Re: [PATCH] bpf: Specify access type of bpf_sysctl_get_name args

[Jerome Marchand <jmarchan@redhat.com>]

On 27/05/2025 23:39, Eduard Zingerman wrote:
> "Jerome Marchand" <jmarchan@redhat.com> writes:
> 
>> The second argument of bpf_sysctl_get_name() helper is a pointer to a
>> buffer that is being written to. However that isn't specify in the
>> prototype.
>>
>> Until commit 37cce22dbd51a ("bpf: verifier: Refactor helper access
>> type tracking"), all helper accesses were considered as a possible
>> write access by the verifier, so no big harm was done. However, since
>> then, the verifier might make wrong asssumption about the content of
>> that address which might lead it to make faulty optimizations (such as
>> removing code that was wrongly labeled dead). This is what happens in
>> test_sysctl selftest to the tests related to sysctl_get_name.
>>
>> Correctly mark the second argument of bpf_sysctl_get_name() as
>> ARG_PTR_TO_UNINIT_MEM.
>>
>> Signed-off-by: Jerome Marchand <jmarchan@redhat.com>
>> ---
> 
> Looks like we don't run bpf_sysctl_get_name tests on the CI.
> CI executes the following binaries:
> - test_progs{,-no_alu32,-cpuv4}
> - test_verifier
> - test_maps
> test_progs is what is actively developed.
> 
> I agree with the reasoning behind this patch, however, could you please
> add a selftest demonstrating unsafe behaviour?

Do you mean to write a selftest that demonstrate the current unsafe 
behavior of the bpf_sysctl_get_name helper? I could write something 
similar as the failing test_sysctl cases.

I'm thinking that a more general test that would check that helpers 
don't access memory in a different way than advertised in their 
prototype would be more useful. But that's quite a different endeavor.


Regards,
Jerome

> You can use tools/testing/selftests/bpf/progs/verifier_and.c as an
> example of verifier test checking for specific log message.
> (framework also supports execution if __retval is specified,
>   tests can be written in plain C as well, e.g. as in .../iters.c).
>