Re: How to map high memory for block io

[Jens Axboe <axboe@suse.de>]

On Fri, Jan 27 2006, Russell King wrote:
> On Fri, Jan 27, 2006 at 09:04:33PM +0100, Pierre Ossman wrote:
> > Russell King wrote:
> > > I don't see what the problem is.  A sg entry is a list of struct page
> > > pointers, an offset, and a size.  As such, it can't describe a transfer
> > > which crosses a page because such a structure does not imply that one
> > > struct page follows another struct page.
> > 
> > If the pages do not strictly follow each other then there is a lot of
> > broken code in the kernel. drivers/mmc/mmci.c and drivers/block/ub.c
> > being two occurences since both assume they can access the entire entry
> > through a single mapping.
> 
> We don't make that assumption.  What we do is:
> 
> - map the current sg using kmap_atomic()
> - copy up to sg->length into or out of that mapping
> - unmap current sg
> - if we have reached the end of this sg, move on to the next
> 
> What this means is that we assume sg->offset + sg->length <= PAGE_SIZE
> in all cases, which is the same assumption architecture DMA code makes.
> If that's invalid, there's likely to be a lot of architecture DMA support
> which is broken.

That is definitely valid, same goes for the bio_vec structure. They map
_a_ page, after all :-)

-- 
Jens Axboe