Fix: deadlock when thread join is issued in read-side C.S.

The transitive dependency between:

RCU read-side C.S. -> synchronize_rcu -> rcu_gp_lock -> rcu_register_thread

and the dependency:

pthread_join -> awaiting for thread completion

Can block a thread on join, and thus have the side-effect of deadlocking
a thread doing a pthread_join while within a RCU read-side critical
section. This join would be awaiting for completion of register_thread or
rcu_unregister_thread, which may never complete because the rcu_gp_lock
is held by synchronize_rcu executed from another thread.

One solution to fix this is to add a new lock, rcu_registry_lock. This
lock now protects the thread registry. It is released between iterations
on the registry by synchronize_rcu, thus allowing thread
registration/unregistration to complete even though synchronize_rcu is
awaiting for RCU read-side critical sections to complete.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
CC: Eugene Ivanov <Eugene.Ivanov@orc-group.com>
CC: Lai Jiangshan <laijs@cn.fujitsu.com>
CC: Stephen Hemminger <stephen@networkplumber.org>

diff --git a/urcu-bp.c b/urcu-bp.c
index 6b2875d2566da7e16e5767c171aee56186eb03be..4dc402865a026800ccef0bb35378be6241c81505 100644 (file)
--- a/urcu-bp.c
+++ b/urcu-bp.c
@@ -99,7 +99,21 @@ void __attribute__((constructor)) rcu_bp_init(void);
 static
 void __attribute__((destructor)) rcu_bp_exit(void);
 
+/*
+ * rcu_gp_lock ensures mutual exclusion between threads calling
+ * synchronize_rcu().
+ */
 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
+/*
+ * rcu_registry_lock ensures mutual exclusion between threads
+ * registering and unregistering themselves to/from the registry, and
+ * with threads reading that registry from synchronize_rcu(). However,
+ * this lock is not held all the way through the completion of awaiting
+ * for the grace period. It is sporadically released between iterations
+ * on the registry.
+ * rcu_registry_lock may nest inside rcu_gp_lock.
+ */
+static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
 
 static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER;
 static int initialized;
@@ -160,6 +174,10 @@ static void mutex_unlock(pthread_mutex_t *mutex)
                urcu_die(ret);
 }
 
+/*
+ * Always called with rcu_registry lock held. Releases this lock between
+ * iterations and grabs it again. Holds the lock when it returns.
+ */
 static void wait_for_readers(struct cds_list_head *input_readers,
                        struct cds_list_head *cur_snap_readers,
                        struct cds_list_head *qsreaders)
@@ -202,10 +220,14 @@ static void wait_for_readers(struct cds_list_head *input_readers,
                if (cds_list_empty(input_readers)) {
                        break;
                } else {
+                       /* Temporarily unlock the registry lock. */
+                       mutex_unlock(&rcu_registry_lock);
                        if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
                                (void) poll(NULL, 0, RCU_SLEEP_DELAY_MS);
                        else
                                caa_cpu_relax();
+                       /* Re-lock the registry lock before the next loop. */
+                       mutex_lock(&rcu_registry_lock);
                }
        }
 }
@@ -224,6 +246,8 @@ void synchronize_rcu(void)
 
        mutex_lock(&rcu_gp_lock);
 
+       mutex_lock(&rcu_registry_lock);
+
        if (cds_list_empty(&registry))
                goto out;
 
@@ -234,6 +258,8 @@ void synchronize_rcu(void)
 
        /*
         * Wait for readers to observe original parity or be quiescent.
+        * wait_for_readers() can release and grab again rcu_registry_lock
+        * interally.
         */
        wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 
@@ -263,6 +289,8 @@ void synchronize_rcu(void)
 
        /*
         * Wait for readers to observe new parity or be quiescent.
+        * wait_for_readers() can release and grab again rcu_registry_lock
+        * interally.
         */
        wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 
@@ -277,6 +305,7 @@ void synchronize_rcu(void)
         */
        cmm_smp_mb();
 out:
+       mutex_unlock(&rcu_registry_lock);
        mutex_unlock(&rcu_gp_lock);
        ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
        assert(!ret);
@@ -485,9 +514,9 @@ void rcu_bp_register(void)
         */
        rcu_bp_init();
 
