libev/ev_linuxaio.c

/*
 * libev linux aio fd activity backend
 *
 * Copyright (c) 2019 Marc Alexander Lehmann <libev@schmorp.de>
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without modifica-
 * tion, are permitted provided that the following conditions are met:
 *
 *   1.  Redistributions of source code must retain the above copyright notice,
 *       this list of conditions and the following disclaimer.
 *
 *   2.  Redistributions in binary form must reproduce the above copyright
 *       notice, this list of conditions and the following disclaimer in the
 *       documentation and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER-
 * CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO
 * EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE-
 * CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH-
 * ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
 * OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 * Alternatively, the contents of this file may be used under the terms of
 * the GNU General Public License ("GPL") version 2 or any later version,
 * in which case the provisions of the GPL are applicable instead of
 * the above. If you wish to allow the use of your version of this file
 * only under the terms of the GPL and not to allow others to use your
 * version of this file under the BSD license, indicate your decision
 * by deleting the provisions above and replace them with the notice
 * and other provisions required by the GPL. If you do not delete the
 * provisions above, a recipient may use your version of this file under
 * either the BSD or the GPL.
 */

/*
 * general notes about linux aio:
 *
 * a) at first, the linux aio IOCB_CMD_POLL functionality introduced in
 *    4.18 looks too good to be true: both watchers and events can be
 *    batched, and events can even be handled in userspace using
 *    a ring buffer shared with the kernel. watchers can be canceled
 *    regardless of whether the fd has been closed. no problems with fork.
 *    ok, the ring buffer is 200% undocumented (there isn't even a
 *    header file), but otherwise, it's pure bliss!
 * b) ok, watchers are one-shot, so you have to re-arm active ones
 *    on every iteration. so much for syscall-less event handling,
 *    but at least these re-arms can be batched, no big deal, right?
 * c) well, linux as usual: the documentation lies to you: io_submit
 *    sometimes returns EINVAL because the kernel doesn't feel like
 *    handling your poll mask - ttys can be polled for POLLOUT,
 *    POLLOUT|POLLIN, but polling for POLLIN fails. just great,
 *    so we have to fall back to something else (hello, epoll),
 *    but at least the fallback can be slow, because these are
 *    exceptional cases, right?
 * d) hmm, you have to tell the kernel the maximum number of watchers
 *    you want to queue when initialising the aio context. but of
 *    course the real limit is magically calculated in the kernel, and
 *    is often higher then we asked for. so we just have to destroy
 *    the aio context and re-create it a bit larger if we hit the limit.
 *    (starts to remind you of epoll? well, it's a bit more deterministic
 *    and less gambling, but still ugly as hell).
 * e) that's when you find out you can also hit an arbitrary system-wide
 *    limit. or the kernel simply doesn't want to handle your watchers.
 *    what the fuck do we do then? you guessed it, in the middle
 *    of event handling we have to switch to 100% epoll polling. and
 *    that better is as fast as normal epoll polling, so you practically
 *    have to use the normal epoll backend with all its quirks.
 * f) end result of this train wreck: it inherits all the disadvantages
 *    from epoll, while adding a number on its own. why even bother to use
 *    it? because if conditions are right and your fds are supported and you
 *    don't hit a limit, this backend is actually faster, doesn't gamble with
 *    your fds, batches watchers and events and doesn't require costly state
 *    recreates. well, until it does.
 * g) all of this makes this backend use almost twice as much code as epoll.
 *    which in turn uses twice as much code as poll. and that#s not counting
 *    the fact that this backend also depends on the epoll backend, making
 *    it three times as much code as poll, or kqueue.
 * h) bleah. why can't linux just do kqueue. sure kqueue is ugly, but by now
 *    it's clear that whatever linux comes up with is far, far, far worse.
 */

#include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */
#include <poll.h>
#include <linux/aio_abi.h>

/*****************************************************************************/
/* syscall wrapdadoop - this section has the raw api/abi definitions */

#include <sys/syscall.h> /* no glibc wrappers */

/* aio_abi.h is not versioned in any way, so we cannot test for its existance */
#define IOCB_CMD_POLL 5

