[PATCH 00/13] Wangxun fixes

[PATCH 00/13] Wangxun fixes

[Jiawen Wu]

Fix some bugs in txgbe/ngbe, and optimize the Tx flow.

Jiawen Wu (13):
  net/txgbe: fix swfw mbox failure
  net/txgbe: fix VF-PF mbox interrupt
  net/txgbe: remove outer UDP checksum capability
  net/txgbe: fix driver load bit to inform firmware
  net/txgbe: enable Tx descriptor error interrupt
  net/txgbe: check length of Tx packets
  net/txgbe: add Tx descriptor error statistics
  net/ngbe: check length of Tx packets
  net/ngbe: add Tx descriptor error statistics
  net/ngbe: fix driver load bit to inform firmware
  net/ngbe: reconfigure more MAC Rx registers
  net/ngbe: fix interrupt lost in legacy or MSI mode
  net/ngbe: restrict configuration of VLAN strip offload

 drivers/net/ngbe/base/ngbe_regs.h       |   2 +
 drivers/net/ngbe/ngbe_ethdev.c          |  79 +++++++++-----
 drivers/net/ngbe/ngbe_rxtx.c            |  35 ++++++
 drivers/net/ngbe/ngbe_rxtx.h            |   1 +
 drivers/net/ngbe/ngbe_rxtx_vec_neon.c   |  11 +-
 drivers/net/ngbe/ngbe_rxtx_vec_sse.c    |  11 +-
 drivers/net/txgbe/base/txgbe_mng.c      |   1 +
 drivers/net/txgbe/base/txgbe_regs.h     |   8 +-
 drivers/net/txgbe/txgbe_ethdev.c        |  75 +++++++++++++
 drivers/net/txgbe/txgbe_ethdev.h        |   3 +
 drivers/net/txgbe/txgbe_rxtx.c          | 139 +++++++++++++++++++++++-
 drivers/net/txgbe/txgbe_rxtx.h          |   2 +
 drivers/net/txgbe/txgbe_rxtx_vec_neon.c |  11 +-
 drivers/net/txgbe/txgbe_rxtx_vec_sse.c  |  11 +-
 14 files changed, 345 insertions(+), 44 deletions(-)

-- 
2.27.0

[PATCH 01/13] net/txgbe: fix swfw mbox failure

[Jiawen Wu]

There is a unknown bug that the register TXGBE_MNGMBX cannot be written
in the loop, when DPDK is built with GCC high version. Access any register
before write TXGBE_MNGMBX can fix it.

Bugzilla ID: 1531
Fixes: 35c90ecccfd4 ("net/txgbe: add EEPROM functions")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/base/txgbe_mng.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/net/txgbe/base/txgbe_mng.c b/drivers/net/txgbe/base/txgbe_mng.c
index 20db982891..7dc8f21183 100644
--- a/drivers/net/txgbe/base/txgbe_mng.c
+++ b/drivers/net/txgbe/base/txgbe_mng.c
@@ -58,6 +58,7 @@ txgbe_hic_unlocked(struct txgbe_hw *hw, u32 *buffer, u32 length, u32 timeout)
 
 	dword_len = length >> 2;
 
+	txgbe_flush(hw);
 	/* The device driver writes the relevant command block
 	 * into the ram area.
 	 */
-- 
2.27.0

[PATCH 02/13] net/txgbe: fix VF-PF mbox interrupt

[Jiawen Wu]

There was a incorrect bit to define TXGBE_ICRMISC_VFMBX that prevents the
interrupt from being handled correctly.

Fixes: a6712cd029a4 ("net/txgbe: add PF module init and uninit for SRIOV")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/base/txgbe_regs.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/txgbe/base/txgbe_regs.h b/drivers/net/txgbe/base/txgbe_regs.h
index 4ea4a2e3d8..b46d65331e 100644
--- a/drivers/net/txgbe/base/txgbe_regs.h
+++ b/drivers/net/txgbe/base/txgbe_regs.h
@@ -1197,7 +1197,7 @@ enum txgbe_5tuple_protocol {
 #define   TXGBE_ICRMISC_ANDONE MS(19, 0x1) /* link auto-nego done */
 #define   TXGBE_ICRMISC_ERRIG  MS(20, 0x1) /* integrity error */
 #define   TXGBE_ICRMISC_SPI    MS(21, 0x1) /* SPI interface */
-#define   TXGBE_ICRMISC_VFMBX  MS(22, 0x1) /* VF-PF message box */
+#define   TXGBE_ICRMISC_VFMBX  MS(23, 0x1) /* VF-PF message box */
 #define   TXGBE_ICRMISC_GPIO   MS(26, 0x1) /* GPIO interrupt */
 #define   TXGBE_ICRMISC_ERRPCI MS(27, 0x1) /* pcie request error */
 #define   TXGBE_ICRMISC_HEAT   MS(28, 0x1) /* overheat detection */
-- 
2.27.0

[PATCH 03/13] net/txgbe: remove outer UDP checksum capability

[Jiawen Wu]

The hardware does not support outer UDP checksum for tunnel packets.
It's wrong to claim this Tx offload capability, so fix it.

Bugzilla ID: 1529
Fixes: b950203be7f1 ("net/txgbe: support VXLAN-GPE")
Fixes: 295968d17407 ("ethdev: add namespace")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/txgbe_rxtx.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
index 5bc0f8772f..c12726553c 100644
--- a/drivers/net/txgbe/txgbe_rxtx.c
+++ b/drivers/net/txgbe/txgbe_rxtx.c
@@ -2284,8 +2284,7 @@ txgbe_get_tx_port_offloads(struct rte_eth_dev *dev)
 
 	tx_offload_capa |= RTE_ETH_TX_OFFLOAD_MACSEC_INSERT;
 
-	tx_offload_capa |= RTE_ETH_TX_OFFLOAD_OUTER_IPV4_CKSUM |
-			   RTE_ETH_TX_OFFLOAD_OUTER_UDP_CKSUM;
+	tx_offload_capa |= RTE_ETH_TX_OFFLOAD_OUTER_IPV4_CKSUM;
 
 #ifdef RTE_LIB_SECURITY
 	if (dev->security_ctx)
-- 
2.27.0

[PATCH 04/13] net/txgbe: fix driver load bit to inform firmware

[Jiawen Wu]

Drv_load bit will be reset to default 0 after hardware LAN reset,
reconfigure it to inform firmware that driver is loaded. And set it to 0
when device is closed.

Fixes: b1f596677d8e ("net/txgbe: support device start")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/txgbe_ethdev.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/net/txgbe/txgbe_ethdev.c b/drivers/net/txgbe/txgbe_ethdev.c
index 2834468764..4aa3bfd0bc 100644
--- a/drivers/net/txgbe/txgbe_ethdev.c
+++ b/drivers/net/txgbe/txgbe_ethdev.c
@@ -331,6 +331,8 @@ txgbe_pf_reset_hw(struct txgbe_hw *hw)
 	status = hw->mac.reset_hw(hw);
 
 	ctrl_ext = rd32(hw, TXGBE_PORTCTL);
+	/* let hardware know driver is loaded */
+	ctrl_ext |= TXGBE_PORTCTL_DRVLOAD;
 	/* Set PF Reset Done bit so PF/VF Mail Ops can work */
 	ctrl_ext |= TXGBE_PORTCTL_RSTDONE;
 	wr32(hw, TXGBE_PORTCTL, ctrl_ext);
@@ -2061,6 +2063,9 @@ txgbe_dev_close(struct rte_eth_dev *dev)
 
 	ret = txgbe_dev_stop(dev);
 
+	/* Let firmware take over control of hardware */
+	wr32m(hw, TXGBE_PORTCTL, TXGBE_PORTCTL_DRVLOAD, 0);
+
 	txgbe_dev_free_queues(dev);
 
 	txgbe_set_pcie_master(hw, false);
-- 
2.27.0

[PATCH 05/13] net/txgbe: enable Tx descriptor error interrupt

[Jiawen Wu]

Enable Tdm_desc_chk then handle the interrupt of TDM desc error. If it's a
non-fatal error, clear the error to reset the queue. If it's a fatal error,
require users to manually restart the port. This flow prevents the hardware
from PCIe pending due to Tx hang, resulting in a reboot to recover.

But remarkably, when packet layer length does not match the packet type in
TX descriptor, it will cause non-fatal error if Tdm_desc_chk is enabled.
But it can be transmitted normally if Tdm_desc_chk is disabled. So in order
to prevent hardware over check, fix the layer length on the basis of packet
type.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/base/txgbe_regs.h |   6 ++
 drivers/net/txgbe/txgbe_ethdev.c    |  64 ++++++++++++++++++
 drivers/net/txgbe/txgbe_ethdev.h    |   3 +
 drivers/net/txgbe/txgbe_rxtx.c      | 100 ++++++++++++++++++++++++++++
 drivers/net/txgbe/txgbe_rxtx.h      |   1 +
 5 files changed, 174 insertions(+)

diff --git a/drivers/net/txgbe/base/txgbe_regs.h b/drivers/net/txgbe/base/txgbe_regs.h
index b46d65331e..7a9ba6976f 100644
--- a/drivers/net/txgbe/base/txgbe_regs.h
+++ b/drivers/net/txgbe/base/txgbe_regs.h
@@ -1197,6 +1197,7 @@ enum txgbe_5tuple_protocol {
 #define   TXGBE_ICRMISC_ANDONE MS(19, 0x1) /* link auto-nego done */
 #define   TXGBE_ICRMISC_ERRIG  MS(20, 0x1) /* integrity error */
 #define   TXGBE_ICRMISC_SPI    MS(21, 0x1) /* SPI interface */
+#define   TXGBE_ICRMISC_TXDESC MS(22, 0x1) /* TDM desc error */
 #define   TXGBE_ICRMISC_VFMBX  MS(23, 0x1) /* VF-PF message box */
 #define   TXGBE_ICRMISC_GPIO   MS(26, 0x1) /* GPIO interrupt */
 #define   TXGBE_ICRMISC_ERRPCI MS(27, 0x1) /* pcie request error */
@@ -1382,6 +1383,11 @@ enum txgbe_5tuple_protocol {
 #define   TXGBE_TXCFG_WTHRESH(v)        LS(v, 16, 0x7F)
 #define   TXGBE_TXCFG_FLUSH             MS(26, 0x1)
 
+#define TXGBE_TDM_DESC_CHK(i)		(0x0180B0 + (i) * 4) /*0-3*/
+#define TXGBE_TDM_DESC_NONFATAL(i)	(0x0180C0 + (i) * 4) /*0-3*/
+#define TXGBE_TDM_DESC_FATAL(i)		(0x0180D0 + (i) * 4) /*0-3*/
+#define TXGBE_TDM_DESC_MASK(v)		MS(v, 0x1)
+
 /* interrupt registers */
 #define TXGBE_ITRI                      0x000180
 #define TXGBE_ITR(i)                    (0x000200 + 4 * (i))
diff --git a/drivers/net/txgbe/txgbe_ethdev.c b/drivers/net/txgbe/txgbe_ethdev.c
index 4aa3bfd0bc..bafa9cf829 100644
--- a/drivers/net/txgbe/txgbe_ethdev.c
+++ b/drivers/net/txgbe/txgbe_ethdev.c
@@ -1936,6 +1936,7 @@ txgbe_dev_stop(struct rte_eth_dev *dev)
 	PMD_INIT_FUNC_TRACE();
 
 	rte_eal_alarm_cancel(txgbe_dev_detect_sfp, dev);
+	rte_eal_alarm_cancel(txgbe_tx_queue_clear_error, dev);
 	txgbe_dev_wait_setup_link_complete(dev, 0);
 
 	/* disable interrupts */
@@ -2838,6 +2839,60 @@ txgbe_dev_setup_link_alarm_handler(void *param)
 	intr->flags &= ~TXGBE_FLAG_NEED_LINK_CONFIG;
 }
 
+static void
+txgbe_do_reset(struct rte_eth_dev *dev)
+{
+	struct txgbe_hw *hw = TXGBE_DEV_HW(dev);
+	struct txgbe_tx_queue *txq;
+	u32 i;
+
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		txq->resetting = true;
+	}
+
+	rte_delay_ms(1);
+	wr32(hw, TXGBE_RST, TXGBE_RST_LAN(hw->bus.lan_id));
+	txgbe_flush(hw);
+
+	PMD_DRV_LOG(ERR, "Please manually restart the port %d",
+		dev->data->port_id);
+}
+
+static void
+txgbe_tx_ring_recovery(struct rte_eth_dev *dev)
+{
+	struct txgbe_hw *hw = TXGBE_DEV_HW(dev);
+	u32 desc_error[4] = {0, 0, 0, 0};
+	struct txgbe_tx_queue *txq;
+	u32 i;
+
+	/* check tdm fatal error */
+	for (i = 0; i < 4; i++) {
+		desc_error[i] = rd32(hw, TXGBE_TDM_DESC_FATAL(i));
+		if (desc_error[i] != 0) {
+			PMD_DRV_LOG(ERR, "TDM fatal error reg[%d]: 0x%x", i, desc_error[i]);
+			txgbe_do_reset(dev);
+			return;
+		}
+	}
+
+	/* check tdm non-fatal error */
+	for (i = 0; i < 4; i++)
+		desc_error[i] = rd32(hw, TXGBE_TDM_DESC_NONFATAL(i));
+
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		if (desc_error[i / 32] & (1 << i % 32)) {
+			PMD_DRV_LOG(ERR, "TDM non-fatal error, reset port[%d] queue[%d]",
+				dev->data->port_id, i);
+			dev->data->tx_queue_state[i] = RTE_ETH_QUEUE_STATE_STOPPED;
+			txq = dev->data->tx_queues[i];
+			txq->resetting = true;
+			rte_eal_alarm_set(1000, txgbe_tx_queue_clear_error, (void *)dev);
+		}
+	}
+}
+
 /*
  * If @timeout_ms was 0, it means that it will not return until link complete.
  * It returns 1 on complete, return 0 on timeout.
@@ -3096,6 +3151,7 @@ txgbe_dev_misc_interrupt_setup(struct rte_eth_dev *dev)
 	intr->mask |= mask;
 	intr->mask_misc |= TXGBE_ICRMISC_GPIO;
 	intr->mask_misc |= TXGBE_ICRMISC_ANDONE;
+	intr->mask_misc |= TXGBE_ICRMISC_TXDESC;
 	return 0;
 }
 
@@ -3191,6 +3247,9 @@ txgbe_dev_interrupt_get_status(struct rte_eth_dev *dev,
 	if (eicr & TXGBE_ICRMISC_HEAT)
 		intr->flags |= TXGBE_FLAG_OVERHEAT;
 
+	if (eicr & TXGBE_ICRMISC_TXDESC)
+		intr->flags |= TXGBE_FLAG_TX_DESC_ERR;
+
 	((u32 *)hw->isb_mem)[TXGBE_ISB_MISC] = 0;
 
 	return 0;
@@ -3310,6 +3369,11 @@ txgbe_dev_interrupt_action(struct rte_eth_dev *dev,
 		intr->flags &= ~TXGBE_FLAG_OVERHEAT;
 	}
 
+	if (intr->flags & TXGBE_FLAG_TX_DESC_ERR) {
+		txgbe_tx_ring_recovery(dev);
+		intr->flags &= ~TXGBE_FLAG_TX_DESC_ERR;
+	}
+
 	PMD_DRV_LOG(DEBUG, "enable intr immediately");
 	txgbe_enable_intr(dev);
 	rte_intr_enable(intr_handle);
diff --git a/drivers/net/txgbe/txgbe_ethdev.h b/drivers/net/txgbe/txgbe_ethdev.h
index f0f4ced5b0..302ea9f037 100644
--- a/drivers/net/txgbe/txgbe_ethdev.h
+++ b/drivers/net/txgbe/txgbe_ethdev.h
@@ -31,6 +31,7 @@
 #define TXGBE_FLAG_NEED_LINK_CONFIG (uint32_t)(1 << 4)
 #define TXGBE_FLAG_NEED_AN_CONFIG   (uint32_t)(1 << 5)
 #define TXGBE_FLAG_OVERHEAT         (uint32_t)(1 << 6)
+#define TXGBE_FLAG_TX_DESC_ERR      (uint32_t)(1 << 7)
 
 /*
  * Defines that were not part of txgbe_type.h as they are not used by the
@@ -474,6 +475,8 @@ int txgbe_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id);
 
 int txgbe_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id);
 
+void txgbe_tx_queue_clear_error(void *param);
+
 void txgbe_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
 	struct rte_eth_rxq_info *qinfo);
 
diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
index c12726553c..2d2b437643 100644
--- a/drivers/net/txgbe/txgbe_rxtx.c
+++ b/drivers/net/txgbe/txgbe_rxtx.c
@@ -728,6 +728,66 @@ txgbe_get_tun_len(struct rte_mbuf *mbuf)
 	return tun_len;
 }
 
+static inline void
+txgbe_fix_offload_len(union txgbe_tx_offload *ol)
+{
+	uint8_t ptid = ol->ptid;
+
+	if (ptid & TXGBE_PTID_PKT_TUN) {
+		if (ol->outer_l2_len == 0)
+			ol->outer_l2_len = sizeof(struct rte_ether_hdr);
+		if (ol->outer_l3_len == 0) {
+			if (ptid & TXGBE_PTID_TUN_IPV6)
+				ol->outer_l3_len = sizeof(struct rte_ipv6_hdr);
+			else
+				ol->outer_l3_len = sizeof(struct rte_ipv4_hdr);
+		}
+		if ((ptid & 0xF) == 0) {
+			ol->l3_len = 0;
+			ol->l4_len = 0;
+		} else {
+			goto inner;
+		}
+	}
+
+	if ((ptid & 0xF0) == TXGBE_PTID_PKT_MAC) {
+		if (ol->l2_len == 0)
+			ol->l2_len = sizeof(struct rte_ether_hdr);
+		ol->l3_len = 0;
+		ol->l4_len = 0;
+	} else if ((ptid & 0xF0) == TXGBE_PTID_PKT_IP) {
+		if (ol->l2_len == 0)
+			ol->l2_len = sizeof(struct rte_ether_hdr);
+inner:
+		if (ol->l3_len == 0) {
+			if (ptid & TXGBE_PTID_PKT_IPV6)
+				ol->l3_len = sizeof(struct rte_ipv6_hdr);
+			else
+				ol->l3_len = sizeof(struct rte_ipv4_hdr);
+		}
+		switch (ptid & 0x7) {
+		case 0x1:
+		case 0x2:
+			ol->l4_len = 0;
+			break;
+		case 0x3:
+			if (ol->l4_len == 0)
+				ol->l4_len =  sizeof(struct rte_udp_hdr);
+			break;
+		case 0x4:
+			if (ol->l4_len == 0)
+				ol->l4_len =  sizeof(struct rte_tcp_hdr);
+			break;
+		case 0x5:
+			if (ol->l4_len == 0)
+				ol->l4_len =  sizeof(struct rte_sctp_hdr);
+			break;
+		default:
+			break;
+		}
+	}
+}
+
 static inline uint8_t
 txgbe_parse_tun_ptid(struct rte_mbuf *tx_pkt, uint8_t tun_len)
 {
@@ -782,6 +842,10 @@ txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 	uint8_t use_ipsec;
 #endif
 
+	txq = tx_queue;
+	if (txq->resetting)
+		return 0;
+
 	tx_offload.data[0] = 0;
 	tx_offload.data[1] = 0;
 	txq = tx_queue;
@@ -826,6 +890,7 @@ txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 			if (tx_offload.ptid & TXGBE_PTID_PKT_TUN)
 				tx_offload.ptid |= txgbe_parse_tun_ptid(tx_pkt,
 							tx_offload.outer_tun_len);
+			txgbe_fix_offload_len(&tx_offload);
 
 #ifdef RTE_LIB_SECURITY
 			if (use_ipsec) {
@@ -4570,6 +4635,11 @@ txgbe_dev_tx_init(struct rte_eth_dev *dev)
 		wr32(hw, TXGBE_TXWP(txq->reg_idx), 0);
 	}
 
+#ifndef RTE_LIB_SECURITY
+	for (i = 0; i < 4; i++)
+		wr32(hw, TXGBE_TDM_DESC_CHK(i), 0xFFFFFFFF);
+#endif
+
 	/* Device configured with multiple TX queues. */
 	txgbe_dev_mq_tx_configure(dev);
 }
@@ -4806,6 +4876,7 @@ txgbe_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
 	rte_wmb();
 	wr32(hw, TXGBE_TXWP(txq->reg_idx), txq->tx_tail);
 	dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
+	txq->resetting = false;
 
 	return 0;
 }
@@ -4863,6 +4934,35 @@ txgbe_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
 	return 0;
 }
 
