suspecting 5.10.0-21 kernel breaking something with futexes on s390x
#1030545 reported against qemu. The issue is rather fundamental for qemu,
it is one of the very basic operations. I can reproduce the same issue on
older versions as well, even installed bullseye chroot on zelenka.d.o and
it fails/hangs the same way there.
A common pattern is the usage of 5.10.0-21 kernel. For example, last
successful builds of libguestfs (which suffered from this issue) were
done on hosts running previous kernel version, and there's no single
successful build on 5.10.0-21. Zelenka is also running 5.10.0-21, -
I asked to reboot it into -20 to see if it changes something.
We tried to run it on other s390x machines, - the bug does not show
itself there. But these run other kernels.
I looked at the changes between -20 and -21, 5.10.150 -> 5.10.162, and
I do see at least one s390x-specific change to futexes (and it is the
futex which behave in a weird way now), plus changes in signal/wakeup
handling. Also, there are other changes past .162 which includes quite
some changes to s390x atomix/cmpxchg stuff. So it *might* be related.
At any rate, with the data points I have so far, everything points to
the kernel.
If anyone know what's going on there, please do share your knowledge :)
Thanks,
/mjt
Reply to: