Hotfix.
* sb/object-store-replace:
get_main_ref_store: BUG() when outside a repository
object.c: clear replace map before freeing it
replace-object.c: remove the_repository from prepare_replace_object
object.c: free replace map in raw_object_store_clear
Configure curl to accept compressed responses when using protocol v2 by
setting `CURLOPT_ENCODING` to "", which indicates that curl should send
an "Accept-Encoding" header with all supported compression encodings.
Signed-off-by: Brandon Williams <bmwill@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Configure curl to accept all encodings which curl supports instead of
only accepting gzip responses.
This fixes an issue when using an installation of curl which is built
without the "zlib" feature. Since aa90b9697 (Enable info/refs gzip
decompression in HTTP client, 2012-09-19) we end up requesting "gzip"
encoding anyway despite libcurl not being able to decode it. Worse,
instead of getting a clear error message indicating so, we end up
falling back to "dumb" http, producing a confusing and difficult to
debug result.
Since curl doesn't do any checking to verify that it supports the a
requested encoding, instead set the curl option `CURLOPT_ENCODING` with
an empty string indicating that curl should send an "Accept-Encoding"
header containing only the encodings supported by curl.
Reported-by: Anton Golubev <anton.golubev@gmail.com>
Signed-off-by: Brandon Williams <bmwill@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* maint-2.16:
Git 2.16.4
Git 2.15.2
Git 2.14.4
Git 2.13.7
verify_path: disallow symlinks in .gitmodules
update-index: stat updated files earlier
verify_dotfile: mention case-insensitivity in comment
verify_path: drop clever fallthrough
skip_prefix: add case-insensitive variant
is_{hfs,ntfs}_dotgitmodules: add tests
is_ntfs_dotgit: match other .git files
is_hfs_dotgit: match other .git files
is_ntfs_dotgit: use a size_t for traversing string
submodule-config: verify submodule names as paths
* maint-2.15:
Git 2.15.2
Git 2.14.4
Git 2.13.7
verify_path: disallow symlinks in .gitmodules
update-index: stat updated files earlier
verify_dotfile: mention case-insensitivity in comment
verify_path: drop clever fallthrough
skip_prefix: add case-insensitive variant
is_{hfs,ntfs}_dotgitmodules: add tests
is_ntfs_dotgit: match other .git files
is_hfs_dotgit: match other .git files
is_ntfs_dotgit: use a size_t for traversing string
submodule-config: verify submodule names as paths
* maint-2.14:
Git 2.14.4
Git 2.13.7
verify_path: disallow symlinks in .gitmodules
update-index: stat updated files earlier
verify_dotfile: mention case-insensitivity in comment
verify_path: drop clever fallthrough
skip_prefix: add case-insensitive variant
is_{hfs,ntfs}_dotgitmodules: add tests
is_ntfs_dotgit: match other .git files
is_hfs_dotgit: match other .git files
is_ntfs_dotgit: use a size_t for traversing string
submodule-config: verify submodule names as paths
* maint-2.13:
Git 2.13.7
verify_path: disallow symlinks in .gitmodules
update-index: stat updated files earlier
verify_dotfile: mention case-insensitivity in comment
verify_path: drop clever fallthrough
skip_prefix: add case-insensitive variant
is_{hfs,ntfs}_dotgitmodules: add tests
is_ntfs_dotgit: match other .git files
is_hfs_dotgit: match other .git files
is_ntfs_dotgit: use a size_t for traversing string
submodule-config: verify submodule names as paths
* jk/submodule-fix-loose:
verify_path: disallow symlinks in .gitmodules
update-index: stat updated files earlier
verify_dotfile: mention case-insensitivity in comment
verify_path: drop clever fallthrough
skip_prefix: add case-insensitive variant
is_{hfs,ntfs}_dotgitmodules: add tests
is_ntfs_dotgit: match other .git files
is_hfs_dotgit: match other .git files
is_ntfs_dotgit: use a size_t for traversing string
submodule-config: verify submodule names as paths
We've recently forbidden .gitmodules to be a symlink in
verify_path(). And it's an easy way to circumvent our fsck
checks for .gitmodules content. So let's complain when we
see it.
