[PATCH 1/9] pasemi_mac: Fix TX interrupt threshold

[PATCH 1/9] pasemi_mac: Fix TX interrupt threshold

[Olof Johansson]

[-- Attachment #1: pasemi_mac-rx-intr-fix --]
[-- Type: text/plain, Size: 779 bytes --]

It was mistakenly set to interrupt on the second packet instead of first, causing
some interesting latency behaviour.


Signed-off-by: Olof Johansson <olof@lixom.net>

Index: netdev-2.6/drivers/net/pasemi_mac.c
===================================================================
--- netdev-2.6.orig/drivers/net/pasemi_mac.c
+++ netdev-2.6/drivers/net/pasemi_mac.c
@@ -755,7 +755,7 @@ static int pasemi_mac_open(struct net_de
 	flags |= PAS_MAC_CFG_PCFG_TSR_1G | PAS_MAC_CFG_PCFG_SPD_1G;
 
 	pci_write_config_dword(mac->iob_pdev, PAS_IOB_DMA_RXCH_CFG(mac->dma_rxch),
-			       PAS_IOB_DMA_RXCH_CFG_CNTTH(1));
+			       PAS_IOB_DMA_RXCH_CFG_CNTTH(0));
 
 	pci_write_config_dword(mac->iob_pdev, PAS_IOB_DMA_TXCH_CFG(mac->dma_txch),
 			       PAS_IOB_DMA_TXCH_CFG_CNTTH(32));

--

Re: [PATCH 1/9] pasemi_mac: Fix TX interrupt threshold

[Jeff Garzik]

Olof Johansson wrote:
> It was mistakenly set to interrupt on the second packet instead of first, causing
> some interesting latency behaviour.
> 
> 
> Signed-off-by: Olof Johansson <olof@lixom.net>

applied

[PATCH 2/9] pasemi_mac: Clean TX ring in poll

[Olof Johansson]

[-- Attachment #1: pasemi_mac-clean-tx-in-poll --]
[-- Type: text/plain, Size: 635 bytes --]

Clean the TX ring in the poll call, to avoid sitting on mapped buffers
for a long time. NFS doesn't seem to like it much, for example.


Signed-off-by: Olof Johansson <olof@lixom.net>

Index: netdev-2.6/drivers/net/pasemi_mac.c
===================================================================
--- netdev-2.6.orig/drivers/net/pasemi_mac.c
+++ netdev-2.6/drivers/net/pasemi_mac.c
@@ -1052,6 +1052,7 @@ static int pasemi_mac_poll(struct net_de
 	int pkts, limit = min(*budget, dev->quota);
 	struct pasemi_mac *mac = netdev_priv(dev);
 
+	pasemi_mac_clean_tx(mac);
 	pkts = pasemi_mac_clean_rx(mac, limit);
 
 	dev->quota -= pkts;

--

Re: [PATCH 2/9] pasemi_mac: Clean TX ring in poll

[Jeff Garzik]

Olof Johansson wrote:
> Clean the TX ring in the poll call, to avoid sitting on mapped buffers
> for a long time. NFS doesn't seem to like it much, for example.
> 
> 
> Signed-off-by: Olof Johansson <olof@lixom.net>
> 
> Index: netdev-2.6/drivers/net/pasemi_mac.c
> ===================================================================
> --- netdev-2.6.orig/drivers/net/pasemi_mac.c
> +++ netdev-2.6/drivers/net/pasemi_mac.c
> @@ -1052,6 +1052,7 @@ static int pasemi_mac_poll(struct net_de
>  	int pkts, limit = min(*budget, dev->quota);
>  	struct pasemi_mac *mac = netdev_priv(dev);
>  
> +	pasemi_mac_clean_tx(mac);
>  	pkts = pasemi_mac_clean_rx(mac, limit);

Why is this needed?  Is your TX interrupt mitigated or delayed somehow?

In general, drivers should only clean TX in one place.  Doing so in 
multiple places tends to indicate a bug somewhere.

You are correct that you should not sit on buffers for too long -- that 
increases latency of all who are waiting on those buffers to be sent.

Re: [PATCH 2/9] pasemi_mac: Clean TX ring in poll

[Olof Johansson]

On Mon, Jul 02, 2007 at 08:34:22AM -0400, Jeff Garzik wrote:
> >Index: netdev-2.6/drivers/net/pasemi_mac.c
> >===================================================================
> >--- netdev-2.6.orig/drivers/net/pasemi_mac.c
> >+++ netdev-2.6/drivers/net/pasemi_mac.c
> >@@ -1052,6 +1052,7 @@ static int pasemi_mac_poll(struct net_de
> > 	int pkts, limit = min(*budget, dev->quota);
> > 	struct pasemi_mac *mac = netdev_priv(dev);
> > 
> >+	pasemi_mac_clean_tx(mac);
> > 	pkts = pasemi_mac_clean_rx(mac, limit);
> 
> Why is this needed?  Is your TX interrupt mitigated or delayed somehow?

On some level hardware, we only have the possibility of taking the
TX interrupt after a certain number of packets, and no way of doing a
(hw) timer-based one (i.e.  after no packets have been sent for x ms,
interrupt). Taking an interrupt for every TX packet isn't an interesting
proposal, and doing it solely based on count > 1 can leave stale packets
there for a long time when the network is otherwise idle.

So, I would need to do a software based timer. The poll function seemed like
a good enough solution for now though, since it's called periodically even
when the NAPI mode is such that interrupts are enabled.

> In general, drivers should only clean TX in one place.  Doing so in 
> multiple places tends to indicate a bug somewhere.

Yes, I wish I could just do it in the interrupt handler. I could do it
with a dedicated timer if you prefer, but I'd still need to have two
ways of entering the tx clean function.


-Olof

[PATCH 3/9] pasemi_mac: Abstract out register access

[Olof Johansson]

[-- Attachment #1: pasemi_mac-regaccess --]
[-- Type: text/plain, Size: 13443 bytes --]

Abstract out the PCI config read/write accesses into reg read/write ones, still
calling the pci accessors on the back end.

