-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
*
- * This file is part of Portals
- * http://sourceforge.net/projects/sandiaportals/
+ * Copyright (c) 2011, 2017, Intel Corporation.
+ *
+ * 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>
-#if defined(__KERNEL__) && defined(LNET_ROUTER)
+#define LNET_NRB_TINY_MIN 512 /* min value for each CPT */
+#define LNET_NRB_TINY (LNET_NRB_TINY_MIN * 4)
+#define LNET_NRB_SMALL_MIN 4096 /* min value for each CPT */
+#define LNET_NRB_SMALL (LNET_NRB_SMALL_MIN * 4)
+#define LNET_NRB_SMALL_PAGES 1
+#define LNET_NRB_LARGE_MIN 256 /* min value for each CPT */
+#define LNET_NRB_LARGE (LNET_NRB_LARGE_MIN * 4)
+#define LNET_NRB_LARGE_PAGES ((LNET_MTU + PAGE_SIZE - 1) >> \
+ PAGE_SHIFT)
static char *forwarding = "";
-CFS_MODULE_PARM(forwarding, "s", charp, 0444,
- "Explicitly enable/disable forwarding between networks");
-
-static int tiny_router_buffers = 1024;
-CFS_MODULE_PARM(tiny_router_buffers, "i", int, 0444,
- "# of 0 payload messages to buffer in the router");
-static int small_router_buffers = 8192;
-CFS_MODULE_PARM(small_router_buffers, "i", int, 0444,
- "# of small (1 page) messages to buffer in the router");
-static int large_router_buffers = 512;
-CFS_MODULE_PARM(large_router_buffers, "i", int, 0444,
- "# of large messages to buffer in the router");
-static int peer_buffer_credits = 0;
-CFS_MODULE_PARM(peer_buffer_credits, "i", int, 0444,
- "# router buffer credits per peer");
+module_param(forwarding, charp, 0444);
+MODULE_PARM_DESC(forwarding, "Explicitly enable/disable forwarding between networks");
+
+static int tiny_router_buffers;
+module_param(tiny_router_buffers, int, 0444);
+MODULE_PARM_DESC(tiny_router_buffers, "# of 0 payload messages to buffer in the router");
+static int small_router_buffers;
+module_param(small_router_buffers, int, 0444);
+MODULE_PARM_DESC(small_router_buffers, "# of small (1 page) messages to buffer in the router");
+static int large_router_buffers;
+module_param(large_router_buffers, int, 0444);
+MODULE_PARM_DESC(large_router_buffers, "# of large messages to buffer in the router");
+static int peer_buffer_credits;
+module_param(peer_buffer_credits, int, 0444);
+MODULE_PARM_DESC(peer_buffer_credits, "# router buffer credits per peer");
static int auto_down = 1;
-CFS_MODULE_PARM(auto_down, "i", int, 0444,
- "Automatically mark peers down on comms error");
-
-int
-lnet_peer_buffer_credits(lnet_ni_t *ni)
-{
- /* NI option overrides LNet default */
- if (ni->ni_peerrtrcredits > 0)
- return ni->ni_peerrtrcredits;
- if (peer_buffer_credits > 0)
- return peer_buffer_credits;
-
- /* As an approximation, allow this peer the same number of router
- * buffers as it is allowed outstanding sends */
- return ni->ni_peertxcredits;
-}
-
-/* forward ref's */
-static int lnet_router_checker(void *);
-#else
+module_param(auto_down, int, 0444);
+MODULE_PARM_DESC(auto_down, "Automatically mark peers down on comms error");
int
-lnet_peer_buffer_credits(lnet_ni_t *ni)
+lnet_peer_buffer_credits(struct lnet_net *net)
{
- return 0;
+ /* NI option overrides LNet default */
+ if (net->net_tunables.lct_peer_rtr_credits > 0)
+ return net->net_tunables.lct_peer_rtr_credits;
+ if (peer_buffer_credits > 0)
+ return peer_buffer_credits;
+
+ /* As an approximation, allow this peer the same number of router
+ * buffers as it is allowed outstanding sends */
+ return net->net_tunables.lct_peer_tx_credits;
}
-#endif
+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");
-static int check_routers_before_use = 0;
-CFS_MODULE_PARM(check_routers_before_use, "i", int, 0444,
- "Assume routers are down and ping them before use");
+int avoid_asym_router_failure = 1;
+module_param(avoid_asym_router_failure, int, 0644);
+MODULE_PARM_DESC(avoid_asym_router_failure, "Avoid asymmetrical router failures (0 to disable)");
-static int avoid_asym_router_failure = 0;
-CFS_MODULE_PARM(avoid_asym_router_failure, "i", int, 0444,
- "Avoid asymmetrical failures: reserved, use at your own risk");
+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 dead_router_check_interval = 0;
-CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0444,
- "Seconds between dead router health checks (<= 0 to disable)");
-
-static int live_router_check_interval = 0;
-CFS_MODULE_PARM(live_router_check_interval, "i", int, 0444,
- "Seconds between live 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)");
static int router_ping_timeout = 50;
-CFS_MODULE_PARM(router_ping_timeout, "i", int, 0444,
- "Seconds to wait for the reply to a router health query");
+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;
+ return check_routers_before_use;
}
void
-lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, cfs_time_t when)
+lnet_notify_locked(struct lnet_peer_ni *lp, int notifylnd, int alive,
+ time64_t when)
{
- if (cfs_time_before(when, lp->lp_timestamp)) { /* out of date information */
- CDEBUG(D_NET, "Out of date\n");
- return;
- }
-
- lp->lp_timestamp = when; /* update timestamp */
- lp->lp_ping_deadline = 0; /* disable ping timeout */
-
- if (lp->lp_alive_count != 0 && /* got old news */
- (!lp->lp_alive) == (!alive)) { /* new date for old news */
- CDEBUG(D_NET, "Old news\n");
- return;
- }
-
- /* Flag that notification is outstanding */
-
- lp->lp_alive_count++;
- lp->lp_alive = !(!alive); /* 1 bit! */
- lp->lp_notify = 1;
- lp->lp_notifylnd |= notifylnd;
-
- CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
+ if (lp->lpni_timestamp > when) { /* out of date information */
+ CDEBUG(D_NET, "Out of date\n");
+ return;
+ }
+
+ /*
+ * 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
+ */
+ spin_lock(&lp->lpni_lock);
+
+ lp->lpni_timestamp = when; /* update timestamp */
+
+ 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;
+ }
+
+ /* Flag that notification is outstanding */
+
+ 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 */
+
+ spin_unlock(&lp->lpni_lock);
+
+ CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lpni_nid), alive);
}
-void
-lnet_do_notify (lnet_peer_t *lp)
+/*
+ * This function will always be called with lp->lpni_cpt lock held.
+ */
+static void
+lnet_ni_notify_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
{
- lnet_ni_t *ni = lp->lp_ni;
- int alive;
- int notifylnd;
-
- LNET_LOCK();
+ int alive;
+ int notifylnd;
- /* 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 */
+ /* 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 */
- if (lp->lp_notifying) {
- LNET_UNLOCK();
- return;
- }
+ spin_lock(&lp->lpni_lock);
- lp->lp_notifying = 1;
+ if (lp->lpni_notifying || ni == NULL) {
+ spin_unlock(&lp->lpni_lock);
+ return;
+ }
- while (lp->lp_notify) {
- alive = lp->lp_alive;
- notifylnd = lp->lp_notifylnd;
+ lp->lpni_notifying = 1;
- lp->lp_notifylnd = 0;
- lp->lp_notify = 0;
+ /*
+ * lp->lpni_notify needs to be protected because it can be set in
+ * lnet_notify_locked().
