X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=rculfhash.c;h=0d8c5d6150ae9fa9eecc05d46b96e2ff4e6ef44c;hp=8058931b386cf9d75f64173d7ff0e3333d486032;hb=9b35f1e4c8561fd8937ab394e65cc9b230ea854d;hpb=9d2614f07691a813a3c560a6c0bcd0a7be854ed5 diff --git a/rculfhash.c b/rculfhash.c index 8058931..0d8c5d6 100644 --- a/rculfhash.c +++ b/rculfhash.c @@ -38,8 +38,16 @@ #include #include -#define BUCKET_SIZE_RESIZE_THRESHOLD 32 -#define MAX_NR_BUCKETS 1048576 /* 1M buckets */ +#define DEBUG /* Test */ + +#ifdef DEBUG +#define dbg_printf(args...) printf(args) +#else +#define dbg_printf(args...) +#endif + +#define CHAIN_LEN_TARGET 1 +#define CHAIN_LEN_RESIZE_THRESHOLD 2 #ifndef max #define max(a, b) ((a) > (b) ? (a) : (b)) @@ -47,6 +55,8 @@ struct rcu_table { unsigned long size; /* always a power of 2 */ + unsigned long resize_target; + int resize_initiated; struct rcu_head head; struct rcu_ht_node *tbl[0]; }; @@ -54,9 +64,9 @@ struct rcu_table { struct rcu_ht { struct rcu_table *t; /* shared */ ht_hash_fct hash_fct; - void *hashseed; + ht_compare_fct compare_fct; + unsigned long hash_seed; pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ - unsigned long target_size; void (*ht_call_rcu)(struct rcu_head *head, void (*func)(struct rcu_head *head)); }; @@ -66,23 +76,12 @@ struct rcu_resize_work { struct rcu_ht *ht; }; -static -void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth); - -static -void check_resize(struct rcu_ht *ht, struct rcu_table *t, - unsigned long chain_len) -{ - //printf("check resize chain len %lu\n", chain_len); - if (chain_len >= BUCKET_SIZE_RESIZE_THRESHOLD) - ht_resize_lazy(ht, t, chain_len / BUCKET_SIZE_RESIZE_THRESHOLD); -} - /* * Algorithm to reverse bits in a word by lookup table, extended to * 64-bit words. - * ref. + * Source: * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable + * Originally from Public Domain. */ static const uint8_t BitReverseTable256[256] = @@ -134,6 +133,45 @@ unsigned long bit_reverse_ulong(unsigned long v) #endif } +/* + * Algorithm to find the log2 of a 32-bit unsigned integer. + * source: http://graphics.stanford.edu/~seander/bithacks.html#IntegerLogLookup + * Originally from Public Domain. + */ +static const char LogTable256[256] = +{ +#define LT(n) n, n, n, n, n, n, n, n, n, n, n, n, n, n, n, n + -1, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, + LT(4), LT(5), LT(5), LT(6), LT(6), LT(6), LT(6), + LT(7), LT(7), LT(7), LT(7), LT(7), LT(7), LT(7), LT(7) +}; + +uint32_t log2_u32(uint32_t v) +{ + uint32_t t, tt; + + if ((tt = (v >> 16))) + return (t = (tt >> 8)) + ? 24 + LogTable256[t] + : 16 + LogTable256[tt]; + else + return (t = (v >> 8)) + ? 8 + LogTable256[t] + : LogTable256[v]; +} + +static +void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth); + +static +void check_resize(struct rcu_ht *ht, struct rcu_table *t, + uint32_t chain_len) +{ + if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) + ht_resize_lazy(ht, t, + log2_u32(chain_len - CHAIN_LEN_TARGET - 1)); +} + static struct rcu_ht_node *clear_flag(struct rcu_ht_node *node) { @@ -153,7 +191,7 @@ struct rcu_ht_node *flag_removed(struct rcu_ht_node *node) } static -void _uatomic_max(unsigned long *ptr, unsigned long v) +unsigned long _uatomic_max(unsigned long *ptr, unsigned long v) { unsigned long old1, old2; @@ -161,103 +199,132 @@ void _uatomic_max(unsigned long *ptr, unsigned long v) do { old2 = old1; if (old2 >= v) - break; + return old2; } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); + return v; } +/* + * Remove all logically deleted nodes from a bucket up to a certain node key. + */ static -void _ht_add(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node) +void _ht_gc_bucket(struct rcu_ht_node *dummy, struct rcu_ht_node *node) { - struct rcu_ht_node *iter_prev = NULL, *iter = NULL; + struct rcu_ht_node *iter_prev, *iter, *next; + + for (;;) { + iter_prev = dummy; + /* We can always skip the dummy node initially */ + iter = rcu_dereference(iter_prev->next); + assert(iter_prev->reverse_hash <= node->reverse_hash); + for (;;) { + if (unlikely(!iter)) + return; + if (clear_flag(iter)->reverse_hash > node->reverse_hash) + return; + next = rcu_dereference(clear_flag(iter)->next); + if (is_removed(next)) + break; + iter_prev = iter; + iter = next; + } + assert(!is_removed(iter)); + (void) uatomic_cmpxchg(&iter_prev->next, iter, clear_flag(next)); + } +} + +static +int _ht_add(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node, + int unique) +{ + struct rcu_ht_node *iter_prev, *dummy, *iter, *next; if (!t->size) - return; + return 0; for (;;) { - unsigned long chain_len = 0; + uint32_t chain_len = 0; + /* + * iter_prev points to the non-removed node prior to the + * insert location. + */ iter_prev = rcu_dereference(t->tbl[node->hash & (t->size - 1)]); - //printf("iter prev %p hash %lu bucket %lu\n", iter_prev, - // node->hash, node->hash & (t->size - 1)); - assert(iter_prev); + /* We can always skip the dummy node initially */ + iter = rcu_dereference(iter_prev->next); assert(iter_prev->reverse_hash <= node->reverse_hash); for (;;) { - iter = clear_flag(rcu_dereference(iter_prev->next)); if (unlikely(!iter)) - break; - if (iter->reverse_hash < node->reverse_hash) - break; - iter_prev = iter; - check_resize(ht, t, ++chain_len); + goto insert; + if (clear_flag(iter)->reverse_hash > node->reverse_hash) + goto insert; + next = rcu_dereference(clear_flag(iter)->next); + if (is_removed(next)) + goto gc_node; + /* Only account for identical reverse hash once */ + if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash) + check_resize(ht, t, ++chain_len); + iter_prev = clear_flag(iter); + iter = next; } - /* add in iter_prev->next */ - if (is_removed(iter)) - continue; - assert(node != iter); - node->next = iter; + insert: + assert(node != clear_flag(iter)); + assert(!is_removed(iter_prev)); assert(iter_prev != node); - if (uatomic_cmpxchg(&iter_prev->next, iter, node) != iter) - continue; - break; + node->next = iter; + if (uatomic_cmpxchg(&iter_prev->next, iter, + node) != iter) + continue; /* retry */ + else + goto gc_end; + gc_node: + assert(!is_removed(iter)); + (void) uatomic_cmpxchg(&iter_prev->next, iter, clear_flag(next)); + /* retry */ } +gc_end: + /* Garbage collect logically removed nodes in the bucket */ + dummy = rcu_dereference(t->tbl[node->hash & (t->size - 1)]); + _ht_gc_bucket(dummy, node); + return 0; } static int _ht_remove(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node) { - struct rcu_ht_node *iter_prev, *iter, *next, *old; - unsigned long chain_len; - int found, ret = 0; + struct rcu_ht_node *dummy, *next, *old; int flagged = 0; -retry: - chain_len = 0; - found = 0; - iter_prev = rcu_dereference(t->tbl[node->hash & (t->size - 1)]); - assert(iter_prev); - assert(iter_prev->reverse_hash <= node->reverse_hash); - for (;;) { - iter = clear_flag(rcu_dereference(iter_prev->next)); - if (unlikely(!iter)) - break; - if (iter->reverse_hash < node->reverse_hash) - break; - if (iter == node) { - found = 1; - break; - } - iter_prev = iter; - } - if (!found) { - ret = -ENOENT; - goto end; - } - next = rcu_dereference(iter->next); - if (!flagged) { - if (is_removed(next)) { - ret = -ENOENT; + /* logically delete the node */ + old = rcu_dereference(node->next); + do { + next = old; + if (is_removed(next)) goto end; - } - /* set deletion flag */ - if ((old = uatomic_cmpxchg(&iter->next, next, - flag_removed(next))) != next) { - if (old == flag_removed(next)) { - ret = -ENOENT; - goto end; - } else { - goto retry; - } - } - flagged = 1; - } + assert(!node->dummy); + old = uatomic_cmpxchg(&node->next, next, + flag_removed(next)); + } while (old != next); + + /* We performed the (logical) deletion. */ + flagged = 1; + /* - * Remove the element from the list. Retry if there has been a - * concurrent add (there cannot be a concurrent delete, because - * we won the deletion flag cmpxchg). + * Ensure that the node is not visible to readers anymore: lookup for + * the node, and remove it (along with any other logically removed node) + * if found. */ - if (uatomic_cmpxchg(&iter_prev->next, iter, clear_flag(next)) != iter) - goto retry; + dummy = rcu_dereference(t->tbl[node->hash & (t->size - 1)]); + _ht_gc_bucket(dummy, node); end: - return ret; + /* + * Only the flagging action indicated that we (and no other) + * removed the node from the hash. + */ + if (flagged) { + assert(is_removed(rcu_dereference(node->next))); + return 0; + } else + return -ENOENT; } static @@ -275,13 +342,15 @@ void init_table(struct rcu_ht *ht, struct rcu_table *t, t->tbl[i]->dummy = 1; t->tbl[i]->hash = i; t->tbl[i]->reverse_hash = bit_reverse_ulong(i); - _ht_add(ht, t, t->tbl[i]); + (void) _ht_add(ht, t, t->tbl[i], 0); } - t->size = end; + t->resize_target = t->size = end; + t->resize_initiated = 0; } struct rcu_ht *ht_new(ht_hash_fct hash_fct, - void *hashseed, + ht_compare_fct compare_fct, + unsigned long hash_seed, unsigned long init_size, void (*ht_call_rcu)(struct rcu_head *head, void (*func)(struct rcu_head *head))) @@ -290,7 +359,8 @@ struct rcu_ht *ht_new(ht_hash_fct hash_fct, ht = calloc(1, sizeof(struct rcu_ht)); ht->hash_fct = hash_fct; - ht->hashseed = hashseed; + ht->compare_fct = compare_fct; + ht->hash_seed = hash_seed; ht->ht_call_rcu = ht_call_rcu; /* this mutex should not nest in read-side C.S. */ pthread_mutex_init(&ht->resize_mutex, NULL); @@ -300,17 +370,16 @@ struct rcu_ht *ht_new(ht_hash_fct hash_fct, pthread_mutex_lock(&ht->resize_mutex); init_table(ht, ht->t, 0, max(init_size, 1)); pthread_mutex_unlock(&ht->resize_mutex); - ht->target_size = ht->t->size; return ht; } -struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key) +struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key, size_t key_len) { struct rcu_table *t; struct rcu_ht_node *node; unsigned long hash, reverse_hash; - hash = ht->hash_fct(ht->hashseed, key); + hash = ht->hash_fct(key, key_len, ht->hash_seed); reverse_hash = bit_reverse_ulong(hash); t = rcu_dereference(ht->t); @@ -318,17 +387,18 @@ struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key) for (;;) { if (unlikely(!node)) break; - if (node->reverse_hash > reverse_hash) { + if (unlikely(node->reverse_hash > reverse_hash)) { node = NULL; break; } - if (node->key == key) { - if (is_removed(rcu_dereference(node->next))) - node = NULL; - break; + if (!ht->compare_fct(node->key, node->key_len, key, key_len)) { + if (likely(!is_removed(rcu_dereference(node->next))) + && likely(!node->dummy)) + break; } node = clear_flag(rcu_dereference(node->next)); } + assert(!node || !node->dummy); return node; } @@ -336,11 +406,22 @@ void ht_add(struct rcu_ht *ht, struct rcu_ht_node *node) { struct rcu_table *t; - node->hash = ht->hash_fct(ht->hashseed, node->key); + node->hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); node->reverse_hash = bit_reverse_ulong((unsigned long) node->hash); t = rcu_dereference(ht->t); - _ht_add(ht, t, node); + (void) _ht_add(ht, t, node, 