knfsd: Fix a race in closing NFSd connections.

If you lose this race, it can iput a socket inode twice and you
get a BUG in fs/inode.c

When I added the option for user-space to close a socket,
I added some cruft to svc_delete_socket so that I could call
that function when closing a socket per user-space request.

This was the wrong thing to do.  I should have just set SK_CLOSE
and let normal mechanisms do the work.

Not only wrong, but buggy.  The locking is all wrong and it openned
up a race where-by a socket could be closed twice.

So this patch:
  Introduces svc_close_socket which sets SK_CLOSE then either leave
  the close up to a thread, or calls svc_delete_socket if it can
  get SK_BUSY.

  Adds a bias to sk_busy which is removed when SK_DEAD is set,
  This avoid races around shutting down the socket.

  Changes several 'spin_lock' to 'spin_lock_bh' where the _bh
  was missing.

Bugzilla-url: http://bugzilla.kernel.org/show_bug.cgi?id=7916

Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>

diff --git a/include/linux/sunrpc/svcsock.h b/include/linux/sunrpc/svcsock.h
index 98b21ad370fdd529c7a35105642276876360ff82..db312a1e2eebc70bca97465b0e28c5f6a7056bf6 100644 (file)
--- a/include/linux/sunrpc/svcsock.h
+++ b/include/linux/sunrpc/svcsock.h
@@ -63,7 +63,7 @@ struct svc_sock {
  * Function prototypes.
  */
 int            svc_makesock(struct svc_serv *, int, unsigned short);
-void           svc_delete_socket(struct svc_sock *);
+void           svc_close_socket(struct svc_sock *);
 int            svc_recv(struct svc_rqst *, long);
 int            svc_send(struct svc_rqst *);
 void           svc_drop(struct svc_rqst *);

diff --git a/net/sunrpc/svc.c b/net/sunrpc/svc.c
index f8179755bd5f39b74398feac9f5227bca8f7e3b0..9359b84cc8d64829b1decee64c07e2aaa0134a7a 100644 (file)
--- a/net/sunrpc/svc.c
+++ b/net/sunrpc/svc.c
@@ -387,7 +387,7 @@ svc_destroy(struct svc_serv *serv)
                svsk = list_entry(serv->sv_tempsocks.next,
                                  struct svc_sock,
                                  sk_list);
-               svc_delete_socket(svsk);
+               svc_close_socket(svsk);
        }
        if (serv->sv_shutdown)
                serv->sv_shutdown(serv);
@@ -396,7 +396,7 @@ svc_destroy(struct svc_serv *serv)
                svsk = list_entry(serv->sv_permsocks.next,
                                  struct svc_sock,
                                  sk_list);
-               svc_delete_socket(svsk);
+               svc_close_socket(svsk);
        }
        
        cache_clean_deferred(serv);

diff --git a/net/sunrpc/svcsock.c b/net/sunrpc/svcsock.c
index 4e86df55a5effb873e9be48757ce304ad1f3b422..80b9540acb883bfc5a918bb94f757ff6802574a3 100644 (file)
--- a/net/sunrpc/svcsock.c
+++ b/net/sunrpc/svcsock.c
@@ -61,6 +61,12 @@
  *             after a clear, the socket must be read/accepted
  *              if this succeeds, it must be set again.
  *     SK_CLOSE can set at any time. It is never cleared.
+ *      sk_inuse contains a bias of '1' until SK_DEAD is set.
+ *             so when sk_inuse hits zero, we know the socket is dead
+ *             and no-one is using it.
+ *      SK_DEAD can only be set while SK_BUSY is held which ensures
+ *             no other thread will be using the socket or will try to
+ *            set SK_DEAD.
  *
  */
 
@@ -69,6 +75,7 @@
 
 static struct svc_sock *svc_setup_socket(struct svc_serv *, struct socket *,
                                         int *errp, int pmap_reg);
+static void            svc_delete_socket(struct svc_sock *svsk);
 static void            svc_udp_data_ready(struct sock *, int);
 static int             svc_udp_recvfrom(struct svc_rqst *);
 static int             svc_udp_sendto(struct svc_rqst *);
@@ -299,8 +306,9 @@ void svc_reserve(struct svc_rqst *rqstp, int space)
 static inline void
 svc_sock_put(struct svc_sock *svsk)
 {
-       if (atomic_dec_and_test(&svsk->sk_inuse) &&
-                       test_bit(SK_DEAD, &svsk->sk_flags)) {
+       if (atomic_dec_and_test(&svsk->sk_inuse)) {
+               BUG_ON(! test_bit(SK_DEAD, &svsk->sk_flags));
+
                dprintk("svc: releasing dead socket\n");
                if (svsk->sk_sock->file)
                        sockfd_put(svsk->sk_sock);
@@ -490,7 +498,7 @@ svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
 