+ */
+ while (lp->lpni_notify) {
+ alive = lp->lpni_alive;
+ notifylnd = lp->lpni_notifylnd;
- if (notifylnd && ni->ni_lnd->lnd_notify != NULL) {
- LNET_UNLOCK();
+ lp->lpni_notifylnd = 0;
+ lp->lpni_notify = 0;
- /* A new notification could happen now; I'll handle it
- * when control returns to me */
+ if (notifylnd && ni->ni_net->net_lnd->lnd_notify != NULL) {
+ spin_unlock(&lp->lpni_lock);
+ lnet_net_unlock(lp->lpni_cpt);
- (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive);
+ /* A new notification could happen now; I'll handle it
+ * when control returns to me */
- LNET_LOCK();
- }
- }
+ (ni->ni_net->net_lnd->lnd_notify)(ni, lp->lpni_nid,
+ alive);
- lp->lp_notifying = 0;
+ lnet_net_lock(lp->lpni_cpt);
+ spin_lock(&lp->lpni_lock);
+ }
+ }
- LNET_UNLOCK();
+ lp->lpni_notifying = 0;
+ spin_unlock(&lp->lpni_lock);
}
-
static void
-lnet_rtr_addref_locked(lnet_peer_t *lp)
+lnet_rtr_addref_locked(struct lnet_peer *lp)
{
- LASSERT (lp->lp_refcount > 0);
- LASSERT (lp->lp_rtr_refcount >= 0);
-
- lp->lp_rtr_refcount++;
- if (lp->lp_rtr_refcount == 1) {
- cfs_list_t *pos;
-
- /* a simple insertion sort */
- cfs_list_for_each_prev(pos, &the_lnet.ln_routers) {
- lnet_peer_t *rtr = cfs_list_entry(pos, lnet_peer_t,
- lp_rtr_list);
-
- if (rtr->lp_nid < lp->lp_nid)
- break;
- }
-
- cfs_list_add(&lp->lp_rtr_list, pos);
- /* addref for the_lnet.ln_routers */
- lnet_peer_addref_locked(lp);
- the_lnet.ln_routers_version++;
- }
+ LASSERT(lp->lp_rtr_refcount >= 0);
+
+ /* 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_addref_locked(lp);
+ the_lnet.ln_routers_version++;
+ }
}
static void
-lnet_rtr_decref_locked(lnet_peer_t *lp)
+lnet_rtr_decref_locked(struct lnet_peer *lp)
{
- LASSERT (lp->lp_refcount > 0);
- LASSERT (lp->lp_rtr_refcount > 0);
-
- lp->lp_rtr_refcount--;
- if (lp->lp_rtr_refcount == 0) {
- if (lp->lp_rcd != NULL) {
- cfs_list_add(&lp->lp_rcd->rcd_list,
- &the_lnet.ln_zombie_rcd);
- lp->lp_rcd = NULL;
- }
-
- cfs_list_del(&lp->lp_rtr_list);
- /* decref for the_lnet.ln_routers */
- lnet_peer_decref_locked(lp);
- the_lnet.ln_routers_version++;
- }
+ LASSERT(atomic_read(&lp->lp_refcount) > 0);
+ LASSERT(lp->lp_rtr_refcount > 0);
+
+ /* lnet_net_lock must be exclusively locked */
+ lp->lp_rtr_refcount--;
+ if (lp->lp_rtr_refcount == 0) {
+ LASSERT(list_empty(&lp->lp_routes));
+
+ list_del(&lp->lp_rtr_list);
+ /* decref for the_lnet.ln_routers */
+ lnet_peer_decref_locked(lp);
+ the_lnet.ln_routers_version++;
+ }
}
-lnet_remotenet_t *
-lnet_find_net_locked (__u32 net)
+struct lnet_remotenet *
+lnet_find_rnet_locked(__u32 net)
{
- lnet_remotenet_t *rnet;
- cfs_list_t *tmp;
+ 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);
- cfs_list_for_each (tmp, &the_lnet.ln_remote_nets) {
- rnet = cfs_list_entry(tmp, lnet_remotenet_t, lrn_list);
+ rn_list = lnet_net2rnethash(net);
+ list_for_each(tmp, rn_list) {
+ rnet = list_entry(tmp, struct lnet_remotenet, lrn_list);
- if (rnet->lrn_net == net)
- return rnet;
- }
- return NULL;
+ if (rnet->lrn_net == net)
+ return rnet;
+ }
+ return NULL;
+}
+
+static void lnet_shuffle_seed(void)
+{
+ static int seeded;
+ struct lnet_ni *ni = NULL;
+
+ if (seeded)
+ return;
+
+ /* 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)))
+ add_device_randomness(&ni->ni_nid, sizeof(ni->ni_nid));
+
+ seeded = 1;
+ return;
}
/* NB expects LNET_LOCK held */
-void
-lnet_add_route_to_rnet (lnet_remotenet_t *rnet, lnet_route_t *route)
+static void
+lnet_add_route_to_rnet(struct lnet_remotenet *rnet, struct lnet_route *route)
{
- unsigned int len = 0;
- unsigned int offset = 0;
- cfs_list_t *e;
- extern __u64 lnet_create_interface_cookie(void);
-
- cfs_list_for_each (e, &rnet->lrn_routes) {
- len++;
- }
-
- /* FIXME use Lustre random function when it's moved to libcfs.
- * See bug 18751 */
- /* len+1 positions to add a new entry, also prevents division by 0 */
- offset = ((unsigned int) lnet_create_interface_cookie()) % (len + 1);
- cfs_list_for_each (e, &rnet->lrn_routes) {
- if (offset == 0)
- break;
- offset--;
- }
- cfs_list_add(&route->lr_list, e);
-
- the_lnet.ln_remote_nets_version++;
- lnet_rtr_addref_locked(route->lr_gateway);
+ unsigned int len = 0;
+ unsigned int offset = 0;
+ struct list_head *e;
+
+ lnet_shuffle_seed();
+
+ list_for_each(e, &rnet->lrn_routes)
+ len++;
+
+ /*
+ * 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)
+ break;
+ offset--;
+ }
+ list_add(&route->lr_list, e);
+ /*
+ * force a router check on the gateway to make sure the route is
+ * alive
+ */
+ route->lr_gateway->lp_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, unsigned int hops, lnet_nid_t gateway)
+lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway,
+ unsigned int priority)
{
- cfs_list_t *e;
- lnet_remotenet_t *rnet;
- lnet_remotenet_t *rnet2;
- lnet_route_t *route;
- lnet_ni_t *ni;
- int add_route;
- int rc;
-
- CDEBUG(D_NET, "Add route: net %s hops %u gw %s\n",
- libcfs_net2str(net), hops, libcfs_nid2str(gateway));
-
- if (gateway == LNET_NID_ANY ||
- LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
- net == LNET_NIDNET(LNET_NID_ANY) ||
- LNET_NETTYP(net) == LOLND ||
- LNET_NIDNET(gateway) == net ||
- hops < 1 || hops > 255)
- return (-EINVAL);
-
- if (lnet_islocalnet(net)) /* it's a local network */
- return 0; /* ignore the route entry */
-
- /* Assume net, route, all new */
- LIBCFS_ALLOC(route, sizeof(*route));
- LIBCFS_ALLOC(rnet, sizeof(*rnet));
- if (route == NULL || rnet == NULL) {
- CERROR("Out of memory creating route %s %d %s\n",
- libcfs_net2str(net), hops, libcfs_nid2str(gateway));
- if (route != NULL)
- LIBCFS_FREE(route, sizeof(*route));
- if (rnet != NULL)
- LIBCFS_FREE(rnet, sizeof(*rnet));
- return -ENOMEM;
- }
-
- CFS_INIT_LIST_HEAD(&rnet->lrn_routes);
- rnet->lrn_net = net;
- route->lr_hops = hops;
-
- LNET_LOCK();
-
- rc = lnet_nid2peer_locked(&route->lr_gateway, gateway);
- if (rc != 0) {
- LNET_UNLOCK();
-
- LIBCFS_FREE(route, sizeof(*route));
- LIBCFS_FREE(rnet, sizeof(*rnet));
-
- if (rc == -EHOSTUNREACH) /* gateway is not on a local net */
- return 0; /* ignore the route entry */
-
- CERROR("Error %d creating route %s %d %s\n", rc,
- libcfs_net2str(net), hops, libcfs_nid2str(gateway));
- return rc;
- }
-
- LASSERT (!the_lnet.ln_shutdown);
-
- rnet2 = lnet_find_net_locked(net);
- if (rnet2 == NULL) {
- /* new network */
- cfs_list_add_tail(&rnet->lrn_list, &the_lnet.ln_remote_nets);
- rnet2 = rnet;
- }
-
- /* Search for a duplicate route (it's a NOOP if it is) */
- add_route = 1;
- cfs_list_for_each (e, &rnet2->lrn_routes) {
- lnet_route_t *route2 = cfs_list_entry(e, lnet_route_t, lr_list);
-
- if (route2->lr_gateway == route->lr_gateway) {
- add_route = 0;
- break;
- }
-
- /* our lookups must be true */
- LASSERT (route2->lr_gateway->lp_nid != gateway);
- }
-
- if (add_route) {
- ni = route->lr_gateway->lp_ni;
- lnet_ni_addref_locked(ni);
-
- lnet_add_route_to_rnet(rnet2, route);
- LNET_UNLOCK();
-
- /* XXX Assume alive */
- if (ni->ni_lnd->lnd_notify != NULL)
- (ni->ni_lnd->lnd_notify)(ni, gateway, 1);
-
- lnet_ni_decref(ni);
- } else {
- lnet_peer_decref_locked(route->lr_gateway);
- LNET_UNLOCK();
- LIBCFS_FREE(route, sizeof(*route));
- }
-
- if (rnet != rnet2)
- LIBCFS_FREE(rnet, sizeof(*rnet));
-
- return 0;
+ 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 ||
+ LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
+ net == LNET_NIDNET(LNET_NID_ANY) ||
+ LNET_NETTYP(net) == LOLND ||
+ LNET_NIDNET(gateway) == net ||
+ (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
+ return -EINVAL;
+
+ /* it's a local network */
+ if (lnet_islocalnet(net))
+ return -EEXIST;
+
+ /* Assume net, route, all new */
+ LIBCFS_ALLOC(route, sizeof(*route));
+ LIBCFS_ALLOC(rnet, sizeof(*rnet));
+ if (route == NULL || rnet == NULL) {
+ CERROR("Out of memory creating route %s %d %s\n",
+ libcfs_net2str(net), hops, libcfs_nid2str(gateway));
+ if (route != NULL)
+ LIBCFS_FREE(route, sizeof(*route));
+ if (rnet != NULL)
+ LIBCFS_FREE(rnet, sizeof(*rnet));
+ return -ENOMEM;
+ }
+
+ INIT_LIST_HEAD(&rnet->lrn_routes);
+ rnet->lrn_net = net;
+ /* store the local and remote net that the route represents */
+ route->lr_lnet = LNET_NIDNET(gateway);
+ route->lr_net = net;
+ 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(route, sizeof(*route));
+ LIBCFS_FREE(rnet, sizeof(*rnet));
+
+ rc = PTR_ERR(lpni);
+ CERROR("Error %d creating route %s %d %s\n", rc,
+ libcfs_net2str(net), hops,
+ libcfs_nid2str(gateway));
+ return rc;
+ }
+
+ 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) {
+ /* new network */
+ list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
+ rnet2 = rnet;
+ }
+
+ /* Search for a duplicate route (it's a NOOP if it is) */
+ add_route = 1;
+ 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->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_add_route_to_rnet(rnet2, route);
+
+ /*
+ * get rid of the reference on the lpni.
