Code

git.git
13 years agoMerge branch 'maint'
Junio C Hamano [Sun, 3 Apr 2011 20:53:11 +0000 (13:53 -0700)]
Merge branch 'maint'

* maint:
  Start preparing for 1.7.4.4
  pull: do not clobber untracked files on initial pull
  compat: add missing #include <sys/resource.h>

Conflicts:
RelNotes

13 years agoStart preparing for 1.7.4.4
Junio C Hamano [Sun, 3 Apr 2011 19:58:23 +0000 (12:58 -0700)]
Start preparing for 1.7.4.4

Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agopull: do not clobber untracked files on initial pull
Jeff King [Fri, 25 Mar 2011 18:13:31 +0000 (14:13 -0400)]
pull: do not clobber untracked files on initial pull

For a pull into an unborn branch, we do not use "git merge"
at all. Instead, we call read-tree directly. However, we
used the --reset parameter instead of "-m", which turns off
the safety features.

Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoMerge branch 'jc/index-update-if-able' into maint
Junio C Hamano [Sun, 3 Apr 2011 19:33:05 +0000 (12:33 -0700)]
Merge branch 'jc/index-update-if-able' into maint

* jc/index-update-if-able:
  update $GIT_INDEX_FILE when there are racily clean entries
  diff/status: refactor opportunistic index update

13 years agoMerge branch 'lt/default-abbrev' into maint
Junio C Hamano [Sun, 3 Apr 2011 19:32:51 +0000 (12:32 -0700)]
Merge branch 'lt/default-abbrev' into maint

* lt/default-abbrev:
  Rename core.abbrevlength back to core.abbrev
  Make the default abbrev length configurable

13 years agoMerge branch 'jc/maint-rev-list-culled-boundary' into maint
Junio C Hamano [Sun, 3 Apr 2011 19:32:35 +0000 (12:32 -0700)]
Merge branch 'jc/maint-rev-list-culled-boundary' into maint

* jc/maint-rev-list-culled-boundary:
  list-objects.c: don't add an unparsed NULL as a pending tree

Conflicts:
list-objects.c

13 years agoMerge branch 'mm/maint-log-n-with-diff-filtering' into maint
Junio C Hamano [Sun, 3 Apr 2011 19:31:59 +0000 (12:31 -0700)]
Merge branch 'mm/maint-log-n-with-diff-filtering' into maint

* mm/maint-log-n-with-diff-filtering:
  log: fix --max-count when used together with -S or -G

13 years agoMerge branch 'jk/format-patch-multiline-header' into maint
Junio C Hamano [Sun, 3 Apr 2011 19:31:19 +0000 (12:31 -0700)]
Merge branch 'jk/format-patch-multiline-header' into maint

* jk/format-patch-multiline-header:
  format-patch: rfc2047-encode newlines in headers
  format-patch: wrap long header lines
  strbuf: add fixed-length version of add_wrapped_text

13 years agoMerge branch 'jn/maint-instaweb-plack-fix' into maint
Junio C Hamano [Sun, 3 Apr 2011 19:30:58 +0000 (12:30 -0700)]
Merge branch 'jn/maint-instaweb-plack-fix' into maint

* jn/maint-instaweb-plack-fix:
  git-instaweb: Change how gitweb.psgi is made runnable as standalone app

13 years agoMerge branch 'lp/config-vername-check' into maint
Junio C Hamano [Sun, 3 Apr 2011 19:29:45 +0000 (12:29 -0700)]
Merge branch 'lp/config-vername-check' into maint

* lp/config-vername-check:
  Disallow empty section and variable names
  Sanity-check config variable names

13 years agocompat: add missing #include <sys/resource.h>
Jonathan Nieder [Fri, 18 Mar 2011 20:23:52 +0000 (15:23 -0500)]
compat: add missing #include <sys/resource.h>

Starting with commit c793430 (Limit file descriptors used by packs,
2011-02-28), git uses getrlimit to tell how many file descriptors it
can use.  Unfortunately it does not include the header declaring that
function, resulting in compilation errors:

 sha1_file.c: In function 'open_packed_git_1':
 sha1_file.c:718: error: storage size of 'lim' isn't known
 sha1_file.c:721: warning: implicit declaration of function 'getrlimit'
 sha1_file.c:721: error: 'RLIMIT_NOFILE' undeclared (first use in this function)
 sha1_file.c:718: warning: unused variable 'lim'

The standard header to include for this is <sys/resource.h> (which on
some systems itself requires declarations from <sys/types.h> or
<sys/time.h>).  Probably the problem was missed until now because in
current glibc sys/resource.h happens to be included by sys/wait.h.

MinGW does not provide sys/resource.h (and compat/mingw takes care of
providing getrlimit some other way), so add the missing #include to
the "#ifndef __MINGW32__" block in git-compat-util.h.

Reported-by: Stefan Sperling <stsp@stsp.name>
Tested-by: Stefan Sperling <stsp@stsp.name> [on OpenBSD]
Tested-by: Arnaud Lacombe <lacombar@gmail.com> [on FreeBSD 8]
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agogit.el: Don't use font-lock-compile-keywords
Lawrence Mitchell [Fri, 4 Feb 2011 10:59:18 +0000 (10:59 +0000)]
git.el: Don't use font-lock-compile-keywords

If font-lock is disabled, font-lock-compile-keywords complains.
Really what we want to do is to replace log-edit's font-lock
definitions with our own, so define a major mode deriving from
log-edit and set up font-lock-defaults there.  We then use the
optional MODE argument to log-edit to set up the major mode of the
commit buffer appropriately.

