net: trans_start cleanups

[deliverable/linux.git] / drivers / net / sunqe.c

/* sunqe.c: Sparc QuadEthernet 10baseT SBUS card driver.
 *          Once again I am out to prove that every ethernet
 *          controller out there can be most efficiently programmed
 *          if you make it look like a LANCE.
 *
 * Copyright (C) 1996, 1999, 2003, 2006, 2008 David S. Miller (davem@davemloft.net)
 */

#include <linux/module.h>
#include <linux/kernel.h>
#include <linux/types.h>
#include <linux/errno.h>
#include <linux/fcntl.h>
#include <linux/interrupt.h>
#include <linux/ioport.h>
#include <linux/in.h>
#include <linux/slab.h>
#include <linux/string.h>
#include <linux/delay.h>
#include <linux/init.h>
#include <linux/crc32.h>
#include <linux/netdevice.h>
#include <linux/etherdevice.h>
#include <linux/skbuff.h>
#include <linux/ethtool.h>
#include <linux/bitops.h>
#include <linux/dma-mapping.h>
#include <linux/of.h>
#include <linux/of_device.h>

#include <asm/system.h>
#include <asm/io.h>
#include <asm/dma.h>
#include <asm/byteorder.h>
#include <asm/idprom.h>
#include <asm/openprom.h>
#include <asm/oplib.h>
#include <asm/auxio.h>
#include <asm/pgtable.h>
#include <asm/irq.h>

#include "sunqe.h"

#define DRV_NAME	"sunqe"
#define DRV_VERSION	"4.1"
#define DRV_RELDATE	"August 27, 2008"
#define DRV_AUTHOR	"David S. Miller (davem@davemloft.net)"

static char version[] =
	DRV_NAME ".c:v" DRV_VERSION " " DRV_RELDATE " " DRV_AUTHOR "\n";

MODULE_VERSION(DRV_VERSION);
MODULE_AUTHOR(DRV_AUTHOR);
MODULE_DESCRIPTION("Sun QuadEthernet 10baseT SBUS card driver");
MODULE_LICENSE("GPL");

static struct sunqec *root_qec_dev;

static void qe_set_multicast(struct net_device *dev);

#define QEC_RESET_TRIES 200

static inline int qec_global_reset(void __iomem *gregs)
{
	int tries = QEC_RESET_TRIES;

	sbus_writel(GLOB_CTRL_RESET, gregs + GLOB_CTRL);
	while (--tries) {
		u32 tmp = sbus_readl(gregs + GLOB_CTRL);
		if (tmp & GLOB_CTRL_RESET) {
			udelay(20);
			continue;
		}
		break;
	}
	if (tries)
		return 0;
	printk(KERN_ERR "QuadEther: AIEEE cannot reset the QEC!\n");
	return -1;
}

#define MACE_RESET_RETRIES 200
#define QE_RESET_RETRIES   200

static inline int qe_stop(struct sunqe *qep)
{
	void __iomem *cregs = qep->qcregs;
	void __iomem *mregs = qep->mregs;
	int tries;

	/* Reset the MACE, then the QEC channel. */
	sbus_writeb(MREGS_BCONFIG_RESET, mregs + MREGS_BCONFIG);
	tries = MACE_RESET_RETRIES;
	while (--tries) {
		u8 tmp = sbus_readb(mregs + MREGS_BCONFIG);
		if (tmp & MREGS_BCONFIG_RESET) {
			udelay(20);
			continue;
		}
		break;
	}
	if (!tries) {
		printk(KERN_ERR "QuadEther: AIEEE cannot reset the MACE!\n");
		return -1;
	}

	sbus_writel(CREG_CTRL_RESET, cregs + CREG_CTRL);
	tries = QE_RESET_RETRIES;
	while (--tries) {
		u32 tmp = sbus_readl(cregs + CREG_CTRL);
		if (tmp & CREG_CTRL_RESET) {
			udelay(20);
			continue;
		}
		break;
	}
	if (!tries) {
		printk(KERN_ERR "QuadEther: Cannot reset QE channel!\n");
		return -1;
	}
	return 0;
}

static void qe_init_rings(struct sunqe *qep)
{
	struct qe_init_block *qb = qep->qe_block;
	struct sunqe_buffers *qbufs = qep->buffers;
	__u32 qbufs_dvma = qep->buffers_dvma;
	int i;

	qep->rx_new = qep->rx_old = qep->tx_new = qep->tx_old = 0;
	memset(qb, 0, sizeof(struct qe_init_block));
	memset(qbufs, 0, sizeof(struct sunqe_buffers));
	for (i = 0; i < RX_RING_SIZE; i++) {
		qb->qe_rxd[i].rx_addr = qbufs_dvma + qebuf_offset(rx_buf, i);
		qb->qe_rxd[i].rx_flags =
			(RXD_OWN | ((RXD_PKT_SZ) & RXD_LENGTH));
	}
}

static int qe_init(struct sunqe *qep, int from_irq)
{
	struct sunqec *qecp = qep->parent;
	void __iomem *cregs = qep->qcregs;
	void __iomem *mregs = qep->mregs;
	void __iomem *gregs = qecp->gregs;
	unsigned char *e = &qep->dev->dev_addr[0];
	u32 tmp;
	int i;

