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 1.32 by root, Tue Jun 25 18:07:44 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. */
… …
115 unsigned incompat_features; 115 unsigned incompat_features;
116 unsigned header_length; /* size of aio_ring */ 116 unsigned header_length; /* size of aio_ring */
117 117
118 struct io_event io_events[0]; 118 struct io_event io_events[0];
119}; 119};
120
121/*
122 * define some syscall wrappers for common architectures
123 * this is mostly for nice looks during debugging, not performance.
124 * our syscalls return < 0, not == -1, on error. which is good
125 * enough for linux aio.
126 * TODO: arm is also common nowadays, maybe even mips and x86
127 * TODO: after implementing this, it suddenly looks like overkill, but its hard to remove...
128 */
129#if __GNUC__ && __linux && ECB_AMD64 && !defined __OPTIMIZE_SIZE__
130 /* the costly errno access probably kills this for size optimisation */
131
132 #define ev_syscall(nr,narg,arg1,arg2,arg3,arg4,arg5) \
133 ({ \
134 long res; \
135 register unsigned long r5 __asm__ ("r8" ); \
136 register unsigned long r4 __asm__ ("r10"); \
137 register unsigned long r3 __asm__ ("rdx"); \
138 register unsigned long r2 __asm__ ("rsi"); \
139 register unsigned long r1 __asm__ ("rdi"); \
140 if (narg >= 5) r5 = (unsigned long)(arg5); \
141 if (narg >= 4) r4 = (unsigned long)(arg4); \
142 if (narg >= 3) r3 = (unsigned long)(arg3); \
143 if (narg >= 2) r2 = (unsigned long)(arg2); \
144 if (narg >= 1) r1 = (unsigned long)(arg1); \
145 __asm__ __volatile__ ( \
146 "syscall\n\t" \
147 : "=a" (res) \
148 : "0" (nr), "r" (r1), "r" (r2), "r" (r3), "r" (r4), "r" (r5) \
149 : "cc", "r11", "cx", "memory"); \
150 errno = -res; \
151 res; \
152 })
153
154#endif
155
156#ifdef ev_syscall
157 #define ev_syscall0(nr) ev_syscall (nr, 0, 0, 0, 0, 0, 0
158 #define ev_syscall1(nr,arg1) ev_syscall (nr, 1, arg1, 0, 0, 0, 0)
159 #define ev_syscall2(nr,arg1,arg2) ev_syscall (nr, 2, arg1, arg2, 0, 0, 0)
160 #define ev_syscall3(nr,arg1,arg2,arg3) ev_syscall (nr, 3, arg1, arg2, arg3, 0, 0)
161 #define ev_syscall4(nr,arg1,arg2,arg3,arg4) ev_syscall (nr, 3, arg1, arg2, arg3, arg4, 0)
162 #define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) ev_syscall (nr, 5, arg1, arg2, arg3, arg4, arg5)
163#else
164 #define ev_syscall0(nr) syscall (nr)
165 #define ev_syscall1(nr,arg1) syscall (nr, arg1)
166 #define ev_syscall2(nr,arg1,arg2) syscall (nr, arg1, arg2)
167 #define ev_syscall3(nr,arg1,arg2,arg3) syscall (nr, arg1, arg2, arg3)
168 #define ev_syscall4(nr,arg1,arg2,arg3,arg4) syscall (nr, arg1, arg2, arg3, arg4)
169 #define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) syscall (nr, arg1, arg2, arg3, arg4, arg5)
170#endif
171 120
172inline_size 121inline_size
173int 122int
174evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) 123evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
175{ 124{
… …
286linuxaio_modify (EV_P_ int fd, int oev, int nev) 235linuxaio_modify (EV_P_ int fd, int oev, int nev)
287{ 236{
288 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);
289 ANIOCBP iocb = linuxaio_iocbps [fd]; 238 ANIOCBP iocb = linuxaio_iocbps [fd];
290 239
291 if (iocb->io.aio_reqprio < 0) 240 if (ecb_expect_false (iocb->io.aio_reqprio < 0))
292 { 241 {
293 /* we handed this fd over to epoll, so undo this first */ 242 /* we handed this fd over to epoll, so undo this first */
