This makes "git push" issue a more detailed instruction when a user pushes
into the current branch of a non-bare repository without having an
explicit configuration set to receive.denycurrentbranch. In such a case,
it will also tell the user that the default will change to refusal in a
future version of git.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Replace all 'git log --graph' calls for history verification with the
combination of 'git log ...| git name-rev' first introduced by a6c7a27
(rebase -i: correctly remember --root flag across --continue,
2009-01-26). This should be less susceptible to format changes than
the --graph code.
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
currently for cases like
From: A U Thor <a.u.thor@example.com> (Comment)
mailinfo extracts the following 'Author:' field:
Author: A U Thor (Comment)
^^
which has two extra spaces left in there after removed email part.
I think this is wrong so here is a fix.
Signed-off-by: Kirill Smelkov <kirr@landau.phys.spbu.ru>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* ks/maint-mailinfo-folded:
mailinfo: tests for RFC2047 examples
mailinfo: add explicit test for mails like '<a.u.thor@example.com> (A U Thor)'
mailinfo: 'From:' header should be unfold as well
mailinfo: correctly handle multiline 'Subject:' header
* jk/signal-cleanup:
t0005: use SIGTERM for sigchain test
pager: do wait_for_pager on signal death
refactor signal handling for cleanup functions
chain kill signals for cleanup functions
diff: refactor tempfile cleanup handling
Windows: Fix signal numbers
The previous fix to mergetool to use checkout-index instead of cat-file
broke running mergetool anywhere except the root of the repository.
This fixes it by using the correct relative paths for temporary files
and index paths.
Signed-off-by: Charles Bailey <charles@hashpling.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
a6c7a27 (rebase -i: correctly remember --root flag across --continue,
2009-01-26) introduced a more portable GIT_EDITOR usage, but left the
old tests unchanged.
Since we never use the editor (all tests run the rebase script as
proposed by rebase -i), just disable it outright, which simplifies the
tests.
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When trying to find out mode changes, we should not access the symlink
targets using stat(); instead we use lstat() so that the diff does
not fail trying to find a non-existing symlink target.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git fsck" used to validate only loose objects that are local and nothing
else by default. This is not just too little when a repository is
borrowing objects from other object stores, but also caused the
connectivity check to mistakenly declare loose objects borrowed from them
to be missing.
The rationale behind the default mode that validates only loose objects is
because these objects are still young and more unlikely to have been
pushed to other repositories yet. That holds for loose objects borrowed
from alternate object stores as well.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
By default we looked at all refs but not HEAD. The only thing that made
fsck not lose sight of commits that are only reachable from a detached
HEAD was the reflog for the HEAD.
This fixes it, with a new test.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The signal tests consists of checking that each of our
handlers is executed, and that the test program was killed
by the final signal. We arbitrarily used SIGINT as the kill
signal.
However, some platforms (notably Solaris) will default
SIGINT to SIG_IGN if there is no controlling terminal. In
that case, we don't end up killing the program with the
final signal and the test fails.
This is a problem since the test script should not depend
on outside factors; let's use SIGTERM instead, which should
behave consistently.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When calling "git symbolic-ref" it is easy to forget that
the target must be a fully qualified ref. E.g., you might
accidentally do:
$ git symbolic-ref HEAD master
Unfortunately, this is very difficult to recover from,
because the bogus contents of HEAD make git believe we are
no longer in a git repository (as is_git_dir explicitly
checks for "^refs/heads/" in the HEAD target). So
immediately trying to fix the situation doesn't work:
$ git symbolic-ref HEAD refs/heads/master
fatal: Not a git repository
and one is left editing the .git/HEAD file manually.
Furthermore, one might be tempted to use symbolic-ref to set
up a detached HEAD:
$ git symbolic-ref HEAD `git rev-parse HEAD`
which sets up an even more bogus HEAD:
$ cat .git/HEAD
ref: 1a9ace4f2ad4176148e61b5a85cd63d5604aac6d
This patch introduces a small safety valve to prevent the
specific case of anything not starting with refs/heads/ to
go into HEAD. The scope of the safety valve is intentionally
very limited, to make sure that we are not preventing any
behavior that would otherwise be valid (like pointing a
different symref than HEAD outside of refs/heads/).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Use the newly introduced test_commit() and test_merge() helpers.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Use test_commit() and test_merge(). This way, it is harder to forget to
tag, or to call test_tick before committing.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Use test_commit() and test_merge(), reducing the code while making the
intent clearer.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Often we just need to add a commit with a given (short) name, that will
be tagged with the same name. Now, relatively complicated graphs can be
constructed easily and in a clear fashion:
test_commit A &&
test_commit B &&
git checkout A &&
test_commit C &&
test_merge D B
will construct this graph:
A - B
\ \
C - D
For simplicity, files whose name is the lower case version of the commit
message (to avoid a warning about ambiguous names) will be committed, with
the corresponding commit messages as contents.
If you need to provide a different file/different contents, you can use
the more explicit form
test_commit $MESSAGE $FILENAME $CONTENTS
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Make it easy for other authors to use rebase tests' fake-editor.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Rather than copying and pasting, which is prone to lead to fixes
missing in one version, move the fake-editor generator to t/t3404/.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
After you resolve a conflicted merge to remove the path, "git add -u"
failed to record the removal. Instead it errored out by saying that the
removed path is not found in the work tree, but that is what the user
already knows, and the wanted to record the removal as the resolution,
so the error does not make sense.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Also as suggested by Junio, in order to try to catch other MIME
problems, test cases from the "8. Examples" section of RFC2047 are added
to t5100 testsuite as well.
Signed-off-by: Kirill Smelkov <kirr@landau.phys.spbu.ru>
Actually, I think the issue is pretty independent of submodules; when
"git commit" gets an empty parameter, it misinterprets it as a file.
So avoid passing an empty parameter to "git commit".
Actually, this is a nice cleanup, as MSG_FILE and EDIT_COMMIT were mutually
exclusive; use one variable instead
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Attempting to rebase three-commit series (two regular changes, followed by
one commit that changes what commit is bound for a submodule path) to
squash the first two results in a failure; not just the first two commits
squashed, but the change to the submodule is also included in the result.
This failure causes the subsequent step to "pick" the change that actually
changes the submodule to be applied, because there is no change left to be
applied.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This functions similarly to "git branch --contains"; it will show all
tags that contain the specified commit, by sharing the same logic.
The patch also adds documentation and tests for the new option.
Signed-off-by: Jake Goulding <goulding@vivisimo.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Most of the existing codepaths were meant to treat missing uninteresting
objects to be a silently ignored non-error, but there were a few places
in handle_commit() and add_parents_to_list(), which are two key functions
in the revision traversal machinery, that cared:
- When a tag refers to an object that we do not have, we barfed. We
ignore such a tag if it is painted as UNINTERESTING with this change.
- When digging deeper into the ancestry chain of a commit that is already
painted as UNINTERESTING, in order to paint its parents UNINTERESTING,
we barfed if parse_parent() for a parent commit object failed. We can
ignore such a parent commit object.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The last test case checks whether unpacked objects receive the time stamp
of the pack file. Due to different implementations of stat(2) by MSYS and
our version in compat/mingw.c, the test fails in about half of the test
runs.
Note the following facts:
- The test uses perl's -M operator to compare the time stamps. Since we
depend on MSYS perl, the result of this operator is based on MSYS's
implementation of the stat(2) call.
- NTFS on Windows records fractional seconds.
- The MSYS implementation of stat(2) *rounds* fractional seconds to full
seconds instead of truncating them. This becomes obvious by comparing the
modification times reported by 'ls --full-time $f' and 'stat $f' for
various files $f.
- Our implementation of stat(2) in compat/mingw.c *truncates* to full
seconds.
The consequence of this is that
- add_packed_git() picks up a truncated whole second modification time
from the pack file time stamp, which is then used for the loose objects,
while the pack file retains its time stamp in fractional seconds;
- but the test case compared the pack file's rounded modification times
to the loose objects' truncated modification times.
And half of the time the rounded modification time is not the same as its
truncated modification time.
