X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu-call-rcu-impl.h;h=65f63ee51670ae73b9c8ead9869554e4f8f35a82;hp=36e3cf4b93a5904b5fc1e3d0ee84eccd1cc3d9ea;hb=a59f39055b5ecb77b68cf78b9839aa9e8e4ec332;hpb=cc6b0c208a9707f8a64609a51167290caeace288

diff --git a/urcu-call-rcu-impl.h b/urcu-call-rcu-impl.h
index 36e3cf4..65f63ee 100644
--- a/urcu-call-rcu-impl.h
+++ b/urcu-call-rcu-impl.h
@@ -21,6 +21,7 @@
  */
 
 #define _GNU_SOURCE
+#define _LGPL_SOURCE
 #include <stdio.h>
 #include <pthread.h>
 #include <signal.h>
@@ -35,37 +36,65 @@
 #include <sched.h>
 
 #include "config.h"
-#include "urcu/wfqueue.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"
+
+#define SET_AFFINITY_CHECK_PERIOD		(1U << 8)	/* 256 */
+#define SET_AFFINITY_CHECK_PERIOD_MASK		(SET_AFFINITY_CHECK_PERIOD - 1)
 
 /* Data structure that identifies a call_rcu thread. */
 
 struct call_rcu_data {
-	struct cds_wfq_queue cbs;
+	/*
+	 * 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 gp_count;
 	struct cds_list_head list;
 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
 
+struct call_rcu_completion {
+	int barrier_count;
+	int32_t futex;
+	struct urcu_ref ref;
+};
+
+struct call_rcu_completion_work {
+	struct rcu_head head;
+	struct call_rcu_completion *completion;
+};
+
 /*
  * List of all call_rcu_data structures to keep valgrind happy.
  * Protected by call_rcu_mutex.
  */
 
-CDS_LIST_HEAD(call_rcu_data_list);
+static CDS_LIST_HEAD(call_rcu_data_list);
 
 /* Link a thread using call_rcu() to its call_rcu thread. */
 
-static __thread struct call_rcu_data *thread_call_rcu_data;
-
-/* Guard call_rcu thread creation. */
+static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data);
 
+/*
+ * Guard call_rcu thread creation and atfork handlers.
+ */
 static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
 
 /* If a given thread does not have its own call_rcu thread, this is default. */
@@ -78,7 +107,7 @@ static struct call_rcu_data *default_call_rcu_data;
  * CPUs rather than only to specific threads.
  */
 
-#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
+#if defined(HAVE_SYSCONF) && (defined(HAVE_SCHED_GETCPU) || defined(HAVE_GETCPUID))
 
 /*
  * Pointer to array of pointers to per-CPU call_rcu_data structures
@@ -121,7 +150,7 @@ static void alloc_cpu_call_rcu_data(void)
 	}
 }
 
-#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
+#else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
 
 /*
  * per_cpu_call_rcu_data should be constant, but some functions below, used both
@@ -139,49 +168,66 @@ static void alloc_cpu_call_rcu_data(void)
 {
 }
 
-static int sched_getcpu(void)
-{
-	return -1;
-}
-
-#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
+#endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
 
 /* Acquire the specified pthread mutex. */
 
 static void call_rcu_lock(pthread_mutex_t *pmp)
 {
-	if (pthread_mutex_lock(pmp) != 0) {
-		perror("pthread_mutex_lock");
-		exit(-1);
-	}
+	int ret;
+
+	ret = pthread_mutex_lock(pmp);
+	if (ret)
+		urcu_die(ret);
 }
 
 /* Release the specified pthread mutex. */
 
 static void call_rcu_unlock(pthread_mutex_t *pmp)
 {
-	if (pthread_mutex_unlock(pmp) != 0) {
-		perror("pthread_mutex_unlock");
-		exit(-1);
-	}
+	int ret;
+
+	ret = pthread_mutex_unlock(pmp);
+	if (ret)
+		urcu_die(ret);
 }
 