Signed-off-by: Lawrence Mitchell <wence@gmx.li>
Acked-by: Alexandre Julliard <julliard@winehq.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agot2019-checkout-ambiguous-ref.sh: depend on C_LOCALE_OUTPUT
Ævar Arnfjörð Bjarmason [Sun, 3 Apr 2011 13:33:39 +0000 (13:33 +0000)]
t2019-checkout-ambiguous-ref.sh: depend on C_LOCALE_OUTPUT

The t2019-checkout-ambiguous-ref.sh tests added in v1.7.4.3~12^2
examines the output for a translatable string, and must be marked
with C_LOCALE_OUTPUT; otherwise, GETTEXT_POISON=YesPlease tests
will break.

Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoFix two unused variable warnings in gcc 4.6
Dan McGee [Mon, 28 Mar 2011 01:32:19 +0000 (20:32 -0500)]
Fix two unused variable warnings in gcc 4.6

Seen with -Wunused-but-set-variable.

Signed-off-by: Dan McGee <dpmcgee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoRemove old binaries from .gitignore
Dan McGee [Thu, 31 Mar 2011 02:26:01 +0000 (21:26 -0500)]
Remove old binaries from .gitignore

These two programs were dumped a while ago.

Signed-off-by: Dan McGee <dpmcgee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agosparse: Fix errors and silence warnings
Stephen Boyd [Sun, 3 Apr 2011 07:06:54 +0000 (00:06 -0700)]
sparse: Fix errors and silence warnings

 * load_file() returns a void pointer but is using 0 for the return
   value

 * builtin/receive-pack.c forgot to include builtin.h

 * packet_trace_prefix can be marked static

 * ll_merge takes a pointer for its last argument, not an int

 * crc32 expects a pointer as the second argument but Z_NULL is defined
   to be 0 (see 38f4d13 sparse fix: Using plain integer as NULL pointer,
   2006-11-18 for more info)

Signed-off-by: Stephen Boyd <bebarino@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoUpdate release notes
Junio C Hamano [Sun, 3 Apr 2011 07:18:50 +0000 (00:18 -0700)]
Update release notes

As 1.7.4.3 has backmerged a handful of fixes from the master,
drop these entries from 1.7.5 release notes.

Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoSync with 1.7.4.3
Junio C Hamano [Sun, 3 Apr 2011 07:14:16 +0000 (00:14 -0700)]
Sync with 1.7.4.3

13 years agoGit 1.7.4.3 v1.7.4.3
Junio C Hamano [Sun, 3 Apr 2011 07:12:31 +0000 (00:12 -0700)]
Git 1.7.4.3

Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoDoc: mention --delta-base-offset is the default for Porcelain commands
Junio C Hamano [Sun, 3 Apr 2011 06:08:13 +0000 (23:08 -0700)]
Doc: mention --delta-base-offset is the default for Porcelain commands

The underlying pack-objects plumbing command still needs an explicit
option from the command line, but these days Porcelain passes the
option, so there is no need for end users to worry about it anymore.

Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoMerge branch 'nd/init-gitdir'
Junio C Hamano [Sat, 2 Apr 2011 00:57:37 +0000 (17:57 -0700)]
Merge branch 'nd/init-gitdir'

* nd/init-gitdir:
  init, clone: support --separate-git-dir for .git file
  git-init.txt: move description section up

Conflicts:
builtin/clone.c

13 years agoMerge branch 'jc/merge-sans-branch'
Junio C Hamano [Sat, 2 Apr 2011 00:57:16 +0000 (17:57 -0700)]
Merge branch 'jc/merge-sans-branch'

* jc/merge-sans-branch:
  merge: merge with the default upstream branch without argument
  merge: match the help text with the documentation

Conflicts:
builtin/merge.c

13 years agoMerge branch 'jr/grep-en-config'
Junio C Hamano [Sat, 2 Apr 2011 00:56:27 +0000 (17:56 -0700)]
Merge branch 'jr/grep-en-config'

* jr/grep-en-config:
  grep: allow -E and -n to be turned on by default via configuration

13 years agoMerge branch 'ab/i18n-st'
Junio C Hamano [Sat, 2 Apr 2011 00:55:55 +0000 (17:55 -0700)]
Merge branch 'ab/i18n-st'

* ab/i18n-st: (69 commits)
  i18n: git-shortlog basic messages
  i18n: git-revert split up "could not revert/apply" message
  i18n: git-revert literal "me" messages
  i18n: git-revert "Your local changes" message
  i18n: git-revert basic messages
  i18n: git-notes GIT_NOTES_REWRITE_MODE error message
  i18n: git-notes basic commands
  i18n: git-gc "Auto packing the repository" message
  i18n: git-gc basic messages
  i18n: git-describe basic messages
  i18n: git-clean clean.requireForce messages
  i18n: git-clean basic messages
  i18n: git-bundle basic messages
  i18n: git-archive basic messages
  i18n: git-status "renamed: " message
  i18n: git-status "Initial commit" message
  i18n: git-status "Changes to be committed" message
  i18n: git-status shortstatus messages
  i18n: git-status "nothing to commit" messages
  i18n: git-status basic messages
  ...

Conflicts:
builtin/branch.c
builtin/checkout.c
builtin/clone.c
builtin/commit.c
builtin/grep.c
builtin/merge.c
builtin/push.c
builtin/revert.c
t/t3507-cherry-pick-conflict.sh
t/t7607-merge-overwrite.sh

13 years agoMerge branch 'jk/pull-into-empty'
Junio C Hamano [Sat, 2 Apr 2011 00:50:54 +0000 (17:50 -0700)]
Merge branch 'jk/pull-into-empty'

* jk/pull-into-empty:
  pull: do not clobber untracked files on initial pull
  merge: merge unborn index before setting ref

13 years agoMerge branch 'sb/sparse-more'
Junio C Hamano [Sat, 2 Apr 2011 00:50:20 +0000 (17:50 -0700)]
Merge branch 'sb/sparse-more'

* sb/sparse-more:
  Makefile: Cover more files with make check

13 years agoMerge branch 'maint'
Junio C Hamano [Sat, 2 Apr 2011 00:45:59 +0000 (17:45 -0700)]
Merge branch 'maint'

* maint:
  docs: fix filter-branch subdir example for exotic repo names

13 years agoMerge branch 'nd/index-doc' into maint
Junio C Hamano [Fri, 1 Apr 2011 23:23:55 +0000 (16:23 -0700)]
Merge branch 'nd/index-doc' into maint

* nd/index-doc:
  doc: technical details about the index file format
  doc: technical details about the index file format

13 years agoMerge branch 'pk/stash-apply-status-relative' into maint
Junio C Hamano [Fri, 1 Apr 2011 23:23:43 +0000 (16:23 -0700)]
Merge branch 'pk/stash-apply-status-relative' into maint

* pk/stash-apply-status-relative:
  Add test: git stash shows status relative to current dir
  git stash: show status relative to current directory

13 years agoMerge branch 'jc/maint-diff-q-filter' into maint
Junio C Hamano [Fri, 1 Apr 2011 23:23:34 +0000 (16:23 -0700)]
Merge branch 'jc/maint-diff-q-filter' into maint

* jc/maint-diff-q-filter:
  diff --quiet: disable optimization when --diff-filter=X is used

13 years agoMerge branch 'js/maint-stash-index-copy' into maint
Junio C Hamano [Fri, 1 Apr 2011 23:23:21 +0000 (16:23 -0700)]
Merge branch 'js/maint-stash-index-copy' into maint

* js/maint-stash-index-copy:
  stash: copy the index using --index-output instead of cp -p
  stash: fix incorrect quoting in cleanup of temporary files

13 years agoMerge branch 'mg/doc-bisect-tweak-worktree' into maint
Junio C Hamano [Fri, 1 Apr 2011 23:23:15 +0000 (16:23 -0700)]
Merge branch 'mg/doc-bisect-tweak-worktree' into maint

* mg/doc-bisect-tweak-worktree:
  git-bisect.txt: example for bisecting with hot-fix
  git-bisect.txt: streamline run presentation

13 years agoMerge branch 'jh/maint-do-not-track-non-branches' into maint
Junio C Hamano [Fri, 1 Apr 2011 23:20:45 +0000 (16:20 -0700)]
Merge branch 'jh/maint-do-not-track-non-branches' into maint

* jh/maint-do-not-track-non-branches:
  branch/checkout --track: Ensure that upstream branch is indeed a branch

13 years agoMerge branch 'fk/maint-cvsimport-early-failure' into maint
Junio C Hamano [Fri, 1 Apr 2011 23:20:24 +0000 (16:20 -0700)]
Merge branch 'fk/maint-cvsimport-early-failure' into maint

* fk/maint-cvsimport-early-failure:
  git-cvsimport.perl: Bail out right away when reading from the server fails

13 years agoMerge branch 'jc/maint-apply-report-offset' into maint
Junio C Hamano [Fri, 1 Apr 2011 23:20:19 +0000 (16:20 -0700)]
Merge branch 'jc/maint-apply-report-offset' into maint

* jc/maint-apply-report-offset:
  apply -v: show offset count when patch did not apply exactly

13 years agoMerge branch 'jc/maint-apply-no-double-patch' into maint
Junio C Hamano [Fri, 1 Apr 2011 23:20:11 +0000 (16:20 -0700)]
Merge branch 'jc/maint-apply-no-double-patch' into maint

* jc/maint-apply-no-double-patch:
  apply: do not patch lines that were already patched

13 years agoMerge branch 'js/checkout-untracked-symlink' into maint
Junio C Hamano [Fri, 1 Apr 2011 23:19:03 +0000 (16:19 -0700)]
Merge branch 'js/checkout-untracked-symlink' into maint

* js/checkout-untracked-symlink:
  do not overwrite untracked symlinks
  Demonstrate breakage: checkout overwrites untracked symlink with directory

13 years agoMerge "checkout ambiguous ref bugfix" into maint
Junio C Hamano [Fri, 1 Apr 2011 23:16:23 +0000 (16:16 -0700)]
Merge "checkout ambiguous ref bugfix" into maint

* commit '0cb6ad3':
  checkout: fix bug with ambiguous refs

13 years agodocs: fix filter-branch subdir example for exotic repo names
Jeff King [Fri, 1 Apr 2011 14:46:27 +0000 (10:46 -0400)]
docs: fix filter-branch subdir example for exotic repo names

The GIT_INDEX_FILE variable we get from git has the full
path to the repo, which may contain spaces. When we use it
in our shell snippet, it needs to be quoted.

Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoGit 1.7.5-rc0 v1.7.5-rc0
Junio C Hamano [Thu, 31 Mar 2011 20:58:16 +0000 (13:58 -0700)]
Git 1.7.5-rc0

Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoMerge branch 'maint'
Junio C Hamano [Thu, 31 Mar 2011 02:33:53 +0000 (19:33 -0700)]
Merge branch 'maint'

* maint:
  parse-remote: typofix

13 years agoparse-remote: typofix
Junio C Hamano [Thu, 31 Mar 2011 02:33:33 +0000 (19:33 -0700)]
parse-remote: typofix

An earlier patch had a trivial typo that two people did not notice.
Pointed out by Michael Schubert.

Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoMerge branch 'maint'
Junio C Hamano [Wed, 30 Mar 2011 21:10:41 +0000 (14:10 -0700)]
Merge branch 'maint'

* maint:
  contrib/thunderbird-patch-inline: do not require bash to run the script
  t8001: check the exit status of the command being tested
  strbuf.h: remove a tad stale docs-in-comment and reference api-doc instead
  Typos: t/README
  Documentation/config.txt: make truth value of numbers more explicit
  git-pack-objects.txt: fix grammatical errors
  parse-remote: replace unnecessary sed invocation

13 years agogrep: allow -E and -n to be turned on by default via configuration
Joe Ratterman [Wed, 30 Mar 2011 19:31:05 +0000 (14:31 -0500)]
grep: allow -E and -n to be turned on by default via configuration

Add two configration variables grep.extendedRegexp and grep.lineNumbers to
allow the user to skip typing -E and -n on the command line, respectively.

