Before a2cd6aeae8, "repo mirror with --current-branch" fetches git data
using command
git fetch --progress --update-head-ok cros --tags
No refspec is specified, thus it fetches default refspec, which is +refs/heads/*:refs/heads/*
After a2cd6aeae8, the fetch command became
git fetch --progress --update-head-ok cros --tags +refs/tags/*:refs/tags/*
It did not only add tags refspec, but also suppressed the fetching of default refspec.
In other words, repo mirrors doesn't work if current_branch_only=True.
This CL explicitly adds the default refspec to command line if none is
specified.
Bug: https://crbug.com/gerrit/11990
Change-Id: Iadcf7b9aa50f53c47132cfe6c53b3fb2076ebca2
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/246632
Reviewed-by: David Pursehouse <dpursehouse@collab.net>
Reviewed-by: Mike Frysinger <vapier@google.com>
Tested-by: Kuang-che Wu <kcwu@chromium.org>
Series of steps:
* Create a local "b1" branch with `repo start b1` that tracks a remote
branch (totally fine)
* Manually create a local "b2" branch with `git branch --track b1 b2`
that tracks the local "b1" (uh-oh...)
* Delete the local "b1" branch manually or via `repo prune` (....)
* Try to process the "b2" branch with `repo prune`
Since b2 tracks a branch that no longer exists, everything blows up
at this point as we try to probe the non-existent ref. Instead, we
should flag this as unknown and leave it up to the user to resolve.
This probably could come up if a local branch was tracking a remote
branch that was deleted from the server, and users ran something like
`repo sync --prune` which cleaned up the remote refs.
Bug: https://crbug.com/gerrit/11485
Change-Id: I6b6b6041943944b8efa6e2ad0b8b10f13a75a5c2
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/236793
Reviewed-by: David Pursehouse <dpursehouse@collab.net>
Reviewed-by: Kirtika Ruchandani <kirtika@google.com>
Reviewed-by: Mike Frysinger <vapier@google.com>
Tested-by: Mike Frysinger <vapier@google.com>
If the local branch changed state since its last upload, the data
cached in .git/config related to the last uploaded CL might not be
that relevant. If we're able to fast forward merge to the latest
tree state, then let's do that. This would be akin to checking
out a detached head before syncing where we already switch state.
If we aren't able to fast forward merge, then it's not a big deal
as we'll continue on to the existing branch checking logic.
This would be easy to reproduce by doing something like:
$ repo start foo .
$ git revert HEAD
$ repo upload --cbr .
$ git reset --hard HEAD^
<CL is merged>
$ repo sync .
<we can fast forward>
Change-Id: I7d62f3d1ba5314a349d85b4dbb0ec8352eca18bb
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/238552
Tested-by: Mike Frysinger <vapier@google.com>
Reviewed-by: Michael Mortensen <mmortensen@google.com>
Reviewed-by: Mike Frysinger <vapier@google.com>
If you run `repo sync -l foo` without first `repo sync -n foo`,
repo sets up an invalid gitdir tree that gets wedged and requires
manual recovery. Add a sanity check to abort cleanly first.
Change-Id: Iad865ea860a3f1fd2f39ce683fe66bd4380745a5
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/244732
Reviewed-by: David Pursehouse <dpursehouse@collab.net>
Tested-by: Mike Frysinger <vapier@google.com>
Use open() as a context manager to simplify the close logic and make
the code easier to read & understand. This is also more Pythonic.
Change-Id: I579d03cca86f99b2c6c6a1f557f6e5704e2515a7
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/244734
Reviewed-by: David Pursehouse <dpursehouse@collab.net>
Tested-by: Mike Frysinger <vapier@google.com>
Hitting Ctrl-C in the middle of this func will leave the .git in a
bad state that requires manual recovery. The code tries to catch
all exceptions and recover by deleting the incomplete .git dir, but
it omits KeyboardInterrupt which Exception misses.
We could add that to the recovery path, but we can make this more
robust with a different approach: set up everything in .git.tmp/
and only move it to .git/ once we've fully initialized it.
Change-Id: I0f5b97f2e19fc39cffc3e5e23993a2da7220f4e3
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/244733
Reviewed-by: David Pursehouse <dpursehouse@collab.net>
Tested-by: Mike Frysinger <vapier@google.com>
If the repo client checkout is in an incomplete sync state, the work
git repo might be in a bad way. Turn errors parsing HEAD into None
since callers of CurrentBranch already need to account for it.
