Commit | Line | Data |
---|---|---|
ecc314ba | 1 | /* |
03244c19 BC |
2 | * windowsaio engine |
3 | * | |
4 | * IO engine using Windows IO Completion Ports. | |
ecc314ba BC |
5 | */ |
6 | ||
ecc314ba BC |
7 | #include <stdio.h> |
8 | #include <stdlib.h> | |
9 | #include <unistd.h> | |
10 | #include <signal.h> | |
11 | #include <errno.h> | |
ecc314ba BC |
12 | |
13 | #include "../fio.h" | |
14 | ||
ea4500d8 BC |
15 | typedef BOOL (WINAPI *CANCELIOEX)(HANDLE hFile, LPOVERLAPPED lpOverlapped); |
16 | ||
66c098b8 BC |
17 | int geterrno_from_win_error (DWORD code, int deferrno); |
18 | ||
67897036 BC |
19 | struct fio_overlapped { |
20 | OVERLAPPED o; | |
21 | struct io_u *io_u; | |
22 | BOOL io_complete; | |
67897036 | 23 | }; |
ecc314ba | 24 | |
9b836561 | 25 | struct windowsaio_data { |
9b836561 | 26 | struct io_u **aio_events; |
5a90bb5f | 27 | HANDLE iocp; |
67897036 | 28 | HANDLE iothread; |
9b836561 | 29 | HANDLE iocomplete_event; |
67897036 | 30 | BOOL iothread_running; |
9b836561 BC |
31 | }; |
32 | ||
ecc314ba | 33 | struct thread_ctx { |
9b836561 | 34 | HANDLE iocp; |
ecc314ba BC |
35 | struct windowsaio_data *wd; |
36 | }; | |
37 | ||
ecc314ba | 38 | static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter); |
ecc314ba | 39 | |
ecc314ba BC |
40 | static int fio_windowsaio_init(struct thread_data *td) |
41 | { | |
ecc314ba | 42 | struct windowsaio_data *wd; |
9b836561 | 43 | int rc = 0; |
ecc314ba | 44 | |
03244c19 BC |
45 | wd = calloc(1, sizeof(struct windowsaio_data)); |
46 | if (wd == NULL) { | |
47 | log_err("windowsaio: failed to allocate memory for engine data\n"); | |
9b836561 | 48 | rc = 1; |
03244c19 | 49 | } |
ecc314ba | 50 | |
9b836561 BC |
51 | if (!rc) { |
52 | wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*)); | |
03244c19 BC |
53 | if (wd->aio_events == NULL) { |
54 | log_err("windowsaio: failed to allocate memory for aio events list\n"); | |
9b836561 | 55 | rc = 1; |
03244c19 | 56 | } |
e4db9fec BC |
57 | } |
58 | ||
9b836561 BC |
59 | if (!rc) { |
60 | /* Create an auto-reset event */ | |
61 | wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL); | |
03244c19 BC |
62 | if (wd->iocomplete_event == NULL) { |
63 | log_err("windowsaio: failed to create io complete event handle\n"); | |
9b836561 | 64 | rc = 1; |
03244c19 | 65 | } |
9b836561 BC |
66 | } |
67 | ||
9b836561 | 68 | if (rc) { |
9b836561 | 69 | if (wd != NULL) { |
9b836561 BC |
70 | if (wd->aio_events != NULL) |
71 | free(wd->aio_events); | |
72 | ||
73 | free(wd); | |
74 | } | |
75 | } | |
ecc314ba | 76 | |
565e784d | 77 | td->io_ops_data = wd; |
93bcfd20 | 78 | |
5a90bb5f BC |
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); | |
03244c19 BC |
85 | if (hFile == INVALID_HANDLE_VALUE) { |
86 | log_err("windowsaio: failed to create io completion port\n"); | |
5a90bb5f | 87 | rc = 1; |
03244c19 | 88 | } |
5a90bb5f | 89 | |
565e784d | 90 | wd = td->io_ops_data; |
5a90bb5f BC |
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 | { | |
03244c19 | 99 | log_err("windowsaio: failed to allocate memory for thread context structure\n"); |
5a90bb5f BC |
100 | CloseHandle(hFile); |
101 | rc = 1; | |
102 | } | |
103 | ||
104 | if (!rc) | |
105 | { | |
438bb1cf BC |
106 | DWORD threadid; |
107 | ||
5a90bb5f BC |
