Bruce Cran | ecc314b | 2011-01-04 10:59:30 +0100 | [diff] [blame^] | 1 | /* |
| 2 | * Native Windows async IO engine |
| 3 | * Copyright (C) 2010 Bruce Cran <bruce@cran.org.uk> |
| 4 | */ |
| 5 | |
| 6 | |
| 7 | #include <stdio.h> |
| 8 | #include <stdlib.h> |
| 9 | #include <unistd.h> |
| 10 | #include <signal.h> |
| 11 | #include <errno.h> |
| 12 | #include <windows.h> |
| 13 | |
| 14 | #include "../fio.h" |
| 15 | |
| 16 | BOOL windowsaio_debug = FALSE; |
| 17 | |
| 18 | struct windowsaio_data { |
| 19 | struct io_u **aio_events; |
| 20 | unsigned int ioFinished; |
| 21 | BOOL running; |
| 22 | BOOL stopped; |
| 23 | HANDLE hThread; |
| 24 | }; |
| 25 | |
| 26 | typedef struct { |
| 27 | OVERLAPPED o; |
| 28 | struct io_u *io_u; |
| 29 | } FIO_OVERLAPPED; |
| 30 | |
| 31 | struct thread_ctx { |
| 32 | HANDLE ioCP; |
| 33 | struct windowsaio_data *wd; |
| 34 | }; |
| 35 | |
| 36 | static void PrintError(LPCSTR lpszFunction); |
| 37 | static int fio_windowsaio_cancel(struct thread_data *td, |
| 38 | struct io_u *io_u); |
| 39 | static DWORD GetEndCount(DWORD startCount, struct timespec *t); |
| 40 | static BOOL TimedOut(DWORD startCount, DWORD endCount); |
| 41 | static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min, |
| 42 | unsigned int max, struct timespec *t); |
| 43 | static struct io_u *fio_windowsaio_event(struct thread_data *td, int event); |
| 44 | static int fio_windowsaio_queue(struct thread_data *td, |
| 45 | struct io_u *io_u); |
| 46 | static void fio_windowsaio_cleanup(struct thread_data *td); |
| 47 | static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter); |
| 48 | static int fio_windowsaio_init(struct thread_data *td); |
| 49 | static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f); |
| 50 | static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f); |
| 51 | |
| 52 | static void PrintError(LPCSTR lpszFunction) |
| 53 | { |
| 54 | // Retrieve the system error message for the last-error code |
| 55 | |
| 56 | LPSTR lpMsgBuf; |
| 57 | DWORD dw = GetLastError(); |
| 58 | |
| 59 | FormatMessage( |
| 60 | FORMAT_MESSAGE_ALLOCATE_BUFFER | |
| 61 | FORMAT_MESSAGE_FROM_SYSTEM | |
| 62 | FORMAT_MESSAGE_IGNORE_INSERTS, |
| 63 | NULL, |
| 64 | dw, |
| 65 | MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), |
| 66 | (LPTSTR)&lpMsgBuf, |
| 67 | 0, NULL ); |
| 68 | |
| 69 | log_err("%s - %s", lpszFunction, lpMsgBuf); |
| 70 | LocalFree(lpMsgBuf); |
| 71 | } |
| 72 | |
| 73 | static int fio_windowsaio_cancel(struct thread_data *td, |
| 74 | struct io_u *io_u) |
| 75 | { |
| 76 | BOOL bSuccess; |
| 77 | int rc = 0; |
| 78 | |
| 79 | bSuccess = CancelIo(io_u->file->hFile); |
| 80 | |
| 81 | if (!bSuccess) |
| 82 | rc = 1; |
| 83 | |
| 84 | return rc; |
| 85 | } |
| 86 | |
| 87 | static DWORD GetEndCount(DWORD startCount, struct timespec *t) |
| 88 | { |
| 89 | DWORD endCount = startCount; |
| 90 | |
| 91 | if (t == NULL) |
| 92 | return 0; |
| 93 | |
| 94 | endCount += (t->tv_sec * 1000) + (t->tv_nsec / 1000000); |
