The `combine` filter takes the intersection of its children, that is:
objects are shown only when all child filters would admit the object.
The preceding patches added support for many individual filter types.
Enable users to compose these filters by implementing support for the
`combine` filter type.
Mapping intersection onto path_walk_info works because every supported
child filter is a monotonic restriction:
- `blob:none`, `tree:0` unconditionally clear `info->blobs` and (for
`tree:0`) `info->trees`; clearing an already-cleared flag is a
no-op.
- `object:type=X` is now expressed as an AND of each type flag with the
filtered type, so applying multiple such filters only refines the
existing set rather than overwrites it.
- `blob:limit=N` has to compose too: the intersection of "size < L1"
and "size < L2" is "size < min(L1, L2)".
Update the `LOFC_BLOB_LIMIT` handler to take the running minimum when
`info->blob_limit` is already set, so a combined filter with, e.g.,
both "blob:limit=10" and "blob:limit=5" produces a limit of 5
regardless of ordering.
- `sparse:oid` is left unchanged. A `combine` filter that includes a
`sparse:oid` is allowed at most once, since the existing handler
refuses to overwrite `info->pl`. Two `sparse:oid` filters in a single
`combine` would be unusual and are rejected with a warning, matching
the standalone `sparse:oid` behavior.
Implementation-wise, the existing `prepare_filters()` called
`list_objects_filter_release()` inside each case branch. That works fine for
top-level filters, but `combine` filters need to recurse over its child
filters without releasing each one in turn (since the parent's release
iterates the sub array). Split `prepare_filters()` into a recursive helper
that performs only the mutation, plus a thin wrapper that calls the helper
and then releases the top-level filter once.
The `LOFC_COMBINE` case in the helper just walks `sub_nr` and recurses;
child filters are released by the wrapper's single
`list_objects_filter_release()` call on the parent (which itself recursively
releases each sub-filter, the same way it always has).
If any sub-filter is unsupported (e.g. "tree:1", "sparse:<path>", or a
not-yet-supported choice), the recursion bubbles a failure up and the
existing pack-objects/backfill fallback paths kick in.
Add coverage in t6601:
- "combine:blob:none+tree:0" collapses to "tree:0"
- "combine:object:type=blob+blob:limit=3" yields only the blobs
smaller than three bytes
- "combine:object:type=blob+object:type=tree" intersects to empty
- "combine:tree:1+blob:none" reports the "tree:1" error.
Update Documentation/git-pack-objects.adoc to add combine to the
list of supported --filter forms.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The `object:type` filter accepts only objects of a single type; it is
the second member of the object-info-only filter family that bitmap
traversal already supports.
Like `blob:none` and `tree:0`, it can be evaluated with nothing more
than the object's type, which is exactly the granularity path-walk's
existing info->{commits,trees,blobs,tags} flags already control.
Map `LOFC_OBJECT_TYPE` in `prepare_filters()` by AND-ing each flag
against the filtered type. A single `object:type=X` filter
applied to the default info (all flags = 1) leaves `info->X = 1` and
all the others 0, which is what we want.
Using an AND rather than straight assignment prepares us for a
subsequent change to implement combined object filters.
The path-walk machinery is mostly already wired for the per-type
distinction:
- `walk_path()` calls `path_fn` for a batch only when the corresponding
`info->X` flag is set, so unwanted types are silently not reported.
- `add_tree_entries()` skips tree entries of type `OBJ_BLOB` when
`info->blobs` is unset, so we don't even allocate paths for them.
- The commit-walk loop short-circuits the root-tree fetch when
`!info->trees && !info->blobs`, so commit-only filters don't descend
into trees at all.
But there are a couple of side effects of the "trees off, blobs on" case
that need fixing:
1. 'setup_pending_objects()' previously skipped pending trees as soon
as `info->trees` was zero. For 'object:type=blob' the call site
needs those pending trees: a lightweight tag pointing to a tree, or
an annotated tag whose peeled target is a tree, can both reach
blobs that are otherwise unreachable from any commit's root tree.
