[userspace-rcu.git] / urcu / workqueue-fifo.h

#ifndef _URCU_WORKQUEUE_FIFO_H
#define _URCU_WORKQUEUE_FIFO_H

/*
 * urcu/workqueue-fifo.h
 *
 * Userspace RCU library - work queue scheme with FIFO semantic
 *
 * Copyright (c) 2014 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
 *
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * This library 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
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with this library; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

#include <urcu/uatomic.h>
#include <urcu/lfstack.h>
#include <urcu/waitqueue-lifo.h>
#include <urcu/wfcqueue.h>
#include <urcu/rculist.h>
#include <pthread.h>
#include <assert.h>

enum urcu_accept_ret {
	URCU_ACCEPT_WORK	= 0,
	URCU_ACCEPT_SHUTDOWN	= 1,
};

/*
 * We use RCU to steal work from siblings. Therefore, one of RCU flavors
 * need to be included before this header. All worker that participate
 * in stealing (initialized with the URCU_WORKER_STEAL flag) need to be
 * registered RCU readers threads.
 */

struct urcu_work {
	struct cds_wfcq_node node;
};

struct urcu_workqueue {
	/* FIFO work queue */
	struct __cds_wfcq_head head;
	struct cds_wfcq_tail tail;

	/* Associated wait queue for LIFO wait/wakeup */
	struct urcu_wait_queue waitqueue;

	/* RCU linked list head of siblings for work stealing. */
	struct cds_list_head sibling_head;
	pthread_mutex_t sibling_lock;	/* Protect sibling list updates */

	bool shutdown;			/* Shutdown performed */
};

struct urcu_worker {
	/* Workqueue which can be either used by worker, or stolen. */
	struct cds_wfcq_head head;
	struct cds_wfcq_tail tail;

	/* Work belonging to worker. Cannot be stolen. */
	struct urcu_work *own;

	struct urcu_wait_node wait_node;
	/* RCU linked list node of siblings for work stealing. */
	struct cds_list_head sibling_node;
	int flags;	/* enum urcu_worker_flags */
};

enum urcu_worker_flags {
	URCU_WORKER_STEAL	= (1 << 0),
};

static inline
void urcu_workqueue_init(struct urcu_workqueue *queue)
{
	__cds_wfcq_init(&queue->head, &queue->tail);
	urcu_wait_queue_init(&queue->waitqueue);
	CDS_INIT_LIST_HEAD(&queue->sibling_head);
	pthread_mutex_init(&queue->sibling_lock, NULL);
	queue->shutdown = false;
}

static inline
void urcu_queue_work(struct urcu_workqueue *queue, struct urcu_work *work)
{
	bool was_empty;

	cds_wfcq_node_init(&work->node);

	/* Enqueue work. */
	was_empty = !cds_wfcq_enqueue(&queue->head, &queue->tail,
			&work->node);
	/*
	 * If workqueue was previously empty, wakeup one worker thread.
	 * It will eventually grab the entire content of the work-queue
	 * (therefore grabbing a "work batch"). After having grabbed the
	 * work batch, while that thread is running and taking care of
	 * that work batch, when we enqueue more work, we will wake
	 * another thread (if there is one waiting), which will
	 * eventually grab the new batch, and so on. This scheme ensures
	 * that contiguous batch of work are handled by the same thread
	 * (for locality), and also ensures that we scale work to many
	 * worker threads when threads are busy enough to still be
	 * running when work is enqueued.
	 */
	if (was_empty) {
		rcu_read_lock();	/* Protect stack dequeue */
		(void) urcu_dequeue_wake_single(&queue->waitqueue);
		rcu_read_unlock();	/* Protect stack dequeue */
	}
}

static inline
void __urcu_workqueue_wakeup_all(struct urcu_workqueue *queue)
{
	struct urcu_waiters waiters;

	rcu_read_lock();	/* Protect stack dequeue */
	urcu_move_waiters(&waiters, &queue->waitqueue);
	rcu_read_unlock();	/* Protect stack dequeue */

