/*
* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
*
- * Copyright (c) 2011, 2016, Intel Corporation.
+ * Copyright (c) 2011, 2017, Intel Corporation.
*
- * This file is part of Lustre, https://wiki.hpdd.intel.com/
+ * This file is part of Lustre, https://wiki.whamcloud.com/
*
* Portals is free software; you can redistribute it and/or
* modify it under the terms of version 2 of the GNU General Public
*/
#define DEBUG_SUBSYSTEM S_LNET
+
+#include <linux/random.h>
#include <lnet/lib-lnet.h>
#define LNET_NRB_TINY_MIN 512 /* min value for each CPT */
#define LNET_NRB_LARGE_PAGES ((LNET_MTU + PAGE_SIZE - 1) >> \
PAGE_SHIFT)
+extern unsigned int lnet_current_net_count;
+
static char *forwarding = "";
module_param(forwarding, charp, 0444);
MODULE_PARM_DESC(forwarding, "Explicitly enable/disable forwarding between networks");
return net->net_tunables.lct_peer_tx_credits;
}
-/* forward ref's */
-static int lnet_router_checker(void *);
-
static int check_routers_before_use;
module_param(check_routers_before_use, int, 0444);
MODULE_PARM_DESC(check_routers_before_use, "Assume routers are down and ping them before use");
module_param(avoid_asym_router_failure, int, 0644);
MODULE_PARM_DESC(avoid_asym_router_failure, "Avoid asymmetrical router failures (0 to disable)");
-static int dead_router_check_interval = 60;
-module_param(dead_router_check_interval, int, 0644);
-MODULE_PARM_DESC(dead_router_check_interval, "Seconds between dead router health checks (<= 0 to disable)");
-
-static int live_router_check_interval = 60;
-module_param(live_router_check_interval, int, 0644);
-MODULE_PARM_DESC(live_router_check_interval, "Seconds between live router health checks (<= 0 to disable)");
+int alive_router_check_interval = 60;
+module_param(alive_router_check_interval, int, 0644);
+MODULE_PARM_DESC(alive_router_check_interval, "Seconds between live router health checks (<= 0 to disable)");
static int router_ping_timeout = 50;
module_param(router_ping_timeout, int, 0644);
MODULE_PARM_DESC(router_ping_timeout, "Seconds to wait for the reply to a router health query");
-int
-lnet_peers_start_down(void)
-{
- return check_routers_before_use;
-}
+/*
+ * A value between 0 and 100. 0 meaning that even if router's interfaces
+ * have the worse health still consider the gateway usable.
+ * 100 means that at least one interface on the route's remote net is 100%
+ * healthy to consider the route alive.
+ * The default is set to 100 to ensure we maintain the original behavior.
+ */
+unsigned int router_sensitivity_percentage = 100;
+static int rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp);
+static struct kernel_param_ops param_ops_rtr_sensitivity = {
+ .set = rtr_sensitivity_set,
+ .get = param_get_int,
+};
+#define param_check_rtr_sensitivity(name, p) \
+ __param_check(name, p, int)
+#ifdef HAVE_KERNEL_PARAM_OPS
+module_param(router_sensitivity_percentage, rtr_sensitivity, S_IRUGO|S_IWUSR);
+#else
+module_param_call(router_sensitivity_percentage, rtr_sensitivity_set, param_get_int,
+ &router_sensitivity_percentage, S_IRUGO|S_IWUSR);
+#endif
+MODULE_PARM_DESC(router_sensitivity_percentage,
+ "How healthy a gateway should be to be used in percent");
-void
-lnet_notify_locked(struct lnet_peer_ni *lp, int notifylnd, int alive,
- cfs_time_t when)
+static int
+rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
{
- if (cfs_time_before(when, lp->lpni_timestamp)) { /* out of date information */
- CDEBUG(D_NET, "Out of date\n");
- return;
+ int rc;
+ unsigned *sen = (unsigned *)kp->arg;
+ unsigned long value;
+
+ rc = kstrtoul(val, 0, &value);
+ if (rc) {
+ CERROR("Invalid module parameter value for 'router_sensitivity_percentage'\n");
+ return rc;
+ }
+
+ if (value < 0 || value > 100) {
+ CERROR("Invalid value: %lu for 'router_sensitivity_percentage'\n", value);
+ return -EINVAL;
}
/*
- * This function can be called with different cpt locks being
- * held. lpni_alive_count modification needs to be properly protected.
- * Significant reads to lpni_alive_count are also protected with
- * the same lock
+ * The purpose of locking the api_mutex here is to ensure that
+ * the correct value ends up stored properly.
*/
- spin_lock(&lp->lpni_lock);
+ mutex_lock(&the_lnet.ln_api_mutex);
- lp->lpni_timestamp = when; /* update timestamp */
- lp->lpni_ping_deadline = 0; /* disable ping timeout */
+ *sen = value;
- if (lp->lpni_alive_count != 0 && /* got old news */
- (!lp->lpni_alive) == (!alive)) { /* new date for old news */
- spin_unlock(&lp->lpni_lock);
- CDEBUG(D_NET, "Old news\n");
- return;
- }
+ mutex_unlock(&the_lnet.ln_api_mutex);
- /* Flag that notification is outstanding */
+ return 0;
+}
- lp->lpni_alive_count++;
- lp->lpni_alive = (alive) ? 1 : 0;
- lp->lpni_notify = 1;
- lp->lpni_notifylnd = notifylnd;
- if (lp->lpni_alive)
- lp->lpni_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
+void
+lnet_rtr_transfer_to_peer(struct lnet_peer *src, struct lnet_peer *target)
+{
+ struct lnet_route *route;
- spin_unlock(&lp->lpni_lock);
+ lnet_net_lock(LNET_LOCK_EX);
+ target->lp_rtr_refcount += src->lp_rtr_refcount;
+ /* move the list of queued messages to the new peer */
+ list_splice_init(&src->lp_rtrq, &target->lp_rtrq);
+ /* move all the routes that reference the peer */
+ list_splice_init(&src->lp_routes, &target->lp_routes);
+ /* update all the routes to point to the new peer */
+ list_for_each_entry(route, &target->lp_routes, lr_gwlist)
+ route->lr_gateway = target;
+ /* remove the old peer from the ln_routers list */
+ list_del_init(&src->lp_rtr_list);
+ /* add the new peer to the ln_routers list */
+ if (list_empty(&target->lp_rtr_list)) {
+ lnet_peer_addref_locked(target);
+ list_add_tail(&target->lp_rtr_list, &the_lnet.ln_routers);
+ }
+ /* reset the ref count on the old peer and decrement its ref count */
+ src->lp_rtr_refcount = 0;
+ lnet_peer_decref_locked(src);
+ /* update the router version */
+ the_lnet.ln_routers_version++;
+ lnet_net_unlock(LNET_LOCK_EX);
+}
- CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lpni_nid), alive);
+int
+lnet_peers_start_down(void)
+{
+ return check_routers_before_use;
}
/*
- * This function will always be called with lp->lpni_cpt lock held.
+ * A net is alive if at least one gateway NI on the network is alive.
*/
-static void
-lnet_ni_notify_locked(lnet_ni_t *ni, struct lnet_peer_ni *lp)
+static bool
+lnet_is_gateway_net_alive(struct lnet_peer_net *lpn)
{
- int alive;
- int notifylnd;
+ struct lnet_peer_ni *lpni;
- /* Notify only in 1 thread at any time to ensure ordered notification.
- * NB individual events can be missed; the only guarantee is that you
- * always get the most recent news */
+ list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
+ if (lnet_is_peer_ni_alive(lpni))
+ return true;
+ }
- spin_lock(&lp->lpni_lock);
+ return false;
+}
- if (lp->lpni_notifying || ni == NULL) {
- spin_unlock(&lp->lpni_lock);
- return;
+/*
+ * a gateway is alive only if all its nets are alive
+ * called with cpt lock held
+ */
+bool lnet_is_gateway_alive(struct lnet_peer *gw)
+{
+ struct lnet_peer_net *lpn;
+
+ list_for_each_entry(lpn, &gw->lp_peer_nets, lpn_peer_nets) {
+ if (!lnet_is_gateway_net_alive(lpn))
+ return false;
}
- lp->lpni_notifying = 1;
+ return true;
+}
+
+/*
+ * lnet_is_route_alive() needs to be called with cpt lock held
+ * A route is alive if the gateway can route between the local network and
+ * the remote network of the route.
