sunrpc/cache: handle missing listeners better.

If no handler (such as rpc.mountd) has opened
a cache 'channel', the sunrpc cache responds to
all lookup requests with -ENOENT.  This is particularly
important for the auth.unix.gid cache which is
optional.

If the channel was open briefly and an upcall was written to it,
this upcall remains pending even when the handler closes the
channel.  When an upcall is pending, the code currently
doesn't check if there are still listeners, it only performs
that check before sending an upcall.

As the cache treads a recently closes channel (closed less than
30 seconds ago) as "potentially still open", there is a
reasonable sized window when a request can become pending
in a closed channel, and thereby block lookups indefinitely.

This can easily be demonstrated by running
  cat /proc/net/rpc/auth.unix.gid/channel

and then trying to mount an NFS filesystem from this host.  It
will block indefinitely (unless mountd is run with --manage-gids,
or krb5 is used).

When cache_check() finds that an upcall is pending, it should
perform the "cache_listeners_exist()" exist test.  If no
listeners do exist, the request should be negated.

With this change in place, there can still be a 30second wait on
mount, until the cache gives up waiting for a handler to come
back, but this is much better than an indefinite wait.

Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

net/sunrpc/cache.c

@@ -40,6 +40,7 @@
 
 static bool cache_defer_req(struct cache_req *req, struct cache_head *item);
 static void cache_revisit_request(struct cache_head *item);
+static bool cache_listeners_exist(struct cache_detail *detail);
 
 static void cache_init(struct cache_head *h, struct cache_detail *detail)
 {
@@ -306,7 +307,8 @@ int cache_check(struct cache_detail *detail,
  cache_fresh_unlocked(h, detail);
  break;
  }
- }
+ } else if (!cache_listeners_exist(detail))
+ rv = try_to_negate_entry(detail, h);
  }
 
  if (rv == -EAGAIN) {