#include <lnet/lib-lnet.h>
#include <uapi/linux/lnet/lnet-dlc.h>
+/* Value indicating that recovery needs to re-check a peer immediately. */
+#define LNET_REDISCOVER_PEER (1)
+
+static int lnet_peer_queue_for_discovery(struct lnet_peer *lp);
+
static void
lnet_peer_remove_from_remote_list(struct lnet_peer_ni *lpni)
{
INIT_LIST_HEAD(&lp->lp_peer_list);
INIT_LIST_HEAD(&lp->lp_peer_nets);
INIT_LIST_HEAD(&lp->lp_dc_list);
+ INIT_LIST_HEAD(&lp->lp_dc_pendq);
init_waitqueue_head(&lp->lp_dc_waitq);
spin_lock_init(&lp->lp_lock);
lp->lp_primary_nid = nid;
LASSERT(atomic_read(&lp->lp_refcount) == 0);
LASSERT(list_empty(&lp->lp_peer_nets));
LASSERT(list_empty(&lp->lp_peer_list));
+ LASSERT(list_empty(&lp->lp_dc_list));
+
+ if (lp->lp_data)
+ lnet_ping_buffer_decref(lp->lp_data);
LIBCFS_FREE(lp, sizeof(*lp));
}
/*
* If there are no more peer nets, make the peer unfindable
* via the peer_tables.
+ *
+ * Otherwise, if the peer is DISCOVERED, tell discovery to
+ * take another look at it. This is a no-op if discovery for
+ * this peer did the detaching.
*/
if (list_empty(&lp->lp_peer_nets)) {
list_del_init(&lp->lp_peer_list);
ptable->pt_peers--;
+ } else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
+ /* Discovery isn't running, nothing to do here. */
+ } else if (lp->lp_state & LNET_PEER_DISCOVERED) {
+ lnet_peer_queue_for_discovery(lp);
+ wake_up(&the_lnet.ln_dc_waitq);
}
CDEBUG(D_NET, "peer %s NID %s\n",
libcfs_nid2str(lp->lp_primary_nid),
return lpni;
}
+struct lnet_peer *
+lnet_find_peer(lnet_nid_t nid)
+{
+ struct lnet_peer_ni *lpni;
+ struct lnet_peer *lp = NULL;
+ int cpt;
+
+ cpt = lnet_net_lock_current();
+ lpni = lnet_find_peer_ni_locked(nid);
+ if (lpni) {
+ lp = lpni->lpni_peer_net->lpn_peer;
+ lnet_peer_addref_locked(lp);
+ lnet_peer_ni_decref_locked(lpni);
+ }
+ lnet_net_unlock(cpt);
+
+ return lp;
+}
+
struct lnet_peer_ni *
lnet_get_peer_ni_idx_locked(int idx, struct lnet_peer_net **lpn,
struct lnet_peer **lp)
}
/*
+ * Start pushes to peers that need to be updated for a configuration
+ * change on this node.
+ */
+void
+lnet_push_update_to_peers(int force)
+{
+ struct lnet_peer_table *ptable;
+ struct lnet_peer *lp;
+ int lncpt;
+ int cpt;
+
+ lnet_net_lock(LNET_LOCK_EX);
+ lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
+ for (cpt = 0; cpt < lncpt; cpt++) {
+ ptable = the_lnet.ln_peer_tables[cpt];
+ list_for_each_entry(lp, &ptable->pt_peer_list, lp_peer_list) {
+ if (force) {
+ spin_lock(&lp->lp_lock);
+ if (lp->lp_state & LNET_PEER_MULTI_RAIL)
+ lp->lp_state |= LNET_PEER_FORCE_PUSH;
+ spin_unlock(&lp->lp_lock);
+ }
+ if (lnet_peer_needs_push(lp))
+ lnet_peer_queue_for_discovery(lp);
+ }
+ }
+ lnet_net_unlock(LNET_LOCK_EX);
+ wake_up(&the_lnet.ln_dc_waitq);
+}
+
+/*
* Test whether a ni is a preferred ni for this peer_ni, e.g, whether
* this is a preferred point-to-point path. Call with lnet_net_lock in
* shared mmode.
lnet_nid_t
LNetPrimaryNID(lnet_nid_t nid)
{
+ struct lnet_peer *lp;
struct lnet_peer_ni *lpni;
lnet_nid_t primary_nid = nid;
int rc = 0;
rc = PTR_ERR(lpni);
goto out_unlock;
}
- primary_nid = lpni->lpni_peer_net->lpn_peer->lp_primary_nid;
+ lp = lpni->lpni_peer_net->lpn_peer;
+ while (!lnet_peer_is_uptodate(lp)) {
+ rc = lnet_discover_peer_locked(lpni, cpt, true);
+ if (rc)
+ goto out_decref;
+ lp = lpni->lpni_peer_net->lpn_peer;
+ }
+ primary_nid = lp->lp_primary_nid;
+out_decref:
lnet_peer_ni_decref_locked(lpni);
out_unlock:
lnet_net_unlock(cpt);
}
/*
+ * Update the primary NID of a peer, if possible.
+ *
+ * Call with the lnet_api_mutex held.
+ */
+static int
+lnet_peer_set_primary_nid(struct lnet_peer *lp, lnet_nid_t nid, unsigned flags)
+{
+ lnet_nid_t old = lp->lp_primary_nid;
+ int rc = 0;
+
+ if (lp->lp_primary_nid == nid)
+ goto out;
+ rc = lnet_peer_add_nid(lp, nid, flags);
+ if (rc)
+ goto out;
+ lp->lp_primary_nid = nid;
+out:
+ CDEBUG(D_NET, "peer %s NID %s: %d\n",
+ libcfs_nid2str(old), libcfs_nid2str(nid), rc);
+ return rc;
+}
+
+/*
* lpni creation initiated due to traffic either sending or receiving.
