self.contains(other) means that the self tree contains the other tree (i.e.
the self path is prefix of the other), but it could be confused the other way
around if we were thinking about the path literal, not the tree. Let's add
.starts_with() instead by copying the std::path::Path definition.
This enables cheap str-to-RepoPath cast, which is useful when sorting and
filtering a large Vec<(String, _)> list by using matcher for example. It
will also eliminate temporary allocation by repo_path.parent().
I'm going to add borrowed RepoPath type, and most from_internal_string()
callers will be migrated to it. For the remaining callers, it makes more
sense to move the ownership of String to RepoPathBuf.
This follows up on 3967f63 (see that commit's description for more
motivation) and e79c8b6.
In a discussion linked below, it was decided that forbidding `-r --skip-empty`
entirely is preferable to the mixed behavior introduced in 3967f63.
3967f637dc (commitcomment-133539911)
RepoPath::from_components() is removed since it is no longer a primitive
function.
The components iterator could be implemented on top of str::split(), but
it's not as we'll probably want to add components.as_path() -> &RepoPath.
Tree walking and tree_states map construction get slightly faster thanks to
fewer allocations and/or better cache locality. If we add a borrowed RepoPath
type, we can also implement a cheap &str to &RepoPath conversion on top. Then,
we can get rid of BTreeMap<RepoPath, FileState> construction at all.
Snapshot without watchman:
```
% hyperfine --sort command --warmup 3 --runs 10 -L bin jj-0,jj-1 \
"target/release-with-debug/{bin} -R ~/mirrors/linux status"
Benchmark 1: target/release-with-debug/jj-0 -R ~/mirrors/linux status
Time (mean ± σ): 950.1 ms ± 24.9 ms [User: 1642.4 ms, System: 681.1 ms]
Range (min … max): 913.8 ms … 990.9 ms 10 runs
Benchmark 2: target/release-with-debug/jj-1 -R ~/mirrors/linux status
Time (mean ± σ): 872.1 ms ± 14.5 ms [User: 1922.3 ms, System: 625.8 ms]
Range (min … max): 853.2 ms … 895.9 ms 10 runs
Relative speed comparison
1.09 ± 0.03 target/release-with-debug/jj-0 -R ~/mirrors/linux status
1.00 target/release-with-debug/jj-1 -R ~/mirrors/linux status
```
Tree walk:
```
% hyperfine --sort command --warmup 3 --runs 10 -L bin jj-0,jj-1 \
"target/release-with-debug/{bin} -R ~/mirrors/linux files --ignore-working-copy"
Benchmark 1: target/release-with-debug/jj-0 -R ~/mirrors/linux files --ignore-working-copy
Time (mean ± σ): 375.3 ms ± 15.4 ms [User: 223.3 ms, System: 151.8 ms]
Range (min … max): 359.4 ms … 394.1 ms 10 runs
Benchmark 2: target/release-with-debug/jj-1 -R ~/mirrors/linux files --ignore-working-copy
Time (mean ± σ): 357.1 ms ± 16.2 ms [User: 214.7 ms, System: 142.6 ms]
Range (min … max): 341.6 ms … 378.9 ms 10 runs
Relative speed comparison
1.05 ± 0.06 target/release-with-debug/jj-0 -R ~/mirrors/linux files --ignore-working-copy
1.00 target/release-with-debug/jj-1 -R ~/mirrors/linux files --ignore-working-copy
```
The added test would fail if paths were purely ordered by concatenated strings.
I'm not sure if we want to preserve the current ordering, but let's not break
it for the moment.
This follows up on @matts1 's #2609.
We still allow the `-r` commit to become empty. I would be more comfortable if
there was a test for that, but I haven't done that (yet?) and it seems pretty
safe. If that's a problem, I'm happy to forbid `-r --skip-empty` entirely,
since it is far less useful than `-s --skip-empty` or `-b --skip-empty`.
I think it is undesired to abandon emptied descendants. As far as descendants
of `A` are concerned, `jj rebase -r A` should be equivalent to `jj abandon A`,
and `jj abandon` does not remove emptied commits. It also doesn't seem very
useful to do that, since I think descendant commits of an abandoned (or moved
with `-r`) commit only become empty in pathological cases.
Additionally, if we did want -r to empty descendants of `A`, we'd have to add
thorough tests and possibly improve the algorithm. I want to refactor `rebase
-r` and add features to it, and having to consider cases of commits becoming
abandoned makes everything harder.
For example, if we have
```
root -> A -> B -> C
```
and `jj rebase -r A -d C` empties commit `B` (or `C`), I do not know whether
the current algorithm will work correctly. It seems possible that it would, but
that depends on the fact that empty merge commits are not abandoned for
descendants. That seems dangerous to rely on without tests.
I hope (but can't promise) that in the near future, making DescendantRebaser
return more information should help make it possible to create such
functionality in a more robust way. I am likely to attempt this as part of
implementing `-r --after`.
This is a step towards introducing a borrowed RepoPath type. The current
RepoPath type is inefficient as each component String is usually short. We
could apply short-string optimization, but still each inlined component would
consume 24 bytes just for e.g. "src", and increase the chance of random memory
access. If the owned RepoPath type is backed by String, we can implement cheap
cast from &str to borrowed &RepoPath type.
Leading/trailing slashes would introduce a bit of complexity if we migrate
the backing type from Vec<String> to String. Empty components are okay, but
let's reject them as they are cryptic and invalid.
