acenic.c

	readl(&regs->ModeStat);		/* PCI write posting */

	mac1 = 0;
	for(i = 0; i < 4; i++) {
		int tmp;

		mac1 = mac1 << 8;
		tmp = read_eeprom_byte(dev, 0x8c+i);
		if (tmp < 0) {
			ecode = -EIO;
			goto init_error;
		} else
			mac1 |= (tmp & 0xff);
	}
	mac2 = 0;
	for(i = 4; i < 8; i++) {
		int tmp;

		mac2 = mac2 << 8;
		tmp = read_eeprom_byte(dev, 0x8c+i);
		if (tmp < 0) {
			ecode = -EIO;
			goto init_error;
		} else
			mac2 |= (tmp & 0xff);
	}

	writel(mac1, &regs->MacAddrHi);
	writel(mac2, &regs->MacAddrLo);

	printk("MAC: %02x:%02x:%02x:%02x:%02x:%02x\n",
	       (mac1 >> 8) & 0xff, mac1 & 0xff, (mac2 >> 24) &0xff,
	       (mac2 >> 16) & 0xff, (mac2 >> 8) & 0xff, mac2 & 0xff);

	dev->dev_addr[0] = (mac1 >> 8) & 0xff;
	dev->dev_addr[1] = mac1 & 0xff;
	dev->dev_addr[2] = (mac2 >> 24) & 0xff;
	dev->dev_addr[3] = (mac2 >> 16) & 0xff;
	dev->dev_addr[4] = (mac2 >> 8) & 0xff;
	dev->dev_addr[5] = mac2 & 0xff;

	/*
	 * Looks like this is necessary to deal with on all architectures,
	 * even this %$#%$# N440BX Intel based thing doesn't get it right.
	 * Ie. having two NICs in the machine, one will have the cache
	 * line set at boot time, the other will not.
	 */
	pdev = ap->pdev;
	pci_read_config_byte(pdev, PCI_CACHE_LINE_SIZE, &cache_size);
	cache_size <<= 2;
	if (cache_size != SMP_CACHE_BYTES) {
		printk(KERN_INFO "  PCI cache line size set incorrectly "
		       "(%i bytes) by BIOS/FW, ", cache_size);
		if (cache_size > SMP_CACHE_BYTES)
			printk("expecting %i\n", SMP_CACHE_BYTES);
		else {
			printk("correcting to %i\n", SMP_CACHE_BYTES);
			pci_write_config_byte(pdev, PCI_CACHE_LINE_SIZE,
					      SMP_CACHE_BYTES >> 2);
		}
	}

	pci_state = readl(&regs->PciState);
	printk(KERN_INFO "  PCI bus width: %i bits, speed: %iMHz, "
	       "latency: %i clks\n",
	       	(pci_state & PCI_32BIT) ? 32 : 64,
		(pci_state & PCI_66MHZ) ? 66 : 33, 
		ap->pci_latency);

	/*
	 * Set the max DMA transfer size. Seems that for most systems
	 * the performance is better when no MAX parameter is
	 * set. However for systems enabling PCI write and invalidate,
	 * DMA writes must be set to the L1 cache line size to get
	 * optimal performance.
	 *
	 * The default is now to turn the PCI write and invalidate off
	 * - that is what Alteon does for NT.
	 */
	tmp = READ_CMD_MEM | WRITE_CMD_MEM;
	if (ap->version >= 2) {
		tmp |= (MEM_READ_MULTIPLE | (pci_state & PCI_66MHZ));
		/*
		 * Tuning parameters only supported for 8 cards
		 */
		if (board_idx == BOARD_IDX_OVERFLOW ||
		    dis_pci_mem_inval[board_idx]) {
			if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
				ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
				pci_write_config_word(pdev, PCI_COMMAND,
						      ap->pci_command);
				printk(KERN_INFO "  Disabling PCI memory "
				       "write and invalidate\n");
			}
		} else if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
			printk(KERN_INFO "  PCI memory write & invalidate "
			       "enabled by BIOS, enabling counter measures\n");

			switch(SMP_CACHE_BYTES) {
			case 16:
				tmp |= DMA_WRITE_MAX_16;
				break;
			case 32:
				tmp |= DMA_WRITE_MAX_32;
				break;
			case 64:
				tmp |= DMA_WRITE_MAX_64;
				break;
			case 128:
				tmp |= DMA_WRITE_MAX_128;
				break;
			default:
				printk(KERN_INFO "  Cache line size %i not "
				       "supported, PCI write and invalidate "
				       "disabled\n", SMP_CACHE_BYTES);
				ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
				pci_write_config_word(pdev, PCI_COMMAND,
						      ap->pci_command);
			}
		}
	}

