Message ID | 3fb334388ac7af755e1f03abb76a0a6335a90ff6.1603824486.git.lorenzo@kernel.org (mailing list archive) |
---|---|
State | Superseded |
Delegated to: | BPF |
Headers | show |
Series | xdp: introduce bulking for page_pool tx return path | expand |
Lorenzo Bianconi <lorenzo@kernel.org> writes: > Convert mlx5 driver to xdp_return_frame_bulk APIs. > > XDP_REDIRECT (upstream codepath): 8.5Mpps > XDP_REDIRECT (upstream codepath + bulking APIs): 10.1Mpps > > Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org> > --- > drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c | 5 ++++- > 1 file changed, 4 insertions(+), 1 deletion(-) > > diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c > b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c > index ae90d533a350..5fdfbf390d5c 100644 > --- a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c > +++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c > @@ -369,8 +369,10 @@ static void mlx5e_free_xdpsq_desc(struct > mlx5e_xdpsq *sq, > bool recycle) > { > struct mlx5e_xdp_info_fifo *xdpi_fifo = &sq->db.xdpi_fifo; > + struct xdp_frame_bulk bq; > u16 i; > > + bq.xa = NULL; > for (i = 0; i < wi->num_pkts; i++) { > struct mlx5e_xdp_info xdpi = > mlx5e_xdpi_fifo_pop(xdpi_fifo); > > @@ -379,7 +381,7 @@ static void mlx5e_free_xdpsq_desc(struct > mlx5e_xdpsq *sq, > /* XDP_TX from the XSK RQ and XDP_REDIRECT > */ > dma_unmap_single(sq->pdev, > xdpi.frame.dma_addr, > xdpi.frame.xdpf->len, > DMA_TO_DEVICE); > - xdp_return_frame(xdpi.frame.xdpf); > + xdp_return_frame_bulk(xdpi.frame.xdpf, > &bq); > break; > case MLX5E_XDP_XMIT_MODE_PAGE: > /* XDP_TX from the regular RQ */ > @@ -393,6 +395,7 @@ static void mlx5e_free_xdpsq_desc(struct > mlx5e_xdpsq *sq, > WARN_ON_ONCE(true); > } > } > + xdp_flush_frame_bulk(&bq); While I understand the rational behind this patchset, using an intermediate buffer void *q[XDP_BULK_QUEUE_SIZE]; means more pressure on the data cache. At the time I ran performance tests on mlx5 to see whether batching skbs before passing them to GRO would improve performance. On some flows I got worse performance. This function seems to have less Dcache contention than RX flow, but maybe some performance testing are needed here. > } > > bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
> > Lorenzo Bianconi <lorenzo@kernel.org> writes: > > > Convert mlx5 driver to xdp_return_frame_bulk APIs. > > > > XDP_REDIRECT (upstream codepath): 8.5Mpps > > XDP_REDIRECT (upstream codepath + bulking APIs): 10.1Mpps > > > > Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org> > > --- > > drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c | 5 ++++- > > 1 file changed, 4 insertions(+), 1 deletion(-) > > > > diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c > > b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c > > index ae90d533a350..5fdfbf390d5c 100644 > > --- a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c > > +++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c > > @@ -369,8 +369,10 @@ static void mlx5e_free_xdpsq_desc(struct > > mlx5e_xdpsq *sq, > > bool recycle) > > { > > struct mlx5e_xdp_info_fifo *xdpi_fifo = &sq->db.xdpi_fifo; > > + struct xdp_frame_bulk bq; > > u16 i; > > + bq.xa = NULL; > > for (i = 0; i < wi->num_pkts; i++) { > > struct mlx5e_xdp_info xdpi = mlx5e_xdpi_fifo_pop(xdpi_fifo); > > @@ -379,7 +381,7 @@ static void mlx5e_free_xdpsq_desc(struct > > mlx5e_xdpsq *sq, > > /* XDP_TX from the XSK RQ and XDP_REDIRECT */ > > dma_unmap_single(sq->pdev, xdpi.frame.dma_addr, > > xdpi.frame.xdpf->len, DMA_TO_DEVICE); > > - xdp_return_frame(xdpi.frame.xdpf); > > + xdp_return_frame_bulk(xdpi.frame.xdpf, &bq); > > break; > > case MLX5E_XDP_XMIT_MODE_PAGE: > > /* XDP_TX from the regular RQ */ > > @@ -393,6 +395,7 @@ static void mlx5e_free_xdpsq_desc(struct mlx5e_xdpsq > > *sq, > > WARN_ON_ONCE(true); > > } > > } > > + xdp_flush_frame_bulk(&bq); > > While I understand the rational behind this patchset, using an intermediate > buffer > void *q[XDP_BULK_QUEUE_SIZE]; > means more pressure on the data cache. > > At the time I ran performance tests on mlx5 to see whether batching skbs > before passing them to GRO would improve performance. On some flows I got > worse performance. > This function seems to have less Dcache contention than RX flow, but maybe > some performance testing are needed here. Hi Shay, this codepath is only activated for "redirected" frames (not for packets forwarded to networking stack). We run performance comparisons with the upstream code for this particular use-case and we reported a nice improvements (8.5Mpps vs 10.1Mpps). Do you have in mind other possible performance tests to run? Regards, Lorenzo > > > } > > bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq) >
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c index ae90d533a350..5fdfbf390d5c 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c @@ -369,8 +369,10 @@ static void mlx5e_free_xdpsq_desc(struct mlx5e_xdpsq *sq, bool recycle) { struct mlx5e_xdp_info_fifo *xdpi_fifo = &sq->db.xdpi_fifo; + struct xdp_frame_bulk bq; u16 i; + bq.xa = NULL; for (i = 0; i < wi->num_pkts; i++) { struct mlx5e_xdp_info xdpi = mlx5e_xdpi_fifo_pop(xdpi_fifo); @@ -379,7 +381,7 @@ static void mlx5e_free_xdpsq_desc(struct mlx5e_xdpsq *sq, /* XDP_TX from the XSK RQ and XDP_REDIRECT */ dma_unmap_single(sq->pdev, xdpi.frame.dma_addr, xdpi.frame.xdpf->len, DMA_TO_DEVICE); - xdp_return_frame(xdpi.frame.xdpf); + xdp_return_frame_bulk(xdpi.frame.xdpf, &bq); break; case MLX5E_XDP_XMIT_MODE_PAGE: /* XDP_TX from the regular RQ */ @@ -393,6 +395,7 @@ static void mlx5e_free_xdpsq_desc(struct mlx5e_xdpsq *sq, WARN_ON_ONCE(true); } } + xdp_flush_frame_bulk(&bq); } bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
Convert mlx5 driver to xdp_return_frame_bulk APIs. XDP_REDIRECT (upstream codepath): 8.5Mpps XDP_REDIRECT (upstream codepath + bulking APIs): 10.1Mpps Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org> --- drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)