1	Detection, Prevention, and Containment: A Study of grsecurity Brad Spengler http://www.grsecurity.net spender@grsecurity.net
2
3	The Problem Bugs in software cause unexpected results Unexpected functionality can result from errors in design, implementation, or configuration Bugs can often be wielded for malicious purposes by an attacker
4	Problems With the Current Solution Avoid / Identify / Fix Current state of security is a never ending rat race Endless cycle of vulnerability discovery and fixing
5	Problems With the Current Solution Ultimate goal of today’s security – removal of software bugs through auditing Security utopia – greatest result, though impossible to achieve
6	Problems With the Current Solution Auditing is expensive, slow, and requires a great deal of knowledge Auditing provides no guarantees about the security of the software Auditing cannot be fully automated EXAMPLE: format-string vulnerabilities
7
8	The (Attainable) Solution Detection Prevention Containment
9	Advantages of the (Attainable) Solution Inexpensive Can be mostly automated Works for known and unknown bugs Allows administrators to focus more on administration (checking logs..etc) instead of rushing for the newest patch
10	Our solution: grsecurity
11
12	Background on grsecurity Started in February 2001 Initial release was for Linux 2.4.1 Originally a port of Openwall to Linux 2.4
13	Goals of grsecurity Configuration-free operation Complete protection against all forms of address space modification bugs Feature-rich ACL and auditing systems Operation on multiple processor architectures and Operating Systems
14	Features of grsecurity A robust ACL system with an intelligent userspace administration tool Extensive auditing capabilities Measures to stop the most common methods of exploiting a system: Address space modification Races (specifically filesystem races, most common of which are /tmp races) Breaking a chroot(2) jail
15	Features of grsecurity Supports sysctl so that it can be included with Linux distributions and allow the user to modify the options to his/her liking Netfilter module that drops connections to unserved TCP and UDP ports Many of the same randomness features as OpenBSD An enhanced implementation of Trusted Path Execution (TPE)
16
17	Detection in grsecurity Implemented in two forms Auditing Logging of real attacks Inode and device numbers used wherever possible Parent process info logged
18	Auditing Audited events include: Exec (with arguments) Chdir(2) Mount(2)/unmount(2) IPC (semaphore, message queue, shared memory) creation and deletion
19	Auditing Signals: SIGSEGV, SIGABRT, SIGBUS, SIGILL Failed forks Ptrace(2) Time changes (stime(2), settimeofday(2)) Execs inside chroot(2) Denied capabilities
20
21	Prevention in grsecurity Prevention is implemented through PaX and hardening certain sections of the kernel Hardened syscalls include: Chroot(2) Ptrace(2) Mmap(2) Link(2)/symlink(2) Sysctl(2)
22	Prevention in grsecurity - PaX What is PaX? PaX implements non-executable VM pages on architectures that do not support the non-executable bit (currently only ia-32, more to come) PaX makes use of hardware-supported non-executable bits (still to be tested, but should work for alpha, parisc, and ia-64) PaX provides full address space layout randomization (ASLR) for ELF binaries
23	Prevention in grsecurity - PaX How does PaX accomplish this? Include/asm-<arch>/processor.h is modified to support executable and non-executable pages (if they don’t already exist) Rest of kernel is modified to use the non-executable pages, applied to ELF and a.out binaries if they carry the required PaX flags (enabled by default)
24	Prevention in grsecurity - PaX Non-executable pages are made supervisor in the TLB; executable pages are left as user If CPU is in user mode, access to the non-executable pages causes a page-fault which PaX handles Makes up the core logic of how PaX works Makes PaX ineffective against kernel overflows Mmap(2) and mprotect(2) restrictions/features Disallows anonymous mappings with PROT_EXEC present – stops one method of arbitrary code execution (another method, mapping a file with PROT_EXEC, is handled by ACL system)
25	Prevention in grsecurity - PaX Causes mmaps (applies to libraries) to be mapped at random locations below 0x01000000 until it’s full, then above 0x40000000 Causes exploits to have to guess the library function address Makes the address contain a NULL byte, which stops ASCII shellcode from calling a library function Keeps non-executable pages from being mprotected to executable No performance impact
26	Prevention in grsecurity - PaX Full Address Space Layout Randomization (ASLR) Randomizes the base of mmaps, stack, and executable (if the binary is ET_DYN) Makes the leftover methods of exploitation a guessing game
27	Prevention in grsecurity - PaX
28	Prevention in grsecurity - PaX Full ASLR can only be bypassed in the case of information leak. While there’s nothing that can be done about software vulnerabilities that allow information leaking without crashing, we’ve implemented the following features to stop local users from obtaining information about the random base addresses: Ptrace(2) restrictions in ACL system Restricted /proc For 64-bit architectures, the randomness provided by full ASLR could be increased to 48/64/80 bits (the amount the attacker has to overcome is determined by the type of exploit)
29	Prevention in grsecurity - PaX What’s in it for me? No more arbitrary code execution No more stack smashing, heap or bss overflow exploitation No more return-to-libc exploitation (Soon) no more arbitrary execution flow redirection
