Travis updates.
* sg/travis-linux32-sanity:
travis-ci: don't fail if user already exists on 32 bit Linux build job
travis-ci: don't run the test suite as root in the 32 bit Linux build
travis-ci: don't repeat the path of the cache directory
travis-ci: use 'set -e' in the 32 bit Linux build job
travis-ci: use 'set -x' for the commands under 'su' in the 32 bit Linux build
Completion of "git merge -s<strategy>" (in contrib/) did not work
well in non-C locale.
* nd/list-merge-strategy:
completion: fix completing merge strategies on non-C locales
The sequencer infrastructure is shared across "git cherry-pick",
"git rebase -i", etc., and has always spawned "git commit" when it
needs to create a commit. It has been taught to do so internally,
when able, by reusing the codepath "git commit" itself uses, which
gives performance boost for a few tens of percents in some sample
scenarios.
* pw/sequencer-in-process-commit:
sequencer: run 'prepare-commit-msg' hook
t7505: add tests for cherry-pick and rebase -i/-p
t7505: style fixes
sequencer: assign only free()able strings to gpg_sign
sequencer: improve config handling
t3512/t3513: remove KNOWN_FAILURE_CHERRY_PICK_SEES_EMPTY_COMMIT=1
sequencer: try to commit without forking 'git commit'
sequencer: load commit related config
sequencer: simplify adding Signed-off-by: trailer
commit: move print_commit_summary() to libgit
commit: move post-rewrite code to libgit
Add a function to update HEAD after creating a commit
commit: move empty message checks to libgit
t3404: check intermediate squash messages
Code clean-up.
* nd/shared-index-fix:
read-cache: don't write index twice if we can't write shared index
read-cache.c: move tempfile creation/cleanup out of write_shared_index
read-cache.c: change type of "temp" in write_shared_index()
Prevent "clang-format" from breaking line after function return type.
* po/clang-format-functype-weight:
clang-format: adjust penalty for return type line break
The split-index mode had a few corner case bugs fixed.
* tg/split-index-fixes:
travis: run tests with GIT_TEST_SPLIT_INDEX
split-index: don't write cache tree with null oid entries
read-cache: fix reading the shared index for other repos
Update Coccinelle rules to catch and optimize strbuf_addf(&buf, "%s", str)
* rs/strbuf-cocci-workaround:
cocci: use format keyword instead of a literal string
Crash fix for a corner case where an error codepath tried to unlock
what it did not acquire lock on.
* mr/packed-ref-store-fix:
files_initial_transaction_commit(): only unlock if locked
The http tracing code, often used to debug connection issues,
learned to redact potentially sensitive information from its output
so that it can be more safely sharable.
* jt/http-redact-cookies:
http: support omitting data from traces
http: support cookie redaction when tracing
The tracing machinery learned to report tweaking of environment
variables as well.
* nd/trace-with-env:
run-command.c: print new cwd in trace_run_command()
run-command.c: print env vars in trace_run_command()
run-command.c: print program 'git' when tracing git_cmd mode
run-command.c: introduce trace_run_command()
trace.c: move strbuf_release() out of print_trace_line()
trace: avoid unnecessary quoting
sq_quote_argv: drop maxlen parameter
Rewrite two more "git submodule" subcommands in C.
* pc/submodule-helper:
submodule: port submodule subcommand 'deinit' from shell to C
submodule: port submodule subcommand 'sync' from shell to C
Avoid showing a warning message in the middle of a line of "git
diff" output.
* nd/diff-flush-before-warning:
diff.c: flush stdout before printing rename warnings
Doc updates.
* ks/submodule-doc-updates:
Doc/git-submodule: improve readability and grammar of a sentence
Doc/gitsubmodules: make some changes to improve readability and syntax
Retire mru API as it does not give enough abstraction over
underlying list API to be worth it.
* gs/retire-mru:
mru: Replace mru.[ch] with list.h implementation
The first step to getting rid of mru API and using the
doubly-linked list API directly instead.
* ot/mru-on-list:
mru: use double-linked list from list.h
The machinery to clone & fetch, which in turn involves packing and
unpacking objects, have been told how to omit certain objects using
the filtering mechanism introduced by the jh/object-filtering
topic, and also mark the resulting pack as a promisor pack to
tolerate missing objects, taking advantage of the mechanism
introduced by the jh/fsck-promisors topic.
