lkml.org 
[lkml]   [2022]   [Jun]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH v2 1/3] mm/swapfile: make security_vm_enough_memory_mm() work as expected
Date
Miaohe Lin <linmiaohe@huawei.com> writes:

> On 2022/6/21 9:35, Huang, Ying wrote:
>> Miaohe Lin <linmiaohe@huawei.com> writes:
>>
>>> On 2022/6/20 15:31, Huang, Ying wrote:
>>>> Miaohe Lin <linmiaohe@huawei.com> writes:
>>>>
>>>>> security_vm_enough_memory_mm() checks whether a process has enough memory
>>>>> to allocate a new virtual mapping. And total_swap_pages is considered as
>>>>> available memory while swapoff tries to make sure there's enough memory
>>>>> that can hold the swapped out memory. But total_swap_pages contains the
>>>>> swap space that is being swapoff. So security_vm_enough_memory_mm() will
>>>>> success even if there's no memory to hold the swapped out memory because
>>>>> total_swap_pages always greater than or equal to p->pages.
>>>>
>>>> Per my understanding, swapoff will not allocate virtual mapping by
>>>> itself. But after swapoff, the overcommit limit could be exceeded.
>>>> security_vm_enough_memory_mm() is used to check that. For example, in a
>>>> system with 4GB memory and 8GB swap, and 10GB is in use,
>>>>
>>>> CommitLimit: 4+8 = 12GB
>>>> Committed_AS: 10GB
>>>>
>>>> security_vm_enough_memory_mm() in swapoff() will fail because
>>>> 10+8 = 18 > 12. This is expected because after swapoff, the overcommit
>>>> limit will be exceeded.
>>>>
>>>> If 3GB is in use,
>>>>
>>>> CommitLimit: 4+8 = 12GB
>>>> Committed_AS: 3GB
>>>>
>>>> security_vm_enough_memory_mm() in swapoff() will succeed because
>>>> 3+8 = 11 < 12. This is expected because after swapoff, the overcommit
>>>> limit will not be exceeded.
>>>
>>> In OVERCOMMIT_NEVER scene, I think you're right.
>>>
>>>>
>>>> So, what's the real problem of the original implementation? Can you
>>>> show it with an example as above?
>>>
>>> In OVERCOMMIT_GUESS scene, in a system with 4GB memory and 8GB swap, and 10GB is in use,
>>> pages below is 8GB, totalram_pages() + total_swap_pages is 12GB, so swapoff() will succeed
>>> instead of expected failure because 8 < 12. The overcommit limit is always *ignored* in the
>>> below case.
>>>
>>> if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
>>> if (pages > totalram_pages() + total_swap_pages)
>>> goto error;
>>> return 0;
>>> }
>>>
>>> Or am I miss something?
>>
>> Per my understanding, with OVERCOMMIT_GUESS, the number of in-use pages
>> isn't checked at all. The only restriction is that the size of the
>> virtual mapping created should be less than total RAM + total swap
>
> Do you mean the only restriction is that the size of the virtual mapping
> *created every time* should be less than total RAM + total swap pages but
> *total virtual mapping* is not limited in OVERCOMMIT_GUESS scene? If so,
> the current behavior should be sane and I will drop this patch.

Yes. This is my understanding.

Best Regards,
Huang, Ying

> Thanks!
>
>> pages. Because swapoff() will not create virtual mapping, so it's
>> expected that security_vm_enough_memory_mm() in swapoff() always
>> succeeds.
>>
>> Best Regards,
>> Huang, Ying
>>
>>>
>>> Thanks!
>>>
>>>>
>>>>> In order to fix it, p->pages should be retracted from total_swap_pages
>>>>> first and then check whether there's enough memory for inuse swap pages.
>>>>>
>>>>> Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
>>>>
>>>> [snip]
>>>>
>>>> .
>>>>
>>
>> .
>>

\
 
 \ /
  Last update: 2022-06-21 09:48    [W:0.069 / U:0.204 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site