Optimized memcmp leaks useful timing differences

One of the main questions raised after our Blackhat talk on exploiting remote timing attacks was how memcmp() optimizations might affect our results.

Our talk focused on a particular timing vulnerability. If an HMAC comparison exits early when a mismatch is found, it reveals information to an attacker about the correct HMAC for the forged message. This early-exit behavior is desirable in general but can leak timing information about secret data, allowing an attacker to iteratively guess the secret.

In C, memcmp() is almost always used for comparisons of binary data. Its API specifies that it compares two fixed-length buffers and returns the difference between them or zero if they are identical. In most implementations, memcmp() exits as soon as a difference is found in the two buffers.

Even if memcmp() exits early, the size of comparisons might affect how exploitable this timing delta is. For example, an optimized memcmp() may work in units of 64-bit quad words. Such an implementation might not be exploitable even if the timing delta of a mismatch was 1 second since you would have to brute-force 64-bits before seeing the timing delta increase due to a match. However, things aren’t so simple and using an optimized memcmp() does not always save you.

To prepare for our talk, we disassembled a few memcmp() implementations and found a surprising number were byte-based. This gives an attacker only 256 possibilities per guess, which is quite tractable. We decided to follow up with more detailed results for various operating systems.

We found two things:

  1. Even an optimized memcmp() usually leaks bytewise timing differences. In fact, it is often required by its API to do so.
  2. OS and compiler settings often select a bytewise compare even when a more-optimized version is available.

Anyone who is assuming particular compiler behavior is advised to review their disassembly. It’s almost always best to use a constant-time comparison instead of memcmp() when working with secret data.

Windows XP SP3 (32-bit)

We disassembled all .exes and DLLs in the system32 directory. All of them import memcmp() from either ntdll.dll or msvcrt.dll. All of them were compiled with some version of MSVC.

The msvcrt.dll memcmp() is 32-bit optimized. It compares 32-bit words if both pointers are 32-bit aligned (rep cmpsd). The remainder is compared with up to a 3-byte series of individual byte compares (cmp/jnz). If either of the input pointers is not 32-bit aligned, it goes into an unrolled loop that compares single bytes and exits early if any mismatch. ntdll.dll has the same memcmp() implementation.

Even with this optimization, the Windows memcmp() leaks bytewise timing information. When a difference is found in the 32-bit compares, it goes through each byte of the 32-bit word to find which byte differs. This means there is a 2-instruction timing delta per byte the attacker got correct, even though the original compare was in 32-bit chunks.

Windows memcmp() doesn’t actually have to do this set of bytewise comparisons. MSDN states the following behavior:

Return value Relationship of first count bytes of buf1 and buf2
< 0 buf1 less than buf2
0 buf1 identical to buf2
> 0 buf1 greater than buf2

The cygwin man page for memcmp() states:

The function returns an integer greater than, equal to or less than zero according to whether the object pointed to by S1 is greater than, equal to or less than the object pointed to by S2.

Since neither implementation requires identifying the bytewise difference, the implementation could just subtract the two 32-bit words. Only the sign of the return value matters, not the magnitude. This is analogous to treating the two buffers as multi-precision integers, just like in an RSA implementation.

We also compiled our own code using the cygwin and mingw gcc compilers. With no optimization flags, the cygwin gcc links to cygwin1.dll, which has a 32-bit optimized memcmp() similar to msvcrt.dll. It returns the bytewise difference and checks the input pointers for 32-bit alignment. The mingw compiler links to the msvcrt.dll version of memcmp() when no optimization flags are specified.

However, if an optimization level higher than default is used (-O1, -O2, -O3), both cygwin and mingw gcc use a 1-byte compare loop (rep cmpsb). This is surprising. If the -fno-builtin flag is specified, both cygwin and mingw gcc revert to the previous behavior, calling the external implementation in their respective DLLs.

We disassembled the default cygwin system utilities and libraries and found they all used a 1-byte compare loop. This is likely because they were compiled with optimization enabled but without the -fno-builtin flag.

Summary

The default Windows memcmp() (msvcrt.dll or ntdll.dll) leaks bytewise timing information even for chunks of 32-bits due to its search for the differing byte once a mismatch is found. It also leaks bytewise information if either input pointer is not 32-bit aligned. MSVC always uses an external DLL memcmp().

Both cygwin and mingw gcc use a 1-byte compare loop if optimization is enabled (-O1 or higher). If no optimization is enabled, gcc links to a DLL (cygwin1.dll or msvcrt.dll) that has a 32-bit optimized memcmp(). This memcmp() leaks the bytewise difference just like msvcrt.dll.

Windows 7 (64-bit)

This memcmp() implementation in msvcrt.dll compares buffers longer than 8 bytes in 64-bit chunks. It does not return the bytewise difference if it finds a mismatch. It returns the difference of the trailing 32-bits. The only bytewise timing leak occurs if the input length is not a multiple of 8. In this case, the remainder of the bytes are compared one at a time.

