[urcu.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 <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/assert.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);
	ret = sched_setaffinity(0, sizeof(mask), &mask);

	/*
	 * 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)
{
	int ret;

	/* Read condition before read futex */
	cmm_smp_mb();
	while (uatomic_read(futex) == -1) {
		if (!futex_async(futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
			/*
			 * Prior queued wakeups queued by unrelated code
			 * using the same address can cause futex wait to
			 * return 0 even through the futex value is still
			 * -1 (spurious wakeups). Check the value again
			 * in user-space to validate whether it really
			 * differs from -1.
			 */
			continue;
		}
		switch (errno) {
		case EAGAIN:
			/* Value already changed. */
			return;
		case EINTR:
			/* Retry if interrupted by signal. */
			break;	/* Get out of switch. Check again. */
		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);
		urcu_posix_assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
		urcu_posix_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);
	}
	urcu_posix_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>
ccacf27f MD	28	#include <stdlib.h>
	29	#include <stdint.h>
	30	#include <string.h>
	31	#include <errno.h>
	32	#include <poll.h>
	33	#include <sys/time.h>
	34	#include <unistd.h>
	35	#include <sched.h>
	36
	37	#include "compat-getcpu.h"
01477510	38	#include <urcu/assert.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	*/
2388c075	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);
ccacf27f	110	ret = sched_setaffinity(0, sizeof(mask), &mask);
0614a2e6	111
ccacf27f MD	112	/*
	113	* EINVAL is fine: can be caused by hotunplugged CPUs, or by
	114	* cpuset(7). This is why we should always retry if we detect
	115	* migration.
	116	*/
	117	if (ret && errno == EINVAL) {
	118	ret = 0;
	119	errno = 0;
	120	}
	121	return ret;
	122	}
	123	#else
a142df4e	124	static int set_thread_cpu_affinity(struct urcu_workqueue *workqueue __attribute__((unused)))
ccacf27f MD	125	{
	126	return 0;
	127	}
	128	#endif
	129
	130	static void futex_wait(int32_t *futex)
	131	{
80c34c2d MD	132	int ret;
80c34c2d MD	133
ccacf27f MD	134	/* Read condition before read futex */
ccacf27f MD	135	cmm_smp_mb();
80c34c2d MD	136	while (uatomic_read(futex) == -1) {
	137	if (!futex_async(futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
	138	/*
	139	* Prior queued wakeups queued by unrelated code
	140	* using the same address can cause futex wait to
	141	* return 0 even through the futex value is still
	142	* -1 (spurious wakeups). Check the value again
	143	* in user-space to validate whether it really
	144	* differs from -1.
	145	*/
	146	continue;
	147	}
ccacf27f	148	switch (errno) {
80c34c2d	149	case EAGAIN:
ccacf27f MD	150	/* Value already changed. */
	151	return;
	152	case EINTR:
	153	/* Retry if interrupted by signal. */
80c34c2d	154	break; /* Get out of switch. Check again. */
ccacf27f MD	155	default:
	156	/* Unexpected error. */
	157	urcu_die(errno);
	158	}
	159	}
	160	}
	161
	162	static void futex_wake_up(int32_t *futex)
	163	{
	164	/* Write to condition before reading/writing futex */
	165	cmm_smp_mb();
	166	if (caa_unlikely(uatomic_read(futex) == -1)) {
	167	uatomic_set(futex, 0);
	168	if (futex_async(futex, FUTEX_WAKE, 1,
	169	NULL, NULL, 0) < 0)
	170	urcu_die(errno);
	171	}
	172	}
	173
	174	/* This is the code run by each worker thread. */
	175
	176	static void workqueue_thread(void arg)
	177	{
	178	unsigned long cbcount;
	179	struct urcu_workqueue workqueue = (struct urcu_workqueue ) arg;
	180	int rt = !!(uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_RT);
	181
	182	if (set_thread_cpu_affinity(workqueue))
	183	urcu_die(errno);
	184
	185	if (workqueue->initialize_worker_fct)
	186	workqueue->initialize_worker_fct(workqueue, workqueue->priv);
	187
	188	if (!rt) {
	189	uatomic_dec(&workqueue->futex);
	190	/* Decrement futex before reading workqueue */
	191	cmm_smp_mb();
	192	}
	193	for (;;) {
	194	struct cds_wfcq_head cbs_tmp_head;
	195	struct cds_wfcq_tail cbs_tmp_tail;
	196	struct cds_wfcq_node cbs, cbs_tmp_n;
	197	enum cds_wfcq_ret splice_ret;
	198
	199	if (set_thread_cpu_affinity(workqueue))
	200	urcu_die(errno);
	201
	202	if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) {
	203	/*
	204	* Pause requested. Become quiescent: remove
	205	* ourself from all global lists, and don't
	206	* process any callback. The callback lists may
	207	* still be non-empty though.