	/* Shut it up. */
	if (qe_stop(qep))
		return -EAGAIN;

	/* Setup initial rx/tx init block pointers. */
	sbus_writel(qep->qblock_dvma + qib_offset(qe_rxd, 0), cregs + CREG_RXDS);
	sbus_writel(qep->qblock_dvma + qib_offset(qe_txd, 0), cregs + CREG_TXDS);

	/* Enable/mask the various irq's. */
	sbus_writel(0, cregs + CREG_RIMASK);
	sbus_writel(1, cregs + CREG_TIMASK);

	sbus_writel(0, cregs + CREG_QMASK);
	sbus_writel(CREG_MMASK_RXCOLL, cregs + CREG_MMASK);

	/* Setup the FIFO pointers into QEC local memory. */
	tmp = qep->channel * sbus_readl(gregs + GLOB_MSIZE);
	sbus_writel(tmp, cregs + CREG_RXRBUFPTR);
	sbus_writel(tmp, cregs + CREG_RXWBUFPTR);

	tmp = sbus_readl(cregs + CREG_RXRBUFPTR) +
		sbus_readl(gregs + GLOB_RSIZE);
	sbus_writel(tmp, cregs + CREG_TXRBUFPTR);
	sbus_writel(tmp, cregs + CREG_TXWBUFPTR);

	/* Clear the channel collision counter. */
	sbus_writel(0, cregs + CREG_CCNT);

	/* For 10baseT, inter frame space nor throttle seems to be necessary. */
	sbus_writel(0, cregs + CREG_PIPG);

	/* Now dork with the AMD MACE. */
	sbus_writeb(MREGS_PHYCONFIG_AUTO, mregs + MREGS_PHYCONFIG);
	sbus_writeb(MREGS_TXFCNTL_AUTOPAD, mregs + MREGS_TXFCNTL);
	sbus_writeb(0, mregs + MREGS_RXFCNTL);

	/* The QEC dma's the rx'd packets from local memory out to main memory,
	 * and therefore it interrupts when the packet reception is "complete".
	 * So don't listen for the MACE talking about it.
	 */
	sbus_writeb(MREGS_IMASK_COLL | MREGS_IMASK_RXIRQ, mregs + MREGS_IMASK);
	sbus_writeb(MREGS_BCONFIG_BSWAP | MREGS_BCONFIG_64TS, mregs + MREGS_BCONFIG);
	sbus_writeb((MREGS_FCONFIG_TXF16 | MREGS_FCONFIG_RXF32 |
		     MREGS_FCONFIG_RFWU | MREGS_FCONFIG_TFWU),
		    mregs + MREGS_FCONFIG);

	/* Only usable interface on QuadEther is twisted pair. */
	sbus_writeb(MREGS_PLSCONFIG_TP, mregs + MREGS_PLSCONFIG);

	/* Tell MACE we are changing the ether address. */
	sbus_writeb(MREGS_IACONFIG_ACHNGE | MREGS_IACONFIG_PARESET,
		    mregs + MREGS_IACONFIG);
	while ((sbus_readb(mregs + MREGS_IACONFIG) & MREGS_IACONFIG_ACHNGE) != 0)
		barrier();
	sbus_writeb(e[0], mregs + MREGS_ETHADDR);
	sbus_writeb(e[1], mregs + MREGS_ETHADDR);
	sbus_writeb(e[2], mregs + MREGS_ETHADDR);
	sbus_writeb(e[3], mregs + MREGS_ETHADDR);
	sbus_writeb(e[4], mregs + MREGS_ETHADDR);
	sbus_writeb(e[5], mregs + MREGS_ETHADDR);

	/* Clear out the address filter. */
	sbus_writeb(MREGS_IACONFIG_ACHNGE | MREGS_IACONFIG_LARESET,
		    mregs + MREGS_IACONFIG);
	while ((sbus_readb(mregs + MREGS_IACONFIG) & MREGS_IACONFIG_ACHNGE) != 0)
		barrier();
	for (i = 0; i < 8; i++)
		sbus_writeb(0, mregs + MREGS_FILTER);