Signed-off-by: Olof Johansson <olof@lixom.net>


Index: netdev-2.6/drivers/net/pasemi_mac.c
===================================================================
--- netdev-2.6.orig/drivers/net/pasemi_mac.c
+++ netdev-2.6/drivers/net/pasemi_mac.c
@@ -81,6 +81,48 @@ MODULE_PARM_DESC(debug, "PA Semi MAC bit
 
 static struct pasdma_status *dma_status;
 
+static unsigned int read_iob_reg(struct pasemi_mac *mac, unsigned int reg)
+{
+	unsigned int val;
+
+	pci_read_config_dword(mac->iob_pdev, reg, &val);
+	return val;
+}
+
+static void write_iob_reg(struct pasemi_mac *mac, unsigned int reg,
+			  unsigned int val)
+{
+	pci_write_config_dword(mac->iob_pdev, reg, val);
+}
+
+static unsigned int read_mac_reg(struct pasemi_mac *mac, unsigned int reg)
+{
+	unsigned int val;
+
+	pci_read_config_dword(mac->pdev, reg, &val);
+	return val;
+}
+
+static void write_mac_reg(struct pasemi_mac *mac, unsigned int reg,
+			  unsigned int val)
+{
+	pci_write_config_dword(mac->pdev, reg, val);
+}
+
+static unsigned int read_dma_reg(struct pasemi_mac *mac, unsigned int reg)
+{
+	unsigned int val;
+
+	pci_read_config_dword(mac->dma_pdev, reg, &val);
+	return val;
+}
+
+static void write_dma_reg(struct pasemi_mac *mac, unsigned int reg,
+			  unsigned int val)
+{
+	pci_write_config_dword(mac->dma_pdev, reg, val);
+}
+
 static int pasemi_get_mac_addr(struct pasemi_mac *mac)
 {
 	struct pci_dev *pdev = mac->pdev;
@@ -166,22 +208,21 @@ static int pasemi_mac_setup_rx_resources
 
 	memset(ring->buffers, 0, RX_RING_SIZE * sizeof(u64));
 
-	pci_write_config_dword(mac->dma_pdev, PAS_DMA_RXCHAN_BASEL(chan_id),
-			       PAS_DMA_RXCHAN_BASEL_BRBL(ring->dma));
+	write_dma_reg(mac, PAS_DMA_RXCHAN_BASEL(chan_id), PAS_DMA_RXCHAN_BASEL_BRBL(ring->dma));
+
+	write_dma_reg(mac, PAS_DMA_RXCHAN_BASEU(chan_id),
+			   PAS_DMA_RXCHAN_BASEU_BRBH(ring->dma >> 32) |
+			   PAS_DMA_RXCHAN_BASEU_SIZ(RX_RING_SIZE >> 2));
+
+	write_dma_reg(mac, PAS_DMA_RXCHAN_CFG(chan_id),
+			   PAS_DMA_RXCHAN_CFG_HBU(1));
 
-	pci_write_config_dword(mac->dma_pdev, PAS_DMA_RXCHAN_BASEU(chan_id),
-			       PAS_DMA_RXCHAN_BASEU_BRBH(ring->dma >> 32) |
-			       PAS_DMA_RXCHAN_BASEU_SIZ(RX_RING_SIZE >> 2));
-
-	pci_write_config_dword(mac->dma_pdev, PAS_DMA_RXCHAN_CFG(chan_id),
-			       PAS_DMA_RXCHAN_CFG_HBU(1));
-
-	pci_write_config_dword(mac->dma_pdev, PAS_DMA_RXINT_BASEL(mac->dma_if),
-			       PAS_DMA_RXINT_BASEL_BRBL(__pa(ring->buffers)));
-
-	pci_write_config_dword(mac->dma_pdev, PAS_DMA_RXINT_BASEU(mac->dma_if),
-			       PAS_DMA_RXINT_BASEU_BRBH(__pa(ring->buffers) >> 32) |
-			       PAS_DMA_RXINT_BASEU_SIZ(RX_RING_SIZE >> 3));
+	write_dma_reg(mac, PAS_DMA_RXINT_BASEL(mac->dma_if),
+			   PAS_DMA_RXINT_BASEL_BRBL(__pa(ring->buffers)));
+
+	write_dma_reg(mac, PAS_DMA_RXINT_BASEU(mac->dma_if),
+			   PAS_DMA_RXINT_BASEU_BRBH(__pa(ring->buffers) >> 32) |
+			   PAS_DMA_RXINT_BASEU_SIZ(RX_RING_SIZE >> 3));
 
 	ring->next_to_fill = 0;
 	ring->next_to_clean = 0;
@@ -233,18 +274,18 @@ static int pasemi_mac_setup_tx_resources
 
 	memset(ring->desc, 0, TX_RING_SIZE * sizeof(struct pas_dma_xct_descr));
 
-	pci_write_config_dword(mac->dma_pdev, PAS_DMA_TXCHAN_BASEL(chan_id),
-			       PAS_DMA_TXCHAN_BASEL_BRBL(ring->dma));
+	write_dma_reg(mac, PAS_DMA_TXCHAN_BASEL(chan_id),
+			   PAS_DMA_TXCHAN_BASEL_BRBL(ring->dma));
 	val = PAS_DMA_TXCHAN_BASEU_BRBH(ring->dma >> 32);
 	val |= PAS_DMA_TXCHAN_BASEU_SIZ(TX_RING_SIZE >> 2);
 
-	pci_write_config_dword(mac->dma_pdev, PAS_DMA_TXCHAN_BASEU(chan_id), val);
+	write_dma_reg(mac, PAS_DMA_TXCHAN_BASEU(chan_id), val);
 
-	pci_write_config_dword(mac->dma_pdev, PAS_DMA_TXCHAN_CFG(chan_id),
-			       PAS_DMA_TXCHAN_CFG_TY_IFACE |
-			       PAS_DMA_TXCHAN_CFG_TATTR(mac->dma_if) |
-			       PAS_DMA_TXCHAN_CFG_UP |
-			       PAS_DMA_TXCHAN_CFG_WT(2));
+	write_dma_reg(mac, PAS_DMA_TXCHAN_CFG(chan_id),
+			   PAS_DMA_TXCHAN_CFG_TY_IFACE |
+			   PAS_DMA_TXCHAN_CFG_TATTR(mac->dma_if) |
+			   PAS_DMA_TXCHAN_CFG_UP |
+			   PAS_DMA_TXCHAN_CFG_WT(2));
 
 	ring->next_to_use = 0;
 	ring->next_to_clean = 0;
@@ -383,12 +424,8 @@ static void pasemi_mac_replenish_rx_ring
 
 	wmb();
 
-	pci_write_config_dword(mac->dma_pdev,
-			       PAS_DMA_RXCHAN_INCR(mac->dma_rxch),
-			       limit - count);
-	pci_write_config_dword(mac->dma_pdev,
-			       PAS_DMA_RXINT_INCR(mac->dma_if),
-			       limit - count);
+	write_dma_reg(mac, PAS_DMA_RXCHAN_INCR(mac->dma_rxch), limit - count);
+	write_dma_reg(mac, PAS_DMA_RXINT_INCR(mac->dma_if), limit - count);
 
 	mac->rx->next_to_fill += limit - count;
 }
@@ -404,9 +441,7 @@ static void pasemi_mac_restart_rx_intr(s
 
 	reg = PAS_IOB_DMA_RXCH_RESET_PCNT(pcnt) | PAS_IOB_DMA_RXCH_RESET_PINTC;
 