Signed-off-by: Jeff King <peff@peff.net>
Now that the internal fsck code has all of the plumbing we
need, we can start checking incoming .gitmodules files.
Naively, it seems like we would just need to add a call to
fsck_finish() after we've processed all of the objects. And
that would be enough to cover the initial test included
here. But there are two extra bits:
1. We currently don't bother calling fsck_object() at all
for blobs, since it has traditionally been a noop. We'd
actually catch these blobs in fsck_finish() at the end,
but it's more efficient to check them when we already
have the object loaded in memory.
2. The second pass done by fsck_finish() needs to access
the objects, but we're actually indexing the pack in
this process. In theory we could give the fsck code a
special callback for accessing the in-pack data, but
it's actually quite tricky:
a. We don't have an internal efficient index mapping
oids to packfile offsets. We only generate it on
the fly as part of writing out the .idx file.
b. We'd still have to reconstruct deltas, which means
we'd basically have to replicate all of the
reading logic in packfile.c.
Instead, let's avoid running fsck_finish() until after
we've written out the .idx file, and then just add it
to our internal packed_git list.
This does mean that the objects are "in the repository"
before we finish our fsck checks. But unpack-objects
already exhibits this same behavior, and it's an
acceptable tradeoff here for the same reason: the
quarantine mechanism means that pushes will be
fully protected.
In addition to a basic push test in t7415, we add a sneaky
pack that reverses the usual object order in the pack,
requiring that index-pack access the tree and blob during
the "finish" step.
This already works for unpack-objects (since it will have
written out loose objects), but we'll check it with this
sneaky pack for good measure.
Signed-off-by: Jeff King <peff@peff.net>
As with the previous commit, we must call fsck's "finish"
function in order to catch any queued objects for
.gitmodules checks.
This second pass will be able to access any incoming
objects, because we will have exploded them to loose objects
by now.
This isn't quite ideal, because it means that bad objects
may have been written to the object database (and a
subsequent operation could then reference them, even if the
other side doesn't send the objects again). However, this is
sufficient when used with receive.fsckObjects, since those
loose objects will all be placed in a temporary quarantine
area that will get wiped if we find any problems.
Signed-off-by: Jeff King <peff@peff.net>
Now that the internal fsck code is capable of checking
.gitmodules files, we just need to teach its callers to use
the "finish" function to check any queued objects.
With this, we can now catch the malicious case in t7415 with
git-fsck.
Signed-off-by: Jeff King <peff@peff.net>
This patch detects and blocks submodule names which do not
match the policy set forth in submodule-config. These should
already be caught by the submodule code itself, but putting
the check here means that newer versions of Git can protect
older ones from malicious entries (e.g., a server with
receive.fsckObjects will block the objects, protecting
clients which fetch from it).
As a side effect, this means fsck will also complain about
.gitmodules files that cannot be parsed (or were larger than
core.bigFileThreshold).
Signed-off-by: Jeff King <peff@peff.net>
If we have a tree that points to a .gitmodules blob but
don't have that blob, we can't check its contents. This
produces an fsck error when we encounter it.
But in the case of a promisor object, this absence is
expected, and we must not complain. Note that this can
technically circumvent our transfer.fsckObjects check.
Imagine a client fetches a tree, but not the matching
.gitmodules blob. An fsck of the incoming objects will show
that we don't have enough information. Later, we do fetch
the actual blob. But we have no idea that it's a .gitmodules
file.
The only ways to get around this would be to re-scan all of
the existing trees whenever new ones enter (which is
expensive), or to somehow persist the gitmodules_found set
between fsck runs (which is complicated).
