[PATCH] A new 10GB Ethernet Driver by Chelsio Communications

A Linux driver for the Chelsio 10Gb Ethernet Network Controller by Chelsio
(http://www.chelsio.com).  This driver supports the Chelsio N210 NIC and is
backward compatible with the Chelsio N110 model 10Gb NICs.  It supports
AMD64, EM64T and x86 systems.

Signed-off-by: Tina Yang <tinay@chelsio.com>
Signed-off-by: Scott Bardone <sbardone@chelsio.com>
Signed-off-by: Christoph Lameter <christoph@lameter.com>

Adrian said:

- my3126.c is unused (because t1_my3126_ops isn't used anywhere)
- what are the EXTRA_CFLAGS in drivers/net/chelsio/Makefile for?
- $(cxgb-y) in drivers/net/chelsio/Makefile seems to be unneeded
- completely unused global functions:
  - espi.c: t1_espi_get_intr_counts
  - sge.c: t1_sge_get_intr_counts
- the following functions can be made static:
  - sge.c: t1_espi_workaround
  - sge.c: t1_sge_tx
  - subr.c: __t1_tpi_read
  - subr.c: __t1_tpi_write
  - subr.c: t1_wait_op_done

shemminger said:

The performance recommendations in cxgb.txt are common to all fast devices,
and should be in one file rather than just for this device. I would rather
see ip-sysctl.txt updated or a new file on tuning recommendations started.
Some of them have consequences that aren't documented well.
For example, turning off TCP timestamps risks data corruption from sequence wrap.

A new driver shouldn't need so may #ifdef's unless you want to putit on older
vendor versions of 2.4

Some accessor and wrapper functions like:
        t1_pci_read_config_4
        adapter_name
        t1_malloc
are just annoying noise.

Why have useless dead code like:

/* Interrupt handler */
+static int pm3393_interrupt_handler(struct cmac *cmac)
+{
+       u32 master_intr_status;
+/*
+    1. Read master interrupt register.
+    2. Read BLOCK's interrupt status registers.
+    3. Handle BLOCK interrupts.
+*/

Jeff said:

step 1:  kill all the OS wrappers.

 And do you really need hooks for multiple MACs, when only one MAC is
 really supported?  Typically these hooks are at a higher level anyway --
 struct net_device.

From: Christoph Lameter <christoph@lameter

Driver modified as suggested by Pekka Enberg, Stephen Hemminger and Andrian
Bunk.  Reduces the size of the driver to ~260k.

- clean up tabs
- removed my3126.c
- removed 85% of suni1x10gexp_regs.h
- removed 80% of regs.h
- removed various calls, renamed variables/functions.
- removed system specific and other wrappers (usleep, msleep)
- removed dead code
- dropped redundant casts in osdep.h
- dropped redundant check of kfree
- dropped weird code (MODVERSIONS stuff)
- reduced number of #ifdefs
- use kcalloc now instead of kmalloc
- Add information about known issues with the driver
- Add information about authors

Signed-off-by: Scott Bardone <sbardone@chelsio.com>
Signed-off-by: Christoph Lameter <christoph@lameter.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>