Loosen the gate to "if (!info->trees && !info->blobs) continue" and
similarly retrieve the root_tree_list whenever either trees or
blobs are wanted.
2. The revision machinery's `handle_commit()` drops pending trees when
`revs->tree_objects` is zero (see the 'OBJ_TREE' handler in
revision.c), so by the time path-walk sees the pending list
after `prepare_revision_walk()` the tree-bearing pendings would
already be gone. Fix this by setting
revs->tree_objects = info->trees || info->blobs
so pending trees survive `prepare_revision_walk()` whenever we
need to walk into them. Path-walk still resets tree_objects to
zero immediately after `prepare_revision_walk()` returns, so the
rev-walk itself never enumerates trees redundantly with
path-walk's own descent.
Add coverage in t6601 for each of the four `object:type` values. The
'object:type=blob' test in particular asserts that file2 and child/file
(both reachable only through tag-pointed trees) show up in the output,
exercising the pending-tree fix.
Update Documentation/git-pack-objects.adoc to add object:type to
the list of supported --filter forms.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The `tree:0` object filter omits all trees and blobs from the result,
keeping only commits and tags. Consequently, this filter type should
has a fairly straightforward integration with path-walk, as the decision
to include an object depends only on its type and does not depend on any
path-sensitive state.
Mapping it onto `path_walk_info` is direct: set `info->trees = 0` and
`info->blobs = 0` in `prepare_filters()` when the `LOFC_TREE_DEPTH`
choice is requested with depth zero. The existing code already plumbs
those flags through the rest of the walk:
- 'walk_objects_by_path()' sets `revs->blob_objects = info->blobs` and
`revs->tree_objects = info->trees` before `prepare_revision_walk()`,
so the revision walk doesn't try to enumerate trees or blobs itself.
- The commit-walk loop short-circuits the root-tree fetch with
"if (!info->trees && !info->blobs) continue;", so we never even
look up the root tree, let alone descend into it.
- `setup_pending_objects()` skips pending trees and blobs based on
the same flags.
This means the path-walk doesn't allocate or expand any tree structures
at all under `tree:0`, which matches the intended behavior of the
filter.
However, this requires first fixing some issues with how the path-walk
API handles directly-requested trees _and_ trees requested through
lightweight tags. These changes create substantial updates to
t6601-path-walk.sh, which the previous change highlighted as a problem
by tagging otherwise-unreachable trees and having them not appear in the
output.
Non-zero tree-depth filters are not supported. Those depend on the depth
at which a tree is visited, which is a path-walk concept the filter
machinery doesn't currently share with the path-walk API. Reject them in
`prepare_filters()` with a helpful error and let pack-objects fall back
to the regular traversal, the same way it already does for unsupported
filters.
Add coverage in t6601 for both `--all` and a single-branch case to
confirm that no trees or blobs are emitted, and a separate test that
`tree:1` is rejected with the expected error message. Place the new
tests before "setup sparse filter blob" so they run on the original set
of refs, before the orphan branch that the sparse-tree tests create.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The --filter=sparse:<oid> option to 'git pack-objects' allows focusing
an object set to a sparse-checkout definition. This reduces the set of
matching blobs while retaining all reachable trees. No server currently
supports fetching with this filter because it is expensive to compute
and reachability bitmaps do not help without a significant effort to
extend the bitmap feature to store bitmaps for each supported sparse-
checkout definition.
Without focusing on serving fetches and clones with these filters, there
are still benefits that could be realized by making this faster. With
the sparse index, it's more realistic now than ever to be able to
operate a local clone that was bootstrapped by a packfile created with
a sparse filter, because the missing trees are not needed to move a
sparse-checkout from one commit to another or to view the history of any
path in scope. Such clones could perhaps be bootstrapped by partial
bundles.
Previously, constructing these sparse packs has been incredibly
computationally inefficient. The revision walk that explores which
objects are in scope spends a lot of time checking each object to see if
it matches the sparse-checkout patterns, causing quadratic behavior
(number of objects times number of sparse-checkout patterns). This
improves somewhat when using cone-mode sparse-checkout patterns that can
use hashtables and prefix matches to determine containment. However, the
check per object is still too expensive for most cases.
