X-Git-Url: https://git.lttng.org/?p=urcu.git;a=blobdiff_plain;f=rculfhash.c;h=98547851b210a64e25788eef3cca2bc87f7ce8e0;hp=606f46e904b2fde4e0c3f4be1e66d5e5a437929d;hb=92cfe223501f564d3bd726a6c580702daeef4198;hpb=24365af70940dac62e3910e566e91f549a045423 diff --git a/rculfhash.c b/rculfhash.c index 606f46e..9854785 100644 --- a/rculfhash.c +++ b/rculfhash.c @@ -20,6 +20,80 @@ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA */ +/* + * Based on the following articles: + * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free + * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405. + * - Michael, M. M. High performance dynamic lock-free hash tables + * and list-based sets. In Proceedings of the fourteenth annual ACM + * symposium on Parallel algorithms and architectures, ACM Press, + * (2002), 73-82. + * + * Some specificities of this Lock-Free Expandable RCU Hash Table + * implementation: + * + * - RCU read-side critical section allows readers to perform hash + * table lookups and use the returned objects safely by delaying + * memory reclaim of a grace period. + * - Add and remove operations are lock-free, and do not need to + * allocate memory. They need to be executed within RCU read-side + * critical section to ensure the objects they read are valid and to + * deal with the cmpxchg ABA problem. + * - add and add_unique operations are supported. add_unique checks if + * the node key already exists in the hash table. It ensures no key + * duplicata exists. + * - The resize operation executes concurrently with add/remove/lookup. + * - Hash table nodes are contained within a split-ordered list. This + * list is ordered by incrementing reversed-bits-hash value. + * - An index of dummy nodes is kept. These dummy nodes are the hash + * table "buckets", and they are also chained together in the + * split-ordered list, which allows recursive expansion. + * - The resize operation only allows expanding the hash table. + * It is triggered either through an API call or automatically by + * detecting long chains in the add operation. + * - Resize operation initiated by long chain detection is executed by a + * call_rcu thread, which keeps lock-freedom of add and remove. + * - Resize operations are protected by a mutex. + * - The removal operation is split in two parts: first, a "removed" + * flag is set in the next pointer within the node to remove. Then, + * a "garbage collection" is performed in the bucket containing the + * removed node (from the start of the bucket up to the removed node). + * All encountered nodes with "removed" flag set in their next + * pointers are removed from the linked-list. If the cmpxchg used for + * removal fails (due to concurrent garbage-collection or concurrent + * add), we retry from the beginning of the bucket. This ensures that + * the node with "removed" flag set is removed from the hash table + * (not visible to lookups anymore) before the RCU read-side critical + * section held across removal ends. Furthermore, this ensures that + * the node with "removed" flag set is removed from the linked-list + * before its memory is reclaimed. Only the thread which removal + * successfully set the "removed" flag (with a cmpxchg) into a node's + * next pointer is considered to have succeeded its removal (and thus + * owns the node to reclaim). Because we garbage-collect starting from + * an invariant node (the start-of-bucket dummy node) up to the + * "removed" node (or find a reverse-hash that is higher), we are sure + * that a successful traversal of the chain leads to a chain that is + * present in the linked-list (the start node is never removed) and + * that is does not contain the "removed" node anymore, even if + * concurrent delete/add operations are changing the structure of the + * list concurrently. + * - The add operation performs gargage collection of buckets if it + * encounters nodes with removed flag set in the bucket where it wants + * to add its new node. This ensures lock-freedom of add operation by + * helping the remover unlink nodes from the list rather than to wait + * for