+ * This means at least one NI is alive on each of the local and remote
+ * networks of the gateway.
+ */
+bool lnet_is_route_alive(struct lnet_route *route)
+{
+ struct lnet_peer *gw = route->lr_gateway;
+ struct lnet_peer_net *llpn;
+ struct lnet_peer_net *rlpn;
+ bool route_alive;
/*
- * lp->lpni_notify needs to be protected because it can be set in
- * lnet_notify_locked().
+ * check the gateway's interfaces on the route rnet to make sure
+ * that the gateway is viable.
*/
- while (lp->lpni_notify) {
- alive = lp->lpni_alive;
- notifylnd = lp->lpni_notifylnd;
+ llpn = lnet_peer_get_net_locked(gw, route->lr_lnet);
+ if (!llpn)
+ return false;
+
+ route_alive = lnet_is_gateway_net_alive(llpn);
+
+ if (avoid_asym_router_failure) {
+ rlpn = lnet_peer_get_net_locked(gw, route->lr_net);
+ if (!rlpn)
+ return false;
+ route_alive = route_alive &&
+ lnet_is_gateway_net_alive(rlpn);
+ }
- lp->lpni_notifylnd = 0;
- lp->lpni_notify = 0;
+ if (!route_alive)
+ return route_alive;
- if (notifylnd && ni->ni_net->net_lnd->lnd_notify != NULL) {
- spin_unlock(&lp->lpni_lock);
- lnet_net_unlock(lp->lpni_cpt);
+ spin_lock(&gw->lp_lock);
+ if (!(gw->lp_state & LNET_PEER_ROUTER_ENABLED)) {
+ if (gw->lp_rtr_refcount > 0)
+ CERROR("peer %s is being used as a gateway but routing feature is not turned on\n",
+ libcfs_nid2str(gw->lp_primary_nid));
+ route_alive = false;
+ }
+ spin_unlock(&gw->lp_lock);
- /* A new notification could happen now; I'll handle it
- * when control returns to me */
+ return route_alive;
+}
- (ni->ni_net->net_lnd->lnd_notify)(ni, lp->lpni_nid,
- alive);
+void
+lnet_consolidate_routes_locked(struct lnet_peer *orig_lp,
+ struct lnet_peer *new_lp)
+{
+ struct lnet_peer_ni *lpni;
+ struct lnet_route *route;
- lnet_net_lock(lp->lpni_cpt);
- spin_lock(&lp->lpni_lock);
+ /*
+ * Although a route is correlated with a peer, but when it's added
+ * a specific NID is used. That NID refers to a peer_ni within
+ * a peer. There could be other peer_nis on the same net, which
+ * can be used to send to that gateway. However when we are
+ * consolidating gateways because of discovery, the nid used to
+ * add the route might've moved between gateway peers. In this
+ * case we want to move the route to the new gateway as well. The
+ * intent here is not to confuse the user who added the route.
+ */
+ list_for_each_entry(route, &orig_lp->lp_routes, lr_gwlist) {
+ lpni = lnet_peer_get_ni_locked(orig_lp, route->lr_nid);
+ if (!lpni) {
+ lnet_net_lock(LNET_LOCK_EX);
+ list_move(&route->lr_gwlist, &new_lp->lp_routes);
+ lnet_net_unlock(LNET_LOCK_EX);
}
}
- lp->lpni_notifying = 0;
- spin_unlock(&lp->lpni_lock);
}
-static void
-lnet_rtr_addref_locked(struct lnet_peer_ni *lp)
+void
+lnet_router_discovery_complete(struct lnet_peer *lp)
{
- LASSERT(atomic_read(&lp->lpni_refcount) > 0);
- LASSERT(lp->lpni_rtr_refcount >= 0);
+ struct lnet_peer_ni *lpni = NULL;
- /* lnet_net_lock must be exclusively locked */
- lp->lpni_rtr_refcount++;
- if (lp->lpni_rtr_refcount == 1) {
- struct list_head *pos;
+ spin_lock(&lp->lp_lock);
+ lp->lp_state &= ~LNET_PEER_RTR_DISCOVERY;
+ spin_unlock(&lp->lp_lock);
- /* a simple insertion sort */
- list_for_each_prev(pos, &the_lnet.ln_routers) {
- struct lnet_peer_ni *rtr =
- list_entry(pos, struct lnet_peer_ni,
- lpni_rtr_list);
+ /*
+ * Router discovery successful? All peer information would've been
+ * updated already. No need to do any more processing
+ */
+ if (!lp->lp_dc_error)
+ return;
+ /*
+ * discovery failed? then we need to set the status of each lpni
+ * to DOWN. It will be updated the next time we discover the
+ * router. For router peer NIs not on local networks, we never send
+ * messages directly to them, so their health will always remain
+ * at maximum. We can only tell if they are up or down from the
+ * status returned in the PING response. If we fail to get that
+ * status in our scheduled router discovery, then we'll assume
+ * it's down until we're told otherwise.
+ */
+ CDEBUG(D_NET, "%s: Router discovery failed %d\n",
+ libcfs_nid2str(lp->lp_primary_nid), lp->lp_dc_error);
+ while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
+ lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
+}
- if (rtr->lpni_nid < lp->lpni_nid)
- break;
- }
+static void
+lnet_rtr_addref_locked(struct lnet_peer *lp)
+{
+ LASSERT(lp->lp_rtr_refcount >= 0);
- list_add(&lp->lpni_rtr_list, pos);
+ /* lnet_net_lock must be exclusively locked */
+ lp->lp_rtr_refcount++;
+ if (lp->lp_rtr_refcount == 1) {
+ list_add_tail(&lp->lp_rtr_list, &the_lnet.ln_routers);
/* addref for the_lnet.ln_routers */
- lnet_peer_ni_addref_locked(lp);
+ lnet_peer_addref_locked(lp);
the_lnet.ln_routers_version++;
}
}
static void
-lnet_rtr_decref_locked(struct lnet_peer_ni *lp)
+lnet_rtr_decref_locked(struct lnet_peer *lp)
{
- LASSERT(atomic_read(&lp->lpni_refcount) > 0);
- LASSERT(lp->lpni_rtr_refcount > 0);
+ LASSERT(atomic_read(&lp->lp_refcount) > 0);
+ LASSERT(lp->lp_rtr_refcount > 0);
/* lnet_net_lock must be exclusively locked */
- lp->lpni_rtr_refcount--;
- if (lp->lpni_rtr_refcount == 0) {
- LASSERT(list_empty(&lp->lpni_routes));
-
- if (lp->lpni_rcd != NULL) {
- list_add(&lp->lpni_rcd->rcd_list,
- &the_lnet.ln_rcd_deathrow);
- lp->lpni_rcd = NULL;
- }
+ lp->lp_rtr_refcount--;
+ if (lp->lp_rtr_refcount == 0) {
+ LASSERT(list_empty(&lp->lp_routes));
- list_del(&lp->lpni_rtr_list);
+ list_del(&lp->lp_rtr_list);
/* decref for the_lnet.ln_routers */
- lnet_peer_ni_decref_locked(lp);
+ lnet_peer_decref_locked(lp);
the_lnet.ln_routers_version++;
}
}
-lnet_remotenet_t *
+struct lnet_remotenet *
lnet_find_rnet_locked(__u32 net)
{
- lnet_remotenet_t *rnet;
+ struct lnet_remotenet *rnet;
struct list_head *tmp;
struct list_head *rn_list;
- LASSERT(!the_lnet.ln_shutdown);
+ LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
rn_list = lnet_net2rnethash(net);
list_for_each(tmp, rn_list) {
- rnet = list_entry(tmp, lnet_remotenet_t, lrn_list);
+ rnet = list_entry(tmp, struct lnet_remotenet, lrn_list);
if (rnet->lrn_net == net)
return rnet;
static void lnet_shuffle_seed(void)
{
static int seeded;
- __u32 lnd_type;
- __u32 seed[2];
- struct timespec64 ts;
- lnet_ni_t *ni = NULL;
+ struct lnet_ni *ni = NULL;
if (seeded)
return;
- cfs_get_random_bytes(seed, sizeof(seed));
-
/* Nodes with small feet have little entropy
* the NID for this node gives the most entropy in the low bits */
- while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
- lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
-
- if (lnd_type != LOLND)
- seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type);
- }
+ while ((ni = lnet_get_next_ni_locked(NULL, ni)))
+ add_device_randomness(&ni->ni_nid, sizeof(ni->ni_nid));
- ktime_get_ts64(&ts);
- cfs_srand(ts.tv_sec ^ seed[0], ts.tv_nsec ^ seed[1]);
seeded = 1;
return;
}
/* NB expects LNET_LOCK held */
static void
-lnet_add_route_to_rnet(lnet_remotenet_t *rnet, lnet_route_t *route)
+lnet_add_route_to_rnet(struct lnet_remotenet *rnet, struct lnet_route *route)
{
- unsigned int len = 0;
- unsigned int offset = 0;
+ struct lnet_peer_net *lpn;
+ unsigned int offset = 0;
+ unsigned int len = 0;
struct list_head *e;
lnet_shuffle_seed();
- list_for_each(e, &rnet->lrn_routes) {
+ list_for_each(e, &rnet->lrn_routes)
len++;
- }
- /* len+1 positions to add a new entry, also prevents division by 0 */
+ /*
+ * Randomly adding routes to the list is done to ensure that when
+ * different nodes are using the same list of routers, they end up
+ * preferring different routers.