This is where the path-walk feature comes in. We can proceed as normal
by placing objects in bins by path and _then_ check a group of objects
all at once. Since sparse:<oid> only restricts blobs, the path-walk must
include all reachable trees while using the cone-mode patterns to skip
blobs at paths outside the sparse scope. This establishes a baseline for
a potential future "treesparse:<oid>" filter that would also restrict
trees, but introducing such a new filter is deferred to a later change.
The implementation here is focused around loading the sparse-checkout
patterns from the provided object ID and checking that the patterns are
indeed cone-mode patterns. We can then load the correct pattern list
into the path walk context and use the logic that already exists from
bff4555767 (backfill: add --sparse option, 2025-02-03), though that
feature loads sparse-checkout patterns from the worktree's local
settings and also restricts tree objects. We use a combination of errors
and warnings to signal problems during this load. The difference is that
errors are likely fatal for the non-path-walk version while the warnings
are probably just implementation details for the path-walk version and
the 'git pack-objects' command can fall back to the revision walk
version.
Now that the SEEN flag is deferred until after pattern checks (from the
previous commit), handle the case where a tree with a shared OID appears
at both an out-of-cone and in-cone path. When trees are not being pruned
(pl_sparse_trees == 0), the path-walk re-walks the tree at the in-cone
path so that in-cone blobs within it are discovered. The new tests in
t5317 and t6601 demonstrate this behavior and would fail without these
changes.
The performance test p5315 shows the impact of this change when using
sparse filters:
Test HEAD~1 HEAD
----------------------------------------------------------------------
5315.10: repack (sparse:oid) 77.98 77.47 -0.7%
5315.11: repack size (sparse:oid) 187.5M 187.4M -0.0%
5315.12: repack (sparse:oid, --path-walk) 77.91 31.41 -59.7%
5315.13: repack size (sparse:oid, --path-walk) 187.5M 161.1M -14.1%
These performance tests were run on the Git repository. The --path-walk
feature shows meaningful space savings (14% smaller for sparse packs)
and dramatic time savings (60% faster) by leveraging the path-walk's
ability to skip blobs outside the sparse scope.
Co-authored-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Taylor Blaue <me@ttaylorr.com>
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The path-walk API prunes trees and blobs when a sparse-checkout pattern
list is provided, which is the correct behavior for 'git backfill
--sparse' since it only needs to fill in objects at paths within the
sparse cone.
However, a future change will use the path-walk API with a sparse:<oid>
filter that restricts only blobs while retaining all reachable trees.
To support both behaviors, add a 'pl_sparse_trees' flag to
path_walk_info. When set (as in 'git backfill --sparse' and the
--stdin-pl test helper mode), the sparse patterns prune both trees and
blobs. When unset, only blobs are filtered and all trees are walked and
reported.
Additionally, move the SEEN flag assignment in add_tree_entries() to
after the sparse pattern and pathspec checks. Previously, SEEN was set
immediately upon discovering an object, before checking whether its path
matched the sparse patterns. When the same object ID appeared at
multiple paths (e.g. sibling directories with identical contents), the
first path to be visited would mark the object as SEEN. If that path was
outside the sparse cone, the object would be skipped there but also
never discovered at its in-cone path.
By deferring the SEEN flag until after the checks pass, objects that are
skipped due to sparse filtering remain discoverable at other paths where
they may be in scope.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Extend the path-walk API to handle the 'blob:limit=<size>' object
filter natively. This filter omits blobs whose size is equal to or
greater than the given limit, matching the semantics used by the
list-objects-filter machinery.
When revs->filter.choice is LOFC_BLOB_LIMIT, the prepare_filters()
method stores the limit value in info->blob_limit and clears the filter
from revs. If the limit is zero, this degenerates to blob:none (all
blobs excluded), so info->blobs is set to 0 instead.
