Whamcloud - gitweb
LU-1425 build: make Lustre build ready for gcov
[fs/lustre-release.git] / lnet / lnet / lib-eq.c
index 5bae602..fec269a 100644 (file)
-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
+/*
+ * GPL HEADER START
  *
- * lib/lib-eq.c
- * Library level Event queue management routines
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
- *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 only,
+ * as published by the Free Software Foundation.
  *
- *   This file is part of Lustre, http://www.lustre.org
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License version 2 for more details (a copy is included
+ * in the LICENSE file that accompanied this code).
  *
- *   Lustre is free software; you can redistribute it and/or
- *   modify it under the terms of version 2 of the GNU General Public
- *   License as published by the Free Software Foundation.
+ * You should have received a copy of the GNU General Public License
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  *
- *   Lustre is distributed in the hope that it will be useful,
- *   but WITHOUT ANY WARRANTY; without even the implied warranty of
- *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- *   GNU General Public License for more details.
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
  *
- *   You should have received a copy of the GNU General Public License
- *   along with Lustre; if not, write to the Free Software
- *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ * GPL HEADER END
+ */
+/*
+ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Use is subject to license terms.
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ * Lustre is a trademark of Sun Microsystems, Inc.
+ *
+ * lnet/lnet/lib-eq.c
+ *
+ * Library level Event queue management routines
  */
 
 #define DEBUG_SUBSYSTEM S_LNET
 #include <lnet/lib-lnet.h>
 
+/**
+ * Create an event queue that has room for \a count number of events.
+ *
+ * The event queue is circular and older events will be overwritten by new
+ * ones if they are not removed in time by the user using the functions
+ * LNetEQGet(), LNetEQWait(), or LNetEQPoll(). It is up to the user to
+ * determine the appropriate size of the event queue to prevent this loss
+ * of events. Note that when EQ handler is specified in \a callback, no
+ * event loss can happen, since the handler is run for each event deposited
+ * into the EQ.
+ *
+ * \param count The number of events to be stored in the event queue. It
+ * will be rounded up to the next power of two.
+ * \param callback A handler function that runs when an event is deposited
+ * into the EQ. The constant value LNET_EQ_HANDLER_NONE can be used to
+ * indicate that no event handler is desired.
+ * \param handle On successful return, this location will hold a handle for
+ * the newly created EQ.
+ *
+ * \retval 0       On success.
+ * \retval -EINVAL If an parameter is not valid.
+ * \retval -ENOMEM If memory for the EQ can't be allocated.
+ *
+ * \see lnet_eq_handler_t for the discussion on EQ handler semantics.
+ */
 int
-LNetEQAlloc(unsigned int count, lnet_eq_handler_t callback, 
+LNetEQAlloc(unsigned int count, lnet_eq_handler_t callback,
             lnet_handle_eq_t *handle)
 {
         lnet_eq_t     *eq;
 
         LASSERT (the_lnet.ln_init);
         LASSERT (the_lnet.ln_refcount > 0);
-        
+
         /* We need count to be a power of 2 so that when eq_{enq,deq}_seq
          * overflow, they don't skip entries, so the queue has the same
-         * apparant capacity at all times */
-
-        if (count != LOWEST_BIT_SET(count)) {   /* not a power of 2 already */
-                do {                    /* knock off all but the top bit... */
-                        count &= ~LOWEST_BIT_SET (count);
-                } while (count != LOWEST_BIT_SET(count));
-
-                count <<= 1;                             /* ...and round up */
-        }
-
-        if (count == 0)        /* catch bad parameter / overflow on roundup */
-                return (-EINVAL);
-        
-        eq = lnet_eq_alloc();
-        if (eq == NULL)
-                return (-ENOMEM);
-
-        LIBCFS_ALLOC(eq->eq_events, count * sizeof(lnet_event_t));
-        if (eq->eq_events == NULL) {
-                LNET_LOCK();
-                lnet_eq_free (eq);
-                LNET_UNLOCK();
-
-                return -ENOMEM;
-        }
-
-        /* NB this resets all event sequence numbers to 0, to be earlier
-         * than eq_deq_seq */
-        memset(eq->eq_events, 0, count * sizeof(lnet_event_t));
+         * apparent capacity at all times */
+
+       count = cfs_power2_roundup(count);
+
+       if (callback != LNET_EQ_HANDLER_NONE && count != 0) {
+               CWARN("EQ callback is guaranteed to get every event, "
+                     "do you still want to set eqcount %d for polling "
+                     "event which will have locking overhead? "
+                     "Please contact with developer to confirm\n", count);
+       }
+
+       /* count can be 0 if only need callback, we can eliminate
+        * overhead of enqueue event */
+       if (count == 0 && callback == LNET_EQ_HANDLER_NONE)
+               return -EINVAL;
+
+       eq = lnet_eq_alloc();
+       if (eq == NULL)
+               return -ENOMEM;
+
+       if (count != 0) {
+               LIBCFS_ALLOC(eq->eq_events, count * sizeof(lnet_event_t));
+               if (eq->eq_events == NULL)
+                       goto failed;
+               /* NB allocator has set all event sequence numbers to 0,
+                * so all them should be earlier than eq_deq_seq */
+       }
 