+ */
offset = cfs_rand() % (len + 1);
list_for_each(e, &rnet->lrn_routes) {
if (offset == 0)
offset--;
}
list_add(&route->lr_list, e);
- list_add(&route->lr_gwlist, &route->lr_gateway->lpni_routes);
+ /*
+ * force a router check on the gateway to make sure the route is
+ * alive
+ */
+ list_for_each_entry(lpn, &route->lr_gateway->lp_peer_nets,
+ lpn_peer_nets) {
+ lpn->lpn_rtrcheck_timestamp = 0;
+ }
the_lnet.ln_remote_nets_version++;
+
+ /* add the route on the gateway list */
+ list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
+
+ /* take a router reference count on the gateway */
lnet_rtr_addref_locked(route->lr_gateway);
}
int
lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway,
- unsigned int priority)
+ __u32 priority, __u32 sensitivity)
{
- struct list_head *e;
- lnet_remotenet_t *rnet;
- lnet_remotenet_t *rnet2;
- lnet_route_t *route;
- lnet_ni_t *ni;
- struct lnet_peer_ni *lpni;
- int add_route;
- int rc;
-
- CDEBUG(D_NET, "Add route: net %s hops %d priority %u gw %s\n",
+ struct list_head *route_entry;
+ struct lnet_remotenet *rnet;
+ struct lnet_remotenet *rnet2;
+ struct lnet_route *route;
+ struct lnet_peer_ni *lpni;
+ struct lnet_peer *gw;
+ int add_route;
+ int rc;
+
+ CDEBUG(D_NET, "Add route: remote net %s hops %d priority %u gw %s\n",
libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
if (gateway == LNET_NID_ANY ||
(hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
return -EINVAL;
- if (lnet_islocalnet(net)) /* it's a local network */
+ /* it's a local network */
+ if (lnet_islocalnet(net))
return -EEXIST;
+ if (!lnet_islocalnet(LNET_NIDNET(gateway))) {
+ CERROR("Cannot add route with gateway %s. There is no local interface configured on LNet %s\n",
+ libcfs_nid2str(gateway),
+ libcfs_net2str(LNET_NIDNET(gateway)));
+ return -EINVAL;
+ }
+
/* Assume net, route, all new */
LIBCFS_ALLOC(route, sizeof(*route));
LIBCFS_ALLOC(rnet, sizeof(*rnet));
INIT_LIST_HEAD(&rnet->lrn_routes);
rnet->lrn_net = net;
- route->lr_hops = hops;
+ /* store the local and remote net that the route represents */
+ route->lr_lnet = LNET_NIDNET(gateway);
route->lr_net = net;
+ route->lr_nid = gateway;
route->lr_priority = priority;
+ route->lr_hops = hops;
lnet_net_lock(LNET_LOCK_EX);
+ /*
+ * lnet_nid2peerni_ex() grabs a ref on the lpni. We will need to
+ * lose that once we're done
+ */
lpni = lnet_nid2peerni_ex(gateway, LNET_LOCK_EX);
if (IS_ERR(lpni)) {
lnet_net_unlock(LNET_LOCK_EX);
LIBCFS_FREE(rnet, sizeof(*rnet));
rc = PTR_ERR(lpni);
- if (rc == -EHOSTUNREACH) /* gateway is not on a local net. */
- return rc; /* ignore the route entry */
CERROR("Error %d creating route %s %d %s\n", rc,
libcfs_net2str(net), hops,
libcfs_nid2str(gateway));
return rc;
}
- route->lr_gateway = lpni;
- LASSERT(!the_lnet.ln_shutdown);
+
+ LASSERT(lpni->lpni_peer_net && lpni->lpni_peer_net->lpn_peer);
+ gw = lpni->lpni_peer_net->lpn_peer;
+
+ route->lr_gateway = gw;
rnet2 = lnet_find_rnet_locked(net);
if (rnet2 == NULL) {
/* Search for a duplicate route (it's a NOOP if it is) */
add_route = 1;
- list_for_each(e, &rnet2->lrn_routes) {
- lnet_route_t *route2 = list_entry(e, lnet_route_t, lr_list);
+ list_for_each(route_entry, &rnet2->lrn_routes) {
+ struct lnet_route *route2;
+ route2 = list_entry(route_entry, struct lnet_route, lr_list);
if (route2->lr_gateway == route->lr_gateway) {
add_route = 0;
break;
}
/* our lookups must be true */
- LASSERT(route2->lr_gateway->lpni_nid != gateway);
+ LASSERT(route2->lr_gateway->lp_primary_nid != gateway);
}
+ /*
+ * It is possible to add multiple routes through the same peer,
+ * but it'll be using a different NID of that peer. When the
+ * gateway is discovered, discovery will consolidate the different
+ * peers into one peer. In this case the discovery code will have
+ * to move the routes from the peer that's being deleted to the
+ * consolidated peer lp_routes list
+ */
if (add_route) {
- lnet_peer_ni_addref_locked(route->lr_gateway); /* +1 for notify */
+ gw->lp_health_sensitivity = sensitivity;
lnet_add_route_to_rnet(rnet2, route);
-
- ni = lnet_get_next_ni_locked(route->lr_gateway->lpni_net, NULL);
- lnet_net_unlock(LNET_LOCK_EX);
-
- /* XXX Assume alive */
- if (ni->ni_net->net_lnd->lnd_notify != NULL)
- (ni->ni_net->net_lnd->lnd_notify)(ni, gateway, 1);
-
- lnet_net_lock(LNET_LOCK_EX);
}
- /* -1 for notify or !add_route */
- lnet_peer_ni_decref_locked(route->lr_gateway);
+ /*
+ * get rid of the reference on the lpni.
