/*
* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
*
- * Copyright (c) 2011, 2013, Intel Corporation.
+ * Copyright (c) 2011, 2015, Intel Corporation.
*
- * This file is part of Portals
- * http://sourceforge.net/projects/sandiaportals/
+ * This file is part of Lustre, https://wiki.hpdd.intel.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 <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_CACHE_SIZE - 1) >> \
+ PAGE_CACHE_SHIFT)
static char *forwarding = "";
CFS_MODULE_PARM(forwarding, "s", charp, 0444,
/* forward ref's */
static int lnet_router_checker(void *);
-#else
-
-int
-lnet_peer_buffer_credits(lnet_ni_t *ni)
-{
- return 0;
-}
-
-#endif
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");
-static int avoid_asym_router_failure = 1;
+int avoid_asym_router_failure = 1;
CFS_MODULE_PARM(avoid_asym_router_failure, "i", int, 0644,
"Avoid asymmetrical router failures (0 to disable)");
CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
}
-void
+static void
lnet_ni_notify_locked(lnet_ni_t *ni, lnet_peer_t *lp)
{
int alive;
* NB individual events can be missed; the only guarantee is that you
* always get the most recent news */
- if (lp->lp_notifying)
+ if (lp->lp_notifying || ni == NULL)
return;
lp->lp_notifying = 1;
static void lnet_shuffle_seed(void)
{
- static int seeded = 0;
- int lnd_type, seed[2];
- struct timeval tv;
- lnet_ni_t *ni;
+ static int seeded;
+ __u32 lnd_type;
+ __u32 seed[2];
+ struct timeval tv;
+ lnet_ni_t *ni;
struct list_head *tmp;
if (seeded)
}
/* 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(lnet_remotenet_t *rnet, lnet_route_t *route)
{
unsigned int len = 0;
unsigned int offset = 0;
lnet_add_route(__u32 net, unsigned int hops, lnet_nid_t gateway,
unsigned int priority)
{
- struct list_head *e;
- lnet_remotenet_t *rnet;
- lnet_remotenet_t *rnet2;
- lnet_route_t *route;
- lnet_ni_t *ni;
- int add_route;
- int rc;
+ struct list_head *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 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 < 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;
- }
+ 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 -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;
- route->lr_hops = hops;
+ rnet->lrn_net = net;
+ route->lr_hops = hops;
route->lr_net = net;
route->lr_priority = priority;
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 */
- } else {
- CERROR("Error %d creating route %s %d %s\n", rc,
- libcfs_net2str(net), hops,
- libcfs_nid2str(gateway));
- }
- return rc;
- }
+ 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;
+ }
- LASSERT (!the_lnet.ln_shutdown);
+ LASSERT(!the_lnet.ln_shutdown);
rnet2 = lnet_find_net_locked(net);
if (rnet2 == NULL) {
list_for_each(e, &rnet2->lrn_routes) {
lnet_route_t *route2 = list_entry(e, lnet_route_t, lr_list);
- if (route2->lr_gateway == route->lr_gateway) {
- add_route = 0;
- break;
- }
+ if (route2->lr_gateway == route->lr_gateway) {
+ add_route = 0;
+ break;
+ }
- /* our lookups must be true */
- LASSERT (route2->lr_gateway->lp_nid != gateway);
- }
+ /* our lookups must be true */
+ LASSERT(route2->lr_gateway->lp_nid != gateway);
+ }
- if (add_route) {
+ if (add_route) {
lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */
lnet_add_route_to_rnet(rnet2, route);
lnet_peer_decref_locked(route->lr_gateway);
lnet_net_unlock(LNET_LOCK_EX);
- if (!add_route)
+ rc = 0;
+
+ if (!add_route) {
+ rc = -EEXIST;
LIBCFS_FREE(route, sizeof(*route));
+ }
if (rnet != rnet2)
LIBCFS_FREE(rnet, sizeof(*rnet));
- return 0;
