diff mbox

crypto: algapi - fix NULL dereference in crypto_remove_spawns()

Message ID 20171229203019.1413-1-ebiggers3@gmail.com (mailing list archive)
State Accepted
Delegated to: Herbert Xu
Headers show

Commit Message

Eric Biggers Dec. 29, 2017, 8:30 p.m. UTC
From: Eric Biggers <ebiggers@google.com>

syzkaller triggered a NULL pointer dereference in crypto_remove_spawns()
via a program that repeatedly and concurrently requests AEADs
"authenc(cmac(des3_ede-asm),pcbc-aes-aesni)" and hashes "cmac(des3_ede)"
through AF_ALG, where the hashes are requested as "untested"
(CRYPTO_ALG_TESTED is set in ->salg_mask but clear in ->salg_feat; this
causes the template to be instantiated for every request).

Although AF_ALG users really shouldn't be able to request an "untested"
algorithm, the NULL pointer dereference is actually caused by a
longstanding race condition where crypto_remove_spawns() can encounter
an instance which has had spawn(s) "grabbed" but hasn't yet been
registered, resulting in ->cra_users still being NULL.

We probably should properly initialize ->cra_users earlier, but that
would require updating many templates individually.  For now just fix
the bug in a simple way that can easily be backported: make
crypto_remove_spawns() treat a NULL ->cra_users list as empty.

Reported-by: syzbot <syzkaller@googlegroups.com>
Cc: stable@vger.kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>
---
 crypto/algapi.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)

Comments

Herbert Xu Jan. 5, 2018, 11:18 a.m. UTC | #1
On Fri, Dec 29, 2017 at 02:30:19PM -0600, Eric Biggers wrote:
> From: Eric Biggers <ebiggers@google.com>
> 
> syzkaller triggered a NULL pointer dereference in crypto_remove_spawns()
> via a program that repeatedly and concurrently requests AEADs
> "authenc(cmac(des3_ede-asm),pcbc-aes-aesni)" and hashes "cmac(des3_ede)"
> through AF_ALG, where the hashes are requested as "untested"
> (CRYPTO_ALG_TESTED is set in ->salg_mask but clear in ->salg_feat; this
> causes the template to be instantiated for every request).
> 
> Although AF_ALG users really shouldn't be able to request an "untested"
> algorithm, the NULL pointer dereference is actually caused by a
> longstanding race condition where crypto_remove_spawns() can encounter
> an instance which has had spawn(s) "grabbed" but hasn't yet been
> registered, resulting in ->cra_users still being NULL.
> 
> We probably should properly initialize ->cra_users earlier, but that
> would require updating many templates individually.  For now just fix
> the bug in a simple way that can easily be backported: make
> crypto_remove_spawns() treat a NULL ->cra_users list as empty.
> 
> Reported-by: syzbot <syzkaller@googlegroups.com>
> Cc: stable@vger.kernel.org
> Signed-off-by: Eric Biggers <ebiggers@google.com>

Patch applied.  Thanks.
diff mbox

Patch

diff --git a/crypto/algapi.c b/crypto/algapi.c
index 9895cafcce7e..395b082d03a9 100644
--- a/crypto/algapi.c
+++ b/crypto/algapi.c
@@ -166,6 +166,18 @@  void crypto_remove_spawns(struct crypto_alg *alg, struct list_head *list,
 
 			spawn->alg = NULL;
 			spawns = &inst->alg.cra_users;
+
+			/*
+			 * We may encounter an unregistered instance here, since
+			 * an instance's spawns are set up prior to the instance
+			 * being registered.  An unregistered instance will have
+			 * NULL ->cra_users.next, since ->cra_users isn't
+			 * properly initialized until registration.  But an
+			 * unregistered instance cannot have any users, so treat
+			 * it the same as ->cra_users being empty.
+			 */
+			if (spawns->next == NULL)
+				break;
 		}
 	} while ((spawns = crypto_more_spawns(alg, &stack, &top,
 					      &secondary_spawns)));