	(void) urcu_wake_all_waiters(&waiters);
}

static inline
void urcu_worker_init(struct urcu_worker *worker, int flags)
{
	cds_wfcq_init(&worker->head, &worker->tail);
	worker->flags = flags;
	urcu_wait_node_init(&worker->wait_node, URCU_WAIT_RUNNING);
	worker->own = NULL;
}

static inline
void urcu_worker_register(struct urcu_workqueue *queue,
		struct urcu_worker *worker)
{
	if (worker->flags & URCU_WORKER_STEAL) {
		pthread_mutex_lock(&queue->sibling_lock);
		cds_list_add_rcu(&worker->sibling_node, &queue->sibling_head);
		pthread_mutex_unlock(&queue->sibling_lock);
	}
}

static inline
void urcu_worker_unregister(struct urcu_workqueue *queue,
		struct urcu_worker *worker)
{
	enum cds_wfcq_ret wfcq_ret;

	if (worker->flags & URCU_WORKER_STEAL) {
		pthread_mutex_lock(&queue->sibling_lock);
		cds_list_del_rcu(&worker->sibling_node);
		pthread_mutex_unlock(&queue->sibling_lock);
	}

	/*
	 * Make sure we are removed from waitqueue.
	 */
	if (CMM_LOAD_SHARED(worker->wait_node.node.next))
		__urcu_workqueue_wakeup_all(queue);

	/*
	 * Put any local work we still have back into the workqueue.
	 */
	wfcq_ret = __cds_wfcq_splice_blocking(&queue->head,
			&queue->tail,
			&worker->head,
			&worker->tail);
	if (wfcq_ret != CDS_WFCQ_RET_SRC_EMPTY
			&& wfcq_ret == CDS_WFCQ_RET_DEST_EMPTY) {
		/*
		 * Wakeup worker thread if we have put work back into
		 * workqueue that was previously empty.
		 */
		rcu_read_lock();	/* Protect stack dequeue */
		(void) urcu_dequeue_wake_single(&queue->waitqueue);
		rcu_read_unlock();	/* Protect stack dequeue */
	}

	/*
	 * Wait for grace period before freeing or reusing
	 * "worker" because used by RCU linked list.
	 * Also prevents ABA for waitqueue stack dequeue: matches RCU
	 * read-side critical sections around dequeue and move all
	 * operations on waitqueue).
	 */
	synchronize_rcu();
}

static inline
bool ___urcu_grab_work(struct urcu_worker *worker,
		cds_wfcq_head_ptr_t src_head,
		struct cds_wfcq_tail *src_tail,
		bool steal)
{
	enum cds_wfcq_ret splice_ret;
	struct __cds_wfcq_head tmp_head;
	struct cds_wfcq_tail tmp_tail;
	struct cds_wfcq_node *node;

	/*
	 * Don't bother grabbing the src queue lock if it is empty.
	 */
	if (cds_wfcq_empty(src_head, src_tail))
		return false;
	__cds_wfcq_init(&tmp_head, &tmp_tail);

	/* Ensure that we preserve FIFO work order. */
	assert(!steal || worker->own == NULL);

	/* Splice to temporary queue. */
	if (steal)
		cds_wfcq_dequeue_lock(src_head.h, src_tail);
	splice_ret = __cds_wfcq_splice_blocking(&tmp_head,
			&tmp_tail,
			src_head,
			src_tail);
	if (steal)
		cds_wfcq_dequeue_unlock(src_head.h, src_tail);
	if (splice_ret == CDS_WFCQ_RET_SRC_EMPTY)
		return false;

	/*
	 * Keep one work entry for ourself. This ensures forward
	 * progress amongst stealing co-workers. This also ensures that
	 * when a worker grab some work from the global workqueue, it
	 * will have at least one work item to deal with.
	 */
	if (worker->own == NULL) {
		if (!steal) {
			/*
			 * Try to grab own work from worker workqueue to
			 * preserve FIFO order.
			 */
			node = cds_wfcq_dequeue_blocking(&worker->head,
				&worker->tail);
			if (node)
				goto got_node;
		}
		node = __cds_wfcq_dequeue_blocking(&tmp_head, &tmp_tail);
		assert(node != NULL);
got_node:
		worker->own = caa_container_of(node, struct urcu_work, node);
	}

	/* Splice into worker workqueue. */
	splice_ret = __cds_wfcq_splice_blocking(&worker->head,
			&worker->tail,
			&tmp_head,
			&tmp_tail);
	/* Ensure that we preserve FIFO work order. */
	assert(!steal || splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
	return true;
}