+ /* indicate to startup the router checker if configured */
+ wake_up(&the_lnet.ln_rc_waitq);
+
+ return rc;
}
int
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 i, rc = -ENOENT, lidx, 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);
+ lidx = idx;
+ cfs_percpt_for_each(rbp, j, the_lnet.ln_rtrpools) {
+ if (lidx-- == 0) {
+ rc = 0;
+ 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;
+ break;
+ }
+ }
+ lnet_net_unlock(LNET_LOCK_EX);
+ }
+
+ lnet_net_lock(LNET_LOCK_EX);
+ pool_cfg->pl_routing = the_lnet.ln_routing;
+ lnet_net_unlock(LNET_LOCK_EX);
+
+ return rc;
+}
+
int
lnet_get_route(int idx, __u32 *net, __u32 *hops,
lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
*hops = route->lr_hops;
*priority = route->lr_priority;
*gateway = route->lr_gateway->lp_nid;
- *alive = route->lr_gateway->lp_alive;
+ *alive = lnet_is_route_alive(route);
lnet_net_unlock(cpt);
return 0;
}
__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;
+ __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;
}
/**
{
lnet_ping_info_t *info = rcd->rcd_pinginfo;
struct lnet_peer *gw = rcd->rcd_gateway;
- lnet_route_t *rtr;
+ lnet_route_t *rte;
if (!gw->lp_alive)
return;
if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0)
return; /* can't carry NI status info */
- list_for_each_entry(rtr, &gw->lp_routes, lr_gwlist) {
+ list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
int down = 0;
int up = 0;
int i;
+ if ((gw->lp_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++) {
lnet_ni_status_t *stat = &info->pi_ni[i];
lnet_nid_t nid = stat->ns_nid;
}
if (stat->ns_status == LNET_NI_STATUS_UP) {
- if (LNET_NIDNET(nid) == rtr->lr_net) {
+ if (LNET_NIDNET(nid) == rte->lr_net) {
up = 1;
break;
}
}
if (up) { /* ignore downed NIs if NI for dest network is up */
- rtr->lr_downis = 0;
+ rte->lr_downis = 0;
continue;
}
- rtr->lr_downis = down;
+ /* 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;
}
}
lnet_net_unlock(lp->lp_cpt);
}
-void
+static void
lnet_wait_known_routerstate(void)
{
lnet_peer_t *rtr;
if (all_known)
return;
-#ifndef __KERNEL__
- lnet_router_checker();
-#endif
- cfs_pause(cfs_time_seconds(1));
- }
+ set_current_state(TASK_UNINTERRUPTIBLE);
+ schedule_timeout(cfs_time_seconds(1));
+ }
}
void
+lnet_router_ni_update_locked(lnet_peer_t *gw, __u32 net)
+{
+ lnet_route_t *rte;
+
+ if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0) {
+ list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
+ if (rte->lr_net == net) {
+ rte->lr_downis = 0;
+ break;
+ }
+ }
+ }
+}
+
+static void
lnet_update_ni_status_locked(void)
{
lnet_ni_t *ni;
}
}
-void
+static void
lnet_destroy_rc_data(lnet_rc_data_t *rcd)
{
LASSERT(list_empty(&rcd->rcd_list));
LIBCFS_FREE(rcd, sizeof(*rcd));
}
-lnet_rc_data_t *
+static lnet_rc_data_t *
lnet_create_rc_data_locked(lnet_peer_t *gateway)
{
lnet_rc_data_t *rcd = NULL;
lnet_handle_md_t mdh;
id.nid = rtr->lp_nid;
- id.pid = LUSTRE_SRV_LNET_PID;
+ id.pid = LNET_PID_LUSTRE;
CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
rtr->lp_ping_notsent = 1;
lnet_router_checker_start(void)
{
int rc;
- int eqsz;
-#ifdef __KERNEL__
+ int eqsz = 0;
struct task_struct *task;
-#else /* __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_net_lock(0);
-
- /* 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;
- 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_net_unlock(0);
-
- rc = LNetSetAsync(id, 1);
- if (rc != 0) {
- CWARN("LNetSetAsync %s failed: %d\n",
- libcfs_id2str(id), rc);
- return rc;
- }
-
- lnet_net_lock(0);
- /* NB router list doesn't change in userspace */
