if (len != 0)
lnet_setpayloadbuffer(msg);
- memset(&msg->msg_hdr, 0, sizeof(msg->msg_hdr));
- msg->msg_hdr.type = cpu_to_le32(type);
- msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
- msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
+ memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
+ msg->msg_hdr.type = cpu_to_le32(type);
+ msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
/* src_nid will be set later */
- msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
+ msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
msg->msg_hdr.payload_length = cpu_to_le32(len);
}
return rc;
}
-/* NB: caller shall hold a ref on 'lp' as I'd drop lnet_net_lock */
+/*
+ * This function can be called from two paths:
+ * 1. when sending a message
+ * 2. when decommiting a message (lnet_msg_decommit_tx())
+ * In both these cases the peer_ni should have it's reference count
+ * acquired by the caller and therefore it is safe to drop the spin
+ * lock before calling lnd_query()
+ */
static void
lnet_ni_query_locked(lnet_ni_t *ni, struct lnet_peer_ni *lp)
{
cfs_time_t last_alive = 0;
+ int cpt = lnet_cpt_of_nid_locked(lp->lpni_nid, ni);
LASSERT(lnet_peer_aliveness_enabled(lp));
LASSERT(ni->ni_net->net_lnd->lnd_query != NULL);
- lnet_net_unlock(lp->lpni_cpt);
+ lnet_net_unlock(cpt);
(ni->ni_net->net_lnd->lnd_query)(ni, lp->lpni_nid, &last_alive);
- lnet_net_lock(lp->lpni_cpt);
+ lnet_net_lock(cpt);
lp->lpni_last_query = cfs_time_current();
* Trust lnet_notify() if it has more recent aliveness news, but
* ignore the initial assumed death (see lnet_peers_start_down()).
*/
+ spin_lock(&lp->lpni_lock);
if (!lp->lpni_alive && lp->lpni_alive_count > 0 &&
- cfs_time_aftereq(lp->lpni_timestamp, lp->lpni_last_alive))
+ cfs_time_aftereq(lp->lpni_timestamp, lp->lpni_last_alive)) {
+ spin_unlock(&lp->lpni_lock);
return 0;
+ }
deadline =
cfs_time_add(lp->lpni_last_alive,
* case, and moreover lpni_last_alive at peer creation is assumed.
*/
if (alive && !lp->lpni_alive &&
- !(lnet_isrouter(lp) && lp->lpni_alive_count == 0))
+ !(lnet_isrouter(lp) && lp->lpni_alive_count == 0)) {
+ spin_unlock(&lp->lpni_lock);
lnet_notify_locked(lp, 0, 1, lp->lpni_last_alive);
+ } else {
+ spin_unlock(&lp->lpni_lock);
+ }
return alive;
}
the_lnet.ln_counters[cpt]->drop_count++;
the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
lnet_net_unlock(cpt);
+ if (msg->msg_txpeer)
+ atomic_inc(&msg->msg_txpeer->lpni_stats.drop_count);
+ if (msg->msg_txni)
+ atomic_inc(&msg->msg_txni->ni_stats.drop_count);
CNETERR("Dropping message for %s: peer not alive\n",
libcfs_id2str(msg->msg_target));
}
if (!msg->msg_peertxcredit) {
+ spin_lock(&lp->lpni_lock);
LASSERT((lp->lpni_txcredits < 0) ==
!list_empty(&lp->lpni_txq));
if (lp->lpni_txcredits < 0) {
msg->msg_tx_delayed = 1;
list_add_tail(&msg->msg_list, &lp->lpni_txq);
+ spin_unlock(&lp->lpni_lock);
return LNET_CREDIT_WAIT;
}
+ spin_unlock(&lp->lpni_lock);
}
if (!msg->msg_txcredit) {
msg->msg_txcredit = 1;
tq->tq_credits--;
+ atomic_dec(&ni->ni_tx_credits);
if (tq->tq_credits < tq->tq_credits_min)
