Phillip Wood [Tue, 4 Oct 2022 10:01:34 +0000 (10:01 +0000)]
ssh signing: return an error when signature cannot be read
If the signature file cannot be read we print an error message but do
not return an error to the caller. In practice it seems unlikely that
the file would be unreadable if the call to ssh-keygen succeeds.
The unlink_or_warn() call is moved to the end of the function so that
we always try and remove the signature file. This isn't strictly
necessary at the moment but it protects us against any extra code
being added between trying to read the signature file and the cleanup
at the end of the function in the future. unlink_or_warn() only prints
a warning if it exists and cannot be removed.
Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Jeff King [Mon, 3 Oct 2022 17:35:02 +0000 (13:35 -0400)]
sequencer: detect author name errors in read_author_script()
As we parse the author-script file, we check for missing or duplicate
lines for GIT_AUTHOR_NAME, etc. But after reading the whole file, our
final error conditional checks "date_i" twice and "name_i" not at all.
This not only leads to us failing to abort, but we may do an
out-of-bounds read on the string_list array.
The bug goes back to 442c36bd08 (am: improve author-script error
reporting, 2018-10-31), though the code was soon after moved to this
spot by bcd33ec25f (add read_author_script() to libgit, 2018-10-31).
It was presumably just a typo in 442c36bd08.
We'll add test coverage for all the error cases here, though only the
GIT_AUTHOR_NAME ones fail (even in a vanilla build they segfault
consistently, but certainly with SANITIZE=address).
Reported-by: Michael V. Scovetta <michael.scovetta@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Phillip Wood [Mon, 3 Oct 2022 09:23:30 +0000 (09:23 +0000)]
mailinfo -b: fix an out of bounds access
To remove bracketed strings containing "PATCH" from the subject line
cleanup_subject() scans the subject for the opening bracket using an
offset from the beginning of the line. It then searches for the
closing bracket with strchr(). To calculate the length of the
bracketed string it unfortunately adds rather than subtracts the
offset from the result of strchr(). This leads to an out of bounds
access in memmem() when looking to see if the brackets contain
"PATCH".
We have tests that trigger this bug that were added in ae52d57f0b
(t5100: add some more mailinfo tests, 2017-05-31). The commit message
mentions that they are marked test_expect_failure as they trigger an
assertion in strbuf_splice(). While it is reassuring that
strbuf_splice() detects the problem and dies in retrospect that should
perhaps have warranted a little more investigation. The bug was
introduced by 17635fc900 (mailinfo: -b option keeps [bracketed]
strings that is not a [PATCH] marker, 2009-07-15). I think the reason
it has survived so long is that '-b' is not a popular option and
without it the offset is always zero.
This was found by the address sanitizer while I was cleaning up the
test_todo idea in [1].
René Scharfe [Sat, 1 Oct 2022 10:28:07 +0000 (12:28 +0200)]
diff: support ^! for merges
revision.c::handle_revision_arg_1() resolves <rev>^! by first adding the
negated parents and then <rev> itself. builtin_diff_combined() expects
the first tree to be the merge and the remaining ones to be the parents,
though. This mismatch results in bogus diff output.
Remember the first tree that doesn't belong to a parent and use it
instead of blindly picking the first one. This makes "git diff <rev>^!"
consistent with "git show <rev>^!".
Reported-by: Tim Jaacks <tim.jaacks@garz-fricke.com> Suggested-by: Junio C Hamano <gitster@pobox.com> Signed-off-by: René Scharfe <l.s.r@web.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
René Scharfe [Sat, 1 Oct 2022 10:26:34 +0000 (12:26 +0200)]
revisions.txt: unspecify order of resolved parts of ^!
gitrevisions(7) says that <rev>^! resolves to <rev> and then all the
parents of <rev>. revision.c::handle_revision_arg_1() actually adds
all parents first, then <rev>. Change the documentation to leave the
order unspecified, to avoid misleading readers.
Signed-off-by: René Scharfe <l.s.r@web.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Taylor Blau [Fri, 30 Sep 2022 20:47:00 +0000 (16:47 -0400)]
t7814: prepare for changing protocol.file.allow
Explicitly cloning over the "file://" protocol in t7814 in preparation
for merging a security release which will change the default value of
this configuration to be "user".
Taylor Blau [Fri, 30 Sep 2022 20:47:00 +0000 (16:47 -0400)]
t5537: prepare for changing protocol.file.allow
Explicitly cloning over the "file://" protocol in t5537 in preparation
for merging a security release which will change the default value of
this configuration to be "user".
Taylor Blau [Fri, 30 Sep 2022 20:47:00 +0000 (16:47 -0400)]
t5516: prepare for changing protocol.file.allow
Explicitly cloning over the "file://" protocol in t5516 in preparation
for merging a security release which will change the default value of
this configuration to be "user".
Taylor Blau [Fri, 30 Sep 2022 20:47:00 +0000 (16:47 -0400)]
t3207: prepare for changing protocol.file.allow
Explicitly cloning over the "file://" protocol in t3207 in preparation
for merging a security release which will change the default value of
this configuration to be "user".
Taylor Blau [Fri, 30 Sep 2022 20:48:56 +0000 (16:48 -0400)]
t2080: prepare for changing protocol.file.allow
Explicitly cloning over the "file://" protocol in t1092 in preparation
for merging a security release which will change the default value of
this configuration to be "user".
Taylor Blau [Fri, 30 Sep 2022 20:47:00 +0000 (16:47 -0400)]
t1092: prepare for changing protocol.file.allow
Explicitly cloning over the "file://" protocol in t1092 in preparation
for merging a security release which will change the default value of
this configuration to be "user".
Taylor Blau [Fri, 30 Sep 2022 20:48:56 +0000 (16:48 -0400)]
t2080: prepare for changing protocol.file.allow
Explicitly cloning over the "file://" protocol in t1092 in preparation
for merging a security release which will change the default value of
this configuration to be "user".
Taylor Blau [Fri, 30 Sep 2022 20:47:00 +0000 (16:47 -0400)]
t1092: prepare for changing protocol.file.allow
Explicitly cloning over the "file://" protocol in t1092 in preparation
for merging a security release which will change the default value of
this configuration to be "user".
