Re: [PATCH bpf-next v1 1/3] bpf: Add skb dynptrs

[Andrii Nakryiko <andrii.nakryiko@gmail.com>]

On Mon, Aug 1, 2022 at 2:16 PM Joanne Koong <joannelkoong@gmail.com> wrote:
>
> On Mon, Aug 1, 2022 at 12:38 PM Martin KaFai Lau <kafai@fb.com> wrote:
> >
> > On Mon, Aug 01, 2022 at 10:52:14AM -0700, Joanne Koong wrote:
> > > > Since we are on bpf_dynptr_write, what is the reason
> > > > on limiting it to the skb_headlen() ?  Not implying one
> > > > way is better than another.  would like to undertand the reason
> > > > behind it since it is not clear in the commit message.
> > > For bpf_dynptr_write, if we don't limit it to skb_headlen() then there
> > > may be writes that pull the skb, so any existing data slices to the
> > > skb must be invalidated. However, in the verifier we can't detect when
> > > the data slice should be invalidated vs. when it shouldn't (eg
> > > detecting when a write goes into the paged area vs when the write is
> > > only in the head). If the prog wants to write into the paged area, I
> > > think the only way it can work is if it pulls the data first with
> > > bpf_skb_pull_data before calling bpf_dynptr_write. I will add this to
> > > the commit message in v2
> > Note that current verifier unconditionally invalidates PTR_TO_PACKET
> > after bpf_skb_store_bytes().  Potentially the same could be done for
> > other new helper like bpf_dynptr_write().  I think this bpf_dynptr_write()
> > behavior cannot be changed later, so want to raise this possibility here
> > just in case it wasn't considered before.
>
> Thanks for raising this possibility. To me, it seems more intuitive
> from the user standpoint to have bpf_dynptr_write() on a paged area
> fail (even if bpf_dynptr_read() on that same offset succeeds) than to
> have bpf_dynptr_write() always invalidate all dynptr slices related to
> that skb. I think most writes will be to the data in the head area,
> which seems unfortunate that bpf_dynptr_writes to the head area would
> invalidate the dynptr slices regardless.

+1. Given bpf_skb_store_bytes() is a more powerful superset of
bpf_dynptr_write(), I'd keep bpf_dynptr_write() in such a form as to
play nicely with bpf_dynptr_data() pointers.

>
> What are your thoughts? Do you think you prefer having
> bpf_dynptr_write() always work regardless of where the data is? If so,
> I'm happy to make that change for v2 :)
>
> >
> > Thinking from the existing bpf_skb_{load,store}_bytes() and skb->data perspective.
> > If the user changes the skb by directly using skb->data to avoid calling
> > load_bytes()/store_bytes(), the user will do the necessary bpf_skb_pull_data()
> > before reading/writing the skb->data.  If load_bytes()+store_bytes() is used instead,
> > it would be hard to reason why the earlier bpf_skb_load_bytes() can load a particular
> > byte but [may] need to make an extra bpf_skb_pull_data() call before it can use
> > bpf_skb_store_bytes() to store a modified byte at the same offset.