ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libev/ev_linuxaio.c
Revision: 1.52
Committed: Wed Aug 28 10:05:23 2019 UTC (4 years, 10 months ago) by root
Content type: text/plain
Branch: MAIN
CVS Tags: rel-4_31, EV-rel-4_28, EV-rel-4_29, EV-rel-4_31, EV-rel-4_30
Changes since 1.51: +1 -1 lines
Log Message:
*** empty log message ***

File Contents

# Content
1 /*
2 * libev linux aio fd activity backend
3 *
4 * Copyright (c) 2019 Marc Alexander Lehmann <libev@schmorp.de>
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without modifica-
8 * tion, are permitted provided that the following conditions are met:
9 *
10 * 1. Redistributions of source code must retain the above copyright notice,
11 * this list of conditions and the following disclaimer.
12 *
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
18 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER-
19 * CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO
20 * EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE-
21 * CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
22 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
23 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
24 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH-
25 * ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
26 * OF THE POSSIBILITY OF SUCH DAMAGE.
27 *
28 * Alternatively, the contents of this file may be used under the terms of
29 * the GNU General Public License ("GPL") version 2 or any later version,
30 * in which case the provisions of the GPL are applicable instead of
31 * the above. If you wish to allow the use of your version of this file
32 * only under the terms of the GPL and not to allow others to use your
33 * version of this file under the BSD license, indicate your decision
34 * by deleting the provisions above and replace them with the notice
35 * and other provisions required by the GPL. If you do not delete the
36 * provisions above, a recipient may use your version of this file under
37 * either the BSD or the GPL.
38 */
39
40 /*
41 * general notes about linux aio:
42 *
43 * a) at first, the linux aio IOCB_CMD_POLL functionality introduced in
44 * 4.18 looks too good to be true: both watchers and events can be
45 * batched, and events can even be handled in userspace using
46 * a ring buffer shared with the kernel. watchers can be canceled
47 * regardless of whether the fd has been closed. no problems with fork.
48 * ok, the ring buffer is 200% undocumented (there isn't even a
49 * header file), but otherwise, it's pure bliss!
50 * b) ok, watchers are one-shot, so you have to re-arm active ones
51 * on every iteration. so much for syscall-less event handling,
52 * but at least these re-arms can be batched, no big deal, right?
53 * c) well, linux as usual: the documentation lies to you: io_submit
54 * sometimes returns EINVAL because the kernel doesn't feel like
55 * handling your poll mask - ttys can be polled for POLLOUT,
56 * POLLOUT|POLLIN, but polling for POLLIN fails. just great,
57 * so we have to fall back to something else (hello, epoll),
58 * but at least the fallback can be slow, because these are
59 * exceptional cases, right?
60 * d) hmm, you have to tell the kernel the maximum number of watchers
61 * you want to queue when initialising the aio context. but of
62 * course the real limit is magically calculated in the kernel, and
63 * is often higher then we asked for. so we just have to destroy
64 * the aio context and re-create it a bit larger if we hit the limit.
65 * (starts to remind you of epoll? well, it's a bit more deterministic
66 * and less gambling, but still ugly as hell).
67 * e) that's when you find out you can also hit an arbitrary system-wide
68 * limit. or the kernel simply doesn't want to handle your watchers.
69 * what the fuck do we do then? you guessed it, in the middle
70 * of event handling we have to switch to 100% epoll polling. and
71 * that better is as fast as normal epoll polling, so you practically
72 * have to use the normal epoll backend with all its quirks.
73 * f) end result of this train wreck: it inherits all the disadvantages
74 * from epoll, while adding a number on its own. why even bother to use
75 * it? because if conditions are right and your fds are supported and you
76 * don't hit a limit, this backend is actually faster, doesn't gamble with
77 * your fds, batches watchers and events and doesn't require costly state
78 * recreates. well, until it does.
79 * g) all of this makes this backend use almost twice as much code as epoll.