+/*
+ * 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 call_rcu_data *crdp)
 {
 	cpu_set_t mask;
+	int ret;
 
 	if (crdp->cpu_affinity < 0)
 		return 0;
+	if (++crdp->gp_count & SET_AFFINITY_CHECK_PERIOD_MASK)
+		return 0;
+	if (urcu_sched_getcpu() == crdp->cpu_affinity)
+		return 0;
 
 	CPU_ZERO(&mask);
 	CPU_SET(crdp->cpu_affinity, &mask);
 #if SCHED_SETAFFINITY_ARGS == 2
-	return sched_setaffinity(0, &mask);
+	ret = sched_setaffinity(0, &mask);
 #else
-	return sched_setaffinity(0, sizeof(mask), &mask);
+	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
@@ -195,9 +241,22 @@ static void call_rcu_wait(struct call_rcu_data *crdp)
 {
 	/* Read call_rcu list before read futex */
 	cmm_smp_mb();
-	if (uatomic_read(&crdp->futex) == -1)
-		futex_async(&crdp->futex, FUTEX_WAIT, -1,
-		      NULL, NULL, 0);
+	if (uatomic_read(&crdp->futex) != -1)
+		return;
+	while (futex_async(&crdp->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 call_rcu_wake_up(struct call_rcu_data *crdp)
@@ -206,8 +265,43 @@ static void call_rcu_wake_up(struct call_rcu_data *crdp)
 	cmm_smp_mb();
 	if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) {
 		uatomic_set(&crdp->futex, 0);
-		futex_async(&crdp->futex, FUTEX_WAKE, 1,
-		      NULL, NULL, 0);
+		if (futex_async(&crdp->futex, FUTEX_WAKE, 1,
+				NULL, NULL, 0) < 0)
+			urcu_die(errno);
+	}
+}
+
+static void call_rcu_completion_wait(struct call_rcu_completion *completion)
+{
+	/* Read completion barrier count before read futex */
+	cmm_smp_mb();
+	if (uatomic_read(&completion->futex) != -1)
+		return;
+	while (futex_async(&completion->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 call_rcu_completion_wake_up(struct call_rcu_completion *completion)
+{
+	/* Write to completion barrier count before reading/writing futex */
+	cmm_smp_mb();
+	if (caa_unlikely(uatomic_read(&completion->futex) == -1)) {
+		uatomic_set(&completion->futex, 0);
+		if (futex_async(&completion->futex, FUTEX_WAKE, 1,
+				NULL, NULL, 0) < 0)
+			urcu_die(errno);
 	}
 }
 
@@ -216,60 +310,76 @@ static void call_rcu_wake_up(struct call_rcu_data *crdp)
 static void *call_rcu_thread(void *arg)
 {
 	unsigned long cbcount;
-	struct cds_wfq_node *cbs;
-	struct cds_wfq_node **cbs_tail;
-	struct call_rcu_data *crdp = (struct call_rcu_data *)arg;
-	struct rcu_head *rhp;
+	struct call_rcu_data *crdp = (struct call_rcu_data *) arg;
 	int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
 
-	if (set_thread_cpu_affinity(crdp) != 0) {
-		perror("pthread_setaffinity_np");
-		exit(-1);
-	}
+	if (set_thread_cpu_affinity(crdp))
+		urcu_die(errno);
 
 	/*
 	 * If callbacks take a read-side lock, we need to be registered.
 	 */
 	rcu_register_thread();
 
