Re: [RFC v2 PATCH 2/2] mm: mmap: zap pages with read mmap_sem for large mapping

From: Yang Shi <yang.shi@linux.alibaba.com>
To: Michal Hocko <mhocko@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>,
	Nadav Amit <nadav.amit@gmail.com>,
	Matthew Wilcox <willy@infradead.org>,
	ldufour@linux.vnet.ibm.com,
	Andrew Morton <akpm@linux-foundation.org>,
	Ingo Molnar <mingo@redhat.com>,
	acme@kernel.org, alexander.shishkin@linux.intel.com,
	jolsa@redhat.com, namhyung@kernel.org,
	"open list:MEMORY MANAGEMENT" <linux-mm@kvack.org>,
	linux-kernel@vger.kernel.org
Subject: Re: [RFC v2 PATCH 2/2] mm: mmap: zap pages with read mmap_sem for large mapping
Date: Thu, 28 Jun 2018 12:10:10 -0700	[thread overview]
Message-ID: <2ecdb667-f4de-673d-6a5f-ee50df505d0c@linux.alibaba.com> (raw)
In-Reply-To: <20180628115101.GE32348@dhcp22.suse.cz>

On 6/28/18 4:51 AM, Michal Hocko wrote:
> On Wed 27-06-18 10:23:39, Yang Shi wrote:
>>
>> On 6/27/18 12:24 AM, Michal Hocko wrote:
>>> On Tue 26-06-18 18:03:34, Yang Shi wrote:
>>>> On 6/26/18 12:43 AM, Peter Zijlstra wrote:
>>>>> On Mon, Jun 25, 2018 at 05:06:23PM -0700, Yang Shi wrote:
>>>>>> By looking this deeper, we may not be able to cover all the unmapping range
>>>>>> for VM_DEAD, for example, if the start addr is in the middle of a vma. We
>>>>>> can't set VM_DEAD to that vma since that would trigger SIGSEGV for still
>>>>>> mapped area.
>>>>>>
>>>>>> splitting can't be done with read mmap_sem held, so maybe just set VM_DEAD
>>>>>> to non-overlapped vmas. Access to overlapped vmas (first and last) will
>>>>>> still have undefined behavior.
>>>>> Acquire mmap_sem for writing, split, mark VM_DEAD, drop mmap_sem. Acquire
>>>>> mmap_sem for reading, madv_free drop mmap_sem. Acquire mmap_sem for
>>>>> writing, free everything left, drop mmap_sem.
>>>>>
>>>>> ?
>>>>>
>>>>> Sure, you acquire the lock 3 times, but both write instances should be
>>>>> 'short', and I suppose you can do a demote between 1 and 2 if you care.
>>>> Thanks, Peter. Yes, by looking the code and trying two different approaches,
>>>> it looks this approach is the most straight-forward one.
>>> Yes, you just have to be careful about the max vma count limit.
>> Yes, we should just need copy what do_munmap does as below:
>>
>> if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
>>              return -ENOMEM;
>>
>> If the mas map count limit has been reached, it will return failure before
>> zapping mappings.
> Yeah, but as soon as you drop the lock and retake it, somebody might
> have changed the adddress space and we might get inconsistency.
>
> So I am wondering whether we really need upgrade_read (to promote read
> to write lock) and do the
> 	down_write
> 	split & set up VM_DEAD
> 	downgrade_write
> 	unmap
> 	upgrade_read
> 	zap ptes
> 	up_write

I'm supposed address space changing just can be done by mmap, mremap, 
mprotect. If so, we may utilize the new VM_DEAD flag. If the VM_DEAD 
flag is set for the vma, just return failure since it is being unmapped.

Does it sounds reasonable?

Thanks,
Yang

>
> looks terrible, no question about that, but we won't drop the mmap sem
> at any time.