Date:   Tue, 15 Feb 2022 13:38:20 -0800 (PST)
From:   Hugh Dickins <hughd@...gle.com>
To:     Matthew Wilcox <willy@...radead.org>
cc:     Hugh Dickins <hughd@...gle.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Michal Hocko <mhocko@...e.com>,
        Vlastimil Babka <vbabka@...e.cz>,
        "Kirill A. Shutemov" <kirill@...temov.name>,
        David Hildenbrand <david@...hat.com>,
        Alistair Popple <apopple@...dia.com>,
        Johannes Weiner <hannes@...xchg.org>,
        Rik van Riel <riel@...riel.com>,
        Suren Baghdasaryan <surenb@...gle.com>,
        Yu Zhao <yuzhao@...gle.com>, Greg Thelen <gthelen@...gle.com>,
        Shakeel Butt <shakeelb@...gle.com>,
        Yang Li <yang.lee@...ux.alibaba.com>,
        linux-kernel@...r.kernel.org, linux-mm@...ck.org
Subject: Re: [PATCH v2 04/13] mm/munlock: rmap call mlock_vma_page()
 munlock_vma_page()

On Tue, 15 Feb 2022, Matthew Wilcox wrote:
> On Mon, Feb 14, 2022 at 06:26:39PM -0800, Hugh Dickins wrote:
> > Add vma argument to mlock_vma_page() and munlock_vma_page(), make them
> > inline functions which check (vma->vm_flags & VM_LOCKED) before calling
> > mlock_page() and munlock_page() in mm/mlock.c.
> > 
> > Add bool compound to mlock_vma_page() and munlock_vma_page(): this is
> > because we have understandable difficulty in accounting pte maps of THPs,
> > and if passed a PageHead page, mlock_page() and munlock_page() cannot
> > tell whether it's a pmd map to be counted or a pte map to be ignored.
> > 
> [...]
> > 
> > Mlock accounting on THPs has been hard to define, differed between anon
> > and file, involved PageDoubleMap in some places and not others, required
> > clear_page_mlock() at some points.  Keep it simple now: just count the
> > pmds and ignore the ptes, there is no reason for ptes to undo pmd mlocks.
> 
> How would you suggest we handle the accounting for folios which are
> intermediate in size between PMDs and PTEs?  eg, an order-4 page?
> Would it make sense to increment mlock_count by HUGE_PMD_NR for
> each PMD mapping and by 1 for each PTE mapping?

I think you're asking the wrong question here, but perhaps you've
already decided there's only one satisfactory answer to the right question.

To answer what you've asked: it doesn't matter at all how you count them
in mlock_count, just so long as they are counted up and down consistently.
Since it's simplest just to count 1 in mlock_count for each pmd or pte,
I prefer that (as I did with THPs); but if you prefer to count pmds up
and down by HUGE_PMD_NR, that works too.

Though, reading again, you're asking about a PMD mapping of an order-4
page?  I don't understand how that could be allowed (except on some
non-x86 architecture where the page table fits only 16 pages).

The question I thought you should be asking is about how to count them
in Mlocked.  That's tough; but I take it for granted that you would not
want per-subpage flags and counts involved (or not unless forced to do
so by some regression that turns out to matter).  And I think the only
satisfactory answer is to count the whole compound_nr() as Mlocked
when any part of it (a single pte, a series of ptes, a pmd) is mlocked;
and (try to) move folio to Unevictable whenever any part of it is mlocked.

That differs from what Kirill decided for THPs (which I cannot
confidently describe, but something like count pmd as Mlocked, don't count
ptes as Mlocked, but uncount pmd if any ptes), and what I simplified it to
in the mm/munlock series (count pmd as Mlocked, ignore ptes), and will
tend to show larger numbers for Mlocked than before; but alternatives
seem unworkable to me.

Hugh

Powered by blists - more mailing lists