openssl/crypto/hashtable/hashtable.c
Tomas Mraz 8951ee06b4 ossl_ht_insert(): Allow for 4 iterations of grow_hashtable()
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Neil Horman <nhorman@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/25273)
2024-08-22 14:52:43 +02:00

761 lines
22 KiB
C

/*
* Copyright 2024 The OpenSSL Project Authors. All Rights Reserved.
*
* Licensed under the Apache License 2.0 (the "License"). You may not use
* this file except in compliance with the License. You can obtain a copy
* in the file LICENSE in the source distribution or at
* https://www.openssl.org/source/license.html
*
*
*
* Notes On hash table design and layout
* This hashtable uses a hopscotch algorithm to do indexing. The data structure
* looks as follows:
*
* hash +--------------+
* value+------->+ HT_VALUE |
* + +--------------+
* +-------+
* | |
* +---------------------------------------------------------+
* | | | | | |
* | entry | entry | entry | entry | |
* | | | | | |
* +---------------------------------------------------------+
* | | |
* | | |
* +---------------------------------------------------------+
* | + + +
* | neighborhood[0] neighborhood[1] |
* | |
* | |
* +---------------------------------------------------------+
* |
* +
* neighborhoods
*
* On lookup/insert/delete, the items key is hashed to a 64 bit value
* and the result is masked to provide an index into the neighborhoods
* table. Once a neighborhood is determined, an in-order search is done
* of the elements in the neighborhood indexes entries for a matching hash
* value, if found, the corresponding HT_VALUE is used for the respective
* operation. The number of entries in a neighborhood is determined at build
* time based on the cacheline size of the target CPU. The intent is for a
* neighborhood to have all entries in the neighborhood fit into a single cache
* line to speed up lookups. If all entries in a neighborhood are in use at the
* time of an insert, the table is expanded and rehashed.
*
* Lockless reads hash table is based on the same design but does not
* allow growing and deletion. Thus subsequent neighborhoods are always
* searched for a match until an empty entry is found.
*/
#include <string.h>
#include <internal/rcu.h>
#include <internal/hashtable.h>
#include <openssl/rand.h>
/*
* gcc defines __SANITIZE_THREAD__
* but clang uses the feature attributes api
* map the latter to the former
*/
#if defined(__clang__) && defined(__has_feature)
# if __has_feature(thread_sanitizer)
# define __SANITIZE_THREADS__
# endif
#endif
#ifdef __SANITIZE_THREADS__
# include <sanitizer/tsan_interface.h>
#endif
#include "internal/numbers.h"
/*
* When we do a lookup/insert/delete, there is a high likelihood
* that we will iterate over at least part of the neighborhood list
* As such, because we design a neighborhood entry to fit into a single
* cache line it is advantageous, when supported to fetch the entire
* structure for faster lookups
*/
#if defined(__GNUC__) || defined(__CLANG__)
# define PREFETCH_NEIGHBORHOOD(x) __builtin_prefetch(x.entries)
# define PREFETCH(x) __builtin_prefetch(x)
#else
