Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Fri, 22 Oct 2021 12:15:16 +0100
From:   Matthew Wilcox <willy@infradead.org>
To:     Andrea Righi <andrea.righi@canonical.com>
Cc:     Andrew Morton <akpm@linux-foundation.org>,
        Yang Shi <yang.shi@linux.alibaba.com>,
        Minchan Kim <minchan@kernel.org>, linux-kernel@vger.kernel.org,
        linux-mm@kvack.org
Subject: Re: [PATCH] mm: fix sleeping copy_huge_page called from atomic
 context
Message-ID: <YXKdRDKp+l6lis/R@casper.infradead.org>
References: <20211022074619.57355-1-andrea.righi@canonical.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20211022074619.57355-1-andrea.righi@canonical.com>
Precedence: bulk

On Fri, Oct 22, 2021 at 09:46:19AM +0200, Andrea Righi wrote:
> copy_huge_page() can be called with mapping->private_lock held from
> __buffer_migrate_page() -> migrate_page_copy(), so it is not safe to
> do a cond_resched() in this context.
> 
> Introduce migrate_page_copy_nowait() and copy_huge_page_nowait()
> variants that can be used from an atomic context.

I think this is a consequence of THPs being created when they should not
be.  This is the wrong way to fix this problem; and I suspect it may
already be fixed at least in -mm.  We should have taken this path:

        if (!page_has_buffers(page))
                return migrate_page(mapping, newpage, page, mode);

but since we didn't, we can infer that there's a THP which has buffers
(this should never occur).  It's the same root cause as the invalidatepage
problem, just with a very different signature.