Re: [PATCH 0/3] map multiple blocks in get_block() and mpage_readpages()

[Badari Pulavarty <pbadari@us.ibm.com>]

On Wed, 2006-02-22 at 16:12 +0100, christoph wrote:
..
> Thanks Badari, with that interface changes the mpage_readpage changes
> look a lot nicer than my original version.  I'd like to second
> the request to put it into -mm. 
> 
> And if the namesys folks could try out whether this works for their
> reiser4 requirements it'd be nice.  If you have an even faster
> ->readpages I'd be interested in that secrete souce receipe for
> further improvement to mpage_readpages.

I don't have any secret receipes, but I was thinking of re-organizing
the code a little. Complexity is due to "confused" case and 
"blocksize < pagesize" cases + going in-and-out of the worker routine
with stored state.

I am thinking of having a "fast path" which doesn't deal with any
of those and "slow" path to deal with all that non-sense.
Something like ..

mpage_readpages()
{
	if (block-size < page-size)
		slow_path;

	while (nr-pages) {
		if (get_block(bh)) 
			slow_path;
		if (uptodate(bh))
			slow_path;
		while (bh.b_size) {
			if (not contig)
				submit bio();
			add all the pages we can to bio();
			bh.b_size -= size-of-pages-added;
			nr_pages -= count-of-pages-added;
		}

	}
}

slow_path is going to be slow & ugly. How important is to handle
1k, 2k filesystems efficiently ? Should I try ?
	
Thanks,
Badari