| 95 | return endCount; |
| 96 | } |
| 97 | |
| 98 | static BOOL TimedOut(DWORD startCount, DWORD endCount) |
| 99 | { |
| 100 | BOOL expired = FALSE; |
| 101 | DWORD currentTime; |
| 102 | |
| 103 | if (startCount == 0 || endCount == 0) |
| 104 | return FALSE; |
| 105 | |
| 106 | currentTime = GetTickCount(); |
| 107 | |
| 108 | if ((endCount > startCount) && currentTime >= endCount) |
| 109 | expired = TRUE; |
| 110 | else if (currentTime < startCount && currentTime > endCount) |
| 111 | expired = TRUE; |
| 112 | |
| 113 | if (windowsaio_debug) |
| 114 | printf("windowsaio: timedout = %d\n", expired); |
| 115 | |
| 116 | return expired; |
| 117 | } |
| 118 | |
| 119 | static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min, |
| 120 | unsigned int max, struct timespec *t) |
| 121 | { |
| 122 | struct windowsaio_data *wd = td->io_ops->data; |
| 123 | struct flist_head *entry; |
| 124 | unsigned int dequeued = 0; |
| 125 | struct io_u *io_u; |
| 126 | DWORD startCount = 0, endCount = 0; |
| 127 | BOOL timedout = FALSE; |
| 128 | unsigned int r = 0; |
| 129 | |
| 130 | if (windowsaio_debug) |
| 131 | printf("getevents (min %d, max %d)\n", min, max); |
| 132 | |
| 133 | if (t != NULL) { |
| 134 | startCount = GetTickCount(); |
| 135 | endCount = GetEndCount(startCount, t); |
| 136 | } |
| 137 | |
| 138 | while (dequeued < min && !timedout) { |
| 139 | |
| 140 | flist_for_each(entry, &td->io_u_busylist) { |
| 141 | io_u = flist_entry(entry, struct io_u, list); |
| 142 | |
| 143 | if (io_u->seen == 0) |
| 144 | continue; |
| 145 | |
| 146 | dequeued++; |
| 147 | |
| 148 | wd->ioFinished--; |
| 149 | wd->aio_events[r] = io_u; |
| 150 | r++; |
| 151 | |
| 152 | if (windowsaio_debug) |
| 153 | printf("dequeued %d\n", dequeued); |
| 154 | |
| 155 | if (dequeued == max) |
| 156 | break; |
| 157 | } |
| 158 | |
| 159 | if (TimedOut(startCount, endCount)) |
| 160 | timedout = TRUE; |
| 161 | |
| 162 | if (dequeued < min && !timedout) |
| 163 | Sleep(250); |
| 164 | } |
| 165 | |
| 166 | if (windowsaio_debug) |
| 167 | printf("leave getevents (%d)\n", dequeued); |
| 168 | |
| 169 | return dequeued; |
| 170 | } |
| 171 | |
| 172 | static struct io_u *fio_windowsaio_event(struct thread_data *td, int event) |
| 173 | { |
| 174 | struct windowsaio_data *wd = td->io_ops->data; |
| 175 | return wd->aio_events[event]; |
| 176 | } |
| 177 | |
| 178 | static int fio_windowsaio_queue(struct thread_data *td, |
| 179 | struct io_u *io_u) |
| 180 | { |
| 181 | FIO_OVERLAPPED *fov; |
| 182 | DWORD ioBytes; |
| 183 | BOOL bSuccess = TRUE; |
| 184 | int rc; |
| 185 | |
| 186 | fio_ro_check(td, io_u); |
| 187 | |
| 188 | if (windowsaio_debug) |
| 189 | printf("enqueue enter\n"); |
| 190 | |
| 191 | fov = malloc(sizeof(FIO_OVERLAPPED)); |
| 192 | ZeroMemory(fov, sizeof(FIO_OVERLAPPED)); |
| 193 | |
| 194 | io_u->seen = 0; |
| 195 | |
| 196 | fov->o.Offset = io_u->offset & 0xFFFFFFFF; |
| 197 | fov->o.OffsetHigh = io_u->offset >> 32; |
