Commit | Line | Data |
---|---|---|
ecc314ba BC |
1 | /* |
2 | * Native Windows async IO engine | |
ea4500d8 | 3 | * Copyright (C) 2011 Bruce Cran <bruce@cran.org.uk> |
ecc314ba BC |
4 | */ |
5 | ||
ecc314ba BC |
6 | #include <stdio.h> |
7 | #include <stdlib.h> | |
8 | #include <unistd.h> | |
9 | #include <signal.h> | |
10 | #include <errno.h> | |
11 | #include <windows.h> | |
12 | ||
13 | #include "../fio.h" | |
14 | ||
ea4500d8 BC |
15 | typedef BOOL (WINAPI *CANCELIOEX)(HANDLE hFile, LPOVERLAPPED lpOverlapped); |
16 | ||
67897036 BC |
17 | struct fio_overlapped { |
18 | OVERLAPPED o; | |
19 | struct io_u *io_u; | |
20 | BOOL io_complete; | |
21 | BOOL io_free; | |
22 | }; | |
ecc314ba | 23 | |
9b836561 | 24 | struct windowsaio_data { |
67897036 | 25 | struct fio_overlapped *ovls; |
9b836561 | 26 | struct io_u **aio_events; |
67897036 | 27 | HANDLE iothread; |
9b836561 | 28 | HANDLE iocomplete_event; |
ea4500d8 | 29 | CANCELIOEX pCancelIoEx; |
67897036 BC |
30 | BOOL iothread_running; |
31 | BOOL use_iocp; | |
9b836561 BC |
32 | }; |
33 | ||
ecc314ba | 34 | struct thread_ctx { |
9b836561 | 35 | HANDLE iocp; |
ecc314ba BC |
36 | struct windowsaio_data *wd; |
37 | }; | |
38 | ||
39 | static void PrintError(LPCSTR lpszFunction); | |
40 | static int fio_windowsaio_cancel(struct thread_data *td, | |
41 | struct io_u *io_u); | |
67897036 | 42 | static BOOL timeout_expired(DWORD start_count, DWORD end_count); |
ecc314ba BC |
43 | static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min, |
44 | unsigned int max, struct timespec *t); | |
45 | static struct io_u *fio_windowsaio_event(struct thread_data *td, int event); | |
46 | static int fio_windowsaio_queue(struct thread_data *td, | |
47 | struct io_u *io_u); | |
48 | static void fio_windowsaio_cleanup(struct thread_data *td); | |
49 | static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter); | |
50 | static int fio_windowsaio_init(struct thread_data *td); | |
51 | static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f); | |
52 | static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f); | |
53 | ||
9b836561 BC |
54 | int sync_file_range(int fd, off64_t offset, off64_t nbytes, |
55 | unsigned int flags) | |
56 | { | |
57 | errno = ENOSYS; | |
58 | return -1; | |
59 | } | |
60 | ||
ecc314ba BC |
61 | static void PrintError(LPCSTR lpszFunction) |
62 | { | |
63 | // Retrieve the system error message for the last-error code | |
64 | ||
65 | LPSTR lpMsgBuf; | |
66 | DWORD dw = GetLastError(); | |
67 | ||
68 | FormatMessage( | |
69 | FORMAT_MESSAGE_ALLOCATE_BUFFER | | |
70 | FORMAT_MESSAGE_FROM_SYSTEM | | |
71 | FORMAT_MESSAGE_IGNORE_INSERTS, | |
72 | NULL, | |
73 | dw, | |
74 | MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), | |
75 | (LPTSTR)&lpMsgBuf, | |
76 | 0, NULL ); | |
77 | ||
78 | log_err("%s - %s", lpszFunction, lpMsgBuf); | |
79 | LocalFree(lpMsgBuf); | |
80 | } | |
81 | ||
ecc314ba BC |
82 | static int fio_windowsaio_init(struct thread_data *td) |
83 | { | |
ecc314ba | 84 | struct windowsaio_data *wd; |
ea4500d8 | 85 | HANDLE hKernel32Dll; |
9b836561 | 86 | int rc = 0; |
67897036 | 87 | int i; |
ecc314ba | 88 | |
ecc314ba | 89 | wd = malloc(sizeof(struct windowsaio_data)); |