* jh/partial-clone:
t5616: test bulk prefetch after partial fetch
fetch: inherit filter-spec from partial clone
t5616: end-to-end tests for partial clone
fetch-pack: restore save_commit_buffer after use
unpack-trees: batch fetching of missing blobs
clone: partial clone
partial-clone: define partial clone settings in config
fetch: support filters
fetch: refactor calculation of remote list
fetch-pack: test support excluding large blobs
fetch-pack: add --no-filter
fetch-pack, index-pack, transport: partial clone
upload-pack: add object filtering for partial clone
In preparation for implementing narrow/partial clone, the machinery
for checking object connectivity used by gc and fsck has been
taught that a missing object is OK when it is referenced by a
packfile specially marked as coming from trusted repository that
promises to make them available on-demand and lazily.
* jh/fsck-promisors:
gc: do not repack promisor packfiles
rev-list: support termination at promisor objects
sha1_file: support lazily fetching missing objects
introduce fetch-object: fetch one promisor object
index-pack: refactor writing of .keep files
fsck: support promisor objects as CLI argument
fsck: support referenced promisor objects
fsck: support refs pointing to promisor objects
fsck: introduce partialclone extension
extension.partialclone: introduce partial clone extension
The build procedure for perl/ part has been greatly simplified by
weaning ourselves off of MakeMaker.
* ab/simplify-perl-makefile:
perl: treat PERLLIB_EXTRA as an extra path again
perl: avoid *.pmc and fix Error.pm further
Makefile: replace perl/Makefile.PL with simple make rules
The 32 bit Linux build job runs in a Docker container, which lends
itself to running and debugging locally, too. Especially during
debugging one usually doesn't want to start with a fresh container
every time, to save time spent on installing a bunch of dependencies.
However, that doesn't work quite smootly, because the script running
in the container always creates a new user, which then must be removed
every time before subsequent executions, or the build script fails.
Make this process more convenient and don't try to create that user if
it already exists and has the right user ID in the container, so
developers don't have to bother with running a 'userdel' each time
before they run the build script.
The build job on Travis CI always starts with a fresh Docker
container, so this change doesn't make a difference there.
Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Travis CI runs the 32 bit Linux build job in a Docker container, where
all commands are executed as root by default. Therefore, ever since
we added this build job in 88dedd5e7 (Travis: also test on 32-bit
Linux, 2017-03-05), we have a bit of code to create a user in the
container matching the ID of the host user and then to run the test
suite as this user. Matching the host user ID is important, because
otherwise the host user would have no access to any files written by
processes running in the container, notably the logs of failed tests
couldn't be included in the build job's trace log.
Alas, this piece of code never worked, because it sets the variable
holding the user name ($CI_USER) in a subshell, meaning it doesn't
have any effect by the time we get to the point to actually use the
variable to switch users with 'su'. So all this time we were running
the test suite as root.
Reorganize that piece of code in 'ci/run-linux32-build.sh' a bit to
avoid that problematic subshell and to ensure that we switch to the
right user. Furthermore, make the script's optional host user ID
option mandatory, so running the build accidentally as root will
become harder when debugging locally. If someone really wants to run
the test suite as root, whatever the reasons might be, it'll still be
possible to do so by explicitly passing '0' as host user ID.
Finally, one last catch: since commit 7e72cfcee (travis-ci: save prove
state for the 32 bit Linux build, 2017-12-27) the 'prove' test harness
has been writing its state to the Travis CI cache directory from
within the Docker container while running as root. After this patch
'prove' will run as a regular user, so in future build jobs it won't
be able overwrite a previously written, still root-owned state file,
resulting in build job failures. To resolve this we should manually
delete caches containing such root-owned files, but that would be a
hassle. Instead, work this around by changing the owner of the whole
contents of the cache directory to the host user ID.
Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Some of our 'ci/*' scripts repeat the name or full path of the Travis
CI cache directory, and the following patches will add new places
using that path.
Use a variable to refer to the path of the cache directory instead, so
it's hard-coded only in a single place.
Pay extra attention to the 32 bit Linux build: it runs in a Docker
container, so pass the path of the cache directory from the host to
the container in an environment variable. Note that an environment
variable passed this way is exported inside the container, therefore
its value is directly available in the 'su' snippet even though that
snippet is single quoted. Furthermore, use the variable in the
container only if it's been assigned a non-empty value, to prevent
errors when someone is running or debugging the Docker build locally,
because in that case the variable won't be set as there won't be any
Travis CI cache.
Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The script 'ci/run-linux32-build.sh' running inside the Docker
container of the 32 bit Linux build job uses an && chain to break the
build if one of the commands fails. This is problematic for two
reasons:
- The && chain is broken, because there is this in the middle:
test -z $HOST_UID || (CI_USER="ci" && useradd -u $HOST_UID $CI_USER) &&
Luckily it is broken in a way that it didn't lead to false
successes. If installing dependencies fails, then the rest of the
first && chain is skipped and execution resumes after the ||
operator. At that point $HOST_UID is still unset, causing
'useradd' to error out with "invalid user ID 'ci'", which in turn
causes the second && chain to abort the script and thus break the
build.
