Recent update to "git daemon" tries to enable the socket-level
KEEPALIVE, but when it is spawned via inetd, the standard input
file descriptor may not necessarily be connected to a socket.
Suppress an ENOTSOCK error from setsockopt().
* ew/daemon-socket-keepalive:
Windows: add missing definition of ENOTSOCK
daemon: ignore ENOTSOCK from setsockopt
"git pack-objects" and "git index-pack" mostly operate with off_t
when talking about the offset of objects in a packfile, but there
were a handful of places that used "unsigned long" to hold that
value, leading to an unintended truncation.
* nd/pack-ofs-4gb-limit:
fsck: use streaming interface for large blobs in pack
pack-objects: do not truncate result in-pack object size on 32-bit systems
index-pack: correct "offset" type in unpack_entry_data()
index-pack: report correct bad object offsets even if they are large
index-pack: correct "len" type in unpack_data()
sha1_file.c: use type off_t* for object_info->disk_sizep
pack-objects: pass length to check_pack_crc() without truncation
"git worktree prune" protected worktrees that are marked as
"locked" by creating a file in a known location. "git worktree"
command learned a dedicated command pair to create and remove such
a file, so that the users do not have to do this with editor.
* nd/worktree-lock:
worktree.c: find_worktree() search by path suffix
worktree: add "unlock" command
worktree: add "lock" command
worktree.c: add is_worktree_locked()
worktree.c: add is_main_worktree()
worktree.c: add find_worktree()
"git notes merge" had a code to see if a path exists (and fails if
it does) and then open the path for writing (when it doesn't).
Replace it with open with O_EXCL.
* rs/notes-merge-no-toctou:
notes-merge: use O_EXCL to avoid overwriting existing files
A few tests that specifically target "git rebase -i" have been
added.
* js/rebase-i-tests:
rebase -i: we allow extra spaces after fixup!/squash!
rebase -i: demonstrate a bug with --autosquash
t3404: add a test for the --gpg-sign option
"git merge" in v2.9 prevents merging unrelated histories.
"git subtree split --rejoin" creates unrelated histories when
creating a split repo from a raw sub-directory that did not
originate from an invocation of "git subtree add".
Restore the original behavior by passing --allow-unrelated-histories
when merging subtrees. This ensures that the synthetic history
created by "git subtree split" can be merged.
Add a test to ensure that this feature works as advertised.
Reported-by: Brett Cundal <brett.cundal@iugome.com>
Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de>
Signed-off-by: David Aguilar <davvid@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Allow whitespace in arguments to subtree_test_create_repo.
Add missing && chains.
Signed-off-by: David Aguilar <davvid@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This brings the short help in line with the documentation.
Signed-off-by: Orgad Shaneh <orgads@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Code cleanup.
* rs/use-strbuf-addbuf:
strbuf: avoid calling strbuf_grow() twice in strbuf_addbuf()
use strbuf_addbuf() for appending a strbuf to another
Existing autoconf generated test for the need to link with pthread
library did not check all the functions from pthread libraries;
recent FreeBSD has some functions in libc but not others, and we
mistakenly thought linking with libc is enough when it is not.
* ew/autoconf-pthread:
configure.ac: stronger test for pthread linkage
"git blame file" allowed the lineage of lines in the uncommitted,
unadded contents of "file" to be inspected, but it refused when
"file" did not appear in the current commit. When "file" was
created by renaming an existing file (but the change has not been
committed), this restriction was unnecessarily tight.
* mh/blame-worktree:
t/t8003-blame-corner-cases.sh: Use here documents
blame: allow to blame paths freshly added to the index
When "git fsck" reports a broken link (e.g. a tree object contains
a blob that does not exist), both containing object and the object
that is referred to were reported with their 40-hex object names.
The command learned the "--name-objects" option to show the path to
the containing object from existing refs (e.g. "HEAD~24^2:file.txt").
