From: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Date: Tue, 30 May 2017 19:14:58 +0000 (-0400)
Subject: Implement urcu workqueues internal API
X-Git-Tag: v0.10.0~4
X-Git-Url: https://git.lttng.org./?a=commitdiff_plain;h=ccacf27fd8d1af069016d7cf3a91f8995af08b8c;p=userspace-rcu.git

Implement urcu workqueues internal API

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
---

diff --git a/src/Makefile.am b/src/Makefile.am
index a801020..60b833d 100644
--- a/src/Makefile.am
+++ b/src/Makefile.am
@@ -54,7 +54,7 @@ liburcu_bp_la_SOURCES = urcu-bp.c urcu-pointer.c $(COMPAT)
 liburcu_bp_la_LIBADD = liburcu-common.la
 
 liburcu_cds_la_SOURCES = rculfqueue.c rculfstack.c lfstack.c \
-	$(RCULFHASH) $(COMPAT)
+	workqueue.c workqueue.h $(RCULFHASH) $(COMPAT)
 liburcu_cds_la_LIBADD = liburcu-common.la
 
 pkgconfigdir = $(libdir)/pkgconfig
diff --git a/src/workqueue.c b/src/workqueue.c
new file mode 100644
index 0000000..17ea835
--- /dev/null
+++ b/src/workqueue.c
@@ -0,0 +1,472 @@
+/*
+ * 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-call-rcu.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 call_rcu callback-invocation 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).
+ */
+#if 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)
+{
+	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 rcu_head *rhp;
+
+				rhp = caa_container_of(cbs,
+					struct rcu_head, next);
+				rhp->func(rhp);
+				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);
+				(void) poll(NULL, 0, 10);
+				uatomic_dec(&workqueue->futex);
+				/*
+				 * Decrement futex before reading
+				 * call_rcu list.
+				 */
+				cmm_smp_mb();
+			} else {
+				(void) poll(NULL, 0, 10);
+			}
+		} else {
+			(void) poll(NULL, 0, 10);
+		}
+		if (workqueue->worker_after_wake_up_fct)
+			workqueue->worker_after_wake_up_fct(workqueue, workqueue->priv);
+	}
+	if (!rt) {
+		/*
+		 * Read call_rcu 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_CALL_RCU_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);
+	}
+}
diff --git a/src/workqueue.h b/src/workqueue.h
new file mode 100644
index 0000000..52b6973
--- /dev/null
+++ b/src/workqueue.h
@@ -0,0 +1,104 @@
+#ifndef _URCU_WORKQUEUE_H
+#define _URCU_WORKQUEUE_H
+
+/*
+ * workqueue.h
+ *
+ * Userspace RCU header - Userspace workqueues
+ *
+ * Copyright (c) 2009,2017 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+ * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
+ *
+ * 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 <stdlib.h>
+#include <pthread.h>
+
+#include <urcu/wfcqueue.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Note that struct urcu_workqueue is opaque to callers. */
+
+struct urcu_workqueue;
+struct urcu_workqueue_completion;
+
+/* Flag values. */
+
+#define URCU_WORKQUEUE_RT	(1U << 0)
+#define URCU_WORKQUEUE_STOP	(1U << 1)
+#define URCU_WORKQUEUE_PAUSE	(1U << 2)
+#define URCU_WORKQUEUE_PAUSED	(1U << 3)
+
+/*
+ * The urcu_work data structure is placed in the structure to be acted
+ * upon via urcu_workqueue_queue_work().
+ */
+
+struct urcu_work {
+	struct cds_wfcq_node next;
+	void (*func)(struct urcu_work *head);
+};
+
+/*
+ * Exported functions
+ */
+
+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));
+void urcu_workqueue_destroy(struct urcu_workqueue *workqueue);
+
+/*
+ * Never fails. Should not be used to enqueue work from worker threads
+ * after the application invokes urcu_workqueue_free.
+ */
+void urcu_workqueue_queue_work(struct urcu_workqueue *workqueue,
+		struct urcu_work *work,
+		void (*func)(struct urcu_work *work));
+
+struct urcu_workqueue_completion *urcu_workqueue_create_completion(void);
+void urcu_workqueue_destroy_completion(struct urcu_workqueue_completion *completion);
+
+void urcu_workqueue_queue_completion(struct urcu_workqueue *workqueue,
+		struct urcu_workqueue_completion *completion);
+void urcu_workqueue_wait_completion(struct urcu_workqueue_completion *completion);
+
+void urcu_workqueue_flush_queued_work(struct urcu_workqueue *workqueue);
+
+/*
+ * pause/resume/create worker threads. Can be used to pause worker
+ * threads across fork/clone while keeping the workqueue in place.
+ * Pause is used in parent pre-fork, resume in parent post-fork, create
+ * in child after-fork.
+ */
+void urcu_workqueue_pause_worker(struct urcu_workqueue *workqueue);
+void urcu_workqueue_resume_worker(struct urcu_workqueue *workqueue);
+void urcu_workqueue_create_worker(struct urcu_workqueue *workqueue);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* _URCU_WORKQUEUE_H */