Re: [PATCH] locking/lockdep: Disable KASAN instrumentation of lockdep.c

[Waiman Long]

On 2/3/25 6:24 AM, Peter Zijlstra wrote:
> On Fri, Jan 31, 2025 at 04:47:06PM -0500, Waiman Long wrote:
> > On 1/31/25 11:50 AM, Waiman Long wrote:
> > > Both KASAN and LOCKDEP are commonly enabled in building a debug kernel.
> > > Each of them can significantly slow down the speed of a debug kernel.
> > > Enabling KASAN instrumentation of the LOCKDEP code will further slow
> > > thing down.
> > >
> > > Since LOCKDEP is a high overhead debugging tool, it will never get
> > > enabled in a production kernel. The LOCKDEP code is also pretty mature
> > > and is unlikely to get major changes. There is also a possibility of
> > > recursion similar to KCSAN. As the small advantage of enabling KASAN
> > > instrumentation to catch potential memory access error is probably
> > > not worth the drawback of further slowing down a debug kernel, disable
> > > KASAN instrumentation to enable a debug kernel to gain a little bit of
> > > speed back.
> > >
> > > With a debug kernel with both LOCKDEP and KASAN enabled running on a
> > > 2-socket 144-thread system, the time to do a "make -j144" kernel build
> > > was 18m40.641s. After applying this patch, the parallel kernel build
> > > time was reduced to 17m35.136s. This is a reduction of about 66s (5.8%).
> > >
> > > Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
> > > ---
> > >    kernel/locking/Makefile | 1 +
> > >    1 file changed, 1 insertion(+)
> > >
> > > diff --git a/kernel/locking/Makefile b/kernel/locking/Makefile
> > > index 0db4093d17b8..8a588b0227b1 100644
> > > --- a/kernel/locking/Makefile
> > > +++ b/kernel/locking/Makefile
> > > @@ -6,6 +6,7 @@ KCOV_INSTRUMENT		:= n
> > >    obj-y += mutex.o semaphore.o rwsem.o percpu-rwsem.o
> > >    # Avoid recursion lockdep -> sanitizer -> ... -> lockdep.
> > > +KASAN_SANITIZE_lockdep.o := n
> > >    KCSAN_SANITIZE_lockdep.o := n
> > >    ifdef CONFIG_FUNCTION_TRACER
> > The rationale behind this patch is due to the fact that a similar configured
> > PREEMPT_RT debug kernel is found to be about 3 times slower than the non-RT
> > debug kernel. For the test same system, the parallel build runtime is
> > 59m56.722s. After applying this patch, it is reduced to 38m3.348s. Its more
> > than 1/3 reduction is more than I would have expected. So the lockdep code
> > is much more heavily used in a PREEMPT_RT debug kernel.
> Perhaps put that in the changelog instead?
>
> Its not like RT is this secret out of tree project :-)
>
> Also, any quick clues as to what causes the extra lockdep overhead?
> Initially I thought perhaps local-lock, but that should also cause
> lockdep on !RT builds.

Yes, I am planning to update the patch with more RT debug kernel 
performance data.

As to why, my guess is that the average nesting depth will be higher 
because spin_lock_irq* no longer disable IRQ and there is an extra wait 
lock underneath the rt-mutex. Also the increase in the number of 
sleep-wake cycles because of the sleeping lock nature of rt-spinlock may 
be a contributing factor.

Cheers,
Longman

Cheers,
Longman