Re: [PATCH bpf] bpf: fix ldx in ld_abs rewrite for large offsets

On Tue, Jul 10, 2018 at 02:13:42PM +0200, Daniel Borkmann wrote:

On 07/10/2018 12:14 PM, Mark Rutland wrote:

quoted

On Tue, Jul 10, 2018 at 12:43:22AM +0200, Daniel Borkmann wrote:

quoted

Mark reported that syzkaller triggered a KASAN detected slab-out-of-bounds
bug in ___bpf_prog_run() with a BPF_LD | BPF_ABS word load at offset 0x8001.
After further investigation it became clear that the issue was the
BPF_LDX_MEM() which takes offset as an argument whereas it cannot encode
larger than S16_MAX offsets into it. For this synthetical case we need to
move the full address into tmp register instead and do the LDX without
immediate value.

Fixes: e0cea7ce988c ("bpf: implement ld_abs/ld_ind in native bpf")
Reported-by: syzbot <redacted>
Reported-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
---
 net/core/filter.c | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)

diff --git a/net/core/filter.c b/net/core/filter.c
index 5fa66a3..a13f5b1 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c

@@ -459,11 +459,21 @@ static bool convert_bpf_ld_abs(struct sock_filter *fp, struct bpf_insn **insnp)
 	     (!unaligned_ok && offset >= 0 &&
 	      offset + ip_align >= 0 &&
 	      offset + ip_align % size == 0))) {
+		bool ldx_off_ok = offset <= S16_MAX;
+

Given offset is a (signed) int, is it possible for that to be a negative
value less than S16_MIN? ... or is that ruled out elsewhere?

This branch here handles only positive offset.

Ah, sorry. I missed the "offset >= 0" check in the context above.

offset can be negative,
but in that case these insns won't be emitted and handling will be done
in 'slow path' via bpf_skb_load_helper_*().

Ok; looks good to me, then.

Thanks,
Mark.