-	pci_write_config_dword(mac->iob_pdev,
-			       PAS_IOB_DMA_RXCH_RESET(mac->dma_rxch),
-			       reg);
+	write_iob_reg(mac, PAS_IOB_DMA_RXCH_RESET(mac->dma_rxch), reg);
 }
 
 static void pasemi_mac_restart_tx_intr(struct pasemi_mac *mac)
@@ -418,8 +453,7 @@ static void pasemi_mac_restart_tx_intr(s
 
 	reg = PAS_IOB_DMA_TXCH_RESET_PCNT(pcnt) | PAS_IOB_DMA_TXCH_RESET_PINTC;
 
-	pci_write_config_dword(mac->iob_pdev,
-			       PAS_IOB_DMA_TXCH_RESET(mac->dma_txch), reg);
+	write_iob_reg(mac, PAS_IOB_DMA_TXCH_RESET(mac->dma_txch), reg);
 }
 
 
@@ -574,8 +608,6 @@ static irqreturn_t pasemi_mac_rx_intr(in
 	 * all others.
 	 */
 
-	pci_read_config_dword(mac->dma_pdev, PAS_DMA_RXINT_RCMDSTA(mac->dma_if), &reg);
-
 	reg = 0;
 	if (*mac->rx_status & PAS_STATUS_SOFT)
 		reg |= PAS_IOB_DMA_RXCH_RESET_SINTC;
@@ -586,9 +618,7 @@ static irqreturn_t pasemi_mac_rx_intr(in
 
 	netif_rx_schedule(dev);
 
-	pci_write_config_dword(mac->iob_pdev,
-			       PAS_IOB_DMA_RXCH_RESET(mac->dma_rxch), reg);
-
+	write_iob_reg(mac, PAS_IOB_DMA_RXCH_RESET(mac->dma_rxch), reg);
 
 	return IRQ_HANDLED;
 }
@@ -613,9 +643,7 @@ static irqreturn_t pasemi_mac_tx_intr(in
 	if (*mac->tx_status & PAS_STATUS_ERROR)
 		reg |= PAS_IOB_DMA_TXCH_RESET_DINTC;
 
-	pci_write_config_dword(mac->iob_pdev,
-			       PAS_IOB_DMA_TXCH_RESET(mac->dma_txch),
-			       reg);
+	write_iob_reg(mac, PAS_IOB_DMA_TXCH_RESET(mac->dma_txch), reg);
 
 	return IRQ_HANDLED;
 }
@@ -641,7 +669,7 @@ static void pasemi_adjust_link(struct ne
 	} else
 		netif_carrier_on(dev);
 
-	pci_read_config_dword(mac->pdev, PAS_MAC_CFG_PCFG, &flags);
+	flags = read_mac_reg(mac, PAS_MAC_CFG_PCFG);
 	new_flags = flags & ~(PAS_MAC_CFG_PCFG_HD | PAS_MAC_CFG_PCFG_SPD_M |
 			      PAS_MAC_CFG_PCFG_TSR_M);
 
@@ -673,7 +701,7 @@ static void pasemi_adjust_link(struct ne
 	mac->link = mac->phydev->link;
 
 	if (new_flags != flags)
-		pci_write_config_dword(mac->pdev, PAS_MAC_CFG_PCFG, new_flags);
+		write_mac_reg(mac, PAS_MAC_CFG_PCFG, new_flags);
 
 	if (msg && netif_msg_link(mac))
 		printk(KERN_INFO "%s: Link is up at %d Mbps, %s duplex.\n",
@@ -736,39 +764,37 @@ static int pasemi_mac_open(struct net_de
 	int ret;
 
 	/* enable rx section */
-	pci_write_config_dword(mac->dma_pdev, PAS_DMA_COM_RXCMD,
-			       PAS_DMA_COM_RXCMD_EN);
+	write_dma_reg(mac, PAS_DMA_COM_RXCMD, PAS_DMA_COM_RXCMD_EN);
 
 	/* enable tx section */
-	pci_write_config_dword(mac->dma_pdev, PAS_DMA_COM_TXCMD,
-			       PAS_DMA_COM_TXCMD_EN);
+	write_dma_reg(mac, PAS_DMA_COM_TXCMD, PAS_DMA_COM_TXCMD_EN);
 
 	flags = PAS_MAC_CFG_TXP_FCE | PAS_MAC_CFG_TXP_FPC(3) |
 		PAS_MAC_CFG_TXP_SL(3) | PAS_MAC_CFG_TXP_COB(0xf) |
 		PAS_MAC_CFG_TXP_TIFT(8) | PAS_MAC_CFG_TXP_TIFG(12);
 
-	pci_write_config_dword(mac->pdev, PAS_MAC_CFG_TXP, flags);
+	write_mac_reg(mac, PAS_MAC_CFG_TXP, flags);
 
 	flags = PAS_MAC_CFG_PCFG_S1 | PAS_MAC_CFG_PCFG_PE |
 		PAS_MAC_CFG_PCFG_PR | PAS_MAC_CFG_PCFG_CE;
 
 	flags |= PAS_MAC_CFG_PCFG_TSR_1G | PAS_MAC_CFG_PCFG_SPD_1G;
 
-	pci_write_config_dword(mac->iob_pdev, PAS_IOB_DMA_RXCH_CFG(mac->dma_rxch),
-			       PAS_IOB_DMA_RXCH_CFG_CNTTH(0));
+	write_iob_reg(mac, PAS_IOB_DMA_RXCH_CFG(mac->dma_rxch),
+			   PAS_IOB_DMA_RXCH_CFG_CNTTH(0));
 
-	pci_write_config_dword(mac->iob_pdev, PAS_IOB_DMA_TXCH_CFG(mac->dma_txch),
-			       PAS_IOB_DMA_TXCH_CFG_CNTTH(32));
+	write_iob_reg(mac, PAS_IOB_DMA_TXCH_CFG(mac->dma_txch),
+			   PAS_IOB_DMA_TXCH_CFG_CNTTH(32));
 
 	/* Clear out any residual packet count state from firmware */
 	pasemi_mac_restart_rx_intr(mac);
 	pasemi_mac_restart_tx_intr(mac);
 
