#include "murmur.h"
#include "mem.h"
#include "struct.h"
+#include "nstring.h"
-#define GET_PTR(x) ((string_t *)((x) & MASK(48))) // low-order 48 bits is a pointer to a string_t
+#define GET_PTR(x) ((nstring_t *)((x) & MASK(48))) // low-order 48 bits is a pointer to a nstring_t
typedef struct ht_entry {
- uint64_t key;
+ uint64_t key; // ptr to nstring_t
uint64_t value;
} entry_t;
-typedef struct string {
- uint32_t len;
- char val[];
-} string_t;
-
-typedef struct hash_table_i {
+typedef struct hti {
volatile entry_t *table;
- hash_table_t *ht; // parent ht;
- struct hash_table_i *next;
- struct hash_table_i *next_free;
+ hashtable_t *ht; // parent ht;
+ struct hti *next;
+ struct hti *next_free;
unsigned int scale;
int max_probe;
int count; // TODO: make these counters distributed
int num_entries_copied;
int scan;
-} hash_table_i_t;
+} hashtable_i_t;
+
+struct ht {
+ hashtable_i_t *hti;
+};
static const uint64_t COPIED_VALUE = -1;
static const uint64_t TOMBSTONE = STRIP_TAG(-1);
static const unsigned MIN_SCALE = 4; // min 16 entries (4 buckets)
static const unsigned MAX_BUCKETS_TO_PROBE = 250;
-static int hti_copy_entry
- (hash_table_i_t *ht1, volatile entry_t *e, uint32_t e_key_hash, hash_table_i_t *ht2);
+static int hti_copy_entry (hashtable_i_t *ht1, volatile entry_t *e, uint32_t e_key_hash, hashtable_i_t *ht2);
// Choose the next bucket to probe using the high-order bits of <key_hash>.
static inline int get_next_ndx(int old_ndx, uint32_t key_hash, int ht_scale) {
static inline int ht_key_equals (uint64_t a, uint32_t b_hash, const char *b_value, uint32_t b_len) {
if ((b_hash >> 16) != (a >> 48)) // high-order 16 bits are from the hash value
return FALSE;
- const string_t *a_key = GET_PTR(a);
- assert(a_key);
- return a_key->len == b_len && memcmp(a_key->val, b_value, b_len) == 0;
+ return ns_cmp_raw(GET_PTR(a), b_value, b_len) == 0;
}
// Lookup <key> in <hti>.
//
// Record if the entry being returned is empty. Otherwise the caller will have to waste time with
// ht_key_equals() to confirm that it did not lose a race to fill an empty entry.
-static volatile entry_t *hti_lookup (hash_table_i_t *hti, uint32_t key_hash, const char *key_val, uint32_t key_len, int *is_empty) {
- TRACE("h2", "hti_lookup(key %p in hti %p)", key_val, hti);
+static volatile entry_t *hti_lookup (hashtable_i_t *hti, uint32_t key_hash, const char *key_data, uint32_t key_len, int *is_empty) {
+ TRACE("h2", "hti_lookup(key %p in hti %p)", key_data, hti);
*is_empty = 0;
// Probe one cache line at a time
uint64_t e_key = e->key;
if (e_key == DOES_NOT_EXIST) {
- TRACE("h1", "hti_lookup: entry %p for key \"%s\" is empty", e, GET_PTR(e_key)->val);
+ TRACE("h1", "hti_lookup: entry %p for key \"%s\" is empty", e, ns_data(GET_PTR(e_key)));
*is_empty = 1; // indicate an empty so the caller avoids an expensive ht_key_equals
return e;
}
- if (ht_key_equals(e_key, key_hash, key_val, key_len)) {
- TRACE("h1", "hti_lookup: entry %p key \"%s\"", e, GET_PTR(e_key)->val);
- TRACE("h2", "hti_lookup: entry key len %llu, value %p", GET_PTR(e_key)->len, e->value);
+ if (ht_key_equals(e_key, key_hash, key_data, key_len)) {
+ TRACE("h1", "hti_lookup: entry %p key \"%s\"", e, ns_data(GET_PTR(e_key)));
+ TRACE("h2", "hti_lookup: entry key len %llu, value %p", ns_len(GET_PTR(e_key)), e->value);
return e;
}
}
return NULL;
}
-// Allocate and initialize a hash_table_i_t with 2^<scale> entries.
