The code to compute the commit-graph has been taught to use a more
robust way to tell if two object directories refer to the same
thing.
* tb/commit-graph-object-dir:
commit-graph.h: use odb in 'load_commit_graph_one_fd_st'
commit-graph.c: remove path normalization, comparison
commit-graph.h: store object directory in 'struct commit_graph'
commit-graph.h: store an odb in 'struct write_commit_graph_context'
t5318: don't pass non-object directory to '--object-dir'
The index-pack code now diagnoses a bad input packstream that
records the same object twice when it is used as delta base; the
code used to declare a software bug when encountering such an
input, but it is an input error.
* jk/index-pack-dupfix:
index-pack: downgrade twice-resolved REF_DELTA to die()
The way "git submodule status" reports an initialized but not yet
populated submodule has not been reimplemented correctly when a
part of the "git submodule" command was rewritten in C, which has
been corrected.
* pk/status-of-uncloned-submodule:
t7400: testcase for submodule status on unregistered inner git repos
submodule: fix status of initialized but not cloned submodules
t7400: add a testcase for submodule status on empty dirs
Some codepaths were given a repository instance as a parameter to
work in the repository, but passed the_repository instance to its
callees, which has been cleaned up (somewhat).
* mt/use-passed-repo-more-in-funcs:
sha1-file: allow check_object_signature() to handle any repo
sha1-file: pass git_hash_algo to hash_object_file()
sha1-file: pass git_hash_algo to write_object_file_prepare()
streaming: allow open_istream() to handle any repo
pack-check: use given repo's hash_algo at verify_packfile()
cache-tree: use given repo's hash_algo at verify_one()
diff: make diff_populate_filespec() honor its repo argument
Some rough edges in the sparse-checkout feature, especially around
the cone mode, have been cleaned up.
* ds/sparse-checkout-harden:
sparse-checkout: fix cone mode behavior mismatch
sparse-checkout: improve docs around 'set' in cone mode
sparse-checkout: escape all glob characters on write
sparse-checkout: use C-style quotes in 'list' subcommand
sparse-checkout: unquote C-style strings over --stdin
sparse-checkout: write escaped patterns in cone mode
sparse-checkout: properly match escaped characters
sparse-checkout: warn on globs in cone patterns
sparse-checkout: detect short patterns
sparse-checkout: cone mode does not recognize "**"
sparse-checkout: fix documentation typo for core.sparseCheckoutCone
clone: fix --sparse option with URLs
sparse-checkout: create leading directories
t1091: improve here-docs
t1091: use check_files to reduce boilerplate
Unneeded connectivity check is now disabled in a partial clone when
fetching into it.
* jt/connectivity-check-optim-in-partial-clone:
fetch: forgo full connectivity check if --filter
connected: verify promisor-ness of partial clone
"git rebase -i" (and friends) used to unnecessarily check out the
tip of the branch to be rebased, which has been corrected.
* ag/rebase-avoid-unneeded-checkout:
rebase -i: stop checking out the tip of the branch to rebase
Traditionally, we avoided threaded grep while searching in objects
(as opposed to files in the working tree) as accesses to the object
layer is not thread-safe. This limitation is getting lifted.
* mt/threaded-grep-in-object-store:
grep: use no. of cores as the default no. of threads
grep: move driver pre-load out of critical section
grep: re-enable threads in non-worktree case
grep: protect packed_git [re-]initialization
grep: allow submodule functions to run in parallel
submodule-config: add skip_if_read option to repo_read_gitmodules()
grep: replace grep_read_mutex by internal obj read lock
object-store: allow threaded access to object reading
replace-object: make replace operations thread-safe
grep: fix racy calls in grep_objects()
grep: fix race conditions at grep_submodule()
grep: fix race conditions on userdiff calls
Two help messages given when "git add" notices the user gave it
nothing to add have been updated to use advise() API.
* hw/advice-add-nothing:
add: change advice config variables used by the add API
add: use advise function to display hints
"git grep --no-index" should not get affected by the contents of
the .gitmodules file but when "--recurse-submodules" is given or
the "submodule.recurse" variable is set, it did. Now these
settings are ignored in the "--no-index" mode.
