X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=rculfhash.c;h=6a80049ddd1f783a21e117d63437eec5771ba514;hp=7605036555f9526cfafbc8a1265ba4d732499a97;hb=53960c9f970e231b1ca5f53913b60d298540cf9a;hpb=df03fab8d93f4778ff06cad3e6aef2a61c65170c diff --git a/rculfhash.c b/rculfhash.c index 7605036..6a80049 100644 --- a/rculfhash.c +++ b/rculfhash.c @@ -1,7 +1,7 @@ /* * rculfhash.c * - * Userspace RCU library - Lock-Free Expandable RCU Hash Table + * Userspace RCU library - Lock-Free Resizable RCU Hash Table * * Copyright 2010-2011 - Mathieu Desnoyers * @@ -20,6 +20,117 @@ * 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 Resizable 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 for small tables only allows expanding the hash table. + * It is triggered automatically by detecting long chains in the add + * operation. + * - The resize operation for larger tables (and available through an + * API) allows both expanding and shrinking the hash table. + * - Per-CPU Split-counters are used to keep track of the number of + * nodes within the hash table for automatic resize triggering. + * - 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. + * + * A bit of ascii art explanation: + * + * Order index is the off-by-one compare to the actual power of 2 because + * we use index 0 to deal with the 0 special-case. + * + * This shows the nodes for a small table ordered by reversed bits: + * + * bits reverse + * 0 000 000 + * 4 100 001 + * 2 010 010 + * 6 110 011 + * 1 001 100 + * 5 101 101 + * 3 011 110 + * 7 111 111 + * + * This shows the nodes in order of non-reversed bits, linked by + * reversed-bit order. + * + * order bits reverse + * 0 0 000 000 + * | + * 1 | 1 001 100 <- <- + * | | | | + * 2 | | 2 010 010 | | + * | | | 3 011 110 | <- | + * | | | | | | | + * 3 -> | | | 4 100 001 | | + * -> | | 5 101 101 | + * -> | 6 110 011 + * -> 7 111 111 + */ + #define _LGPL_SOURCE #include #include @@ -28,6 +139,7 @@ #include #include +#include "config.h" #include #include #include @@ -38,16 +150,34 @@ #include #include -//#define DEBUG /* Test */ - #ifdef DEBUG -#define dbg_printf(args...) printf(args) +#define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args) #else -#define dbg_printf(args...) +#define dbg_printf(fmt, args...) #endif -#define CHAIN_LEN_TARGET 4 -#define CHAIN_LEN_RESIZE_THRESHOLD 8 +/* For testing */ +#define POISON_FREE + +/* + * Per-CPU split-counters lazily update the global counter each 1024 + * addition/removal. It automatically keeps track of resize required. + * We use the bucket length as indicator for need to expand for small + * tables and machines lacking per-cpu data suppport. + */ +#define COUNT_COMMIT_ORDER 10 +#define CHAIN_LEN_TARGET 1 +#define CHAIN_LEN_RESIZE_THRESHOLD 3 + +/* + * Define the minimum table size. Protects against hash table resize overload + * when too many entries are added quickly before the resize can complete. + * This is especially the case if the table could be shrinked to a size of 1. + * TODO: we might want to make the add/remove operations help the resize to + * add or remove dummy nodes when a resize is ongoing to ensure upper-bound on + * chain length. + */ +#define MIN_TABLE_SIZE 128 #ifndef max #define max(a, b) ((a) > (b) ? (a) : (b)) @@ -62,28 +192,41 @@ #define DUMMY_FLAG (1UL << 1) #define FLAGS_MASK ((1UL << 2) - 1) +struct ht_items_count { + unsigned long add, remove; +} __attribute__((aligned(CAA_CACHE_LINE_SIZE))); + +struct rcu_level { + struct rcu_head head; + struct _cds_lfht_node nodes[0]; +}; + 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]; + struct rcu_level *tbl[0]; }; -struct rcu_ht { +struct cds_lfht { struct rcu_table *t; /* shared */ - ht_hash_fct hash_fct; - ht_compare_fct compare_fct; + cds_lfht_hash_fct hash_fct; + cds_lfht_compare_fct