ev->status = 0;
ev->unlinked = 1;
ev->type = LNET_EVENT_UNLINK;
- lnet_md_deconstruct(md, &ev->md);
+ lnet_md_deconstruct(md, ev);
lnet_md2handle(&ev->md_handle, md);
EXIT;
}
/* build umd in event */
lnet_md2handle(&msg->msg_ev.md_handle, md);
- lnet_md_deconstruct(md, &msg->msg_ev.md);
+ lnet_md_deconstruct(md, &msg->msg_ev);
}
static int
msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
- /* NB: we probably want to use NID of msg::msg_from as 3rd
- * parameter (router NID) if it's routed message */
rc = lnet_send(msg->msg_ev.target.nid, msg, msg->msg_from);
lnet_net_lock(cpt);
*/
if (list_empty(&local_ni->ni_recovery) &&
atomic_read(&local_ni->ni_healthv) < LNET_MAX_HEALTH_VALUE) {
- CERROR("ni %s added to recovery queue. Health = %d\n",
+ CDEBUG(D_NET, "ni %s added to recovery queue. Health = %d\n",
libcfs_nid2str(local_ni->ni_nid),
atomic_read(&local_ni->ni_healthv));
list_add_tail(&local_ni->ni_recovery,
__u32 sensitivity = lnet_health_sensitivity;
__u32 lp_sensitivity;
- /* lpni could be NULL if we're in the LOLND case */
+ /*
+ * NO-OP if:
+ * 1. lpni could be NULL if we're in the LOLND case
+ * 2. this is a recovery message
+ */
if (!lpni)
return;
sensitivity = lp_sensitivity;
lnet_dec_healthv_locked(&lpni->lpni_healthv, sensitivity);
+
+ /* update the peer_net's health value */
+ lnet_update_peer_net_healthv(lpni);
+
/*
* add the peer NI to the recovery queue if it's not already there
* and it's health value is actually below the maximum. It's
}
static void
-lnet_incr_hstats(struct lnet_msg *msg, enum lnet_msg_hstatus hstatus)
+lnet_incr_hstats(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
+ enum lnet_msg_hstatus hstatus)
{
- struct lnet_ni *ni = msg->msg_txni;
- struct lnet_peer_ni *lpni = msg->msg_txpeer;
struct lnet_counters_health *health;
health = &the_lnet.ln_counters[0]->lct_health;
}
}
+static void
+lnet_resend_msg_locked(struct lnet_msg *msg)
+{
+ msg->msg_retry_count++;
+
+ /*
+ * remove message from the active list and reset it to prepare
+ * for a resend. Two exceptions to this
+ *
+ * 1. the router case. When a message is being routed it is
+ * committed for rx when received and committed for tx when
+ * forwarded. We don't want to remove it from the active list, since
+ * code which handles receiving expects it to remain on the active
+ * list.
+ *
+ * 2. The REPLY case. Reply messages use the same message
+ * structure for the GET that was received.
+ */
+ if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
+ list_del_init(&msg->msg_activelist);
+ msg->msg_onactivelist = 0;
+ }
+ /*
+ * The msg_target.nid which was originally set
+ * when calling LNetGet() or LNetPut() might've
+ * been overwritten if we're routing this message.
+ * Call lnet_msg_decommit_tx() to return the credit
+ * this message consumed. The message will
+ * consume another credit when it gets resent.