* pb/do-not-recurse-grep-no-index:
grep: ignore --recurse-submodules if --no-index is given
advice.addNothing config variable is used to control the visibility of
two advice messages in the add library. This config variable is
replaced by two new variables, whose names are more clear and relevant
to the two cases.
Also add the two new variables to the documentation.
Signed-off-by: Heba Waly <heba.waly@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git checkout X" did not correctly fail when X is not a local
branch but could name more than one remote-tracking branches
(i.e. to be dwimmed as the starting point to create a corresponding
local branch), which has been corrected.
* am/checkout-file-and-ref-ref-ambiguity:
checkout: don't revert file on ambiguous tracking branches
parse_branchname_arg(): extract part as new function
The effort to move "git-add--interactive" to C continues.
* js/patch-mode-in-others-in-c:
commit --interactive: make it work with the built-in `add -i`
built-in add -p: implement the "worktree" patch modes
built-in add -p: implement the "checkout" patch modes
built-in stash: use the built-in `git add -p` if so configured
legacy stash -p: respect the add.interactive.usebuiltin setting
built-in add -p: implement the "stash" and "reset" patch modes
built-in add -p: prepare for patch modes other than "stage"
When we're resolving a REF_DELTA, we compare-and-swap its type from
REF_DELTA to whatever real type the base object has, as discussed in
ab791dd138 (index-pack: fix race condition with duplicate bases,
2014-08-29). If the old type wasn't a REF_DELTA, we consider that a
BUG(). But as discussed in that commit, we might see this case whenever
we try to resolve an object twice, which may happen because we have
multiple copies of the base object.
So this isn't a bug at all, but rather a sign that the input pack is
broken. And indeed, this case is triggered already in t5309.5 and
t5309.6, which create packs with delta cycles and duplicate bases. But
we never noticed because those tests are marked expect_failure.
Those tests were added by b2ef3d9ebb (test index-pack on packs with
recoverable delta cycles, 2013-08-23), which was leaving the door open
for cases that we theoretically _could_ handle. And when we see an
already-resolved object like this, in theory we could keep going after
confirming that the previously resolved child->real_type matches
base->obj->real_type. But:
- enforcing the "only resolve once" rule here saves us from an
infinite loop in other parts of the code. If we keep going, then the
delta cycle in t5309.5 causes us to loop infinitely, as
find_ref_delta_children() doesn't realize which objects have already
been resolved. So there would be more changes needed to make this
case work, and in the meantime we'd be worse off.
- any pack that triggers this is broken anyway. It either has a
duplicate base object, or it has a cycle which causes us to bring in
a duplicate via --fix-thin. In either case, we'd end up rejecting
the pack in write_idx_file(), which also detects duplicates.
So the tests have little value in documenting what we _could_ be doing
(and have been neglected for 6+ years). Let's switch them to confirming
that we handle this case cleanly (and switch out the BUG() for a more
informative die() so that we do so).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Apply a similar treatment as in the previous patch to pass a 'struct
object_directory *' through the 'load_commit_graph_one_fd_st'
initializer, too.
This prevents a potential bug where a pointer comparison is made to a
NULL 'g->odb', which would cause the commit-graph machinery to think
that a pair of commit-graphs belonged to different alternates when in
fact they do not (i.e., in the case of no '--object-dir').
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
As of the previous patch, all calls to 'commit-graph.c' functions which
perform path normalization (for e.g., 'get_commit_graph_filename()') are
of the form 'ctx->odb->path', which is always in normalized form.
Now that there are no callers passing non-normalized paths to these
functions, ensure that future callers are bound by the same restrictions
by making these functions take a 'struct object_directory *' instead of
a 'const char *'. To match, replace all calls with arguments of the form
'ctx->odb->path' with 'ctx->odb' To recover the path, functions that
perform path manipulation simply use 'odb->path'.
Further, avoid string comparisons with arguments of the form
'odb->path', and instead prefer raw pointer comparisons, which
accomplish the same effect, but are far less brittle.
This has a pleasant side-effect of making these functions much more
robust to paths that cannot be normalized by 'normalize_path_copy()',
i.e., because they are outside of the current working directory.
