Eric Sandeen wrote:
There are still a few large stack users in xfs even when it's used by
itself - one that comes to mind is code that runs when you use xfs_fsr
(hm, I really must check in my fix for that....) And figuring out
callchain depth is tricky, so hard to say how these things will all add up.
I guess I would suggest that you just test it out on one non-critical
box, with a "normal" workload for you, and see how it fares. It's
tight, but in many cases it might be just fine.
I used default fedora core 3 kernels on my xfs only laptop for a quite
a while without problems. I went to building my own kernels only
because I ended up with a wireless card which requires one of those
darn windows only drivers, and windows drivers appear to eat
stack space for breakfast.
If so, then it's not hard to recompile the fc3 kernel (or a stock
kernel) for 8k stacks, just we're not too sure how risky the default
fc3 setup is.
You'll know after you figure out how often you hit problems. :)
We're also thinking of running RHEL AS4 instead of fc3 (they are very
similar), but that definitely needs a recompile as XFS isn't included in
the default AS4 kernel :-/
Yep... that's something I've been meaning to look into as well -
although the 4k stack issue remains.
AS 4.0 uses the 8K stack model, but yes, their filesystem selection
seems a little limited. Suse is actually looking pretty good right
now, and I have an ubuntu install running on a box right now just
to check it out.