9b836561 BC |
90 | if (wd != NULL) |
91 | ZeroMemory(wd, sizeof(struct windowsaio_data)); | |
92 | else | |
93 | rc = 1; | |
ecc314ba | 94 | |
9b836561 BC |
95 | if (!rc) { |
96 | wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*)); | |
97 | if (wd->aio_events == NULL) | |
98 | rc = 1; | |
e4db9fec BC |
99 | } |
100 | ||
9b836561 | 101 | if (!rc) { |
67897036 BC |
102 | wd->ovls = malloc(td->o.iodepth * sizeof(struct fio_overlapped)); |
103 | if (wd->ovls == NULL) | |
9b836561 | 104 | rc = 1; |
e4db9fec BC |
105 | } |
106 | ||
9b836561 | 107 | if (!rc) { |
67897036 BC |
108 | for (i = 0; i < td->o.iodepth; i++) { |
109 | wd->ovls[i].io_free = TRUE; | |
110 | wd->ovls[i].io_complete = FALSE; | |
111 | ||
112 | wd->ovls[i].o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL); | |
113 | if (wd->ovls[i].o.hEvent == NULL) { | |
114 | rc = 1; | |
115 | break; | |
116 | } | |
117 | } | |
9b836561 | 118 | } |
e4db9fec | 119 | |
9b836561 BC |
120 | if (!rc) { |
121 | /* Create an auto-reset event */ | |
122 | wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL); | |
123 | if (wd->iocomplete_event == NULL) | |
124 | rc = 1; | |
125 | } | |
126 | ||
9b836561 BC |
127 | if (rc) { |
128 | PrintError(__func__); | |
129 | if (wd != NULL) { | |
130 | if (wd->ovls != NULL) | |
131 | free(wd->ovls); | |
9b836561 BC |
132 | if (wd->aio_events != NULL) |
133 | free(wd->aio_events); | |
134 | ||
135 | free(wd); | |
136 | } | |
137 | } | |
ecc314ba | 138 | |
ea4500d8 BC |
139 | hKernel32Dll = GetModuleHandle("kernel32.dll"); |
140 | wd->pCancelIoEx = GetProcAddress(hKernel32Dll, "CancelIoEx"); | |
141 | ||
ecc314ba | 142 | td->io_ops->data = wd; |
e4db9fec | 143 | return 0; |
ecc314ba BC |
144 | } |
145 | ||
67897036 BC |
146 | static void fio_windowsaio_cleanup(struct thread_data *td) |
147 | { | |
148 | int i; | |
149 | struct windowsaio_data *wd; | |
150 | ||
151 | wd = td->io_ops->data; | |
152 | ||
153 | if (wd != NULL) { | |
154 | wd->iothread_running = FALSE; | |
155 | WaitForSingleObject(wd->iothread, INFINITE); | |
156 | ||
157 | CloseHandle(wd->iothread); | |
158 | CloseHandle(wd->iocomplete_event); | |
159 | ||
160 | for (i = 0; i < td->o.iodepth; i++) { | |
161 | CloseHandle(wd->ovls[i].o.hEvent); | |
162 | } | |
163 | ||
164 | free(wd->aio_events); | |
165 | free(wd->ovls); | |
166 | free(wd); | |
167 | ||
168 | td->io_ops->data = NULL; | |
169 | } | |
170 | } | |
171 | ||
172 | ||
ecc314ba BC |
173 | static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f) |
174 | { | |
175 | int rc = 0; | |
176 | HANDLE hFile; | |
ea4500d8 | 177 | DWORD flags = FILE_FLAG_POSIX_SEMANTICS; |
ecc314ba BC |
178 | DWORD sharemode = FILE_SHARE_READ | FILE_SHARE_WRITE; |
179 | DWORD openmode = OPEN_ALWAYS; | |
180 | DWORD access; | |
181 | ||
182 | dprint(FD_FILE, "fd open %s\n", f->file_name); | |
183 | ||
ecc314ba BC |
184 | if (f->filetype == FIO_TYPE_PIPE) { |
185 | log_err("fio: windowsaio doesn't support pipes\n"); | |
186 | return 1; | |
187 | } | |
188 | ||
189 | if (!strcmp(f->file_name, "-")) { | |
190 | log_err("fio: can't read/write to stdin/out\n"); | |
191 | return 1; | |
192 | } | |
193 | ||
ea4500d8 BC |
194 | if (!td->o.odirect && !td->o.sync_io && td->io_ops->data != NULL) |