/*
 * Try stealing work from siblings when we have nothing to do.
 */
static inline
bool ___urcu_steal_work(struct urcu_worker *worker,
		struct urcu_worker *sibling)
{
	return ___urcu_grab_work(worker, &sibling->head, &sibling->tail, 1);
}

static inline
bool __urcu_steal_work(struct urcu_workqueue *queue,
		struct urcu_worker *worker)
{
	struct urcu_worker *sibling_prev, *sibling_next;
	struct cds_list_head *sibling_node;
	bool steal_performed = 0;

	if (!(worker->flags & URCU_WORKER_STEAL))
		return false;

	rcu_read_lock();

	sibling_node = rcu_dereference(worker->sibling_node.next);
	if (sibling_node == &queue->sibling_head)
		sibling_node = rcu_dereference(sibling_node->next);
	sibling_next = caa_container_of(sibling_node, struct urcu_worker,
			sibling_node);
	if (sibling_next != worker)
		steal_performed = ___urcu_steal_work(worker, sibling_next);
	if (steal_performed)
		goto end;

	sibling_node = rcu_dereference(worker->sibling_node.prev);
	if (sibling_node == &queue->sibling_head)
		sibling_node = rcu_dereference(sibling_node->prev);
	sibling_prev = caa_container_of(sibling_node, struct urcu_worker,
			sibling_node);
	if (sibling_prev != worker && sibling_prev != sibling_next)
		steal_performed =  ___urcu_steal_work(worker, sibling_prev);
end:
	rcu_read_unlock();

	return steal_performed;
}

static inline
bool ___urcu_wakeup_sibling(struct urcu_worker *sibling)
{
	return urcu_adaptative_wake_up(&sibling->wait_node);
}

static inline
bool __urcu_wakeup_siblings(struct urcu_workqueue *queue,
		struct urcu_worker *worker)
{
	struct urcu_worker *sibling_prev, *sibling_next;
	struct cds_list_head *sibling_node;
	bool wakeup_performed = 0;

	if (!(worker->flags & URCU_WORKER_STEAL))
		return;

	/* Only wakeup siblings if we have work in our own queue. */
	if (cds_wfcq_empty(&worker->head, &worker->tail))
		return;

	rcu_read_lock();

	sibling_node = rcu_dereference(worker->sibling_node.next);
	if (sibling_node == &queue->sibling_head)
		sibling_node = rcu_dereference(sibling_node->next);
	sibling_next = caa_container_of(sibling_node, struct urcu_worker,
			sibling_node);
	if (sibling_next != worker)
		wakeup_performed = ___urcu_wakeup_sibling(sibling_next);
	if (wakeup_performed)
		goto end;

	sibling_node = rcu_dereference(worker->sibling_node.prev);
	if (sibling_node == &queue->sibling_head)
		sibling_node = rcu_dereference(sibling_node->prev);
	sibling_prev = caa_container_of(sibling_node, struct urcu_worker,
			sibling_node);
	if (sibling_prev != worker && sibling_prev != sibling_next)
		wakeup_performed = ___urcu_wakeup_sibling(sibling_prev);
end:
	rcu_read_unlock();

	return wakeup_performed;
}

static inline
enum urcu_accept_ret urcu_accept_work(struct urcu_workqueue *queue,
		struct urcu_worker *worker)
{
	enum cds_wfcq_ret wfcq_ret;
	bool has_work;

	has_work = ___urcu_grab_work(worker, &queue->head, &queue->tail, 0);
	/* Don't wait if we have work to do. */
	if (has_work || !cds_wfcq_empty(&worker->head, &worker->tail))
		goto do_work;
	/* Try to steal work from sibling instead of blocking */
	if (__urcu_steal_work(queue, worker))
		goto do_work;
	/* No more work to do, check shutdown state */
	if (CMM_LOAD_SHARED(queue->shutdown))
		return URCU_ACCEPT_SHUTDOWN;
	urcu_wait_set_state(&worker->wait_node,
			URCU_WAIT_WAITING);
	if (!CMM_LOAD_SHARED(worker->wait_node.node.next)) {
		int was_empty;

		/*
		 * NULL next pointer. We are therefore not in
		 * the queue.
		 */
		cds_lfs_node_init(&worker->wait_node.node);
		/* Protect stack dequeue against ABA */
		synchronize_rcu();
		was_empty = !urcu_wait_add(&queue->waitqueue,
				&worker->wait_node);
		/*
		 * If the wait queue was empty, it means we are the
		 * first thread to be put back into an otherwise empty
		 * wait queue. Re-check if work queue is empty after
		 * adding ourself to wait queue, so we can wakeup the
		 * top of wait queue since new work have appeared, and
		 * work enqueuer may not have seen that it needed to do
		 * a wake up.
		 */
		if (was_empty && !cds_wfcq_empty(&queue->head,
						&queue->tail)) {
			rcu_read_lock();	/* Protect stack dequeue */
			(void) urcu_dequeue_wake_single(&queue->waitqueue);
			rcu_read_unlock();	/* Protect stack dequeue */
		}
	} else {
		/*
		 * Non-NULL next pointer. We are therefore in
		 * the queue, or the dispatcher just removed us
		 * from it (after we read the next pointer), and
		 * is therefore awakening us. The state will
		 * therefore have been changed from WAITING to
		 * some other state, which will let the busy
		 * wait pass through.
		 */
	}
	urcu_adaptative_busy_wait(&worker->wait_node);
	return;