# define PREFETCH_NEIGHBORHOOD(x)
# define PREFETCH(x)
#endif
static ossl_unused uint64_t fnv1a_hash(uint8_t *key, size_t len)
{
uint64_t hash = 0xcbf29ce484222325ULL;
size_t i;
for (i = 0; i < len; i++) {
hash ^= key[i];
hash *= 0x00000100000001B3ULL;
}
return hash;
}
/*
* Define our neighborhood list length
* Note: It should always be a power of 2
*/
#define DEFAULT_NEIGH_LEN_LOG 4
#define DEFAULT_NEIGH_LEN (1 << DEFAULT_NEIGH_LEN_LOG)
/*
* For now assume cache line size is 64 bytes
*/
#define CACHE_LINE_BYTES 64
#define CACHE_LINE_ALIGNMENT CACHE_LINE_BYTES
#define NEIGHBORHOOD_LEN (CACHE_LINE_BYTES / sizeof(struct ht_neighborhood_entry_st))
/*
* Defines our chains of values
*/
struct ht_internal_value_st {
HT_VALUE value;
HT *ht;
};
struct ht_neighborhood_entry_st {
uint64_t hash;
struct ht_internal_value_st *value;
};
struct ht_neighborhood_st {
struct ht_neighborhood_entry_st entries[NEIGHBORHOOD_LEN];
};
/*
* Updates to data in this struct
* require an rcu sync after modification
* prior to free
*/
struct ht_mutable_data_st {
struct ht_neighborhood_st *neighborhoods;
void *neighborhood_ptr_to_free;
uint64_t neighborhood_mask;
};
/*
* Private data may be updated on the write
* side only, and so do not require rcu sync
*/
struct ht_write_private_data_st {
size_t neighborhood_len;
size_t value_count;
int need_sync;
};
struct ht_internal_st {
HT_CONFIG config;
CRYPTO_RCU_LOCK *lock;
CRYPTO_RWLOCK *atomic_lock;
struct ht_mutable_data_st *md;
struct ht_write_private_data_st wpd;
};
static void free_value(struct ht_internal_value_st *v);
static struct ht_neighborhood_st *alloc_new_neighborhood_list(size_t len,
void **freeptr)
{
struct ht_neighborhood_st *ret;
ret = OPENSSL_aligned_alloc(sizeof(struct ht_neighborhood_st) * len,
CACHE_LINE_BYTES, freeptr);
/* fall back to regular malloc */
if (ret == NULL) {
ret = *freeptr = OPENSSL_malloc(sizeof(struct ht_neighborhood_st) * len);
if (ret == NULL)
return NULL;
}
memset(ret, 0, sizeof(struct ht_neighborhood_st) * len);
return ret;
}
static void internal_free_nop(HT_VALUE *v)
{
return;
}
HT *ossl_ht_new(const HT_CONFIG *conf)
{
HT *new = OPENSSL_zalloc(sizeof(*new));
if (new == NULL)
return NULL;
new->atomic_lock = CRYPTO_THREAD_lock_new();
if (new->atomic_lock == NULL)
goto err;
memcpy(&new->config, conf, sizeof(*conf));
if (new->config.init_neighborhoods != 0) {
new->wpd.neighborhood_len = new->config.init_neighborhoods;
/* round up to the next power of 2 */
new->wpd.neighborhood_len--;
new->wpd.neighborhood_len |= new->wpd.neighborhood_len >> 1;
new->wpd.neighborhood_len |= new->wpd.neighborhood_len >> 2;
new->wpd.neighborhood_len |= new->wpd.neighborhood_len >> 4;
new->wpd.neighborhood_len |= new->wpd.neighborhood_len >> 8;
new->wpd.neighborhood_len |= new->wpd.neighborhood_len >> 16;
new->wpd.neighborhood_len++;
} else {
new->wpd.neighborhood_len = DEFAULT_NEIGH_LEN;
}
if (new->config.ht_free_fn == NULL)
new->config.ht_free_fn = internal_free_nop;
new->md = OPENSSL_zalloc(sizeof(*new->md));
if (new->md == NULL)
goto err;
