Mirror of :pserver:anonymous@cvs.schmorp.de/schmorpforge libev http://software.schmorp.de/pkg/libev.html
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

621 lines
21 KiB

2 years ago
2 years ago
2 years ago
2 years ago
2 years ago
2 years ago
2 years ago
  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. * general notes about linux aio:
  41. *
  42. * a) at first, the linux aio IOCB_CMD_POLL functionality introduced in
  43. * 4.18 looks too good to be true: both watchers and events can be
  44. * batched, and events can even be handled in userspace using
  45. * a ring buffer shared with the kernel. watchers can be canceled
  46. * regardless of whether the fd has been closed. no problems with fork.
  47. * ok, the ring buffer is 200% undocumented (there isn't even a
  48. * header file), but otherwise, it's pure bliss!
  49. * b) ok, watchers are one-shot, so you have to re-arm active ones
  50. * on every iteration. so much for syscall-less event handling,
  51. * but at least these re-arms can be batched, no big deal, right?
  52. * c) well, linux as usual: the documentation lies to you: io_submit
  53. * sometimes returns EINVAL because the kernel doesn't feel like
  54. * handling your poll mask - ttys can be polled for POLLOUT,
  55. * POLLOUT|POLLIN, but polling for POLLIN fails. just great,
  56. * so we have to fall back to something else (hello, epoll),
  57. * but at least the fallback can be slow, because these are
  58. * exceptional cases, right?
  59. * d) hmm, you have to tell the kernel the maximum number of watchers
  60. * you want to queue when initialising the aio context. but of
  61. * course the real limit is magically calculated in the kernel, and
  62. * is often higher then we asked for. so we just have to destroy
  63. * the aio context and re-create it a bit larger if we hit the limit.
  64. * (starts to remind you of epoll? well, it's a bit more deterministic
  65. * and less gambling, but still ugly as hell).
  66. * e) that's when you find out you can also hit an arbitrary system-wide
  67. * limit. or the kernel simply doesn't want to handle your watchers.
  68. * what the fuck do we do then? you guessed it, in the middle
  69. * of event handling we have to switch to 100% epoll polling. and
  70. * that better is as fast as normal epoll polling, so you practically
  71. * have to use the normal epoll backend with all its quirks.
  72. * f) end result of this train wreck: it inherits all the disadvantages
  73. * from epoll, while adding a number on its own. why even bother to use
  74. * it? because if conditions are right and your fds are supported and you
  75. * don't hit a limit, this backend is actually faster, doesn't gamble with
  76. * your fds, batches watchers and events and doesn't require costly state
  77. * recreates. well, until it does.
  78. * g) all of this makes this backend use almost twice as much code as epoll.
  79. * which in turn uses twice as much code as poll. and that#s not counting
  80. * the fact that this backend also depends on the epoll backend, making
  81. * it three times as much code as poll, or kqueue.
  82. * h) bleah. why can't linux just do kqueue. sure kqueue is ugly, but by now
  83. * it's clear that whatever linux comes up with is far, far, far worse.
  84. */
  85. #include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */
  86. #include <poll.h>
  87. #include <linux/aio_abi.h>
  88. /*****************************************************************************/
  89. /* syscall wrapdadoop - this section has the raw api/abi definitions */
  90. #include <sys/syscall.h> /* no glibc wrappers */
  91. /* aio_abi.h is not versioned in any way, so we cannot test for its existance */
  92. #define IOCB_CMD_POLL 5
  93. /* taken from linux/fs/aio.c. yup, that's a .c file.
  94. * not only is this totally undocumented, not even the source code
  95. * can tell you what the future semantics of compat_features and
  96. * incompat_features are, or what header_length actually is for.
