X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=rculfhash.c;h=cf822fcd620e98b47d6a718582f1385887f34496;hp=1487980c832c115154937cda4b185024636f9e44;hb=1a401918bbce4b3be849340ac75a597f58db0f14;hpb=b198f0fda77c816d80f63cfa2e71a7e0b4496736 diff --git a/rculfhash.c b/rculfhash.c index 1487980..cf822fc 100644 --- a/rculfhash.c +++ b/rculfhash.c @@ -144,7 +144,6 @@ #include #include #include -#include #include #include #include @@ -246,7 +245,7 @@ struct cds_lfht { 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 */ + long count; /* global approximate item count */ struct ht_items_count *percpu_count; /* per-cpu item count */ }; @@ -256,7 +255,7 @@ struct rcu_resize_work { }; 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, @@ -556,7 +555,7 @@ void ht_count_add(struct cds_lfht *ht, unsigned long size) return; percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1); if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { - unsigned long count; + long count; dbg_printf("add percpu %lu\n", percpu_count); count = uatomic_add_return(&ht->count, @@ -565,7 +564,7 @@ void ht_count_add(struct cds_lfht *ht, unsigned long size) if (!(count & (count - 1))) { if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size) return; - dbg_printf("add set global %lu\n", count); + dbg_printf("add set global %ld\n", count); cds_lfht_resize_lazy_count(ht, size, count >> (CHAIN_LEN_TARGET - 1)); } @@ -583,9 +582,9 @@ void ht_count_del(struct cds_lfht *ht, unsigned long size) cpu = ht_get_cpu(); if (unlikely(cpu < 0)) return; - percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, -1); + percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1); if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { - unsigned long count; + long count; dbg_printf("del percpu %lu\n", percpu_count); count = uatomic_add_return(&ht->count, @@ -594,7 +593,13 @@ void ht_count_del(struct cds_lfht *ht, unsigned long size) if (!(count & (count - 1))) { if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size) return; - dbg_printf("del set global %lu\n", count); + dbg_printf("del set global %ld\n", count); + /* + * Don't shrink table if the number of nodes is below a + * certain threshold. + */ + if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1)) + return; cds_lfht_resize_lazy_count(ht, size, count >> (CHAIN_LEN_TARGET - 1)); } @@ -603,7 +608,7 @@ void ht_count_del(struct cds_lfht *ht, unsigned long size) #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) @@ -762,6 +767,80 @@ void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node return; } +static +int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, + struct cds_lfht_node *old_node, + struct cds_lfht_node *ret_next, + struct cds_lfht_node *new_node) +{ + struct cds_lfht_node *dummy, *old_next; + struct _cds_lfht_node *lookup; + int flagged = 0; + unsigned long hash, index, order; + + 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); + do { + /* Insert after node to be replaced */ + old_next = ret_next; + 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)); + } while (ret_next != old_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. + */ + hash = bit_reverse_ulong(old_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; + _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, @@ -846,36 +925,13 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, } 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: @@ -894,6 +950,7 @@ gc_end: 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; } @@ -907,6 +964,9 @@ int _cds_lfht_del(struct cds_lfht *ht, unsigned long size, 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)); assert(!is_removed(node)); @@ -948,8 +1008,9 @@ end: if (flagged) { assert(is_removed(rcu_dereference(node->p.next))); return 0; - } else + } else { return -ENOENT; + } } static @@ -973,18 +1034,20 @@ void partition_resize_helper(struct cds_lfht *ht, unsigned long i, 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; @@ -992,16 +1055,15 @@ void partition_resize_helper(struct cds_lfht *ht, unsigned long i, 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); } /* @@ -1032,8 +1094,6 @@ void init_table_populate_partition(struct cds_lfht *ht, unsigned long i, 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(); } @@ -1135,8 +1195,6 @@ void remove_table_partition(struct cds_lfht *ht, unsigned long 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(); } @@ -1274,11 +1332,11 @@ void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len, node = clear_flag(node); for (;;) { if (unlikely(is_end(node))) { - node = NULL; + node = next = NULL; break; } if (unlikely(node->p.reverse_hash > reverse_hash)) { - node = NULL; + node = next = NULL; break; } next = rcu_dereference(node->p.next); @@ -1294,7 +1352,7 @@ void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len, 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; @@ -1310,11 +1368,11 @@ void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter) for (;;) { if (unlikely(is_end(node))) { - node = NULL; + node = next = NULL; break; } if (unlikely(node->p.reverse_hash > reverse_hash)) { - node = NULL; + node = next = NULL; break; } next = rcu_dereference(node->p.next); @@ -1330,6 +1388,41 @@ void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter) 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; @@ -1358,7 +1451,7 @@ struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht, 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; @@ -1374,13 +1467,23 @@ struct cds_lfht_node *cds_lfht_replace(struct cds_lfht *ht, 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; @@ -1432,7 +1535,8 @@ int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr) 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); @@ -1446,13 +1550,25 @@ int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr) } void cds_lfht_count_nodes(struct cds_lfht *ht, + long *approx_before, unsigned long *count, - unsigned long *removed) + unsigned long *removed, + long *approx_after) { struct cds_lfht_node *node, *next; struct _cds_lfht_node *lookup; unsigned long nr_dummy = 0; + *approx_before = 0; + if (nr_cpus_mask >= 0) { + int i; + + for (i = 0; i < nr_cpus_mask + 1; i++) { + *approx_before += uatomic_read(&ht->percpu_count[i].add); + *approx_before -= uatomic_read(&ht->percpu_count[i].del); + } + } + *count = 0; *removed = 0; @@ -1462,8 +1578,10 @@ void cds_lfht_count_nodes(struct cds_lfht *ht, do { next = rcu_dereference(node->p.next); if (is_removed(next)) { - assert(!is_dummy(next)); - (*removed)++; + if (!is_dummy(next)) + (*removed)++; + else + (nr_dummy)++; } else if (!is_dummy(next)) (*count)++; else @@ -1471,6 +1589,15 @@ void cds_lfht_count_nodes(struct cds_lfht *ht, node = clear_flag(next); } while (!is_end(node)); dbg_printf("number of dummy nodes: %lu\n", nr_dummy); + *approx_after = 0; + if (nr_cpus_mask >= 0) { + int i; + + for (i = 0; i < nr_cpus_mask + 1; i++) { + *approx_after += uatomic_read(&ht->percpu_count[i].add); + *approx_after -= uatomic_read(&ht->percpu_count[i].del); + } + } } /* called with resize mutex held */ @@ -1482,8 +1609,8 @@ void _do_cds_lfht_grow(struct cds_lfht *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); } @@ -1498,8 +1625,8 @@ void _do_cds_lfht_shrink(struct cds_lfht *ht, 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. */ @@ -1517,6 +1644,9 @@ void _do_cds_lfht_resize(struct cds_lfht *ht) * 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); @@ -1585,7 +1715,11 @@ void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth) 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); @@ -1608,7 +1742,11 @@ void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, 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);