[userspace-rcu.git] / src / workqueue.c

/*
 * workqueue.c
 *
 * Userspace RCU library - Userspace workqeues
 *
 * Copyright (c) 2010 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
 * Copyright (c) 2017 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
 */

#define _LGPL_SOURCE
#include <stdio.h>
#include <pthread.h>
#include <signal.h>
#include <assert.h>
#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include <errno.h>
#include <poll.h>
#include <sys/time.h>
#include <unistd.h>
#include <sched.h>

#include "compat-getcpu.h"
#include <urcu/wfcqueue.h>
#include <urcu/pointer.h>
#include <urcu/list.h>
#include <urcu/futex.h>
#include <urcu/tls-compat.h>
#include <urcu/ref.h>
#include "urcu-die.h"

#include "workqueue.h"

#define SET_AFFINITY_CHECK_PERIOD		(1U << 8)	/* 256 */
#define SET_AFFINITY_CHECK_PERIOD_MASK		(SET_AFFINITY_CHECK_PERIOD - 1)

/* Data structure that identifies a workqueue. */

struct urcu_workqueue {
	/*
	 * We do not align head on a different cache-line than tail
	 * mainly because workqueue threads use batching ("splice") to
	 * get an entire list of callbacks, which effectively empties
	 * the queue, and requires to touch the tail anyway.
	 */
	struct cds_wfcq_tail cbs_tail;
	struct cds_wfcq_head cbs_head;
	unsigned long flags;
	int32_t futex;
	unsigned long qlen; /* maintained for debugging. */
	pthread_t tid;
	int cpu_affinity;
	unsigned long loop_count;
	void *priv;
	void (*grace_period_fct)(struct urcu_workqueue *workqueue, void *priv);
	void (*initialize_worker_fct)(struct urcu_workqueue *workqueue, void *priv);
	void (*finalize_worker_fct)(struct urcu_workqueue *workqueue, void *priv);
	void (*worker_before_pause_fct)(struct urcu_workqueue *workqueue, void *priv);
	void (*worker_after_resume_fct)(struct urcu_workqueue *workqueue, void *priv);
	void (*worker_before_wait_fct)(struct urcu_workqueue *workqueue, void *priv);
	void (*worker_after_wake_up_fct)(struct urcu_workqueue *workqueue, void *priv);
} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));

struct urcu_workqueue_completion {
	int barrier_count;
	int32_t futex;
	struct urcu_ref ref;
};

struct urcu_workqueue_completion_work {
	struct urcu_work work;
	struct urcu_workqueue_completion *completion;
};

/*
 * Periodically retry setting CPU affinity if we migrate.
 * Losing affinity can be caused by CPU hotunplug/hotplug, or by
 * cpuset(7).
 */
#ifdef HAVE_SCHED_SETAFFINITY
static int set_thread_cpu_affinity(struct urcu_workqueue *workqueue)
{
	cpu_set_t mask;
	int ret;

	if (workqueue->cpu_affinity < 0)
		return 0;
	if (++workqueue->loop_count & SET_AFFINITY_CHECK_PERIOD_MASK)
		return 0;
	if (urcu_sched_getcpu() == workqueue->cpu_affinity)
		return 0;

	CPU_ZERO(&mask);
	CPU_SET(workqueue->cpu_affinity, &mask);
#if SCHED_SETAFFINITY_ARGS == 2
	ret = sched_setaffinity(0, &mask);
#else
	ret = sched_setaffinity(0, sizeof(mask), &mask);
#endif
	/*
	 * EINVAL is fine: can be caused by hotunplugged CPUs, or by
	 * cpuset(7). This is why we should always retry if we detect
	 * migration.
	 */
	if (ret && errno == EINVAL) {
		ret = 0;
		errno = 0;
	}
	return ret;
}
#else
static int set_thread_cpu_affinity(struct urcu_workqueue *workqueue __attribute__((unused)))
{
	return 0;
}
#endif

static void futex_wait(int32_t *futex)
{
	/* Read condition before read futex */
	cmm_smp_mb();
	if (uatomic_read(futex) != -1)
		return;
	while (futex_async(futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
		switch (errno) {
		case EWOULDBLOCK:
			/* Value already changed. */
			return;
		case EINTR:
			/* Retry if interrupted by signal. */
			break;	/* Get out of switch. */
		default:
			/* Unexpected error. */
			urcu_die(errno);
		}
	}
}

static void futex_wake_up(int32_t *futex)
{
	/* Write to condition before reading/writing futex */
	cmm_smp_mb();
	if (caa_unlikely(uatomic_read(futex) == -1)) {
		uatomic_set(futex, 0);
		if (futex_async(futex, FUTEX_WAKE, 1,
				NULL, NULL, 0) < 0)
			urcu_die(errno);
	}
}

