/*
* The removed flag needs to be updated atomically with the pointer.
* It indicates that no node must attach to the node scheduled for
- * removal. The gc flag also needs to be updated atomically with the
- * pointer. It indicates that node garbage collection must be performed.
- * "removed" and "gc" flags are separate for the benefit of replacement
- * operation.
+ * removal, and that node garbage collection must be performed.
* The dummy flag does not require to be updated atomically with the
* pointer, but it is added as a pointer low bit flag to save space.
*/
#define REMOVED_FLAG (1UL << 0)
-#define GC_FLAG (1UL << 1)
-#define DUMMY_FLAG (1UL << 2)
-#define FLAGS_MASK ((1UL << 3) - 1)
+#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 cds_lfht_node *node,
enum add_mode mode, int dummy);
-static
-int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
- struct cds_lfht_node *node,
- int dummy_removal, int do_gc);
-
/*
* Algorithm to reverse bits in a word by lookup table, extended to
* 64-bit words.
return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
}
-static
-int is_gc(struct cds_lfht_node *node)
-{
- return ((unsigned long) node) & GC_FLAG;
-}
-
-static
-struct cds_lfht_node *flag_gc(struct cds_lfht_node *node)
-{
- return (struct cds_lfht_node *) (((unsigned long) node) | GC_FLAG);
-}
-
static
int is_dummy(struct cds_lfht_node *node)
{
struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
assert(!is_dummy(dummy));
- assert(!is_gc(dummy));
assert(!is_removed(dummy));
assert(!is_dummy(node));
- assert(!is_gc(node));
assert(!is_removed(node));
for (;;) {
iter_prev = dummy;
if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
return;
next = rcu_dereference(clear_flag(iter)->p.next);
- if (likely(is_gc(next)))
+ if (likely(is_removed(next)))
break;
iter_prev = clear_flag(iter);
iter = next;
}
- assert(!is_gc(iter));
+ assert(!is_removed(iter));
if (is_dummy(iter))
new_next = flag_dummy(clear_flag(next));
else
enum add_mode mode, int dummy)
{
struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
- *dummy_node, *return_node, *replace_pinned = NULL;
+ *dummy_node, *return_node;
struct _cds_lfht_node *lookup;
unsigned long hash, index, order;
assert(!is_dummy(node));
- assert(!is_gc(node));
assert(!is_removed(node));
if (!size) {
assert(dummy);
}
hash = bit_reverse_ulong(node->p.reverse_hash);
for (;;) {
- uint32_t chain_len;
+ uint32_t chain_len = 0;
- retry:
- chain_len = 0;
/*
* iter_prev points to the non-removed node prior to the
* insert location.
if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
goto insert;
next = rcu_dereference(clear_flag(iter)->p.next);
- if (unlikely(is_gc(next)))
- goto gc_node;
- if (unlikely(replace_pinned)) {
- /*
- * We're in the retry of a node
- * replacement. Only get exact iter
- * pointer match. We own it, so it
- * _needs_ to be there at some point.
- */
- if (clear_flag(iter) == replace_pinned)
- goto replace;
- }
- /*
- * Next is removed but not gc'd. We need to
- * busy-loop, because a concurrent replacement
- * is keeping it temporarily pinned there but we
- * cannot attach to it. The easiest solution is
- * to retry.
- */
if (unlikely(is_removed(next)))
- goto retry;
+ goto gc_node;
if ((mode == ADD_UNIQUE || mode == ADD_REPLACE)
&& !is_dummy(next)
&& !ht->compare_fct(node->key, node->key_len,
assert(node != clear_flag(iter));
assert(!is_removed(iter_prev));
assert(!is_removed(iter));
- assert(!is_gc(iter_prev));
- assert(!is_gc(iter));
assert(iter_prev != node);
- assert(!replace_pinned);
if (!dummy)
node->p.next = clear_flag(iter);
else
}
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(!is_gc(iter_prev));
- assert(!is_gc(iter));
assert(iter_prev != node);
assert(!dummy);
- node->p.next = clear_flag(next);
+ node->p.next = clear_flag(iter);
if (is_dummy(iter))
new_node = flag_dummy(node);
else
new_node = node;
/*
- * Try to delete to-be-replaced node. Don't gc yet. Not
- * performing gc here is important, because this lets
- * concurrent lookups see the old node until we
- * atomically swap the new node into its place.
- *
- * This algorithm is _not_ strictly lock-free between
- * _cds_lfht_del and the uatomic_cmpxchg of the
- * replacement operation, so a replacement should _not_
- * crash here (which means: don't do replacements if you
- * need strict lock-free guarantees).