tq->tq_credits_min = tq->tq_credits;
LASSERT(!do_recv || msg->msg_rx_delayed);
if (!msg->msg_peerrtrcredit) {
+ spin_lock(&lp->lpni_lock);
LASSERT((lp->lpni_rtrcredits < 0) ==
!list_empty(&lp->lpni_rtrq));
LASSERT(msg->msg_rx_ready_delay);
msg->msg_rx_delayed = 1;
list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
+ spin_unlock(&lp->lpni_lock);
return LNET_CREDIT_WAIT;
}
+ spin_unlock(&lp->lpni_lock);
}
rbp = lnet_msg2bufpool(msg);
!list_empty(&tq->tq_delayed));
tq->tq_credits++;
+ atomic_inc(&ni->ni_tx_credits);
if (tq->tq_credits <= 0) {
msg2 = list_entry(tq->tq_delayed.next,
lnet_msg_t, msg_list);
LASSERT(msg2->msg_txni == ni);
LASSERT(msg2->msg_tx_delayed);
+ LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
(void) lnet_post_send_locked(msg2, 1);
}
/* give back peer txcredits */
msg->msg_peertxcredit = 0;
+ spin_lock(&txpeer->lpni_lock);
LASSERT((txpeer->lpni_txcredits < 0) ==
!list_empty(&txpeer->lpni_txq));
txpeer->lpni_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
- LASSERT (txpeer->lpni_txqnob >= 0);
+ LASSERT(txpeer->lpni_txqnob >= 0);
txpeer->lpni_txcredits++;
if (txpeer->lpni_txcredits <= 0) {
msg2 = list_entry(txpeer->lpni_txq.next,
lnet_msg_t, msg_list);
list_del(&msg2->msg_list);
+ spin_unlock(&txpeer->lpni_lock);
LASSERT(msg2->msg_txpeer == txpeer);
LASSERT(msg2->msg_tx_delayed);
- (void) lnet_post_send_locked(msg2, 1);
+ if (msg2->msg_tx_cpt != msg->msg_tx_cpt) {
+ lnet_net_unlock(msg->msg_tx_cpt);
+ lnet_net_lock(msg2->msg_tx_cpt);
+ }
+ (void) lnet_post_send_locked(msg2, 1);
+ if (msg2->msg_tx_cpt != msg->msg_tx_cpt) {
+ lnet_net_unlock(msg2->msg_tx_cpt);
+ lnet_net_lock(msg->msg_tx_cpt);
+ }
+ } else {
+ spin_unlock(&txpeer->lpni_lock);
}
- }
+ }
if (txni != NULL) {
msg->msg_txni = NULL;
}
if (txpeer != NULL) {
+ /*
+ * TODO:
+ * Once the patch for the health comes in we need to set
+ * the health of the peer ni to bad when we fail to send
+ * a message.
+ * int status = msg->msg_ev.status;
+ * if (status != 0)
+ * lnet_set_peer_ni_health_locked(txpeer, false)
+ */
msg->msg_txpeer = NULL;
lnet_peer_ni_decref_locked(txpeer);
}
void
lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
{
- lnet_msg_t *msg;
- lnet_msg_t *tmp;
- struct list_head drop;
-
- INIT_LIST_HEAD(&drop);
-
- list_splice_init(list, &drop);
+ lnet_msg_t *msg;
+ lnet_msg_t *tmp;
lnet_net_unlock(cpt);
- list_for_each_entry_safe(msg, tmp, &drop, msg_list) {
+ list_for_each_entry_safe(msg, tmp, list, msg_list) {
lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
0, 0, 0, msg->msg_hdr.payload_length);
list_del_init(&msg->msg_list);
/* give back peer router credits */
msg->msg_peerrtrcredit = 0;
+ spin_lock(&rxpeer->lpni_lock);
LASSERT((rxpeer->lpni_rtrcredits < 0) ==
!list_empty(&rxpeer->lpni_rtrq));
/* drop all messages which are queued to be routed on that
* peer. */
if (!the_lnet.ln_routing) {
- lnet_drop_routed_msgs_locked(&rxpeer->lpni_rtrq,
- msg->msg_rx_cpt);
+ struct list_head drop;
+ INIT_LIST_HEAD(&drop);
+ list_splice_init(&rxpeer->lpni_rtrq, &drop);
+ spin_unlock(&rxpeer->lpni_lock);
+ lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
} else if (rxpeer->lpni_rtrcredits <= 0) {
msg2 = list_entry(rxpeer->lpni_rtrq.next,
lnet_msg_t, msg_list);
list_del(&msg2->msg_list);
-
+ spin_unlock(&rxpeer->lpni_lock);
(void) lnet_post_routed_recv_locked(msg2, 1);
+ } else {
+ spin_unlock(&rxpeer->lpni_lock);
}
}
if (rxni != NULL) {
}
static int
+lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
+{
+ if (p1->lpni_txqnob < p2->lpni_txqnob)
+ return 1;
+
+ if (p1->lpni_txqnob > p2->lpni_txqnob)
+ return -1;
+
+ if (p1->lpni_txcredits > p2->lpni_txcredits)
+ return 1;
+
+ if (p1->lpni_txcredits < p2->lpni_txcredits)
+ return -1;
+
+ return 0;
+}
+
+static int
lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
{
struct lnet_peer_ni *p1 = r1->lr_gateway;
struct lnet_peer_ni *p2 = r2->lr_gateway;
int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
+ int rc;
if (r1->lr_priority < r2->lr_priority)
return 1;
if (r1->lr_priority > r2->lr_priority)
- return -ERANGE;
+ return -1;
if (r1_hops < r2_hops)
return 1;
if (r1_hops > r2_hops)
- return -ERANGE;
-
- if (p1->lpni_txqnob < p2->lpni_txqnob)
- return 1;
+ return -1;
- if (p1->lpni_txqnob > p2->lpni_txqnob)
- return -ERANGE;
-
- if (p1->lpni_txcredits > p2->lpni_txcredits)
- return 1;
-
- if (p1->lpni_txcredits < p2->lpni_txcredits)
- return -ERANGE;
+ rc = lnet_compare_peers(p1, p2);
+ if (rc)
+ return rc;
if (r1->lr_seq - r2->lr_seq <= 0)
return 1;
- return -ERANGE;
+ return -1;
}
static struct lnet_peer_ni *
return lpni_best;
}
-int
-lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
+static int
+lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
+ struct lnet_msg *msg, lnet_nid_t rtr_nid)
{
- lnet_nid_t dst_nid = msg->msg_target.nid;
- struct lnet_ni *src_ni;
- struct lnet_ni *local_ni;
- struct lnet_peer_ni *lp;
- int cpt;
- int cpt2;
- int rc;
+ struct lnet_ni *best_ni;
+ struct lnet_peer_ni *best_lpni;
+ struct lnet_peer_ni *best_gw;
+ struct lnet_peer_ni *lpni;
+ struct lnet_peer *peer;
+ struct lnet_peer_net *peer_net;
+ struct lnet_net *local_net;
+ struct lnet_ni *ni;
+ __u32 seq;
+ int cpt, cpt2, rc;
+ bool routing;
+ bool ni_is_pref;
+ bool preferred;
+ int best_credits;
+ int best_lpni_credits;
+ int md_cpt;
+ int shortest_distance;
- /* NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
- * but we might want to use pre-determined router for ACK/REPLY
- * in the future */
- /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
- LASSERT(msg->msg_txpeer == NULL);
- LASSERT(!msg->msg_sending);
- LASSERT(!msg->msg_target_is_router);
- LASSERT(!msg->msg_receiving);
+ /*
+ * get an initial CPT to use for locking. The idea here is not to
+ * serialize the calls to select_pathway, so that as many
+ * operations can run concurrently as possible. To do that we use
+ * the CPT where this call is being executed. Later on when we
+ * determine the CPT to use in lnet_message_commit, we switch the
+ * lock and check if there was any configuration change. If none,
+ * then we proceed, if there is, then we restart the operation.