/* taken from linux/fs/aio.c. yup, that's a .c file.
 * not only is this totally undocumented, not even the source code
 * can tell you what the future semantics of compat_features and
 * incompat_features are, or what header_length actually is for.
 */
#define AIO_RING_MAGIC                  0xa10a10a1
#define AIO_RING_INCOMPAT_FEATURES      0
struct aio_ring
{
  unsigned id;    /* kernel internal index number */
  unsigned nr;    /* number of io_events */
  unsigned head;  /* Written to by userland or by kernel. */
  unsigned tail;

  unsigned magic;
  unsigned compat_features;
  unsigned incompat_features;
  unsigned header_length;  /* size of aio_ring */

  struct io_event io_events[0];
};

inline_size
int
evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
{
  return syscall (SYS_io_setup, nr_events, ctx_idp);
}

inline_size
int
evsys_io_destroy (aio_context_t ctx_id)
{
  return syscall (SYS_io_destroy, ctx_id);
}

inline_size
int
evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
{
  return syscall (SYS_io_submit, ctx_id, nr, cbp);
}

inline_size
int
evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result)
{
  return syscall (SYS_io_cancel, ctx_id, cbp, result);
}

inline_size
int
evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout)
{
  return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
}

/*****************************************************************************/
/* actual backed implementation */

ecb_cold
static int
linuxaio_nr_events (EV_P)
{
  /* we start with 16 iocbs and incraese from there
   * that's tiny, but the kernel has a rather low system-wide
   * limit that can be reached quickly, so let's be parsimonious
   * with this resource.
   * Rest assured, the kernel generously rounds up small and big numbers
   * in different ways (but doesn't seem to charge you for it).
   * The 15 here is because the kernel usually has a power of two as aio-max-nr,
   * and this helps to take advantage of that limit.
   */

  /* we try to fill 4kB pages exactly.
   * the ring buffer header is 32 bytes, every io event is 32 bytes.
   * the kernel takes the io requests number, doubles it, adds 2
   * and adds the ring buffer.
   * the way we use this is by starting low, and then roughly doubling the
   * size each time we hit a limit.
   */

  int requests   = 15 << linuxaio_iteration;
  int one_page   =  (4096
                    / sizeof (struct io_event)    ) / 2; /* how many fit into one page */
  int first_page = ((4096 - sizeof (struct aio_ring))
                    / sizeof (struct io_event) - 2) / 2; /* how many fit into the first page */

  /* if everything fits into one page, use count exactly */
  if (requests > first_page)
    /* otherwise, round down to full pages and add the first page */
    requests = requests / one_page * one_page + first_page;

  return requests;
}

/* we use out own wrapper structure in case we ever want to do something "clever" */
typedef struct aniocb
{
  struct iocb io;
  /*int inuse;*/
} *ANIOCBP;

inline_size
void
linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
{
  while (count--)
    {
      /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */
      ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));

      /* full zero initialise is probably not required at the moment, but
       * this is not well documented, so we better do it.
       */
      memset (iocb, 0, sizeof (*iocb));

      iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
      iocb->io.aio_data       = offset;
      iocb->io.aio_fildes     = offset;

      base [offset++] = iocb;
    }
}

ecb_cold
static void
linuxaio_free_iocbp (EV_P)
{
  while (linuxaio_iocbpmax--)
    ev_free (linuxaio_iocbps [linuxaio_iocbpmax]);

  linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */
}

static void
linuxaio_modify (EV_P_ int fd, int oev, int nev)
{
  array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
  ANIOCBP iocb = linuxaio_iocbps [fd];

  if (iocb->io.aio_reqprio < 0)
    {
      /* we handed this fd over to epoll, so undo this first */
      /* we do it manually because the optimisations on epoll_modfy won't do us any good */
      epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
      iocb->io.aio_reqprio = 0;
    }

  if (iocb->io.aio_buf)
    /* io_cancel always returns some error on relevant kernels, but works */
    evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0);

  if (nev)
    {
      iocb->io.aio_buf =
          (nev & EV_READ ? POLLIN : 0)
          | (nev & EV_WRITE ? POLLOUT : 0);

      /* queue iocb up for io_submit */
      /* this assumes we only ever get one call per fd per loop iteration */
      ++linuxaio_submitcnt;
      array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
      linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
    }
}

static void
linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
{
  epoll_poll (EV_A_ 0);
}

static void
linuxaio_fd_rearm (EV_P_ int fd)
{
  anfds [fd].events = 0;
  linuxaio_iocbps [fd]->io.aio_buf = 0;
  fd_change (EV_A_ fd, EV_ANFD_REIFY);
}

static void
linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
{
  while (nr)
    {
      int fd  = ev->data;
      int res = ev->res;

      assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));

      /* feed events, we do not expect or handle POLLNVAL */
      fd_event (
        EV_A_
        fd,
        (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
        | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
      );

      /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
      linuxaio_fd_rearm (EV_A_ fd);

      --nr;
      ++ev;
    }
}

/* get any events from ring buffer, return true if any were handled */
static int
linuxaio_get_events_from_ring (EV_P)
{
  struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;

  /* the kernel reads and writes both of these variables, */
  /* as a C extension, we assume that volatile use here */
  /* both makes reads atomic and once-only */
  unsigned head = *(volatile unsigned *)&ring->head;
  unsigned tail = *(volatile unsigned *)&ring->tail;

  if (head == tail)
    return 0;

  /* bail out if the ring buffer doesn't match the expected layout */
  if (expect_false (ring->magic != AIO_RING_MAGIC)
                    || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
                    || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
    return 0;

  /* make sure the events up to tail are visible */
  ECB_MEMORY_FENCE_ACQUIRE;

  /* parse all available events, but only once, to avoid starvation */
  if (tail > head) /* normal case around */
    linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
  else /* wrapped around */
    {
      linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
      linuxaio_parse_events (EV_A_ ring->io_events, tail);
    }

  ECB_MEMORY_FENCE_RELEASE;
  /* as an extension to C, we hope that the volatile will make this atomic and once-only */
  *(volatile unsigned *)&ring->head = tail;
  /* make sure kernel can see our new head value - probably not required */

  return 1;
}

/* read at least one event from kernel, or timeout */
inline_size
void
linuxaio_get_events (EV_P_ ev_tstamp timeout)
{
  struct timespec ts;
  struct io_event ioev[1];
  int res;

  if (linuxaio_get_events_from_ring (EV_A))
    return;

  /* no events, so wait for at least one, then poll ring buffer again */
  /* this degrades to one event per loop iteration */
  /* if the ring buffer changes layout, but so be it */

  EV_RELEASE_CB;

  ts.tv_sec  = (long)timeout;
  ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);

  res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);

  EV_ACQUIRE_CB;

  if (res < 0)
    if (errno == EINTR)
      /* ignored */;
    else
      ev_syserr ("(libev) linuxaio io_getevents");
  else if (res)
    {
      /* at least one event received, handle it and any remaining ones in the ring buffer */
      linuxaio_parse_events (EV_A_ ioev, res);
      linuxaio_get_events_from_ring (EV_A);
    }
}

static int
linuxaio_io_setup (EV_P)
{
  linuxaio_ctx = 0;
  return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
}

static void
linuxaio_poll (EV_P_ ev_tstamp timeout)
{
  int submitted;

  /* first phase: submit new iocbs */

  /* io_submit might return less than the requested number of iocbs */
  /* this is, afaics, only because of errors, but we go by the book and use a loop, */
  /* which allows us to pinpoint the erroneous iocb */
  for (submitted = 0; submitted < linuxaio_submitcnt; )
    {
      int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);

      if (expect_false (res < 0))
        if (errno == EINVAL)
          {
            /* This happens for unsupported fds, officially, but in my testing,
             * also randomly happens for supported fds. We fall back to good old
             * poll() here, under the assumption that this is a very rare case.
             * See https://lore.kernel.org/patchwork/patch/1047453/ to see
             * discussion about such a case (ttys) where polling for POLLIN
             * fails but POLLIN|POLLOUT works.
             */
            struct iocb *iocb = linuxaio_submits [submitted];
            epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
            iocb->aio_reqprio = -1; /* mark iocb as epoll */

            res = 1; /* skip this iocb - another iocb, another chance */
          }
        else if (errno == EAGAIN)
          {
            /* This happens when the ring buffer is full, or some other shit we
             * don't know and isn't documented. Most likely because we have too
             * many requests and linux aio can't be assed to handle them.
             * In this case, we try to allocate a larger ring buffer, freeing
             * ours first. This might fail, in which case we have to fall back to 100%
             * epoll.
             * God, how I hate linux not getting its act together. Ever.
             */
            evsys_io_destroy (linuxaio_ctx);
            linuxaio_submitcnt = 0;

            /* rearm all fds with active iocbs */
            {
              int fd;
              for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
                if (linuxaio_iocbps [fd]->io.aio_buf)
                  linuxaio_fd_rearm (EV_A_ fd);
            }

            ++linuxaio_iteration;
            if (linuxaio_io_setup (EV_A) < 0)
              {
                /* to bad, we can't get a new aio context, go 100% epoll */
                linuxaio_free_iocbp (EV_A);
                ev_io_stop (EV_A_ &linuxaio_epoll_w);
                ev_ref (EV_A);
                linuxaio_ctx = 0;
                backend_modify = epoll_modify;
                backend_poll   = epoll_poll;
              }

            timeout = 0;
            /* it's easiest to handle this mess in another iteration */
            return;
          }
        else if (errno == EBADF)
          {
            fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);

            res = 1; /* skip this iocb */
          }
        else
          ev_syserr ("(libev) linuxaio io_submit");

      submitted += res;
    }

  linuxaio_submitcnt = 0;

  /* second phase: fetch and parse events */

  linuxaio_get_events (EV_A_ timeout);
}

inline_size
int
linuxaio_init (EV_P_ int flags)
{
  /* would be great to have a nice test for IOCB_CMD_POLL instead */
  /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
  /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
  if (ev_linux_version () < 0x041300)
    return 0;

  if (!epoll_init (EV_A_ 0))
    return 0;

  linuxaio_iteration = 0;

  if (linuxaio_io_setup (EV_A) < 0)
    {
      epoll_destroy (EV_A);
      return 0;
    }

  ev_io_init  (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
  ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
  ev_io_start (EV_A_ &linuxaio_epoll_w);
  ev_unref (EV_A); /* watcher should not keep loop alive */

  backend_modify  = linuxaio_modify;
  backend_poll    = linuxaio_poll;

  linuxaio_iocbpmax = 0;
  linuxaio_iocbps = 0;

  linuxaio_submits = 0;
  linuxaio_submitmax = 0;
  linuxaio_submitcnt = 0;

  return EVBACKEND_LINUXAIO;
}

inline_size
void
linuxaio_destroy (EV_P)
{
  epoll_destroy (EV_A);
  linuxaio_free_iocbp (EV_A);
  evsys_io_destroy (linuxaio_ctx);
}

inline_size
void
linuxaio_fork (EV_P)
{
  /* this frees all iocbs, which is very heavy-handed */
  linuxaio_destroy (EV_A);
  linuxaio_submitcnt = 0; /* all pointers were invalidated */

  linuxaio_iteration = 0; /* we start over in the child */

  while (linuxaio_io_setup (EV_A) < 0)
    ev_syserr ("(libev) linuxaio io_setup");

  epoll_fork (EV_A);

  ev_io_stop  (EV_A_ &linuxaio_epoll_w);
  ev_io_set   (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
  ev_io_start (EV_A_ &linuxaio_epoll_w);

  /* epoll_fork already did this. hopefully */
  /*fd_rearm_all (EV_A);*/
}

Revision:	1.28
Committed:	Tue Jun 25 06:36:59 2019 UTC (4 years, 10 months ago) by root
Content type:	text/plain
Branch:	MAIN
Changes since 1.27:	+1 -2 lines
Log Message:	* empty log message *
#	User	Rev	Content
1	root	1.1	/*
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	root	1.25	/*
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	root	1.27	* you want to queue when initialising the aio context. but of
62	root	1.25	* 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	root	1.27	* f) end result of this train wreck: it inherits all the disadvantages
74	root	1.25	* 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	root	1.27	* which in turn uses twice as much code as poll. and that#s not counting
81	root	1.25	* 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	root	1.27	* it's clear that whatever linux comes up with is far, far, far worse.
85	root	1.25	*/
86	root	1.10
87	root	1.1	#include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */
88	root	1.2	#include <poll.h>
89	root	1.1	#include <linux/aio_abi.h>
90
91			/*****************************************************************************/
92	root	1.25	/* syscall wrapdadoop - this section has the raw api/abi definitions */
93	root	1.1
94			#include <sys/syscall.h> /* no glibc wrappers */
95
96	root	1.5	/* aio_abi.h is not versioned in any way, so we cannot test for its existance */
97	root	1.1	#define IOCB_CMD_POLL 5
98
99	root	1.25	/* 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	root	1.1	#define AIO_RING_MAGIC 0xa10a10a1
105			#define 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	root	1.6	inline_size
122			int
123	root	1.20	evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
124	root	1.1	{
125			return syscall (SYS_io_setup, nr_events, ctx_idp);
126			}
127
128	root	1.6	inline_size
129			int
130	root	1.20	evsys_io_destroy (aio_context_t ctx_id)
131	root	1.1	{
132			return syscall (SYS_io_destroy, ctx_id);
133			}
134
135	root	1.6	inline_size
136			int
137	root	1.20	evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
138	root	1.1	{
139			return syscall (SYS_io_submit, ctx_id, nr, cbp);
140			}
141
142	root	1.6	inline_size
143			int
144	root	1.20	evsys_io_cancel (aio_context_t ctx_id, struct iocb cbp, struct io_event result)
145	root	1.1	{
146			return syscall (SYS_io_cancel, ctx_id, cbp, result);
147			}
148
149	root	1.6	inline_size
150			int
151	root	1.20	evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event events, struct timespec timeout)
152	root	1.1	{
153			return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
154			}
155
156			/*****************************************************************************/
157			/* actual backed implementation */
158
159	root	1.25	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	root	1.27	/* we use out own wrapper structure in case we ever want to do something "clever" */
196	root	1.1	typedef struct aniocb
197			{
198			struct iocb io;
199			/int inuse;/
200			} *ANIOCBP;
201
202			inline_size
203			void
204	root	1.22	linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
205	root	1.1	{
206			while (count--)
207			{
208	root	1.27	/* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */
209	root	1.22	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_data = offset;
218			iocb->io.aio_fildes = offset;
219
220			base [offset++] = iocb;
221	root	1.1	}
222			}
223
224	root	1.6	ecb_cold
225	root	1.1	static void
226			linuxaio_free_iocbp (EV_P)
227			{
228			while (linuxaio_iocbpmax--)
229			ev_free (linuxaio_iocbps [linuxaio_iocbpmax]);
230
231	root	1.6	linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */
232	root	1.1	}
233
234			static void
235			linuxaio_modify (EV_P_ int fd, int oev, int nev)
236			{
237			array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
238	root	1.22	ANIOCBP iocb = linuxaio_iocbps [fd];
239	root	1.1
240	root	1.10	if (iocb->io.aio_reqprio < 0)
241			{
242	root	1.25	/* we handed this fd over to epoll, so undo this first */
243	root	1.27	/* we do it manually because the optimisations on epoll_modfy won't do us any good */
244	root	1.10	epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
245			iocb->io.aio_reqprio = 0;
246			}
247
248	root	1.1	if (iocb->io.aio_buf)
249	root	1.25	/* io_cancel always returns some error on relevant kernels, but works */
250			evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0);
251	root	1.1
252			if (nev)
253			{
254	root	1.22	iocb->io.aio_buf =
255	root	1.1	(nev & EV_READ ? POLLIN : 0)
256			\| (nev & EV_WRITE ? POLLOUT : 0);
257
258			/* queue iocb up for io_submit */