It seems generally useful to optimize revset expressions in
`evaluate_programmatic()` so the caller doesn't have to remember to do
it. It should generally be cheap to do so even if it's often not
needed.
`RevsetExpression::resolve()` is meant for programmatically created
expressions. In particular, it may not contain symbols. Let's try to
clarify that by renaming the function and documenting it.
While it got faster to build a large BTreeMap<RepoPath, _>, there's still
a measurable cost. Let's eliminate it if watchman is enabled and the working
copy is clean. Perhaps, we should introduce new serialization format that
supports instant loading and lookup, but this hack works for the moment.
I'm not sure if the new tree_state format should be flat (RepoPath, _) list,
or tree like the backend storage btw.
In my "linux" repo (watchman enabled):
% hyperfine --sort command --warmup 3 --runs 10 -L bin jj-0,jj-1 \
"target/release-with-debug/{bin} -R ~/mirrors/linux status"
Benchmark 1: target/release-with-debug/jj-0 -R ~/mirrors/linux status
Time (mean ± σ): 768.9 ms ± 14.2 ms [User: 630.7 ms, System: 131.2 ms]
Range (min … max): 742.3 ms … 783.1 ms 10 runs
Benchmark 2: target/release-with-debug/jj-1 -R ~/mirrors/linux status
Time (mean ± σ): 713.0 ms ± 16.8 ms [User: 587.9 ms, System: 116.2 ms]
Range (min … max): 681.5 ms … 731.1 ms 10 runs
Relative speed comparison
1.08 ± 0.03 target/release-with-debug/jj-0 -R ~/mirrors/linux status
1.00 target/release-with-debug/jj-1 -R ~/mirrors/linux status
This ensures that the root fsmonitor_matcher matches nothing if there are no
working-copy changes. The query result can be observed by "jj debug watchman
query-changed-files".
I don't have expertise on watchman query language, but using the watchman API
is probably better than .filter()-ing the result manually.
Suppose the input list is presorted, sorting a sorted vec would be cheaper
than .insert()-ing sorted items one by one.
In my "linux" repo (watchman eanbled):
- jj-0: baseline
- jj-1: previous (don't randomize by HashMap)
- jj-2: this
% hyperfine --sort command --warmup 3 --runs 10 -L bin jj-0,jj-1,jj-2 \
"target/release-with-debug/{bin} -R ~/mirrors/linux status"
Benchmark 1: target/release-with-debug/jj-0 -R ~/mirrors/linux status
Time (mean ± σ): 1.034 s ± 0.020 s [User: 0.881 s, System: 0.212 s]
Range (min … max): 1.011 s … 1.068 s 10 runs
Benchmark 2: target/release-with-debug/jj-1 -R ~/mirrors/linux status
Time (mean ± σ): 849.3 ms ± 13.8 ms [User: 710.7 ms, System: 199.3 ms]
Range (min … max): 821.7 ms … 870.2 ms 10 runs
Benchmark 3: target/release-with-debug/jj-2 -R ~/mirrors/linux status
Time (mean ± σ): 786.2 ms ± 16.7 ms [User: 650.7 ms, System: 204.1 ms]
Range (min … max): 760.8 ms … 805.2 ms 10 runs
Relative speed comparison
1.32 ± 0.04 target/release-with-debug/jj-0 -R ~/mirrors/linux status
1.08 ± 0.03 target/release-with-debug/jj-1 -R ~/mirrors/linux status
1.00 target/release-with-debug/jj-2 -R ~/mirrors/linux status
According to the doc, this is compatible with the map syntax.
https://protobuf.dev/programming-guides/proto3/#maps
This change means that the serialized file states are sorted by RepoPath,
so BTreeMap<RepoPath, _> can be reconstructed with fewer cache misses.
In my "linux" repo (watchman enabled):
- jj-0: baseline
- jj-1: this
% hyperfine --sort command --warmup 3 --runs 10 -L bin jj-0,jj-1,jj-2 \
"target/release-with-debug/{bin} -R ~/mirrors/linux status"
Benchmark 1: target/release-with-debug/jj-0 -R ~/mirrors/linux status
Time (mean ± σ): 1.034 s ± 0.020 s [User: 0.881 s, System: 0.212 s]
Range (min … max): 1.011 s … 1.068 s 10 runs
Benchmark 2: target/release-with-debug/jj-1 -R ~/mirrors/linux status
Time (mean ± σ): 849.3 ms ± 13.8 ms [User: 710.7 ms, System: 199.3 ms]
Range (min … max): 821.7 ms … 870.2 ms 10 runs
Relative speed comparison
1.32 ± 0.04 target/release-with-debug/jj-0 -R ~/mirrors/linux status
1.08 ± 0.03 target/release-with-debug/jj-1 -R ~/mirrors/linux status
Cache-misses got reduced:
% perf stat -e task-clock,cycles,instructions,cache-references,cache-misses \
-- ./target/release-with-debug/jj-0 -R ~/mirrors/linux --no-pager status
1,091.68 msec task-clock # 1.032 CPUs utilized
4,179,596,978 cycles # 3.829 GHz
6,166,231,489 instructions # 1.48 insn per cycle
134,032,047 cache-references # 122.776 M/sec
29,322,707 cache-misses # 21.88% of all cache refs
1.057474164 seconds time elapsed
0.897042000 seconds user
0.194819000 seconds sys
% perf stat -e task-clock,cycles,instructions,cache-references,cache-misses \
-- ./target/release-with-debug/jj-1 -R ~/mirrors/linux --no-pager status
927.05 msec task-clock # 1.083 CPUs utilized
3,451,299,198 cycles # 3.723 GHz
6,222,418,272 instructions # 1.80 insn per cycle
98,499,363 cache-references # 106.251 M/sec
11,998,523 cache-misses # 12.18% of all cache refs
0.855938336 seconds time elapsed
0.720568000 seconds user
0.207924000 seconds sys
The idea is the same as the heads_pos() change in 9832ee205d. While
IndexEntry::position() should be cheap, saving 20 bytes per entry appears to
improve the performance in mid-size repos.
