Message-ID: <20221103005151.wfil26boeemar6uk@MacBook-Pro-5.local.dhcp.thefacebook.com>
Date:   Wed, 2 Nov 2022 17:51:51 -0700
From:   Alexei Starovoitov <alexei.starovoitov@...il.com>
To:     Kumar Kartikeya Dwivedi <memxor@...il.com>
Cc:     David Vernet <void@...ifault.com>, bpf <bpf@...r.kernel.org>,
        Alexei Starovoitov <ast@...nel.org>,
        Daniel Borkmann <daniel@...earbox.net>,
        Andrii Nakryiko <andrii@...nel.org>,
        Martin KaFai Lau <martin.lau@...ux.dev>,
        Song Liu <song@...nel.org>, Yonghong Song <yhs@...com>,
        John Fastabend <john.fastabend@...il.com>,
        KP Singh <kpsingh@...nel.org>,
        Stanislav Fomichev <sdf@...gle.com>,
        Hao Luo <haoluo@...gle.com>, Jiri Olsa <jolsa@...nel.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Kernel Team <kernel-team@...com>, Tejun Heo <tj@...nel.org>
Subject: Re: [PATCH bpf-next v6 1/3] bpf: Allow trusted pointers to be passed
 to KF_TRUSTED_ARGS kfuncs

On Thu, Nov 03, 2022 at 02:18:35AM +0530, Kumar Kartikeya Dwivedi wrote:
> On Wed, Nov 02, 2022 at 08:04:57AM IST, Alexei Starovoitov wrote:
> > On Tue, Nov 1, 2022 at 6:01 PM Kumar Kartikeya Dwivedi <memxor@...il.com> wrote:
> > >
> > > On Wed, 2 Nov 2022 at 06:02, Alexei Starovoitov
> > > <alexei.starovoitov@...il.com> wrote:
> > > >
> > > > On Wed, Nov 02, 2022 at 04:01:11AM +0530, Kumar Kartikeya Dwivedi wrote:
> > > > > On Wed, 2 Nov 2022 at 03:06, David Vernet <void@...ifault.com> wrote:
> > > > > >
> > > > > > On Tue, Nov 01, 2022 at 01:22:39PM -0700, Alexei Starovoitov wrote:
> > > > > > > On Tue, Nov 1, 2022 at 11:11 AM David Vernet <void@...ifault.com> wrote:
> > > > > > > >
> > > > > > > > > What kind of bpf prog will be able to pass 'struct nf_conn___init *' into these bpf_ct_* ?
> > > > > > > > > We've introduced / vs nf_conf specifically to express the relationship
> > > > > > > > > between allocated nf_conn and other nf_conn-s via different types.
> > > > > > > > > Why is this not enough?
> > > > > > > >
> > > > > > > > Kumar should have more context here (he originally suggested this in
> > > > > > > > [0]),
> > > > > > >
> > > > > > > Quoting:
> > > > > > > "
> > > > > > > Unfortunately a side effect of this change is that now since
> > > > > > > PTR_TO_BTF_ID without ref_obj_id is considered trusted, the bpf_ct_*
> > > > > > > functions would begin working with tp_btf args.
> > > > > > > "
> > > > > > > I couldn't find any tracepoint that has nf_conn___init as an argument.
> > > > > > > The whole point of that new type was to return it to bpf prog,
> > > > > > > so the verifier type matches it when it's passed into bpf_ct_*
> > > > > > > in turn.
> > > > > > > So I don't see a need for a new OWNED flag still.
> > > > > > > If nf_conn___init is passed into tracepoint it's a bug and
> > > > > > > we gotta fix it.
