libceph: handle connection reopen race with callbacks
authorSage Weil <sage@newdream.net>
Thu, 19 May 2011 18:21:05 +0000 (11:21 -0700)
committerSage Weil <sage@newdream.net>
Thu, 19 May 2011 18:21:05 +0000 (11:21 -0700)
If a connection is closed and/or reopened (ceph_con_close, ceph_con_open)
it can race with a callback.  con_work does various state checks for
closed or reopened sockets at the beginning, but drops con->mutex before
making callbacks.  We need to check for state bit changes after retaking
the lock to ensure we restart con_work and execute those CLOSED/OPENING
tests or else we may end up operating under stale assumptions.

In Jim's case, this was causing 'bad tag' errors.

There are four cases where we re-take the con->mutex inside con_work: catch
them all and return EAGAIN from try_{read,write} so that we can restart
con_work.

Reported-by: Jim Schutt <jaschut@sandia.gov>
Tested-by: Jim Schutt <jaschut@sandia.gov>
Signed-off-by: Sage Weil <sage@newdream.net>
net/ceph/messenger.c

index e15a82ccc05f69789e162c5caa51d4bb2cc082a5..b140dd3515de1b46177e862a495067a98e7b1837 100644 (file)
@@ -598,7 +598,7 @@ static void prepare_write_keepalive(struct ceph_connection *con)
  * Connection negotiation.
  */
 
-static void prepare_connect_authorizer(struct ceph_connection *con)
+static int prepare_connect_authorizer(struct ceph_connection *con)
 {
        void *auth_buf;
        int auth_len = 0;
@@ -612,6 +612,10 @@ static void prepare_connect_authorizer(struct ceph_connection *con)
                                         con->auth_retry);
        mutex_lock(&con->mutex);
 
+       if (test_bit(CLOSED, &con->state) ||
+           test_bit(OPENING, &con->state))
+               return -EAGAIN;
+
        con->out_connect.authorizer_protocol = cpu_to_le32(auth_protocol);
        con->out_connect.authorizer_len = cpu_to_le32(auth_len);
 
@@ -619,6 +623,8 @@ static void prepare_connect_authorizer(struct ceph_connection *con)
        con->out_kvec[con->out_kvec_left].iov_len = auth_len;
        con->out_kvec_left++;
        con->out_kvec_bytes += auth_len;
+
+       return 0;
 }
 
 /*
@@ -640,9 +646,9 @@ static void prepare_write_banner(struct ceph_messenger *msgr,
        set_bit(WRITE_PENDING, &con->state);
 }
 
-static void prepare_write_connect(struct ceph_messenger *msgr,
-                                 struct ceph_connection *con,
-                                 int after_banner)
+static int prepare_write_connect(struct ceph_messenger *msgr,
+                                struct ceph_connection *con,
+                                int after_banner)
 {
        unsigned global_seq = get_global_seq(con->msgr, 0);
        int proto;
@@ -683,7 +689,7 @@ static void prepare_write_connect(struct ceph_messenger *msgr,
        con->out_more = 0;
        set_bit(WRITE_PENDING, &con->state);
 
-       prepare_connect_authorizer(con);
+       return prepare_connect_authorizer(con);
 }
 
 
@@ -1216,6 +1222,7 @@ static int process_connect(struct ceph_connection *con)
        u64 sup_feat = con->msgr->supported_features;
        u64 req_feat = con->msgr->required_features;
        u64 server_feat = le64_to_cpu(con->in_reply.features);
+       int ret;
 
        dout("process_connect on %p tag %d\n", con, (int)con->in_tag);
 
@@ -1250,7 +1257,9 @@ static int process_connect(struct ceph_connection *con)
                        return -1;
                }
                con->auth_retry = 1;
-               prepare_write_connect(con->msgr, con, 0);
+               ret = prepare_write_connect(con->msgr, con, 0);
+               if (ret < 0)
+                       return ret;
                prepare_read_connect(con);
                break;
 
@@ -1277,6 +1286,9 @@ static int process_connect(struct ceph_connection *con)
                if (con->ops->peer_reset)
                        con->ops->peer_reset(con);
                mutex_lock(&con->mutex);
+               if (test_bit(CLOSED, &con->state) ||
+                   test_bit(OPENING, &con->state))
+                       return -EAGAIN;
                break;
 
        case CEPH_MSGR_TAG_RETRY_SESSION:
@@ -1810,6 +1822,17 @@ static int try_read(struct ceph_connection *con)
 more:
        dout("try_read tag %d in_base_pos %d\n", (int)con->in_tag,
             con->in_base_pos);
+
+       /*
+        * process_connect and process_message drop and re-take
+        * con->mutex.  make sure we handle a racing close or reopen.
+        */
+       if (test_bit(CLOSED, &con->state) ||
+           test_bit(OPENING, &con->state)) {
+               ret = -EAGAIN;
+               goto out;
+       }
+
        if (test_bit(CONNECTING, &con->state)) {
                if (!test_bit(NEGOTIATING, &con->state)) {
                        dout("try_read connecting\n");
@@ -1938,8 +1961,10 @@ static void con_work(struct work_struct *work)
 {
        struct ceph_connection *con = container_of(work, struct ceph_connection,
                                                   work.work);
+       int ret;
 
        mutex_lock(&con->mutex);
+restart:
        if (test_and_clear_bit(BACKOFF, &con->state)) {
                dout("con_work %p backing off\n", con);
                if (queue_delayed_work(ceph_msgr_wq, &con->work,
@@ -1969,18 +1994,31 @@ static void con_work(struct work_struct *work)
                con_close_socket(con);
        }
 
-       if (test_and_clear_bit(SOCK_CLOSED, &con->state) ||
-           try_read(con) < 0 ||
-           try_write(con) < 0) {
-               mutex_unlock(&con->mutex);
-               ceph_fault(con);     /* error/fault path */
-               goto done_unlocked;
-       }
+       if (test_and_clear_bit(SOCK_CLOSED, &con->state))
+               goto fault;
+
+       ret = try_read(con);
+       if (ret == -EAGAIN)
+               goto restart;
+       if (ret < 0)
+               goto fault;
+
+       ret = try_write(con);
+       if (ret == -EAGAIN)
+               goto restart;
+       if (ret < 0)
+               goto fault;
 
 done:
        mutex_unlock(&con->mutex);
 done_unlocked:
        con->ops->put(con);
+       return;
+
+fault:
+       mutex_unlock(&con->mutex);
+       ceph_fault(con);     /* error/fault path */
+       goto done_unlocked;
 }