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 1.28 by root, Tue Jun 25 06:36:59 2019 UTC vs.
Revision 1.44 by root, Tue Jul 2 06:07:54 2019 UTC

… …
100 * not only is this totally undocumented, not even the source code 100 * not only is this totally undocumented, not even the source code
101 * can tell you what the future semantics of compat_features and 101 * can tell you what the future semantics of compat_features and
102 * incompat_features are, or what header_length actually is for. 102 * incompat_features are, or what header_length actually is for.
103 */ 103 */
104#define AIO_RING_MAGIC 0xa10a10a1 104#define AIO_RING_MAGIC 0xa10a10a1
105#define AIO_RING_INCOMPAT_FEATURES 0 105#define EV_AIO_RING_INCOMPAT_FEATURES 0
106struct aio_ring 106struct aio_ring
107{ 107{
108 unsigned id; /* kernel internal index number */ 108 unsigned id; /* kernel internal index number */
109 unsigned nr; /* number of io_events */ 109 unsigned nr; /* number of io_events */
110 unsigned head; /* Written to by userland or by kernel. */ 110 unsigned head; /* Written to by userland or by kernel. */
… …
120 120
121inline_size 121inline_size
122int 122int
123evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 123evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
124{ 124{
125 return syscall (SYS_io_setup, nr_events, ctx_idp); 125 return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp);
126} 126}
127 127
128inline_size 128inline_size
129int 129int
130evsys_io_destroy (aio_context_t ctx_id) 130evsys_io_destroy (aio_context_t ctx_id)
131{ 131{
132 return syscall (SYS_io_destroy, ctx_id); 132 return ev_syscall1 (SYS_io_destroy, ctx_id);
133} 133}
134 134
135inline_size 135inline_size
136int 136int
137evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) 137evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
138{ 138{
139 return syscall (SYS_io_submit, ctx_id, nr, cbp); 139 return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp);
140} 140}
141 141
142inline_size 142inline_size
143int 143int
144evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) 144evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result)
145{ 145{
146 return syscall (SYS_io_cancel, ctx_id, cbp, result); 146 return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result);
147} 147}
148 148
149inline_size 149inline_size
150int 150int
151evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) 151evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout)
152{ 152{
153 return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); 153 return ev_syscall5 (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
154} 154}
155 155
156/*****************************************************************************/ 156/*****************************************************************************/
157/* actual backed implementation */ 157/* actual backed implementation */
158 158
… …
235linuxaio_modify (EV_P_ int fd, int oev, int nev) 235linuxaio_modify (EV_P_ int fd, int oev, int nev)
236{ 236{
237 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); 237 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
238 ANIOCBP iocb = linuxaio_iocbps [fd]; 238 ANIOCBP iocb = linuxaio_iocbps [fd];
239 239
240 if (iocb->io.aio_reqprio < 0) 240 if (ecb_expect_false (iocb->io.aio_reqprio < 0))
241 { 241 {
242 /* we handed this fd over to epoll, so undo this first */ 242 /* we handed this fd over to epoll, so undo this first */
243 /* we do it manually because the optimisations on epoll_modfy won't do us any good */ 243 /* we do it manually because the optimisations on epoll_modify won't do us any good */
244 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); 244 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
245 anfds [fd].emask = 0;
245 iocb->io.aio_reqprio = 0; 246 iocb->io.aio_reqprio = 0;
246 } 247 }
247 248
249 if (ecb_expect_false (iocb->io.aio_buf))
250 {
251 /* iocb active, so cancel it first before resubmit */
252 for (;;)
253 {
254 /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */
255 if (ecb_expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0))
256 break;
257
258 if (ecb_expect_true (errno == EINPROGRESS))
259 break;
260
261 /* the EINPROGRESS test is for nicer error message. clumsy. */
262 assert (("libev: linuxaio unexpected io_cancel failed", errno != EINPROGRESS && errno != EINTR));
