From mboxrd@z Thu Jan 1 00:00:00 1970 From: Pekka Pietikainen Subject: b44 and 4g4g Date: Mon, 31 May 2004 23:21:04 +0300 Sender: netdev-bounce@oss.sgi.com Message-ID: <20040531202104.GA8301@ee.oulu.fi> Mime-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 Return-path: To: netdev@oss.sgi.com Content-Disposition: inline Errors-to: netdev-bounce@oss.sgi.com List-Id: netdev.vger.kernel.org After diagnosing https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=118165 for a while it seems that bcm4401 is quite broken when DMA:ing from/to addresses that are > 1GB. This only is a problem with > 1GB of physical memory and when using a 4:4 split. It appears the network core code doesn't have a way of dealing with this situation cleanly, Arjan found this: static inline int illegal_highdma(struct net_device *dev, struct sk_buff *skb) { int i; if (dev->features & NETIF_F_HIGHDMA) return 0; for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) if (skb_shinfo(skb)->frags[i].page >= highmem_start_page) return 1; return 0; } which doesn't take the pci dma mask into account at all, which it probably should. (Please correct me if I got details wrong, this is a bit over my head :-) ) Now, close your eyes and imagine you never saw this patch (which does go a bit overkill with the GFP_DMA's just-in-case), which demonstrates what's required to make even receives work :-( (trying to transmit causes the chip to do a link down/link up thing after a few packets, which would suggest it gets very confused). --- ../linux-2.6.6-1.391/net/core/dev.c 2004-05-26 15:29:13.000000000 +0300 +++ net/core/dev.c 2004-05-31 20:05:45.000000000 +0300 @@ -1197,7 +1197,7 @@ skb->ip_summed = CHECKSUM_NONE; return skb; } - +#if 0 #ifdef CONFIG_HIGHMEM /* Actually, we should eliminate this check as soon as we know, that: * 1. IOMMU is present and allows to map all the memory. @@ -1220,7 +1220,8 @@ #else #define illegal_highdma(dev, skb) (0) #endif - +#endif +#define illegal_highdma(dev, skb) (1) extern void skb_release_data(struct sk_buff *); /* Keep head the same: replace data */ --- ../linux-2.6.6-1.391/drivers/net/b44.c 2004-05-26 15:29:11.000000000 +0300 +++ drivers/net/b44.c 2004-05-31 22:38:20.249808096 +0300 @@ -634,7 +634,7 @@ src_map = &bp->rx_buffers[src_idx]; dest_idx = dest_idx_unmasked & (B44_RX_RING_SIZE - 1); map = &bp->rx_buffers[dest_idx]; - skb = dev_alloc_skb(RX_PKT_BUF_SZ); + skb = __dev_alloc_skb(RX_PKT_BUF_SZ,GFP_DMA); if (skb == NULL) return -ENOMEM; @@ -762,7 +762,7 @@ struct sk_buff *copy_skb; b44_recycle_rx(bp, cons, bp->rx_prod); - copy_skb = dev_alloc_skb(len + 2); + copy_skb = __dev_alloc_skb(len + 2,GFP_DMA); if (copy_skb == NULL) goto drop_it_no_recycle; @@ -1077,13 +1077,13 @@ int size; size = B44_RX_RING_SIZE * sizeof(struct ring_info); - bp->rx_buffers = kmalloc(size, GFP_KERNEL); + bp->rx_buffers = kmalloc(size, GFP_DMA|GFP_KERNEL); if (!bp->rx_buffers) goto out_err; memset(bp->rx_buffers, 0, size); size = B44_TX_RING_SIZE * sizeof(struct ring_info); - bp->tx_buffers = kmalloc(size, GFP_KERNEL); + bp->tx_buffers = kmalloc(size, GFP_DMA|GFP_KERNEL); if (!bp->tx_buffers) goto out_err; memset(bp->tx_buffers, 0, size); @@ -1727,7 +1727,13 @@ pci_set_master(pdev); - err = pci_set_dma_mask(pdev, (u64) 0xffffffff); + err = pci_set_dma_mask(pdev, (u64) 0x3fffffff); + if (err) { + printk(KERN_ERR PFX "No usable DMA configuration, " + "aborting.\n"); + goto err_out_free_res; + } + err = pci_set_consistent_dma_mask(pdev, (u64) 0x3fffffff); if (err) { printk(KERN_ERR PFX "No usable DMA configuration, " "aborting.\n");