108 | ctx->iocp = hFile; |
109 | ctx->wd = wd; | |
438bb1cf BC |
110 | wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, &threadid); |
111 | ||
112 | if (wd->iothread != NULL) | |
113 | fio_setaffinity(threadid, td->o.cpumask); | |
114 | else | |
03244c19 | 115 | log_err("windowsaio: failed to create io completion thread\n"); |
5a90bb5f BC |
116 | } |
117 | ||
118 | if (rc || wd->iothread == NULL) | |
119 | rc = 1; | |
120 | } | |
121 | ||
4e79098f | 122 | return rc; |
ecc314ba BC |
123 | } |
124 | ||
67897036 BC |
125 | static void fio_windowsaio_cleanup(struct thread_data *td) |
126 | { | |
67897036 BC |
127 | struct windowsaio_data *wd; |
128 | ||
565e784d | 129 | wd = td->io_ops_data; |
67897036 BC |
130 | |
131 | if (wd != NULL) { | |
40c5db35 JA |
132 | wd->iothread_running = FALSE; |
133 | WaitForSingleObject(wd->iothread, INFINITE); | |
67897036 BC |
134 | |
135 | CloseHandle(wd->iothread); | |
136 | CloseHandle(wd->iocomplete_event); | |
137 | ||
67897036 | 138 | free(wd->aio_events); |
67897036 BC |
139 | free(wd); |
140 | ||
565e784d | 141 | td->io_ops_data = NULL; |
67897036 BC |
142 | } |
143 | } | |
144 | ||
8300eba5 SW |
145 | static int windowsaio_invalidate_cache(struct fio_file *f) |
146 | { | |
147 | DWORD error; | |
148 | DWORD isharemode = (FILE_SHARE_DELETE | FILE_SHARE_READ | | |
149 | FILE_SHARE_WRITE); | |
150 | HANDLE ihFile; | |
151 | int rc = 0; | |
152 | ||
153 | /* | |
154 | * Encourage Windows to drop cached parts of a file by temporarily | |
155 | * opening it for non-buffered access. Note: this will only work when | |
156 | * the following is the only thing with the file open on the whole | |
157 | * system. | |
158 | */ | |
159 | dprint(FD_IO, "windowaio: attempt invalidate cache for %s\n", | |
160 | f->file_name); | |
161 | ihFile = CreateFile(f->file_name, 0, isharemode, NULL, OPEN_EXISTING, | |
162 | FILE_FLAG_NO_BUFFERING, NULL); | |
163 | ||
164 | if (ihFile != INVALID_HANDLE_VALUE) { | |
165 | if (!CloseHandle(ihFile)) { | |
166 | error = GetLastError(); | |
167 | log_info("windowsaio: invalidation fd close %s " | |
168 | "failed: error %d\n", f->file_name, error); | |
169 | rc = 1; | |
170 | } | |
171 | } else { | |
172 | error = GetLastError(); | |
173 | if (error != ERROR_FILE_NOT_FOUND) { | |
174 | log_info("windowsaio: cache invalidation of %s failed: " | |
175 | "error %d\n", f->file_name, error); | |
176 | rc = 1; | |
177 | } | |
178 | } | |
179 | ||
180 | return rc; | |
181 | } | |
182 | ||
ecc314ba BC |
183 | static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f) |
184 | { | |
185 | int rc = 0; | |
4e79098f | 186 | DWORD flags = FILE_FLAG_POSIX_SEMANTICS | FILE_FLAG_OVERLAPPED; |
ecc314ba BC |
187 | DWORD sharemode = FILE_SHARE_READ | FILE_SHARE_WRITE; |
188 | DWORD openmode = OPEN_ALWAYS; | |
189 | DWORD access; | |
190 | ||
191 | dprint(FD_FILE, "fd open %s\n", f->file_name); | |
192 | ||
ecc314ba | 193 | if (f->filetype == FIO_TYPE_PIPE) { |
03244c19 | 194 | log_err("windowsaio: pipes are not supported\n"); |
ecc314ba BC |
195 | return 1; |
196 | } | |
197 | ||
198 | if (!strcmp(f->file_name, "-")) { | |
03244c19 | 199 | log_err("windowsaio: can't read/write to stdin/out\n"); |
ecc314ba BC |
200 | return 1; |
201 | } | |
202 | ||
203 | if (td->o.odirect) | |
