* IBM's contributions to this file may be relicensed under LGPLv2 or later.
*/
+#define URCU_NO_COMPAT_IDENTIFIERS
#define _LGPL_SOURCE
#include <stdio.h>
#include <pthread.h>
#include <errno.h>
#include <poll.h>
#include <unistd.h>
+#include <stdbool.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"
-#include "urcu-pointer.h"
-#include "urcu/tls-compat.h"
+#include <urcu/config.h>
+#include <urcu/arch.h>
+#include <urcu/wfcqueue.h>
+#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"
+#include "urcu-utils.h"
+#define URCU_API_MAP
/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
#undef _LGPL_SOURCE
-#include "urcu-bp.h"
+#include <urcu/urcu-bp.h>
#define _LGPL_SOURCE
#ifndef MAP_ANONYMOUS
* This is not generic.
*/
static
-void *mremap_wrapper(void *old_address, size_t old_size,
- size_t new_size, int flags)
+void *mremap_wrapper(void *old_address __attribute__((unused)),
+ size_t old_size __attribute__((unused)),
+ size_t new_size __attribute__((unused)),
+ int flags)
{
assert(!(flags & MREMAP_MAYMOVE));
/* Sleep delay in ms */
#define RCU_SLEEP_DELAY_MS 10
-#define INIT_NR_THREADS 8
-#define ARENA_INIT_ALLOC \
- sizeof(struct registry_chunk) \
- + INIT_NR_THREADS * sizeof(struct rcu_reader)
+#define INIT_READER_COUNT 8
/*
* Active attempts to check for reader Q.S. before calling sleep().
#define RCU_QS_ACTIVE_ATTEMPTS 100
static
-int rcu_bp_refcount;
+int urcu_bp_refcount;
/* If the headers do not support membarrier system call, fall back smp_mb. */
#ifdef __NR_membarrier
#endif
enum membarrier_cmd {
- MEMBARRIER_CMD_QUERY = 0,
- MEMBARRIER_CMD_SHARED = (1 << 0),
+ MEMBARRIER_CMD_QUERY = 0,
+ MEMBARRIER_CMD_SHARED = (1 << 0),
+ /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */
+ /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */
+ MEMBARRIER_CMD_PRIVATE_EXPEDITED = (1 << 3),
+ MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED = (1 << 4),
};
static
-void __attribute__((constructor)) rcu_bp_init(void);
+void __attribute__((constructor)) _urcu_bp_init(void);
static
-void __attribute__((destructor)) rcu_bp_exit(void);
+void urcu_bp_exit(void);
+static
+void __attribute__((destructor)) urcu_bp_exit_destructor(void);
+static void urcu_call_rcu_exit(void);
#ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
int urcu_bp_has_sys_membarrier;
static pthread_key_t urcu_bp_key;
-struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT };
+struct urcu_bp_gp urcu_bp_gp = { .ctr = URCU_BP_GP_COUNT };
/*
* Pointer to registry elements. Written to only by each individual reader. Read
* by both the reader and the writers.
*/
-DEFINE_URCU_TLS_IE(struct rcu_reader *, rcu_reader);
+DEFINE_URCU_TLS(struct urcu_bp_reader *, urcu_bp_reader);
static CDS_LIST_HEAD(registry);
struct registry_chunk {
- size_t data_len; /* data length */
- size_t used; /* amount of data used */
+ size_t capacity; /* capacity of this chunk (in elements) */
+ size_t used; /* count of elements used */
struct cds_list_head node; /* chunk_list node */
- char data[];
+ struct urcu_bp_reader readers[];
};
struct registry_arena {
static void smp_mb_master(void)
{
- if (caa_likely(urcu_bp_has_sys_membarrier))
- (void) membarrier(MEMBARRIER_CMD_SHARED, 0);
- else
+ if (caa_likely(urcu_bp_has_sys_membarrier)) {
+ if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0))
+ urcu_die(errno);
+ } else {
cmm_smp_mb();
+ }
+}
+
+/* Get the size of a chunk's allocation from its capacity (an element count). */
+static size_t chunk_allocation_size(size_t capacity)
+{
+ return (capacity * sizeof(struct urcu_bp_reader)) +
+ sizeof(struct registry_chunk);
}
/*
struct cds_list_head *qsreaders)
{
unsigned int wait_loops = 0;
- struct rcu_reader *index, *tmp;
+ struct urcu_bp_reader *index, *tmp;
/*
- * Wait for each thread URCU_TLS(rcu_reader).ctr to either
+ * Wait for each thread URCU_TLS(urcu_bp_reader).ctr to either
* indicate quiescence (not nested), or observe the current
* rcu_gp.ctr value.