- All other 'ci/*' scripts use 'set -e' to break the build if one of
the commands fails. This inconsistency among these scripts is
asking for trouble: I forgot about the && chain more than once
while working on this patch series.
Enable 'set -e' for the whole script and for the commands executed
under 'su' as well.
While touching every line in the 'su' command block anyway, change
their indentation to use a tab instead of spaces.
Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
It seems necessary to control destruction ordering to avoid a
segfault with SVN 1.9.5 when using "git svn branch". I've also
reported the problem against libsvn-perl to Debian [Bug #888791],
but releasing the SVN::Client instance can be beneficial anyways to
save memory.
ref: https://bugs.debian.org/888791
Tested-by: Todd Zullinger <tmz@pobox.com>
Reported-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Eric Wong <e@80x24.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When 'git show' is called without any object it defaults to HEAD. This
has been true since d4ed9793fd ("Simplify common default options setup
for built-in log family.", 2006-04-16).
The SYNOPSIS suggests that the object argument is required. Clarify
that it is not required and note the default.
Signed-off-by: Todd Zullinger <tmz@pobox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The anchor string "Available strategies are:" is translatable so
__git_list_merge_strategies may fail to collect available strategies
from 'git merge' on non-C locales. Force C locale on this command.
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When git-daemon gets a pktline request, we strip off any
trailing newline, replacing it with a NUL. Clients prior to
5ad312bede (in git v1.4.0) would send:
git-upload-pack repo.git\n
and we need to strip it off to understand their request.
After 5ad312bede, we send the host attribute but no newline,
like:
git-upload-pack repo.git\0host=example.com\0
Both of these are parsed correctly by git-daemon. But if
some client were to combine the two:
git-upload-pack repo.git\n\0host=example.com\0
we don't parse it correctly. The problem is that we use the
"len" variable to record the position of the NUL separator,
but then decrement it when we strip the newline. So we start
with:
git-upload-pack repo.git\n\0host=example.com\0
^-- len
and end up with:
git-upload-pack repo.git\0\0host=example.com\0
^-- len
This is arguably correct, since "len" tells us the length of
the initial string, but we don't actually use it for that.
What we do use it for is finding the offset of the extended
attributes; they used to be at len+1, but are now at len+2.
We can solve that by just leaving "len" where it is. We
don't have to care about the length of the shortened string,
since we just treat it like a C string.
No version of Git ever produced such a string, but it seems
like the daemon code meant to handle this case (and it seems
like a reasonable thing for somebody to do in a 3rd-party
implementation).
Reported-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
All of our git-protocol tests rely on invoking the client
and having it make a request of a server. That gives a nice
real-world test of how the two behave together, but it
doesn't leave any room for testing how a server might react
to _other_ clients.
Let's add a few test helper functions which can be used to
manually conduct a git-protocol conversation with a remote
git-daemon:
1. To connect to a remote git-daemon, we need something
like "netcat". But not everybody will have netcat. And
even if they do, the behavior with respect to
half-duplex shutdowns is not portable (openbsd netcat
has "-N", with others you must rely on "-q 1", which is
racy).
Here we provide a "fake_nc" that is capable of doing
a client-side netcat, with sane half-duplex semantics.
It relies on perl's IO::Socket::INET. That's been in
the base distribution since 5.6.0, so it's probably
available everywhere. But just to be on the safe side,
we'll add a prereq.
2. To help tests speak and read pktline, this patch adds
packetize() and depacketize() functions.
I've put fake_nc() into lib-git-daemon.sh, since that's
really the only server where we'd need to use a network
socket. Whereas the pktline helpers may be of more general
use, so I've added them to test-lib-functions.sh. Programs
like upload-pack speak pktline, but can talk directly over
stdio without a network socket.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If we receive a request with extended attributes after the
NUL, we try to write those attributes to the log. We do so
with a "%s" format specifier, which will only show
characters up to the first NUL.
That's enough for printing a "host=" specifier. But since
dfe422d04d (daemon: recognize hidden request arguments,
2017-10-16) we may have another NUL, followed by protocol
parameters, and those are not logged at all.
Let's cut out the attempt to show the whole string, and
instead log when we parse individual attributes. We could
leave the "extended attributes (%d bytes) exist" part of the
log, which in theory could alert us to attributes that fail
to parse. But anything we don't parse as a "host=" parameter
gets blindly added to the "protocol" attribute, so we'd see
it in that part of the log.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If receive a request like:
git-upload-pack /foo.git\0host=localhost
we mark the offset of the NUL byte as "len", and then log
the bytes after the NUL with a "%.*s" placeholder, using
"pktlen - len" as the length, and "line + len + 1" as the
start of the string.