         eq->eq_deq_seq = 1;
         eq->eq_enq_seq = 1;
         eq->eq_size = count;
-        eq->eq_refcount = 0;
         eq->eq_callback = callback;
 
-        LNET_LOCK();
+       eq->eq_refs = cfs_percpt_alloc(lnet_cpt_table(),
+                                      sizeof(*eq->eq_refs[0]));
+       if (eq->eq_refs == NULL)
+               goto failed;
 
-        lnet_initialise_handle (&eq->eq_lh, LNET_COOKIE_TYPE_EQ);
-        list_add (&eq->eq_list, &the_lnet.ln_active_eqs);
+       /* MUST hold both exclusive lnet_res_lock */
+       lnet_res_lock(LNET_LOCK_EX);
+       /* NB: hold lnet_eq_wait_lock for EQ link/unlink, so we can do
+        * both EQ lookup and poll event with only lnet_eq_wait_lock */
+       lnet_eq_wait_lock();
 
-        LNET_UNLOCK();
+       lnet_res_lh_initialize(&the_lnet.ln_eq_container, &eq->eq_lh);
+       cfs_list_add(&eq->eq_list, &the_lnet.ln_eq_container.rec_active);
 
-        lnet_eq2handle(handle, eq);
-        return (0);
+       lnet_eq_wait_unlock();
+       lnet_res_unlock(LNET_LOCK_EX);
+
+       lnet_eq2handle(handle, eq);
+       return 0;
+
+failed:
+       if (eq->eq_events != NULL)
+               LIBCFS_FREE(eq->eq_events, count * sizeof(lnet_event_t));
+
+       if (eq->eq_refs != NULL)
+               cfs_percpt_free(eq->eq_refs);
+
+       lnet_eq_free(eq);
+       return -ENOMEM;
 }
+EXPORT_SYMBOL(LNetEQAlloc);
 
