Re: [PATCH bpf-next v2 05/10] bpf: lsm: BTF API for LSM hooks

Thanks for reviewing!

On 16-Jan 16:28, Andrii Nakryiko wrote:

On Wed, Jan 15, 2020 at 9:14 AM KP Singh [off-list ref] wrote:

quoted

From: KP Singh <redacted>

The BTF API provides information required by the BPF verifier to
attach eBPF programs to the LSM hooks by using the BTF information of
two types:

- struct security_hook_heads: This type provides the offset which
  a new dynamically allocated security hook must be attached to.
- union security_list_options: This provides the information about the
  function prototype required by the hook.

When the program is loaded:

- The verifier receives the index of a member in struct
  security_hook_heads to which a program must be attached as
  prog->aux->lsm_hook_index. The index is one-based for better
  verification.
- bpf_lsm_type_by_index is used to determine the func_proto of
  the LSM hook and updates prog->aux->attach_func_proto
- bpf_lsm_head_by_index is used to determine the hlist_head to which
  the BPF program must be attached.

Signed-off-by: KP Singh <redacted>
---
 include/linux/bpf_lsm.h |  12 +++++
 security/bpf/Kconfig    |   1 +
 security/bpf/hooks.c    | 104 ++++++++++++++++++++++++++++++++++++++++
 3 files changed, 117 insertions(+)

diff --git a/include/linux/bpf_lsm.h b/include/linux/bpf_lsm.h
index 9883cf25241c..a9b4f7b41c65 100644
--- a/include/linux/bpf_lsm.h
+++ b/include/linux/bpf_lsm.h

@@ -19,6 +19,8 @@ extern struct security_hook_heads bpf_lsm_hook_heads;

 int bpf_lsm_srcu_read_lock(void);
 void bpf_lsm_srcu_read_unlock(int idx);
+const struct btf_type *bpf_lsm_type_by_index(struct btf *btf, u32 offset);
+const struct btf_member *bpf_lsm_head_by_index(struct btf *btf, u32 id);

 #define CALL_BPF_LSM_VOID_HOOKS(FUNC, ...)                     \
        do {                                                    \

@@ -65,6 +67,16 @@ static inline int bpf_lsm_srcu_read_lock(void)
        return 0;
 }
 static inline void bpf_lsm_srcu_read_unlock(int idx) {}
+static inline const struct btf_type *bpf_lsm_type_by_index(
+       struct btf *btf, u32 index)
+{
+       return ERR_PTR(-EOPNOTSUPP);
+}
+static inline const struct btf_member *bpf_lsm_head_by_index(
+       struct btf *btf, u32 id)
+{
+       return ERR_PTR(-EOPNOTSUPP);
+}

 #endif /* CONFIG_SECURITY_BPF */

diff --git a/security/bpf/Kconfig b/security/bpf/Kconfig
index 595e4ad597ae..9438d899b618 100644
--- a/security/bpf/Kconfig
+++ b/security/bpf/Kconfig

@@ -7,6 +7,7 @@ config SECURITY_BPF
        depends on SECURITY
        depends on BPF_SYSCALL
        depends on SRCU
+       depends on DEBUG_INFO_BTF
        help
          This enables instrumentation of the security hooks with
          eBPF programs.

diff --git a/security/bpf/hooks.c b/security/bpf/hooks.c
index b123d9cb4cd4..82725611693d 100644
--- a/security/bpf/hooks.c
+++ b/security/bpf/hooks.c

@@ -5,6 +5,8 @@
  */

 #include <linux/bpf_lsm.h>
+#include <linux/bpf.h>
+#include <linux/btf.h>
 #include <linux/srcu.h>

 DEFINE_STATIC_SRCU(security_hook_srcu);

@@ -18,3 +20,105 @@ void bpf_lsm_srcu_read_unlock(int idx)
 {
        return srcu_read_unlock(&security_hook_srcu, idx);
 }
+
+static inline int validate_hlist_head(struct btf *btf, u32 type_id)
+{
+       s32 hlist_id;
+
+       hlist_id = btf_find_by_name_kind(btf, "hlist_head", BTF_KIND_STRUCT);
+       if (hlist_id < 0 || hlist_id != type_id)
+               return -EINVAL;

This feels backwards and expensive. You already have type_id you want
to check. Do a quick look up, check type and other attributes, if you
want. There is no need to do linear search for struct named
"hlist_head".

But in reality, you should trust kernel BTF, you already know that you
found correct "security_hook_heads" struct, so its member has to be
hlist_head, no?

We had a discussion internally and also came the same conclusion (it's
okay to trust the BTF) and will remove sone of the "over-cautious"
checks in the next revision.