During walk_path(), blob batches are filtered before being delivered to
the callback: each blob's size is checked via odb_read_object_info(),
and only blobs strictly smaller than the limit are included. Blobs whose
size cannot be determined (e.g. missing in a partial clone) are
conservatively included, matching the existing filter behavior. Empty
batches after filtering are skipped entirely.
The check for inclusion in the path batch looks a little strange at
first glance. We use odb_read_object_info() to read the object's size.
Based on all of the assumptions to this point, this _should_ return
OBJ_BLOB. Since we are focused on the size filter, we use a
short-circuited OR (||) to skip the size check if that method returns a
different object type.
Notice that this inspection of object sizes requires the content to be
present in the repository. The odb_read_object_info() call will download
a missing blob on-demand. This means that the use of the path-walk API
within 'git backfill' would not operate nicely with this filter type.
The intention of that command is to download missing blobs in batches.
Downloading objects one-by-one would go against the point. Update the
validation in 'git backfill' to add its own compatibility check on top
of path_walk_filter_compatible().
Add tests for blob:limit=0 (equivalent to blob:none) and blob:limit=3
(which exercises partial filtering within a batch where some blobs are
kept and others are excluded).
Co-authored-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The 'git pack-objects' command can opt-in to using the path-walk API for
scanning the objects. Currently, this option is dynamically disabled if
combined with '--filter=<X>', even when using a simple filter such as
'blob:none' to signal a blobless packfile. This is a common scenario for
repos at scale, so is worth integrating.
Also, users can opt-in to the '--path-walk' option by default through
the pack.usePathWalk=true config option. When using that in a blobless
partial clone, the following warning can appear even though the user did
not specify either option directly:
warning: cannot use --filter with --path-walk
Teach the path-walk API to handle the 'blob:none' object filter
natively. When revs->filter.choice is LOFC_BLOB_NONE, the path-walk
sets info->blobs to 0 (skipping all blob objects) and clears the
filter from revs so that prepare_revision_walk() does not reject the
configuration.
This check is implemented in the static prepare_filters() method, which
will simultaneously check if the input filters are compatible and will
make the appropriate mutations to the path_walk_info and filters if the
path_walk_info is non-NULL. This allows us to use this logic both in the
API method path_walk_filter_compatible() for use in
builtin/pack-objects.c and as a prep step in walk_objects_by_path().
Update the test helper (test-path-walk) to accept --filter=<spec>
as a test-tool option (before '--'), applying it to revs after
setup_revisions() to avoid the --objects requirement check. We can also
revert recent GIT_TEST_PACK_PATH_WALK overrides in t5620.
Also switch test-path-walk from REV_INFO_INIT with manual repo
assignment to repo_init_revisions(), which properly initializes
the filter_spec strbuf needed for filter parsing.
Add tests for blob:none with --all and with a single branch.
The performance test p5315 shows the impact of this change when using
blobless filters:
Test HEAD~1 HEAD
---------------------------------------------------------------------
5315.6: repack (blob:none) 13.53 13.87 +2.5%
5315.7: repack size (blob:none) 137.7M 137.8M +0.1%
5315.8: repack (blob:none, --path-walk) 13.51 23.43 +73.4%
5315.9: repack size (blob:none, --path-walk) 137.7M 115.2M -16.3%
These performance tests were run on the Git repository. The --path-walk
feature shows meaningful space savings (16% smaller for blobless packs)
at the cost of increased computation time due to the two compression
passes. This data demonstrates that the feature is engaged and provides
real compression benefits when --no-reuse-delta forces fresh deltas.
Co-Authored-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We are preparing to integrate the path-walk API with some --filter options
in 'git pack-objects', but there is a subtle issue that is revealed when
those are put together and the test suite is run with
GIT_TEST_PACK_PATH_WALK=1.
When a filter reduces the set of requested objects, this results in
filtering out directly-requested objects, such as in the download of needed
blobs in a blobless partial clone.
The root cause is that the scan of pending objects in the path-walk API
respects the filters set in the path_walk_info instead of overriding them
for pending objects.