compare_fct; unsigned long hash_seed; + int flags; pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ - unsigned int in_progress_resize; - void (*ht_call_rcu)(struct rcu_head *head, + unsigned int in_progress_resize, in_progress_destroy; + void (*cds_lfht_call_rcu)(struct rcu_head *head, void (*func)(struct rcu_head *head)); + void (*cds_lfht_synchronize_rcu)(void); + unsigned long count; /* global approximate item count */ + struct ht_items_count *percpu_count; /* per-cpu item count */ }; struct rcu_resize_work { struct rcu_head head; - struct rcu_ht *ht; + struct cds_lfht *ht; }; /* @@ -278,49 +421,221 @@ int get_count_order_ulong(unsigned long x) return order; } +#ifdef POISON_FREE +#define poison_free(ptr) \ + do { \ + memset(ptr, 0x42, sizeof(*(ptr))); \ + free(ptr); \ + } while (0) +#else +#define poison_free(ptr) free(ptr) +#endif + +static +void cds_lfht_resize_lazy(struct cds_lfht *ht, struct rcu_table *t, int growth); + +/* + * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are + * available, then we support hash table item accounting. + * In the unfortunate event the number of CPUs reported would be + * inaccurate, we use modulo arithmetic on the number of CPUs we got. + */ +#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) + +static +void cds_lfht_resize_lazy_count(struct cds_lfht *ht, struct rcu_table *t, + unsigned long count); + +static long nr_cpus_mask = -1; + +static +struct ht_items_count *alloc_per_cpu_items_count(void) +{ + struct ht_items_count *count; + + switch (nr_cpus_mask) { + case -2: + return NULL; + case -1: + { + long maxcpus; + + maxcpus = sysconf(_SC_NPROCESSORS_CONF); + if (maxcpus <= 0) { + nr_cpus_mask = -2; + return NULL; + } + /* + * round up number of CPUs to next power of two, so we + * can use & for modulo. + */ + maxcpus = 1UL << get_count_order_ulong(maxcpus); + nr_cpus_mask = maxcpus - 1; + } + /* Fall-through */ + default: + return calloc(nr_cpus_mask + 1, sizeof(*count)); + } +} + static -void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth); +void free_per_cpu_items_count(struct ht_items_count *count) +{ + poison_free(count); +} static -void check_resize(struct rcu_ht *ht, struct rcu_table *t, +int ht_get_cpu(void) +{ + int cpu; + + assert(nr_cpus_mask >= 0); + cpu = sched_getcpu(); + if (unlikely(cpu < 0)) + return cpu; + else + return cpu & nr_cpus_mask; +} + +static +void ht_count_add(struct cds_lfht *ht, struct rcu_table *t) +{ + unsigned long percpu_count; + int cpu; + + if (unlikely(!ht->percpu_count)) + return; + cpu = ht_get_cpu(); + if (unlikely(cpu < 0)) + return; + percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1); + if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { + unsigned long count; + + dbg_printf("add percpu %lu\n", percpu_count); + count = uatomic_add_return(&ht->count, + 1UL << COUNT_COMMIT_ORDER); + /* If power of 2 */ + if (!(count & (count - 1))) { + if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) + < t->size) + return; + dbg_printf("add set global %lu\n", count); + cds_lfht_resize_lazy_count(ht, t, + count >> (CHAIN_LEN_TARGET - 1)); + } + } +} + +static +void ht_count_remove(struct cds_lfht *ht, struct rcu_table *t) +{ + unsigned long percpu_count; + int cpu; + + if (unlikely(!ht->percpu_count)) + return; + cpu = ht_get_cpu(); + if (unlikely(cpu < 0)) + return; + percpu_count = uatomic_add_return(&ht->percpu_count[cpu].remove, -1); + if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { + unsigned long count; + + dbg_printf("remove percpu %lu\n", percpu_count); + count = uatomic_add_return(&ht->count, + -(1UL << COUNT_COMMIT_ORDER)); + /* If power of 2 */ + if (!