On Tue, Apr 12, 2022 at 1:19 PM Martin KaFai Lau <kafai@xxxxxx> wrote:
>
> On Tue, Apr 12, 2022 at 12:01:41PM -0700, Stanislav Fomichev wrote:
> > On Tue, Apr 12, 2022 at 11:13 AM Martin KaFai Lau <kafai@xxxxxx> wrote:
> > >
> > > On Tue, Apr 12, 2022 at 09:42:40AM -0700, Stanislav Fomichev wrote:
> > > > On Mon, Apr 11, 2022 at 6:36 PM Martin KaFai Lau <kafai@xxxxxx> wrote:
> > > > >
> > > > > On Mon, Apr 11, 2022 at 11:46:20AM -0700, Stanislav Fomichev wrote:
> > > > > > On Fri, Apr 8, 2022 at 3:57 PM Martin KaFai Lau <kafai@xxxxxx> wrote:
> > > > > > >
> > > > > > > On Thu, Apr 07, 2022 at 03:31:08PM -0700, Stanislav Fomichev wrote:
> > > > > > > > Previous patch adds 1:1 mapping between all 211 LSM hooks
> > > > > > > > and bpf_cgroup program array. Instead of reserving a slot per
> > > > > > > > possible hook, reserve 10 slots per cgroup for lsm programs.
> > > > > > > > Those slots are dynamically allocated on demand and reclaimed.
> > > > > > > > This still adds some bloat to the cgroup and brings us back to
> > > > > > > > roughly pre-cgroup_bpf_attach_type times.
> > > > > > > >
> > > > > > > > It should be possible to eventually extend this idea to all hooks if
> > > > > > > > the memory consumption is unacceptable and shrink overall effective
> > > > > > > > programs array.
> > > > > > > >
> > > > > > > > Signed-off-by: Stanislav Fomichev <sdf@xxxxxxxxxx>
> > > > > > > > ---
> > > > > > > > include/linux/bpf-cgroup-defs.h | 4 +-
> > > > > > > > include/linux/bpf_lsm.h | 6 ---
> > > > > > > > kernel/bpf/bpf_lsm.c | 9 ++--
> > > > > > > > kernel/bpf/cgroup.c | 96 ++++++++++++++++++++++++++++-----
> > > > > > > > 4 files changed, 90 insertions(+), 25 deletions(-)
> > > > > > > >
> > > > > > > > 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 ?
> > > > > > >
> > > > > > > > enum cgroup_bpf_attach_type {
> > > > > > > > CGROUP_BPF_ATTACH_TYPE_INVALID = -1,
> > > > > > > > diff --git a/include/linux/bpf_lsm.h b/include/linux/bpf_lsm.h
> > > > > > > > index 7f0e59f5f9be..613de44aa429 100644
> > > > > > > > --- a/include/linux/bpf_lsm.h
> > > > > > > > +++ b/include/linux/bpf_lsm.h
> > > > > > > > @@ -43,7 +43,6 @@ extern const struct bpf_func_proto bpf_inode_storage_delete_proto;
> > > > > > > > void bpf_inode_storage_free(struct inode *inode);
> > > > > > > >
> > > > > > > > int bpf_lsm_find_cgroup_shim(const struct bpf_prog *prog, bpf_func_t *bpf_func);
> > > > > > > > -int bpf_lsm_hook_idx(u32 btf_id);
> > > > > > > >
> > > > > > > > #else /* !CONFIG_BPF_LSM */
> > > > > > > >
> > > > > > > > @@ -74,11 +73,6 @@ static inline int bpf_lsm_find_cgroup_shim(const struct bpf_prog *prog,
> > > > > > > > return -ENOENT;
> > > > > > > > }
> > > > > > > >
> > > > > > > > -static inline int bpf_lsm_hook_idx(u32 btf_id)
> > > > > > > > -{
> > > > > > > > - return -EINVAL;
> > > > > > > > -}
> > > > > > > > -
> > > > > > > > #endif /* CONFIG_BPF_LSM */
> > > > > > > >
> > > > > > > > #endif /* _LINUX_BPF_LSM_H */
> > > > > > > > diff --git a/kernel/bpf/bpf_lsm.c b/kernel/bpf/bpf_lsm.c
> > > > > > > > index eca258ba71d8..8b948ec9ab73 100644
> > > > > > > > --- a/kernel/bpf/bpf_lsm.c
> > > > > > > > +++ b/kernel/bpf/bpf_lsm.c
> > > > > > > > @@ -57,10 +57,12 @@ static unsigned int __cgroup_bpf_run_lsm_socket(const void *ctx,
> > > > > > > > if (unlikely(!sk))
> > > > > > > > return 0;
> > > > > > > >
> > > > > > > > + rcu_read_lock(); /* See bpf_lsm_attach_type_get(). */
> > > > > > > > cgrp = sock_cgroup_ptr(&sk->sk_cgrp_data);
> > > > > > > > if (likely(cgrp))
> > > > > > > > ret = BPF_PROG_RUN_ARRAY_CG(cgrp->bpf.effective[prog->aux->cgroup_atype],
> > > > > > > > ctx, bpf_prog_run, 0);
> > > > > > > > + rcu_read_unlock();
> > > > > > > > return ret;
> > > > > > > > }
> > > > > > > >
> > > > > > > > @@ -77,7 +79,7 @@ static unsigned int __cgroup_bpf_run_lsm_current(const void *ctx,
> > > > > > > > /*prog = container_of(insn, struct bpf_prog, insnsi);*/
> > > > > > > > prog = (const struct bpf_prog *)((void *)insn - offsetof(struct bpf_prog, insnsi));
> > > > > > > >
> > > > > > > > - rcu_read_lock();
> > > > > > > > + rcu_read_lock(); /* See bpf_lsm_attach_type_get(). */
> > > > > > > I think this is also needed for task_dfl_cgroup(). If yes,
> > > > > > > will be a good idea to adjust the comment if it ends up
> > > > > > > using the 'CGROUP_LSM_NUM 10' scheme.