This is off-by-one, since the start of the string skips past
the separating NUL byte, but the adjusted length includes
it. Fortunately this doesn't actually read past the end of
the buffer, since "%.*s" will stop when it hits a NUL. And
regardless of what is in the buffer, packet_read() will
always add an extra NUL terminator for safety.
As an aside, the git.git client sends an extra NUL after a
"host" field, too, so we'd generally hit that one first, not
the one added by packet_read(). You can see this in the test
output which reports 15 bytes, even though the string has
only 14 bytes of visible data. But the point is that even a
client sending unusual data could not get us to read past
the end of the buffer, so this is purely a cosmetic fix.
Reported-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we start git-daemon for our tests, we send its stderr
log stream to a named pipe. We synchronously read the first
line to make sure that the daemon started, and then dump the
rest to descriptor 4. This is handy for debugging test
output with "--verbose", but the tests themselves can't
access the log data.
Let's dump the log into a file, as well, so that future
tests can check the log. There are a few subtleties worth
calling out here:
- we'll continue to send output to descriptor 4 for
viewing/debugging, which would imply swapping out "cat"
for "tee". But we want to ensure that there's no
buffering, and "tee" doesn't have a standard way to
ask for that. So we'll use a shell loop around "read"
and "printf" instead. That ensures that after a request
has been served, the matching log entries will have made
it to the file.
- the existing first-line shell loop used read/echo. We'll
switch to consistently using "read -r" and "printf" to
relay data as faithfully as possible.
- we open the logfile for append, rather than just output.
That makes it OK for tests to truncate the logfile
without restarting the daemon (the OS will atomically
seek to the end of the file when outputting each line).
That allows tests to look at the log without worrying
about pollution from earlier tests.
Helped-by: Lucas Werkmeister <mail@lucaswerkmeister.de>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Separating out the implementation of the handshake when starting a
long-running subprocess (for example, as is done for a clean/smudge
filter) was done in commit fa64a2fdbe ("sub-process: refactor
handshake to common function", 2017-07-26), but its documentation still
resides in gitattributes. Split out the documentation as well.
Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We don't actually care about the clone operation here; we
just want to know if we were able to actually contact the
remote repository. Using ls-remote does that more
efficiently, and without us having to worry about managing
the tmp.git directory.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When git push fails due to server-side WebDAV error, it's not easy to
point to the main culprit. Additional information about exact cURL
error and HTTP server response is helpful for debugging purpose.
New error log helped me pinpoint failing test t5540-http-push-webdav
to a missing Apache dependency in Fedora 27:
https://bugzilla.redhat.com/show_bug.cgi?id=1491151
Signed-off-by: Patryk Obara <patryk.obara@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The penalty of 5 makes clang-format very eager to put even short type
declarations (e.g. "extern int") into a separate line, even when
breaking parameters list is sufficient.
Signed-off-by: Patryk Obara <patryk.obara@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Commit 356ee4659b ("sequencer: try to commit without forking 'git
commit'", 2017-11-24) forgot to run the 'prepare-commit-msg' hook when
creating the commit. Fix this by writing the commit message to a
different file and running the hook. Using a different file means that
if the commit is cancelled the original message file is
unchanged. Also move the checks for an empty commit so the order
matches 'git commit'.
Reported-by: Dmitry Torokhov <dmitry.torokhov@gmail.com>
Helped-by: Ramsay Jones <ramsay@ramsayjones.plus.com>
Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Check that cherry-pick and rebase call the 'prepare-commit-msg' hook
correctly. The expected values for the hook arguments are taken to
match the current master branch. I think there is scope for improving
the arguments passed so they make a bit more sense - for instance
cherry-pick currently passes different arguments depending on whether
the commit message is being edited. Also the arguments for rebase
could be improved. Commit 7c4188360a ("rebase -i: proper
prepare-commit-msg hook argument when squashing", 2008-10-3) apparently
changed things so that when squashing rebase would pass 'squash' as
the argument to the hook but that has been lost.
I think that it would make more sense to pass 'message' for revert and
cherry-pick -x/-s (i.e. cases where there is a new message or the
current message in modified by the command), 'squash' when squashing
with a new message and 'commit HEAD/CHERRY_PICK_HEAD'
otherwise (picking and squashing without a new message).
Helped-by: Eric Sunshine <sunshine@sunshineco.com>
Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Junio C Hamano