Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / drivers / net / ethernet / adaptec / starfire.c
diff --git a/kernel/drivers/net/ethernet/adaptec/starfire.c b/kernel/drivers/net/ethernet/adaptec/starfire.c
new file mode 100644 (file)
index 0000000..ac72882
--- /dev/null
@@ -0,0 +1,2060 @@
+/* starfire.c: Linux device driver for the Adaptec Starfire network adapter. */
+/*
+       Written 1998-2000 by Donald Becker.
+
+       Current maintainer is Ion Badulescu <ionut ta badula tod org>. Please
+       send all bug reports to me, and not to Donald Becker, as this code
+       has been heavily modified from Donald's original version.
+
+       This software may be used and distributed according to the terms of
+       the GNU General Public License (GPL), incorporated herein by reference.
+       Drivers based on or derived from this code fall under the GPL and must
+       retain the authorship, copyright and license notice.  This file is not
+       a complete program and may only be used when the entire operating
+       system is licensed under the GPL.
+
+       The information below comes from Donald Becker's original driver:
+
+       The author may be reached as becker@scyld.com, or C/O
+       Scyld Computing Corporation
+       410 Severn Ave., Suite 210
+       Annapolis MD 21403
+
+       Support and updates available at
+       http://www.scyld.com/network/starfire.html
+       [link no longer provides useful info -jgarzik]
+
+*/
+
+#define DRV_NAME       "starfire"
+#define DRV_VERSION    "2.1"
+#define DRV_RELDATE    "July  6, 2008"
+
+#include <linux/interrupt.h>
+#include <linux/module.h>
+#include <linux/kernel.h>
+#include <linux/pci.h>
+#include <linux/netdevice.h>
+#include <linux/etherdevice.h>
+#include <linux/init.h>
+#include <linux/delay.h>
+#include <linux/crc32.h>
+#include <linux/ethtool.h>
+#include <linux/mii.h>
+#include <linux/if_vlan.h>
+#include <linux/mm.h>
+#include <linux/firmware.h>
+#include <asm/processor.h>             /* Processor type for cache alignment. */
+#include <asm/uaccess.h>
+#include <asm/io.h>
+
+/*
+ * The current frame processor firmware fails to checksum a fragment
+ * of length 1. If and when this is fixed, the #define below can be removed.
+ */
+#define HAS_BROKEN_FIRMWARE
+
+/*
+ * If using the broken firmware, data must be padded to the next 32-bit boundary.
+ */
+#ifdef HAS_BROKEN_FIRMWARE
+#define PADDING_MASK 3
+#endif
+
+/*
+ * Define this if using the driver with the zero-copy patch
+ */
+#define ZEROCOPY
+
+#if defined(CONFIG_VLAN_8021Q) || defined(CONFIG_VLAN_8021Q_MODULE)
+#define VLAN_SUPPORT
+#endif
+
+/* The user-configurable values.
+   These may be modified when a driver module is loaded.*/
+
+/* Used for tuning interrupt latency vs. overhead. */
+static int intr_latency;
+static int small_frames;
+
+static int debug = 1;                  /* 1 normal messages, 0 quiet .. 7 verbose. */
+static int max_interrupt_work = 20;
+static int mtu;
+/* Maximum number of multicast addresses to filter (vs. rx-all-multicast).
+   The Starfire has a 512 element hash table based on the Ethernet CRC. */
+static const int multicast_filter_limit = 512;
+/* Whether to do TCP/UDP checksums in hardware */
+static int enable_hw_cksum = 1;
+
+#define PKT_BUF_SZ     1536            /* Size of each temporary Rx buffer.*/
+/*
+ * Set the copy breakpoint for the copy-only-tiny-frames scheme.
+ * Setting to > 1518 effectively disables this feature.
+ *
+ * NOTE:
+ * The ia64 doesn't allow for unaligned loads even of integers being
+ * misaligned on a 2 byte boundary. Thus always force copying of
+ * packets as the starfire doesn't allow for misaligned DMAs ;-(
+ * 23/10/2000 - Jes
+ *
+ * The Alpha and the Sparc don't like unaligned loads, either. On Sparc64,
+ * at least, having unaligned frames leads to a rather serious performance
+ * penalty. -Ion
+ */
+#if defined(__ia64__) || defined(__alpha__) || defined(__sparc__)
+static int rx_copybreak = PKT_BUF_SZ;
+#else
+static int rx_copybreak /* = 0 */;
+#endif
+
+/* PCI DMA burst size -- on sparc64 we want to force it to 64 bytes, on the others the default of 128 is fine. */
+#ifdef __sparc__
+#define DMA_BURST_SIZE 64
+#else
+#define DMA_BURST_SIZE 128
+#endif
+
+/* Operational parameters that are set at compile time. */
+
+/* The "native" ring sizes are either 256 or 2048.
+   However in some modes a descriptor may be marked to wrap the ring earlier.
+*/
+#define RX_RING_SIZE   256
+#define TX_RING_SIZE   32
+/* The completion queues are fixed at 1024 entries i.e. 4K or 8KB. */
+#define DONE_Q_SIZE    1024
+/* All queues must be aligned on a 256-byte boundary */
+#define QUEUE_ALIGN    256
+
+#if RX_RING_SIZE > 256
+#define RX_Q_ENTRIES Rx2048QEntries
+#else
+#define RX_Q_ENTRIES Rx256QEntries
+#endif
+
+/* Operational parameters that usually are not changed. */
+/* Time in jiffies before concluding the transmitter is hung. */
+#define TX_TIMEOUT     (2 * HZ)
+
+#ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
+/* 64-bit dma_addr_t */
+#define ADDR_64BITS    /* This chip uses 64 bit addresses. */
+#define netdrv_addr_t __le64
+#define cpu_to_dma(x) cpu_to_le64(x)
+#define dma_to_cpu(x) le64_to_cpu(x)
+#define RX_DESC_Q_ADDR_SIZE RxDescQAddr64bit
+#define TX_DESC_Q_ADDR_SIZE TxDescQAddr64bit
+#define RX_COMPL_Q_ADDR_SIZE RxComplQAddr64bit
+#define TX_COMPL_Q_ADDR_SIZE TxComplQAddr64bit
+#define RX_DESC_ADDR_SIZE RxDescAddr64bit
+#else  /* 32-bit dma_addr_t */
+#define netdrv_addr_t __le32
+#define cpu_to_dma(x) cpu_to_le32(x)
+#define dma_to_cpu(x) le32_to_cpu(x)
+#define RX_DESC_Q_ADDR_SIZE RxDescQAddr32bit
+#define TX_DESC_Q_ADDR_SIZE TxDescQAddr32bit
+#define RX_COMPL_Q_ADDR_SIZE RxComplQAddr32bit
+#define TX_COMPL_Q_ADDR_SIZE TxComplQAddr32bit
+#define RX_DESC_ADDR_SIZE RxDescAddr32bit
+#endif
+
+#define skb_first_frag_len(skb)        skb_headlen(skb)
+#define skb_num_frags(skb) (skb_shinfo(skb)->nr_frags + 1)
+
+/* Firmware names */
+#define FIRMWARE_RX    "adaptec/starfire_rx.bin"
+#define FIRMWARE_TX    "adaptec/starfire_tx.bin"
+
+/* These identify the driver base version and may not be removed. */
+static const char version[] =
+KERN_INFO "starfire.c:v1.03 7/26/2000  Written by Donald Becker <becker@scyld.com>\n"
+" (unofficial 2.2/2.4 kernel port, version " DRV_VERSION ", " DRV_RELDATE ")\n";
+
+MODULE_AUTHOR("Donald Becker <becker@scyld.com>");
+MODULE_DESCRIPTION("Adaptec Starfire Ethernet driver");
+MODULE_LICENSE("GPL");
+MODULE_VERSION(DRV_VERSION);
+MODULE_FIRMWARE(FIRMWARE_RX);
+MODULE_FIRMWARE(FIRMWARE_TX);
+
+module_param(max_interrupt_work, int, 0);
+module_param(mtu, int, 0);
+module_param(debug, int, 0);
+module_param(rx_copybreak, int, 0);
+module_param(intr_latency, int, 0);
+module_param(small_frames, int, 0);
+module_param(enable_hw_cksum, int, 0);
+MODULE_PARM_DESC(max_interrupt_work, "Maximum events handled per interrupt");
+MODULE_PARM_DESC(mtu, "MTU (all boards)");
+MODULE_PARM_DESC(debug, "Debug level (0-6)");
+MODULE_PARM_DESC(rx_copybreak, "Copy breakpoint for copy-only-tiny-frames");
+MODULE_PARM_DESC(intr_latency, "Maximum interrupt latency, in microseconds");
+MODULE_PARM_DESC(small_frames, "Maximum size of receive frames that bypass interrupt latency (0,64,128,256,512)");
+MODULE_PARM_DESC(enable_hw_cksum, "Enable/disable hardware cksum support (0/1)");
+
+/*
+                               Theory of Operation
+
+I. Board Compatibility
+
+This driver is for the Adaptec 6915 "Starfire" 64 bit PCI Ethernet adapter.
+
+II. Board-specific settings
+
+III. Driver operation
+
+IIIa. Ring buffers
+
+The Starfire hardware uses multiple fixed-size descriptor queues/rings.  The
+ring sizes are set fixed by the hardware, but may optionally be wrapped
+earlier by the END bit in the descriptor.
+This driver uses that hardware queue size for the Rx ring, where a large
+number of entries has no ill effect beyond increases the potential backlog.
+The Tx ring is wrapped with the END bit, since a large hardware Tx queue
+disables the queue layer priority ordering and we have no mechanism to
+utilize the hardware two-level priority queue.  When modifying the
+RX/TX_RING_SIZE pay close attention to page sizes and the ring-empty warning
+levels.
+
+IIIb/c. Transmit/Receive Structure
+
+See the Adaptec manual for the many possible structures, and options for
+each structure.  There are far too many to document all of them here.
+
+For transmit this driver uses type 0/1 transmit descriptors (depending
+on the 32/64 bitness of the architecture), and relies on automatic
+minimum-length padding.  It does not use the completion queue
+consumer index, but instead checks for non-zero status entries.
+
+For receive this driver uses type 2/3 receive descriptors.  The driver
+allocates full frame size skbuffs for the Rx ring buffers, so all frames
+should fit in a single descriptor.  The driver does not use the completion
+queue consumer index, but instead checks for non-zero status entries.
+
+When an incoming frame is less than RX_COPYBREAK bytes long, a fresh skbuff
+is allocated and the frame is copied to the new skbuff.  When the incoming
+frame is larger, the skbuff is passed directly up the protocol stack.
+Buffers consumed this way are replaced by newly allocated skbuffs in a later
+phase of receive.
+
+A notable aspect of operation is that unaligned buffers are not permitted by
+the Starfire hardware.  Thus the IP header at offset 14 in an ethernet frame
+isn't longword aligned, which may cause problems on some machine
+e.g. Alphas and IA64. For these architectures, the driver is forced to copy
+the frame into a new skbuff unconditionally. Copied frames are put into the
+skbuff at an offset of "+2", thus 16-byte aligning the IP header.
