Re: [PATCH tip/core/urgent] fix several lockdep splats, allow multiple splats

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, Apr 30, 2010 at 02:12:01PM -0400, Mathieu Desnoyers wrote:
> * Paul E. McKenney (paulmck@linux.vnet.ibm.com) wrote:
> > On Fri, Apr 30, 2010 at 12:16:45PM -0400, Valdis.Kletnieks@vt.edu wrote:
> > > On Fri, 30 Apr 2010 11:33:34 EDT, Mathieu Desnoyers said:
> > > 
> > > > I recommend creating a kernel command line parameter that would tweak
> > > > the number of messages printed by lockdep. The default would indeed by 1
> > > > message, but people in a debugging marathon can specify a larger value
> > > > so they won't have to reboot between each individual lockdep error.
> > > 
> > > Yeah, that would rock for development kernels - playing whack-a-mole with
> > > a half-dozen new lockdep whinges can easily stretch out for quite some time.
> > 
> > The RCU-lockdep splats are a bit different in nature than the
> > deadlock-related splats that lockdep normally prints.  The RCU-lockdep
> > splats are transient in nature, and it is easy to apply WARN_ON_ONCE().
> > In contrast, if you permit multiple deadlock-related lockdep splats,
> > you tend to get lots of warnings about the same deadlock cycle.
> > 
> > So how about an additional kernel configuration variable, default
> > disabled, perhaps named CONFIG_PROVE_RCU_MULTIPLE, that allows a
> > single boot to see multiple messages?  Unlike the dyntick-idle
> > WARN_ON()s that generated multi-gigabyte console logs in a great
> > hurry, I haven't yet seen excessive quantities of RCU-lockdep splats,
> > so I don't see the need for an integer limit.
> > 
> > Thoughts?
> 
> Ideally we don't want to flood the console with thousands of instances
> of the same RCU-lockdep splat (think of a missing read lock on a common
> code path). Therefore I think keeping an integer limit is relevant here.
> I agree that this integer limit could be selected by a CONFIG_ option
> rather than by a kernel parameter, as it will typically only be used on
> development kernels with "kernel hacking" enabled anyway. There is not
> much point in bloating the kernel code with an extra debug-only kernel
> parameter parsing.

We already limit via WARN_ON_ONCE(), and there are fewer than 500 lines
of code in the kernel that can give RCU lockdep splats, so I really believe
that we are OK without an overall limit for the foreseeable future.

							Thanx, Paul