In my "linux" repo:
revsets/all()
-------------
baseline 1.24 156.0±1.06ms
this 1.00 126.0±0.51ms
I don't see significant difference in small-sized repos like "jj" or "git".
IndexEntryByPosition isn't removed since it's still used by the revset engine.
This removes the last use of `ouroboros`. Since `TreeEntriesDirItem`
is only used in "legacy trees" (before tree-level conflicts), I didn't
bother to check the performance impact. I also didn't bother to check
the matcher before adding the entries to the list, instead leaving
that where it is in `Iterator::next()`.
This removes the last use of `ouroboros` in `merged_tree.rs`. The set
of conflicts to iterate is usually so small that I didn't bother
checking the performance impact.
While importing the `ouroboros` crate and the `aliasable` crate it
depends on, the "unsafe Rust reviewer" expressed some concern that
they contain a lot of unsafe code that's hard to review. We can avoid
the unsafe code altogether by making `TreeEntriesIterator` not
self-refential. Instead, we can collect the matching entries in an
individual tree up front. It does have some performance cost:
```
❯ hyperfine --warmup 3 --runs 30 \
'/tmp/jj-before --ignore-working-copy files -r v6.0' \
'/tmp/jj-after --ignore-working-copy files -r v6.0'
Benchmark 1: /tmp/jj-before --ignore-working-copy files -r v6.0
Time (mean ± σ): 461.4 ms ± 14.3 ms [User: 232.1 ms, System: 229.4 ms]
Range (min … max): 443.4 ms … 496.3 ms 30 runs
Benchmark 2: /tmp/jj-after --ignore-working-copy files -r v6.0
Time (mean ± σ): 482.0 ms ± 14.3 ms [User: 257.2 ms, System: 224.9 ms]
Range (min … max): 461.8 ms … 513.3 ms 30 runs
Summary
'/tmp/jj-before --ignore-working-copy files -r v6.0' ran
1.04 ± 0.04 times faster than '/tmp/jj-after --ignore-working-copy files -r v6.0'
```
I think that's acceptable.
This is much faster (maybe because of better cache locality?) Another option
is to use BTreeSet, but the BinaryHeap version is slightly faster.
"bench revset" result in my linux repo:
revsets/heads(tags())
---------------------
baseline 3.28 560.6±4.01ms
1 2.92 500.0±2.99ms
2 1.98 339.6±1.64ms
3 (this) 1.00 171.2±0.30ms
Apparently, IndexEntry::generation_number() isn't cheap probably because it
involves random access to larger memory region, and the u32 value might not
be aligned. Let's instead store the generation numbers in BinaryHeap.
Also, heads_pos() becomes slightly faster by keeping the BinaryHeap entries
small, so I've removed the IndexEntry at all.
This makes the default log and disambiguation revsets fast, which evaluate
'heads(immutable_heads())'.
"bench revset" result in my linux repo:
revsets/heads(tags())
---------------------
baseline 3.28 560.6±4.01ms
1 2.92 500.0±2.99ms
2 (this) 1.98 339.6±1.64ms
All callers just iterate over the parent entries.
"bench revset" result in my linux repo:
revsets/heads(tags())
---------------------
baseline 3.28 560.6±4.01ms
1 (this) 2.92 500.0±2.99ms
For loose refs, uninteresting directories can be just skipped. For packed refs,
gix will have to do binary search for each prefix to find the starting point.
Still it's better overall if the repository contains tons of refs/jj/keep refs.
With my linux repo containing ~5k loose jj refs, this saves ~40ms:
% hyperfine --warmup 3 --runs 10 \
"/tmp/jj-gix --ignore-working-copy git import -R ~/mirrors/linux" \
"/tmp/jj-gix-iter --ignore-working-copy git import -R ~/mirrors/linux"
Benchmark 1: /tmp/jj-gix --ignore-working-copy git import -R ~/mirrors/linux
Time (mean ± σ): 151.6 ms ± 11.4 ms [User: 38.8 ms, System: 111.6 ms]
Range (min … max): 129.8 ms … 159.5 ms 10 runs
Benchmark 2: /tmp/jj-gix-iter --ignore-working-copy git import -R ~/mirrors/linux
Time (mean ± σ): 109.9 ms ± 11.6 ms [User: 27.5 ms, System: 82.4 ms]
Range (min … max): 89.4 ms … 117.8 ms 10 runs
Gitoxide errors are boxed since there are various error types and they tend
to exceed the clippy size limit.
