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.
 */

#define EPOLL_FALLBACK 1

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

#if EPOLL_FALLBACK
# include <sys/epoll.h>
#endif

/* 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 event number, doubles it, adds 2, adds the ring buffer.
 * therefore the calculation below will use "exactly" 4kB for the ring buffer
 */
#define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */

/*****************************************************************************/
/* syscall wrapdadoop */

#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 */
#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
ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
{
  return syscall (SYS_io_setup, nr_events, ctx_idp);
}

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

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

inline_size
int
ev_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
ev_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 */

/* we use out own wrapper structure in acse 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 count)
{
  /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
  while (count--)
    {
      *base = (ANIOCBP)ev_malloc (sizeof (**base));
      /* TODO: full zero initialize required? */
      memset (*base, 0, sizeof (**base));
      /* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */
      (*base)->io.aio_lio_opcode = IOCB_CMD_POLL;
      ++base;
    }
}

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);
  struct aniocb *iocb = linuxaio_iocbps [fd];

#if EPOLL_FALLBACK
  if (iocb->io.aio_reqprio < 0)
    {
      epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
      iocb->io.aio_reqprio = 0;
    }
#endif

  if (iocb->io.aio_buf)
    ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */

  if (nev)
    {
      iocb->io.aio_data       = fd;
      iocb->io.aio_fildes     = fd;
      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_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));

      /* linux aio is oneshot: rearm fd */
      linuxaio_iocbps [fd]->io.aio_buf = 0;
      anfds [fd].events = 0;
      fd_change (EV_A_ fd, 0);

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

      --nr;
      ++ev;
    }
}

/* get any events from ringbuffer, 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 (ecb_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);
    }

  /* TODO: we only need a compiler barrier here, not a read fence */
  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 */
  ECB_MEMORY_FENCE_RELEASE;

  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;
  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 */

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

  res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts);

  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, 1);
      linuxaio_get_events_from_ring (EV_A);
    }
}

#if EPOLL_FALLBACK
static void
linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op)
{
  struct epoll_event eev;

  eev.events = EPOLLONESHOT;
  if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ;
  if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT;
  eev.data.fd = iocb->aio_fildes;

  if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
    ev_syserr ("(libeio) linuxaio epoll_ctl");
}
#endif

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 errornous iocb */
  for (submitted = 0; submitted < linuxaio_submitcnt; )
    {
      int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);

      if (ecb_expect_false (res < 0))
        if (errno == EAGAIN)
          {
            /* This happens when the ring buffer is full, at least. I assume this means
             * that the event was queued synchronously during io_submit, and thus
             * the buffer overflowed.
             * In this case, we just try in next loop iteration.
             * This should not result in a few fds taking priority, as the interface
             * is one-shot, and we submit iocb's in a round-robin fashion.
             */
            memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
            linuxaio_submitcnt -= submitted;
            timeout = 0;
            break;
          }
#if EPOLL_FALLBACK
        else 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/ for evidence
             * that the problem is known, but ignored.
             */
            struct iocb *iocb = linuxaio_submits [submitted];
            res = 1; /* skip this iocb */

            linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD);
            iocb->aio_reqprio = -1; /* mark iocb as epoll */
          }
#endif
        else
          ev_syserr ("(libev) linuxaio io_submit");

      submitted += res;
    }

  linuxaio_submitcnt = 0;

  /* second phase: fetch and parse events */

  linuxaio_get_events (EV_A_ timeout);
}

#if EPOLL_FALLBACK

static void
linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
{
  struct epoll_event events[16];

  for (;;)
    {
      int idx;
      int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);

      if (ecb_expect_false (res < 0))
        ev_syserr ("(libev) linuxaio epoll_wait");
      else if (!res)
        break;

      for (idx = res; idx--; )
        {
          int      fd = events [idx].data.fd;
          uint32_t ev = events [idx].events;

