On Mon, Nov 26, 2018 at 6:17 PM David Sterba <dste...@suse.cz> wrote:
>
> On Fri, Nov 23, 2018 at 06:25:40PM +0000, fdman...@kernel.org wrote:
> > From: Filipe Manana <fdman...@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 <fdman...@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 <dste...@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.