*/
static int
LNET_PEER_FORCE_PING |
LNET_PEER_FORCE_PUSH)) {
rc = false;
+ } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
+ rc = true;
} else if (lp->lp_state & LNET_PEER_REDISCOVER) {
if (lnet_peer_discovery_disabled)
rc = true;
else
rc = false;
+ } else if (lnet_peer_needs_push(lp)) {
+ rc = false;
} else if (lp->lp_state & LNET_PEER_DISCOVERED) {
if (lp->lp_state & LNET_PEER_NIDS_UPTODATE)
rc = true;
rc = -EALREADY;
}
+ CDEBUG(D_NET, "Queue peer %s: %d\n",
+ libcfs_nid2str(lp->lp_primary_nid), rc);
+
return rc;
}
*/
static void lnet_peer_discovery_complete(struct lnet_peer *lp)
{
+ struct lnet_msg *msg = NULL;
+ int rc = 0;
+ struct list_head pending_msgs;
+
+ INIT_LIST_HEAD(&pending_msgs);
+
+ CDEBUG(D_NET, "Discovery complete. Dequeue peer %s\n",
+ libcfs_nid2str(lp->lp_primary_nid));
+
list_del_init(&lp->lp_dc_list);
+ list_splice_init(&lp->lp_dc_pendq, &pending_msgs);
wake_up_all(&lp->lp_dc_waitq);
+
+ lnet_net_unlock(LNET_LOCK_EX);
+
+ /* iterate through all pending messages and send them again */
+ list_for_each_entry(msg, &pending_msgs, msg_list) {
+ if (lp->lp_dc_error) {
+ lnet_finalize(msg, lp->lp_dc_error);
+ continue;
+ }
+
+ CDEBUG(D_NET, "sending pending message %s to target %s\n",
+ lnet_msgtyp2str(msg->msg_type),
+ libcfs_id2str(msg->msg_target));
+ rc = lnet_send(msg->msg_src_nid_param, msg,
+ msg->msg_rtr_nid_param);
+ if (rc < 0) {
+ CNETERR("Error sending %s to %s: %d\n",
+ lnet_msgtyp2str(msg->msg_type),
+ libcfs_id2str(msg->msg_target), rc);
+ lnet_finalize(msg, rc);
+ }
+ }
+ lnet_net_lock(LNET_LOCK_EX);
+ lnet_peer_decref_locked(lp);
+}
+
+/*
+ * Handle inbound push.
+ * Like any event handler, called with lnet_res_lock/CPT held.
+ */
+void lnet_peer_push_event(struct lnet_event *ev)
+{
+ struct lnet_ping_buffer *pbuf = ev->md.user_ptr;
+ struct lnet_peer *lp;
+
+ /* lnet_find_peer() adds a refcount */
+ lp = lnet_find_peer(ev->source.nid);
+ if (!lp) {
+ CERROR("Push Put from unknown %s (source %s)\n",
+ libcfs_nid2str(ev->initiator.nid),
+ libcfs_nid2str(ev->source.nid));
+ return;
+ }
+
+ /* Ensure peer state remains consistent while we modify it. */
+ spin_lock(&lp->lp_lock);
+
+ /*
+ * If some kind of error happened the contents of the message
+ * cannot be used. Clear the NIDS_UPTODATE and set the
+ * FORCE_PING flag to trigger a ping.
+ */
+ if (ev->status) {
+ lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
+ lp->lp_state |= LNET_PEER_FORCE_PING;
+ CDEBUG(D_NET, "Push Put error %d from %s (source %s)\n",
+ ev->status,
+ libcfs_nid2str(lp->lp_primary_nid),
+ libcfs_nid2str(ev->source.nid));
+ goto out;
+ }
+
+ /*
+ * A push with invalid or corrupted info. Clear the UPTODATE
+ * flag to trigger a ping.
+ */
+ if (lnet_ping_info_validate(&pbuf->pb_info)) {
+ lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
+ lp->lp_state |= LNET_PEER_FORCE_PING;
+ CDEBUG(D_NET, "Corrupted Push from %s\n",
+ libcfs_nid2str(lp->lp_primary_nid));
+ goto out;
+ }
+
+ /*
+ * Make sure we'll allocate the correct size ping buffer when
+ * pinging the peer.
+ */
+ if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
+ lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
+
+ /*
+ * A non-Multi-Rail peer is not supposed to be capable of
+ * sending a push.
+ */
+ if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL)) {
+ CERROR("Push from non-Multi-Rail peer %s dropped\n",
+ libcfs_nid2str(lp->lp_primary_nid));
+ goto out;
+ }
+
+ /*
+ * Check the MULTIRAIL flag. Complain if the peer was DLC
+ * configured without it.
+ */
+ if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
+ if (lp->lp_state & LNET_PEER_CONFIGURED) {
+ CERROR("Push says %s is Multi-Rail, DLC says not\n",
+ libcfs_nid2str(lp->lp_primary_nid));
+ } else {
+ lp->lp_state |= LNET_PEER_MULTI_RAIL;
+ lnet_peer_clr_non_mr_pref_nids(lp);
+ }
+ }
+
+ /*
+ * The peer may have discovery disabled at its end. Set
+ * NO_DISCOVERY as appropriate.
+ */
+ if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
+ CDEBUG(D_NET, "Peer %s has discovery disabled\n",
+ libcfs_nid2str(lp->lp_primary_nid));
+ lp->lp_state |= LNET_PEER_NO_DISCOVERY;
+ } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
+ CDEBUG(D_NET, "Peer %s has discovery enabled\n",
+ libcfs_nid2str(lp->lp_primary_nid));
+ lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
+ }
+
+ /*
+ * Check for truncation of the Put message. Clear the
+ * NIDS_UPTODATE flag and set FORCE_PING to trigger a ping,
+ * and tell discovery to allocate a bigger buffer.
