[E/F/Unstable][PATCH 0/1] crypto: fix regression/use-after-free in af_alg_accept()

[Mauricio Faria de Oliveira]

Regarding the submission policies for this kernel SRU cycle:
this patch does not necessarily have to be applied for now;
just review/ack for B/E would be useful if at all possible.

It has only been merged on Linus' tree yesterday.

The patch applies cleanly on Unstable/Focal/Eoan,
and is a trivial backport on Disco/Bionic/Xenial.
(it's on all series because it's a fix to stable.)

[Impact]

 * Users of the Linux kernel's crypto userspace API
   reported BUG() / kernel NULL pointer dereference
   errors after kernel upgrades.

 * The stack trace signature is an accept() syscall
   going through af_alg_accept() and hitting errors
   usually in one of:
   - apparmor_sk_clone_security()
   - apparmor_sock_graft()
   - release_sock()

[Fix]
      
 * This is a regression introduced by upstream commit
   37f96694cf73 ("crypto: af_alg - Use bh_lock_sock
   in sk_destruct") which made its way through stable.
   
 * The offending patch allows the critical regions
   of af_alg_accept() and af_alg_release_parent() to
   run concurrently; now with the "right" events on 2
   CPUs it might drop the non-atomic reference counter
   of the alg_sock then the sock, thus release a sock
   that is still in use.

 * The fix is upstream commit 34c86f4c4a7b ("crypto:
   af_alg - fix use-after-free in af_alg_accept() due
   to bh_lock_sock()") [1]. It changes alg_sock's ref
   counter to atomic, which addresses the root cause.
   
[Test Case]

 * There is a synthetic test case available, which
   uses a kprobes kernel module to synchronize the
   concurrent CPUs on the instructions responsible
   for the problem; and a userspace part to run it.

 * The organic reproducer is the Varnish Cache Plus
   software with the Crypto vmod (which uses kernel
   crypto userspace API) under long, very high load.
   
 * The patch has been verified on both reproducers
   with the 4.15 and 5.7 kernels.
   
 * More tests performed with 'stress-ng --af-alg'
   with 11 CPUs on Xenial/Bionic/Disco/Eoan/Focal
   (all on same version of stress-ng, V0.11.14)
   
   No regressions observed from original kernel.
   (the af-alg stressor can exercise almost all
   kernel crypto modules shipped with the kernel;
   so it checks more paths/crypto alg interfaces.)
   
[Regression Potential]

 * The fix patch does a fundamental change in how
   alg_sock reference counters work, plus another
   change to the 'nokey' counting. This of course
   *has* a risk of regression.

 * Regressions theoretically could manifest as use
   after free errors (in case of undercounting) in
   the af_alg functions or silent memory leaks (in
   case of overcounting), but also other behaviors
   since reference counting is key to many things.
   
 * FWIW, this patch has been written by the crypto
   subsystem maintainer, who certainly knows a lot
   of the normal and corner cases, thus giving the
   patch more credit.
   
 * Testing with the organic reproducer ran as long
   as 5 days, without issues, so it does look good.

[Other Info]

 * Not sending for Groovy (should get via Unstable).
 
 * [1] Patch: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=34c86f4c4a7be3b3e35aa48bd18299d4c756064d
 
[Stack Trace Examples]

Examples:

    BUG: unable to handle kernel NULL pointer dereference at 0000000000000000
    ...
    RIP: 0010:apparmor_sk_clone_security+0x26/0x70
    ...
    Call Trace:
     security_sk_clone+0x33/0x50
     af_alg_accept+0x81/0x1c0 [af_alg]
     alg_accept+0x15/0x20 [af_alg]
     SYSC_accept4+0xff/0x210
     SyS_accept+0x10/0x20
     do_syscall_64+0x73/0x130
     entry_SYSCALL_64_after_hwframe+0x3d/0xa2

    general protection fault: 0000 [#1] SMP PTI
    ...
    RIP: 0010:__release_sock+0x54/0xe0
    ...
    Call Trace:
     release_sock+0x30/0xa0
     af_alg_accept+0x122/0x1c0 [af_alg]
     alg_accept+0x15/0x20 [af_alg]
     SYSC_accept4+0xff/0x210
     SyS_accept+0x10/0x20
     do_syscall_64+0x73/0x130
     entry_SYSCALL_64_after_hwframe+0x3d/0xa2

Herbert Xu (1):
  crypto: af_alg - fix use-after-free in af_alg_accept() due to
    bh_lock_sock()

 crypto/af_alg.c         | 26 +++++++++++---------------
 crypto/algif_aead.c     |  9 +++------
 crypto/algif_hash.c     |  9 +++------
 crypto/algif_skcipher.c |  9 +++------
 include/crypto/if_alg.h |  4 ++--
 5 files changed, 22 insertions(+), 35 deletions(-)

-- 
2.25.1