Re: [PATCH bpf-next v7 1/3] bpf: Allow trusted pointers to be passed to KF_TRUSTED_ARGS kfuncs

[David Vernet]

On Fri, Nov 18, 2022 at 08:13:37PM -0800, Alexei Starovoitov wrote:
> On Fri, Nov 18, 2022 at 03:44:42PM -0600, David Vernet wrote:
> > > > if it's a release arg it should always have a refcount on it.
> > > > PTR_UNTRUSTED | PTR_TRUSTED would also make no sense. MEM_FIXED_SIZE
> > > > though seems fine? In general, I thought it was prudent for us to take
> > > > the most conservative possible approach here, which is that PTR_TRUSTED
> > > > only applies when no other modifiers are present, and it applies for all
> > > > obj_ptr types (other than PTR_TO_CTX which does its own thing).
> > > 
> > > Probably worth refining when PTR_TRUSTED is cleared.
> > > For example adding PTR_UNTRUSTED should definitely clear it.



> > 
> > That makes sense for PTR_UNTRUSTED, what about the other type modifiers
> > like PTR_MAYBE_NULL? We set and unset if a ptr is NULL throughout a
> > function, so we'd have to record if it was previously trusted in order
> > to properly re-OR after a NULL check.
> 
> PTR_MAYBE_NULL is another bit and I don't think it conflicts with PTR_TRUSTED.
> PTR_TO_BTF_ID | PTR_TRUSTED is a valid pointer.
> PTR_TO_BTF_ID | PTR_TRUSTED | PTR_MAYBE_NULL is a valid pointer or NULL.
> 
> PTR_TO_BTF_ID | PTR_MAYBE_NULL is a legacy "valid pointer" or NULL.
> That legacy pointer cannot be passed to KF_TRUSTED_ARGS kfuncs.
> 
> KF_TRUSTED_ARGS kfuncs should not accept PTR_TO_BTF_ID | PTR_TRUSTED | PTR_MAYBE_NULL.

Indeed -- my point was that I don't think e.g. clearing PTR_TRUSTED when
we set PTR_UNTRUSTED will work, at least not yet. It's still too tricky
to find all the places where we'd have to &= ~PTR_TRUSTED or |=
PTR_TRUSTED when setting specific type modifiers. As described below, we
first have to clarify the general workflow to enable the presence of
PTR_TRUSTED to be the single source of truth for trust.

> It's a job of the prog to do != NULL check.
> Otherwise all such != NULL checks would need to move inside kfuncs which is not good.
> 
> > > MEM_ALLOC flag is probably equivalent to PTR_TRUSTED.
> > > Maybe the bit:
> > > regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
> > > should set PTR_TRUSTED as well?
> > 
> > We could, but that changes the meaning of PTR_TRUSTED and IMO makes it
> > harder to reason about. Before it was just "the kernel passed this arg
> > to the program and promises the program that it was trusted when it was
> > first passed". Now it's that plus it could mean that it points to an
> > allocated object from bpf_obj_new()". IMO we should keep all of these
> > modifiers separate so that the presence of a modifier has a well-defined
> > meaning that we can interpret in each context as needed.  In this case,
> > we can make trust opt-in, so a KF_TRUSTED_ARGS BTF pointer either of the
> > following:
> > 
> > 1. reg->ref_obj_id > 0
> > 2. Either one of PTR_TRUSTED | MEM_ALLOC type modifiers are set, and no
> >    others.
> 
> I don't think MEM_ALLOC conflicts with PTR_TRUSTED.
> MEM_ALLOC flags means that it came from bpf_obj_new() and that's what
> bpf_spin_lock and bpf_obj_drop() want to see.
> 
> Adding PTR_TRUSTED to MEM_ALLOC looks necessary to me.
> It doesn't have to be done right now, but eventually feels right.

I think I agree. MEM_ALLOC should always imply PTR_TRUSTED. Ideally we
shouldn't have to check MEM_ALLOC for KF_TRUSTED_ARGS at all, and
PTR_TRUSTED should be the only modifier representing if something is
safe. For now I'd prefer to keep them separate until we have a clear
plan, especially with respect to clearing PTR_TRUSTED for when something
unsafe happens like PTR_UNTRUSTED or PTR_MAYBE_NULL. It's all too
muddied still.

> I've been thinking whether reg->ref_obj_id > 0 condition should be converted
> to PTR_TRUSTED too...
> On one side it will simplify the check for KF_TRUSTED_ARGS.
> The only thing check_kfunc_args() would need to do is:
> is_kfunc_trusted_args(meta)
> && type_flag(reg->type) & PTR_TRUSTED
> && !(type_flag(reg->type) & PTR_MAYBE_NULL)
> 
> On the other side fixing all places where we set ref_obj_id
> and adding |= PTR_TRUSTED may be too cumbersome ?

I think it's probably too cumbersome now, but yeah, as mentioned above,
I think it's the right direction. I think it will require a lot of
thought to do it right, though. With the code the way that it is now, I
can't convince myself that we wouldn't do something like |= PTR_TRUSTED
when we observe ref_obj_id > 0, and then later &= ~PTR_TRUSTED when
setting PTR_MAYBE_NULL. I think Kumar's latest patch set is a nice step
towards achieving this clearer state. Hopefully we can continue to
improve.

> Right now we're saying PTR_TO_CTX is implicitly trusted, but we can OR
> PTR_TO_CTX with PTR_TRUSTED to make it explicit and truly generalize the check.

Further agreed, this is the correct longer-term direction.

> > Agreed that after the rebase this would no longer be correct. I think we
> > should make it opt-in, though. PTR_TRUSTED | MEM_ALLOC is fine.
> > PTR_TRUSTED | MEM_ALLOC | PTR_MAYBE_NULL would not be.
> 
> to pass into KF_TRUSTED_ARGS kfunc? Agree.
> I guess we can tighten the check a bit:
> is_kfunc_trusted_args(meta)
> && type_flag(reg->type) & PTR_TRUSTED
> && !(type_flag(reg->type) & ~(PTR_TRUSTED | MEM_ALLOC))
> 
> In english: the pointer should have PTR_TRUSTED flag and
> no other flags other than PTR_TRUSTED and MEM_ALLOC should be set.

Yeah, I think this is the correct way to model this for now. Later on
once we refactor things, the presence of PTR_TRUSTED on its own should
be sufficient. A good north star to aim towards.

I'll send this out as v8 tomorrow.