In that case, emit_udiv() will be called with rn == ARM_R0 (r_scratch)
and loading rm first into ARM_R0 will result in jit_udiv() function
being called the same dividend and divisor. Fix that by loading rn
first into ARM_R1 and then rm into ARM_R0.
Signed-off-by: Nicolas Schichan <nschichan@xxxxxxxxxx>
Cc: <stable@xxxxxxxxxxxxxxx> # v3.13+
Fixes: aee636c4809f (bpf: do not use reciprocal divide)
arch/arm/net/bpf_jit_32.c | 15 +++++++++++++--
1 file changed, 13 insertions(+), 2 deletions(-)
This patch was part of a serie adding support to JITed seccomp filters
to ARM but as it is an unrelated fix, it is more appropriate to send
Changes from previous version: add a comment clarifying how
emit_udiv() is called when translating BPF_ALU | BPF_DIV | BPF_K and
BPF_ALU | BPF_DIV | BPF_X with respect to the rm and rn parameters.
diff --git a/arch/arm/net/bpf_jit_32.c b/arch/arm/net/bpf_jit_32.c
index e1268f9..f412b53 100644
@@ -449,10 +449,21 @@ static inline void emit_udiv(u8 rd, u8 rm, u8 rn, struct jit_ctx *ctx)
- if (rm != ARM_R0)
- emit(ARM_MOV_R(ARM_R0, rm), ctx);
+ * For BPF_ALU | BPF_DIV | BPF_K instructions, rm is ARM_R4
+ * (r_A) and rn is ARM_R0 (r_scratch) so load rn first into
+ * ARM_R1 to avoid accidentally overwriting ARM_R0 with rm
+ * before using it as a source for ARM_R1.
+ * For BPF_ALU | BPF_DIV | BPF_X rm is ARM_R4 (r_A) and rn is
+ * ARM_R5 (r_X) so there is no particular register overlap
+ * issues.
if (rn != ARM_R1)
emit(ARM_MOV_R(ARM_R1, rn), ctx);
+ if (rm != ARM_R0)
+ emit(ARM_MOV_R(ARM_R0, rm), ctx);
ctx->seen |= SEEN_CALL;
emit_mov_i(ARM_R3, (u32)jit_udiv, ctx);