In the past, we used to execute `git gc --auto` as part of our automatic
housekeeping routines. As git-gc(1) may require quite some time to
perform the housekeeping, it knows to detach itself and run in the
background so that the user can continue their work.
Eventually, we refactored our automatic housekeeping to instead use the
more flexible git-maintenance(1) command. The upside of this new infra
is that the user can configure which maintenance tasks are performed, at
least to a certain degree. So while it continues to run git-gc(1) by
default, it can also be adapted to e.g. use git-multi-pack-index(1) for
maintenance of the object database.
The auto-detach of the new infra is somewhat broken though once the user
configures non-standard tasks. The problem is essentially that we detach
at the wrong level in the process hierarchy: git-maintenance(1) never
detaches itself, but instead it continues to be git-gc(1) which does.
When configured to only run the git-gc(1) maintenance task, then the
result is basically the same as before. But when configured to run other
tasks, then git-maintenance(1) will wait for these to run to completion.
Even worse, it may be that git-gc(1) runs concurrently with other
housekeeping tasks, stomping on each others feet.
Fix this bug by asking git-gc(1) to not detach when it is being invoked
via git-maintenance(1). Instead, git-maintenance(1) now respects a new
config "maintenance.autoDetach", the equivalent of "gc.autoDetach", and
detaches itself into the background when running as part of our auto
maintenance. This should continue to behave the same for all users which
use the git-gc(1) task, only. For others though, it means that we now
properly perform all tasks in the background. The default behaviour of
git-maintenance(1) when executed by the user does not change, it will
remain in the foreground unless they pass the `--detach` option.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Same as the preceding commit, add a `--[no-]detach` flag to the
git-maintenance(1) command. This will be used in a subsequent commit to
fix backgrounding of that command when configured with a non-standard
set of tasks.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When running `git gc --auto`, the command will by default detach and
continue running in the background. This behaviour can be tweaked via
the `gc.autoDetach` config, but not via a command line switch. We need
that in a subsequent commit though, where git-maintenance(1) will want
to ask its git-gc(1) child process to not detach anymore.
Add a `--[no-]detach` flag that does this for us.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When detaching, git-gc(1) will redirect its stderr to a "gc.log" log
file, which is then used to surface errors of a backgrounded process to
the user. To ensure that the file is properly managed on abnormal exit
paths, we install both signal and exit handlers that try to either
commit the underlying lock file or roll it back in case there wasn't any
error.
This logic is severly broken when handling signals though, as we end up
calling all kinds of functions that are not signal safe. This includes
malloc(3P) via `git_path()`, fprintf(3P), fflush(3P) and many more
functions. The consequence can be anything, from deadlocks to crashes.
Unfortunately, we cannot really do much about this without a larger
refactoring.
The least-worst thing we can do is to not set up the signal handler in
the first place. This will still cause us to remove the lockfile, as the
underlying tempfile subsystem already knows to unlink locks when
receiving a signal. But it may cause us to remove the lock even in the
case where it would have contained actual errors, which is a change in
behaviour.
The consequence is that "gc.log" will not be committed, and thus
subsequent calls to `git gc --auto` won't bail out because of this.
Arguably though, it is better to retry garbage collection rather than
having the process run into a potentially-corrupted state.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We're leaking config values in git-gc(1) when those values are tracked
as strings. Introduce a new `gc_config_release()` function that releases
this memory to plug those leaks and release old values before populating
the config fields via `git_config_string()` et al.
Note that there is one small gotcha here with the "--prune" option. Next
to passing a string, this option also accepts the "--no-prune" option
that overrides the default or configured value. We thus need to discern
between the option not having been passed by the user and the negative
variant of it. This is done by using a simple sentinel value that lets
us discern these cases.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The git-gc(1) command knows to read a bunch of config keys to tweak its
own behaviour. The values are parsed into global variables, which makes
it hard to correctly manage the lifecycle of values that may require a
memory allocation.
Refactor the code to use a `struct gc_config` that gets populated and
passed around. For one, this makes previously-implicit dependencies on
these config values clear. Second, it will allow us to properly manage
the lifecycle in the next commit.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The type of the out parameter of `git_config_get_expiry()` is a pointer
to a constant string, which creates the impression that ownership of the
returned data wasn't transferred to the caller. This isn't true though
and thus quite misleading.
Adapt the parameter to be of type `char **` and adjust callers
accordingly. While at it, refactor `get_shared_index_expire_date()` to
drop the static `shared_index_expire` variable. It is only used in that
function, and furthermore we would only hit the code where we parse the
expiry date a single time because we already use a static `prepared`
variable to track whether we did parse it.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git rev-list ... | git diff-tree -p --remerge-diff --stdin" should
behave more or less like "git log -p --remerge-diff" but instead it
crashed, forgetting to prepare a temporary object store needed.
* xx/diff-tree-remerge-diff-fix:
diff-tree: fix crash when used with --remerge-diff
The refs API has been taught to give symref target information to
the users of ref iterators, allowing for-each-ref and friends to
avoid an extra ref_resolve_* API call per a symbolic ref.
