Re: [PATCH] sockets: update SOCKET_ADDRESS_TYPE_FD listen(2) backlog

[Eric Blake]

On 3/16/21 4:10 AM, Stefan Hajnoczi wrote:
> On Wed, Mar 10, 2021 at 05:30:04PM +0000, Stefan Hajnoczi wrote:
>> socket_get_fd() fails with the error "socket_get_fd: too many
>> connections" if the given listen backlog value is not 1.
>>
>> Not all callers set the backlog to 1. For example, commit
>> 582d4210eb2f2ab5baac328fe4b479cd86da1647 ("qemu-nbd: Use SOMAXCONN for
>> socket listen() backlog") uses SOMAXCONN. This will always fail with in
>> socket_get_fd().
>>
>> This patch calls listen(2) on the fd to update the backlog value. The
>> socket may already be in the listen state. I have tested that this works
>> on Linux 5.10 and macOS Catalina.
>>
>> As a bonus this allows us to detect when the fd cannot listen. Now we'll
>> be able to catch unbound or connected fds in socket_listen().
>>
>> Drop the num argument from socket_get_fd() since this function is also
>> called by socket_connect() where a listen backlog value does not make
>> sense.
>>
>> Fixes: e5b6353cf25c99c3f08bf51e29933352f7140e8f ("socket: Add backlog 
>> parameter to socket_listen")
>> Reported-by: Richard W.M. Jones <rjones@redhat.com>
>> Cc: Juan Quintela <quintela@redhat.com>
>> Cc: Eric Blake <eblake@redhat.com>
>> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
>> ---
>>  util/qemu-sockets.c | 29 ++++++++++++++++++++++-------
>>  1 file changed, 22 insertions(+), 7 deletions(-)
> 
> Dan and Gerd: Can this go via one of your trees?
> 

As it showed up as a regression in qemu-nbd, I can also consider queuing
it in my NBD tree.  However, I claim it counts as a bug fix, so it is
fine for -rc1 even if it misses soft freeze.

I'm fine whichever maintainer takes this, although I've now flagged it
to go through an NBD pull request if it doesn't land elsewhere sooner.

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org