Re: [PATCH V11 0/4] BPF: New helper to obtain namespace data from current task

[Yonghong Song <yhs@fb.com>]

On 9/26/19 9:16 AM, John Fastabend wrote:
> Eric W. Biederman wrote:
>> Carlos Neira <cneirabustos@gmail.com> writes:
>>
>>> Currently bpf_get_current_pid_tgid(), is used to do pid filtering in bcc's
>>> scripts but this helper returns the pid as seen by the root namespace which is
>>> fine when a bcc script is not executed inside a container.
>>> When the process of interest is inside a container, pid filtering will not work
>>> if bpf_get_current_pid_tgid() is used.
>>> This helper addresses this limitation returning the pid as it's seen by the current
>>> namespace where the script is executing.
>>>
>>> In the future different pid_ns files may belong to different devices, according to the
>>> discussion between Eric Biederman and Yonghong in 2017 Linux plumbers conference.
>>> To address that situation the helper requires inum and dev_t from /proc/self/ns/pid.
>>> This helper has the same use cases as bpf_get_current_pid_tgid() as it can be
>>> used to do pid filtering even inside a container.
>>
>> I think I may have asked this before.  If I am repeating old gound
>> please excuse me.
>>
>> Am I correct in understanding these new helpers are designed to be used
>> when programs running in ``conainers'' call it inside pid namespaces
>> register bpf programs for tracing?
>>
>> If so would it be possible to change how the existing bpf opcodes
>> operate when they are used in the context of a pid namespace?
>>
>> That later would seem to allow just moving an existing application into
>> a pid namespace with no modifications.   If we can do this with trivial
>> cost at bpf compile time and with no userspace changes that would seem
>> a better approach.
>>
>> If not can someone point me to why we can't do that?  What am I missing?
> 
> We have some management/observabiliity bpf programs loaded from privileged
> containers that end up getting triggered in multiple container context. Here
> we want the root namespace pid otherwise there would be collisions (same pid
> in multiple containers) when its used as a key and we would have difficulty
> finding the pid from the root namespace.

Yes, using root namespace pid will work.

I am referring to a priviledged container (current root, and future may
just CAP_BPF and CAP_TRACIING) where you do not need to go to root
to check root pids. Also, there are cases, we do pid namespace-scope 
statistics collecting, filtering based on namespace "id" is also needed.

> 
> I guess at load time if its an unprivileged program we could convert it to
> use the pid of the current namespace?

This way we will need to helper to get current namespace pid.

> 
> Or if the application is moved into a unprivileged container?

Ya. A helper will be needed.

> 
> Our code is outside bcc so not sure exactly how the bcc case works. Just
> wanted to point out we use the root namespace pid for various things
> so I think it might need to be a bit smarter than just the moving an
> existing application into a pid namespace.

As a workaround, we do this as well. The goal is to improve usability.
So we do not need to go to root to find these pids.
Sometimes if filtering at namespace level, we have to approximate as 
sometimes it is impossible to track all pids in the container.

> 
> .John
>