|Please consider subscribing to LWN|
Subscriptions are the lifeblood of LWN.net. If you appreciate this content and would like to see more of it, your subscription will help to ensure that LWN continues to thrive. Please visit this page to join up and keep LWN on the net.
H. Peter Anvin has been involved with Linux for more than 20 years, and is currently one of the x86 architecture maintainers. During his work on Linux, one of his areas of interest has been the generation and use of random numbers, and his talk at LinuxCon Europe 2012 was designed to address a lot of misunderstandings that he has encountered regarding random numbers. The topic is complex, and so, for the benefit of experts in the area, he noted that "I will be making some simplifications". (In addition, your editor has attempted to fill out some details that Peter mentioned only briefly, and so may have added some "simplifications" of his own.)
Possibly the first use of random numbers in computer programs was for games. Later, random numbers were used in Monte Carlo simulations, where randomness can be used to mimic physical processes. More recently, random numbers have become an essential component in security protocols.
Randomness is a subtle property. To illustrate this, Peter displayed a photograph of three icosahedral dice that he'd thrown at home, saying "here, if you need a random number, you can use 846". Why doesn't this work, he asked. First of all, a random number is only random once. In addition, it is only random until we know what it is. These facts are not the same thing. Peter noted that it is possible to misuse a random number by reusing it; this can lead to breaches in security protocols.
There are no tests for randomness. Indeed, there is a yet-to-be-proved mathematical conjecture that there are no tractable tests of randomness. On the other hand, there are tests for some kinds non-randomness. Peter noted that, for example, we can probably quickly deduce that the bit stream 101010101010… is not random. However, tests don't prove randomness: they simply show that we haven't detected any non-randomness. Writing reliability tests for random numbers requires an understanding of the random number source and the possible failure modes for that source.
Most games that require some randomness will work fine even if the source of randomness is poor. However, for some applications, the quality of the randomness source "matters a lot".
Why does getting good randomness matter? If the random numbers used to generate cryptographic keys are not really random, then the keys will be weak and subject to easier discovery by an attacker. Peter noted a couple of recent cases of poor random number handling in the Linux ecosystem. In one of these cases, a well-intentioned Debian developer reacted to a warning from a code analysis tool by removing a fundamental part of the random number generation in OpenSSL. As a result, Debian for a long time generated only one of 32,767 SSH keys. Enumerating that set of keys is, of course, a trivial task. The resulting security bug went unnoticed for more than a year. The problem is, of course, that unless you are testing for this sort of failure in the randomness source, good random numbers are hard to distinguish from bad random numbers. In another case, certain embedded Linux devices have been known to generate a key before they could generate good random numbers. A weakness along these lines allowed the Sony PS3 root key to be cracked [PDF] (see pages 122 to 130 of that presentation, and also this video of the presentation, starting at about 35'24").
Poor randomness can also be a problem for storage systems that depend on some form of probabilistically unique identifiers. Universally unique IDs (UUIDs) are the classic example. There is no theoretical guarantee that UUIDs are unique. However, if they are properly generated from truly random numbers, then, for all practical purposes, the chance of two UUIDs being the same is virtually zero. But if the source of random numbers is poor, this is no longer guaranteed.
Of course, computers are not random; hardware manufacturers go to great lengths to ensure that computers behave reliably and deterministically. So, applications need methods to generate random numbers. Peter then turned to a discussion of two types of random number generators (RNGs): pseudo-random number generators and hardware random number generators.
The classical solution for generating random numbers is the so-called pseudo-random number generator (PRNG):
A PRNG has two parts: a state, which is some set of bits determined by a "seed" value, and a chaotic process that operates on the state and updates it, at the same time producing an output sequence of numbers. In early PRNGs, the size of the state was very small, and the chaotic process simply multiplied the state by a constant and discarded the high bits. Modern PRNGs have a larger state, and the chaotic process is usually a cryptographic primitive. Because cryptographic primitives are usually fairly slow, PRNGs using non-cryptographic primitives are still sometimes employed in cases where speed is important.
The quality of PRNGs is evaluated on a number of criteria. For example, without knowing the seed and algorithm of the PRNG, is it possible to derive any statistical patterns in or make any predictions about the output stream? One statistical property of particular interest in a PRNG is its cycle length. The cycle length tells us how many numbers can be generated before the state returns to its initial value; from that point on, the PRNG will repeat its output sequence. Modern PRNGs generally have extremely long cycle lengths. However, some applications still use hardware-implemented PRNGs with short cycle lengths, because they don't really need high-quality randomness. Another property of PRNGs that is of importance in security applications is whether the PRNG algorithm is resistant to analysis: given the output stream, is it possible to figure out what the state is? If an attacker can do that, then it is possible to predict future output of the PRNG.
The output of a PRNG is only as good as its seed and algorithm, and while it may pass all known tests for non-randomness, it is not truly random. But, Peter noted, there is a source of true randomness in the world: quantum mechanics. This fact can be used to build hardware "true" random number generators.
