This is a beginning of resurrecting the multi-head pulling support
for git-fetch-pack command. The git-fetch-script wrapper still
only knows about fetching a single head, without renaming, so it is
not very useful unless you directly call git-fetch-pack itself yet.
It also fixes a longstanding obsolete description of how the command
discovers the list of local commits.
In the case where we don't know from context what type an object is, but
we don't have to fetch it, we need to parse it to determine the type
before processing it.
Signed-off-by: Daniel Barkalow <barkalow@iabervon.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
[PATCH] Trapping exit in tests, using return for errors: further fixes.
"return" from a test would leave the exit trap set, which could cause a
spurious error message if it's the last test in the script or
--immediate is used.
The easiest solution would be to have a global trap that is set when
test-lib.sh is sourced and unset either by test_done(), error() or by
test_failure_() with --immediate.
Signed-off-by: Pavel Roskin <proski@gnu.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
[PATCH] Trapping exit in tests, using return for errors
I have noticed that "make test" fails without any explanations when the
"merge" utility is missing. I don't think tests should be silent in
case of failure.
It turned out that the particular test was using "exit" to interrupt the
test in case of an error. This caused the whole test script to exit.
No further tests would be run even if "--immediate" wasn't specified.
No error message was printed.
This patch does following:
All instances of "exit", "exit 1" and "(exit 1)" in tests have been
replaced with "return 1". In fact, "(exit 1)" had no effect.
File descriptor 5 is duplicated from file descriptor 1. This is needed
to print important error messages from tests.
New function test_run_() has been introduced. Any "return" in the test
would merely cause that function to return without skipping calls to
test_failure_() and test_ok_(). The new function also traps "exit" and
treats it like a fatal error (in case somebody reintroduces "exit" in
the tests).
test_expect_failure() and test_expect_success() check both the result of
eval and the return value of test_run_(). If the later is not 0, it's
always a failure because it indicates the the test didn't complete.
Signed-off-by: Pavel Roskin <proski@gnu.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
"t5400-send-pack.sh --verbose" stops waiting for user input. It happens
because "git log" uses less for output now. To prevent this, PAGER
should be set to cat.
Signed-off-by: Pavel Roskin <proski@gnu.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
All test scripts should end with test_done, which reports the test
results. In the future, it could be used for other purposes, e.g. to
distinguish graceful end from "exit" in a test. This patch fixes
scripts that don't call test_done.
Signed-off-by: Pavel Roskin <proski@gnu.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
- Split gitk off to its own package;
it needs tk installed, but nothing else does.
- Refer to GPL properly, don't install COPYING.
- Fix maintainer.
- Use dh_movefiles instead of dh_install;
we don't want to list everything *except* gitk.
In commit 4f7eb2e5a351e0d1f19fd4eab7e92834cc4528c2 I fixed git-merge-base
getting confused by datestamps that caused it to traverse things in a
non-obvious order.
However, my fix was a very brute-force one, and it had some really
horrible implications for more complex trees with lots of parallell
development. It might end up traversing all the way to the root commit.
Now, normally that isn't that horrible: it's used mainly for merging, and
the bad cases really tend to happen fairly rarely, so if it takes a few
seconds, we're not in too bad shape.
However, gitk will also do the git-merge-base for every merge it shows,
because it basically re-does the trivial merge in order to show the
"interesting" parts. And there we'd really like the result to be
instantaneous.
This patch does that by walking the tree more completely, and using the
same heuristic as git-rev-list to decide "ok, the rest is uninteresting".
In one - hopefully fairly extreme - case, it made a git-merge-base go from
just under five seconds(!) to a tenth of a second on my machine.
Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
Refine the update heuristic to improve responsiveness a bit.
The previous commit improved performance a lot but also meant that
we waited longer to see something drawn. This refines the heuristics
for when to call update so that (1) when we have finished processing
a bufferfull of information from git-rev-list, we call update if
enough time has elapsed, regardless of how many commits we've drawn,
and (2) the number of commits drawn between updates scales with the
total number of commits drawn: 1 for 1-99 commits, 10 for 100-9999
commits, or 100 for >= 10000 commits.
git-rev-parse HEAD^1 would fail, because of an off-by-one bug (but HEAD^
would yield the expected result). Also, when the parent does not exist, do
not silently return an incorrect SHA1. Of course, this no longer applies
to git-rev-parse alone, but every user of get_sha1().
