| 1 | /* |
| 2 | * windowsaio engine |
| 3 | * |
| 4 | * IO engine using Windows IO Completion Ports. |
| 5 | */ |
| 6 | |
| 7 | #include <stdio.h> |
| 8 | #include <stdlib.h> |
| 9 | #include <unistd.h> |
| 10 | #include <signal.h> |
| 11 | #include <errno.h> |
| 12 | |
| 13 | #include "../fio.h" |
| 14 | |
| 15 | typedef BOOL (WINAPI *CANCELIOEX)(HANDLE hFile, LPOVERLAPPED lpOverlapped); |
| 16 | |
| 17 | int geterrno_from_win_error (DWORD code, int deferrno); |
| 18 | |
| 19 | struct fio_overlapped { |
| 20 | OVERLAPPED o; |
| 21 | struct io_u *io_u; |
| 22 | BOOL io_complete; |
| 23 | }; |
| 24 | |
| 25 | struct windowsaio_data { |
| 26 | struct io_u **aio_events; |
| 27 | HANDLE iocp; |
| 28 | HANDLE iothread; |
| 29 | HANDLE iocomplete_event; |
| 30 | BOOL iothread_running; |
| 31 | }; |
| 32 | |
| 33 | struct thread_ctx { |
| 34 | HANDLE iocp; |
| 35 | struct windowsaio_data *wd; |
| 36 | }; |
| 37 | |
| 38 | static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter); |
| 39 | |
| 40 | static int fio_windowsaio_init(struct thread_data *td) |
| 41 | { |
| 42 | struct windowsaio_data *wd; |
| 43 | int rc = 0; |
| 44 | |
| 45 | wd = calloc(1, sizeof(struct windowsaio_data)); |
| 46 | if (wd == NULL) { |
| 47 | log_err("windowsaio: failed to allocate memory for engine data\n"); |
| 48 | rc = 1; |
| 49 | } |
| 50 | |
| 51 | if (!rc) { |
| 52 | wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*)); |
| 53 | if (wd->aio_events == NULL) { |
| 54 | log_err("windowsaio: failed to allocate memory for aio events list\n"); |
| 55 | rc = 1; |
| 56 | } |
| 57 | } |
| 58 | |
| 59 | if (!rc) { |
| 60 | /* Create an auto-reset event */ |
| 61 | wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL); |
| 62 | if (wd->iocomplete_event == NULL) { |
| 63 | log_err("windowsaio: failed to create io complete event handle\n"); |
| 64 | rc = 1; |
| 65 | } |
| 66 | } |
| 67 | |
| 68 | if (rc) { |
| 69 | if (wd != NULL) { |
| 70 | if (wd->aio_events != NULL) |
| 71 | free(wd->aio_events); |
| 72 | |
| 73 | free(wd); |
| 74 | } |
| 75 | } |
| 76 | |
| 77 | td->io_ops_data = wd; |
| 78 | |
| 79 | if (!rc) { |
| 80 | struct thread_ctx *ctx; |
| 81 | struct windowsaio_data *wd; |
| 82 | HANDLE hFile; |
| 83 | |
| 84 | hFile = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 0); |
| 85 | if (hFile == INVALID_HANDLE_VALUE) { |
| 86 | log_err("windowsaio: failed to create io completion port\n"); |
| 87 | rc = 1; |
| 88 | } |
| 89 | |
| 90 | wd = td->io_ops_data; |
| 91 | wd->iothread_running = TRUE; |
| 92 | wd->iocp = hFile; |
| 93 | |
| 94 | if (!rc) |
| 95 | ctx = malloc(sizeof(struct thread_ctx)); |
| 96 | |
| 97 | if (!rc && ctx == NULL) { |
| 98 | log_err("windowsaio: failed to allocate memory for thread context structure\n"); |
| 99 | CloseHandle(hFile); |
| 100 | rc = 1; |
| 101 | } |
| 102 | |
| 103 | if (!rc) { |
| 104 | DWORD threadid; |
| 105 | |
| 106 | ctx->iocp = hFile; |
| 107 | ctx->wd = wd; |
| 108 | wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, &threadid); |
| 109 | |
| 110 | if (wd->iothread != NULL) |
