[PATCH 0/4] Improvements to octeon_ethernet.

[PATCH 0/4] Improvements to octeon_ethernet.

[David Daney]

Here are a couple of improvements to the octeon_ethernet in
drivers/staging/octeon.  The first patch is just cleanup, the rest are
genuine bug fixes.

I will reply with the four patches.

We may want to merge via Ralf's linux-mips.org tree as Octeon is
infact a MIPS based SOC and he has a bunch of other patches queued
there that these depend on.

David Daney (4):
   Staging: octeon: remove unneeded includes
   Staging: Octeon:  Run phy bus accesses on a workqueue.
   MIPS: Octeon: Do proper acknowledgment of CIU timer interrupts.
   Staging: Octeon:  Free transmit SKBs in a timely manner.

  arch/mips/cavium-octeon/octeon-irq.c      |   67 +++++++++++++-
  drivers/staging/octeon/Kconfig            |    1 -
  drivers/staging/octeon/ethernet-defines.h |    5 +-
  drivers/staging/octeon/ethernet-mdio.h    |    1 -
  drivers/staging/octeon/ethernet-rgmii.c   |   56 +++++++++---
  drivers/staging/octeon/ethernet-sgmii.c   |    1 -
  drivers/staging/octeon/ethernet-spi.c     |    1 -
  drivers/staging/octeon/ethernet-tx.c      |  137 
++++++++++++++++++++++------
  drivers/staging/octeon/ethernet-tx.h      |    6 +-
  drivers/staging/octeon/ethernet-xaui.c    |    1 -
  drivers/staging/octeon/ethernet.c         |  141 
++++++++++++++++-------------
  drivers/staging/octeon/octeon-ethernet.h  |   11 ++-
  12 files changed, 303 insertions(+), 125 deletions(-)

[PATCH 1/4] Staging: octeon: remove unneeded includes

[David Daney]

Signed-off-by: David Daney <ddaney@caviumnetworks.com>
---
 drivers/staging/octeon/ethernet-mdio.h  |    1 -
 drivers/staging/octeon/ethernet-rgmii.c |    1 -
 drivers/staging/octeon/ethernet-sgmii.c |    1 -
 drivers/staging/octeon/ethernet-spi.c   |    1 -
 drivers/staging/octeon/ethernet-xaui.c  |    1 -
 drivers/staging/octeon/ethernet.c       |    1 -
 6 files changed, 0 insertions(+), 6 deletions(-)

diff --git a/drivers/staging/octeon/ethernet-mdio.h b/drivers/staging/octeon/ethernet-mdio.h
index 55d0614..a417d4f 100644
--- a/drivers/staging/octeon/ethernet-mdio.h
+++ b/drivers/staging/octeon/ethernet-mdio.h
@@ -32,7 +32,6 @@
 #include <linux/ip.h>
 #include <linux/string.h>
 #include <linux/ethtool.h>
-#include <linux/mii.h>
 #include <linux/seq_file.h>
 #include <linux/proc_fs.h>
 #include <net/dst.h>
diff --git a/drivers/staging/octeon/ethernet-rgmii.c b/drivers/staging/octeon/ethernet-rgmii.c
index 3820f1e..f90d46e 100644
--- a/drivers/staging/octeon/ethernet-rgmii.c
+++ b/drivers/staging/octeon/ethernet-rgmii.c
@@ -26,7 +26,6 @@
 **********************************************************************/
 #include <linux/kernel.h>
 #include <linux/netdevice.h>
-#include <linux/mii.h>
 #include <net/dst.h>
 
 #include <asm/octeon/octeon.h>
diff --git a/drivers/staging/octeon/ethernet-sgmii.c b/drivers/staging/octeon/ethernet-sgmii.c
index 6061d01..2d8589e 100644
--- a/drivers/staging/octeon/ethernet-sgmii.c
+++ b/drivers/staging/octeon/ethernet-sgmii.c
@@ -26,7 +26,6 @@
 **********************************************************************/
 #include <linux/kernel.h>
 #include <linux/netdevice.h>
-#include <linux/mii.h>
 #include <net/dst.h>
 
 #include <asm/octeon/octeon.h>
diff --git a/drivers/staging/octeon/ethernet-spi.c b/drivers/staging/octeon/ethernet-spi.c
index 00dc0f4..b58b897 100644
--- a/drivers/staging/octeon/ethernet-spi.c
+++ b/drivers/staging/octeon/ethernet-spi.c
@@ -26,7 +26,6 @@
 **********************************************************************/
 #include <linux/kernel.h>
 #include <linux/netdevice.h>
-#include <linux/mii.h>
 #include <net/dst.h>
 
 #include <asm/octeon/octeon.h>
diff --git a/drivers/staging/octeon/ethernet-xaui.c b/drivers/staging/octeon/ethernet-xaui.c
index ee3dc41..3fca1cc 100644
--- a/drivers/staging/octeon/ethernet-xaui.c
+++ b/drivers/staging/octeon/ethernet-xaui.c
@@ -26,7 +26,6 @@
 **********************************************************************/
 #include <linux/kernel.h>
 #include <linux/netdevice.h>
-#include <linux/mii.h>
 #include <net/dst.h>
 
 #include <asm/octeon/octeon.h>
diff --git a/drivers/staging/octeon/ethernet.c b/drivers/staging/octeon/ethernet.c
index 9d63202..5afece0 100644
--- a/drivers/staging/octeon/ethernet.c
+++ b/drivers/staging/octeon/ethernet.c
@@ -29,7 +29,6 @@
 #include <linux/module.h>
 #include <linux/netdevice.h>
 #include <linux/etherdevice.h>
-#include <linux/delay.h>
 #include <linux/phy.h>
 
 #include <net/dst.h>
-- 
1.6.6

[PATCH 2/4] Staging: Octeon: Run phy bus accesses on a workqueue.

[David Daney]

When directly accessing a phy, we must acquire the mdio bus lock.  To
do that we cannot be in interrupt context, so we need to move these
operations to a workqueue.

Signed-off-by: David Daney <ddaney@caviumnetworks.com>
---
 drivers/staging/octeon/ethernet-rgmii.c  |   55 +++++++++++----
 drivers/staging/octeon/ethernet.c        |  113 +++++++++++++++++-------------
 drivers/staging/octeon/octeon-ethernet.h |    4 +
 3 files changed, 109 insertions(+), 63 deletions(-)

diff --git a/drivers/staging/octeon/ethernet-rgmii.c b/drivers/staging/octeon/ethernet-rgmii.c
index f90d46e..a0d4d4b 100644
--- a/drivers/staging/octeon/ethernet-rgmii.c
+++ b/drivers/staging/octeon/ethernet-rgmii.c
@@ -26,6 +26,7 @@
 **********************************************************************/
 #include <linux/kernel.h>
 #include <linux/netdevice.h>
+#include <linux/phy.h>
 #include <net/dst.h>
 
 #include <asm/octeon/octeon.h>
@@ -47,14 +48,20 @@ static int number_rgmii_ports;
 static void cvm_oct_rgmii_poll(struct net_device *dev)
 {
 	struct octeon_ethernet *priv = netdev_priv(dev);
-	unsigned long flags;
+	unsigned long flags = 0;
 	cvmx_helper_link_info_t link_info;
+	int use_global_register_lock = (priv->phydev == NULL);
 
-	/*
-	 * Take the global register lock since we are going to touch
-	 * registers that affect more than one port.
-	 */
-	spin_lock_irqsave(&global_register_lock, flags);
+	BUG_ON(in_interrupt());
+	if (use_global_register_lock) {
+		/*
+		 * Take the global register lock since we are going to
+		 * touch registers that affect more than one port.
+		 */
+		spin_lock_irqsave(&global_register_lock, flags);
+	} else {
+		mutex_lock(&priv->phydev->bus->mdio_lock);
+	}
 
 	link_info = cvmx_helper_link_get(priv->port);
 	if (link_info.u64 == priv->link_info) {
@@ -114,7 +121,11 @@ static void cvm_oct_rgmii_poll(struct net_device *dev)
 				     dev->name);
 			}
 		}
-		spin_unlock_irqrestore(&global_register_lock, flags);
+
+		if (use_global_register_lock)
+			spin_unlock_irqrestore(&global_register_lock, flags);
+		else
+			mutex_unlock(&priv->phydev->bus->mdio_lock);
 		return;
 	}
 
@@ -150,7 +161,12 @@ static void cvm_oct_rgmii_poll(struct net_device *dev)
 		link_info = cvmx_helper_link_autoconf(priv->port);
 		priv->link_info = link_info.u64;
 	}
-	spin_unlock_irqrestore(&global_register_lock, flags);
+
+	if (use_global_register_lock)
+		spin_unlock_irqrestore(&global_register_lock, flags);
+	else {
+		mutex_unlock(&priv->phydev->bus->mdio_lock);
+	}
 
 	if (priv->phydev == NULL) {
 		/* Tell core. */
@@ -212,8 +228,11 @@ static irqreturn_t cvm_oct_rgmii_rml_interrupt(int cpl, void *dev_id)
 				struct net_device *dev =
 				    cvm_oct_device[cvmx_helper_get_ipd_port
 						   (interface, index)];
-				if (dev)
-					cvm_oct_rgmii_poll(dev);
+				struct octeon_ethernet *priv = netdev_priv(dev);
+
+				if (dev && !atomic_read(&cvm_oct_poll_queue_stopping))
+					queue_work(cvm_oct_poll_queue, &priv->port_work);
+
 				gmx_rx_int_reg.u64 = 0;
 				gmx_rx_int_reg.s.phy_dupx = 1;
 				gmx_rx_int_reg.s.phy_link = 1;
@@ -251,8 +270,11 @@ static irqreturn_t cvm_oct_rgmii_rml_interrupt(int cpl, void *dev_id)
 				struct net_device *dev =
 				    cvm_oct_device[cvmx_helper_get_ipd_port
 						   (interface, index)];
-				if (dev)
-					cvm_oct_rgmii_poll(dev);
+				struct octeon_ethernet *priv = netdev_priv(dev);
+
+				if (dev && !atomic_read(&cvm_oct_poll_queue_stopping))
+					queue_work(cvm_oct_poll_queue, &priv->port_work);
+
 				gmx_rx_int_reg.u64 = 0;
 				gmx_rx_int_reg.s.phy_dupx = 1;
 				gmx_rx_int_reg.s.phy_link = 1;
@@ -301,6 +323,12 @@ int cvm_oct_rgmii_stop(struct net_device *dev)
 	return 0;
 }
 