294 /* 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 */
295 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); 244 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
296 anfds [fd].emask = 0; 245 anfds [fd].emask = 0;
297 iocb->io.aio_reqprio = 0; 246 iocb->io.aio_reqprio = 0;
298 } 247 }
299 248
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
300 if (iocb->io.aio_buf) 266 iocb->io.aio_buf =
301 /* io_cancel always returns some error on relevant kernels, but works */ 267 (nev & EV_READ ? POLLIN : 0)
302 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); 268 | (nev & EV_WRITE ? POLLOUT : 0);
303 269
304 if (nev) 270 if (nev)
305 { 271 {
306 iocb->io.aio_buf =
307 (nev & EV_READ ? POLLIN : 0)
308 | (nev & EV_WRITE ? POLLOUT : 0);
309
310 /* queue iocb up for io_submit */ 272 /* queue iocb up for io_submit */
311 /* 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 */
312 ++linuxaio_submitcnt; 274 ++linuxaio_submitcnt;
313 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);
314 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; 276 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
… …
319linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) 281linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
320{ 282{
321 epoll_poll (EV_A_ 0); 283 epoll_poll (EV_A_ 0);
322} 284}
323 285
324static void 286inline_speed
287void
325linuxaio_fd_rearm (EV_P_ int fd) 288linuxaio_fd_rearm (EV_P_ int fd)
326{ 289{
327 anfds [fd].events = 0; 290 anfds [fd].events = 0;
328 linuxaio_iocbps [fd]->io.aio_buf = 0; 291 linuxaio_iocbps [fd]->io.aio_buf = 0;
329 fd_change (EV_A_ fd, EV_ANFD_REIFY); 292 fd_change (EV_A_ fd, EV_ANFD_REIFY);
… …
345 fd, 308 fd,
346 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) 309 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
347 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) 310 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
348 ); 311 );
349 312
350 /* 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 */
351 linuxaio_fd_rearm (EV_A_ fd); 314 linuxaio_fd_rearm (EV_A_ fd);
352 315
353 --nr; 316 --nr;
354 ++ev; 317 ++ev;
355 } 318 }
… …
358/* 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 */
359static int 322static int
360linuxaio_get_events_from_ring (EV_P) 323linuxaio_get_events_from_ring (EV_P)
361{ 324{
362 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; 325 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
326 unsigned head, tail;
363 327
364 /* the kernel reads and writes both of these variables, */ 328 /* the kernel reads and writes both of these variables, */
365 /* as a C extension, we assume that volatile use here */ 329 /* as a C extension, we assume that volatile use here */
366 /* both makes reads atomic and once-only */ 330 /* both makes reads atomic and once-only */
367 unsigned head = *(volatile unsigned *)&ring->head; 331 head = *(volatile unsigned *)&ring->head;
332 ECB_MEMORY_FENCE_ACQUIRE;
368 unsigned tail = *(volatile unsigned *)&ring->tail; 333 tail = *(volatile unsigned *)&ring->tail;
369 334
370 if (head == tail) 335 if (head == tail)
371 return 0; 336 return 0;
372
373 /* bail out if the ring buffer doesn't match the expected layout */
374 if (expect_false (ring->magic != AIO_RING_MAGIC)
375 || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
376 || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
377 return 0;
378
379 /* make sure the events up to tail are visible */