80 * which in turn uses twice as much code as poll. and that#s not counting
81 * the fact that this backend also depends on the epoll backend, making
82 * it three times as much code as poll, or kqueue.
83 * h) bleah. why can't linux just do kqueue. sure kqueue is ugly, but by now
84 * it's clear that whatever linux comes up with is far, far, far worse.
85 */
86
87 #include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */
88 #include <poll.h>
89 #include <linux/aio_abi.h>
90
91 /*****************************************************************************/
92 /* syscall wrapdadoop - this section has the raw api/abi definitions */
93
94 #include <sys/syscall.h> /* no glibc wrappers */
95
96 /* aio_abi.h is not versioned in any way, so we cannot test for its existance */
97 #define IOCB_CMD_POLL 5
98
99 /* taken from linux/fs/aio.c. yup, that's a .c file.
100 * not only is this totally undocumented, not even the source code
101 * can tell you what the future semantics of compat_features and
102 * incompat_features are, or what header_length actually is for.
103 */
104 #define AIO_RING_MAGIC 0xa10a10a1
105 #define EV_AIO_RING_INCOMPAT_FEATURES 0
106 struct aio_ring
107 {
108 unsigned id; /* kernel internal index number */
109 unsigned nr; /* number of io_events */
110 unsigned head; /* Written to by userland or by kernel. */
111 unsigned tail;
112
113 unsigned magic;
114 unsigned compat_features;
115 unsigned incompat_features;
116 unsigned header_length; /* size of aio_ring */
117
118 struct io_event io_events[0];
119 };
120
121 inline_size
122 int
123 evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
124 {
125 return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp);
126 }
127
128 inline_size
129 int
130 evsys_io_destroy (aio_context_t ctx_id)
131 {
132 return ev_syscall1 (SYS_io_destroy, ctx_id);
133 }
134
135 inline_size
136 int
137 evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
138 {
139 return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp);
140 }
141
142 inline_size
143 int
144 evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result)
145 {
146 return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result);
147 }
148
149 inline_size
150 int
151 evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout)
152 {
153 return ev_syscall5 (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
154 }
155
156 /*****************************************************************************/
157 /* actual backed implementation */
158
159 ecb_cold
160 static int
161 linuxaio_nr_events (EV_P)
162 {
163 /* we start with 16 iocbs and incraese from there
164 * that's tiny, but the kernel has a rather low system-wide
165 * limit that can be reached quickly, so let's be parsimonious
166 * with this resource.
167 * Rest assured, the kernel generously rounds up small and big numbers
168 * in different ways (but doesn't seem to charge you for it).
169 * The 15 here is because the kernel usually has a power of two as aio-max-nr,
170 * and this helps to take advantage of that limit.
171 */
172
173 /* we try to fill 4kB pages exactly.
174 * the ring buffer header is 32 bytes, every io event is 32 bytes.
175 * the kernel takes the io requests number, doubles it, adds 2
176 * and adds the ring buffer.
177 * the way we use this is by starting low, and then roughly doubling the
178 * size each time we hit a limit.
179 */
180
181 int requests = 15 << linuxaio_iteration;
182 int one_page = (4096
183 / sizeof (struct io_event) ) / 2; /* how many fit into one page */
184 int first_page = ((4096 - sizeof (struct aio_ring))
185 / sizeof (struct io_event) - 2) / 2; /* how many fit into the first page */
186
187 /* if everything fits into one page, use count exactly */
188 if (requests > first_page)
189 /* otherwise, round down to full pages and add the first page */
190 requests = requests / one_page * one_page + first_page;
191
192 return requests;
193 }
194
195 /* we use out own wrapper structure in case we ever want to do something "clever" */
196 typedef struct aniocb
197 {
198 struct iocb io;
199 /*int inuse;*/
200 } *ANIOCBP;
201
202 inline_size
203 void
204 linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
205 {
206 while (count--)
207 {
208 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */
209 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
210
211 /* full zero initialise is probably not required at the moment, but
212 * this is not well documented, so we better do it.
