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

[Zifei Tong]

On Tue, Sep 16, 2014 at 11:30 PM, Kirill Batuzov <address@hidden> wrote:
> 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;

Hi Kirill,

Thanks for your review and detailed analysis.

> 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.

This sounds good to me. I'll send a V2 patch based on this.

Thanks,
Zifei

> 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