| 111 | fio_setaffinity(threadid, td->o.cpumask); |
| 112 | else |
| 113 | log_err("windowsaio: failed to create io completion thread\n"); |
| 114 | } |
| 115 | |
| 116 | if (rc || wd->iothread == NULL) |
| 117 | rc = 1; |
| 118 | } |
| 119 | |
| 120 | return rc; |
| 121 | } |
| 122 | |
| 123 | static void fio_windowsaio_cleanup(struct thread_data *td) |
| 124 | { |
| 125 | struct windowsaio_data *wd; |
| 126 | |
| 127 | wd = td->io_ops_data; |
| 128 | |
| 129 | if (wd != NULL) { |
| 130 | wd->iothread_running = FALSE; |
| 131 | WaitForSingleObject(wd->iothread, INFINITE); |
| 132 | |
| 133 | CloseHandle(wd->iothread); |
| 134 | CloseHandle(wd->iocomplete_event); |
| 135 | |
| 136 | free(wd->aio_events); |
| 137 | free(wd); |
| 138 | |
| 139 | td->io_ops_data = NULL; |
| 140 | } |
| 141 | } |
| 142 | |
| 143 | static int windowsaio_invalidate_cache(struct fio_file *f) |
| 144 | { |
| 145 | DWORD error; |
| 146 | DWORD isharemode = (FILE_SHARE_DELETE | FILE_SHARE_READ | |
| 147 | FILE_SHARE_WRITE); |
| 148 | HANDLE ihFile; |
| 149 | int rc = 0; |
| 150 | |
| 151 | /* |
| 152 | * Encourage Windows to drop cached parts of a file by temporarily |
| 153 | * opening it for non-buffered access. Note: this will only work when |
| 154 | * the following is the only thing with the file open on the whole |
| 155 | * system. |
| 156 | */ |
| 157 | dprint(FD_IO, "windowaio: attempt invalidate cache for %s\n", |
| 158 | f->file_name); |
| 159 | ihFile = CreateFile(f->file_name, 0, isharemode, NULL, OPEN_EXISTING, |
| 160 | FILE_FLAG_NO_BUFFERING, NULL); |
| 161 | |
| 162 | if (ihFile != INVALID_HANDLE_VALUE) { |
| 163 | if (!CloseHandle(ihFile)) { |
| 164 | error = GetLastError(); |
| 165 | log_info("windowsaio: invalidation fd close %s " |
| 166 | "failed: error %d\n", f->file_name, error); |
| 167 | rc = 1; |
| 168 | } |
| 169 | } else { |
| 170 | error = GetLastError(); |
| 171 | if (error != ERROR_FILE_NOT_FOUND) { |
| 172 | log_info("windowsaio: cache invalidation of %s failed: " |
| 173 | "error %d\n", f->file_name, error); |
| 174 | rc = 1; |
| 175 | } |
| 176 | } |
| 177 | |
| 178 | return rc; |
| 179 | } |
| 180 | |
| 181 | static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f) |
| 182 | { |
| 183 | int rc = 0; |
| 184 | DWORD flags = FILE_FLAG_POSIX_SEMANTICS | FILE_FLAG_OVERLAPPED; |
| 185 | DWORD sharemode = FILE_SHARE_READ | FILE_SHARE_WRITE; |
| 186 | DWORD openmode = OPEN_ALWAYS; |
| 187 | DWORD access; |
| 188 | |
| 189 | dprint(FD_FILE, "fd open %s\n", f->file_name); |
| 190 | |
| 191 | if (f->filetype == FIO_TYPE_PIPE) { |
| 192 | log_err("windowsaio: pipes are not supported\n"); |
| 193 | return 1; |
| 194 | } |
| 195 | |
| 196 | if (!strcmp(f->file_name, "-")) { |
| 197 | log_err("windowsaio: can't read/write to stdin/out\n"); |
| 198 | return 1; |
| 199 | } |
| 200 | |
| 201 | if (td->o.odirect) |
| 202 | flags |= FILE_FLAG_NO_BUFFERING; |
| 203 | if (td->o.sync_io) |
| 204 | flags |= FILE_FLAG_WRITE_THROUGH; |
| 205 | |
| 206 | /* |
| 207 | * Inform Windows whether we're going to be doing sequential or |
