/*
* Define the minimum table size.
*/
-#define MIN_TABLE_SIZE 128
+#define MIN_TABLE_SIZE 1
#if (CAA_BITS_PER_LONG == 32)
#define MAX_TABLE_ORDER 32
#define MAX_TABLE_ORDER 64
#endif
+/*
+ * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
+ */
+#define MIN_PARTITION_PER_THREAD 4096
+
#ifndef min
#define min(a, b) ((a) < (b) ? (a) : (b))
#endif
#define DUMMY_FLAG (1UL << 1)
#define FLAGS_MASK ((1UL << 2) - 1)
+/* Value of the end pointer. Should not interact with flags. */
+#define END_VALUE NULL
+
struct ht_items_count {
unsigned long add, remove;
} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
void (*cds_lfht_rcu_read_unlock)(void);
void (*cds_lfht_rcu_thread_offline)(void);
void (*cds_lfht_rcu_thread_online)(void);
+ void (*cds_lfht_rcu_register_thread)(void);
+ void (*cds_lfht_rcu_unregister_thread)(void);
+ pthread_attr_t *resize_attr; /* Resize threads attributes */
unsigned long count; /* global approximate item count */
struct ht_items_count *percpu_count; /* per-cpu item count */
};
struct cds_lfht *ht;
};
+struct partition_resize_work {
+ struct rcu_head head;
+ struct cds_lfht *ht;
+ unsigned long i, start, len;
+ void (*fct)(struct cds_lfht *ht, unsigned long i,
+ unsigned long start, unsigned long len);
+};
+
static
struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
unsigned long size,
{
return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
}
-
+
+static
+struct cds_lfht_node *get_end(void)
+{
+ return (struct cds_lfht_node *) END_VALUE;
+}
+
+static
+int is_end(struct cds_lfht_node *node)
+{
+ return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
+}
+
static
unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
{
* Remove all logically deleted nodes from a bucket up to a certain node key.
*/
static
-int _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
+void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
{
struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
iter_prev = dummy;
/* We can always skip the dummy node initially */
iter = rcu_dereference(iter_prev->p.next);
- if (unlikely(iter == NULL)) {
- /*
- * We are executing concurrently with a hash table
- * expand, so we see a dummy node with NULL next value.
- * Help expand by linking this node into the list and
- * retry.
- */
- return 1;
- }
assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
/*
* We should never be called with dummy (start of chain)
*/
assert(dummy != node);
for (;;) {
- if (unlikely(!clear_flag(iter)))
- return 0;
+ if (unlikely(is_end(iter)))
+ return;
if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
- return 0;
+ return;
next = rcu_dereference(clear_flag(iter)->p.next);
if (likely(is_removed(next)))
break;
new_next = clear_flag(next);
(void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
}
- return 0;
+ return;
}
static
assert(!is_removed(node));
if (!size) {
assert(dummy);
- node->p.next = flag_dummy(NULL);
+ node->p.next = flag_dummy(get_end());
return node; /* Initial first add (head) */
}
hash = bit_reverse_ulong(node->p.reverse_hash);
iter_prev = (struct cds_lfht_node *) lookup;
/* We can always skip the dummy node initially */
iter = rcu_dereference(iter_prev->p.next);
- if (unlikely(iter == NULL)) {
- /*
- * We are executing concurrently with a hash table
- * expand, so we see a dummy node with NULL next value.
- * Help expand by linking this node into the list and
- * retry.
- */
- (void) _cds_lfht_add(ht, size >> 1, iter_prev, 0, 1);
- continue; /* retry */
- }
assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
for (;;) {
- /*
- * When adding a dummy node, we allow concurrent
- * add/removal to help. If we find the dummy node in
- * place, skip its insertion.
- */
- if (unlikely(dummy && clear_flag(iter) == node))
- return node;
- if (unlikely(!clear_flag(iter)))
+ if (unlikely(is_end(iter)))
goto insert;
if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
goto insert;
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;
- if (_cds_lfht_gc_bucket(dummy_node, node)) {
- /* Help expand */
- (void) _cds_lfht_add(ht, size >> 1, dummy_node, 0, 1);
- goto gc_end; /* retry */
- }
+ _cds_lfht_gc_bucket(dummy_node, node);
return node;
}
* the node, and remove it (along with any other logically removed node)
* if found.
*/
-gc_retry:
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))];
dummy = (struct cds_lfht_node *) lookup;
- if (_cds_lfht_gc_bucket(dummy, node)) {
- /* Help expand */
- (void) _cds_lfht_add(ht, size >> 1, dummy, 0, 1);
- goto gc_retry; /* retry */
- }
+ _cds_lfht_gc_bucket(dummy, node);
end:
/*
* Only the flagging action indicated that we (and no other)
}
static
-void init_table_hash(struct cds_lfht *ht, unsigned long i,
- unsigned long len)
+void *partition_resize_thread(void *arg)
{
- unsigned long j;
+ struct partition_resize_work *work = arg;
- for (j = 0; j < len; j++) {
- struct cds_lfht_node *new_node =
- (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
+ work->ht->cds_lfht_rcu_register_thread();
+ work->fct(work->ht, work->i, work->start, work->len);
+ work->ht->cds_lfht_rcu_unregister_thread();
+ return NULL;
+}
- dbg_printf("init hash entry: i %lu j %lu hash %lu\n",
- i, j, !i ? 0 : (1UL << (i - 1)) + j);
- new_node->p.reverse_hash =
- bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
- if (CMM_LOAD_SHARED(ht->in_progress_destroy))
- break;
+static
+void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
+ unsigned long len,
+ void (*fct)(struct cds_lfht *ht, unsigned long i,
+ unsigned long start, unsigned long len))
+{
+ unsigned long partition_len;
+ struct partition_resize_work *work;
+ int cpu, ret;
+ pthread_t *thread_id;
+
+ /* Note: nr_cpus_mask + 1 is always power of 2 */
+ partition_len = len >> get_count_order_ulong(nr_cpus_mask + 1);
+ work = calloc(nr_cpus_mask + 1, sizeof(*work));
+ thread_id = calloc(nr_cpus_mask + 1, sizeof(*thread_id));
+ assert(work);
+ for (cpu = 0; cpu < nr_cpus_mask + 1; cpu++) {
+ work[cpu].ht = ht;
+ work[cpu].i = i;
+ work[cpu].len = partition_len;
+ work[cpu].start = cpu * partition_len;
+ work[cpu].fct = fct;
+ ret = pthread_create(&thread_id[cpu], ht->resize_attr,
+ partition_resize_thread, &work[cpu]);
+ assert(!ret);
}
+ for (cpu = 0; cpu < nr_cpus_mask + 1; cpu++) {
+ ret = pthread_join(thread_id[cpu], NULL);
+ assert(!ret);
+ }
+ free(work);
+ free(thread_id);
}
/*
* Holding RCU read lock to protect _cds_lfht_add against memory
* reclaim that could be performed by other call_rcu worker threads (ABA
* problem).
+ *
+ * When we reach a certain length, we can split this population phase over
+ * many worker threads, based on the number of CPUs available in the system.
+ * This should therefore take care of not having the expand lagging behind too
+ * many concurrent insertion threads by using the scheduler's ability to
+ * schedule dummy node population fairly with insertions.
*/
static
-void init_table_link(struct cds_lfht *ht, unsigned long i, unsigned long len)
+void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
+ unsigned long start, unsigned long len)
{
unsigned long j;
- ht->cds_lfht_rcu_thread_online();
ht->cds_lfht_rcu_read_lock();
- for (j = 0; j < len; j++) {
+ for (j = start; j < start + len; j++) {
struct cds_lfht_node *new_node =
(struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
- dbg_printf("init link: i %lu j %lu hash %lu\n",
+ dbg_printf("init populate: i %lu j %lu hash %lu\n",
i, j, !i ? 0 : (1UL << (i - 1)) + j);
+ new_node->p.reverse_hash =
+ bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
(void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
new_node, 0, 1);
if (CMM_LOAD_SHARED(ht->in_progress_destroy))
break;
}
ht->cds_lfht_rcu_read_unlock();
- ht->cds_lfht_rcu_thread_offline();
+}
+
+static
+void init_table_populate(struct cds_lfht *ht, unsigned long i,
+ unsigned long len)
+{
+ assert(nr_cpus_mask != -1);
+ if (nr_cpus_mask < 0 || len < (nr_cpus_mask + 1) * MIN_PARTITION_PER_THREAD) {
+ ht->cds_lfht_rcu_thread_online();
+ init_table_populate_partition(ht, i, 0, len);
+ ht->cds_lfht_rcu_thread_offline();
+ return;
+ }
+ partition_resize_helper(ht, i, len, init_table_populate_partition);
}
static
ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level)
+ (len * sizeof(struct _cds_lfht_node)));
-
- /* Set all dummy nodes reverse hash values for a level */
- init_table_hash(ht, i, len);
+ assert(ht->t.tbl[i]);
/*
- * Update table size. At this point, concurrent add/remove see
- * dummy nodes with correctly initialized reverse hash value,
- * but with NULL next pointers. If they do, they can help us
- * link the dummy nodes into the list and retry.
+ * Set all dummy nodes reverse hash values for a level and
+ * link all dummy nodes into the table.
*/
- cmm_smp_wmb(); /* populate data before RCU size */
- CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
+ init_table_populate(ht, i, len);
/*
- * Link all dummy nodes into the table. Concurrent
- * add/remove are helping us.