John Cai [Fri, 30 Sep 2022 20:47:11 +0000 (20:47 +0000)]
tmp-objdir: skip clean up when handling a signal
In the tmp-objdir api, tmp_objdir_create will create a temporary
directory but also register signal handlers responsible for removing
the directory's contents and the directory itself. However, the
function responsible for recursively removing the contents and
directory, remove_dir_recurse() calls opendir(3) and closedir(3).
This can be problematic because these functions allocate and free
memory, which are not async-signal-safe functions. This can lead to
deadlocks.
One place we call tmp_objdir_create() is in git-receive-pack, where
we create a temporary quarantine directory "incoming". Incoming
objects will be written to this directory before they get moved to
the object directory.
We have observed this code leading to a deadlock:
Thread 1 (Thread 0x7f621ba0b200 (LWP 326305)):
#0 __lll_lock_wait_private (futex=futex@entry=0x7f621bbf8b80
<main_arena>) at ./lowlevellock.c:35
#1 0x00007f621baa635b in __GI___libc_malloc
(bytes=bytes@entry=32816) at malloc.c:3064
#2 0x00007f621bae9f49 in __alloc_dir (statp=0x7fff2ea7ed60,
flags=0, close_fd=true, fd=5)
at ../sysdeps/posix/opendir.c:118
#3 opendir_tail (fd=5) at ../sysdeps/posix/opendir.c:69
#4 __opendir (name=<optimized out>)
at ../sysdeps/posix/opendir.c:92
#5 0x0000557c19c77de1 in remove_dir_recurse ()
git#6 0x0000557c19d81a4f in remove_tmp_objdir_on_signal ()
#7 <signal handler called>
git#8 _int_malloc (av=av@entry=0x7f621bbf8b80 <main_arena>,
bytes=bytes@entry=7160) at malloc.c:4116
git#9 0x00007f621baa62c9 in __GI___libc_malloc (bytes=7160)
at malloc.c:3066
git#10 0x00007f621bd1e987 in inflateInit2_ ()
from /opt/gitlab/embedded/lib/libz.so.1
git#11 0x0000557c19dbe5f4 in git_inflate_init ()
git#12 0x0000557c19cee02a in unpack_compressed_entry ()
git#13 0x0000557c19cf08cb in unpack_entry ()
git#14 0x0000557c19cf0f32 in packed_object_info ()
git#15 0x0000557c19cd68cd in do_oid_object_info_extended ()
git#16 0x0000557c19cd6e2b in read_object_file_extended ()
git#17 0x0000557c19cdec2f in parse_object ()
git#18 0x0000557c19c34977 in lookup_commit_reference_gently ()
git#19 0x0000557c19d69309 in mark_uninteresting ()
git#20 0x0000557c19d2d180 in do_for_each_repo_ref_iterator ()
git#21 0x0000557c19d21678 in for_each_ref ()
git#22 0x0000557c19d6a94f in assign_shallow_commits_to_refs ()
git#23 0x0000557c19bc02b2 in cmd_receive_pack ()
git#24 0x0000557c19b29fdd in handle_builtin ()
git#25 0x0000557c19b2a526 in cmd_main ()
git#26 0x0000557c19b28ea2 in main ()
Since we can't do the cleanup in a portable and signal-safe way, skip
the cleanup when we're handling a signal.
This means that when signal handling, the temporary directory may not
get cleaned up properly. This is mitigated by b3cecf49ea (tmp-objdir: new
API for creating temporary writable databases, 2021-12-06) which changed
the default name and allows gc to clean up these temporary directories.
In the event of a normal exit, we should still be cleaning up via the
atexit() handler.
Helped-by: Jeff King <peff@peff.net> Signed-off-by: John Cai <johncai86@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Kevin Backhouse [Wed, 28 Sep 2022 22:53:32 +0000 (18:53 -0400)]
alias.c: reject too-long cmdline strings in split_cmdline()
This function improperly uses an int to represent the number of entries
in the resulting argument array. This allows a malicious actor to
intentionally overflow the return value, leading to arbitrary heap
writes.
Because the resulting argv array is typically passed to execv(), it may
be possible to leverage this attack to gain remote code execution on a
victim machine. This was almost certainly the case for certain
configurations of git-shell until the previous commit limited the size
of input it would accept. Other calls to split_cmdline() are typically
limited by the size of argv the OS is willing to hand us, so are
similarly protected.
So this is not strictly fixing a known vulnerability, but is a hardening
of the function that is worth doing to protect against possible unknown
vulnerabilities.
One approach to fixing this would be modifying the signature of
`split_cmdline()` to look something like:
int split_cmdline(char *cmdline, const char ***argv, size_t *argc);
Where the return value of `split_cmdline()` is negative for errors, and
zero otherwise. If non-NULL, the `*argc` pointer is modified to contain
the size of the `**argv` array.
But this implies an absurdly large `argv` array, which more than likely
larger than the system's argument limit. So even if split_cmdline()
allowed this, it would fail immediately afterwards when we called
execv(). So instead of converting all of `split_cmdline()`'s callers to
work with `size_t` types in this patch, instead pursue the minimal fix
here to prevent ever returning an array with more than INT_MAX entries
in it.
Signed-off-by: Kevin Backhouse <kevinbackhouse@github.com> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com>
Jeff King [Wed, 28 Sep 2022 22:52:48 +0000 (18:52 -0400)]
shell: limit size of interactive commands
When git-shell is run in interactive mode (which must be enabled by
creating $HOME/git-shell-commands), it reads commands from stdin, one
per line, and executes them.
We read the commands with git_read_line_interactively(), which uses a
strbuf under the hood. That means we'll accept an input of arbitrary
size (limited only by how much heap we can allocate). That creates two
problems:
- the rest of the code is not prepared to handle large inputs. The
most serious issue here is that split_cmdline() uses "int" for most
of its types, which can lead to integer overflow and out-of-bounds
array reads and writes. But even with that fixed, we assume that we
can feed the command name to snprintf() (via xstrfmt()), which is
stuck for historical reasons using "int", and causes it to fail (and
even trigger a BUG() call).
- since the point of git-shell is to take input from untrusted or
semi-trusted clients, it's a mild denial-of-service. We'll allocate
as many bytes as the client sends us (actually twice as many, since
we immediately duplicate the buffer).
