NFS: Fix an ABBA spinlock issue in pnfs_update_layout()

Commit Message

Trond Myklebust April 13, 2020, 8:36 p.m. UTC

From: Trond Myklebust <[e-mail]>

We need to drop the inode spinlock while calling nfs4_select_rw_stateid(),
since nfs4_copy_delegation_stateid() could take the delegation lock.
Note that it is safe to do this, since all other calls to
pnfs_update_layout() for that inode will find themselves blocked by
the lock we hold on NFS_LAYOUT_FIRST_LAYOUTGET.

Fixes: fc51b1cf391d ("NFS: Beware when dereferencing the delegation cred")
Signed-off-by: Trond Myklebust <[e-mail]>
---
 fs/nfs/pnfs.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

Patch

diff --git a/fs/nfs/pnfs.c b/fs/nfs/pnfs.c
index f2dc35c22964..b8d78f393365 100644
--- a/fs/nfs/pnfs.c
+++ b/fs/nfs/pnfs.c
@@ -2023,6 +2023,7 @@  pnfs_update_layout(struct inode *ino,
 			goto lookup_again;
 		}
 
+		spin_unlock(&ino->i_lock);
 		first = true;
 		status = nfs4_select_rw_stateid(ctx->state,
 					iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
@@ -2032,12 +2033,12 @@  pnfs_update_layout(struct inode *ino,
 			trace_pnfs_update_layout(ino, pos, count,
 					iomode, lo, lseg,
 					PNFS_UPDATE_LAYOUT_INVALID_OPEN);
-			spin_unlock(&ino->i_lock);
 			nfs4_schedule_stateid_recovery(server, ctx->state);
 			pnfs_clear_first_layoutget(lo);
 			pnfs_put_layout_hdr(lo);
 			goto lookup_again;
 		}
+		spin_lock(&ino->i_lock);
 	} else {
 		nfs4_stateid_copy(&stateid, &lo->plh_stateid);
 	}