(count & (count - 1))) { + if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) + >= t->size) + return; + dbg_printf("remove set global %lu\n", count); + cds_lfht_resize_lazy_count(ht, t, + count >> (CHAIN_LEN_TARGET - 1)); + } + } +} + +#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ + +static const long nr_cpus_mask = -1; + +static +struct ht_items_count *alloc_per_cpu_items_count(void) +{ + return NULL; +} + +static +void free_per_cpu_items_count(struct ht_items_count *count) +{ +} + +static +void ht_count_add(struct cds_lfht *ht, struct rcu_table *t) +{ +} + +static +void ht_count_remove(struct cds_lfht *ht, struct rcu_table *t) +{ +} + +#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ + + +static +void check_resize(struct cds_lfht *ht, struct rcu_table *t, uint32_t chain_len) { + unsigned long count; + + if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) + return; + count = uatomic_read(&ht->count); + /* + * Use bucket-local length for small table expand and for + * environments lacking per-cpu data support. + */ + if (count >= (1UL << COUNT_COMMIT_ORDER)) + return; if (chain_len > 100) - dbg_printf("rculfhash: WARNING: large chain length: %u.\n", + dbg_printf("WARNING: large chain length: %u.\n", chain_len); if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) - ht_resize_lazy(ht, t, + cds_lfht_resize_lazy(ht, t, get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1))); } static -struct rcu_ht_node *clear_flag(struct rcu_ht_node *node) +struct cds_lfht_node *clear_flag(struct cds_lfht_node *node) { - return (struct rcu_ht_node *) (((unsigned long) node) & ~FLAGS_MASK); + return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK); } static -int is_removed(struct rcu_ht_node *node) +int is_removed(struct cds_lfht_node *node) { return ((unsigned long) node) & REMOVED_FLAG; } static -struct rcu_ht_node *flag_removed(struct rcu_ht_node *node) +struct cds_lfht_node *flag_removed(struct cds_lfht_node *node) { - return (struct rcu_ht_node *) (((unsigned long) node) | REMOVED_FLAG); + return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG); } static -int is_dummy(struct rcu_ht_node *node) +int is_dummy(struct cds_lfht_node *node) { return ((unsigned long) node) & DUMMY_FLAG; } static -struct rcu_ht_node *flag_dummy(struct rcu_ht_node *node) +struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node) { - return (struct rcu_ht_node *) (((unsigned long) node) | DUMMY_FLAG); + return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG); } static @@ -337,29 +652,62 @@ unsigned long _uatomic_max(unsigned long *ptr, unsigned long v) return v; } +static +void cds_lfht_free_table_cb(struct rcu_head *head) +{ + struct rcu_table *t = + caa_container_of(head, struct rcu_table, head); + poison_free(t); +} + +static +void cds_lfht_free_level(struct rcu_head *head) +{ + struct rcu_level *l = + caa_container_of(head, struct rcu_level, head); + poison_free(l); +} + /* * Remove all logically deleted nodes from a bucket up to a certain node key. */ static -void _ht_gc_bucket(struct rcu_ht_node *dummy, struct rcu_ht_node *node) +void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node) { - struct rcu_ht_node *iter_prev, *iter, *next, *new_next; - + struct cds_lfht_node *iter_prev, *iter, *next, *new_next; + struct cds_lfht_node *iter_trace[64]; + unsigned long trace_idx = 0; + + memset(iter_trace, 0, sizeof(iter_trace)); + assert(!is_dummy(dummy)); + assert(!is_removed(dummy)); + assert(!is_dummy(node)); + assert(!is_removed(node)); for (;;) { + iter_trace[trace_idx++ & (64 - 1)] = (void *) 0x1; iter_prev = dummy; /* We can always skip the dummy node initially */ iter = rcu_dereference(iter_prev->p.next); + iter_trace[trace_idx++ & (64 - 1)] = iter; assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); + /* + * We should never be called with dummy (start of chain) + * and logically removed node (end of path compression + * marker) being the actual same node. This would