> > > > > > >
> > > > > > > While at rcu_read_lock(), have you thought about what major things are
> > > > > > > needed to make BPF_LSM_CGROUP sleepable ?
> > > > > > >
> > > > > > > The cgroup local storage could be one that require changes but it seems
> > > > > > > the cgroup local storage is not available to BPF_LSM_GROUP in this change set.
> > > > > > > The current use case doesn't need it?
> > > > > >
> > > > > > No, I haven't thought about sleepable at all yet :-( But seems like
> > > > > > having that rcu lock here might be problematic if we want to sleep? In
> > > > > > this case, Jakub's suggestion seems better.
> > > > > The new rcu_read_lock() here seems fine after some thoughts.
> > > > >
> > > > > I was looking at the helpers in cgroup_base_func_proto() to get a sense
> > > > > on sleepable support. Only the bpf_get_local_storage caught my eyes for
> > > > > now because it uses a call_rcu to free the storage. That will be the
> > > > > major one to change for sleepable that I can think of for now.
> > > >
> > > > That rcu_read_lock should be switched over to rcu_read_lock_trace in
> > > > the sleepable case I'm assuming? Are we allowed to sleep while holding
> > > > rcu_read_lock_trace?
> > > Ah. right, suddenly forgot the obvious in between emails :(
> > >
> > > In that sense, may as well remove the rcu_read_lock() here and let
> > > the trampoline to decide which one (rcu_read_lock or rcu_read_lock_trace)
> > > to call before calling the shim_prog. The __bpf_prog_enter(_sleepable) will
> > > call the right rcu_read_lock(_trace) based on the prog is sleepable or not.
> >
> > Removing rcu_read_lock in __cgroup_bpf_run_lsm_current might be
> > problematic because we also want to guarantee current's cgroup doesn't
> > go away. I'm assuming things like task migrating to a new cgroup and
> > the old one being freed can happen while we are trying to get cgroup's
> > effective array.
> Right, sleepable one may need a short rcu_read_lock only upto
> a point that the cgrp->bpf.effective[...] is obtained.
> call_rcu_tasks_trace() is then needed to free the bpf_prog_array.
>
> The future sleepable one may be better off to have a different shim func,
> not sure. rcu_read_lock() can be added back later if it ends up reusing
> the same shim func is cleaner.
In this case I'll probably have rcu_read_lock for
cgroup+bpf_lsm_attach_type_get for the current shim.
> > I guess BPF_PROG_RUN_ARRAY_CG will also need some work before
> > sleepable can happen (it calls rcu_read_lock unconditionally).
> Yep. I think so.
>
> >
> > Also, it doesn't seem like BPF_PROG_RUN_ARRAY_CG rcu usage is correct.
> > It receives __rcu array_rcu, takes rcu read lock and does deref. I'm
> > assuming that array_rcu can be free'd before we even get to
> > BPF_PROG_RUN_ARRAY_CG's rcu_read_lock? (so having rcu_read_lock around
> > BPF_PROG_RUN_ARRAY_CG makes sense)
> BPF_PROG_RUN_ARRAY_CG is __always_inline though.
Does it help? This should still expand to the following, right?
array_rcu = cgrp->bpf.effective[atype];
/* theoretically, array_rcu can be freed here? */
rcu_read_lock();
array = rcu_dereference(array_rcu);
...
Feels like the callers of BPF_PROG_RUN_ARRAY_CG really have to care
about rcu locking, not the BPF_PROG_RUN_ARRAY_CG itself.
