Re: [Qemu-devel] [Qemu-stable] Recent patches for 2.4

[Peter Lieven <pl@kamp.de>]

Am 04.08.2015 um 14:29 schrieb Peter Lieven:
> Am 04.08.2015 um 14:09 schrieb Paolo Bonzini:
>>
>> On 04/08/2015 13:57, Peter Lieven wrote:
>>> Okay, what I found out is that in aio_poll I get revents = POLLIN for
>>> the nfs file descriptor. But there is no data available on the socket.
>> Does read return 0 or EAGAIN?
>>
>> If it returns EAGAIN, the bug is in the QEMU main loop or the kernel.
>> It should never happen that poll returns POLLIN and read returns EAGAIN.
>>
>> If it returns 0, it means the other side called shutdown(fd, SHUT_WR).
>> Then I think the bug is in the libnfs driver or more likely libnfs.  You
>> should stop polling the POLLIN event after read has returned 0 once.
>
> You might be right. Ronnie originally used the FIONREAD ioctl before every read and considered
> the socket as disconnected if the available bytes returned where 0.
> I found that I get available bytes == 0 from that ioctl even if the socket was not closed.
> This seems to be some kind of bug in Linux - at least what I have thought.
>
> See BUGS in the select(2) manpage.
>
>        Under Linux, select() may report a socket file descriptor as "ready for reading", while nevertheless a subsequent read blocks. This could for example happen when data  has  arrived  but
>        upon  examination  has  wrong checksum and is discarded. There may be other circumstances in which a file descriptor is spuriously reported as ready.  Thus it may be safer to use O_NON‐
>        BLOCK on sockets that should not block.
>
> I will debug further, but it seems to be that I receive a POLLIN even if there is no data available. I see 0 bytes from the recv call inside libnfs and continue without a deadlock - at least
> so far.
>
> Would it be a good idea to count the number of 0 bytes from recv and react after I received 0 bytes for a number of consecutive times?
>
> And then: stop polling POLLIN or reconnect?

Okay, got it. Ronnie was using FIONREAD without checking for EAGAIN or EINTR.

I will send a patch for libnfs to reconnect if count == 0. Libiscsi is not affected, it reconnects if count is 0.

Peter