X-Git-Url: https://pd.if.org/git/?p=nbds;a=blobdiff_plain;f=map%2Fhashtable.c;h=85ffe41f20226f920a31bed8dabc5d3e7d55525d;hp=ba62692b8414cefdcbf685c6bab5ded5d1723904;hb=778b8c8ca708b082a1192acfb114a6751b2ad7c9;hpb=a03cf3b0c40e6c3b8b4877b49a64288cb3fcb919

diff --git a/map/hashtable.c b/map/hashtable.c
index ba62692..85ffe41 100644
--- a/map/hashtable.c
+++ b/map/hashtable.c
@@ -9,7 +9,9 @@
  * 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>
@@ -34,6 +36,9 @@ typedef struct hti {
     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 ref_count;
@@ -52,8 +57,8 @@ struct ht {
     const datatype_t *key_type;
 };
 
-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 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;
@@ -95,7 +100,7 @@ static volatile entry_t *hti_lookup (hti_t *hti, map_key_t key, uint32_t key_has
             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;
             }
@@ -136,16 +141,19 @@ static volatile entry_t *hti_lookup (hti_t *hti, map_key_t key, uint32_t key_has
 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;
     }
@@ -169,8 +177,8 @@ static void hti_start_copy (hti_t *hti) {
     // 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);
@@ -178,7 +186,11 @@ static void hti_start_copy (hti_t *hti) {
     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);
@@ -193,7 +205,7 @@ static int hti_copy_entry (hti_t *ht1, volatile entry_t *ht1_ent, uint32_t key_h
     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
@@ -233,9 +245,11 @@ static int hti_copy_entry (hti_t *ht1, volatile entry_t *ht1_ent, uint32_t key_h
     // 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 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;
@@ -276,8 +290,8 @@ static int hti_copy_entry (hti_t *ht1, volatile entry_t *ht1_ent, uint32_t key_h
     // 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;
     }
 
@@ -362,9 +376,9 @@ static map_val_t hti_cas (hti_t *hti, map_key_t key, uint32_t key_hash, map_val_
     map_val_t ent_val = ent->val;
     if (EXPECT_FALSE(IS_TAGGED(ent_val, TAG1))) {
         if (ent_val != COPIED_VALUE && ent_val != TAG_VALUE(TOMBSTONE, TAG1)) {
-            int did_copy = hti_copy_entry(hti, ent, key_hash, ((volatile hti_t *)hti)->next);
+            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);
@@ -398,9 +412,9 @@ static map_val_t hti_cas (hti_t *hti, map_key_t key, uint32_t key_hash, map_val_
 
     // 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.
@@ -417,7 +431,7 @@ static map_val_t hti_get (hti_t *hti, map_key_t key, uint32_t key_hash) {
     // 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;
     }
@@ -429,12 +443,12 @@ static map_val_t hti_get (hti_t *hti, map_key_t key, uint32_t key_hash) {
     map_val_t ent_val = ent->val;
     if (EXPECT_FALSE(IS_TAGGED(ent_val, TAG1))) {
         if (EXPECT_FALSE(ent_val != COPIED_VALUE && ent_val != TAG_VALUE(TOMBSTONE, TAG1))) {
-            int did_copy = hti_copy_entry(hti, ent, key_hash, ((volatile hti_t *)hti)->next);
+            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;
@@ -442,7 +456,11 @@ static map_val_t hti_get (hti_t *hti, map_key_t key, uint32_t key_hash) {
 
 //
 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);
 }
 
@@ -455,9 +473,9 @@ static int hti_help_copy (hti_t *hti) {
     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;
 
@@ -473,26 +491,26 @@ static int hti_help_copy (hti_t *hti) {
             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 < (1 << hti->scale); ++i) {
+    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)
@@ -502,7 +520,11 @@ static void hti_defer_free (hti_t *hti) {
             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);
 }
 
@@ -538,7 +560,11 @@ map_val_t ht_cas (hashtable_t *ht, map_key_t key, map_val_t expected_val, map_va
     }
 
     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;
@@ -552,7 +578,11 @@ map_val_t ht_cas (hashtable_t *ht, map_key_t key, map_val_t expected_val, map_va
 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)
@@ -598,7 +628,7 @@ void ht_print (hashtable_t *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, (uint64_t)ent->val);
             if (i > 30) {
@@ -637,7 +667,7 @@ map_val_t ht_iter_next (ht_iter_t *iter, map_key_t *key_ptr) {
     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) {
@@ -649,16 +679,23 @@ map_val_t ht_iter_next (ht_iter_t *iter, map_key_t *key_ptr) {
 
     } 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;
 }