For example, prior to this patch, Valgrind reports that the following
uninitialized memory read [1]:
$ ( cd t && GIT_DIR=../.git valgrind git rev-parse HEAD^ )
because 'normalize_path_copy()' can't normalize '../.git' (since it's
relative to but above of the current working directory) [2].
By using a 'struct object_directory *' directly,
'get_commit_graph_filename()' does not need to normalize, because all
paths are relative to the current working directory since they are
always read from the '->path' of an object directory.
[1]: https://lore.kernel.org/git/20191027042116.GA5801@sigill.intra.peff.net.
[2]: The bug here is that 'get_commit_graph_filename()' returns the
result of 'normalize_path_copy()' without checking the return
value.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In a previous patch, the 'char *object_dir' in 'struct commit_graph' was
replaced with a 'struct object_directory'. This patch applies the same
treatment to 'struct commit_graph', which is another intermediate step
towards getting rid of all path normalization in 'commit-graph.c'.
Instead of taking a 'char *object_dir', functions that construct a
'struct commit_graph' now take a 'struct object_directory *'. Any code
that needs an object directory path use '->path' instead.
This ensures that all calls to functions that perform path normalization
are given arguments which do not themselves require normalization. This
prepares those functions to drop their normalization entirely, which
will occur in the subsequent patch.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There are lots of places in 'commit-graph.h' where a function either has
(or almost has) a full 'struct object_directory *', accesses '->path',
and then throws away the rest of the struct.
This can cause headaches when comparing the locations of object
directories across alternates (e.g., in the case of deciding if two
commit-graph layers can be merged). These paths are normalized with
'normalize_path_copy()' which mitigates some comparison issues, but not
all [1].
Replace usage of 'char *object_dir' with 'odb->path' by storing a
'struct object_directory *' in the 'write_commit_graph_context'
structure. This is an intermediate step towards getting rid of all path
normalization in 'commit-graph.c'.
Resolving a user-provided '--object-dir' argument now requires that we
compare it to the known alternates for equality. Prior to this patch,
an unknown '--object-dir' argument would silently exit with status zero.
This can clearly lead to unintended behavior, such as verifying
commit-graphs that aren't in a repository's own object store (or one of
its alternates), or causing a typo to mask a legitimate commit-graph
verification failure. Make this error non-silent by 'die()'-ing when the
given '--object-dir' does not match any known alternate object store.
[1]: In my testing, for example, I can get one side of the commit-graph
code to fill object_dir with "./objects" and the other with just
"objects".
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The sparse-checkout patterns allow special globs according to
fnmatch(3). When writing cone-mode patterns for paths containing
these characters, they must be escaped.
Use is_glob_special() to check which characters must be escaped
this way, and add a path to the tests that contains all glob
characters at once. Note that ']' is not special, since the
initial bracket '[' is escaped.
Reported-by: Jeff King <peff@peff.net>
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When in cone mode, the 'git sparse-checkout list' subcommand lists
the directories included in the sparse cone. When these directories
contain odd characters, such as a backslash, then we need to use
C-style quotes similar to 'git ls-tree'.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If a user somehow creates a directory with an asterisk (*) or backslash
(\), then the "git sparse-checkout set" command will struggle to provide
the correct pattern in the sparse-checkout file. When not in cone mode,
the provided pattern is written directly into the sparse-checkout file.
However, in cone mode we expect a list of paths to directories and then
we convert those into patterns.
Even more specifically, the goal is to always allow the following from
the root of a repo:
git ls-tree --name-only -d HEAD | git sparse-checkout set --stdin
The ls-tree command provides directory names with an unescaped asterisk.
It also quotes the directories that contain an escaped backslash. We
must remove these quotes, then keep the escaped backslashes.
Use unquote_c_style() when parsing lines from stdin. Command-line
arguments will be parsed as-is, assuming the user can do the correct
level of escaping from their environment to match the exact directory
names.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If a user somehow creates a directory with an asterisk (*) or backslash
(\), then the "git sparse-checkout set" command will struggle to provide
the correct pattern in the sparse-checkout file. When not in cone mode,
the provided pattern is written directly into the sparse-checkout file.
However, in cone mode we expect a list of paths to directories and then
we convert those into patterns.
However, there is some care needed for the timing of these escapes. The
in-memory pattern list is used to update the working directory before
writing the patterns to disk. Thus, we need the command to have the
unescaped names in the hashsets for the cone comparisons, then escape
the patterns later.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We often skip an optional prefix in a string with a hardcoded
constant, e.g.
if (starts_with(string, "prefix"))
string += 6;
which is less error prone when written
skip_prefix(string, "prefix", &string);
Note that this changes a few error messages from "git reflog expire
--expire=nonsense.timestamp", which used to complain by saying
'--expire=nonsense.timestamp' is not a valid timestamp
but with this change, we say
'nonsense.timestamp' is not a valid timestamp
which is more technically correct (the string with --expire= as
a prefix obviously cannot be a valid timestamp, but the error is
about the part of the input without that prefix).
Helped-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Some callers of check_object_signature() can work on arbitrary
repositories, but the repo does not get passed to this function.
Instead, the_repository is always used internally. To fix possible
inconsistencies, allow the function to receive a struct repository and
make those callers pass on the repo being handled.
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Allow hash_object_file() to work on arbitrary repos by introducing a
git_hash_algo parameter. Change callers which have a struct repository
pointer in their scope to pass on the git_hash_algo from the said repo.
For all other callers, pass on the_hash_algo, which was already being
used internally at hash_object_file(). This functionality will be used
in the following patch to make check_object_signature() be able to work
on arbitrary repos (which, in turn, will be used to fix an
inconsistency at object.c:parse_object()).
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Some callers of open_istream() at archive-tar.c and archive-zip.c are
capable of working on arbitrary repositories but the repo struct is not
passed down to open_istream(), which uses the_repository internally. For
now, that's not a problem since the said callers are only being called
with the_repository. But to be consistent and avoid future problems,
let's allow open_istream() to receive a struct repository and use that
instead of the_repository. This parameter addition will also be used in
a future patch to make sha1-file.c:check_object_signature() be able to
work on arbitrary repos.
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
gpg.minTrustLevel configuration variable has been introduced to
tell various signature verification codepaths the required minimum
trust level.
* hi/gpg-mintrustlevel:
gpg-interface: add minTrustLevel as a configuration option
If a filter is specified, we do not need a full connectivity check on
the contents of the packfile we just fetched; we only need to check that
the objects referenced are promisor objects.
This significantly speeds up fetches into repositories that have many
promisor objects, because during the connectivity check, all promisor
objects are enumerated (to mark them UNINTERESTING), and that takes a
significant amount of time.
Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Commit dfa33a298d ("clone: do faster object check for partial clones",
2019-04-21) optimized the connectivity check done when cloning with
--filter to check only the existence of objects directly pointed to by
refs. But this is not sufficient: they also need to be promisor objects.
Make this check more robust by instead checking that these objects are
promisor objects, that is, they appear in a promisor pack.
Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since grep learned to recurse into submodules in 0281e487fd
(grep: optionally recurse into submodules, 2016-12-16),
using --recurse-submodules along with --no-index makes Git
die().
This is unfortunate because if submodule.recurse is set in a user's
~/.gitconfig, invoking `git grep --no-index` either inside or outside
a Git repository results in
fatal: option not supported with --recurse-submodules
Let's allow using these options together, so that setting submodule.recurse
globally does not prevent using `git grep --no-index`.
Using `--recurse-submodules` should not have any effect if `--no-index`
is used inside a repository, as Git will recurse into the checked out
submodule directories just like into regular directories.
Helped-by: Junio C Hamano <gitster@pobox.com>
Signed-off-by: Philippe Blain <levraiphilippeblain@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Original bash helper for "submodule status" was doing a check for
initialized but not cloned submodules and prefixed the status with
a minus sign in case no .git file or folder was found inside the
submodule directory.
This check was missed when the original port of the functionality
from bash to C was done.