+static void cvm_oct_rgmii_immediate_poll(struct work_struct *work)
+{
+	struct octeon_ethernet *priv = container_of(work, struct octeon_ethernet, port_work);
+	cvm_oct_rgmii_poll(cvm_oct_device[priv->port]);
+}
+
 int cvm_oct_rgmii_init(struct net_device *dev)
 {
 	struct octeon_ethernet *priv = netdev_priv(dev);
@@ -308,7 +336,7 @@ int cvm_oct_rgmii_init(struct net_device *dev)
 
 	cvm_oct_common_init(dev);
 	dev->netdev_ops->ndo_stop(dev);
-
+	INIT_WORK(&priv->port_work, cvm_oct_rgmii_immediate_poll);
 	/*
 	 * Due to GMX errata in CN3XXX series chips, it is necessary
 	 * to take the link down immediately when the PHY changes
@@ -396,4 +424,5 @@ void cvm_oct_rgmii_uninit(struct net_device *dev)
 	number_rgmii_ports--;
 	if (number_rgmii_ports == 0)
 		free_irq(OCTEON_IRQ_RML, &number_rgmii_ports);
+	cancel_work_sync(&priv->port_work);
 }
diff --git a/drivers/staging/octeon/ethernet.c b/drivers/staging/octeon/ethernet.c
index 5afece0..1771c10 100644
--- a/drivers/staging/octeon/ethernet.c
+++ b/drivers/staging/octeon/ethernet.c
@@ -123,9 +123,16 @@ MODULE_PARM_DESC(rx_napi_weight, "The NAPI WEIGHT parameter.");
 static unsigned int cvm_oct_mac_addr_offset;
 
 /**
- * Periodic timer to check auto negotiation
+ * cvm_oct_poll_queue - Workqueue for polling operations.
  */
-static struct timer_list cvm_oct_poll_timer;
+struct workqueue_struct *cvm_oct_poll_queue;
+
+/**
+ * cvm_oct_poll_queue_stopping - flag to indicate polling should stop.
+ *
+ * Set to one right before cvm_oct_poll_queue is destroyed.
+ */
+atomic_t cvm_oct_poll_queue_stopping = ATOMIC_INIT(0);
 
 /**
  * Array of every ethernet device owned by this driver indexed by
@@ -133,47 +140,39 @@ static struct timer_list cvm_oct_poll_timer;
  */
 struct net_device *cvm_oct_device[TOTAL_NUMBER_OF_PORTS];
 
-/**
- * Periodic timer tick for slow management operations
- *
- * @arg:    Device to check
- */
-static void cvm_do_timer(unsigned long arg)
+static void cvm_oct_rx_refill_worker(struct work_struct *work);
+static DECLARE_DELAYED_WORK(cvm_oct_rx_refill_work, cvm_oct_rx_refill_worker);
+
+static void cvm_oct_rx_refill_worker(struct work_struct *work)
 {
-	static int port;
-	if (port < CVMX_PIP_NUM_INPUT_PORTS) {
-		if (cvm_oct_device[port]) {
-			struct octeon_ethernet *priv = netdev_priv(cvm_oct_device[port]);
-			if (priv->poll)
-				priv->poll(cvm_oct_device[port]);
-			cvm_oct_free_tx_skbs(priv);
-			cvm_oct_device[port]->netdev_ops->ndo_get_stats(cvm_oct_device[port]);
-		}
-		port++;
-		/*
-		 * Poll the next port in a 50th of a second.  This
-		 * spreads the polling of ports out a little bit.
-		 */
-		mod_timer(&cvm_oct_poll_timer, jiffies + HZ/50);
-	} else {
-		port = 0;
-		/*
-		 * FPA 0 may have been drained, try to refill it if we
-		 * need more than num_packet_buffers / 2, otherwise
-		 * normal receive processing will refill it.  If it
-		 * were drained, no packets could be received so
-		 * cvm_oct_napi_poll would never be invoked to do the
-		 * refill.
-		 */
-		cvm_oct_rx_refill_pool(num_packet_buffers / 2);
-		/*
-		 * All ports have been polled. Start the next iteration through
-		 * the ports in one second.
-		 */
-		mod_timer(&cvm_oct_poll_timer, jiffies + HZ);
-	}
+	/*
+	 * FPA 0 may have been drained, try to refill it if we need
+	 * more than num_packet_buffers / 2, otherwise normal receive
+	 * processing will refill it.  If it were drained, no packets
+	 * could be received so cvm_oct_napi_poll would never be
+	 * invoked to do the refill.
+	 */
+	cvm_oct_rx_refill_pool(num_packet_buffers / 2);
+
+	if (!atomic_read(&cvm_oct_poll_queue_stopping))
+		queue_delayed_work(cvm_oct_poll_queue,
+				   &cvm_oct_rx_refill_work, HZ);
 }
 
+static void cvm_oct_tx_clean_worker(struct work_struct *work)
+{
+	struct octeon_ethernet *priv = container_of(work,
+						    struct octeon_ethernet,
+						    tx_clean_work.work);
+
+	if (priv->poll)
+		priv->poll(cvm_oct_device[priv->port]);
+	cvm_oct_free_tx_skbs(priv);
+	cvm_oct_device[priv->port]->netdev_ops->ndo_get_stats(cvm_oct_device[priv->port]);
+	if (!atomic_read(&cvm_oct_poll_queue_stopping))
+		queue_delayed_work(cvm_oct_poll_queue, &priv->tx_clean_work, HZ);
+ }
+
 /**
  * Configure common hardware for all interfaces
  */
@@ -624,6 +623,12 @@ static int __init cvm_oct_init_module(void)
 	else
 		cvm_oct_mac_addr_offset = 0;
 
+	cvm_oct_poll_queue = create_singlethread_workqueue("octeon-ethernet");
+	if (cvm_oct_poll_queue == NULL) {
+		pr_err("octeon-ethernet: Cannot create workqueue");
+		return -ENOMEM;
+	}
+
 	cvm_oct_proc_initialize();
 	cvm_oct_configure_common_hw();
 
@@ -719,7 +724,9 @@ static int __init cvm_oct_init_module(void)
 
 			/* Initialize the device private structure. */
 			priv = netdev_priv(dev);
-			memset(priv, 0, sizeof(struct octeon_ethernet));
+
+			INIT_DELAYED_WORK(&priv->tx_clean_work,
+					  cvm_oct_tx_clean_worker);
 
 			priv->imode = imode;
 			priv->port = port;
@@ -785,17 +792,15 @@ static int __init cvm_oct_init_module(void)
 				fau -=
 				    cvmx_pko_get_num_queues(priv->port) *
 				    sizeof(uint32_t);
+				queue_delayed_work(cvm_oct_poll_queue,
+						   &priv->tx_clean_work, HZ);
 			}
 		}
 	}
 
 	cvm_oct_rx_initialize();
 
-	/* Enable the poll timer for checking RGMII status */
-	init_timer(&cvm_oct_poll_timer);
-	cvm_oct_poll_timer.data = 0;
-	cvm_oct_poll_timer.function = cvm_do_timer;
-	mod_timer(&cvm_oct_poll_timer, jiffies + HZ);
+	queue_delayed_work(cvm_oct_poll_queue, &cvm_oct_rx_refill_work, HZ);
 
 	return 0;
 }
@@ -817,20 +822,28 @@ static void __exit cvm_oct_cleanup_module(void)
 	/* Free the interrupt handler */
 	free_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group, cvm_oct_device);
 
-	del_timer(&cvm_oct_poll_timer);
+	atomic_inc_return(&cvm_oct_poll_queue_stopping);
+	cancel_delayed_work_sync(&cvm_oct_rx_refill_work);
+
 	cvm_oct_rx_shutdown();
 	cvmx_pko_disable();
 
 	/* Free the ethernet devices */
 	for (port = 0; port < TOTAL_NUMBER_OF_PORTS; port++) {
 		if (cvm_oct_device[port]) {
-			cvm_oct_tx_shutdown(cvm_oct_device[port]);
-			unregister_netdev(cvm_oct_device[port]);
-			kfree(cvm_oct_device[port]);
+			struct net_device *dev = cvm_oct_device[port];
+			struct octeon_ethernet *priv = netdev_priv(dev);
+			cancel_delayed_work_sync(&priv->tx_clean_work);
+
+			cvm_oct_tx_shutdown(dev);
+			unregister_netdev(dev);
+			kfree(dev);
 			cvm_oct_device[port] = NULL;
 		}
 	}
 
+	destroy_workqueue(cvm_oct_poll_queue);
+
 	cvmx_pko_shutdown();
 	cvm_oct_proc_shutdown();
 
diff --git a/drivers/staging/octeon/octeon-ethernet.h b/drivers/staging/octeon/octeon-ethernet.h
index 40b6956..8d09210 100644
--- a/drivers/staging/octeon/octeon-ethernet.h
+++ b/drivers/staging/octeon/octeon-ethernet.h
@@ -61,6 +61,8 @@ struct octeon_ethernet {
 	void (*poll) (struct net_device *dev);
 	struct hrtimer		tx_restart_timer;
 	ktime_t			tx_restart_interval;
+	struct delayed_work	tx_clean_work;
+	struct work_struct	port_work;	/* may be unused. */
 };
 
 /**
@@ -97,6 +99,8 @@ extern int pow_send_group;
 extern int pow_receive_group;
 extern char pow_send_list[];
 extern struct net_device *cvm_oct_device[];
+extern struct workqueue_struct *cvm_oct_poll_queue;
+extern atomic_t cvm_oct_poll_queue_stopping;
 
 extern int max_rx_cpus;
 extern int rx_napi_weight;
-- 
1.6.6

[PATCH 3/4] MIPS: Octeon: Do proper acknowledgment of CIU timer interrupts.