/* This is the code run by each worker thread. */

static void *workqueue_thread(void *arg)
{
	unsigned long cbcount;
	struct urcu_workqueue *workqueue = (struct urcu_workqueue *) arg;
	int rt = !!(uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_RT);

	if (set_thread_cpu_affinity(workqueue))
		urcu_die(errno);

	if (workqueue->initialize_worker_fct)
		workqueue->initialize_worker_fct(workqueue, workqueue->priv);

	if (!rt) {
		uatomic_dec(&workqueue->futex);
		/* Decrement futex before reading workqueue */
		cmm_smp_mb();
	}
	for (;;) {
		struct cds_wfcq_head cbs_tmp_head;
		struct cds_wfcq_tail cbs_tmp_tail;
		struct cds_wfcq_node *cbs, *cbs_tmp_n;
		enum cds_wfcq_ret splice_ret;

		if (set_thread_cpu_affinity(workqueue))
			urcu_die(errno);

		if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) {
			/*
			 * Pause requested. Become quiescent: remove
			 * ourself from all global lists, and don't
			 * process any callback. The callback lists may
			 * still be non-empty though.
			 */
			if (workqueue->worker_before_pause_fct)
				workqueue->worker_before_pause_fct(workqueue, workqueue->priv);
			cmm_smp_mb__before_uatomic_or();
			uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSED);
			while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) != 0)
				(void) poll(NULL, 0, 1);
			uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSED);
			cmm_smp_mb__after_uatomic_and();
			if (workqueue->worker_after_resume_fct)
				workqueue->worker_after_resume_fct(workqueue, workqueue->priv);
		}

		cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
		splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
			&cbs_tmp_tail, &workqueue->cbs_head, &workqueue->cbs_tail);
		assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
		assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
		if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
			if (workqueue->grace_period_fct)
				workqueue->grace_period_fct(workqueue, workqueue->priv);
			cbcount = 0;
			__cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
					&cbs_tmp_tail, cbs, cbs_tmp_n) {
				struct urcu_work *uwp;

				uwp = caa_container_of(cbs,
					struct urcu_work, next);
				uwp->func(uwp);
				cbcount++;
			}
			uatomic_sub(&workqueue->qlen, cbcount);
		}
		if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_STOP)
			break;
		if (workqueue->worker_before_wait_fct)
			workqueue->worker_before_wait_fct(workqueue, workqueue->priv);
		if (!rt) {
			if (cds_wfcq_empty(&workqueue->cbs_head,
					&workqueue->cbs_tail)) {
				futex_wait(&workqueue->futex);
				uatomic_dec(&workqueue->futex);
				/*
				 * Decrement futex before reading
				 * urcu_work list.
				 */
				cmm_smp_mb();
			}
		} else {
			if (cds_wfcq_empty(&workqueue->cbs_head,
					&workqueue->cbs_tail)) {
				(void) poll(NULL, 0, 10);
			}
		}
		if (workqueue->worker_after_wake_up_fct)
			workqueue->worker_after_wake_up_fct(workqueue, workqueue->priv);
	}
	if (!rt) {
		/*
		 * Read urcu_work list before write futex.
		 */
		cmm_smp_mb();
		uatomic_set(&workqueue->futex, 0);
	}
	if (workqueue->finalize_worker_fct)
		workqueue->finalize_worker_fct(workqueue, workqueue->priv);
	return NULL;
}

struct urcu_workqueue *urcu_workqueue_create(unsigned long flags,
		int cpu_affinity, void *priv,
		void (*grace_period_fct)(struct urcu_workqueue *workqueue, void *priv),
		void (*initialize_worker_fct)(struct urcu_workqueue *workqueue, void *priv),
		void (*finalize_worker_fct)(struct urcu_workqueue *workqueue, void *priv),
		void (*worker_before_wait_fct)(struct urcu_workqueue *workqueue, void *priv),
		void (*worker_after_wake_up_fct)(struct urcu_workqueue *workqueue, void *priv),
		void (*worker_before_pause_fct)(struct urcu_workqueue *workqueue, void *priv),
		void (*worker_after_resume_fct)(struct urcu_workqueue *workqueue, void *priv))
{
	struct urcu_workqueue *workqueue;
	int ret;

	workqueue = malloc(sizeof(*workqueue));
	if (workqueue == NULL)
		urcu_die(errno);
	memset(workqueue, '\0', sizeof(*workqueue));
	cds_wfcq_init(&workqueue->cbs_head, &workqueue->cbs_tail);
	workqueue->qlen = 0;
	workqueue->futex = 0;
	workqueue->flags = flags;
	workqueue->priv = priv;
	workqueue->grace_period_fct = grace_period_fct;
	workqueue->initialize_worker_fct = initialize_worker_fct;
	workqueue->finalize_worker_fct = finalize_worker_fct;
	workqueue->worker_before_wait_fct = worker_before_wait_fct;
	workqueue->worker_after_wake_up_fct = worker_after_wake_up_fct;
	workqueue->worker_before_pause_fct = worker_before_pause_fct;
	workqueue->worker_after_resume_fct = worker_after_resume_fct;
	workqueue->cpu_affinity = cpu_affinity;
	workqueue->loop_count = 0;
	cmm_smp_mb();  /* Structure initialized before pointer is planted. */
	ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
	if (ret) {
		urcu_die(ret);
	}
	return workqueue;
}

static void wake_worker_thread(struct urcu_workqueue *workqueue)
{
	if (!(_CMM_LOAD_SHARED(workqueue->flags) & URCU_WORKQUEUE_RT))
		futex_wake_up(&workqueue->futex);
}

static int urcu_workqueue_destroy_worker(struct urcu_workqueue *workqueue)
{
	int ret;
	void *retval;

	uatomic_or(&workqueue->flags, URCU_WORKQUEUE_STOP);
	wake_worker_thread(workqueue);

	ret = pthread_join(workqueue->tid, &retval);
	if (ret) {
		urcu_die(ret);
	}
	if (retval != NULL) {
		urcu_die(EINVAL);
	}
	workqueue->flags &= ~URCU_WORKQUEUE_STOP;
	workqueue->tid = 0;
	return 0;
}

void urcu_workqueue_destroy(struct urcu_workqueue *workqueue)
{
	if (workqueue == NULL) {
		return;
	}
	if (urcu_workqueue_destroy_worker(workqueue)) {
		urcu_die(errno);
	}
	assert(cds_wfcq_empty(&workqueue->cbs_head, &workqueue->cbs_tail));
	free(workqueue);
}

void urcu_workqueue_queue_work(struct urcu_workqueue *workqueue,
		      struct urcu_work *work,
		      void (*func)(struct urcu_work *work))
{
	cds_wfcq_node_init(&work->next);
	work->func = func;
	cds_wfcq_enqueue(&workqueue->cbs_head, &workqueue->cbs_tail, &work->next);
	uatomic_inc(&workqueue->qlen);
	wake_worker_thread(workqueue);
}

static
void free_completion(struct urcu_ref *ref)
{
	struct urcu_workqueue_completion *completion;

	completion = caa_container_of(ref, struct urcu_workqueue_completion, ref);
	free(completion);
}

static
void _urcu_workqueue_wait_complete(struct urcu_work *work)
{
	struct urcu_workqueue_completion_work *completion_work;
	struct urcu_workqueue_completion *completion;

	completion_work = caa_container_of(work, struct urcu_workqueue_completion_work, work);
	completion = completion_work->completion;
	if (!uatomic_sub_return(&completion->barrier_count, 1))
		futex_wake_up(&completion->futex);
	urcu_ref_put(&completion->ref, free_completion);
	free(completion_work);
}

struct urcu_workqueue_completion *urcu_workqueue_create_completion(void)
{
	struct urcu_workqueue_completion *completion;

	completion = calloc(sizeof(*completion), 1);
	if (!completion)
		urcu_die(errno);
	urcu_ref_set(&completion->ref, 1);
	completion->barrier_count = 0;
	return completion;
}

void urcu_workqueue_destroy_completion(struct urcu_workqueue_completion *completion)
{
	urcu_ref_put(&completion->ref, free_completion);
}

void urcu_workqueue_wait_completion(struct urcu_workqueue_completion *completion)
{
	/* Wait for them */
	for (;;) {
		uatomic_dec(&completion->futex);
		/* Decrement futex before reading barrier_count */
		cmm_smp_mb();
		if (!uatomic_read(&completion->barrier_count))
			break;
		futex_wait(&completion->futex);
	}
}

void urcu_workqueue_queue_completion(struct urcu_workqueue *workqueue,
		struct urcu_workqueue_completion *completion)
{
	struct urcu_workqueue_completion_work *work;

	work = calloc(sizeof(*work), 1);
	if (!work)
		urcu_die(errno);
	work->completion = completion;
	urcu_ref_get(&completion->ref);
	uatomic_inc(&completion->barrier_count);
	urcu_workqueue_queue_work(workqueue, &work->work, _urcu_workqueue_wait_complete);
}

/*
 * Wait for all in-flight work to complete execution.
 */
void urcu_workqueue_flush_queued_work(struct urcu_workqueue *workqueue)
{
	struct urcu_workqueue_completion *completion;

	completion = urcu_workqueue_create_completion();
	if (!completion)
		urcu_die(ENOMEM);
	urcu_workqueue_queue_completion(workqueue, completion);
	urcu_workqueue_wait_completion(completion);
	urcu_workqueue_destroy_completion(completion);
}

/* To be used in before fork handler. */
void urcu_workqueue_pause_worker(struct urcu_workqueue *workqueue)
{
	uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSE);
	cmm_smp_mb__after_uatomic_or();
	wake_worker_thread(workqueue);

	while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) == 0)
		(void) poll(NULL, 0, 1);
}

/* To be used in after fork parent handler. */
void urcu_workqueue_resume_worker(struct urcu_workqueue *workqueue)
{
	uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSE);
	while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) != 0)
		(void) poll(NULL, 0, 1);
}

void urcu_workqueue_create_worker(struct urcu_workqueue *workqueue)
{
	int ret;

	/* Clear workqueue state from parent. */
	workqueue->flags &= ~URCU_WORKQUEUE_PAUSED;
	workqueue->flags &= ~URCU_WORKQUEUE_PAUSE;
	workqueue->tid = 0;
	ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
	if (ret) {
		urcu_die(ret);
	}
}
Commit	Line	Data
ccacf27f MD	1	/*
	2	* workqueue.c
	3	*
	4	* Userspace RCU library - Userspace workqeues
	5	*
	6	* Copyright (c) 2010 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
	7	* Copyright (c) 2017 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
	8	*
	9	* This library is free software; you can redistribute it and/or
	10	* modify it under the terms of the GNU Lesser General Public
	11	* License as published by the Free Software Foundation; either
	12	* version 2.1 of the License, or (at your option) any later version.
	13	*
	14	* This library is distributed in the hope that it will be useful,
	15	* but WITHOUT ANY WARRANTY; without even the implied warranty of
	16	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
	17	* Lesser General Public License for more details.
	18	*
	19	* You should have received a copy of the GNU Lesser General Public
	20	* License along with this library; if not, write to the Free Software
	21	* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
	22	*/
	23
	24	#define _LGPL_SOURCE
	25	#include <stdio.h>
	26	#include <pthread.h>
	27	#include <signal.h>
	28	#include <assert.h>
	29	#include <stdlib.h>
	30	#include <stdint.h>
	31	#include <string.h>
	32	#include <errno.h>
	33	#include <poll.h>
	34	#include <sys/time.h>
	35	#include <unistd.h>
	36	#include <sched.h>
	37
	38	#include "compat-getcpu.h"
6cd23d47 MD	39	#include <urcu/wfcqueue.h>
	40	#include <urcu/pointer.h>
	41	#include <urcu/list.h>
	42	#include <urcu/futex.h>
	43	#include <urcu/tls-compat.h>
	44	#include <urcu/ref.h>
ccacf27f MD	45	#include "urcu-die.h"
	46
	47	#include "workqueue.h"
	48
	49	#define SET_AFFINITY_CHECK_PERIOD (1U << 8) /* 256 */
	50	#define SET_AFFINITY_CHECK_PERIOD_MASK (SET_AFFINITY_CHECK_PERIOD - 1)
	51
	52	/* Data structure that identifies a workqueue. */
	53
	54	struct urcu_workqueue {
	55	/*
	56	* We do not align head on a different cache-line than tail
1b6fc173 MD	57	* mainly because workqueue threads use batching ("splice") to
	58	* get an entire list of callbacks, which effectively empties
	59	* the queue, and requires to touch the tail anyway.
ccacf27f MD	60	*/
	61	struct cds_wfcq_tail cbs_tail;
	62	struct cds_wfcq_head cbs_head;
	63	unsigned long flags;
	64	int32_t futex;
	65	unsigned long qlen; /* maintained for debugging. */
	66	pthread_t tid;
	67	int cpu_affinity;
	68	unsigned long loop_count;
	69	void *priv;
	70	void (grace_period_fct)(struct urcu_workqueue workqueue, void *priv);
	71	void (initialize_worker_fct)(struct urcu_workqueue workqueue, void *priv);
	72	void (finalize_worker_fct)(struct urcu_workqueue workqueue, void *priv);
	73	void (worker_before_pause_fct)(struct urcu_workqueue workqueue, void *priv);
	74	void (worker_after_resume_fct)(struct urcu_workqueue workqueue, void *priv);
	75	void (worker_before_wait_fct)(struct urcu_workqueue workqueue, void *priv);
	76	void (worker_after_wake_up_fct)(struct urcu_workqueue workqueue, void *priv);
	77	} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
	78
	79	struct urcu_workqueue_completion {
	80	int barrier_count;
	81	int32_t futex;
	82	struct urcu_ref ref;
	83	};
	84
	85	struct urcu_workqueue_completion_work {
	86	struct urcu_work work;
	87	struct urcu_workqueue_completion *completion;
	88	};
	89
	90	/*
	91	* Periodically retry setting CPU affinity if we migrate.
	92	* Losing affinity can be caused by CPU hotunplug/hotplug, or by
	93	* cpuset(7).
	94	*/
a9e31859	95	#ifdef HAVE_SCHED_SETAFFINITY
ccacf27f MD	96	static int set_thread_cpu_affinity(struct urcu_workqueue *workqueue)
	97	{
	98	cpu_set_t mask;
	99	int ret;
	100
	101	if (workqueue->cpu_affinity < 0)
	102	return 0;
	103	if (++workqueue->loop_count & SET_AFFINITY_CHECK_PERIOD_MASK)
	104	return 0;
