yellowfin.c

来自「linux 内核源代码」· C语言 代码 · 共 1,437 行 · 第 1/4 页

{
	struct yellowfin_private *yp = netdev_priv(dev);
	int i;

	yp->tx_full = 0;
	yp->cur_rx = yp->cur_tx = 0;
	yp->dirty_tx = 0;

	yp->rx_buf_sz = (dev->mtu <= 1500 ? PKT_BUF_SZ : dev->mtu + 32);

	for (i = 0; i < RX_RING_SIZE; i++) {
		yp->rx_ring[i].dbdma_cmd =
			cpu_to_le32(CMD_RX_BUF | INTR_ALWAYS | yp->rx_buf_sz);
		yp->rx_ring[i].branch_addr = cpu_to_le32(yp->rx_ring_dma +
			((i+1)%RX_RING_SIZE)*sizeof(struct yellowfin_desc));
	}

	for (i = 0; i < RX_RING_SIZE; i++) {
		struct sk_buff *skb = dev_alloc_skb(yp->rx_buf_sz);
		yp->rx_skbuff[i] = skb;
		if (skb == NULL)
			break;
		skb->dev = dev;		/* Mark as being used by this device. */
		skb_reserve(skb, 2);	/* 16 byte align the IP header. */
		yp->rx_ring[i].addr = cpu_to_le32(pci_map_single(yp->pci_dev,
			skb->data, yp->rx_buf_sz, PCI_DMA_FROMDEVICE));
	}
	yp->rx_ring[i-1].dbdma_cmd = cpu_to_le32(CMD_STOP);
	yp->dirty_rx = (unsigned int)(i - RX_RING_SIZE);

#define NO_TXSTATS
#ifdef NO_TXSTATS
	/* In this mode the Tx ring needs only a single descriptor. */
	for (i = 0; i < TX_RING_SIZE; i++) {
		yp->tx_skbuff[i] = NULL;
		yp->tx_ring[i].dbdma_cmd = cpu_to_le32(CMD_STOP);
		yp->tx_ring[i].branch_addr = cpu_to_le32(yp->tx_ring_dma +
			((i+1)%TX_RING_SIZE)*sizeof(struct yellowfin_desc));
	}
	/* Wrap ring */
	yp->tx_ring[--i].dbdma_cmd = cpu_to_le32(CMD_STOP | BRANCH_ALWAYS);
#else
{
	int j;

	/* Tx ring needs a pair of descriptors, the second for the status. */
	for (i = 0; i < TX_RING_SIZE; i++) {
		j = 2*i;
		yp->tx_skbuff[i] = 0;
		/* Branch on Tx error. */
		yp->tx_ring[j].dbdma_cmd = cpu_to_le32(CMD_STOP);
		yp->tx_ring[j].branch_addr = cpu_to_le32(yp->tx_ring_dma +
			(j+1)*sizeof(struct yellowfin_desc);
		j++;
		if (yp->flags & FullTxStatus) {
			yp->tx_ring[j].dbdma_cmd =
				cpu_to_le32(CMD_TXSTATUS | sizeof(*yp->tx_status));
			yp->tx_ring[j].request_cnt = sizeof(*yp->tx_status);
			yp->tx_ring[j].addr = cpu_to_le32(yp->tx_status_dma +
				i*sizeof(struct tx_status_words);
		} else {
			/* Symbios chips write only tx_errs word. */
			yp->tx_ring[j].dbdma_cmd =
				cpu_to_le32(CMD_TXSTATUS | INTR_ALWAYS | 2);
			yp->tx_ring[j].request_cnt = 2;
			/* Om pade ummmmm... */
			yp->tx_ring[j].addr = cpu_to_le32(yp->tx_status_dma +
				i*sizeof(struct tx_status_words) +
				&(yp->tx_status[0].tx_errs) -
				&(yp->tx_status[0]));
		}
		yp->tx_ring[j].branch_addr = cpu_to_le32(yp->tx_ring_dma +
			((j+1)%(2*TX_RING_SIZE))*sizeof(struct yellowfin_desc));
	}
	/* Wrap ring */
	yp->tx_ring[++j].dbdma_cmd |= cpu_to_le32(BRANCH_ALWAYS | INTR_ALWAYS);
}
#endif
	yp->tx_tail_desc = &yp->tx_status[0];
	return;
}

static int yellowfin_start_xmit(struct sk_buff *skb, struct net_device *dev)
{
	struct yellowfin_private *yp = netdev_priv(dev);
	unsigned entry;
	int len = skb->len;

	netif_stop_queue (dev);

