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));
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;
!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);
int best_credits = 0;
__u32 seq, seq2;
int best_lpni_credits = INT_MIN;
+ int md_cpt = 0;
+ int shortest_distance = INT_MAX;
+ int distance = 0;
+ bool found_ir = false;
again:
/*
routing = false;
local_net = NULL;
best_ni = NULL;
+ shortest_distance = INT_MAX;
+ found_ir = false;
if (the_lnet.ln_shutdown) {
lnet_net_unlock(cpt);
return -ESHUTDOWN;
}
+ 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;
+
/*
* initialize the variables which could be reused if we go to
* again
lpni = NULL;
seq = lnet_get_dlc_seq_locked();
- rc = lnet_find_or_create_peer_locked(dst_nid, cpt, &peer);
- if (rc != 0) {
+ peer = lnet_find_or_create_peer_locked(dst_nid, cpt);
+ if (IS_ERR(peer)) {
lnet_net_unlock(cpt);
- return rc;
+ return PTR_ERR(peer);
}
/* If peer is not healthy then can not send anything to it */
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
}
}
- if (best_ni == the_lnet.ln_loni) {
- /* No send credit hassles with LOLND */
- 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);
-
- lnet_ni_addref_locked(best_ni, cpt);
- lnet_net_unlock(cpt);
- msg->msg_txni = best_ni;
- lnet_ni_send(best_ni, msg);
-
- *lo_sent = true;
- return 0;
- }
-
if (best_ni)
goto pick_peer;
continue;
/*
- * Second jab at determining best_ni
- * if we get here then the peer we're trying to send
- * to is on a directly connected network, and we'll
- * need to pick the local_ni on that network to send
- * from
+ * 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
*/
while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
+ int ni_credits;
+
if (!lnet_is_ni_healthy_locked(ni))
continue;
- /* TODO: compare NUMA distance */
- if (ni->ni_tx_queues[cpt]->tq_credits <=
- best_credits) {
+
+ 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);
+
+ /*
+ * If we already have a closer NI within the NUMA
+ * range provided, then there is no need to
+ * consider the current NI. Move on to the next
+ * one.
+ */
+ if (distance > shortest_distance &&
+ distance > lnet_get_numa_range())
+ continue;
+
+ if (distance < shortest_distance &&
+ distance > lnet_get_numa_range()) {
/*
- * all we want is to read tq_credits
- * value as an approximation of how
- * busy the NI is. No need to grab a lock
+ * The current NI is the closest one that we
+ * have found, even though it's not in the
+ * NUMA range specified. This occurs if
+ * the NUMA range is less than the least
+ * of the distances in the system.
+ * In effect NUMA range consideration is
+ * turned off.
*/
- continue;
- } else if (best_ni) {
- if ((best_ni)->ni_seq - ni->ni_seq <= 0)
+ shortest_distance = distance;
+ } else if ((distance <= shortest_distance &&
+ distance < lnet_get_numa_range()) ||
+ distance == shortest_distance) {
+ /*
+ * This NI is either within range or it's
+ * equidistant. In both of these cases we
+ * would want to select the NI based on
+ * its available credits first, and then
+ * via Round Robin.
+ */
+ if (distance <= shortest_distance &&
+ distance < lnet_get_numa_range()) {
+ /*
+ * If this is the first NI that's
+ * within range, then set the
+ * shortest distance to the range
+ * specified by the user. In
+ * effect we're saying that all
+ * NIs that fall within this NUMA
+ * range shall be dealt with as
+ * having equal NUMA weight. Which
+ * will mean that we should select
+ * through that set by their
+ * available credits first
+ * followed by Round Robin.
+ *
+ * And since this is the first NI
+ * in the range, let's just set it
+ * as our best_ni for now. The
+ * following NIs found in the
+ * range will be dealt with as
+ * mentioned previously.
+ */
+ shortest_distance = lnet_get_numa_range();
+ if (!found_ir) {
+ found_ir = true;
+ goto set_ni;
+ }
+ }
+ /*
+ * This NI is NUMA equidistant let's
+ * select using credits followed by Round
+ * Robin.
+ */
+ if (ni_credits < best_credits) {
continue;
- (best_ni)->ni_seq = ni->ni_seq + 1;
+ } else if (ni_credits == best_credits) {
+ if (best_ni) {
+ if (best_ni->ni_seq <= ni->ni_seq)
+ continue;
+ }
+ }
}
-
+set_ni:
best_ni = ni;
- best_credits = ni->ni_tx_queues[cpt]->tq_credits;
+ best_credits = ni_credits;
}
}
+ /*
+ * 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);
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);
+
+ lnet_net_unlock(cpt);
+ msg->msg_txni = best_ni;
+ lnet_ni_send(best_ni, msg);
+
+ *lo_sent = true;
+ return 0;
+ }
+
lpni = NULL;
if (msg->msg_type == LNET_MSG_REPLY ||
}
CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
- libcfs_nid2str(lpni->lpni_nid),
+ libcfs_nid2str(dst_nid),
libcfs_nid2str(best_gw->lpni_nid),
lnet_msgtyp2str(msg->msg_type), msg->msg_len);
best_lpni = NULL;
while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
/*
- * if this peer ni is not healty just skip it, no point in
+ * 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;
- } if (lpni->lpni_txcredits <= best_lpni_credits)
+ } 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 (best_lpni) {
- if (best_lpni->lpni_seq - lpni->lpni_seq <= 0)
- continue;
- best_lpni->lpni_seq = lpni->lpni_seq + 1;
+ 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;
+ }
}
best_lpni = lpni;
best_lpni_credits = lpni->lpni_txcredits;
}
+ /*
+ * Increment sequence number of the peer selected so that we can
+ * pick the next one in Round Robin.
+ */
+ best_lpni->lpni_seq++;
+
/* 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));
- goto again;
+ return -EHOSTUNREACH;
}
send:
/*
- * determine the cpt to use and if it has changed then
- * lock the new cpt and check if the config has changed.
- * If it has changed then repeat the algorithm since the
- * ni or peer list could have changed and the algorithm
- * would endup picking a different ni/peer_ni pair.
+ * 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
+ * or best_peer_ni yet, and they may have vanished.
*/
- cpt2 = best_lpni->lpni_cpt;
+ cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
if (cpt != cpt2) {
lnet_net_unlock(cpt);
cpt = cpt2;
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;