+
+IIId. Synchronization
+
+The driver runs as two independent, single-threaded flows of control.  One
+is the send-packet routine, which enforces single-threaded use by the
+dev->tbusy flag.  The other thread is the interrupt handler, which is single
+threaded by the hardware and interrupt handling software.
+
+The send packet thread has partial control over the Tx ring and the netif_queue
+status. If the number of free Tx slots in the ring falls below a certain number
+(currently hardcoded to 4), it signals the upper layer to stop the queue.
+
+The interrupt handler has exclusive control over the Rx ring and records stats
+from the Tx ring.  After reaping the stats, it marks the Tx queue entry as
+empty by incrementing the dirty_tx mark. Iff the netif_queue is stopped and the
+number of free Tx slow is above the threshold, it signals the upper layer to
+restart the queue.
+
+IV. Notes
+
+IVb. References
+
+The Adaptec Starfire manuals, available only from Adaptec.
+http://www.scyld.com/expert/100mbps.html
+http://www.scyld.com/expert/NWay.html
+
+IVc. Errata
+
+- StopOnPerr is broken, don't enable
+- Hardware ethernet padding exposes random data, perform software padding
+  instead (unverified -- works correctly for all the hardware I have)
+
+*/
+
+
+
+enum chip_capability_flags {CanHaveMII=1, };
+
+enum chipset {
+       CH_6915 = 0,
+};
+
+static const struct pci_device_id starfire_pci_tbl[] = {
+       { PCI_VDEVICE(ADAPTEC, 0x6915), CH_6915 },
+       { 0, }
+};
+MODULE_DEVICE_TABLE(pci, starfire_pci_tbl);
+
+/* A chip capabilities table, matching the CH_xxx entries in xxx_pci_tbl[] above. */
+static const struct chip_info {
+       const char *name;
+       int drv_flags;
+} netdrv_tbl[] = {
+       { "Adaptec Starfire 6915", CanHaveMII },
+};
+
+
+/* Offsets to the device registers.
+   Unlike software-only systems, device drivers interact with complex hardware.
+   It's not useful to define symbolic names for every register bit in the
+   device.  The name can only partially document the semantics and make
+   the driver longer and more difficult to read.
+   In general, only the important configuration values or bits changed
+   multiple times should be defined symbolically.
+*/
+enum register_offsets {
+       PCIDeviceConfig=0x50040, GenCtrl=0x50070, IntrTimerCtrl=0x50074,
+       IntrClear=0x50080, IntrStatus=0x50084, IntrEnable=0x50088,
+       MIICtrl=0x52000, TxStationAddr=0x50120, EEPROMCtrl=0x51000,
+       GPIOCtrl=0x5008C, TxDescCtrl=0x50090,
+       TxRingPtr=0x50098, HiPriTxRingPtr=0x50094, /* Low and High priority. */
+       TxRingHiAddr=0x5009C,           /* 64 bit address extension. */
+       TxProducerIdx=0x500A0, TxConsumerIdx=0x500A4,
+       TxThreshold=0x500B0,
+       CompletionHiAddr=0x500B4, TxCompletionAddr=0x500B8,
+       RxCompletionAddr=0x500BC, RxCompletionQ2Addr=0x500C0,
+       CompletionQConsumerIdx=0x500C4, RxDMACtrl=0x500D0,
+       RxDescQCtrl=0x500D4, RxDescQHiAddr=0x500DC, RxDescQAddr=0x500E0,
+       RxDescQIdx=0x500E8, RxDMAStatus=0x500F0, RxFilterMode=0x500F4,
+       TxMode=0x55000, VlanType=0x55064,
+       PerfFilterTable=0x56000, HashTable=0x56100,
+       TxGfpMem=0x58000, RxGfpMem=0x5a000,
+};
+
+/*
+ * Bits in the interrupt status/mask registers.
+ * Warning: setting Intr[Ab]NormalSummary in the IntrEnable register
+ * enables all the interrupt sources that are or'ed into those status bits.
+ */
+enum intr_status_bits {
+       IntrLinkChange=0xf0000000, IntrStatsMax=0x08000000,
+       IntrAbnormalSummary=0x02000000, IntrGeneralTimer=0x01000000,
+       IntrSoftware=0x800000, IntrRxComplQ1Low=0x400000,
+       IntrTxComplQLow=0x200000, IntrPCI=0x100000,
+       IntrDMAErr=0x080000, IntrTxDataLow=0x040000,
+       IntrRxComplQ2Low=0x020000, IntrRxDescQ1Low=0x010000,
+       IntrNormalSummary=0x8000, IntrTxDone=0x4000,
+       IntrTxDMADone=0x2000, IntrTxEmpty=0x1000,
+       IntrEarlyRxQ2=0x0800, IntrEarlyRxQ1=0x0400,
+       IntrRxQ2Done=0x0200, IntrRxQ1Done=0x0100,
+       IntrRxGFPDead=0x80, IntrRxDescQ2Low=0x40,
+       IntrNoTxCsum=0x20, IntrTxBadID=0x10,
+       IntrHiPriTxBadID=0x08, IntrRxGfp=0x04,
+       IntrTxGfp=0x02, IntrPCIPad=0x01,
+       /* not quite bits */
+       IntrRxDone=IntrRxQ2Done | IntrRxQ1Done,
+       IntrRxEmpty=IntrRxDescQ1Low | IntrRxDescQ2Low,
+       IntrNormalMask=0xff00, IntrAbnormalMask=0x3ff00fe,
+};
+
+/* Bits in the RxFilterMode register. */
+enum rx_mode_bits {
+       AcceptBroadcast=0x04, AcceptAllMulticast=0x02, AcceptAll=0x01,
+       AcceptMulticast=0x10, PerfectFilter=0x40, HashFilter=0x30,
+       PerfectFilterVlan=0x80, MinVLANPrio=0xE000, VlanMode=0x0200,
+       WakeupOnGFP=0x0800,
+};
+
+/* Bits in the TxMode register */
+enum tx_mode_bits {
+       MiiSoftReset=0x8000, MIILoopback=0x4000,
+       TxFlowEnable=0x0800, RxFlowEnable=0x0400,
+       PadEnable=0x04, FullDuplex=0x02, HugeFrame=0x01,
+};
+
+/* Bits in the TxDescCtrl register. */
+enum tx_ctrl_bits {
+       TxDescSpaceUnlim=0x00, TxDescSpace32=0x10, TxDescSpace64=0x20,
+       TxDescSpace128=0x30, TxDescSpace256=0x40,
+       TxDescType0=0x00, TxDescType1=0x01, TxDescType2=0x02,
+       TxDescType3=0x03, TxDescType4=0x04,
+       TxNoDMACompletion=0x08,
+       TxDescQAddr64bit=0x80, TxDescQAddr32bit=0,
+       TxHiPriFIFOThreshShift=24, TxPadLenShift=16,
+       TxDMABurstSizeShift=8,
+};
+
+/* Bits in the RxDescQCtrl register. */
+enum rx_ctrl_bits {
+       RxBufferLenShift=16, RxMinDescrThreshShift=0,
+       RxPrefetchMode=0x8000, RxVariableQ=0x2000,
+       Rx2048QEntries=0x4000, Rx256QEntries=0,
+       RxDescAddr64bit=0x1000, RxDescAddr32bit=0,
+       RxDescQAddr64bit=0x0100, RxDescQAddr32bit=0,
+       RxDescSpace4=0x000, RxDescSpace8=0x100,
+       RxDescSpace16=0x200, RxDescSpace32=0x300,
+       RxDescSpace64=0x400, RxDescSpace128=0x500,
+       RxConsumerWrEn=0x80,
+};
+
+/* Bits in the RxDMACtrl register. */
+enum rx_dmactrl_bits {
+       RxReportBadFrames=0x80000000, RxDMAShortFrames=0x40000000,
+       RxDMABadFrames=0x20000000, RxDMACrcErrorFrames=0x10000000,
+       RxDMAControlFrame=0x08000000, RxDMAPauseFrame=0x04000000,
+       RxChecksumIgnore=0, RxChecksumRejectTCPUDP=0x02000000,
+       RxChecksumRejectTCPOnly=0x01000000,
+       RxCompletionQ2Enable=0x800000,
+       RxDMAQ2Disable=0, RxDMAQ2FPOnly=0x100000,
+       RxDMAQ2SmallPkt=0x200000, RxDMAQ2HighPrio=0x300000,
+       RxDMAQ2NonIP=0x400000,
+       RxUseBackupQueue=0x080000, RxDMACRC=0x040000,
+       RxEarlyIntThreshShift=12, RxHighPrioThreshShift=8,
+       RxBurstSizeShift=0,
+};
+
+/* Bits in the RxCompletionAddr register */
+enum rx_compl_bits {
+       RxComplQAddr64bit=0x80, RxComplQAddr32bit=0,
+       RxComplProducerWrEn=0x40,
+       RxComplType0=0x00, RxComplType1=0x10,
+       RxComplType2=0x20, RxComplType3=0x30,
+       RxComplThreshShift=0,
+};
+
+/* Bits in the TxCompletionAddr register */
+enum tx_compl_bits {
+       TxComplQAddr64bit=0x80, TxComplQAddr32bit=0,
+       TxComplProducerWrEn=0x40,
+       TxComplIntrStatus=0x20,
+       CommonQueueMode=0x10,
+       TxComplThreshShift=0,
+};
+
+/* Bits in the GenCtrl register */
+enum gen_ctrl_bits {
+       RxEnable=0x05, TxEnable=0x0a,
+       RxGFPEnable=0x10, TxGFPEnable=0x20,
+};
+
+/* Bits in the IntrTimerCtrl register */
+enum intr_ctrl_bits {
+       Timer10X=0x800, EnableIntrMasking=0x60, SmallFrameBypass=0x100,
+       SmallFrame64=0, SmallFrame128=0x200, SmallFrame256=0x400, SmallFrame512=0x600,
+       IntrLatencyMask=0x1f,
+};
+
+/* The Rx and Tx buffer descriptors. */
+struct starfire_rx_desc {
+       netdrv_addr_t rxaddr;
+};
+enum rx_desc_bits {
+       RxDescValid=1, RxDescEndRing=2,
+};
+
+/* Completion queue entry. */
+struct short_rx_done_desc {
+       __le32 status;                  /* Low 16 bits is length. */
+};
+struct basic_rx_done_desc {
+       __le32 status;                  /* Low 16 bits is length. */
+       __le16 vlanid;
+       __le16 status2;
+};
+struct csum_rx_done_desc {
+       __le32 status;                  /* Low 16 bits is length. */
+       __le16 csum;                    /* Partial checksum */
+       __le16 status2;
+};
+struct full_rx_done_desc {
+       __le32 status;                  /* Low 16 bits is length. */
+       __le16 status3;
+       __le16 status2;
+       __le16 vlanid;
+       __le16 csum;                    /* partial checksum */
+       __le32 timestamp;
+};
+/* XXX: this is ugly and I'm not sure it's worth the trouble -Ion */
+#ifdef VLAN_SUPPORT
+typedef struct full_rx_done_desc rx_done_desc;
+#define RxComplType RxComplType3
+#else  /* not VLAN_SUPPORT */
+typedef struct csum_rx_done_desc rx_done_desc;
+#define RxComplType RxComplType2
+#endif /* not VLAN_SUPPORT */
+
+enum rx_done_bits {
+       RxOK=0x20000000, RxFIFOErr=0x10000000, RxBufQ2=0x08000000,
+};
+
+/* Type 1 Tx descriptor. */
+struct starfire_tx_desc_1 {
+       __le32 status;                  /* Upper bits are status, lower 16 length. */
