| 1 | /* |
| 2 | * sync/psync engine |
| 3 | * |
| 4 | * IO engine that does regular read(2)/write(2) with lseek(2) to transfer |
| 5 | * data and IO engine that does regular pread(2)/pwrite(2) to transfer data. |
| 6 | * |
| 7 | */ |
| 8 | #include <stdio.h> |
| 9 | #include <stdlib.h> |
| 10 | #include <unistd.h> |
| 11 | #include <sys/uio.h> |
| 12 | #include <errno.h> |
| 13 | #include <assert.h> |
| 14 | |
| 15 | #include "../fio.h" |
| 16 | |
| 17 | /* |
| 18 | * Sync engine uses engine_data to store last offset |
| 19 | */ |
| 20 | #define LAST_POS(f) ((f)->engine_data) |
| 21 | |
| 22 | struct syncio_data { |
| 23 | struct iovec *iovecs; |
| 24 | struct io_u **io_us; |
| 25 | unsigned int queued; |
| 26 | unsigned int events; |
| 27 | unsigned long queued_bytes; |
| 28 | |
| 29 | unsigned long long last_offset; |
| 30 | struct fio_file *last_file; |
| 31 | enum fio_ddir last_ddir; |
| 32 | }; |
| 33 | |
| 34 | static int fio_syncio_prep(struct thread_data *td, struct io_u *io_u) |
| 35 | { |
| 36 | struct fio_file *f = io_u->file; |
| 37 | |
| 38 | if (!ddir_rw(io_u->ddir)) |
| 39 | return 0; |
| 40 | |
| 41 | if (LAST_POS(f) != -1ULL && LAST_POS(f) == io_u->offset) |
| 42 | return 0; |
| 43 | |
| 44 | if (lseek(f->fd, io_u->offset, SEEK_SET) == -1) { |
| 45 | td_verror(td, errno, "lseek"); |
| 46 | return 1; |
| 47 | } |
| 48 | |
| 49 | return 0; |
| 50 | } |
| 51 | |
| 52 | static int fio_io_end(struct thread_data *td, struct io_u *io_u, int ret) |
| 53 | { |
| 54 | if (io_u->file && ret >= 0 && ddir_rw(io_u->ddir)) |
| 55 | LAST_POS(io_u->file) = io_u->offset + ret; |
| 56 | |
| 57 | if (ret != (int) io_u->xfer_buflen) { |
| 58 | if (ret >= 0) { |
| 59 | io_u->resid = io_u->xfer_buflen - ret; |
| 60 | io_u->error = 0; |
| 61 | return FIO_Q_COMPLETED; |
| 62 | } else |
| 63 | io_u->error = errno; |
| 64 | } |
| 65 | |
| 66 | if (io_u->error) { |
| 67 | io_u_log_error(td, io_u); |
| 68 | td_verror(td, io_u->error, "xfer"); |
| 69 | } |
| 70 | |
| 71 | return FIO_Q_COMPLETED; |
| 72 | } |
| 73 | |
| 74 | #ifdef CONFIG_PWRITEV |
| 75 | static int fio_pvsyncio_queue(struct thread_data *td, struct io_u *io_u) |
| 76 | { |
| 77 | struct syncio_data *sd = td->io_ops->data; |
| 78 | struct iovec *iov = &sd->iovecs[0]; |
| 79 | struct fio_file *f = io_u->file; |
| 80 | int ret; |
| 81 | |
| 82 | fio_ro_check(td, io_u); |
| 83 | |
| 84 | iov->iov_base = io_u->xfer_buf; |
| 85 | iov->iov_len = io_u->xfer_buflen; |
| 86 | |
| 87 | if (io_u->ddir == DDIR_READ) |
| 88 | ret = preadv(f->fd, iov, 1, io_u->offset); |
| 89 | else if (io_u->ddir == DDIR_WRITE) |
| 90 | ret = pwritev(f->fd, iov, 1, io_u->offset); |
| 91 | else if (io_u->ddir == DDIR_TRIM) { |
| 92 | do_io_u_trim(td, io_u); |
| 93 | return FIO_Q_COMPLETED; |
| 94 | } else |
| 95 | ret = do_io_u_sync(td, io_u); |
| 96 | |
| 97 | return fio_io_end(td, io_u, ret); |
| 98 | } |
| 99 | #endif |
| 100 | |
| 101 | static int fio_psyncio_queue(struct thread_data *td, struct io_u *io_u) |
| 102 | { |
| 103 | struct fio_file *f = io_u->file; |
| 104 | int ret; |
| 105 | |
| 106 | fio_ro_check(td, io_u); |
| 107 | |
| 108 | if (io_u->ddir == DDIR_READ) |
| 109 | ret = pread(f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
| 110 | else if (io_u->ddir == DDIR_WRITE) |
| 111 | ret = pwrite(f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
| 112 | else if (io_u->ddir == DDIR_TRIM) { |
| 113 | do_io_u_trim(td, io_u); |
| 114 | return FIO_Q_COMPLETED; |
| 115 | } else |
| 116 | ret = do_io_u_sync(td, io_u); |
| 117 | |
| 118 | return fio_io_end(td, io_u, ret); |
| 119 | } |
| 120 | |
| 121 | static int fio_syncio_queue(struct thread_data *td, struct io_u *io_u) |
| 122 | { |
| 123 | struct fio_file *f = io_u->file; |
| 124 | int ret; |
| 125 | |
| 126 | fio_ro_check(td, io_u); |
| 127 | |
| 128 | if (io_u->ddir == DDIR_READ) |
| 129 | ret = read(f->fd, io_u->xfer_buf, io_u->xfer_buflen); |
| 130 | else if (io_u->ddir == DDIR_WRITE) |
| 131 | ret = write(f->fd, io_u->xfer_buf, io_u->xfer_buflen); |
| 132 | else if (io_u->ddir == DDIR_TRIM) { |
| 133 | do_io_u_trim(td, io_u); |
| 134 | return FIO_Q_COMPLETED; |
| 135 | } else |
| 136 | ret = do_io_u_sync(td, io_u); |
| 137 | |
| 138 | return fio_io_end(td, io_u, ret); |
| 139 | } |
| 140 | |
| 141 | static int fio_vsyncio_getevents(struct thread_data *td, unsigned int min, |
| 142 | unsigned int max, |
| 143 | const struct timespec fio_unused *t) |
| 144 | { |
| 145 | struct syncio_data *sd = td->io_ops->data; |
| 146 | int ret; |
| 147 | |
| 148 | if (min) { |
| 149 | ret = sd->events; |
| 150 | sd->events = 0; |
| 151 | } else |
| 152 | ret = 0; |
| 153 | |
| 154 | dprint(FD_IO, "vsyncio_getevents: min=%d,max=%d: %d\n", min, max, ret); |
| 155 | return ret; |
| 156 | } |
| 157 | |
| 158 | static struct io_u *fio_vsyncio_event(struct thread_data *td, int event) |
| 159 | { |
| 160 | struct syncio_data *sd = td->io_ops->data; |
| 161 | |
| 162 | return sd->io_us[event]; |
| 163 | } |
| 164 | |
| 165 | static int fio_vsyncio_append(struct thread_data *td, struct io_u *io_u) |
| 166 | { |
| 167 | struct syncio_data *sd = td->io_ops->data; |
| 168 | |
| 169 | if (ddir_sync(io_u->ddir)) |
| 170 | return 0; |
| 171 | |
| 172 | if (io_u->offset == sd->last_offset && io_u->file == sd->last_file && |
| 173 | io_u->ddir == sd->last_ddir) |
| 174 | return 1; |
| 175 | |
| 176 | return 0; |
| 177 | } |
| 178 | |
| 179 | static void fio_vsyncio_set_iov(struct syncio_data *sd, struct io_u *io_u, |
| 180 | int idx) |
| 181 | { |
| 182 | sd->io_us[idx] = io_u; |
| 183 | sd->iovecs[idx].iov_base = io_u->xfer_buf; |
| 184 | sd->iovecs[idx].iov_len = io_u->xfer_buflen; |
| 185 | sd->last_offset = io_u->offset + io_u->xfer_buflen; |
| 186 | sd->last_file = io_u->file; |
| 187 | sd->last_ddir = io_u->ddir; |
| 188 | sd->queued_bytes += io_u->xfer_buflen; |
| 189 | sd->queued++; |
| 190 | } |
| 191 | |
| 192 | static int fio_vsyncio_queue(struct thread_data *td, struct io_u *io_u) |
| 193 | { |
| 194 | struct syncio_data *sd = td->io_ops->data; |
| 195 | |
| 196 | fio_ro_check(td, io_u); |
| 197 | |