- LASSERT(version == the_lnet.ln_routers_version);
- }
-
- lnet_net_unlock(0);
-
- 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);
+ LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
if (check_routers_before_use &&
dead_router_check_interval <= 0) {
return -EINVAL;
}
- if (!the_lnet.ln_routing &&
- live_router_check_interval <= 0 &&
- dead_router_check_interval <= 0)
- return 0;
-
-#ifdef __KERNEL__
sema_init(&the_lnet.ln_rc_signal, 0);
- /* EQ size doesn't matter; the callback is guaranteed to get every
- * event */
- eqsz = 0;
- 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;
- }
- the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
-#ifdef __KERNEL__
+ 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);
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
void
lnet_router_checker_stop (void)
{
- int rc;
+ int rc;
- if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
- return;
+ if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
+ return;
- LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
+ 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);
-#ifdef __KERNEL__
/* block until event callback signals exit */
down(&the_lnet.ln_rc_signal);
-#else
- lnet_router_checker();
-#endif
LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
rc = LNetEQFree(the_lnet.ln_rc_eqh);
i++;
CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
"Waiting for rc buffers to unlink\n");
- cfs_pause(cfs_time_seconds(1) / 4);
+ 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
+ */
+static 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;
-#if defined(__KERNEL__) && defined(LNET_ROUTER)
+ return !list_empty(&the_lnet.ln_routers) &&
+ (live_router_check_interval > 0 ||
+ dead_router_check_interval > 0);
+}
static int
lnet_router_checker(void *arg)
{
- lnet_peer_t *rtr;
+ lnet_peer_t *rtr;
struct list_head *entry;
- cfs_block_allsigs();
-
- LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
+ cfs_block_allsigs();
- while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
+ while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
__u64 version;
int cpt;
int cpt2;
goto rescan;
}
- lnet_ping_router_locked(rtr);
+ lnet_ping_router_locked(rtr);
- /* NB dropped lock */
- if (version != the_lnet.ln_routers_version) {
- /* the routers list has changed */
- goto rescan;
- }
- }
+ /* NB dropped lock */
+ if (version != the_lnet.ln_routers_version) {
+ /* the routers list has changed */
+ goto rescan;
+ }
+ }
if (the_lnet.ln_routing)
lnet_update_ni_status_locked();
lnet_prune_rc_data(0); /* don't wait for UNLINK */
- /* Call cfs_pause() here always adds 1 to load average
+ /* Call schedule_timeout() here always adds 1 to load average
* because kernel counts # active tasks as nr_running
* + nr_uninterruptible. */
- schedule_timeout_and_set_state(TASK_INTERRUPTIBLE,
- cfs_time_seconds(1));
+ /* 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());
+ else
+ wait_event_interruptible_timeout(the_lnet.ln_rc_waitq,
+ false,
+ cfs_time_seconds(1));
}
- LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING);
-
lnet_prune_rc_data(1); /* wait for UNLINK */
the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
LIBCFS_FREE(rb, sz);
}
-lnet_rtrbuf_t *
+static lnet_rtrbuf_t *
lnet_new_rtrbuf(lnet_rtrbufpool_t *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;
+ int npages = rbp->rbp_npages;
+ int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
+ struct page *page;
+ lnet_rtrbuf_t *rb;
+ int i;
LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
if (rb == NULL)