+       __le32 addr;
+};
+
+/* Type 2 Tx descriptor. */
+struct starfire_tx_desc_2 {
+       __le32 status;                  /* Upper bits are status, lower 16 length. */
+       __le32 reserved;
+       __le64 addr;
+};
+
+#ifdef ADDR_64BITS
+typedef struct starfire_tx_desc_2 starfire_tx_desc;
+#define TX_DESC_TYPE TxDescType2
+#else  /* not ADDR_64BITS */
+typedef struct starfire_tx_desc_1 starfire_tx_desc;
+#define TX_DESC_TYPE TxDescType1
+#endif /* not ADDR_64BITS */
+#define TX_DESC_SPACING TxDescSpaceUnlim
+
+enum tx_desc_bits {
+       TxDescID=0xB0000000,
+       TxCRCEn=0x01000000, TxDescIntr=0x08000000,
+       TxRingWrap=0x04000000, TxCalTCP=0x02000000,
+};
+struct tx_done_desc {
+       __le32 status;                  /* timestamp, index. */
+#if 0
+       __le32 intrstatus;              /* interrupt status */
+#endif
+};
+
+struct rx_ring_info {
+       struct sk_buff *skb;
+       dma_addr_t mapping;
+};
+struct tx_ring_info {
+       struct sk_buff *skb;
+       dma_addr_t mapping;
+       unsigned int used_slots;
+};
+
+#define PHY_CNT                2
+struct netdev_private {
+       /* Descriptor rings first for alignment. */
+       struct starfire_rx_desc *rx_ring;
+       starfire_tx_desc *tx_ring;
+       dma_addr_t rx_ring_dma;
+       dma_addr_t tx_ring_dma;
+       /* The addresses of rx/tx-in-place skbuffs. */
+       struct rx_ring_info rx_info[RX_RING_SIZE];
+       struct tx_ring_info tx_info[TX_RING_SIZE];
+       /* Pointers to completion queues (full pages). */
+       rx_done_desc *rx_done_q;
+       dma_addr_t rx_done_q_dma;
+       unsigned int rx_done;
+       struct tx_done_desc *tx_done_q;
+       dma_addr_t tx_done_q_dma;
+       unsigned int tx_done;
+       struct napi_struct napi;
+       struct net_device *dev;
+       struct pci_dev *pci_dev;
+#ifdef VLAN_SUPPORT
+       unsigned long active_vlans[BITS_TO_LONGS(VLAN_N_VID)];
+#endif
+       void *queue_mem;
+       dma_addr_t queue_mem_dma;
+       size_t queue_mem_size;
+
+       /* Frequently used values: keep some adjacent for cache effect. */
+       spinlock_t lock;
+       unsigned int cur_rx, dirty_rx;  /* Producer/consumer ring indices */
+       unsigned int cur_tx, dirty_tx, reap_tx;
+       unsigned int rx_buf_sz;         /* Based on MTU+slack. */
+       /* These values keep track of the transceiver/media in use. */
+       int speed100;                   /* Set if speed == 100MBit. */
+       u32 tx_mode;
+       u32 intr_timer_ctrl;
+       u8 tx_threshold;
+       /* MII transceiver section. */
+       struct mii_if_info mii_if;              /* MII lib hooks/info */
+       int phy_cnt;                    /* MII device addresses. */
+       unsigned char phys[PHY_CNT];    /* MII device addresses. */
+       void __iomem *base;
+};
+
+
+static int     mdio_read(struct net_device *dev, int phy_id, int location);
+static void    mdio_write(struct net_device *dev, int phy_id, int location, int value);
+static int     netdev_open(struct net_device *dev);
+static void    check_duplex(struct net_device *dev);
+static void    tx_timeout(struct net_device *dev);
+static void    init_ring(struct net_device *dev);
+static netdev_tx_t start_tx(struct sk_buff *skb, struct net_device *dev);
+static irqreturn_t intr_handler(int irq, void *dev_instance);
+static void    netdev_error(struct net_device *dev, int intr_status);
+static int     __netdev_rx(struct net_device *dev, int *quota);
+static int     netdev_poll(struct napi_struct *napi, int budget);
+static void    refill_rx_ring(struct net_device *dev);
+static void    netdev_error(struct net_device *dev, int intr_status);
+static void    set_rx_mode(struct net_device *dev);
+static struct net_device_stats *get_stats(struct net_device *dev);
+static int     netdev_ioctl(struct net_device *dev, struct ifreq *rq, int cmd);
+static int     netdev_close(struct net_device *dev);
+static void    netdev_media_change(struct net_device *dev);
+static const struct ethtool_ops ethtool_ops;
+
+
+#ifdef VLAN_SUPPORT
+static int netdev_vlan_rx_add_vid(struct net_device *dev,
+                                 __be16 proto, u16 vid)
+{
+       struct netdev_private *np = netdev_priv(dev);
+
+       spin_lock(&np->lock);
+       if (debug > 1)
+               printk("%s: Adding vlanid %d to vlan filter\n", dev->name, vid);
+       set_bit(vid, np->active_vlans);
+       set_rx_mode(dev);
+       spin_unlock(&np->lock);
+
+       return 0;
+}
+
+static int netdev_vlan_rx_kill_vid(struct net_device *dev,
+                                  __be16 proto, u16 vid)
+{
+       struct netdev_private *np = netdev_priv(dev);
+
+       spin_lock(&np->lock);
+       if (debug > 1)
+               printk("%s: removing vlanid %d from vlan filter\n", dev->name, vid);
+       clear_bit(vid, np->active_vlans);
+       set_rx_mode(dev);
+       spin_unlock(&np->lock);
+
+       return 0;
+}
+#endif /* VLAN_SUPPORT */
+
+
+static const struct net_device_ops netdev_ops = {
+       .ndo_open               = netdev_open,
+       .ndo_stop               = netdev_close,
+       .ndo_start_xmit         = start_tx,
+       .ndo_tx_timeout         = tx_timeout,
+       .ndo_get_stats          = get_stats,
+       .ndo_set_rx_mode        = set_rx_mode,
+       .ndo_do_ioctl           = netdev_ioctl,
+       .ndo_change_mtu         = eth_change_mtu,
+       .ndo_set_mac_address    = eth_mac_addr,
+       .ndo_validate_addr      = eth_validate_addr,
+#ifdef VLAN_SUPPORT
+       .ndo_vlan_rx_add_vid    = netdev_vlan_rx_add_vid,
+       .ndo_vlan_rx_kill_vid   = netdev_vlan_rx_kill_vid,
+#endif
+};
+
+static int starfire_init_one(struct pci_dev *pdev,
+                            const struct pci_device_id *ent)
+{
+       struct device *d = &pdev->dev;
+       struct netdev_private *np;
+       int i, irq, chip_idx = ent->driver_data;
+       struct net_device *dev;
+       long ioaddr;
+       void __iomem *base;
+       int drv_flags, io_size;
+       int boguscnt;
+
+/* when built into the kernel, we only print version if device is found */
+#ifndef MODULE
+       static int printed_version;
+       if (!printed_version++)
+               printk(version);
+#endif
+
+       if (pci_enable_device (pdev))
+               return -EIO;
+
+       ioaddr = pci_resource_start(pdev, 0);
+       io_size = pci_resource_len(pdev, 0);
+       if (!ioaddr || ((pci_resource_flags(pdev, 0) & IORESOURCE_MEM) == 0)) {
+               dev_err(d, "no PCI MEM resources, aborting\n");
+               return -ENODEV;
+       }
+
+       dev = alloc_etherdev(sizeof(*np));
+       if (!dev)
+               return -ENOMEM;
+
+       SET_NETDEV_DEV(dev, &pdev->dev);
+
+       irq = pdev->irq;
+
+       if (pci_request_regions (pdev, DRV_NAME)) {
+               dev_err(d, "cannot reserve PCI resources, aborting\n");
+               goto err_out_free_netdev;
+       }
+
+       base = ioremap(ioaddr, io_size);
+       if (!base) {
+               dev_err(d, "cannot remap %#x @ %#lx, aborting\n",
+                       io_size, ioaddr);
+               goto err_out_free_res;
+       }
+
+       pci_set_master(pdev);
+
+       /* enable MWI -- it vastly improves Rx performance on sparc64 */
+       pci_try_set_mwi(pdev);
+
+#ifdef ZEROCOPY
+       /* Starfire can do TCP/UDP checksumming */
+       if (enable_hw_cksum)
+               dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG;
+#endif /* ZEROCOPY */
+
+#ifdef VLAN_SUPPORT
+       dev->features |= NETIF_F_HW_VLAN_CTAG_RX | NETIF_F_HW_VLAN_CTAG_FILTER;
+#endif /* VLAN_RX_KILL_VID */
+#ifdef ADDR_64BITS
+       dev->features |= NETIF_F_HIGHDMA;
+#endif /* ADDR_64BITS */
+
+       /* Serial EEPROM reads are hidden by the hardware. */
+       for (i = 0; i < 6; i++)
+               dev->dev_addr[i] = readb(base + EEPROMCtrl + 20 - i);
+
+#if ! defined(final_version) /* Dump the EEPROM contents during development. */
+       if (debug > 4)
+               for (i = 0; i < 0x20; i++)
+                       printk("%2.2x%s",
+                              (unsigned int)readb(base + EEPROMCtrl + i),
+                              i % 16 != 15 ? " " : "\n");
+#endif
+
+       /* Issue soft reset */
+       writel(MiiSoftReset, base + TxMode);
+       udelay(1000);
+       writel(0, base + TxMode);
+
+       /* Reset the chip to erase previous misconfiguration. */
+       writel(1, base + PCIDeviceConfig);
+       boguscnt = 1000;
+       while (--boguscnt > 0) {
+               udelay(10);
+               if ((readl(base + PCIDeviceConfig) & 1) == 0)
+                       break;
+       }
+       if (boguscnt == 0)
+               printk("%s: chipset reset never completed!\n", dev->name);
+       /* wait a little longer */
+       udelay(1000);
+
+       np = netdev_priv(dev);
+       np->dev = dev;
+       np->base = base;
+       spin_lock_init(&np->lock);
+       pci_set_drvdata(pdev, dev);