+/**
+ * Release the resources associated with an event queue if it's idle;
+ * otherwise do nothing and it's up to the user to try again.
+ *
+ * \param eqh A handle for the event queue to be released.
+ *
+ * \retval 0 If the EQ is not in use and freed.
+ * \retval -ENOENT If \a eqh does not point to a valid EQ.
+ * \retval -EBUSY  If the EQ is still in use by some MDs.
+ */
 int
 LNetEQFree(lnet_handle_eq_t eqh)
 {
-        lnet_eq_t     *eq;
-        int            size;
-        lnet_event_t  *events;
-
-        LASSERT (the_lnet.ln_init);
-        LASSERT (the_lnet.ln_refcount > 0);
-        
-        LNET_LOCK();
+       struct lnet_eq  *eq;
+       lnet_event_t    *events = NULL;
+       int             **refs = NULL;
+       int             *ref;
+       int             rc = 0;
+       int             size = 0;
+       int             i;
+
+       LASSERT(the_lnet.ln_init);
+       LASSERT(the_lnet.ln_refcount > 0);
+
+       lnet_res_lock(LNET_LOCK_EX);
+       /* NB: hold lnet_eq_wait_lock for EQ link/unlink, so we can do
+        * both EQ lookup and poll event with only lnet_eq_wait_lock */
+       lnet_eq_wait_lock();
+
+       eq = lnet_handle2eq(&eqh);
+       if (eq == NULL) {
+               rc = -ENOENT;
+               goto out;
+       }
+
+       cfs_percpt_for_each(ref, i, eq->eq_refs) {
+               LASSERT(*ref >= 0);
+               if (*ref == 0)
+                       continue;
+
+               CDEBUG(D_NET, "Event equeue (%d: %d) busy on destroy.\n",
+                      i, *ref);
+               rc = -EBUSY;
+               goto out;
+       }
+
+       /* stash for free after lock dropped */
+       events  = eq->eq_events;
+       size    = eq->eq_size;
+       refs    = eq->eq_refs;
+
+       lnet_res_lh_invalidate(&eq->eq_lh);
+       cfs_list_del(&eq->eq_list);
+       lnet_eq_free_locked(eq);
+ out:
+       lnet_eq_wait_unlock();
+       lnet_res_unlock(LNET_LOCK_EX);
+
+       if (events != NULL)
+               LIBCFS_FREE(events, size * sizeof(lnet_event_t));
+       if (refs != NULL)
+               cfs_percpt_free(refs);
+
+       return rc;
+}
+EXPORT_SYMBOL(LNetEQFree);
 
-        eq = lnet_handle2eq(&eqh);
-        if (eq == NULL) {
-                LNET_UNLOCK();
-                return (-ENOENT);
-        }
+void
+lnet_eq_enqueue_event(lnet_eq_t *eq, lnet_event_t *ev)
+{
+       /* MUST called with resource lock hold but w/o lnet_eq_wait_lock */
+       int index;
 
-        if (eq->eq_refcount != 0) {
-                LNET_UNLOCK();
-                return (-EBUSY);
-        }
+       if (eq->eq_size == 0) {
+               LASSERT(eq->eq_callback != LNET_EQ_HANDLER_NONE);
+               eq->eq_callback(ev);
+               return;
+       }
 
-        /* stash for free after lock dropped */
-        events  = eq->eq_events;
-        size    = eq->eq_size;
+       lnet_eq_wait_lock();
+       ev->sequence = eq->eq_enq_seq++;
 
-        lnet_invalidate_handle (&eq->eq_lh);
-        list_del (&eq->eq_list);
-        lnet_eq_free (eq);
+       LASSERT(eq->eq_size == LOWEST_BIT_SET(eq->eq_size));
+       index = ev->sequence & (eq->eq_size - 1);
 
-        LNET_UNLOCK();
+       eq->eq_events[index] = *ev;
 
-        LIBCFS_FREE(events, size * sizeof (lnet_event_t));
+       if (eq->eq_callback != LNET_EQ_HANDLER_NONE)
+               eq->eq_callback(ev);
 
-        return 0;
+#ifdef __KERNEL__
+       /* Wake anyone waiting in LNetEQPoll() */
+       if (cfs_waitq_active(&the_lnet.ln_eq_waitq))
+               cfs_waitq_broadcast(&the_lnet.ln_eq_waitq);
+#else
+# ifndef HAVE_LIBPTHREAD
+       /* LNetEQPoll() calls into _the_ LND to wait for action */
+# else
+       /* Wake anyone waiting in LNetEQPoll() */
+       pthread_cond_broadcast(&the_lnet.ln_eq_cond);
+# endif
+#endif
+       lnet_eq_wait_unlock();
 }
 
 int
-lib_get_event (lnet_eq_t *eq, lnet_event_t *ev)
+lnet_eq_dequeue_event(lnet_eq_t *eq, lnet_event_t *ev)
 {
-        int           new_index = eq->eq_deq_seq & (eq->eq_size - 1);
-        lnet_event_t *new_event = &eq->eq_events[new_index];
-        int           rc;
-        ENTRY;
+       int             new_index = eq->eq_deq_seq & (eq->eq_size - 1);
+       lnet_event_t    *new_event = &eq->eq_events[new_index];
+       int             rc;
+       ENTRY;
 