The fix is that we replace perl by 'test-chmtime -v +0', which prints the
truncated whole-second mtime without modifying it.
Signed-off-by: Johannes Sixt <j6t@kdbg.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* maint:
send-pack: do not send unknown object name from ".have" to pack-objects
test-path-utils: Fix off by one, found by valgrind
get_sha1_basic(): fix invalid memory access, found by valgrind
v1.6.1 introduced ".have" extension to the protocol to allow the receiving
side to advertise objects that are reachable from refs in the repositories
it borrows from. This was meant to be used by the sending side to avoid
sending such objects; they are already available through the alternates
mechanism.
The client side implementation in v1.6.1, which was introduced with
40c155f (push: prepare sender to receive extended ref information from the
receiver, 2008-09-09) aka v1.6.1-rc1~203^2~1, were faulty in that it did
not consider the possiblity that the repository receiver borrows from
might have objects it does not know about.
This fixes it by refraining from passing missing commits to underlying
pack-objects. Revision machinery may need to be tightened further to
treat missing uninteresting objects as non-error events, but this is an
obvious and safe fix for a maintenance release that is almost good enough.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Previously, git mergetool used cat-file which does not perform git to
worktree conversion. This changes mergetool to use git checkout-index
instead which means that the temporary files used for mergetool use the
correct line endings for the platform.
Signed-off-by: Charles Bailey <charles@hashpling.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This is an evil merge, as a test added since 1.6.0 expects an incorrect
behaviour the merged commit fixes.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A patch that changes the filetype (e.g. regular file to symlink) of a path
must be split into a deletion event followed by a creation event, which
means that we need to have two independent metainfo lines for each.
However, the code reused the single set of metainfo lines.
As the blob object names recorded on the index lines are usually not used
nor validated on the receiving end, this is not an issue with normal use
of the resulting patch. However, when accepting a binary patch to delete
a blob, git-apply verified that the postimage blob object name on the
index line is 0{40}, hence a patch that deletes a regular file blob that
records binary contents to create a blob with different filetype (e.g. a
symbolic link) failed to apply. "git am -3" also uses the blob object
names recorded on the index line, so it would also misbehave when
synthesizing a preimage tree.
This moves the code to generate metainfo lines around, so that two
independent sets of metainfo lines are used for the split halves.
Additional tests by Jeff King.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Some shells have issues with a single-shot environment variable export
when invoking a shell function. This fixes the ones I found that invoke
test_must_fail that way.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
d911d14 (rebase -i: learn to rebase root commit, 2009-01-02) tried to
remember the --root flag across a merge conflict in a broken way.
Introduce a flag file $DOTEST/rebase-root to fix and clarify.
While at it, also make sure $UPSTREAM is always initialized to guard
against existing values in the environment.
[tr: added tests]
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This new option tells 'git-am' to ignore the date header field
recorded in the format-patch output. The commits will have the
timestamp when they are created instead.
You can work a lot in one day to accumulate many changes, but
apply and push to the public repository only some of them at
the end of the first day. Then next day you can spend all your
working hours reading comics or chatting with your coworkers,
and apply your remaining patches from the previous day using
this option to pretend that you have been working at the end
of the day.
Signed-off-by: しらいしななこ <nanako3@lavabit.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This new option tells 'git-am' to use the timestamp recorded
in the Email message as both author and committer date.
Signed-off-by: しらいしななこ <nanako3@lavabit.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* js/diff-color-words:
Change the spelling of "wordregex".
color-words: Support diff.wordregex config option
color-words: make regex configurable via attributes
color-words: expand docs with precise semantics
color-words: enable REG_NEWLINE to help user
color-words: take an optional regular expression describing words
color-words: change algorithm to allow for 0-character word boundaries
color-words: refactor word splitting and use ALLOC_GROW()
Add color_fwrite_lines(), a function coloring each line individually
Added a test for this option, similar to (and based on) t9133 about
ignorance of .git directories
Signed-off-by: Vitaly "_Vi" Shukela <public_vi@tut.by>
Acked-by: Eric Wong <normalperson@yhbt.net>
[ew: replaced 'echo -e' with printf so it works on POSIX shells]
[ew: added Vitaly to copyright even though it's based on my test]
Remove a call to git-log that I introduced for debugging and that
accidentally made it into d18ba22 (sha1_name: support @{-N} syntax in
get_sha1(), 2009-01-17).
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* tr/maint-no-index-fixes:
diff --no-index -q: fix endless loop
diff --no-index: test for pager after option parsing
diff: accept -- when using --no-index
With --reject, git-am simply passes the --reject option to git-apply and thus
allows people to work with reject files if they so prefer.
Signed-off-by: martin f. krafft <madduck@madduck.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add testcases for 'git log --diff-filter=[CM]' (copies and renames).
Also add a testcase for 'git log --follow'.
Signed-off-by: Arjen Laarhoven <arjen@yaph.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Cloning an empty repository manually (that is, doing 'git init' and
then doing all configuration by hand) can be a lot of work. Save the
user this work by allowing the cloning of empty repositories.
Signed-off-by: Sverre Rabbelier <srabbelier@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Use "wordRegex" for configuration variable names. Use "word_regex" for C
language tokens.
Signed-off-by: Boyd Stephen Smith Jr. <bss@iguanasuicide.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If a piece of code wanted to do some cleanup before exiting
(e.g., cleaning up a lockfile or a tempfile), our usual
strategy was to install a signal handler that did something
like this:
do_cleanup(); /* actual work */
signal(signo, SIG_DFL); /* restore previous behavior */
raise(signo); /* deliver signal, killing ourselves */
For a single handler, this works fine. However, if we want
to clean up two _different_ things, we run into a problem.
The most recently installed handler will run, but when it
removes itself as a handler, it doesn't put back the first
handler.
This patch introduces sigchain, a tiny library for handling
a stack of signal handlers. You sigchain_push each handler,
and use sigchain_pop to restore whoever was before you in
the stack.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When diff is invoked with --color-words (w/o =regex), use the regular
expression the user has configured as diff.wordregex.
diff drivers configured via attributes take precedence over the
diff.wordregex-words setting. If the user wants to change them, they have
their own configuration variables.
Signed-off-by: Boyd Stephen Smith Jr <bss@iguanasuicide.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
All the other config variables use CamelCase. This config variable should
not be an exception.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In 'man 1p trap' there is written:
"Implementations may permit names with the SIG prefix or ignore case
in signal names as an extension."
So change the lowercase signals to uppercase, which is POSIX compliant
instead of being an extension.
There wasn't anybody claiming that it doesn't work, but there was a bug
with using a signal with the SIG prefix, which is an extension as well.
So let's play it safe and change it, since it doesn't hurt anyone.
While at it, also convert 8 indentation spaces to 1 tab character.
Signed-off-by: Markus Heidelberg <markus.heidelberg@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The code used to misbehave when options to ignore certain whitespaces
(-w -b and --ignore-at-eol) were combined.
Signed-off-by: Keith Cascio <keith@cs.ucla.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There are three flags involved (-w -b and --ignore-space-at-eol) which
makes 8 combinations possible in total, but only 3 cases are tested (none,
-w alone and -b alone).
This adds the other 5 cases.
Signed-off-by: Keith Cascio <keith@cs.ucla.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
To do that, Git no longer looks forward for the '@{' corresponding to the
closing '}' but backward, and dwim_ref() as well as dwim_log() learnt
about the @{-<N>} notation.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
You can have quite a many reflog entries, but you typically won't recall
which branch you were on after switching branches for more than several
times.
Instead of reading the reflog twice, this reads the branch switching event
and keeps as many entries as the user asked from the latest such entries,
which is the minimum required to be able to switch back to the branch we
were recently on.
[jc: improvements from Dscho squashed in]
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Some SVN repositories contain git repositories within them
(hopefully accidentally checked in). Since git refuses to track
nested ".git" repositories, this can be a problem when fetching
updates from SVN.
Thanks to Morgan Christiansson for the report and testing.
Signed-off-by: Eric Wong <normalperson@yhbt.net>
The get_log() function in the Perl SVN API introduced the limit
parameter in 1.2.0. However, this got discarded in our SVN::Ra
compatibility layer when used with SVN 1.1.x. We now emulate
the limit functionality in older SVN versions by preventing the
original callback from being called if the given limit has been
reached. This emulation is less bandwidth efficient, but SVN
1.1.x is becoming rarer now.
Additionally, the --limit parameter in svn(1) uses the
aforementioned get_log() functionality change in SVN 1.2.x.
t9129 no longer depends on --limit to work and instead uses
Perl to parse out the commit message.
Thanks to Tom G. Christensen for the bug report.
Signed-off-by: Eric Wong <normalperson@yhbt.net>
This is a followup to 7fc35e0e94,
(workaround a for broken symlinks in SVN).
Since broken SVN clients can commit svn:special files without
the magic "link " prefix, this can affect delta application
when we update the broken svn:special file. So now we fall
back and retry the delta application on symlinks if having
a "link " prefix fails.
Our behavior differs from svn(1) (v1.5.1) slightly:
When a svn:special file is created w/o a "link " prefix, svn
will create a regular file (mode 100644 to git) with the
contents of the blob as-is.
Our behavior is to continue creating the symlink (mode 120000
to git) with the contents of the blob as-is. While this
differs from current svn(1) behavior, this is easier and more
efficient to implement (and the correctness of the svn(1) is
debatable, since it's a workaround for a bug in the first
place).
More information on this SVN bug is described here:
http://subversion.tigris.org/issues/show_bug.cgi?id=2692
Signed-off-by: Eric Wong <normalperson@yhbt.net>
Broken SVN clients generate empty files with the svn:special set
to '*'. This attempts to denote a symlink pointing to a file
with an empty path (""), which cannot be generated on a POSIX
system.
Thus, we mimic the behavior of svn(1) and create a zero-byte
file in our tree.
Signed-off-by: Eric Wong <normalperson@yhbt.net>
The ability to "...fatten [the] local repository by packing everything that
is needed by the local ref into a single new pack, including things that are
borrowed from alternates"[1] is supposed to be provided by the '-a' or '-A'
options to repack when '-l' is not used, but there is a flaw. For each
pack in the local repository without a .keep file, repack supplies a
--unpacked=<pack> argument to pack-objects.
The --unpacked option to pack-objects, with or without an argument, causes
pack-objects to ignore any object which is packed in a pack not mentioned
in an argument to --unpacked=. So, if there are local packs, and
'repack -a' is called, then any objects which reside in packs accessible
through alternates will _not_ be packed. If there are no local packs, then
no --unpacked argument will be supplied, and repack will behave as expected.
[1] http://mid.gmane.org/7v8wrwidi3.fsf@gitster.siamese.dyndns.org
Signed-off-by: Brandon Casey <casey@nrlssc.navy.mil>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There are three flags involved (-w -b and --ignore-space-at-eol) which
makes 8 combinations possible in total, but only 3 cases are tested (none,
-w alone and -b alone).
This adds the other 5 cases.
Signed-off-by: Keith Cascio <keith@cs.ucla.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* tr/rebase-root:
rebase: update documentation for --root
rebase -i: learn to rebase root commit
rebase: learn to rebase root commit
rebase -i: execute hook only after argument checking
* tr/maint-no-index-fixes:
diff --no-index -q: fix endless loop
diff --no-index: test for pager after option parsing
diff: accept -- when using --no-index
* maint:
Update draft release notes for 1.6.1.1
builtin-fsck: fix off by one head count
t5540: clarify that http-push does not handle packed-refs on the remote
http-push: when making directories, have a trailing slash in the path name
http-push: fix off-by-path_len
Documentation: let asciidoc align related options
githooks.txt: add missing word
builtin-commit.c: do not remove COMMIT_EDITMSG
"git bundle create x master master" used to create a bundle that lists
the same branch (master) twice. Cloning from such a bundle resulted in
a needless warning "warning: Duplicated ref: refs/remotes/origin/master".
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When HEAD is detached, --all should list it, too, logically, as a
detached HEAD is by definition a temporary, unnamed branch.
It is especially necessary to list it when garbage collecting, as
the detached HEAD would be trashed.
Noticed by Thomas Rast.
Note that this affects creating bundles with --all; I contend that it
is a good change to add the HEAD, so that cloning from such a bundle
will give you a current branch. However, I had to fix t5701 as it
assumed that --all does not imply HEAD.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If the current working directory is a subdirectory of the gitdir (e.g.
<repo>/.git/refs/), then setup_git_directory_gently() will climb its
parent directories until it finds itself in a gitdir. However, no
matter how many parent directories it climbs, it sets
'GIT_DIR_ENVIRONMENT' to ".", which is obviously wrong.
This behaviour affected at least 'git rev-parse --git-dir' and hence
caused some errors in bash completion (e.g. customized command prompt
when on a detached head and completion of refs).
To fix this, we set the absolute path of the found gitdir instead.
Signed-off-by: SZEDER Gábor <szeder@ira.uka.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Have '-' mean the same as '@{-1}', i.e., the last branch we were on.
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Let get_sha1() parse the @{-N} syntax, with docs and tests.
Note that while @{-1}^2, @{-2}~5 and such are supported, @{-1}@{1} is
currently not allowed.
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Manipulating the character class table in ctype.c by hand is error prone.
To ensure that typos are found quickly, add a test program and script.
test-ctype checks the output of the character class macros isspace() et.
al. by applying them on all possible char values and consulting a list of
all characters in the particular class. It doesn't check tolower() and
toupper(); this could be added later.
The test script t0070-fundamental.sh is created because there is no good
place for the ctype test, yet -- except for t0000-basic.sh perhaps, but
it doesn't run well on Windows, yet.
Signed-off-by: Rene Scharfe <rene.scharfe@lsrfire.ath.cx>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The function lock_remote() sends MKCOL requests to make leading
directories; However, if it does not put a forward slash '/' at the end of
the path, the server sends a 301 redirect.
By leaving the '/' in place, we can avoid this additional step.
Incidentally, at least one version of Curl (7.16.3) does not resend
credentials when it follows a 301 redirect, so this commit also fixes
a bug.
Original patch by Tay Ray Chuan <rctay89@gmail.com>.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Currently, the only colors available to --pretty=format
users are red, green, and blue. Rather than expand it with a
few new colors, this patch makes the usual config color
syntax available, including more colors, backgrounds, and
attributes.
Because colors are no longer bounded to a single word (e.g.,
%Cred), this uses a more advanced syntax that features a
beginning and end delimiter (but the old syntax still
works). So you can now do:
git log --pretty=tformat:'%C(yellow)%h%C(reset) %s'
to emulate --pretty=oneline, or even
git log --pretty=tformat:'%C(cyan magenta bold)%s%C(reset)'
if you want to relive the awesomeness of 4-color CGA.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Make the --color-words splitting regular expression configurable via
the diff driver's 'wordregex' attribute. The user can then set the
driver on a file in .gitattributes. If a regex is given on the
command line, it overrides the driver's setting.
We also provide built-in regexes for the languages that already had
funcname patterns, and add an appropriate diff driver entry for C/++.
(The patterns are designed to run UTF-8 sequences into a single chunk
to make sure they remain readable.)
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In some applications, words are not delimited by white space. To
allow for that, you can specify a regular expression describing
what makes a word with
git diff --color-words='[A-Za-z0-9]+'
Note that words cannot contain newline characters.
As suggested by Thomas Rast, the words are the exact matches of the
regular expression.
Note that a regular expression beginning with a '^' will match only
a word at the beginning of the hunk, not a word at the beginning of
a line, and is probably not what you want.
This commit contains a quoting fix by Thomas Rast.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Up until now, the color-words code assumed that word boundaries are
identical to white space characters.
Therefore, it could get away with a very simple scheme: it copied the
hunks, substituted newlines for each white space character, called
libxdiff with the processed text, and then identified the text to
output by the offsets (which agreed since the original text had the
same length).
This code was ugly, for a number of reasons:
- it was impossible to introduce 0-character word boundaries,
- we had to print everything word by word, and
- the code needed extra special handling of newlines in the removed part.
Fix all of these issues by processing the text such that
- we build word lists, separated by newlines,
- we remember the original offsets for every word, and
- after calling libxdiff on the wordlists, we parse the hunk headers, and
find the corresponding offsets, and then
- we print the removed/added parts in one go.
The pre and post samples in the test were provided by Santi Béjar.
Note that there is some strange special handling of hunk headers where
one line range is 0 due to POSIX: in this case, the start is one too
low. In other words a hunk header '@@ -1,0 +2 @@' actually means that
the line must be added after the _second_ line of the pre text, _not_
the first.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* maint:
t3404: Add test case for auto-amending only edited commits after "edit"
t3404: Add test case for aborted --continue after "edit"
t3501: check that commits are actually done
* maint-1.6.0:
t3404: Add test case for auto-amending only edited commits after "edit"
t3404: Add test case for aborted --continue after "edit"
t3501: check that commits are actually done
Add a test case for the bugfix introduced by commit c14c3c82d
"git-rebase--interactive: auto amend only edited commit".
Signed-off-by: Stephan Beyer <s-beyer@gmx.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add a test case for the bugfix introduced by commit 8beb1f33d
"git-rebase-interactive: do not squash commits on abort".
Signed-off-by: Stephan Beyer <s-beyer@gmx.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The basic idea of t3501 is to check whether revert
and cherry-pick works on renamed files.
But as there is no pure cherry-pick/revert test, it is
good to also check if commits are actually done in that
scenario.
Signed-off-by: Stephan Beyer <s-beyer@gmx.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* maint:
Update draft release notes to 1.6.1.1
Make t3411 executable
fix handling of multiple untracked files for git mv -k
add test cases for "git mv -k"
The line length was read from the same position every time,
causing mangled output when printing notes with multiple lines.
Also, adding new-line manually for each line ensures that we
get a new-line between commits, matching git-log for commits
without notes.
Signed-off-by: Tor Arne Vestbø <tavestbo@trolltech.com>
Acked-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The "-k" option to "git mv" should allow specifying multiple untracked
files. Currently, multiple untracked files raise an assertion if they
appear consecutively as arguments. Fix this by decrementing the loop
index after removing one entry from the array of arguments.
Signed-off-by: Michael J Gruber <git@drmicha.warpmail.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add test cases for ignoring nonexisting and untracked files using the -k
option to "git mv". There is one known breakage related to multiple
untracked files specfied as consecutive arguments.
Signed-off-by: Michael J Gruber <git@drmicha.warpmail.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* cb/maint-merge-recursive-fix:
merge-recursive: do not clobber untracked working tree garbage
modify/delete conflict resolution overwrites untracked file
Conflicts:
builtin-merge-recursive.c
Without any explicit -o parameter, we correctly avoided putting the
resulting patch output to the toplevel. We should do the same when
the user gave a relative pathname to be consistent with this case.
Noticed by Cesar Eduardo Barros.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
At present we do headers unfolding (see RFC822 3.1.1. LONG HEADER FIELDS) for
all fields except 'From' (always) and 'Subject' (when keep_subject is set)
Not unfolding 'From' is a bug -- see above-mentioned RFC link.
Signed-off-by: Kirill Smelkov <kirr@landau.phys.spbu.ru>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Thanks to a200337 (git-am: propagate -C<n>, -p<n> options as well,
2008-12-04) and commits around it, "git am" is equipped to correctly
propagate the command line flags such as -C/-p/-whitespace across a patch
failure and restart.
It is trivial to support --directory option now, resurrecting previous
attempts by Kevin and Simon.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Teach git-rebase -i a new option --root, which instructs it to rebase
the entire history leading up to <branch>. This is mainly for
symmetry with ordinary git-rebase; it cannot be used to edit the root
commit in-place (it requires --onto <newbase>). Commits that already
exist in <newbase> are skipped.
In the normal mode of operation, this is fairly straightforward. We
run cherry-pick in a loop, and cherry-pick has supported picking the
root commit since f95ebf7 (Allow cherry-picking root commits,
2008-07-04).
In --preserve-merges mode, we track the mapping from old to rewritten
commits and use it to update the parent list of each commit. In this
case, we define 'rebase -i -p --root --onto $onto $branch' to rewrite
the parent list of all root commit(s) on $branch to contain $onto
instead.
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Teach git-rebase a new option --root, which instructs it to rebase the
entire history leading up to <branch>. This option must be used with
--onto <newbase>, and causes commits that already exist in <newbase>
to be skipped. (Normal operation skips commits that already exist in
<upstream> instead.)
One possible use-case is with git-svn: suppose you start hacking
(perhaps offline) on a new project, but later notice you want to
commit this work to SVN. You will have to rebase the entire history,
including the root commit, on a (possibly empty) commit coming from
git-svn, to establish a history connection. This previously had to
be done by cherry-picking the root commit manually.
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
git-notes have the potential of being pretty expensive, so test with
a lot of commits. A lot. So to make things cheaper, you have to
opt-in explicitely, by setting the environment variable
GIT_NOTES_TIMING_TESTS.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
I'm not sure how often this functionality is used, but in case
it's not, having an extra test here will help catch breakage
sooner.
Signed-off-by: Eric Wong <normalperson@yhbt.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The die() message updated accordingly.
The previous behaviour was to only allow cloning when the destination
directory doesn't exist.
[jc: added trivial tests]
Signed-off-by: Alexander Potashev <aspotashev@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When native language (RU) is in use, subject header usually contains several
parts, e.g.
Subject: [Navy-patches] [PATCH]
=?utf-8?b?0JjQt9C80LXQvdGR0L0g0YHQv9C40YHQvtC6INC/0LA=?=
=?utf-8?b?0LrQtdGC0L7QsiDQvdC10L7QsdGF0L7QtNC40LzRi9GFINC00LvRjyA=?=
=?utf-8?b?0YHQsdC+0YDQutC4?=
This exposes several bugs in builtin-mailinfo.c:
1. decode_b_segment: do not append explicit NUL -- explicit NUL was preventing
correct header construction on parts concatenation via strbuf_addbuf in
decode_header_bq. Fixes:
-Subject: Изменён список пакетов необходимых для сборки
+Subject: Изменён список па
Then
2. Do not emit '\n' between "encoded-word" where RFC2046 says that linear
white space between them are ignored when displaying. Fixes:
-Subject: Изменён список пакетов необходимых для сборки
+Subject: Изменён список па кетов необходимых для сборки
Signed-off-by: Kirill Smelkov <kirr@mns.spb.ru>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The "--signoff" test case in t7500-commit.sh was setting VISUAL while
using -F -, which indeed tested that the editor is not spawned with -F.
However, having it there was confusing, since there was no obvious reason
to the casual reader for it to be there.
This commits removes the setting of VISUAL from the --signoff test, and
adds in t7501-commit.sh a dedicated test case, where the rest of tests for
-F are.
Signed-off-by: Adeodato Simó <dato@net.com.org.es>
Okay-then-by: Johannes Schindelin <Johannes.Schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
git_commit_non_empty_tree is added to the functions that can be run from
commit filters. Its effect is to commit only commits actually touching the
tree and that are not merge points either.
The option --prune-empty is added. It defaults the commit-filter to
'git_commit_non_empty_tree "$@"', and can be used with any other
combination of filters, except --commit-hook that must used
'git_commit_non_empty_tree "$@"' where one puts 'git commit-tree "$@"'
usually to achieve the same result.
Signed-off-by: Pierre Habouzit <madcoder@debian.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous "parse-opt"ification broke git-apply reading from the
standard input. "git apply A - C <B" is supposed to read patches from
files A, B and C in this order.
Before "parse-opt"ification, we used be able to:
git apply --stat - --apply <A B
to read the patch from file A, showing only the diffstat, and then read the
patch from file B, showing the diffstat and actually applying it. Even
with this fix we cannot do that anymore, but that is so crazy use case I
do not think anybody sane relied on such a broken behaviour.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This commit teaches Git to produce diff output using the patience diff
algorithm with the diff option '--patience'.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Accept -- as an "end of options" marker even when using --no-index.
