pNFS: Fix a deadlock when coalescing writes and returning the layout

[ Upstream commit 61f454e3 ]

Consider the following deadlock:

Process P1	Process P2		Process P3
==========	==========		==========
					lock_page(page)

		lseg = pnfs_update_layout(inode)

lo = NFS_I(inode)->layout
pnfs_error_mark_layout_for_return(lo)

		lock_page(page)

					lseg = pnfs_update_layout(inode)

In this scenario,
- P1 has declared the layout to be in error, but P2 holds a reference to
  a layout segment on that inode, so the layoutreturn is deferred.
- P2 is waiting for a page lock held by P3.
- P3 is asking for a new layout segment, but is blocked waiting
  for the layoutreturn.

The fix is to ensure that pnfs_error_mark_layout_for_return() does
not set the NFS_LAYOUT_RETURN flag, which blocks P3. Instead, we allow
the latter to call LAYOUTGET so that it can make progress and unblock
P2.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Sasha Levin <alexander.levin@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

fs/nfs/pnfs.c

@@ -1953,8 +1953,6 @@ void pnfs_error_mark_layout_for_return(struct inode *inode,
 
 	spin_lock(&inode->i_lock);
 	pnfs_set_plh_return_info(lo, range.iomode, 0);
-	/* Block LAYOUTGET */
-	set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
 	/*
 	 * mark all matching lsegs so that we are sure to have no live
 	 * segments at hand when sending layoutreturn. See pnfs_put_lseg()