This patch fixes the invariant violations that can happen after we
refine ranges & tnum based on an incorrectly-detected branch condition.
For example, the branch is always true, but we miss it in
is_branch_taken; we then refine based on the branch being false and end
up with incoherent ranges (e.g. umax < umin).
To avoid this, we can simulate the refinement on both branches. More
specifically, this patch simulates both branches taken using
regs_refine_cond_op and reg_bounds_sync. If the resulting register
states are ill-formed on one of the branches, is_branch_taken can mark
that branch as "never taken".
On a more formal note, we can deduce a branch is not taken when
regs_refine_cond_op or reg_bounds_sync returns an ill-formed state
because the branch operators are sound (verified with Agni [1]).
Soundness means that the verifier is guaranteed to produce sound
outputs on the taken branches. On the non-taken branch (explored
because of imprecision in the bounds), the verifier is free to produce
any output. We use ill-formedness as a signal that the branch is dead
and prune that branch.
This patch moves the refinement logic for both branches from
reg_set_min_max to their own function, simulate_both_branches_taken,
which is called from is_scalar_branch_taken. As a result,
reg_set_min_max now only runs sanity checks and has been renamed to
reg_bounds_sanity_check_branches to reflect that.
We have had five patches fixing specific cases of invariant violations
in the past, all added with selftests:
- commit fbc7aef517 ("bpf: Fix u32/s32 bounds when ranges cross
min/max boundary")
- commit efc11a6678 ("bpf: Improve bounds when tnum has a single
possible value")
- commit f41345f47f ("bpf: Use tnums for JEQ/JNE is_branch_taken
logic")
- commit 00bf8d0c6c ("bpf: Improve bounds when s64 crosses sign
boundary")
- commit 6279846b9b ("bpf: Forget ranges when refining tnum after
JSET")
To confirm that this patch addresses all invariant violations, we have
also reverted those five commits and verified that their related
selftests don't cause any invariant violation warnings anymore. Those
selftests still fail but only because of misdetected branches or
less-precise bounds than expected. This demonstrates that the current
patch is enough to avoid the invariant violation warning AND that the
previous five patches are still useful to improve branch detection.
In addition to the selftests, this change was also tested with the
Cilium complexity test suite: all programs were successfully loaded and
it didn't change the number of processed instructions.
Link: https://github.com/bpfverif/agni [1]
Reported-by: syzbot+c950cc277150935cc0b5@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=c950cc277150935cc0b5
Co-developed-by: Paul Chaignon <paul.chaignon@gmail.com>
Signed-off-by: Paul Chaignon <paul.chaignon@gmail.com>
Co-developed-by: Srinivas Narayana <srinivas.narayana@rutgers.edu>
Signed-off-by: Srinivas Narayana <srinivas.narayana@rutgers.edu>
Co-developed-by: Santosh Nagarakatte <santosh.nagarakatte@rutgers.edu>
Signed-off-by: Santosh Nagarakatte <santosh.nagarakatte@rutgers.edu>
Signed-off-by: Harishankar Vishwanathan <harishankar.vishwanathan@gmail.com>
Acked-by: Shung-Hsi Yu <shung-hsi.yu@suse.com>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://lore.kernel.org/r/a166b54a3cbbbdbcdf8a87f53045f1097176218b.1775142354.git.paul.chaignon@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
b254c6d816