On Thu, Nov 5, 2020 at 7:00 PM Hugh Dickins <hughd@xxxxxxxxxx> wrote:
>
> I don't know why this was addressed to me in particular (easy to imagine
> I've made a mod at some time that bears on this, but I haven't found it);
> but have spent longer considering the patch than I should have done -
> apologies to everyone else I should be replying to.
>
I really appreciate your insights and historical anecdotes. I always
learn something new.
> On Wed, 4 Nov 2020, Shakeel Butt wrote:
>
> > Since the commit 369ea8242c0f ("mm/rmap: update to new mmu_notifier
> > semantic v2"), the code to check the secondary MMU's page table access
> > bit is broken for !(TTU_IGNORE_ACCESS) because the page is unmapped from
> > the secondary MMU's page table before the check. More specifically for
> > those secondary MMUs which unmap the memory in
> > mmu_notifier_invalidate_range_start() like kvm.
>
> Well, "broken" seems a bit unfair to 369ea8242c0f. It put a warning
> mmu_notifier_invalidate_range_start() at the beginning, and matching
> mmu_notifier_invalidate_range_end() at the end of try_to_unmap_one();
> with its mmu_notifier_invalidate_range() exactly where the
> mmu_notifier_invalidate_page() was before (I think the story gets
> more complicated later). Yes, if notifiee takes invalidate_range_start()
> as signal to invalidate all their own range, then that will sometimes
> cause them unnecessary invalidations.
>
> Not just for !TTU_IGNORE_ACCESS: there's also the !TTU_IGNORE_MLOCK
> case meeting a VM_LOCKED vma and setting PageMlocked where that had
> been missed earlier (and page_check_references() has intentionally but
> confusingly marked this case as PAGEREF_RECLAIM, not to reclaim the page,
> but to reach the try_to_unmap_one() which will recognize and fix it up -
> historically easier to do there than in page_referenced_one()).
>
> But I think mmu_notifier is a diversion from what needs thinking about.
>
> >
> > However memory reclaim is the only user of !(TTU_IGNORE_ACCESS) or the
> > absence of TTU_IGNORE_ACCESS and it explicitly performs the page table
> > access check before trying to unmap the page. So, at worst the reclaim
> > will miss accesses in a very short window if we remove page table access
> > check in unmapping code.
>
> I agree with you and Johannes that the short race window when the page
> might be re-referenced is no issue at all: the functional issue is the
> one in your next paragraph. If that's agreed by memcg guys, great,
> then this patch is a nice observation and a welcome cleanup.
>
> >
> > There is an unintented consequence of !(TTU_IGNORE_ACCESS) for the memcg
> > reclaim. From memcg reclaim the page_referenced() only account the
> > accesses from the processes which are in the same memcg of the target
> > page but the unmapping code is considering accesses from all the
> > processes, so, decreasing the effectiveness of memcg reclaim.
>
> Are you sure it was unintended?
>
> Since the dawn of memcg reclaim, it has been the case that a recent
> reference in a "foreign" vma has rescued that page from being reclaimed:
> now you propose to change that. I expect some workflows will benefit
> and others be disadvantaged. I have no objection myself to the change,
> but I do think it needs to be better highlighted here, and explicitly
> agreed by those more familiar with memcg reclaim.
The reason I said unintended was due to bed7161a519a2 ("Memory
controller: make page_referenced() cgroup aware"). From the commit
message it seems like the intention was to not be influenced by
foreign accesses during memcg reclaim but it missed to make
try_to_unmap_one() memcg aware.
I agree with you that this is a behavior change and we have explicitly
agree to not let memcg reclaim be influenced by foreign accesses.