Signed-off-by: Peter Kaestle <peter.kaestle@nokia.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The --sparse option was added to the clone builtin in d89f09c (clone:
add --sparse mode, 2019-11-21) and was tested with a local path clone
in t1091-sparse-checkout-builtin.sh. However, due to a difference in
how local paths are handled versus URLs, this mechanism does not work
with URLs.
Modify the test to use a "file://" URL, which would output this error
before the code change:
Cloning into 'clone'...
fatal: cannot change to 'file://.../repo': No such file or directory
error: failed to initialize sparse-checkout
These errors are due to using a "-C <path>" option to call 'git -C
<path> sparse-checkout init' but the URL is being given instead of
the target directory.
Update that target directory to evaluate this correctly. I have also
manually tested that https:// URLs are handled correctly as well.
Acked-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The 'git init' command creates the ".git/info" directory and fills it
with some default files. However, 'git worktree add' does not create
the info directory for that worktree. This causes a problem when running
"git sparse-checkout init" inside a worktree. While care was taken to
allow the sparse-checkout config to be specific to a worktree, this
initialization was untested.
Safely create the leading directories for the sparse-checkout file. This
is the safest thing to do even without worktrees, as a user could delete
their ".git/info" directory and expect Git to recover safely.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
One of the first things done when using a sequencer-based
rebase (ie. `rebase -i', `rebase -r', or `rebase -m') is to make a todo
list. This requires knowledge of the commit range to rebase. To get
the oid of the last commit of the range, the tip of the branch to rebase
is checked out with prepare_branch_to_be_rebased(), then the oid of the
head is read. After this, the tip of the branch is not even modified.
The `am' backend, on the other hand, does not check out the branch.
On big repositories, it's a performance penalty: with `rebase -i', the
user may have to wait before editing the todo list while git is
extracting the branch silently, and "quiet" rebases will be slower than
`am'.
Since we already have the oid of the tip of the branch in
`opts->orig_head', it's useless to switch to this commit.
This removes the call to prepare_branch_to_be_rebased() in
do_interactive_rebase(), and adds a `orig_head' parameter to
get_revision_ranges(). prepare_branch_to_be_rebased() is removed as it
is no longer used.
This introduces a visible change: as we do not switch on the tip of the
branch to rebase, no reflog entry is created at the beginning of the
rebase for it.
Unscientific performance measurements, performed on linux.git, are as
follow:
Before this patch:
$ time git rebase -m --onto v4.18 463fa44eec2fef50~ 463fa44eec2fef50
real 0m8,940s
user 0m6,830s
sys 0m2,121s
After this patch:
$ time git rebase -m --onto v4.18 463fa44eec2fef50~ 463fa44eec2fef50
real 0m1,834s
user 0m0,916s
sys 0m0,206s
Reported-by: SZEDER Gábor <szeder.dev@gmail.com>
Signed-off-by: Alban Gruin <alban.gruin@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Additional checks are added in have_duplicate_entry() and
obj_is_packed() to avoid duplicate objects in the reuse
bitmap. It was probably buggy to not have such a check
before.
Git as a client would never both asks for a tag by sha1 and
specify "include-tag", but libgit2 will, so a libgit2 client
cloning from a Git server would trigger the bug.
If a client both asks for a tag by sha1 and specifies
"include-tag", we may end up including the tag in the reuse
bitmap (due to the first thing), and then later adding it to
the packlist (due to the second). This results in duplicate
objects in the pack, which git chokes on. We should notice
that we are already including it when doing the include-tag
portion, and avoid adding it to the packlist.
The simplest place to fix this is right in add_ref_tag(),
where we could avoid peeling the tag at all if we know that
we are already including it. However, this pushes the check
instead into have_duplicate_entry(). This fixes not only
this case, but also means that we cannot have any similar
problems lurking in other code.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The old code to reuse deltas from an existing packfile
just tried to dump a whole segment of the pack verbatim.
That's faster than the traditional way of actually adding
objects to the packing list, but it didn't kick in very
often. This new code is really going for a middle ground:
do _some_ per-object work, but way less than we'd
traditionally do.
The general strategy of the new code is to make a bitmap
of objects from the packfile we'll include, and then
iterate over it, writing out each object exactly as it is
in our on-disk pack, but _not_ adding it to our packlist
(which costs memory, and increases the search space for
deltas).