| 208 | * random IO so it can tune the Cache Manager |
| 209 | */ |
| 210 | switch (td->o.fadvise_hint) { |
| 211 | case F_ADV_TYPE: |
| 212 | if (td_random(td)) |
| 213 | flags |= FILE_FLAG_RANDOM_ACCESS; |
| 214 | else |
| 215 | flags |= FILE_FLAG_SEQUENTIAL_SCAN; |
| 216 | break; |
| 217 | case F_ADV_RANDOM: |
| 218 | flags |= FILE_FLAG_RANDOM_ACCESS; |
| 219 | break; |
| 220 | case F_ADV_SEQUENTIAL: |
| 221 | flags |= FILE_FLAG_SEQUENTIAL_SCAN; |
| 222 | break; |
| 223 | case F_ADV_NONE: |
| 224 | break; |
| 225 | default: |
| 226 | log_err("fio: unknown fadvise type %d\n", td->o.fadvise_hint); |
| 227 | } |
| 228 | |
| 229 | if (!td_write(td) || read_only) |
| 230 | access = GENERIC_READ; |
| 231 | else |
| 232 | access = (GENERIC_READ | GENERIC_WRITE); |
| 233 | |
| 234 | if (td->o.create_on_open) |
| 235 | openmode = OPEN_ALWAYS; |
| 236 | else |
| 237 | openmode = OPEN_EXISTING; |
| 238 | |
| 239 | /* If we're going to use direct I/O, Windows will try and invalidate |
| 240 | * its cache at that point so there's no need to do it here */ |
| 241 | if (td->o.invalidate_cache && !td->o.odirect) |
| 242 | windowsaio_invalidate_cache(f); |
| 243 | |
| 244 | f->hFile = CreateFile(f->file_name, access, sharemode, |
| 245 | NULL, openmode, flags, NULL); |
| 246 | |
| 247 | if (f->hFile == INVALID_HANDLE_VALUE) { |
| 248 | log_err("windowsaio: failed to open file \"%s\"\n", f->file_name); |
| 249 | rc = 1; |
| 250 | } |
| 251 | |
| 252 | /* Only set up the completion port and thread if we're not just |
| 253 | * querying the device size */ |
| 254 | if (!rc && td->io_ops_data != NULL) { |
| 255 | struct windowsaio_data *wd; |
| 256 | |
| 257 | wd = td->io_ops_data; |
| 258 | |
| 259 | if (CreateIoCompletionPort(f->hFile, wd->iocp, 0, 0) == NULL) { |
| 260 | log_err("windowsaio: failed to create io completion port\n"); |
| 261 | rc = 1; |
| 262 | } |
| 263 | } |
| 264 | |
| 265 | return rc; |
| 266 | } |
| 267 | |
| 268 | static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f) |
| 269 | { |
| 270 | int rc = 0; |
| 271 | |
| 272 | dprint(FD_FILE, "fd close %s\n", f->file_name); |
| 273 | |
| 274 | if (f->hFile != INVALID_HANDLE_VALUE) { |
| 275 | if (!CloseHandle(f->hFile)) { |
| 276 | log_info("windowsaio: failed to close file handle for \"%s\"\n", f->file_name); |
| 277 | rc = 1; |
| 278 | } |
| 279 | } |
| 280 | |
| 281 | f->hFile = INVALID_HANDLE_VALUE; |
| 282 | return rc; |
| 283 | } |
| 284 | |
| 285 | static BOOL timeout_expired(DWORD start_count, DWORD end_count) |
| 286 | { |
| 287 | BOOL expired = FALSE; |
| 288 | DWORD current_time; |
| 289 | |
| 290 | current_time = GetTickCount(); |
| 291 | |
| 292 | if ((end_count > start_count) && current_time >= end_count) |
| 293 | expired = TRUE; |
| 294 | else if (current_time < start_count && current_time > end_count) |
| 295 | expired = TRUE; |
| 296 | |
| 297 | return expired; |
| 298 | } |
| 299 | |
| 300 | static struct io_u* fio_windowsaio_event(struct thread_data *td, int event) |
| 301 | { |
| 302 | struct windowsaio_data *wd = td->io_ops_data; |
