#include <urcu-call-rcu.h>
#include <urcu/arch.h>
#include <urcu/uatomic.h>
-#include <urcu/jhash.h>
#include <urcu/compiler.h>
#include <urcu/rculfhash.h>
#include <stdio.h>
} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
struct rcu_level {
- struct rcu_head head;
+ /* Note: manually update allocation length when adding a field */
struct _cds_lfht_node nodes[0];
};
};
struct partition_resize_work {
- struct rcu_head head;
+ pthread_t thread_id;
struct cds_lfht *ht;
unsigned long i, start, len;
void (*fct)(struct cds_lfht *ht, unsigned long i,
#endif
}
+/*
+ * Return the minimum order for which x <= (1UL << order).
+ * Return -1 if x is 0.
+ */
int get_count_order_u32(uint32_t x)
{
- int order;
+ if (!x)
+ return -1;
- order = fls_u32(x) - 1;
- if (x & (x - 1))
- order++;
- return order;
+ return fls_u32(x - 1);
}
+/*
+ * Return the minimum order for which x <= (1UL << order).
+ * Return -1 if x is 0.
+ */
int get_count_order_ulong(unsigned long x)
{
- int order;
+ if (!x)
+ return -1;
- order = fls_ulong(x) - 1;
- if (x & (x - 1))
- order++;
- return order;
+ return fls_ulong(x - 1);
}
#ifdef POISON_FREE
#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
-static const long nr_cpus_mask = -1;
+static const long nr_cpus_mask = -2;
static
struct ht_items_count *alloc_per_cpu_items_count(void)
}
static
-void cds_lfht_free_level(struct rcu_head *head)
+struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
+ unsigned long hash)
{
- struct rcu_level *l =
- caa_container_of(head, struct rcu_level, head);
- poison_free(l);
+ unsigned long index, order;
+
+ assert(size > 0);
+ index = hash & (size - 1);
+ /*
+ * equivalent to get_count_order_ulong(index + 1), but optimizes
+ * away the non-existing 0 special-case for
+ * get_count_order_ulong.
+ */
+ order = fls_ulong(index);
+
+ dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
+ hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
+
+ return &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
}
/*
return;
}
+static
+int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
+ struct cds_lfht_node *old_node,
+ struct cds_lfht_node *old_next,
+ struct cds_lfht_node *new_node)
+{
+ struct cds_lfht_node *dummy, *ret_next;
+ struct _cds_lfht_node *lookup;
+ int flagged = 0;
+
+ if (!old_node) /* Return -ENOENT if asked to replace NULL node */
+ goto end;
+
+ assert(!is_removed(old_node));
+ assert(!is_dummy(old_node));
+ assert(!is_removed(new_node));
+ assert(!is_dummy(new_node));
+ assert(new_node != old_node);
+ for (;;) {
+ /* Insert after node to be replaced */
+ if (is_removed(old_next)) {
+ /*
+ * Too late, the old node has been removed under us
+ * between lookup and replace. Fail.
+ */
+ goto end;
+ }
+ assert(!is_dummy(old_next));
+ assert(new_node != clear_flag(old_next));
+ new_node->p.next = clear_flag(old_next);
+ /*
+ * Here is the whole trick for lock-free replace: we add
+ * the replacement node _after_ the node we want to
+ * replace by atomically setting its next pointer at the
+ * same time we set its removal flag. Given that
+ * the lookups/get next use an iterator aware of the
+ * next pointer, they will either skip the old node due
+ * to the removal flag and see the new node, or use
+ * the old node, but will not see the new one.
+ */
+ ret_next = uatomic_cmpxchg(&old_node->p.next,
+ old_next, flag_removed(new_node));
+ if (ret_next == old_next)
+ break;
+ old_next = ret_next;
+ }
+
+ /* We performed the replacement. */
+ flagged = 1;
+
+ /*
+ * Ensure that the old node is not visible to readers anymore:
+ * lookup for the node, and remove it (along with any other
+ * logically removed node) if found.
+ */
+ lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash));
+ dummy = (struct cds_lfht_node *) lookup;
+ _cds_lfht_gc_bucket(dummy, new_node);
+end:
+ /*
+ * Only the flagging action indicated that we (and no other)
+ * replaced the node from the hash table.
