When the "git pull" command updates the branch head you are
currently on, before doing anything else, first update your
index file and the working tree contents to that of the new
branch head. Otherwise, the later resolving steps would think
your index file is attempting to revert the change between the
original head commit and the updated head commit.
It uses two-tree fast-forward form of "read-tree -m -u" to
prevent losing whatever local changes you may have in the
working tree to do this update. I think this would at least
make things safer (a lot safer), and prevent mistakes.
Also "git fetch" command is forbidden from fetching and fast
forwarding the current branch head unless --update-head-ok flag
is given. "git pull" passes the flag when it internally calls
"git fetch".
You can define WITH_SEND_EMAIL to include the send-email command as
part of the installation. Since Debian, unlike RPM/Fedora, has the
two necessary Perl modules available as part of the mainline
distribution, there is no reason for us to shy away from shipping
send-email.
Gitzilla updated bunch of undocumented command pages, so move the
entries in the main documentation index around to put them in proper
category. Ordering within category will be fixed later.
Various updates and cleanups for my howto on using branches in GIT
as a Linux subsystem maintainer. Three categories of changes:
1) Updates for new features in GIT 0.99.5
2) Changes to use "git fetch" rather than "git pull" to update
local linus branch.
3) Cleanups suggested by Len Brown
Signed-off-by: Tony Luck <tony.luck@intel.com> Signed-off-by: Junio C Hamano <junkio@cox.net>
The "verify_pathspec()" function doesn't test for ending NUL character in
the pathspec, causing some really funky and unexpected behaviour. It just
happened to work in the cases I had tested.
Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
[PATCH] Allow "+remote:local" refspec to cause --force when fetching.
With this we could say:
Pull: master:ko-master +pu:ko-pu
to mean "fast forward ko-master with master, overwrite ko-pu with pu",
and the latter one does not require the remote "pu" to be descendant
of local "ko-pu".
[PATCH] Make "git pull" and "git fetch" default to origin
Amos Waterland sent in a patch for the pre-multi-head aware
version of "git pull" to do this, but the code changed quite a
bit since then. If there is no argument given to pull from, and
if "origin" makes sense, default to fetch/pull from "origin"
instead of barfing.
[jc: besides, the patch by Amos broke the non-default case where
explicit refspecs are specified, and did not make sure we know
what "origin" means before defaulting to it.]
This script uses the list of heads and their origin multi-head "git
fetch" left in the $GIT_DIR/FETCH_HEAD file, and makes an octopus
merge on top of the current HEAD using them.
The implementation tries to be strict for the sake of safety. It
insists that your working tree is clean (no local changes) and matches
the HEAD, and when any of the merged heads does not automerge, the
whole process is aborted and tries to rewind your working tree is to
the original state.
Update git-pull to match updated git-fetch and allow pull to
fetch from multiple remote references. There is no support for
resolving more than two heads, which will be done with "git
octopus".
Update "git ls-remote" to use git-parse-remote-script.
- A <refspec> of form "<src>:<dst>" is to fetch the objects
needed for the remote ref that matches <src>, and if <dst>
is not empty, store it as a local <dst>.
- "tag" followed by <next> is just an old way of saying
"refs/tags/<next>:refs/tags/<next>"; this mimics the
current behaviour of the third form above and means "fetch
that tag and store it under the same name".
- A single token <refspec> without colon is a shorthand for
"<refspec>:" That is, "fetch that ref but do not store
anywhere".
- when there is no <refspec> specified
- if <remote> is the name of a file under $GIT_DIR/remotes/
(i.e. a new-style shorthand), then it is the same as giving
the <refspec>s listed on Pull: line in that file.
- if <remote> is the name of a file under $GIT_DIR/branches/
(i.e. an old-style shorthand, without trailing path), then it
is the same as giving a single <refspec>
"<remote-name>:refs/heads/<remote>" on the command line, where
<remote-name> is the remote branch name (defaults to HEAD, but
can be overridden by .git/branches/<remote> file having the
URL fragment notation). That is, "fetch that branch head and
store it in refs/heads/<remote>".
- otherwise, it is the same as giving a single <refspec>
that is "HEAD:".