+ */
+ lnet_peer_ni_decref_locked(lpni);
+ lnet_net_unlock(LNET_LOCK_EX);
+
+ rc = 0;
+
+ if (!add_route) {
+ rc = -EEXIST;
+ LIBCFS_FREE(route, sizeof(*route));
+ }
+
+ if (rnet != rnet2)
+ LIBCFS_FREE(rnet, sizeof(*rnet));
+
+ /* 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;
- cfs_list_t *e1;
- cfs_list_t *e2;
-
- LNET_LOCK();
-
- cfs_list_for_each (e1, &the_lnet.ln_remote_nets) {
- rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list);
-
- route2 = NULL;
- cfs_list_for_each (e2, &rnet->lrn_routes) {
- route = cfs_list_entry(e2, lnet_route_t, lr_list);
-
- if (route2 == NULL)
- route2 = route;
- else if (route->lr_gateway->lp_ni !=
- route2->lr_gateway->lp_ni) {
- LNET_UNLOCK();
-
- CERROR("Routes to %s via %s and %s not supported\n",
- libcfs_net2str(rnet->lrn_net),
- libcfs_nid2str(route->lr_gateway->lp_nid),
- libcfs_nid2str(route2->lr_gateway->lp_nid));
- return -EINVAL;
- }
- }
- }
-
- LNET_UNLOCK();
- return 0;
+ 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;
+
+ /*
+ * 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++;
+
+ list_del(&route->lr_gwlist);
+ lnet_rtr_decref_locked(gateway);
+ }
}
int
-lnet_del_route (__u32 net, lnet_nid_t gw_nid)
+lnet_del_route(__u32 net, lnet_nid_t gw_nid)
{
- lnet_remotenet_t *rnet;
- lnet_route_t *route;
- cfs_list_t *e1;
- cfs_list_t *e2;
- int rc = -ENOENT;
-
- CDEBUG(D_NET, "Del route: net %s : gw %s\n",
- libcfs_net2str(net), libcfs_nid2str(gw_nid));
-
- /* NB Caller may specify either all routes via the given gateway
- * or a specific route entry actual NIDs) */
-
- again:
- LNET_LOCK();
-
- cfs_list_for_each (e1, &the_lnet.ln_remote_nets) {
- rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list);
-
- if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
- net == rnet->lrn_net))
- continue;
-
- cfs_list_for_each (e2, &rnet->lrn_routes) {
- route = cfs_list_entry(e2, lnet_route_t, lr_list);
-
- if (!(gw_nid == LNET_NID_ANY ||
- gw_nid == route->lr_gateway->lp_nid))
- continue;
-
- cfs_list_del(&route->lr_list);
- the_lnet.ln_remote_nets_version++;
-
- if (cfs_list_empty(&rnet->lrn_routes))
- cfs_list_del(&rnet->lrn_list);
- else
- rnet = NULL;
-
- lnet_rtr_decref_locked(route->lr_gateway);
- lnet_peer_decref_locked(route->lr_gateway);
- LNET_UNLOCK();
-
- LIBCFS_FREE(route, sizeof (*route));
-
- if (rnet != NULL)
- LIBCFS_FREE(rnet, sizeof(*rnet));
-
- rc = 0;
- goto again;
- }
- }
-
- LNET_UNLOCK();
- return rc;
+ 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));
+
+ /* NB Caller may specify either all routes via the given gateway
+ * or a specific route entry actual NIDs) */
+
+ lnet_net_lock(LNET_LOCK_EX);
+
+ 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);
+ }
+
+ 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;
+ }
+
+ 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_move(&rnet->lrn_list, &rnet_zombies);
+ }
+ }
+
+delete_zombies:
+ lnet_net_unlock(LNET_LOCK_EX);
+
+ while (!list_empty(&zombies)) {
+ route = list_first_entry(&zombies, struct lnet_route, lr_list);
+ list_del(&route->lr_list);
+ LIBCFS_FREE(route, sizeof(*route));
+ }
+
+ 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));
+ }
+
+ return 0;
}
void
lnet_destroy_routes (void)
{
- lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
-}
-
-int
-lnet_get_route (int idx, __u32 *net, __u32 *hops,
- lnet_nid_t *gateway, __u32 *alive)
-{
- cfs_list_t *e1;
- cfs_list_t *e2;
- lnet_remotenet_t *rnet;
- lnet_route_t *route;
-
- LNET_LOCK();
-
- cfs_list_for_each (e1, &the_lnet.ln_remote_nets) {
- rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list);
-
- cfs_list_for_each (e2, &rnet->lrn_routes) {
- route = cfs_list_entry(e2, lnet_route_t, lr_list);
-
- if (idx-- == 0) {
- *net = rnet->lrn_net;
- *hops = route->lr_hops;
- *gateway = route->lr_gateway->lp_nid;
- *alive = route->lr_gateway->lp_alive;
- LNET_UNLOCK();
- return 0;
- }
- }
- }
-
- LNET_UNLOCK();
- return -ENOENT;
+ lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
}
-void
-lnet_swap_pinginfo(lnet_ping_info_t *info)
+int lnet_get_rtr_pool_cfg(int cpt, struct lnet_ioctl_pool_cfg *pool_cfg)
{
- int i;
- lnet_ni_status_t *stat;
-
- __swab32s(&info->pi_version);
- __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;
+ struct lnet_rtrbufpool *rbp;
+ int i, rc = -ENOENT, j;
+
+ if (the_lnet.ln_rtrpools == NULL)
+ return rc;
+
+
+ cfs_percpt_for_each(rbp, i, the_lnet.ln_rtrpools) {
+ if (i != cpt)
+ continue;
+
+ 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(i);
+ rc = 0;
+ break;
+ }
+
+ lnet_net_lock(LNET_LOCK_EX);
+ pool_cfg->pl_routing = the_lnet.ln_routing;
+ lnet_net_unlock(LNET_LOCK_EX);
+
+ return rc;
}
-/* Returns # of down NIs, or negative error codes; ignore downed NIs
- * if a NI in 'net' is up */
int
-lnet_router_down_ni(lnet_peer_t *rtr, __u32 net)
+lnet_get_route(int idx, __u32 *net, __u32 *hops,
+ lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
{
- int i;
- int down = 0;
- int ptl_up = 0;
- int ptl_down = 0;
- lnet_ping_info_t *info;
-
- if (!avoid_asym_router_failure)
- return -ENOENT;
-
- if (rtr->lp_rcd == NULL)
- return -EINVAL;
-
- if (!rtr->lp_alive)
- return -EINVAL; /* stale lp_rcd */
-
- info = rtr->lp_rcd->rcd_pinginfo;
- LASSERT (info != NULL);
-
- /* NB always racing with network! */
- if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
- lnet_swap_pinginfo(info);
- } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
- CDEBUG(D_NETERROR, "%s: Unexpected magic %08x\n",
- libcfs_nid2str(rtr->lp_nid), info->pi_magic);
- return -EPROTO;
- }
-
- if (info->pi_version == LNET_PROTO_PING_VERSION1)
- return -ENOENT; /* v1 doesn't carry NI status info */
-
- if (info->pi_version != LNET_PROTO_PING_VERSION) {
- CDEBUG(D_NETERROR, "%s: Unexpected version 0x%x\n",
- libcfs_nid2str(rtr->lp_nid), info->pi_version);
- return -EPROTO;
- }
-
- for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
- lnet_ni_status_t *stat = &info->pi_ni[i];
- lnet_nid_t nid = stat->ns_nid;
-
- if (nid == LNET_NID_ANY) {
- CDEBUG(D_NETERROR, "%s: unexpected LNET_NID_ANY\n",
- libcfs_nid2str(rtr->lp_nid));
- return -EPROTO;
- }
-
- if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
- continue;
-
- if (stat->ns_status == LNET_NI_STATUS_DOWN) {
- if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND)
- ptl_down = 1;
- else
- down++;
- continue;
- }
-
- if (stat->ns_status != LNET_NI_STATUS_UP) {
- CDEBUG(D_NETERROR, "%s: Unexpected status 0x%x\n",
- libcfs_nid2str(rtr->lp_nid), stat->ns_status);
- return -EPROTO;
- }
-
- /* ignore downed NIs if there's a NI up for dest network */
- if (LNET_NIDNET(nid) == net)
- return 0;
-
- if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND)
- ptl_up = 1;
- }
-
- /* ptl NIs are considered down only when they're all down */
- return down + (ptl_up ? 0 : ptl_down);
