Re: [PATCH bpf-next v2 4/6] bpf: Introduce css_descendant open-coded iterator kfuncs

[Andrii Nakryiko]

On Fri, Sep 15, 2023 at 4:57 AM Chuyi Zhou <zhouchuyi@xxxxxxxxxxxxx> wrote:
>
> Hello.
>
> 在 2023/9/15 07:26, Andrii Nakryiko 写道:
> > On Tue, Sep 12, 2023 at 12:02 AM Chuyi Zhou <zhouchuyi@xxxxxxxxxxxxx> wrote:
> >>
> >> This Patch adds kfuncs bpf_iter_css_{pre,post}_{new,next,destroy} which
> >> allow creation and manipulation of struct bpf_iter_css in open-coded
> >> iterator style. These kfuncs actually wrapps css_next_descendant_{pre,
> >> post}. BPF programs can use these kfuncs through bpf_for_each macro for
> >> iteration of all descendant css under a root css.
> >>
> >> Signed-off-by: Chuyi Zhou <zhouchuyi@xxxxxxxxxxxxx>
> >> ---
> >>   include/uapi/linux/bpf.h       |  8 +++++
> >>   kernel/bpf/helpers.c           |  6 ++++
> >>   kernel/bpf/task_iter.c         | 53 ++++++++++++++++++++++++++++++++++
> >>   tools/include/uapi/linux/bpf.h |  8 +++++
> >>   tools/lib/bpf/bpf_helpers.h    | 12 ++++++++
> >>   5 files changed, 87 insertions(+)
> >>
> >> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> >> index befa55b52e29..57760afc13d0 100644
> >> --- a/include/uapi/linux/bpf.h
> >> +++ b/include/uapi/linux/bpf.h
> >> @@ -7326,4 +7326,12 @@ struct bpf_iter_process {
> >>          __u64 __opaque[1];
> >>   } __attribute__((aligned(8)));
> >>
> >> +struct bpf_iter_css_pre {
> >> +       __u64 __opaque[2];
> >> +} __attribute__((aligned(8)));
> >> +
> >> +struct bpf_iter_css_post {
> >> +       __u64 __opaque[2];
> >> +} __attribute__((aligned(8)));
> >> +
> >>   #endif /* _UAPI__LINUX_BPF_H__ */
> >> diff --git a/kernel/bpf/helpers.c b/kernel/bpf/helpers.c
> >> index 9b7d2c6f99d1..ca1f6404af9e 100644
> >> --- a/kernel/bpf/helpers.c
> >> +++ b/kernel/bpf/helpers.c
> >> @@ -2510,6 +2510,12 @@ BTF_ID_FLAGS(func, bpf_iter_css_task_destroy, KF_ITER_DESTROY)
> >>   BTF_ID_FLAGS(func, bpf_iter_process_new, KF_ITER_NEW)
> >>   BTF_ID_FLAGS(func, bpf_iter_process_next, KF_ITER_NEXT | KF_RET_NULL)
> >>   BTF_ID_FLAGS(func, bpf_iter_process_destroy, KF_ITER_DESTROY)
> >> +BTF_ID_FLAGS(func, bpf_iter_css_pre_new, KF_ITER_NEW)
> >> +BTF_ID_FLAGS(func, bpf_iter_css_pre_next, KF_ITER_NEXT | KF_RET_NULL)
> >> +BTF_ID_FLAGS(func, bpf_iter_css_pre_destroy, KF_ITER_DESTROY)
> >> +BTF_ID_FLAGS(func, bpf_iter_css_post_new, KF_ITER_NEW)
> >> +BTF_ID_FLAGS(func, bpf_iter_css_post_next, KF_ITER_NEXT | KF_RET_NULL)
> >> +BTF_ID_FLAGS(func, bpf_iter_css_post_destroy, KF_ITER_DESTROY)
> >>   BTF_ID_FLAGS(func, bpf_dynptr_adjust)
> >>   BTF_ID_FLAGS(func, bpf_dynptr_is_null)
> >>   BTF_ID_FLAGS(func, bpf_dynptr_is_rdonly)
> >> diff --git a/kernel/bpf/task_iter.c b/kernel/bpf/task_iter.c
> >> index 9d1927dc3a06..8963fc779b87 100644
> >> --- a/kernel/bpf/task_iter.c
> >> +++ b/kernel/bpf/task_iter.c
> >> @@ -880,6 +880,59 @@ __bpf_kfunc void bpf_iter_process_destroy(struct bpf_iter_process *it)
> >>   {
> >>   }
> >>
> >> +struct bpf_iter_css_kern {
> >> +       struct cgroup_subsys_state *root;
> >> +       struct cgroup_subsys_state *pos;
> >> +} __attribute__((aligned(8)));
> >> +
> >> +__bpf_kfunc int bpf_iter_css_pre_new(struct bpf_iter_css_pre *it,
> >> +               struct cgroup_subsys_state *root)
> >
> > similar to my comment on previous patches, please see
> > kernel/bpf/cgroup_iter.c for iter/cgroup iterator program. Let's stay
> > consistent. We have one iterator that accepts parameters defining
> > iteration order and starting cgroup. Unless there are some technical
> > reasons we can't follow similar approach with this open-coded iter,
> > let's use the same approach. We can even reuse
> > BPF_CGROUP_ITER_DESCENDANTS_PRE, BPF_CGROUP_ITER_DESCENDANTS_POST,
> > BPF_CGROUP_ITER_ANCESTORS_UP enums.
> >
>
> I know your concern. It would be nice if we keep consistent with
> kernel/bpf/cgroup_iter.c
>
> But this patch actually want to support iterating css
> (cgroup_subsys_state) not cgroup (css is more low lever).
> With css_iter we can do something like
> "for_each_mem_cgroup_tree/cpuset_for_each_descendant_pre"
> in BPF Progs which is hard for cgroup_iter. In the future we can use
> this iterator to plug some customizable policy in other resource control
> system.

That's fine if it's not exactly cgroup iter and returns a different
kernel object. But let's at least consistently use
BPF_CGROUP_ITER_DESCENDANTS_PRE/BPF_CGROUP_ITER_DESCENDANTS_POST/BPF_CGROUP_ITER_ANCESTORS_UP
approach as a way to specify iteration order?

>
> BTW, what I did in RFC actually very similar with the approach of
> cgroup_iter.
> (https://lore.kernel.org/all/20230827072057.1591929-4-zhouchuyi@xxxxxxxxxxxxx/).
>
> Thanks.