[David Daney]

Signed-off-by: David Daney <ddaney@caviumnetworks.com>
---
 arch/mips/cavium-octeon/octeon-irq.c |   67 ++++++++++++++++++++++++++++++++--
 1 files changed, 63 insertions(+), 4 deletions(-)

diff --git a/arch/mips/cavium-octeon/octeon-irq.c b/arch/mips/cavium-octeon/octeon-irq.c
index 1460d08..0bc79dc 100644
--- a/arch/mips/cavium-octeon/octeon-irq.c
+++ b/arch/mips/cavium-octeon/octeon-irq.c
@@ -51,9 +51,6 @@ static void octeon_irq_core_eoi(unsigned int irq)
 	 */
 	if (desc->status & IRQ_DISABLED)
 		return;
-
-	/* There is a race here.  We should fix it.  */
-
 	/*
 	 * We don't need to disable IRQs to make these atomic since
 	 * they are already disabled earlier in the low level
@@ -202,6 +199,29 @@ static void octeon_irq_ciu0_ack_v2(unsigned int irq)
 }
 
 /*
+ * CIU timer type interrupts must be acknoleged by writing a '1' bit
+ * to their sum0 bit.
+ */
+static void octeon_irq_ciu0_timer_ack(unsigned int irq)
+{
+	int index = cvmx_get_core_num() * 2;
+	uint64_t mask = 1ull << (irq - OCTEON_IRQ_WORKQ0);
+	cvmx_write_csr(CVMX_CIU_INTX_SUM0(index), mask);
+}
+
+static void octeon_irq_ciu0_timer_ack_v1(unsigned int irq)
+{
+	octeon_irq_ciu0_timer_ack(irq);
+	octeon_irq_ciu0_ack(irq);
+}
+
+static void octeon_irq_ciu0_timer_ack_v2(unsigned int irq)
+{
+	octeon_irq_ciu0_timer_ack(irq);
+	octeon_irq_ciu0_ack_v2(irq);
+}
+
+/*
  * Enable the irq on the current core for chips that have the EN*_W1{S,C}
  * registers.
  */
@@ -304,6 +324,28 @@ static struct irq_chip octeon_irq_chip_ciu0 = {
 #endif
 };
 
+static struct irq_chip octeon_irq_chip_ciu0_timer_v2 = {
+	.name = "CIU0-T",
+	.enable = octeon_irq_ciu0_enable_v2,
+	.disable = octeon_irq_ciu0_disable_all_v2,
+	.ack = octeon_irq_ciu0_timer_ack_v2,
+	.eoi = octeon_irq_ciu0_eoi_v2,
+#ifdef CONFIG_SMP
+	.set_affinity = octeon_irq_ciu0_set_affinity_v2,
+#endif
+};
+
+static struct irq_chip octeon_irq_chip_ciu0_timer = {
+	.name = "CIU0-T",
+	.enable = octeon_irq_ciu0_enable,
+	.disable = octeon_irq_ciu0_disable,
+	.ack = octeon_irq_ciu0_timer_ack_v1,
+	.eoi = octeon_irq_ciu0_eoi,
+#ifdef CONFIG_SMP
+	.set_affinity = octeon_irq_ciu0_set_affinity,
+#endif
+};
+
 
 static void octeon_irq_ciu1_ack(unsigned int irq)
 {
@@ -587,6 +629,7 @@ void __init arch_init_irq(void)
 {
 	int irq;
 	struct irq_chip *chip0;
+	struct irq_chip *chip0_timer;
 	struct irq_chip *chip1;
 
 #ifdef CONFIG_SMP
@@ -602,9 +645,11 @@ void __init arch_init_irq(void)
 	    OCTEON_IS_MODEL(OCTEON_CN56XX_PASS2_X) ||
 	    OCTEON_IS_MODEL(OCTEON_CN52XX_PASS2_X)) {
 		chip0 = &octeon_irq_chip_ciu0_v2;
+		chip0_timer = &octeon_irq_chip_ciu0_timer_v2;
 		chip1 = &octeon_irq_chip_ciu1_v2;
 	} else {
 		chip0 = &octeon_irq_chip_ciu0;
+		chip0_timer = &octeon_irq_chip_ciu0_timer;
 		chip1 = &octeon_irq_chip_ciu1;
 	}
 
@@ -618,7 +663,21 @@ void __init arch_init_irq(void)
 
 	/* 24 - 87 CIU_INT_SUM0 */
 	for (irq = OCTEON_IRQ_WORKQ0; irq <= OCTEON_IRQ_BOOTDMA; irq++) {
-		set_irq_chip_and_handler(irq, chip0, handle_percpu_irq);
+		switch (irq) {
+		case OCTEON_IRQ_GMX_DRP0:
+		case OCTEON_IRQ_GMX_DRP1:
+		case OCTEON_IRQ_IPD_DRP:
+		case OCTEON_IRQ_KEY_ZERO:
+		case OCTEON_IRQ_TIMER0:
+		case OCTEON_IRQ_TIMER1:
+		case OCTEON_IRQ_TIMER2:
+		case OCTEON_IRQ_TIMER3:
+			set_irq_chip_and_handler(irq, chip0_timer, handle_percpu_irq);
+			break;
+		default:
+			set_irq_chip_and_handler(irq, chip0, handle_percpu_irq);
+			break;
+		}
 	}
 
 	/* 88 - 151 CIU_INT_SUM1 */
-- 
1.6.6

[PATCH 4/4] Staging: Octeon: Free transmit SKBs in a timely manner.

[David Daney]

If we wait for the once-per-second cleanup to free transmit SKBs,
sockets with small transmit buffer sizes might spend most of their
time blocked waiting for the cleanup.

Normally we do a cleanup for each transmitted packet.  We add a
watchdog type timer so that we also schedule a timeout for 150uS after
a packet is transmitted.  The watchdog is reset for each transmitted
packet, so for high packet rates, it never expires.  At these high
rates, the cleanups are done for each packet so the extra watchdog
initiated cleanups are not needed.

Signed-off-by: David Daney <ddaney@caviumnetworks.com>
---
 drivers/staging/octeon/Kconfig            |    1 -
 drivers/staging/octeon/ethernet-defines.h |    5 +-
 drivers/staging/octeon/ethernet-tx.c      |  137 +++++++++++++++++++++++------
 drivers/staging/octeon/ethernet-tx.h      |    6 +-
 drivers/staging/octeon/ethernet.c         |   47 +++++-----
 drivers/staging/octeon/octeon-ethernet.h  |    9 +--
 6 files changed, 142 insertions(+), 63 deletions(-)

diff --git a/drivers/staging/octeon/Kconfig b/drivers/staging/octeon/Kconfig
index 579b8f1..638ad6b 100644
--- a/drivers/staging/octeon/Kconfig
+++ b/drivers/staging/octeon/Kconfig
@@ -3,7 +3,6 @@ config OCTEON_ETHERNET
 	depends on CPU_CAVIUM_OCTEON
 	select PHYLIB
 	select MDIO_OCTEON
-	select HIGH_RES_TIMERS
 	help
 	  This driver supports the builtin ethernet ports on Cavium
 	  Networks' products in the Octeon family. This driver supports the
diff --git a/drivers/staging/octeon/ethernet-defines.h b/drivers/staging/octeon/ethernet-defines.h
index 00a8561..6a2cd50 100644
--- a/drivers/staging/octeon/ethernet-defines.h
+++ b/drivers/staging/octeon/ethernet-defines.h
@@ -95,10 +95,11 @@
 /*#define DONT_WRITEBACK(x)         0   */
 
 /* Maximum number of SKBs to try to free per xmit packet. */
-#define MAX_SKB_TO_FREE 10
 #define MAX_OUT_QUEUE_DEPTH 1000
 
-#define FAU_NUM_PACKET_BUFFERS_TO_FREE (CVMX_FAU_REG_END - sizeof(uint32_t))
+#define FAU_TOTAL_TX_TO_CLEAN (CVMX_FAU_REG_END - sizeof(uint32_t))
+#define FAU_NUM_PACKET_BUFFERS_TO_FREE (FAU_TOTAL_TX_TO_CLEAN - sizeof(uint32_t))
+
 #define TOTAL_NUMBER_OF_PORTS       (CVMX_PIP_NUM_INPUT_PORTS+1)
 
 
diff --git a/drivers/staging/octeon/ethernet-tx.c b/drivers/staging/octeon/ethernet-tx.c
index 62258bd..5175247 100644
--- a/drivers/staging/octeon/ethernet-tx.c
+++ b/drivers/staging/octeon/ethernet-tx.c
@@ -48,6 +48,7 @@
 