	/* Address changes are now complete. */
	sbus_writeb(0, mregs + MREGS_IACONFIG);

	qe_init_rings(qep);

	/* Wait a little bit for the link to come up... */
	mdelay(5);
	if (!(sbus_readb(mregs + MREGS_PHYCONFIG) & MREGS_PHYCONFIG_LTESTDIS)) {
		int tries = 50;

		while (--tries) {
			u8 tmp;

			mdelay(5);
			barrier();
			tmp = sbus_readb(mregs + MREGS_PHYCONFIG);
			if ((tmp & MREGS_PHYCONFIG_LSTAT) != 0)
				break;
		}
		if (tries == 0)
			printk(KERN_NOTICE "%s: Warning, link state is down.\n", qep->dev->name);
	}

	/* Missed packet counter is cleared on a read. */
	sbus_readb(mregs + MREGS_MPCNT);

	/* Reload multicast information, this will enable the receiver
	 * and transmitter.
	 */
	qe_set_multicast(qep->dev);

	/* QEC should now start to show interrupts. */
	return 0;
}

/* Grrr, certain error conditions completely lock up the AMD MACE,
 * so when we get these we _must_ reset the chip.
 */
static int qe_is_bolixed(struct sunqe *qep, u32 qe_status)
{
	struct net_device *dev = qep->dev;
	int mace_hwbug_workaround = 0;

	if (qe_status & CREG_STAT_EDEFER) {
		printk(KERN_ERR "%s: Excessive transmit defers.\n", dev->name);
		dev->stats.tx_errors++;
	}

	if (qe_status & CREG_STAT_CLOSS) {
		printk(KERN_ERR "%s: Carrier lost, link down?\n", dev->name);
		dev->stats.tx_errors++;
		dev->stats.tx_carrier_errors++;
	}

	if (qe_status & CREG_STAT_ERETRIES) {
		printk(KERN_ERR "%s: Excessive transmit retries (more than 16).\n", dev->name);
		dev->stats.tx_errors++;
		mace_hwbug_workaround = 1;
	}

	if (qe_status & CREG_STAT_LCOLL) {
		printk(KERN_ERR "%s: Late transmit collision.\n", dev->name);
		dev->stats.tx_errors++;
		dev->stats.collisions++;
		mace_hwbug_workaround = 1;
	}

	if (qe_status & CREG_STAT_FUFLOW) {
		printk(KERN_ERR "%s: Transmit fifo underflow, driver bug.\n", dev->name);
		dev->stats.tx_errors++;
		mace_hwbug_workaround = 1;
	}

	if (qe_status & CREG_STAT_JERROR) {
		printk(KERN_ERR "%s: Jabber error.\n", dev->name);
	}

	if (qe_status & CREG_STAT_BERROR) {
		printk(KERN_ERR "%s: Babble error.\n", dev->name);
	}

	if (qe_status & CREG_STAT_CCOFLOW) {
		dev->stats.tx_errors += 256;
		dev->stats.collisions += 256;
	}

	if (qe_status & CREG_STAT_TXDERROR) {
		printk(KERN_ERR "%s: Transmit descriptor is bogus, driver bug.\n", dev->name);
		dev->stats.tx_errors++;
		dev->stats.tx_aborted_errors++;
		mace_hwbug_workaround = 1;
	}

	if (qe_status & CREG_STAT_TXLERR) {
		printk(KERN_ERR "%s: Transmit late error.\n", dev->name);
		dev->stats.tx_errors++;
		mace_hwbug_workaround = 1;
	}

	if (qe_status & CREG_STAT_TXPERR) {
		printk(KERN_ERR "%s: Transmit DMA parity error.\n", dev->name);
		dev->stats.tx_errors++;
		dev->stats.tx_aborted_errors++;
		mace_hwbug_workaround = 1;
	}

	if (qe_status & CREG_STAT_TXSERR) {
		printk(KERN_ERR "%s: Transmit DMA sbus error ack.\n", dev->name);
		dev->stats.tx_errors++;
		dev->stats.tx_aborted_errors++;
		mace_hwbug_workaround = 1;
	}

	if (qe_status & CREG_STAT_RCCOFLOW) {
		dev->stats.rx_errors += 256;
		dev->stats.collisions += 256;
	}

	if (qe_status & CREG_STAT_RUOFLOW) {
		dev->stats.rx_errors += 256;
		dev->stats.rx_over_errors += 256;
	}

	if (qe_status & CREG_STAT_MCOFLOW) {
		dev->stats.rx_errors += 256;
		dev->stats.rx_missed_errors += 256;
	}