| 198 | fov->o.hEvent = CreateEvent(NULL, FALSE, FALSE, NULL); |
| 199 | fov->io_u = io_u; |
| 200 | |
| 201 | if (fov->o.hEvent == NULL) { |
| 202 | PrintError(__func__); |
| 203 | return 1; |
| 204 | } |
| 205 | |
| 206 | if (io_u->ddir == DDIR_WRITE) |
| 207 | bSuccess = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &ioBytes, &fov->o); |
| 208 | else if (io_u->ddir == DDIR_READ) |
| 209 | bSuccess = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &ioBytes, &fov->o); |
| 210 | else if (io_u->ddir == DDIR_SYNC || |
| 211 | io_u->ddir == DDIR_DATASYNC || |
| 212 | io_u->ddir == DDIR_SYNC_FILE_RANGE) |
| 213 | { |
| 214 | FlushFileBuffers(io_u->file->hFile); |
| 215 | return FIO_Q_COMPLETED; |
| 216 | } else if (io_u->ddir == DDIR_TRIM) { |
| 217 | log_info("explicit TRIM isn't supported on Windows"); |
| 218 | return FIO_Q_COMPLETED; |
| 219 | } |
| 220 | |
| 221 | if (bSuccess) { |
| 222 | io_u->seen = 1; |
| 223 | io_u->resid = io_u->xfer_buflen - fov->o.InternalHigh; |
| 224 | io_u->error = 0; |
| 225 | rc = FIO_Q_COMPLETED; |
| 226 | } else if (!bSuccess && GetLastError() == ERROR_IO_PENDING) { |
| 227 | rc = FIO_Q_QUEUED; |
| 228 | } else { |
| 229 | PrintError(__func__); |
| 230 | io_u->error = GetLastError(); |
| 231 | io_u->resid = io_u->xfer_buflen; |
| 232 | rc = FIO_Q_COMPLETED; |
| 233 | } |
| 234 | |
| 235 | if (windowsaio_debug) |
| 236 | printf("enqueue - leave (offset %llu)\n", io_u->offset); |
| 237 | |
| 238 | return rc; |
| 239 | } |
| 240 | |
| 241 | static void fio_windowsaio_cleanup(struct thread_data *td) |
| 242 | { |
| 243 | struct windowsaio_data *wd; |
| 244 | |
| 245 | if (windowsaio_debug) |
| 246 | printf("windowsaio: cleanup - enter\n"); |
| 247 | |
| 248 | wd = td->io_ops->data; |
| 249 | wd->running = FALSE; |
| 250 | |
| 251 | while (wd->stopped == FALSE) |
| 252 | Sleep(5); |
| 253 | |
| 254 | if (wd != NULL) { |
| 255 | CloseHandle(wd->hThread); |
| 256 | free(wd->aio_events); |
| 257 | wd->aio_events = NULL; |
| 258 | free(wd); |
| 259 | td->io_ops->data = NULL; |
| 260 | } |
| 261 | |
| 262 | if (windowsaio_debug) |
| 263 | printf("windowsaio: cleanup - leave\n"); |
| 264 | } |
| 265 | |
| 266 | static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter) |
| 267 | { |
| 268 | OVERLAPPED *ovl; |
| 269 | FIO_OVERLAPPED *fov; |
| 270 | struct io_u *io_u; |
| 271 | struct windowsaio_data *wd; |
| 272 | |
| 273 | struct thread_ctx *ctx; |
| 274 | ULONG_PTR ulKey = 0; |
| 275 | BOOL bSuccess; |
| 276 | DWORD bytes; |
| 277 | |
| 278 | |
| 279 | ctx = (struct thread_ctx*)lpParameter; |
| 280 | wd = ctx->wd; |
| 281 | bSuccess = TRUE; |
| 282 | |
| 283 | if (windowsaio_debug) |
| 284 | printf("windowsaio: IoCompletionRoutine - enter\n"); |
| 285 | |
| 286 | while (ctx->wd->running) { |
| 287 | bSuccess = GetQueuedCompletionStatus(ctx->ioCP, &bytes, &ulKey, &ovl, 500); |
| 288 | |
| 289 | if (windowsaio_debug) |
| 290 | printf("GetQueuedCompletionStatus returned %d\n", bSuccess); |
