+ struct lnet_peer *lp = event->md.user_ptr;
+ struct lnet_ping_buffer *pbuf;
+ int rc;
+
+ /* discovery needs to take another look */
+ rc = LNET_REDISCOVER_PEER;
+
+ switch (event->type) {
+ case LNET_EVENT_ACK:
+ lnet_discovery_event_ack(lp, event);
+ break;
+ case LNET_EVENT_REPLY:
+ lnet_discovery_event_reply(lp, event);
+ break;
+ case LNET_EVENT_SEND:
+ /* Only send failure triggers a retry. */
+ rc = lnet_discovery_event_send(lp, event);
+ break;
+ case LNET_EVENT_UNLINK:
+ /* Valid event, nothing to do here. */
+ break;
+ default:
+ /* Invalid events. */
+ LBUG();
+ }
+ lnet_net_lock(LNET_LOCK_EX);
+ if (event->unlinked) {
+ pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
+ lnet_ping_buffer_decref(pbuf);
+ lnet_peer_decref_locked(lp);
+ }
+ if (rc == LNET_REDISCOVER_PEER) {
+ list_move_tail(&lp->lp_dc_list, &the_lnet.ln_dc_request);
+ wake_up(&the_lnet.ln_dc_waitq);
+ }
+ lnet_net_unlock(LNET_LOCK_EX);
+}
+
+/*
+ * Build a peer from incoming data.
+ *
+ * The NIDs in the incoming data are supposed to be structured as follows:
+ * - loopback
+ * - primary NID
+ * - other NIDs in same net
+ * - NIDs in second net
+ * - NIDs in third net
+ * - ...
+ * This due to the way the list of NIDs in the data is created.
+ *
+ * Note that this function will mark the peer uptodate unless an
+ * ENOMEM is encontered. All other errors are due to a conflict
+ * between the DLC configuration and what discovery sees. We treat DLC
+ * as binding, and therefore set the NIDS_UPTODATE flag to prevent the
+ * peer from becoming stuck in discovery.
+ */
+static int lnet_peer_merge_data(struct lnet_peer *lp,
+ struct lnet_ping_buffer *pbuf)
+{
+ struct lnet_peer_ni *lpni;
+ lnet_nid_t *curnis = NULL;
+ lnet_nid_t *addnis = NULL;
+ lnet_nid_t *delnis = NULL;
+ unsigned flags;
+ int ncurnis;
+ int naddnis;
+ int ndelnis;
+ int nnis = 0;
+ int i;
+ int j;
+ int rc;
+
+ flags = LNET_PEER_DISCOVERED;
+ if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
+ flags |= LNET_PEER_MULTI_RAIL;
+
+ nnis = MAX(lp->lp_nnis, pbuf->pb_info.pi_nnis);
+ LIBCFS_ALLOC(curnis, nnis * sizeof(lnet_nid_t));
+ LIBCFS_ALLOC(addnis, nnis * sizeof(lnet_nid_t));
+ LIBCFS_ALLOC(delnis, nnis * sizeof(lnet_nid_t));
+ if (!curnis || !addnis || !delnis) {
+ rc = -ENOMEM;
+ goto out;
+ }
+ ncurnis = 0;
+ naddnis = 0;
+ ndelnis = 0;
+
+ /* Construct the list of NIDs present in peer. */
+ lpni = NULL;
+ while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
+ curnis[ncurnis++] = lpni->lpni_nid;
+
+ /*
+ * Check for NIDs in pbuf not present in curnis[].
+ * The loop starts at 1 to skip the loopback NID.
+ */
+ for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
+ for (j = 0; j < ncurnis; j++)
+ if (pbuf->pb_info.pi_ni[i].ns_nid == curnis[j])
+ break;
+ if (j == ncurnis)
+ addnis[naddnis++] = pbuf->pb_info.pi_ni[i].ns_nid;
+ }
+ /*
+ * Check for NIDs in curnis[] not present in pbuf.
+ * The nested loop starts at 1 to skip the loopback NID.
+ *
+ * But never add the loopback NID to delnis[]: if it is
+ * present in curnis[] then this peer is for this node.
