The code to accept shallow "git push" has been optimized.
* ps/receive-pack-shallow-optim:
commit: use commit graph in `lookup_commit_reference_gently()`
commit: make `repo_parse_commit_no_graph()` more robust
commit: avoid parsing non-commits in `lookup_commit_reference_gently()`
The function `lookup_commit_reference_gently()` can be used to look up a
committish by object ID. As such, the function knows to peel for example
tag objects so that we eventually end up with the commit.
The function is used quite a lot throughout our tree. One such user is
"shallow.c" via `assign_shallow_commits_to_refs()`. The intent of this
function is to figure out whether a shallow push is missing any objects
that are required to satisfy the ref updates, and if so, which of the
ref updates is missing objects.
This is done by painting the tree with `UNINTERESTING`. We start
painting by calling `refs_for_each_ref()` so that we can mark all
existing referenced objects as the boundary of objects that we already
have, and which are supposed to be fully connected. The reference tips
are then parsed via `lookup_commit_reference_gently()`, and the commit
is then marked as uninteresting.
But references may not necessarily point to a committish, and if a lot
of them aren't then this step takes a lot of time. This is mostly due to
the way that `lookup_commit_reference_gently()` is implemented: before
we learn about the type of the object we already call `parse_object()`
on the object ID. This has two consequences:
- We parse all objects, including trees and blobs, even though we
don't even need the contents of them.
- More importantly though, `parse_object()` will cause us to check
whether the object ID matches its contents.
Combined this means that we deflate and hash every non-committish
object, and that of course ends up being both CPU- and memory-intensive.
Improve the logic so that we first use `peel_object()`. This function
won't parse the object for us, and thus it allows us to learn about the
object's type before we parse and return it.
The following benchmark pushes a single object from a shallow clone into
a repository that has 100,000 refs. These refs were created by listing
all objects via `git rev-list(1) --objects --all` and creating refs for
a subset of them, so lots of those refs will cover non-commit objects.
Benchmark 1: git-receive-pack (rev = HEAD~)
Time (mean ± σ): 62.571 s ± 0.413 s [User: 58.331 s, System: 4.053 s]
Range (min … max): 62.191 s … 63.010 s 3 runs
Benchmark 2: git-receive-pack (rev = HEAD)
Time (mean ± σ): 38.339 s ± 0.192 s [User: 36.220 s, System: 1.992 s]
Range (min … max): 38.176 s … 38.551 s 3 runs
Summary
git-receive-pack . </tmp/input (rev = HEAD) ran
1.63 ± 0.01 times faster than git-receive-pack . </tmp/input (rev = HEAD~)
This leads to a sizeable speedup as we now skip reading and parsing
non-commit objects. Before this change we spent around 40% of the time
in `assign_shallow_commits_to_refs()`, after the change we only spend
around 1.2% of the time in there. Almost the entire remainder of the
time is spent in git-rev-list(1) to perform the connectivity checks.
Despite the speedup though, this also leads to a massive reduction in
allocations. Before:
HEAP SUMMARY:
in use at exit: 352,480,441 bytes in 97,185 blocks
total heap usage: 2,793,820 allocs, 2,696,635 frees, 67,271,456,983 bytes allocated
And after:
HEAP SUMMARY:
in use at exit: 17,524,978 bytes in 22,393 blocks
total heap usage: 33,313 allocs, 10,920 frees, 407,774,251 bytes allocated
Note that when all references refer to commits performance stays roughly
the same, as expected. The following benchmark was executed with 600k
commits:
Benchmark 1: git-receive-pack (rev = HEAD~)
Time (mean ± σ): 9.101 s ± 0.006 s [User: 8.800 s, System: 0.520 s]
Range (min … max): 9.095 s … 9.106 s 3 runs
Benchmark 2: git-receive-pack (rev = HEAD)
Time (mean ± σ): 9.128 s ± 0.094 s [User: 8.820 s, System: 0.522 s]
Range (min … max): 9.019 s … 9.188 s 3 runs
Summary
git-receive-pack (rev = HEAD~) ran
1.00 ± 0.01 times faster than git-receive-pack (rev = HEAD)
This will be improved in the next commit.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When inspecting a range of commits from some set of starting references, it
is sometimes useful to learn which commits are not reachable from any other
commits in the selected range.
One such application is in the creation of a sequence of bundles for the
bundle URI feature. Creating a stack of bundles representing different
slices of time includes defining which references to include. If all
references are used, then this may be overwhelming or redundant. Instead,
selecting commits that are maximal to the range could help defining a
smaller reference set to use in the bundle header.
Add a new '--maximal-only' option to restrict the output of a revision range
to be only the commits that are not reachable from any other commit in the
range, based on the reachability definition of the walk.
This is accomplished by adding a new 28th bit flag, CHILD_VISITED, that is
set as we walk. This does extend the bit range in object.h, but using an
earlier bit may collide with another feature.
The tests demonstrate the behavior of the feature with a positive-only
range, ranges with negative references, and walk-modifying flags like
--first-parent and --exclude-first-parent-only.
Since the --boundary option would not increase any results when used with
the --maximal-only option, mark them as incompatible.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git maintenance" command learned "is-needed" subcommand to tell if
it is necessary to perform various maintenance tasks.
* kn/maintenance-is-needed:
maintenance: add 'is-needed' subcommand
maintenance: add checking logic in `pack_refs_condition()`
refs: add a `optimize_required` field to `struct ref_storage_be`
reftable/stack: add function to check if optimization is required
reftable/stack: return stack segments directly
Some ref backend storage can hold not just the object name of an
annotated tag, but the object name of the object the tag points at.
The code to handle this information has been streamlined.
* ps/ref-peeled-tags:
t7004: do not chdir around in the main process
ref-filter: fix stale parsed objects
ref-filter: parse objects on demand
ref-filter: detect broken tags when dereferencing them
refs: don't store peeled object IDs for invalid tags
object: add flag to `peel_object()` to verify object type
refs: drop infrastructure to peel via iterators
refs: drop `current_ref_iter` hack
builtin/show-ref: convert to use `reference_get_peeled_oid()`
ref-filter: propagate peeled object ID
upload-pack: convert to use `reference_get_peeled_oid()`
refs: expose peeled object ID via the iterator
refs: refactor reference status flags
refs: fully reset `struct ref_iterator::ref` on iteration
refs: introduce `.ref` field for the base iterator
refs: introduce wrapper struct for `each_ref_fn`
The 'git-maintenance(1)' command supports an '--auto' flag. Usage of the
flag ensures to run maintenance tasks only if certain thresholds are
met. The heuristic is defined on a task level, wherein each task defines
an 'auto_condition', which states if the task should be run.
