On Wed, 20 May 2020 16:34:05 +0200
Björn Töpel <bjorn.topel@xxxxxxxxx> wrote:
> On 2020-05-20 15:18, Jesper Dangaard Brouer wrote:
> > On Wed, 20 May 2020 11:47:28 +0200
> > Björn Töpel <bjorn.topel@xxxxxxxxx> wrote:
> >
> >> From: Björn Töpel <bjorn.topel@xxxxxxxxx>
> >>
> >> Calculating the "data_hard_end" for an XDP buffer coming from AF_XDP
> >> zero-copy mode, the return value of xsk_umem_xdp_frame_sz() is added
> >> to "data_hard_start".
> >>
> >> Currently, the chunk size of the UMEM is returned by
> >> xsk_umem_xdp_frame_sz(). This is not correct, if the fixed UMEM
> >> headroom is non-zero. Fix this by returning the chunk_size without the
> >> UMEM headroom.
> >>
> >> Fixes: 2a637c5b1aaf ("xdp: For Intel AF_XDP drivers add XDP frame_sz")
> >> Signed-off-by: Björn Töpel <bjorn.topel@xxxxxxxxx>
> >> ---
> >> include/net/xdp_sock.h | 2 +-
> >> 1 file changed, 1 insertion(+), 1 deletion(-)
> >>
> >> diff --git a/include/net/xdp_sock.h b/include/net/xdp_sock.h
> >> index abd72de25fa4..6b1137ce1692 100644
> >> --- a/include/net/xdp_sock.h
> >> +++ b/include/net/xdp_sock.h
> >> @@ -239,7 +239,7 @@ static inline u64 xsk_umem_adjust_offset(struct xdp_umem *umem, u64 address,
> >>
> >> static inline u32 xsk_umem_xdp_frame_sz(struct xdp_umem *umem)
> >> {
> >> - return umem->chunk_size_nohr + umem->headroom;
> >> + return umem->chunk_size_nohr;
> >
> > Hmm, is this correct?
> >
> > As you write "xdp_data_hard_end" is calculated as an offset from
> > xdp->data_hard_start pointer based on the frame_sz. Will your
> > xdp->data_hard_start + frame_sz point to packet end?
> >
>
> Yes, I believe this is correct.
>
> Say that a user uses a chunk size of 2k, and a umem headroom of, say,
> 64. This means that the kernel should (at least) leave 64B which the
> kernel shouldn't touch.
>
> umem->headroom | XDP_PACKET_HEADROOM | packet | |
> ^ ^ ^ ^ ^
> a b c d e
>
> a: data_hard_start
> b: data
> c: data_end
> d: data_hard_end, (e - 320)
> e: hardlimit of chunk, a + umem->chunk_size_nohr
>
> Prior this fix the umem->headroom was *included* in frame_sz.
Thanks for the nice ascii art description. I can now see that you are
right. We should add this kind of documentation, perhaps as a comment
in the code?
> > #define xdp_data_hard_end(xdp) \
> > ((xdp)->data_hard_start + (xdp)->frame_sz - \
> > SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
> >
> > Note the macro reserves the last 320 bytes (for skb_shared_info), but
> > for AF_XDP zero-copy mode, it will never create an SKB that use this
> > area. Thus, in principle we can allow XDP-progs to extend/grow tail
> > into this area, but I don't think there is any use-case for this, as
> > it's much easier to access packet-data in userspace application.
> > (Thus, it might not be worth the complexity to give AF_XDP
> > bpf_xdp_adjust_tail access to this area, by e.g. "lying" via adding 320
> > bytes to frame_sz).
> >
>
> I agree, and in the picture (well...) above that would be "d". IOW
> data_hard_end is 320 "off" the real end.
Yes, we agree.
--
Best regards,
Jesper Dangaard Brouer
MSc.CS, Principal Kernel Engineer at Red Hat
LinkedIn: http://www.linkedin.com/in/brouer
