+lnet_handle_remote_failure_locked(struct lnet_peer_ni *lpni)
+{
+ __u32 sensitivity = lnet_health_sensitivity;
+ __u32 lp_sensitivity;
+
+ /*
+ * If there is a health sensitivity in the peer then use that
+ * instead of the globally set one.
+ */
+ lp_sensitivity = lpni->lpni_peer_net->lpn_peer->lp_health_sensitivity;
+ if (lp_sensitivity)
+ sensitivity = lp_sensitivity;
+
+ lnet_dec_healthv_locked(&lpni->lpni_healthv, sensitivity);
+
+ /* update the peer_net's health value */
+ lnet_update_peer_net_healthv(lpni);
+
+ /*
+ * add the peer NI to the recovery queue if it's not already there
+ * and it's health value is actually below the maximum. It's
+ * possible that the sensitivity might be set to 0, and the health
+ * value will not be reduced. In this case, there is no reason to
+ * invoke recovery
+ */
+ lnet_peer_ni_add_to_recoveryq_locked(lpni,
+ &the_lnet.ln_mt_peerNIRecovq,
+ ktime_get_seconds());
+}
+
+static void
+lnet_handle_remote_failure(struct lnet_peer_ni *lpni)
+{
+ /* lpni could be NULL if we're in the LOLND case */
+ if (!lpni)
+ return;
+
+ lnet_net_lock(0);
+ /* the mt could've shutdown and cleaned up the queues */
+ if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
+ lnet_net_unlock(0);
+ return;
+ }
+ lnet_handle_remote_failure_locked(lpni);
+ lnet_net_unlock(0);
+}
+
+static void
+lnet_incr_hstats(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
+ enum lnet_msg_hstatus hstatus)
+{
+ struct lnet_counters_health *health;
+
+ health = &the_lnet.ln_counters[0]->lct_health;
+
+ switch (hstatus) {
+ case LNET_MSG_STATUS_LOCAL_INTERRUPT:
+ atomic_inc(&ni->ni_hstats.hlt_local_interrupt);
+ health->lch_local_interrupt_count++;
+ break;
+ case LNET_MSG_STATUS_LOCAL_DROPPED:
+ atomic_inc(&ni->ni_hstats.hlt_local_dropped);
+ health->lch_local_dropped_count++;
+ break;
+ case LNET_MSG_STATUS_LOCAL_ABORTED:
+ atomic_inc(&ni->ni_hstats.hlt_local_aborted);
+ health->lch_local_aborted_count++;
+ break;
+ case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
+ atomic_inc(&ni->ni_hstats.hlt_local_no_route);
+ health->lch_local_no_route_count++;
+ break;
+ case LNET_MSG_STATUS_LOCAL_TIMEOUT:
+ atomic_inc(&ni->ni_hstats.hlt_local_timeout);
+ health->lch_local_timeout_count++;
+ break;
+ case LNET_MSG_STATUS_LOCAL_ERROR:
+ atomic_inc(&ni->ni_hstats.hlt_local_error);
+ health->lch_local_error_count++;
+ break;
+ case LNET_MSG_STATUS_REMOTE_DROPPED:
+ if (lpni)
+ atomic_inc(&lpni->lpni_hstats.hlt_remote_dropped);
+ health->lch_remote_dropped_count++;
+ break;
+ case LNET_MSG_STATUS_REMOTE_ERROR:
+ if (lpni)
+ atomic_inc(&lpni->lpni_hstats.hlt_remote_error);
+ health->lch_remote_error_count++;
+ break;
+ case LNET_MSG_STATUS_REMOTE_TIMEOUT:
+ if (lpni)
+ atomic_inc(&lpni->lpni_hstats.hlt_remote_timeout);
+ health->lch_remote_timeout_count++;
+ break;
+ case LNET_MSG_STATUS_NETWORK_TIMEOUT:
+ if (lpni)
+ atomic_inc(&lpni->lpni_hstats.hlt_network_timeout);
+ health->lch_network_timeout_count++;
+ break;
+ case LNET_MSG_STATUS_OK:
+ break;
+ default:
+ LBUG();
+ }
+}
+
+static void
+lnet_resend_msg_locked(struct lnet_msg *msg)
+{
+ msg->msg_retry_count++;
+
+ /*
+ * remove message from the active list and reset it to prepare
+ * for a resend. Two exceptions to this
+ *
+ * 1. the router case. When a message is being routed it is
+ * committed for rx when received and committed for tx when
+ * forwarded. We don't want to remove it from the active list, since
+ * code which handles receiving expects it to remain on the active
+ * list.
+ *
+ * 2. The REPLY case. Reply messages use the same message
+ * structure for the GET that was received.
+ */
+ if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
+ list_del_init(&msg->msg_activelist);
+ msg->msg_onactivelist = 0;
+ }
+ /*
+ * The msg_target.nid which was originally set
+ * when calling LNetGet() or LNetPut() might've
+ * been overwritten if we're routing this message.