| 198 | if (!fio_vsyncio_append(td, io_u)) { |
| 199 | dprint(FD_IO, "vsyncio_queue: no append (%d)\n", sd->queued); |
| 200 | /* |
| 201 | * If we can't append and have stuff queued, tell fio to |
| 202 | * commit those first and then retry this io |
| 203 | */ |
| 204 | if (sd->queued) |
| 205 | return FIO_Q_BUSY; |
| 206 | if (ddir_sync(io_u->ddir)) { |
| 207 | int ret = do_io_u_sync(td, io_u); |
| 208 | |
| 209 | return fio_io_end(td, io_u, ret); |
| 210 | } |
| 211 | |
| 212 | sd->queued = 0; |
| 213 | sd->queued_bytes = 0; |
| 214 | fio_vsyncio_set_iov(sd, io_u, 0); |
| 215 | } else { |
| 216 | if (sd->queued == td->o.iodepth) { |
| 217 | dprint(FD_IO, "vsyncio_queue: max depth %d\n", sd->queued); |
| 218 | return FIO_Q_BUSY; |
| 219 | } |
| 220 | |
| 221 | dprint(FD_IO, "vsyncio_queue: append\n"); |
| 222 | fio_vsyncio_set_iov(sd, io_u, sd->queued); |
| 223 | } |
| 224 | |
| 225 | dprint(FD_IO, "vsyncio_queue: depth now %d\n", sd->queued); |
| 226 | return FIO_Q_QUEUED; |
| 227 | } |
| 228 | |
| 229 | /* |
| 230 | * Check that we transferred all bytes, or saw an error, etc |
| 231 | */ |
| 232 | static int fio_vsyncio_end(struct thread_data *td, ssize_t bytes) |
| 233 | { |
| 234 | struct syncio_data *sd = td->io_ops->data; |
| 235 | struct io_u *io_u; |
| 236 | unsigned int i; |
| 237 | int err; |
| 238 | |
| 239 | /* |
| 240 | * transferred everything, perfect |
| 241 | */ |
| 242 | if (bytes == sd->queued_bytes) |
| 243 | return 0; |
| 244 | |
| 245 | err = errno; |
| 246 | for (i = 0; i < sd->queued; i++) { |
| 247 | io_u = sd->io_us[i]; |
| 248 | |
| 249 | if (bytes == -1) { |
| 250 | io_u->error = err; |
| 251 | } else { |
| 252 | unsigned int this_io; |
| 253 | |
| 254 | this_io = bytes; |
| 255 | if (this_io > io_u->xfer_buflen) |
| 256 | this_io = io_u->xfer_buflen; |
| 257 | |
| 258 | io_u->resid = io_u->xfer_buflen - this_io; |
| 259 | io_u->error = 0; |
| 260 | bytes -= this_io; |
| 261 | } |
| 262 | } |
| 263 | |
| 264 | if (bytes == -1) { |
| 265 | td_verror(td, err, "xfer vsync"); |
| 266 | return -err; |
| 267 | } |
| 268 | |
| 269 | return 0; |
| 270 | } |
| 271 | |
| 272 | static int fio_vsyncio_commit(struct thread_data *td) |
| 273 | { |
| 274 | struct syncio_data *sd = td->io_ops->data; |
| 275 | struct fio_file *f; |
| 276 | ssize_t ret; |
| 277 | |
| 278 | if (!sd->queued) |
| 279 | return 0; |
| 280 | |
| 281 | io_u_mark_submit(td, sd->queued); |
| 282 | f = sd->last_file; |
| 283 | |
| 284 | if (lseek(f->fd, sd->io_us[0]->offset, SEEK_SET) == -1) { |
| 285 | int err = -errno; |
| 286 | |
| 287 | td_verror(td, errno, "lseek"); |
| 288 | return err; |
| 289 | } |
| 290 | |
| 291 | if (sd->last_ddir == DDIR_READ) |
| 292 | ret = readv(f->fd, sd->iovecs, sd->queued); |
| 293 | else |
| 294 | ret = writev(f->fd, sd->iovecs, sd->queued); |
| 295 | |
| 296 | dprint(FD_IO, "vsyncio_commit: %d\n", (int) ret); |
| 297 | sd->events = sd->queued; |
| 298 | sd->queued = 0; |
| 299 | return fio_vsyncio_end(td, ret); |
| 300 | } |
| 301 | |
| 302 | static int fio_vsyncio_init(struct thread_data *td) |
| 303 | { |
| 304 | struct syncio_data *sd; |
| 305 | |
| 306 | sd = malloc(sizeof(*sd)); |
| 307 | memset(sd, 0, sizeof(*sd)); |
| 308 | sd->last_offset = -1ULL; |
| 309 | sd->iovecs = malloc(td->o.iodepth * sizeof(struct iovec)); |
| 310 | sd->io_us = malloc(td->o.iodepth * sizeof(struct io_u *)); |
| 311 | |
| 312 | td->io_ops->data = sd; |
| 313 | return 0; |
| 314 | } |
| 315 | |
| 316 | static void fio_vsyncio_cleanup(struct thread_data *td) |
| 317 | { |
| 318 | struct syncio_data *sd = td->io_ops->data; |
| 319 | |
| 320 | free(sd->iovecs); |
| 321 | free(sd->io_us); |
| 322 | free(sd); |
| 323 | } |
| 324 | |
| 325 | static struct ioengine_ops ioengine_rw = { |
| 326 | .name = "sync", |
| 327 | .version = FIO_IOOPS_VERSION, |
| 328 | .prep = fio_syncio_prep, |
| 329 | .queue = fio_syncio_queue, |
| 330 | .open_file = generic_open_file, |
| 331 | .close_file = generic_close_file, |
| 332 | .get_file_size = generic_get_file_size, |
| 333 | .flags = FIO_SYNCIO, |
| 334 | }; |
| 335 | |
| 336 | static struct ioengine_ops ioengine_prw = { |
| 337 | .name = "psync", |
| 338 | .version = FIO_IOOPS_VERSION, |
| 339 | .queue = fio_psyncio_queue, |
| 340 | .open_file = generic_open_file, |
| 341 | .close_file = generic_close_file, |
| 342 | .get_file_size = generic_get_file_size, |
| 343 | .flags = FIO_SYNCIO, |
| 344 | }; |
| 345 | |
| 346 | static struct ioengine_ops ioengine_vrw = { |
| 347 | .name = "vsync", |
| 348 | .version = FIO_IOOPS_VERSION, |
| 349 | .init = fio_vsyncio_init, |
| 350 | .cleanup = fio_vsyncio_cleanup, |
| 351 | .queue = fio_vsyncio_queue, |
| 352 | .commit = fio_vsyncio_commit, |
| 353 | .event = fio_vsyncio_event, |
| 354 | .getevents = fio_vsyncio_getevents, |
| 355 | .open_file = generic_open_file, |
| 356 | .close_file = generic_close_file, |
| 357 | .get_file_size = generic_get_file_size, |
| 358 | .flags = FIO_SYNCIO, |
| 359 | }; |
| 360 | |
| 361 | #ifdef CONFIG_PWRITEV |
| 362 | static struct ioengine_ops ioengine_pvrw = { |
| 363 | .name = "pvsync", |
| 364 | .version = FIO_IOOPS_VERSION, |
| 365 | .init = fio_vsyncio_init, |
| 366 | .cleanup = fio_vsyncio_cleanup, |
| 367 | .queue = fio_pvsyncio_queue, |
| 368 | .open_file = generic_open_file, |
| 369 | .close_file = generic_close_file, |
| 370 | .get_file_size = generic_get_file_size, |
| 371 | .flags = FIO_SYNCIO, |
| 372 | }; |
| 373 | #endif |
| 374 | |
| 375 | static void fio_init fio_syncio_register(void) |
| 376 | { |
| 377 | register_ioengine(&ioengine_rw); |
| 378 | register_ioengine(&ioengine_prw); |
| 379 | register_ioengine(&ioengine_vrw); |
| 380 | #ifdef CONFIG_PWRITEV |
| 381 | register_ioengine(&ioengine_pvrw); |
| 382 | #endif |
| 383 | } |
| 384 | |
| 385 | static void fio_exit fio_syncio_unregister(void) |
| 386 | { |
| 387 | unregister_ioengine(&ioengine_rw); |
| 388 | unregister_ioengine(&ioengine_prw); |
| 389 | unregister_ioengine(&ioengine_vrw); |
| 390 | #ifdef CONFIG_PWRITEV |
| 391 | unregister_ioengine(&ioengine_pvrw); |
| 392 | #endif |
| 393 | } |