| Age | Commit message (Collapse) | Author |
|---|
|
Use git verify-commit to verify gpg signatures, rather than the old method
of parsing git log output.
These two methods should always have the same result. Note that
git verify-commit allows signatures with unknown validity, the same as
git log's "U" output which was accepted. So any key in the gpg keyring
is allowed to sign the commit. Propellor provides gpg with a keyring
containing only the allowed keys.
Needs git 2.0, which is in even debian oldstable.
This commit was sponsored by Ewen McNeill on Patreon.
|
|
now that compatability with ghc 7 is no longer needed.
Data.Type.Bool contains effectively the same stuff that was implemented
here, so removed my code.
Tried to use Data.Type.Equality instead of my EqT, but it seems to be some
other type of (type level) equality, and didn't compile. Instead went with
the simpler EqT implementation that newer ghc versions allow.
The rest of the changes are simply better syntax for defining type
families.
And upon using that syntax, ghc noticed that `type family a + b`
does not have kind "ab" like I wrote before, but is kind *.
Tested on debian stable with ghc 8.0.1.
This commit was sponsored by John Pellman on Patreon.
|
|
|
|
|
|
Apt.installedBackport would do this:
apt-get install -t stretch-backports foo bar
Apt.backportInstalled does this:
apt-get install foo/stretch-backports bar/stretch-backports
The Apt.installedBackport behaviour can install the dependencies of foo and bar
from stretch-backports even when the versions in stretch will satisfy the
dependencies of the backports of foo and bar. So this property can result in
very many more backports being installed on the host when intended. But the
number of installed backports should always be minimised.
Worse, whether this happens is highly dependent on the system state, and the
order in which other properties get ensured. For example,
& Apt.installed ["dgit"]
& Apt.installedBackport ["dgit"]
will install only dgit from stretch-backports, but unless debhelper and
devscripts happen to already be installed,
& Apt.installedBackport ["dgit"]
& Apt.installed ["dgit"]
will install dgit, debhelper, devscripts and maybe more from backports. This is
surprising, difficult to debug, and breaks the expectation that when the order
in which properties are ensured is not specified with connectives like
`requires` and `before`, ensuring them in any order will produce the same
result.
Property renamed because user configs should not silently break, as they would
if they did not list dependencies that must be installed from stable-backports.
Signed-off-by: Sean Whitton <spwhitton@spwhitton.name>
|
|
No such backport exists in the archive.
Signed-off-by: Sean Whitton <spwhitton@spwhitton.name>
|
|
|
|
|
|
version changes.
Surprised this didn't come up before, since propellor was run against
progressively old versions of libraries when propelling eg, a chroot.
It finally broke on an arm box, where libm got updated to a new version
and propellor used a symbol from the new version.
The comment says that propellor may be running from an existing shim
in which case it's reused. That could be a nested chroot or other
container, or perhaps propellor was deployed via a precompiled tarball
which is built using a shim. The code used to use "checkAlreadyShimmed shim"
which I don't see how it deals with either of those scenarios.
Changed to "checkAlreadyShimmed propellorbin", which I think will deal with
them, but I've not tested such scenarios.
Added code to delete old versions of libraries out of the shim directory
to avoid masses of old ones piling up over time. Property.Chroot sets up
the shim directory and then bind mounts it into the chroot. To avoid
deleting the source of a bind mount, made this only delete files in the
shim directory, but not the shim directory itself.
|
|
|
|
This reverts commit 02eca2ae4cf51d8e83d94d8359e15ac053451109.
This seems to have broken propellor badly, in testing I'm seeing it
crash at the end of a run with "thread blocked indefinitely in an STM
transaction" and also during the run it printed out some odd output
like:
apache2:
apache2:
dummy IN SSHFP 4 1 35df80973f5877e4041f1b70947385eb2f6a0822
dummy IN SSHFP 4 2 3a0bb426e76eebc5c56e3b0f1428aa9d18539e9621bf8f9e3b7f56a4e7d81c85
Which seems like it might be output of commands that
propellor is supposed to be reading?
Seems likely that there's a bug or two that have crept
into then concurrent-output library since the version embedded in
propellor.
|
|
Also tested with lts-11.6 / ghc 8.2.2 and it does build, however the stack
in debian stable (and even unstable currently) does not support that
version's data.
|
|
Removed deps on transformers, text, stm. Updated debian/control and
Propellor.Bootstrap accordingly. Sorted the lists of deps to make it easier
to keep them in sync.
This commit was sponsored by Nick Daly on Patreon.
|
|
Fix build with ghc 8.4, which broke due to the Semigroup Monoid change.
See https://prime.haskell.org/wiki/Libraries/Proposals/SemigroupMonoid
Dropped support for building propellor with ghc 7 (as in debian
oldstable), to avoid needing to depend on the semigroups transitional
package, but also because it's just too old to be worth supporting.
If we indeed drop ghc 7 support entirely, some code to support "jessie"
can be removed; concurrent-output can be de-embedded, and the Singletons
code can be simplified.
This commit was sponsored by Jack Hill on Patreon.
|
|
|
|
This is a trick I only just learned about, see
https://stackoverflow.com/questions/6711151/how-to-avoid-recompiling-in-this-cabal-file#6711739
Significantly increased propellor build speed when your config.hs is in
a fork of the propellor repository, by avoiding redundant builds of
propellor library.
Also avoids needing to list all the build deps 3 times.
Also avoids cabal 2.x wanting every module to be listed 3 times.
Note that the bulk of wrapper.hs had to move into the propellor library,
since that code depended on stuff not exposed by the library.
This commit was sponsored by Henrik Riomar on Patreon.
|
|
|
|
|
|
--to-dest and --to-source work. Thanks, Russell Sim
|
|
which seems to not work anymore.
Thanks, Russell Sim.
|
|
|
|
|
|
Sipma.
|
|
|
|
Debian git bundle using an older version of propellor that set up an upstream remote.
This commit was sponsored by Jake Vosloo on Patreon.
|
|
|
|
|
|
Added Propellor.Property.Atomic, which can make a non-atomic property
that operates on a directory into an atomic property.
Also has a generic version that could be used for things other than
directories that can be updated atomically.
(Inspired by Vaibhav Sagar's talk on Functional Devops in a
Dysfunctional World at LCA 2018.)
This commit was sponsored by Fernando Jimenez on Patreon.
|
|
This commit was sponsored by Nick Daly on Patreon.
|
|
|
|
* Last release mistakenly contained my personal branch not master.
* contrib/post-merge-hook documentation updated to recommend also using
it as a post-checkout hook, to avoid such problems.
|
|
|
|
|
|
This commit was sponsored by Brock Spratlen on Patreon.
|
|
This commit was sponsored by Ewen McNeill on Patreon.
|
|
includes the caller.
This commit was sponsored by Jochen Bartl on Patreon.
|
|
To avoid inheriting some problematic environment variables, such as TMP,
from the caller.
The only potential breakage from this change would be if something used
setEnv before one of the affected properties. Audited propellor's source
for that, and nothing does. Anything that did would could fail in a
concurrent context anyway.
|
|
* Parted: Allow partitions to have no filesystem, for eg, GPT BIOS boot
partitions. (API change)
* Added rawPartition to PartSpec, for specifying partitions with no
filesystem.
* Added BiosGrubFlag to PartFlag.
Note that man parted does not list the "bios_boot" flag, but I found it in
its html documentation. Other flags may also be missing.
This commit was sponsored by Boyd Stephen Smith Jr. on Patreon.
|
|
Avoid bogus warning about new upstream version when /usr/bin/propellor is
run on a Debian system, but ~/.propellor was not cloned from the Debian git
bundle.
|
|
|
|
Added Propellor.Property.Installer modules, which can be used to create
bootable installer disk images, which then run propellor to install a
system.
This code was extracted from the demo I gave at my talk at DebConf 2017,
from the secret-project repository. That repository was AGPL licensed.
I hereby relicense the code committed here under the same 2-clause BSD
license as the rest of propellor.
Changes from secret-project: Generalized UserInput to a type class,
and added a lot more documentation and examples.
This commit was sponsored by Brock Spratlen on Patreon.
|
|
Signed-off-by: Sean Whitton <spwhitton@spwhitton.name>
|
|
This is to support eg, coreboot. The GrubTarget passed to Grub.installed
is introspected to determine --target. If multiple grubs are installed,
it currently doesn't pass any --target. Might make more sense to run
grub-install repeatedly, but I don't know if that case is sane at all.
The Xen -> "x86_64-xen" mapping is kind of arbitrarily
chosen since there's a i386-xen available too. I don't know when that
case would be used in any case though; chainPVGrub uses installed Xen,
but it does not run grub-install. If this does become a problem,
would probably need to split it into Xen64 and Xen32.
Renamed BIOS to GrubTarget in passing to match grub's terminology; BIOS was
kind of a joke term for this in propellor.
This commit was sponsored by Francois Marier on Patreon.
|
|
* DiskImage: Fix rsync crash when a mount point does not exist in the
chroot.
* Fix bug in unmountBelow that caused unmounting of nested mounts to
fail.
This commit was sponsored by Jack Hill on Patreon.
|
|
Cheap flash drives need partitions aligned to 4 MiB in order to not be slow
(and to avoid extra writes). <https://lwn.net/Articles/428584/>
And at least 1 MiB alignment is generally a good idea, and most people
seem to think 4 MiB is for all drives.
I noticed that Parted.partitioned does not do that; the first
partition started at an offset of 1 MB, and subsequent partitions
from where it ends. (The 1 MB offset came from the PartedVal PartSize
instance, and note that it was not 1 MiB.)
* Parted: Add an Alignment parameter. (API change)
A good default to use is safeAlignment, which is 4MiB,
well suited for inexpensive flash drives, and fine for other disks too.
Previously, a very non-optimial 1MB (not 1MiB) alignment had been used.
* DiskImage: Use safeAlignment. It didn't seem worth making the
alignment configurable here.
Alignment is implemented by offsetting the first partition's start
position so it's aligned (making sure to leave room for the partition
table). Each partition is then extended as needed so the next partition
will start properly aligned.
Note that parted rejects partition tables that don't fit in cylinder
bounderies. Before, propellor let parted deal with the fine details
of layout, so that was not a problem. Now it's possible to set some
wacky Alignment not divisible by 512, or use Byte sizes for partitions
and create a partition table that parted rejects. But, using
safeAlignment and MegaBytes should always be safe.
Also, this fixes a rounding bug in Parted.calcPartTable.
It was rounding up to the nearest MegaByte when allocating remaining
disk space, so returned partition table that was actually larger than
the disk size.
This commit was sponsored by an anonymous bitcoiner.
|
|
Use "fat32" for both; these are not distinguished at the partition table
level.
This commit was sponsored by Henrik Riomar on Patreon.
|
|
Avoid needing to --spin host deployed from disk image to let its propellor
access its privdata.
This commit was sponsored by Boyd Stephen Smith Jr. on Patreon.
|
|
|
|
|
|
|