Linux (32 and 64-bit, Ubuntu 10.04 with glibc 2.12.1)

The x86 32-bit memcmp() is a long unrolled set of 32-bit compares. When it finds some difference, it jumps to find_diff which, like Windows, identifies exactly which byte differs. This leaks a bytewise timing delta.

Like Windows, the Linux memcmp() does not require finding the differing byte:

The memcmp() function returns an integer less than, equal to, or greater than zero if the first n bytes of s1 is found, respectively, to be less than, to match, or be greater than the first n bytes of s2.

The 64-bit memcmp() uses a tricky combination of a bit-set instruction (bsf) and shifts to identify the first byte difference without branching. So if the input is a multiple of 4, 8, or 16, it should not reveal a timing delta for any smaller units. If the input length is not a multiple of these values, smaller differences are visible.

Linux has optimized versions for other architectures as well, but we did not examine them.

Summary

Only the Linux 64-bit memcmp() does not reveal bytewise timing deltas. The 32-bit version behaves the same as Windows.

FreeBSD 8.0 (32 and 64-bit, gcc 4.2.1)

The system libc.so and libc.a have an optimized memcmp() for the most popular CPU architectures (i386, amd64, and arm). For i386, it consists of a 32-bit compare for the length in chunks of 4 (rep cmpsd), and a single-byte rep cmpsb for the remaining up to 3 bytes. There is no check for alignment first, so behavior with unaligned pointers is identical. The amd64 version has the same behavior as i386 but the libc memcmp() works in units of 8-byte quadwords (rep cmpsq).

If a difference is found in the word comparison, all implementations of memcmp() back up and find the first different byte using rep cmpsb. This required by the man page:

The memcmp() function returns zero if the two strings are identical, otherwise returns the difference between the first two differing bytes (treated as unsigned char values, so that ‘\200’ is greater than ‘\0’, for example). Zero-length strings are always identical.

For all other platforms (sparc, mips, ia64), the generic implementation of memcmp() is used. It is a loop of 1-byte compares, implemented in C.

GCC does not use the libc version of memcmp() by default. It uses a built-in version that consists of a rep cmpsb for the entire input. This occurred no matter what the optimization level (-O1, -O2, -O3). If we added the -fno-builtin flag, gcc generated calls to the libc memcmp(). Nothing changed based on optimization level.

We disassembled the default system utilities on FreeBSD 8.0 and found none of them were compiled with -fno-builtin, and thus all used the 1-byte memcmp.

Summary

By default, gcc on FreeBSD uses a completely timeable memcmp() that works in 1-byte units. This is regardless of optimization level. If the -fno-builtin flag is used, the libc memcmp() is linked in instead. The libc memcmp() is not more resistant to timing attacks because it finds the first differing byte. The sparc, mips, and ia64 platforms always use a 1-byte loop.

Exploiting remote timing attacks

We’ll be giving a Blackhat talk on exploiting remote timing attacks. Our goal is to convince developers that this class of attack is exploitable and worth fixing. This article in Computer World gives a decent background.

The attack is very simple. You repeatedly send guesses about a secret value to the server, which rejects them as incorrect. However, if your first byte of the guess is correct, it takes a very slightly longer time to return the error. With many measurements and some filtering, you can distinguish this difference.

While any comparison against a secret is a potential target, we chose HMAC comparison for a few reasons. An HMAC is a message authenticator, similar to a digital signature. The primary difference is that the the verifying party must keep the true HMAC secret since it gives the attacker the correct authenticator for their forged message. HMACs are used in many protocols, including most web authentication frameworks such as OAuth and OpenID and HTTP session cookies.

Guessing the correct HMAC for an arbitrary message is game over for these authentication frameworks. The token grants access to resources or allows the attacker to assume a user’s identity on various websites.

This is not a new attack. Remote timing attacks on OpenSSL were shown to be practical in 2003. Further research in 2007 showed that differences as small as 20 microseconds over the Internet and 100 nanoseconds over the LAN could be distinguished with about 1000 samples.

We (and others) have been reporting these flaws for over a year and raising developer awareness. In 2009, we found a timing leak in Google Keyczar‘s HMAC verification that was quickly fixed. Coda Hale found a similar flaw in Java’s MessageDigest implementation. The OAuth group discussed his bug back then and some maintainers decided to fix it in their code too. But many didn’t.

A quick review of OAuth and OpenID implementations showed many had timing leaks that were potentially exploitable. Either developers knew about the bug and gave it a low priority or they weren’t aware of it. Either way, we thought some concrete research was needed to show exactly how easy or hard it was to exploit these flaws in various environments.

Exploiting timing attacks depends on extracting a timing difference from many samples by filtering out the effect of noise. If there is too much noise (the difference is too small), this attack may take too long to be practical. But an attacker who can control the environment to decrease noise (say, by blocking competing users of the server), accurately model the noise and thus filter it better, or just wait longer because their target is so valuable might be successful.

Our talk builds most closely on the Crosby 2007 paper mentioned above. We have tested many configurations to find how different variables influence an attacker. The most obvious analysis is how small a time delta can be distinguished for a given number of samples. This was performed from various vantage points (guest-to-guest VM, LAN, Internet) and for various languages (C, Python, Ruby, Java, etc.)

We applied various filtering methods to the samples to see how much unfiltered jitter remained. This would determine how small a difference could be distinguished. We added in other variables such as competing load, power management, and other factors.

The talk will have the full results. Both the proponents and skeptics should be surprised in some way. We have found some configurations that are almost certainly not exploitable and others that certainly are. If you’re the maintainer of a software package, don’t count on your users being safe from timing attacks because of your assumptions. Cryptographic software, especially open-source, is deployed in everything from slow embedded servers on up to multi-Ghz clusters.

Likewise, attackers often have a better vantage point than you’d first assume. With shared hosting providers and cloud computing, you have to assume attackers can locate themselves on the same host as their target. Even in a shared datacenter, you may assume the attacker has a LAN-equivalent vantage point.

Given that it is difficult to rule out timing attacks and the fix is so simple, we think it’s best to fix them proactively. The approach that is easiest to gain assurance is to use a constant-time compare function. (That post also gives reasons why other approaches don’t work or are too complicated to verify).

We hope our talk will give some concrete results so that more developers will take this flaw seriously. See you in Vegas!

SSL gives point-to-point, not end-to-end security

At Summercon, Jon Oberheide gave a talk on Android security. He described a trojan app called RootStrap that checks for native code updates on a remote server and executes them. This is possible because Android doesn’t place any restrictions on native code that is installed with or later downloaded by an app. The only limitation is that the code runs as the app’s unprivileged UID, but there are no additional restrictions.

The more interesting part gives an overview how the app installation process interacts with GTalkService. In a followup post today, Jon gave more analysis of this installation mechanism. Unlike other parts of Android, this service is not open source so you have to disassemble the DEX file to see how it works.

When you select “install” on the Market app, the phone doesn’t download and install the app immediately. Instead, Google’s server sends an INSTALL_APP message to GTalkService on the phone, which downloads and installs the app. The message is delivered over SSL and includes a SHA-1 hash of the app’s installer (.apk). While this is better than no authentication, the link between the user’s action and the actual code installed is tenuous.

SSL provides good point-to-point privacy and integrity protection. However, there is no guarantee to upper layers that SSL is indeed in use. Few, if any, programs query the SSL layer to check the state of the connection, do additional cert validation, etc. Even when active, SSL provides point-to-point, not end-to-end security.

In today’s computing environment, there are seldom only two systems involved in a transaction. Even if the apps were stored on a single Google server, they are still compiled and signed on other systems. Anywhere along that production chain, a compromise could lead to apps being trojaned and surreptitiously pushed to many Android phones.

Android does provide some security in its code signing model. The developer’s signature on the .apk is basically a JAR signature. The hash of the APK cert is used to determine if a new app can access the same data as the previous app since it determines which UID an app gets. However, this only protects data created by existing apps from being accessed by other apps that are not signed with the same key. It also doesn’t say anything about the legitimacy of the code since the developer signs it themselves, often with a self-signed cert.

Since it appears that the INSTALL_APP message does not have any signature on itself, this message is not protected other than with SSL. Could an attacker who could inject some messages into the Google server replay this message, causing phones everywhere to install their malware? Will phones install apps without the Market service requesting it?

We’ll have to see what happens as more info is found out about GTalkService. The installation process should include a challenge/response value for liveness (perhaps this is the “tickle_id” field?) The installed APK should be linked to the phone’s install challenge with a Google signature. After all, Android ships with a list of CAs. Why can’t Google include some limited CA for their own domains to enable this signing?

This is a good example of how SSL only provides point-to-point, not end-to-end security. While SSL is great for transactions, additional protection is needed for application-level functions such as updates, especially in today’s multi-server environment.

Don’t use included version of IDAPython

I ran into a problem the other day with IDAPython. It’s a great tool but the version that comes by default with IDA is often out-of-date. Lesson: always update your version of IDAPython after installing or upgrading IDA.

The problem I saw was that the idautils.Functions() generator would never return any functions. I added lots of debugging prints and found that Segments() worked but Functions() never found a function, no matter what the range of addresses was. I then found that the Functions() routine would never location any functions if the first address was not the exact EA of a function.

This was in IDA 5.5 with its default IDAPython. Here’s the commit that fixed the bug. Since there are other bugs in older releases, I recommend updating to 1.3.2 (Windows binary or SVN source).