The SHA1 object names of fetched refs are stored in FETCH_HEAD,
one name per line, with a comment to describe where it came from.
This is later used by "git resolve" and "git octopus".
Recent changes in git have broken cg-log. git-rev-list no longer
prints "commit" in front of commit hashes. It turn out a local
"prefix" variable in main() shadows a file-scoped "prefix" variable.
The patch removed the local "prefix" variable since its value is never
used (in the intended way, that is). The call to
setup_git_directory() is kept since it has useful side effects.
The file-scoped "prefix" variable is renamed to "commit_prefix" just
in case someone reintroduces "prefix" to hold the return value of
setup_git_directory().
Signed-off-by: Pavel Roskin <proski@gnu.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
Earlier rounds broke 'whatchanged -p'. In attempting to fix this,
make two axis of output selection in rev-parse orthogonal:
--revs-only tells it not to output things that are not revisions nor
flags that rev-list would take.
--no-revs tells it not to output things that are revisions or
flags that rev-list would take.
--flags tells it not to output parameters that do not start with
a '-'.
--no-flags tells it not to output parameters that starts with a '-'.
So for example 'rev-parse --no-revs -p arch/i386' would yield '-p arch/i386',
while 'rev-parse --no-revs --flags -p archi/i386' would give just '-p'.
Also the meaning of --verify has been made stronger. It now rejects
anything but a single valid rev argument. Earlier it passed some flags
through without complaining.
Pulling from a packed repository over dumb transport without the
server info file fails, so run update-server-info automatically
after a repack by default. This can be disabled with the '-n'
flag.
Try to find the optimum merge base while resolving.
The merge-base command acquires a new option, '--all', that causes it
to output all the common ancestor candidates. The "git resolve"
command then uses it to pick the optimum merge base by picking the one
that results in the smallest number of nontrivial merges.
This trivial patch makes "git-rev-list" able to handle not being in
the top-level directory. This magically also makes "git-whatchanged"
do the right thing.
Trivial scripting fix to make sure that "git log" also works.
Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
I have been feeling that the current behaviour of "git reset" is
not quite optimal, but so far could not express exactly what I
felt was wrong with it. This patch clarifies it.
There are at least two situations you may want to "reset" your
working tree.
1. You made a mess in your working tree. You want to switch
back to a known good state and start over. This mess may be
a result of your own editing, a merge that had too many
conflicting changes that you do not feel like to resolve by
hand at this moment, or a botched application of a patch you
received from somewhere.
In this case, you would want to have "git reset HEAD" reset
the index file to the tree read from the HEAD commit and the
files in the working tree to match index (i.e. "git status"
should say "Nothing to commit", without any "unrecorded
changes").
The current behaviour leaves the files in the working tree
intact, which requires you to run "git checkout -f". Also
you need to remember "rm -f" any files that the botched patch
may have left in the working tree if the purpose of this
"reset" is to attempt to apply it again; most likely the
patch would fail if such a file is left behind.
2. You have discovered that commits you made earlier need to be
reorganized. The simplest example is to undo the last
commit, re-edit some files, and redo the commit. Another
simple eample is to undo the last two commits, and commit the
changes in those two commits as a single commit.
In this case, you would want to have "git reset HEAD^" reset
the $GIT_DIR/HEAD to the commit object name of the parent
commit of the current commit (i.e. rewinding one commit),
leave the index file and the files in the working tree in a
state where you can easily make a commit that records a tree
that resembles what you have in the current index file and
the working tree.
The current behaviour is almost OK for this purpose, except
that you need to find which files you need to manually run
"git add" yourself. They are files that are in the original
HEAD commit and not in the commit you are resetting to.
The default without the type flag is to do "--mixed", which is
the current behaviour.
A hard reset would be used for 1 and works in this way:
(1) remember the set of paths that appear in the current
index file (which may even have unmerged entries) and
the current $GIT_DIR/HEAD commit.
(2) "read-tree --reset" the specified <commit-ish> (default
to HEAD), followed by "checkout-cache -f -u -a".
(3) remove any files that appear in (1) but not in
<commit-ish> from the working tree.