new->md->neighborhoods =
alloc_new_neighborhood_list(new->wpd.neighborhood_len,
&new->md->neighborhood_ptr_to_free);
if (new->md->neighborhoods == NULL)
goto err;
new->md->neighborhood_mask = new->wpd.neighborhood_len - 1;
new->lock = ossl_rcu_lock_new(1, conf->ctx);
if (new->lock == NULL)
goto err;
if (new->config.ht_hash_fn == NULL)
new->config.ht_hash_fn = fnv1a_hash;
return new;
err:
CRYPTO_THREAD_lock_free(new->atomic_lock);
ossl_rcu_lock_free(new->lock);
if (new->md != NULL)
OPENSSL_free(new->md->neighborhood_ptr_to_free);
OPENSSL_free(new->md);
OPENSSL_free(new);
return NULL;
}
void ossl_ht_read_lock(HT *htable)
{
ossl_rcu_read_lock(htable->lock);
}
void ossl_ht_read_unlock(HT *htable)
{
ossl_rcu_read_unlock(htable->lock);
}
void ossl_ht_write_lock(HT *htable)
{
ossl_rcu_write_lock(htable->lock);
htable->wpd.need_sync = 0;
}
void ossl_ht_write_unlock(HT *htable)
{
int need_sync = htable->wpd.need_sync;
htable->wpd.need_sync = 0;
ossl_rcu_write_unlock(htable->lock);
if (need_sync)
ossl_synchronize_rcu(htable->lock);
}
static void free_oldmd(void *arg)
{
struct ht_mutable_data_st *oldmd = arg;
size_t i, j;
size_t neighborhood_len = (size_t)oldmd->neighborhood_mask + 1;
struct ht_internal_value_st *v;
for (i = 0; i < neighborhood_len; i++) {
PREFETCH_NEIGHBORHOOD(oldmd->neighborhoods[i + 1]);
for (j = 0; j < NEIGHBORHOOD_LEN; j++) {
if (oldmd->neighborhoods[i].entries[j].value != NULL) {
v = oldmd->neighborhoods[i].entries[j].value;
v->ht->config.ht_free_fn((HT_VALUE *)v);
free_value(v);
}
}
}
OPENSSL_free(oldmd->neighborhood_ptr_to_free);
OPENSSL_free(oldmd);
}
static int ossl_ht_flush_internal(HT *h)
{
struct ht_mutable_data_st *newmd = NULL;
struct ht_mutable_data_st *oldmd = NULL;
newmd = OPENSSL_zalloc(sizeof(*newmd));
if (newmd == NULL)
return 0;
newmd->neighborhoods = alloc_new_neighborhood_list(DEFAULT_NEIGH_LEN,
&newmd->neighborhood_ptr_to_free);
if (newmd->neighborhoods == NULL) {
OPENSSL_free(newmd);
return 0;
}
newmd->neighborhood_mask = DEFAULT_NEIGH_LEN - 1;
/* Swap the old and new mutable data sets */
oldmd = ossl_rcu_deref(&h->md);
ossl_rcu_assign_ptr(&h->md, &newmd);
/* Set the number of entries to 0 */
h->wpd.value_count = 0;
h->wpd.neighborhood_len = DEFAULT_NEIGH_LEN;
ossl_rcu_call(h->lock, free_oldmd, oldmd);
h->wpd.need_sync = 1;
return 1;
}
int ossl_ht_flush(HT *h)
{
return ossl_ht_flush_internal(h);
}
void ossl_ht_free(HT *h)
{
if (h == NULL)
return;
ossl_ht_write_lock(h);
ossl_ht_flush_internal(h);
ossl_ht_write_unlock(h);
/* Freeing the lock does a final sync for us */
CRYPTO_THREAD_lock_free(h->atomic_lock);
ossl_rcu_lock_free(h->lock);
OPENSSL_free(h->md->neighborhood_ptr_to_free);
OPENSSL_free(h->md);
OPENSSL_free(h);
return;
}
size_t ossl_ht_count(HT *h)
{
size_t count;
count = h->wpd.value_count;
return count;
}
void ossl_ht_foreach_until(HT *h, int (*cb)(HT_VALUE *obj, void *arg),
void *arg)
{
size_t i, j;
struct ht_mutable_data_st *md;
md = ossl_rcu_deref(&h->md);
for (i = 0; i < md->neighborhood_mask + 1; i++) {
PREFETCH_NEIGHBORHOOD(md->neighborhoods[i + 1]);