30	Prevention in grsecurity - PaX What’s coming for this section of grsecurity? New segmentation-based implementation of non-executable pages with an insignificant performance hit Increased stack base address randomness to 24 bits Binary instrumentation Stops ret-to-libc by checkpointing execution flow changes Ability to handle other vulnerabilities (eg. Stack based overflows, format string, info-leak)
31	Prevention in grsecurity OpenBSD randomness features Random IP IDs Random RPC XIDs Random RPC privileged ports Random PIDs
32	Prevention in grsecurity Random IP IDs Uses Niels Provos’ random IP ID generation function ported to Linux Little entropy use Values are not reused quickly Useful for preventing OS fingerprinting and spoofed scans
33	Prevention in grsecurity Random RPC XIDs Uses same random IP ID code Useful for preventing RPC connection hijacking Random PIDs Uses same random IP ID code Properties of returned values make the function almost always return an unused PID even on heavily loaded servers
34	Prevention in grsecurity Prevents filesystem races since getpid() is sometimes used as part of a temporary filename Adds additional randomness to programs that use getpid(2) for srandom(3) seeding
35	Prevention in grsecurity Stealth netfilter module Based on the fact that OS fingerprinting relies greatly on the packets sent in reply to those sent to unserved TCP or UDP ports Matches unserved ports dynamically, so it can be used in shell-server environments Slows down blocking port-scanners
36	Prevention in grsecurity Problems with chroot(2) Easy to use it insecurely Generally only filesystem-related functions care if a process is chrooted Easy for a root user in chroot to break out
37	Prevention in grsecurity How we strengthen chroot(2): Make syscalls unrelated to the filesystem chroot-aware Deny double-chroots, pivot_root(2) Restrict signals outside of chroot Deny fchdir(2) outside of chroot Deny mounting Enforce chdir(“/”) upon chroot Lower capabilities upon chroot
38
39	Containment in grsecurity Trusted Path Execution (TPE) Keeps users from executing untrusted binaries (those not in root-owned non-world writable directories) Hardened against evasion Silent removal of glibc environment variables that allow arbitrary code execution (eg. LD_PRELOAD) TPE checks implemented in mmap(2) (stops /lib/ld.so <executable> evasion)
40	Containment in grsecurity Grsecurity’s ACL system Process-based : Allowed for a large reduction in code base ACL parsing handled via userspace, interacts with kernel via a /proc entry Include directive ACL analysis $PATH /etc/ld.so.conf Auto-add libraries for ELF executables /etc/lilo.conf
41	Containment in grsecurity Uses LEX/YACC Sends data to kernel in ready-to-use structures – further reduces necessary kernel code size Enable, disable, and administrator modes Hidden and protected processes Read, write, append and execute modes for file objects Inherit and hidden flags for file objects
42	Containment in grsecurity Capability support (including inheritance) Hardened against ACL evasion and privilege leaking Ptrace restriction – user can only ptrace a process if the default ACL allows writing to it Glibc environment variable handling Performs correct handling, not just a denied exec if LD_ is found Checks each path in glibc environment to see if the default ACL allows writing to it; if so, deny the exec and log pathname and environment variable used Applies executable restrictions in mmap(2), not just execve
43	Containment in grsecurity Human readable configuration files Insignificant performance impact due to efficient searching algorithms (hash tables == O(1) )
44	Containment in grsecurity What’s coming for the ACL system? Redesign to become more modular and allow quicker implementation of new features Intelligent learning mode resulting in a least-privilege system with little or no configuration necessary Support of fine-grained resource restrictions and something similar to nergal’s segvguard Time-based ACLs Merging of GID-based grsecurity features Role-Based Access Control (RBAC)
45	Containment in grsecurity Domain-based authentication support
46
47	Performance of ACL system Completed 150 runs of 16 dbench processes Average throughput with ACL system was larger than a clean kernel Standard deviation was 5MB/s, which was larger than the difference of throughput RESULT: The ACL system causes no noticeable performance hit on filesystem access
48	Performance of ACL system Results of kernel compile benchmark: Total time with ACL system – 265.86 seconds Total time w/o ACL system – 264.94 seconds .4% performance hit Performance hit only due to execs in compiling and making – search is called twice, acl label is copied, acl label is set, checks are performed on the environment
49	Performance with PaX Memory load latency microbenchmarks MySQL benchmarks (real life example) Test system: Dual AMD XP 1600+ 512MB PC2100 ECC DDR registered RAM 266mhz FSB 80GB ATA100 5400RPM HD
50	Performance with PaX
51	Performance with PaX
52	Performance with PaX Athlons encounter less performance hit partially due to their 256 entry DTLB (4KB page x 256 = 1MB) PaX starts showing its performance impact when the DTLB is full and expired entries are replaced Performance with PaX can only be determined by the size and type of memory accesses performed by an application
53	Performance with PaX
54	Performance with PaX
55	Performance with PaX
56	Performance with PaX A result weighted according to an actual system’s load shows that for MySQL, PaX caused an overall performance hit of 13% Since the memory access patterns of each test were different, the performance hits for each test ranged from 3% - 20%
57	For More Information… grsecurity’s ACL documentation: http://www.grsecurity.net/gracldoc.htm PaX http://pageexec.virtualave.net
58	THANKS PaX Team Tim Yardley Michael Dalton - grsecurity