 	/* 0xffffff is max value, about 16ms */
-	pci_write_config_dword(mac->iob_pdev, PAS_IOB_DMA_COM_TIMEOUTCFG,
-			       PAS_IOB_DMA_COM_TIMEOUTCFG_TCNT(0xffffff));
+	write_iob_reg(mac, PAS_IOB_DMA_COM_TIMEOUTCFG,
+			   PAS_IOB_DMA_COM_TIMEOUTCFG_TCNT(0xffffff));
 
-	pci_write_config_dword(mac->pdev, PAS_MAC_CFG_PCFG, flags);
+	write_mac_reg(mac, PAS_MAC_CFG_PCFG, flags);
 
 	ret = pasemi_mac_setup_rx_resources(dev);
 	if (ret)
@@ -778,25 +804,22 @@ static int pasemi_mac_open(struct net_de
 	if (ret)
 		goto out_tx_resources;
 
-	pci_write_config_dword(mac->pdev, PAS_MAC_IPC_CHNL,
-			       PAS_MAC_IPC_CHNL_DCHNO(mac->dma_rxch) |
-			       PAS_MAC_IPC_CHNL_BCH(mac->dma_rxch));
+	write_mac_reg(mac, PAS_MAC_IPC_CHNL,
+			   PAS_MAC_IPC_CHNL_DCHNO(mac->dma_rxch) |
+			   PAS_MAC_IPC_CHNL_BCH(mac->dma_rxch));
 
 	/* enable rx if */
-	pci_write_config_dword(mac->dma_pdev,
-			       PAS_DMA_RXINT_RCMDSTA(mac->dma_if),
-			       PAS_DMA_RXINT_RCMDSTA_EN);
+	write_dma_reg(mac, PAS_DMA_RXINT_RCMDSTA(mac->dma_if),
+			   PAS_DMA_RXINT_RCMDSTA_EN);
 
 	/* enable rx channel */
-	pci_write_config_dword(mac->dma_pdev,
-			       PAS_DMA_RXCHAN_CCMDSTA(mac->dma_rxch),
-			       PAS_DMA_RXCHAN_CCMDSTA_EN |
-			       PAS_DMA_RXCHAN_CCMDSTA_DU);
+	write_dma_reg(mac, PAS_DMA_RXCHAN_CCMDSTA(mac->dma_rxch),
+			   PAS_DMA_RXCHAN_CCMDSTA_EN |
+			   PAS_DMA_RXCHAN_CCMDSTA_DU);
 
 	/* enable tx channel */
-	pci_write_config_dword(mac->dma_pdev,
-			       PAS_DMA_TXCHAN_TCMDSTA(mac->dma_txch),
-			       PAS_DMA_TXCHAN_TCMDSTA_EN);
+	write_dma_reg(mac, PAS_DMA_TXCHAN_TCMDSTA(mac->dma_txch),
+			   PAS_DMA_TXCHAN_TCMDSTA_EN);
 
 	pasemi_mac_replenish_rx_ring(dev);
 
@@ -875,20 +898,12 @@ static int pasemi_mac_close(struct net_d
 	pasemi_mac_clean_rx(mac, RX_RING_SIZE);
 
 	/* Disable interface */
-	pci_write_config_dword(mac->dma_pdev,
-			       PAS_DMA_TXCHAN_TCMDSTA(mac->dma_txch),
-			       PAS_DMA_TXCHAN_TCMDSTA_ST);
-	pci_write_config_dword(mac->dma_pdev,
-		      PAS_DMA_RXINT_RCMDSTA(mac->dma_if),
-		      PAS_DMA_RXINT_RCMDSTA_ST);
-	pci_write_config_dword(mac->dma_pdev,
-		      PAS_DMA_RXCHAN_CCMDSTA(mac->dma_rxch),
-		      PAS_DMA_RXCHAN_CCMDSTA_ST);
+	write_dma_reg(mac, PAS_DMA_TXCHAN_TCMDSTA(mac->dma_txch), PAS_DMA_TXCHAN_TCMDSTA_ST);
+	write_dma_reg(mac, PAS_DMA_RXINT_RCMDSTA(mac->dma_if), PAS_DMA_RXINT_RCMDSTA_ST);
+	write_dma_reg(mac, PAS_DMA_RXCHAN_CCMDSTA(mac->dma_rxch), PAS_DMA_RXCHAN_CCMDSTA_ST);
 
 	for (retries = 0; retries < MAX_RETRIES; retries++) {
-		pci_read_config_dword(mac->dma_pdev,
-				      PAS_DMA_TXCHAN_TCMDSTA(mac->dma_txch),
-				      &stat);
+		stat = read_dma_reg(mac, PAS_DMA_TXCHAN_TCMDSTA(mac->dma_txch));
 		if (!(stat & PAS_DMA_TXCHAN_TCMDSTA_ACT))
 			break;
 		cond_resched();
@@ -898,9 +913,7 @@ static int pasemi_mac_close(struct net_d
 		dev_err(&mac->dma_pdev->dev, "Failed to stop tx channel\n");
 
 	for (retries = 0; retries < MAX_RETRIES; retries++) {
-		pci_read_config_dword(mac->dma_pdev,
-				      PAS_DMA_RXCHAN_CCMDSTA(mac->dma_rxch),
-				      &stat);
+		stat = read_dma_reg(mac, PAS_DMA_RXCHAN_CCMDSTA(mac->dma_rxch));
 		if (!(stat & PAS_DMA_RXCHAN_CCMDSTA_ACT))
 			break;
 		cond_resched();
@@ -910,9 +923,7 @@ static int pasemi_mac_close(struct net_d
 		dev_err(&mac->dma_pdev->dev, "Failed to stop rx channel\n");
 
 	for (retries = 0; retries < MAX_RETRIES; retries++) {
-		pci_read_config_dword(mac->dma_pdev,
-				      PAS_DMA_RXINT_RCMDSTA(mac->dma_if),
-				      &stat);
+		stat = read_dma_reg(mac, PAS_DMA_RXINT_RCMDSTA(mac->dma_if));
 		if (!(stat & PAS_DMA_RXINT_RCMDSTA_ACT))
 			break;
 		cond_resched();
@@ -925,12 +936,9 @@ static int pasemi_mac_close(struct net_d
 	 * stopping, since you can't disable when active.
 	 */
 