  97. */
  98. #define AIO_RING_MAGIC 0xa10a10a1
  99. #define EV_AIO_RING_INCOMPAT_FEATURES 0
  100. struct aio_ring
  101. {
  102. unsigned id; /* kernel internal index number */
  103. unsigned nr; /* number of io_events */
  104. unsigned head; /* Written to by userland or by kernel. */
  105. unsigned tail;
  106. unsigned magic;
  107. unsigned compat_features;
  108. unsigned incompat_features;
  109. unsigned header_length; /* size of aio_ring */
  110. struct io_event io_events[0];
  111. };
  112. inline_size
  113. int
  114. evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
  115. {
  116. return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp);
  117. }
  118. inline_size
  119. int
  120. evsys_io_destroy (aio_context_t ctx_id)
  121. {
  122. return ev_syscall1 (SYS_io_destroy, ctx_id);
  123. }
  124. inline_size
  125. int
  126. evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
  127. {
  128. return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp);
  129. }
  130. inline_size
  131. int
  132. evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result)
  133. {
  134. return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result);
  135. }
  136. inline_size
  137. int
  138. evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout)
  139. {
  140. return ev_syscall5 (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
  141. }
  142. /*****************************************************************************/
  143. /* actual backed implementation */
  144. ecb_cold
  145. static int
  146. linuxaio_nr_events (EV_P)
  147. {
  148. /* we start with 16 iocbs and incraese from there
  149. * that's tiny, but the kernel has a rather low system-wide
  150. * limit that can be reached quickly, so let's be parsimonious
  151. * with this resource.
  152. * Rest assured, the kernel generously rounds up small and big numbers
  153. * in different ways (but doesn't seem to charge you for it).
  154. * The 15 here is because the kernel usually has a power of two as aio-max-nr,
  155. * and this helps to take advantage of that limit.
  156. */
  157. /* we try to fill 4kB pages exactly.
  158. * the ring buffer header is 32 bytes, every io event is 32 bytes.
  159. * the kernel takes the io requests number, doubles it, adds 2
  160. * and adds the ring buffer.
  161. * the way we use this is by starting low, and then roughly doubling the
  162. * size each time we hit a limit.
  163. */
  164. int requests = 15 << linuxaio_iteration;
  165. int one_page = (4096
  166. / sizeof (struct io_event) ) / 2; /* how many fit into one page */
  167. int first_page = ((4096 - sizeof (struct aio_ring))
  168. / sizeof (struct io_event) - 2) / 2; /* how many fit into the first page */
  169. /* if everything fits into one page, use count exactly */
  170. if (requests > first_page)
  171. /* otherwise, round down to full pages and add the first page */
  172. requests = requests / one_page * one_page + first_page;
  173. return requests;
  174. }
  175. /* we use out own wrapper structure in case we ever want to do something "clever" */
  176. typedef struct aniocb
  177. {
  178. struct iocb io;
  179. /*int inuse;*/
  180. } *ANIOCBP;
  181. inline_size
  182. void
  183. linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
  184. {
  185. while (count--)
  186. {
  187. /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */
  188. ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
  189. /* full zero initialise is probably not required at the moment, but
  190. * this is not well documented, so we better do it.