In practice, it's probably OK to ignore the problem. Any
repository which has all of the objects (including the one
serving the promisor packs) can perform the checks. Since
promisor packs are inherently about a hierarchical topology
in which clients rely on upstream repositories, those
upstream repositories can protect all of their downstream
clients from broken objects.
Signed-off-by: Jeff King <peff@peff.net>
In preparation for performing fsck checks on .gitmodules
files, this commit plumbs in the actual detection of the
files. Note that unlike most other fsck checks, this cannot
be a property of a single object: we must know that the
object is found at a ".gitmodules" path at the root tree of
a commit.
Since the fsck code only sees one object at a time, we have
to mark the related objects to fit the puzzle together. When
we see a commit we mark its tree as a root tree, and when
we see a root tree with a .gitmodules file, we mark the
corresponding blob to be checked.
In an ideal world, we'd check the objects in topological
order: commits followed by trees followed by blobs. In that
case we can avoid ever loading an object twice, since all
markings would be complete by the time we get to the marked
objects. And indeed, if we are checking a single packfile,
this is the order in which Git will generally write the
objects. But we can't count on that:
1. git-fsck may show us the objects in arbitrary order
(loose objects are fed in sha1 order, but we may also
have multiple packs, and we process each pack fully in
sequence).
2. The type ordering is just what git-pack-objects happens
to write now. The pack format does not require a
specific order, and it's possible that future versions
of Git (or a custom version trying to fool official
Git's fsck checks!) may order it differently.
3. We may not even be fscking all of the relevant objects
at once. Consider pushing with transfer.fsckObjects,
where one push adds a blob at path "foo", and then a
second push adds the same blob at path ".gitmodules".
The blob is not part of the second push at all, but we
need to mark and check it.
So in the general case, we need to make up to three passes
over the objects: once to make sure we've seen all commits,
then once to cover any trees we might have missed, and then
a final pass to cover any .gitmodules blobs we found in the
second pass.
We can simplify things a bit by loosening the requirement
that we find .gitmodules only at root trees. Technically
a file like "subdir/.gitmodules" is not parsed by Git, but
it's not unreasonable for us to declare that Git is aware of
all ".gitmodules" files and make them eligible for checking.
That lets us drop the root-tree requirement, which
eliminates one pass entirely. And it makes our worst case
much better: instead of potentially queueing every root tree
to be re-examined, the worst case is that we queue each
unique .gitmodules blob for a second look.
This patch just adds the boilerplate to find .gitmodules
files. The actual content checks will come in a subsequent
commit.
Signed-off-by: Jeff King <peff@peff.net>
Because fscking a blob has always been a noop, we didn't
bother passing around the blob data. In preparation for
content-level checks, let's fix up a few things:
1. The fsck_object() function just returns success for any
blob. Let's a noop fsck_blob(), which we can fill in
with actual logic later.
2. The fsck_loose() function in builtin/fsck.c
just threw away blob content after loading it. Let's
hold onto it until after we've called fsck_object().
The easiest way to do this is to just drop the
parse_loose_object() helper entirely. Incidentally,
this also fixes a memory leak: if we successfully
loaded the object data but did not parse it, we would
have left the function without freeing it.
3. When fsck_loose() loads the object data, it
does so with a custom read_loose_object() helper. This
function streams any blobs, regardless of size, under
the assumption that we're only checking the sha1.
Instead, let's actually load blobs smaller than
big_file_threshold, as the normal object-reading
code-paths would do. This lets us fsck small files, and
a NULL return is an indication that the blob was so big
that it needed to be streamed, and we can pass that
information along to fsck_blob().
Signed-off-by: Jeff King <peff@peff.net>
There's no need for us to manually check for ".git"; it's a
subset of the other filesystem-specific tests. Dropping it
makes our code slightly shorter. More importantly, the
existing code may make a reader wonder why ".GIT" is not
covered here, and whether that is a bug (it isn't, as it's
also covered in the filesystem-specific tests).