Previously, the -- triggered a "normal" index/tree diff and subsequently
failed because of the unrecognized (in that mode) --no-index.
Note that the second loop can treat '--' as a normal option, because
the preceding checks ensure it is the third-to-last argument.
While at it, fix the parsing of "-q" option in --no-index mode as well.
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Commit 0cf73755 (unpack-trees.c: assume submodules are clean during
check-out) changed an argument to verify_absent from 'path' to 'ce',
which is however shadowed by a local variable of the same name.
The bug triggers if verify_absent is used on a tree entry, for which
the index contains one or more subsequent directories of the same
length. The affected subdirectories are removed from the index. The
testcase included in this commit bisects to 55218834 (checkout: do not
lose staged removal), which reveals the bug in this case, but is
otherwise unrelated.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Commit 203a2fe1 (Allow callers of unpack_trees() to handle failure)
changed the "die on error" behavior to "return failure code".
verify_absent did not handle errors returned by
verify_clean_subdirectory, however.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
On Mac OS X and possibly BSDs, /bin/pwd reads PWD from the environment if
available and shows the logical path by default rather than the physical
one.
Unset PWD before running /bin/pwd in both cd_to_toplevel and its test.
Still use the external /bin/pwd because in my Bash on Linux, the builtin
pwd prints the same result whether or not PWD is set.
Signed-off-by: Marcel M. Cary <marcel@oak.homeunix.org>
Tested-by: Wincent Colaiuta <win@wincent.com> (on Mac OS X 10.5.5)
Tested-by: Marcel Koeppen <git-dev@marzelpan.de> (on Mac OS X 10.5.6)
Signed-off-by: Junio C Hamano <gitster@pobox.com>
It comes quite as a surprise to an unsuspecting Git user that calling
"git add submodule/file" (which is a mistake, alright) _removes_
the submodule in the index, and adds the file. Instead, complain loudly.
While at it, be nice when the user said "git add submodule/" which is
most likely the consequence of tab-completion, and stage the submodule,
instead of trying to add the contents of that directory.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
With options such as "--all --since=2.weeks.ago", annotated tags used to
be included, when they should have been excluded. The reason is that we
heavily abuse the revision walker to determine what needs to be included
or excluded. And the revision walker does not show tags at all (and
therefore never marks tags as uninteresting).
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git checkout -b newbranch $commit^{tree}" mistakenly created a new branch
rooted at the current HEAD, because in that case, the two structure fields
used to see if the command was invoked without any argument (hence it
needs to default to checking out the HEAD) were populated incorrectly.
Upon seeing a command line argument that we took as a rev, we should store
that string in new.name, even if that does not name a commit. This will
correctly trigger the existing safety logic.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Acked-by: Daniel Barkalow <barkalow@iabervon.org>
A git patch that does not change the executable bit records the mode bits
on its "index" line. "git apply" used to interpret this mode exactly the
same way as it interprets the mode recorded on "new mode" line, as the
wish by the patch submitter to set the mode to the one recorded on the
line.
The reason the mode does not agree between the submitter and the receiver
in the first place is because there is _another_ commit that only appears
on one side but not the other since their histories diverged, and that
commit changes the mode. The patch has "index" line but not "new mode"
line because its change is about updating the contents without affecting
the mode. The application of such a patch is an explicit wish by the
submitter to only cherry-pick the commit that updates the contents without
cherry-picking the commit that modifies the mode. Viewed this way, the
current behaviour is problematic, even though the command does warn when
the mode of the path being patched does not match this mode, and a careful
user could detect this inconsistencies between the patch submitter and the
patch receiver.
This changes the semantics of the mode recorded on the "index" line;
instead of interpreting it as the submitter's wish to set the mode to the
recorded value, it merely informs what the mode submitter happened to
have, and the presense of the "index" line is taken as submitter's wish to
keep whatever the mode is on the receiving end.
This is based on the patch originally done by Alexander Potashev with a
minor fix; the tests are mine.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* np/auto-thread:
Force t5302 to use a single thread
pack-objects: don't use too many threads with few objects
autodetect number of CPUs by default when using threads
Merge two hunks if there is only the specified number of otherwise unshown
context between them. For --inter-hunk-context=1, the resulting patch has
the same number of lines but shows uninterrupted context instead of a
context header line in between.
Patches generated with this option are easier to read but are also more
likely to conflict if the file to be patched contains other changes.
This patch keeps the default for this option at 0. It is intended to just
make the feature available in order to see its advantages and downsides.
Signed-off-by: Rene Scharfe <rene.scharfe@lsrfire.ath.cx>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"Assume unchanged" bit means "please pretend that I have never touched
this file", so if user removes the file, we should not care.
This patch teaches "git grep" to use cache version in such
situations. External grep case has not been fixed yet. But given that
on the platform that CE_VALID bit may be used like Windows, external
grep is not available anyway, I would wait for people to raise their
hands before touching it.
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In 212945d4 ("Teach git-describe to verify annotated tag names
before output") git-describe learned how to output a warning if
an annotated tag object was matched but its internal name doesn't
match the local ref name.
However, "git describe --all" causes the local ref name to be
prefixed with "tags/", so we need to skip over this prefix before
comparing the local ref name with the name recorded inside of the
tag object.
Patch-by: René Scharfe <rene.scharfe@lsrfire.ath.cx>
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When a file was renamed in one branch, but deleted in the other, one
should expect the index to contain an unmerged entry, namely the
target of the rename. Make it so.
Noticed by Constantine Plotnikov.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* js/rebase-i-p:
rebase -i -p: leave a --cc patch when a merge could not be redone
rebase -i -p: Fix --continue after a merge could not be redone
Show a failure of rebase -p if the merge had a conflict
The script 'git notes' allows you to edit and show commit notes, by
calling either
git notes show <commit>
or
git notes edit <commit>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We cleaned up lockfiles upon receiving the usual suspects HUP, TERM, QUIT
but a wicked user could kill us of asphyxiation by piping our output to a
pipe that does not read. Protect ourselves by catching SIGPIPE and clean
up the lockfiles as well in such a case.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
I want directories of my working tree to be linked to from various
paths on my filesystem where third-party components expect them, both
in development and production environments. A build system's install
step could solve this, but I develop scripts and web pages that don't
need to be built. Git's submodule system could solve this, but we
tend to develop, branch, and test those directories all in unison, so
one big repository feels more natural. We prefer to edit and commit
on the symlinked paths, not the canonical ones, and in that setting,
"git pull" fails to find the top-level directory of the repository
while other commands work fine.
"git pull" fails because POSIX shells have a notion of current working
directory that is different from getcwd(). The shell stores this path
in PWD. As a result, "cd ../" can be interpreted differently in a
shell script than chdir("../") in a C program. The shell interprets
"../" by essentially stripping the last textual path component from
PWD, whereas C chdir() follows the ".." link in the current directory
on the filesystem. When PWD is a symlink, these are different
destinations. As a result, Git's C commands find the correct
top-level working tree, and shell scripts do not.
Changes:
* When interpreting a relative upward (../) path in cd_to_toplevel,
prepend the cwd without symlinks, given by /bin/pwd
* Add tests for cd_to_toplevel and "git pull" in a symlinked
directory that failed before this fix, plus contrasting scenarios
that already worked
Signed-off-by: Marcel M. Cary <marcel@oak.homeunix.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When a merge that has a conflict was rebased, then rebase stopped to let
the user resolve the conflicts. However, thereafter --continue failed
because the author-script was not saved. (This is rebase -i's way to
preserve a commit's authorship.) This fixes it by doing taking the same
failure route after a merge that is also taken after a normal cherry-pick.
Signed-off-by: Johannes Sixt <j6t@kdbg.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This extends t3409-rebase-preserve-merges by a case where the merge that
is rebased has a conflict. Therefore, the rebase stops and expects that
the user resolves the conflict. However, currently rebase --continue
fails because .git/rebase-merge/author-script is missing.
The test script had allocated two identical clones, but only one of them
(clone2) was used. Now we use both as indicated in the comment. Also,
two instances of && was missing in the setup part.
