Re: [PATCH v3 0/3] locking/spinlock_debug: Change it to a mostly fair lock

From: Paul E. McKenney
Date: Tue Feb 07 2017 - 14:53:20 EST


On Tue, Feb 07, 2017 at 02:46:31PM -0500, Waiman Long wrote:
> On 02/07/2017 04:48 AM, Ingo Molnar wrote:
> > * Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> >> On Thu, Feb 02, 2017 at 09:15:26AM -0500, Waiman Long wrote:
> >>> v2->v3:
> >>> - Keep the original v1 patches but move patch 3 of v2 in front so
> >>> as to disable GENERIC_LOCKBREAK when DEBUG_LOCK_ALLOC is on.
> >>>
> >>> v1->v2:
> >>> - Pack lockup and break_lock into a single 4-byte slot so as not
> >>> to in increase spinlock size when GENERIC_LOCKBREAK is
> >>> on. Hopefully that will be enough to fix a frame size too large
> >>> warning in 0-day build.
> >>> - Add a new patch to disable GENERIC_LOCKBREAK when DEBUG_LOCK_ALLOC
> >>> is on.
> >>>
> >>> The current debug spinlock implementation is a TATAS unfair lock. This
> >>> can occasionally lead to system lockup with a debug kernel because
> >>> of the unfairness of the lock rather than inherent locking problem.
> >>>
> >>> This patch set changes the debug spinlock implementation to a
> >>> mostly fair spinlock based on the MCS lock similar to what is done
> >>> in qspinlock. It also includes a patch that disable GENERIC_LOCKBREAK
> >>> when DEBUG_LOCK_ALLOC is on.
> >>
> >> An alternative is to just delete the thing entirely..
> >>
> >> Ingo, much of what this thing does seems to be superseded by both
> >> lockdep and a reliable NMI watchdog. Is there still value in
> >> spinlock_debug?
> > So there's still early stages when the NMI watchdog is not running, and
> > spinlock-debug can detect lockups in raw locks that lockdep does not cover, right?
> >
> > But yeah ... it would simplify things all around, so I'm not unsympathetic to the
> > idea...
> >
> > I've Cc:-ed a few other locking gents.
> >
> > Thanks,
> >
> > Ingo
>
> I have no problem deleting the debug_spinlock code entirely. I can
> update the patch to delete the code if you guys think that is the right
> thing to do.
>
> Cheers,
> Longman

My usual question is "how often does the spinlock_debug code find a
problem that would be hard to find otherwise?" Probably unanswerable
given the nature of Linux-kernel development, but I figured I would ask
anyway. ;-)

Thanx, Paul