for (j = 0; j < NEIGHBORHOOD_LEN; j++) {
if (md->neighborhoods[i].entries[j].value != NULL) {
if (!cb((HT_VALUE *)md->neighborhoods[i].entries[j].value, arg))
goto out;
}
}
}
out:
return;
}
HT_VALUE_LIST *ossl_ht_filter(HT *h, size_t max_len,
int (*filter)(HT_VALUE *obj, void *arg),
void *arg)
{
struct ht_mutable_data_st *md;
HT_VALUE_LIST *list = OPENSSL_zalloc(sizeof(HT_VALUE_LIST)
+ (sizeof(HT_VALUE *) * max_len));
size_t i, j;
struct ht_internal_value_st *v;
if (list == NULL)
return NULL;
/*
* The list array lives just beyond the end of
* the struct
*/
list->list = (HT_VALUE **)(list + 1);
md = ossl_rcu_deref(&h->md);
for (i = 0; i < md->neighborhood_mask + 1; i++) {
PREFETCH_NEIGHBORHOOD(md->neighborhoods[i+1]);
for (j = 0; j < NEIGHBORHOOD_LEN; j++) {
v = md->neighborhoods[i].entries[j].value;
if (v != NULL && filter((HT_VALUE *)v, arg)) {
list->list[list->list_len++] = (HT_VALUE *)v;
if (list->list_len == max_len)
goto out;
}
}
}
out:
return list;
}
void ossl_ht_value_list_free(HT_VALUE_LIST *list)
{
OPENSSL_free(list);
}
static int compare_hash(uint64_t hash1, uint64_t hash2)
{
return (hash1 == hash2);
}
static void free_old_neigh_table(void *arg)
{
struct ht_mutable_data_st *oldmd = arg;
OPENSSL_free(oldmd->neighborhood_ptr_to_free);
OPENSSL_free(oldmd);
}
/*
* Increase hash table bucket list
* must be called with write_lock held
*/
static int grow_hashtable(HT *h, size_t oldsize)
{
struct ht_mutable_data_st *newmd;
struct ht_mutable_data_st *oldmd = ossl_rcu_deref(&h->md);
int rc = 0;
uint64_t oldi, oldj, newi, newj;
uint64_t oldhash;
struct ht_internal_value_st *oldv;
int rehashed;
size_t newsize = oldsize * 2;
if (h->config.lockless_reads)
goto out;
if ((newmd = OPENSSL_zalloc(sizeof(*newmd))) == NULL)
goto out;
/* bucket list is always a power of 2 */
newmd->neighborhoods = alloc_new_neighborhood_list(oldsize * 2,
&newmd->neighborhood_ptr_to_free);
if (newmd->neighborhoods == NULL)
goto out_free;
/* being a power of 2 makes for easy mask computation */
newmd->neighborhood_mask = (newsize - 1);
/*
* Now we need to start rehashing entries
* Note we don't need to use atomics here as the new
* mutable data hasn't been published
*/
for (oldi = 0; oldi < h->wpd.neighborhood_len; oldi++) {
PREFETCH_NEIGHBORHOOD(oldmd->neighborhoods[oldi + 1]);
for (oldj = 0; oldj < NEIGHBORHOOD_LEN; oldj++) {
oldv = oldmd->neighborhoods[oldi].entries[oldj].value;
if (oldv == NULL)
continue;
oldhash = oldmd->neighborhoods[oldi].entries[oldj].hash;
newi = oldhash & newmd->neighborhood_mask;
rehashed = 0;
for (newj = 0; newj < NEIGHBORHOOD_LEN; newj++) {
if (newmd->neighborhoods[newi].entries[newj].value == NULL) {
newmd->neighborhoods[newi].entries[newj].value = oldv;
newmd->neighborhoods[newi].entries[newj].hash = oldhash;
rehashed = 1;
break;
}
}
if (rehashed == 0) {
/* we ran out of space in a neighborhood, grow again */
OPENSSL_free(newmd->neighborhoods);
OPENSSL_free(newmd);
return grow_hashtable(h, newsize);
}
}
}
/*
* Now that our entries are all hashed into the new bucket list
* update our bucket_len and target_max_load
*/
