OXYGENHUB
# c3d2ef6104c2 oxygen / crates / oxy-core / src / repo / git.rs
74.4 KiB · Rust · last changed Git import: a large merge renders its tree in batches 1 month ago source blame history raw
79ab8528 Andy McCall ↶ 1 
//! Git bridge (spec §23), export side: each state the current bookmark
2 
//! passed through becomes a Git commit through `git fast-import`, with the
3 
//! state's full tree and the previous state's commit as parent. A state
4 
//! that added exactly one change carries that change's message, author,
5 
//! and `Change-Id`/`Oxy-Hash` trailers; every commit carries `Oxy-State`.
6 
//!
7 
//! Marks make re-exports incremental: `.oxy/git/<target>/git-marks` is
8 
//! git's own marks file and `oxy-marks` maps state refs to marks.
9
10 
use super::*;
11 
use std::process::{Command, Stdio};
62794ecc Andy McCall ↶ 12 
use std::sync::Arc;
79ab8528 Andy McCall ↶ 13
14 
#[derive(Debug, Clone, Default, PartialEq, Eq)]
c3069537 Andy McCall ↶ 15 
pub struct GitImportReport {
16 
    pub commits: usize,
17 
    pub merges: usize,
4dee277c Andy McCall ↶ 18 
    /// Gitlink entries imported as subrepo nodes with Git pins.
19 
    pub submodules: usize,
c3069537 Andy McCall ↶ 20 
    pub heads: usize,
21 
}
22
23 
/// A running `git cat-file --batch`, for reading many blobs cheaply.
24 
struct CatFile {
25 
    child: std::process::Child,
26 
    stdin: std::process::ChildStdin,
27 
    stdout: std::io::BufReader<std::process::ChildStdout>,
28 
}
29
30 
impl CatFile {
31 
    fn start(git_dir: &Path) -> Result<CatFile> {
32 
        let mut child = Command::new("git")
33 
            .arg("-C")
34 
            .arg(git_dir)
35 
            .args(["cat-file", "--batch"])
36 
            .stdin(Stdio::piped())
37 
            .stdout(Stdio::piped())
38 
            .stderr(Stdio::null())
39 
            .spawn()
40 
            .map_err(|e| RepoError::Msg(format!("running git cat-file: {e}")))?;
41 
        let stdin = child.stdin.take().unwrap();
42 
        let stdout = std::io::BufReader::new(child.stdout.take().unwrap());
75167672 Andy McCall 43 
        Ok(CatFile {
44 
            child,
45 
            stdin,
46 
            stdout,
47 
        })
c3069537 Andy McCall ↶ 48 
    }
49
50 
    fn read(&mut self, sha: &str) -> Result<Vec<u8>> {
51 
        use std::io::{BufRead, Read};
52 
        writeln!(self.stdin, "{sha}")?;
53 
        self.stdin.flush()?;
54 
        let mut header = String::new();
55 
        self.stdout.read_line(&mut header)?;
56 
        let parts: Vec<&str> = header.trim_end().split(' ').collect();
57 
        if parts.len() != 3 {
58 
            return fail(format!("git cat-file: object {sha} not found"));
59 
        }
75167672 Andy McCall 60 
        let size: usize = parts[2]
61 
            .parse()
62 
            .map_err(|_| RepoError::Msg("git cat-file: bad size".into()))?;
c3069537 Andy McCall ↶ 63 
        let mut body = vec![0u8; size];
64 
        self.stdout.read_exact(&mut body)?;
65 
        let mut nl = [0u8; 1];
66 
        self.stdout.read_exact(&mut nl)?;
67 
        Ok(body)
68 
    }
69 
}
70
71 
impl Drop for CatFile {
72 
    fn drop(&mut self) {
73 
        let _ = self.child.kill();
74 
        let _ = self.child.wait();
75 
    }
76 
}
77
2306a4d2 Andy McCall ↶ 78 
/// One path a commit changes against its first parent: its new mode and
79 
/// blob, or `None` when the commit deletes it.
80 
type TreeDiff = Vec<(String, Option<(String, String)>)>;
81
82 
/// A running `git diff-tree --stdin`, fed every `commit parent` pair up
83 
/// front, whose diffs are read in the same order as they are needed: a
84 
/// commit's tree is its first parent's plus its diff, at a cost that
85 
/// follows what it changes rather than the whole tree (oxygen#25).
86 
struct DiffTree {
87 
    child: std::process::Child,
88 
    stdout: std::io::BufReader<std::process::ChildStdout>,
89 
    feeder: Option<std::thread::JoinHandle<()>>,
90 
    /// The header read at the end of the previous diff.
91 
    next: Option<String>,
92 
}
93
94 
impl DiffTree {
95 
    fn start(git_dir: &Path, pairs: Vec<(String, String)>) -> Result<DiffTree> {
96 
        let mut child = Command::new("git")
97 
            .arg("-C")
98 
            .arg(git_dir)
99 
            .args([
100 
                "diff-tree",
101 
                "--stdin",
102 
                "--always",
103 
                "-r",
104 
                "-z",
105 
                "--no-renames",
106 
            ])
107 
            .stdin(Stdio::piped())
108 
            .stdout(Stdio::piped())
109 
            .stderr(Stdio::null())
110 
            .spawn()
111 
            .map_err(|e| RepoError::Msg(format!("running git diff-tree: {e}")))?;
112 
        let mut stdin = child.stdin.take().unwrap();
113 
        let stdout = std::io::BufReader::new(child.stdout.take().unwrap());
114 
        // A thread writes, so git never waits on a full pipe for us to read.
115 
        let feeder = std::thread::spawn(move || {
116 
            let mut w = std::io::BufWriter::new(&mut stdin);
117 
            for (c, p) in pairs {
118 
                if writeln!(w, "{c} {p}").is_err() {
119 
                    return;
120 
                }
121 
            }
122 
            let _ = w.flush();
123 
        });
124 
        Ok(DiffTree {
125 
            child,
126 
            stdout,
127 
            feeder: Some(feeder),
128 
            next: None,
129 
        })
130 
    }
131
132 
    fn field(&mut self) -> Result<Option<String>> {
133 
        use std::io::BufRead;
134 
        let mut buf = Vec::new();
135 
        if self.stdout.read_until(0, &mut buf)? == 0 {
136 
            return Ok(None);
137 
        }
138 
        if buf.last() == Some(&0) {
139 
            buf.pop();
140 
        }
141 
        Ok(Some(String::from_utf8_lossy(&buf).into_owned()))
142 
    }
143
144 
    /// The diff of `sha` against the parent it was fed with.
