Many places in "git apply" knew that "/dev/null" that signals
"there is no such file on this side of the diff" can be followed by
whitespace and garbage when parsing a patch, except for one, which
made an otherwise valid patch (e.g. ones from subversion) rejected.
* tk/apply-dev-null-verify-name-fix:
apply: handle Subversion diffs with /dev/null gracefully
apply: demonstrate a problem applying svn diffs
"git am" has learned the "--quit" option, in addition to the existing
"--abort" option; having the pair mirrors a few other commands like
"rebase" and "cherry-pick".
Merge branch 'sb/submodule-update-reset-fix' into maint
When resetting the working tree files recursively, the working tree
of submodules are now also reset to match.
* sb/submodule-update-reset-fix:
submodule: submodule_move_head omits old argument in forced case
unpack-trees: oneway_merge to update submodules
t/lib-submodule-update.sh: fix test ignoring ignored files in submodules
t/lib-submodule-update.sh: clarify test
Merge branch 'nd/ita-wt-renames-in-status' into maint
"git status" after moving a path in the working tree (hence making
it appear "removed") and then adding with the -N option (hence
making that appear "added") detected it as a rename, but did not
report the old and new pathnames correctly.
* nd/ita-wt-renames-in-status:
wt-status.c: handle worktree renames
wt-status.c: rename rename-related fields in wt_status_change_data
wt-status.c: catch unhandled diff status codes
wt-status.c: coding style fix
Use DIFF_DETECT_RENAME for detect_rename assignments
t2203: test status output with porcelain v2 format
* jk/test-hashmap-updates:
test-hashmap: use "unsigned int" for hash storage
test-hashmap: simplify alloc_test_entry
test-hashmap: use strbuf_getline rather than fgets
test-hashmap: use xsnprintf rather than snprintf
test-hashmap: check allocation computation for overflow
test-hashmap: use ALLOC_ARRAY rather than bare malloc
Code to unquote single-quoted string (used in the parser for
configuration files, etc.) did not diagnose bogus input correctly
and produced bogus results instead.
* jk/sq-dequote-on-bogus-input:
sq_dequote: fix extra consumption of source string
"git add" files in the same directory, but spelling the directory
path in different cases on case insensitive filesystem, corrupted
the name hash data structure and led to unexpected results. This
has been corrected.
* bp/name-hash-dirname-fix:
name-hash: properly fold directory names in adjust_dirname_case()
Doc update to warn against remaining bugs in untracked cache.
* ab/untracked-cache-invalidation-docs:
update-index doc: note the caveat with "could not open..."
update-index doc: note a fixed bug in the untracked cache
Some bugs around "untracked cache" feature have been fixed.
* nd/fix-untracked-cache-invalidation:
dir.c: ignore paths containing .git when invalidating untracked cache
dir.c: stop ignoring opendir() error in open_cached_dir()
dir.c: fix missing dir invalidation in untracked code
dir.c: avoid stat() in valid_cached_dir()
status: add a failing test showing a core.untrackedCache bug
* sg/test-i18ngrep:
t: make 'test_i18ngrep' more informative on failure
t: validate 'test_i18ngrep's parameters
t: move 'test_i18ncmp' and 'test_i18ngrep' to 'test-lib-functions.sh'
t5536: let 'test_i18ngrep' read the file without redirection
t5510: consolidate 'grep' and 'test_i18ngrep' patterns
t4001: don't run 'git status' upstream of a pipe
t6022: don't run 'git merge' upstream of a pipe
t5812: add 'test_i18ngrep's missing filename parameter
t5541: add 'test_i18ngrep's missing filename parameter
The log from "git daemon" can be redirected with a new option; one
relevant use case is to send the log to standard error (instead of
syslog) when running it from inetd.
"git format-patch" learned to give 72-cols to diffstat, which is
consistent with other line length limits the subcommand uses for
its output meant for e-mails.