* js/fsck-name-object:
fsck: optionally show more helpful info for broken links
fsck: give the error function a chance to see the fsck_options
fsck_walk(): optionally name objects on the go
fsck: refactor how to describe objects
"git add -N dir/file && git write-tree" produced an incorrect tree
when there are other paths in the same directory that sorts after
"file".
* nd/cache-tree-ita:
cache-tree: do not generate empty trees as a result of all i-t-a subentries
cache-tree.c: fix i-t-a entry skipping directory updates sometimes
test-lib.sh: introduce and use $EMPTY_BLOB
test-lib.sh: introduce and use $EMPTY_TREE
"git fetch http://user:pass@host/repo..." scrubbed the userinfo
part, but "git push" didn't.
* jk/push-scrub-url:
t5541: fix url scrubbing test when GPG is not set
push: anonymize URL in status output
Build clean-up.
* nd/test-helpers:
t/test-lib.sh: fix running tests with --valgrind
Makefile: use VCSSVN_LIB to refer to svn library
Makefile: drop extra dependencies for test helpers
"git merge" with renormalization did not work well with
merge-recursive, due to "safer crlf" conversion kicking in when it
shouldn't.
* jc/renormalize-merge-kill-safer-crlf:
merge: avoid "safer crlf" during recording of merge results
convert: unify the "auto" handling of CRLF
The use of strbuf in "git rm" to build filename to remove was a bit
suboptimal, which has been fixed.
* rs/rm-strbuf-optim:
rm: reuse strbuf for all remove_dir_recursively() calls
Makefile assumed that -lrt is always available on platforms that
want to use clock_gettime() and CLOCK_MONOTONIC, which is not a
case for recent Mac OS X. The necessary symbols are often found in
libc on many modern systems and having -lrt on the command line, as
long as the library exists, had no effect, but when the platform
removes librt.a that is a different matter--having -lrt will break
the linkage.
This change could be seen as a regression for those who do need to
specify -lrt, as they now specifically ask for NEEDS_LIBRT when
building. Hopefully they are in the minority these days.
* rw/make-needs-librt:
config.mak.uname: define NEEDS_LIBRT under Linux, for now
Makefile: add NEEDS_LIBRT to optionally link with librt
An age old bug that caused "git diff --ignore-space-at-eol"
misbehave has been fixed.
* js/ignore-space-at-eol:
diff: fix a double off-by-one with --ignore-space-at-eol
diff: demonstrate a bug with --patience and --ignore-space-at-eol
The API to iterate over all the refs (i.e. for_each_ref(), etc.)
has been revamped.
* mh/ref-iterators:
for_each_reflog(): reimplement using iterators
dir_iterator: new API for iterating over a directory tree
for_each_reflog(): don't abort for bad references
do_for_each_ref(): reimplement using reference iteration
refs: introduce an iterator interface
ref_resolves_to_object(): new function
entry_resolves_to_object(): rename function from ref_resolves_to_object()
get_ref_cache(): only create an instance if there is a submodule
remote rm: handle symbolic refs correctly
delete_refs(): add a flags argument
refs: use name "prefix" consistently
do_for_each_ref(): move docstring to the header file
refs: remove unnecessary "extern" keywords
Error handling in the codepaths that updates refs has been
improved.
* mh/update-ref-errors:
lock_ref_for_update(): avoid a symref resolution
lock_ref_for_update(): make error handling more uniform
t1404: add more tests of update-ref error handling
t1404: document function test_update_rejected
t1404: remove "prefix" argument to test_update_rejected
t1404: rename file to t1404-update-ref-errors.sh
Further preparatory work on the refs API before the pluggable
backend series can land.
