php的GC机制

在php5.3版本之前, php变量的回收机制只是简单的通过计数来处理(当refcount=0时，会回收内存),但这样会出现一个问题

$a=array("str");

$a[]=&$a;

unset($a);

执行unset之前,$a的refcount 为2,执行unset之后,$a的refcout为1,因为是1不等于0,不能被回收内存,即为垃圾,当然，在php脚本执行完毕后,所分配的内存将全部被回收,但是现在php除了应用于脚本以外，更多的地方用于写守护服务(当然我不推荐),可能长达一个月，两个月才结束脚本,这期间例如上面的程序会产生内存溢出

 注：unset并不能释放内存，需要看zval的refcount是否为0

php5.3以后增了GC垃圾回收机制

在分配zval时，以zval_gc_info为单位

 

#define ALLOC_ZVAL(z)                                     \
    do {                                                \
        (z) = (zval*)emalloc(sizeof(zval_gc_info));        \
        GC_ZVAL_INIT(z);                                \
    } while (0)

typedef struct _zval_gc_info {
    zval z;
    union {
        gc_root_buffer       *buffered;
        struct _zval_gc_info *next;
    } u;
} zval_gc_info;

#define FREE_ZVAL(z)                                     \
    do {                                                \
        GC_REMOVE_ZVAL_FROM_BUFFER(z);                    \
        efree(z);                                        \
    } while (0)

 

 

typedef struct _gc_root_buffer {
    struct _gc_root_buffer   *prev;        /* double-linked list               */
    struct _gc_root_buffer   *next;
    zend_object_handle        handle;    /* must be 0 for zval               */
    union {
        zval                 *pz;
        zend_object_handlers *handlers;
    } u;
} gc_root_buffer;

 

 

 

php的GC回收机制启动时,会分配10000个gc_root_buffer的空间

ZEND_API void gc_init(TSRMLS_D) { if (GC_G(buf) == NULL && GC_G(gc_enabled)) { GC_G(buf) = (gc_root_buffer*) malloc(sizeof(gc_root_buffer) * GC_ROOT_BUFFER_MAX_ENTRIES); GC_G(last_unused) = &GC_G(buf)[GC_ROOT_BUFFER_MAX_ENTRIES]; gc_reset(TSRMLS_C); } } 

 

在unset($a)时，详见这里 在active_systom_table找到key为a的zval后，对其value执行析构函数，将其zval的refcount-1，若减1后的值为0，说明可能直接释放内存，若为大于0，放到gc_root_buffer中

ZEND_API void _zval_ptr_dtor(zval **zval_ptr ZEND_FILE_LINE_DC) /* {{{ */ {   Z_DELREF_PP(zval_ptr); if (Z_REFCOUNT_PP(zval_ptr) == 0) { TSRMLS_FETCH(); if (*zval_ptr != &EG(uninitialized_zval)) { GC_REMOVE_ZVAL_FROM_BUFFER(*zval_ptr); zval_dtor(*zval_ptr); efree_rel(*zval_ptr); } } else { TSRMLS_FETCH(); if (Z_REFCOUNT_PP(zval_ptr) == 1) { Z_UNSET_ISREF_PP(zval_ptr); } GC_ZVAL_CHECK_POSSIBLE_ROOT(*zval_ptr); } } 

 

 

 

