| Commit message (Collapse) | Author | Age | Files | Lines |
... | |
|
|
|
|
|
|
|
|
| |
Use dd and sfdisk instead of qemu-img, parted and install-mbr
Write files without using tee, that was only required when we were
not expecting morph to be run as root
partx has not been touched, I don't know enough about how to process
the output in python and I don't speak awk
|
| |
|
| |
|
|
|
|
| |
We can use self.app.settings instead.
|
|
|
|
|
| |
This is better done in morph, I think, since it is clear policy,
and should not be buried deep in the call stack.
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
|
|
|
|
|
|
|
|
| |
The "cache prefix" concept is not immediately clear, and it's
generally speaking best to centralize the construction of filenames
in the cache to the CacheDir class anyway.
Also, use SaveFile for writing the log file, instead of writing it
to a temporary file and then copying, which is unnecessary I/O.
|
|
|
|
|
|
|
| |
It needs to be different for each subclass, yet SystemBuilder was
sharing the implementation with ChunkBuilder, which worked, but only
by happenstance. Now each class has their own implementation and the
base class has a NotImplemented implementation.
|
|
|
|
|
|
| |
It's not really the blob builder's job to unpack chunks, or determine
whether something needs building or not. Moved those things to better
places.
|
| |
|
|
|
|
|
|
|
|
| |
The helper class, Factory, has unit tests, which is why it's currently
separate. It may later get integrated with BlobBuilder, or the other
way around.
Classes that don't have unit tests are marked out of coverage.
|
| |
|
| |
|
| |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Also, change builder.py to always obey the --max-jobs setting, unless
a morphology has a max-jobs field.
The defaults have been chosen so that they work for everyone equally
well. It may be useful to have a local mirror and then set the options
to point there, but it's not reasonable to try to guess such things,
so the defaults can be adapated to that.
Collect the defaults into one place so they're easier to overview.
The cliapp interface for adding settings is verbose enough that the
defaults were getting buried.
|
| |
|
|
|
|
|
| |
This avoids problems with files with the right names but partial
content, if morph is killed in the middle of writing the file.
|
|
|
|
|
|
|
|
|
|
|
|
| |
This makes it easier to check the output with vgrep for errors or
problems. Voluminous git output made that hard.
The key is that "git bundle create" has no --quiet optio (meh),
so we have to redirect stderr to /dev/null. However, that would
hide real errors, which we want to catch. Luckily, we can just
replace the call to subprocess.call with a call to
subprocess.check_call, and then we'll at least abort if there's
an error, instead of silently ignoring it.
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Either all .gits need to be left alone, including submodules, or
all of them get changed.
Doing all gets surprisingly complicated, as bottom-up traversal is
needed for directories to have their mtimes set, which means that
either we check the path for a .git component, or we do our own
traversal function, to top-down exclude .git and bottom-up set
mtimes.
This is more effort than it is worth when git probably doesn't care
so let's just drop it entirely.
|
| |
|
|
|
|
|
|
|
|
| |
Setting mtimes can fail if it is a symbolic link and it
points to a non-existant file.
This has happened when we had symbolic links pointing to files
in a submodule, so we set mtimes after submodules have been
checked out.
|
| |
|
| |
|
|
|
|
|
|
|
|
|
| |
This is done to keep the git repository intact with history and all
that. cp + checkout turns out to be faster than a regular "git clone".
In order to avoid roundtrips to the internet whenever a chunk tries to
do something with submodules, we force the submodule URLs to point to
our locally cached repos.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Putting the commands in a file made it more difficult to tell
what happened in scripts, until the self.msg was added.
After then it made the commands more difficult to reproduce
So if it can safely be done in a single command it should be more
readable.
This approach uses sh -c to cd then evaluate the remaining options
The command is 'cd "$1" && shift && eval "$@"', then passes the command
as arguments to the shell script, so the program does not need to do any
escaping of strings.
If the commands were a list of strings instead of a quoted string then
it would be a bit safer, but less flexible for use.
|
| |
|
| |
|
| |
|
|\ |
|
| | |
|
| |
| |
| |
| | |
Now it should actually work.
|
|/
|
|
|
|
| |
This adds options --staging-filler and --staging-chroot.
The wisdom of these options needs to be re-considered at
some point, but for now they're OK.
|
|
|
|
|
| |
This time, with testing. For a five line method, this was surprisingly
tricky to get right. Should sleep properly at night...
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This allows us to do this:
"build-system": "autotools",
"configure-commands": [
"./configure --prefix=$PREFIX --disable-nls --enable-foo"
]
The explicit command is used for configure, but the commands provided
by the autotools build-system will be used for build-commands,
test-commands, and install-commands.
|
|
|
|
| |
This reverts commit 63cab3e0e2f550a0bfe470d018063a8b91a759bd.
|
|
|
|
| |
This reverts commit e0fb97f08156af97ec2341ec8061fe514ae358e5.
|
|
|
|
| |
This only includes PKG_CONFIG_PATH and LD_LIBRARY_PATH at the moment.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This change requires the latest version of cliapp to be installed.
Rather than creating an archive from the cached git repositories,
moving this tarball archive into the cache and then extracting it from
there into the build directory, we now directly extract the sources into
the build directory.
We also recursively extract all submodules into the corresponding paths
inside the build directory. (Recursing has not been tested yet,
however.)
|
|
|
|
|
| |
This change is necessary because upstream git no longer allows --remote
to be combined with an arbitrary SHA1.
|
|
|
|
|
| |
This may not be the optimal way of doing it but it is a start towards
making debugging failed builds a little easier.
|
|
|
|
|
|
|
|
|
|
|
|
| |
This requires build-single to take a dependency context tuple when
building chunks of a stratum. This context tuple is the surrounding
stratum which is used to construct the dependency graph in the worker
and then do a breadth-first search to collect all dependencies that need
to be added to the staging area.
Implementing this required a few hash/eq changes in Blob, Morphology
and Treeish as well as a few adjustments in the corresponding unit
tests.
|
|
|
|
|
| |
...while at the same time making sure that all dependencies are marked
for staging prior to building the blob itself.
|
| |
|