Home About USENIX Events Membership Publications Students
BSDCon 2002 Paper    [BSDCon '02 Tech Program Index]

Pp. 47–54 of the Proceedings

An implementation of the Yarrow PRNG for FreeBSD

Mark R. V. Murray
FreeBSD Services, Ltd
Amersham, Buckinghamshire, UK


Computers are by their definition predictable. The problem of obtaining good-quality random numbers is well known.

There is a great need for entropy in the running kernel, as well as in user-space. The kernel needs to randomise TCP sequences, seed keys for IPSec, randomise PIDs, and so on. Starvation of these random numbers is a critical problem. Users need random keys, random filenames, nondeterministic games, random numbers for Monte-Carlo simulation and so on.

Kelsey, Schneier and Ferguson proposed an improved algorithm for providing statistically random numbers, at the same time cryptographically protecting their sequence and state. This is the Yarrow algorithm.

This work presents an implementation of this algorithm as the entropy device (/dev/random) in FreeBSD's kernel.

1  Introduction

In an earlier work[Mur00], the author introduced the new entropy device to FreeBSD-CURRENT as a work-in progress. In that work, attack methodologies were briefly discussed, and the difference between the older entropy device and this device were discussed. Yarrow[KSF99] was briefly explained.

It is important to remember that this device is not designed to produce pure1 random numbers. Computers do not produce enough natural randomness for that approach to be useful in entropy-consuming environments.

Instead, this device is a free-running pseudo-random number generator (PRNG), one in which great effort has been made to cryptographically protect the state of the generator. Further, the internal state is constantly perturbed with ``harvested'' entropy to thwart attackers.

The algorithm is divided into four parts (see Figure 1):

#1 Simplified Yarrow Structure

  • Entropy Accumulator and Pools
    These are used to ``harvest'' entropy from the running kernel. The API provided by the author is intended to be simple to use anywhere in the kernel.
  • Reseed
    Reseeding is entirely internal to Yarrow. The author has attempted to stay as close as possible to the published algorithm.
  • Reseed Control
    Reseeds happen in response to harvested entropy, and to reads from the entropy device. There are statistical requirements to these reseeds that are unimplemented.
  • Output Generator
    The generator is similar to ``classic'' PRNG's, excepting:

    1. It uses a large, cryptographically secure hash instead of a simple feedback formula.
    2. It is perturbed on a regular basis by harvested entropy.

2  Design Issues

An API for ``harvesting'' entropy was needed, so that kernel programmers could easily provide such randomness their subsystem could produce. The requirements were that the API should be extensible, fast, simple and able to operate in interrupt context. Where practical, entropy sources needed the ability to be disabled at the whim of the system administrator.

256-Bit storage pools were desired, as this was deemed to hold a reasonable amount of entropy without being overly expensive. It should be remembered that Yarrow uses two accumulation ``pools'' (fast and slow), so this meant that up to 512 bits of environmental entropy could be held.

This decision meant that a 256-bit hashing algorithm and a 256-bit block cipher were needed. The need for a 256-bit hash ruled out using MD2, MD4, MD5 or SHA-1 unless a lengthening algorithm was also used. There were a few choices for 256-bit block ciphers, however availability (or potential availability) in the FreeBSD kernel was a limiting factor. As a suitable ``natural'' hash did not exist, a hash had to be constructed using block ciphers. Likely candidates were initially Blowfish and DES (reluctantly, as a block-lengthening process would be needed). Other AES candidates were considered, but as a finalist had not been selected they were not initially used.

The output generator needed to be fast, and also needed good key-setup speed, as the key is changed often. In order to preserve the strength of Yarrow, its block size was deemed to be the same size as the hash buffer. This made the choice of the encryption cipher simple, as the hash cipher could be used.

Further research[Sch96a] indicated that lengthening algorithms were most probably unwise.

2.1  Entropy Harvesting

As entropy could be found in any part of the kernel, both bottom-half and top-half, the entropy harvesting needed to be cheap, non-invasive and non-blocking.

A fixed-size circular buffer is used to accumulate entropy for later processing. If the buffer becomes full, further attempts to add entropy are ignored. The buffer is never locked when written to; this does not matter, as data corruption would be beneficial.

Entropy is added to the buffer by a subsystem calling the random_harvest(9) function. This is declared in sys/random.h as follows:

enum esource { \
void random_harvest(void *data, \
    u_int count, u_int bits, \
    u_int frac, enum esource source);

Entropy is accumulated in up to HARVESTSIZE2 byte chunks.