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

          linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);

          fd_event (EV_A_ fd,
            (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0)
            | (ev & (EPOLLIN  | EPOLLERR | EPOLLHUP) ? EV_READ  : 0));
        }

      if (res < sizeof (events) / sizeof (events [0]))
        break;
    }
}

#endif

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 */
#if EPOLL_FALLBACK
  /* 4.19 made epoll work */
  if (ev_linux_version () < 0x041300)
    return 0;
#else
  /* 4.18 introduced IOCB_CMD_POLL */
  if (ev_linux_version () < 0x041200)
    return 0;
#endif

  linuxaio_ctx = 0;
  if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
    return 0;

#if EPOLL_FALLBACK
  backend_fd = ev_epoll_create ();
  if (backend_fd < 0)
    {
      ev_io_destroy (linuxaio_ctx);
      return 0;
    }

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

  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)
{
#if EPOLL_FALLBACK
  close (backend_fd);
#endif
  linuxaio_free_iocbp (EV_A);
  ev_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_ctx = 0;
  while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
    ev_syserr ("(libev) linuxaio io_setup");

#if EPOLL_FALLBACK
  while ((backend_fd = ev_epoll_create ()) < 0)
   ev_syserr ("(libev) linuxaio epoll_create");

  ev_io_stop  (EV_A_ &linuxaio_epoll_w);
  ev_io_init  (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
  ev_io_start (EV_A_ &linuxaio_epoll_w);
#endif