Apparently, gitoxide is faster than git2:
% hyperfine --warmup 3 --runs 10 \
"/tmp/jj-baseline --ignore-working-copy git import -R ~/mirrors/linux" \
"/tmp/jj-gix --ignore-working-copy git import -R ~/mirrors/linux"
Benchmark 1: /tmp/jj-baseline --ignore-working-copy git import -R ~/mirrors/linux
Time (mean ± σ): 205.4 ms ± 15.7 ms [User: 59.6 ms, System: 144.6 ms]
Range (min … max): 189.7 ms … 223.9 ms 10 runs
Benchmark 2: /tmp/jj-gix --ignore-working-copy git import -R ~/mirrors/linux
Time (mean ± σ): 176.2 ms ± 13.7 ms [User: 41.2 ms, System: 134.0 ms]
Range (min … max): 155.4 ms … 186.5 ms 10 runs
We need to .collect_vec() the parents iterator to temporary buffer since the
borrowed iterator can't be returned back to the dag_walk functions. Another
option is to clone op_store and parent ids to remove &self lifetime from the
iterator, but that also means a temporary Vec is created.
Unlike dfs_ok(), this function short-circuits at an Err as we use non-lazy
topo_order_forward() internally. I think that's good enough. If we implement
GC on operation log, deleted parents will be excluded (or mapped to tombstone)
by caller. An Err shouldn't mean it's GC-ed.
This unblocks the use of Result<T, E> in op.parents().
There are two ways to encode errors:
a. impl IntoIterator<Item = Result<T, E>>
b. Result<V, E> where V: FromIterator<Item = T>
I think (a) is more natural to algorithms like dfs(), which can process error
nodes transparently.
Still the caller might have to collect the source iterator to temporary Vec
to conform to the neighbors_fn signature. It's not easy for neighbors_fn to
return an iterator borrowing the input node. We already have GAT, but doesn't
have return-position impl Trait in trait yet.
Recognize signature metadata from git commit objects, implement
a basic version of that for the native backend.
Extract the signed data (a commit binary repr without the signature) to
be verified later.
Otherwise, ref updates would fail if we port git::export_refs() to gitoxide.
This change isn't strictly needed for the backend itself, but we'll reuse the
gix::Repository instance created by the backend when importing and exporting
Git refs.
GitBackend will use it to configure gix::Repository. I think UserSettings
is generally useful to pass store-specific parameters, so I've updated all
factory functions.
While the safe implementation is a bit more complex (and probably more branchy),
I don't think the runtime overhead would matter here. Let's remove one more
unsafe for better code maintainability.
We have a few places where we have a `MergedTreeValue` and need to
read the data associated with it so we can write to the working copy
or include it in a diff. Let's extract some of that shared logic to a
function so we can reuse it. I plan to use it for reading file
contents in advance while streaming a diff in `local_working_copy`
soon (and probably in `jj diff` thereafter), but I think it seems like
an improvement on its own.
I'd like to read N files ahead from the backend, to avoid serializing
too many server calls on backends that are backed by a server. Moving
the reads a little earlier is a little step towards that.
The `TreeState::write_*()` functions can now be made into free/static
functions if we prefer.
We no longer have "unsafe" in this function, so let's use the iterator API
instead of recursion. Apparently I haven't pushed this change before because
unsafe in .find_map() looked scary.
Remove a couple of unnecessary unsafes:
- The NonZeroUsize is a constant where the unwrap will optimize away
anyway and we don't have an unsafe without any good reason there :)
- The other two were simply not needed, lifetimes worked fine, maybe
Rust became better since that code was written? NLL? Anyway, they're
gone now
As discussed in Discord, it's less useful if remote_branches() included
Git-tracking branches. Users wouldn't consider the backing Git repo as
a remote.
We could allow explicit 'remote_branches(remote=exact:"git")' query by changing
the default remote pattern to something like 'remote=~exact:"git"'. I don't
know which will be better overall, but we don't have support for negative
patterns anyway.
It's super common that a Merge object holds a resolved value, so let's inline
up to 1 element. T of Merge<T> usually consists of a couple of pointer-sized
fields. I don't see any measurable speed up, but it's no worse than the
original.
Since the concurrent diff algorithm is significantly slower when using
the Git backend, I think we'll have to use switch between the two
algorithms depending on backend. Even if the concurrent version always
performed as well as the sequential version, exactly how concurrent it
should be probably still depends on the backend. This commit therefore
adds a function to the `Backend` trait, so each backend can say how
much concurrency they deal well with. I then use that number for
choosing between the sequential and concurrent versions in
`MergedTree::diff_stream()`, and also to decide the number of
concurrent reads to do in the concurrent version.
When diffing two trees, we currently start at the root and diff those
trees. Then we diff each subtree, one at a time, recursively. When
using a commit backend that uses remote storage, like our backend at
Google does, diffing the subtrees one at a time gets very slow. We
should be able to diff subtrees concurrently. That way, the number of
roundtrips to a server becomes determined by the depth of the deepest
difference instead of by the number of differing trees (times 2,
even). This patch implements such an algorithm behind a `Stream`
interface. It's not hooked in to `MergedTree::diff_stream()` yet; that
will happen in the next commit.
I timed the new implementation by updating `jj diff -s` to use the new
diff stream and then ran it on the Linux repo with `jj diff
--ignore-working-copy -s --from v5.0 --to v6.0`. That slowed down by
~20%, from ~750 ms to ~900 ms. Maybe we can get some of that
performance back but I think it'll be hard to match
`MergedTree::diff()`. We can decide later if we're okay with the
difference (after hopefully reducing the gap a bit) or if we want to
keep both implementations.
