Version 0.9.5

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

diff --git a/urcu-bp.c b/urcu-bp.c
index 4dc402865a026800ccef0bb35378be6241c81505..9c7da3d8d235355246de95dfb5de87c8dcc1fec8 100644 (file)
--- a/urcu-bp.c
+++ b/urcu-bp.c
@@ -36,6 +36,7 @@
 #include <unistd.h>
 #include <sys/mman.h>
 
+#include "urcu/arch.h"
 #include "urcu/wfcqueue.h"
 #include "urcu/map/urcu-bp.h"
 #include "urcu/static/urcu-bp.h"
@@ -94,11 +95,25 @@ void *mremap_wrapper(void *old_address, size_t old_size,
 static
 int rcu_bp_refcount;
 
+/* If the headers do not support membarrier system call, fall back smp_mb. */
+#ifdef __NR_membarrier
+# define membarrier(...)               syscall(__NR_membarrier, __VA_ARGS__)
+#else
+# define membarrier(...)               -ENOSYS
+#endif
+
+enum membarrier_cmd {
+       MEMBARRIER_CMD_QUERY = 0,
+       MEMBARRIER_CMD_SHARED = (1 << 0),
+};
+
 static
 void __attribute__((constructor)) rcu_bp_init(void);
 static
 void __attribute__((destructor)) rcu_bp_exit(void);
 
+int urcu_bp_has_sys_membarrier;
+
 /*
  * rcu_gp_lock ensures mutual exclusion between threads calling
  * synchronize_rcu().
@@ -174,6 +189,14 @@ static void mutex_unlock(pthread_mutex_t *mutex)
                urcu_die(ret);
 }
 
+static void smp_mb_master(void)
+{
+       if (caa_likely(urcu_bp_has_sys_membarrier))
+               (void) membarrier(MEMBARRIER_CMD_SHARED, 0);
+       else
+               cmm_smp_mb();
+}
+
 /*
  * Always called with rcu_registry lock held. Releases this lock between
  * iterations and grabs it again. Holds the lock when it returns.
@@ -254,7 +277,7 @@ void synchronize_rcu(void)
        /* All threads should read qparity before accessing data structure
         * where new ptr points to. */
        /* Write new ptr before changing the qparity */
-       cmm_smp_mb();
+       smp_mb_master();
 
        /*
         * Wait for readers to observe original parity or be quiescent.
@@ -303,7 +326,7 @@ void synchronize_rcu(void)
         * Finish waiting for reader threads before letting the old ptr being
         * freed.
         */
-       cmm_smp_mb();
+       smp_mb_master();
 out:
        mutex_unlock(&rcu_registry_lock);
        mutex_unlock(&rcu_gp_lock);
@@ -350,13 +373,14 @@ 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);
                if (new_chunk == MAP_FAILED)
                        abort();
-               bzero(new_chunk, new_chunk_len);
+               memset(new_chunk, 0, new_chunk_len);
                new_chunk->data_len =
                        new_chunk_len - sizeof(struct registry_chunk);
                cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
@@ -376,7 +400,7 @@ void expand_arena(struct registry_arena *arena)
        if (new_chunk != MAP_FAILED) {
                /* Should not have moved. */
                assert(new_chunk == last_chunk);
-               bzero((char *) last_chunk + old_chunk_len,
+               memset((char *) last_chunk + old_chunk_len, 0,
                        new_chunk_len - old_chunk_len);
                last_chunk->data_len =
                        new_chunk_len - sizeof(struct registry_chunk);
@@ -384,13 +408,14 @@ 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);
        if (new_chunk == MAP_FAILED)
                abort();
-       bzero(new_chunk, new_chunk_len);
+       memset(new_chunk, 0, new_chunk_len);
        new_chunk->data_len =
                new_chunk_len - sizeof(struct registry_chunk);
        cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
@@ -556,6 +581,17 @@ void urcu_bp_thread_exit_notifier(void *rcu_key)
        rcu_bp_unregister(rcu_key);
 }
 
+static
+void rcu_sys_membarrier_status(int available)
+{
+       /*
+        * membarrier has blocking behavior, which changes the
+        * application behavior too much compared to using barriers when
+        * synchronize_rcu is used repeatedly (without using call_rcu).
+        * Don't use membarrier for now.
+        */
+}
+
 static
 void rcu_bp_init(void)
 {
@@ -567,6 +603,9 @@ void rcu_bp_init(void)
                                urcu_bp_thread_exit_notifier);
                if (ret)
                        abort();
+               ret = membarrier(MEMBARRIER_CMD_QUERY, 0);
+               rcu_sys_membarrier_status(ret >= 0
+                               && (ret & MEMBARRIER_CMD_SHARED));
                initialized = 1;
        }
        mutex_unlock(&init_lock);
@@ -582,9 +621,10 @@ 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();