be a + * bug in the algorithm implementation. + */ + assert(dummy != 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; + iter_trace[trace_idx++ & (64 - 1)] = iter; } assert(!is_removed(iter)); if (is_dummy(iter)) @@ -367,18 +715,21 @@ void _ht_gc_bucket(struct rcu_ht_node *dummy, struct rcu_ht_node *node) else new_next = clear_flag(next); (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); + iter_trace[trace_idx++ & (64 - 1)] = (void *) 0x2; } } static -struct rcu_ht_node *_ht_add(struct rcu_ht *ht, struct rcu_table *t, - struct rcu_ht_node *node, int unique, int dummy) +struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, struct rcu_table *t, + struct cds_lfht_node *node, int unique, int dummy) { - struct rcu_ht_node *iter_prev, *iter, *next, *new_node, *new_next, + struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next, *dummy_node; - struct _rcu_ht_node *lookup; + struct _cds_lfht_node *lookup; unsigned long hash, index, order; + assert(!is_dummy(node)); + assert(!is_removed(node)); if (!t->size) { assert(dummy); node->p.next = flag_dummy(NULL); @@ -394,18 +745,20 @@ struct rcu_ht_node *_ht_add(struct rcu_ht *ht, struct rcu_table *t, */ index = hash & (t->size - 1); order = get_count_order_ulong(index + 1); - lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)]; - iter_prev = (struct rcu_ht_node *) lookup; + lookup = &t->tbl[order]->nodes[index & ((!order ? 0 : (1UL << (order - 1))) - 1)]; + iter_prev = (struct cds_lfht_node *) lookup; /* We can always skip the dummy node initially */ iter = rcu_dereference(iter_prev->p.next); assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); for (;;) { + /* TODO: check if removed */ if (unlikely(!clear_flag(iter))) goto insert; - if (clear_flag(iter)->p.reverse_hash > node->p.reverse_hash) + /* TODO: check if removed */ + 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) @@ -423,6 +776,7 @@ struct rcu_ht_node *_ht_add(struct rcu_ht *ht, struct rcu_table *t, insert: assert(node != clear_flag(iter)); assert(!is_removed(iter_prev)); + assert(!is_removed(iter)); assert(iter_prev != node); if (!dummy) node->p.next = clear_flag(iter); @@ -450,27 +804,33 @@ gc_end: /* Garbage collect logically removed nodes in the bucket */ index = hash & (t->size - 1); order = get_count_order_ulong(index + 1); - lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)]; - dummy_node = (struct rcu_ht_node *) lookup; - _ht_gc_bucket(dummy_node, node); + lookup = &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); return node; } static -int _ht_remove(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node) +int _cds_lfht_remove(struct cds_lfht *ht, struct rcu_table *t, + struct cds_lfht_node *node, int dummy_removal) { - struct rcu_ht_node *dummy, *next, *old; - struct _rcu_ht_node *lookup; + struct cds_lfht_node *dummy, *next, *old; + struct _cds_lfht_node *lookup; int flagged = 0; unsigned long hash, index, order; /* logically delete the node */ + assert(!is_dummy(node)); + assert(!is_removed(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)); + if (dummy_removal) + assert(is_dummy(next)); + else + assert(!is_dummy(next)); old = uatomic_cmpxchg(&node->p.next, next, flag_removed(next)); } while (old != next); @@ -484,11 +844,12 @@ int _ht_remove(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node) * if found. */ hash = bit_reverse_ulong(node->p.reverse_hash); + assert(t->size > 0); index = hash & (t->size - 1); order = get_count_order_ulong(index + 1); - lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)]; - dummy = (struct rcu_ht_node *) lookup; - _ht_gc_bucket(dummy, node); + lookup = &t->tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))]; + dummy = (struct cds_lfht_node *) lookup; + _cds_lfht_gc_bucket(dummy, node); end: /* * Only the flagging action indicated that we (and no other) @@ -502,12 +863,12 @@ end: } static -void init_table(struct rcu_ht *ht, struct rcu_table *t, +void init_table(struct cds_lfht *ht, struct rcu_table *t, unsigned long first_order, unsigned long len_order) { unsigned long i, end_order; - dbg_printf("rculfhash: init table: first_order %lu end_order %lu\n", + dbg_printf("init table: first_order %lu end_order %lu\n", first_order, first_order + len_order); end_order = first_order + len_order; t->size = !first_order ? 