Re: [PATCH] mm/mmap: Map MAP_STACK to VM_STACK

[Waiman Long <longman@redhat.com>]

On 4/18/23 21:36, Hugh Dickins wrote:
> On Tue, 18 Apr 2023, Waiman Long wrote:
>> On 4/18/23 17:18, Andrew Morton wrote:
>>> On Tue, 18 Apr 2023 17:02:30 -0400 Waiman Long <longman@redhat.com> wrote:
>>>
>>>> One of the flags of mmap(2) is MAP_STACK to request a memory segment
>>>> suitable for a process or thread stack. The kernel currently ignores
>>>> this flags. Glibc uses MAP_STACK when mmapping a thread stack. However,
>>>> selinux has an execstack check in selinux_file_mprotect() which disallows
>>>> a stack VMA to be made executable.
>>>>
>>>> Since MAP_STACK is a noop, it is possible for a stack VMA to be merged
>>>> with an adjacent anonymous VMA. With that merging, using mprotect(2)
>>>> to change a part of the merged anonymous VMA to make it executable may
>>>> fail. This can lead to sporadic failure of applications that need to
>>>> make those changes.
>>> "Sporadic failure of applications" sounds quite serious.  Can you
>>> provide more details?
>> The problem boils down to the fact that it is possible for user code to mmap a
>> region of memory and then for the kernel to merge the VMA for that memory with
>> the VMA for one of the application's thread stacks. This is causing random
>> SEGVs with one of our large customer application.
>>
>> At a high level, this is what's happening:
>>
>>   1) App runs creating lots of threads.
>>   2) It mmap's 256K pages of anonymous memory.
>>   3) It writes executable code to that memory.
>>   4) It calls mprotect() with PROT_EXEC on that memory so
>>      it can subsequently execute the code.
>>
>> The above mprotect() will fail if the mmap'd region's VMA gets merged with the
>> VMA for one of the thread stacks.  That's because the default RHEL SELinux
>> policy is to not allow executable stacks.
> Then wouldn't the bug be at the SELinux end?  VMAs may have been merged
> already, but the mprotect() with PROT_EXEC of the good non-stack range
> will then split that area off from the stack again - maybe the SELinux
> check does not understand that must happen?

The SELinux check is done per VMA, not a region within a VMA. After VMA 
merging, SELinux is probably not able to determine which part of a VMA 
is a stack unless we keep that information somewhere and provide an API 
for SELinux to query. That can be quite a lot of work. So the easiest 
way to prevent this problem is to avoid merging a stack VMA with a 
regular anonymous VMA.

Cheers,
Longman