	/* Note: Ordering is important here, set the field with the
	   "ownership" bit last, and only then increment cur_tx. */

	/* Calculate the next Tx descriptor entry. */
	entry = yp->cur_tx % TX_RING_SIZE;

	if (gx_fix) {	/* Note: only works for paddable protocols e.g.  IP. */
		int cacheline_end = ((unsigned long)skb->data + skb->len) % 32;
		/* Fix GX chipset errata. */
		if (cacheline_end > 24  || cacheline_end == 0) {
			len = skb->len + 32 - cacheline_end + 1;
			if (skb_padto(skb, len)) {
				yp->tx_skbuff[entry] = NULL;
				netif_wake_queue(dev);
				return 0;
			}
		}
	}
	yp->tx_skbuff[entry] = skb;

#ifdef NO_TXSTATS
	yp->tx_ring[entry].addr = cpu_to_le32(pci_map_single(yp->pci_dev,
		skb->data, len, PCI_DMA_TODEVICE));
	yp->tx_ring[entry].result_status = 0;
	if (entry >= TX_RING_SIZE-1) {
		/* New stop command. */
		yp->tx_ring[0].dbdma_cmd = cpu_to_le32(CMD_STOP);
		yp->tx_ring[TX_RING_SIZE-1].dbdma_cmd =
			cpu_to_le32(CMD_TX_PKT|BRANCH_ALWAYS | len);
	} else {
		yp->tx_ring[entry+1].dbdma_cmd = cpu_to_le32(CMD_STOP);
		yp->tx_ring[entry].dbdma_cmd =
			cpu_to_le32(CMD_TX_PKT | BRANCH_IFTRUE | len);
	}
	yp->cur_tx++;
#else
	yp->tx_ring[entry<<1].request_cnt = len;
	yp->tx_ring[entry<<1].addr = cpu_to_le32(pci_map_single(yp->pci_dev,
		skb->data, len, PCI_DMA_TODEVICE));
	/* The input_last (status-write) command is constant, but we must
	   rewrite the subsequent 'stop' command. */

	yp->cur_tx++;
	{
		unsigned next_entry = yp->cur_tx % TX_RING_SIZE;
		yp->tx_ring[next_entry<<1].dbdma_cmd = cpu_to_le32(CMD_STOP);
	}
	/* Final step -- overwrite the old 'stop' command. */

	yp->tx_ring[entry<<1].dbdma_cmd =
		cpu_to_le32( ((entry % 6) == 0 ? CMD_TX_PKT|INTR_ALWAYS|BRANCH_IFTRUE :
					  CMD_TX_PKT | BRANCH_IFTRUE) | len);
#endif

	/* Non-x86 Todo: explicitly flush cache lines here. */

	/* Wake the potentially-idle transmit channel. */
	iowrite32(0x10001000, yp->base + TxCtrl);

	if (yp->cur_tx - yp->dirty_tx < TX_QUEUE_SIZE)
		netif_start_queue (dev);		/* Typical path */
	else
		yp->tx_full = 1;
	dev->trans_start = jiffies;

	if (yellowfin_debug > 4) {
		printk(KERN_DEBUG "%s: Yellowfin transmit frame #%d queued in slot %d.\n",
			   dev->name, yp->cur_tx, entry);
	}
	return 0;
}