#ifdef __sparc__
	/*
	 * On this platform, we know what the best dma settings
	 * are.  We use 64-byte maximum bursts, because if we
	 * burst larger than the cache line size (or even cross
	 * a 64byte boundary in a single burst) the UltraSparc
	 * PCI controller will disconnect at 64-byte multiples.
	 *
	 * Read-multiple will be properly enabled above, and when
	 * set will give the PCI controller proper hints about
	 * prefetching.
	 */
	tmp &= ~DMA_READ_WRITE_MASK;
	tmp |= DMA_READ_MAX_64;
	tmp |= DMA_WRITE_MAX_64;
#endif
#ifdef __alpha__
	tmp &= ~DMA_READ_WRITE_MASK;
	tmp |= DMA_READ_MAX_128;
	/*
	 * All the docs say MUST NOT. Well, I did.
	 * Nothing terrible happens, if we load wrong size.
	 * Bit w&i still works better!
	 */
	tmp |= DMA_WRITE_MAX_128;
#endif
	writel(tmp, &regs->PciState);

#if 0
	/*
	 * The Host PCI bus controller driver has to set FBB.
	 * If all devices on that PCI bus support FBB, then the controller
	 * can enable FBB support in the Host PCI Bus controller (or on
	 * the PCI-PCI bridge if that applies).
	 * -ggg
	 */
	/*
	 * I have received reports from people having problems when this
	 * bit is enabled.
	 */
	if (!(ap->pci_command & PCI_COMMAND_FAST_BACK)) {
		printk(KERN_INFO "  Enabling PCI Fast Back to Back\n");
		ap->pci_command |= PCI_COMMAND_FAST_BACK;
		pci_write_config_word(pdev, PCI_COMMAND, ap->pci_command);
	}
#endif
		
	/*
	 * Configure DMA attributes.
	 */
	if (!pci_set_dma_mask(pdev, DMA_64BIT_MASK)) {
		ap->pci_using_dac = 1;
	} else if (!pci_set_dma_mask(pdev, DMA_32BIT_MASK)) {
		ap->pci_using_dac = 0;
	} else {
		ecode = -ENODEV;
		goto init_error;
	}

	/*
	 * Initialize the generic info block and the command+event rings
	 * and the control blocks for the transmit and receive rings
	 * as they need to be setup once and for all.
	 */
	if (!(info = pci_alloc_consistent(ap->pdev, sizeof(struct ace_info),
					  &ap->info_dma))) {
		ecode = -EAGAIN;
		goto init_error;
	}
	ap->info = info;

	/*
	 * Get the memory for the skb rings.
	 */
	if (!(ap->skb = kmalloc(sizeof(struct ace_skb), GFP_KERNEL))) {
		ecode = -EAGAIN;
		goto init_error;
	}

	ecode = request_irq(pdev->irq, ace_interrupt, SA_SHIRQ,
			    DRV_NAME, dev);
	if (ecode) {
		printk(KERN_WARNING "%s: Requested IRQ %d is busy\n",
		       DRV_NAME, pdev->irq);
		goto init_error;
	} else
		dev->irq = pdev->irq;

#ifdef INDEX_DEBUG
	spin_lock_init(&ap->debug_lock);
	ap->last_tx = ACE_TX_RING_ENTRIES(ap) - 1;
	ap->last_std_rx = 0;
	ap->last_mini_rx = 0;
#endif

	memset(ap->info, 0, sizeof(struct ace_info));
	memset(ap->skb, 0, sizeof(struct ace_skb));

	ace_load_firmware(dev);
	ap->fw_running = 0;

	tmp_ptr = ap->info_dma;
	writel(tmp_ptr >> 32, &regs->InfoPtrHi);
	writel(tmp_ptr & 0xffffffff, &regs->InfoPtrLo);

	memset(ap->evt_ring, 0, EVT_RING_ENTRIES * sizeof(struct event));

	set_aceaddr(&info->evt_ctrl.rngptr, ap->evt_ring_dma);
	info->evt_ctrl.flags = 0;

	*(ap->evt_prd) = 0;
	wmb();
	set_aceaddr(&info->evt_prd_ptr, ap->evt_prd_dma);
	writel(0, &regs->EvtCsm);

	set_aceaddr(&info->cmd_ctrl.rngptr, 0x100);
	info->cmd_ctrl.flags = 0;
	info->cmd_ctrl.max_len = 0;

	for (i = 0; i < CMD_RING_ENTRIES; i++)
		writel(0, &regs->CmdRng[i]);

	writel(0, &regs->CmdPrd);
	writel(0, &regs->CmdCsm);

	tmp_ptr = ap->info_dma;
	tmp_ptr += (unsigned long) &(((struct ace_info *)0)->s.stats);
	set_aceaddr(&info->stats2_ptr, (dma_addr_t) tmp_ptr);

	set_aceaddr(&info->rx_std_ctrl.rngptr, ap->rx_ring_base_dma);
	info->rx_std_ctrl.max_len = ACE_STD_BUFSIZE;
	info->rx_std_ctrl.flags =
	  RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | ACE_RCB_VLAN_FLAG;

	memset(ap->rx_std_ring, 0,
	       RX_STD_RING_ENTRIES * sizeof(struct rx_desc));

	for (i = 0; i < RX_STD_RING_ENTRIES; i++)
		ap->rx_std_ring[i].flags = BD_FLG_TCP_UDP_SUM;

	ap->rx_std_skbprd = 0;
	atomic_set(&ap->cur_rx_bufs, 0);

	set_aceaddr(&info->rx_jumbo_ctrl.rngptr,
		    (ap->rx_ring_base_dma +
		     (sizeof(struct rx_desc) * RX_STD_RING_ENTRIES)));
	info->rx_jumbo_ctrl.max_len = 0;
	info->rx_jumbo_ctrl.flags =
	  RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | ACE_RCB_VLAN_FLAG;

	memset(ap->rx_jumbo_ring, 0,
	       RX_JUMBO_RING_ENTRIES * sizeof(struct rx_desc));

	for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++)
		ap->rx_jumbo_ring[i].flags = BD_FLG_TCP_UDP_SUM | BD_FLG_JUMBO;

	ap->rx_jumbo_skbprd = 0;
	atomic_set(&ap->cur_jumbo_bufs, 0);

	memset(ap->rx_mini_ring, 0,
	       RX_MINI_RING_ENTRIES * sizeof(struct rx_desc));

	if (ap->version >= 2) {
		set_aceaddr(&info->rx_mini_ctrl.rngptr,
			    (ap->rx_ring_base_dma +
			     (sizeof(struct rx_desc) *
			      (RX_STD_RING_ENTRIES +
			       RX_JUMBO_RING_ENTRIES))));
		info->rx_mini_ctrl.max_len = ACE_MINI_SIZE;
		info->rx_mini_ctrl.flags = 
		  RCB_FLG_TCP_UDP_SUM|RCB_FLG_NO_PSEUDO_HDR|ACE_RCB_VLAN_FLAG;

		for (i = 0; i < RX_MINI_RING_ENTRIES; i++)
			ap->rx_mini_ring[i].flags =
				BD_FLG_TCP_UDP_SUM | BD_FLG_MINI;
	} else {
		set_aceaddr(&info->rx_mini_ctrl.rngptr, 0);
		info->rx_mini_ctrl.flags = RCB_FLG_RNG_DISABLE;
		info->rx_mini_ctrl.max_len = 0;
	}

	ap->rx_mini_skbprd = 0;
	atomic_set(&ap->cur_mini_bufs, 0);

	set_aceaddr(&info->rx_return_ctrl.rngptr,
		    (ap->rx_ring_base_dma +
		     (sizeof(struct rx_desc) *
		      (RX_STD_RING_ENTRIES +
		       RX_JUMBO_RING_ENTRIES +
		       RX_MINI_RING_ENTRIES))));
	info->rx_return_ctrl.flags = 0;
	info->rx_return_ctrl.max_len = RX_RETURN_RING_ENTRIES;

	memset(ap->rx_return_ring, 0,
	       RX_RETURN_RING_ENTRIES * sizeof(struct rx_desc));

	set_aceaddr(&info->rx_ret_prd_ptr, ap->rx_ret_prd_dma);
	*(ap->rx_ret_prd) = 0;

	writel(TX_RING_BASE, &regs->WinBase);

	if (ACE_IS_TIGON_I(ap)) {
		ap->tx_ring = (struct tx_desc *) regs->Window;
		for (i = 0; i < (TIGON_I_TX_RING_ENTRIES 
				 * sizeof(struct tx_desc)) / sizeof(u32); i++)
			writel(0, (void __iomem *)ap->tx_ring  + i * 4);

		set_aceaddr(&info->tx_ctrl.rngptr, TX_RING_BASE);
	} else {
		memset(ap->tx_ring, 0,
		       MAX_TX_RING_ENTRIES * sizeof(struct tx_desc));

		set_aceaddr(&info->tx_ctrl.rngptr, ap->tx_ring_dma);
	}

	info->tx_ctrl.max_len = ACE_TX_RING_ENTRIES(ap);
	tmp = RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | ACE_RCB_VLAN_FLAG;