* mh/split-under-lock: (33 commits)
lock_ref_sha1_basic(): only handle REF_NODEREF mode
commit_ref_update(): remove the flags parameter
lock_ref_for_update(): don't resolve symrefs
lock_ref_for_update(): don't re-read non-symbolic references
refs: resolve symbolic refs first
ref_transaction_update(): check refname_is_safe() at a minimum
unlock_ref(): move definition higher in the file
lock_ref_for_update(): new function
add_update(): initialize the whole ref_update
verify_refname_available(): adjust constness in declaration
refs: don't dereference on rename
refs: allow log-only updates
delete_branches(): use resolve_refdup()
ref_transaction_commit(): correctly report close_ref() failure
ref_transaction_create(): disallow recursive pruning
refs: make error messages more consistent
lock_ref_sha1_basic(): remove unneeded local variable
read_raw_ref(): move docstring to header file
read_raw_ref(): improve docstring
read_raw_ref(): rename symref argument to referent
...
When c5c31d33 (grep: move pattern-type bits support to top-level
grep.[ch], 2012-10-03) introduced grep_commit_pattern_type() helper
function, the intention was to allow the users of grep API to having
to fiddle only with .pattern_type_option (which can be set to "fixed",
"basic", "extended", and "pcre"), and then immediately before compiling
the pattern strings for use, call grep_commit_pattern_type() to have
it prepare various bits in the grep_opt structure (like .fixed,
.regflags, etc.).
However, grep_set_pattern_type_option() helper function the grep API
internally uses were left as an external function by mistake. This
function shouldn't have been made callable by the users of the API.
Later when the grep API was used in revision traversal machinery,
the caller then mistakenly started calling the function around
34a4ae55 (log --grep: use the same helper to set -E/-F options as
"git grep", 2012-10-03), instead of setting the .pattern_type_option
field and letting the grep_commit_pattern_type() to take care of the
details.
This caused an unnecessary bug that made a configured
grep.patternType take precedence over the command line options
(e.g. --basic-regexp, --fixed-strings) in "git log" family of
commands.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
'git submodule--helper update-clone' has logic to retry failed clones
a second time. For this purpose, there is a list of submodules to clone,
and a second list that is filled with the submodules to retry. Within
these lists, the submodules are identified by an index as if both lists
were just appended.
This works nicely except when the second clone attempt fails as well. To
report an error, the identifying index must be adjusted by an offset so
that it can be used as an index into the second list. However, the
calculation uses the logical total length of the lists so that the result
always points one past the end of the second list.
Pick the correct index.
Signed-off-by: Johannes Sixt <j6t@kdbg.org>
Acked-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
git-submodule--helper is invoked as the upstream of a pipe in several
places. Usually, the failure of a program in this position is not
detected by the shell. For this reason, the code inserts a token in the
output stream when git-submodule--helper fails that is detected
downstream, where the shell script is quit with exit code 1.
There happens to be a bug in git-submodule--helper that leads to a
segmentation fault. The test suite triggers the crash in several places,
all of which are protected by 'test_must_fail'. But due to the inspecific
exit code 1, the crash remains undiagnosed.
Extend the failure protocol such that git-submodule--helper's exit code
is passed downstream (only in the case of failure). This enables the
downstream to use it as its own exit code, and 'test_must_fail' to
identify the segmentation fault as an unexpected failure.
The bug itself is fixed in the next commit.
Signed-off-by: Johannes Sixt <j6t@kdbg.org>
Acked-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Checking the version of the installed SVN libraries should not
require a git repository at all. This matches the behavior of
"git --version".
Add a test for "git svn help" for the same behavior while we're
at it, too.
Signed-off-by: Eric Wong <e@80x24.org>
In the transport protocol we use NAK to signal the non existence of a
common base, so fix the documentation. This helps readers of the document,
as they don't have to wonder about the difference between NAK and NACK.
As NACK is used in git archive and upload-archive, this is easy to get
wrong.
Signed-off-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If you have whitespace errors in lines you've introduced, it
can be convenient to be able to jump directly to them for
fixing. You can't quite use "git jump diff" for this,
because though it passes arbitrary options to "git diff", it
expects to see an actual unified diff in the output.