> > > > > >
> > > > > > Yep, this is what I'm seeing as well. I think you're right that
> > > > > > KF_OWNED_ARGS is just strictly unnecessary and that creating wrapper
> > > > > > types is the way to enable an ownership model like this.
> > > > > >
> > > > >
> > > > > It's not just nf_conn___init. Some CT helpers also take nf_conn.
> > > > > e.g. bpf_ct_change_timeout, bpf_ct_change_status.
> > > > > Right now they are only allowed in XDP and TC programs, so the tracing
> > > > > args part is not a problem _right now_.
> > > >
> > > > ... and it will be fine to use bpf_ct_change_timeout from tp_btf as well.
> > > >
> > > > > So currently it may not be possible to pass such a trusted but
> > > > > ref_obj_id == 0 nf_conn to those helpers.
> > > > > But based on changes unrelated to this, it may become possible in the
> > > > > future to obtain such a trusted nf_conn pointer.
> > > >
> > > > From kfunc's pov trusted pointer means valid pointer.
> > > > It doesn't need to be ref_obj_id refcounted from the verifier pov.
> > > > It can be refcounted on the kernel side and it will be trusted.
> > > > The code that calls trace_*() passes only trusted pointers into tp-s.
> > > > If there is a tracepoint somewhere in the kernel that uses a volatile
> > > > pointer to potentially uaf kernel object it's a bug that should be fixed.
> > > >
> > >
> > > This is all fine. I'm asking you to distinguish between
> > > trusted-not-refcounted and trusted-and-refcounted.
> >
> > That's not what you're asking :)
> >
> > > It is necessary for nf_conn, since the object can be reused if the
> > > refcount is not held.
> >
> > of course. No one argues the opposite.
> >
> > > Some other CPU could be reusing the same memory and allocating a new
> > > nf_conn on it while we change its status.
> > > So it's not ok to call bpf_ct_change_timeout/status on trusted
> > > nf_conn, but only on trusted+refcounted nf_conn.
> >
> > and here we start to disagree.
> >
> > > Trusted doesn't capture the difference between 'valid' vs 'valid and
> > > owned by prog' anymore with the new definition
> > > for PTR_TO_BTF_ID.
> >
> > and here we disagree completely.
> > You're asking to distinguish refcnt++ done by the program
> > and recognized by the verifier as ref_obj_id > 0 vs
> > refcnt++ done by the kernel code before it calls into tracepoint.
> > That's odd, right?
> > I don't think people adding kfuncs should care what piece
> > of code before kfunc did refcnt++.
> >
> 
> I think we're talking past each other. Your point is that whenever PTR_TRUSTED
> is set (tracepoint args, etc.) the kernel itself should ensure it holds the
> refcount on behalf of the program. There shouldn't be a scenario where
> PTR_TRUSTED nf_conn is obtained and the program or the kernel isn't holding
> an active refcount during its use with helpers/kfuncs.