-	pci_write_config_dword(mac->dma_pdev,
-			       PAS_DMA_TXCHAN_TCMDSTA(mac->dma_txch), 0);
-	pci_write_config_dword(mac->dma_pdev,
-			       PAS_DMA_RXCHAN_CCMDSTA(mac->dma_rxch), 0);
-	pci_write_config_dword(mac->dma_pdev,
-			       PAS_DMA_RXINT_RCMDSTA(mac->dma_if), 0);
+	write_dma_reg(mac, PAS_DMA_TXCHAN_TCMDSTA(mac->dma_txch), 0);
+	write_dma_reg(mac, PAS_DMA_RXCHAN_CCMDSTA(mac->dma_rxch), 0);
+	write_dma_reg(mac, PAS_DMA_RXINT_RCMDSTA(mac->dma_if), 0);
 
 	free_irq(mac->tx_irq, dev);
 	free_irq(mac->rx_irq, dev);
@@ -1011,8 +1019,7 @@ static int pasemi_mac_start_tx(struct sk
 
 	spin_unlock_irqrestore(&txring->lock, flags);
 
-	pci_write_config_dword(mac->dma_pdev,
-			       PAS_DMA_TXCHAN_INCR(mac->dma_txch), 1);
+	write_dma_reg(mac, PAS_DMA_TXCHAN_INCR(mac->dma_txch), 1);
 
 	return NETDEV_TX_OK;
 
@@ -1035,7 +1042,7 @@ static void pasemi_mac_set_rx_mode(struc
 	struct pasemi_mac *mac = netdev_priv(dev);
 	unsigned int flags;
 
-	pci_read_config_dword(mac->pdev, PAS_MAC_CFG_PCFG, &flags);
+	flags = read_mac_reg(mac, PAS_MAC_CFG_PCFG);
 
 	/* Set promiscuous */
 	if (dev->flags & IFF_PROMISC)
@@ -1043,7 +1050,7 @@ static void pasemi_mac_set_rx_mode(struc
 	else
 		flags &= ~PAS_MAC_CFG_PCFG_PR;
 
-	pci_write_config_dword(mac->pdev, PAS_MAC_CFG_PCFG, flags);
+	write_mac_reg(mac, PAS_MAC_CFG_PCFG, flags);
 }
 
 

--

Re: [PATCH 3/9] pasemi_mac: Abstract out register access

[Jeff Garzik]

Olof Johansson wrote:
> Abstract out the PCI config read/write accesses into reg read/write ones, still
> calling the pci accessors on the back end.
> 
> Signed-off-by: Olof Johansson <olof@lixom.net>
> 
> 
> Index: netdev-2.6/drivers/net/pasemi_mac.c
> ===================================================================
> --- netdev-2.6.orig/drivers/net/pasemi_mac.c
> +++ netdev-2.6/drivers/net/pasemi_mac.c
> @@ -81,6 +81,48 @@ MODULE_PARM_DESC(debug, "PA Semi MAC bit
>  
>  static struct pasdma_status *dma_status;
>  
> +static unsigned int read_iob_reg(struct pasemi_mac *mac, unsigned int reg)
> +{
> +	unsigned int val;
> +
> +	pci_read_config_dword(mac->iob_pdev, reg, &val);
> +	return val;
> +}
> +
> +static void write_iob_reg(struct pasemi_mac *mac, unsigned int reg,
> +			  unsigned int val)
> +{
> +	pci_write_config_dword(mac->iob_pdev, reg, val);
> +}
> +
> +static unsigned int read_mac_reg(struct pasemi_mac *mac, unsigned int reg)
> +{
> +	unsigned int val;
> +
> +	pci_read_config_dword(mac->pdev, reg, &val);
> +	return val;
> +}
> +
> +static void write_mac_reg(struct pasemi_mac *mac, unsigned int reg,
> +			  unsigned int val)
> +{
> +	pci_write_config_dword(mac->pdev, reg, val);
> +}
> +
> +static unsigned int read_dma_reg(struct pasemi_mac *mac, unsigned int reg)
> +{
> +	unsigned int val;
> +
> +	pci_read_config_dword(mac->dma_pdev, reg, &val);
> +	return val;
> +}
> +
> +static void write_dma_reg(struct pasemi_mac *mac, unsigned int reg,
> +			  unsigned int val)
> +{
> +	pci_write_config_dword(mac->dma_pdev, reg, val);
> +}

The general concept is fine, but you should use 'u32' and similar 
size-based types for your input and output values.

[PATCH 4/9] pasemi_mac: Use MMIO instead of pci config accessors

[Olof Johansson]

[-- Attachment #1: pasemi_mac-kill-pci-config-access --]
[-- Type: text/plain, Size: 6918 bytes --]

Move away from using the pci config access functions for simple register
access.  Our device has all of the registers in the config space (hey,
from the hardware point of view it looks reasonable :-), so we need to
somehow get to it. Newer firmwares have it in the device tree such that
we can just get it and ioremap it there (in case it ever moves in future
products). For now, provide a hardcoded fallback for older firmwares.


Signed-off-by: Olof Johansson <olof@lixom.net>


Index: netdev-2.6/drivers/net/pasemi_mac.c
===================================================================
--- netdev-2.6.orig/drivers/net/pasemi_mac.c
+++ netdev-2.6/drivers/net/pasemi_mac.c
@@ -81,46 +81,47 @@ MODULE_PARM_DESC(debug, "PA Semi MAC bit
 
 static struct pasdma_status *dma_status;
 
-static unsigned int read_iob_reg(struct pasemi_mac *mac, unsigned int reg)
+static inline unsigned int read_iob_reg(struct pasemi_mac *mac, unsigned int reg)
 {
 	unsigned int val;
 
-	pci_read_config_dword(mac->iob_pdev, reg, &val);
+	val = in_le32(mac->iob_regs+reg);
+
 	return val;
 }
 
-static void write_iob_reg(struct pasemi_mac *mac, unsigned int reg,
+static inline void write_iob_reg(struct pasemi_mac *mac, unsigned int reg,
 			  unsigned int val)
 {
-	pci_write_config_dword(mac->iob_pdev, reg, val);
+	out_le32(mac->iob_regs+reg, val);
 }
 
-static unsigned int read_mac_reg(struct pasemi_mac *mac, unsigned int reg)
+static inline unsigned int read_mac_reg(struct pasemi_mac *mac, unsigned int reg)
 {
 	unsigned int val;
 
-	pci_read_config_dword(mac->pdev, reg, &val);
+	val = in_le32(mac->regs+reg);
 	return val;
 }
 
-static void write_mac_reg(struct pasemi_mac *mac, unsigned int reg,
+static inline void write_mac_reg(struct pasemi_mac *mac, unsigned int reg,
 			  unsigned int val)
 {
-	pci_write_config_dword(mac->pdev, reg, val);
+	out_le32(mac->regs+reg, val);
 }
 
-static unsigned int read_dma_reg(struct pasemi_mac *mac, unsigned int reg)
+static inline unsigned int read_dma_reg(struct pasemi_mac *mac, unsigned int reg)
 {
 	unsigned int val;
 