diff -puN /dev/null Documentation/networking/cxgb.txt

1 files changed, 1451 insertions, 0 deletions

diff --git a/drivers/net/chelsio/sge.c b/drivers/net/chelsio/sge.c
new file mode 100644
index 000000000000..bcf8b1e939b0
--- /dev/null
+++ b/drivers/net/chelsio/sge.c
@@ -0,0 +1,1451 @@
+/*****************************************************************************
+ *                                                                           *
+ * File: sge.c                                                               *
+ * $Revision: 1.13 $                                                         *
+ * $Date: 2005/03/23 07:41:27 $                                              *
+ * Description:                                                              *
+ *  DMA engine.                                                              *
+ *  part of the Chelsio 10Gb Ethernet Driver.                                *
+ *                                                                           *
+ * This program is free software; you can redistribute it and/or modify      *
+ * it under the terms of the GNU General Public License, version 2, as       *
+ * published by the Free Software Foundation.                                *
+ *                                                                           *
+ * You should have received a copy of the GNU General Public License along   *
+ * with this program; if not, write to the Free Software Foundation, Inc.,   *
+ * 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.                 *
+ *                                                                           *
+ * THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR IMPLIED    *
+ * WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED WARRANTIES OF      *
+ * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.                     *
+ *                                                                           *
+ * http://www.chelsio.com                                                    *
+ *                                                                           *
+ * Copyright (c) 2003 - 2005 Chelsio Communications, Inc.                    *
+ * All rights reserved.                                                      *
+ *                                                                           *
+ * Maintainers: maintainers@chelsio.com                                      *
+ *                                                                           *
+ * Authors: Dimitrios Michailidis   <dm@chelsio.com>                         *
+ *          Tina Yang               <tainay@chelsio.com>                     *
+ *          Felix Marti             <felix@chelsio.com>                      *
+ *          Scott Bardone           <sbardone@chelsio.com>                   *
+ *          Kurt Ottaway            <kottaway@chelsio.com>                   *
+ *          Frank DiMambro          <frank@chelsio.com>                      *
+ *                                                                           *
+ * History:                                                                  *
+ *                                                                           *
+ ****************************************************************************/
+
+#include "common.h"
+
+#include <linux/config.h>
+#include <linux/types.h>
+#include <linux/errno.h>
+#include <linux/pci.h>
+#include <linux/netdevice.h>
+#include <linux/etherdevice.h>
+#include <linux/if_vlan.h>
+#include <linux/skbuff.h>
+#include <linux/init.h>
+#include <linux/mm.h>
+#include <linux/ip.h>
+#include <linux/in.h>
+#include <linux/if_arp.h>
+
+#include "cpl5_cmd.h"
+#include "sge.h"
+#include "regs.h"
+#include "espi.h"
+
+#include <linux/tcp.h>
+
+#define SGE_CMDQ_N		2
+#define SGE_FREELQ_N		2
+#define SGE_CMDQ0_E_N		512
+#define SGE_CMDQ1_E_N		128
+#define SGE_FREEL_SIZE		4096
+#define SGE_JUMBO_FREEL_SIZE	512
+#define SGE_FREEL_REFILL_THRESH	16
+#define SGE_RESPQ_E_N		1024
+#define SGE_INTR_BUCKETSIZE	100
+#define SGE_INTR_LATBUCKETS	5
+#define SGE_INTR_MAXBUCKETS	11
+#define SGE_INTRTIMER0		1
+#define SGE_INTRTIMER1		50
+#define SGE_INTRTIMER_NRES	10000
+#define SGE_RX_COPY_THRESHOLD	256
+#define SGE_RX_SM_BUF_SIZE	1536
+
+#define SGE_RESPQ_REPLENISH_THRES ((3 * SGE_RESPQ_E_N) / 4)
+
+#define SGE_RX_OFFSET 2
+#ifndef NET_IP_ALIGN
+# define NET_IP_ALIGN SGE_RX_OFFSET
+#endif
+
+/*
+ * Memory Mapped HW Command, Freelist and Response Queue Descriptors
+ */
+#if defined(__BIG_ENDIAN_BITFIELD)
+struct cmdQ_e {
+	u32 AddrLow;
+	u32 GenerationBit	: 1;
+	u32 BufferLength	: 31;
+	u32 RespQueueSelector	: 4;
+	u32 ResponseTokens	: 12;
+	u32 CmdId		: 8;
+	u32 Reserved		: 3;
+	u32 TokenValid		: 1;
+	u32 Eop			: 1;
+	u32 Sop			: 1;
+	u32 DataValid		: 1;
+	u32 GenerationBit2	: 1;
+	u32 AddrHigh;
+};
+
+struct freelQ_e {
+	u32 AddrLow;
+	u32 GenerationBit	: 1;
+	u32 BufferLength	: 31;
+	u32 Reserved		: 31;
+	u32 GenerationBit2	: 1;
+	u32 AddrHigh;
+};
+
+struct respQ_e {
+	u32 Qsleeping		: 4;
+	u32 Cmdq1CreditReturn	: 5;
+	u32 Cmdq1DmaComplete	: 5;
+	u32 Cmdq0CreditReturn	: 5;
+	u32 Cmdq0DmaComplete	: 5;
+	u32 FreelistQid		: 2;
+	u32 CreditValid		: 1;
+	u32 DataValid		: 1;
+	u32 Offload		: 1;
+	u32 Eop			: 1;
+	u32 Sop			: 1;
+	u32 GenerationBit	: 1;
+	u32 BufferLength;
+};
+
+#elif defined(__LITTLE_ENDIAN_BITFIELD)
+struct cmdQ_e {
+	u32 BufferLength	: 31;
+	u32 GenerationBit	: 1;
+	u32 AddrLow;
+	u32 AddrHigh;
+	u32 GenerationBit2	: 1;
+	u32 DataValid		: 1;
+	u32 Sop			: 1;
+	u32 Eop			: 1;
+	u32 TokenValid		: 1;
+	u32 Reserved		: 3;
+	u32 CmdId		: 8;
+	u32 ResponseTokens	: 12;
+	u32 RespQueueSelector	: 4;
+};
+
+struct freelQ_e {
+	u32 BufferLength	: 31;
+	u32 GenerationBit	: 1;
+	u32 AddrLow;
+	u32 AddrHigh;
+	u32 GenerationBit2	: 1;
+	u32 Reserved		: 31;
+};
+
+struct respQ_e {
+	u32 BufferLength;
+	u32 GenerationBit	: 1;
+	u32 Sop			: 1;
+	u32 Eop			: 1;
+	u32 Offload		: 1;
+	u32 DataValid		: 1;
+	u32 CreditValid		: 1;
+	u32 FreelistQid		: 2;
+	u32 Cmdq0DmaComplete	: 5;
+	u32 Cmdq0CreditReturn	: 5;
+	u32 Cmdq1DmaComplete	: 5;
+	u32 Cmdq1CreditReturn	: 5;
+	u32 Qsleeping		: 4;
+} ;
+#endif
+
+/*
+ * SW Context Command and Freelist Queue Descriptors
+ */
+struct cmdQ_ce {
+	struct sk_buff *skb;
+	DECLARE_PCI_UNMAP_ADDR(dma_addr);
+	DECLARE_PCI_UNMAP_LEN(dma_len);
+	unsigned int single;
+};
+
+struct freelQ_ce {
+	struct sk_buff *skb;
+	DECLARE_PCI_UNMAP_ADDR(dma_addr);
+	DECLARE_PCI_UNMAP_LEN(dma_len);
+};
+
+/*
+ * SW Command, Freelist and Response Queue
+ */
+struct cmdQ {
+	atomic_t	asleep;		/* HW DMA Fetch status */
+	atomic_t	credits;	/* # available descriptors for TX */
+	atomic_t	pio_pidx;	/* Variable updated on Doorbell */
+	u16		entries_n;	/* # descriptors for TX	*/
+	u16		pidx;		/* producer index (SW) */
+	u16		cidx;		/* consumer index (HW) */
+	u8		genbit;		/* current generation (=valid) bit */
+	struct cmdQ_e  *entries;	/* HW command descriptor Q */
+	struct cmdQ_ce *centries;	/* SW command context descriptor Q */
+	spinlock_t	Qlock;		/* Lock to protect cmdQ enqueuing */
+	dma_addr_t	dma_addr;	/* DMA addr HW command descriptor Q */
+};
+
+struct freelQ {
+	unsigned int	credits;	/* # of available RX buffers */
+	unsigned int	entries_n;	/* free list capacity */
+	u16		pidx;		/* producer index (SW) */
+	u16		cidx;		/* consumer index (HW) */
+	u16		rx_buffer_size; /* Buffer size on this free list */
+	u16		dma_offset;     /* DMA offset to align IP headers */
+	u8		genbit;		/* current generation (=valid) bit */
+	struct freelQ_e	*entries;	/* HW freelist descriptor Q */
+	struct freelQ_ce *centries;	/* SW freelist conext descriptor Q */
+	dma_addr_t	dma_addr;	/* DMA addr HW freelist descriptor Q */
+};
+
+struct respQ {
+	u16		credits;	/* # of available respQ descriptors */
+	u16		credits_pend;	/* # of not yet returned descriptors */
+	u16		entries_n;	/* # of response Q descriptors */
+	u16		pidx;		/* producer index (HW) */
+	u16		cidx;		/* consumer index (SW) */
+	u8		genbit;		/* current generation(=valid) bit */
+	struct respQ_e *entries;        /* HW response descriptor Q */
+	dma_addr_t	dma_addr;	/* DMA addr HW response descriptor Q */
+};
+
+/*
+ * Main SGE data structure
+ *
+ * Interrupts are handled by a single CPU and it is likely that on a MP system
+ * the application is migrated to another CPU. In that scenario, we try to
+ * seperate the RX(in irq context) and TX state in order to decrease memory
+ * contention.
+ */
+struct sge {
+	struct adapter *adapter; 	/* adapter backpointer */
+	struct freelQ 	freelQ[SGE_FREELQ_N]; /* freelist Q(s) */
+	struct respQ 	respQ;		/* response Q instatiation */
+	unsigned int	rx_pkt_pad;     /* RX padding for L2 packets */
+	unsigned int	jumbo_fl;       /* jumbo freelist Q index */
+	u32		intrtimer[SGE_INTR_MAXBUCKETS];	/* ! */
+	u32		currIndex;	/* current index into intrtimer[] */
+	u32		intrtimer_nres;	/* no resource interrupt timer value */
+	u32		sge_control;	/* shadow content of sge control reg */
+	struct sge_intr_counts intr_cnt;
+	struct timer_list ptimer;
+	struct sk_buff  *pskb;
+	u32		ptimeout;
+	struct cmdQ cmdQ[SGE_CMDQ_N] ____cacheline_aligned; /* command Q(s)*/
+};
+
+static unsigned int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter,
+			unsigned int qid);
+
+/*
+ * PIO to indicate that memory mapped Q contains valid descriptor(s).
+ */
+static inline void doorbell_pio(struct sge *sge, u32 val)
+{
+	wmb();
+	t1_write_reg_4(sge->adapter, A_SG_DOORBELL, val);
+}
+
+/*
+ * Disables the DMA engine.
+ */
+void t1_sge_stop(struct sge *sge)
+{
+	t1_write_reg_4(sge->adapter, A_SG_CONTROL, 0);
+	t1_read_reg_4(sge->adapter, A_SG_CONTROL);     /* flush write */
+	if (is_T2(sge->adapter))
+		del_timer_sync(&sge->ptimer);
+}
+
+static u8 ch_mac_addr[ETH_ALEN] = {0x0, 0x7, 0x43, 0x0, 0x0, 0x0};
+static void t1_espi_workaround(void *data)
+{
+	struct adapter *adapter = (struct adapter *)data;
+	struct sge *sge = adapter->sge;
+
+	if (netif_running(adapter->port[0].dev) &&
+		atomic_read(&sge->cmdQ[0].asleep)) {
+
+		u32 seop = t1_espi_get_mon(adapter, 0x930, 0);
+
+		if ((seop & 0xfff0fff) == 0xfff && sge->pskb) {
+			struct sk_buff *skb = sge->pskb;
+			if (!skb->cb[0]) {
+				memcpy(skb->data+sizeof(struct cpl_tx_pkt), ch_mac_addr, ETH_ALEN);
+				memcpy(skb->data+skb->len-10, ch_mac_addr, ETH_ALEN);
+
+				skb->cb[0] = 0xff;
+			}
+			t1_sge_tx(skb, adapter,0);
+		}
+	}
+	mod_timer(&adapter->sge->ptimer, jiffies + sge->ptimeout);
+}
+
+/*
+ * Enables the DMA engine.
+ */
+void t1_sge_start(struct sge *sge)
+{
+	t1_write_reg_4(sge->adapter, A_SG_CONTROL, sge->sge_control);
+	t1_read_reg_4(sge->adapter, A_SG_CONTROL);     /* flush write */
+	if (is_T2(sge->adapter)) {
+		init_timer(&sge->ptimer);
+		sge->ptimer.function = (void *)&t1_espi_workaround;
+		sge->ptimer.data = (unsigned long)sge->adapter;
+		sge->ptimer.expires = jiffies + sge->ptimeout;
+		add_timer(&sge->ptimer);
+	}
+}
+
+/*
+ * Creates a t1_sge structure and returns suggested resource parameters.
+ */
+struct sge * __devinit t1_sge_create(struct adapter *adapter,
+				     struct sge_params *p)
+{
+	struct sge *sge = kmalloc(sizeof(*sge), GFP_KERNEL);
+
+	if (!sge)
+		return NULL;
+	memset(sge, 0, sizeof(*sge));
+
+	if (is_T2(adapter))
+		sge->ptimeout = 1;	/* finest allowed */
+
+	sge->adapter = adapter;
+	sge->rx_pkt_pad = t1_is_T1B(adapter) ? 0 : SGE_RX_OFFSET;
+	sge->jumbo_fl = t1_is_T1B(adapter) ? 1 : 0;
+
+	p->cmdQ_size[0] = SGE_CMDQ0_E_N;
+	p->cmdQ_size[1] = SGE_CMDQ1_E_N;
+	p->freelQ_size[!sge->jumbo_fl] = SGE_FREEL_SIZE;
+	p->freelQ_size[sge->jumbo_fl] = SGE_JUMBO_FREEL_SIZE;
+	p->rx_coalesce_usecs = SGE_INTRTIMER1;
+	p->last_rx_coalesce_raw = SGE_INTRTIMER1 *
+	  (board_info(sge->adapter)->clock_core / 1000000);
+	p->default_rx_coalesce_usecs = SGE_INTRTIMER1;
+	p->coalesce_enable = 0;	/* Turn off adaptive algorithm by default */
+	p->sample_interval_usecs = 0;
+	return sge;
+}
+
+/*
+ * Frees all RX buffers on the freelist Q. The caller must make sure that
+ * the SGE is turned off before calling this function.
+ */
+static void free_freelQ_buffers(struct pci_dev *pdev, struct freelQ *Q)
+{
+	unsigned int cidx = Q->cidx, credits = Q->credits;
+
+	while (credits--) {
+		struct freelQ_ce *ce = &Q->centries[cidx];
+
+		pci_unmap_single(pdev, pci_unmap_addr(ce, dma_addr),
+				 pci_unmap_len(ce, dma_len),
+				 PCI_DMA_FROMDEVICE);
+		dev_kfree_skb(ce->skb);
+		ce->skb = NULL;
+		if (++cidx == Q->entries_n)
+			cidx = 0;
+	}
+}
+
+/*
+ * Free RX free list and response queue resources.
+ */
+static void free_rx_resources(struct sge *sge)
+{
+	struct pci_dev *pdev = sge->adapter->pdev;
+	unsigned int size, i;
+
+	if (sge->respQ.entries) {
+		size = sizeof(struct respQ_e) * sge->respQ.entries_n;
+		pci_free_consistent(pdev, size, sge->respQ.entries,
+				    sge->respQ.dma_addr);
+	}
+
+	for (i = 0; i < SGE_FREELQ_N; i++) {
+		struct freelQ *Q = &sge->freelQ[i];
+
+		if (Q->centries) {
+			free_freelQ_buffers(pdev, Q);
+			kfree(Q->centries);
+		}
+		if (Q->entries) {
+			size = sizeof(struct freelQ_e) * Q->entries_n;
+			pci_free_consistent(pdev, size, Q->entries,
+					    Q->dma_addr);
+		}
+	}
+}
+
+/*
+ * Allocates basic RX resources, consisting of memory mapped freelist Qs and a
+ * response Q.
+ */
+static int alloc_rx_resources(struct sge *sge, struct sge_params *p)
+{
+	struct pci_dev *pdev = sge->adapter->pdev;
+	unsigned int size, i;
+
+	for (i = 0; i < SGE_FREELQ_N; i++) {
+		struct freelQ *Q = &sge->freelQ[i];
+
+		Q->genbit = 1;
+		Q->entries_n = p->freelQ_size[i];
+		Q->dma_offset = SGE_RX_OFFSET - sge->rx_pkt_pad;
+		size = sizeof(struct freelQ_e) * Q->entries_n;
+		Q->entries = (struct freelQ_e *)
+			      pci_alloc_consistent(pdev, size, &Q->dma_addr);
+		if (!Q->entries)
+			goto err_no_mem;
+		memset(Q->entries, 0, size);
+		Q->centries = kcalloc(Q->entries_n, sizeof(struct freelQ_ce),
+				      GFP_KERNEL);
+		if (!Q->centries)
+			goto err_no_mem;
+	}
+
+	/*
+	 * Calculate the buffer sizes for the two free lists.  FL0 accommodates
+	 * regular sized Ethernet frames, FL1 is sized not to exceed 16K,
+	 * including all the sk_buff overhead.
+	 *
+	 * Note: For T2 FL0 and FL1 are reversed.
+	 */
+	sge->freelQ[!sge->jumbo_fl].rx_buffer_size = SGE_RX_SM_BUF_SIZE +
+		sizeof(struct cpl_rx_data) +
+		sge->freelQ[!sge->jumbo_fl].dma_offset;
+	sge->freelQ[sge->jumbo_fl].rx_buffer_size = (16 * 1024) -
+		SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
+
+	sge->respQ.genbit = 1;
+	sge->respQ.entries_n = SGE_RESPQ_E_N;
+	sge->respQ.credits = SGE_RESPQ_E_N;
+	size = sizeof(struct respQ_e) * sge->respQ.entries_n;
+	sge->respQ.entries = (struct respQ_e *)
+		pci_alloc_consistent(pdev, size, &sge->respQ.dma_addr);
+	if (!sge->respQ.entries)
+		goto err_no_mem;
+	memset(sge->respQ.entries, 0, size);
+	return 0;
+
+err_no_mem:
+	free_rx_resources(sge);
+	return -ENOMEM;
+}
+
+/*
+ * Frees 'credits_pend' TX buffers and returns the credits to Q->credits.
+ *
+ * The adaptive algorithm receives the total size of the buffers freed
+ * accumulated in @*totpayload. No initialization of this argument here.
+ *
+ */
+static void free_cmdQ_buffers(struct sge *sge, struct cmdQ *Q,
+			      unsigned int credits_pend, unsigned int *totpayload)
+{
+	struct pci_dev *pdev = sge->adapter->pdev;
+	struct sk_buff *skb;
+	struct cmdQ_ce *ce, *cq = Q->centries;
+	unsigned int entries_n = Q->entries_n, cidx = Q->cidx,
+		     i = credits_pend;
+
+
+	ce = &cq[cidx];
+	while (i--) {
+		if (ce->single)
+			pci_unmap_single(pdev, pci_unmap_addr(ce, dma_addr),
+					 pci_unmap_len(ce, dma_len),
+					 PCI_DMA_TODEVICE);
+		else
+			pci_unmap_page(pdev, pci_unmap_addr(ce, dma_addr),
+				       pci_unmap_len(ce, dma_len),
+				       PCI_DMA_TODEVICE);
+		if (totpayload)
+			*totpayload += pci_unmap_len(ce, dma_len);
+
+		skb = ce->skb;
+		if (skb)
+			dev_kfree_skb_irq(skb);
+
+		ce++;
+		if (++cidx == entries_n) {
+			cidx = 0;
+			ce = cq;
+		}
+	}
+
+	Q->cidx = cidx;
+	atomic_add(credits_pend, &Q->credits);
+}
+
+/*
+ * Free TX resources.
+ *
+ * Assumes that SGE is stopped and all interrupts are disabled.
+ */
+static void free_tx_resources(struct sge *sge)
+{
+	struct pci_dev *pdev = sge->adapter->pdev;
+	unsigned int size, i;
+
+	for (i = 0; i < SGE_CMDQ_N; i++) {
+		struct cmdQ *Q = &sge->cmdQ[i];
+
+		if (Q->centries) {
+			unsigned int pending = Q->entries_n -
+					       atomic_read(&Q->credits);
+
+			if (pending)
+				free_cmdQ_buffers(sge, Q, pending, NULL);
+			kfree(Q->centries);
+		}
+		if (Q->entries) {
+			size = sizeof(struct cmdQ_e) * Q->entries_n;
+			pci_free_consistent(pdev, size, Q->entries,
+					    Q->dma_addr);
+		}
+	}
+}
+
+/*
+ * Allocates basic TX resources, consisting of memory mapped command Qs.
+ */
+static int alloc_tx_resources(struct sge *sge, struct sge_params *p)
+{
+	struct pci_dev *pdev = sge->adapter->pdev;
+	unsigned int size, i;
+
+	for (i = 0; i < SGE_CMDQ_N; i++) {
+		struct cmdQ *Q = &sge->cmdQ[i];
+
+		Q->genbit = 1;
+		Q->entries_n = p->cmdQ_size[i];
+		atomic_set(&Q->credits, Q->entries_n);
+		atomic_set(&Q->asleep, 1);
+		spin_lock_init(&Q->Qlock);
+		size = sizeof(struct cmdQ_e) * Q->entries_n;
+		Q->entries = (struct cmdQ_e *)
+			      pci_alloc_consistent(pdev, size, &Q->dma_addr);
+		if (!Q->entries)
+			goto err_no_mem;
+		memset(Q->entries, 0, size);
+		Q->centries = kcalloc(Q->entries_n, sizeof(struct cmdQ_ce),
+				      GFP_KERNEL);
+		if (!Q->centries)
+			goto err_no_mem;
+	}
+
+	return 0;
+
+err_no_mem:
+	free_tx_resources(sge);
+	return -ENOMEM;
+}
+
+static inline void setup_ring_params(struct adapter *adapter, u64 addr,
+				     u32 size, int base_reg_lo,
+				     int base_reg_hi, int size_reg)
+{
+	t1_write_reg_4(adapter, base_reg_lo, (u32)addr);
+	t1_write_reg_4(adapter, base_reg_hi, addr >> 32);
+	t1_write_reg_4(adapter, size_reg, size);
+}
+
+/*
+ * Enable/disable VLAN acceleration.
+ */
+void t1_set_vlan_accel(struct adapter *adapter, int on_off)
+{
+	struct sge *sge = adapter->sge;
+
+	sge->sge_control &= ~F_VLAN_XTRACT;
+	if (on_off)
+		sge->sge_control |= F_VLAN_XTRACT;
+	if (adapter->open_device_map) {
+		t1_write_reg_4(adapter, A_SG_CONTROL, sge->sge_control);
+		t1_read_reg_4(adapter, A_SG_CONTROL);	/* flush */
+	}
+}
+
+/*
+ * Sets the interrupt latency timer when the adaptive Rx coalescing
+ * is turned off. Do nothing when it is turned on again.
+ *
+ * This routine relies on the fact that the caller has already set
+ * the adaptive policy in adapter->sge_params before calling it.
+*/
+int t1_sge_set_coalesce_params(struct sge *sge, struct sge_params *p)
+{
+	if (!p->coalesce_enable) {
+		u32 newTimer = p->rx_coalesce_usecs *
+			(board_info(sge->adapter)->clock_core / 1000000);
