Btrfs: fix our reservations for updating an inode when completing io

People have been reporting ENOSPC crashes in finish_ordered_io.  This is because
we try to steal from the delalloc block rsv to satisfy a reservation to update
the inode.  The problem with this is we don't explicitly save space for updating
the inode when doing delalloc.  This is kind of a problem and we've gotten away
with this because way back when we just stole from the delalloc reserve without
any questions, and this worked out fine because generally speaking the leaf had
been modified either by the mtime update when we did the original write or
because we just updated the leaf when we inserted the file extent item, only on
rare occasions had the leaf not actually been modified, and that was still ok
because we'd just use a block or two out of the over-reservation that is
delalloc.

Then came the delayed inode stuff.  This is amazing, except it wants a full
reservation for updating the inode since it may do it at some point down the
road after we've written the blocks and we have to recow everything again.  This
worked out because the delayed inode stuff just stole from the global reserve,
that is until recently when I changed that because it caused other problems.

So here we are, we're doing everything right and being screwed for it.  So take
an extra reservation for the inode at delalloc reservation time and carry it
through the life of the delalloc reservation.  If we need it we can steal it in
the delayed inode stuff.  If we have already stolen it try and do a normal
metadata reservation.  If that fails try to steal from the delalloc reservation.
If _that_ fails we'll get a WARN_ON() so I can start thinking of a better way to
solve this and in the meantime we'll steal from the global reserve.

With this patch I ran xfstests 13 in a loop for a couple of hours and didn't see
any problems.

Signed-off-by: Josef Bacik <josef@redhat.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>

1 files changed, 62 insertions, 3 deletions

diff --git a/fs/btrfs/delayed-inode.c b/fs/btrfs/delayed-inode.c
index bbe8496..313ee14 100644
--- a/fs/btrfs/delayed-inode.c
+++ b/fs/btrfs/delayed-inode.c
@@ -617,12 +617,14 @@ static void btrfs_delayed_item_release_metadata(struct btrfs_root *root,
 static int btrfs_delayed_inode_reserve_metadata(
 					struct btrfs_trans_handle *trans,
 					struct btrfs_root *root,
+					struct inode *inode,
 					struct btrfs_delayed_node *node)
 {
 	struct btrfs_block_rsv *src_rsv;
 	struct btrfs_block_rsv *dst_rsv;
 	u64 num_bytes;
 	int ret;
+	int release = false;
 
 	src_rsv = trans->block_rsv;
 	dst_rsv = &root->fs_info->delayed_block_rsv;
@@ -652,11 +654,67 @@ static int btrfs_delayed_inode_reserve_metadata(
 		if (!ret)
 			node->bytes_reserved = num_bytes;
 		return ret;
+	} else if (src_rsv == &root->fs_info->delalloc_block_rsv) {
+		spin_lock(&BTRFS_I(inode)->lock);
+		if (BTRFS_I(inode)->delalloc_meta_reserved) {
+			BTRFS_I(inode)->delalloc_meta_reserved = 0;
+			spin_unlock(&BTRFS_I(inode)->lock);
+			release = true;
+			goto migrate;
+		}
+		spin_unlock(&BTRFS_I(inode)->lock);
+
+		/* Ok we didn't have space pre-reserved.  This shouldn't happen
+		 * too often but it can happen if we do delalloc to an existing
+		 * inode which gets dirtied because of the time update, and then
+		 * isn't touched again until after the transaction commits and
+		 * then we try to write out the data.  First try to be nice and
+		 * reserve something strictly for us.  If not be a pain and try
+		 * to steal from the delalloc block rsv.
+		 */
+		ret = btrfs_block_rsv_add_noflush(root, dst_rsv, num_bytes);
+		if (!ret)
+			goto out;
+
+		ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes);
+		if (!ret)
+			goto out;
+
+		/*
+		 * Ok this is a problem, let's just steal from the global rsv
+		 * since this really shouldn't happen that often.
+		 */
+		WARN_ON(1);
+		ret = btrfs_block_rsv_migrate(&root->fs_info->global_block_rsv,
+					      dst_rsv, num_bytes);
+		goto out;
 	}
 
+migrate:
 	ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes);
-	if (!ret)
-		node->bytes_reserved = num_bytes;
+	if (unlikely(ret)) {
+		/* This shouldn't happen */
+		BUG_ON(release);
+		return ret;
+	}
+
+out:
+	/*
+	 * Migrate only takes a reservation, it doesn't touch the size of the
+	 * block_rsv.  This is to simplify people who don't normally have things
+	 * migrated from their block rsv.  If they go to release their
+	 * reservation, that will decrease the size as well, so if migrate
+	 * reduced size we'd end up with a negative size.  But for the
+	 * delalloc_meta_reserved stuff we will only know to drop 1 reservation,
+	 * but we could in fact do this reserve/migrate dance several times
+	 * between the time we did the original reservation and we'd clean it
+	 * up.  So to take care of this, release the space for the meta
+	 * reservation here.  I think it may be time for a documentation page on
+	 * how block rsvs. work.
+	 */
+	if (release)
+		btrfs_block_rsv_release(root, src_rsv, num_bytes);
+	node->bytes_reserved = num_bytes;
 
 	return ret;
 }
@@ -1708,7 +1766,8 @@ int btrfs_delayed_update_inode(struct btrfs_trans_handle *trans,
 		goto release_node;
 	}
 
-	ret = btrfs_delayed_inode_reserve_metadata(trans, root, delayed_node);
+	ret = btrfs_delayed_inode_reserve_metadata(trans, root, inode,
+						   delayed_node);
 	if (ret)
 		goto release_node;