On Wed, Nov 12, 2025 at 5:11 AM Aditya Garg
<[email protected]> wrote:
The MANA hardware supports a maximum of 30 scatter-gather entries (SGEs)
per TX WQE. Exceeding this limit can cause TX failures.
Add ndo_features_check() callback to validate SKB layout before
transmission. For GSO SKBs that would exceed the hardware SGE limit, clear
NETIF_F_GSO_MASK to enforce software segmentation in the stack.
Add a fallback in mana_start_xmit() to linearize non-GSO SKBs that still
exceed the SGE limit.
Also, Add ethtool counter for SKBs linearized
Co-developed-by: Dipayaan Roy <[email protected]>
Signed-off-by: Dipayaan Roy <[email protected]>
Signed-off-by: Aditya Garg <[email protected]>
---
drivers/net/ethernet/microsoft/mana/mana_en.c | 37 ++++++++++++++++++-
.../ethernet/microsoft/mana/mana_ethtool.c | 2 +
include/net/mana/gdma.h | 6 ++-
include/net/mana/mana.h | 1 +
4 files changed, 43 insertions(+), 3 deletions(-)
diff --git a/drivers/net/ethernet/microsoft/mana/mana_en.c
b/drivers/net/ethernet/microsoft/mana/mana_en.c
index cccd5b63cee6..67ae5421f9ee 100644
--- a/drivers/net/ethernet/microsoft/mana/mana_en.c
+++ b/drivers/net/ethernet/microsoft/mana/mana_en.c
@@ -11,6 +11,7 @@
#include <linux/mm.h>
#include <linux/pci.h>
#include <linux/export.h>
+#include <linux/skbuff.h>
#include <net/checksum.h>
#include <net/ip6_checksum.h>
@@ -329,6 +330,20 @@ netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct
net_device *ndev)
cq = &apc->tx_qp[txq_idx].tx_cq;
tx_stats = &txq->stats;
+ if (MAX_SKB_FRAGS + 2 > MAX_TX_WQE_SGL_ENTRIES &&
+ skb_shinfo(skb)->nr_frags + 2 > MAX_TX_WQE_SGL_ENTRIES) {
+ /* GSO skb with Hardware SGE limit exceeded is not expected here
+ * as they are handled in mana_features_check() callback
+ */
+ if (skb_linearize(skb)) {
+ netdev_warn_once(ndev, "Failed to linearize skb with
nr_frags=%d and is_gso=%d\n",
+ skb_shinfo(skb)->nr_frags,
+ skb_is_gso(skb));
+ goto tx_drop_count;
+ }
+ apc->eth_stats.linear_pkt_tx_cnt++;
+ }
+
pkg.tx_oob.s_oob.vcq_num = cq->gdma_id;
pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame;
@@ -442,8 +457,6 @@ netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct
net_device *ndev)
}
}
- WARN_ON_ONCE(pkg.wqe_req.num_sge > MAX_TX_WQE_SGL_ENTRIES);
-
if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) {
pkg.wqe_req.sgl = pkg.sgl_array;
} else {
@@ -518,6 +531,25 @@ netdev_tx_t mana_start_xmit(struct sk_buff *skb, struct
net_device *ndev)
return NETDEV_TX_OK;
}
#if MAX_SKB_FRAGS + 2 > MAX_TX_WQE_SGL_ENTRIES
+static netdev_features_t mana_features_check(struct sk_buff *skb,
+ struct net_device *ndev,
+ netdev_features_t features)
+{
+ if (MAX_SKB_FRAGS + 2 > MAX_TX_WQE_SGL_ENTRIES &&
+ skb_shinfo(skb)->nr_frags + 2 > MAX_TX_WQE_SGL_ENTRIES) {
+ /* Exceeds HW SGE limit.
+ * GSO case:
+ * Disable GSO so the stack will software-segment the skb
+ * into smaller skbs that fit the SGE budget.
+ * Non-GSO case:
+ * The xmit path will attempt skb_linearize() as a fallback.
+ */
+ if (skb_is_gso(skb))
No need to test skb_is_gso(skb), you can clear bits, this will be a
NOP if the packet is non GSO anyway.