+
+       np->pci_dev = pdev;
+
+       np->mii_if.dev = dev;
+       np->mii_if.mdio_read = mdio_read;
+       np->mii_if.mdio_write = mdio_write;
+       np->mii_if.phy_id_mask = 0x1f;
+       np->mii_if.reg_num_mask = 0x1f;
+
+       drv_flags = netdrv_tbl[chip_idx].drv_flags;
+
+       np->speed100 = 1;
+
+       /* timer resolution is 128 * 0.8us */
+       np->intr_timer_ctrl = (((intr_latency * 10) / 1024) & IntrLatencyMask) |
+               Timer10X | EnableIntrMasking;
+
+       if (small_frames > 0) {
+               np->intr_timer_ctrl |= SmallFrameBypass;
+               switch (small_frames) {
+               case 1 ... 64:
+                       np->intr_timer_ctrl |= SmallFrame64;
+                       break;
+               case 65 ... 128:
+                       np->intr_timer_ctrl |= SmallFrame128;
+                       break;
+               case 129 ... 256:
+                       np->intr_timer_ctrl |= SmallFrame256;
+                       break;
+               default:
+                       np->intr_timer_ctrl |= SmallFrame512;
+                       if (small_frames > 512)
+                               printk("Adjusting small_frames down to 512\n");
+                       break;
+               }
+       }
+
+       dev->netdev_ops = &netdev_ops;
+       dev->watchdog_timeo = TX_TIMEOUT;
+       dev->ethtool_ops = &ethtool_ops;
+
+       netif_napi_add(dev, &np->napi, netdev_poll, max_interrupt_work);
+
+       if (mtu)
+               dev->mtu = mtu;
+
+       if (register_netdev(dev))
+               goto err_out_cleardev;
+
+       printk(KERN_INFO "%s: %s at %p, %pM, IRQ %d.\n",
+              dev->name, netdrv_tbl[chip_idx].name, base,
+              dev->dev_addr, irq);
+
+       if (drv_flags & CanHaveMII) {
+               int phy, phy_idx = 0;
+               int mii_status;
+               for (phy = 0; phy < 32 && phy_idx < PHY_CNT; phy++) {
+                       mdio_write(dev, phy, MII_BMCR, BMCR_RESET);
+                       mdelay(100);
+                       boguscnt = 1000;
+                       while (--boguscnt > 0)
+                               if ((mdio_read(dev, phy, MII_BMCR) & BMCR_RESET) == 0)
+                                       break;
+                       if (boguscnt == 0) {
+                               printk("%s: PHY#%d reset never completed!\n", dev->name, phy);
+                               continue;
+                       }
+                       mii_status = mdio_read(dev, phy, MII_BMSR);
+                       if (mii_status != 0) {
+                               np->phys[phy_idx++] = phy;
+                               np->mii_if.advertising = mdio_read(dev, phy, MII_ADVERTISE);
+                               printk(KERN_INFO "%s: MII PHY found at address %d, status "
+                                          "%#4.4x advertising %#4.4x.\n",
+                                          dev->name, phy, mii_status, np->mii_if.advertising);
+                               /* there can be only one PHY on-board */
+                               break;
+                       }
+               }
+               np->phy_cnt = phy_idx;
+               if (np->phy_cnt > 0)
+                       np->mii_if.phy_id = np->phys[0];
+               else
+                       memset(&np->mii_if, 0, sizeof(np->mii_if));
+       }
+
+       printk(KERN_INFO "%s: scatter-gather and hardware TCP cksumming %s.\n",
+              dev->name, enable_hw_cksum ? "enabled" : "disabled");
+       return 0;
+
+err_out_cleardev:
+       iounmap(base);
+err_out_free_res:
+       pci_release_regions (pdev);
+err_out_free_netdev:
+       free_netdev(dev);
+       return -ENODEV;
+}
+
+
+/* Read the MII Management Data I/O (MDIO) interfaces. */
+static int mdio_read(struct net_device *dev, int phy_id, int location)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       void __iomem *mdio_addr = np->base + MIICtrl + (phy_id<<7) + (location<<2);
+       int result, boguscnt=1000;
+       /* ??? Should we add a busy-wait here? */
+       do {
+               result = readl(mdio_addr);
+       } while ((result & 0xC0000000) != 0x80000000 && --boguscnt > 0);
+       if (boguscnt == 0)
+               return 0;
+       if ((result & 0xffff) == 0xffff)
+               return 0;
+       return result & 0xffff;
+}
+
+
+static void mdio_write(struct net_device *dev, int phy_id, int location, int value)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       void __iomem *mdio_addr = np->base + MIICtrl + (phy_id<<7) + (location<<2);
+       writel(value, mdio_addr);
+       /* The busy-wait will occur before a read. */
+}
+
+
+static int netdev_open(struct net_device *dev)
+{
+       const struct firmware *fw_rx, *fw_tx;
+       const __be32 *fw_rx_data, *fw_tx_data;
+       struct netdev_private *np = netdev_priv(dev);
+       void __iomem *ioaddr = np->base;
+       const int irq = np->pci_dev->irq;
+       int i, retval;
+       size_t tx_size, rx_size;
+       size_t tx_done_q_size, rx_done_q_size, tx_ring_size, rx_ring_size;
+
+       /* Do we ever need to reset the chip??? */
+
+       retval = request_irq(irq, intr_handler, IRQF_SHARED, dev->name, dev);
+       if (retval)
+               return retval;
+
+       /* Disable the Rx and Tx, and reset the chip. */
+       writel(0, ioaddr + GenCtrl);
+       writel(1, ioaddr + PCIDeviceConfig);
+       if (debug > 1)
+               printk(KERN_DEBUG "%s: netdev_open() irq %d.\n",
+                      dev->name, irq);
+
+       /* Allocate the various queues. */
+       if (!np->queue_mem) {
+               tx_done_q_size = ((sizeof(struct tx_done_desc) * DONE_Q_SIZE + QUEUE_ALIGN - 1) / QUEUE_ALIGN) * QUEUE_ALIGN;
+               rx_done_q_size = ((sizeof(rx_done_desc) * DONE_Q_SIZE + QUEUE_ALIGN - 1) / QUEUE_ALIGN) * QUEUE_ALIGN;
+               tx_ring_size = ((sizeof(starfire_tx_desc) * TX_RING_SIZE + QUEUE_ALIGN - 1) / QUEUE_ALIGN) * QUEUE_ALIGN;
+               rx_ring_size = sizeof(struct starfire_rx_desc) * RX_RING_SIZE;
+               np->queue_mem_size = tx_done_q_size + rx_done_q_size + tx_ring_size + rx_ring_size;
+               np->queue_mem = pci_alloc_consistent(np->pci_dev, np->queue_mem_size, &np->queue_mem_dma);
+               if (np->queue_mem == NULL) {
+                       free_irq(irq, dev);
+                       return -ENOMEM;
+               }
+
+               np->tx_done_q     = np->queue_mem;
+               np->tx_done_q_dma = np->queue_mem_dma;
+               np->rx_done_q     = (void *) np->tx_done_q + tx_done_q_size;
+               np->rx_done_q_dma = np->tx_done_q_dma + tx_done_q_size;
+               np->tx_ring       = (void *) np->rx_done_q + rx_done_q_size;
+               np->tx_ring_dma   = np->rx_done_q_dma + rx_done_q_size;
+               np->rx_ring       = (void *) np->tx_ring + tx_ring_size;
+               np->rx_ring_dma   = np->tx_ring_dma + tx_ring_size;
+       }
+
+       /* Start with no carrier, it gets adjusted later */
+       netif_carrier_off(dev);
+       init_ring(dev);
+       /* Set the size of the Rx buffers. */
+       writel((np->rx_buf_sz << RxBufferLenShift) |
+              (0 << RxMinDescrThreshShift) |
+              RxPrefetchMode | RxVariableQ |
+              RX_Q_ENTRIES |
+              RX_DESC_Q_ADDR_SIZE | RX_DESC_ADDR_SIZE |
+              RxDescSpace4,
+              ioaddr + RxDescQCtrl);
+
+       /* Set up the Rx DMA controller. */
+       writel(RxChecksumIgnore |
+              (0 << RxEarlyIntThreshShift) |
+              (6 << RxHighPrioThreshShift) |
+              ((DMA_BURST_SIZE / 32) << RxBurstSizeShift),
+              ioaddr + RxDMACtrl);
+
+       /* Set Tx descriptor */
+       writel((2 << TxHiPriFIFOThreshShift) |
+              (0 << TxPadLenShift) |
+              ((DMA_BURST_SIZE / 32) << TxDMABurstSizeShift) |
+              TX_DESC_Q_ADDR_SIZE |
+              TX_DESC_SPACING | TX_DESC_TYPE,
+              ioaddr + TxDescCtrl);
+
+       writel( (np->queue_mem_dma >> 16) >> 16, ioaddr + RxDescQHiAddr);
+       writel( (np->queue_mem_dma >> 16) >> 16, ioaddr + TxRingHiAddr);
+       writel( (np->queue_mem_dma >> 16) >> 16, ioaddr + CompletionHiAddr);
+       writel(np->rx_ring_dma, ioaddr + RxDescQAddr);
+       writel(np->tx_ring_dma, ioaddr + TxRingPtr);
+
+       writel(np->tx_done_q_dma, ioaddr + TxCompletionAddr);
+       writel(np->rx_done_q_dma |
+              RxComplType |
+              (0 << RxComplThreshShift),
+              ioaddr + RxCompletionAddr);
+
+       if (debug > 1)
+               printk(KERN_DEBUG "%s: Filling in the station address.\n", dev->name);
+
+       /* Fill both the Tx SA register and the Rx perfect filter. */
+       for (i = 0; i < 6; i++)
+               writeb(dev->dev_addr[i], ioaddr + TxStationAddr + 5 - i);
+       /* The first entry is special because it bypasses the VLAN filter.
