Re: [PATCH 08/17] net: convert sk_filter.refcnt from atomic_t to refcount_t

On 03/16/2017 04:28 PM, Elena Reshetova wrote:

quoted

refcount_t type and corresponding API should be
used instead of atomic_t when the variable is used as
a reference counter. This allows to avoid accidental
refcounter overflows that might lead to use-after-free
situations.

Signed-off-by: Elena Reshetova <elena.reshetova@intel.com>
Signed-off-by: Hans Liljestrand <redacted>
Signed-off-by: Kees Cook <redacted>
Signed-off-by: David Windsor <redacted>
---
  include/linux/filter.h | 3 ++-
  net/core/filter.c      | 7 ++++---
  2 files changed, 6 insertions(+), 4 deletions(-)

diff --git a/include/linux/filter.h b/include/linux/filter.h
index 8053c38..20247e7 100644
--- a/include/linux/filter.h
+++ b/include/linux/filter.h

@@ -7,6 +7,7 @@
  #include <stdarg.h>

  #include <linux/atomic.h>
+#include <linux/refcount.h>
  #include <linux/compat.h>
  #include <linux/skbuff.h>
  #include <linux/linkage.h>

@@ -431,7 +432,7 @@ struct bpf_prog {
  };

  struct sk_filter {
-	atomic_t	refcnt;
+	refcount_t	refcnt;
  	struct rcu_head	rcu;
  	struct bpf_prog	*prog;
  };

diff --git a/net/core/filter.c b/net/core/filter.c
index ebaeaf2..62267e2 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c

@@ -928,7 +928,7 @@ static void sk_filter_release_rcu(struct rcu_head *rcu)
   */
  static void sk_filter_release(struct sk_filter *fp)
  {
-	if (atomic_dec_and_test(&fp->refcnt))
+	if (refcount_dec_and_test(&fp->refcnt))
  		call_rcu(&fp->rcu, sk_filter_release_rcu);
  }

@@ -950,7 +950,7 @@ bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
  	/* same check as in sock_kmalloc() */
  	if (filter_size <= sysctl_optmem_max &&
  	    atomic_read(&sk->sk_omem_alloc) + filter_size <

sysctl_optmem_max) {

quoted

-		atomic_inc(&fp->refcnt);
+		refcount_inc(&fp->refcnt);
  		atomic_add(filter_size, &sk->sk_omem_alloc);
  		return true;
  	}

@@ -1179,12 +1179,13 @@ static int __sk_attach_prog(struct bpf_prog *prog,

struct sock *sk)

quoted

  		return -ENOMEM;

  	fp->prog = prog;
-	atomic_set(&fp->refcnt, 0);
+	refcount_set(&fp->refcnt, 1);

  	if (!sk_filter_charge(sk, fp)) {
  		kfree(fp);
  		return -ENOMEM;
  	}
+	refcount_set(&fp->refcnt, 1);

Regarding the two subsequent refcount_set(, 1) that look a bit strange
due to the sk_filter_charge() having refcount_inc() I presume ... can't
the refcount API handle such corner case? 

Yes, it was exactly because of recount_inc() from zero in sk_filter_charge(). 
refcount_inc() would refuse to do an inc from zero for security reasons. At some 
point in past we discussed refcount_inc_not_one() but it was decided to be too special case
to support (we really have very little of such cases).


Or alternatively the let the

sk_filter_charge() handle it, for example:

bool __sk_filter_charge(struct sock *sk, struct sk_filter *fp)
{
	u32 filter_size = bpf_prog_size(fp->prog->len);

	/* same check as in sock_kmalloc() */
	if (filter_size <= sysctl_optmem_max &&
	    atomic_read(&sk->sk_omem_alloc) + filter_size <
sysctl_optmem_max) {
		atomic_add(filter_size, &sk->sk_omem_alloc);
		return true;
	}
	return false;
}

And this goes to filter.h:

bool __sk_filter_charge(struct sock *sk, struct sk_filter *fp);

bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
{
	bool ret = __sk_filter_charge(sk, fp);
	if (ret)
		refcount_inc(&fp->refcnt);
	return ret;
}

... and let __sk_attach_prog() call __sk_filter_charge() and only fo
the second refcount_set()?

quoted

  	old_fp = rcu_dereference_protected(sk->sk_filter,

lockdep_sock_is_held(sk));

quoted

Oh, yes, this would make it look less awkward. Thank you for the suggestion Daniel! 
I guess we try to be less invasive for code changes overall, maybe even too careful... 

I will update the patch and send a new version. 

Best Regards,
Elena.