 #include "cvmx-wqe.h"
 #include "cvmx-fau.h"
+#include "cvmx-pip.h"
 #include "cvmx-pko.h"
 #include "cvmx-helper.h"
 
@@ -66,6 +67,11 @@
 #define GET_SKBUFF_QOS(skb) 0
 #endif
 
+static void cvm_oct_tx_do_cleanup(unsigned long arg);
+static DECLARE_TASKLET(cvm_oct_tx_cleanup_tasklet, cvm_oct_tx_do_cleanup, 0);
+
+/* Maximum number of SKBs to try to free per xmit packet. */
+#define MAX_SKB_TO_FREE (MAX_OUT_QUEUE_DEPTH * 2)
 
 static inline int32_t cvm_oct_adjust_skb_to_free(int32_t skb_to_free, int fau)
 {
@@ -77,10 +83,24 @@ static inline int32_t cvm_oct_adjust_skb_to_free(int32_t skb_to_free, int fau)
 	return skb_to_free;
 }
 
-void cvm_oct_free_tx_skbs(struct octeon_ethernet *priv)
+static void cvm_oct_kick_tx_poll_watchdog(void)
+{
+	union cvmx_ciu_timx ciu_timx;
+	ciu_timx.u64 = 0;
+	ciu_timx.s.one_shot = 1;
+	ciu_timx.s.len = cvm_oct_tx_poll_interval;
+	cvmx_write_csr(CVMX_CIU_TIMX(1), ciu_timx.u64);
+}
+
+void cvm_oct_free_tx_skbs(struct net_device *dev)
 {
 	int32_t skb_to_free;
 	int qos, queues_per_port;
+	int total_freed = 0;
+	int total_remaining = 0;
+	unsigned long flags;
+	struct octeon_ethernet *priv = netdev_priv(dev);
+
 	queues_per_port = cvmx_pko_get_num_queues(priv->port);
 	/* Drain any pending packets in the free list */
 	for (qos = 0; qos < queues_per_port; qos++) {
@@ -89,24 +109,31 @@ void cvm_oct_free_tx_skbs(struct octeon_ethernet *priv)
 		skb_to_free = cvmx_fau_fetch_and_add32(priv->fau+qos*4, MAX_SKB_TO_FREE);
 		skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free, priv->fau+qos*4);
 
-		while (skb_to_free > 0) {
-			dev_kfree_skb_any(skb_dequeue(&priv->tx_free_list[qos]));
-			skb_to_free--;
+
+		total_freed += skb_to_free;
+		if (skb_to_free > 0) {
+			struct sk_buff *to_free_list = NULL;
+			spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
+			while (skb_to_free > 0) {
+				struct sk_buff *t = __skb_dequeue(&priv->tx_free_list[qos]);
+				t->next = to_free_list;
+				to_free_list = t;
+				skb_to_free--;
+			}
+			spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
+			/* Do the actual freeing outside of the lock. */
+			while (to_free_list) {
+				struct sk_buff *t = to_free_list;
+				to_free_list = to_free_list->next;
+				dev_kfree_skb_any(t);
+			}
 		}
+		total_remaining += skb_queue_len(&priv->tx_free_list[qos]);
 	}
-}
-
-enum hrtimer_restart cvm_oct_restart_tx(struct hrtimer *timer)
-{
-	struct octeon_ethernet *priv = container_of(timer, struct octeon_ethernet, tx_restart_timer);
-	struct net_device *dev = cvm_oct_device[priv->port];
-
-	cvm_oct_free_tx_skbs(priv);
-
-	if (netif_queue_stopped(dev))
+	if (total_freed >= 0 && netif_queue_stopped(dev))
 		netif_wake_queue(dev);
-
-	return HRTIMER_NORESTART;
+	if (total_remaining)
+		cvm_oct_kick_tx_poll_watchdog();
 }
 
 /**
@@ -129,6 +156,7 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 	struct sk_buff *to_free_list;
 	int32_t skb_to_free;
 	int32_t buffers_to_free;
+	u32 total_to_clean;
 	unsigned long flags;
 #if REUSE_SKBUFFS_WITHOUT_FREE
 	unsigned char *fpa_head;
@@ -232,7 +260,6 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 	pko_command.s.subone0 = 1;
 
 	pko_command.s.dontfree = 1;
-	pko_command.s.reg0 = priv->fau + qos * 4;
 
 	/* Build the PKO buffer pointer */
 	hw_buffer.u64 = 0;
@@ -327,7 +354,6 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 	 * We can use this buffer in the FPA.  We don't need the FAU
 	 * update anymore
 	 */
-	pko_command.s.reg0 = 0;
 	pko_command.s.dontfree = 0;
 
 	hw_buffer.s.back = ((unsigned long)skb->data >> 7) - ((unsigned long)fpa_head >> 7);
@@ -384,15 +410,17 @@ dont_put_skbuff_in_hw:
 	 * If we're sending faster than the receive can free them then
 	 * don't do the HW free.
 	 */
-	if ((buffers_to_free < -100) && !pko_command.s.dontfree) {
+	if ((buffers_to_free < -100) && !pko_command.s.dontfree)
 		pko_command.s.dontfree = 1;
-		pko_command.s.reg0 = priv->fau + qos * 4;
-	}
 
-	if (pko_command.s.dontfree)
+	if (pko_command.s.dontfree) {
 		queue_type = QUEUE_CORE;
-	else
+		pko_command.s.reg0 = priv->fau+qos*4;
+	} else {
 		queue_type = QUEUE_HW;
+	}
+	if (USE_ASYNC_IOBDMA)
+		cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH, FAU_TOTAL_TX_TO_CLEAN, 1);
 
 	spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 
@@ -402,10 +430,7 @@ dont_put_skbuff_in_hw:
 			/* Drop the lock when notifying the core.  */
 			spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 			netif_stop_queue(dev);
-			hrtimer_start(&priv->tx_restart_timer,
-				      priv->tx_restart_interval, HRTIMER_MODE_REL);
 			spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
-
 		} else {
 			/* If not using normal queueing.  */
 			queue_type = QUEUE_DROP;
@@ -460,11 +485,27 @@ skip_xmit:
 	}
 
 	if (USE_ASYNC_IOBDMA) {
+		CVMX_SYNCIOBDMA;
+		total_to_clean = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 		/* Restore the scratch area */
 		cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 		cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
+	} else {
+		total_to_clean = cvmx_fau_fetch_and_add32(FAU_TOTAL_TX_TO_CLEAN, 1);
 	}
 
+	if (total_to_clean & 0x3ff) {
+		/*
+		 * Schedule the cleanup tasklet every 1024 packets for
+		 * the pathological case of high traffic on one port
+		 * delaying clean up of packets on a different port
+		 * that is blocked waiting for the cleanup.
+		 */
+		tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
+	}
+
+	cvm_oct_kick_tx_poll_watchdog();
+
 	return NETDEV_TX_OK;
 }
 
@@ -624,7 +665,7 @@ int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
  *
  * @dev:    Device being shutdown
  */
-void cvm_oct_tx_shutdown(struct net_device *dev)
+void cvm_oct_tx_shutdown_dev(struct net_device *dev)
 {
 	struct octeon_ethernet *priv = netdev_priv(dev);
 	unsigned long flags;
@@ -638,3 +679,45 @@ void cvm_oct_tx_shutdown(struct net_device *dev)
 		spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 	}
 }
+
+static void cvm_oct_tx_do_cleanup(unsigned long arg)
+{
+	int port;
+
+	for (port = 0; port < TOTAL_NUMBER_OF_PORTS; port++) {
+		if (cvm_oct_device[port]) {
+			struct net_device *dev = cvm_oct_device[port];
+			cvm_oct_free_tx_skbs(dev);
+		}
+	}
+}
+
+static irqreturn_t cvm_oct_tx_cleanup_watchdog(int cpl, void *dev_id)
+{
+	/* Disable the interrupt.  */
+	cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
+	/* Do the work in the tasklet.  */
+	tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
+	return IRQ_HANDLED;
+}
+
+void cvm_oct_tx_initialize(void)
+{
+	int i;
+
+	/* Disable the interrupt.  */
+	cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
+	/* Register an IRQ hander for to receive CIU_TIMX(1) interrupts */
+	i = request_irq(OCTEON_IRQ_TIMER1,
+			cvm_oct_tx_cleanup_watchdog, 0,
+			"Ethernet", cvm_oct_device);
+
+	if (i)
+		panic("Could not acquire Ethernet IRQ %d\n", OCTEON_IRQ_TIMER1);
+}
+
+void cvm_oct_tx_shutdown(void)
+{
+	/* Free the interrupt handler */
+	free_irq(OCTEON_IRQ_TIMER1, cvm_oct_device);
+}
diff --git a/drivers/staging/octeon/ethernet-tx.h b/drivers/staging/octeon/ethernet-tx.h
index b628d8c..547680c 100644
--- a/drivers/staging/octeon/ethernet-tx.h
+++ b/drivers/staging/octeon/ethernet-tx.h
@@ -29,6 +29,6 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev);
 int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev);
 int cvm_oct_transmit_qos(struct net_device *dev, void *work_queue_entry,
 			 int do_free, int qos);
-void cvm_oct_tx_shutdown(struct net_device *dev);
-void cvm_oct_free_tx_skbs(struct octeon_ethernet *priv);
-enum hrtimer_restart cvm_oct_restart_tx(struct hrtimer *timer);
+void cvm_oct_tx_initialize(void);
+void cvm_oct_tx_shutdown(void);
+void cvm_oct_tx_shutdown_dev(struct net_device *dev);
diff --git a/drivers/staging/octeon/ethernet.c b/drivers/staging/octeon/ethernet.c
index 1771c10..f2ec8a6 100644
--- a/drivers/staging/octeon/ethernet.c
+++ b/drivers/staging/octeon/ethernet.c
@@ -140,6 +140,8 @@ atomic_t cvm_oct_poll_queue_stopping = ATOMIC_INIT(0);
  */
 struct net_device *cvm_oct_device[TOTAL_NUMBER_OF_PORTS];
 