-static hash_table_i_t *hti_alloc (hash_table_t *parent, int scale) {
+// Allocate and initialize a hashtable_i_t with 2^<scale> entries.
+static hashtable_i_t *hti_alloc (hashtable_t *parent, int scale) {
// Include enough slop to align the actual table on a cache line boundry
- size_t n = sizeof(hash_table_i_t)
+ size_t n = sizeof(hashtable_i_t)
+ sizeof(entry_t) * (1 << scale)
+ (CACHE_LINE_SIZE - 1);
- hash_table_i_t *hti = (hash_table_i_t *)calloc(n, 1);
+ hashtable_i_t *hti = (hashtable_i_t *)calloc(n, 1);
// Align the table of hash entries on a cache line boundry.
- hti->table = (entry_t *)(((uint64_t)hti + sizeof(hash_table_i_t) + (CACHE_LINE_SIZE-1))
+ hti->table = (entry_t *)(((uint64_t)hti + sizeof(hashtable_i_t) + (CACHE_LINE_SIZE-1))
& ~(CACHE_LINE_SIZE-1));
hti->scale = scale;
// Called when <hti> runs out of room for new keys.
//
-// Initiates a copy by creating a larger hash_table_i_t and installing it in <hti->next>.
-static void hti_start_copy (hash_table_i_t *hti) {
+// Initiates a copy by creating a larger hashtable_i_t and installing it in <hti->next>.
+static void hti_start_copy (hashtable_i_t *hti) {
TRACE("h0", "hti_start_copy(hti %p scale %llu)", hti, hti->scale);
// heuristics to determine the size of the new table
new_scale += (count > (1 << (new_scale - 2))); // double size again if more than 1/2 full
// Allocate the new table and attempt to install it.
- hash_table_i_t *next = hti_alloc(hti->ht, new_scale);
- hash_table_i_t *old_next = SYNC_CAS(&hti->next, NULL, next);
+ hashtable_i_t *next = hti_alloc(hti->ht, new_scale);
+ hashtable_i_t *old_next = SYNC_CAS(&hti->next, NULL, next);
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);
//
// Return 1 unless <ht1_e> is already copied (then return 0), so the caller can account for the total
// number of entries left to copy.
-static int hti_copy_entry (hash_table_i_t *ht1, volatile entry_t *ht1_e, uint32_t key_hash,
- hash_table_i_t *ht2) {
+static int hti_copy_entry (hashtable_i_t *ht1, volatile entry_t *ht1_e, uint32_t key_hash,
+ hashtable_i_t *ht2) {
TRACE("h2", "hti_copy_entry: entry %p to table %p", ht1_e, ht2);
assert(ht1);
assert(ht1->next);
// to be freed.
assert(COPIED_VALUE == TAG_VALUE(TOMBSTONE));
if (ht1_e_value == TOMBSTONE) {
- TRACE("h1", "hti_copy_entry: entry %p old value was deleted, now freeing key %p", ht1_e,
- GET_PTR(ht1_e->key));
+ TRACE("h1", "hti_copy_entry: entry %p old value was deleted, now freeing key %p", ht1_e, GET_PTR(ht1_e->key));
nbd_defer_free(GET_PTR(ht1_e->key));
return TRUE;
}
// Install the key in the new table.
uint64_t key = ht1_e->key;
- string_t *key_string = GET_PTR(key);
+ nstring_t *key_string = GET_PTR(key);
uint64_t value = STRIP_TAG(ht1_e_value);
// We use 0 to indicate that <key_hash> isn't initiallized. Occasionally the <key_hash> will
// really be 0 and we will waste time recomputing it. That is rare enough that it is OK.
if (key_hash == 0) {
- key_hash = murmur32(key_string->val, key_string->len);
+ key_hash = murmur32(ns_data(key_string), ns_len(key_string));
}
int is_empty;
- volatile entry_t *ht2_e = hti_lookup(ht2, key_hash, key_string->val, key_string->len, &is_empty);
+ volatile entry_t *ht2_e = hti_lookup(ht2, key_hash, ns_data(key_string), ns_len(key_string), &is_empty);
TRACE("h0", "hti_copy_entry: copy entry %p to entry %p", ht1_e, ht2_e);
// it is possible that there is not any room in the new table either
// Update the count if we were the one that completed the copy.
