Re: [Qemu-devel] [PATCH] qemu-char: Do not disconnect when there's data for reading

[Kirill Batuzov]

On Tue, 16 Sep 2014, Markus Armbruster wrote:

> 
> Kirill, you added the code being changed.  Could you review the patch?
>

I'll try but this is more about GIOConditions which I do not understand
well. See below.

Zifei Tong <address@hidden> writes:

> After commit 812c1057f6175ac9a9829fa2920a2b5783814193 (Handle G_IO_HUP
> in tcp_chr_read for tcp chardev), the connection is disconnected when in
> G_IO_HUP condition.
>
> However, it's possible that the channel is in G_IO_IN condition at the
> same time, meaning there is data for reading. In that case, the
> remaining data is not handled.
>
> I saw a related bug when running socat in write-only mode, with
>
>   $ echo "quit" | socat -u - UNIX-CONNECT:qemu-monitor
>
> the monitor won't not run the 'quit' command.
> CC: Kirill Batuzov <address@hidden>
> CC: Nikolay Nikolaev <address@hidden>
> CC: Anthony Liguori <address@hidden>
> Signed-off-by: Zifei Tong <address@hidden>
> ---
>  qemu-char.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/qemu-char.c b/qemu-char.c
> index 1a8d9aa..5018c3a 100644
> --- a/qemu-char.c
> +++ b/qemu-char.c
> @@ -2706,7 +2706,7 @@ static gboolean tcp_chr_read(GIOChannel *chan, 
> GIOCondition cond, void *opaque)
>      uint8_t buf[READ_BUF_LEN];
>      int len, size;
>  
> -    if (cond & G_IO_HUP) {
> +    if (!(cond & G_IO_IN) && (cond & G_IO_HUP)) {
>          /* connection closed */
>          tcp_chr_disconnect(chr);
>          return TRUE;

I've tried running the above code and watching in debugger what is
happening. I wanted to know that tcp_chr_disconnect is called properly
so I replaced the 'quit' command with 'help'. From the way code works I
have a feeling that we are using some undocumented linux-specific
behaviour here.

What I saw:
 - Sometimes all three (G_IO_IN, G_IO_HUP and G_IO_ERR) conditions are
   asserted, sometimes only two of them (G_IO_IN and G_IO_HUP).
 - G_IO_IN is *never* de-asserted. Even after all data is depleted it is
   still up.
 - When G_IO_ERR is asserted and all data have been read one call to
   tcp_chr_recv returns -1. Subsequent calls return 0.

GIOCondition behaviour in corner cases is puzzling and can differ from OS
to OS (commit 812c105 is an example, there also were freebsd-specific
bugfixes if I remember correctly).

I suggest we remove condition checks completely and use more reliable
and better documented source of information - return value of
tcp_chr_recv (which is just return value of recvmsg). All we need to do
is to handle 'size < 0' and not forget about EAGAIN case.

Currently we have a mix of GLib conditions and POSIX return values to
handle all cases and we can not do it with GLib alone (I do not know a
way to tell if there is data in channel or not when G_IO_HUP is asserted).

All these problems were before this patch, but I think it is better to
fix it once than add patch over patch fighting GIOCondition's weird
behaviour.

-- 
Kirill