The 'pack-refs' task is hard-coded to return 1 as:
1. There was never a way to check if the reference backend needs to be
optimized without actually performing the optimization.
2. We can pass in the '--auto' flag to 'git-pack-refs(1)' which would
optimize based on heuristics.
The previous commit added a `refs_optimize_required()` function, which
can be used to check if a reference backend required optimization. Use
this within `pack_refs_condition()`.
This allows us to add a 'git maintenance is-needed' subcommand which can
notify the user if maintenance is needed without actually performing the
optimization. Without this change, the reference backend would always
state that optimization is needed.
Since we import 'revision.h', we need to remove the definition for
'SEEN' which is duplicated in the included header.
Signed-off-by: Karthik Nayak <karthik.188@gmail.com>
Acked-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Our Bencher dashboards [1] have recently alerted us about a bunch of
performance regressions when writing references, specifically with the
reftable backend. There is a 3x regression when writing many refs with
preexisting refs in the reftable format, and a 10x regression when
migrating refs between backends in either of the formats.
Bisecting the issue lands us at 6ec4c0b45b (refs: don't store peeled
object IDs for invalid tags, 2025-10-23). The gist of the commit is that
we may end up storing peeled objects in both reftables and packed-refs
for corrupted tags, where the claimed tagged object type is different
than the actual tagged object type. This will then cause us to create
the `struct object *` with a wrong type, as well, and obviously nothing
good comes out of that.
The fix for this issue was to introduce a new flag to `peel_object()`
that causes us to verify the tagged object's type before writing it into
the refdb -- if the tag is corrupt, we skip writing the peeled value.
To verify whether the peeled value is correct we have to look up the
object type via the ODB and compare the actual type with the claimed
type, and that additional object lookup is costly.
This also explains why we see the regression only when writing refs with
the reftable backend, but we see the regression with both backends when
migrating refs:
- The reftable backend knows to store peeled values in the new table
immediately, so it has to try and peel each ref it's about to write
to the transaction. So the performance regression is visible for all
writes.
- The files backend only stores peeled values when writing the
packed-refs file, so it wouldn't hit the performance regression for
normal writes. But on ref migrations we know to write all new values
into the packed-refs file immediately, and that's why we see the
regression for both backends there.
Taking a step back though reveals an oddity in the new verification
logic: we not only verify the _tagged_ object's type, but we also verify
the type of the tag itself. But this isn't really needed, as we wouldn't
hit the bug in such a case anyway, as we only hit the issue with corrupt
tags claiming an invalid type for the tagged object.
The consequence of this is that we now started to look up the target
object of every single reference we're about to write, regardless of
whether it even is a tag or not. And that is of course quite costly.
Fix the issue by only verifying the type of the tagged objects. This
means that we of course still have a performance hit for actual tags.
But this only happens for writes anyway, and I'd claim it's preferable
to not store corrupted data in the refdb than to be fast here. Rename
the flag accordingly to clarify that we only verify the tagged object's
type.
This fix brings performance back to previous levels:
Benchmark 1: baseline
Time (mean ± σ): 46.0 ms ± 0.4 ms [User: 40.0 ms, System: 5.7 ms]
Range (min … max): 45.0 ms … 47.1 ms 54 runs
Benchmark 2: regression
Time (mean ± σ): 140.2 ms ± 1.3 ms [User: 77.5 ms, System: 60.5 ms]
Range (min … max): 138.0 ms … 142.7 ms 20 runs
Benchmark 3: fix
Time (mean ± σ): 46.2 ms ± 0.4 ms [User: 40.2 ms, System: 5.7 ms]
Range (min … max): 45.0 ms … 47.3 ms 55 runs
Summary
update-ref: baseline
1.00 ± 0.01 times faster than fix
3.05 ± 0.04 times faster than regression
[1]: https://bencher.dev/perf/git/plots
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When peeling a tag to a non-tag object we repeatedly call
`parse_object()` on the tagged object until we find the first object
that isn't a tag. While this feels sensible at first, there is a big
catch here: `parse_object()` doesn't actually verify the type of the
tagged object.
The relevant code path here eventually ends up in `parse_tag_buffer()`.
Here, we parse the various fields of the tag, including the "type". Once
we've figured out the type and the tagged object ID, we call one of the
`lookup_${type}()` functions for whatever type we have found. There is
two possible outcomes in the successful case:
1. The object is already part of our cached objects. In that case we
double-check whether the type we're trying to look up matches the
type that was cached.
2. The object is _not_ part of our cached objects. In that case, we
simply create a new object with the expected type, but we don't
parse that object.
In the first case we might notice type mismatches, but only in the case
where our cache has the object with the correct type. In the second
case, we'll blindly assume that the type is correct and then go with it.
We'll only notice that the type might be wrong when we try to parse the
object at a later point.
Now arguably, we could change `parse_tag_buffer()` to verify the tagged
object's type for us. But that would have the effect that such a tag
cannot be parsed at all anymore, and we have a small bunch of tests for
exactly this case that assert we still can open such tags. So this
change does not feel like something we can retroactively tighten, even
though one shouldn't ever hit such corrupted tags.
Instead, add a new `flags` field to `peel_object()` that allows the
caller to opt in to strict object verification. This will be wired up at
a subset of callsites over the next few commits.
Note that this change also inlines `deref_tag_noverify()`. There's only
been two callsites of that function, the one we're changing and one in
our test helpers. The latter callsite can trivially use `deref_tag()`
instead, so by inlining the function we avoid having to pass down the
flag.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The current implementation of git-last-modified(1) works by doing a
revision walk, and inspecting the diff at each level of that walk to
annotate entries remaining in the hashmap of paths. In other words, if
the diff at some level touches a path which has not yet been associated
with a commit, then that commit becomes associated with the path.
While a perfectly reasonable implementation, it can perform poorly in
either one of two scenarios:
1. There are many entries of interest, in which case there is simply
a lot of work to do.
2. Or, there are (even a few) entries which have not been updated in a
long time, and so we must walk through a lot of history in order to
find a commit that touches that path.
This patch rewrites the last-modified implementation that addresses the
second point. The idea behind the algorithm is to propagate a set of
'active' paths (a path is 'active' if it does not yet belong to a
commit) up to parents and do a truncated revision walk.
