Re: [PATCH] vsock: use sock_error() to consume sk_err after connect timeout

[Michal Luczaj <mhal@rbox.co>]

On 7/29/26 15:29, Stefano Garzarella wrote:
> On Wed, Jul 29, 2026 at 03:20:38PM +0200, Michal Luczaj wrote:
>> On 7/29/26 15:03, Stefano Garzarella wrote:
>>> On Wed, Jul 29, 2026 at 05:46:29PM +0800, Nguyen Dinh Phi [SG] wrote:
>>>> On 28/7/26 16:21, Stefano Garzarella wrote:
>>>>> On Fri, Jul 24, 2026 at 03:34:23PM +0800, Nguyen Dinh Phi [SG] wrote:
>>>>>> On 24/7/26 05:43, Michal Luczaj wrote:
>>>>>>> On 7/23/26 12:26, Nguyen Dinh Phi [SG] wrote:
>>>>>>>>>>>> ...
>>>>>>>>>>>> Yeah, we need to handle that part better, I think it's
>>>>>>>>>>>> a leftover when
>>>>>>>>>>>> we generalized AF_VSOCK to support more transport than vmci.
>>>>>>>>>>
>>>>>>>>>> Speaking of leftovers, I have trouble understanding where
>>>>>>>>>> does vsock set
>>>>>>>>>> sk_err on listener sockets anyway. If it doesn't, why vsock_accept()
>>>>>>>>>> checks for it?
>>>>>>>>>
>>>>>>>>> I can't also see where it can be set TBH. Should we remove it ?
>>>>>>>>
>>>>>>>> I couldn't find it for listener side too.
>>>>>>>
>>>>>>> Removing sk_err handling from vsock_accept() solves the problem, right?
>>>>>>>
>>>>>>> thanks,
>>>>>>> Michal
>>>>>>
>>>>>> Yes, confirmed, removing sk_err checks from vsock_accept() does
>>>>>> solve the problem.
>>>>>
>>>>> Okay, so maybe better on going on this direction. WDYT?
>>>>>
>>>>> Stefano
>>>>>
>>>>
>>>> I'm still a bit concerned about how connect() and poll() interact
>>>> here, even with the sk_err checks removed from vsock_accept().
>>>>
>>>> For example:
>>>> vsock_accept() now lets us reuse a socket whose connect() failed (call
>>>> it r0) as syzbot reproducer does. After listen(), r0 becomes a
>>>> listener (sk_state == TCP_LISTEN) and works correctly -- it accepts
>>>> connections.
>>>>
>>>> But poll() on r0 still marks POLLERR, even though there is no error on
>>>> that socket at that point.
>>>>
>>>> As I understand it, sk_err holds an error that has not yet been
>>>> reported to userspace. In the blocking vsock_connect() case we have
>>>> already read that error and returned it to the caller, so it is no
>>>> longer pending
>>>> Shouldn't sk_err be consumed/cleared when vsock_connect() returns it
>>>> to userspace?
>>>
>>> Yeah, makes sense to me, I'll ack the v2.
>>> @Michal WDYT?
>>
>> I'm worried this patch does not address the non-blocking connect() case.
>> Could vsock_connect_timeout() set `sk->sk_err = ETIMEDOUT` after connect()
>> returns?
> 
> This is a good point!
> 
> So we still need to remove `sk_err` check in vsock_accept(), or set 
> `sk->sk_err = 0` in vsock_listen() to have a complete fix, right?

Yup, that's my understanding as well.

That said, I'm not against flushing sk_err on a failed blocking connect()
(this patch) as a follow up, if you guys agree that makes vsock behave like
other socket families.