h->wpd.neighborhood_len = newsize;
/*
* Now we replace the old mutable data with the new
*/
ossl_rcu_assign_ptr(&h->md, &newmd);
ossl_rcu_call(h->lock, free_old_neigh_table, oldmd);
h->wpd.need_sync = 1;
/*
* And we're done
*/
rc = 1;
out:
return rc;
out_free:
OPENSSL_free(newmd->neighborhoods);
OPENSSL_free(newmd);
goto out;
}
static void free_old_ht_value(void *arg)
{
HT_VALUE *h = (HT_VALUE *)arg;
/*
* Note, this is only called on replacement,
* the caller is responsible for freeing the
* held data, we just need to free the wrapping
* struct here
*/
OPENSSL_free(h);
}
static ossl_inline int match_key(HT_KEY *a, HT_KEY *b)
{
/*
* keys match if they are both present, the same size
* and compare equal in memory
*/
PREFETCH(a->keybuf);
PREFETCH(b->keybuf);
if (a->keybuf != NULL && b->keybuf != NULL && a->keysize == b->keysize)
return !memcmp(a->keybuf, b->keybuf, a->keysize);
return 1;
}
static int ossl_ht_insert_locked(HT *h, uint64_t hash,
struct ht_internal_value_st *newval,
HT_VALUE **olddata)
{
struct ht_mutable_data_st *md = h->md;
uint64_t neigh_idx_start = hash & md->neighborhood_mask;
uint64_t neigh_idx = neigh_idx_start;
size_t j;
uint64_t ihash;
HT_VALUE *ival;
size_t empty_idx = SIZE_MAX;
int lockless_reads = h->config.lockless_reads;
do {
PREFETCH_NEIGHBORHOOD(md->neighborhoods[neigh_idx]);
for (j = 0; j < NEIGHBORHOOD_LEN; j++) {
ival = ossl_rcu_deref(&md->neighborhoods[neigh_idx].entries[j].value);
if (ival == NULL) {
empty_idx = j;
/* lockless_reads implies no deletion, we can break out */
if (lockless_reads)
goto not_found;
continue;
}
if (!CRYPTO_atomic_load(&md->neighborhoods[neigh_idx].entries[j].hash,
&ihash, h->atomic_lock))
return 0;
if (compare_hash(hash, ihash) && match_key(&newval->value.key,
&ival->key)) {
if (olddata == NULL) {
/* This would insert a duplicate -> fail */
return 0;
}
/* Do a replacement */
if (!CRYPTO_atomic_store(&md->neighborhoods[neigh_idx].entries[j].hash,
hash, h->atomic_lock))
return 0;
*olddata = (HT_VALUE *)md->neighborhoods[neigh_idx].entries[j].value;
ossl_rcu_assign_ptr(&md->neighborhoods[neigh_idx].entries[j].value,
&newval);
ossl_rcu_call(h->lock, free_old_ht_value, *olddata);
h->wpd.need_sync = 1;
return 1;
}
}
if (!lockless_reads)
break;
/* Continue search in subsequent neighborhoods */
neigh_idx = (neigh_idx + 1) & md->neighborhood_mask;
} while (neigh_idx != neigh_idx_start);
not_found:
/* If we get to here, its just an insert */
if (empty_idx == SIZE_MAX)
return -1; /* out of space */
if (!CRYPTO_atomic_store(&md->neighborhoods[neigh_idx].entries[empty_idx].hash,
hash, h->atomic_lock))
return 0;
h->wpd.value_count++;
ossl_rcu_assign_ptr(&md->neighborhoods[neigh_idx].entries[empty_idx].value,
&newval);
return 1;
}
static struct ht_internal_value_st *alloc_new_value(HT *h, HT_KEY *key,
void *data,
uintptr_t *type)
{
struct ht_internal_value_st *tmp;
size_t nvsize = sizeof(*tmp);
if (h->config.collision_check == 1)
nvsize += key->keysize;
tmp = OPENSSL_malloc(nvsize);
if (tmp == NULL)
return NULL;
tmp->ht = h;
tmp->value.value = data;
tmp->value.type_id = type;