213 */
214 memset (iocb, 0, sizeof (*iocb));
215
216 iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
217 iocb->io.aio_fildes = offset;
218
219 base [offset++] = iocb;
220 }
221 }
222
223 ecb_cold
224 static void
225 linuxaio_free_iocbp (EV_P)
226 {
227 while (linuxaio_iocbpmax--)
228 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]);
229
230 linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */
231 }
232
233 static void
234 linuxaio_modify (EV_P_ int fd, int oev, int nev)
235 {
236 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
237 ANIOCBP iocb = linuxaio_iocbps [fd];
238 ANFD *anfd = &anfds [fd];
239
240 if (ecb_expect_false (iocb->io.aio_reqprio < 0))
241 {
242 /* we handed this fd over to epoll, so undo this first */
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);
245 anfd->emask = 0;
246 iocb->io.aio_reqprio = 0;
247 }
248 else if (ecb_expect_false (iocb->io.aio_buf))
249 {
250 /* iocb active, so cancel it first before resubmit */
251 /* this assumes we only ever get one call per fd per loop iteration */
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 if (errno != EINTR)
263 {
264 assert (("libev: linuxaio unexpected io_cancel failed", errno != EINTR && errno != EINPROGRESS));
265 break;
266 }
267 }
268
269 /* increment generation counter to avoid handling old events */
270 ++anfd->egen;
271 }
272
273 iocb->io.aio_buf =
274 (nev & EV_READ ? POLLIN : 0)
275 | (nev & EV_WRITE ? POLLOUT : 0);
276
277 if (nev)
278 {
279 iocb->io.aio_data = (uint32_t)fd | ((__u64)(uint32_t)anfd->egen << 32);
280
281 /* queue iocb up for io_submit */
282 /* this assumes we only ever get one call per fd per loop iteration */
283 ++linuxaio_submitcnt;
284 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
285 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
286 }
287 }
288
289 static void
290 linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
291 {
292 epoll_poll (EV_A_ 0);
293 }
294
295 inline_speed
296 void
297 linuxaio_fd_rearm (EV_P_ int fd)
298 {
299 anfds [fd].events = 0;
300 linuxaio_iocbps [fd]->io.aio_buf = 0;
301 fd_change (EV_A_ fd, EV_ANFD_REIFY);
302 }
303
304 static void
305 linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
306 {
307 while (nr)
308 {
309 int fd = ev->data & 0xffffffff;
310 uint32_t gen = ev->data >> 32;
311 int res = ev->res;
312
313 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
314
315 /* only accept events if generation counter matches */
316 if (ecb_expect_true (gen == (uint32_t)anfds [fd].egen))
317 {
318 /* feed events, we do not expect or handle POLLNVAL */
319 fd_event (
320 EV_A_
321 fd,
322 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
323 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
324 );
325
326 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
327 linuxaio_fd_rearm (EV_A_ fd);
328 }
329
330 --nr;
331 ++ev;
332 }
333 }
334
335 /* get any events from ring buffer, return true if any were handled */
336 static int
337 linuxaio_get_events_from_ring (EV_P)
338 {
339 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
340 unsigned head, tail;
341
342 /* the kernel reads and writes both of these variables, */
343 /* as a C extension, we assume that volatile use here */
344 /* both makes reads atomic and once-only */
345 head = *(volatile unsigned *)&ring->head;
346 ECB_MEMORY_FENCE_ACQUIRE;
347 tail = *(volatile unsigned *)&ring->tail;
348
349 if (head == tail)
350 return 0;
351
352 /* parse all available events, but only once, to avoid starvation */
353 if (ecb_expect_true (tail > head)) /* normal case around */
354 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
355 else /* wrapped around */
356 {
357 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
358 linuxaio_parse_events (EV_A_ ring->io_events, tail);
359 }
360
361 ECB_MEMORY_FENCE_RELEASE;
362 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
363 *(volatile unsigned *)&ring->head = tail;
