Random number generation

From ArchWiki
The printable version is no longer supported and may have rendering errors. Please update your browser bookmarks and please use the default browser print function instead.

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).

/dev/random

Tango-view-refresh-red.pngThis article or section is out of date.Tango-view-refresh-red.png

Reason: /dev/random works similarly to /dev/urandom since v5.6[1][2] (Discuss in Talk:Random number generation)

/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.

Warning: Do not use even /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.[3]

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.

Warning: /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:

  1. Use the kernel's getrandom() system call, whenever possible. It is provided to ensure that /dev/urandom has initialized appropriately. See getrandom(2).
  2. If you use the /dev/urandom device for generating entropy for any long-term purpose, do ensure it has initialized properly.
  3. 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.[4]

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

Tango-view-fullscreen.pngThis article or section needs expansion.Tango-view-fullscreen.png

Reason: No explanation of the actual issues. (Discuss in Talk:Random number generation)

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