/* The interrupt handler does all of the Rx thread work and cleans up
   after the Tx thread. */
static irqreturn_t yellowfin_interrupt(int irq, void *dev_instance)
{
	struct net_device *dev = dev_instance;
	struct yellowfin_private *yp;
	void __iomem *ioaddr;
	int boguscnt = max_interrupt_work;
	unsigned int handled = 0;

	yp = netdev_priv(dev);
	ioaddr = yp->base;

	spin_lock (&yp->lock);

	do {
		u16 intr_status = ioread16(ioaddr + IntrClear);

		if (yellowfin_debug > 4)
			printk(KERN_DEBUG "%s: Yellowfin interrupt, status %4.4x.\n",
				   dev->name, intr_status);

		if (intr_status == 0)
			break;
		handled = 1;

		if (intr_status & (IntrRxDone | IntrEarlyRx)) {
			yellowfin_rx(dev);
			iowrite32(0x10001000, ioaddr + RxCtrl);		/* Wake Rx engine. */
		}

#ifdef NO_TXSTATS
		for (; yp->cur_tx - yp->dirty_tx > 0; yp->dirty_tx++) {
			int entry = yp->dirty_tx % TX_RING_SIZE;
			struct sk_buff *skb;

			if (yp->tx_ring[entry].result_status == 0)
				break;
			skb = yp->tx_skbuff[entry];
			dev->stats.tx_packets++;
			dev->stats.tx_bytes += skb->len;
			/* Free the original skb. */
			pci_unmap_single(yp->pci_dev, le32_to_cpu(yp->tx_ring[entry].addr),
				skb->len, PCI_DMA_TODEVICE);
			dev_kfree_skb_irq(skb);
			yp->tx_skbuff[entry] = NULL;
		}
		if (yp->tx_full
			&& yp->cur_tx - yp->dirty_tx < TX_QUEUE_SIZE - 4) {
			/* The ring is no longer full, clear tbusy. */
			yp->tx_full = 0;
			netif_wake_queue(dev);
		}
#else
		if ((intr_status & IntrTxDone) || (yp->tx_tail_desc->tx_errs)) {
			unsigned dirty_tx = yp->dirty_tx;

			for (dirty_tx = yp->dirty_tx; yp->cur_tx - dirty_tx > 0;
				 dirty_tx++) {
				/* Todo: optimize this. */
				int entry = dirty_tx % TX_RING_SIZE;
				u16 tx_errs = yp->tx_status[entry].tx_errs;
				struct sk_buff *skb;

#ifndef final_version
				if (yellowfin_debug > 5)
					printk(KERN_DEBUG "%s: Tx queue %d check, Tx status "
						   "%4.4x %4.4x %4.4x %4.4x.\n",
						   dev->name, entry,
						   yp->tx_status[entry].tx_cnt,
						   yp->tx_status[entry].tx_errs,
						   yp->tx_status[entry].total_tx_cnt,
						   yp->tx_status[entry].paused);
#endif
				if (tx_errs == 0)
					break;	/* It still hasn't been Txed */
				skb = yp->tx_skbuff[entry];
				if (tx_errs & 0xF810) {
					/* There was an major error, log it. */
#ifndef final_version
					if (yellowfin_debug > 1)
						printk(KERN_DEBUG "%s: Transmit error, Tx status %4.4x.\n",
							   dev->name, tx_errs);
#endif
					dev->stats.tx_errors++;
					if (tx_errs & 0xF800) dev->stats.tx_aborted_errors++;
					if (tx_errs & 0x0800) dev->stats.tx_carrier_errors++;
					if (tx_errs & 0x2000) dev->stats.tx_window_errors++;
					if (tx_errs & 0x8000) dev->stats.tx_fifo_errors++;
				} else {
#ifndef final_version
					if (yellowfin_debug > 4)
						printk(KERN_DEBUG "%s: Normal transmit, Tx status %4.4x.\n",
							   dev->name, tx_errs);
#endif
					dev->stats.tx_bytes += skb->len;
					dev->stats.collisions += tx_errs & 15;
					dev->stats.tx_packets++;
				}
				/* Free the original skb. */
				pci_unmap_single(yp->pci_dev,
					yp->tx_ring[entry<<1].addr, skb->len,
					PCI_DMA_TODEVICE);
				dev_kfree_skb_irq(skb);
				yp->tx_skbuff[entry] = 0;
				/* Mark status as empty. */
				yp->tx_status[entry].tx_errs = 0;
			}