While at it, add a test.
Signed-off-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Junio C Hamano <junkio@cox.net>
Only do an update every 100 commits when drawing the graph.
On a large repository with > 60,000 commits, each call to the Tk
update primitive (which gives Tk a chance to respond to events and
redraw the screen) was taking up to 0.2 seconds. Because the logic
was to call update after drawing a commit if 0.1 seconds had passed
since the last update call, we were calling it for every commit,
which was slowing us down enormously. Now we also require that we
have drawn 100 commits since the last update (as well as it being
at least 0.1 seconds since the last update). Drawing 100 commits
takes around 0.1 - 0.2 seconds (even in this large repo) on my G5.
Although it is uncertain if we would keep .git/branches for
long, the shorthand stored there can be used for pushing if it
is host:path/to/git format, so let's make use of it. This does
not use git-parse-remote because that script will be rewritten
quite a bit for updated pulling.
Although these commands take only begin and end, not necessarily
generic SHA1 expressions rev-parse supports, supporting a..b
notation is good for consistency. This commit adds such without
breaking backward compatibility.
Given one existing commit, revert the change the patch
introduces, and record a new commit that records it. This
requires your working tree to be clean (no modifications from
the HEAD commit).
This is based on what Linus posted to the list, with
enhancements he suggested, including the use of -M to attempt
reverting renames.
While moving '-m' to make room for CVS compatible "here is the
log message", enhance source of log parameters.
-m 'message': a command line parameter.
-F <file> : a file (use '-' to read from stdin).
-C <commit> : message in existing commit.
-c <commit> : message in existing commit (allows further editing).
Longer option names for these options are also available.
While we are at it, get rid of shell array bashism.
I haven't audited the rev-parse users, but I am having a feeling
that many of them would choke when they expect a couple of SHA1
object names and malicious user feeds them "--max-count=6" or
somesuch to shoot himself in the foot. Anyway, this adds a
couple of missing parameters that affect the list of revs to be
returned from rev-list, not the flags that affect how they are
presented by rev-list. I think that is the intention, but I am
not quite sure.
Use GIT_SSH environment to specify alternate ssh binary.
[jc: I ended up rewriting Martin's patch due to whitespace
breakage, but the credit goes to Martin for doing the initial
patch to identify what needs to be changed.]
Signed-off-by: Martin Sivak <mars@nomi.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
GCC's format __attribute__ is good for checking errors, especially
with -Wformat=2 parameter. This fixes most of the reported problems
against 2005-08-09 snapshot.
RPM folks have problem installing the package otherwise. Since
its usefulness does have much to do with GIT, downgrade it to
"contrib" status for now. We may want to move it to contrib/
subdirectory after auditing other programs when we reorganize
the source tree.
[PATCH] "Child" information in commit window - and cleanups
This adds "Child: " lines to the commit window, which tells what children
a commit has.
It also cleans things up: it marks the text widget as no-wrap, which means
that it doesn't need to truncate the commit description arbitrarily by
hand. Also, the description itself is now done by a common helper routine
that handles both the parent and the children.
Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Paul Mackerras <paulus@samba.org>
Sergey Vlasov says we do not pre-require (i.e. required packages
during installation) the dependencies, and should use Requires
instead of Prereq. Knowing nothing about RPM, I just believe
him.
My understanding is that having my name there is just as wrong
as having name of Linus, since neither of us is a debian
maintainer, but at least this would prevent people from bugging
Linus.
Better graph line details display and expand history coverage.
Now the history remembers when we have clicked on a graph line
and when we have asked for a diff between two commits, as well
as when we have displayed a commit.
The display when you click on a graph line now uses clickable
SHA1 IDs instead of the embedded "Go" buttons. Also made the
IDs clickable in the header for a diff between two commits.
Update git-apply-patch-script for symbolic links.
Make git-prune-script executable again.
Do not write out new index if nothing has changed.
diff-cache shows differences for unmerged paths without --cache.
Update diff engine for symlinks stored in the cache.
