Re: [PATCH bpf-next v2 4/7] net: sched: add lightweight update path for cls_bpf

[Kumar Kartikeya Dwivedi <memxor@gmail.com>]

On Tue, Jun 08, 2021 at 05:02:04AM IST, Andrii Nakryiko wrote:
> On Thu, Jun 3, 2021 at 11:32 PM Kumar Kartikeya Dwivedi
> <memxor@gmail.com> wrote:
> >
> > This is used by BPF_LINK_UPDATE to replace the attach SCHED_CLS bpf prog
> > effectively changing the classifier implementation for a given filter
> > owned by a bpf_link.
> >
> > Note that READ_ONCE suffices in this case as the ordering for loads from
> > the filter are implicitly provided by the data dependency on BPF prog
> > pointer.
> >
> > On the writer side we can just use a relaxed WRITE_ONCE store to make
> > sure one or the other value is visible to a reader in cls_bpf_classify.
> > Lifetime is managed using RCU so bpf_prog_put path should wait until
> > readers are done for old_prog.
> >
> > All other parties accessing the BPF prog are under RTNL protection, so
> > need no changes.
> >
> > Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>.
> > Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
> > ---
> >  net/sched/cls_bpf.c | 55 +++++++++++++++++++++++++++++++++++++++++++--
> >  1 file changed, 53 insertions(+), 2 deletions(-)
> >
> > diff --git a/net/sched/cls_bpf.c b/net/sched/cls_bpf.c
> > index bf61ffbb7fd0..f23304685c48 100644
> > --- a/net/sched/cls_bpf.c
> > +++ b/net/sched/cls_bpf.c
> > @@ -9,6 +9,7 @@
> >   * (C) 2013 Daniel Borkmann <dborkman@redhat.com>
> >   */
> >
> > +#include <linux/atomic.h>
> >  #include <linux/module.h>
> >  #include <linux/types.h>
> >  #include <linux/skbuff.h>
> > @@ -104,11 +105,11 @@ static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
> >                         /* It is safe to push/pull even if skb_shared() */
> >                         __skb_push(skb, skb->mac_len);
> >                         bpf_compute_data_pointers(skb);
> > -                       filter_res = BPF_PROG_RUN(prog->filter, skb);
> > +                       filter_res = BPF_PROG_RUN(READ_ONCE(prog->filter), skb);
> >                         __skb_pull(skb, skb->mac_len);
> >                 } else {
> >                         bpf_compute_data_pointers(skb);
> > -                       filter_res = BPF_PROG_RUN(prog->filter, skb);
> > +                       filter_res = BPF_PROG_RUN(READ_ONCE(prog->filter), skb);
> >                 }
> >
> >                 if (prog->exts_integrated) {
> > @@ -775,6 +776,55 @@ static int cls_bpf_link_detach(struct bpf_link *link)
> >         return 0;
> >  }
> >
> > +static int cls_bpf_link_update(struct bpf_link *link, struct bpf_prog *new_prog,
> > +                              struct bpf_prog *old_prog)
> > +{
> > +       struct cls_bpf_link *cls_link;
> > +       struct cls_bpf_prog cls_prog;
> > +       struct cls_bpf_prog *prog;
> > +       int ret;
> > +
> > +       rtnl_lock();
> > +
> > +       cls_link = container_of(link, struct cls_bpf_link, link);
> > +       if (!cls_link->prog) {
> > +               ret = -ENOLINK;
> > +               goto out;
> > +       }
> > +
> > +       prog = cls_link->prog;
> > +
> > +       /* BPF_F_REPLACEing? */
> > +       if (old_prog && prog->filter != old_prog) {
> > +               ret = -EINVAL;
> > +               goto out;
> > +       }
> > +
> > +       old_prog = prog->filter;
> > +
> > +       if (new_prog == old_prog) {
> > +               ret = 0;
>
> So the contract is that if update is successful, new_prog's refcount
> taken by link_update() in kernel/bpf/syscall.c is transferred here. On
> error, it will be bpf_prog_put() by link_update(). So here you don't
> need extra refcnt, but it's also not an error, so you need to
> bpf_prog_put(new_prog) explicitly to balance out refcnt. See how it's
> done for XDP, for example.
>

Yes, thanks for spotting this.

>
> > +               goto out;
> > +       }
> > +
> > +       cls_prog = *prog;
> > +       cls_prog.filter = new_prog;
> > +
> > +       ret = cls_bpf_offload(prog->tp, &cls_prog, prog, NULL);
> > +       if (ret < 0)
> > +               goto out;
> > +
> > +       WRITE_ONCE(prog->filter, new_prog);
> > +
> > +       bpf_prog_inc(new_prog);
>
> and you don't need this, you already got the reference from link_update()
>

So the reason I still keep an extra refcount is because the existing code on the
netlink side assumes that. Even though the link itself holds a refcount for us,
the actual freeing of cls_bpf_prog may happen independent of bpf_link.

I'll add a comment for this.

> > +       /* release our reference */
> > +       bpf_prog_put(old_prog);
> > +
> > +out:
> > +       rtnl_unlock();
> > +       return ret;
> > +}
> > +
> >  static void __bpf_fill_link_info(struct cls_bpf_link *link,
> >                                  struct bpf_link_info *info)
> >  {
> > @@ -859,6 +909,7 @@ static const struct bpf_link_ops cls_bpf_link_ops = {
> >         .show_fdinfo = cls_bpf_link_show_fdinfo,
> >  #endif
> >         .fill_link_info = cls_bpf_link_fill_link_info,
> > +       .update_prog = cls_bpf_link_update,
> >  };
> >
> >  static inline char *cls_bpf_link_name(u32 prog_id, const char *name)
> > --
> > 2.31.1
> >

--
Kartikeya