nfs: opendir/closedir for every readdir

Revert "nfs3: Unref & unbind dir fd with inode lock on EOF"

This reverts commit 4e6fb304ce41acbaf7c9ba67c06bf443e65082e8.

The above commit (which unbinds fds at EOF) does not fix the original
bug (1619) because a readdir from a second app could have already
started before the readdir_cbk of the first app's readdir reaches
NFS code. Hence the race still exists.

Performing extra unrefs when EOF is received is not a reliable way
of detecting that a client has performed a closedir (and to close
the fd ourselves). Neither is interpreting a 0 cookies a new opendir.
Clients can always use telldir/seekdir and hit EOFs twice.

Due to the way NFS3 protocol is designed, it is just not possible
for the server to reliably detect opendirs/closedirs performed by
the client and map the corresponding readdirs to the same dir fd
on the server side.

The only reliable way of fixing this is to perform opendir/closedir
at the cost of performance. Any optimization towards keeping dir fds
open attempting to map them with application's opendir/closedir will
either result in fd leaks or extra fd unrefs.

Signed-off-by: Anand V. Avati <avati@blackhole.gluster.com>
Signed-off-by: Anand V. Avati <avati@dev.gluster.com>

BUG: 2061 (NFS server crashes in readdir_fstat_cbk due to extra fd unref)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=2061

1 files changed, 1 insertions, 1 deletions

diff --git a/xlators/nfs/server/src/nfs3-helpers.c b/xlators/nfs/server/src/nfs3-helpers.c
index cff029d6e72..8e4ca372027 100644
--- a/xlators/nfs/server/src/nfs3-helpers.c
+++ b/xlators/nfs/server/src/nfs3-helpers.c
@@ -1775,7 +1775,7 @@ nfs3_dir_open_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                 goto err;
         }
 
-        cs->fd = fd_ref (fd);
+        cs->fd = fd;     /* Gets unrefd when the call state is wiped. */
         nfs3_set_inode_opened (cs->nfsx, cs->resolvedloc.inode);
         gf_log (GF_NFS3, GF_LOG_TRACE, "FD_REF: %d", fd->refcount);
         nfs3_call_resume (cs);