We can fix both by just limiting the amount of per-command input we're
willing to receive.
We should also fix split_cmdline(), of course, which is an accident
waiting to happen, but that can come on top. Most calls to
split_cmdline(), including the other one in git-shell, are OK because
they are reading from an OS-provided argv, which is limited in practice.
This patch should eliminate the immediate vulnerabilities.
I picked 4MB as an arbitrary limit. It's big enough that nobody should
ever run into it in practice (since the point is to run the commands via
exec, we're subject to OS limits which are typically much lower). But
it's small enough that allocating it isn't that big a deal.
The code is mostly just swapping out fgets() for the strbuf call, but we
have to add a few niceties like flushing and trimming line endings. We
could simplify things further by putting the buffer on the stack, but
4MB is probably a bit much there. Note that we'll _always_ allocate 4MB,
which for normal, non-malicious requests is more than we would before
this patch. But on the other hand, other git programs are happy to use
96MB for a delta cache. And since we'd never touch most of those pages,
on a lazy-allocating OS like Linux they won't even get allocated to
actual RAM.
The ideal would be a version of strbuf_getline() that accepted a maximum
value. But for a minimal vulnerability fix, let's keep things localized
and simple. We can always refactor further on top.
The included test fails in an obvious way with ASan or UBSan (which
notice the integer overflow and out-of-bounds reads). Without them, it
fails in a less obvious way: we may segfault, or we may try to xstrfmt()
a long string, leading to a BUG(). Either way, it fails reliably before
this patch, and passes with it. Note that we don't need an EXPENSIVE
prereq on it. It does take 10-15s to fail before this patch, but with
the new limit, we fail almost immediately (and the perl process
generating 2GB of data exits via SIGPIPE).
Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com>
Jeff King [Wed, 28 Sep 2022 22:50:36 +0000 (18:50 -0400)]
shell: add basic tests
We have no tests of even basic functionality of git-shell. Let's add a
couple of obvious ones. This will serve as a framework for adding tests
for new things we fix, as well as making sure we don't screw anything up
too badly while doing so.
Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com>
Taylor Blau [Fri, 29 Jul 2022 19:22:13 +0000 (15:22 -0400)]
transport: make `protocol.file.allow` be "user" by default
An earlier patch discussed and fixed a scenario where Git could be used
as a vector to exfiltrate sensitive data through a Docker container when
a potential victim clones a suspicious repository with local submodules
that contain symlinks.
That security hole has since been plugged, but a similar one still
exists. Instead of convincing a would-be victim to clone an embedded
submodule via the "file" protocol, an attacker could convince an
individual to clone a repository that has a submodule pointing to a
valid path on the victim's filesystem.
For example, if an individual (with username "foo") has their home
directory ("/home/foo") stored as a Git repository, then an attacker
could exfiltrate data by convincing a victim to clone a malicious
repository containing a submodule pointing at "/home/foo/.git" with
`--recurse-submodules`. Doing so would expose any sensitive contents in
stored in "/home/foo" tracked in Git.
For systems (such as Docker) that consider everything outside of the
immediate top-level working directory containing a Dockerfile as
inaccessible to the container (with the exception of volume mounts, and
so on), this is a violation of trust by exposing unexpected contents in
the working copy.
To mitigate the likelihood of this kind of attack, adjust the "file://"
protocol's default policy to be "user" to prevent commands that execute
without user input (including recursive submodule initialization) from
taking place by default.
Suggested-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com>
Taylor Blau [Fri, 29 Jul 2022 19:21:53 +0000 (15:21 -0400)]
t/t9NNN: allow local submodules
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that interact with submodules a handful of times use
`test_config_global`.
Taylor Blau [Fri, 29 Jul 2022 19:21:40 +0000 (15:21 -0400)]
t/t7NNN: allow local submodules
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that only need to interact with submodules in a limited capacity
have individual Git commands annotated with the appropriate
configuration via `-c`. Tests that interact with submodules a handful of
times use `test_config_global` instead. Test scripts that rely on
submodules throughout use a `git config --global` during a setup test
towards the beginning of the script.
Taylor Blau [Fri, 29 Jul 2022 19:21:18 +0000 (15:21 -0400)]
t/t6NNN: allow local submodules
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that only need to interact with submodules in a limited capacity
have individual Git commands annotated with the appropriate
configuration via `-c`.
Taylor Blau [Fri, 29 Jul 2022 19:21:06 +0000 (15:21 -0400)]
t/t5NNN: allow local submodules
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that only need to interact with submodules in a limited capacity
have individual Git commands annotated with the appropriate
configuration via `-c`. Tests that interact with submodules a handful of
times use `test_config_global` instead. Test scripts that rely on
submodules throughout use a `git config --global` during a setup test
towards the beginning of the script.
Taylor Blau [Fri, 29 Jul 2022 19:20:43 +0000 (15:20 -0400)]
t/t4NNN: allow local submodules
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that only need to interact with submodules in a limited capacity
have individual Git commands annotated with the appropriate
configuration via `-c`. Tests that interact with submodules a handful of
times use `test_config_global` instead. Test scripts that rely on
submodules throughout use a `git config --global` during a setup test
towards the beginning of the script.
Taylor Blau [Fri, 29 Jul 2022 19:20:28 +0000 (15:20 -0400)]
t/t3NNN: allow local submodules
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that only need to interact with submodules in a limited capacity
have individual Git commands annotated with the appropriate
configuration via `-c`. Tests that interact with submodules a handful of
times use `test_config_global` instead. Test scripts that rely on
submodules throughout use a `git config --global` during a setup test
towards the beginning of the script.
Taylor Blau [Fri, 29 Jul 2022 19:20:03 +0000 (15:20 -0400)]
t/2NNNN: allow local submodules
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that only need to interact with submodules in a limited capacity
have individual Git commands annotated with the appropriate
configuration via `-c`. Tests that interact with submodules a handful of
times use `test_config_global` instead. Test scripts that rely on
submodules throughout use a `git config --global` during a setup test
towards the beginning of the script.
Taylor Blau [Fri, 29 Jul 2022 19:16:10 +0000 (15:16 -0400)]
t/t1NNN: allow local submodules
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that only need to interact with submodules in a limited capacity
have individual Git commands annotated with the appropriate
configuration via `-c`. Tests that interact with submodules a handful of
times use `test_config_global` instead.