Signed-off-by: Johannes Sixt <j6t@kdbg.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Even though newer Porcelain tools always record the tagger information
when creating new tags, export/import pair should be able to faithfully
reproduce ancient tag objects that lack tagger information.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Acked-by: Shawn O. Pearce <spearce@spearce.org>
When no tagger was found (old Git produced tags like this),
no "tagger" line is printed (but this is incompatible with the current
git fast-import).
Alternatively, you can pass the option --fake-missing-tagger, forcing
fast-export to fake a tagger
Unspecified Tagger <no-tagger>
with a tag date of the beginning of (Unix) time in the case of a missing
tagger, so that fast-import is still able to import the result.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add a test which checks that negated patterns such as "!foo.html" can
override previous patterns such as "*.html". This is documented
behaviour but had not been tested so far.
Signed-off-by: Michael J Gruber <git@drmicha.warpmail.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* maint:
fast-import: close pack before unlinking it
pager: do not dup2 stderr if it is already redirected
git-show: do not segfault when showing a bad tag
If the packs are made using multiple threads, they are no longer identical
on the 4-core Xeon I tested on.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Acked-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* cb/maint-merge-recursive-fix:
merge-recursive: do not clobber untracked working tree garbage
modify/delete conflict resolution overwrites untracked file
When merge-recursive wanted to create a new file in the work tree (either
as the final result, or a hint for reference purposes while delete/modify
conflicts), it unconditionally overwrote an untracked file in the working
tree. Be careful not to lose whatever the user has that is not tracked.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If a file was removed in HEAD, but modified in MERGE_HEAD, recursive merge
will result in a "CONFLICT (delete/modify)". If the (now untracked) file
already exists and was not added to the index, it is overwritten with the
conflict resolution contents.
In similar situations (cf. test 2), the merge would abort with
"error: Untracked working tree 'file' would be overwritten by merge."
The same should happen in this case.
Signed-off-by: Clemens Buchacher <drizzd@aon.at>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When a tag points at a bad or nonexistent object, we should diagnose the
breakage and exit. An earlier commit 4f3dcc2 (Fix 'git show' on signed
tag of signed tag of commit, 2008-07-01) lost this check and made it
segfault instead; not good.
This fixes it.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In dc871831(Only use GIT_CONFIG in "git config", not other programs),
GIT_CONFIG_LOCAL was rested in peace, in favor of not reading
/etc/gitconfig and $HOME/.gitconfig at all when GIT_CONFIG is set.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Make it easier to recover from a mistaken branch deletion by displaying the
sha1 of the branch's tip commit.
Update t3200 test to match the change in output.
Signed-off-by: Brandon Casey <casey@nrlssc.navy.mil>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When I tweaked the patch to use $SHELL_PATH instead of a hard-coded
"#!/bin/sh" to produce 3aa1f7c (diff: respect textconv in rewrite diffs,
2008-12-09), I screwed up. This should fix it.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Currently we just skip rewrite diffs for binary files; this
patch makes an exception for files which will be textconv'd,
and actually performs the textconv before generating the
diff.
Conceptually, rewrite diffs should be in the exact same
format as the a non-rewrite diff, except that we refuse to
share any context. Thus it makes very little sense for "git
diff" to show a textconv'd diff, but for "git diff -B" to
show "Binary files differ".
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The current emit_rewrite_diff code always writes a text patch without
checking whether the content is binary. This means that if you end up with
a rewrite diff for a binary file, you get lots of raw binary goo in your
patch.
Instead, if we have binary files, then let's just skip emit_rewrite_diff
altogether. We will already have shown the "dissimilarity index" line, so
it is really about the diff contents. If binary diffs are turned off, the
"Binary files a/file and b/file differ" message should be the same in
either case. If we do have binary patches turned on, there isn't much
point in making a less-efficient binary patch that does a total rewrite;
no human is going to read it, and since binary patches don't apply with
any fuzz anyway, the result of application should be the same.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
find_parent_branch generates branch@rev type branches when one has to
look back through SVN history to properly get the history for a branch
copied from somewhere not already being tracked by git-svn. If in the
process of fetching this history, git-svn is interrupted, then when one
fetches again, it will use whatever was last fetched as the parent
commit and fail to fetch any more history which it didn't get to before
being terminated. This is especially troubling in that different
git-svn copies of the same SVN repository can end up with different
commit sha1s, incorrectly showing the history as divergent and
precluding easy collaboration using git push and fetch.
To fix this, when we initialise the Git::SVN object $gs to search for
and perhaps fetch history, we check if there are any commits in SVN in
the range between the current revision $gs is at, and the top revision
for which we were asked to fill history. If there are commits we're
missing in that range, we continue the fetch from the current revision
to the top, properly getting all history before using it as the parent
for the branch we're trying to create.
Signed-off-by: Deskin Miller <deskinm@umich.edu>
Acked-by: Eric Wong <normalperson@yhbt.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In insert_file() subroutine (which is used to insert HTML fragments as
custom header, footer, hometext (for projects list view), and per
project README.html (for summary view)) we used:
map(to_utf8, <$fd>);
This doesn't work, and other form has to be used:
map { to_utf8($_) } <$fd>;
Now with test for t9600 added, for $GIT_DIR/README.html.
Signed-off-by: Jakub Narebski <jnareb@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
According to the message of commit 0fe7c1de16,
"git diff" with three or more trees expects the merged tree first followed by
the parents, in order. However, this command reversed the order of its
arguments, resulting in confusing diffs. A comment /* Again, the revs are all
reverse */ suggested there was a reason for this, but I can't figure out the
reason, so I removed the reversal of the arguments. Test case included.
Signed-off-by: Matt McCutchen <matt@mattmccutchen.net>
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
All other state files use dash in their names, not underscores.
Also, there is no reason to call this "extra". Drop it.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
These tests make sure that "git am" does not lose command line options
specified when it was started, after it is interrupted by a patch that
does not apply earlier in the series.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The 'branch' subcommand incorrectly had the svn-remote to use hardcoded
as 'svn', the default remote name. This meant that branches derived
from other svn-remotes would try to use the branch and tag configuration
for the 'svn' remote, potentially copying would-be branches to the wrong
place in SVN, into the branch namespace for another project.
Fix this by using the remote name extracted from the svn info for the
specified git ref. Add a testcase for this behaviour.
[jc: squashed in a fix to test from Michael J Gruber for older svn (1.4)]
Signed-off-by: Deskin Miller <deskinm@umich.edu>
Acked-by: Eric Wong <normalperson@yhbt.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* jc/rm-i-t-a:
git add --intent-to-add: do not let an empty blob be committed by accident
git add --intent-to-add: fix removal of cached emptiness
builtin-rm.c: explain and clarify the "local change" logic
Extend index to save more flags
* bc/maint-keep-pack:
repack: only unpack-unreachable if we are deleting redundant packs
t7700: test that 'repack -a' packs alternate packed objects
pack-objects: extend --local to mean ignore non-local loose objects too
sha1_file.c: split has_loose_object() into local and non-local counterparts
t7700: demonstrate mishandling of loose objects in an alternate ODB
builtin-gc.c: use new pack_keep bitfield to detect .keep file existence
repack: do not fall back to incremental repacking with [-a|-A]
repack: don't repack local objects in packs with .keep file
pack-objects: new option --honor-pack-keep
packed_git: convert pack_local flag into a bitfield and add pack_keep
t7700: demonstrate mishandling of objects in packs with a .keep file
There are printfs around that do not grok '\1', but need '\01'.
Discovered on AIX 4.3.x.
Signed-off-by: Johannes Sixt <j6t@kdbg.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The patch that allows "git bisect skip" to be passed a range of
commits using the "<commit1>..<commit2>" notation is flawed because
it introduces a regression when it was passed a simple rev or commit.
"git bisect skip <commit>" doesn't work any more, because <commit>
is quoted but not properly unquoted.
This patch fixes that and add tests cases to better check when it is
passed commits and range of commits.
While at it, this patch also properly quotes the non range arguments
using the "sq" function.