Scripts that are meant to be used by random users and/or in random
repositories now have use -G and/or --no-line-number options as
appropriately to override the settings in the repository or user's
~/.gitconfig settings. Just because the script didn't say "git grep -n" no
longer guarantees that the output from the command will not have line
numbers.

Signed-off-by: Joe Ratterman <jratt0@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agocontrib/thunderbird-patch-inline: do not require bash to run the script
Maxin john [Wed, 30 Mar 2011 08:52:23 +0000 (09:52 +0100)]
contrib/thunderbird-patch-inline: do not require bash to run the script

The script does not have to be run under bash, but any POSIX compliant
shell would do, as it does not use any bash-isms.

It may be written under a different style than what is recommended in
Documentation/CodingGuidelines, but that is a different matter.

While at it, fix obvious typos in the comment.

Signed-off-by: Maxin B. John <maxin@maxinbjohn.info>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agot8001: check the exit status of the command being tested
Junio C Hamano [Wed, 30 Mar 2011 19:22:05 +0000 (12:22 -0700)]
t8001: check the exit status of the command being tested

Avoid running the command being tested as an upstream of a pipe;
doing so will lose its exit status.

While at it, modernise the style of the script.

Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agostrbuf.h: remove a tad stale docs-in-comment and reference api-doc instead
Michael Witten [Tue, 15 Feb 2011 23:12:04 +0000 (23:12 +0000)]
strbuf.h: remove a tad stale docs-in-comment and reference api-doc instead

Signed-off-by: Michael Witten <mfwitten@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoTypos: t/README
Michael Witten [Tue, 22 Feb 2011 17:15:00 +0000 (17:15 +0000)]
Typos: t/README

Signed-off-by: Michael Witten <mfwitten@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoDocumentation/config.txt: make truth value of numbers more explicit
Carlos Martín Nieto [Wed, 30 Mar 2011 10:22:32 +0000 (12:22 +0200)]
Documentation/config.txt: make truth value of numbers more explicit

Change the order to 1/0 to have the same true/false order as the rest
of the possibilities for a boolean variable in order not not confuse
users.

Signed-off-by: Carlos Martín Nieto <cmn@elego.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agogit-pack-objects.txt: fix grammatical errors
Stephen Boyd [Wed, 30 Mar 2011 09:00:06 +0000 (02:00 -0700)]
git-pack-objects.txt: fix grammatical errors

Signed-off-by: Stephen Boyd <bebarino@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoparse-remote: replace unnecessary sed invocation
Stephen Boyd [Wed, 30 Mar 2011 08:48:40 +0000 (01:48 -0700)]
parse-remote: replace unnecessary sed invocation

Just use parameter expansion instead.

Signed-off-by: Stephen Boyd <bebarino@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoMerge branch 'svn-fe' of git://repo.or.cz/git/jrn
Junio C Hamano [Wed, 30 Mar 2011 17:49:13 +0000 (10:49 -0700)]
Merge branch 'svn-fe' of git://repo.or.cz/git/jrn

* 'svn-fe' of git://repo.or.cz/git/jrn:
  tests: kill backgrounded processes more robustly
  vcs-svn: a void function shouldn't try to return something
  tests: make sure input to sed is newline terminated
  vcs-svn: add missing cast to printf argument

13 years agotests: kill backgrounded processes more robustly
Jeff King [Wed, 30 Mar 2011 03:30:17 +0000 (23:30 -0400)]
tests: kill backgrounded processes more robustly

t0081 creates several background processes that write to a fifo and
then go to sleep for a while (so the reader of the fifo does not see
EOF).

Each background process is made in a curly-braced block in the shell,
and after we are done reading from the fifo, we use "kill $!" to kill
it off.

For a simple, single-command process, this works reliably and kills
the child sleep process. But for more complex commands like
"make_some_output && sleep", the results are less predictable. When
executing under bash, we end up with a subshell that gets killed by
the $! but leaves the sleep process still alive.

This is bad not only for process hygeine (we are leaving random sleep
processes to expire after a while), but also interacts badly with the
"prove" command. When prove executes a test, it does not realize the
test is done when it sees SIGCHLD, but rather waits until the test's
stdout pipe is closed. The orphaned sleep process may keep that pipe
open via test-lib's file descriptor 5, causing prove to hang for 100
seconds.

The solution is to explicitly use a subshell and to exec the final
sleep process, so that when we "kill $!" we get the process id of the
sleep process.

[jn: original patch by Jeff had some additional bits:

   1. Wrap the "kill" in a test_when_finished, since we want
      to clean up the process whether the test succeeds or not.

   2. The "kill" is part of our && chain for test success. It
      probably won't fail, but it can if the process has
      expired before we manage to kill it. So let's mark it
      as OK to fail.

 I'm postponing that for now.]

Reported-by: Junio C Hamano <gitster@pobox.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
13 years agoMerge branch 'maint'
Junio C Hamano [Tue, 29 Mar 2011 21:09:15 +0000 (14:09 -0700)]
Merge branch 'maint'

* maint:
  HOME must be set before calling git-init when creating test repositories

13 years agoMerge branch 'jc/fetch-progressive-stride'
Junio C Hamano [Tue, 29 Mar 2011 21:09:08 +0000 (14:09 -0700)]
Merge branch 'jc/fetch-progressive-stride'

* jc/fetch-progressive-stride:
  Fix potential local deadlock during fetch-pack

13 years agoMerge branches 'sp/maint-fetch-pack-stop-early' and 'sp/maint-upload-pack-stop-early'
Junio C Hamano [Tue, 29 Mar 2011 21:09:02 +0000 (14:09 -0700)]
Merge branches 'sp/maint-fetch-pack-stop-early' and 'sp/maint-upload-pack-stop-early'

* sp/maint-fetch-pack-stop-early:
  enable "no-done" extension only when fetching over smart-http

* sp/maint-upload-pack-stop-early:
  enable "no-done" extension only when serving over smart-http

13 years agovcs-svn: a void function shouldn't try to return something
Michael Witten [Tue, 29 Mar 2011 17:31:30 +0000 (17:31 +0000)]
vcs-svn: a void function shouldn't try to return something

As v1.7.4-rc0~184 (2010-10-04) and C99 §6.8.6.4.1 remind us, standard
C does not permit returning an expression of type void, even for a
tail call.

