On Tue, 3 Mar 2009, James Bottomley wrote:

On Tue, 2009-03-03 at 17:08 +0100, Jan Engelhardt wrote:

quoted

On Tuesday 2009-03-03 16:21, James Bottomley wrote:

quoted

quoted

quoted

$ slabtop
  OBJS ACTIVE  USE OBJ SIZE  SLABS OBJ/SLAB CACHE SIZE NAME                   
818616 818616 100%    0.16K  34109       24    136436K sgpool-8
253692 253692 100%    0.62K  42282        6    169128K sgpool-32
 52017  52016  99%    2.50K  17339        3    138712K sgpool-128
 26220  26219  99%    0.31K   2185       12      8740K sgpool-16
  8927   8574  96%    0.03K     79      113       316K size-32

Looks like a leak, by failing to call scsi_release_buffers()
somehow. (Which was changed recently)

Firstly, I have to say I don't see this in the mainline tree, so could
you try that with your setup just to verify (git head at 2.6.29-rc6).

Yes, looking at the rt patch (in broken-out it's in origin.diff),
it seems a bit obvious - the scsi_release_buffers is not called anymore:

OK, this is a bad patch, so just revert it.  It was posted to linux-scsi
initially in this form before the author posted a new one with the
missing release buffers added.  It looks like the first incarnation got
pulled into the -rt tree for some reasons.

So the real question is why does the -rt tree even have patches not in
the vanilla SCSI tree?  This type of cockup clearly demonstrates why
it's a bad idea.

My bad. I was playing with that to get rid of the aic7xxx wreckage on
one of my test boxen and forgot to remove it.

Thanks,

	tglx