The walk is truncated because it does not produce a revision for every
change in the original pathspec, but rather only for active paths.
More specifically, consider a priority queue of commits sorted by
generation number. First, enqueue the set of boundary commits with all
paths in the original spec marked as interesting.
Then, while the queue is not empty, do the following:
1. Pop an element, say, 'c', off of the queue, making sure that 'c'
isn't reachable by anything in the '--not' set.
2. For each parent 'p' (with index 'parent_i') of 'c', do the
following:
a. Compute the diff between 'c' and 'p'.
b. Pass any active paths that are TREESAME from 'c' to 'p'.
c. If 'p' has any active paths, push it onto the queue.
3. Any path that remains active on 'c' is associated to that commit.
This ends up being equivalent to doing something like 'git log -1 --
$path' for each path simultaneously. But, it allows us to go much faster
than the original implementation by limiting the number of diffs we
compute, since we can avoid parts of history that would have been
considered by the revision walk in the original implementation, but are
known to be uninteresting to us because we have already marked all paths
in that area to be inactive.
To avoid computing many first-parent diffs, add another trick on top of
this and check if all paths active in 'c' are DEFINITELY NOT in c's
Bloom filter. Since the commit-graph only stores first-parent diffs in
the Bloom filters, we can only apply this trick to first-parent diffs.
Comparing the performance of this new algorithm shows about a 2.5x
improvement on git.git:
Benchmark 1: master no bloom
Time (mean ± σ): 2.868 s ± 0.023 s [User: 2.811 s, System: 0.051 s]
Range (min … max): 2.847 s … 2.926 s 10 runs
Benchmark 2: master with bloom
Time (mean ± σ): 949.9 ms ± 15.2 ms [User: 907.6 ms, System: 39.5 ms]
Range (min … max): 933.3 ms … 971.2 ms 10 runs
Benchmark 3: HEAD no bloom
Time (mean ± σ): 782.0 ms ± 6.3 ms [User: 740.7 ms, System: 39.2 ms]
Range (min … max): 776.4 ms … 798.2 ms 10 runs
Benchmark 4: HEAD with bloom
Time (mean ± σ): 307.1 ms ± 1.7 ms [User: 276.4 ms, System: 29.9 ms]
Range (min … max): 303.7 ms … 309.5 ms 10 runs
Summary
HEAD with bloom ran
2.55 ± 0.02 times faster than HEAD no bloom
3.09 ± 0.05 times faster than master with bloom
9.34 ± 0.09 times faster than master no bloom
In short, the existing implementation is comparably fast *with* Bloom
filters as the new implementation is *without* Bloom filters. So, most
repositories should get a dramatic speed-up by just deploying this (even
without computing Bloom filters), and all repositories should get faster
still when computing Bloom filters.
When comparing a more extreme example of
`git last-modified -- COPYING t`, the difference is even 5 times better:
Benchmark 1: master
Time (mean ± σ): 4.372 s ± 0.057 s [User: 4.286 s, System: 0.062 s]
Range (min … max): 4.308 s … 4.509 s 10 runs
Benchmark 2: HEAD
Time (mean ± σ): 826.3 ms ± 22.3 ms [User: 784.1 ms, System: 39.2 ms]
Range (min … max): 810.6 ms … 881.2 ms 10 runs
Summary
HEAD ran
5.29 ± 0.16 times faster than master
As an added benefit, results are more consistent now. For example
implementation in 'master' gives:
$ git log --max-count=1 --format=%H -- pkt-line.h
15df15fe07
$ git last-modified -- pkt-line.h
15df15fe07 pkt-line.h
$ git last-modified | grep pkt-line.h
5b49c1af03 pkt-line.h
With the changes in this patch the results of git-last-modified(1)
always match those of `git log --max-count=1`.
One thing to note though, the results might be outputted in a different
order than before. This is not considerd to be an issue because nowhere
is documented the order is guaranteed.
Based-on-patches-by: Derrick Stolee <stolee@gmail.com>
Based-on-patches-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Toon Claes <toon@iotcl.com>
Acked-by: Taylor Blau <me@ttaylorr.com>
[jc: tweaked use of xcalloc() to unbreak coccicheck]
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Optimize the code to dedup references recorded in a bundle file.
* kn/bundle-dedup-optim:
bundle: fix non-linear performance scaling with refs
t6020: test for duplicate refnames in bundle creation
The 'git bundle create' command has non-linear performance with the
number of refs in the repository. Benchmarking the command shows that
a large portion of the time (~75%) is spent in the
`object_array_remove_duplicates()` function.
The `object_array_remove_duplicates()` function was added in
b2a6d1c686 (bundle: allow the same ref to be given more than once,
2009-01-17) to skip duplicate refs provided by the user from being
written to the bundle. Since this is an O(N^2) algorithm, in repos with
large number of references, this can take up a large amount of time.
Let's instead use a 'strset' to skip duplicates inside
`write_bundle_refs()`. This improves the performance by around 6 times
when tested against in repository with 100000 refs:
Benchmark 1: bundle (refcount = 100000, revision = master)
Time (mean ± σ): 14.653 s ± 0.203 s [User: 13.940 s, System: 0.762 s]
Range (min … max): 14.237 s … 14.920 s 10 runs
Benchmark 2: bundle (refcount = 100000, revision = HEAD)
Time (mean ± σ): 2.394 s ± 0.023 s [User: 1.684 s, System: 0.798 s]
Range (min … max): 2.364 s … 2.425 s 10 runs
Summary
bundle (refcount = 100000, revision = HEAD) ran
6.12 ± 0.10 times faster than bundle (refcount = 100000, revision = master)
Previously, `object_array_remove_duplicates()` ensured that both the
refname and the object it pointed to were checked for duplicates. The
new approach, implemented within `write_bundle_refs()`, eliminates
duplicate refnames without comparing the objects they reference. This
works because, for bundle creation, we only need to prevent duplicate
refs from being written to the bundle header. The `revs->pending` array
can contain duplicates of multiple types.
First, references which resolve to the same refname. For e.g. "git
bundle create out.bdl master master" or "git bundle create out.bdl
refs/heads/master refs/heads/master" or "git bundle create out.bdl
master refs/heads/master". In these scenarios we want to prevent writing
"refs/heads/master" twice to the bundle header. Since both the refnames
here would point to the same object (unless there is a race), we do not
need to check equality of the object.
