| Commit message (Collapse) | Author | Files | Lines |
|---|
|
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Cherry pick commit d3775de0 (Makefile: force -O0 when compiling with
SANITIZE=leak, 2022-10-18), as otherwise the leak checker at GitHub
Actions CI seems to fail with a false positive.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Since c2d17ba3db0d (branch --edit-description: protect against mistyped
branch name, 2012-02-05) we return -1 on error editing the branch
description.
Let's change to 1, which follows the established convention and it is
better for portability reasons.
Signed-off-by: Rubén Justo <rjusto@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
In c847f53712 (Detached HEAD (experimental), 2007-01-01) an error
condition was introduced in rename_branch() to prevent renaming, later
also copying, a detached HEAD.
The condition used was checking for NULL in oldname, the source branch
to rename/copy. That condition cannot be satisfied because if no source
branch is specified, HEAD is going to be used in the call.
The error issued instead is:
fatal: Invalid branch name: 'HEAD'
Let's remove the condition in copy_or_rename_branch() (the current
function name) and check for HEAD before calling it, dying with the
original intended error if we're in a detached HEAD.
Signed-off-by: Rubén Justo <rjusto@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
mark_common() in negotiator/skipping.c may overflow the stack due to
recursive function calls. Avoid this by instead recursing using a
heap-allocated data structure.
Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
|
|
During the initial development of the fsck-msgids.txt feature, it
has become apparent that it is very much error prone to make sure
the description in the documentation file are sorted and correctly
match what is in the fsck.h header file.
Add a quick-and-dirty Perl script and doc-lint target to sanity
check that the fsck-msgids.txt is consistent with the error type
list in the fsck.h header file.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
The documentation lacks mention of specific <msg-id> that are supported.
While git-help --config will display a list of these options, often
developers' first instinct is to consult the git docs to find valid
config values.
Add a list of fsck error messages, and link to it from the git-fsck
documentation.
Signed-off-by: John Cai <johncai86@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
This error type has never been used since it was introduced at
159e7b08 (fsck: detect gitmodules files, 2018-05-02).
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
2175a0c6 (fsck: stop checking tag->tagged, 2019-10-18) stopped
checking the tagged object referred to by a tag object, which is what the
error message BAD_TAG_OBJECT was for. Since then the BAD_TAG_OBJECT
message is no longer used anywhere.
Remove the BAD_TAG_OBJECT msg-id.
Signed-off-by: John Cai <johncai86@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
The apply code is not prepared to handle extremely large files. It uses
"int" in some places, and "unsigned long" in others.
This combination leads to unfortunate problems when switching between
the two types. Using "int" prevents us from handling large files, since
large offsets will wrap around and spill into small negative values,
which can result in wrong behavior (like accessing the patch buffer with
a negative offset).
Converting from "unsigned long" to "int" also has truncation problems
even on LLP64 platforms where "long" is the same size as "int", since
the former is unsigned but the latter is not.
To avoid potential overflow and truncation issues in `git apply`, apply
similar treatment as in dcd1742e56 (xdiff: reject files larger than
~1GB, 2015-09-24), where the xdiff code was taught to reject large
files for similar reasons.
The maximum size was chosen somewhat arbitrarily, but picking a value
just shy of a gigabyte allows us to double it without overflowing 2^31-1
(after which point our value would wrap around to a negative number).
To give ourselves a bit of extra margin, the maximum patch size is a MiB
smaller than a full GiB, which gives us some slop in case we allocate
"(records + 1) * sizeof(int)" or similar.
Luckily, the security implications of these conversion issues are
relatively uninteresting, because a victim needs to be convinced to
apply a malicious patch.
Reported-by: 정재우 <thebound7@gmail.com>
Suggested-by: Johannes Schindelin <Johannes.Schindelin@gmx.de>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
With the recent turnover on the git-security list, questions came up how
things are usually run. Rather than answering questions individually,
extend Git's existing documentation about security vulnerabilities to
describe the git-security mailing list, how things are run on that list,
and what to expect throughout the process from the time a security bug
is reported all the way to the time when a fix is released.
Helped-by: Junio C Hamano <gitster@pobox.com>
Helped-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Julia Ramer <gitprplr@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
The last git version that had "diff-tree" in the header text
of "git diff-tree" output was v1.3.0 from 2006. The header text
was changed from "diff-tree" to "commit" in 91539833
("Log message printout cleanups").
