| 1 | /* |
| 2 | * posixaio engine |
| 3 | * |
| 4 | * IO engine that uses the posix defined aio interface. |
| 5 | * |
| 6 | */ |
| 7 | #include <stdio.h> |
| 8 | #include <stdlib.h> |
| 9 | #include <unistd.h> |
| 10 | #include <errno.h> |
| 11 | #include <fcntl.h> |
| 12 | |
| 13 | #include "../fio.h" |
| 14 | |
| 15 | #ifdef FIO_HAVE_POSIXAIO |
| 16 | |
| 17 | struct posixaio_data { |
| 18 | struct io_u **aio_events; |
| 19 | unsigned int queued; |
| 20 | }; |
| 21 | |
| 22 | static int fill_timespec(struct timespec *ts) |
| 23 | { |
| 24 | #ifdef _POSIX_TIMERS |
| 25 | if (!clock_gettime(CLOCK_MONOTONIC, ts)) |
| 26 | return 0; |
| 27 | |
| 28 | perror("clock_gettime"); |
| 29 | #endif |
| 30 | return 1; |
| 31 | } |
| 32 | |
| 33 | static unsigned long long ts_utime_since_now(struct timespec *t) |
| 34 | { |
| 35 | long long sec, nsec; |
| 36 | struct timespec now; |
| 37 | |
| 38 | if (fill_timespec(&now)) |
| 39 | return 0; |
| 40 | |
| 41 | sec = now.tv_sec - t->tv_sec; |
| 42 | nsec = now.tv_nsec - t->tv_nsec; |
| 43 | if (sec > 0 && nsec < 0) { |
| 44 | sec--; |
| 45 | nsec += 1000000000; |
| 46 | } |
| 47 | |
| 48 | sec *= 1000000; |
| 49 | nsec /= 1000; |
| 50 | return sec + nsec; |
| 51 | } |
| 52 | |
| 53 | static int fio_posixaio_cancel(struct thread_data fio_unused *td, |
| 54 | struct io_u *io_u) |
| 55 | { |
| 56 | struct fio_file *f = io_u->file; |
| 57 | int r = aio_cancel(f->fd, &io_u->aiocb); |
| 58 | |
| 59 | if (r == 1 || r == AIO_CANCELED) |
| 60 | return 0; |
| 61 | |
| 62 | return 1; |
| 63 | } |
| 64 | |
| 65 | static int fio_posixaio_prep(struct thread_data fio_unused *td, |
| 66 | struct io_u *io_u) |
| 67 | { |
| 68 | struct aiocb *aiocb = &io_u->aiocb; |
| 69 | struct fio_file *f = io_u->file; |
| 70 | |
| 71 | aiocb->aio_fildes = f->fd; |
| 72 | aiocb->aio_buf = io_u->xfer_buf; |
| 73 | aiocb->aio_nbytes = io_u->xfer_buflen; |
| 74 | aiocb->aio_offset = io_u->offset; |
| 75 | |
| 76 | io_u->seen = 0; |
| 77 | return 0; |
| 78 | } |
| 79 | |
| 80 | #define SUSPEND_ENTRIES 8 |
| 81 | |
| 82 | static int fio_posixaio_getevents(struct thread_data *td, unsigned int min, |
| 83 | unsigned int max, struct timespec *t) |
| 84 | { |
| 85 | struct posixaio_data *pd = td->io_ops->data; |
| 86 | struct aiocb *suspend_list[SUSPEND_ENTRIES]; |
| 87 | struct flist_head *entry; |
| 88 | struct timespec start; |
| 89 | int have_timeout = 0; |
| 90 | int suspend_entries = 0; |
| 91 | unsigned int r; |
| 92 | |
| 93 | if (t && !fill_timespec(&start)) |
| 94 | have_timeout = 1; |
| 95 | |
| 96 | r = 0; |
| 97 | memset(suspend_list, 0, sizeof(*suspend_list)); |
| 98 | restart: |
| 99 | flist_for_each(entry, &td->io_u_busylist) { |
| 100 | struct io_u *io_u = flist_entry(entry, struct io_u, list); |
| 101 | int err; |
| 102 | |
| 103 | if (io_u->seen) |
| 104 | continue; |
| 105 | |
| 106 | err = aio_error(&io_u->aiocb); |
| 107 | if (err == EINPROGRESS) { |
| 108 | if (suspend_entries < SUSPEND_ENTRIES) { |
| 109 | suspend_list[suspend_entries] = &io_u->aiocb; |
| 110 | suspend_entries++; |
| 111 | } |
| 112 | continue; |
| 113 | } |
| 114 | |
| 115 | io_u->seen = 1; |
| 116 | pd->queued--; |
| 117 | pd->aio_events[r++] = io_u; |
| 118 | |
| 119 | if (err == ECANCELED) |
| 120 | io_u->resid = io_u->xfer_buflen; |
| 121 | else if (!err) { |
| 122 | ssize_t retval = aio_return(&io_u->aiocb); |
| 123 | |
| 124 | io_u->resid = io_u->xfer_buflen - retval; |
| 125 | } else |
| 126 | io_u->error = err; |
| 127 | |
| 128 | if (r >= max) |
| 129 | break; |
| 130 | } |
| 131 | |
| 132 | if (r >= min) |
| 133 | return r; |
| 134 | |
| 135 | if (have_timeout) { |
| 136 | unsigned long long usec; |
| 137 | |
| 138 | usec = (t->tv_sec * 1000000) + (t->tv_nsec / 1000); |
