[PATCH bpf-next 0/3] bpf: Fix and test aux usage after do_check_insn()

[Luis Gerhorst]

Fix cur_aux()->nospec_result test after do_check_insn() referring to the
to-be-analyzed (potentially unsafe) instruction, not the
already-analyzed (safe) instruction. This might allow a unsafe insn to
slip through on a speculative path. Create some tests from the
reproducer [1].

Commit d6f1c85f2253 ("bpf: Fall back to nospec for Spectre v1") should
not be in any stable kernel yet, therefore bpf-next should suffice.

[1] https://lore.kernel.org/bpf/685b3c1b.050a0220.2303ee.0010.GAE@xxxxxxxxxx/

Changes since RFC:
- Introduce prev_aux() as suggested by Alexei. For this, we must move
  the env->prev_insn_idx assignment to happen directly after
  do_check_insn(), for which I have created a separate commit. This
  patch could be simplified by using a local prev_aux variable as
  sugested by Eduard, but I figured one might find the new
  assignment-strategy easier to understand (before, prev_insn_idx and
  env->prev_insn_idx were out-of-sync for the latter part of the loop).
  Also, like this we do not have an additional prev_* variable that must
  be kept in-sync and the local variable's usage (old prev_insn_idx, new
  tmp) is much more local. If you think it would be better to not take
  the risk and keep the fix simple by just introducing the prev_aux
  variable, let me know.
- Change WARN_ON_ONCE() to verifier_bug_if() as suggested by Alexei
- Change assertion to check instruction is BPF_JMP[32] as suggested by
  Eduard
- RFC: https://lore.kernel.org/bpf/8734bmoemx.fsf@xxxxxx/

Luis Gerhorst (3):
  bpf: Update env->prev_insn_idx after do_check_insn()
  bpf: Fix aux usage after do_check_insn()
  selftests/bpf: Add Spectre v4 tests

 kernel/bpf/verifier.c                         |  30 ++--
 tools/testing/selftests/bpf/progs/bpf_misc.h  |   4 +
 .../selftests/bpf/progs/verifier_unpriv.c     | 149 ++++++++++++++++++
 3 files changed, 174 insertions(+), 9 deletions(-)


base-commit: d69bafe6ee2b5eff6099fa26626ecc2963f0f363
-- 
2.49.0