-       mutex_lock(&rcu_gp_lock);
+       mutex_lock(&rcu_registry_lock);
        add_thread();
-       mutex_unlock(&rcu_gp_lock);
+       mutex_unlock(&rcu_registry_lock);
 end:
        ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
        if (ret)
@@ -508,9 +537,9 @@ void rcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
        if (ret)
                abort();
 
-       mutex_lock(&rcu_gp_lock);
+       mutex_lock(&rcu_registry_lock);
        remove_thread(rcu_reader_reg);
-       mutex_unlock(&rcu_gp_lock);
+       mutex_unlock(&rcu_registry_lock);
        ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
        if (ret)
                abort();
@@ -564,9 +593,10 @@ void rcu_bp_exit(void)
 }
 
 /*
- * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
- * a concurrent thread executing with this same lock held. This ensures that the
- * registry is in a coherent state in the child.
+ * Holding the rcu_gp_lock and rcu_registry_lock across fork will make
+ * sure we fork() don't race with a concurrent thread executing with
+ * any of those locks held. This ensures that the registry and data
+ * protected by rcu_gp_lock are in a coherent state in the child.
  */
 void rcu_bp_before_fork(void)
 {
@@ -578,6 +608,7 @@ void rcu_bp_before_fork(void)
        ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
        assert(!ret);
        mutex_lock(&rcu_gp_lock);
+       mutex_lock(&rcu_registry_lock);
        saved_fork_signal_mask = oldmask;
 }
 
@@ -587,6 +618,7 @@ void rcu_bp_after_fork_parent(void)
        int ret;
 
        oldmask = saved_fork_signal_mask;
+       mutex_unlock(&rcu_registry_lock);
        mutex_unlock(&rcu_gp_lock);
        ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
        assert(!ret);
@@ -594,7 +626,7 @@ void rcu_bp_after_fork_parent(void)
 
 /*
  * Prune all entries from registry except our own thread. Fits the Linux
- * fork behavior. Called with rcu_gp_lock held.
+ * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held.
  */
 static
 void urcu_bp_prune_registry(void)
@@ -622,6 +654,7 @@ void rcu_bp_after_fork_child(void)
 
        urcu_bp_prune_registry();
        oldmask = saved_fork_signal_mask;
+       mutex_unlock(&rcu_registry_lock);
        mutex_unlock(&rcu_gp_lock);
        ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
        assert(!ret);

diff --git a/urcu-qsbr.c b/urcu-qsbr.c
index a35dcfc5aea3674e9747ef69daea0f49def7a972..685efb5297dca06632697d810f5f078bc0aa835d 100644 (file)
--- a/urcu-qsbr.c
+++ b/urcu-qsbr.c
@@ -52,7 +52,21 @@
 
 void __attribute__((destructor)) rcu_exit(void);
 
+/*
+ * rcu_gp_lock ensures mutual exclusion between threads calling
+ * synchronize_rcu().
+ */
 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
+/*
+ * rcu_registry_lock ensures mutual exclusion between threads
+ * registering and unregistering themselves to/from the registry, and
+ * with threads reading that registry from synchronize_rcu(). However,
+ * this lock is not held all the way through the completion of awaiting
+ * for the grace period. It is sporadically released between iterations
+ * on the registry.
+ * rcu_registry_lock may nest inside rcu_gp_lock.
+ */
+static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
 struct rcu_gp rcu_gp = { .ctr = RCU_GP_ONLINE };
 
 /*
@@ -112,6 +126,10 @@ static void wait_gp(void)
                      NULL, NULL, 0);
 }
 
+/*
+ * Always called with rcu_registry lock held. Releases this lock between
+ * iterations and grabs it again. Holds the lock when it returns.
+ */
 static void wait_for_readers(struct cds_list_head *input_readers,
                        struct cds_list_head *cur_snap_readers,
                        struct cds_list_head *qsreaders)
@@ -171,6 +189,8 @@ static void wait_for_readers(struct cds_list_head *input_readers,
                        }
                        break;
                } else {
+                       /* Temporarily unlock the registry lock. */
+                       mutex_unlock(&rcu_registry_lock);
                        if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
                                wait_gp();
                        } else {
@@ -180,6 +200,8 @@ static void wait_for_readers(struct cds_list_head *input_readers,
                                cmm_smp_mb();
 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
                        }
+                       /* Re-lock the registry lock before the next loop. */
+                       mutex_lock(&rcu_registry_lock);
                }
        }
 }
@@ -233,11 +255,15 @@ void synchronize_rcu(void)
         */
        urcu_move_waiters(&waiters, &gp_waiters);
 
