Re: + shmem-fix-faulting-into-a-hole-while-its-punched-take-2.patch added to -mm tree

[Hugh Dickins <hughd@google.com>]

On Thu, 10 Jul 2014, Sasha Levin wrote:
> On 07/10/2014 03:06 PM, Hugh Dickins wrote:
> > On Thu, 10 Jul 2014, Sasha Levin wrote:
> >> > On 07/10/2014 02:52 PM, Hugh Dickins wrote:
> >>> > > On Thu, 10 Jul 2014, Sasha Levin wrote:
> >>>>> > >> > On 07/10/2014 01:55 PM, Hugh Dickins wrote:
> >>>>>>>>> > >>>> > >> And finally, (not) holding the i_mmap_mutex:
> >>>>>>> > >>> > > I don't understand what prompts you to show this particular task.
> >>>>>>> > >>> > > I imagine the dump shows lots of other tasks which are waiting to get an
> >>>>>>> > >>> > > i_mmap_mutex, and quite a lot of other tasks which are neither waiting
> >>>>>>> > >>> > > for nor holding an i_mmap_mutex.
> >>>>>>> > >>> > > 
> >>>>>>> > >>> > > Why are you showing this one in particular?  Because it looks like the
> >>>>>>> > >>> > > one you fingered yesterday?  But I didn't see a good reason to finger
> >>>>>>> > >>> > > that one either.
> >>>>> > >> > 
> >>>>> > >> > There are a few more tasks like this one, my criteria was tasks that lockdep
> >>>>> > >> > claims were holding i_mmap_mutex, but are actually not.
> >>> > > You and Vlastimil enlightened me yesterday that lockdep shows tasks as
> >>> > > holding i_mmap_mutex when they are actually waiting to get i_mmap_mutex.
> >>> > > Hundreds of those in yesterday's log, hundreds of them in today's.
> >> > 
> >> > What if we move lockdep's acquisition point to after it actually got the
> >> > lock?
> >> > 
> >> > We'd miss deadlocks, but we don't care about them right now. Anyways, doesn't
> >> > lockdep have anything built in to allow us to separate between locks which
> >> > we attempt to acquire and locks that are actually acquired?
> >> > 
> >> > (cc PeterZ)
> >> > 
> >> > We can treat locks that are in the process of being acquired the same as
> >> > acquired locks to avoid races, but when we print something out it would
> >> > be nice to have annotation of the read state of the lock.
> > I certainly hope someone can work on improving that.  I imagine it would
> > be easy, and well worth doing.  But won't be looking into it myself.
> 
> I'd be happy to work on that, just want Peter to confirm that there's no reason
> that this is missing right now.

Great, thanks.  And for this bug (and many others?) it would also be very
helpful if those waiting on a mutex show the current mutex owner's pid.

Don't worry about getting a final mergeable patch, covering all lock types:
just something hacked up to show that i_mmap_mutex owner would help a lot.

But be careful, maybe owner is corrupted, or contains a now-invalid
address, or points to something no longer a task_struct.

Hugh

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>