Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:18 as permitted sender) client-ip=2620:137:e000::1:18;
Date:   Tue, 15 Feb 2022 23:21:10 +0000
From:   Matthew Wilcox <willy@infradead.org>
To:     Hugh Dickins <hughd@google.com>
Cc:     Andrew Morton <akpm@linux-foundation.org>,
        Michal Hocko <mhocko@suse.com>,
        Vlastimil Babka <vbabka@suse.cz>,
        "Kirill A. Shutemov" <kirill@shutemov.name>,
        David Hildenbrand <david@redhat.com>,
        Alistair Popple <apopple@nvidia.com>,
        Johannes Weiner <hannes@cmpxchg.org>,
        Rik van Riel <riel@surriel.com>,
        Suren Baghdasaryan <surenb@google.com>,
        Yu Zhao <yuzhao@google.com>, Greg Thelen <gthelen@google.com>,
        Shakeel Butt <shakeelb@google.com>,
        Yang Li <yang.lee@linux.alibaba.com>,
        linux-kernel@vger.kernel.org, linux-mm@kvack.org
Subject: Re: [PATCH v2 04/13] mm/munlock: rmap call mlock_vma_page()
 munlock_vma_page()
Message-ID: <Ygw1Zn+WUuY5WkZy@casper.infradead.org>
References: <55a49083-37f9-3766-1de9-9feea7428ac@google.com>
 <501673c-a5a-6c5f-ab65-38545dfb723d@google.com>
 <YgvFMjWPITbD1o64@casper.infradead.org>
 <3c6097a7-df8c-f39c-36e8-8b5410e76c8a@google.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <3c6097a7-df8c-f39c-36e8-8b5410e76c8a@google.com>
Precedence: bulk

On Tue, Feb 15, 2022 at 01:38:20PM -0800, Hugh Dickins wrote:
> 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.

Or I've gravely misunderstood the situation.  Or explained my concern
badly.  The possibilities are endless!

My concern is that a filesystem may create an order-4 folio, an
application mmaps the folio and then calls mlock() (either over a portion
or the entirety of the folio).  As far as I can tell, we then do not
move the folio onto the unevictable list because it is of order >0 and
is only mapped by PTEs.  This presumably then has performance problems
(or we wouldn't need to have an unevictable list in the first place).

> 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.

I think that makes sense.  As with so many other things, we choose to
manage memory in >PAGE_SIZE chunks.  If you mlock() a part of a folio,
we lock the whole folio in memory, and it all counts as being locked.