lkml.org 
[lkml]   [2004]   [Jan]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [BUG] x86_64 pci_map_sg modifies sg list - fails multiple map/unmaps
Hi,

Andi Kleen reworked pci-gart code.
Would you try Andi's x86-64 patch and see if the problems still exist ?

ftp://ftp.x86-64.org/pub/linux/v2.6/x86_64-2.6.1rc2-1.bz2

And also, Can you try with "iommu=noforce,nomerge" ?

Fixing the sg-list in the upperlayer (by re-creating) it in case of retry
worked for me. I am still not sure why you are running into "len==0"
panics.

Thanks,
Badari

Berkley Shands wrote:

> Running with the force segment merge OFF panics the processor after
> about 1000 scsi retries. the error given, also in pci-gart.c, is
> pci_map_area overflow 4096 bytes
> So a brain dead repair kills the kernel. Someone clearly needs to figure
> out where to correct the merge of the sg lists. A bit of doc on the iommu
> and the 4096 byte limit would be nice too :-)
>
> I see that is is the aborting of an SCB that causes the sg list halt.
>
> Jan 7 09:18:32 typhoon kernel: DevQ(0:6:0): 0 waiting
> Jan 7 09:18:32 typhoon kernel: (scsi0:A:2:0): SCB 0x46 - timed out
> Jan 7 09:18:32 typhoon kernel: Recovery SCB completes
> Jan 7 09:18:32 typhoon kernel: scsi0: Issued Channel A Bus Reset. 3 SCBs aborted
> Jan 7 09:18:46 typhoon kernel: Did it again, boss 0000:01:03.0
>
> Since the sg list merges into one i/o list, simply adding s->length = 4096
> back into the list seems to keep the kernel up. a better if slightly less
> stupid fix is to add up the remaining sg list lengths, and ajust
> the sg[0] entry to sum to the correct value.
>
> /* BUG_ON(s->length == 0); */
> if (! s->length)
> {
> unsigned long zero = sg[0].length;
> unsigned long remain = 0;
> int t = 0;
>
> BUG_ON(i != 1); /* some other error here */
>
> for (t = i + 1; t < nents; t++)
> remain += sg[t].length; /* collect remaining sizes */
> zero -= remain; /* deduct what is left on the list */
> sg[0].length = zero / 2;
> sg[1].length = zero / 2; /* allocate uniformly */
> size = zero / 2; /* reduce oversize first entry */
> printk(KERN_WARNING "Did it again, boss %s\n", dev->slot_name);
> }
>
> The better solution is to have the upper layer fix the sg list, or
> have some marker that the list was diddled, and save the old entries
> to put it back.
>
> berkley

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 13:59    [W:0.032 / U:1.048 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site