+
+void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
+{
+ ptlrpc_update_next_ping(imp);
+}
+
+int ptlrpc_pinger_add_import(struct obd_import *imp)
+{
+ ENTRY;
+ if (!list_empty(&imp->imp_pinger_chain))
+ RETURN(-EALREADY);
+
+ mutex_down(&pinger_sem);
+ CDEBUG(D_HA, "adding pingable import %s->%s\n",
+ imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
+ /* if we add to pinger we want recovery on this import */
+ imp->imp_obd->obd_no_recov = 0;
+ ptlrpc_update_next_ping(imp);
+ /* XXX sort, blah blah */
+ list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
+ class_import_get(imp);
+
+ ptlrpc_pinger_wake_up();
+ mutex_up(&pinger_sem);
+
+ RETURN(0);
+}
+
+int ptlrpc_pinger_del_import(struct obd_import *imp)
+{
+ ENTRY;
+ if (list_empty(&imp->imp_pinger_chain))
+ RETURN(-ENOENT);
+
+ mutex_down(&pinger_sem);
+ list_del_init(&imp->imp_pinger_chain);
+ CDEBUG(D_HA, "removing pingable import %s->%s\n",
+ imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
+ /* if we remove from pinger we don't want recovery on this import */
+ imp->imp_obd->obd_no_recov = 1;
+ class_import_put(imp);
+ mutex_up(&pinger_sem);
+ RETURN(0);
+}
+
+void ptlrpc_pinger_wake_up()
+{
+#ifdef ENABLE_PINGER
+ pinger_thread->t_flags |= SVC_EVENT;
+ cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
+#endif
+}
+
+/* Ping evictor thread */
+#define PET_READY 1
+#define PET_TERMINATE 2
+
+static int pet_refcount = 0;
+static int pet_state;
+static wait_queue_head_t pet_waitq;
+static struct obd_export *pet_exp = NULL;
+static spinlock_t pet_lock = SPIN_LOCK_UNLOCKED;
+
+int ping_evictor_wake(struct obd_export *exp)
+{
+ spin_lock(&pet_lock);
+ if (pet_exp || (pet_state != PET_READY)) {
+ /* eventually the new obd will call here again. */
+ spin_unlock(&pet_lock);
+ return 1;
+ }
+
+ /* We have to make sure the obd isn't destroyed between now and when
+ * the ping evictor runs. We'll take a reference here, and drop it
+ * when we finish in the evictor. We don't really care about this
+ * export in particular; we just need one to keep the obd alive. */
+ pet_exp = class_export_get(exp);
+ spin_unlock(&pet_lock);
+
+ wake_up(&pet_waitq);
+ return 0;
+}
+
+static int ping_evictor_main(void *arg)
+{
+ struct obd_device *obd;
+ struct obd_export *exp;
+ struct l_wait_info lwi = { 0 };
+ time_t expire_time;
+ ENTRY;
+
+ ptlrpc_daemonize("ll_evictor");
+
+ CDEBUG(D_HA, "Starting Ping Evictor\n");
+ pet_exp = NULL;
+ pet_state = PET_READY;
+ while (1) {
+ l_wait_event(pet_waitq, pet_exp ||
+ (pet_state == PET_TERMINATE), &lwi);
+ if (pet_state == PET_TERMINATE)
+ break;
+
+ /* we only get here if pet_exp != NULL, and the end of this
+ * loop is the only place which sets it NULL again, so lock
+ * is not strictly necessary. */
+ spin_lock(&pet_lock);
+ obd = pet_exp->exp_obd;
+ spin_unlock(&pet_lock);
+
+ expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
+
+ CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
+ obd->obd_name, expire_time);
+
+ /* Exports can't be deleted out of the list while we hold
+ * the obd lock (class_unlink_export), which means we can't
+ * lose the last ref on the export. If they've already been
+ * removed from the list, we won't find them here. */
+ spin_lock(&obd->obd_dev_lock);
+ while (!list_empty(&obd->obd_exports_timed)) {
+ exp = list_entry(obd->obd_exports_timed.next,
+ struct obd_export,exp_obd_chain_timed);
+ if (expire_time > exp->exp_last_request_time) {
+ class_export_get(exp);
+ spin_unlock(&obd->obd_dev_lock);
+ LCONSOLE_WARN("%s: haven't heard from client %s"
+ " (at %s) in %ld seconds. I think"
+ " it's dead, and I am evicting"
+ " it. exp %p, cur %ld expire %ld"
+ " last %ld\n",
+ obd->obd_name,
+ obd_uuid2str(&exp->exp_client_uuid),
+ obd_export_nid2str(exp),
+ (long)(cfs_time_current_sec() -
+ exp->exp_last_request_time),
