Re: [PATCH bpf-next 1/2] bpf: Take return from set_memory_ro() into account with bpf_prog_lock_ro()

[Daniel Borkmann <daniel@iogearbox.net>]

On 2/19/24 7:39 AM, Christophe Leroy wrote:
> 
> 
> Le 19/02/2024 à 02:40, Hengqi Chen a écrit :
>> [Vous ne recevez pas souvent de courriers de hengqi.chen@gmail.com. Découvrez pourquoi ceci est important à https://aka.ms/LearnAboutSenderIdentification ]
>>
>> Hello Christophe,
>>
>> On Sun, Feb 18, 2024 at 6:55 PM Christophe Leroy
>> <christophe.leroy@csgroup.eu> wrote:
>>>
>>> set_memory_ro() can fail, leaving memory unprotected.
>>>
>>> Check its return and take it into account as an error.
>>>
>>
>> I don't see a cover letter for this series, could you describe how
>> set_memory_ro() could fail.
>> (Most callsites of set_memory_ro() didn't check the return values)
> 
> Yeah, there is no cover letter because as explained in patch 2 the two
> patches are autonomous. The only reason why I sent it as a series is
> because the patches both modify include/linux/filter.h in two places
> that are too close to each other.
> 
> I should have added a link to https://github.com/KSPP/linux/issues/7
> See that link for detailed explanation.
> 
> If we take powerpc as an exemple, set_memory_ro() is a frontend to
> change_memory_attr(). When you look at change_memory_attr() you see it
> can return -EINVAL in two cases. Then it calls
> apply_to_existing_page_range(). When you go down the road you see you
> can get -EINVAL or -ENOMEM from that function or its callees.

By that logic, don't you have the same issue when undoing all of this?
E.g. take arch_protect_bpf_trampoline() / arch_unprotect_bpf_trampoline()
which is not covered in here, but what happens if you set it first to ro
and later the setting back to rw fails? How would the error path there
look like? It's something you cannot recover.

Thanks,
Daniel