Re: mmu notifier calls in apply_to_page_range()

[Andrea Arcangeli <aarcange@redhat.com>]

On Fri, Jul 09, 2010 at 08:06:20AM -0700, Jeremy Fitzhardinge wrote:
> I just noticed that the original mmu notifier change (cddb8a5c14a) adds
> calls to mmu_notifier_invalidate_range_start/end to
> apply_to_page_range().  This doesn't seem correct to me, since
> apply_to_page_range can perform arbitrary operations to the range of
> pages, not just invalidation of the pages.  It seems to me that the
> appropriate mmu notifiers should be called either around the call to
> apply_to_page_range(), or from within the callback function.
> 
> Andrea, what's the rationale for mmu_notifier_invalidate_range_start/end
> here?

As long as the secondary mappings are teardown in range_start and
allowed to be established again only after range_end, all
modifications will be picked up by the secondary mmu. Imagine
secondary mmu like a tlb, that you only invalidate, then it'll be
refilled later (after range_end).

The exception is set_pte_at_notify that is called by ksm to establish
a readonly secondary pte in KVM, KVM only calls
get_user_pages(write=1) (never write=0 even for reads) so until that
is optimized set_pte_at_notify allows guest to access readonly data
without breaking the cow. set_pte_at_notify invokes a change_pte
method, if not implemented it'll just fallback to the invalidate_page
method that is backwards compatible, so no mmu notifier user is
required to call change_pte (especially if the secondary page fault -
kind of secondary tlb-miss software handler invokes get_user_pages
with write=0 for reads, ->change_pte can only eliminate one minor
fault so no big deal).