	/*
	 * The Tigon I does not like having the TX ring in host memory ;-(
	 */
	if (!ACE_IS_TIGON_I(ap))
		tmp |= RCB_FLG_TX_HOST_RING;
#if TX_COAL_INTS_ONLY
	tmp |= RCB_FLG_COAL_INT_ONLY;
#endif
	info->tx_ctrl.flags = tmp;

	set_aceaddr(&info->tx_csm_ptr, ap->tx_csm_dma);

	/*
	 * Potential item for tuning parameter
	 */
#if 0 /* NO */
	writel(DMA_THRESH_16W, &regs->DmaReadCfg);
	writel(DMA_THRESH_16W, &regs->DmaWriteCfg);
#else
	writel(DMA_THRESH_8W, &regs->DmaReadCfg);
	writel(DMA_THRESH_8W, &regs->DmaWriteCfg);
#endif

	writel(0, &regs->MaskInt);
	writel(1, &regs->IfIdx);
#if 0
	/*
	 * McKinley boxes do not like us fiddling with AssistState
	 * this early
	 */
	writel(1, &regs->AssistState);
#endif

	writel(DEF_STAT, &regs->TuneStatTicks);
	writel(DEF_TRACE, &regs->TuneTrace);

	ace_set_rxtx_parms(dev, 0);

	if (board_idx == BOARD_IDX_OVERFLOW) {
		printk(KERN_WARNING "%s: more than %i NICs detected, "
		       "ignoring module parameters!\n",
		       ap->name, ACE_MAX_MOD_PARMS);
	} else if (board_idx >= 0) {
		if (tx_coal_tick[board_idx])
			writel(tx_coal_tick[board_idx],
			       &regs->TuneTxCoalTicks);
		if (max_tx_desc[board_idx])
			writel(max_tx_desc[board_idx], &regs->TuneMaxTxDesc);

		if (rx_coal_tick[board_idx])
			writel(rx_coal_tick[board_idx],
			       &regs->TuneRxCoalTicks);
		if (max_rx_desc[board_idx])
			writel(max_rx_desc[board_idx], &regs->TuneMaxRxDesc);

		if (trace[board_idx])
			writel(trace[board_idx], &regs->TuneTrace);

		if ((tx_ratio[board_idx] > 0) && (tx_ratio[board_idx] < 64))
			writel(tx_ratio[board_idx], &regs->TxBufRat);
	}

	/*
	 * Default link parameters
	 */
	tmp = LNK_ENABLE | LNK_FULL_DUPLEX | LNK_1000MB | LNK_100MB |
		LNK_10MB | LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL | LNK_NEGOTIATE;
	if(ap->version >= 2)
		tmp |= LNK_TX_FLOW_CTL_Y;

	/*
	 * Override link default parameters
	 */
	if ((board_idx >= 0) && link[board_idx]) {
		int option = link[board_idx];

		tmp = LNK_ENABLE;

		if (option & 0x01) {
			printk(KERN_INFO "%s: Setting half duplex link\n",
			       ap->name);
			tmp &= ~LNK_FULL_DUPLEX;
		}
		if (option & 0x02)
			tmp &= ~LNK_NEGOTIATE;
		if (option & 0x10)
			tmp |= LNK_10MB;
		if (option & 0x20)
			tmp |= LNK_100MB;
		if (option & 0x40)
			tmp |= LNK_1000MB;
		if ((option & 0x70) == 0) {
			printk(KERN_WARNING "%s: No media speed specified, "
			       "forcing auto negotiation\n", ap->name);
			tmp |= LNK_NEGOTIATE | LNK_1000MB |
				LNK_100MB | LNK_10MB;
		}
		if ((option & 0x100) == 0)
			tmp |= LNK_NEG_FCTL;
		else
			printk(KERN_INFO "%s: Disabling flow control "
			       "negotiation\n", ap->name);
		if (option & 0x200)
			tmp |= LNK_RX_FLOW_CTL_Y;
		if ((option & 0x400) && (ap->version >= 2)) {
			printk(KERN_INFO "%s: Enabling TX flow control\n",
			       ap->name);
			tmp |= LNK_TX_FLOW_CTL_Y;
		}
	}

	ap->link = tmp;
	writel(tmp, &regs->TuneLink);
	if (ap->version >= 2)
		writel(tmp, &regs->TuneFastLink);

	if (ACE_IS_TIGON_I(ap))
		writel(tigonFwStartAddr, &regs->Pc);
	if (ap->version == 2)
		writel(tigon2FwStartAddr, &regs->Pc);

	writel(0, &regs->Mb0Lo);