* jc/refs-symref-referent:
ref-filter: populate symref from iterator
refs: add referent to each_ref_fn
refs: keep track of unresolved reference value in iterators
Support to specify ref backend for submodules has been enhanced.
* ps/submodule-ref-format:
object: fix leaking packfiles when closing object store
submodule: fix leaking seen submodule names
submodule: fix leaking fetch tasks
builtin/submodule: allow "add" to use different ref storage format
refs: fix ref storage format for submodule ref stores
builtin/clone: propagate ref storage format to submodules
builtin/submodule: allow cloning with different ref storage format
git-submodule.sh: break overly long command lines
Coding style fixes to a test script.
* ag/t7004-modernize:
t7004: make use of write_script
t7004: use single quotes instead of double quotes
t7004: begin the test body on the same line as test_expect_success
t7004: description on the same line as test_expect_success
t7004: do not prepare things outside test_expect_success
t7004: use indented here-doc
t7004: one command per line
t7004: remove space after redirect operators
The code paths to compact multiple reftable files have been updated
to correctly deal with multiple compaction triggering at the same
time.
* ps/reftable-stack-compaction:
reftable/stack: handle locked tables during auto-compaction
reftable/stack: fix corruption on concurrent compaction
reftable/stack: use lock_file when adding table to "tables.list"
reftable/stack: do not die when fsyncing lock file files
reftable/stack: simplify tracking of table locks
reftable/stack: update stats on failed full compaction
reftable/stack: test compaction with already-locked tables
reftable/stack: extract function to setup stack with N tables
reftable/stack: refactor function to gather table sizes
A policy document that describes platform support levels and
expectation on platform stakeholders has been introduced.
* es/doc-platform-support-policy:
Documentation: add platform support policy
An existing test of hashmap API has been rewritten with the
unit-test framework.
* gt/unit-test-hashmap:
t: port helper/test-hashmap.c to unit-tests/t-hashmap.c
When generating pseudo-merge bitmaps, it's possible that concurrent
reference updates may reveal some pseudo-merge candidates which reach
objects that are not contained in the bitmap's pack or pseudo-pack
order (in the case of MIDX bitmaps).
The latter case is relatively easy to demonstrate: if we generate a MIDX
bitmap with only half of the repository packed, then the unpacked
contents are not part of the MIDX's object order.
If we happen to select one or more commit(s) from the unpacked portion
of the repository for inclusion in a pseudo-merge, we'll get the
following message when trying to generate its bitmap:
$ git multi-pack-index write --bitmap
[...]
Selecting pseudo-merge commits: 100% (1/1), done.
warning: Failed to write bitmap index. Packfile doesn't have full closure (object ... is missing)
Building bitmaps: 50% (1/2), done.
error: could not write multi-pack bitmap
, and the attempted bitmap write will fail, leaving the repository
without a current bitmap.
Rectify this by ensuring that the commits which are pseudo-merge
candidates can only be so if they appear somewhere in the packing order.
This is sufficient, since we know that the original packing order is
closed under reachability, so if a commit appears in that list as a
potential pseudo-merge candidate, we know that everything reachable from
it also appears in the list (and thus the candidate is a good one).
Noticed-by: Jeff King <peff@peff.net>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous commit demonstrated it is possible to generate empty
pseudo-merge commits, which is not useful as such pseudo-merges carry no
information.
Ensure that we only generate non-empty groups by not pushing a new
commit onto the bitmap_writer when that commit has no parents.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Demonstrate that it is possible to generate empty pseudo-merge commits
in certain cases.
In the below instance, we generate one non-empty pseudo-merge
(containing commit "base"), and one empty pseudo-merge group
(corresponding to the unstable commits within that group).
(In my testing, the pseudo-merge machinery seems to handle empty groups
just fine, but generating them is pointless as they carry no
information.)
This commit (introducing a deliberate "test_expect_failure") is split
out from the actual fix (which will appear in the following commit) to
demonstrate that the failure is correctly induced.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Ordinarily, the pack-bitmap machinery will select some subset of
reachable commits to receive bitmaps. But when there are fewer than 100
commits indexed in the first place, they will all receive bitmaps as a
special case.
When this happens, pseudo-merges are not generated, making it impossible
to test pseudo-merge corner cases with fewer than 100 commits.
Select pseudo-merges even for bitmaps with fewer than 100 commits to
make such testing easier. In practice, this should not make a difference
to non-testing bitmaps, as they are unlikely to be used when a
repository has so few commits to begin with.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In a similar fashion as the previous commit, drop a redundant argument
from the `bitmap_writer_finish()` function.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In a similar fashion as the previous commit, drop a redundant argument
from the `bitmap_writer_build()` function.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous commit ensures that the bitmap_writer's "to_pack" field is
initialized early on, so the "to_pack" and "index_nr" arguments to
`bitmap_writer_build_type_index()` are redundant.
Drop them and adjust the callers accordingly.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In order to determine its object order, the pack-bitmap machinery keeps
a 'struct packing_data' corresponding to the pack or pseudo-pack (when
writing a MIDX bitmap) being written.