I also timed the new implementation on our cloud-based repo at
Google. As expected, it made some diffs much faster (I'm not sure if
I'm allowed to share figures).
I'm going to add a Merge method that removes negative/positive terms pair, and
swap_remove() is the easiest option. The order of the conflicted ref targets
doesn't matter.
Many callers use interleaved iterators, and recently-added serialization code
is built on top of that, so I think it's better to store terms in that format.
map() functions no longer use MergeBuilder as we know the mapped values are
ordered properly. flatten() and simplify() are reimplemented to work with the
interleaved values. The other changes are trivial.
This motivation for this is so we can easily skip calling the function
if the user has opted out of the propagation of abandoned commits we
usually do (#2504). However, it seems like a good piece of code to
extract regardless of that feature.
One less git2 API use in CLI.
The function name GitBackend::init_colocated() is a bit odd, but we need to
specify the work-tree path, not the ".git" repo path. So we can't eliminate
the notion of the working copy path anyway.
What make rebase_to_dest_parent a good candidate for jj_lib::rewrite module:
- It is used both in obslog and interdiff. It's a sign that it may be moved to a lower layer
- CommandError is returned by converting from TreeMergeError. Not explicitly.
- It only use jj_lib::rewrite fonctions.
This will make it a little faster to update the working copy at Google
once we've made `MergedTree::diff_stream()` fetch trees
concurrently. (It only makes it a little faster because we still fetch
files serially.)
I'm going to implement a `Stream`-based version optimized for
high-latency (RPC-based) commit backends. So far, that implementation
is about 20% slower in the Linux repo when running `jj diff
--ignore-working-copy -s --from v5.0 --to v6.0`. I think that's almost
only because the algorithm is different, not because it's async per
se.
This commit adds a `Stream`-based version of `MergedTree::diff()` that
just wraps the regular iterator in stream. I updated `jj diff` to use
it. I couldn't measure any difference on the command above in the
Linux repo. I think that means we can safely use the same
`Stream`-based interface regardless of backend, even if we end up
needing two different implementations of the `Stream`. We would then
be using the wrapped iterator from this commit for local backends, and
the new implementation for remote backends. But ideally we can make
the remote-friendly implementation fast enough that we don't need two
implementations.
During the transition to using more async code, I keep running into
https://github.com/rust-lang/futures-rs/issues/2090. Right now, I want
to convert `MergedTree::diff()` into a `Stream`. I don't want to
update all call sites at once, so instead I'm adding a
`MergedTree::diff_stream()` method, which just wraps
`MergedTree::diff()` in a `Stream. However, since the iterator is
synchronous, it needs to block on the async `Backend::read_tree()`
calls. If we then also block on the `Stream` in the CLI, we run into
the panic.
We had similar code in two places for restoring paths from one tree to
another. Let's reuse it instead.
I put the new function in the `rewrite` module. I'm not sure if that's
right place. Maybe it belongs in `tree`?
Since gix::Repository::config_snapshot() borrows the repo instance, it has to
be allocated in caller's stack. That's why GitBackend::git_config() is removed.
My gut feeling is that gitoxide aims to be more transparent than libgit2. We'll
need to know more about the underlying Git data model.
Random comments on gix API:
* gix::Repository provides API similar to git2::Repository, but has less
"convenient" functions. For example, we need to use .find_object() +
.try_to/into_<kind>() instead of .find_<kind>().
* gix::Object, Blob, etc. own raw data as bytes. gix::object and gix::objs
types provide high-level views on such data.
* Tree building is pretty low-level compared to git2.
* gix leverages bstr (i.e. bytes) extensively.
It's probably not difficult to migrate git::import/export_refs(). It might
help eliminate the startup overhead of libssl initialization. The gix-based
GitBackend appears to be a bit faster, but that wouldn't practically matter.
#2316
Otherwise, the initialized repo could have a different work-dir path than the
load()-ed one. libgit2 appears to do some normalization somewhere, but gix
won't.
This avoids https://github.com/rust-lang/futures-rs/issues/2090. I
don't think we need to worry about reading legacy conflicts
asynchronously - async is really only useful for Google's backend
right now, and we don't use the legacy format at Google. In
particular, I don't want `MergedTree::value()` to have to be async.
I want to fix error propagation before I start using async in this
code. This makes the diff iterator propagate errors from reading tree
objects.
Errors include the path and don't stop the iteration. The idea is that
we should be able to show the user an error inline in diff output if
we failed to read a tree. That's going to be especially useful for
backends that can return `BackendError::AccessDenied`. That error
variant doesn't yet exist, but I plan to add it, and use it in
Google's internal backend.
Reasons to introduce this alias:
* Reduces complexity of a type, to silence Clippy warnings in the
future if we use this type as a type parameter
* The type is used quite frequently, so it makes sense to have a name
for it
* It's easier to visually scan for the end of the type when you don't
have to match opening and closing angle brackets
I'm going to add `MergedTreeValue` as an alias for
`Merge<Option<TreeValue>>`, but we already have a type by that name in
`merged_tree`. This patch renames it away, to make room for the new
alias. I used `MergedTreeVal` for this borrowing version to be a bit
like how `str` is a borrowed version of `String`.
Since "jj git fetch --branch" supports glob patterns, users would expect that
"jj git push --branch glob:.." also works.
