安全
在线手册:中文 英文
PHP手册

文件系统安全

Table of Contents

PHP 遵从大多数服务器系统中关于文件和目录权限的安全机制。这就使管理员可以控制哪些文件在文件系统内是可读的。必须特别注意的是全局的可读文件,并确保每一个有权限的用户对这些文件的读取动作都是安全的。

PHP 被设计为以用户级别来访问文件系统,所以完全有可能通过编写一段 PHP 代码来读取系统文件如 /etc/passwd,更改网络连接以及发送大量打印任务等等。因此必须确保 PHP 代码读取和写入的是合适的文件。

请看下面的代码,用户想要删除自己主目录中的一个文件。假设此情形是通过 web 界面来管理文件系统,因此 Apache 用户有权删除用户目录下的文件。

Example #1 不对变量进行安全检查会导致……

<?php
// 从用户目录中删除指定的文件
$username $_POST['user_submitted_name'];
$userfile $_POST['user_submitted_filename'];
$homedir "/home/$username";
unlink ("$homedir/$userfile");
echo 
"The file has been deleted!";
?>
既然 username 和 filename 变量可以通过用户表单来提交,那就可以提交别人的用户名和文件名,甚至可以删掉本来不应该让他们删除的文件。这种情况下,就要考虑其它方式的认证。想想看如果提交的变量是“../etc/”和“passwd”会发生什么。上面的代码就等同于:

Example #2 ……文件系统攻击

<?php
// 删除硬盘中任何 PHP 有访问权限的文件。如果 PHP 有 root 权限:
$username $_POST['user_submitted_name']; // "../etc"
$userfile $_POST['user_submitted_filename']; // "passwd"
$homedir  "/home/$username"// "/home/../etc"

unlink("$homedir/$userfile"); // "/home/../etc/passwd"

echo "The file has been deleted!";
?>
有两个重要措施来防止此类问题。 下面是改进的脚本:

Example #3 更安全的文件名检查

<?php
// 删除硬盘中 PHP 有权访问的文件
$username $_SERVER['REMOTE_USER']; // 使用认证机制
$userfile basename($_POST['user_submitted_filename']);
$homedir "/home/$username";

$filepath "$homedir/$userfile";


if (
file_exists($filepath) && unlink($filepath)) {
    
$logstring "Deleted $filepath\n";
} else {
    
$logstring "Failed to delete $filepath\n";
}
$fp fopen("/home/logging/filedelete.log""a");
fwrite ($fp$logstring);
fclose($fp);

echo 
htmlentities($logstringENT_QUOTES);
?>
然而,这样做仍然是有缺陷的。如果认证系统允许用户建立自己的登录用户名,而用户选择用“../etc/”作为用户名,系统又一次沦陷了。所以,需要加强检查:

Example #4 更安全的文件名检查

<?php
$username 
$_SERVER['REMOTE_USER']; // 使用认证机制
$userfile     $_POST['user_submitted_filename'];
$homedir "/home/$username";

$filepath     "$homedir/$userfile";
if (!
ctype_alnum($username) || !preg_match('/^(?:[a-z0-9_-]|\.(?!\.))+$/iD'$userfile)) {
    die(
"Bad username/filename");
}
//后略……
?>

根据操作系统的不同,存在着各种各样需要注意的文件,包括联系到系统的设备(/dev/ 或者 COM1)、配置文件(/ect/ 文件和 .ini文件)、常用的存储区域(/home/ 或者 My Documents)等等。由于此原因,建立一个策略禁止所有权限而只开放明确允许的通常更容易些。


安全
在线手册:中文 英文
PHP手册
PHP手册 - N: 文件系统安全

用户评论:

bk2 at nospam me dot com (05-Feb-2012 07:48)

I note that the php standard error printing includes "filename", which appears to always be the /full/folder/path/and/filename. And on all unix systems and by common practice on many windows systems, the /home/(control panel login id)/ is part of the folder path, and on most hosts the same account login id is the control panel login id. So any error message effectively exposes to the world the control panel login id. From there a hacker and brute force can break any password in time.  PHP does not appear to have a way to disable displaying the actual physical directory structure from appearing in error messages, which even is no login id were there, is a security problem. One can direct the error to a log file, but that still exposes critical information to contract web programmers. Typical practice is to give web programmers limited ftp access to a web directory, but then php defeats this restriction by displaying the entire directory structure in the error message to the web developer who you tried to give restricted access to. And 9 time out of ten also gave away the control panel login in id. So all the wise script writing in the world cannot obscure the private information the php error printing is giving out. Can it?

