Fix rcu_assign_pointer() dynamic linking behavior
[urcu.git] / urcu.c
diff --git a/urcu.c b/urcu.c
index c07df2e8cf21384a07da3bb47066cd2b2bb44084..c13ad634e7cb7b98f24d8c9a5a756eb26487bf7b 100644 (file)
--- a/urcu.c
+++ b/urcu.c
 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
 #include "urcu.h"
 
-#ifndef DEBUG_FULL_MB
+#ifndef URCU_MB
+static int init_done;
+
 void __attribute__((constructor)) urcu_init(void);
 void __attribute__((destructor)) urcu_exit(void);
 #else
-static inline urcu_init(void)
+void urcu_init(void)
 {
 }
 #endif
 
-int init_done;
+static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
 
-pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
+int gp_futex;
 
 /*
  * Global grace period counter.
  * Contains the current RCU_GP_CTR_BIT.
- * Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path.
+ * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
  * Written to only by writer with mutex taken. Read by both writer and readers.
  */
 long urcu_gp_ctr = RCU_GP_COUNT;
@@ -61,27 +63,16 @@ long urcu_gp_ctr = RCU_GP_COUNT;
  * Written to only by each individual reader. Read by both the reader and the
  * writers.
  */
-long __thread urcu_active_readers;
-
-/* Thread IDs of registered readers */
-#define INIT_NUM_THREADS 4
-
-struct reader_registry {
-       pthread_t tid;
-       long *urcu_active_readers;
-       char *need_mb;
-};
+struct urcu_reader __thread urcu_reader;
 
 #ifdef DEBUG_YIELD
 unsigned int yield_active;
 unsigned int __thread rand_yield;
 #endif
 
-static struct reader_registry *registry;
-static char __thread need_mb;
-static int num_readers, alloc_readers;
+static LIST_HEAD(registry);
 
-void internal_urcu_lock(void)
+static void internal_urcu_lock(void)
 {
        int ret;
 
@@ -98,9 +89,9 @@ void internal_urcu_lock(void)
                        perror("Error in pthread mutex lock");
                        exit(-1);
                }
-               if (need_mb) {
+               if (urcu_reader.need_mb) {
                        smp_mb();
-                       need_mb = 0;
+                       urcu_reader.need_mb = 0;
                        smp_mb();
                }
                poll(NULL,0,10);
@@ -108,7 +99,7 @@ void internal_urcu_lock(void)
 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 }
 
-void internal_urcu_unlock(void)
+static void internal_urcu_unlock(void)
 {
        int ret;
 
@@ -127,30 +118,30 @@ static void switch_next_urcu_qparity(void)
        STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
 }
 
-#ifdef DEBUG_FULL_MB
-#ifdef HAS_INCOHERENT_CACHES
-static void force_mb_single_thread(struct reader_registry *index)
+#ifdef URCU_MB
+#if 0 /* unused */
+static void force_mb_single_thread(struct urcu_reader *index)
 {
        smp_mb();
 }
-#endif /* #ifdef HAS_INCOHERENT_CACHES */
+#endif //0
 
 static void force_mb_all_threads(void)
 {
        smp_mb();
 }
-#else /* #ifdef DEBUG_FULL_MB */
-#ifdef HAS_INCOHERENT_CACHES
-static void force_mb_single_thread(struct reader_registry *index)
+#else /* #ifdef URCU_MB */
+#if 0 /* unused */
+static void force_mb_single_thread(struct urcu_reader *index)
 {
-       assert(registry);
+       assert(!list_empty(&registry));
        /*
         * pthread_kill has a smp_mb(). But beware, we assume it performs
         * a cache flush on architectures with non-coherent cache. Let's play
         * safe and don't assume anything : we use smp_mc() to make sure the
         * cache flush is enforced.
         */
-       *index->need_mb = 1;
+       index->need_mb = 1;
        smp_mc();       /* write ->need_mb before sending the signals */
        pthread_kill(index->tid, SIGURCU);
        smp_mb();
@@ -158,21 +149,22 @@ static void force_mb_single_thread(struct reader_registry *index)
         * Wait for sighandler (and thus mb()) to execute on every thread.
         * BUSY-LOOP.
         */
-       while (*index->need_mb) {
+       while (index->need_mb) {
                poll(NULL, 0, 1);
        }
        smp_mb();       /* read ->need_mb before ending the barrier */
 }
-#endif /* #ifdef HAS_INCOHERENT_CACHES */
+#endif //0
 
 static void force_mb_all_threads(void)
 {
-       struct reader_registry *index;
+       struct urcu_reader *index;
+
        /*
         * Ask for each threads to execute a smp_mb() so we can consider the
         * compiler barriers around rcu read lock as real memory barriers.
         */
-       if (!registry)
+       if (list_empty(&registry))
                return;
        /*
         * pthread_kill has a smp_mb(). But beware, we assume it performs
@@ -180,8 +172,8 @@ static void force_mb_all_threads(void)
         * safe and don't assume anything : we use smp_mc() to make sure the
         * cache flush is enforced.
         */
-       for (index = registry; index < registry + num_readers; index++) {
-               *index->need_mb = 1;
+       list_for_each_entry(index, &registry, head) {
+               index->need_mb = 1;
                smp_mc();       /* write need_mb before sending the signal */
                pthread_kill(index->tid, SIGURCU);
        }
@@ -198,45 +190,95 @@ static void force_mb_all_threads(void)
         * relevant bug report.  For Linux kernels, we recommend getting
         * the Linux Test Project (LTP).
         */
-       for (index = registry; index < registry + num_readers; index++) {
-               while (*index->need_mb) {
+       list_for_each_entry(index, &registry, head) {
+               while (index->need_mb) {
                        pthread_kill(index->tid, SIGURCU);
                        poll(NULL, 0, 1);
                }
        }
        smp_mb();       /* read ->need_mb before ending the barrier */
 }
-#endif /* #else #ifdef DEBUG_FULL_MB */
+#endif /* #else #ifdef URCU_MB */
+
+/*
+ * synchronize_rcu() waiting. Single thread.
+ */
+static void wait_gp(void)
+{
+       /* Read reader_gp before read futex */
+       force_mb_all_threads();
+       if (uatomic_read(&gp_futex) == -1)
+               futex(&gp_futex, FUTEX_WAIT, -1,
+                     NULL, NULL, 0);
+}
 
 void wait_for_quiescent_state(void)
 {
-       struct reader_registry *index;
+       LIST_HEAD(qsreaders);
+       int wait_loops = 0;
+       struct urcu_reader *index, *tmp;
 
-       if (!registry)
+       if (list_empty(&registry))
                return;
        /*
-        * Wait for each thread urcu_active_readers count to become 0.
+        * Wait for each thread urcu_reader.ctr count to become 0.
         */
-       for (index = registry; index < registry + num_readers; index++) {
+       for (;;) {
+               wait_loops++;
+               if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
+                       uatomic_dec(&gp_futex);
+                       /* Write futex before read reader_gp */
+                       force_mb_all_threads();
+               }
+
+               list_for_each_entry_safe(index, tmp, &registry, head) {
+                       if (!rcu_old_gp_ongoing(&index->ctr))
+                               list_move(&index->head, &qsreaders);
+               }
+
 #ifndef HAS_INCOHERENT_CACHES
-               while (rcu_old_gp_ongoing(index->urcu_active_readers))
-                       cpu_relax();
+               if (list_empty(&registry)) {
+                       if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
+                               /* Read reader_gp before write futex */
+                               force_mb_all_threads();
+                               uatomic_set(&gp_futex, 0);
+                       }
+                       break;
+               } else {
+                       if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
+                               wait_gp();
+                       else
+                               cpu_relax();
+               }
 #else /* #ifndef HAS_INCOHERENT_CACHES */
-               int wait_loops = 0;
                /*
                 * BUSY-LOOP. Force the reader thread to commit its
-                * urcu_active_readers update to memory if we wait for too long.
+                * urcu_reader.ctr update to memory if we wait for too long.
                 */
-               while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
-                       if (wait_loops++ == KICK_READER_LOOPS) {
-                               force_mb_single_thread(index);
+               if (list_empty(&registry)) {
+                       if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
+                               /* Read reader_gp before write futex */
+                               force_mb_all_threads();
+                               uatomic_set(&gp_futex, 0);
+                       }
+                       break;
+               } else {
+                       switch (wait_loops) {
+                       case RCU_QS_ACTIVE_ATTEMPTS:
+                               wait_gp();
+                               break; /* only escape switch */
+                       case KICK_READER_LOOPS:
+                               force_mb_all_threads();
                                wait_loops = 0;
-                       } else {
+                               break; /* only escape switch */
+                       default:
                                cpu_relax();
                        }
                }
 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
        }
+       /* put back the reader list in the registry */
+       list_splice(&qsreaders, &registry);
 }
 