+ */
+ lnet_peer_ni_decref_locked(lpni);
lnet_net_unlock(LNET_LOCK_EX);
rc = 0;
if (rnet != rnet2)
LIBCFS_FREE(rnet, sizeof(*rnet));
- /* indicate to startup the router checker if configured */
- wake_up(&the_lnet.ln_rc_waitq);
+ /* kick start the monitor thread to handle the added route */
+ wake_up(&the_lnet.ln_mt_waitq);
return rc;
}
-int
-lnet_check_routes(void)
+static void
+lnet_del_route_from_rnet(lnet_nid_t gw_nid, struct list_head *route_list,
+ struct list_head *zombies)
{
- lnet_remotenet_t *rnet;
- lnet_route_t *route;
- lnet_route_t *route2;
- struct list_head *e1;
- struct list_head *e2;
- int cpt;
- struct list_head *rn_list;
- int i;
-
- cpt = lnet_net_lock_current();
-
- for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
- rn_list = &the_lnet.ln_remote_nets_hash[i];
- list_for_each(e1, rn_list) {
- rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
-
- route2 = NULL;
- list_for_each(e2, &rnet->lrn_routes) {
- lnet_nid_t nid1;
- lnet_nid_t nid2;
- int net;
-
- route = list_entry(e2, lnet_route_t,
- lr_list);
-
- if (route2 == NULL) {
- route2 = route;
- continue;
- }
-
- if (route->lr_gateway->lpni_net ==
- route2->lr_gateway->lpni_net)
- continue;
-
- nid1 = route->lr_gateway->lpni_nid;
- nid2 = route2->lr_gateway->lpni_nid;
- net = rnet->lrn_net;
+ struct lnet_peer *gateway;
+ struct lnet_route *route;
+ struct lnet_route *tmp;
+
+ list_for_each_entry_safe(route, tmp, route_list, lr_list) {
+ gateway = route->lr_gateway;
+ if (gw_nid != LNET_NID_ANY &&
+ gw_nid != gateway->lp_primary_nid)
+ continue;
- lnet_net_unlock(cpt);
+ /*
+ * move to zombie to delete outside the lock
+ * Note that this function is called with the
+ * ln_api_mutex held as well as the exclusive net
+ * lock. Adding to the remote net list happens
+ * under the same conditions. Same goes for the
+ * gateway router list
+ */
+ list_move(&route->lr_list, zombies);
+ the_lnet.ln_remote_nets_version++;
- CERROR("Routes to %s via %s and %s not "
- "supported\n",
- libcfs_net2str(net),
- libcfs_nid2str(nid1),
- libcfs_nid2str(nid2));
- return -EINVAL;
- }
- }
+ list_del(&route->lr_gwlist);
+ lnet_rtr_decref_locked(gateway);
}
-
- lnet_net_unlock(cpt);
- return 0;
}
int
lnet_del_route(__u32 net, lnet_nid_t gw_nid)
{
- struct lnet_peer_ni *gateway;
- lnet_remotenet_t *rnet;
- lnet_route_t *route;
- struct list_head *e1;
- struct list_head *e2;
- int rc = -ENOENT;
- struct list_head *rn_list;
- int idx = 0;
+ struct list_head rnet_zombies;
+ struct lnet_remotenet *rnet;
+ struct lnet_remotenet *tmp;
+ struct list_head *rn_list;
+ struct lnet_peer_ni *lpni;
+ struct lnet_route *route;
+ struct list_head zombies;
+ struct lnet_peer *lp;
+ int i = 0;
+
+ INIT_LIST_HEAD(&rnet_zombies);
+ INIT_LIST_HEAD(&zombies);
CDEBUG(D_NET, "Del route: net %s : gw %s\n",
libcfs_net2str(net), libcfs_nid2str(gw_nid));
* or a specific route entry actual NIDs) */
lnet_net_lock(LNET_LOCK_EX);
- if (net == LNET_NIDNET(LNET_NID_ANY))
- rn_list = &the_lnet.ln_remote_nets_hash[0];
- else
- rn_list = lnet_net2rnethash(net);
-
-again:
- list_for_each(e1, rn_list) {
- rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
-
- if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
- net == rnet->lrn_net))
- continue;
- list_for_each(e2, &rnet->lrn_routes) {
- route = list_entry(e2, lnet_route_t, lr_list);
+ lpni = lnet_find_peer_ni_locked(gw_nid);
+ if (lpni) {
+ lp = lpni->lpni_peer_net->lpn_peer;
+ LASSERT(lp);
+ gw_nid = lp->lp_primary_nid;
+ lnet_peer_ni_decref_locked(lpni);
+ }
- gateway = route->lr_gateway;
- if (!(gw_nid == LNET_NID_ANY ||
- gw_nid == gateway->lpni_nid))
- continue;
+ if (net != LNET_NIDNET(LNET_NID_ANY)) {
+ rnet = lnet_find_rnet_locked(net);
+ if (!rnet) {
+ lnet_net_unlock(LNET_LOCK_EX);
+ return -ENOENT;
+ }
+ lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes,
+ &zombies);
+ if (list_empty(&rnet->lrn_routes))
+ list_move(&rnet->lrn_list, &rnet_zombies);
+ goto delete_zombies;
+ }
- list_del(&route->lr_list);
- list_del(&route->lr_gwlist);
- the_lnet.ln_remote_nets_version++;
+ for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
+ rn_list = &the_lnet.ln_remote_nets_hash[i];
+ list_for_each_entry_safe(rnet, tmp, rn_list, lrn_list) {
+ lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes,
+ &zombies);
if (list_empty(&rnet->lrn_routes))
- list_del(&rnet->lrn_list);
- else
- rnet = NULL;
-
- lnet_rtr_decref_locked(gateway);
- lnet_peer_ni_decref_locked(gateway);
-
- lnet_net_unlock(LNET_LOCK_EX);
+ list_move(&rnet->lrn_list, &rnet_zombies);
+ }
+ }
- LIBCFS_FREE(route, sizeof(*route));
+delete_zombies:
+ /*
+ * check if there are any routes remaining on the gateway
+ * If there are no more routes make sure to set the peer's
+ * lp_disc_net_id to 0 (invalid), in case we add more routes in
+ * the future on that gateway, then we start our discovery process
+ * from scratch
+ */
+ if (lpni) {
+ if (list_empty(&lp->lp_routes))
+ lp->lp_disc_net_id = 0;
+ }
- if (rnet != NULL)
- LIBCFS_FREE(rnet, sizeof(*rnet));
+ lnet_net_unlock(LNET_LOCK_EX);
- rc = 0;
- lnet_net_lock(LNET_LOCK_EX);
- goto again;
- }
+ while (!list_empty(&zombies)) {
+ route = list_first_entry(&zombies, struct lnet_route, lr_list);
+ list_del(&route->lr_list);
+ LIBCFS_FREE(route, sizeof(*route));
}
- if (net == LNET_NIDNET(LNET_NID_ANY) &&
- ++idx < LNET_REMOTE_NETS_HASH_SIZE) {
- rn_list = &the_lnet.ln_remote_nets_hash[idx];
- goto again;
+ while (!list_empty(&rnet_zombies)) {
+ rnet = list_first_entry(&rnet_zombies, struct lnet_remotenet,
+ lrn_list);
+ list_del(&rnet->lrn_list);
+ LIBCFS_FREE(rnet, sizeof(*rnet));
}
- lnet_net_unlock(LNET_LOCK_EX);
- return rc;
+ return 0;
}
void
lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
}
-int lnet_get_rtr_pool_cfg(int idx, struct lnet_ioctl_pool_cfg *pool_cfg)
+int lnet_get_rtr_pool_cfg(int cpt, struct lnet_ioctl_pool_cfg *pool_cfg)
{
+ struct lnet_rtrbufpool *rbp;
int i, rc = -ENOENT, j;
if (the_lnet.ln_rtrpools == NULL)
return rc;
- for (i = 0; i < LNET_NRBPOOLS; i++) {
- lnet_rtrbufpool_t *rbp;
- lnet_net_lock(LNET_LOCK_EX);
- cfs_percpt_for_each(rbp, j, the_lnet.ln_rtrpools) {
- if (i++ != idx)
- continue;
+ cfs_percpt_for_each(rbp, i, the_lnet.ln_rtrpools) {
+ if (i != cpt)
+ continue;
- pool_cfg->pl_pools[i].pl_npages = rbp[i].rbp_npages;
- pool_cfg->pl_pools[i].pl_nbuffers = rbp[i].rbp_nbuffers;
- pool_cfg->pl_pools[i].pl_credits = rbp[i].rbp_credits;
- pool_cfg->pl_pools[i].pl_mincredits = rbp[i].rbp_mincredits;
- rc = 0;
- break;
+ lnet_net_lock(i);
+ for (j = 0; j < LNET_NRBPOOLS; j++) {
+ pool_cfg->pl_pools[j].pl_npages = rbp[j].rbp_npages;
+ pool_cfg->pl_pools[j].pl_nbuffers = rbp[j].rbp_nbuffers;
+ pool_cfg->pl_pools[j].pl_credits = rbp[j].rbp_credits;
+ pool_cfg->pl_pools[j].pl_mincredits = rbp[j].rbp_mincredits;
}
- lnet_net_unlock(LNET_LOCK_EX);
+ lnet_net_unlock(i);
+ rc = 0;
+ break;
}
lnet_net_lock(LNET_LOCK_EX);
int
lnet_get_route(int idx, __u32 *net, __u32 *hops,
- lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
+ lnet_nid_t *gateway, __u32 *alive, __u32 *priority, __u32 *sensitivity)
{
+ struct lnet_remotenet *rnet;
+ struct list_head *rn_list;
+ struct lnet_route *route;
struct list_head *e1;
struct list_head *e2;
- lnet_remotenet_t *rnet;
- lnet_route_t *route;
- int cpt;
- int i;
- struct list_head *rn_list;
+ int cpt;
+ int i;
cpt = lnet_net_lock_current();
for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
rn_list = &the_lnet.ln_remote_nets_hash[i];
list_for_each(e1, rn_list) {
- rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
+ rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
list_for_each(e2, &rnet->lrn_routes) {
- route = list_entry(e2, lnet_route_t,
+ route = list_entry(e2, struct lnet_route,
lr_list);
if (idx-- == 0) {
*net = rnet->lrn_net;
+ *gateway = route->lr_nid;
*hops = route->lr_hops;
*priority = route->lr_priority;
- *gateway = route->lr_gateway->lpni_nid;
+ *sensitivity = route->lr_gateway->
+ lp_health_sensitivity;
*alive = lnet_is_route_alive(route);
lnet_net_unlock(cpt);
return 0;
return -ENOENT;
}
-void
-lnet_swap_pinginfo(struct lnet_ping_info *info)
-{
- int i;
- struct lnet_ni_status *stat;
-
- __swab32s(&info->pi_magic);
- __swab32s(&info->pi_features);
- __swab32s(&info->pi_pid);
- __swab32s(&info->pi_nnis);
- for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
- stat = &info->pi_ni[i];
- __swab64s(&stat->ns_nid);
- __swab32s(&stat->ns_status);
- }
- return;
-}
-
-/**
- * parse router-checker pinginfo, record number of down NIs for remote
- * networks on that router.
