08:51 <qyliss> Hmm, small issue

08:51 <qyliss> cloud-hypervisor uses non-blocking eventfds everywhere, and crosvm uses blocking ones

08:51 <qyliss> which is going to be a problem if I pass eventfds from crosvm code into cloud-hypervisor code

08:52 <qyliss> It's also not possible to specify on each operation whether it should block or not, because eventfd doesn't implement send(2) and recv(2) (and therefore MSG_DONTWAIT)

08:53 <qyliss> The obvious option here is to try to rewrite the cloud hypervisor code to block, but I don't like the idea of that as a prerequisite to getting it to work at all

08:54 <qyliss> So I might instead attempt to just implement send() and recv() for eventfd in the kernel, which feels like it shouldn't be a very big job at all, and would be generally useful and probably likely to be accepted upstream.

08:55 <qyliss> But I will have a bit of a think first before I embark on a kernel development adventure, even if it should be a short one.

10:38 <qyliss> Aww, looks like implementing that would be harder than I thought because socket operations are seperate to normal file operations

23:04 <qyliss> This Week in Spectrum: https://spectrum-os.org/lists/archives/spectrum-discuss/87wo2glkg0.fsf@alyssa.is/T/#u

23:07 <cole-h> Nice, reading!

23:09 <cole-h> multi-memfd memory backing seems interesting -- interested to see how that pans out.

23:10 <qyliss> It's not all that complicated in principle

23:10 <qyliss> I think the hard part is working out how big they need to be