Re: [PATCH bpf-next v3 3/4] bpf: cgroup: properly use bpf_prog_array api

On 05/28, Roman Gushchin wrote:

On Tue, May 28, 2019 at 11:29:45AM -0700, Stanislav Fomichev wrote:

quoted

Now that we don't have __rcu markers on the bpf_prog_array helpers,
let's use proper rcu_dereference_protected to obtain array pointer
under mutex.

We also don't need __rcu annotations on cgroup_bpf.inactive since
it's not read/updated concurrently.

v3:
* amend cgroup_rcu_dereference to include percpu_ref_is_dying;
  cgroup_bpf is now reference counted and we don't hold cgroup_mutex
  anymore in cgroup_bpf_release

v2:
* replace xchg with rcu_swap_protected

Cc: Roman Gushchin <redacted>
Signed-off-by: Stanislav Fomichev <redacted>
---
 include/linux/bpf-cgroup.h |  2 +-
 kernel/bpf/cgroup.c        | 32 +++++++++++++++++++++-----------
 2 files changed, 22 insertions(+), 12 deletions(-)

diff --git a/include/linux/bpf-cgroup.h b/include/linux/bpf-cgroup.h
index 9f100fc422c3..b631ee75762d 100644
--- a/include/linux/bpf-cgroup.h
+++ b/include/linux/bpf-cgroup.h

@@ -72,7 +72,7 @@ struct cgroup_bpf {
 	u32 flags[MAX_BPF_ATTACH_TYPE];
 
 	/* temp storage for effective prog array used by prog_attach/detach */
-	struct bpf_prog_array __rcu *inactive;
+	struct bpf_prog_array *inactive;
 
 	/* reference counter used to detach bpf programs after cgroup removal */
 	struct percpu_ref refcnt;

diff --git a/kernel/bpf/cgroup.c b/kernel/bpf/cgroup.c
index d995edbe816d..118b70175dd9 100644
--- a/kernel/bpf/cgroup.c
+++ b/kernel/bpf/cgroup.c

@@ -22,6 +22,13 @@
 DEFINE_STATIC_KEY_FALSE(cgroup_bpf_enabled_key);
 EXPORT_SYMBOL(cgroup_bpf_enabled_key);
 
+#define cgroup_rcu_dereference(cgrp, p)					\
+	rcu_dereference_protected(p, lockdep_is_held(&cgroup_mutex) ||	\
+				  percpu_ref_is_dying(&cgrp->bpf.refcnt))

Some comments why percpu_ref_is_dying(&cgrp->bpf.refcnt) is enough here will
be appreciated.

I was actually debating whether to just use raw
rcu_dereference_protected(p, lockdep_is_held()) in __cgroup_bpf_query and
rcu_dereference_protected(p, percpu_ref_is_dying()) in cgroup_bpf_release
instead of having a cgroup_rcu_dereference which covers both cases.

Maybe that should make it more clear (and doesn't require any comment)?