-	pci_read_config_dword(mac->dma_pdev, reg, &val);
+	val = in_le32(mac->dma_regs+reg);
 	return val;
 }
 
-static void write_dma_reg(struct pasemi_mac *mac, unsigned int reg,
+static inline void write_dma_reg(struct pasemi_mac *mac, unsigned int reg,
 			  unsigned int val)
 {
-	pci_write_config_dword(mac->dma_pdev, reg, val);
+	out_le32(mac->dma_regs+reg, val);
 }
 
 static int pasemi_get_mac_addr(struct pasemi_mac *mac)
@@ -585,7 +586,6 @@ static int pasemi_mac_clean_tx(struct pa
 	}
 	mac->tx->next_to_clean += count;
 	spin_unlock_irqrestore(&mac->tx->lock, flags);
-
 	netif_wake_queue(mac->netdev);
 
 	return count;
@@ -1077,6 +1077,73 @@ static int pasemi_mac_poll(struct net_de
 	}
 }
 
+static inline void __iomem * __devinit map_onedev(struct pci_dev *p, int index)
+{
+	struct device_node *dn;
+	void __iomem *ret;
+
+	dn = pci_device_to_OF_node(p);
+	if (!dn)
+		goto fallback;
+
+	ret = of_iomap(dn, index);
+	if (!ret)
+		goto fallback;
+
+	return ret;
+fallback:
+	/* This is hardcoded and ugly, but we have some firmware versions
+	 * who don't provide the register space in the device tree. Luckily
+	 * they are at well-known locations so we can just do the math here.
+	 */
+	return ioremap(0xe0000000 + (p->devfn << 12), 0x1000);
+}
+
+static int __devinit pasemi_mac_map_regs(struct pasemi_mac *mac)
+{
+	struct resource res;
+	struct device_node *dn;
+	int err;
+
+	mac->dma_pdev = pci_get_device(PCI_VENDOR_ID_PASEMI, 0xa007, NULL);
+	if (!mac->dma_pdev) {
+		dev_err(&mac->pdev->dev, "Can't find DMA Controller\n");
+		return -ENODEV;
+	}
+
+	mac->iob_pdev = pci_get_device(PCI_VENDOR_ID_PASEMI, 0xa001, NULL);
+	if (!mac->iob_pdev) {
+		dev_err(&mac->pdev->dev, "Can't find I/O Bridge\n");
+		return -ENODEV;
+	}
+
+	mac->regs = map_onedev(mac->pdev, 0);
+	mac->dma_regs = map_onedev(mac->dma_pdev, 0);
+	mac->iob_regs = map_onedev(mac->iob_pdev, 0);
+
+	if (!mac->regs || !mac->dma_regs || !mac->iob_regs) {
+		dev_err(&mac->pdev->dev, "Can't map registers\n");
+		return -ENODEV;
+	}
+
+	/* The dma status structure is located in the I/O bridge, and
+	 * is cache coherent.
+	 */
+	if (!dma_status) {
+		dn = pci_device_to_OF_node(mac->iob_pdev);
+		if (dn)
+			err = of_address_to_resource(dn, 1, &res);
+		if (!dn || err) {
+			/* Fallback for old firmware */
+			res.start = 0xfd800000;
+			res.end = res.start + 0x1000;
+		}
+		dma_status = __ioremap(res.start, res.end-res.start, 0);
+	}
+
+	return 0;
+}
+
 static int __devinit
 pasemi_mac_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
 {
@@ -1105,21 +1172,6 @@ pasemi_mac_probe(struct pci_dev *pdev, c
 
 	mac->pdev = pdev;
 	mac->netdev = dev;
-	mac->dma_pdev = pci_get_device(PCI_VENDOR_ID_PASEMI, 0xa007, NULL);
-
-	if (!mac->dma_pdev) {
-		dev_err(&pdev->dev, "Can't find DMA Controller\n");
-		err = -ENODEV;
-		goto out_free_netdev;
-	}
-
-	mac->iob_pdev = pci_get_device(PCI_VENDOR_ID_PASEMI, 0xa001, NULL);
-
-	if (!mac->iob_pdev) {
-		dev_err(&pdev->dev, "Can't find I/O Bridge\n");
-		err = -ENODEV;
-		goto out_put_dma_pdev;
-	}
 
 	/* These should come out of the device tree eventually */
 	mac->dma_txch = index;
@@ -1162,12 +1214,9 @@ pasemi_mac_probe(struct pci_dev *pdev, c
 	dev->poll = pasemi_mac_poll;
 	dev->features = NETIF_F_HW_CSUM;
 
-	/* The dma status structure is located in the I/O bridge, and
-	 * is cache coherent.
-	 */
-	if (!dma_status)
-		/* XXXOJN This should come from the device tree */
-		dma_status = __ioremap(0xfd800000, 0x1000, 0);
+	err = pasemi_mac_map_regs(mac);
+	if (err)
+		goto out;
 
 	mac->rx_status = &dma_status->rx_sta[mac->dma_rxch];
 	mac->tx_status = &dma_status->tx_sta[mac->dma_txch];
@@ -1194,10 +1243,17 @@ pasemi_mac_probe(struct pci_dev *pdev, c
 	return err;
 
 out:
-	pci_dev_put(mac->iob_pdev);
-out_put_dma_pdev:
-	pci_dev_put(mac->dma_pdev);
-out_free_netdev:
+	if (mac->iob_pdev)
+		pci_dev_put(mac->iob_pdev);
+	if (mac->dma_pdev)
+		pci_dev_put(mac->dma_pdev);
+	if (mac->dma_regs)
+		iounmap(mac->dma_regs);
+	if (mac->iob_regs)
+		iounmap(mac->iob_regs);
+	if (mac->regs)
+		iounmap(mac->regs);
+
 	free_netdev(dev);
 out_disable_device:
 	pci_disable_device(pdev);
@@ -1221,6 +1277,10 @@ static void __devexit pasemi_mac_remove(
 	pci_dev_put(mac->dma_pdev);
 	pci_dev_put(mac->iob_pdev);
 
+	iounmap(mac->regs);
+	iounmap(mac->dma_regs);
+	iounmap(mac->iob_regs);
+
 	pci_set_drvdata(pdev, NULL);
 	free_netdev(netdev);
 }
Index: netdev-2.6/drivers/net/pasemi_mac.h
===================================================================
--- netdev-2.6.orig/drivers/net/pasemi_mac.h
+++ netdev-2.6/drivers/net/pasemi_mac.h
@@ -52,6 +52,9 @@ struct pasemi_mac_rxring {
 
 struct pasemi_mac {
 	struct net_device *netdev;
+	void __iomem *regs;
+	void __iomem *dma_regs;
+	void __iomem *iob_regs;
 	struct pci_dev *pdev;
 	struct pci_dev *dma_pdev;
 	struct pci_dev *iob_pdev;

