Re: [GIT PULL] dax fixes for 4.20-rc6

[Jan Kara]

On Sun 09-12-18 10:26:54, Dan Williams wrote:
> [ add Willy and Jan ]
> 
> On Sun, Dec 9, 2018 at 10:02 AM Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > On Sat, Dec 8, 2018 at 10:26 PM Williams, Dan J
> > <dan.j.williams@xxxxxxxxx> wrote:
> > >
> > >   git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm/nvdimm tags/dax-fixes-4.20-rc6
> >
> > What's going on with the odd non-exclusive exclusive wait?
> >
> >         prepare_to_wait_exclusive(wq, &ewait.wait, TASK_UNINTERRUPTIBLE);
> >         ...
> >         /*
> >          * Entry lock waits are exclusive. Wake up the next waiter since
> >          * we aren't sure we will acquire the entry lock and thus wake
> >          * the next waiter up on unlock.
> >          */
> >         if (waitqueue_active(wq))
> >                 __wake_up(wq, TASK_NORMAL, 1, &ewait.key);
> >
> > that seems to make little or no sense.
> >
> > Why isn't that prepare_to_wait_exclusive() just a regular
> > prepare_to_wait(), and then the whole "let's wake up anybody else" can
> > be removed?
> >
> > I've pulled it, but am awaiting explanation of what looks like some
> > pretty crazy code. I *suspect* it's a copy-and-paste situation where
> > you took the exclusive wait from somewhere else.
> 
> Yes, I believe that's true. In the other instances of waiting for an
> entry to be in unlocked there is a guarantee that the waiter will
> attain the lock and perform an unlock + wakeup. In the dax_lock_page()
> path there is the possibility that the inode dies before the lock is
> attained and a subsequent unlock sequence is not guaranteed. So, I
> believe the intent, Willy correct me if I am wrong, was to keep all
> waits "exclusive" for some sense of symmetry, but this one can and
> should be a non-exclusive wait.

Agreed. I didn't realize this when reviewing Matthew's patch and
misunderstood his comment to this end.

								Honza

-- 
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR