On 12/30/22 14:18, Joel Fernandes wrote:
> I think the patch from Matthew Wilcox will address it because the
> read side section already acquires the dep_map. So lockdep subsystem
> should be able to nail the dependency. (...)
Perhaps it's something misconfigured on my side, but I still don't see any
lockdep splats, just the usual task hang warning after 120s.
If that's any help, here's a crude selftest (actually a severed version of
xen_shinfo_test). Under current mainline 6.2.0-rc1 it results in exactly
the type of deadlocks described by David.