bpf: Fix an incorrect verification success with movsx insn

syzbot reports a verifier bug which triggers a runtime panic.
The test bpf program is:
   0: (62) *(u32 *)(r10 -8) = 553656332
   1: (bf) r1 = (s16)r10
   2: (07) r1 += -8
   3: (b7) r2 = 3
   4: (bd) if r2 <= r1 goto pc+0
   5: (85) call bpf_trace_printk#-138320
   6: (b7) r0 = 0
   7: (95) exit

At insn 1, the current implementation keeps 'r1' as a frame pointer,
which caused later bpf_trace_printk helper call crash since frame
pointer address is not valid any more. Note that at insn 4,
the 'pointer vs. scalar' comparison is allowed for privileged
prog run.

To fix the problem with above insn 1, the fix in the patch adopts
similar pattern to existing 'R1 = (u32) R2' handling. For unprivileged
prog run, verification will fail with 'R<num> sign-extension part of pointer'.
For privileged prog run, the dst_reg 'r1' will be marked as
an unknown scalar, so later 'bpf_trace_pointk' helper will complain
since it expected certain pointers.

Reported-by: syzbot+d61b595e9205573133b3@syzkaller.appspotmail.com
Fixes: 8100928c8814 ("bpf: Support new sign-extension mov insns")
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://lore.kernel.org/r/20230807175721.671696-1-yonghong.song@linux.dev
Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>

0 files changed, 0 insertions, 0 deletions