Re: [PATCH bpf-next v7 4/8] bpf: Introduce cgroup iter

From: Andrii Nakryiko
Date: Tue Aug 16 2022 - 03:36:04 EST


On Mon, Aug 15, 2022 at 9:12 PM Andrii Nakryiko
<andrii.nakryiko@xxxxxxxxx> wrote:
>
> On Thu, Aug 11, 2022 at 7:10 AM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote:
> >
> > On Wed, Aug 10, 2022 at 8:10 PM Hao Luo <haoluo@xxxxxxxxxx> wrote:
> > >
> > > On Tue, Aug 9, 2022 at 11:38 AM Hao Luo <haoluo@xxxxxxxxxx> wrote:
> > > >
> > > > On Tue, Aug 9, 2022 at 9:23 AM Alexei Starovoitov
> > > > <alexei.starovoitov@xxxxxxxxx> wrote:
> > > > >
> > > > > On Mon, Aug 08, 2022 at 05:56:57PM -0700, Hao Luo wrote:
> > > > > > On Mon, Aug 8, 2022 at 5:19 PM Andrii Nakryiko
> > > > > > <andrii.nakryiko@xxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Fri, Aug 5, 2022 at 2:49 PM Hao Luo <haoluo@xxxxxxxxxx> wrote:
> > > > > > > >
> > > > > > > > Cgroup_iter is a type of bpf_iter. It walks over cgroups in four modes:
> > > > > > > >
> > > > > > > > - walking a cgroup's descendants in pre-order.
> > > > > > > > - walking a cgroup's descendants in post-order.
> > > > > > > > - walking a cgroup's ancestors.
> > > > > > > > - process only the given cgroup.
> > > > > > > >
> > > > [...]
> > > > > > > > diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> > > > > > > > index 59a217ca2dfd..4d758b2e70d6 100644
> > > > > > > > --- a/include/uapi/linux/bpf.h
> > > > > > > > +++ b/include/uapi/linux/bpf.h
> > > > > > > > @@ -87,10 +87,37 @@ struct bpf_cgroup_storage_key {
> > > > > > > > __u32 attach_type; /* program attach type (enum bpf_attach_type) */
> > > > > > > > };
> > > > > > > >
> > > > > > > > +enum bpf_iter_order {
> > > > > > > > + BPF_ITER_ORDER_DEFAULT = 0, /* default order. */
> > > > > > >
> > > > > > > why is this default order necessary? It just adds confusion (I had to
> > > > > > > look up source code to know what is default order). I might have
> > > > > > > missed some discussion, so if there is some very good reason, then
> > > > > > > please document this in commit message. But I'd rather not do some
> > > > > > > magical default order instead. We can set 0 to mean invalid and error
> > > > > > > out, or just do SELF as the very first value (and if user forgot to
> > > > > > > specify more fancy mode, they hopefully will quickly discover this in
> > > > > > > their testing).
> > > > > > >
> > > > > >
> > > > > > PRE/POST/UP are tree-specific orders. SELF applies on all iters and
> > > > > > yields only a single object. How does task_iter express a non-self
> > > > > > order? By non-self, I mean something like "I don't care about the
> > > > > > order, just scan _all_ the objects". And this "don't care" order, IMO,
> > > > > > may be the common case. I don't think everyone cares about walking
> > > > > > order for tasks. The DEFAULT is intentionally put at the first value,
> > > > > > so that if users don't care about order, they don't have to specify
> > > > > > this field.
> > > > > >
> > > > > > If that sounds valid, maybe using "UNSPEC" instead of "DEFAULT" is better?
> > > > >
> > > > > I agree with Andrii.
> > > > > This:
> > > > > + if (order == BPF_ITER_ORDER_DEFAULT)
> > > > > + order = BPF_ITER_DESCENDANTS_PRE;
> > > > >
> > > > > looks like an arbitrary choice.
> > > > > imo
> > > > > BPF_ITER_DESCENDANTS_PRE = 0,
> > > > > would have been more obvious. No need to dig into definition of "default".
> > > > >
> > > > > UNSPEC = 0
> > > > > is fine too if we want user to always be conscious about the order
> > > > > and the kernel will error if that field is not initialized.
> > > > > That would be my preference, since it will match the rest of uapi/bpf.h
> > > > >
> > > >
> > > > Sounds good. In the next version, will use
> > > >
> > > > enum bpf_iter_order {
> > > > BPF_ITER_ORDER_UNSPEC = 0,
> > > > BPF_ITER_SELF_ONLY, /* process only a single object. */
> > > > BPF_ITER_DESCENDANTS_PRE, /* walk descendants in pre-order. */
> > > > BPF_ITER_DESCENDANTS_POST, /* walk descendants in post-order. */
> > > > BPF_ITER_ANCESTORS_UP, /* walk ancestors upward. */
> > > > };
> > > >
> > >
> > > Sigh, I find that having UNSPEC=0 and erroring out when seeing UNSPEC
> > > doesn't work. Basically, if we have a non-iter prog and a cgroup_iter
> > > prog written in the same source file, I can't use
> > > bpf_object__attach_skeleton to attach them. Because the default
> > > prog_attach_fn for iter initializes `order` to 0 (that is, UNSPEC),
> > > which is going to be rejected by the kernel. In order to make
> > > bpf_object__attach_skeleton work on cgroup_iter, I think I need to use
> > > the following
> > >
> > > enum bpf_iter_order {
>
> so first of all, this can't be called "bpf_iter_order" as it doesn't
> apply to BPF iterators in general. I think this should be called
> bpf_iter_cgroup_order (or maybe bpf_cgroup_iter_order) and if/when we
> add ability to iterate tasks within cgroups then we'll just reuse enum
> bpf_iter_cgroup_order as an extra parameter for task iterator.
>
> And with that future case in mind I do think that we should have 0
> being "UNSPEC" case.
>
> > > BPF_ITER_DESCENDANTS_PRE, /* walk descendants in pre-order. */
> > > BPF_ITER_DESCENDANTS_POST, /* walk descendants in post-order. */
> > > BPF_ITER_ANCESTORS_UP, /* walk ancestors upward. */
> > > BPF_ITER_SELF_ONLY, /* process only a single object. */
> > > };
> > >
> > > So that when calling bpf_object__attach_skeleton() on cgroup_iter, a
> > > link can be generated and the generated link defaults to pre-order
> > > walk on the whole hierarchy. Is there a better solution?
> > >
>
> I was actually surprised that we specify these additional parameters
> at attach (LINK_CREATE) time, and not at bpf_iter_create() call time.
> It seems more appropriate to allow to specify such runtime parameters
> very late, when we create a specific instance of seq_file. But I guess
> this was done because one of the initial motivations for iterators was
> to be pinned in BPFFS and read as a file, so it was more convenient to
> store such parameters upfront at link creation time to keep
> BPF_OBJ_PIN simpler. I guess it makes sense, worst case you'll need to
> create multiple bpf_link files, one for each cgroup hierarchy you'd
> like to query with the same single BPF program.
>
> But I digress.
>
> As for not being able to auto-attach cgroup iterator. I think that's
> sort of expected and is in line with not being able to auto-attach
> cgroup programs, as you need cgroup FD at runtime. So even if you had
> some reasonable default order, you still would need to specify target
> cgroup (either through FD or ID).
>
> So... either don't do skeleton auto-attach, or let's teach libbpf code
> to not auto-attach some iter types?
>
> Alternatively, we could teach libbpf to parse some sort of cgroup
> iterator spec, like:
>
> SEC("iter/cgroup:/path/to/cgroup:descendants_pre")
>
> But this approach won't work for a bunch of other parameterized
> iterators (e.g., task iter, or map elem iter), so I'm hesitant about
> adding this to libbpf as a generic functionality.