+ */
+ msg->msg_target.nid = msg->msg_hdr.dest_nid;
+ lnet_msg_decommit_tx(msg, -EAGAIN);
+ msg->msg_sending = 0;
+ msg->msg_receiving = 0;
+ msg->msg_target_is_router = 0;
+
+ CDEBUG(D_NET, "%s->%s:%s:%s - queuing msg (%p) for resend\n",
+ libcfs_nid2str(msg->msg_hdr.src_nid),
+ libcfs_nid2str(msg->msg_hdr.dest_nid),
+ lnet_msgtyp2str(msg->msg_type),
+ lnet_health_error2str(msg->msg_health_status), msg);
+
+ list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
+
+ complete(&the_lnet.ln_mt_wait_complete);
+}
+
+int
+lnet_check_finalize_recursion_locked(struct lnet_msg *msg,
+ struct list_head *containerq,
+ int nworkers, void **workers)
+{
+ int my_slot = -1;
+ int i;
+
+ list_add_tail(&msg->msg_list, containerq);
+
+ for (i = 0; i < nworkers; i++) {
+ if (workers[i] == current)
+ break;
+
+ if (my_slot < 0 && workers[i] == NULL)
+ my_slot = i;
+ }
+
+ if (i < nworkers || my_slot < 0)
+ return -1;
+
+ workers[my_slot] = current;
+
+ return my_slot;
+}
+
+int
+lnet_attempt_msg_resend(struct lnet_msg *msg)
+{
+ struct lnet_msg_container *container;
+ int my_slot;
+ int cpt;
+
+ /* we can only resend tx_committed messages */
+ LASSERT(msg->msg_tx_committed);
+
+ /* don't resend recovery messages */
+ if (msg->msg_recovery) {
+ CDEBUG(D_NET, "msg %s->%s is a recovery ping. retry# %d\n",
+ libcfs_nid2str(msg->msg_from),
+ libcfs_nid2str(msg->msg_target.nid),
+ msg->msg_retry_count);
+ return -ENOTRECOVERABLE;
+ }
+
+ /*
+ * if we explicitly indicated we don't want to resend then just
+ * return
+ */
+ if (msg->msg_no_resend) {
+ CDEBUG(D_NET, "msg %s->%s requested no resend. retry# %d\n",
+ libcfs_nid2str(msg->msg_from),
+ libcfs_nid2str(msg->msg_target.nid),
+ msg->msg_retry_count);
+ return -ENOTRECOVERABLE;
+ }
+
+ /* check if the message has exceeded the number of retries */
+ if (msg->msg_retry_count >= lnet_retry_count) {
+ CNETERR("msg %s->%s exceeded retry count %d\n",
+ libcfs_nid2str(msg->msg_from),
+ libcfs_nid2str(msg->msg_target.nid),
+ msg->msg_retry_count);
+ return -ENOTRECOVERABLE;
+ }
+
+ cpt = msg->msg_tx_cpt;
+ lnet_net_lock(cpt);
+
+ /* check again under lock */
+ if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
+ lnet_net_unlock(cpt);
+ return -ESHUTDOWN;
+ }
+
+ container = the_lnet.ln_msg_containers[cpt];
+ my_slot =
+ lnet_check_finalize_recursion_locked(msg,
+ &container->msc_resending,
+ container->msc_nfinalizers,
+ container->msc_resenders);
+
+ /* enough threads are resending */
+ if (my_slot == -1) {
+ lnet_net_unlock(cpt);
+ return 0;
+ }
+
+ while (!list_empty(&container->msc_resending)) {
+ msg = list_entry(container->msc_resending.next,
+ struct lnet_msg, msg_list);
+ list_del(&msg->msg_list);
+
+ /*
+ * resending the message will require us to call
+ * lnet_msg_decommit_tx() which will return the credit
+ * which this message holds. This could trigger another
+ * queued message to be sent. If that message fails and
+ * requires a resend we will recurse.
+ * But since at this point the slot is taken, the message
+ * will be queued in the container and dealt with
+ * later. This breaks the recursion.
+ */
+ lnet_resend_msg_locked(msg);
+ }
+
+ /*
+ * msc_resenders is an array of process pointers. Each entry holds
+ * a pointer to the current process operating on the message. An
+ * array entry is created per CPT. If the array slot is already
+ * set, then it means that there is a thread on the CPT currently
+ * resending a message.
+ * Once the thread finishes clear the slot to enable the thread to
+ * take on more resend work.
+ */
+ container->msc_resenders[my_slot] = NULL;
+ lnet_net_unlock(cpt);
+
+ return 0;
+}
+
/*
* Do a health check on the message:
* return -1 if we're not going to handle the error or
lnet_health_check(struct lnet_msg *msg)
{
enum lnet_msg_hstatus hstatus = msg->msg_health_status;
- bool lo = false;
- struct lnet_ni *ni;
struct lnet_peer_ni *lpni;
+ struct lnet_ni *ni;
+ bool lo = false;
+ bool attempt_local_resend;
+ bool attempt_remote_resend;
+ bool handle_local_health;
+ bool handle_remote_health;
/* if we're shutting down no point in handling health. */
if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
* if we're sending to the LOLND then the msg_txpeer will not be
* set. So no need to sanity check it.
