Re: [PATCH v3 2/2][BNX2]: Add iSCSI support to BNX2 devices.

[FUJITA Tomonori <tomof@acm.org>]

CC'ed Jens, James, and linux-scsi.

On Thu, 27 Sep 2007 03:31:55 -0400
Jeff Garzik <jeff@garzik.org> wrote:

> FUJITA Tomonori wrote:
> > Yeah, we could nicely handle lld's restrictions (especially with
> > stacking devices). But iommu code needs only max_segment_size and
> > seg_boundary_mask, right? If so, the first simple approach to add two
> > values to device structure is not so bad, I think.
> 
> (replying to slightly older email in the thread)
> (added benh, since we've discussed this issue in the past)
> 
> dumb question, what happened to seg_boundary_mask?

I'll work on it too after finishing max_seg_size.


> If you look at drivers/ata/libata-core.c:ata_fill_sg(), you will note 
> that we split s/g segments after DMA-mapping.  Looking at libata LLDD's, 
> you will also note judicious use of ATA_DMA_BOUNDARY (0xffff).

I know the workaround since I fixed libata's sg chaining patch.


> It was drilled into my head by James and benh that I cannot rely on the 
> DMA boundary + block/scsi + dma_map_sg() to ensure that my S/G segments 
> never cross a 64K boundary, a legacy IDE requirement.  Thus the 
> additional code in ata_fill_sg() to split S/G segments straddling 64K, 
> in addition to setting dma boundary to 0xffff.

I think that the block layer can handle both max_segment_size and
seg_boundary_mask properly (and SCSI-ml just uses the block layer). So
if we fix iommu, then we can remove a workaround to fix sg lists in
llds.


> A key problem I was hoping would be solved with your work here was the 
> elimination of that post dma_map_sg() split.

Yeah, that's my goal too.


> If I understood James and Ben correctly, one of the key problems was 
> always in communicating libata's segment boundary needs to the IOMMU layers?
> 
> 	Jeff
>