bit-babbler (0.8) unstable; urgency=medium
* Support hotplugging devices into libvirt guest domains which have names
containing characters that are not valid as part of a shell variable name.
Another reminder that the important part of keeping things as simple as
possible is always the "as possible" bit.
* Support reading seedd(1) options from a configuration file. The original
design plan explicitly avoided this, partly just to keep the code as
simple and easy to audit as possible, and partly because it was desirable
to make invocation as simple and foolproof as possible. The more options
that something has, the easier it is to make some mistake with running it
which could have subtle and even serious consequences. But we are at the
point now where there are enough real alternative options which are either
genuinely desirable or needed for some use case, that the balance becomes
weighted toward being able to keep persistent configuration settings in a
file rather than having to spell them out on the command line each time.
The final straw for making this change now was the inability of systemd to
sanely support the existing simplified configuration interface that was
provided in /etc/default/seedd for the SysV init script. When given the
alternative choices available to us of either adding a shell wrapper to
do what systemd could not, or forcing people to manually edit or override
the systemd unit directly to make any configuration change, this was
clearly the Lesser Evil to embrace if we were going to provide a native
systemd unit for the system daemon. The former gains us nothing over the
existing LSB init script, and the latter would require every user to first
have a solid grasp of all the non-obvious consequences which can come into
play when configuring a system which (according to systemd.directives(7))
"contains 2464 entries in 13 sections, referring to 241 individual manual
pages" - and where even package maintainers and systemd upstream still
make mistakes that can take a long time for the real consequences to be
noticed. So if we were to provide a systemd unit, it needs to be well
tested and give people few, if any, reasons to ever need to modify it.
* Preserve existing configuration on package upgrades. The new default
configuration file behaves the same way as the old defaults did. If the
settings in /etc/default/seedd have been customised, then on upgrade we
generate a custom /etc/bit-babbler/seedd.conf implementing the same set
of options. The old customised file content will be retained, and can
be found in /etc/default/seedd.dpkg-old, in case there was anything else
in it which people might also want to keep, but after checking for that
it can safely be removed by the system admin. Nothing from this package
uses files in /etc/default from this version onward.
* Two systemd unit files are now included in this package, but only one is
enabled by default.
The seedd.service unit provides the same functionality as the SysV init
script does, and will be used instead of it on systems where systemd is
running as the init process. It will start the seedd(1) daemon as soon
as possible during boot, reading its options from the new configuration
file, and if feeding entropy to the kernel it will begin doing so as soon
as the available USB devices are announced to the system by udev.
The seedd-wait.service oneshot unit is not enabled by default. It provides
a simple sequence point which may be used to ensure that QA checked seed
entropy from available BitBabbler devices can be mixed into the kernel's
pool before other ordinary services which might rely upon it are started.
This is its default behaviour if it is simply enabled, and ordinarily it
will not delay the boot for very long, only until udev announces a device
that we can read some good seed bits from. By default this will time out
after 30 seconds if good entropy cannot be obtained, which should be more
than enough time to get a good seed if that was going to be possible, but
won't completely cripple the system when it is acceptable for it to still
be running without having a working BitBabbler attached.
Additionally, the seedd-wait.service can also be used to place a harder
constraint on individual services, if there are particular things which
the local admin does not want started at all if good seed entropy was not
obtained. Or it can be configured to divert the boot to a degraded mode
(such as the single-user mode emergency.target) if the availability of
good entropy from a BitBabbler should be a hard requirement for the whole
system. For more details of its use see the BOOT SEQUENCING section of
the seedd(1) manual page.
-- Ron Lee <firstname.lastname@example.org> Thu, 08 Feb 2018 10:26:52 +1030
bit-babbler (0.7) unstable; urgency=medium
* Handle the oddball case of a RHEL/CentOS 6 kernel being used with libusb
version 1.0.13 or later. They backported the USBDEVFS_GET_CAPABILITIES
API, but not the patch which was applied to the mainline kernel at the
same time to add USBDEVFS_CAP_BULK_SCATTER_GATHER. With the default
libusb 1.0.9 that it normally ships with this doesn't matter because none
of it is supported there anyway, but updating it makes this become a real
problem that we need to deal with.
* Test for SIGRTMIN, MacOSX doesn't have it.
* Test for pthread_setname_np by signature, it's implemented differently on
different platforms. Also support pthread_set_name_np which is what is
used by OpenBSD and FreeBSD.