* nd/format-patch-stat-width:
format-patch: reduce patch diffstat width to 72
format-patch: keep cover-letter diffstat wrapped in 72 columns
The transport-helper protocol c-style quotes the value of
any options passed to the helper via the "option <key> <value>"
directive. However, remote-curl doesn't actually unquote the
push-option values, meaning that we will send the quoted
version to the other side (whereas git-over-ssh would send
the raw value).
The pack-protocol.txt documentation defines the push-options
as a series of VCHARs, which excludes most characters that
would need quoting. But:
1. You can still see the bug with a valid push-option that
starts with a double-quote (since that triggers
quoting).
2. We do currently handle any non-NUL characters correctly
in git-over-ssh. So even though the spec does not say
that we need to handle most quoted characters, it's
nice if our behavior is consistent between protocols.
There are two new tests: the "direct" one shows that this
already works in the non-http case, and the http one covers
this bugfix.
Reported-by: Jon Simons <jon@jonsimons.org> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
We repeat many lines of setup code in the two http tests,
and further tests would need to repeat it again. Let's
factor this out into a function.
Incidentally, this also fixes an unlikely bug: if the httpd
root path contains a double-quote, our test_when_finished
would barf due to improper quoting (we escape the embedded
quotes, but not the $, meaning we expand the variable before
the eval).
Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Support for generating an rpm was dropped in ab214331cf ("Makefile: stop
pretending to support rpmbuild", 2016-04-04). We don't generate any
*.spec files so there is no need to clean them up.
Signed-off-by: Todd Zullinger <tmz@pobox.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
git-worktree.txt: fix indentation of example and text of 'add' command
When 4e85333197 (worktree: make add <path> <branch> dwim, 2017-11-26)
added an example command in a literal code block, it neglected to
insert a mandatory "+" line before the block. This omission resulted
in both the literal code block and the (existing) paragraph following
the block to be outdented, even though they should be indented under
the 'add' sub-command along with the rest of the text pertaining to
that command. Furthermore, the mandatory "+" line separating the code
block from the following text got rendered as a leading character on
the line ("+ If <commit-ish>...") rather than being treated as a
formatting directive.
Fix these problems by adding the missing "+" line before the example
code block.
Signed-off-by: Eric Sunshine <sunshine@sunshineco.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add the closing ")" to a parenthetical phrase introduced by 4e85333197
(worktree: make add <path> <branch> dwim, 2017-11-26).
While at it, add a missing ":" at the end of the same sentence since
it precedes an example literal command block.
Reported-by: Mike Nordell <tamlin.thefirst@gmail.com> Signed-off-by: Eric Sunshine <sunshine@sunshineco.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
git respects XDG_CACHE_HOME for the credential cache. So, we should
unset XDG_CACHE_HOME for the test environment, lest a user's custom one
cause failure in the test.
For example, t/t0301-credential-cache.sh expects a default directory
to be used if it hasn't explicitly set XDG_CACHE_HOME.
Signed-off-by: Genki Sky <sky@genki.is> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Merge branch 'jk/abort-clone-with-existing-dest' into maint
"git clone $there $here" is allowed even when here directory exists
as long as it is an empty directory, but the command incorrectly
removed it upon a failure of the operation.
* jk/abort-clone-with-existing-dest:
clone: do not clean up directories we didn't create
clone: factor out dir_exists() helper
t5600: modernize style
t5600: fix outdated comment about unborn HEAD
* rs/lose-leak-pending:
commit: remove unused function clear_commit_marks_for_object_array()
revision: remove the unused flag leak_pending
checkout: avoid using the rev_info flag leak_pending
bundle: avoid using the rev_info flag leak_pending
bisect: avoid using the rev_info flag leak_pending
object: add clear_commit_marks_all()
ref-filter: use clear_commit_marks_many() in do_merge_filter()
commit: use clear_commit_marks_many() in remove_redundant()
commit: avoid allocation in clear_commit_marks_many()
Merge branch 'jm/svn-pushmergeinfo-fix' into maint
"git svn dcommit" did not take into account the fact that a
svn+ssh:// URL with a username@ (typically used for pushing) refers
to the same SVN repository without the username@ and failed when
svn.pushmergeinfo option is set.
