root / posix-aio-compat.c @ 67bb172f
History | View | Annotate | Download (15.9 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 | dc786bc9 | Jes Sorensen | #include "sysemu.h" |
28 | f141eafe | aliguori | #include "qemu-common.h" |
29 | 6d519a5f | Stefan Hajnoczi | #include "trace.h" |
30 | 9ef91a67 | Christoph Hellwig | #include "block_int.h" |
31 | 9ef91a67 | Christoph Hellwig | |
32 | 9ef91a67 | Christoph Hellwig | #include "block/raw-posix-aio.h" |
33 | 9ef91a67 | Christoph Hellwig | |
34 | 9ef91a67 | Christoph Hellwig | |
35 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb {
|
36 | 9ef91a67 | Christoph Hellwig | BlockDriverAIOCB common; |
37 | 9ef91a67 | Christoph Hellwig | int aio_fildes;
|
38 | 9ef91a67 | Christoph Hellwig | union {
|
39 | 9ef91a67 | Christoph Hellwig | struct iovec *aio_iov;
|
40 | b587a52c | Stefan Hajnoczi | void *aio_ioctl_buf;
|
41 | 9ef91a67 | Christoph Hellwig | }; |
42 | 9ef91a67 | Christoph Hellwig | int aio_niov;
|
43 | 9ef91a67 | Christoph Hellwig | size_t aio_nbytes; |
44 | 9ef91a67 | Christoph Hellwig | #define aio_ioctl_cmd aio_nbytes /* for QEMU_AIO_IOCTL */ |
45 | 9ef91a67 | Christoph Hellwig | int ev_signo;
|
46 | 9ef91a67 | Christoph Hellwig | off_t aio_offset; |
47 | 9ef91a67 | Christoph Hellwig | |
48 | 72cf2d4f | Blue Swirl | QTAILQ_ENTRY(qemu_paiocb) node; |
49 | 9ef91a67 | Christoph Hellwig | int aio_type;
|
50 | 9ef91a67 | Christoph Hellwig | ssize_t ret; |
51 | 9ef91a67 | Christoph Hellwig | int active;
|
52 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *next;
|
53 | e5f37649 | Kevin Wolf | |
54 | e5f37649 | Kevin Wolf | int async_context_id;
|
55 | 9ef91a67 | Christoph Hellwig | }; |
56 | 9ef91a67 | Christoph Hellwig | |
57 | 9ef91a67 | Christoph Hellwig | typedef struct PosixAioState { |
58 | 9ef91a67 | Christoph Hellwig | int rfd, wfd;
|
59 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *first_aio;
|
60 | 9ef91a67 | Christoph Hellwig | } PosixAioState; |
61 | 3c529d93 | aliguori | |
62 | 3c529d93 | aliguori | |
63 | 3c529d93 | aliguori | static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
|
64 | 3c529d93 | aliguori | static pthread_cond_t cond = PTHREAD_COND_INITIALIZER;
|
65 | 3c529d93 | aliguori | static pthread_t thread_id;
|
66 | a8227a5a | malc | static pthread_attr_t attr;
|
67 | 3c529d93 | aliguori | static int max_threads = 64; |
68 | 3c529d93 | aliguori | static int cur_threads = 0; |
69 | 3c529d93 | aliguori | static int idle_threads = 0; |
70 | 72cf2d4f | Blue Swirl | static QTAILQ_HEAD(, qemu_paiocb) request_list;
|
71 | 3c529d93 | aliguori | |
72 | 2341f9a1 | Juan Quintela | #ifdef CONFIG_PREADV
|
73 | ceb42de8 | aliguori | static int preadv_present = 1; |
74 | ceb42de8 | aliguori | #else
|
75 | ceb42de8 | aliguori | static int preadv_present = 0; |
76 | ceb42de8 | aliguori | #endif
|
77 | ceb42de8 | aliguori | |
78 | 8653c015 | malc | static void die2(int err, const char *what) |
79 | 8653c015 | malc | { |
80 | 8653c015 | malc | fprintf(stderr, "%s failed: %s\n", what, strerror(err));
|
81 | 8653c015 | malc | abort(); |
82 | 8653c015 | malc | } |
83 | 8653c015 | malc | |
84 | 8653c015 | malc | static void die(const char *what) |
85 | 8653c015 | malc | { |
86 | 8653c015 | malc | die2(errno, what); |
87 | 8653c015 | malc | } |
88 | 8653c015 | malc | |
89 | 8653c015 | malc | static void mutex_lock(pthread_mutex_t *mutex) |
90 | 8653c015 | malc | { |
91 | 8653c015 | malc | int ret = pthread_mutex_lock(mutex);
|
92 | 8653c015 | malc | if (ret) die2(ret, "pthread_mutex_lock"); |
93 | 8653c015 | malc | } |
94 | 8653c015 | malc | |
95 | 8653c015 | malc | static void mutex_unlock(pthread_mutex_t *mutex) |
96 | 8653c015 | malc | { |
97 | 8653c015 | malc | int ret = pthread_mutex_unlock(mutex);
|
98 | 8653c015 | malc | if (ret) die2(ret, "pthread_mutex_unlock"); |
99 | 8653c015 | malc | } |
100 | 8653c015 | malc | |
101 | 8653c015 | malc | static int cond_timedwait(pthread_cond_t *cond, pthread_mutex_t *mutex, |
102 | 8653c015 | malc | struct timespec *ts)
|
103 | 8653c015 | malc | { |
104 | 8653c015 | malc | int ret = pthread_cond_timedwait(cond, mutex, ts);
|
