Version 0.8.11

[userspace-rcu.git] / urcu-bp.c

diff --git a/urcu-bp.c b/urcu-bp.c
index 7a591a20da3ac8a58adc8c1ca1c0ffbf982350fe..ac8c6806e1008e5314f61e013b31e6b454e3fb69 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;
@@ -165,6 +179,10 @@ 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)
@@ -207,10 +225,14 @@ 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);
                }
        }
 }
@@ -229,6 +251,8 @@ void synchronize_rcu(void)
 
        mutex_lock(&rcu_gp_lock);
 
+       mutex_lock(&rcu_registry_lock);
+
        if (cds_list_empty(&registry))
                goto out;
 
@@ -239,6 +263,8 @@ 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);
 
@@ -268,6 +294,8 @@ 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);
 
@@ -282,6 +310,7 @@ 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);
@@ -326,7 +355,8 @@ void expand_arena(struct registry_arena *arena)
                        sizeof(struct registry_chunk)
                        + sizeof(struct rcu_reader));
                new_chunk_len = ARENA_INIT_ALLOC;
-               new_chunk = mmap(NULL, new_chunk_len,
+               new_chunk = (struct registry_chunk *) mmap(NULL,
+                       new_chunk_len,
                        PROT_READ | PROT_WRITE,
                        MAP_ANONYMOUS | MAP_PRIVATE,
                        -1, 0);
@@ -360,7 +390,8 @@ void expand_arena(struct registry_arena *arena)
        }
 
        /* Remap did not succeed, we need to add a new chunk. */
-       new_chunk = mmap(NULL, new_chunk_len,
+       new_chunk = (struct registry_chunk *) mmap(NULL,
+               new_chunk_len,
                PROT_READ | PROT_WRITE,
                MAP_ANONYMOUS | MAP_PRIVATE,
                -1, 0);
@@ -490,9 +521,9 @@ 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)
@@ -513,9 +544,9 @@ 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();
@@ -558,9 +589,10 @@ void _rcu_bp_exit(void)
 
                cds_list_for_each_entry_safe(chunk, tmp,
                                &registry_arena.chunk_list, node) {
-                       munmap(chunk, chunk->data_len
+                       munmap((void *) chunk, chunk->data_len
                                        + sizeof(struct registry_chunk));
                }
+               CDS_INIT_LIST_HEAD(&registry_arena.chunk_list);
                ret = pthread_key_delete(urcu_bp_key);
                if (ret)
                        abort();
@@ -578,9 +610,10 @@ 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)
 {
@@ -592,6 +625,7 @@ 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;
 }
 
@@ -601,6 +635,7 @@ 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);
@@ -608,7 +643,7 @@ 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)
@@ -636,6 +671,7 @@ 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);