}
#endif
-/* Sleep delay in us */
-#define RCU_SLEEP_DELAY 1000
+/* Sleep delay in ms */
+#define RCU_SLEEP_DELAY_MS 10
#define INIT_NR_THREADS 8
#define ARENA_INIT_ALLOC \
sizeof(struct registry_chunk) \
*/
#define RCU_QS_ACTIVE_ATTEMPTS 100
-void __attribute__((destructor)) rcu_bp_exit(void);
+static
+int rcu_bp_refcount;
+
+static
+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;
+
+static pthread_key_t urcu_bp_key;
#ifdef DEBUG_YIELD
unsigned int yield_active;
-DEFINE_URCU_TLS(unsigned int, rand_yield);
+__DEFINE_URCU_TLS_GLOBAL(unsigned int, rand_yield);
#endif
/*
* Pointer to registry elements. Written to only by each individual reader. Read
* by both the reader and the writers.
*/
-DEFINE_URCU_TLS(struct rcu_reader *, rcu_reader);
+__DEFINE_URCU_TLS_GLOBAL(struct rcu_reader *, rcu_reader);
static CDS_LIST_HEAD(registry);
struct registry_chunk {
size_t data_len; /* data length */
- size_t used; /* data used */
+ size_t used; /* amount of data used */
struct cds_list_head node; /* chunk_list node */
char data[];
};
/* Saved fork signal mask, protected by rcu_gp_lock */
static sigset_t saved_fork_signal_mask;
-static void rcu_gc_registry(void);
-
static void mutex_lock(pthread_mutex_t *mutex)
{
int ret;
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.
+ */
void update_counter_and_wait(void)
{
CDS_LIST_HEAD(qsreaders);
- int wait_loops = 0;
+ unsigned int wait_loops = 0;
struct rcu_reader *index, *tmp;
/* Switch parity: 0 -> 1, 1 -> 0 */
* Wait for each thread rcu_reader.ctr count to become 0.
*/
for (;;) {
- wait_loops++;
+ if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
+ wait_loops++;
+
cds_list_for_each_entry_safe(index, tmp, ®istry, node) {
if (!rcu_old_gp_ongoing(&index->ctr))
cds_list_move(&index->node, &qsreaders);
if (cds_list_empty(®istry)) {
break;
} else {
- if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
- usleep(RCU_SLEEP_DELAY);
+ /* 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);
}
}
/* put back the reader list in the registry */
assert(!ret);
mutex_lock(&rcu_gp_lock);
+ mutex_lock(&rcu_registry_lock);
if (cds_list_empty(®istry))
goto out;
/* Write new ptr before changing the qparity */
cmm_smp_mb();
- /* Remove old registry elements */
- rcu_gc_registry();
-
/*
* Wait for previous parity to be empty of readers.
+ * update_counter_and_wait() can release and grab again
+ * rcu_registry_lock interally.
*/
update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */
/*
* Wait for previous parity to be empty of readers.
+ * update_counter_and_wait() can release and grab again
+ * rcu_registry_lock interally.
*/
update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */
*/
cmm_smp_mb();
out:
+ mutex_unlock(&rcu_registry_lock);
mutex_unlock(&rcu_gp_lock);
ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
assert(!ret);
void add_thread(void)
{
struct rcu_reader *rcu_reader_reg;
+ int ret;
rcu_reader_reg = arena_alloc(®istry_arena);
if (!rcu_reader_reg)
abort();
+ ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg);
+ if (ret)
+ abort();
/* Add to registry */
rcu_reader_reg->tid = pthread_self();
URCU_TLS(rcu_reader) = rcu_reader_reg;
}
-/* Called with signals off and mutex locked */
-static void rcu_gc_registry(void)
+/* Called with mutex locked */
+static
+void cleanup_thread(struct registry_chunk *chunk,
+ struct rcu_reader *rcu_reader_reg)
+{
+ rcu_reader_reg->ctr = 0;
+ cds_list_del(&rcu_reader_reg->node);
+ rcu_reader_reg->tid = 0;
+ rcu_reader_reg->alloc = 0;
+ chunk->used -= sizeof(struct rcu_reader);
+}
+
+static
+struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg)
{
struct registry_chunk *chunk;
- struct rcu_reader *rcu_reader_reg;
cds_list_for_each_entry(chunk, ®istry_arena.chunk_list, node) {
- for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0];
- rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len];
- rcu_reader_reg++) {
- pthread_t tid;
- int ret;
-
- if (!rcu_reader_reg->alloc)
- continue;
- tid = rcu_reader_reg->tid;
- ret = pthread_kill(tid, 0);
- assert(ret != EINVAL);
- if (ret == ESRCH) {
- cds_list_del(&rcu_reader_reg->node);
- rcu_reader_reg->ctr = 0;
- rcu_reader_reg->alloc = 0;
- chunk->used -= sizeof(struct rcu_reader);
- }
-
- }
+ if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0])
+ continue;
+ if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len])
+ continue;
+ return chunk;
}
+ return NULL;
+}
+
+/* Called with signals off and mutex locked */
+static
+void remove_thread(struct rcu_reader *rcu_reader_reg)
+{
+ cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
+ URCU_TLS(rcu_reader) = NULL;
}
/* Disable signals, take mutex, add to registry */
int ret;
ret = sigfillset(&newmask);
- assert(!ret);
+ if (ret)
+ abort();
ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
- assert(!ret);
+ if (ret)
+ abort();
/*
* Check if a signal concurrently registered our thread since
- * the check in rcu_read_lock(). */
+ * the check in rcu_read_lock().
+ */
if (URCU_TLS(rcu_reader))
goto end;
- mutex_lock(&rcu_gp_lock);
+ /*
+ * Take care of early registration before urcu_bp constructor.
+ */
+ rcu_bp_init();
+
+ 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);
- assert(!ret);
+ if (ret)
+ abort();
}
-void rcu_bp_exit(void)
+/* Disable signals, take mutex, remove from registry */
+static
+void rcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
{
- struct registry_chunk *chunk, *tmp;
+ sigset_t newmask, oldmask;
+ int ret;
- cds_list_for_each_entry_safe(chunk, tmp,
- ®istry_arena.chunk_list, node) {
- munmap(chunk, chunk->data_len + sizeof(struct registry_chunk));
+ ret = sigfillset(&newmask);
+ if (ret)
+ abort();
+ ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
+ if (ret)
+ abort();
+
+ mutex_lock(&rcu_registry_lock);
+ remove_thread(rcu_reader_reg);
+ mutex_unlock(&rcu_registry_lock);
+ ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
+ if (ret)
+ abort();
+ _rcu_bp_exit();
+}
+
+/*
+ * Remove thread from the registry when it exits, and flag it as
+ * destroyed so garbage collection can take care of it.
+ */
+static
+void urcu_bp_thread_exit_notifier(void *rcu_key)
+{
+ rcu_bp_unregister(rcu_key);
+}
+
+static
+void rcu_bp_init(void)
+{
+ mutex_lock(&init_lock);
+ if (!rcu_bp_refcount++) {
+ int ret;
+
+ ret = pthread_key_create(&urcu_bp_key,
+ urcu_bp_thread_exit_notifier);
+ if (ret)
+ abort();
+ initialized = 1;
}
+ mutex_unlock(&init_lock);
+}
+
+static
+void _rcu_bp_exit(void)
+{
+ mutex_lock(&init_lock);
+ if (!--rcu_bp_refcount) {
+ struct registry_chunk *chunk, *tmp;
+ int ret;
+
+ cds_list_for_each_entry_safe(chunk, tmp,
+ ®istry_arena.chunk_list, node) {
+ munmap(chunk, chunk->data_len
+ + sizeof(struct registry_chunk));
+ }
+ ret = pthread_key_delete(urcu_bp_key);
+ if (ret)
+ abort();
+ }
+ mutex_unlock(&init_lock);
+}
+
+/*
+ * Keep ABI compability within stable versions. This has never been
+ * exposed through a header, but needs to stay in the .so until the
+ * soname is bumped.
+ */
+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)
{
ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
assert(!ret);
mutex_lock(&rcu_gp_lock);
+ mutex_lock(&rcu_registry_lock);
saved_fork_signal_mask = oldmask;
}
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);
}
+/*
+ * Prune all entries from registry except our own thread. Fits the Linux
+ * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held.
+ */
+static
+void urcu_bp_prune_registry(void)
+{
+ struct registry_chunk *chunk;
+ struct rcu_reader *rcu_reader_reg;
+
+ cds_list_for_each_entry(chunk, ®istry_arena.chunk_list, node) {
+ for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0];
+ rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len];
+ rcu_reader_reg++) {
+ if (!rcu_reader_reg->alloc)
+ continue;
+ if (rcu_reader_reg->tid == pthread_self())
+ continue;
+ cleanup_thread(chunk, rcu_reader_reg);
+ }
+ }
+}
+
void rcu_bp_after_fork_child(void)
{
sigset_t oldmask;
int ret;
- rcu_gc_registry();
+ 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);
projects / userspace-rcu.git / blobdiff