Re: [11/36] Use page_cache_xxx in fs/buffer.c

[Jens Axboe <jens.axboe@oracle.com>]

On Fri, Aug 31 2007, Christoph Lameter wrote:
> On Fri, 31 Aug 2007, Jens Axboe wrote:
> 
> > > A DMA boundary cannot be crossed AFAIK. The compound pages are aligned to 
> > > the power of two boundaries and the page allocator will not create pages 
> > > that cross the zone boundaries.
> > 
> > With a 64k page and a dma boundary of 0x7fff, that's two segments.
> 
> Ok so DMA memory restrictions not conforming to the DMA zones? The 
> example is a bit weird. DMA only to the first 32k of memory? If the limit 
> would be higher like 16MB then we would not have an issue. Is there really 
> a device that can only do I/O to the first 32k of memory?

They have nothing to do with each other, you are mixing things up. It
has nothing to do with the device being able to dma into that memory or
not, we have fine existing infrastructure to handle that. But different
hardware have different characteristics on what a single segment is. You
can say "a single segment cannot cross a 32kb boundary". So from the
example above, your single 64k page may need to be split into two
segments. Or it could have a maximum segment size of 32k, in which case
it would have to be split as well.

Do you see what I mean now?

> How do we split that up today? We could add processing to submit_bio
> to check for the boundary and create two bios.

But we do not split them up today - see what I wrote! Today we impose
the restriction that a device must be able to handle a single "normal"
page, and if it can't do that, it has to split it up itself.

But yes, you would have to create some out-of-line function to use
bio_split() until you have chopped things down enough. It's not a good
thing for performance naturally, but if we consider this a "just make it
work" fallback, I don't think it's too bad. You want to make a note of
that it is happening though, so people realize that it is happening.

> > > submit_bh() is used to submit a single buffer and I think that was
> > > our main concern here.
> > 
> > And how large can that be?
> 
> As large as mkxxxfs allowed it to be. For XFS and extX with the
> current patchset 32k is the limit (64k with the fixes to ext2) but a
> new filesystem could theoretically use a larger blocksize.

OK, since it goes direct to bio anyway, it can be handled there.

-- 
Jens Axboe