	208	*/
	209	if (workqueue->worker_before_pause_fct)
	210	workqueue->worker_before_pause_fct(workqueue, workqueue->priv);
	211	cmm_smp_mb__before_uatomic_or();
	212	uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSED);
	213	while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) != 0)
	214	(void) poll(NULL, 0, 1);
	215	uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSED);
	216	cmm_smp_mb__after_uatomic_and();
	217	if (workqueue->worker_after_resume_fct)
	218	workqueue->worker_after_resume_fct(workqueue, workqueue->priv);
219	}
220
221	cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
222	splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
223	&cbs_tmp_tail, &workqueue->cbs_head, &workqueue->cbs_tail);
01477510 FD	224	urcu_posix_assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
01477510 FD	225	urcu_posix_assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
ccacf27f MD	226	if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
	227	if (workqueue->grace_period_fct)
	228	workqueue->grace_period_fct(workqueue, workqueue->priv);
	229	cbcount = 0;
	230	__cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
	231	&cbs_tmp_tail, cbs, cbs_tmp_n) {
d25f25df	232	struct urcu_work *uwp;
ccacf27f	233
d25f25df MD	234	uwp = caa_container_of(cbs,
	235	struct urcu_work, next);
	236	uwp->func(uwp);
ccacf27f MD	237	cbcount++;
	238	}
	239	uatomic_sub(&workqueue->qlen, cbcount);
	240	}
	241	if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_STOP)
	242	break;
	243	if (workqueue->worker_before_wait_fct)
	244	workqueue->worker_before_wait_fct(workqueue, workqueue->priv);
	245	if (!rt) {
	246	if (cds_wfcq_empty(&workqueue->cbs_head,
	247	&workqueue->cbs_tail)) {
	248	futex_wait(&workqueue->futex);
ccacf27f MD	249	uatomic_dec(&workqueue->futex);
	250	/*
	251	* Decrement futex before reading
1b6fc173	252	* urcu_work list.
ccacf27f MD	253	*/
ccacf27f MD	254	cmm_smp_mb();
ccacf27f MD	255	}
ccacf27f MD	256	} else {
5bbe95bb JG	257	if (cds_wfcq_empty(&workqueue->cbs_head,
	258	&workqueue->cbs_tail)) {
	259	(void) poll(NULL, 0, 10);
	260	}
ccacf27f MD	261	}
	262	if (workqueue->worker_after_wake_up_fct)
	263	workqueue->worker_after_wake_up_fct(workqueue, workqueue->priv);
	264	}
	265	if (!rt) {
	266	/*
1b6fc173	267	* Read urcu_work list before write futex.
ccacf27f MD	268	*/
	269	cmm_smp_mb();
	270	uatomic_set(&workqueue->futex, 0);
	271	}
	272	if (workqueue->finalize_worker_fct)
	273	workqueue->finalize_worker_fct(workqueue, workqueue->priv);
	274	return NULL;
	275	}
	276
	277	struct urcu_workqueue *urcu_workqueue_create(unsigned long flags,
	278	int cpu_affinity, void *priv,
	279	void (grace_period_fct)(struct urcu_workqueue workqueue, void *priv),
	280	void (initialize_worker_fct)(struct urcu_workqueue workqueue, void *priv),
	281	void (finalize_worker_fct)(struct urcu_workqueue workqueue, void *priv),
	282	void (worker_before_wait_fct)(struct urcu_workqueue workqueue, void *priv),
	283	void (worker_after_wake_up_fct)(struct urcu_workqueue workqueue, void *priv),
	284	void (worker_before_pause_fct)(struct urcu_workqueue workqueue, void *priv),
	285	void (worker_after_resume_fct)(struct urcu_workqueue workqueue, void *priv))
	286	{
	287	struct urcu_workqueue *workqueue;
	288	int ret;
	289
	290	workqueue = malloc(sizeof(*workqueue));
	291	if (workqueue == NULL)
	292	urcu_die(errno);
	293	memset(workqueue, '\0', sizeof(*workqueue));
	294	cds_wfcq_init(&workqueue->cbs_head, &workqueue->cbs_tail);
	295	workqueue->qlen = 0;
	296	workqueue->futex = 0;
	297	workqueue->flags = flags;
	298	workqueue->priv = priv;
	299	workqueue->grace_period_fct = grace_period_fct;
	300	workqueue->initialize_worker_fct = initialize_worker_fct;
	301	workqueue->finalize_worker_fct = finalize_worker_fct;
	302	workqueue->worker_before_wait_fct = worker_before_wait_fct;
	303	workqueue->worker_after_wake_up_fct = worker_after_wake_up_fct;
	304	workqueue->worker_before_pause_fct = worker_before_pause_fct;
	305	workqueue->worker_after_resume_fct = worker_after_resume_fct;
	306	workqueue->cpu_affinity = cpu_affinity;
	307	workqueue->loop_count = 0;
	308	cmm_smp_mb(); /* Structure initialized before pointer is planted. */
	309	ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
	310	if (ret) {
	311	urcu_die(ret);
	312	}
	313	return workqueue;
	314	}
	315
	316	static void wake_worker_thread(struct urcu_workqueue *workqueue)
	317	{
9db9edcc	318	if (!(_CMM_LOAD_SHARED(workqueue->flags) & URCU_WORKQUEUE_RT))
ccacf27f MD	319	futex_wake_up(&workqueue->futex);
	320	}
	321
	322	static int urcu_workqueue_destroy_worker(struct urcu_workqueue *workqueue)
	323	{
	324	int ret;
	325	void *retval;
	326
	327	uatomic_or(&workqueue->flags, URCU_WORKQUEUE_STOP);
	328	wake_worker_thread(workqueue);
	329
	330	ret = pthread_join(workqueue->tid, &retval);
	331	if (ret) {
	332	urcu_die(ret);
	333	}
	334	if (retval != NULL) {
	335	urcu_die(EINVAL);
	336	}
	337	workqueue->flags &= ~URCU_WORKQUEUE_STOP;
	338	workqueue->tid = 0;
	339	return 0;
	340	}
	341
	342	void urcu_workqueue_destroy(struct urcu_workqueue *workqueue)
	343	{
	344	if (workqueue == NULL) {
	345	return;
	346	}
	347	if (urcu_workqueue_destroy_worker(workqueue)) {
	348	urcu_die(errno);
	349	}
01477510	350	urcu_posix_assert(cds_wfcq_empty(&workqueue->cbs_head, &workqueue->cbs_tail));
ccacf27f MD	351	free(workqueue);
	352	}
	353
	354	void urcu_workqueue_queue_work(struct urcu_workqueue *workqueue,
	355	struct urcu_work *work,
	356	void (func)(struct urcu_work work))
	357	{
	358	cds_wfcq_node_init(&work->next);
	359	work->func = func;
	360	cds_wfcq_enqueue(&workqueue->cbs_head, &workqueue->cbs_tail, &work->next);
	361	uatomic_inc(&workqueue->qlen);
	362	wake_worker_thread(workqueue);
	363	}
	364
	365	static
	366	void free_completion(struct urcu_ref *ref)
	367	{
	368	struct urcu_workqueue_completion *completion;