Signed-off-by: Junio C Hamano <junkio@cox.net>
where all the parent commit ID's are clickable, because the new lines are
added as part of the "comment" string, and thus the regular clickability
thing will match them automatically.
I think this is good. And my random-tcl-monkey-skills are clearly getting
better (although it's perfectly possible that somebody who actually knows
what he is doing would have done things differently).
Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Paul Mackerras <paulus@samba.org>
Use lf translation rather than binary when reading commit data.
The effect of this is that it allows Tcl to do the locale-specific
conversion of the input data to its internal unicode representation.
That means that commit messages in Russian or other languages should
be displayed correctly now (according to the locale that is in effect.)
Change cursor to a hand cursor when over a SHA1 ID link.
This is based on suggestions by Jeff Epler and Linus Torvalds, but
extended so that we do the switching between the watch cursor and
the normal cursor correctly as well.
Also fixed a bug pointed out by Junio Hamano - I wasn't incrementing
the link number (duh!).
This was triggered by a query by Sam Ravnborg, and extends "git reset" to
reset the index and the .git/HEAD pointer to an arbitrarily named point.
For example
git reset HEAD^
will just reset the current HEAD to its own parent - leaving the working
directory untouched, but effectively un-doing the top-most commit. You
might want to do this if you realize after you committed that you made a
mistake that you want to fix up: reset your HEAD back to its previous
state, fix up the working directory and re-do the commit.
If you want to totally un-do the commit (and reset your working directory
to that point too), you'd first use "git reset HEAD^" to reset to the
parent, and then do a "git checkout -f" to reset the working directory
state to that point in time too.
Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
This is my first attempt to adjust Debian and RPM to pass
prefix, to prepare the 0.99.4 release.
It updates debian/rules and git-core.spec.in to properly pass
prefix when building binary packages. It also updates
debian/changelog to make the resulting binary package name
0.99.4; this is not needed on the RPM side (it takes the version
number from the main Makefile).
Per discussion with people interested in binary packaging,
change the default template location from /etc/git-core to
/usr/share/git-core hierarchy. If a user wants to run git
before installing for whatever reason, in addition to adding
$src to the PATH environment variable, git-init-db can be run
with --template=$src/templates/blt/ parameter.
send-pack: allow generic sha1 expression on the source side.
This extends the source side semantics to match what Linus
suggested.
An example:
$ git-send-pack kernel.org:/pub/scm/git/git.git pu^^:master pu
would allow me to push the current pu into pu, and the
commit two commits before it into master, on my public
repository.
The revised rule for updating remote heads is as follows.
$ git-send-pack [--all] <remote> [<ref>...]
- When no <ref> is specified:
- with '--all', it is the same as specifying the full refs/*
path for all local refs;
- without '--all', it is the same as specifying the full
refs/* path for refs that exist on both ends;
- When one or more <ref>s are specified:
- a single token <ref> (i.e. no colon) must be a pattern that
tail-matches refs/* path for an existing local ref. It is
an error for the pattern to match no local ref, or more
than one local refs. The matching ref is pushed to the
remote end under the same name.
- <src>:<dst> can have different cases. <src> is first tried
as the tail-matching pattern for refs/* path.
- If more than one matches are found, it is an error.
- If one match is found, <dst> must either match no remote
ref and start with "refs/", or match exactly one remote
ref. That remote ref is updated with the sha1 value
obtained from the <src> sha1.
- If no match is found, it is given to get_extended_sha1();
it is an error if get_extended_sha1() does not find an
object name. If it succeeds, <dst> must either match
no remote ref and start with "refs/" or match exactly
one remote ref. That remote ref is updated with the sha1
value.
send-pack: allow the same source to be pushed more than once.
The revised code accidentally inherited the restriction that a
reference can be pushed only once, only because the original did
not allow renaming. This is no longer necessary so lift it.
Add forward and back buttons and make SHA1 IDs clickable links.
When we display the commit message in the details pane, any string
of 40 [0-9a-f] characters that corresponds to a SHA1 ID that we
know about gets turned into a clickable link, and displayed in
blue and underlined.
We now keep a history of commits that we have looked at, and we
have forward and back buttons for moving within the history list.