| 291 | |
| 292 | if (!bSuccess) { |
| 293 | if (GetLastError() == WAIT_TIMEOUT) { |
| 294 | continue; |
| 295 | } else { |
| 296 | PrintError(__func__); |
| 297 | continue; |
| 298 | } |
| 299 | } |
| 300 | |
| 301 | fov = CONTAINING_RECORD(ovl, FIO_OVERLAPPED, o); |
| 302 | io_u = fov->io_u; |
| 303 | |
| 304 | if (windowsaio_debug) { |
| 305 | if (io_u->seen == 1) |
| 306 | printf("IoCompletionRoutine - got already completed IO\n"); |
| 307 | else |
| 308 | printf("IoCompletionRoutine - completed %d IO\n", ctx->wd->ioFinished); |
| 309 | } |
| 310 | |
| 311 | if (io_u->seen == 1) |
| 312 | continue; |
| 313 | |
| 314 | ctx->wd->ioFinished++; |
| 315 | |
| 316 | if (ovl->Internal == ERROR_SUCCESS) { |
| 317 | io_u->resid = io_u->xfer_buflen - ovl->InternalHigh; |
| 318 | io_u->error = 0; |
| 319 | } else { |
| 320 | io_u->resid = io_u->xfer_buflen; |
| 321 | io_u->error = 1; |
| 322 | } |
| 323 | |
| 324 | io_u->seen = 1; |
| 325 | CloseHandle(ovl->hEvent); |
| 326 | free(ovl); |
| 327 | } |
| 328 | |
| 329 | bSuccess = CloseHandle(ctx->ioCP); |
| 330 | if (!bSuccess) |
| 331 | PrintError(__func__); |
| 332 | |
| 333 | if (windowsaio_debug) |
| 334 | printf("windowsaio: IoCompletionRoutine - leave\n"); |
| 335 | |
| 336 | ctx->wd->stopped = TRUE; |
| 337 | free(ctx); |
| 338 | return 0; |
| 339 | } |
| 340 | |
| 341 | static int fio_windowsaio_init(struct thread_data *td) |
| 342 | { |
| 343 | int rc = 0; |
| 344 | struct windowsaio_data *wd; |
| 345 | |
| 346 | if (windowsaio_debug) |
| 347 | printf("windowsaio: init\n"); |
| 348 | |
| 349 | wd = malloc(sizeof(struct windowsaio_data)); |
| 350 | |
| 351 | ZeroMemory(wd, sizeof(*wd)); |
| 352 | wd->aio_events = malloc((td->o.iodepth + 1) * sizeof(struct io_u *)); |
| 353 | ZeroMemory(wd->aio_events, (td->o.iodepth + 1) * sizeof(struct io_u *)); |
| 354 | |
| 355 | td->io_ops->data = wd; |
| 356 | return rc; |
| 357 | } |
| 358 | |
| 359 | static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f) |
| 360 | { |
| 361 | int rc = 0; |
| 362 | HANDLE hFile; |
| 363 | DWORD flags = FILE_ATTRIBUTE_TEMPORARY | FILE_FLAG_POSIX_SEMANTICS | FILE_FLAG_OVERLAPPED; |
| 364 | DWORD sharemode = FILE_SHARE_READ | FILE_SHARE_WRITE; |
| 365 | DWORD openmode = OPEN_ALWAYS; |
| 366 | DWORD access; |
| 367 | |
| 368 | dprint(FD_FILE, "fd open %s\n", f->file_name); |
| 369 | |
| 370 | if (windowsaio_debug) |
| 371 | printf("windowsaio: open file %s - enter\n", f->file_name); |
| 372 | |
| 373 | if (f->filetype == FIO_TYPE_PIPE) { |
| 374 | log_err("fio: windowsaio doesn't support pipes\n"); |
| 375 | return 1; |
| 376 | } |
| 377 | |
| 378 | if (!strcmp(f->file_name, "-")) { |
| 379 | log_err("fio: can't read/write to stdin/out\n"); |
| 380 | return 1; |
| 381 | } |
| 382 | |
| 383 | if (td->o.odirect) |
| 384 | flags |= FILE_FLAG_NO_BUFFERING; |
| 385 | if (td->o.sync_io) |
| 386 | flags |= FILE_FLAG_WRITE_THROUGH; |
| 387 | |
| 388 | |
| 389 | if (td->o.td_ddir == TD_DDIR_READ || |
