James Bottomley wrote:
> On Mon, 2006-02-06 at 16:45 +0000, Hugh Dickins wrote:
>> But I'd also want James or someone to clarify the paragraph
>> "Please note that the sg cannot be mapped again if it has been mapped once.
>>  The mapping process is allowed to destroy information in the sg."
>> which I took as explicitly allowing what x86_64 does in gart_map_sg.
>> I thought James had a scenario in mind which demands this wholesale
>> destruction, but it seems not; and I now read that first sentence as
>> saying the sg must be unmapped before it can be mapped a second time,
>> not that it can only be mapped once.
>>
>> And add a paragraph explaining that really the one array of scatterlist
>> entries should be regarded as two arrays of possibly different lengths,
>> the page,offset,length array and the dma_address,dma_length array:
>> because entries of the latter may be coalesced, so that in the end
>> the dma_address in a scatterlength entry may bear no relation to the
>> page pointer in that same entry, but to the page pointer in a later entry.
>> Though it gets hard to explain given that not all architectures coalesce,
>> so may not even have a separate dma_length field; or use different naming.
>> If you can express this better than I, please do!
> 
> Yes, I added that piece after the x86_64 problem.  The original x86_64
> bug was that you couldn't do dma_map_sg() then dma_unmap_sg() then
> dma_map_sg() on the same scatterlist (the map was destroying information
> which wasn't restored on the unmap, so the second map produced an
> incorrect scatterlist), which was causing a bug in all SCSI drivers
> (because that's the way SCSI requeueing works).

Here is a second try at a documentation update. Any takers to fixup x86_64?

Brian

-- 
Brian King
eServer Storage I/O
IBM Linux Technology Center