Given how long ago this change was made, it is highly unlikely that
anyone is still feeding in outputs from that git version.
Remove the handling of the "diff-tree" prefix and document the
source of the other prefixes so that the overall functionality
is more clear.
Signed-off-by: Jerry Zhang <Jerry@skydio.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
There are situations where the user might not want the default
setting where patch-id strips all whitespace. They might be working
in a language where white space is syntactically important, or they
might have CI testing that enforces strict whitespace linting. In
these cases, a whitespace change would result in the patch
fundamentally changing, and thus deserving of a different id.
Add a new mode that is exclusive of --stable and --unstable called
--verbatim. It also corresponds to the config
patchid.verbatim = true. In this mode, the stable algorithm is
used and whitespace is not stripped from the patch text.
Users of --unstable mainly care about compatibility with old git
versions, which unstripping the whitespace would break. Thus there
isn't a usecase for the combination of --verbatim and --unstable,
and we don't expose this so as to not add maintainence burden.
Signed-off-by: Jerry Zhang <jerry@skydio.com>
fixes https://github.com/Skydio/revup/issues/2
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Currently patch-id as used in rebase and cherry-pick does not account
for file modes if the file is modified. One consequence of this is
that if you have a local patch that changes modes, but upstream
has applied an outdated version of the patch that doesn't include
that mode change, "git rebase" will drop your local version of the
patch along with your mode changes. It also means that internal
patch-id doesn't produce the same output as the builtin, which does
account for mode changes due to them being part of diff output.
Fix by adding mode to the patch-id if it has changed, in the same
format that would be produced by diff, so that it is compatible
with builtin patch-id.
Signed-off-by: Jerry Zhang <Jerry@skydio.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
"git patch-id" currently doesn't produce correct output if the
incoming diff has any binary files. Add logic to get_one_patchid
to handle the different possible styles of binary diff. This
attempts to keep resulting patch-ids identical to what would be
produced by the counterpart logic in diff.c, that is it produces
the id by hashing the a and b oids in succession.
In general we handle binary diffs by first caching the object ids from
the "index" line and using those if we then find an indication
that the diff is binary.
The input could contain patches generated with "git diff --binary". This
currently breaks the parse logic and results in multiple patch-ids
output for a single commit. Here we have to skip the contents of the
patch itself since those do not go into the patch id. --binary
implies --full-index so the object ids are always available.
When the diff is generated with --full-index there is no patch content
to skip over.
When a diff is generated without --full-index or --binary, it will
contain abbreviated object ids. This will still result in a sufficiently
unique patch-id when hashed, but does not match internal patch id
output. We'll call this ok for now as we already need specialized
arguments to diff in order to match internal patch id (namely -U3).
Signed-off-by: Jerry Zhang <Jerry@skydio.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Git doesn't persist patch-ids during the rebase process, so there is
no need to specifically invoke the unstable variant. Use the stable
logic for all internal patch-id calculations to minimize the number of
code paths and improve test coverage.
Signed-off-by: Jerry Zhang <jerry@skydio.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Patch-ids for binary patches are found by hashing the object
ids of the before and after objects in succession. However in
the --stable case, there is a bug where hunks are not flushed
for binary and header-only patch ids, which would always result
in a patch-id of 0000. The --unstable case is currently correct.
Reorder the logic to branch into 3 cases for populating the
patch body: header-only which populates nothing, binary which
populates the object ids, and normal which populates the text
diff. All branches will end up flushing the hunk.
Don't populate the ---a/ and +++b/ lines for binary diffs, to correspond
to those lines not being present in the "git diff" text output.
This is necessary because we advertise that the patch-id calculated
internally and used in format-patch is the same that what the
builtin "git patch-id" would produce when piped from a diff.
Update the test to run on both binary and normal files.
Signed-off-by: Jerry Zhang <jerry@skydio.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
In the same spirit as the previous commit, reimplement
`--group=committer` as a special case of `--group=<format>`, too.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Instead of handling SHORTLOG_GROUP_AUTHOR separately, reimplement it as
a special case of the new `--group=<format>` mode, where the author mode
is a shorthand for `--group='%aN <%aE>'.