#define GC_ZVAL_CHECK_POSSIBLE_ROOT(z) / gc_zval_check_possible_root((z) TSRMLS_CC) static zend_always_inline void gc_zval_check_possible_root(zval *z TSRMLS_DC) { if (z->type == IS_ARRAY || z->type == IS_OBJECT) { gc_zval_possible_root(z TSRMLS_CC); } } ZEND_API void gc_zval_possible_root(zval *zv TSRMLS_DC) { if (UNEXPECTED(GC_G(free_list) != NULL && GC_ZVAL_ADDRESS(zv) != NULL && GC_ZVAL_GET_COLOR(zv) == GC_BLACK) && (GC_ZVAL_ADDRESS(zv) < GC_G(buf) || GC_ZVAL_ADDRESS(zv) >= GC_G(last_unused))) { /* The given zval is a garbage that is going to be deleted by * currently running GC */
        return; } if (zv->type == IS_OBJECT) { GC_ZOBJ_CHECK_POSSIBLE_ROOT(zv); return; } GC_BENCH_INC(zval_possible_root); //如果zv中的gc_root_buffer最后两位不是紫色,则进行处理 if (GC_ZVAL_GET_COLOR(zv) != GC_PURPLE) { GC_ZVAL_SET_PURPLE(zv); //设置为紫色 if (!GC_ZVAL_ADDRESS(zv)) { gc_root_buffer *newRoot = GC_G(unused); if (newRoot) { GC_G(unused) = newRoot->prev; } else if (GC_G(first_unused) != GC_G(last_unused)) { newRoot = GC_G(first_unused); GC_G(first_unused)++; } else { if (!GC_G(gc_enabled)) { GC_ZVAL_SET_BLACK(zv); return; } zv->refcount__gc++; gc_collect_cycles(TSRMLS_C); zv->refcount__gc--; newRoot = GC_G(unused); if (!newRoot) { return; } GC_ZVAL_SET_PURPLE(zv); GC_G(unused) = newRoot->prev; } newRoot->next = GC_G(roots).next; newRoot->prev = &GC_G(roots); GC_G(roots).next->prev = newRoot; GC_G(roots).next = newRoot; GC_ZVAL_SET_ADDRESS(zv, newRoot); //将gc_root_buffer放到zval_gc_info结构体中 newRoot->handle = 0; newRoot->u.pz = zv; GC_BENCH_INC(zval_buffered); GC_BENCH_INC(root_buf_length); GC_BENCH_PEAK(root_buf_peak, root_buf_length); } } }

 

 

将当前zval放到gc_root_bufer中，每个zval只放一次,依据是该zval所在的zval_gc_info中的gc_root_buffer的颜色 是否是 紫色

 

#define GC_ZVAL_GET_COLOR(v) \
GC_GET_COLOR(((zval_gc_info*)(v))->u.buffered)

#define GC_GET_COLOR(v) \
(((zend_uintptr_t)(v)) & GC_COLOR)

 

若不是紫色，则设置为紫色

#define GC_ZVAL_SET_PURPLE(v) \
GC_SET_PURPLE(((zval_gc_info*)(v))->u.buffered)

#define GC_SET_PURPLE(v) \
(v) = ((gc_root_buffer*)(((zend_uintptr_t)(v)) | GC_PURPLE))

 

宏GC_ZVAL_SET_ADDRESS(zv, newRoot);用来将gc_root_buffer newRoot 放到zv相应位置

#define GC_COLOR  0x03

#define GC_BLACK  0x00
#define GC_WHITE  0x01
#define GC_GREY   0x02
#define GC_PURPLE 0x03

#define GC_ZVAL_SET_ADDRESS(v, a) \ GC_SET_ADDRESS(((zval_gc_info*)(v))->u.buffered, (a)) #define GC_SET_ADDRESS(v, a) \ (v) = ((gc_root_buffer*)((((zend_uintptr_t)(v)) & GC_COLOR) | ((zend_uintptr_t)(a))))

 

 

GC_ZVAL_SET_ADDRESS宏中先将v强制转为zval_gc_info类型,本身在为zval分配内存时, 就是以zval_gc_info为单位的,将强转为zval*,因为只需要对zval结构体填充数据,不需要 gc_root_buffer *bufer这个样成员
因为在结构体zval_gc_val中,zval z是第一个成员,那么z的地址也是zval_gc_info本身的内存地址