(4) backup $GIT_DIR/HEAD to $GIT_DIR/ORIG_HEAD and update
$GIT_DIR/HEAD with the specified <commit-ish>.
(5) remove leftover $GIT_DIR/MERGE_HEAD
A soft reset would be used for 2 and works in this way:
(1) Make sure that the index file is merged and we do not
have MERGE_HEAD; otherwise it does not make sense to do
soft reset.
(2) backup $GIT_DIR/HEAD to $GIT_DIR/ORIG_HEAD and update
$GIT_DIR/HEAD with the specified <commit-ish>.
Note that with the current behaviour, "git diff" is the way to
see what could be committed immediately after "git reset". With
the "soft reset" described here you would need to say "git diff
HEAD" to find that out.
I am not sure what mixed reset (the current behaviour) is good
for. If nobody comes up with a good use case it may not be a
bad idea to remove it.
Clean-up output from"git show-branch" and document it.
When showing only one branch a lot of default output becomes redundant,
so clean it up a bit, and document what is shown. Retire the earlier
implementation "git-show-branches-script".
The 'git show-branches' command turns out to be reasonably useful,
but painfully slow. So rewrite it in C, using ideas from merge-base
while enhancing it a bit more.
- Unlike show-branches, it can take --heads (show me all my
heads), --tags (show me all my tags), or --all (both).
- It can take --more=<number> to show beyond the merge-base.
- It shows the short name for each commit in the extended SHA1
syntax.
- It can find merge-base for more than two heads.
Examples:
$ git show-branch --more=6 HEAD
is almost the same as "git log --pretty=oneline --max-count=6".
$ git show-branch --merge-base master mhf misc
finds the merge base of the three given heads.
$ git show-branch master mhf misc
shows logs from the top of these three branch heads, up to their
common ancestor commit is shown.
$ git show-branch --all --more=10
is poor-man's gitk, showing all the tags and heads, and
going back 10 commits beyond the merge base of those refs.
[PATCH] Add a new extended SHA1 syntax <name>~<num>
The new notation is a short-hand for <name> followed by <num>
caret ('^') characters. E.g. "master~4" is the fourth
generation ancestor of the current "master" branch head,
following the first parents; same as "master^^^^" but a bit
more readable.
This will be used in the updated "git show-branch" command.
When "git-diff-script A..B" notation was introduced, it ended up breaking
the traditional two revisions notation.
[jc: there are other issues with the current "git diff" I would like to
address, but they would be left to later rounds. For example, -M and -p flags
should not be hardcoded default, and it shouldn't be too hard to rewrite
the script without using shell arrays.]
This generalizes the git "glob" string to be a lot more like the
git-diff-* pathspecs (but there are still differences: the diff family
doesn't do any globbing, and because the diff family always generates the
full native pathname, it doesn't have the issue with "..").
It does three things:
- it allows multiple matching strings, ie you can do things like
- the "matching" criteria is a combination of "exact path component
match" (the same as the git-diff-* family), and "fnmatch()". However,
you should be careful with the confusion between the git-ls-files
internal globbing and the standard shell globbing, ie
git-ls-files fs/*.c
does globbing in the shell, and does something totally different from
git-ls-files 'fs/*.c'
which does the globbing inside git-ls-files.
The latter has _one_ pathspec with a wildcard, and will match any .c
file anywhere under the fs/ directory, while the former has been
expanded by the shell into having _lots_ of pathspec entries, all of
which are just in the top-level fs/ subdirectory. They will happily
be matched exactly, but we will thus miss all the subdirectories under
fs/.
As a result, the first one will (on the current kernel) match 55 files,
while the second one will match 664 files!
- it uses the generic path prefixing, so that ".." and friends at the
beginning of the path spec work automatically
NOTE! When generating relative pathname output (the default), a
pathspec that causes the base to be outside the current working
directory will be rejected with an error message like:
[PATCH] Make "git-ls-files" work in subdirectories
This makes git-ls-files work inside a relative directory, and also adds
some rudimentary filename globbing support. For example, in the kernel you
can now do
cd arch/i386
git-ls-files
and it will show all files under that subdirectory (and it will have
removed the "arch/i386/" prefix unless you give it the "--full-name"
option, so that you can feed the result to "xargs grep" or similar).