We can tell that a path is part of the directly-referenced objects if its
path name starts with '/' (other paths, including root trees never have this
starting character). Create a path_is_for_direct_objects() to make this
meaning clear, especially as we add more references in the future as we
integrate the path-walk API with partial clone filter options.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The value of a wrong pointer variable was referenced in an error
message that reported that it shouldn't be NULL.
* yc/path-walk-fix-error-reporting:
path-walk: fix NULL pointer dereference in error message
Previously, walk_objects_by_path() silently ignored pathspecs containing
wildcards or magic by clearing them. This caused all blobs to be
downloaded regardless of the given pathspec. Wildcard pathspecs like
"d/file.*.txt" are useful for narrowing which blobs to process (e.g.,
during 'git backfill').
Support wildcard pathspecs by making two changes:
1. Add an 'exact_pathspecs' flag to path_walk_context. When the
pathspec has no wildcards or magic, set this flag and use the
existing fast-path prefix matching in add_tree_entries(). When
wildcards are present, skip that block since prefix matching
cannot handle glob patterns.
2. Add a match_pathspec() check in walk_path() to filter out blobs
whose full path does not match the pathspec. This provides the
actual blob-level filtering for wildcard pathspecs.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous change allowed specifying revision arguments over the 'git
backfill' command-line. This created the opportunity for restricting the
initial commit set by filtering the revision walk through a pathspec. Other
than filtering the commit set (and thereby the root trees), this did not
restrict the path-walk implementation of 'git backfill' and did not restrict
the blobs that were downloaded to only those matching the pathspec.
Update the path-walk API to accept certain kinds of pathspecs and to
silently ignore anything too complex, for now. We will update this in the
next change to properly restrict to even complex pathspecs.
The current behavior focuses on pathspecs that match paths exactly. This
includes exact filenames, including directory names as prefixes. Pathspecs
containing wildcards or magic are cleared so the path walk downloads all
blobs, as before.
The reason for this restriction is to allow for a faster execution by
pruning the path walk to only trees that could contribute towards one of
those paths as a parent directory.
The test directory 'd/f/' (next to 'd/file*.txt') was prepared in a
previous commit to exercise the subtlety in prefix matching.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When lookup_tree() or lookup_blob() cannot find a tree entry's object,
'o' is set to NULL via:
o = child ? &child->object : NULL;
The subsequent null-check catches this correctly, but then dereferences
'o' to format the error message:
error(_("failed to find object %s"), oid_to_hex(&o->oid));
This causes a segfault instead of the intended diagnostic output.
Fix this by using &entry.oid instead. 'entry' is the struct name_entry
populated by tree_entry() on each loop iteration and holds the OID of
the failing lookup -- which is exactly what the error should report.
This crash is reachable via git-backfill(1) when a tree entry's object
is absent from the local object database.
Signed-off-by: Yuvraj Singh Chauhan <ysinghcin@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Use the passed in repository instead of the implicit the_repository when
parsing the tree.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous change fixed a bug in 'git repack -adf --path-walk' that
was due to an update to how path lists are initialized and missing some
important cases when processing the pending objects.
This change takes the three critical places where path lists are
initialized and combines them into a static method. This simplifies the
callers somewhat while also helping to avoid a missed update in the
future.
The other places where a path list (struct type_and_oid_list) is
initialized is for the following "fixed" lists:
* Tag objects.
* Commit objects.
* Root trees.
* Tagged trees.
* Tagged blobs.
These lists are created and consumed in different ways, with only the
root trees being passed into the logic that cares about the
"maybe_interesting" bit. It is appropriate to keep these uses separate.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Users reported an issue where objects were missing from their local
repositories after a full repack using 'git repack -adf --path-walk'.
This was alarming and took a while to create a reproducer. Here, we fix
the bug and include a test case that would fail without this fix.
