Re: [RFC v3 PATCH 5/5] x86: check VM_DEAD flag in page fault

From: Yang Shi <yang.shi@linux.alibaba.com>
To: Michal Hocko <mhocko@kernel.org>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>,
	willy@infradead.org, akpm@linux-foundation.org,
	peterz@infradead.org, mingo@redhat.com, acme@kernel.org,
	alexander.shishkin@linux.intel.com, jolsa@redhat.com,
	namhyung@kernel.org, tglx@linutronix.de, hpa@zytor.com,
	linux-mm@kvack.org, x86@kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [RFC v3 PATCH 5/5] x86: check VM_DEAD flag in page fault
Date: Mon, 2 Jul 2018 11:10:23 -0700	[thread overview]
Message-ID: <a5b4888c-6518-df47-bc0d-d4173984daa9@linux.alibaba.com> (raw)
In-Reply-To: <20180702175749.GG19043@dhcp22.suse.cz>

On 7/2/18 10:57 AM, Michal Hocko wrote:
> On Mon 02-07-18 10:24:27, Yang Shi wrote:
>>
>> On 7/2/18 6:37 AM, Michal Hocko wrote:
>>> On Mon 02-07-18 15:33:11, Laurent Dufour wrote:
>>>> On 02/07/2018 14:45, Michal Hocko wrote:
>>>>> On Mon 02-07-18 14:26:09, Laurent Dufour wrote:
>>>>>> On 02/07/2018 14:15, Michal Hocko wrote:
>>> [...]
>>>>>>> We already do have a model for that. Have a look at MMF_UNSTABLE.
>>>>>> MMF_UNSTABLE is a mm's flag, here this is a VMA's flag which is checked.
>>>>> Yeah, and we have the VMA ready for all places where we do check the
>>>>> flag. check_stable_address_space can be made to get vma rather than mm.
>>>> Yeah, this would have been more efficient to check that flag at the beginning
>>>> of the page fault handler rather than the end, but this way it will be easier
>>>> to handle the speculative page fault too ;)
>>> The thing is that it doesn't really need to be called earlier. You are
>>> not risking data corruption on file backed mappings.
>> OK, I just think it could save a few cycles to check the flag earlier.
> This should be an extremely rare case. Just think about it. It should
> only ever happen when an access races with munmap which itself is
> questionable if not an outright bug.
>
>> If nobody think it is necessary, we definitely could re-use
>> check_stable_address_space(),
> If we really need this whole VM_DEAD thing then it should be better
> handled at the same place rather than some ad-hoc places.
>
>> just return VM_FAULT_SIGSEGV for VM_DEAD vma,
>> and check for both shared and non-shared.
> Why would you even care about shared mappings?

Just thought about we are dealing with VM_DEAD, which means the vma will 
be tore down soon regardless it is shared or non-shared.

MMF_UNSTABLE doesn't care about !shared case.