+ */
+ cpt = lnet_net_lock_current();
+again:
+ best_ni = NULL;
+ best_lpni = NULL;
+ best_gw = NULL;
+ local_net = NULL;
+ routing = false;
- msg->msg_sending = 1;
+ seq = lnet_get_dlc_seq_locked();
- LASSERT(!msg->msg_tx_committed);
- local_ni = lnet_net2ni(LNET_NIDNET(dst_nid));
- cpt = lnet_cpt_of_nid(rtr_nid == LNET_NID_ANY ? dst_nid : rtr_nid,
- local_ni);
- again:
- if (the_lnet.ln_shutdown)
+ if (the_lnet.ln_shutdown) {
+ lnet_net_unlock(cpt);
return -ESHUTDOWN;
- lnet_net_lock(cpt);
+ }
- if (src_nid == LNET_NID_ANY) {
- src_ni = NULL;
- } else {
- src_ni = lnet_nid2ni_locked(src_nid, cpt);
- if (src_ni == NULL) {
+ if (msg->msg_md != NULL)
+ /* get the cpt of the MD, used during NUMA based selection */
+ md_cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
+ else
+ md_cpt = CFS_CPT_ANY;
+
+ peer = lnet_find_or_create_peer_locked(dst_nid, cpt);
+ if (IS_ERR(peer)) {
+ lnet_net_unlock(cpt);
+ return PTR_ERR(peer);
+ }
+
+ /* If peer is not healthy then can not send anything to it */
+ if (!lnet_is_peer_healthy_locked(peer)) {
+ lnet_net_unlock(cpt);
+ return -EHOSTUNREACH;
+ }
+
+ if (!peer->lp_multi_rail && lnet_get_num_peer_nis(peer) > 1) {
+ CERROR("peer %s is declared to be non MR capable, "
+ "yet configured with more than one NID\n",
+ libcfs_nid2str(dst_nid));
+ return -EINVAL;
+ }
+
+ /*
+ * STEP 1: first jab at determineing best_ni
+ * if src_nid is explicitly specified, then best_ni is already
+ * pre-determiend for us. Otherwise we need to select the best
+ * one to use later on
+ */
+ if (src_nid != LNET_NID_ANY) {
+ best_ni = lnet_nid2ni_locked(src_nid, cpt);
+ if (!best_ni) {
lnet_net_unlock(cpt);
LCONSOLE_WARN("Can't send to %s: src %s is not a "
"local nid\n", libcfs_nid2str(dst_nid),
libcfs_nid2str(src_nid));
return -EINVAL;
}
- LASSERT(!msg->msg_routing);
- }
-
- /* Is this for someone on a local network? */
- local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid), cpt);
- if (local_ni != NULL) {
- if (src_ni == NULL) {
- src_ni = local_ni;
- src_nid = src_ni->ni_nid;
- } else if (src_ni != local_ni) {
+ if (best_ni->ni_net->net_id != LNET_NIDNET(dst_nid)) {
lnet_net_unlock(cpt);
LCONSOLE_WARN("No route to %s via from %s\n",
libcfs_nid2str(dst_nid),
libcfs_nid2str(src_nid));
return -EINVAL;
}
+ goto pick_peer;
+ }
- LASSERT(src_nid != LNET_NID_ANY);
- lnet_msg_commit(msg, cpt);
+ /*
+ * Decide whether we need to route to peer_ni.
+ * Get the local net that I need to be on to be able to directly
+ * send to that peer.
+ *
+ * a. Find the peer which the dst_nid belongs to.
+ * b. Iterate through each of the peer_nets/nis to decide
+ * the best peer/local_ni pair to use
+ */
+ list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_on_peer_list) {
+ if (!lnet_is_peer_net_healthy_locked(peer_net))
+ continue;
- if (!msg->msg_routing)
- msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
+ local_net = lnet_get_net_locked(peer_net->lpn_net_id);
+ if (!local_net) {
+ struct lnet_peer_ni *net_gw;
+ /*
+ * go through each peer_ni on that peer_net and
+ * determine the best possible gw to go through
+ */
+ list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
+ lpni_on_peer_net_list) {
+ net_gw = lnet_find_route_locked(NULL,
+ lpni->lpni_nid,
+ rtr_nid);
+
+ /*
+ * if no route is found for that network then
+ * move onto the next peer_ni in the peer
+ */
+ if (!net_gw)
+ continue;
+
+ if (!best_gw) {
+ best_gw = net_gw;
+ best_lpni = lpni;
+ } else {
+ rc = lnet_compare_peers(net_gw,
+ best_gw);
+ if (rc > 0) {
+ best_gw = net_gw;
+ best_lpni = lpni;
+ }
+ }
+ }
- if (src_ni == the_lnet.ln_loni) {
- /* No send credit hassles with LOLND */
- lnet_net_unlock(cpt);
- lnet_ni_send(src_ni, msg);
- return 0;
- }
+ if (!best_gw)
+ continue;
- rc = lnet_nid2peerni_locked(&lp, dst_nid, cpt);
- if (rc != 0) {
- lnet_net_unlock(cpt);
- LCONSOLE_WARN("Error %d finding peer %s\n", rc,
- libcfs_nid2str(dst_nid));
- /* ENOMEM or shutting down */
- return rc;
+ local_net = lnet_get_net_locked
+ (LNET_NIDNET(best_gw->lpni_nid));
+ routing = true;
+ } else {
+ routing = false;
+ best_gw = NULL;
}
- LASSERT (lp->lpni_net == src_ni->ni_net);
- } else {
- /* sending to a remote network */
- lp = lnet_find_route_locked(src_ni != NULL ?
- src_ni->ni_net : NULL,
- dst_nid, rtr_nid);
- if (lp == NULL) {
- lnet_net_unlock(cpt);
- LCONSOLE_WARN("No route to %s via %s "
- "(all routers down)\n",
- libcfs_id2str(msg->msg_target),
- libcfs_nid2str(src_nid));
- return -EHOSTUNREACH;
- }
+ /* no routable net found go on to a different net */
+ if (!local_net)
+ continue;
+
+ /*
+ * Iterate through the NIs in this local Net and select
+ * the NI to send from. The selection is determined by
+ * these 3 criterion in the following priority:
+ * 1. NUMA
+ * 2. NI available credits
+ * 3. Round Robin
+ */
+ shortest_distance = INT_MAX;
+ best_credits = INT_MIN;
+ ni = NULL;
+ while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
+ int ni_credits;
+ int distance;
+
+ if (!lnet_is_ni_healthy_locked(ni))
+ continue;
+
+ ni_credits = atomic_read(&ni->ni_tx_credits);
+
+ /*
+ * calculate the distance from the CPT on which
+ * the message memory is allocated to the CPT of
+ * the NI's physical device
+ */
+ distance = cfs_cpt_distance(lnet_cpt_table(),
+ md_cpt,
+ ni->dev_cpt);
- /* rtr_nid is LNET_NID_ANY or NID of pre-determined router,
- * it's possible that rtr_nid isn't LNET_NID_ANY and lp isn't
- * pre-determined router, this can happen if router table
- * was changed when we release the lock */
- if (rtr_nid != lp->lpni_nid) {
- cpt2 = lp->lpni_cpt;
- if (cpt2 != cpt) {
- lnet_net_unlock(cpt);
-
- rtr_nid = lp->lpni_nid;
- cpt = cpt2;
- goto again;
+ /*
+ * All distances smaller than the NUMA range
+ * are treated equally.
+ */
+ if (distance < lnet_get_numa_range())
+ distance = lnet_get_numa_range();
+
+ /*
+ * Select on shorter distance, then available
+ * credits, then round-robin.
+ */
+ if (distance > shortest_distance) {
+ continue;
+ } else if (distance < shortest_distance) {
+ shortest_distance = distance;
+ } else if (ni_credits < best_credits) {
+ continue;
+ } else if (ni_credits == best_credits) {
+ if (best_ni && best_ni->ni_seq <= ni->ni_seq)
+ continue;
}
+ best_ni = ni;
+ best_credits = ni_credits;
}
+ }
- CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
- libcfs_nid2str(dst_nid), libcfs_nid2str(lp->lpni_nid),
- lnet_msgtyp2str(msg->msg_type), msg->msg_len);
+ /*
+ * if the peer is not MR capable, then we should always send to it
+ * using the first NI in the NET we determined.
+ */
+ if (!peer->lp_multi_rail && local_net != NULL)
+ best_ni = lnet_net2ni_locked(local_net->net_id, cpt);
+
+ if (!best_ni) {
+ lnet_net_unlock(cpt);
+ LCONSOLE_WARN("No local ni found to send from to %s\n",
+ libcfs_nid2str(dst_nid));
+ return -EINVAL;
+ }
+
+ /*
+ * Now that we selected the NI to use increment its sequence
+ * number so the Round Robin algorithm will detect that it has
+ * been used and pick the next NI.