	105	if (urcu_sched_getcpu() == workqueue->cpu_affinity)
	106	return 0;
	107
	108	CPU_ZERO(&mask);
	109	CPU_SET(workqueue->cpu_affinity, &mask);
	110	#if SCHED_SETAFFINITY_ARGS == 2
	111	ret = sched_setaffinity(0, &mask);
	112	#else
	113	ret = sched_setaffinity(0, sizeof(mask), &mask);
	114	#endif
	115	/*
	116	* EINVAL is fine: can be caused by hotunplugged CPUs, or by
	117	* cpuset(7). This is why we should always retry if we detect
	118	* migration.
	119	*/
	120	if (ret && errno == EINVAL) {
	121	ret = 0;
	122	errno = 0;
	123	}
	124	return ret;
	125	}
	126	#else
0d657320	127	static int set_thread_cpu_affinity(struct urcu_workqueue *workqueue __attribute__((unused)))
ccacf27f MD	128	{
	129	return 0;
	130	}
	131	#endif
	132
	133	static void futex_wait(int32_t *futex)
	134	{
	135	/* Read condition before read futex */
	136	cmm_smp_mb();
	137	if (uatomic_read(futex) != -1)
	138	return;
	139	while (futex_async(futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
	140	switch (errno) {
	141	case EWOULDBLOCK:
	142	/* Value already changed. */
	143	return;
	144	case EINTR:
	145	/* Retry if interrupted by signal. */
	146	break; /* Get out of switch. */
	147	default:
	148	/* Unexpected error. */
	149	urcu_die(errno);
	150	}
	151	}
	152	}
	153
	154	static void futex_wake_up(int32_t *futex)
	155	{
	156	/* Write to condition before reading/writing futex */
	157	cmm_smp_mb();
	158	if (caa_unlikely(uatomic_read(futex) == -1)) {
	159	uatomic_set(futex, 0);
	160	if (futex_async(futex, FUTEX_WAKE, 1,
	161	NULL, NULL, 0) < 0)
	162	urcu_die(errno);
	163	}
	164	}
	165
	166	/* This is the code run by each worker thread. */
	167
	168	static void workqueue_thread(void arg)
	169	{
	170	unsigned long cbcount;
	171	struct urcu_workqueue workqueue = (struct urcu_workqueue ) arg;
	172	int rt = !!(uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_RT);
	173
	174	if (set_thread_cpu_affinity(workqueue))
	175	urcu_die(errno);
	176
	177	if (workqueue->initialize_worker_fct)
	178	workqueue->initialize_worker_fct(workqueue, workqueue->priv);
	179
	180	if (!rt) {
	181	uatomic_dec(&workqueue->futex);
	182	/* Decrement futex before reading workqueue */
	183	cmm_smp_mb();
	184	}
	185	for (;;) {
	186	struct cds_wfcq_head cbs_tmp_head;
	187	struct cds_wfcq_tail cbs_tmp_tail;
	188	struct cds_wfcq_node cbs, cbs_tmp_n;
	189	enum cds_wfcq_ret splice_ret;
	190
	191	if (set_thread_cpu_affinity(workqueue))
192	urcu_die(errno);
193
194	if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) {
195	/*
196	* Pause requested. Become quiescent: remove
197	* ourself from all global lists, and don't
198	* process any callback. The callback lists may
199	* still be non-empty though.
200	*/
201	if (workqueue->worker_before_pause_fct)
202	workqueue->worker_before_pause_fct(workqueue, workqueue->priv);
203	cmm_smp_mb__before_uatomic_or();
204	uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSED);
205	while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) != 0)
206	(void) poll(NULL, 0, 1);
207	uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSED);
208	cmm_smp_mb__after_uatomic_and();
209	if (workqueue->worker_after_resume_fct)
210	workqueue->worker_after_resume_fct(workqueue, workqueue->priv);
211	}
212
213	cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
214	splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
215	&cbs_tmp_tail, &workqueue->cbs_head, &workqueue->cbs_tail);
216	assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
217	assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
218	if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
219	if (workqueue->grace_period_fct)
220	workqueue->grace_period_fct(workqueue, workqueue->priv);
221	cbcount = 0;
222	__cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
223	&cbs_tmp_tail, cbs, cbs_tmp_n) {
d25f25df	224	struct urcu_work *uwp;
ccacf27f	225
d25f25df MD	226	uwp = caa_container_of(cbs,
	227	struct urcu_work, next);
	228	uwp->func(uwp);