+ struct list_head *e1;
+ struct list_head *e2;
+ struct lnet_remotenet *rnet;
+ struct lnet_route *route;
+ int cpt;
+ int i;
+ struct list_head *rn_list;
+
+ 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, struct lnet_remotenet, lrn_list);
+
+ list_for_each(e2, &rnet->lrn_routes) {
+ route = list_entry(e2, struct lnet_route,
+ lr_list);
+
+ if (idx-- == 0) {
+ *net = rnet->lrn_net;
+ *hops = route->lr_hops;
+ *priority = route->lr_priority;
+ *gateway = route->lr_gateway->lp_primary_nid;
+ *alive = lnet_is_route_alive(route);
+ lnet_net_unlock(cpt);
+ return 0;
+ }
+ }
+ }
+ }
+
+ lnet_net_unlock(cpt);
+ return -ENOENT;
}
void
-lnet_wait_known_routerstate(void)
+lnet_swap_pinginfo(struct lnet_ping_buffer *pbuf)
{
- lnet_peer_t *rtr;
- cfs_list_t *entry;
- int all_known;
-
- LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
-
- for (;;) {
- LNET_LOCK();
-
- all_known = 1;
- cfs_list_for_each (entry, &the_lnet.ln_routers) {
- rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list);
-
- if (rtr->lp_alive_count == 0) {
- all_known = 0;
- break;
- }
- }
-
- LNET_UNLOCK();
-
- if (all_known)
- return;
+ struct lnet_ni_status *stat;
+ int nnis;
+ int i;
+
+ __swab32s(&pbuf->pb_info.pi_magic);
+ __swab32s(&pbuf->pb_info.pi_features);
+ __swab32s(&pbuf->pb_info.pi_pid);
+ __swab32s(&pbuf->pb_info.pi_nnis);
+ nnis = pbuf->pb_info.pi_nnis;
+ if (nnis > pbuf->pb_nnis)
+ nnis = pbuf->pb_nnis;
+ for (i = 0; i < nnis; i++) {
+ stat = &pbuf->pb_info.pi_ni[i];
+ __swab64s(&stat->ns_nid);
+ __swab32s(&stat->ns_status);
+ }
+ return;
+}
-#ifndef __KERNEL__
- lnet_router_checker();
-#endif
- cfs_pause(cfs_time_seconds(1));
- }
+/**
+ * TODO: re-implement
+ */
+static void
+lnet_parse_rc_info(struct lnet_rc_data *rcd)
+{
+ rcd = rcd;
}
static void
-lnet_router_checker_event (lnet_event_t *event)
+lnet_router_checker_event(struct lnet_event *event)
{
- /* CAVEAT EMPTOR: I'm called with LNET_LOCKed and I'm not allowed to
- * drop it (that's how come I see _every_ event, even ones that would
- * overflow my EQ) */
- lnet_rc_data_t *rcd = event->md.user_ptr;
- lnet_peer_t *lp;
- lnet_nid_t nid;
-
- if (event->unlinked) {
- if (rcd != NULL) {
- LNetInvalidateHandle(&rcd->rcd_mdh);
- return;
- }
-
- /* The router checker thread has unlinked the default rc_md
- * and exited. */
- LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKING);
- the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKED;
-#ifdef __KERNEL__
- cfs_mutex_up(&the_lnet.ln_rc_signal);
-#endif
- return;
- }
-
- LASSERT (event->type == LNET_EVENT_SEND ||
- event->type == LNET_EVENT_REPLY);
-
- nid = (event->type == LNET_EVENT_SEND) ?
- event->target.nid : event->initiator.nid;
-
- lp = lnet_find_peer_locked(nid);
- if (lp == NULL) {
- /* router may have been removed */
- CDEBUG(D_NET, "Router %s not found\n", libcfs_nid2str(nid));
- return;
- }
-
- if (event->type == LNET_EVENT_SEND) /* re-enable another ping */
- lp->lp_ping_notsent = 0;
-
- if (lnet_isrouter(lp) && /* ignore if no longer a router */
- (event->status != 0 ||
- event->type == 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!!! */
- }
-
- /* This decref will NOT drop LNET_LOCK (it had to have 1 ref when it
- * was in the peer table and I've not dropped the lock, so no-one else
- * can have reduced the refcount) */
- LASSERT(lp->lp_refcount > 1);
-
- lnet_peer_decref_locked(lp);
+ struct lnet_rc_data *rcd = event->md.user_ptr;
+ struct lnet_peer_ni *lp;
+
+ LASSERT(rcd != NULL);
+
+ if (event->unlinked) {
+ LNetInvalidateMDHandle(&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) {
+ 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, ktime_get_seconds());
+ /* 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);
}
-void
-lnet_update_ni_status(void)
+static void
+lnet_wait_known_routerstate(void)
{
- cfs_time_t now = cfs_time_current();
- lnet_ni_t *ni;
- int status;
- int timeout;
+ struct lnet_peer *rtr;
+ struct list_head *entry;
+ int all_known;
- LASSERT (the_lnet.ln_routing);
+ LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
- timeout = router_ping_timeout +
- MAX(live_router_check_interval, dead_router_check_interval);
+ for (;;) {
+ int cpt = lnet_net_lock_current();
- LNET_LOCK();
+ all_known = 1;
+ list_for_each(entry, &the_lnet.ln_routers) {
+ rtr = list_entry(entry, struct lnet_peer,
+ lp_rtr_list);
- cfs_list_for_each_entry (ni, &the_lnet.ln_nis, ni_list) {
- lnet_ni_status_t *ns = ni->ni_status;
+ spin_lock(&rtr->lp_lock);
- LASSERT (ns != NULL);
+ if ((rtr->lp_state & LNET_PEER_DISCOVERED) == 0) {
+ all_known = 0;
+ spin_unlock(&rtr->lp_lock);
+ break;
+ }
+ spin_unlock(&rtr->lp_lock);
+ }
- status = LNET_NI_STATUS_UP;
- if (ni->ni_lnd->lnd_type != LOLND && /* @lo forever alive */
- cfs_time_after(now, cfs_time_add(ni->ni_last_alive,
- cfs_time_seconds(timeout))))
- status = LNET_NI_STATUS_DOWN;
+ lnet_net_unlock(cpt);
- if (ns->ns_status != status) {
- ns->ns_status = status;
- CDEBUG(D_NET, "NI(%s:%d) status changed to %s\n",
- libcfs_nid2str(ni->ni_nid), timeout,
- status == LNET_NI_STATUS_UP ? "up" : "down");
- }
- }
+ if (all_known)
+ return;
- LNET_UNLOCK();
+ set_current_state(TASK_UNINTERRUPTIBLE);
+ schedule_timeout(cfs_time_seconds(1));
+ }
}
+/* TODO: reimplement */
void
-lnet_destroy_rc_data (lnet_rc_data_t *rcd)
+lnet_router_ni_update_locked(struct lnet_peer_ni *gw, __u32 net)
{
- LASSERT (cfs_list_empty(&rcd->rcd_list));
- /* detached from network */
- LASSERT (LNetHandleIsInvalid(rcd->rcd_mdh));
-
- LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
- LIBCFS_FREE(rcd, sizeof(*rcd));
- return;
+ struct lnet_route *rte;
+ struct lnet_peer *lp;
+
+ if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0)
+ lp = gw->lpni_peer_net->lpn_peer;
+ else
+ return;
+
+ list_for_each_entry(rte, &lp->lp_routes, lr_gwlist) {
+ if (rte->lr_net == net) {
+ rte->lr_downis = 0;
+ break;
+ }
+ }
}
-lnet_rc_data_t *
-lnet_create_rc_data (void)
+static void
+lnet_update_ni_status_locked(void)
{
- int i;
- int rc;
- lnet_ping_info_t *pi;
- lnet_rc_data_t *rcd;
-
- LIBCFS_ALLOC(rcd, sizeof(*rcd));
- if (rcd == NULL)
- return NULL;
-
- LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
- if (pi == NULL) {
- LIBCFS_FREE(rcd, sizeof(*rcd));
- return NULL;
- }
-
- memset(pi, 0, LNET_PINGINFO_SIZE);
- 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;
- LNetInvalidateHandle(&rcd->rcd_mdh);
- CFS_INIT_LIST_HEAD(&rcd->rcd_list);
-
- 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);
- lnet_destroy_rc_data(rcd);
- return NULL;
- }
- LASSERT (rc == 0);
- return rcd;
+ struct lnet_ni *ni = NULL;
+ time64_t now;
+ time64_t timeout;
+
+ LASSERT(the_lnet.ln_routing);
+
+ timeout = router_ping_timeout +
+ MAX(live_router_check_interval, dead_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)
