| 1 | /* |
| 2 | * libaio engine |
| 3 | * |
| 4 | * IO engine using the Linux native aio interface. |
| 5 | * |
| 6 | */ |
| 7 | #include <stdio.h> |
| 8 | #include <stdlib.h> |
| 9 | #include <unistd.h> |
| 10 | #include <errno.h> |
| 11 | #include <assert.h> |
| 12 | |
| 13 | #include "../fio.h" |
| 14 | |
| 15 | #ifdef FIO_HAVE_LIBAIO |
| 16 | |
| 17 | #define ev_to_iou(ev) (struct io_u *) ((unsigned long) (ev)->obj) |
| 18 | |
| 19 | struct libaio_data { |
| 20 | io_context_t aio_ctx; |
| 21 | struct io_event *aio_events; |
| 22 | struct iocb **iocbs; |
| 23 | struct io_u **io_us; |
| 24 | int iocbs_nr; |
| 25 | }; |
| 26 | |
| 27 | static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u) |
| 28 | { |
| 29 | struct fio_file *f = io_u->file; |
| 30 | |
| 31 | if (io_u->ddir == DDIR_READ) |
| 32 | io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
| 33 | else if (io_u->ddir == DDIR_WRITE) |
| 34 | io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
| 35 | else if (ddir_sync(io_u->ddir)) |
| 36 | io_prep_fsync(&io_u->iocb, f->fd); |
| 37 | else |
| 38 | return 1; |
| 39 | |
| 40 | return 0; |
| 41 | } |
| 42 | |
| 43 | static struct io_u *fio_libaio_event(struct thread_data *td, int event) |
| 44 | { |
| 45 | struct libaio_data *ld = td->io_ops->data; |
| 46 | struct io_event *ev; |
| 47 | struct io_u *io_u; |
| 48 | |
| 49 | ev = ld->aio_events + event; |
| 50 | io_u = ev_to_iou(ev); |
| 51 | |
| 52 | if (ev->res != io_u->xfer_buflen) { |
| 53 | if (ev->res > io_u->xfer_buflen) |
| 54 | io_u->error = -ev->res; |
| 55 | else |
| 56 | io_u->resid = io_u->xfer_buflen - ev->res; |
| 57 | } else |
| 58 | io_u->error = 0; |
| 59 | |
| 60 | return io_u; |
| 61 | } |
| 62 | |
| 63 | static int fio_libaio_getevents(struct thread_data *td, unsigned int min, |
| 64 | unsigned int max, struct timespec *t) |
| 65 | { |
| 66 | struct libaio_data *ld = td->io_ops->data; |
| 67 | int r; |
| 68 | |
| 69 | do { |
| 70 | r = io_getevents(ld->aio_ctx, min, max, ld->aio_events, t); |
| 71 | if (r >= (int) min) |
| 72 | break; |
| 73 | else if (r == -EAGAIN) { |
| 74 | usleep(100); |
| 75 | continue; |
| 76 | } else if (r == -EINTR) |
| 77 | continue; |
| 78 | else if (r != 0) |
| 79 | break; |
| 80 | } while (1); |
| 81 | |
| 82 | return r; |
| 83 | } |
| 84 | |
| 85 | static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u) |
| 86 | { |
| 87 | struct libaio_data *ld = td->io_ops->data; |
| 88 | |
| 89 | fio_ro_check(td, io_u); |
| 90 | |
| 91 | if (ld->iocbs_nr == (int) td->o.iodepth) |
| 92 | return FIO_Q_BUSY; |
| 93 | |
| 94 | /* |
| 95 | * fsync is tricky, since it can fail and we need to do it |
| 96 | * serialized with other io. the reason is that linux doesn't |
| 97 | * support aio fsync yet. So return busy for the case where we |
| 98 | * have pending io, to let fio complete those first. |
| 99 | */ |
| 100 | if (io_u->ddir == DDIR_SYNC) { |
| 101 | if (ld->iocbs_nr) |
| 102 | return FIO_Q_BUSY; |
| 103 | if (fsync(io_u->file->fd) < 0) |
| 104 | io_u->error = errno; |
| 105 | |
| 106 | return FIO_Q_COMPLETED; |
| 107 | } else if (io_u->ddir == DDIR_DATASYNC) { |
| 108 | if (ld->iocbs_nr) |
| 109 | return FIO_Q_BUSY; |
| 110 | if (fdatasync(io_u->file->fd) < 0) |
| 111 | io_u->error = errno; |
| 112 | |
| 113 | return FIO_Q_COMPLETED; |
| 114 | } |
| 115 | |
| 116 | ld->iocbs[ld->iocbs_nr] = &io_u->iocb; |
| 117 | ld->io_us[ld->iocbs_nr] = io_u; |
| 118 | ld->iocbs_nr++; |
| 119 | return FIO_Q_QUEUED; |
| 120 | } |
| 121 | |
| 122 | static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us, |
| 123 | unsigned int nr) |
| 124 | { |
| 125 | struct timeval now; |
| 126 | unsigned int i; |
| 127 | |
| 128 | if (!fio_fill_issue_time(td)) |
| 129 | return; |
| 130 | |
| 131 | fio_gettime(&now, NULL); |
| 132 | |
| 133 | for (i = 0; i < nr; i++) { |
| 134 | struct io_u *io_u = io_us[i]; |
| 135 | |
| 136 | memcpy(&io_u->issue_time, &now, sizeof(now)); |
| 137 | io_u_queued(td, io_u); |
| 138 | } |