The to_pack field is provided to the bitmap machinery by callers of
bitmap_writer_build() and assigned to the bitmap_writer struct at that
point.
But a subsequent commit will want to have access to that data earlier on
during commit selection. Prepare for that by adding a 'to_pack' argument
to 'bitmap_writer_init()', and initializing the field during that
function.
Subsequent commits will clean up other functions which take
now-redundant arguments (like nr_objects, which is equivalent to
pdata->objects_nr, or pdata itself).
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This patch fixes a racy test failure in t4129.
The deletion test added by e95d515141 (apply: canonicalize modes read
from patches, 2024-08-05) wants to make sure that git-apply does not
complain about a non-canonical mode in the patch, even if that mode does
not match the working tree file. So it does this:
echo content >non-canon &&
git add non-canon &&
chmod 666 non-canon &&
This is wrong, because running chmod will update the ctime on the file,
making it stat-dirty and causing git-apply to refuse to apply the patch.
But this only happens sometimes, since it depends on the timestamps
crossing a second boundary (but it triggers pretty quickly when run with
--stress).
We can fix this by doing the chmod before updating the index. The order
isn't important here, as the mode will be canonicalized to 100644 in the
index anyway (in fact, the chmod is not even that important in the first
place, since git-apply will only look at the index; I only added it as
an extra confirmation that git-apply would not be confused by it).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* ps/reftable-stack-compaction:
reftable/stack: handle locked tables during auto-compaction
reftable/stack: fix corruption on concurrent compaction
reftable/stack: use lock_file when adding table to "tables.list"
reftable/stack: do not die when fsyncing lock file files
reftable/stack: simplify tracking of table locks
reftable/stack: update stats on failed full compaction
reftable/stack: test compaction with already-locked tables
reftable/stack: extract function to setup stack with N tables
reftable/stack: refactor function to gather table sizes
We would strip all leading and trailing whitespace, which git commit
does not. Let's be consistent here.
Signed-off-by: Oswald Buddenhagen <oswald.buddenhagen@gmx.de>
Signed-off-by: Johannes Sixt <j6t@kdbg.org>
This is also known as "washing". This is consistent with the behavior of
interactive git commit, which we should emulate as closely as possible
to avoid usability problems. This way commit message templates and
prepare hooks can be used properly, and comments from conflicted rebases
and merges are cleaned up without having to introduce special handling
for them.
Signed-off-by: Oswald Buddenhagen <oswald.buddenhagen@gmx.de>
Signed-off-by: Johannes Sixt <j6t@kdbg.org>
Forgot to mention that the preformatted documentation repositories
are updated every time the master branch of the project advances.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Git-SVN was previously taught to use the svn:global-ignores property as
well as svn:ignore when creating or showing .gitignore files from a
Subversion repository. However, the documentation and help message still
only mentioned svn:ignore. Update Git-SVN's documentation and help
command to mention support for the new property. Also capitalize the
help message for the 'mkdirs' command, for consistency.
Signed-off-by: Alex Galvin <agalvin@comqi.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A test that fails on an unusually slow machine was found, and made
less likely to cause trouble by lengthening the expiry value it
uses.
* tb/t7704-deflake:
t/t7704-repack-cruft.sh: avoid failures during long-running tests
A test in reftable library has been rewritten using the unit test
framework.
* cp/unit-test-reftable-tree:
t-reftable-tree: improve the test for infix_walk()
t-reftable-tree: add test for non-existent key
t-reftable-tree: split test_tree() into two sub-test functions
t: move reftable/tree_test.c to the unit testing framework
reftable: remove unnecessary curly braces in reftable/tree.c
A flakey test and incorrect calls to strtoX() functions have been
fixed.
* kl/test-fixes:
t6421: fix test to work when repo dir contains d0
set errno=0 before strtoX calls
The patch parser in "git patch-id" has been tightened to avoid
getting confused by lines that look like a patch header in the log
message.
* jc/patch-id:
patch-id: tighten code to detect the patch header
patch-id: rewrite code that detects the beginning of a patch
patch-id: make get_one_patchid() more extensible
patch-id: call flush_current_id() only when needed
t4204: patch-id supports various input format
In the refs subsystem, implicit reliance of the_repository has been
eliminated; the repository associated with the ref store object is
used instead.
* ps/refs-wo-the-repository:
refs/reftable: stop using `the_repository`
refs/packed: stop using `the_repository`
refs/files: stop using `the_repository`
refs/files: stop using `the_repository` in `parse_loose_ref_contents()`
refs: stop using `the_repository`
"git config --value=foo --fixed-value section.key newvalue" barfed
when the existing value in the configuration file used the
valueless true syntax, which has been corrected.
* tb/config-fixed-value-with-valueless-true:
config.c: avoid segfault with --fixed-value and valueless config
The patch parser in 'git apply' has been a bit more lenient against
unexpected mode bits, like 100664, recorded on extended header lines.
* jk/apply-patch-mode-check-fix:
apply: canonicalize modes read from patches
A recent update broke "git ls-remote" used outside a repository,
which has been corrected.
* ps/ls-remote-out-of-repo-fix:
builtin/ls-remote: fall back to SHA1 outside of a repo
