Received-SPF: pass (google.com: domain of linux-kernel+bounces-138105-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45d1:ec00::1 as permitted sender) client-ip=2604:1380:45d1:ec00::1;
Subject: Re: [PATCH] mm/memory-failure: fix deadlock when
 hugetlb_optimize_vmemmap is enabled
To: Oscar Salvador <osalvador@suse.de>
CC: <akpm@linux-foundation.org>, <naoya.horiguchi@nec.com>,
	<linux-mm@kvack.org>, <linux-kernel@vger.kernel.org>
References: <20240407085456.2798193-1-linmiaohe@huawei.com>
 <ZhVMThr9TNeP6SWj@localhost.localdomain>
 <ZhVoatdJZ1RWu2r3@localhost.localdomain>
From: Miaohe Lin <linmiaohe@huawei.com>
Message-ID: <13aa38af-46a1-3894-32bd-c3eb6ef67359@huawei.com>
Date: Wed, 10 Apr 2024 15:52:14 +0800
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:78.0) Gecko/20100101
 Thunderbird/78.6.0
Precedence: bulk
MIME-Version: 1.0
In-Reply-To: <ZhVoatdJZ1RWu2r3@localhost.localdomain>
Content-Type: text/plain; charset="utf-8"
Content-Language: en-US
Content-Transfer-Encoding: 7bit

On 2024/4/10 0:10, Oscar Salvador wrote:
> On Tue, Apr 09, 2024 at 04:10:22PM +0200, Oscar Salvador wrote:
>> On Sun, Apr 07, 2024 at 04:54:56PM +0800, Miaohe Lin wrote:
>>> In short, below scene breaks the lock dependency chain:
>>>
>>>  memory_failure
>>>   __page_handle_poison
>>>    zone_pcp_disable -- lock(pcp_batch_high_lock)
>>>    dissolve_free_huge_page
>>>     __hugetlb_vmemmap_restore_folio
>>>      static_key_slow_dec
>>>       cpus_read_lock -- rlock(cpu_hotplug_lock)
>>>
>>> Fix this by calling drain_all_pages() instead.
>>>
>>> Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
>>
>> Acked-by: Oscar Salvador <osalvador@suse.de>

Thanks.

> 
> On a second though,
> 
> disabling pcp via zone_pcp_disable() was a deterministic approach.
> Now, with drain_all_pages() we drain PCP queues to buddy, but nothing
> guarantees that those pages do not end up in a PCP queue again before we
> the call to take_page_off_budy() if we
> need refilling, right?

AFAICS, iff check_pages_enabled static key is enabled and in hard offline mode,
check_new_pages() will prevent those pages from ending up in a PCP queue again
when refilling PCP list. Because PageHWPoison pages will be taken as 'bad' pages
and skipped when refill PCP list.

> 
> I guess we can live with that because we will let the system know that we
> failed to isolate that page.

We're trying best to isolate that page anyway. :)

Thanks for your thought.
.

> 
>