Re: [PATCH] workaround minor lockdep bug triggered by mm_take_all_locks

[Andrea Arcangeli <andrea@qumranet.com>]

On Mon, Aug 04, 2008 at 07:00:03PM -0700, Roland Dreier wrote:
>  > The point is that this is a runtime evaluation of lock orders, if
>  > runtime isn't the lucky one that reproduces the deadlock, it'll find
>  > nothing at all.
> 
> I think the point you miss is that lockdep can report a potential
> deadlock, even if the deadlock does not actually occur.  For example
> suppose there is an AB-BA deadlock somewhere.  For this to actually
> trigger, we have to have one CPU running the AB code path at exactly the
> moment another CPU runs the BA code path, with the right timing so one
> CPU holds A and tries to grab B while the other CPU already holds B.
> 
> With lockdep, we just have to have the AB code path run once at any
> point, and then the BA code path run at any later time (even days after
> the AB code path has released all the locks).  And then we get a
> warning dump that explains the exact potential deadlock.

Thanks a lot for the detailed explanation of check_noncircular. I
agree check_noncircular is surely a good argument not to get rid of
prove-locking as a whole. But check_noncircular is also a red-herring
in this context. It's not check_noncircular trapping here,
check_deadlock traps with false positives instead. The question is
what are those false positives buying us? To avoid a developer to
press sysrq+p or break on kgdb?

Let's focus on check_deadlock->print_deadlock_bug and somebody who's
not beyond the point please explain what print_deadlock_bug reports
that does not actually occur and why it's a good idea to change the
common code to accommodate for its false positives instead of getting
rid of it for good.