diff mbox series

dax: Fix missed wakeup in put_unlocked_entry()

Message ID 20210416173524.GA1379987@redhat.com (mailing list archive)
State Superseded
Headers show
Series dax: Fix missed wakeup in put_unlocked_entry() | expand

Commit Message

Vivek Goyal April 16, 2021, 5:35 p.m. UTC
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 32M to reproduce
the problem consistently.

This is not a complete patch. I am just proposing this partial fix to
highlight the issue and trying to figure out how it should be fixed.
Should it be fixed in generic dax code or should filesystem (fuse/virtiofs)
take care of this.

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.

In my case I am seeing that dax page fault path itself is waiting
on grab_mapping_entry() and also invalidate_inode_page2() is 
waiting in get_unlocked_entry() but entry has already been cleaned
up and nobody woke up these processes. Atleast I think that's what
is happening.

This patch wakes up a process even if entry=0. And deadlock does not
happen. I am running into some OOM issues, that will debug.

So my question is that is it a dax issue and should it be fixed in
dax layer. Or should it be handled in fuse to make sure that
multiple instances of invalidate_inode_pages2() on same inode
don't make progress in parallel and introduce enough locking
around it.

Right now fuse_finish_open() calls invalidate_inode_pages2() without
any locking. That allows it to make progress in parallel to dax
fault path as well as allows multiple instances of invalidate_inode_pages2()
to run in parallel.

Not-yet-signed-off-by: Vivek Goyal <vgoyal@redhat.com>
---
 fs/dax.c |    7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

Comments

Dan Williams April 16, 2021, 7:56 p.m. UTC | #1
On Fri, Apr 16, 2021 at 10:35 AM Vivek Goyal <vgoyal@redhat.com> wrote:
>
> 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 32M to reproduce
> the problem consistently.
>
> This is not a complete patch. I am just proposing this partial fix to
> highlight the issue and trying to figure out how it should be fixed.
> Should it be fixed in generic dax code or should filesystem (fuse/virtiofs)
> take care of this.
>
> 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.
>
> In my case I am seeing that dax page fault path itself is waiting
> on grab_mapping_entry() and also invalidate_inode_page2() is
> waiting in get_unlocked_entry() but entry has already been cleaned
> up and nobody woke up these processes. Atleast I think that's what
> is happening.
>
> This patch wakes up a process even if entry=0. And deadlock does not
> happen. I am running into some OOM issues, that will debug.
>
> So my question is that is it a dax issue and should it be fixed in
> dax layer. Or should it be handled in fuse to make sure that
> multiple instances of invalidate_inode_pages2() on same inode
> don't make progress in parallel and introduce enough locking
> around it.
>
> Right now fuse_finish_open() calls invalidate_inode_pages2() without
> any locking. That allows it to make progress in parallel to dax
> fault path as well as allows multiple instances of invalidate_inode_pages2()
> to run in parallel.
>
> Not-yet-signed-off-by: Vivek Goyal <vgoyal@redhat.com>
> ---
>  fs/dax.c |    7 ++++---
>  1 file changed, 4 insertions(+), 3 deletions(-)
>
> Index: redhat-linux/fs/dax.c
> ===================================================================
> --- redhat-linux.orig/fs/dax.c  2021-04-16 12:50:40.141363317 -0400
> +++ redhat-linux/fs/dax.c       2021-04-16 12:51:42.385926390 -0400
> @@ -266,9 +266,10 @@ static void wait_entry_unlocked(struct x
>
>  static void put_unlocked_entry(struct xa_state *xas, void *entry)
>  {
> -       /* If we were the only waiter woken, wake the next one */
> -       if (entry && !dax_is_conflict(entry))
> -               dax_wake_entry(xas, entry, false);
> +       if (dax_is_conflict(entry))
> +               return;
> +
> +       dax_wake_entry(xas, entry, false);

How does this work if entry is NULL? dax_entry_waitqueue() will not
know if it needs to adjust the index. I think the fix might be to
specify that put_unlocked_entry() in the invalidate path needs to do a
wake_up_all().
Vivek Goyal April 16, 2021, 9:24 p.m. UTC | #2
On Fri, Apr 16, 2021 at 12:56:05PM -0700, Dan Williams wrote:
> On Fri, Apr 16, 2021 at 10:35 AM Vivek Goyal <vgoyal@redhat.com> wrote:
> >
> > 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 32M to reproduce
> > the problem consistently.
> >
> > This is not a complete patch. I am just proposing this partial fix to
> > highlight the issue and trying to figure out how it should be fixed.
> > Should it be fixed in generic dax code or should filesystem (fuse/virtiofs)
> > take care of this.
> >
> > 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.
> >
> > In my case I am seeing that dax page fault path itself is waiting
> > on grab_mapping_entry() and also invalidate_inode_page2() is
> > waiting in get_unlocked_entry() but entry has already been cleaned
> > up and nobody woke up these processes. Atleast I think that's what
> > is happening.
> >
> > This patch wakes up a process even if entry=0. And deadlock does not
> > happen. I am running into some OOM issues, that will debug.
> >
> > So my question is that is it a dax issue and should it be fixed in
> > dax layer. Or should it be handled in fuse to make sure that
> > multiple instances of invalidate_inode_pages2() on same inode
> > don't make progress in parallel and introduce enough locking
> > around it.
> >
> > Right now fuse_finish_open() calls invalidate_inode_pages2() without
> > any locking. That allows it to make progress in parallel to dax
> > fault path as well as allows multiple instances of invalidate_inode_pages2()
> > to run in parallel.
> >
> > Not-yet-signed-off-by: Vivek Goyal <vgoyal@redhat.com>
> > ---
> >  fs/dax.c |    7 ++++---
> >  1 file changed, 4 insertions(+), 3 deletions(-)
> >
> > Index: redhat-linux/fs/dax.c
> > ===================================================================
> > --- redhat-linux.orig/fs/dax.c  2021-04-16 12:50:40.141363317 -0400
> > +++ redhat-linux/fs/dax.c       2021-04-16 12:51:42.385926390 -0400
> > @@ -266,9 +266,10 @@ static void wait_entry_unlocked(struct x
> >
> >  static void put_unlocked_entry(struct xa_state *xas, void *entry)
> >  {
> > -       /* If we were the only waiter woken, wake the next one */
> > -       if (entry && !dax_is_conflict(entry))
> > -               dax_wake_entry(xas, entry, false);
> > +       if (dax_is_conflict(entry))
> > +               return;
> > +
> > +       dax_wake_entry(xas, entry, false);
> 

Hi Dan,

> How does this work if entry is NULL? dax_entry_waitqueue() will not
> know if it needs to adjust the index.

Wake waiters both at current index as well PMD adjusted index. It feels
little ugly though.

> I think the fix might be to
> specify that put_unlocked_entry() in the invalidate path needs to do a
> wake_up_all().

Doing a wake_up_all() when we invalidate an entry, sounds good. I will give
it a try.

Thanks
Vivek
Jan Kara April 19, 2021, 10:07 a.m. UTC | #3
On Fri 16-04-21 17:24:49, Vivek Goyal wrote:
> On Fri, Apr 16, 2021 at 12:56:05PM -0700, Dan Williams wrote:
> > On Fri, Apr 16, 2021 at 10:35 AM Vivek Goyal <vgoyal@redhat.com> wrote:
> > >
> > > 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 32M to reproduce
> > > the problem consistently.
> > >
> > > This is not a complete patch. I am just proposing this partial fix to
> > > highlight the issue and trying to figure out how it should be fixed.
> > > Should it be fixed in generic dax code or should filesystem (fuse/virtiofs)
> > > take care of this.
> > >
> > > 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.
> > >
> > > In my case I am seeing that dax page fault path itself is waiting
> > > on grab_mapping_entry() and also invalidate_inode_page2() is
> > > waiting in get_unlocked_entry() but entry has already been cleaned
> > > up and nobody woke up these processes. Atleast I think that's what
> > > is happening.
> > >
> > > This patch wakes up a process even if entry=0. And deadlock does not
> > > happen. I am running into some OOM issues, that will debug.
> > >
> > > So my question is that is it a dax issue and should it be fixed in
> > > dax layer. Or should it be handled in fuse to make sure that
> > > multiple instances of invalidate_inode_pages2() on same inode
> > > don't make progress in parallel and introduce enough locking
> > > around it.
> > >
> > > Right now fuse_finish_open() calls invalidate_inode_pages2() without
> > > any locking. That allows it to make progress in parallel to dax
> > > fault path as well as allows multiple instances of invalidate_inode_pages2()
> > > to run in parallel.
> > >
> > > Not-yet-signed-off-by: Vivek Goyal <vgoyal@redhat.com>
> > > ---
> > >  fs/dax.c |    7 ++++---
> > >  1 file changed, 4 insertions(+), 3 deletions(-)
> > >
> > > Index: redhat-linux/fs/dax.c
> > > ===================================================================
> > > --- redhat-linux.orig/fs/dax.c  2021-04-16 12:50:40.141363317 -0400
> > > +++ redhat-linux/fs/dax.c       2021-04-16 12:51:42.385926390 -0400
> > > @@ -266,9 +266,10 @@ static void wait_entry_unlocked(struct x
> > >
> > >  static void put_unlocked_entry(struct xa_state *xas, void *entry)
> > >  {
> > > -       /* If we were the only waiter woken, wake the next one */
> > > -       if (entry && !dax_is_conflict(entry))
> > > -               dax_wake_entry(xas, entry, false);
> > > +       if (dax_is_conflict(entry))
> > > +               return;
> > > +
> > > +       dax_wake_entry(xas, entry, false);
> > 
> 
> Hi Dan,
> 
> > How does this work if entry is NULL? dax_entry_waitqueue() will not
> > know if it needs to adjust the index.
> 
> Wake waiters both at current index as well PMD adjusted index. It feels
> little ugly though.
> 
> > I think the fix might be to
> > specify that put_unlocked_entry() in the invalidate path needs to do a
> > wake_up_all().
> 
> Doing a wake_up_all() when we invalidate an entry, sounds good. I will give
> it a try.

Yeah, that's what I'd suggest as well. After invalidating entry, there's no
point to let other waiters sleep. Trying to optimize for thundering herd
problems in face of entry invalidation is really fragile as you noticed.

								Honza
diff mbox series

Patch

Index: redhat-linux/fs/dax.c
===================================================================
--- redhat-linux.orig/fs/dax.c	2021-04-16 12:50:40.141363317 -0400
+++ redhat-linux/fs/dax.c	2021-04-16 12:51:42.385926390 -0400
@@ -266,9 +266,10 @@  static void wait_entry_unlocked(struct x
 
 static void put_unlocked_entry(struct xa_state *xas, void *entry)
 {
-	/* If we were the only waiter woken, wake the next one */
-	if (entry && !dax_is_conflict(entry))
-		dax_wake_entry(xas, entry, false);
+	if (dax_is_conflict(entry))
+		return;
+
+	dax_wake_entry(xas, entry, false);
 }
 
 /*