+
+		t1_write_reg_4(sge->adapter, A_SG_INTRTIMER, newTimer);
+	}
+	return 0;
+}
+
+/*
+ * Programs the various SGE registers. However, the engine is not yet enabled,
+ * but sge->sge_control is setup and ready to go.
+ */
+static void configure_sge(struct sge *sge, struct sge_params *p)
+{
+	struct adapter *ap = sge->adapter;
+	int i;
+
+	t1_write_reg_4(ap, A_SG_CONTROL, 0);
+	setup_ring_params(ap, sge->cmdQ[0].dma_addr, sge->cmdQ[0].entries_n,
+			  A_SG_CMD0BASELWR, A_SG_CMD0BASEUPR, A_SG_CMD0SIZE);
+	setup_ring_params(ap, sge->cmdQ[1].dma_addr, sge->cmdQ[1].entries_n,
+			  A_SG_CMD1BASELWR, A_SG_CMD1BASEUPR, A_SG_CMD1SIZE);
+	setup_ring_params(ap, sge->freelQ[0].dma_addr,
+			  sge->freelQ[0].entries_n, A_SG_FL0BASELWR,
+			  A_SG_FL0BASEUPR, A_SG_FL0SIZE);
+	setup_ring_params(ap, sge->freelQ[1].dma_addr,
+			  sge->freelQ[1].entries_n, A_SG_FL1BASELWR,
+			  A_SG_FL1BASEUPR, A_SG_FL1SIZE);
+
+	/* The threshold comparison uses <. */
+	t1_write_reg_4(ap, A_SG_FLTHRESHOLD, SGE_RX_SM_BUF_SIZE + 1);
+
+	setup_ring_params(ap, sge->respQ.dma_addr, sge->respQ.entries_n,
+			A_SG_RSPBASELWR, A_SG_RSPBASEUPR, A_SG_RSPSIZE);
+	t1_write_reg_4(ap, A_SG_RSPQUEUECREDIT, (u32)sge->respQ.entries_n);
+
+	sge->sge_control = F_CMDQ0_ENABLE | F_CMDQ1_ENABLE | F_FL0_ENABLE |
+		F_FL1_ENABLE | F_CPL_ENABLE | F_RESPONSE_QUEUE_ENABLE |
+		V_CMDQ_PRIORITY(2) | F_DISABLE_CMDQ1_GTS | F_ISCSI_COALESCE |
+		V_RX_PKT_OFFSET(sge->rx_pkt_pad);
+
+#if defined(__BIG_ENDIAN_BITFIELD)
+	sge->sge_control |= F_ENABLE_BIG_ENDIAN;
+#endif
+
+	/*
+	 * Initialize the SGE Interrupt Timer arrray:
+	 * intrtimer[0]	      = (SGE_INTRTIMER0) usec
+	 * intrtimer[0<i<5]   = (SGE_INTRTIMER0 + i*2) usec
+	 * intrtimer[4<i<10]  = ((i - 3) * 6) usec
+	 * intrtimer[10]      = (SGE_INTRTIMER1) usec
+	 *
+	 */
+	sge->intrtimer[0] = board_info(sge->adapter)->clock_core / 1000000;
+	for (i = 1; i < SGE_INTR_LATBUCKETS; ++i) {
+		sge->intrtimer[i] = SGE_INTRTIMER0 + (2 * i);
+		sge->intrtimer[i] *= sge->intrtimer[0];
+	}
+	for (i = SGE_INTR_LATBUCKETS; i < SGE_INTR_MAXBUCKETS - 1; ++i) {
+		sge->intrtimer[i] = (i - 3) * 6;
+		sge->intrtimer[i] *= sge->intrtimer[0];
+	}
+	sge->intrtimer[SGE_INTR_MAXBUCKETS - 1] =
+	  sge->intrtimer[0] * SGE_INTRTIMER1;
+	/* Initialize resource timer */
+	sge->intrtimer_nres = sge->intrtimer[0] * SGE_INTRTIMER_NRES;
+	/* Finally finish initialization of intrtimer[0] */
+	sge->intrtimer[0] *= SGE_INTRTIMER0;
+	/* Initialize for a throughput oriented workload */
+	sge->currIndex = SGE_INTR_MAXBUCKETS - 1;
+
+	if (p->coalesce_enable)
+		t1_write_reg_4(ap, A_SG_INTRTIMER,
+			       sge->intrtimer[sge->currIndex]);
+	else
+		t1_sge_set_coalesce_params(sge, p);
+}
+
+/*
+ * Return the payload capacity of the jumbo free-list buffers.
+ */
+static inline unsigned int jumbo_payload_capacity(const struct sge *sge)
+{
+	return sge->freelQ[sge->jumbo_fl].rx_buffer_size -
+		sizeof(struct cpl_rx_data) - SGE_RX_OFFSET + sge->rx_pkt_pad;
+}
+
+/*
+ * Allocates both RX and TX resources and configures the SGE. However,
+ * the hardware is not enabled yet.
+ */
+int t1_sge_configure(struct sge *sge, struct sge_params *p)
+{
+	if (alloc_rx_resources(sge, p))
+		return -ENOMEM;
+	if (alloc_tx_resources(sge, p)) {
+		free_rx_resources(sge);
+		return -ENOMEM;
+	}
+	configure_sge(sge, p);
+
+	/*
+	 * Now that we have sized the free lists calculate the payload
+	 * capacity of the large buffers.  Other parts of the driver use
+	 * this to set the max offload coalescing size so that RX packets
+	 * do not overflow our large buffers.
+	 */
+	p->large_buf_capacity = jumbo_payload_capacity(sge);
+	return 0;
+}
+
+/*
+ * Frees all SGE related resources and the sge structure itself
+ */
+void t1_sge_destroy(struct sge *sge)
+{
+	if (sge->pskb)
+		dev_kfree_skb(sge->pskb);
+	free_tx_resources(sge);
+	free_rx_resources(sge);
+	kfree(sge);
+}
+
+/*
+ * Allocates new RX buffers on the freelist Q (and tracks them on the freelist
+ * context Q) until the Q is full or alloc_skb fails.
+ *
+ * It is possible that the generation bits already match, indicating that the
+ * buffer is already valid and nothing needs to be done. This happens when we
+ * copied a received buffer into a new sk_buff during the interrupt processing.
+ *
+ * If the SGE doesn't automatically align packets properly (!sge->rx_pkt_pad),
+ * we specify a RX_OFFSET in order to make sure that the IP header is 4B
+ * aligned.
+ */
+static void refill_free_list(struct sge *sge, struct freelQ *Q)
+{
+	struct pci_dev *pdev = sge->adapter->pdev;
+	struct freelQ_ce *ce = &Q->centries[Q->pidx];
+	struct freelQ_e *e = &Q->entries[Q->pidx];
+	unsigned int dma_len = Q->rx_buffer_size - Q->dma_offset;
+
+
+	while (Q->credits < Q->entries_n) {
+		if (e->GenerationBit != Q->genbit) {
+			struct sk_buff *skb;
+			dma_addr_t mapping;
+
+			skb = alloc_skb(Q->rx_buffer_size, GFP_ATOMIC);
+			if (!skb)
+				break;
+			if (Q->dma_offset)
+				skb_reserve(skb, Q->dma_offset);
+			mapping = pci_map_single(pdev, skb->data, dma_len,
+						 PCI_DMA_FROMDEVICE);
+			ce->skb = skb;
+			pci_unmap_addr_set(ce, dma_addr, mapping);
+			pci_unmap_len_set(ce, dma_len, dma_len);
+			e->AddrLow = (u32)mapping;
+			e->AddrHigh = (u64)mapping >> 32;
+			e->BufferLength = dma_len;
+			e->GenerationBit = e->GenerationBit2 = Q->genbit;
+		}
+
+		e++;
+		ce++;
+		if (++Q->pidx == Q->entries_n) {
+			Q->pidx = 0;
+			Q->genbit ^= 1;
+			ce = Q->centries;
+			e = Q->entries;
+		}
+		Q->credits++;
+	}
+
+}
+
+/*
+ * Calls refill_free_list for both freelist Qs. If we cannot
+ * fill at least 1/4 of both Qs, we go into 'few interrupt mode' in order
+ * to give the system time to free up resources.
+ */
+static void freelQs_empty(struct sge *sge)
+{
+	u32 irq_reg = t1_read_reg_4(sge->adapter, A_SG_INT_ENABLE);
+	u32 irqholdoff_reg;
+
+	refill_free_list(sge, &sge->freelQ[0]);
+	refill_free_list(sge, &sge->freelQ[1]);
+
+	if (sge->freelQ[0].credits > (sge->freelQ[0].entries_n >> 2) &&
+	    sge->freelQ[1].credits > (sge->freelQ[1].entries_n >> 2)) {
+		irq_reg |= F_FL_EXHAUSTED;
+		irqholdoff_reg = sge->intrtimer[sge->currIndex];
+	} else {
+		/* Clear the F_FL_EXHAUSTED interrupts for now */
+		irq_reg &= ~F_FL_EXHAUSTED;