204 | flags |= FILE_FLAG_NO_BUFFERING; | |
205 | if (td->o.sync_io) | |
206 | flags |= FILE_FLAG_WRITE_THROUGH; | |
207 | ||
93bcfd20 BC |
208 | /* |
209 | * Inform Windows whether we're going to be doing sequential or | |
f20a86a7 | 210 | * random IO so it can tune the Cache Manager |
93bcfd20 | 211 | */ |
f20a86a7 SW |
212 | switch (td->o.fadvise_hint) { |
213 | case F_ADV_TYPE: | |
214 | if (td_random(td)) | |
215 | flags |= FILE_FLAG_RANDOM_ACCESS; | |
216 | else | |
217 | flags |= FILE_FLAG_SEQUENTIAL_SCAN; | |
218 | break; | |
219 | case F_ADV_RANDOM: | |
ecc314ba | 220 | flags |= FILE_FLAG_RANDOM_ACCESS; |
f20a86a7 SW |
221 | break; |
222 | case F_ADV_SEQUENTIAL: | |
223 | flags |= FILE_FLAG_SEQUENTIAL_SCAN; | |
224 | break; | |
225 | case F_ADV_NONE: | |
226 | break; | |
227 | default: | |
228 | log_err("fio: unknown fadvise type %d\n", td->o.fadvise_hint); | |
229 | } | |
ecc314ba | 230 | |
ea4500d8 | 231 | if (!td_write(td) || read_only) |
ecc314ba BC |
232 | access = GENERIC_READ; |
233 | else | |
234 | access = (GENERIC_READ | GENERIC_WRITE); | |
235 | ||
93bcfd20 | 236 | if (td->o.create_on_open) |
ecc314ba BC |
237 | openmode = OPEN_ALWAYS; |
238 | else | |
239 | openmode = OPEN_EXISTING; | |
240 | ||
8300eba5 SW |
241 | /* If we're going to use direct I/O, Windows will try and invalidate |
242 | * its cache at that point so there's no need to do it here */ | |
3e9ae304 | 243 | if (td->o.invalidate_cache && !td->o.odirect) |
8300eba5 | 244 | windowsaio_invalidate_cache(f); |
8300eba5 | 245 | |
ecc314ba BC |
246 | f->hFile = CreateFile(f->file_name, access, sharemode, |
247 | NULL, openmode, flags, NULL); | |
248 | ||
03244c19 BC |
249 | if (f->hFile == INVALID_HANDLE_VALUE) { |
250 | log_err("windowsaio: failed to open file \"%s\"\n", f->file_name); | |
ecc314ba | 251 | rc = 1; |
03244c19 | 252 | } |
ecc314ba | 253 | |
93bcfd20 | 254 | /* Only set up the completion port and thread if we're not just |
ecc314ba | 255 | * querying the device size */ |
565e784d | 256 | if (!rc && td->io_ops_data != NULL) { |
40c5db35 | 257 | struct windowsaio_data *wd; |
ecc314ba | 258 | |
565e784d | 259 | wd = td->io_ops_data; |
ecc314ba | 260 | |
03244c19 BC |
261 | if (CreateIoCompletionPort(f->hFile, wd->iocp, 0, 0) == NULL) { |
262 | log_err("windowsaio: failed to create io completion port\n"); | |
ecc314ba | 263 | rc = 1; |
03244c19 | 264 | } |
ecc314ba BC |
265 | } |
266 | ||
ecc314ba BC |
267 | return rc; |
268 | } | |
269 | ||
270 | static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f) | |
271 | { | |
4e79098f | 272 | int rc = 0; |
1e42cc3d | 273 | |
9b836561 BC |
274 | dprint(FD_FILE, "fd close %s\n", f->file_name); |
275 | ||
ecc314ba | 276 | if (f->hFile != INVALID_HANDLE_VALUE) { |
03244c19 BC |
277 | if (!CloseHandle(f->hFile)) { |
278 | log_info("windowsaio: failed to close file handle for \"%s\"\n", f->file_name); | |
4e79098f | 279 | rc = 1; |
03244c19 | 280 | } |
ecc314ba BC |
281 | } |
282 | ||
283 | f->hFile = INVALID_HANDLE_VALUE; | |
4e79098f | 284 | return rc; |
ecc314ba BC |
285 | } |
286 | ||
67897036 BC |
287 | static BOOL timeout_expired(DWORD start_count, DWORD end_count) |
288 | { | |
289 | BOOL expired = FALSE; | |
290 | DWORD current_time; | |
291 | ||
292 | current_time = GetTickCount(); | |
