ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libev/ev_linuxaio.c
(Generate patch)

Comparing libev/ev_linuxaio.c (file contents):
Revision 1.17 by root, Mon Jun 24 02:02:35 2019 UTC vs.
Revision 1.23 by root, Mon Jun 24 21:34:47 2019 UTC

53 * therefore the calculation below will use "exactly" 4kB for the ring buffer 53 * therefore the calculation below will use "exactly" 4kB for the ring buffer
54 */ 54 */
55#define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */ 55#define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */
56 56
57/*****************************************************************************/ 57/*****************************************************************************/
58/* syscall wrapdadoop */ 58/* syscall wrapdadoop - this section has the raw syscall definitions */
59 59
60#include <sys/syscall.h> /* no glibc wrappers */ 60#include <sys/syscall.h> /* no glibc wrappers */
61 61
62/* aio_abi.h is not versioned in any way, so we cannot test for its existance */ 62/* aio_abi.h is not versioned in any way, so we cannot test for its existance */
63#define IOCB_CMD_POLL 5 63#define IOCB_CMD_POLL 5
80 struct io_event io_events[0]; 80 struct io_event io_events[0];
81}; 81};
82 82
83inline_size 83inline_size
84int 84int
85ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 85evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
86{ 86{
87 return syscall (SYS_io_setup, nr_events, ctx_idp); 87 return syscall (SYS_io_setup, nr_events, ctx_idp);
88} 88}
89 89
90inline_size 90inline_size
91int 91int
92ev_io_destroy (aio_context_t ctx_id) 92evsys_io_destroy (aio_context_t ctx_id)
93{ 93{
94 return syscall (SYS_io_destroy, ctx_id); 94 return syscall (SYS_io_destroy, ctx_id);
95} 95}
96 96
97inline_size 97inline_size
98int 98int
99ev_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) 99evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
100{ 100{
101 return syscall (SYS_io_submit, ctx_id, nr, cbp); 101 return syscall (SYS_io_submit, ctx_id, nr, cbp);
102} 102}
103 103
104inline_size 104inline_size
105int 105int
106ev_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) 106evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result)
107{ 107{
108 return syscall (SYS_io_cancel, ctx_id, cbp, result); 108 return syscall (SYS_io_cancel, ctx_id, cbp, result);
109} 109}
110 110
111inline_size 111inline_size
112int 112int
113ev_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) 113evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout)
114{ 114{
115 return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); 115 return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
116} 116}
117 117
118/*****************************************************************************/ 118/*****************************************************************************/
125 /*int inuse;*/ 125 /*int inuse;*/
126} *ANIOCBP; 126} *ANIOCBP;
127 127
128inline_size 128inline_size
129void 129void
130linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) 130linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
131{ 131{
132 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
133 while (count--) 132 while (count--)
134 { 133 {
134 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
135 *base = (ANIOCBP)ev_malloc (sizeof (**base)); 135 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
136 /* TODO: full zero initialize required? */ 136
137 /* full zero initialise is probably not required at the moment, but
138 * this is not well documented, so we better do it.