- */
-static void
-lnet_parse_rc_info(lnet_rc_data_t *rcd)
-{
- struct lnet_ping_info *info = rcd->rcd_pinginfo;
- struct lnet_peer_ni *gw = rcd->rcd_gateway;
- lnet_route_t *rte;
-
- if (!gw->lpni_alive)
- return;
-
- /*
- * Protect gw->lpni_ping_feats. This can be set from
- * lnet_notify_locked with different locks being held
- */
- spin_lock(&gw->lpni_lock);
-
- if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
- lnet_swap_pinginfo(info);
-
- /* NB always racing with network! */
- if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
- CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
- libcfs_nid2str(gw->lpni_nid), info->pi_magic);
- gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
- spin_unlock(&gw->lpni_lock);
- return;
- }
-
- gw->lpni_ping_feats = info->pi_features;
- if ((gw->lpni_ping_feats & LNET_PING_FEAT_MASK) == 0) {
- CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
- libcfs_nid2str(gw->lpni_nid), gw->lpni_ping_feats);
- spin_unlock(&gw->lpni_lock);
- return; /* nothing I can understand */
- }
-
- if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0) {
- spin_unlock(&gw->lpni_lock);
- return; /* can't carry NI status info */
- }
-
- list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) {
- int down = 0;
- int up = 0;
- int i;
-
- if ((gw->lpni_ping_feats & LNET_PING_FEAT_RTE_DISABLED) != 0) {
- rte->lr_downis = 1;
- continue;
- }
-
- for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
- struct lnet_ni_status *stat = &info->pi_ni[i];
- lnet_nid_t nid = stat->ns_nid;
-
- if (nid == LNET_NID_ANY) {
- CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
- libcfs_nid2str(gw->lpni_nid));
- gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
- spin_unlock(&gw->lpni_lock);
- return;
- }
-
- if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
- continue;
-
- if (stat->ns_status == LNET_NI_STATUS_DOWN) {
- down++;
- continue;
- }
-
- if (stat->ns_status == LNET_NI_STATUS_UP) {
- if (LNET_NIDNET(nid) == rte->lr_net) {
- up = 1;
- break;
- }
- continue;
- }
-
- CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
- libcfs_nid2str(gw->lpni_nid), stat->ns_status);
- gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
- spin_unlock(&gw->lpni_lock);
- return;
- }
-
- if (up) { /* ignore downed NIs if NI for dest network is up */
- rte->lr_downis = 0;
- continue;
- }
- /* if @down is zero and this route is single-hop, it means
- * we can't find NI for target network */
- if (down == 0 && rte->lr_hops == 1)
- down = 1;
-
- rte->lr_downis = down;
- }
-
- spin_unlock(&gw->lpni_lock);
-}
-
-static void
-lnet_router_checker_event(lnet_event_t *event)
-{
- lnet_rc_data_t *rcd = event->md.user_ptr;
- struct lnet_peer_ni *lp;
-
- LASSERT(rcd != NULL);
-
- if (event->unlinked) {
- LNetInvalidateHandle(&rcd->rcd_mdh);
- return;
- }
-
- LASSERT(event->type == LNET_EVENT_SEND ||
- event->type == LNET_EVENT_REPLY);
-
- lp = rcd->rcd_gateway;
- LASSERT(lp != NULL);
-
- /* NB: it's called with holding lnet_res_lock, we have a few
- * places need to hold both locks at the same time, please take
- * care of lock ordering */
- lnet_net_lock(lp->lpni_cpt);
- if (!lnet_isrouter(lp) || lp->lpni_rcd != rcd) {
- /* ignore if no longer a router or rcd is replaced */
- goto out;
- }
-
- if (event->type == LNET_EVENT_SEND) {
- lp->lpni_ping_notsent = 0;
- if (event->status == 0)
- goto out;
- }
-
- /* LNET_EVENT_REPLY */
- /* A successful REPLY means the router is up. If _any_ comms
- * to the router fail I assume it's down (this will happen if
- * we ping alive routers to try to detect router death before
- * apps get burned). */
-
- lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
- /* The router checker will wake up very shortly and do the
- * actual notification.
- * XXX If 'lp' stops being a router before then, it will still
- * have the notification pending!!! */
-
- if (avoid_asym_router_failure && event->status == 0)
- lnet_parse_rc_info(rcd);
-
- out:
- lnet_net_unlock(lp->lpni_cpt);
-}
-
static void
lnet_wait_known_routerstate(void)
{
- struct lnet_peer_ni *rtr;
+ struct lnet_peer *rtr;
struct list_head *entry;
int all_known;
- LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
+ LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
for (;;) {
int cpt = lnet_net_lock_current();
all_known = 1;
list_for_each(entry, &the_lnet.ln_routers) {
- rtr = list_entry(entry, struct lnet_peer_ni,
- lpni_rtr_list);
+ rtr = list_entry(entry, struct lnet_peer,
+ lp_rtr_list);
- spin_lock(&rtr->lpni_lock);
+ spin_lock(&rtr->lp_lock);
- if (rtr->lpni_alive_count == 0) {
+ if ((rtr->lp_state & LNET_PEER_DISCOVERED) == 0) {
all_known = 0;
- spin_unlock(&rtr->lpni_lock);
+ spin_unlock(&rtr->lp_lock);
break;
}
- spin_unlock(&rtr->lpni_lock);
+ spin_unlock(&rtr->lp_lock);
}
lnet_net_unlock(cpt);
}
}
-void
-lnet_router_ni_update_locked(struct lnet_peer_ni *gw, __u32 net)
+static inline bool
+lnet_net_set_status_locked(struct lnet_net *net, __u32 status)
{
- lnet_route_t *rte;
+ struct lnet_ni *ni;
+ bool update = false;
- if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0) {
- list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) {
- if (rte->lr_net == net) {
- rte->lr_downis = 0;
- break;
- }
+ list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
+ lnet_ni_lock(ni);
+ if (ni->ni_status &&
+ ni->ni_status->ns_status != status) {
+ ni->ni_status->ns_status = status;
+ update = true;
}
+ lnet_ni_unlock(ni);
}
+
+ return update;
}
-static void
+static bool
lnet_update_ni_status_locked(void)
{
- lnet_ni_t *ni = NULL;
- time64_t now;
- int timeout;
+ struct lnet_net *net;
+ bool push = false;
+ time64_t now;
+ time64_t timeout;
LASSERT(the_lnet.ln_routing);
- timeout = router_ping_timeout +
- MAX(live_router_check_interval, dead_router_check_interval);
+ timeout = router_ping_timeout + alive_router_check_interval;
now = ktime_get_real_seconds();
- while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
- if (ni->ni_net->net_lnd->lnd_type == LOLND)
+ list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
+ if (net->net_lnd->lnd_type == LOLND)
continue;
- if (now < ni->ni_last_alive + timeout)
+ if (now < net->net_last_alive + timeout)
continue;
- lnet_ni_lock(ni);
+ spin_lock(&net->net_lock);
/* re-check with lock */
- if (now < ni->ni_last_alive + timeout) {
- lnet_ni_unlock(ni);
+ if (now < net->net_last_alive + timeout) {
+ spin_unlock(&net->net_lock);
continue;
}
+ spin_unlock(&net->net_lock);
- LASSERT(ni->ni_status != NULL);
-
- if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
- CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
- libcfs_nid2str(ni->ni_nid), timeout);
- /* NB: so far, this is the only place to set
- * NI status to "down" */
- ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
- }
- lnet_ni_unlock(ni);
- }
-}
-
-static void
-lnet_destroy_rc_data(lnet_rc_data_t *rcd)
-{
- LASSERT(list_empty(&rcd->rcd_list));
- /* detached from network */
- LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
-
- if (rcd->rcd_gateway != NULL) {
- int cpt = rcd->rcd_gateway->lpni_cpt;
-
- lnet_net_lock(cpt);
- lnet_peer_ni_decref_locked(rcd->rcd_gateway);
- lnet_net_unlock(cpt);
- }
-
- if (rcd->rcd_pinginfo != NULL)
- LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
-
- LIBCFS_FREE(rcd, sizeof(*rcd));
-}
-
-static lnet_rc_data_t *
-lnet_create_rc_data_locked(struct lnet_peer_ni *gateway)
-{
- lnet_rc_data_t *rcd = NULL;
- struct lnet_ping_info *pi;
- int rc;
- int i;
-
- lnet_net_unlock(gateway->lpni_cpt);
-
- LIBCFS_ALLOC(rcd, sizeof(*rcd));
- if (rcd == NULL)
- goto out;
-
- LNetInvalidateHandle(&rcd->rcd_mdh);
- INIT_LIST_HEAD(&rcd->rcd_list);
-
- LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
- if (pi == NULL)
- goto out;
-
- for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
- pi->pi_ni[i].ns_nid = LNET_NID_ANY;
- pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
- }
- rcd->rcd_pinginfo = pi;
-
- LASSERT(!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
- rc = LNetMDBind((lnet_md_t){.start = pi,
- .user_ptr = rcd,
- .length = LNET_PINGINFO_SIZE,
- .threshold = LNET_MD_THRESH_INF,
- .options = LNET_MD_TRUNCATE,
- .eq_handle = the_lnet.ln_rc_eqh},
- LNET_UNLINK,
- &rcd->rcd_mdh);
- if (rc < 0) {
- CERROR("Can't bind MD: %d\n", rc);
- goto out;
- }
- LASSERT(rc == 0);
-
- lnet_net_lock(gateway->lpni_cpt);
- /* router table changed or someone has created rcd for this gateway */
- if (!lnet_isrouter(gateway) || gateway->lpni_rcd != NULL) {
- lnet_net_unlock(gateway->lpni_cpt);
- goto out;
- }
-
- lnet_peer_ni_addref_locked(gateway);
- rcd->rcd_gateway = gateway;
- gateway->lpni_rcd = rcd;
- gateway->lpni_ping_notsent = 0;
-
- return rcd;
-
-out:
- if (rcd != NULL) {
- if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
- rc = LNetMDUnlink(rcd->rcd_mdh);
- LASSERT(rc == 0);
- }
- lnet_destroy_rc_data(rcd);
- }
-
- lnet_net_lock(gateway->lpni_cpt);
- return gateway->lpni_rcd;
-}
-
-static int
-lnet_router_check_interval (struct lnet_peer_ni *rtr)
-{
- int secs;
-
- secs = rtr->lpni_alive ? live_router_check_interval :
- dead_router_check_interval;
- if (secs < 0)
- secs = 0;
-
- return secs;
-}
-
-static void
-lnet_ping_router_locked (struct lnet_peer_ni *rtr)
-{
- lnet_rc_data_t *rcd = NULL;
- cfs_time_t now = cfs_time_current();
- int secs;
- struct lnet_ni *ni;
-
- lnet_peer_ni_addref_locked(rtr);
-
- if (rtr->lpni_ping_deadline != 0 && /* ping timed out? */
- cfs_time_after(now, rtr->lpni_ping_deadline))
- lnet_notify_locked(rtr, 1, 0, now);
-
- /* Run any outstanding notifications */
- ni = lnet_get_next_ni_locked(rtr->lpni_net, NULL);
- lnet_ni_notify_locked(ni, rtr);
-
- if (!lnet_isrouter(rtr) ||
- the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
- /* router table changed or router checker is shutting down */
- lnet_peer_ni_decref_locked(rtr);
- return;
+ /*
+ * if the net didn't receive any traffic for past the
+ * timeout on any of its constituent NIs, then mark all
+ * the NIs down.
+ */
+ push = lnet_net_set_status_locked(net, LNET_NI_STATUS_DOWN);
}
- rcd = rtr->lpni_rcd != NULL ?
- rtr->lpni_rcd : lnet_create_rc_data_locked(rtr);
-
- if (rcd == NULL)
- return;
-
- secs = lnet_router_check_interval(rtr);
-
- CDEBUG(D_NET,
- "rtr %s %d: deadline %lu ping_notsent %d alive %d "
- "alive_count %d lpni_ping_timestamp %lu\n",
- libcfs_nid2str(rtr->lpni_nid), secs,
- rtr->lpni_ping_deadline, rtr->lpni_ping_notsent,
- rtr->lpni_alive, rtr->lpni_alive_count, rtr->lpni_ping_timestamp);
-
- if (secs != 0 && !rtr->lpni_ping_notsent &&
- cfs_time_after(now, cfs_time_add(rtr->lpni_ping_timestamp,
- cfs_time_seconds(secs)))) {
- int rc;
- lnet_process_id_t id;
- lnet_handle_md_t mdh;
-
- id.nid = rtr->lpni_nid;
- id.pid = LNET_PID_LUSTRE;
- CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
-
- rtr->lpni_ping_notsent = 1;
- rtr->lpni_ping_timestamp = now;
-
- mdh = rcd->rcd_mdh;
-
- if (rtr->lpni_ping_deadline == 0) {
- rtr->lpni_ping_deadline =
- cfs_time_shift(router_ping_timeout);
- }
-
- lnet_net_unlock(rtr->lpni_cpt);
-
- rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
- LNET_PROTO_PING_MATCHBITS, 0);
-
- lnet_net_lock(rtr->lpni_cpt);
- if (rc != 0)
- rtr->lpni_ping_notsent = 0; /* no event pending */
- }
-
- lnet_peer_ni_decref_locked(rtr);
- return;
+ return push;
}
-int
-lnet_router_checker_start(void)
+void lnet_wait_router_start(void)
{
- int rc;
- int eqsz = 0;
- struct task_struct *task;
-
- LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
-
- if (check_routers_before_use &&
- dead_router_check_interval <= 0) {
- LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
- " set if 'check_routers_before_use' is set"
- "\n");
- return -EINVAL;
- }
-
- sema_init(&the_lnet.ln_rc_signal, 0);
-
- rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
- if (rc != 0) {
- CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
- return -ENOMEM;
- }
-
- the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
- task = kthread_run(lnet_router_checker, NULL, "router_checker");
- if (IS_ERR(task)) {
- rc = PTR_ERR(task);
- CERROR("Can't start router checker thread: %d\n", rc);
- /* block until event callback signals exit */
- down(&the_lnet.ln_rc_signal);
- rc = LNetEQFree(the_lnet.ln_rc_eqh);
- LASSERT(rc == 0);
- the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
- return -ENOMEM;
- }
-
if (check_routers_before_use) {
/* Note that a helpful side-effect of pinging all known routers
* at startup is that it makes them drop stale connections they
* may have to a previous instance of me. */
lnet_wait_known_routerstate();
}
-
- return 0;
-}
-
-void
-lnet_router_checker_stop (void)
-{
- int rc;
-
- if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
- return;
-
- LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
- the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
- /* wakeup the RC thread if it's sleeping */
- wake_up(&the_lnet.ln_rc_waitq);
-
- /* block until event callback signals exit */
- down(&the_lnet.ln_rc_signal);
- LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
-
- rc = LNetEQFree(the_lnet.ln_rc_eqh);
- LASSERT(rc == 0);
- return;
-}
-
-static void
-lnet_prune_rc_data(int wait_unlink)
-{
- lnet_rc_data_t *rcd;
- lnet_rc_data_t *tmp;
- struct lnet_peer_ni *lp;
- struct list_head head;
- int i = 2;
-
- if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
- list_empty(&the_lnet.ln_rcd_deathrow) &&
- list_empty(&the_lnet.ln_rcd_zombie)))
- return;
-
- INIT_LIST_HEAD(&head);
-
