cassini.c

来自「linux 内核源代码」· C语言 代码 · 共 2,335 行 · 第 1/5 页

			entry = TX_DESC_NEXT(ring, entry);
			continue;
		}

		/* however, we might get only a partial skb release. */
		count -= skb_shinfo(skb)->nr_frags +
			+ cp->tx_tiny_use[ring][entry].nbufs + 1;
		if (count < 0)
			break;

		if (netif_msg_tx_done(cp))
			printk(KERN_DEBUG "%s: tx[%d] done, slot %d\n",
			       cp->dev->name, ring, entry);

		skbs[entry] = NULL;
		cp->tx_tiny_use[ring][entry].nbufs = 0;

		for (frag = 0; frag <= skb_shinfo(skb)->nr_frags; frag++) {
			struct cas_tx_desc *txd = txds + entry;

			daddr = le64_to_cpu(txd->buffer);
			dlen = CAS_VAL(TX_DESC_BUFLEN,
				       le64_to_cpu(txd->control));
			pci_unmap_page(cp->pdev, daddr, dlen,
				       PCI_DMA_TODEVICE);
			entry = TX_DESC_NEXT(ring, entry);

			/* tiny buffer may follow */
			if (cp->tx_tiny_use[ring][entry].used) {
				cp->tx_tiny_use[ring][entry].used = 0;
				entry = TX_DESC_NEXT(ring, entry);
			}
		}

		spin_lock(&cp->stat_lock[ring]);
		cp->net_stats[ring].tx_packets++;
		cp->net_stats[ring].tx_bytes += skb->len;
		spin_unlock(&cp->stat_lock[ring]);
		dev_kfree_skb_irq(skb);
	}
	cp->tx_old[ring] = entry;

	/* this is wrong for multiple tx rings. the net device needs
	 * multiple queues for this to do the right thing.  we wait
	 * for 2*packets to be available when using tiny buffers
	 */
	if (netif_queue_stopped(dev) &&
	    (TX_BUFFS_AVAIL(cp, ring) > CAS_TABORT(cp)*(MAX_SKB_FRAGS + 1)))
		netif_wake_queue(dev);
	spin_unlock(&cp->tx_lock[ring]);
}

static void cas_tx(struct net_device *dev, struct cas *cp,
		   u32 status)
{
        int limit, ring;
#ifdef USE_TX_COMPWB
	u64 compwb = le64_to_cpu(cp->init_block->tx_compwb);
#endif
	if (netif_msg_intr(cp))
		printk(KERN_DEBUG "%s: tx interrupt, status: 0x%x, %llx\n",
			cp->dev->name, status, (unsigned long long)compwb);
	/* process all the rings */
	for (ring = 0; ring < N_TX_RINGS; ring++) {
#ifdef USE_TX_COMPWB
		/* use the completion writeback registers */
		limit = (CAS_VAL(TX_COMPWB_MSB, compwb) << 8) |
			CAS_VAL(TX_COMPWB_LSB, compwb);
		compwb = TX_COMPWB_NEXT(compwb);
#else
		limit = readl(cp->regs + REG_TX_COMPN(ring));
#endif
		if (cp->tx_old[ring] != limit)
			cas_tx_ringN(cp, ring, limit);
	}
}


static int cas_rx_process_pkt(struct cas *cp, struct cas_rx_comp *rxc,
			      int entry, const u64 *words,
			      struct sk_buff **skbref)
{
	int dlen, hlen, len, i, alloclen;
	int off, swivel = RX_SWIVEL_OFF_VAL;
	struct cas_page *page;
	struct sk_buff *skb;
	void *addr, *crcaddr;
	__sum16 csum;
	char *p;

	hlen = CAS_VAL(RX_COMP2_HDR_SIZE, words[1]);
	dlen = CAS_VAL(RX_COMP1_DATA_SIZE, words[0]);
	len  = hlen + dlen;

	if (RX_COPY_ALWAYS || (words[2] & RX_COMP3_SMALL_PKT))
		alloclen = len;
	else
		alloclen = max(hlen, RX_COPY_MIN);

	skb = dev_alloc_skb(alloclen + swivel + cp->crc_size);
	if (skb == NULL)
		return -1;