| 303 | return wd->aio_events[event]; |
| 304 | } |
| 305 | |
| 306 | static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min, |
| 307 | unsigned int max, |
| 308 | const struct timespec *t) |
| 309 | { |
| 310 | struct windowsaio_data *wd = td->io_ops_data; |
| 311 | unsigned int dequeued = 0; |
| 312 | struct io_u *io_u; |
| 313 | int i; |
| 314 | struct fio_overlapped *fov; |
| 315 | DWORD start_count = 0; |
| 316 | DWORD end_count = 0; |
| 317 | DWORD status; |
| 318 | DWORD mswait = 250; |
| 319 | |
| 320 | if (t != NULL) { |
| 321 | mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000); |
| 322 | start_count = GetTickCount(); |
| 323 | end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000); |
| 324 | } |
| 325 | |
| 326 | do { |
| 327 | io_u_qiter(&td->io_u_all, io_u, i) { |
| 328 | if (!(io_u->flags & IO_U_F_FLIGHT)) |
| 329 | continue; |
| 330 | |
| 331 | fov = (struct fio_overlapped*)io_u->engine_data; |
| 332 | |
| 333 | if (fov->io_complete) { |
| 334 | fov->io_complete = FALSE; |
| 335 | wd->aio_events[dequeued] = io_u; |
| 336 | dequeued++; |
| 337 | } |
| 338 | |
| 339 | } |
| 340 | if (dequeued >= min) |
| 341 | break; |
| 342 | |
| 343 | if (dequeued < min) { |
| 344 | status = WaitForSingleObject(wd->iocomplete_event, mswait); |
| 345 | if (status != WAIT_OBJECT_0 && dequeued >= min) |
| 346 | break; |
| 347 | } |
| 348 | |
| 349 | if (dequeued >= min || |
| 350 | (t != NULL && timeout_expired(start_count, end_count))) |
| 351 | break; |
| 352 | } while (1); |
| 353 | |
| 354 | return dequeued; |
| 355 | } |
| 356 | |
| 357 | static enum fio_q_status fio_windowsaio_queue(struct thread_data *td, |
| 358 | struct io_u *io_u) |
| 359 | { |
| 360 | struct fio_overlapped *o = io_u->engine_data; |
| 361 | LPOVERLAPPED lpOvl = &o->o; |
| 362 | BOOL success = FALSE; |
| 363 | int rc = FIO_Q_COMPLETED; |
| 364 | |
| 365 | fio_ro_check(td, io_u); |
| 366 | |
| 367 | lpOvl->Internal = 0; |
| 368 | lpOvl->InternalHigh = 0; |
| 369 | lpOvl->Offset = io_u->offset & 0xFFFFFFFF; |
| 370 | lpOvl->OffsetHigh = io_u->offset >> 32; |
| 371 | |
| 372 | switch (io_u->ddir) { |
| 373 | case DDIR_WRITE: |
| 374 | success = WriteFile(io_u->file->hFile, io_u->xfer_buf, |
| 375 | io_u->xfer_buflen, NULL, lpOvl); |
| 376 | break; |
| 377 | case DDIR_READ: |
| 378 | success = ReadFile(io_u->file->hFile, io_u->xfer_buf, |
| 379 | io_u->xfer_buflen, NULL, lpOvl); |
| 380 | break; |
| 381 | case DDIR_SYNC: |
| 382 | case DDIR_DATASYNC: |
| 383 | case DDIR_SYNC_FILE_RANGE: |
| 384 | success = FlushFileBuffers(io_u->file->hFile); |
| 385 | if (!success) { |
| 386 | log_err("windowsaio: failed to flush file buffers\n"); |
| 387 | io_u->error = win_to_posix_error(GetLastError()); |
| 388 | } |
| 389 | |
| 390 | return FIO_Q_COMPLETED; |
| 391 | case DDIR_TRIM: |
| 392 | log_err("windowsaio: manual TRIM isn't supported on Windows\n"); |
| 393 | io_u->error = 1; |
| 394 | io_u->resid = io_u->xfer_buflen; |
| 395 | return FIO_Q_COMPLETED; |
| 396 | default: |
| 397 | assert(0); |
| 398 | break; |
| 399 | } |