Second, refnames which are duplicates but do not point to the same
object. This can happen when we use an exclusion criteria. For e.g. "git
bundle create out.bdl master master^!", Here `revs->pending` would
contain two elements, both with refname set to "master". However, each
of them would be pointing to an INTERESTING and UNINTERESTING object
respectively. Since we only write refnames with INTERESTING objects to
the bundle header, we perform our duplicate checks only on such objects.
Signed-off-by: Karthik Nayak <karthik.188@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There are a couple of functions exposed by "object.c" that implicitly
depend on `the_repository`. Remove this dependency by injecting the
repository via a parameter. Adapt callers accordingly by simply using
`the_repository`, except in cases where the subsystem is already free of
the repository. In that case, we instead pass the repository provided by
the caller's context.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We do not clear grafts part of the parsed object pool when clearing the
pool itself, which can lead to memory leaks when a repository is being
cleared.
Fix this by moving `reset_commit_grafts()` into "object.c" and making it
part of the `struct parsed_object_pool` interface such that we can call
it from `parsed_object_pool_clear()`. Adapt `parsed_object_pool_new()`
to take and store a reference to its owning repository, which is needed
by `unparse_commit()`.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The Bloom filter used for path limited history traversal was broken
on systems whose "char" is unsigned; update the implementation and
bump the format version to 2.
* tb/path-filter-fix:
bloom: introduce `deinit_bloom_filters()`
commit-graph: reuse existing Bloom filters where possible
object.h: fix mis-aligned flag bits table
commit-graph: new Bloom filter version that fixes murmur3
commit-graph: unconditionally load Bloom filters
bloom: prepare to discard incompatible Bloom filters
bloom: annotate filters with hash version
repo-settings: introduce commitgraph.changedPathsVersion
t4216: test changed path filters with high bit paths
t/helper/test-read-graph: implement `bloom-filters` mode
bloom.h: make `load_bloom_filter_from_graph()` public
t/helper/test-read-graph.c: extract `dump_graph_info()`
gitformat-commit-graph: describe version 2 of BDAT
commit-graph: ensure Bloom filters are read with consistent settings
revision.c: consult Bloom filters for root commits
t/t4216-log-bloom.sh: harden `test_bloom_filters_not_used()`
A CPP macro USE_THE_REPOSITORY_VARIABLE is introduced to help
transition the codebase to rely less on the availability of the
singleton the_repository instance.
* ps/use-the-repository:
hex: guard declarations with `USE_THE_REPOSITORY_VARIABLE`
t/helper: remove dependency on `the_repository` in "proc-receive"
t/helper: fix segfault in "oid-array" command without repository
t/helper: use correct object hash in partial-clone helper
compat/fsmonitor: fix socket path in networked SHA256 repos
replace-object: use hash algorithm from passed-in repository
protocol-caps: use hash algorithm from passed-in repository
oidset: pass hash algorithm when parsing file
http-fetch: don't crash when parsing packfile without a repo
hash-ll: merge with "hash.h"
refs: avoid include cycle with "repository.h"
global: introduce `USE_THE_REPOSITORY_VARIABLE` macro
hash: require hash algorithm in `empty_tree_oid_hex()`
hash: require hash algorithm in `is_empty_{blob,tree}_oid()`
hash: make `is_null_oid()` independent of `the_repository`
hash: convert `oidcmp()` and `oideq()` to compare whole hash
global: ensure that object IDs are always padded
hash: require hash algorithm in `oidread()` and `oidclr()`
hash: require hash algorithm in `hasheq()`, `hashcmp()` and `hashclr()`
hash: drop (mostly) unused `is_empty_{blob,tree}_sha1()` functions
In an earlier commit, a bug was described where it's possible for Git to
produce non-murmur3 hashes when the platform's "char" type is signed,
and there are paths with characters whose highest bit is set (i.e. all
characters >= 0x80).
That patch allows the caller to control which version of Bloom filters
are read and written. However, even on platforms with a signed "char"
type, it is possible to reuse existing Bloom filters if and only if
there are no changed paths in any commit's first parent tree-diff whose
characters have their highest bit set.
When this is the case, we can reuse the existing filter without having
to compute a new one. This is done by marking trees which are known to
have (or not have) any such paths. When a commit's root tree is verified
to not have any such paths, we mark it as such and declare that the
commit's Bloom filter is reusable.
Note that this heuristic only goes in one direction. If neither a commit
nor its first parent have any paths in their trees with non-ASCII
characters, then we know for certain that a path with non-ASCII
characters will not appear in a tree-diff against that commit's first
parent. The reverse isn't necessarily true: just because the tree-diff
doesn't contain any such paths does not imply that no such paths exist
in either tree.
So we end up recomputing some Bloom filters that we don't strictly have
to (i.e. their bits are the same no matter which version of murmur3 we
use). But culling these out is impossible, since we'd have to perform
the full tree-diff, which is the same effort as computing the Bloom
filter from scratch.
But because we can cache our results in each tree's flag bits, we can
often avoid recomputing many filters, thereby reducing the time it takes
to run
$ git commit-graph write --changed-paths --reachable
when upgrading from v1 to v2 Bloom filters.
To benchmark this, let's generate a commit-graph in linux.git with v1
changed-paths in generation order[^1]:
$ git clone git@github.com:torvalds/linux.git
$ cd linux
$ git commit-graph write --reachable --changed-paths
$ graph=".git/objects/info/commit-graph"
$ mv $graph{,.bak}
Then let's time how long it takes to go from v1 to v2 filters (with and
without the upgrade path enabled), resetting the state of the
commit-graph each time:
$ git config commitGraph.changedPathsVersion 2
$ hyperfine -p 'cp -f $graph.bak $graph' -L v 0,1 \
'GIT_TEST_UPGRADE_BLOOM_FILTERS={v} git.compile commit-graph write --reachable --changed-paths'
On linux.git (where there aren't any non-ASCII paths), the timings
indicate that this patch represents a speed-up over recomputing all
Bloom filters from scratch:
Benchmark 1: GIT_TEST_UPGRADE_BLOOM_FILTERS=0 git.compile commit-graph write --reachable --changed-paths
Time (mean ± σ): 124.873 s ± 0.316 s [User: 124.081 s, System: 0.643 s]
Range (min … max): 124.621 s … 125.227 s 3 runs
Benchmark 2: GIT_TEST_UPGRADE_BLOOM_FILTERS=1 git.compile commit-graph write --reachable --changed-paths
Time (mean ± σ): 79.271 s ± 0.163 s [User: 74.611 s, System: 4.521 s]
Range (min … max): 79.112 s … 79.437 s 3 runs
Summary
'GIT_TEST_UPGRADE_BLOOM_FILTERS=1 git.compile commit-graph write --reachable --changed-paths' ran
1.58 ± 0.01 times faster than 'GIT_TEST_UPGRADE_BLOOM_FILTERS=0 git.compile commit-graph write --reachable --changed-paths'
On git.git, we do have some non-ASCII paths, giving us a more modest
improvement from 4.163 seconds to 3.348 seconds, for a 1.24x speed-up.
