Re: [PATCH v2] Btrfs: fix race between send and deduplication that lead to failures and crashes

[Filipe Manana <fdmanana@kernel.org>]

On Mon, May 13, 2019 at 5:06 PM David Sterba <dsterba@suse.cz> wrote:
>
> On Mon, May 13, 2019 at 05:56:07PM +0200, David Sterba wrote:
> > On Mon, Apr 22, 2019 at 04:43:42PM +0100, fdmanana@kernel.org wrote:
> > > +           btrfs_warn_rl(root_dst->fs_info,
> > > +"Can not deduplicate to root %llu while send operations are using it (%d in progress)",
> > > +                         root_dst->root_key.objectid,
> > > +                         root_dst->send_in_progress);
> >
> > The test btrfs/187 stresses this code and the logs are flooded by the
> > messages, even ratelimited.
> >
> > I wonder if the test is rather artificail (and that's fine for the testing
> > purposes) or if the number of messages would repeat under normal conditions.
> >
> > We don't need to print the message each time the dedup tries to acces a
> > snapshot under send, so keeping track if the message has been sent already
> > would be less intrusive and still provide the information.
>
> Untested:
>
> --- a/fs/btrfs/ctree.h
> +++ b/fs/btrfs/ctree.h
> @@ -1205,6 +1205,8 @@ enum {
>         BTRFS_ROOT_DEAD_RELOC_TREE,
>         /* Mark dead root stored on device whose cleanup needs to be resumed */
>         BTRFS_ROOT_DEAD_TREE,
> +       /* Track if dedupe was attempted under a current send */
> +       BTRFS_ROOT_NOTIFIED_DEDUPE_DURING_SEND,
>  };
>
>  /*
> diff --git a/fs/btrfs/ioctl.c b/fs/btrfs/ioctl.c
> index 6dafa857bbb9..23677cf12afc 100644
> --- a/fs/btrfs/ioctl.c
> +++ b/fs/btrfs/ioctl.c
> @@ -3263,7 +3263,9 @@ static int btrfs_extent_same(struct inode *src, u64 loff, u64 olen,
>
>         spin_lock(&root_dst->root_item_lock);
>         if (root_dst->send_in_progress) {
> -               btrfs_warn_rl(root_dst->fs_info,
> +               if (!test_and_set_bit(BTRFS_ROOT_NOTIFIED_DEDUPE_DURING_SEND,
> +                                       &root_dst->state))
> +                       btrfs_warn(root_dst->fs_info,
>  "cannot deduplicate to root %llu while send operations are using it (%d in progress)",
>                               root_dst->root_key.objectid,
>                               root_dst->send_in_progress);
> diff --git a/fs/btrfs/send.c b/fs/btrfs/send.c
> index dd38dfe174df..cc85ae903368 100644
> --- a/fs/btrfs/send.c
> +++ b/fs/btrfs/send.c
> @@ -6637,6 +6637,8 @@ static void btrfs_root_dec_send_in_progress(struct btrfs_root* root)
>                 btrfs_err(root->fs_info,
>                           "send_in_progress unbalanced %d root %llu",
>                           root->send_in_progress, root->root_key.objectid);
> +       if (root->send_in_progress == 0)
> +               clear_bit(BTRFS_ROOT_NOTIFIED_DEDUPE_DURING_SEND, &root->state);
>         spin_unlock(&root->root_item_lock);

I would leave it as it is unless users start to complain. Yes, the
test does this on purpose.
Adding such code/state seems weird to me, instead I would change the
rate limit state so that the messages would repeat much less
frequently.

>  }
>