Re: PATCH: Further aacraid work

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

On Thu, 2004-06-17 at 16:25, Andi Kleen wrote:
> On Thu, Jun 17, 2004 at 04:54:14PM -0400, Alan Cox wrote:
> > I would rather see it below the I/O layer for things like AMD64. The
> > reason I say this is that many drivers would suffer from iommu merging not
> > gain, and others may have limits.
> > 
> > Something like
> > 
> > 	new_sglist = sg_squash(old_sglist, [target max segments], [max per seg])
> > 
> > could be used by drivers when appropriate to hand back a better sg list
> > (or if not possible the existing one). That would put control rather closer
> > to the driver. 
> 
> My understanding was that it was too late in the driver because the SG lists
> are already sized, because higher layer manage this. That is why
> the BIO_VMERGE_BOUNDARY define is checked by BIO, not the driver.
> 
> The input of sg_squash should not be an already mapped list
> (that would be too costly) better would be probably
> a pci_map_sg_merge() with hints that tries to merge and other
> than that works like normal pci_map_sg()

Well, that's why we don't enable BIO_VMERGE_BOUNDARY nor any of the merging
at the BIO level, at least we didn't on ppc64 when I last worked on the code.

We let the BIO generate things that will always fit. We just have pci_map_*
do merging on a "best it can" basis. Most of the time, it does end up merging
a lot.

I don't think any driver control would help much here ...

Ben.