Home > Articles

Memory Management

  • Print
  • + Share This
This chapter is from the book

This chapter is from the book

ONE OF THE MOST JARRING DIFFERENCES BETWEEN A MANAGED language like PHP, and an unmanaged language like C is control over memory pointers.


In PHP, populating a string variable is as simple as <?php $str = 'hello world'; ?> and the string can be freely modified, copied, and moved around. In C, on the other hand, although you could start with a simple static string such as char *str = "hello world";, that string cannot be modified because it lives in program space. To create a manipulable string, you'd have to allocate a block of memory and copy the contents in using a function such as strdup().

   char *str;

   str = strdup("hello world");
   if (!str) {
       fprintf(stderr, "Unable to allocate memory!");

For reasons you'll explore through the course of this chapter, the traditional memory management functions (malloc(), free(), strdup(), realloc(), calloc(), and so on) are almost never used directly by the PHP source code.

Free the Mallocs

Memory management on nearly all platforms is handled in a request and release fashion. An application says to the layer above it (usually the operating system) "I want some number of bytes of memory to use as I please." If there is space available, the operating system offers it to the program and makes a note not to give that chunk of memory out to anyone else.

When the application is done using the memory, it's expected to give it back to the OS so that it can be allocated elsewhere. If the program doesn't give the memory back, the OS has no way of knowing that it's no longer being used and can be allocated again by another process. If a block of memory is not freed, and the owning application has lost track of it, then it's said to have "leaked" because it's simply no longer available to anyone.

In a typical client application, small infrequent leaks are sometimes tolerated with the knowledge that the process will end after a short period of time and the leaked memory will be implicitly returned to the OS. This is no great feat as the OS knows which program it gave that memory to, and it can be certain that the memory is no longer needed when the program terminates.

With long running server daemons, including web servers like Apache and by extension mod_php, the process is designed to run for much longer periods, often indefinitely. Because the OS can't clean up memory usage, any degree of leakage—no matter how small—will tend to build up over time and eventually exhaust all system resources.

Consider the userspace stristr() function; in order to find a string using a caseinsensitive search, it actually creates a lowercase copy of both the haystack and the needle, and then performs a more traditional case-sensitive search to find the relative offset. After the offset of the string has been located, however, it no longer has use for the lowercase versions of the haystack and needle strings. If it didn't free these copies, then every script that used stristr() would leak some memory every time it was called. Eventually the web server process would own all the system memory, but not be able to use it.

The ideal solution, I can hear you shouting, is to write good, clean, consistent code, and that's absolutely true. In an environment like the PHP interpreter, however, that's only half the solution.

Error Handling

In order to provide the ability to bail out of an active request to userspace scripts and the extension functions they rely on, a means needs to exist to jump out of an active request entirely. The way this is handled within the Zend Engine is to set a bailout address at the beginning of a request, and then on any die() or exit() call, or on encountering any critical error (E_ERROR) perform a longjmp() to that bailout address.

Although this bailout process simplifies program flow, it almost invariably means that resource cleanup code (such as free() calls) will be skipped and memory could get leaked. Consider this simplified version of the engine code that handles function calls:

void call_function(const char *fname, int fname_len TSRMLS_DC)
    zend_function *fe;
    char *lcase_fname;
    /* PHP function names are case-insensitive
     * to simplify locating them in the function tables
     * all function names are implicitly
     * translated to lowercase
    lcase_fname = estrndup(fname, fname_len);
    zend_str_tolower(lcase_fname, fname_len);

    if (zend_hash_find(EG(function_table),
            lcase_fname, fname_len + 1, (void **)&fe) == FAILURE) {
        zend_execute(fe->op_array TSRMLS_CC);
    } else {
        php_error_docref(NULL TSRMLS_CC, E_ERROR,
                         "Call to undefined function: %s()", fname);

When the php_error_docref() line is encountered, the internal error handler sees that the error level is critical and invokes longjmp() to interrupt the current program flow and leave call_function() without ever reaching the efree(lcase_fname) line. Again, you're probably thinking that the efree() line could just be moved above the zend_error() line, but what about the code that called this call_function() routine in the first place? Most likely fname itself was an allocated string and you can't free that before it has been used in the error message.

Zend Memory Manager

The solution to memory leaks during request bailout is the Zend Memory Management (ZendMM) layer. This portion of the engine acts in much the same way the operating system would normally act, allocating memory to calling applications. The difference is that it is low enough in the process space to be request-aware so that when one request dies, it can perform the same action the OS would perform when a process dies. That is, it implicitly frees all the memory owned by that request. Figure 3.1 shows ZendMM in relation to the OS and the PHP process.


