[PATCH v2 net-next] net: ethernet: enetc: unlock XDP_REDIRECT for XDP non-linear buffers

Even if full XDP_REDIRECT is not supported yet for non-linear XDP buffers
since we allow redirecting just into CPUMAPs, unlock XDP_REDIRECT for
S/G XDP buffer and rely on XDP stack to properly take care of the
frames.

Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
---
Changes since v1:
- drop Fixes tag
- unlock XDP_REDIRECT
- populate missing XDP metadata

Please note this patch is just compile tested
---
 drivers/net/ethernet/freescale/enetc/enetc.c | 27 +++++++++-----------
 1 file changed, 12 insertions(+), 15 deletions(-)

diff --git a/drivers/net/ethernet/freescale/enetc/enetc.c b/drivers/net/ethernet/freescale/enetc/enetc.c
index 8671591cb750..9fd15e1e692d 100644
--- a/drivers/net/ethernet/freescale/enetc/enetc.c
+++ b/drivers/net/ethernet/freescale/enetc/enetc.c

@@ -1412,6 +1412,16 @@ static void enetc_add_rx_buff_to_xdp(struct enetc_bdr *rx_ring, int i,
 	/* To be used for XDP_TX */
 	rx_swbd->len = size;
 
+	if (!xdp_buff_has_frags(xdp_buff)) {
+		xdp_buff_set_frags_flag(xdp_buff);
+		shinfo->xdp_frags_size = size;
+	} else {
+		shinfo->xdp_frags_size += size;
+	}
+
+	if (page_is_pfmemalloc(rx_swbd->page))
+		xdp_buff_set_frag_pfmemalloc(xdp_buff);
+
 	skb_frag_off_set(frag, rx_swbd->page_offset);
 	skb_frag_size_set(frag, size);
 	__skb_frag_set_page(frag, rx_swbd->page);

@@ -1601,22 +1611,7 @@ static int enetc_clean_rx_ring_xdp(struct enetc_bdr *rx_ring,
 			}
 			break;
 		case XDP_REDIRECT:
-			/* xdp_return_frame does not support S/G in the sense
-			 * that it leaks the fragments (__xdp_return should not
-			 * call page_frag_free only for the initial buffer).
-			 * Until XDP_REDIRECT gains support for S/G let's keep
-			 * the code structure in place, but dead. We drop the
-			 * S/G frames ourselves to avoid memory leaks which
-			 * would otherwise leave the kernel OOM.
-			 */
-			if (unlikely(cleaned_cnt - orig_cleaned_cnt != 1)) {
-				enetc_xdp_drop(rx_ring, orig_i, i);
-				rx_ring->stats.xdp_redirect_sg++;
-				break;
-			}
-
 			tmp_orig_i = orig_i;
-
 			while (orig_i != i) {
 				enetc_flip_rx_buff(rx_ring,
 						   &rx_ring->rx_swbd[orig_i]);

@@ -1628,6 +1623,8 @@ static int enetc_clean_rx_ring_xdp(struct enetc_bdr *rx_ring,
 				enetc_xdp_free(rx_ring, tmp_orig_i, i);
 			} else {
 				xdp_redirect_frm_cnt++;
+				if (xdp_buff_has_frags(&xdp_buff))
+					rx_ring->stats.xdp_redirect_sg++;
 				rx_ring->stats.xdp_redirect++;
 			}
 		}

-- 
2.38.1

On Fri, Dec 09, 2022 at 01:19:44AM +0100, Lorenzo Bianconi wrote:

Even if full XDP_REDIRECT is not supported yet for non-linear XDP buffers
since we allow redirecting just into CPUMAPs, unlock XDP_REDIRECT for
S/G XDP buffer and rely on XDP stack to properly take care of the
frames.

Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
---
Changes since v1:
- drop Fixes tag
- unlock XDP_REDIRECT
- populate missing XDP metadata

Please note this patch is just compile tested
---

How would you like me to test this patch?

