Re: [PATCH memcg v6] net: set proper memcg for net_init hooks allocations

[Vasily Averin]

On 6/7/22 08:58, Shakeel Butt wrote:
> On Mon, Jun 6, 2022 at 11:45 AM Vasily Averin <vvs@xxxxxxxxxx> wrote:
>>
> [...]
>>
>> As far as I understand this report means that 'init_net' have incorrect
>> virtual address on arm64.
> 
> So, the two call stacks tell the addresses belong to the kernel
> modules (nfnetlink and nf_tables) whose underlying memory is allocated
> through vmalloc and virt_to_page() does not work on vmalloc()
> addresses.

However in both these cases get_mem_cgroup_from_obj() -> mem_cgroup_from_obj() ->
virt_to_folio() -> virt_to_page() -> virt_to_pfn() -> __virt_to_phys() 
handles address of struct net taken from for_each_net().
The only net namespace that exists at this stage is init_net,
and dmesg output confirms this:
"virt_to_phys used for non-linear address: ffffd8efe2d2fe00 (init_net)"

>> Roman, Shakeel, I need your help
>>
>> Should we perhaps verify kaddr via virt_addr_valid() before using virt_to_page()
>> If so, where it should be checked?
> 
> I think virt_addr_valid() check in mem_cgroup_from_obj() should work
> but I think it is expensive on the arm64 platform. The cheaper and a
> bit hacky way to avoid such addresses is to directly use
> is_vmalloc_addr() directly.

I do not understand why you mean that processed address is vmalloc-specific.
As far as I understand it is valid address of static variable, and for some reason
arm64 does not consider them valid virtual addresses.

Thank you,
	Vasily Averin