Taylor Blau [Fri, 29 Jul 2022 19:13:58 +0000 (15:13 -0400)]
t/lib-submodule-update.sh: allow local submodules
To prepare for changing the default value of `protocol.file.allow` to
"user", update the `prolog()` function in lib-submodule-update to allow
submodules to be cloned over the file protocol.
This is used by a handful of submodule-related test scripts, which
themselves will have to tweak the value of `protocol.file.allow` in
certain locations. Those will be done in subsequent commits.
Taylor Blau [Thu, 28 Jul 2022 21:35:17 +0000 (17:35 -0400)]
builtin/clone.c: disallow `--local` clones with symlinks
When cloning a repository with `--local`, Git relies on either making a
hardlink or copy to every file in the "objects" directory of the source
repository. This is done through the callpath `cmd_clone()` ->
`clone_local()` -> `copy_or_link_directory()`.
The way this optimization works is by enumerating every file and
directory recursively in the source repository's `$GIT_DIR/objects`
directory, and then either making a copy or hardlink of each file. The
only exception to this rule is when copying the "alternates" file, in
which case paths are rewritten to be absolute before writing a new
"alternates" file in the destination repo.
One quirk of this implementation is that it dereferences symlinks when
cloning. This behavior was most recently modified in 36596fd2df (clone:
better handle symlinked files at .git/objects/, 2019-07-10), which
attempted to support `--local` clones of repositories with symlinks in
their objects directory in a platform-independent way.
Unfortunately, this behavior of dereferencing symlinks (that is,
creating a hardlink or copy of the source's link target in the
destination repository) can be used as a component in attacking a
victim by inadvertently exposing the contents of file stored outside of
the repository.
Take, for example, a repository that stores a Dockerfile and is used to
build Docker images. When building an image, Docker copies the directory
contents into the VM, and then instructs the VM to execute the
Dockerfile at the root of the copied directory. This protects against
directory traversal attacks by copying symbolic links as-is without
dereferencing them.
That is, if a user has a symlink pointing at their private key material
(where the symlink is present in the same directory as the Dockerfile,
but the key itself is present outside of that directory), the key is
unreadable to a Docker image, since the link will appear broken from the
container's point of view.
This behavior enables an attack whereby a victim is convinced to clone a
repository containing an embedded submodule (with a URL like
"file:///proc/self/cwd/path/to/submodule") which has a symlink pointing
at a path containing sensitive information on the victim's machine. If a
user is tricked into doing this, the contents at the destination of
those symbolic links are exposed to the Docker image at runtime.
One approach to preventing this behavior is to recreate symlinks in the
destination repository. But this is problematic, since symlinking the
objects directory are not well-supported. (One potential problem is that
when sharing, e.g. a "pack" directory via symlinks, different writers
performing garbage collection may consider different sets of objects to
be reachable, enabling a situation whereby garbage collecting one
repository may remove reachable objects in another repository).
Instead, prohibit the local clone optimization when any symlinks are
present in the `$GIT_DIR/objects` directory of the source repository.
Users may clone the repository again by prepending the "file://" scheme
to their clone URL, or by adding the `--no-local` option to their `git
clone` invocation.
The directory iterator used by `copy_or_link_directory()` must no longer
dereference symlinks (i.e., it *must* call `lstat()` instead of `stat()`
in order to discover whether or not there are symlinks present). This has
no bearing on the overall behavior, since we will immediately `die()` on
encounter a symlink.
Note that t5604.33 suggests that we do support local clones with
symbolic links in the source repository's objects directory, but this
was likely unintentional, or at least did not take into consideration
the problem with sharing parts of the objects directory with symbolic
links at the time. Update this test to reflect which options are and
aren't supported.
Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com>
Junio C Hamano [Fri, 30 Sep 2022 18:06:22 +0000 (11:06 -0700)]
branch: do not fail a no-op --edit-desc
Imagine running "git branch --edit-description" while on a branch
without the branch description, and then exit the editor after
emptying the edit buffer, which is the way to tell the command that
you changed your mind and you do not want the description after all.
The command should just happily oblige, adding no branch description
for the current branch, and exit successfully. But it fails to do
so:
$ git init -b main
$ git commit --allow-empty -m commit
$ GIT_EDITOR=: git branch --edit-description
fatal: could not unset 'branch.main.description'
The end result is OK in that the configuration variable does not
exist in the resulting repository, but we should do better. If we
know we didn't have a description, and if we are asked not to have a
description by the editor, we can just return doing nothing.
This of course introduces TOCTOU. If you add a branch description
to the same branch from another window, while you had the editor
open to edit the description, and then exit the editor without
writing anything there, we'd end up not removing the description you
added in the other window. But you are fooling yourself in your own
repository at that point, and if it hurts, you'd be better off not
doing so ;-).
test-lib: have SANITIZE=leak imply TEST_NO_MALLOC_CHECK
Since 131b94a10a7 (test-lib.sh: Use GLIBC_TUNABLES instead of
MALLOC_CHECK_ on glibc >= 2.34, 2022-03-04) compiling with
SANITIZE=leak has missed reporting some leaks. The old MALLOC_CHECK
method used before glibc 2.34 seems to have been (mostly?) compatible
with it, but after 131b94a10a7 e.g. running:
TEST_NO_MALLOC_CHECK=1 make SANITIZE=leak test T=t6437-submodule-merge.sh
Would report a leak in builtin/commit.c, but this would not:
TEST_NO_MALLOC_CHECK= make SANITIZE=leak test T=t6437-submodule-merge.sh
Since the interaction is clearly breaking the SANITIZE=leak mode,
let's mark them as explicitly incompatible.
A related regression for SANITIZE=address was fixed in 067109a5e7d (tests: make SANITIZE=address imply TEST_NO_MALLOC_CHECK,
2022-04-09).
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Acked-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Alex Henrie [Wed, 28 Sep 2022 05:58:11 +0000 (23:58 -0600)]
push: improve grammar of branch.autoSetupMerge advice
"upstream branches" is plural but "name" and "local branch" are
singular. Make them all singular. And because we're talking about a
hypothetical branch that doesn't exist yet, use the future tense.
