*
* Copyright (c) 2011, 2014, 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 */
/* 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,
* 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)
if (rnet != rnet2)
LIBCFS_FREE(rnet, sizeof(*rnet));
+ /* indicate to startup the router checker if configured */
+ wake_up(&the_lnet.ln_rc_waitq);
+
return rc;
}
rte->lr_downis = 0;
continue;
}
+ /* if @down is zero and this route is single-hop, it means
+ * we can't find NI for target network */
+ if (down == 0 && rte->lr_hops == 1)
+ down = 1;
+
rte->lr_downis = down;
}
}
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_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 = LNET_PID_LUSTRE;
-
- 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;
-#if defined(__KERNEL__) && defined(LNET_ROUTER)
+ /* Router Checker thread needs to run when routing is enabled in
+ * order to call lnet_update_ni_status_locked() */
+ if (the_lnet.ln_routing)
+ return true;
+
+ return !list_empty(&the_lnet.ln_routers) &&
+ (live_router_check_interval > 0 ||
+ dead_router_check_interval > 0);
+}
static int
lnet_router_checker(void *arg)
{
- lnet_peer_t *rtr;
+ lnet_peer_t *rtr;
struct list_head *entry;
- cfs_block_allsigs();
+ cfs_block_allsigs();
- LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
+ LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
- 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);
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(int keep_pools)
-{
-}
-
-int
-lnet_rtrpools_alloc(int im_a_arouter)
-{
- return 0;
-}
-
-#endif