (ret < 0) |
412 | 3c529d93 | aliguori | ret = -ret; |
413 | 3c529d93 | aliguori | else
|
414 | 3c529d93 | aliguori | ret = 0;
|
415 | 3c529d93 | aliguori | |
416 | 3c529d93 | aliguori | return ret;
|
417 | 3c529d93 | aliguori | } |
418 | 3c529d93 | aliguori | |
419 | 59c7b155 | Kevin Wolf | static int posix_aio_process_queue(void *opaque) |
420 | 3c529d93 | aliguori | { |
421 | 9ef91a67 | Christoph Hellwig | PosixAioState *s = opaque; |
422 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *acb, **pacb;
|
423 | 3c529d93 | aliguori | int ret;
|
424 | 59c7b155 | Kevin Wolf | int result = 0; |
425 | e5f37649 | Kevin Wolf | int async_context_id = get_async_context_id();
|
426 | 9ef91a67 | Christoph Hellwig | |
427 | 9ef91a67 | Christoph Hellwig | for(;;) {
|
428 | 9ef91a67 | Christoph Hellwig | pacb = &s->first_aio; |
429 | 9ef91a67 | Christoph Hellwig | for(;;) {
|
430 | 9ef91a67 | Christoph Hellwig | acb = *pacb; |
431 | 9ef91a67 | Christoph Hellwig | if (!acb)
|
432 | 59c7b155 | Kevin Wolf | return result;
|
433 | e5f37649 | Kevin Wolf | |
434 | e5f37649 | Kevin Wolf | /* we're only interested in requests in the right context */
|
435 | e5f37649 | Kevin Wolf | if (acb->async_context_id != async_context_id) {
|
436 | e5f37649 | Kevin Wolf | pacb = &acb->next; |
437 | e5f37649 | Kevin Wolf | continue;
|
438 | e5f37649 | Kevin Wolf | } |
439 | e5f37649 | Kevin Wolf | |
440 | 9ef91a67 | Christoph Hellwig | ret = qemu_paio_error(acb); |
441 | 9ef91a67 | Christoph Hellwig | if (ret == ECANCELED) {
|
442 | 9ef91a67 | Christoph Hellwig | /* remove the request */
|
443 | 9ef91a67 | Christoph Hellwig | *pacb = acb->next; |
444 | 9ef91a67 | Christoph Hellwig | qemu_aio_release(acb); |
445 | 59c7b155 | Kevin Wolf | result = 1;
|
446 | 9ef91a67 | Christoph Hellwig | } else if (ret != EINPROGRESS) { |
447 | 9ef91a67 | Christoph Hellwig | /* end of aio */
|
448 | 9ef91a67 | Christoph Hellwig | if (ret == 0) { |
449 | 9ef91a67 | Christoph Hellwig | ret = qemu_paio_return(acb); |
450 | 9ef91a67 | Christoph Hellwig | if (ret == acb->aio_nbytes)
|
451 | 9ef91a67 | Christoph Hellwig | ret = 0;
|
452 | 9ef91a67 | Christoph Hellwig | else
|
453 | 9ef91a67 | Christoph Hellwig | ret = -EINVAL; |
454 | 9ef91a67 | Christoph Hellwig | } else {
|
455 | 9ef91a67 | Christoph Hellwig | ret = -ret; |
456 | 9ef91a67 | Christoph Hellwig | } |
457 | 9ef91a67 | Christoph Hellwig | /* remove the request */
|
458 | 9ef91a67 | Christoph Hellwig | *pacb = acb->next; |
459 | 9ef91a67 | Christoph Hellwig | /* call the callback */
|
460 | 9ef91a67 | Christoph Hellwig | acb->common.cb(acb->common.opaque, ret); |
461 | 9ef91a67 | Christoph Hellwig | qemu_aio_release(acb); |
462 | 59c7b155 | Kevin Wolf | result = 1;
|
463 | 9ef91a67 | Christoph Hellwig | break;
|
464 | 9ef91a67 | Christoph Hellwig | } else {
