#ifdef __KERNEL__
static void
-check_and_extend_recovery_timer(struct obd_device *obd,
- struct ptlrpc_request *req);
+check_and_start_recovery_timer(struct obd_device *obd,
+ struct ptlrpc_request *req, int new_client);
#else
static inline void
-check_and_extend_recovery_timer(struct obd_device *obd,
- struct ptlrpc_request *req)
+check_and_start_recovery_timer(struct obd_device *obd,
+ struct ptlrpc_request *req, int new_client)
{
}
#endif
/* If this is the first time a client connects, reset the recovery
* timer */
- if (rc == 0 && target->obd_recovering && export)
- check_and_extend_recovery_timer(target, req);
+ if (rc == 0 && target->obd_recovering)
+ check_and_start_recovery_timer(target, req, export == NULL);
/* We want to handle EALREADY but *not* -EALREADY from
* target_handle_reconnect(), return reconnection state in a flag */
&export->exp_nid_hash);
}
- cfs_spin_lock(&target->obd_recovery_task_lock);
if (target->obd_recovering && !export->exp_in_recovery) {
+ int has_transno;
+ __u64 transno = data->ocd_transno;
+
cfs_spin_lock(&export->exp_lock);
export->exp_in_recovery = 1;
export->exp_req_replay_needed = 1;
export->exp_lock_replay_needed = 1;
cfs_spin_unlock(&export->exp_lock);
- if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
- && (data->ocd_transno == 0))
+
+ has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
+ MSG_CONNECT_TRANSNO);
+ if (has_transno && transno == 0)
CWARN("Connect with zero transno!\n");
- if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
- && data->ocd_transno < target->obd_next_recovery_transno &&
- data->ocd_transno > target->obd_last_committed)
- target->obd_next_recovery_transno = data->ocd_transno;
- target->obd_connected_clients++;
+ if (has_transno && transno > 0 &&
+ transno < target->obd_next_recovery_transno &&
+ transno > target->obd_last_committed) {
+ /* another way is to use cmpxchg() so it will be
+ * lock free */
+ cfs_spin_lock(&target->obd_recovery_task_lock);
+ if (transno < target->obd_next_recovery_transno)
+ target->obd_next_recovery_transno = transno;
+ cfs_spin_unlock(&target->obd_recovery_task_lock);
+ }
+
cfs_atomic_inc(&target->obd_req_replay_clients);
cfs_atomic_inc(&target->obd_lock_replay_clients);
- if (target->obd_connected_clients ==
+ if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
target->obd_max_recoverable_clients)
cfs_waitq_signal(&target->obd_next_transno_waitq);
}
- cfs_spin_unlock(&target->obd_recovery_task_lock);
/* Tell the client we're in recovery, when client is involved in it. */
if (target->obd_recovering)
"%d recovered and %d %s evicted.\n", obd->obd_name,
(int)elapsed_time / 60, (int)elapsed_time % 60,
obd->obd_max_recoverable_clients,
- obd->obd_connected_clients, obd->obd_stale_clients,
+ cfs_atomic_read(&obd->obd_connected_clients),
+ obd->obd_stale_clients,
obd->obd_stale_clients == 1 ? "was" : "were");
ldlm_reprocess_all_ns(obd->obd_namespace);
cfs_timer_disarm(&obd->obd_recovery_timer);
}
-void target_start_recovery_timer(struct obd_device *obd)
+static void target_start_recovery_timer(struct obd_device *obd)
{
+ if (obd->obd_recovery_start != 0)
+ return;
+
cfs_spin_lock(&obd->obd_dev_lock);
if (!obd->obd_recovering || obd->obd_abort_recovery) {
cfs_spin_unlock(&obd->obd_dev_lock);
return;
}
- if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
+ LASSERT(obd->obd_recovery_timeout != 0);
+
+ if (obd->obd_recovery_start != 0) {
cfs_spin_unlock(&obd->obd_dev_lock);
return;
}
cfs_time_shift(obd->obd_recovery_timeout));
obd->obd_recovery_start = cfs_time_current_sec();
cfs_spin_unlock(&obd->obd_dev_lock);
- CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
+
+ LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
+ "or until %d client%s reconnect%s\n",
+ obd->obd_name,
+ obd->obd_recovery_timeout / 60,
+ obd->obd_recovery_timeout % 60,
+ obd->obd_max_recoverable_clients,
+ (obd->obd_max_recoverable_clients == 1) ? "" : "s",
+ (obd->obd_max_recoverable_clients == 1) ? "s": "");
}
-EXPORT_SYMBOL(target_start_recovery_timer);
/* extend recovery window to have extra @duration seconds at least. */
static void extend_recovery_timer(struct obd_device *obd, int drt)
{
- cfs_time_t now = cfs_time_current_sec();
+ cfs_time_t now;
+ cfs_time_t end;
cfs_duration_t left;
- if (!cfs_timer_is_armed(&obd->obd_recovery_timer)) {
- cfs_spin_lock(&obd->obd_dev_lock);
- if (obd->obd_recovery_timeout < drt)
- obd->obd_recovery_timeout = drt;
+ cfs_spin_lock(&obd->obd_dev_lock);
+ if (!obd->obd_recovering || obd->obd_abort_recovery) {
cfs_spin_unlock(&obd->obd_dev_lock);
return;
}
- left = obd->obd_recovery_timeout;
- left -= cfs_time_sub(now, obd->obd_recovery_start);
- if (drt > left) {
- cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(drt));
- CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
- obd->obd_name, (unsigned)drt);
+ LASSERT(obd->obd_recovery_start != 0);
+
+ now = cfs_time_current_sec();
+ end = obd->obd_recovery_start + obd->obd_recovery_timeout;
+ left = cfs_time_sub(end, now);
+ if (left < 0) {
+ obd->obd_recovery_timeout += drt - left;
+ } else if (left < drt) {
+ drt -= left;
+ obd->obd_recovery_timeout += drt;
+ } else {
+ drt = left;
}
+
+ cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(drt));
+ cfs_spin_unlock(&obd->obd_dev_lock);
+
+ CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
+ obd->obd_name, (unsigned)drt);
}
/* Reset the timer with each new client connection */
*/
static void
-check_and_extend_recovery_timer(struct obd_device *obd,
- struct ptlrpc_request *req)
+check_and_start_recovery_timer(struct obd_device *obd,
+ struct ptlrpc_request *req,
+ int new_client)
{
int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
struct obd_device_target *obt = &obd->u.obt;
struct lustre_sb_info *lsi;
- if (service_time)
+ if (!new_client && service_time)
/* Teach server about old server's estimates, as first guess
* at how long new requests will take. */
at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
service_time);
+ target_start_recovery_timer(obd);
+
/* convert the service time to rpc timeout,
* reuse service_time to limit stack usage */
service_time = at_est2timeout(service_time);
/** Checking routines for recovery */
static int check_for_clients(struct obd_device *obd)
{
+ unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
+
if (obd->obd_abort_recovery || obd->obd_recovery_expired)
return 1;
- LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
+ LASSERT(clnts <= obd->obd_max_recoverable_clients);
if (obd->obd_no_conn == 0 &&
- obd->obd_connected_clients + obd->obd_stale_clients ==
- obd->obd_max_recoverable_clients)
+ clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
return 1;
return 0;
}
req_transno = 0;
}
- connected = obd->obd_connected_clients;
+ connected = cfs_atomic_read(&obd->obd_connected_clients);
completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
queue_len = obd->obd_requests_queued_for_recovery;
next_transno = obd->obd_next_recovery_transno;
" after %lds (%d clients connected)\n",
obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
cfs_time_current_sec()- obd->obd_recovery_start,
- obd->obd_connected_clients);
+ cfs_atomic_read(&obd->obd_connected_clients));
obd->obd_recovery_expired = 1;
cfs_waitq_signal(&obd->obd_next_transno_waitq);