145 
    fn diff(&mut self, sha: &str) -> Result<TreeDiff> {
146 
        let header = match self.next.take() {
147 
            Some(h) => Some(h),
148 
            None => self.field()?,
149 
        };
150 
        if header.as_deref() != Some(sha) {
151 
            return fail(format!("git diff-tree: expected {sha}, got {header:?}"));
152 
        }
153 
        let mut out = Vec::new();
154 
        while let Some(f) = self.field()? {
155 
            let Some(meta) = f.strip_prefix(':') else {
156 
                self.next = Some(f);
157 
                break;
158 
            };
159 
            let Some(path) = self.field()? else {
160 
                return fail("git diff-tree: record without a path");
161 
            };
162 
            // `oldmode newmode oldsha newsha status`
163 
            let m: Vec<&str> = meta.split(' ').collect();
164 
            if m.len() != 5 {
165 
                return fail(format!("git diff-tree: bad record {f:?}"));
166 
            }
167 
            let entry = if m[4].starts_with('D') {
168 
                None
169 
            } else {
170 
                Some((m[1].to_string(), m[3].to_string()))
171 
            };
172 
            out.push((path, entry));
173 
        }
174 
        Ok(out)
175 
    }
176 
}
177
178 
impl Drop for DiffTree {
179 
    fn drop(&mut self) {
180 
        let _ = self.child.kill();
181 
        let _ = self.child.wait();
182 
        if let Some(f) = self.feeder.take() {
183 
            let _ = f.join();
184 
        }
185 
    }
186 
}
187
c3069537 Andy McCall ↶ 188 
/// One commit as `git log` describes it.
189 
struct GitCommit {
190 
    sha: String,
191 
    parents: Vec<String>,
192 
    author_secs: u64,
193 
    author: String,
194 
    message: String,
195 
}
196
197 
fn list_commits(git_dir: &Path, branch: &str) -> Result<Vec<GitCommit>> {
198 
    let out = Command::new("git")
199 
        .arg("-C")
200 
        .arg(git_dir)
75167672 Andy McCall 201 
        .args([
202 
            "log",
203 
            "--reverse",
204 
            "--topo-order",
205 
            "--format=%H%x00%P%x00%at%x00%an <%ae>%x00%B%x1e",
206 
            branch,
207 
        ])
c3069537 Andy McCall ↶ 208 
        .output()
209 
        .map_err(|e| RepoError::Msg(format!("running git log: {e}")))?;
210 
    if !out.status.success() {
75167672 Andy McCall 211 
        return fail(format!(
212 
            "git log {branch} failed: {}",
213 
            String::from_utf8_lossy(&out.stderr).trim()
214 
        ));
c3069537 Andy McCall ↶ 215 
    }
216 
    let text = String::from_utf8_lossy(&out.stdout);
217 
    let mut commits = Vec::new();
218 
    for record in text.split('\x1e') {
219 
        let record = record.trim_start_matches('\n');
220 
        if record.trim().is_empty() {
221 
            continue;
222 
        }
223 
        let f: Vec<&str> = record.splitn(5, '\0').collect();
224 
        if f.len() != 5 {
225 
            return fail("unexpected git log record");
226 
        }
227 
        commits.push(GitCommit {
228 
            sha: f[0].to_string(),
229 
            parents: f[1].split_whitespace().map(str::to_string).collect(),
230 
            author_secs: f[2].parse().unwrap_or(0),
231 
            author: f[3].to_string(),
232 
            message: f[4].trim_end().to_string(),
233 
        });
234 
    }
235 
    Ok(commits)
236 
}
237
238 
/// `path → (mode, blob sha)` for a commit's tree.
239 
fn list_tree(git_dir: &Path, sha: &str) -> Result<BTreeMap<String, (String, String)>> {
240 
    let text = git(git_dir, &["ls-tree", "-r", "-z", sha])?;
241 
    let mut out = BTreeMap::new();
242 
    for entry in text.split('\0') {
243 
        if entry.is_empty() {
244 
            continue;
245 
        }
75167672 Andy McCall 246 
        let Some((meta, path)) = entry.split_once('\t') else {
247 
            continue;
248 
        };
c3069537 Andy McCall ↶ 249 
        let m: Vec<&str> = meta.split(' ').collect();
250 
        if m.len() == 3 {
251 
            out.insert(path.to_string(), (m[0].to_string(), m[2].to_string()));
252 
        }
253 
    }
254 
    Ok(out)
255 
}
256
257 
/// The ChangeId for a Git commit: its `Change-Id` trailer if it has a
258 
/// 32-hex one, else BLAKE3-128 of the commit hash (spec §23).
259 
fn change_id_for(sha: &str, message: &str) -> ChangeId {
260 
    for line in message.lines().rev() {
261 
        if let Some(v) = line.strip_prefix("Change-Id:") {
262 
            if let Some(id) = ChangeId::from_hex(v.trim()) {
263 
                return id;
264 
            }
265 
        }
266 
    }
267 
    let h = blake3::hash(sha.as_bytes());
268 
    ChangeId(h.as_bytes()[..16].try_into().unwrap())
269 
}
270
30ea9c1e Andy McCall ↶ 271 
/// Git commits with a known Oxygen counterpart.
272 
struct GitKnown {
273 
    /// Imported commit → its change.
274 
    imported: HashMap<String, ChangeHash>,
275 
    /// Exported commit (or a fetched tip) → the bookmark state it equals.
276 
    exported: HashMap<String, ObjectRef>,
277 
}
278
c3069537 Andy McCall ↶ 279 
#[derive(Debug, Clone, Default, PartialEq, Eq)]
79ab8528 Andy McCall ↶ 280 
pub struct GitExportReport {
281 
    pub exported: usize,
282 
    pub skipped: usize,
4dee277c Andy McCall ↶ 283 
    /// Subrepo nodes with Oxygen pins, which Git cannot represent.
284 
    pub skipped_subrepos: usize,
79ab8528 Andy McCall ↶ 285 
    pub branch: String,
286 
    pub heads: usize,
287 
}
288
fd280228 Andy McCall ↶ 289 
/// A merge bringing more commits than this sheds its base's graphs first.
290 
const MERGE_SHED_COMMITS: usize = 100;
291 
/// A kept state with more built graphs than this sheds them.
292 
const KEPT_GRAPHS: usize = 2000;
293
c3d2ef61 Andy McCall ↶ 294 
/// Text files of a merge rendered this many at a time (oxygen#25): a
295 
/// batch's graphs are built beside the state and dropped once rendered,
296 
/// so what is held for the whole tree is the renderings, not the graphs
297 
/// with every deleted line of their history.
298 
const MERGE_BATCH: usize = 500;
299
300 
thread_local! {
301 
    /// A smaller batch for this thread's imports: tests set 1, so a small
302 
    /// merge takes the batched path.
303 
    static MERGE_BATCH_OVERRIDE: std::cell::Cell<Option<usize>> = const { std::cell::Cell::new(None) };
304 
}
305
306 
#[cfg(test)]
307 
pub(crate) fn set_merge_batch_for_tests(n: Option<usize>) {
308 
    MERGE_BATCH_OVERRIDE.with(|c| c.set(n));
309 
}
310
311 
fn merge_batch() -> usize {
312 
    MERGE_BATCH_OVERRIDE
313 
        .with(|c| c.get())
314 
        .filter(|n| *n > 0)
315 
        .unwrap_or(MERGE_BATCH)
316 
}
317
318 
impl Repo {
319 
    /// As `State::materialize_except`, for a tree whose graphs would not
320 
    /// fit in memory at once. The few files the change builder reads
321 
    /// graphs for (conflicted ones, and renderings whose lines are not
322 
    /// one per live atom) are built in the state afterwards.