105 | 8653c015 | malc | if (ret && ret != ETIMEDOUT) die2(ret, "pthread_cond_timedwait"); |
106 | 8653c015 | malc | return ret;
|
107 | 8653c015 | malc | } |
108 | 8653c015 | malc | |
109 | 5d47e372 | malc | static void cond_signal(pthread_cond_t *cond) |
110 | 8653c015 | malc | { |
111 | 5d47e372 | malc | int ret = pthread_cond_signal(cond);
|
112 | 5d47e372 | malc | if (ret) die2(ret, "pthread_cond_signal"); |
113 | 8653c015 | malc | } |
114 | 8653c015 | malc | |
115 | 8653c015 | malc | static void thread_create(pthread_t *thread, pthread_attr_t *attr, |
116 | 8653c015 | malc | void *(*start_routine)(void*), void *arg) |
117 | 8653c015 | malc | { |
118 | 8653c015 | malc | int ret = pthread_create(thread, attr, start_routine, arg);
|
119 | 8653c015 | malc | if (ret) die2(ret, "pthread_create"); |
120 | 8653c015 | malc | } |
121 | 8653c015 | malc | |
122 | 6769da29 | Kevin Wolf | static ssize_t handle_aiocb_ioctl(struct qemu_paiocb *aiocb) |
123 | f141eafe | aliguori | { |
124 | b587a52c | Stefan Hajnoczi | int ret;
|
125 | b587a52c | Stefan Hajnoczi | |
126 | b587a52c | Stefan Hajnoczi | ret = ioctl(aiocb->aio_fildes, aiocb->aio_ioctl_cmd, aiocb->aio_ioctl_buf); |
127 | b587a52c | Stefan Hajnoczi | if (ret == -1) |
128 | b587a52c | Stefan Hajnoczi | return -errno;
|
129 | b587a52c | Stefan Hajnoczi | |
130 | b587a52c | Stefan Hajnoczi | /*
|
131 | b587a52c | Stefan Hajnoczi | * This looks weird, but the aio code only consideres a request
|
132 | b0cd712c | Stefan Weil | * successful if it has written the number full number of bytes.
|
133 | b587a52c | Stefan Hajnoczi | *
|
134 | b587a52c | Stefan Hajnoczi | * Now we overload aio_nbytes as aio_ioctl_cmd for the ioctl command,
|
135 | b587a52c | Stefan Hajnoczi | * so in fact we return the ioctl command here to make posix_aio_read()
|
136 | b587a52c | Stefan Hajnoczi | * happy..
|
137 | b587a52c | Stefan Hajnoczi | */
|
138 | b587a52c | Stefan Hajnoczi | return aiocb->aio_nbytes;
|
139 | f141eafe | aliguori | } |
140 | f141eafe | aliguori | |
141 | 6769da29 | Kevin Wolf | static ssize_t handle_aiocb_flush(struct qemu_paiocb *aiocb) |
142 | b2e12bc6 | Christoph Hellwig | { |
143 | b2e12bc6 | Christoph Hellwig | int ret;
|
144 | b2e12bc6 | Christoph Hellwig | |
145 | 47faadc6 | Blue Swirl | ret = qemu_fdatasync(aiocb->aio_fildes); |
146 | b2e12bc6 | Christoph Hellwig | if (ret == -1) |
147 | b2e12bc6 | Christoph Hellwig | return -errno;
|
148 | b2e12bc6 | Christoph Hellwig | return 0; |
149 | b2e12bc6 | Christoph Hellwig | } |
150 | b2e12bc6 | Christoph Hellwig | |
151 | 2341f9a1 | Juan Quintela | #ifdef CONFIG_PREADV
|
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 preadv(fd, iov, nr_iov, offset);
|
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 pwritev(fd, iov, nr_iov, offset);
|
163 | ceb42de8 | aliguori | } |
164 | ceb42de8 | aliguori | |
165 | ceb42de8 | aliguori | #else
|
166 | ceb42de8 | aliguori | |
167 | ceb42de8 | aliguori | static ssize_t
|
168 | ceb42de8 | aliguori | qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset) |
169 | ceb42de8 | aliguori | { |
170 | ceb42de8 | aliguori | return -ENOSYS;
|
171 | ceb42de8 | aliguori | } |
172 | ceb42de8 | aliguori | |
173 | ceb42de8 | aliguori | static ssize_t
|
174 | ceb42de8 | aliguori | qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset) |
175 | ceb42de8 | aliguori | { |
176 | ceb42de8 | aliguori | return -ENOSYS;
|
177 | ceb42de8 | aliguori | } |
178 | ceb42de8 | aliguori | |
179 | ceb42de8 | aliguori | #endif
|
180 | ceb42de8 | aliguori | |
181 | 6769da29 | Kevin Wolf | static ssize_t handle_aiocb_rw_vector(struct qemu_paiocb *aiocb) |
182 | ceb42de8 | aliguori | { |
183 | ceb42de8 | aliguori | size_t offset = 0;
|
184 | ceb42de8 | aliguori | ssize_t len; |
185 | ceb42de8 | aliguori | |
186 | ceb42de8 | aliguori | do {
|
187 | 9ef91a67 | Christoph Hellwig | if (aiocb->aio_type & QEMU_AIO_WRITE)
|
188 | ceb42de8 | aliguori | len = qemu_pwritev(aiocb->aio_fildes, |
189 | ceb42de8 | aliguori | aiocb->aio_iov, |
190 | ceb42de8 | aliguori | aiocb->aio_niov, |
191 | ceb42de8 | aliguori | aiocb->aio_offset + offset); |
192 | ceb42de8 | aliguori | else
