* Note: This is code uses synchronous atomic operations because that is all that x86 provides.
* Every atomic operation is also an implicit full memory barrier. The upshot is that it simplifies
* the code a bit, but it won't be as fast as it could be on platforms that provide weaker
- * operations like and unfenced CAS which would still do the job.
+ * operations like unfenced CAS which would still do the job.
+ *
+ * 11FebO9 - Bug fix in ht_iter_next() from Rui Ueyama
*/
#include <stdio.h>
#include "common.h"
#include "murmur.h"
#include "mem.h"
+#include "rcu.h"
#include "hashtable.h"
+#ifndef NBD32
#define GET_PTR(x) ((void *)((x) & MASK(48))) // low-order 48 bits is a pointer to a nstring_t
+#else
+#define GET_PTR(x) ((void *)(x))
+#endif
typedef struct entry {
map_key_t key;
volatile entry_t *table;
hashtable_t *ht; // parent ht;
struct hti *next;
+#ifdef USE_SYSTEM_MALLOC
+ void *unaligned_table_ptr; // system malloc doesn't guarentee cache-line alignment
+#endif
unsigned scale;
int max_probe;
- int references;
+ int ref_count;
int count; // TODO: make these counters distributed
int num_entries_copied;
int copy_scan;
const datatype_t *key_type;
};
-static const map_val_t COPIED_VALUE = -1;
-static const map_val_t TOMBSTONE = STRIP_TAG(-1, TAG1);
+static const map_val_t COPIED_VALUE = TAG_VALUE(DOES_NOT_EXIST, TAG1);
+static const map_val_t TOMBSTONE = STRIP_TAG(-1, TAG1);
static const unsigned ENTRIES_PER_BUCKET = CACHE_LINE_SIZE/sizeof(entry_t);
static const unsigned ENTRIES_PER_COPY_CHUNK = CACHE_LINE_SIZE/sizeof(entry_t)*2;
map_key_t ent_key = ent->key;
if (ent_key == DOES_NOT_EXIST) {
TRACE("h1", "hti_lookup: entry %p for key %p is empty", ent,
- (hti->ht->key_type == NULL) ? (void *)ent_key : GET_PTR(ent_key));
+ (hti->ht->key_type == NULL) ? (void *)key : GET_PTR(key));
*is_empty = 1; // indicate an empty so the caller avoids an expensive key compare
return ent;
}
return ent;
}
} else {
+#ifndef NBD32
// The key in <ent> is made up of two parts. The 48 low-order bits are a pointer. The
// high-order 16 bits are taken from the hash. The bits from the hash are used as a
// quick check to rule out non-equal keys without doing a complete compare.
if ((key_hash >> 16) == (ent_key >> 48)) {
+#endif
if (hti->ht->key_type->cmp(GET_PTR(ent_key), (void *)key) == 0) {
TRACE("h1", "hti_lookup: found entry %p with key %p", ent, GET_PTR(ent_key));
return ent;
+#ifndef NBD32
}
+#endif
}
}
}
static hti_t *hti_alloc (hashtable_t *parent, int scale) {
hti_t *hti = (hti_t *)nbd_malloc(sizeof(hti_t));
memset(hti, 0, sizeof(hti_t));
-
- size_t sz = sizeof(entry_t) * (1 << scale);
- entry_t *table = nbd_malloc(sz);
- memset(table, 0, sz);
- hti->table = table;
-
hti->scale = scale;
+ size_t sz = sizeof(entry_t) * (1ULL << scale);
+#ifdef USE_SYSTEM_MALLOC
+ hti->unaligned_table_ptr = nbd_malloc(sz + CACHE_LINE_SIZE - 1);
+ hti->table = (void *)(((size_t)hti->unaligned_table_ptr + CACHE_LINE_SIZE - 1) & ~(CACHE_LINE_SIZE - 1));
+#else
+ hti->table = nbd_malloc(sz);
+#endif
+ memset((void *)hti->table, 0, sz);
+
// When searching for a key probe a maximum of 1/4 of the buckets up to 1000 buckets.
- hti->max_probe = ((1 << (hti->scale - 2)) / ENTRIES_PER_BUCKET) + 4;
+ hti->max_probe = ((1ULL << (hti->scale - 2)) / ENTRIES_PER_BUCKET) + 4;
if (hti->max_probe > MAX_BUCKETS_TO_PROBE) {
hti->max_probe = MAX_BUCKETS_TO_PROBE;
}
hti->ht = parent;
+ hti->ref_count = 1; // one for the parent
assert(hti->scale >= MIN_SCALE && hti->scale < 63); // size must be a power of 2
assert(sizeof(entry_t) * ENTRIES_PER_BUCKET % CACHE_LINE_SIZE == 0); // divisible into cache
// heuristics to determine the size of the new table
size_t count = ht_count(hti->ht);
unsigned int new_scale = hti->scale;
- new_scale += (count > (1 << (new_scale - 2))); // double size if more than 1/4 full
- new_scale += (count > (1 << (new_scale - 2))); // double size again if more than 1/2 full
+ new_scale += (count > (1ULL << (new_scale - 2))); // double size if more than 1/4 full
+ new_scale += (count > (1ULL << (new_scale - 2))); // double size again if more than 1/2 full
// Allocate the new table and attempt to install it.
hti_t *next = hti_alloc(hti->ht, new_scale);
if (old_next != NULL) {
// Another thread beat us to it.
TRACE("h0", "hti_start_copy: lost race to install new hti; found %p", old_next, 0);
- nbd_free(next);
+#ifdef USE_SYSTEM_MALLOC
+ nbd_free(next->unaligned_table_ptr);
+#else
+ nbd_free((void *)next->table);
+#endif
return;
}
TRACE("h0", "hti_start_copy: new hti %p scale %llu", next, next->scale);
assert(ht1);
assert(ht1->next);
assert(ht2);
- assert(ht1_ent >= ht1->table && ht1_ent < ht1->table + (1 << ht1->scale));
+ assert(ht1_ent >= ht1->table && ht1_ent < ht1->table + (1ULL << ht1->scale));
+#ifndef NBD32
assert(key_hash == 0 || ht1->ht->key_type == NULL || (key_hash >> 16) == (ht1_ent->key >> 48));
+#endif
map_val_t ht1_ent_val = ht1_ent->val;
- if (EXPECT_FALSE(ht1_ent_val == COPIED_VALUE)) {
+ if (EXPECT_FALSE(ht1_ent_val == COPIED_VALUE || ht1_ent_val == TAG_VALUE(TOMBSTONE, TAG1))) {
TRACE("h1", "hti_copy_entry: entry %p already copied to table %p", ht1_ent, ht2);
return FALSE; // already copied
}
TRACE("h1", "hti_copy_entry: empty entry %p killed", ht1_ent, 0);
return TRUE;
}
- if (ht1_ent_val == COPIED_VALUE) {
- TRACE("h0", "hti_copy_entry: lost race to kill empty entry %p; the entry is already killed", ht1_ent, 0);
- return FALSE; // another thread beat us to it
- }
TRACE("h0", "hti_copy_entry: lost race to kill empty entry %p; the entry is not empty", ht1_ent, 0);
}
// Tag the value in the old entry to indicate a copy is in progress.
ht1_ent_val = SYNC_FETCH_AND_OR(&ht1_ent->val, TAG_VALUE(0, TAG1));
TRACE("h2", "hti_copy_entry: tagged the value %p in old entry %p", ht1_ent_val, ht1_ent);
- if (ht1_ent_val == COPIED_VALUE) {
+ if (ht1_ent_val == COPIED_VALUE || ht1_ent_val == TAG_VALUE(TOMBSTONE, TAG1)) {
TRACE("h1", "hti_copy_entry: entry %p already copied to table %p", ht1_ent, ht2);
return FALSE; // <value> was already copied by another thread.
}
+ // The old table's dead entries don't need to be copied to the new table
+ if (ht1_ent_val == TOMBSTONE)
+ return TRUE;
+
// Install the key in the new table.
map_key_t ht1_ent_key = ht1_ent->key;
map_key_t key = (ht1->ht->key_type == NULL) ? (map_key_t)ht1_ent_key : (map_key_t)GET_PTR(ht1_ent_key);
- // The old table's dead entries don't need to be copied to the new table, but their keys need to be freed.
- assert(COPIED_VALUE == TAG_VALUE(TOMBSTONE, TAG1));
- if (ht1_ent_val == TOMBSTONE) {
- TRACE("h1", "hti_copy_entry: entry %p old value was deleted, now freeing key %p", ht1_ent, key);
- if (EXPECT_FALSE(ht1->ht->key_type != NULL)) {
- nbd_defer_free((void *)key);
- }
- return TRUE;
- }
-
// We use 0 to indicate that <key_hash> is uninitiallized. Occasionally the key's hash will really be 0 and we
- // waste time recomputing it every time. It is rare enough (1 in 65k) that it won't hurt performance.
