"Fossies" - the Fresh Open Source Software Archive

Member "libev-4.33/ev_linuxaio.c" (27 Dec 2019, 21494 Bytes) of package /linux/misc/libev-4.33.tar.gz:


As a special service "Fossies" has tried to format the requested source page into HTML format using (guessed) C and C++ source code syntax highlighting (style: standard) with prefixed line numbers and code folding option. Alternatively you can here view or download the uninterpreted source code file. For more information about "ev_linuxaio.c" see the Fossies "Dox" file reference documentation and the latest Fossies "Diffs" side-by-side code changes report: 4.31_vs_4.33.

    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 = (nev & EV_READ  ? POLLIN  : 0)
  274                    | (nev & EV_WRITE ? POLLOUT : 0);
  275 
  276   if (nev)
  277     {
  278       iocb->io.aio_data = (uint32_t)fd | ((__u64)(uint32_t)anfd->egen << 32);
  279 
  280       /* queue iocb up for io_submit */
  281       /* this assumes we only ever get one call per fd per loop iteration */
  282       ++linuxaio_submitcnt;
  283       array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
  284       linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
  285     }
  286 }
  287 
  288 static void
  289 linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
  290 {
  291   epoll_poll (EV_A_ 0);
  292 }
  293 
  294 inline_speed
  295 void
  296 linuxaio_fd_rearm (EV_P_ int fd)
  297 {
  298   anfds [fd].events = 0;
  299   linuxaio_iocbps [fd]->io.aio_buf = 0;
  300   fd_change (EV_A_ fd, EV_ANFD_REIFY);
  301 }
  302 
  303 static void
  304 linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
  305 {
  306   while (nr)
  307     {
  308       int fd       = ev->data & 0xffffffff;
  309       uint32_t gen = ev->data >> 32;
  310       int res      = ev->res;
  311 
  312       assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
  313 
  314       /* only accept events if generation counter matches */
  315       if (ecb_expect_true (gen == (uint32_t)anfds [fd].egen))
  316         {
  317           /* feed events, we do not expect or handle POLLNVAL */
  318           fd_event (
  319             EV_A_
  320             fd,
  321             (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
  322             | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
  323           );
  324 
  325           /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
  326           linuxaio_fd_rearm (EV_A_ fd);
  327         }
  328 
  329       --nr;
  330       ++ev;
  331     }
  332 }
  333 
  334 /* get any events from ring buffer, return true if any were handled */
  335 static int
  336 linuxaio_get_events_from_ring (EV_P)
  337 {
  338   struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
  339   unsigned head, tail;
  340 
  341   /* the kernel reads and writes both of these variables, */
  342   /* as a C extension, we assume that volatile use here */
  343   /* both makes reads atomic and once-only */
  344   head = *(volatile unsigned *)&ring->head;
  345   ECB_MEMORY_FENCE_ACQUIRE;
  346   tail = *(volatile unsigned *)&ring->tail;
  347 
  348   if (head == tail)
  349     return 0;
  350 
  351   /* parse all available events, but only once, to avoid starvation */
  352   if (ecb_expect_true (tail > head)) /* normal case around */
  353     linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
  354   else /* wrapped around */
  355     {
  356       linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
  357       linuxaio_parse_events (EV_A_ ring->io_events, tail);
  358     }
  359 
  360   ECB_MEMORY_FENCE_RELEASE;
  361   /* as an extension to C, we hope that the volatile will make this atomic and once-only */
  362   *(volatile unsigned *)&ring->head = tail;
  363 
  364   return 1;
  365 }
  366 
  367 inline_size
  368 int
  369 linuxaio_ringbuf_valid (EV_P)
  370 {
  371   struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
  372 
  373   return ecb_expect_true (ring->magic == AIO_RING_MAGIC)
  374                       && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES
  375                       && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */
  376 }
  377 
  378 /* read at least one event from kernel, or timeout */
  379 inline_size
  380 void
  381 linuxaio_get_events (EV_P_ ev_tstamp timeout)
  382 {
  383   struct timespec ts;
  384   struct io_event ioev[8]; /* 256 octet stack space */
  385   int want = 1; /* how many events to request */
  386   int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
  387 
  388   if (ecb_expect_true (ringbuf_valid))
  389     {
  390       /* if the ring buffer has any events, we don't wait or call the kernel at all */
  391       if (linuxaio_get_events_from_ring (EV_A))
  392         return;
  393 
  394       /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */
  395       if (!timeout)
  396         return;
  397     }
  398   else
  399     /* no ringbuffer, request slightly larger batch */
  400     want = sizeof (ioev) / sizeof (ioev [0]);
  401 
  402   /* no events, so wait for some
  403    * for fairness reasons, we do this in a loop, to fetch all events
  404    */
  405   for (;;)
  406     {
  407       int res;
  408 
  409       EV_RELEASE_CB;
  410 
  411       EV_TS_SET (ts, timeout);
  412       res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
  413 
  414       EV_ACQUIRE_CB;
  415 
  416       if (res < 0)
  417         if (errno == EINTR)
  418           /* ignored, retry */;
  419         else
  420           ev_syserr ("(libev) linuxaio io_getevents");
  421       else if (res)
  422         {
  423           /* at least one event available, handle them */
  424           linuxaio_parse_events (EV_A_ ioev, res);
  425 
  426           if (ecb_expect_true (ringbuf_valid))
  427             {
  428               /* if we have a ring buffer, handle any remaining events in it */
  429               linuxaio_get_events_from_ring (EV_A);
  430 
  431               /* at this point, we should have handled all outstanding events */
  432               break;
  433             }
  434           else if (res < want)
  435             /* otherwise, if there were fewere events than we wanted, we assume there are no more */
  436             break;
  437         }
  438       else
  439         break; /* no events from the kernel, we are done */
  440 
  441       timeout = EV_TS_CONST (0.); /* only wait in the first iteration */
  442     }
  443 }
  444 
  445 inline_size
  446 int
  447 linuxaio_io_setup (EV_P)
  448 {
  449   linuxaio_ctx = 0;
  450   return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
  451 }
  452 
  453 static void
  454 linuxaio_poll (EV_P_ ev_tstamp timeout)
  455 {
  456   int submitted;
  457 
  458   /* first phase: submit new iocbs */
  459 
  460   /* io_submit might return less than the requested number of iocbs */
  461   /* this is, afaics, only because of errors, but we go by the book and use a loop, */
  462   /* which allows us to pinpoint the erroneous iocb */
  463   for (submitted = 0; submitted < linuxaio_submitcnt; )
  464     {
  465       int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
  466 
  467       if (ecb_expect_false (res < 0))
  468         if (errno == EINVAL)
  469           {
  470             /* This happens for unsupported fds, officially, but in my testing,
  471              * also randomly happens for supported fds. We fall back to good old
  472              * poll() here, under the assumption that this is a very rare case.
  473              * See https://lore.kernel.org/patchwork/patch/1047453/ to see
  474              * discussion about such a case (ttys) where polling for POLLIN
  475              * fails but POLLIN|POLLOUT works.
  476              */
  477             struct iocb *iocb = linuxaio_submits [submitted];
  478             epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
  479             iocb->aio_reqprio = -1; /* mark iocb as epoll */
  480 
  481             res = 1; /* skip this iocb - another iocb, another chance */
  482           }
  483         else if (errno == EAGAIN)
  484           {
  485             /* This happens when the ring buffer is full, or some other shit we
  486              * don't know and isn't documented. Most likely because we have too
  487              * many requests and linux aio can't be assed to handle them.
  488              * In this case, we try to allocate a larger ring buffer, freeing
  489              * ours first. This might fail, in which case we have to fall back to 100%
  490              * epoll.
  491              * God, how I hate linux not getting its act together. Ever.