do_work:
	/*
	 * We will be busy handling the work batch, awaken siblings so
	 * they can steal from us.
	 */
	(void) __urcu_wakeup_siblings(queue, worker);
	return URCU_ACCEPT_WORK;
}

static inline
struct urcu_work *urcu_dequeue_work(struct urcu_worker *worker)
{
	struct cds_wfcq_node *node;

	if (worker->own) {
		struct urcu_work *work;

		/* Process our own work entry. */
		work = worker->own;
		worker->own = NULL;
		return work;
	}
	/*
	 * If we are registered for work stealing, we need to dequeue
	 * safely against siblings.
	 */
	if (worker->flags & URCU_WORKER_STEAL) {
		/*
		 * Don't bother grabbing the worker queue lock if it is
		 * empty.
		 */
		if (cds_wfcq_empty(&worker->head, &worker->tail))
			return NULL;
		node = cds_wfcq_dequeue_blocking(&worker->head,
				&worker->tail);
	} else {
		node = ___cds_wfcq_dequeue_with_state(&worker->head,
				&worker->tail, NULL, 1, 0);
	}
	if (!node)
		return NULL;
	return caa_container_of(node, struct urcu_work, node);
}

static inline
void urcu_workqueue_shutdown(struct urcu_workqueue *queue)
{
	/* Set shutdown */
	CMM_STORE_SHARED(queue->shutdown, true);
	/* Wakeup all workers */
	__urcu_workqueue_wakeup_all(queue);
}

