Re: [PATCH v4] Btrfs: fix deadlock with memory reclaim during scrub

[David Sterba <dsterba@suse.cz>]

On Mon, Nov 26, 2018 at 08:10:30PM +0000, Filipe Manana wrote:
> On Mon, Nov 26, 2018 at 6:17 PM David Sterba <dsterba@suse.cz> wrote:
> >
> > On Fri, Nov 23, 2018 at 06:25:40PM +0000, fdmanana@kernel.org wrote:
> > > From: Filipe Manana <fdmanana@suse.com>
> > >
> > > When a transaction commit starts, it attempts to pause scrub and it blocks
> > > until the scrub is paused. So while the transaction is blocked waiting for
> > > scrub to pause, we can not do memory allocation with GFP_KERNEL from scrub,
> > > otherwise we risk getting into a deadlock with reclaim.
> > >
> > > Checking for scrub pause requests is done early at the beginning of the
> > > while loop of scrub_stripe() and later in the loop, scrub_extent() and
> > > scrub_raid56_parity() are called, which in turn call scrub_pages() and
> > > scrub_pages_for_parity() respectively. These last two functions do memory
> > > allocations using GFP_KERNEL. Same problem could happen while scrubbing
> > > the super blocks, since it calls scrub_pages().
> > >
> > > So make sure GFP_NOFS is used for the memory allocations because at any
> > > time a scrub pause request can happen from another task that started to
> > > commit a transaction.
> > >
> > > Fixes: 58c4e173847a ("btrfs: scrub: use GFP_KERNEL on the submission path")
> > > Signed-off-by: Filipe Manana <fdmanana@suse.com>
> > > ---
> > >
> > > V2: Make using GFP_NOFS unconditionial. Previous version was racy, as pausing
> > > requests migth happen just after we checked for them.
> > >
> > > V3: Use memalloc_nofs_save() just like V1 did.
> > >
> > > V4: Similar problem happened for raid56, which was previously missed, so
> > >     deal with it as well as the case for scrub_supers().
> >
> > Enclosing the whole scrub to 'nofs' seems like the best option and
> > future proof. What I missed in 58c4e173847a was the "don't hold big lock
> > under GFP_KERNEL allocation" pattern.
> >
> > >  fs/btrfs/scrub.c | 12 ++++++++++++
> > >  1 file changed, 12 insertions(+)
> > >
> > > diff --git a/fs/btrfs/scrub.c b/fs/btrfs/scrub.c
> > > index 3be1456b5116..e08b7502d1f0 100644
> > > --- a/fs/btrfs/scrub.c
> > > +++ b/fs/btrfs/scrub.c
> > > @@ -3779,6 +3779,7 @@ int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
> > >       struct scrub_ctx *sctx;
> > >       int ret;
> > >       struct btrfs_device *dev;
> > > +     unsigned int nofs_flag;
> > >
> > >       if (btrfs_fs_closing(fs_info))
> > >               return -EINVAL;
> > > @@ -3882,6 +3883,16 @@ int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
> > >       atomic_inc(&fs_info->scrubs_running);
> > >       mutex_unlock(&fs_info->scrub_lock);
> > >
> > > +     /*
> > > +      * In order to avoid deadlock with reclaim when there is a transaction
> > > +      * trying to pause scrub, make sure we use GFP_NOFS for all the
> > > +      * allocations done at btrfs_scrub_pages() and scrub_pages_for_parity()
> > > +      * invoked by our callees. The pausing request is done when the
> > > +      * transaction commit starts, and it blocks the transaction until scrub
> > > +      * is paused (done at specific points at scrub_stripe() or right above
> > > +      * before incrementing fs_info->scrubs_running).
> >
> > This hilights why there's perhaps no point in trying to make the nofs
> > section smaller, handling all the interactions between scrub and
> > transaction would be too complex.
> >
> > Reviewed-by: David Sterba <dsterba@suse.com>
> 
> Well, the worker tasks can also not use gfp_kernel, since the scrub
> task waits for them to complete before pausing.
> I missed this, and 2 reviewers as well, so perhaps it wasn't that
> trivial and I shouldn't feel that I miserably failed to identify all
> cases for something rather trivial. V5 sent.

You can say that you left it there intentionally, such cookies are a
good drill for reviewers to stay sharp.

When I started the conversions of GFP_NOFS -> GFP_KERNEL, scrub looked
quite safe in this respect but turns out it's not. I was wondering if we
could add some lock assertions before GFP_KERNEL allocations, like:

	assert_lock_not_held(fs_info->device_list_mutex)
	kmalloc(GFP_KERNEL)

and add more locks per subsystem (eg. the scrub lock) and possibly some
convenience wrappers. Michal's scope GFS_NOFS patch series has a
debugging warning where NOFS is used in context where it does not need
to, while having the 'must not hold an important lock' would be a good
debugging helper too.