-	thread_call_rcu_data = crdp;
+	URCU_TLS(thread_call_rcu_data) = crdp;
 	if (!rt) {
 		uatomic_dec(&crdp->futex);
 		/* Decrement futex before reading call_rcu list */
 		cmm_smp_mb();
 	}
 	for (;;) {
-		if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
-			while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
-				poll(NULL, 0, 1);
-			_CMM_STORE_SHARED(crdp->cbs.head, NULL);
-			cbs_tail = (struct cds_wfq_node **)
-				uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
+		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(crdp))
+			urcu_die(errno);
+
+		if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_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.
+			 */
+			rcu_unregister_thread();
+			cmm_smp_mb__before_uatomic_or();
+			uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED);
+			while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0)
+				(void) poll(NULL, 0, 1);
+			uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED);
+			cmm_smp_mb__after_uatomic_and();
+			rcu_register_thread();
+		}
+
+		cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
+		splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
+			&cbs_tmp_tail, &crdp->cbs_head, &crdp->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) {
 			synchronize_rcu();
 			cbcount = 0;
-			do {
-				while (cbs->next == NULL &&
-				       &cbs->next != cbs_tail)
-				       	poll(NULL, 0, 1);
-				if (cbs == &crdp->cbs.dummy) {
-					cbs = cbs->next;
-					continue;
-				}
-				rhp = (struct rcu_head *)cbs;
-				cbs = cbs->next;
+			__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++;
-			} while (cbs != NULL);
+			}
 			uatomic_sub(&crdp->qlen, cbcount);
 		}
 		if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
 			break;
 		rcu_thread_offline();
 		if (!rt) {
-			if (&crdp->cbs.head
-			    == _CMM_LOAD_SHARED(crdp->cbs.tail)) {
+			if (cds_wfcq_empty(&crdp->cbs_head,
+					&crdp->cbs_tail)) {
 				call_rcu_wait(crdp);
-				poll(NULL, 0, 10);
+				(void) poll(NULL, 0, 10);
 				uatomic_dec(&crdp->futex);
 				/*
 				 * Decrement futex before reading
@@ -277,10 +387,10 @@ static void *call_rcu_thread(void *arg)
 				 */
 				cmm_smp_mb();
 			} else {
-				poll(NULL, 0, 10);
+				(void) poll(NULL, 0, 10);
 			}
 		} else {
-			poll(NULL, 0, 10);
+			(void) poll(NULL, 0, 10);
 		}
 		rcu_thread_online();
 	}
@@ -307,32 +417,31 @@ static void call_rcu_data_init(struct call_rcu_data **crdpp,
 			       int cpu_affinity)
 {
 	struct call_rcu_data *crdp;
+	int ret;
 
 	crdp = malloc(sizeof(*crdp));
-	if (crdp == NULL) {
-		fprintf(stderr, "Out of memory.\n");
-		exit(-1);
-	}
+	if (crdp == NULL)
+		urcu_die(errno);
 	memset(crdp, '\0', sizeof(*crdp));
-	cds_wfq_init(&crdp->cbs);
+	cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail);
 	crdp->qlen = 0;
 	crdp->futex = 0;
 	crdp->flags = flags;
 	cds_list_add(&crdp->list, &call_rcu_data_list);
 	crdp->cpu_affinity = cpu_affinity;
+	crdp->gp_count = 0;
 	cmm_smp_mb();  /* Structure initialized before pointer is planted. */
 	*crdpp = crdp;
-	if (pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp) != 0) {
-		perror("pthread_create");
-		exit(-1);
-	}
+	ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp);
+	if (ret)
+		urcu_die(ret);
 }
 
 /*
  * Return a pointer to the call_rcu_data structure for the specified
  * CPU, returning NULL if there is none.  We cannot automatically
  * created it because the platform we are running on might not define
- * sched_getcpu().
+ * urcu_sched_getcpu().
  *
  * The call to this function and use of the returned call_rcu_data
  * should be protected by RCU read-side lock.
@@ -470,11 +579,11 @@ struct call_rcu_data *get_call_rcu_data(void)
 {
 	struct call_rcu_data *crd;
 
-	if (thread_call_rcu_data != NULL)
-		return thread_call_rcu_data;
+	if (URCU_TLS(thread_call_rcu_data) != NULL)
+		return URCU_TLS(thread_call_rcu_data);
 
 	if (maxcpus > 0) {
-		crd = get_cpu_call_rcu_data(sched_getcpu());
+		crd = get_cpu_call_rcu_data(urcu_sched_getcpu());
 		if (crd)
 			return crd;
 	}
@@ -488,7 +597,7 @@ struct call_rcu_data *get_call_rcu_data(void)
 
 struct call_rcu_data *get_thread_call_rcu_data(void)
 {
-	return thread_call_rcu_data;
+	return URCU_TLS(thread_call_rcu_data);
 }
 
