[deliverable/linux.git] / drivers / staging / lustre / lnet / lnet / lib-eq.c

/*
 * GPL HEADER START
 *
 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
 *
 * 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 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).
 *
 * 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
 *
 * 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.
 *
 * GPL HEADER END
 */
/*
 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
 * Use is subject to license terms.
 *
 * Copyright (c) 2012, Intel Corporation.
 */
/*
 * 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 <linux/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,
	    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
	 * 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_callback = callback;

	eq->eq_refs = cfs_percpt_alloc(lnet_cpt_table(),
				       sizeof(*eq->eq_refs[0]));
	if (eq->eq_refs == NULL)
		goto failed;

	/* 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_res_lh_initialize(&the_lnet.ln_eq_container, &eq->eq_lh);
	list_add(&eq->eq_list, &the_lnet.ln_eq_container.rec_active);

	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)
{
	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);
	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);

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_size == 0) {
		LASSERT(eq->eq_callback != LNET_EQ_HANDLER_NONE);
		eq->eq_callback(ev);
		return;
	}

	lnet_eq_wait_lock();
	ev->sequence = eq->eq_enq_seq++;

	LASSERT(eq->eq_size == LOWEST_BIT_SET(eq->eq_size));
	index = ev->sequence & (eq->eq_size - 1);

	eq->eq_events[index] = *ev;

	if (eq->eq_callback != LNET_EQ_HANDLER_NONE)
		eq->eq_callback(ev);

	/* Wake anyone waiting in LNetEQPoll() */
	if (waitqueue_active(&the_lnet.ln_eq_waitq))
		wake_up_all(&the_lnet.ln_eq_waitq);
	lnet_eq_wait_unlock();
}

int
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;

	/* must called with lnet_eq_wait_lock hold */
	if (LNET_SEQ_GT(eq->eq_deq_seq, new_event->sequence))
		return 0;

	/* 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) {
		rc = 1;
	} else {
		/* 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 = -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);
}
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)
{
	int which;

	return LNetEQPoll(&eventq, 1, LNET_TIME_FOREVER,
			 event, &which);
}
EXPORT_SYMBOL(LNetEQWait);


static int
lnet_eq_wait_locked(int *timeout_ms)
{
	int		tms = *timeout_ms;
	int		wait;
	wait_queue_t  wl;
	cfs_time_t      now;

	if (tms == 0)
		return -1; /* don't want to wait and no new event */

	init_waitqueue_entry_current(&wl);
	set_current_state(TASK_INTERRUPTIBLE);
	add_wait_queue(&the_lnet.ln_eq_waitq, &wl);

	lnet_eq_wait_unlock();

	if (tms < 0) {
		waitq_wait(&wl, TASK_INTERRUPTIBLE);

	} else {
		struct timeval tv;

		now = cfs_time_current();
		waitq_timedwait(&wl, 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();
	remove_wait_queue(&the_lnet.ln_eq_waitq, &wl);

	return wait;
}


/**
 * 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;

	LASSERT (the_lnet.ln_init);
	LASSERT (the_lnet.ln_refcount > 0);

	if (neq < 1)
		return -ENOENT;

	lnet_eq_wait_lock();

