Re: [PATCH][v2] dax: Fix missed wakeup during dax entry invalidation

[Vivek Goyal]

On Mon, Apr 19, 2021 at 04:39:47PM -0400, Vivek Goyal wrote:
> On Mon, Apr 19, 2021 at 12:48:58PM -0700, Dan Williams wrote:
> > On Mon, Apr 19, 2021 at 11:45 AM Vivek Goyal <vgoyal@xxxxxxxxxx> wrote:
> > >
> > > This is V2 of the patch. Posted V1 here.
> > >
> > > https://lore.kernel.org/linux-fsdevel/20210416173524.GA1379987@xxxxxxxxxx/
> > >
> > > Based on feedback from Dan and Jan, modified the patch to wake up
> > > all waiters when dax entry is invalidated. This solves the issues
> > > of missed wakeups.
> > 
> > Care to send a formal patch with this commentary moved below the --- line?
> > 
> > One style fixup below...
> > 
> > >
> > > I am seeing missed wakeups which ultimately lead to a deadlock when I am
> > > using virtiofs with DAX enabled and running "make -j". I had to mount
> > > virtiofs as rootfs and also reduce to dax window size to 256M to reproduce
> > > the problem consistently.
> > >
> > > So here is the problem. put_unlocked_entry() wakes up waiters only
> > > if entry is not null as well as !dax_is_conflict(entry). But if I
> > > call multiple instances of invalidate_inode_pages2() in parallel,
> > > then I can run into a situation where there are waiters on
> > > this index but nobody will wait these.
> > >
> > > invalidate_inode_pages2()
> > >   invalidate_inode_pages2_range()
> > >     invalidate_exceptional_entry2()
> > >       dax_invalidate_mapping_entry_sync()
> > >         __dax_invalidate_entry() {
> > >                 xas_lock_irq(&xas);
> > >                 entry = get_unlocked_entry(&xas, 0);
> > >                 ...
> > >                 ...
> > >                 dax_disassociate_entry(entry, mapping, trunc);
> > >                 xas_store(&xas, NULL);
> > >                 ...
> > >                 ...
> > >                 put_unlocked_entry(&xas, entry);
> > >                 xas_unlock_irq(&xas);
> > >         }
> > >
> > > Say a fault in in progress and it has locked entry at offset say "0x1c".
> > > Now say three instances of invalidate_inode_pages2() are in progress
> > > (A, B, C) and they all try to invalidate entry at offset "0x1c". Given
> > > dax entry is locked, all tree instances A, B, C will wait in wait queue.
> > >
> > > When dax fault finishes, say A is woken up. It will store NULL entry
> > > at index "0x1c" and wake up B. When B comes along it will find "entry=0"
> > > at page offset 0x1c and it will call put_unlocked_entry(&xas, 0). And
> > > this means put_unlocked_entry() will not wake up next waiter, given
> > > the current code. And that means C continues to wait and is not woken
> > > up.
> > >
> > > This patch fixes the issue by waking up all waiters when a dax entry
> > > has been invalidated. This seems to fix the deadlock I am facing
> > > and I can make forward progress.
> > >
> > > Reported-by: Sergio Lopez <slp@xxxxxxxxxx>
> > > Signed-off-by: Vivek Goyal <vgoyal@xxxxxxxxxx>
> > > ---
> > >  fs/dax.c |   12 ++++++------
> > >  1 file changed, 6 insertions(+), 6 deletions(-)
> > >
> > > Index: redhat-linux/fs/dax.c
> > > ===================================================================
> > > --- redhat-linux.orig/fs/dax.c  2021-04-16 14:16:44.332140543 -0400
> > > +++ redhat-linux/fs/dax.c       2021-04-19 11:24:11.465213474 -0400
> > > @@ -264,11 +264,11 @@ static void wait_entry_unlocked(struct x
> > >         finish_wait(wq, &ewait.wait);
> > >  }
> > >
> > > -static void put_unlocked_entry(struct xa_state *xas, void *entry)
> > > +static void put_unlocked_entry(struct xa_state *xas, void *entry, bool wake_all)
> > >  {
> > >         /* If we were the only waiter woken, wake the next one */
> > >         if (entry && !dax_is_conflict(entry))
> > > -               dax_wake_entry(xas, entry, false);
> > > +               dax_wake_entry(xas, entry, wake_all);
> > >  }
> > >
> > >  /*
> > > @@ -622,7 +622,7 @@ struct page *dax_layout_busy_page_range(
> > >                         entry = get_unlocked_entry(&xas, 0);
> > >                 if (entry)
> > >                         page = dax_busy_page(entry);
> > > -               put_unlocked_entry(&xas, entry);
> > > +               put_unlocked_entry(&xas, entry, false);
> > 
> > I'm not a fan of raw true/false arguments because if you read this
> > line in isolation you need to go read put_unlocked_entry() to recall
> > what that argument means. So lets add something like:
> > 
> > /**
> >  * enum dax_entry_wake_mode: waitqueue wakeup toggle
> >  * @WAKE_NEXT: entry was not mutated
> >  * @WAKE_ALL: entry was invalidated, or resized
> >  */
> > enum dax_entry_wake_mode {
> >         WAKE_NEXT,
> >         WAKE_ALL,
> > }
> > 
> > ...and use that as the arg for dax_wake_entry(). So I'd expect this to
> > be a 3 patch series, introduce dax_entry_wake_mode for
> > dax_wake_entry(), introduce the argument for put_unlocked_entry()
> > without changing the logic, and finally this bug fix. Feel free to add
> > 'Fixes: ac401cc78242 ("dax: New fault locking")' in case you feel this
> > needs to be backported.
> 
> Hi Dan,
> 
> I will make changes as you suggested and post another version.
> 
> I am wondering what to do with dax_wake_entry(). It also has a boolean
> parameter wake_all. Should that be converted as well to make use of
> enum dax_entry_wake_mode?

oops, you already mentioned dax_wake_entry(). I read too fast. Sorry for
the noise

Vivek