* jm/svn-pushmergeinfo-fix:
git-svn: fix svn.pushmergeinfo handling of svn+ssh usernames.
More abstraction of hash function from the codepath.
* bc/hash-algo:
hash: update obsolete reference to SHA1_HEADER
bulk-checkin: abstract SHA-1 usage
csum-file: abstract uses of SHA-1
csum-file: rename sha1file to hashfile
read-cache: abstract away uses of SHA-1
pack-write: switch various SHA-1 values to abstract forms
pack-check: convert various uses of SHA-1 to abstract forms
fast-import: switch various uses of SHA-1 to the_hash_algo
sha1_file: switch uses of SHA-1 to the_hash_algo
builtin/unpack-objects: switch uses of SHA-1 to the_hash_algo
builtin/index-pack: improve hash function abstraction
hash: create union for hash context allocation
hash: move SHA-1 macros to hash.h
The way "git reset --hard" reports the commit the updated HEAD
points at is made consistent with the way how the commit title is
generated by the other parts of the system. This matters when the
title is spread across physically multiple lines.
* tg/reset-hard-show-head-with-pretty:
reset --hard: make use of the pretty machinery
* ab/wildmatch-tests:
wildmatch test: mark test as EXPENSIVE_ON_WINDOWS
test-lib: add an EXPENSIVE_ON_WINDOWS prerequisite
wildmatch test: create & test files on disk in addition to in-memory
wildmatch test: perform all tests under all wildmatch() modes
wildmatch test: use test_must_fail, not ! for test-wildmatch
wildmatch test: remove dead fnmatch() test code
wildmatch test: use a paranoia pattern from nul_match()
wildmatch test: don't try to vertically align our output
wildmatch test: use more standard shell style
wildmatch test: indent with tabs, not spaces
Avoid mmapping small files while using packed refs (especially ones
with zero size, which would cause later munmap() to fail).
* kg/packed-ref-cache-fix:
packed_ref_cache: don't use mmap() for small files
load_contents(): don't try to mmap an empty file
packed_ref_iterator_begin(): make optimization more general
find_reference_location(): make function safe for empty snapshots
create_snapshot(): use `xmemdupz()` rather than a strbuf
struct snapshot: store `start` rather than `header_len`
* en/merge-recursive-fixes:
merge-recursive: add explanation for src_entry and dst_entry
merge-recursive: fix logic ordering issue
Tighten and correct a few testcases for merging and cherry-picking
Push the submodule version of collision-detecting SHA-1 hash
implementation a bit harder on builders.
* ab/sha1dc-build:
sha1dc_git.h: re-arrange an ifdef chain for a subsequent change
Makefile: under "make dist", include the sha1collisiondetection submodule
Makefile: don't error out under DC_SHA1_EXTERNAL if DC_SHA1_SUBMODULE=auto
worktree: add: fix 'post-checkout' not knowing new worktree location
Although "git worktree add" learned to run the 'post-checkout' hook in ade546be47 (worktree: invoke post-checkout hook, 2017-12-07), it
neglected to change to the directory of the newly-created worktree
before running the hook. Instead, the hook runs within the directory
from which the "git worktree add" command itself was invoked, which
effectively neuters the hook since it knows nothing about the new
worktree directory.
Further, ade546be47 failed to sanitize the environment before running
the hook, which means that user-assigned values of GIT_DIR and
GIT_WORK_TREE could mislead the hook about the location of the new
worktree. In the case of "git worktree add" being run from a bare
repository, the GIT_DIR="." assigned by Git itself leaks into the hook's
environment and breaks Git commands; this is so even when the working
directory is correctly changed to the new worktree before the hook runs
since ".", relative to the new worktree directory, does not point at the
bare repository.
Fix these problems by (1) changing to the new worktree's directory
before running the hook, and (2) sanitizing the environment of GIT_DIR
and GIT_WORK_TREE so hooks can't be confused by misleading values.
Enhance the t2025 'post-checkout' tests to verify that the hook is
indeed run within the correct directory and that Git commands invoked by
the hook compute Git-dir and top-level worktree locations correctly.
While at it, also add two new tests: (1) verify that the hook is run
within the correct directory even when the new worktree is created from
a sibling worktree (as opposed to the main worktree); (2) verify that
the hook is provided with correct context when the new worktree is
created from a bare repository (test provided by Lars Schneider).
Implementation Notes:
Rather than sanitizing the environment of GIT_DIR and GIT_WORK_TREE, an
alternative would be to set them explicitly, as is already done for
other Git commands run internally by "git worktree add". This patch opts
instead to sanitize the environment in order to clearly document that
the worktree is fully functional by the time the hook is run, thus does
not require special environmental overrides.
The hook is run manually, rather than via run_hook_le(), since it needs
to change the working directory to that of the worktree, and
run_hook_le() does not provide such functionality. As this is a one-off
case, adding 'run_hook' overloads which allow the directory to be set
does not seem warranted at this time.
Reported-by: Lars Schneider <larsxschneider@gmail.com> Signed-off-by: Eric Sunshine <sunshine@sunshineco.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Subsequent patches will introduce file formats that make use of a fanout
array and a sorted table containing hashes, just like packfiles.
Refactor the hash search in packfile.c into its own function, so that
those patches can make use of it as well.
Signed-off-by: Jonathan Tan <jonathantanmy@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
In commit 628522ec1439 ("sha1-lookup: more memory efficient search in
sorted list of SHA-1", 2008-04-09), a different algorithm for searching
a sorted list was introduced, together with a set of log statements
guarded by GIT_DEBUG_LOOKUP that are invoked both when using that
algorithm and when using the existing binary search. Those log
statements was meant for experiments and debugging, but with the removal
of the aforementioned different algorithm in commit f1068efefe6d
("sha1_file: drop experimental GIT_USE_LOOKUP search", 2017-08-09),
those log statements are probably no longer necessary.
Remove those statements.
Signed-off-by: Jonathan Tan <jonathantanmy@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
apply: handle Subversion diffs with /dev/null gracefully
Subversion generates diffs that can contain lines like this one:
--- /dev/null (nonexistent)
Let's teach Git's apply machinery to handle such a line gracefully.
This fixes https://github.com/git-for-windows/git/isues/1489
Signed-off-by: Tatyana Krasnukha <tatyana@synopsys.com> Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Documentation/git-status: clarify status table for porcelain mode
It is possible to have the output ' A' from 'git status --porcelain'
by adding a file using the '--intend-to-add' flag. Make this clear by
adding the pattern in the table of the documentation.
However the mode 'DM' (deleted in the index, modified in the working tree)
is not possible in the non-merge case in which the file only shows
as 'D ' (and adding it back to the worktree would show an additional line
of an '??' untracked file). It is also not possible in the merge case as
then the mode involves a 'U' on one side of the merge.
Remove that pattern.
Reported-by: Ross Light <light@google.com> Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
sq_dequote: fix extra consumption of source string
This fixes a (probably harmless) parsing problem in
sq_dequote_step(), in which we parse some bogus input
incorrectly rather than complaining that it's bogus.
Our shell-dequoting function is very strict: it can unquote
everything generated by sq_quote(), but not arbitrary
strings. In particular, it only allows characters outside of
the single-quoted string if they are immediately backslashed
and then the single-quoted string is resumed. So:
'foo'\''bar'
is OK. But these are not:
'foo'\'bar
'foo'\'
'foo'\'\''bar'
even though they are all valid shell. The parser has a funny
corner case here. When we see a backslashed character, we
keep incrementing the "src" pointer as we parse it. For a
single sq_dequote() call, that's OK; our next step is to
bail with an error, and we don't care where "src" points.
But if we're parsing multiple strings with sq_dequote_to_argv(),
then our next step is to see if the string is followed by
whitespace. Because we erroneously incremented the "src"
pointer, we don't barf on the bogus backslash that we
skipped. Instead, we may find whitespace that immediately
follows it, and continue as if all is well (skipping the
backslashed character completely!).
