diff --git a/tools/testing/selftests/bpf/progs/verifier_linked_scalars.c b/tools/testing/selftests/bpf/progs/verifier_linked_scalars.c index 2ef346c827c2..7bf7dbfd237d 100644 --- a/tools/testing/selftests/bpf/progs/verifier_linked_scalars.c +++ b/tools/testing/selftests/bpf/progs/verifier_linked_scalars.c @@ -363,4 +363,68 @@ void alu32_negative_offset(void) __sink(path[0]); } +void dummy_calls(void) +{ + bpf_iter_num_new(0, 0, 0); + bpf_iter_num_next(0); + bpf_iter_num_destroy(0); +} + +SEC("socket") +__success +__flag(BPF_F_TEST_STATE_FREQ) +int spurious_precision_marks(void *ctx) +{ + struct bpf_iter_num iter; + + asm volatile( + "r1 = %[iter];" + "r2 = 0;" + "r3 = 10;" + "call %[bpf_iter_num_new];" + "1:" + "r1 = %[iter];" + "call %[bpf_iter_num_next];" + "if r0 == 0 goto 4f;" + "r7 = *(u32 *)(r0 + 0);" + "r8 = *(u32 *)(r0 + 0);" + /* This jump can't be predicted and does not change r7 or r8 state. */ + "if r7 > r8 goto 2f;" + /* Branch explored first ties r2 and r7 as having the same id. */ + "r2 = r7;" + "goto 3f;" + "2:" + /* Branch explored second does not tie r2 and r7 but has a function call. */ + "call %[bpf_get_prandom_u32];" + "3:" + /* + * A checkpoint. + * When first branch is explored, this would inject linked registers + * r2 and r7 into the jump history. + * When second branch is explored, this would be a cache hit point, + * triggering propagate_precision(). + */ + "if r7 <= 42 goto +0;" + /* + * Mark r7 as precise using an if condition that is always true. + * When reached via the second branch, this triggered a bug in the backtrack_insn() + * because r2 (tied to r7) was propagated as precise to a call. + */ + "if r7 <= 0xffffFFFF goto +0;" + "goto 1b;" + "4:" + "r1 = %[iter];" + "call %[bpf_iter_num_destroy];" + : + : __imm_ptr(iter), + __imm(bpf_iter_num_new), + __imm(bpf_iter_num_next), + __imm(bpf_iter_num_destroy), + __imm(bpf_get_prandom_u32) + : __clobber_common, "r7", "r8" + ); + + return 0; +} + char _license[] SEC("license") = "GPL";