	*skbref = skb;
	skb_reserve(skb, swivel);

	p = skb->data;
	addr = crcaddr = NULL;
	if (hlen) { /* always copy header pages */
		i = CAS_VAL(RX_COMP2_HDR_INDEX, words[1]);
		page = cp->rx_pages[CAS_VAL(RX_INDEX_RING, i)][CAS_VAL(RX_INDEX_NUM, i)];
		off = CAS_VAL(RX_COMP2_HDR_OFF, words[1]) * 0x100 +
			swivel;

		i = hlen;
		if (!dlen) /* attach FCS */
			i += cp->crc_size;
		pci_dma_sync_single_for_cpu(cp->pdev, page->dma_addr + off, i,
				    PCI_DMA_FROMDEVICE);
		addr = cas_page_map(page->buffer);
		memcpy(p, addr + off, i);
		pci_dma_sync_single_for_device(cp->pdev, page->dma_addr + off, i,
				    PCI_DMA_FROMDEVICE);
		cas_page_unmap(addr);
		RX_USED_ADD(page, 0x100);
		p += hlen;
		swivel = 0;
	}


	if (alloclen < (hlen + dlen)) {
		skb_frag_t *frag = skb_shinfo(skb)->frags;

		/* normal or jumbo packets. we use frags */
		i = CAS_VAL(RX_COMP1_DATA_INDEX, words[0]);
		page = cp->rx_pages[CAS_VAL(RX_INDEX_RING, i)][CAS_VAL(RX_INDEX_NUM, i)];
		off = CAS_VAL(RX_COMP1_DATA_OFF, words[0]) + swivel;

		hlen = min(cp->page_size - off, dlen);
		if (hlen < 0) {
			if (netif_msg_rx_err(cp)) {
				printk(KERN_DEBUG "%s: rx page overflow: "
				       "%d\n", cp->dev->name, hlen);
			}
			dev_kfree_skb_irq(skb);
			return -1;
		}
		i = hlen;
		if (i == dlen)  /* attach FCS */
			i += cp->crc_size;
		pci_dma_sync_single_for_cpu(cp->pdev, page->dma_addr + off, i,
				    PCI_DMA_FROMDEVICE);

		/* make sure we always copy a header */
		swivel = 0;
		if (p == (char *) skb->data) { /* not split */
			addr = cas_page_map(page->buffer);
			memcpy(p, addr + off, RX_COPY_MIN);
			pci_dma_sync_single_for_device(cp->pdev, page->dma_addr + off, i,
					PCI_DMA_FROMDEVICE);
			cas_page_unmap(addr);
			off += RX_COPY_MIN;
			swivel = RX_COPY_MIN;
			RX_USED_ADD(page, cp->mtu_stride);
		} else {
			RX_USED_ADD(page, hlen);
		}
		skb_put(skb, alloclen);

		skb_shinfo(skb)->nr_frags++;
		skb->data_len += hlen - swivel;
		skb->truesize += hlen - swivel;
		skb->len      += hlen - swivel;

		get_page(page->buffer);
		frag->page = page->buffer;
		frag->page_offset = off;
		frag->size = hlen - swivel;

		/* any more data? */
		if ((words[0] & RX_COMP1_SPLIT_PKT) && ((dlen -= hlen) > 0)) {
			hlen = dlen;
			off = 0;

			i = CAS_VAL(RX_COMP2_NEXT_INDEX, words[1]);
			page = cp->rx_pages[CAS_VAL(RX_INDEX_RING, i)][CAS_VAL(RX_INDEX_NUM, i)];
			pci_dma_sync_single_for_cpu(cp->pdev, page->dma_addr,
					    hlen + cp->crc_size,
					    PCI_DMA_FROMDEVICE);
			pci_dma_sync_single_for_device(cp->pdev, page->dma_addr,
					    hlen + cp->crc_size,
					    PCI_DMA_FROMDEVICE);

			skb_shinfo(skb)->nr_frags++;
			skb->data_len += hlen;
			skb->len      += hlen;
			frag++;

			get_page(page->buffer);
			frag->page = page->buffer;
			frag->page_offset = 0;
			frag->size = hlen;
			RX_USED_ADD(page, hlen + cp->crc_size);
		}

		if (cp->crc_size) {
			addr = cas_page_map(page->buffer);
			crcaddr  = addr + off + hlen;
		}

	} else {
		/* copying packet */
		if (!dlen)
			goto end_copy_pkt;

		i = CAS_VAL(RX_COMP1_DATA_INDEX, words[0]);
		page = cp->rx_pages[CAS_VAL(RX_INDEX_RING, i)][CAS_VAL(RX_INDEX_NUM, i)];
		off = CAS_VAL(RX_COMP1_DATA_OFF, words[0]) + swivel;
		hlen = min(cp->page_size - off, dlen);
		if (hlen < 0) {
			if (netif_msg_rx_err(cp)) {
				printk(KERN_DEBUG "%s: rx page overflow: "
				       "%d\n", cp->dev->name, hlen);
			}
			dev_kfree_skb_irq(skb);
			return -1;
		}
		i = hlen;
		if (i == dlen) /* attach FCS */
			i += cp->crc_size;
		pci_dma_sync_single_for_cpu(cp->pdev, page->dma_addr + off, i,
				    PCI_DMA_FROMDEVICE);
		addr = cas_page_map(page->buffer);
		memcpy(p, addr + off, i);
		pci_dma_sync_single_for_device(cp->pdev, page->dma_addr + off, i,
				    PCI_DMA_FROMDEVICE);
		cas_page_unmap(addr);
		if (p == (char *) skb->data) /* not split */
			RX_USED_ADD(page, cp->mtu_stride);
		else
			RX_USED_ADD(page, i);