	if (qe_status & CREG_STAT_RXFOFLOW) {
		printk(KERN_ERR "%s: Receive fifo overflow.\n", dev->name);
		dev->stats.rx_errors++;
		dev->stats.rx_over_errors++;
	}

	if (qe_status & CREG_STAT_RLCOLL) {
		printk(KERN_ERR "%s: Late receive collision.\n", dev->name);
		dev->stats.rx_errors++;
		dev->stats.collisions++;
	}

	if (qe_status & CREG_STAT_FCOFLOW) {
		dev->stats.rx_errors += 256;
		dev->stats.rx_frame_errors += 256;
	}

	if (qe_status & CREG_STAT_CECOFLOW) {
		dev->stats.rx_errors += 256;
		dev->stats.rx_crc_errors += 256;
	}

	if (qe_status & CREG_STAT_RXDROP) {
		printk(KERN_ERR "%s: Receive packet dropped.\n", dev->name);
		dev->stats.rx_errors++;
		dev->stats.rx_dropped++;
		dev->stats.rx_missed_errors++;
	}

	if (qe_status & CREG_STAT_RXSMALL) {
		printk(KERN_ERR "%s: Receive buffer too small, driver bug.\n", dev->name);
		dev->stats.rx_errors++;
		dev->stats.rx_length_errors++;
	}

	if (qe_status & CREG_STAT_RXLERR) {
		printk(KERN_ERR "%s: Receive late error.\n", dev->name);
		dev->stats.rx_errors++;
		mace_hwbug_workaround = 1;
	}

	if (qe_status & CREG_STAT_RXPERR) {
		printk(KERN_ERR "%s: Receive DMA parity error.\n", dev->name);
		dev->stats.rx_errors++;
		dev->stats.rx_missed_errors++;
		mace_hwbug_workaround = 1;
	}

	if (qe_status & CREG_STAT_RXSERR) {
		printk(KERN_ERR "%s: Receive DMA sbus error ack.\n", dev->name);
		dev->stats.rx_errors++;
		dev->stats.rx_missed_errors++;
		mace_hwbug_workaround = 1;
	}

	if (mace_hwbug_workaround)
		qe_init(qep, 1);
	return mace_hwbug_workaround;
}

/* Per-QE receive interrupt service routine.  Just like on the happy meal
 * we receive directly into skb's with a small packet copy water mark.
 */
static void qe_rx(struct sunqe *qep)
{
	struct qe_rxd *rxbase = &qep->qe_block->qe_rxd[0];
	struct net_device *dev = qep->dev;
	struct qe_rxd *this;
	struct sunqe_buffers *qbufs = qep->buffers;
	__u32 qbufs_dvma = qep->buffers_dvma;
	int elem = qep->rx_new, drops = 0;
	u32 flags;

	this = &rxbase[elem];
	while (!((flags = this->rx_flags) & RXD_OWN)) {
		struct sk_buff *skb;
		unsigned char *this_qbuf =
			&qbufs->rx_buf[elem & (RX_RING_SIZE - 1)][0];
		__u32 this_qbuf_dvma = qbufs_dvma +
			qebuf_offset(rx_buf, (elem & (RX_RING_SIZE - 1)));
		struct qe_rxd *end_rxd =
			&rxbase[(elem+RX_RING_SIZE)&(RX_RING_MAXSIZE-1)];
		int len = (flags & RXD_LENGTH) - 4;  /* QE adds ether FCS size to len */

		/* Check for errors. */
		if (len < ETH_ZLEN) {
			dev->stats.rx_errors++;
			dev->stats.rx_length_errors++;
			dev->stats.rx_dropped++;
		} else {
			skb = dev_alloc_skb(len + 2);
			if (skb == NULL) {
				drops++;
				dev->stats.rx_dropped++;
			} else {
				skb_reserve(skb, 2);
				skb_put(skb, len);
				skb_copy_to_linear_data(skb, (unsigned char *) this_qbuf,
						 len);
				skb->protocol = eth_type_trans(skb, qep->dev);
				netif_rx(skb);
				dev->stats.rx_packets++;
				dev->stats.rx_bytes += len;
			}
		}
		end_rxd->rx_addr = this_qbuf_dvma;
		end_rxd->rx_flags = (RXD_OWN | ((RXD_PKT_SZ) & RXD_LENGTH));

		elem = NEXT_RX(elem);
		this = &rxbase[elem];
	}
	qep->rx_new = elem;
	if (drops)
		printk(KERN_NOTICE "%s: Memory squeeze, deferring packet.\n", qep->dev->name);
}

static void qe_tx_reclaim(struct sunqe *qep);