Change-Id: Ia7682e29ef4182006b1fb5f5e57800f8ab67a9f4
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/239239
Reviewed-by: David Pursehouse <dpursehouse@collab.net>
Tested-by: Mike Frysinger <vapier@google.com>
If `git diff` fails in any project checkout (e.g. an incomplete
sync), make sure we print that error clearly rather than blowing
up, and exit non-zero in the process.
Bug: https://crbug.com/gerrit/11613
Change-Id: I12f278427cced20f23f8047e7e3dba8f442ee25e
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/239236
Reviewed-by: David Pursehouse <dpursehouse@collab.net>
Tested-by: Mike Frysinger <vapier@google.com>
When displaying progress bars, we use \r to reset the cursor to the
start of the line before showing the new update. This assumes the
new line will fully erase whatever was displayed there previously.
The "done" codepath tries to handle this by including a few extra
spaces at the end of the message to "white out" what was there.
Lets replace that hack with the standard ECMA escape sequence that
clears the current line completely. This is the CSI "erase in line"
sequence that the terminal will use to delete all content. The \r
is still needed to move the cursor to the start of the line. Using
this sequence should be OK since we're already assuming the terminal
is ECMA compliant with our use of coloring sequences. We also put
the \r after the CSI sequence on the off chance the terminal can't
process it and displays a few bytes of garbage.
The other improvement is to the syncbuffer API. When it dumps its
status information, it almost always comes after a progress bar
update which leads to confusing comingled output. Something like:
Fetching projects: 100% (2/2) error: src/platform2/: branch ...
Since the progress bar is "throw away", have the syncbuffer reset
the current output to the start of the line before showing whatever
messages it has queued.
Bug: https://crbug.com/gerrit/11293
Change-Id: I6544d073fe993d98ee7e91fca5e501ba5fecfe4c
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/236615
Reviewed-by: David Pursehouse <dpursehouse@collab.net>
Tested-by: Mike Frysinger <vapier@google.com>
This makes it way easier to recover from forgetting to run repo start
before committing: just run `repo start -b new-branch`, instead of
all that tedious mucking around with reflogs.
Change-Id: I56d49dce5d027e28fbba0507ac10cd763ccfc36d
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/232712
Reviewed-by: Mike Frysinger <vapier@google.com>
Reviewed-by: David Pursehouse <dpursehouse@collab.net>
Tested-by: Mike Frysinger <vapier@google.com>
There is a standard Python "trace" module, so having a local trace.py
prevents us being able to import that. Rename the module to avoid.
Change-Id: I23e29ec95a2204bb168a641323d05e76968d9b57
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/234832
Reviewed-by: David Pursehouse <dpursehouse@collab.net>
Tested-by: Mike Frysinger <vapier@google.com>
The GitCommand Wait helper takes care of decoding bytes to strings
for us. That means we don't have to decode stdout ourselves which
is what our local rev list, ls-remote, and generic get_attr helpers
were doing.
If we don't use Wait though to capture the output but instead go
directly to the subprocess stdout, we do have to handle decoding
ourselves. This is what the diff helpers were doing.
Bug: https://crbug.com/gerrit/10418
Change-Id: I057ca245af3ff18d6b4a074e3900887f06a5617d
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/233076
Reviewed-by: David Pursehouse <dpursehouse@collab.net>
Tested-by: Mike Frysinger <vapier@google.com>
As we convert repo to support Python 3, the version of Python that we
use might not be the version that repo hooks users have written for.
Since repo upgrades are not immediate, and not easily under direct
control of end users (relative to the projects maintaining the hook
code), allow hook authors to declare the version of Python that they
want to use.
Now repo will read the shebang from the hook script and compare it
against the version of Python repo itself is running under. If they
differ, we'll try to execute a separate instance of Python and have
it load & execute the hook. If things are compatible, then we still
use the inprocess execution logic that we have today.
This allows repo hook users to upgrade on their own schedule (they
could even upgrade to Python 3 ahead of us) without having to worry
about their supported version being exactly in sync with repo's.
Bug: https://crbug.com/gerrit/10418
Change-Id: I97c7c96b64fb2ee465c39b90e9bdcc76394a146a
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/228432
Reviewed-by: David Pursehouse <dpursehouse@collab.net>
Tested-by: Mike Frysinger <vapier@google.com>
A new option, --partial-clone is added to 'repo init' which tells repo
to utilize git's partial clone functionality, which reduces disk and
bandwidth usage when downloading by omitting blob downloads initially.
Different from restricting clone-depth, the user will have full access
to change history, etc., as the objects are downloaded on demand.
Change-Id: I60326744875eac16521a007bd7d5481112a98749
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/229532
Reviewed-by: Mike Frysinger <vapier@google.com>
Tested-by: Xin Li <delphij@google.com>
This fix exception with python3 with stack-trace:
error: Cannot fetch platform_external_grpc-grpc-java.git (UnicodeDecodeError: 'utf-8' codec can't decode byte 0x96 in position 640: invalid start byte)
[...]