+ */
+ if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
+ if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
+ the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
+ lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
+ lp->lp_state |= LNET_PEER_FORCE_PING;
+ CDEBUG(D_NET, "Truncated Push from %s (%d nids)\n",
+ libcfs_nid2str(lp->lp_primary_nid),
+ pbuf->pb_info.pi_nnis);
+ goto out;
+ }
+
+ /*
+ * Check whether the Put data is stale. Stale data can just be
+ * dropped.
+ */
+ if (pbuf->pb_info.pi_nnis > 1 &&
+ lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid &&
+ LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno) {
+ CDEBUG(D_NET, "Stale Push from %s: got %u have %u\n",
+ libcfs_nid2str(lp->lp_primary_nid),
+ LNET_PING_BUFFER_SEQNO(pbuf),
+ lp->lp_peer_seqno);
+ goto out;
+ }
+
+ /*
+ * Check whether the Put data is new, in which case we clear
+ * the UPTODATE flag and prepare to process it.
+ *
+ * If the Put data is current, and the peer is UPTODATE then
+ * we assome everything is all right and drop the data as
+ * stale.
+ */
+ if (LNET_PING_BUFFER_SEQNO(pbuf) > lp->lp_peer_seqno) {
+ lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
+ lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
+ } else if (lp->lp_state & LNET_PEER_NIDS_UPTODATE) {
+ CDEBUG(D_NET, "Stale Push from %s: got %u have %u\n",
+ libcfs_nid2str(lp->lp_primary_nid),
+ LNET_PING_BUFFER_SEQNO(pbuf),
+ lp->lp_peer_seqno);
+ goto out;
+ }
+
+ /*
+ * If there is data present that hasn't been processed yet,
+ * we'll replace it if the Put contained newer data and it
+ * fits. We're racing with a Ping or earlier Push in this
+ * case.
+ */
+ if (lp->lp_state & LNET_PEER_DATA_PRESENT) {
+ if (LNET_PING_BUFFER_SEQNO(pbuf) >
+ LNET_PING_BUFFER_SEQNO(lp->lp_data) &&
+ pbuf->pb_info.pi_nnis <= lp->lp_data->pb_nnis) {
+ memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
+ LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
+ CDEBUG(D_NET, "Ping/Push race from %s: %u vs %u\n",
+ libcfs_nid2str(lp->lp_primary_nid),
+ LNET_PING_BUFFER_SEQNO(pbuf),
+ LNET_PING_BUFFER_SEQNO(lp->lp_data));
+ }
+ goto out;
+ }
+
+ /*
+ * Allocate a buffer to copy the data. On a failure we drop
+ * the Push and set FORCE_PING to force the discovery
+ * thread to fix the problem by pinging the peer.
+ */
+ lp->lp_data = lnet_ping_buffer_alloc(lp->lp_data_nnis, GFP_ATOMIC);
+ if (!lp->lp_data) {
+ lp->lp_state |= LNET_PEER_FORCE_PING;
+ CDEBUG(D_NET, "Cannot allocate Push buffer for %s %u\n",
+ libcfs_nid2str(lp->lp_primary_nid),
+ LNET_PING_BUFFER_SEQNO(pbuf));
+ goto out;
+ }
+
+ /* Success */
+ memcpy(&lp->lp_data->pb_info, &pbuf->pb_info,
+ LNET_PING_INFO_SIZE(pbuf->pb_info.pi_nnis));
+ lp->lp_state |= LNET_PEER_DATA_PRESENT;
+ CDEBUG(D_NET, "Received Push %s %u\n",
+ libcfs_nid2str(lp->lp_primary_nid),
+ LNET_PING_BUFFER_SEQNO(pbuf));
+
+out:
+ /*
+ * Queue the peer for discovery, and wake the discovery thread
+ * if the peer was already queued, because its status changed.
+ */
+ spin_unlock(&lp->lp_lock);
+ lnet_net_lock(LNET_LOCK_EX);
+ if (lnet_peer_queue_for_discovery(lp))
+ wake_up(&the_lnet.ln_dc_waitq);
+ /* Drop refcount from lookup */
lnet_peer_decref_locked(lp);
+ lnet_net_unlock(LNET_LOCK_EX);
+}
+
+/*
+ * Clear the discovery error state, unless we're already discovering
+ * this peer, in which case the error is current.
+ */
+static void lnet_peer_clear_discovery_error(struct lnet_peer *lp)
+{
+ spin_lock(&lp->lp_lock);
+ if (!(lp->lp_state & LNET_PEER_DISCOVERING))
+ lp->lp_dc_error = 0;
+ spin_unlock(&lp->lp_lock);
}
/*
* because discovery could tear down an lnet_peer.
*/
int
-lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt)
+lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
{
DEFINE_WAIT(wait);
struct lnet_peer *lp;
again:
lnet_net_unlock(cpt);
lnet_net_lock(LNET_LOCK_EX);
+ lp = lpni->lpni_peer_net->lpn_peer;
+ lnet_peer_clear_discovery_error(lp);
- /* We're willing to be interrupted. */
+ /*
+ * We're willing to be interrupted. The lpni can become a
+ * zombie if we race with DLC, so we must check for that.
+ */
for (;;) {
- lp = lpni->lpni_peer_net->lpn_peer;
prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
if (signal_pending(current))
break;
if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
break;
+ if (lp->lp_dc_error)
+ break;
if (lnet_peer_is_uptodate(lp))
break;
lnet_peer_queue_for_discovery(lp);
lnet_peer_addref_locked(lp);
+ /*
+ * if caller requested a non-blocking operation then
+ * return immediately. Once discovery is complete then the
+ * peer ref will be decremented and any pending messages
+ * that were stopped due to discovery will be transmitted.