+ */
+ if (flagged) {
+ assert(is_removed(rcu_dereference(old_node->p.next)));
+ return 0;
+ } else {
+ return -ENOENT;
+ }
+}
+
static
struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
unsigned long size,
enum add_mode mode, int dummy)
{
struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
- *dummy_node, *return_node;
+ *return_node;
struct _cds_lfht_node *lookup;
- unsigned long hash, index, order;
assert(!is_dummy(node));
assert(!is_removed(node));
node->p.next = flag_dummy(get_end());
return node; /* Initial first add (head) */
}
- hash = bit_reverse_ulong(node->p.reverse_hash);
+ lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
for (;;) {
uint32_t chain_len = 0;
* iter_prev points to the non-removed node prior to the
* insert location.
*/
- index = hash & (size - 1);
- order = get_count_order_ulong(index + 1);
- lookup = &ht->t.tbl[order]->nodes[index & ((!order ? 0 : (1UL << (order - 1))) - 1)];
iter_prev = (struct cds_lfht_node *) lookup;
/* We can always skip the dummy node initially */
iter = rcu_dereference(iter_prev->p.next);
goto insert;
if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
goto insert;
+ /* dummy node is the first node of the identical-hash-value chain */
+ if (dummy && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash)
+ goto insert;
next = rcu_dereference(clear_flag(iter)->p.next);
if (unlikely(is_removed(next)))
goto gc_node;
if ((mode == ADD_UNIQUE || mode == ADD_REPLACE)
&& !is_dummy(next)
+ && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash
&& !ht->compare_fct(node->key, node->key_len,
clear_flag(iter)->key,
clear_flag(iter)->key_len)) {
return_node = NULL;
else /* ADD_DEFAULT and ADD_UNIQUE */
return_node = node;
- goto gc_end;
+ goto end;
}
replace:
- /* Insert after node to be replaced */
- iter_prev = clear_flag(iter);
- iter = next;
- assert(node != clear_flag(iter));
- assert(!is_removed(iter_prev));
- assert(!is_removed(iter));
- assert(iter_prev != node);
- assert(!dummy);
- node->p.next = clear_flag(iter);
- if (is_dummy(iter))
- new_node = flag_dummy(node);
- else
- new_node = node;
- /*
- * Here is the whole trick for lock-free replace: we add
- * the replacement node _after_ the node we want to
- * replace by atomically setting its next pointer at the
- * same time we set its removal flag. Given that
- * the lookups/get next use an iterator aware of the
- * next pointer, they will either skip the old node due
- * to the removal flag and see the new node, or use
- * the old node, but will not see the new one.
- */
- new_node = flag_removed(new_node);
- if (uatomic_cmpxchg(&iter_prev->p.next,
- iter, new_node) != iter) {
- continue; /* retry */
+
+ if (!_cds_lfht_replace(ht, size, clear_flag(iter), next,
+ node)) {
+ return_node = clear_flag(iter);
+ goto end; /* gc already done */
} else {
- return_node = iter_prev;
- goto gc_end;
+ continue; /* retry */
}
gc_node:
(void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
/* retry */
}
-gc_end:
- /* Garbage collect logically removed nodes in the bucket */
- index = hash & (size - 1);
- order = get_count_order_ulong(index + 1);
- lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
- dummy_node = (struct cds_lfht_node *) lookup;
- _cds_lfht_gc_bucket(dummy_node, node);
+end:
return return_node;
}
struct cds_lfht_node *dummy, *next, *old;
struct _cds_lfht_node *lookup;
int flagged = 0;
- unsigned long hash, index, order;
+
+ if (!node) /* Return -ENOENT if asked to delete NULL node */
+ goto end;
/* logically delete the node */
assert(!is_dummy(node));
* the node, and remove it (along with any other logically removed node)
* if found.
