[PATCH net 0/5] dwc_eth_qos: stability fixes and support for CMA

[PATCH net 0/5] dwc_eth_qos: stability fixes and support for CMA

[Lars Persson]

This series has bug fixes for the dwc_eth_qos ethernet driver.

Mainly two stability fixes for problems found by Rabin Vincent:
- Successive starts and stops of the interface would trigger a DMA reset timeout.
- A race condition in the TX DMA handling could trigger a netdev watchdog
 timeout.

The memory allocation was improved to support use of the CMA as DMA allocator
backend.

Lars Persson (2):
  dwc_eth_qos: release descriptors outside netif_tx_lock
  dwc_eth_qos: do phy_start before resetting hardware

Rabin Vincent (3):
  dwc_eth_qos: fix race condition in dwceqos_start_xmit
  dwc_eth_qos: use GFP_KERNEL in dma_alloc_coherent()
  dwc_eth_qos: use DWCEQOS_MSG_DEFAULT

 drivers/net/ethernet/synopsys/dwc_eth_qos.c | 45 ++++++++++++++++++++---------
 1 file changed, 31 insertions(+), 14 deletions(-)

-- 
2.1.4

[PATCH net 1/5] dwc_eth_qos: fix race condition in dwceqos_start_xmit

[Lars Persson]

From: Rabin Vincent <rabinv@axis.com>

The xmit handler and the tx_reclaim tasklet had a race on the tx_free
variable which could lead to a tx timeout if tx_free was updated after
the tx complete interrupt.

Signed-off-by: Rabin Vincent <rabinv@axis.com>
Signed-off-by: Lars Persson <larper@axis.com>
---
 drivers/net/ethernet/synopsys/dwc_eth_qos.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/drivers/net/ethernet/synopsys/dwc_eth_qos.c b/drivers/net/ethernet/synopsys/dwc_eth_qos.c
index fc8bbff..926db2d 100644
--- a/drivers/net/ethernet/synopsys/dwc_eth_qos.c
+++ b/drivers/net/ethernet/synopsys/dwc_eth_qos.c
@@ -2178,12 +2178,10 @@ static int dwceqos_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 		((trans.initial_descriptor + trans.nr_descriptors) %
 		 DWCEQOS_TX_DCNT));
 
-	dwceqos_tx_finalize(skb, lp, &trans);
-
-	netdev_sent_queue(ndev, skb->len);
-
 	spin_lock_bh(&lp->tx_lock);
 	lp->tx_free -= trans.nr_descriptors;
+	dwceqos_tx_finalize(skb, lp, &trans);
+	netdev_sent_queue(ndev, skb->len);
 	spin_unlock_bh(&lp->tx_lock);
 
 	ndev->trans_start = jiffies;
-- 
2.1.4

[PATCH net 2/5] dwc_eth_qos: release descriptors outside netif_tx_lock

[Lars Persson]

To prepare for using the CMA, we can not be in atomic context when
de-allocating DMA buffers.

The tx lock was needed only to protect the hw reset against the xmit
handler. Now we briefly grab the tx lock while stopping the queue to
make sure no thread is inside or will enter the xmit handler.

Signed-off-by: Lars Persson <larper@axis.com>
---
 drivers/net/ethernet/synopsys/dwc_eth_qos.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/drivers/net/ethernet/synopsys/dwc_eth_qos.c b/drivers/net/ethernet/synopsys/dwc_eth_qos.c
index 926db2d..53d48c0 100644
--- a/drivers/net/ethernet/synopsys/dwc_eth_qos.c
+++ b/drivers/net/ethernet/synopsys/dwc_eth_qos.c
@@ -1918,15 +1918,17 @@ static int dwceqos_stop(struct net_device *ndev)
 	phy_stop(lp->phy_dev);
 
 	tasklet_disable(&lp->tx_bdreclaim_tasklet);
-	netif_stop_queue(ndev);
 	napi_disable(&lp->napi);
 
-	dwceqos_drain_dma(lp);
+	/* Stop all tx before we drain the tx dma. */
+	netif_tx_lock_bh(lp->ndev);
+	netif_stop_queue(ndev);
+	netif_tx_unlock_bh(lp->ndev);
 