+ features &= ~NETIF_F_GSO_MASK;
+ }
+ return features;
+}
#endif
+
static void mana_get_stats64(struct net_device *ndev,
struct rtnl_link_stats64 *st)
{
@@ -878,6 +910,7 @@ static const struct net_device_ops mana_devops = {
.ndo_open = mana_open,
.ndo_stop = mana_close,
.ndo_select_queue = mana_select_queue,
+ .ndo_features_check = mana_features_check,
Note that if your mana_features_check() is a nop if MAX_SKB_FRAGS is
small enough,
you could set a non NULL .ndo_features_check based on a preprocessor condition
#if MAX_SKB_FRAGS + 2 > MAX_TX_WQE_SGL_ENTRIES
.ndo_features_check = ....
#endif
This would avoid an expensive indirect call when possible.
.ndo_start_xmit = mana_start_xmit,
.ndo_validate_addr = eth_validate_addr,
.ndo_get_stats64 = mana_get_stats64,
diff --git a/drivers/net/ethernet/microsoft/mana/mana_ethtool.c
b/drivers/net/ethernet/microsoft/mana/mana_ethtool.c
index a1afa75a9463..fa5e1a2f06a9 100644
--- a/drivers/net/ethernet/microsoft/mana/mana_ethtool.c
+++ b/drivers/net/ethernet/microsoft/mana/mana_ethtool.c
@@ -71,6 +71,8 @@ static const struct mana_stats_desc mana_eth_stats[] = {
{"tx_cq_err", offsetof(struct mana_ethtool_stats, tx_cqe_err)},
{"tx_cqe_unknown_type", offsetof(struct mana_ethtool_stats,
tx_cqe_unknown_type)},
+ {"linear_pkt_tx_cnt", offsetof(struct mana_ethtool_stats,
+ linear_pkt_tx_cnt)},
{"rx_coalesced_err", offsetof(struct mana_ethtool_stats,
rx_coalesced_err)},
{"rx_cqe_unknown_type", offsetof(struct mana_ethtool_stats,
diff --git a/include/net/mana/gdma.h b/include/net/mana/gdma.h
index 637f42485dba..84614ebe0f4c 100644
--- a/include/net/mana/gdma.h
+++ b/include/net/mana/gdma.h
@@ -592,6 +592,9 @@ enum {
#define GDMA_DRV_CAP_FLAG_1_HANDLE_RECONFIG_EQE BIT(17)
#define GDMA_DRV_CAP_FLAG_1_HW_VPORT_LINK_AWARE BIT(6)
+/* Driver supports linearizing the skb when num_sge exceeds hardware limit */
+#define GDMA_DRV_CAP_FLAG_1_SKB_LINEARIZE BIT(20)
+
#define GDMA_DRV_CAP_FLAGS1 \
(GDMA_DRV_CAP_FLAG_1_EQ_SHARING_MULTI_VPORT | \
GDMA_DRV_CAP_FLAG_1_NAPI_WKDONE_FIX | \
@@ -601,7 +604,8 @@ enum {
GDMA_DRV_CAP_FLAG_1_DYNAMIC_IRQ_ALLOC_SUPPORT | \
GDMA_DRV_CAP_FLAG_1_SELF_RESET_ON_EQE | \
GDMA_DRV_CAP_FLAG_1_HANDLE_RECONFIG_EQE | \
- GDMA_DRV_CAP_FLAG_1_HW_VPORT_LINK_AWARE)
+ GDMA_DRV_CAP_FLAG_1_HW_VPORT_LINK_AWARE | \
+ GDMA_DRV_CAP_FLAG_1_SKB_LINEARIZE)
#define GDMA_DRV_CAP_FLAGS2 0
diff --git a/include/net/mana/mana.h b/include/net/mana/mana.h
index 8906901535f5..50a532fb30d6 100644
--- a/include/net/mana/mana.h
+++ b/include/net/mana/mana.h
@@ -404,6 +404,7 @@ struct mana_ethtool_stats {
u64 hc_tx_err_gdma;
u64 tx_cqe_err;
u64 tx_cqe_unknown_type;
+ u64 linear_pkt_tx_cnt;
u64 rx_coalesced_err;
u64 rx_cqe_unknown_type;
};
--
2.43.0