Note that we still need to keep the SHORTLOG_GROUP_AUTHOR enum since it
has a different meaning in `read_from_stdin()`, where it is still used
for a different purpose.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Extract a function which finishes setting up the shortlog struct for
use. The caller in `make_cover_letter()` does not care about trailer
sorting, so it isn't strictly necessary to add a call there in this
patch.
But the next patch will add additional functionality to the new
`shortlog_finish_setup()` function, which the caller in
`make_cover_letter()` will care about.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
In addition to generating a shortlog based on committer, author, or the
identity in one or more specified trailers, it can be useful to generate
a shortlog based on an arbitrary commit format.
This can be used, for example, to generate a distribution of commit
activity over time, like so:
$ git shortlog --group='%cd' --date='format:%Y-%m' -s v2.37.0..
117 2022-06
274 2022-07
324 2022-08
263 2022-09
7 2022-10
Arbitrary commit formats can be used. In fact, `git shortlog`'s default
behavior (to count by commit authors) can be emulated as follows:
$ git shortlog --group='%aN <%aE>' ...
and future patches will make the default behavior (as well as
`--committer`, and `--group=trailer:<trailer>`) special cases of the
more flexible `--group` option.
Note also that the SHORTLOG_GROUP_FORMAT enum value is used only to
designate that `--group:<format>` is in use when in stdin mode to
declare that the combination is invalid.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
The subsequent commit will add another unhandled case in
`read_from_stdin()` which will want to use the same message as with
`--group=trailer`.
Extract the "--group=trailer" part from this message so the same
translation key can be used for both cases.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
When there are no trailers to insert, it is natural that
insert_records_from_trailers() should return without having done any
work.
But instead we guard this call unnecessarily by first checking whether
`log->groups` has the `SHORTLOG_GROUP_TRAILER` bit set.
Prepare to match a similar pattern in the future where a function which
inserts records of a certain type does no work when no specifiers
matching that type are given.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Prepare for a future patch which will introduce arbitrary pretty formats
via the `--group` argument.
To allow additional customizability (for example, to support something
like `git shortlog -s --group='%aD' --date='format:%Y-%m' ...` (which
groups commits by the datestring 'YYYY-mm' according to author date), we
must store off the `--date` parsed from calling `parse_revision_opt()`.
Note that this also affects custom output `--format` strings in `git
shortlog`. Though this is a behavior change, this is arguably fixing a
long-standing bug (ie., that `--format` strings are not affected by
`--date` specifiers as they should be).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Add global counters mechanism to Trace2.
The Trace2 counters mechanism adds the ability to create a set of
global counter variables and an API to increment them efficiently.
Counters can optionally report per-thread usage in addition to the sum
across all threads.
Counter events are emitted to the Trace2 logs when a thread exits and
at process exit.
Counters are an alternative to `data` and `data_json` events.
Counters are useful when you want to measure something across the life
of the process, when you don't want per-measurement events for
performance reasons, when the data does not fit conveniently within a
region, or when your control flow does not easily let you write the
final total. For example, you might use this to report the number of
calls to unzip() or the number of de-delta steps during a checkout.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Add stopwatch timer mechanism to Trace2.
Timers are an alternative to Trace2 Regions. Regions are useful for
measuring the time spent in various computation phases, such as the
time to read the index, time to scan for unstaged files, time to scan
for untracked files, and etc.
However, regions are not appropriate in all places. For example,
during a checkout, it would be very inefficient to use regions to
measure the total time spent inflating objects from the ODB from
across the entire lifetime of the process; a per-unzip() region would
flood the output and significantly slow the command; and some form of
post-processing would be requried to compute the time spent in unzip().
Timers can be used to measure a series of timer intervals and emit
a single summary event (at thread and/or process exit).
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Convert the `tr2tls_thread_ctx.thread_name` field from a `strbuf`
to a "const char*" pointer.
The `thread_name` field is a constant string that is constructed when
the context is created. Using a (non-const) `strbuf` structure for it
caused some confusion in the past because it implied that someone
could rename a thread after it was created. That usage was not
intended. Change it to a const pointer to make the intent more clear.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Improve the documentation of the tr2tls_thread_ctx.thread_name field
and its relation to the tr2tls_thread_ctx.thread_id field.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Rename the `thread_name` argument in `tr2tls_create_self()` and
`trace2_thread_start()` to be `thread_base_name` to make it clearer
that the passed argument is a component used in the construction of
the actual `struct tr2tls_thread_ctx.thread_name` variable.