+ */
+ for (i = 0; i < ncurnis; i++) {
+ if (LNET_NETTYP(LNET_NIDNET(curnis[i])) == LOLND)
+ continue;
+ for (j = 1; j < pbuf->pb_info.pi_nnis; j++)
+ if (curnis[i] == pbuf->pb_info.pi_ni[j].ns_nid)
+ break;
+ if (j == pbuf->pb_info.pi_nnis)
+ delnis[ndelnis++] = curnis[i];
+ }
+
+ for (i = 0; i < naddnis; i++) {
+ rc = lnet_peer_add_nid(lp, addnis[i], flags);
+ if (rc) {
+ CERROR("Error adding NID %s to peer %s: %d\n",
+ libcfs_nid2str(addnis[i]),
+ libcfs_nid2str(lp->lp_primary_nid), rc);
+ if (rc == -ENOMEM)
+ goto out;
+ }
+ }
+ for (i = 0; i < ndelnis; i++) {
+ rc = lnet_peer_del_nid(lp, delnis[i], flags);
+ if (rc) {
+ CERROR("Error deleting NID %s from peer %s: %d\n",
+ libcfs_nid2str(delnis[i]),
+ libcfs_nid2str(lp->lp_primary_nid), rc);
+ if (rc == -ENOMEM)
+ goto out;
+ }
+ }
+ /*
+ * Errors other than -ENOMEM are due to peers having been
+ * configured with DLC. Ignore these because DLC overrides
+ * Discovery.
+ */
+ rc = 0;
+out:
+ LIBCFS_FREE(curnis, nnis * sizeof(lnet_nid_t));
+ LIBCFS_FREE(addnis, nnis * sizeof(lnet_nid_t));
+ LIBCFS_FREE(delnis, nnis * sizeof(lnet_nid_t));
+ lnet_ping_buffer_decref(pbuf);
+ CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
+
+ if (rc) {
+ spin_lock(&lp->lp_lock);
+ lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
+ lp->lp_state |= LNET_PEER_PING_REQUIRED;
+ spin_unlock(&lp->lp_lock);
+ }
+ return rc;
+}
+
+/*
+ * The data in pbuf says lp is its primary peer, but the data was
+ * received by a different peer. Try to update lp with the data.
+ */
+static int
+lnet_peer_set_primary_data(struct lnet_peer *lp, struct lnet_ping_buffer *pbuf)
+{
+ lnet_handle_md_t mdh;
+
+ /* Queue lp for discovery, and force it on the request queue. */
+ lnet_net_lock(LNET_LOCK_EX);
+ if (lnet_peer_queue_for_discovery(lp))
+ list_move(&lp->lp_dc_list, &the_lnet.ln_dc_request);
+ lnet_net_unlock(LNET_LOCK_EX);
+
+ LNetInvalidateHandle(&mdh);
+
+ /*
+ * Decide whether we can move the peer to the DATA_PRESENT state.
+ *
+ * We replace stale data for a multi-rail peer, repair PING_FAILED
+ * status, and preempt PING_REQUIRED.
+ *
+ * If after that we have DATA_PRESENT, we merge it into this peer.
+ */
+ spin_lock(&lp->lp_lock);
+ if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
+ if (lp->lp_peer_seqno < LNET_PING_BUFFER_SEQNO(pbuf)) {
+ lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
+ } else if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
+ lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
+ lnet_ping_buffer_decref(pbuf);
+ pbuf = lp->lp_data;
+ lp->lp_data = NULL;
+ }
+ }
+ if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
+ lnet_ping_buffer_decref(lp->lp_data);
+ lp->lp_data = NULL;
+ lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
+ }
+ if (lp->lp_state & LNET_PEER_PING_FAILED) {
+ mdh = lp->lp_ping_mdh;
+ LNetInvalidateHandle(&lp->lp_ping_mdh);
+ lp->lp_state &= ~LNET_PEER_PING_FAILED;
+ lp->lp_ping_error = 0;
+ }
+ if (lp->lp_state & LNET_PEER_PING_REQUIRED) {
+ lp->lp_state &= ~LNET_PEER_PING_REQUIRED;
+ }
+ lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
+ spin_unlock(&lp->lp_lock);
+
+ if (!LNetHandleIsInvalid(mdh))
+ LNetMDUnlink(mdh);
+
+ if (pbuf)
+ return lnet_peer_merge_data(lp, pbuf);
+
+ CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
+ return 0;
+}
+
+/*
+ * Update a peer using the data received.