+ exp, (long)cfs_time_current_sec(),
+ (long)expire_time,
+ (long)exp->exp_last_request_time);
+ CDEBUG(D_HA, "Last request was at %ld\n",
+ exp->exp_last_request_time);
+ class_fail_export(exp);
+ class_export_put(exp);
+ spin_lock(&obd->obd_dev_lock);
+ } else {
+ /* List is sorted, so everyone below is ok */
+ break;
+ }
+ }
+ spin_unlock(&obd->obd_dev_lock);
+
+ class_export_put(pet_exp);
+
+ spin_lock(&pet_lock);
+ pet_exp = NULL;
+ spin_unlock(&pet_lock);
+ }
+ CDEBUG(D_HA, "Exiting Ping Evictor\n");
+
+ RETURN(0);
+}
+
+void ping_evictor_start(void)
+{
+ int rc;
+
+ if (++pet_refcount > 1)
+ return;
+
+ init_waitqueue_head(&pet_waitq);
+
+ rc = cfs_kernel_thread(ping_evictor_main, NULL, CLONE_VM | CLONE_FILES);
+ if (rc < 0) {
+ pet_refcount--;
+ CERROR("Cannot start ping evictor thread: %d\n", rc);
+ }
+}
+EXPORT_SYMBOL(ping_evictor_start);
+
+void ping_evictor_stop(void)
+{
+ if (--pet_refcount > 0)
+ return;
+
+ pet_state = PET_TERMINATE;
+ wake_up(&pet_waitq);
+}
+EXPORT_SYMBOL(ping_evictor_stop);
+#else /* !__KERNEL__ */
+
+/* XXX
+ * the current implementation of pinger in liblustre is not optimized
+ */
+
+#ifdef ENABLE_PINGER
+static struct pinger_data {
+ int pd_recursion;
+ cfs_time_t pd_this_ping; /* jiffies */
+ cfs_time_t pd_next_ping; /* jiffies */
+ struct ptlrpc_request_set *pd_set;
+} pinger_args;
+
+static int pinger_check_rpcs(void *arg)
+{
+ cfs_time_t curtime = cfs_time_current();
+ struct ptlrpc_request *req;
+ struct ptlrpc_request_set *set;
+ struct list_head *iter;
+ struct pinger_data *pd = &pinger_args;
+ int rc;
+
+ /* prevent recursion */
+ if (pd->pd_recursion++) {
+ CDEBUG(D_HA, "pinger: recursion! quit\n");
+ LASSERT(pd->pd_set);
+ pd->pd_recursion--;
+ return 0;
+ }
+
+ /* have we reached ping point? */
+ if (!pd->pd_set && time_before(curtime, pd->pd_next_ping)) {
+ pd->pd_recursion--;
+ return 0;
+ }
+
+ /* if we have rpc_set already, continue processing it */
+ if (pd->pd_set) {
+ LASSERT(pd->pd_this_ping);
+ set = pd->pd_set;
+ goto do_check_set;
+ }
+
+ pd->pd_this_ping = curtime;
+ pd->pd_set = ptlrpc_prep_set();
+ if (pd->pd_set == NULL)
+ goto out;
+ set = pd->pd_set;
+
+ /* add rpcs into set */
+ mutex_down(&pinger_sem);
+ list_for_each(iter, &pinger_imports) {
+ struct obd_import *imp =
+ list_entry(iter, struct obd_import, imp_pinger_chain);
+ int generation, level;
+
+ if (cfs_time_aftereq(pd->pd_this_ping,
+ imp->imp_next_ping - 5 * CFS_TICK)) {
+ /* Add a ping. */
+ spin_lock(&imp->imp_lock);
+ generation = imp->imp_generation;
+ level = imp->imp_state;
+ spin_unlock(&imp->imp_lock);
+
+ if (level != LUSTRE_IMP_FULL) {
+ CDEBUG(D_HA,
+ "not pinging %s (in recovery)\n",
+ obd2cli_tgt(imp->imp_obd));
+ continue;
+ }
+
+ req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
+ LUSTRE_OBD_VERSION,
+ OBD_PING);
+ if (req == NULL) {
+ CERROR("OOM trying to ping %s->%s\n",
+ imp->imp_obd->obd_uuid.uuid,
+ obd2cli_tgt(imp->imp_obd));
+ break;
+ }
+
+ req->rq_no_resend = 1;
+ ptlrpc_request_set_replen(req);
+ req->rq_send_state = LUSTRE_IMP_FULL;
+ req->rq_phase = RQ_PHASE_RPC;
+ req->rq_import_generation = generation;
+ ptlrpc_set_add_req(set, req);
+ } else {
+ CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
+ " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
+ imp->imp_next_ping, pd->pd_this_ping);
+ }
+ }
+ pd->pd_this_ping = curtime;
+ mutex_up(&pinger_sem);
+
+ /* Might be empty, that's OK. */
+ if (set->set_remaining == 0)
+ CDEBUG(D_RPCTRACE, "nothing to ping\n");
+
+ list_for_each(iter, &set->set_requests) {
+ struct ptlrpc_request *req =
+ list_entry(iter, struct ptlrpc_request,
+ rq_set_chain);
+ DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
+ req->rq_import->imp_obd->obd_uuid.uuid,
+ obd2cli_tgt(req->rq_import->imp_obd));
+ (void)ptl_send_rpc(req, 0);
+ }