correct.

> I was not making that assumption, but it's not wrong. Which is why I wanted to
> limit their use to only when verifier is tracking a reference in the program
> for the nf_conn returned from an acquire kfunc.

limit the use for PTR_TRUSTED to verifier only?
and introduce PTR_TRUSTED_BECAUSE_IT_CAME_FROM_KERNEL as well?
But the patch we're discussing will be treating them the same way...
Or you're proposing the kfunc writer also suppose to add two flags to a kfunc:
KF_TRUSTED_ARGS | KF_TRUSTED_BECAUSE_KERNEL_ARGS ?

and that's where it comes back to the point I keep stressing in this thread:
The kfunc writer should not be exposed to such details.
The little people need to understand about the verifier when the write bpf
programs or provide kfuncs the better.

> > > Yes, in most cases the tracepoints/tracing functions whitelisted will
> > > have the caller ensure that,
> > > but we should then allow trusted nf_conn in those hooks explicitly,
> > > not implicitly by default everywhere.
> > > Until then it should be restricted to ref_obj_id > 0 IMO as it is right now.
> > >
> > > > > It is a requirement of those kfuncs that the nf_conn has its refcount
> > > > > held while they are called.
> > > >
> > > > and it will be. Just not by the verifier.
> > > >
> > > > > KF_TRUSTED_ARGS was encoding this requirement before, but it wouldn't anymore.
> > > > > It seems better to me to keep that restriction instead of relaxing it,
> > > > > if it is part of the contract.
> > > >
> > > > Disagree as explained above.
> > > >
> > > > > It is fine to not require people to dive into these details and just
> > > > > use KF_TRUSTED_ARGS in general, but we need something to cover special
> > > > > cases like these where the object is only stable while we hold an
> > > > > active refcount, RCU protection is not enough against reuse.
> > > >
> > > > This is not related to RCU. Let's not mix RCU concerns in here.
> > > > It's a different topic.
> > > >
> > >
> > > What I meant is that in the normal case, usually objects aren't reused
> > > while the RCU read lock is held.
> > > In case of nf_conn, the refcount needs to be held to ensure that,
> > > since it uses SLAB_TYPESAFE_BY_RCU.
> > > This is why bpf_ct_lookup needs to bump the refcount and match the key
> > > after that again, and cannot just return the looked up ct directly.
> >
> > bpf_ct_lookup needs to bump a refcnt?!
> > bpf_skb_ct_lookup calls __bpf_nf_ct_lookup
> > that calls nf_conntrack_find_get() that does
> > the search and incs the refcnt in a generic kernel code.
> > There is nothing bpf specific stuff here. bpf kfunc didn't
> > add any special refcnt incs.
> >
> > There are no tracepoints in netfilter, so this discussion
> > is all theoretical, but if there was then the code
> > should have made sure that refcnt is held before
> > passing nf_conn into tracepoint.
> >
> 
> Right, if so, it makes sense. Thanks for explaining.
> 
> > > > > It could be 'expert only' __ref suffix on the nf_conn arg, or
> > > > > KF_OWNED_ARGS, or something else.
> > > >
> > > > I'm still against that.
> > > >
> > >
> > > I understand (and agree) that you don't want to complicate things further.
> > > It's fine if you want to deal with this later when the above concern
> > > materializes. But it will be yet another thing to keep in mind for the
> > > future.
> >
> > I don't share the concern.
> > With nf_conn there is none, right?
> > But imagine there is only RCU protected pointer that
> > is passed into tracepoint somewhere.
> > The verifier doesn't recognize refcnt++ on it and ref_obj_id == 0
> > the kernel code doesn't do refcnt++ either.
> > But it's still safe and this arg should still be
> > PTR_TO_BTF_ID | PTR_TRUSTED.
> > The bpf prog can pass it further into kfunc that has KF_TRUSTED_ARGS.
> > Since RCU is held before calling into tracepoint the bpf prog
> > has to be non sleepable. Additional rcu_read_lock done by
> > the prog is redundant, but doesn't hurt.
> > When prog is calling kfunc the pointer is still valid and
> > kfunc can safely operate on it assuming that object is not going away.
> > That is the definition of KF_TRUSTED_ARGS from pov of kfunc.
> > You documented it yourself :)
> > "
> > The KF_TRUSTED_ARGS flag is used for kfuncs taking pointer arguments. It
> > indicates that the all pointer arguments will always have a guaranteed lifetime,
> > "
> 
> So based on the above, verifier should only mark nf_conn as PTR_TRUSTED in cases
> where the kernel already holds the refcount on behalf of the program, 

yes, but 'holding the reference' is not the only option.
Holding rcu_read_lock is fine too.
Or any other mechanism that ensures that the pointer has guaranteed lifetime.

and back to my earlier point...
using your logic we'd need PTR_TRUSTED_BECAUSE_RCU in the verifier and
corresponding KF_TRUSTED_ARGS_BECAUSE_RCU flag on kfunc.
That just doesn't scale.

PTR_RCU flag might actually be needed for some other reason, but
KF_TRUSTED_ARGS_BECAUSE_RCU is "too much info" for kfunc providers.

> otherwise
> for the non-PTR_TRUSTED case program must have ref_obj_id > 0?
> 
> I.e. 'the guaranteed lifetime' for the nf_conn case also means the refcount is
> held in context where program can obtain PTR_TRUSTED nf_conn pointer.

yes.

> It will surely be true in the tracepoint case if there ever is one. Won't work
> easily for fentry/fexit but you already said that PTR_TRUSTED shouldn't be set
> in that case, atleast by default and without allowlisting those BTF IDs.

Correct. allowlisting kernel func's btf_ids is one approach.
I might argue that fentry hook on any kernel func that has 'struct nf_conn *'
as an argument is also safe (barring recursion into bpf_ct_* kfuncs).
Dual compiler of the kernel (native + bpf) will help this kind of safety analysis.

Powered by blists - more mailing lists