diff mbox

[v3,1/1] qemu-char: socket backend: disconnect on write error

Message ID 1486045589-8074-1-git-send-email-den@openvz.org
State New
Headers show

Commit Message

Denis V. Lunev Feb. 2, 2017, 2:26 p.m. UTC
From: Anton Nefedov <anton.nefedov@virtuozzo.com>

Socket backend read handler should normally perform a disconnect, however
the read handler may not get a chance to run if the frontend is not ready
(qemu_chr_be_can_write() == 0).

This means that in virtio-serial frontend case if
 - the host has disconnected (giving EPIPE on socket write)
 - and the guest has disconnected (-> frontend not ready -> backend
   will not read)
 - and there is still data (frontend->backend) to flush (has to be a really
   tricky timing but nevertheless, we have observed the case in production)

This results in virtio-serial trying to flush this data continiously forming
a busy loop.

Solution: react on write error in the socket write handler.
errno is not reliable after qio_channel_writev_full(), so we may not get
the exact EPIPE, so disconnect on any error but QIO_CHANNEL_ERR_BLOCK which
io_channel_send_full() converts to errno EAGAIN.
We must not disconnect right away though, there still may be data to read
(see 4bf1cb0).

Signed-off-by: Anton Nefedov <anton.nefedov@virtuozzo.com>
Signed-off-by: Denis V. Lunev <den@openvz.org>
CC: Paolo Bonzini <pbonzini@redhat.com>
CC: Daniel P. Berrange <berrange@redhat.com>
CC: Marc-André Lureau <marcandre.lureau@redhat.com>
---
Changes from v2:
- fixed prototype of tcp_chr_disconnect

Changes from v1:
- we do not rely on EPIPE anynore. Socket should be closed on all errors
  except EAGAIN

 chardev/char-socket.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

Comments

Paolo Bonzini Feb. 3, 2017, 6:59 p.m. UTC | #1
On 02/02/2017 06:26, Denis V. Lunev wrote:
> From: Anton Nefedov <anton.nefedov@virtuozzo.com>
> 
> Socket backend read handler should normally perform a disconnect, however
> the read handler may not get a chance to run if the frontend is not ready
> (qemu_chr_be_can_write() == 0).
> 
> This means that in virtio-serial frontend case if
>  - the host has disconnected (giving EPIPE on socket write)
>  - and the guest has disconnected (-> frontend not ready -> backend
>    will not read)
>  - and there is still data (frontend->backend) to flush (has to be a really
>    tricky timing but nevertheless, we have observed the case in production)
> 
> This results in virtio-serial trying to flush this data continiously forming
> a busy loop.
> 
> Solution: react on write error in the socket write handler.
> errno is not reliable after qio_channel_writev_full(), so we may not get
> the exact EPIPE, so disconnect on any error but QIO_CHANNEL_ERR_BLOCK which
> io_channel_send_full() converts to errno EAGAIN.
> We must not disconnect right away though, there still may be data to read
> (see 4bf1cb0).
> 
> Signed-off-by: Anton Nefedov <anton.nefedov@virtuozzo.com>
> Signed-off-by: Denis V. Lunev <den@openvz.org>
> CC: Paolo Bonzini <pbonzini@redhat.com>
> CC: Daniel P. Berrange <berrange@redhat.com>
> CC: Marc-André Lureau <marcandre.lureau@redhat.com>
> ---
> Changes from v2:
> - fixed prototype of tcp_chr_disconnect
> 
> Changes from v1:
> - we do not rely on EPIPE anynore. Socket should be closed on all errors
>   except EAGAIN
> 
>  chardev/char-socket.c | 10 ++++++++++
>  1 file changed, 10 insertions(+)
> 
> diff --git a/chardev/char-socket.c b/chardev/char-socket.c
> index 4068dc5..865c527 100644
> --- a/chardev/char-socket.c
> +++ b/chardev/char-socket.c
> @@ -97,6 +97,9 @@ static gboolean tcp_chr_accept(QIOChannel *chan,
>                                 GIOCondition cond,
>                                 void *opaque);
>  
> +static int tcp_chr_read_poll(void *opaque);
> +static void tcp_chr_disconnect(Chardev *chr);
> +
>  /* Called with chr_write_lock held.  */
>  static int tcp_chr_write(Chardev *chr, const uint8_t *buf, int len)
>  {
> @@ -114,6 +117,13 @@ static int tcp_chr_write(Chardev *chr, const uint8_t *buf, int len)
>              s->write_msgfds_num = 0;
>          }
>  
> +        if (ret < 0 && errno != EAGAIN) {
> +            if (tcp_chr_read_poll(chr) <= 0) {
> +                tcp_chr_disconnect(chr);
> +                return len;
> +            } /* else let the read handler finish it properly */
> +        }
> +
>          return ret;
>      } else {
>          /* XXX: indicate an error ? */
> 

Queued, thanks.

Paolo
diff mbox

Patch

diff --git a/chardev/char-socket.c b/chardev/char-socket.c
index 4068dc5..865c527 100644
--- a/chardev/char-socket.c
+++ b/chardev/char-socket.c
@@ -97,6 +97,9 @@  static gboolean tcp_chr_accept(QIOChannel *chan,
                                GIOCondition cond,
                                void *opaque);
 
+static int tcp_chr_read_poll(void *opaque);
+static void tcp_chr_disconnect(Chardev *chr);
+
 /* Called with chr_write_lock held.  */
 static int tcp_chr_write(Chardev *chr, const uint8_t *buf, int len)
 {
@@ -114,6 +117,13 @@  static int tcp_chr_write(Chardev *chr, const uint8_t *buf, int len)
             s->write_msgfds_num = 0;
         }
 
+        if (ret < 0 && errno != EAGAIN) {
+            if (tcp_chr_read_poll(chr) <= 0) {
+                tcp_chr_disconnect(chr);
+                return len;
+            } /* else let the read handler finish it properly */
+        }
+
         return ret;
     } else {
         /* XXX: indicate an error ? */