[dpdk-dev] [v7] net/e1000: fix i219 hang on reset/close

[Ye Xiaolong]

On 07/22, Xiao Zhang wrote:
>Unit hang may occur if multiple descriptors are available in the rings
>during reset or close. This state can be detected by configure status
>by bit 8 in register. If the bit is set and there are pending
>descriptors in one of the rings, we must flush them before reset or
>close.
>
>Fixes: 80580344("e1000: support EM devices (also known as e1000/e1000e)")
>Cc: stable at dpdk.org
>
>Signed-off-by: Xiao Zhang <xiao.zhang at intel.com>
>---
>v7 Add fix line.
>v6 Change the fix on em driver instead of igb driver and update the 
>register address according to C-Spec.
>v5 Change the subject.
>v4 Correct the tail descriptor of tx ring.
>v3 Add loop to handle all tx and rx queues.
>v2 Use configuration register instead of NVM7 to get the hang state.
>---
> drivers/net/e1000/e1000_ethdev.h |   4 ++
> drivers/net/e1000/em_ethdev.c    |   5 ++
> drivers/net/e1000/em_rxtx.c      | 108 +++++++++++++++++++++++++++++++++++++++
> 3 files changed, 117 insertions(+)
>
>diff --git a/drivers/net/e1000/e1000_ethdev.h b/drivers/net/e1000/e1000_ethdev.h
>index 67acb73..01ff943 100644
>--- a/drivers/net/e1000/e1000_ethdev.h
>+++ b/drivers/net/e1000/e1000_ethdev.h
>@@ -35,6 +35,9 @@
> #define IGB_MAX_RX_QUEUE_NUM           8
> #define IGB_MAX_RX_QUEUE_NUM_82576     16
> 
>+#define E1000_I219_MAX_RX_QUEUE_NUM		2
>+#define E1000_I219_MAX_TX_QUEUE_NUM		2
>+
> #define E1000_SYN_FILTER_ENABLE        0x00000001 /* syn filter enable field */
> #define E1000_SYN_FILTER_QUEUE         0x0000000E /* syn filter queue field */
> #define E1000_SYN_FILTER_QUEUE_SHIFT   1          /* syn filter queue field */
>@@ -522,5 +525,6 @@ int igb_action_rss_same(const struct rte_flow_action_rss *comp,
> int igb_config_rss_filter(struct rte_eth_dev *dev,
> 			struct igb_rte_flow_rss_conf *conf,
> 			bool add);
>+void em_flush_desc_rings(struct rte_eth_dev *dev);
> 
> #endif /* _E1000_ETHDEV_H_ */
>diff --git a/drivers/net/e1000/em_ethdev.c b/drivers/net/e1000/em_ethdev.c
>index dc88661..62d3a95 100644
>--- a/drivers/net/e1000/em_ethdev.c
>+++ b/drivers/net/e1000/em_ethdev.c
>@@ -738,6 +738,11 @@ eth_em_stop(struct rte_eth_dev *dev)
> 	em_lsc_intr_disable(hw);
> 
> 	e1000_reset_hw(hw);
>+
>+	/* Flush desc rings for i219 */
>+	if (hw->mac.type >= e1000_pch_spt)
>+		em_flush_desc_rings(dev);
>+
> 	if (hw->mac.type >= e1000_82544)
> 		E1000_WRITE_REG(hw, E1000_WUC, 0);
> 
>diff --git a/drivers/net/e1000/em_rxtx.c b/drivers/net/e1000/em_rxtx.c
>index 708f832..96c10cd 100644
>--- a/drivers/net/e1000/em_rxtx.c
>+++ b/drivers/net/e1000/em_rxtx.c
>@@ -18,6 +18,7 @@
> #include <rte_log.h>
> #include <rte_debug.h>
> #include <rte_pci.h>
>+#include <rte_bus_pci.h>
> #include <rte_memory.h>
> #include <rte_memcpy.h>
> #include <rte_memzone.h>
>@@ -59,6 +60,11 @@
> #define E1000_TX_OFFLOAD_NOTSUP_MASK \
> 		(PKT_TX_OFFLOAD_MASK ^ E1000_TX_OFFLOAD_MASK)
> 
>+/* PCI offset for querying configuration status register */
>+#define PCI_CFG_STATUS_REG                 0x06
>+#define FLUSH_DESC_REQUIRED               0x100
>+
>+
> /**
>  * Structure associated with each descriptor of the RX ring of a RX queue.
>  */
>@@ -2000,3 +2006,105 @@ em_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
> 	qinfo->conf.tx_rs_thresh = txq->tx_rs_thresh;
> 	qinfo->conf.offloads = txq->offloads;
> }
>+
>+static void e1000_flush_tx_ring(struct rte_eth_dev *dev)

Minor nit, according to dpdk community's coding style, the function type should
be on a line by itself preceding the function, like

static void
e1000_flush_tx_ring(struct rte_eth_dev *dev)