/* Interrupts for all QE's get filtered out via the QEC master controller,
 * so we just run through each qe and check to see who is signaling
 * and thus needs to be serviced.
 */
static irqreturn_t qec_interrupt(int irq, void *dev_id)
{
	struct sunqec *qecp = dev_id;
	u32 qec_status;
	int channel = 0;

	/* Latch the status now. */
	qec_status = sbus_readl(qecp->gregs + GLOB_STAT);
	while (channel < 4) {
		if (qec_status & 0xf) {
			struct sunqe *qep = qecp->qes[channel];
			u32 qe_status;

			qe_status = sbus_readl(qep->qcregs + CREG_STAT);
			if (qe_status & CREG_STAT_ERRORS) {
				if (qe_is_bolixed(qep, qe_status))
					goto next;
			}
			if (qe_status & CREG_STAT_RXIRQ)
				qe_rx(qep);
			if (netif_queue_stopped(qep->dev) &&
			    (qe_status & CREG_STAT_TXIRQ)) {
				spin_lock(&qep->lock);
				qe_tx_reclaim(qep);
				if (TX_BUFFS_AVAIL(qep) > 0) {
					/* Wake net queue and return to
					 * lazy tx reclaim.
					 */
					netif_wake_queue(qep->dev);
					sbus_writel(1, qep->qcregs + CREG_TIMASK);
				}
				spin_unlock(&qep->lock);
			}
	next:
			;
		}
		qec_status >>= 4;
		channel++;
	}

	return IRQ_HANDLED;
}

static int qe_open(struct net_device *dev)
{
	struct sunqe *qep = netdev_priv(dev);

	qep->mconfig = (MREGS_MCONFIG_TXENAB |
			MREGS_MCONFIG_RXENAB |
			MREGS_MCONFIG_MBAENAB);
	return qe_init(qep, 0);
}

static int qe_close(struct net_device *dev)
{
	struct sunqe *qep = netdev_priv(dev);

	qe_stop(qep);
	return 0;
}

/* Reclaim TX'd frames from the ring.  This must always run under
 * the IRQ protected qep->lock.
 */
static void qe_tx_reclaim(struct sunqe *qep)
{
	struct qe_txd *txbase = &qep->qe_block->qe_txd[0];
	int elem = qep->tx_old;

	while (elem != qep->tx_new) {
		u32 flags = txbase[elem].tx_flags;

		if (flags & TXD_OWN)
			break;
		elem = NEXT_TX(elem);
	}
	qep->tx_old = elem;
}

static void qe_tx_timeout(struct net_device *dev)
{
	struct sunqe *qep = netdev_priv(dev);
	int tx_full;

	spin_lock_irq(&qep->lock);

	/* Try to reclaim, if that frees up some tx
	 * entries, we're fine.
	 */
	qe_tx_reclaim(qep);
	tx_full = TX_BUFFS_AVAIL(qep) <= 0;

	spin_unlock_irq(&qep->lock);

	if (! tx_full)
		goto out;

	printk(KERN_ERR "%s: transmit timed out, resetting\n", dev->name);
	qe_init(qep, 1);

out:
	netif_wake_queue(dev);
}

/* Get a packet queued to go onto the wire. */
static int qe_start_xmit(struct sk_buff *skb, struct net_device *dev)
{
	struct sunqe *qep = netdev_priv(dev);
	struct sunqe_buffers *qbufs = qep->buffers;
	__u32 txbuf_dvma, qbufs_dvma = qep->buffers_dvma;
	unsigned char *txbuf;
	int len, entry;

	spin_lock_irq(&qep->lock);

	qe_tx_reclaim(qep);

	len = skb->len;
	entry = qep->tx_new;

	txbuf = &qbufs->tx_buf[entry & (TX_RING_SIZE - 1)][0];
	txbuf_dvma = qbufs_dvma +
		qebuf_offset(tx_buf, (entry & (TX_RING_SIZE - 1)));

	/* Avoid a race... */
	qep->qe_block->qe_txd[entry].tx_flags = TXD_UPDATE;

	skb_copy_from_linear_data(skb, txbuf, len);

	qep->qe_block->qe_txd[entry].tx_addr = txbuf_dvma;
	qep->qe_block->qe_txd[entry].tx_flags =
		(TXD_OWN | TXD_SOP | TXD_EOP | (len & TXD_LENGTH));
	qep->tx_new = NEXT_TX(entry);