tmp->value.key.keybuf = NULL;
if (h->config.collision_check) {
tmp->value.key.keybuf = (uint8_t *)(tmp + 1);
tmp->value.key.keysize = key->keysize;
memcpy(tmp->value.key.keybuf, key->keybuf, key->keysize);
}
return tmp;
}
static void free_value(struct ht_internal_value_st *v)
{
OPENSSL_free(v);
}
int ossl_ht_insert(HT *h, HT_KEY *key, HT_VALUE *data, HT_VALUE **olddata)
{
struct ht_internal_value_st *newval = NULL;
uint64_t hash;
int rc = 0;
int i;
if (data->value == NULL)
goto out;
newval = alloc_new_value(h, key, data->value, data->type_id);
if (newval == NULL)
goto out;
/*
* we have to take our lock here to prevent other changes
* to the bucket list
*/
hash = h->config.ht_hash_fn(key->keybuf, key->keysize);
for (i = 0;
(rc = ossl_ht_insert_locked(h, hash, newval, olddata)) == -1
&& i < 4;
++i)
if (!grow_hashtable(h, h->wpd.neighborhood_len)) {
rc = -1;
break;
}
if (rc <= 0)
free_value(newval);
out:
return rc;
}
HT_VALUE *ossl_ht_get(HT *h, HT_KEY *key)
{
struct ht_mutable_data_st *md;
uint64_t hash;
uint64_t neigh_idx_start;
uint64_t neigh_idx;
struct ht_internal_value_st *ival = NULL;
size_t j;
uint64_t ehash;
int lockless_reads = h->config.lockless_reads;
hash = h->config.ht_hash_fn(key->keybuf, key->keysize);
md = ossl_rcu_deref(&h->md);
neigh_idx = neigh_idx_start = hash & md->neighborhood_mask;
do {
PREFETCH_NEIGHBORHOOD(md->neighborhoods[neigh_idx]);
for (j = 0; j < NEIGHBORHOOD_LEN; j++) {
ival = ossl_rcu_deref(&md->neighborhoods[neigh_idx].entries[j].value);
if (ival == NULL) {
if (lockless_reads)
/* lockless_reads implies no deletion, we can break out */
return NULL;
continue;
}
if (!CRYPTO_atomic_load(&md->neighborhoods[neigh_idx].entries[j].hash,
&ehash, h->atomic_lock))
return NULL;
if (compare_hash(hash, ehash) && match_key(&ival->value.key, key))
return (HT_VALUE *)ival;
}
if (!lockless_reads)
break;
/* Continue search in subsequent neighborhoods */
neigh_idx = (neigh_idx + 1) & md->neighborhood_mask;
} while (neigh_idx != neigh_idx_start);
return NULL;
}
static void free_old_entry(void *arg)
{
struct ht_internal_value_st *v = arg;
v->ht->config.ht_free_fn((HT_VALUE *)v);
free_value(v);
}
int ossl_ht_delete(HT *h, HT_KEY *key)
{
uint64_t hash;
uint64_t neigh_idx;
size_t j;
struct ht_internal_value_st *v = NULL;
HT_VALUE *nv = NULL;
int rc = 0;
if (h->config.lockless_reads)
return 0;
hash = h->config.ht_hash_fn(key->keybuf, key->keysize);
neigh_idx = hash & h->md->neighborhood_mask;
PREFETCH_NEIGHBORHOOD(h->md->neighborhoods[neigh_idx]);
for (j = 0; j < NEIGHBORHOOD_LEN; j++) {
v = (struct ht_internal_value_st *)h->md->neighborhoods[neigh_idx].entries[j].value;
if (v == NULL)
continue;
if (compare_hash(hash, h->md->neighborhoods[neigh_idx].entries[j].hash)
&& match_key(key, &v->value.key)) {
if (!CRYPTO_atomic_store(&h->md->neighborhoods[neigh_idx].entries[j].hash,
0, h->atomic_lock))
break;
h->wpd.value_count--;
ossl_rcu_assign_ptr(&h->md->neighborhoods[neigh_idx].entries[j].value,
&nv);
rc = 1;
break;
}
}
if (rc == 1) {
ossl_rcu_call(h->lock, free_old_entry, v);
h->wpd.need_sync = 1;
}
return rc;
}