Commit | Line | Data |
---|---|---|
2866c82d | 1 | /* |
da751ca9 JA |
2 | * posixaio engine |
3 | * | |
4 | * IO engine that uses the posix defined aio interface. | |
2866c82d JA |
5 | * |
6 | */ | |
7 | #include <stdio.h> | |
8 | #include <stdlib.h> | |
9 | #include <unistd.h> | |
10 | #include <errno.h> | |
bc733f3b | 11 | #include <fcntl.h> |
5f350952 JA |
12 | |
13 | #include "../fio.h" | |
2866c82d JA |
14 | |
15 | struct posixaio_data { | |
16 | struct io_u **aio_events; | |
207cb0f0 | 17 | unsigned int queued; |
2866c82d JA |
18 | }; |
19 | ||
20 | static int fill_timespec(struct timespec *ts) | |
21 | { | |
22 | #ifdef _POSIX_TIMERS | |
23 | if (!clock_gettime(CLOCK_MONOTONIC, ts)) | |
24 | return 0; | |
25 | ||
26 | perror("clock_gettime"); | |
27 | #endif | |
28 | return 1; | |
29 | } | |
30 | ||
31 | static unsigned long long ts_utime_since_now(struct timespec *t) | |
32 | { | |
33 | long long sec, nsec; | |
34 | struct timespec now; | |
35 | ||
36 | if (fill_timespec(&now)) | |
37 | return 0; | |
38 | ||
39 | sec = now.tv_sec - t->tv_sec; | |
40 | nsec = now.tv_nsec - t->tv_nsec; | |
41 | if (sec > 0 && nsec < 0) { | |
42 | sec--; | |
43 | nsec += 1000000000; | |
44 | } | |
45 | ||
46 | sec *= 1000000; | |
47 | nsec /= 1000; | |
48 | return sec + nsec; | |
49 | } | |
50 | ||
7a16dd02 JA |
51 | static int fio_posixaio_cancel(struct thread_data fio_unused *td, |
52 | struct io_u *io_u) | |
2866c82d | 53 | { |
53cdc686 JA |
54 | struct fio_file *f = io_u->file; |
55 | int r = aio_cancel(f->fd, &io_u->aiocb); | |
2866c82d | 56 | |
2faf9ec8 | 57 | if (r == AIO_ALLDONE || r == AIO_CANCELED) |
2866c82d JA |
58 | return 0; |
59 | ||
60 | return 1; | |
61 | } | |
62 | ||
7a16dd02 JA |
63 | static int fio_posixaio_prep(struct thread_data fio_unused *td, |
64 | struct io_u *io_u) | |
2866c82d | 65 | { |
e97c1442 | 66 | os_aiocb_t *aiocb = &io_u->aiocb; |
53cdc686 | 67 | struct fio_file *f = io_u->file; |
2866c82d | 68 | |
53cdc686 | 69 | aiocb->aio_fildes = f->fd; |
cec6b55d JA |
70 | aiocb->aio_buf = io_u->xfer_buf; |
71 | aiocb->aio_nbytes = io_u->xfer_buflen; | |
2866c82d | 72 | aiocb->aio_offset = io_u->offset; |
9918be5a | 73 | aiocb->aio_sigevent.sigev_notify = SIGEV_NONE; |
2866c82d JA |
74 | |
75 | io_u->seen = 0; | |
76 | return 0; | |
77 | } | |
78 | ||
3c77037e JA |
79 | #define SUSPEND_ENTRIES 8 |
80 | ||
e7d2e616 JA |
81 | static int fio_posixaio_getevents(struct thread_data *td, unsigned int min, |
82 | unsigned int max, struct timespec *t) | |
2866c82d JA |
83 | { |
84 | struct posixaio_data *pd = td->io_ops->data; | |
e97c1442 | 85 | os_aiocb_t *suspend_list[SUSPEND_ENTRIES]; |
01743ee1 | 86 | struct flist_head *entry; |
2866c82d | 87 | struct timespec start; |
a3cc770d | 88 | int have_timeout = 0; |
3c77037e | 89 | int suspend_entries = 0; |
a3cc770d | 90 | unsigned int r; |
2866c82d JA |
91 | |
92 | if (t && !fill_timespec(&start)) | |
93 | have_timeout = 1; | |
94 | ||
95 | r = 0; | |
565cc35f | 96 | memset(suspend_list, 0, sizeof(*suspend_list)); |
2866c82d | 97 | restart: |
01743ee1 JA |
98 | flist_for_each(entry, &td->io_u_busylist) { |
99 | struct io_u *io_u = flist_entry(entry, struct io_u, list); | |
2866c82d JA |
100 | int err; |
101 | ||
102 | if (io_u->seen) | |
103 | continue; | |
104 | ||
105 | err = aio_error(&io_u->aiocb); | |
3c77037e JA |
106 | if (err == EINPROGRESS) { |
107 | if (suspend_entries < SUSPEND_ENTRIES) { | |
108 | suspend_list[suspend_entries] = &io_u->aiocb; | |
109 | suspend_entries++; | |
110 | } | |
3f344316 | 111 | continue; |
3c77037e | 112 | } |
3f344316 JA |
113 | |
114 | io_u->seen = 1; | |
207cb0f0 | 115 | pd->queued--; |
3f344316 JA |
116 | pd->aio_events[r++] = io_u; |
117 | ||
118 | if (err == ECANCELED) | |
119 | io_u->resid = io_u->xfer_buflen; | |
120 | else if (!err) { | |
121 | ssize_t retval = aio_return(&io_u->aiocb); | |
122 | ||
123 | io_u->resid = io_u->xfer_buflen - retval; | |
124 | } else | |
125 | io_u->error = err; | |
2866c82d JA |
126 | } |
127 | ||
128 | if (r >= min) | |
129 | return r; | |