+ continue;
+
+ if (now < ni->ni_last_alive + timeout)
+ continue;
+
+ lnet_ni_lock(ni);
+ /* re-check with lock */
+ if (now < ni->ni_last_alive + timeout) {
+ lnet_ni_unlock(ni);
+ continue;
+ }
+
+ LASSERT(ni->ni_status != NULL);
+
+ if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
+ CDEBUG(D_NET, "NI(%s:%lld) 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 int
-lnet_router_check_interval (lnet_peer_t *rtr)
+int lnet_router_pre_mt_start(void)
{
- int secs;
-
- secs = rtr->lp_alive ? live_router_check_interval :
- dead_router_check_interval;
- if (secs < 0)
- secs = 0;
-
- return secs;
+ int rc;
+
+ 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;
+ }
+
+ rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
+ if (rc != 0) {
+ CERROR("Can't allocate EQ(0): %d\n", rc);
+ return -ENOMEM;
+ }
+
+ return 0;
}
-static void
-lnet_ping_router_locked (lnet_peer_t *rtr)
+void lnet_router_post_mt_start(void)
{
- int newrcd = 0;
- lnet_rc_data_t *rcd = NULL;
- cfs_time_t now = cfs_time_current();
- int secs;
-
- lnet_peer_addref_locked(rtr);
-
- if (rtr->lp_ping_deadline != 0 && /* ping timed out? */
- cfs_time_after(now, rtr->lp_ping_deadline))
- lnet_notify_locked(rtr, 1, 0, now);
-
- if (avoid_asym_router_failure && rtr->lp_rcd == NULL)
- newrcd = 1;
-
- LNET_UNLOCK();
-
- /* Run any outstanding notifications */
- lnet_do_notify(rtr);
-
- if (newrcd)
- rcd = lnet_create_rc_data();
-
- LNET_LOCK();
-
- if (!lnet_isrouter(rtr)) {
- lnet_peer_decref_locked(rtr);
- if (rcd != NULL)
- cfs_list_add(&rcd->rcd_list, &the_lnet.ln_zombie_rcd);
- return; /* router table changed! */
- }
-
- if (rcd != NULL) {
- LASSERT (rtr->lp_rcd == NULL);
- rtr->lp_rcd = rcd;
- }
-
- secs = lnet_router_check_interval(rtr);
-
- CDEBUG(D_NET,
- "rtr %s %d: deadline %lu ping_notsent %d alive %d "
- "alive_count %d lp_ping_timestamp %lu\n",
- libcfs_nid2str(rtr->lp_nid), secs,
- rtr->lp_ping_deadline, rtr->lp_ping_notsent,
- rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
-
- if (secs != 0 && !rtr->lp_ping_notsent &&
- cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
- cfs_time_seconds(secs)))) {
- int rc;
- lnet_process_id_t id;
- lnet_handle_md_t mdh;
-
- id.nid = rtr->lp_nid;
- id.pid = LUSTRE_SRV_LNET_PID;
- CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
-
- rtr->lp_ping_notsent = 1;
- rtr->lp_ping_timestamp = now;
- mdh = (rtr->lp_rcd == NULL) ? the_lnet.ln_rc_mdh :
- rtr->lp_rcd->rcd_mdh;
-
- if (rtr->lp_ping_deadline == 0)
- rtr->lp_ping_deadline = cfs_time_shift(router_ping_timeout);
-
- LNET_UNLOCK();
-
- rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
- LNET_PROTO_PING_MATCHBITS, 0);
-
- LNET_LOCK();
- if (rc != 0)
- rtr->lp_ping_notsent = 0; /* no event pending */
- }
-
- lnet_peer_decref_locked(rtr);
- return;
+ 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();
+ }
}
-int
-lnet_router_checker_start(void)
+void
+lnet_router_cleanup(void)
{
- static lnet_ping_info_t pinginfo;
-
- lnet_md_t md;
- int rc;
- int eqsz;
-#ifndef __KERNEL__
- lnet_peer_t *rtr;
- __u64 version;
- int nrtr = 0;
- int router_checker_max_eqsize = 10240;
-
- LASSERT (check_routers_before_use);
- LASSERT (dead_router_check_interval > 0);
-
- LNET_LOCK();
-
- /* As an approximation, allow each router the same number of
- * outstanding events as it is allowed outstanding sends */
- eqsz = 0;
- version = the_lnet.ln_routers_version;
- cfs_list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) {
- lnet_ni_t *ni = rtr->lp_ni;
- lnet_process_id_t id;
-
- nrtr++;
- eqsz += ni->ni_peertxcredits;
-
- /* one async ping reply per router */
- id.nid = rtr->lp_nid;
- id.pid = LUSTRE_SRV_LNET_PID;
-
- LNET_UNLOCK();
-
- rc = LNetSetAsync(id, 1);
- if (rc != 0) {
- CWARN("LNetSetAsync %s failed: %d\n",
- libcfs_id2str(id), rc);
- return rc;
- }
-
- LNET_LOCK();
- /* NB router list doesn't change in userspace */
- LASSERT (version == the_lnet.ln_routers_version);
- }
-
- LNET_UNLOCK();
-
- if (nrtr == 0) {
- CDEBUG(D_NET,
- "No router found, not starting router checker\n");
- return 0;
- }
-
- /* at least allow a SENT and a REPLY per router */
- if (router_checker_max_eqsize < 2 * nrtr)
- router_checker_max_eqsize = 2 * nrtr;
-
- LASSERT (eqsz > 0);
- if (eqsz > router_checker_max_eqsize)
- eqsz = router_checker_max_eqsize;
-#endif
-
- 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;
- }
-
- if (!the_lnet.ln_routing &&
- live_router_check_interval <= 0 &&
- dead_router_check_interval <= 0)
- return 0;
-
-#ifdef __KERNEL__
- cfs_init_mutex_locked(&the_lnet.ln_rc_signal);
- /* EQ size doesn't matter; the callback is guaranteed to get every
- * event */
- eqsz = 1;
- rc = LNetEQAlloc(eqsz, lnet_router_checker_event,
- &the_lnet.ln_rc_eqh);
-#else
- rc = LNetEQAlloc(eqsz, LNET_EQ_HANDLER_NONE,
- &the_lnet.ln_rc_eqh);
-#endif
- if (rc != 0) {
- CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
- return -ENOMEM;
- }
-
- memset(&md, 0, sizeof(md));
- md.user_ptr = NULL;
- md.start = &pinginfo;
- md.length = sizeof(pinginfo);
- md.options = LNET_MD_TRUNCATE;
- md.threshold = LNET_MD_THRESH_INF;
- md.eq_handle = the_lnet.ln_rc_eqh;
- rc = LNetMDBind(md, LNET_UNLINK, &the_lnet.ln_rc_mdh);
- if (rc < 0) {
- CERROR("Can't bind MD: %d\n", rc);
- rc = LNetEQFree(the_lnet.ln_rc_eqh);
- LASSERT (rc == 0);
- return -ENOMEM;
- }
- LASSERT (rc == 0);
-
- the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
-#ifdef __KERNEL__
- rc = (int)cfs_kernel_thread(lnet_router_checker, NULL, 0);
- if (rc < 0) {
- CERROR("Can't start router checker thread: %d\n", rc);
- the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKING;
- rc = LNetMDUnlink(the_lnet.ln_rc_mdh);
- LASSERT (rc == 0);
- /* block until event callback signals exit */
- cfs_mutex_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;
- }
-#endif
-
- 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;
+ int rc;
+
+ rc = LNetEQFree(the_lnet.ln_rc_eqh);
+ LASSERT(rc == 0);
+ return;
}
void
-lnet_router_checker_stop (void)
+lnet_prune_rc_data(int wait_unlink)
{
- 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_STOPTHREAD;
-
-#ifdef __KERNEL__
- /* block until event callback signals exit */
- cfs_mutex_down(&the_lnet.ln_rc_signal);
-#else
- while (the_lnet.ln_rc_state != LNET_RC_STATE_UNLINKED) {
- lnet_router_checker();
- cfs_pause(cfs_time_seconds(1));
- }
-#endif
- LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKED);
-
- rc = LNetEQFree(the_lnet.ln_rc_eqh);
- LASSERT (rc == 0);
- the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
- return;
+ wait_unlink = wait_unlink;
}
-#if defined(__KERNEL__) && defined(LNET_ROUTER)
-
-static void
-lnet_prune_zombie_rcd (int wait_unlink)
+/*
+ * This function is called from the monitor thread to check if there are
+ * any active routers that need to be checked.
