Re: [SUGGESTION]: drop virtual merge accounting in I/O requests

[FUJITA Tomonori]

On Tue, 15 Jul 2008 20:10:42 -0700 (PDT)
David Miller <davem@xxxxxxxxxxxxx> wrote:

> From: FUJITA Tomonori <fujita.tomonori@xxxxxxxxxxxxx>
> Date: Tue, 15 Jul 2008 02:45:03 +0900
> 
> > dma_4u_map_sg() has:
> > 
> > if (segstart != s) {
> > 	/* We cannot merge if:
> > 	 * - allocated dma_addr isn't contiguous to previous allocation
> > 	 */
> > 	if ((dma_addr != dma_next) ||
> > 	    (outs->dma_length + s->length > max_seg_size) ||
> > 	    (is_span_boundary(out_entry, base_shift,
> > 		/* Can't merge: create a new segment */
> > 		segstart = s;
> > 		outcount++;
> > 		outs = sg_next(outs);
> > 
> > So if the IOMMU allocated dma_addr isn't contiguous to previous
> > allocation, it might not merge segments that the block layer expected
> > the IOMMU to merge.
> > 
> > We need kinda two phase merging code such as the old SPARC64 IOMMU
> > code and PARISC IOMMUs though I like the new simple SPARC64 IOMMU
> > code.
> 
> I see.
> 
> I wonder if all that complexity is really worth it.  Also, all of this
> IOMMU allocation and mapping code runs under a spinlock with hw IRQs
> disabled.
> 
> More and more I'm seeing that it's likely better to remove the VMERGE
> code.  I can't see what it really buys us anymore, and to make it work
> requires quite a large amount of complexity in the IOMMU layer.

Agreed, especially with modern HBAs, the VMERGE accounting isn't
useful, I think. The recent IOMMU implementations, Intel VT-d and AMD
virtualization one, don't do even virtual merging.

I'm fine with removing the VMERGE accounting in the block layer if
Jens and the users are happy about it.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/