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

[Mikulas Patocka <mpatocka@redhat.com>]

On Fri, 11 Jul 2008, FUJITA Tomonori wrote:

> On Thu, 10 Jul 2008 17:56:08 -0400 (EDT)
> Mikulas Patocka <mpatocka@redhat.com> wrote:
>
>> When I thought about it more, I realized that this accounting of virtual
>> segments in I/O layer can't work correctly at all. If an architecture
>> defines symbols BIOVEC_VIRT_MERGEABLE and BIOVEC_VIRT_OVERSIZE, it
>> declares that it's IOMMU must merge any two regions satisfying these
>> conditions. But in an IOMMU, it is impossible to guarantee, because:
>
> Yeah, IOMMUs can't guarantee that. The majority of architectures set
> BIO_VMERGE_BOUNDARY to 0 so they don't hit this, I think.

Yes, the architectures without IOMMU don't hit this problem.

>> * the bus address is allocated basiclly randomly, so we can hit
>> dev->dma_parms->segment_boundary_mask any time. This will prevent virtual
>> merging from happenning. I/O layer doesn't know the bus address at the
>> time it merges requests, so it can't predict when this happens.
>>
>> * the IOMMU isn't guaranteed to find a continuous space in it's bus
>> address space. If it skips over already mapped regions, it can't perform
>> virtual merging.
>>
>> * when creating the mapping, we can hit per-device limit
>> "dev->dma_parms->max_segment_size" --- but the I/O layer checks only
>> against global limit BIOVEC_VIRT_OVERSIZE. (this last issue is fixable,
>> the previous two are not).
>
> I think that the block layer can handle this properly via
> q->max_segment_size. We have the same value at two different
> places. Yeah, it's not good...
>
>
> BTW, inia100_template sets sg_tablesize to SG_ALL. If the controller
> has at most 32 SG entries per request, we need to fix that.

Later, it sets that to shost->sg_tablesize = TOTAL_SG_ENTRY; I don't know 
why in inia100_template there is SG_ALL.

Mikulas