Analyzing the evolution of the garbage collection mechanism in PHP5-PHP Tutorial-php.cn

Preface

PHP is a managed language. In PHP programming, programmers do not need to manually handle the allocation and release of memory resources (except when using C to write PHP or Zend extensions), which means that PHP It implements the garbage collection mechanism (Garbage Collection) itself. Now if you go to the official PHP website (php.net) you can see that the current two branch versions of PHP5, PHP5.2 and PHP5.3, are updated separately. This is because many projects still use the 5.2 version of PHP, and the 5.3 version is 5.2 is not fully compatible. PHP5.3 has made many improvements based on PHP5.2, among which the garbage collection algorithm is a relatively big change. This article will discuss the garbage collection mechanisms of PHP5.2 and PHP5.3 respectively, and discuss the impact of this evolution and improvement on programmers writing PHP and the issues they should pay attention to.

Internal representation of PHP variables and associated memory objects

Garbage collection is ultimately an operation on variables and their associated memory objects, so before discussing PHP’s garbage collection mechanism, let’s briefly introduce it The internal representation of variables and their memory objects in PHP (their representation in C source code).

The PHP official documentation divides variables in PHP into two categories: scalar types and complex types. Scalar types include booleans, integers, floating point types and strings; complex types include arrays, objects and resources; there is also a special NULL, which is not divided into any type, but becomes a separate category.

All these types are uniformly represented by a structure called zval within PHP. In the PHP source code, the name of this structure is "_zval_struct". The specific definition of zval is in the "Zend/zend.h" file of the PHP source code. The following is an excerpt of the relevant code.

typedef union _zvalue_value {
    long lval;                  /* long value */
    double dval;                /* double value */
    struct {
        char *val;
        int len;
    } str;
    HashTable *ht;              /* hash table value */
    zend_object_value obj;
} zvalue_value;
 
struct _zval_struct {
    /* Variable information */
    zvalue_value value;     /* value */
    zend_uint refcount__gc;
    zend_uchar type;    /* active type */
    zend_uchar is_ref__gc;
};

Copy after login

The union "_zvalue_value" is used to represent the values of all variables in PHP. The reason why union is used here is because a zval can only represent one type of variable at a time. You can see that there are only 5 fields in _zvalue_value, but there are 8 data types in PHP including NULL. So how does PHP use 5 fields to represent 8 types internally? This is one of the more clever aspects of PHP design. It achieves the purpose of reducing fields by reusing fields. For example, within PHP, Boolean types, integers and resources (as long as the identifier of the resource is stored) are stored through the lval field; dval is used to store floating point types; str stores strings; ht stores arrays (note that in PHP The array is actually a hash table); and obj stores the object type; if all fields are set to 0 or NULL, it means NULL in PHP, so that 5 fields are used to store 8 types of values.

The type of value in the current zval (the type of value is _zvalue_value) is determined by the type in "_zval_struct". _zval_struct is the specific implementation of zval in C language. Each zval represents a memory object of a variable. In addition to value and type, you can see that there are two fields refcount__gc and is_ref__gc in _zval_struct. From their suffixes, you can conclude that these two guys are related to garbage collection. That's right, PHP's garbage collection relies entirely on these two fields. Among them, refcount__gc indicates that there are several variables currently referencing this zval, and is_ref__gc indicates whether the current zval is referenced by reference. This sounds very confusing. This is related to the "Write-On-Copy" mechanism of zval in PHP. Since this topic is not This article is the focus, so I won’t go into details here. Readers only need to remember the role of the refcount__gc field.

Garbage collection algorithm in PHP5.2 - Reference Counting

The memory recycling algorithm used in PHP5.2 is the famous Reference Counting. The Chinese translation of this algorithm is called "reference counting". The idea is very intuitive and concise: assign a counter to each memory object. When a memory object is created, the counter is initialized to 1 (so there is always a variable referencing this object at this time). Every time a new variable refers to this memory object, The counter is incremented by 1, and every time a variable that references this memory object is reduced, the counter is decremented by 1. When the garbage collection mechanism operates, all memory objects with a counter of 0 are destroyed and the memory they occupy is recycled. The memory object in PHP is zval, and the counter is refcount__gc.