139 */
137 memset (*base, 0, sizeof (**base)); 140 memset (iocb, 0, sizeof (*iocb));
138 /* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */ 141
139 (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; 142 iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
140 ++base; 143 iocb->io.aio_data = offset;
144 iocb->io.aio_fildes = offset;
145
146 base [offset++] = iocb;
141 } 147 }
142} 148}
143 149
144ecb_cold 150ecb_cold
145static void 151static void
153 159
154static void 160static void
155linuxaio_modify (EV_P_ int fd, int oev, int nev) 161linuxaio_modify (EV_P_ int fd, int oev, int nev)
156{ 162{
157 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); 163 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
158 struct aniocb *iocb = linuxaio_iocbps [fd]; 164 ANIOCBP iocb = linuxaio_iocbps [fd];
159 165
160#if EPOLL_FALLBACK 166#if EPOLL_FALLBACK
161 if (iocb->io.aio_reqprio < 0) 167 if (iocb->io.aio_reqprio < 0)
162 { 168 {
163 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); 169 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
164 iocb->io.aio_reqprio = 0; 170 iocb->io.aio_reqprio = 0;
165 } 171 }
166#endif 172#endif
167 173
168 if (iocb->io.aio_buf) 174 if (iocb->io.aio_buf)
169 ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ 175 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */
170 176
171 if (nev) 177 if (nev)
172 { 178 {
173 iocb->io.aio_data = fd;
174 iocb->io.aio_fildes = fd;
175 iocb->io.aio_buf = 179 iocb->io.aio_buf =
176 (nev & EV_READ ? POLLIN : 0) 180 (nev & EV_READ ? POLLIN : 0)
177 | (nev & EV_WRITE ? POLLOUT : 0); 181 | (nev & EV_WRITE ? POLLOUT : 0);
178 182
179 /* queue iocb up for io_submit */ 183 /* queue iocb up for io_submit */
180 /* this assumes we only ever get one call per fd per loop iteration */ 184 /* this assumes we only ever get one call per fd per loop iteration */
182 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); 186 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
183 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 187 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
184 } 188 }
185} 189}
186 190
191#if EPOLL_FALLBACK
192
193static void
194linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op)
195{
196 struct epoll_event eev;
197
198 eev.events = EPOLLONESHOT;
199 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
200 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
201 eev.data.fd = iocb->aio_fildes;
202
203 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
204 ev_syserr ("(libeio) linuxaio epoll_ctl");
205}
206
207static void
208linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
209{
210 struct epoll_event events[16];
211
212 for (;;)
213 {
214 int idx;
215 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
216
217 if (expect_false (res < 0))
218 ev_syserr ("(libev) linuxaio epoll_wait");
219 else if (!res)
220 break;
221
222 for (idx = res; idx--; )
223 {
224 int fd = events [idx].data.fd;
225 uint32_t ev = events [idx].events;
226
227 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
228
229 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
230
231 fd_event (EV_A_ fd,
232 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
233 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
234 }
235
236 if (res < sizeof (events) / sizeof (events [0]))
237 break;
238 }
239}
240
241#endif
242
187static void 243static void
188linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) 244linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
189{ 245{
190 while (nr) 246 while (nr)
191 { 247 {
192 int fd = ev->data; 248 int fd = ev->data;
193 int res = ev->res; 249 int res = ev->res;
194 250
195 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); 251 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
196 252
197 /* linux aio is oneshot: rearm fd */ 253 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
198 linuxaio_iocbps [fd]->io.aio_buf = 0; 254 linuxaio_iocbps [fd]->io.aio_buf = 0;
199 anfds [fd].events = 0; 255 anfds [fd].events = 0;
200 fd_change (EV_A_ fd, 0); 256 fd_change (EV_A_ fd, 0);
201 257
202 /* feed events, we do not expect or handle POLLNVAL */ 258 /* feed events, we do not expect or handle POLLNVAL */
203 if (expect_false (res & POLLNVAL))
204 fd_kill (EV_A_ fd);
205 else
206 fd_event ( 259 fd_event (
207 EV_A_ 260 EV_A_
208 fd, 261 fd,
209 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 262 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
210 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 263 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
211 ); 264 );
212 265
213 --nr; 266 --nr;
214 ++ev; 267 ++ev;
215 } 268 }
216} 269}
262inline_size 315inline_size
263void 316void
264linuxaio_get_events (EV_P_ ev_tstamp timeout) 317linuxaio_get_events (EV_P_ ev_tstamp timeout)
265{ 318{
266 struct timespec ts; 319 struct timespec ts;
267 struct io_event ioev; 320 struct io_event ioev[1];
268 int res; 321 int res;
269 322
270 if (linuxaio_get_events_from_ring (EV_A)) 323 if (linuxaio_get_events_from_ring (EV_A))