*/
- if (msg->msg_tx_committed &&
- LNET_NETTYP(LNET_NIDNET(msg->msg_txni->ni_nid)) != LOLND)
+ if (msg->msg_tx_committed && msg->msg_txni->ni_nid != LNET_NID_LO_0)
LASSERT(msg->msg_txpeer);
else if (msg->msg_tx_committed &&
- LNET_NETTYP(LNET_NIDNET(msg->msg_txni->ni_nid)) == LOLND)
+ msg->msg_txni->ni_nid == LNET_NID_LO_0)
lo = true;
if (hstatus != LNET_MSG_STATUS_OK &&
return -1;
/*
- * stats are only incremented for errors so avoid wasting time
- * incrementing statistics if there is no error.
- */
- if (hstatus != LNET_MSG_STATUS_OK) {
- lnet_net_lock(0);
- lnet_incr_hstats(msg, hstatus);
- lnet_net_unlock(0);
- }
-
- /*
* always prefer txni/txpeer if they message is committed for both
* directions.
*/
if (msg->msg_tx_committed) {
ni = msg->msg_txni;
lpni = msg->msg_txpeer;
+ attempt_local_resend = attempt_remote_resend = true;
} else {
ni = msg->msg_rxni;
lpni = msg->msg_rxpeer;
+ attempt_local_resend = attempt_remote_resend = false;
+ }
+
+ /* Don't further decrement the health value if a recovery message
+ * failed.
+ */
+ if (msg->msg_recovery)
+ handle_local_health = handle_remote_health = false;
+ else
+ handle_local_health = handle_remote_health = true;
+
+ /* For local failures, health/recovery/resends are not needed if I only
+ * have a single (non-lolnd) interface. NB: pb_nnis includes the lolnd
+ * interface, so a single-rail node would have pb_nnis == 2.
+ */
+ if (the_lnet.ln_ping_target->pb_nnis <= 2) {
+ handle_local_health = false;
+ attempt_local_resend = false;
+ }
+
+ /* For remote failures, health/recovery/resends are not needed if the
+ * peer only has a single interface. Special case for routers where we
+ * rely on health feature to manage route aliveness. NB: unlike pb_nnis
+ * above, lp_nnis does _not_ include the lolnd, so a single-rail node
+ * would have lp_nnis == 1.
+ */
+ if (lpni && lpni->lpni_peer_net->lpn_peer->lp_nnis <= 1) {
+ attempt_remote_resend = false;
+ if (!lnet_isrouter(lpni))
+ handle_remote_health = false;
}
if (!lo)
lnet_msgtyp2str(msg->msg_type),
lnet_health_error2str(hstatus));
+ /*
+ * stats are only incremented for errors so avoid wasting time
+ * incrementing statistics if there is no error.
+ */
+ if (hstatus != LNET_MSG_STATUS_OK) {
+ lnet_net_lock(0);
+ lnet_incr_hstats(ni, lpni, hstatus);
+ lnet_net_unlock(0);
+ }
+
switch (hstatus) {
case LNET_MSG_STATUS_OK:
/*
* increment the local ni health weather we successfully
* received or sent a message on it.