+void
+txgbe_tx_queue_clear_error(void *param)
+{
+	struct rte_eth_dev *dev = (struct rte_eth_dev *)param;
+	struct txgbe_hw *hw = TXGBE_DEV_HW(dev);
+	struct txgbe_tx_queue *txq;
+	u32 i;
+
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		if (!txq->resetting)
+			continue;
+
+		txgbe_dev_save_tx_queue(hw, i);
+
+		/* tx ring reset */
+		wr32(hw, TXGBE_TDM_DESC_NONFATAL(i / 32),
+			TXGBE_TDM_DESC_MASK(i % 32));
+
+		if (txq->ops != NULL) {
+			txq->ops->release_mbufs(txq);
+			txq->ops->reset(txq);
+		}
+
+		txgbe_dev_store_tx_queue(hw, i);
+		txgbe_dev_tx_queue_start(dev, i);
+	}
+}
+
 void
 txgbe_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
 	struct rte_eth_rxq_info *qinfo)
diff --git a/drivers/net/txgbe/txgbe_rxtx.h b/drivers/net/txgbe/txgbe_rxtx.h
index 9155eb1f70..e668b60b1e 100644
--- a/drivers/net/txgbe/txgbe_rxtx.h
+++ b/drivers/net/txgbe/txgbe_rxtx.h
@@ -412,6 +412,7 @@ struct txgbe_tx_queue {
 	/**< indicates that IPsec TX feature is in use */
 #endif
 	const struct rte_memzone *mz;
+	bool		    resetting;
 };
 
 struct txgbe_txq_ops {
-- 
2.27.0

[PATCH 06/13] net/txgbe: check length of Tx packets

[Jiawen Wu]

Add checking of the Tx packet length to avoid TDM fatal error as far as
possible. Set the pkt_len=1518 for invalid packet in simple Tx code path,
and drop it directly in featured Tx code path.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/txgbe_rxtx.c          | 33 +++++++++++++++++++++++++
 drivers/net/txgbe/txgbe_rxtx_vec_neon.c | 11 ++++++---
 drivers/net/txgbe/txgbe_rxtx_vec_sse.c  | 11 ++++++---
 3 files changed, 49 insertions(+), 6 deletions(-)

diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
index 2d2b437643..06acbd0881 100644
--- a/drivers/net/txgbe/txgbe_rxtx.c
+++ b/drivers/net/txgbe/txgbe_rxtx.c
@@ -160,6 +160,8 @@ tx4(volatile struct txgbe_tx_desc *txdp, struct rte_mbuf **pkts)
 	for (i = 0; i < 4; ++i, ++txdp, ++pkts) {
 		buf_dma_addr = rte_mbuf_data_iova(*pkts);
 		pkt_len = (*pkts)->data_len;
+		if (pkt_len < RTE_ETHER_HDR_LEN)
+			pkt_len = TXGBE_FRAME_SIZE_DFT;
 
 		/* write data to descriptor */
 		txdp->qw0 = rte_cpu_to_le_64(buf_dma_addr);
@@ -180,6 +182,8 @@ tx1(volatile struct txgbe_tx_desc *txdp, struct rte_mbuf **pkts)
 
 	buf_dma_addr = rte_mbuf_data_iova(*pkts);
 	pkt_len = (*pkts)->data_len;
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = TXGBE_FRAME_SIZE_DFT;
 
 	/* write data to descriptor */
 	txdp->qw0 = cpu_to_le64(buf_dma_addr);
@@ -813,6 +817,30 @@ txgbe_parse_tun_ptid(struct rte_mbuf *tx_pkt, uint8_t tun_len)
 	return ptid;
 }
 
+static inline bool
+txgbe_check_pkt_err(struct rte_mbuf *tx_pkt)
+{
+	uint32_t total_len = 0, nb_seg = 0;
+	struct rte_mbuf *mseg;
+
+	mseg = tx_pkt;
+	do {
+		if (mseg->data_len == 0)
+			return true;
+		total_len += mseg->data_len;
+		nb_seg++;
+		mseg = mseg->next;
+	} while (mseg != NULL);
+
+	if (tx_pkt->pkt_len != total_len || tx_pkt->pkt_len == 0)
+		return true;
+
+	if (tx_pkt->nb_segs != nb_seg || tx_pkt->nb_segs > 64)
+		return true;
+
+	return false;
+}
+
 uint16_t
 txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 		uint16_t nb_pkts)
@@ -864,6 +892,11 @@ txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
 		new_ctx = 0;
 		tx_pkt = *tx_pkts++;
+		if (txgbe_check_pkt_err(tx_pkt)) {
+			rte_pktmbuf_free(tx_pkt);
+			continue;
+		}
+
 		pkt_len = tx_pkt->pkt_len;
 
 		/*
diff --git a/drivers/net/txgbe/txgbe_rxtx_vec_neon.c b/drivers/net/txgbe/txgbe_rxtx_vec_neon.c
index a96baf9b1d..14be7886ef 100644
--- a/drivers/net/txgbe/txgbe_rxtx_vec_neon.c
+++ b/drivers/net/txgbe/txgbe_rxtx_vec_neon.c
@@ -476,9 +476,14 @@ static inline void
 vtx1(volatile struct txgbe_tx_desc *txdp,
 		struct rte_mbuf *pkt, uint64_t flags)
 {
-	uint64x2_t descriptor = {
-			pkt->buf_iova + pkt->data_off,
-			(uint64_t)pkt->pkt_len << 45 | flags | pkt->data_len};
+	uint16_t pkt_len = pkt->data_len;
+	uint64x2_t descriptor;
+
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = TXGBE_FRAME_SIZE_DFT;
+
+	descriptor = {pkt->buf_iova + pkt->data_off,
+			(uint64_t)pkt_len << 45 | flags | pkt_len};
 
 	vst1q_u64((uint64_t *)(uintptr_t)txdp, descriptor);
 }
diff --git a/drivers/net/txgbe/txgbe_rxtx_vec_sse.c b/drivers/net/txgbe/txgbe_rxtx_vec_sse.c
index 1a3f2ce3cd..8ecce33471 100644
--- a/drivers/net/txgbe/txgbe_rxtx_vec_sse.c
+++ b/drivers/net/txgbe/txgbe_rxtx_vec_sse.c
@@ -607,9 +607,14 @@ static inline void
 vtx1(volatile struct txgbe_tx_desc *txdp,
 		struct rte_mbuf *pkt, uint64_t flags)
 {
-	__m128i descriptor = _mm_set_epi64x((uint64_t)pkt->pkt_len << 45 |
-			flags | pkt->data_len,
-			pkt->buf_iova + pkt->data_off);
+	uint16_t pkt_len = pkt->data_len;
+	__m128i descriptor;
+
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = TXGBE_FRAME_SIZE_DFT;
+
+	descriptor = _mm_set_epi64x((uint64_t)pkt_len << 45 | flags | pkt_len,
+				pkt->buf_iova + pkt->data_off);
 	_mm_store_si128((__m128i *)(uintptr_t)txdp, descriptor);
 }
 
-- 
2.27.0

[PATCH 07/13] net/txgbe: add Tx descriptor error statistics

[Jiawen Wu]

Count the number of packets not sent due to Tx descriptor error.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/txgbe_ethdev.c | 6 ++++++
 drivers/net/txgbe/txgbe_rxtx.c   | 3 +++
 drivers/net/txgbe/txgbe_rxtx.h   | 1 +
 3 files changed, 10 insertions(+)

diff --git a/drivers/net/txgbe/txgbe_ethdev.c b/drivers/net/txgbe/txgbe_ethdev.c
index bafa9cf829..0c76e986f4 100644
--- a/drivers/net/txgbe/txgbe_ethdev.c
+++ b/drivers/net/txgbe/txgbe_ethdev.c
@@ -2344,6 +2344,7 @@ txgbe_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 	struct txgbe_hw_stats *hw_stats = TXGBE_DEV_STATS(dev);
 	struct txgbe_stat_mappings *stat_mappings =
 			TXGBE_DEV_STAT_MAPPINGS(dev);
+	struct txgbe_tx_queue *txq;
 	uint32_t i, j;
 
 	txgbe_read_stats_registers(hw, hw_stats);
@@ -2398,6 +2399,11 @@ txgbe_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 
 	/* Tx Errors */
 	stats->oerrors  = 0;
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		stats->oerrors += txq->desc_error;
+	}
+
 	return 0;
 }
 
diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
index 06acbd0881..fc9e7b14f5 100644
--- a/drivers/net/txgbe/txgbe_rxtx.c
+++ b/drivers/net/txgbe/txgbe_rxtx.c
@@ -894,6 +894,7 @@ txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 		tx_pkt = *tx_pkts++;
 		if (txgbe_check_pkt_err(tx_pkt)) {
 			rte_pktmbuf_free(tx_pkt);
+			txq->desc_error++;
 			continue;
 		}
 
@@ -2523,6 +2524,7 @@ txgbe_dev_tx_queue_setup(struct rte_eth_dev *dev,
 	txgbe_set_tx_function(dev, txq);
 
 	txq->ops->reset(txq);
+	txq->desc_error = 0;
 
 	dev->data->tx_queues[queue_idx] = txq;
 
@@ -4980,6 +4982,7 @@ txgbe_tx_queue_clear_error(void *param)
 		if (!txq->resetting)
 			continue;
 
+		txq->desc_error++;
 		txgbe_dev_save_tx_queue(hw, i);
 
 		/* tx ring reset */
diff --git a/drivers/net/txgbe/txgbe_rxtx.h b/drivers/net/txgbe/txgbe_rxtx.h
index e668b60b1e..622a0d3981 100644
--- a/drivers/net/txgbe/txgbe_rxtx.h
+++ b/drivers/net/txgbe/txgbe_rxtx.h
@@ -412,6 +412,7 @@ struct txgbe_tx_queue {
 	/**< indicates that IPsec TX feature is in use */
 #endif
 	const struct rte_memzone *mz;
+	uint64_t	    desc_error;
 	bool		    resetting;
 };
 
-- 
2.27.0

[PATCH 08/13] net/ngbe: check length of Tx packets

[Jiawen Wu]

Add checking of the Tx packet length to avoid TDM fatal error as far as
possible. Set the pkt_len=1518 for invalid packet in simple Tx code path,
and drop it directly in featured Tx code path. Althrough the hardware
does not support TDM desc check.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_rxtx.c          | 33 +++++++++++++++++++++++++++
 drivers/net/ngbe/ngbe_rxtx_vec_neon.c | 11 ++++++---
 drivers/net/ngbe/ngbe_rxtx_vec_sse.c  | 11 ++++++---
 3 files changed, 49 insertions(+), 6 deletions(-)

diff --git a/drivers/net/ngbe/ngbe_rxtx.c b/drivers/net/ngbe/ngbe_rxtx.c
index f3eb797d0c..25a314f10a 100644
--- a/drivers/net/ngbe/ngbe_rxtx.c
+++ b/drivers/net/ngbe/ngbe_rxtx.c
@@ -113,6 +113,8 @@ tx4(volatile struct ngbe_tx_desc *txdp, struct rte_mbuf **pkts)
 	for (i = 0; i < 4; ++i, ++txdp, ++pkts) {
 		buf_dma_addr = rte_mbuf_data_iova(*pkts);
 		pkt_len = (*pkts)->data_len;
+		if (pkt_len < RTE_ETHER_HDR_LEN)
+			pkt_len = NGBE_FRAME_SIZE_DFT;
 
 		/* write data to descriptor */
 		txdp->qw0 = rte_cpu_to_le_64(buf_dma_addr);
@@ -133,6 +135,8 @@ tx1(volatile struct ngbe_tx_desc *txdp, struct rte_mbuf **pkts)
 
 	buf_dma_addr = rte_mbuf_data_iova(*pkts);
 	pkt_len = (*pkts)->data_len;
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = NGBE_FRAME_SIZE_DFT;
 
 	/* write data to descriptor */
 	txdp->qw0 = cpu_to_le64(buf_dma_addr);
@@ -555,6 +559,30 @@ ngbe_xmit_cleanup(struct ngbe_tx_queue *txq)
 	return 0;
 }
 
+static inline bool
+ngbe_check_pkt_err(struct rte_mbuf *tx_pkt)
+{
+	uint32_t total_len = 0, nb_seg = 0;
+	struct rte_mbuf *mseg;
+
+	mseg = tx_pkt;
+	do {
+		if (mseg->data_len == 0)
+			return true;
+		total_len += mseg->data_len;
+		nb_seg++;
+		mseg = mseg->next;
+	} while (mseg != NULL);
+
+	if (tx_pkt->pkt_len != total_len || tx_pkt->pkt_len == 0)
+		return true;
+
+	if (tx_pkt->nb_segs != nb_seg || tx_pkt->nb_segs > 64)
+		return true;
+
+	return false;
+}
+
 uint16_t
 ngbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 		uint16_t nb_pkts)
@@ -599,6 +627,11 @@ ngbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
 		new_ctx = 0;
 		tx_pkt = *tx_pkts++;
+		if (ngbe_check_pkt_err(tx_pkt)) {
+			rte_pktmbuf_free(tx_pkt);
+			continue;
+		}
+
 		pkt_len = tx_pkt->pkt_len;
 
 		/*
diff --git a/drivers/net/ngbe/ngbe_rxtx_vec_neon.c b/drivers/net/ngbe/ngbe_rxtx_vec_neon.c
index dcf12b7070..d870191792 100644
--- a/drivers/net/ngbe/ngbe_rxtx_vec_neon.c
+++ b/drivers/net/ngbe/ngbe_rxtx_vec_neon.c
@@ -476,9 +476,14 @@ static inline void
 vtx1(volatile struct ngbe_tx_desc *txdp,
 		struct rte_mbuf *pkt, uint64_t flags)
 {
-	uint64x2_t descriptor = {
-			pkt->buf_iova + pkt->data_off,
-			(uint64_t)pkt->pkt_len << 45 | flags | pkt->data_len};
+	uint16_t pkt_len = pkt->data_len;
+	uint64x2_t descriptor;
+
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = NGBE_FRAME_SIZE_DFT;
+
+	descriptor = {pkt->buf_iova + pkt->data_off,
+			(uint64_t)pkt_len << 45 | flags | pkt_len};
 
 	vst1q_u64((uint64_t *)(uintptr_t)txdp, descriptor);
 }
diff --git a/drivers/net/ngbe/ngbe_rxtx_vec_sse.c b/drivers/net/ngbe/ngbe_rxtx_vec_sse.c
index b128bd3a67..19c69cdfa6 100644
--- a/drivers/net/ngbe/ngbe_rxtx_vec_sse.c
+++ b/drivers/net/ngbe/ngbe_rxtx_vec_sse.c
@@ -563,9 +563,14 @@ static inline void
 vtx1(volatile struct ngbe_tx_desc *txdp,
 		struct rte_mbuf *pkt, uint64_t flags)
 {
-	__m128i descriptor = _mm_set_epi64x((uint64_t)pkt->pkt_len << 45 |
-			flags | pkt->data_len,
-			pkt->buf_iova + pkt->data_off);
+	uint16_t pkt_len = pkt->data_len;
+	__m128i descriptor;
+
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = NGBE_FRAME_SIZE_DFT;
+
+	descriptor = _mm_set_epi64x((uint64_t)pkt_len << 45 | flags | pkt_len,
+				pkt->buf_iova + pkt->data_off);
 	_mm_store_si128((__m128i *)(uintptr_t)txdp, descriptor);
 }
 
-- 
2.27.0

[PATCH 09/13] net/ngbe: add Tx descriptor error statistics

[Jiawen Wu]

Count the number of packets not sent due to Tx descriptor error.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_ethdev.c | 6 ++++++
 drivers/net/ngbe/ngbe_rxtx.c   | 2 ++
 drivers/net/ngbe/ngbe_rxtx.h   | 1 +
 3 files changed, 9 insertions(+)

diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index 6c45ffaad3..d9d2daf656 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -1507,6 +1507,7 @@ ngbe_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 	struct ngbe_hw_stats *hw_stats = NGBE_DEV_STATS(dev);
 	struct ngbe_stat_mappings *stat_mappings =
 			NGBE_DEV_STAT_MAPPINGS(dev);
+	struct ngbe_tx_queue *txq;
 	uint32_t i, j;
 
 	ngbe_read_stats_registers(hw, hw_stats);
@@ -1559,6 +1560,11 @@ ngbe_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 
 	/* Tx Errors */
 	stats->oerrors  = 0;
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		stats->oerrors += txq->desc_error;
+	}
+
 	return 0;
 }
 
diff --git a/drivers/net/ngbe/ngbe_rxtx.c b/drivers/net/ngbe/ngbe_rxtx.c
index 25a314f10a..8d31d47de9 100644
--- a/drivers/net/ngbe/ngbe_rxtx.c
+++ b/drivers/net/ngbe/ngbe_rxtx.c
@@ -629,6 +629,7 @@ ngbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 		tx_pkt = *tx_pkts++;
 		if (ngbe_check_pkt_err(tx_pkt)) {
 			rte_pktmbuf_free(tx_pkt);
+			txq->desc_error++;
 			continue;
 		}
 
@@ -2100,6 +2101,7 @@ ngbe_dev_tx_queue_setup(struct rte_eth_dev *dev,
 	ngbe_set_tx_function(dev, txq);
 
 	txq->ops->reset(txq);
+	txq->desc_error = 0;
 
 	dev->data->tx_queues[queue_idx] = txq;
 
diff --git a/drivers/net/ngbe/ngbe_rxtx.h b/drivers/net/ngbe/ngbe_rxtx.h
index 7574db32d8..8534ec123a 100644
--- a/drivers/net/ngbe/ngbe_rxtx.h
+++ b/drivers/net/ngbe/ngbe_rxtx.h
@@ -375,6 +375,7 @@ struct ngbe_tx_queue {
 
 	const struct ngbe_txq_ops *ops;       /**< txq ops */
 	const struct rte_memzone *mz;
+	uint64_t            desc_error;
 };
 
 struct ngbe_txq_ops {
-- 
2.27.0

[PATCH 10/13] net/ngbe: fix driver load bit to inform firmware

[Jiawen Wu]

Drv_load bit will be reset to default 0 after hardware LAN reset,
reconfigure it to inform firmware that driver is loaded. And set it to 0
when device is closed.

Fixes: 3518df5774c7 ("net/ngbe: support device start/stop")
Fixes: cc63194e89cb ("net/ngbe: support close and reset device")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_ethdev.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index d9d2daf656..ba46dcf2a5 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -263,6 +263,8 @@ ngbe_pf_reset_hw(struct ngbe_hw *hw)
 	status = hw->mac.reset_hw(hw);
 
 	ctrl_ext = rd32(hw, NGBE_PORTCTL);
+	/* let hardware know driver is loaded */
+	ctrl_ext |= NGBE_PORTCTL_DRVLOAD;
 	/* Set PF Reset Done bit so PF/VF Mail Ops can work */
 	ctrl_ext |= NGBE_PORTCTL_RSTDONE;
 	wr32(hw, NGBE_PORTCTL, ctrl_ext);
@@ -1277,6 +1279,9 @@ ngbe_dev_close(struct rte_eth_dev *dev)
 
 	ngbe_dev_stop(dev);
 
+	/* Let firmware take over control of hardware */
+	wr32m(hw, NGBE_PORTCTL, NGBE_PORTCTL_DRVLOAD, 0);
+
 	ngbe_dev_free_queues(dev);
 
 	ngbe_set_pcie_master(hw, false);
-- 
2.27.0

[PATCH 11/13] net/ngbe: reconfigure more MAC Rx registers

[Jiawen Wu]

When link status changes, there is a probability that no more packets
can be received on the port, due to hardware defects. These MAC Rx
registers should be reconfigured to fix this problem.

Fixes: b9246b8fa280 ("net/ngbe: support link update")
Fixes: a7c5f95ed9c2 ("net/ngbe: reconfigure MAC Rx when link update")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/base/ngbe_regs.h | 2 ++
 drivers/net/ngbe/ngbe_ethdev.c    | 6 ++++++
 2 files changed, 8 insertions(+)

diff --git a/drivers/net/ngbe/base/ngbe_regs.h b/drivers/net/ngbe/base/ngbe_regs.h
index 8a6776b0e6..b1295280a7 100644
--- a/drivers/net/ngbe/base/ngbe_regs.h
+++ b/drivers/net/ngbe/base/ngbe_regs.h
@@ -712,6 +712,8 @@ enum ngbe_5tuple_protocol {
 #define   NGBE_MACRXFLT_CTL_PASS       LS(3, 6, 0x3)
 #define   NGBE_MACRXFLT_RXALL          MS(31, 0x1)
 
+#define NGBE_MAC_WDG_TIMEOUT           0x01100C
+
 /******************************************************************************
  * Statistic Registers
  ******************************************************************************/
diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index ba46dcf2a5..3ea7ed43ff 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -1934,6 +1934,7 @@ ngbe_dev_link_update_share(struct rte_eth_dev *dev,
 	bool link_up;
 	int err;
 	int wait = 1;
+	u32 reg;
 
 	memset(&link, 0, sizeof(link));
 	link.link_status = RTE_ETH_LINK_DOWN;
@@ -1991,8 +1992,13 @@ ngbe_dev_link_update_share(struct rte_eth_dev *dev,
 			wr32m(hw, NGBE_MACTXCFG, NGBE_MACTXCFG_SPEED_MASK,
 				NGBE_MACTXCFG_SPEED_1G | NGBE_MACTXCFG_TE);
 		}
+		/* Re configure MAC RX */
+		reg = rd32(hw, NGBE_MACRXCFG);
+		wr32(hw, NGBE_MACRXCFG, reg);
 		wr32m(hw, NGBE_MACRXFLT, NGBE_MACRXFLT_PROMISC,
 			NGBE_MACRXFLT_PROMISC);
+		reg = rd32(hw, NGBE_MAC_WDG_TIMEOUT);
+		wr32(hw, NGBE_MAC_WDG_TIMEOUT, reg);
 	}
 
 	return rte_eth_linkstatus_set(dev, &link);
-- 
2.27.0

[PATCH 12/13] net/ngbe: fix interrupt lost in legacy or MSI mode

[Jiawen Wu]

When interrupt is legacy or MSI mode, shared interrupt may cause the
interrupt cannot be re-enabled. So fix to read the shared interrupt.