Signed-off-by: Alex Henrie <alexhenrie24@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Victoria Dye [Wed, 28 Sep 2022 17:19:00 +0000 (17:19 +0000)]
read-cache: avoid misaligned reads in index v4
The process for reading the index into memory from disk is to first read its
contents into a single memory-mapped file buffer (type 'char *'), then
sequentially convert each on-disk index entry into a corresponding incore
'cache_entry'. To access the contents of the on-disk entry for processing, a
moving pointer within the memory-mapped file is cast to type 'struct
ondisk_cache_entry *'.
In index v4, the entries in the on-disk index file are written *without*
aligning their first byte to a 4-byte boundary; entries are a variable
length (depending on the entry name and whether or not extended flags are
used). As a result, casting the 'char *' buffer pointer to 'struct
ondisk_cache_entry *' then accessing its contents in a 'SANITIZE=undefined'
build can trigger the following error:
read-cache.c:1886:46: runtime error: member access within misaligned
address <address> for type 'struct ondisk_cache_entry', which requires 4
byte alignment
Avoid this error by reading fields directly from the 'char *' buffer, using
the 'offsetof' individual fields in 'struct ondisk_cache_entry'.
Additionally, add documentation describing why the new approach avoids the
misaligned address error, as well as advice on how to improve the
implementation in the future.
Reported-by: Jeff King <peff@peff.net> Signed-off-by: Victoria Dye <vdye@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
merge-ort: return early when failing to write a blob
In the previous commit, we fixed a segmentation fault when a tree object
could not be written.
However, before the tree object is written, `merge-ort` wants to write
out a blob object (except in cases where the merge results in a blob
that already exists in the database). And this can fail, too, but we
ignore that write failure so far.
Let's pay close attention and error out early if the blob could not be
written. This reduces the error output of t4301.25 ("merge-ort fails
gracefully in a read-only repository") from:
error: insufficient permission for adding an object to repository database ./objects
error: error: Unable to add numbers to database
error: insufficient permission for adding an object to repository database ./objects
error: error: Unable to add greeting to database
error: insufficient permission for adding an object to repository database ./objects
fatal: failure to merge
to:
error: insufficient permission for adding an object to repository database ./objects
error: error: Unable to add numbers to database
fatal: failure to merge
This is _not_ just a cosmetic change: Even though one might assume that
the operation would have failed anyway at the point when the new tree
object is written (and the corresponding tree object _will_ be new if it
contains a blob that is new), but that is not so: As pointed out by
Elijah Newren, when Git has previously been allowed to add loose objects
via `sudo` calls, it is very possible that the blob object cannot be
written (because the corresponding `.git/objects/??/` directory may be
owned by `root`) but the tree object can be written (because the
corresponding objects directory is owned by the current user). This
would result in a corrupt repository because it is missing the blob
object, and with this here patch we prevent that.
Note: This patch adjusts two variable declarations from `unsigned` to
`int` because their purpose is to hold the return value of
`handle_content_merge()`, which is of type `int`. The existing users of
those variables are only interested whether that variable is zero or
non-zero, therefore this type change does not affect the existing code.
Reviewed-by: Elijah Newren <newren@gmail.com> Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
merge-ort: fix segmentation fault in read-only repositories
If the blob/tree objects cannot be written, we really need the merge
operations to fail, and not to continue (and then try to access the tree
object which is however still set to `NULL`).
Let's stop ignoring the return value of `write_object_file()` and
`write_tree()` and set `clean = -1` in the error case.
Reviewed-by: Elijah Newren <newren@gmail.com> Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
string-list: document iterator behavior on NULL input
The for_each_string_list_item() macro takes a string_list and
automatically constructs a for loop to iterate over its contents. This
macro will segfault if the list is non-NULL.
We cannot change the macro to be careful around NULL values because
there are many callers that use the address of a local variable, which
will never be NULL and will cause compile errors with -Werror=address.
For now, leave a documentation comment to try to avoid mistakes in the
future where a caller does not check for a NULL list.
Signed-off-by: Derrick Stolee <derrickstolee@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
The 'git maintenance [un]register' commands set or unset the multi-
valued maintenance.repo config key with the absolute path of the current
repository. These are set in the global config file.
Instead of calling a subcommand and creating a new process, create the
proper API calls to git_config_set_multivar_in_file_gently(). It
requires loading the filename for the global config file (and erroring
out if now $HOME value is set). We also need to be careful about using
CONFIG_REGEX_NONE when adding the value and using
CONFIG_FLAGS_FIXED_VALUE when removing the value. In both cases, we
check that the value already exists (this check already existed for
'unregister').
Also, remove the transparent translation of the error code from the
config API to the exit code of 'git maintenance'. Instead, use die() to
recover from failures at that level. In the case of 'unregister
--force', allow the CONFIG_NOTHING_SET error code to be a success. This
allows a possible race where another process removes the config value.
The end result is that the config value is not set anymore, so we can
treat this as a success.
Reported-by: SZEDER Gábor <szeder.dev@gmail.com> Signed-off-by: Derrick Stolee <derrickstolee@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
The 'scalar unregister' command removes a repository from the list of
registered Scalar repositories and removes it from the list of
repositories registered for background maintenance. If the repository
was not already registered for background maintenance, then the command
fails, even if the repository was still registered as a Scalar
repository.
After using 'scalar clone' or 'scalar register', the repository would be
enrolled in background maintenance since those commands run 'git
maintenance start'. If the user runs 'git maintenance unregister' on
that repository, then it is still in the list of repositories which get
new config updates from 'scalar reconfigure'. The 'scalar unregister'
command would fail since 'git maintenance unregister' would fail.
Further, the add_or_remove_enlistment() method in scalar.c already has
this idempotent nature built in as an expectation since it returns zero
when the scalar.repo list already has the proper containment of the
repository.
The previous change added the 'git maintenance unregister --force'
option, so use it within 'scalar unregister' to make it idempotent.
Signed-off-by: Derrick Stolee <derrickstolee@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
The 'git maintenance unregister' subcommand has a step that removes the
current repository from the multi-valued maitenance.repo config key.
This fails if the repository is not listed in that key. This makes
running 'git maintenance unregister' twice result in a failure in the
second instance.