Noticed with gcc -pedantic:

 vcs-svn/svndump.c: In function 'handle_node':
 vcs-svn/svndump.c:213:3: warning: ISO C forbids 'return' with expression,
  in function returning void [-pedantic]

[jn: with simplified log message]

Signed-off-by: Michael Witten <mfwitten@gmail.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
13 years agoRevert two "no-done" reverts
Junio C Hamano [Tue, 29 Mar 2011 19:29:10 +0000 (12:29 -0700)]
Revert two "no-done" reverts

Last night I had to make these two emergency reverts, but now we have a
better understanding of which part of the topic was broken, let's get rid
of the revert to fix it correctly.

Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoenable "no-done" extension only when serving over smart-http
Junio C Hamano [Tue, 29 Mar 2011 17:24:59 +0000 (10:24 -0700)]
enable "no-done" extension only when serving over smart-http

Do not advertise no-done capability when upload-pack is not serving over
smart-http, as there is no way for this server to know when it should stop
reading in-flight data from the client, even though it is necessary to
drain all the in-flight data in order to unblock the client.

Signed-off-by: Junio C Hamano <gitster@pobox.com>
Acked-by: Shawn O. Pearce <spearce@spearce.org>
13 years agoFix potential local deadlock during fetch-pack
Junio C Hamano [Tue, 29 Mar 2011 17:06:19 +0000 (10:06 -0700)]
Fix potential local deadlock during fetch-pack

The fetch-pack/upload-pack protocol relies on the underlying transport
(local pipe or TCP socket) to have enough slack to allow one window worth
of data in flight without blocking the writer.  Traditionally we always
relied on being able to have two windows of 32 "have"s in flight (roughly
3k bytes) to stream.

The recent "progressive-stride" change allows "fetch-pack" to send up to
1024 "have"s without reading any response from "upload-pack".  The
outgoing pipe of "upload-pack" can be clogged with many ACK and NAK that
are unread, while "fetch-pack" is still stuffing its outgoing pipe with
more "have"s, leading to a deadlock.

Revert the change unless we are in stateless rpc (aka smart-http) mode, as
using a large window full of "have"s is still a good way to help reduce
the number of back-and-forth, and there is no buffering issue there (it is
strictly "ping-pong" without an overlap).

Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoenable "no-done" extension only when fetching over smart-http
Junio C Hamano [Tue, 29 Mar 2011 17:16:29 +0000 (10:16 -0700)]
enable "no-done" extension only when fetching over smart-http

When 'no-done' protocol extension is used, the upload-pack (i.e. the
server side) process stops listening to the fetch-pack after issuing the
final NAK, and starts sending the generated pack data back, but there may
be more "have" send by the latter in flight that the fetch-pack is
expecting to be responded with ACK/NAK.  This will typically result in a
deadlock (both will block on write that the other end never reads) or
SIGPIPE on the fetch-pack end (upload-pack will finish writing a small
pack and goes away).

Disable it unless fetch-pack is running under smart-http, where there is
no such streaming issue.

Signed-off-by: Junio C Hamano <gitster@pobox.com>
Acked-by: Shawn O. Pearce <spearce@spearce.org>
13 years agoHOME must be set before calling git-init when creating test repositories
Alex Riesen [Sat, 26 Mar 2011 18:46:34 +0000 (19:46 +0100)]
HOME must be set before calling git-init when creating test repositories

Otherwise the created test repositories will be affected by users ~/.gitconfig.
For example, setting core.logAllrefupdates in users config will make all
calls to "git config --unset core.logAllrefupdates" fail which will break
the first test which uses the statement and expects it to succeed.

Signed-off-by: Alex Riesen <raa.lkml@gmail.com>
Acked-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agotests: make sure input to sed is newline terminated
Jonathan Nieder [Tue, 29 Mar 2011 03:08:20 +0000 (22:08 -0500)]
tests: make sure input to sed is newline terminated

POSIX only requires sed to work on text files and because it does
not end with a newline, this commit's content is not a text file.
Add a newline to fix it.  Without this change, OS X sed helpfully
adds a newline to actual.message, causing t9010.13 to fail.

Reported-by: Torsten Bögershausen <tboegi@web.de>
Tested-by: Brian Gernhardt <benji@silverinsanity.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
13 years agoRevert "fetch-pack: Implement no-done capability"
Junio C Hamano [Tue, 29 Mar 2011 06:35:39 +0000 (23:35 -0700)]
Revert "fetch-pack: Implement no-done capability"

This reverts commit 761ecf0bc7b6cddf311f00877c59e6381cdbdeea.

13 years agoRevert "upload-pack: Implement no-done capability"
Junio C Hamano [Tue, 29 Mar 2011 06:33:51 +0000 (23:33 -0700)]
Revert "upload-pack: Implement no-done capability"

This reverts 3e63b21 (upload-pack: Implement no-done capability,
2011-03-14).  Together with 761ecf0 (fetch-pack: Implement no-done
capability, 2011-03-14) it seems to make the fetch-pack process out of
sync and makes it keep talking long after upload-pack stopped listening to
it, terminating the process with SIGPIPE.