  191. */
  192. memset (iocb, 0, sizeof (*iocb));
  193. iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
  194. iocb->io.aio_fildes = offset;
  195. base [offset++] = iocb;
  196. }
  197. }
  198. ecb_cold
  199. static void
  200. linuxaio_free_iocbp (EV_P)
  201. {
  202. while (linuxaio_iocbpmax--)
  203. ev_free (linuxaio_iocbps [linuxaio_iocbpmax]);
  204. linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */
  205. }
  206. static void
  207. linuxaio_modify (EV_P_ int fd, int oev, int nev)
  208. {
  209. array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
  210. ANIOCBP iocb = linuxaio_iocbps [fd];
  211. ANFD *anfd = &anfds [fd];
  212. if (ecb_expect_false (iocb->io.aio_reqprio < 0))
  213. {
  214. /* we handed this fd over to epoll, so undo this first */
  215. /* we do it manually because the optimisations on epoll_modify won't do us any good */
  216. epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
  217. anfd->emask = 0;
  218. iocb->io.aio_reqprio = 0;
  219. }
  220. else if (ecb_expect_false (iocb->io.aio_buf))
  221. {
  222. /* iocb active, so cancel it first before resubmit */
  223. /* this assumes we only ever get one call per fd per loop iteration */
  224. for (;;)
  225. {
  226. /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */
  227. if (ecb_expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0))
  228. break;
  229. if (ecb_expect_true (errno == EINPROGRESS))
  230. break;
  231. /* the EINPROGRESS test is for nicer error message. clumsy. */
  232. if (errno != EINTR)
  233. {
  234. assert (("libev: linuxaio unexpected io_cancel failed", errno != EINTR && errno != EINPROGRESS));
  235. break;
  236. }
  237. }
  238. /* increment generation counter to avoid handling old events */
  239. ++anfd->egen;
  240. }
  241. iocb->io.aio_buf =
  242. (nev & EV_READ ? POLLIN : 0)
  243. | (nev & EV_WRITE ? POLLOUT : 0);
  244. if (nev)
  245. {
  246. iocb->io.aio_data = (uint32_t)fd | ((__u64)(uint32_t)anfd->egen << 32);
  247. /* queue iocb up for io_submit */
  248. /* this assumes we only ever get one call per fd per loop iteration */
  249. ++linuxaio_submitcnt;
  250. array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
  251. linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
  252. }
  253. }
  254. static void
  255. linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
  256. {
  257. epoll_poll (EV_A_ 0);
  258. }
  259. inline_speed
  260. void
  261. linuxaio_fd_rearm (EV_P_ int fd)
  262. {
  263. anfds [fd].events = 0;
  264. linuxaio_iocbps [fd]->io.aio_buf = 0;
  265. fd_change (EV_A_ fd, EV_ANFD_REIFY);
  266. }
  267. static void
  268. linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
  269. {
  270. while (nr)
  271. {
  272. int fd = ev->data & 0xffffffff;
  273. uint32_t gen = ev->data >> 32;
  274. int res = ev->res;
  275. assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
  276. /* only accept events if generation counter matches */
  277. if (ecb_expect_true (gen == (uint32_t)anfds [fd].egen))
  278. {
  279. /* feed events, we do not expect or handle POLLNVAL */
  280. fd_event (
  281. EV_A_
  282. fd,
  283. (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
  284. | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
  285. );
  286. /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
  287. linuxaio_fd_rearm (EV_A_ fd);
  288. }
  289. --nr;
  290. ++ev;
  291. }
  292. }
  293. /* get any events from ring buffer, return true if any were handled */
  294. static int
  295. linuxaio_get_events_from_ring (EV_P)
  296. {
  297. struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
  298. unsigned head, tail;
  299. /* the kernel reads and writes both of these variables, */
  300. /* as a C extension, we assume that volatile use here */
  301. /* both makes reads atomic and once-only */
  302. head = *(volatile unsigned *)&ring->head;
  303. ECB_MEMORY_FENCE_ACQUIRE;
  304. tail = *(volatile unsigned *)&ring->tail;
  305. if (head == tail)
  306. return 0;
  307. /* parse all available events, but only once, to avoid starvation */
  308. if (ecb_expect_true (tail > head)) /* normal case around */
  309. linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
  310. else /* wrapped around */
  311. {
  312. linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
  313. linuxaio_parse_events (EV_A_ ring->io_events, tail);