0 : (1UL << (first_order - 1)); @@ -515,58 +876,116 @@ void init_table(struct rcu_ht *ht, struct rcu_table *t, unsigned long j, len; len = !i ? 1 : 1UL << (i - 1); - dbg_printf("rculfhash: init order %lu len: %lu\n", i, len); - t->tbl[i] = calloc(len, sizeof(struct _rcu_ht_node)); + dbg_printf("init order %lu len: %lu\n", i, len); + t->tbl[i] = calloc(1, sizeof(struct rcu_level) + + (len * sizeof(struct _cds_lfht_node))); for (j = 0; j < len; j++) { - dbg_printf("rculfhash: init entry: i %lu j %lu hash %lu\n", + struct cds_lfht_node *new_node = + (struct cds_lfht_node *) &t->tbl[i]->nodes[j]; + + dbg_printf("init entry: i %lu j %lu hash %lu\n", i, j, !i ? 0 : (1UL << (i - 1)) + j); - struct rcu_ht_node *new_node = - (struct rcu_ht_node *) &t->tbl[i][j]; new_node->p.reverse_hash = bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); - (void) _ht_add(ht, t, new_node, 0, 1); + (void) _cds_lfht_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); + dbg_printf("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; +} + +static +void fini_table(struct cds_lfht *ht, struct rcu_table *t, + unsigned long first_order, unsigned long len_order) +{ + long i, end_order; + + dbg_printf("fini table: first_order %lu end_order %lu\n", + first_order, first_order + len_order); + end_order = first_order + len_order; + assert(first_order > 0); + assert(t->size == (1UL << (end_order - 1))); + for (i = end_order - 1; i >= first_order; i--) { + unsigned long j, len; + + len = !i ? 1 : 1UL << (i - 1); + dbg_printf("fini order %lu len: %lu\n", i, len); + /* + * Update table size. Need to shrink this table prior to + * removal so gc lookups use non-logically-removed dummy + * nodes. + */ + t->size = 1UL << (i - 1); + /* Unlink */ + for (j = 0; j < len; j++) { + struct cds_lfht_node *fini_node = + (struct cds_lfht_node *) &t->tbl[i]->nodes[j]; + + dbg_printf("fini entry: i %lu j %lu hash %lu\n", + i, j, !i ? 0 : (1UL << (i - 1)) + j); + fini_node->p.reverse_hash = + bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); + (void) _cds_lfht_remove(ht, t, fini_node, 1); + if (CMM_LOAD_SHARED(ht->in_progress_destroy)) + break; + } + ht->cds_lfht_call_rcu(&t->tbl[i]->head, cds_lfht_free_level); + dbg_printf("fini new size: %lu\n", t->size); + if (CMM_LOAD_SHARED(ht->in_progress_destroy)) + break; } t->resize_target = t->size; t->resize_initiated = 0; } -struct rcu_ht *ht_new(ht_hash_fct hash_fct, - 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))) +struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct, + cds_lfht_compare_fct compare_fct, + unsigned long hash_seed, + unsigned long init_size, + int flags, + void (*cds_lfht_call_rcu)(struct rcu_head *head, + void (*func)(struct rcu_head *head)), + void (*cds_lfht_synchronize_rcu)(void)) { - struct rcu_ht *ht; + struct cds_lfht *ht; unsigned long order; - ht = calloc(1, sizeof(struct rcu_ht)); + /* init_size must be power of two */ + if (init_size && (init_size & (init_size - 1))) + return NULL; + ht = calloc(1, sizeof(struct cds_lfht)); ht->hash_fct = hash_fct; ht->compare_fct = compare_fct; ht->hash_seed = hash_seed; - ht->ht_call_rcu = ht_call_rcu; + ht->cds_lfht_call_rcu = cds_lfht_call_rcu; + ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu; ht->in_progress_resize = 0; + ht->percpu_count = alloc_per_cpu_items_count(); /* this mutex should not nest in read-side C.S. */ pthread_mutex_init(&ht->resize_mutex, NULL); - order = get_count_order_ulong(max(init_size, 1)) + 1; - ht->t = calloc(1, sizeof(struct rcu_table) - + (order * sizeof(struct _rcu_ht_node *))); + order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1; + ht->t = calloc(1, sizeof(struct cds_lfht) + + (order * sizeof(struct rcu_level *))); ht->t->size = 0; + ht->flags = flags; pthread_mutex_lock(&ht->resize_mutex); init_table(ht, ht->t, 0, order); pthread_mutex_unlock(&ht->resize_mutex); return ht; } -struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key, size_t key_len) +struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len) { struct rcu_table *t; - struct rcu_ht_node *node, *next; - struct _rcu_ht_node *lookup; + struct cds_lfht_node *node, *next; + struct _cds_lfht_node *lookup; unsigned long hash, reverse_hash, index, order; hash = ht->hash_fct(key, key_len, ht->hash_seed); @@ -575,10 +994,43 @@ struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key, size_t key_len) t = rcu_dereference(ht->t); index = hash & (t->size - 1); order = get_count_order_ulong(index + 1); - lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)]; - dbg_printf("rculfhash: lookup hash %lu index %lu order %lu aridx %lu\n", - hash, index, order, index & ((1UL << (order - 1)) - 1)); - node = (struct rcu_ht_node *) lookup; + lookup = &t->tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)]; + dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n", + hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1))); + node = (struct cds_lfht_node *) lookup; + for (;;) { + if (unlikely(!node)) + break; + if (unlikely(node->p.reverse_hash > reverse_hash)) { + node = NULL; + break; + } + next = rcu_dereference(node->p.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; +} + +struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht, + struct cds_lfht_node *node) +{ + struct cds_lfht_node *next; + unsigned long reverse_hash; + void *key; + size_t key_len; + + reverse_hash = node->p.reverse_hash; + key = node->key; + key_len = node->key_len; + next = rcu_dereference(node->p.next); + node = clear_flag(next); + for (;;) { if (unlikely(!node)) break; @@ -598,7 +1050,7 @@ struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key, size_t key_len) return node; } -void ht_add(struct rcu_ht *ht, struct rcu_ht_node *node) +void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node) { struct rcu_table *t; unsigned long hash; @@ -607,41 +1059,51 @@ void ht_add(struct rcu_ht *ht, struct rcu_ht_node *node) node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); t = rcu_dereference(ht->t); - (void) _ht_add(ht, t, node, 0, 0); + (void) _cds_lfht_add(ht, t, node, 0, 0); + ht_count_add(ht, t); } -struct rcu_ht_node *ht_add_unique(struct rcu_ht *ht, struct rcu_ht_node *node) +struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht, + struct cds_lfht_node *node) { struct rcu_table *t; unsigned long hash; + struct cds_lfht_node *ret; hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); t = rcu_dereference(ht->t); - return _ht_add(ht, t, node, 1, 0); + ret = _cds_lfht_add(ht, t, node, 1, 0); + if (ret != node) + ht_count_add(ht, t); + return ret; } -int ht_remove(struct rcu_ht *ht, struct rcu_ht_node *node) +int cds_lfht_remove(struct cds_lfht *ht, struct cds_lfht_node *node) { struct rcu_table *t; + int ret; t = rcu_dereference(ht->t); - return _ht_remove(ht, t, node); + ret = _cds_lfht_remove(ht, t, node, 0); + if (!ret) + ht_count_remove(ht, t); + return ret; } static -int ht_delete_dummy(struct rcu_ht *ht) +int cds_lfht_delete_dummy(struct cds_lfht *ht) { struct rcu_table *t; - struct rcu_ht_node *node; - struct _rcu_ht_node *lookup; + struct cds_lfht_node *node; + struct _cds_lfht_node *lookup; unsigned long order, i; t = ht->t; /* Check that the table is empty */ - lookup = &t->tbl[0][0]; - node = (struct rcu_ht_node *) lookup; + lookup = &t->tbl[0]->nodes[0]; + node = (struct cds_lfht_node *) lookup; do { node = clear_flag(node)->p.next; if (!is_dummy(node)) @@ -654,12 +1116,12 @@ int ht_delete_dummy(struct rcu_ht *ht) len = !order ? 