263 }
264 }
265
248 if (iocb->io.aio_buf) 266 iocb->io.aio_buf =
249 /* io_cancel always returns some error on relevant kernels, but works */ 267 (nev & EV_READ ? POLLIN : 0)
250 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); 268 | (nev & EV_WRITE ? POLLOUT : 0);
251 269
252 if (nev) 270 if (nev)
253 { 271 {
254 iocb->io.aio_buf =
255 (nev & EV_READ ? POLLIN : 0)
256 | (nev & EV_WRITE ? POLLOUT : 0);
257
258 /* queue iocb up for io_submit */ 272 /* queue iocb up for io_submit */
259 /* this assumes we only ever get one call per fd per loop iteration */ 273 /* this assumes we only ever get one call per fd per loop iteration */
260 ++linuxaio_submitcnt; 274 ++linuxaio_submitcnt;
261 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); 275 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
262 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 276 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
… …
267linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) 281linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
268{ 282{
269 epoll_poll (EV_A_ 0); 283 epoll_poll (EV_A_ 0);
270} 284}
271 285
272static void 286inline_speed
287void
273linuxaio_fd_rearm (EV_P_ int fd) 288linuxaio_fd_rearm (EV_P_ int fd)
274{ 289{
275 anfds [fd].events = 0; 290 anfds [fd].events = 0;
276 linuxaio_iocbps [fd]->io.aio_buf = 0; 291 linuxaio_iocbps [fd]->io.aio_buf = 0;
277 fd_change (EV_A_ fd, EV_ANFD_REIFY); 292 fd_change (EV_A_ fd, EV_ANFD_REIFY);
… …
293 fd, 308 fd,
294 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 309 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
295 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 310 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
296 ); 311 );
297 312
298 /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */ 313 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
299 linuxaio_fd_rearm (EV_A_ fd); 314 linuxaio_fd_rearm (EV_A_ fd);
300 315
301 --nr; 316 --nr;
302 ++ev; 317 ++ev;
303 } 318 }
… …
306/* get any events from ring buffer, return true if any were handled */ 321/* get any events from ring buffer, return true if any were handled */
307static int 322static int
308linuxaio_get_events_from_ring (EV_P) 323linuxaio_get_events_from_ring (EV_P)
309{ 324{
310 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 325 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
326 unsigned head, tail;
311 327
312 /* the kernel reads and writes both of these variables, */ 328 /* the kernel reads and writes both of these variables, */
313 /* as a C extension, we assume that volatile use here */ 329 /* as a C extension, we assume that volatile use here */
314 /* both makes reads atomic and once-only */ 330 /* both makes reads atomic and once-only */
315 unsigned head = *(volatile unsigned *)&ring->head; 331 head = *(volatile unsigned *)&ring->head;
332 ECB_MEMORY_FENCE_ACQUIRE;
316 unsigned tail = *(volatile unsigned *)&ring->tail; 333 tail = *(volatile unsigned *)&ring->tail;
317 334
318 if (head == tail) 335 if (head == tail)
319 return 0; 336 return 0;
320
321 /* bail out if the ring buffer doesn't match the expected layout */
322 if (expect_false (ring->magic != AIO_RING_MAGIC)
323 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
324 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
325 return 0;
326
327 /* make sure the events up to tail are visible */
328 ECB_MEMORY_FENCE_ACQUIRE;
329 337
330 /* parse all available events, but only once, to avoid starvation */ 338 /* parse all available events, but only once, to avoid starvation */
331 if (tail > head) /* normal case around */ 339 if (tail > head) /* normal case around */
332 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 340 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
333 else /* wrapped around */ 341 else /* wrapped around */
… …
337 } 345 }
338 346
339 ECB_MEMORY_FENCE_RELEASE; 347 ECB_MEMORY_FENCE_RELEASE;
340 /* as an extension to C, we hope that the volatile will make this atomic and once-only */ 348 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
341 *(volatile unsigned *)&ring->head = tail; 349 *(volatile unsigned *)&ring->head = tail;
342 /* make sure kernel can see our new head value - probably not required */
343 350
344 return 1; 351 return 1;