+		irqholdoff_reg = sge->intrtimer_nres;
+	}
+	t1_write_reg_4(sge->adapter, A_SG_INTRTIMER, irqholdoff_reg);
+	t1_write_reg_4(sge->adapter, A_SG_INT_ENABLE, irq_reg);
+
+	/* We reenable the Qs to force a freelist GTS interrupt later */
+	doorbell_pio(sge, F_FL0_ENABLE | F_FL1_ENABLE);
+}
+
+#define SGE_PL_INTR_MASK (F_PL_INTR_SGE_ERR | F_PL_INTR_SGE_DATA)
+#define SGE_INT_FATAL (F_RESPQ_OVERFLOW | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
+#define SGE_INT_ENABLE (F_RESPQ_EXHAUSTED | F_RESPQ_OVERFLOW | \
+			F_FL_EXHAUSTED | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
+
+/*
+ * Disable SGE Interrupts
+ */
+void t1_sge_intr_disable(struct sge *sge)
+{
+	u32 val = t1_read_reg_4(sge->adapter, A_PL_ENABLE);
+
+	t1_write_reg_4(sge->adapter, A_PL_ENABLE, val & ~SGE_PL_INTR_MASK);
+	t1_write_reg_4(sge->adapter, A_SG_INT_ENABLE, 0);
+}
+
+/*
+ * Enable SGE interrupts.
+ */
+void t1_sge_intr_enable(struct sge *sge)
+{
+	u32 en = SGE_INT_ENABLE;
+	u32 val = t1_read_reg_4(sge->adapter, A_PL_ENABLE);
+
+	if (sge->adapter->flags & TSO_CAPABLE)
+		en &= ~F_PACKET_TOO_BIG;
+	t1_write_reg_4(sge->adapter, A_SG_INT_ENABLE, en);
+	t1_write_reg_4(sge->adapter, A_PL_ENABLE, val | SGE_PL_INTR_MASK);
+}
+
+/*
+ * Clear SGE interrupts.
+ */
+void t1_sge_intr_clear(struct sge *sge)
+{
+	t1_write_reg_4(sge->adapter, A_PL_CAUSE, SGE_PL_INTR_MASK);
+	t1_write_reg_4(sge->adapter, A_SG_INT_CAUSE, 0xffffffff);
+}
+
+/*
+ * SGE 'Error' interrupt handler
+ */
+int t1_sge_intr_error_handler(struct sge *sge)
+{
+	struct adapter *adapter = sge->adapter;
+	u32 cause = t1_read_reg_4(adapter, A_SG_INT_CAUSE);
+
+	if (adapter->flags & TSO_CAPABLE)
+		cause &= ~F_PACKET_TOO_BIG;
+	if (cause & F_RESPQ_EXHAUSTED)
+		sge->intr_cnt.respQ_empty++;
+	if (cause & F_RESPQ_OVERFLOW) {
+		sge->intr_cnt.respQ_overflow++;
+		CH_ALERT("%s: SGE response queue overflow\n",
+			 adapter->name);
+	}
+	if (cause & F_FL_EXHAUSTED) {
+		sge->intr_cnt.freelistQ_empty++;
+		freelQs_empty(sge);
+	}
+	if (cause & F_PACKET_TOO_BIG) {
+		sge->intr_cnt.pkt_too_big++;
+		CH_ALERT("%s: SGE max packet size exceeded\n",
+			 adapter->name);
+	}
+	if (cause & F_PACKET_MISMATCH) {
+		sge->intr_cnt.pkt_mismatch++;
+		CH_ALERT("%s: SGE packet mismatch\n", adapter->name);
+	}
+	if (cause & SGE_INT_FATAL)
+		t1_fatal_err(adapter);
+
+	t1_write_reg_4(adapter, A_SG_INT_CAUSE, cause);
+	return 0;
+}
+
+/*
+ * The following code is copied from 2.6, where the skb_pull is doing the
+ * right thing and only pulls ETH_HLEN.
+ *
+ *	Determine the packet's protocol ID. The rule here is that we
+ *	assume 802.3 if the type field is short enough to be a length.
+ *	This is normal practice and works for any 'now in use' protocol.
+ */
+static unsigned short sge_eth_type_trans(struct sk_buff *skb,
+					 struct net_device *dev)
+{
+	struct ethhdr *eth;
+	unsigned char *rawp;
+
+	skb->mac.raw = skb->data;
+	skb_pull(skb, ETH_HLEN);
+	eth = (struct ethhdr *)skb->mac.raw;
+
+	if (*eth->h_dest&1) {
+		if(memcmp(eth->h_dest, dev->broadcast, ETH_ALEN) == 0)
+			skb->pkt_type = PACKET_BROADCAST;
+		else
+			skb->pkt_type = PACKET_MULTICAST;
+	}
+
+	/*
+	 *	This ALLMULTI check should be redundant by 1.4
+	 *	so don't forget to remove it.
+	 *
+	 *	Seems, you forgot to remove it. All silly devices
+	 *	seems to set IFF_PROMISC.
+	 */
+
+	else if (1 /*dev->flags&IFF_PROMISC*/)
+	{
+		if(memcmp(eth->h_dest,dev->dev_addr, ETH_ALEN))
+			skb->pkt_type=PACKET_OTHERHOST;
+	}
+
+	if (ntohs(eth->h_proto) >= 1536)
+		return eth->h_proto;
+
+	rawp = skb->data;
+
+	/*
+	 * This is a magic hack to spot IPX packets. Older Novell breaks
+	 * the protocol design and runs IPX over 802.3 without an 802.2 LLC
+	 * layer. We look for FFFF which isn't a used 802.2 SSAP/DSAP. This
+	 * won't work for fault tolerant netware but does for the rest.
+	 */
+	if (*(unsigned short *)rawp == 0xFFFF)
+		return htons(ETH_P_802_3);
+
+	/*
+	 * Real 802.2 LLC
+	 */
+	return htons(ETH_P_802_2);
+}
+
+/*
+ * Prepare the received buffer and pass it up the stack. If it is small enough
+ * and allocation doesn't fail, we use a new sk_buff and copy the content.
+ */
+static unsigned int t1_sge_rx(struct sge *sge, struct freelQ *Q,
+			      unsigned int len, unsigned int offload)
+{
+	struct sk_buff *skb;
+	struct adapter *adapter = sge->adapter;
+	struct freelQ_ce *ce = &Q->centries[Q->cidx];
+
+	if (len <= SGE_RX_COPY_THRESHOLD &&
+	    (skb = alloc_skb(len + NET_IP_ALIGN, GFP_ATOMIC))) {
+		struct freelQ_e *e;
+		char *src = ce->skb->data;
+
+		pci_dma_sync_single_for_cpu(adapter->pdev,
+					    pci_unmap_addr(ce, dma_addr),
+					    pci_unmap_len(ce, dma_len),
+					    PCI_DMA_FROMDEVICE);
+		if (!offload) {
+			skb_reserve(skb, NET_IP_ALIGN);
+			src += sge->rx_pkt_pad;
+		}
+		memcpy(skb->data, src, len);
+
+		/* Reuse the entry. */
+		e = &Q->entries[Q->cidx];
+		e->GenerationBit  ^= 1;
+		e->GenerationBit2 ^= 1;
+	} else {
+		pci_unmap_single(adapter->pdev, pci_unmap_addr(ce, dma_addr),
+				 pci_unmap_len(ce, dma_len),
+				 PCI_DMA_FROMDEVICE);
+		skb = ce->skb;
+		if (!offload && sge->rx_pkt_pad)
+			__skb_pull(skb, sge->rx_pkt_pad);
+	}
+
+	skb_put(skb, len);
+
+
+	if (unlikely(offload)) {
+		{
+			printk(KERN_ERR
+			       "%s: unexpected offloaded packet, cmd %u\n",
+			       adapter->name, *skb->data);
+			dev_kfree_skb_any(skb);
+		}
+	} else {
+		struct cpl_rx_pkt *p = (struct cpl_rx_pkt *)skb->data;
+
+		skb_pull(skb, sizeof(*p));
+		skb->dev = adapter->port[p->iff].dev;
+		skb->dev->last_rx = jiffies;
+		skb->protocol = sge_eth_type_trans(skb, skb->dev);
+		if ((adapter->flags & RX_CSUM_ENABLED) && p->csum == 0xffff &&
+		    skb->protocol == htons(ETH_P_IP) &&
+		    (skb->data[9] == IPPROTO_TCP ||
+		     skb->data[9] == IPPROTO_UDP))
+			skb->ip_summed = CHECKSUM_UNNECESSARY;
+		else
+			skb->ip_summed = CHECKSUM_NONE;
+		if (adapter->vlan_grp && p->vlan_valid)
+			vlan_hwaccel_rx(skb, adapter->vlan_grp,
+					ntohs(p->vlan));
+		else
+			netif_rx(skb);
+	}
+
+	if (++Q->cidx == Q->entries_n)
+		Q->cidx = 0;
+
+	if (unlikely(--Q->credits < Q->entries_n - SGE_FREEL_REFILL_THRESH))
+		refill_free_list(sge, Q);
+	return 1;
+}
+
+
+/*
+ * Adaptive interrupt timer logic to keep the CPU utilization to