  fd_rearm_all (EV_A);
}

Revision:	1.16
Committed:	Mon Jun 24 00:30:24 2019 UTC (4 years, 10 months ago) by root
Content type:	text/plain
Branch:	MAIN
Changes since 1.15:	+3 -1 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	1.10	#define EPOLL_FALLBACK 1
41
42	root	1.1	#include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */
43	root	1.2	#include <poll.h>
44	root	1.1	#include <linux/aio_abi.h>
45
46	root	1.10	#if EPOLL_FALLBACK
47			# include <sys/epoll.h>
48			#endif
49
50	root	1.6	/* we try to fill 4kB pages exactly.
51	root	1.1	* the ring buffer header is 32 bytes, every io event is 32 bytes.
52	root	1.6	* the kernel takes the io event number, doubles it, adds 2, adds the ring buffer.
53	root	1.9	* therefore the calculation below will use "exactly" 4kB for the ring buffer
54	root	1.1	*/
55	root	1.9	#define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */
56	root	1.1
57			/*****************************************************************************/
58			/* syscall wrapdadoop */
59
60			#include <sys/syscall.h> /* no glibc wrappers */
61
62	root	1.5	/* aio_abi.h is not versioned in any way, so we cannot test for its existance */
63	root	1.1	#define IOCB_CMD_POLL 5
64
65			/* taken from linux/fs/aio.c */
66			#define AIO_RING_MAGIC 0xa10a10a1
67			#define AIO_RING_INCOMPAT_FEATURES 0
68			struct aio_ring
69			{
70			unsigned id; /* kernel internal index number */
71			unsigned nr; /* number of io_events */
72			unsigned head; /* Written to by userland or by kernel. */
73			unsigned tail;
74
75			unsigned magic;
76			unsigned compat_features;
77			unsigned incompat_features;
78			unsigned header_length; /* size of aio_ring */
79
80			struct io_event io_events[0];
81			};
82
83	root	1.6	inline_size
84			int
85	root	1.1	ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
86			{
87			return syscall (SYS_io_setup, nr_events, ctx_idp);
88			}
89
90	root	1.6	inline_size
91			int
92	root	1.1	ev_io_destroy (aio_context_t ctx_id)
93			{
94			return syscall (SYS_io_destroy, ctx_id);
95			}
96
97	root	1.6	inline_size
98			int
99	root	1.1	ev_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
100			{
101			return syscall (SYS_io_submit, ctx_id, nr, cbp);
102			}
103
104	root	1.6	inline_size
105			int
106	root	1.1	ev_io_cancel (aio_context_t ctx_id, struct iocb cbp, struct io_event result)
107			{
108			return syscall (SYS_io_cancel, ctx_id, cbp, result);
109			}
110
111	root	1.6	inline_size
112			int
113	root	1.1	ev_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event events, struct timespec timeout)
114			{
115			return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
116			}
117
118			/*****************************************************************************/
119			/* actual backed implementation */
120
121	root	1.7	/* we use out own wrapper structure in acse we ever want to do something "clever" */
122	root	1.1	typedef struct aniocb
123			{
124			struct iocb io;
125			/int inuse;/
126			} *ANIOCBP;
127
128			inline_size
129			void
130			linuxaio_array_needsize_iocbp (ANIOCBP *base, int count)
131			{
132	root	1.7	/* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */
133	root	1.1	while (count--)
134			{
135			base = (ANIOCBP)ev_malloc (sizeof (*base));
136	root	1.2	/* TODO: full zero initialize required? */
137	root	1.1	memset (base, 0, sizeof (*base));
138	root	1.7	/* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */
139	root	1.1	(*base)->io.aio_lio_opcode = IOCB_CMD_POLL;
140			++base;
141			}
142			}
143
144	root	1.6	ecb_cold
145	root	1.1	static void
146			linuxaio_free_iocbp (EV_P)
147			{
148			while (linuxaio_iocbpmax--)
149			ev_free (linuxaio_iocbps [linuxaio_iocbpmax]);
150
151	root	1.6	linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */
152	root	1.1	}
153
154			static void
155			linuxaio_modify (EV_P_ int fd, int oev, int nev)
156			{
157			array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
158			struct aniocb *iocb = linuxaio_iocbps [fd];
159
160	root	1.10	#if EPOLL_FALLBACK
161			if (iocb->io.aio_reqprio < 0)
162			{
163			epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
164			iocb->io.aio_reqprio = 0;
165			}
166			#endif
167
168	root	1.1	if (iocb->io.aio_buf)
169	root	1.4	ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event )0); / always returns an error relevant kernels */
170	root	1.1
171			if (nev)
172			{
173			iocb->io.aio_data = fd;
174			iocb->io.aio_fildes = fd;
175			iocb->io.aio_buf =
176			(nev & EV_READ ? POLLIN : 0)
177			\| (nev & EV_WRITE ? POLLOUT : 0);
178
179			/* queue iocb up for io_submit */
180			/* this assumes we only ever get one call per fd per loop iteration */
181			++linuxaio_submitcnt;
182			array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