*/
- hash = bit_reverse_ulong(node->p.reverse_hash);
- assert(size > 0);
- index = hash & (size - 1);
- order = get_count_order_ulong(index + 1);
- lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
+ lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
dummy = (struct cds_lfht_node *) lookup;
_cds_lfht_gc_bucket(dummy, node);
end:
if (flagged) {
assert(is_removed(rcu_dereference(node->p.next)));
return 0;
- } else
+ } else {
return -ENOENT;
+ }
}
static
struct partition_resize_work *work;
int thread, ret;
unsigned long nr_threads;
- pthread_t *thread_id;
/*
* Note: nr_cpus_mask + 1 is always power of 2.
* We spawn just the number of threads we need to satisfy the minimum
* partition size, up to the number of CPUs in the system.
*/
- nr_threads = min(nr_cpus_mask + 1,
- len >> MIN_PARTITION_PER_THREAD_ORDER);
+ if (nr_cpus_mask > 0) {
+ nr_threads = min(nr_cpus_mask + 1,
+ len >> MIN_PARTITION_PER_THREAD_ORDER);
+ } else {
+ nr_threads = 1;
+ }
partition_len = len >> get_count_order_ulong(nr_threads);
work = calloc(nr_threads, sizeof(*work));
- thread_id = calloc(nr_threads, sizeof(*thread_id));
assert(work);
for (thread = 0; thread < nr_threads; thread++) {
work[thread].ht = ht;
work[thread].len = partition_len;
work[thread].start = thread * partition_len;
work[thread].fct = fct;
- ret = pthread_create(&thread_id[thread], ht->resize_attr,
+ ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
partition_resize_thread, &work[thread]);
assert(!ret);
}
for (thread = 0; thread < nr_threads; thread++) {
- ret = pthread_join(thread_id[thread], NULL);
+ ret = pthread_join(work[thread].thread_id, NULL);
assert(!ret);
}
free(work);
- free(thread_id);
}
/*
bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
(void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
new_node, ADD_DEFAULT, 1);
- if (CMM_LOAD_SHARED(ht->in_progress_destroy))
- break;
}
ht->cds_lfht_rcu_read_unlock();
}
if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
break;
- ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level)
- + (len * sizeof(struct _cds_lfht_node)));
+ ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node));
assert(ht->t.tbl[i]);
/*
bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
(void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
fini_node, 1);
- if (CMM_LOAD_SHARED(ht->in_progress_destroy))
- break;
}
ht->cds_lfht_rcu_read_unlock();
}
unsigned long first_order, unsigned long len_order)
{
long i, end_order;
+ void *free_by_rcu = NULL;
dbg_printf("fini table: first_order %lu end_order %lu\n",
first_order, first_order + len_order);
* return a logically removed node as insert position.
*/
ht->cds_lfht_synchronize_rcu();
+ if (free_by_rcu)
+ free(free_by_rcu);
/*
* Set "removed" flag in dummy nodes about to be removed.
*/
remove_table(ht, i, len);
- ht->cds_lfht_call_rcu(&ht->t.tbl[i]->head, cds_lfht_free_level);
+ free_by_rcu = ht->t.tbl[i];
dbg_printf("fini new size: %lu\n", 1UL << i);
if (CMM_LOAD_SHARED(ht->in_progress_destroy))
break;
}
+
+ if (free_by_rcu) {
+ ht->cds_lfht_synchronize_rcu();
+ free(free_by_rcu);
+ }
}
struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
{
struct cds_lfht_node *node, *next, *dummy_node;
struct _cds_lfht_node *lookup;
- unsigned long hash, reverse_hash, index, order, size;
+ unsigned long hash, reverse_hash, size;
hash = ht->hash_fct(key, key_len, ht->hash_seed);
reverse_hash = bit_reverse_ulong(hash);
size = rcu_dereference(ht->t.size);
- index = hash & (size - 1);
- order = get_count_order_ulong(index + 1);
- lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)];
- dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
- hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
+ lookup = lookup_bucket(ht, size, hash);
dummy_node = (struct cds_lfht_node *) lookup;
/* We can always skip the dummy node initially */
node = rcu_dereference(dummy_node->p.next);
next = rcu_dereference(node->p.next);
if (likely(!is_removed(next))
&& !is_dummy(next)
+ && clear_flag(node)->p.reverse_hash == reverse_hash
&& likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
break;
}
iter->next = next;
}
-void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
+void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter)
{
struct cds_lfht_node *node, *next;
unsigned long reverse_hash;
iter->next = next;
}
+void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
+{
+ struct cds_lfht_node *node, *next;
+
+ node = clear_flag(iter->next);
+ for (;;) {
+ if (unlikely(is_end(node))) {
+ node = next = NULL;
+ break;
+ }
+ next = rcu_dereference(node->p.next);
+ if (likely(!is_removed(next))
+ && !is_dummy(next)) {
+ break;
+ }
+ node = clear_flag(next);
+ }
+ assert(!node || !is_dummy(rcu_dereference(node->p.next)));
+ iter->node = node;
+ iter->next = next;
+}
+
+void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
+{
+ struct _cds_lfht_node *lookup;
+
+ /*
+ * Get next after first dummy node. The first dummy node is the
+ * first node of the linked list.