271 return; 324 return;
272 325
273 /* no events, so wait for at least one, then poll ring buffer again */ 326 /* no events, so wait for at least one, then poll ring buffer again */
274 /* this degrades to one event per loop iteration */ 327 /* this degrades to one event per loop iteration */
275 /* if the ring buffer changes layout, but so be it */ 328 /* if the ring buffer changes layout, but so be it */
276 329
330 EV_RELEASE_CB;
331
277 ts.tv_sec = (long)timeout; 332 ts.tv_sec = (long)timeout;
278 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 333 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
279 334
280 res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); 335 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
336
337 EV_ACQUIRE_CB;
281 338
282 if (res < 0) 339 if (res < 0)
283 if (errno == EINTR) 340 if (errno == EINTR)
284 /* ignored */; 341 /* ignored */;
285 else 342 else
286 ev_syserr ("(libev) linuxaio io_getevents"); 343 ev_syserr ("(libev) linuxaio io_getevents");
287 else if (res) 344 else if (res)
288 { 345 {
289 /* at least one event received, handle it and any remaining ones in the ring buffer */ 346 /* at least one event received, handle it and any remaining ones in the ring buffer */
290 linuxaio_parse_events (EV_A_ &ioev, 1); 347 linuxaio_parse_events (EV_A_ ioev, res);
291 linuxaio_get_events_from_ring (EV_A); 348 linuxaio_get_events_from_ring (EV_A);
292 } 349 }
293} 350}
294
295#if EPOLL_FALLBACK
296static void
297linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op)
298{
299 struct epoll_event eev;
300
301 eev.events = EPOLLONESHOT;
302 if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
303 if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
304 eev.data.fd = iocb->aio_fildes;
305
306 if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
307 ev_syserr ("(libeio) linuxaio epoll_ctl");
308}
309#endif
310 351
311static void 352static void
312linuxaio_poll (EV_P_ ev_tstamp timeout) 353linuxaio_poll (EV_P_ ev_tstamp timeout)
313{ 354{
314 int submitted; 355 int submitted;
318 /* io_submit might return less than the requested number of iocbs */ 359 /* io_submit might return less than the requested number of iocbs */
319 /* this is, afaics, only because of errors, but we go by the book and use a loop, */ 360 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
320 /* which allows us to pinpoint the errornous iocb */ 361 /* which allows us to pinpoint the errornous iocb */
321 for (submitted = 0; submitted < linuxaio_submitcnt; ) 362 for (submitted = 0; submitted < linuxaio_submitcnt; )
322 { 363 {
364#if 0
365 int res;
366 if (linuxaio_submits[submitted]->aio_fildes == backend_fd)
367 res = evsys_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted);
368 else
369 { res = -1; errno = EINVAL; };
370#else
323 int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 371 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
372#endif
324 373
325 if (expect_false (res < 0)) 374 if (expect_false (res < 0))
326 if (errno == EAGAIN) 375 if (errno == EAGAIN)
327 { 376 {
328 /* This happens when the ring buffer is full, at least. I assume this means 377 /* This happens when the ring buffer is full, at least. I assume this means
329 * that the event was queued synchronously during io_submit, and thus 378 * that the event was queued synchronously during io_submit, and thus
330 * the buffer overflowed. 379 * the buffer overflowed.
331 * In this case, we just try in next loop iteration. 380 * In this case, we just try in next loop iteration.
332 * This should not result in a few fds taking priority, as the interface 381 * This should not result in a few fds taking priority, as the interface
333 * is one-shot, and we submit iocb's in a round-robin fashion. 382 * is one-shot, and we submit iocb's in a round-robin fashion.
383 * TODO: maybe make "submitted" persistent, so we don't have to memmove?
334 */ 384 */
385 if (ecb_expect_false (submitted))
386 {
335 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); 387 memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
336 linuxaio_submitcnt -= submitted; 388 linuxaio_submitcnt -= submitted;
389 }
390
337 timeout = 0; 391 timeout = 0;
338 break; 392 break;
339 } 393 }
340#if EPOLL_FALLBACK 394#if EPOLL_FALLBACK
341 else if (errno == EINVAL) 395 else if (errno == EINVAL)
342 { 396 {
343 /* This happens for unsupported fds, officially, but in my testing, 397 /* This happens for unsupported fds, officially, but in my testing,
344 * also randomly happens for supported fds. We fall back to good old 398 * also randomly happens for supported fds. We fall back to good old
345 * poll() here, under the assumption that this is a very rare case. 399 * poll() here, under the assumption that this is a very rare case.
346 * See https://lore.kernel.org/patchwork/patch/1047453/ for evidence 400 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
347 * that the problem is known, but ignored. 401 * discussion about such a case (ttys) where polling for POLLIN
402 * fails but POLLIN|POLLOUT works.