259			/* this assumes we only ever get one call per fd per loop iteration */
260			++linuxaio_submitcnt;
261			array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
262			linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
263			}
264			}
265
266	root	1.19	static void
267	root	1.25	linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
268	root	1.19	{
269	root	1.25	epoll_poll (EV_A_ 0);
270	root	1.19	}
271
272			static void
273	root	1.25	linuxaio_fd_rearm (EV_P_ int fd)
274	root	1.19	{
275	root	1.25	anfds [fd].events = 0;
276			linuxaio_iocbps [fd]->io.aio_buf = 0;
277			fd_change (EV_A_ fd, EV_ANFD_REIFY);
278	root	1.19	}
279
280	root	1.1	static void
281			linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
282			{
283			while (nr)
284			{
285			int fd = ev->data;
286			int res = ev->res;
287
288	root	1.2	assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
289	root	1.1
290			/* feed events, we do not expect or handle POLLNVAL */
291	root	1.21	fd_event (
292			EV_A_
293			fd,
294			(res & (POLLOUT \| POLLERR \| POLLHUP) ? EV_WRITE : 0)
295			\| (res & (POLLIN \| POLLERR \| POLLHUP) ? EV_READ : 0)
296			);
297	root	1.1
298	root	1.25	/* linux aio is oneshot: rearm fd. TODO: this does more work than needed */
299			linuxaio_fd_rearm (EV_A_ fd);
300
301	root	1.1	--nr;
302			++ev;
303			}
304			}
305
306	root	1.27	/* get any events from ring buffer, return true if any were handled */
307	root	1.1	static int
308			linuxaio_get_events_from_ring (EV_P)
309			{
310			struct aio_ring ring = (struct aio_ring )linuxaio_ctx;
311
312	root	1.13	/* the kernel reads and writes both of these variables, */
313			/* as a C extension, we assume that volatile use here */
314			/* both makes reads atomic and once-only */
315			unsigned head = (volatile unsigned )&ring->head;
316	root	1.1	unsigned tail = (volatile unsigned )&ring->tail;
317
318	root	1.6	if (head == tail)
319			return 0;
320
321	root	1.7	/* bail out if the ring buffer doesn't match the expected layout */
322	root	1.17	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	root	1.1	return 0;
326
327	root	1.12	/* make sure the events up to tail are visible */
328	root	1.9	ECB_MEMORY_FENCE_ACQUIRE;
329
330	root	1.1	/* parse all available events, but only once, to avoid starvation */
331			if (tail > head) /* normal case around */
332			linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
333	root	1.6	else /* wrapped around */
334	root	1.1	{
335			linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
336			linuxaio_parse_events (EV_A_ ring->io_events, tail);
337			}
338
339	root	1.28	ECB_MEMORY_FENCE_RELEASE;
340	root	1.16	/* as an extension to C, we hope that the volatile will make this atomic and once-only */
341	root	1.11	(volatile unsigned )&ring->head = tail;
342	root	1.12	/* make sure kernel can see our new head value - probably not required */
343	root	1.1
344			return 1;
345			}
346
347			/* read at least one event from kernel, or timeout */
348			inline_size
349			void
350			linuxaio_get_events (EV_P_ ev_tstamp timeout)
351			{
352			struct timespec ts;
353	root	1.18	struct io_event ioev[1];
354	root	1.1	int res;
355
356			if (linuxaio_get_events_from_ring (EV_A))
357			return;
358
359			/* no events, so wait for at least one, then poll ring buffer again */
360	root	1.7	/* this degrades to one event per loop iteration */
361	root	1.1	/* if the ring buffer changes layout, but so be it */
362
363	root	1.19	EV_RELEASE_CB;
364
365	root	1.1	ts.tv_sec = (long)timeout;
366			ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
367
368	root	1.20	res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts);
369	root	1.1
370	root	1.19	EV_ACQUIRE_CB;
371
372	root	1.1	if (res < 0)
373	root	1.10	if (errno == EINTR)
374			/* ignored */;
375			else
376			ev_syserr ("(libev) linuxaio io_getevents");
377	root	1.1	else if (res)
378			{
379			/* at least one event received, handle it and any remaining ones in the ring buffer */
380	root	1.18	linuxaio_parse_events (EV_A_ ioev, res);
381	root	1.1	linuxaio_get_events_from_ring (EV_A);
382			}
383			}
384
385	root	1.25	static int
386			linuxaio_io_setup (EV_P)
387			{
388			linuxaio_ctx = 0;
389			return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
390			}
391
392	root	1.1	static void
393			linuxaio_poll (EV_P_ ev_tstamp timeout)
394			{
395			int submitted;
396
397			/* first phase: submit new iocbs */
398
399			/* io_submit might return less than the requested number of iocbs */
400			/* this is, afaics, only because of errors, but we go by the book and use a loop, */
401	root	1.27	/* which allows us to pinpoint the erroneous iocb */
402	root	1.1	for (submitted = 0; submitted < linuxaio_submitcnt; )
403			{