This one, however, in particular is to protect against the case when a
new member which is not a hlist_head is added to security_hook_heads
and the user-space tries to attach at that index.

I admit that the likelyhood of that happening is very little  but I
think it's worth checking. I do, like your idea and will update the
code to use the type_id and do a quick check rather than a linear
search to look for the type_id.

This is what remains of all the pedantic checks pertaining to
hlist_head:

	t = btf_type_by_id(btf, member->type);
	if (unlikely(!t))
		return -EINVAL;

	if (BTF_INFO_KIND(t->info) != BTF_KIND_STRUCT)
		return -EINVAL;

	if (t->size != sizeof(struct hlist_head))
		return -EINVAL;

quoted

+
+       return 0;
+}
+
+/* Find the BTF representation of the security_hook_heads member for a member
+ * with a given index in struct security_hook_heads.
+ */
+const struct btf_member *bpf_lsm_head_by_index(struct btf *btf, u32 index)
+{
+       const struct btf_member *member;
+       const struct btf_type *t;
+       u32 off, i;
+       int ret;
+
+       t = btf_type_by_name_kind(btf, "security_hook_heads", BTF_KIND_STRUCT);
+       if (WARN_ON_ONCE(IS_ERR(t)))
+               return ERR_CAST(t);
+
+       for_each_member(i, t, member) {
+               /* We've found the id requested and need to check the

quoted

+                * the following:
+                *
+                * - Is it at a valid alignment for struct hlist_head?
+                *
+                * - Is it a valid hlist_head struct?
+                */
+               if (index == i) {

Also not efficient. Check index to be < vlen(t), then member =
btf_type_member(t) + index;

Neat! Updated.

quoted

+                       off = btf_member_bit_offset(t, member);
+                       if (off % 8)
+                               /* valid c code cannot generate such btf */
+                               return ERR_PTR(-EINVAL);
+                       off /= 8;
+
+                       if (off % __alignof__(struct hlist_head))
+                               return ERR_PTR(-EINVAL);
+
+                       ret = validate_hlist_head(btf, member->type);
+                       if (ret < 0)
+                               return ERR_PTR(ret);
+
+                       return member;

This feels a bit over-cautious to double-check this. If
security_hook_heads definition is controlled by kernel sources, then
we could just trust vmlinux BTF?

Yep, makes sense. Removed some of these checks.

quoted

+               }
+       }
+
+       return ERR_PTR(-ENOENT);
+}
+
+/* Given an index of a member in security_hook_heads return the
+ * corresponding type for the LSM hook. The members of the union
+ * security_list_options have the same name as the security_hook_heads which
+ * is ensured by the LSM_HOOK_INIT macro defined in include/linux/lsm_hooks.h
+ */
+const struct btf_type *bpf_lsm_type_by_index(struct btf *btf, u32 index)
+{
+       const struct btf_member *member, *hook_head = NULL;
+       const struct btf_type *t, *hook_type = NULL;
+       u32 i;
+
+       hook_head = bpf_lsm_head_by_index(btf, index);
+       if (IS_ERR(hook_head))
+               return ERR_PTR(PTR_ERR(hook_head));
+
+       t = btf_type_by_name_kind(btf, "security_list_options", BTF_KIND_UNION);
+       if (WARN_ON_ONCE(IS_ERR(t)))
+               return ERR_CAST(t);

btf_type_by_name_kind() is a linear search (at least right now), so it
might be a good idea to cache found type_id's of security_list_options
and security_hook_heads?

I am already caching these types in the next patch (struct
bpf_lsm_info) of the series which implements attachment. I moved it to
this patch so that it's clearer.

quoted

+
+       for_each_member(i, t, member) {
+               if (hook_head->name_off == member->name_off) {
+                       /* There should be only one member with the same name
+                        * as the LSM hook. This should never really happen
+                        * and either indicates malformed BTF or someone trying
+                        * trick the LSM.
+                        */
+                       if (WARN_ON(hook_type))
+                               return ERR_PTR(-EINVAL);
+
+                       hook_type = btf_type_by_id(btf, member->type);
+                       if (unlikely(!hook_type))
+                               return ERR_PTR(-EINVAL);
+
+                       if (!btf_type_is_ptr(hook_type))
+                               return ERR_PTR(-EINVAL);
+               }
+       }
+
+       if (!hook_type)
+               return ERR_PTR(-ENOENT);
+
+       t = btf_type_by_id(btf, hook_type->type);
+       if (unlikely(!t))
+               return ERR_PTR(-EINVAL);

why not do this inside the loop when you find correct member and not
continue processing all the fields?

Updated.

- KP

quoted

+
+       return t;
+}
--
2.20.1