Latchezar Tzvetkoff (10-Mar-2009 11:06)

A basic filename/directory/symlink checking may be done (and I personally do) via realpath() ...

<?php

if (isset($_GET['file'])) {
   
$base = '/home/polizei/public_html/'// it seems this one is good to be realpath too.. meaning not a symlinked path..
   
if (strpos($file = realpath($base.$_GET['file']), $base) === 0 && is_file($file)) {
       
unlink($file);
    } else {
        die(
'blah!');
    }
}
?>

BAT Svensson (27-Mar-2008 10:20)

A note on mapping for added security.

The principle with mapping is: WHAT YOU SEE MAY NOT BE.

Map each entity that can be manipulated to a token. Only the token will be displayed to the end user. For added security, map allowed functionality to each token. Changing the token will at most cause a user to manipulate some else already mapped file. This achieve the goal of control and limited access.

There is a variety to implement this but the main point here is that a physical name should never be displayed to the end user, thus restricting an evil user knowledge of the system and it architecture.

This principle is also know under names such as "black box" or "need to know" or "abstraction", etc etc depending on the context.

The point here is that one does not need to know under lying structure in order to do operation one high levels such as "add/move/remove/edit". Present a view - the view will provide the user with what the user need to know.

It is all about creating an illusion of existence at higher level of something that does not exist on a lover level.

steelchords at yahoo dot com (17-Apr-2007 09:12)

It seems to me in this particular instance that a simple check to make sure that name or partial pathname doesn't already exist would prevent this attack... if a 'passwd/etc/...' existed as the password directory, you couldn't create a username to exploit the hole in the first place.  But that's only from a 'script user' perspective, it still doesn't protect your server from other sub-admin's badly written code.

Don For

anonymous (17-Nov-2005 10:58)

(A) Better not to create files or folders with user-supplied names. If you do not validate enough, you can have trouble. Instead create files and folders with randomly generated names like fg3754jk3h and store the username and this file or folder name in a table named, say, user_objects. This will ensure that whatever the user may type, the command going to the shell will contain values from a specific set only and no mischief can be done.

(B) The same applies to commands executed based on an operation that the user chooses. Better not to allow any part of the user's input to go to the command that you will execute. Instead, keep a fixed set of commands and based on what the user has input, and run those only.

For example,
(A) Keep a table named, say, user_objects with values like:
username|chosen_name   |actual_name|file_or_dir
--------|--------------|-----------|-----------
jdoe    |trekphotos    |m5fg767h67 |D
jdoe    |notes.txt     |nm4b6jh756 |F
tim1997 |_imp_ folder  |45jkh64j56 |D

and always use the actual_name in the filesystem operations rather than the user supplied names.

(B)
<?php
$op
= $_POST['op'];//after a lot of validations
$dir = $_POST['dirname'];//after a lot of validations or maybe you can use technique (A)
switch($op){
    case
"cd":
       
chdir($dir);
        break;
    case
"rd":
       
rmdir($dir);
        break;
    .....
    default:
       
mail("webmaster@example.com", "Mischief", $_SERVER['REMOTE_ADDR']." is probably attempting an attack.");
}

fmrose at ncsu dot edu (10-Oct-2005 12:31)

All of the fixes here assume that it is necessary to allow the user to enter system sensitive information to begin with. The proper way to handle this would be to provide something like a numbered list of files to perform an unlink action on and then the chooses the matching number. There is no way for the user to specify a clever attack circumventing whatever pattern matching filename exclusion syntax that you may have.

Anytime you have a security issue, the proper behaviour is to deny all then allow specific instances, not allow all and restrict. For the simple reason that you may not think of every possible restriction.

joshudson';DROP TABLE EMAILS;' gmail.com (01-Sep-2005 05:50)

I keep application configuration files in the document root. I found the most effective trick to prevent access to them is to
1. Give them no code that actually runs when included (except for variable assignments),
2. Don't use register globals so nobody can do anything weird,
3. Name them *.php so PHP runs them when asked for
4. Don't have anything before <?php
5. Don
't have a ?>

1 at 234 dot cx (23-Jun-2005 01:24)

I don't think the filename validation solution from Jones at partykel is complete.  It certainly helps, but it doesn't address the case where the user is able to create a symlink pointing from his home directory to the root.  He might then ask to unlink "foo/etc/passwd" which would be in his home directory, except that foo is a symlink pointing to /.

Personally I wouldn't feel confident that any solution to this problem would keep my system secure.  Running PHP as root (or some equivalent which can unlink files in all users' home directories) is asking for trouble.

If you have a multi-user system and you are afraid that users may install scripts like this, try security-enhanced Linux.  It won't give total protection, but it at least makes sure that an insecure user script can only affect files which the web server is meant to have access to.  Whatever script someone installs, outsiders are not going to be able to read your password file---or remove it.

Jones at partykel dot de (10-Dec-2004 10:00)

What about:

<?php
  $file_to_delete
= '/home/'.$username.'/'.$userfile;
 
  if ((!
ereg('\.\.', $file_to_delete)) and (file_exists($file_to_delete))) {
     
unlink($file_to_delete);
  }
?>

I think this should prevent every attempt to go outside the user-directory.
Additionally you should check usernames at the registration.

Another way whould be to use the user-ID as home-directory - so, this can't be changed and every registered user as an unique one (if it's a primary key in your database). But then you still have to check the given $userfile.

So the code above could be taken as a "last instance check" directly before finally deleting of the file.

akul at inbox dot ru (06-May-2004 05:59)

Common and simple way to avoid path attack is separating objectname and filesystem space when possible. For example if I have users on my site and directory per user solution is not "httpdocs/users/$login", but "httpdocs/users/".md5($login) or "httpdocs/users/".$userId

tim at correctclick dot com (31-Mar-2002 12:48)

One more thing --

whenever you connect to a database with a password hard coded into your script, make sure you put the script off of your web document tree.  Put the script somewhere where apache won't serve documents, and then include/require this file in your other scripts.  That way, if the server ever gets misconfigured, it won't serve your PHP scripts with passwords, etc. as plain text for all to see.

-Tim

cronos586(AT)caramail(DOT)com (05-Jan-2002 11:48)

when using Apache you might consider a apache_lookup_uri on the path, to discover the real path, regardless of any directory trickery.
then, look at the prefix, and compare with a list of allowed prefixes.
for example, my source.php for my website includes:
if(isset($doc)) {
    $apacheres = apache_lookup_uri($doc);
    $really = realpath($apacheres->filename);
    if(substr($really, 0, strlen($DOCUMENT_ROOT)) == $DOCUMENT_ROOT) {
        if(is_file($really)) {
            show_source($really);
        }
    }
}
hope this helps
regards,
KAT44

devik at cdi dot cz (21-Aug-2001 03:52)

Well, the fact that all users run under the same UID is a big problem. Userspace  security hacks (ala safe_mode) should not be substitution for proper kernel level security checks/accounting.
Good news: Apache 2 allows you to assign UIDs for different vhosts.
devik

eLuddite at not dot a dot real dot addressnsky dot ru (11-Nov-2000 05:44)

I think the lesson is clear:

(1) Forbit path separators in usernames.
(2) map username to a physical home directory - /home/username is fine
(3) read the home directory
(4) present only results of (3) as an option for deletion.

I have discovered a marvelous method of doing the above in php but this submission box is too small to contain it.

:-)