+          Don't use it. */
+       writew(0, ioaddr + PerfFilterTable);
+       writew(0, ioaddr + PerfFilterTable + 4);
+       writew(0, ioaddr + PerfFilterTable + 8);
+       for (i = 1; i < 16; i++) {
+               __be16 *eaddrs = (__be16 *)dev->dev_addr;
+               void __iomem *setup_frm = ioaddr + PerfFilterTable + i * 16;
+               writew(be16_to_cpu(eaddrs[2]), setup_frm); setup_frm += 4;
+               writew(be16_to_cpu(eaddrs[1]), setup_frm); setup_frm += 4;
+               writew(be16_to_cpu(eaddrs[0]), setup_frm); setup_frm += 8;
+       }
+
+       /* Initialize other registers. */
+       /* Configure the PCI bus bursts and FIFO thresholds. */
+       np->tx_mode = TxFlowEnable|RxFlowEnable|PadEnable;      /* modified when link is up. */
+       writel(MiiSoftReset | np->tx_mode, ioaddr + TxMode);
+       udelay(1000);
+       writel(np->tx_mode, ioaddr + TxMode);
+       np->tx_threshold = 4;
+       writel(np->tx_threshold, ioaddr + TxThreshold);
+
+       writel(np->intr_timer_ctrl, ioaddr + IntrTimerCtrl);
+
+       napi_enable(&np->napi);
+
+       netif_start_queue(dev);
+
+       if (debug > 1)
+               printk(KERN_DEBUG "%s: Setting the Rx and Tx modes.\n", dev->name);
+       set_rx_mode(dev);
+
+       np->mii_if.advertising = mdio_read(dev, np->phys[0], MII_ADVERTISE);
+       check_duplex(dev);
+
+       /* Enable GPIO interrupts on link change */
+       writel(0x0f00ff00, ioaddr + GPIOCtrl);
+
+       /* Set the interrupt mask */
+       writel(IntrRxDone | IntrRxEmpty | IntrDMAErr |
+              IntrTxDMADone | IntrStatsMax | IntrLinkChange |
+              IntrRxGFPDead | IntrNoTxCsum | IntrTxBadID,
+              ioaddr + IntrEnable);
+       /* Enable PCI interrupts. */
+       writel(0x00800000 | readl(ioaddr + PCIDeviceConfig),
+              ioaddr + PCIDeviceConfig);
+
+#ifdef VLAN_SUPPORT
+       /* Set VLAN type to 802.1q */
+       writel(ETH_P_8021Q, ioaddr + VlanType);
+#endif /* VLAN_SUPPORT */
+
+       retval = request_firmware(&fw_rx, FIRMWARE_RX, &np->pci_dev->dev);
+       if (retval) {
+               printk(KERN_ERR "starfire: Failed to load firmware \"%s\"\n",
+                      FIRMWARE_RX);
+               goto out_init;
+       }
+       if (fw_rx->size % 4) {
+               printk(KERN_ERR "starfire: bogus length %zu in \"%s\"\n",
+                      fw_rx->size, FIRMWARE_RX);
+               retval = -EINVAL;
+               goto out_rx;
+       }
+       retval = request_firmware(&fw_tx, FIRMWARE_TX, &np->pci_dev->dev);
+       if (retval) {
+               printk(KERN_ERR "starfire: Failed to load firmware \"%s\"\n",
+                      FIRMWARE_TX);
+               goto out_rx;
+       }
+       if (fw_tx->size % 4) {
+               printk(KERN_ERR "starfire: bogus length %zu in \"%s\"\n",
+                      fw_tx->size, FIRMWARE_TX);
+               retval = -EINVAL;
+               goto out_tx;
+       }
+       fw_rx_data = (const __be32 *)&fw_rx->data[0];
+       fw_tx_data = (const __be32 *)&fw_tx->data[0];
+       rx_size = fw_rx->size / 4;
+       tx_size = fw_tx->size / 4;
+
+       /* Load Rx/Tx firmware into the frame processors */
+       for (i = 0; i < rx_size; i++)
+               writel(be32_to_cpup(&fw_rx_data[i]), ioaddr + RxGfpMem + i * 4);
+       for (i = 0; i < tx_size; i++)
+               writel(be32_to_cpup(&fw_tx_data[i]), ioaddr + TxGfpMem + i * 4);
+       if (enable_hw_cksum)
+               /* Enable the Rx and Tx units, and the Rx/Tx frame processors. */
+               writel(TxEnable|TxGFPEnable|RxEnable|RxGFPEnable, ioaddr + GenCtrl);
+       else
+               /* Enable the Rx and Tx units only. */
+               writel(TxEnable|RxEnable, ioaddr + GenCtrl);
+
+       if (debug > 1)
+               printk(KERN_DEBUG "%s: Done netdev_open().\n",
+                      dev->name);
+
+out_tx:
+       release_firmware(fw_tx);
+out_rx:
+       release_firmware(fw_rx);
+out_init:
+       if (retval)
+               netdev_close(dev);
+       return retval;
+}
+
+
+static void check_duplex(struct net_device *dev)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       u16 reg0;
+       int silly_count = 1000;
+
+       mdio_write(dev, np->phys[0], MII_ADVERTISE, np->mii_if.advertising);
+       mdio_write(dev, np->phys[0], MII_BMCR, BMCR_RESET);
+       udelay(500);
+       while (--silly_count && mdio_read(dev, np->phys[0], MII_BMCR) & BMCR_RESET)
+               /* do nothing */;
+       if (!silly_count) {
+               printk("%s: MII reset failed!\n", dev->name);
+               return;
+       }
+
+       reg0 = mdio_read(dev, np->phys[0], MII_BMCR);
+
+       if (!np->mii_if.force_media) {
+               reg0 |= BMCR_ANENABLE | BMCR_ANRESTART;
+       } else {
+               reg0 &= ~(BMCR_ANENABLE | BMCR_ANRESTART);
+               if (np->speed100)
+                       reg0 |= BMCR_SPEED100;
+               if (np->mii_if.full_duplex)
+                       reg0 |= BMCR_FULLDPLX;
+               printk(KERN_DEBUG "%s: Link forced to %sMbit %s-duplex\n",
+                      dev->name,
+                      np->speed100 ? "100" : "10",
+                      np->mii_if.full_duplex ? "full" : "half");
+       }
+       mdio_write(dev, np->phys[0], MII_BMCR, reg0);
+}
+
+
+static void tx_timeout(struct net_device *dev)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       void __iomem *ioaddr = np->base;
+       int old_debug;
+
+       printk(KERN_WARNING "%s: Transmit timed out, status %#8.8x, "
+              "resetting...\n", dev->name, (int) readl(ioaddr + IntrStatus));
+
+       /* Perhaps we should reinitialize the hardware here. */
+
+       /*
+        * Stop and restart the interface.
+        * Cheat and increase the debug level temporarily.
+        */
+       old_debug = debug;
+       debug = 2;
+       netdev_close(dev);
+       netdev_open(dev);
+       debug = old_debug;
+
+       /* Trigger an immediate transmit demand. */
+
+       dev->trans_start = jiffies; /* prevent tx timeout */
+       dev->stats.tx_errors++;
+       netif_wake_queue(dev);
+}
+
+
+/* Initialize the Rx and Tx rings, along with various 'dev' bits. */
+static void init_ring(struct net_device *dev)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       int i;
+
+       np->cur_rx = np->cur_tx = np->reap_tx = 0;
+       np->dirty_rx = np->dirty_tx = np->rx_done = np->tx_done = 0;
+
+       np->rx_buf_sz = (dev->mtu <= 1500 ? PKT_BUF_SZ : dev->mtu + 32);
+
+       /* Fill in the Rx buffers.  Handle allocation failure gracefully. */
+       for (i = 0; i < RX_RING_SIZE; i++) {
+               struct sk_buff *skb = netdev_alloc_skb(dev, np->rx_buf_sz);
+               np->rx_info[i].skb = skb;
+               if (skb == NULL)
+                       break;
+               np->rx_info[i].mapping = pci_map_single(np->pci_dev, skb->data, np->rx_buf_sz, PCI_DMA_FROMDEVICE);
+               /* Grrr, we cannot offset to correctly align the IP header. */
+               np->rx_ring[i].rxaddr = cpu_to_dma(np->rx_info[i].mapping | RxDescValid);
+       }
+       writew(i - 1, np->base + RxDescQIdx);
+       np->dirty_rx = (unsigned int)(i - RX_RING_SIZE);
+
+       /* Clear the remainder of the Rx buffer ring. */
+       for (  ; i < RX_RING_SIZE; i++) {
+               np->rx_ring[i].rxaddr = 0;
+               np->rx_info[i].skb = NULL;
+               np->rx_info[i].mapping = 0;
+       }
+       /* Mark the last entry as wrapping the ring. */
+       np->rx_ring[RX_RING_SIZE - 1].rxaddr |= cpu_to_dma(RxDescEndRing);
+
+       /* Clear the completion rings. */
+       for (i = 0; i < DONE_Q_SIZE; i++) {
+               np->rx_done_q[i].status = 0;
+               np->tx_done_q[i].status = 0;
+       }
+
+       for (i = 0; i < TX_RING_SIZE; i++)
+               memset(&np->tx_info[i], 0, sizeof(np->tx_info[i]));
+}
+
+
+static netdev_tx_t start_tx(struct sk_buff *skb, struct net_device *dev)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       unsigned int entry;
+       u32 status;
+       int i;
+
+       /*
+        * be cautious here, wrapping the queue has weird semantics
+        * and we may not have enough slots even when it seems we do.