1 : 1UL << (order - 1); for (i = 0; i < len; i++) { - dbg_printf("rculfhash: delete order %lu i %lu hash %lu\n", + dbg_printf("delete order %lu i %lu hash %lu\n", order, i, - bit_reverse_ulong(t->tbl[order][i].reverse_hash)); - assert(is_dummy(t->tbl[order][i].next)); + bit_reverse_ulong(t->tbl[order]->nodes[i].reverse_hash)); + assert(is_dummy(t->tbl[order]->nodes[i].next)); } - free(t->tbl[order]); + poison_free(t->tbl[order]); } return 0; } @@ -668,28 +1130,30 @@ int ht_delete_dummy(struct rcu_ht *ht) * Should only be called when no more concurrent readers nor writers can * possibly access the table. */ -int ht_destroy(struct rcu_ht *ht) +int cds_lfht_destroy(struct cds_lfht *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); + ret = cds_lfht_delete_dummy(ht); if (ret) return ret; - free(ht->t); - free(ht); + poison_free(ht->t); + free_per_cpu_items_count(ht->percpu_count); + poison_free(ht); return ret; } -void ht_count_nodes(struct rcu_ht *ht, +void cds_lfht_count_nodes(struct cds_lfht *ht, unsigned long *count, unsigned long *removed) { struct rcu_table *t; - struct rcu_ht_node *node, *next; - struct _rcu_ht_node *lookup; + struct cds_lfht_node *node, *next; + struct _cds_lfht_node *lookup; unsigned long nr_dummy = 0; *count = 0; @@ -697,8 +1161,8 @@ void ht_count_nodes(struct rcu_ht *ht, t = rcu_dereference(ht->t); /* Count non-dummy nodes in the table */ - lookup = &t->tbl[0][0]; - node = (struct rcu_ht_node *) lookup; + lookup = &t->tbl[0]->nodes[0]; + node = (struct cds_lfht_node *) lookup; do { next = rcu_dereference(node->p.next); if (is_removed(next)) { @@ -710,43 +1174,88 @@ void ht_count_nodes(struct rcu_ht *ht, (nr_dummy)++; node = clear_flag(next); } while (node); - dbg_printf("rculfhash: number of dummy nodes: %lu\n", nr_dummy); + dbg_printf("number of dummy nodes: %lu\n", nr_dummy); } +/* called with resize mutex held */ static -void ht_free_table_cb(struct rcu_head *head) +void _do_cds_lfht_grow(struct cds_lfht *ht, struct rcu_table *old_t, + unsigned long old_size, unsigned long new_size) { - struct rcu_table *t = - caa_container_of(head, struct rcu_table, head); - free(t); + unsigned long old_order, new_order; + struct rcu_table *new_t; + + 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); + new_t = malloc(sizeof(struct cds_lfht) + + (new_order * sizeof(struct rcu_level *))); + assert(new_size > old_size); + memcpy(&new_t->tbl, &old_t->tbl, + old_order * sizeof(struct rcu_level *)); + init_table(ht, new_t, old_order, new_order - old_order); + /* Changing table and size atomically wrt lookups */ + rcu_assign_pointer(ht->t, new_t); + ht->cds_lfht_call_rcu(&old_t->head, cds_lfht_free_table_cb); } /* called with resize mutex held */ static -void _do_ht_resize(struct rcu_ht *ht) +void _do_cds_lfht_shrink(struct cds_lfht *ht, struct rcu_table *old_t, + unsigned long old_size, unsigned long new_size) { - unsigned long new_size, old_size, old_order, new_order; - struct rcu_table *new_t, *old_t; + unsigned long old_order, new_order; + struct rcu_table *new_t; - old_t = ht->t; - old_size = old_t->size; + new_size = max(new_size, MIN_TABLE_SIZE); old_order = get_count_order_ulong(old_size) + 1; - - new_size = CMM_LOAD_SHARED(old_t->resize_target); - if (old_size == new_size) - return; new_order = get_count_order_ulong(new_size) + 1; - printf("rculfhash: resize from %lu (order %lu) to %lu (order %lu) buckets\n", + printf("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); + new_t = malloc(sizeof(struct cds_lfht) + + (new_order * sizeof(struct rcu_level *))); + assert(new_size < old_size); memcpy(&new_t->tbl, &old_t->tbl, - old_order * sizeof(struct _rcu_ht_node *)); - init_table(ht, new_t, old_order, new_order - old_order); + new_order * sizeof(struct rcu_level *)); + new_t->size = !new_order ? 