--

Re: [PATCH 4/9] pasemi_mac: Use MMIO instead of pci config accessors

[Jeff Garzik]

Olof Johansson wrote:
> Move away from using the pci config access functions for simple register
> access.  Our device has all of the registers in the config space (hey,
> from the hardware point of view it looks reasonable :-), so we need to
> somehow get to it. Newer firmwares have it in the device tree such that
> we can just get it and ioremap it there (in case it ever moves in future
> products). For now, provide a hardcoded fallback for older firmwares.
> 
> 
> Signed-off-by: Olof Johansson <olof@lixom.net>

patches 4-9 seen sane.  dropped due to previous patches being dropped

[PATCH 5/9] pasemi_mac: Enable L2 caching of packet headers

[Olof Johansson]

[-- Attachment #1: pasemi_mac-cache-headers --]
[-- Type: text/plain, Size: 1923 bytes --]

Enable settings to target L2 for the first few cachelines of the packet,
since we'll access them to get to the various headers.


Signed-off-by: Olof Johansson <olof@lixom.net>


Index: netdev-2.6/drivers/net/pasemi_mac.c
===================================================================
--- netdev-2.6.orig/drivers/net/pasemi_mac.c
+++ netdev-2.6/drivers/net/pasemi_mac.c
@@ -216,7 +216,7 @@ static int pasemi_mac_setup_rx_resources
 			   PAS_DMA_RXCHAN_BASEU_SIZ(RX_RING_SIZE >> 2));
 
 	write_dma_reg(mac, PAS_DMA_RXCHAN_CFG(chan_id),
-			   PAS_DMA_RXCHAN_CFG_HBU(1));
+			   PAS_DMA_RXCHAN_CFG_HBU(2));
 
 	write_dma_reg(mac, PAS_DMA_RXINT_BASEL(mac->dma_if),
 			   PAS_DMA_RXINT_BASEL_BRBL(__pa(ring->buffers)));
@@ -225,6 +225,9 @@ static int pasemi_mac_setup_rx_resources
 			   PAS_DMA_RXINT_BASEU_BRBH(__pa(ring->buffers) >> 32) |
 			   PAS_DMA_RXINT_BASEU_SIZ(RX_RING_SIZE >> 3));
 
+	write_dma_reg(mac, PAS_DMA_RXINT_CFG(mac->dma_if),
+			   PAS_DMA_RXINT_CFG_DHL(2));
+
 	ring->next_to_fill = 0;
 	ring->next_to_clean = 0;
 
Index: netdev-2.6/drivers/net/pasemi_mac.h
===================================================================
--- netdev-2.6.orig/drivers/net/pasemi_mac.h
+++ netdev-2.6/drivers/net/pasemi_mac.h
@@ -218,6 +218,14 @@ enum {
 #define    PAS_DMA_RXINT_RCMDSTA_ACT	0x00010000
 #define    PAS_DMA_RXINT_RCMDSTA_DROPS_M	0xfffe0000
 #define    PAS_DMA_RXINT_RCMDSTA_DROPS_S	17
+#define PAS_DMA_RXINT_CFG(i)		(0x204+(i)*_PAS_DMA_RXINT_STRIDE)
+#define    PAS_DMA_RXINT_CFG_DHL_M	0x07000000
+#define    PAS_DMA_RXINT_CFG_DHL_S	24
+#define    PAS_DMA_RXINT_CFG_DHL(x)	(((x) << PAS_DMA_RXINT_CFG_DHL_S) & \
+					 PAS_DMA_RXINT_CFG_DHL_M)
+#define    PAS_DMA_RXINT_CFG_WIF	0x00000002
+#define    PAS_DMA_RXINT_CFG_WIL	0x00000001
+
 #define PAS_DMA_RXINT_INCR(i)		(0x210+(i)*_PAS_DMA_RXINT_STRIDE)
 #define    PAS_DMA_RXINT_INCR_INCR_M	0x0000ffff
 #define    PAS_DMA_RXINT_INCR_INCR_S	0

--

[PATCH 6/9] pasemi_mac: Simplify memcpy for short receives

[Olof Johansson]

[-- Attachment #1: pasemi_mac-bugfix --]
[-- Type: text/plain, Size: 732 bytes --]

No need to copy over the skipped align bytes (besides, NET_IP_ALIGN is
0 on ppc64).

Signed-off-by: Olof Johansson <olof@lixom.net>


Index: netdev-2.6/drivers/net/pasemi_mac.c
===================================================================
--- netdev-2.6.orig/drivers/net/pasemi_mac.c
+++ netdev-2.6/drivers/net/pasemi_mac.c
@@ -516,9 +516,7 @@ static int pasemi_mac_clean_rx(struct pa
 			    netdev_alloc_skb(mac->netdev, len + NET_IP_ALIGN);
 			if (new_skb) {
 				skb_reserve(new_skb, NET_IP_ALIGN);
-				memcpy(new_skb->data - NET_IP_ALIGN,
-					skb->data - NET_IP_ALIGN,
-					len + NET_IP_ALIGN);
+				memcpy(new_skb->data, skb->data, len);
 				/* save the skb in buffer_info as good */
 				skb = new_skb;
 			}

--

[PATCH 7/9] pasemi_mac: Minor performance tweaks

[Olof Johansson]

[-- Attachment #1: pasemi_mac-rxperf --]
[-- Type: text/plain, Size: 2471 bytes --]

Various minor performance tweaks, do some explicit prefetching of packet
data, etc.