|
465 | 9ef91a67 | Christoph Hellwig | pacb = &acb->next; |
466 | 9ef91a67 | Christoph Hellwig | } |
467 | 9ef91a67 | Christoph Hellwig | } |
468 | 9ef91a67 | Christoph Hellwig | } |
469 | 59c7b155 | Kevin Wolf | |
470 | 59c7b155 | Kevin Wolf | return result;
|
471 | 59c7b155 | Kevin Wolf | } |
472 | 59c7b155 | Kevin Wolf | |
473 | 59c7b155 | Kevin Wolf | static void posix_aio_read(void *opaque) |
474 | 59c7b155 | Kevin Wolf | { |
475 | 59c7b155 | Kevin Wolf | PosixAioState *s = opaque; |
476 | 59c7b155 | Kevin Wolf | ssize_t len; |
477 | 59c7b155 | Kevin Wolf | |
478 | 59c7b155 | Kevin Wolf | /* read all bytes from signal pipe */
|
479 | 59c7b155 | Kevin Wolf | for (;;) {
|
480 | 59c7b155 | Kevin Wolf | char bytes[16]; |
481 | 59c7b155 | Kevin Wolf | |
482 | 59c7b155 | Kevin Wolf | len = read(s->rfd, bytes, sizeof(bytes));
|
483 | 59c7b155 | Kevin Wolf | if (len == -1 && errno == EINTR) |
484 | 59c7b155 | Kevin Wolf | continue; /* try again */ |
485 | 59c7b155 | Kevin Wolf | if (len == sizeof(bytes)) |
486 | 59c7b155 | Kevin Wolf | continue; /* more to read */ |
487 | 59c7b155 | Kevin Wolf | break;
|
488 | 59c7b155 | Kevin Wolf | } |
489 | 59c7b155 | Kevin Wolf | |
490 | 59c7b155 | Kevin Wolf | posix_aio_process_queue(s); |
491 | 9ef91a67 | Christoph Hellwig | } |
492 | 9ef91a67 | Christoph Hellwig | |
493 | 9ef91a67 | Christoph Hellwig | static int posix_aio_flush(void *opaque) |
494 | 9ef91a67 | Christoph Hellwig | { |
495 | 9ef91a67 | Christoph Hellwig | PosixAioState *s = opaque; |
496 | 9ef91a67 | Christoph Hellwig | return !!s->first_aio;
|
497 | 9ef91a67 | Christoph Hellwig | } |
498 | 9ef91a67 | Christoph Hellwig | |
499 | 9ef91a67 | Christoph Hellwig | static PosixAioState *posix_aio_state;
|
500 | 9ef91a67 | Christoph Hellwig | |
501 | 9ef91a67 | Christoph Hellwig | static void aio_signal_handler(int signum) |
502 | 9ef91a67 | Christoph Hellwig | { |
503 | 9ef91a67 | Christoph Hellwig | if (posix_aio_state) {
|
504 | 9ef91a67 | Christoph Hellwig | char byte = 0; |
505 | 4817d327 | Kirill A. Shutemov | ssize_t ret; |
506 | 9ef91a67 | Christoph Hellwig | |
507 | 4817d327 | Kirill A. Shutemov | ret = write(posix_aio_state->wfd, &byte, sizeof(byte));
|
508 | 4817d327 | Kirill A. Shutemov | if (ret < 0 && errno != EAGAIN) |
509 | 4817d327 | Kirill A. Shutemov | die("write()");
|
510 | 9ef91a67 | Christoph Hellwig | } |
511 | 9ef91a67 | Christoph Hellwig | |
512 | 9ef91a67 | Christoph Hellwig | qemu_service_io(); |
513 | 9ef91a67 | Christoph Hellwig | } |
514 | 9ef91a67 | Christoph Hellwig | |
515 | 9ef91a67 | Christoph Hellwig | static void paio_remove(struct qemu_paiocb *acb) |
516 | 9ef91a67 | Christoph Hellwig | { |
517 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb **pacb;
|
518 | 9ef91a67 | Christoph Hellwig | |