	/* Get it going. */
	sbus_writel(CREG_CTRL_TWAKEUP, qep->qcregs + CREG_CTRL);

	dev->stats.tx_packets++;
	dev->stats.tx_bytes += len;

	if (TX_BUFFS_AVAIL(qep) <= 0) {
		/* Halt the net queue and enable tx interrupts.
		 * When the tx queue empties the tx irq handler
		 * will wake up the queue and return us back to
		 * the lazy tx reclaim scheme.
		 */
		netif_stop_queue(dev);
		sbus_writel(0, qep->qcregs + CREG_TIMASK);
	}
	spin_unlock_irq(&qep->lock);

	dev_kfree_skb(skb);

	return NETDEV_TX_OK;
}

static void qe_set_multicast(struct net_device *dev)
{
	struct sunqe *qep = netdev_priv(dev);
	struct netdev_hw_addr *ha;
	u8 new_mconfig = qep->mconfig;
	char *addrs;
	int i;
	u32 crc;

	/* Lock out others. */
	netif_stop_queue(dev);

	if ((dev->flags & IFF_ALLMULTI) || (netdev_mc_count(dev) > 64)) {
		sbus_writeb(MREGS_IACONFIG_ACHNGE | MREGS_IACONFIG_LARESET,
			    qep->mregs + MREGS_IACONFIG);
		while ((sbus_readb(qep->mregs + MREGS_IACONFIG) & MREGS_IACONFIG_ACHNGE) != 0)
			barrier();
		for (i = 0; i < 8; i++)
			sbus_writeb(0xff, qep->mregs + MREGS_FILTER);
		sbus_writeb(0, qep->mregs + MREGS_IACONFIG);
	} else if (dev->flags & IFF_PROMISC) {
		new_mconfig |= MREGS_MCONFIG_PROMISC;
	} else {
		u16 hash_table[4];
		u8 *hbytes = (unsigned char *) &hash_table[0];

		memset(hash_table, 0, sizeof(hash_table));
		netdev_for_each_mc_addr(ha, dev) {
			addrs = ha->addr;

			if (!(*addrs & 1))
				continue;
			crc = ether_crc_le(6, addrs);
			crc >>= 26;
			hash_table[crc >> 4] |= 1 << (crc & 0xf);
		}
		/* Program the qe with the new filter value. */
		sbus_writeb(MREGS_IACONFIG_ACHNGE | MREGS_IACONFIG_LARESET,
			    qep->mregs + MREGS_IACONFIG);
		while ((sbus_readb(qep->mregs + MREGS_IACONFIG) & MREGS_IACONFIG_ACHNGE) != 0)
			barrier();
		for (i = 0; i < 8; i++) {
			u8 tmp = *hbytes++;
			sbus_writeb(tmp, qep->mregs + MREGS_FILTER);
		}
		sbus_writeb(0, qep->mregs + MREGS_IACONFIG);
	}

	/* Any change of the logical address filter, the physical address,
	 * or enabling/disabling promiscuous mode causes the MACE to disable
	 * the receiver.  So we must re-enable them here or else the MACE
	 * refuses to listen to anything on the network.  Sheesh, took
	 * me a day or two to find this bug.
	 */
	qep->mconfig = new_mconfig;
	sbus_writeb(qep->mconfig, qep->mregs + MREGS_MCONFIG);

	/* Let us get going again. */
	netif_wake_queue(dev);
}

/* Ethtool support... */
static void qe_get_drvinfo(struct net_device *dev, struct ethtool_drvinfo *info)
{
	const struct linux_prom_registers *regs;
	struct sunqe *qep = netdev_priv(dev);
	struct of_device *op;

	strcpy(info->driver, "sunqe");
	strcpy(info->version, "3.0");

	op = qep->op;
	regs = of_get_property(op->node, "reg", NULL);
	if (regs)
		sprintf(info->bus_info, "SBUS:%d", regs->which_io);

}

static u32 qe_get_link(struct net_device *dev)
{
	struct sunqe *qep = netdev_priv(dev);
	void __iomem *mregs = qep->mregs;
	u8 phyconfig;

	spin_lock_irq(&qep->lock);
	phyconfig = sbus_readb(mregs + MREGS_PHYCONFIG);
	spin_unlock_irq(&qep->lock);

	return (phyconfig & MREGS_PHYCONFIG_LSTAT);
}

static const struct ethtool_ops qe_ethtool_ops = {
	.get_drvinfo		= qe_get_drvinfo,
	.get_link		= qe_get_link,
};

/* This is only called once at boot time for each card probed. */
static void qec_init_once(struct sunqec *qecp, struct of_device *op)
{
	u8 bsizes = qecp->qec_bursts;

	if (sbus_can_burst64() && (bsizes & DMA_BURST64)) {
		sbus_writel(GLOB_CTRL_B64, qecp->gregs + GLOB_CTRL);
	} else if (bsizes & DMA_BURST32) {
		sbus_writel(GLOB_CTRL_B32, qecp->gregs + GLOB_CTRL);
	} else {
		sbus_writel(GLOB_CTRL_B16, qecp->gregs + GLOB_CTRL);
	}