0); +} + +int ht_add_unique(struct rcu_ht *ht, struct rcu_ht_node *node) +{ + struct rcu_table *t; + + node->hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); + node->reverse_hash = bit_reverse_ulong((unsigned long) node->hash); + + t = rcu_dereference(ht->t); + return _ht_add(ht, t, node, 1); } int ht_remove(struct rcu_ht *ht, struct rcu_ht_node *node) @@ -365,6 +446,7 @@ int ht_delete_dummy(struct rcu_ht *ht) if (!node->dummy) return -EPERM; node = node->next; + assert(!is_removed(node)); } while (node); /* Internal sanity check: all nodes left should be dummy */ for (i = 0; i < t->size; i++) { @@ -390,6 +472,30 @@ int ht_destroy(struct rcu_ht *ht) return ret; } +void ht_count_nodes(struct rcu_ht *ht, + unsigned long *count, + unsigned long *removed) +{ + struct rcu_table *t; + struct rcu_ht_node *node, *next; + + *count = 0; + *removed = 0; + + t = rcu_dereference(ht->t); + /* Check that the table is empty */ + node = rcu_dereference(t->tbl[0]); + do { + next = rcu_dereference(node->next); + if (is_removed(next)) { + assert(!node->dummy); + (*removed)++; + } else if (!node->dummy) + (*count)++; + node = clear_flag(next); + } while (node); +} + static void ht_free_table_cb(struct rcu_head *head) { @@ -405,12 +511,12 @@ void _do_ht_resize(struct rcu_ht *ht) unsigned long new_size, old_size; struct rcu_table *new_t, *old_t; - //return; //TEST - old_t = ht->t; old_size = old_t->size; - new_size = CMM_LOAD_SHARED(ht->target_size); + new_size = CMM_LOAD_SHARED(old_t->resize_target); + dbg_printf("rculfhash: resize from %lu to %lu buckets\n", + old_size, new_size); if (old_size == new_size) return; new_t = malloc(sizeof(struct rcu_table) @@ -419,31 +525,31 @@ void _do_ht_resize(struct rcu_ht *ht) memcpy(&new_t->tbl, &old_t->tbl, old_size * sizeof(struct rcu_ht_node *)); init_table(ht, new_t, old_size, new_size - old_size); - new_t->size = new_size; /* Changing table and size atomically wrt lookups */ rcu_assign_pointer(ht->t, new_t); ht->ht_call_rcu(&old_t->head, ht_free_table_cb); } static -void resize_target_update(struct rcu_ht *ht, struct rcu_table *t, - int growth_order) +unsigned long resize_target_update(struct rcu_table *t, + int growth_order) { - unsigned long new_size = t->size << growth_order; - - if (new_size > MAX_NR_BUCKETS) - new_size = MAX_NR_BUCKETS; - //printf("resize update prevtarget %lu current %lu order %d\n", - // ht->target_size, t->size, growth_order); - _uatomic_max(&ht->target_size, new_size); + return _uatomic_max(&t->resize_target, + t->size << growth_order); } void ht_resize(struct rcu_ht *ht, int growth) { - resize_target_update(ht, rcu_dereference(ht->t), growth); - pthread_mutex_lock(&ht->resize_mutex); - _do_ht_resize(ht); - pthread_mutex_unlock(&ht->resize_mutex); + struct rcu_table *t = rcu_dereference(ht->t); + unsigned long target_size; + + target_size = resize_target_update(t, growth); + if (t->size < target_size) { + CMM_STORE_SHARED(t->resize_initiated, 1); + pthread_mutex_lock(&ht->resize_mutex); + _do_ht_resize(ht); + pthread_mutex_unlock(&ht->resize_mutex); + } } static @@ -463,9 +569,13 @@ static void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth) { struct rcu_resize_work *work; + unsigned long target_size; - work = malloc(sizeof(*work)); - work->ht = ht; - resize_target_update(ht, t, growth); - ht->ht_call_rcu(&work->head, do_resize_cb); + target_size = resize_target_update(t, growth); + if (!CMM_LOAD_SHARED(t->resize_initiated) && t->size < target_size) { + work = malloc(sizeof(*work)); + work->ht = ht; + ht->ht_call_rcu(&work->head, do_resize_cb); + CMM_STORE_SHARED(t->resize_initiated, 1); + } }