Re: RCU latency regression in 2.6.16-rc1

[Lee Revell <rlrevell@joe-job.com>]

On Sun, 2006-01-29 at 21:11 -0800, Paul E. McKenney wrote:
> > Well, if as a bonus we are able to expand the size of the hash
> table, it 
> > could be very very good : As of today, the boot time sizing of this
> hash 
> > table is somewhat problematic.
> > 
> > If the size is expanded by a 2 factor (or a power of too), can your 
> > proposal works ?
> 
> Yep!!!
> 
> Add the following:
> 
> o       Add a size variable for each of the tables.  It works best
>         if the per-table state is stored with the table itself, for
>         example:
> 
>         struct hashtbl {
>                 int size;
>                 int fvl;
>                 struct hash_param params;
>                 struct list_head buckets[0];
>         };
> 
> o       When switching tables, allocate a new one of the desired size
>         and free up the non-current one.  (But remember to wait at
> least
>         one grace period after the last switch before starting
> this!!!)
> 
> o       Compute hash parameters suitable for the new table size.
> 
> o       Continue as before.
> 
> Note that you are not restricted to power-of-two expansion -- the
> hash parameters should handle any desired difference, and in fact
> handle contraction as well as expansion. 

I'd be glad to test any patches whenever someone gets around to working
on this.

Lee