libev/ev_iouring.c

/*
 * libev linux io_uring 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 io_uring:
 *
 * a) it's the best interface I have seen so far. on linux.
 * b) best is not necessarily very good.
 * c) it's better than the aio mess, doesn't suffer from the fork problems
 *    of linux aio or epoll and so on and so on. and you could do event stuff
 *    without any syscalls. what's not to like?
 * d) ok, it's vastly more complex, but that's ok, really.
 * e) why 3 mmaps instead of one? one would be more space-efficient,
 *    and I can't see what benefit three would have (other than being
 *    somehow resizable/relocatable, but that's apparently not possible).
 *    (FIXME: newer kernels can use 2 mmaps only, need to look into this).
 * f) hmm, it's practiclaly undebuggable (gdb can't access the memory, and
 *    the bizarre way structure offsets are communicated makes it hard to
 *    just print the ring buffer heads, even *iff* the memory were visible
 *    in gdb. but then, that's also ok, really.
 * g) well, you cannot specify a timeout when waiting for events. no,
 *    seriously, the interface doesn't support a timeout. never seen _that_
 *    before. sure, you can use a timerfd, but that's another syscall
 *    you could have avoided. overall, this bizarre omission smells
 *    like a µ-optimisation by the io_uring author for his personal
 *    applications, to the detriment of everybody else who just wants
 *    an event loop. but, umm, ok, if that's all, it could be worse.
 *    (FIXME: jens mentioned timeout commands, need to investigate)
 * h) there is a hardcoded limit of 4096 outstanding events. okay,
 *    at least there is no arbitrary low system-wide limit...
 *    (FIXME: apparently, this was increased to 32768 in later kernels(
 * i) unlike linux aio, you *can* register more then the limit
 *    of fd events, and the kernel will "gracefully" signal an
 *    overflow, after which you could destroy and recreate the kernel
 *    state, a bit bigger, or fall back to e.g. poll. thats not
 *    totally insane, but kind of questions the point a high
 *    performance I/O framework when it doesn't really work
 *    under stress.
 *    (FIXME: iouring should no longer drop events, need to investigate)
 * j) but, oh my! is has exactly the same bugs as the linux aio backend,
 *    where some undocumented poll combinations just fail.
 *    so we need epoll AGAIN as a fallback. AGAIN! epoll!! and of course,
 *    this is completely undocumented, have I mantioned this already?
 * k) overall, the *API* itself is, I dare to say, not a total trainwreck.
 *    the big isuess with it are the bugs requiring epoll, which might
 *    or might not get fixed (do I hold my breath?).
 */

/* TODO: use internal TIMEOUT */
/* TODO: take advantage of single mmap, NODROP etc. */
/* TODO: resize cq/sq size independently */

#include <sys/timerfd.h>
#include <sys/mman.h>
#include <poll.h>
#include <stdint.h>

#define IOURING_INIT_ENTRIES 32

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

#include <linux/fs.h>
#include <linux/types.h>

/* mostly directly taken from the kernel or documentation */

struct io_uring_sqe
{
  __u8 opcode;
  __u8 flags;
  __u16 ioprio;
  __s32 fd;
  union {
    __u64 off;
    __u64 addr2;
  };
  __u64 addr;
  __u32 len;
  union {
    __kernel_rwf_t rw_flags;
    __u32 fsync_flags;
    __u16 poll_events;
    __u32 sync_range_flags;
    __u32 msg_flags;
    __u32 timeout_flags;
    __u32 accept_flags;
    __u32 cancel_flags;
    __u32 open_flags;
    __u32 statx_flags;
  };
  __u64 user_data;
  union {
    __u16 buf_index;
    __u64 __pad2[3];
  };
};

struct io_uring_cqe
{
  __u64 user_data;
  __s32 res;
  __u32 flags;
};

struct io_sqring_offsets
{
  __u32 head;
  __u32 tail;
  __u32 ring_mask;
  __u32 ring_entries;
  __u32 flags;
  __u32 dropped;
  __u32 array;
  __u32 resv1;
  __u64 resv2;
};

struct io_cqring_offsets
{
  __u32 head;
  __u32 tail;
  __u32 ring_mask;
  __u32 ring_entries;
  __u32 overflow;
  __u32 cqes;
  __u64 resv[2];
};

struct io_uring_params
{
  __u32 sq_entries;
  __u32 cq_entries;
  __u32 flags;
  __u32 sq_thread_cpu;
  __u32 sq_thread_idle;
  __u32 features;
  __u32 resv[4];
  struct io_sqring_offsets sq_off;
  struct io_cqring_offsets cq_off;
};

#define IORING_SETUP_CQSIZE 0x00000008

#define IORING_OP_POLL_ADD        6
#define IORING_OP_POLL_REMOVE     7
#define IORING_OP_TIMEOUT        11
#define IORING_OP_TIMEOUT_REMOVE 12

/* relative or absolute, reference clock is CLOCK_MONOTONIC */
struct iouring_kernel_timespec
{
  int64_t tv_sec;
  long long tv_nsec;
};

#define IORING_TIMEOUT_ABS 0x00000001

#define IORING_ENTER_GETEVENTS 0x01

#define IORING_OFF_SQ_RING 0x00000000ULL
#define IORING_OFF_CQ_RING 0x08000000ULL
#define IORING_OFF_SQES    0x10000000ULL