195 | flags |= FILE_FLAG_OVERLAPPED; | |
196 | ||
ecc314ba BC |
197 | if (td->o.odirect) |
198 | flags |= FILE_FLAG_NO_BUFFERING; | |
199 | if (td->o.sync_io) | |
200 | flags |= FILE_FLAG_WRITE_THROUGH; | |
201 | ||
202 | ||
203 | if (td->o.td_ddir == TD_DDIR_READ || | |
ea4500d8 | 204 | td->o.td_ddir == TD_DDIR_WRITE) |
ecc314ba | 205 | flags |= FILE_FLAG_SEQUENTIAL_SCAN; |
ecc314ba | 206 | else |
ecc314ba | 207 | flags |= FILE_FLAG_RANDOM_ACCESS; |
ecc314ba | 208 | |
ea4500d8 | 209 | if (!td_write(td) || read_only) |
ecc314ba BC |
210 | access = GENERIC_READ; |
211 | else | |
212 | access = (GENERIC_READ | GENERIC_WRITE); | |
213 | ||
214 | if (td->o.create_on_open > 0) | |
215 | openmode = OPEN_ALWAYS; | |
216 | else | |
217 | openmode = OPEN_EXISTING; | |
218 | ||
219 | f->hFile = CreateFile(f->file_name, access, sharemode, | |
220 | NULL, openmode, flags, NULL); | |
221 | ||
222 | if (f->hFile == INVALID_HANDLE_VALUE) { | |
ecc314ba BC |
223 | PrintError(__func__); |
224 | rc = 1; | |
225 | } | |
226 | ||
227 | /* Only set up the competion port and thread if we're not just | |
228 | * querying the device size */ | |
ea4500d8 | 229 | if (!rc && td->io_ops->data != NULL && !td->o.odirect && !td->o.sync_io) { |
ecc314ba | 230 | struct thread_ctx *ctx; |
ea4500d8 | 231 | struct windowsaio_data *wd; |
ecc314ba BC |
232 | hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0); |
233 | ||
ea4500d8 BC |
234 | wd = td->io_ops->data; |
235 | ||
236 | if (!td->o.odirect && !td->o.sync_io) | |
67897036 | 237 | wd->use_iocp = 1; |
ea4500d8 | 238 | else |
67897036 | 239 | wd->use_iocp = 0; |
ecc314ba | 240 | |
9b836561 | 241 | wd->iothread_running = TRUE; |
ecc314ba | 242 | |
9b836561 BC |
243 | if (!rc) { |
244 | ctx = malloc(sizeof(struct thread_ctx)); | |
245 | ctx->iocp = hFile; | |
246 | ctx->wd = wd; | |
247 | ||
248 | wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL); | |
249 | } | |
ecc314ba | 250 | |
9b836561 | 251 | if (rc || wd->iothread == NULL) { |
ecc314ba BC |
252 | PrintError(__func__); |
253 | rc = 1; | |
254 | } | |
255 | } | |
256 | ||
ecc314ba BC |
257 | return rc; |
258 | } | |
259 | ||
260 | static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f) | |
261 | { | |
9b836561 BC |
262 | dprint(FD_FILE, "fd close %s\n", f->file_name); |
263 | ||
ecc314ba | 264 | if (f->hFile != INVALID_HANDLE_VALUE) { |
9b836561 | 265 | if (!CloseHandle(f->hFile)) |
ecc314ba BC |
266 | PrintError(__func__); |
267 | } | |
268 | ||
269 | f->hFile = INVALID_HANDLE_VALUE; | |
270 | return 0; | |
271 | } | |
272 | ||
67897036 BC |
273 | static BOOL timeout_expired(DWORD start_count, DWORD end_count) |
274 | { | |
275 | BOOL expired = FALSE; | |
276 | DWORD current_time; | |
277 | ||
278 | current_time = GetTickCount(); | |
279 | ||
280 | if ((end_count > start_count) && current_time >= end_count) | |
281 | expired = TRUE; | |
282 | else if (current_time < start_count && current_time > end_count) | |
283 | expired = TRUE; | |
284 | ||
285 | return expired; | |
286 | } | |
287 | ||
288 | static struct io_u* fio_windowsaio_event(struct thread_data *td, int event) | |
289 | { | |
290 | struct windowsaio_data *wd = td->io_ops->data; | |