#ifndef final_version
			if (yp->cur_tx - dirty_tx > TX_RING_SIZE) {
				printk(KERN_ERR "%s: Out-of-sync dirty pointer, %d vs. %d, full=%d.\n",
					   dev->name, dirty_tx, yp->cur_tx, yp->tx_full);
				dirty_tx += TX_RING_SIZE;
			}
#endif

			if (yp->tx_full
				&& yp->cur_tx - dirty_tx < TX_QUEUE_SIZE - 2) {
				/* The ring is no longer full, clear tbusy. */
				yp->tx_full = 0;
				netif_wake_queue(dev);
			}

			yp->dirty_tx = dirty_tx;
			yp->tx_tail_desc = &yp->tx_status[dirty_tx % TX_RING_SIZE];
		}
#endif

		/* Log errors and other uncommon events. */
		if (intr_status & 0x2ee)	/* Abnormal error summary. */
			yellowfin_error(dev, intr_status);

		if (--boguscnt < 0) {
			printk(KERN_WARNING "%s: Too much work at interrupt, "
				   "status=0x%4.4x.\n",
				   dev->name, intr_status);
			break;
		}
	} while (1);

	if (yellowfin_debug > 3)
		printk(KERN_DEBUG "%s: exiting interrupt, status=%#4.4x.\n",
			   dev->name, ioread16(ioaddr + IntrStatus));

	spin_unlock (&yp->lock);
	return IRQ_RETVAL(handled);
}

/* This routine is logically part of the interrupt handler, but separated
   for clarity and better register allocation. */
static int yellowfin_rx(struct net_device *dev)
{
	struct yellowfin_private *yp = netdev_priv(dev);
	int entry = yp->cur_rx % RX_RING_SIZE;
	int boguscnt = yp->dirty_rx + RX_RING_SIZE - yp->cur_rx;

	if (yellowfin_debug > 4) {
		printk(KERN_DEBUG " In yellowfin_rx(), entry %d status %8.8x.\n",
			   entry, yp->rx_ring[entry].result_status);
		printk(KERN_DEBUG "   #%d desc. %8.8x %8.8x %8.8x.\n",
			   entry, yp->rx_ring[entry].dbdma_cmd, yp->rx_ring[entry].addr,
			   yp->rx_ring[entry].result_status);
	}

	/* If EOP is set on the next entry, it's a new packet. Send it up. */
	while (1) {
		struct yellowfin_desc *desc = &yp->rx_ring[entry];
		struct sk_buff *rx_skb = yp->rx_skbuff[entry];
		s16 frame_status;
		u16 desc_status;
		int data_size;
		u8 *buf_addr;

		if(!desc->result_status)
			break;
		pci_dma_sync_single_for_cpu(yp->pci_dev, le32_to_cpu(desc->addr),
			yp->rx_buf_sz, PCI_DMA_FROMDEVICE);
		desc_status = le32_to_cpu(desc->result_status) >> 16;
		buf_addr = rx_skb->data;
		data_size = (le32_to_cpu(desc->dbdma_cmd) -
			le32_to_cpu(desc->result_status)) & 0xffff;
		frame_status = le16_to_cpu(get_unaligned((__le16*)&(buf_addr[data_size - 2])));
		if (yellowfin_debug > 4)
			printk(KERN_DEBUG "  yellowfin_rx() status was %4.4x.\n",
				   frame_status);
		if (--boguscnt < 0)
			break;
		if ( ! (desc_status & RX_EOP)) {
			if (data_size != 0)
				printk(KERN_WARNING "%s: Oversized Ethernet frame spanned multiple buffers,"
					   " status %4.4x, data_size %d!\n", dev->name, desc_status, data_size);
			dev->stats.rx_length_errors++;
		} else if ((yp->drv_flags & IsGigabit)  &&  (frame_status & 0x0038)) {
			/* There was a error. */
			if (yellowfin_debug > 3)
				printk(KERN_DEBUG "  yellowfin_rx() Rx error was %4.4x.\n",
					   frame_status);