[PATCH] e1000, errata 2{3, 4} - possible EEH or memory corruption when DMA crosses a 64k boundary

[PATCH] e1000, errata 2{3, 4} - possible EEH or memory corruption when DMA crosses a 64k boundary

[Olaf Hering]

We have this patch in SLES9 SP1.
I asked google about 'fix for errata 23, cant cross 64kB boundary', and
it shhows such a patch is also part of RH 2.6.9.
It still applies to current Linus tree.
Can you check wether this is still required for the current driver?


References: SUSE48368 LTC12567

Need to check 64k boundary on DMA address as well.

We also need to have 64k boundary checking on the DMA address
that comes back from pci_map_single(). This address is what will
be passed to the adapter on ppc64 for it to DMA into.  It's the
address that the adapter sees which will trip erratum 23. 

The so patched driver passed a quick netperf run and a weekend
long stress test.

diff -puN drivers/net/e1000-new/e1000_main.c~64k-align-check-dma-suse drivers/net/e1000-new/e1000_main.c
--- linux-2.6.5-7.127/drivers/net/e1000-new/e1000_main.c~64k-align-check-dma-suse	Wed Dec  8 16:55:46 2004
+++ linux-2.6.5-7.127-moilanen/drivers/net/e1000-new/e1000_main.c	Thu Dec  9 15:46:04 2004
@@ -2579,6 +2579,29 @@ e1000_alloc_rx_buffers(struct e1000_adap
 						  adapter->rx_buffer_len,
 						  PCI_DMA_FROMDEVICE);
 
+		if(adapter->hw.mac_type == e1000_82545 ||
+		   adapter->hw.mac_type == e1000_82546)   {
+			/* fix for errata 23, cant cross 64kB boundary */
+			begin = (unsigned long)buffer_info->dma;
+			end   = (unsigned long)(adapter->rx_buffer_len) - 1;
+
+			if(!e1000_check_64k_alignment(adapter, begin, end)) {
+
+				DPRINTK(RX_ERR,ERR,"dma align check failed: "
+				        "begin: 0x%lx, end: 0x%lx\n", begin, end);
+
+				dev_kfree_skb(skb);
+				buffer_info->skb = NULL;
+				
+				pci_unmap_single(pdev,
+						 buffer_info->dma,
+						 adapter->rx_buffer_len,
+						 PCI_DMA_FROMDEVICE);
+
+				break; /* while !buffer_info->skb */
+			}
+		}
+
 		rx_desc = E1000_RX_DESC(*rx_ring, i);
 		rx_desc->buffer_addr = cpu_to_le64(buffer_info->dma);

Re: [PATCH] e1000, errata 2{3, 4} - possible EEH or memory corruption when DMA crosses a 64k boundary

[Jon Mason]

On Tuesday 01 February 2005 01:27 pm, Olaf Hering wrote:
> 
> We have this patch in SLES9 SP1.
> I asked google about 'fix for errata 23, cant cross 64kB boundary', and
> it shhows such a patch is also part of RH 2.6.9.
> It still applies to current Linus tree.
> Can you check wether this is still required for the current driver?

This patch is still lacking from the latest e1000 driver.  Intel has the patch in their queue, so it will be needed until they release their next version of the e1000 driver.

RE: [PATCH] e1000, errata 2{3, 4} - possible EEH or memory corruption when DMA crosses a 64k boundary

[Venkatesan, Ganesh]

The patch attached to your mail does not seem to be complete. Did the
mail application truncate?

In any case, this fix is required in e1000. It is *not* in the latest
driver that is released but *is* in the driver that is lined up for
release in a couple of weeks.

Thanks,
Ganesh.

>-----Original Message-----
>From: Olaf Hering [mailto:olh@suse.de]
>Sent: Tuesday, February 01, 2005 11:28 AM
>To: Venkatesan, Ganesh; netdev@oss.sgi.com
>Cc: linuxppc64-dev@ozlabs.org
>Subject: [PATCH] e1000, errata 2{3,4} - possible EEH or memory
corruption
>when DMA crosses a 64k boundary
>
>
>We have this patch in SLES9 SP1.
>I asked google about 'fix for errata 23, cant cross 64kB boundary', and
>it shhows such a patch is also part of RH 2.6.9.
>It still applies to current Linus tree.
>Can you check wether this is still required for the current driver?
>
>
>References: SUSE48368 LTC12567
>
>Need to check 64k boundary on DMA address as well.
>
>We also need to have 64k boundary checking on the DMA address
>that comes back from pci_map_single(). This address is what will
>be passed to the adapter on ppc64 for it to DMA into.  It's the
>address that the adapter sees which will trip erratum 23.
>
>The so patched driver passed a quick netperf run and a weekend
>long stress test.
>
>diff -puN drivers/net/e1000-new/e1000_main.c~64k-align-check-dma-suse
>drivers/net/e1000-new/e1000_main.c
>---
linux-2.6.5-7.127/drivers/net/e1000-new/e1000_main.c~64k-align-check-
>dma-suse	Wed Dec  8 16:55:46 2004
>+++ linux-2.6.5-7.127-moilanen/drivers/net/e1000-new/e1000_main.c
Thu Dec
>9 15:46:04 2004
>@@ -2579,6 +2579,29 @@ e1000_alloc_rx_buffers(struct e1000_adap
>
adapter->rx_buffer_len,
> 						  PCI_DMA_FROMDEVICE);
>
>+		if(adapter->hw.mac_type == e1000_82545 ||
>+		   adapter->hw.mac_type == e1000_82546)   {
>+			/* fix for errata 23, cant cross 64kB boundary
*/
>+			begin = (unsigned long)buffer_info->dma;
>+			end   = (unsigned long)(adapter->rx_buffer_len)
- 1;
>+
>+			if(!e1000_check_64k_alignment(adapter, begin,
end)) {
>+
>+				DPRINTK(RX_ERR,ERR,"dma align check
failed: "
>+				        "begin: 0x%lx, end: 0x%lx\n",
begin, end);
>+
>+				dev_kfree_skb(skb);
>+				buffer_info->skb = NULL;
>+
>+				pci_unmap_single(pdev,
>+						 buffer_info->dma,
>+						 adapter->rx_buffer_len,
>+						 PCI_DMA_FROMDEVICE);
>+
>+				break; /* while !buffer_info->skb */
>+			}
>+		}
>+
> 		rx_desc = E1000_RX_DESC(*rx_ring, i);
> 		rx_desc->buffer_addr = cpu_to_le64(buffer_info->dma);
>

Re: [PATCH] e1000, errata 2{3, 4} - possible EEH or memory corruption when DMA crosses a 64k boundary

[Olaf Hering]

 On Tue, Feb 01, Venkatesan, Ganesh wrote:

> In any case, this fix is required in e1000. It is *not* in the latest
> driver that is released but *is* in the driver that is lined up for
> release in a couple of weeks.

Thats good enough, just that things dont get lost. There are probably a
few separate patches for each problem found.
We are still in the process of sorting out our +2k patch mess.