#define IORING_FEAT_SINGLE_MMAP   0x00000001
#define IORING_FEAT_NODROP        0x00000002
#define IORING_FEAT_SUBMIT_STABLE 0x00000004

inline_size
int
evsys_io_uring_setup (unsigned entries, struct io_uring_params *params)
{
  return ev_syscall2 (SYS_io_uring_setup, entries, params);
}

inline_size
int
evsys_io_uring_enter (int fd, unsigned to_submit, unsigned min_complete, unsigned flags, const sigset_t *sig, size_t sigsz)
{
  return ev_syscall6 (SYS_io_uring_enter, fd, to_submit, min_complete, flags, sig, sigsz);
}

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

/* we hope that volatile will make the compiler access this variables only once */
#define EV_SQ_VAR(name) *(volatile unsigned *)((char *)iouring_sq_ring + iouring_sq_ ## name)
#define EV_CQ_VAR(name) *(volatile unsigned *)((char *)iouring_cq_ring + iouring_cq_ ## name)

/* the index array */
#define EV_SQ_ARRAY     ((unsigned *)((char *)iouring_sq_ring + iouring_sq_array))

/* the submit/completion queue entries */
#define EV_SQES         ((struct io_uring_sqe *)         iouring_sqes)
#define EV_CQES         ((struct io_uring_cqe *)((char *)iouring_cq_ring + iouring_cq_cqes))

static
struct io_uring_sqe *
iouring_sqe_get (EV_P)
{
  unsigned tail = EV_SQ_VAR (tail);

  if (tail + 1 - EV_SQ_VAR (head) > EV_SQ_VAR (ring_entries))
    {
      /* queue full, flush */
      evsys_io_uring_enter (iouring_fd, iouring_to_submit, 0, 0, 0, 0);
      iouring_to_submit = 0;
    }

  assert (("libev: io_uring queue full after flush", tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)));

  return EV_SQES + (tail & EV_SQ_VAR (ring_mask));
}

inline_size
struct io_uring_sqe *
iouring_sqe_submit (EV_P_ struct io_uring_sqe *sqe)
{
  unsigned idx = sqe - EV_SQES;

  EV_SQ_ARRAY [idx] = idx;
  ECB_MEMORY_FENCE_RELEASE;
  ++EV_SQ_VAR (tail);
  /*ECB_MEMORY_FENCE_RELEASE; /* for the time being we assume this is not needed */
  ++iouring_to_submit;
}

/*****************************************************************************/

/* when the timerfd expires we simply note the fact,
 * as the purpose of the timerfd is to wake us up, nothing else.
 * the next iteration should re-set it.
 */
static void
iouring_tfd_cb (EV_P_ struct ev_io *w, int revents)
{
  iouring_tfd_to = EV_TSTAMP_HUGE;
}

/* called for full and partial cleanup */
ecb_cold
static int
iouring_internal_destroy (EV_P)
{
  close (iouring_tfd);
  close (iouring_fd);

  if (iouring_sq_ring != MAP_FAILED) munmap (iouring_sq_ring, iouring_sq_ring_size);
  if (iouring_cq_ring != MAP_FAILED) munmap (iouring_cq_ring, iouring_cq_ring_size);
  if (iouring_sqes    != MAP_FAILED) munmap (iouring_sqes   , iouring_sqes_size   );

  if (ev_is_active (&iouring_tfd_w))
    {
      ev_ref (EV_A);
      ev_io_stop (EV_A_ &iouring_tfd_w);
    }
}

ecb_cold
static int
iouring_internal_init (EV_P)
{
  struct io_uring_params params = { 0 };

  iouring_to_submit = 0;

  iouring_tfd     = -1;
  iouring_sq_ring = MAP_FAILED;
  iouring_cq_ring = MAP_FAILED;
  iouring_sqes    = MAP_FAILED;

  if (!have_monotonic) /* cannot really happen, but what if11 */
    return -1;

  for (;;)
    {
      iouring_fd = evsys_io_uring_setup (iouring_entries, &params);

      if (iouring_fd >= 0)
        break; /* yippie */

      if (errno != EINVAL)
        return -1; /* we failed */

#if TODO
      if ((~params.features) & (IORING_FEAT_NODROP | IORING_FEATURE_SINGLE_MMAP))
        return -1; /* we require the above features */
#endif

      /* EINVAL: lots of possible reasons, but maybe
       * it is because we hit the unqueryable hardcoded size limit
       */

      /* we hit the limit already, give up */
      if (iouring_max_entries)
        return -1;

      /* first time we hit EINVAL? assume we hit the limit, so go back and retry */
      iouring_entries >>= 1;
      iouring_max_entries = iouring_entries;
    }

  iouring_sq_ring_size = params.sq_off.array + params.sq_entries * sizeof (unsigned);
  iouring_cq_ring_size = params.cq_off.cqes  + params.cq_entries * sizeof (struct io_uring_cqe);
  iouring_sqes_size    =                       params.sq_entries * sizeof (struct io_uring_sqe);

  iouring_sq_ring = mmap (0, iouring_sq_ring_size, PROT_READ | PROT_WRITE,
                          MAP_SHARED | MAP_POPULATE, iouring_fd, IORING_OFF_SQ_RING);
  iouring_cq_ring = mmap (0, iouring_cq_ring_size, PROT_READ | PROT_WRITE,
                          MAP_SHARED | MAP_POPULATE, iouring_fd, IORING_OFF_CQ_RING);
  iouring_sqes    = mmap (0, iouring_sqes_size, PROT_READ | PROT_WRITE,
                          MAP_SHARED | MAP_POPULATE, iouring_fd, IORING_OFF_SQES);

  if (iouring_sq_ring == MAP_FAILED || iouring_cq_ring == MAP_FAILED || iouring_sqes == MAP_FAILED)
    return -1;