+ * Update table size.
*/
- init_table_link(ht, i, len);
+ cmm_smp_wmb(); /* populate data before RCU size */
+ CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
if (CMM_LOAD_SHARED(ht->in_progress_destroy))
*
* Logical removal and garbage collection can therefore be done in batch or on a
* node-per-node basis, as long as the guarantee above holds.
+ *
+ * When we reach a certain length, we can split this removal over many worker
+ * threads, based on the number of CPUs available in the system. This should
+ * take care of not letting resize process lag behind too many concurrent
+ * updater threads actively inserting into the hash table.
*/
static
-void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
+void remove_table_partition(struct cds_lfht *ht, unsigned long i,
+ unsigned long start, unsigned long len)
{
unsigned long j;
- ht->cds_lfht_rcu_thread_online();
ht->cds_lfht_rcu_read_lock();
- for (j = 0; j < len; j++) {
+ for (j = start; j < start + len; j++) {
struct cds_lfht_node *fini_node =
(struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
break;
}
ht->cds_lfht_rcu_read_unlock();
- ht->cds_lfht_rcu_thread_offline();
+}
+
+static
+void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
+{
+
+ assert(nr_cpus_mask != -1);
+ if (nr_cpus_mask < 0 || len < (nr_cpus_mask + 1) * MIN_PARTITION_PER_THREAD) {
+ ht->cds_lfht_rcu_thread_online();
+ remove_table_partition(ht, i, 0, len);
+ ht->cds_lfht_rcu_thread_offline();
+ return;
+ }
+ partition_resize_helper(ht, i, len, remove_table_partition);
}
static
}
}
-struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct,
+struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
cds_lfht_compare_fct compare_fct,
unsigned long hash_seed,
unsigned long init_size,
void (*cds_lfht_rcu_read_lock)(void),
void (*cds_lfht_rcu_read_unlock)(void),
void (*cds_lfht_rcu_thread_offline)(void),
- void (*cds_lfht_rcu_thread_online)(void))
+ void (*cds_lfht_rcu_thread_online)(void),
+ void (*cds_lfht_rcu_register_thread)(void),
+ void (*cds_lfht_rcu_unregister_thread)(void),
+ pthread_attr_t *attr)
{
struct cds_lfht *ht;
unsigned long order;
if (init_size && (init_size & (init_size - 1)))
return NULL;
ht = calloc(1, sizeof(struct cds_lfht));
+ assert(ht);
ht->hash_fct = hash_fct;
ht->compare_fct = compare_fct;
ht->hash_seed = hash_seed;
ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
+ ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
+ ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
+ ht->resize_attr = attr;
ht->percpu_count = alloc_per_cpu_items_count();
/* this mutex should not nest in read-side C.S. */
pthread_mutex_init(&ht->resize_mutex, NULL);
struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len)
{
- struct cds_lfht_node *node, *next;
+ struct cds_lfht_node *node, *next, *dummy_node;
struct _cds_lfht_node *lookup;
unsigned long hash, reverse_hash, index, order, size;
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)));
- node = (struct cds_lfht_node *) lookup;
+ dummy_node = (struct cds_lfht_node *) lookup;
+ /* We can always skip the dummy node initially */
+ node = rcu_dereference(dummy_node->p.next);
+ node = clear_flag(node);
for (;;) {
- if (unlikely(!node))
+ if (unlikely(is_end(node))) {
+ node = NULL;
break;
+ }
if (unlikely(node->p.reverse_hash > reverse_hash)) {
node = NULL;
break;
node = clear_flag(next);
for (;;) {
- if (unlikely(!node))
+ if (unlikely(is_end(node))) {
+ node = NULL;
break;
+ }
if (unlikely(node->p.reverse_hash > reverse_hash)) {
node = NULL;
break;
size = rcu_dereference(ht->t.size);
ret = _cds_lfht_add(ht, size, node, 1, 0);
- if (ret != node)
+ if (ret == node)
ht_count_add(ht, size);
return ret;
}
if (!is_dummy(node))
return -EPERM;
assert(!is_removed(node));
- } while (clear_flag(node));
+ } while (!is_end(node));
/*
* size accessed without rcu_dereference because hash table is
* being destroyed.
* Should only be called when no more concurrent readers nor writers can
* possibly access the table.
*/
-int cds_lfht_destroy(struct cds_lfht *ht)
+int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
{
int ret;
if (ret)
return ret;
free_per_cpu_items_count(ht->percpu_count);
+ if (attr)
+ *attr = ht->resize_attr;
poison_free(ht);
return ret;
}
else
(nr_dummy)++;
node = clear_flag(next);
- } while (node);
+ } while (!is_end(node));
dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
}
projects / urcu.git / blobdiff