>+{
>+	struct e1000_hw *hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
>+	volatile struct e1000_data_desc *tx_desc;
>+	volatile uint32_t *tdt_reg_addr;
>+	uint32_t tdt, tctl, txd_lower = E1000_TXD_CMD_IFCS;
>+	uint16_t size = 512;
>+	struct em_tx_queue *txq;
>+	int i;
>+
>+	if (dev->data->tx_queues == NULL)
>+		return;
>+	tctl = E1000_READ_REG(hw, E1000_TCTL);
>+	E1000_WRITE_REG(hw, E1000_TCTL, tctl | E1000_TCTL_EN);
>+	for (i = 0; i < dev->data->nb_tx_queues &&
>+		i < E1000_I219_MAX_TX_QUEUE_NUM; i++) {
>+		txq = dev->data->tx_queues[i];
>+		tdt = E1000_READ_REG(hw, E1000_TDT(i));
>+		if (tdt != txq->tx_tail)
>+			return;
>+		tx_desc = &txq->tx_ring[txq->tx_tail];
>+		tx_desc->buffer_addr = rte_cpu_to_le_64(txq->tx_ring_phys_addr);
>+		tx_desc->lower.data = rte_cpu_to_le_32(txd_lower | size);
>+		tx_desc->upper.data = 0;
>+
>+		rte_wmb();
>+		txq->tx_tail++;
>+		if (txq->tx_tail == txq->nb_tx_desc)
>+			txq->tx_tail = 0;
>+		rte_io_wmb();
>+		tdt_reg_addr = E1000_PCI_REG_ADDR(hw, E1000_TDT(i));
>+		E1000_PCI_REG_WRITE_RELAXED(tdt_reg_addr, txq->tx_tail);
>+		usec_delay(250);
>+	}
>+}
>+
>+static void e1000_flush_rx_ring(struct rte_eth_dev *dev)

Ditto.

>+{
>+	uint32_t rctl, rxdctl;
>+	struct e1000_hw *hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
>+	int i;
>+
>+	rctl = E1000_READ_REG(hw, E1000_RCTL);
>+	E1000_WRITE_REG(hw, E1000_RCTL, rctl & ~E1000_RCTL_EN);
>+	E1000_WRITE_FLUSH(hw);
>+	usec_delay(150);
>+
>+	for (i = 0; i < dev->data->nb_rx_queues &&
>+		i < E1000_I219_MAX_RX_QUEUE_NUM; i++) {
>+		rxdctl = E1000_READ_REG(hw, E1000_RXDCTL(i));
>+		/* zero the lower 14 bits (prefetch and host thresholds) */
>+		rxdctl &= 0xffffc000;
>+
>+		/* update thresholds: prefetch threshold to 31,
>+		 * host threshold to 1 and make sure the granularity
>+		 * is "descriptors" and not "cache lines"
>+		 */
>+		rxdctl |= (0x1F | (1UL << 8) | E1000_RXDCTL_THRESH_UNIT_DESC);
>+
>+		E1000_WRITE_REG(hw, E1000_RXDCTL(i), rxdctl);
>+	}
>+	/* momentarily enable the RX ring for the changes to take effect */
>+	E1000_WRITE_REG(hw, E1000_RCTL, rctl | E1000_RCTL_EN);
>+	E1000_WRITE_FLUSH(hw);
>+	usec_delay(150);
>+	E1000_WRITE_REG(hw, E1000_RCTL, rctl & ~E1000_RCTL_EN);
>+}
>+
>+/**
>+ * em_flush_desc_rings - remove all descriptors from the descriptor rings
>+ *
>+ * In i219, the descriptor rings must be emptied before resetting/closing the
>+ * HW. Failure to do this will cause the HW to enter a unit hang state which
>+ * can only be released by PCI reset on the device
>+ *
>+ */
>+
>+void em_flush_desc_rings(struct rte_eth_dev *dev)

Ditto.


For the rest part, 
Reviewed-by: Xiaolong Ye <xiaolong.ye at intel.com>

>+{
>+	uint32_t fextnvm11, tdlen;
>+	struct e1000_hw *hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
>+	struct rte_pci_device *pci_dev = RTE_ETH_DEV_TO_PCI(dev);
>+	uint16_t pci_cfg_status = 0;
>+
>+	fextnvm11 = E1000_READ_REG(hw, E1000_FEXTNVM11);
>+	E1000_WRITE_REG(hw, E1000_FEXTNVM11,
>+			fextnvm11 | E1000_FEXTNVM11_DISABLE_MULR_FIX);
>+	tdlen = E1000_READ_REG(hw, E1000_TDLEN(0));
>+	rte_pci_read_config(pci_dev, &pci_cfg_status, sizeof(pci_cfg_status),
>+				PCI_CFG_STATUS_REG);
>+
>+	/* do nothing if we're not in faulty state, or if the queue is empty */
>+	if ((pci_cfg_status & FLUSH_DESC_REQUIRED) && tdlen) {
>+		/* flush desc ring */
>+		e1000_flush_tx_ring(dev);
>+		rte_pci_read_config(pci_dev, &pci_cfg_status,
>+				sizeof(pci_cfg_status), PCI_CFG_STATUS_REG);
>+		if (pci_cfg_status & FLUSH_DESC_REQUIRED)
>+			e1000_flush_rx_ring(dev);
>+	}
>+}
>-- 
>2.7.4
>