  iouring_sq_head         = params.sq_off.head;
  iouring_sq_tail         = params.sq_off.tail;
  iouring_sq_ring_mask    = params.sq_off.ring_mask;
  iouring_sq_ring_entries = params.sq_off.ring_entries;
  iouring_sq_flags        = params.sq_off.flags;
  iouring_sq_dropped      = params.sq_off.dropped;
  iouring_sq_array        = params.sq_off.array;

  iouring_cq_head         = params.cq_off.head;
  iouring_cq_tail         = params.cq_off.tail;
  iouring_cq_ring_mask    = params.cq_off.ring_mask;
  iouring_cq_ring_entries = params.cq_off.ring_entries;
  iouring_cq_overflow     = params.cq_off.overflow;
  iouring_cq_cqes         = params.cq_off.cqes;

  iouring_tfd = timerfd_create (CLOCK_MONOTONIC, TFD_CLOEXEC);

  if (iouring_tfd < 0)
    return iouring_tfd;

  iouring_tfd_to = EV_TSTAMP_HUGE;

  return 0;
}

ecb_cold
static void
iouring_fork (EV_P)
{
  iouring_internal_destroy (EV_A);

  while (iouring_internal_init (EV_A) < 0)
    ev_syserr ("(libev) io_uring_setup");

  fd_rearm_all (EV_A);

  ev_io_stop  (EV_A_ &iouring_tfd_w);
  ev_io_set   (EV_A_ &iouring_tfd_w, iouring_tfd, EV_READ);
  ev_io_start (EV_A_ &iouring_tfd_w);
}

/*****************************************************************************/

static void
iouring_modify (EV_P_ int fd, int oev, int nev)
{
  if (oev)
    {
      /* we assume the sqe's are all "properly" initialised */
      struct io_uring_sqe *sqe = iouring_sqe_get (EV_A);
      sqe->opcode    = IORING_OP_POLL_REMOVE;
      sqe->fd        = fd;
      /* Jens Axboe notified me that user_data is not what is documented, but is
       * some kind of unique ID that has to match, otherwise the request cannot
       * be removed. Since we don't *really* have that, we pass in the old
       * generation counter - if that fails, too bad, it will hopefully be removed
       * at close time and then be ignored. */
      sqe->user_data = (uint32_t)fd | ((__u64)(uint32_t)anfds [fd].egen << 32);
      iouring_sqe_submit (EV_A_ sqe);

      /* increment generation counter to avoid handling old events */
      ++anfds [fd].egen;
    }

  if (nev)
    {
      struct io_uring_sqe *sqe = iouring_sqe_get (EV_A);
      sqe->opcode      = IORING_OP_POLL_ADD;
      sqe->fd          = fd;
      sqe->user_data   = (uint32_t)fd | ((__u64)(uint32_t)anfds [fd].egen << 32);
      sqe->poll_events =
        (nev & EV_READ ? POLLIN : 0)
        | (nev & EV_WRITE ? POLLOUT : 0);
      iouring_sqe_submit (EV_A_ sqe);
    }
}

inline_size
void
iouring_tfd_update (EV_P_ ev_tstamp timeout)
{
  ev_tstamp tfd_to = mn_now + timeout;

  /* we assume there will be many iterations per timer change, so
   * we only re-set the timerfd when we have to because its expiry
   * is too late.
   */
  if (ecb_expect_false (tfd_to < iouring_tfd_to))
    {
       struct itimerspec its;

       iouring_tfd_to = tfd_to;
       EV_TS_SET (its.it_interval, 0.);
       EV_TS_SET (its.it_value, tfd_to);

       if (timerfd_settime (iouring_tfd, TFD_TIMER_ABSTIME, &its, 0) < 0)
         assert (("libev: iouring timerfd_settime failed", 0));
    }
}

inline_size
void
iouring_process_cqe (EV_P_ struct io_uring_cqe *cqe)
{
  int      fd  = cqe->user_data & 0xffffffffU;
  uint32_t gen = cqe->user_data >> 32;
  int      res = cqe->res;

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

  /* documentation lies, of course. the result value is NOT like
   * normal syscalls, but like linux raw syscalls, i.e. negative
   * error numbers. fortunate, as otherwise there would be no way
   * to get error codes at all. still, why not document this?
   */

  /* ignore event if generation doesn't match */
  /* other than skipping removal events, */
  /* this should actually be very rare */
  if (ecb_expect_false (gen != (uint32_t)anfds [fd].egen))
    return;

  if (ecb_expect_false (res < 0))
    {
      /*TODO: EINVAL handling (was something failed with this fd)*/
      /*TODO: EBUSY happens when?*/

      if (res == -EBADF)
        {
          assert (("libev: event loop rejected bad fd", res != -EBADF));
          fd_kill (EV_A_ fd);
        }
      else
        {
          errno = -res;
          ev_syserr ("(libev) IORING_OP_POLL_ADD");
        }

      return;
    }

  /* 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)
  );

  /* io_uring is oneshot, so we need to re-arm the fd next iteration */
  /* this also means we usually have to do at least one syscall per iteration */
  anfds [fd].events = 0;
  fd_change (EV_A_ fd, EV_ANFD_REIFY);
}

/* called when the event queue overflows */
ecb_cold
static void
iouring_overflow (EV_P)
{
  /* we have two options, resize the queue (by tearing down
   * everything and recreating it, or living with it
   * and polling.
   * we implement this by resizing the queue, and, if that fails,
   * we just recreate the state on every failure, which
   * kind of is a very inefficient poll.
   * one danger is, due to the bios toward lower fds,
   * we will only really get events for those, so
   * maybe we need a poll() fallback, after all.
   */
  /*EV_CQ_VAR (overflow) = 0;*/ /* need to do this if we keep the state and poll manually */

  fd_rearm_all (EV_A);