	for (;;) {
		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;
}
Commit	Line	Data
d7e09d03 PT	1	/*
	2	* GPL HEADER START
	3	*
	4	* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
	5	*
	6	* This program is free software; you can redistribute it and/or modify
	7	* it under the terms of the GNU General Public License version 2 only,
	8	* as published by the Free Software Foundation.
	9	*
	10	* This program is distributed in the hope that it will be useful, but
	11	* WITHOUT ANY WARRANTY; without even the implied warranty of
	12	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
	13	* General Public License version 2 for more details (a copy is included
	14	* in the LICENSE file that accompanied this code).
	15	*
	16	* You should have received a copy of the GNU General Public License
	17	* version 2 along with this program; If not, see
	18	* http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
	19	*
	20	* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
	21	* CA 95054 USA or visit www.sun.com if you need additional information or
	22	* have any questions.
	23	*
	24	* GPL HEADER END
	25	*/
	26	/*
	27	* Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
	28	* Use is subject to license terms.
	29	*
	30	* Copyright (c) 2012, Intel Corporation.
	31	*/
	32	/*
	33	* This file is part of Lustre, http://www.lustre.org/
	34	* Lustre is a trademark of Sun Microsystems, Inc.
	35	*
	36	* lnet/lnet/lib-eq.c
	37	*
	38	* Library level Event queue management routines
	39	*/
	40
	41	#define DEBUG_SUBSYSTEM S_LNET
	42	#include <linux/lnet/lib-lnet.h>
	43
	44	/**
	45	* Create an event queue that has room for \a count number of events.
	46	*
	47	* The event queue is circular and older events will be overwritten by new
	48	* ones if they are not removed in time by the user using the functions
	49	* LNetEQGet(), LNetEQWait(), or LNetEQPoll(). It is up to the user to
	50	* determine the appropriate size of the event queue to prevent this loss
	51	* of events. Note that when EQ handler is specified in \a callback, no
	52	* event loss can happen, since the handler is run for each event deposited
	53	* into the EQ.
	54	*
	55	* \param count The number of events to be stored in the event queue. It
	56	* will be rounded up to the next power of two.
	57	* \param callback A handler function that runs when an event is deposited
	58	* into the EQ. The constant value LNET_EQ_HANDLER_NONE can be used to
	59	* indicate that no event handler is desired.
	60	* \param handle On successful return, this location will hold a handle for
	61	* the newly created EQ.
	62	*
	63	* \retval 0 On success.
	64	* \retval -EINVAL If an parameter is not valid.
65	* \retval -ENOMEM If memory for the EQ can't be allocated.
66	*
67	* \see lnet_eq_handler_t for the discussion on EQ handler semantics.
68	*/
69	int
70	LNetEQAlloc(unsigned int count, lnet_eq_handler_t callback,
71	lnet_handle_eq_t *handle)
72	{
73	lnet_eq_t *eq;
74
75	LASSERT (the_lnet.ln_init);
76	LASSERT (the_lnet.ln_refcount > 0);
77
78	/* We need count to be a power of 2 so that when eq_{enq,deq}_seq
79	* overflow, they don't skip entries, so the queue has the same
80	* apparent capacity at all times */
81
82	count = cfs_power2_roundup(count);
83
84	if (callback != LNET_EQ_HANDLER_NONE && count != 0) {
85	CWARN("EQ callback is guaranteed to get every event, "
86	"do you still want to set eqcount %d for polling "
87	"event which will have locking overhead? "
88	"Please contact with developer to confirm\n", count);