	/*
	 * Set tx_csm before we start receiving interrupts, otherwise
	 * the interrupt handler might think it is supposed to process
	 * tx ints before we are up and running, which may cause a null
	 * pointer access in the int handler.
	 */
	ap->cur_rx = 0;
	ap->tx_prd = *(ap->tx_csm) = ap->tx_ret_csm = 0;

	wmb();
	ace_set_txprd(regs, ap, 0);
	writel(0, &regs->RxRetCsm);

	/*
	 * Zero the stats before starting the interface
	 */
	memset(&ap->stats, 0, sizeof(ap->stats));

       /*
	* Enable DMA engine now.
	* If we do this sooner, Mckinley box pukes.
	* I assume it's because Tigon II DMA engine wants to check
	* *something* even before the CPU is started.
	*/
       writel(1, &regs->AssistState);  /* enable DMA */

	/*
	 * Start the NIC CPU
	 */
	writel(readl(&regs->CpuCtrl) & ~(CPU_HALT|CPU_TRACE), &regs->CpuCtrl);
	readl(&regs->CpuCtrl);

	/*
	 * Wait for the firmware to spin up - max 3 seconds.
	 */
	myjif = jiffies + 3 * HZ;
	while (time_before(jiffies, myjif) && !ap->fw_running)
		cpu_relax();

	if (!ap->fw_running) {
		printk(KERN_ERR "%s: Firmware NOT running!\n", ap->name);

		ace_dump_trace(ap);
		writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
		readl(&regs->CpuCtrl);

		/* aman@sgi.com - account for badly behaving firmware/NIC:
		 * - have observed that the NIC may continue to generate
		 *   interrupts for some reason; attempt to stop it - halt
		 *   second CPU for Tigon II cards, and also clear Mb0
		 * - if we're a module, we'll fail to load if this was
		 *   the only GbE card in the system => if the kernel does
		 *   see an interrupt from the NIC, code to handle it is
		 *   gone and OOps! - so free_irq also
		 */
		if (ap->version >= 2)
			writel(readl(&regs->CpuBCtrl) | CPU_HALT,
			       &regs->CpuBCtrl);
		writel(0, &regs->Mb0Lo);
		readl(&regs->Mb0Lo);

		ecode = -EBUSY;
		goto init_error;
	}

	/*
	 * We load the ring here as there seem to be no way to tell the
	 * firmware to wipe the ring without re-initializing it.
	 */
	if (!test_and_set_bit(0, &ap->std_refill_busy))
		ace_load_std_rx_ring(ap, RX_RING_SIZE);
	else
		printk(KERN_ERR "%s: Someone is busy refilling the RX ring\n",
		       ap->name);
	if (ap->version >= 2) {
		if (!test_and_set_bit(0, &ap->mini_refill_busy))
			ace_load_mini_rx_ring(ap, RX_MINI_SIZE);
		else
			printk(KERN_ERR "%s: Someone is busy refilling "
			       "the RX mini ring\n", ap->name);
	}
	return 0;

 init_error:
	ace_init_cleanup(dev);
	return ecode;
}


static void ace_set_rxtx_parms(struct net_device *dev, int jumbo)
{
	struct ace_private *ap = netdev_priv(dev);
	struct ace_regs __iomem *regs = ap->regs;
	int board_idx = ap->board_idx;

	if (board_idx >= 0) {
		if (!jumbo) {
			if (!tx_coal_tick[board_idx])
				writel(DEF_TX_COAL, &regs->TuneTxCoalTicks);
			if (!max_tx_desc[board_idx])
				writel(DEF_TX_MAX_DESC, &regs->TuneMaxTxDesc);
			if (!rx_coal_tick[board_idx])
				writel(DEF_RX_COAL, &regs->TuneRxCoalTicks);
			if (!max_rx_desc[board_idx])
				writel(DEF_RX_MAX_DESC, &regs->TuneMaxRxDesc);
			if (!tx_ratio[board_idx])
				writel(DEF_TX_RATIO, &regs->TxBufRat);
		} else {
			if (!tx_coal_tick[board_idx])
				writel(DEF_JUMBO_TX_COAL,
				       &regs->TuneTxCoalTicks);
			if (!max_tx_desc[board_idx])
				writel(DEF_JUMBO_TX_MAX_DESC,
				       &regs->TuneMaxTxDesc);
			if (!rx_coal_tick[board_idx])
				writel(DEF_JUMBO_RX_COAL,
				       &regs->TuneRxCoalTicks);
			if (!max_rx_desc[board_idx])
				writel(DEF_JUMBO_RX_MAX_DESC,
				       &regs->TuneMaxRxDesc);
			if (!tx_ratio[board_idx])
				writel(DEF_JUMBO_TX_RATIO, &regs->TxBufRat);
		}
	}
}


static void ace_watchdog(struct net_device *data)
{
	struct net_device *dev = data;
	struct ace_private *ap = netdev_priv(dev);
	struct ace_regs __iomem *regs = ap->regs;