This failure exit code is helpful, but its message is not. Add a new
die() message that explicitly calls out the failure due to the
repository not being registered.
In some cases, users may want to run 'git maintenance unregister' just
to make sure that background jobs will not start on this repository, but
they do not want to check to see if it is registered first. Add a new
'--force' option that will siltently succeed if the repository is not
already registered.
Also add an extra test of 'git maintenance unregister' at a point where
there are no registered repositories. This should fail without --force.
Signed-off-by: Derrick Stolee <derrickstolee@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
The trace2 region around the call to lazy_bitmap_for_commit() in
bitmap_for_commit() was added in 28cd730680d (pack-bitmap: prepare to
read lookup table extension, 2022-08-14). While adding trace2 regions is
typically helpful for tracking performance, this method is called
possibly thousands of times as a commit walk explores commit history
looking for a matching bitmap. When trace2 output is enabled, this
region is emitted many times and performance is throttled by that
output.
For now, remove these regions entirely.
This is a critical path, and it would be valuable to measure that the
time spent in bitmap_for_commit() does not increase when using the
commit lookup table. The best way to do that would be to use a mechanism
that sums the time spent in a region and reports a single value at the
end of the process. This technique was introduced but not merged by [1]
so maybe this example presents some justification to revisit that
approach.
To help with the 'git blame' output in this region, add a comment that
warns against adding a trace2 region. Delete a test from t5310 that used
that trace output to check that this lookup optimization was activated.
To create this kind of test again in the future, the stopwatch traces
mentioned earlier could be used as a signal that we activated this code
path.
Helpedy-by: Junio C Hamano <gitster@pobox.com> Signed-off-by: Derrick Stolee <derrickstolee@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
In 2708ce62d2 (branch: sort detached HEAD based on a flag, 2021-01-07) a
call to wt_status_state_free_buffers, responsible of freeing the
resources that could be allocated in the local struct wt_status_state
state, was eliminated.
The call to wt_status_state_free_buffers was introduced in 962dd7ebc3
(wt-status: introduce wt_status_state_free_buffers(), 2020-09-27). This
commit brings back that call in get_head_description.
Signed-off-by: Rubén Justo <rjusto@gmail.com> Reviewed-by: Martin Ågren <martin.agren@gmail.com> Acked-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
sequencer: avoid dropping fixup commit that targets self via commit-ish
Commit 68d5d03bc4 (rebase: teach --autosquash to match on sha1 in
addition to message, 2010-11-04) taught autosquash to recognize
subjects like "fixup! 7a235b" where 7a235b is an OID-prefix. It
actually did more than advertised: 7a235b can be an arbitrary
commit-ish (as long as it's not trailed by spaces).
Accidental(?) use of this secret feature revealed a bug where we
would silently drop a fixup commit. The bug can also be triggered
when using an OID-prefix but that's unlikely in practice.
Let the commit with subject "fixup! main" be the tip of the "main"
branch. When computing the fixup target for this commit, we find
the commit itself. This is wrong because, by definition, a fixup
target must be an earlier commit in the todo list. We wrongly find
the current commit because we added it to the todo list prematurely.
Avoid these fixup-cycles by only adding the current commit to the
todo list after we have finished looking for the fixup target.
Reported-by: Erik Cervin Edin <erik@cervined.in> Signed-off-by: Johannes Altmanninger <aclopte@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Alex Henrie [Mon, 27 Jun 2022 04:17:24 +0000 (22:17 -0600)]
l10n: fr: don't say that merge is "the default strategy"
The text of this message was changed in commit 71076d0edde43a7672a9a0f555753ff078602a64 to avoid making any
suggestion about which strategy is better for the situation at hand.
Update the Franch translation to match.
Signed-off-by: Alex Henrie <alexhenrie24@gmail.com>
Martin Ågren [Fri, 23 Sep 2022 08:07:33 +0000 (10:07 +0200)]
cmd-list.perl: fix identifying man sections
We attribute each documentation text file to a man section by finding a
line in the file that looks like "gitfoo(<digit>)". Commit cc75e556a9
("scalar: add to 'git help -a' command list", 2022-09-02) updated this
logic to look not only for "gitfoo" but also "scalarfoo". In doing so,
it forgot to account for the fact that after the updated regex has found
a match, the man section is no longer to be found in `$1` but now lives
in `$2`.
This makes our git(1) manpage look as follows:
Main porcelain commands
git-add(git)
Add file contents to the index.
[...]
gitk(git)
The Git repository browser.
scalar(scalar)
A tool for managing large Git repositories.
Restore the man sections by not capturing the (git|scalar) part of the
match into `$1`.
As noted by Ævar [1], we could even match any "foo" rather than just
"gitfoo" and "scalarfoo", but that's a larger change. For now, just fix
the regression in cc75e556a9.
Helped-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Martin Ågren <martin.agren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Turn on sparse index and remove ensure_full_index().
Before this patch, `git-grep` utilizes the ensure_full_index() method to
expand the index and search all the entries. Because this method
requires walking all the trees and constructing the index, it is the
slow part within the whole command.
To achieve better performance, this patch uses grep_tree() to search the
sparse directory entries and get rid of the ensure_full_index() method.
Why grep_tree() is a better choice over ensure_full_index()?
1) grep_tree() is as correct as ensure_full_index(). grep_tree() looks
into every sparse-directory entry (represented by a tree) recursively
when looping over the index, and the result of doing so matches the
result of expanding the index.
2) grep_tree() utilizes pathspecs to limit the scope of searching.
ensure_full_index() always expands the index, which means it will
always walk all the trees and blobs in the repo without caring if
the user only wants a subset of the content, i.e. using a pathspec.
On the other hand, grep_tree() will only search the contents that
match the pathspec, and thus possibly walking fewer trees.
3) grep_tree() does not construct and copy back a new index, while
ensure_full_index() does. This also saves some time.
----------------
Performance test
- Summary:
p2000 tests demonstrate a ~71% execution time reduction for
`git grep --cached bogus -- "f2/f1/f1/*"` using tree-walking logic.
However, notice that this result varies depending on the pathspec
given. See below "Command used for testing" for more details.