+       mutex_lock(&rcu_registry_lock);
+
        if (cds_list_empty(&registry))
                goto out;
 
        /*
         * Wait for readers to observe original parity or be quiescent.
+        * wait_for_readers() can release and grab again rcu_registry_lock
+        * interally.
         */
        wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 
@@ -279,6 +305,8 @@ void synchronize_rcu(void)
 
        /*
         * Wait for readers to observe new parity or be quiescent.
+        * wait_for_readers() can release and grab again rcu_registry_lock
+        * interally.
         */
        wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 
@@ -287,6 +315,7 @@ void synchronize_rcu(void)
         */
        cds_list_splice(&qsreaders, &registry);
 out:
+       mutex_unlock(&rcu_registry_lock);
        mutex_unlock(&rcu_gp_lock);
        urcu_wake_all_waiters(&waiters);
 gp_end:
@@ -339,6 +368,8 @@ void synchronize_rcu(void)
         */
        urcu_move_waiters(&waiters, &gp_waiters);
 
+       mutex_lock(&rcu_registry_lock);
+
        if (cds_list_empty(&registry))
                goto out;
 
@@ -363,6 +394,8 @@ void synchronize_rcu(void)
 
        /*
         * Wait for readers to observe new count of be quiescent.
+        * wait_for_readers() can release and grab again rcu_registry_lock
+        * interally.
         */
        wait_for_readers(&registry, NULL, &qsreaders);
 
@@ -371,6 +404,7 @@ void synchronize_rcu(void)
         */
        cds_list_splice(&qsreaders, &registry);
 out:
+       mutex_unlock(&rcu_registry_lock);
        mutex_unlock(&rcu_gp_lock);
        urcu_wake_all_waiters(&waiters);
 gp_end:
@@ -420,9 +454,9 @@ void rcu_register_thread(void)
        URCU_TLS(rcu_reader).tid = pthread_self();
        assert(URCU_TLS(rcu_reader).ctr == 0);
 
-       mutex_lock(&rcu_gp_lock);
+       mutex_lock(&rcu_registry_lock);
        cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
-       mutex_unlock(&rcu_gp_lock);
+       mutex_unlock(&rcu_registry_lock);
        _rcu_thread_online();
 }
 
@@ -433,9 +467,9 @@ void rcu_unregister_thread(void)
         * with a waiting writer.
         */
        _rcu_thread_offline();
-       mutex_lock(&rcu_gp_lock);
+       mutex_lock(&rcu_registry_lock);
        cds_list_del(&URCU_TLS(rcu_reader).node);
-       mutex_unlock(&rcu_gp_lock);
+       mutex_unlock(&rcu_registry_lock);
 }
 
 void rcu_exit(void)

diff --git a/urcu.c b/urcu.c
index ae3490f7cb2cdc1f7cdb1d26f77c0ee22620e9eb..1429f6dfbbd3f392b307a9139931bf24eeebd67a 100644 (file)
--- a/urcu.c
+++ b/urcu.c
@@ -100,7 +100,21 @@ void __attribute__((constructor)) rcu_init(void);
 void __attribute__((destructor)) rcu_exit(void);
 #endif
 
+/*
+ * rcu_gp_lock ensures mutual exclusion between threads calling
+ * synchronize_rcu().
+ */
 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
+/*
+ * rcu_registry_lock ensures mutual exclusion between threads
+ * registering and unregistering themselves to/from the registry, and
+ * with threads reading that registry from synchronize_rcu(). However,
+ * this lock is not held all the way through the completion of awaiting
+ * for the grace period. It is sporadically released between iterations
+ * on the registry.
+ * rcu_registry_lock may nest inside rcu_gp_lock.
+ */
+static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
 struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT };
 
 /*
@@ -226,6 +240,10 @@ static void wait_gp(void)
                      NULL, NULL, 0);
 }
 