File "[...]project.py", line 2255, in _IsValidBundle
if f.read(16) == '# v2 git bundle\n':
File "/usr/lib/python3.5/codecs.py", line 321, in decode
(result, consumed) = self._buffer_decode(data, self.errors, final)
Even if we ask 16 characters, python buffered decoder will try to decode more in the buffer
The patch works for python2 and python3, and open the file in byte mode so that decoding is not attemped
Signed-off-by: Pierre Tardy <tardyp@gmail.com>
Change-Id: I837ae3c5cd724b34670fc2a84e853431f482b20d
Reviewed-on: https://gerrit-review.googlesource.com/c/git-repo/+/224642
Reviewed-by: Mike Frysinger <vapier@google.com>
Reviewed-by: David Pursehouse <dpursehouse@collab.net>
Tested-by: Mike Frysinger <vapier@google.com>
Currently we read the binary stream from the subprocess code directly
before waiting for it to finish, but there's no need to do so as we
aren't streaming the output to the user. This also means we pass up
binary data to the caller as we don't go through GitCommand's internal
logic which decodes the stream as utf-8.
Simplify the code by calling Wait first, then splitting the entire
captured output in one line.
Bug: https://crbug.com/gerrit/10418
Change-Id: I7a57904be8cb546a229980fb79c829fc3df31e7d
There's no reason to support any other encoding in these files.
This only affects the files themselves and not streams they open.
Bug: https://crbug.com/gerrit/10418
Change-Id: I053cb40cd3666ce5c8a0689b9dd938f24ca765bf
During sync, repo runs `git read-tree --reset -u -v HEAD` which causes
git-lfs's smudge filter to run, which fails because git-lfs does not
work with bare repositories.
This was fixed in I091ff37998131e2e6bbc59aa37ee352fe12d7fcd to
automatically disable this smudge filter. However, later versions of
Git (2.11.0) introduced a new filter protocol [1], to avoid spawning
a new command for each filtered file. This was implemented in Git-LFS
1.5.0 [2].
This patch fixes the issue by setting the git lfs process filter, in
addition to the smudge filter. For any projects that have LFS objects,
`git lfs pull` must still be executed manually afterwards.
[1] edcc85814c
[2] https://github.com/git-lfs/git-lfs/pull/1617
Bug: https://crbug.com/gerrit/10911
Change-Id: I277fc68fdefc91514a2412b3887e3be9106cab48
When syncing a lot of projects in parallel, it is not otherwise
clear which one of them has failed to init work tree.
Change-Id: I8edfb4955023389a499e99cfa511bdc0d2850ba2
This allows projects to include submodules inside of
projects that use repo without repo incorrectly believing
the area is dirty just because a submodule has updates.
This is in line with git porcelain commands which generally
require a commandline flag to include submodules (git add,
git rebase).
Change-Id: Ide8a292162a42ab35145b5c4ca8ca0d020cdfe81
The newly introduced "Already have persistent ref" message prevents
repo from overwriting the last line when syncing quietly. Omit the
message when syncing quietly to clean up the output and to restore
the previous behaviour.
Change-Id: Idf42751c67f95924d6de50092ba54d4c6fe93096
If a tag is rewritten on the server (which is bad), trying to recover
locally with `repo sync --force-sync` doesn't actually work. People
have to manually delete things themselves to fix syncing. While tags
should never be rewritten in practice, allow users to easily recover
from broken servers.
We updated some of these code paths already (see commit 6e53844f1e
"Allow clobbering of existing tags from remote."), but the incremental
update flow was missed.
Bug: b/120778183
Bug: chromium:932651
Test: delete local tag & recreate to diff commit, then check
`repo sync` & `repo sync --force-sync` behavior
Change-Id: I3648f7d2526732c06016b691a9a36c003157618d
Even if dots are used as separators for Git config keys, they are not
forbidden as part of submodule names. This fixes the issue of submodules
with a name like e.g. "long.js" to be skipped from checkout.
Change-Id: I77da07925ad207fa3d043067dfbbcb4a1ebdac4d
When --ne/--no-emails is added to 'repo upload' command line, gerrit
server will not generate notification emails.
project.py:Project.UploadForReview method is modified to accept a
string recognizable by gerrit to indicate different sets of destination
email addressees, but the upload command line allows only one option -
disable sending emails completely.
Default repo upload behavior is not being changed.
TEST=tried in the Chrome OS repo, observed that patches uploaded with
--ne or --no-emails indeed do not trigger any emails, while
patches uploaded without these command line options still trigger
email notifications.