+ */
+ if (!block)
+ break;
lnet_net_unlock(LNET_LOCK_EX);
schedule();
finish_wait(&lp->lp_dc_waitq, &wait);
lnet_net_lock(LNET_LOCK_EX);
lnet_peer_decref_locked(lp);
- /* Do not use lp beyond this point. */
+ /* Peer may have changed */
+ lp = lpni->lpni_peer_net->lpn_peer;
}
finish_wait(&lp->lp_dc_waitq, &wait);
rc = -EINTR;
else if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
rc = -ESHUTDOWN;
+ else if (lp->lp_dc_error)
+ rc = lp->lp_dc_error;
+ else if (!block)
+ CDEBUG(D_NET, "non-blocking discovery\n");
else if (!lnet_peer_is_uptodate(lp))
goto again;
+ CDEBUG(D_NET, "peer %s NID %s: %d. %s\n",
+ (lp ? libcfs_nid2str(lp->lp_primary_nid) : "(none)"),
+ libcfs_nid2str(lpni->lpni_nid), rc,
+ (!block) ? "pending discovery" : "discovery complete");
+
return rc;
}
-/*
- * Event handler for the discovery EQ.
- *
- * Called with lnet_res_lock(cpt) held. The cpt is the
- * lnet_cpt_of_cookie() of the md handle cookie.
- */
-static void lnet_discovery_event_handler(lnet_event_t *event)
+/* Handle an incoming ack for a push. */
+static void
+lnet_discovery_event_ack(struct lnet_peer *lp, struct lnet_event *ev)
{
- wake_up(&the_lnet.ln_dc_waitq);
+ struct lnet_ping_buffer *pbuf;
+
+ pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
+ spin_lock(&lp->lp_lock);
+ lp->lp_state &= ~LNET_PEER_PUSH_SENT;
+ lp->lp_push_error = ev->status;
+ if (ev->status)
+ lp->lp_state |= LNET_PEER_PUSH_FAILED;
+ else
+ lp->lp_node_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
+ spin_unlock(&lp->lp_lock);
+
+ CDEBUG(D_NET, "peer %s ev->status %d\n",
+ libcfs_nid2str(lp->lp_primary_nid), ev->status);
}
-/*
- * Wait for work to be queued or some other change that must be
- * attended to. Returns non-zero if the discovery thread should shut
- * down.
- */
-static int lnet_peer_discovery_wait_for_work(void)
+/* Handle a Reply message. This is the reply to a Ping message. */
+static void
+lnet_discovery_event_reply(struct lnet_peer *lp, struct lnet_event *ev)
{
- int cpt;
- int rc = 0;
+ struct lnet_ping_buffer *pbuf;
+ int rc;
- DEFINE_WAIT(wait);
+ spin_lock(&lp->lp_lock);
- cpt = lnet_net_lock_current();
- for (;;) {
- prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
- TASK_INTERRUPTIBLE);
- if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
- break;
- if (lnet_push_target_resize_needed())
- break;
- if (!list_empty(&the_lnet.ln_dc_request))
- break;
- lnet_net_unlock(cpt);
- schedule();
- finish_wait(&the_lnet.ln_dc_waitq, &wait);
- cpt = lnet_net_lock_current();
+ /*
+ * If some kind of error happened the contents of message
+ * cannot be used. Set PING_FAILED to trigger a retry.
+ */
+ if (ev->status) {
+ lp->lp_state |= LNET_PEER_PING_FAILED;
+ lp->lp_ping_error = ev->status;
+ CDEBUG(D_NET, "Ping Reply error %d from %s (source %s)\n",
+ ev->status,
+ libcfs_nid2str(lp->lp_primary_nid),
+ libcfs_nid2str(ev->source.nid));
+ goto out;
}
- finish_wait(&the_lnet.ln_dc_waitq, &wait);
-
- if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
- rc = -ESHUTDOWN;
- lnet_net_unlock(cpt);
+ pbuf = LNET_PING_INFO_TO_BUFFER(ev->md.start);
+ if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
+ lnet_swap_pinginfo(pbuf);
- CDEBUG(D_NET, "woken: %d\n", rc);
+ /*
+ * A reply with invalid or corrupted info. Set PING_FAILED to
+ * trigger a retry.
+ */
+ rc = lnet_ping_info_validate(&pbuf->pb_info);
+ if (rc) {
+ lp->lp_state |= LNET_PEER_PING_FAILED;
+ lp->lp_ping_error = 0;
+ CDEBUG(D_NET, "Corrupted Ping Reply from %s: %d\n",
+ libcfs_nid2str(lp->lp_primary_nid), rc);
+ goto out;
+ }
- return rc;
-}
+ /*
+ * Update the MULTI_RAIL flag based on the reply. If the peer
+ * was configured with DLC then the setting should match what
+ * DLC put in.
+ */
+ if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL) {
+ if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
+ /* Everything's fine */
+ } else if (lp->lp_state & LNET_PEER_CONFIGURED) {
+ CWARN("Reply says %s is Multi-Rail, DLC says not\n",
+ libcfs_nid2str(lp->lp_primary_nid));
+ } else {
+ lp->lp_state |= LNET_PEER_MULTI_RAIL;
+ lnet_peer_clr_non_mr_pref_nids(lp);
+ }
+ } else if (lp->lp_state & LNET_PEER_MULTI_RAIL) {
+ if (lp->lp_state & LNET_PEER_CONFIGURED) {
+ CWARN("DLC says %s is Multi-Rail, Reply says not\n",
+ libcfs_nid2str(lp->lp_primary_nid));
+ } else {
+ CERROR("Multi-Rail state vanished from %s\n",
+ libcfs_nid2str(lp->lp_primary_nid));
+ lp->lp_state &= ~LNET_PEER_MULTI_RAIL;
+ }
+ }
-/* The discovery thread. */
-static int lnet_peer_discovery(void *arg)
-{
- struct lnet_peer *lp;
+ /*
+ * Make sure we'll allocate the correct size ping buffer when
+ * pinging the peer.
