On Wed, Sep 22, 2021 at 08:49:25PM -0400, John Snow wrote:
> The single space is indeed required to successfully transmit the file
> descriptor to QEMU.
Sending fds requires a payload of at least one byte, but I don't think
that qemu cares which byte. Thus, while your choice of space is fine,
the commit message may be a bit misleading at implying it must be
space.
OK, I'll rephrase. (Space winds up being useful in particular because it doesn't mess with the parsing for subsequent JSON objects sent over the wire.)
(Idle curiosity: Is it possible to make QEMU accept an empty payload here? I understand that for compatibility reasons it wouldn't change much for the python lib even if we did, but I'm curious.)
>
> Python 3.11 removes support for calling sendmsg directly from a
> transport's socket. There is no other interface for doing this, our use
> case is, I suspect, "quite unique".
>
> As far as I can tell, this is safe to do -- send_fd_scm is a synchronous
> function and we can be guaranteed that the async coroutines will *not* be
> running when it is invoked. In testing, it works correctly.
>
> I investigated quite thoroughly the possibility of creating my own
> asyncio Transport (The class that ultimately manages the raw socket
> object) so that I could manage the socket myself, but this is so wildly
> invasive and unportable I scrapped the idea. It would involve a lot of
> copy-pasting of various python utilities and classes just to re-create
> the same infrastructure, and for extremely little benefit. Nah.
>
> Just boldly void the warranty instead, while I try to follow up on
> https://bugs.python.org/issue43232
Bummer that we have to do that, but at least you are documenting the
problems and pursuing a remedy upstream.
Yeah. I suspect our use case is so niche that it's not likely to get traction, but I'll try again. This sort of thing might make it harder to use projects like pypy, so it does feel like a defeat. Still, where there's a will, there's a way, right? :)
--js