Re: [PATCH v3] mm/gup: Allow real explicit breaking of COW

From: Peter Xu
Date: Thu Aug 20 2020 - 17:55:00 EST


On Tue, Aug 11, 2020 at 04:10:57PM -0700, Linus Torvalds wrote:
> On Tue, Aug 11, 2020 at 2:43 PM Peter Xu <peterx@xxxxxxxxxx> wrote:
> >
> > I don't know good enough on the reuse refactoring patch (which at least looks
> > functionally correct), but... IMHO we still need the enforced cow logic no
> > matter we refactor the page reuse logic or not, am I right?
> >
> > Example:
> >
> > - Process A & B shares private anonymous page P0
> >
> > - Process A does READ of get_user_pages() on page P0
> >
> > - Process A (e.g., another thread of process A, or as long as process A still
> > holds the page P0 somehow) writes to page P0 which triggers cow, so for
> > process A the page P0 is replaced by P1 with identical content
> >
> > Then process A still keeps the reference to page P0 that potentially belongs to
> > process B or others?
>
> The COW from process A will indeed keep a reference to page P0 (for
> whatever nefarious kernel use it did the GUP for). And yes, that page
> is still mapped into process B.
>
> HOWEVER.
>
> Since the GUP will be incrementing the reference count of said page,
> the actual problem has gone away. Because the GUP copy won't be
> modifying the page (it's a read lookup), and as long as process B only
> reads from the page, we're happily sharing a read-only page.
>
> And if process B now starts writing to it, the "page_count()" check at
> fault time will trigger, and process B will do a COW copy.
>
> So now we'll have three copies of the page: the original one is being
> kept for the GUP, and both A and B did their COW copies in their page
> tables.
>
> And that's exactly what we wanted - they are all now containing
> different data, after all.
>
> The problem with the *current* code is that we don't actually look at
> the page count at all, only the mapping count, so the GUP reference
> count is basically invisible.
>
> And the reason we don't look too closely at the page count is that
> there's a lot of incidental things that can affect it, like the whole
> KSM reference, the swap cache reference, other GUP users etc etc. So
> we historically have tried to maximize the amount of sharing we can
> do.
>
> But that "maximize sharing" is really complicated.
>
> That's the big change of that simplification patch - it's basically
> saying that "whenever _anything_ else has a reference to that page,
> we'll just copy and not even try to share".

Sorry for the late reply, and thanks for the explanations. That definitely
helped me to understand.

So, which way should we go?

I kind of prefer the new suggestion to remove code rather than adding new
codes. I definitely don't know enough on the side effect of it, especially
performance-wise on either ksm or swap, but... IIUC the worst case is we'll get
some perf report later on, and it seems also not hard to revert the patch later
if we want.

Thanks,

--
Peter Xu