+ */
+static int lnet_peer_data_present(struct lnet_peer *lp)
+__must_hold(&lp->lp_lock)
+{
+ struct lnet_ping_buffer *pbuf;
+ struct lnet_peer_ni *lpni;
+ lnet_nid_t nid = LNET_NID_ANY;
+ unsigned flags;
+ int rc = 0;
+
+ pbuf = lp->lp_data;
+ lp->lp_data = NULL;
+ lp->lp_state &= ~LNET_PEER_DATA_PRESENT;
+ lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
+ spin_unlock(&lp->lp_lock);
+
+ /*
+ * Modifications of peer structures are done while holding the
+ * ln_api_mutex. A global lock is required because we may be
+ * modifying multiple peer structures, and a mutex greatly
+ * simplifies memory management.
+ *
+ * The actual changes to the data structures must also protect
+ * against concurrent lookups, for which the lnet_net_lock in
+ * LNET_LOCK_EX mode is used.
+ */
+ mutex_lock(&the_lnet.ln_api_mutex);
+ if (the_lnet.ln_state != LNET_STATE_RUNNING) {
+ rc = -ESHUTDOWN;
+ goto out;
+ }
+
+ /*
+ * If this peer is not on the peer list then it is being torn
+ * down, and our reference count may be all that is keeping it
+ * alive. Don't do any work on it.
+ */
+ if (list_empty(&lp->lp_peer_list))
+ goto out;
+
+ flags = LNET_PEER_DISCOVERED;
+ if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)
+ flags |= LNET_PEER_MULTI_RAIL;
+
+ /*
+ * Check whether the primary NID in the message matches the
+ * primary NID of the peer. If it does, update the peer, if
+ * it it does not, check whether there is already a peer with
+ * that primary NID. If no such peer exists, try to update
+ * the primary NID of the current peer (allowed if it was
+ * created due to message traffic) and complete the update.
+ * If the peer did exist, hand off the data to it.
+ *
+ * The peer for the loopback interface is a special case: this
+ * is the peer for the local node, and we want to set its
+ * primary NID to the correct value here.
+ */
+ if (pbuf->pb_info.pi_nnis > 1)
+ nid = pbuf->pb_info.pi_ni[1].ns_nid;
+ if (LNET_NETTYP(LNET_NIDNET(lp->lp_primary_nid)) == LOLND) {
+ rc = lnet_peer_set_primary_nid(lp, nid, flags);
+ if (!rc)
+ rc = lnet_peer_merge_data(lp, pbuf);
+ } else if (lp->lp_primary_nid == nid) {
+ rc = lnet_peer_merge_data(lp, pbuf);
+ } else {
+ lpni = lnet_find_peer_ni_locked(nid);
+ if (!lpni) {
+ rc = lnet_peer_set_primary_nid(lp, nid, flags);
+ if (rc) {
+ CERROR("Primary NID error %s versus %s: %d\n",
+ libcfs_nid2str(lp->lp_primary_nid),
+ libcfs_nid2str(nid), rc);
+ } else {
+ rc = lnet_peer_merge_data(lp, pbuf);
+ }
+ } else {
+ rc = lnet_peer_set_primary_data(
+ lpni->lpni_peer_net->lpn_peer, pbuf);
+ lnet_peer_ni_decref_locked(lpni);
+ }
+ }
+out:
+ CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
+ mutex_unlock(&the_lnet.ln_api_mutex);
+
+ spin_lock(&lp->lp_lock);
+ /* Tell discovery to re-check the peer immediately. */
+ if (!rc)
+ rc = LNET_REDISCOVER_PEER;
+ return rc;
+}
+
+/*
+ * A ping failed. Clear the PING_FAILED state and set the
+ * PING_REQUIRED state, to ensure a retry even if discovery is
+ * disabled. This avoids being left with incorrect state.
+ */
+static int lnet_peer_ping_failed(struct lnet_peer *lp)
+__must_hold(&lp->lp_lock)
+{
+ lnet_handle_md_t mdh;
+ int rc;
+
+ mdh = lp->lp_ping_mdh;
+ LNetInvalidateHandle(&lp->lp_ping_mdh);
+ lp->lp_state &= ~LNET_PEER_PING_FAILED;
+ lp->lp_state |= LNET_PEER_PING_REQUIRED;
+ rc = lp->lp_ping_error;
+ lp->lp_ping_error = 0;
+ spin_unlock(&lp->lp_lock);
+
+ LNetMDUnlink(mdh);
+
+ CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
+
+ spin_lock(&lp->lp_lock);
+ return rc ? rc : LNET_REDISCOVER_PEER;
+}
+
+/*
+ * Select NID to send a Ping or Push to.