Signed-off-by: Jeff King <peff@peff.net>
If fsck reports an error, we say only "Error in object".
This isn't quite as bad as it might seem, since the fsck
code would have dumped some errors to stderr already. But it
might help to give a little more context. The earlier output
would not have even mentioned "fsck", and that may be a clue
that the "fsck.*" or "*.fsckObjects" config may be relevant.
Signed-off-by: Jeff King <peff@peff.net>
* jk/submodule-name-verify-fix:
verify_path: disallow symlinks in .gitmodules
update-index: stat updated files earlier
verify_path: drop clever fallthrough
skip_prefix: add icase-insensitive variant
is_{hfs,ntfs}_dotgitmodules: add tests
path: match NTFS short names for more .git files
is_hfs_dotgit: match other .git files
is_ntfs_dotgit: use a size_t for traversing string
submodule-config: verify submodule names as paths
Note that this includes two bits of evil-merge:
- there's a new call to verify_path() that doesn't actually
have a mode available. It should be OK to pass "0" here,
since we're just manipulating the untracked cache, not an
actual index entry.
- the lstat() in builtin/update-index.c:update_one() needs
to be updated to handle the fsmonitor case (without this
it still behaves correctly, but does an unnecessary
lstat).
There are a few reasons it's not a good idea to make
.gitmodules a symlink, including:
1. It won't be portable to systems without symlinks.
2. It may behave inconsistently, since Git may look at
this file in the index or a tree without bothering to
resolve any symbolic links. We don't do this _yet_, but
the config infrastructure is there and it's planned for
the future.
With some clever code, we could make (2) work. And some
people may not care about (1) if they only work on one
platform. But there are a few security reasons to simply
disallow it:
a. A symlinked .gitmodules file may circumvent any fsck
checks of the content.
b. Git may read and write from the on-disk file without
sanity checking the symlink target. So for example, if
you link ".gitmodules" to "../oops" and run "git
submodule add", we'll write to the file "oops" outside
the repository.
Again, both of those are problems that _could_ be solved
with sufficient code, but given the complications in (1) and
(2), we're better off just outlawing it explicitly.
Note the slightly tricky call to verify_path() in
update-index's update_one(). There we may not have a mode if
we're not updating from the filesystem (e.g., we might just
be removing the file). Passing "0" as the mode there works
fine; since it's not a symlink, we'll just skip the extra
checks.
Signed-off-by: Jeff King <peff@peff.net>
In the update_one(), we check verify_path() on the proposed
path before doing anything else. In preparation for having
verify_path() look at the file mode, let's stat the file
earlier, so we can check the mode accurately.
This is made a bit trickier by the fact that this function
only does an lstat in a few code paths (the ones that flow
down through process_path()). So we can speculatively do the
lstat() here and pass the results down, and just use a dummy
mode for cases where we won't actually be updating the index
from the filesystem.
Signed-off-by: Jeff King <peff@peff.net>
We're more restrictive than we need to be in matching ".GIT"
on case-sensitive filesystems; let's make a note that this
is intentional.
Signed-off-by: Jeff King <peff@peff.net>
We check ".git" and ".." in the same switch statement, and
fall through the cases to share the end-of-component check.
While this saves us a line or two, it makes modifying the
function much harder. Let's just write it out.
Signed-off-by: Jeff King <peff@peff.net>
We have the convenient skip_prefix() helper, but if you want
to do case-insensitive matching, you're stuck doing it by
hand. We could add an extra parameter to the function to
let callers ask for this, but the function is small and
somewhat performance-critical. Let's just re-implement it
for the case-insensitive version.
Signed-off-by: Jeff King <peff@peff.net>
This tests primarily for NTFS issues, but also adds one example of an
HFS+ issue.
Thanks go to Congyi Wu for coming up with the list of examples where
NTFS would possibly equate the filename with `.gitmodules`.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Jeff King <peff@peff.net>
When we started to catch NTFS short names that clash with .git, we only
looked for GIT~1. This is sufficient because we only ever clone into an
empty directory, so .git is guaranteed to be the first subdirectory or
file in that directory.