Ideally we'd remove this counter altogether. Nothing interesting to see.

 				rx_ring->stats.xdp_redirect++;
 			}
 		}
-- 
2.38.1

On Fri, Dec 09, 2022 at 01:19:44AM +0100, Lorenzo Bianconi wrote:

quoted

Even if full XDP_REDIRECT is not supported yet for non-linear XDP buffers
since we allow redirecting just into CPUMAPs, unlock XDP_REDIRECT for
S/G XDP buffer and rely on XDP stack to properly take care of the
frames.

Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
---
Changes since v1:
- drop Fixes tag
- unlock XDP_REDIRECT
- populate missing XDP metadata

Please note this patch is just compile tested
---

How would you like me to test this patch?

Hi Vladimir,

you can use xdp_redirect_cpu sample available in the kernel source tree.

quoted

 drivers/net/ethernet/freescale/enetc/enetc.c | 27 +++++++++-----------
 1 file changed, 12 insertions(+), 15 deletions(-)

diff --git a/drivers/net/ethernet/freescale/enetc/enetc.c b/drivers/net/ethernet/freescale/enetc/enetc.c
index 8671591cb750..9fd15e1e692d 100644
--- a/drivers/net/ethernet/freescale/enetc/enetc.c
+++ b/drivers/net/ethernet/freescale/enetc/enetc.c

@@ -1412,6 +1412,16 @@ static void enetc_add_rx_buff_to_xdp(struct enetc_bdr *rx_ring, int i,
 	/* To be used for XDP_TX */
 	rx_swbd->len = size;
 
+	if (!xdp_buff_has_frags(xdp_buff)) {
+		xdp_buff_set_frags_flag(xdp_buff);
+		shinfo->xdp_frags_size = size;
+	} else {
+		shinfo->xdp_frags_size += size;
+	}
+
+	if (page_is_pfmemalloc(rx_swbd->page))
+		xdp_buff_set_frag_pfmemalloc(xdp_buff);
+
 	skb_frag_off_set(frag, rx_swbd->page_offset);
 	skb_frag_size_set(frag, size);
 	__skb_frag_set_page(frag, rx_swbd->page);

@@ -1601,22 +1611,7 @@ static int enetc_clean_rx_ring_xdp(struct enetc_bdr *rx_ring,
 			}
 			break;
 		case XDP_REDIRECT:
-			/* xdp_return_frame does not support S/G in the sense
-			 * that it leaks the fragments (__xdp_return should not
-			 * call page_frag_free only for the initial buffer).
-			 * Until XDP_REDIRECT gains support for S/G let's keep
-			 * the code structure in place, but dead. We drop the
-			 * S/G frames ourselves to avoid memory leaks which
-			 * would otherwise leave the kernel OOM.
-			 */
-			if (unlikely(cleaned_cnt - orig_cleaned_cnt != 1)) {
-				enetc_xdp_drop(rx_ring, orig_i, i);
-				rx_ring->stats.xdp_redirect_sg++;
-				break;
-			}
-
 			tmp_orig_i = orig_i;
-
 			while (orig_i != i) {
 				enetc_flip_rx_buff(rx_ring,
 						   &rx_ring->rx_swbd[orig_i]);

@@ -1628,6 +1623,8 @@ static int enetc_clean_rx_ring_xdp(struct enetc_bdr *rx_ring,
 				enetc_xdp_free(rx_ring, tmp_orig_i, i);
 			} else {
 				xdp_redirect_frm_cnt++;
+				if (xdp_buff_has_frags(&xdp_buff))
+					rx_ring->stats.xdp_redirect_sg++;

Ideally we'd remove this counter altogether. Nothing interesting to see.

ok, I will get rid of it in the next version.

Regards,
Lorenzo

quoted

 				rx_ring->stats.xdp_redirect++;
 			}
 		}
-- 
2.38.1