		/* any more data? */
		if ((words[0] & RX_COMP1_SPLIT_PKT) && ((dlen -= hlen) > 0)) {
			p += hlen;
			i = CAS_VAL(RX_COMP2_NEXT_INDEX, words[1]);
			page = cp->rx_pages[CAS_VAL(RX_INDEX_RING, i)][CAS_VAL(RX_INDEX_NUM, i)];
			pci_dma_sync_single_for_cpu(cp->pdev, page->dma_addr,
					    dlen + cp->crc_size,
					    PCI_DMA_FROMDEVICE);
			addr = cas_page_map(page->buffer);
			memcpy(p, addr, dlen + cp->crc_size);
			pci_dma_sync_single_for_device(cp->pdev, page->dma_addr,
					    dlen + cp->crc_size,
					    PCI_DMA_FROMDEVICE);
			cas_page_unmap(addr);
			RX_USED_ADD(page, dlen + cp->crc_size);
		}
end_copy_pkt:
		if (cp->crc_size) {
			addr    = NULL;
			crcaddr = skb->data + alloclen;
		}
		skb_put(skb, alloclen);
	}

	csum = (__force __sum16)htons(CAS_VAL(RX_COMP4_TCP_CSUM, words[3]));
	if (cp->crc_size) {
		/* checksum includes FCS. strip it out. */
		csum = csum_fold(csum_partial(crcaddr, cp->crc_size,
					      csum_unfold(csum)));
		if (addr)
			cas_page_unmap(addr);
	}
	skb->csum = csum_unfold(~csum);
	skb->ip_summed = CHECKSUM_COMPLETE;
	skb->protocol = eth_type_trans(skb, cp->dev);
	return len;
}


/* we can handle up to 64 rx flows at a time. we do the same thing
 * as nonreassm except that we batch up the buffers.
 * NOTE: we currently just treat each flow as a bunch of packets that
 *       we pass up. a better way would be to coalesce the packets
 *       into a jumbo packet. to do that, we need to do the following:
 *       1) the first packet will have a clean split between header and
 *          data. save both.
 *       2) each time the next flow packet comes in, extend the
 *          data length and merge the checksums.
 *       3) on flow release, fix up the header.
 *       4) make sure the higher layer doesn't care.
 * because packets get coalesced, we shouldn't run into fragment count
 * issues.
 */
static inline void cas_rx_flow_pkt(struct cas *cp, const u64 *words,
				   struct sk_buff *skb)
{
	int flowid = CAS_VAL(RX_COMP3_FLOWID, words[2]) & (N_RX_FLOWS - 1);
	struct sk_buff_head *flow = &cp->rx_flows[flowid];

	/* this is protected at a higher layer, so no need to
	 * do any additional locking here. stick the buffer
	 * at the end.
	 */
	__skb_insert(skb, flow->prev, (struct sk_buff *) flow, flow);
	if (words[0] & RX_COMP1_RELEASE_FLOW) {
		while ((skb = __skb_dequeue(flow))) {
			cas_skb_release(skb);
		}
	}
}

/* put rx descriptor back on ring. if a buffer is in use by a higher
 * layer, this will need to put in a replacement.
 */
static void cas_post_page(struct cas *cp, const int ring, const int index)
{
	cas_page_t *new;
	int entry;

	entry = cp->rx_old[ring];

	new = cas_page_swap(cp, ring, index);
	cp->init_rxds[ring][entry].buffer = cpu_to_le64(new->dma_addr);
	cp->init_rxds[ring][entry].index  =
		cpu_to_le64(CAS_BASE(RX_INDEX_NUM, index) |
			    CAS_BASE(RX_INDEX_RING, ring));

	entry = RX_DESC_ENTRY(ring, entry + 1);
	cp->rx_old[ring] = entry;

	if (entry % 4)
		return;

	if (ring == 0)
		writel(entry, cp->regs + REG_RX_KICK);
	else if ((N_RX_DESC_RINGS > 1) &&
		 (cp->cas_flags & CAS_FLAG_REG_PLUS))
		writel(entry, cp->regs + REG_PLUS_RX_KICK1);
}