|
193 | ceb42de8 | aliguori | len = qemu_preadv(aiocb->aio_fildes, |
194 | ceb42de8 | aliguori | aiocb->aio_iov, |
195 | ceb42de8 | aliguori | aiocb->aio_niov, |
196 | ceb42de8 | aliguori | aiocb->aio_offset + offset); |
197 | ceb42de8 | aliguori | } while (len == -1 && errno == EINTR); |
198 | ceb42de8 | aliguori | |
199 | ceb42de8 | aliguori | if (len == -1) |
200 | ceb42de8 | aliguori | return -errno;
|
201 | ceb42de8 | aliguori | return len;
|
202 | ceb42de8 | aliguori | } |
203 | ceb42de8 | aliguori | |
204 | 6769da29 | Kevin Wolf | static ssize_t handle_aiocb_rw_linear(struct qemu_paiocb *aiocb, char *buf) |
205 | 221f715d | aliguori | { |
206 | 6769da29 | Kevin Wolf | ssize_t offset = 0;
|
207 | 6769da29 | Kevin Wolf | ssize_t len; |
208 | 221f715d | aliguori | |
209 | 221f715d | aliguori | while (offset < aiocb->aio_nbytes) {
|
210 | 9ef91a67 | Christoph Hellwig | if (aiocb->aio_type & QEMU_AIO_WRITE)
|
211 | f141eafe | aliguori | len = pwrite(aiocb->aio_fildes, |
212 | f141eafe | aliguori | (const char *)buf + offset, |
213 | f141eafe | aliguori | aiocb->aio_nbytes - offset, |
214 | f141eafe | aliguori | aiocb->aio_offset + offset); |
215 | f141eafe | aliguori | else
|
216 | f141eafe | aliguori | len = pread(aiocb->aio_fildes, |
217 | f141eafe | aliguori | buf + offset, |
218 | 221f715d | aliguori | aiocb->aio_nbytes - offset, |
219 | 221f715d | aliguori | aiocb->aio_offset + offset); |
220 | 221f715d | aliguori | |
221 | f141eafe | aliguori | if (len == -1 && errno == EINTR) |
222 | f141eafe | aliguori | continue;
|
223 | f141eafe | aliguori | else if (len == -1) { |
224 | f141eafe | aliguori | offset = -errno; |
225 | f141eafe | aliguori | break;
|
226 | f141eafe | aliguori | } else if (len == 0) |
227 | f141eafe | aliguori | break;
|
228 | f141eafe | aliguori | |
229 | f141eafe | aliguori | offset += len; |
230 | 221f715d | aliguori | } |
231 | 221f715d | aliguori | |
232 | 221f715d | aliguori | return offset;
|
233 | 221f715d | aliguori | } |
234 | 221f715d | aliguori | |
235 | 6769da29 | Kevin Wolf | static ssize_t handle_aiocb_rw(struct qemu_paiocb *aiocb) |
236 | 221f715d | aliguori | { |
237 | 6769da29 | Kevin Wolf | ssize_t nbytes; |
238 | f141eafe | aliguori | char *buf;
|
239 | f141eafe | aliguori | |
240 | 9ef91a67 | Christoph Hellwig | if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
|
241 | f141eafe | aliguori | /*
|
242 | f141eafe | aliguori | * If there is just a single buffer, and it is properly aligned
|
243 | f141eafe | aliguori | * we can just use plain pread/pwrite without any problems.
|
244 | f141eafe | aliguori | */
|
245 | ceb42de8 | aliguori | if (aiocb->aio_niov == 1) |
246 | ceb42de8 | aliguori | return handle_aiocb_rw_linear(aiocb, aiocb->aio_iov->iov_base);
|
247 | ceb42de8 | aliguori | |
248 | ceb42de8 | aliguori | /*
|
249 | ceb42de8 | aliguori | * We have more than one iovec, and all are properly aligned.
|
250 | ceb42de8 | aliguori | *
|
251 | ceb42de8 | aliguori | * Try preadv/pwritev first and fall back to linearizing the
|
252 | ceb42de8 | aliguori | * buffer if it's not supported.
|
253 | ceb42de8 | aliguori | */
|
254 | b587a52c | Stefan Hajnoczi | if (preadv_present) {
|
255 | ceb42de8 | aliguori | nbytes = handle_aiocb_rw_vector(aiocb); |
256 | ceb42de8 | aliguori | if (nbytes == aiocb->aio_nbytes)
|
257 | b587a52c | Stefan Hajnoczi | return nbytes;
|
258 | ceb42de8 | aliguori | if (nbytes < 0 && nbytes != -ENOSYS) |
259 | ceb42de8 | aliguori | return nbytes;
|
260 | ceb42de8 | aliguori | preadv_present = 0;
|
261 | ceb42de8 | aliguori | } |
262 | ceb42de8 | aliguori | |
263 | ceb42de8 | aliguori | /*
|
264 | ceb42de8 | aliguori | * XXX(hch): short read/write. no easy way to handle the reminder
|
265 | ceb42de8 | aliguori | * using these interfaces. For now retry using plain
|
266 | ceb42de8 | aliguori | * pread/pwrite?
|
267 | ceb42de8 | aliguori | */
|
268 | f141eafe | aliguori | } |
269 | 221f715d | aliguori | |
270 | f141eafe | aliguori | /*
|
271 | f141eafe | aliguori | * Ok, we have to do it the hard way, copy all segments into
|
272 | f141eafe | aliguori | * a single aligned buffer.