#endif /* _URCU_WORKQUEUE_FIFO_H */
Commit	Line	Data
13652c4b MD	1	#ifndef _URCU_WORKQUEUE_FIFO_H
	2	#define _URCU_WORKQUEUE_FIFO_H
	3
	4	/*
	5	* urcu/workqueue-fifo.h
	6	*
	7	* Userspace RCU library - work queue scheme with FIFO semantic
	8	*
	9	* Copyright (c) 2014 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
	10	*
	11	* This library is free software; you can redistribute it and/or
	12	* modify it under the terms of the GNU Lesser General Public
	13	* License as published by the Free Software Foundation; either
	14	* version 2.1 of the License, or (at your option) any later version.
	15	*
	16	* This library is distributed in the hope that it will be useful,
	17	* but WITHOUT ANY WARRANTY; without even the implied warranty of
	18	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
	19	* Lesser General Public License for more details.
	20	*
	21	* You should have received a copy of the GNU Lesser General Public
	22	* License along with this library; if not, write to the Free Software
	23	* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
	24	*/
	25
	26	#include <urcu/uatomic.h>
7a618cf7	27	#include <urcu/lfstack.h>
13652c4b MD	28	#include <urcu/waitqueue-lifo.h>
	29	#include <urcu/wfcqueue.h>
	30	#include <urcu/rculist.h>
	31	#include <pthread.h>
e10c65b3	32	#include <assert.h>
13652c4b	33
1b0a9891 MD	34	enum urcu_accept_ret {
	35	URCU_ACCEPT_WORK = 0,
	36	URCU_ACCEPT_SHUTDOWN = 1,
	37	};
	38
13652c4b MD	39	/*
	40	* We use RCU to steal work from siblings. Therefore, one of RCU flavors
	41	* need to be included before this header. All worker that participate
	42	* in stealing (initialized with the URCU_WORKER_STEAL flag) need to be
	43	* registered RCU readers threads.
	44	*/
	45
	46	struct urcu_work {
	47	struct cds_wfcq_node node;
	48	};
	49
	50	struct urcu_workqueue {
	51	/* FIFO work queue */
	52	struct __cds_wfcq_head head;
	53	struct cds_wfcq_tail tail;
	54
	55	/* Associated wait queue for LIFO wait/wakeup */
	56	struct urcu_wait_queue waitqueue;
	57
	58	/* RCU linked list head of siblings for work stealing. */
	59	struct cds_list_head sibling_head;
	60	pthread_mutex_t sibling_lock; /* Protect sibling list updates */
1b0a9891 MD	61
1b0a9891 MD	62	bool shutdown; /* Shutdown performed */
13652c4b MD	63	};
	64
	65	struct urcu_worker {
a6492159	66	/* Workqueue which can be either used by worker, or stolen. */
13652c4b MD	67	struct cds_wfcq_head head;
	68	struct cds_wfcq_tail tail;
	69
a6492159 MD	70	/* Work belonging to worker. Cannot be stolen. */
	71	struct urcu_work *own;
	72
13652c4b MD	73	struct urcu_wait_node wait_node;
	74	/* RCU linked list node of siblings for work stealing. */
	75	struct cds_list_head sibling_node;
	76	int flags; /* enum urcu_worker_flags */
	77	};
	78
	79	enum urcu_worker_flags {
	80	URCU_WORKER_STEAL = (1 << 0),
	81	};
	82
	83	static inline
	84	void urcu_workqueue_init(struct urcu_workqueue *queue)
	85	{
	86	__cds_wfcq_init(&queue->head, &queue->tail);
	87	urcu_wait_queue_init(&queue->waitqueue);
	88	CDS_INIT_LIST_HEAD(&queue->sibling_head);
8313fa62	89	pthread_mutex_init(&queue->sibling_lock, NULL);
1b0a9891	90	queue->shutdown = false;
13652c4b MD	91	}
	92
	93	static inline
	94	void urcu_queue_work(struct urcu_workqueue queue, struct urcu_work work)
	95	{
	96	bool was_empty;
	97
	98	cds_wfcq_node_init(&work->node);
	99
	100	/* Enqueue work. */
	101	was_empty = !cds_wfcq_enqueue(&queue->head, &queue->tail,
	102	&work->node);
	103	/*
	104	* If workqueue was previously empty, wakeup one worker thread.
	105	* It will eventually grab the entire content of the work-queue
	106	* (therefore grabbing a "work batch"). After having grabbed the
	107	* work batch, while that thread is running and taking care of
	108	* that work batch, when we enqueue more work, we will wake
	109	* another thread (if there is one waiting), which will
	110	* eventually grab the new batch, and so on. This scheme ensures
	111	* that contiguous batch of work are handled by the same thread
	112	* (for locality), and also ensures that we scale work to many
	113	* worker threads when threads are busy enough to still be
	114	* running when work is enqueued.