The base name will be used along with the thread id to create a
unique thread name.
This commit does not change how the `thread_name` field is
allocated or stored within the `tr2tls_thread_ctx` structure.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Eliminate the mostly obsolete `Public API` sub-section from the
`Trace2 API` section in the documentation. Strengthen the referral
to `trace2.h`.
Most of the technical information in this sub-section was moved to
`trace2.h` in 6c51cb525d (trace2: move doc to trace2.h, 2019-11-17) to
be adjacent to the function prototypes. The remaining text wasn't
that useful by itself.
Furthermore, the text would need a bit of overhaul to add routines
that do not immediately generate a message, such as stopwatch timers.
So it seemed simpler to just get rid of it.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Reduce or eliminate use of the term "TLS" in the Trace2 code.
The term "TLS" has two popular meanings: "thread-local storage" and
"transport layer security". In the Trace2 source, the term is associated
with the former. There was concern on the mailing list about it refering
to the latter.
Update the source and documentation to eliminate the use of the "TLS" term
or replace it with the phrase "thread-local storage" to reduce ambiguity.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Use "size_t" rather than "int" for the "alloc" and "nr_open_regions"
fields in the "tr2tls_thread_ctx". These are used by ALLOC_GROW().
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
absorb_git_dir_into_superproject() uses a strbuf and strvec_pushl() to
build and add the --super-prefix option and its argument. Use a single
strvec_pushf() call to add the stuck form instead, which reduces the
code size and avoids a strbuf allocation and release. The same is
already done in submodule_reset_index() and submodule_move_head().
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
One of our tests intentionally causes the cruft-pack generation phase of
repack to fail, in order to stimulate an exit from repack at the desired
moment. It does so by feeding a bogus option argument to pack-objects.
This is a simple and reliable way to get pack-objects to fail, but it
has one downside: pack-objects will die before reading its stdin, which
means the caller repack may racily get SIGPIPE writing to it.
For the purposes of this test, that's OK. We are checking whether repack
cleans up already-created .tmp files, and it will do so whether it exits
or dies by signal (because the tempfile API hooks both).
But we have to tell test_must_fail that either outcome is OK, or it
complains about the signal. Arguably this is a workaround (compared to
fixing repack), as repack dying to SIGPIPE means that it loses the
opportunity to give a more detailed message. But we don't actually write
such a message anyway; we rely on pack-objects to have written something
useful to stderr, and it does. In either case (signal or exit), that is
the main thing the user will see.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
Add an option, --stdin, to merge-tree which will accept lines of input
with two branches to merge per line, and which will perform all the
merges and give output for each in turn. This option implies -z, and
modifies the output to also include a merge status since the exit code
of the program can no longer convey that information now that multiple
merges are involved.
This could be useful, for example, by Git hosting providers. When one
branch is updated, one may want to check whether all code reviews
targetting that branch can still cleanly merge. Avoiding the overhead
of starting up a separate process for each of those code reviews might
provide significant savings in a repository with many code reviews.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
The Informational Messages was updated in de90581141 ("merge-ort:
optionally produce machine-readable output", 2022-06-18) to provide more
detailed and machine parseable output when `-z` is passed, but the
Documentation was not updated to reflect these changes. Update it now.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
When initializing a repository object, we run "git rev-parse --git-dir"
to let the C version of Git find the correct directory. But curiously,
if this fails we don't automatically say "not a git repository".
Instead, we do our own pure-perl check to see if we're in a bare
repository.
This makes little sense, as rev-parse will report both bare and non-bare
directories. This logic comes from d5c7721d58 (Git.pm: Add support for
subdirectories inside of working copies, 2006-06-24), but I don't see
any reason given why we can't just rely on rev-parse. Worse, because we
treat any non-error response from rev-parse as a non-bare repository,
we'll erroneously set the object's WorkingCopy, even in a bare
repository.
But it gets worse. Since 8959555cee (setup_git_directory(): add an owner
check for the top-level directory, 2022-03-02), it's actively wrong (and
dangerous). The perl code doesn't implement the same ownership checks.
And worse, after "finding" the bare repository, it sets GIT_DIR in the
environment, which tells any subsequent Git commands that we've
confirmed the directory is OK, and to trust us. I.e., it re-opens the
vulnerability plugged by 8959555cee when using Git.pm's repository
discovery code.
We can fix this by just relying on rev-parse to tell us when we're not
in a repository, which fixes the vulnerability. Furthermore, we'll ask
its --is-bare-repository function to tell us if we're bare or not, and
rely on that.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
When changing a directory to a symlink on one side of history, and
renaming the parent of that directory to a different directory name
on the other side, e.g. with this kind of setup:
Base commit: Has a file named dir/subdir/file
Side1: Rename dir/ -> renamed-dir/
Side2: delete dir/subdir/file, add dir/subdir as symlink
Then merge-ort was running into an assertion failure:
git: merge-ort.c:2622: apply_directory_rename_modifications: Assertion `ci->dirmask == 0' failed
merge-recursive did not have as obvious an issue handling this case,
likely because we never fixed it to handle the case from commit
902c521a35 ("t6423: more involved directory rename test", 2020-10-15)
where we need to be careful about nested renames when a directory rename
occurs (dir/ -> renamed-dir/ implies dir/subdir/ ->
renamed-dir/subdir/). However, merge-recursive does have multiple
problems with this testcase:
* Incorrect stages for the file: merge-recursive omits the stage in
the index corresponding to the base stage, making `git status`
report "added by us" for renamed-dir/subdir/file instead of the
expected "deleted by them".
* Poor directory/file conflict handling: For the renamed-dir/subdir
symlink, instead of reporting a file/directory conflict as
expected, it reports "Error: Refusing to lose untracked file at
renamed-dir/subdir". This is a lie because there is no untracked
file at that location. It then does the normal suboptimal
merge-recursive thing of having the symlink be tracked in the index
at a location where it can't be written due to D/F conflicts
(namely, renamed-dir/subdir), but writes it to the working tree at
a different location as a new untracked file (namely,
renamed-dir/subdir~B^0)
Technically, these problems don't prevent the user from resolving the
merge if they can figure out to ignore the confusion, but because both
pieces of output are quite confusing I don't want to modify the test
to claim the recursive also passes it even if it doesn't have the bug
that ort did.
So, fix the bug in ort by splitting the conflict_info for "dir/subdir"
into two, one for the directory part, one for the file (i.e. symlink)
part, since the symlink is being renamed by directory rename detection.
The directory part is needed for proper nesting, since there are still
conflict_info fields for files underneath it (though those are marked
as is_null, they are still present until the entries are processed,
and the entry processing wants every non-toplevel entry to have a
parent directory).
Reported-by: Stefano Rivera <stefano@rivera.za.net>
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
After we've successfully finished the repack, we call
remove_temporary_files(), which looks for and removes any files matching
".tmp-$$-pack-*", where $$ is the pid of the current process. But this
is pointless. If we make it this far in the process, we've already
renamed these tempfiles into place, and there is nothing left to delete.
Nor is there a point in trying to call it to clean up when we _aren't_
successful. It's not safe for using in a signal handler, and the
previous commit already handed that job over to the tempfile API.
It might seem like it would be useful to clean up stray .tmp files left
by other invocations of git-repack. But it won't clean those files; it
only matches ones with its pid, and leaves the rest. Fortunately, those
are cleaned up naturally by successive calls to git-repack; we'll
consider .tmp-*.pack the same as normal packfiles, so "repack -ad", etc,
will roll up their contents and eventually delete them.
The one case that could matter is if pack-objects generates an extension
we don't know about, like ".tmp-pack-$$-$hash.some-new-ext". The current
code will quietly delete such a file, while after this patch we'd leave
it in place. In practice this doesn't happen, and would be indicative of
a bug. Leaving the file as cruft is arguably a better behavior, as it
means somebody is more likely to eventually notice and fix the bug. If
we really wanted to be paranoid, we could scan for and warn about such
files, but that seems like overkill.
There's nothing to test with regard to the removal of this function. It
was doing nothing, so the behavior should be the same. However, we can
verify (and protect) our assumption that "repack -ad" will eventually
remove stray files by adding a test for that.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
When git-repack exits due to a signal, it tries to clean up by calling
its remove_temporary_files() function, which walks through the packs dir
looking for ".tmp-$$-pack-*" files to delete (where "$$" is the pid of
the current process).
The biggest problem here is that remove_temporary_files() is not safe to
call in a signal handler. It uses opendir(), which isn't on the POSIX
async-signal-safe list. The details will be platform-specific, but a
likely issue is that it needs to allocate memory; if we receive a signal
while inside malloc(), etc, we'll conflict on the allocator lock and
deadlock with ourselves.
We can fix this by just cleaning up the files directly, without walking
the directory. We already know the complete list of .tmp-* files that
were generated, because we recorded them via populate_pack_exts(). When
we find files there, we can use register_tempfile() to record the
filenames. If we receive a signal, then the tempfile API will clean them
up for us, and it's async-safe and pretty battle-tested.
Note that this is slightly racier than the existing scheme. We don't
record the filenames until pack-objects tells us the hash over stdout.
So during the period between it generating the file and reporting the
hash, we'd fail to clean up. However, that period is very small. During
most of the pack generation process pack-objects is using its own
internal tempfiles. It's only at the very end that it moves them into
the names git-repack expects, and then it immediately reports the name
to us. Given that cleanup like this is best effort (after all, we may
get SIGKILL), this level of race is acceptable.
When we register the tempfiles, we'll record them locally and use the
result to call rename_tempfile(), rather than renaming by hand. This
isn't strictly necessary, as once we've renamed the files they're gone,
and the tempfile API's cleanup unlink() would simply become a pointless
noop. But managing the lifetimes of the tempfile objects is the cleanest
thing to do, and the tempfile pointers naturally fill the same role as
the old booleans.
This patch also fixes another small problem. We only hook signals, and
don't set up an atexit handler. So if we see an error that causes us to
die(), we'll leave the .tmp-* files in place. But since the tempfile API
handles this for us, this is now fixed for free. The new test covers
this by stimulating a failure of pack-objects when generating a cruft
pack. Before this patch, the .tmp-* file for the main pack would have
been left, but now we correctly clean it up.
Two small subtleties on the implementation:
- in the renaming loop, we can stop re-constructing fname_old; we only
use it when we have a tempfile to rename, so we can just ask the
tempfile for its path (which, barring bugs, should be identical)
- when renaming fails, our error message mentions fname_old. But since
a failed rename_tempfile() invalidates the tempfile struct, we'll
lose access to that string. Instead, let's mention the destination
filename, which is what most other callers do.
Reported-by: Jan Pokorný <poki@fnusa.cz>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
If pack-objects tells us it generated pack $hash, we expect to find
.tmp-$$-pack-$hash.pack, .idx, .rev, and so on. Some of these files are
optional, but others are not. For the required ones, we'll bail with an
error if any of them is missing.
The error message is just "missing required file", which is a bit vague.
We should be more clear that it is not the user's fault, but rather that
the sub-pgoram we called is not operating as expected. In practice,
nobody should ever see this message, as it would generally only be
caused by a bug in Git.
It probably doesn't make sense to convert this to a BUG(), though, as
there are other (unlikely) possibilities, such as somebody else racily
deleting the files, filesystem errors causing stat() to fail, and so on.
A nice side effect here is that we stop relying on fname_old in this
code path, which will let us deal with it only in the first part of the
conditional.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
After generating the main pack and then any additional cruft packs, we
iterate over the "names" list (which contains hashes of packs generated
by pack-objects), and call populate_pack_exts() for each.
There's one small problem with this. In repack_promisor_objects(), we
may add entries to "names" and call populate_pack_exts() for them.
Calling it again is mostly just wasteful, as we'll stat() the filename
with each possible extension, get the same result, and just overwrite
our bits.
So we could drop the call there, and leave the final loop to populate
all of the bits. But instead, this patch does the reverse: drops the
final loop, and teaches the other two sites to populate the bits as they
add entries.
This makes the code easier to reason about, as you never have to worry
about when the util field is valid; it is always valid for each entry.