-	netif_tx_lock(lp->ndev);
+	dwceqos_drain_dma(lp);
 	dwceqos_reset_hw(lp);
+
 	dwceqos_descriptor_free(lp);
-	netif_tx_unlock(lp->ndev);
 
 	return 0;
 }
-- 
2.1.4

[PATCH net 3/5] dwc_eth_qos: use GFP_KERNEL in dma_alloc_coherent()

[Lars Persson]

From: Rabin Vincent <rabinv@axis.com>

Since we are in non-atomic context here we can pass GFP_KERNEL to
dma_alloc_coherent(). This enables use of the CMA.

Signed-off-by: Rabin Vincent <rabinv@axis.com>
Signed-off-by: Lars Persson <larper@axis.com>
---
 drivers/net/ethernet/synopsys/dwc_eth_qos.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/net/ethernet/synopsys/dwc_eth_qos.c b/drivers/net/ethernet/synopsys/dwc_eth_qos.c
index 53d48c0..3ca2d5c 100644
--- a/drivers/net/ethernet/synopsys/dwc_eth_qos.c
+++ b/drivers/net/ethernet/synopsys/dwc_eth_qos.c
@@ -1113,7 +1113,7 @@ static int dwceqos_descriptor_init(struct net_local *lp)
 	/* Allocate DMA descriptors */
 	size = DWCEQOS_RX_DCNT * sizeof(struct dwceqos_dma_desc);
 	lp->rx_descs = dma_alloc_coherent(lp->ndev->dev.parent, size,
-			&lp->rx_descs_addr, 0);
+			&lp->rx_descs_addr, GFP_KERNEL);
 	if (!lp->rx_descs)
 		goto err_out;
 	lp->rx_descs_tail_addr = lp->rx_descs_addr +
@@ -1121,7 +1121,7 @@ static int dwceqos_descriptor_init(struct net_local *lp)
 
 	size = DWCEQOS_TX_DCNT * sizeof(struct dwceqos_dma_desc);
 	lp->tx_descs = dma_alloc_coherent(lp->ndev->dev.parent, size,
-			&lp->tx_descs_addr, 0);
+			&lp->tx_descs_addr, GFP_KERNEL);
 	if (!lp->tx_descs)
 		goto err_out;
 	lp->tx_descs_tail_addr = lp->tx_descs_addr +
-- 
2.1.4

[PATCH net 4/5] dwc_eth_qos: use DWCEQOS_MSG_DEFAULT

[Lars Persson]

From: Rabin Vincent <rabinv@axis.com>

Since debug is hardcoded to 3, the defaults in the DWCEQOS_MSG_DEFAULT
macro are never used, which does not seem to be the intended behaviour
here.  Set debug to -1 like other drivers so that DWCEQOS_MSG_DEFAULT is
actually used by default.

Signed-off-by: Rabin Vincent <rabinv@axis.com>
Signed-off-by: Lars Persson <larper@axis.com>
---
 drivers/net/ethernet/synopsys/dwc_eth_qos.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/synopsys/dwc_eth_qos.c b/drivers/net/ethernet/synopsys/dwc_eth_qos.c
index 3ca2d5c..6897c1d 100644
--- a/drivers/net/ethernet/synopsys/dwc_eth_qos.c
+++ b/drivers/net/ethernet/synopsys/dwc_eth_qos.c
@@ -426,7 +426,7 @@
 #define DWC_MMC_RXOCTETCOUNT_GB          0x0784
 #define DWC_MMC_RXPACKETCOUNT_GB         0x0780
 
-static int debug = 3;
+static int debug = -1;
 module_param(debug, int, 0);
 MODULE_PARM_DESC(debug, "DWC_eth_qos debug level (0=none,...,16=all)");
 
-- 
2.1.4

[PATCH net 5/5] dwc_eth_qos: do phy_start before resetting hardware

[Lars Persson]

This reverts the changed init order from commit 3647bc35bd42
("dwc_eth_qos: Reset hardware before PHY start") and makes another fix
for the race.

It turned out that the reset state machine of the dwceqos hardware
requires PHY clocks to be present in order to complete the reset
cycle.