	115	*/
d3afe039 MD	116	if (was_empty) {
d3afe039 MD	117	rcu_read_lock(); /* Protect stack dequeue */
13652c4b	118	(void) urcu_dequeue_wake_single(&queue->waitqueue);
d3afe039 MD	119	rcu_read_unlock(); /* Protect stack dequeue */
d3afe039 MD	120	}
13652c4b MD	121	}
	122
	123	static inline
1b0a9891	124	void __urcu_workqueue_wakeup_all(struct urcu_workqueue *queue)
13652c4b MD	125	{
	126	struct urcu_waiters waiters;
	127
d3afe039	128	rcu_read_lock(); /* Protect stack dequeue */
13652c4b	129	urcu_move_waiters(&waiters, &queue->waitqueue);
d3afe039 MD	130	rcu_read_unlock(); /* Protect stack dequeue */
d3afe039 MD	131
13652c4b MD	132	(void) urcu_wake_all_waiters(&waiters);
	133	}
	134
	135	static inline
	136	void urcu_worker_init(struct urcu_worker *worker, int flags)
	137	{
	138	cds_wfcq_init(&worker->head, &worker->tail);
	139	worker->flags = flags;
	140	urcu_wait_node_init(&worker->wait_node, URCU_WAIT_RUNNING);
a6492159	141	worker->own = NULL;
13652c4b MD	142	}
	143
	144	static inline
	145	void urcu_worker_register(struct urcu_workqueue *queue,
	146	struct urcu_worker *worker)
	147	{
	148	if (worker->flags & URCU_WORKER_STEAL) {
	149	pthread_mutex_lock(&queue->sibling_lock);
	150	cds_list_add_rcu(&worker->sibling_node, &queue->sibling_head);
	151	pthread_mutex_unlock(&queue->sibling_lock);
	152	}
	153	}
	154
	155	static inline
	156	void urcu_worker_unregister(struct urcu_workqueue *queue,
	157	struct urcu_worker *worker)
	158	{
	159	enum cds_wfcq_ret wfcq_ret;
	160
	161	if (worker->flags & URCU_WORKER_STEAL) {
	162	pthread_mutex_lock(&queue->sibling_lock);
	163	cds_list_del_rcu(&worker->sibling_node);
	164	pthread_mutex_unlock(&queue->sibling_lock);
13652c4b MD	165	}
13652c4b MD	166
d3afe039	167	/*
6e17009c	168	* Make sure we are removed from waitqueue.
d3afe039	169	*/
6e17009c MD	170	if (CMM_LOAD_SHARED(worker->wait_node.node.next))
6e17009c MD	171	__urcu_workqueue_wakeup_all(queue);
d3afe039	172
13652c4b MD	173	/*
	174	* Put any local work we still have back into the workqueue.
	175	*/
	176	wfcq_ret = __cds_wfcq_splice_blocking(&queue->head,
	177	&queue->tail,
	178	&worker->head,
	179	&worker->tail);
	180	if (wfcq_ret != CDS_WFCQ_RET_SRC_EMPTY
	181	&& wfcq_ret == CDS_WFCQ_RET_DEST_EMPTY) {
	182	/*
	183	* Wakeup worker thread if we have put work back into
	184	* workqueue that was previously empty.
	185	*/
d3afe039	186	rcu_read_lock(); /* Protect stack dequeue */
13652c4b	187	(void) urcu_dequeue_wake_single(&queue->waitqueue);
d3afe039	188	rcu_read_unlock(); /* Protect stack dequeue */
13652c4b	189	}
6e17009c MD	190
	191	/*
	192	* Wait for grace period before freeing or reusing
	193	* "worker" because used by RCU linked list.
	194	* Also prevents ABA for waitqueue stack dequeue: matches RCU
	195	* read-side critical sections around dequeue and move all
	196	* operations on waitqueue).
	197	*/
	198	synchronize_rcu();
13652c4b MD	199	}
13652c4b MD	200
13652c4b	201	static inline
a6492159 MD	202	bool ___urcu_grab_work(struct urcu_worker *worker,
	203	cds_wfcq_head_ptr_t src_head,
	204	struct cds_wfcq_tail *src_tail,
	205	bool steal)
13652c4b	206	{
e10c65b3	207	enum cds_wfcq_ret splice_ret;
8313fa62	208	struct __cds_wfcq_head tmp_head;
a6492159 MD	209	struct cds_wfcq_tail tmp_tail;
a6492159 MD	210	struct cds_wfcq_node *node;
e10c65b3	211
30926570	212	/*
a6492159	213	* Don't bother grabbing the src queue lock if it is empty.
30926570	214	*/
a6492159	215	if (cds_wfcq_empty(src_head, src_tail))
e10c65b3	216	return false;
8313fa62	217	__cds_wfcq_init(&tmp_head, &tmp_tail);
a6492159 MD	218
	219	/* Ensure that we preserve FIFO work order. */
	220	assert(!steal \|\| worker->own == NULL);
	221
	222	/* Splice to temporary queue. */
	223	if (steal)
	224	cds_wfcq_dequeue_lock(src_head.h, src_tail);
	225	splice_ret = __cds_wfcq_splice_blocking(&tmp_head,
	226	&tmp_tail,
	227	src_head,
	228	src_tail);
	229	if (steal)
	230	cds_wfcq_dequeue_unlock(src_head.h, src_tail);
	231	if (splice_ret == CDS_WFCQ_RET_SRC_EMPTY)
	232	return false;
	233
	234	/*
	235	* Keep one work entry for ourself. This ensures forward
	236	* progress amongst stealing co-workers. This also ensures that
