Re: [PATCH] dma-debug: hash_bucket_find needs to allow for offsets within an entry

["Roedel, Joerg" <Joerg.Roedel@amd.com>]

On Wed, Jul 20, 2011 at 10:32:22AM -0400, Neil Horman wrote:
> On Wed, Jul 20, 2011 at 03:29:25PM +0200, Roedel, Joerg wrote:
> > You are right. We need to scan
> > 
> > 	0 <= idx <= hash_fn(rstart)
> > 
> > Probably we can fix that with a better hash-function. Any ideas? Using
> > the device is not an option because then all entries would end up in
> > only a few buckets. This will impact scanning performance too much.
> > 
> Unfortunately I don't have any ideas for a better hash function here, but I had
> been thinking about fixing this in add_dma_entry.  We could detect there that a
> debug entry to be added crossed one or more hash bucket boundaries, and, if it
> did, split it along those boundaries into multiple entries, hashing each of them
> in separately.  The check_unmap and check_sync routines would of course then
> potentially have to do multiple lookups as well to ensure that they found all of
> the correct entries to validate/remove.  It would work in all cases, but it
> might be overkill.  What do you think?

Interesting. I discussed that with a colleague an hour ago and he came
up with the same idea :-)
I like it because we still need to scan only one hash-bucket, so this
seems like the best solution.

> > For now, the partial syncs seem to happen rarely enough so that we can
> > make it a slow-path. It is probably the best to do the exact scan first
> > and do the full scan only if exact-scan failed (until we come up with a
> > better solution).
> > 
> Agreed, if you don't like my above idea, I'll get to work on this this
> afternoon.

I think this idea is a better solution.

Thanks,

	Joerg

-- 
AMD Operating System Research Center

Advanced Micro Devices GmbH Einsteinring 24 85609 Dornach
General Managers: Alberto Bozzo, Andrew Bowd
Registration: Dornach, Landkr. Muenchen; Registerger. Muenchen, HRB Nr. 43632