404	root	1.20	int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
405	root	1.1
406	root	1.17	if (expect_false (res < 0))
407	root	1.25	if (errno == EINVAL)
408	root	1.10	{
409	root	1.15	/* This happens for unsupported fds, officially, but in my testing,
410	root	1.10	* 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.
412	root	1.19	* See https://lore.kernel.org/patchwork/patch/1047453/ to see
413			* discussion about such a case (ttys) where polling for POLLIN
414			* fails but POLLIN\|POLLOUT works.
415	root	1.10	*/
416			struct iocb *iocb = linuxaio_submits [submitted];
417	root	1.25	epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
418			iocb->aio_reqprio = -1; /* mark iocb as epoll */
419	root	1.10
420	root	1.25	res = 1; /* skip this iocb - another iocb, another chance */
421			}
422			else if (errno == EAGAIN)
423			{
424			/* This happens when the ring buffer is full, or some other shit we
425	root	1.27	* don't know and isn't documented. Most likely because we have too
426	root	1.25	* many requests and linux aio can't be assed to handle them.
427			* In this case, we try to allocate a larger ring buffer, freeing
428			* ours first. This might fail, in which case we have to fall back to 100%
429			* epoll.
430			* God, how I hate linux not getting its act together. Ever.
431			*/
432			evsys_io_destroy (linuxaio_ctx);
433			linuxaio_submitcnt = 0;
434
435			/* rearm all fds with active iocbs */
436			{
437			int fd;
438			for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
439			if (linuxaio_iocbps [fd]->io.aio_buf)
440			linuxaio_fd_rearm (EV_A_ fd);
441			}
442
443			++linuxaio_iteration;
444			if (linuxaio_io_setup (EV_A) < 0)
445			{
446			/* to bad, we can't get a new aio context, go 100% epoll */
447			linuxaio_free_iocbp (EV_A);
448			ev_io_stop (EV_A_ &linuxaio_epoll_w);
449			ev_ref (EV_A);
450			linuxaio_ctx = 0;
451			backend_modify = epoll_modify;
452			backend_poll = epoll_poll;
453			}
454	root	1.21
455	root	1.25	timeout = 0;
456			/* it's easiest to handle this mess in another iteration */
457			return;
458	root	1.10	}
459	root	1.21	else if (errno == EBADF)
460			{
461			fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
462
463			res = 1; /* skip this iocb */
464			}
465	root	1.1	else
466	root	1.8	ev_syserr ("(libev) linuxaio io_submit");
467	root	1.1
468			submitted += res;
469			}
470
471			linuxaio_submitcnt = 0;
472
473			/* second phase: fetch and parse events */
474
475			linuxaio_get_events (EV_A_ timeout);
476			}
477
478			inline_size
479			int
480			linuxaio_init (EV_P_ int flags)
481			{
482			/* would be great to have a nice test for IOCB_CMD_POLL instead */
483	root	1.2	/* also: test some semi-common fd types, such as files and ttys in recommended_backends */
484	root	1.27	/* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
485	root	1.15	if (ev_linux_version () < 0x041300)
486			return 0;
487	root	1.25
488			if (!epoll_init (EV_A_ 0))
489	root	1.1	return 0;
490
491	root	1.25	linuxaio_iteration = 0;
492	root	1.1
493	root	1.25	if (linuxaio_io_setup (EV_A) < 0)
494	root	1.10	{
495	root	1.25	epoll_destroy (EV_A);
496	root	1.10	return 0;
497			}
498
499			ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
500	root	1.19	ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
501	root	1.10	ev_io_start (EV_A_ &linuxaio_epoll_w);
502	root	1.14	ev_unref (EV_A); /* watcher should not keep loop alive */
503	root	1.10
504	root	1.1	backend_modify = linuxaio_modify;
505			backend_poll = linuxaio_poll;
506
507			linuxaio_iocbpmax = 0;
508			linuxaio_iocbps = 0;
509
510			linuxaio_submits = 0;
511			linuxaio_submitmax = 0;
512			linuxaio_submitcnt = 0;
513
514			return EVBACKEND_LINUXAIO;
515			}
516
517			inline_size
518			void
519			linuxaio_destroy (EV_P)
520			{
521	root	1.25	epoll_destroy (EV_A);
522	root	1.1	linuxaio_free_iocbp (EV_A);
523	root	1.21	evsys_io_destroy (linuxaio_ctx);
524	root	1.1	}
525
526			inline_size
527			void
528			linuxaio_fork (EV_P)
529			{
530	root	1.6	/* this frees all iocbs, which is very heavy-handed */
531	root	1.2	linuxaio_destroy (EV_A);
532	root	1.6	linuxaio_submitcnt = 0; /* all pointers were invalidated */
533	root	1.2
534	root	1.25	linuxaio_iteration = 0; /* we start over in the child */
535
536			while (linuxaio_io_setup (EV_A) < 0)
537	root	1.8	ev_syserr ("(libev) linuxaio io_setup");
538	root	1.2
539	root	1.25	epoll_fork (EV_A);
540	root	1.10
541			ev_io_stop (EV_A_ &linuxaio_epoll_w);
542	root	1.25	ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
543	root	1.10	ev_io_start (EV_A_ &linuxaio_epoll_w);
544
545	root	1.25	/* epoll_fork already did this. hopefully */
546			/fd_rearm_all (EV_A);/
547	root	1.1	}
548