Re: [PATCH bpf-next v2 2/3] bpf, netns: Keep attached programs in bpf_prog_array

[Jakub Sitnicki <jakub@cloudflare.com>]

On Wed, Jun 24, 2020 at 07:47 PM CEST, Andrii Nakryiko wrote:
> On Wed, Jun 24, 2020 at 10:19 AM Jakub Sitnicki <jakub@cloudflare.com> wrote:
>>
>> On Tue, Jun 23, 2020 at 12:34 PM CEST, Jakub Sitnicki wrote:
>> > Prepare for having multi-prog attachments for new netns attach types by
>> > storing programs to run in a bpf_prog_array, which is well suited for
>> > iterating over programs and running them in sequence.
>> >
>> > Because bpf_prog_array is dynamically resized, after this change a
>> > potentially blocking memory allocation in bpf(PROG_QUERY) callback can
>> > happen, in order to collect program IDs before copying the values to
>> > user-space supplied buffer. This forces us to adapt how we protect access
>> > to the attached program in the callback. As bpf_prog_array_copy_to_user()
>> > helper can sleep, we switch from an RCU read lock to holding a mutex that
>> > serializes updaters.
>> >
>> > To handle bpf(PROG_ATTACH) scenario when we are replacing an already
>> > attached program, we introduce a new bpf_prog_array helper called
>> > bpf_prog_array_replace_item that will exchange the old program with a new
>> > one. bpf-cgroup does away with such helper by computing an index into the
>> > array from a program position in an external list of attached
>> > programs/links. Such approach fails when a dummy prog is left in the array
>> > after a memory allocation failure on link release, but is necessary in
>> > bpf-cgroup case because the same BPF program can be present in the array
>> > multiple times due to inheritance, and attachment cannot be reliably
>> > identified by bpf_prog pointer comparison.
>> >
>> > No functional changes intended.
>> >
>> > Acked-by: Andrii Nakryiko <andriin@fb.com>
>> > Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
>> > ---
>> >  include/linux/bpf.h        |   3 +
>> >  include/net/netns/bpf.h    |   5 +-
>> >  kernel/bpf/core.c          |  20 ++++--
>> >  kernel/bpf/net_namespace.c | 137 +++++++++++++++++++++++++++----------
>> >  net/core/flow_dissector.c  |  21 +++---
>> >  5 files changed, 132 insertions(+), 54 deletions(-)
>> >
>>
>> [...]
>>
>> > diff --git a/kernel/bpf/net_namespace.c b/kernel/bpf/net_namespace.c
>> > index b951dab2687f..593523a22168 100644
>> > --- a/kernel/bpf/net_namespace.c
>> > +++ b/kernel/bpf/net_namespace.c
>>
>> [...]
>>
>> > @@ -93,8 +108,16 @@ static int bpf_netns_link_update_prog(struct bpf_link *link,
>> >               goto out_unlock;
>> >       }
>> >
>> > +     run_array = rcu_dereference_protected(net->bpf.run_array[type],
>> > +                                           lockdep_is_held(&netns_bpf_mutex));
>> > +     if (run_array)
>> > +             ret = bpf_prog_array_replace_item(run_array, link->prog, new_prog);
>>
>> Thinking about this some more, link update should fail with -EINVAL if
>> new_prog already exists in run_array. Same as PROG_ATTACH fails with
>> -EINVAL when trying to attach the same prog for the second time.
>>
>> Otherwise, LINK_UPDATE can lead to having same BPF prog present multiple
>> times in the prog_array, once attaching more than one prog gets enabled.
>>
>> Then we would we end up with the same challenge as bpf-cgroup, that is
>> how to find the program index into the prog_array in presence of
>> dummy_prog's.
>
> If you attach 5 different links having the same bpf_prog, it should be
> allowed and all five bpf_progs should be attached and called 5 times.
> They are independent links, that's the main thing. What specific BPF
> program is attached by the link (or later updated to) shouldn't be of
> any concern here (relative to other attached links/programs).
>
> Attaching the same *link* twice shouldn't be allowed, though.

Thanks for clarifying. I need to change the approach then:

 1) find the prog index into prog_array by iterating the list of links,
 2) adjust the index for any dummy progs in prog_array below the index.

That might work for bpf-cgroup too.

The only other alternative I can think of it to copy the prog array to
filter out dummy_progs, before replacing the prog on link update.

>
>>
>> > +     else
>> > +             ret = -ENOENT;
>> > +     if (ret)
>> > +             goto out_unlock;
>> > +
>> >       old_prog = xchg(&link->prog, new_prog);
>> > -     rcu_assign_pointer(net->bpf.progs[type], new_prog);
>> >       bpf_prog_put(old_prog);
>> >
>> >  out_unlock:
>>
>> [...]
>>
>> > @@ -217,14 +249,25 @@ int netns_bpf_prog_attach(const union bpf_attr *attr, struct bpf_prog *prog)
>> >       if (ret)
>> >               goto out_unlock;
>> >
>> > -     attached = rcu_dereference_protected(net->bpf.progs[type],
>> > -                                          lockdep_is_held(&netns_bpf_mutex));
>> > +     attached = net->bpf.progs[type];
>> >       if (attached == prog) {
>> >               /* The same program cannot be attached twice */
>> >               ret = -EINVAL;
>> >               goto out_unlock;
>> >       }
>> > -     rcu_assign_pointer(net->bpf.progs[type], prog);
>> > +
>> > +     run_array = rcu_dereference_protected(net->bpf.run_array[type],
>> > +                                           lockdep_is_held(&netns_bpf_mutex));
>> > +     if (run_array) {
>> > +             ret = bpf_prog_array_replace_item(run_array, attached, prog);
>>
>> I didn't consider here that there can be a run_array with a dummy_prog
>> from a link release that failed to allocate memory.
>>
>> In such case bpf_prog_array_replace_item will fail, while we actually
>> want to replace the dummy_prog.
>>
>> The right thing to do is to replace the first item in prog array:
>>
>>         if (run_array) {
>>                 WRITE_ONCE(run_array->items[0].prog, prog);
>>         } else {
>>                 /* allocate a bpf_prog_array */
>>         }
>>
>> This leaves just one user of bpf_prog_array_replace_item(), so I think
>> I'm just going to fold it into its only caller, that is the update_prog
>> callback.
>
> That will change relative order of BPF programs, which I think is bad.
> So I agree that bpf_prog_array_replace_item is not all that useful and
> probably should be dropped. And the right way is to know the position
> of bpf_prog you are trying to replace/delete, just like cgroup case.
> Except cgroup case is even more complicated due to inheritance and
> hierarchy, which luckily you don't have to deal with here.

I think we are on the same page.

The write to first item could change the relative prog order, and even
replace the wrong program, but only if we had to deal with a prog_array
larger > 1.

In this case, direct attachment with PROG_ATTACH to netns, the
prog_array size is always 1, if it has been allocated already.

That is because I did not plan to enable multi-prog attachment for
flow_dissector with links. And only flow_dissector uses PROG_ATTACH to
netns.

>
>>
>> > +     } else {
>> > +             ret = bpf_prog_array_copy(NULL, NULL, prog, &run_array);
>> > +             rcu_assign_pointer(net->bpf.run_array[type], run_array);
>> > +     }
>> > +     if (ret)
>> > +             goto out_unlock;
>> > +
>> > +     net->bpf.progs[type] = prog;
>> >       if (attached)
>> >               bpf_prog_put(attached);
>> >
>>
>> [...]