A hardware random number generator (HWRNG) consists of a number of components, as shown in the following diagram:
Entropy is a measure of the disorder, or randomness, in a system. An entropy source is a device that "harvests" quantum randomness from a physical system. The process of harvesting the randomness may be simple or complex, but regardless of that, Peter said, you should have a good argument as to why the harvested information truly derives from a source of quantum randomness. Within a HWRNG, the entropy source is necessarily a hardware component, but the other components may be implemented in hardware or software. (In Peter's diagrams, redrawn and slightly modified for this article, yellow indicated a hardware component, and blue indicated a software component.)
Most entropy sources don't produce "good" random numbers. The source may, for example, produce ones only 25% of the time. This doesn't negate the value of the source. However, the "obvious" non-randomness must be eliminated; that is the task of the conditioner.
The output of the conditioner is then fed into a cryptographically secure pseudo-random number generator (CSPRNG). The reason for doing this is that we can better reason about the output of a CSPRNG; by contrast, it is difficult to reason about the output of the entropy source. Thus, it is possible to say that the resulting device is at least as secure as a CSPRNG, but, since we have a constant stream of new seeds, we can be confident that it is actually a better source of random numbers than a CSPRNG that is seeded less frequently.
The job of the integrity monitor is to detect failures in the entropy source. It addresses the problem that entropy sources can fail silently. For example, a circuit in the entropy source might pick up an induced signal from a wire on the same chip, with the result that the source outputs a predictable pattern. So, the job of the integrity monitor is to look for the kinds of failure modes that are typical of this kind of source; if failures are detected, the integrity monitor produces an error indication.
There are various properties of a HWRNG that are important to users. One of these is bandwidth—the rate at which output bits are produced. HWRNGs vary widely in bandwidth. At one end, the physical drum-and-ball systems used in some public lottery draws produce at most a few numbers per minute. At the other end, some electronic hardware random number sources can generate output at the rate of gigabits per second. Another important property of HWRNGs is resistance to observation. An attacker should not be able to look into the conditioner or CSPRNG and figure out the future state.
Peter then briefly looked at a couple of examples of entropy sources. One of these is the recent Bull Mountain Technology digital random number generator produced by his employer (Intel). This device contains a logic circuit that is forced into an impossible state between 0 and 1, until the circuit is released by a CPU clock cycle, at which point quantum thermal noise forces the circuit randomly to zero or one. Another example of a hardware random number source—one that has actually been used—is a lava lamp. The motion of the liquids inside a lava lamp is a random process driven by thermal noise. A digital camera can be used to extract that randomness.
The Linux kernel RNG has the following basic structure:
The structure consists of a two-level cascaded sequence of pools coupled with CSPRNGs. Each pool is a large group of bits which represents the current state of the random number generator. The CSPRNGs are currently based on SHA-1, but the kernel developers are considering a switch to SHA-3.
The kernel RNG produces two user-space output streams. One of these goes to /dev/urandom and also to the kernel itself; the latter is useful because there are uses for random numbers within the kernel. The other output stream goes to /dev/random. The difference between the two is that /dev/random tries to estimate how much entropy is coming into the system, and will throttle its output if there is insufficient entropy. By contrast, the /dev/urandom stream does not throttle output, and if users consume all of the available entropy, the interface degrades to a pure CSPRNG.
Starting with Linux 3.6, if the system provides an architectural HWRNG, then the kernel will XOR the output of the HWRNG with the output of each CSPRNG. (An architectural HWRNG is a complete random number generator designed into the hardware and guaranteed to be available in future generations of the chip. Such a HWRNG makes its output stream directly available to user space via dedicated assembler instructions.) Consequently, Peter said, the output will be even more secure. A member of the audience asked why the kernel couldn't just do away with the existing system and use the HWRNG directly. Peter responded that some people had been concerned that if the HWRNG turned out not to be good enough, then this would result in a security regression in the kernel. (It is also worth noting that some people have wondered whether the design of HWRNGs may have been influenced by certain large government agencies.)
The inputs for the kernel RNG are shown in this diagram:
In the absence of anything else, the kernel RNG will use the timings of events such as hardware interrupts as inputs. In addition, certain fixed values such as the network card's MAC address may be used to initialize the RNG, in order to ensure that, in the absence of any other input, different machines will at least seed a unique input value.
The rngd program is a user-space daemon whose job is to feed inputs (normally from the HWRNG driver, /dev/hwrng) to the kernel RNG input pool, after first performing some tests for non-randomness in the input. If there is HWRNG with a kernel driver, then rngd will use that as input source. In a virtual machine, the driver is also capable of taking random numbers from the host system via the virtio system. Starting with Linux 3.7, randomness can also be harvested from the HWRNG of the Trusted Platform Module (TPM) if the system has one. (In kernels before 3.7, rngd can access the TPM directly to obtain randomness, but doing so means that the TPM can't be used for any other purpose.) If the system has an architectural HWRNG, then rngd can harvest randomness from it directly, rather than going through the HWRNG driver.
"You really, really want to run rngd", Peter said. It should be started as early as possible during system boot-up, so that the applications have early access to the randomness that it provides.
One thing you should not do is the following:
rngd -r /dev/urandom
Peter noted that he had seen this command in several places on the web. Its effect is to connect the output of the kernel's RNG back into itself, fooling the kernel into believing it has an endless supply of entropy.
HAVEGE (HArdware Volatile Entropy Gathering and Expansion) is a piece of user-space software that claims to extract entropy from CPU nondeterminism. Having read a number of papers about HAVEGE, Peter said he had been unable to work out whether this was a "real thing". Most of the papers that he has read run along the lines, "we took the output from HAVEGE, and ran some tests on it and all of the tests passed". The problem with this sort of reasoning is the point that Peter made earlier: there are no tests for randomness, only for non-randomness.
One of Peter's colleagues replaced the random input source employed by HAVEGE with a constant stream of ones. All of the same tests passed. In other words, all that the test results are guaranteeing is that the HAVEGE developers have built a very good PRNG. It is possible that HAVEGE does generate some amount of randomness, Peter said. But the problem is that the proposed source of randomness is simply too complex to analyze; thus it is not possible to make a definitive statement about whether it is truly producing randomness. (By contrast, the HWRNGs that Peter described earlier have been analyzed to produce a quantum theoretical justification that they are producing true randomness.) "So, while I can't really recommend it, I can't not recommend it either." If you are going to run HAVEGE, Peter strongly recommended running it together with rngd, rather than as a replacement for it.
If you are writing applications that need to use a lot of randomness, you really want to use a cryptographic library such as OpenSSL, Peter said. Every cryptographic library has a component for dealing with random numbers. If you need just a little bit of randomness, then just use /dev/random or /dev/urandom. The difference between the two is how they behave when entropy is in short supply. Reads from /dev/random will block until further entropy is available. By contrast, reads from /dev/urandom will always immediately return data, but that data will degrade to a CSPRNG when entropy is exhausted. So, if you prefer that your application would fail rather than getting randomness that is not of the highest qualify, then use /dev/random. On the other hand, if you want to always be able to get a non-blocking stream of (possibly pseudo-) random data, then use /dev/urandom.
"Please conserve randomness!", Peter said. If you are running recent hardware that has a HWRNG, then there is a virtually unlimited supply of randomness. But, the majority of existing hardware does not have the benefit of a HWRNG. Don't use buffered I/O on /dev/random or /dev/urandom. The effect of performing a buffered read on one of these devices is to consume a large amount of the possibly limited entropy. For example, the C library stdio functions operate in buffered mode by default, and an initial read will consume 8192 bytes as the library fills its input buffer. A well written application should use non-buffered I/O and read only as much randomness as it needs.
Where possible, defer the extraction of randomness as late as possible in an application. For example, in a network server that needs randomness, it is preferable to defer extraction of that randomness until (say) the first client connect, rather extracting the randomness when the server starts. The problem is that most servers start early in the boot process, and at that point, there may be little entropy available, and many applications may be fighting to obtain some randomness. If the randomness is being extracted from /dev/urandom, then the effect will be that the randomness may degrade for a CSPRNG stream. If the randomness is being extracted from /dev/random, then reads will block until the system has generated enough entropy.
Another piece of work is the completion of the virtio-rng system. This feature is useful in a virtual-machine environment. Guest operating systems typically have few sources of entropy. The virtio-rng system is a mechanism for the host operating system to provide entropy to a guest operating system. The guest side of this work was already completed in 2008. However, work on the host side (i.e., QEMU and KVM) got stalled for various reasons; hopefully, patches to complete the host side will be merged quite soon, Peter said.
A couple of questions at the end of the talk concerned the problem of live cloning a virtual machine image, including its memory (i.e., the virtual machine is not shut down and rebooted). In this case, the randomness pool of the cloned kernel will be duplicated in each virtual machine, which is a security problem. There is currently no way to invalidate the randomness pool in the clones, by (say) setting the clone's entropy count to zero. Peter thinks a (kernel) solution to this problem is probably needed.
Interest in high-quality random numbers has increased in parallel with the increasing demands to secure stored data and network communications with high-quality cryptographic keys. However, as various examples in Peter's talk demonstrated, there are many pitfalls to be wary of when dealing with random numbers. As HWRNGs become more prevalent, the task of obtaining high-quality randomness will become easier. But even if such hardware becomes universally available, it will still be necessary to deal with legacy systems that don't have a HWRNG for a long time. Furthermore, even with a near infinite supply of randomness it is still possible to make subtle but dangerous errors when dealing with random numbers.
Copyright © 2012, Eklektix, Inc.
This article may be redistributed under the terms of the Creative Commons CC BY-SA 4.0 license
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds