Merge branch 'bpf-xsk-rx-batch'

Merge branch 'bpf-xsk-rx-batch'

Magnus Karlsson says:

====================
This patch set introduces a batched interface for Rx buffer allocation
in AF_XDP buffer pool. Instead of using xsk_buff_alloc(*pool), drivers
can now use xsk_buff_alloc_batch(*pool, **xdp_buff_array,
max). Instead of returning a pointer to an xdp_buff, it returns the
number of xdp_buffs it managed to allocate up to the maximum value of
the max parameter in the function call. Pointers to the allocated
xdp_buff:s are put in the xdp_buff_array supplied in the call. This
could be a SW ring that already exists in the driver or a new
structure that the driver has allocated.

  u32 xsk_buff_alloc_batch(struct xsk_buff_pool *pool,
                           struct xdp_buff **xdp,
                           u32 max);

When using this interface, the driver should also use the new
interface below to set the relevant fields in the struct xdp_buff. The
reason for this is that xsk_buff_alloc_batch() does not fill in the
data and data_meta fields for you as is the case with
xsk_buff_alloc(). So it is not sufficient to just set data_end
(effectively the size) anymore in the driver. The reason for this is
performance as explained in detail in the commit message.

  void xsk_buff_set_size(struct xdp_buff *xdp, u32 size);

Patch 6 also optimizes the buffer allocation in the aligned case. In
this case, we can skip the reinitialization of most fields in the
xdp_buff_xsk struct at allocation time. As the number of elements in
the heads array is equal to the number of possible buffers in the
umem, we can initialize them once and for all at bind time and then
just point to the correct one in the xdp_buff_array that is returned
to the driver. No reason to have a stack of free head entries. In the
unaligned case, the buffers can reside anywhere in the umem, so this
optimization is not possible as we still have to fill in the right
information in the xdp_buff every single time one is allocated.

I have updated i40e and ice to use this new batched interface.

These are the throughput results on my 2.1 GHz Cascade Lake system:

Aligned mode:
ice: +11% / -9 cycles/pkt
i40e: +12% / -9 cycles/pkt

Unaligned mode:
ice: +1.5% / -1 cycle/pkt
i40e: +1% / -1 cycle/pkt

For the aligned case, batching provides around 40% of the performance
improvement and the aligned optimization the rest, around 60%. Would
have expected a ~4% boost for unaligned with this data, but I only get
around 1%. Do not know why. Note that memory consumption in aligned
mode is also reduced by this patch set.
====================

Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>