- lnet_net_lock(LNET_LOCK_EX);
-
- if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
- /* router checker is stopping, prune all */
- list_for_each_entry(lp, &the_lnet.ln_routers,
- lpni_rtr_list) {
- if (lp->lpni_rcd == NULL)
- continue;
-
- LASSERT(list_empty(&lp->lpni_rcd->rcd_list));
- list_add(&lp->lpni_rcd->rcd_list,
- &the_lnet.ln_rcd_deathrow);
- lp->lpni_rcd = NULL;
- }
- }
-
- /* unlink all RCDs on deathrow list */
- list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
-
- if (!list_empty(&head)) {
- lnet_net_unlock(LNET_LOCK_EX);
-
- list_for_each_entry(rcd, &head, rcd_list)
- LNetMDUnlink(rcd->rcd_mdh);
-
- lnet_net_lock(LNET_LOCK_EX);
- }
-
- list_splice_init(&head, &the_lnet.ln_rcd_zombie);
-
- /* release all zombie RCDs */
- while (!list_empty(&the_lnet.ln_rcd_zombie)) {
- list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
- rcd_list) {
- if (LNetHandleIsInvalid(rcd->rcd_mdh))
- list_move(&rcd->rcd_list, &head);
- }
-
- wait_unlink = wait_unlink &&
- !list_empty(&the_lnet.ln_rcd_zombie);
-
- lnet_net_unlock(LNET_LOCK_EX);
-
- while (!list_empty(&head)) {
- rcd = list_entry(head.next,
- lnet_rc_data_t, rcd_list);
- list_del_init(&rcd->rcd_list);
- lnet_destroy_rc_data(rcd);
- }
-
- if (!wait_unlink)
- return;
-
- i++;
- CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
- "Waiting for rc buffers to unlink\n");
- set_current_state(TASK_UNINTERRUPTIBLE);
- schedule_timeout(cfs_time_seconds(1) / 4);
-
- lnet_net_lock(LNET_LOCK_EX);
- }
-
- lnet_net_unlock(LNET_LOCK_EX);
}
/*
- * This function is called to check if the RC should block indefinitely.
- * It's called from lnet_router_checker() as well as being passed to
- * wait_event_interruptible() to avoid the lost wake_up problem.
- *
- * When it's called from wait_event_interruptible() it is necessary to
- * also not sleep if the rc state is not running to avoid a deadlock
- * when the system is shutting down
+ * This function is called from the monitor thread to check if there are
+ * any active routers that need to be checked.
*/
-static inline bool
+inline bool
lnet_router_checker_active(void)
{
- if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING)
- return true;
-
/* Router Checker thread needs to run when routing is enabled in
* order to call lnet_update_ni_status_locked() */
if (the_lnet.ln_routing)
return true;
return !list_empty(&the_lnet.ln_routers) &&
- (live_router_check_interval > 0 ||
- dead_router_check_interval > 0);
+ alive_router_check_interval > 0;
}
-static int
-lnet_router_checker(void *arg)
+void
+lnet_check_routers(void)
{
- struct lnet_peer_ni *rtr;
+ struct lnet_peer_net *first_lpn = NULL;
+ struct lnet_peer_net *lpn;
+ struct lnet_peer_ni *lpni;
struct list_head *entry;
+ struct lnet_peer *rtr;
+ bool push = false;
+ bool found_lpn;
+ __u64 version;
+ __u32 net_id;
+ time64_t now;
+ int cpt;
+ int rc;
- cfs_block_allsigs();
-
- while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
- __u64 version;
- int cpt;
- int cpt2;
-
- cpt = lnet_net_lock_current();
+ cpt = lnet_net_lock_current();
rescan:
- version = the_lnet.ln_routers_version;
-
- list_for_each(entry, &the_lnet.ln_routers) {
- rtr = list_entry(entry, struct lnet_peer_ni,
- lpni_rtr_list);
-
- cpt2 = rtr->lpni_cpt;
- if (cpt != cpt2) {
- lnet_net_unlock(cpt);
- cpt = cpt2;
- lnet_net_lock(cpt);
- /* the routers list has changed */
- if (version != the_lnet.ln_routers_version)
- goto rescan;
+ version = the_lnet.ln_routers_version;
+
+ list_for_each(entry, &the_lnet.ln_routers) {
+ rtr = list_entry(entry, struct lnet_peer,
+ lp_rtr_list);
+
+ now = ktime_get_real_seconds();
+
+ /*
+ * only discover the router if we've passed
+ * alive_router_check_interval seconds. Some of the router
+ * interfaces could be down and in that case they would be
+ * undergoing recovery separately from this discovery.
+ */
+ /* find next peer net which is also local */
+ net_id = rtr->lp_disc_net_id;
+ do {
+ lpn = lnet_get_next_peer_net_locked(rtr, net_id);
+ if (!lpn) {
+ CERROR("gateway %s has no networks\n",
+ libcfs_nid2str(rtr->lp_primary_nid));
+ break;
}
+ if (first_lpn == lpn)
+ break;
+ if (!first_lpn)
+ first_lpn = lpn;
+ found_lpn = lnet_islocalnet_locked(lpn->lpn_net_id);
+ net_id = lpn->lpn_net_id;
+ } while (!found_lpn);
+
+ if (!found_lpn || !lpn) {
+ CERROR("no local network found for gateway %s\n",
+ libcfs_nid2str(rtr->lp_primary_nid));
+ continue;
+ }
- lnet_ping_router_locked(rtr);
-
- /* NB dropped lock */
- if (version != the_lnet.ln_routers_version) {
- /* the routers list has changed */
- goto rescan;
- }
+ if (now - lpn->lpn_rtrcheck_timestamp <
+ alive_router_check_interval / lnet_current_net_count)
+ continue;
+
+ /*
+ * If we're currently discovering the peer then don't
+ * issue another discovery
+ */
+ spin_lock(&rtr->lp_lock);
+ if (rtr->lp_state & LNET_PEER_RTR_DISCOVERY) {
+ spin_unlock(&rtr->lp_lock);
+ continue;
+ }
+ /* make sure we actively discover the router */
+ rtr->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
+ rtr->lp_state |= LNET_PEER_RTR_DISCOVERY;
+ spin_unlock(&rtr->lp_lock);
+
+ /* find the peer_ni associated with the primary NID */
+ lpni = lnet_peer_get_ni_locked(rtr, rtr->lp_primary_nid);
+ if (!lpni) {
+ CDEBUG(D_NET, "Expected to find an lpni for %s, but non found\n",
+ libcfs_nid2str(rtr->lp_primary_nid));
+ continue;
}
+ lnet_peer_ni_addref_locked(lpni);
- if (the_lnet.ln_routing)
- lnet_update_ni_status_locked();
+ /* specify the net to use */
+ rtr->lp_disc_net_id = lpn->lpn_net_id;
- lnet_net_unlock(cpt);
+ /* discover the router */
+ CDEBUG(D_NET, "discover %s, cpt = %d\n",
+ libcfs_nid2str(lpni->lpni_nid), cpt);
+ rc = lnet_discover_peer_locked(lpni, cpt, false);
+
+ /* decrement ref count acquired by find_peer_ni_locked() */
+ lnet_peer_ni_decref_locked(lpni);
- lnet_prune_rc_data(0); /* don't wait for UNLINK */
-
- /* Call schedule_timeout() here always adds 1 to load average
- * because kernel counts # active tasks as nr_running
- * + nr_uninterruptible. */
- /* if there are any routes then wakeup every second. If
- * there are no routes then sleep indefinitely until woken
- * up by a user adding a route */
- if (!lnet_router_checker_active())
- wait_event_interruptible(the_lnet.ln_rc_waitq,
- lnet_router_checker_active());
+ if (!rc)
+ lpn->lpn_rtrcheck_timestamp = now;
else
- wait_event_interruptible_timeout(the_lnet.ln_rc_waitq,
- false,
- cfs_time_seconds(1));
+ CERROR("Failed to discover router %s\n",
+ libcfs_nid2str(rtr->lp_primary_nid));
+
+ /* NB dropped lock */
+ if (version != the_lnet.ln_routers_version) {
+ /* the routers list has changed */
+ goto rescan;
+ }
}
- lnet_prune_rc_data(1); /* wait for UNLINK */
+ if (the_lnet.ln_routing)
+ push = lnet_update_ni_status_locked();