+ */
+ best_ni->ni_seq++;
+
+ if (routing)
+ goto send;
+
+pick_peer:
+ if (best_ni == the_lnet.ln_loni) {
+ /* No send credit hassles with LOLND */
+ lnet_ni_addref_locked(best_ni, cpt);
+ msg->msg_hdr.dest_nid = cpu_to_le64(best_ni->ni_nid);
+ if (!msg->msg_routing)
+ msg->msg_hdr.src_nid = cpu_to_le64(best_ni->ni_nid);
+ msg->msg_target.nid = best_ni->ni_nid;
+ lnet_msg_commit(msg, cpt);
+ msg->msg_txni = best_ni;
+ lnet_net_unlock(cpt);
+
+ return LNET_CREDIT_OK;
+ }
- if (src_ni == NULL) {
- src_ni = lnet_get_next_ni_locked(lp->lpni_net, NULL);
- LASSERT(src_ni != NULL);
- src_nid = src_ni->ni_nid;
+ if (msg->msg_type == LNET_MSG_REPLY ||
+ msg->msg_type == LNET_MSG_ACK) {
+ /*
+ * for replies we want to respond on the same peer_ni we
+ * received the message on if possible. If not, then pick
+ * a peer_ni to send to
+ */
+ best_lpni = lnet_find_peer_ni_locked(dst_nid);
+ if (best_lpni) {
+ lnet_peer_ni_decref_locked(best_lpni);
+ goto send;
} else {
- LASSERT (src_ni->ni_net == lp->lpni_net);
+ CDEBUG(D_NET, "unable to send msg_type %d to "
+ "originating %s\n", msg->msg_type,
+ libcfs_nid2str(dst_nid));
}
+ }
- lnet_peer_ni_addref_locked(lp);
+ peer_net = lnet_peer_get_net_locked(peer,
+ best_ni->ni_net->net_id);
+ /*
+ * peer_net is not available or the src_nid is explicitly defined
+ * and the peer_net for that src_nid is unhealthy. find a route to
+ * the destination nid.
+ */
+ if (!peer_net ||
+ (src_nid != LNET_NID_ANY &&
+ !lnet_is_peer_net_healthy_locked(peer_net))) {
+ best_gw = lnet_find_route_locked(best_ni->ni_net,
+ dst_nid,
+ rtr_nid);
+ /*
+ * if no route is found for that network then
+ * move onto the next peer_ni in the peer
+ */
+ if (!best_gw) {
+ lnet_net_unlock(cpt);
+ LCONSOLE_WARN("No route to peer from %s\n",
+ libcfs_nid2str(best_ni->ni_nid));
+ return -EHOSTUNREACH;
+ }
- LASSERT(src_nid != LNET_NID_ANY);
- lnet_msg_commit(msg, cpt);
+ CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
+ libcfs_nid2str(dst_nid),
+ libcfs_nid2str(best_gw->lpni_nid),
+ lnet_msgtyp2str(msg->msg_type), msg->msg_len);
+
+ best_lpni = lnet_find_peer_ni_locked(dst_nid);
+ LASSERT(best_lpni != NULL);
+ lnet_peer_ni_decref_locked(best_lpni);
- if (!msg->msg_routing) {
- /* I'm the source and now I know which NI to send on */
- msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
+ routing = true;
+
+ goto send;
+ } else if (!lnet_is_peer_net_healthy_locked(peer_net)) {
+ /*
+ * this peer_net is unhealthy but we still have an opportunity
+ * to find another peer_net that we can use
+ */
+ __u32 net_id = peer_net->lpn_net_id;
+ LCONSOLE_WARN("peer net %s unhealthy\n",
+ libcfs_net2str(net_id));
+ goto again;
+ }
+
+ /*
+ * Look at the peer NIs for the destination peer that connect
+ * to the chosen net. If a peer_ni is preferred when using the
+ * best_ni to communicate, we use that one. If there is no
+ * preferred peer_ni, or there are multiple preferred peer_ni,
+ * the available transmit credits are used. If the transmit
+ * credits are equal, we round-robin over the peer_ni.
+ */
+ lpni = NULL;
+ best_lpni_credits = INT_MIN;
+ preferred = false;
+ while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
+ /*
+ * if this peer ni is not healthy just skip it, no point in
+ * examining it further
+ */
+ if (!lnet_is_peer_ni_healthy_locked(lpni))
+ continue;
+ ni_is_pref = lnet_peer_is_ni_pref_locked(lpni, best_ni);
+
+ /* if this is a preferred peer use it */
+ if (!preferred && ni_is_pref) {
+ preferred = true;
+ } else if (preferred && !ni_is_pref) {
+ /*
+ * this is not the preferred peer so let's ignore
+ * it.
+ */
+ continue;
+ } else if (lpni->lpni_txcredits < best_lpni_credits) {
+ /*
+ * We already have a peer that has more credits
+ * available than this one. No need to consider
+ * this peer further.