| 139 | } |
| 140 | |
| 141 | static int fio_libaio_commit(struct thread_data *td) |
| 142 | { |
| 143 | struct libaio_data *ld = td->io_ops->data; |
| 144 | struct iocb **iocbs; |
| 145 | struct io_u **io_us; |
| 146 | int ret; |
| 147 | |
| 148 | if (!ld->iocbs_nr) |
| 149 | return 0; |
| 150 | |
| 151 | io_us = ld->io_us; |
| 152 | iocbs = ld->iocbs; |
| 153 | do { |
| 154 | ret = io_submit(ld->aio_ctx, ld->iocbs_nr, iocbs); |
| 155 | if (ret > 0) { |
| 156 | fio_libaio_queued(td, io_us, ret); |
| 157 | io_u_mark_submit(td, ret); |
| 158 | ld->iocbs_nr -= ret; |
| 159 | io_us += ret; |
| 160 | iocbs += ret; |
| 161 | ret = 0; |
| 162 | } else if (!ret || ret == -EAGAIN || ret == -EINTR) { |
| 163 | if (!ret) |
| 164 | io_u_mark_submit(td, ret); |
| 165 | continue; |
| 166 | } else |
| 167 | break; |
| 168 | } while (ld->iocbs_nr); |
| 169 | |
| 170 | return ret; |
| 171 | } |
| 172 | |
| 173 | static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u) |
| 174 | { |
| 175 | struct libaio_data *ld = td->io_ops->data; |
| 176 | |
| 177 | return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events); |
| 178 | } |
| 179 | |
| 180 | static void fio_libaio_cleanup(struct thread_data *td) |
| 181 | { |
| 182 | struct libaio_data *ld = td->io_ops->data; |
| 183 | |
| 184 | if (ld) { |
| 185 | io_destroy(ld->aio_ctx); |
| 186 | free(ld->aio_events); |
| 187 | free(ld->iocbs); |
| 188 | free(ld->io_us); |
| 189 | free(ld); |
| 190 | } |
| 191 | } |
| 192 | |
| 193 | static int fio_libaio_init(struct thread_data *td) |
| 194 | { |
| 195 | struct libaio_data *ld = malloc(sizeof(*ld)); |
| 196 | int err; |
| 197 | |
| 198 | memset(ld, 0, sizeof(*ld)); |
| 199 | |
| 200 | err = io_queue_init(td->o.iodepth, &ld->aio_ctx); |
| 201 | if (err) { |
| 202 | td_verror(td, -err, "io_queue_init"); |
| 203 | log_err("fio: check /proc/sys/fs/aio-max-nr\n"); |
| 204 | free(ld); |
| 205 | return 1; |
| 206 | } |
| 207 | |
| 208 | ld->aio_events = malloc(td->o.iodepth * sizeof(struct io_event)); |
| 209 | memset(ld->aio_events, 0, td->o.iodepth * sizeof(struct io_event)); |
| 210 | ld->iocbs = malloc(td->o.iodepth * sizeof(struct iocb *)); |
| 211 | memset(ld->iocbs, 0, sizeof(struct iocb *)); |
| 212 | ld->io_us = malloc(td->o.iodepth * sizeof(struct io_u *)); |
| 213 | memset(ld->io_us, 0, td->o.iodepth * sizeof(struct io_u *)); |
| 214 | ld->iocbs_nr = 0; |
| 215 | |
| 216 | td->io_ops->data = ld; |
| 217 | return 0; |
| 218 | } |
| 219 | |
| 220 | static struct ioengine_ops ioengine = { |
| 221 | .name = "libaio", |
| 222 | .version = FIO_IOOPS_VERSION, |
| 223 | .init = fio_libaio_init, |
| 224 | .prep = fio_libaio_prep, |
| 225 | .queue = fio_libaio_queue, |
| 226 | .commit = fio_libaio_commit, |
| 227 | .cancel = fio_libaio_cancel, |
| 228 | .getevents = fio_libaio_getevents, |
| 229 | .event = fio_libaio_event, |
| 230 | .cleanup = fio_libaio_cleanup, |
| 231 | .open_file = generic_open_file, |
| 232 | .close_file = generic_close_file, |
| 233 | .get_file_size = generic_get_file_size, |
| 234 | }; |
| 235 | |
| 236 | #else /* FIO_HAVE_LIBAIO */ |
| 237 | |
| 238 | /* |
| 239 | * When we have a proper configure system in place, we simply wont build |
| 240 | * and install this io engine. For now install a crippled version that |
| 241 | * just complains and fails to load. |
| 242 | */ |
| 243 | static int fio_libaio_init(struct thread_data fio_unused *td) |
| 244 | { |
| 245 | fprintf(stderr, "fio: libaio not available\n"); |
| 246 | return 1; |
| 247 | } |
| 248 | |
| 249 | static struct ioengine_ops ioengine = { |
| 250 | .name = "libaio", |
| 251 | .version = FIO_IOOPS_VERSION, |
| 252 | .init = fio_libaio_init, |
| 253 | }; |
| 254 | |
| 255 | #endif |
| 256 | |
| 257 | static void fio_init fio_libaio_register(void) |
| 258 | { |
| 259 | register_ioengine(&ioengine); |
| 260 | } |
| 261 | |
| 262 | static void fio_exit fio_libaio_unregister(void) |
| 263 | { |
| 264 | unregister_ioengine(&ioengine); |
| 265 | } |