519 | 9ef91a67 | Christoph Hellwig | /* remove the callback from the queue */
|
520 | 9ef91a67 | Christoph Hellwig | pacb = &posix_aio_state->first_aio; |
521 | 9ef91a67 | Christoph Hellwig | for(;;) {
|
522 | 9ef91a67 | Christoph Hellwig | if (*pacb == NULL) { |
523 | 9ef91a67 | Christoph Hellwig | fprintf(stderr, "paio_remove: aio request not found!\n");
|
524 | 9ef91a67 | Christoph Hellwig | break;
|
525 | 9ef91a67 | Christoph Hellwig | } else if (*pacb == acb) { |
526 | 9ef91a67 | Christoph Hellwig | *pacb = acb->next; |
527 | 9ef91a67 | Christoph Hellwig | qemu_aio_release(acb); |
528 | 9ef91a67 | Christoph Hellwig | break;
|
529 | 9ef91a67 | Christoph Hellwig | } |
530 | 9ef91a67 | Christoph Hellwig | pacb = &(*pacb)->next; |
531 | 9ef91a67 | Christoph Hellwig | } |
532 | 9ef91a67 | Christoph Hellwig | } |
533 | 9ef91a67 | Christoph Hellwig | |
534 | 9ef91a67 | Christoph Hellwig | static void paio_cancel(BlockDriverAIOCB *blockacb) |
535 | 9ef91a67 | Christoph Hellwig | { |
536 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *acb = (struct qemu_paiocb *)blockacb; |
537 | 9ef91a67 | Christoph Hellwig | int active = 0; |
538 | 3c529d93 | aliguori | |
539 | 8653c015 | malc | mutex_lock(&lock); |
540 | 9ef91a67 | Christoph Hellwig | if (!acb->active) {
|
541 | 72cf2d4f | Blue Swirl | QTAILQ_REMOVE(&request_list, acb, node); |
542 | 9ef91a67 | Christoph Hellwig | acb->ret = -ECANCELED; |
543 | 9ef91a67 | Christoph Hellwig | } else if (acb->ret == -EINPROGRESS) { |
544 | 9ef91a67 | Christoph Hellwig | active = 1;
|
545 | 9ef91a67 | Christoph Hellwig | } |
546 | 8653c015 | malc | mutex_unlock(&lock); |
547 | 3c529d93 | aliguori | |
548 | 9ef91a67 | Christoph Hellwig | if (active) {
|
549 | 9ef91a67 | Christoph Hellwig | /* fail safe: if the aio could not be canceled, we wait for
|
550 | 9ef91a67 | Christoph Hellwig | it */
|
551 | 9ef91a67 | Christoph Hellwig | while (qemu_paio_error(acb) == EINPROGRESS)
|
552 | 9ef91a67 | Christoph Hellwig | ; |
553 | 9ef91a67 | Christoph Hellwig | } |
554 | 9ef91a67 | Christoph Hellwig | |
555 | 9ef91a67 | Christoph Hellwig | paio_remove(acb); |
556 | 9ef91a67 | Christoph Hellwig | } |
557 | 9ef91a67 | Christoph Hellwig | |
558 | 9ef91a67 | Christoph Hellwig | static AIOPool raw_aio_pool = {
|
559 | 9ef91a67 | Christoph Hellwig | .aiocb_size = sizeof(struct qemu_paiocb), |
560 | 9ef91a67 | Christoph Hellwig | .cancel = paio_cancel, |
561 | 9ef91a67 | Christoph Hellwig | }; |
562 | 9ef91a67 | Christoph Hellwig | |
563 | 1e5b9d2f | Kevin Wolf | BlockDriverAIOCB *paio_submit(BlockDriverState *bs, int fd,
|
564 | 9ef91a67 | Christoph Hellwig | int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
|
565 | 9ef91a67 | Christoph Hellwig | BlockDriverCompletionFunc *cb, void *opaque, int type) |
566 | 9ef91a67 | Christoph Hellwig | { |