- the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
- up(&the_lnet.ln_rc_signal);
- /* The unlink event callback will signal final completion */
- return 0;
+ lnet_net_unlock(cpt);
+
+ /* if the status of the ni changed update the peers */
+ if (push)
+ lnet_push_update_to_peers(1);
}
void
-lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
+lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
{
- int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
+ int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
while (--npages >= 0)
__free_page(rb->rb_kiov[npages].kiov_page);
LIBCFS_FREE(rb, sz);
}
-static lnet_rtrbuf_t *
-lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
+static struct lnet_rtrbuf *
+lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
{
int npages = rbp->rbp_npages;
- int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
+ int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
struct page *page;
- lnet_rtrbuf_t *rb;
+ struct lnet_rtrbuf *rb;
int i;
LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
}
static void
-lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp, int cpt)
+lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
{
- int npages = rbp->rbp_npages;
- lnet_rtrbuf_t *rb;
+ int npages = rbp->rbp_npages;
+ struct lnet_rtrbuf *rb;
struct list_head tmp;
if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
/* Free buffers on the free list. */
while (!list_empty(&tmp)) {
- rb = list_entry(tmp.next, lnet_rtrbuf_t, rb_list);
+ rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
list_del(&rb->rb_list);
lnet_destroy_rtrbuf(rb, npages);
}
}
static int
-lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
+lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
{
struct list_head rb_list;
- lnet_rtrbuf_t *rb;
+ struct lnet_rtrbuf *rb;
int num_rb;
int num_buffers = 0;
int old_req_nbufs;
failed:
while (!list_empty(&rb_list)) {
- rb = list_entry(rb_list.next, lnet_rtrbuf_t, rb_list);
+ rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
list_del(&rb->rb_list);
lnet_destroy_rtrbuf(rb, npages);
}
}
static void
-lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
+lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
{
INIT_LIST_HEAD(&rbp->rbp_msgs);
INIT_LIST_HEAD(&rbp->rbp_bufs);
void
lnet_rtrpools_free(int keep_pools)
{
- lnet_rtrbufpool_t *rtrp;
+ struct lnet_rtrbufpool *rtrp;
int i;
if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
int
lnet_rtrpools_alloc(int im_a_router)
{
- lnet_rtrbufpool_t *rtrp;
+ struct lnet_rtrbufpool *rtrp;
int nrb_tiny;
int nrb_small;
int nrb_large;
the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
LNET_NRBPOOLS *
- sizeof(lnet_rtrbufpool_t));
+ sizeof(struct lnet_rtrbufpool));
if (the_lnet.ln_rtrpools == NULL) {
LCONSOLE_ERROR_MSG(0x10c,
"Failed to initialize router buffe pool\n");
lnet_net_lock(LNET_LOCK_EX);
the_lnet.ln_routing = 1;
lnet_net_unlock(LNET_LOCK_EX);
+ wake_up(&the_lnet.ln_mt_waitq);
return 0;
failed:
int nrb = 0;
int rc = 0;
int i;
- lnet_rtrbufpool_t *rtrp;
+ struct lnet_rtrbufpool *rtrp;
/* If the provided values for each buffer pool are different than the
* configured values, we need to take action. */
lnet_net_lock(LNET_LOCK_EX);
the_lnet.ln_routing = 1;
- the_lnet.ln_ping_info->pi_features &= ~LNET_PING_FEAT_RTE_DISABLED;
+ the_lnet.ln_ping_target->pb_info.pi_features &=
+ ~LNET_PING_FEAT_RTE_DISABLED;
lnet_net_unlock(LNET_LOCK_EX);
return rc;
lnet_net_lock(LNET_LOCK_EX);
the_lnet.ln_routing = 0;
- the_lnet.ln_ping_info->pi_features |= LNET_PING_FEAT_RTE_DISABLED;
+ the_lnet.ln_ping_target->pb_info.pi_features |=
+ LNET_PING_FEAT_RTE_DISABLED;
tiny_router_buffers = 0;
small_router_buffers = 0;
lnet_rtrpools_free(1);
}
+static inline void
+lnet_notify_peer_down(struct lnet_ni *ni, lnet_nid_t nid)
+{
+ if (ni->ni_net->net_lnd->lnd_notify_peer_down != NULL)
+ (ni->ni_net->net_lnd->lnd_notify_peer_down)(nid);
+}
+
+/*
+ * ni: local NI used to communicate with the peer
+ * nid: peer NID
+ * alive: true if peer is alive, false otherwise
+ * reset: reset health value. This is requested by the LND.
+ * when: notificaiton time.
+ */
int
-lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
+lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, bool alive, bool reset,
+ time64_t when)
{
- struct lnet_peer_ni *lp = NULL;
- cfs_time_t now = cfs_time_current();
- int cpt = lnet_cpt_of_nid(nid, ni);
+ struct lnet_peer_ni *lpni = NULL;
+ time64_t now = ktime_get_seconds();
+ int cpt;
LASSERT (!in_interrupt ());
}
/* can't do predictions... */
- if (cfs_time_after(when, now)) {
+ if (when > now) {
CWARN("Ignoring prediction from %s of %s %s "
- "%ld seconds in the future\n",
+ "%lld seconds in the future\n",
(ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
- libcfs_nid2str(nid), alive ? "up" : "down",
- cfs_duration_sec(cfs_time_sub(when, now)));
+ libcfs_nid2str(nid), alive ? "up" : "down", when - now);
return -EINVAL;
}
return 0;
}
- lnet_net_lock(cpt);
+ /* must lock 0 since this is used for synchronization */
+ lnet_net_lock(0);
- if (the_lnet.ln_shutdown) {
- lnet_net_unlock(cpt);
+ if (the_lnet.ln_state != LNET_STATE_RUNNING) {
+ lnet_net_unlock(0);
return -ESHUTDOWN;
}
- lp = lnet_find_peer_ni_locked(nid);
- if (lp == NULL) {
+ lpni = lnet_find_peer_ni_locked(nid);
+ if (lpni == NULL) {
/* nid not found */
- lnet_net_unlock(cpt);
+ lnet_net_unlock(0);
CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
return 0;
}
- /*
- * It is possible for this function to be called for the same peer
- * but with different NIs. We want to synchronize the notification
- * between the different calls. So we will use the lpni_cpt to
- * grab the net lock.
- */
- if (lp->lpni_cpt != cpt) {
- lnet_net_unlock(cpt);
- cpt = lp->lpni_cpt;
- lnet_net_lock(cpt);
+ if (alive) {
+ if (reset)
+ lnet_set_healthv(&lpni->lpni_healthv,
+ LNET_MAX_HEALTH_VALUE);
+ else
+ lnet_inc_healthv(&lpni->lpni_healthv);
+ } else {
+ lnet_handle_remote_failure_locked(lpni);
}
- /* We can't fully trust LND on reporting exact peer last_alive
- * if he notifies us about dead peer. For example ksocklnd can
- * call us with when == _time_when_the_node_was_booted_ if
- * no connections were successfully established */
- if (ni != NULL && !alive && when < lp->lpni_last_alive)
- when = lp->lpni_last_alive;
-
- lnet_notify_locked(lp, ni == NULL, alive, when);
+ /* recalculate aliveness */
+ alive = lnet_is_peer_ni_alive(lpni);
+ lnet_net_unlock(0);
- if (ni != NULL)
- lnet_ni_notify_locked(ni, lp);
-
- lnet_peer_ni_decref_locked(lp);
+ if (ni != NULL && !alive)
+ lnet_notify_peer_down(ni, lpni->lpni_nid);
+ cpt = lpni->lpni_cpt;
+ lnet_net_lock(cpt);
+ lnet_peer_ni_decref_locked(lpni);
lnet_net_unlock(cpt);
+
return 0;
}
EXPORT_SYMBOL(lnet_notify);