For example, the following PHP code demonstrates the working principle of the PHP5.2 counter (the counter value is obtained through xdebug):

<?php
 
$val1 = 100; //zval(val1).refcount_gc = 1;
$val2 = $val1; //zval(val1).refcount_gc = 2,zval(val2).refcount_gc = 2(因为是Write on copy，当前val2与val1共同引用一个zval)
$val2 = 200; //zval(val1).refcount_gc = 1,zval(val2).refcount_gc = 1(此处val2新建了一个zval)
unset($val1); //zval(val1).refcount_gc = 0($val1引用的zval再也不可用，会被GC回收)
 
?>
Reference Counting简单直观，实现方便，但却存在一个致命的缺陷，就是容易造成内存泄露。很多朋友可能已经意识到了，如果存在循环引用，那么Reference Counting就可能导致内存泄露。例如下面的代码：
<?php
$a = array();
$a[] = & $a;
unset($a);
 
?>

Copy after login

This code first creates the array a, and then lets the first An element points to a by reference. At this time, the refcount of a's zval becomes 2. Then we destroy the variable a. At this time, the refcount of the zval initially pointed to by a is 1, but we can no longer operate on it because it A circular self-reference is formed, as shown in the following figure:

Analyzing the evolution of the garbage collection mechanism in PHP5

The gray part indicates that it no longer exists. Since the refcount of the zval pointed to by a is 1 (referenced by the first element of its HashTable), this zval will not be destroyed by GC, and this part of the memory will be leaked.

这里特别要指出的是，PHP是通过符号表（Symbol Table）存储变量符号的，全局有一个符号表，而每个复杂类型如数组或对象有自己的符号表，因此上面代码中，a和a[0]是两个符号，但是a储存在全局符号表中，而a[0]储存在数组本身的符号表中，且这里a和a[0]引用同一个zval（当然符号a后来被销毁了）。希望读者朋友注意分清符号（Symbol）的zval的关系。

在PHP只用于做动态页面脚本时，这种泄露也许不是很要紧，因为动态页面脚本的生命周期很短，PHP会保证当脚本执行完毕后，释放其所有资源。但是PHP发展到目前已经不仅仅用作动态页面脚本这么简单，如果将PHP用在生命周期较长的场景中，例如自动化测试脚本或deamon进程，那么经过多次循环后积累下来的内存泄露可能就会很严重。这并不是我在耸人听闻，我曾经实习过的一个公司就通过PHP写的deamon进程来与数据存储服务器交互。

由于Reference Counting的这个缺陷，PHP5.3改进了垃圾回收算法。

PHP5.3中的垃圾回收算法——Concurrent Cycle Collection in Reference Counted Systems

PHP5.3的垃圾回收算法仍然以引用计数为基础，但是不再是使用简单计数作为回收准则，而是使用了一种同步回收算法，这个算法由IBM的工程师在论文Concurrent Cycle Collection in Reference Counted Systems中提出。

这个算法可谓相当复杂，从论文29页的数量我想大家也能看出来，所以我不打算（也没有能力）完整论述此算法，有兴趣的朋友可以阅读上面的提到的论文（强烈推荐，这篇论文非常精彩）。

我在这里，只能大体描述一下此算法的基本思想。

首先PHP会分配一个固定大小的“根缓冲区”，这个缓冲区用于存放固定数量的zval，这个数量默认是10,000，如果需要修改则需要修改源代码Zend/zend_gc.c中的常量GC_ROOT_BUFFER_MAX_ENTRIES然后重新编译。

由上文我们可以知道，一个zval如果有引用，要么被全局符号表中的符号引用，要么被其它表示复杂类型的zval中的符号引用。因此在zval中存在一些可能根（root）。这里我们暂且不讨论PHP是如何发现这些可能根的，这是个很复杂的问题，总之PHP有办法发现这些可能根zval并将它们投入根缓冲区。

当根缓冲区满额时，PHP就会执行垃圾回收，此回收算法如下：

1、对每个根缓冲区中的根zval按照深度优先遍历算法遍历所有能遍历到的zval，并将每个zval的refcount减1，同时为了避免对同一zval多次减1（因为可能不同的根能遍历到同一个zval），每次对某个zval减1后就对其标记为“已减”。

2、再次对每个缓冲区中的根zval深度优先遍历，如果某个zval的refcount不为0，则对其加1，否则保持其为0。

3、清空根缓冲区中的所有根（注意是把这些zval从缓冲区中清除而不是销毁它们），然后销毁所有refcount为0的zval，并收回其内存。

如果不能完全理解也没有关系，只需记住PHP5.3的垃圾回收算法有以下几点特性：

1、并不是每次refcount减少时都进入回收周期，只有根缓冲区满额后在开始垃圾回收。

2、可以解决循环引用问题。

3、可以总将内存泄露保持在一个阈值以下。

PHP5.2与PHP5.3垃圾回收算法的性能比较

由于我目前条件所限，我就不重新设计试验了，而是直接引用PHP Manual中的实验，关于两者的性能比较请参考PHP Manual中的相关章节：http://www.php.net/manual/en/features.gc.performance-considerations.php。

首先是内存泄露试验，下面直接引用PHP Manual中的实验代码和试验结果图：

<?php
class Foo
{
    public $var = &#39;3.1415962654&#39;;
}
 
$baseMemory = memory_get_usage();
 
for ( $i = 0; $i <= 100000; $i++ )
{
    $a = new Foo;
    $a->self = $a;
    if ( $i % 500 === 0 )
    {
        echo sprintf( &#39;%8d: &#39;, $i ), memory_get_usage() - $baseMemory, "\n";
    }
}
?>

Copy after login

Analyzing the evolution of the garbage collection mechanism in PHP5

可以看到在可能引发累积性内存泄露的场景下，PHP5.2发生持续累积性内存泄露，而PHP5.3则总能将内存泄露控制在一个阈值以下（与根缓冲区大小有关）。

另外是关于性能方面的对比：

<?php
class Foo
{
    public $var = &#39;3.1415962654&#39;;
}
 
for ( $i = 0; $i <= 1000000; $i++ )
{
    $a = new Foo;
    $a->self = $a;
}
 
echo memory_get_peak_usage(), "\n";
?>

Copy after login

这个脚本执行1000000次循环，使得延迟时间足够进行对比。

然后使用CLI方式分别在打开内存回收和关闭内存回收的的情况下运行此脚本：

time php -dzend.enable_gc=0 -dmemory_limit=-1 -n example2.php
# and
time php -dzend.enable_gc=1 -dmemory_limit=-1 -n example2.php

Copy after login

在我的机器环境下，运行时间分别为6.4s和7.2s，可以看到PHP5.3的垃圾回收机制会慢一些，但是影响并不大。

PHP configuration related to garbage collection algorithm

You can turn on or off PHP's garbage collection mechanism by modifying zend.enable_gc in php.ini, or you can turn it on by calling gc_enable() or gc_disable() Or turn off PHP's garbage collection mechanism. Even if the garbage collection mechanism is turned off in PHP5.3, PHP will still record possible roots to the root buffer, but when the root buffer is full, PHP will not automatically run garbage collection. Of course, you can manually call gc_collect_cycles at any time. () function forces memory recycling.

The above is the detailed content of Analyzing the evolution of the garbage collection mechanism in PHP5. For more information, please follow other related articles on the PHP Chinese website!