+ /*
+ * 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_nidstr(&msg->msg_hdr.src_nid),
+ libcfs_nidstr(&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_nidstr(&msg->msg_from),
+ libcfs_nidstr(&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_nidstr(&msg->msg_from),
+ libcfs_nidstr(&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_nidstr(&msg->msg_from),
+ libcfs_nidstr(&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 ((msg = list_first_entry_or_null(&container->msc_resending,
+ struct lnet_msg,
+ msg_list)) != NULL) {
+ 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;