- */
- if (!replace_pinned) {
- if (_cds_lfht_del(ht, size, clear_flag(iter), 0, 0))
- continue; /* concurrently removed. retry. */
- }
- /*
- * After _cds_lfht_del succeeds, we have pinned the
- * to-be-removed node in place by setting its removed
- * flag, but not its gc flag. If we fail to cmpxchg our
- * new node with this node, we need to retry everything
- * from the initial lookup, and only stop when we reach
- * the node we pinned into place.
+ * 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.
*/
- return_node = uatomic_cmpxchg(&iter_prev->p.next,
- iter, new_node);
- if (return_node != iter) {
- /*
- * If cmpxchg fails, we need to do path
- * compression, but end it by placing our own
- * node into place.
- */
- replace_pinned = clear_flag(iter);
+ new_node = flag_removed(new_node);
+ if (uatomic_cmpxchg(&iter_prev->p.next,
+ iter, new_node) != iter) {
continue; /* retry */
} else {
- /*
- * cmpxchg succeeded. gc unnecessary, because we
- * unlinked the return_node ourself with the
- * cmpxchg.
- */
- return_node = clear_flag(return_node);
- goto end;
+ return_node = iter_prev;
+ goto gc_end;
}
gc_node:
assert(!is_removed(iter));
- assert(!is_gc(iter));
if (is_dummy(iter))
new_next = flag_dummy(clear_flag(next));
else
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;
}
static
int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
struct cds_lfht_node *node,
- int dummy_removal, int do_gc)
+ int dummy_removal)
{
struct cds_lfht_node *dummy, *next, *old;
struct _cds_lfht_node *lookup;
/* logically delete the node */
assert(!is_dummy(node));
- assert(!is_gc(node));
assert(!is_removed(node));
old = rcu_dereference(node->p.next);
do {
else
assert(!is_dummy(next));
new_next = flag_removed(next);
- if (do_gc)
- new_next = flag_gc(new_next);
old = uatomic_cmpxchg(&node->p.next, next, new_next);
} while (old != next);
/* We performed the (logical) deletion. */
flagged = 1;
- if (!do_gc)
- goto end;
-
/*
* Ensure that the node is not visible to readers anymore: lookup for
* the node, and remove it (along with any other logically removed node)
fini_node->p.reverse_hash =
bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
(void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
- fini_node, 1, 1);
+ fini_node, 1);
if (CMM_LOAD_SHARED(ht->in_progress_destroy))
break;
}
return ht;
}
-struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len)
+void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
+ struct cds_lfht_iter *iter)
{
struct cds_lfht_node *node, *next, *dummy_node;
struct _cds_lfht_node *lookup;
break;
}
next = rcu_dereference(node->p.next);
- /*
- * We consider return nodes marked removed but not gc as
- * hits for lookup vs replacement consistency.
- */
- if (likely(!is_gc(next))
+ if (likely(!is_removed(next))
&& !is_dummy(next)
&& likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
break;
node = clear_flag(next);
}
assert(!node || !is_dummy(rcu_dereference(node->p.next)));
- return node;
+ iter->node = node;
+ iter->next = next;
}
-struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht,
- struct cds_lfht_node *node)
+void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
{
- struct cds_lfht_node *next;
+ struct cds_lfht_node *node, *next;
unsigned long reverse_hash;
void *key;
size_t key_len;
+ node = iter->node;
reverse_hash = node->p.reverse_hash;
key = node->key;
key_len = node->key_len;
- next = rcu_dereference(node->p.next);
+ next = iter->next;
node = clear_flag(next);
for (;;) {
break;
}
next = rcu_dereference(node->p.next);
- /*
- * We consider return nodes marked removed but not gc as
- * hits for lookup vs replacement consistency.
- */
- if (likely(!is_gc(next))
+ if (likely(!is_removed(next))
&& !is_dummy(next)
&& likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
break;
node = clear_flag(next);
}
assert(!node || !is_dummy(rcu_dereference(node->p.next)));
- return node;
+ iter->node = node;
+ iter->next = next;
}
void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
int ret;
size = rcu_dereference(ht->t.size);
- ret = _cds_lfht_del(ht, size, node, 0, 1);
+ ret = _cds_lfht_del(ht, size, node, 0);
if (!ret)
ht_count_del(ht, size);
return ret;
if (!is_dummy(node))
return -EPERM;
assert(!is_removed(node));
- assert(!is_gc(node));
} while (!is_end(node));
/*
* size accessed without rcu_dereference because hash table is
node = (struct cds_lfht_node *) lookup;
do {
next = rcu_dereference(node->p.next);
- if (is_removed(next) || is_gc(next)) {
+ if (is_removed(next)) {
assert(!is_dummy(next));
(*removed)++;
} else if (!is_dummy(next))