The error handling bits are copied from "branch" sub commands. We might want to
extract it to a common helper function, but I haven't figured out a reasonable
boundary point yet.
AFAICT, all callers of `Merge::to_file_merge()` are already well
prepared for working with executable files. It's called from these
places:
* `local_working_copy.rs`: Materialized conflicts are correctly
updated using `Merge::with_new_file_ids()`.
* `merge_tools/`: Same as above.
* `cmd_cat()`: We already ignore the executable bit when we print
non-conflicted files, so it makes sense to also ignore it for
conflicted files.
* `git_diff_part()`: We print all conflicts with mode "100644" (the
mode for regular files). Maybe it's best to use "100755" for
conflicts that are unambiguously executable, or maybe it's better to
use a fake mode like "000000" for all conflicts. Either way, the
current behavior seems fine.
* `diff_content()`: We use the diff content in various diff
formats. We could add more detail about the executable bits in some
of them, but I think the current output is fine. For example,
instead of our current "Created conflict in my-file", we could say
"Created conflict in executable file my-file" or "Created conflict
in ambiguously executable file my-file". That's getting verbose,
though.
So, I think all we need to do is to make `Merge::to_file_merge()` not
require its inputs to be non-executable.
Closes#1279.
Resolves states are most common and the current format is pretty
verbose. Let's print it as if `Merge` were an enum with `Resolved` and
`Conflicted` variants instead.
Since local/remote branches are now of different types, it doesn't make much
sense to dispatch merging through RefName. Let's add merge_<kind>() methods
instead.
MutableRepo handles merging of the other kind of refs internally, and the
merge function is short enough to inline. I also removed early returns since
most callers provide non-identical ref targets, and merge_ref_targets() should
be cheap if the inputs can be trivially merged.
This partially reverts the change in 30fb7995c2 "view: make local/remote
branches iterator yield RemoteRef instead of RefTarget." As I'm going to add
diff function for RemoteRef pairs, we'll need a generic version of merge-join
iterator anyway.
We need to let async-ness propagate up from the backend because
`block_on()` doesn't like to be called recursively. The conflict
materialization code is a good place to make async because it doesn't
depends on anything that isn't already async-ready.
It seems we'll end up using `block_on()` quite a bit, at least until
we're done transitioning to async, and the function name doesn't
conflict with anything else, so let's always import it when we need
it.
We can provide more actionable error message than "not fast-forwardable". If
the push was fast-forwardable, "jj branch track" should be able to merge the
remote branch without conflicts, so the added step would be minimal.
Although this is logically correct, the error message is a bit cryptic. It's
probably better to reject push if non-tracking remote branches exist.
#1136
We'll use remote_ref.tracking_target() to classify push action, but not all
callers of local_remote_branches() need tracking_target() instead of target.
This means that the commits previously pinned by remote branches are no longer
abandoned. I think that's more correct since "push" is the operation to
propagate local view to remote, and uninteresting commits should have been
locally abandoned.
Since I'm going to make git::push_branches() update the repo view internally,
it should fail fast if the remote name is reserved. Before, the problem was
detected on git::import_refs().
Since pushed remote branches will share the common base targets with locals,
these branches should be marked as tracking. git::push_branches() will handle
that. It looks ugly that the public GitBranchPushTargets type keeps "force"-d
branches as a separate set, but we'll need to rework that anyway when we
implement --force-with-lease behavior. So let's leave it for now.
Some of the git::push_updates() tests have been migrated to the new function.
I left a couple of basic tests for git::push_updates() because push_updates()
will be used to implement a low-level "jj git push-refs" command.
I made import_refs() not preserve commits referenced by remote branches at
520f692a46 "git: on import_refs(), don't preserve old branches referenced by
remote refs." The idea is that remote branches are weak, and commits referenced
by these refs can be freely rewritten by future local changes without moving
the refs. I don't think that's wrong, but 520f692a46 also made "new" remote
changes be abandoned by old remote refs. This problem occurs only when
git.auto-local-branch is off.
I think there are two ways to fix the problem:
a. pin non-tracking remote branches just like local refs
b. pin newly fetched refs in addition to local refs
This patch implements (b) because it's simpler and more obvious that the
fetched commits would never be abandoned immediately.
This add support for custom `jj` binaries to use custom working-copy
backends. It works in the same way as with the other backends, i.e. we
write a `.jj/working_copy/type` file when the working copy is
initialized, and then we let that file control which implementation to
use (see previous commit).
I included an example of a (useless) working-copy implementation. I
hope we can figure out a way to test the examples some day.
This makes `Workspace::load()` look a new `.jj/working_copy/type` file
in order to load the right working copy implementation, just like
`Repo::load()` picks the right backends based on `.jj/store/type`,
`.jj/op_store/type`, etc. We don't write the file yet, and we don't
have a way of adding alternative working copy implementations, so it
will always be `LocalWorkingCopy` for now.
Our internal working copy implementations at Google will need the
commit so they can walk history backwards until they get to a "public"
commit. They'll then use that to tell build tools and virtual file
systems to present that as a base.
I'm not sure if we'll need to update `reset()` too. It's currently
only used by `jj untrack`, which doesn't change the commit's parent,
so it wouldn't affect any history walks.
`ReadonlyRepo::init()` takes callbacks for initializing each kind of
backend. We called these things like `op_store_initializer`. I found
that confusing because it is not a `OpStoreFactory` (which is for
loading an existing backend). This patch tries to clarify that by
renaming the arguments and adding types for each kind of callback
function.
This patch adds MutableRepo::track_remote_branch() as we'll probably need to
track the default branch on "jj git clone". untrack_remote_branch() is also
added for consistency.
We could instead migrate the storage types to (local_branches, remote_views),
but that would be more involved and break forward compatibility with little
benefit. Maybe we can do that later when we introduce remote tags.
The state field isn't saved yet. git import/export code paths are migrated,
but new tracking state is always calculated based on git.auto-local-branch
setting. So the tracking state is effectively a global flag.
As we don't know whether the existing remote branches have been merged in to
local branches, we assume that remote branches are "tracking" so long as the
local counterparts exist. This means existing locally-deleted branch won't
be pushed without re-tracking it. I think it's rare to leave locally-deleted
branches for long. For "git.auto-local-branch = false" setup, users might have
to untrack branches if they've manually "merged" remote branches and want to
continue that workflow. I considered using git.auto-local-branch setting in the
migration path, but I don't think that would give a better result. The setting
may be toggled after the branches got merged, and I'm planning to change it
default off for better Git interop.
Implementation-wise, the state enum can be a simple bool. It's enum just
because I originally considered to pack "forgotten" concept into it. I have
no idea which will be better for future extension.
It's going to be easier to define a `LockedWorkingCopy` trait if it
doesn't need to borrow from `WorkingCopy`, so let's remove the
reference we currently have and have
`LockedLocalWorkingCopy::finish()` return the new `LocalWorkingCopy`
instead.
I think the main disadvantage is that we now have to remember to
replace the old `LocalWorkingCopy` instance by the new one, whereas
the compiler would remind us before this commit. We could make
`start_modification()` take an owned `self`, but that would be a bit
annoying to work with when we have the instance stored in a field.
I'm about to make `LockedLocalWorkingCopy` not borrow from
`LocalWorkingCopy`. That will make it easier to forget to update any
`LocalWorkingCopy` variables when the modifications have been
committed. This patch introduces a wrapper around
`LockedLocalWorkingCopy` to help prevent that.
Thanks to Yuya for the suggestion.
`LocalWorkingCopy::check_out()` can be expressed using the planned
`WorkingCopy` trait, so it doesn't need to be in the trait itself
`WorkingCopy`. I wasn't sure if I should make it a free function in
`working_copy`, but I ended up moving it onto `Workspace`.
Since set_remote_branch_target() is called while merging refs, its tracking
state shouldn't be reinitialized. The other callers are migrated to new setter
to keep the story simple.
This isn't important, but I'm going to change remote_targets to store RemoteRef
instead of RefTarget, so I went ahead and change the other field types as well.
We could fix do_git_clone() instead, but it seemed a bit weird that the
git_repo_path is relative to the store path which is unknown to callers.
Fixes#2374
There's a subtle behavior change. Unlike the original remove_remote_branch(),
remote_views entry is not discarded when the branches map becomes empty. The
reasoning here is that the remote view can be added/removed when the remote
is added/removed respectively, though that's not implemented yet. Since the
serialized data cannot represent an empty remote, such view may generate
non-unique content hash.
These functions depend heavily on the underlying data structure, and I haven't
decided abstract View API to access to per-remote data types. Let's use the
underlying data type for now.
I'm planning to add support for untracked remote branches, and under that
model, there will be many remote branches without local counterparts. That's
the main reason why remote branches are grouped by remote, not by branch name.
The added helper functions will be used by simple_op_store and view.
get_branch() would need to reconstruct the remote_targets map if we migrate
the underlying data structure to per-remote views. Let's remove the method as
it is only used in tests.
It seems pretty clear from the context. Turns out we only use the
function in a test case. Maybe we don't even need it. It's easy to
provide it, though.
The `TreeStateError` type is specific to the current local-disk
working-copy backend, so it should not be part of the generic
working-copy interface I'm trying to create.
I think some of the errors variants in `CheckoutError` are too
specific to the local-disk implementation. Let's merge them and make
them less specific, so it's easier to define a reasonable trait for
the working copy.
As I'm going to split branches into per-remote map, .get_branch(name) will need
to gather remote branches by name to construct remote_targets map. Let's instead
iterate local/remote branches separately. I also migrated diffing of the other
kinds of refs to filter out unchanged entries upfront.
As we now diff incoming git refs against our known remote branches, the problem
described in the comment no longer occurs. test_branch_forget_fetched_branch()
passes, and the inline comments in the test are still valid.
As we need to build a set of all branch names anyway, we can also put old/new
targets there. InvalidGitName is moved to caller since the diff function no
longer converts RefName to "refs/" string.
The idea is that the "remote" refs could have been "op restore"-d whereas
view.git_refs() will never be. The next commit will update known_remote_refs
to be constructed from the current remote branches.
Instead of building these lists in a single loop, we could load new git_refs
to the view first, and then build diffs of the remote refs. I considered that,
but I feel it would be a bit awkward to update refs before importing commits
to the view.
The "remote" refs are stored in BTreeMap since merging order should be stable.
As I'm going to add separate lists of changed git_refs/remote_refs, it'll
become a bit unclear which one we should check for reserved remotes. The
diff might also be reorganized as a list of (remote, name, kind, old_target,
new_target) where remote == "git" means the git-tracking branch. In this
data structure, the notion of reserved remote name would be lost.