+ */
+ if (lp->lp_data_nnis < pbuf->pb_info.pi_nnis)
+ lp->lp_data_nnis = pbuf->pb_info.pi_nnis;
- CDEBUG(D_NET, "started\n");
- cfs_block_allsigs();
+ /*
+ * The peer may have discovery disabled at its end. Set
+ * NO_DISCOVERY as appropriate.
+ */
+ if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_DISCOVERY)) {
+ CDEBUG(D_NET, "Peer %s has discovery disabled\n",
+ libcfs_nid2str(lp->lp_primary_nid));
+ lp->lp_state |= LNET_PEER_NO_DISCOVERY;
+ } else if (lp->lp_state & LNET_PEER_NO_DISCOVERY) {
+ CDEBUG(D_NET, "Peer %s has discovery enabled\n",
+ libcfs_nid2str(lp->lp_primary_nid));
+ lp->lp_state &= ~LNET_PEER_NO_DISCOVERY;
+ }
- for (;;) {
+ /*
+ * Check for truncation of the Reply. Clear PING_SENT and set
+ * PING_FAILED to trigger a retry.
+ */
+ if (pbuf->pb_nnis < pbuf->pb_info.pi_nnis) {
+ if (the_lnet.ln_push_target_nnis < pbuf->pb_info.pi_nnis)
+ the_lnet.ln_push_target_nnis = pbuf->pb_info.pi_nnis;
+ lp->lp_state |= LNET_PEER_PING_FAILED;
+ lp->lp_ping_error = 0;
+ CDEBUG(D_NET, "Truncated Reply from %s (%d nids)\n",
+ libcfs_nid2str(lp->lp_primary_nid),
+ pbuf->pb_info.pi_nnis);
+ goto out;
+ }
+
+ /*
+ * Check the sequence numbers in the reply. These are only
+ * available if the reply came from a Multi-Rail peer.
+ */
+ if (pbuf->pb_info.pi_features & LNET_PING_FEAT_MULTI_RAIL &&
+ pbuf->pb_info.pi_nnis > 1 &&
+ lp->lp_primary_nid == pbuf->pb_info.pi_ni[1].ns_nid) {
+ if (LNET_PING_BUFFER_SEQNO(pbuf) < lp->lp_peer_seqno) {
+ CDEBUG(D_NET, "Stale Reply from %s: got %u have %u\n",
+ libcfs_nid2str(lp->lp_primary_nid),
+ LNET_PING_BUFFER_SEQNO(pbuf),
+ lp->lp_peer_seqno);
+ goto out;
+ }
+
+ if (LNET_PING_BUFFER_SEQNO(pbuf) > lp->lp_peer_seqno)
+ lp->lp_peer_seqno = LNET_PING_BUFFER_SEQNO(pbuf);
+ }
+
+ /* We're happy with the state of the data in the buffer. */
+ CDEBUG(D_NET, "peer %s data present %u\n",
+ libcfs_nid2str(lp->lp_primary_nid), lp->lp_peer_seqno);
+ if (lp->lp_state & LNET_PEER_DATA_PRESENT)
+ lnet_ping_buffer_decref(lp->lp_data);
+ else
+ lp->lp_state |= LNET_PEER_DATA_PRESENT;
+ lnet_ping_buffer_addref(pbuf);
+ lp->lp_data = pbuf;
+out:
+ lp->lp_state &= ~LNET_PEER_PING_SENT;
+ spin_unlock(&lp->lp_lock);
+}
+
+/*
+ * Send event handling. Only matters for error cases, where we clean
+ * up state on the peer and peer_ni that would otherwise be updated in
+ * the REPLY event handler for a successful Ping, and the ACK event
+ * handler for a successful Push.
+ */
+static int
+lnet_discovery_event_send(struct lnet_peer *lp, struct lnet_event *ev)
+{
+ int rc = 0;
+
+ if (!ev->status)
+ goto out;
+
+ spin_lock(&lp->lp_lock);
+ if (ev->msg_type == LNET_MSG_GET) {
+ lp->lp_state &= ~LNET_PEER_PING_SENT;
+ lp->lp_state |= LNET_PEER_PING_FAILED;
+ lp->lp_ping_error = ev->status;
+ } else { /* ev->msg_type == LNET_MSG_PUT */
+ lp->lp_state &= ~LNET_PEER_PUSH_SENT;
+ lp->lp_state |= LNET_PEER_PUSH_FAILED;
+ lp->lp_push_error = ev->status;
+ }
+ spin_unlock(&lp->lp_lock);
+ rc = LNET_REDISCOVER_PEER;
+out:
+ CDEBUG(D_NET, "%s Send to %s: %d\n",
+ (ev->msg_type == LNET_MSG_GET ? "Ping" : "Push"),
+ libcfs_nid2str(ev->target.nid), rc);
+ return rc;
+}
+
+/*
+ * Unlink event handling. This event is only seen if a call to
+ * LNetMDUnlink() caused the event to be unlinked. If this call was
+ * made after the event was set up in LNetGet() or LNetPut() then we
+ * assume the Ping or Push timed out.