323 
    fn materialize_batched(
324 
        &self,
325 
        base: &mut State,
326 
        need: &[crate::ids::NodeId],
327 
        skip: &HashSet<crate::ids::NodeId>,
328 
    ) -> Result<crate::state::Snapshot> {
329 
        use crate::graph::Materialized;
330 
        use crate::state::{FileContent, FileState};
331 
        use rayon::prelude::*;
332 
        let view = base.tree().view();
333 
        let mut rendered: HashMap<crate::ids::NodeId, (Vec<u8>, Materialized)> =
334 
            HashMap::with_capacity(need.len());
335 
        let mut keep_built: Vec<crate::ids::NodeId> = Vec::new();
336 
        for batch in need.chunks(merge_batch()) {
337 
            let t = std::time::Instant::now();
338 
            let graphs = self.build_graphs_detached(base, batch)?;
339 
            crate::timing::add("render: build graphs, batch", t.elapsed());
340 
            let t = std::time::Instant::now();
341 
            let batch_rendered: Vec<(crate::ids::NodeId, Vec<u8>, Materialized)> = graphs
342 
                .par_iter()
343 
                .map(|(node, g)| {
344 
                    let (bytes, m) = base.render_graph(g);
345 
                    (*node, bytes, m)
346 
                })
347 
                .collect();
348 
            drop(graphs);
349 
            crate::timing::add("render: render, batch", t.elapsed());
350 
            for (node, bytes, m) in batch_rendered {
351 
                if m.is_conflicted()
352 
                    || m.live_atoms().len() != crate::diff::split_lines(&bytes).len()
353 
                {
354 
                    keep_built.push(node);
355 
                }
356 
                rendered.insert(node, (bytes, m));
357 
            }
358 
        }
359 
        if !keep_built.is_empty() {
360 
            self.load_graphs_of(base, &keep_built)?;
361 
        }
362 
        let mut files = BTreeMap::new();
363 
        let mut unrendered = BTreeSet::new();
364 
        for (path, e) in view.entries.iter() {
365 
            let f = if e.kind == NodeKind::File && !e.mode.opaque {
366 
                match rendered.remove(&e.node) {
367 
                    Some((bytes, materialized)) => FileState {
368 
                        node: e.node,
369 
                        kind: e.kind,
370 
                        mode: e.mode,
371 
                        content: FileContent::Text {
372 
                            bytes,
373 
                            materialized,
374 
                        },
375 
                    },
376 
                    None if skip.contains(&e.node) => {
377 
                        unrendered.insert(path.clone());
378 
                        FileState {
379 
                            node: e.node,
380 
                            kind: e.kind,
381 
                            mode: e.mode,
382 
                            content: FileContent::Text {
383 
                                bytes: Vec::new(),
384 
                                materialized: Materialized::default(),
385 
                            },
386 
                        }
387 
                    }
388 
                    None => {
389 
                        // Neither asked for nor known unchanged: render it
390 
                        // from a built graph, as the whole-tree path does.
391 
                        self.load_graphs_of(base, &[e.node])?;
392 
                        base.entry_state(e)
393 
                    }
394 
                }
395 
            } else {
396 
                base.entry_state(e)
397 
            };
398 
            files.insert(path.clone(), f);
399 
        }
400 
        Ok(crate::state::Snapshot {
401 
            tree: view,
402 
            files,
403 
            unrendered,
404 
        })
405 
    }
406 
}
407
79ab8528 Andy McCall ↶ 408 
fn git(dir: &Path, args: &[&str]) -> Result<String> {
75167672 Andy McCall 409 
    let out = Command::new("git")
410 
        .arg("-C")
411 
        .arg(dir)
412 
        .args(args)
413 
        .output()
414 
        .map_err(|e| RepoError::Msg(format!("running git: {e}")))?;
79ab8528 Andy McCall ↶ 415 
    if !out.status.success() {
75167672 Andy McCall 416 
        return fail(format!(
417 
            "git {} failed: {}",
418 
            args.join(" "),
419 
            String::from_utf8_lossy(&out.stderr).trim()
420 
        ));
79ab8528 Andy McCall ↶ 421 
    }
422 
    Ok(String::from_utf8_lossy(&out.stdout).trim().to_string())
423 
}
424
425 
impl Repo {
426 
    fn git_marks_dir(&self, git_dir: &Path) -> Result<PathBuf> {
75167672 Andy McCall 427 
        let canonical = git_dir
428 
            .canonicalize()
429 
            .unwrap_or_else(|_| git_dir.to_path_buf());
79ab8528 Andy McCall ↶ 430 
        let key = hash_object(ObjectType::Chunk, canonical.to_string_lossy().as_bytes()).short(16);
431 
        let dir = self.oxy.join("git").join(key);
432 
        fs::create_dir_all(&dir)?;
433 
        Ok(dir)
434 
    }
435
30ea9c1e Andy McCall ↶ 436 
    /// `state ref → Git commit sha` for bookmark states that equal a Git
437 
    /// commit's tree (recorded by import and fetch).
438 
    fn read_git_states(path: &Path) -> Result<BTreeMap<ObjectRef, String>> {
439 
        let mut out = BTreeMap::new();
440 
        if let Ok(text) = fs::read_to_string(path) {
441 
            for line in text.lines() {
442 
                if let Some((h, sha)) = line.split_once(' ') {
443 
                    if let Some(h) = ObjectRef::from_hex(h) {
444 
                        out.insert(h, sha.trim().to_string());
445 
                    }
446 
                }
447 
            }
448 
        }
449 
        Ok(out)
450 
    }
451
452 
    fn record_git_state(&self, git_dir: &Path, state: &ObjectRef, sha: &str) -> Result<()> {
453 
        let path = self.git_marks_dir(git_dir)?.join("git-states");
454 
        let mut states = Self::read_git_states(&path)?;
455 
        states.insert(*state, sha.to_string());
75167672 Andy McCall 456 
        let text: String = states
457 
            .iter()
458 
            .map(|(h, s)| format!("{} {s}\n", h.to_hex()))
459 
            .collect();
30ea9c1e Andy McCall ↶ 460 
        write_atomic(&path, text.as_bytes())
461 
    }
462
463 
    /// The bare mirror this repository keeps for the Git remote `url`.
464 
    fn git_mirror(&self, url: &str) -> Result<PathBuf> {
465 
        let key = hash_object(ObjectType::Chunk, url.as_bytes()).short(16);
466 
        let dir = self.oxy.join("git").join(key).join("mirror.git");
467 
        if !dir.join("HEAD").exists() {
468 
            fs::create_dir_all(&dir)?;
469 
            git(&dir, &["init", "--bare", "-q"])?;
470 
        }
471 
        Ok(dir)
472 
    }
473
474 
    /// Git commits already imported into the bookmark, by sha.
7891e21c Andy McCall ↶ 475 
    pub(crate) fn imported_commits(&self) -> Result<HashMap<String, ChangeHash>> {
30ea9c1e Andy McCall ↶ 476 
        let mut map = HashMap::new();
477 
        for e in self.log()? {
478 
            if e.is_working_copy {
479 
                continue;
480 
            }
75167672 Andy McCall 481 
            if let Some(sha) = e
482 
                .change
483 
                .message
484 
                .lines()
485 
                .rev()
486 
                .find_map(|l| l.strip_prefix("Git-Commit: "))
487 
            {
30ea9c1e Andy McCall ↶ 488 
                map.insert(sha.trim().to_string(), e.hash);
489 
            }
490 
        }
491 
        Ok(map)
492 
    }
493
494 
    /// Bring the mirror's `branch` up to date with the Git remote `url`.
