Common to all test runs: stress-ng version,
and command to load as many crypto modules
as found/possible in the system.

$ ./stress-ng --version
stress-ng, version 0.11.14 (<build system details>) ūüíĽūüĒ•

$ sudo modprobe -a \
          $(modinfo \
              /lib/modules/$(uname -r)/kernel/crypto/*.ko \
              /lib/modules/$(uname -r)/kernel/arch/*/crypto/*.ko \
              | grep -ow 'crypto-.*')

You received this bug notification because you are a member of Kernel
Packages, which is subscribed to linux in Ubuntu.

  use-after-free in af_alg_accept() due to bh_lock_sock()

Status in linux package in Ubuntu:
  In Progress
Status in linux source package in Xenial:
  In Progress
Status in linux source package in Bionic:
  In Progress
Status in linux source package in Eoan:
  In Progress
Status in linux source package in Focal:
  In Progress
Status in linux source package in Groovy:
  Won't Fix

Bug description:

   * 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()


   * 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:

  [Stack Trace Examples]


      BUG: unable to handle kernel NULL pointer dereference at 0000000000000000
      RIP: 0010:apparmor_sk_clone_security+0x26/0x70
      Call Trace:
       af_alg_accept+0x81/0x1c0 [af_alg]
       alg_accept+0x15/0x20 [af_alg]

      general protection fault: 0000 [#1] SMP PTI
      RIP: 0010:__release_sock+0x54/0xe0
      Call Trace:
       af_alg_accept+0x122/0x1c0 [af_alg]
       alg_accept+0x15/0x20 [af_alg]

To manage notifications about this bug go to:

Mailing list:
Post to     :
Unsubscribe :
More help   :

Reply via email to