+ * manageable levels. Basically, as the Average Packet Size (APS)
+ * gets higher, the interrupt latency setting gets longer. Every
+ * SGE_INTR_BUCKETSIZE (of 100B) causes a bump of 2usec to the
+ * base value of SGE_INTRTIMER0. At large values of payload the
+ * latency hits the ceiling value of SGE_INTRTIMER1 stored at
+ * index SGE_INTR_MAXBUCKETS-1 in sge->intrtimer[].
+ *
+ * sge->currIndex caches the last index to save unneeded PIOs.
+ */
+static inline void update_intr_timer(struct sge *sge, unsigned int avg_payload)
+{
+	unsigned int newIndex;
+
+	newIndex = avg_payload / SGE_INTR_BUCKETSIZE;
+	if (newIndex > SGE_INTR_MAXBUCKETS - 1) {
+		newIndex = SGE_INTR_MAXBUCKETS - 1;
+	}
+	/* Save a PIO with this check....maybe */
+	if (newIndex != sge->currIndex) {
+		t1_write_reg_4(sge->adapter, A_SG_INTRTIMER,
+			       sge->intrtimer[newIndex]);
+		sge->currIndex = newIndex;
+		sge->adapter->params.sge.last_rx_coalesce_raw =
+			sge->intrtimer[newIndex];
+	}
+}
+
+/*
+ * Returns true if command queue q_num has enough available descriptors that
+ * we can resume Tx operation after temporarily disabling its packet queue.
+ */
+static inline int enough_free_Tx_descs(struct sge *sge, int q_num)
+{
+	return atomic_read(&sge->cmdQ[q_num].credits) >
+		(sge->cmdQ[q_num].entries_n >> 2);
+}
+
+/*
+ * Main interrupt handler, optimized assuming that we took a 'DATA'
+ * interrupt.
+ *
+ * 1. Clear the interrupt
+ * 2. Loop while we find valid descriptors and process them; accumulate
+ *	information that can be processed after the loop
+ * 3. Tell the SGE at which index we stopped processing descriptors
+ * 4. Bookkeeping; free TX buffers, ring doorbell if there are any
+ *	outstanding TX buffers waiting, replenish RX buffers, potentially
+ *	reenable upper layers if they were turned off due to lack of TX
+ *	resources which are available again.
+ * 5. If we took an interrupt, but no valid respQ descriptors was found we
+ *	let the slow_intr_handler run and do error handling.
+ */
+irqreturn_t t1_interrupt(int irq, void *cookie, struct pt_regs *regs)
+{
+	struct net_device *netdev;
+	struct adapter *adapter = cookie;
+	struct sge *sge = adapter->sge;
+	struct respQ *Q = &sge->respQ;
+	unsigned int credits = Q->credits, flags = 0, ret = 0;
+	unsigned int tot_rxpayload = 0, tot_txpayload = 0, n_rx = 0, n_tx = 0;
+	unsigned int credits_pend[SGE_CMDQ_N] = { 0, 0 };
+
+	struct respQ_e *e = &Q->entries[Q->cidx];
+	prefetch(e);
+
+	t1_write_reg_4(adapter, A_PL_CAUSE, F_PL_INTR_SGE_DATA);
+
+
+	while (e->GenerationBit == Q->genbit) {
+		if (--credits < SGE_RESPQ_REPLENISH_THRES) {
+			u32 n = Q->entries_n - credits - 1;
+
+			t1_write_reg_4(adapter, A_SG_RSPQUEUECREDIT, n);
+			credits += n;
+		}
+		if (likely(e->DataValid)) {
+			if (!e->Sop || !e->Eop)
+				BUG();
+			t1_sge_rx(sge, &sge->freelQ[e->FreelistQid],
+				  e->BufferLength, e->Offload);
+			tot_rxpayload += e->BufferLength;
+			++n_rx;
+		}
+		flags |= e->Qsleeping;
+		credits_pend[0] += e->Cmdq0CreditReturn;
+		credits_pend[1] += e->Cmdq1CreditReturn;
+
+#ifdef CONFIG_SMP
+		/*
+		 * If enough cmdQ0 buffers have finished DMAing free them so
+		 * anyone that may be waiting for their release can continue.
+		 * We do this only on MP systems to allow other CPUs to proceed
+		 * promptly.  UP systems can wait for the free_cmdQ_buffers()
+		 * calls after this loop as the sole CPU is currently busy in
+		 * this loop.
+		 */
+		if (unlikely(credits_pend[0] > SGE_FREEL_REFILL_THRESH)) {
+			free_cmdQ_buffers(sge, &sge->cmdQ[0], credits_pend[0],
+					  &tot_txpayload);
+			n_tx += credits_pend[0];
+			credits_pend[0] = 0;
+		}
+#endif
+		ret++;
+		e++;
+		if (unlikely(++Q->cidx == Q->entries_n)) {
+			Q->cidx = 0;
+			Q->genbit ^= 1;
+			e = Q->entries;
+		}
+	}
+
+	Q->credits = credits;
+	t1_write_reg_4(adapter, A_SG_SLEEPING, Q->cidx);
+
+	if (credits_pend[0])
+		free_cmdQ_buffers(sge, &sge->cmdQ[0], credits_pend[0], &tot_txpayload);
+	if (credits_pend[1])
+		free_cmdQ_buffers(sge, &sge->cmdQ[1], credits_pend[1], &tot_txpayload);
+
+	/* Do any coalescing and interrupt latency timer adjustments */
+	if (adapter->params.sge.coalesce_enable) {
+		unsigned int	avg_txpayload = 0, avg_rxpayload = 0;
+
+		n_tx += credits_pend[0] + credits_pend[1];
+
+		/*
+		 * Choose larger avg. payload size to increase
+		 * throughput and reduce [CPU util., intr/s.]
+		 *
+		 * Throughput behavior favored in mixed-mode.
+		 */
+		if (n_tx)
+			avg_txpayload = tot_txpayload/n_tx;
+		if (n_rx)
+			avg_rxpayload = tot_rxpayload/n_rx;
+
+		if (n_tx && avg_txpayload > avg_rxpayload){
+			update_intr_timer(sge, avg_txpayload);
+		} else if (n_rx) {
+			update_intr_timer(sge, avg_rxpayload);
+		}
+	}
+
+	if (flags & F_CMDQ0_ENABLE) {
+		struct cmdQ *cmdQ = &sge->cmdQ[0];
+
+		atomic_set(&cmdQ->asleep, 1);
+		if (atomic_read(&cmdQ->pio_pidx) != cmdQ->pidx) {
+			doorbell_pio(sge, F_CMDQ0_ENABLE);
+			atomic_set(&cmdQ->pio_pidx, cmdQ->pidx);
+		}
+	}
+	if (unlikely(flags & (F_FL0_ENABLE | F_FL1_ENABLE)))
+		freelQs_empty(sge);
+
+	netdev = adapter->port[0].dev;
+	if (unlikely(netif_queue_stopped(netdev) && netif_carrier_ok(netdev) &&
+		     enough_free_Tx_descs(sge, 0) &&
+		     enough_free_Tx_descs(sge, 1))) {
+		netif_wake_queue(netdev);
+	}
+	if (unlikely(!ret))
+		ret = t1_slow_intr_handler(adapter);
+
+	return IRQ_RETVAL(ret != 0);
+}
+
+/*
+ * Enqueues the sk_buff onto the cmdQ[qid] and has hardware fetch it.
+ *
+ * The code figures out how many entries the sk_buff will require in the
+ * cmdQ and updates the cmdQ data structure with the state once the enqueue
+ * has complete. Then, it doesn't access the global structure anymore, but
+ * uses the corresponding fields on the stack. In conjuction with a spinlock
+ * around that code, we can make the function reentrant without holding the
+ * lock when we actually enqueue (which might be expensive, especially on
+ * architectures with IO MMUs).
+ */
+static unsigned int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter,
+		       unsigned int qid)
+{
+	struct sge *sge = adapter->sge;
+	struct cmdQ *Q = &sge->cmdQ[qid];
+	struct cmdQ_e *e;
+	struct cmdQ_ce *ce;
+	dma_addr_t mapping;
+	unsigned int credits, pidx, genbit;
+
+	unsigned int count = 1 + skb_shinfo(skb)->nr_frags;