	237	* when a worker grab some work from the global workqueue, it
	238	* will have at least one work item to deal with.
	239	*/
	240	if (worker->own == NULL) {
	241	if (!steal) {
	242	/*
	243	* Try to grab own work from worker workqueue to
	244	* preserve FIFO order.
	245	*/
	246	node = cds_wfcq_dequeue_blocking(&worker->head,
	247	&worker->tail);
	248	if (node)
	249	goto got_node;
	250	}
	251	node = __cds_wfcq_dequeue_blocking(&tmp_head, &tmp_tail);
	252	assert(node != NULL);
	253	got_node:
	254	worker->own = caa_container_of(node, struct urcu_work, node);
	255	}
	256
	257	/* Splice into worker workqueue. */
8313fa62	258	splice_ret = __cds_wfcq_splice_blocking(&worker->head,
13652c4b	259	&worker->tail,
a6492159 MD	260	&tmp_head,
a6492159 MD	261	&tmp_tail);
e10c65b3	262	/* Ensure that we preserve FIFO work order. */
a6492159 MD	263	assert(!steal \|\| splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
	264	return true;
	265	}
	266
	267	/*
	268	* Try stealing work from siblings when we have nothing to do.
	269	*/
	270	static inline
	271	bool ___urcu_steal_work(struct urcu_worker *worker,
	272	struct urcu_worker *sibling)
	273	{
	274	return ___urcu_grab_work(worker, &sibling->head, &sibling->tail, 1);
13652c4b MD	275	}
	276
	277	static inline
e10c65b3	278	bool __urcu_steal_work(struct urcu_workqueue *queue,
13652c4b MD	279	struct urcu_worker *worker)
	280	{
	281	struct urcu_worker sibling_prev, sibling_next;
	282	struct cds_list_head *sibling_node;
e10c65b3	283	bool steal_performed = 0;
13652c4b MD	284
13652c4b MD	285	if (!(worker->flags & URCU_WORKER_STEAL))
e10c65b3	286	return false;
13652c4b MD	287
	288	rcu_read_lock();
	289
	290	sibling_node = rcu_dereference(worker->sibling_node.next);
	291	if (sibling_node == &queue->sibling_head)
	292	sibling_node = rcu_dereference(sibling_node->next);
	293	sibling_next = caa_container_of(sibling_node, struct urcu_worker,
	294	sibling_node);
	295	if (sibling_next != worker)
e10c65b3 MD	296	steal_performed = ___urcu_steal_work(worker, sibling_next);
	297	if (steal_performed)
	298	goto end;
13652c4b MD	299
	300	sibling_node = rcu_dereference(worker->sibling_node.prev);
	301	if (sibling_node == &queue->sibling_head)
	302	sibling_node = rcu_dereference(sibling_node->prev);
	303	sibling_prev = caa_container_of(sibling_node, struct urcu_worker,
	304	sibling_node);
	305	if (sibling_prev != worker && sibling_prev != sibling_next)
e10c65b3 MD	306	steal_performed = ___urcu_steal_work(worker, sibling_prev);
e10c65b3 MD	307	end:
13652c4b MD	308	rcu_read_unlock();
13652c4b MD	309
e10c65b3	310	return steal_performed;
13652c4b MD	311	}
	312
	313	static inline
5d30bf32	314	bool ___urcu_wakeup_sibling(struct urcu_worker *sibling)
13652c4b	315	{
5d30bf32	316	return urcu_adaptative_wake_up(&sibling->wait_node);
13652c4b MD	317	}
	318
	319	static inline
5d30bf32	320	bool __urcu_wakeup_siblings(struct urcu_workqueue *queue,
13652c4b MD	321	struct urcu_worker *worker)
	322	{
	323	struct urcu_worker sibling_prev, sibling_next;
	324	struct cds_list_head *sibling_node;
5d30bf32	325	bool wakeup_performed = 0;
13652c4b MD	326
	327	if (!(worker->flags & URCU_WORKER_STEAL))
	328	return;
	329
	330	/* Only wakeup siblings if we have work in our own queue. */
	331	if (cds_wfcq_empty(&worker->head, &worker->tail))
	332	return;
	333
	334	rcu_read_lock();
	335
	336	sibling_node = rcu_dereference(worker->sibling_node.next);
	337	if (sibling_node == &queue->sibling_head)
	338	sibling_node = rcu_dereference(sibling_node->next);
	339	sibling_next = caa_container_of(sibling_node, struct urcu_worker,
	340	sibling_node);
	341	if (sibling_next != worker)
5d30bf32 MD	342	wakeup_performed = ___urcu_wakeup_sibling(sibling_next);
	343	if (wakeup_performed)
	344	goto end;
13652c4b MD	345
	346	sibling_node = rcu_dereference(worker->sibling_node.prev);
	347	if (sibling_node == &queue->sibling_head)
	348	sibling_node = rcu_dereference(sibling_node->prev);
	349	sibling_prev = caa_container_of(sibling_node, struct urcu_worker,
	350	sibling_node);
	351	if (sibling_prev != worker && sibling_prev != sibling_next)