1 : (1UL << (new_order - 1)); + assert(new_t->size == new_size); + new_t->resize_target = new_t->size; + new_t->resize_initiated = 0; + /* 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); + + /* + * We need to wait for all add operations to reach Q.S. (and + * thus use the new table for lookups) before we can start + * releasing the old dummy nodes. Otherwise their lookup will + * return a logically removed node as insert position. + */ + ht->cds_lfht_synchronize_rcu(); + + /* Unlink and remove all now-unused dummy node pointers. */ + fini_table(ht, old_t, new_order, old_order - new_order); + ht->cds_lfht_call_rcu(&old_t->head, cds_lfht_free_table_cb); +} + + +/* called with resize mutex held */ +static +void _do_cds_lfht_resize(struct cds_lfht *ht) +{ + unsigned long new_size, old_size; + struct rcu_table *old_t; + + old_t = ht->t; + old_size = old_t->size; + new_size = CMM_LOAD_SHARED(old_t->resize_target); + if (old_size < new_size) + _do_cds_lfht_grow(ht, old_t, old_size, new_size); + else if (old_size > new_size) + _do_cds_lfht_shrink(ht, old_t, old_size, new_size); + else + CMM_STORE_SHARED(old_t->resize_initiated, 0); } static @@ -757,18 +1266,23 @@ unsigned long resize_target_update(struct rcu_table *t, t->size << growth_order); } -void ht_resize(struct rcu_ht *ht, int growth) +static +void resize_target_update_count(struct rcu_table *t, + unsigned long count) +{ + count = max(count, MIN_TABLE_SIZE); + uatomic_set(&t->resize_target, count); +} + +void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size) { 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); - } + resize_target_update_count(t, new_size); + CMM_STORE_SHARED(t->resize_initiated, 1); + pthread_mutex_lock(&ht->resize_mutex); + _do_cds_lfht_resize(ht); + pthread_mutex_unlock(&ht->resize_mutex); } static @@ -776,18 +1290,18 @@ void do_resize_cb(struct rcu_head *head) { struct rcu_resize_work *work = caa_container_of(head, struct rcu_resize_work, head); - struct rcu_ht *ht = work->ht; + struct cds_lfht *ht = work->ht; pthread_mutex_lock(&ht->resize_mutex); - _do_ht_resize(ht); + _do_cds_lfht_resize(ht); pthread_mutex_unlock(&ht->resize_mutex); - free(work); + poison_free(work); cmm_smp_mb(); /* finish resize before decrement */ uatomic_dec(&ht->in_progress_resize); } static -void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth) +void cds_lfht_resize_lazy(struct cds_lfht *ht, struct rcu_table *t, int growth) { struct rcu_resize_work *work; unsigned long target_size; @@ -798,7 +1312,30 @@ void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth) cmm_smp_mb(); /* increment resize count before calling it */ work = malloc(sizeof(*work)); work->ht = ht; - ht->ht_call_rcu(&work->head, do_resize_cb); + ht->cds_lfht_call_rcu(&work->head, do_resize_cb); + CMM_STORE_SHARED(t->resize_initiated, 1); + } +} + +#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) + +static +void cds_lfht_resize_lazy_count(struct cds_lfht *ht, struct rcu_table *t, + unsigned long count) +{ + struct rcu_resize_work *work; + + if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) + return; + resize_target_update_count(t, count); + if (!CMM_LOAD_SHARED(t->resize_initiated)) { + uatomic_inc(&ht->in_progress_resize); + cmm_smp_mb(); /* increment resize count before calling it */ + work = malloc(sizeof(*work)); + work->ht = ht; + ht->cds_lfht_call_rcu(&work->head, do_resize_cb); CMM_STORE_SHARED(t->resize_initiated, 1); } } + +#endif