*/
wait_loops++;
cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
- switch (rcu_reader_state(&index->ctr)) {
- case RCU_READER_ACTIVE_CURRENT:
+ switch (urcu_bp_reader_state(&index->ctr)) {
+ case URCU_BP_READER_ACTIVE_CURRENT:
if (cur_snap_readers) {
cds_list_move(&index->node,
cur_snap_readers);
break;
}
/* Fall-through */
- case RCU_READER_INACTIVE:
+ case URCU_BP_READER_INACTIVE:
cds_list_move(&index->node, qsreaders);
break;
- case RCU_READER_ACTIVE_OLD:
+ case URCU_BP_READER_ACTIVE_OLD:
/*
* Old snapshot. Leaving node in
* input_readers will make us busy-loop
}
}
-void synchronize_rcu(void)
+void urcu_bp_synchronize_rcu(void)
{
CDS_LIST_HEAD(cur_snap_readers);
CDS_LIST_HEAD(qsreaders);
cmm_smp_mb();
/* Switch parity: 0 -> 1, 1 -> 0 */
- CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR_PHASE);
+ CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ URCU_BP_GP_CTR_PHASE);
/*
* Must commit qparity update to memory before waiting for other parity
* library wrappers to be used by non-LGPL compatible source code.
*/
-void rcu_read_lock(void)
+void urcu_bp_read_lock(void)
{
- _rcu_read_lock();
+ _urcu_bp_read_lock();
}
-void rcu_read_unlock(void)
+void urcu_bp_read_unlock(void)
{
- _rcu_read_unlock();
+ _urcu_bp_read_unlock();
}
-int rcu_read_ongoing(void)
+int urcu_bp_read_ongoing(void)
{
- return _rcu_read_ongoing();
+ return _urcu_bp_read_ongoing();
}
/*
void expand_arena(struct registry_arena *arena)
{
struct registry_chunk *new_chunk, *last_chunk;
- size_t old_chunk_len, new_chunk_len;
+ size_t old_chunk_size_bytes, new_chunk_size_bytes, new_capacity;
/* 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_size_bytes = chunk_allocation_size(INIT_READER_COUNT);
new_chunk = (struct registry_chunk *) mmap(NULL,
- new_chunk_len,
+ new_chunk_size_bytes,
PROT_READ | PROT_WRITE,
MAP_ANONYMOUS | MAP_PRIVATE,
-1, 0);
if (new_chunk == MAP_FAILED)
abort();
- memset(new_chunk, 0, new_chunk_len);
- new_chunk->data_len =
- new_chunk_len - sizeof(struct registry_chunk);
+ memset(new_chunk, 0, new_chunk_size_bytes);
+ new_chunk->capacity = INIT_READER_COUNT;
cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
return; /* We're done. */
}
/* 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;
+ old_chunk_size_bytes = chunk_allocation_size(last_chunk->capacity);
+ new_capacity = last_chunk->capacity << 1;
+ new_chunk_size_bytes = chunk_allocation_size(new_capacity);
/* Don't allow memory mapping to move, just expand. */
- new_chunk = mremap_wrapper(last_chunk, old_chunk_len,
- new_chunk_len, 0);
+ new_chunk = mremap_wrapper(last_chunk, old_chunk_size_bytes,
+ new_chunk_size_bytes, 0);
if (new_chunk != MAP_FAILED) {
/* Should not have moved. */
assert(new_chunk == last_chunk);
- 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);
+ memset((char *) last_chunk + old_chunk_size_bytes, 0,
+ new_chunk_size_bytes - old_chunk_size_bytes);
+ last_chunk->capacity = new_capacity;
return; /* We're done. */
}
/* Remap did not succeed, we need to add a new chunk. */
new_chunk = (struct registry_chunk *) mmap(NULL,
- new_chunk_len,
+ new_chunk_size_bytes,
PROT_READ | PROT_WRITE,
MAP_ANONYMOUS | MAP_PRIVATE,
-1, 0);