*/
- lnet_inc_healthv(&ni->ni_healthv);
+ lnet_inc_healthv(&ni->ni_healthv, lnet_health_sensitivity);
/*
* It's possible msg_txpeer is NULL in the LOLND
* case. Only increment the peer's health if we're
* I'm a router, then set that lpni's health to
* maximum so we can commence communication
*/
- if (lnet_isrouter(lpni) || the_lnet.ln_routing)
- lnet_set_healthv(&lpni->lpni_healthv,
- LNET_MAX_HEALTH_VALUE);
- else
- lnet_inc_healthv(&lpni->lpni_healthv);
+ lnet_net_lock(0);
+ if (lnet_isrouter(lpni) || the_lnet.ln_routing) {
+ lnet_set_lpni_healthv_locked(lpni,
+ LNET_MAX_HEALTH_VALUE);
+ } else {
+ __u32 sensitivity = lpni->lpni_peer_net->
+ lpn_peer->lp_health_sensitivity;
+
+ lnet_inc_lpni_healthv_locked(lpni,
+ (sensitivity) ? sensitivity :
+ lnet_health_sensitivity);
+ }
+ lnet_net_unlock(0);
}
/* we can finalize this message */
case LNET_MSG_STATUS_LOCAL_ABORTED:
case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
case LNET_MSG_STATUS_LOCAL_TIMEOUT:
- lnet_handle_local_failure(ni);
- if (msg->msg_tx_committed)
- /* add to the re-send queue */
- goto resend;
+ if (handle_local_health)
+ lnet_handle_local_failure(ni);
+ if (attempt_local_resend)
+ return lnet_attempt_msg_resend(msg);
break;
-
- /*
- * These errors will not trigger a resend so simply
- * finalize the message
- */
case LNET_MSG_STATUS_LOCAL_ERROR:
- lnet_handle_local_failure(ni);
+ if (handle_local_health)
+ lnet_handle_local_failure(ni);
return -1;
-
- /*
- * TODO: since the remote dropped the message we can
- * attempt a resend safely.
- */
case LNET_MSG_STATUS_REMOTE_DROPPED:
- lnet_handle_remote_failure(lpni);
- if (msg->msg_tx_committed)
- goto resend;
+ if (handle_remote_health)
+ lnet_handle_remote_failure(lpni);
+ if (attempt_remote_resend)
+ return lnet_attempt_msg_resend(msg);
break;
-
case LNET_MSG_STATUS_REMOTE_ERROR:
case LNET_MSG_STATUS_REMOTE_TIMEOUT:
case LNET_MSG_STATUS_NETWORK_TIMEOUT:
- lnet_handle_remote_failure(lpni);
+ if (handle_remote_health)
+ lnet_handle_remote_failure(lpni);
return -1;
default:
LBUG();
}
-resend:
- /* we can only resend tx_committed messages */
- LASSERT(msg->msg_tx_committed);
-
- /* don't resend recovery messages */
- if (msg->msg_recovery) {
- CDEBUG(D_NET, "msg %s->%s is a recovery ping. retry# %d\n",
- libcfs_nid2str(msg->msg_from),
- libcfs_nid2str(msg->msg_target.nid),
- msg->msg_retry_count);
- return -1;
- }
-
- /*
- * if we explicitly indicated we don't want to resend then just
- * return
- */
- if (msg->msg_no_resend) {
- CDEBUG(D_NET, "msg %s->%s requested no resend. retry# %d\n",
- libcfs_nid2str(msg->msg_from),
- libcfs_nid2str(msg->msg_target.nid),
- msg->msg_retry_count);
- return -1;
- }
-
- /* check if the message has exceeded the number of retries */
- if (msg->msg_retry_count >= lnet_retry_count) {
- CNETERR("msg %s->%s exceeded retry count %d\n",
- libcfs_nid2str(msg->msg_from),
- libcfs_nid2str(msg->msg_target.nid),
- msg->msg_retry_count);
- return -1;
- }
- msg->msg_retry_count++;
-
- lnet_net_lock(msg->msg_tx_cpt);
-
- /* check again under lock */
- if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
- lnet_net_unlock(msg->msg_tx_cpt);
- return -1;
- }
-
- /*
- * remove message from the active list and reset it in preparation
- * for a resend. Two exception to this
- *
- * 1. the router case, whe a message is committed for rx when
- * received, then tx when it is sent. When committed to both tx and
- * rx we don't want to remove it from the active list.
- *
- * 2. The REPLY case since it uses the same msg block for the GET
- * that was received.
- */
- if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
- list_del_init(&msg->msg_activelist);
- msg->msg_onactivelist = 0;
- }
- /*
- * The msg_target.nid which was originally set
- * when calling LNetGet() or LNetPut() might've
- * been overwritten if we're routing this message.
- * Call lnet_return_tx_credits_locked() to return
- * the credit this message consumed. The message will
- * consume another credit when it gets resent.