The filename globbing is kind of strange: it does _not_ follow normal
globbing rules, although it does look "almost" like a normal file glob
(and it uses the POSIX.2 "fnmatch()" function).
The glob pattern (there can be only one) is always split into a "directory
part" and a "glob part", where the directory part is defined as any full
directory path without any '*' or '?' characters. The "glob" part is
whatever is left over.
For example, when doing
git-ls-files 'arch/i386/p*/*.c'
the "directory part" is is "arch/i386/", and the "glob part" is "p*/*.c".
The directory part will be added to the prefix, and handled efficiently
(ie we will not be searching outside of that subdirectory), while the glob
part (if anything is left over) will be used to trigger "fnmatch()"
matches.
This is efficient and very useful, but can result in somewhat
non-intuitive behaviour.
For example:
git-ls-files 'arch/i386/*.[ch]'
will find all .c and .h files under arch/i386/, _including_ things in
lower subdirectories (ie it will match "arch/i386/kernel/process.c",
because "kernel/process.c" will match the "*.c" specifier).
Also, while
git-ls-files arch/i386/
will show all files under that subdirectory, doing the same without the
final slash would try to show the file "i386" under the "arch/"
subdirectory, and since there is no such file (even if there is such a
_directory_) it will not match anything at all.
These semantics may not seem intuitive, but they are actually very
practical. In particular, it makes it very simple to do
git-ls-files fs/*.c | xargs grep some_pattern
and it does what you want.
Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
I think Linus did a cut & paste from an early JIT code while
developing the current extended SHA1 notation, and left it there as a
courtesy, but the directory does not deserve to be treated any more
specially than, say, .git/refs/bisect.
If the subdirectories under .git/refs proliferate, we may want to
switch to scanning that hierarchy at runtime, instead of the current
hard-coded set, although I think that would be overkill.
Signed-off-by: Junio C Hamano <junkio@cox.net>
From nobody Mon Sep 17 00:00:00 2001
Subject: [PATCH] Add a new extended SHA1 syntax <name>:<num>
From: Junio C Hamano <junkio@cox.net>
Date: 1124617434 -0700
The new notation is a short-hand for <name> followed by <num>
caret ('^') characters. E.g. "master:4" is the fourth
generation ancestor of the current "master" branch head,
following the first parents; same as "master^^^^" but a bit more
readable.
This will be used in the updated "git show-branch" command.
Signed-off-by: Junio C Hamano <junkio@cox.net>
---
+static int get_nth_ancestor(const char *name, int len,
+ unsigned char *result, int generation)
+{
+ unsigned char sha1[20];
+ int ret = get_sha1_1(name, len, sha1);
+ if (ret)
+ return ret;
+
+ while (generation--) {
+ struct commit *commit = lookup_commit_reference(sha1);
+
+ if (!commit || parse_commit(commit) || !commit->parents)
+ return -1;
+ memcpy(sha1, commit->parents->item->object.sha1, 20);
+ }
+ memcpy(result, sha1, 20);
+ return 0;
+}
+
static int get_sha1_1(const char *name, int len, unsigned char *sha1)
{
int parent, ret;
+ const char *cp;
/* foo^[0-9] or foo^ (== foo^1); we do not do more than 9 parents. */
if (len > 2 && name[len-2] == '^' &&
@@ -210,6 +230,27 @@ static int get_sha1_1(const char *name,
if (parent >= 0)
return get_parent(name, len, sha1, parent);
+ /* name:3 is name^^^,
+ * name:12 is name^^^^^^^^^^^^, and
+ * name: is name
+ */
+ parent = 0;
+ for (cp = name + len - 1; name <= cp; cp--) {
+ int ch = *cp;
+ if ('0' <= ch && ch <= '9')
+ continue;
+ if (ch != ':')
+ parent = -1;
+ break;
+ }
+ if (!parent && *cp == ':') {
+ int len1 = cp - name;
+ cp++;
+ while (cp < name + len)
+ parent = parent * 10 + *cp++ - '0';
+ return get_nth_ancestor(name, len1, sha1, parent);
+ }
+
ret = get_sha1_basic(name, len, sha1);
if (!ret)
return 0;
[PATCH] Fix git-commit-script to output on stderr when -v fails
When git-commit-script is called with -v option and
verify test fails result is print on stdout
instead of stderr.