|
273 | f141eafe | aliguori | */
|
274 | 72aef731 | Christoph Hellwig | buf = qemu_blockalign(aiocb->common.bs, aiocb->aio_nbytes); |
275 | 9ef91a67 | Christoph Hellwig | if (aiocb->aio_type & QEMU_AIO_WRITE) {
|
276 | f141eafe | aliguori | char *p = buf;
|
277 | f141eafe | aliguori | int i;
|
278 | f141eafe | aliguori | |
279 | f141eafe | aliguori | for (i = 0; i < aiocb->aio_niov; ++i) { |
280 | f141eafe | aliguori | memcpy(p, aiocb->aio_iov[i].iov_base, aiocb->aio_iov[i].iov_len); |
281 | f141eafe | aliguori | p += aiocb->aio_iov[i].iov_len; |
282 | f141eafe | aliguori | } |
283 | f141eafe | aliguori | } |
284 | f141eafe | aliguori | |
285 | f141eafe | aliguori | nbytes = handle_aiocb_rw_linear(aiocb, buf); |
286 | 9ef91a67 | Christoph Hellwig | if (!(aiocb->aio_type & QEMU_AIO_WRITE)) {
|
287 | f141eafe | aliguori | char *p = buf;
|
288 | f141eafe | aliguori | size_t count = aiocb->aio_nbytes, copy; |
289 | f141eafe | aliguori | int i;
|
290 | f141eafe | aliguori | |
291 | f141eafe | aliguori | for (i = 0; i < aiocb->aio_niov && count; ++i) { |
292 | f141eafe | aliguori | copy = count; |
293 | f141eafe | aliguori | if (copy > aiocb->aio_iov[i].iov_len)
|
294 | f141eafe | aliguori | copy = aiocb->aio_iov[i].iov_len; |
295 | f141eafe | aliguori | memcpy(aiocb->aio_iov[i].iov_base, p, copy); |
296 | f141eafe | aliguori | p += copy; |
297 | f141eafe | aliguori | count -= copy; |
298 | f141eafe | aliguori | } |
299 | f141eafe | aliguori | } |
300 | f141eafe | aliguori | qemu_vfree(buf); |
301 | f141eafe | aliguori | |
302 | f141eafe | aliguori | return nbytes;
|
303 | 221f715d | aliguori | } |
304 | 221f715d | aliguori | |
305 | 3c529d93 | aliguori | static void *aio_thread(void *unused) |
306 | 3c529d93 | aliguori | { |
307 | a8227a5a | malc | pid_t pid; |
308 | 3c529d93 | aliguori | |
309 | a8227a5a | malc | pid = getpid(); |
310 | a8227a5a | malc | |
311 | 3c529d93 | aliguori | while (1) { |
312 | 3c529d93 | aliguori | struct qemu_paiocb *aiocb;
|
313 | 6769da29 | Kevin Wolf | ssize_t ret = 0;
|
314 | 30525aff | malc | qemu_timeval tv; |
315 | 30525aff | malc | struct timespec ts;
|
316 | 30525aff | malc | |
317 | 30525aff | malc | qemu_gettimeofday(&tv); |
318 | 30525aff | malc | ts.tv_sec = tv.tv_sec + 10;
|
319 | 30525aff | malc | ts.tv_nsec = 0;
|
320 | 3c529d93 | aliguori | |
321 | 8653c015 | malc | mutex_lock(&lock); |
322 | 3c529d93 | aliguori | |
323 | 72cf2d4f | Blue Swirl | while (QTAILQ_EMPTY(&request_list) &&
|
324 | 3c529d93 | aliguori | !(ret == ETIMEDOUT)) { |
325 | 8653c015 | malc | ret = cond_timedwait(&cond, &lock, &ts); |
326 | 3c529d93 | aliguori | } |
327 | 3c529d93 | aliguori | |
328 | 72cf2d4f | Blue Swirl | if (QTAILQ_EMPTY(&request_list))
|
329 | 3c529d93 | aliguori | break;
|
330 | 3c529d93 | aliguori | |
331 | 72cf2d4f | Blue Swirl | aiocb = QTAILQ_FIRST(&request_list); |
332 | 72cf2d4f | Blue Swirl | QTAILQ_REMOVE(&request_list, aiocb, node); |
333 | 3c529d93 | aliguori | aiocb->active = 1;
|
334 | 3c529d93 | aliguori | idle_threads--; |
335 | 8653c015 | malc | mutex_unlock(&lock); |
336 | 3c529d93 | aliguori | |
337 | 9ef91a67 | Christoph Hellwig | switch (aiocb->aio_type & QEMU_AIO_TYPE_MASK) {
|
338 | 9ef91a67 | Christoph Hellwig | case QEMU_AIO_READ:
|
339 | 9ef91a67 | Christoph Hellwig | case QEMU_AIO_WRITE:
|
340 | b587a52c | Stefan Hajnoczi | ret = handle_aiocb_rw(aiocb); |
341 | b587a52c | Stefan Hajnoczi | break;
|
342 | b2e12bc6 | Christoph Hellwig | case QEMU_AIO_FLUSH:
|
343 | b587a52c | Stefan Hajnoczi | ret = handle_aiocb_flush(aiocb); |
344 | b587a52c | Stefan Hajnoczi | break;
|
345 | 9ef91a67 | Christoph Hellwig | case QEMU_AIO_IOCTL:
|
346 | b587a52c | Stefan Hajnoczi | ret = handle_aiocb_ioctl(aiocb); |
347 | b587a52c | Stefan Hajnoczi | break;
|
348 | b587a52c | Stefan Hajnoczi | default:
|
349 | b587a52c | Stefan Hajnoczi | fprintf(stderr, "invalid aio request (0x%x)\n", aiocb->aio_type);
|
350 | b587a52c | Stefan Hajnoczi | ret = -EINVAL; |
351 | b587a52c | Stefan Hajnoczi | break;
|
352 | b587a52c | Stefan Hajnoczi | } |
353 | 3c529d93 | aliguori | |
354 | 8653c015 | malc | mutex_lock(&lock); |
355 | 221f715d | aliguori | aiocb->ret = ret; |
356 | 3c529d93 | aliguori | idle_threads++; |