在PHP GC中,使用颜色来标明垃圾的处理过程

指针无论在32位机或64位机,最后两位均为0,

gc_collect_cycles处理垃圾

ZEND_API int gc_collect_cycles(TSRMLS_D) { int count = 0; if (GC_G(roots).next != &GC_G(roots)) { zval_gc_info *p, *q, *orig_free_list, *orig_next_to_free; if (GC_G(gc_active)) { return 0; } GC_G(gc_runs)++; GC_G(zval_to_free) = FREE_LIST_END; GC_G(gc_active) = 1; gc_mark_roots(TSRMLS_C); gc_scan_roots(TSRMLS_C); gc_collect_roots(TSRMLS_C); orig_free_list = GC_G(free_list); orig_next_to_free = GC_G(next_to_free); p = GC_G(free_list) = GC_G(zval_to_free); GC_G(zval_to_free) = NULL; GC_G(gc_active) = 0; /* First call destructors */
        while (p != FREE_LIST_END) { if (Z_TYPE(p->z) == IS_OBJECT) { if (EG(objects_store).object_buckets && EG(objects_store).object_buckets[Z_OBJ_HANDLE(p->z)].valid && EG(objects_store).object_buckets[Z_OBJ_HANDLE(p->z)].bucket.obj.refcount <= 0 && EG(objects_store).object_buckets[Z_OBJ_HANDLE(p->z)].bucket.obj.dtor &&
                    !EG(objects_store).object_buckets[Z_OBJ_HANDLE(p->z)].destructor_called) { 　　EG(objects_store).object_buckets[Z_OBJ_HANDLE(p->z)].destructor_called = 1; 　　EG(objects_store).object_buckets[Z_OBJ_HANDLE(p->z)].bucket.obj.refcount++; 　　EG(objects_store).object_buckets[Z_OBJ_HANDLE(p->z)].bucket.obj.dtor(EG(objects_store).object_buckets[Z_OBJ_HANDLE(p->z)].bucket.obj.object, Z_OBJ_HANDLE(p->z) TSRMLS_CC); 　　EG(objects_store).object_buckets[Z_OBJ_HANDLE(p->z)].bucket.obj.refcount--; 　　} } count++; p = p->u.next; } /* Destroy zvals */ p = GC_G(free_list); while (p != FREE_LIST_END) { GC_G(next_to_free) = p->u.next; if (Z_TYPE(p->z) == IS_OBJECT) { if (EG(objects_store).object_buckets && EG(objects_store).object_buckets[Z_OBJ_HANDLE(p->z)].valid && EG(objects_store).object_buckets[Z_OBJ_HANDLE(p->z)].bucket.obj.refcount <= 0) { EG(objects_store).object_buckets[Z_OBJ_HANDLE(p->z)].bucket.obj.refcount = 1; Z_TYPE(p->z) = IS_NULL; zend_objects_store_del_ref_by_handle_ex(Z_OBJ_HANDLE(p->z), Z_OBJ_HT(p->z) TSRMLS_CC); } } else if (Z_TYPE(p->z) == IS_ARRAY) { Z_TYPE(p->z) = IS_NULL; zend_hash_destroy(Z_ARRVAL(p->z)); FREE_HASHTABLE(Z_ARRVAL(p->z)); } else { zval_dtor(&p->z); Z_TYPE(p->z) = IS_NULL; } p = GC_G(next_to_free); } /* Free zvals */ p = GC_G(free_list); while (p != FREE_LIST_END) { q = p->u.next; FREE_ZVAL_EX(&p->z); p = q; } GC_G(collected) += count; GC_G(free_list) = orig_free_list; GC_G(next_to_free) = orig_next_to_free; } return count; }

 

gc_mark_roots(TSRMLS_C); 为垃圾打下颜色标记，遍历gc_root_buffer，其中的u.pz将紫色变更为灰色，遍历u.pz该array zval的每个元素,将其refcount-1,
gc_collect_roots遍历gc_root_buffer，如果refcount==0再设置为白色，表示为垃圾,若refcount >0,表示别人在使用，设置为黑色

static void gc_mark_roots(TSRMLS_D) { gc_root_buffer *current = GC_G(roots).next; while (current != &GC_G(roots)) { if (current->handle) { if (EG(objects_store).object_buckets) { //处理对象,暂时不用看 } } else { if (GC_ZVAL_GET_COLOR(current->u.pz) == GC_PURPLE) { zval_mark_grey(current->u.pz TSRMLS_CC); } else { GC_ZVAL_SET_ADDRESS(current->u.pz, NULL); GC_REMOVE_FROM_BUFFER(current); } } current = current->next; } }

 

static void zval_mark_grey(zval *pz TSRMLS_DC) { Bucket *p; tail_call: if (GC_ZVAL_GET_COLOR(pz) != GC_GREY) { p = NULL; GC_BENCH_INC(zval_marked_grey); GC_ZVAL_SET_COLOR(pz, GC_GREY); if (Z_TYPE_P(pz) == IS_OBJECT && EG(objects_store).object_buckets) { //对象的处理, 暂时不用管 } else if (Z_TYPE_P(pz) == IS_ARRAY) { if (Z_ARRVAL_P(pz) == &EG(symbol_table)) { GC_ZVAL_SET_BLACK(pz); } else { p = Z_ARRVAL_P(pz)->pListHead; } } while (p != NULL) { pz = *(zval**)p->pData; if (Z_TYPE_P(pz) != IS_ARRAY || Z_ARRVAL_P(pz) != &EG(symbol_table)) { pz->refcount__gc--; } if (p->pListNext == NULL) { goto tail_call; } else { zval_mark_grey(pz TSRMLS_CC); } p = p->pListNext; } } }

 

 第二次遍历gc_root_buffer,如果zv中的颜色为灰色，且refcount=0,再置为白色，若refcount>0,置为黑色（不是垃圾）

static void gc_scan_roots(TSRMLS_D)
{
    gc_root_buffer *current = GC_G(roots).next;

    while (current != &GC_G(roots)) {
        if (current->handle) {
            zval z;

            INIT_PZVAL(&z);
            Z_OBJ_HANDLE(z) = current->handle;
            Z_OBJ_HT(z) = current->u.handlers;
            zobj_scan(&z TSRMLS_CC);
        } else {
            zval_scan(current->u.pz TSRMLS_CC);
        }
        current = current->next;
    }
}

 

 

static int zval_scan(zval *pz TSRMLS_DC)
{
    Bucket *p;

tail_call:    
    if (GC_ZVAL_GET_COLOR(pz) == GC_GREY) {
        p = NULL;
        if (pz->refcount__gc > 0) {
            zval_scan_black(pz TSRMLS_CC);
        } else {
            GC_ZVAL_SET_COLOR(pz, GC_WHITE); if (Z_TYPE_P(pz) == IS_OBJECT && EG(objects_store).object_buckets) {
               //处理object
            } else if (Z_TYPE_P(pz) == IS_ARRAY) {
                if (Z_ARRVAL_P(pz) == &EG(symbol_table)) {
                    GC_ZVAL_SET_BLACK(pz);
                } else {
                    p = Z_ARRVAL_P(pz)->pListHead;
                }
            }
        }
        while (p != NULL) {
            if (p->pListNext == NULL) {
                pz = *(zval**)p->pData;
                goto tail_call;
            } else {
                zval_scan(*(zval**)p->pData TSRMLS_CC);
            }
            p = p->pListNext;
        }
    }
    return 0;
}

 

 

遍历gc_root_buffer链表，将zv颜色为白色的数据放置单独一个链表，全部回收

static void gc_collect_roots(TSRMLS_D) { gc_root_buffer *current = GC_G(roots).next; while (current != &GC_G(roots)) { if (current->handle) { if (EG(objects_store).object_buckets) { struct _store_object *obj = &EG(objects_store).object_buckets[current->handle].bucket.obj; zval z; GC_SET_ADDRESS(obj->buffered, NULL); INIT_PZVAL(&z); Z_OBJ_HANDLE(z) = current->handle; Z_OBJ_HT(z) = current->u.handlers; zobj_collect_white(&z TSRMLS_CC); } } else { GC_ZVAL_SET_ADDRESS(current->u.pz, NULL); zval_collect_white(current->u.pz TSRMLS_CC); } GC_REMOVE_FROM_BUFFER(current); current = current->next; } }

 

 

static void zval_collect_white(zval *pz TSRMLS_DC) { Bucket *p; tail_call: if (((zval_gc_info*)(pz))->u.buffered == (gc_root_buffer*)GC_WHITE) { p = NULL; GC_ZVAL_SET_BLACK(pz); if (Z_TYPE_P(pz) == IS_OBJECT && EG(objects_store).object_buckets) { ... } else { if (Z_TYPE_P(pz) == IS_ARRAY) { p = Z_ARRVAL_P(pz)->pListHead; } } /* restore refcount and put into list to free */ pz->refcount__gc++; ((zval_gc_info*)pz)->u.next = GC_G(zval_to_free); GC_G(zval_to_free) = (zval_gc_info*)pz; while (p != NULL) { pz = *(zval**)p->pData; if (Z_TYPE_P(pz) != IS_ARRAY || Z_ARRVAL_P(pz) != &EG(symbol_table)) { pz->refcount__gc++; } if (p->pListNext == NULL) { goto tail_call; } else { zval_collect_white(pz TSRMLS_CC); } p = p->pListNext; } } }