293 | ||
294 | if ((end_count > start_count) && current_time >= end_count) | |
295 | expired = TRUE; | |
296 | else if (current_time < start_count && current_time > end_count) | |
297 | expired = TRUE; | |
298 | ||
299 | return expired; | |
300 | } | |
301 | ||
302 | static struct io_u* fio_windowsaio_event(struct thread_data *td, int event) | |
303 | { | |
565e784d | 304 | struct windowsaio_data *wd = td->io_ops_data; |
67897036 BC |
305 | return wd->aio_events[event]; |
306 | } | |
307 | ||
308 | static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min, | |
1f440ece JA |
309 | unsigned int max, |
310 | const struct timespec *t) | |
67897036 | 311 | { |
565e784d | 312 | struct windowsaio_data *wd = td->io_ops_data; |
67897036 BC |
313 | unsigned int dequeued = 0; |
314 | struct io_u *io_u; | |
7b2dfab1 | 315 | int i; |
67897036 BC |
316 | struct fio_overlapped *fov; |
317 | DWORD start_count = 0; | |
318 | DWORD end_count = 0; | |
319 | DWORD status; | |
320 | DWORD mswait = 250; | |
321 | ||
322 | if (t != NULL) { | |
323 | mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000); | |
324 | start_count = GetTickCount(); | |
325 | end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000); | |
326 | } | |
327 | ||
328 | do { | |
7b2dfab1 BC |
329 | io_u_qiter(&td->io_u_all, io_u, i) { |
330 | if (!(io_u->flags & IO_U_F_FLIGHT)) | |
331 | continue; | |
332 | ||
67897036 BC |
333 | fov = (struct fio_overlapped*)io_u->engine_data; |
334 | ||
335 | if (fov->io_complete) { | |
40c5db35 | 336 | fov->io_complete = FALSE; |
67897036 BC |
337 | wd->aio_events[dequeued] = io_u; |
338 | dequeued++; | |
339 | } | |
340 | ||
67897036 | 341 | } |
77e7b330 JR |
342 | if (dequeued >= min) |
343 | break; | |
67897036 | 344 | |
40c5db35 | 345 | if (dequeued < min) { |
67897036 | 346 | status = WaitForSingleObject(wd->iocomplete_event, mswait); |
f9a58c2a | 347 | if (status != WAIT_OBJECT_0 && dequeued >= min) |
03244c19 | 348 | break; |
67897036 BC |
349 | } |
350 | ||
351 | if (dequeued >= min || (t != NULL && timeout_expired(start_count, end_count))) | |
352 | break; | |
353 | } while (1); | |
354 | ||
355 | return dequeued; | |
356 | } | |
357 | ||
40c5db35 | 358 | static int fio_windowsaio_queue(struct thread_data *td, struct io_u *io_u) |
67897036 | 359 | { |
c73ed246 JA |
360 | struct fio_overlapped *o = io_u->engine_data; |
361 | LPOVERLAPPED lpOvl = &o->o; | |
93bcfd20 | 362 | BOOL success = FALSE; |
67897036 BC |
363 | int rc = FIO_Q_COMPLETED; |
364 | ||
365 | fio_ro_check(td, io_u); | |
366 | ||
77e7b330 | 367 | lpOvl->Internal = 0; |
4e79098f BC |
368 | lpOvl->InternalHigh = 0; |
369 | lpOvl->Offset = io_u->offset & 0xFFFFFFFF; | |
370 | lpOvl->OffsetHigh = io_u->offset >> 32; | |
67897036 BC |
371 | |
372 | switch (io_u->ddir) { | |
40c5db35 | 373 | case DDIR_WRITE: |
77e7b330 | 374 | success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, NULL, lpOvl); |
67897036 BC |
375 | break; |
376 | case DDIR_READ: | |
77e7b330 | 377 | success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, NULL, lpOvl); |
67897036 BC |
378 | break; |
379 | case DDIR_SYNC: | |
380 | case DDIR_DATASYNC: | |
381 | case DDIR_SYNC_FILE_RANGE: | |
382 | success = FlushFileBuffers(io_u->file->hFile); | |
03244c19 BC |
383 | if (!success) { |
384 | log_err("windowsaio: failed to flush file buffers\n"); | |
385 | io_u->error = win_to_posix_error(GetLastError()); | |
386 | } | |
67897036 BC |
387 | |
388 | return FIO_Q_COMPLETED; | |
389 | break; | |
390 | case DDIR_TRIM: | |
03244c19 | 391 | log_err("windowsaio: manual TRIM isn't supported on Windows\n"); |
67897036 BC |
392 | io_u->error = 1; |
393 | io_u->resid = io_u->xfer_buflen; | |
394 | return FIO_Q_COMPLETED; | |
395 | break; | |
396 | default: | |
397 | assert(0); | |
93bcfd20 | 398 | break; |
67897036 BC |
399 | } |
400 | ||
40c5db35 | 401 | if (success || GetLastError() == ERROR_IO_PENDING) |
67897036 | 402 | rc = FIO_Q_QUEUED; |
40c5db35 | 403 | else { |
2277d5d5 | 404 | io_u->error = win_to_posix_error(GetLastError()); |
67897036 BC |
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 { | |
66c098b8 | 426 | if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250) && ovl == NULL) |
67897036 BC |
427 | continue; |
428 | ||
429 | fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o); | |
430 | io_u = fov->io_u; | |
431 | ||
432 | if (ovl->Internal == ERROR_SUCCESS) { | |
433 | io_u->resid = io_u->xfer_buflen - ovl->InternalHigh; | |
434 | io_u->error = 0; | |
435 | } else { | |
436 | io_u->resid = io_u->xfer_buflen; | |
2277d5d5 | 437 | io_u->error = win_to_posix_error(GetLastError()); |
67897036 BC |
438 | } |
439 | ||
40c5db35 | 440 | fov->io_complete = TRUE; |
67897036 BC |
441 | SetEvent(wd->iocomplete_event); |
442 | } while (ctx->wd->iothread_running); | |
443 | ||
444 | CloseHandle(ctx->iocp); | |
445 | free(ctx); | |
446 | return 0; | |
447 | } | |
448 | ||
c73ed246 JA |
449 | static void fio_windowsaio_io_u_free(struct thread_data *td, struct io_u *io_u) |
450 | { | |
451 | struct fio_overlapped *o = io_u->engine_data; | |
452 | ||
453 | if (o) { | |
c73ed246 JA |
454 | io_u->engine_data = NULL; |
455 | free(o); | |
456 | } | |
457 | } | |
458 | ||
459 | static int fio_windowsaio_io_u_init(struct thread_data *td, struct io_u *io_u) | |
460 | { | |
461 | struct fio_overlapped *o; | |
462 | ||
463 | o = malloc(sizeof(*o)); | |
b0106419 | 464 | o->io_complete = FALSE; |
c73ed246 | 465 | o->io_u = io_u; |
77e7b330 | 466 | o->o.hEvent = NULL; |
c73ed246 JA |
467 | io_u->engine_data = o; |
468 | return 0; | |
469 | } | |
470 | ||
ecc314ba BC |
471 | static struct ioengine_ops ioengine = { |
472 | .name = "windowsaio", | |
473 | .version = FIO_IOOPS_VERSION, | |
474 | .init = fio_windowsaio_init, | |
475 | .queue = fio_windowsaio_queue, | |
ecc314ba BC |
476 | .getevents = fio_windowsaio_getevents, |
477 | .event = fio_windowsaio_event, | |
478 | .cleanup = fio_windowsaio_cleanup, | |
479 | .open_file = fio_windowsaio_open_file, | |
480 | .close_file = fio_windowsaio_close_file, | |
c73ed246 JA |
481 | .get_file_size = generic_get_file_size, |
482 | .io_u_init = fio_windowsaio_io_u_init, | |
483 | .io_u_free = fio_windowsaio_io_u_free, | |
ecc314ba BC |
484 | }; |
485 | ||
c874d188 | 486 | static void fio_init fio_windowsaio_register(void) |
ecc314ba BC |
487 | { |
488 | register_ioengine(&ioengine); | |
489 | } | |
490 | ||
c874d188 | 491 | static void fio_exit fio_windowsaio_unregister(void) |
ecc314ba BC |
492 | { |
493 | unregister_ioengine(&ioengine); | |
494 | } |