X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=rculfhash.c;h=bff26a8d4fe668e2d4330927dcc812384aea039e;hp=30c5c8a2e75390e72623e45af1f0889bbbcccd84;hb=15cfbec77d2c573110cc936d5b33745d44207b50;hpb=48ed1c184a1aa1da7d84c2d0bc1b9528627248ac diff --git a/rculfhash.c b/rculfhash.c index 30c5c8a..bff26a8 100644 --- a/rculfhash.c +++ b/rculfhash.c @@ -144,7 +144,6 @@ #include #include #include -#include #include #include #include @@ -194,17 +193,13 @@ /* * 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 @@ -250,7 +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 */ }; @@ -279,11 +274,6 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, 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. @@ -565,7 +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, @@ -574,7 +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)); } @@ -592,9 +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, @@ -603,7 +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)); } @@ -678,18 +674,6 @@ struct cds_lfht_node *flag_removed(struct cds_lfht_node *node) 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) { @@ -745,10 +729,8 @@ 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; 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; @@ -768,12 +750,12 @@ void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node 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 @@ -785,6 +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, @@ -792,12 +848,11 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, 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); @@ -806,10 +861,8 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, } 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. @@ -827,27 +880,8 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, 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, @@ -870,10 +904,7 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, 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 @@ -894,65 +925,17 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, } replace: - 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); - 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. - */ - 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); - continue; /* retry */ + + if (!_cds_lfht_replace(ht, size, clear_flag(iter), next, + node)) { + return_node = clear_flag(iter); + goto end; /* gc already done */ } else { - /* - * cmpxchg succeeded. gc unnecessary, because we - * unlinked the return_node ourself with the - * cmpxchg. - */ - return_node = clear_flag(return_node); - goto end; + continue; /* retry */ } gc_node: assert(!is_removed(iter)); - assert(!is_gc(iter)); if (is_dummy(iter)) new_next = flag_dummy(clear_flag(next)); else @@ -974,16 +957,18 @@ end: 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; 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_gc(node)); assert(!is_removed(node)); old = rcu_dereference(node->p.next); do { @@ -997,17 +982,12 @@ int _cds_lfht_del(struct cds_lfht *ht, unsigned long size, 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) @@ -1028,8 +1008,9 @@ end: if (flagged) { assert(is_removed(rcu_dereference(node->p.next))); return 0; - } else + } else { return -ENOENT; + } } static @@ -1214,7 +1195,7 @@ void remove_table_partition(struct cds_lfht *ht, unsigned long i, 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; } @@ -1332,7 +1313,8 @@ struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct, 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; @@ -1353,19 +1335,15 @@ struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key 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); - /* - * 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; @@ -1373,38 +1351,35 @@ struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key 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_duplicate(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 (;;) { 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); - /* - * 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; @@ -1412,7 +1387,46 @@ struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht, node = clear_flag(next); } assert(!node || !is_dummy(rcu_dereference(node->p.next))); - return node; + iter->node = node; + iter->next = next; +} + +void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter) +{ + struct cds_lfht_node *node, *next; + + node = iter->node; + next = iter->next; + node = clear_flag(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->node = (struct cds_lfht_node *) lookup; + cds_lfht_next(ht, iter); } void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node) @@ -1443,7 +1457,7 @@ 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; @@ -1459,13 +1473,23 @@ 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, 1); + ret = _cds_lfht_del(ht, size, iter->node, 0); if (!ret) ht_count_del(ht, size); return ret; @@ -1486,7 +1510,6 @@ int cds_lfht_delete_dummy(struct cds_lfht *ht) 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 @@ -1532,13 +1555,25 @@ 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; @@ -1547,9 +1582,11 @@ void cds_lfht_count_nodes(struct cds_lfht *ht, node = (struct cds_lfht_node *) lookup; do { next = rcu_dereference(node->p.next); - if (is_removed(next) || is_gc(next)) { - assert(!is_dummy(next)); - (*removed)++; + if (is_removed(next)) { + if (!is_dummy(next)) + (*removed)++; + else + (nr_dummy)++; } else if (!is_dummy(next)) (*count)++; else @@ -1557,6 +1594,15 @@ 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 */