  314. }
  315. ECB_MEMORY_FENCE_RELEASE;
  316. /* as an extension to C, we hope that the volatile will make this atomic and once-only */
  317. *(volatile unsigned *)&ring->head = tail;
  318. return 1;
  319. }
  320. inline_size
  321. int
  322. linuxaio_ringbuf_valid (EV_P)
  323. {
  324. struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
  325. return ecb_expect_true (ring->magic == AIO_RING_MAGIC)
  326. && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES
  327. && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */
  328. }
  329. /* read at least one event from kernel, or timeout */
  330. inline_size
  331. void
  332. linuxaio_get_events (EV_P_ ev_tstamp timeout)
  333. {
  334. struct timespec ts;
  335. struct io_event ioev[8]; /* 256 octet stack space */
  336. int want = 1; /* how many events to request */
  337. int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
  338. if (ecb_expect_true (ringbuf_valid))
  339. {
  340. /* if the ring buffer has any events, we don't wait or call the kernel at all */
  341. if (linuxaio_get_events_from_ring (EV_A))
  342. return;
  343. /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */
  344. if (!timeout)
  345. return;
  346. }
  347. else
  348. /* no ringbuffer, request slightly larger batch */
  349. want = sizeof (ioev) / sizeof (ioev [0]);
  350. /* no events, so wait for some
  351. * for fairness reasons, we do this in a loop, to fetch all events
  352. */
  353. for (;;)
  354. {
  355. int res;
  356. EV_RELEASE_CB;
  357. EV_TS_SET (ts, timeout);
  358. res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
  359. EV_ACQUIRE_CB;
  360. if (res < 0)
  361. if (errno == EINTR)
  362. /* ignored, retry */;
  363. else
  364. ev_syserr ("(libev) linuxaio io_getevents");
  365. else if (res)
  366. {
  367. /* at least one event available, handle them */
  368. linuxaio_parse_events (EV_A_ ioev, res);
  369. if (ecb_expect_true (ringbuf_valid))
  370. {
  371. /* if we have a ring buffer, handle any remaining events in it */
  372. linuxaio_get_events_from_ring (EV_A);
  373. /* at this point, we should have handled all outstanding events */
  374. break;
  375. }
  376. else if (res < want)
  377. /* otherwise, if there were fewere events than we wanted, we assume there are no more */
  378. break;
  379. }
  380. else
  381. break; /* no events from the kernel, we are done */
  382. timeout = EV_TS_CONST (0.); /* only wait in the first iteration */
  383. }
  384. }
  385. inline_size
  386. int
  387. linuxaio_io_setup (EV_P)
  388. {
  389. linuxaio_ctx = 0;
  390. return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
  391. }
  392. static void
  393. linuxaio_poll (EV_P_ ev_tstamp timeout)
  394. {
  395. int submitted;
  396. /* first phase: submit new iocbs */
  397. /* io_submit might return less than the requested number of iocbs */
  398. /* this is, afaics, only because of errors, but we go by the book and use a loop, */
  399. /* which allows us to pinpoint the erroneous iocb */
  400. for (submitted = 0; submitted < linuxaio_submitcnt; )
  401. {
  402. int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
  403. if (ecb_expect_false (res < 0))
  404. if (errno == EINVAL)
  405. {
  406. /* This happens for unsupported fds, officially, but in my testing,
  407. * also randomly happens for supported fds. We fall back to good old
  408. * poll() here, under the assumption that this is a very rare case.
  409. * See https://lore.kernel.org/patchwork/patch/1047453/ to see
  410. * discussion about such a case (ttys) where polling for POLLIN
  411. * fails but POLLIN|POLLOUT works.
  412. */
  413. struct iocb *iocb = linuxaio_submits [submitted];
  414. epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
  415. iocb->aio_reqprio = -1; /* mark iocb as epoll */
  416. res = 1; /* skip this iocb - another iocb, another chance */
  417. }
  418. else if (errno == EAGAIN)
  419. {
  420. /* This happens when the ring buffer is full, or some other shit we
  421. * don't know and isn't documented. Most likely because we have too
  422. * many requests and linux aio can't be assed to handle them.
  423. * In this case, we try to allocate a larger ring buffer, freeing
  424. * ours first. This might fail, in which case we have to fall back to 100%
  425. * epoll.
  426. * God, how I hate linux not getting its act together. Ever.