Whereas "git diff --check" actually produces lines that look
like compiler quickfix lines already, meaning we just need
to run it and feed the output directly to the editor.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The hunk-header regex looks for "\+\d+" to find the
post-image line numbers, but it skips the pre-image line
numbers with a simple ".*". That means we may greedily eat
the post-image numbers and match a "\+\d" further on, in the
funcname text.
For example, commit 6b9c38e has this hunk header:
diff --git a/t/t0006-date.sh b/t/t0006-date.sh
[...]
@@ -50,8 +50,8 @@ check_show iso-local "$TIME" '2016-06-15 14:13:20 +0000'
If you run:
git checkout 6b9c38e
git jump diff HEAD^ t/
it will erroneously match "+0000" as the starting line
number and jump there, rather than line 50.
We can fix it by just making the "skip" regex non-greedy,
taking the first "+" we see, which should be the post-image
line information.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When accessing a blob for a diff, we may try to reuse file
contents in the working tree, under the theory that it is
faster to mmap those file contents than it would be to
extract the content from the object database.
When we have to filter those contents, though, that
assumption does not hold. Even for our internal conversions
like CRLF, we have to allocate and fill a new buffer anyway.
But much worse, for external clean filters we have to exec
an arbitrary script, and we have no idea how expensive it
may be to run.
So let's skip this optimization when conversion into git's
"clean" form is required. This applies whenever the
"want_file" flag is false. When it's true, the caller
actually wants the smudged worktree contents, which the
reused file by definition already has (in fact, this is a
key optimization going the other direction, since reusing
the worktree file there lets us skip smudge filters).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous commit introduced the first use of ENOTSOCK. This macro is
not available on Windows. Define it as WSAENOTSOCK because that is the
corresponding error value reported by the Windows versions of socket
functions.
Signed-off-by: Johannes Sixt <j6t@kdbg.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous method simply used the UNIX timestamp of when the binary was
built as its build label.
$ make && ./git-remote-persistent-http -print_label
1469061546
This patch aims to align the label for this binary with the Git version
contained in the GIT-VERSION-FILE. This gives a better sense of the version
of the binary as it can be mapped to a particular revision or release of
Git itself. For example:
$ make && ./git-remote-persistent-http -print_label
2.9.1.275.g75676c8
Discussion of this patch is available on a related thread in the mailing
list surrounding this package called "contrib/persistent-https: update
ldflags syntax for Go 1.7+". The gmane.org link is:
http://article.gmane.org/gmane.comp.version-control.git/299653/
Signed-off-by: Parker Moore <parkrmoore@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Running `make all` in `contrib/persistent-https` results in a
failure on Go 1.7 and above.
Specifically, the error is:
go build -o git-remote-persistent-https \
-ldflags "-X main._BUILD_EMBED_LABEL 1468613136"
# _/Users/parkr/github/git/contrib/persistent-https
/usr/local/Cellar/go/1.7rc1/libexec/pkg/tool/darwin_amd64/link: -X
flag requires argument of the form importpath.name=value
make: *** [git-remote-persistent-https] Error 2
This `name=value` syntax for the -X flag was introduced in Go v1.5
(released Aug 19, 2015):
- release notes: https://golang.org/doc/go1.5#link
- commit: 12795c02f3
In Go v1.7, support for the old syntax was removed:
- release notes: https://tip.golang.org/doc/go1.7#compiler
- commit: 51b624e6a2
Add '=' between the symbol and its value for recent versions of Go,
while leaving it out for older ones.
Signed-off-by: Parker Moore <parkrmoore@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We already suggest 'git rebase --abort' during a conflicted rebase.
Similarly, suggest 'git merge --abort' during conflict resolution on
'git merge'.