On my machine, the stats for git.git are:
- 8,285 Bloom filters computed from scratch
- 10 Bloom filters generated as empty
- 4 Bloom filters generated as truncated due to too many changed paths
- 65,114 Bloom filters were reused when transitioning from v1 to v2.
[^1]: Note that this is is important, since `--stdin-packs` or
`--stdin-commits` orders commits in the commit-graph by their pack
position (with `--stdin-packs`) or in the raw input (with
`--stdin-commits`).
Since we compute Bloom filters in the same order that commits appear
in the graph, we must see a commit's (first) parent before we process
the commit itself. This is only guaranteed to happen when sorting
commits by their generation number.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The pseudo-merge reachability bitmap to help more efficient storage
of the reachability bitmap in a repository with too many refs has
been added.
* tb/pseudo-merge-reachability-bitmap: (26 commits)
pack-bitmap.c: ensure pseudo-merge offset reads are bounded
Documentation/technical/bitmap-format.txt: add missing position table
t/perf: implement performance tests for pseudo-merge bitmaps
pseudo-merge: implement support for finding existing merges
ewah: `bitmap_equals_ewah()`
pack-bitmap: extra trace2 information
pack-bitmap.c: use pseudo-merges during traversal
t/test-lib-functions.sh: support `--notick` in `test_commit_bulk()`
pack-bitmap: implement test helpers for pseudo-merge
ewah: implement `ewah_bitmap_popcount()`
pseudo-merge: implement support for reading pseudo-merge commits
pack-bitmap.c: read pseudo-merge extension
pseudo-merge: scaffolding for reads
pack-bitmap: extract `read_bitmap()` function
pack-bitmap-write.c: write pseudo-merge table
pseudo-merge: implement support for selecting pseudo-merge commits
config: introduce `git_config_double()`
pack-bitmap: make `bitmap_writer_push_bitmapped_commit()` public
pack-bitmap: implement `bitmap_writer_has_bitmapped_object_id()`
pack-bitmap-write: support storing pseudo-merge commits
...
The "hash-ll.h" header was introduced via d1cbe1e6d8 (hash-ll.h: split
out of hash.h to remove dependency on repository.h, 2023-04-22) to make
explicit the split between hash-related functions that rely on the
global `the_repository`, and those that don't. This split is no longer
necessary now that we we have removed the reliance on `the_repository`.
Merge "hash-ll.h" back into "hash.h". This causes some code units to not
include "repository.h" anymore, which requires us to add some forward
declarations.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Further clean-up the refs subsystem to stop relying on
the_repository, and instead use the repository associated to the
ref_store object.
* ps/refs-without-the-repository-updates:
refs/packed: remove references to `the_hash_algo`
refs/files: remove references to `the_hash_algo`
refs/files: use correct repository
refs: remove `dwim_log()`
refs: drop `git_default_branch_name()`
refs: pass repo when peeling objects
refs: move object peeling into "object.c"
refs: pass ref store when detecting dangling symrefs
refs: convert iteration over replace refs to accept ref store
refs: retrieve worktree ref stores via associated repository
refs: refactor `resolve_gitlink_ref()` to accept a repository
refs: pass repo when retrieving submodule ref store
refs: track ref stores via strmap
refs: implement releasing ref storages
refs: rename `init_db` callback to avoid confusion
refs: adjust names for `init` and `init_db` callbacks
Prepare to write pseudo-merge bitmaps by annotating individual bitmapped
commits (which are represented by the `bitmapped_commit` structure) with
an extra bit indicating whether or not they are a pseudo-merge.
In subsequent commits, pseudo-merge bitmaps will be generated by
allocating a fake commit node with parents covering the full set of
commits represented by the pseudo-merge bitmap. These commits will be
added to the set of "selected" commits as usual, but will be written
specially instead of being included with the rest of the selected
commits.
Mechanically speaking, there are two parts of this change:
- The bitmapped_commit struct gets a new bit indicating whether it is
a pseudo-merge, or an ordinary commit selected for bitmaps.
- A handful of changes to only write out the non-pseudo-merge commits
when enumerating through the selected array (see the new
`bitmap_writer_selected_nr()` function). Pseudo-merge commits appear
after all non-pseudo-merge commits, so it is safe to enumerate
through the selected array like so:
for (i = 0; i < bitmap_writer_selected_nr(); i++)
if (writer.selected[i].pseudo_merge)
BUG("unexpected pseudo-merge");
without encountering the BUG().
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Both `peel_object()` and `peel_iterated_oid()` implicitly rely on
`the_repository` to look up objects. Despite the fact that we want to
get rid of `the_repository`, it also leads to some restrictions in our
ref iterators when trying to retrieve the peeled value for a repository
other than `the_repository`.
Refactor these functions such that both take a repository as argument
and remove the now-unnecessary restrictions.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Peeling an object has nothing to do with refs, but we still have the
code in "refs.c". Move it over into "object.c", which is a more natural
place to put it.
Ideally, we'd also move `peel_iterated_oid()` over into "object.c". But
this function is tied to the refs interfaces because it uses a global
ref iterator variable to optimize peeling when the iterator already has
the peeled object ID readily available.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In commit 7cc8f97108 (pack-objects: implement bitmap writing,
2013-12-21) the NEEDS_BITMAP flag was introduced into pack-bitmap.h, but
no object flags allocation table existed at the time.
In 208acbfb82 (object.h: centralize object flag allocation, 2014-03-25)
when that table was first introduced, we never added the flags from
7cc8f97108, which has remained the case since.
Rectify this by including the flag bit used by pack-bitmap.h into the
centralized table in object.h.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Work to support a repository that work with both SHA-1 and SHA-256
hash algorithms has started.
