On Mon, Apr 11, 2022 at 6:19 PM Martin KaFai Lau <kafai@xxxxxx> wrote:
>
> On Sat, Apr 09, 2022 at 07:04:05PM +0200, Jakub Sitnicki wrote:
> > >> diff --git a/include/linux/bpf-cgroup-defs.h b/include/linux/bpf-cgroup-defs.h
> > >> index 6c661b4df9fa..d42516e86b3a 100644
> > >> --- a/include/linux/bpf-cgroup-defs.h
> > >> +++ b/include/linux/bpf-cgroup-defs.h
> > >> @@ -10,7 +10,9 @@
> > >>
> > >> struct bpf_prog_array;
> > >>
> > >> -#define CGROUP_LSM_NUM 211 /* will be addressed in the next patch */
> > >> +/* Maximum number of concurrently attachable per-cgroup LSM hooks.
> > >> + */
> > >> +#define CGROUP_LSM_NUM 10
> > > hmm...only 10 different lsm hooks (or 10 different attach_btf_ids) can
> > > have BPF_LSM_CGROUP programs attached. This feels quite limited but having
> > > a static 211 (and potentially growing in the future) is not good either.
> > > I currently do not have a better idea also. :/
> > >
> > > Have you thought about other dynamic schemes or they would be too slow ?
> >
> > As long as we're talking ideas - how about a 2-level lookup?
> >
> > L1: 0..255 -> { 0..31, -1 }, where -1 is inactive cgroup_bp_attach_type
> > L2: 0..31 -> struct bpf_prog_array * for cgroup->bpf.effective[],
> > struct hlist_head [^1] for cgroup->bpf.progs[],
> > u32 for cgroup->bpf.flags[],
> >
> > This way we could have 32 distinct _active_ attachment types for each
> > cgroup instance, to be shared among regular cgroup attach types and BPF
> > LSM attach types.
> >
> > It is 9 extra slots in comparison to today, so if anyone has cgroups
> > that make use of all available attach types at the same time, we don't
> > break their setup.
> >
> > The L1 lookup table would still a few slots for new cgroup [^2] or LSM
> > hooks:
> >
> > 256 - 23 (cgroup attach types) - 211 (LSM hooks) = 22
> >
> > Memory bloat:
> >
> > +256 B - L1 lookup table
> Does L1 need to be per cgroup ?
>
> or different cgroups usually have a very different active(/effective) set ?
I'm assuming the suggestion is to have it per cgroup. Otherwise, if it's
global, it's close to whatever I'm proposing in the original patch. As I
mentioned in the commit message, in theory, all cgroup_bpf can be managed
the way I propose to manage 10 lsm slots if we get to the point where
10 slots is not enough.
I've played with this mode a bit and it looks a bit complicated :-( Since it's
per cgroup, we have to be careful to preserve this mapping during
cgroup_bpf_inherit.
I'll see what I can do, but I'll most likely revert to my initial
version for now (I'll also include list_head->hlist_head conversion
patch, very nice idea).
> > + 72 B - extra effective[] slots
> > + 72 B - extra progs[] slots
> > + 36 B - extra flags[] slots
> > -184 B - savings from switching to hlist_head
> > ------
> > +252 B per cgroup instance
> >
> > Total cgroup_bpf{} size change - 720 B -> 968 B.
> >
> > WDYT?
> >
> > [^1] It looks like we can easily switch from cgroup->bpf.progs[] from
> > list_head to hlist_head and save some bytes!
> >
> > We only access the list tail in __cgroup_bpf_attach(). We can
> > either iterate over the list and eat the cost there or push the new
> > prog onto the front.
> >
> > I think we treat cgroup->bpf.progs[] everywhere like an unordered
> > set. Except for __cgroup_bpf_query, where the user might notice the
> > order change in the BPF_PROG_QUERY dump.
> >
> > [^2] Unrelated, but we would like to propose a
> > CGROUP_INET[46]_POST_CONNECT hook in the near future to make it
> > easier to bind UDP sockets to 4-tuple without creating conflicts:
> >
> > https://github.com/cloudflare/cloudflare-blog/tree/master/2022-02-connectx/ebpf_connect4
> >
> > [...]