it do to so. + * - A RCU "order table" indexed by log2(hash index) is copied and + * expanded by the resize operation. This order table allows finding + * the "dummy node" tables. + * - There is one dummy node table per hash index order. The size of + * each dummy node table is half the number of hashes contained in + * this order. + * - call_rcu is used to garbage-collect the old order table. + * - The per-order dummy node tables contain a compact version of the + * hash table nodes. These tables are invariant after they are + * populated into the hash table. + */ + #define _LGPL_SOURCE #include #include @@ -38,16 +112,14 @@ #include #include -//#define DEBUG /* Test */ - #ifdef DEBUG -#define dbg_printf(args...) printf(args) +#define dbg_printf(fmt, args...) printf(fmt, ## args) #else -#define dbg_printf(args...) +#define dbg_printf(fmt, args...) #endif -#define CHAIN_LEN_TARGET 1 -#define CHAIN_LEN_RESIZE_THRESHOLD 2 +#define CHAIN_LEN_TARGET 4 +#define CHAIN_LEN_RESIZE_THRESHOLD 8 #ifndef max #define max(a, b) ((a) > (b) ? (a) : (b)) @@ -76,7 +148,7 @@ struct rcu_ht { ht_compare_fct compare_fct; unsigned long hash_seed; pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ - unsigned int in_progress_resize; + unsigned int in_progress_resize, in_progress_destroy; void (*ht_call_rcu)(struct rcu_head *head, void (*func)(struct rcu_head *head)); }; @@ -290,7 +362,7 @@ void check_resize(struct rcu_ht *ht, struct rcu_table *t, chain_len); if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) ht_resize_lazy(ht, t, - get_count_order_u32(chain_len - CHAIN_LEN_TARGET + 1)); + get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1))); } static @@ -353,10 +425,10 @@ void _ht_gc_bucket(struct rcu_ht_node *dummy, struct rcu_ht_node *node) for (;;) { if (unlikely(!clear_flag(iter))) return; - if (clear_flag(iter)->p.reverse_hash > node->p.reverse_hash) + if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) return; next = rcu_dereference(clear_flag(iter)->p.next); - if (is_removed(next)) + if (likely(is_removed(next))) break; iter_prev = clear_flag(iter); iter = next; @@ -402,10 +474,10 @@ struct rcu_ht_node *_ht_add(struct rcu_ht *ht, struct rcu_table *t, for (;;) { if (unlikely(!clear_flag(iter))) goto insert; - if (clear_flag(iter)->p.reverse_hash > node->p.reverse_hash) + if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) goto insert; next = rcu_dereference(clear_flag(iter)->p.next); - if (is_removed(next)) + if (unlikely(is_removed(next))) goto gc_node; if (unique && !is_dummy(next) @@ -468,7 +540,7 @@ int _ht_remove(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node) old = rcu_dereference(node->p.next); do { next = old; - if (is_removed(next)) + if (unlikely(is_removed(next))) goto end; assert(!is_dummy(next)); old = uatomic_cmpxchg(&node->p.next, next, @@ -525,10 +597,14 @@ void init_table(struct rcu_ht *ht, struct rcu_table *t, new_node->p.reverse_hash = bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); (void) _ht_add(ht, t, new_node, 0, 1); + if (CMM_LOAD_SHARED(ht->in_progress_destroy)) + break; } /* Update table size */ t->size = !i ? 1 : (1UL << i); dbg_printf("rculfhash: init new size: %lu\n", t->size); + if (CMM_LOAD_SHARED(ht->in_progress_destroy)) + break; } t->resize_target = t->size; t->resize_initiated = 0; @@ -673,6 +749,7 @@ int ht_destroy(struct rcu_ht *ht) int ret; /* Wait for in-flight resize operations to complete */ + CMM_STORE_SHARED(ht->in_progress_destroy, 1); while (uatomic_read(&ht->in_progress_resize)) poll(NULL, 0, 100); /* wait for 100ms */ ret = ht_delete_dummy(ht); @@ -736,8 +813,8 @@ void _do_ht_resize(struct rcu_ht *ht) if (old_size == new_size) return; new_order = get_count_order_ulong(new_size) + 1; - dbg_printf("rculfhash: resize from %lu (order %lu) to %lu (order %lu) buckets\n", - old_size, old_order, new_size, new_order); + printf("rculfhash: resize from %lu (order %lu) to %lu (order %lu) buckets\n", + old_size, old_order, new_size, new_order); new_t = malloc(sizeof(struct rcu_table) + (new_order * sizeof(struct _rcu_ht_node *))); assert(new_size > old_size);