13 years agoMerge branch 'maint'
Junio C Hamano [Mon, 28 Mar 2011 21:17:17 +0000 (14:17 -0700)]
Merge branch 'maint'

* maint:
  git tag documentation grammar fixes and readability updates
  grep: Add the option '--line-number'

13 years agogit tag documentation grammar fixes and readability updates
Michael Witten [Mon, 28 Mar 2011 15:11:49 +0000 (15:11 +0000)]
git tag documentation grammar fixes and readability updates

... with help from Eric Raible.

In addition, describe the use of GIT_COMMITTER_DATE more comprehensively
by including "date-formats.txt"

Signed-off-by: Michael Witten <mfwitten@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agotests: fix overeager scrubbing of environment variables
Jens Lehmann [Mon, 28 Mar 2011 19:16:09 +0000 (21:16 +0200)]
tests: fix overeager scrubbing of environment variables

In commit 95a1d12e9b9f ("tests: scrub environment of GIT_* variables") all
environment variables starting with "GIT_" were unset for the tests using
a perl script rather than unsetting them one by one. Only three exceptions
were made to make them work as before: "GIT_TRACE*", "GIT_DEBUG*" and
"GIT_USE_LOOKUP".

Unfortunately some environment variables used by the test framework itself
were not added to the exceptions and thus stopped working when given
before the make command instead of after it. Those are:

- GIT_NOTES_TIMING_TESTS
- GIT_PATCHID_TIMING_TESTS
- GIT_PROVE_OPTS
- GIT_REMOTE_SVN_TEST_BIG_FILES
- GIT_SKIP_TESTS
- GIT_TEST*
- GIT_VALGRIND_OPTIONS

I noticed that when skipping a test the way I was used to suddenly failed:

GIT_SKIP_TESTS='t1234' GIT_TEST_OPTS='--root=/dev/shm' make -j10 test

This should work according to t/README, but didn't anymore, so let's fix
that by adding them to the exception list. And to avoid having a long
regexp put the exceptions in a separate variable using nicer formatting.

Signed-off-by: Jens Lehmann <Jens.Lehmann@web.de>
Acked-by: Jonathan Nieder <jrnieder@gmail.com>
Thanks-to: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agogrep: Add the option '--line-number'
Joe Ratterman [Mon, 28 Mar 2011 18:11:55 +0000 (13:11 -0500)]
grep: Add the option '--line-number'

This is a synonym for the existing '-n' option, matching GNU grep.

Signed-off-by: Joe Ratterman <jratt0@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoImprove test for pthreads flag
Giuseppe Bilotta [Mon, 28 Mar 2011 07:39:59 +0000 (09:39 +0200)]
Improve test for pthreads flag

When compiling with CC=clang using Clang 1.1 as shipped by Debian
unstable (package version 2.7-3), the -mt flag is sufficient to compile
during the `configure` test. However, building git would then fail at
link time complaining about missing symbols such as `pthread_key_create'
and `pthread_create'.

Work around this issue by adding pthread key creation to the pthreads
configure test source.

Signed-off-by: Giuseppe Bilotta <giuseppe.bilotta@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agovcs-svn: add missing cast to printf argument
Jonathan Nieder [Mon, 28 Mar 2011 00:38:15 +0000 (19:38 -0500)]
vcs-svn: add missing cast to printf argument

gcc -m32 correctly warns:

 vcs-svn/fast_export.c: In function 'fast_export_commit':
 vcs-svn/fast_export.c:54:2: warning: format '%llu' expects
   argument of type 'long long unsigned int', but argument 2
   has type 'unsigned int' [-Wformat]

Fix it.

Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agovcs-svn: add missing cast to printf argument
Jonathan Nieder [Sun, 27 Mar 2011 17:19:14 +0000 (12:19 -0500)]
vcs-svn: add missing cast to printf argument

gcc -m32 correctly warns:

 vcs-svn/fast_export.c: In function 'fast_export_commit':
 vcs-svn/fast_export.c:54:2: warning: format '%llu' expects
   argument of type 'long long unsigned int', but argument 2
   has type 'unsigned int' [-Wformat]

Fix it.

Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
13 years agoUpdate draft release note to 1.7.5
Junio C Hamano [Sun, 27 Mar 2011 03:17:15 +0000 (20:17 -0700)]
Update draft release note to 1.7.5

Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoMerge branch 'mg/rev-list-n-parents'
Junio C Hamano [Sun, 27 Mar 2011 03:13:17 +0000 (20:13 -0700)]
Merge branch 'mg/rev-list-n-parents'

* mg/rev-list-n-parents:
  tests: avoid nonportable {foo,bar} glob
  rev-list --min-parents,--max-parents: doc, test and completion
  revision.c: introduce --min-parents and --max-parents options
  t6009: use test_commit() from test-lib.sh

13 years agoMerge branch 'js/remove-unused-variables'
Junio C Hamano [Sun, 27 Mar 2011 03:13:17 +0000 (20:13 -0700)]
Merge branch 'js/remove-unused-variables'

* js/remove-unused-variables:
  Remove unused variables

13 years agoMerge branch 'jp/completion-help-alias'
Junio C Hamano [Sun, 27 Mar 2011 03:13:17 +0000 (20:13 -0700)]
Merge branch 'jp/completion-help-alias'

* jp/completion-help-alias:
  git-completion: Add git help completion for aliases

13 years agoMerge branch 'jc/index-update-if-able'
Junio C Hamano [Sun, 27 Mar 2011 03:13:16 +0000 (20:13 -0700)]
Merge branch 'jc/index-update-if-able'

* jc/index-update-if-able:
  update $GIT_INDEX_FILE when there are racily clean entries
  diff/status: refactor opportunistic index update

13 years agoMerge branch 'jk/checkout-orphan-warning'
Junio C Hamano [Sun, 27 Mar 2011 03:13:16 +0000 (20:13 -0700)]
Merge branch 'jk/checkout-orphan-warning'