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
The loop picks elements from @ARGV one by one, sifts them into arguments
meant for format-patch and the script itself, and pushes them to @files
and @rev_list_opts arrays. Pick elements from @ARGV starting at the
beginning using shift, instead of at the end using pop, as push appends
them to the end of the array.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Writing a tree out of an index with an "intent to add" entry is blocked.
This implies that you cannot "git commit" from such a state; however you
can still do "git commit -a" or "git commit $that_path".
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This uses the extended index flag mechanism introduced earlier to mark
the entries added to the index via "git add -N" with CE_INTENT_TO_ADD.
The logic to detect an "intent to add" entry for the purpose of allowing
"git rm --cached $path" is tightened to check not just for a staged empty
blob, but with the CE_INTENT_TO_ADD bit. This protects an empty blob that
was explicitly added and then modified in the work tree from being dropped
with this sequence:
$ >empty
$ git add empty
$ echo "non empty" >empty
$ git rm --cached empty
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* ph/send-email:
git send-email: ask less questions when --compose is used.
git send-email: add --annotate option
git send-email: interpret unknown files as revision lists
git send-email: make the message file name more specific.
* maint:
sha1_file.c: resolve confusion EACCES vs EPERM
sha1_file: avoid bogus "file exists" error message
git checkout: don't warn about unborn branch if -f is already passed
bash: offer refs instead of filenames for 'git revert'
bash: remove dashed command leftovers
git-p4: fix keyword-expansion regex
fast-export: use an unsorted string list for extra_refs
Add new testcase to show fast-export does not always exports all tags
The sed call used in compare_svn_head_with() uses the + quantifier, which
is not supported in the OSX version of sed. It is replaced by the
equivalent \{1,\}.
Signed-off-by: Marcel Koeppen <git-dev@marzelpan.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The original intention of 72909be (Add diff-option --ext-diff, 2007-06-30)
was to optionally allow the use of external diff viewer in "git log"
family (while keeping them disabled by default). It exposed the "allow
external diff" bit to the UI, but forgot to adjust the "git diff" codepath
that was set up to always allow use of the external diff viewer.
Noticed by Nazri Ramliy; tests by René Scharfe squashed in.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The list extra_refs contains tags and the objects referenced by them,
so that they can be handled at the end. When a tag references a
commit, that commit is added to the list using the same name.
Also, the function handle_tags_and_duplicates() relies on the order
the items were added to extra_refs, so clearly we do not want to
use a sorted list here.
Noticed by Miklos Vajna.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Tested-by: Miklos Vajna <vmiklos@frugalware.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Avoid passing cygwin pathnames to Perl. Some Perls have problems using them
Signed-off-by: Alex Riesen <raa.lkml@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The -A option calls pack-objects with the --unpack-unreachable option so
that the unreachable objects in local packs are left in the local object
store loose. But if the -d option to repack was _not_ used, then these
unpacked loose objects are redundant and unnecessary.
Update tests in t7701.
Signed-off-by: Brandon Casey <casey@nrlssc.navy.mil>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Implement git-pull --quiet and git-pull --verbose by
adding the options to git-pull and fixing verbosity
handling in git-fetch.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* np/pack-safer:
t5303: fix printf format string for portability
t5303: work around printf breakage in dash
pack-objects: don't leak pack window reference when splitting packs
extend test coverage for latest pack corruption resilience improvements
pack-objects: allow "fixing" a corrupted pack without a full repack
make find_pack_revindex() aware of the nasty world
make check_object() resilient to pack corruptions
make packed_object_info() resilient to pack corruptions
make unpack_object_header() non fatal
better validation on delta base object offsets
close another possibility for propagating pack corruption
* bc/maint-keep-pack:
t7700: test that 'repack -a' packs alternate packed objects
pack-objects: extend --local to mean ignore non-local loose objects too
sha1_file.c: split has_loose_object() into local and non-local counterparts
t7700: demonstrate mishandling of loose objects in an alternate ODB
builtin-gc.c: use new pack_keep bitfield to detect .keep file existence
repack: do not fall back to incremental repacking with [-a|-A]
repack: don't repack local objects in packs with .keep file
pack-objects: new option --honor-pack-keep
packed_git: convert pack_local flag into a bitfield and add pack_keep
t7700: demonstrate mishandling of objects in packs with a .keep file
* mv/remote-rename:
git-remote: document the migration feature of the rename subcommand
git-remote rename: migrate from remotes/ and branches/
remote: add a new 'origin' variable to the struct
Implement git remote rename
Previously, when 'repack -a' was called and there were no packs in the local
repository without a .keep file, the repack would fall back to calling
pack-objects with '--unpacked --incremental'. This resulted in the created
pack file, if any, to be missing the packed objects in the alternate object
store. Test that this specific case has been fixed.
Signed-off-by: Brandon Casey <casey@nrlssc.navy.mil>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* maint:
Start 1.6.0.5 cycle
Fix pack.packSizeLimit and --max-pack-size handling
checkout: Fix "initial checkout" detection
Remove the period after the git-check-attr summary
Conflicts:
RelNotes
If the limit was sufficiently low, having a single object written
could bust the limit (by design), but caused the remaining allowed
size to go negative for subsequent objects, which for an unsigned
variable is a rather huge limit.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since we can use the same "diff against empty tree" trick as
we do for the non-initial case, it is trivial to make this
work.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When the "-v" option is given, we put diff of what is to be committed into
the commit template, and then strip it back out again after the user has
edited it.
We used to look for the diff by searching for the "diff --git a/"
header. With diff.mnemonicprefix set in the configuration, however, this
pattern does not match. The pattern is loosened to cover this case.
Also, if the user puts their own diff in the message (e.g., as a sample
output), then we will accidentally trigger the pattern, removing part of
their output.
We can avoid doing this stripping altogether if the user didn't use "-v"
in the first place, so we know that any match we find will be a false
positive.
[jc: this fix was split out of a series originally meant for master.]
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
With this patch, --local means pack only local objects that are not already
packed.
Additionally, this fixes t7700 testing whether loose objects in an alternate
object database are repacked.
Signed-off-by: Brandon Casey <drafnel@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Loose objects residing in an alternate object database should not be packed
when the -l option to repack is used.
Signed-off-by: Brandon Casey <drafnel@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If the user created a .keep file for a local pack, then it can be inferred
that the user does not want those objects repacked.
This fixes the repack bug tested by t7700.
Signed-off-by: Brandon Casey <casey@nrlssc.navy.mil>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Objects residing in pack files that have an associated .keep file are not
supposed to be repacked into new pack files, but they are.
Signed-off-by: Brandon Casey <casey@nrlssc.navy.mil>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Filter out all the arguments git-send-email doesn't like to a
git format-patch command, that dumps its content to a safe directory.
Barf when a file/revision conflict occurs, allow it to be overriden
--[no-]format-patch.
Signed-off-by: Pierre Habouzit <madcoder@debian.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Remote definition that came from $GIT_DIR/remotes/nick and
$GIT_DIR/branches/nick are migrated to [remotes "nick"] section in the
configuration file.
Signed-off-by: Miklos Vajna <vmiklos@frugalware.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Current git versions ignore everything after # (called <head> in the
following) when pushing. Older versions (before cf818348f1),
interpret #<head> as part of the URL, which make git bail out.
As branches origin from Cogito, it is the best to correct this by using
the behaviour of cg-push, that is to push HEAD to remote refs/heads/<head>.
Signed-off-by: Martin Koegler <mkoegler@auto.tuwien.ac.at>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
printf "\x01" is bad; write printf "\001" for portability.
Testing with dash is a good way to find this kind of POSIX.1 violation
breakages.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* maint:
Documentation: bisect: change a few instances of "git-cmd" to "git cmd"
Documentation: rev-list: change a few instances of "git-cmd" to "git cmd"
checkout: Don't crash when switching away from an invalid branch.
Pushing into the currently checked out branch of a non-bare
repository can be dangerous; the HEAD then loses sync with
the index and working tree, and it looks in the receiving
repo as if the pushed changes have been reverted in the
index (since they were never there in the first place).
This patch adds a safety valve that checks for this
condition and either generates a warning or denies the
update. We trigger the check only on a non-bare repository,
since a bare repo does not have a working tree (and in fact,
pushing to the HEAD branch is a common workflow for
publishing repositories).