if (old_ht2_e_value == DOES_NOT_EXIST) {
- TRACE("h0", "hti_copy_entry: key \"%s\" value %p copied to new entry", key_string->val, value);
+ TRACE("h0", "hti_copy_entry: key \"%s\" value %p copied to new entry", ns_data(key_string), value);
SYNC_ADD(&ht1->count, -1);
SYNC_ADD(&ht2->count, 1);
return TRUE;
//
// NOTE: the returned value matches <expected> iff the set succeeds
//
-// Certain values of <expected> have special meaning. If <expected> is HT_EXPECT_EXISTS then any
+// Certain values of <expected> have special meaning. If <expected> is EXPECT_EXISTS then any
// real value matches (i.e. not a TOMBSTONE or DOES_NOT_EXIST) as long as <key> is in the table. If
-// <expected> is HT_EXPECT_WHATEVER then skip the test entirely.
+// <expected> is EXPECT_WHATEVER then skip the test entirely.
//
-static uint64_t hti_compare_and_set (hash_table_i_t *hti, uint32_t key_hash, const char *key_val,
+static uint64_t hti_compare_and_set (hashtable_i_t *hti, uint32_t key_hash, const char *key_data,
uint32_t key_len, uint64_t expected, uint64_t new) {
- TRACE("h1", "hti_compare_and_set: hti %p key %p", hti, key_val);
+ TRACE("h1", "hti_compare_and_set: hti %p key %p", hti, key_data);
TRACE("h1", "hti_compare_and_set: value %p expect %p", new, expected);
assert(hti);
assert(new != DOES_NOT_EXIST && !IS_TAGGED(new));
- assert(key_val);
+ assert(key_data);
int is_empty;
- volatile entry_t *e = hti_lookup(hti, key_hash, key_val, key_len, &is_empty);
+ volatile entry_t *e = hti_lookup(hti, key_hash, key_data, key_len, &is_empty);
// There is no room for <key>, grow the table and try again.
if (e == NULL) {
// Install <key> in the table if it doesn't exist.
if (is_empty) {
TRACE("h0", "hti_compare_and_set: entry %p is empty", e, 0);
- if (expected != HT_EXPECT_WHATEVER && expected != HT_EXPECT_NOT_EXISTS)
+ if (expected != EXPECT_WHATEVER && expected != EXPECT_DOES_NOT_EXIST)
return DOES_NOT_EXIST;
// No need to do anything, <key> is already deleted.
return DOES_NOT_EXIST;
// Allocate <key>.
- string_t *key = nbd_malloc(sizeof(uint32_t) + key_len);
- key->len = key_len;
- memcpy(key->val, key_val, key_len);
+ nstring_t *key = ns_alloc(key_data, key_len);
// Combine <key> pointer with bits from its hash, CAS it into the table.
uint64_t temp = ((uint64_t)(key_hash >> 16) << 48) | (uint64_t)key;
TRACE("h0", "hti_compare_and_set: lost race to install key %p in entry %p", key, e);
TRACE("h0", "hti_compare_and_set: found %p instead of NULL", GET_PTR(e_key), 0);
nbd_free(key);
- return hti_compare_and_set(hti, key_hash, key_val, key_len, expected, new); // tail-call
+ return hti_compare_and_set(hti, key_hash, key_data, key_len, expected, new); // tail-call
}
TRACE("h2", "hti_compare_and_set: installed key %p in entry %p", key, e);
}
- TRACE("h0", "hti_compare_and_set: entry for key \"%s\" is %p", GET_PTR(e->key)->val, e);
+ TRACE("h0", "hti_compare_and_set: entry for key \"%s\" is %p", ns_data(GET_PTR(e->key)), e);
// If the entry is in the middle of a copy, the copy must be completed first.
uint64_t e_value = e->value;
if (EXPECT_FALSE(IS_TAGGED(e_value))) {
if (e_value != COPIED_VALUE) {
- int did_copy = hti_copy_entry(hti, e, key_hash, ((volatile hash_table_i_t *)hti)->next);
+ int did_copy = hti_copy_entry(hti, e, key_hash, ((volatile hashtable_i_t *)hti)->next);
if (did_copy) {
SYNC_ADD(&hti->num_entries_copied, 1);
}
// Fail if the old value is not consistent with the caller's expectation.
int old_existed = (e_value != TOMBSTONE && e_value != DOES_NOT_EXIST);
- if (EXPECT_FALSE(expected != HT_EXPECT_WHATEVER && expected != e_value)) {
- if (EXPECT_FALSE(expected != (old_existed ? HT_EXPECT_EXISTS : HT_EXPECT_NOT_EXISTS))) {
+ if (EXPECT_FALSE(expected != EXPECT_WHATEVER && expected != e_value)) {
+ if (EXPECT_FALSE(expected != (old_existed ? EXPECT_EXISTS : EXPECT_DOES_NOT_EXIST))) {
TRACE("h1", "hti_compare_and_set: value %p expected by caller not found; found value %p",
expected, e_value);
return e_value;
uint64_t v = SYNC_CAS(&e->value, e_value, new);
if (EXPECT_FALSE(v != e_value)) {
TRACE("h0", "hti_compare_and_set: value CAS failed; expected %p found %p", e_value, v);
- return hti_compare_and_set(hti, key_hash, key_val, key_len, expected, new); // recursive tail-call
+ return hti_compare_and_set(hti, key_hash, key_data, key_len, expected, new); // recursive tail-call
}
// The set succeeded. Adjust the value count.
}
//
-static uint64_t hti_get (hash_table_i_t *hti, uint32_t key_hash, const char *key_val, uint32_t key_len) {
- assert(key_val);
+static uint64_t hti_get (hashtable_i_t *hti, uint32_t key_hash, const char *key_data, uint32_t key_len) {
+ assert(key_data);
int is_empty;
- volatile entry_t *e = hti_lookup(hti, key_hash, key_val, key_len, &is_empty);
+ volatile entry_t *e = hti_lookup(hti, key_hash, key_data, key_len, &is_empty);
// When hti_lookup() returns NULL it means we hit the reprobe limit while
// searching the table. In that case, if a copy is in progress the key
// might exist in the copy.
if (EXPECT_FALSE(e == NULL)) {
- if (((volatile hash_table_i_t *)hti)->next != NULL)
- return hti_get(hti->next, key_hash, key_val, key_len); // recursive tail-call
+ if (((volatile hashtable_i_t *)hti)->next != NULL)
+ return hti_get(hti->next, key_hash, key_data, key_len); // recursive tail-call
return DOES_NOT_EXIST;
}
uint64_t e_value = e->value;
if (EXPECT_FALSE(IS_TAGGED(e_value))) {
if (EXPECT_FALSE(e_value != COPIED_VALUE)) {
- int did_copy = hti_copy_entry(hti, e, key_hash, ((volatile hash_table_i_t *)hti)->next);
+ int did_copy = hti_copy_entry(hti, e, key_hash, ((volatile hashtable_i_t *)hti)->next);
if (did_copy) {
SYNC_ADD(&hti->num_entries_copied, 1);
}
}
- return hti_get(((volatile hash_table_i_t *)hti)->next, key_hash, key_val, key_len); // tail-call
+ return hti_get(((volatile hashtable_i_t *)hti)->next, key_hash, key_data, key_len); // tail-call
}
return (e_value == TOMBSTONE) ? DOES_NOT_EXIST : e_value;
}
//
-uint64_t ht_get (hash_table_t *ht, const char *key_val, uint32_t key_len) {
- return hti_get(*ht, murmur32(key_val, key_len), key_val, key_len);
+uint64_t ht_get (hashtable_t *ht, const char *key_data, uint32_t key_len) {
+ return hti_get(ht->hti, murmur32(key_data, key_len), key_data, key_len);
}
//
-uint64_t ht_compare_and_set (hash_table_t *ht, const char *key_val, uint32_t key_len,
+uint64_t ht_compare_and_set (hashtable_t *ht, const char *key_data, uint32_t key_len,
uint64_t expected_val, uint64_t new_val) {
- TRACE("h2", "ht_compare_and_set: key %p len %u", key_val, key_len);
+ TRACE("h2", "ht_compare_and_set: key %p len %u", key_data, key_len);
TRACE("h2", "ht_compare_and_set: expected val %p new val %p", expected_val, new_val);
- assert(key_val);
+ assert(key_data);
assert(!IS_TAGGED(new_val) && new_val != DOES_NOT_EXIST);
- hash_table_i_t *hti = *ht;
+ hashtable_i_t *hti = ht->hti;
// Help with an ongoing copy.
if (EXPECT_FALSE(hti->next != NULL)) {
// Dispose of fully copied tables.
if (hti->num_entries_copied == (1 << hti->scale) || panic) {
assert(hti->next);
- if (SYNC_CAS(ht, hti, hti->next) == hti) {
+ if (SYNC_CAS(&ht->hti, hti, hti->next) == hti) {
nbd_defer_free(hti);
}
}
}
uint64_t old_val;
- uint32_t key_hash = murmur32(key_val, key_len);
- while ((old_val = hti_compare_and_set(hti, key_hash, key_val, key_len, expected_val, new_val))
+ uint32_t key_hash = murmur32(key_data, key_len);
+ while ((old_val = hti_compare_and_set(hti, key_hash, key_data, key_len, expected_val, new_val))
== COPIED_VALUE) {
assert(hti->next);
hti = hti->next;
return old_val == TOMBSTONE ? DOES_NOT_EXIST : old_val;
}
-// Remove the value in <ht> associated with <key_val>. Returns the value removed, or
+// Remove the value in <ht> associated with <key_data>. Returns the value removed, or
// DOES_NOT_EXIST if there was no value for that key.
-uint64_t ht_remove (hash_table_t *ht, const char *key_val, uint32_t key_len) {
- hash_table_i_t *hti = *ht;
+uint64_t ht_remove (hashtable_t *ht, const char *key_data, uint32_t key_len) {
+ hashtable_i_t *hti = ht->hti;
uint64_t val;
- uint32_t key_hash = murmur32(key_val, key_len);
+ uint32_t key_hash = murmur32(key_data, key_len);
do {
- val = hti_compare_and_set(hti, key_hash, key_val, key_len, HT_EXPECT_WHATEVER, TOMBSTONE);
+ val = hti_compare_and_set(hti, key_hash, key_data, key_len, EXPECT_WHATEVER, TOMBSTONE);
if (val != COPIED_VALUE)
return val == TOMBSTONE ? DOES_NOT_EXIST : val;
assert(hti->next);
}
// Returns the number of key-values pairs in <ht>
-uint64_t ht_count (hash_table_t *ht) {
- hash_table_i_t *hti = *ht;
+uint64_t ht_count (hashtable_t *ht) {
+ hashtable_i_t *hti = ht->hti;
uint64_t count = 0;
while (hti) {
count += hti->count;
}
// Allocate and initialize a new hash table.
-hash_table_t *ht_alloc (void) {
- hash_table_t *ht = nbd_malloc(sizeof(hash_table_t));
- *ht = (hash_table_i_t *)hti_alloc(ht, MIN_SCALE);
+hashtable_t *ht_alloc (void) {
+ hashtable_t *ht = nbd_malloc(sizeof(hashtable_t));
+ ht->hti = (hashtable_i_t *)hti_alloc(ht, MIN_SCALE);
return ht;
}
// Free <ht> and its internal structures.
-void ht_free (hash_table_t *ht) {
- hash_table_i_t *hti = *ht;
+void ht_free (hashtable_t *ht) {
+ hashtable_i_t *hti = ht->hti;
do {
for (uint32_t i = 0; i < (1 << hti->scale); ++i) {
assert(hti->table[i].value == COPIED_VALUE || !IS_TAGGED(hti->table[i].value));
nbd_free(GET_PTR(hti->table[i].key));
}
}
- hash_table_i_t *next = hti->next;
+ hashtable_i_t *next = hti->next;
nbd_free(hti);
hti = next;
} while (hti);