364
365 return 1;
366 }
367
368 inline_size
369 int
370 linuxaio_ringbuf_valid (EV_P)
371 {
372 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
373
374 return ecb_expect_true (ring->magic == AIO_RING_MAGIC)
375 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES
376 && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */
377 }
378
379 /* read at least one event from kernel, or timeout */
380 inline_size
381 void
382 linuxaio_get_events (EV_P_ ev_tstamp timeout)
383 {
384 struct timespec ts;
385 struct io_event ioev[8]; /* 256 octet stack space */
386 int want = 1; /* how many events to request */
387 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
388
389 if (ecb_expect_true (ringbuf_valid))
390 {
391 /* if the ring buffer has any events, we don't wait or call the kernel at all */
392 if (linuxaio_get_events_from_ring (EV_A))
393 return;
394
395 /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */
396 if (!timeout)
397 return;
398 }
399 else
400 /* no ringbuffer, request slightly larger batch */
401 want = sizeof (ioev) / sizeof (ioev [0]);
402
403 /* no events, so wait for some
404 * for fairness reasons, we do this in a loop, to fetch all events
405 */
406 for (;;)
407 {
408 int res;
409
410 EV_RELEASE_CB;
411
412 EV_TS_SET (ts, timeout);
413 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
414
415 EV_ACQUIRE_CB;
416
417 if (res < 0)
418 if (errno == EINTR)
419 /* ignored, retry */;
420 else
421 ev_syserr ("(libev) linuxaio io_getevents");
422 else if (res)
423 {
424 /* at least one event available, handle them */
425 linuxaio_parse_events (EV_A_ ioev, res);
426
427 if (ecb_expect_true (ringbuf_valid))
428 {
429 /* if we have a ring buffer, handle any remaining events in it */
430 linuxaio_get_events_from_ring (EV_A);
431
432 /* at this point, we should have handled all outstanding events */
433 break;
434 }
435 else if (res < want)
436 /* otherwise, if there were fewere events than we wanted, we assume there are no more */
437 break;
438 }
439 else
440 break; /* no events from the kernel, we are done */
441
442 timeout = EV_TS_CONST (0.); /* only wait in the first iteration */
443 }
444 }
445
446 inline_size
447 int
448 linuxaio_io_setup (EV_P)
449 {
450 linuxaio_ctx = 0;
451 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
452 }
453
454 static void
455 linuxaio_poll (EV_P_ ev_tstamp timeout)
456 {
457 int submitted;
458
459 /* first phase: submit new iocbs */
460
461 /* io_submit might return less than the requested number of iocbs */
462 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
463 /* which allows us to pinpoint the erroneous iocb */
464 for (submitted = 0; submitted < linuxaio_submitcnt; )
465 {
466 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
467
468 if (ecb_expect_false (res < 0))
469 if (errno == EINVAL)
470 {
471 /* This happens for unsupported fds, officially, but in my testing,
472 * also randomly happens for supported fds. We fall back to good old
473 * poll() here, under the assumption that this is a very rare case.
474 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
475 * discussion about such a case (ttys) where polling for POLLIN
476 * fails but POLLIN|POLLOUT works.
477 */
478 struct iocb *iocb = linuxaio_submits [submitted];
479 epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
480 iocb->aio_reqprio = -1; /* mark iocb as epoll */
481
482 res = 1; /* skip this iocb - another iocb, another chance */
483 }
484 else if (errno == EAGAIN)
485 {
486 /* This happens when the ring buffer is full, or some other shit we
487 * don't know and isn't documented. Most likely because we have too
488 * many requests and linux aio can't be assed to handle them.
489 * In this case, we try to allocate a larger ring buffer, freeing
490 * ours first. This might fail, in which case we have to fall back to 100%
491 * epoll.
492 * God, how I hate linux not getting its act together. Ever.
