Commit | Line | Data |
---|---|---|
2866c82d | 1 | /* |
da751ca9 JA |
2 | * libaio engine |
3 | * | |
4 | * IO engine using the Linux native aio interface. | |
2866c82d JA |
5 | * |
6 | */ | |
7 | #include <stdio.h> | |
8 | #include <stdlib.h> | |
9 | #include <unistd.h> | |
10 | #include <errno.h> | |
11 | #include <assert.h> | |
5f350952 JA |
12 | |
13 | #include "../fio.h" | |
2866c82d | 14 | |
34cfcdaf JA |
15 | #ifdef FIO_HAVE_LIBAIO |
16 | ||
2866c82d JA |
17 | struct libaio_data { |
18 | io_context_t aio_ctx; | |
19 | struct io_event *aio_events; | |
755200a3 | 20 | struct iocb **iocbs; |
7e77dd02 | 21 | struct io_u **io_us; |
755200a3 | 22 | int iocbs_nr; |
2866c82d JA |
23 | }; |
24 | ||
de890a1e SL |
25 | struct libaio_options { |
26 | struct thread_data *td; | |
27 | unsigned int userspace_reap; | |
28 | }; | |
29 | ||
30 | static struct fio_option options[] = { | |
31 | { | |
32 | .name = "userspace_reap", | |
33 | .type = FIO_OPT_STR_SET, | |
34 | .off1 = offsetof(struct libaio_options, userspace_reap), | |
35 | .help = "Use alternative user-space reap implementation", | |
36 | }, | |
37 | { | |
38 | .name = NULL, | |
39 | }, | |
40 | }; | |
41 | ||
7a16dd02 | 42 | static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u) |
2866c82d | 43 | { |
53cdc686 JA |
44 | struct fio_file *f = io_u->file; |
45 | ||
2866c82d | 46 | if (io_u->ddir == DDIR_READ) |
cec6b55d | 47 | io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
87dc1ab1 | 48 | else if (io_u->ddir == DDIR_WRITE) |
cec6b55d | 49 | io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
5f9099ea | 50 | else if (ddir_sync(io_u->ddir)) |
87dc1ab1 | 51 | io_prep_fsync(&io_u->iocb, f->fd); |
2866c82d JA |
52 | |
53 | return 0; | |
54 | } | |
55 | ||
56 | static struct io_u *fio_libaio_event(struct thread_data *td, int event) | |
57 | { | |
58 | struct libaio_data *ld = td->io_ops->data; | |
f423479d JA |
59 | struct io_event *ev; |
60 | struct io_u *io_u; | |
2866c82d | 61 | |
f423479d | 62 | ev = ld->aio_events + event; |
0cc6ee59 | 63 | io_u = container_of(ev->obj, struct io_u, iocb); |
f423479d JA |
64 | |
65 | if (ev->res != io_u->xfer_buflen) { | |
66 | if (ev->res > io_u->xfer_buflen) | |
67 | io_u->error = -ev->res; | |
68 | else | |
69 | io_u->resid = io_u->xfer_buflen - ev->res; | |
70 | } else | |
71 | io_u->error = 0; | |
72 | ||
73 | return io_u; | |
2866c82d JA |
74 | } |
75 | ||
675012f0 DE |
76 | struct aio_ring { |
77 | unsigned id; /** kernel internal index number */ | |
78 | unsigned nr; /** number of io_events */ | |
79 | unsigned head; | |
80 | unsigned tail; | |
c44b1ff5 | 81 | |
675012f0 DE |
82 | unsigned magic; |
83 | unsigned compat_features; | |
84 | unsigned incompat_features; | |
85 | unsigned header_length; /** size of aio_ring */ | |
86 | ||
87 | struct io_event events[0]; | |
88 | }; | |
89 | ||
90 | #define AIO_RING_MAGIC 0xa10a10a1 | |
91 | ||
92 | static int user_io_getevents(io_context_t aio_ctx, unsigned int max, | |
c44b1ff5 | 93 | struct io_event *events) |
675012f0 DE |
94 | { |
95 | long i = 0; | |
96 | unsigned head; | |
c44b1ff5 | 97 | struct aio_ring *ring = (struct aio_ring*) aio_ctx; |
675012f0 DE |
98 | |
99 | while (i < max) { | |
100 | head = ring->head; | |
101 | ||
102 | if (head == ring->tail) { | |
103 | /* There are no more completions */ | |
104 | break; | |
105 | } else { | |
106 | /* There is another completion to reap */ | |
107 | events[i] = ring->events[head]; | |
108 | read_barrier(); | |
c44b1ff5 | 109 | ring->head = (head + 1) % ring->nr; |
675012f0 DE |
110 | i++; |
111 | } | |
112 | } | |
113 | ||
114 | return i; | |
115 | } | |
116 | ||
e7d2e616 JA |
117 | static int fio_libaio_getevents(struct thread_data *td, unsigned int min, |
118 | unsigned int max, struct timespec *t) | |
2866c82d JA |
119 | { |