+ */
+static lnet_nid_t lnet_peer_select_nid(struct lnet_peer *lp)
+{
+ struct lnet_peer_ni *lpni;
+
+ /* Look for a direct-connected NID for this peer. */
+ lpni = NULL;
+ while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
+ if (!lnet_is_peer_ni_healthy_locked(lpni))
+ continue;
+ if (!lnet_get_net_locked(lpni->lpni_peer_net->lpn_net_id))
+ continue;
+ break;
+ }
+ if (lpni)
+ return lpni->lpni_nid;
+
+ /* Look for a routed-connected NID for this peer. */
+ lpni = NULL;
+ while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
+ if (!lnet_is_peer_ni_healthy_locked(lpni))
+ continue;
+ if (!lnet_find_rnet_locked(lpni->lpni_peer_net->lpn_net_id))
+ continue;
+ break;
+ }
+ if (lpni)
+ return lpni->lpni_nid;
+
+ return LNET_NID_ANY;
+}
+
+/* Active side of ping. */
+static int lnet_peer_send_ping(struct lnet_peer *lp)
+__must_hold(&lp->lp_lock)
+{
+ lnet_md_t md = { NULL };
+ lnet_process_id_t id;
+ struct lnet_ping_buffer *pbuf;
+ int nnis;
+ int rc;
+ int cpt;
+
+ lp->lp_state |= LNET_PEER_PING_SENT;
+ lp->lp_state &= ~LNET_PEER_PING_REQUIRED;
+ spin_unlock(&lp->lp_lock);
+
+ nnis = MAX(lp->lp_data_nnis, LNET_MIN_INTERFACES);
+ pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
+ if (!pbuf) {
+ rc = -ENOMEM;
+ goto fail_error;
+ }
+
+ /* initialize md content */
+ md.start = &pbuf->pb_info;
+ md.length = LNET_PING_INFO_SIZE(nnis);
+ md.threshold = 2; /* GET/REPLY */
+ md.max_size = 0;
+ md.options = LNET_MD_TRUNCATE;
+ md.user_ptr = lp;
+ md.eq_handle = the_lnet.ln_dc_eqh;
+
+ rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_ping_mdh);
+ if (rc != 0) {
+ lnet_ping_buffer_decref(pbuf);
+ CERROR("Can't bind MD: %d\n", rc);
+ goto fail_error;
+ }
+ cpt = lnet_net_lock_current();
+ /* Refcount for MD. */
+ lnet_peer_addref_locked(lp);
+ id.pid = LNET_PID_LUSTRE;
+ id.nid = lnet_peer_select_nid(lp);
+ lnet_net_unlock(cpt);
+
+ if (id.nid == LNET_NID_ANY) {
+ rc = -EHOSTUNREACH;
+ goto fail_unlink_md;
+ }
+
+ rc = LNetGet(LNET_NID_ANY, lp->lp_ping_mdh, id,
+ LNET_RESERVED_PORTAL,
+ LNET_PROTO_PING_MATCHBITS, 0);
+
+ if (rc)
+ goto fail_unlink_md;
+
+ CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
+
+ spin_lock(&lp->lp_lock);
+ return 0;
+
+fail_unlink_md:
+ LNetMDUnlink(lp->lp_ping_mdh);
+ LNetInvalidateHandle(&lp->lp_ping_mdh);
+fail_error:
+ CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
+ /*
+ * The errors that get us here are considered hard errors and
+ * cause Discovery to terminate. Se we clear PING_SENT, but do
+ * not set either PING_FAILED or PING_REQUIRED.
+ */
+ spin_lock(&lp->lp_lock);
+ lp->lp_state &= ~LNET_PEER_PING_SENT;
+ return rc;
+}
+
+/*
+ * This function exists because you cannot call LNetMDUnlink() from an
+ * event handler.