In practice, this shouldn't be a big deal. The input is
bogus, and our sq_quote() would never generate this bogus
input. In all but one callers, we are parsing input created
by an earlier call to sq_quote(). That final case is "git
shell", which parses shell-quoting generated by the client.
And in that case we use the singular sq_quote(), which has
always behaved correctly.
One might also wonder if you could provoke a read past the
end of the string. But the answer is no; we still parse
character by character, and would never advance past a NUL.
This patch implements the minimal fix, along with
documenting the restriction (which confused at least me
while reading the code). We should possibly consider
being more liberal in accepting valid shell-quoted words. I
suspect the code may actually be simpler, and it would be
more friendly to anybody generating or editing input by
hand. But I wanted to fix just the immediate bug in this
patch.
We don't have a direct way to unit-test the sq_dequote()
functions, but we can do this by feeding input to
GIT_CONFIG_PARAMETERS (which is not normally a user-facing
interface, but serves here as it expects to see sq_quote()
input from "git -c"). I've included both a bogus example,
and a related "good" one to confirm that we still parse it
correctly.
Noticed-by: Michael Haggerty <mhagger@alum.mit.edu> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
The hashmap API always use an unsigned value for storing
and comparing hashes. Whereas this test code uses "int".
This works out in practice since one can typically
round-trip between "int" and "unsigned int". But since this
is essentially reference code for the hashmap API, we should
model using the correct types.
Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
This function takes two ptr/len pairs, which implies that
they can be arbitrary buffers. But internally, it assumes
that each "ptr" is NUL-terminated at "len" (because we
memcpy an extra byte to pick up the NUL terminator).
In practice this works because each caller only ever passes
strlen(ptr) as the length. But let's drop the "len"
parameters to make our expectations clear.
Note that we can get rid of the "l1" and "l2" variables from
cmd_main() as a further cleanup, since they are now mostly
used to check whether the p1 and p2 arguments are present
(technically the length parameters conflated NULL with the
empty string, which we no longer do, but I think that is
actually an improvement).
Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
test-hashmap: use strbuf_getline rather than fgets
Using fgets() with a fixed-size buffer can lead to lines
being accidentally split across two calls if they are larger
than the buffer size.
As this is just a test helper, this is unlikely to be a
problem in practice. But since people may look at test
helpers as reference code, it's a good idea for them to
model the preferred behavior.
Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
In general, using a bare snprintf can truncate the resulting
buffer, leading to confusing results. In this case we know
that our buffer is sized large enough to accommodate our
loop, so there's no bug. However, we should use xsnprintf()
to document (and check) that assumption, and to model good
practice to people reading the code.
Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
test-hashmap: check allocation computation for overflow
When we allocate the test_entry flex-struct, we have to add
up all of the elements that go into the flex array. If these
were to overflow a size_t, this would allocate a too-small
buffer, which we would then overflow in our memcpy steps.
Since this is just a test-helper, it probably doesn't matter
in practice, but we should model the correct technique by
using the st_add() macros.
Unfortunately, we cannot use the FLEX_ALLOC() macros here,
because we are stuffing two different buffers into a single
flex array.
While we're here, let's also swap out "malloc" for our
error-checking "xmalloc", and use the preferred
"sizeof(*var)" instead of "sizeof(type)".
Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
test-hashmap: use ALLOC_ARRAY rather than bare malloc
These two array allocations have several minor flaws:
- they use bare malloc, rather than our error-checking
xmalloc
- they do a bare multiplication to determine the total
size (which in theory can overflow, though in this case
the sizes are all constants)
- they use sizeof(type), but the type in the second one
doesn't match the actual array (though it's "int" versus
"unsigned int", which are guaranteed by C99 to have the
same size)
None of these are likely to be problems in practice, and
this is just a test helper. But since people often look at
test helpers as reference code, we should do our best to
model the recommended techniques.
Switching to ALLOC_ARRAY fixes all three.
Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
* 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
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