ccacf27f MD	229	cbcount++;
	230	}
	231	uatomic_sub(&workqueue->qlen, cbcount);
	232	}
	233	if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_STOP)
	234	break;
	235	if (workqueue->worker_before_wait_fct)
	236	workqueue->worker_before_wait_fct(workqueue, workqueue->priv);
	237	if (!rt) {
	238	if (cds_wfcq_empty(&workqueue->cbs_head,
	239	&workqueue->cbs_tail)) {
	240	futex_wait(&workqueue->futex);
ccacf27f MD	241	uatomic_dec(&workqueue->futex);
	242	/*
	243	* Decrement futex before reading
1b6fc173	244	* urcu_work list.
ccacf27f MD	245	*/
ccacf27f MD	246	cmm_smp_mb();
ccacf27f MD	247	}
ccacf27f MD	248	} else {
5bbe95bb JG	249	if (cds_wfcq_empty(&workqueue->cbs_head,
	250	&workqueue->cbs_tail)) {
	251	(void) poll(NULL, 0, 10);
	252	}
ccacf27f MD	253	}
	254	if (workqueue->worker_after_wake_up_fct)
	255	workqueue->worker_after_wake_up_fct(workqueue, workqueue->priv);
	256	}
	257	if (!rt) {
	258	/*
1b6fc173	259	* Read urcu_work list before write futex.
ccacf27f MD	260	*/
	261	cmm_smp_mb();
	262	uatomic_set(&workqueue->futex, 0);
	263	}
	264	if (workqueue->finalize_worker_fct)
	265	workqueue->finalize_worker_fct(workqueue, workqueue->priv);
	266	return NULL;
	267	}
	268
	269	struct urcu_workqueue *urcu_workqueue_create(unsigned long flags,
	270	int cpu_affinity, void *priv,
	271	void (grace_period_fct)(struct urcu_workqueue workqueue, void *priv),
	272	void (initialize_worker_fct)(struct urcu_workqueue workqueue, void *priv),
	273	void (finalize_worker_fct)(struct urcu_workqueue workqueue, void *priv),
	274	void (worker_before_wait_fct)(struct urcu_workqueue workqueue, void *priv),
	275	void (worker_after_wake_up_fct)(struct urcu_workqueue workqueue, void *priv),
	276	void (worker_before_pause_fct)(struct urcu_workqueue workqueue, void *priv),
	277	void (worker_after_resume_fct)(struct urcu_workqueue workqueue, void *priv))
	278	{
	279	struct urcu_workqueue *workqueue;
	280	int ret;
	281
	282	workqueue = malloc(sizeof(*workqueue));
	283	if (workqueue == NULL)
	284	urcu_die(errno);
	285	memset(workqueue, '\0', sizeof(*workqueue));
	286	cds_wfcq_init(&workqueue->cbs_head, &workqueue->cbs_tail);
	287	workqueue->qlen = 0;
	288	workqueue->futex = 0;
	289	workqueue->flags = flags;
	290	workqueue->priv = priv;
	291	workqueue->grace_period_fct = grace_period_fct;
	292	workqueue->initialize_worker_fct = initialize_worker_fct;
	293	workqueue->finalize_worker_fct = finalize_worker_fct;
	294	workqueue->worker_before_wait_fct = worker_before_wait_fct;
	295	workqueue->worker_after_wake_up_fct = worker_after_wake_up_fct;
	296	workqueue->worker_before_pause_fct = worker_before_pause_fct;
	297	workqueue->worker_after_resume_fct = worker_after_resume_fct;
	298	workqueue->cpu_affinity = cpu_affinity;
	299	workqueue->loop_count = 0;
	300	cmm_smp_mb(); /* Structure initialized before pointer is planted. */
	301	ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
	302	if (ret) {
	303	urcu_die(ret);
	304	}
	305	return workqueue;
	306	}
	307
	308	static void wake_worker_thread(struct urcu_workqueue *workqueue)
	309	{
9db9edcc	310	if (!(_CMM_LOAD_SHARED(workqueue->flags) & URCU_WORKQUEUE_RT))
ccacf27f MD	311	futex_wake_up(&workqueue->futex);
	312	}
	313
	314	static int urcu_workqueue_destroy_worker(struct urcu_workqueue *workqueue)
	315	{
	316	int ret;
	317	void *retval;
	318
	319	uatomic_or(&workqueue->flags, URCU_WORKQUEUE_STOP);
	320	wake_worker_thread(workqueue);
	321
	322	ret = pthread_join(workqueue->tid, &retval);
	323	if (ret) {
	324	urcu_die(ret);
	325	}
	326	if (retval != NULL) {
	327	urcu_die(EINVAL);
	328	}
	329	workqueue->flags &= ~URCU_WORKQUEUE_STOP;
	330	workqueue->tid = 0;
	331	return 0;
	332	}
	333
	334	void urcu_workqueue_destroy(struct urcu_workqueue *workqueue)
	335	{
	336	if (workqueue == NULL) {
	337	return;
	338	}
	339	if (urcu_workqueue_destroy_worker(workqueue)) {
	340	urcu_die(errno);
	341	}
	342	assert(cds_wfcq_empty(&workqueue->cbs_head, &workqueue->cbs_tail));
	343	free(workqueue);
	344	}
	345