The reason for specifying a pathspec is that, if we don't specify a
pathspec, then grep_tree() will walk all the trees and blobs to find the
pattern, and the time consumed doing so is not too different from using
the original ensure_full_index() method, which also spends most of the
time walking trees. However, when a pathspec is specified, this latest
logic will only walk the area of trees enclosed by the pathspec, and the
time consumed is reasonably a lot less.
Generally speaking, because the performance gain is acheived by walking
less trees, which are specified by the pathspec, the HEAD time v.s.
HEAD~ time in sparse-v[3|4], should be proportional to
"pathspec enclosed area" v.s. "all area", respectively. Namely, the
wider the <pathspec> is encompassing, the less the performance
difference between HEAD~ and HEAD, and vice versa.
That is, if we don't specify a pathspec, the performance difference [1]
is indistinguishable: both methods walk all the trees and take generally
same amount of time (even with the index construction time included for
ensure_full_index()).
[1] Performance test result without pathspec (hence walking all trees):
--------------------------
NEEDSWORK about submodules
There are a few NEEDSWORKs that belong to improvements beyond this
topic. See the NEEDSWORK in builtin/grep.c::grep_submodule() for
more context. The other two NEEDSWORKs in t1092 are also relative.
Suggested-by: Derrick Stolee <derrickstolee@github.com> Helped-by: Derrick Stolee <derrickstolee@github.com> Helped-by: Victoria Dye <vdye@github.com> Helped-by: Elijah Newren <newren@gmail.com> Signed-off-by: Shaoxuan Yuan <shaoxuan.yuan02@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
GNU grep deprecated `egrep` and `fgrep` with release 2.5.3 in 2007.
As of release 3.8 in 2022, those commands warn[1] that they are
obsolescent. Now that all the Git test scripts have been scrubbed of
uses of `egrep` and `fgrep`, make `check-non-portable-shell` complain
about them to prevent new instances from creeping back into the project.
* 'main' of github.com:git/git:
list-objects-filter: initialize sub-filter structs
Git 2.38-rc1
Final batch before -rc1
builtin/diagnose.c: don't translate the two mode values
t/Makefile: remove 'test-results' on 'make clean'
gc: don't translate literal commands
Documentation: clean up various typos in technical docs
Documentation: clean up a few misspelled word typos
version: fix builtin linking & documentation
diagnose: add to command-list.txt
Documentation: add ReviewingGuidelines
commit-graph: Fix missing closedir in expire_commit_graphs
diagnose.c: refactor to safely use 'd_type'
help: fix doubled words in explanation for developer interfaces
api docs: link to html version of api-trace2
docs: fix a few recently broken links
reftable: use a pointer for pq_entry param
Contrary to the documentation on credential helpers, as well as the help
text for git-credential-netrc itself, this helper will `die` when
presented with an unknown property/attribute/token.
Correct the behaviour here by skipping and ignoring any tokens that are
unknown. This means all helpers in the tree are consistent and ignore
any unknown credential properties/attributes.
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
It is the expectation that credential helpers be liberal in what they
accept and conservative in what they return, to allow for future growth
and evolution of the protocol/interaction.
All of the other helpers (store, cache, osxkeychain, libsecret,
gnome-keyring) except `netrc` currently ignore any credential lines
that are not recognised, whereas the Windows helper (wincred) instead
dies.
Fix the discrepancy and ignore unknown lines in the wincred helper.
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Jeff King [Thu, 22 Sep 2022 05:33:29 +0000 (01:33 -0400)]
remote: handle rename of remote without fetch refspec
We return an error when trying to rename a remote that has no fetch
refspec:
$ git config --unset-all remote.origin.fetch
$ git remote rename origin foo
fatal: could not unset 'remote.foo.fetch'
To make things even more confusing, we actually _do_ complete the config
modification, via git_config_rename_section(). After that we try to
rewrite the fetch refspec (to say refs/remotes/foo instead of origin).
But our call to git_config_set_multivar() to remove the existing entries
fails, since there aren't any, and it calls die().
We could fix this by using the "gently" form of the config call, and
checking the error code. But there is an even simpler fix: if we know
that there are no refspecs to rewrite, then we can skip that part
entirely.
Reported-by: John A. Leuenhagen <john@zlima12.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Jeff King [Thu, 22 Sep 2022 05:32:05 +0000 (01:32 -0400)]
clone: allow "--bare" with "-o"
We explicitly forbid the combination of "--bare" with "-o", but there
doesn't seem to be any good reason to do so. The original logic came as
part of e6489a1bdf (clone: do not accept more than one -o option.,
2006-01-22), but that commit does not give any reason.
Furthermore, the equivalent combination via config is allowed:
git -c clone.defaultRemoteName=foo clone ...
and works as expected. It may be that this combination was considered
useless, because a bare clone does not set remote.origin.fetch (and
hence there is no refs/remotes/origin hierarchy). But it does set
remote.origin.url, and that name is visible to the user via "git fetch
origin", etc.
Let's allow the options to be used together, and switch the "forbid"
test in t5606 to check that we use the requested name. That test came
much later in 349cff76de (clone: add tests for --template and some
disallowed option pairs, 2020-09-29), and does not offer any logic
beyond "let's test what the code currently does".
Reported-by: John A. Leuenhagen <john@zlima12.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since commit c54980ab83 (list-objects-filter: convert filter_spec to a
strbuf, 2022-09-11), building with SANITIZE=undefined triggers an error
in t5616.
The problem is that we end up with a strbuf that has been
zero-initialized instead of via STRBUF_INIT. Feeding that strbuf to
strbuf_addbuf() in list_objects_filter_copy() means we will call memcpy
like:
memcpy(some_actual_buffer, NULL, 0);
This works on most systems because we're copying zero bytes, but it is
technically undefined behavior to ever pass NULL to memcpy.
Even though c54980ab83 is where the bug manifests, that is only because
we switched away from a string_list, which is OK with being
zero-initialized (though it may cause other problems by not duplicating
the strings, it happened to be OK in this instance).
The actual bug is caused by the commit before that, 2a01bdedf8
(list-objects-filter: add and use initializers, 2022-09-11). There we
consistently initialize the top-level filter structs, but we forgot the
dynamically allocated ones we stick in filter_options->sub when creating
combined filters.
