Re: [PATCH] arm64: pageattr: Explicitly bail out when changing permissions for vmalloc_huge mappings

[Ryan Roberts]

On 01/04/2025 06:12, Mike Rapoport wrote:
> Hi Ryan,
> 
> On Tue, Apr 01, 2025 at 10:43:01AM +0100, Ryan Roberts wrote:
>> On 30/03/2025 03:32, Mike Rapoport wrote:
>>> On Sat, Mar 29, 2025 at 09:46:56AM +0000, Ryan Roberts wrote:
>>>> On 28/03/2025 18:50, Mike Rapoport wrote:
>>>>> On Fri, Mar 28, 2025 at 11:51:03AM +0530, Dev Jain wrote:
>>>>>> arm64 uses apply_to_page_range to change permissions for kernel VA mappings,
>>>>>
>>>>>                                                      for vmalloc mappings ^
>>>>>
>>>>> arm64 does not allow changing permissions to any VA address right now.
>>>>>
>>>>>> which does not support changing permissions for leaf mappings. This function
>>>>>> will change permissions until it encounters a leaf mapping, and will bail
>>>>>> out. To avoid this partial change, explicitly disallow changing permissions
>>>>>> for VM_ALLOW_HUGE_VMAP mappings.
>>>>>>
>>>>>> Signed-off-by: Dev Jain <dev.jain@xxxxxxx>
>>>>
>>>> I wonder if we want a Fixes: tag here? It's certainly a *latent* bug.
>>>
>>> We have only a few places that use vmalloc_huge() or VM_ALLOW_HUGE_VMAP and
>>> if there was a code that plays permission games on these allocations, x86
>>> set_memory would blow up immediately, so I don't think Fixes: is needed
>>> here.
>>
>> Hi Mike,
>>
>> I think I may have misunderstood your comments when we spoke at LSF/MM the other
>> day, as this statement seems to contradict. I thought you said that on x86 BPF
>> allocates memory using vmalloc_huge()/VM_ALLOW_HUGE_VMAP and then it's
>> sub-allocator will set_memory_*() on a sub-region of that allocation? (And we
>> then agreed that it would be good for arm64 to eventually support this with BBML2).
> 
> I misremembered :)
> They do allocate several PMD_SIZE chunks at once, but they don't use
> vmalloc_huge(), so everything there is mapped with base pages.
> 
> And now they are using execmem rather than vmalloc directly, and execmem
> doesn't use VM_ALLOW_HUGE_VMAP anywhere except modules on x86.

Ahh OK! Like I said, I don't think it reduces the value of this patch though.

>  
>> Anyway, regardless, I think this change is useful first step to improving
>> vmalloc as it makes us more defensive against any future attempt to change
>> permissions on a huge allocation. In the long term I'd like to get to the point
>> where arm64 (with BBML2) can map with VM_ALLOW_HUGE_VMAP by default.
>>
>> Thanks,
>> Ryan
>