Fixes: b9246b8fa280 ("net/ngbe: support link update")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_ethdev.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index 3ea7ed43ff..e7dc1c0f94 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -2186,6 +2186,19 @@ ngbe_dev_interrupt_get_status(struct rte_eth_dev *dev)
 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
 	struct ngbe_interrupt *intr = ngbe_dev_intr(dev);
 
+	eicr = ((u32 *)hw->isb_mem)[NGBE_ISB_VEC0];
+	if (!eicr) {
+		/*
+		 * shared interrupt alert!
+		 * make sure interrupts are enabled because the read will
+		 * have disabled interrupts.
+		 */
+		if (!hw->adapter_stopped)
+			ngbe_enable_intr(dev);
+		return 0;
+	}
+	((u32 *)hw->isb_mem)[NGBE_ISB_VEC0] = 0;
+
 	/* read-on-clear nic registers here */
 	eicr = ((u32 *)hw->isb_mem)[NGBE_ISB_MISC];
 	PMD_DRV_LOG(DEBUG, "eicr %x", eicr);
-- 
2.27.0

[PATCH 13/13] net/ngbe: restrict configuration of VLAN strip offload

[Jiawen Wu]

There is a hardware limitation that Rx ring config register is not
writable when Rx ring is enabled, i.e. the NGBE_RXCFG_ENA bit is set.
But disabling the ring when there is traffic will cause ring get stuck.
So restrict the configuration of VLAN strip offload only if device is
started.

Fixes: 59b46438fdaa ("net/ngbe: support VLAN offload and VLAN filter")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_ethdev.c | 49 ++++++++++++++--------------------
 1 file changed, 20 insertions(+), 29 deletions(-)

diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index e7dc1c0f94..eef31af233 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -586,41 +586,25 @@ ngbe_vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
 }
 
 static void
-ngbe_vlan_strip_queue_set(struct rte_eth_dev *dev, uint16_t queue, int on)
+ngbe_vlan_strip_q_set(struct rte_eth_dev *dev, uint16_t queue, int on)
 {
-	struct ngbe_hw *hw = ngbe_dev_hw(dev);
-	struct ngbe_rx_queue *rxq;
-	bool restart;
-	uint32_t rxcfg, rxbal, rxbah;
-
 	if (on)
 		ngbe_vlan_hw_strip_enable(dev, queue);
 	else
 		ngbe_vlan_hw_strip_disable(dev, queue);
+}
 
-	rxq = dev->data->rx_queues[queue];
-	rxbal = rd32(hw, NGBE_RXBAL(rxq->reg_idx));
-	rxbah = rd32(hw, NGBE_RXBAH(rxq->reg_idx));
-	rxcfg = rd32(hw, NGBE_RXCFG(rxq->reg_idx));
-	if (rxq->offloads & RTE_ETH_RX_OFFLOAD_VLAN_STRIP) {
-		restart = (rxcfg & NGBE_RXCFG_ENA) &&
-			!(rxcfg & NGBE_RXCFG_VLAN);
-		rxcfg |= NGBE_RXCFG_VLAN;
-	} else {
-		restart = (rxcfg & NGBE_RXCFG_ENA) &&
-			(rxcfg & NGBE_RXCFG_VLAN);
-		rxcfg &= ~NGBE_RXCFG_VLAN;
-	}
-	rxcfg &= ~NGBE_RXCFG_ENA;
+static void
+ngbe_vlan_strip_queue_set(struct rte_eth_dev *dev, uint16_t queue, int on)
+{
+	struct ngbe_hw *hw = ngbe_dev_hw(dev);
 
-	if (restart) {
-		/* set vlan strip for ring */
-		ngbe_dev_rx_queue_stop(dev, queue);
-		wr32(hw, NGBE_RXBAL(rxq->reg_idx), rxbal);
-		wr32(hw, NGBE_RXBAH(rxq->reg_idx), rxbah);
-		wr32(hw, NGBE_RXCFG(rxq->reg_idx), rxcfg);
-		ngbe_dev_rx_queue_start(dev, queue);
+	if (!hw->adapter_stopped) {
+		PMD_DRV_LOG(ERR, "Please stop port first");
+		return;
 	}
+
+	ngbe_vlan_strip_q_set(dev, queue, on);
 }
 
 static int
@@ -846,9 +830,9 @@ ngbe_vlan_hw_strip_config(struct rte_eth_dev *dev)
 		rxq = dev->data->rx_queues[i];
 
 		if (rxq->offloads & RTE_ETH_RX_OFFLOAD_VLAN_STRIP)
-			ngbe_vlan_hw_strip_enable(dev, i);
+			ngbe_vlan_strip_q_set(dev, i, 1);
 		else
-			ngbe_vlan_hw_strip_disable(dev, i);
+			ngbe_vlan_strip_q_set(dev, i, 0);
 	}
 }
 
@@ -910,6 +894,13 @@ ngbe_vlan_offload_config(struct rte_eth_dev *dev, int mask)
 static int
 ngbe_vlan_offload_set(struct rte_eth_dev *dev, int mask)
 {
+	struct ngbe_hw *hw = ngbe_dev_hw(dev);
+
+	if (!hw->adapter_stopped && (mask & RTE_ETH_VLAN_STRIP_MASK)) {
+		PMD_DRV_LOG(ERR, "Please stop port first");
+		return -EPERM;
+	}
+
 	ngbe_config_vlan_strip_on_all_queues(dev, mask);
 
 	ngbe_vlan_offload_config(dev, mask);
-- 
2.27.0

[PATCH v2 00/13] Wangxun fixes

[Jiawen Wu]

Fix some bugs in txgbe/ngbe, and optimize the Tx flow.

v1 -> v2:
- Fix build errors on ARM platform.

Jiawen Wu (13):
  net/txgbe: fix swfw mbox failure
  net/txgbe: fix VF-PF mbox interrupt
  net/txgbe: remove outer UDP checksum capability
  net/txgbe: fix driver load bit to inform firmware
  net/txgbe: enable Tx descriptor error interrupt
  net/txgbe: check length of Tx packets
  net/txgbe: add Tx descriptor error statistics
  net/ngbe: check length of Tx packets
  net/ngbe: add Tx descriptor error statistics
  net/ngbe: fix driver load bit to inform firmware
  net/ngbe: reconfigure more MAC Rx registers
  net/ngbe: fix interrupt lost in legacy or MSI mode
  net/ngbe: restrict configuration of VLAN strip offload

 drivers/net/ngbe/base/ngbe_regs.h       |   2 +
 drivers/net/ngbe/ngbe_ethdev.c          |  79 +++++++++-----
 drivers/net/ngbe/ngbe_rxtx.c            |  35 ++++++
 drivers/net/ngbe/ngbe_rxtx.h            |   1 +
 drivers/net/ngbe/ngbe_rxtx_vec_neon.c   |  10 +-
 drivers/net/ngbe/ngbe_rxtx_vec_sse.c    |  11 +-
 drivers/net/txgbe/base/txgbe_mng.c      |   1 +
 drivers/net/txgbe/base/txgbe_regs.h     |   8 +-
 drivers/net/txgbe/txgbe_ethdev.c        |  75 +++++++++++++
 drivers/net/txgbe/txgbe_ethdev.h        |   3 +
 drivers/net/txgbe/txgbe_rxtx.c          | 139 +++++++++++++++++++++++-
 drivers/net/txgbe/txgbe_rxtx.h          |   2 +
 drivers/net/txgbe/txgbe_rxtx_vec_neon.c |  10 +-
 drivers/net/txgbe/txgbe_rxtx_vec_sse.c  |  11 +-
 14 files changed, 343 insertions(+), 44 deletions(-)

-- 
2.27.0

[PATCH v2 01/13] net/txgbe: fix swfw mbox failure

[Jiawen Wu]

There is a unknown bug that the register TXGBE_MNGMBX cannot be written
in the loop, when DPDK is built with GCC high version. Access any register
before write TXGBE_MNGMBX can fix it.

Bugzilla ID: 1531
Fixes: 35c90ecccfd4 ("net/txgbe: add EEPROM functions")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/base/txgbe_mng.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/net/txgbe/base/txgbe_mng.c b/drivers/net/txgbe/base/txgbe_mng.c
index 20db982891..7dc8f21183 100644
--- a/drivers/net/txgbe/base/txgbe_mng.c
+++ b/drivers/net/txgbe/base/txgbe_mng.c
@@ -58,6 +58,7 @@ txgbe_hic_unlocked(struct txgbe_hw *hw, u32 *buffer, u32 length, u32 timeout)
 
 	dword_len = length >> 2;
 
+	txgbe_flush(hw);
 	/* The device driver writes the relevant command block
 	 * into the ram area.
 	 */
-- 
2.27.0

[PATCH v2 02/13] net/txgbe: fix VF-PF mbox interrupt

[Jiawen Wu]

There was a incorrect bit to define TXGBE_ICRMISC_VFMBX that prevents the
interrupt from being handled correctly.

Fixes: a6712cd029a4 ("net/txgbe: add PF module init and uninit for SRIOV")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/base/txgbe_regs.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/txgbe/base/txgbe_regs.h b/drivers/net/txgbe/base/txgbe_regs.h
index 4ea4a2e3d8..b46d65331e 100644
--- a/drivers/net/txgbe/base/txgbe_regs.h
+++ b/drivers/net/txgbe/base/txgbe_regs.h
@@ -1197,7 +1197,7 @@ enum txgbe_5tuple_protocol {
 #define   TXGBE_ICRMISC_ANDONE MS(19, 0x1) /* link auto-nego done */
 #define   TXGBE_ICRMISC_ERRIG  MS(20, 0x1) /* integrity error */
 #define   TXGBE_ICRMISC_SPI    MS(21, 0x1) /* SPI interface */
-#define   TXGBE_ICRMISC_VFMBX  MS(22, 0x1) /* VF-PF message box */
+#define   TXGBE_ICRMISC_VFMBX  MS(23, 0x1) /* VF-PF message box */
 #define   TXGBE_ICRMISC_GPIO   MS(26, 0x1) /* GPIO interrupt */
 #define   TXGBE_ICRMISC_ERRPCI MS(27, 0x1) /* pcie request error */
 #define   TXGBE_ICRMISC_HEAT   MS(28, 0x1) /* overheat detection */
-- 
2.27.0

[PATCH v2 03/13] net/txgbe: remove outer UDP checksum capability

[Jiawen Wu]

The hardware does not support outer UDP checksum for tunnel packets.
It's wrong to claim this Tx offload capability, so fix it.

Bugzilla ID: 1529
Fixes: b950203be7f1 ("net/txgbe: support VXLAN-GPE")
Fixes: 295968d17407 ("ethdev: add namespace")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/txgbe_rxtx.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
index 5bc0f8772f..c12726553c 100644
--- a/drivers/net/txgbe/txgbe_rxtx.c
+++ b/drivers/net/txgbe/txgbe_rxtx.c
@@ -2284,8 +2284,7 @@ txgbe_get_tx_port_offloads(struct rte_eth_dev *dev)
 
 	tx_offload_capa |= RTE_ETH_TX_OFFLOAD_MACSEC_INSERT;
 
-	tx_offload_capa |= RTE_ETH_TX_OFFLOAD_OUTER_IPV4_CKSUM |
-			   RTE_ETH_TX_OFFLOAD_OUTER_UDP_CKSUM;
+	tx_offload_capa |= RTE_ETH_TX_OFFLOAD_OUTER_IPV4_CKSUM;
 
 #ifdef RTE_LIB_SECURITY
 	if (dev->security_ctx)
-- 
2.27.0

[PATCH v2 04/13] net/txgbe: fix driver load bit to inform firmware

[Jiawen Wu]

Drv_load bit will be reset to default 0 after hardware LAN reset,
reconfigure it to inform firmware that driver is loaded. And set it to 0
when device is closed.

Fixes: b1f596677d8e ("net/txgbe: support device start")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/txgbe_ethdev.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/net/txgbe/txgbe_ethdev.c b/drivers/net/txgbe/txgbe_ethdev.c
index 2834468764..4aa3bfd0bc 100644
--- a/drivers/net/txgbe/txgbe_ethdev.c
+++ b/drivers/net/txgbe/txgbe_ethdev.c
@@ -331,6 +331,8 @@ txgbe_pf_reset_hw(struct txgbe_hw *hw)
 	status = hw->mac.reset_hw(hw);
 
 	ctrl_ext = rd32(hw, TXGBE_PORTCTL);
+	/* let hardware know driver is loaded */
+	ctrl_ext |= TXGBE_PORTCTL_DRVLOAD;
 	/* Set PF Reset Done bit so PF/VF Mail Ops can work */
 	ctrl_ext |= TXGBE_PORTCTL_RSTDONE;
 	wr32(hw, TXGBE_PORTCTL, ctrl_ext);
@@ -2061,6 +2063,9 @@ txgbe_dev_close(struct rte_eth_dev *dev)
 
 	ret = txgbe_dev_stop(dev);
 
+	/* Let firmware take over control of hardware */
+	wr32m(hw, TXGBE_PORTCTL, TXGBE_PORTCTL_DRVLOAD, 0);
+
 	txgbe_dev_free_queues(dev);
 
 	txgbe_set_pcie_master(hw, false);
-- 
2.27.0

[PATCH v2 05/13] net/txgbe: enable Tx descriptor error interrupt

[Jiawen Wu]

Enable Tdm_desc_chk then handle the interrupt of TDM desc error. If it's a
non-fatal error, clear the error to reset the queue. If it's a fatal error,
require users to manually restart the port. This flow prevents the hardware
from PCIe pending due to Tx hang, resulting in a reboot to recover.

But remarkably, when packet layer length does not match the packet type in
TX descriptor, it will cause non-fatal error if Tdm_desc_chk is enabled.
But it can be transmitted normally if Tdm_desc_chk is disabled. So in order
to prevent hardware over check, fix the layer length on the basis of packet
type.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/base/txgbe_regs.h |   6 ++
 drivers/net/txgbe/txgbe_ethdev.c    |  64 ++++++++++++++++++
 drivers/net/txgbe/txgbe_ethdev.h    |   3 +
 drivers/net/txgbe/txgbe_rxtx.c      | 100 ++++++++++++++++++++++++++++
 drivers/net/txgbe/txgbe_rxtx.h      |   1 +
 5 files changed, 174 insertions(+)

diff --git a/drivers/net/txgbe/base/txgbe_regs.h b/drivers/net/txgbe/base/txgbe_regs.h
index b46d65331e..7a9ba6976f 100644
--- a/drivers/net/txgbe/base/txgbe_regs.h
+++ b/drivers/net/txgbe/base/txgbe_regs.h
@@ -1197,6 +1197,7 @@ enum txgbe_5tuple_protocol {
 #define   TXGBE_ICRMISC_ANDONE MS(19, 0x1) /* link auto-nego done */
 #define   TXGBE_ICRMISC_ERRIG  MS(20, 0x1) /* integrity error */
 #define   TXGBE_ICRMISC_SPI    MS(21, 0x1) /* SPI interface */
+#define   TXGBE_ICRMISC_TXDESC MS(22, 0x1) /* TDM desc error */
 #define   TXGBE_ICRMISC_VFMBX  MS(23, 0x1) /* VF-PF message box */
 #define   TXGBE_ICRMISC_GPIO   MS(26, 0x1) /* GPIO interrupt */
 #define   TXGBE_ICRMISC_ERRPCI MS(27, 0x1) /* pcie request error */
@@ -1382,6 +1383,11 @@ enum txgbe_5tuple_protocol {
 #define   TXGBE_TXCFG_WTHRESH(v)        LS(v, 16, 0x7F)
 #define   TXGBE_TXCFG_FLUSH             MS(26, 0x1)
 
+#define TXGBE_TDM_DESC_CHK(i)		(0x0180B0 + (i) * 4) /*0-3*/
+#define TXGBE_TDM_DESC_NONFATAL(i)	(0x0180C0 + (i) * 4) /*0-3*/
+#define TXGBE_TDM_DESC_FATAL(i)		(0x0180D0 + (i) * 4) /*0-3*/
+#define TXGBE_TDM_DESC_MASK(v)		MS(v, 0x1)
+
 /* interrupt registers */
 #define TXGBE_ITRI                      0x000180
 #define TXGBE_ITR(i)                    (0x000200 + 4 * (i))
diff --git a/drivers/net/txgbe/txgbe_ethdev.c b/drivers/net/txgbe/txgbe_ethdev.c
index 4aa3bfd0bc..bafa9cf829 100644
--- a/drivers/net/txgbe/txgbe_ethdev.c
+++ b/drivers/net/txgbe/txgbe_ethdev.c
@@ -1936,6 +1936,7 @@ txgbe_dev_stop(struct rte_eth_dev *dev)
 	PMD_INIT_FUNC_TRACE();
 
 	rte_eal_alarm_cancel(txgbe_dev_detect_sfp, dev);
+	rte_eal_alarm_cancel(txgbe_tx_queue_clear_error, dev);
 	txgbe_dev_wait_setup_link_complete(dev, 0);
 
 	/* disable interrupts */
@@ -2838,6 +2839,60 @@ txgbe_dev_setup_link_alarm_handler(void *param)
 	intr->flags &= ~TXGBE_FLAG_NEED_LINK_CONFIG;
 }
 
+static void
+txgbe_do_reset(struct rte_eth_dev *dev)
+{
+	struct txgbe_hw *hw = TXGBE_DEV_HW(dev);
+	struct txgbe_tx_queue *txq;
+	u32 i;
+
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		txq->resetting = true;
+	}
+
+	rte_delay_ms(1);
+	wr32(hw, TXGBE_RST, TXGBE_RST_LAN(hw->bus.lan_id));
+	txgbe_flush(hw);
+
+	PMD_DRV_LOG(ERR, "Please manually restart the port %d",
+		dev->data->port_id);
+}
+
+static void
+txgbe_tx_ring_recovery(struct rte_eth_dev *dev)
+{
+	struct txgbe_hw *hw = TXGBE_DEV_HW(dev);
+	u32 desc_error[4] = {0, 0, 0, 0};
+	struct txgbe_tx_queue *txq;
+	u32 i;
+
+	/* check tdm fatal error */
+	for (i = 0; i < 4; i++) {
+		desc_error[i] = rd32(hw, TXGBE_TDM_DESC_FATAL(i));
+		if (desc_error[i] != 0) {
+			PMD_DRV_LOG(ERR, "TDM fatal error reg[%d]: 0x%x", i, desc_error[i]);
+			txgbe_do_reset(dev);
+			return;
+		}
+	}
+
+	/* check tdm non-fatal error */
+	for (i = 0; i < 4; i++)
+		desc_error[i] = rd32(hw, TXGBE_TDM_DESC_NONFATAL(i));
+
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		if (desc_error[i / 32] & (1 << i % 32)) {
+			PMD_DRV_LOG(ERR, "TDM non-fatal error, reset port[%d] queue[%d]",
+				dev->data->port_id, i);
+			dev->data->tx_queue_state[i] = RTE_ETH_QUEUE_STATE_STOPPED;
+			txq = dev->data->tx_queues[i];
+			txq->resetting = true;
+			rte_eal_alarm_set(1000, txgbe_tx_queue_clear_error, (void *)dev);
+		}
+	}
+}
+
 /*
  * If @timeout_ms was 0, it means that it will not return until link complete.
  * It returns 1 on complete, return 0 on timeout.
@@ -3096,6 +3151,7 @@ txgbe_dev_misc_interrupt_setup(struct rte_eth_dev *dev)
 	intr->mask |= mask;
 	intr->mask_misc |= TXGBE_ICRMISC_GPIO;
 	intr->mask_misc |= TXGBE_ICRMISC_ANDONE;
+	intr->mask_misc |= TXGBE_ICRMISC_TXDESC;
 	return 0;
 }
 