357 | 8653c015 | malc | mutex_unlock(&lock); |
358 | 3c529d93 | aliguori | |
359 | a8227a5a | malc | if (kill(pid, aiocb->ev_signo)) die("kill failed"); |
360 | 3c529d93 | aliguori | } |
361 | 3c529d93 | aliguori | |
362 | 3c529d93 | aliguori | idle_threads--; |
363 | 3c529d93 | aliguori | cur_threads--; |
364 | 8653c015 | malc | mutex_unlock(&lock); |
365 | 3c529d93 | aliguori | |
366 | 3c529d93 | aliguori | return NULL; |
367 | 3c529d93 | aliguori | } |
368 | 3c529d93 | aliguori | |
369 | 8653c015 | malc | static void spawn_thread(void) |
370 | 3c529d93 | aliguori | { |
371 | ee399306 | malc | sigset_t set, oldset; |
372 | ee399306 | malc | |
373 | 3c529d93 | aliguori | cur_threads++; |
374 | 3c529d93 | aliguori | idle_threads++; |
375 | ee399306 | malc | |
376 | ee399306 | malc | /* block all signals */
|
377 | ee399306 | malc | if (sigfillset(&set)) die("sigfillset"); |
378 | ee399306 | malc | if (sigprocmask(SIG_SETMASK, &set, &oldset)) die("sigprocmask"); |
379 | ee399306 | malc | |
380 | 8653c015 | malc | thread_create(&thread_id, &attr, aio_thread, NULL);
|
381 | ee399306 | malc | |
382 | ee399306 | malc | if (sigprocmask(SIG_SETMASK, &oldset, NULL)) die("sigprocmask restore"); |
383 | 3c529d93 | aliguori | } |
384 | 3c529d93 | aliguori | |
385 | 9ef91a67 | Christoph Hellwig | static void qemu_paio_submit(struct qemu_paiocb *aiocb) |
386 | 3c529d93 | aliguori | { |
387 | 3c529d93 | aliguori | aiocb->ret = -EINPROGRESS; |
388 | 3c529d93 | aliguori | aiocb->active = 0;
|
389 | 8653c015 | malc | mutex_lock(&lock); |
390 | 3c529d93 | aliguori | if (idle_threads == 0 && cur_threads < max_threads) |
391 | 3c529d93 | aliguori | spawn_thread(); |
392 | 72cf2d4f | Blue Swirl | QTAILQ_INSERT_TAIL(&request_list, aiocb, node); |
393 | 8653c015 | malc | mutex_unlock(&lock); |
394 | 5d47e372 | malc | cond_signal(&cond); |
395 | 3c529d93 | aliguori | } |
396 | 3c529d93 | aliguori | |
397 | 9ef91a67 | Christoph Hellwig | static ssize_t qemu_paio_return(struct qemu_paiocb *aiocb) |
398 | 3c529d93 | aliguori | { |
399 | 3c529d93 | aliguori | ssize_t ret; |
400 | 3c529d93 | aliguori | |
401 | 8653c015 | malc | mutex_lock(&lock); |
402 | 3c529d93 | aliguori | ret = aiocb->ret; |
403 | 8653c015 | malc | mutex_unlock(&lock); |
404 | 3c529d93 | aliguori | |
405 | 3c529d93 | aliguori | return ret;
|
406 | 3c529d93 | aliguori | } |
407 | 3c529d93 | aliguori | |
408 | 9ef91a67 | Christoph Hellwig | static int qemu_paio_error(struct qemu_paiocb *aiocb) |
409 | 3c529d93 | aliguori | { |
410 | 3c529d93 | aliguori | ssize_t ret = qemu_paio_return(aiocb); |
411 | 3c529d93 | aliguori | |
412 | 3c529d93 | aliguori | if (ret < 0) |
413 | 3c529d93 | aliguori | ret = -ret; |
414 | 3c529d93 | aliguori | else
|
415 | 3c529d93 | aliguori | ret = 0;
|
416 | 3c529d93 | aliguori | |
417 | 3c529d93 | aliguori | return ret;
|
418 | 3c529d93 | aliguori | } |
419 | 3c529d93 | aliguori | |
420 | 59c7b155 | Kevin Wolf | static int posix_aio_process_queue(void *opaque) |
421 | 3c529d93 | aliguori | { |
422 | 9ef91a67 | Christoph Hellwig | PosixAioState *s = opaque; |
423 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *acb, **pacb;
|
424 | 3c529d93 | aliguori | int ret;
|
425 | 59c7b155 | Kevin Wolf | int result = 0; |
426 | e5f37649 | Kevin Wolf | int async_context_id = get_async_context_id();
|
427 | 9ef91a67 | Christoph Hellwig | |
428 | 9ef91a67 | Christoph Hellwig | for(;;) {
|
429 | 9ef91a67 | Christoph Hellwig | pacb = &s->first_aio; |
430 | 9ef91a67 | Christoph Hellwig | for(;;) {
|
431 | 9ef91a67 | Christoph Hellwig | acb = *pacb; |
432 | 9ef91a67 | Christoph Hellwig | if (!acb)
|
433 | 59c7b155 | Kevin Wolf | return result;
|
434 | e5f37649 | Kevin Wolf | |
435 | e5f37649 | Kevin Wolf | /* we're only interested in requests in the right context */
|
436 | e5f37649 | Kevin Wolf | if (acb->async_context_id != async_context_id) {
|
437 | e5f37649 | Kevin Wolf | pacb = &acb->next; |
438 | e5f37649 | Kevin Wolf | continue;
|
439 | e5f37649 | Kevin Wolf | } |
440 | e5f37649 | Kevin Wolf | |
441 | 9ef91a67 | Christoph Hellwig | ret = qemu_paio_error(acb); |
442 | 9ef91a67 | Christoph Hellwig | if (ret == ECANCELED) {
|
443 | 9ef91a67 | Christoph Hellwig | /* remove the request */
|