	369
	370	completion = caa_container_of(ref, struct urcu_workqueue_completion, ref);
	371	free(completion);
	372	}
	373
	374	static
	375	void _urcu_workqueue_wait_complete(struct urcu_work *work)
	376	{
	377	struct urcu_workqueue_completion_work *completion_work;
	378	struct urcu_workqueue_completion *completion;
	379
	380	completion_work = caa_container_of(work, struct urcu_workqueue_completion_work, work);
	381	completion = completion_work->completion;
	382	if (!uatomic_sub_return(&completion->barrier_count, 1))
	383	futex_wake_up(&completion->futex);
	384	urcu_ref_put(&completion->ref, free_completion);
	385	free(completion_work);
	386	}
	387
	388	struct urcu_workqueue_completion *urcu_workqueue_create_completion(void)
	389	{
	390	struct urcu_workqueue_completion *completion;
	391
	392	completion = calloc(sizeof(*completion), 1);
	393	if (!completion)
	394	urcu_die(errno);
	395	urcu_ref_set(&completion->ref, 1);
	396	completion->barrier_count = 0;
	397	return completion;
	398	}
	399
	400	void urcu_workqueue_destroy_completion(struct urcu_workqueue_completion *completion)
	401	{
	402	urcu_ref_put(&completion->ref, free_completion);
	403	}
	404
	405	void urcu_workqueue_wait_completion(struct urcu_workqueue_completion *completion)
	406	{
	407	/* Wait for them */
	408	for (;;) {
	409	uatomic_dec(&completion->futex);
	410	/* Decrement futex before reading barrier_count */
	411	cmm_smp_mb();
	412	if (!uatomic_read(&completion->barrier_count))
	413	break;
	414	futex_wait(&completion->futex);
415	}
416	}
417
418	void urcu_workqueue_queue_completion(struct urcu_workqueue *workqueue,
419	struct urcu_workqueue_completion *completion)
420	{
421	struct urcu_workqueue_completion_work *work;
422
423	work = calloc(sizeof(*work), 1);
424	if (!work)
425	urcu_die(errno);
426	work->completion = completion;
427	urcu_ref_get(&completion->ref);
428	uatomic_inc(&completion->barrier_count);
429	urcu_workqueue_queue_work(workqueue, &work->work, _urcu_workqueue_wait_complete);
430	}
431
432	/*
433	* Wait for all in-flight work to complete execution.
434	*/
435	void urcu_workqueue_flush_queued_work(struct urcu_workqueue *workqueue)
436	{
437	struct urcu_workqueue_completion *completion;
438
439	completion = urcu_workqueue_create_completion();
440	if (!completion)
441	urcu_die(ENOMEM);
442	urcu_workqueue_queue_completion(workqueue, completion);
443	urcu_workqueue_wait_completion(completion);
444	urcu_workqueue_destroy_completion(completion);
445	}
446
447	/* To be used in before fork handler. */
448	void urcu_workqueue_pause_worker(struct urcu_workqueue *workqueue)
449	{
450	uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSE);
451	cmm_smp_mb__after_uatomic_or();
452	wake_worker_thread(workqueue);
453
454	while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) == 0)
455	(void) poll(NULL, 0, 1);
456	}
457
458	/* To be used in after fork parent handler. */
459	void urcu_workqueue_resume_worker(struct urcu_workqueue *workqueue)
460	{
461	uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSE);
462	while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) != 0)
463	(void) poll(NULL, 0, 1);
464	}
465
466	void urcu_workqueue_create_worker(struct urcu_workqueue *workqueue)
467	{
468	int ret;
469
470	/* Clear workqueue state from parent. */
471	workqueue->flags &= ~URCU_WORKQUEUE_PAUSED;
472	workqueue->flags &= ~URCU_WORKQUEUE_PAUSE;
473	workqueue->tid = 0;
474	ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
475	if (ret) {
476	urcu_die(ret);
477	}
478	}