The root cause is that certain objects existed in the index and had no
second versions. These objects are usually blobs, though trees can be
included if a cache-tree exists. The issue is that the revision walk
adds these objects to the "pending" list and the path-walk API forgets
to mark the lists it creates at this point as "maybe_interesting". If
these paths only ever have a single version in the history of the repo
(including the current staged version) then the parent directory never
tries to add a new object to the list and mark the list as
"maybe_interesting". Thus, when walking the list later, the group is
skipped as it is expected that no objects are interesting. This happens
even when there are actually no UNINTERESTING objects at all! This is
based on the optimization enabled by the pack.useSparse=true config
option, which is the default.
Thus, we create a test case that demonstrates the many cases of this
issue for reproducibility:
1. File a/b/c has only one committed version.
2. Files a/i and x/y only exist as staged changes.
3. Tree x/ only exists in the cache-tree.
After performing a non-path-walk repack to force all loose objects into
packfiles, run a --path-walk repack followed by 'git fsck'. This fsck is
what fails with the following errors:
error: invalid object 100644 f2e41136... for 'a/b/c'
This is the dropped instance of the single-versioned a/b/c file.
broken link from tree cfda31d8...
to tree 3f725fcd...
This is the missing tree for the single-versioned a/b/ directory.
missing blob 0ddf2bae... (a/i)
missing blob 975fbec8... (x/y)
missing blob a60d869d... (file)
missing blob f2e41136... (a/b/c)
missing tree 3f725fcd... (a/b/)
dangling tree 5896d7e... (staged root tree)
Note that since the staged root tree is missing, the fsck output cannot
even report that the staged x/ tree is missing as well.
The core problem here is that the "maybe_interesting" member of 'struct
type_and_oid_list' is not initialized to '1'. This member was added in
6333e7ae0b (path-walk: mark trees and blobs as UNINTERESTING,
2024-12-20) in a way to help when creating packfiles for a small commit
range using the sparse path algorithm (enabled by pack.useSparse=true).
The idea here is that the list is marked as "maybe_interesting" if an
object is added that does not have the UNINTERESTING flag on it. Later,
this is checked again in case all objects in the list were marked
UNINTERESTING after that point in time. In this case, the algorithm
skips the list as there is no reason to visit it.
This leads to the problem where the "maybe_interesting" member was not
appropriately initialized when the list is created from pending objects.
Initializing this in the correct places fixes the bug.
To reduce risk of similar bugs around initializing this structure, a
follow-up change will make initializing lists use a shared method.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In preparation for allowing both the --shallow and --path-walk options
in the 'git pack-objects' builtin, create a new 'edge_aggressive' option
in the path-walk API. This option will help walk the boundary more
thoroughly and help avoid sending extra objects during fetches and
pushes.
The only use of the 'edge_hint_aggressive' option in the revision API is
within mark_edges_uninteresting(), which is usually called before
between prepare_revision_walk() and before visiting commits with
get_revision(). In prepare_revision_walk(), the UNINTERESTING commits
are walked until a boundary is found.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
One way to significantly reduce the cost of a Git clone and later fetches is
to use a blobless partial clone and combine that with a sparse-checkout that
reduces the paths that need to be populated in the working directory. Not
only does this reduce the cost of clones and fetches, the sparse-checkout
reduces the number of objects needed to download from a promisor remote.
However, history investigations can be expensive as computing blob diffs
will trigger promisor remote requests for one object at a time. This can be
avoided by downloading the blobs needed for the given sparse-checkout using
'git backfill' and its new '--sparse' mode, at a time that the user is
willing to pay that extra cost.
Note that this is distinctly different from the '--filter=sparse:<oid>'
option, as this assumes that the partial clone has all reachable trees and
we are using client-side logic to avoid downloading blobs outside of the
sparse-checkout cone. This avoids the server-side cost of walking trees
while also achieving a similar goal. It also downloads in batches based on
similar path names, presenting a resumable download if things are
interrupted.
This augments the path-walk API to have a possibly-NULL 'pl' member that may
point to a 'struct pattern_list'. This could be more general than the
sparse-checkout definition at HEAD, but 'git backfill --sparse' is currently
the only consumer.