444 | 9ef91a67 | Christoph Hellwig | *pacb = acb->next; |
445 | 9ef91a67 | Christoph Hellwig | qemu_aio_release(acb); |
446 | 59c7b155 | Kevin Wolf | result = 1;
|
447 | 9ef91a67 | Christoph Hellwig | } else if (ret != EINPROGRESS) { |
448 | 9ef91a67 | Christoph Hellwig | /* end of aio */
|
449 | 9ef91a67 | Christoph Hellwig | if (ret == 0) { |
450 | 9ef91a67 | Christoph Hellwig | ret = qemu_paio_return(acb); |
451 | 9ef91a67 | Christoph Hellwig | if (ret == acb->aio_nbytes)
|
452 | 9ef91a67 | Christoph Hellwig | ret = 0;
|
453 | 9ef91a67 | Christoph Hellwig | else
|
454 | 9ef91a67 | Christoph Hellwig | ret = -EINVAL; |
455 | 9ef91a67 | Christoph Hellwig | } else {
|
456 | 9ef91a67 | Christoph Hellwig | ret = -ret; |
457 | 9ef91a67 | Christoph Hellwig | } |
458 | ddca9fb2 | Stefan Hajnoczi | |
459 | ddca9fb2 | Stefan Hajnoczi | trace_paio_complete(acb, acb->common.opaque, ret); |
460 | ddca9fb2 | Stefan Hajnoczi | |
461 | 9ef91a67 | Christoph Hellwig | /* remove the request */
|
462 | 9ef91a67 | Christoph Hellwig | *pacb = acb->next; |
463 | 9ef91a67 | Christoph Hellwig | /* call the callback */
|
464 | 9ef91a67 | Christoph Hellwig | acb->common.cb(acb->common.opaque, ret); |
465 | 9ef91a67 | Christoph Hellwig | qemu_aio_release(acb); |
466 | 59c7b155 | Kevin Wolf | result = 1;
|
467 | 9ef91a67 | Christoph Hellwig | break;
|
468 | 9ef91a67 | Christoph Hellwig | } else {
|
469 | 9ef91a67 | Christoph Hellwig | pacb = &acb->next; |
470 | 9ef91a67 | Christoph Hellwig | } |
471 | 9ef91a67 | Christoph Hellwig | } |
472 | 9ef91a67 | Christoph Hellwig | } |
473 | 59c7b155 | Kevin Wolf | |
474 | 59c7b155 | Kevin Wolf | return result;
|
475 | 59c7b155 | Kevin Wolf | } |
476 | 59c7b155 | Kevin Wolf | |
477 | 59c7b155 | Kevin Wolf | static void posix_aio_read(void *opaque) |
478 | 59c7b155 | Kevin Wolf | { |
479 | 59c7b155 | Kevin Wolf | PosixAioState *s = opaque; |
480 | 59c7b155 | Kevin Wolf | ssize_t len; |
481 | 59c7b155 | Kevin Wolf | |
482 | 59c7b155 | Kevin Wolf | /* read all bytes from signal pipe */
|
483 | 59c7b155 | Kevin Wolf | for (;;) {
|
484 | 59c7b155 | Kevin Wolf | char bytes[16]; |
485 | 59c7b155 | Kevin Wolf | |
486 | 59c7b155 | Kevin Wolf | len = read(s->rfd, bytes, sizeof(bytes));
|
487 | 59c7b155 | Kevin Wolf | if (len == -1 && errno == EINTR) |
488 | 59c7b155 | Kevin Wolf | continue; /* try again */ |
489 | 59c7b155 | Kevin Wolf | if (len == sizeof(bytes)) |
490 | 59c7b155 | Kevin Wolf | continue; /* more to read */ |
491 | 59c7b155 | Kevin Wolf | break;
|
492 | 59c7b155 | Kevin Wolf | } |
493 | 59c7b155 | Kevin Wolf | |
494 | 59c7b155 | Kevin Wolf | posix_aio_process_queue(s); |
495 | 9ef91a67 | Christoph Hellwig | } |
496 | 9ef91a67 | Christoph Hellwig | |
497 | 9ef91a67 | Christoph Hellwig | static int posix_aio_flush(void *opaque) |
498 | 9ef91a67 | Christoph Hellwig | { |
499 | 9ef91a67 | Christoph Hellwig | PosixAioState *s = opaque; |
500 | 9ef91a67 | Christoph Hellwig | return !!s->first_aio;
|
501 | 9ef91a67 | Christoph Hellwig | } |
502 | 9ef91a67 | Christoph Hellwig | |
503 | 9ef91a67 | Christoph Hellwig | static PosixAioState *posix_aio_state;
|
504 | 9ef91a67 | Christoph Hellwig | |
505 | 9ef91a67 | Christoph Hellwig | static void aio_signal_handler(int signum) |
506 | 9ef91a67 | Christoph Hellwig | { |
507 | 9ef91a67 | Christoph Hellwig | if (posix_aio_state) {
|
508 | 9ef91a67 | Christoph Hellwig | char byte = 0; |
509 | 4817d327 | Kirill A. Shutemov | ssize_t ret; |
510 | 9ef91a67 | Christoph Hellwig | |
511 | 4817d327 | Kirill A. Shutemov | ret = write(posix_aio_state->wfd, &byte, sizeof(byte));
|
512 | 4817d327 | Kirill A. Shutemov | if (ret < 0 && errno != EAGAIN) |
513 | 4817d327 | Kirill A. Shutemov | die("write()");
|
514 | 9ef91a67 | Christoph Hellwig | } |
515 | 9ef91a67 | Christoph Hellwig | |
516 | 9ef91a67 | Christoph Hellwig | qemu_service_io(); |
517 | 9ef91a67 | Christoph Hellwig | } |
518 | 9ef91a67 | Christoph Hellwig | |
519 | 9ef91a67 | Christoph Hellwig | static void paio_remove(struct qemu_paiocb *acb) |
520 | 9ef91a67 | Christoph Hellwig | { |
521 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb **pacb;
|
522 | 9ef91a67 | Christoph Hellwig | |
523 | 9ef91a67 | Christoph Hellwig | /* remove the callback from the queue */