It also serves my ulterior purpose: recording the generated filenames as
soon as possible will make it easier for a future patch to use them for
cleaning up from a failed operation.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
We keep a string_list "names" containing the hashes of packs generated
on our behalf by pack-objects. The util field of each item is treated as
a bitfield that tells us which extensions (.pack, .idx, .rev, etc) are
present for each name.
Let's switch this to allocating a real array. That will give us room in
a future patch to store more data than just a single bit per extension.
And it makes the code a little easier to read, as we avoid casting back
and forth between uintptr_t and a void pointer.
Since the only thing we're storing is an array, we could just allocate
it directly. But instead I've put it into a named struct here. That
further increases readability around the casts, and in particular helps
differentiate us from other string_lists in the same file which use
their util field differently. E.g., the existing_*_packs lists still do
bit-twiddling, but their bits have different meaning than the ones in
"names". This makes it hard to grep around the code to see how the util
fields are used; now you can look for "generated_pack_data".
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
The previous step made an attempt to correctly compute display
columns allocated and padded different parts of diffstat output.
There are at least two known codepaths in the function that still
mixes up display widths and byte length that need to be fixed.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
The previous commit fixed a failure in 'git subtree merge --squash' when
the previous squash-merge merged an annotated tag of the subtree
repository which is missing locally.
The same failure happens in 'git subtree split', either directly or when
called by 'git subtree push', under the same circumstances: 'cmd_split'
invokes 'find_existing_splits', which loops through previous commits and
invokes 'git rev-parse' (via 'process_subtree_split_trailer') on the
value of any 'git subtree-split' trailer it finds. This fails if this
value is the hash of an annotated tag which is missing locally.
Add a new optional argument 'repository' to 'cmd_split' and
'find_existing_splits', and invoke 'cmd_split' with that argument from
'cmd_push'. This allows 'process_subtree_split_trailer' to try to fetch
the missing tag from the 'repository' if it's not available locally,
mirroring the new behaviour of 'git subtree pull' and 'git subtree
merge'.
Signed-off-by: Philippe Blain <levraiphilippeblain@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
|
When 'git subtree merge --squash $ref' is invoked, either directly or
through 'git subtree pull --squash $repo $ref', the code looks for the
latest squash merge of the subtree in order to create the new merge
commit as a child of the previous squash merge.
This search is done in function 'process_subtree_split_trailer', invoked
by 'find_latest_squash', which looks for the most recent commit with a
'git-subtree-split' trailer; that trailer's value is the object name in
the subtree repository of the ref that was last squash-merged. The
function verifies that this object is present locally with 'git
rev-parse', and aborts if it's not.
The hash referenced by the 'git-subtree-split' trailer is guaranteed to
correspond to a commit since it is the result of running 'git rev-parse
-q --verify "$1^{commit}"' on the first argument of 'cmd_merge' (this
corresponds to 'rev' in 'cmd_merge' which is passed through to
'new_squash_commit' and 'squash_msg').
But this is only the case since e4f8baa88a (subtree: parse revs in
individual cmd_ functions, 2021-04-27), which went into Git 2.32. Before
that commit, 'cmd_merge' verified the revision it was given using 'git
rev-parse --revs-only "$@"'. Such an invocation, when fed the name of an
annotated tag, would return the hash of the tag, not of the commit
referenced by the tag.
This leads to a failure in 'find_latest_squash' when squash-merging if
the most recent squash-merge merged an annotated tag of the subtree
repository, using a pre-2.32 version of 'git subtree', unless that
previous annotated tag is present locally (which is not usually the
case).
We can fix this by fetching the object directly by its hash in
'process_subtree_split_trailer' when 'git rev-parse' fails, but in order
to do so we need to know the name or URL of the subtree repository.
This is not possible in general for 'git subtree merge', but is easy
when it is invoked through 'git subtree pull' since in that case the
subtree repository is passed by the user at the command line.
Allow the 'git subtree pull' scenario to work out-of-the-box by adding
an optional 'repository' argument to functions 'cmd_merge',
'find_latest_squash' and 'process_subtree_split_trailer', and invoke
'cmd_merge' with that 'repository' argument in 'cmd_pull'.
If 'repository' is absent in 'process_subtree_split_trailer', instruct
the user to try fetching the missing object directly.
Signed-off-by: Philippe Blain <levraiphilippeblain@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
|