	/*
	 * We haven't received a stats update event for more than 2.5
	 * seconds and there is data in the transmit queue, thus we
	 * asume the card is stuck.
	 */
	if (*ap->tx_csm != ap->tx_ret_csm) {
		printk(KERN_WARNING "%s: Transmitter is stuck, %08x\n",
		       dev->name, (unsigned int)readl(&regs->HostCtrl));
		/* This can happen due to ieee flow control. */
	} else {
		printk(KERN_DEBUG "%s: BUG... transmitter died. Kicking it.\n",
		       dev->name);
#if 0
		netif_wake_queue(dev);
#endif
	}
}


static void ace_tasklet(unsigned long dev)
{
	struct ace_private *ap = netdev_priv((struct net_device *)dev);
	int cur_size;

	cur_size = atomic_read(&ap->cur_rx_bufs);
	if ((cur_size < RX_LOW_STD_THRES) &&
	    !test_and_set_bit(0, &ap->std_refill_busy)) {
#ifdef DEBUG
		printk("refilling buffers (current %i)\n", cur_size);
#endif
		ace_load_std_rx_ring(ap, RX_RING_SIZE - cur_size);
	}

	if (ap->version >= 2) {
		cur_size = atomic_read(&ap->cur_mini_bufs);
		if ((cur_size < RX_LOW_MINI_THRES) &&
		    !test_and_set_bit(0, &ap->mini_refill_busy)) {
#ifdef DEBUG
			printk("refilling mini buffers (current %i)\n",
			       cur_size);
#endif
			ace_load_mini_rx_ring(ap, RX_MINI_SIZE - cur_size);
		}
	}

	cur_size = atomic_read(&ap->cur_jumbo_bufs);
	if (ap->jumbo && (cur_size < RX_LOW_JUMBO_THRES) &&
	    !test_and_set_bit(0, &ap->jumbo_refill_busy)) {
#ifdef DEBUG
		printk("refilling jumbo buffers (current %i)\n", cur_size);
#endif
		ace_load_jumbo_rx_ring(ap, RX_JUMBO_SIZE - cur_size);
	}
	ap->tasklet_pending = 0;
}


/*
 * Copy the contents of the NIC's trace buffer to kernel memory.
 */
static void ace_dump_trace(struct ace_private *ap)
{
#if 0
	if (!ap->trace_buf)
		if (!(ap->trace_buf = kmalloc(ACE_TRACE_SIZE, GFP_KERNEL)))
		    return;
#endif
}


/*
 * Load the standard rx ring.
 *
 * Loading rings is safe without holding the spin lock since this is
 * done only before the device is enabled, thus no interrupts are
 * generated and by the interrupt handler/tasklet handler.
 */
static void ace_load_std_rx_ring(struct ace_private *ap, int nr_bufs)
{
	struct ace_regs __iomem *regs = ap->regs;
	short i, idx;
	

	prefetchw(&ap->cur_rx_bufs);

	idx = ap->rx_std_skbprd;

	for (i = 0; i < nr_bufs; i++) {
		struct sk_buff *skb;
		struct rx_desc *rd;
		dma_addr_t mapping;

		skb = alloc_skb(ACE_STD_BUFSIZE + NET_IP_ALIGN, GFP_ATOMIC);
		if (!skb)
			break;

		skb_reserve(skb, NET_IP_ALIGN);
		mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
				       offset_in_page(skb->data),
				       ACE_STD_BUFSIZE,
				       PCI_DMA_FROMDEVICE);
		ap->skb->rx_std_skbuff[idx].skb = skb;
		pci_unmap_addr_set(&ap->skb->rx_std_skbuff[idx],
				   mapping, mapping);

		rd = &ap->rx_std_ring[idx];
		set_aceaddr(&rd->addr, mapping);
		rd->size = ACE_STD_BUFSIZE;
		rd->idx = idx;
		idx = (idx + 1) % RX_STD_RING_ENTRIES;
	}

	if (!i)
		goto error_out;

	atomic_add(i, &ap->cur_rx_bufs);
	ap->rx_std_skbprd = idx;