Be sure to test this in both cone mode and not cone mode. Cone mode has the
benefit that the path-walk can skip certain paths once they would expand
beyond the sparse-checkout. Non-cone mode can describe the included files
using both positive and negative patterns, which changes the possible return
values of path_matches_pattern_list(). Test both kinds of matches for
increased coverage.
To test this, we can create a blobless sparse clone, expand the
sparse-checkout slightly, and then run 'git backfill --sparse' to see
how much data is downloaded. The general steps are
1. git clone --filter=blob:none --sparse <url>
2. git sparse-checkout set <dir1> ... <dirN>
3. git backfill --sparse
For the Git repository with the 'builtin' directory in the
sparse-checkout, we get these results for various batch sizes:
| Batch Size | Pack Count | Pack Size | Time |
|-----------------|------------|-----------|-------|
| (Initial clone) | 3 | 110 MB | |
| 10K | 12 | 192 MB | 17.2s |
| 15K | 9 | 192 MB | 15.5s |
| 20K | 8 | 192 MB | 15.5s |
| 25K | 7 | 192 MB | 14.7s |
This case matters less because a full clone of the Git repository from
GitHub is currently at 277 MB.
Using a copy of the Linux repository with the 'kernel/' directory in the
sparse-checkout, we get these results:
| Batch Size | Pack Count | Pack Size | Time |
|-----------------|------------|-----------|------|
| (Initial clone) | 2 | 1,876 MB | |
| 10K | 11 | 2,187 MB | 46s |
| 25K | 7 | 2,188 MB | 43s |
| 50K | 5 | 2,194 MB | 44s |
| 100K | 4 | 2,194 MB | 48s |
This case is more meaningful because a full clone of the Linux
repository is currently over 6 GB, so this is a valuable way to download
a fraction of the repository and no longer need network access for all
reachable objects within the sparse-checkout.
Choosing a batch size will depend on a lot of factors, including the
user's network speed or reliability, the repository's file structure,
and how many versions there are of the file within the sparse-checkout
scope. There will not be a one-size-fits-all solution.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This call to parse_tree() was flagged by Coverity for ignoring the
return value. But if we look a little further up the function, we can
see that there is already a call to parse_tree_gently(), and we'll
return early if that fails. So by this point the tree will always be
parsed, and the call is redundant.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The path-walk API currently uses a stack-based approach to recursing
through the list of paths within the repository. This guarantees that
after a tree path is explored, all paths contained within that tree path
will be explored before continuing to explore siblings of that tree
path.
The initial motivation of this depth-first approach was to minimize
memory pressure while exploring the repository. A breadth-first approach
would have too many "active" paths being stored in the paths_to_lists
map.
We can take this approach one step further by making sure that blob
paths are visited before tree paths. This allows the API to free the
memory for these blob objects before continuing to perform the
depth-first search. This modifies the order in which we visit siblings,
but does not change the fact that we are performing depth-first search.
To achieve this goal, use a priority queue with a custom sorting method.
The sort needs to handle tags, blobs, and trees (commits are handled
slightly differently). When objects share a type, we can sort by path
name. This will keep children of the latest path to leave the stack be
preferred over the rest of the paths in the stack, since they agree in
prefix up to and including a directory separator. When the types are
different, we can prefer tags over other types and blobs over trees.
This causes significant adjustments to t6601-path-walk.sh to rearrange
the order of the visited paths.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When the input rev_info has UNINTERESTING starting points, we want to be
sure that the UNINTERESTING flag is passed appropriately through the
objects. To match how this is done in places such as 'git pack-objects', we
use the mark_edges_uninteresting() method.
This method has an option for using the "sparse" walk, which is similar in
spirit to the path-walk API's walk. To be sure to keep it independent, add a
new 'prune_all_uninteresting' option to the path_walk_info struct.
To check how the UNINTERSTING flag is spread through our objects, extend the
'test-tool path-walk' command to output whether or not an object has that
flag. This changes our tests significantly, including the removal of some
objects that were previously visited due to the incomplete implementation.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The rev_info that is specified for a path-walk traversal may specify
visiting tag refs (both lightweight and annotated) and also may specify
indexed objects (blobs and trees). Update the path-walk API to walk
these objects as well.