However, even with a fresh clone, .gitmodules is *not* necessarily the
first file to be written that would want the NTFS short name GITMOD~1: a
malicious repository can add .gitmodul0000 and friends, which sorts
before `.gitmodules` and is therefore checked out *first*. For that
reason, we have to test not only for ~1 short names, but for others,
too.
It's hard to just adapt the existing checks in is_ntfs_dotgit(): since
Windows 2000 (i.e., in all Windows versions still supported by Git),
NTFS short names are only generated in the <prefix>~<number> form up to
number 4. After that, a *different* prefix is used, calculated from the
long file name using an undocumented, but stable algorithm.
For example, the short name of .gitmodules would be GITMOD~1, but if it
is taken, and all of ~2, ~3 and ~4 are taken, too, the short name
GI7EBA~1 will be used. From there, collisions are handled by
incrementing the number, shortening the prefix as needed (until ~9999999
is reached, in which case NTFS will not allow the file to be created).
We'd also want to handle .gitignore and .gitattributes, which suffer
from a similar problem, using the fall-back short names GI250A~1 and
GI7D29~1, respectively.
To accommodate for that, we could reimplement the hashing algorithm, but
it is just safer and simpler to provide the known prefixes. This
algorithm has been reverse-engineered and described at
https://usn.pw/blog/gen/2015/06/09/filenames/, which is defunct but
still available via https://web.archive.org/.
These can be recomputed by running the following Perl script:
-- snip --
use warnings;
use strict;
sub compute_short_name_hash ($) {
my $checksum = 0;
foreach (split('', $_[0])) {
$checksum = ($checksum * 0x25 + ord($_)) & 0xffff;
}
$checksum = ($checksum * 314159269) & 0xffffffff;
$checksum = 1 + (~$checksum & 0x7fffffff) if ($checksum & 0x80000000);
$checksum -= (($checksum * 1152921497) >> 60) * 1000000007;
return scalar reverse sprintf("%x", $checksum & 0xffff);
}
print compute_short_name_hash($ARGV[0]);
-- snap --
E.g., running that with the argument ".gitignore" will
result in "250a" (which then becomes "gi250a" in the code).
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Jeff King <peff@peff.net>
Both verify_path() and fsck match ".git", ".GIT", and other
variants specific to HFS+. Let's allow matching other
special files like ".gitmodules", which we'll later use to
enforce extra restrictions via verify_path() and fsck.
Signed-off-by: Jeff King <peff@peff.net>
We walk through the "name" string using an int, which can
wrap to a negative value and cause us to read random memory
before our array (e.g., by creating a tree with a name >2GB,
since "int" is still 32 bits even on most 64-bit platforms).
Worse, this is easy to trigger during the fsck_tree() check,
which is supposed to be protecting us from malicious
garbage.
Note one bit of trickiness in the existing code: we
sometimes assign -1 to "len" at the end of the loop, and
then rely on the "len++" in the for-loop's increment to take
it back to 0. This is still legal with a size_t, since
assigning -1 will turn into SIZE_MAX, which then wraps
around to 0 on increment.
Signed-off-by: Jeff King <peff@peff.net>
Submodule "names" come from the untrusted .gitmodules file,
but we blindly append them to $GIT_DIR/modules to create our
on-disk repo paths. This means you can do bad things by
putting "../" into the name (among other things).
Let's sanity-check these names to avoid building a path that
can be exploited. There are two main decisions:
1. What should the allowed syntax be?
It's tempting to reuse verify_path(), since submodule
names typically come from in-repo paths. But there are
two reasons not to:
a. It's technically more strict than what we need, as
we really care only about breaking out of the
$GIT_DIR/modules/ hierarchy. E.g., having a
submodule named "foo/.git" isn't actually
dangerous, and it's possible that somebody has
manually given such a funny name.
b. Since we'll eventually use this checking logic in
fsck to prevent downstream repositories, it should
be consistent across platforms. Because
verify_path() relies on is_dir_sep(), it wouldn't
block "foo\..\bar" on a non-Windows machine.