495 
    /// A remote without that branch yet leaves the mirror as it is.
496 
    fn git_mirror_fetch(&self, mirror: &Path, url: &str, branch: &str) -> Result<()> {
497 
        let spec = format!("+refs/heads/{branch}:refs/heads/{branch}");
498 
        match git(mirror, &["fetch", "-q", url, &spec]) {
499 
            Ok(_) => Ok(()),
500 
            Err(e) => {
501 
                let ls = git(mirror, &["ls-remote", "--heads", url, branch]).unwrap_or_default();
75167672 Andy McCall 502 
                if ls.trim().is_empty() {
503 
                    Ok(())
504 
                } else {
505 
                    Err(e)
506 
                }
30ea9c1e Andy McCall ↶ 507 
            }
508 
        }
509 
    }
510
511 
    /// Fetch `branch` of the Git repository at `url` (a path or a Git URL)
512 
    /// through a mirror and import the commits not yet in the bookmark
513 
    /// (spec §23). Works on a repository that already has history: the
514 
    /// new changes join the bookmark like a pull.
515 
    pub fn git_fetch(&self, url: &str, branch: &str) -> Result<GitImportReport> {
516 
        let mirror = self.git_mirror(url)?;
517 
        self.git_mirror_fetch(&mirror, url, branch)?;
75167672 Andy McCall 518 
        if git(
519 
            &mirror,
520 
            &[
521 
                "rev-parse",
522 
                "--verify",
523 
                "-q",
524 
                &format!("refs/heads/{branch}"),
525 
            ],
526 
        )
527 
        .is_err()
528 
        {
30ea9c1e Andy McCall ↶ 529 
            return fail(format!("{url} has no branch {branch:?}"));
530 
        }
531 
        self.import_git_commits(&mirror, branch, &format!("oxy git fetch {url} {branch}"))
532 
    }
533
534 
    /// Export the bookmark to `branch` of the Git repository at `url`
535 
    /// through the mirror and push it. The remote's tip must already have
536 
    /// been fetched (`git_fetch`), so the push fast-forwards.
537 
    pub fn git_push(&self, url: &str, branch: &str) -> Result<GitExportReport> {
538 
        let mirror = self.git_mirror(url)?;
539 
        self.git_mirror_fetch(&mirror, url, branch)?;
75167672 Andy McCall 540 
        if let Ok(tip) = git(
541 
            &mirror,
542 
            &[
543 
                "rev-parse",
544 
                "--verify",
545 
                "-q",
546 
                &format!("refs/heads/{branch}"),
547 
            ],
548 
        ) {
30ea9c1e Andy McCall ↶ 549 
            let known = self.git_known(&mirror)?;
550 
            if !known.imported.contains_key(&tip) && !known.exported.contains_key(&tip) {
75167672 Andy McCall 551 
                return fail(format!(
552 
                    "{url} {branch} has commits not yet fetched; run `oxy git fetch {url}` first"
553 
                ));
30ea9c1e Andy McCall ↶ 554 
            }
555 
        }
556 
        let report = self.git_export_as(&mirror, branch)?;
557 
        let spec = format!("refs/heads/{branch}:refs/heads/{branch}");
558 
        git(&mirror, &["push", "-q", url, &spec])?;
559 
        Ok(report)
560 
    }
561
79ab8528 Andy McCall ↶ 562 
    fn read_oxy_marks(path: &Path) -> Result<BTreeMap<ObjectRef, u64>> {
563 
        let mut out = BTreeMap::new();
564 
        if let Ok(text) = fs::read_to_string(path) {
565 
            for line in text.lines() {
566 
                if let Some((h, m)) = line.split_once(' ') {
567 
                    if let (Some(h), Ok(m)) = (ObjectRef::from_hex(h), m.parse()) {
568 
                        out.insert(h, m);
569 
                    }
570 
                }
571 
            }
572 
        }
573 
        Ok(out)
574 
    }
575
576 
    /// Export the current bookmark to the Git repository at `git_dir`,
577 
    /// creating a bare one if nothing is there. Each state the bookmark
578 
    /// passed through (per the operation log) becomes one commit whose tree
579 
    /// is that state, so the Git branch is a linear history of full trees.
580 
    pub fn git_export(&self, git_dir: &Path) -> Result<GitExportReport> {
30ea9c1e Andy McCall ↶ 581 
        let bookmark = self.config.bookmark.clone();
582 
        self.git_export_as(git_dir, &bookmark)
583 
    }
584
585 
    /// `git_export` to the Git branch `branch`. A state that came from a
586 
    /// Git commit (`oxy git fetch`) is not re-exported: later states chain
587 
    /// onto that commit, so a push back to the Git remote fast-forwards.
588 
    pub fn git_export_as(&self, git_dir: &Path, branch: &str) -> Result<GitExportReport> {
79ab8528 Andy McCall ↶ 589 
        if !git_dir.join("HEAD").exists() && !git_dir.join(".git").exists() {
590 
            fs::create_dir_all(git_dir)?;
591 
            git(git_dir, &["init", "--bare", "-q"])?;
592 
        }
593 
        let marks_dir = self.git_marks_dir(git_dir)?;
594 
        let git_marks = marks_dir.join("git-marks");
595 
        let oxy_marks_path = marks_dir.join("oxy-marks");
596 
        let mut marks = Self::read_oxy_marks(&oxy_marks_path)?;
30ea9c1e Andy McCall ↶ 597 
        let git_states = Self::read_git_states(&marks_dir.join("git-states"))?;
79ab8528 Andy McCall ↶ 598 
        let mut next_mark = marks.values().copied().max().unwrap_or(0) + 1;
599
600 
        let bookmark = self.config.bookmark.clone();
601 
        let mut stream: Vec<u8> = Vec::new();
75167672 Andy McCall 602 
        let mut report = GitExportReport {
603 
            branch: branch.to_string(),
604 
            heads: self.bookmark_heads()?.len(),
605 
            ..Default::default()
606 
        };
79ab8528 Andy McCall ↶ 607
608 
        // States along the bookmark, oldest first, skipping unchanged and pruned ones.
609 
        let mut ops = self.op_log()?;
610 
        ops.reverse();
611 
        let mut prev: Option<(ObjectRef, Vec<ChangeHash>)> = None;
30ea9c1e Andy McCall ↶ 612 
        // The fast-import parent of the next commit: `:mark` or a commit sha.