| 390 | td->o.td_ddir == TD_DDIR_WRITE || |
| 391 | td->o.td_ddir == TD_DDIR_RANDRW) |
| 392 | { |
| 393 | flags |= FILE_FLAG_SEQUENTIAL_SCAN; |
| 394 | } |
| 395 | else |
| 396 | { |
| 397 | flags |= FILE_FLAG_RANDOM_ACCESS; |
| 398 | } |
| 399 | |
| 400 | if (td_read(td) || read_only) |
| 401 | access = GENERIC_READ; |
| 402 | else |
| 403 | access = (GENERIC_READ | GENERIC_WRITE); |
| 404 | |
| 405 | if (td->o.create_on_open > 0) |
| 406 | openmode = OPEN_ALWAYS; |
| 407 | else |
| 408 | openmode = OPEN_EXISTING; |
| 409 | |
| 410 | f->hFile = CreateFile(f->file_name, access, sharemode, |
| 411 | NULL, openmode, flags, NULL); |
| 412 | |
| 413 | if (f->hFile == INVALID_HANDLE_VALUE) { |
| 414 | log_err("Failed to open %s\n", f->file_name); |
| 415 | PrintError(__func__); |
| 416 | rc = 1; |
| 417 | } |
| 418 | |
| 419 | /* Only set up the competion port and thread if we're not just |
| 420 | * querying the device size */ |
| 421 | if (!rc && td->io_ops->data != NULL) { |
| 422 | struct windowsaio_data *wd; |
| 423 | struct thread_ctx *ctx; |
| 424 | hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0); |
| 425 | |
| 426 | wd = td->io_ops->data; |
| 427 | wd->running = TRUE; |
| 428 | wd->stopped = FALSE; |
| 429 | |
| 430 | ctx = malloc(sizeof(struct thread_ctx)); |
| 431 | ctx->ioCP = hFile; |
| 432 | ctx->wd = wd; |
| 433 | |
| 434 | wd->hThread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL); |
| 435 | |
| 436 | if (wd->hThread == NULL) { |
| 437 | PrintError(__func__); |
| 438 | rc = 1; |
| 439 | } |
| 440 | } |
| 441 | |
| 442 | if (windowsaio_debug) |
| 443 | printf("windowsaio: open file - leave (%d)\n", rc); |
| 444 | |
| 445 | return rc; |
| 446 | } |
| 447 | |
| 448 | static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f) |
| 449 | { |
| 450 | BOOL bSuccess; |
| 451 | |
| 452 | if (windowsaio_debug) |
| 453 | printf("windowsaio: close file\n"); |
| 454 | |
| 455 | if (f->hFile != INVALID_HANDLE_VALUE) { |
| 456 | bSuccess = CloseHandle(f->hFile); |
| 457 | if (!bSuccess) |
| 458 | PrintError(__func__); |
| 459 | } |
| 460 | |
| 461 | f->hFile = INVALID_HANDLE_VALUE; |
| 462 | return 0; |
| 463 | } |
| 464 | |
| 465 | static struct ioengine_ops ioengine = { |
| 466 | .name = "windowsaio", |
| 467 | .version = FIO_IOOPS_VERSION, |
| 468 | .init = fio_windowsaio_init, |
| 469 | .queue = fio_windowsaio_queue, |
| 470 | .cancel = fio_windowsaio_cancel, |
| 471 | .getevents = fio_windowsaio_getevents, |
| 472 | .event = fio_windowsaio_event, |
| 473 | .cleanup = fio_windowsaio_cleanup, |
| 474 | .open_file = fio_windowsaio_open_file, |
| 475 | .close_file = fio_windowsaio_close_file, |
| 476 | .get_file_size = generic_get_file_size |
| 477 | }; |
| 478 | |
| 479 | static void fio_init fio_posixaio_register(void) |
| 480 | { |
| 481 | register_ioengine(&ioengine); |
| 482 | } |
| 483 | |
| 484 | static void fio_exit fio_posixaio_unregister(void) |
| 485 | { |
| 486 | unregister_ioengine(&ioengine); |
| 487 | } |