+        */
+       if ((np->cur_tx - np->dirty_tx) + skb_num_frags(skb) * 2 > TX_RING_SIZE) {
+               netif_stop_queue(dev);
+               return NETDEV_TX_BUSY;
+       }
+
+#if defined(ZEROCOPY) && defined(HAS_BROKEN_FIRMWARE)
+       if (skb->ip_summed == CHECKSUM_PARTIAL) {
+               if (skb_padto(skb, (skb->len + PADDING_MASK) & ~PADDING_MASK))
+                       return NETDEV_TX_OK;
+       }
+#endif /* ZEROCOPY && HAS_BROKEN_FIRMWARE */
+
+       entry = np->cur_tx % TX_RING_SIZE;
+       for (i = 0; i < skb_num_frags(skb); i++) {
+               int wrap_ring = 0;
+               status = TxDescID;
+
+               if (i == 0) {
+                       np->tx_info[entry].skb = skb;
+                       status |= TxCRCEn;
+                       if (entry >= TX_RING_SIZE - skb_num_frags(skb)) {
+                               status |= TxRingWrap;
+                               wrap_ring = 1;
+                       }
+                       if (np->reap_tx) {
+                               status |= TxDescIntr;
+                               np->reap_tx = 0;
+                       }
+                       if (skb->ip_summed == CHECKSUM_PARTIAL) {
+                               status |= TxCalTCP;
+                               dev->stats.tx_compressed++;
+                       }
+                       status |= skb_first_frag_len(skb) | (skb_num_frags(skb) << 16);
+
+                       np->tx_info[entry].mapping =
+                               pci_map_single(np->pci_dev, skb->data, skb_first_frag_len(skb), PCI_DMA_TODEVICE);
+               } else {
+                       const skb_frag_t *this_frag = &skb_shinfo(skb)->frags[i - 1];
+                       status |= skb_frag_size(this_frag);
+                       np->tx_info[entry].mapping =
+                               pci_map_single(np->pci_dev,
+                                              skb_frag_address(this_frag),
+                                              skb_frag_size(this_frag),
+                                              PCI_DMA_TODEVICE);
+               }
+
+               np->tx_ring[entry].addr = cpu_to_dma(np->tx_info[entry].mapping);
+               np->tx_ring[entry].status = cpu_to_le32(status);
+               if (debug > 3)
+                       printk(KERN_DEBUG "%s: Tx #%d/#%d slot %d status %#8.8x.\n",
+                              dev->name, np->cur_tx, np->dirty_tx,
+                              entry, status);
+               if (wrap_ring) {
+                       np->tx_info[entry].used_slots = TX_RING_SIZE - entry;
+                       np->cur_tx += np->tx_info[entry].used_slots;
+                       entry = 0;
+               } else {
+                       np->tx_info[entry].used_slots = 1;
+                       np->cur_tx += np->tx_info[entry].used_slots;
+                       entry++;
+               }
+               /* scavenge the tx descriptors twice per TX_RING_SIZE */
+               if (np->cur_tx % (TX_RING_SIZE / 2) == 0)
+                       np->reap_tx = 1;
+       }
+
+       /* Non-x86: explicitly flush descriptor cache lines here. */
+       /* Ensure all descriptors are written back before the transmit is
+          initiated. - Jes */
+       wmb();
+
+       /* Update the producer index. */
+       writel(entry * (sizeof(starfire_tx_desc) / 8), np->base + TxProducerIdx);
+
+       /* 4 is arbitrary, but should be ok */
+       if ((np->cur_tx - np->dirty_tx) + 4 > TX_RING_SIZE)
+               netif_stop_queue(dev);
+
+       return NETDEV_TX_OK;
+}
+
+
+/* The interrupt handler does all of the Rx thread work and cleans up
+   after the Tx thread. */
+static irqreturn_t intr_handler(int irq, void *dev_instance)
+{
+       struct net_device *dev = dev_instance;
+       struct netdev_private *np = netdev_priv(dev);
+       void __iomem *ioaddr = np->base;
+       int boguscnt = max_interrupt_work;
+       int consumer;
+       int tx_status;
+       int handled = 0;
+
+       do {
+               u32 intr_status = readl(ioaddr + IntrClear);
+
+               if (debug > 4)
+                       printk(KERN_DEBUG "%s: Interrupt status %#8.8x.\n",
+                              dev->name, intr_status);
+
+               if (intr_status == 0 || intr_status == (u32) -1)
+                       break;
+
+               handled = 1;
+
+               if (intr_status & (IntrRxDone | IntrRxEmpty)) {
+                       u32 enable;
+
+                       if (likely(napi_schedule_prep(&np->napi))) {
+                               __napi_schedule(&np->napi);
+                               enable = readl(ioaddr + IntrEnable);
+                               enable &= ~(IntrRxDone | IntrRxEmpty);
+                               writel(enable, ioaddr + IntrEnable);
+                               /* flush PCI posting buffers */
+                               readl(ioaddr + IntrEnable);
+                       } else {
+                               /* Paranoia check */
+                               enable = readl(ioaddr + IntrEnable);
+                               if (enable & (IntrRxDone | IntrRxEmpty)) {
+                                       printk(KERN_INFO
+                                              "%s: interrupt while in poll!\n",
+                                              dev->name);
+                                       enable &= ~(IntrRxDone | IntrRxEmpty);
+                                       writel(enable, ioaddr + IntrEnable);
+                               }
+                       }
+               }
+
+               /* Scavenge the skbuff list based on the Tx-done queue.
+                  There are redundant checks here that may be cleaned up
+                  after the driver has proven to be reliable. */
+               consumer = readl(ioaddr + TxConsumerIdx);
+               if (debug > 3)
+                       printk(KERN_DEBUG "%s: Tx Consumer index is %d.\n",
+                              dev->name, consumer);
+
+               while ((tx_status = le32_to_cpu(np->tx_done_q[np->tx_done].status)) != 0) {
+                       if (debug > 3)
+                               printk(KERN_DEBUG "%s: Tx completion #%d entry %d is %#8.8x.\n",
+                                      dev->name, np->dirty_tx, np->tx_done, tx_status);
+                       if ((tx_status & 0xe0000000) == 0xa0000000) {
+                               dev->stats.tx_packets++;
+                       } else if ((tx_status & 0xe0000000) == 0x80000000) {
+                               u16 entry = (tx_status & 0x7fff) / sizeof(starfire_tx_desc);
+                               struct sk_buff *skb = np->tx_info[entry].skb;
+                               np->tx_info[entry].skb = NULL;
+                               pci_unmap_single(np->pci_dev,
+                                                np->tx_info[entry].mapping,
+                                                skb_first_frag_len(skb),
+                                                PCI_DMA_TODEVICE);
+                               np->tx_info[entry].mapping = 0;
+                               np->dirty_tx += np->tx_info[entry].used_slots;
+                               entry = (entry + np->tx_info[entry].used_slots) % TX_RING_SIZE;
+                               {
+                                       int i;
+                                       for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
+                                               pci_unmap_single(np->pci_dev,
+                                                                np->tx_info[entry].mapping,
+                                                                skb_frag_size(&skb_shinfo(skb)->frags[i]),
+                                                                PCI_DMA_TODEVICE);
+                                               np->dirty_tx++;
+                                               entry++;
+                                       }
+                               }
+
+                               dev_kfree_skb_irq(skb);
+                       }
+                       np->tx_done_q[np->tx_done].status = 0;
+                       np->tx_done = (np->tx_done + 1) % DONE_Q_SIZE;
+               }
+               writew(np->tx_done, ioaddr + CompletionQConsumerIdx + 2);
+
+               if (netif_queue_stopped(dev) &&
+                   (np->cur_tx - np->dirty_tx + 4 < TX_RING_SIZE)) {
+                       /* The ring is no longer full, wake the queue. */
+                       netif_wake_queue(dev);
+               }
+
+               /* Stats overflow */
+               if (intr_status & IntrStatsMax)
+                       get_stats(dev);
+
+               /* Media change interrupt. */
+               if (intr_status & IntrLinkChange)
+                       netdev_media_change(dev);
+
+               /* Abnormal error summary/uncommon events handlers. */
+               if (intr_status & IntrAbnormalSummary)
+                       netdev_error(dev, intr_status);
+
+               if (--boguscnt < 0) {
+                       if (debug > 1)
+                               printk(KERN_WARNING "%s: Too much work at interrupt, "
+                                      "status=%#8.8x.\n",
+                                      dev->name, intr_status);
+                       break;
+               }
+       } while (1);
+
+       if (debug > 4)
+               printk(KERN_DEBUG "%s: exiting interrupt, status=%#8.8x.\n",
+                      dev->name, (int) readl(ioaddr + IntrStatus));
+       return IRQ_RETVAL(handled);
+}
+
+
+/*
+ * This routine is logically part of the interrupt/poll handler, but separated
+ * for clarity and better register allocation.
+ */
+static int __netdev_rx(struct net_device *dev, int *quota)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       u32 desc_status;
+       int retcode = 0;
+
+       /* If EOP is set on the next entry, it's a new packet. Send it up. */
+       while ((desc_status = le32_to_cpu(np->rx_done_q[np->rx_done].status)) != 0) {
+               struct sk_buff *skb;
+               u16 pkt_len;
+               int entry;
+               rx_done_desc *desc = &np->rx_done_q[np->rx_done];
+
+               if (debug > 4)
+                       printk(KERN_DEBUG "  netdev_rx() status of %d was %#8.8x.\n", np->rx_done, desc_status);
+               if (!(desc_status & RxOK)) {
+                       /* There was an error. */
+                       if (debug > 2)
+                               printk(KERN_DEBUG "  netdev_rx() Rx error was %#8.8x.\n", desc_status);
+                       dev->stats.rx_errors++;
+                       if (desc_status & RxFIFOErr)
+                               dev->stats.rx_fifo_errors++;
+                       goto next_rx;
+               }
+
+               if (*quota <= 0) {      /* out of rx quota */
+                       retcode = 1;
+                       goto out;
+               }
+               (*quota)--;
+
+               pkt_len = desc_status;  /* Implicitly Truncate */
+               entry = (desc_status >> 16) & 0x7ff;
+
+               if (debug > 4)
+                       printk(KERN_DEBUG "  netdev_rx() normal Rx pkt length %d, quota %d.\n", pkt_len, *quota);
+               /* Check if the packet is long enough to accept without copying
+                  to a minimally-sized skbuff. */
+               if (pkt_len < rx_copybreak &&
+                   (skb = netdev_alloc_skb(dev, pkt_len + 2)) != NULL) {
+                       skb_reserve(skb, 2);    /* 16 byte align the IP header */
+                       pci_dma_sync_single_for_cpu(np->pci_dev,
+                                                   np->rx_info[entry].mapping,
+                                                   pkt_len, PCI_DMA_FROMDEVICE);
+                       skb_copy_to_linear_data(skb, np->rx_info[entry].skb->data, pkt_len);
+                       pci_dma_sync_single_for_device(np->pci_dev,
+                                                      np->rx_info[entry].mapping,
+                                                      pkt_len, PCI_DMA_FROMDEVICE);
+                       skb_put(skb, pkt_len);
+               } else {
+                       pci_unmap_single(np->pci_dev, np->rx_info[entry].mapping, np->rx_buf_sz, PCI_DMA_FROMDEVICE);
+                       skb = np->rx_info[entry].skb;
+                       skb_put(skb, pkt_len);
+                       np->rx_info[entry].skb = NULL;
+                       np->rx_info[entry].mapping = 0;
+               }
+#ifndef final_version                  /* Remove after testing. */
+               /* You will want this info for the initial debug. */
+               if (debug > 5) {
+                       printk(KERN_DEBUG "  Rx data %pM %pM %2.2x%2.2x.\n",
+                              skb->data, skb->data + 6,
+                              skb->data[12], skb->data[13]);
+               }
+#endif
+
+               skb->protocol = eth_type_trans(skb, dev);
+#ifdef VLAN_SUPPORT
+               if (debug > 4)
+                       printk(KERN_DEBUG "  netdev_rx() status2 of %d was %#4.4x.\n", np->rx_done, le16_to_cpu(desc->status2));
+#endif
+               if (le16_to_cpu(desc->status2) & 0x0100) {
+                       skb->ip_summed = CHECKSUM_UNNECESSARY;
+                       dev->stats.rx_compressed++;
+               }
+               /*
+                * This feature doesn't seem to be working, at least
+                * with the two firmware versions I have. If the GFP sees
+                * an IP fragment, it either ignores it completely, or reports
+                * "bad checksum" on it.
+                *
+                * Maybe I missed something -- corrections are welcome.