613 
        let mut last_parent: Option<String> = None;
79ab8528 Andy McCall ↶ 614 
        for (_, op) in &ops {
75167672 Andy McCall 615 
            let Some(state_ref) = op.view.bookmarks.get(&bookmark).copied() else {
616 
                continue;
617 
            };
79ab8528 Andy McCall ↶ 618 
            if prev.as_ref().is_some_and(|(p, _)| *p == state_ref) {
619 
                continue;
620 
            }
75167672 Andy McCall 621 
            let Ok(state) = self.store.get_state(&state_ref) else {
622 
                continue;
623 
            };
79ab8528 Andy McCall ↶ 624 
            if state.heads.is_empty() {
625 
                prev = Some((state_ref, vec![]));
626 
                continue;
627 
            }
30ea9c1e Andy McCall ↶ 628 
            if let Some(sha) = git_states.get(&state_ref) {
629 
                report.skipped += 1;
630 
                last_parent = Some(sha.clone());
631 
                prev = Some((state_ref, state.heads.clone()));
632 
                continue;
633 
            }
79ab8528 Andy McCall ↶ 634 
            if let Some(m) = marks.get(&state_ref) {
635 
                report.skipped += 1;
30ea9c1e Andy McCall ↶ 636 
                last_parent = Some(format!(":{m}"));
79ab8528 Andy McCall ↶ 637 
                prev = Some((state_ref, state.heads.clone()));
638 
                continue;
639 
            }
640 
            let prev_heads = prev.as_ref().map(|(_, h)| h.clone()).unwrap_or_default();
75167672 Andy McCall 641 
            let added = self
642 
                .negotiate(&prev_heads, &state.heads)
643 
                .unwrap_or_default();
79ab8528 Andy McCall ↶ 644 
            let loaded = self.load_state(&state.heads)?;
645 
            let snapshot = loaded.materialize();
646 
            let (author, secs, mut message) = match added.as_slice() {
647 
                [h] => {
648 
                    let c = self.store.get_change(h)?;
649 
                    let mut m = c.message.clone();
650 
                    if !m.ends_with('\n') {
651 
                        m.push('\n');
652 
                    }
75167672 Andy McCall 653 
                    m.push_str(&format!(
654 
                        "\nChange-Id: {}\nOxy-Hash: {}\n",
655 
                        c.change_id.to_hex(),
656 
                        h.to_hex()
657 
                    ));
79ab8528 Andy McCall ↶ 658 
                    (self.git_author(&c), c.hlc.wall_ms / 1000, m)
659 
                }
75167672 Andy McCall 660 
                [] => (
661 
                    "oxy <oxy@localhost>".to_string(),
662 
                    op.hlc.wall_ms / 1000,
663 
                    format!("oxy: {}\n\n", op.command),
664 
                ),
79ab8528 Andy McCall ↶ 665 
                many => {
666 
                    let mut m = format!("oxy: {} ({} changes)\n\n", op.command, many.len());
667 
                    for h in many {
668 
                        let c = self.store.get_change(h)?;
75167672 Andy McCall 669 
                        m.push_str(&format!(
670 
                            "- {} {}\n",
671 
                            c.change_id.short(8),
672 
                            c.message.lines().next().unwrap_or("")
673 
                        ));
79ab8528 Andy McCall ↶ 674 
                    }
75167672 Andy McCall 675 
                    m.push_str(&format!(
676 
                        "\nOxy-Changes: {}\n",
677 
                        many.iter()
678 
                            .map(|h| h.to_hex())
679 
                            .collect::<Vec<_>>()
680 
                            .join(" ")
681 
                    ));
79ab8528 Andy McCall ↶ 682 
                    ("oxy <oxy@localhost>".to_string(), op.hlc.wall_ms / 1000, m)
683 
                }
684 
            };
685 
            message.push_str(&format!("Oxy-State: {}\n", state_ref.to_hex()));
686 
            if snapshot.is_conflicted() {
687 
                message.push_str("Oxy-Conflicted: true\n");
688 
            }
689 
            let mark = next_mark;
690 
            next_mark += 1;
30ea9c1e Andy McCall ↶ 691 
            let parents: Vec<String> = last_parent.clone().into_iter().collect();
75167672 Andy McCall 692 
            self.write_commit(
693 
                &mut stream,
694 
                "refs/oxy/export",
695 
                mark,
696 
                &author,
697 
                secs,
698 
                &message,
699 
                &parents,
700 
                &snapshot,
701 
            )?;
79ab8528 Andy McCall ↶ 702 
            marks.insert(state_ref, mark);
30ea9c1e Andy McCall ↶ 703 
            last_parent = Some(format!(":{mark}"));
79ab8528 Andy McCall ↶ 704 
            prev = Some((state_ref, state.heads.clone()));
705 
            report.exported += 1;
706 
        }
30ea9c1e Andy McCall ↶ 707 
        if let Some(p) = &last_parent {
708 
            stream.extend_from_slice(format!("reset refs/heads/{branch}\nfrom {p}\n\n").as_bytes());
79ab8528 Andy McCall ↶ 709 
        }
132b884b Andy McCall ↶ 710 
        // Tags: a tag whose state was exported as a bookmark state points at
711 
        // that commit; otherwise its state gets a commit of its own.
712 
        for t in self.tags()? {
30ea9c1e Andy McCall ↶ 713 
            let target = match (git_states.get(&t.state), marks.get(&t.state)) {
714 
                (Some(sha), _) => sha.clone(),
715 
                (None, Some(m)) => format!(":{m}"),
716 
                (None, None) => {
132b884b Andy McCall ↶ 717 
                    let loaded = self.load_state(&t.heads)?;
718 
                    let snapshot = loaded.materialize();
719 
                    let mark = next_mark;
720 
                    next_mark += 1;
75167672 Andy McCall 721 
                    let message = format!(
722 
                        "oxy: tag {}\n\n{}Oxy-State: {}\n",
723 
                        t.name,
724 
                        if t.message.is_empty() {
725 
                            String::new()
726 
                        } else {
727 
                            format!("{}\n\n", t.message)
728 
                        },
729 
                        t.state.to_hex()
730 
                    );
731 
                    self.write_commit(
732 
                        &mut stream,
733 
                        "refs/oxy/export",
734 
                        mark,
735 
                        "oxy <oxy@localhost>",
736 
                        Hlc::now().wall_ms / 1000,
737 
                        &message,
738 
                        &[],
739 
                        &snapshot,
740 
                    )?;
132b884b Andy McCall ↶ 741 
                    marks.insert(t.state, mark);
742 
                    report.exported += 1;
30ea9c1e Andy McCall ↶ 743 
                    format!(":{mark}")
132b884b Andy McCall ↶ 744 
                }
745 
            };
75167672 Andy McCall 746 
            stream.extend_from_slice(
747 
                format!("reset refs/tags/{}\nfrom {target}\n\n", t.name).as_bytes(),
748 
            );
132b884b Andy McCall ↶ 749 
        }
750 
        stream.extend_from_slice(b"done\n");
79ab8528 Andy McCall ↶ 751
752 
        let mut cmd = Command::new("git");
753 
        // `--force`: the scratch ref may jump between unrelated states.
