On 9/11/19 8:03 PM, Andy Lutomirski wrote:
That distinction is important because if it ever comes to a choice
between adding a new lock to protect vm_page_prot (and consequently slow
down the whole vm system) and using the WRITE_ONCE solution in TTM, we
should know what the implications are. As it turns out previous choices
in this area actually seem to have opted for the lockless WRITE_ONCE /
READ_ONCE / ptl solution. See __split_huge_pmd_locked() and
vma_set_page_prot().
I think it would be even better if the whole thing could work without
ever writing to vm_page_prot. This would be a requirement for vvar in
the unlikely event that the vvar vma ever supported splittable huge
pages. Fortunately, that seems unlikely :)
Yeah, for TTM the situation is different since we want huge vm pages at
some point.
But I re-read __split_huge_pmd_locked() and it actually looks like
vm_page_prot is only accessed for anonymous vmas. For other vmas, it
appears it just simply zaps the PMD, relying on re-faulting the page
table enries if necessary (as also suggested by Christian in another
thread).
So perhaps we should be good never writing to vm_page_prot.
/Thomas