567 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *acb;
|
568 | 9ef91a67 | Christoph Hellwig | |
569 | 9ef91a67 | Christoph Hellwig | acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque); |
570 | 9ef91a67 | Christoph Hellwig | if (!acb)
|
571 | 9ef91a67 | Christoph Hellwig | return NULL; |
572 | 9ef91a67 | Christoph Hellwig | acb->aio_type = type; |
573 | 9ef91a67 | Christoph Hellwig | acb->aio_fildes = fd; |
574 | 9ef91a67 | Christoph Hellwig | acb->ev_signo = SIGUSR2; |
575 | e5f37649 | Kevin Wolf | acb->async_context_id = get_async_context_id(); |
576 | e5f37649 | Kevin Wolf | |
577 | b2e12bc6 | Christoph Hellwig | if (qiov) {
|
578 | b2e12bc6 | Christoph Hellwig | acb->aio_iov = qiov->iov; |
579 | b2e12bc6 | Christoph Hellwig | acb->aio_niov = qiov->niov; |
580 | b2e12bc6 | Christoph Hellwig | } |
581 | 9ef91a67 | Christoph Hellwig | acb->aio_nbytes = nb_sectors * 512;
|
582 | 9ef91a67 | Christoph Hellwig | acb->aio_offset = sector_num * 512;
|
583 | 9ef91a67 | Christoph Hellwig | |
584 | 9ef91a67 | Christoph Hellwig | acb->next = posix_aio_state->first_aio; |
585 | 9ef91a67 | Christoph Hellwig | posix_aio_state->first_aio = acb; |
586 | 9ef91a67 | Christoph Hellwig | |
587 | 6d519a5f | Stefan Hajnoczi | trace_paio_submit(acb, opaque, sector_num, nb_sectors, type); |
588 | 9ef91a67 | Christoph Hellwig | qemu_paio_submit(acb); |
589 | 9ef91a67 | Christoph Hellwig | return &acb->common;
|
590 | 9ef91a67 | Christoph Hellwig | } |
591 | 9ef91a67 | Christoph Hellwig | |
592 | 9ef91a67 | Christoph Hellwig | BlockDriverAIOCB *paio_ioctl(BlockDriverState *bs, int fd,
|
593 | 9ef91a67 | Christoph Hellwig | unsigned long int req, void *buf, |
594 | 9ef91a67 | Christoph Hellwig | BlockDriverCompletionFunc *cb, void *opaque)
|
595 | 9ef91a67 | Christoph Hellwig | { |
596 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *acb;
|
597 | 9ef91a67 | Christoph Hellwig | |
598 | 9ef91a67 | Christoph Hellwig | acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque); |
599 | 9ef91a67 | Christoph Hellwig | if (!acb)
|
600 | 9ef91a67 | Christoph Hellwig | return NULL; |
601 | 9ef91a67 | Christoph Hellwig | acb->aio_type = QEMU_AIO_IOCTL; |
602 | 9ef91a67 | Christoph Hellwig | acb->aio_fildes = fd; |
603 | 9ef91a67 | Christoph Hellwig | acb->ev_signo = SIGUSR2; |
604 | 34cf0081 | Andrew de Quincey | acb->async_context_id = get_async_context_id(); |
605 | 9ef91a67 | Christoph Hellwig | acb->aio_offset = 0;
|
606 | 9ef91a67 | Christoph Hellwig | acb->aio_ioctl_buf = buf; |
607 | 9ef91a67 | Christoph Hellwig | acb->aio_ioctl_cmd = req; |
608 | 9ef91a67 | Christoph Hellwig | |
609 | 9ef91a67 | Christoph Hellwig | acb->next = posix_aio_state->first_aio; |
610 | 9ef91a67 | Christoph Hellwig | posix_aio_state->first_aio = acb; |