291 | return wd->aio_events[event]; | |
292 | } | |
293 | ||
294 | static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min, | |
295 | unsigned int max, struct timespec *t) | |
296 | { | |
297 | struct windowsaio_data *wd = td->io_ops->data; | |
298 | struct flist_head *entry; | |
299 | unsigned int dequeued = 0; | |
300 | struct io_u *io_u; | |
301 | struct fio_overlapped *fov; | |
302 | DWORD start_count = 0; | |
303 | DWORD end_count = 0; | |
304 | DWORD status; | |
305 | DWORD mswait = 250; | |
306 | ||
307 | if (t != NULL) { | |
308 | mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000); | |
309 | start_count = GetTickCount(); | |
310 | end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000); | |
311 | } | |
312 | ||
313 | do { | |
314 | flist_for_each(entry, &td->io_u_busylist) { | |
315 | io_u = flist_entry(entry, struct io_u, list); | |
316 | fov = (struct fio_overlapped*)io_u->engine_data; | |
317 | ||
318 | if (fov->io_complete) { | |
319 | fov->io_complete = FALSE; | |
320 | fov->io_free = TRUE; | |
321 | wd->aio_events[dequeued] = io_u; | |
322 | dequeued++; | |
323 | } | |
324 | ||
325 | if (dequeued >= min) | |
326 | break; | |
327 | } | |
328 | ||
329 | if (dequeued < min) { | |
330 | status = WaitForSingleObject(wd->iocomplete_event, mswait); | |
331 | if (status != WAIT_OBJECT_0 && dequeued > 0) | |
332 | break; | |
333 | } | |
334 | ||
335 | if (dequeued >= min || (t != NULL && timeout_expired(start_count, end_count))) | |
336 | break; | |
337 | } while (1); | |
338 | ||
339 | return dequeued; | |
340 | } | |
341 | ||
342 | static int fio_windowsaio_queue(struct thread_data *td, | |
343 | struct io_u *io_u) | |
344 | { | |
345 | LPOVERLAPPED lpOvl = NULL; | |
346 | struct windowsaio_data *wd; | |
347 | DWORD iobytes; | |
348 | BOOL success; | |
349 | int index; | |
350 | int rc = FIO_Q_COMPLETED; | |
351 | ||
352 | fio_ro_check(td, io_u); | |
353 | ||
354 | wd = td->io_ops->data; | |
355 | ||
356 | if (wd->use_iocp) { | |
357 | for (index = 0; index < td->o.iodepth; index++) { | |
358 | if (wd->ovls[index].io_free) { | |
359 | wd->ovls[index].io_free = FALSE; | |
360 | ResetEvent(wd->ovls[index].o.hEvent); | |
361 | break; | |
362 | } | |
363 | } | |
364 | ||
365 | assert(index < td->o.iodepth); | |
366 | ||
367 | lpOvl = &wd->ovls[index].o; | |
368 | wd->ovls[index].io_u = io_u; | |
369 | lpOvl->Internal = STATUS_PENDING; | |
370 | lpOvl->InternalHigh = 0; | |
371 | lpOvl->Offset = io_u->offset & 0xFFFFFFFF; | |
372 | lpOvl->OffsetHigh = io_u->offset >> 32; | |
373 | lpOvl->Pointer = NULL; | |
374 | io_u->engine_data = &wd->ovls[index]; | |
375 | } | |
376 | ||
377 | switch (io_u->ddir) { | |
378 | case DDIR_WRITE: | |
379 | success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl); | |
380 | break; | |
381 | case DDIR_READ: | |
382 | success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl); | |
383 | break; | |
384 | case DDIR_SYNC: | |
385 | case DDIR_DATASYNC: | |
386 | case DDIR_SYNC_FILE_RANGE: | |
387 | success = FlushFileBuffers(io_u->file->hFile); | |
388 | if (!success) | |
389 | io_u->error = GetLastError(); | |
390 | ||
391 | return FIO_Q_COMPLETED; | |
392 | break; | |
393 | case DDIR_TRIM: | |
