powerpc_mmu_radix: add leaf page for wired mappings when pmap_enter(psind=1)

[rew]

Summary

This applies the fix to powerpc's pmap as was done in commit aa3bcaad51076ceb
and d0941ed9b5c39d92 for amd64 and riscv pmaps, respectively.

Reported by: Jenkins
Fixes: e4078494f344bcba8709216bd601efa3dd05f6b3

[markj]

Thank you! I didn't know that this pmap implemented support for pmap_enter(psind=1).

sys/powerpc/aim/mmu_radix.c

4972

Is this part of the change intentional? I don't think it's right. This pmap doesn't use the valid field to store state; its use in other pmaps is a relatively new change.

[bnovkov]

Aside from the issue @markj pointed out, LGTM.

[rew]

sys/powerpc/aim/mmu_radix.c

4972

It was - after reading through the code and git log a bit more, I can see that it is wrong.

The issue I ran into was initializing the ref_count with the correct value when the leaf page is added. Setting the ref_count to 1 would fix the reported panic but would hit a KASSERT() in a different test when trying to unmap a superpage with ref_count == 1.

The two ideas I thought of was to relax the KASSERT() for ref_count == 1 or set ref_count = NPTEPG and the fill the leaf page with mappings when the page is allocated. I'll update this revision to illustrate the latter.

[rew]

• drop vm_page_all_valid() change
• fill added leaf page with mappings

[markj]

sys/powerpc/aim/mmu_radix.c

3259

I think this line is too long.

4972

I see now, thank you. I think your approach is ok. During demotion, we'll update the attributes in the cached page table page if necessary.