+ */
+ lookup = &ht->t.tbl[0]->nodes[0];
+ iter->next = lookup->next;
+ cds_lfht_next(ht, iter);
+}
+
void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
{
unsigned long hash, size;
return ret;
}
-struct cds_lfht_node *cds_lfht_replace(struct cds_lfht *ht,
+struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
struct cds_lfht_node *node)
{
unsigned long hash, size;
return ret;
}
-int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
+int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
+ struct cds_lfht_node *new_node)
+{
+ unsigned long size;
+
+ size = rcu_dereference(ht->t.size);
+ return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
+ new_node);
+}
+
+int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
{
unsigned long size;
int ret;
size = rcu_dereference(ht->t.size);
- ret = _cds_lfht_del(ht, size, node, 0);
+ ret = _cds_lfht_del(ht, size, iter->node, 0);
if (!ret)
ht_count_del(ht, size);
return ret;
int ret;
/* Wait for in-flight resize operations to complete */
- CMM_STORE_SHARED(ht->in_progress_destroy, 1);
+ _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
+ cmm_smp_mb(); /* Store destroy before load resize */
while (uatomic_read(&ht->in_progress_resize))
poll(NULL, 0, 100); /* wait for 100ms */
ret = cds_lfht_delete_dummy(ht);
old_order = get_count_order_ulong(old_size) + 1;
new_order = get_count_order_ulong(new_size) + 1;
- printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
- old_size, old_order, new_size, new_order);
+ dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
+ old_size, old_order, new_size, new_order);
assert(new_size > old_size);
init_table(ht, old_order, new_order - old_order);
}
new_size = max(new_size, MIN_TABLE_SIZE);
old_order = get_count_order_ulong(old_size) + 1;
new_order = get_count_order_ulong(new_size) + 1;
- printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
- old_size, old_order, new_size, new_order);
+ dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
+ old_size, old_order, new_size, new_order);
assert(new_size < old_size);
/* Remove and unlink all dummy nodes to remove. */
* Resize table, re-do if the target size has changed under us.
*/
do {
+ assert(uatomic_read(&ht->in_progress_resize));
+ if (CMM_LOAD_SHARED(ht->in_progress_destroy))
+ break;
ht->t.resize_initiated = 1;
old_size = ht->t.size;
new_size = CMM_LOAD_SHARED(ht->t.resize_target);
cmm_smp_mb();
if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
uatomic_inc(&ht->in_progress_resize);
- cmm_smp_mb(); /* increment resize count before calling it */
+ cmm_smp_mb(); /* increment resize count before load destroy */
+ if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
+ uatomic_dec(&ht->in_progress_resize);
+ return;
+ }
work = malloc(sizeof(*work));
work->ht = ht;
ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
cmm_smp_mb();
if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
uatomic_inc(&ht->in_progress_resize);
- cmm_smp_mb(); /* increment resize count before calling it */
+ cmm_smp_mb(); /* increment resize count before load destroy */
+ if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
+ uatomic_dec(&ht->in_progress_resize);
+ return;
+ }
work = malloc(sizeof(*work));
work->ht = ht;
ht->cds_lfht_call_rcu(&work->head, do_resize_cb);