To plug the race with the phy state machine we defer link speed
setting until the hardware init has finished.

Signed-off-by: Lars Persson <larper@axis.com>
---
 drivers/net/ethernet/synopsys/dwc_eth_qos.c | 23 ++++++++++++++++++++---
 1 file changed, 20 insertions(+), 3 deletions(-)

diff --git a/drivers/net/ethernet/synopsys/dwc_eth_qos.c b/drivers/net/ethernet/synopsys/dwc_eth_qos.c
index 6897c1d..af11ed1 100644
--- a/drivers/net/ethernet/synopsys/dwc_eth_qos.c
+++ b/drivers/net/ethernet/synopsys/dwc_eth_qos.c
@@ -650,6 +650,11 @@ struct net_local {
 	u32 mmc_tx_counters_mask;
 
 	struct dwceqos_flowcontrol flowcontrol;
+
+	/* Tracks the intermediate state of phy started but hardware
+	 * init not finished yet.
+	 */
+	bool phy_defer;
 };
 
 static void dwceqos_read_mmc_counters(struct net_local *lp, u32 rx_mask,
@@ -901,6 +906,9 @@ static void dwceqos_adjust_link(struct net_device *ndev)
 	struct phy_device *phydev = lp->phy_dev;
 	int status_change = 0;
 
+	if (lp->phy_defer)
+		return;
+
 	if (phydev->link) {
 		if ((lp->speed != phydev->speed) ||
 		    (lp->duplex != phydev->duplex)) {
@@ -1635,6 +1643,12 @@ static void dwceqos_init_hw(struct net_local *lp)
 	regval = dwceqos_read(lp, REG_DWCEQOS_MAC_CFG);
 	dwceqos_write(lp, REG_DWCEQOS_MAC_CFG,
 		      regval | DWCEQOS_MAC_CFG_TE | DWCEQOS_MAC_CFG_RE);
+
+	lp->phy_defer = false;
+	mutex_lock(&lp->phy_dev->lock);
+	phy_read_status(lp->phy_dev);
+	dwceqos_adjust_link(lp->ndev);
+	mutex_unlock(&lp->phy_dev->lock);
 }
 
 static void dwceqos_tx_reclaim(unsigned long data)
@@ -1880,9 +1894,13 @@ static int dwceqos_open(struct net_device *ndev)
 	}
 	netdev_reset_queue(ndev);
 
+	/* The dwceqos reset state machine requires all phy clocks to complete,
+	 * hence the unusual init order with phy_start first.
+	 */
+	lp->phy_defer = true;
+	phy_start(lp->phy_dev);
 	dwceqos_init_hw(lp);
 	napi_enable(&lp->napi);
-	phy_start(lp->phy_dev);
 
 	netif_start_queue(ndev);
 	tasklet_enable(&lp->tx_bdreclaim_tasklet);
@@ -1915,8 +1933,6 @@ static int dwceqos_stop(struct net_device *ndev)
 {
 	struct net_local *lp = netdev_priv(ndev);
 
-	phy_stop(lp->phy_dev);
-
 	tasklet_disable(&lp->tx_bdreclaim_tasklet);
 	napi_disable(&lp->napi);
 
@@ -1927,6 +1943,7 @@ static int dwceqos_stop(struct net_device *ndev)
 
 	dwceqos_drain_dma(lp);
 	dwceqos_reset_hw(lp);
+	phy_stop(lp->phy_dev);
 
 	dwceqos_descriptor_free(lp);
 
-- 
2.1.4

Re: [PATCH net 0/5] dwc_eth_qos: stability fixes and support for CMA

[David Miller]

From: Lars Persson <lars.persson@axis.com>
Date: Mon, 29 Feb 2016 16:22:29 +0100

> This series has bug fixes for the dwc_eth_qos ethernet driver.
> 
> Mainly two stability fixes for problems found by Rabin Vincent:
> - Successive starts and stops of the interface would trigger a DMA reset timeout.
> - A race condition in the TX DMA handling could trigger a netdev watchdog
>  timeout.
> 
> The memory allocation was improved to support use of the CMA as DMA allocator
> backend.

Series applied, thanks.