Re: [PATCH v4 72/73] xfs: Convert mru cache to XArray

[Byungchul Park]

On Fri, Dec 08, 2017 at 09:22:16AM +1100, Dave Chinner wrote:
> On Thu, Dec 07, 2017 at 11:06:34AM -0500, Theodore Ts'o wrote:
> > On Wed, Dec 06, 2017 at 06:06:48AM -0800, Matthew Wilcox wrote:
> > > > Unfortunately for you, I don't find arguments along the lines of
> > > > "lockdep will save us" at all convincing.  lockdep already throws
> > > > too many false positives to be useful as a tool that reliably and
> > > > accurately points out rare, exciting, complex, intricate locking
> > > > problems.
> > > 
> > > But it does reliably and accurately point out "dude, you forgot to take
> > > the lock".  It's caught a number of real problems in my own testing that
> > > you never got to see.
> > 
> > The problem is that if it has too many false positives --- and it's
> > gotten *way* worse with the completion callback "feature", people will
> > just stop using Lockdep as being too annyoing and a waste of developer
> > time when trying to figure what is a legitimate locking bug versus
> > lockdep getting confused.
> > 
> > <Rant>I can't even disable the new Lockdep feature which is throwing
> > lots of new false positives --- it's just all or nothing.</Rant>
> > 
> > Dave has just said he's already stopped using Lockdep, as a result.
> 
> This is compeltely OT, but FYI I stopped using lockdep a long time
> ago.  We've spend orders of magnitude more time and effort to shut
> up lockdep false positives in the XFS code than we ever have on
> locking problems that lockdep has uncovered. And still lockdep
> throws too many false positives on XFS workloads to be useful to me.
> 
> But it's more than that: I understand just how much lockdep *doesn't
> check* and that means *I know I can't rely on lockdep* for potential
> deadlock detection. e.g.  it doesn't cover semaphores, which means

Hello,

I'm careful in saying the following since you seem to feel not good at
crossrelease and even lockdep. Now that cross-release has been
introduced, semaphores can be covered as you might know. Actually, all
general waiters can.

> it has zero coverage of the entire XFS metadata buffer subsystem and
> the complex locking orders we have for metadata updates.
> 
> Put simply: lockdep doesn't provide me with any benefit, so I don't
> use it...

Sad..

> Cheers,
> 
> Dave.
> -- 
> Dave Chinner
> david@xxxxxxxxxxxxx