+/*
+ * Always called with rcu_registry lock held. Releases this lock between
+ * iterations and grabs it again. Holds the lock when it returns.
+ */
 static void wait_for_readers(struct cds_list_head *input_readers,
                        struct cds_list_head *cur_snap_readers,
                        struct cds_list_head *qsreaders)
@@ -282,10 +300,14 @@ static void wait_for_readers(struct cds_list_head *input_readers,
                        }
                        break;
                } else {
+                       /* Temporarily unlock the registry lock. */
+                       mutex_unlock(&rcu_registry_lock);
                        if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
                                wait_gp();
                        else
                                caa_cpu_relax();
+                       /* Re-lock the registry lock before the next loop. */
+                       mutex_lock(&rcu_registry_lock);
                }
 #else /* #ifndef HAS_INCOHERENT_CACHES */
                /*
@@ -305,12 +327,16 @@ static void wait_for_readers(struct cds_list_head *input_readers,
                                smp_mb_master(RCU_MB_GROUP);
                                wait_gp_loops = 0;
                        }
+                       /* Temporarily unlock the registry lock. */
+                       mutex_unlock(&rcu_registry_lock);
                        if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
                                wait_gp();
                                wait_gp_loops++;
                        } else {
                                caa_cpu_relax();
                        }
+                       /* Re-lock the registry lock before the next loop. */
+                       mutex_lock(&rcu_registry_lock);
                }
 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
        }
@@ -348,17 +374,23 @@ void synchronize_rcu(void)
         */
        urcu_move_waiters(&waiters, &gp_waiters);
 
+       mutex_lock(&rcu_registry_lock);
+
        if (cds_list_empty(&registry))
                goto out;
 
-       /* All threads should read qparity before accessing data structure
-        * where new ptr points to. Must be done within rcu_gp_lock because it
-        * iterates on reader threads.*/
+       /*
+        * All threads should read qparity before accessing data structure
+        * where new ptr points to. Must be done within rcu_registry_lock
+        * because it iterates on reader threads.
+        */
        /* Write new ptr before changing the qparity */
        smp_mb_master(RCU_MB_GROUP);
 
        /*
         * Wait for readers to observe original parity or be quiescent.
+        * wait_for_readers() can release and grab again rcu_registry_lock
+        * interally.
         */
        wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 
@@ -399,6 +431,8 @@ void synchronize_rcu(void)
 
        /*
         * Wait for readers to observe new parity or be quiescent.
+        * wait_for_readers() can release and grab again rcu_registry_lock
+        * interally.
         */
        wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 
@@ -407,11 +441,14 @@ void synchronize_rcu(void)
         */
        cds_list_splice(&qsreaders, &registry);
 
-       /* Finish waiting for reader threads before letting the old ptr being
-        * freed. Must be done within rcu_gp_lock because it iterates on reader
-        * threads. */
+       /*
+        * Finish waiting for reader threads before letting the old ptr
+        * being freed. Must be done within rcu_registry_lock because it
+        * iterates on reader threads.
+        */
        smp_mb_master(RCU_MB_GROUP);
 out:
+       mutex_unlock(&rcu_registry_lock);
        mutex_unlock(&rcu_gp_lock);
 
        /*
@@ -447,17 +484,17 @@ void rcu_register_thread(void)
        assert(URCU_TLS(rcu_reader).need_mb == 0);
        assert(!(URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK));
 
-       mutex_lock(&rcu_gp_lock);
+       mutex_lock(&rcu_registry_lock);
        rcu_init();     /* In case gcc does not support constructor attribute */
        cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
-       mutex_unlock(&rcu_gp_lock);
+       mutex_unlock(&rcu_registry_lock);
 }
 
 void rcu_unregister_thread(void)
 {
-       mutex_lock(&rcu_gp_lock);
+       mutex_lock(&rcu_registry_lock);
        cds_list_del(&URCU_TLS(rcu_reader).node);
-       mutex_unlock(&rcu_gp_lock);
+       mutex_unlock(&rcu_registry_lock);
 }
 
 #ifdef RCU_MEMBARRIER
@@ -488,9 +525,9 @@ static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
  * rcu_init constructor. Called when the library is linked, but also when
  * reader threads are calling rcu_register_thread().
  * Should only be called by a single thread at a given time. This is ensured by
- * holing the rcu_gp_lock from rcu_register_thread() or by running at library
- * load time, which should not be executed by multiple threads nor concurrently
- * with rcu_register_thread() anyway.
+ * holing the rcu_registry_lock from rcu_register_thread() or by running
+ * at library load time, which should not be executed by multiple
+ * threads nor concurrently with rcu_register_thread() anyway.
  */
 void rcu_init(void)
 {