+u64 cvm_oct_tx_poll_interval;
+
 static void cvm_oct_rx_refill_worker(struct work_struct *work);
 static DECLARE_DELAYED_WORK(cvm_oct_rx_refill_work, cvm_oct_rx_refill_worker);
 
@@ -159,18 +161,19 @@ static void cvm_oct_rx_refill_worker(struct work_struct *work)
 				   &cvm_oct_rx_refill_work, HZ);
 }
 
-static void cvm_oct_tx_clean_worker(struct work_struct *work)
+static void cvm_oct_preiodic_worker(struct work_struct *work)
 {
 	struct octeon_ethernet *priv = container_of(work,
 						    struct octeon_ethernet,
-						    tx_clean_work.work);
+						    port_periodic_work.work);
 
 	if (priv->poll)
 		priv->poll(cvm_oct_device[priv->port]);
-	cvm_oct_free_tx_skbs(priv);
+
 	cvm_oct_device[priv->port]->netdev_ops->ndo_get_stats(cvm_oct_device[priv->port]);
+
 	if (!atomic_read(&cvm_oct_poll_queue_stopping))
-		queue_delayed_work(cvm_oct_poll_queue, &priv->tx_clean_work, HZ);
+		queue_delayed_work(cvm_oct_poll_queue, &priv->port_periodic_work, HZ);
  }
 
 /**
@@ -662,6 +665,9 @@ static int __init cvm_oct_init_module(void)
 	 */
 	cvmx_fau_atomic_write32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
 
+	/* Initialize the FAU used for counting tx SKBs that need to be freed */
+	cvmx_fau_atomic_write32(FAU_TOTAL_TX_TO_CLEAN, 0);
+
 	if ((pow_send_group != -1)) {
 		struct net_device *dev;
 		pr_info("\tConfiguring device for POW only access\n");
@@ -670,18 +676,6 @@ static int __init cvm_oct_init_module(void)
 			/* Initialize the device private structure. */
 			struct octeon_ethernet *priv = netdev_priv(dev);
 
-			hrtimer_init(&priv->tx_restart_timer,
-				     CLOCK_MONOTONIC,
-				     HRTIMER_MODE_REL);
-			priv->tx_restart_timer.function = cvm_oct_restart_tx;
-
-			/*
-			 * Default for 10GE 5000nS enough time to
-			 * transmit about 100 64byte packtes.  1GE
-			 * interfaces will get 50000nS below.
-			 */
-			priv->tx_restart_interval = ktime_set(0, 5000);
-
 			dev->netdev_ops = &cvm_oct_pow_netdev_ops;
 			priv->imode = CVMX_HELPER_INTERFACE_MODE_DISABLED;
 			priv->port = CVMX_PIP_NUM_INPUT_PORTS;
@@ -725,9 +719,8 @@ static int __init cvm_oct_init_module(void)
 			/* Initialize the device private structure. */
 			priv = netdev_priv(dev);
 
-			INIT_DELAYED_WORK(&priv->tx_clean_work,
-					  cvm_oct_tx_clean_worker);
-
+			INIT_DELAYED_WORK(&priv->port_periodic_work,
+					  cvm_oct_preiodic_worker);
 			priv->imode = imode;
 			priv->port = port;
 			priv->queue = cvmx_pko_get_base_queue(priv->port);
@@ -763,7 +756,6 @@ static int __init cvm_oct_init_module(void)
 
 			case CVMX_HELPER_INTERFACE_MODE_SGMII:
 				dev->netdev_ops = &cvm_oct_sgmii_netdev_ops;
-				priv->tx_restart_interval = ktime_set(0, 50000);
 				strcpy(dev->name, "eth%d");
 				break;
 
@@ -775,7 +767,6 @@ static int __init cvm_oct_init_module(void)
 			case CVMX_HELPER_INTERFACE_MODE_RGMII:
 			case CVMX_HELPER_INTERFACE_MODE_GMII:
 				dev->netdev_ops = &cvm_oct_rgmii_netdev_ops;
-				priv->tx_restart_interval = ktime_set(0, 50000);
 				strcpy(dev->name, "eth%d");
 				break;
 			}
@@ -793,13 +784,19 @@ static int __init cvm_oct_init_module(void)
 				    cvmx_pko_get_num_queues(priv->port) *
 				    sizeof(uint32_t);
 				queue_delayed_work(cvm_oct_poll_queue,
-						   &priv->tx_clean_work, HZ);
+						   &priv->port_periodic_work, HZ);
 			}
 		}
 	}
 
+	cvm_oct_tx_initialize();
 	cvm_oct_rx_initialize();
 
+	/*
+	 * 150 uS: about 10 1500-byte packtes at 1GE.
+	 */
+	cvm_oct_tx_poll_interval = 150 * (octeon_get_clock_rate() / 1000000);
+
 	queue_delayed_work(cvm_oct_poll_queue, &cvm_oct_rx_refill_work, HZ);
 
 	return 0;
@@ -826,6 +823,8 @@ static void __exit cvm_oct_cleanup_module(void)
 	cancel_delayed_work_sync(&cvm_oct_rx_refill_work);
 
 	cvm_oct_rx_shutdown();
+	cvm_oct_tx_shutdown();
+
 	cvmx_pko_disable();
 
 	/* Free the ethernet devices */
@@ -833,9 +832,9 @@ static void __exit cvm_oct_cleanup_module(void)
 		if (cvm_oct_device[port]) {
 			struct net_device *dev = cvm_oct_device[port];
 			struct octeon_ethernet *priv = netdev_priv(dev);
-			cancel_delayed_work_sync(&priv->tx_clean_work);
+			cancel_delayed_work_sync(&priv->port_periodic_work);
 
-			cvm_oct_tx_shutdown(dev);
+			cvm_oct_tx_shutdown_dev(dev);
 			unregister_netdev(dev);
 			kfree(dev);
 			cvm_oct_device[port] = NULL;
diff --git a/drivers/staging/octeon/octeon-ethernet.h b/drivers/staging/octeon/octeon-ethernet.h
index 8d09210..db2a3cc 100644
--- a/drivers/staging/octeon/octeon-ethernet.h
+++ b/drivers/staging/octeon/octeon-ethernet.h
@@ -4,7 +4,7 @@
  * Contact: support@caviumnetworks.com
  * This file is part of the OCTEON SDK
  *
- * Copyright (c) 2003-2007 Cavium Networks
+ * Copyright (c) 2003-2010 Cavium Networks
  *
  * This file is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License, Version 2, as
@@ -31,8 +31,6 @@
 #ifndef OCTEON_ETHERNET_H
 #define OCTEON_ETHERNET_H
 
-#include <linux/hrtimer.h>
-
 /**
  * This is the definition of the Ethernet driver's private
  * driver state stored in netdev_priv(dev).
@@ -59,9 +57,7 @@ struct octeon_ethernet {
 	uint64_t link_info;
 	/* Called periodically to check link status */
 	void (*poll) (struct net_device *dev);
-	struct hrtimer		tx_restart_timer;
-	ktime_t			tx_restart_interval;
-	struct delayed_work	tx_clean_work;
+	struct delayed_work	port_periodic_work;
 	struct work_struct	port_work;	/* may be unused. */
 };
 
@@ -101,6 +97,7 @@ extern char pow_send_list[];
 extern struct net_device *cvm_oct_device[];
 extern struct workqueue_struct *cvm_oct_poll_queue;
 extern atomic_t cvm_oct_poll_queue_stopping;
+extern u64 cvm_oct_tx_poll_interval;
 
 extern int max_rx_cpus;
 extern int rx_napi_weight;
-- 
1.6.6

Re: [PATCH 4/4] Staging: Octeon: Free transmit SKBs in a timely manner.

[Eric Dumazet]

Le lundi 15 février 2010 à 12:13 -0800, David Daney a écrit :
> If we wait for the once-per-second cleanup to free transmit SKBs,
> sockets with small transmit buffer sizes might spend most of their
> time blocked waiting for the cleanup.
> 
> Normally we do a cleanup for each transmitted packet.  We add a
> watchdog type timer so that we also schedule a timeout for 150uS after
> a packet is transmitted.  The watchdog is reset for each transmitted
> packet, so for high packet rates, it never expires.  At these high
> rates, the cleanups are done for each packet so the extra watchdog
> initiated cleanups are not needed.

s/needed/fired/

Hmm, but re-arming a timer for each transmited packet must have a cost ?

> 
> Signed-off-by: David Daney <ddaney@caviumnetworks.com>

Is there any particular reason periodic is spelled preiodic ?

> ---
>  }
>  
> -static void cvm_oct_tx_clean_worker(struct work_struct *work)
> +static void cvm_oct_preiodic_worker(struct work_struct *work)
>  {



> -			INIT_DELAYED_WORK(&priv->tx_clean_work,
> -					  cvm_oct_tx_clean_worker);
> -
> +			INIT_DELAYED_WORK(&priv->port_periodic_work,
> +					  cvm_oct_preiodic_worker);

Re: [PATCH 4/4] Staging: Octeon: Free transmit SKBs in a timely manner.