Note that we need to fix two spots here: where we parse a "combine:"
filter, but also where we transform from a single-filter into a combined
one after seeing multiple "--filter" options. In the second spot, we'll
do some minor refactoring to avoid repeating our very-long array index.
Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Jeff King [Thu, 22 Sep 2022 10:15:38 +0000 (06:15 -0400)]
parse_object_buffer(): respect save_commit_buffer
If the global variable "save_commit_buffer" is set to 0, then
parse_commit() will throw away the commit object data after parsing it,
rather than sticking it into a commit slab. This goes all the way back
to 60ab26de99 ([PATCH] Avoid wasting memory in git-rev-list,
2005-09-15).
But there's another code path which may similarly stash the buffer:
parse_object_buffer(). This is where we end up if we parse a commit via
parse_object(), and it's used directly in a few other code paths like
git-fsck.
The original goal of 60ab26de99 was avoiding extra memory usage for
rev-list. And there it's not all that important to catch parse_object().
We use that function only for looking at the tips of the traversal, and
the majority of the commits are parsed by following parent links, where
we use parse_commit() directly. So we were wasting some memory, but only
a small portion.
It's much easier to see the effect with fsck. Since we now turn off
save_commit_buffer by default there, we _should_ be able to drop the
freeing of the commit buffer in fsck_obj(). But if we do so (taking the
first hunk of this patch without the rest), then the peak heap of "git
fsck" in a clone of git.git goes from 136MB to 194MB. Teaching
parse_object_buffer() to respect save_commit_buffer brings that down to
134.5MB (it's hard to tell from massif's output, but I suspect the
savings comes from avoiding the overhead of the mostly-empty commit
slab).
Other programs should see a small improvement. Both "rev-list --all" and
"fsck --connectivity-only" improve by a few hundred kilobytes, as they'd
avoid loading the tip objects of their traversals.
Most importantly, no code should be hurt by doing this. Any program that
turns off save_commit_buffer is already making the assumption that any
commit it sees may need to have its object data loaded on demand, as it
doesn't know which ones were parsed by parse_commit() versus
parse_object(). Not to mention that anything parsed by the commit graph
may be in the same boat, even if save_commit_buffer was not disabled.
This should be the only spot that needs to be fixed. Grepping for
set_commit_buffer() shows that this and parse_commit() are the only
relevant calls.
Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Jeff King [Thu, 22 Sep 2022 10:13:36 +0000 (06:13 -0400)]
fsck: turn off save_commit_buffer
When parsing a commit, the default behavior is to stuff the original
buffer into a commit_slab (which takes ownership of it). But for a tool
like fsck, this isn't useful. While we may look at the buffer further as
part of fsck_commit(), we'll always do so through a separate pointer;
attaching the buffer to the slab doesn't help.
Worse, it means we have to remember to free the commit buffer in all
call paths. We do so in fsck_obj(), which covers a regular "git fsck".
But with "--connectivity-only", we forget to do so in both
traverse_one_object(), which covers reachable objects, and
mark_unreachable_referents(), which covers unreachable ones. As a
result, that mode ends up storing an uncompressed copy of every commit
on the heap at once.
We could teach the code paths for --connectivity-only to also free
commit buffers. But there's an even easier fix: we can just turn off the
save_commit_buffer flag, and then we won't attach them to the commits in
the first place.
This reduces the peak heap of running "git fsck --connectivity-only" in
a clone of linux.git from ~2GB to ~1GB. According to massif, the
remaining memory goes where you'd expect: the object structs themselves,
the obj_hash containing them, and the delta base cache.
Note that we'll leave the call to free commit buffers in fsck_obj() for
now; it's not quite redundant because of a related bug that we'll fix in
a subsequent commit.
Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
Jeff King [Thu, 22 Sep 2022 10:11:43 +0000 (06:11 -0400)]
fsck: free tree buffers after walking unreachable objects
After calling fsck_walk(), a tree object struct may be left in the
parsed state, with the full tree contents available via tree->buffer.
It's the responsibility of the caller to free these when it's done with
the object to avoid having many trees allocated at once.
In a regular "git fsck", we hit fsck_walk() only from fsck_obj(), which
does call free_tree_buffer(). Likewise for "--connectivity-only", we see
most objects via traverse_one_object(), which makes a similar call.
The exception is in mark_unreachable_referents(). When using both
"--connectivity-only" and "--dangling" (the latter of which is the
default), we walk all of the unreachable objects, and there we forget to
free. Most cases would not notice this, because they don't have a lot of
unreachable objects, but you can make a pathological case like this:
git clone --bare /path/to/linux.git repo.git
cd repo.git
rm packed-refs ;# now everything is unreachable!
git fsck --connectivity-only
That ends up with peak heap usage ~18GB, which is (not coincidentally)
close to the size of all uncompressed trees in the repository. After
this patch, the peak heap is only ~2GB.
A few things to note:
- it might seem like fsck_walk(), if it is parsing the trees, should
be responsible for freeing them. But the situation is quite tricky.
In the non-connectivity mode, after we call fsck_walk() we then
proceed with fsck_object() which actually does the type-specific
sanity checks on the object contents. We do pass our own separate
buffer to fsck_object(), but there's a catch: our earlier call to
parse_object_buffer() may have attached that buffer to the object
struct! So by freeing it, we leave the rest of the code with a
dangling pointer.
Likewise, the call to fsck_walk() in index-pack is subtle. It
attaches a buffer to the tree object that must not be freed! And
so rather than calling free_tree_buffer(), it actually detaches it
by setting tree->buffer to NULL.
These cases would _probably_ be fixable by having fsck_walk() free
the tree buffer only when it was the one who allocated it via
parse_tree(). But that would still leave the callers responsible for
freeing other cases, so they wouldn't be simplified. While the
current semantics for fsck_walk() make it easy to accidentally leak
in new callers, at least they are simple to explain, and it's not a
function that's likely to get a lot of new call-sites.
And in any case, it's probably sensible to fix the leak first with
this simple patch, and try any more complicated refactoring
separately.
- a careful reader may notice that fsck_obj() also frees commit
buffers, but neither the call in traverse_one_object() nor the one
touched in this patch does so. And indeed, this is another problem
for --connectivity-only (and accounts for most of the 2GB heap after
this patch), but it's one we'll fix in a separate commit.
Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
projects / thirdparty / git.git / log