+                * Until then, the printk stays. :-) -Ion
+                */
+               else if (le16_to_cpu(desc->status2) & 0x0040) {
+                       skb->ip_summed = CHECKSUM_COMPLETE;
+                       skb->csum = le16_to_cpu(desc->csum);
+                       printk(KERN_DEBUG "%s: checksum_hw, status2 = %#x\n", dev->name, le16_to_cpu(desc->status2));
+               }
+#ifdef VLAN_SUPPORT
+               if (le16_to_cpu(desc->status2) & 0x0200) {
+                       u16 vlid = le16_to_cpu(desc->vlanid);
+
+                       if (debug > 4) {
+                               printk(KERN_DEBUG "  netdev_rx() vlanid = %d\n",
+                                      vlid);
+                       }
+                       __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlid);
+               }
+#endif /* VLAN_SUPPORT */
+               netif_receive_skb(skb);
+               dev->stats.rx_packets++;
+
+       next_rx:
+               np->cur_rx++;
+               desc->status = 0;
+               np->rx_done = (np->rx_done + 1) % DONE_Q_SIZE;
+       }
+
+       if (*quota == 0) {      /* out of rx quota */
+               retcode = 1;
+               goto out;
+       }
+       writew(np->rx_done, np->base + CompletionQConsumerIdx);
+
+ out:
+       refill_rx_ring(dev);
+       if (debug > 5)
+               printk(KERN_DEBUG "  exiting netdev_rx(): %d, status of %d was %#8.8x.\n",
+                      retcode, np->rx_done, desc_status);
+       return retcode;
+}
+
+static int netdev_poll(struct napi_struct *napi, int budget)
+{
+       struct netdev_private *np = container_of(napi, struct netdev_private, napi);
+       struct net_device *dev = np->dev;
+       u32 intr_status;
+       void __iomem *ioaddr = np->base;
+       int quota = budget;
+
+       do {
+               writel(IntrRxDone | IntrRxEmpty, ioaddr + IntrClear);
+
+               if (__netdev_rx(dev, &quota))
+                       goto out;
+
+               intr_status = readl(ioaddr + IntrStatus);
+       } while (intr_status & (IntrRxDone | IntrRxEmpty));
+
+       napi_complete(napi);
+       intr_status = readl(ioaddr + IntrEnable);
+       intr_status |= IntrRxDone | IntrRxEmpty;
+       writel(intr_status, ioaddr + IntrEnable);
+
+ out:
+       if (debug > 5)
+               printk(KERN_DEBUG "  exiting netdev_poll(): %d.\n",
+                      budget - quota);
+
+       /* Restart Rx engine if stopped. */
+       return budget - quota;
+}
+
+static void refill_rx_ring(struct net_device *dev)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       struct sk_buff *skb;
+       int entry = -1;
+
+       /* Refill the Rx ring buffers. */
+       for (; np->cur_rx - np->dirty_rx > 0; np->dirty_rx++) {
+               entry = np->dirty_rx % RX_RING_SIZE;
+               if (np->rx_info[entry].skb == NULL) {
+                       skb = netdev_alloc_skb(dev, np->rx_buf_sz);
+                       np->rx_info[entry].skb = skb;
+                       if (skb == NULL)
+                               break;  /* Better luck next round. */
+                       np->rx_info[entry].mapping =
+                               pci_map_single(np->pci_dev, skb->data, np->rx_buf_sz, PCI_DMA_FROMDEVICE);
+                       np->rx_ring[entry].rxaddr =
+                               cpu_to_dma(np->rx_info[entry].mapping | RxDescValid);
+               }
+               if (entry == RX_RING_SIZE - 1)
+                       np->rx_ring[entry].rxaddr |= cpu_to_dma(RxDescEndRing);
+       }
+       if (entry >= 0)
+               writew(entry, np->base + RxDescQIdx);
+}
+
+
+static void netdev_media_change(struct net_device *dev)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       void __iomem *ioaddr = np->base;
+       u16 reg0, reg1, reg4, reg5;
+       u32 new_tx_mode;
+       u32 new_intr_timer_ctrl;
+
+       /* reset status first */
+       mdio_read(dev, np->phys[0], MII_BMCR);
+       mdio_read(dev, np->phys[0], MII_BMSR);
+
+       reg0 = mdio_read(dev, np->phys[0], MII_BMCR);
+       reg1 = mdio_read(dev, np->phys[0], MII_BMSR);
+
+       if (reg1 & BMSR_LSTATUS) {
+               /* link is up */
+               if (reg0 & BMCR_ANENABLE) {
+                       /* autonegotiation is enabled */
+                       reg4 = mdio_read(dev, np->phys[0], MII_ADVERTISE);
+                       reg5 = mdio_read(dev, np->phys[0], MII_LPA);
+                       if (reg4 & ADVERTISE_100FULL && reg5 & LPA_100FULL) {
+                               np->speed100 = 1;
+                               np->mii_if.full_duplex = 1;
+                       } else if (reg4 & ADVERTISE_100HALF && reg5 & LPA_100HALF) {
+                               np->speed100 = 1;
+                               np->mii_if.full_duplex = 0;
+                       } else if (reg4 & ADVERTISE_10FULL && reg5 & LPA_10FULL) {
+                               np->speed100 = 0;
+                               np->mii_if.full_duplex = 1;
+                       } else {
+                               np->speed100 = 0;
+                               np->mii_if.full_duplex = 0;
+                       }
+               } else {
+                       /* autonegotiation is disabled */
+                       if (reg0 & BMCR_SPEED100)
+                               np->speed100 = 1;
+                       else
+                               np->speed100 = 0;
+                       if (reg0 & BMCR_FULLDPLX)
+                               np->mii_if.full_duplex = 1;
+                       else
+                               np->mii_if.full_duplex = 0;
+               }
+               netif_carrier_on(dev);
+               printk(KERN_DEBUG "%s: Link is up, running at %sMbit %s-duplex\n",
+                      dev->name,
+                      np->speed100 ? "100" : "10",
+                      np->mii_if.full_duplex ? "full" : "half");
+
+               new_tx_mode = np->tx_mode & ~FullDuplex;        /* duplex setting */
+               if (np->mii_if.full_duplex)
+                       new_tx_mode |= FullDuplex;
+               if (np->tx_mode != new_tx_mode) {
+                       np->tx_mode = new_tx_mode;
+                       writel(np->tx_mode | MiiSoftReset, ioaddr + TxMode);
+                       udelay(1000);
+                       writel(np->tx_mode, ioaddr + TxMode);
+               }
+
+               new_intr_timer_ctrl = np->intr_timer_ctrl & ~Timer10X;
+               if (np->speed100)
+                       new_intr_timer_ctrl |= Timer10X;
+               if (np->intr_timer_ctrl != new_intr_timer_ctrl) {
+                       np->intr_timer_ctrl = new_intr_timer_ctrl;
+                       writel(new_intr_timer_ctrl, ioaddr + IntrTimerCtrl);
+               }
+       } else {
+               netif_carrier_off(dev);
+               printk(KERN_DEBUG "%s: Link is down\n", dev->name);
+       }
+}
+
+
+static void netdev_error(struct net_device *dev, int intr_status)
+{
+       struct netdev_private *np = netdev_priv(dev);
+
+       /* Came close to underrunning the Tx FIFO, increase threshold. */
+       if (intr_status & IntrTxDataLow) {
+               if (np->tx_threshold <= PKT_BUF_SZ / 16) {
+                       writel(++np->tx_threshold, np->base + TxThreshold);
+                       printk(KERN_NOTICE "%s: PCI bus congestion, increasing Tx FIFO threshold to %d bytes\n",
+                              dev->name, np->tx_threshold * 16);
+               } else
+                       printk(KERN_WARNING "%s: PCI Tx underflow -- adapter is probably malfunctioning\n", dev->name);
+       }
+       if (intr_status & IntrRxGFPDead) {
+               dev->stats.rx_fifo_errors++;
+               dev->stats.rx_errors++;
+       }
+       if (intr_status & (IntrNoTxCsum | IntrDMAErr)) {
+               dev->stats.tx_fifo_errors++;
+               dev->stats.tx_errors++;
+       }
+       if ((intr_status & ~(IntrNormalMask | IntrAbnormalSummary | IntrLinkChange | IntrStatsMax | IntrTxDataLow | IntrRxGFPDead | IntrNoTxCsum | IntrPCIPad)) && debug)
+               printk(KERN_ERR "%s: Something Wicked happened! %#8.8x.\n",
+                      dev->name, intr_status);
+}
+
+
+static struct net_device_stats *get_stats(struct net_device *dev)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       void __iomem *ioaddr = np->base;
+
+       /* This adapter architecture needs no SMP locks. */
+       dev->stats.tx_bytes = readl(ioaddr + 0x57010);
+       dev->stats.rx_bytes = readl(ioaddr + 0x57044);
+       dev->stats.tx_packets = readl(ioaddr + 0x57000);
+       dev->stats.tx_aborted_errors =
+               readl(ioaddr + 0x57024) + readl(ioaddr + 0x57028);
+       dev->stats.tx_window_errors = readl(ioaddr + 0x57018);
+       dev->stats.collisions =
+               readl(ioaddr + 0x57004) + readl(ioaddr + 0x57008);
+
+       /* The chip only need report frame silently dropped. */
+       dev->stats.rx_dropped += readw(ioaddr + RxDMAStatus);
+       writew(0, ioaddr + RxDMAStatus);
+       dev->stats.rx_crc_errors = readl(ioaddr + 0x5703C);
+       dev->stats.rx_frame_errors = readl(ioaddr + 0x57040);
+       dev->stats.rx_length_errors = readl(ioaddr + 0x57058);
+       dev->stats.rx_missed_errors = readl(ioaddr + 0x5707C);
+
+       return &dev->stats;
+}
+
+#ifdef VLAN_SUPPORT
+static u32 set_vlan_mode(struct netdev_private *np)
+{
+       u32 ret = VlanMode;
+       u16 vid;
+       void __iomem *filter_addr = np->base + HashTable + 8;
+       int vlan_count = 0;
+
+       for_each_set_bit(vid, np->active_vlans, VLAN_N_VID) {
+               if (vlan_count == 32)
+                       break;
+               writew(vid, filter_addr);
+               filter_addr += 16;
+               vlan_count++;
+       }
+       if (vlan_count == 32) {
+               ret |= PerfectFilterVlan;
+               while (vlan_count < 32) {
+                       writew(0, filter_addr);
+                       filter_addr += 16;
+                       vlan_count++;
+               }
+       }
+       return ret;
+}
+#endif /* VLAN_SUPPORT */
+
+static void set_rx_mode(struct net_device *dev)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       void __iomem *ioaddr = np->base;
+       u32 rx_mode = MinVLANPrio;
+       struct netdev_hw_addr *ha;
+       int i;
+
+#ifdef VLAN_SUPPORT
+       rx_mode |= set_vlan_mode(np);
+#endif /* VLAN_SUPPORT */
+
+       if (dev->flags & IFF_PROMISC) { /* Set promiscuous. */
+               rx_mode |= AcceptAll;
