nfs: don't try to migrate pages with active requests
authorJeff Layton <jlayton@redhat.com>
Wed, 12 Oct 2011 14:57:42 +0000 (10:57 -0400)
committerGreg Kroah-Hartman <gregkh@suse.de>
Fri, 11 Nov 2011 17:43:14 +0000 (09:43 -0800)
commitc56f5c1b3e815614d5bc92e19d1c7633fc1d07e2
tree323b9335d9c2b1e53de2905ea4d64c4b96deb53b
parentcce574ff161d0f5e4263ed28a4a3785220802922
nfs: don't try to migrate pages with active requests

commit 2da956523526e440ef4f4dd174e26f5ac06fe011 upstream.

nfs_find_and_lock_request will take a reference to the nfs_page and
will then put it if the req is already locked. It's possible though
that the reference will be the last one. That put then can kick off
a whole series of reference puts:

nfs_page
   nfs_open_context
      dentry
          inode

If the inode ends up being deleted, then the VFS will call
truncate_inode_pages. That function will try to take the page lock, but
it was already locked when migrate_page was called. The code
deadlocks.

Fix this by simply refusing the migration request if PagePrivate is
already set, indicating that the page is already associated with an
active read or write request.

We've had a customer test a backported version of this patch and
the preliminary results seem good.

Cc: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Harshula Jayasuriya <harshula@redhat.com>
Signed-off-by: Jeff Layton <jlayton@redhat.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
fs/nfs/write.c