/* only when things are bad */
static int cas_post_rxds_ringN(struct cas *cp, int ring, int num)
{
	unsigned int entry, last, count, released;
	int cluster;
	cas_page_t **page = cp->rx_pages[ring];

	entry = cp->rx_old[ring];

	if (netif_msg_intr(cp))
		printk(KERN_DEBUG "%s: rxd[%d] interrupt, done: %d\n",
		       cp->dev->name, ring, entry);

	cluster = -1;
	count = entry & 0x3;
	last = RX_DESC_ENTRY(ring, num ? entry + num - 4: entry - 4);
	released = 0;
	while (entry != last) {
		/* make a new buffer if it's still in use */
		if (page_count(page[entry]->buffer) > 1) {
			cas_page_t *new = cas_page_dequeue(cp);
			if (!new) {
				/* let the timer know that we need to
				 * do this again
				 */
				cp->cas_flags |= CAS_FLAG_RXD_POST(ring);
				if (!timer_pending(&cp->link_timer))
					mod_timer(&cp->link_timer, jiffies +
						  CAS_LINK_FAST_TIMEOUT);
				cp->rx_old[ring]  = entry;
				cp->rx_last[ring] = num ? num - released : 0;
				return -ENOMEM;
			}
			spin_lock(&cp->rx_inuse_lock);
			list_add(&page[entry]->list, &cp->rx_inuse_list);
			spin_unlock(&cp->rx_inuse_lock);
			cp->init_rxds[ring][entry].buffer =
				cpu_to_le64(new->dma_addr);
			page[entry] = new;

		}

		if (++count == 4) {
			cluster = entry;
			count = 0;
		}
		released++;
		entry = RX_DESC_ENTRY(ring, entry + 1);
	}
	cp->rx_old[ring] = entry;

	if (cluster < 0)
		return 0;

	if (ring == 0)
		writel(cluster, cp->regs + REG_RX_KICK);
	else if ((N_RX_DESC_RINGS > 1) &&
		 (cp->cas_flags & CAS_FLAG_REG_PLUS))
		writel(cluster, cp->regs + REG_PLUS_RX_KICK1);
	return 0;
}


/* process a completion ring. packets are set up in three basic ways:
 * small packets: should be copied header + data in single buffer.
 * large packets: header and data in a single buffer.
 * split packets: header in a separate buffer from data.
 *                data may be in multiple pages. data may be > 256
 *                bytes but in a single page.
 *
 * NOTE: RX page posting is done in this routine as well. while there's
 *       the capability of using multiple RX completion rings, it isn't
 *       really worthwhile due to the fact that the page posting will
 *       force serialization on the single descriptor ring.
 */
static int cas_rx_ringN(struct cas *cp, int ring, int budget)
{
	struct cas_rx_comp *rxcs = cp->init_rxcs[ring];
	int entry, drops;
	int npackets = 0;

	if (netif_msg_intr(cp))
		printk(KERN_DEBUG "%s: rx[%d] interrupt, done: %d/%d\n",
		       cp->dev->name, ring,
		       readl(cp->regs + REG_RX_COMP_HEAD),
		       cp->rx_new[ring]);

	entry = cp->rx_new[ring];
	drops = 0;
	while (1) {
		struct cas_rx_comp *rxc = rxcs + entry;
		struct sk_buff *skb;
		int type, len;
		u64 words[4];
		int i, dring;

		words[0] = le64_to_cpu(rxc->word1);
		words[1] = le64_to_cpu(rxc->word2);
		words[2] = le64_to_cpu(rxc->word3);
		words[3] = le64_to_cpu(rxc->word4);

		/* don't touch if still owned by hw */
		type = CAS_VAL(RX_COMP1_TYPE, words[0]);
		if (type == 0)
			break;

		/* hw hasn't cleared the zero bit yet */
		if (words[3] & RX_COMP4_ZERO) {
			break;
		}

		/* get info on the packet */
		if (words[3] & (RX_COMP4_LEN_MISMATCH | RX_COMP4_BAD)) {
			spin_lock(&cp->stat_lock[ring]);
			cp->net_stats[ring].rx_errors++;
			if (words[3] & RX_COMP4_LEN_MISMATCH)
				cp->net_stats[ring].rx_length_errors++;
			if (words[3] & RX_COMP4_BAD)
				cp->net_stats[ring].rx_crc_errors++;
			spin_unlock(&cp->stat_lock[ring]);

			/* We'll just return it to Cassini. */
		drop_it:
			spin_lock(&cp->stat_lock[r