-        CDEBUG(D_INFO, "event: %p, sequence: %lu, eq->size: %u\n",
-               new_event, eq->eq_deq_seq, eq->eq_size);
+       /* must called with lnet_eq_wait_lock hold */
+       if (LNET_SEQ_GT(eq->eq_deq_seq, new_event->sequence))
+               RETURN(0);
 
-        if (LNET_SEQ_GT (eq->eq_deq_seq, new_event->sequence)) {
-                RETURN(0);
-        }
+       /* We've got a new event... */
+       *ev = *new_event;
 
-        /* We've got a new event... */
-        *ev = *new_event;
+       CDEBUG(D_INFO, "event: %p, sequence: %lu, eq->size: %u\n",
+              new_event, eq->eq_deq_seq, eq->eq_size);
 
         /* ...but did it overwrite an event we've not seen yet? */
         if (eq->eq_deq_seq == new_event->sequence) {
@@ -154,16 +278,48 @@ lib_get_event (lnet_eq_t *eq, lnet_event_t *ev)
         RETURN(rc);
 }
 
-
+/**
+ * A nonblocking function that can be used to get the next event in an EQ.
+ * If an event handler is associated with the EQ, the handler will run before
+ * this function returns successfully. The event is removed from the queue.
+ *
+ * \param eventq A handle for the event queue.
+ * \param event On successful return (1 or -EOVERFLOW), this location will
+ * hold the next event in the EQ.
+ *
+ * \retval 0          No pending event in the EQ.
+ * \retval 1          Indicates success.
+ * \retval -ENOENT    If \a eventq does not point to a valid EQ.
+ * \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that
+ * at least one event between this event and the last event obtained from the
+ * EQ has been dropped due to limited space in the EQ.
+ */
 int
 LNetEQGet (lnet_handle_eq_t eventq, lnet_event_t *event)
 {
         int which;
 
-        return LNetEQPoll(&eventq, 1, 0, 
+        return LNetEQPoll(&eventq, 1, 0,
                          event, &which);
 }
+EXPORT_SYMBOL(LNetEQGet);
 
+/**
+ * Block the calling process until there is an event in the EQ.
+ * If an event handler is associated with the EQ, the handler will run before
+ * this function returns successfully. This function returns the next event
+ * in the EQ and removes it from the EQ.
+ *
+ * \param eventq A handle for the event queue.
+ * \param event On successful return (1 or -EOVERFLOW), this location will
+ * hold the next event in the EQ.
+ *
+ * \retval 1          Indicates success.
+ * \retval -ENOENT    If \a eventq does not point to a valid EQ.
+ * \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that
+ * at least one event between this event and the last event obtained from the
+ * EQ has been dropped due to limited space in the EQ.
+ */
 int
 LNetEQWait (lnet_handle_eq_t eventq, lnet_event_t *event)
 {
@@ -172,24 +328,189 @@ LNetEQWait (lnet_handle_eq_t eventq, lnet_event_t *event)
         return LNetEQPoll(&eventq, 1, LNET_TIME_FOREVER,
                          event, &which);
 }
+EXPORT_SYMBOL(LNetEQWait);
 