  /* we double the size until we hit the hard-to-probe maximum */
  if (!iouring_max_entries)
    {
      iouring_entries <<= 1;
      iouring_fork (EV_A);
    }
  else
    {
      /* we hit the kernel limit, we should fall back to something else.
       * we can either poll() a few times and hope for the best,
       * poll always, or switch to epoll.
       * TODO: is this necessary with newer kernels?
       */

      iouring_internal_destroy (EV_A);

      /* this should make it so that on return, we don't call any uring functions */
      iouring_to_submit = 0;

      for (;;)
        {
          backend = epoll_init (EV_A_ 0);

          if (backend)
            break;

          ev_syserr ("(libev) iouring switch to epoll");
        }
    }
}

/* handle any events in the completion queue, return true if there were any */
static int
iouring_handle_cq (EV_P)
{
  unsigned head, tail, mask;
  
  head = EV_CQ_VAR (head);
  ECB_MEMORY_FENCE_ACQUIRE;
  tail = EV_CQ_VAR (tail);

  if (head == tail)
    return 0;

  /* it can only overflow if we have events, yes, yes? */
  if (ecb_expect_false (EV_CQ_VAR (overflow)))
    {
      iouring_overflow (EV_A);
      return 1;
    }

  mask = EV_CQ_VAR (ring_mask);

  do
    iouring_process_cqe (EV_A_ &EV_CQES [head++ & mask]);
  while (head != tail);

  EV_CQ_VAR (head) = head;
  ECB_MEMORY_FENCE_RELEASE;

  return 1;
}

static void
iouring_poll (EV_P_ ev_tstamp timeout)
{
  /* if we have events, no need for extra syscalls, but we might have to queue events */
  if (iouring_handle_cq (EV_A))
    timeout = EV_TS_CONST (0.);
  else
    /* no events, so maybe wait for some */
    iouring_tfd_update (EV_A_ timeout);

  /* only enter the kernel if we have something to submit, or we need to wait */
  if (timeout || iouring_to_submit)
    {
      int res;

      EV_RELEASE_CB;

      res = evsys_io_uring_enter (iouring_fd, iouring_to_submit, 1,
                                  timeout > EV_TS_CONST (0.) ? IORING_ENTER_GETEVENTS : 0, 0, 0);
      iouring_to_submit = 0;

      EV_ACQUIRE_CB;

      if (ecb_expect_false (res < 0))
        if (errno == EINTR)
          /* ignore */;
        else
          ev_syserr ("(libev) iouring setup");
      else
        iouring_handle_cq (EV_A);
    }
}

inline_size
int
iouring_init (EV_P_ int flags)
{
  iouring_entries     = IOURING_INIT_ENTRIES;
  iouring_max_entries = 0;

  if (iouring_internal_init (EV_A) < 0)
    {
      iouring_internal_destroy (EV_A);
      return 0;
    }

  ev_io_init  (&iouring_tfd_w, iouring_tfd_cb, iouring_tfd, EV_READ);
  ev_set_priority (&iouring_tfd_w, EV_MINPRI);
  ev_io_start (EV_A_ &iouring_tfd_w);
  ev_unref (EV_A); /* watcher should not keep loop alive */

  backend_modify = iouring_modify;
  backend_poll   = iouring_poll;