@@ -3191,6 +3247,9 @@ txgbe_dev_interrupt_get_status(struct rte_eth_dev *dev,
 	if (eicr & TXGBE_ICRMISC_HEAT)
 		intr->flags |= TXGBE_FLAG_OVERHEAT;
 
+	if (eicr & TXGBE_ICRMISC_TXDESC)
+		intr->flags |= TXGBE_FLAG_TX_DESC_ERR;
+
 	((u32 *)hw->isb_mem)[TXGBE_ISB_MISC] = 0;
 
 	return 0;
@@ -3310,6 +3369,11 @@ txgbe_dev_interrupt_action(struct rte_eth_dev *dev,
 		intr->flags &= ~TXGBE_FLAG_OVERHEAT;
 	}
 
+	if (intr->flags & TXGBE_FLAG_TX_DESC_ERR) {
+		txgbe_tx_ring_recovery(dev);
+		intr->flags &= ~TXGBE_FLAG_TX_DESC_ERR;
+	}
+
 	PMD_DRV_LOG(DEBUG, "enable intr immediately");
 	txgbe_enable_intr(dev);
 	rte_intr_enable(intr_handle);
diff --git a/drivers/net/txgbe/txgbe_ethdev.h b/drivers/net/txgbe/txgbe_ethdev.h
index f0f4ced5b0..302ea9f037 100644
--- a/drivers/net/txgbe/txgbe_ethdev.h
+++ b/drivers/net/txgbe/txgbe_ethdev.h
@@ -31,6 +31,7 @@
 #define TXGBE_FLAG_NEED_LINK_CONFIG (uint32_t)(1 << 4)
 #define TXGBE_FLAG_NEED_AN_CONFIG   (uint32_t)(1 << 5)
 #define TXGBE_FLAG_OVERHEAT         (uint32_t)(1 << 6)
+#define TXGBE_FLAG_TX_DESC_ERR      (uint32_t)(1 << 7)
 
 /*
  * Defines that were not part of txgbe_type.h as they are not used by the
@@ -474,6 +475,8 @@ int txgbe_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id);
 
 int txgbe_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id);
 
+void txgbe_tx_queue_clear_error(void *param);
+
 void txgbe_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
 	struct rte_eth_rxq_info *qinfo);
 
diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
index c12726553c..2d2b437643 100644
--- a/drivers/net/txgbe/txgbe_rxtx.c
+++ b/drivers/net/txgbe/txgbe_rxtx.c
@@ -728,6 +728,66 @@ txgbe_get_tun_len(struct rte_mbuf *mbuf)
 	return tun_len;
 }
 
+static inline void
+txgbe_fix_offload_len(union txgbe_tx_offload *ol)
+{
+	uint8_t ptid = ol->ptid;
+
+	if (ptid & TXGBE_PTID_PKT_TUN) {
+		if (ol->outer_l2_len == 0)
+			ol->outer_l2_len = sizeof(struct rte_ether_hdr);
+		if (ol->outer_l3_len == 0) {
+			if (ptid & TXGBE_PTID_TUN_IPV6)
+				ol->outer_l3_len = sizeof(struct rte_ipv6_hdr);
+			else
+				ol->outer_l3_len = sizeof(struct rte_ipv4_hdr);
+		}
+		if ((ptid & 0xF) == 0) {
+			ol->l3_len = 0;
+			ol->l4_len = 0;
+		} else {
+			goto inner;
+		}
+	}
+
+	if ((ptid & 0xF0) == TXGBE_PTID_PKT_MAC) {
+		if (ol->l2_len == 0)
+			ol->l2_len = sizeof(struct rte_ether_hdr);
+		ol->l3_len = 0;
+		ol->l4_len = 0;
+	} else if ((ptid & 0xF0) == TXGBE_PTID_PKT_IP) {
+		if (ol->l2_len == 0)
+			ol->l2_len = sizeof(struct rte_ether_hdr);
+inner:
+		if (ol->l3_len == 0) {
+			if (ptid & TXGBE_PTID_PKT_IPV6)
+				ol->l3_len = sizeof(struct rte_ipv6_hdr);
+			else
+				ol->l3_len = sizeof(struct rte_ipv4_hdr);
+		}
+		switch (ptid & 0x7) {
+		case 0x1:
+		case 0x2:
+			ol->l4_len = 0;
+			break;
+		case 0x3:
+			if (ol->l4_len == 0)
+				ol->l4_len =  sizeof(struct rte_udp_hdr);
+			break;
+		case 0x4:
+			if (ol->l4_len == 0)
+				ol->l4_len =  sizeof(struct rte_tcp_hdr);
+			break;
+		case 0x5:
+			if (ol->l4_len == 0)
+				ol->l4_len =  sizeof(struct rte_sctp_hdr);
+			break;
+		default:
+			break;
+		}
+	}
+}
+
 static inline uint8_t
 txgbe_parse_tun_ptid(struct rte_mbuf *tx_pkt, uint8_t tun_len)
 {
@@ -782,6 +842,10 @@ txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 	uint8_t use_ipsec;
 #endif
 
+	txq = tx_queue;
+	if (txq->resetting)
+		return 0;
+
 	tx_offload.data[0] = 0;
 	tx_offload.data[1] = 0;
 	txq = tx_queue;
@@ -826,6 +890,7 @@ txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 			if (tx_offload.ptid & TXGBE_PTID_PKT_TUN)
 				tx_offload.ptid |= txgbe_parse_tun_ptid(tx_pkt,
 							tx_offload.outer_tun_len);
+			txgbe_fix_offload_len(&tx_offload);
 
 #ifdef RTE_LIB_SECURITY
 			if (use_ipsec) {
@@ -4570,6 +4635,11 @@ txgbe_dev_tx_init(struct rte_eth_dev *dev)
 		wr32(hw, TXGBE_TXWP(txq->reg_idx), 0);
 	}
 
+#ifndef RTE_LIB_SECURITY
+	for (i = 0; i < 4; i++)
+		wr32(hw, TXGBE_TDM_DESC_CHK(i), 0xFFFFFFFF);
+#endif
+
 	/* Device configured with multiple TX queues. */
 	txgbe_dev_mq_tx_configure(dev);
 }
@@ -4806,6 +4876,7 @@ txgbe_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
 	rte_wmb();
 	wr32(hw, TXGBE_TXWP(txq->reg_idx), txq->tx_tail);
 	dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
+	txq->resetting = false;
 
 	return 0;
 }
@@ -4863,6 +4934,35 @@ txgbe_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
 	return 0;
 }
 
+void
+txgbe_tx_queue_clear_error(void *param)
+{
+	struct rte_eth_dev *dev = (struct rte_eth_dev *)param;
+	struct txgbe_hw *hw = TXGBE_DEV_HW(dev);
+	struct txgbe_tx_queue *txq;
+	u32 i;
+
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		if (!txq->resetting)
+			continue;
+
+		txgbe_dev_save_tx_queue(hw, i);
+
+		/* tx ring reset */
+		wr32(hw, TXGBE_TDM_DESC_NONFATAL(i / 32),
+			TXGBE_TDM_DESC_MASK(i % 32));
+
+		if (txq->ops != NULL) {
+			txq->ops->release_mbufs(txq);
+			txq->ops->reset(txq);
+		}
+
+		txgbe_dev_store_tx_queue(hw, i);
+		txgbe_dev_tx_queue_start(dev, i);
+	}
+}
+
 void
 txgbe_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
 	struct rte_eth_rxq_info *qinfo)
diff --git a/drivers/net/txgbe/txgbe_rxtx.h b/drivers/net/txgbe/txgbe_rxtx.h
index 9155eb1f70..e668b60b1e 100644
--- a/drivers/net/txgbe/txgbe_rxtx.h
+++ b/drivers/net/txgbe/txgbe_rxtx.h
@@ -412,6 +412,7 @@ struct txgbe_tx_queue {
 	/**< indicates that IPsec TX feature is in use */
 #endif
 	const struct rte_memzone *mz;
+	bool		    resetting;
 };
 
 struct txgbe_txq_ops {
-- 
2.27.0

[PATCH v2 06/13] net/txgbe: check length of Tx packets

[Jiawen Wu]

Add checking of the Tx packet length to avoid TDM fatal error as far as
possible. Set the pkt_len=1518 for invalid packet in simple Tx code path,
and drop it directly in featured Tx code path.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/txgbe_rxtx.c          | 33 +++++++++++++++++++++++++
 drivers/net/txgbe/txgbe_rxtx_vec_neon.c | 10 +++++---
 drivers/net/txgbe/txgbe_rxtx_vec_sse.c  | 11 ++++++---
 3 files changed, 48 insertions(+), 6 deletions(-)

diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
index 2d2b437643..06acbd0881 100644
--- a/drivers/net/txgbe/txgbe_rxtx.c
+++ b/drivers/net/txgbe/txgbe_rxtx.c
@@ -160,6 +160,8 @@ tx4(volatile struct txgbe_tx_desc *txdp, struct rte_mbuf **pkts)
 	for (i = 0; i < 4; ++i, ++txdp, ++pkts) {
 		buf_dma_addr = rte_mbuf_data_iova(*pkts);
 		pkt_len = (*pkts)->data_len;
+		if (pkt_len < RTE_ETHER_HDR_LEN)
+			pkt_len = TXGBE_FRAME_SIZE_DFT;
 
 		/* write data to descriptor */
 		txdp->qw0 = rte_cpu_to_le_64(buf_dma_addr);
@@ -180,6 +182,8 @@ tx1(volatile struct txgbe_tx_desc *txdp, struct rte_mbuf **pkts)
 
 	buf_dma_addr = rte_mbuf_data_iova(*pkts);
 	pkt_len = (*pkts)->data_len;
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = TXGBE_FRAME_SIZE_DFT;
 
 	/* write data to descriptor */
 	txdp->qw0 = cpu_to_le64(buf_dma_addr);
@@ -813,6 +817,30 @@ txgbe_parse_tun_ptid(struct rte_mbuf *tx_pkt, uint8_t tun_len)
 	return ptid;
 }
 
+static inline bool
+txgbe_check_pkt_err(struct rte_mbuf *tx_pkt)
+{
+	uint32_t total_len = 0, nb_seg = 0;
+	struct rte_mbuf *mseg;
+
+	mseg = tx_pkt;
+	do {
+		if (mseg->data_len == 0)
+			return true;
+		total_len += mseg->data_len;
+		nb_seg++;
+		mseg = mseg->next;
+	} while (mseg != NULL);
+
+	if (tx_pkt->pkt_len != total_len || tx_pkt->pkt_len == 0)
+		return true;
+
+	if (tx_pkt->nb_segs != nb_seg || tx_pkt->nb_segs > 64)
+		return true;
+
+	return false;
+}
+
 uint16_t
 txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 		uint16_t nb_pkts)
@@ -864,6 +892,11 @@ txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
 		new_ctx = 0;
 		tx_pkt = *tx_pkts++;
+		if (txgbe_check_pkt_err(tx_pkt)) {
+			rte_pktmbuf_free(tx_pkt);
+			continue;
+		}
+
 		pkt_len = tx_pkt->pkt_len;
 
 		/*
diff --git a/drivers/net/txgbe/txgbe_rxtx_vec_neon.c b/drivers/net/txgbe/txgbe_rxtx_vec_neon.c
index a96baf9b1d..d4d647fab5 100644
--- a/drivers/net/txgbe/txgbe_rxtx_vec_neon.c
+++ b/drivers/net/txgbe/txgbe_rxtx_vec_neon.c
@@ -476,9 +476,13 @@ static inline void
 vtx1(volatile struct txgbe_tx_desc *txdp,
 		struct rte_mbuf *pkt, uint64_t flags)
 {
-	uint64x2_t descriptor = {
-			pkt->buf_iova + pkt->data_off,
-			(uint64_t)pkt->pkt_len << 45 | flags | pkt->data_len};
+	uint16_t pkt_len = pkt->data_len;
+
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = TXGBE_FRAME_SIZE_DFT;
+
+	uint64x2_t descriptor = {pkt->buf_iova + pkt->data_off,
+				(uint64_t)pkt_len << 45 | flags | pkt_len};
 
 	vst1q_u64((uint64_t *)(uintptr_t)txdp, descriptor);
 }
diff --git a/drivers/net/txgbe/txgbe_rxtx_vec_sse.c b/drivers/net/txgbe/txgbe_rxtx_vec_sse.c
index 1a3f2ce3cd..8ecce33471 100644
--- a/drivers/net/txgbe/txgbe_rxtx_vec_sse.c
+++ b/drivers/net/txgbe/txgbe_rxtx_vec_sse.c
@@ -607,9 +607,14 @@ static inline void
 vtx1(volatile struct txgbe_tx_desc *txdp,
 		struct rte_mbuf *pkt, uint64_t flags)
 {
-	__m128i descriptor = _mm_set_epi64x((uint64_t)pkt->pkt_len << 45 |
-			flags | pkt->data_len,
-			pkt->buf_iova + pkt->data_off);
+	uint16_t pkt_len = pkt->data_len;
+	__m128i descriptor;
+
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = TXGBE_FRAME_SIZE_DFT;
+
+	descriptor = _mm_set_epi64x((uint64_t)pkt_len << 45 | flags | pkt_len,
+				pkt->buf_iova + pkt->data_off);
 	_mm_store_si128((__m128i *)(uintptr_t)txdp, descriptor);
 }
 
-- 
2.27.0

Re: [PATCH v2 06/13] net/txgbe: check length of Tx packets

[Ferruh Yigit]

On 10/28/2024 2:31 AM, Jiawen Wu wrote:
> Add checking of the Tx packet length to avoid TDM fatal error as far as
> possible. Set the pkt_len=1518 for invalid packet in simple Tx code path,
> and drop it directly in featured Tx code path.
> 
> Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
> ---
>  drivers/net/txgbe/txgbe_rxtx.c          | 33 +++++++++++++++++++++++++
>  drivers/net/txgbe/txgbe_rxtx_vec_neon.c | 10 +++++---
>  drivers/net/txgbe/txgbe_rxtx_vec_sse.c  | 11 ++++++---
>  3 files changed, 48 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
> index 2d2b437643..06acbd0881 100644
> --- a/drivers/net/txgbe/txgbe_rxtx.c
> +++ b/drivers/net/txgbe/txgbe_rxtx.c
> @@ -160,6 +160,8 @@ tx4(volatile struct txgbe_tx_desc *txdp, struct rte_mbuf **pkts)
>  	for (i = 0; i < 4; ++i, ++txdp, ++pkts) {
>  		buf_dma_addr = rte_mbuf_data_iova(*pkts);
>  		pkt_len = (*pkts)->data_len;
> +		if (pkt_len < RTE_ETHER_HDR_LEN)
> +			pkt_len = TXGBE_FRAME_SIZE_DFT;
>  
>  		/* write data to descriptor */
>  		txdp->qw0 = rte_cpu_to_le_64(buf_dma_addr);
> @@ -180,6 +182,8 @@ tx1(volatile struct txgbe_tx_desc *txdp, struct rte_mbuf **pkts)
>  
>  	buf_dma_addr = rte_mbuf_data_iova(*pkts);
>  	pkt_len = (*pkts)->data_len;
> +	if (pkt_len < RTE_ETHER_HDR_LEN)
> +		pkt_len = TXGBE_FRAME_SIZE_DFT;
>  
>  	/* write data to descriptor */
>  	txdp->qw0 = cpu_to_le64(buf_dma_addr);
> @@ -813,6 +817,30 @@ txgbe_parse_tun_ptid(struct rte_mbuf *tx_pkt, uint8_t tun_len)
>  	return ptid;
>  }
>  
> +static inline bool
> +txgbe_check_pkt_err(struct rte_mbuf *tx_pkt)
> +{
> +	uint32_t total_len = 0, nb_seg = 0;
> +	struct rte_mbuf *mseg;
> +
> +	mseg = tx_pkt;
> +	do {
> +		if (mseg->data_len == 0)
> +			return true;
> +		total_len += mseg->data_len;
> +		nb_seg++;
> +		mseg = mseg->next;
> +	} while (mseg != NULL);
> +
> +	if (tx_pkt->pkt_len != total_len || tx_pkt->pkt_len == 0)
> +		return true;
> +
> +	if (tx_pkt->nb_segs != nb_seg || tx_pkt->nb_segs > 64)
> +		return true;
> +
> +	return false;
> +}
> +
>

Hi Jiawen,

Above are generic checks, we may add this function to ethdev driver
header (ethdev_driver.h) so that any PMD can use it, what do you think?

RE: [PATCH v2 06/13] net/txgbe: check length of Tx packets

[Jiawen Wu]

On Fri, Nov 1, 2024 9:23 AM, Ferruh Yigit wrote:
> On 10/28/2024 2:31 AM, Jiawen Wu wrote:
> > Add checking of the Tx packet length to avoid TDM fatal error as far as
> > possible. Set the pkt_len=1518 for invalid packet in simple Tx code path,
> > and drop it directly in featured Tx code path.
> >
> > Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
> > ---
> >  drivers/net/txgbe/txgbe_rxtx.c          | 33 +++++++++++++++++++++++++
> >  drivers/net/txgbe/txgbe_rxtx_vec_neon.c | 10 +++++---
> >  drivers/net/txgbe/txgbe_rxtx_vec_sse.c  | 11 ++++++---
> >  3 files changed, 48 insertions(+), 6 deletions(-)
> >
> > diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
> > index 2d2b437643..06acbd0881 100644
> > --- a/drivers/net/txgbe/txgbe_rxtx.c
> > +++ b/drivers/net/txgbe/txgbe_rxtx.c
> > @@ -160,6 +160,8 @@ tx4(volatile struct txgbe_tx_desc *txdp, struct rte_mbuf **pkts)
> >  	for (i = 0; i < 4; ++i, ++txdp, ++pkts) {
> >  		buf_dma_addr = rte_mbuf_data_iova(*pkts);
> >  		pkt_len = (*pkts)->data_len;
> > +		if (pkt_len < RTE_ETHER_HDR_LEN)
> > +			pkt_len = TXGBE_FRAME_SIZE_DFT;
> >
> >  		/* write data to descriptor */
> >  		txdp->qw0 = rte_cpu_to_le_64(buf_dma_addr);
> > @@ -180,6 +182,8 @@ tx1(volatile struct txgbe_tx_desc *txdp, struct rte_mbuf **pkts)
> >
> >  	buf_dma_addr = rte_mbuf_data_iova(*pkts);
> >  	pkt_len = (*pkts)->data_len;
> > +	if (pkt_len < RTE_ETHER_HDR_LEN)
> > +		pkt_len = TXGBE_FRAME_SIZE_DFT;
> >
> >  	/* write data to descriptor */
> >  	txdp->qw0 = cpu_to_le64(buf_dma_addr);
> > @@ -813,6 +817,30 @@ txgbe_parse_tun_ptid(struct rte_mbuf *tx_pkt, uint8_t tun_len)
> >  	return ptid;
> >  }
> >
> > +static inline bool
> > +txgbe_check_pkt_err(struct rte_mbuf *tx_pkt)
> > +{
> > +	uint32_t total_len = 0, nb_seg = 0;
> > +	struct rte_mbuf *mseg;
> > +
> > +	mseg = tx_pkt;
> > +	do {
> > +		if (mseg->data_len == 0)
> > +			return true;
> > +		total_len += mseg->data_len;
> > +		nb_seg++;
> > +		mseg = mseg->next;
> > +	} while (mseg != NULL);
> > +
> > +	if (tx_pkt->pkt_len != total_len || tx_pkt->pkt_len == 0)
> > +		return true;
> > +
> > +	if (tx_pkt->nb_segs != nb_seg || tx_pkt->nb_segs > 64)
> > +		return true;
> > +
> > +	return false;
> > +}
> > +
> >
> 
> Hi Jiawen,
> 
> Above are generic checks, we may add this function to ethdev driver
> header (ethdev_driver.h) so that any PMD can use it, what do you think?

In fact, the limitation of tx_pkt->nb_segs is already implemented in
.tx_pkt_prepare. But users developing their own apps don't necessarily call
this function. So I'd like to add these in txgbe_xmit_pkts(). What are you
going to do about it?