+ // waste time recomputing it every time. It is rare enough that it won't hurt performance.
if (key_hash == 0) {
- key_hash = (ht1->ht->key_type == NULL)
- ? murmur32_8b(ht1_ent_key)
- : ht1->ht->key_type->hash((void *)key);
+#ifdef NBD32
+ key_hash = (ht1->ht->key_type == NULL) ? murmur32_4b(ht1_ent_key) : ht1->ht->key_type->hash((void *)key);
+#else
+ key_hash = (ht1->ht->key_type == NULL) ? murmur32_8b(ht1_ent_key) : ht1->ht->key_type->hash((void *)key);
+#endif
}
int ht2_ent_is_empty;
// Update the count if we were the one that completed the copy.
if (old_ht2_ent_val == DOES_NOT_EXIST) {
TRACE("h0", "hti_copy_entry: key %p value %p copied to new entry", key, ht1_ent_val);
- SYNC_ADD(&ht1->count, -1);
- SYNC_ADD(&ht2->count, 1);
+ (void)SYNC_ADD(&ht1->count, -1);
+ (void)SYNC_ADD(&ht2->count, 1);
return TRUE;
}
map_key_t new_key = (hti->ht->key_type == NULL)
? (map_key_t)key
: (map_key_t)hti->ht->key_type->clone((void *)key);
+#ifndef NBD32
if (EXPECT_FALSE(hti->ht->key_type != NULL)) {
// Combine <new_key> pointer with bits from its hash
new_key = ((uint64_t)(key_hash >> 16) << 48) | new_key;
}
+#endif
// CAS the key into the table.
map_key_t old_ent_key = SYNC_CAS(&ent->key, DOES_NOT_EXIST, new_key);
// If the entry is in the middle of a copy, the copy must be completed first.
map_val_t ent_val = ent->val;
if (EXPECT_FALSE(IS_TAGGED(ent_val, TAG1))) {
- if (ent_val != COPIED_VALUE) {
- int did_copy = hti_copy_entry(hti, ent, key_hash, ((volatile hti_t *)hti)->next);
+ if (ent_val != COPIED_VALUE && ent_val != TAG_VALUE(TOMBSTONE, TAG1)) {
+ int did_copy = hti_copy_entry(hti, ent, key_hash, VOLATILE_DEREF(hti).next);
if (did_copy) {
- SYNC_ADD(&hti->num_entries_copied, 1);
+ (void)SYNC_ADD(&hti->num_entries_copied, 1);
}
TRACE("h0", "hti_cas: value in the middle of a copy, copy completed by %s",
(did_copy ? "self" : "other"), 0);
// The set succeeded. Adjust the value count.
if (old_existed && new == DOES_NOT_EXIST) {
- SYNC_ADD(&hti->count, -1);
+ (void)SYNC_ADD(&hti->count, -1);
} else if (!old_existed && new != DOES_NOT_EXIST) {
- SYNC_ADD(&hti->count, 1);
+ (void)SYNC_ADD(&hti->count, 1);
}
// Return the previous value.
// searching the table. In that case, if a copy is in progress the key
// might exist in the copy.
if (EXPECT_FALSE(ent == NULL)) {
- if (((volatile hti_t *)hti)->next != NULL)
+ if (VOLATILE_DEREF(hti).next != NULL)
return hti_get(hti->next, key, key_hash); // recursive tail-call
return DOES_NOT_EXIST;
}
// If the entry is being copied, finish the copy and retry on the next table.
