Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Fri, 4 Mar 2022 04:35:04 +0000
From:   Matthew Wilcox <willy@infradead.org>
To:     Hugh Dickins <hughd@google.com>
Cc:     Andrew Morton <akpm@linux-foundation.org>,
        linux-kernel@vger.kernel.org, linux-mm@kvack.org
Subject: Re: [PATCH mmotm] mm: filemap_unaccount_folio() large skip mapcount
 fixup
Message-ID: <YiGW+J4EtjK7KgYn@casper.infradead.org>
References: <879c4426-4122-da9c-1a86-697f2c9a083@google.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <879c4426-4122-da9c-1a86-697f2c9a083@google.com>
Precedence: bulk

On Thu, Mar 03, 2022 at 08:21:19PM -0800, Hugh Dickins wrote:
> The page_mapcount_reset() when folio_mapped() while mapping_exiting()
> was devised long before there were huge or compound pages in the cache.
> It is still valid for small pages, but not at all clear what's right to
> check and reset on large pages.  Just don't try when folio_test_large().

Thanks for bringing this up!  I was really unsure about this chunk of code
when converting unaccount_page_cache_page() to filemap_unaccount_folio().

Part of me wants to just delete the whole thing.  I'm unconvinced by
the argument; surely it's better to leak memory than perhaps reuse a
page which should not have been freed yet?

Also, the code doesn't take into account that folio_mapped() is freaking
expensive for THP (512 cache lines, blowing away 32kB of your L1 cache!),
and we may as well calculate folio_mapcount() while we're doing it.

Do you see this report often on machines that don't have
VM_BUG_ON_FOLIO() enabled?