oj 3.13.2 → 3.13.6

data/README.md

@@ -64,7 +64,7 @@ links.
 
 ## Releases
 
-See [{file:CHANGELOG.md}](CHANGELOG.md)
+See [{file:CHANGELOG.md}](CHANGELOG.md) and [{file:RELEASE_NOTES.md}](RELEASE_NOTES.md)
 
 ## Links
 

data/RELEASE_NOTES.md

@@ -0,0 +1,55 @@
+# RELEASE NOTES
+
+The release notes here are organized by release. For a list of changes
+see the See [{file:CHANGELOG.md}](CHANGELOG.md) file. In this file are
+the steps to take to aid in keeping things rolling after updating to
+the latest version.
+
+## 3.13.x
+
+This release included a new cache that performs better than the
+earlier cache and a new high performance parser.
+
+### Cache
+
+The new cache includes a least recently used expiration to reduce
+memory use. The cache is also self adjusting and will expand as needed
+for better performance. It also handles Hash keys and string values
+with two options, `:cache_keys`, a boolean and `:cache_str` an
+integer. The `:cache_str` if set to more than zero is the limit for
+the length of string values to cache. The maximum value is 35 which
+allows strings up to 34 bytes to be cached.
+
+One interesting aspect of the cache is not so much the string caching
+which performs similar to the Ruby intern functions but the caching of
+symbols and object attribute names. There is a significant gain for
+symbols and object attributes.
+
+If the cache is not desired then setting the default options to turn
+it off can be done with this line:
+
+``` ruby
+Oj.default_options = { cache_keys: false, cache_str: 0 }
+```
+
+### Oj::Parser
+
+The new parser uses a different core that follows the approach taken
+by [OjC](https://github.com/ohler55/ojc) and
+[OjG](https://github.com/ohler55/ojg). It also takes advantage of the
+bulk Array and Hash functions. Another issue the new parser addresses
+is option management. Instead of a single global default_options each
+parser instance maintains it's own options.
+
+There is a price to be paid when using the Oj::Parser. The API is not
+the same the older parser. A single parser can only be used in a
+single thread. This allows reuse of internal buffers for additional
+improvements in performance.
+
+The performane advantage of the Oj::Parse is that it is more than 3
+times faster than the Oj::compat_load call and 6 times faster than the
+JSON gem.
+
+### Dump Performance
+
+Thanks to Watson1978 Oj.dump also received a speed boost.

data/ext/oj/cache.c

@@ -1,65 +1,74 @@
 // Copyright (c) 2011, 2021 Peter Ohler. All rights reserved.
 // Licensed under the MIT License. See LICENSE file in the project root for license details.
 
+#if HAVE_PTHREAD_MUTEX_INIT
+#include <pthread.h>
+#endif
+#include <stdlib.h>
+
 #include "cache.h"
 
-#define REHASH_LIMIT 64
+// The stdlib calloc, realloc, and free are used instead of the Ruby ALLOC,
+// ALLOC_N, REALLOC, and xfree since the later could trigger a GC which will
+// either corrupt memory or if the mark function locks will deadlock.
+
+#define REHASH_LIMIT 4
 #define MIN_SHIFT 8
+#define REUSE_MAX 8192
+
+#if HAVE_PTHREAD_MUTEX_INIT
+#define CACHE_LOCK(c) pthread_mutex_lock(&((c)->mutex))
+#define CACHE_UNLOCK(c) pthread_mutex_unlock(&((c)->mutex))
+#else
+#define CACHE_LOCK(c) rb_mutex_lock((c)->mutex)
+#define CACHE_UNLOCK(c) rb_mutex_unlock((c)->mutex)
+#endif
+
+// almost the Murmur hash algorithm
+#define M 0x5bd1e995
 
 typedef struct _slot {
-    struct _slot *next;
-    VALUE         val;
-    uint32_t      hash;
-    uint8_t       klen;
-    char          key[CACHE_MAX_KEY];
+    struct _slot *    next;
+    VALUE             val;
+    uint64_t          hash;
+    volatile uint32_t use_cnt;
+    uint8_t           klen;
+    char              key[CACHE_MAX_KEY];
 } * Slot;
 
 typedef struct _cache {
-    Slot * slots;
-    size_t cnt;
+    volatile Slot * slots;
+    volatile size_t cnt;
     VALUE (*form)(const char *str, size_t len);
-    uint32_t size;
-    uint32_t mask;
-    bool     mark;
+    uint64_t size;
+    uint64_t mask;
+    VALUE (*intern)(struct _cache *c, const char *key, size_t len);
+    volatile Slot reuse;
+    size_t        rcnt;
+#if HAVE_PTHREAD_MUTEX_INIT
+    pthread_mutex_t mutex;
+#else
+    VALUE mutex;
+#endif
+    uint8_t xrate;
+    bool    mark;
 } * Cache;
 
-// almost the Murmur hash algorithm
-#define M 0x5bd1e995
-#define C1 0xCC9E2D51
-#define C2 0x1B873593
-#define N 0xE6546B64
-
 void cache_set_form(Cache c, VALUE (*form)(const char *str, size_t len)) {
     c->form = form;
 }
 
-#if 0
-// For debugging only.
-static void cache_print(Cache c) {
-    for (uint32_t i = 0; i < c->size; i++) {
-        printf("%4d:", i);
-        for (Slot s = c->slots[i]; NULL != s; s = s->next) {
-            char buf[40];
-            strncpy(buf, s->key, s->klen);
-            buf[s->klen] = '\0';
-            printf(" %s", buf);
-        }
-        printf("\n");
-    }
-}
-#endif
-
-static uint32_t hash_calc(const uint8_t *key, size_t len) {
+static uint64_t hash_calc(const uint8_t *key, size_t len) {
     const uint8_t *end     = key + len;
     const uint8_t *endless = key + (len & 0xFFFFFFFC);
-    uint32_t       h       = (uint32_t)len;
-    uint32_t       k;
+    uint64_t       h       = (uint64_t)len;
+    uint64_t       k;
 
     while (key < endless) {
-        k = (uint32_t)*key++;
-        k |= (uint32_t)*key++ << 8;
-        k |= (uint32_t)*key++ << 16;
-        k |= (uint32_t)*key++ << 24;
+        k = (uint64_t)*key++;
+        k |= (uint64_t)*key++ << 8;
+        k |= (uint64_t)*key++ << 16;
+        k |= (uint64_t)*key++ << 24;
 
         k *= M;
         k ^= k >> 24;
@@ -83,43 +92,25 @@ static uint32_t hash_calc(const uint8_t *key, size_t len) {
     return h;
 }
 
-Cache cache_create(size_t size, VALUE (*form)(const char *str, size_t len), bool mark) {
-    Cache c     = ALLOC(struct _cache);
-    int   shift = 0;
-
-    for (; REHASH_LIMIT < size; size /= 2, shift++) {
-    }
-    if (shift < MIN_SHIFT) {
-        shift = MIN_SHIFT;
-    }
-    c->size  = 1 << shift;
-    c->mask  = c->size - 1;
-    c->slots = ALLOC_N(Slot, c->size);
-    memset(c->slots, 0, sizeof(Slot) * c->size);
-    c->form = form;
-    c->cnt  = 0;
-    c->mark = mark;
-
-    return c;
-}
-
 static void rehash(Cache c) {
-    uint32_t osize = c->size;
-    Slot *   end   = c->slots + osize;
+    uint64_t osize;
+    Slot *   end;
     Slot *   sp;
 
-    c->size = osize * 4;
-    c->mask = c->size - 1;
-    REALLOC_N(c->slots, Slot, c->size);
-    memset(c->slots + osize, 0, sizeof(Slot) * osize * 3);
-    for (sp = c->slots; sp < end; sp++) {
+    osize    = c->size;
+    c->size  = osize * 4;
+    c->mask  = c->size - 1;
+    c->slots = realloc((void *)c->slots, sizeof(Slot) * c->size);
+    memset((Slot *)c->slots + osize, 0, sizeof(Slot) * osize * 3);
+    end = (Slot *)c->slots + osize;
+    for (sp = (Slot *)c->slots; sp < end; sp++) {
         Slot s    = *sp;
         Slot next = NULL;
 
         *sp = NULL;
         for (; NULL != s; s = next) {
-            uint32_t h      = s->hash & c->mask;
-            Slot *   bucket = c->slots + h;
+            uint64_t h      = s->hash & c->mask;
+            Slot *   bucket = (Slot *)c->slots + h;
 
             next    = s->next;
             s->next = *bucket;
@@ -128,8 +119,148 @@ static void rehash(Cache c) {
     }
 }
 
+static VALUE lockless_intern(Cache c, const char *key, size_t len) {
+    uint64_t       h      = hash_calc((const uint8_t *)key, len);
+    Slot *         bucket = (Slot *)c->slots + (h & c->mask);
+    Slot           b;
+    volatile VALUE rkey;
+
+    while (REUSE_MAX < c->rcnt) {
+        if (NULL != (b = c->reuse)) {
+            c->reuse = b->next;
+            free(b);
+            c->rcnt--;
+        } else {
+            // An accounting error occured somewhere so correct it.
+            c->rcnt = 0;
+        }
+    }
+    for (b = *bucket; NULL != b; b = b->next) {
+        if ((uint8_t)len == b->klen && 0 == strncmp(b->key, key, len)) {
+            b->use_cnt += 16;
+            return b->val;
+        }
+    }
+    rkey = c->form(key, len);
+    if (NULL == (b = c->reuse)) {
+        b = calloc(1, sizeof(struct _slot));
+    } else {
+        c->reuse = b->next;
+        c->rcnt--;
+    }
+    b->hash = h;
+    memcpy(b->key, key, len);
+    b->klen     = (uint8_t)len;
+    b->key[len] = '\0';
+    b->val      = rkey;
+    b->use_cnt  = 4;
+    b->next     = *bucket;
+    *bucket     = b;
+    c->cnt++;  // Don't worry about wrapping. Worse case is the entry is removed and recreated.
+    if (REHASH_LIMIT < c->cnt / c->size) {
+        rehash(c);
+    }
+    return rkey;
+}
+
+static VALUE locking_intern(Cache c, const char *key, size_t len) {
+    uint64_t       h;
+    Slot *         bucket;
+    Slot           b;
+    uint64_t       old_size;
+    volatile VALUE rkey;
+
+    CACHE_LOCK(c);
+    while (REUSE_MAX < c->rcnt) {
+        if (NULL != (b = c->reuse)) {
+            c->reuse = b->next;
+            free(b);
+            c->rcnt--;
+        } else {
+            // An accounting error occured somewhere so correct it.
+            c->rcnt = 0;
+        }
+    }
+    h      = hash_calc((const uint8_t *)key, len);
+    bucket = (Slot *)c->slots + (h & c->mask);
+    for (b = *bucket; NULL != b; b = b->next) {
+        if ((uint8_t)len == b->klen && 0 == strncmp(b->key, key, len)) {
+            b->use_cnt += 4;
+            CACHE_UNLOCK(c);
+
+            return b->val;
+        }
+    }
+    old_size = c->size;
+    // The creation of a new value may trigger a GC which be a problem if the
+    // cache is locked so make sure it is unlocked for the key value creation.
+    if (NULL != (b = c->reuse)) {
+        c->reuse = b->next;
+        c->rcnt--;
+    }
+    CACHE_UNLOCK(c);
+    if (NULL == b) {
+        b = calloc(1, sizeof(struct _slot));
+    }
+    rkey    = c->form(key, len);
+    b->hash = h;
+    memcpy(b->key, key, len);
+    b->klen     = (uint8_t)len;
+    b->key[len] = '\0';
+    b->val      = rkey;
+    b->use_cnt  = 16;
+
+    // Lock again to add the new entry.
+    CACHE_LOCK(c);
+    if (old_size != c->size) {
+        h      = hash_calc((const uint8_t *)key, len);
+        bucket = (Slot *)c->slots + (h & c->mask);
+    }
+    b->next = *bucket;
+    *bucket = b;
+    c->cnt++;  // Don't worry about wrapping. Worse case is the entry is removed and recreated.
+    if (REHASH_LIMIT < c->cnt / c->size) {
+        rehash(c);
+    }
+    CACHE_UNLOCK(c);
+
+    return rkey;
+}
+
+Cache cache_create(size_t size, VALUE (*form)(const char *str, size_t len), bool mark, bool locking) {
+    Cache c     = calloc(1, sizeof(struct _cache));
+    int   shift = 0;
+
+    for (; REHASH_LIMIT < size; size /= 2, shift++) {
+    }
+    if (shift < MIN_SHIFT) {
+        shift = MIN_SHIFT;
+    }
+#if HAVE_PTHREAD_MUTEX_INIT
+    pthread_mutex_init(&c->mutex, NULL);
+#else
+    c->mutex = rb_mutex_new();
+#endif
+    c->size    = 1 << shift;
+    c->mask    = c->size - 1;
+    c->slots   = calloc(c->size, sizeof(Slot));
+    c->form    = form;
+    c->xrate   = 1;  // low
+    c->mark    = mark;
+    if (locking) {
+        c->intern = locking_intern;
+    } else {
+        c->intern = lockless_intern;
+    }
+    return c;
+}
+
+void cache_set_expunge_rate(Cache c, int rate) {
+    c->xrate = (uint8_t)rate;
+}
+
 void cache_free(Cache c) {
-    uint32_t i;
+    uint64_t i;
 
     for (i = 0; i < c->size; i++) {
         Slot next;
@@ -137,57 +268,59 @@ void cache_free(Cache c) {
 
         for (s = c->slots[i]; NULL != s; s = next) {
             next = s->next;
-            xfree(s);
+            free(s);
         }
     }
-    xfree(c->slots);
-    xfree(c);
+    free((void *)c->slots);
+    free(c);
 }
 
 void cache_mark(Cache c) {
-    if (c->mark) {
-	uint32_t i;
+    uint64_t i;
+
+#if !HAVE_PTHREAD_MUTEX_INIT
+    rb_gc_mark(c->mutex);
+#endif
+    if (0 == c->cnt) {
+        return;
+    }
+    for (i = 0; i < c->size; i++) {
+        Slot s;
+        Slot prev = NULL;
+        Slot next;
 
-        for (i = 0; i < c->size; i++) {
-            Slot s;
-            for (s = c->slots[i]; NULL != s; s = s->next) {
+        for (s = c->slots[i]; NULL != s; s = next) {
+            next = s->next;
+            if (0 == s->use_cnt) {
+                if (NULL == prev) {
+                    c->slots[i] = next;
+                } else {
+                    prev->next = next;
+                }
+                c->cnt--;
+                s->next  = c->reuse;
+                c->reuse = s;
+                c->rcnt++;
+                continue;
+            }
+            switch (c->xrate) {
+            case 0: break;
+            case 2: s->use_cnt -= 2; break;
+            case 3: s->use_cnt /= 2; break;
+            default: s->use_cnt--; break;
+            }
+            if (c->mark) {
                 rb_gc_mark(s->val);
             }
+            prev = s;
         }
     }
 }
 
 VALUE
 cache_intern(Cache c, const char *key, size_t len) {
-    if (CACHE_MAX_KEY < len) {
+    if (CACHE_MAX_KEY <= len) {
         return c->form(key, len);
     }
-    uint32_t h      = hash_calc((const uint8_t *)key, len);
-    Slot *   bucket = c->slots + (h & c->mask);
-    Slot     b;
-    Slot     tail = NULL;
-
-    for (b = *bucket; NULL != b; b = b->next) {
-        if ((uint8_t)len == b->klen && 0 == strncmp(b->key, key, len)) {
-            return b->val;
-        }
-        tail = b;
-    }
-    b       = ALLOC(struct _slot);
-    b->hash = h;
-    b->next = NULL;
-    memcpy(b->key, key, len);
-    b->klen     = (uint8_t)len;
-    b->key[len] = '\0';
-    b->val      = c->form(key, len);
-    if (NULL == tail) {
-        *bucket = b;
-    } else {
-        tail->next = b;
-    }
-    c->cnt++;
-    if (REHASH_LIMIT < c->cnt / c->size) {
-        rehash(c);
-    }
-    return b->val;
+    return c->intern(c, key, len);
 }