if (new_chunk == MAP_FAILED)
abort();
- memset(new_chunk, 0, new_chunk_len);
- new_chunk->data_len =
- new_chunk_len - sizeof(struct registry_chunk);
+ memset(new_chunk, 0, new_chunk_size_bytes);
+ new_chunk->capacity = new_capacity;
cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
}
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);
retry:
cds_list_for_each_entry(chunk, &arena->chunk_list, node) {
- if (chunk->data_len - chunk->used < len)
+ size_t spot_idx;
+
+ /* Skip fully used chunks. */
+ if (chunk->used == chunk->capacity) {
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;
+ }
+
+ /* Find a spot. */
+ for (spot_idx = 0; spot_idx < chunk->capacity; spot_idx++) {
+ if (!chunk->readers[spot_idx].alloc) {
+ chunk->readers[spot_idx].alloc = 1;
+ chunk->used++;
+ return &chunk->readers[spot_idx];
}
}
}
* Reader threads are pointing to the reader registry. This is
* why its memory should never be relocated.
*/
- URCU_TLS(rcu_reader) = rcu_reader_reg;
+ URCU_TLS(urcu_bp_reader) = rcu_reader_reg;
}
/* Called with mutex locked */
cds_list_del(&rcu_reader_reg->node);
rcu_reader_reg->tid = 0;
rcu_reader_reg->alloc = 0;
- chunk->used -= sizeof(struct rcu_reader);
+ chunk->used--;
}
static
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])
+ if (rcu_reader_reg < (struct urcu_bp_reader *) &chunk->readers[0])
continue;
- if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len])
+ if (rcu_reader_reg >= (struct urcu_bp_reader *) &chunk->readers[chunk->capacity])
continue;
return chunk;
}
void remove_thread(struct rcu_reader *rcu_reader_reg)
{
cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
- URCU_TLS(rcu_reader) = NULL;
+ URCU_TLS(urcu_bp_reader) = NULL;
}
/* Disable signals, take mutex, add to registry */
-void rcu_bp_register(void)
+void urcu_bp_register(void)
{
sigset_t newmask, oldmask;
int ret;
* Check if a signal concurrently registered our thread since
* the check in rcu_read_lock().
*/
- if (URCU_TLS(rcu_reader))
+ if (URCU_TLS(urcu_bp_reader))
goto end;
/*
* Take care of early registration before urcu_bp constructor.
*/
- rcu_bp_init();
+ _urcu_bp_init();
mutex_lock(&rcu_registry_lock);
add_thread();
abort();
}
+void urcu_bp_register_thread(void)
+{
+ if (caa_unlikely(!URCU_TLS(urcu_bp_reader)))
+ urcu_bp_register(); /* If not yet registered. */
+}
+
/* Disable signals, take mutex, remove from registry */
static
-void rcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
+void urcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
{
sigset_t newmask, oldmask;
int ret;
ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
if (ret)
abort();
- rcu_bp_exit();
+ urcu_bp_exit();
}
/*
static
void urcu_bp_thread_exit_notifier(void *rcu_key)
{
- rcu_bp_unregister(rcu_key);
+ urcu_bp_unregister(rcu_key);
}
#ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
static
-void rcu_sys_membarrier_status(int available)
+void urcu_bp_sys_membarrier_status(bool available)
{
if (!available)
abort();
}
#else
static
-void rcu_sys_membarrier_status(int available)
+void urcu_bp_sys_membarrier_status(bool 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, unless its use has been
- * explicitly forced when building liburcu.