611 | 9ef91a67 | Christoph Hellwig | |
612 | 9ef91a67 | Christoph Hellwig | qemu_paio_submit(acb); |
613 | 9ef91a67 | Christoph Hellwig | return &acb->common;
|
614 | 9ef91a67 | Christoph Hellwig | } |
615 | 9ef91a67 | Christoph Hellwig | |
616 | 1e5b9d2f | Kevin Wolf | int paio_init(void) |
617 | 9ef91a67 | Christoph Hellwig | { |
618 | 9ef91a67 | Christoph Hellwig | struct sigaction act;
|
619 | 9ef91a67 | Christoph Hellwig | PosixAioState *s; |
620 | 9ef91a67 | Christoph Hellwig | int fds[2]; |
621 | 9ef91a67 | Christoph Hellwig | int ret;
|
622 | 9ef91a67 | Christoph Hellwig | |
623 | 9ef91a67 | Christoph Hellwig | if (posix_aio_state)
|
624 | 1e5b9d2f | Kevin Wolf | return 0; |
625 | 9ef91a67 | Christoph Hellwig | |
626 | 9ef91a67 | Christoph Hellwig | s = qemu_malloc(sizeof(PosixAioState));
|
627 | 9ef91a67 | Christoph Hellwig | |
628 | 9ef91a67 | Christoph Hellwig | sigfillset(&act.sa_mask); |
629 | 9ef91a67 | Christoph Hellwig | act.sa_flags = 0; /* do not restart syscalls to interrupt select() */ |
630 | 9ef91a67 | Christoph Hellwig | act.sa_handler = aio_signal_handler; |
631 | 9ef91a67 | Christoph Hellwig | sigaction(SIGUSR2, &act, NULL);
|
632 | 9ef91a67 | Christoph Hellwig | |
633 | 9ef91a67 | Christoph Hellwig | s->first_aio = NULL;
|
634 | 40ff6d7e | Kevin Wolf | if (qemu_pipe(fds) == -1) { |
635 | 9ef91a67 | Christoph Hellwig | fprintf(stderr, "failed to create pipe\n");
|
636 | 1e5b9d2f | Kevin Wolf | return -1; |
637 | 9ef91a67 | Christoph Hellwig | } |
638 | 9ef91a67 | Christoph Hellwig | |
639 | 9ef91a67 | Christoph Hellwig | s->rfd = fds[0];
|
640 | 9ef91a67 | Christoph Hellwig | s->wfd = fds[1];
|
641 | 9ef91a67 | Christoph Hellwig | |
642 | 9ef91a67 | Christoph Hellwig | fcntl(s->rfd, F_SETFL, O_NONBLOCK); |
643 | 9ef91a67 | Christoph Hellwig | fcntl(s->wfd, F_SETFL, O_NONBLOCK); |
644 | 9ef91a67 | Christoph Hellwig | |
645 | 8febfa26 | Kevin Wolf | qemu_aio_set_fd_handler(s->rfd, posix_aio_read, NULL, posix_aio_flush,
|
646 | 8febfa26 | Kevin Wolf | posix_aio_process_queue, s); |
647 | 9ef91a67 | Christoph Hellwig | |
648 | 9ef91a67 | Christoph Hellwig | ret = pthread_attr_init(&attr); |
649 | 9ef91a67 | Christoph Hellwig | if (ret)
|
650 | 9ef91a67 | Christoph Hellwig | die2(ret, "pthread_attr_init");
|
651 | 9ef91a67 | Christoph Hellwig | |
652 | 9ef91a67 | Christoph Hellwig | ret = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED); |
653 | 9ef91a67 | Christoph Hellwig | if (ret)
|
654 | 9ef91a67 | Christoph Hellwig | die2(ret, "pthread_attr_setdetachstate");
|
655 | 9ef91a67 | Christoph Hellwig | |
656 | 72cf2d4f | Blue Swirl | QTAILQ_INIT(&request_list); |
657 | 9ef91a67 | Christoph Hellwig | |
658 | 9ef91a67 | Christoph Hellwig | posix_aio_state = s; |
659 | 1e5b9d2f | Kevin Wolf | return 0; |
660 | 3c529d93 | aliguori | } |