| 139 | if (ts_utime_since_now(&start) > usec) |
| 140 | return r; |
| 141 | } |
| 142 | |
| 143 | /* |
| 144 | * must have some in-flight, wait for at least one |
| 145 | */ |
| 146 | aio_suspend((const struct aiocb * const *)suspend_list, |
| 147 | suspend_entries, t); |
| 148 | goto restart; |
| 149 | } |
| 150 | |
| 151 | static struct io_u *fio_posixaio_event(struct thread_data *td, int event) |
| 152 | { |
| 153 | struct posixaio_data *pd = td->io_ops->data; |
| 154 | |
| 155 | return pd->aio_events[event]; |
| 156 | } |
| 157 | |
| 158 | static int fio_posixaio_queue(struct thread_data fio_unused *td, |
| 159 | struct io_u *io_u) |
| 160 | { |
| 161 | struct posixaio_data *pd = td->io_ops->data; |
| 162 | struct aiocb *aiocb = &io_u->aiocb; |
| 163 | int ret; |
| 164 | |
| 165 | fio_ro_check(td, io_u); |
| 166 | |
| 167 | if (io_u->ddir == DDIR_READ) |
| 168 | ret = aio_read(aiocb); |
| 169 | else if (io_u->ddir == DDIR_WRITE) |
| 170 | ret = aio_write(aiocb); |
| 171 | else { |
| 172 | #ifdef FIO_HAVE_POSIXAIO_FSYNC |
| 173 | ret = aio_fsync(O_SYNC, aiocb); |
| 174 | #else |
| 175 | if (pd->queued) |
| 176 | return FIO_Q_BUSY; |
| 177 | |
| 178 | if (fsync(io_u->file->fd) < 0) |
| 179 | io_u->error = errno; |
| 180 | |
| 181 | return FIO_Q_COMPLETED; |
| 182 | #endif |
| 183 | } |
| 184 | |
| 185 | if (ret) { |
| 186 | io_u->error = errno; |
| 187 | td_verror(td, io_u->error, "xfer"); |
| 188 | return FIO_Q_COMPLETED; |
| 189 | } |
| 190 | |
| 191 | pd->queued++; |
| 192 | return FIO_Q_QUEUED; |
| 193 | } |
| 194 | |
| 195 | static void fio_posixaio_cleanup(struct thread_data *td) |
| 196 | { |
| 197 | struct posixaio_data *pd = td->io_ops->data; |
| 198 | |
| 199 | if (pd) { |
| 200 | free(pd->aio_events); |
| 201 | free(pd); |
| 202 | } |
| 203 | } |
| 204 | |
| 205 | static int fio_posixaio_init(struct thread_data *td) |
| 206 | { |
| 207 | struct posixaio_data *pd = malloc(sizeof(*pd)); |
| 208 | |
| 209 | memset(pd, 0, sizeof(*pd)); |
| 210 | pd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u *)); |
| 211 | memset(pd->aio_events, 0, td->o.iodepth * sizeof(struct io_u *)); |
| 212 | |
| 213 | td->io_ops->data = pd; |
| 214 | return 0; |
| 215 | } |
| 216 | |
| 217 | static struct ioengine_ops ioengine = { |
| 218 | .name = "posixaio", |
| 219 | .version = FIO_IOOPS_VERSION, |
| 220 | .init = fio_posixaio_init, |
| 221 | .prep = fio_posixaio_prep, |
| 222 | .queue = fio_posixaio_queue, |
| 223 | .cancel = fio_posixaio_cancel, |
| 224 | .getevents = fio_posixaio_getevents, |
| 225 | .event = fio_posixaio_event, |
| 226 | .cleanup = fio_posixaio_cleanup, |
| 227 | .open_file = generic_open_file, |
| 228 | .close_file = generic_close_file, |
| 229 | .get_file_size = generic_get_file_size, |
| 230 | }; |
| 231 | |
| 232 | #else /* FIO_HAVE_POSIXAIO */ |
| 233 | |
| 234 | /* |
| 235 | * When we have a proper configure system in place, we simply wont build |
| 236 | * and install this io engine. For now install a crippled version that |
| 237 | * just complains and fails to load. |
| 238 | */ |
| 239 | static int fio_posixaio_init(struct thread_data fio_unused *td) |
| 240 | { |
| 241 | fprintf(stderr, "fio: posixaio not available\n"); |
| 242 | return 1; |
| 243 | } |
| 244 | |
| 245 | static struct ioengine_ops ioengine = { |
| 246 | .name = "posixaio", |
| 247 | .version = FIO_IOOPS_VERSION, |
| 248 | .init = fio_posixaio_init, |
| 249 | }; |
| 250 | |
| 251 | #endif |
| 252 | |
| 253 | static void fio_init fio_posixaio_register(void) |
| 254 | { |
| 255 | register_ioengine(&ioengine); |
| 256 | } |
| 257 | |
| 258 | static void fio_exit fio_posixaio_unregister(void) |
| 259 | { |
| 260 | unregister_ioengine(&ioengine); |
| 261 | } |