 void synchronize_rcu(void)
@@ -338,13 +380,19 @@ void *rcu_dereference(void *p)
 void *rcu_assign_pointer_sym(void **p, void *v)
 {
        wmb();
-       return STORE_SHARED(p, v);
+       return STORE_SHARED(*p, v);
 }
 
 void *rcu_xchg_pointer_sym(void **p, void *v)
 {
        wmb();
-       return xchg(p, v);
+       return uatomic_xchg(p, v);
+}
+
+void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new)
+{
+       wmb();
+       return uatomic_cmpxchg(p, old, _new);
 }
 
 void *rcu_publish_content_sym(void **p, void *v)
@@ -356,71 +404,26 @@ void *rcu_publish_content_sym(void **p, void *v)
        return oldptr;
 }
 
-static void rcu_add_reader(pthread_t id)
-{
-       struct reader_registry *oldarray;
-
-       if (!registry) {
-               alloc_readers = INIT_NUM_THREADS;
-               num_readers = 0;
-               registry =
-                       malloc(sizeof(struct reader_registry) * alloc_readers);
-       }
-       if (alloc_readers < num_readers + 1) {
-               oldarray = registry;
-               registry = malloc(sizeof(struct reader_registry)
-                               * (alloc_readers << 1));
-               memcpy(registry, oldarray,
-                       sizeof(struct reader_registry) * alloc_readers);
-               alloc_readers <<= 1;
-               free(oldarray);
-       }
-       registry[num_readers].tid = id;
-       /* reference to the TLS of _this_ reader thread. */
-       registry[num_readers].urcu_active_readers = &urcu_active_readers;
-       registry[num_readers].need_mb = &need_mb;
-       num_readers++;
-}
-
-/*
- * Never shrink (implementation limitation).
- * This is O(nb threads). Eventually use a hash table.
- */
-static void rcu_remove_reader(pthread_t id)
-{
-       struct reader_registry *index;
-
-       assert(registry != NULL);
-       for (index = registry; index < registry + num_readers; index++) {
-               if (pthread_equal(index->tid, id)) {
-                       memcpy(index, &registry[num_readers - 1],
-                               sizeof(struct reader_registry));
-                       registry[num_readers - 1].tid = 0;
-                       registry[num_readers - 1].urcu_active_readers = NULL;
-                       num_readers--;
-                       return;
-               }
-       }
-       /* Hrm not found, forgot to register ? */
-       assert(0);
-}
-
 void rcu_register_thread(void)
 {
+       urcu_reader.tid = pthread_self();
+       assert(urcu_reader.need_mb == 0);
+       assert(urcu_reader.ctr == 0);
+
        internal_urcu_lock();
        urcu_init();    /* In case gcc does not support constructor attribute */
-       rcu_add_reader(pthread_self());
+       list_add(&urcu_reader.head, &registry);
        internal_urcu_unlock();
 }
 
 void rcu_unregister_thread(void)
 {
        internal_urcu_lock();
-       rcu_remove_reader(pthread_self());
+       list_del(&urcu_reader.head);
        internal_urcu_unlock();
 }
 
-#ifndef DEBUG_FULL_MB
+#ifndef URCU_MB
 static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 {
        /*
@@ -429,7 +432,7 @@ static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
         * executed on.
         */
        smp_mb();
-       need_mb = 0;
+       urcu_reader.need_mb = 0;
        smp_mb();
 }
 
@@ -451,7 +454,7 @@ void urcu_init(void)
        init_done = 1;
 
        act.sa_sigaction = sigurcu_handler;
-       act.sa_flags = SA_SIGINFO;
+       act.sa_flags = SA_SIGINFO | SA_RESTART;
        sigemptyset(&act.sa_mask);
        ret = sigaction(SIGURCU, &act, NULL);
        if (ret) {
@@ -471,6 +474,6 @@ void urcu_exit(void)
                exit(-1);
        }
        assert(act.sa_sigaction == sigurcu_handler);
-       free(registry);
+       assert(list_empty(&registry));
 }
-#endif /* #ifndef DEBUG_FULL_MB */
+#endif /* #ifndef URCU_MB */
This page took 0.027472 seconds and 4 git commands to generate.