[David Daney]

On 02/15/2010 12:27 PM, Eric Dumazet wrote:
> Le lundi 15 février 2010 à 12:13 -0800, David Daney a écrit :
>> If we wait for the once-per-second cleanup to free transmit SKBs,
>> sockets with small transmit buffer sizes might spend most of their
>> time blocked waiting for the cleanup.
>>
>> Normally we do a cleanup for each transmitted packet.  We add a
>> watchdog type timer so that we also schedule a timeout for 150uS after
>> a packet is transmitted.  The watchdog is reset for each transmitted
>> packet, so for high packet rates, it never expires.  At these high
>> rates, the cleanups are done for each packet so the extra watchdog
>> initiated cleanups are not needed.
>
> s/needed/fired/
>

or perhaps s/are not needed/are neither needed nor fired/

> Hmm, but re-arming a timer for each transmited packet must have a cost ?
>

The cost is fairly low (less than 10 processor clock cycles).  We didn't 
add this for amusement, people actually do things like only send UDP 
packets from userspace.  Since we can fill the transmit queue faster 
than it is emptied, the socket transmit buffer is quickly consumed.  If 
we don't free the SKBs in short order, the transmitting process get to 
take a long sleep (until our previous once per second clean up task was 
run).

>>
>> Signed-off-by: David Daney<ddaney@caviumnetworks.com>
>
> Is there any particular reason periodic is spelled preiodic ?

Ha!  No good reason.  I will correct the spelling.

Thanks,
David Daney


>
>> ---
>>   }
>>
>> -static void cvm_oct_tx_clean_worker(struct work_struct *work)
>> +static void cvm_oct_preiodic_worker(struct work_struct *work)
>>   {
>
>
>
>> -			INIT_DELAYED_WORK(&priv->tx_clean_work,
>> -					  cvm_oct_tx_clean_worker);
>> -
>> +			INIT_DELAYED_WORK(&priv->port_periodic_work,
>> +					  cvm_oct_preiodic_worker);
>
>
>
>

Re: [PATCH 0/4] Improvements to octeon_ethernet.

[Greg KH]

On Mon, Feb 15, 2010 at 12:12:22PM -0800, David Daney wrote:
> Here are a couple of improvements to the octeon_ethernet in
> drivers/staging/octeon.  The first patch is just cleanup, the rest are
> genuine bug fixes.
>
> I will reply with the four patches.
>
> We may want to merge via Ralf's linux-mips.org tree as Octeon is
> infact a MIPS based SOC and he has a bunch of other patches queued
> there that these depend on.

That's fine with me for them to go through that way.

thanks,

greg k-h

Re: [PATCH 4/4] Staging: Octeon: Free transmit SKBs in a timely manner.

[Eric Dumazet]

Le lundi 15 février 2010 à 12:41 -0800, David Daney a écrit :
> On 02/15/2010 12:27 PM, Eric Dumazet wrote:
> > Le lundi 15 février 2010 à 12:13 -0800, David Daney a écrit :
> >> If we wait for the once-per-second cleanup to free transmit SKBs,
> >> sockets with small transmit buffer sizes might spend most of their
> >> time blocked waiting for the cleanup.
> >>
> >> Normally we do a cleanup for each transmitted packet.  We add a
> >> watchdog type timer so that we also schedule a timeout for 150uS after
> >> a packet is transmitted.  The watchdog is reset for each transmitted
> >> packet, so for high packet rates, it never expires.  At these high
> >> rates, the cleanups are done for each packet so the extra watchdog
> >> initiated cleanups are not needed.
> >
> > s/needed/fired/
> >
> 
> or perhaps s/are not needed/are neither needed nor fired/
> 
> > Hmm, but re-arming a timer for each transmited packet must have a cost ?
> >
> 
> The cost is fairly low (less than 10 processor clock cycles).  We didn't 
> add this for amusement, people actually do things like only send UDP 
> packets from userspace.  Since we can fill the transmit queue faster 
> than it is emptied, the socket transmit buffer is quickly consumed.  If 
> we don't free the SKBs in short order, the transmitting process get to 
> take a long sleep (until our previous once per second clean up task was 
> run).

I understand this, but traditionaly, NIC drivers dont use a timer, but a
'TX complete' interrupt, that usually fires a few us after packet
submission on Gigabit speed.

A fast program could try to send X small udp packets in less than 150
us, X being greater than the size of your TX ring.

So your patch makes the window smaller, but it still is there (at
physical layer, we'll see a burst of packets, a ~100us delay, then a
second burst)

Re: [PATCH 4/4] Staging: Octeon: Free transmit SKBs in a timely manner.

[David Daney]

On 02/15/2010 01:11 PM, Eric Dumazet wrote:
> Le lundi 15 février 2010 à 12:41 -0800, David Daney a écrit :
>> On 02/15/2010 12:27 PM, Eric Dumazet wrote:
>>> Le lundi 15 février 2010 à 12:13 -0800, David Daney a écrit :
>>>> If we wait for the once-per-second cleanup to free transmit SKBs,
>>>> sockets with small transmit buffer sizes might spend most of their
>>>> time blocked waiting for the cleanup.
>>>>
>>>> Normally we do a cleanup for each transmitted packet.  We add a
>>>> watchdog type timer so that we also schedule a timeout for 150uS after
>>>> a packet is transmitted.  The watchdog is reset for each transmitted
>>>> packet, so for high packet rates, it never expires.  At these high
>>>> rates, the cleanups are done for each packet so the extra watchdog
>>>> initiated cleanups are not needed.
>>>
>>> s/needed/fired/
>>>
>>
>> or perhaps s/are not needed/are neither needed nor fired/
>>
>>> Hmm, but re-arming a timer for each transmited packet must have a cost ?
>>>
>>
>> The cost is fairly low (less than 10 processor clock cycles).  We didn't
>> add this for amusement, people actually do things like only send UDP
>> packets from userspace.  Since we can fill the transmit queue faster
>> than it is emptied, the socket transmit buffer is quickly consumed.  If
>> we don't free the SKBs in short order, the transmitting process get to
>> take a long sleep (until our previous once per second clean up task was
>> run).
>
> I understand this, but traditionaly, NIC drivers dont use a timer, but a
> 'TX complete' interrupt, that usually fires a few us after packet
> submission on Gigabit speed.
>

Indeed.  Lacking this type of interrupt, the watchdog seemed the best 
short term solution.

I am investigating the possibility of feeding TX complete notifications 
back through the RX path where it is possible to generate interrupts. 
The drawback to this is that it takes a lot more CPU cycles as well as 
added cache pressure.

> A fast program could try to send X small udp packets in less than 150
> us, X being greater than the size of your TX ring.

My TX queue (it is not a ring) size can be made arbitrarily large 
(currently 1000).  64bytes * 1000 packets * 10 bits/packet / 10e9 
bits/sec  == 640uS.  My watchdog will fire after less than 1/4 of the 
ring capacity is freed.

>
> So your patch makes the window smaller, but it still is there (at
> physical layer, we'll see a burst of packets, a ~100us delay, then a
> second burst)
>

With this patch, there will be no burstiness using default socket buffer 
sizes and packets of arbitrary size on a standard 1gig port.

On the 10gig ports there is the possibility for burstiness as you aptly 
explain.  However, in practice it would be difficult to arrange things 
to achieve sufficiently high packet rates, so we can live with it like this.

David Daney

[PATCH 4/4] Staging: Octeon: Free transmit SKBs in a timely manner (v2).

[David Daney]

If we wait for the once-per-second cleanup to free transmit SKBs,
sockets with small transmit buffer sizes might spend most of their
time blocked waiting for the cleanup.

Normally we do a cleanup for each transmitted packet.  We add a
watchdog type timer so that we also schedule a timeout for 150uS after
a packet is transmitted.  The watchdog is reset for each transmitted
packet, so for high packet rates, it never expires.  At these high
rates, the cleanups are done for each packet so the extra watchdog
initiated cleanups are neither needed nor triggered.

Signed-off-by: David Daney <ddaney@caviumnetworks.com>
CC: Eric Dumazet <eric.dumazet@gmail.com>
---

This version has spelling and comment changes based on feedback from
Eric Dumazet.

 drivers/staging/octeon/Kconfig            |    1 -
 drivers/staging/octeon/ethernet-defines.h |    5 +-
 drivers/staging/octeon/ethernet-tx.c      |  137 +++++++++++++++++++++++------
 drivers/staging/octeon/ethernet-tx.h      |    6 +-
 drivers/staging/octeon/ethernet.c         |   47 +++++-----
 drivers/staging/octeon/octeon-ethernet.h  |    9 +--
 6 files changed, 142 insertions(+), 63 deletions(-)

diff --git a/drivers/staging/octeon/Kconfig b/drivers/staging/octeon/Kconfig
index 579b8f1..638ad6b 100644
--- a/drivers/staging/octeon/Kconfig
+++ b/drivers/staging/octeon/Kconfig
@@ -3,7 +3,6 @@ config OCTEON_ETHERNET
 	depends on CPU_CAVIUM_OCTEON
 	select PHYLIB
 	select MDIO_OCTEON
-	select HIGH_RES_TIMERS
 	help
 	  This driver supports the builtin ethernet ports on Cavium
 	  Networks' products in the Octeon family. This driver supports the
diff --git a/drivers/staging/octeon/ethernet-defines.h b/drivers/staging/octeon/ethernet-defines.h
index 00a8561..6a2cd50 100644
--- a/drivers/staging/octeon/ethernet-defines.h
+++ b/drivers/staging/octeon/ethernet-defines.h
@@ -95,10 +95,11 @@
 /*#define DONT_WRITEBACK(x)         0   */
 
 /* Maximum number of SKBs to try to free per xmit packet. */
-#define MAX_SKB_TO_FREE 10
 #define MAX_OUT_QUEUE_DEPTH 1000
 