|
524 | 9ef91a67 | Christoph Hellwig | pacb = &posix_aio_state->first_aio; |
525 | 9ef91a67 | Christoph Hellwig | for(;;) {
|
526 | 9ef91a67 | Christoph Hellwig | if (*pacb == NULL) { |
527 | 9ef91a67 | Christoph Hellwig | fprintf(stderr, "paio_remove: aio request not found!\n");
|
528 | 9ef91a67 | Christoph Hellwig | break;
|
529 | 9ef91a67 | Christoph Hellwig | } else if (*pacb == acb) { |
530 | 9ef91a67 | Christoph Hellwig | *pacb = acb->next; |
531 | 9ef91a67 | Christoph Hellwig | qemu_aio_release(acb); |
532 | 9ef91a67 | Christoph Hellwig | break;
|
533 | 9ef91a67 | Christoph Hellwig | } |
534 | 9ef91a67 | Christoph Hellwig | pacb = &(*pacb)->next; |
535 | 9ef91a67 | Christoph Hellwig | } |
536 | 9ef91a67 | Christoph Hellwig | } |
537 | 9ef91a67 | Christoph Hellwig | |
538 | 9ef91a67 | Christoph Hellwig | static void paio_cancel(BlockDriverAIOCB *blockacb) |
539 | 9ef91a67 | Christoph Hellwig | { |
540 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *acb = (struct qemu_paiocb *)blockacb; |
541 | 9ef91a67 | Christoph Hellwig | int active = 0; |
542 | 3c529d93 | aliguori | |
543 | ddca9fb2 | Stefan Hajnoczi | trace_paio_cancel(acb, acb->common.opaque); |
544 | ddca9fb2 | Stefan Hajnoczi | |
545 | 8653c015 | malc | mutex_lock(&lock); |
546 | 9ef91a67 | Christoph Hellwig | if (!acb->active) {
|
547 | 72cf2d4f | Blue Swirl | QTAILQ_REMOVE(&request_list, acb, node); |
548 | 9ef91a67 | Christoph Hellwig | acb->ret = -ECANCELED; |
549 | 9ef91a67 | Christoph Hellwig | } else if (acb->ret == -EINPROGRESS) { |
550 | 9ef91a67 | Christoph Hellwig | active = 1;
|
551 | 9ef91a67 | Christoph Hellwig | } |
552 | 8653c015 | malc | mutex_unlock(&lock); |
553 | 3c529d93 | aliguori | |
554 | 9ef91a67 | Christoph Hellwig | if (active) {
|
555 | 9ef91a67 | Christoph Hellwig | /* fail safe: if the aio could not be canceled, we wait for
|
556 | 9ef91a67 | Christoph Hellwig | it */
|
557 | 9ef91a67 | Christoph Hellwig | while (qemu_paio_error(acb) == EINPROGRESS)
|
558 | 9ef91a67 | Christoph Hellwig | ; |
559 | 9ef91a67 | Christoph Hellwig | } |
560 | 9ef91a67 | Christoph Hellwig | |
561 | 9ef91a67 | Christoph Hellwig | paio_remove(acb); |
562 | 9ef91a67 | Christoph Hellwig | } |
563 | 9ef91a67 | Christoph Hellwig | |
564 | 9ef91a67 | Christoph Hellwig | static AIOPool raw_aio_pool = {
|
565 | 9ef91a67 | Christoph Hellwig | .aiocb_size = sizeof(struct qemu_paiocb), |
566 | 9ef91a67 | Christoph Hellwig | .cancel = paio_cancel, |
567 | 9ef91a67 | Christoph Hellwig | }; |
568 | 9ef91a67 | Christoph Hellwig | |
569 | 1e5b9d2f | Kevin Wolf | BlockDriverAIOCB *paio_submit(BlockDriverState *bs, int fd,
|
570 | 9ef91a67 | Christoph Hellwig | int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
|
571 | 9ef91a67 | Christoph Hellwig | BlockDriverCompletionFunc *cb, void *opaque, int type) |
572 | 9ef91a67 | Christoph Hellwig | { |
573 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *acb;
|
574 | 9ef91a67 | Christoph Hellwig | |
575 | 9ef91a67 | Christoph Hellwig | acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque); |
576 | 9ef91a67 | Christoph Hellwig | if (!acb)
|
577 | 9ef91a67 | Christoph Hellwig | return NULL; |
578 | 9ef91a67 | Christoph Hellwig | acb->aio_type = type; |
579 | 9ef91a67 | Christoph Hellwig | acb->aio_fildes = fd; |
580 | 9ef91a67 | Christoph Hellwig | acb->ev_signo = SIGUSR2; |
581 | e5f37649 | Kevin Wolf | acb->async_context_id = get_async_context_id(); |
582 | e5f37649 | Kevin Wolf | |
583 | b2e12bc6 | Christoph Hellwig | if (qiov) {
|
584 | b2e12bc6 | Christoph Hellwig | acb->aio_iov = qiov->iov; |
585 | b2e12bc6 | Christoph Hellwig | acb->aio_niov = qiov->niov; |
586 | b2e12bc6 | Christoph Hellwig | } |
587 | 9ef91a67 | Christoph Hellwig | acb->aio_nbytes = nb_sectors * 512;
|
588 | 9ef91a67 | Christoph Hellwig | acb->aio_offset = sector_num * 512;
|
589 | 9ef91a67 | Christoph Hellwig | |
590 | 9ef91a67 | Christoph Hellwig | acb->next = posix_aio_state->first_aio; |
591 | 9ef91a67 | Christoph Hellwig | posix_aio_state->first_aio = acb; |
592 | 9ef91a67 | Christoph Hellwig | |
593 | 6d519a5f | Stefan Hajnoczi | trace_paio_submit(acb, opaque, sector_num, nb_sectors, type); |
594 | 9ef91a67 | Christoph Hellwig | qemu_paio_submit(acb); |