        if (!serv)
                return 0;
-       spin_lock(&serv->sv_lock);
+       spin_lock_bh(&serv->sv_lock);
        list_for_each_entry(svsk, &serv->sv_permsocks, sk_list) {
                int onelen = one_sock_name(buf+len, svsk);
                if (toclose && strcmp(toclose, buf+len) == 0)
@@ -498,12 +506,12 @@ svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
                else
                        len += onelen;
        }
-       spin_unlock(&serv->sv_lock);
+       spin_unlock_bh(&serv->sv_lock);
        if (closesk)
                /* Should unregister with portmap, but you cannot
                 * unregister just one protocol...
                 */
-               svc_delete_socket(closesk);
+               svc_close_socket(closesk);
        else if (toclose)
                return -ENOENT;
        return len;
@@ -653,6 +661,11 @@ svc_udp_recvfrom(struct svc_rqst *rqstp)
                return svc_deferred_recv(rqstp);
        }
 
+       if (test_bit(SK_CLOSE, &svsk->sk_flags)) {
+               svc_delete_socket(svsk);
+               return 0;
+       }
+
        clear_bit(SK_DATA, &svsk->sk_flags);
        while ((skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err)) == NULL) {
                if (err == -EAGAIN) {
@@ -1142,7 +1155,8 @@ svc_tcp_sendto(struct svc_rqst *rqstp)
                       rqstp->rq_sock->sk_server->sv_name,
                       (sent<0)?"got error":"sent only",
                       sent, xbufp->len);
-               svc_delete_socket(rqstp->rq_sock);
+               set_bit(SK_CLOSE, &rqstp->rq_sock->sk_flags);
+               svc_sock_enqueue(rqstp->rq_sock);
                sent = -EAGAIN;
        }
        return sent;
@@ -1461,7 +1475,7 @@ svc_setup_socket(struct svc_serv *serv, struct socket *sock,
        svsk->sk_odata = inet->sk_data_ready;
        svsk->sk_owspace = inet->sk_write_space;
        svsk->sk_server = serv;
-       atomic_set(&svsk->sk_inuse, 0);
+       atomic_set(&svsk->sk_inuse, 1);
        svsk->sk_lastrecv = get_seconds();
        spin_lock_init(&svsk->sk_defer_lock);
        INIT_LIST_HEAD(&svsk->sk_deferred);
@@ -1582,7 +1596,7 @@ bummer:
 /*
  * Remove a dead socket
  */
-void
+static void
 svc_delete_socket(struct svc_sock *svsk)
 {
        struct svc_serv *serv;
@@ -1608,16 +1622,26 @@ svc_delete_socket(struct svc_sock *svsk)
         * while still attached to a queue, the queue itself
         * is about to be destroyed (in svc_destroy).
         */
-       if (!test_and_set_bit(SK_DEAD, &svsk->sk_flags))
+       if (!test_and_set_bit(SK_DEAD, &svsk->sk_flags)) {
+               BUG_ON(atomic_read(&svsk->sk_inuse)<2);
+               atomic_dec(&svsk->sk_inuse);
                if (test_bit(SK_TEMP, &svsk->sk_flags))
                        serv->sv_tmpcnt--;
+       }
 
-       /* This atomic_inc should be needed - svc_delete_socket
-        * should have the semantic of dropping a reference.
-        * But it doesn't yet....
-        */
-       atomic_inc(&svsk->sk_inuse);
        spin_unlock_bh(&serv->sv_lock);
+}
+
+void svc_close_socket(struct svc_sock *svsk)
+{
+       set_bit(SK_CLOSE, &svsk->sk_flags);
+       if (test_and_set_bit(SK_BUSY, &svsk->sk_flags))
+               /* someone else will have to effect the close */
+               return;
+
+       atomic_inc(&svsk->sk_inuse);
+       svc_delete_socket(svsk);
+       clear_bit(SK_BUSY, &svsk->sk_flags);
        svc_sock_put(svsk);
 }