Figure 3.1 Zend Memory Manager replaces system calls for per-request allocations.

In addition to providing implicit memory cleanup, ZendMM also controls the perrequest memory usage according to the php.ini setting: memory_limit. If a script attempts to ask for more memory than is available to the system as a whole, or more than is remaining in its per-request limit, ZendMM will automatically issue an E_ERROR message and begin the bailout process. An added benefit of this is that the return value of most memory allocation calls doesn't need to be checked because failure results in an immediate longjmp() to the shutdown part of the engine.

Hooking itself in between PHP internal code and the OS's actual memory management layer is accomplished by nothing more complex than requiring that all memory allocated internally is requested using an alternative set of functions. For example, rather than allocate a 16-byte block of memory using malloc(16), PHP code will use emalloc(16). In addition to performing the actual memory allocation task, ZendMM will flag that block with information concerning what request it's bound to so that when a request bails out, ZendMM can implicitly free it.

Often, memory needs to be allocated for longer than the duration of a single request. These types of allocations, called persistent allocations because they persist beyond the end of a request, could be performed using the traditional memory allocators because these do not add the additional per-request information used by ZendMM. Sometimes, however, it's not known until runtime whether a particular allocation will need to be persistent or not, so ZendMM exports a set of helper macros that act just like the other memory allocation functions, but have an additional parameter at the end to indicate persistence.

If you genuinely want a persistent allocation, this parameter should be set to one, in which case the request will be passed through to the traditional malloc() family of allocators. If runtime logic has determined that this block does not need to be persistent however, this parameter may be set to zero, and the call will be channeled to the perrequest memory allocator functions.

For example, pemalloc(buffer_len, 1) maps to malloc(buffer_len), whereas pemalloc(buffer_len, 0) maps to emalloc(buffer_len) using the following

#define in Zend/zend_alloc.h:

#define pemalloc(size, persistent)             ((persistent)?malloc(size): emalloc(size))

Each of the allocator functions found in ZendMM can be found below along with their more traditional counterparts.

Table 3.1 shows each of the allocator functions supported by ZendMM and their e/pe counterparts:

Table 3.1. Traditional versus PHP-specific allocators

Allocator funtion

e/pe counterpart

void *malloc(size_t count);

void *emalloc(size_t count);

void *pemalloc(size_t count, char persistent);

void *calloc(size_t count);

void *ecalloc(size_t count);

void *pecalloc(size_t count, char persistent);

void *realloc(void *ptr, size_t count);

void *erealloc(void *ptr, size_t count);

void *perealloc(void *ptr, size_t count, char persistent);

void *strdup(void *ptr);

void *estrdup(void *ptr);

void *pestrdup(void *ptr, char persistent);

void free(void *ptr);

void efree(void *ptr);

void pefree(void *ptr, char persistent);

You'll notice that even pefree() requires the persistency flag. This is because at the time that pefree() is called, it doesn't actually know if ptr was a persistent allocation or not. Calling free() on a non-persistent allocation could lead to a messy double free, whereas calling efree() on a persistent one will most likely lead to a segmentation fault as the memory manager attempts to look for management information that doesn't exist. Your code is expected to remember whether the data structure it allocated was persistent or not.

In addition to the core set of allocator functions, a few additional and quite handy ZendMM specific functions exist:

void *estrndup(void *ptr, int len);

Allocate len+1 bytes of memory and copy len bytes from ptr to the newly allocated block. The behavior of estrndup() is roughly the following:

void *estrndup(void *ptr, int len)
    char *dst = emalloc(len + 1);
    memcpy(dst, ptr, len);
    dst[len] = 0;
    return dst;

The terminating NULL byte implicitly placed at the end of the buffer here ensures that any function that uses estrndup() for string duplication doesn't need to worry about passing the resulting buffer to a function that expects NULL terminated strings such as printf(). When using estrndup() to copy non-string data, this last byte is essentially wasted, but more often than not, the convenience outweighs the minor inefficiency.

void *safe_emalloc(size_t size, size_t count, size_t addtl);
void *safe_pemalloc(size_t size, size_t count, size_t addtl, char persistent);

The amount of memory allocated by these functions is the result of ((size * count) + addtl). You may be asking, "Why an extra function at all? Why not just use emalloc/pemalloc and do the math myself?"The reason comes in the name: safe. Although the circumstances leading up to it would be exceedingly unlikely, it's possible that the end result of such an equation might overflow the integer limits of the host platform. This could result in an allocation for a negative number of bytes, or worse, a positive number that is significantly smaller than what the calling program believed it requested. safe_emalloc() avoids this type of trap by checking for integer overflow and explicitly failing if such an overflow occurs.

  • + Share This
  • 🔖 Save To Your Account