75167672 Andy McCall 754 
        cmd.arg("-C")
755 
            .arg(git_dir)
756 
            .args(["fast-import", "--quiet", "--done", "--force"])
757 
            .arg(format!("--export-marks={}", git_marks.display()));
79ab8528 Andy McCall ↶ 758 
        if git_marks.exists() {
759 
            cmd.arg(format!("--import-marks={}", git_marks.display()));
760 
        }
75167672 Andy McCall 761 
        let mut child = cmd
762 
            .stdin(Stdio::piped())
763 
            .stdout(Stdio::null())
764 
            .stderr(Stdio::piped())
765 
            .spawn()
766 
            .map_err(|e| RepoError::Msg(format!("running git fast-import: {e}")))?;
79ab8528 Andy McCall ↶ 767 
        child.stdin.take().unwrap().write_all(&stream)?;
768 
        let out = child.wait_with_output()?;
769 
        if !out.status.success() {
75167672 Andy McCall 770 
            return fail(format!(
771 
                "git fast-import failed: {}",
772 
                String::from_utf8_lossy(&out.stderr).trim()
773 
            ));
79ab8528 Andy McCall ↶ 774 
        }
75167672 Andy McCall 775 
        let text: String = marks
776 
            .iter()
777 
            .map(|(h, m)| format!("{} {m}\n", h.to_hex()))
778 
            .collect();
79ab8528 Andy McCall ↶ 779 
        write_atomic(&oxy_marks_path, text.as_bytes())?;
780 
        let _ = git(git_dir, &["update-ref", "-d", "refs/oxy/export"]);
4dee277c Andy McCall ↶ 781 
        report.skipped_subrepos = self.git_export_skipped.replace(0);
79ab8528 Andy McCall ↶ 782 
        Ok(report)
783 
    }
784
c3069537 Andy McCall ↶ 785 
    /// Import branch `branch` of the Git repository at `git_dir` into this
786 
    /// (empty) repository's bookmark (spec §23): one change per commit, its
787 
    /// edits the difference between the union of its parents' states and its
788 
    /// tree, so merge commits become resolutions. Submodules are skipped.
789 
    pub fn import_git(&self, git_dir: &Path, branch: &str) -> Result<GitImportReport> {
790 
        if !self.bookmark_heads()?.is_empty() {
791 
            return fail("git import needs an empty bookmark; import into a fresh repository");
792 
        }
75167672 Andy McCall 793 
        self.import_git_commits(
794 
            git_dir,
795 
            branch,
796 
            &format!("oxy git import {} {branch}", git_dir.display()),
797 
        )
30ea9c1e Andy McCall ↶ 798 
    }
799
800 
    /// Import the commits of `branch` that the bookmark does not have yet,
801 
    /// then move the bookmark to the union of its heads and theirs.
75167672 Andy McCall 802 
    fn import_git_commits(
803 
        &self,
804 
        git_dir: &Path,
805 
        branch: &str,
806 
        command: &str,
807 
    ) -> Result<GitImportReport> {
8f10a07a Andy McCall ↶ 808 
        let batch = self.store.begin_batch()?;
809 
        let result = self.import_git_commits_batched(git_dir, branch, command);
810 
        batch.finish()?;
811 
        result
812 
    }
813
75167672 Andy McCall 814 
    fn import_git_commits_batched(
815 
        &self,
816 
        git_dir: &Path,
817 
        branch: &str,
818 
        command: &str,
819 
    ) -> Result<GitImportReport> {
c3069537 Andy McCall ↶ 820 
        let commits = list_commits(git_dir, branch)?;
821 
        let mut cat = CatFile::start(git_dir)?;
822 
        let mut report = GitImportReport::default();
30ea9c1e Andy McCall ↶ 823 
        let mut known = self.git_known(git_dir)?;
824 
        let old_heads = self.bookmark_heads()?;
dfb23a3a Andy McCall ↶ 825 
        // The bookmark's heads as changes arrive: a change stops being a
826 
        // head when another depends on it, so no full state of every
827 
        // imported change is needed at the end.
828 
        let mut heads_now: BTreeSet<ChangeHash> = old_heads.iter().copied().collect();
958fb27d Andy McCall ↶ 829 
        // The state and Git tree of imported commits whose first-parent
830 
        // children are still to come, so a child starts from its parent's
831 
        // state instead of rebuilding it from history. The last child takes
dfb23a3a Andy McCall ↶ 832 
        // it; earlier ones clone it. At most `max_kept` are kept (a later
958fb27d Andy McCall ↶ 833 
        // child of an evicted one rebuilds).
dfb23a3a Andy McCall ↶ 834 
        // OXY_IMPORT_KEEP overrides it, for measuring.
835 
        let max_kept: usize = std::env::var("OXY_IMPORT_KEEP")
836 
            .ok()
837 
            .and_then(|v| v.parse().ok())
838 
            .unwrap_or(8);
958fb27d Andy McCall ↶ 839 
        let mut waiting: HashMap<&str, usize> = HashMap::new();
840 
        for gc in &commits {
841 
            if let Some(p1) = gc.parents.first() {
842 
                *waiting.entry(p1.as_str()).or_default() += 1;
843 
            }
844 
        }
845 
        type Kept = (State, BTreeMap<String, (String, String)>);
846 
        let mut kept: HashMap<String, Kept> = HashMap::new();
2306a4d2 Andy McCall ↶ 847 
        // Every commit to import that has a parent, diffed against its first.
848 
        let pairs: Vec<(String, String)> = commits
849 
            .iter()
850 
            .filter(|gc| {
851 
                !known.imported.contains_key(&gc.sha) && !known.exported.contains_key(&gc.sha)
852 
            })
853 
            .filter_map(|gc| Some((gc.sha.clone(), gc.parents.first()?.clone())))
854 
            .collect();
855 
        let mut diffs = DiffTree::start(git_dir, pairs)?;
856 
        let mut rebuilt = 0usize;
958fb27d Andy McCall ↶ 857 
        let mut kept_order: std::collections::VecDeque<String> = std::collections::VecDeque::new();
858 
        let mut last: Option<String> = None;
c3069537 Andy McCall ↶ 859 
        let mut blob_cache: HashMap<String, Vec<u8>> = HashMap::new();
860 
        let mut cache_bytes: usize = 0;
958fb27d Andy McCall ↶ 861 
        // Time spent per phase over all commits (`OXY_TIMINGS=1`).
862 
        let mut spent = [std::time::Duration::ZERO; 6];
863 
        let mut clock = std::time::Instant::now();
fb5b6181 Andy McCall ↶ 864 
        // One line per commit on stderr with OXY_IMPORT_TRACE=1: where its
865 
        // base came from, what a merge added, the graphs its state holds,
866 
        // the kept states and the process's resident memory.
867 
        let trace = std::env::var_os("OXY_IMPORT_TRACE").is_some();
868 
        let rss_mb = || -> u64 {
869 
            std::fs::read_to_string("/proc/self/statm")
870 
                .ok()
871 
                .and_then(|s| s.split_whitespace().nth(1)?.parse::<u64>().ok())
872 
                .map(|pages| pages * 4096 >> 20)
873 
                .unwrap_or(0)
874 
        };
875 
        let mut done = 0usize;
fd280228 Andy McCall ↶ 876 
        // Graphs turned back into history (see `State::shed_graphs`).
877 
        let mut shed = 0usize;
958fb27d Andy McCall ↶ 878 
        let mut lap = |i: usize, clock: &mut std::time::Instant| {
879 
            let now = std::time::Instant::now();
880 
            spent[i] += now - *clock;
881 
            *clock = now;
882 
        };
c3069537 Andy McCall ↶ 883
884 
        for gc in &commits {
30ea9c1e Andy McCall ↶ 885 
            if known.imported.contains_key(&gc.sha) || known.exported.contains_key(&gc.sha) {
886 
                continue;
887 
            }
fb5b6181 Andy McCall ↶ 888 
            let mut base_how = "root";
958fb27d Andy McCall ↶ 889 
            // Base: union of every parent's state, starting from the first
890 
            // parent's kept state when there is one; its tree tells which
891 
            // paths this commit leaves alone.