89	}
90
91	/* count can be 0 if only need callback, we can eliminate
92	* overhead of enqueue event */
93	if (count == 0 && callback == LNET_EQ_HANDLER_NONE)
94	return -EINVAL;
95
96	eq = lnet_eq_alloc();
97	if (eq == NULL)
98	return -ENOMEM;
99
100	if (count != 0) {
101	LIBCFS_ALLOC(eq->eq_events, count * sizeof(lnet_event_t));
102	if (eq->eq_events == NULL)
103	goto failed;
104	/* NB allocator has set all event sequence numbers to 0,
105	* so all them should be earlier than eq_deq_seq */
106	}
107
108	eq->eq_deq_seq = 1;
109	eq->eq_enq_seq = 1;
110	eq->eq_size = count;
111	eq->eq_callback = callback;
112
113	eq->eq_refs = cfs_percpt_alloc(lnet_cpt_table(),
114	sizeof(*eq->eq_refs[0]));
115	if (eq->eq_refs == NULL)
116	goto failed;
117
118	/* MUST hold both exclusive lnet_res_lock */
119	lnet_res_lock(LNET_LOCK_EX);
120	/* NB: hold lnet_eq_wait_lock for EQ link/unlink, so we can do
121	* both EQ lookup and poll event with only lnet_eq_wait_lock */
122	lnet_eq_wait_lock();
123
124	lnet_res_lh_initialize(&the_lnet.ln_eq_container, &eq->eq_lh);
125	list_add(&eq->eq_list, &the_lnet.ln_eq_container.rec_active);
126
127	lnet_eq_wait_unlock();
128	lnet_res_unlock(LNET_LOCK_EX);
129
130	lnet_eq2handle(handle, eq);
131	return 0;
132
133	failed:
134	if (eq->eq_events != NULL)
135	LIBCFS_FREE(eq->eq_events, count * sizeof(lnet_event_t));
136
137	if (eq->eq_refs != NULL)
138	cfs_percpt_free(eq->eq_refs);
139
140	lnet_eq_free(eq);
141	return -ENOMEM;
142	}
143	EXPORT_SYMBOL(LNetEQAlloc);
144
145	/**
146	* Release the resources associated with an event queue if it's idle;
147	* otherwise do nothing and it's up to the user to try again.
148	*
149	* \param eqh A handle for the event queue to be released.
150	*
151	* \retval 0 If the EQ is not in use and freed.
152	* \retval -ENOENT If \a eqh does not point to a valid EQ.
153	* \retval -EBUSY If the EQ is still in use by some MDs.
154	*/
155	int
156	LNetEQFree(lnet_handle_eq_t eqh)
157	{
158	struct lnet_eq *eq;
159	lnet_event_t *events = NULL;
160	int **refs = NULL;
161	int *ref;
162	int rc = 0;
163	int size = 0;
164	int i;
165
166	LASSERT(the_lnet.ln_init);
167	LASSERT(the_lnet.ln_refcount > 0);
168
169	lnet_res_lock(LNET_LOCK_EX);
170	/* NB: hold lnet_eq_wait_lock for EQ link/unlink, so we can do
171	* both EQ lookup and poll event with only lnet_eq_wait_lock */
172	lnet_eq_wait_lock();
173
174	eq = lnet_handle2eq(&eqh);
175	if (eq == NULL) {
176	rc = -ENOENT;
177	goto out;
178	}
179
180	cfs_percpt_for_each(ref, i, eq->eq_refs) {
181	LASSERT(*ref >= 0);
182	if (*ref == 0)
183	continue;
184
185	CDEBUG(D_NET, "Event equeue (%d: %d) busy on destroy.\n",
186	i, *ref);
187	rc = -EBUSY;
188	goto out;
189	}
190
191	/* stash for free after lock dropped */
192	events = eq->eq_events;
193	size = eq->eq_size;
194	refs = eq->eq_refs;
195
196	lnet_res_lh_invalidate(&eq->eq_lh);
197	list_del(&eq->eq_list);
198	lnet_eq_free_locked(eq);
199	out:
200	lnet_eq_wait_unlock();
201	lnet_res_unlock(LNET_LOCK_EX);
202
203	if (events != NULL)
204	LIBCFS_FREE(events, size * sizeof(lnet_event_t));
205	if (refs != NULL)
206	cfs_percpt_free(refs);
207
208	return rc;
209	}
210	EXPORT_SYMBOL(LNetEQFree);
211
212	void
213	lnet_eq_enqueue_event(lnet_eq_t eq, lnet_event_t ev)
214	{
215	/* MUST called with resource lock hold but w/o lnet_eq_wait_lock */
216	int index;
217
218	if (eq->eq_size == 0) {
219	LASSERT(eq->eq_callback != LNET_EQ_HANDLER_NONE);
220	eq->eq_callback(ev);
221	return;
222	}
223