- */
+ if (!available)
+ return;
+ urcu_bp_has_sys_membarrier = 1;
}
#endif
static
-void rcu_bp_init(void)
+void urcu_bp_sys_membarrier_init(void)
+{
+ bool available = false;
+ int mask;
+
+ mask = membarrier(MEMBARRIER_CMD_QUERY, 0);
+ if (mask >= 0) {
+ if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) {
+ if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0))
+ urcu_die(errno);
+ available = true;
+ }
+ }
+ urcu_bp_sys_membarrier_status(available);
+}
+
+static
+void _urcu_bp_init(void)
{
mutex_lock(&init_lock);
- if (!rcu_bp_refcount++) {
+ if (!urcu_bp_refcount++) {
int ret;
ret = pthread_key_create(&urcu_bp_key,
urcu_bp_thread_exit_notifier);
if (ret)
abort();
- ret = membarrier(MEMBARRIER_CMD_QUERY, 0);
- rcu_sys_membarrier_status(ret >= 0
- && (ret & MEMBARRIER_CMD_SHARED));
+ urcu_bp_sys_membarrier_init();
initialized = 1;
}
mutex_unlock(&init_lock);
}
static
-void rcu_bp_exit(void)
+void urcu_bp_exit(void)
{
mutex_lock(&init_lock);
- if (!--rcu_bp_refcount) {
+ if (!--urcu_bp_refcount) {
struct registry_chunk *chunk, *tmp;
int ret;
cds_list_for_each_entry_safe(chunk, tmp,
®istry_arena.chunk_list, node) {
- munmap((void *) chunk, chunk->data_len
- + sizeof(struct registry_chunk));
+ munmap((void *) chunk, chunk_allocation_size(chunk->capacity));
}
CDS_INIT_LIST_HEAD(®istry_arena.chunk_list);
ret = pthread_key_delete(urcu_bp_key);
mutex_unlock(&init_lock);
}
+static
+void urcu_bp_exit_destructor(void)
+{
+ urcu_call_rcu_exit();
+ urcu_bp_exit();
+}
+
/*
* 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)
+void urcu_bp_before_fork(void)
{
sigset_t newmask, oldmask;
int ret;
saved_fork_signal_mask = oldmask;
}
-void rcu_bp_after_fork_parent(void)
+void urcu_bp_after_fork_parent(void)
{
sigset_t oldmask;
int ret;
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)
+ size_t spot_idx;
+
+ for (spot_idx = 0; spot_idx < chunk->capacity; spot_idx++) {
+ struct urcu_bp_reader *reader = &chunk->readers[spot_idx];
+
+ if (!reader->alloc)
continue;
- if (rcu_reader_reg->tid == pthread_self())
+ if (reader->tid == pthread_self())
continue;
- cleanup_thread(chunk, rcu_reader_reg);
+ cleanup_thread(chunk, reader);
}
}
}
-void rcu_bp_after_fork_child(void)
+void urcu_bp_after_fork_child(void)
{
sigset_t oldmask;
int ret;
assert(!ret);
}
-void *rcu_dereference_sym_bp(void *p)
+void *urcu_bp_dereference_sym(void *p)
{
return _rcu_dereference(p);
}
-void *rcu_set_pointer_sym_bp(void **p, void *v)
+void *urcu_bp_set_pointer_sym(void **p, void *v)
{
cmm_wmb();
uatomic_set(p, v);
return v;
}
-void *rcu_xchg_pointer_sym_bp(void **p, void *v)
+void *urcu_bp_xchg_pointer_sym(void **p, void *v)
{
cmm_wmb();
return uatomic_xchg(p, v);
}
-void *rcu_cmpxchg_pointer_sym_bp(void **p, void *old, void *_new)
+void *urcu_bp_cmpxchg_pointer_sym(void **p, void *old, void *_new)
{
cmm_wmb();
return uatomic_cmpxchg(p, old, _new);
projects / urcu.git / blobdiff