/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
* vim:expandtab:shiftwidth=8:tabstop=8:
*
- * lib/lib-eq.c
- * Library level Event queue management routines
+ * GPL HEADER START
+ *
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
- * Copyright (c) 2001-2003 Cluster File Systems, Inc.
- * Copyright (c) 2001-2002 Sandia National Laboratories
+ * 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.sf.net/projects/lustre/
+ * 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_PORTALS
-#include <portals/lib-p30.h>
+#define DEBUG_SUBSYSTEM S_LNET
+#include <lnet/lib-lnet.h>
-int
-lib_api_eq_alloc (nal_t *apinal, ptl_size_t count,
- ptl_eq_handler_t callback,
- ptl_handle_eq_t *handle)
+/**
+ * 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,
+ lnet_handle_eq_t *handle)
{
- lib_nal_t *nal = apinal->nal_data;
- lib_eq_t *eq;
- unsigned long flags;
- int rc;
+ 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 */
+ * apparent capacity at all times */
if (count != LOWEST_BIT_SET(count)) { /* not a power of 2 already */
do { /* knock off all but the top bit... */
}
if (count == 0) /* catch bad parameter / overflow on roundup */
- return (PTL_VAL_FAILED);
-
- eq = lib_eq_alloc (nal);
+ return (-EINVAL);
+
+ eq = lnet_eq_alloc();
if (eq == NULL)
- return (PTL_NO_SPACE);
+ return (-ENOMEM);
- PORTAL_ALLOC(eq->eq_events, count * sizeof(ptl_event_t));
+ LIBCFS_ALLOC(eq->eq_events, count * sizeof(lnet_event_t));
if (eq->eq_events == NULL) {
- LIB_LOCK(nal, flags);
- lib_eq_free (nal, eq);
- LIB_UNLOCK(nal, flags);
- }
+ LNET_LOCK();
+ lnet_eq_free (eq);
+ LNET_UNLOCK();
- if (nal->libnal_map != NULL) {
- struct iovec iov = {
- .iov_base = eq->eq_events,
- .iov_len = count * sizeof(ptl_event_t)};
-
- rc = nal->libnal_map(nal, 1, &iov, &eq->eq_addrkey);
- if (rc != PTL_OK) {
- LIB_LOCK(nal, flags);
- lib_eq_free (nal, eq);
- LIB_UNLOCK(nal, flags);
- return (rc);
- }
+ 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(ptl_event_t));
+ memset(eq->eq_events, 0, count * sizeof(lnet_event_t));
eq->eq_deq_seq = 1;
eq->eq_enq_seq = 1;
eq->eq_refcount = 0;
eq->eq_callback = callback;
- LIB_LOCK(nal, flags);
+ LNET_LOCK();
- lib_initialise_handle (nal, &eq->eq_lh, PTL_COOKIE_TYPE_EQ);
- list_add (&eq->eq_list, &nal->libnal_ni.ni_active_eqs);
+ lnet_initialise_handle (&eq->eq_lh, LNET_COOKIE_TYPE_EQ);
+ cfs_list_add (&eq->eq_list, &the_lnet.ln_active_eqs);
- LIB_UNLOCK(nal, flags);
+ LNET_UNLOCK();
- ptl_eq2handle(handle, nal, eq);
- return (PTL_OK);
+ lnet_eq2handle(handle, eq);
+ return (0);
}
-int
-lib_api_eq_free(nal_t *apinal, ptl_handle_eq_t *eqh)
+/**
+ * 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)
{
- lib_nal_t *nal = apinal->nal_data;
- lib_eq_t *eq;
+ lnet_eq_t *eq;
int size;
- ptl_event_t *events;
- void *addrkey;
- unsigned long flags;
+ lnet_event_t *events;
- LIB_LOCK(nal, flags);
+ LASSERT (the_lnet.ln_init);
+ LASSERT (the_lnet.ln_refcount > 0);
- eq = ptl_handle2eq(eqh, nal);
+ LNET_LOCK();
+
+ eq = lnet_handle2eq(&eqh);
if (eq == NULL) {
- LIB_UNLOCK(nal, flags);
- return (PTL_EQ_INVALID);
+ LNET_UNLOCK();
+ return (-ENOENT);
}
if (eq->eq_refcount != 0) {
- LIB_UNLOCK(nal, flags);
- return (PTL_EQ_IN_USE);
+ CDEBUG(D_NET, "Event queue (%d) busy on destroy.\n",
+ eq->eq_refcount);
+ LNET_UNLOCK();
+ return (-EBUSY);
}
/* stash for free after lock dropped */
events = eq->eq_events;
size = eq->eq_size;
- addrkey = eq->eq_addrkey;
- lib_invalidate_handle (nal, &eq->eq_lh);
- list_del (&eq->eq_list);
- lib_eq_free (nal, eq);
+ lnet_invalidate_handle (&eq->eq_lh);
+ cfs_list_del (&eq->eq_list);
+ lnet_eq_free (eq);
- LIB_UNLOCK(nal, flags);
+ LNET_UNLOCK();
- if (nal->libnal_unmap != NULL) {
- struct iovec iov = {
- .iov_base = events,
- .iov_len = size * sizeof(ptl_event_t)};
+ LIBCFS_FREE(events, size * sizeof (lnet_event_t));