2. Where should we enforce it? These days most of the
.gitmodules reads go through submodule-config.c, so
I've put it there in the reading step. That should
cover all of the C code.
We also construct the name for "git submodule add"
inside the git-submodule.sh script. This is probably
not a big deal for security since the name is coming
from the user anyway, but it would be polite to remind
them if the name they pick is invalid (and we need to
expose the name-checker to the shell anyway for our
test scripts).
This patch issues a warning when reading .gitmodules
and just ignores the related config entry completely.
This will generally end up producing a sensible error,
as it works the same as a .gitmodules file which is
missing a submodule entry (so "submodule update" will
barf, but "git clone --recurse-submodules" will print
an error but not abort the clone.
There is one minor oddity, which is that we print the
warning once per malformed config key (since that's how
the config subsystem gives us the entries). So in the
new test, for example, the user would see three
warnings. That's OK, since the intent is that this case
should never come up outside of malicious repositories
(and then it might even benefit the user to see the
message multiple times).
Credit for finding this vulnerability and the proof of
concept from which the test script was adapted goes to
Etienne Stalmans.
Signed-off-by: Jeff King <peff@peff.net>
We use the lockfile API to avoid multiple Git processes from writing to
the commit-graph file in the .git/objects/info directory. In some cases,
this directory may not exist, so we check for its existence.
The existing code does the following when acquiring the lock:
1. Try to acquire the lock.
2. If it fails, try to create the .git/object/info directory.
3. Try to acquire the lock, failing if necessary.
The problem is that if the lockfile exists, then the mkdir fails, giving
an error that doesn't help the user:
"fatal: cannot mkdir .git/objects/info: File exists"
While technically this honors the lockfile, it does not help the user.
Instead, do the following:
1. Check for existence of .git/objects/info; create if necessary.
2. Try to acquire the lock, failing if necessary.
The new output looks like:
fatal: Unable to create
'<dir>/.git/objects/info/commit-graph.lock': File exists.
Another git process seems to be running in this repository, e.g.
an editor opened by 'git commit'. Please make sure all processes
are terminated then try again. If it still fails, a git process
may have crashed in this repository earlier:
remove the file manually to continue.
Helped-by: Jeff King <peff@peff.net>
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We now calculate generation numbers in the commit-graph file and use
them in paint_down_to_common().
Expand the section on generation numbers to discuss how the three
special generation numbers GENERATION_NUMBER_INFINITY, _ZERO, and
_MAX interact with other generation numbers.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Now that we use generation numbers from the commit-graph, we must
ensure that all commits that exist in the commit-graph are loaded
from that file instead of from the object database. Since the
commit-graph file is only checked if core.commitGraph is true, we
must check the default config before we load any commits.
In the merge builtin, the config was checked after loading the HEAD
commit. This was due to the use of the global 'branch' when checking
merge-specific config settings.
Move the config load to be between the initialization of 'branch' and
the commit lookup.
Without this change, a fast-forward merge would hit a BUG("bad
generation skip") statement in commit.c during paint_down_to_common().
This is because the HEAD commit would be loaded with "infinite"
generation but then reached by commits with "finite" generation
numbers.
Add a test to t5318-commit-graph.sh that exercises this code path to
prevent a regression.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The static remove_redundant() method is used to filter a list
of commits by removing those that are reachable from another
commit in the list. This is used to remove all possible merge-
bases except a maximal, mutually independent set.
To determine these commits are independent, we use a number of
paint_down_to_common() walks and use the PARENT1, PARENT2 flags
to determine reachability. Since we only care about reachability
and not the full set of merge-bases between 'one' and 'twos', we
can use the 'min_generation' parameter to short-circuit the walk.