+ */
+inline bool
+lnet_router_checker_active(void)
{
- lnet_rc_data_t *rcd;
- lnet_rc_data_t *tmp;
- cfs_list_t free_rcd;
- int i;
- __u64 version;
-
- CFS_INIT_LIST_HEAD(&free_rcd);
-
- LNET_LOCK();
-rescan:
- version = the_lnet.ln_routers_version;
- cfs_list_for_each_entry_safe (rcd, tmp, &the_lnet.ln_zombie_rcd,
- rcd_list) {
- if (LNetHandleIsInvalid(rcd->rcd_mdh)) {
- cfs_list_del(&rcd->rcd_list);
- cfs_list_add(&rcd->rcd_list, &free_rcd);
- continue;
- }
-
- LNET_UNLOCK();
-
- LNetMDUnlink(rcd->rcd_mdh);
-
- LNET_LOCK();
- if (version != the_lnet.ln_routers_version)
- goto rescan;
- }
-
- i = 2;
- while (wait_unlink && !cfs_list_empty(&the_lnet.ln_zombie_rcd)) {
- rcd = cfs_list_entry(the_lnet.ln_zombie_rcd.next,
- lnet_rc_data_t, rcd_list);
- if (LNetHandleIsInvalid(rcd->rcd_mdh)) {
- cfs_list_del(&rcd->rcd_list);
- cfs_list_add(&rcd->rcd_list, &free_rcd);
- continue;
- }
-
- LNET_UNLOCK();
-
- LNetMDUnlink(rcd->rcd_mdh);
-
- i++;
- CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
- "Waiting for rc buffers to unlink\n");
- cfs_pause(cfs_time_seconds(1));
-
- LNET_LOCK();
- }
-
- LNET_UNLOCK();
-
- while (!cfs_list_empty(&free_rcd)) {
- rcd = cfs_list_entry(free_rcd.next, lnet_rc_data_t, rcd_list);
- cfs_list_del_init(&rcd->rcd_list);
- lnet_destroy_rc_data(rcd);
- }
- return;
+ if (the_lnet.ln_mt_state != LNET_MT_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;
+
+ /* if there are routers that need to be cleaned up then do so */
+ if (!list_empty(&the_lnet.ln_rcd_deathrow) ||
+ !list_empty(&the_lnet.ln_rcd_zombie))
+ return true;
+
+ return !list_empty(&the_lnet.ln_routers) &&
+ (live_router_check_interval > 0 ||
+ dead_router_check_interval > 0);
}
-static int
-lnet_router_checker(void *arg)
+void
+lnet_check_routers(void)
{
- int rc;
- lnet_peer_t *rtr;
- cfs_list_t *entry;
- lnet_process_id_t rtr_id;
-
- cfs_daemonize("router_checker");
- cfs_block_allsigs();
-
- rtr_id.pid = LUSTRE_SRV_LNET_PID;
-
- LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
+ struct lnet_peer *rtr;
+ struct list_head *entry;
+ __u64 version;
+ int cpt;
- while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
- __u64 version;
-
- LNET_LOCK();
+ cpt = lnet_net_lock_current();
rescan:
- version = the_lnet.ln_routers_version;
-
- cfs_list_for_each (entry, &the_lnet.ln_routers) {
- rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list);
- lnet_ping_router_locked(rtr);
-
- /* NB dropped lock */
- if (version != the_lnet.ln_routers_version) {
- /* the routers list has changed */
- goto rescan;
- }
- }
-
- LNET_UNLOCK();
-
- if (the_lnet.ln_routing)
- lnet_update_ni_status();
+ version = the_lnet.ln_routers_version;
- lnet_prune_zombie_rcd(0); /* don't wait for UNLINK */
+ list_for_each(entry, &the_lnet.ln_routers) {
+ rtr = list_entry(entry, struct lnet_peer,
+ lp_rtr_list);
- /* Call cfs_pause() here always adds 1 to load average
- * because kernel counts # active tasks as nr_running
- * + nr_uninterruptible. */
- cfs_schedule_timeout_and_set_state(CFS_TASK_INTERRUPTIBLE,
- cfs_time_seconds(1));
- }
+ /* TODO use discovery to determine if router is alive */
- LNET_LOCK();
+ /* NB dropped lock */
+ if (version != the_lnet.ln_routers_version) {
+ /* the routers list has changed */
+ goto rescan;
+ }
+ }
- cfs_list_for_each (entry, &the_lnet.ln_routers) {
- rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list);
+ if (the_lnet.ln_routing)
+ lnet_update_ni_status_locked();
- if (rtr->lp_rcd == NULL)
- continue;
+ lnet_net_unlock(cpt);
- LASSERT (cfs_list_empty(&rtr->lp_rcd->rcd_list));
- cfs_list_add(&rtr->lp_rcd->rcd_list, &the_lnet.ln_zombie_rcd);
- rtr->lp_rcd = NULL;
- }
-
- LNET_UNLOCK();
-
- lnet_prune_zombie_rcd(1); /* wait for UNLINK */
-
- LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_STOPTHREAD);
- the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKING;
-
- rc = LNetMDUnlink(the_lnet.ln_rc_mdh);
- LASSERT (rc == 0);
-
- /* The unlink event callback will signal final completion */
- return 0;
+ lnet_prune_rc_data(0); /* don't wait for UNLINK */
}
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)
- cfs_free_page(rb->rb_kiov[npages].kiov_page);
+ while (--npages >= 0)
+ __free_page(rb->rb_kiov[npages].kiov_page);
- LIBCFS_FREE(rb, sz);
+ LIBCFS_FREE(rb, sz);
}
-lnet_rtrbuf_t *
-lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp)
+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]);
- struct page *page;
- lnet_rtrbuf_t *rb;
- int i;
-
- LIBCFS_ALLOC(rb, sz);
- if (rb == NULL)
- return NULL;
-
- rb->rb_pool = rbp;
-
- for (i = 0; i < npages; i++) {
- page = cfs_alloc_page(CFS_ALLOC_ZERO | CFS_ALLOC_STD);
- if (page == NULL) {
- while (--i >= 0)
- cfs_free_page(rb->rb_kiov[i].kiov_page);
-
- LIBCFS_FREE(rb, sz);
- return NULL;
- }
-
- rb->rb_kiov[i].kiov_len = CFS_PAGE_SIZE;
- rb->rb_kiov[i].kiov_offset = 0;
- rb->rb_kiov[i].kiov_page = page;
- }
-
- return rb;
+ int npages = rbp->rbp_npages;
+ int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
+ struct page *page;
+ struct lnet_rtrbuf *rb;
+ int i;
+
+ LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
+ if (rb == NULL)
+ return NULL;
+
+ rb->rb_pool = rbp;
+
+ for (i = 0; i < npages; i++) {
+ page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
+ GFP_KERNEL | __GFP_ZERO);
+ if (page == NULL) {
+ while (--i >= 0)
+ __free_page(rb->rb_kiov[i].kiov_page);
+
+ LIBCFS_FREE(rb, sz);
+ return NULL;
+ }
+
+ rb->rb_kiov[i].kiov_len = PAGE_SIZE;
+ rb->rb_kiov[i].kiov_offset = 0;
+ rb->rb_kiov[i].kiov_page = page;
+ }
+
+ return rb;
}
-void
-lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp)
+static void
+lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
{
- int npages = rbp->rbp_npages;
- int nbuffers = 0;
- lnet_rtrbuf_t *rb;
-
- LASSERT (cfs_list_empty(&rbp->rbp_msgs));
- LASSERT (rbp->rbp_credits == rbp->rbp_nbuffers);
-
- while (!cfs_list_empty(&rbp->rbp_bufs)) {
- LASSERT (rbp->rbp_credits > 0);
-
- rb = cfs_list_entry(rbp->rbp_bufs.next,
- lnet_rtrbuf_t, rb_list);
- cfs_list_del(&rb->rb_list);
- lnet_destroy_rtrbuf(rb, npages);
- nbuffers++;
- }
-
- LASSERT (rbp->rbp_nbuffers == nbuffers);
- LASSERT (rbp->rbp_credits == nbuffers);
-
- rbp->rbp_nbuffers = rbp->rbp_credits = 0;
+ int npages = rbp->rbp_npages;
+ struct lnet_rtrbuf *rb;
+ struct list_head tmp;
+
+ if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
+ return;
+
+ INIT_LIST_HEAD(&tmp);
+
+ lnet_net_lock(cpt);
+ list_splice_init(&rbp->rbp_msgs, &tmp);
+ lnet_drop_routed_msgs_locked(&tmp, cpt);
+ list_splice_init(&rbp->rbp_bufs, &tmp);
+ rbp->rbp_req_nbuffers = 0;
+ rbp->rbp_nbuffers = rbp->rbp_credits = 0;
+ rbp->rbp_mincredits = 0;
+ lnet_net_unlock(cpt);
+
+ /* Free buffers on the free list. */
+ while (!list_empty(&tmp)) {
+ rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
+ list_del(&rb->rb_list);
+ lnet_destroy_rtrbuf(rb, npages);
+ }
}
-int
-lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs)
+static int
+lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
{
- lnet_rtrbuf_t *rb;
- int i;
-
- if (rbp->rbp_nbuffers != 0) {
- LASSERT (rbp->rbp_nbuffers == nbufs);
- return 0;
- }
-
- for (i = 0; i < nbufs; i++) {
- rb = lnet_new_rtrbuf(rbp);
-
- if (rb == NULL) {
- CERROR("Failed to allocate %d router bufs of %d pages\n",
- nbufs, rbp->rbp_npages);
- return -ENOMEM;
- }
-
- rbp->rbp_nbuffers++;
- rbp->rbp_credits++;
- rbp->rbp_mincredits++;
- cfs_list_add(&rb->rb_list, &rbp->rbp_bufs);
-
- /* No allocation "under fire" */
- /* Otherwise we'd need code to schedule blocked msgs etc */
- LASSERT (!the_lnet.ln_routing);
- }
-
- LASSERT (rbp->rbp_credits == nbufs);
- return 0;
+ struct list_head rb_list;
+ struct lnet_rtrbuf *rb;
+ int num_rb;
+ int num_buffers = 0;
+ int old_req_nbufs;
+ int npages = rbp->rbp_npages;
+
+ lnet_net_lock(cpt);
+ /* If we are called for less buffers than already in the pool, we
+ * just lower the req_nbuffers number and excess buffers will be
+ * thrown away as they are returned to the free list. Credits
+ * then get adjusted as well.
+ * If we already have enough buffers allocated to serve the
+ * increase requested, then we can treat that the same way as we
+ * do the decrease. */
+ num_rb = nbufs - rbp->rbp_nbuffers;
+ if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
+ rbp->rbp_req_nbuffers = nbufs;
+ lnet_net_unlock(cpt);
+ return 0;
+ }
+ /* store the older value of rbp_req_nbuffers and then set it to
+ * the new request to prevent lnet_return_rx_credits_locked() from
+ * freeing buffers that we need to keep around */
+ old_req_nbufs = rbp->rbp_req_nbuffers;
+ rbp->rbp_req_nbuffers = nbufs;
+ lnet_net_unlock(cpt);
+
+ INIT_LIST_HEAD(&rb_list);
+
+ /* allocate the buffers on a local list first. If all buffers are
+ * allocated successfully then join this list to the rbp buffer
+ * list. If not then free all allocated buffers. */
+ while (num_rb-- > 0) {
+ rb = lnet_new_rtrbuf(rbp, cpt);
+ if (rb == NULL) {
+ CERROR("Failed to allocate %d route bufs of %d pages\n",
+ nbufs, npages);
+
+ lnet_net_lock(cpt);
+ rbp->rbp_req_nbuffers = old_req_nbufs;
+ lnet_net_unlock(cpt);
+
+ goto failed;
+ }
+
+ list_add(&rb->rb_list, &rb_list);
+ num_buffers++;
+ }
+
+ lnet_net_lock(cpt);
+
+ list_splice_tail(&rb_list, &rbp->rbp_bufs);
+ rbp->rbp_nbuffers += num_buffers;
+ rbp->rbp_credits += num_buffers;
+ rbp->rbp_mincredits = rbp->rbp_credits;
+ /* We need to schedule blocked msg using the newly
+ * added buffers. */
+ while (!list_empty(&rbp->rbp_bufs) &&
+ !list_empty(&rbp->rbp_msgs))
+ lnet_schedule_blocked_locked(rbp);
+
+ lnet_net_unlock(cpt);
+
+ return 0;
+
+failed:
+ while (!list_empty(&rb_list)) {
+ rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
+ list_del(&rb->rb_list);
+ lnet_destroy_rtrbuf(rb, npages);
+ }
+
+ return -ENOMEM;
}
-void
-lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
+static void
+lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
{
- CFS_INIT_LIST_HEAD(&rbp->rbp_msgs);
- CFS_INIT_LIST_HEAD(&rbp->rbp_bufs);
+ INIT_LIST_HEAD(&rbp->rbp_msgs);
+ INIT_LIST_HEAD(&rbp->rbp_bufs);
- rbp->rbp_npages = npages;
- rbp->rbp_credits = 0;
- rbp->rbp_mincredits = 0;
+ rbp->rbp_npages = npages;
+ rbp->rbp_credits = 0;
+ rbp->rbp_mincredits = 0;
}
void
-lnet_free_rtrpools(void)
+lnet_rtrpools_free(int keep_pools)
{
- lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[0]);
- lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[1]);
- lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[2]);
+ struct lnet_rtrbufpool *rtrp;
+ int i;
+
+ if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
+ return;
+
+ cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
+ lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
+ lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
+ lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
+ }
+
+ if (!keep_pools) {
+ cfs_percpt_free(the_lnet.ln_rtrpools);
+ the_lnet.ln_rtrpools = NULL;
+ }
}
-void
-lnet_init_rtrpools(void)
+static int
+lnet_nrb_tiny_calculate(void)
{
- int small_pages = 1;
- int large_pages = (LNET_MTU + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
+ int nrbs = LNET_NRB_TINY;
- lnet_rtrpool_init(&the_lnet.ln_rtrpools[0], 0);
- lnet_rtrpool_init(&the_lnet.ln_rtrpools[1], small_pages);
- lnet_rtrpool_init(&the_lnet.ln_rtrpools[2], large_pages);
-}
+ if (tiny_router_buffers < 0) {
+ LCONSOLE_ERROR_MSG(0x10c,
+ "tiny_router_buffers=%d invalid when "
+ "routing enabled\n", tiny_router_buffers);
+ return -EINVAL;
+ }
+ if (tiny_router_buffers > 0)
+ nrbs = tiny_router_buffers;
-int
-lnet_alloc_rtrpools(int im_a_router)
-{
- int rc;
-
- if (!strcmp(forwarding, "")) {
- /* not set either way */
- if (!im_a_router)
- return 0;
- } else if (!strcmp(forwarding, "disabled")) {
- /* explicitly disabled */
- return 0;
- } else if (!strcmp(forwarding, "enabled")) {
- /* explicitly enabled */
- } else {
- LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
- "'enabled' or 'disabled'\n");
- return -EINVAL;
- }
-
- if (tiny_router_buffers <= 0) {
- LCONSOLE_ERROR_MSG(0x10c, "tiny_router_buffers=%d invalid when "
- "routing enabled\n", tiny_router_buffers);
- rc = -EINVAL;
- goto failed;
- }
-
- rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[0],
- tiny_router_buffers);
- if (rc != 0)
- goto failed;
-
- if (small_router_buffers <= 0) {
- LCONSOLE_ERROR_MSG(0x10d, "small_router_buffers=%d invalid when"
- " routing enabled\n", small_router_buffers);
- rc = -EINVAL;
- goto failed;
- }
-
- rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[1],
- small_router_buffers);
- if (rc != 0)
- goto failed;
-
- if (large_router_buffers <= 0) {
- LCONSOLE_ERROR_MSG(0x10e, "large_router_buffers=%d invalid when"
- " routing enabled\n", large_router_buffers);
- rc = -EINVAL;
- goto failed;
- }
-
- rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[2],
- large_router_buffers);
- if (rc != 0)
- goto failed;
-
- LNET_LOCK();
- the_lnet.ln_routing = 1;
- LNET_UNLOCK();
-
- return 0;
-
- failed:
- lnet_free_rtrpools();
- return rc;
+ nrbs /= LNET_CPT_NUMBER;
+ return max(nrbs, LNET_NRB_TINY_MIN);
}
-int
-lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
+static int
+lnet_nrb_small_calculate(void)
{
- lnet_peer_t *lp = NULL;
- cfs_time_t now = cfs_time_current();
-
- LASSERT (!cfs_in_interrupt ());
-
- CDEBUG (D_NET, "%s notifying %s: %s\n",
- (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
- libcfs_nid2str(nid),
- alive ? "up" : "down");
-
- if (ni != NULL &&
- LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
- CWARN ("Ignoring notification of %s %s by %s (different net)\n",
- libcfs_nid2str(nid), alive ? "birth" : "death",
- libcfs_nid2str(ni->ni_nid));
- return -EINVAL;
- }
-
- /* can't do predictions... */
- if (cfs_time_after(when, now)) {
- CWARN ("Ignoring prediction from %s of %s %s "
- "%ld 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)));
- return -EINVAL;
- }
-
- if (ni != NULL && !alive && /* LND telling me she's down */
- !auto_down) { /* auto-down disabled */
- CDEBUG(D_NET, "Auto-down disabled\n");
- return 0;
- }
-
- LNET_LOCK();
-
- lp = lnet_find_peer_locked(nid);
- if (lp == NULL) {
- /* nid not found */
- LNET_UNLOCK();
- CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
- return 0;
- }
-
- /* 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->lp_last_alive)
- when = lp->lp_last_alive;
-
- lnet_notify_locked(lp, ni == NULL, alive, when);
-
- LNET_UNLOCK();
-
- lnet_do_notify(lp);
-
- LNET_LOCK();
-
- lnet_peer_decref_locked(lp);
-
- LNET_UNLOCK();
- return 0;
+ int nrbs = LNET_NRB_SMALL;
+
+ if (small_router_buffers < 0) {
+ LCONSOLE_ERROR_MSG(0x10c,
+ "small_router_buffers=%d invalid when "
+ "routing enabled\n", small_router_buffers);
+ return -EINVAL;
+ }
+
+ if (small_router_buffers > 0)
+ nrbs = small_router_buffers;
+
+ nrbs /= LNET_CPT_NUMBER;
+ return max(nrbs, LNET_NRB_SMALL_MIN);
}
-EXPORT_SYMBOL(lnet_notify);
-void
-lnet_get_tunables (void)
+static int
+lnet_nrb_large_calculate(void)
{
- return;
-}
+ int nrbs = LNET_NRB_LARGE;
+
+ if (large_router_buffers < 0) {
+ LCONSOLE_ERROR_MSG(0x10c,
+ "large_router_buffers=%d invalid when "
+ "routing enabled\n", large_router_buffers);
+ return -EINVAL;
+ }
-#else
+ if (large_router_buffers > 0)
+ nrbs = large_router_buffers;
+
+ nrbs /= LNET_CPT_NUMBER;
+ return max(nrbs, LNET_NRB_LARGE_MIN);
+}
int
-lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
+lnet_rtrpools_alloc(int im_a_router)
{
- return -EOPNOTSUPP;
+ struct lnet_rtrbufpool *rtrp;
+ int nrb_tiny;
+ int nrb_small;
+ int nrb_large;
+ int rc;
+ int i;
+
+ if (!strcmp(forwarding, "")) {
+ /* not set either way */
+ if (!im_a_router)
+ return 0;
+ } else if (!strcmp(forwarding, "disabled")) {
+ /* explicitly disabled */
+ return 0;
+ } else if (!strcmp(forwarding, "enabled")) {
+ /* explicitly enabled */
+ } else {
+ LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
+ "'enabled' or 'disabled'\n");
+ return -EINVAL;
+ }
+
+ nrb_tiny = lnet_nrb_tiny_calculate();
+ if (nrb_tiny < 0)
+ return -EINVAL;
+
+ nrb_small = lnet_nrb_small_calculate();
+ if (nrb_small < 0)
+ return -EINVAL;
+
+ nrb_large = lnet_nrb_large_calculate();
+ if (nrb_large < 0)
+ return -EINVAL;
+
+ the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
+ LNET_NRBPOOLS *
+ sizeof(struct lnet_rtrbufpool));
+ if (the_lnet.ln_rtrpools == NULL) {
+ LCONSOLE_ERROR_MSG(0x10c,
+ "Failed to initialize router buffe pool\n");
+ return -ENOMEM;
+ }
+
+ cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
+ lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
+ rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
+ nrb_tiny, i);
+ if (rc != 0)
+ goto failed;
+
+ lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
+ LNET_NRB_SMALL_PAGES);
+ rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
+ nrb_small, i);
+ if (rc != 0)
+ goto failed;
+
+ lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
+ LNET_NRB_LARGE_PAGES);
+ rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
+ nrb_large, i);
+ if (rc != 0)
+ goto failed;
+ }
+
+ 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:
+ lnet_rtrpools_free(0);
+ return rc;
}
-void
-lnet_router_checker (void)
+static int
+lnet_rtrpools_adjust_helper(int tiny, int small, int large)
{
- static time_t last = 0;
- static int running = 0;
-
- time_t now = cfs_time_current_sec();
- int interval = now - last;
- int rc;
- __u64 version;
- lnet_peer_t *rtr;
-
- /* It's no use to call me again within a sec - all intervals and
- * timeouts are measured in seconds */
- if (last != 0 && interval < 2)
- return;
-
- if (last != 0 &&
- interval > MAX(live_router_check_interval,
- dead_router_check_interval))
- CDEBUG(D_NETERROR, "Checker(%d/%d) not called for %d seconds\n",
- live_router_check_interval, dead_router_check_interval,
- interval);
-
- LNET_LOCK();
- LASSERT (!running); /* recursion check */
- running = 1;
- LNET_UNLOCK();
-
- last = now;
-
- if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPTHREAD) {
- the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKING;
- rc = LNetMDUnlink(the_lnet.ln_rc_mdh);
- LASSERT (rc == 0);
- }
-
- /* consume all pending events */
- while (1) {
- int i;
- lnet_event_t ev;
-
- /* NB ln_rc_eqh must be the 1st in 'eventqs' otherwise the
- * recursion breaker in LNetEQPoll would fail */
- rc = LNetEQPoll(&the_lnet.ln_rc_eqh, 1, 0, &ev, &i);
- if (rc == 0) /* no event pending */
- break;
-
- /* NB a lost SENT prevents me from pinging a router again */
- if (rc == -EOVERFLOW) {
- CERROR("Dropped an event!!!\n");
- abort();
- }
-
- LASSERT (rc == 1);
-
- LNET_LOCK();
- lnet_router_checker_event(&ev);
- LNET_UNLOCK();
- }
-
- if (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKED ||
- the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKING) {
- running = 0;
- return;
- }
-
- LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
-
- LNET_LOCK();
-
- version = the_lnet.ln_routers_version;
- cfs_list_for_each_entry (rtr, &the_lnet.ln_routers, lp_rtr_list) {
- lnet_ping_router_locked(rtr);
- LASSERT (version == the_lnet.ln_routers_version);
- }
-
- LNET_UNLOCK();
-
- running = 0; /* lock only needed for the recursion check */
- return;
+ int nrb = 0;
+ int rc = 0;
+ int i;
+ struct lnet_rtrbufpool *rtrp;
+
+ /* If the provided values for each buffer pool are different than the
+ * configured values, we need to take action. */
+ if (tiny >= 0) {
+ tiny_router_buffers = tiny;
+ nrb = lnet_nrb_tiny_calculate();
+ cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
+ rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
+ nrb, i);
+ if (rc != 0)
+ return rc;
+ }
+ }
+ if (small >= 0) {
+ small_router_buffers = small;
+ nrb = lnet_nrb_small_calculate();
+ cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
+ rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
+ nrb, i);
+ if (rc != 0)
+ return rc;
+ }
+ }
+ if (large >= 0) {
+ large_router_buffers = large;
+ nrb = lnet_nrb_large_calculate();
+ cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
+ rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
+ nrb, i);
+ if (rc != 0)
+ return rc;
+ }
+ }
+
+ return 0;
}
-/* NB lnet_peers_start_down depends on me,
- * so must be called before any peer creation */
-void
-lnet_get_tunables (void)
+int
+lnet_rtrpools_adjust(int tiny, int small, int large)
{
- char *s;
-
- s = getenv("LNET_ROUTER_PING_TIMEOUT");
- if (s != NULL) router_ping_timeout = atoi(s);
-
- s = getenv("LNET_LIVE_ROUTER_CHECK_INTERVAL");
- if (s != NULL) live_router_check_interval = atoi(s);
+ /* this function doesn't revert the changes if adding new buffers
+ * failed. It's up to the user space caller to revert the
+ * changes. */
- s = getenv("LNET_DEAD_ROUTER_CHECK_INTERVAL");
- if (s != NULL) dead_router_check_interval = atoi(s);
+ if (!the_lnet.ln_routing)
+ return 0;
- /* This replaces old lnd_notify mechanism */
- check_routers_before_use = 1;
- if (dead_router_check_interval <= 0)
- dead_router_check_interval = 30;
+ return lnet_rtrpools_adjust_helper(tiny, small, large);
}
-void
-lnet_free_rtrpools (void)
+int
+lnet_rtrpools_enable(void)
{
+ int rc = 0;
+
+ if (the_lnet.ln_routing)
+ return 0;
+
+ if (the_lnet.ln_rtrpools == NULL)
+ /* If routing is turned off, and we have never
+ * initialized the pools before, just call the
+ * standard buffer pool allocation routine as
+ * if we are just configuring this for the first
+ * time. */
+ rc = lnet_rtrpools_alloc(1);
+ else
+ rc = lnet_rtrpools_adjust_helper(0, 0, 0);
+ if (rc != 0)
+ return rc;
+
+ lnet_net_lock(LNET_LOCK_EX);
+ the_lnet.ln_routing = 1;
+
+ the_lnet.ln_ping_target->pb_info.pi_features &=
+ ~LNET_PING_FEAT_RTE_DISABLED;
+ lnet_net_unlock(LNET_LOCK_EX);
+
+ return rc;
}
void
-lnet_init_rtrpools (void)
+lnet_rtrpools_disable(void)
{
+ if (!the_lnet.ln_routing)
+ return;
+
+ lnet_net_lock(LNET_LOCK_EX);
+ the_lnet.ln_routing = 0;
+ the_lnet.ln_ping_target->pb_info.pi_features |=
+ LNET_PING_FEAT_RTE_DISABLED;
+
+ tiny_router_buffers = 0;
+ small_router_buffers = 0;
+ large_router_buffers = 0;
+ lnet_net_unlock(LNET_LOCK_EX);
+ lnet_rtrpools_free(1);
}
int
-lnet_alloc_rtrpools (int im_a_arouter)
+lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, int alive, time64_t when)
{
- return 0;
+ struct lnet_peer_ni *lp = NULL;
+ time64_t now = ktime_get_seconds();
+ int cpt = lnet_cpt_of_nid(nid, ni);
+
+ LASSERT (!in_interrupt ());
+
+ CDEBUG (D_NET, "%s notifying %s: %s\n",
+ (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
+ libcfs_nid2str(nid),
+ alive ? "up" : "down");
+
+ if (ni != NULL &&
+ LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
+ CWARN("Ignoring notification of %s %s by %s (different net)\n",
+ libcfs_nid2str(nid), alive ? "birth" : "death",
+ libcfs_nid2str(ni->ni_nid));
+ return -EINVAL;
+ }
+
+ /* can't do predictions... */
+ if (when > now) {
+ CWARN("Ignoring prediction from %s of %s %s "
+ "%lld seconds in the future\n",
+ (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
+ libcfs_nid2str(nid), alive ? "up" : "down", when - now);
+ return -EINVAL;
+ }
+
+ if (ni != NULL && !alive && /* LND telling me she's down */
+ !auto_down) { /* auto-down disabled */
+ CDEBUG(D_NET, "Auto-down disabled\n");
+ return 0;
+ }
+
+ lnet_net_lock(cpt);
+
+ if (the_lnet.ln_state != LNET_STATE_RUNNING) {
+ lnet_net_unlock(cpt);
+ return -ESHUTDOWN;
+ }
+
+ lp = lnet_find_peer_ni_locked(nid);
+ if (lp == NULL) {
+ /* nid not found */
+ lnet_net_unlock(cpt);
+ 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);
+ }
+
+ /* 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);
+
+ if (ni != NULL)
+ lnet_ni_notify_locked(ni, lp);
+
+ lnet_peer_ni_decref_locked(lp);
+
+ lnet_net_unlock(cpt);
+ return 0;
}
-
-#endif
+EXPORT_SYMBOL(lnet_notify);