+ */
+static void
+lnet_discovery_event_unlink(struct lnet_peer *lp, struct lnet_event *ev)
+{
+ spin_lock(&lp->lp_lock);
+ /* We've passed through LNetGet() */
+ if (lp->lp_state & LNET_PEER_PING_SENT) {
+ lp->lp_state &= ~LNET_PEER_PING_SENT;
+ lp->lp_state |= LNET_PEER_PING_FAILED;
+ lp->lp_ping_error = -ETIMEDOUT;
+ CDEBUG(D_NET, "Ping Unlink for message to peer %s\n",
+ libcfs_nid2str(lp->lp_primary_nid));
+ }
+ /* We've passed through LNetPut() */
+ if (lp->lp_state & LNET_PEER_PUSH_SENT) {
+ lp->lp_state &= ~LNET_PEER_PUSH_SENT;
+ lp->lp_state |= LNET_PEER_PUSH_FAILED;
+ lp->lp_push_error = -ETIMEDOUT;
+ CDEBUG(D_NET, "Push Unlink for message to peer %s\n",
+ libcfs_nid2str(lp->lp_primary_nid));
+ }
+ spin_unlock(&lp->lp_lock);
+}
+
+/*
+ * Event handler for the discovery EQ.
+ *
+ * Called with lnet_res_lock(cpt) held. The cpt is the
+ * lnet_cpt_of_cookie() of the md handle cookie.
+ */
+static void lnet_discovery_event_handler(lnet_event_t *event)
+{
+ 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;
+
+ CDEBUG(D_NET, "Received event: %d\n", event->type);
+
+ 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:
+ /* LNetMDUnlink() was called */
+ lnet_discovery_event_unlink(lp, event);
+ 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_FORCE_PING;
+ 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);
+
+ LNetInvalidateMDHandle(&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 FORCE_PING.
+ *
+ * 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;
+ LNetInvalidateMDHandle(&lp->lp_ping_mdh);
+ lp->lp_state &= ~LNET_PEER_PING_FAILED;
+ lp->lp_ping_error = 0;
+ }
+ if (lp->lp_state & LNET_PEER_FORCE_PING)
+ lp->lp_state &= ~LNET_PEER_FORCE_PING;
+ lp->lp_state |= LNET_PEER_NIDS_UPTODATE;
+ spin_unlock(&lp->lp_lock);
+
+ if (!LNetMDHandleIsInvalid(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
+ * FORCE_PING 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;
+ LNetInvalidateMDHandle(&lp->lp_ping_mdh);
+ lp->lp_state &= ~LNET_PEER_PING_FAILED;
+ lp->lp_state |= LNET_PEER_FORCE_PING;
+ rc = lp->lp_ping_error;
+ lp->lp_ping_error = 0;
+ spin_unlock(&lp->lp_lock);
+
+ if (!LNetMDHandleIsInvalid(mdh))
+ LNetMDUnlink(mdh);
+
+ CDEBUG(D_NET, "peer %s:%d\n",
+ libcfs_nid2str(lp->lp_primary_nid), rc);
+
+ 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_FORCE_PING;
+ spin_unlock(&lp->lp_lock);
+
+ nnis = MAX(lp->lp_data_nnis, LNET_INTERFACES_MIN);
+ 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);
+ LNetInvalidateMDHandle(&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. So we clear PING_SENT, but do
+ * not set either PING_FAILED or FORCE_PING. In fact we need
+ * to clear PING_FAILED, because the unlink event handler will
+ * have set it if we called LNetMDUnlink() above.
+ */
+ spin_lock(&lp->lp_lock);
+ lp->lp_state &= ~(LNET_PEER_PING_SENT | LNET_PEER_PING_FAILED);
+ 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;
+ LNetInvalidateMDHandle(&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);
+
+ if (!LNetMDHandleIsInvalid(mdh))
+ 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;
+
+ /* Don't push to a non-multi-rail peer. */
+ if (!(lp->lp_state & LNET_PEER_MULTI_RAIL)) {
+ lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
+ return 0;
+ }
+
+ lp->lp_state |= LNET_PEER_PUSH_SENT;
+ lp->lp_state &= ~LNET_PEER_FORCE_PUSH;
+ 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);
+ LNetInvalidateMDHandle(&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. So we clear PUSH_SENT, but do
+ * not set PUSH_FAILED. In fact we need to clear PUSH_FAILED,
+ * because the unlink event handler will have set it if we
+ * called LNetMDUnlink() above.
+ */
+ spin_lock(&lp->lp_lock);
+ lp->lp_state &= ~(LNET_PEER_PUSH_SENT | LNET_PEER_PUSH_FAILED);
+ 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_REDISCOVER;
+ spin_unlock(&lp->lp_lock);
+}
+
+/*
+ * 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 |
+ LNET_PEER_REDISCOVER);
+
+ CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
+
+ return 0;
+}
+
+/*
+ * Mark the peer as to be rediscovered.
+ */
+static int lnet_peer_rediscover(struct lnet_peer *lp)
+__must_hold(&lp->lp_lock)
+{
+ lp->lp_state |= LNET_PEER_REDISCOVER;
+ lp->lp_state &= ~LNET_PEER_DISCOVERING;
+
+ CDEBUG(D_NET, "peer %s\n", libcfs_nid2str(lp->lp_primary_nid));
+
+ return 0;
+}
+
+/*
+ * Returns the first peer on the ln_dc_working queue if its timeout
+ * has expired. Takes the current time as an argument so as to not
+ * obsessively re-check the clock. The oldest discovery request will
+ * be at the head of the queue.