-int
-LNetEQPoll (lnet_handle_eq_t *eventqs, int neq, int timeout_ms,
-            lnet_event_t *event, int *which)
-{
-        int              i;
-        int              rc;
 #ifdef __KERNEL__
-        cfs_waitlink_t   wl;
-        cfs_time_t       now;
-#else
-        struct timeval   then;
-        struct timeval   now;
+
+static int
+lnet_eq_wait_locked(int *timeout_ms)
+{
+       int             tms = *timeout_ms;
+       int             wait;
+       cfs_waitlink_t  wl;
+       cfs_time_t      now;
+
+       if (tms == 0)
+               return -1; /* don't want to wait and no new event */
+
+       cfs_waitlink_init(&wl);
+       cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
+       cfs_waitq_add(&the_lnet.ln_eq_waitq, &wl);
+
+       lnet_eq_wait_unlock();
+
+       if (tms < 0) {
+               cfs_waitq_wait(&wl, CFS_TASK_INTERRUPTIBLE);
+
+       } else {
+               struct timeval tv;
+
+               now = cfs_time_current();
+               cfs_waitq_timedwait(&wl, CFS_TASK_INTERRUPTIBLE,
+                                   cfs_time_seconds(tms) / 1000);
+               cfs_duration_usec(cfs_time_sub(cfs_time_current(), now), &tv);
+               tms -= (int)(tv.tv_sec * 1000 + tv.tv_usec / 1000);
+               if (tms < 0) /* no more wait but may have new event */
+                       tms = 0;
+       }
+
+       wait = tms != 0; /* might need to call here again */
+       *timeout_ms = tms;
+
+       lnet_eq_wait_lock();
+       cfs_waitq_del(&the_lnet.ln_eq_waitq, &wl);
+
+       return wait;
+}
+
+#else /* !__KERNEL__ */
+
 # ifdef HAVE_LIBPTHREAD
-        struct timespec  ts;
+static void
+lnet_eq_cond_wait(struct timespec *ts)
+{
+       if (ts == NULL) {
+               pthread_cond_wait(&the_lnet.ln_eq_cond,
+                                 &the_lnet.ln_eq_wait_lock);
+       } else {
+               pthread_cond_timedwait(&the_lnet.ln_eq_cond,
+                                      &the_lnet.ln_eq_wait_lock, ts);
+       }
+}
 # endif
-        lnet_ni_t       *eqwaitni = the_lnet.ln_eqwaitni;
-#endif
+
+static int
+lnet_eq_wait_locked(int *timeout_ms)
+{
+       lnet_ni_t       *eq_waitni = NULL;
+       int             tms = *timeout_ms;
+       int             wait;
+       struct timeval  then;
+       struct timeval  now;
+
+       if (the_lnet.ln_eq_waitni != NULL) {
+               /* I have a single NI that I have to call into, to get
+                * events queued, or to block. */
+               lnet_eq_wait_unlock();
+
+               lnet_net_lock(0);
+               eq_waitni = the_lnet.ln_eq_waitni;
+               if (unlikely(eq_waitni == NULL)) {
+                       lnet_net_unlock(0);
+
+                       lnet_eq_wait_lock();
+                       return -1;
+               }
+
+               lnet_ni_addref_locked(eq_waitni, 0);
+               lnet_net_unlock(0);
+
+               if (tms <= 0) { /* even for tms == 0 */
+                       (eq_waitni->ni_lnd->lnd_wait)(eq_waitni, tms);
+
+               } else {
+                       gettimeofday(&then, NULL);
+
+                       (eq_waitni->ni_lnd->lnd_wait)(eq_waitni, tms);
+
+                       gettimeofday(&now, NULL);
+                       tms -= (now.tv_sec - then.tv_sec) * 1000 +
+                              (now.tv_usec - then.tv_usec) / 1000;
+                       if (tms < 0)
+                               tms = 0;
+               }
+
+               lnet_ni_decref(eq_waitni);
+               lnet_eq_wait_lock();
+       } else { /* w/o eq_waitni */
+# ifndef HAVE_LIBPTHREAD
+               /* If I'm single-threaded, LNET fails at startup if it can't
+                * set the_lnet.ln_eqwaitni correctly.  */
+               LBUG();
+# else /* HAVE_LIBPTHREAD */
+               struct timespec  ts;
+
+               if (tms == 0) /* don't want to wait and new event */
+                       return -1;
+
+               if (tms < 0) {
+                       lnet_eq_cond_wait(NULL);
+
+               } else {
+
+                       gettimeofday(&then, NULL);
+
+                       ts.tv_sec = then.tv_sec + tms / 1000;
+                       ts.tv_nsec = then.tv_usec * 1000 +
+                                    (tms % 1000) * 1000000;
+                       if (ts.tv_nsec >= 1000000000) {
+                               ts.tv_sec++;
+                               ts.tv_nsec -= 1000000000;
+                       }
+
+                       lnet_eq_cond_wait(&ts);
+
+                       gettimeofday(&now, NULL);
+                       tms -= (now.tv_sec - then.tv_sec) * 1000 +
+                              (now.tv_usec - then.tv_usec) / 1000;
+                       if (tms < 0)
+                               tms = 0;
+               }
+# endif /* HAVE_LIBPTHREAD */
+       }
+
+       wait = tms != 0;
+       *timeout_ms = tms;
+
+       return wait;
+}
+
+#endif /* __KERNEL__ */
+
+
+/**
+ * Block the calling process until there's an event from a set of EQs or
+ * timeout happens.
+ *
+ * If an event handler is associated with the EQ, the handler will run before
+ * this function returns successfully, in which case the corresponding event
+ * is consumed.
+ *
+ * LNetEQPoll() provides a timeout to allow applications to poll, block for a
+ * fixed period, or block indefinitely.
+ *
+ * \param eventqs,neq An array of EQ handles, and size of the array.
+ * \param timeout_ms Time in milliseconds to wait for an event to occur on
+ * one of the EQs. The constant LNET_TIME_FOREVER can be used to indicate an
+ * infinite timeout.
+ * \param event,which On successful return (1 or -EOVERFLOW), \a event will
+ * hold the next event in the EQs, and \a which will contain the index of the
+ * EQ from which the event was taken.
+ *
+ * \retval 0          No pending event in the EQs after timeout.
+ * \retval 1          Indicates success.
+ * \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that
+ * at least one event between this event and the last event obtained from the
+ * EQ indicated by \a which has been dropped due to limited space in the EQ.
+ * \retval -ENOENT    If there's an invalid handle in \a eventqs.
+ */
+int
+LNetEQPoll(lnet_handle_eq_t *eventqs, int neq, int timeout_ms,
+          lnet_event_t *event, int *which)
+{
+       int     wait = 1;
+       int     rc;
+       int     i;
         ENTRY;
 