When no commit-graph exists, there is no change in behavior.
For a copy of the Linux repository, we measured the following
performance improvements:
git merge-base v3.3 v4.5
Before: 234 ms
After: 208 ms
Rel %: -11%
git merge-base v4.3 v4.5
Before: 102 ms
After: 83 ms
Rel %: -19%
The experiments above were chosen to demonstrate that we are
improving the filtering of the merge-base set. In the first
example, more time is spent walking the history to find the
set of merge bases before the remove_redundant() call. The
starting commits are closer together in the second example,
therefore more time is spent in remove_redundant(). The relative
change in performance differs as expected.
Reported-by: Jakub Narebski <jnareb@gmail.com>
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When running 'git branch --contains', the in_merge_bases_many()
method calls paint_down_to_common() to discover if a specific
commit is reachable from a set of branches. Commits with lower
generation number are not needed to correctly answer the
containment query of in_merge_bases_many().
Add a new parameter, min_generation, to paint_down_to_common() that
prevents walking commits with generation number strictly less than
min_generation. If 0 is given, then there is no functional change.
For in_merge_bases_many(), we can pass commit->generation as the
cutoff, and this saves time during 'git branch --contains' queries
that would otherwise walk "around" the commit we are inspecting.
For a copy of the Linux repository, where HEAD is checked out at
v4.13~100, we get the following performance improvement for
'git branch --contains' over the previous commit:
Before: 0.21s
After: 0.13s
Rel %: -38%
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The containment algorithm for 'git branch --contains' is different
from that for 'git tag --contains' in that it uses is_descendant_of()
instead of contains_tag_algo(). The expensive portion of the branch
algorithm is computing merge bases.
When a commit-graph file exists with generation numbers computed,
we can avoid this merge-base calculation when the target commit has
a larger generation number than the initial commits.
Performance tests were run on a copy of the Linux repository where
HEAD is contained in v4.13 but no earlier tag. Also, all tags were
copied to branches and 'git branch --contains' was tested:
Before: 60.0s
After: 0.4s
Rel %: -99.3%
Reported-by: Jeff King <peff@peff.net>
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A commit A can reach a commit B only if the generation number of A
is strictly larger than the generation number of B. This condition
allows significantly short-circuiting commit-graph walks.
Use generation number for '--contains' type queries.
On a copy of the Linux repository where HEAD is contained in v4.13
but no earlier tag, the command 'git tag --contains HEAD' had the
following peformance improvement:
Before: 0.81s
After: 0.04s
Rel %: -95%
Helped-by: Jeff King <peff@peff.net>
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Most code paths load commits using lookup_commit() and then
parse_commit(). In some cases, including some branch lookups, the commit
is parsed using parse_object_buffer() which side-steps parse_commit() in
favor of parse_commit_buffer().
With generation numbers in the commit-graph, we need to ensure that any
commit that exists in the commit-graph file has its generation number
loaded.
Create new load_commit_graph_info() method to fill in the information
for a commit that exists only in the commit-graph file. Call it from
parse_commit_buffer() after loading the other commit information from
the given buffer. Only fill this information when specified by the
'check_graph' parameter.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Define compare_commits_by_gen_then_commit_date(), which uses generation
numbers as a primary comparison and commit date to break ties (or as a
comparison when both commits do not have computed generation numbers).
Since the commit-graph file is closed under reachability, we know that
all commits in the file have generation at most GENERATION_NUMBER_MAX
which is less than GENERATION_NUMBER_INFINITY.
This change does not affect the number of commits that are walked during
the execution of paint_down_to_common(), only the order that those
commits are inspected. In the case that commit dates violate topological
order (i.e. a parent is "newer" than a child), the previous code could
walk a commit twice: if a commit is reached with the PARENT1 bit, but
later is re-visited with the PARENT2 bit, then that PARENT2 bit must be
propagated to its parents. Using generation numbers avoids this extra
effort, even if it is somewhat rare.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
