Re: [PATCH 1/2] lockdep: Introduce CROSSRELEASE_STACK_TRACE and make it not unwind as default

[Ingo Molnar]

* Byungchul Park <byungchul.park@xxxxxxx> wrote:

> On Thu, Oct 19, 2017 at 03:11:12PM +0900, Byungchul Park wrote:
> > On Thu, Oct 19, 2017 at 07:57:30AM +0200, Ingo Molnar wrote:
> > > 
> > > * Byungchul Park <byungchul.park@xxxxxxx> wrote:
> > > 
> > > > On Wed, Oct 18, 2017 at 12:09:44PM +0200, Ingo Molnar wrote:
> > > > > BTW., have you attempted limiting the depth of the stack traces? I suspect more 
> > > > > than 2-4 are rarely required to disambiguate the calling context.
> > > > 
> > > > I did it for you. Let me show you the result.
> > > > 
> > > > 1. No lockdep:				2.756558155 seconds time elapsed                ( +-  0.09% )
> > > > 2. Lockdep:					2.968710420 seconds time elapsed		( +-  0.12% )
> > > > 3. Lockdep + Crossrelease 5 entries:		3.153839636 seconds time elapsed                ( +-  0.31% )
> > > > 4. Lockdep + Crossrelease 3 entries:		3.137205534 seconds time elapsed                ( +-  0.87% )
> > > > 5. Lockdep + Crossrelease + This patch:	2.963669551 seconds time elapsed		( +-  0.11% )
> > > 
> > > I think the lockdep + crossrelease + full-stack numbers are missing?
> > 
> > Ah, the last version of crossrelease merged into vanilla, records 5
> > entries, since I thought it overloads too much if full stack is used,
> > and 5 entries are enough. Don't you think so?
> > 
> > > But yeah, looks like single-entry-stacktrace crossrelease only has a +0.2% 
> > > performance cost (with 0.1% noise), while lockdep itself has a +7.7% cost.
> > > 
> > > That's very reasonable and we can keep the single-entry cross-release feature 
> > > enabled by default as part of CONFIG_PROVE_LOCKING=y - assuming all the crashes 
> > 
> > BTW, is there any crash by cross-release I don't know? Of course, I know
> > cases of false positives, but I don't about crash.
> 
> Are you talking about the oops by 'null pointer dereference' by unwinder a
> few weeks ago?
> 
> At the time, cross-release was falsely accused. AFAIK, cross-release has
> not crashed system yet.

I'm talking about the crash fixed here:

  8b405d5c5d09: locking/lockdep: Fix stacktrace mess

Which was introduced by your patch:

  ce07a9415f26: locking/lockdep: Make check_prev_add() able to handle external stack_trace

... which was a preparatory patch for cross-release. So 'technically' it's not a 
cross-release crash, but was very much related. It even says so in the changelog:

  Actually crossrelease needs to do other than saving a stack_trace.
  So pass a stack_trace and callback to handle it, to check_prev_add().

... so let's not pretend it wasn't related, ok?

Thanks,

	Ingo