	346	void urcu_workqueue_queue_work(struct urcu_workqueue *workqueue,
	347	struct urcu_work *work,
	348	void (func)(struct urcu_work work))
	349	{
	350	cds_wfcq_node_init(&work->next);
	351	work->func = func;
	352	cds_wfcq_enqueue(&workqueue->cbs_head, &workqueue->cbs_tail, &work->next);
	353	uatomic_inc(&workqueue->qlen);
	354	wake_worker_thread(workqueue);
	355	}
	356
	357	static
	358	void free_completion(struct urcu_ref *ref)
	359	{
	360	struct urcu_workqueue_completion *completion;
	361
	362	completion = caa_container_of(ref, struct urcu_workqueue_completion, ref);
	363	free(completion);
	364	}
	365
	366	static
	367	void _urcu_workqueue_wait_complete(struct urcu_work *work)
	368	{
	369	struct urcu_workqueue_completion_work *completion_work;
	370	struct urcu_workqueue_completion *completion;
	371
	372	completion_work = caa_container_of(work, struct urcu_workqueue_completion_work, work);
	373	completion = completion_work->completion;
	374	if (!uatomic_sub_return(&completion->barrier_count, 1))
375	futex_wake_up(&completion->futex);
376	urcu_ref_put(&completion->ref, free_completion);
377	free(completion_work);
378	}
379
380	struct urcu_workqueue_completion *urcu_workqueue_create_completion(void)
381	{
382	struct urcu_workqueue_completion *completion;
383
384	completion = calloc(sizeof(*completion), 1);
385	if (!completion)
386	urcu_die(errno);
387	urcu_ref_set(&completion->ref, 1);
388	completion->barrier_count = 0;
389	return completion;
390	}
391
392	void urcu_workqueue_destroy_completion(struct urcu_workqueue_completion *completion)
393	{
394	urcu_ref_put(&completion->ref, free_completion);
395	}
396
397	void urcu_workqueue_wait_completion(struct urcu_workqueue_completion *completion)
398	{
399	/* Wait for them */
400	for (;;) {
401	uatomic_dec(&completion->futex);
402	/* Decrement futex before reading barrier_count */
403	cmm_smp_mb();
404	if (!uatomic_read(&completion->barrier_count))
405	break;
406	futex_wait(&completion->futex);
407	}
408	}
409
410	void urcu_workqueue_queue_completion(struct urcu_workqueue *workqueue,
411	struct urcu_workqueue_completion *completion)
412	{
413	struct urcu_workqueue_completion_work *work;
414
415	work = calloc(sizeof(*work), 1);
416	if (!work)
417	urcu_die(errno);
418	work->completion = completion;
419	urcu_ref_get(&completion->ref);
420	uatomic_inc(&completion->barrier_count);
421	urcu_workqueue_queue_work(workqueue, &work->work, _urcu_workqueue_wait_complete);
422	}
423
424	/*
425	* Wait for all in-flight work to complete execution.
426	*/
427	void urcu_workqueue_flush_queued_work(struct urcu_workqueue *workqueue)
428	{
429	struct urcu_workqueue_completion *completion;
430
431	completion = urcu_workqueue_create_completion();
432	if (!completion)
433	urcu_die(ENOMEM);
434	urcu_workqueue_queue_completion(workqueue, completion);
435	urcu_workqueue_wait_completion(completion);
436	urcu_workqueue_destroy_completion(completion);
437	}
438
439	/* To be used in before fork handler. */
440	void urcu_workqueue_pause_worker(struct urcu_workqueue *workqueue)
441	{
442	uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSE);
443	cmm_smp_mb__after_uatomic_or();
444	wake_worker_thread(workqueue);
445
446	while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) == 0)
447	(void) poll(NULL, 0, 1);
448	}
449
450	/* To be used in after fork parent handler. */
451	void urcu_workqueue_resume_worker(struct urcu_workqueue *workqueue)
452	{
453	uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSE);
454	while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) != 0)
455	(void) poll(NULL, 0, 1);
456	}
457
458	void urcu_workqueue_create_worker(struct urcu_workqueue *workqueue)
459	{
460	int ret;
461
462	/* Clear workqueue state from parent. */
463	workqueue->flags &= ~URCU_WORKQUEUE_PAUSED;
464	workqueue->flags &= ~URCU_WORKQUEUE_PAUSE;
465	workqueue->tid = 0;
466	ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
467	if (ret) {
468	urcu_die(ret);
469	}
470	}