Re: [PATCH v2 06/13] net/txgbe: check length of Tx packets

[Ferruh Yigit]

On 11/1/2024 1:52 AM, Jiawen Wu wrote:
> On Fri, Nov 1, 2024 9:23 AM, Ferruh Yigit wrote:
>> On 10/28/2024 2:31 AM, Jiawen Wu wrote:
>>> Add checking of the Tx packet length to avoid TDM fatal error as far as
>>> possible. Set the pkt_len=1518 for invalid packet in simple Tx code path,
>>> and drop it directly in featured Tx code path.
>>>
>>> Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
>>> ---
>>>  drivers/net/txgbe/txgbe_rxtx.c          | 33 +++++++++++++++++++++++++
>>>  drivers/net/txgbe/txgbe_rxtx_vec_neon.c | 10 +++++---
>>>  drivers/net/txgbe/txgbe_rxtx_vec_sse.c  | 11 ++++++---
>>>  3 files changed, 48 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
>>> index 2d2b437643..06acbd0881 100644
>>> --- a/drivers/net/txgbe/txgbe_rxtx.c
>>> +++ b/drivers/net/txgbe/txgbe_rxtx.c
>>> @@ -160,6 +160,8 @@ tx4(volatile struct txgbe_tx_desc *txdp, struct rte_mbuf **pkts)
>>>  	for (i = 0; i < 4; ++i, ++txdp, ++pkts) {
>>>  		buf_dma_addr = rte_mbuf_data_iova(*pkts);
>>>  		pkt_len = (*pkts)->data_len;
>>> +		if (pkt_len < RTE_ETHER_HDR_LEN)
>>> +			pkt_len = TXGBE_FRAME_SIZE_DFT;
>>>
>>>  		/* write data to descriptor */
>>>  		txdp->qw0 = rte_cpu_to_le_64(buf_dma_addr);
>>> @@ -180,6 +182,8 @@ tx1(volatile struct txgbe_tx_desc *txdp, struct rte_mbuf **pkts)
>>>
>>>  	buf_dma_addr = rte_mbuf_data_iova(*pkts);
>>>  	pkt_len = (*pkts)->data_len;
>>> +	if (pkt_len < RTE_ETHER_HDR_LEN)
>>> +		pkt_len = TXGBE_FRAME_SIZE_DFT;
>>>
>>>  	/* write data to descriptor */
>>>  	txdp->qw0 = cpu_to_le64(buf_dma_addr);
>>> @@ -813,6 +817,30 @@ txgbe_parse_tun_ptid(struct rte_mbuf *tx_pkt, uint8_t tun_len)
>>>  	return ptid;
>>>  }
>>>
>>> +static inline bool
>>> +txgbe_check_pkt_err(struct rte_mbuf *tx_pkt)
>>> +{
>>> +	uint32_t total_len = 0, nb_seg = 0;
>>> +	struct rte_mbuf *mseg;
>>> +
>>> +	mseg = tx_pkt;
>>> +	do {
>>> +		if (mseg->data_len == 0)
>>> +			return true;
>>> +		total_len += mseg->data_len;
>>> +		nb_seg++;
>>> +		mseg = mseg->next;
>>> +	} while (mseg != NULL);
>>> +
>>> +	if (tx_pkt->pkt_len != total_len || tx_pkt->pkt_len == 0)
>>> +		return true;
>>> +
>>> +	if (tx_pkt->nb_segs != nb_seg || tx_pkt->nb_segs > 64)
>>> +		return true;
>>> +
>>> +	return false;
>>> +}
>>> +
>>>
>>
>> Hi Jiawen,
>>
>> Above are generic checks, we may add this function to ethdev driver
>> header (ethdev_driver.h) so that any PMD can use it, what do you think?
> 
> In fact, the limitation of tx_pkt->nb_segs is already implemented in
> .tx_pkt_prepare. But users developing their own apps don't necessarily call
> this function. So I'd like to add these in txgbe_xmit_pkts(). What are you
> going to do about it?
> 
> 

I had the same thought, that is why I am not suggesting to add these
checks to .tx_pkt_prepare() function.

But still these generic checks can go into a new static inline function
for drivers to use, not for end users, and any other driver can call
this function. Your two drivers already using the same function.

This is not hard requirement, but if there is a common code for multiple
drivers, that can go into a helper function in ethdev_driver.h

Re: [PATCH v2 06/13] net/txgbe: check length of Tx packets

[Stephen Hemminger]

On Fri, 1 Nov 2024 01:22:47 +0000
Ferruh Yigit <ferruh.yigit@amd.com> wrote:

> Hi Jiawen,
> 
> Above are generic checks, we may add this function to ethdev driver
> header (ethdev_driver.h) so that any PMD can use it, what do you think?

This is in the fastpath, and additional checks should not be added there.
Or at least put them under RTE_ETHDEV_DEBUG.

Re: [PATCH v2 06/13] net/txgbe: check length of Tx packets

[Ferruh Yigit]

On 11/1/2024 2:56 AM, Stephen Hemminger wrote:
> On Fri, 1 Nov 2024 01:22:47 +0000
> Ferruh Yigit <ferruh.yigit@amd.com> wrote:
> 
>> Hi Jiawen,
>>
>> Above are generic checks, we may add this function to ethdev driver
>> header (ethdev_driver.h) so that any PMD can use it, what do you think?
> 
> This is in the fastpath, and additional checks should not be added there.
> Or at least put them under RTE_ETHDEV_DEBUG.
>

It is in the datapath, and ideally there shouldn't be error checks in
datapath.
But if HW does not check for errors and doesn't recover from error
conditions, driver may prefer to get the hit and do the checks.

Not for all drivers, but for the drivers that intentionally wants do the
check, I thought having a common helper function helps but I can see it
can create confusion.

@Jiawen, please continue as it is in next version, I mean have the
checks function in the driver itself.

[PATCH v2 07/13] net/txgbe: add Tx descriptor error statistics

[Jiawen Wu]

Count the number of packets not sent due to Tx descriptor error.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/txgbe_ethdev.c | 6 ++++++
 drivers/net/txgbe/txgbe_rxtx.c   | 3 +++
 drivers/net/txgbe/txgbe_rxtx.h   | 1 +
 3 files changed, 10 insertions(+)

diff --git a/drivers/net/txgbe/txgbe_ethdev.c b/drivers/net/txgbe/txgbe_ethdev.c
index bafa9cf829..0c76e986f4 100644
--- a/drivers/net/txgbe/txgbe_ethdev.c
+++ b/drivers/net/txgbe/txgbe_ethdev.c
@@ -2344,6 +2344,7 @@ txgbe_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 	struct txgbe_hw_stats *hw_stats = TXGBE_DEV_STATS(dev);
 	struct txgbe_stat_mappings *stat_mappings =
 			TXGBE_DEV_STAT_MAPPINGS(dev);
+	struct txgbe_tx_queue *txq;
 	uint32_t i, j;
 
 	txgbe_read_stats_registers(hw, hw_stats);
@@ -2398,6 +2399,11 @@ txgbe_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 
 	/* Tx Errors */
 	stats->oerrors  = 0;
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		stats->oerrors += txq->desc_error;
+	}
+
 	return 0;
 }
 
diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
index 06acbd0881..fc9e7b14f5 100644
--- a/drivers/net/txgbe/txgbe_rxtx.c
+++ b/drivers/net/txgbe/txgbe_rxtx.c
@@ -894,6 +894,7 @@ txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 		tx_pkt = *tx_pkts++;
 		if (txgbe_check_pkt_err(tx_pkt)) {
 			rte_pktmbuf_free(tx_pkt);
+			txq->desc_error++;
 			continue;
 		}
 
@@ -2523,6 +2524,7 @@ txgbe_dev_tx_queue_setup(struct rte_eth_dev *dev,
 	txgbe_set_tx_function(dev, txq);
 
 	txq->ops->reset(txq);
+	txq->desc_error = 0;
 
 	dev->data->tx_queues[queue_idx] = txq;
 
@@ -4980,6 +4982,7 @@ txgbe_tx_queue_clear_error(void *param)
 		if (!txq->resetting)
 			continue;
 
+		txq->desc_error++;
 		txgbe_dev_save_tx_queue(hw, i);
 
 		/* tx ring reset */
diff --git a/drivers/net/txgbe/txgbe_rxtx.h b/drivers/net/txgbe/txgbe_rxtx.h
index e668b60b1e..622a0d3981 100644
--- a/drivers/net/txgbe/txgbe_rxtx.h
+++ b/drivers/net/txgbe/txgbe_rxtx.h
@@ -412,6 +412,7 @@ struct txgbe_tx_queue {
 	/**< indicates that IPsec TX feature is in use */
 #endif
 	const struct rte_memzone *mz;
+	uint64_t	    desc_error;
 	bool		    resetting;
 };
 
-- 
2.27.0

Re: [PATCH v2 07/13] net/txgbe: add Tx descriptor error statistics

[Ferruh Yigit]

On 10/28/2024 2:31 AM, Jiawen Wu wrote:
> Count the number of packets not sent due to Tx descriptor error.
> 
> Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
> ---
>  drivers/net/txgbe/txgbe_ethdev.c | 6 ++++++
>  drivers/net/txgbe/txgbe_rxtx.c   | 3 +++
>  drivers/net/txgbe/txgbe_rxtx.h   | 1 +
>  3 files changed, 10 insertions(+)
> 
> diff --git a/drivers/net/txgbe/txgbe_ethdev.c b/drivers/net/txgbe/txgbe_ethdev.c
> index bafa9cf829..0c76e986f4 100644
> --- a/drivers/net/txgbe/txgbe_ethdev.c
> +++ b/drivers/net/txgbe/txgbe_ethdev.c
> @@ -2344,6 +2344,7 @@ txgbe_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
>  	struct txgbe_hw_stats *hw_stats = TXGBE_DEV_STATS(dev);
>  	struct txgbe_stat_mappings *stat_mappings =
>  			TXGBE_DEV_STAT_MAPPINGS(dev);
> +	struct txgbe_tx_queue *txq;
>  	uint32_t i, j;
>  
>  	txgbe_read_stats_registers(hw, hw_stats);
> @@ -2398,6 +2399,11 @@ txgbe_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
>  
>  	/* Tx Errors */
>  	stats->oerrors  = 0;
> +	for (i = 0; i < dev->data->nb_tx_queues; i++) {
> +		txq = dev->data->tx_queues[i];
> +		stats->oerrors += txq->desc_error;
> +	}
> +

stats_get is implemented, but stats_reset also needs to be implemented
for this stat.

>  	return 0;
>  }
>  
> diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
> index 06acbd0881..fc9e7b14f5 100644
> --- a/drivers/net/txgbe/txgbe_rxtx.c
> +++ b/drivers/net/txgbe/txgbe_rxtx.c
> @@ -894,6 +894,7 @@ txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
>  		tx_pkt = *tx_pkts++;
>  		if (txgbe_check_pkt_err(tx_pkt)) {
>  			rte_pktmbuf_free(tx_pkt);
> +			txq->desc_error++;
>  			continue;
>  		}
>  
> @@ -2523,6 +2524,7 @@ txgbe_dev_tx_queue_setup(struct rte_eth_dev *dev,
>  	txgbe_set_tx_function(dev, txq);
>  
>  	txq->ops->reset(txq);
> +	txq->desc_error = 0;
>  
>  	dev->data->tx_queues[queue_idx] = txq;
>  
> @@ -4980,6 +4982,7 @@ txgbe_tx_queue_clear_error(void *param)
>  		if (!txq->resetting)
>  			continue;
>  
> +		txq->desc_error++;
>

Why error value is increased in this function, which resets the Tx queue?
Is the intention to reset the error value here?

>  		txgbe_dev_save_tx_queue(hw, i);
>  
>  		/* tx ring reset */
> diff --git a/drivers/net/txgbe/txgbe_rxtx.h b/drivers/net/txgbe/txgbe_rxtx.h
> index e668b60b1e..622a0d3981 100644
> --- a/drivers/net/txgbe/txgbe_rxtx.h
> +++ b/drivers/net/txgbe/txgbe_rxtx.h
> @@ -412,6 +412,7 @@ struct txgbe_tx_queue {
>  	/**< indicates that IPsec TX feature is in use */
>  #endif
>  	const struct rte_memzone *mz;
> +	uint64_t	    desc_error;
>  	bool		    resetting;
>  };
>  

RE: [PATCH v2 07/13] net/txgbe: add Tx descriptor error statistics

[Jiawen Wu]

> > @@ -4980,6 +4982,7 @@ txgbe_tx_queue_clear_error(void *param)
> >  		if (!txq->resetting)
> >  			continue;
> >
> > +		txq->desc_error++;
> >
> 
> Why error value is increased in this function, which resets the Tx queue?
> Is the intention to reset the error value here?

When there is a desc error to cause Tx ring hang,  the interrupt directs to
reset of the queue. So we increase the error count in the specific queue
reset function. The queue is reset, but the error that led to the resetting
is recorded. The error count is only reset at the setup function.

Re: [PATCH v2 07/13] net/txgbe: add Tx descriptor error statistics

[Ferruh Yigit]

On 11/1/2024 2:06 AM, Jiawen Wu wrote:
>>> @@ -4980,6 +4982,7 @@ txgbe_tx_queue_clear_error(void *param)
>>>  		if (!txq->resetting)
>>>  			continue;
>>>
>>> +		txq->desc_error++;
>>>
>>
>> Why error value is increased in this function, which resets the Tx queue?
>> Is the intention to reset the error value here?
> 
> When there is a desc error to cause Tx ring hang,  the interrupt directs to
> reset of the queue. So we increase the error count in the specific queue
> reset function. The queue is reset, but the error that led to the resetting
> is recorded. The error count is only reset at the setup function.
> 

Got it. But this variable counts number of bad packets. Increasing it
when Tx ring hang too may be confusing as there are two different
reasons to increase the counter.

Btw, .stats_reset() needs to reset this error stat.

RE: [PATCH v2 07/13] net/txgbe: add Tx descriptor error statistics

[Jiawen Wu]

> -----Original Message-----
> From: Ferruh Yigit <ferruh.yigit@amd.com>
> Sent: Friday, November 1, 2024 10:46 AM
> To: Jiawen Wu <jiawenwu@trustnetic.com>; dev@dpdk.org
> Subject: Re: [PATCH v2 07/13] net/txgbe: add Tx descriptor error statistics
> 
> On 11/1/2024 2:06 AM, Jiawen Wu wrote:
> >>> @@ -4980,6 +4982,7 @@ txgbe_tx_queue_clear_error(void *param)
> >>>  		if (!txq->resetting)
> >>>  			continue;
> >>>
> >>> +		txq->desc_error++;
> >>>
> >>
> >> Why error value is increased in this function, which resets the Tx queue?
> >> Is the intention to reset the error value here?
> >
> > When there is a desc error to cause Tx ring hang,  the interrupt directs to
> > reset of the queue. So we increase the error count in the specific queue
> > reset function. The queue is reset, but the error that led to the resetting
> > is recorded. The error count is only reset at the setup function.
> >
> 
> Got it. But this variable counts number of bad packets. Increasing it
> when Tx ring hang too may be confusing as there are two different
> reasons to increase the counter.

I think I should add a comment here:
/* Increase the count of Tx desc error since it causes the queue reset. */

> Btw, .stats_reset() needs to reset this error stat.

I'll add it into .stats_reset() in patch set v3.

[PATCH v2 08/13] net/ngbe: check length of Tx packets

[Jiawen Wu]

Add checking of the Tx packet length to avoid TDM fatal error as far as
possible. Set the pkt_len=1518 for invalid packet in simple Tx code path,
and drop it directly in featured Tx code path. Althrough the hardware
does not support TDM desc check.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_rxtx.c          | 33 +++++++++++++++++++++++++++
 drivers/net/ngbe/ngbe_rxtx_vec_neon.c | 10 +++++---
 drivers/net/ngbe/ngbe_rxtx_vec_sse.c  | 11 ++++++---
 3 files changed, 48 insertions(+), 6 deletions(-)

diff --git a/drivers/net/ngbe/ngbe_rxtx.c b/drivers/net/ngbe/ngbe_rxtx.c
index f3eb797d0c..25a314f10a 100644
--- a/drivers/net/ngbe/ngbe_rxtx.c
+++ b/drivers/net/ngbe/ngbe_rxtx.c
@@ -113,6 +113,8 @@ tx4(volatile struct ngbe_tx_desc *txdp, struct rte_mbuf **pkts)
 	for (i = 0; i < 4; ++i, ++txdp, ++pkts) {
 		buf_dma_addr = rte_mbuf_data_iova(*pkts);
 		pkt_len = (*pkts)->data_len;
+		if (pkt_len < RTE_ETHER_HDR_LEN)
+			pkt_len = NGBE_FRAME_SIZE_DFT;
 
 		/* write data to descriptor */
 		txdp->qw0 = rte_cpu_to_le_64(buf_dma_addr);
@@ -133,6 +135,8 @@ tx1(volatile struct ngbe_tx_desc *txdp, struct rte_mbuf **pkts)
 
 	buf_dma_addr = rte_mbuf_data_iova(*pkts);
 	pkt_len = (*pkts)->data_len;
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = NGBE_FRAME_SIZE_DFT;
 
 	/* write data to descriptor */
 	txdp->qw0 = cpu_to_le64(buf_dma_addr);
@@ -555,6 +559,30 @@ ngbe_xmit_cleanup(struct ngbe_tx_queue *txq)
 	return 0;
 }
 
+static inline bool
+ngbe_check_pkt_err(struct rte_mbuf *tx_pkt)
+{
+	uint32_t total_len = 0, nb_seg = 0;
+	struct rte_mbuf *mseg;
+
+	mseg = tx_pkt;
+	do {
+		if (mseg->data_len == 0)
+			return true;
+		total_len += mseg->data_len;
+		nb_seg++;
+		mseg = mseg->next;
+	} while (mseg != NULL);
+
+	if (tx_pkt->pkt_len != total_len || tx_pkt->pkt_len == 0)
+		return true;
+
+	if (tx_pkt->nb_segs != nb_seg || tx_pkt->nb_segs > 64)
+		return true;
+
+	return false;
+}
+
 uint16_t
 ngbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 		uint16_t nb_pkts)
@@ -599,6 +627,11 @@ ngbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
 		new_ctx = 0;
 		tx_pkt = *tx_pkts++;
+		if (ngbe_check_pkt_err(tx_pkt)) {
+			rte_pktmbuf_free(tx_pkt);
+			continue;
+		}
+
 		pkt_len = tx_pkt->pkt_len;
 
 		/*
diff --git a/drivers/net/ngbe/ngbe_rxtx_vec_neon.c b/drivers/net/ngbe/ngbe_rxtx_vec_neon.c
index dcf12b7070..37075ea5e7 100644
--- a/drivers/net/ngbe/ngbe_rxtx_vec_neon.c
+++ b/drivers/net/ngbe/ngbe_rxtx_vec_neon.c
@@ -476,9 +476,13 @@ static inline void
 vtx1(volatile struct ngbe_tx_desc *txdp,
 		struct rte_mbuf *pkt, uint64_t flags)
 {
-	uint64x2_t descriptor = {
-			pkt->buf_iova + pkt->data_off,
-			(uint64_t)pkt->pkt_len << 45 | flags | pkt->data_len};
+	uint16_t pkt_len = pkt->data_len;
+
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = NGBE_FRAME_SIZE_DFT;
+
+	uint64x2_t descriptor = {pkt->buf_iova + pkt->data_off,
+				(uint64_t)pkt_len << 45 | flags | pkt_len};
 
 	vst1q_u64((uint64_t *)(uintptr_t)txdp, descriptor);
 }
diff --git a/drivers/net/ngbe/ngbe_rxtx_vec_sse.c b/drivers/net/ngbe/ngbe_rxtx_vec_sse.c
index b128bd3a67..19c69cdfa6 100644
--- a/drivers/net/ngbe/ngbe_rxtx_vec_sse.c
+++ b/drivers/net/ngbe/ngbe_rxtx_vec_sse.c
@@ -563,9 +563,14 @@ static inline void
 vtx1(volatile struct ngbe_tx_desc *txdp,
 		struct rte_mbuf *pkt, uint64_t flags)
 {
-	__m128i descriptor = _mm_set_epi64x((uint64_t)pkt->pkt_len << 45 |
-			flags | pkt->data_len,
-			pkt->buf_iova + pkt->data_off);
+	uint16_t pkt_len = pkt->data_len;
+	__m128i descriptor;
+
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = NGBE_FRAME_SIZE_DFT;
+
+	descriptor = _mm_set_epi64x((uint64_t)pkt_len << 45 | flags | pkt_len,
+				pkt->buf_iova + pkt->data_off);
 	_mm_store_si128((__m128i *)(uintptr_t)txdp, descriptor);
 }
 
-- 
2.27.0

[PATCH v2 09/13] net/ngbe: add Tx descriptor error statistics

[Jiawen Wu]

Count the number of packets not sent due to Tx descriptor error.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_ethdev.c | 6 ++++++
 drivers/net/ngbe/ngbe_rxtx.c   | 2 ++
 drivers/net/ngbe/ngbe_rxtx.h   | 1 +
 3 files changed, 9 insertions(+)

diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index 6c45ffaad3..d9d2daf656 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -1507,6 +1507,7 @@ ngbe_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 	struct ngbe_hw_stats *hw_stats = NGBE_DEV_STATS(dev);
 	struct ngbe_stat_mappings *stat_mappings =
 			NGBE_DEV_STAT_MAPPINGS(dev);
+	struct ngbe_tx_queue *txq;
 	uint32_t i, j;
 
 	ngbe_read_stats_registers(hw, hw_stats);
@@ -1559,6 +1560,11 @@ ngbe_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 
 	/* Tx Errors */
 	stats->oerrors  = 0;
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		stats->oerrors += txq->desc_error;
+	}
+
 	return 0;
 }
 
diff --git a/drivers/net/ngbe/ngbe_rxtx.c b/drivers/net/ngbe/ngbe_rxtx.c
index 25a314f10a..8d31d47de9 100644
--- a/drivers/net/ngbe/ngbe_rxtx.c
+++ b/drivers/net/ngbe/ngbe_rxtx.c
@@ -629,6 +629,7 @@ ngbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 		tx_pkt = *tx_pkts++;
 		if (ngbe_check_pkt_err(tx_pkt)) {
 			rte_pktmbuf_free(tx_pkt);
+			txq->desc_error++;
 			continue;
 		}
 
@@ -2100,6 +2101,7 @@ ngbe_dev_tx_queue_setup(struct rte_eth_dev *dev,
 	ngbe_set_tx_function(dev, txq);
 
 	txq->ops->reset(txq);
+	txq->desc_error = 0;
 
 	dev->data->tx_queues[queue_idx] = txq;
 
diff --git a/drivers/net/ngbe/ngbe_rxtx.h b/drivers/net/ngbe/ngbe_rxtx.h
index 7574db32d8..8534ec123a 100644
--- a/drivers/net/ngbe/ngbe_rxtx.h
+++ b/drivers/net/ngbe/ngbe_rxtx.h
@@ -375,6 +375,7 @@ struct ngbe_tx_queue {
 
 	const struct ngbe_txq_ops *ops;       /**< txq ops */
 	const struct rte_memzone *mz;
+	uint64_t            desc_error;
 };
 
 struct ngbe_txq_ops {
-- 
2.27.0

[PATCH v2 10/13] net/ngbe: fix driver load bit to inform firmware

[Jiawen Wu]

Drv_load bit will be reset to default 0 after hardware LAN reset,
reconfigure it to inform firmware that driver is loaded. And set it to 0
when device is closed.