 /*
@@ -504,7 +613,7 @@ struct call_rcu_data *get_thread_call_rcu_data(void)
 
 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
 {
-	thread_call_rcu_data = crdp;
+	URCU_TLS(thread_call_rcu_data) = crdp;
 }
 
 /*
@@ -568,6 +677,17 @@ static void wake_call_rcu_thread(struct call_rcu_data *crdp)
 		call_rcu_wake_up(crdp);
 }
 
+static void _call_rcu(struct rcu_head *head,
+		      void (*func)(struct rcu_head *head),
+		      struct call_rcu_data *crdp)
+{
+	cds_wfcq_node_init(&head->next);
+	head->func = func;
+	cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next);
+	uatomic_inc(&crdp->qlen);
+	wake_call_rcu_thread(crdp);
+}
+
 /*
  * Schedule a function to be invoked after a following grace period.
  * This is the only function that must be called -- the others are
@@ -582,21 +702,16 @@ static void wake_call_rcu_thread(struct call_rcu_data *crdp)
  *
  * call_rcu must be called by registered RCU read-side threads.
  */
-
 void call_rcu(struct rcu_head *head,
 	      void (*func)(struct rcu_head *head))
 {
 	struct call_rcu_data *crdp;
 
-	cds_wfq_node_init(&head->next);
-	head->func = func;
 	/* Holding rcu read-side lock across use of per-cpu crdp */
-	rcu_read_lock();
+	_rcu_read_lock();
 	crdp = get_call_rcu_data();
-	cds_wfq_enqueue(&crdp->cbs, &head->next);
-	uatomic_inc(&crdp->qlen);
-	wake_call_rcu_thread(crdp);
-	rcu_read_unlock();
+	_call_rcu(head, func, crdp);
+	_rcu_read_unlock();
 }
 
 /*
@@ -620,13 +735,13 @@ void call_rcu(struct rcu_head *head,
  * The caller must wait for a grace-period to pass between return from
  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
  * previous call rcu data as argument.
+ *
+ * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
+ * a list corruption bug in the 0.7.x series. The equivalent fix
+ * appeared in 0.6.8 for the stable-0.6 branch.
  */
 void call_rcu_data_free(struct call_rcu_data *crdp)
 {
-	struct cds_wfq_node *cbs;
-	struct cds_wfq_node **cbs_tail;
-	struct cds_wfq_node **cbs_endprev;
-
 	if (crdp == NULL || crdp == default_call_rcu_data) {
 		return;
 	}
@@ -634,19 +749,14 @@ void call_rcu_data_free(struct call_rcu_data *crdp)
 		uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
 		wake_call_rcu_thread(crdp);
 		while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
-			poll(NULL, 0, 1);
-	}
-	if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
-		while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
-			poll(NULL, 0, 1);
-		_CMM_STORE_SHARED(crdp->cbs.head, NULL);
-		cbs_tail = (struct cds_wfq_node **)
-			uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
+			(void) poll(NULL, 0, 1);
+	}
+	if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
 		/* Create default call rcu data if need be */
 		(void) get_default_call_rcu_data();
-		cbs_endprev = (struct cds_wfq_node **)
-			uatomic_xchg(&default_call_rcu_data, cbs_tail);
-		*cbs_endprev = cbs;
+		__cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head,
+			&default_call_rcu_data->cbs_tail,
+			&crdp->cbs_head, &crdp->cbs_tail);
 		uatomic_add(&default_call_rcu_data->qlen,
 			    uatomic_read(&crdp->qlen));
 		wake_call_rcu_thread(default_call_rcu_data);
@@ -699,14 +809,118 @@ void free_all_cpu_call_rcu_data(void)
 	free(crdp);
 }
 