[jc: Johannes spent time and effort to see how consistent our
use of terminilogy is, and as a byproduct made these corrections
not related to the terminology unification. I really appreciate
it.]
Signed-off-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Junio C Hamano <junkio@cox.net>
[PATCH] git-commit-script fix for degenerated merge
If merging results in an unchanged tree, git-commit-script should not
complain that there's nothing to commit.
Also, add "[--all]" to usage().
[jc: usually there is no reason to record an unchanging merge,
but this code path is triggered only when there is a nontrivial
merge that needed to be resolved by hand, and we should be able
to record the fact that these two tree heads are dealt with as a
regular two-parent commit in order to help later merges.]
When more than two references need to be checked with
ref_newer() function, the second and later calls did not work
correctly. This was because the later calls found commits
retained by the "struct object" layer that still had smudges
made by earlier calls.
Every time after servicing the connection, select() first fails
with EINTR and ends up waiting for one second before serving the
next client. The sleep() was placed by the original author per
suggestion from the list to avoid spinning on failing select,
but at least this EINTR situation should not result in "at most
one client per second" service limit.
I am not sure if this is the right fix, but WTH. The king
penguin says that serious people would run the daemon under
inetd anyway, and I agree with that.
Now, for extra bonus points, maybe you should make "git-rev-list" also
understand the "rev..rev" format (which you can't do with just the
get_sha1() interface, since it expands into more).
Everybody envies rev-parse, who is the only one that can grok
the extended sha1 format. Move the get_extended_sha1() out of
rev-parse, rename it to get_sha1() and make it available to
everybody else.
The one I posted earlier to the list had one bug where it did
not handle a name that ends with a digit correctly (it
incorrectly tried the "Nth parent" path). This commit fixes it.
This commit makes sure that we do not barf when pushing a ref
that is a non-commitish tag. You can update a remote ref under
the following conditions:
* You can always use --force.
* Creating a brand new ref is OK.
* If the remote ref is exactly the same as what you are
pushing, it is OK (nothing is pushed).
* You can replace a commitish with another commitish which is a
descendant of it, if you can verify the ancestry between them;
this and the above means you have to have what you are replacing.
* Otherwise you cannot update; you need to use --force.
Compress the graph horizontally if it gets too wide.
If the graph gets to use more than a certain percentage (default 50%)
of the width of the top-left pane, we now reduce the amount of space
allowed for each graph line. This means it doesn't look quite as
nice but you can still see the headline for the commit. (Currently
the only way to customize the percentage is to edit your ~/.gitk
file manually.)
When I munged the original from Linus, which did not terminate
when the last bisect to check happened to be a bad one, to
terminate, I seem to have botched the end result to pick.
Thanks for Sanjoy Mahajan for a good reproduction recipe to
diagnose this.
This allows git-send-pack to push local refs to a destination
repository under different names.
Here is the name mapping rules for refs.
* If there is no ref mapping on the command line:
- if '--all' is specified, it is equivalent to specifying
<local> ":" <local> for all the existing local refs on the
command line
- otherwise, it is equivalent to specifying <ref> ":" <ref> for
all the refs that exist on both sides.
* <name> is just a shorthand for <name> ":" <name>
* <src> ":" <dst>
push ref that matches <src> to ref that matches <dst>.
- It is an error if <src> does not match exactly one of local
refs.
- It is an error if <dst> matches more than one remote refs.
- If <dst> does not match any remote refs, either
- it has to start with "refs/"; <dst> is used as the
destination literally in this case.
- <src> == <dst> and the ref that matched the <src> must not
exist in the set of remote refs; the ref matched <src>
locally is used as the name of the destination.
For example,
- "git-send-pack --all <remote>" works exactly as before;
- "git-send-pack <remote> master:upstream" pushes local master
to remote ref that matches "upstream". If there is no such
ref, it is an error.
- "git-send-pack <remote> master:refs/heads/upstream" pushes
local master to remote refs/heads/upstream, even when
refs/heads/upstream does not exist.
- "git-send-pack <remote> master" into an empty remote
repository pushes the local ref/heads/master to the remote
ref/heads/master.
A template mechanism to populate newly initialized repository
with default set of files is introduced. Use it to ship example
hooks that can be used for update and post update checks, as
Josef Weidendorfer suggests.