  return EVBACKEND_IOURING;
}

inline_size
void
iouring_destroy (EV_P)
{
  iouring_internal_destroy (EV_A);
}

Revision:	1.13
Committed:	Sat Dec 28 03:29:50 2019 UTC (4 years, 4 months ago) by root
Content type:	text/plain
Branch:	MAIN
Changes since 1.12:	+7 -6 lines
Log Message:	* empty log message *
#	User	Rev	Content
1	root	1.1	/*
2			* libev linux io_uring 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 io_uring:
42			*
43			* a) it's the best interface I have seen so far. on linux.
44			* b) best is not necessarily very good.
45			* c) it's better than the aio mess, doesn't suffer from the fork problems
46			* of linux aio or epoll and so on and so on. and you could do event stuff
47			* without any syscalls. what's not to like?
48			* d) ok, it's vastly more complex, but that's ok, really.
49			* e) why 3 mmaps instead of one? one would be more space-efficient,
50			* and I can't see what benefit three would have (other than being
51			* somehow resizable/relocatable, but that's apparently not possible).
52	root	1.7	* (FIXME: newer kernels can use 2 mmaps only, need to look into this).
53	root	1.1	* f) hmm, it's practiclaly undebuggable (gdb can't access the memory, and
54	root	1.7	* the bizarre way structure offsets are communicated makes it hard to
55	root	1.1	* just print the ring buffer heads, even iff the memory were visible
56			* in gdb. but then, that's also ok, really.
57			* g) well, you cannot specify a timeout when waiting for events. no,
58			* seriously, the interface doesn't support a timeout. never seen _that_
59			* before. sure, you can use a timerfd, but that's another syscall
60			* you could have avoided. overall, this bizarre omission smells
61			* like a µ-optimisation by the io_uring author for his personal
62			* applications, to the detriment of everybody else who just wants
63			* an event loop. but, umm, ok, if that's all, it could be worse.
64	root	1.7	* (FIXME: jens mentioned timeout commands, need to investigate)
65	root	1.1	* h) there is a hardcoded limit of 4096 outstanding events. okay,
66			* at least there is no arbitrary low system-wide limit...
67	root	1.7	* (FIXME: apparently, this was increased to 32768 in later kernels(
68	root	1.1	* i) unlike linux aio, you can register more then the limit
69			* of fd events, and the kernel will "gracefully" signal an
70			* overflow, after which you could destroy and recreate the kernel
71			* state, a bit bigger, or fall back to e.g. poll. thats not
72			* totally insane, but kind of questions the point a high
73			* performance I/O framework when it doesn't really work
74			* under stress.
75	root	1.7	* (FIXME: iouring should no longer drop events, need to investigate)
76	root	1.1	* j) but, oh my! is has exactly the same bugs as the linux aio backend,
77			* where some undocumented poll combinations just fail.
78			* so we need epoll AGAIN as a fallback. AGAIN! epoll!! and of course,
79			* this is completely undocumented, have I mantioned this already?
80			* k) overall, the API itself is, I dare to say, not a total trainwreck.
81			* the big isuess with it are the bugs requiring epoll, which might
82			* or might not get fixed (do I hold my breath?).
83			*/
84
85	root	1.9	/* TODO: use internal TIMEOUT */
86			/* TODO: take advantage of single mmap, NODROP etc. */
87			/* TODO: resize cq/sq size independently */
88
89	root	1.1	#include <sys/timerfd.h>
90			#include <sys/mman.h>
91			#include <poll.h>
92	root	1.12	#include <stdint.h>
93	root	1.1
94			#define IOURING_INIT_ENTRIES 32
95
96			/*****************************************************************************/
97			/* syscall wrapdadoop - this section has the raw api/abi definitions */
98
99			#include <linux/fs.h>
100			#include <linux/types.h>
101
102			/* mostly directly taken from the kernel or documentation */
103
104			struct io_uring_sqe
105			{
106			__u8 opcode;
107			__u8 flags;
108			__u16 ioprio;
109			__s32 fd;
110	root	1.8	union {
111			__u64 off;
112			__u64 addr2;
113			};
114	root	1.1	__u64 addr;
115			__u32 len;
116			union {
117			__kernel_rwf_t rw_flags;
118			__u32 fsync_flags;
119			__u16 poll_events;
120			__u32 sync_range_flags;
121			__u32 msg_flags;
122	root	1.8	__u32 timeout_flags;
123			__u32 accept_flags;
124			__u32 cancel_flags;
125			__u32 open_flags;
126			__u32 statx_flags;
127	root	1.1	};
128			__u64 user_data;
129			union {
130			__u16 buf_index;
131			__u64 __pad2[3];
132			};
133			};
134
135			struct io_uring_cqe
136			{
137			__u64 user_data;
138			__s32 res;
139			__u32 flags;
140			};
141
142			struct io_sqring_offsets
143			{
144			__u32 head;
145			__u32 tail;
146			__u32 ring_mask;
147			__u32 ring_entries;
148			__u32 flags;
149			__u32 dropped;
150			__u32 array;
151			__u32 resv1;
152			__u64 resv2;
153			};
154
155			struct io_cqring_offsets
156			{
157			__u32 head;
158			__u32 tail;
159			__u32 ring_mask;
160			__u32 ring_entries;
161			__u32 overflow;
162			__u32 cqes;
163			__u64 resv[2];
164			};
165
166			struct io_uring_params
167			{
168			__u32 sq_entries;
169			__u32 cq_entries;
170			__u32 flags;
171			__u32 sq_thread_cpu;
172			__u32 sq_thread_idle;
173	root	1.8	__u32 features;
174			__u32 resv[4];
175	root	1.1	struct io_sqring_offsets sq_off;
176			struct io_cqring_offsets cq_off;
177			};
178
179	root	1.12	#define IORING_SETUP_CQSIZE 0x00000008
180
181			#define IORING_OP_POLL_ADD 6
182			#define IORING_OP_POLL_REMOVE 7
183			#define IORING_OP_TIMEOUT 11
184			#define IORING_OP_TIMEOUT_REMOVE 12
185
186			/* relative or absolute, reference clock is CLOCK_MONOTONIC */
187			struct iouring_kernel_timespec
188			{
189			int64_t tv_sec;
190			long long tv_nsec;
191			};
192
193			#define IORING_TIMEOUT_ABS 0x00000001
194	root	1.1
195			#define IORING_ENTER_GETEVENTS 0x01
196
197			#define IORING_OFF_SQ_RING 0x00000000ULL
198			#define IORING_OFF_CQ_RING 0x08000000ULL
199			#define IORING_OFF_SQES 0x10000000ULL
200
201	root	1.12	#define IORING_FEAT_SINGLE_MMAP 0x00000001
202			#define IORING_FEAT_NODROP 0x00000002
203			#define IORING_FEAT_SUBMIT_STABLE 0x00000004
204	root	1.8
205	root	1.1	inline_size
206			int
207			evsys_io_uring_setup (unsigned entries, struct io_uring_params *params)
208			{
209			return ev_syscall2 (SYS_io_uring_setup, entries, params);
210			}
211
212			inline_size
213			int
214			evsys_io_uring_enter (int fd, unsigned to_submit, unsigned min_complete, unsigned flags, const sigset_t *sig, size_t sigsz)
215			{
216			return ev_syscall6 (SYS_io_uring_enter, fd, to_submit, min_complete, flags, sig, sigsz);
217			}
218
219			/*****************************************************************************/
220			/* actual backed implementation */
221
222			/* we hope that volatile will make the compiler access this variables only once */
223			#define EV_SQ_VAR(name) (volatile unsigned )((char *)iouring_sq_ring + iouring_sq_ ## name)
224			#define EV_CQ_VAR(name) (volatile unsigned )((char *)iouring_cq_ring + iouring_cq_ ## name)
225
226			/* the index array */
227			#define EV_SQ_ARRAY ((unsigned )((char )iouring_sq_ring + iouring_sq_array))
228
229			/* the submit/completion queue entries */
230			#define EV_SQES ((struct io_uring_sqe *) iouring_sqes)
231			#define EV_CQES ((struct io_uring_cqe )((char )iouring_cq_ring + iouring_cq_cqes))
232
233			static
234			struct io_uring_sqe *
235			iouring_sqe_get (EV_P)
236			{
237			unsigned tail = EV_SQ_VAR (tail);
238
239			if (tail + 1 - EV_SQ_VAR (head) > EV_SQ_VAR (ring_entries))
240			{
241			/* queue full, flush */
242			evsys_io_uring_enter (iouring_fd, iouring_to_submit, 0, 0, 0, 0);
243			iouring_to_submit = 0;
244			}
245
246			assert (("libev: io_uring queue full after flush", tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)));
247
248			return EV_SQES + (tail & EV_SQ_VAR (ring_mask));
249			}
250
251			inline_size
252			struct io_uring_sqe *
253			iouring_sqe_submit (EV_P_ struct io_uring_sqe *sqe)
254			{
255			unsigned idx = sqe - EV_SQES;
256
257			EV_SQ_ARRAY [idx] = idx;
258			ECB_MEMORY_FENCE_RELEASE;
259			++EV_SQ_VAR (tail);
260			/ECB_MEMORY_FENCE_RELEASE; / for the time being we assume this is not needed */
261			++iouring_to_submit;
262			}
263
264			/*****************************************************************************/
265
266			/* when the timerfd expires we simply note the fact,
267			* as the purpose of the timerfd is to wake us up, nothing else.
268			* the next iteration should re-set it.
269			*/
270			static void
271			iouring_tfd_cb (EV_P_ struct ev_io *w, int revents)
272			{
273			iouring_tfd_to = EV_TSTAMP_HUGE;
274			}
275
276			/* called for full and partial cleanup */
277			ecb_cold
278			static int
279			iouring_internal_destroy (EV_P)
280			{
281			close (iouring_tfd);
282			close (iouring_fd);
283
284			if (iouring_sq_ring != MAP_FAILED) munmap (iouring_sq_ring, iouring_sq_ring_size);
285			if (iouring_cq_ring != MAP_FAILED) munmap (iouring_cq_ring, iouring_cq_ring_size);
286			if (iouring_sqes != MAP_FAILED) munmap (iouring_sqes , iouring_sqes_size );
287
288	root	1.8	if (ev_is_active (&iouring_tfd_w))
289			{
290			ev_ref (EV_A);
291			ev_io_stop (EV_A_ &iouring_tfd_w);
292			}
293	root	1.1	}
294
295			ecb_cold
296			static int
297			iouring_internal_init (EV_P)
298			{
299			struct io_uring_params params = { 0 };
300
301			iouring_to_submit = 0;
302
303			iouring_tfd = -1;
304			iouring_sq_ring = MAP_FAILED;
305			iouring_cq_ring = MAP_FAILED;
306			iouring_sqes = MAP_FAILED;
307
308	root	1.12	if (!have_monotonic) /* cannot really happen, but what if11 */
309			return -1;
310
311	root	1.1	for (;;)
312			{
313			iouring_fd = evsys_io_uring_setup (iouring_entries, &params);
314
315			if (iouring_fd >= 0)
316			break; /* yippie */
317
318			if (errno != EINVAL)
319			return -1; /* we failed */
320
321	root	1.11	#if TODO
322			if ((~params.features) & (IORING_FEAT_NODROP \| IORING_FEATURE_SINGLE_MMAP))
323			return -1; /* we require the above features */
324			#endif
325
326	root	1.1	/* EINVAL: lots of possible reasons, but maybe
327			* it is because we hit the unqueryable hardcoded size limit
328			*/
329
330			/* we hit the limit already, give up */
