On Sat, Jun 14, 2003 at 02:38:43AM -0700, David S. Miller wrote:
> From: Andi Kleen <ak@xxxxxxx>
> Date: Sat, 14 Jun 2003 11:36:30 +0200
> Also when you do use it generically you will hopefully
> discard some old code (like the rt cache?) which may make
> up for the additional bloat. But until that happens having
> both even when not needed doesn't make too much sense.
> The rtcache will likely be retained as a flow cache lookup
> miss handler even once we use the flowcache for all lookups.
> Actually, that entire area is in flux, I still do not know the
> fate of the rtcache even without the flow cache :)
In that case you could really apply the patch. It doesn't close
any future options for you, just makes live a bit better for
some users today.
> > How about working on making the xfrm layer more lean instead? :)
> My last proposal for this (using hlists in the hash tables) was
> rejected, so I don't see much chance to do this.
> Because hlists cannot retain the behavior we need, specifically
> because we need the ability to add to the tail.
> If it's some in-kernel-image table, why not dynamically allocate the
> table in question?
Allocating it at first lookup would be racy (would need a nasty spinlock
at least). It may be possible at first policy setup, but it's not guaranteed
can still get two 32K continuous areas. You could fall back to vmalloc I guess.
Allocating it at bootup would be equivalent to the current BSS allocation.
Advantage of the dynamic allocation is that it would work for vendor kernels