Fixes: 3518df5774c7 ("net/ngbe: support device start/stop")
Fixes: cc63194e89cb ("net/ngbe: support close and reset device")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_ethdev.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index d9d2daf656..ba46dcf2a5 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -263,6 +263,8 @@ ngbe_pf_reset_hw(struct ngbe_hw *hw)
 	status = hw->mac.reset_hw(hw);
 
 	ctrl_ext = rd32(hw, NGBE_PORTCTL);
+	/* let hardware know driver is loaded */
+	ctrl_ext |= NGBE_PORTCTL_DRVLOAD;
 	/* Set PF Reset Done bit so PF/VF Mail Ops can work */
 	ctrl_ext |= NGBE_PORTCTL_RSTDONE;
 	wr32(hw, NGBE_PORTCTL, ctrl_ext);
@@ -1277,6 +1279,9 @@ ngbe_dev_close(struct rte_eth_dev *dev)
 
 	ngbe_dev_stop(dev);
 
+	/* Let firmware take over control of hardware */
+	wr32m(hw, NGBE_PORTCTL, NGBE_PORTCTL_DRVLOAD, 0);
+
 	ngbe_dev_free_queues(dev);
 
 	ngbe_set_pcie_master(hw, false);
-- 
2.27.0

[PATCH v2 11/13] net/ngbe: reconfigure more MAC Rx registers

[Jiawen Wu]

When link status changes, there is a probability that no more packets
can be received on the port, due to hardware defects. These MAC Rx
registers should be reconfigured to fix this problem.

Fixes: b9246b8fa280 ("net/ngbe: support link update")
Fixes: a7c5f95ed9c2 ("net/ngbe: reconfigure MAC Rx when link update")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/base/ngbe_regs.h | 2 ++
 drivers/net/ngbe/ngbe_ethdev.c    | 6 ++++++
 2 files changed, 8 insertions(+)

diff --git a/drivers/net/ngbe/base/ngbe_regs.h b/drivers/net/ngbe/base/ngbe_regs.h
index 8a6776b0e6..b1295280a7 100644
--- a/drivers/net/ngbe/base/ngbe_regs.h
+++ b/drivers/net/ngbe/base/ngbe_regs.h
@@ -712,6 +712,8 @@ enum ngbe_5tuple_protocol {
 #define   NGBE_MACRXFLT_CTL_PASS       LS(3, 6, 0x3)
 #define   NGBE_MACRXFLT_RXALL          MS(31, 0x1)
 
+#define NGBE_MAC_WDG_TIMEOUT           0x01100C
+
 /******************************************************************************
  * Statistic Registers
  ******************************************************************************/
diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index ba46dcf2a5..3ea7ed43ff 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -1934,6 +1934,7 @@ ngbe_dev_link_update_share(struct rte_eth_dev *dev,
 	bool link_up;
 	int err;
 	int wait = 1;
+	u32 reg;
 
 	memset(&link, 0, sizeof(link));
 	link.link_status = RTE_ETH_LINK_DOWN;
@@ -1991,8 +1992,13 @@ ngbe_dev_link_update_share(struct rte_eth_dev *dev,
 			wr32m(hw, NGBE_MACTXCFG, NGBE_MACTXCFG_SPEED_MASK,
 				NGBE_MACTXCFG_SPEED_1G | NGBE_MACTXCFG_TE);
 		}
+		/* Re configure MAC RX */
+		reg = rd32(hw, NGBE_MACRXCFG);
+		wr32(hw, NGBE_MACRXCFG, reg);
 		wr32m(hw, NGBE_MACRXFLT, NGBE_MACRXFLT_PROMISC,
 			NGBE_MACRXFLT_PROMISC);
+		reg = rd32(hw, NGBE_MAC_WDG_TIMEOUT);
+		wr32(hw, NGBE_MAC_WDG_TIMEOUT, reg);
 	}
 
 	return rte_eth_linkstatus_set(dev, &link);
-- 
2.27.0

[PATCH v2 12/13] net/ngbe: fix interrupt lost in legacy or MSI mode

[Jiawen Wu]

When interrupt is legacy or MSI mode, shared interrupt may cause the
interrupt cannot be re-enabled. So fix to read the shared interrupt.

Fixes: b9246b8fa280 ("net/ngbe: support link update")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_ethdev.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index 3ea7ed43ff..e7dc1c0f94 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -2186,6 +2186,19 @@ ngbe_dev_interrupt_get_status(struct rte_eth_dev *dev)
 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
 	struct ngbe_interrupt *intr = ngbe_dev_intr(dev);
 
+	eicr = ((u32 *)hw->isb_mem)[NGBE_ISB_VEC0];
+	if (!eicr) {
+		/*
+		 * shared interrupt alert!
+		 * make sure interrupts are enabled because the read will
+		 * have disabled interrupts.
+		 */
+		if (!hw->adapter_stopped)
+			ngbe_enable_intr(dev);
+		return 0;
+	}
+	((u32 *)hw->isb_mem)[NGBE_ISB_VEC0] = 0;
+
 	/* read-on-clear nic registers here */
 	eicr = ((u32 *)hw->isb_mem)[NGBE_ISB_MISC];
 	PMD_DRV_LOG(DEBUG, "eicr %x", eicr);
-- 
2.27.0

[PATCH v2 13/13] net/ngbe: restrict configuration of VLAN strip offload

[Jiawen Wu]

There is a hardware limitation that Rx ring config register is not
writable when Rx ring is enabled, i.e. the NGBE_RXCFG_ENA bit is set.
But disabling the ring when there is traffic will cause ring get stuck.
So restrict the configuration of VLAN strip offload only if device is
started.

Fixes: 59b46438fdaa ("net/ngbe: support VLAN offload and VLAN filter")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_ethdev.c | 49 ++++++++++++++--------------------
 1 file changed, 20 insertions(+), 29 deletions(-)

diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index e7dc1c0f94..eef31af233 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -586,41 +586,25 @@ ngbe_vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
 }
 
 static void
-ngbe_vlan_strip_queue_set(struct rte_eth_dev *dev, uint16_t queue, int on)
+ngbe_vlan_strip_q_set(struct rte_eth_dev *dev, uint16_t queue, int on)
 {
-	struct ngbe_hw *hw = ngbe_dev_hw(dev);
-	struct ngbe_rx_queue *rxq;
-	bool restart;
-	uint32_t rxcfg, rxbal, rxbah;
-
 	if (on)
 		ngbe_vlan_hw_strip_enable(dev, queue);
 	else
 		ngbe_vlan_hw_strip_disable(dev, queue);
+}
 
-	rxq = dev->data->rx_queues[queue];
-	rxbal = rd32(hw, NGBE_RXBAL(rxq->reg_idx));
-	rxbah = rd32(hw, NGBE_RXBAH(rxq->reg_idx));
-	rxcfg = rd32(hw, NGBE_RXCFG(rxq->reg_idx));
-	if (rxq->offloads & RTE_ETH_RX_OFFLOAD_VLAN_STRIP) {
-		restart = (rxcfg & NGBE_RXCFG_ENA) &&
-			!(rxcfg & NGBE_RXCFG_VLAN);
-		rxcfg |= NGBE_RXCFG_VLAN;
-	} else {
-		restart = (rxcfg & NGBE_RXCFG_ENA) &&
-			(rxcfg & NGBE_RXCFG_VLAN);
-		rxcfg &= ~NGBE_RXCFG_VLAN;
-	}
-	rxcfg &= ~NGBE_RXCFG_ENA;
+static void
+ngbe_vlan_strip_queue_set(struct rte_eth_dev *dev, uint16_t queue, int on)
+{
+	struct ngbe_hw *hw = ngbe_dev_hw(dev);
 
-	if (restart) {
-		/* set vlan strip for ring */
-		ngbe_dev_rx_queue_stop(dev, queue);
-		wr32(hw, NGBE_RXBAL(rxq->reg_idx), rxbal);
-		wr32(hw, NGBE_RXBAH(rxq->reg_idx), rxbah);
-		wr32(hw, NGBE_RXCFG(rxq->reg_idx), rxcfg);
-		ngbe_dev_rx_queue_start(dev, queue);
+	if (!hw->adapter_stopped) {
+		PMD_DRV_LOG(ERR, "Please stop port first");
+		return;
 	}
+
+	ngbe_vlan_strip_q_set(dev, queue, on);
 }
 
 static int
@@ -846,9 +830,9 @@ ngbe_vlan_hw_strip_config(struct rte_eth_dev *dev)
 		rxq = dev->data->rx_queues[i];
 
 		if (rxq->offloads & RTE_ETH_RX_OFFLOAD_VLAN_STRIP)
-			ngbe_vlan_hw_strip_enable(dev, i);
+			ngbe_vlan_strip_q_set(dev, i, 1);
 		else
-			ngbe_vlan_hw_strip_disable(dev, i);
+			ngbe_vlan_strip_q_set(dev, i, 0);
 	}
 }
 
@@ -910,6 +894,13 @@ ngbe_vlan_offload_config(struct rte_eth_dev *dev, int mask)
 static int
 ngbe_vlan_offload_set(struct rte_eth_dev *dev, int mask)
 {
+	struct ngbe_hw *hw = ngbe_dev_hw(dev);
+
+	if (!hw->adapter_stopped && (mask & RTE_ETH_VLAN_STRIP_MASK)) {
+		PMD_DRV_LOG(ERR, "Please stop port first");
+		return -EPERM;
+	}
+
 	ngbe_config_vlan_strip_on_all_queues(dev, mask);
 
 	ngbe_vlan_offload_config(dev, mask);
-- 
2.27.0

[PATCH v3 00/13] Wangxun fixes

[Jiawen Wu]

Fix some bugs in txgbe/ngbe, and optimize the Tx flow.

v2 -> v3:
- Reset Tx desc error statistics in .stats_reset().
- Add comment for Tx desc error count.

v1 -> v2:
- Fix build errors on ARM platform.

Jiawen Wu (13):
  net/txgbe: fix swfw mbox failure
  net/txgbe: fix VF-PF mbox interrupt
  net/txgbe: remove outer UDP checksum capability
  net/txgbe: fix driver load bit to inform firmware
  net/txgbe: enable Tx descriptor error interrupt
  net/txgbe: check length of Tx packets
  net/txgbe: add Tx descriptor error statistics
  net/ngbe: check length of Tx packets
  net/ngbe: add Tx descriptor error statistics
  net/ngbe: fix driver load bit to inform firmware
  net/ngbe: reconfigure more MAC Rx registers
  net/ngbe: fix interrupt lost in legacy or MSI mode
  net/ngbe: restrict configuration of VLAN strip offload

 drivers/net/ngbe/base/ngbe_regs.h       |   2 +
 drivers/net/ngbe/ngbe_ethdev.c          |  86 +++++++++-----
 drivers/net/ngbe/ngbe_rxtx.c            |  35 ++++++
 drivers/net/ngbe/ngbe_rxtx.h            |   1 +
 drivers/net/ngbe/ngbe_rxtx_vec_neon.c   |  10 +-
 drivers/net/ngbe/ngbe_rxtx_vec_sse.c    |  11 +-
 drivers/net/txgbe/base/txgbe_mng.c      |   1 +
 drivers/net/txgbe/base/txgbe_regs.h     |   8 +-
 drivers/net/txgbe/txgbe_ethdev.c        |  82 ++++++++++++++
 drivers/net/txgbe/txgbe_ethdev.h        |   3 +
 drivers/net/txgbe/txgbe_rxtx.c          | 142 +++++++++++++++++++++++-
 drivers/net/txgbe/txgbe_rxtx.h          |   2 +
 drivers/net/txgbe/txgbe_rxtx_vec_neon.c |  10 +-
 drivers/net/txgbe/txgbe_rxtx_vec_sse.c  |  11 +-
 14 files changed, 360 insertions(+), 44 deletions(-)

-- 
2.27.0

[PATCH v3 01/13] net/txgbe: fix swfw mbox failure

[Jiawen Wu]

There is a unknown bug that the register TXGBE_MNGMBX cannot be written
in the loop, when DPDK is built with GCC high version. Access any register
before write TXGBE_MNGMBX can fix it.

Bugzilla ID: 1531
Fixes: 35c90ecccfd4 ("net/txgbe: add EEPROM functions")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/base/txgbe_mng.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/net/txgbe/base/txgbe_mng.c b/drivers/net/txgbe/base/txgbe_mng.c
index 20db982891..7dc8f21183 100644
--- a/drivers/net/txgbe/base/txgbe_mng.c
+++ b/drivers/net/txgbe/base/txgbe_mng.c
@@ -58,6 +58,7 @@ txgbe_hic_unlocked(struct txgbe_hw *hw, u32 *buffer, u32 length, u32 timeout)
 
 	dword_len = length >> 2;
 
+	txgbe_flush(hw);
 	/* The device driver writes the relevant command block
 	 * into the ram area.
 	 */
-- 
2.27.0

[PATCH v3 02/13] net/txgbe: fix VF-PF mbox interrupt

[Jiawen Wu]

There was a incorrect bit to define TXGBE_ICRMISC_VFMBX that prevents the
interrupt from being handled correctly.

Fixes: a6712cd029a4 ("net/txgbe: add PF module init and uninit for SRIOV")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/base/txgbe_regs.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/txgbe/base/txgbe_regs.h b/drivers/net/txgbe/base/txgbe_regs.h
index 4ea4a2e3d8..b46d65331e 100644
--- a/drivers/net/txgbe/base/txgbe_regs.h
+++ b/drivers/net/txgbe/base/txgbe_regs.h
@@ -1197,7 +1197,7 @@ enum txgbe_5tuple_protocol {
 #define   TXGBE_ICRMISC_ANDONE MS(19, 0x1) /* link auto-nego done */
 #define   TXGBE_ICRMISC_ERRIG  MS(20, 0x1) /* integrity error */
 #define   TXGBE_ICRMISC_SPI    MS(21, 0x1) /* SPI interface */
-#define   TXGBE_ICRMISC_VFMBX  MS(22, 0x1) /* VF-PF message box */
+#define   TXGBE_ICRMISC_VFMBX  MS(23, 0x1) /* VF-PF message box */
 #define   TXGBE_ICRMISC_GPIO   MS(26, 0x1) /* GPIO interrupt */
 #define   TXGBE_ICRMISC_ERRPCI MS(27, 0x1) /* pcie request error */
 #define   TXGBE_ICRMISC_HEAT   MS(28, 0x1) /* overheat detection */
-- 
2.27.0

[PATCH v3 03/13] net/txgbe: remove outer UDP checksum capability

[Jiawen Wu]

The hardware does not support outer UDP checksum for tunnel packets.
It's wrong to claim this Tx offload capability, so fix it.

Bugzilla ID: 1529
Fixes: b950203be7f1 ("net/txgbe: support VXLAN-GPE")
Fixes: 295968d17407 ("ethdev: add namespace")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/txgbe_rxtx.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
index 5bc0f8772f..c12726553c 100644
--- a/drivers/net/txgbe/txgbe_rxtx.c
+++ b/drivers/net/txgbe/txgbe_rxtx.c
@@ -2284,8 +2284,7 @@ txgbe_get_tx_port_offloads(struct rte_eth_dev *dev)
 
 	tx_offload_capa |= RTE_ETH_TX_OFFLOAD_MACSEC_INSERT;
 
-	tx_offload_capa |= RTE_ETH_TX_OFFLOAD_OUTER_IPV4_CKSUM |
-			   RTE_ETH_TX_OFFLOAD_OUTER_UDP_CKSUM;
+	tx_offload_capa |= RTE_ETH_TX_OFFLOAD_OUTER_IPV4_CKSUM;
 
 #ifdef RTE_LIB_SECURITY
 	if (dev->security_ctx)
-- 
2.27.0

[PATCH v3 04/13] net/txgbe: fix driver load bit to inform firmware

[Jiawen Wu]

Drv_load bit will be reset to default 0 after hardware LAN reset,
reconfigure it to inform firmware that driver is loaded. And set it to 0
when device is closed.

Fixes: b1f596677d8e ("net/txgbe: support device start")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/txgbe_ethdev.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/net/txgbe/txgbe_ethdev.c b/drivers/net/txgbe/txgbe_ethdev.c
index 2834468764..4aa3bfd0bc 100644
--- a/drivers/net/txgbe/txgbe_ethdev.c
+++ b/drivers/net/txgbe/txgbe_ethdev.c
@@ -331,6 +331,8 @@ txgbe_pf_reset_hw(struct txgbe_hw *hw)
 	status = hw->mac.reset_hw(hw);
 
 	ctrl_ext = rd32(hw, TXGBE_PORTCTL);
+	/* let hardware know driver is loaded */
+	ctrl_ext |= TXGBE_PORTCTL_DRVLOAD;
 	/* Set PF Reset Done bit so PF/VF Mail Ops can work */
 	ctrl_ext |= TXGBE_PORTCTL_RSTDONE;
 	wr32(hw, TXGBE_PORTCTL, ctrl_ext);
@@ -2061,6 +2063,9 @@ txgbe_dev_close(struct rte_eth_dev *dev)
 
 	ret = txgbe_dev_stop(dev);
 
+	/* Let firmware take over control of hardware */
+	wr32m(hw, TXGBE_PORTCTL, TXGBE_PORTCTL_DRVLOAD, 0);
+
 	txgbe_dev_free_queues(dev);
 
 	txgbe_set_pcie_master(hw, false);
-- 
2.27.0

[PATCH v3 05/13] net/txgbe: enable Tx descriptor error interrupt

[Jiawen Wu]

Enable Tdm_desc_chk then handle the interrupt of TDM desc error. If it's a
non-fatal error, clear the error to reset the queue. If it's a fatal error,
require users to manually restart the port. This flow prevents the hardware
from PCIe pending due to Tx hang, resulting in a reboot to recover.