[jc: The original patch from Marco updated git-commit-script that
still had the piece of code in question, which has been moved to
an example hook script on its own, so I transplanted the patch to
that new file instead.]
Signed-off-by: Marco Costalba <mcostalba@yahoo.it> Signed-off-by: Junio C Hamano <junkio@cox.net>
[PATCH] Allow file removal when "git commit --all" is used.
After you deleted files from your working tree, automatic
git-update-cache used when the "--all" flag is given to "git
commit" barfs because it lacks the --remove flag.
It can be argued that this is a feature; people should be
careful and something with a grave consequence like removing
files should be done manually, in which case the current
behaviour may be OK.
The patch is for people who thinks the user who uses the "--all"
flag deserves the danger that comes with the convenience.
[PATCH] Fix git-format-patch-script to handle empty messages
In case of a commit with an empty message there is no
mandatory empty line between headers and body
[jc: This makes --mbox output valid even when the commit message does
not have anything but its first line, which the one I wrote botched.
One side-effect is that it adds an extra blank line at the end even if
it has more than one lines, which will be eaten by the receiving end.
As Marco says, this is a stop-gap measure. This script needs to be
split into two, one that gets the format specifier and a commit ID to
write to its standard output, and another that drives that one reading
from rev-list. I'll fix things properly when that happens by
rewriting the former part in Perl or something more reasonable than
the current shell, sed and grep mishmash.]
Signed-off-by: Marco Costalba <mcostalba@yahoo.it> Signed-off-by: Junio C Hamano <junkio@cox.net>
Display the contents of a tag when the user clicks on it.
This just displays the result of git-cat-file on the tag in the
details pane. If the tag is a "direct" tag (the tag file contains
the SHA1 ID of a commit rather than a tag), we show the tag name
and SHA1 ID.
These are features requested by Junio. Any plain file under .git/refs
whose contents start with 40 hex characters is taken as a reference
and displayed like a head but with a light blue background (unless it
is in .git/refs/tags or .git/refs/heads, in which case it is displayed
as before). There is now a "Reread references" menu item in the File
menu which re-reads all the plain files under .git/refs and redisplays
any references that have changed.
This makes git-applymbox verify that the index matches the current HEAD
before it starts applying patches.
Otherwise, you might have updated the index with unrelated changes, and
the first patch will commit not just the patch from the mbox, but also any
changes you had in your index.
Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
With --parents, git-rev-list gives us the list of parents on the
first line of each commit. We use that rather than looking for
the parent: lines in the commit body, since this way we get to
know about the grafts for free.
Based on the discussion on the git list, here are some important changes
to the glossary. (There is no cache, but an index. Use "object name"
rather than "SHA1". Reorder. Clarify.)
Signed-off-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Junio C Hamano <junkio@cox.net>
When the graph gets too wide (as defined by the maxwidth variable,
which can be set in ~/.gitk), we can now terminate graph lines with
an arrow pointing downwards, and reintroduce them later with an
arrow pointing upwards when we need them. This makes the graph much
less cluttered on large repositories such as the linux kernel.
Unfortunately this has made it slower; it takes about 10 seconds
user time on the linux-2.6 repository on my machine now, compared
to 6 seconds before. I'll have to work on optimizing that. Also
on the todo list are making the arrow heads active (so if you click
on them you jump to the other end) and improving the placement of
the null entry.
It did not check to see if the working tree was clean and matched
the commit we were starting out as, resulting in the initial rebased
commit including whatever dirty state the working tree has had.
Make sure alternates are carried over from the original repository.
When we create a cheap local clone by pointing at the object databse
of the original repository, we forgot to take the alternates the original
repository might have had into account.
The earlier one to grab output from diff-files --name-only has a grave
bug that when no paths are given it ended up doing the equivalent of
"git-commit --all", which was not what I intended.