+ */
+ continue;
+ } else if (lpni->lpni_txcredits == best_lpni_credits) {
+ /*
+ * The best peer found so far and the current peer
+ * have the same number of available credits let's
+ * make sure to select between them using Round
+ * Robin
+ */
+ if (best_lpni) {
+ if (best_lpni->lpni_seq <= lpni->lpni_seq)
+ continue;
+ }
}
- msg->msg_target_is_router = 1;
- msg->msg_target.nid = lp->lpni_nid;
- msg->msg_target.pid = LNET_PID_LUSTRE;
+ best_lpni = lpni;
+ best_lpni_credits = lpni->lpni_txcredits;
}
- /* 'lp' is our best choice of peer */
+ /* if we still can't find a peer ni then we can't reach it */
+ if (!best_lpni) {
+ __u32 net_id = peer_net->lpn_net_id;
+ lnet_net_unlock(cpt);
+ LCONSOLE_WARN("no peer_ni found on peer net %s\n",
+ libcfs_net2str(net_id));
+ return -EHOSTUNREACH;
+ }
+
+send:
+ /*
+ * Increment sequence number of the peer selected so that we
+ * pick the next one in Round Robin.
+ */
+ best_lpni->lpni_seq++;
+
+ /*
+ * When routing the best gateway found acts as the best peer
+ * NI to send to.
+ */
+ if (routing)
+ best_lpni = best_gw;
+
+ /*
+ * grab a reference on the peer_ni so it sticks around even if
+ * we need to drop and relock the lnet_net_lock below.
+ */
+ lnet_peer_ni_addref_locked(best_lpni);
+
+ /*
+ * Use lnet_cpt_of_nid() to determine the CPT used to commit the
+ * message. This ensures that we get a CPT that is correct for
+ * the NI when the NI has been restricted to a subset of all CPTs.
+ * If the selected CPT differs from the one currently locked, we
+ * must unlock and relock the lnet_net_lock(), and then check whether
+ * the configuration has changed. We don't have a hold on the best_ni
+ * yet, and it may have vanished.
+ */
+ cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
+ if (cpt != cpt2) {
+ lnet_net_unlock(cpt);
+ cpt = cpt2;
+ lnet_net_lock(cpt);
+ if (seq != lnet_get_dlc_seq_locked()) {
+ lnet_peer_ni_decref_locked(best_lpni);
+ goto again;
+ }
+ }
- LASSERT(!msg->msg_peertxcredit);
- LASSERT(!msg->msg_txcredit);
- LASSERT(msg->msg_txpeer == NULL);
+ /*
+ * store the best_lpni in the message right away to avoid having
+ * to do the same operation under different conditions
+ */
+ msg->msg_txpeer = best_lpni;
+ msg->msg_txni = best_ni;
- msg->msg_txpeer = lp; /* msg takes my ref on lp */
- /* set the NI for this message */
- msg->msg_txni = src_ni;
+ /*
+ * grab a reference for the best_ni since now it's in use in this
+ * send. the reference will need to be dropped when the message is
+ * finished in lnet_finalize()
+ */
lnet_ni_addref_locked(msg->msg_txni, cpt);
+ /*
+ * set the destination nid in the message here because it's
+ * possible that we'd be sending to a different nid than the one
+ * originaly given.
+ */
+ msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
+
+ /*
+ * Always set the target.nid to the best peer picked. Either the
+ * nid will be one of the preconfigured NIDs, or the same NID as
+ * what was originaly set in the target or it will be the NID of
+ * a router if this message should be routed
+ */
+ msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
+
+ /*
+ * lnet_msg_commit assigns the correct cpt to the message, which
+ * is used to decrement the correct refcount on the ni when it's
+ * time to return the credits
+ */
+ lnet_msg_commit(msg, cpt);
+
+ /*
+ * If we are routing the message then we don't need to overwrite
+ * the src_nid since it would've been set at the origin. Otherwise
+ * we are the originator so we need to set it.