The arguments are:

a pointer to the stochastic data
the number of bytes of data
an estimate of the random bits
as above, except fractional ([ frac/1024] bits)
the source of the entropy

The stochastic events added to the buffer are stored in a structure:

struct harvest {
        u_int64_t somecounter;
        u_char entropy[HARVESTSIZE];
        u_int size, bits, frac;
        enum esource source;

The structure holds all of the information provided by random_harvest plus a timestamp.

The timestamp is taken from the CPU's fast counter register (like the Intel Pentium(tm) processor's TSC register). CPUs that do not have this register (like the Intel i386) use nanotime(9) instead. This has an unfortunate time penalty.

It is not important that this timestamp is an accurate reflection of real-world time, nor is it important that multiple CPUs in an SMP environment would have different values. It is important that the counter/timestamp increase quickly and linearly with time.

A count of accumulated entropy is kept, and this is used to reseed the output generator on occasion. The fractional entropy count supplied in the frac parameter is used in very low entropy situations. For example, a particular device can be said to produce 1 bit of randomness every 20 events.

Kernel programmers wishing to supply entropy from their code should extend the enum esource list, leaving the constant at the end of the list. Then, the randomness should be gathered and supplied as efficiently as possible.

In sys/random.h:

enum esource {
    RANDOM_MYSTUFF, /* New */

In the code to be harvested:

#include <sys/types>
#include <sys/random>

    struct {
        u_int32_t junk;
        u_int32_t garbage;
        u_char    rubbish[8];
    } randomstuff;

    randomstuff.junk = somelocaljunk;
    randomstuff.garbage = otherjunk;
    strncpy(randomstuff.rubbish, dirt, 8);
    /* harvest the entropy in
     * randomstuff. Be really
     * conservative and estimate the
     * the random bit count as only 4.
        sizeof(randomstuff), 4, 0,

If control over the new harvesting is required, then a sysctl may be added to src/sys/dev/random/randomdev.[ch]:

    OID_AUTO, interrupt,
    &harvest.mystuff, 0,
    random_check_boolean, "I",
    "Harvest mystuff entropy");

The call to random_harvest should then be made conditional on harvest.mystuff:

    if (random.mystuff)
            sizeof(randomstuff), 4, 0,

Writing to the entropy device from the user's perspective (ie, writing to /dev/random) is similar to writing to /dev/null; it has no discernible effect. In actual fact, the data written is ``harvested'' using the harvesting calls, with the proviso that the entropy is estimated to be nothing. This has the effect of not causing reseeds, but perturbing the internal state anyway. If the user is the superuser, then closing the device after a write will cause an explicit reseed.

A kernel thread ``kthread'' constantly runs, polling the circular buffer, and if data is present, it accumulates each event alternately into the two accumulation hashes (or ``entropy pools'').

2.2  Accumulation Pools

An initial version of the 256-bit accumulation hash was constructed using a Davies-Meyer[Sch96b] hash with Blowfish[Sch96c] as the block cipher.

The hash works by repeatedly encrypting an initial (zero) state while cycling the hash data through the key. At each iteration, the previous value of the hash is exclusive-or-ed into the newly encrypted value.

This can be represented as:
Hi = EMi(Hi-1) XOR Hi-1

where Hn is the nth iteration of the hash result, Mj is the jth fragment of the data to be hashed and Ek(m) is the result of encrypting m with block cipher E() and key k.

Davies-Meyer diagram

Davies-Meyer hash from block cipher.

While this worked, it was unbearably slow as Blowfish has an extremely expensive key schedule. Slowness was experienced as very bad kernel latency, and a kernel thread running with unacceptably high CPU usage.

The (by this time) newly released AES (``Rijndael'')[NIS] algorithm was then tried, and a crude benchmark produced extremely promising results. (Here, Blowfish was replaced with Rijndael.)

The benchmark is a timed 16MB read from each device:

$ dd if=${DEVICE} of=/dev/null \
        count=16 bs=1048576

For comparison, /dev/zero was also read.

The time is the time in seconds for the 16MB read, and the rate is measured in KB/s.

$device Time (s) Rate (kB/s)
Blowfish 137.7 122
AES 6.5 2595
Zero 0.2 81861

After consulting literature [SKW+][WSB][FKL+], it was suspected that AES was the ideal algorithm, but further investigation was considered prudent, particularly as the benchmark measured output performance, not hashing performance.