But remarkably, when packet layer length does not match the packet type in
TX descriptor, it will cause non-fatal error if Tdm_desc_chk is enabled.
But it can be transmitted normally if Tdm_desc_chk is disabled. So in order
to prevent hardware over check, fix the layer length on the basis of packet
type.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/base/txgbe_regs.h |   6 ++
 drivers/net/txgbe/txgbe_ethdev.c    |  64 ++++++++++++++++++
 drivers/net/txgbe/txgbe_ethdev.h    |   3 +
 drivers/net/txgbe/txgbe_rxtx.c      | 100 ++++++++++++++++++++++++++++
 drivers/net/txgbe/txgbe_rxtx.h      |   1 +
 5 files changed, 174 insertions(+)

diff --git a/drivers/net/txgbe/base/txgbe_regs.h b/drivers/net/txgbe/base/txgbe_regs.h
index b46d65331e..7a9ba6976f 100644
--- a/drivers/net/txgbe/base/txgbe_regs.h
+++ b/drivers/net/txgbe/base/txgbe_regs.h
@@ -1197,6 +1197,7 @@ enum txgbe_5tuple_protocol {
 #define   TXGBE_ICRMISC_ANDONE MS(19, 0x1) /* link auto-nego done */
 #define   TXGBE_ICRMISC_ERRIG  MS(20, 0x1) /* integrity error */
 #define   TXGBE_ICRMISC_SPI    MS(21, 0x1) /* SPI interface */
+#define   TXGBE_ICRMISC_TXDESC MS(22, 0x1) /* TDM desc error */
 #define   TXGBE_ICRMISC_VFMBX  MS(23, 0x1) /* VF-PF message box */
 #define   TXGBE_ICRMISC_GPIO   MS(26, 0x1) /* GPIO interrupt */
 #define   TXGBE_ICRMISC_ERRPCI MS(27, 0x1) /* pcie request error */
@@ -1382,6 +1383,11 @@ enum txgbe_5tuple_protocol {
 #define   TXGBE_TXCFG_WTHRESH(v)        LS(v, 16, 0x7F)
 #define   TXGBE_TXCFG_FLUSH             MS(26, 0x1)
 
+#define TXGBE_TDM_DESC_CHK(i)		(0x0180B0 + (i) * 4) /*0-3*/
+#define TXGBE_TDM_DESC_NONFATAL(i)	(0x0180C0 + (i) * 4) /*0-3*/
+#define TXGBE_TDM_DESC_FATAL(i)		(0x0180D0 + (i) * 4) /*0-3*/
+#define TXGBE_TDM_DESC_MASK(v)		MS(v, 0x1)
+
 /* interrupt registers */
 #define TXGBE_ITRI                      0x000180
 #define TXGBE_ITR(i)                    (0x000200 + 4 * (i))
diff --git a/drivers/net/txgbe/txgbe_ethdev.c b/drivers/net/txgbe/txgbe_ethdev.c
index 4aa3bfd0bc..bafa9cf829 100644
--- a/drivers/net/txgbe/txgbe_ethdev.c
+++ b/drivers/net/txgbe/txgbe_ethdev.c
@@ -1936,6 +1936,7 @@ txgbe_dev_stop(struct rte_eth_dev *dev)
 	PMD_INIT_FUNC_TRACE();
 
 	rte_eal_alarm_cancel(txgbe_dev_detect_sfp, dev);
+	rte_eal_alarm_cancel(txgbe_tx_queue_clear_error, dev);
 	txgbe_dev_wait_setup_link_complete(dev, 0);
 
 	/* disable interrupts */
@@ -2838,6 +2839,60 @@ txgbe_dev_setup_link_alarm_handler(void *param)
 	intr->flags &= ~TXGBE_FLAG_NEED_LINK_CONFIG;
 }
 
+static void
+txgbe_do_reset(struct rte_eth_dev *dev)
+{
+	struct txgbe_hw *hw = TXGBE_DEV_HW(dev);
+	struct txgbe_tx_queue *txq;
+	u32 i;
+
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		txq->resetting = true;
+	}
+
+	rte_delay_ms(1);
+	wr32(hw, TXGBE_RST, TXGBE_RST_LAN(hw->bus.lan_id));
+	txgbe_flush(hw);
+
+	PMD_DRV_LOG(ERR, "Please manually restart the port %d",
+		dev->data->port_id);
+}
+
+static void
+txgbe_tx_ring_recovery(struct rte_eth_dev *dev)
+{
+	struct txgbe_hw *hw = TXGBE_DEV_HW(dev);
+	u32 desc_error[4] = {0, 0, 0, 0};
+	struct txgbe_tx_queue *txq;
+	u32 i;
+
+	/* check tdm fatal error */
+	for (i = 0; i < 4; i++) {
+		desc_error[i] = rd32(hw, TXGBE_TDM_DESC_FATAL(i));
+		if (desc_error[i] != 0) {
+			PMD_DRV_LOG(ERR, "TDM fatal error reg[%d]: 0x%x", i, desc_error[i]);
+			txgbe_do_reset(dev);
+			return;
+		}
+	}
+
+	/* check tdm non-fatal error */
+	for (i = 0; i < 4; i++)
+		desc_error[i] = rd32(hw, TXGBE_TDM_DESC_NONFATAL(i));
+
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		if (desc_error[i / 32] & (1 << i % 32)) {
+			PMD_DRV_LOG(ERR, "TDM non-fatal error, reset port[%d] queue[%d]",
+				dev->data->port_id, i);
+			dev->data->tx_queue_state[i] = RTE_ETH_QUEUE_STATE_STOPPED;
+			txq = dev->data->tx_queues[i];
+			txq->resetting = true;
+			rte_eal_alarm_set(1000, txgbe_tx_queue_clear_error, (void *)dev);
+		}
+	}
+}
+
 /*
  * If @timeout_ms was 0, it means that it will not return until link complete.
  * It returns 1 on complete, return 0 on timeout.
@@ -3096,6 +3151,7 @@ txgbe_dev_misc_interrupt_setup(struct rte_eth_dev *dev)
 	intr->mask |= mask;
 	intr->mask_misc |= TXGBE_ICRMISC_GPIO;
 	intr->mask_misc |= TXGBE_ICRMISC_ANDONE;
+	intr->mask_misc |= TXGBE_ICRMISC_TXDESC;
 	return 0;
 }
 
@@ -3191,6 +3247,9 @@ txgbe_dev_interrupt_get_status(struct rte_eth_dev *dev,
 	if (eicr & TXGBE_ICRMISC_HEAT)
 		intr->flags |= TXGBE_FLAG_OVERHEAT;
 
+	if (eicr & TXGBE_ICRMISC_TXDESC)
+		intr->flags |= TXGBE_FLAG_TX_DESC_ERR;
+
 	((u32 *)hw->isb_mem)[TXGBE_ISB_MISC] = 0;
 
 	return 0;
@@ -3310,6 +3369,11 @@ txgbe_dev_interrupt_action(struct rte_eth_dev *dev,
 		intr->flags &= ~TXGBE_FLAG_OVERHEAT;
 	}
 
+	if (intr->flags & TXGBE_FLAG_TX_DESC_ERR) {
+		txgbe_tx_ring_recovery(dev);
+		intr->flags &= ~TXGBE_FLAG_TX_DESC_ERR;
+	}
+
 	PMD_DRV_LOG(DEBUG, "enable intr immediately");
 	txgbe_enable_intr(dev);
 	rte_intr_enable(intr_handle);
diff --git a/drivers/net/txgbe/txgbe_ethdev.h b/drivers/net/txgbe/txgbe_ethdev.h
index f0f4ced5b0..302ea9f037 100644
--- a/drivers/net/txgbe/txgbe_ethdev.h
+++ b/drivers/net/txgbe/txgbe_ethdev.h
@@ -31,6 +31,7 @@
 #define TXGBE_FLAG_NEED_LINK_CONFIG (uint32_t)(1 << 4)
 #define TXGBE_FLAG_NEED_AN_CONFIG   (uint32_t)(1 << 5)
 #define TXGBE_FLAG_OVERHEAT         (uint32_t)(1 << 6)
+#define TXGBE_FLAG_TX_DESC_ERR      (uint32_t)(1 << 7)
 
 /*
  * Defines that were not part of txgbe_type.h as they are not used by the
@@ -474,6 +475,8 @@ int txgbe_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id);
 
 int txgbe_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id);
 
+void txgbe_tx_queue_clear_error(void *param);
+
 void txgbe_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
 	struct rte_eth_rxq_info *qinfo);
 
diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
index c12726553c..2d2b437643 100644
--- a/drivers/net/txgbe/txgbe_rxtx.c
+++ b/drivers/net/txgbe/txgbe_rxtx.c
@@ -728,6 +728,66 @@ txgbe_get_tun_len(struct rte_mbuf *mbuf)
 	return tun_len;
 }
 
+static inline void
+txgbe_fix_offload_len(union txgbe_tx_offload *ol)
+{
+	uint8_t ptid = ol->ptid;
+
+	if (ptid & TXGBE_PTID_PKT_TUN) {
+		if (ol->outer_l2_len == 0)
+			ol->outer_l2_len = sizeof(struct rte_ether_hdr);
+		if (ol->outer_l3_len == 0) {
+			if (ptid & TXGBE_PTID_TUN_IPV6)
+				ol->outer_l3_len = sizeof(struct rte_ipv6_hdr);
+			else
+				ol->outer_l3_len = sizeof(struct rte_ipv4_hdr);
+		}
+		if ((ptid & 0xF) == 0) {
+			ol->l3_len = 0;
+			ol->l4_len = 0;
+		} else {
+			goto inner;
+		}
+	}
+
+	if ((ptid & 0xF0) == TXGBE_PTID_PKT_MAC) {
+		if (ol->l2_len == 0)
+			ol->l2_len = sizeof(struct rte_ether_hdr);
+		ol->l3_len = 0;
+		ol->l4_len = 0;
+	} else if ((ptid & 0xF0) == TXGBE_PTID_PKT_IP) {
+		if (ol->l2_len == 0)
+			ol->l2_len = sizeof(struct rte_ether_hdr);
+inner:
+		if (ol->l3_len == 0) {
+			if (ptid & TXGBE_PTID_PKT_IPV6)
+				ol->l3_len = sizeof(struct rte_ipv6_hdr);
+			else
+				ol->l3_len = sizeof(struct rte_ipv4_hdr);
+		}
+		switch (ptid & 0x7) {
+		case 0x1:
+		case 0x2:
+			ol->l4_len = 0;
+			break;
+		case 0x3:
+			if (ol->l4_len == 0)
+				ol->l4_len =  sizeof(struct rte_udp_hdr);
+			break;
+		case 0x4:
+			if (ol->l4_len == 0)
+				ol->l4_len =  sizeof(struct rte_tcp_hdr);
+			break;
+		case 0x5:
+			if (ol->l4_len == 0)
+				ol->l4_len =  sizeof(struct rte_sctp_hdr);
+			break;
+		default:
+			break;
+		}
+	}
+}
+
 static inline uint8_t
 txgbe_parse_tun_ptid(struct rte_mbuf *tx_pkt, uint8_t tun_len)
 {
@@ -782,6 +842,10 @@ txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 	uint8_t use_ipsec;
 #endif
 
+	txq = tx_queue;
+	if (txq->resetting)
+		return 0;
+
 	tx_offload.data[0] = 0;
 	tx_offload.data[1] = 0;
 	txq = tx_queue;
@@ -826,6 +890,7 @@ txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 			if (tx_offload.ptid & TXGBE_PTID_PKT_TUN)
 				tx_offload.ptid |= txgbe_parse_tun_ptid(tx_pkt,
 							tx_offload.outer_tun_len);
+			txgbe_fix_offload_len(&tx_offload);
 
 #ifdef RTE_LIB_SECURITY
 			if (use_ipsec) {
@@ -4570,6 +4635,11 @@ txgbe_dev_tx_init(struct rte_eth_dev *dev)
 		wr32(hw, TXGBE_TXWP(txq->reg_idx), 0);
 	}
 
+#ifndef RTE_LIB_SECURITY
+	for (i = 0; i < 4; i++)
+		wr32(hw, TXGBE_TDM_DESC_CHK(i), 0xFFFFFFFF);
+#endif
+
 	/* Device configured with multiple TX queues. */
 	txgbe_dev_mq_tx_configure(dev);
 }
@@ -4806,6 +4876,7 @@ txgbe_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
 	rte_wmb();
 	wr32(hw, TXGBE_TXWP(txq->reg_idx), txq->tx_tail);
 	dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
+	txq->resetting = false;
 
 	return 0;
 }
@@ -4863,6 +4934,35 @@ txgbe_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
 	return 0;
 }
 
+void
+txgbe_tx_queue_clear_error(void *param)
+{
+	struct rte_eth_dev *dev = (struct rte_eth_dev *)param;
+	struct txgbe_hw *hw = TXGBE_DEV_HW(dev);
+	struct txgbe_tx_queue *txq;
+	u32 i;
+
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		if (!txq->resetting)
+			continue;
+
+		txgbe_dev_save_tx_queue(hw, i);
+
+		/* tx ring reset */
+		wr32(hw, TXGBE_TDM_DESC_NONFATAL(i / 32),
+			TXGBE_TDM_DESC_MASK(i % 32));
+
+		if (txq->ops != NULL) {
+			txq->ops->release_mbufs(txq);
+			txq->ops->reset(txq);
+		}
+
+		txgbe_dev_store_tx_queue(hw, i);
+		txgbe_dev_tx_queue_start(dev, i);
+	}
+}
+
 void
 txgbe_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
 	struct rte_eth_rxq_info *qinfo)
diff --git a/drivers/net/txgbe/txgbe_rxtx.h b/drivers/net/txgbe/txgbe_rxtx.h
index 9155eb1f70..e668b60b1e 100644
--- a/drivers/net/txgbe/txgbe_rxtx.h
+++ b/drivers/net/txgbe/txgbe_rxtx.h
@@ -412,6 +412,7 @@ struct txgbe_tx_queue {
 	/**< indicates that IPsec TX feature is in use */
 #endif
 	const struct rte_memzone *mz;
+	bool		    resetting;
 };
 
 struct txgbe_txq_ops {
-- 
2.27.0

[PATCH v3 06/13] net/txgbe: check length of Tx packets

[Jiawen Wu]

Add checking of the Tx packet length to avoid TDM fatal error as far as
possible. Set the pkt_len=1518 for invalid packet in simple Tx code path,
and drop it directly in featured Tx code path.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/txgbe_rxtx.c          | 33 +++++++++++++++++++++++++
 drivers/net/txgbe/txgbe_rxtx_vec_neon.c | 10 +++++---
 drivers/net/txgbe/txgbe_rxtx_vec_sse.c  | 11 ++++++---
 3 files changed, 48 insertions(+), 6 deletions(-)

diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
index 2d2b437643..06acbd0881 100644
--- a/drivers/net/txgbe/txgbe_rxtx.c
+++ b/drivers/net/txgbe/txgbe_rxtx.c
@@ -160,6 +160,8 @@ tx4(volatile struct txgbe_tx_desc *txdp, struct rte_mbuf **pkts)
 	for (i = 0; i < 4; ++i, ++txdp, ++pkts) {
 		buf_dma_addr = rte_mbuf_data_iova(*pkts);
 		pkt_len = (*pkts)->data_len;
+		if (pkt_len < RTE_ETHER_HDR_LEN)
+			pkt_len = TXGBE_FRAME_SIZE_DFT;
 
 		/* write data to descriptor */
 		txdp->qw0 = rte_cpu_to_le_64(buf_dma_addr);
@@ -180,6 +182,8 @@ tx1(volatile struct txgbe_tx_desc *txdp, struct rte_mbuf **pkts)
 
 	buf_dma_addr = rte_mbuf_data_iova(*pkts);
 	pkt_len = (*pkts)->data_len;
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = TXGBE_FRAME_SIZE_DFT;
 
 	/* write data to descriptor */
 	txdp->qw0 = cpu_to_le64(buf_dma_addr);
@@ -813,6 +817,30 @@ txgbe_parse_tun_ptid(struct rte_mbuf *tx_pkt, uint8_t tun_len)
 	return ptid;
 }
 
+static inline bool
+txgbe_check_pkt_err(struct rte_mbuf *tx_pkt)
+{
+	uint32_t total_len = 0, nb_seg = 0;
+	struct rte_mbuf *mseg;
+
+	mseg = tx_pkt;
+	do {
+		if (mseg->data_len == 0)
+			return true;
+		total_len += mseg->data_len;
+		nb_seg++;
+		mseg = mseg->next;
+	} while (mseg != NULL);
+
+	if (tx_pkt->pkt_len != total_len || tx_pkt->pkt_len == 0)
+		return true;
+
+	if (tx_pkt->nb_segs != nb_seg || tx_pkt->nb_segs > 64)
+		return true;
+
+	return false;
+}
+
 uint16_t
 txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 		uint16_t nb_pkts)
@@ -864,6 +892,11 @@ txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
 		new_ctx = 0;
 		tx_pkt = *tx_pkts++;
+		if (txgbe_check_pkt_err(tx_pkt)) {
+			rte_pktmbuf_free(tx_pkt);
+			continue;
+		}
+
 		pkt_len = tx_pkt->pkt_len;
 
 		/*
diff --git a/drivers/net/txgbe/txgbe_rxtx_vec_neon.c b/drivers/net/txgbe/txgbe_rxtx_vec_neon.c
index a96baf9b1d..d4d647fab5 100644
--- a/drivers/net/txgbe/txgbe_rxtx_vec_neon.c
+++ b/drivers/net/txgbe/txgbe_rxtx_vec_neon.c
@@ -476,9 +476,13 @@ static inline void
 vtx1(volatile struct txgbe_tx_desc *txdp,
 		struct rte_mbuf *pkt, uint64_t flags)
 {
-	uint64x2_t descriptor = {
-			pkt->buf_iova + pkt->data_off,
-			(uint64_t)pkt->pkt_len << 45 | flags | pkt->data_len};
+	uint16_t pkt_len = pkt->data_len;
+
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = TXGBE_FRAME_SIZE_DFT;
+
+	uint64x2_t descriptor = {pkt->buf_iova + pkt->data_off,
+				(uint64_t)pkt_len << 45 | flags | pkt_len};
 
 	vst1q_u64((uint64_t *)(uintptr_t)txdp, descriptor);
 }
diff --git a/drivers/net/txgbe/txgbe_rxtx_vec_sse.c b/drivers/net/txgbe/txgbe_rxtx_vec_sse.c
index 1a3f2ce3cd..8ecce33471 100644
--- a/drivers/net/txgbe/txgbe_rxtx_vec_sse.c
+++ b/drivers/net/txgbe/txgbe_rxtx_vec_sse.c
@@ -607,9 +607,14 @@ static inline void
 vtx1(volatile struct txgbe_tx_desc *txdp,
 		struct rte_mbuf *pkt, uint64_t flags)
 {
-	__m128i descriptor = _mm_set_epi64x((uint64_t)pkt->pkt_len << 45 |
-			flags | pkt->data_len,
-			pkt->buf_iova + pkt->data_off);
+	uint16_t pkt_len = pkt->data_len;
+	__m128i descriptor;
+
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = TXGBE_FRAME_SIZE_DFT;
+
+	descriptor = _mm_set_epi64x((uint64_t)pkt_len << 45 | flags | pkt_len,
+				pkt->buf_iova + pkt->data_off);
 	_mm_store_si128((__m128i *)(uintptr_t)txdp, descriptor);
 }
 
-- 
2.27.0

[PATCH v3 07/13] net/txgbe: add Tx descriptor error statistics

[Jiawen Wu]

Count the number of packets not sent due to Tx descriptor error.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/txgbe/txgbe_ethdev.c | 13 +++++++++++++
 drivers/net/txgbe/txgbe_rxtx.c   |  6 ++++++
 drivers/net/txgbe/txgbe_rxtx.h   |  1 +
 3 files changed, 20 insertions(+)

diff --git a/drivers/net/txgbe/txgbe_ethdev.c b/drivers/net/txgbe/txgbe_ethdev.c
index bafa9cf829..a956216abb 100644
--- a/drivers/net/txgbe/txgbe_ethdev.c
+++ b/drivers/net/txgbe/txgbe_ethdev.c
@@ -2344,6 +2344,7 @@ txgbe_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 	struct txgbe_hw_stats *hw_stats = TXGBE_DEV_STATS(dev);
 	struct txgbe_stat_mappings *stat_mappings =
 			TXGBE_DEV_STAT_MAPPINGS(dev);
+	struct txgbe_tx_queue *txq;
 	uint32_t i, j;
 
 	txgbe_read_stats_registers(hw, hw_stats);
@@ -2398,6 +2399,11 @@ txgbe_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 
 	/* Tx Errors */
 	stats->oerrors  = 0;
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		stats->oerrors += txq->desc_error;
+	}
+
 	return 0;
 }
 
@@ -2406,6 +2412,13 @@ txgbe_dev_stats_reset(struct rte_eth_dev *dev)
 {
 	struct txgbe_hw *hw = TXGBE_DEV_HW(dev);
 	struct txgbe_hw_stats *hw_stats = TXGBE_DEV_STATS(dev);
+	struct txgbe_tx_queue *txq;
+	uint32_t i;
+
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		txq->desc_error = 0;
+	}
 
 	/* HW registers are cleared on read */
 	hw->offset_loaded = 0;
diff --git a/drivers/net/txgbe/txgbe_rxtx.c b/drivers/net/txgbe/txgbe_rxtx.c
index 06acbd0881..4e4b78fb43 100644
--- a/drivers/net/txgbe/txgbe_rxtx.c
+++ b/drivers/net/txgbe/txgbe_rxtx.c
@@ -894,6 +894,7 @@ txgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 		tx_pkt = *tx_pkts++;
 		if (txgbe_check_pkt_err(tx_pkt)) {
 			rte_pktmbuf_free(tx_pkt);
+			txq->desc_error++;
 			continue;
 		}
 
@@ -2523,6 +2524,7 @@ txgbe_dev_tx_queue_setup(struct rte_eth_dev *dev,
 	txgbe_set_tx_function(dev, txq);
 
 	txq->ops->reset(txq);
+	txq->desc_error = 0;
 
 	dev->data->tx_queues[queue_idx] = txq;
 
@@ -4980,6 +4982,10 @@ txgbe_tx_queue_clear_error(void *param)
 		if (!txq->resetting)
 			continue;
 
+		/* Increase the count of Tx desc error since
+		 * it causes the queue reset.
+		 */
+		txq->desc_error++;
 		txgbe_dev_save_tx_queue(hw, i);
 
 		/* tx ring reset */
diff --git a/drivers/net/txgbe/txgbe_rxtx.h b/drivers/net/txgbe/txgbe_rxtx.h
index e668b60b1e..622a0d3981 100644
--- a/drivers/net/txgbe/txgbe_rxtx.h
+++ b/drivers/net/txgbe/txgbe_rxtx.h
@@ -412,6 +412,7 @@ struct txgbe_tx_queue {
 	/**< indicates that IPsec TX feature is in use */
 #endif
 	const struct rte_memzone *mz;
+	uint64_t	    desc_error;
 	bool		    resetting;
 };
 
-- 
2.27.0

[PATCH v3 08/13] net/ngbe: check length of Tx packets

[Jiawen Wu]

Add checking of the Tx packet length to avoid TDM fatal error as far as
possible. Set the pkt_len=1518 for invalid packet in simple Tx code path,
and drop it directly in featured Tx code path. Althrough the hardware
does not support TDM desc check.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_rxtx.c          | 33 +++++++++++++++++++++++++++
 drivers/net/ngbe/ngbe_rxtx_vec_neon.c | 10 +++++---
 drivers/net/ngbe/ngbe_rxtx_vec_sse.c  | 11 ++++++---
 3 files changed, 48 insertions(+), 6 deletions(-)

diff --git a/drivers/net/ngbe/ngbe_rxtx.c b/drivers/net/ngbe/ngbe_rxtx.c
index f3eb797d0c..25a314f10a 100644
--- a/drivers/net/ngbe/ngbe_rxtx.c
+++ b/drivers/net/ngbe/ngbe_rxtx.c
@@ -113,6 +113,8 @@ tx4(volatile struct ngbe_tx_desc *txdp, struct rte_mbuf **pkts)
 	for (i = 0; i < 4; ++i, ++txdp, ++pkts) {
 		buf_dma_addr = rte_mbuf_data_iova(*pkts);
 		pkt_len = (*pkts)->data_len;
+		if (pkt_len < RTE_ETHER_HDR_LEN)
+			pkt_len = NGBE_FRAME_SIZE_DFT;
 
 		/* write data to descriptor */
 		txdp->qw0 = rte_cpu_to_le_64(buf_dma_addr);
@@ -133,6 +135,8 @@ tx1(volatile struct ngbe_tx_desc *txdp, struct rte_mbuf **pkts)
 
 	buf_dma_addr = rte_mbuf_data_iova(*pkts);
 	pkt_len = (*pkts)->data_len;
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = NGBE_FRAME_SIZE_DFT;
 
 	/* write data to descriptor */
 	txdp->qw0 = cpu_to_le64(buf_dma_addr);
@@ -555,6 +559,30 @@ ngbe_xmit_cleanup(struct ngbe_tx_queue *txq)
 	return 0;
 }
 
+static inline bool
+ngbe_check_pkt_err(struct rte_mbuf *tx_pkt)
+{
+	uint32_t total_len = 0, nb_seg = 0;
+	struct rte_mbuf *mseg;
+
+	mseg = tx_pkt;
+	do {
+		if (mseg->data_len == 0)
+			return true;
+		total_len += mseg->data_len;
+		nb_seg++;
+		mseg = mseg->next;
+	} while (mseg != NULL);
+
+	if (tx_pkt->pkt_len != total_len || tx_pkt->pkt_len == 0)
+		return true;
+
+	if (tx_pkt->nb_segs != nb_seg || tx_pkt->nb_segs > 64)
+		return true;
+
+	return false;
+}
+
 uint16_t
 ngbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 		uint16_t nb_pkts)
@@ -599,6 +627,11 @@ ngbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
 		new_ctx = 0;
 		tx_pkt = *tx_pkts++;
+		if (ngbe_check_pkt_err(tx_pkt)) {
+			rte_pktmbuf_free(tx_pkt);
+			continue;
+		}
+
 		pkt_len = tx_pkt->pkt_len;
 
 		/*
diff --git a/drivers/net/ngbe/ngbe_rxtx_vec_neon.c b/drivers/net/ngbe/ngbe_rxtx_vec_neon.c
index dcf12b7070..37075ea5e7 100644
--- a/drivers/net/ngbe/ngbe_rxtx_vec_neon.c
+++ b/drivers/net/ngbe/ngbe_rxtx_vec_neon.c
@@ -476,9 +476,13 @@ static inline void
 vtx1(volatile struct ngbe_tx_desc *txdp,
 		struct rte_mbuf *pkt, uint64_t flags)
 {
-	uint64x2_t descriptor = {
-			pkt->buf_iova + pkt->data_off,
-			(uint64_t)pkt->pkt_len << 45 | flags | pkt->data_len};
+	uint16_t pkt_len = pkt->data_len;
+
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = NGBE_FRAME_SIZE_DFT;
+
+	uint64x2_t descriptor = {pkt->buf_iova + pkt->data_off,
+				(uint64_t)pkt_len << 45 | flags | pkt_len};
 
 	vst1q_u64((uint64_t *)(uintptr_t)txdp, descriptor);
 }
diff --git a/drivers/net/ngbe/ngbe_rxtx_vec_sse.c b/drivers/net/ngbe/ngbe_rxtx_vec_sse.c
index b128bd3a67..19c69cdfa6 100644
--- a/drivers/net/ngbe/ngbe_rxtx_vec_sse.c
+++ b/drivers/net/ngbe/ngbe_rxtx_vec_sse.c
@@ -563,9 +563,14 @@ static inline void
 vtx1(volatile struct ngbe_tx_desc *txdp,
 		struct rte_mbuf *pkt, uint64_t flags)
 {
-	__m128i descriptor = _mm_set_epi64x((uint64_t)pkt->pkt_len << 45 |
-			flags | pkt->data_len,
-			pkt->buf_iova + pkt->data_off);
+	uint16_t pkt_len = pkt->data_len;
+	__m128i descriptor;
+
+	if (pkt_len < RTE_ETHER_HDR_LEN)
+		pkt_len = NGBE_FRAME_SIZE_DFT;
+
+	descriptor = _mm_set_epi64x((uint64_t)pkt_len << 45 | flags | pkt_len,
+				pkt->buf_iova + pkt->data_off);
 	_mm_store_si128((__m128i *)(uintptr_t)txdp, descriptor);
 }
 
-- 
2.27.0

[PATCH v3 09/13] net/ngbe: add Tx descriptor error statistics

[Jiawen Wu]

Count the number of packets not sent due to Tx descriptor error.

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_ethdev.c | 13 +++++++++++++
 drivers/net/ngbe/ngbe_rxtx.c   |  2 ++
 drivers/net/ngbe/ngbe_rxtx.h   |  1 +
 3 files changed, 16 insertions(+)

diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index 6c45ffaad3..353d17acc8 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -1507,6 +1507,7 @@ ngbe_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 	struct ngbe_hw_stats *hw_stats = NGBE_DEV_STATS(dev);
 	struct ngbe_stat_mappings *stat_mappings =
 			NGBE_DEV_STAT_MAPPINGS(dev);
+	struct ngbe_tx_queue *txq;
 	uint32_t i, j;
 
 	ngbe_read_stats_registers(hw, hw_stats);
@@ -1559,6 +1560,11 @@ ngbe_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 
 	/* Tx Errors */
 	stats->oerrors  = 0;
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		stats->oerrors += txq->desc_error;
+	}
+
 	return 0;
 }
 
@@ -1567,6 +1573,13 @@ ngbe_dev_stats_reset(struct rte_eth_dev *dev)
 {
 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
 	struct ngbe_hw_stats *hw_stats = NGBE_DEV_STATS(dev);
+	struct ngbe_tx_queue *txq;
+	uint32_t i;
+
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		txq->desc_error = 0;
+	}
 
 	/* HW registers are cleared on read */
 	hw->offset_loaded = 0;
diff --git a/drivers/net/ngbe/ngbe_rxtx.c b/drivers/net/ngbe/ngbe_rxtx.c
index 25a314f10a..8d31d47de9 100644
--- a/drivers/net/ngbe/ngbe_rxtx.c
+++ b/drivers/net/ngbe/ngbe_rxtx.c
@@ -629,6 +629,7 @@ ngbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 		tx_pkt = *tx_pkts++;
 		if (ngbe_check_pkt_err(tx_pkt)) {
 			rte_pktmbuf_free(tx_pkt);
+			txq->desc_error++;
 			continue;
 		}
 
@@ -2100,6 +2101,7 @@ ngbe_dev_tx_queue_setup(struct rte_eth_dev *dev,
 	ngbe_set_tx_function(dev, txq);
 
 	txq->ops->reset(txq);
+	txq->desc_error = 0;
 
 	dev->data->tx_queues[queue_idx] = txq;
 
diff --git a/drivers/net/ngbe/ngbe_rxtx.h b/drivers/net/ngbe/ngbe_rxtx.h
index 7574db32d8..8534ec123a 100644
--- a/drivers/net/ngbe/ngbe_rxtx.h
+++ b/drivers/net/ngbe/ngbe_rxtx.h
@@ -375,6 +375,7 @@ struct ngbe_tx_queue {
 
 	const struct ngbe_txq_ops *ops;       /**< txq ops */
 	const struct rte_memzone *mz;
+	uint64_t            desc_error;
 };
 
 struct ngbe_txq_ops {
-- 
2.27.0

[PATCH v3 10/13] net/ngbe: fix driver load bit to inform firmware

[Jiawen Wu]

Drv_load bit will be reset to default 0 after hardware LAN reset,
reconfigure it to inform firmware that driver is loaded. And set it to 0
when device is closed.

Fixes: 3518df5774c7 ("net/ngbe: support device start/stop")
Fixes: cc63194e89cb ("net/ngbe: support close and reset device")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_ethdev.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index 353d17acc8..238533f2b8 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -263,6 +263,8 @@ ngbe_pf_reset_hw(struct ngbe_hw *hw)
 	status = hw->mac.reset_hw(hw);
 
 	ctrl_ext = rd32(hw, NGBE_PORTCTL);
+	/* let hardware know driver is loaded */
+	ctrl_ext |= NGBE_PORTCTL_DRVLOAD;
 	/* Set PF Reset Done bit so PF/VF Mail Ops can work */
 	ctrl_ext |= NGBE_PORTCTL_RSTDONE;
 	wr32(hw, NGBE_PORTCTL, ctrl_ext);
@@ -1277,6 +1279,9 @@ ngbe_dev_close(struct rte_eth_dev *dev)
 
 	ngbe_dev_stop(dev);
 
+	/* Let firmware take over control of hardware */
+	wr32m(hw, NGBE_PORTCTL, NGBE_PORTCTL_DRVLOAD, 0);
+
 	ngbe_dev_free_queues(dev);
 
 	ngbe_set_pcie_master(hw, false);
-- 
2.27.0

[PATCH v3 11/13] net/ngbe: reconfigure more MAC Rx registers

[Jiawen Wu]

When link status changes, there is a probability that no more packets
can be received on the port, due to hardware defects. These MAC Rx
registers should be reconfigured to fix this problem.

Fixes: b9246b8fa280 ("net/ngbe: support link update")
Fixes: a7c5f95ed9c2 ("net/ngbe: reconfigure MAC Rx when link update")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/base/ngbe_regs.h | 2 ++
 drivers/net/ngbe/ngbe_ethdev.c    | 6 ++++++
 2 files changed, 8 insertions(+)

diff --git a/drivers/net/ngbe/base/ngbe_regs.h b/drivers/net/ngbe/base/ngbe_regs.h
index 8a6776b0e6..b1295280a7 100644
--- a/drivers/net/ngbe/base/ngbe_regs.h
+++ b/drivers/net/ngbe/base/ngbe_regs.h
@@ -712,6 +712,8 @@ enum ngbe_5tuple_protocol {
 #define   NGBE_MACRXFLT_CTL_PASS       LS(3, 6, 0x3)
 #define   NGBE_MACRXFLT_RXALL          MS(31, 0x1)
 
+#define NGBE_MAC_WDG_TIMEOUT           0x01100C
+
 /******************************************************************************
  * Statistic Registers
  ******************************************************************************/
diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index 238533f2b8..c372fd928c 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -1941,6 +1941,7 @@ ngbe_dev_link_update_share(struct rte_eth_dev *dev,
 	bool link_up;
 	int err;
 	int wait = 1;
+	u32 reg;
 
 	memset(&link, 0, sizeof(link));
 	link.link_status = RTE_ETH_LINK_DOWN;
@@ -1998,8 +1999,13 @@ ngbe_dev_link_update_share(struct rte_eth_dev *dev,
 			wr32m(hw, NGBE_MACTXCFG, NGBE_MACTXCFG_SPEED_MASK,
 				NGBE_MACTXCFG_SPEED_1G | NGBE_MACTXCFG_TE);
 		}
+		/* Re configure MAC RX */
+		reg = rd32(hw, NGBE_MACRXCFG);
+		wr32(hw, NGBE_MACRXCFG, reg);
 		wr32m(hw, NGBE_MACRXFLT, NGBE_MACRXFLT_PROMISC,
 			NGBE_MACRXFLT_PROMISC);
+		reg = rd32(hw, NGBE_MAC_WDG_TIMEOUT);
+		wr32(hw, NGBE_MAC_WDG_TIMEOUT, reg);
 	}
 
 	return rte_eth_linkstatus_set(dev, &link);
-- 
2.27.0

[PATCH v3 12/13] net/ngbe: fix interrupt lost in legacy or MSI mode

[Jiawen Wu]

When interrupt is legacy or MSI mode, shared interrupt may cause the
interrupt cannot be re-enabled. So fix to read the shared interrupt.

Fixes: b9246b8fa280 ("net/ngbe: support link update")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_ethdev.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index c372fd928c..325a9d1eaf 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -2193,6 +2193,19 @@ ngbe_dev_interrupt_get_status(struct rte_eth_dev *dev)
 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
 	struct ngbe_interrupt *intr = ngbe_dev_intr(dev);
 
+	eicr = ((u32 *)hw->isb_mem)[NGBE_ISB_VEC0];
+	if (!eicr) {
+		/*
+		 * shared interrupt alert!
+		 * make sure interrupts are enabled because the read will
+		 * have disabled interrupts.
+		 */
+		if (!hw->adapter_stopped)
+			ngbe_enable_intr(dev);
+		return 0;
+	}
+	((u32 *)hw->isb_mem)[NGBE_ISB_VEC0] = 0;
+
 	/* read-on-clear nic registers here */
 	eicr = ((u32 *)hw->isb_mem)[NGBE_ISB_MISC];
 	PMD_DRV_LOG(DEBUG, "eicr %x", eicr);
-- 
2.27.0

[PATCH v3 13/13] net/ngbe: restrict configuration of VLAN strip offload

[Jiawen Wu]

There is a hardware limitation that Rx ring config register is not
writable when Rx ring is enabled, i.e. the NGBE_RXCFG_ENA bit is set.
But disabling the ring when there is traffic will cause ring get stuck.
So restrict the configuration of VLAN strip offload only if device is
started.

Fixes: 59b46438fdaa ("net/ngbe: support VLAN offload and VLAN filter")
Cc: stable@dpdk.org

Signed-off-by: Jiawen Wu <jiawenwu@trustnetic.com>
---
 drivers/net/ngbe/ngbe_ethdev.c | 49 ++++++++++++++--------------------
 1 file changed, 20 insertions(+), 29 deletions(-)

diff --git a/drivers/net/ngbe/ngbe_ethdev.c b/drivers/net/ngbe/ngbe_ethdev.c
index 325a9d1eaf..08e87471f6 100644
--- a/drivers/net/ngbe/ngbe_ethdev.c
+++ b/drivers/net/ngbe/ngbe_ethdev.c
@@ -586,41 +586,25 @@ ngbe_vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
 }
 
 static void
-ngbe_vlan_strip_queue_set(struct rte_eth_dev *dev, uint16_t queue, int on)
+ngbe_vlan_strip_q_set(struct rte_eth_dev *dev, uint16_t queue, int on)
 {
-	struct ngbe_hw *hw = ngbe_dev_hw(dev);
-	struct ngbe_rx_queue *rxq;
-	bool restart;
-	uint32_t rxcfg, rxbal, rxbah;
-
 	if (on)
 		ngbe_vlan_hw_strip_enable(dev, queue);
 	else
 		ngbe_vlan_hw_strip_disable(dev, queue);
+}
 
-	rxq = dev->data->rx_queues[queue];
-	rxbal = rd32(hw, NGBE_RXBAL(rxq->reg_idx));
-	rxbah = rd32(hw, NGBE_RXBAH(rxq->reg_idx));
-	rxcfg = rd32(hw, NGBE_RXCFG(rxq->reg_idx));
-	if (rxq->offloads & RTE_ETH_RX_OFFLOAD_VLAN_STRIP) {
-		restart = (rxcfg & NGBE_RXCFG_ENA) &&
-			!(rxcfg & NGBE_RXCFG_VLAN);
-		rxcfg |= NGBE_RXCFG_VLAN;
-	} else {
-		restart = (rxcfg & NGBE_RXCFG_ENA) &&
-			(rxcfg & NGBE_RXCFG_VLAN);
-		rxcfg &= ~NGBE_RXCFG_VLAN;
-	}
-	rxcfg &= ~NGBE_RXCFG_ENA;
+static void
+ngbe_vlan_strip_queue_set(struct rte_eth_dev *dev, uint16_t queue, int on)
+{
+	struct ngbe_hw *hw = ngbe_dev_hw(dev);
 
-	if (restart) {
-		/* set vlan strip for ring */
-		ngbe_dev_rx_queue_stop(dev, queue);
-		wr32(hw, NGBE_RXBAL(rxq->reg_idx), rxbal);
-		wr32(hw, NGBE_RXBAH(rxq->reg_idx), rxbah);
-		wr32(hw, NGBE_RXCFG(rxq->reg_idx), rxcfg);
-		ngbe_dev_rx_queue_start(dev, queue);
+	if (!hw->adapter_stopped) {
+		PMD_DRV_LOG(ERR, "Please stop port first");
+		return;
 	}
+
+	ngbe_vlan_strip_q_set(dev, queue, on);
 }
 
 static int
@@ -846,9 +830,9 @@ ngbe_vlan_hw_strip_config(struct rte_eth_dev *dev)
 		rxq = dev->data->rx_queues[i];
 
 		if (rxq->offloads & RTE_ETH_RX_OFFLOAD_VLAN_STRIP)
-			ngbe_vlan_hw_strip_enable(dev, i);
+			ngbe_vlan_strip_q_set(dev, i, 1);
 		else
-			ngbe_vlan_hw_strip_disable(dev, i);
+			ngbe_vlan_strip_q_set(dev, i, 0);
 	}
 }
 
@@ -910,6 +894,13 @@ ngbe_vlan_offload_config(struct rte_eth_dev *dev, int mask)
 static int
 ngbe_vlan_offload_set(struct rte_eth_dev *dev, int mask)
 {
+	struct ngbe_hw *hw = ngbe_dev_hw(dev);
+
+	if (!hw->adapter_stopped && (mask & RTE_ETH_VLAN_STRIP_MASK)) {
+		PMD_DRV_LOG(ERR, "Please stop port first");
+		return -EPERM;
+	}
+
 	ngbe_config_vlan_strip_on_all_queues(dev, mask);
 
 	ngbe_vlan_offload_config(dev, mask);
-- 
2.27.0

Re: [PATCH v3 00/13] Wangxun fixes

[Ferruh Yigit]

On 11/4/2024 2:29 AM, Jiawen Wu wrote:
> Fix some bugs in txgbe/ngbe, and optimize the Tx flow.
> 
> v2 -> v3:
> - Reset Tx desc error statistics in .stats_reset().
> - Add comment for Tx desc error count.
> 
> v1 -> v2:
> - Fix build errors on ARM platform.
> 
> Jiawen Wu (13):
>   net/txgbe: fix swfw mbox failure
>   net/txgbe: fix VF-PF mbox interrupt
>   net/txgbe: remove outer UDP checksum capability
>   net/txgbe: fix driver load bit to inform firmware
>   net/txgbe: enable Tx descriptor error interrupt
>   net/txgbe: check length of Tx packets
>   net/txgbe: add Tx descriptor error statistics
>   net/ngbe: check length of Tx packets
>   net/ngbe: add Tx descriptor error statistics
>   net/ngbe: fix driver load bit to inform firmware
>   net/ngbe: reconfigure more MAC Rx registers
>   net/ngbe: fix interrupt lost in legacy or MSI mode
>   net/ngbe: restrict configuration of VLAN strip offload
>

Series applied to dpdk-next-net/main, thanks.