* Explicitly set the pthread stack size on (more) platforms where it is tiny
by default. We do create some large structures on the stack, and it's
probably better to have a consistent size on all platforms than have some
of them smash it "unexpectedly".
* Querying the string length with strftime is a GNU extension, so only use
that where it's actually available.
* Provide an implementation of FeedKernelEntropy for MacOS. It does have a
documented interface for that, even if the implementation of it that is in
its kernel source as of Sierra is ... let's go with enlightening.
* Rename _P(), our convenience alias to ngettext, to P_(). On OpenBSD 6.1
_P is defined in ctype.h, and regardless of what you might think of that,
symbols starting with an underscore are reserved. So they win this one.
* Test for LOG_MAKEPRI, it isn't required by POSIX, and OpenBSD 6.1 doesn't
* Add explicit tests for strtod_l and newlocale. The newlocale function is
specified by POSIX, but OpenBSD doesn't provide it. We can get away with
falling back to strtod there because it only has very limited support for
locales anyway, and right now it will always be either C or en_US.UTF8,
both of which use '.' as the decimal point.
* Support systems without abi::__forced_unwind for thread cancellation stack
unwinding and clean up. The GCC toolchain on OpenBSD doesn't support it,
and neither does clang on FreeBSD or MacOS.
* Disable thread cancellation around calls to vfprintf on OpenBSD. On that
platform it is a cancellation point (as expected), but if cancellation
does occur there on 6.1, it can leave its internal _thread_flockfile mutex
locked which means any future calls to vfprintf (or anything else needing
that lock) will deadlock. We can't easily test for that bug, so we just
always provide our own safe cancellation point instead on that platform.
* Support unordered_map in both std and std::tr1 namespaces. Normally it
isn't available in std:: unless C++11 support is enabled, but clang on
FreeBSD provides the tr1 header as a symlink to the std one, and only
provides the template in the std namespace.
* Test for libusb_has_capability. FreeBSD 11 bumped LIBUSB_API_VERSION to
0x01000102, but didn't add the libusb_has_capability() function which was
part of that interface version. It did however add the hotplug API, and
it does mostly work, so if the API version is sufficient, but we can't
test for this capability, then we'll assume it's available and that it
will handle a call to it gracefully if for some reason it really isn't.
* Deal with FreeBSD 11 libusb start up and shut down delays. In our testing
there was a delay of around 4 seconds before it would report any existing
devices when LIBUSB_HOTPLUG_ENUMERATE was used with the hotplug callback
(and a similar delay when new devices really were hotplugged later). That
makes things awkward for the --scan option, which would see no devices at
all before returning its results, unless it too had an arbitrary (and user
unfriendly long) delay before responding. So we explicitly enumerate the
initial set ourselves on FreeBSD now even when hotplug support is enabled,
and handle any duplicates when the hotplug events finally do arrive.
Likewise, when libusb_exit is called, it also blocks for around 4 seconds
before returning - which delays our code from being able to do a clean
exit quickly. There's not much we can do about that one except add some
extra debug logging when verbosity is turned up, so that users can see it
is not actually our code that has them twiddling their thumbs waiting.
Hopefully later releases of FreeBSD will improve on this.
* Work around FreeBSD 11 deadlocking on device unplug. If a device is ever
removed while we are in the middle of a call to libusb_bulk_transfer(),
then that call may deadlock and never return, and the thread which called
it will not be able to be cancelled. We can limit the impact that has on
our code (since we will already always be cancelling that thread once we
get the hotplug notification of the removal) by using pthread_timedjoin_np
and if the join times out, bark about it and just ignore the zombie thread
that we've been left with. It's not ideal, and if it happens often enough
in a single process run, then leaked resources will start to be exhausted.
But most people don't replug things all that frequently, this won't happen
every time they do, and it is about the best we can do until the FreeBSD
bug is fixed. There's no downside to the workaround if the bug doesn't
actually occur when a device is unplugged.
* Disable some buggy optimisations (normally enabled by -O2) when using GCC
on FreeBSD 11. On that platform (we've not seen this anywhere else), they
appear to miscompile some code in a way that stack unwinding details are
lost and a thrown exception will invoke terminate rather than being caught
by the handler it should have unwound to. Some of these were in obscure
corners of the code that are only seen when unlikely errors occur, so it's
not impossible that there may be a few more lurking. But for now, we'll
just disable the known problem ones rather than falling all the way back
to compiling with -O0 on that platform.
* Add a --limit-max-xfer option to seedd and bbcheck. This gives people a
runtime workaround for systems where, despite having an ostensibly new
enough Linux kernel and libusb version to support large bulk URBs, the
hardware chipset has some quirk that isn't yet fixed in the kernel driver
which makes using them troublesome (I'm looking at you RPI3). If passed,
this will force the old 16kB limit on individual requests to usbfs. That
doesn't have any effect on the size of requests that users can make from
our code, it just might be slightly slower at obtaining huge numbers of
bits, since we need to get them in smaller chunks internally.
* Add an example of how to obtain random integers within an arbitrary range
where every value in that range remains equally probable. This isn't a
difficult thing to do, but enough people have asked to provide a working
example, and there's enough history elsewhere of people doing something
naive, like using a modulus to limit the range (thus making some values
more probable than others), to have a good example easily available that
users can refer to.
* Search for where udevadm is found in the qemu-hook script. Its installed
path was changed by systemd 204 in Jessie from /sbin to /bin, so now we
need to deal with multiple interfaces to stay portable, since there are
supported distro releases that aren't EOL yet which still have it in the
old location. Jessie's systemd had a compat link to make it available in
both places, but the systemd maintainers want to drop that for Buster (and
not all other distros with newer systemd provide that). Closes: #852582
* Tweak the udev rules to work around a bug in udev versions up to at least
232-25 which makes testing ATTR keys with != become a perilous folly. If
an event occurs for a device which does not have that attribute at all,
then the rule will be skipped, the same as if it was actually equal to the
value being tested for.
-- Ron Lee <email@example.com> Mon, 19 Jun 2017 13:31:17 +0930
bit-babbler (0.6) unstable; urgency=medium
* Update the libvirt Suggests to include libvirt-clients, since libvirt-bin
got split into a bunch of pieces in Jessie (because GNOME, see #679074),
and -clients is where virsh has moved to now.
* Add configure tests for the functions udev_device_get_tags_list_entry and
udev_device_get_sysattr_list_entry which were added in udev 154 and 167
respectively. The RHEL/CentOS 6 release appears to ship with udev 147,
so we can't use them there, and it won't be fully EOL for some time yet.
That's not a big deal, we only use them to output extra device information
when the debug level is cranked up, so we can just omit that code on any
systems where they aren't supported.
* Include xlocale.h explicitly on platforms where we need that for strtod_l.
* Initialise struct addrinfo more portably. There's little point to trying
to be clever with static initialisers while g++ has limited support for
them in C++ code, just do it in a way that will work everywhere.
* Add an explicit guard rather than a platform check around the code to
append a wstring as UTF-8 and disable it by default for all platforms now.
We don't actually need or use that anywhere here right now, and there are
more platforms than MSW where wchar_t is unspecified and/or insane. So
we'll worry about enabling it again if we ever do need it here, since we
do want this to be portable to those still.
* Automatically add -D_REENTRANT for compilers that don't do that themselves
when -pthread is used.
* Preserve the Chi^2 statistic when long term results need to be normalised
to prevent wrap around, and improve the precision on scaling the other
metrics to minimise discontinuities, especially on machines with a 32-bit
size_t where the normalisation is more likely to be needed in practice.
Many thanks to George Tsegas for his extensive testing, and careful and
critical attention to questioning all of the results of that.
* Fix the framing/status check to handle devices plugged into USB 1.0 ports.
Apparently there are still a few of those left in the wild.
-- Ron Lee <firstname.lastname@example.org> Fri, 17 Jun 2016 23:36:55 +0930
bit-babbler (0.5) unstable; urgency=medium
* Add more options to optimise for minimal power consumption. The defaults
before now were mostly focussed on keeping a good supply of fresh entropy
being regularly mixed into the kernel pool, and on minimising the risk of
starvation delays when demand is high. But there's an equally important
group of users who not only want good entropy, but also want to minimise
idle power usage as much as possible. So we now have some extra tunables
to better support that too.
The rate at which new entropy is mixed into the kernel pool even when it
has not fallen below its low water mark is now directly configurable, as
is the rate at which we throttle down requesting more entropy from the
hardware when real demand for it falls. Tuning these can minimise how
often we are responsible for waking the CPU on an otherwise idle system.
It is also now possible to configure the devices to be released when we
expect to be idle beyond a given period of time, which will allow them to
be powered down and suspended, and only woken again when we do need more
entropy from them. There are new udev rules which automatically enable
the USB autosuspend feature of the Linux kernel for them when they are
plugged in, which means this will work without needing to manually set
all that up (unless you want to further tweak the parameters there too).
* Don't create the control socket by default when only a limited number of
output --bytes are requested. It can still be enabled explicitly if you
do want it available while they are being read, but that's normally of
fairly limited use, and it's otherwise just annoying to have to remember
to explicitly disable it when extracting a block of entropy in this way,
and confusing to users if it complains they don't have permission to
(re)create it in the default location.
* Defer device initialisation until the pool threads have been started.
Most users won't really notice any difference from that, but when you
have 100 devices in a machine together then even small delays quickly
add up to become a thumb twiddling pause if they are serialised rather
than being run in parallel.
* Better support for pass-through to libvirt managed virtual machines when
there is more than one BitBabbler device in the host.
This is still more painful than it really ought to be, but we now have a
big enough hammer pounding on enough of the rough edges in libvirt support
for things to work like USB devices should be expected to work. They can
be hotplugged dynamically without admin intervention to the guest machines
you want them assigned to, and assigned to guest machines without fragile
hacks based on which USB port they are plugged into.
-- Ron Lee <email@example.com> Wed, 23 Dec 2015 00:38:47 +1030
bit-babbler (0.4) unstable; urgency=medium
* Switch to using libusb-1.0 now. It turns out that libusb-0.1 doesn't
actually work on kFreeBSD, it only builds there ... which isn't very
helpful. The kFreeBSD port actually uses FreeBSD's own libusb which also
provides a compatibility API for libusb-1.0 - and we need to jump through
a few small extra hoops to use it, but it has the advantage of actually
working, which is a plus. This also means we immediately get much better
support and lots of bugfixes for non-Debian platforms too, so this should
work everywhere that current releases of libusb do now.
* Drop libftdi. This is partly a consequence of the above, since a version
of it built with libusb-1.0 isn't widely available, and partly a result of
realising we weren't really using anything from it that we couldn't just
do more easily and more directly though libusb ourselves anyway. Our code
ended up being significantly refactored and simplified as a result of this
and it opened the way for a number of additional easy improvements too.
* Drop the --device-num option for selecting devices. Having an arbitrary
enumeration isn't really all that useful in hotplug environments, and the
--device-id option now transparently supports selecting devices by their
serial number, or by either their logical or physical address on the bus,
so the duplication there was only becoming a source of confusion.
* More speed and efficiency tuning. As a result of now having more direct
control over the device we've been able to notably reduce some of the
overheads of streaming data out of it, which means we're now using less
CPU cycles with an increase in throughput for the same device clock rate.
* Make the libudev build dependency conditional on linux-any so the kFreeBSD
buildds will actually want to build it. We can't do much for Hurd until
someone actually ports libusb-1.0 to it.
* Make the use of signals which may not exist on all platforms conditional,
which should enable this to build on MIPS, Sparc, and Alpha too.
-- Ron Lee <firstname.lastname@example.org> Sat, 05 Dec 2015 04:40:11 +1030
bit-babbler (0.3) unstable; urgency=medium
* Include a simple example script for reading from the UDP socket.
* Include the documentation for configuring virtual machine pass-through
in the binary package as well.
* Document how to deal with cgroups mandatory access control when using
devices inside libvirt managed virtual machines.
* Initial upload to Debian, Closes: #805979
-- Ron Lee <email@example.com> Tue, 24 Nov 2015 23:15:24 +1030
bit-babbler (0.2) unstable; urgency=low
* Add the option to read entropy directly from a UDP socket too.
* Permit TCP to be used for the control socket. Not all platforms have
unix domain sockets, and some people might actually want to be able to
access it remotely anyway.
* Guard the Linux specific code (for feeding the kernel entropy) to only
build on Linux, and get it to build with the mingw-w64 cross compiler.
* Add a system group bit-babbler, and a udev rule which grants permission
to access the device(s) to people in that group. This is mostly useful
for people who want to stream bits directly out of the devices and don't
need or want the privilege required to be able to write entropy directly
to the kernel pool.
-- Ron Lee <firstname.lastname@example.org> Sat, 27 Jun 2015 01:17:17 +0930
bit-babbler (0.1) unstable; urgency=low
* Initial release
-- Ron Lee <email@example.com> Tue, 24 Feb 2015 09:23:18 +1030