	/* Packetsize only used in 100baseT BigMAC configurations,
	 * set it to zero just to be on the safe side.
	 */
	sbus_writel(GLOB_PSIZE_2048, qecp->gregs + GLOB_PSIZE);

	/* Set the local memsize register, divided up to one piece per QE channel. */
	sbus_writel((resource_size(&op->resource[1]) >> 2),
		    qecp->gregs + GLOB_MSIZE);

	/* Divide up the local QEC memory amongst the 4 QE receiver and
	 * transmitter FIFOs.  Basically it is (total / 2 / num_channels).
	 */
	sbus_writel((resource_size(&op->resource[1]) >> 2) >> 1,
		    qecp->gregs + GLOB_TSIZE);
	sbus_writel((resource_size(&op->resource[1]) >> 2) >> 1,
		    qecp->gregs + GLOB_RSIZE);
}

static u8 __devinit qec_get_burst(struct device_node *dp)
{
	u8 bsizes, bsizes_more;

	/* Find and set the burst sizes for the QEC, since it
	 * does the actual dma for all 4 channels.
	 */
	bsizes = of_getintprop_default(dp, "burst-sizes", 0xff);
	bsizes &= 0xff;
	bsizes_more = of_getintprop_default(dp->parent, "burst-sizes", 0xff);

	if (bsizes_more != 0xff)
		bsizes &= bsizes_more;
	if (bsizes == 0xff || (bsizes & DMA_BURST16) == 0 ||
	    (bsizes & DMA_BURST32)==0)
		bsizes = (DMA_BURST32 - 1);

	return bsizes;
}

static struct sunqec * __devinit get_qec(struct of_device *child)
{
	struct of_device *op = to_of_device(child->dev.parent);
	struct sunqec *qecp;

	qecp = dev_get_drvdata(&op->dev);
	if (!qecp) {
		qecp = kzalloc(sizeof(struct sunqec), GFP_KERNEL);
		if (qecp) {
			u32 ctrl;

			qecp->op = op;
			qecp->gregs = of_ioremap(&op->resource[0], 0,
						 GLOB_REG_SIZE,
						 "QEC Global Registers");
			if (!qecp->gregs)
				goto fail;

			/* Make sure the QEC is in MACE mode. */
			ctrl = sbus_readl(qecp->gregs + GLOB_CTRL);
			ctrl &= 0xf0000000;
			if (ctrl != GLOB_CTRL_MMODE) {
				printk(KERN_ERR "qec: Not in MACE mode!\n");
				goto fail;
			}

			if (qec_global_reset(qecp->gregs))
				goto fail;

			qecp->qec_bursts = qec_get_burst(op->node);

			qec_init_once(qecp, op);

			if (request_irq(op->irqs[0], qec_interrupt,
					IRQF_SHARED, "qec", (void *) qecp)) {
				printk(KERN_ERR "qec: Can't register irq.\n");
				goto fail;
			}

			dev_set_drvdata(&op->dev, qecp);

			qecp->next_module = root_qec_dev;
			root_qec_dev = qecp;
		}
	}

	return qecp;

fail:
	if (qecp->gregs)
		of_iounmap(&op->resource[0], qecp->gregs, GLOB_REG_SIZE);
	kfree(qecp);
	return NULL;
}

static const struct net_device_ops qec_ops = {
	.ndo_open		= qe_open,
	.ndo_stop		= qe_close,
	.ndo_start_xmit		= qe_start_xmit,
	.ndo_set_multicast_list	= qe_set_multicast,
	.ndo_tx_timeout		= qe_tx_timeout,
	.ndo_change_mtu		= eth_change_mtu,
	.ndo_set_mac_address	= eth_mac_addr,
	.ndo_validate_addr	= eth_validate_addr,
};

static int __devinit qec_ether_init(struct of_device *op)
{
	static unsigned version_printed;
	struct net_device *dev;
	struct sunqec *qecp;
	struct sunqe *qe;
	int i, res;

	if (version_printed++ == 0)
		printk(KERN_INFO "%s", version);

	dev = alloc_etherdev(sizeof(struct sunqe));
	if (!dev)
		return -ENOMEM;

	memcpy(dev->dev_addr, idprom->id_ethaddr, 6);

	qe = netdev_priv(dev);

	res = -ENODEV;

	i = of_getintprop_default(op->node, "channel#", -1);
	if (i == -1)
		goto fail;
	qe->channel = i;
	spin_lock_init(&qe->lock);

	qecp = get_qec(op);
	if (!qecp)
		goto fail;

	qecp->qes[qe->channel] = qe;
	qe->dev = dev;
	qe->parent = qecp;
	qe->op = op;

	res = -ENOMEM;
	qe->qcregs = of_ioremap(&op->resource[0], 0,
				CREG_REG_SIZE, "QEC Channel Registers");
	if (!qe->qcregs) {
		printk(KERN_ERR "qe: Cannot map channel registers.\n");
		goto fail;
	}

	qe->mregs = of_ioremap(&op->resource[1], 0,
			       MREGS_REG_SIZE, "QE MACE Registers");
	if (!qe->mregs) {
		printk(KERN_ERR "qe: Cannot map MACE registers.\n");
		goto fail;
	}

	qe->qe_block = dma_alloc_coherent(&op->dev, PAGE_SIZE,
					  &qe->qblock_dvma, GFP_ATOMIC);
	qe->buffers = dma_alloc_coherent(&op->dev, sizeof(struct sunqe_buffers),
					 &qe->buffers_dvma, GFP_ATOMIC);
	if (qe->qe_block == NULL || qe->qblock_dvma == 0 ||
	    qe->buffers == NULL || qe->buffers_dvma == 0)
		goto fail;

	/* Stop this QE. */
	qe_stop(qe);

	SET_NETDEV_DEV(dev, &op->dev);

	dev->watchdog_timeo = 5*HZ;
	dev->irq = op->irqs[0];
	dev->dma = 0;
	dev->ethtool_ops = &qe_ethtool_ops;
	dev->netdev_ops = &qec_ops;

	res = register_netdev(dev);
	if (res)
		goto fail;

	dev_set_drvdata(&op->dev, qe);

	printk(KERN_INFO "%s: qe channel[%d] ", dev->name, qe->channel);
	for (i = 0; i < 6; i++)
		printk ("%2.2x%c",
			dev->dev_addr[i],
			i == 5 ? ' ': ':');
	printk("\n");


	return 0;

fail:
	if (qe->qcregs)
		of_iounmap(&op->resource[0], qe->qcregs, CREG_REG_SIZE);
	if (qe->mregs)
		of_iounmap(&op->resource[1], qe->mregs, MREGS_REG_SIZE);
	if (qe->qe_block)
		dma_free_coherent(&op->dev, PAGE_SIZE,
				  qe->qe_block, qe->qblock_dvma);
	if (qe->buffers)
		dma_free_coherent(&op->dev,
				  sizeof(struct sunqe_buffers),
				  qe->buffers,
				  qe->buffers_dvma);

	free_netdev(dev);

	return res;
}

static int __devinit qec_sbus_probe(struct of_device *op, const struct of_device_id *match)
{
	return qec_ether_init(op);
}

static int __devexit qec_sbus_remove(struct of_device *op)
{
	struct sunqe *qp = dev_get_drvdata(&op->dev);
	struct net_device *net_dev = qp->dev;

	unregister_netdev(net_dev);

	of_iounmap(&op->resource[0], qp->qcregs, CREG_REG_SIZE);
	of_iounmap(&op->resource[1], qp->mregs, MREGS_REG_SIZE);
	dma_free_coherent(&op->dev, PAGE_SIZE,
			  qp->qe_block, qp->qblock_dvma);
	dma_free_coherent(&op->dev, sizeof(struct sunqe_buffers),
			  qp->buffers, qp->buffers_dvma);

	free_netdev(net_dev);

	dev_set_drvdata(&op->dev, NULL);

	return 0;
}

static const struct of_device_id qec_sbus_match[] = {
	{
		.name = "qe",
	},
	{},
};

MODULE_DEVICE_TABLE(of, qec_sbus_match);

static struct of_platform_driver qec_sbus_driver = {
	.name		= "qec",
	.match_table	= qec_sbus_match,
	.probe		= qec_sbus_probe,
	.remove		= __devexit_p(qec_sbus_remove),
};

static int __init qec_init(void)
{
	return of_register_driver(&qec_sbus_driver, &of_bus_type);
}

static void __exit qec_exit(void)
{
	of_unregister_driver(&qec_sbus_driver);

	while (root_qec_dev) {
		struct sunqec *next = root_qec_dev->next_module;
		struct of_device *op = root_qec_dev->op;

		free_irq(op->irqs[0], (void *) root_qec_dev);
		of_iounmap(&op->resource[0], root_qec_dev->gregs,
			   GLOB_REG_SIZE);
		kfree(root_qec_dev);

		root_qec_dev = next;
	}
}

module_init(qec_init);
module_exit(qec_exit);