Signed-off-by: Matthieu Moy <Matthieu.Moy@imag.fr>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Implement strbuf_addbuf() as a normal function in order to avoid calling
strbuf_grow() twice, with the second callinside strbud_add() being a
no-op. This is slightly faster and also reduces the text size a bit.
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When the GPG prereq is not set, we do not run test 34. That
test changes the directory of the test script as a side
effect (something we usually frown on, but which matches the
style of the rest of this script). When test 35 (the
url-scrubbing test) runs, it expects to be in the directory
from test 34. If it's not, the test fails; we are in a
different sub-repo, our test-commit is built on a different
history, and the push becomes a non-fast-forward.
We can fix this by unconditionally moving to the directory
we expect (again, against our usual style but matching how
the rest of the script operates).
As an additional protection, let's also switch from "make a
new commit and push to master" to just "push to a new
branch". We don't care about the branch name; we just want
_some_ ref update to trigger the status output. Pushing to a
new branch is less likely to run into problems with
force-updates, changing the checked-out branch, etc.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
After a client has sent us the complete pack, we may spend
some time processing the data and running hooks. If the
client asked us to be quiet, receive-pack won't send any
progress data during the index-pack or connectivity-check
steps. And hooks may or may not produce their own progress
output. In these cases, the network connection is totally
silent from both ends.
Git itself doesn't care about this (it will wait forever),
but other parts of the system (e.g., firewalls,
load-balancers, etc) might hang up the connection. So we'd
like to send some sort of keepalive to let the network and
the client side know that we're still alive and processing.
We can use the same trick we did in 05e9515 (upload-pack:
send keepalive packets during pack computation, 2013-09-08).
Namely, we will send an empty sideband data packet every `N`
seconds that we do not relay any stderr data over the
sideband channel. As with 05e9515, this means that we won't
bother sending keepalives when there's actual progress data,
but will kick in when it has been disabled (or if there is a
lull in the progress data).
The concept is simple, but the details are subtle enough
that they need discussing here.
Before the client sends us the pack, we don't want to do any
keepalives. We'll have sent our ref advertisement, and we're
waiting for them to send us the pack (and tell us that they
support sidebands at all).
While we're receiving the pack from the client (or waiting
for it to start), there's no need for keepalives; it's up to
them to keep the connection active by sending data.
Moreover, it would be wrong for us to do so. When we are the
server in the smart-http protocol, we must treat our
connection as half-duplex. So any keepalives we send while
receiving the pack would potentially be buffered by the
webserver. Not only does this make them useless (since they
would not be delivered in a timely manner), but it could
actually cause a deadlock if we fill up the buffer with
keepalives. (It wouldn't be wrong to send keepalives in this
phase for a full-duplex connection like ssh; it's simply
pointless, as it is the client's responsibility to speak).
As soon as we've gotten all of the pack data, then the
client is waiting for us to speak, and we should start
keepalives immediately. From here until the end of the
connection, we send one any time we are not otherwise
sending data.
But there's a catch. Receive-pack doesn't know the moment
we've gotten all the data. It passes the descriptor to
index-pack, who reads all of the data, and then starts
resolving the deltas. We have to communicate that back.
To make this work, we instruct the sideband muxer to enable
keepalives in three phases:
1. In the beginning, not at all.
2. While reading from index-pack, wait for a signal
indicating end-of-input, and then start them.
3. Afterwards, always.
The signal from index-pack in phase 2 has to come over the
stderr channel which the muxer is reading. We can't use an
extra pipe because the portable run-command interface only
gives us stderr and stdout.
Stdout is already used to pass the .keep filename back to
receive-pack. We could also send a signal there, but then we
would find out about it in the main thread. And the
keepalive needs to be done by the async muxer thread (since
it's the one writing sideband data back to the client). And
we can't reliably signal the async thread from the main
thread, because the async code sometimes uses threads and
sometimes uses forked processes.