331			if (iouring_max_entries)
332			return -1;
333
334			/* first time we hit EINVAL? assume we hit the limit, so go back and retry */
335			iouring_entries >>= 1;
336			iouring_max_entries = iouring_entries;
337			}
338
339			iouring_sq_ring_size = params.sq_off.array + params.sq_entries * sizeof (unsigned);
340			iouring_cq_ring_size = params.cq_off.cqes + params.cq_entries * sizeof (struct io_uring_cqe);
341			iouring_sqes_size = params.sq_entries * sizeof (struct io_uring_sqe);
342
343			iouring_sq_ring = mmap (0, iouring_sq_ring_size, PROT_READ \| PROT_WRITE,
344			MAP_SHARED \| MAP_POPULATE, iouring_fd, IORING_OFF_SQ_RING);
345			iouring_cq_ring = mmap (0, iouring_cq_ring_size, PROT_READ \| PROT_WRITE,
346			MAP_SHARED \| MAP_POPULATE, iouring_fd, IORING_OFF_CQ_RING);
347			iouring_sqes = mmap (0, iouring_sqes_size, PROT_READ \| PROT_WRITE,
348			MAP_SHARED \| MAP_POPULATE, iouring_fd, IORING_OFF_SQES);
349
350			if (iouring_sq_ring == MAP_FAILED \|\| iouring_cq_ring == MAP_FAILED \|\| iouring_sqes == MAP_FAILED)
351			return -1;
352
353			iouring_sq_head = params.sq_off.head;
354			iouring_sq_tail = params.sq_off.tail;
355			iouring_sq_ring_mask = params.sq_off.ring_mask;
356			iouring_sq_ring_entries = params.sq_off.ring_entries;
357			iouring_sq_flags = params.sq_off.flags;
358			iouring_sq_dropped = params.sq_off.dropped;
359			iouring_sq_array = params.sq_off.array;
360
361			iouring_cq_head = params.cq_off.head;
362			iouring_cq_tail = params.cq_off.tail;
363			iouring_cq_ring_mask = params.cq_off.ring_mask;
364			iouring_cq_ring_entries = params.cq_off.ring_entries;
365			iouring_cq_overflow = params.cq_off.overflow;
366			iouring_cq_cqes = params.cq_off.cqes;
367
368			iouring_tfd = timerfd_create (CLOCK_MONOTONIC, TFD_CLOEXEC);
369
370			if (iouring_tfd < 0)
371			return iouring_tfd;
372
373			iouring_tfd_to = EV_TSTAMP_HUGE;
374
375			return 0;
376			}
377
378			ecb_cold
379			static void
380			iouring_fork (EV_P)
381			{
382			iouring_internal_destroy (EV_A);
383
384			while (iouring_internal_init (EV_A) < 0)
385			ev_syserr ("(libev) io_uring_setup");
386
387	root	1.8	fd_rearm_all (EV_A);
388	root	1.1
389			ev_io_stop (EV_A_ &iouring_tfd_w);
390			ev_io_set (EV_A_ &iouring_tfd_w, iouring_tfd, EV_READ);
391			ev_io_start (EV_A_ &iouring_tfd_w);
392			}
393
394			/*****************************************************************************/
395
396			static void
397			iouring_modify (EV_P_ int fd, int oev, int nev)
398			{
399			if (oev)
400			{
401			/* we assume the sqe's are all "properly" initialised */
402			struct io_uring_sqe *sqe = iouring_sqe_get (EV_A);
403			sqe->opcode = IORING_OP_POLL_REMOVE;
404			sqe->fd = fd;
405	root	1.13	/* Jens Axboe notified me that user_data is not what is documented, but is
406			* some kind of unique ID that has to match, otherwise the request cannot
407			* be removed. Since we don't really have that, we pass in the old
408			* generation counter - if that fails, too bad, it will hopefully be removed
409			* at close time and then be ignored. */
410			sqe->user_data = (uint32_t)fd \| ((__u64)(uint32_t)anfds [fd].egen << 32);
411	root	1.1	iouring_sqe_submit (EV_A_ sqe);
412	root	1.2
413			/* increment generation counter to avoid handling old events */
414			++anfds [fd].egen;
415	root	1.1	}
416
417			if (nev)
418			{
419			struct io_uring_sqe *sqe = iouring_sqe_get (EV_A);
420			sqe->opcode = IORING_OP_POLL_ADD;
421			sqe->fd = fd;
422	root	1.2	sqe->user_data = (uint32_t)fd \| ((__u64)(uint32_t)anfds [fd].egen << 32);
423	root	1.1	sqe->poll_events =
424			(nev & EV_READ ? POLLIN : 0)
425			\| (nev & EV_WRITE ? POLLOUT : 0);
426			iouring_sqe_submit (EV_A_ sqe);
427			}
428			}
429
430			inline_size
431			void
432			iouring_tfd_update (EV_P_ ev_tstamp timeout)
433			{
434			ev_tstamp tfd_to = mn_now + timeout;
435
436			/* we assume there will be many iterations per timer change, so
437			* we only re-set the timerfd when we have to because its expiry
438			* is too late.
439			*/
440			if (ecb_expect_false (tfd_to < iouring_tfd_to))
441			{
442			struct itimerspec its;
443
444			iouring_tfd_to = tfd_to;
445			EV_TS_SET (its.it_interval, 0.);
446			EV_TS_SET (its.it_value, tfd_to);
447
448			if (timerfd_settime (iouring_tfd, TFD_TIMER_ABSTIME, &its, 0) < 0)
449			assert (("libev: iouring timerfd_settime failed", 0));
450			}
451			}
452
453			inline_size
454			void
455			iouring_process_cqe (EV_P_ struct io_uring_cqe *cqe)
456			{
457			int fd = cqe->user_data & 0xffffffffU;
458			uint32_t gen = cqe->user_data >> 32;
459			int res = cqe->res;
460
461			assert (("libev: io_uring fd must be in-bounds", fd >= 0 && fd < anfdmax));
462
463			/* documentation lies, of course. the result value is NOT like
464			* normal syscalls, but like linux raw syscalls, i.e. negative
465			* error numbers. fortunate, as otherwise there would be no way
466			* to get error codes at all. still, why not document this?
467			*/
468
469			/* ignore event if generation doesn't match */
470	root	1.13	/* other than skipping removal events, */