120 | struct libaio_data *ld = td->io_ops->data; | |
de890a1e | 121 | struct libaio_options *o = td->eo; |
0b7fdba7 DE |
122 | unsigned actual_min = td->o.iodepth_batch_complete == 0 ? 0 : min; |
123 | int r, events = 0; | |
2866c82d JA |
124 | |
125 | do { | |
de890a1e | 126 | if (o->userspace_reap == 1 |
675012f0 DE |
127 | && actual_min == 0 |
128 | && ((struct aio_ring *)(ld->aio_ctx))->magic | |
129 | == AIO_RING_MAGIC) { | |
130 | r = user_io_getevents(ld->aio_ctx, max, | |
131 | ld->aio_events + events); | |
132 | } else { | |
133 | r = io_getevents(ld->aio_ctx, actual_min, | |
134 | max, ld->aio_events + events, t); | |
135 | } | |
0b7fdba7 DE |
136 | if (r >= 0) |
137 | events += r; | |
138 | else if (r == -EAGAIN) | |
2866c82d | 139 | usleep(100); |
0b7fdba7 DE |
140 | } while (events < min); |
141 | ||
142 | return r < 0 ? r : events; | |
2866c82d JA |
143 | } |
144 | ||
145 | static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u) | |
146 | { | |
147 | struct libaio_data *ld = td->io_ops->data; | |
2866c82d | 148 | |
7101d9c2 JA |
149 | fio_ro_check(td, io_u); |
150 | ||
2dc1bbeb | 151 | if (ld->iocbs_nr == (int) td->o.iodepth) |
755200a3 JA |
152 | return FIO_Q_BUSY; |
153 | ||
154 | /* | |
155 | * fsync is tricky, since it can fail and we need to do it | |
156 | * serialized with other io. the reason is that linux doesn't | |
157 | * support aio fsync yet. So return busy for the case where we | |
158 | * have pending io, to let fio complete those first. | |
159 | */ | |
f011531e | 160 | if (ddir_sync(io_u->ddir)) { |
755200a3 JA |
161 | if (ld->iocbs_nr) |
162 | return FIO_Q_BUSY; | |
5f9099ea | 163 | |
f011531e | 164 | do_io_u_sync(td, io_u); |
755200a3 JA |
165 | return FIO_Q_COMPLETED; |
166 | } | |
167 | ||
a5f3027c JA |
168 | if (io_u->ddir == DDIR_TRIM) { |
169 | if (ld->iocbs_nr) | |
170 | return FIO_Q_BUSY; | |
171 | ||
172 | do_io_u_trim(td, io_u); | |
173 | return FIO_Q_COMPLETED; | |
174 | } | |
175 | ||
755200a3 | 176 | ld->iocbs[ld->iocbs_nr] = &io_u->iocb; |
7e77dd02 | 177 | ld->io_us[ld->iocbs_nr] = io_u; |
755200a3 JA |
178 | ld->iocbs_nr++; |
179 | return FIO_Q_QUEUED; | |
180 | } | |
181 | ||
7e77dd02 JA |
182 | static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us, |
183 | unsigned int nr) | |
184 | { | |
185 | struct timeval now; | |
186 | unsigned int i; | |
187 | ||
12d9d841 JA |
188 | if (!fio_fill_issue_time(td)) |
189 | return; | |
190 | ||
7e77dd02 JA |
191 | fio_gettime(&now, NULL); |
192 | ||
193 | for (i = 0; i < nr; i++) { | |
194 | struct io_u *io_u = io_us[i]; | |
195 | ||
196 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
197 | io_u_queued(td, io_u); | |
198 | } | |
199 | } | |
200 | ||
755200a3 JA |
201 | static int fio_libaio_commit(struct thread_data *td) |
202 | { | |
203 | struct libaio_data *ld = td->io_ops->data; | |
204 | struct iocb **iocbs; | |
7e77dd02 | 205 | struct io_u **io_us; |
5e00c2c4 | 206 | int ret; |
755200a3 JA |
207 | |
208 | if (!ld->iocbs_nr) | |
209 | return 0; | |
210 | ||
7e77dd02 | 211 | io_us = ld->io_us; |
755200a3 | 212 | iocbs = ld->iocbs; |
2866c82d | 213 | do { |
5e00c2c4 JA |
214 | ret = io_submit(ld->aio_ctx, ld->iocbs_nr, iocbs); |
215 | if (ret > 0) { | |
7e77dd02 | 216 | fio_libaio_queued(td, io_us, ret); |
838bc709 | 217 | io_u_mark_submit(td, ret); |
5e00c2c4 | 218 | ld->iocbs_nr -= ret; |
7e77dd02 | 219 | io_us += ret; |
755200a3 | 220 | iocbs += ret; |
5e00c2c4 | 221 | ret = 0; |
838bc709 JA |
222 | } else if (!ret || ret == -EAGAIN || ret == -EINTR) { |
223 | if (!ret) | |
224 | io_u_mark_submit(td, ret); | |
2866c82d | 225 | continue; |
838bc709 | 226 | } else |
2866c82d | 227 | break; |
5e00c2c4 | 228 | } while (ld->iocbs_nr); |
2866c82d | 229 | |
36167d82 | 230 | return ret; |
2866c82d JA |
231 | } |
232 | ||