380 ECB_MEMORY_FENCE_ACQUIRE;
381 337
382 /* parse all available events, but only once, to avoid starvation */ 338 /* parse all available events, but only once, to avoid starvation */
383 if (tail > head) /* normal case around */ 339 if (tail > head) /* normal case around */
384 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); 340 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
385 else /* wrapped around */ 341 else /* wrapped around */
… …
393 *(volatile unsigned *)&ring->head = tail; 349 *(volatile unsigned *)&ring->head = tail;
394 350
395 return 1; 351 return 1;
396} 352}
397 353
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]? */
363}
364
398/* read at least one event from kernel, or timeout */ 365/* read at least one event from kernel, or timeout */
399inline_size 366inline_size
400void 367void
401linuxaio_get_events (EV_P_ ev_tstamp timeout) 368linuxaio_get_events (EV_P_ ev_tstamp timeout)
402{ 369{
403 struct timespec ts; 370 struct timespec ts;
404 struct io_event ioev[1]; 371 struct io_event ioev[8]; /* 256 octet stack space */
405 int res; 372 int want = 1; /* how many events to request */
373 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
406 374
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 */
407 if (linuxaio_get_events_from_ring (EV_A)) 378 if (linuxaio_get_events_from_ring (EV_A))
408 return; 379 return;
409 380
410 /* 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 */
411 /* this degrades to one event per loop iteration */ 382 if (!timeout)
412 /* 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]);
413 388
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
414 EV_RELEASE_CB; 396 EV_RELEASE_CB;
415 397
416 ts.tv_sec = (long)timeout; 398 EV_TS_SET (ts, timeout);
417 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); 399 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
418 400
419 res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
420
421 EV_ACQUIRE_CB; 401 EV_ACQUIRE_CB;
422 402
423 if (res < 0) 403 if (res < 0)
424 if (errno == EINTR) 404 if (errno == EINTR)
425 /* ignored */; 405 /* ignored, retry */;
426 else 406 else
427 ev_syserr ("(libev) linuxaio io_getevents"); 407 ev_syserr ("(libev) linuxaio io_getevents");
428 else if (res) 408 else if (res)
429 { 409 {
430 /* at least one event available, handle it and any remaining ones in the ring buffer */ 410 /* at least one event available, handle them */
431 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 */
432 linuxaio_get_events_from_ring (EV_A); 416 linuxaio_get_events_from_ring (EV_A);
417
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 */
433 } 429 }
434} 430}
435 431
436inline_size 432inline_size
437int 433int
… …
453 /* which allows us to pinpoint the erroneous iocb */ 449 /* which allows us to pinpoint the erroneous iocb */
454 for (submitted = 0; submitted < linuxaio_submitcnt; ) 450 for (submitted = 0; submitted < linuxaio_submitcnt; )
455 { 451 {
456 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);
457 453
458 if (expect_false (res < 0)) 454 if (ecb_expect_false (res < 0))
459 if (errno == EINVAL) 455 if (errno == EINVAL)
460 { 456 {
461 /* This happens for unsupported fds, officially, but in my testing, 457 /* This happens for unsupported fds, officially, but in my testing,
462 * 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
463 * 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.
… …
493 } 489 }
494 490
495 ++linuxaio_iteration; 491 ++linuxaio_iteration;
496 if (linuxaio_io_setup (EV_A) < 0) 492 if (linuxaio_io_setup (EV_A) < 0)
497 { 493 {
494 /* TODO: rearm all and recreate epoll backend from scratch */
495 /* TODO: might be more prudent? */
496
498 /* 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 */
499 linuxaio_free_iocbp (EV_A); 498 linuxaio_free_iocbp (EV_A);
500 ev_io_stop (EV_A_ &linuxaio_epoll_w); 499 ev_io_stop (EV_A_ &linuxaio_epoll_w);
501 ev_ref (EV_A); 500 ev_ref (EV_A);
502 linuxaio_ctx = 0; 501 linuxaio_ctx = 0;
502
503 backend = EVBACKEND_EPOLL;
503 backend_modify = epoll_modify; 504 backend_modify = epoll_modify;
504 backend_poll = epoll_poll; 505 backend_poll = epoll_poll;
505 } 506 }
506 507
507 timeout = 0; 508 timeout = 0;
508 /* it's easiest to handle this mess in another iteration */ 509 /* it's easiest to handle this mess in another iteration */
509 return; 510 return;
510 } 511 }
511 else if (errno == EBADF) 512 else if (errno == EBADF)
512 { 513 {
514 assert (("libev: event loop rejected bad fd", errno != EBADF));
513 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); 515 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
514 516
515 res = 1; /* skip this iocb */ 517 res = 1; /* skip this iocb */
516 } 518 }
519 else if (errno == EINTR) /* not seen in reality, not documented */
520 res = 0; /* silently ignore and retry */
517 else 521 else
522 {
518 ev_syserr ("(libev) linuxaio io_submit"); 523 ev_syserr ("(libev) linuxaio io_submit");
524 res = 0;
525 }
519 526
520 submitted += res; 527 submitted += res;
521 } 528 }
522 529
523 linuxaio_submitcnt = 0; 530 linuxaio_submitcnt = 0;
… …
551 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);
552 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); 559 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
553 ev_io_start (EV_A_ &linuxaio_epoll_w); 560 ev_io_start (EV_A_ &linuxaio_epoll_w);
554 ev_unref (EV_A); /* watcher should not keep loop alive */ 561 ev_unref (EV_A); /* watcher should not keep loop alive */
555 562
556 backend_modify = linuxaio_modify; 563 backend_modify = linuxaio_modify;
557 backend_poll = linuxaio_poll; 564 backend_poll = linuxaio_poll;
558 565
559 linuxaio_iocbpmax = 0; 566 linuxaio_iocbpmax = 0;
560 linuxaio_iocbps = 0; 567 linuxaio_iocbps = 0;
561 568
562 linuxaio_submits = 0; 569 linuxaio_submits = 0;
… …
570void 577void
571linuxaio_destroy (EV_P) 578linuxaio_destroy (EV_P)
572{ 579{
573 epoll_destroy (EV_A); 580 epoll_destroy (EV_A);
574 linuxaio_free_iocbp (EV_A); 581 linuxaio_free_iocbp (EV_A);
575 evsys_io_destroy (linuxaio_ctx); 582 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
576} 583}
577 584
578inline_size 585ecb_cold
579void 586static void
580linuxaio_fork (EV_P) 587linuxaio_fork (EV_P)
581{ 588{
582 /* this frees all iocbs, which is very heavy-handed */ 589 /* this frees all iocbs, which is very heavy-handed */
583 linuxaio_destroy (EV_A); 590 linuxaio_destroy (EV_A);
584 linuxaio_submitcnt = 0; /* all pointers were invalidated */ 591 linuxaio_submitcnt = 0; /* all pointers were invalidated */
… …
586 linuxaio_iteration = 0; /* we start over in the child */ 593 linuxaio_iteration = 0; /* we start over in the child */
587 594
588 while (linuxaio_io_setup (EV_A) < 0) 595 while (linuxaio_io_setup (EV_A) < 0)
589 ev_syserr ("(libev) linuxaio io_setup"); 596 ev_syserr ("(libev) linuxaio io_setup");
590 597
598 /* forking epoll should also effectively unregister all fds from the backend */
591 epoll_fork (EV_A); 599 epoll_fork (EV_A);
600 /* epoll_fork already did this. hopefully */
601 /*fd_rearm_all (EV_A);*/
592 602
593 ev_io_stop (EV_A_ &linuxaio_epoll_w); 603 ev_io_stop (EV_A_ &linuxaio_epoll_w);
594 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);
595 ev_io_start (EV_A_ &linuxaio_epoll_w); 605 ev_io_start (EV_A_ &linuxaio_epoll_w);
596
597 /* epoll_fork already did this. hopefully */
598 /*fd_rearm_all (EV_A);*/
599} 606}
600 607

Diff Legend

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