Date: Fri, 19 Feb 1999 16:08:06 -0500 >From: "Simson L. Garfinkel" <simsongat_private> Subject: Process-table attack Wide-ranging attack works against almost any UNIX systems on the Internet ABSTRACT: The Process Table Attack is a [relatively] new kind of denial-of-service attack that can be waged against numerous network services on a variety of different UNIX systems. The attack is launched against network services which fork() or otherwise allocate a new process for each incoming TCP/IP connection. Although the standard UNIX operating system places limits on the number of processes that any one user may launch, there are no limits on the number of processes that the superuser can create other than the hard limits imposed by the operating system. Since incoming TCP/IP connections are usually handled by servers that run as root, it is possible to completely fill a target machine's process table with multiple instantiations of network servers. Properly executed, this attack prevents any other command from being executed on the target machine. DETAILS In the book Practical UNIX and Internet Security, Gene Spafford and I observed that the UNIX operating system originally contained few defenses to protect it from a denial-of-service attack. This is changing. With the growth of the Internet, there has been a concerted effort in recent years to strengthen the operating system and its network services to these attacks. Each time a network client makes a connection to a network server, a number of resources on the server are consumed. The most important resources consumed are memory, disk space, and CPU time. Some network services, such as sendmail, now monitor system resources and will not accept incoming network connections if accepting them would place the system in jeopardy. One system resource that has escaped monitoring is the number of processes that are currently running on a computer. Most versions of UNIX will only allow a certain number of processes to be running at one time. Each process takes up a slot in the system's process table. By filling this table, it is possible to prevent the operating system from creating new processes, even when other resources (such as memory, disk space, and CPU time) are widely available. The implementation of many network services leaves them open to a process table attack ? that is, an attack in which the attacker fills up the target computer's process table so that no new programs can be executed. The design of some network protocols actually leads the programmer into making these mistakes. An example of such a protocol is the finger protocol (TCP port 79). The protocol follows this sequence: 1. The client makes a connection to the server. 2. The server accepts the connection, and creates a process to service the request. 3. The client sends a single line to the server consisting of the name of the entity that the client wishes to finger. 4. The server performs the necessary database lookup and sends the information back to the client. 5. The server closes the connection. To launch a process table attack, the client need only open a connection to the server and not send any information. As long as the client holds the connection open, the server's process will occupy a slot in the server's process table. On most computers, finger is launched by inetd. The authors of inetd placed several checks into the program's source code which must be bypassed in order to initiate a successful process attack. If the inetd receives more than 40 connections to a particular service within 1 minute, that service is disabled for 10 minutes. The purpose of these checks was not to protect the server against a process table attack, but to protect the server against buggy code that might create many connections in rapid-fire sequence. To launch a successful process table attack against a computer running inetd and finger, the following sequence may be followed: 1. Open a connection to the target's finger port. 2. Wait for 4 seconds. 3. Repeat steps 1-2. The attack program is not without technical difficulty. Many systems limit the number TCP connections that may be initiated by a single process. Thus, it may be necessary to launch the attack from multiple processes, perhaps running on multiple computers. We have tested a variety of network services on a variety of operating systems. We believe that the UW imap and sendmail servers are also vulnerable. The UW imap contains no checks for rapid-fire connections. Thus, it is possible to shut down a computer by opening multiple connections to the imap server in rapid succession. With sendmail the situation is reversed. Normally, sendmail will not accept connections after the system load has jumped above a predefined level. Thus, to initiate a successful sendmail attack it is necessary to open the connections very slowly, so that the process table keeps growing in size while the system load remains more or less constant. We have also seen a variety of problems on BSD-based machines being used as the attacker. When the target machine freezes or crashes, the attacker machine sometimes crashes as well. Apparently the TCP stack does not gracefully handle hundreds of connections to the same port on the same machine simultaneously going into the FIN_WAIT_2 state. There are variants of this attack: 1. Use IP spoofing so that the incoming connections appear to come from many different locations on the Internet. This makes tracking considerably harder to do. 2. Begin the attack by sending 50 requests in rapid fire to the telnet, rlogin and rsh ports on the target machine. This will cause inetd to shut down those services on the target machine, which will deny administrative access during the attack. 3. Instead of initiating a new connection every 4 seconds, initiate one every minute or so. The attack slowly builds, making it more difficult to detect on packet traces. There are several ways to defend against the attack: 1. inetd and other programs should check to see the number of free slots in the process table before accepting new connections. If there is less than a predefined number of free slots, new connections should be accepted. 2. Alternatively, if there are more than a preset number of network daemons for the service running, incoming requests should be queued rather than serviced. 3. Network services (such as finger) should implement timeouts. For example, the statement alarm(30) could be inserted into the finger daemon source code so that the program would stop running after 30 seconds of execution. Simson L. Garfinkel, Sandstorm Enterprises, Inc. <www.sandstorm.net> [Simson informed me over a year ago that he had discovered this attack and had notified many relevant operating system vendors. To the best of my knowledge, no one has addressed the problem in the intervening year. We thus include this item in the hopes of spurring some action, or at least awareness and public discussion. On the other hand, we of course do not recommend conducting experiments to demonstrate this flaw on other people's systems. PGN]
This archive was generated by hypermail 2b30 : Fri Apr 13 2001 - 14:36:41 PDT