+
+do_check_set:
+ rc = ptlrpc_check_set(NULL, set);
+
+ /* not finished, and we are not expired, simply return */
+ if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
+ cfs_time_seconds(PING_INTERVAL)))) {
+ CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
+ pd->pd_recursion--;
+ return 0;
+ }
+
+ /* Expire all the requests that didn't come back. */
+ mutex_down(&pinger_sem);
+ list_for_each(iter, &set->set_requests) {
+ req = list_entry(iter, struct ptlrpc_request,
+ rq_set_chain);
+
+ if (req->rq_phase == RQ_PHASE_COMPLETE)
+ continue;
+
+ req->rq_phase = RQ_PHASE_COMPLETE;
+ atomic_dec(&req->rq_import->imp_inflight);
+ set->set_remaining--;
+ /* If it was disconnected, don't sweat it. */
+ if (list_empty(&req->rq_import->imp_pinger_chain)) {
+ ptlrpc_unregister_reply(req);
+ continue;
+ }
+
+ CDEBUG(D_RPCTRACE, "pinger initiate expire_one_request\n");
+ ptlrpc_expire_one_request(req);
+ }
+ mutex_up(&pinger_sem);
+
+ ptlrpc_set_destroy(set);
+ pd->pd_set = NULL;
+
+out:
+ pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
+ cfs_time_seconds(PING_INTERVAL));
+ pd->pd_this_ping = 0; /* XXX for debug */
+
+ CDEBUG(D_INFO, "finished a round ping\n");
+ pd->pd_recursion--;
+ return 0;
+}
+
+static void *pinger_callback = NULL;
+#endif /* ENABLE_PINGER */
+
+int ptlrpc_start_pinger(void)
+{
+#ifdef ENABLE_PINGER
+ memset(&pinger_args, 0, sizeof(pinger_args));
+ pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
+ &pinger_check_rpcs,
+ &pinger_args);
+#endif
+ return 0;
+}
+
+int ptlrpc_stop_pinger(void)
+{
+#ifdef ENABLE_PINGER
+ if (pinger_callback)
+ liblustre_deregister_wait_callback(pinger_callback);
+#endif
+ return 0;
+}
+
+void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
+{
+#ifdef ENABLE_PINGER
+ mutex_down(&pinger_sem);
+ ptlrpc_update_next_ping(imp);
+ if (pinger_args.pd_set == NULL &&
+ time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
+ CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
+ imp->imp_next_ping, cfs_time_current());
+ pinger_args.pd_next_ping = imp->imp_next_ping;
+ }
+ mutex_up(&pinger_sem);
+#endif
+}
+
+int ptlrpc_pinger_add_import(struct obd_import *imp)
+{
+ ENTRY;
+ if (!list_empty(&imp->imp_pinger_chain))
+ RETURN(-EALREADY);
+
+ CDEBUG(D_HA, "adding pingable import %s->%s\n",
+ imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
+ ptlrpc_pinger_sending_on_import(imp);
+
+ mutex_down(&pinger_sem);
+ list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
+ class_import_get(imp);
+ mutex_up(&pinger_sem);
+
+ RETURN(0);
+}
+
+int ptlrpc_pinger_del_import(struct obd_import *imp)
+{
+ ENTRY;
+ if (list_empty(&imp->imp_pinger_chain))
+ RETURN(-ENOENT);
+
+ mutex_down(&pinger_sem);
+ list_del_init(&imp->imp_pinger_chain);
+ CDEBUG(D_HA, "removing pingable import %s->%s\n",
+ imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
+ class_import_put(imp);
+ mutex_up(&pinger_sem);
+ RETURN(0);
+}
+
+void ptlrpc_pinger_wake_up()
+{
+#ifdef ENABLE_PINGER
+ /* XXX force pinger to run, if needed */
+ struct obd_import *imp;
+ ENTRY;
+ list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
+ CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
+ imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
+#ifdef ENABLE_LIBLUSTRE_RECOVERY
+ if (imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive)
+#else
+ /*XXX only recover for the initial connection */
+ if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
+ imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive)
+#endif
+ ptlrpc_initiate_recovery(imp);
+ else if (imp->imp_state != LUSTRE_IMP_FULL)
+ CDEBUG(D_HA, "Refused to recover import %s->%s "
+ "state %d, deactive %d\n",
+ imp->imp_obd->obd_uuid.uuid,
+ obd2cli_tgt(imp->imp_obd), imp->imp_state,
+ imp->imp_deactive);
+ }
+ EXIT;
+#endif
+}
+#endif /* !__KERNEL__ */