Re: [PATCH bpf-next v3 1/2] bpf, cgroup: Don't populate prog_attach_flags array when effective query

[Martin KaFai Lau]

On 9/19/22 6:32 AM, Pu Lehui wrote:
> On 2022/9/17 8:03, Martin KaFai Lau wrote:
> > On 9/14/22 9:17 AM, Pu Lehui wrote:
> > > From: Pu Lehui <pulehui@xxxxxxxxxx>
> > >
> > > Attach flags is only valid for attached progs of this layer cgroup,
> > > but not for effective progs. For querying with EFFECTIVE flags,
> > > exporting attach flags does not make sense. so we don't need to
> > > populate prog_attach_flags array when effective query.
> >
> > prog_attach_flags has been added to 6.0 which is in rc5.  It is still 
> > doable (and cleaner) to reject prog_attach_flags when it is an 
> > effective_query.  This should be done regardless of 'type == 
> > BPF_LSM_CGROUP' or not.  Something like:
> >
> > if (effective_query && prog_attach_flags)
> >      return -EINVAL;
> >
> > Otherwise, the whole prog_attach_flags needs to be set to 0 during 
> > effective_query.  Please target the change to the bpf tree instead of 
> > bpf-next such that this uapi bit can be fixed before 6.0.
>
> Okay, will handle in next version.

Thanks.  It will also be useful to comment the uapi's bpf.h and mention
prog_attach_flags should not be set during effective_query.

> > Also, the effective_query issue is not limited to the 
> > prog_attach_flags? For the older uattr->query.attach_flags, it should 
> > be set to 0 also when it is an effective_query, right?
>
> For output uattr->query.attach_flags, we certainly don't need to copy it 
> to userspace when effective query. Since we do not utilize 
> uattr->query.attach_flags in the cgroup query function, should we need 
> to take it as input and reject when it is non-zero in effective query? 
> Something like:
> if (effective_query && (prog_attach_flags || attr->query.attach_flags))

No.  I don't think the zero attr->query.attach_flags can be enforced 
now.  It is used as an output value only and its input value has never 
been checked.  Although the bpftool always sets it to 0 before the 
query, checking zero now does not gain much while there is a slight 
chance of breaking other users.

Only need to set/output uattr->query.attach_flags as 0 during 
effective_query.

> For both output and input scenarios, we are faced with the problem that 
> there is a ambiguity in attach_flags being 0. When we do not copy to the 
> userspace, libbpf will set it to 0 by default, and 0 can mean NONE flag 
> attach, or no attach prog. The same is true for input scenarios.
>
> So should we need to define NONE attach flag and redefine the others? 
> Such as follow:
> #define BPF_F_ALLOW_NONE        (1U << 0)

I would not change the uapi for this.  0 implicitly means no flags or 
none.  Regardless, this change does not belong to the bpf tree where 
this fix will be landing.

> #define BPF_F_ALLOW_OVERRIDE    (1U << 1)
> #define BPF_F_ALLOW_MULTI       (1U << 2)
> #define BPF_F_REPLACE           (1U << 3)
>
> And then attach flags being 0 certainly means no attach any prog.