* eb/hash-transition: (30 commits)
t1016-compatObjectFormat: add tests to verify the conversion between objects
t1006: test oid compatibility with cat-file
t1006: rename sha1 to oid
test-lib: compute the compatibility hash so tests may use it
builtin/ls-tree: let the oid determine the output algorithm
object-file: handle compat objects in check_object_signature
tree-walk: init_tree_desc take an oid to get the hash algorithm
builtin/cat-file: let the oid determine the output algorithm
rev-parse: add an --output-object-format parameter
repository: implement extensions.compatObjectFormat
object-file: update object_info_extended to reencode objects
object-file-convert: convert commits that embed signed tags
object-file-convert: convert commit objects when writing
object-file-convert: don't leak when converting tag objects
object-file-convert: convert tag objects when writing
object-file-convert: add a function to convert trees between algorithms
object: factor out parse_mode out of fast-import and tree-walk into in object.h
cache: add a function to read an OID of a specific algorithm
tag: sign both hashes
commit: export add_header_signature to support handling signatures on tags
...
When a client sends us a "want" or "have" line, we call parse_object()
to get an object struct. If the object is a tree, then the parsed state
means that tree->buffer points to the uncompressed contents of the tree.
But we don't really care about it. We only really need to parse commits
and tags; for trees and blobs, the important output is just a "struct
object" with the correct type.
But much worse, we do not ever free that tree buffer. It's not leaked in
the traditional sense, in that we still have a pointer to it from the
global object hash. But if the client requests many trees, we'll hold
all of their contents in memory at the same time.
Nobody really noticed because it's rare for clients to directly request
a tree. It might happen for a lightweight tag pointing straight at a
tree, or it might happen for a "tree:depth" partial clone filling in
missing trees.
But it's also possible for a malicious client to request a lot of trees,
causing upload-pack's memory to balloon. For example, without this
patch, requesting every tree in git.git like:
pktline() {
local msg="$*"
printf "%04x%s\n" $((1+4+${#msg})) "$msg"
}
want_trees() {
pktline command=fetch
printf 0001
git cat-file --batch-all-objects --batch-check='%(objectname) %(objecttype)' |
while read oid type; do
test "$type" = "tree" || continue
pktline want $oid
done
pktline done
printf 0000
}
want_trees | GIT_PROTOCOL=version=2 valgrind --tool=massif ./git upload-pack . >/dev/null
shows a peak heap usage of ~3.7GB. Which is just about the sum of the
sizes of all of the uncompressed trees. For linux.git, it's closer to
17GB.
So the obvious thing to do is to call free_tree_buffer() after we
realize that we've parsed a tree. We know that upload-pack won't need it
later. But let's push the logic into parse_object_with_flags(), telling
it to discard the tree buffer immediately. There are two reasons for
this. One, all of the relevant call-sites already call the with_options
variant to pass the SKIP_HASH flag. So it actually ends up as less code
than manually free-ing in each spot. And two, it enables an extra
optimization that I'll discuss below.
I've touched all of the sites that currently use SKIP_HASH in
upload-pack. That drops the peak heap of the upload-pack invocation
above from 3.7GB to ~24MB.
I've also modified the caller in get_reference(); a partial clone
benefits from its use in pack-objects for the reasons given in
0bc2557951 (upload-pack: skip parse-object re-hashing of "want" objects,
2022-09-06), where we were measuring blob requests. But note that the
results of get_reference() are used for traversing, as well; so we
really would _eventually_ use the tree contents. That makes this at
first glance a space/time tradeoff: we won't hold all of the trees in
memory at once, but we'll have to reload them each when it comes time to
traverse.
And here's where our extra optimization comes in. If the caller is not
going to immediately look at the tree contents, and it doesn't care
about checking the hash, then parse_object() can simply skip loading the
tree entirely, just like we do for blobs! And now it's not a space/time
tradeoff in get_reference() anymore. It's just a lazy-load: we're
delaying reading the tree contents until it's time to actually traverse
them one by one.
And of course for upload-pack, this optimization means we never load the
trees at all, saving lots of CPU time. Timing the "every tree from
git.git" request above shows upload-pack dropping from 32 seconds of CPU
to 19 (the remainder is mostly due to pack-objects actually sending the
pack; timing just the upload-pack portion shows we go from 13s to
~0.28s).
These are all highly gamed numbers, of course. For real-world
partial-clone requests we're saving only a small bit of time in
practice. But it does help harden upload-pack against malicious
denial-of-service attacks.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
builtin/fast-import.c and tree-walk.c have almost identical version of
get_mode. The two functions started out the same but have diverged
slightly. The version in fast-import changed mode to a uint16_t to
save memory. The version in tree-walk started erroring if no mode was
present.
As far as I can tell both of these changes are valid for both of the
callers, so add the both changes and place the common parsing helper
in object.h
Rename the helper from get_mode to parse_mode so it does not
conflict with another helper named get_mode in diff-no-index.c
This will be used shortly in a new helper decode_tree_entry_raw
which is used to compute cmpatibility objects as part of
the sha256 transition.
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The object traversal using reachability bitmap done by
"pack-object" has been tweaked to take advantage of the fact that
using "boundary" commits as representative of all the uninteresting
ones can save quite a lot of object enumeration.
* tb/pack-bitmap-traversal-with-boundary:
pack-bitmap.c: use commit boundary during bitmap traversal
pack-bitmap.c: extract `fill_in_bitmap()`
object: add object_array initializer helper function
The object_array API has an OBJECT_ARRAY_INIT macro, but lacks a
function to initialize an object_array at a given location in memory.
Introduce `object_array_init()` to implement such a function.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
hash.h depends upon and includes repository.h, due to the definition and
use of the_hash_algo (defined as the_repository->hash_algo). However,
most headers trying to include hash.h are only interested in the layout
of the structs like object_id. Move the parts of hash.h that do not
depend upon repository.h into a new file hash-ll.h (the "low level"
parts of hash.h), and adjust other files to use this new header where
the convenience inline functions aren't needed.
This allows hash.h and object.h to be fairly small, minimal headers. It
also exposes a lot of hidden dependencies on both path.h (which was
brought in by repository.h) and repository.h (which was previously
implicitly brought in by object.h), so also adjust other files to be
more explicit about what they depend upon.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The object_type() inline function is very tied to the enum object_type
declaration within object.h, and just seemed to make more sense to live
there. That makes S_ISGITLINK and some other defines make sense to go
with it, as well as the create_ce_mode() and canon_mode() inline
functions. Move all these inline functions and defines from cache.h to
object.h.
Signed-off-by: Elijah Newren <newren@gmail.com>
Acked-by: Calvin Wan <calvinwan@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Things should be able to depend on object.h without pulling in all of
cache.h. Move an enum to allow this.
Note that a couple files previously depended on things brought in
through cache.h indirectly (revision.h -> commit.h -> object.h ->
cache.h). As such, this change requires making existing dependencies
more explicit in half a dozen files. The inclusion of strbuf.h in
some headers if of particular note: these headers directly embedded a
strbuf in some new structs, meaning they should have been including
strbuf.h all along but were indirectly getting the necessary
definitions.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The parse_object() function checks the object hash of any object it
parses. This is a nice feature, as it means we may catch bit corruption
during normal use, rather than waiting for specific fsck operations.
But it also can be slow. It's particularly noticeable for blobs, where
except for the hash check, we could return without loading the object
contents at all. Now one may wonder what is the point of calling
parse_object() on a blob in the first place then, but usually it's not
intentional: we were fed an oid from somewhere, don't know the type, and
want an object struct. For commits and trees, the parsing is usually
helpful; we're about to look at the contents anyway. But this is less
true for blobs, where we may be collecting them as part of a
reachability traversal, etc, and don't actually care what's in them. And
blobs, of course, tend to be larger.
We don't want to just throw out the hash-checks for blobs, though. We do
depend on them in some circumstances (e.g., rev-list --verify-objects
uses parse_object() to check them). It's only the callers that know
how they're going to use the result. And so we can help them by
providing a special flag to skip the hash check.
We could just apply this to blobs, as they're going to be the main
source of performance improvement. But if a caller doesn't care about
checking the hash, we might as well skip it for other object types, too.
Even though we can't avoid reading the object contents, we can still
skip the actual hash computation.
If this seems like it is making Git a little bit less safe against
corruption, it may be. But it's part of a series of tradeoffs we're
already making. For instance, "rev-list --objects" does not open the
contents of blobs it prints. And when a commit graph is present, we skip
opening most commits entirely. The important thing will be to use this
flag in cases where it's safe to skip the check. For instance, when
serving a pack for a fetch, we know the client will fully index the
objects and do a connectivity check itself. There's little to be gained
from the server side re-hashing a blob itself. And indeed, most of the
time we don't! The revision machinery won't open up a blob reached by
traversal, but only one requested directly with a "want" line. So
applied properly, this new feature shouldn't make anything less safe in
practice.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We have long allowed users to run e.g.
git log --ancestry-path master..seen
which shows all commits which satisfy all three of these criteria:
* are an ancestor of seen
* are not an ancestor of master
* have master as an ancestor
This commit allows another variant:
git log --ancestry-path=$TOPIC master..seen
which shows all commits which satisfy all of these criteria:
* are an ancestor of seen
* are not an ancestor of master
* have $TOPIC in their ancestry-path
that last bullet can be defined as commits meeting any of these
criteria:
* are an ancestor of $TOPIC
* have $TOPIC as an ancestor
* are $TOPIC
This also allows multiple --ancestry-path arguments, which can be
used to find commits with any of the given topics in their ancestry
path.
Signed-off-by: Elijah Newren <newren@gmail.com>
Acked-by: Derrick Stolee <derrickstolee@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Currently stash shells out to reflog in order to delete refs. In an
effort to reduce how much we shell out to a subprocess, libify the
functionality that stash needs into reflog.c.
Add a reflog_delete function that is pretty much the logic in the while
loop in builtin/reflog.c cmd_reflog_delete(). This is a function that
builtin/reflog.c and builtin/stash.c can both call.
Also move functions needed by reflog_delete and export them.
Helped-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: John Cai <johncai86@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In C it isn't required to specify that all members of a struct are
zero'd out to 0, NULL or '\0', just providing a "{ 0 }" will
accomplish that.
Let's also change code that provided N zero'd fields to just
provide one, and change e.g. "{ NULL }" to "{ 0 }" for
consistency. I.e. even if the first member is a pointer let's use "0"
instead of "NULL". The point of using "0" consistently is to pick one,
and to not have the reader wonder why we're not using the same pattern
everywhere.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In the original code from 08cdfb1337 (pack-objects --keep-unreachable,
2007-09-16), we add each object to the packing list with type
`obj->type`, where `obj` comes from `lookup_unknown_object()`. Unless we
had already looked up and parsed the object, this will be `OBJ_NONE`.
That's fine, since oe_set_type() sets the type_valid bit to '0', and we
determine the real type later on.
So the only thing we need from the object lookup is access to the
`flags` field so that we can mark that we've added the object with
`OBJECT_ADDED` to avoid adding it again (we can just pass `OBJ_NONE`
directly instead of grabbing it from the object).
But add_object_entry() already rejects duplicates! This has been the
behavior since 7a979d99ba (Thin pack - create packfile with missing
delta base., 2006-02-19), but 08cdfb1337 didn't take advantage of it.
Moreover, to do the OBJECT_ADDED check, we have to do a hash lookup in
`obj_hash`.
So we can drop the lookup_unknown_object() call completely, *and* the
OBJECT_ADDED flag, too, since the spot we're touching here is the only
location that checks it.
In the end, we perform the same number of hash lookups, but with the
added bonus that we don't waste memory allocating an OBJ_NONE object (if
we were traversing, we'd need it eventually, but the whole point of this
code path is not to traverse).
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In some cases it's useful for efficiency reasons to get the type of an
object before deciding whether to parse it, but we still want an object
struct. E.g., in reachable.c, bitmaps give us the type, but we just want
to mark flags on each object. Likewise, we may loop over every object
and only parse tags in order to peel them; checking the type first lets
us avoid parsing the non-tags.
But our lookup_blob(), etc, functions make getting an object struct
annoying: we have to call the right function for every type. And we
cannot just use the generic lookup_object(), because it only returns an
already-seen object; it won't allocate a new object struct.
Let's provide a function that dispatches to the correct lookup_*
function based on a run-time type. In fact, reachable.c already has such
a helper, so we'll just make that public.
I did change the return type from "void *" to "struct object *". While
the former is a clever way to avoid casting inside the function, it's
less safe and less informative to people reading the function
declaration.
The next commit will add a new caller.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The lookup_unknown_object() system is not often used and is somewhat
confusing. Let's try to explain it a bit more (which is especially
important as I'm adding a related but slightly different function in the
next commit).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git push" learns to discover common ancestor with the receiving
end over protocol v2.
* jt/push-negotiation:
send-pack: support push negotiation
fetch: teach independent negotiation (no packfile)
fetch-pack: refactor command and capability write
fetch-pack: refactor add_haves()
fetch-pack: refactor process_acks()
Currently, the packfile negotiation step within a Git fetch cannot be
done independent of sending the packfile, even though there is at least
one application wherein this is useful. Therefore, make it possible for
this negotiation step to be done independently. A subsequent commit will
use this for one such application - push negotiation.
This feature is for protocol v2 only. (An implementation for protocol v0
would require a separate implementation in the fetch, transport, and
transport helper code.)
In the protocol, the main hindrance towards independent negotiation is
that the server can unilaterally decide to send the packfile. This is
solved by a "wait-for-done" argument: the server will then wait for the
client to say "done". In practice, the client will never say it; instead
it will cease requests once it is satisfied.
In the client, the main change lies in the transport and transport
helper code. fetch_refs_via_pack() performs everything needed - protocol
version and capability checks, and the negotiation itself.
There are 2 code paths that do not go through fetch_refs_via_pack() that
needed to be individually excluded: the bundle transport (excluded
through requiring smart_options, which the bundle transport doesn't
support) and transport helpers that do not support takeover. If or when
we support independent negotiation for protocol v0, we will need to
modify these 2 code paths to support it. But for now, report failure if
independent negotiation is requested in these cases.
Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
All of the other lookup_foo() functions take a repository argument, but
lookup_unknown_object() was never converted, and it uses the_repository
internally. Let's fix that.
We could leave a wrapper that uses the_repository, but there aren't that
many calls, so we'll just convert them all. I looked briefly at each
site to see if we had a repository struct (besides the_repository) we
could pass, but none of them do (so this conversion to pass
the_repository is a pure noop in each case, though it does take us one
step closer to eventually getting rid of the_repository).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Allow callers to specify the repository to use. Rename the function to
repo_clear_commit_marks to document its new scope. No functional change
intended.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Instead of writing a new commit-graph in every 'git maintenance run
--auto' process (when maintenance.commit-graph.enalbed is configured to
be true), only write when there are "enough" commits not in a
commit-graph file.
This count is controlled by the maintenance.commit-graph.auto config
option.
To compute the count, use a depth-first search starting at each ref, and
leaving markers using the SEEN flag. If this count reaches the limit,
then terminate early and start the task. Otherwise, this operation will
peel every ref and parse the commit it points to. If these are all in
the commit-graph, then this is typically a very fast operation. Users
with many refs might feel a slow-down, and hence could consider updating
their limit to be very small. A negative value will force the step to
run every time.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When "fetch.writeCommitGraph" configuration is set in a shallow
repository and a fetch moves the shallow boundary, we wrote out
broken commit-graph files that do not match the reality, which has
been corrected.
* tb/fix-persistent-shallow:
commit.c: don't persist substituted parents when unshallowing
Since 37b9dcabfc (shallow.c: use '{commit,rollback}_shallow_file',
2020-04-22), Git knows how to reset stat-validity checks for the
$GIT_DIR/shallow file, allowing it to change between a shallow and
non-shallow state in the same process (e.g., in the case of 'git fetch
--unshallow').
However, when $GIT_DIR/shallow changes, Git does not alter or remove any
grafts (nor substituted parents) in memory.
This comes up in a "git fetch --unshallow" with fetch.writeCommitGraph
set to true. Ordinarily in a shallow repository (and before 37b9dcabfc,
even in this case), commit_graph_compatible() would return false,
indicating that the repository should not be used to write a
commit-graphs (since commit-graph files cannot represent a shallow
history). But since 37b9dcabfc, in an --unshallow operation that check
succeeds.
Thus even though the repository isn't shallow any longer (that is, we
have all of the objects), the in-core representation of those objects
still has munged parents at the shallow boundaries. When the
commit-graph write proceeds, we use the incorrect parentage, producing
wrong results.
There are two ways for a user to work around this: either (1) set
'fetch.writeCommitGraph' to 'false', or (2) drop the commit-graph after
unshallowing.
One way to fix this would be to reset the parsed object pool entirely
(flushing the cache and thus preventing subsequent reads from modifying
their parents) after unshallowing. That would produce a problem when
callers have a now-stale reference to the old pool, and so this patch
implements a different approach. Instead, attach a new bit to the pool,
'substituted_parent', which indicates if the repository *ever* stored a
commit which had its parents modified (i.e., the shallow boundary
prior to unshallowing).
This bit needs to be sticky because all reads subsequent to modifying a
commit's parents are unreliable when unshallowing. Modify the check in
'commit_graph_compatible' to take this bit into account, and correctly
avoid generating commit-graphs in this case, thus solving the bug.
Helped-by: Derrick Stolee <dstolee@microsoft.com>
Helped-by: Jonathan Nieder <jrnieder@gmail.com>
Reported-by: Jay Conrod <jayconrod@google.com>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The code to push changes over "dumb" HTTP had a bad interaction
with the commit reachability code due to incorrect allocation of
object flag bits, which has been corrected.
* bc/http-push-flagsfix:
http-push: ensure unforced pushes fail when data would be lost
The bit fields in struct object have an unfortunate layout. Here's what
pahole reports on x86_64 GNU/Linux:
struct object {
unsigned int parsed:1; /* 0: 0 4 */
unsigned int type:3; /* 0: 1 4 */
/* XXX 28 bits hole, try to pack */
/* Force alignment to the next boundary: */
unsigned int :0;
unsigned int flags:29; /* 4: 0 4 */
/* XXX 3 bits hole, try to pack */
struct object_id oid; /* 8 32 */
/* size: 40, cachelines: 1, members: 4 */
/* sum members: 32 */
/* sum bitfield members: 33 bits, bit holes: 2, sum bit holes: 31 bits */
/* last cacheline: 40 bytes */
};
Notice the 1+3+29=33 bits in bit fields and 28+3=31 bits in holes.
There are holes inside the flags bit field as well -- while some object
flags are used for more than one purpose, 22, 23 and 24 are still free.
Use 23 and 24 instead of 27 and 28 for TOPO_WALK_EXPLORED and
TOPO_WALK_INDEGREE. This allows us to reduce FLAG_BITS by one so that
all bitfields combined fit into a single 32-bit slot:
struct object {
unsigned int parsed:1; /* 0: 0 4 */
unsigned int type:3; /* 0: 1 4 */
unsigned int flags:28; /* 0: 4 4 */
struct object_id oid; /* 4 32 */
/* size: 36, cachelines: 1, members: 4 */
/* last cacheline: 36 bytes */
};
With this tight packing the size of struct object is reduced by 10%.
Other architectures probably benefit as well.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