183			linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
184			}
185			}
186
187			static void
188			linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
189			{
190			while (nr)
191			{
192			int fd = ev->data;
193			int res = ev->res;
194
195	root	1.2	assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
196	root	1.1
197			/* linux aio is oneshot: rearm fd */
198			linuxaio_iocbps [fd]->io.aio_buf = 0;
199			anfds [fd].events = 0;
200			fd_change (EV_A_ fd, 0);
201
202			/* feed events, we do not expect or handle POLLNVAL */
203			if (ecb_expect_false (res & POLLNVAL))
204			fd_kill (EV_A_ fd);
205			else
206			fd_event (
207			EV_A_
208			fd,
209			(res & (POLLOUT \| POLLERR \| POLLHUP) ? EV_WRITE : 0)
210			\| (res & (POLLIN \| POLLERR \| POLLHUP) ? EV_READ : 0)
211			);
212
213			--nr;
214			++ev;
215			}
216			}
217
218			/* get any events from ringbuffer, return true if any were handled */
219			static int
220			linuxaio_get_events_from_ring (EV_P)
221			{
222			struct aio_ring ring = (struct aio_ring )linuxaio_ctx;
223
224	root	1.13	/* the kernel reads and writes both of these variables, */
225			/* as a C extension, we assume that volatile use here */
226			/* both makes reads atomic and once-only */
227			unsigned head = (volatile unsigned )&ring->head;
228	root	1.1	unsigned tail = (volatile unsigned )&ring->tail;
229
230	root	1.6	if (head == tail)
231			return 0;
232
233	root	1.7	/* bail out if the ring buffer doesn't match the expected layout */
234	root	1.6	if (ecb_expect_false (ring->magic != AIO_RING_MAGIC)
235			\|\| ring->incompat_features != AIO_RING_INCOMPAT_FEATURES
236			\|\| ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */
237	root	1.1	return 0;
238
239	root	1.12	/* make sure the events up to tail are visible */
240	root	1.9	ECB_MEMORY_FENCE_ACQUIRE;
241
242	root	1.1	/* parse all available events, but only once, to avoid starvation */
243			if (tail > head) /* normal case around */
244			linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
245	root	1.6	else /* wrapped around */
246	root	1.1	{
247			linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
248			linuxaio_parse_events (EV_A_ ring->io_events, tail);
249			}
250
251	root	1.16	/* TODO: we only need a compiler barrier here, not a read fence */
252			ECB_MEMORY_FENCE_RELEASE;
253			/* as an extension to C, we hope that the volatile will make this atomic and once-only */
254	root	1.11	(volatile unsigned )&ring->head = tail;
255	root	1.12	/* make sure kernel can see our new head value - probably not required */
256	root	1.11	ECB_MEMORY_FENCE_RELEASE;
257	root	1.1
258			return 1;
259			}
260
261			/* read at least one event from kernel, or timeout */
262			inline_size
263			void
264			linuxaio_get_events (EV_P_ ev_tstamp timeout)
265			{
266			struct timespec ts;
267			struct io_event ioev;
268			int res;
269
270			if (linuxaio_get_events_from_ring (EV_A))
271			return;
272
273			/* no events, so wait for at least one, then poll ring buffer again */
274	root	1.7	/* this degrades to one event per loop iteration */
275	root	1.1	/* if the ring buffer changes layout, but so be it */
276
277			ts.tv_sec = (long)timeout;
278			ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
279
280			res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts);
281
282			if (res < 0)
283	root	1.10	if (errno == EINTR)
284			/* ignored */;
285			else
286			ev_syserr ("(libev) linuxaio io_getevents");
287	root	1.1	else if (res)
288			{
289			/* at least one event received, handle it and any remaining ones in the ring buffer */
290			linuxaio_parse_events (EV_A_ &ioev, 1);
291			linuxaio_get_events_from_ring (EV_A);
292			}
293			}
294
295	root	1.10	#if EPOLL_FALLBACK
296			static void
297			linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op)
298			{
299			struct epoll_event eev;
300
301			eev.events = EPOLLONESHOT;
302			if (iocb->aio_buf & POLLIN ) eev.events \|= EPOLLIN ;
303			if (iocb->aio_buf & POLLOUT) eev.events \|= EPOLLOUT;
304			eev.data.fd = iocb->aio_fildes;
305
306			if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0)
307			ev_syserr ("(libeio) linuxaio epoll_ctl");
308			}
309			#endif
310
311	root	1.1	static void
312			linuxaio_poll (EV_P_ ev_tstamp timeout)
313			{
314			int submitted;
315
316			/* first phase: submit new iocbs */
317
318			/* io_submit might return less than the requested number of iocbs */
319			/* this is, afaics, only because of errors, but we go by the book and use a loop, */
320			/* which allows us to pinpoint the errornous iocb */
321			for (submitted = 0; submitted < linuxaio_submitcnt; )
322			{
323			int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
324
325	root	1.6	if (ecb_expect_false (res < 0))
326	root	1.1	if (errno == EAGAIN)
327			{
