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

[Andrea Arcangeli <andrea@qumranet.com>]

On Mon, Aug 04, 2008 at 04:38:16PM -0700, Arjan van de Ven wrote:
> interesting; lockdep has been working for the last.. 2 1/2 years at
> least, and I don't remember seeing bugreports against it from you that
> would describe it as totally non-functional.

I reported it to Peter. If you see David's email, I guess it can be
implied that I wasn't the only one aware that prove-locking made
certain systems non functional, I thought it was widespread knowledge
maybe not.

It's amazing that things seem to have improved on that side, it surely
gives me more confidence in prove-locking!

> Oh well.. seems you're rather preoccupied about it; that's ok, you're
> entitled to your opinion even if I don't agree with it ;-)

So let me understand better: your opinion is that all of lockdep is
useful, not just the AB BA detection?

By reading the source again after 11 months to me it still looks
check_deadlock() only has knowledge of the current context. It loops
over the task struct checking all the locks of the current task!
Combine the great feature that check_deadlock provides, with crashing
at boot, and I hope that better explains my feeling about
lockdep-prove-locking.

This check_deadlock() thing is the real core of my dislike of
prove-locking! The check_noncircular part I totally agree it's useful
now that I see it works differently than check_deadlock (when I read
it last time I thought it worked the same as check_deadlock).

check_noncircular being useful doesn't automatically make
check_deadlock useful.

And incidentally it's exactly this check_deadlock part that is
trapping on my code and that is now requiring silly changes to the
common code (the ones I did) or to make the common code even more
complex (what Peter is planning I guess).