         LASSERT (the_lnet.ln_init);
@@ -198,121 +519,50 @@ LNetEQPoll (lnet_handle_eq_t *eventqs, int neq, int timeout_ms,
         if (neq < 1)
                 RETURN(-ENOENT);
 
-        LNET_LOCK();
-
-        for (;;) {
-                for (i = 0; i < neq; i++) {
-                        lnet_eq_t *eq = lnet_handle2eq(&eventqs[i]);
-
-                        if (eq == NULL) {
-                                LNET_UNLOCK();
-                                RETURN(-ENOENT);
-                        }
-
-                        rc = lib_get_event (eq, event);
-                        if (rc != 0) {
-                                LNET_UNLOCK();
-                                *which = i;
-                                RETURN(rc);
-                        }
-                }
-                
-#ifdef __KERNEL__
-                if (timeout_ms == 0) {
-                        LNET_UNLOCK ();
-                        RETURN (0);
-                }
-
-                cfs_waitlink_init(&wl);
-                set_current_state(TASK_INTERRUPTIBLE);
-                cfs_waitq_add(&the_lnet.ln_waitq, &wl);
-
-                LNET_UNLOCK();
-
-                if (timeout_ms < 0) {
-                        cfs_waitq_wait (&wl, CFS_TASK_INTERRUPTIBLE);
-                } else { 
-                        struct timeval tv;
-
-                        now = cfs_time_current();
-                        cfs_waitq_timedwait(&wl, CFS_TASK_INTERRUPTIBLE,
-                                            cfs_time_seconds(timeout_ms)/1000);
-                        cfs_duration_usec(cfs_time_sub(cfs_time_current(), now), 
-                                          &tv); 
-                        timeout_ms -= tv.tv_sec * 1000 + tv.tv_usec / 1000;
-                        if (timeout_ms < 0)
-                                timeout_ms = 0;
-                }
-                
-                LNET_LOCK();
-                cfs_waitq_del(&the_lnet.ln_waitq, &wl);
-#else
-                if (eqwaitni != NULL) {
-                        /* I have a single NI that I have to call into, to get
-                         * events queued, or to block. */
-                        lnet_ni_addref_locked(eqwaitni);
-                        LNET_UNLOCK();
-
-                        if (timeout_ms <= 0) {
-                                (eqwaitni->ni_lnd->lnd_wait)(eqwaitni, timeout_ms);
-                        } else {
-                                gettimeofday(&then, NULL);
-
-                                (eqwaitni->ni_lnd->lnd_wait)(eqwaitni, timeout_ms);
-                                
-                                gettimeofday(&now, NULL);
-                                timeout_ms -= (now.tv_sec - then.tv_sec) * 1000 +
-                                              (now.tv_usec - then.tv_usec) / 1000;
-                                if (timeout_ms < 0)
-                                        timeout_ms = 0;
-                        }
-
-                        LNET_LOCK();
-                        lnet_ni_decref_locked(eqwaitni);
-
-                        /* don't call into eqwaitni again if timeout has
-                         * expired */
-                        if (timeout_ms == 0)
-                                eqwaitni = NULL;
-
-                        continue;               /* go back and check for events */
-                }
-
-                if (timeout_ms == 0) {
-                        LNET_UNLOCK();
-                        RETURN (0);
-                }
+       lnet_eq_wait_lock();
 