348 */ 403 */
349 struct iocb *iocb = linuxaio_submits [submitted]; 404 struct iocb *iocb = linuxaio_submits [submitted];
405
406 linuxaio_rearm_epoll (EV_A_ linuxaio_submits [submitted], EPOLL_CTL_ADD);
407 iocb->aio_reqprio = -1; /* mark iocb as epoll */
408
350 res = 1; /* skip this iocb */ 409 res = 1; /* skip this iocb */
351
352 linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD);
353 iocb->aio_reqprio = -1; /* mark iocb as epoll */
354 } 410 }
355#endif 411#endif
412 else if (errno == EBADF)
413 {
414 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
415
416 res = 1; /* skip this iocb */
417 }
356 else 418 else
357 ev_syserr ("(libev) linuxaio io_submit"); 419 ev_syserr ("(libev) linuxaio io_submit");
358 420
359 submitted += res; 421 submitted += res;
360 } 422 }
363 425
364 /* second phase: fetch and parse events */ 426 /* second phase: fetch and parse events */
365 427
366 linuxaio_get_events (EV_A_ timeout); 428 linuxaio_get_events (EV_A_ timeout);
367} 429}
368
369#if EPOLL_FALLBACK
370
371static void
372linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
373{
374 struct epoll_event events[16];
375
376 for (;;)
377 {
378 int idx;
379 int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
380
381 if (expect_false (res < 0))
382 ev_syserr ("(libev) linuxaio epoll_wait");
383 else if (!res)
384 break;
385
386 for (idx = res; idx--; )
387 {
388 int fd = events [idx].data.fd;
389 uint32_t ev = events [idx].events;
390
391 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
392
393 linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
394
395 fd_event (EV_A_ fd,
396 (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
397 | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0));
398 }
399
400 if (res < sizeof (events) / sizeof (events [0]))
401 break;
402 }
403}
404
405#endif
406 430
407inline_size 431inline_size
408int 432int
409linuxaio_init (EV_P_ int flags) 433linuxaio_init (EV_P_ int flags)
410{ 434{
419 if (ev_linux_version () < 0x041200) 443 if (ev_linux_version () < 0x041200)
420 return 0; 444 return 0;
421#endif 445#endif
422 446
423 linuxaio_ctx = 0; 447 linuxaio_ctx = 0;
424 if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 448 if (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
425 return 0; 449 return 0;
426 450
427#if EPOLL_FALLBACK 451#if EPOLL_FALLBACK
428 backend_fd = ev_epoll_create (); 452 backend_fd = ev_epoll_create ();
429 if (backend_fd < 0) 453 if (backend_fd < 0)
430 { 454 {
431 ev_io_destroy (linuxaio_ctx); 455 evsys_io_destroy (linuxaio_ctx);
432 return 0; 456 return 0;
433 } 457 }
434 458
435 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 459 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
460 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
436 ev_io_start (EV_A_ &linuxaio_epoll_w); 461 ev_io_start (EV_A_ &linuxaio_epoll_w);
437 ev_unref (EV_A); /* watcher should not keep loop alive */ 462 ev_unref (EV_A); /* watcher should not keep loop alive */
438#endif 463#endif
439 464
440 backend_modify = linuxaio_modify; 465 backend_modify = linuxaio_modify;
456{ 481{
457#if EPOLL_FALLBACK 482#if EPOLL_FALLBACK
458 close (backend_fd); 483 close (backend_fd);
459#endif 484#endif
460 linuxaio_free_iocbp (EV_A); 485 linuxaio_free_iocbp (EV_A);
461 ev_io_destroy (linuxaio_ctx); 486 evsys_io_destroy (linuxaio_ctx);
462} 487}
463 488
464inline_size 489inline_size
465void 490void
466linuxaio_fork (EV_P) 491linuxaio_fork (EV_P)
468 /* this frees all iocbs, which is very heavy-handed */ 493 /* this frees all iocbs, which is very heavy-handed */
469 linuxaio_destroy (EV_A); 494 linuxaio_destroy (EV_A);
470 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 495 linuxaio_submitcnt = 0; /* all pointers were invalidated */
471 496
472 linuxaio_ctx = 0; 497 linuxaio_ctx = 0;
473 while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) 498 while (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
474 ev_syserr ("(libev) linuxaio io_setup"); 499 ev_syserr ("(libev) linuxaio io_setup");
475 500
476#if EPOLL_FALLBACK 501#if EPOLL_FALLBACK
477 while ((backend_fd = ev_epoll_create ()) < 0) 502 while ((backend_fd = ev_epoll_create ()) < 0)
478 ev_syserr ("(libev) linuxaio epoll_create"); 503 ev_syserr ("(libev) linuxaio epoll_create");

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines