On Mon, Dec 12, 2011 at 08:43:57AM -0500, Ryan C. England wrote:
> On Mon, Dec 12, 2011 at 4:00 AM, Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> > On Mon, Dec 12, 2011 at 06:13:11AM +0100, Andi Kleen wrote:
> > > BTW I suppose it wouldn't be all that hard to add more stacks and
> > > switch to them too, similar to what the 32bit do_IRQ does.
> > > Perhaps XFS could just allocate its own stack per thread
> > > (or maybe only if it detects some specific configuration that
> > > is known to need much stack)
> > That's possible, but rather complex, I think.
> > > It would need to be per thread if you could sleep inside them.
> > Yes, we'd need to sleep, do IO, possibly operate within a
> > transaction context, etc, and a workqueue handles all these cases
> > without having to do anything special. Splitting the stack at a
> > logical point is probably better, such as this patch:
> > http://oss.sgi.com/archives/xfs/2011-07/msg00443.html
> Is it possible to apply this patch to my current installation? We use this
> box in production and the reboots that we're experiencing are an
Not easily. The problem with a backport is that the workqueue
infrastructure changed around 2.6.36, allowing workqueues to act
like an (almost) infinite pool of worker threads and so by using a
workqueue we can have effectively unlimited numbers of concurrent
allocations in progress at once.
The workqueue implementation in 2.6.32 only allows a single work
instance per workqueue thread, and so even with per-CPU worker
threads, would only allow one allocation at a time per CPU. This
adds additional serialisation within a filesystem, between
filesystem and potentially adds new deadlock conditions as well.
So it's not exactly obvious whether it can be backported in a sane
manner or not.
> Is there is a walkthrough on how to apply this patch? If not, could your
> provide the steps necessary to apply successfully? I would greatly
> appreciate it.
It would probably need redesigning and re-implementing from scratch
because of the above reasons. It'd then need a lot of testing and
review. As a workaround, you might be better off doing what Andi
first suggested - recompiling your kernel to use 16k stacks.