130 | ||
131 | if (have_timeout) { | |
132 | unsigned long long usec; | |
133 | ||
134 | usec = (t->tv_sec * 1000000) + (t->tv_nsec / 1000); | |
135 | if (ts_utime_since_now(&start) > usec) | |
136 | return r; | |
137 | } | |
138 | ||
139 | /* | |
3c77037e | 140 | * must have some in-flight, wait for at least one |
2866c82d | 141 | */ |
e97c1442 | 142 | aio_suspend((const os_aiocb_t * const *)suspend_list, |
3c77037e | 143 | suspend_entries, t); |
2866c82d JA |
144 | goto restart; |
145 | } | |
146 | ||
147 | static struct io_u *fio_posixaio_event(struct thread_data *td, int event) | |
148 | { | |
149 | struct posixaio_data *pd = td->io_ops->data; | |
150 | ||
151 | return pd->aio_events[event]; | |
152 | } | |
153 | ||
03e20d68 | 154 | static int fio_posixaio_queue(struct thread_data *td, |
2866c82d JA |
155 | struct io_u *io_u) |
156 | { | |
207cb0f0 | 157 | struct posixaio_data *pd = td->io_ops->data; |
e97c1442 | 158 | os_aiocb_t *aiocb = &io_u->aiocb; |
2866c82d JA |
159 | int ret; |
160 | ||
7101d9c2 JA |
161 | fio_ro_check(td, io_u); |
162 | ||
2866c82d JA |
163 | if (io_u->ddir == DDIR_READ) |
164 | ret = aio_read(aiocb); | |
87dc1ab1 | 165 | else if (io_u->ddir == DDIR_WRITE) |
2866c82d | 166 | ret = aio_write(aiocb); |
a5f3027c JA |
167 | else if (io_u->ddir == DDIR_TRIM) { |
168 | if (pd->queued) | |
169 | return FIO_Q_BUSY; | |
170 | ||
171 | do_io_u_trim(td, io_u); | |
172 | return FIO_Q_COMPLETED; | |
173 | } else { | |
67bf9823 | 174 | #ifdef CONFIG_POSIXAIO_FSYNC |
87dc1ab1 | 175 | ret = aio_fsync(O_SYNC, aiocb); |
207cb0f0 JA |
176 | #else |
177 | if (pd->queued) | |
178 | return FIO_Q_BUSY; | |
179 | ||
f011531e | 180 | do_io_u_sync(td, io_u); |
207cb0f0 JA |
181 | return FIO_Q_COMPLETED; |
182 | #endif | |
183 | } | |
184 | ||
95bcd815 | 185 | if (ret) { |
afa16407 JA |
186 | /* |
187 | * At least OSX has a very low limit on the number of pending | |
def1d8e3 JA |
188 | * IOs, so if it returns EAGAIN, we are out of resources |
189 | * to queue more. Just return FIO_Q_BUSY to naturally | |
190 | * drop off at this depth. | |
afa16407 JA |
191 | */ |
192 | if (errno == EAGAIN) | |
193 | return FIO_Q_BUSY; | |
194 | ||
2866c82d | 195 | io_u->error = errno; |
e1161c32 | 196 | td_verror(td, io_u->error, "xfer"); |
36167d82 | 197 | return FIO_Q_COMPLETED; |
95bcd815 | 198 | } |
36167d82 | 199 | |
207cb0f0 | 200 | pd->queued++; |
36167d82 | 201 | return FIO_Q_QUEUED; |
2866c82d JA |
202 | } |
203 | ||
204 | static void fio_posixaio_cleanup(struct thread_data *td) | |
205 | { | |
206 | struct posixaio_data *pd = td->io_ops->data; | |
207 | ||
208 | if (pd) { | |
209 | free(pd->aio_events); | |
210 | free(pd); | |
2866c82d JA |
211 | } |
212 | } | |
213 | ||
214 | static int fio_posixaio_init(struct thread_data *td) | |
215 | { | |
216 | struct posixaio_data *pd = malloc(sizeof(*pd)); | |
217 | ||
cb781c75 | 218 | memset(pd, 0, sizeof(*pd)); |
2dc1bbeb JA |
219 | pd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u *)); |
220 | memset(pd->aio_events, 0, td->o.iodepth * sizeof(struct io_u *)); | |
2866c82d JA |
221 | |
222 | td->io_ops->data = pd; | |
223 | return 0; | |
224 | } | |
225 | ||
5f350952 | 226 | static struct ioengine_ops ioengine = { |
2866c82d JA |
227 | .name = "posixaio", |
228 | .version = FIO_IOOPS_VERSION, | |
229 | .init = fio_posixaio_init, | |
230 | .prep = fio_posixaio_prep, | |
231 | .queue = fio_posixaio_queue, | |
232 | .cancel = fio_posixaio_cancel, | |
233 | .getevents = fio_posixaio_getevents, | |
234 | .event = fio_posixaio_event, | |
235 | .cleanup = fio_posixaio_cleanup, | |
b5af8293 JA |
236 | .open_file = generic_open_file, |
237 | .close_file = generic_close_file, | |
df9c26b1 | 238 | .get_file_size = generic_get_file_size, |
2866c82d | 239 | }; |
34cfcdaf | 240 | |
5f350952 JA |
241 | static void fio_init fio_posixaio_register(void) |
242 | { | |
243 | register_ioengine(&ioengine); | |
244 | } | |
245 | ||
246 | static void fio_exit fio_posixaio_unregister(void) | |
247 | { | |
248 | unregister_ioengine(&ioengine); | |
249 | } |