	if (ACE_IS_TIGON_I(ap)) {
		struct cmd cmd;
		cmd.evt = C_SET_RX_PRD_IDX;
		cmd.code = 0;
		cmd.idx = ap->rx_std_skbprd;
		ace_issue_cmd(regs, &cmd);
	} else {
		writel(idx, &regs->RxStdPrd);
		wmb();
	}

 out:
	clear_bit(0, &ap->std_refill_busy);
	return;

 error_out:
	printk(KERN_INFO "Out of memory when allocating "
	       "standard receive buffers\n");
	goto out;
}


static void ace_load_mini_rx_ring(struct ace_private *ap, int nr_bufs)
{
	struct ace_regs __iomem *regs = ap->regs;
	short i, idx;

	prefetchw(&ap->cur_mini_bufs);

	idx = ap->rx_mini_skbprd;
	for (i = 0; i < nr_bufs; i++) {
		struct sk_buff *skb;
		struct rx_desc *rd;
		dma_addr_t mapping;

		skb = alloc_skb(ACE_MINI_BUFSIZE + NET_IP_ALIGN, GFP_ATOMIC);
		if (!skb)
			break;

		skb_reserve(skb, NET_IP_ALIGN);
		mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
				       offset_in_page(skb->data),
				       ACE_MINI_BUFSIZE,
				       PCI_DMA_FROMDEVICE);
		ap->skb->rx_mini_skbuff[idx].skb = skb;
		pci_unmap_addr_set(&ap->skb->rx_mini_skbuff[idx],
				   mapping, mapping);

		rd = &ap->rx_mini_ring[idx];
		set_aceaddr(&rd->addr, mapping);
		rd->size = ACE_MINI_BUFSIZE;
		rd->idx = idx;
		idx = (idx + 1) % RX_MINI_RING_ENTRIES;
	}

	if (!i)
		goto error_out;

	atomic_add(i, &ap->cur_mini_bufs);

	ap->rx_mini_skbprd = idx;

	writel(idx, &regs->RxMiniPrd);
	wmb();

 out:
	clear_bit(0, &ap->mini_refill_busy);
	return;
 error_out:
	printk(KERN_INFO "Out of memory when allocating "
	       "mini receive buffers\n");
	goto out;
}


/*
 * Load the jumbo rx ring, this may happen at any time if the MTU
 * is changed to a value > 1500.
 */
static void ace_load_jumbo_rx_ring(struct ace_private *ap, int nr_bufs)
{
	struct ace_regs __iomem *regs = ap->regs;
	short i, idx;

	idx = ap->rx_jumbo_skbprd;

	for (i = 0; i < nr_bufs; i++) {
		struct sk_buff *skb;
		struct rx_desc *rd;
		dma_addr_t mapping;

		skb = alloc_skb(ACE_JUMBO_BUFSIZE + NET_IP_ALIGN, GFP_ATOMIC);
		if (!skb)
			break;

		skb_reserve(skb, NET_IP_ALIGN);
		mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
				       offset_in_page(skb->data),
				       ACE_JUMBO_BUFSIZE,
				       PCI_DMA_FROMDEVICE);
		ap->skb->rx_jumbo_skbuff[idx].skb = skb;
		pci_unmap_addr_set(&ap->skb->rx_jumbo_skbuff[idx],
				   mapping, mapping);

		rd = &ap->rx_jumbo_ring[idx];
		set_aceaddr(&rd->addr, mapping);
		rd->size = ACE_JUMBO_BUFSIZE;
		rd->idx = idx;
		idx = (idx + 1) % RX_JUMBO_RING_ENTRIES;
	}

	if (!i)
		goto error_out;

	atomic_add(i, &ap->cur_jumbo_bufs);
	ap->rx_jumbo_skbprd = idx;

	if (ACE_IS_TIGON_I(ap)) {
		struct cmd cmd;
		cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
		cmd.code = 0;
		cmd.idx = ap->rx_jumbo_skbprd;
		ace_issue_cmd(regs, &cmd);
	} else {
		writel(idx, &regs->RxJumboPrd);
		wmb();
	}

 out:
	clear_bit(0, &ap->jumbo_refill_busy);
	return;
 error_out:
	if (net_ratelimit())
		printk(KERN_INFO "Out of memory when allocating "
		       "jumbo receive buffers\n");
	goto out;
}