The behavior is configurable via receive.denyCurrentBranch,
defaulting to "warn" so as not to break existing setups
(though it may, after a deprecation period, switch to
"refuse" by default). For users who know what they are doing
and want to silence the warning (e.g., because they have a
post-receive hook that reconciles the HEAD and working
tree), they can turn off the warning by setting it to false
or "ignore".
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
t5516 sets up some utility functions for starting each test
with a clean slate. However, there were a few tests added
that do not use these functions, but instead make their own
repositories.
Let's bring these in line with the rest of the tests. Not
only do we reduce the number of lines, but these tests will
benefit from any further enhancements to the utility
scripts.
The conversion is pretty straightforward. Most of the tests
created a parent/child clone relationship, for which we now
use 'testrepo' as the parent. One test looked in testrepo,
but relied on previous tests to have set it up; it now sets
up testrepo explicitly, which makes it a bit more robust to
changes in the script, as well.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When using alternates, it is possible for HEAD to end up pointing to
an invalid commit. git checkout should be able to recover from that
situation without crashing.
Signed-off-by: Alexandre Julliard <julliard@winehq.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* maint:
GIT 1.6.0.4
Update RPM spec for the new location of git-cvsserver.
push: fix local refs update if already up-to-date
do not force write of packed refs
Conflicts:
builtin-revert.c
* mv/maint-branch-m-symref:
update-ref --no-deref -d: handle the case when the pointed ref is packed
git branch -m: forbid renaming of a symref
Fix git update-ref --no-deref -d.
rename_ref(): handle the case when the reflog of a ref does not exist
Fix git branch -m for symrefs.
Commit a240de11 introduced this test and the code to make it
successful.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When in a bare repository (or .git, for that matter), git-svn would fail
to initialise properly, since git rev-parse --show-cdup would not output
anything. However, git rev-parse --show-cdup actually returns an error
code if it's really not in a git directory.
Fix the issue by checking for an explicit error from git rev-parse, and
setting $git_dir appropriately if instead it just does not output.
Signed-off-by: Deskin Miller <deskinm@umich.edu>
Acked-by: Eric Wong <normalperson@yhbt.net>
SVN itself always stores log messages in the repository as
UTF-8. git always stores/retrieves everything as raw binary
data with no transformations whatsoever.
To interact with SVN, we need to encode log messages as UTF-8
before sending them to SVN, as SVN cannot do it for us. When
retrieving log messages from SVN, we also need to (attempt to)
reencode the UTF-8 log message back to the user-specified commit
encoding.
Note, handling i18n.logoutputencoding for "git svn log" also
needs to be done in a future change.
Also, this change only deals with the encoding of commit
messages and nothing else (path names, blob content, ...).
In-Reply-To: <8b168cfb0810282014r789ac01dnec51824de1078f0@mail.gmail.com>
James North <tocapicha@gmail.com> wrote:
> Hi,
>
> I'm using git-svn on a system with ISO-8859-1 encoding. The problem is
> when I try to use "git svn dcommit" to send changes to a remote svn
> (also ISO-8859-1).
>
> Seems like git-svn is sending commit messages with utf-8 (just a
> guessing...) and they look bad on the remote svn log. E.g. "Ca?\241a
> de cami?\243n"
>
> I have tried using i18n.commitencoding=ISO-8859-1 as suggested by the
> warning when doing "git svn dcommit" but messages still are sent with
> wrong encoding.
Signed-off-by: Eric Wong <normalperson@yhbt.net>
git push normally updates local refs only after a successful push. If the
remote already has the updates -- pushed indirectly through another repository,
for example -- we forget to update local tracking refs.
Signed-off-by: Clemens Buchacher <drizzd@aon.at>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We force writing a ref if it does not exist. Originally, we only had to look
for the ref file to check if it existed. Now we have to look for a packed ref
as well. Luckily, resolve_ref already does all the work for us.
Signed-off-by: Clemens Buchacher <drizzd@aon.at>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The new rename subcommand does the followings:
1) Renames the remote.foo configuration section to remote.bar
2) Updates the remote.bar.fetch refspecs
3) Updates the branch.*.remote settings
4) Renames the tracking branches: renames the normal refs and rewrites
the symrefs to point to the new refs.
Signed-off-by: Miklos Vajna <vmiklos@frugalware.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* mv/maint-branch-m-symref:
update-ref --no-deref -d: handle the case when the pointed ref is packed
git branch -m: forbid renaming of a symref
Fix git update-ref --no-deref -d.
rename_ref(): handle the case when the reflog of a ref does not exist
Fix git branch -m for symrefs.
* sh/rebase-i-p:
git-rebase--interactive.sh: comparision with == is bashism
rebase-i-p: minimum fix to obvious issues
rebase-i-p: if todo was reordered use HEAD as the rewritten parent
rebase-i-p: do not include non-first-parent commits touching UPSTREAM
rebase-i-p: only list commits that require rewriting in todo
rebase-i-p: fix 'no squashing merges' tripping up non-merges
rebase-i-p: delay saving current-commit to REWRITTEN if squashing
rebase-i-p: use HEAD for updating the ref instead of mapping OLDHEAD
rebase-i-p: test to exclude commits from todo based on its parents
Abstract
--------
With index v2 we have a per object CRC to allow quick and safe reuse of
pack data when repacking. This, however, doesn't currently prevent a
stealth corruption from being propagated into a new pack when _not_
reusing pack data as demonstrated by the modification to t5302 included
here.
The Context
-----------
The Git database is all checksummed with SHA1 hashes. Any kind of
corruption can be confirmed by verifying this per object hash against
corresponding data. However this can be costly to perform systematically
and therefore this check is often not performed at run time when
accessing the object database.
First, the loose object format is entirely compressed with zlib which
already provide a CRC verification of its own when inflating data. Any
disk corruption would be caught already in this case.
Then, packed objects are also compressed with zlib but only for their
actual payload. The object headers and delta base references are not
deflated for obvious performance reasons, however this leave them
vulnerable to potentially undetected disk corruptions. Object types
are often validated against the expected type when they're requested,
and deflated size must always match the size recorded in the object header,
so those cases are pretty much covered as well.
Where corruptions could go unnoticed is in the delta base reference.
Of course, in the OBJ_REF_DELTA case, the odds for a SHA1 reference to
get corrupted so it actually matches the SHA1 of another object with the
same size (the delta header stores the expected size of the base object
to apply against) are virtually zero. In the OBJ_OFS_DELTA case, the
reference is a pack offset which would have to match the start boundary
of a different base object but still with the same size, and although this
is relatively much more "probable" than in the OBJ_REF_DELTA case, the
probability is also about zero in absolute terms. Still, the possibility
exists as demonstrated in t5302 and is certainly greater than a SHA1
collision, especially in the OBJ_OFS_DELTA case which is now the default
when repacking.
Again, repacking by reusing existing pack data is OK since the per object
CRC provided by index v2 guards against any such corruptions. What t5302
failed to test is a full repack in such case.
The Solution
------------
As unlikely as this kind of stealth corruption can be in practice, it
certainly isn't acceptable to propagate it into a freshly created pack.
But, because this is so unlikely, we don't want to pay the run time cost
associated with extra validation checks all the time either. Furthermore,
consequences of such corruption in anything but repacking should be rather
visible, and even if it could be quite unpleasant, it still has far less
severe consequences than actively creating bad packs.
So the best compromize is to check packed object CRC when unpacking
objects, and only during the compression/writing phase of a repack, and
only when not streaming the result. The cost of this is minimal (less
than 1% CPU time), and visible only with a full repack.
Someone with a stats background could provide an objective evaluation of
this, but I suspect that it's bad RAM that has more potential for data
corruptions at this point, even in those cases where this extra check
is not performed. Still, it is best to prevent a known hole for
corruption when recreating object data into a new pack.
What about the streamed pack case? Well, any client receiving a pack
must always consider that pack as untrusty and perform full validation
anyway, hence no such stealth corruption could be propagated to remote
repositoryes already. It is therefore worthless doing local validation
in that case.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