The hash routines were broken out of the kernel, and various speeds were measured using alternative block ciphers. A Null algorithm and 160-bit SHA-1 were included for comparison.

The ``Null'' cipher simply duplicates the input data, ignoring the key:

Nk(m) = m

This reduced the Davies-Meyer algorithm to the XOR and data-movement parts only.

Each result represents the time taken to hash 2MB of pseudo-random data.

Algorithm Time (s) Rate (kB/s)
AES 3.1 461.6
Blowfish 40.2 35.2
DES 2.9 491.7
SHA-1 2.0 693.3
Null 1.8 786.7

It can be seen that AES with 256-bit keys and 256-bit blocks is approximately as fast as DES with 56-bit keys and 64-bit blocks.

160-bit SHA-1 is about 50% faster than the AES hash, but the AES hash has an approximately 50% larger capacity for storing bits.

The ``Null'' algorithm confirms that encryption overhead is acceptably low in comparison with other code overhead.

2.3  Output Generator

The output generator is a counter that is repeatedly encrypted, producing the output:

Output Generator diagram

The Output Generator

Oi = Ek(Ct)
Ct+1 = Ct+1

where Ct is the (256-bit) counter3 at time t, Oi is the ith output, and Ek(Ct) is the result of encrypting counter Ct using cipher E() and key k.

The dashed line represents the data path during a gate event. The key ``k'' is inserted during a reseed. This is the point at which environmental noise (``harvested'' entropy) is used.

To compromise the output generator, a key compromise of the cipher is necessary. This is computationally difficult; nevertheless, to thwart this, the counter is regularly replaced with data from the output stream:

Ct+1 = Ct+1
Ct+1 = Ek(C)

The data thus used is not used as part of the output. This is called a gate event, and it happens at a time configurable by the system administrator via sysctl(9). It defaults to happening every 10 blocks. If a user process reads less than a 256-bit block, the remainder is cached for future reads.

To show that the output was statistically acceptable, some tests were done.

A simple histogram of 8M single-byte values was plotted:

Plot 1

Spectrum of 8M 8-bit values

A straight line was fitted to this data, and was found to substantiate the fact that the slope was 0 and the mean value was 32k.

The spread of values around 32k was plotted, and the distribution found to be reassuringly normal:

Plot 2

Distribution of values around expected norm of 32k

This corresponded to a mean (m) of 32759.5 and a s of 187.2

Further tests were done using a more sophisticated random number ``torture chamber'' called Diehard[Mar]. Its use produced voluminous output which indicated, on careful perusal, that the generator's output was statistically acceptable.

It must be noted that the output generator does not block. This is intentional.

2.4  Reseed Control

This is the trickiest part of the algorithm to write. The Yarrow specification mandates three separate estimates of incoming entropy ``harvest-units'':

  1. A programmer-supplied estimate. This has been very conservatively set. This is given as a constant to each entropy-harvesting call.
  2. A system-wide ``density''. This is set at [ 1/2], meaning no sample of N bits can supply more than [ N/2] bits of entropy.
  3. A statistically determined, per-source continuous estimate. This is unimplemented, as the mechanism for doing the statistical estimation has been deemed too expensive for the kernel.

The algorithm states that the lowest of these three is taken as the entropy supplied for the individual unit. The author has endeavoured to ensure that the programmer-supplied estimate will always be low enough.

3  Impact on the Running System

The running device has great potential to be very invasive to the running kernel, as early experiments with slow ciphers showed. In the current code, however, the system is proving to be no such hindrance.


last pid: 19524;  load averages:  0.25,  0.22,  0.18 up 3+09:01:43  21:52:53
92 processes:  3 running, 74 sleeping, 15 waiting
CPU states:  4.3% user,  0.0% nice,  2.3% system,  0.4% interrupt, 93.0% idle
Mem: 27M Active, 5536K Inact, 18M Wired, 4348K Cache, 14M Buf, 4856K Free
Swap: 68M Total, 34M Used, 33M Free, 50% Inuse
   10 root     -16    0     0K    12K RUN    59.9H 86.47% 86.47% idle
18128 root      96    0 16636K  5932K select 15:37  1.61%  1.61% XFree86
18169 mark      96    0 15884K  3216K select  4:53  1.61%  1.61% kdeinit
18217 mark      96    0 17116K  4540K select  1:47  0.39%  0.39% kdeinit
18227 mark      96    0 10872K  5916K select  1:59  0.29%  0.29% xemacs-21.1
19524 mark      96    0  2096K  1144K RUN     0:00  0.75%  0.20% top
   22 root     -64 -183     0K    12K WAIT   40:57  0.10%  0.10% irq14: ata0
   12 root     -48 -167     0K    12K RUN    20:06  0.10%  0.10% swi6: tty:sio
18205 mark      96    0 17836K  5928K select  1:38  0.10%  0.10% kdeinit
18203 mark      96    0 21428K  4080K select  1:07  0.10%  0.10% kdeinit
    6 root      20    0     0K    12K syncer  4:02  0.00%  0.00% syncer
   14 root      76    0     0K    12K sleep   3:19  0.00%  0.00% random
18183 mark      60  -36  5244K  2044K select  3:09  0.00%  0.00% artsd
   15 root     -28 -147     0K    12K WAIT    2:58  0.00%  0.00% swi5: task qu
18810 mark      96    0  9728K  3076K select  1:30  0.00%  0.00% acroread
18208 mark      96    0 16372K  3964K select  1:21  0.00%  0.00% kdeinit

#1Snapshot of a running system

This snapshot of a running FreeBSD workstation shows that the random process (the kthread that runs the reseed process) has approximately the same impact on the system as the syncer process, ie negligible.

The use of random numbers by security-conscious engineers has been taken into account over and above the concerns of the professional cryptographic community. Speed was deemed to be more important than the production of number-theoretic-quality random numbers (eg: suited to generating one-time-pads). It is believed that FreeBSD is used by many more system-administrators than professional cryptographers.

The author is, however, appreciative of the concerns of those who would want a more austere presentation of random numbers from the operating system. Those members of the community are considered to be a specialist minority, though.

4  Future plans

There are two main expansion areas in the FreeBSD entropy device.

  1. More entropy harvesting. Any ``cheap'' entropy that may be found in the kernel may be used in the future. The user community is encouraged to submit likely sources. The author has provisional code to harvest entropy from Intel chipsets fitted with hardware random number generators.
  2. Provision of a ``distilled'' device for those who wish to be assured of an ``entropy-in = entropy-out'' conservation-of-entropy device. This needs to be conservative enough to not provide a denial-of-service attack by its very existence.

5  Thanks

Thanks are due to Sue Bourne and Brian Somers for proofreading and helpful comments.

Thanks are also due to FreeBSD Services, Ltd for giving me the time to produce this work.

My fondest thanks are also given to my father. Thanks, Dad. I'll miss you.


Niels Ferguson, John Kelsey, Stefan Lucks, Bruce Schneier, Mike Stay, David Wagner, and Doug Whiting. Improved cryptanalysis of rijndael. http://www.counterpane.com.

John Kelsey, Bruce Schneier, and Niels Ferguson. Yarrow-160: Notes on the design and analysis of the yarrow cryptographic pseudorandom number generator. Sixth Annual Workshop on Selected Areas in Cryptography, August 1999.

George Marsaglia. Diehard. http://www/stat.fsu.edu/[ \tilde]geo/diehard.html.

Mark R. V. Murray. Effective entropy from the freebsd kernel. In BSDCon, pages 92-98, 2000.

NIST. The aes algorithm (rijndael) information. http://csrc.nist.gov/encryption/aes/rijndael/.

Bruce Schneier. Applied Cryptography, pages 430-431. Wiley, second edition, 1996.

Bruce Schneier. Applied Cryptography, pages 446-455. Wiley, second edition, 1996.

Bruce Schneier. Applied Cryptography, pages 336-339. Wiley, second edition, 1996.

Bruce Schneier, John Kelsey, Doug Whiting, David Wagner, and Chris Hall. Performance comparison of the aes submissions. http://www.counterpane.com.

Doug Whiting, Bruce Schneier, and Steve Bellovin. Aes key agility issues in high-speed ipsec implementations. http://www.counterpane.com.


1In the number-theoretic sense; the numbers remain statistically random and include environmental noise

2Currently 16

3Internal to the FreeBSD kernel, the 256-bit value is represented as a structure containing four 64-bit unsigned integers. Only 64 bits are incremented. The author does not believe this is a problem.

This paper was originally published in the Proceedings of the BSDCon '02 Conference on File and Storage Technologies, February 11-14, 2002, Cathedral Hill Hotel, San Francisco, California, USA.
Last changed: 28 Dec. 2001 ml
Technical Program
BSDCon 2002 Home