On Thu, Mar 09, 2006 at 01:55:51PM -0700, Scott Long wrote: > Ruslan Ermilov wrote: > >On Thu, Mar 09, 2006 at 01:57:21PM -0500, Brian Fundakowski Feldman wrote: > > > >>On Tue, Mar 07, 2006 at 02:32:00PM +0100, Pawel Jakub Dawidek wrote: > >> > >>>On Mon, Mar 06, 2006 at 07:39:11PM -0500, Brian Fundakowski Feldman > >>>wrote: > >>>+> On Wed, Feb 22, 2006 at 10:21:05AM +0000, Pawel Jakub Dawidek wrote: > >>>+> > pjd 2006-02-22 10:21:05 UTC > >>>+> > > >>>+> > FreeBSD src repository > >>>+> > > >>>+> > Modified files: > >>>+> > sys/geom/raid3 g_raid3.c > >>>+> > Log: > >>>+> > Do not use bio structure after g_io_deliver(), it may not longer > >>>by valid. > >>>+> > > >>>+> > Found and fixed by: Vsevolod Lobko <[EMAIL PROTECTED]> > >>>+> > MFC after: 3 days > >>>+> > >>>+> I actually found and fixed it half a year ago... could you please > >>>+> integrate the rest of the fixes from my changes back then? A > >>>+> short-term low-memory deadlock is still possible (observed in > >>>+> practice). I think the changes also improve readability -- see > >>>+> for example the reason r1.46 existed. > >>> > >>>Heh. I own you apology. I haven't had time to work on graid3 back then > >>>and I also overlooked fix of this very problem. > >>> > >>>I integrated you fixes to my last patch which I'm planning to commit > >>>after receiving some feedback: > >>> > >>> http://people.freebsd.org/~pjd/patches/graid3.patch > >> > >>Everything seems good to go, running -CURRENT's graid3 + these changes > >>on 6-STABLE; are there particular areas where you think I should try > >>testing for improved performance? > >> > > > >I was fighting a problem today when our geom_cache attempts to fill > >256K of bio_data coming from graid3, and this obviously fails because > >graid3 only allocates memory for first 64k from its 64k UMA zone. > >This patch doesn't fix this problem. I have MAXPHYS bumped to 512K, > >which is important to reproduce the problem. Can you please fix this? > >Temporarily, I've added 128/256/512 zones to satisfy my needs. > > > > > >Cheers, > > MAXPHYS was never meant to go above 128K. You'll run a very real risk > of exhausting KVA under load. And, all block drivers for the past 10 > years have been written with the assumption that MAXPHYS will never be > above 128K. In other words, don't shoot yourself in the foot and then > cry that it hurts =-) > We've fixed some defines that abused MAXPHYS to mean 64k. I'm compiling the kernel with default MAXPHYS now, but still, there's no 128k zone in graid3 to allocate from.
Cheers, -- Ruslan Ermilov [EMAIL PROTECTED] FreeBSD committer
pgprRiFaNS1fK.pgp
Description: PGP signature