Random number generation
From wikipedia:Random number generation:
- A random number generator (RNG) is a computational or physical device designed to generate a sequence of numbers or symbols that lack any pattern, i.e. appear random.
Generation of random data is crucial for several applications like making cryptographic keys (e.g. for data-at-rest encryption), securely wiping disks, running encrypted Software access points.
Kernel built-in RNG
The Linux kernel's built-in RNGs /dev/{u}random are highly acclaimed for producing reliable random data providing the same security level that is used for the creation of cryptographic keys. The random number generator gathers environmental noise from device drivers and other sources into an entropy pool.
Note that the man random
command will misdirect to the library function manpage random(3) while for information about the /dev/random
device files you should run man 4 random
to read random(4). Additionally, random(7) gives an overview of the kernel's RNG interfaces.
/dev/random
/dev/random
uses an entropy pool of 4096 bits (512 Bytes) to generate random data and stops when the pool is exhausted until it gets (slowly) refilled. /dev/random
is designed for generating cryptographic keys (e.g. SSL, SSH, dm-crypt's LUKS), but it is impractical to use for wiping current HDD capacities: what makes disk wiping take so long is waiting for the system to gather enough true entropy. In an entropy-starved situation (e.g. a remote server) this might never end. While doing search operations on large directories or moving the mouse in X can slowly refill the entropy pool, it is designated pool size alone will be indication enough of the inadequacy for wiping a disk.
You can always compare /proc/sys/kernel/random/entropy_avail
against /proc/sys/kernel/random/poolsize
to keep an eye on the system's entropy pool.
While Linux kernel 2.4 did have writable /proc
entries for controlling the entropy pool size, in newer kernels only read_wakeup_threshold
and write_wakeup_threshold
are writable. The pool size is now hardcoded in kernel line 275 of /drivers/char/random.c
:
/* * Configuration information */ #define INPUT_POOL_WORDS 128 #define OUTPUT_POOL_WORDS 32 ...
The kernel's pool size is given by INPUT_POOL_WORDS * OUTPUT_POOL_WORDS
which makes, as already stated, 4096 bits.
/dev/random
to generate critical cryptographic keys on a system you do not control. If in doubt, for example in shared server environments, rather choose to create the keys on another system and transfer them. The cryptographer D. J. Bernstein illustrates the control problem with a Mark Twain quotation./dev/urandom
In contrast to /dev/random
, /dev/urandom
takes (since kernel 4.8) an initial random seed to feed a cryptographic stream cipher.[4]
The stream cipher is used because it can provide a high volume of high quality pseudo-random data suitable for many purposes like, for example, preparing for block device encryption, wiping LUKS keyslots, etc.
/dev/urandom
is not recommended for the generation of long-term cryptographic keys.The warning above keeps it simple, i.e. whether it applies or not largely depends on the system's status and purpose. For example, it matters considerably whether the system has just booted up, or the kernel has had time to gather entropy and fill the entropy pool.
Some posts which explain and discuss the different aspects are:
- Myths about urandom - highly informative but (unfortunately) also containing fallacies, and
- Safely generate random numbers - a post which illustrates what may happen when trying to single-handedly change the kernel infrastructure for gathering entropy.
For an Arch system, the above implies the following for /dev/urandom
:
- Use the kernel's
getrandom()
system call, whenever possible. It is provided to ensure that/dev/urandom
has initialized appropriately. See getrandom(2). - If you use the
/dev/urandom
device for generating entropy for any long-term purpose, do ensure it has initialized properly. - Take care to troubleshoot critical services which rely on
/dev/urandom
entropy during boot/startup. Even on a modern system the kernel's entropy initialization may take many seconds and some factors (e.g. virtualization) may delay it further. The kernel does warn, but not indefinetely.[5]
An example output:
# dmesg | grep -e random: ... [ 0.844818] random: systemd-udevd: uninitialized urandom read (16 bytes read) [ 7.022171] random: fast init done [ 12.916705] random: crng init done
In above example, the system has long reached its default boot target before the kernel gathered enough entropy to initialize the pool. Due to systemd requiring entropy at an early stage, it may happen though that the pool is depleted in the boot process without further kernel warnings.
Hence, a problem at this boot stage is that another service, e.g. an OpenSSL session for a web server or any other cryptographic service, may start without receiving a sufficient quality entropy seed from the depleted /dev/urandom
pool. Moving on, in a typical configuration OpenSSL may use that seed during the whole session's runtime, and only request a fresh seed for a new session.
Alternatives
For applications other than the generation of long-term cryptographic keys, a practical compromise between performance and security is the use of a pseudorandom number generator. In Arch Linux repositories for example:
There are also cryptographically secure pseudorandom number generators like Yarrow (macOS) or Fortuna (FreeBSD, the successor of Yarrow).
OpenSSL issues
OpenSSL supports RDRAND as rdrand
engine (use openssl engine
to view available OpenSSL engines), but OpenSSL and related applications (such as Nginx, Stunnel and CURL) may not use RDRAND by default. OpenSSL does not use the /dev/urandom
directly.
For example, you can pass -engine rdrand
option to the openssl
command line tools, or you can add ssl_engine rdrand;
line to the /etc/nginx/nginx.conf
for nginx.
See also
- RFC 4086 - Randomness Requirements for Security (Section 7.1.2 for /dev/random)
- Linux Kernel ML - discussion on patching /dev/random for higher throughput (February 2013)
- A challenge on /dev/random robustness (June 2013)
- An analysis of low entropy state behaviour of /dev/random, Yarrow, Fortuna and new model approach (March 2014)
- Randomness - A popular science article explaining different RNGs
- ENT - A simple program for testing random sequences (entropy, Chi square test, Monte Carlo, correlation, etc.)
- DIY HRNG - One example of a low-cost, DIY Arduino HRNG
- An Analysis of OpenSSL's Random Number Generator - Paper on RNG reseeding risks in OpenSSL functionality
- Linux Random Number Generator – A New Approach - Paper discussing updates to the devices (March 2017)