#include "urcu/map/urcu-bp.h"
#include "urcu/static/urcu-bp.h"
#include "urcu-pointer.h"
+#include "urcu/tls-compat.h"
+
+#include "urcu-die.h"
/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
#undef _LGPL_SOURCE
#define MAP_ANONYMOUS MAP_ANON
#endif
-#ifndef __linux__
+#ifdef __linux__
+static
+void *mremap_wrapper(void *old_address, size_t old_size,
+ size_t new_size, int flags)
+{
+ return mremap(old_address, old_size, new_size, flags);
+}
+#else
#define MREMAP_MAYMOVE 1
#define MREMAP_FIXED 2
/*
- * mremap wrapper for non-Linux systems. Maps a RW, anonymous private mapping.
+ * mremap wrapper for non-Linux systems not allowing MAYMOVE.
* This is not generic.
*/
-void *mremap(void *old_address, size_t old_size, size_t new_size, int flags)
+static
+void *mremap_wrapper(void *old_address, size_t old_size,
+ size_t new_size, int flags)
{
- void *new_address;
-
- assert(flags & MREMAP_MAYMOVE);
- assert(!(flags & MREMAP_FIXED));
- new_address = mmap(old_address, new_size,
- PROT_READ | PROT_WRITE,
- MAP_ANONYMOUS | MAP_PRIVATE,
- -1, 0);
- if (new_address == MAP_FAILED)
- return MAP_FAILED;
- if (old_address) {
- memcpy(new_address, old_address, old_size);
- munmap(old_address, old_size);
- }
- return new_address;
+ assert(!(flags & MREMAP_MAYMOVE));
+
+ return MAP_FAILED;
}
#endif
/* Sleep delay in us */
#define RCU_SLEEP_DELAY 1000
-#define ARENA_INIT_ALLOC 16
+#define INIT_NR_THREADS 8
+#define ARENA_INIT_ALLOC \
+ sizeof(struct registry_chunk) \
+ + INIT_NR_THREADS * sizeof(struct rcu_reader)
/*
* Active attempts to check for reader Q.S. before calling sleep().
*/
#define RCU_QS_ACTIVE_ATTEMPTS 100
+static
+void __attribute__((constructor)) rcu_bp_init(void);
+static
void __attribute__((destructor)) rcu_bp_exit(void);
static pthread_mutex_t rcu_gp_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;
-unsigned int __thread rand_yield;
+DEFINE_URCU_TLS(unsigned int, rand_yield);
#endif
/*
* Pointer to registry elements. Written to only by each individual reader. Read
* by both the reader and the writers.
*/
-struct rcu_reader __thread *rcu_reader;
+DEFINE_URCU_TLS(struct rcu_reader *, rcu_reader);
static CDS_LIST_HEAD(registry);
+struct registry_chunk {
+ size_t data_len; /* data length */
+ size_t used; /* amount of data used */
+ struct cds_list_head node; /* chunk_list node */
+ char data[];
+};
+
struct registry_arena {
- void *p;
- size_t len;
- size_t used;
+ struct cds_list_head chunk_list;
};
-static struct registry_arena registry_arena;
+static struct registry_arena registry_arena = {
+ .chunk_list = CDS_LIST_HEAD_INIT(registry_arena.chunk_list),
+};
/* 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;
#ifndef DISTRUST_SIGNALS_EXTREME
ret = pthread_mutex_lock(mutex);
- if (ret) {
- perror("Error in pthread mutex lock");
- exit(-1);
- }
+ if (ret)
+ urcu_die(ret);
#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
while ((ret = pthread_mutex_trylock(mutex)) != 0) {
- if (ret != EBUSY && ret != EINTR) {
- printf("ret = %d, errno = %d\n", ret, errno);
- perror("Error in pthread mutex lock");
- exit(-1);
- }
+ if (ret != EBUSY && ret != EINTR)
+ urcu_die(ret);
poll(NULL,0,10);
}
#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
int ret;
ret = pthread_mutex_unlock(mutex);
- if (ret) {
- perror("Error in pthread mutex unlock");
- exit(-1);
- }
+ if (ret)
+ urcu_die(ret);
}
void update_counter_and_wait(void)
sigset_t newmask, oldmask;
int ret;
- ret = sigemptyset(&newmask);
+ ret = sigfillset(&newmask);
assert(!ret);
- ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
+ ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
assert(!ret);
mutex_lock(&rcu_gp_lock);
/* 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.
*/
}
/*
- * only grow for now.
+ * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk.
+ * Else, try expanding the last chunk. If this fails, allocate a new
+ * chunk twice as big as the last chunk.
+ * Memory used by chunks _never_ moves. A chunk could theoretically be
+ * freed when all "used" slots are released, but we don't do it at this
+ * point.
*/
-static void resize_arena(struct registry_arena *arena, size_t len)
+static
+void expand_arena(struct registry_arena *arena)
{
- void *new_arena;
-
- if (!arena->p)
- new_arena = mmap(arena->p, len,
- PROT_READ | PROT_WRITE,
- MAP_ANONYMOUS | MAP_PRIVATE,
- -1, 0);
- else
- new_arena = mremap(arena->p, arena->len,
- len, MREMAP_MAYMOVE);
- assert(new_arena != MAP_FAILED);
+ struct registry_chunk *new_chunk, *last_chunk;
+ size_t old_chunk_len, new_chunk_len;
+
+ /* No chunk. */
+ if (cds_list_empty(&arena->chunk_list)) {
+ assert(ARENA_INIT_ALLOC >=
+ sizeof(struct registry_chunk)
+ + sizeof(struct rcu_reader));
+ new_chunk_len = ARENA_INIT_ALLOC;
+ new_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);
+ new_chunk->data_len =
+ new_chunk_len - sizeof(struct registry_chunk);
+ cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
+ return; /* We're done. */
+ }
- /*
- * re-used the same region ?
- */
- if (new_arena == arena->p)
- return;
+ /* Try expanding last chunk. */
+ last_chunk = cds_list_entry(arena->chunk_list.prev,
+ struct registry_chunk, node);
+ old_chunk_len =
+ last_chunk->data_len + sizeof(struct registry_chunk);
+ new_chunk_len = old_chunk_len << 1;
+
+ /* Don't allow memory mapping to move, just expand. */
+ new_chunk = mremap_wrapper(last_chunk, old_chunk_len,
+ new_chunk_len, 0);
+ if (new_chunk != MAP_FAILED) {
+ /* Should not have moved. */
+ assert(new_chunk == last_chunk);
+ bzero((char *) last_chunk + old_chunk_len,
+ new_chunk_len - old_chunk_len);
+ last_chunk->data_len =
+ new_chunk_len - sizeof(struct registry_chunk);
+ return; /* We're done. */
+ }
- bzero(new_arena + arena->len, len - arena->len);
- arena->p = new_arena;
+ /* Remap did not succeed, we need to add a new chunk. */
+ new_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);
+ new_chunk->data_len =
+ new_chunk_len - sizeof(struct registry_chunk);
+ cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
}
-/* Called with signals off and mutex locked */
-static void add_thread(void)
+static
+struct rcu_reader *arena_alloc(struct registry_arena *arena)
{
+ struct registry_chunk *chunk;
struct rcu_reader *rcu_reader_reg;
+ int expand_done = 0; /* Only allow to expand once per alloc */
+ size_t len = sizeof(struct rcu_reader);
- if (registry_arena.len
- < registry_arena.used + sizeof(struct rcu_reader))
- resize_arena(®istry_arena,
- caa_max(registry_arena.len << 1, ARENA_INIT_ALLOC));
- /*
- * Find a free spot.
- */
- for (rcu_reader_reg = registry_arena.p;
- (void *)rcu_reader_reg < registry_arena.p + registry_arena.len;
- rcu_reader_reg++) {
- if (!rcu_reader_reg->alloc)
- break;
+retry:
+ cds_list_for_each_entry(chunk, &arena->chunk_list, node) {
+ if (chunk->data_len - chunk->used < len)
+ continue;
+ /* Find spot */
+ 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) {
+ rcu_reader_reg->alloc = 1;
+ chunk->used += len;
+ return rcu_reader_reg;
+ }
+ }
+ }
+
+ if (!expand_done) {
+ expand_arena(arena);
+ expand_done = 1;
+ goto retry;
}
- rcu_reader_reg->alloc = 1;
- registry_arena.used += sizeof(struct rcu_reader);
+
+ return NULL;
+}
+
+/* Called with signals off and mutex locked */
+static
+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();
assert(rcu_reader_reg->ctr == 0);
cds_list_add(&rcu_reader_reg->node, ®istry);
- rcu_reader = rcu_reader_reg;
+ /*
+ * Reader threads are pointing to the reader registry. This is
+ * why its memory should never be relocated.
+ */
+ 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)
{
- struct rcu_reader *rcu_reader_reg;
- pthread_t tid;
- int ret;
+ 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);
+}
- for (rcu_reader_reg = registry_arena.p;
- (void *)rcu_reader_reg < registry_arena.p + registry_arena.len;
- rcu_reader_reg++) {
- if (!rcu_reader_reg->alloc)
+static
+struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg)
+{
+ struct registry_chunk *chunk;
+
+ cds_list_for_each_entry(chunk, ®istry_arena.chunk_list, node) {
+ if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0])
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;
- registry_arena.used -= sizeof(struct rcu_reader);
- }
+ 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(void)
+{
+ struct rcu_reader *rcu_reader_reg;
+
+ rcu_reader_reg = URCU_TLS(rcu_reader);
+ cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
+ URCU_TLS(rcu_reader) = NULL;
}
/* Disable signals, take mutex, add to registry */
sigset_t newmask, oldmask;
int ret;
- ret = sigemptyset(&newmask);
- assert(!ret);
- ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
- assert(!ret);
+ ret = sigfillset(&newmask);
+ if (ret)
+ abort();
+ ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
+ if (ret)
+ abort();
/*
* Check if a signal concurrently registered our thread since
- * the check in rcu_read_lock(). */
- if (rcu_reader)
+ * the check in rcu_read_lock().
+ */
+ if (URCU_TLS(rcu_reader))
goto end;
+ /*
+ * Take care of early registration before urcu_bp constructor.
+ */
+ rcu_bp_init();
+
mutex_lock(&rcu_gp_lock);
add_thread();
mutex_unlock(&rcu_gp_lock);
end:
ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
- assert(!ret);
+ if (ret)
+ abort();
+}
+
+/* Disable signals, take mutex, remove from registry */
+static
+void rcu_bp_unregister(void)
+{
+ sigset_t newmask, oldmask;
+ int ret;
+
+ ret = sigfillset(&newmask);
+ if (ret)
+ abort();
+ ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
+ if (ret)
+ abort();
+
+ mutex_lock(&rcu_gp_lock);
+ remove_thread();
+ mutex_unlock(&rcu_gp_lock);
+ ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
+ if (ret)
+ abort();
+}
+
+/*
+ * 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)
+{
+ assert(rcu_key == URCU_TLS(rcu_reader));
+ rcu_bp_unregister();
+}
+
+static
+void rcu_bp_init(void)
+{
+ mutex_lock(&init_lock);
+ if (!initialized) {
+ 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)
{
- if (registry_arena.p)
- munmap(registry_arena.p, registry_arena.len);
+ 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();
}
/*
sigset_t newmask, oldmask;
int ret;
- ret = sigemptyset(&newmask);
+ ret = sigfillset(&newmask);
assert(!ret);
- ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
+ ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
assert(!ret);
mutex_lock(&rcu_gp_lock);
saved_fork_signal_mask = oldmask;
assert(!ret);
}
+/*
+ * Prune all entries from registry except our own thread. Fits the Linux
+ * fork behavior. Called with rcu_gp_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_gp_lock);
ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
assert(!ret);
}
+void *rcu_dereference_sym_bp(void *p)
+{
+ return _rcu_dereference(p);
+}
+
void *rcu_set_pointer_sym_bp(void **p, void *v)
{
cmm_wmb();
- return uatomic_set(p, v);
+ uatomic_set(p, v);
+ return v;
}
void *rcu_xchg_pointer_sym_bp(void **p, void *v)
return uatomic_cmpxchg(p, old, _new);
}
+DEFINE_RCU_FLAVOR(rcu_flavor);
+
#include "urcu-call-rcu-impl.h"
#include "urcu-defer-impl.h"