+ */
+static struct lnet_peer *lnet_peer_dc_timed_out(time64_t now)
+{
+ struct lnet_peer *lp;
+
+ if (list_empty(&the_lnet.ln_dc_working))
+ return NULL;
+ lp = list_first_entry(&the_lnet.ln_dc_working,
+ struct lnet_peer, lp_dc_list);
+ if (now < lp->lp_last_queued + DISCOVERY_TIMEOUT)
+ return NULL;
+ return lp;
+}
+
+/*
+ * Discovering this peer is taking too long. Cancel any Ping or Push
+ * that discovery is waiting on by unlinking the relevant MDs. The
+ * lnet_discovery_event_handler() will proceed from here and complete
+ * the cleanup.
+ */
+static void lnet_peer_discovery_timeout(struct lnet_peer *lp)
+{
+ struct lnet_handle_md ping_mdh;
+ struct lnet_handle_md push_mdh;
+
+ LNetInvalidateMDHandle(&ping_mdh);
+ LNetInvalidateMDHandle(&push_mdh);
+
+ spin_lock(&lp->lp_lock);
+ if (lp->lp_state & LNET_PEER_PING_SENT) {
+ ping_mdh = lp->lp_ping_mdh;
+ LNetInvalidateMDHandle(&lp->lp_ping_mdh);
+ }
+ if (lp->lp_state & LNET_PEER_PUSH_SENT) {
+ push_mdh = lp->lp_push_mdh;
+ LNetInvalidateMDHandle(&lp->lp_push_mdh);
+ }
+ spin_unlock(&lp->lp_lock);
+
+ if (!LNetMDHandleIsInvalid(ping_mdh))
+ LNetMDUnlink(ping_mdh);
+ if (!LNetMDHandleIsInvalid(push_mdh))
+ LNetMDUnlink(push_mdh);
+}
+
+/*
+ * Wait for work to be queued or some other change that must be
+ * attended to. Returns non-zero if the discovery thread should shut
+ * down.
+ */
+static int lnet_peer_discovery_wait_for_work(void)
+{
+ int cpt;
+ int rc = 0;
+
+ DEFINE_WAIT(wait);
+
+ cpt = lnet_net_lock_current();
+ for (;;) {
+ prepare_to_wait(&the_lnet.ln_dc_waitq, &wait,
+ TASK_INTERRUPTIBLE);
+ if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
+ break;
+ if (lnet_push_target_resize_needed())
+ break;
+ if (!list_empty(&the_lnet.ln_dc_request))
+ break;
+ if (lnet_peer_dc_timed_out(ktime_get_real_seconds()))
+ break;
+ lnet_net_unlock(cpt);
+
+ /*
+ * wakeup max every second to check if there are peers that
+ * have been stuck on the working queue for greater than
+ * the peer timeout.
+ */
+ schedule_timeout(cfs_time_seconds(1));
+ finish_wait(&the_lnet.ln_dc_waitq, &wait);
+ cpt = lnet_net_lock_current();
+ }
+ finish_wait(&the_lnet.ln_dc_waitq, &wait);
+
+ if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
+ rc = -ESHUTDOWN;
+
+ lnet_net_unlock(cpt);
+
+ CDEBUG(D_NET, "woken: %d\n", rc);
+
+ return rc;
+}
+
+/* The discovery thread. */
+static int lnet_peer_discovery(void *arg)
+{
+ struct lnet_peer *lp;
+ time64_t now;
+ int rc;
+
+ CDEBUG(D_NET, "started\n");
+ cfs_block_allsigs();
+
+ for (;;) {
if (lnet_peer_discovery_wait_for_work())
break;
lnet_net_lock(LNET_LOCK_EX);
if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
break;
+
+ /*
+ * Process all incoming discovery work requests. When
+ * discovery must wait on a peer to change state, it
+ * is added to the tail of the ln_dc_working queue. A
+ * timestamp keeps track of when the peer was added,
+ * so we can time out discovery requests that take too
+ * long.
+ */
while (!list_empty(&the_lnet.ln_dc_request)) {
lp = list_first_entry(&the_lnet.ln_dc_request,
struct lnet_peer, lp_dc_list);
list_move(&lp->lp_dc_list, &the_lnet.ln_dc_working);
+ /*
+ * set the time the peer was put on the dc_working
+ * queue. It shouldn't remain on the queue
+ * forever, in case the GET message (for ping)
+ * doesn't get a REPLY or the PUT message (for
+ * push) doesn't get an ACK.
+ *
+ * TODO: LNet Health will deal with this scenario
+ * in a generic way.
+ */
+ lp->lp_last_queued = ktime_get_real_seconds();
lnet_net_unlock(LNET_LOCK_EX);
- /* Just tag and release for now. */
+ /*
+ * Select an action depending on the state of
+ * the peer and whether discovery is disabled.
+ * The check whether discovery is disabled is
+ * done after the code that handles processing
+ * for arrived data, cleanup for failures, and
+ * forcing a Ping or Push.
