[ViewVC] Diff of: cvs/libev/ev

Comparing libev/ev_linuxaio.c (file contents):
Revision 1.6 by root, Fri Jun 21 03:52:29 2019 UTC vs.
Revision 1.11 by root, Sun Jun 23 22:57:25 2019 UTC

  * 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" 8kB for the ring buffer
+ * therefore the calculation below will use "exactly" 4kB for the ring buffer
  */
-#define EV_LINUXAIO_DEPTH (256 / 2 - 2 - 1) /* max. number of io events per batch */
+#define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */
 /*****************************************************************************/
 /* syscall wrapdadoop */
 #include <sys/syscall.h> /* no glibc wrappers */
 }
 /*****************************************************************************/
 /* 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 */
+  /* 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 */
+      /* 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;
     }
 }
 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)
 static int
 linuxaio_get_events_from_ring (EV_P)
 {
   struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
-  ECB_MEMORY_FENCE_ACQUIRE;
   unsigned head = 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;
+  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);
     }
-  ring->head = tail;
+  *(volatile unsigned *)&ring->head = tail;
+  /* again, other implementations don't do this, but I think it's required for the kernel to see free slots */
+  ECB_MEMORY_FENCE_RELEASE;
   return 1;
 }
 /* read at least one event from kernel, or timeout */
   if (linuxaio_get_events_from_ring (EV_A))
     return;
   /* no events, so wait for at least one, then poll ring buffer again */
-  /* this degraded to one event per loop iteration */
+  /* 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) io_getevents");
+      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;
       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 overflowd.
+             * the buffer overflowed.
-             * In this case, we just try next loop iteration.
+             * 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 hapüpens 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.
+             */
+            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
-          /* TODO: we get EAGAIN when the ring buffer is full for some reason */
-          /* TODO: should we always just try next time? */
-          ev_syserr ("(libev) io_submit");
+          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)
 {
   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);
+#endif
   backend_modify  = linuxaio_modify;
   backend_poll    = linuxaio_poll;
   linuxaio_iocbpmax = 0;
   linuxaio_iocbps = 0;
 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)
 {
-  /* TODO: verify and test */
   /* 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) io_setup");
+    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);
+  ev_unref (EV_A); /* watcher should not keep loop alive */
+#endif
   fd_rearm_all (EV_A);
 }

Diff Legend

-–
+Removed lines
-+
+Added lines
-<
+Changed lines
->
+Changed lines

Comparing libev/ev_linuxaio.c (file contents): Revision 1.6 by root, Fri Jun 21 03:52:29 2019 UTC vs. Revision 1.11 by root, Sun Jun 23 22:57:25 2019 UTC

Diff Legend

Comparing libev/ev_linuxaio.c (file contents):
Revision 1.6 by root, Fri Jun 21 03:52:29 2019 UTC vs.
Revision 1.11 by root, Sun Jun 23 22:57:25 2019 UTC