-#define FAU_NUM_PACKET_BUFFERS_TO_FREE (CVMX_FAU_REG_END - sizeof(uint32_t))
+#define FAU_TOTAL_TX_TO_CLEAN (CVMX_FAU_REG_END - sizeof(uint32_t))
+#define FAU_NUM_PACKET_BUFFERS_TO_FREE (FAU_TOTAL_TX_TO_CLEAN - sizeof(uint32_t))
+
 #define TOTAL_NUMBER_OF_PORTS       (CVMX_PIP_NUM_INPUT_PORTS+1)
 
 
diff --git a/drivers/staging/octeon/ethernet-tx.c b/drivers/staging/octeon/ethernet-tx.c
index 62258bd..5175247 100644
--- a/drivers/staging/octeon/ethernet-tx.c
+++ b/drivers/staging/octeon/ethernet-tx.c
@@ -48,6 +48,7 @@
 
 #include "cvmx-wqe.h"
 #include "cvmx-fau.h"
+#include "cvmx-pip.h"
 #include "cvmx-pko.h"
 #include "cvmx-helper.h"
 
@@ -66,6 +67,11 @@
 #define GET_SKBUFF_QOS(skb) 0
 #endif
 
+static void cvm_oct_tx_do_cleanup(unsigned long arg);
+static DECLARE_TASKLET(cvm_oct_tx_cleanup_tasklet, cvm_oct_tx_do_cleanup, 0);
+
+/* Maximum number of SKBs to try to free per xmit packet. */
+#define MAX_SKB_TO_FREE (MAX_OUT_QUEUE_DEPTH * 2)
 
 static inline int32_t cvm_oct_adjust_skb_to_free(int32_t skb_to_free, int fau)
 {
@@ -77,10 +83,24 @@ static inline int32_t cvm_oct_adjust_skb_to_free(int32_t skb_to_free, int fau)
 	return skb_to_free;
 }
 
-void cvm_oct_free_tx_skbs(struct octeon_ethernet *priv)
+static void cvm_oct_kick_tx_poll_watchdog(void)
+{
+	union cvmx_ciu_timx ciu_timx;
+	ciu_timx.u64 = 0;
+	ciu_timx.s.one_shot = 1;
+	ciu_timx.s.len = cvm_oct_tx_poll_interval;
+	cvmx_write_csr(CVMX_CIU_TIMX(1), ciu_timx.u64);
+}
+
+void cvm_oct_free_tx_skbs(struct net_device *dev)
 {
 	int32_t skb_to_free;
 	int qos, queues_per_port;
+	int total_freed = 0;
+	int total_remaining = 0;
+	unsigned long flags;
+	struct octeon_ethernet *priv = netdev_priv(dev);
+
 	queues_per_port = cvmx_pko_get_num_queues(priv->port);
 	/* Drain any pending packets in the free list */
 	for (qos = 0; qos < queues_per_port; qos++) {
@@ -89,24 +109,31 @@ void cvm_oct_free_tx_skbs(struct octeon_ethernet *priv)
 		skb_to_free = cvmx_fau_fetch_and_add32(priv->fau+qos*4, MAX_SKB_TO_FREE);
 		skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free, priv->fau+qos*4);
 
-		while (skb_to_free > 0) {
-			dev_kfree_skb_any(skb_dequeue(&priv->tx_free_list[qos]));
-			skb_to_free--;
+
+		total_freed += skb_to_free;
+		if (skb_to_free > 0) {
+			struct sk_buff *to_free_list = NULL;
+			spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
+			while (skb_to_free > 0) {
+				struct sk_buff *t = __skb_dequeue(&priv->tx_free_list[qos]);
+				t->next = to_free_list;
+				to_free_list = t;
+				skb_to_free--;
+			}
+			spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
+			/* Do the actual freeing outside of the lock. */
+			while (to_free_list) {
+				struct sk_buff *t = to_free_list;
+				to_free_list = to_free_list->next;
+				dev_kfree_skb_any(t);
+			}
 		}
+		total_remaining += skb_queue_len(&priv->tx_free_list[qos]);
 	}
-}
-
-enum hrtimer_restart cvm_oct_restart_tx(struct hrtimer *timer)
-{
-	struct octeon_ethernet *priv = container_of(timer, struct octeon_ethernet, tx_restart_timer);
-	struct net_device *dev = cvm_oct_device[priv->port];
-
-	cvm_oct_free_tx_skbs(priv);
-
-	if (netif_queue_stopped(dev))
+	if (total_freed >= 0 && netif_queue_stopped(dev))
 		netif_wake_queue(dev);
-
-	return HRTIMER_NORESTART;
+	if (total_remaining)
+		cvm_oct_kick_tx_poll_watchdog();
 }
 
 /**
@@ -129,6 +156,7 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 	struct sk_buff *to_free_list;
 	int32_t skb_to_free;
 	int32_t buffers_to_free;
+	u32 total_to_clean;
 	unsigned long flags;
 #if REUSE_SKBUFFS_WITHOUT_FREE
 	unsigned char *fpa_head;
@@ -232,7 +260,6 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 	pko_command.s.subone0 = 1;
 
 	pko_command.s.dontfree = 1;
-	pko_command.s.reg0 = priv->fau + qos * 4;
 
 	/* Build the PKO buffer pointer */
 	hw_buffer.u64 = 0;
@@ -327,7 +354,6 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 	 * We can use this buffer in the FPA.  We don't need the FAU
 	 * update anymore
 	 */
-	pko_command.s.reg0 = 0;
 	pko_command.s.dontfree = 0;
 
 	hw_buffer.s.back = ((unsigned long)skb->data >> 7) - ((unsigned long)fpa_head >> 7);
@@ -384,15 +410,17 @@ dont_put_skbuff_in_hw:
 	 * If we're sending faster than the receive can free them then
 	 * don't do the HW free.
 	 */
-	if ((buffers_to_free < -100) && !pko_command.s.dontfree) {
+	if ((buffers_to_free < -100) && !pko_command.s.dontfree)
 		pko_command.s.dontfree = 1;
-		pko_command.s.reg0 = priv->fau + qos * 4;
-	}
 
-	if (pko_command.s.dontfree)
+	if (pko_command.s.dontfree) {
 		queue_type = QUEUE_CORE;
-	else
+		pko_command.s.reg0 = priv->fau+qos*4;
+	} else {
 		queue_type = QUEUE_HW;
+	}
+	if (USE_ASYNC_IOBDMA)
+		cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH, FAU_TOTAL_TX_TO_CLEAN, 1);
 
 	spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 
@@ -402,10 +430,7 @@ dont_put_skbuff_in_hw:
 			/* Drop the lock when notifying the core.  */
 			spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 			netif_stop_queue(dev);
-			hrtimer_start(&priv->tx_restart_timer,
-				      priv->tx_restart_interval, HRTIMER_MODE_REL);
 			spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
-
 		} else {
 			/* If not using normal queueing.  */
 			queue_type = QUEUE_DROP;
@@ -460,11 +485,27 @@ skip_xmit:
 	}
 
 	if (USE_ASYNC_IOBDMA) {
+		CVMX_SYNCIOBDMA;
+		total_to_clean = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 		/* Restore the scratch area */
 		cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 		cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
+	} else {
+		total_to_clean = cvmx_fau_fetch_and_add32(FAU_TOTAL_TX_TO_CLEAN, 1);
 	}
 
+	if (total_to_clean & 0x3ff) {
+		/*
+		 * Schedule the cleanup tasklet every 1024 packets for
+		 * the pathological case of high traffic on one port
+		 * delaying clean up of packets on a different port
+		 * that is blocked waiting for the cleanup.
+		 */
+		tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
+	}
+
+	cvm_oct_kick_tx_poll_watchdog();
+
 	return NETDEV_TX_OK;
 }
 
@@ -624,7 +665,7 @@ int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
  *
  * @dev:    Device being shutdown
  */
-void cvm_oct_tx_shutdown(struct net_device *dev)
+void cvm_oct_tx_shutdown_dev(struct net_device *dev)
 {
 	struct octeon_ethernet *priv = netdev_priv(dev);
 	unsigned long flags;
@@ -638,3 +679,45 @@ void cvm_oct_tx_shutdown(struct net_device *dev)
 		spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 	}
 }
+
+static void cvm_oct_tx_do_cleanup(unsigned long arg)
+{
+	int port;
+
+	for (port = 0; port < TOTAL_NUMBER_OF_PORTS; port++) {
+		if (cvm_oct_device[port]) {
+			struct net_device *dev = cvm_oct_device[port];
+			cvm_oct_free_tx_skbs(dev);
+		}
+	}
+}
+
+static irqreturn_t cvm_oct_tx_cleanup_watchdog(int cpl, void *dev_id)
+{
+	/* Disable the interrupt.  */
+	cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
+	/* Do the work in the tasklet.  */
+	tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
+	return IRQ_HANDLED;
+}
+
+void cvm_oct_tx_initialize(void)
+{
+	int i;
+
+	/* Disable the interrupt.  */
+	cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
+	/* Register an IRQ hander for to receive CIU_TIMX(1) interrupts */
+	i = request_irq(OCTEON_IRQ_TIMER1,
+			cvm_oct_tx_cleanup_watchdog, 0,
+			"Ethernet", cvm_oct_device);
+
+	if (i)
+		panic("Could not acquire Ethernet IRQ %d\n", OCTEON_IRQ_TIMER1);
+}
+
+void cvm_oct_tx_shutdown(void)
+{
+	/* Free the interrupt handler */
+	free_irq(OCTEON_IRQ_TIMER1, cvm_oct_device);
+}
diff --git a/drivers/staging/octeon/ethernet-tx.h b/drivers/staging/octeon/ethernet-tx.h
index b628d8c..547680c 100644
--- a/drivers/staging/octeon/ethernet-tx.h
+++ b/drivers/staging/octeon/ethernet-tx.h
@@ -29,6 +29,6 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev);
 int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev);
 int cvm_oct_transmit_qos(struct net_device *dev, void *work_queue_entry,
 			 int do_free, int qos);
-void cvm_oct_tx_shutdown(struct net_device *dev);
-void cvm_oct_free_tx_skbs(struct octeon_ethernet *priv);
-enum hrtimer_restart cvm_oct_restart_tx(struct hrtimer *timer);
+void cvm_oct_tx_initialize(void);
+void cvm_oct_tx_shutdown(void);
+void cvm_oct_tx_shutdown_dev(struct net_device *dev);
diff --git a/drivers/staging/octeon/ethernet.c b/drivers/staging/octeon/ethernet.c
index 1771c10..5ee60ab 100644
--- a/drivers/staging/octeon/ethernet.c
+++ b/drivers/staging/octeon/ethernet.c
@@ -140,6 +140,8 @@ atomic_t cvm_oct_poll_queue_stopping = ATOMIC_INIT(0);
  */
 struct net_device *cvm_oct_device[TOTAL_NUMBER_OF_PORTS];
 
