Re: can: isotp: epoll breaks isotp_sendmsg

[Marc Kleine-Budde <mkl@pengutronix.de>]

[-- Attachment #1: Type: text/plain, Size: 2704 bytes --]

On 01.07.2023 00:35:18, Michal Sojka wrote:
> Hi Maxime,
> 
> On Fri, Jun 30 2023, Maxime Jayat wrote:
> > Hi,
> >
> > There is something not clear happening with the non-blocking behavior
> > of ISO-TP sockets in the TX path, but more importantly, using epoll now
> > completely breaks isotp_sendmsg.
> > I believe it is related to
> > 79e19fa79c ("can: isotp: isotp_ops: fix poll() to not report false 
> > EPOLLOUT events"),
> > but actually is probably deeper than that.
> >
> > I don't completely understand what is exactly going on, so I am sharing
> > the problem I face:
> >
> > With an ISO-TP socket in non-blocking mode, using epoll seems to make
> > isotp_sendmsg always return -EAGAIN.
> 
> That's definitely not expected behavior. I tested the patch only with
> poll, hoping that epoll would behave the same.
> 
> [...]
> 
> >
> > By reverting 79e19fa79c, I get better results but still incorrect:
> 
> [...]
> 
> > It is then possible to write on the socket but the write is blocking,
> > which is not the expected behavior for a non-blocking socket.
> 
> Yes, incorrect behavior was why we made the commit in question, however
> we saw write() returning -EAGAIN when it shouldn't.
> 
> > I don't know how to solve the problem. To me, using wq_has_sleeper seems 
> > weird.
> 
> Agreed. I've never tried to understand how synchronization works here.
> Hopefully, Oliver knows more.
> 
> > The implementation of isotp_poll feels weird too (calling both 
> > datagram_poll and
> > poll_wait?). But I am not sure what would be the correct
> > implementation.
> 
> I understand it as follows (which might be wrong - someone, please
> correct me), isotp_poll() should register the file with all waitqueues
> it can wait on. so->wait is one and sock->sq.wait (used by
> datagram_poll) is another. The former is definitely used for TX, the
> latter is probably used because skb_recv_datagram() is called for RX.
> But so->wait is also used for RX and there might proabbly be be some
> inconsistency between those.
> 
> > My actual use-case is in Async Rust using tokio.
> 
> Our initial motivation was also Rust and tokio however than I did
> testing only with simple C programs. I'm definitely interested in having
> this working.
> 
> I'll try to look at this in more detail during the weekend. It's too
> late for me today.

Any progress on this issue?

Marc

-- 
Pengutronix e.K.                 | Marc Kleine-Budde          |
Embedded Linux                   | https://www.pengutronix.de |
Vertretung Nürnberg              | Phone: +49-5121-206917-129 |
Amtsgericht Hildesheim, HRA 2686 | Fax:   +49-5121-206917-9   |

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]