qemu-system-ppc failures with glibc 2.38

qemu-system-ppc failures with glibc 2.38

[Richard Purdie]

Hi,

Yocto Project's CI is noticing a lot of issues with qemu-system-ppc
emulation on loaded systems after we switch glibc to 2.38.

This is manifesting as hangs in the emulated system and for example,
systemd units then timeout and restart. If we have long running
commands running over ssh (e.g. configure or kernel module builds),
these are then are then disconnected.

It does appear the system does 'freeze' periodically. It also seems
related to the load on the system running the emulation. On an
otherwise idle system it is fine, when there is other load, we hit the
freezes, watchdog resets and resulting faiilures.

We do sometimes see kernel rcu stalls too, the frequency of them may be
higher with the new glibc, it is hard to tell.

Can anyone think of something that changed in glibc 2.38 which might
affect qemu-system-ppc in this way? Everything else we test seems ok.

I've tested with qemu 8.0.0, 8.0.3, 8.0.4 and 8.1.0 with similar
results. The latter has other problems unfortunately which we're still
trying to debug (x86 hangs, we then tried the "softmmu: Assert data in
bounds in iotlb_to_section" patches but that breaks mips). 

Cheers,

Richard