return rb;
}
-void
-lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp)
+static void
+lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp, int cpt)
{
- int npages = rbp->rbp_npages;
- int nbuffers = 0;
- lnet_rtrbuf_t *rb;
+ int npages = rbp->rbp_npages;
+ lnet_rtrbuf_t *rb;
+ struct list_head tmp;
if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
return;
- LASSERT(list_empty(&rbp->rbp_msgs));
- LASSERT(rbp->rbp_credits == rbp->rbp_nbuffers);
+ INIT_LIST_HEAD(&tmp);
- while (!list_empty(&rbp->rbp_bufs)) {
- LASSERT(rbp->rbp_credits > 0);
+ lnet_net_lock(cpt);
+ lnet_drop_routed_msgs_locked(&rbp->rbp_msgs, 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);
- rb = list_entry(rbp->rbp_bufs.next,
- lnet_rtrbuf_t, rb_list);
+ /* Free buffers on the free list. */
+ while (!list_empty(&tmp)) {
+ rb = list_entry(tmp.next, lnet_rtrbuf_t, rb_list);
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
-lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
+static int
+lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
{
- lnet_rtrbuf_t *rb;
- int i;
+ struct list_head rb_list;
+ lnet_rtrbuf_t *rb;
+ int num_rb;
+ int num_buffers = 0;
+ int old_req_nbufs;
+ int npages = rbp->rbp_npages;
- if (rbp->rbp_nbuffers != 0) {
- LASSERT (rbp->rbp_nbuffers == nbufs);
- return 0;
- }
+ 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);
- for (i = 0; i < nbufs; i++) {
+ 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);
- if (rb == NULL) {
- CERROR("Failed to allocate %d router bufs of %d pages\n",
- nbufs, rbp->rbp_npages);
- return -ENOMEM;
- }
+ lnet_net_lock(cpt);
+ rbp->rbp_req_nbuffers = old_req_nbufs;
+ lnet_net_unlock(cpt);
- rbp->rbp_nbuffers++;
- rbp->rbp_credits++;
- rbp->rbp_mincredits++;
- list_add(&rb->rb_list, &rbp->rbp_bufs);
+ goto failed;
+ }
- /* No allocation "under fire" */
- /* Otherwise we'd need code to schedule blocked msgs etc */
- LASSERT(!the_lnet.ln_routing);
+ list_add(&rb->rb_list, &rb_list);
+ num_buffers++;
}
- LASSERT(rbp->rbp_credits == nbufs);
+ 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, lnet_rtrbuf_t, rb_list);
+ list_del(&rb->rb_list);
+ lnet_destroy_rtrbuf(rb, npages);
+ }
+
+ return -ENOMEM;
}
-void
+static void
lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
{
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_rtrpools_free(void)
+lnet_rtrpools_free(int keep_pools)
{
lnet_rtrbufpool_t *rtrp;
int i;
return;
cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
- lnet_rtrpool_free_bufs(&rtrp[0]);
- lnet_rtrpool_free_bufs(&rtrp[1]);
- lnet_rtrpool_free_bufs(&rtrp[2]);
+ 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);
}
- cfs_percpt_free(the_lnet.ln_rtrpools);
- the_lnet.ln_rtrpools = NULL;
+ if (!keep_pools) {
+ cfs_percpt_free(the_lnet.ln_rtrpools);
+ the_lnet.ln_rtrpools = NULL;
+ }
}
static int
-lnet_nrb_tiny_calculate(int npages)
+lnet_nrb_tiny_calculate(void)
{
int nrbs = LNET_NRB_TINY;
}
static int
-lnet_nrb_small_calculate(int npages)
+lnet_nrb_small_calculate(void)
{
int nrbs = LNET_NRB_SMALL;
}
static int
-lnet_nrb_large_calculate(int npages)
+lnet_nrb_large_calculate(void)
{
int nrbs = LNET_NRB_LARGE;
lnet_rtrpools_alloc(int im_a_router)
{
lnet_rtrbufpool_t *rtrp;
- int large_pages = (LNET_MTU + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
- int small_pages = 1;
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;
- }
+ 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(0);
+ nrb_tiny = lnet_nrb_tiny_calculate();
if (nrb_tiny < 0)
return -EINVAL;
- nrb_small = lnet_nrb_small_calculate(small_pages);
+ nrb_small = lnet_nrb_small_calculate();
if (nrb_small < 0)
return -EINVAL;
- nrb_large = lnet_nrb_large_calculate(large_pages);
+ nrb_large = lnet_nrb_large_calculate();
if (nrb_large < 0)
return -EINVAL;
}
cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
- lnet_rtrpool_init(&rtrp[0], 0);
- rc = lnet_rtrpool_alloc_bufs(&rtrp[0], nrb_tiny, i);
+ 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[1], small_pages);
- rc = lnet_rtrpool_alloc_bufs(&rtrp[1], nrb_small, i);
+ 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[2], large_pages);
- rc = lnet_rtrpool_alloc_bufs(&rtrp[2], nrb_large, i);
+ 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);
-
return 0;
failed:
- lnet_rtrpools_free();
+ lnet_rtrpools_free(0);
return rc;
}
+static int
+lnet_rtrpools_adjust_helper(int tiny, int small, int large)
+{
+ int nrb = 0;
+ int rc = 0;
+ int i;
+ lnet_rtrbufpool_t *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;
+}
+
+int
+lnet_rtrpools_adjust(int tiny, int small, int large)
+{
+ /* this function doesn't revert the changes if adding new buffers
+ * failed. It's up to the user space caller to revert the
+ * changes. */
+
+ if (!the_lnet.ln_routing)
+ return 0;
+
+ return lnet_rtrpools_adjust_helper(tiny, small, large);
+}
+
+int
+lnet_rtrpools_enable(void)
+{
+ int rc;
+
+ 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. */
+ return lnet_rtrpools_alloc(1);
+
+ 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_info->pi_features &= ~LNET_PING_FEAT_RTE_DISABLED;
+ lnet_net_unlock(LNET_LOCK_EX);
+
+ return 0;
+}
+
+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_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_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
{
return 0;
}
EXPORT_SYMBOL(lnet_notify);
-
-void
-lnet_get_tunables (void)
-{
- return;
-}
-
-#else
-
-int
-lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
-{
- return -EOPNOTSUPP;
-}
-
-void
-lnet_router_checker (void)
-{
- 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))
- CNETERR("Checker(%d/%d) not called for %d seconds\n",
- live_router_check_interval, dead_router_check_interval,
- interval);
-
- LASSERT(LNET_CPT_NUMBER == 1);
-
- lnet_net_lock(0);
- LASSERT(!running); /* recursion check */
- running = 1;
- lnet_net_unlock(0);
-
- last = now;
-
- if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING)
- lnet_prune_rc_data(0); /* unlink all rcd and nowait */
-
- /* 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_router_checker_event(&ev);
- }
-
- if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING) {
- lnet_prune_rc_data(1); /* release rcd */
- the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
- running = 0;
- return;
- }
-
- LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
-
- lnet_net_lock(0);
-
- version = the_lnet.ln_routers_version;
- 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_net_unlock(0);
-
- running = 0; /* lock only needed for the recursion check */
- return;
-}
-
-/* NB lnet_peers_start_down depends on me,
- * so must be called before any peer creation */
-void
-lnet_get_tunables (void)
-{
- 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);
-
- s = getenv("LNET_DEAD_ROUTER_CHECK_INTERVAL");
- if (s != NULL) dead_router_check_interval = atoi(s);
-
- /* This replaces old lnd_notify mechanism */
- check_routers_before_use = 1;
- if (dead_router_check_interval <= 0)
- dead_router_check_interval = 30;
-}
-
-void
-lnet_rtrpools_free(void)
-{
-}
-
-int
-lnet_rtrpools_alloc(int im_a_arouter)
-{
- return 0;
-}
-
-#endif