On Wed, Jul 29, 2020 at 9:25 AM John Fastabend <john.fastabend@xxxxxxxxx> wrote:
>
> Similar to patch ("bpf: sock_ops ctx access may stomp registers") if the
> src_reg = dst_reg when reading the sk field of a sock_ops struct we
> generate xlated code,
>
> 53: (61) r9 = *(u32 *)(r9 +28)
> 54: (15) if r9 == 0x0 goto pc+3
> 56: (79) r9 = *(u64 *)(r9 +0)
>
> This stomps on the r9 reg to do the sk_fullsock check and then when
> reading the skops->sk field instead of the sk pointer we get the
> sk_fullsock. To fix use similar pattern noted in the previous fix
> and use the temp field to save/restore a register used to do
> sk_fullsock check.
>
> After the fix the generated xlated code reads,
>
> 52: (7b) *(u64 *)(r9 +32) = r8
> 53: (61) r8 = *(u32 *)(r9 +28)
> 54: (15) if r9 == 0x0 goto pc+3
> 55: (79) r8 = *(u64 *)(r9 +32)
> 56: (79) r9 = *(u64 *)(r9 +0)
> 57: (05) goto pc+1
> 58: (79) r8 = *(u64 *)(r9 +32)
>
> Here r9 register was in-use so r8 is chosen as the temporary register.
> In line 52 r8 is saved in temp variable and at line 54 restored in case
> fullsock != 0. Finally we handle fullsock == 0 case by restoring at
> line 58.
>
> This adds a new macro SOCK_OPS_GET_SK it is almost possible to merge
> this with SOCK_OPS_GET_FIELD, but I found the extra branch logic a
> bit more confusing than just adding a new macro despite a bit of
> duplicating code.
>
> Fixes: 1314ef561102e ("bpf: export bpf_sock for BPF_PROG_TYPE_SOCK_OPS prog type")
> Signed-off-by: John Fastabend <john.fastabend@xxxxxxxxx>
Acked-by: Song Liu <songliubraving@xxxxxx>
[...]