As yet another alternative (given you seem to default to root cgroup
when cgroup_fd or cgroup_id is not specified), we can teach libbpf to
just specify BPF_ITER_DESCENDANTS_PRE (or whichever mode seems like
best default) only for auto-attach case. If that makes most sense.

>
> >
> > I think this can be handled by userspace? We can attach the
> > cgroup_iter separately first (and maybe we will need to set prog->link
> > as well) so that bpf_object__attach_skeleton() doesn't try to attach
> > it? I am following this pattern in the selftest in the final patch,
> > although I think I might be missing setting prog->link, so I am
> > wondering why there are no issues in that selftest which has the same
> > scenario that you are talking about.
> >
> > I think such a pattern will need to be used anyway if the users need
> > to set any non-default arguments for the cgroup_iter prog (like the
> > selftest), right? The only case we are discussing here is the case
> > where the user wants to attach the cgroup_iter with all default
> > options (in which case the default order will fail).
> > I agree that it might be inconvenient if the default/uninitialized
> > options don't work for cgroup_iter, but Alexei pointed out that this
> > matches other bpf uapis.
> >
> > My concern is that in the future we try to reuse enum bpf_iter_order
> > to set ordering for other iterators, and then the
> > default/uninitialized value (BPF_ITER_DESCENDANTS_PRE) doesn't make
> > sense for that iterator (e.g. not a tree). In this case, the same
> > problem that we are avoiding for cgroup_iter here will show up for
> > that iterator, and we can't easily change it at this point because
> > it's uapi.
>
> Yep, valid concern, I agree.
>
> >
> >
> > > > and explicitly list the values acceptable by cgroup_iter, error out if
> > > > UNSPEC is detected.
> > > >
> > > > Also, following Andrii's comments, will change BPF_ITER_SELF to
> > > > BPF_ITER_SELF_ONLY, which does seem a little bit explicit in
> > > > comparison.
> > > >
> > > > > I applied the first 3 patches to ease respin.
> > > >
> > > > Thanks! This helps!
> > > >
> > > > > Thanks!