/*
 * All events are considered to be slow (RX/TX ints do not generate
 * events) and are handled here, outside the main interrupt handler,
 * to reduce the size of the handler.
 */
static u32 ace_handle_event(struct net_device *dev, u32 evtcsm, u32 evtprd)
{
	struct ace_private *ap;

	ap = netdev_priv(dev);

	while (evtcsm != evtprd) {
		switch (ap->evt_ring[evtcsm].evt) {
		case E_FW_RUNNING:
			printk(KERN_INFO "%s: Firmware up and running\n",
			       ap->name);
			ap->fw_running = 1;
			wmb();
			break;
		case E_STATS_UPDATED:
			break;
		case E_LNK_STATE:
		{
			u16 code = ap->evt_ring[evtcsm].code;
			switch (code) {
			case E_C_LINK_UP:
			{
				u32 state = readl(&ap->regs->GigLnkState);
				printk(KERN_WARNING "%s: Optical link UP "
				       "(%s Duplex, Flow Control: %s%s)\n",
				       ap->name,
				       state & LNK_FULL_DUPLEX ? "Full":"Half",
				       state & LNK_TX_FLOW_CTL_Y ? "TX " : "",
				       state & LNK_RX_FLOW_CTL_Y ? "RX" : "");
				break;
			}
			case E_C_LINK_DOWN:
				printk(KERN_WARNING "%s: Optical link DOWN\n",
				       ap->name);
				break;
			case E_C_LINK_10_100:
				printk(KERN_WARNING "%s: 10/100BaseT link "
				       "UP\n", ap->name);
				break;
			default:
				printk(KERN_ERR "%s: Unknown optical link "
				       "state %02x\n", ap->name, code);
			}
			break;
		}
		case E_ERROR:
			switch(ap->evt_ring[evtcsm].code) {
			case E_C_ERR_INVAL_CMD:
				printk(KERN_ERR "%s: invalid command error\n",
				       ap->name);
				break;
			case E_C_ERR_UNIMP_CMD:
				printk(KERN_ERR "%s: unimplemented command "
				       "error\n", ap->name);
				break;
			case E_C_ERR_BAD_CFG:
				printk(KERN_ERR "%s: bad config error\n",
				       ap->name);
				break;
			default:
				printk(KERN_ERR "%s: unknown error %02x\n",
				       ap->name, ap->evt_ring[evtcsm].code);
			}
			break;
		case E_RESET_JUMBO_RNG:
		{
			int i;
			for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
				if (ap->skb->rx_jumbo_skbuff[i].skb) {
					ap->rx_jumbo_ring[i].size = 0;
					set_aceaddr(&ap->rx_jumbo_ring[i].addr, 0);
					dev_kfree_skb(ap->skb->rx_jumbo_skbuff[i].skb);
					ap->skb->rx_jumbo_skbuff[i].skb = NULL;
				}
			}

 			if (ACE_IS_TIGON_I(ap)) {
 				struct cmd cmd;
 				cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
 				cmd.code = 0;
 				cmd.idx = 0;
 				ace_issue_cmd(ap->regs, &cmd);
 			} else {
 				writel(0, &((ap->regs)->RxJumboPrd));
 				wmb();
 			}

			ap->jumbo = 0;
			ap->rx_jumbo_skbprd = 0;
			printk(KERN_INFO "%s: Jumbo ring flushed\n",
			       ap->name);
			clear_bit(0, &ap->jumbo_refill_busy);
			break;
		}
		default:
			printk(KERN_ERR "%s: Unhandled event 0x%02x\n",
			       ap->name, ap->evt_ring[evtcsm].evt);
		}
		evtcsm = (evtcsm + 1) % EVT_RING_ENTRIES;
	}

	return evtcsm;
}


static void ace_rx_int(struct net_device *dev, u32 rxretprd, u32 rxretcsm)
{
	struct ace_private *ap = netdev_priv(dev);
	u32 idx;
	int mini_count = 0, std_count = 0;

	idx = rxretcsm;

	prefetchw(&ap->cur_rx_bufs);
	prefetchw(&ap->cur_mini_bufs);
	
	while (idx != rxretprd) {
		struct ring_info *rip;
		struct sk_buff *skb;
		struct rx_desc *rxdesc, *retdesc;
		u32 skbidx;
		int bd_flags, desc_type, mapsize;
		u16 csum;


		/* make sure the rx descriptor isn't read before rxretprd */
		if (idx == rxretcsm) 
			rmb();

		retdesc = &ap->rx_return_ring[idx];
		skbidx = retdesc->idx;
		bd_flags = retdesc->flags;
		desc_type = bd_flags & (BD_FLG_JUMBO | BD_FLG_MINI);

		switch(desc_type) {
			/*
			 * Normal frames do not have any flags set
			 *
			 * Mini and normal frames arrive frequently,
			 * so use a local counter to avoid doing
			 * atomic operations for each packet arriving.
			 */