328			/* This happens when the ring buffer is full, at least. I assume this means
329			* that the event was queued synchronously during io_submit, and thus
330	root	1.10	* the buffer overflowed.
331			* In this case, we just try in next loop iteration.
332	root	1.7	* This should not result in a few fds taking priority, as the interface
333			* is one-shot, and we submit iocb's in a round-robin fashion.
334	root	1.1	*/
335	root	1.5	memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits));
336	root	1.1	linuxaio_submitcnt -= submitted;
337			timeout = 0;
338			break;
339			}
340	root	1.10	#if EPOLL_FALLBACK
341			else if (errno == EINVAL)
342			{
343	root	1.15	/* This happens for unsupported fds, officially, but in my testing,
344	root	1.10	* also randomly happens for supported fds. We fall back to good old
345			* poll() here, under the assumption that this is a very rare case.
346	root	1.15	* See https://lore.kernel.org/patchwork/patch/1047453/ for evidence
347			* that the problem is known, but ignored.
348	root	1.10	*/
349			struct iocb *iocb = linuxaio_submits [submitted];
350			res = 1; /* skip this iocb */
351
352			linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD);
353			iocb->aio_reqprio = -1; /* mark iocb as epoll */
354			}
355			#endif
356	root	1.1	else
357	root	1.8	ev_syserr ("(libev) linuxaio io_submit");
358	root	1.1
359			submitted += res;
360			}
361
362			linuxaio_submitcnt = 0;
363
364			/* second phase: fetch and parse events */
365
366			linuxaio_get_events (EV_A_ timeout);
367			}
368
369	root	1.10	#if EPOLL_FALLBACK
370
371			static void
372			linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
373			{
374			struct epoll_event events[16];
375
376			for (;;)
377			{
378			int idx;
379			int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0);
380
381			if (ecb_expect_false (res < 0))
382			ev_syserr ("(libev) linuxaio epoll_wait");
383			else if (!res)
384			break;
385
386			for (idx = res; idx--; )
387			{
388			int fd = events [idx].data.fd;
389			uint32_t ev = events [idx].events;
390
391			assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
392
393			linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD);
394
395			fd_event (EV_A_ fd,
396			(ev & (EPOLLOUT \| EPOLLERR \| EPOLLHUP) ? EV_WRITE : 0)
397			\| (ev & (EPOLLIN \| EPOLLERR \| EPOLLHUP) ? EV_READ : 0));
398			}
399
400			if (res < sizeof (events) / sizeof (events [0]))
401			break;
402			}
403			}
404
405			#endif
406
407	root	1.1	inline_size
408			int
409			linuxaio_init (EV_P_ int flags)
410			{
411			/* would be great to have a nice test for IOCB_CMD_POLL instead */
412	root	1.2	/* also: test some semi-common fd types, such as files and ttys in recommended_backends */
413	root	1.15	#if EPOLL_FALLBACK
414			/* 4.19 made epoll work */
415			if (ev_linux_version () < 0x041300)
416			return 0;
417			#else
418			/* 4.18 introduced IOCB_CMD_POLL */
419			if (ev_linux_version () < 0x041200)
420	root	1.1	return 0;
421	root	1.15	#endif
422	root	1.1
423	root	1.3	linuxaio_ctx = 0;
424	root	1.1	if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
425			return 0;
426
427	root	1.10	#if EPOLL_FALLBACK
428			backend_fd = ev_epoll_create ();
429			if (backend_fd < 0)
430			{
431			ev_io_destroy (linuxaio_ctx);
432			return 0;
433			}
434
435			ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
436			ev_io_start (EV_A_ &linuxaio_epoll_w);
437	root	1.14	ev_unref (EV_A); /* watcher should not keep loop alive */
438	root	1.10	#endif
439
440	root	1.1	backend_modify = linuxaio_modify;
441			backend_poll = linuxaio_poll;
442
443			linuxaio_iocbpmax = 0;
444			linuxaio_iocbps = 0;
445
446			linuxaio_submits = 0;
447			linuxaio_submitmax = 0;
448			linuxaio_submitcnt = 0;
449
450			return EVBACKEND_LINUXAIO;
451			}
452
453			inline_size
454			void
455			linuxaio_destroy (EV_P)
456			{
457	root	1.10	#if EPOLL_FALLBACK
458			close (backend_fd);
459			#endif
460	root	1.1	linuxaio_free_iocbp (EV_A);
461			ev_io_destroy (linuxaio_ctx);
462			}
463
464			inline_size
465			void
466			linuxaio_fork (EV_P)
467			{
468	root	1.6	/* this frees all iocbs, which is very heavy-handed */
469	root	1.2	linuxaio_destroy (EV_A);
470	root	1.6	linuxaio_submitcnt = 0; /* all pointers were invalidated */
471	root	1.2
472	root	1.3	linuxaio_ctx = 0;
473	root	1.2	while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0)
474	root	1.8	ev_syserr ("(libev) linuxaio io_setup");
475	root	1.2
476	root	1.10	#if EPOLL_FALLBACK
477			while ((backend_fd = ev_epoll_create ()) < 0)
478			ev_syserr ("(libev) linuxaio epoll_create");
479
480			ev_io_stop (EV_A_ &linuxaio_epoll_w);
481			ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
482			ev_io_start (EV_A_ &linuxaio_epoll_w);
483			#endif
484
485	root	1.2	fd_rearm_all (EV_A);
486	root	1.1	}
487