892 
            let (mut base, prev_tree) = match gc.parents.first() {
893 
                Some(p1) => {
894 
                    let left = waiting.get_mut(p1.as_str()).map(|n| {
895 
                        *n -= 1;
896 
                        *n
897 
                    });
898 
                    let found = if left == Some(0) {
899 
                        kept_order.retain(|k| k != p1);
900 
                        kept.remove(p1.as_str())
901 
                    } else {
902 
                        kept.get(p1.as_str()).cloned()
fb5b6181 Andy McCall ↶ 903 
                    };
904 
                    let how = match (&found, left) {
905 
                        (Some(_), Some(0)) => "kept",
906 
                        (Some(_), _) => "cloned",
907 
                        (None, _) => "rebuilt",
958fb27d Andy McCall ↶ 908 
                    };
fb5b6181 Andy McCall ↶ 909 
                    if trace {
910 
                        base_how = how;
911 
                    }
958fb27d Andy McCall ↶ 912 
                    match found {
913 
                        Some((st, tree)) => (st, Some(tree)),
914 
                        None => {
b493d993 Andy McCall ↶ 915 
                            // History only: graphs are built for the files a
916 
                            // commit changes, so memory follows what changes,
917 
                            // not the whole tree.
918 
                            let mut st = State::new_history_only();
958fb27d Andy McCall ↶ 919 
                            self.add_git_ancestors(&mut st, git_dir, &[p1.as_str()], &known)?;
2306a4d2 Andy McCall ↶ 920 
                            // The parent's tree, so this commit still
921 
                            // renders only what it changes.
922 
                            rebuilt += 1;
923 
                            (st, Some(list_tree(git_dir, p1)?))
958fb27d Andy McCall ↶ 924 
                        }
925 
                    }
30ea9c1e Andy McCall ↶ 926 
                }
b493d993 Andy McCall ↶ 927 
                None => (State::new_history_only(), None),
c3069537 Andy McCall ↶ 928 
            };
958fb27d Andy McCall ↶ 929 
            // Nodes the other parents' changes touch: the first parent's
930 
            // tree cannot vouch for them.
931 
            let mut touched: HashSet<NodeId> = HashSet::new();
fb5b6181 Andy McCall ↶ 932 
            let mut merged = 0usize;
c3069537 Andy McCall ↶ 933 
            if gc.parents.len() > 1 {
934 
                report.merges += 1;
935 
                let mut args: Vec<String> = gc.parents[1..].to_vec();
936 
                args.push(format!("^{}", gc.parents[0]));
30ea9c1e Andy McCall ↶ 937 
                let refs: Vec<&str> = args.iter().map(String::as_str).collect();
fd280228 Andy McCall ↶ 938 
                // A merge of a long-lived branch brings thousands of
939 
                // changes: added to built graphs, which kept states share,
940 
                // they were copied and applied whole (vscode's import passed
941 
                // 14 GB on such merges). Shed the graphs first, so the
942 
                // changes go in as skeletons and only what this commit
943 
                // touches is built again.
944 
                if base.built_count() > 0 {
945 
                    let mut count_args = vec!["rev-list", "--count"];
946 
                    count_args.extend(refs.iter().copied());
947 
                    let others: usize = git(git_dir, &count_args)?.trim().parse().unwrap_or(0);
948 
                    if others > MERGE_SHED_COMMITS {
949 
                        shed += base.shed_graphs(|b| self.store_blob(b))?;
950 
                    }
951 
                }
fb5b6181 Andy McCall ↶ 952 
                let before = base.len();
958fb27d Andy McCall ↶ 953 
                self.add_git_ancestors_touching(&mut base, git_dir, &refs, &known, &mut touched)?;
fb5b6181 Andy McCall ↶ 954 
                merged = base.len() - before;
03571027 Andy McCall ↶ 955 
            }
958fb27d Andy McCall ↶ 956 
            lap(0, &mut clock);
2306a4d2 Andy McCall ↶ 957 
            // The commit's tree: its first parent's with the diff applied,
958 
            // or listed whole when that tree is not kept.
959 
            let diff = match gc.parents.first() {
960 
                Some(_) => Some(diffs.diff(&gc.sha)?),
961 
                None => None,
962 
            };
963 
            let (tree, changed) = match (prev_tree, diff) {
964 
                (Some(mut t), Some(diff)) => {
965 
                    let mut changed: HashSet<String> = HashSet::new();
966 
                    for (path, entry) in diff {
967 
                        match entry {
968 
                            Some(v) => {
969 
                                t.insert(path.clone(), v);
970 
                            }
971 
                            None => {
972 
                                t.remove(&path);
973 
                            }
974 
                        }
975 
                        changed.insert(path);
976 
                    }
977 
                    (t, Some(changed))
978 
                }
979 
                _ => (list_tree(git_dir, &gc.sha)?, None),
980 
            };
958fb27d Andy McCall ↶ 981 
            lap(1, &mut clock);
03571027 Andy McCall ↶ 982 
            // Text files with the same blob and mode as in the first
983 
            // parent's tree equal the base's rendering: they are neither
984 
            // rendered nor read nor compared (spec §12.3, as for a clean
985 
            // working copy), so a commit costs what it changes.
986 
            let mut same: HashMap<&str, NodeId> = HashMap::new();
2306a4d2 Andy McCall ↶ 987 
            // Opaque files and symlinks likewise, when the base has one
988 
            // blob for them: they are neither read nor stored again.
989 
            let mut same_opaque: HashSet<&str> = HashSet::new();
990 
            // An ordinary commit in a tree without submodules is built from
991 
            // the paths it changes alone (oxygen#25): with the `.oxyattributes`
992 
            // files that decide how they are stored, and the directories its
993 
            // removals may leave empty. A merge, or a commit whose parent's
994 
            // tree is not at hand, is built from its whole tree.
995 
            let partial = changed.is_some()
996 
                && gc.parents.len() == 1
997 
                && !tree.values().any(|(mode, _)| mode == "160000");
998 
            let mut rel: BTreeSet<&str> = BTreeSet::new();
999 
            let base_snapshot = if let (true, Some(changed)) = (partial, &changed) {
03571027 Andy McCall ↶ 1000 
                let view = base.tree().view();
2306a4d2 Andy McCall ↶ 1001 
                rel.extend(changed.iter().map(String::as_str));
1002 
                rel.extend(
1003 
                    tree.keys()
1004 
                        .map(String::as_str)
1005 
                        .filter(|p| *p == ".oxyattributes" || p.ends_with("/.oxyattributes")),
1006 
                );
1007 
                let mut dirs: BTreeSet<&str> = BTreeSet::new();
1008 
                for p in changed.iter().filter(|p| !tree.contains_key(*p)) {
1009 
                    let mut end = p.len();
1010 
                    while let Some(i) = p[..end].rfind('/') {
1011 
                        if !dirs.insert(&p[..i]) {
1012 
                            break;
1013 
                        }
1014 
                        end = i;
03571027 Andy McCall ↶ 1015 
                    }
2306a4d2 Andy McCall ↶ 1016 
                }
1017 
                let need: Vec<NodeId> = rel
1018 
                    .iter()
1019 
                    .filter_map(|p| view.entries.get(*p))
1020 
                    .filter(|e| e.kind == NodeKind::File && !e.mode.opaque)
1021 
                    .map(|e| e.node)
1022 
                    .collect();
1023 
                crate::timing::add("render: same and need", clock.elapsed());
1024 
                let t = std::time::Instant::now();
1025 
                self.load_graphs_of(&mut base, &need)?;
1026 
                crate::timing::add("render: load graphs", t.elapsed());
1027 
                let t = std::time::Instant::now();
1028 
                let snap = base.materialize_paths(view, rel.iter().chain(dirs.iter()).copied());
1029 
                crate::timing::add("render: materialize", t.elapsed());
1030 
                snap
1031 
            } else {
1032 
                if let Some(changed) = &changed {
1033 
                    let view = base.tree().view();
1034 
                    for path in tree.keys() {
1035 
                        // Submodule handling reads these two whatever changed.