- */
- msg->msg_target.nid = msg->msg_hdr.dest_nid;
- lnet_msg_decommit_tx(msg, -EAGAIN);
- msg->msg_sending = 0;
- msg->msg_receiving = 0;
- msg->msg_target_is_router = 0;
-
- CDEBUG(D_NET, "%s->%s:%s:%s - queuing for resend\n",
- libcfs_nid2str(msg->msg_hdr.src_nid),
- libcfs_nid2str(msg->msg_hdr.dest_nid),
- lnet_msgtyp2str(msg->msg_type),
- lnet_health_error2str(hstatus));
-
- list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
- lnet_net_unlock(msg->msg_tx_cpt);
-
- wake_up(&the_lnet.ln_mt_waitq);
- return 0;
+ /* no resend is needed */
+ return -1;
}
static void
LASSERT(md->md_refcount >= 0);
unlink = lnet_md_unlinkable(md);
- if (md->md_eq != NULL) {
- msg->msg_ev.status = status;
+ if (md->md_handler) {
+ if ((md->md_flags & LNET_MD_FLAG_ABORTED) && !status) {
+ msg->msg_ev.status = -ETIMEDOUT;
+ CDEBUG(D_NET, "md 0x%p already unlinked\n", md);
+ } else {
+ msg->msg_ev.status = status;
+ }
msg->msg_ev.unlinked = unlink;
- lnet_eq_enqueue_event(md->md_eq, &msg->msg_ev);
+ md->md_handler(&msg->msg_ev);
}
- if (unlink) {
+ if (unlink || (md->md_refcount == 0 &&
+ md->md_threshold == LNET_MD_THRESH_INF))
lnet_detach_rsp_tracker(md, cpt);
+
+ if (unlink)
lnet_md_unlink(md);
- }
msg->msg_md = NULL;
}
int my_slot;
int cpt;
int rc;
- int i;
LASSERT(!in_interrupt());
* put on the resend queue.
*/
if (!lnet_health_check(msg))
- /* Message is queued for resend */
return;
}
lnet_net_lock(cpt);
container = the_lnet.ln_msg_containers[cpt];
- list_add_tail(&msg->msg_list, &container->msc_finalizing);
/* Recursion breaker. Don't complete the message here if I am (or
* enough other threads are) already completing messages */
+ my_slot = lnet_check_finalize_recursion_locked(msg,
+ &container->msc_finalizing,
+ container->msc_nfinalizers,
+ container->msc_finalizers);
- my_slot = -1;
- for (i = 0; i < container->msc_nfinalizers; i++) {
- if (container->msc_finalizers[i] == current)
- break;
-
- if (my_slot < 0 && container->msc_finalizers[i] == NULL)
- my_slot = i;
- }
-
- if (i < container->msc_nfinalizers || my_slot < 0) {
+ /* enough threads are resending */
+ if (my_slot == -1) {
lnet_net_unlock(cpt);
return;
}
- container->msc_finalizers[my_slot] = current;
-
rc = 0;
while (!list_empty(&container->msc_finalizing)) {
msg = list_entry(container->msc_finalizing.next,
CERROR("%d active msg on exit\n", count);
if (container->msc_finalizers != NULL) {
- LIBCFS_FREE(container->msc_finalizers,
- container->msc_nfinalizers *
- sizeof(*container->msc_finalizers));
+ CFS_FREE_PTR_ARRAY(container->msc_finalizers,
+ container->msc_nfinalizers);
container->msc_finalizers = NULL;
}
+
+ if (container->msc_resenders != NULL) {
+ CFS_FREE_PTR_ARRAY(container->msc_resenders,
+ container->msc_nfinalizers);
+ container->msc_resenders = NULL;
+ }
container->msc_init = 0;
}
INIT_LIST_HEAD(&container->msc_active);
INIT_LIST_HEAD(&container->msc_finalizing);
+ INIT_LIST_HEAD(&container->msc_resending);
/* number of CPUs */
container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
return -ENOMEM;
}
+ LIBCFS_CPT_ALLOC(container->msc_resenders, lnet_cpt_table(), cpt,
+ container->msc_nfinalizers *
+ sizeof(*container->msc_resenders));
+
+ if (container->msc_resenders == NULL) {
+ CERROR("Failed to allocate message resenders\n");
+ lnet_msg_container_cleanup(container);
+ return -ENOMEM;
+ }
+
return rc;
}