When walking tags, we need to peel the annotated objects until reaching
a non-tag object. If we reach a commit, then we can add it to the
pending objects to make sure we visit in the commit walk portion. If we
reach a tree, then we will assume that it is a root tree. If we reach a
blob, then we have no good path name and so add it to a new list of
"tagged blobs".
When the rev_info includes the "--indexed-objects" flag, then the
pending set includes blobs and trees found in the cache entries and
cache-tree. The cache entries are usually blobs, though they could be
trees in the case of a sparse index. The cache-tree stores
previously-hashed tree objects but these are cleared out when staging
objects below those paths. We add tests that demonstrate this.
The indexed objects come with a non-NULL 'path' value in the pending
item. This allows us to prepopulate the 'path_to_lists' strmap with
lists for these paths.
The tricky thing about this walk is that we will want to combine the
indexed objects walk with the commit walk, especially in the future case
of walking objects during a command like 'git repack'.
Whenever possible, we want the objects from the index to be grouped with
similar objects in history. We don't want to miss any paths that appear
only in the index and not in the commit history.
Thus, we need to be careful to let the path stack be populated initially
with only the root tree path (and possibly tags and tagged blobs) and go
through the normal depth-first search. Afterwards, if there are other
paths that are remaining in the paths_to_lists strmap, we should then
iterate through the stack and visit those objects recursively.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We add the ability to filter the object types in the path-walk API so
the callback function is called fewer times.
This adds the ability to ask for the commits in a list, as well. We
re-use the empty string for this set of objects because these are passed
directly to the callback function instead of being part of the
'path_stack'.
Future changes will add the ability to visit annotated tags.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In anticipation of a few planned applications, introduce the most basic form
of a path-walk API. It currently assumes that there are no UNINTERESTING
objects, and does not include any complicated filters. It calls a function
pointer on groups of tree and blob objects as grouped by path. This only
includes objects the first time they are discovered, so an object that
appears at multiple paths will not be included in two batches.
These batches are collected in 'struct type_and_oid_list' objects, which
store an object type and an oid_array of objects.
The data structures are documented in 'struct path_walk_context', but in
summary the most important are:
* 'paths_to_lists' is a strmap that connects a path to a
type_and_oid_list for that path. To avoid conflicts in path names,
we make sure that tree paths end in "/" (except the root path with
is an empty string) and blob paths do not end in "/".
* 'path_stack' is a string list that is added to in an append-only
way. This stores the stack of our depth-first search on the heap
instead of using recursion.
* 'path_stack_pushed' is a strmap that stores path names that were
already added to 'path_stack', to avoid repeating paths in the
stack. Mostly, this saves us from quadratic lookups from doing
unsorted checks into the string_list.
The coupling of 'path_stack' and 'path_stack_pushed' is protected by the
push_to_stack() method. Call this instead of inserting into these
structures directly.
The walk_objects_by_path() method initializes these structures and
starts walking commits from the given rev_info struct. The commits are
used to find the list of root trees which populate the start of our
depth-first search.
The core of our depth-first search is in a while loop that continues
while we have not indicated an early exit and our 'path_stack' still has
entries in it. The loop body pops a path off of the stack and "visits"
the path via the walk_path() method.
The walk_path() method gets the list of OIDs from the 'path_to_lists'
strmap and executes the callback method on that list with the given path
and type. If the OIDs correspond to tree objects, then iterate over all
trees in the list and run add_children() to add the child objects to
their own lists, adding new entries to the stack if necessary.
In testing, this depth-first search approach was the one that used the
least memory while iterating over the object lists. There is still a
chance that repositories with too-wide path patterns could cause memory
pressure issues. Limiting the stack size could be done in the future by
limiting how many objects are being considered in-progress, or by
visiting blob paths earlier than trees.
There are many future adaptations that could be made, but they are left for
future updates when consumers are ready to take advantage of those features.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Taylor Blau