Signed-off-by: Olof Johansson <olof@lixom.net>

Index: netdev-2.6/drivers/net/pasemi_mac.c
===================================================================
--- netdev-2.6.orig/drivers/net/pasemi_mac.c
+++ netdev-2.6/drivers/net/pasemi_mac.c
@@ -481,6 +481,7 @@ static int pasemi_mac_clean_rx(struct pa
 		rmb();
 
 		dp = &RX_DESC(mac, n);
+		prefetchw(dp);
 		macrx = dp->macrx;
 
 		if (!(macrx & XCT_MACRX_O))
@@ -502,8 +503,10 @@ static int pasemi_mac_clean_rx(struct pa
 			if (info->dma == dma)
 				break;
 		}
+		prefetchw(info);
 
 		skb = info->skb;
+		prefetchw(skb);
 		info->dma = 0;
 
 		pci_unmap_single(mac->dma_pdev, dma, skb->len,
@@ -526,9 +529,7 @@ static int pasemi_mac_clean_rx(struct pa
 
 		skb_put(skb, len);
 
-		skb->protocol = eth_type_trans(skb, mac->netdev);
-
-		if ((macrx & XCT_MACRX_HTY_M) == XCT_MACRX_HTY_IPV4_OK) {
+		if (likely((macrx & XCT_MACRX_HTY_M) == XCT_MACRX_HTY_IPV4_OK)) {
 			skb->ip_summed = CHECKSUM_COMPLETE;
 			skb->csum = (macrx & XCT_MACRX_CSUM_M) >>
 					   XCT_MACRX_CSUM_S;
@@ -538,6 +539,7 @@ static int pasemi_mac_clean_rx(struct pa
 		mac->stats.rx_bytes += len;
 		mac->stats.rx_packets++;
 
+		skb->protocol = eth_type_trans(skb, mac->netdev);
 		netif_receive_skb(skb);
 
 		dp->ptr = 0;
@@ -569,7 +571,7 @@ static int pasemi_mac_clean_tx(struct pa
 
 	for (i = start; i < mac->tx->next_to_use; i++) {
 		dp = &TX_DESC(mac, i);
-		if (!dp || (dp->mactx & XCT_MACTX_O))
+		if (unlikely(dp->mactx & XCT_MACTX_O))
 			break;
 
 		count++;
@@ -957,7 +959,7 @@ static int pasemi_mac_start_tx(struct sk
 	struct pasemi_mac_txring *txring;
 	struct pasemi_mac_buffer *info;
 	struct pas_dma_xct_descr *dp;
-	u64 dflags;
+	u64 dflags, mactx, ptr;
 	dma_addr_t map;
 	int flags;
 
@@ -985,6 +987,9 @@ static int pasemi_mac_start_tx(struct sk
 	if (dma_mapping_error(map))
 		return NETDEV_TX_BUSY;
 
+	mactx = dflags | XCT_MACTX_LLEN(skb->len);
+	ptr   = XCT_PTR_LEN(skb->len) | XCT_PTR_ADDR(map);
+
 	txring = mac->tx;
 
 	spin_lock_irqsave(&txring->lock, flags);
@@ -1005,12 +1010,11 @@ static int pasemi_mac_start_tx(struct sk
 		}
 	}
 
-
 	dp = &TX_DESC(mac, txring->next_to_use);
 	info = &TX_DESC_INFO(mac, txring->next_to_use);
 
-	dp->mactx = dflags | XCT_MACTX_LLEN(skb->len);
-	dp->ptr   = XCT_PTR_LEN(skb->len) | XCT_PTR_ADDR(map);
+	dp->mactx = mactx;
+	dp->ptr = ptr;
 	info->dma = map;
 	info->skb = skb;
 

--

[PATCH 8/9] pasemi_mac: Reduce locking when cleaning TX ring

[Olof Johansson]

[-- Attachment #1: pasemi_mac-tx-lock-reduction --]
[-- Type: text/plain, Size: 1780 bytes --]

Postpone pci unmap and skb free of the transmitted buffers to outside
of the tx ring lock, batching them up 32 at a time.


Signed-off-by: Olof Johansson <olof@lixom.net>

Index: netdev-2.6/drivers/net/pasemi_mac.c
===================================================================
--- netdev-2.6.orig/drivers/net/pasemi_mac.c
+++ netdev-2.6/drivers/net/pasemi_mac.c
@@ -562,37 +562,56 @@ static int pasemi_mac_clean_tx(struct pa
 	int i;
 	struct pasemi_mac_buffer *info;
 	struct pas_dma_xct_descr *dp;
-	int start, count;
+	unsigned int start, count, limit;
+	unsigned int total_count;
 	int flags;
+	struct sk_buff *skbs[32];
+	dma_addr_t dmas[32];
 
+	total_count = 0;
+restart:
 	spin_lock_irqsave(&mac->tx->lock, flags);
 
 	start = mac->tx->next_to_clean;
+	limit = min(mac->tx->next_to_use, start+32);
+
 	count = 0;
 
-	for (i = start; i < mac->tx->next_to_use; i++) {
+	for (i = start; i < limit; i++) {
 		dp = &TX_DESC(mac, i);
+
 		if (unlikely(dp->mactx & XCT_MACTX_O))
+			/* Not yet transmitted */
 			break;
 
-		count++;
-
 		info = &TX_DESC_INFO(mac, i);
-
-		pci_unmap_single(mac->dma_pdev, info->dma,
-				 info->skb->len, PCI_DMA_TODEVICE);
-		dev_kfree_skb_irq(info->skb);
+		skbs[count] = info->skb;
+		dmas[count] = info->dma;
 
 		info->skb = NULL;
 		info->dma = 0;
 		dp->mactx = 0;
 		dp->ptr = 0;
+
+		count++;
 	}
 	mac->tx->next_to_clean += count;
 	spin_unlock_irqrestore(&mac->tx->lock, flags);
 	netif_wake_queue(mac->netdev);
 
-	return count;
+	for (i = 0; i < count; i++) {
+		pci_unmap_single(mac->dma_pdev, dmas[i],
+				 skbs[i]->len, PCI_DMA_TODEVICE);
+		dev_kfree_skb_irq(skbs[i]);
+	}
+
+	total_count += count;
+
+	/* If the batch was full, try to clean more */
+	if (count == 32)
+		goto restart;
+
+	return total_count;
 }
 
 

--

[PATCH 9/9] pasemi_mac: Enable LLTX

[Olof Johansson]

[-- Attachment #1: pasemi_mac-enable-lltx --]
[-- Type: text/plain, Size: 642 bytes --]

Enable LLTX on pasemi_mac: we're already doing sufficient locking
in the driver to enable it.


Signed-off-by: Olof Johansson <olof@lixom.net>

Index: netdev-2.6/drivers/net/pasemi_mac.c
===================================================================
--- netdev-2.6.orig/drivers/net/pasemi_mac.c
+++ netdev-2.6/drivers/net/pasemi_mac.c
@@ -1239,7 +1239,7 @@ pasemi_mac_probe(struct pci_dev *pdev, c
 	dev->set_multicast_list = pasemi_mac_set_rx_mode;
 	dev->weight = 64;
 	dev->poll = pasemi_mac_poll;
-	dev->features = NETIF_F_HW_CSUM;
+	dev->features = NETIF_F_HW_CSUM | NETIF_F_LLTX;
 
 	err = pasemi_mac_map_regs(mac);
 	if (err)

--