Arguments to a raw tracepoint are tagged as trusted, which carries the
semantics that the pointer will be non-NULL. However, in certain cases,
a raw tracepoint argument may end up being NULL. More context about this
issue is available in [0].
Thus, there is a discrepancy between the reality, that raw_tp arguments can
actually be NULL, and the verifier's knowledge, that they are never NULL,
causing explicit NULL check branch to be dead code eliminated.
A previous attempt [1], i.e. the second fixed commit, was made to
simulate symbolic execution as if in most accesses, the argument is a
non-NULL raw_tp, except for conditional jumps. This tried to suppress
branch prediction while preserving compatibility, but surfaced issues
with production programs that were difficult to solve without increasing
verifier complexity. A more complete discussion of issues and fixes is
available at [2].
Fix this by maintaining an explicit list of tracepoints where the
arguments are known to be NULL, and mark the positional arguments as
PTR_MAYBE_NULL. Additionally, capture the tracepoints where arguments
are known to be ERR_PTR, and mark these arguments as scalar values to
prevent potential dereference.
Each hex digit is used to encode NULL-ness (0x1) or ERR_PTR-ness (0x2),
shifted by the zero-indexed argument number x 4. This can be represented
as follows:
1st arg: 0x1
2nd arg: 0x10
3rd arg: 0x100
... and so on (likewise for ERR_PTR case).
In the future, an automated pass will be used to produce such a list, or
insert __nullable annotations automatically for tracepoints. Each
compilation unit will be analyzed and results will be collated to find
whether a tracepoint pointer is definitely not null, maybe null, or an
unknown state where verifier conservatively marks it PTR_MAYBE_NULL.
A proof of concept of this tool from Eduard is available at [3].
Note that in case we don't find a specification in the raw_tp_null_args
array and the tracepoint belongs to a kernel module, we will
conservatively mark the arguments as PTR_MAYBE_NULL. This is because
unlike for in-tree modules, out-of-tree module tracepoints may pass NULL
freely to the tracepoint. We don't protect against such tracepoints
passing ERR_PTR (which is uncommon anyway), lest we mark all such
arguments as SCALAR_VALUE.
While we are it, let's adjust the test raw_tp_null to not perform
dereference of the skb->mark, as that won't be allowed anymore, and make
it more robust by using inline assembly to test the dead code
elimination behavior, which should still stay the same.
[0]: https://lore.kernel.org/bpf/ZrCZS6nisraEqehw@jlelli-thinkpadt14gen4.remote.csb
[1]: https://lore.kernel.org/all/20241104171959.2938862-1-memxor@gmail.com
[2]: https://lore.kernel.org/bpf/20241206161053.809580-1-memxor@gmail.com
[3]: https://github.com/eddyz87/llvm-project/tree/nullness-for-tracepoint-params
Reported-by: Juri Lelli <juri.lelli@redhat.com> # original bug
Reported-by: Manu Bretelle <chantra@meta.com> # bugs in masking fix
Fixes: 3f00c52393 ("bpf: Allow trusted pointers to be passed to KF_TRUSTED_ARGS kfuncs")
Fixes: cb4158ce8e ("bpf: Mark raw_tp arguments with PTR_MAYBE_NULL")
Reviewed-by: Eduard Zingerman <eddyz87@gmail.com>
Co-developed-by: Jiri Olsa <jolsa@kernel.org>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20241213221929.3495062-3-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
838a10bd2e