471	root	1.1	/* this should actually be very rare */
472	root	1.2	if (ecb_expect_false (gen != (uint32_t)anfds [fd].egen))
473	root	1.1	return;
474
475			if (ecb_expect_false (res < 0))
476			{
477	root	1.11	/TODO: EINVAL handling (was something failed with this fd)/
478			/TODO: EBUSY happens when?/
479	root	1.1
480	root	1.8	if (res == -EBADF)
481	root	1.1	{
482			assert (("libev: event loop rejected bad fd", res != -EBADF));
483			fd_kill (EV_A_ fd);
484			}
485			else
486			{
487			errno = -res;
488			ev_syserr ("(libev) IORING_OP_POLL_ADD");
489			}
490
491			return;
492			}
493
494			/* feed events, we do not expect or handle POLLNVAL */
495			fd_event (
496			EV_A_
497			fd,
498			(res & (POLLOUT \| POLLERR \| POLLHUP) ? EV_WRITE : 0)
499			\| (res & (POLLIN \| POLLERR \| POLLHUP) ? EV_READ : 0)
500			);
501
502			/* io_uring is oneshot, so we need to re-arm the fd next iteration */
503			/* this also means we usually have to do at least one syscall per iteration */
504			anfds [fd].events = 0;
505			fd_change (EV_A_ fd, EV_ANFD_REIFY);
506			}
507
508			/* called when the event queue overflows */
509			ecb_cold
510			static void
511			iouring_overflow (EV_P)
512			{
513			/* we have two options, resize the queue (by tearing down
514			* everything and recreating it, or living with it
515			* and polling.
516	root	1.10	* we implement this by resizing the queue, and, if that fails,
517	root	1.1	* we just recreate the state on every failure, which
518			* kind of is a very inefficient poll.
519			* one danger is, due to the bios toward lower fds,
520			* we will only really get events for those, so
521			* maybe we need a poll() fallback, after all.
522			*/
523			/EV_CQ_VAR (overflow) = 0;/ /* need to do this if we keep the state and poll manually */
524
525			fd_rearm_all (EV_A);
526
527			/* we double the size until we hit the hard-to-probe maximum */
528			if (!iouring_max_entries)
529			{
530			iouring_entries <<= 1;
531			iouring_fork (EV_A);
532			}
533			else
534			{
535			/* we hit the kernel limit, we should fall back to something else.
536			* we can either poll() a few times and hope for the best,
537			* poll always, or switch to epoll.
538	root	1.10	* TODO: is this necessary with newer kernels?
539	root	1.1	*/
540
541			iouring_internal_destroy (EV_A);
542
543	root	1.10	/* this should make it so that on return, we don't call any uring functions */
544	root	1.1	iouring_to_submit = 0;
545
546			for (;;)
547			{
548			backend = epoll_init (EV_A_ 0);
549
550			if (backend)
551			break;
552
553			ev_syserr ("(libev) iouring switch to epoll");
554			}
555			}
556			}
557
558			/* handle any events in the completion queue, return true if there were any */
559			static int
560			iouring_handle_cq (EV_P)
561			{
562			unsigned head, tail, mask;
563
564			head = EV_CQ_VAR (head);
565			ECB_MEMORY_FENCE_ACQUIRE;
566			tail = EV_CQ_VAR (tail);
567
568			if (head == tail)
569			return 0;
570
571			/* it can only overflow if we have events, yes, yes? */
572			if (ecb_expect_false (EV_CQ_VAR (overflow)))
573			{
574			iouring_overflow (EV_A);
575			return 1;
576			}
577
578			mask = EV_CQ_VAR (ring_mask);
579
580			do
581			iouring_process_cqe (EV_A_ &EV_CQES [head++ & mask]);
582			while (head != tail);
583
584			EV_CQ_VAR (head) = head;
585			ECB_MEMORY_FENCE_RELEASE;
586
587			return 1;
588			}
589
590			static void
591			iouring_poll (EV_P_ ev_tstamp timeout)
592			{
593			/* if we have events, no need for extra syscalls, but we might have to queue events */
594			if (iouring_handle_cq (EV_A))
595	root	1.3	timeout = EV_TS_CONST (0.);
596	root	1.1	else
597			/* no events, so maybe wait for some */
598			iouring_tfd_update (EV_A_ timeout);
599
600	root	1.6	/* only enter the kernel if we have something to submit, or we need to wait */
601	root	1.1	if (timeout \|\| iouring_to_submit)
602			{
603			int res;
604
605			EV_RELEASE_CB;
606
607			res = evsys_io_uring_enter (iouring_fd, iouring_to_submit, 1,
608	root	1.3	timeout > EV_TS_CONST (0.) ? IORING_ENTER_GETEVENTS : 0, 0, 0);
609	root	1.1	iouring_to_submit = 0;
610
611			EV_ACQUIRE_CB;
612
613			if (ecb_expect_false (res < 0))
614			if (errno == EINTR)
615			/* ignore */;
616			else
617			ev_syserr ("(libev) iouring setup");
618			else
619			iouring_handle_cq (EV_A);
620			}
621			}
622
623			inline_size
624			int
625			iouring_init (EV_P_ int flags)
626			{
627			iouring_entries = IOURING_INIT_ENTRIES;
628			iouring_max_entries = 0;
629
630			if (iouring_internal_init (EV_A) < 0)
631			{
632			iouring_internal_destroy (EV_A);
633			return 0;
634			}
635
636	root	1.6	ev_io_init (&iouring_tfd_w, iouring_tfd_cb, iouring_tfd, EV_READ);
637	root	1.8	ev_set_priority (&iouring_tfd_w, EV_MINPRI);
638	root	1.1	ev_io_start (EV_A_ &iouring_tfd_w);
639			ev_unref (EV_A); /* watcher should not keep loop alive */
640
641			backend_modify = iouring_modify;
642			backend_poll = iouring_poll;
643
644			return EVBACKEND_IOURING;
645			}
646
647			inline_size
648			void
649			iouring_destroy (EV_P)
650			{
651			iouring_internal_destroy (EV_A);
652			}
653