I'm going to rewrite `TreeDiffIterator` to fetch one level (depth) of
the tree at a time and concurrently. One step towards that is to
convert the iterator to a `Stream`. I'd like to do that by making the
current `Iterator` implementation call the new `Stream`
implementation. However, we can't call `futures::executor::block_on()`
on a future that itself calls `futures::executor::block_on()` (as
`Store::read_tree()` does), so the first step is to bubble up the
async-ness a bit. This patch does that by fetching both sides of the
diff concurrently. That should give close to a 2x speedup on
high-latency backends. (It doesn't help with our backend at Google,
however, because we have a daemon process that does some speculative
prefetching that usually downloads the child trees anyway.)
`futures::stream::Stream::collect()` requires a collection that
implements `Default` and `Extend`, and I would like to to be able to
collect a stream of trees.
The commit backend at Google is cloud-based (and so are the other
backends); it reads and writes commits from/to a server, which stores
them in a database. That makes latency much higher than for disk-based
backends. To reduce the latency, we have a local daemon process that
caches and prefetches objects. There are still many cases where
latency is high, such as when diffing two uncached commits. We can
improve that by changing some of our (jj's) algorithms to read many
objects concurrently from the backend. In the case of tree-diffing, we
can fetch one level (depth) of the tree at a time. There are several
ways of doing that:
* Make the backend methods `async`
* Use many threads for reading from the backend
* Add backend methods for batch reading
I don't think we typically need CPU parallelism, so it's wasteful to
have hundreds of threads running in order to fetch hundreds of objects
in parallel (especially when using a synchronous backend like the Git
backend). Batching would work well for the tree-diffing case, but it's
not as composable as `async`. For example, if we wanted to fetch some
commits at the same time as we were doing a diff, it's hard to see how
to do that with batching. Using async seems like our best bet.
I didn't make the backend interface's write functions async because
writes are already async with the daemon we have at Google. That
daemon will hash the object and immediately return, and then send the
object to the server in the background. I think any cloud-based
solution will need a similar daemon process. However, we may need to
reconsider this if/when jj gets used on a server with a custom backend
that writes directly to a database (i.e. no async daemon in between).
I've tried to measure the performance impact. That's the largest
difference I've been able to measure was on `jj diff
--ignore-working-copy -s --from v5.0 --to v6.0` in the Linux repo,
which increases from 749 ms to 773 ms (3.3%). In most cases I've
tested, there's no measurable difference. I've tried diffing from the
root commit, as well as `jj --ignore-working-copy log --no-graph -r
'::v3.0 & author(torvalds)' -T 'commit_id ++ "\n"'` (to test a
commit-heavy load).
This effectively undoes d8a313cdd4, which is no longer needed since
we just changed that error handling. It should make it easier to share
some of the current if/else blocks.
Before this patch, when updating to a commit that has a file that's
currently an ignored file on disk, jj would crash. After this patch,
we instead leave the conflicting files or directories on disk. We
print a helpful message about how to inspect the differences between
the intended working copy and the actual working copy, and how to
discard the unintended changes.
Closes#976.
I'm about to add handling of parent dirs that are existing ignored
files, so it's better to have it in one place. The only functional
difference should be that we now create parent directories for git
submodules. I don't think that matters.
It's about time we make the working copy a pluggable backend like we
have for the other storage. We will use it at Google for at least two
reasons:
* To support our virtual file system. That will be a completely
separate working copy backend, which will interact with the virtual
file system to update and snapshot the working copy.
* On local disk, we need to tell our build system where to find the
paths that are not in the sparse patterns. We plan to do that by
wrapping the standard local working copy backend (the one moved in
this commit), writing a symlink that points to the mainline commit
where the "background" files can be read from.
Let's start by renaming the exising implementation to
`local_working_copy`.
I've added a boolean flag to the store to ensure that the migration never runs
more than once after the view gets "op restore"-d. I'll probably reorganize the
branches structure to support non-tracking branches later, but updating the
storage format in a single commit would be too involved.
If jj is downgraded, these "git" remote refs would be exported to the Git repo.
Users might have to remove them manually.
I'm going to migrate "refs/heads/" branches to .remote_targets["git"]. This
commit will simplify the story as we won't have to exclude "refs/remotes/git/"
refs when diffing or renaming/removing remote.
Since both has_id() and resolve_prefix() do binary search, their costs are
practically the same. I think has_id() would complete with fewer ops, but such
level of optimization wouldn't be needed here. More importantly, this ensures
that unreachable commits aren't imported by GitBackend::read_commit().
One problematic scenario is that we have commits imported by old jj, and all
of their descendant commits are created by jj. Therefore import_head_commits()
wouldn't reach the old ancestor commits.
This change might bury a real bug, but I don't have a better alternative. Maybe
we can remove this hack after a couple of jj releases, and add a debug command
that imports all reachable Git commits from all historical heads.
Closes#2343
As we can set HEAD to an arbitrary ref by using .reference_symbolic(), we don't
have to manage a ref that can also be valid as a branch name.
Fixes#1495
I'll add a workaround for the root parent issue #1495 there. We can pass in
the wc parent id instead of the wc_commit object, but we might want to use
wc_commit.id() to generate a unique placeholder ref name.
While debugging git issues, I often ended up creating a deadlock by adding
debug prints. It's also not obvious that git::export_refs() works even if the
git_repo() has already been locked, whereas git::import_refs() wouldn't. Let's
consolidate lock handling to the backend implementation.