224	lnet_eq_wait_lock();
225	ev->sequence = eq->eq_enq_seq++;
226
227	LASSERT(eq->eq_size == LOWEST_BIT_SET(eq->eq_size));
228	index = ev->sequence & (eq->eq_size - 1);
229
230	eq->eq_events[index] = *ev;
231
232	if (eq->eq_callback != LNET_EQ_HANDLER_NONE)
233	eq->eq_callback(ev);
234
235	/* Wake anyone waiting in LNetEQPoll() */
236	if (waitqueue_active(&the_lnet.ln_eq_waitq))
237	wake_up_all(&the_lnet.ln_eq_waitq);
238	lnet_eq_wait_unlock();
239	}
240
241	int
242	lnet_eq_dequeue_event(lnet_eq_t eq, lnet_event_t ev)
243	{
244	int new_index = eq->eq_deq_seq & (eq->eq_size - 1);
245	lnet_event_t *new_event = &eq->eq_events[new_index];
246	int rc;
d7e09d03 PT	247
	248	/* must called with lnet_eq_wait_lock hold */
	249	if (LNET_SEQ_GT(eq->eq_deq_seq, new_event->sequence))
0a3bdb00	250	return 0;
d7e09d03 PT	251
	252	/* We've got a new event... */
	253	ev = new_event;
	254
	255	CDEBUG(D_INFO, "event: %p, sequence: %lu, eq->size: %u\n",
	256	new_event, eq->eq_deq_seq, eq->eq_size);
	257
	258	/* ...but did it overwrite an event we've not seen yet? */
	259	if (eq->eq_deq_seq == new_event->sequence) {
	260	rc = 1;
	261	} else {
	262	/* don't complain with CERROR: some EQs are sized small
	263	* anyway; if it's important, the caller should complain */
	264	CDEBUG(D_NET, "Event Queue Overflow: eq seq %lu ev seq %lu\n",
	265	eq->eq_deq_seq, new_event->sequence);
	266	rc = -EOVERFLOW;
	267	}
	268
	269	eq->eq_deq_seq = new_event->sequence + 1;
0a3bdb00	270	return rc;
d7e09d03 PT	271	}
	272
	273	/**
	274	* A nonblocking function that can be used to get the next event in an EQ.
	275	* If an event handler is associated with the EQ, the handler will run before
	276	* this function returns successfully. The event is removed from the queue.
	277	*
	278	* \param eventq A handle for the event queue.
	279	* \param event On successful return (1 or -EOVERFLOW), this location will
	280	* hold the next event in the EQ.
	281	*
	282	* \retval 0 No pending event in the EQ.
	283	* \retval 1 Indicates success.
	284	* \retval -ENOENT If \a eventq does not point to a valid EQ.
	285	* \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that
	286	* at least one event between this event and the last event obtained from the
	287	* EQ has been dropped due to limited space in the EQ.
	288	*/
	289	int
	290	LNetEQGet (lnet_handle_eq_t eventq, lnet_event_t *event)
	291	{
	292	int which;
	293
	294	return LNetEQPoll(&eventq, 1, 0,
	295	event, &which);
	296	}
	297	EXPORT_SYMBOL(LNetEQGet);
	298
	299	/**
	300	* Block the calling process until there is an event in the EQ.
	301	* If an event handler is associated with the EQ, the handler will run before
	302	* this function returns successfully. This function returns the next event
	303	* in the EQ and removes it from the EQ.
	304	*
	305	* \param eventq A handle for the event queue.
	306	* \param event On successful return (1 or -EOVERFLOW), this location will
	307	* hold the next event in the EQ.
	308	*
	309	* \retval 1 Indicates success.
	310	* \retval -ENOENT If \a eventq does not point to a valid EQ.
	311	* \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that
	312	* at least one event between this event and the last event obtained from the
	313	* EQ has been dropped due to limited space in the EQ.
	314	*/
	315	int
	316	LNetEQWait (lnet_handle_eq_t eventq, lnet_event_t *event)
	317	{
	318	int which;
	319
	320	return LNetEQPoll(&eventq, 1, LNET_TIME_FOREVER,
	321	event, &which);
	322	}
	323	EXPORT_SYMBOL(LNetEQWait);
	324
	325