595 | 9ef91a67 | Christoph Hellwig | return &acb->common;
|
596 | 9ef91a67 | Christoph Hellwig | } |
597 | 9ef91a67 | Christoph Hellwig | |
598 | 9ef91a67 | Christoph Hellwig | BlockDriverAIOCB *paio_ioctl(BlockDriverState *bs, int fd,
|
599 | 9ef91a67 | Christoph Hellwig | unsigned long int req, void *buf, |
600 | 9ef91a67 | Christoph Hellwig | BlockDriverCompletionFunc *cb, void *opaque)
|
601 | 9ef91a67 | Christoph Hellwig | { |
602 | 9ef91a67 | Christoph Hellwig | struct qemu_paiocb *acb;
|
603 | 9ef91a67 | Christoph Hellwig | |
604 | 9ef91a67 | Christoph Hellwig | acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque); |
605 | 9ef91a67 | Christoph Hellwig | if (!acb)
|
606 | 9ef91a67 | Christoph Hellwig | return NULL; |
607 | 9ef91a67 | Christoph Hellwig | acb->aio_type = QEMU_AIO_IOCTL; |
608 | 9ef91a67 | Christoph Hellwig | acb->aio_fildes = fd; |
609 | 9ef91a67 | Christoph Hellwig | acb->ev_signo = SIGUSR2; |
610 | 34cf0081 | Andrew de Quincey | acb->async_context_id = get_async_context_id(); |
611 | 9ef91a67 | Christoph Hellwig | acb->aio_offset = 0;
|
612 | 9ef91a67 | Christoph Hellwig | acb->aio_ioctl_buf = buf; |
613 | 9ef91a67 | Christoph Hellwig | acb->aio_ioctl_cmd = req; |
614 | 9ef91a67 | Christoph Hellwig | |
615 | 9ef91a67 | Christoph Hellwig | acb->next = posix_aio_state->first_aio; |
616 | 9ef91a67 | Christoph Hellwig | posix_aio_state->first_aio = acb; |
617 | 9ef91a67 | Christoph Hellwig | |
618 | 9ef91a67 | Christoph Hellwig | qemu_paio_submit(acb); |
619 | 9ef91a67 | Christoph Hellwig | return &acb->common;
|
620 | 9ef91a67 | Christoph Hellwig | } |
621 | 9ef91a67 | Christoph Hellwig | |
622 | 1e5b9d2f | Kevin Wolf | int paio_init(void) |
623 | 9ef91a67 | Christoph Hellwig | { |
624 | 9ef91a67 | Christoph Hellwig | struct sigaction act;
|
625 | 9ef91a67 | Christoph Hellwig | PosixAioState *s; |
626 | 9ef91a67 | Christoph Hellwig | int fds[2]; |
627 | 9ef91a67 | Christoph Hellwig | int ret;
|
628 | 9ef91a67 | Christoph Hellwig | |
629 | 9ef91a67 | Christoph Hellwig | if (posix_aio_state)
|
630 | 1e5b9d2f | Kevin Wolf | return 0; |
631 | 9ef91a67 | Christoph Hellwig | |
632 | 9ef91a67 | Christoph Hellwig | s = qemu_malloc(sizeof(PosixAioState));
|
633 | 9ef91a67 | Christoph Hellwig | |
634 | 9ef91a67 | Christoph Hellwig | sigfillset(&act.sa_mask); |
635 | 9ef91a67 | Christoph Hellwig | act.sa_flags = 0; /* do not restart syscalls to interrupt select() */ |
636 | 9ef91a67 | Christoph Hellwig | act.sa_handler = aio_signal_handler; |
637 | 9ef91a67 | Christoph Hellwig | sigaction(SIGUSR2, &act, NULL);
|
638 | 9ef91a67 | Christoph Hellwig | |
639 | 9ef91a67 | Christoph Hellwig | s->first_aio = NULL;
|
640 | 40ff6d7e | Kevin Wolf | if (qemu_pipe(fds) == -1) { |
641 | 9ef91a67 | Christoph Hellwig | fprintf(stderr, "failed to create pipe\n");
|
642 | 1e5b9d2f | Kevin Wolf | return -1; |
643 | 9ef91a67 | Christoph Hellwig | } |
644 | 9ef91a67 | Christoph Hellwig | |
645 | 9ef91a67 | Christoph Hellwig | s->rfd = fds[0];
|
646 | 9ef91a67 | Christoph Hellwig | s->wfd = fds[1];
|
647 | 9ef91a67 | Christoph Hellwig | |
648 | 9ef91a67 | Christoph Hellwig | fcntl(s->rfd, F_SETFL, O_NONBLOCK); |
649 | 9ef91a67 | Christoph Hellwig | fcntl(s->wfd, F_SETFL, O_NONBLOCK); |
650 | 9ef91a67 | Christoph Hellwig | |
651 | 8febfa26 | Kevin Wolf | qemu_aio_set_fd_handler(s->rfd, posix_aio_read, NULL, posix_aio_flush,
|
652 | 8febfa26 | Kevin Wolf | posix_aio_process_queue, s); |
653 | 9ef91a67 | Christoph Hellwig | |
654 | 9ef91a67 | Christoph Hellwig | ret = pthread_attr_init(&attr); |
655 | 9ef91a67 | Christoph Hellwig | if (ret)
|
656 | 9ef91a67 | Christoph Hellwig | die2(ret, "pthread_attr_init");
|
657 | 9ef91a67 | Christoph Hellwig | |
658 | 9ef91a67 | Christoph Hellwig | ret = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED); |
659 | 9ef91a67 | Christoph Hellwig | if (ret)
|
660 | 9ef91a67 | Christoph Hellwig | die2(ret, "pthread_attr_setdetachstate");
|
661 | 9ef91a67 | Christoph Hellwig | |
662 | 72cf2d4f | Blue Swirl | QTAILQ_INIT(&request_list); |
663 | 9ef91a67 | Christoph Hellwig | |
664 | 9ef91a67 | Christoph Hellwig | posix_aio_state = s; |
665 | 1e5b9d2f | Kevin Wolf | return 0; |
666 | 3c529d93 | aliguori | } |