Received-SPF: pass (google.com: domain of linux-kernel+bounces-160566-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
Date: Fri, 26 Apr 2024 19:27:23 +0100
From: Matthew Wilcox <willy@infradead.org>
To: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: linux-kernel@vger.kernel.org, linux-mm@kvack.org,
	akpm@linux-foundation.org, linmiaohe@huawei.com,
	jane.chu@oracle.com, nao.horiguchi@gmail.com, osalvador@suse.de
Subject: Re: [PATCH] mm/memory-failure: remove shake_page()
Message-ID: <ZivyC3vqa2BIBoMj@casper.infradead.org>
References: <20240426171511.122887-1-sidhartha.kumar@oracle.com>
 <ZivlrMAwRI6xJhc-@casper.infradead.org>
 <d05cbf22-f3e8-414a-a2e3-03e0b857eaca@oracle.com>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <d05cbf22-f3e8-414a-a2e3-03e0b857eaca@oracle.com>

On Fri, Apr 26, 2024 at 10:57:31AM -0700, Sidhartha Kumar wrote:
> On 4/26/24 10:34 AM, Matthew Wilcox wrote:
> > On Fri, Apr 26, 2024 at 10:15:11AM -0700, Sidhartha Kumar wrote:
> > > Use a folio in get_any_page() to save 5 calls to compound head and
> > > convert the last user of shake_page() to shake_folio(). This allows us
> > > to remove the shake_page() definition.
> > 
> > So I didn't do this before because I wasn't convinced it was safe.
> > We don't have a refcount on the folio, so the page might no longer
> > be part of this folio by the time we get the refcount on the folio.
> > 
> > I'd really like to see some argumentation for why this is safe.
> 
> If I moved down the folio = page_folio() line to after we verify
> __get_hwpoison_page() has returned 1, which indicates the reference count
> was successfully incremented via foliO_try_get(), that means the folio
> conversion would happen after we have a refcount. In the case we don't call
> __get_hwpoison_page(), that means the MF_COUNT_INCREASED flag is set. This
> means the page has existing users so that path would be safe as well. So I
> think this is safe after moving page_folio() after __get_hwpoison_page().

See if you can find a hole in this chain of reasoning ...

memory_failure()
        p = pfn_to_online_page(pfn);
        res = try_memory_failure_hugetlb(pfn, flags, &hugetlb);
(not a hugetlb)
        if (TestSetPageHWPoison(p)) {
(not already poisoned)
        if (!(flags & MF_COUNT_INCREASED)) {
                res = get_hwpoison_page(p, flags);

get_hwpoison_page()
                ret = get_any_page(p, flags);

get_any_page()
	folio = page_folio(page)

Because we don't have a reference on the folio at this point (how could
we?), the folio might be split, and now we have a pointer to a folio
which no longer contains the page (assuming we had a hwerror in what
was a tail page at this time).