	326	static int
	327	lnet_eq_wait_locked(int *timeout_ms)
	328	{
	329	int tms = *timeout_ms;
	330	int wait;
	331	wait_queue_t wl;
	332	cfs_time_t now;
	333
	334	if (tms == 0)
335	return -1; /* don't want to wait and no new event */
336
337	init_waitqueue_entry_current(&wl);
338	set_current_state(TASK_INTERRUPTIBLE);
339	add_wait_queue(&the_lnet.ln_eq_waitq, &wl);
340
341	lnet_eq_wait_unlock();
342
343	if (tms < 0) {
344	waitq_wait(&wl, TASK_INTERRUPTIBLE);
345
346	} else {
347	struct timeval tv;
348
349	now = cfs_time_current();
350	waitq_timedwait(&wl, TASK_INTERRUPTIBLE,
351	cfs_time_seconds(tms) / 1000);
352	cfs_duration_usec(cfs_time_sub(cfs_time_current(), now), &tv);
353	tms -= (int)(tv.tv_sec * 1000 + tv.tv_usec / 1000);
354	if (tms < 0) /* no more wait but may have new event */
355	tms = 0;
356	}
357
358	wait = tms != 0; /* might need to call here again */
359	*timeout_ms = tms;
360
361	lnet_eq_wait_lock();
362	remove_wait_queue(&the_lnet.ln_eq_waitq, &wl);
363
364	return wait;
365	}
366
367
368
369	/**
370	* Block the calling process until there's an event from a set of EQs or
371	* timeout happens.
372	*
373	* If an event handler is associated with the EQ, the handler will run before
374	* this function returns successfully, in which case the corresponding event
375	* is consumed.
376	*
377	* LNetEQPoll() provides a timeout to allow applications to poll, block for a
378	* fixed period, or block indefinitely.
379	*
380	* \param eventqs,neq An array of EQ handles, and size of the array.
381	* \param timeout_ms Time in milliseconds to wait for an event to occur on
382	* one of the EQs. The constant LNET_TIME_FOREVER can be used to indicate an
383	* infinite timeout.
384	* \param event,which On successful return (1 or -EOVERFLOW), \a event will
385	* hold the next event in the EQs, and \a which will contain the index of the
386	* EQ from which the event was taken.
387	*
388	* \retval 0 No pending event in the EQs after timeout.
389	* \retval 1 Indicates success.
390	* \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that
391	* at least one event between this event and the last event obtained from the
392	* EQ indicated by \a which has been dropped due to limited space in the EQ.
393	* \retval -ENOENT If there's an invalid handle in \a eventqs.
394	*/
395	int
396	LNetEQPoll(lnet_handle_eq_t *eventqs, int neq, int timeout_ms,
397	lnet_event_t event, int which)
398	{
399	int wait = 1;
400	int rc;
401	int i;
d7e09d03 PT	402
	403	LASSERT (the_lnet.ln_init);
	404	LASSERT (the_lnet.ln_refcount > 0);
	405
	406	if (neq < 1)
0a3bdb00	407	return -ENOENT;
d7e09d03 PT	408
	409	lnet_eq_wait_lock();
	410
	411	for (;;) {
	412	for (i = 0; i < neq; i++) {
	413	lnet_eq_t *eq = lnet_handle2eq(&eventqs[i]);
	414
	415	if (eq == NULL) {
	416	lnet_eq_wait_unlock();
0a3bdb00	417	return -ENOENT;
d7e09d03 PT	418	}
	419
	420	rc = lnet_eq_dequeue_event(eq, event);
	421	if (rc != 0) {
	422	lnet_eq_wait_unlock();
	423	*which = i;
0a3bdb00	424	return rc;
d7e09d03 PT	425	}
	426	}
	427
	428	if (wait == 0)
	429	break;
	430
	431	/*
	432	* return value of lnet_eq_wait_locked:
	433	* -1 : did nothing and it's sure no new event
	434	* 1 : sleep inside and wait until new event
	435	* 0 : don't want to wait anymore, but might have new event
	436	* so need to call dequeue again
	437	*/
	438	wait = lnet_eq_wait_locked(&timeout_ms);
	439	if (wait < 0) /* no new event */
	440	break;
	441	}
	442
	443	lnet_eq_wait_unlock();
0a3bdb00	444	return 0;
d7e09d03	445	}