233 | static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u) | |
234 | { | |
235 | struct libaio_data *ld = td->io_ops->data; | |
236 | ||
237 | return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events); | |
238 | } | |
239 | ||
240 | static void fio_libaio_cleanup(struct thread_data *td) | |
241 | { | |
242 | struct libaio_data *ld = td->io_ops->data; | |
243 | ||
244 | if (ld) { | |
245 | io_destroy(ld->aio_ctx); | |
7e77dd02 JA |
246 | free(ld->aio_events); |
247 | free(ld->iocbs); | |
248 | free(ld->io_us); | |
2866c82d | 249 | free(ld); |
2866c82d JA |
250 | } |
251 | } | |
252 | ||
253 | static int fio_libaio_init(struct thread_data *td) | |
254 | { | |
255 | struct libaio_data *ld = malloc(sizeof(*ld)); | |
c90e1017 JA |
256 | struct libaio_options *o = td->eo; |
257 | int err = 0; | |
2866c82d JA |
258 | |
259 | memset(ld, 0, sizeof(*ld)); | |
c1db2dce | 260 | |
c90e1017 JA |
261 | /* |
262 | * First try passing in 0 for queue depth, since we don't | |
263 | * care about the user ring. If that fails, the kernel is too old | |
264 | * and we need the right depth. | |
265 | */ | |
266 | if (!o->userspace_reap) | |
c224ec05 | 267 | err = io_queue_init(INT_MAX, &ld->aio_ctx); |
c90e1017 JA |
268 | if (o->userspace_reap || err == -EINVAL) |
269 | err = io_queue_init(td->o.iodepth, &ld->aio_ctx); | |
c1db2dce JA |
270 | if (err) { |
271 | td_verror(td, -err, "io_queue_init"); | |
75de55ac | 272 | log_err("fio: check /proc/sys/fs/aio-max-nr\n"); |
cb781c75 | 273 | free(ld); |
2866c82d JA |
274 | return 1; |
275 | } | |
276 | ||
2dc1bbeb JA |
277 | ld->aio_events = malloc(td->o.iodepth * sizeof(struct io_event)); |
278 | memset(ld->aio_events, 0, td->o.iodepth * sizeof(struct io_event)); | |
279 | ld->iocbs = malloc(td->o.iodepth * sizeof(struct iocb *)); | |
755200a3 | 280 | memset(ld->iocbs, 0, sizeof(struct iocb *)); |
2dc1bbeb JA |
281 | ld->io_us = malloc(td->o.iodepth * sizeof(struct io_u *)); |
282 | memset(ld->io_us, 0, td->o.iodepth * sizeof(struct io_u *)); | |
755200a3 JA |
283 | ld->iocbs_nr = 0; |
284 | ||
2866c82d JA |
285 | td->io_ops->data = ld; |
286 | return 0; | |
287 | } | |
288 | ||
5f350952 | 289 | static struct ioengine_ops ioengine = { |
de890a1e SL |
290 | .name = "libaio", |
291 | .version = FIO_IOOPS_VERSION, | |
292 | .init = fio_libaio_init, | |
293 | .prep = fio_libaio_prep, | |
294 | .queue = fio_libaio_queue, | |
295 | .commit = fio_libaio_commit, | |
296 | .cancel = fio_libaio_cancel, | |
297 | .getevents = fio_libaio_getevents, | |
298 | .event = fio_libaio_event, | |
299 | .cleanup = fio_libaio_cleanup, | |
300 | .open_file = generic_open_file, | |
301 | .close_file = generic_close_file, | |
302 | .get_file_size = generic_get_file_size, | |
303 | .options = options, | |
304 | .option_struct_size = sizeof(struct libaio_options), | |
2866c82d | 305 | }; |
34cfcdaf JA |
306 | |
307 | #else /* FIO_HAVE_LIBAIO */ | |
308 | ||
309 | /* | |
310 | * When we have a proper configure system in place, we simply wont build | |
311 | * and install this io engine. For now install a crippled version that | |
312 | * just complains and fails to load. | |
313 | */ | |
314 | static int fio_libaio_init(struct thread_data fio_unused *td) | |
315 | { | |
a3edaf76 | 316 | log_err("fio: libaio not available\n"); |
34cfcdaf JA |
317 | return 1; |
318 | } | |
319 | ||
5f350952 | 320 | static struct ioengine_ops ioengine = { |
34cfcdaf JA |
321 | .name = "libaio", |
322 | .version = FIO_IOOPS_VERSION, | |
323 | .init = fio_libaio_init, | |
324 | }; | |
325 | ||
326 | #endif | |
5f350952 JA |
327 | |
328 | static void fio_init fio_libaio_register(void) | |
329 | { | |
330 | register_ioengine(&ioengine); | |
331 | } | |
332 | ||
333 | static void fio_exit fio_libaio_unregister(void) | |
334 | { | |
335 | unregister_ioengine(&ioengine); | |
336 | } |