352}
353
354inline_size
355int
356linuxaio_ringbuf_valid (EV_P)
357{
358 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
359
360 return ecb_expect_true (ring->magic == AIO_RING_MAGIC)
361 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES
362 && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */
345} 363}
346 364
347/* read at least one event from kernel, or timeout */ 365/* read at least one event from kernel, or timeout */
348inline_size 366inline_size
349void 367void
350linuxaio_get_events (EV_P_ ev_tstamp timeout) 368linuxaio_get_events (EV_P_ ev_tstamp timeout)
351{ 369{
352 struct timespec ts; 370 struct timespec ts;
353 struct io_event ioev[1]; 371 struct io_event ioev[8]; /* 256 octet stack space */
354 int res; 372 int want = 1; /* how many events to request */
373 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
355 374
375 if (ecb_expect_true (ringbuf_valid))
376 {
377 /* if the ring buffer has any events, we don't wait or call the kernel at all */
356 if (linuxaio_get_events_from_ring (EV_A)) 378 if (linuxaio_get_events_from_ring (EV_A))
357 return; 379 return;
358 380
359 /* no events, so wait for at least one, then poll ring buffer again */ 381 /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */
360 /* this degrades to one event per loop iteration */ 382 if (!timeout)
361 /* if the ring buffer changes layout, but so be it */ 383 return;
384 }
385 else
386 /* no ringbuffer, request slightly larger batch */
387 want = sizeof (ioev) / sizeof (ioev [0]);
362 388
389 /* no events, so wait for some
390 * for fairness reasons, we do this in a loop, to fetch all events
391 */
392 for (;;)
393 {
394 int res;
395
363 EV_RELEASE_CB; 396 EV_RELEASE_CB;
364 397
365 ts.tv_sec = (long)timeout; 398 EV_TS_SET (ts, timeout);
366 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 399 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
367 400
368 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
369
370 EV_ACQUIRE_CB; 401 EV_ACQUIRE_CB;
371 402
372 if (res < 0) 403 if (res < 0)
373 if (errno == EINTR) 404 if (errno == EINTR)
374 /* ignored */; 405 /* ignored, retry */;
375 else 406 else
376 ev_syserr ("(libev) linuxaio io_getevents"); 407 ev_syserr ("(libev) linuxaio io_getevents");
377 else if (res) 408 else if (res)
378 { 409 {
379 /* at least one event received, handle it and any remaining ones in the ring buffer */ 410 /* at least one event available, handle them */
380 linuxaio_parse_events (EV_A_ ioev, res); 411 linuxaio_parse_events (EV_A_ ioev, res);
412
413 if (ecb_expect_true (ringbuf_valid))
414 {
415 /* if we have a ring buffer, handle any remaining events in it */
381 linuxaio_get_events_from_ring (EV_A); 416 linuxaio_get_events_from_ring (EV_A);
382 }
383}
384 417
385static int 418 /* at this point, we should have handled all outstanding events */
419 break;
420 }
421 else if (res < want)
422 /* otherwise, if there were fewere events than we wanted, we assume there are no more */
423 break;
424 }
425 else
426 break; /* no events from the kernel, we are done */
427
428 timeout = 0; /* only wait in the first iteration */
429 }
430}
431
432inline_size
433int
386linuxaio_io_setup (EV_P) 434linuxaio_io_setup (EV_P)
387{ 435{
388 linuxaio_ctx = 0; 436 linuxaio_ctx = 0;
389 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx); 437 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
390} 438}
… …
401 /* which allows us to pinpoint the erroneous iocb */ 449 /* which allows us to pinpoint the erroneous iocb */
402 for (submitted = 0; submitted < linuxaio_submitcnt; ) 450 for (submitted = 0; submitted < linuxaio_submitcnt; )
403 { 451 {
404 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); 452 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
405 453
406 if (expect_false (res < 0)) 454 if (ecb_expect_false (res < 0))
407 if (errno == EINVAL) 455 if (errno == EINVAL)
408 { 456 {
409 /* This happens for unsupported fds, officially, but in my testing, 457 /* This happens for unsupported fds, officially, but in my testing,
410 * also randomly happens for supported fds. We fall back to good old 458 * also randomly happens for supported fds. We fall back to good old
411 * poll() here, under the assumption that this is a very rare case. 459 * poll() here, under the assumption that this is a very rare case.