+ */
spin_lock(&lp->lp_lock);
- if (lnet_peer_discovery_disabled) {
- lp->lp_state |= LNET_PEER_REDISCOVER;
- lp->lp_state &= ~(LNET_PEER_DISCOVERED |
- LNET_PEER_NIDS_UPTODATE |
- LNET_PEER_DISCOVERING);
- } else {
- lp->lp_state |= (LNET_PEER_DISCOVERED |
- LNET_PEER_NIDS_UPTODATE);
- lp->lp_state &= ~(LNET_PEER_REDISCOVER |
- LNET_PEER_DISCOVERING);
- }
+ CDEBUG(D_NET, "peer %s state %#x\n",
+ libcfs_nid2str(lp->lp_primary_nid),
+ lp->lp_state);
+ if (lp->lp_state & LNET_PEER_DATA_PRESENT)
+ rc = lnet_peer_data_present(lp);
+ else if (lp->lp_state & LNET_PEER_PING_FAILED)
+ rc = lnet_peer_ping_failed(lp);
+ else if (lp->lp_state & LNET_PEER_PUSH_FAILED)
+ rc = lnet_peer_push_failed(lp);
+ else if (lp->lp_state & LNET_PEER_FORCE_PING)
+ rc = lnet_peer_send_ping(lp);
+ else if (lp->lp_state & LNET_PEER_FORCE_PUSH)
+ rc = lnet_peer_send_push(lp);
+ else if (lnet_peer_discovery_disabled)
+ rc = lnet_peer_rediscover(lp);
+ else if (!(lp->lp_state & LNET_PEER_NIDS_UPTODATE))
+ rc = lnet_peer_send_ping(lp);
+ else if (lnet_peer_needs_push(lp))
+ rc = lnet_peer_send_push(lp);
+ else
+ rc = lnet_peer_discovered(lp);
+ CDEBUG(D_NET, "peer %s state %#x rc %d\n",
+ libcfs_nid2str(lp->lp_primary_nid),
+ lp->lp_state, rc);
spin_unlock(&lp->lp_lock);
lnet_net_lock(LNET_LOCK_EX);
+ if (rc == LNET_REDISCOVER_PEER) {
+ list_move(&lp->lp_dc_list,
+ &the_lnet.ln_dc_request);
+ } else if (rc) {
+ lnet_peer_discovery_error(lp, rc);
+ }
if (!(lp->lp_state & LNET_PEER_DISCOVERING))
lnet_peer_discovery_complete(lp);
if (the_lnet.ln_dc_state == LNET_DC_STATE_STOPPING)
break;
}
+
+ /*
+ * Now that the ln_dc_request queue has been emptied
+ * check the ln_dc_working queue for peers that are
+ * taking too long. Move all that are found to the
+ * ln_dc_expired queue and time out any pending
+ * Ping or Push. We have to drop the lnet_net_lock
+ * in the loop because lnet_peer_discovery_timeout()
+ * calls LNetMDUnlink().
+ */
+ now = ktime_get_real_seconds();
+ while ((lp = lnet_peer_dc_timed_out(now)) != NULL) {
+ list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
+ lnet_net_unlock(LNET_LOCK_EX);
+ lnet_peer_discovery_timeout(lp);
+ lnet_net_lock(LNET_LOCK_EX);
+ }
+
lnet_net_unlock(LNET_LOCK_EX);
}
LNetEQFree(the_lnet.ln_dc_eqh);
LNetInvalidateEQHandle(&the_lnet.ln_dc_eqh);
+ /* Queue cleanup 1: stop all pending pings and pushes. */
lnet_net_lock(LNET_LOCK_EX);
- list_for_each_entry(lp, &the_lnet.ln_dc_request, lp_dc_list) {
- spin_lock(&lp->lp_lock);
- lp->lp_state |= LNET_PEER_REDISCOVER;
- lp->lp_state &= ~(LNET_PEER_DISCOVERED |
- LNET_PEER_DISCOVERING |
- LNET_PEER_NIDS_UPTODATE);
- spin_unlock(&lp->lp_lock);
- lnet_peer_discovery_complete(lp);
+ while (!list_empty(&the_lnet.ln_dc_working)) {
+ lp = list_first_entry(&the_lnet.ln_dc_working,
+ struct lnet_peer, lp_dc_list);
+ list_move(&lp->lp_dc_list, &the_lnet.ln_dc_expired);
+ lnet_net_unlock(LNET_LOCK_EX);
+ lnet_peer_discovery_timeout(lp);
+ lnet_net_lock(LNET_LOCK_EX);
}
- list_for_each_entry(lp, &the_lnet.ln_dc_working, lp_dc_list) {
- spin_lock(&lp->lp_lock);
- lp->lp_state |= LNET_PEER_REDISCOVER;
- lp->lp_state &= ~(LNET_PEER_DISCOVERED |
- LNET_PEER_DISCOVERING |
- LNET_PEER_NIDS_UPTODATE);
- spin_unlock(&lp->lp_lock);
+ lnet_net_unlock(LNET_LOCK_EX);
+
+ /* Queue cleanup 2: wait for the expired queue to clear. */
+ while (!list_empty(&the_lnet.ln_dc_expired))
+ schedule_timeout(cfs_time_seconds(1));
+
+ /* Queue cleanup 3: clear the request queue. */
+ lnet_net_lock(LNET_LOCK_EX);
+ while (!list_empty(&the_lnet.ln_dc_request)) {
+ lp = list_first_entry(&the_lnet.ln_dc_request,
+ struct lnet_peer, lp_dc_list);
+ lnet_peer_discovery_error(lp, -ESHUTDOWN);
lnet_peer_discovery_complete(lp);
}
lnet_net_unlock(LNET_LOCK_EX);
if (the_lnet.ln_dc_state != LNET_DC_STATE_SHUTDOWN)
return -EALREADY;
- INIT_LIST_HEAD(&the_lnet.ln_dc_request);
- INIT_LIST_HEAD(&the_lnet.ln_dc_working);
- init_waitqueue_head(&the_lnet.ln_dc_waitq);
-
rc = LNetEQAlloc(0, lnet_discovery_event_handler, &the_lnet.ln_dc_eqh);
if (rc != 0) {
CERROR("Can't allocate discovery EQ: %d\n", rc);
the_lnet.ln_dc_state = LNET_DC_STATE_SHUTDOWN;
}
+ CDEBUG(D_NET, "discovery start: %d\n", rc);
+
return rc;
}
LASSERT(list_empty(&the_lnet.ln_dc_request));
LASSERT(list_empty(&the_lnet.ln_dc_working));
+ LASSERT(list_empty(&the_lnet.ln_dc_expired));
+
+ CDEBUG(D_NET, "discovery stopped\n");
}
/* Debugging */