+ */
+static int lnet_peer_push_failed(struct lnet_peer *lp)
+__must_hold(&lp->lp_lock)
+{
+ lnet_handle_md_t mdh;
+ int rc;
+
+ mdh = lp->lp_push_mdh;
+ LNetInvalidateHandle(&lp->lp_push_mdh);
+ lp->lp_state &= ~LNET_PEER_PUSH_FAILED;
+ rc = lp->lp_push_error;
+ lp->lp_push_error = 0;
+ spin_unlock(&lp->lp_lock);
+
+ LNetMDUnlink(mdh);
+
+ CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
+ spin_lock(&lp->lp_lock);
+ return rc ? rc : LNET_REDISCOVER_PEER;
+}
+
+/* Active side of push. */
+static int lnet_peer_send_push(struct lnet_peer *lp)
+__must_hold(&lp->lp_lock)
+{
+ struct lnet_ping_buffer *pbuf;
+ lnet_process_id_t id;
+ lnet_md_t md;
+ int cpt;
+ int rc;
+
+ LASSERT(lp->lp_state & LNET_PEER_MULTI_RAIL);
+
+ lp->lp_state |= LNET_PEER_PUSH_SENT;
+ spin_unlock(&lp->lp_lock);
+
+ cpt = lnet_net_lock_current();
+ pbuf = the_lnet.ln_ping_target;
+ lnet_ping_buffer_addref(pbuf);
+ lnet_net_unlock(cpt);
+
+ /* Push source MD */
+ md.start = &pbuf->pb_info;
+ md.length = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
+ md.threshold = 2; /* Put/Ack */
+ md.max_size = 0;
+ md.options = 0;
+ md.eq_handle = the_lnet.ln_dc_eqh;
+ md.user_ptr = lp;
+
+ rc = LNetMDBind(md, LNET_UNLINK, &lp->lp_push_mdh);
+ if (rc) {
+ lnet_ping_buffer_decref(pbuf);
+ CERROR("Can't bind push source MD: %d\n", rc);
+ goto fail_error;
+ }
+ cpt = lnet_net_lock_current();
+ /* Refcount for MD. */
+ lnet_peer_addref_locked(lp);
+ id.pid = LNET_PID_LUSTRE;
+ id.nid = lnet_peer_select_nid(lp);
+ lnet_net_unlock(cpt);
+
+ if (id.nid == LNET_NID_ANY) {
+ rc = -EHOSTUNREACH;
+ goto fail_unlink;
+ }
+
+ rc = LNetPut(LNET_NID_ANY, lp->lp_push_mdh,
+ LNET_ACK_REQ, id, LNET_RESERVED_PORTAL,
+ LNET_PROTO_PING_MATCHBITS, 0, 0);
+
+ if (rc)
+ goto fail_unlink;
+
+ CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
+
+ spin_lock(&lp->lp_lock);
+ return 0;
+
+fail_unlink:
+ LNetMDUnlink(lp->lp_push_mdh);
+ LNetInvalidateHandle(&lp->lp_push_mdh);
+fail_error:
+ CDEBUG(D_NET, "peer %s: %d\n", libcfs_nid2str(lp->lp_primary_nid), rc);
+ /*
+ * The errors that get us here are considered hard errors and
+ * cause Discovery to terminate. Se we clear PUSH_SENT, but do
+ * not set PUSH_FAILED.
+ */
+ spin_lock(&lp->lp_lock);
+ lp->lp_state &= ~LNET_PEER_PUSH_SENT;
+ return rc;
+}
+
+/*
+ * An unrecoverable error was encountered during discovery.
+ * Set error status in peer and abort discovery.
+ */
+static void lnet_peer_discovery_error(struct lnet_peer *lp, int error)
+{
+ CDEBUG(D_NET, "Discovery error %s: %d\n",
+ libcfs_nid2str(lp->lp_primary_nid), error);
+
+ spin_lock(&lp->lp_lock);
+ lp->lp_dc_error = error;
+ lp->lp_state &= ~LNET_PEER_DISCOVERING;
+ lp->lp_state |= LNET_PEER_UNDISCOVERED;
+ spin_unlock(&lp->lp_lock);
+}
+
+/*
+ * Mark the peer as undiscovered because discovery was disabled.
+ */
+static int lnet_peer_undiscovered(struct lnet_peer *lp)
+__must_hold(&lp->lp_lock)
+{
+
+ lp->lp_state &= ~(LNET_PEER_DISCOVERED | LNET_PEER_DISCOVERING);
+ lp->lp_state |= LNET_PEER_UNDISCOVERED;
+
+ CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
+
+ return 0;
+}
+
+/*
+ * Mark the peer as discovered.
+ */
+static int lnet_peer_discovered(struct lnet_peer *lp)
+__must_hold(&lp->lp_lock)
+{
+ lp->lp_state |= LNET_PEER_DISCOVERED;
+ lp->lp_state &= ~LNET_PEER_DISCOVERING;
+
+ CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
+
+ return 0;