+ * Call lnet_msg_decommit_tx() to return the credit
+ * this message consumed. The message will
+ * consume another credit when it gets resent.
+ */
+ msg->msg_target.nid = msg->msg_hdr.dest_nid;
+ lnet_msg_decommit_tx(msg, -EAGAIN);
+ msg->msg_sending = 0;
+ msg->msg_receiving = 0;
+ msg->msg_target_is_router = 0;
+
+ CDEBUG(D_NET, "%s->%s:%s:%s - queuing msg (%p) for resend\n",
+ libcfs_nid2str(msg->msg_hdr.src_nid),
+ libcfs_nid2str(msg->msg_hdr.dest_nid),
+ lnet_msgtyp2str(msg->msg_type),
+ lnet_health_error2str(msg->msg_health_status), msg);
+
+ list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
+
+ complete(&the_lnet.ln_mt_wait_complete);
+}
+
+int
+lnet_check_finalize_recursion_locked(struct lnet_msg *msg,
+ struct list_head *containerq,
+ int nworkers, void **workers)
+{
+ int my_slot = -1;
+ int i;
+
+ list_add_tail(&msg->msg_list, containerq);
+
+ for (i = 0; i < nworkers; i++) {
+ if (workers[i] == current)
+ break;
+
+ if (my_slot < 0 && workers[i] == NULL)
+ my_slot = i;
+ }
+
+ if (i < nworkers || my_slot < 0)
+ return -1;
+
+ workers[my_slot] = current;
+
+ return my_slot;
+}
+
+int
+lnet_attempt_msg_resend(struct lnet_msg *msg)
+{
+ struct lnet_msg_container *container;
+ int my_slot;
+ int cpt;
+
+ /* we can only resend tx_committed messages */
+ LASSERT(msg->msg_tx_committed);
+
+ /* don't resend recovery messages */
+ if (msg->msg_recovery) {
+ CDEBUG(D_NET, "msg %s->%s is a recovery ping. retry# %d\n",
+ libcfs_nid2str(msg->msg_from),
+ libcfs_nid2str(msg->msg_target.nid),
+ msg->msg_retry_count);
+ return -ENOTRECOVERABLE;
+ }
+
+ /*
+ * if we explicitly indicated we don't want to resend then just
+ * return
+ */
+ if (msg->msg_no_resend) {
+ CDEBUG(D_NET, "msg %s->%s requested no resend. retry# %d\n",
+ libcfs_nid2str(msg->msg_from),
+ libcfs_nid2str(msg->msg_target.nid),
+ msg->msg_retry_count);
+ return -ENOTRECOVERABLE;
+ }
+
+ /* check if the message has exceeded the number of retries */
+ if (msg->msg_retry_count >= lnet_retry_count) {
+ CNETERR("msg %s->%s exceeded retry count %d\n",
+ libcfs_nid2str(msg->msg_from),
+ libcfs_nid2str(msg->msg_target.nid),
+ msg->msg_retry_count);
+ return -ENOTRECOVERABLE;
+ }
+
+ cpt = msg->msg_tx_cpt;
+ lnet_net_lock(cpt);
+
+ /* check again under lock */
+ if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
+ lnet_net_unlock(cpt);
+ return -ESHUTDOWN;
+ }
+
+ container = the_lnet.ln_msg_containers[cpt];
+ my_slot =
+ lnet_check_finalize_recursion_locked(msg,
+ &container->msc_resending,
+ container->msc_nfinalizers,
+ container->msc_resenders);
+
+ /* enough threads are resending */
+ if (my_slot == -1) {
+ lnet_net_unlock(cpt);
+ return 0;
+ }
+
+ while (!list_empty(&container->msc_resending)) {
+ msg = list_entry(container->msc_resending.next,
+ struct lnet_msg, msg_list);
+ list_del(&msg->msg_list);
+
+ /*
+ * resending the message will require us to call
+ * lnet_msg_decommit_tx() which will return the credit
+ * which this message holds. This could trigger another
+ * queued message to be sent. If that message fails and
+ * requires a resend we will recurse.
+ * But since at this point the slot is taken, the message
+ * will be queued in the container and dealt with
+ * later. This breaks the recursion.
+ */
+ lnet_resend_msg_locked(msg);
+ }
+
+ /*
+ * msc_resenders is an array of process pointers. Each entry holds
+ * a pointer to the current process operating on the message. An
+ * array entry is created per CPT. If the array slot is already
+ * set, then it means that there is a thread on the CPT currently
+ * resending a message.
+ * Once the thread finishes clear the slot to enable the thread to
+ * take on more resend work.
+ */
+ container->msc_resenders[my_slot] = NULL;
+ lnet_net_unlock(cpt);
+
+ return 0;
+}
+
+/*
+ * Do a health check on the message:
+ * return -1 if we're not going to handle the error or
+ * if we've reached the maximum number of retries.
+ * success case will return -1 as well
+ * return 0 if it the message is requeued for send
+ */
+static int
+lnet_health_check(struct lnet_msg *msg)