- nal->libnal_unmap(nal, 1, &iov, &addrkey);
- }
-
- PORTAL_FREE(events, size * sizeof (ptl_event_t));
-
- return (PTL_OK);
+ return 0;
}
int
-lib_get_event (lib_eq_t *eq, ptl_event_t *ev)
+lib_get_event (lnet_eq_t *eq, lnet_event_t *ev)
{
- int new_index = eq->eq_deq_seq & (eq->eq_size - 1);
- ptl_event_t *new_event = &eq->eq_events[new_index];
- int rc;
+ 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);
- if (PTL_SEQ_GT (eq->eq_deq_seq, new_event->sequence)) {
- RETURN(PTL_EQ_EMPTY);
+ if (LNET_SEQ_GT (eq->eq_deq_seq, new_event->sequence)) {
+ RETURN(0);
}
/* We've got a new event... */
/* ...but did it overwrite an event we've not seen yet? */
if (eq->eq_deq_seq == new_event->sequence) {
- rc = PTL_OK;
+ rc = 1;
} else {
- CERROR("Event Queue Overflow: eq seq %lu ev seq %lu\n",
+ /* don't complain with CERROR: some EQs are sized small
+ * anyway; if it's important, the caller should complain */
+ CDEBUG(D_NET, "Event Queue Overflow: eq seq %lu ev seq %lu\n",
eq->eq_deq_seq, new_event->sequence);
- rc = PTL_EQ_DROPPED;
+ rc = -EOVERFLOW;
}
eq->eq_deq_seq = new_event->sequence + 1;
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,
+ event, &which);
+}
+/**
+ * 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
-lib_api_eq_poll (nal_t *apinal,
- ptl_handle_eq_t *eventqs, int neq, int timeout_ms,
- ptl_event_t *event, int *which)
+LNetEQWait (lnet_handle_eq_t eventq, lnet_event_t *event)
+{
+ int which;
+
+ return LNetEQPoll(&eventq, 1, LNET_TIME_FOREVER,
+ event, &which);
+}
+
+/**
+ * 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)
{
- lib_nal_t *nal = apinal->nal_data;
- lib_ni_t *ni = &nal->libnal_ni;
- unsigned long flags;
int i;
int rc;
#ifdef __KERNEL__
- wait_queue_t wq;
- unsigned long now;
+ cfs_waitlink_t wl;
+ cfs_time_t now;
#else
struct timeval then;
struct timeval now;
+# ifdef HAVE_LIBPTHREAD
struct timespec ts;
+# endif
+ lnet_ni_t *eqwaitni = the_lnet.ln_eqwaitni;
#endif
ENTRY;
- LIB_LOCK(nal, flags);
+ LASSERT (the_lnet.ln_init);
+ LASSERT (the_lnet.ln_refcount > 0);
+
+ if (neq < 1)
+ RETURN(-ENOENT);
+
+ LNET_LOCK();
for (;;) {
+#ifndef __KERNEL__
+ LNET_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_LOCK();
+#endif
for (i = 0; i < neq; i++) {
- lib_eq_t *eq = ptl_handle2eq(&eventqs[i], nal);
+ lnet_eq_t *eq = lnet_handle2eq(&eventqs[i]);
+
+ if (eq == NULL) {
+ LNET_UNLOCK();
+ RETURN(-ENOENT);
+ }
rc = lib_get_event (eq, event);
- if (rc != PTL_EQ_EMPTY) {
- LIB_UNLOCK(nal, flags);
+ if (rc != 0) {
+ LNET_UNLOCK();
*which = i;
RETURN(rc);
}
}
-
+
+#ifdef __KERNEL__
if (timeout_ms == 0) {
- LIB_UNLOCK (nal, flags);
- RETURN (PTL_EQ_EMPTY);
+ LNET_UNLOCK();
+ RETURN (0);
}
- /* Some architectures force us to do spin locking/unlocking
- * in the same stack frame, means we can abstract the
- * locking here */
-#ifdef __KERNEL__
- init_waitqueue_entry(&wq, current);
- set_current_state(TASK_INTERRUPTIBLE);
- add_wait_queue(&ni->ni_waitq, &wq);
+ cfs_waitlink_init(&wl);
+ cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
+ cfs_waitq_add(&the_lnet.ln_waitq, &wl);
- LIB_UNLOCK(nal, flags);
+ LNET_UNLOCK();
if (timeout_ms < 0) {
- schedule ();
+ cfs_waitq_wait (&wl, CFS_TASK_INTERRUPTIBLE);
} else {
- now = jiffies;
- schedule_timeout((timeout_ms * HZ)/1000);
- timeout_ms -= ((jiffies - now) * 1000)/HZ;
+ 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 -= (int)(tv.tv_sec * 1000 + tv.tv_usec / 1000);
if (timeout_ms < 0)
timeout_ms = 0;
}
-
- LIB_LOCK(nal, flags);
+
+ 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);
+ }
+
+# 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(&ni->ni_cond, &ni->ni_mutex);
+ 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 +
+ 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(&ni->ni_cond,
- &ni->ni_mutex, &ts);
-
+
+ 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
#endif
}
}