5d30bf32 MD	352	wakeup_performed = ___urcu_wakeup_sibling(sibling_prev);
5d30bf32 MD	353	end:
13652c4b	354	rcu_read_unlock();
5d30bf32 MD	355
5d30bf32 MD	356	return wakeup_performed;
13652c4b MD	357	}
	358
	359	static inline
1b0a9891 MD	360	enum urcu_accept_ret urcu_accept_work(struct urcu_workqueue *queue,
1b0a9891 MD	361	struct urcu_worker *worker)
13652c4b MD	362	{
13652c4b MD	363	enum cds_wfcq_ret wfcq_ret;
a6492159	364	bool has_work;
13652c4b	365
a6492159	366	has_work = ___urcu_grab_work(worker, &queue->head, &queue->tail, 0);
13652c4b	367	/* Don't wait if we have work to do. */
a6492159	368	if (has_work \|\| !cds_wfcq_empty(&worker->head, &worker->tail))
13652c4b MD	369	goto do_work;
	370	/* Try to steal work from sibling instead of blocking */
	371	if (__urcu_steal_work(queue, worker))
	372	goto do_work;
1b0a9891 MD	373	/* No more work to do, check shutdown state */
	374	if (CMM_LOAD_SHARED(queue->shutdown))
	375	return URCU_ACCEPT_SHUTDOWN;
13652c4b MD	376	urcu_wait_set_state(&worker->wait_node,
	377	URCU_WAIT_WAITING);
	378	if (!CMM_LOAD_SHARED(worker->wait_node.node.next)) {
	379	int was_empty;
	380
	381	/*
	382	* NULL next pointer. We are therefore not in
	383	* the queue.
	384	*/
7a618cf7	385	cds_lfs_node_init(&worker->wait_node.node);
d3afe039 MD	386	/* Protect stack dequeue against ABA */
d3afe039 MD	387	synchronize_rcu();
13652c4b MD	388	was_empty = !urcu_wait_add(&queue->waitqueue,
	389	&worker->wait_node);
	390	/*
	391	* If the wait queue was empty, it means we are the
	392	* first thread to be put back into an otherwise empty
	393	* wait queue. Re-check if work queue is empty after
	394	* adding ourself to wait queue, so we can wakeup the
	395	* top of wait queue since new work have appeared, and
	396	* work enqueuer may not have seen that it needed to do
	397	* a wake up.
	398	*/
	399	if (was_empty && !cds_wfcq_empty(&queue->head,
d3afe039 MD	400	&queue->tail)) {
d3afe039 MD	401	rcu_read_lock(); /* Protect stack dequeue */
13652c4b	402	(void) urcu_dequeue_wake_single(&queue->waitqueue);
d3afe039 MD	403	rcu_read_unlock(); /* Protect stack dequeue */
d3afe039 MD	404	}
13652c4b MD	405	} else {
	406	/*
	407	* Non-NULL next pointer. We are therefore in
	408	* the queue, or the dispatcher just removed us
	409	* from it (after we read the next pointer), and
	410	* is therefore awakening us. The state will
	411	* therefore have been changed from WAITING to
	412	* some other state, which will let the busy
	413	* wait pass through.
	414	*/
	415	}
	416	urcu_adaptative_busy_wait(&worker->wait_node);
	417	return;
	418
	419	do_work:
	420	/*
	421	* We will be busy handling the work batch, awaken siblings so
	422	* they can steal from us.
	423	*/
5d30bf32	424	(void) __urcu_wakeup_siblings(queue, worker);
1b0a9891	425	return URCU_ACCEPT_WORK;
13652c4b MD	426	}
	427
	428	static inline
	429	struct urcu_work urcu_dequeue_work(struct urcu_worker worker)
	430	{
	431	struct cds_wfcq_node *node;
	432
a6492159 MD	433	if (worker->own) {
	434	struct urcu_work *work;
	435
	436	/* Process our own work entry. */
	437	work = worker->own;
	438	worker->own = NULL;
	439	return work;
	440	}
13652c4b MD	441	/*
	442	* If we are registered for work stealing, we need to dequeue
	443	* safely against siblings.
	444	*/
30926570 MD	445	if (worker->flags & URCU_WORKER_STEAL) {
	446	/*
	447	* Don't bother grabbing the worker queue lock if it is
	448	* empty.
	449	*/
	450	if (cds_wfcq_empty(&worker->head, &worker->tail))
	451	return NULL;
13652c4b MD	452	node = cds_wfcq_dequeue_blocking(&worker->head,
13652c4b MD	453	&worker->tail);
30926570	454	} else {
13652c4b MD	455	node = ___cds_wfcq_dequeue_with_state(&worker->head,
13652c4b MD	456	&worker->tail, NULL, 1, 0);
30926570	457	}
13652c4b MD	458	if (!node)
	459	return NULL;
	460	return caa_container_of(node, struct urcu_work, node);
	461	}
	462
1b0a9891 MD	463	static inline
	464	void urcu_workqueue_shutdown(struct urcu_workqueue *queue)
	465	{
	466	/* Set shutdown */
	467	CMM_STORE_SHARED(queue->shutdown, true);
	468	/* Wakeup all workers */
	469	__urcu_workqueue_wakeup_all(queue);
	470	}
	471
13652c4b	472	#endif /* _URCU_WORKQUEUE_FIFO_H */