+
+	/*
+	 * Coming from the timer
+	 */
+	if ((skb == sge->pskb)) {
+		/*
+		 * Quit if any cmdQ activities
+		 */
+		if (!spin_trylock(&Q->Qlock))
+			return 0;
+		if (atomic_read(&Q->credits) != Q->entries_n) {
+			spin_unlock(&Q->Qlock);
+			return 0;
+		}
+	}
+	else
+		spin_lock(&Q->Qlock);
+
+	genbit = Q->genbit;
+	pidx = Q->pidx;
+	credits = atomic_read(&Q->credits);
+
+	credits -= count;
+	atomic_sub(count, &Q->credits);
+	Q->pidx += count;
+	if (Q->pidx >= Q->entries_n) {
+		Q->pidx -= Q->entries_n;
+		Q->genbit ^= 1;
+	}
+
+	if (unlikely(credits < (MAX_SKB_FRAGS + 1))) {
+		sge->intr_cnt.cmdQ_full[qid]++;
+		netif_stop_queue(adapter->port[0].dev);
+	}
+	spin_unlock(&Q->Qlock);
+
+	mapping = pci_map_single(adapter->pdev, skb->data,
+				 skb->len - skb->data_len, PCI_DMA_TODEVICE);
+	ce = &Q->centries[pidx];
+	ce->skb = NULL;
+	pci_unmap_addr_set(ce, dma_addr, mapping);
+	pci_unmap_len_set(ce, dma_len, skb->len - skb->data_len);
+	ce->single = 1;
+
+	e = &Q->entries[pidx];
+	e->Sop =  1;
+	e->DataValid = 1;
+	e->BufferLength = skb->len - skb->data_len;
+	e->AddrHigh = (u64)mapping >> 32;
+	e->AddrLow = (u32)mapping;
+
+	if (--count > 0) {
+		unsigned int i;
+
+		e->Eop = 0;
+		wmb();
+		e->GenerationBit = e->GenerationBit2 = genbit;
+
+		for (i = 0; i < count; i++) {
+			skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
+
+			ce++; e++;
+			if (++pidx == Q->entries_n) {
+				pidx = 0;
+				genbit ^= 1;
+				ce = Q->centries;
+				e = Q->entries;
+			}
+
+			mapping = pci_map_page(adapter->pdev, frag->page,
+					       frag->page_offset,
+					       frag->size,
+					       PCI_DMA_TODEVICE);
+			ce->skb = NULL;
+			pci_unmap_addr_set(ce, dma_addr, mapping);
+			pci_unmap_len_set(ce, dma_len, frag->size);
+			ce->single = 0;
+
+			e->Sop = 0;
+			e->DataValid = 1;
+			e->BufferLength = frag->size;
+			e->AddrHigh = (u64)mapping >> 32;
+			e->AddrLow = (u32)mapping;
+
+			if (i < count - 1) {
+				e->Eop = 0;
+				wmb();
+				e->GenerationBit = e->GenerationBit2 = genbit;
+			}
+		}
+	}
+
+	if (skb != sge->pskb)
+		ce->skb = skb;
+	e->Eop = 1;
+	wmb();
+	e->GenerationBit = e->GenerationBit2 = genbit;
+
+	/*
+	 * We always ring the doorbell for cmdQ1.  For cmdQ0, we only ring
+	 * the doorbell if the Q is asleep. There is a natural race, where
+	 * the hardware is going to sleep just after we checked, however,
+	 * then the interrupt handler will detect the outstanding TX packet
+	 * and ring the doorbell for us.
+	 */
+	if (qid) {
+		doorbell_pio(sge, F_CMDQ1_ENABLE);
+	} else if (atomic_read(&Q->asleep)) {
+		atomic_set(&Q->asleep, 0);
+		doorbell_pio(sge, F_CMDQ0_ENABLE);
+		atomic_set(&Q->pio_pidx, Q->pidx);
+	}
+	return 0;
+}
+
+#define MK_ETH_TYPE_MSS(type, mss) (((mss) & 0x3FFF) | ((type) << 14))
+
+/*
+ * Adds the CPL header to the sk_buff and passes it to t1_sge_tx.
+ */
+int t1_start_xmit(struct sk_buff *skb, struct net_device *dev)
+{
+	struct adapter *adapter = dev->priv;
+	struct cpl_tx_pkt *cpl;
+	struct ethhdr *eth;
+	size_t max_len;
+
+	/*
+	 * We are using a non-standard hard_header_len and some kernel
+	 * components, such as pktgen, do not handle it right.  Complain
+	 * when this happens but try to fix things up.
+	 */
+	if (unlikely(skb_headroom(skb) < dev->hard_header_len - ETH_HLEN)) {
+		struct sk_buff *orig_skb = skb;
+
+		if (net_ratelimit())
+			printk(KERN_ERR
+			       "%s: Tx packet has inadequate headroom\n",
+			       dev->name);
+		skb = skb_realloc_headroom(skb, sizeof(struct cpl_tx_pkt_lso));
+		dev_kfree_skb_any(orig_skb);
+		if (!skb)
+			return -ENOMEM;
+	}
+
+	if (skb_shinfo(skb)->tso_size) {
+		int eth_type;
+		struct cpl_tx_pkt_lso *hdr;
+
+		eth_type = skb->nh.raw - skb->data == ETH_HLEN ?
+			CPL_ETH_II : CPL_ETH_II_VLAN;
+
+		hdr = (struct cpl_tx_pkt_lso *)skb_push(skb, sizeof(*hdr));
+		hdr->opcode = CPL_TX_PKT_LSO;
+		hdr->ip_csum_dis = hdr->l4_csum_dis = 0;
+		hdr->ip_hdr_words = skb->nh.iph->ihl;
+		hdr->tcp_hdr_words = skb->h.th->doff;
+		hdr->eth_type_mss = htons(MK_ETH_TYPE_MSS(eth_type,
+						skb_shinfo(skb)->tso_size));
+		hdr->len = htonl(skb->len - sizeof(*hdr));
+		cpl = (struct cpl_tx_pkt *)hdr;
+	} else
+	{
+		/*
+		 * An Ethernet packet must have at least space for
+		 * the DIX Ethernet header and be no greater than
+		 * the device set MTU. Otherwise trash the packet.
+		 */
+		if (skb->len < ETH_HLEN)
+			goto t1_start_xmit_fail2;
+		eth = (struct ethhdr *)skb->data;
+		if (eth->h_proto == htons(ETH_P_8021Q))
+			max_len = dev->mtu + VLAN_ETH_HLEN;
+		else
+			max_len = dev->mtu + ETH_HLEN;
+		if (skb->len > max_len)
+			goto t1_start_xmit_fail2;
+
+		if (!(adapter->flags & UDP_CSUM_CAPABLE) &&
+		    skb->ip_summed == CHECKSUM_HW &&
+		    skb->nh.iph->protocol == IPPROTO_UDP &&
+		    skb_checksum_help(skb, 0))
+			goto t1_start_xmit_fail3;
+
+
+		if (!adapter->sge->pskb) {
+			if (skb->protocol == htons(ETH_P_ARP) &&
+			    skb->nh.arph->ar_op == htons(ARPOP_REQUEST))
+				adapter->sge->pskb = skb;
+		}
+		cpl = (struct cpl_tx_pkt *)skb_push(skb, sizeof(*cpl));
+		cpl->opcode = CPL_TX_PKT;
+		cpl->ip_csum_dis = 1;    /* SW calculates IP csum */
+		cpl->l4_csum_dis = skb->ip_summed == CHECKSUM_HW ? 0 : 1;
+		/* the length field isn't used so don't bother setting it */
+	}
+	cpl->iff = dev->if_port;
+
+#if defined(CONFIG_VLAN_8021Q) || defined(CONFIG_VLAN_8021Q_MODULE)
+	if (adapter->vlan_grp && vlan_tx_tag_present(skb)) {
+		cpl->vlan_valid = 1;
+		cpl->vlan = htons(vlan_tx_tag_get(skb));
+	} else
+#endif
+		cpl->vlan_valid = 0;
+
+	dev->trans_start = jiffies;
+	return t1_sge_tx(skb, adapter, 0);
+
+t1_start_xmit_fail3:
+	printk(KERN_INFO "%s: Unable to complete checksum\n", dev->name);
+	goto t1_start_xmit_fail1;
+
+t1_start_xmit_fail2:
+	printk(KERN_INFO "%s: Invalid packet length %d, dropping\n",
+			dev->name, skb->len);
+
+t1_start_xmit_fail1:
+	dev_kfree_skb_any(skb);
+	return 0;
+}
+
+void t1_sge_set_ptimeout(adapter_t *adapter, u32 val)
+{
+	struct sge *sge = adapter->sge;
+
+	if (is_T2(adapter))
+		sge->ptimeout = max((u32)((HZ * val) / 1000), (u32)1);
+}
+
+u32 t1_sge_get_ptimeout(adapter_t *adapter)
+{
+	struct sge *sge = adapter->sge;
+
+	return (is_T2(adapter) ? ((sge->ptimeout * 1000) / HZ) : 0);
+}
+