nfsd4: fix open downgrade, again

commit 3d02fa29 upstream.

Yet another open-management regression:

	- nfs4_file_downgrade() doesn't remove the BOTH access bit on
	  downgrade, so the server's idea of the stateid's access gets
	  out of sync with the client's.  If we want to keep an O_RDWR
	  open in this case, we should do that in the file_put_access
	  logic rather than here.
	- We forgot to convert v4 access to an open mode here.

This logic has proven too hard to get right.  In the future we may
consider:
	- reexamining the lock/openowner relationship (locks probably
	  don't really need to take their own references here).
	- adding open upgrade/downgrade support to the vfs.
	- removing the atomic operations.  They're redundant as long as
	  this is all under some other lock.

Also, maybe some kind of additional static checking would help catch
O_/NFS4_SHARE_ACCESS confusion.
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>

fs/nfsd/nfs4state.c

@@ -192,8 +192,15 @@ static void nfs4_file_put_fd(struct nfs4_file *fp, int oflag)
 static void __nfs4_file_put_access(struct nfs4_file *fp, int oflag)
 {
 	if (atomic_dec_and_test(&fp->fi_access[oflag])) {
-		nfs4_file_put_fd(fp, O_RDWR);
 		nfs4_file_put_fd(fp, oflag);
+		/*
+		 * It's also safe to get rid of the RDWR open *if*
+		 * we no longer have need of the other kind of access
+		 * or if we already have the other kind of open:
+		 */
+		if (fp->fi_fds[1-oflag]
+			|| atomic_read(&fp->fi_access[1 - oflag]) == 0)
+			nfs4_file_put_fd(fp, O_RDWR);
 	}
 }
 
@@ -3530,8 +3537,9 @@ static inline void nfs4_file_downgrade(struct nfs4_stateid *stp, unsigned int to
 	int i;
 
 	for (i = 1; i < 4; i++) {
-		if (test_bit(i, &stp->st_access_bmap) && !(i & to_access)) {
-			nfs4_file_put_access(stp->st_file, i);
+		if (test_bit(i, &stp->st_access_bmap)
+					&& ((i & to_access) != i)) {
+			nfs4_file_put_access(stp->st_file, nfs4_access_to_omode(i));
 			__clear_bit(i, &stp->st_access_bmap);
 		}
 	}