+static
+void free_completion(struct urcu_ref *ref)
+{
+	struct call_rcu_completion *completion;
+
+	completion = caa_container_of(ref, struct call_rcu_completion, ref);
+	free(completion);
+}
+
+static
+void _rcu_barrier_complete(struct rcu_head *head)
+{
+	struct call_rcu_completion_work *work;
+	struct call_rcu_completion *completion;
+
+	work = caa_container_of(head, struct call_rcu_completion_work, head);
+	completion = work->completion;
+	if (!uatomic_sub_return(&completion->barrier_count, 1))
+		call_rcu_completion_wake_up(completion);
+	urcu_ref_put(&completion->ref, free_completion);
+	free(work);
+}
+
+/*
+ * Wait for all in-flight call_rcu callbacks to complete execution.
+ */
+void rcu_barrier(void)
+{
+	struct call_rcu_data *crdp;
+	struct call_rcu_completion *completion;
+	int count = 0;
+	int was_online;
+
+	/* Put in offline state in QSBR. */
+	was_online = _rcu_read_ongoing();
+	if (was_online)
+		rcu_thread_offline();
+	/*
+	 * Calling a rcu_barrier() within a RCU read-side critical
+	 * section is an error.
+	 */
+	if (_rcu_read_ongoing()) {
+		static int warned = 0;
+
+		if (!warned) {
+			fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
+		}
+		warned = 1;
+		goto online;
+	}
+
+	completion = calloc(sizeof(*completion), 1);
+	if (!completion)
+		urcu_die(errno);
+
+	call_rcu_lock(&call_rcu_mutex);
+	cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
+		count++;
+
+	/* Referenced by rcu_barrier() and each call_rcu thread. */
+	urcu_ref_set(&completion->ref, count + 1);
+	completion->barrier_count = count;
+
+	cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
+		struct call_rcu_completion_work *work;
+
+		work = calloc(sizeof(*work), 1);
+		if (!work)
+			urcu_die(errno);
+		work->completion = completion;
+		_call_rcu(&work->head, _rcu_barrier_complete, crdp);
+	}
+	call_rcu_unlock(&call_rcu_mutex);
+
+	/* Wait for them */
+	for (;;) {
+		uatomic_dec(&completion->futex);
+		/* Decrement futex before reading barrier_count */
+		cmm_smp_mb();
+		if (!uatomic_read(&completion->barrier_count))
+			break;
+		call_rcu_completion_wait(completion);
+	}
+
+	urcu_ref_put(&completion->ref, free_completion);
+
+online:
+	if (was_online)
+		rcu_thread_online();
+}
+
 /*
  * Acquire the call_rcu_mutex in order to ensure that the child sees
- * all of the call_rcu() data structures in a consistent state.
+ * all of the call_rcu() data structures in a consistent state. Ensure
+ * that all call_rcu threads are in a quiescent state across fork.
  * Suitable for pthread_atfork() and friends.
  */
 void call_rcu_before_fork(void)
 {
+	struct call_rcu_data *crdp;
+
 	call_rcu_lock(&call_rcu_mutex);
+
+	cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
+		uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE);
+		cmm_smp_mb__after_uatomic_or();
+		wake_call_rcu_thread(crdp);
+	}
+	cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
+		while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0)
+			(void) poll(NULL, 0, 1);
+	}
 }
 
 /*
@@ -716,6 +930,14 @@ void call_rcu_before_fork(void)
  */
 void call_rcu_after_fork_parent(void)
 {
+	struct call_rcu_data *crdp;
+
+	cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
+		uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE);
+	cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
+		while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0)
+			(void) poll(NULL, 0, 1);
+	}
 	call_rcu_unlock(&call_rcu_mutex);
 }
 
@@ -746,9 +968,13 @@ void call_rcu_after_fork_child(void)
 	maxcpus_reset();
 	free(per_cpu_call_rcu_data);
 	rcu_set_pointer(&per_cpu_call_rcu_data, NULL);
-	thread_call_rcu_data = NULL;
+	URCU_TLS(thread_call_rcu_data) = NULL;
 
-	/* Dispose of all of the rest of the call_rcu_data structures. */
+	/*
+	 * Dispose of all of the rest of the call_rcu_data structures.
+	 * Leftover call_rcu callbacks will be merged into the new
+	 * default call_rcu thread queue.
+	 */
 	cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) {
 		if (crdp == default_call_rcu_data)
 			continue;