  492              */
  493             evsys_io_destroy (linuxaio_ctx);
  494             linuxaio_submitcnt = 0;
  495 
  496             /* rearm all fds with active iocbs */
  497             {
  498               int fd;
  499           for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
  500                 if (linuxaio_iocbps [fd]->io.aio_buf)
  501                   linuxaio_fd_rearm (EV_A_ fd);
  502             }
  503 
  504             ++linuxaio_iteration;
  505             if (linuxaio_io_setup (EV_A) < 0)
  506               {
  507                 /* TODO: rearm all and recreate epoll backend from scratch */
  508                 /* TODO: might be more prudent? */
  509 
  510                 /* to bad, we can't get a new aio context, go 100% epoll */
  511                 linuxaio_free_iocbp (EV_A);
  512                 ev_io_stop (EV_A_ &linuxaio_epoll_w);
  513                 ev_ref (EV_A);
  514                 linuxaio_ctx = 0;
  515 
  516                 backend        = EVBACKEND_EPOLL;
  517                 backend_modify = epoll_modify;
  518                 backend_poll   = epoll_poll;
  519               }
  520 
  521             timeout = EV_TS_CONST (0.);
  522             /* it's easiest to handle this mess in another iteration */
  523             return;
  524           }
  525         else if (errno == EBADF)
  526           {
  527             assert (("libev: event loop rejected bad fd", errno != EBADF));
  528             fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
  529 
  530             res = 1; /* skip this iocb */
  531           }
  532         else if (errno == EINTR) /* not seen in reality, not documented */
  533           res = 0; /* silently ignore and retry */
  534         else
  535           {
  536             ev_syserr ("(libev) linuxaio io_submit");
  537             res = 0;
  538           }
  539 
  540       submitted += res;
  541     }
  542 
  543   linuxaio_submitcnt = 0;
  544 
  545   /* second phase: fetch and parse events */
  546 
  547   linuxaio_get_events (EV_A_ timeout);
  548 }
  549 
  550 inline_size
  551 int
  552 linuxaio_init (EV_P_ int flags)
  553 {
  554   /* would be great to have a nice test for IOCB_CMD_POLL instead */
  555   /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
  556   /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
  557   if (ev_linux_version () < 0x041300)
  558     return 0;
  559 
  560   if (!epoll_init (EV_A_ 0))
  561     return 0;
  562 
  563   linuxaio_iteration = 0;
  564 
  565   if (linuxaio_io_setup (EV_A) < 0)
  566     {
  567       epoll_destroy (EV_A);
  568       return 0;
  569     }
  570 
  571   ev_io_init  (&linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
  572   ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
  573   ev_io_start (EV_A_ &linuxaio_epoll_w);
  574   ev_unref (EV_A); /* watcher should not keep loop alive */
  575 
  576   backend_modify = linuxaio_modify;
  577   backend_poll   = linuxaio_poll;
  578 
  579   linuxaio_iocbpmax = 0;
  580   linuxaio_iocbps = 0;
  581 
  582   linuxaio_submits = 0;
  583   linuxaio_submitmax = 0;
  584   linuxaio_submitcnt = 0;
  585 
  586   return EVBACKEND_LINUXAIO;
  587 }
  588 
  589 inline_size
  590 void
  591 linuxaio_destroy (EV_P)
  592 {
  593   epoll_destroy (EV_A);
  594   linuxaio_free_iocbp (EV_A);
  595   evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
  596 }
  597 
  598 ecb_cold
  599 static void
  600 linuxaio_fork (EV_P)
  601 {
  602   linuxaio_submitcnt = 0; /* all pointers were invalidated */
  603   linuxaio_free_iocbp (EV_A); /* this frees all iocbs, which is very heavy-handed */
  604   evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
  605 
  606   linuxaio_iteration = 0; /* we start over in the child */
  607 
  608   while (linuxaio_io_setup (EV_A) < 0)
  609     ev_syserr ("(libev) linuxaio io_setup");
  610 
  611   /* forking epoll should also effectively unregister all fds from the backend */
  612   epoll_fork (EV_A);
  613   /* epoll_fork already did this. hopefully */
  614   /*fd_rearm_all (EV_A);*/
  615 
  616   ev_io_stop  (EV_A_ &linuxaio_epoll_w);
  617   ev_io_set   (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
  618   ev_io_start (EV_A_ &linuxaio_epoll_w);
  619 }
  620