+ */
+ if (!msg->msg_routing)
+ msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
+
+ if (routing) {
+ msg->msg_target_is_router = 1;
+ msg->msg_target.pid = LNET_PID_LUSTRE;
+ }
+
rc = lnet_post_send_locked(msg, 0);
+
lnet_net_unlock(cpt);
+ return rc;
+}
+
+int
+lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
+{
+ lnet_nid_t dst_nid = msg->msg_target.nid;
+ int rc;
+
+ /*
+ * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
+ * but we might want to use pre-determined router for ACK/REPLY
+ * in the future
+ */
+ /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
+ LASSERT (msg->msg_txpeer == NULL);
+ LASSERT (!msg->msg_sending);
+ LASSERT (!msg->msg_target_is_router);
+ LASSERT (!msg->msg_receiving);
+
+ msg->msg_sending = 1;
+
+ LASSERT(!msg->msg_tx_committed);
+
+ rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
if (rc < 0)
return rc;
if (rc == LNET_CREDIT_OK)
- lnet_ni_send(src_ni, msg);
+ lnet_ni_send(msg->msg_txni, msg);
- return 0; /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT */
+ /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT */
+ return 0;
}
void
hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
- info.mi_id.nid = hdr->src_nid;
+ /* Primary peer NID. */
+ info.mi_id.nid = msg->msg_initiator;
info.mi_id.pid = hdr->src_pid;
info.mi_opc = LNET_MD_OP_PUT;
info.mi_portal = hdr->msg.put.ptl_index;
info.mi_rlength = hdr->payload_length;
info.mi_roffset = hdr->msg.put.offset;
info.mi_mbits = hdr->msg.put.match_bits;
- info.mi_cpt = msg->msg_rxpeer->lpni_cpt;
+ info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
ready_delay = msg->msg_rx_ready_delay;
{
struct lnet_match_info info;
lnet_hdr_t *hdr = &msg->msg_hdr;
+ lnet_process_id_t source_id;
struct lnet_handle_wire reply_wmd;
int rc;
hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
- info.mi_id.nid = hdr->src_nid;
+ source_id.nid = hdr->src_nid;
+ source_id.pid = hdr->src_pid;
+ /* Primary peer NID */
+ info.mi_id.nid = msg->msg_initiator;
info.mi_id.pid = hdr->src_pid;
info.mi_opc = LNET_MD_OP_GET;
info.mi_portal = hdr->msg.get.ptl_index;
reply_wmd = hdr->msg.get.return_wmd;
- lnet_prep_send(msg, LNET_MSG_REPLY, info.mi_id,
+ lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
msg->msg_offset, msg->msg_wanted);
msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
lnet_pid_t dest_pid;
lnet_nid_t dest_nid;
lnet_nid_t src_nid;
- __u32 payload_length;
- __u32 type;
+ struct lnet_peer_ni *lpni;
+ __u32 payload_length;
+ __u32 type;
LASSERT (!in_interrupt ());
msg->msg_hdr.dest_pid = dest_pid;
msg->msg_hdr.payload_length = payload_length;
}
+ /* Multi-Rail: Primary NID of source. */
+ msg->msg_initiator = lnet_peer_primary_nid(src_nid);
lnet_net_lock(cpt);
- rc = lnet_nid2peerni_locked(&msg->msg_rxpeer, from_nid, cpt);
- if (rc != 0) {
+ lpni = lnet_nid2peerni_locked(from_nid, cpt);
+ if (IS_ERR(lpni)) {
lnet_net_unlock(cpt);
CERROR("%s, src %s: Dropping %s "
- "(error %d looking up sender)\n",
+ "(error %ld looking up sender)\n",
libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
- lnet_msgtyp2str(type), rc);
+ lnet_msgtyp2str(type), PTR_ERR(lpni));
lnet_msg_free(msg);
if (rc == -ESHUTDOWN)
/* We are shutting down. Don't do anything more */
return 0;
goto drop;
}
+ msg->msg_rxpeer = lpni;
msg->msg_rxni = ni;
lnet_ni_addref_locked(ni, cpt);
* called lnet_drop_message(), so I just hang onto msg as well
* until that's done */
- lnet_drop_message(msg->msg_rxni,
- msg->msg_rxpeer->lpni_cpt,
+ lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
msg->msg_private, msg->msg_len);
/*
* NB: message will not generate event because w/o attached MD,
libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
/* setup information for lnet_build_msg_event */
+ msg->msg_initiator = lnet_peer_primary_nid(peer_id.nid);
+ /* Cheaper: msg->msg_initiator = getmsg->msg_txpeer->lp_nid; */
msg->msg_from = peer_id.nid;
msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
msg->msg_hdr.src_nid = peer_id.nid;