+       } else if ((netdev_mc_count(dev) > multicast_filter_limit) ||
+                  (dev->flags & IFF_ALLMULTI)) {
+               /* Too many to match, or accept all multicasts. */
+               rx_mode |= AcceptBroadcast|AcceptAllMulticast|PerfectFilter;
+       } else if (netdev_mc_count(dev) <= 14) {
+               /* Use the 16 element perfect filter, skip first two entries. */
+               void __iomem *filter_addr = ioaddr + PerfFilterTable + 2 * 16;
+               __be16 *eaddrs;
+               netdev_for_each_mc_addr(ha, dev) {
+                       eaddrs = (__be16 *) ha->addr;
+                       writew(be16_to_cpu(eaddrs[2]), filter_addr); filter_addr += 4;
+                       writew(be16_to_cpu(eaddrs[1]), filter_addr); filter_addr += 4;
+                       writew(be16_to_cpu(eaddrs[0]), filter_addr); filter_addr += 8;
+               }
+               eaddrs = (__be16 *)dev->dev_addr;
+               i = netdev_mc_count(dev) + 2;
+               while (i++ < 16) {
+                       writew(be16_to_cpu(eaddrs[0]), filter_addr); filter_addr += 4;
+                       writew(be16_to_cpu(eaddrs[1]), filter_addr); filter_addr += 4;
+                       writew(be16_to_cpu(eaddrs[2]), filter_addr); filter_addr += 8;
+               }
+               rx_mode |= AcceptBroadcast|PerfectFilter;
+       } else {
+               /* Must use a multicast hash table. */
+               void __iomem *filter_addr;
+               __be16 *eaddrs;
+               __le16 mc_filter[32] __attribute__ ((aligned(sizeof(long))));   /* Multicast hash filter */
+
+               memset(mc_filter, 0, sizeof(mc_filter));
+               netdev_for_each_mc_addr(ha, dev) {
+                       /* The chip uses the upper 9 CRC bits
+                          as index into the hash table */
+                       int bit_nr = ether_crc_le(ETH_ALEN, ha->addr) >> 23;
+                       __le32 *fptr = (__le32 *) &mc_filter[(bit_nr >> 4) & ~1];
+
+                       *fptr |= cpu_to_le32(1 << (bit_nr & 31));
+               }
+               /* Clear the perfect filter list, skip first two entries. */
+               filter_addr = ioaddr + PerfFilterTable + 2 * 16;
+               eaddrs = (__be16 *)dev->dev_addr;
+               for (i = 2; i < 16; i++) {
+                       writew(be16_to_cpu(eaddrs[0]), filter_addr); filter_addr += 4;
+                       writew(be16_to_cpu(eaddrs[1]), filter_addr); filter_addr += 4;
+                       writew(be16_to_cpu(eaddrs[2]), filter_addr); filter_addr += 8;
+               }
+               for (filter_addr = ioaddr + HashTable, i = 0; i < 32; filter_addr+= 16, i++)
+                       writew(mc_filter[i], filter_addr);
+               rx_mode |= AcceptBroadcast|PerfectFilter|HashFilter;
+       }
+       writel(rx_mode, ioaddr + RxFilterMode);
+}
+
+static int check_if_running(struct net_device *dev)
+{
+       if (!netif_running(dev))
+               return -EINVAL;
+       return 0;
+}
+
+static void get_drvinfo(struct net_device *dev, struct ethtool_drvinfo *info)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       strlcpy(info->driver, DRV_NAME, sizeof(info->driver));
+       strlcpy(info->version, DRV_VERSION, sizeof(info->version));
+       strlcpy(info->bus_info, pci_name(np->pci_dev), sizeof(info->bus_info));
+}
+
+static int get_settings(struct net_device *dev, struct ethtool_cmd *ecmd)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       spin_lock_irq(&np->lock);
+       mii_ethtool_gset(&np->mii_if, ecmd);
+       spin_unlock_irq(&np->lock);
+       return 0;
+}
+
+static int set_settings(struct net_device *dev, struct ethtool_cmd *ecmd)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       int res;
+       spin_lock_irq(&np->lock);
+       res = mii_ethtool_sset(&np->mii_if, ecmd);
+       spin_unlock_irq(&np->lock);
+       check_duplex(dev);
+       return res;
+}
+
+static int nway_reset(struct net_device *dev)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       return mii_nway_restart(&np->mii_if);
+}
+
+static u32 get_link(struct net_device *dev)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       return mii_link_ok(&np->mii_if);
+}
+
+static u32 get_msglevel(struct net_device *dev)
+{
+       return debug;
+}
+
+static void set_msglevel(struct net_device *dev, u32 val)
+{
+       debug = val;
+}
+
+static const struct ethtool_ops ethtool_ops = {
+       .begin = check_if_running,
+       .get_drvinfo = get_drvinfo,
+       .get_settings = get_settings,
+       .set_settings = set_settings,
+       .nway_reset = nway_reset,
+       .get_link = get_link,
+       .get_msglevel = get_msglevel,
+       .set_msglevel = set_msglevel,
+};
+
+static int netdev_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       struct mii_ioctl_data *data = if_mii(rq);
+       int rc;
+
+       if (!netif_running(dev))
+               return -EINVAL;
+
+       spin_lock_irq(&np->lock);
+       rc = generic_mii_ioctl(&np->mii_if, data, cmd, NULL);
+       spin_unlock_irq(&np->lock);
+
+       if ((cmd == SIOCSMIIREG) && (data->phy_id == np->phys[0]))
+               check_duplex(dev);
+
+       return rc;
+}
+
+static int netdev_close(struct net_device *dev)
+{
+       struct netdev_private *np = netdev_priv(dev);
+       void __iomem *ioaddr = np->base;
+       int i;
+
+       netif_stop_queue(dev);
+
+       napi_disable(&np->napi);
+
+       if (debug > 1) {
+               printk(KERN_DEBUG "%s: Shutting down ethercard, Intr status %#8.8x.\n",
+                          dev->name, (int) readl(ioaddr + IntrStatus));
+               printk(KERN_DEBUG "%s: Queue pointers were Tx %d / %d, Rx %d / %d.\n",
+                      dev->name, np->cur_tx, np->dirty_tx,
+                      np->cur_rx, np->dirty_rx);
+       }
+
+       /* Disable interrupts by clearing the interrupt mask. */
+       writel(0, ioaddr + IntrEnable);
+
+       /* Stop the chip's Tx and Rx processes. */
+       writel(0, ioaddr + GenCtrl);
+       readl(ioaddr + GenCtrl);
+
+       if (debug > 5) {
+               printk(KERN_DEBUG"  Tx ring at %#llx:\n",
+                      (long long) np->tx_ring_dma);
+               for (i = 0; i < 8 /* TX_RING_SIZE is huge! */; i++)
+                       printk(KERN_DEBUG " #%d desc. %#8.8x %#llx -> %#8.8x.\n",
+                              i, le32_to_cpu(np->tx_ring[i].status),
+                              (long long) dma_to_cpu(np->tx_ring[i].addr),
+                              le32_to_cpu(np->tx_done_q[i].status));
+               printk(KERN_DEBUG "  Rx ring at %#llx -> %p:\n",
+                      (long long) np->rx_ring_dma, np->rx_done_q);
+               if (np->rx_done_q)
+                       for (i = 0; i < 8 /* RX_RING_SIZE */; i++) {
+                               printk(KERN_DEBUG " #%d desc. %#llx -> %#8.8x\n",
+                                      i, (long long) dma_to_cpu(np->rx_ring[i].rxaddr), le32_to_cpu(np->rx_done_q[i].status));
+               }
+       }
+
+       free_irq(np->pci_dev->irq, dev);
+
+       /* Free all the skbuffs in the Rx queue. */
+       for (i = 0; i < RX_RING_SIZE; i++) {
+               np->rx_ring[i].rxaddr = cpu_to_dma(0xBADF00D0); /* An invalid address. */
+               if (np->rx_info[i].skb != NULL) {
+                       pci_unmap_single(np->pci_dev, np->rx_info[i].mapping, np->rx_buf_sz, PCI_DMA_FROMDEVICE);
+                       dev_kfree_skb(np->rx_info[i].skb);
+               }
+               np->rx_info[i].skb = NULL;
+               np->rx_info[i].mapping = 0;
+       }
+       for (i = 0; i < TX_RING_SIZE; i++) {
+               struct sk_buff *skb = np->tx_info[i].skb;
+               if (skb == NULL)
+                       continue;
+               pci_unmap_single(np->pci_dev,
+                                np->tx_info[i].mapping,
+                                skb_first_frag_len(skb), PCI_DMA_TODEVICE);
+               np->tx_info[i].mapping = 0;
+               dev_kfree_skb(skb);
+               np->tx_info[i].skb = NULL;
+       }
+
+       return 0;
+}
+
+#ifdef CONFIG_PM
+static int starfire_suspend(struct pci_dev *pdev, pm_message_t state)
+{
+       struct net_device *dev = pci_get_drvdata(pdev);
+
+       if (netif_running(dev)) {
+               netif_device_detach(dev);
+               netdev_close(dev);
+       }
+
+       pci_save_state(pdev);
+       pci_set_power_state(pdev, pci_choose_state(pdev,state));
+
+       return 0;
+}
+
+static int starfire_resume(struct pci_dev *pdev)
+{
+       struct net_device *dev = pci_get_drvdata(pdev);
+
+       pci_set_power_state(pdev, PCI_D0);
+       pci_restore_state(pdev);
+
+       if (netif_running(dev)) {
+               netdev_open(dev);
+               netif_device_attach(dev);
+       }
+
+       return 0;
+}
+#endif /* CONFIG_PM */
+
+
+static void starfire_remove_one(struct pci_dev *pdev)
+{
+       struct net_device *dev = pci_get_drvdata(pdev);
+       struct netdev_private *np = netdev_priv(dev);
+
+       BUG_ON(!dev);
+
+       unregister_netdev(dev);
+
+       if (np->queue_mem)
+               pci_free_consistent(pdev, np->queue_mem_size, np->queue_mem, np->queue_mem_dma);
+
+
+       /* XXX: add wakeup code -- requires firmware for MagicPacket */
+       pci_set_power_state(pdev, PCI_D3hot);   /* go to sleep in D3 mode */
+       pci_disable_device(pdev);
+
+       iounmap(np->base);
+       pci_release_regions(pdev);
+
+       free_netdev(dev);                       /* Will also free np!! */
+}
+
+
+static struct pci_driver starfire_driver = {
+       .name           = DRV_NAME,
+       .probe          = starfire_init_one,
+       .remove         = starfire_remove_one,
+#ifdef CONFIG_PM
+       .suspend        = starfire_suspend,
+       .resume         = starfire_resume,
+#endif /* CONFIG_PM */
+       .id_table       = starfire_pci_tbl,
+};
+
+
+static int __init starfire_init (void)
+{
+/* when a module, this is printed whether or not devices are found in probe */
+#ifdef MODULE
+       printk(version);
+
+       printk(KERN_INFO DRV_NAME ": polling (NAPI) enabled\n");
+#endif
+
+       BUILD_BUG_ON(sizeof(dma_addr_t) != sizeof(netdrv_addr_t));
+
+       return pci_register_driver(&starfire_driver);
+}
+
+
+static void __exit starfire_cleanup (void)
+{
+       pci_unregister_driver (&starfire_driver);
+}
+
+
+module_init(starfire_init);
+module_exit(starfire_cleanup);
+
+
+/*
+ * Local variables:
+ *  c-basic-offset: 8
+ *  tab-width: 8
+ * End:
+ */