  427. */
  428. evsys_io_destroy (linuxaio_ctx);
  429. linuxaio_submitcnt = 0;
  430. /* rearm all fds with active iocbs */
  431. {
  432. int fd;
  433. for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
  434. if (linuxaio_iocbps [fd]->io.aio_buf)
  435. linuxaio_fd_rearm (EV_A_ fd);
  436. }
  437. ++linuxaio_iteration;
  438. if (linuxaio_io_setup (EV_A) < 0)
  439. {
  440. /* TODO: rearm all and recreate epoll backend from scratch */
  441. /* TODO: might be more prudent? */
  442. /* to bad, we can't get a new aio context, go 100% epoll */
  443. linuxaio_free_iocbp (EV_A);
  444. ev_io_stop (EV_A_ &linuxaio_epoll_w);
  445. ev_ref (EV_A);
  446. linuxaio_ctx = 0;
  447. backend = EVBACKEND_EPOLL;
  448. backend_modify = epoll_modify;
  449. backend_poll = epoll_poll;
  450. }
  451. timeout = EV_TS_CONST (0.);
  452. /* it's easiest to handle this mess in another iteration */
  453. return;
  454. }
  455. else if (errno == EBADF)
  456. {
  457. assert (("libev: event loop rejected bad fd", errno != EBADF));
  458. fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
  459. res = 1; /* skip this iocb */
  460. }
  461. else if (errno == EINTR) /* not seen in reality, not documented */
  462. res = 0; /* silently ignore and retry */
  463. else
  464. {
  465. ev_syserr ("(libev) linuxaio io_submit");
  466. res = 0;
  467. }
  468. submitted += res;
  469. }
  470. linuxaio_submitcnt = 0;
  471. /* second phase: fetch and parse events */
  472. linuxaio_get_events (EV_A_ timeout);
  473. }
  474. inline_size
  475. int
  476. linuxaio_init (EV_P_ int flags)
  477. {
  478. /* would be great to have a nice test for IOCB_CMD_POLL instead */
  479. /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
  480. /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
  481. if (ev_linux_version () < 0x041300)
  482. return 0;
  483. if (!epoll_init (EV_A_ 0))
  484. return 0;
  485. linuxaio_iteration = 0;
  486. if (linuxaio_io_setup (EV_A) < 0)
  487. {
  488. epoll_destroy (EV_A);
  489. return 0;
  490. }
  491. ev_io_init (&linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
  492. ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
  493. ev_io_start (EV_A_ &linuxaio_epoll_w);
  494. ev_unref (EV_A); /* watcher should not keep loop alive */
  495. backend_modify = linuxaio_modify;
  496. backend_poll = linuxaio_poll;
  497. linuxaio_iocbpmax = 0;
  498. linuxaio_iocbps = 0;
  499. linuxaio_submits = 0;
  500. linuxaio_submitmax = 0;
  501. linuxaio_submitcnt = 0;
  502. return EVBACKEND_LINUXAIO;
  503. }
  504. inline_size
  505. void
  506. linuxaio_destroy (EV_P)
  507. {
  508. epoll_destroy (EV_A);
  509. linuxaio_free_iocbp (EV_A);
  510. evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
  511. }
  512. ecb_cold
  513. static void
  514. linuxaio_fork (EV_P)
  515. {
  516. linuxaio_submitcnt = 0; /* all pointers were invalidated */
  517. linuxaio_free_iocbp (EV_A); /* this frees all iocbs, which is very heavy-handed */
  518. evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
  519. linuxaio_iteration = 0; /* we start over in the child */
  520. while (linuxaio_io_setup (EV_A) < 0)
  521. ev_syserr ("(libev) linuxaio io_setup");
  522. /* forking epoll should also effectively unregister all fds from the backend */
  523. epoll_fork (EV_A);
  524. /* epoll_fork already did this. hopefully */
  525. /*fd_rearm_all (EV_A);*/
  526. ev_io_stop (EV_A_ &linuxaio_epoll_w);
  527. ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
  528. ev_io_start (EV_A_ &linuxaio_epoll_w);
  529. }