[Bug 203923] Running a nested freedos on AMD Athlon i686-pae results in NULL pointer dereference in L0 (kvm_mmu_load)

[bugzilla-daemon@bugzilla.kernel.org]

https://bugzilla.kernel.org/show_bug.cgi?id=203923

--- Comment #7 from Jiri Palecek (jpalecek@web.de) ---
Created attachment 283393
  --> https://bugzilla.kernel.org/attachment.cgi?id=283393&action=edit
Patch that fixes this problem on my system

So, I had a look around the code and found that SVM initialized the nested
vcpus in such a way that ->arch.mmu points to ->arch.guest_mmu. The code in
mmu.c then uses ->arch.mmu->pae_root which crashes.

This patch really takes the path of the least resistance. If they want to have
pae_root allocated even for guest_mmu, let them have it and just allocate it.
Maybe if this is specific to AMD the whole business should be in svm.c though?
Or do it lazily only when actually doing the nesting?

The patch fixes 5.1 kernel on my machine, kvm guest start and the nested guest
start as well. However, in 5.2 there will probably be more problems ahead
because I got a different error there (kvm_spurious_fault in L1).

What are your thoughts on this?

-- 
You are receiving this mail because:
You are watching the assignee of the bug.