haveged - Generate random numbers and feed Linux's random device.
generates an unpredictable stream of random numbers harvested
from the indirect effects of hardware events on hidden processor state
(caches, branch predictors, memory translation tables, etc) using the HAVEGE
(HArdware Volatile Entropy Gathering and Expansion) algorithm. The algorithm
operates in user space, no special privilege is required for file system
access to the output stream.
Linux pools randomness for distribution by the /dev/random and /dev/urandom
device interfaces. The standard mechanisms of filling the /dev/random pool may
not be sufficient to meet demand on systems with high needs or limited user
interaction. In those circumstances, haveged
may be run as a privileged
daemon to fill the /dev/random pool whenever the supply of random bits in
/dev/random falls below the low water mark of the device.
tunes itself to its environment and provides the same built-in
test suite for the output stream as used on certified hardware security
devices. See NOTES
below for further information.
- -b nnn, --buffer=nnn
- Set collection buffer size to nnn KW. Default is 128KW (or
- -d nnn, --data=nnn
- Set data cache size to nnn KB. Default is 16 or as
- -f file, --file=file
- Set output file path for non-daemon use. Default is
"sample", use "-" for stdout.
- -F , --Foreground
- Run daemon in foreground. Do not fork and detach.
- -i nnn, --inst=nnn
- Set instruction cache size to nnn KB. Default is 16 or as
- -n nnn, --number=nnn
- Set number of bytes written to the output file. The value
may be specified using one of the suffixes k, m, g, or t. The upper bound
of this value is "16t" (2^44 Bytes = 16TB). A value of 0
indicates unbounded output and forces output to stdout. This argument is
required if the daemon interface is not present. If the daemon interface
is present, this setting takes precedence over any --run value.
- -o <spec>, --onlinetest=<spec>
- Specify online tests to run. The <spec> consists of
optional "t"ot and "c"ontinuous groups, each group
indicates the procedures to be run, using "a<n>" to
indicate a AIS-31 procedure A variant, and "b" to indicate AIS
procedure B. The specifications are order independent (procedure B always
runs first in each group) and case insensitive. The a<n> variations
exist to mitigate the a slow autocorrelation test (test5). Normally all
procedure A tests, except the first are iterated 257 times. An a<n>
option indicates test5 should only be executed every modulo <n>
times during the procedure's 257 repetitions. The effect is so noticeable
that A8 is the usual choice.
The "tot" tests run only at initialization - there are no negative
performance consequences except for a slight increase in the time required
to initialize. The "tot" tests guarantee haveged has initialized
properly. The use of both test procedures in the "tot" test is
highly recommended because the two test emphasize different aspects of RNG
In continuous testing, the test sequence is cycled repeatedly. For example,
the string "tbca8b" (suitable for an AIS NTG.1 device) would run
procedure B for the "tot" test, then cycle between procedure A8
and procedure B continuously for all further output. Continuous testing
does not come for free, impacting both throughput and resource
consumption. Continual testing also opens up the possibility of a test
failure. A strict retry procedure recovers from spurious failure in all
but the most extreme circumstances. When the retry fails, operation will
terminate unless a "w" has been appended to the test token to
make the test advisory only. In our example above, the string
"tbca8wbw" would make all continuous tests advisory. For more
detailed information on AIS retries see NOTES below.
Complete control over the test configuration is provided for flexibility.
The defaults (ta8bcb" if run as a daemon and "ta8b"
otherwise) are suitable for most circumstances.
- -p file, --pidfile=file
- Set file path for the daemon pid file. Default is
- -r n, --run=n
- Set run level for daemon interface:
n = 0 Run as daemon - must be root. Fills /dev/random when the supply of
falls below the low water mark of the device.
n = 1 Display configuration info and terminate.
n > 1 Write <n> kb of output. Deprecated (use --number instead),
only provided for backward compatibility.
If --number is specified, values other than 0,1 are ignored. Default is
- -v n, --verbose=n
- Set diagnostic bitmap as sum of following options:
1=Show build/tuning summary on termination, summary for online test retries.
2=Show online test retry details
4=Show timing for collections
8=Show collection loop layout
16=Show collection loop code offsets
32=Show all online test completion detail
Default is 0. Use -1 for all diagnostics.
- -w nnn, --write=nnn
- Set write_wakeup_threshold of daemon interface to nnn bits.
Applies only to run level 0.
- -?, --help
- This summary of program options.
haveged tunes the HAVEGE algorithm for maximum effectiveness using a hierarchy
of defaults, command line options, virtual file system information, and cpuid
information where available. Under most circumstances, user input is not
required for excellent results.
Run-time testing provides assurance of correct haveged operation. The run-time
test suite is modeled upon the AIS-31 specification of the German Common
Criteria body, BIS. This specification is typically applied to hardware
devices, requiring formal certification and mandated start-up and continuous
operational testing. Because haveged runs on many different hardware
platforms, certification cannot be a goal, but the AIS-31 test suite provides
the means to assess haveged output with the same operational tests applied to
certified hardware devices.
AIS test procedure A performs 6 tests to check for statistically inconspicuous
behavior. AIS test procedure B performs more theoretical tests such as
checking multi-step transition probabilities and making an empirical entropy
estimate. Procedure A is the much more resource and compute intensive of the
two but is still recommended for the haveged start-up tests. Procedure B is
well suited to use of haveged as a daemon because the test entropy estimate
confirms the entropy estimate haveged uses when adding entropy to the
No test is perfect. There is a 10e-4 probability that a perfect generator will
fail either of the test procedures. AIS-31 mandates a strict retry policy to
filter out false alarms and haveged always logs test procedure failures.
Retries are expected but rarely observed except when large data sets are
generated with continuous testing. See the libhavege(3)
notes for more
If running as a daemon, access to the following files is required
Haveged returns 0 for success and non-zero for failure. The failure return code
is 1 "general failure" unless execution is terminated by signal
<n>, in which case the return code will be 128 + <n>. The
following diagnostics are issued to stderr upon non-zero termination:
Cannot fork into the background
Call to daemon(3) failed.
Cannot open file <s> for writing.
Could not open sample file <s> for
Cannot write data in file:
Could not write data to the sample file.
Couldn't get pool size.
Unable to read
Couldn't initialize HAVEGE rng
Invalid data or instruction cache size.
Couldn't open PID file <s> for writing
Unable to write daemon PID
Couldn't open random device
Could not open /dev/random for read-write.
Couldn't query entropy-level from kernel: error
Couldn't open PID file <path> for writing
Error writing /var/run/haveged.pid
Unable to write to
Call to ioctl(2) to add entropy failed
The random number generator failed self-test
or encountered a fatal error.
Call to select(2) failed.
Stopping due to signal <n>
Unable to setup online tests
Memory unavailable for online test resources.
- Write 1.5MB of random data to the file /tmp/random
- haveged -n 1.5M -f /tmp/random
- Generate a /tmp/keyfile for disk encryption with LUKS
- haveged -n 2048 -f /tmp/keyfile
- Overwrite partition /dev/sda1 with random data. Be careful,
all data on the partition will be lost!
- haveged -n 0 | dd of=/dev/sda1
- Generate random ASCII passwords of the length 16
- (haveged -n 1000 -f - 2>/dev/null | tr -cd '[:graph:]' |
fold -w 16 && echo ) | head
- Write endless stream of random bytes to the pipe. Utility
pv measures the speed by which data are written to the pipe.
- haveged -n 0 | pv > /dev/null
- Evaluate speed of haveged to generate 1GB of random
- haveged -n 1g -f - | dd of=/dev/null
- Create a random key file containing 65 random keys for the
encryption program aespipe.
- haveged -n 3705 -f - 2>/dev/null | uuencode -m - | head
-n 66 | tail -n 65
- Test the randomness of the generated data with dieharder
- haveged -n 0 | dieharder -g 200 -a
- Generate 16k of data, testing with procedure A and B with
detailed test results. No c result seen because a single buffer fill did not
contain enough data to complete the test.
- haveged -n 16k -o tba8ca8 -v 33
- Generate 16k of data as above with larger buffer. The c
test now completes - enough data now generated to complete the test.
- haveged -n 16k -o tba8ca8 -v 33 -b 512
- Generate 16m of data as above, observe many c test
completions with default buffer size.
- haveged -n 16m -o tba8ca8 -v 33
- Generate large amounts of data - in this case 16TB. Enable
initialization test but made continuous tests advisory only to avoid a
possible situation that program will terminate because of procedureB failing
two times in a row. The probability of procedureB to fail two times in a row
can be estimated as <TB to generate>/3000 which yields 0.5% for
- haveged -n 16T -o tba8cbw -f - | pv > /dev/null
- Generate large amounts of data (16TB). Disable continuous
tests for the maximum throughput but run the online tests at the startup to
make sure that generator for properly initialized:
- haveged -n 16T -o tba8c -f - | pv > /dev/null
HArdware Volatile Entropy Gathering and Expansion: generating unpredictable
random numbers at user level
- cryptsetup(8), aespipe(1), pv(1),
by A. Seznec, N. Sendrier, INRIA Research
Report, RR-4592, October 2002
A proposal for: Functionality classes for random number generators
Killmann and W. Schindler, version 2.0, Bundesamt fur Sicherheit in der
Informationstechnik (BSI), September, 2011
A Statistical Test Suite for the Validation of Random NUmber Generators and
Pseudorandom Number Generators for Cryptographic Applications,
publication SP800-22, National Institute of Standards and Technology, revised
Additional information can also be found at
Gary Wuertz <firstname.lastname@example.org> and Jirka Hladky <hladky jiri AT gmail