493 */
494 evsys_io_destroy (linuxaio_ctx);
495 linuxaio_submitcnt = 0;
496
497 /* rearm all fds with active iocbs */
498 {
499 int fd;
500 for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
501 if (linuxaio_iocbps [fd]->io.aio_buf)
502 linuxaio_fd_rearm (EV_A_ fd);
503 }
504
505 ++linuxaio_iteration;
506 if (linuxaio_io_setup (EV_A) < 0)
507 {
508 /* TODO: rearm all and recreate epoll backend from scratch */
509 /* TODO: might be more prudent? */
510
511 /* to bad, we can't get a new aio context, go 100% epoll */
512 linuxaio_free_iocbp (EV_A);
513 ev_io_stop (EV_A_ &linuxaio_epoll_w);
514 ev_ref (EV_A);
515 linuxaio_ctx = 0;
516
517 backend = EVBACKEND_EPOLL;
518 backend_modify = epoll_modify;
519 backend_poll = epoll_poll;
520 }
521
522 timeout = EV_TS_CONST (0.);
523 /* it's easiest to handle this mess in another iteration */
524 return;
525 }
526 else if (errno == EBADF)
527 {
528 assert (("libev: event loop rejected bad fd", errno != EBADF));
529 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
530
531 res = 1; /* skip this iocb */
532 }
533 else if (errno == EINTR) /* not seen in reality, not documented */
534 res = 0; /* silently ignore and retry */
535 else
536 {
537 ev_syserr ("(libev) linuxaio io_submit");
538 res = 0;
539 }
540
541 submitted += res;
542 }
543
544 linuxaio_submitcnt = 0;
545
546 /* second phase: fetch and parse events */
547
548 linuxaio_get_events (EV_A_ timeout);
549 }
550
551 inline_size
552 int
553 linuxaio_init (EV_P_ int flags)
554 {
555 /* would be great to have a nice test for IOCB_CMD_POLL instead */
556 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
557 /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
558 if (ev_linux_version () < 0x041300)
559 return 0;
560
561 if (!epoll_init (EV_A_ 0))
562 return 0;
563
564 linuxaio_iteration = 0;
565
566 if (linuxaio_io_setup (EV_A) < 0)
567 {
568 epoll_destroy (EV_A);
569 return 0;
570 }
571
572 ev_io_init (&linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
573 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
574 ev_io_start (EV_A_ &linuxaio_epoll_w);
575 ev_unref (EV_A); /* watcher should not keep loop alive */
576
577 backend_modify = linuxaio_modify;
578 backend_poll = linuxaio_poll;
579
580 linuxaio_iocbpmax = 0;
581 linuxaio_iocbps = 0;
582
583 linuxaio_submits = 0;
584 linuxaio_submitmax = 0;
585 linuxaio_submitcnt = 0;
586
587 return EVBACKEND_LINUXAIO;
588 }
589
590 inline_size
591 void
592 linuxaio_destroy (EV_P)
593 {
594 epoll_destroy (EV_A);
595 linuxaio_free_iocbp (EV_A);
596 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
597 }
598
599 ecb_cold
600 static void
601 linuxaio_fork (EV_P)
602 {
603 linuxaio_submitcnt = 0; /* all pointers were invalidated */
604 linuxaio_free_iocbp (EV_A); /* this frees all iocbs, which is very heavy-handed */
605 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
606
607 linuxaio_iteration = 0; /* we start over in the child */
608
609 while (linuxaio_io_setup (EV_A) < 0)
610 ev_syserr ("(libev) linuxaio io_setup");
611
612 /* forking epoll should also effectively unregister all fds from the backend */
613 epoll_fork (EV_A);
614 /* epoll_fork already did this. hopefully */
615 /*fd_rearm_all (EV_A);*/
616
617 ev_io_stop (EV_A_ &linuxaio_epoll_w);
618 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
619 ev_io_start (EV_A_ &linuxaio_epoll_w);
620 }
621