394 | log_err("manual TRIM isn't supported on Windows"); | |
395 | io_u->error = 1; | |
396 | io_u->resid = io_u->xfer_buflen; | |
397 | return FIO_Q_COMPLETED; | |
398 | break; | |
399 | default: | |
400 | assert(0); | |
401 | } | |
402 | ||
403 | if (wd->use_iocp && (success || GetLastError() == ERROR_IO_PENDING)) { | |
404 | rc = FIO_Q_QUEUED; | |
405 | } else if (success && !wd->use_iocp) { | |
406 | io_u->resid = io_u->xfer_buflen - iobytes; | |
407 | io_u->error = 0; | |
408 | } else { | |
409 | PrintError(__func__); | |
410 | io_u->error = GetLastError(); | |
411 | io_u->resid = io_u->xfer_buflen; | |
412 | } | |
413 | ||
414 | return rc; | |
415 | } | |
416 | ||
417 | /* Runs as a thread and waits for queued IO to complete */ | |
418 | static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter) | |
419 | { | |
420 | OVERLAPPED *ovl; | |
421 | struct fio_overlapped *fov; | |
422 | struct io_u *io_u; | |
423 | struct windowsaio_data *wd; | |
424 | struct thread_ctx *ctx; | |
425 | ULONG_PTR ulKey = 0; | |
426 | DWORD bytes; | |
427 | ||
428 | ctx = (struct thread_ctx*)lpParameter; | |
429 | wd = ctx->wd; | |
430 | ||
431 | do { | |
432 | if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250)) | |
433 | continue; | |
434 | ||
435 | fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o); | |
436 | io_u = fov->io_u; | |
437 | ||
438 | if (ovl->Internal == ERROR_SUCCESS) { | |
439 | io_u->resid = io_u->xfer_buflen - ovl->InternalHigh; | |
440 | io_u->error = 0; | |
441 | } else { | |
442 | io_u->resid = io_u->xfer_buflen; | |
443 | io_u->error = ovl->Internal; | |
444 | } | |
445 | ||
446 | fov->io_complete = TRUE; | |
447 | SetEvent(wd->iocomplete_event); | |
448 | } while (ctx->wd->iothread_running); | |
449 | ||
450 | CloseHandle(ctx->iocp); | |
451 | free(ctx); | |
452 | return 0; | |
453 | } | |
454 | ||
455 | static int fio_windowsaio_cancel(struct thread_data *td, | |
456 | struct io_u *io_u) | |
457 | { | |
458 | int rc = 0; | |
459 | ||
460 | struct windowsaio_data *wd = td->io_ops->data; | |
461 | ||
462 | /* If we're running on Vista or newer, we can cancel individual IO requests */ | |
463 | if (wd->pCancelIoEx != NULL) { | |
464 | struct fio_overlapped *ovl = io_u->engine_data; | |
465 | if (!wd->pCancelIoEx(io_u->file->hFile, &ovl->o)) | |
466 | rc = 1; | |
467 | } else | |
468 | rc = 1; | |
469 | ||
470 | return rc; | |
471 | } | |
472 | ||
ecc314ba BC |
473 | static struct ioengine_ops ioengine = { |
474 | .name = "windowsaio", | |
475 | .version = FIO_IOOPS_VERSION, | |
476 | .init = fio_windowsaio_init, | |
477 | .queue = fio_windowsaio_queue, | |
478 | .cancel = fio_windowsaio_cancel, | |
479 | .getevents = fio_windowsaio_getevents, | |
480 | .event = fio_windowsaio_event, | |
481 | .cleanup = fio_windowsaio_cleanup, | |
482 | .open_file = fio_windowsaio_open_file, | |
483 | .close_file = fio_windowsaio_close_file, | |
484 | .get_file_size = generic_get_file_size | |
485 | }; | |
486 | ||
487 | static void fio_init fio_posixaio_register(void) | |
488 | { | |
489 | register_ioengine(&ioengine); | |
490 | } | |
491 | ||
492 | static void fio_exit fio_posixaio_unregister(void) | |
493 | { | |
494 | unregister_ioengine(&ioengine); | |
495 | } |