Re: [PATCH] smbfs async readpage

[Andrew Morton <akpm@zip.com.au>]

Urban Widmark wrote:
> 
> On Tue, 6 Aug 2002, Andrew Morton wrote:
> 
> > What you can do (please, I'd be interested in checking that it's
> > workable) is to create an smbfs_backing_dev_info, and arrange
> > (in smb_iget) for each new inode's mapping->backing_dev_info
> > to point at it.
> 
> Yes, the highest number of pages in action appears to be (mostly) limited
> by what I set ra_pages to. If I set ra_pages to 1 or 2 my printks still
> report up to 4 requests issued, but with 10 I never get more than 10.
> 40 gives something that looks like 40, without actually counting the lines
> ...

That sounds promising.

> 
> > Then implement some way of diddling smbfs_backing_dev_info.ra_pages
> > and voila, tunable readahead.
> 
> Other fs' seems to be using the default too. New 2.5 feature or does block
> device based filesystems get this value from somewhere else?
> (like the block device they live on?)

Yes, there's a backing_dev_info structure in each block request queue and
a pointer to that is placed in each address_space which is hosted
by that queue.

> If being able to tune readahead is important, I could imagine a desire to
> tune differently on a per-mount basis. So I'd stick it in the smbfs
> superblock and have a readahead=N mount option.

Tunability may be important for preventing bad behaviour rather
than for giving users something good ;)   Does readahead across
a n/w filesystem help much?