Change-Id: I0301edec984907aedac277d883bd0e6d3099aedc
Pass options through the refspec for all transports, including ssh.
This means the behavior will be more consistent between the ssh and
https cases.
A downside is that this prevents passing special characters in
reviewer options. That already didn't work over https, so it seems
okay. It could be fixed by using push options instead.
Change-Id: Ia38d16e350cb8cb0de14463bfb3d9724e13bc4bf
* Add more file i/o wrappers in platform_utils to allow using
long paths (length > MAX_PATH) on Windows.
* Paths using the long path syntax ("\\?\" prefix) should never
escape the platform_utils API surface area, so that this
specific syntax is not visible to the rest of the repo code base.
* Forward many calls from os.xxx to platform_utils.xxx in various place
to ensure long paths support, specifically when repo decides to delete
obsolete directories.
* There are more places that need to be converted to support long paths,
this commit is an initial effort to unblock a few common use cases.
* Also, fix remove function to handle directory symlinks
Change-Id: If82ccc408e516e96ff7260be25f8fd2fe3f9571a
"repo init --reference" has two purposes: to decrease bandwidth used
at clone time, and to decrease disk usage afterward, by using the
reference repositories as an alternate store of objects even after
the clone. The downside is that it makes the borrowing repositories
dependent on the reference repositories, so it is easy to end up
with missing objects by mistake after a cleanup operation like "git
gc".
To prevent that, v2.3.0-rc0~30^2 (clone: --dissociate option to mark
that reference is only temporary, 2014-10-14), "git clone" gained a
--dissociate option that makes --reference reuse objects from the
reference repository at clone time but copy them over instead of
using the reference as an alternate. This is more straightforward to
use than plain --reference, at the cost of higher disk usage.
Introduce a --dissociate to "repo init" that brings the same benefits
to repo. The option is simply passed on to "git clone".
Change-Id: Ib50a549eb71e0a2b3e234aea57537923962a80d4
pylint is not used since bb5b1a0. The pyflakes cleanup mentioned in that
commit has not been done, but given that this project is no longer being
actively developed I don't think it's worth spending time doing it.
Leaving the pylint suppressions causes confusion because it leads people
to think that we are still using pylint.
Change-Id: If7d9f280a0f408c780f15915ffdb80579ae21f69
When someone does "repo download -c <project> <change>"
without specifying a patch number, by default patch 1 is
downloaded. An alternative is to look for the latest patch
and download the same when no explicit patch is given.
This commit does the same by identifying the latest patch
using "git ls-remote".
Change-Id: Ia5fa7364415f53a3d9436df4643e38f3c90ded58
Allow the 'default' and 'project' element in the manifest
file to apply "--no-tags" option equivalent.
Change-Id: I7e0f8c17a0e25cca744d45df049076d203c52ff5
Signed-off-by: YOUNG HO CHA <ganadist@gmail.com>
Put the correctly-expanded relative paths in objects/info/alternates.
From gitrepository-layout(5), this path should be "relative to the
object database, not to the repository".
Change-Id: I7b2027ae23cf7d367b80f5a187603c4cbacdb2de
Issue: when subproject url is a relative in .gitmodules
repo tool cannot handle this and cause:
"fatal: '***' does not appear to be a git repository
fatal: Could not read from remote repository."
issue.
Signed-off-by: Shouheng Zhang <shouheng.zhang@intel.com>
Change-Id: I2a24c291ea0074ba13a740b32a11c0c25975e72b
os.remove raises an exception when deleting read-only files on
Windows. Replace all calls with calls to platform_utils.remove,
which deals with deals with that issue.
Change-Id: I4dc9e0c9a36b4238880520c69f5075eca40f3e66
Without this change, '.git\HEAD' files, for examples, are sometime
read incorrectly resulting in the current branch to be reset to
"master" when running a "repo init -b xxx" on an already initialized
repository.
Change-Id: I48c7ef85ff81626edf156914329a560e14252f2a
* changes:
Port os.rename calls to work on Windows
Workaround shutil.rmtree limitation on Windows
Add support for creating symbolic links on Windows
Make "git command" and "forall" work on Windows
The output indicates that fetching happens even when it is skipped.
To avoid confusion, print a message when fetching is skipped for
an immutable ref so that the user knows when and why a fetch is skipped.
Change-Id: Id6e4812cebc5e57d379feb76a9d034af0b93043b
This option allow to bypass verification ssl certification while
establishing connection with Gerrit to upload review.
Change-Id: If2e15f5a273c18a700eb5093ca8a4d5a4cbf80cd