* jk/checkout-orphan-warning:
  checkout: tweak detached-orphan warning format
  checkout: clear commit marks after detached-orphan check
  checkout: add basic tests for detached-orphan warning

13 years agoMerge branch 'jc/maint-rerere-in-workdir'
Junio C Hamano [Sun, 27 Mar 2011 03:13:16 +0000 (20:13 -0700)]
Merge branch 'jc/maint-rerere-in-workdir'

* jc/maint-rerere-in-workdir:
  rerere: make sure it works even in a workdir attached to a young repository

13 years agoMerge branch 'nd/index-doc'
Junio C Hamano [Sun, 27 Mar 2011 03:13:15 +0000 (20:13 -0700)]
Merge branch 'nd/index-doc'

* nd/index-doc:
  doc: technical details about the index file format
  doc: technical details about the index file format

13 years agoMerge branch 'jc/fetch-progressive-stride'
Junio C Hamano [Sun, 27 Mar 2011 03:13:15 +0000 (20:13 -0700)]
Merge branch 'jc/fetch-progressive-stride'

* jc/fetch-progressive-stride:
  fetch-pack: use smaller handshake window for initial request
  fetch-pack: progressively use larger handshake windows
  fetch-pack: factor out hardcoded handshake window size

Conflicts:
builtin/fetch-pack.c

13 years agoMerge branch 'svn-fe' of git://repo.or.cz/git/jrn
Junio C Hamano [Sat, 26 Mar 2011 18:35:41 +0000 (11:35 -0700)]
Merge branch 'svn-fe' of git://repo.or.cz/git/jrn

* 'svn-fe' of git://repo.or.cz/git/jrn:
  vcs-svn: handle log message with embedded NUL
  vcs-svn: avoid unnecessary copying of log message and author
  vcs-svn: remove buffer_read_string
  vcs-svn: make reading of properties binary-safe

13 years agoMerge git-gui 0.14.0
Junio C Hamano [Sat, 26 Mar 2011 17:42:26 +0000 (10:42 -0700)]
Merge git-gui 0.14.0

13 years agovcs-svn: handle log message with embedded NUL
Jonathan Nieder [Sat, 26 Mar 2011 05:49:37 +0000 (00:49 -0500)]
vcs-svn: handle log message with embedded NUL

Pass the log message by strbuf instead of as a C-style string and use
fwrite instead of printf to write it to fast-import so embedded '\0'
bytes can be preserved.

Currently "git log" doesn't show the embedded NULs but "git cat-file
commit" can.

While at it, stop including system headers from repo_tree.h.  git
source files need to include git-compat-util.h (or cache.h or
builtin.h) sooner to ensure the appropriate feature test macros are
defined.

Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
13 years agovcs-svn: avoid unnecessary copying of log message and author
Jonathan Nieder [Fri, 25 Mar 2011 04:10:00 +0000 (23:10 -0500)]
vcs-svn: avoid unnecessary copying of log message and author

Use strbuf_swap when storing the svn:log and svn:author properties, so
pointers to rather than the contents of buffers get copied.  The main
effect should be to make the code a little easier to read.

Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
13 years agovcs-svn: remove buffer_read_string
Jonathan Nieder [Fri, 25 Mar 2011 04:09:19 +0000 (23:09 -0500)]
vcs-svn: remove buffer_read_string

All previous users of buffer_read_string have already been converted
to use the more intuitive buffer_read_binary, so remove the old API to
avoid some confusion.

Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
13 years agovcs-svn: make reading of properties binary-safe
Jonathan Nieder [Sat, 26 Mar 2011 05:15:10 +0000 (00:15 -0500)]
vcs-svn: make reading of properties binary-safe

svn-fe errors out on revision 59151 of the ASF repository:

 fatal: invalid dump: unexpected end of file

The proximate cause is a property with an embedded NUL character.
Previously such anomalies were ignored but commit c9d1c8ba
(2010-12-28) introduced a check strlen(val) == len to avoid reading
uninitialized data when a property list ends early and unfortunately
this test does not distinguish between "foo" followed by EOF and the
string "foo\0bar\0baz".

Fix it by using buffer_read_binary to read to a strbuf and checking
the actual length read.  Most consumers of properties still use
C-style strings, so in practice an author or log message with embedded
NULs will be truncated, but a least this way svn-fe won't error out
(fixing the regression).

Reported-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
13 years agoSync with 1.7.4.2
Junio C Hamano [Sat, 26 Mar 2011 00:57:08 +0000 (17:57 -0700)]
Sync with 1.7.4.2

Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoGit 1.7.4.2 v1.7.4.2
Junio C Hamano [Sat, 26 Mar 2011 00:56:14 +0000 (17:56 -0700)]
Git 1.7.4.2

Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agopull: do not clobber untracked files on initial pull
Jeff King [Fri, 25 Mar 2011 18:13:31 +0000 (14:13 -0400)]
pull: do not clobber untracked files on initial pull

For a pull into an unborn branch, we do not use "git merge"
at all. Instead, we call read-tree directly. However, we
used the --reset parameter instead of "-m", which turns off
the safety features.

Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agomerge: merge unborn index before setting ref
Jeff King [Fri, 25 Mar 2011 18:10:38 +0000 (14:10 -0400)]
merge: merge unborn index before setting ref

When we merge into an unborn branch, there are basically two
steps:

  1. Write the sha1 of the new commit into the ref pointed
     to by HEAD.

  2. Update the index with the new content, and check it out
     to the working tree.

We currently do them in this order. However, (2) is the step
that is much more likely to fail, since it can be blocked by
things like untracked working tree files. When it does, the
merge fails and we are left with an empty index but an
updated HEAD.

This patch switches the order, so that a failure in updating
the index leaves us unchanged. Of course, a failure in
updating the ref now leaves us with an updated index and
mis-matched HEAD. That is arguably not much better, but it
is probably less likely to actually happen.

Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agot7607: clean up stray untracked file
Jeff King [Fri, 25 Mar 2011 18:09:03 +0000 (14:09 -0400)]
t7607: clean up stray untracked file

This file ends up conflicting with the test just after it
(causing the "git merge" to fail).  Neither test is to blame
for the bug, though. It looks like the merge in 1a9fe45
(Merge branch 'tr/merge-unborn-clobber', 2011-02-09) is what
caused the conflict.

We didn't notice because the follow-on test is already
marked as expect_failure (even though it has since been
fixed, and now succeeds once the untracked file is moved out
of the way).

Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agot7607: mark known breakage in test 11 as fixed
Jeff King [Fri, 25 Mar 2011 18:08:36 +0000 (14:08 -0400)]
t7607: mark known breakage in test 11 as fixed

This was fixed by 1d718a51 (do not overwrite untracked
symlinks, 2011-02-20).

Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agogitweb: Fix handling of fractional timezones in parse_date
Jakub Narebski [Fri, 25 Mar 2011 19:20:49 +0000 (20:20 +0100)]
gitweb: Fix handling of fractional timezones in parse_date

Fractional timezones, like -0330 (NST used in Canada) or +0430
(Afghanistan, Iran DST), were not handled properly in parse_date; this
means values such as 'minute_local' and 'iso-tz' were not generated
correctly.

This was caused by two mistakes:

* sign of timezone was applied only to hour part of offset, and not
  as it should be also to minutes part (this affected only negative
  fractional timezones).

* 'int $h + $m/60' is 'int($h + $m/60)' and not 'int($h) + $m/60',
  so fractional part was discarded altogether ($h is hours, $m is
  minutes, which is always less than 60).

Note that positive fractional timezones +0430, +0530 and +1030 can be
found as authortime in git.git repository itself.

For example http://repo.or.cz/w/git.git/commit/88d50e7 had authortime
of "Fri, 8 Jan 2010 18:48:07 +0000 (23:48 +0530)", which is not marked
with 'atnight', when "git show 88d50e7" gives correct author date of
"Sat Jan 9 00:18:07 2010 +0530".

Signed-off-by: Jakub Narebski <jnareb@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agoMerge branch 'ss/git-gui-mergetool' gitgui-0.14.0
Pat Thoyts [Fri, 25 Mar 2011 08:26:47 +0000 (08:26 +0000)]
Merge branch 'ss/git-gui-mergetool'

13 years agogit-gui: detect the use of MUI langauge packs on Windows
Pat Thoyts [Fri, 18 Feb 2011 13:42:54 +0000 (13:42 +0000)]
git-gui: detect the use of MUI langauge packs on Windows

The Tcl msgcat package doesn't detect the use of a multi-lingual language
pack on Windows 7. This means that a user may have their display language
set to Japanese but the system installed langauge was English.
This patch reads the relevent registry key to fix this before loading in
the locale specific parts of git-gui.

Signed-off-by: Pat Thoyts <patthoyts@users.sourceforge.net>
13 years agotests: avoid nonportable {foo,bar} glob
Jonathan Nieder [Thu, 24 Mar 2011 08:21:24 +0000 (03:21 -0500)]
tests: avoid nonportable {foo,bar} glob

Unlike bash and ksh, dash and busybox ash do not support brace
expansion (as in 'echo {hello,world}').  So when dash is sh,
t6009.13 (set up dodecapus) ends up pass a string beginning with
"root{1,2," to "git merge" verbatim and the test fails.

Fix it by introducing a variable to hold the list of parents for
the dodecapus and populating it in a more low-tech way.

While at it, simplify a little by combining this setup code with the
test it sets up for.

Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agomerge: merge with the default upstream branch without argument
Junio C Hamano [Thu, 24 Mar 2011 06:48:24 +0000 (23:48 -0700)]
merge: merge with the default upstream branch without argument

"git merge" without specifying any commit is a no-op by default.

A new option merge.defaultupstream can be set to true to cause such an
invocation of the command to merge the upstream branches configured for
the current branch by using their last observed values stored in their
remote tracking branches.

Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agomerge: match the help text with the documentation
Jared Hance [Thu, 10 Feb 2011 23:52:41 +0000 (18:52 -0500)]
merge: match the help text with the documentation

We used to be very casual in terminology and used <branch>, <ref> and
<rev> more or less interchangeably with <commit>.  Match the help text
given by "git merge -h" with that of the documentation.

Signed-off-by: Jared Hance <jaredhance@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
13 years agorerere: make sure it works even in a workdir attached to a young repository
Junio C Hamano [Fri, 11 Mar 2011 00:02:50 +0000 (16:02 -0800)]
rerere: make sure it works even in a workdir attached to a young repository

The git-new-workdir script in contrib/ makes a new work tree by sharing
many subdirectories of the .git directory with the original repository.
When rerere.enabled is set in the original repository, but the user has
not encountered any conflicts yet, the original repository may not yet
have .git/rr-cache directory.

When rerere wants to run in a new work tree created from such a young
original repository, it fails to mkdir(2) .git/rr-cache that is a symlink
to a yet-to-be-created directory.

There are three possible approaches to this:

 - A naive solution is not to create a symlink in the git-new-workdir
   script to a directory the original does not have (yet).  This is not a
   solution, as we tend to lazily create subdirectories of .git/, and
   having rerere.enabled configuration set is a strong indication that the
   user _wants_ to have this lazy creation to happen;

 - We could always create .git/rr-cache upon repository creation.  This is
   tempting but will not help people with existing repositories.

 - Detect this case by seeing that mkdir(2) failed with EEXIST, checking
   that the path is a symlink, and try running mkdir(2) on the link
   target.

This patch solves the issue by doing the third one.

Strictly speaking, this is incomplete.  It does not attempt to handle
relative symbolic link that points into the original repository, but this
is good enough to help people who use contrib/workdir/git-new-workdir
script.

Signed-off-by: Junio C Hamano <gitster@pobox.com>