map_val_t ent_val = ent->val;
if (EXPECT_FALSE(IS_TAGGED(ent_val, TAG1))) {
- if (EXPECT_FALSE(ent_val != COPIED_VALUE)) {
- int did_copy = hti_copy_entry(hti, ent, key_hash, ((volatile hti_t *)hti)->next);
+ if (EXPECT_FALSE(ent_val != COPIED_VALUE && ent_val != TAG_VALUE(TOMBSTONE, TAG1))) {
+ int did_copy = hti_copy_entry(hti, ent, key_hash, VOLATILE_DEREF(hti).next);
if (did_copy) {
- SYNC_ADD(&hti->num_entries_copied, 1);
+ (void)SYNC_ADD(&hti->num_entries_copied, 1);
}
}
- return hti_get(((volatile hti_t *)hti)->next, key, key_hash); // tail-call
+ return hti_get(VOLATILE_DEREF(hti).next, key, key_hash); // tail-call
}
return (ent_val == TOMBSTONE) ? DOES_NOT_EXIST : ent_val;
//
map_val_t ht_get (hashtable_t *ht, map_key_t key) {
+#ifdef NBD32
+ uint32_t hash = (ht->key_type == NULL) ? murmur32_4b((uint64_t)key) : ht->key_type->hash((void *)key);
+#else
uint32_t hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash((void *)key);
+#endif
return hti_get(ht->hti, key, hash);
}
// returns TRUE if copy is done
-int hti_help_copy (hti_t *hti) {
+static int hti_help_copy (hti_t *hti) {
volatile entry_t *ent;
size_t limit;
size_t total_copied = hti->num_entries_copied;
size_t x = hti->copy_scan;
TRACE("h1", "ht_cas: help copy. scan is %llu, size is %llu", x, 1<<hti->scale);
- if (total_copied != (1 << hti->scale)) {
+ if (total_copied != (1ULL << hti->scale)) {
// Panic if we've been around the array twice and still haven't finished the copy.
- int panic = (x >= (1 << (hti->scale + 1)));
+ int panic = (x >= (1ULL << (hti->scale + 1)));
if (!panic) {
limit = ENTRIES_PER_COPY_CHUNK;
TRACE("h1", "ht_cas: help copy panic", 0, 0);
// scan the whole table
ent = hti->table;
- limit = (1 << hti->scale);
+ limit = (1ULL << hti->scale);
}
// Copy the entries
for (int i = 0; i < limit; ++i) {
num_copied += hti_copy_entry(hti, ent++, 0, hti->next);
- assert(ent <= hti->table + (1 << hti->scale));
+ assert(ent <= hti->table + (1ULL << hti->scale));
}
if (num_copied != 0) {
total_copied = SYNC_ADD(&hti->num_entries_copied, num_copied);
}
}
- return (total_copied == (1 << hti->scale));
+ return (total_copied == (1ULL << hti->scale));
+}
+
+static void hti_defer_free (hti_t *hti) {
+ assert(hti->ref_count == 0);
+
+ for (uint32_t i = 0; i < (1ULL << hti->scale); ++i) {
+ map_key_t key = hti->table[i].key;
+ map_val_t val = hti->table[i].val;
+ if (val == COPIED_VALUE)
+ continue;
+ assert(!IS_TAGGED(val, TAG1) || val == TAG_VALUE(TOMBSTONE, TAG1)); // copy not in progress
+ if (hti->ht->key_type != NULL && key != DOES_NOT_EXIST) {
+ rcu_defer_free(GET_PTR(key));
+ }
+ }
+#ifdef USE_SYSTEM_MALLOC
+ rcu_defer_free(hti->unaligned_table_ptr);
+#else
+ rcu_defer_free((void *)hti->table);
+#endif
+ rcu_defer_free(hti);
+}
+
+static void hti_release (hti_t *hti) {
+ assert(hti->ref_count > 0);
+ int ref_count = SYNC_ADD(&hti->ref_count, -1);
+ if (ref_count == 0) {
+ hti_defer_free(hti);
+ }
}
//
if (EXPECT_FALSE(hti->next != NULL)) {
int done = hti_help_copy(hti);
- // Dispose of fully copied tables.
- if (done && hti->references == 0) {
-
- int r = SYNC_CAS(&hti->references, 0, -1);
- if (r == 0) {
- assert(hti->next);
- if (SYNC_CAS(&ht->hti, hti, hti->next) == hti) {
- nbd_defer_free((void *)hti->table);
- nbd_defer_free(hti);
- }
+ // Unlink fully copied tables.
+ if (done) {
+ assert(hti->next);
+ if (SYNC_CAS(&ht->hti, hti, hti->next) == hti) {
+ hti_release(hti);
}
}
}
map_val_t old_val;
+#ifdef NBD32
+ uint32_t key_hash = (ht->key_type == NULL) ? murmur32_4b((uint64_t)key) : ht->key_type->hash((void *)key);
+#else
uint32_t key_hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash((void *)key);
+#endif
while ((old_val = hti_cas(hti, key, key_hash, expected_val, new_val)) == COPIED_VALUE) {
assert(hti->next);
hti = hti->next;
map_val_t ht_remove (hashtable_t *ht, map_key_t key) {
hti_t *hti = ht->hti;
map_val_t val;
+#ifdef NBD32
+ uint32_t key_hash = (ht->key_type == NULL) ? murmur32_4b((uint64_t)key) : ht->key_type->hash((void *)key);
+#else
uint32_t key_hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash((void *)key);
+#endif
do {
val = hti_cas(hti, key, key_hash, CAS_EXPECT_WHATEVER, DOES_NOT_EXIST);
if (val != COPIED_VALUE)
void ht_free (hashtable_t *ht) {
hti_t *hti = ht->hti;
do {
- for (uint32_t i = 0; i < (1 << hti->scale); ++i) {
- assert(hti->table[i].val == COPIED_VALUE || !IS_TAGGED(hti->table[i].val, TAG1));
- if (ht->key_type != NULL && hti->table[i].key != DOES_NOT_EXIST) {
- nbd_free(GET_PTR(hti->table[i].key));
- }
- }
hti_t *next = hti->next;
- nbd_free((void *)hti->table);
- nbd_free(hti);
+ assert(hti->ref_count == 1);
+ hti_release(hti);
hti = next;
} while (hti);
nbd_free(ht);
hti_t *hti = ht->hti;
while (hti) {
printf("hti:%p scale:%u count:%d copied:%d\n", hti, hti->scale, hti->count, hti->num_entries_copied);
- for (int i = 0; i < (1 << hti->scale); ++i) {
+ for (int i = 0; i < (1ULL << hti->scale); ++i) {
volatile entry_t *ent = hti->table + i;
- printf("[0x%x] 0x%llx:0x%llx\n", i, (uint64_t)ent->key, ent->val);
+ printf("[0x%x] 0x%llx:0x%llx\n", i, (uint64_t)ent->key, (uint64_t)ent->val);
if (i > 30) {
printf("...\n");
break;
}
ht_iter_t *ht_iter_begin (hashtable_t *ht, map_key_t key) {
- hti_t *hti = ht->hti;
- int rcount;
+ hti_t *hti;
+ int ref_count;
do {
- while (((volatile hti_t *)hti)->next != NULL) {
+ hti = ht->hti;
+ while (hti->next != NULL) {
do { } while (hti_help_copy(hti) != TRUE);
hti = hti->next;
}
-
- int old = hti->references;
do {
- rcount = old;
- if (rcount != -1) {
- old = SYNC_CAS(&hti->references, rcount, rcount + 1);
- }
- } while (rcount != old);
- } while (rcount == -1);
+ ref_count = hti->ref_count;
+ if(ref_count == 0)
+ break;
+ } while (ref_count != SYNC_CAS(&hti->ref_count, ref_count, ref_count + 1));
+ } while (ref_count == 0);
ht_iter_t *iter = nbd_malloc(sizeof(ht_iter_t));
iter->hti = hti;
volatile entry_t *ent;
map_key_t key;
map_val_t val;
- size_t table_size = (1 << iter->hti->scale);
+ size_t table_size = (1ULL << iter->hti->scale);
do {
iter->idx++;
if (iter->idx == table_size) {
} while (key == DOES_NOT_EXIST || val == DOES_NOT_EXIST || val == TOMBSTONE);
- if (key_ptr) {
- *key_ptr = key;
- }
if (val == COPIED_VALUE) {
- uint32_t hash = (iter->hti->ht->key_type == NULL)
- ? murmur32_8b((uint64_t)key)
- : iter->hti->ht->key_type->hash((void *)key);
+ const datatype_t *key_type = iter->hti->ht->key_type;
+#ifdef NBD32
+ uint32_t hash = (key_type == NULL) ? murmur32_4b((uint64_t)key) : key_type->hash((void *)key);
+#else
+ uint32_t hash = (key_type == NULL) ? murmur32_8b((uint64_t)key) : key_type->hash((void *)key);
+#endif
val = hti_get(iter->hti->next, (map_key_t)ent->key, hash);
+
+ // Go to the next entry if key is already deleted.
+ if (val == DOES_NOT_EXIST)
+ return ht_iter_next(iter, key_ptr); // recursive tail-call
}
+ if (key_ptr) {
+ *key_ptr = key;
+ }
return val;
}
void ht_iter_free (ht_iter_t *iter) {
- SYNC_ADD(&iter->hti->references, -1);
+ hti_release(iter->hti);
nbd_free(iter);
}
projects / nbds / blobdiff