| 400 | |
| 401 | if (success || GetLastError() == ERROR_IO_PENDING) |
| 402 | rc = FIO_Q_QUEUED; |
| 403 | else { |
| 404 | io_u->error = win_to_posix_error(GetLastError()); |
| 405 | io_u->resid = io_u->xfer_buflen; |
| 406 | } |
| 407 | |
| 408 | return rc; |
| 409 | } |
| 410 | |
| 411 | /* Runs as a thread and waits for queued IO to complete */ |
| 412 | static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter) |
| 413 | { |
| 414 | OVERLAPPED *ovl; |
| 415 | struct fio_overlapped *fov; |
| 416 | struct io_u *io_u; |
| 417 | struct windowsaio_data *wd; |
| 418 | struct thread_ctx *ctx; |
| 419 | ULONG_PTR ulKey = 0; |
| 420 | DWORD bytes; |
| 421 | |
| 422 | ctx = (struct thread_ctx*)lpParameter; |
| 423 | wd = ctx->wd; |
| 424 | |
| 425 | do { |
| 426 | BOOL ret; |
| 427 | |
| 428 | ret = GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, |
| 429 | &ovl, 250); |
| 430 | if (!ret && ovl == NULL) |
| 431 | continue; |
| 432 | |
| 433 | fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o); |
| 434 | io_u = fov->io_u; |
| 435 | |
| 436 | if (ovl->Internal == ERROR_SUCCESS) { |
| 437 | io_u->resid = io_u->xfer_buflen - ovl->InternalHigh; |
| 438 | io_u->error = 0; |
| 439 | } else { |
| 440 | io_u->resid = io_u->xfer_buflen; |
| 441 | io_u->error = win_to_posix_error(GetLastError()); |
| 442 | } |
| 443 | |
| 444 | fov->io_complete = TRUE; |
| 445 | SetEvent(wd->iocomplete_event); |
| 446 | } while (ctx->wd->iothread_running); |
| 447 | |
| 448 | CloseHandle(ctx->iocp); |
| 449 | free(ctx); |
| 450 | return 0; |
| 451 | } |
| 452 | |
| 453 | static void fio_windowsaio_io_u_free(struct thread_data *td, struct io_u *io_u) |
| 454 | { |
| 455 | struct fio_overlapped *o = io_u->engine_data; |
| 456 | |
| 457 | if (o) { |
| 458 | io_u->engine_data = NULL; |
| 459 | free(o); |
| 460 | } |
| 461 | } |
| 462 | |
| 463 | static int fio_windowsaio_io_u_init(struct thread_data *td, struct io_u *io_u) |
| 464 | { |
| 465 | struct fio_overlapped *o; |
| 466 | |
| 467 | o = malloc(sizeof(*o)); |
| 468 | o->io_complete = FALSE; |
| 469 | o->io_u = io_u; |
| 470 | o->o.hEvent = NULL; |
| 471 | io_u->engine_data = o; |
| 472 | return 0; |
| 473 | } |
| 474 | |
| 475 | static struct ioengine_ops ioengine = { |
| 476 | .name = "windowsaio", |
| 477 | .version = FIO_IOOPS_VERSION, |
| 478 | .init = fio_windowsaio_init, |
| 479 | .queue = fio_windowsaio_queue, |
| 480 | .getevents = fio_windowsaio_getevents, |
| 481 | .event = fio_windowsaio_event, |
| 482 | .cleanup = fio_windowsaio_cleanup, |
| 483 | .open_file = fio_windowsaio_open_file, |
| 484 | .close_file = fio_windowsaio_close_file, |
| 485 | .get_file_size = generic_get_file_size, |
| 486 | .io_u_init = fio_windowsaio_io_u_init, |
| 487 | .io_u_free = fio_windowsaio_io_u_free, |
| 488 | }; |
| 489 | |
| 490 | static void fio_init fio_windowsaio_register(void) |
| 491 | { |
| 492 | register_ioengine(&ioengine); |
| 493 | } |
| 494 | |
| 495 | static void fio_exit fio_windowsaio_unregister(void) |
| 496 | { |
| 497 | unregister_ioengine(&ioengine); |
| 498 | } |