One complication is that if we're omitting some objects,
we can't set a delta against a base that we're not
sending. So we have to check each object in
try_partial_reuse() to make sure we have its delta.
About performance, in the worst case we might have
interleaved objects that we are sending or not sending,
and we'd have as many chunks as objects. But in practice
we send big chunks.
For instance, packing torvalds/linux on GitHub servers
now reused 6.5M objects, but only needed ~50k chunks.
Helped-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Let's refactor the way we check if an object is packed by
introducing obj_is_packed(). This function is now a simple
wrapper around packlist_find(), but it will evolve in a
following commit.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Let's make it possible to configure if we want pack reuse or not.
The main reason it might not be wanted is probably debugging and
performance testing, though pack reuse _might_ cause larger packs,
because we wouldn't consider the reused objects as bases for
finding new deltas.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git restore --staged" did not correctly update the cache-tree
structure, resulting in bogus trees to be written afterwards, which
has been corrected.
* nd/switch-and-restore:
restore: invalidate cache-tree when removing entries with --staged
"git commit" gives output similar to "git status" when there is
nothing to commit, but without honoring the advise.statusHints
configuration variable, which has been corrected.
* hw/commit-advise-while-rejecting:
commit: honor advice.statusHints when rejecting an empty commit
In builtin/grep.c:add_work() we pre-load the userdiff drivers before
adding the grep_source in the todo list. This operation is currently
being performed after acquiring the grep_mutex, but as it's already
thread-safe, we don't need to protect it here. So let's move it out of
the critical section which should avoid thread contention and improve
performance.
Running[1] `git grep --threads=8 abcd[02] HEAD` on chromium's
repository[2], I got the following mean times for 30 executions after 2
warmups:
Original | 6.2886s
-------------------------|-----------
Out of critical section | 5.7852s
[1]: Tests performed on an i7-7700HQ with 16GB of RAM and SSD, running
Manjaro Linux.
[2]: chromium’s repo at commit 03ae96f (“Add filters testing at DSF=2”,
04-06-2019), after a 'git gc' execution.
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
They were disabled at 53b8d93 ("grep: disable threading in non-worktree
case", 12-12-2011), due to observable performance drops (to the point
that using a single thread would be faster than multiple threads). But
now that zlib inflation can be performed in parallel we can regain the
speedup, so let's re-enable threads in non-worktree grep.
Grepping 'abcd[02]' ("Regex 1") and '(static|extern) (int|double) \*'
("Regex 2") at chromium's repository[1] I got:
Threads | Regex 1 | Regex 2
---------|------------|-----------
1 | 17.2920s | 20.9624s
2 | 9.6512s | 11.3184s
4 | 6.7723s | 7.6268s
8** | 6.2886s | 6.9843s
These are all means of 30 executions after 2 warmup runs. All tests were
executed on an i7-7700HQ (quad-core w/ hyper-threading), 16GB of RAM and
SSD, running Manjaro Linux. But to make sure the optimization also
performs well on HDD, the tests were repeated on another machine with an
i5-4210U (dual-core w/ hyper-threading), 8GB of RAM and HDD (SATA III,
5400 rpm), also running Manjaro Linux:
Threads | Regex 1 | Regex 2
---------|------------|-----------
1 | 18.4035s | 22.5368s
2 | 12.5063s | 14.6409s
4** | 10.9136s | 12.7106s
** Note that in these cases we relied on hyper-threading, and that's
probably why we don't see a big difference in time.
Unfortunately, multithreaded git-grep might be slow in the non-worktree
case when --textconv is used and there're too many text conversions.
Probably the reason for this is that the object read lock is used to
protect fill_textconv() and therefore there is a mutual exclusion
between textconv execution and object reading. Because both are
time-consuming operations, not being able to perform them in parallel
can cause performance drops. To inform the users about this (and other
threading details), let's also add a "NOTES ON THREADS" section to
Documentation/git-grep.txt.
[1]: chromium’s repo at commit 03ae96f (“Add filters testing at DSF=2”,
04-06-2019), after a 'git gc' execution.
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Some fields in struct raw_object_store are lazy initialized by the
thread-unsafe packfile.c:prepare_packed_git(). Although this function is
present in the call stack of git-grep threads, all paths to it are
currently protected by obj_read_lock() (and the main thread usually
indirectly calls it before firing the worker threads, anyway). However,
it's possible that future modifications add new unprotected paths to it,
introducing a race condition. Because errors derived from it wouldn't
happen often, it could be hard to detect. So to prevent future
headaches, let's force eager initialization of packed_git when setting
git-grep up. There'll be a small overhead in the cases where we didn't
really need to prepare packed_git during execution but this shouldn't be
very noticeable.
Also, packed_git may be re-initialized by
packfile.c:reprepare_packed_git(). Again, all paths to it in git-grep
are already protected by obj_read_lock() but it may suffer from the same
problem in the future. So let's also internally protect it with
obj_read_lock() (which is a recursive mutex).
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Now that object reading operations are internally protected, the
submodule initialization functions at builtin/grep.c:grep_submodule()
are very close to being thread-safe. Let's take a look at each call and
remove from the critical section what we can, for better performance:
- submodule_from_path() and is_submodule_active() cannot be called in
parallel yet only because they call repo_read_gitmodules() which
contains, in its call stack, operations that would otherwise be in
race condition with object reading (for example parse_object() and
is_promisor_remote()). However, they only call repo_read_gitmodules()
if it wasn't read before. So let's pre-read it before firing the
threads and allow these two functions to safely be called in
parallel.
- repo_submodule_init() is already thread-safe, so remove it from the
critical section without other necessary changes.
- The repo_read_gitmodules(&subrepo) call at grep_submodule() is safe as
no other thread is performing object reading operations in the subrepo
yet. However, threads might be working in the superproject, and this
function calls add_to_alternates_memory() internally, which is racy
with object readings in the superproject. So it must be kept
protected for now. Let's add a "NEEDSWORK" to it, informing why it
cannot be removed from the critical section yet.
- Finally, add_to_alternates_memory() must be kept protected for the
same reason as the item above.
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Currently, submodule-config.c doesn't have an externally accessible
function to read gitmodules only if it wasn't already read. But this
exact behavior is internally implemented by gitmodules_read_check(), to
perform a lazy load. Let's merge this function with
repo_read_gitmodules() adding a 'skip_if_read' which allows both
internal and external callers to access this functionality. This
simplifies a little the code. The added option will also be used in
the following patch.
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
git-grep uses 'grep_read_mutex' to protect its calls to object reading
operations. But these have their own internal lock now, which ensures a
better performance (allowing parallel access to more regions). So, let's
remove the former and, instead, activate the latter with
enable_obj_read_lock().
Sections that are currently protected by 'grep_read_mutex' but are not
internally protected by the object reading lock should be surrounded by
obj_read_lock() and obj_read_unlock(). These guarantee mutual exclusion
with object reading operations, keeping the current behavior and
avoiding race conditions. Namely, these places are:
In grep.c:
- fill_textconv() at fill_textconv_grep().
- userdiff_get_textconv() at grep_source_1().
In builtin/grep.c:
- parse_object_or_die() and the submodule functions at
grep_submodule().
- deref_tag() and gitmodules_config_oid() at grep_objects().
If these functions become thread-safe, in the future, we might remove
the locking and probably get some speedup.
Note that some of the submodule functions will already be thread-safe
(or close to being thread-safe) with the internal object reading lock.
However, as some of them will require additional modifications to be
removed from the critical section, this will be done in its own patch.
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
deref_tag() calls is_promisor_object() and parse_object(), both of which
perform lazy initializations and other thread-unsafe operations. If it
was only called by grep_objects() this wouldn't be a problem as the
latter is only executed by the main thread. However, deref_tag() is also
present in read_object_file()'s call stack. So calling deref_tag() in
grep_objects() without acquiring the grep_read_mutex may incur in a race
condition with object reading operations (such as the ones internally
performed by fill_textconv(), called at fill_textconv_grep()). The same
problem happens with the call to gitmodules_config_oid() which also has
parse_object() in its call stack. Fix that protecting both calls with
the said grep_read_mutex.
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