+u64 cvm_oct_tx_poll_interval;
+
 static void cvm_oct_rx_refill_worker(struct work_struct *work);
 static DECLARE_DELAYED_WORK(cvm_oct_rx_refill_work, cvm_oct_rx_refill_worker);
 
@@ -159,18 +161,19 @@ static void cvm_oct_rx_refill_worker(struct work_struct *work)
 				   &cvm_oct_rx_refill_work, HZ);
 }
 
-static void cvm_oct_tx_clean_worker(struct work_struct *work)
+static void cvm_oct_periodic_worker(struct work_struct *work)
 {
 	struct octeon_ethernet *priv = container_of(work,
 						    struct octeon_ethernet,
-						    tx_clean_work.work);
+						    port_periodic_work.work);
 
 	if (priv->poll)
 		priv->poll(cvm_oct_device[priv->port]);
-	cvm_oct_free_tx_skbs(priv);
+
 	cvm_oct_device[priv->port]->netdev_ops->ndo_get_stats(cvm_oct_device[priv->port]);
+
 	if (!atomic_read(&cvm_oct_poll_queue_stopping))
-		queue_delayed_work(cvm_oct_poll_queue, &priv->tx_clean_work, HZ);
+		queue_delayed_work(cvm_oct_poll_queue, &priv->port_periodic_work, HZ);
  }
 
 /**
@@ -662,6 +665,9 @@ static int __init cvm_oct_init_module(void)
 	 */
 	cvmx_fau_atomic_write32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
 
+	/* Initialize the FAU used for counting tx SKBs that need to be freed */
+	cvmx_fau_atomic_write32(FAU_TOTAL_TX_TO_CLEAN, 0);
+
 	if ((pow_send_group != -1)) {
 		struct net_device *dev;
 		pr_info("\tConfiguring device for POW only access\n");
@@ -670,18 +676,6 @@ static int __init cvm_oct_init_module(void)
 			/* Initialize the device private structure. */
 			struct octeon_ethernet *priv = netdev_priv(dev);
 
-			hrtimer_init(&priv->tx_restart_timer,
-				     CLOCK_MONOTONIC,
-				     HRTIMER_MODE_REL);
-			priv->tx_restart_timer.function = cvm_oct_restart_tx;
-
-			/*
-			 * Default for 10GE 5000nS enough time to
-			 * transmit about 100 64byte packtes.  1GE
-			 * interfaces will get 50000nS below.
-			 */
-			priv->tx_restart_interval = ktime_set(0, 5000);
-
 			dev->netdev_ops = &cvm_oct_pow_netdev_ops;
 			priv->imode = CVMX_HELPER_INTERFACE_MODE_DISABLED;
 			priv->port = CVMX_PIP_NUM_INPUT_PORTS;
@@ -725,9 +719,8 @@ static int __init cvm_oct_init_module(void)
 			/* Initialize the device private structure. */
 			priv = netdev_priv(dev);
 
-			INIT_DELAYED_WORK(&priv->tx_clean_work,
-					  cvm_oct_tx_clean_worker);
-
+			INIT_DELAYED_WORK(&priv->port_periodic_work,
+					  cvm_oct_periodic_worker);
 			priv->imode = imode;
 			priv->port = port;
 			priv->queue = cvmx_pko_get_base_queue(priv->port);
@@ -763,7 +756,6 @@ static int __init cvm_oct_init_module(void)
 
 			case CVMX_HELPER_INTERFACE_MODE_SGMII:
 				dev->netdev_ops = &cvm_oct_sgmii_netdev_ops;
-				priv->tx_restart_interval = ktime_set(0, 50000);
 				strcpy(dev->name, "eth%d");
 				break;
 
@@ -775,7 +767,6 @@ static int __init cvm_oct_init_module(void)
 			case CVMX_HELPER_INTERFACE_MODE_RGMII:
 			case CVMX_HELPER_INTERFACE_MODE_GMII:
 				dev->netdev_ops = &cvm_oct_rgmii_netdev_ops;
-				priv->tx_restart_interval = ktime_set(0, 50000);
 				strcpy(dev->name, "eth%d");
 				break;
 			}
@@ -793,13 +784,19 @@ static int __init cvm_oct_init_module(void)
 				    cvmx_pko_get_num_queues(priv->port) *
 				    sizeof(uint32_t);
 				queue_delayed_work(cvm_oct_poll_queue,
-						   &priv->tx_clean_work, HZ);
+						   &priv->port_periodic_work, HZ);
 			}
 		}
 	}
 
+	cvm_oct_tx_initialize();
 	cvm_oct_rx_initialize();
 
+	/*
+	 * 150 uS: about 10 1500-byte packtes at 1GE.
+	 */
+	cvm_oct_tx_poll_interval = 150 * (octeon_get_clock_rate() / 1000000);
+
 	queue_delayed_work(cvm_oct_poll_queue, &cvm_oct_rx_refill_work, HZ);
 
 	return 0;
@@ -826,6 +823,8 @@ static void __exit cvm_oct_cleanup_module(void)
 	cancel_delayed_work_sync(&cvm_oct_rx_refill_work);
 
 	cvm_oct_rx_shutdown();
+	cvm_oct_tx_shutdown();
+
 	cvmx_pko_disable();
 
 	/* Free the ethernet devices */
@@ -833,9 +832,9 @@ static void __exit cvm_oct_cleanup_module(void)
 		if (cvm_oct_device[port]) {
 			struct net_device *dev = cvm_oct_device[port];
 			struct octeon_ethernet *priv = netdev_priv(dev);
-			cancel_delayed_work_sync(&priv->tx_clean_work);
+			cancel_delayed_work_sync(&priv->port_periodic_work);
 
-			cvm_oct_tx_shutdown(dev);
+			cvm_oct_tx_shutdown_dev(dev);
 			unregister_netdev(dev);
 			kfree(dev);
 			cvm_oct_device[port] = NULL;
diff --git a/drivers/staging/octeon/octeon-ethernet.h b/drivers/staging/octeon/octeon-ethernet.h
index 8d09210..db2a3cc 100644
--- a/drivers/staging/octeon/octeon-ethernet.h
+++ b/drivers/staging/octeon/octeon-ethernet.h
@@ -4,7 +4,7 @@
  * Contact: support@caviumnetworks.com
  * This file is part of the OCTEON SDK
  *
- * Copyright (c) 2003-2007 Cavium Networks
+ * Copyright (c) 2003-2010 Cavium Networks
  *
  * This file is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License, Version 2, as
@@ -31,8 +31,6 @@
 #ifndef OCTEON_ETHERNET_H
 #define OCTEON_ETHERNET_H
 
-#include <linux/hrtimer.h>
-
 /**
  * This is the definition of the Ethernet driver's private
  * driver state stored in netdev_priv(dev).
@@ -59,9 +57,7 @@ struct octeon_ethernet {
 	uint64_t link_info;
 	/* Called periodically to check link status */
 	void (*poll) (struct net_device *dev);
-	struct hrtimer		tx_restart_timer;
-	ktime_t			tx_restart_interval;
-	struct delayed_work	tx_clean_work;
+	struct delayed_work	port_periodic_work;
 	struct work_struct	port_work;	/* may be unused. */
 };
 
@@ -101,6 +97,7 @@ extern char pow_send_list[];
 extern struct net_device *cvm_oct_device[];
 extern struct workqueue_struct *cvm_oct_poll_queue;
 extern atomic_t cvm_oct_poll_queue_stopping;
+extern u64 cvm_oct_tx_poll_interval;
 
 extern int max_rx_cpus;
 extern int rx_napi_weight;
-- 
1.6.6

Re: [PATCH 1/4] Staging: octeon: remove unneeded includes

[Ralf Baechle]

Thanks, queued for 2.6.34,

  Ralf

Re: [PATCH 2/4] Staging: Octeon: Run phy bus accesses on a workqueue.

[Ralf Baechle]

Thanks, queued for 2.6.34,

  Ralf

Re: [PATCH 3/4] MIPS: Octeon: Do proper acknowledgment of CIU timer interrupts.

[Ralf Baechle]

Thanks, queued for 2.6.34,

  Ralf

Re: [PATCH 4/4] Staging: Octeon: Free transmit SKBs in a timely manner (v2).

[Ralf Baechle]

Thanks, queued for 2.6.34,

  Ralf