Therefore the signal must come over the stderr channel,
where it may be interspersed with other random
human-readable messages from index-pack. This patch makes
the signal a single NUL byte. This is easy to parse, should
not appear in any normal stderr output, and we don't have to
worry about any timing issues (like seeing half the signal
bytes in one read(), and half in a subsequent one).
This is a bit ugly, but it's simple to code and should work
reliably.
Another option would be to stop using an async thread for
muxing entirely, and just poll() both stderr and stdout of
index-pack from the main thread. This would work for
index-pack (because we aren't doing anything useful in the
main thread while it runs anyway). But it would make the
connectivity check and the hook muxers much more
complicated, as they need to simultaneously feed the
sub-programs while reading their stderr.
The index-pack phase is the only one that needs this
signaling, so it could simply behave differently than the
other two. That would mean having two separate
implementations of copy_to_sideband (and the keepalive
code), though. And it still doesn't get rid of the
signaling; it just means we can write a nicer message like
"END_OF_INPUT" or something on stdout, since we don't have
to worry about separating it from the stderr cruft.
One final note: this signaling trick is only done with
index-pack, not with unpack-objects. There's no point in
doing it for the latter, because by definition it only kicks
in for a small number of objects, where keepalives are not
as useful (and this conveniently lets us avoid duplicating
the implementation).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we receive a large push, the server side of the
connection may spend a lot of time (30s or more for a full
push of linux.git) walking the object graph without
producing any output. Let's give the user some indication
that we're actually working.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If the connectivity check encounters a problem when
receiving a push, the error output goes to receive-pack's
stderr, whose destination depends on the protocol used
(ssh tends to send it to the user, though without a "remote"
prefix; http will generally eat it in the server's error
log).
The information should consistently go back to the user, as
there is a reasonable chance their client is buggy and
generating a bad pack.
We can do so by muxing it over the sideband as we do with
other sub-process stderr.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we receive a large push, the server side may have to
spend a lot of CPU processing the incoming packfile.
During the "receiving" phase, we are typically network
bound, and the client is writing its own progress to the
user. But during the delta resolution phase, we may spend
minutes (e.g., for a full push of linux.git) without
making any indication to the user that the connection has
not hung.
Let's ask index-pack to produce progress output for this
phase (unless the client asked us to be quiet, of course).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The index-pack command has two progress meters: one for
"receiving objects", and one for "resolving deltas". You get
neither by default, or both with "-v".
But for a push through receive-pack, we would want only the
"resolving deltas" phase, _not_ the "receiving objects"
progress. There are two reasons for this.
One is simply that existing clients are already printing
"writing objects" progress at the same time. Arguably
"receiving" from the far end is more useful, because it
tells you what has actually gotten there, as opposed to what
might be stuck in a buffer somewhere between the client and
server. But that would require a protocol extension to tell
clients not to print their progress. Possible, but
complexity for little gain.
The second reason is much more important. In a full-duplex
connection like git-over-ssh, we can print progress while
the pack is incoming, and it will immediately get to the
client. But for a half-duplex connection like git-over-http,
we should not say anything until we have received the full
request. Anything we write is subject to being stuck in a
buffer by the webserver. Worse, we can end up in a deadlock
if that buffer fills up.
So our best bet is to avoid writing anything that isn't a
small fixed size until we've received the full pack.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Because the initial connectivity check for a cloned
repository can be slow, 0781aa4 (clone: let the user know
when check_everything_connected is run, 2013-05-03) added a
"fake" progress meter; we simply say "Checking connectivity"
when it starts, and "done" at the end, with nothing between.
Since check_connected() now knows how to do a real progress
meter, we can drop our fake one and use that one instead.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Connectivity checks have to traverse the entire object graph
in the worst case (e.g., a full clone or a full push). For
large repositories like linux.git, this can take 30-60
seconds, during which time git may produce little or no
output.
Let's add the option of showing progress, which is taken
care of by rev-list.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>