… …
441 } 489 }
442 490
443 ++linuxaio_iteration; 491 ++linuxaio_iteration;
444 if (linuxaio_io_setup (EV_A) < 0) 492 if (linuxaio_io_setup (EV_A) < 0)
445 { 493 {
494 /* TODO: rearm all and recreate epoll backend from scratch */
495 /* TODO: might be more prudent? */
496
446 /* to bad, we can't get a new aio context, go 100% epoll */ 497 /* to bad, we can't get a new aio context, go 100% epoll */
447 linuxaio_free_iocbp (EV_A); 498 linuxaio_free_iocbp (EV_A);
448 ev_io_stop (EV_A_ &linuxaio_epoll_w); 499 ev_io_stop (EV_A_ &linuxaio_epoll_w);
449 ev_ref (EV_A); 500 ev_ref (EV_A);
450 linuxaio_ctx = 0; 501 linuxaio_ctx = 0;
502
503 backend = EVBACKEND_EPOLL;
451 backend_modify = epoll_modify; 504 backend_modify = epoll_modify;
452 backend_poll = epoll_poll; 505 backend_poll = epoll_poll;
453 } 506 }
454 507
455 timeout = 0; 508 timeout = 0;
456 /* it's easiest to handle this mess in another iteration */ 509 /* it's easiest to handle this mess in another iteration */
457 return; 510 return;
458 } 511 }
459 else if (errno == EBADF) 512 else if (errno == EBADF)
460 { 513 {
514 assert (("libev: event loop rejected bad fd", errno != EBADF));
461 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); 515 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
462 516
463 res = 1; /* skip this iocb */ 517 res = 1; /* skip this iocb */
464 } 518 }
519 else if (errno == EINTR) /* not seen in reality, not documented */
520 res = 0; /* silently ignore and retry */
465 else 521 else
522 {
466 ev_syserr ("(libev) linuxaio io_submit"); 523 ev_syserr ("(libev) linuxaio io_submit");
524 res = 0;
525 }
467 526
468 submitted += res; 527 submitted += res;
469 } 528 }
470 529
471 linuxaio_submitcnt = 0; 530 linuxaio_submitcnt = 0;
… …
499 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); 558 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
500 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); 559 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
501 ev_io_start (EV_A_ &linuxaio_epoll_w); 560 ev_io_start (EV_A_ &linuxaio_epoll_w);
502 ev_unref (EV_A); /* watcher should not keep loop alive */ 561 ev_unref (EV_A); /* watcher should not keep loop alive */
503 562
504 backend_modify = linuxaio_modify; 563 backend_modify = linuxaio_modify;
505 backend_poll = linuxaio_poll; 564 backend_poll = linuxaio_poll;
506 565
507 linuxaio_iocbpmax = 0; 566 linuxaio_iocbpmax = 0;
508 linuxaio_iocbps = 0; 567 linuxaio_iocbps = 0;
509 568
510 linuxaio_submits = 0; 569 linuxaio_submits = 0;
… …
518void 577void
519linuxaio_destroy (EV_P) 578linuxaio_destroy (EV_P)
520{ 579{
521 epoll_destroy (EV_A); 580 epoll_destroy (EV_A);
522 linuxaio_free_iocbp (EV_A); 581 linuxaio_free_iocbp (EV_A);
523 evsys_io_destroy (linuxaio_ctx); 582 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
524} 583}
525 584
526inline_size 585ecb_cold
527void 586static void
528linuxaio_fork (EV_P) 587linuxaio_fork (EV_P)
529{ 588{
530 /* this frees all iocbs, which is very heavy-handed */ 589 /* this frees all iocbs, which is very heavy-handed */
531 linuxaio_destroy (EV_A); 590 linuxaio_destroy (EV_A);
532 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 591 linuxaio_submitcnt = 0; /* all pointers were invalidated */
… …
534 linuxaio_iteration = 0; /* we start over in the child */ 593 linuxaio_iteration = 0; /* we start over in the child */
535 594
536 while (linuxaio_io_setup (EV_A) < 0) 595 while (linuxaio_io_setup (EV_A) < 0)
537 ev_syserr ("(libev) linuxaio io_setup"); 596 ev_syserr ("(libev) linuxaio io_setup");
538 597
598 /* forking epoll should also effectively unregister all fds from the backend */
539 epoll_fork (EV_A); 599 epoll_fork (EV_A);
600 /* epoll_fork already did this. hopefully */
601 /*fd_rearm_all (EV_A);*/
540 602
541 ev_io_stop (EV_A_ &linuxaio_epoll_w); 603 ev_io_stop (EV_A_ &linuxaio_epoll_w);
542 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); 604 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
543 ev_io_start (EV_A_ &linuxaio_epoll_w); 605 ev_io_start (EV_A_ &linuxaio_epoll_w);
544
545 /* epoll_fork already did this. hopefully */
546 /*fd_rearm_all (EV_A);*/
547} 606}
548 607

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines