| Age | Commit message (Collapse) | Author |
|---|
|
... by avoiding reusing temp dirs for execute. While we are
at it, also refactor LocalFetchViaTmpRepo() to create its
own empty temp dirs, that cannot be reused by the caller.
|
|
|
|
For archives and Git repositories we should ensure that not finding
the witnessing entity (archive content blob or Git commit,
respectively) results in a distinct status in the response to a
request that sets up roots on the serve endpoint. This will allow
just-mr to better handle its interaction with the serve endpoint.
|
|
|
|
... and improve log messages in case of failure.
|
|
|
|
...to make it available also for setting up 'just serve' builds.
|
|
|
|
|
|
|
|
Ignore-special git-tree-based roots are still content defined, so
one should use the correct marker in the JSON description of the
root that will be stored in the repository description of target
cache keys. This commit fixes the issue and improves documentation.
|
|
The man page for open(2) says the following to the O_SYNC flag: 'O_SYNC
provides synchronized I/O file integrity completion, meaning write operations
will flush data and all associated metadata to the underlying hardware.' This
flag results in a high delay when files are stored in casx, e.g., several
seconds for medium-sized files such as 23 MB. Since just does not care about
persistency, this strong synchronization mechanism is not required and is
deactivated.
|
|
Absent roots are characterised only by a Git tree hash, so a new
variant of the underlying stored information was added in the form
of a plain string.
In order to avoid unwanted implicit conversions when instantiating
via literal strings, we force callers of the constructors to
explicitly differentiate between plain strings and filesystem
paths. Existing tests were updated to reflect this.
Co-authored-by: Alberto Sartori <alberto.sartori@huawei.com>
|
|
TargetCache...
...backed by the same CAS, but the FileStorage uses the given
shard. This is particularly useful for the just-serve server
implementation, since the sharding must be performed according to the
client's request and not following the server configuration.
|
|
Make sure that all CopyFile, WriteFile, and CreateSymlink
functions properly unlink the target file (if it exists and
overwrite requested) to avoid interferences of the install
command. With this change, the clean up step for install-cas
and the within GraphTraverser can new be omitted.
|
|
This is required in order to make them available to 'just serve'
in a minimal just installation.
|
|
This avoids using the more geenric GitRepoRemote method which
has libcurl as a dependency, something that is not needed for this
Git operation.
|
|
Also fixes a small typo in tree existence checker log messages.
|
|
Upwards symlinks should still be collected from actions, even if
only the non-upwards symlinks are supported artifact types. The
client side is thus the one responsible with enforcing the
non-upwardness condition.
|
|
|
|
...that std::filesystem::* calls produce. This is because existence
and type checks use almost exclusively std::filesystem::status,
which follows symbolic links, when being called with path arguments.
Instead, one should instead use these methods with the value
returned by a call of std::filesystem::symlink_status.
This commit also streamlines the FileSystemManager tests, as well
as replace bare calls to std::filesystem with their FileSystemManager
counterparts (where suitable).
|
|
The introduction of non-upwards symlinks as first-class objects
should have updated the handling of known git tree artifacts
containing symlinks. In particular, one should consider trees in
their entirety when uploading (irrespective of the ignore_special
flag), and git trees should only be reported as known only if
the ignore_special flag is set to false.
|
|
The command 'git add .' does not include paths found in .gitignore
files in the directory tree where the command is issued. This is
not the desired behaviour, as we expect for a tree with a given
commit id to contain all of the entries, irrespective of their
meaning to Git.
This commit addresses the issue as described.
For the just-mr.py script we modified the staging command to
'git add -f .'.
For the compiled just-mr, simply adding the force flag to
'git_index_add_all' did not work as intended for files found in
ignored subdirectories. This is a known libgit2 issue which has
been fixed in v1.6.3. Until we can upgrade our libgit2 version,
a workaround was implemented: we recursively read the directory
entries ourselves and add each of them iteratively using
'git_index_add_bypath', making sure to ignore the root '.git'
subtree (which cannot be staged).
At the moment the handling of Git submodules remains an open issue,
as Git does not allow '.git' subtrees to be forcefully added to the
index, and thus such directory entries will currently not be
considered as part of a git tree. This however is consistent
behavior between Git and libgit2.
|
|
...allowing the skipping of certain subtrees if needed. This is
useful, e.g., in simulating what a 'git add' call would do,
which ignores all '.git' subdirectories.
Also adds a corresponding test for the new method.
|
|
|
|
|
|
|
|
...and update tests accordingly.
|
|
...as early as possible. This ensures that callers always receive
only the tree entries for the supported object types.
For the symlinks non-upwardness check we pass a lambda capturing
the real backend of the tree entries, such that the symlinks can
be read.
Updates git_tree tests accordingly.
|
|
...but make sure it is still considered a special type.
The only non-special entry types remain file, executable, and tree.
|
|
|
|
A git CAS ist just a fall back, so it is OK if it is absent (e.g.,
the specificed directory does not exist). Therefore only log at
debug level, not at error level if we cannot open it.
|
|
... which became obsolete with the new fdless write/copy
implementations.
|
|
... to remove the risk of deadlocks on certain combinations
of C++ standard library and libc when performing the
copy/write in a child process. For 'fdless' copy/write, a
child process is used to prevent the parent from getting
polluted with open writable file descriptors (which might
get inherited by other children that keep them open and can
cause EBUSY errors).
|
|
This feature has been introduced with C++20.
|
|
... and in this way, continue to work correctly in the absence
of a current working directory.
|
|
|
|
|
|
...caused by incorrectly setting and resetting the library internal
state and the misuse of pthreads in libgit2.
Normally, git_libgit2_init and git_libgit2_shutdown should span the
life of a worker thread in order to be safely used. However, due to
an incorrect implementation of libgit2's threadstate with pthreads,
on unix systems there is a race condition.
Until the use of pthread_key_t is corrected in libgit2, we need to
apply a workaround by always ensuring that the main thread is the
first thread reaching the GitContext constructor.
|
|
... and drop unecessary IsTree() check.
|
|
... with two minor code base changes compared to previous
use of gsl-lite:
- dag.hpp: ActionNode::Ptr and ArtifactNode::Ptr are not
wrapped in gsl::not_null<> anymore, due to lack of support
for wrapping std::unique_ptr<>. More specifically, the
move constructor is missing, rendering it impossible to
use std::vector<>::emplace_back().
- utils/cpp/gsl.hpp: New header file added to implement the
macros ExpectsAudit() and EnsureAudit(), asserts running
only in debug builds, which were available in gsl-lite but
are missing in MS GSL.
|
|
... as this is only an internal functionality, and the caller will
take care of a proper error message if the absence of that entry
is not expected.
|
|
In the current libgit2 implementation, a fake repository wrapped
around an existing odb is being registered as owner the same way
as a normal repository object. Therefore, one has to guard both
the creation and destruction of the fake repository against all
other git operations that might access the internal cache during
this transfer of ownership.
|
|
Use a similar logic as for repository initialisation: first check
if tag has not already been created in another process, and only
then try creation; make more tries with more wait in between; only
retry if failure was due to internal locking.
|
|
|
|
As the initialisation of Git repositories is something that only
takes place once, we should check early and cheaply whether the
repository is already there before trying to initialize it.
If we do need to initilize a repo, we can afford more attempts and
longer wait times between tries to initalize if the failure to
initialize happens due to the internal Git locking mechanism.
|
|
|
|
Opening a repo should not check parent directories, only try to
open at given path.
|
|
|
|
...due to limited SSH support in libgit2. In order to allow the
fetches to still be parallel, we execute:
git fetch --no-auto-gc --no-write-fetch-head <repo> [<branch>]
This only fetches the packs without updating any refs, at the slight
cost of sometimes fetching some redundant information, which for our
purposes is practically a non-issue.
(If really needed, a 'git gc' call can be done eventually to try to
compact the fetched packs, although a save in disk space is not
actually guaranteed.)
|