-# ifndef HAVE_LIBPTHREAD
-                /* If I'm single-threaded, LNET fails at startup if it can't
-                 * set the_lnet.ln_eqwaitni correctly.  */
-                LBUG();
-# else
-                if (timeout_ms < 0) {
-                        pthread_cond_wait(&the_lnet.ln_cond, 
-                                          &the_lnet.ln_lock);
-                } else {
-                        gettimeofday(&then, NULL);
-                        
-                        ts.tv_sec = then.tv_sec + timeout_ms/1000;
-                        ts.tv_nsec = then.tv_usec * 1000 + 
-                                     (timeout_ms%1000) * 1000000;
-                        if (ts.tv_nsec >= 1000000000) {
-                                ts.tv_sec++;
-                                ts.tv_nsec -= 1000000000;
-                        }
-                        
-                        pthread_cond_timedwait(&the_lnet.ln_cond,
-                                               &the_lnet.ln_lock, &ts);
-                        
-                        gettimeofday(&now, NULL);
-                        timeout_ms -= (now.tv_sec - then.tv_sec) * 1000 +
-                                      (now.tv_usec - then.tv_usec) / 1000;
-                        
-                        if (timeout_ms < 0)
-                                timeout_ms = 0;
-                }
-# endif
+       for (;;) {
+#ifndef __KERNEL__
+               lnet_eq_wait_unlock();
+
+               /* Recursion breaker */
+               if (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
+                   !LNetHandleIsEqual(eventqs[0], the_lnet.ln_rc_eqh))
+                       lnet_router_checker();
+
+               lnet_eq_wait_lock();
 #endif
-        }
+               for (i = 0; i < neq; i++) {
+                       lnet_eq_t *eq = lnet_handle2eq(&eventqs[i]);
+
+                       if (eq == NULL) {
+                               lnet_eq_wait_unlock();
+                               RETURN(-ENOENT);
+                       }
+
+                       rc = lnet_eq_dequeue_event(eq, event);
+                       if (rc != 0) {
+                               lnet_eq_wait_unlock();
+                               *which = i;
+                               RETURN(rc);
+                       }
+               }
+
+               if (wait == 0)
+                       break;
+
+               /*
+                * return value of lnet_eq_wait_locked:
+                * -1 : did nothing and it's sure no new event
+                *  1 : sleep inside and wait until new event
+                *  0 : don't want to wait anymore, but might have new event
+                *      so need to call dequeue again
+                */
+               wait = lnet_eq_wait_locked(&timeout_ms);
+               if (wait < 0) /* no new event */
+                       break;
+       }
+
+       lnet_eq_wait_unlock();
+       RETURN(0);
 }