1036 
                        if changed.contains(path) || path == ".gitmodules" || path == ".oxymodules"
958fb27d Andy McCall ↶ 1037 
                        {
2306a4d2 Andy McCall ↶ 1038 
                            continue;
1039 
                        }
1040 
                        if let Some(e) = view.entries.get(path) {
1041 
                            if touched.contains(&e.node) {
1042 
                                continue;
1043 
                            }
1044 
                            let opaque = e.kind == NodeKind::Symlink
1045 
                                || (e.kind == NodeKind::File && e.mode.opaque);
1046 
                            if e.kind == NodeKind::File && !e.mode.opaque {
1047 
                                same.insert(path.as_str(), e.node);
1048 
                            } else if opaque && base.live_blobs(&e.node).len() == 1 {
1049 
                                same_opaque.insert(path.as_str());
1050 
                            }
03571027 Andy McCall ↶ 1051 
                        }
1052 
                    }
1053 
                }
2306a4d2 Andy McCall ↶ 1054 
                // Build the graphs of every text file the commit may change.
1055 
                let skip: HashSet<NodeId> = same.values().copied().collect();
1056 
                let need: Vec<NodeId> = base
1057 
                    .tree()
1058 
                    .view()
1059 
                    .entries
1060 
                    .values()
1061 
                    .filter(|e| {
1062 
                        e.kind == NodeKind::File && !e.mode.opaque && !skip.contains(&e.node)
1063 
                    })
1064 
                    .map(|e| e.node)
1065 
                    .collect();
1066 
                crate::timing::add("render: same and need", clock.elapsed());
1067 
                let t = std::time::Instant::now();
c3d2ef61 Andy McCall ↶ 1068 
                let snap = if need.len() > merge_batch() {
1069 
                    // A merge onto a base with no built graphs: thousands
1070 
                    // of files, whose graphs would not fit at once.
1071 
                    self.materialize_batched(&mut base, &need, &skip)?
1072 
                } else {
1073 
                    self.load_graphs_of(&mut base, &need)?;
1074 
                    base.materialize_except(&skip)
1075 
                };
1076 
                crate::timing::add("render: whole tree", t.elapsed());
2306a4d2 Andy McCall ↶ 1077 
                snap
1078 
            };
c3069537 Andy McCall ↶ 1079
958fb27d Andy McCall ↶ 1080 
            lap(2, &mut clock);
c3069537 Andy McCall ↶ 1081 
            // The commit's tree as in-memory entries.
1082 
            let mut entries: BTreeMap<String, FsEntry> = BTreeMap::new();
4dee277c Andy McCall ↶ 1083 
            let mut gitlinks: Vec<(String, String)> = Vec::new();
2306a4d2 Andy McCall ↶ 1084 
            let listed: Box<dyn Iterator<Item = (&String, &(String, String))>> = if partial {
1085 
                Box::new(rel.iter().filter_map(|p| tree.get_key_value(*p)))
1086 
            } else {
1087 
                Box::new(tree.iter())
1088 
            };
1089 
            for (path, (mode, sha)) in listed {
03571027 Andy McCall ↶ 1090 
                let (path, sha) = (path.clone(), sha.clone());
c3069537 Andy McCall ↶ 1091 
                let (kind, exec) = match mode.as_str() {
1092 
                    "100644" | "100664" => (NodeKind::File, false),
1093 
                    "100755" => (NodeKind::File, true),
1094 
                    "120000" => (NodeKind::Symlink, false),
1095 
                    "160000" => {
4dee277c Andy McCall ↶ 1096 
                        gitlinks.push((path, sha));
c3069537 Andy McCall ↶ 1097 
                        continue;
1098 
                    }
1099 
                    _ => continue,
1100 
                };
2306a4d2 Andy McCall ↶ 1101 
                if same.contains_key(path.as_str()) || same_opaque.contains(path.as_str()) {
03571027 Andy McCall ↶ 1102 
                    let mut e = FsEntry::new(kind, exec, Vec::new());
1103 
                    e.unchanged = true;
1104 
                    entries.insert(path, e);
1105 
                    continue;
1106 
                }
c3069537 Andy McCall ↶ 1107 
                let content = match blob_cache.get(&sha) {
1108 
                    Some(c) => c.clone(),
1109 
                    None => {
1110 
                        let c = cat.read(&sha)?;
dfb23a3a Andy McCall ↶ 1111 
                        if cache_bytes + c.len() > 64 * 1024 * 1024 {
c3069537 Andy McCall ↶ 1112 
                            blob_cache.clear();
1113 
                            cache_bytes = 0;
1114 
                        }
1115 
                        cache_bytes += c.len();
1116 
                        blob_cache.insert(sha.clone(), c.clone());
1117 
                        c
1118 
                    }
1119 
                };
b7f102fc Andy McCall ↶ 1120 
                entries.insert(path, FsEntry::new(kind, exec, content));
c3069537 Andy McCall ↶ 1121 
            }
1122
4dee277c Andy McCall ↶ 1123 
            // Submodules (spec §23): gitlinks become subrepo nodes with Git
1124 
            // pins, and `.gitmodules` also yields `.oxymodules`.
1125 
            if !gitlinks.is_empty() {
1126 
                report.submodules += gitlinks.len();
75167672 Andy McCall 1127 
                let gm = entries
1128 
                    .get(".gitmodules")
1129 
                    .map(|e| {
1130 
                        crate::subrepo::Modules::from_gitmodules(&String::from_utf8_lossy(
1131 
                            &e.content,
1132 
                        ))
1133 
                    })
1134 
                    .unwrap_or_default();
1135 
                let mut modules = entries
1136 
                    .get(".oxymodules")
1137 
                    .map(|e| crate::subrepo::Modules::parse(&String::from_utf8_lossy(&e.content)))
1138 
                    .unwrap_or_default();
4dee277c Andy McCall ↶ 1139 
                for (path, sha) in &gitlinks {
1140 
                    let m = gm.entries.get(path);
75167672 Andy McCall 1141 
                    let branch = m
1142 
                        .and_then(|m| m.bookmark.clone())
1143 
                        .unwrap_or_else(|| "HEAD".into());
1144 
                    entries.insert(
1145 
                        path.clone(),
1146 
                        FsEntry::new(
1147 
                            NodeKind::Subrepo,
1148 
                            false,
1149 
                            crate::subrepo::Pin::Git {
1150 
                                sha: sha.clone(),
1151 
                                branch,
1152 
                            }
1153 
                            .encode(),
1154 
                        ),
1155 
                    );
4dee277c Andy McCall ↶ 1156 
                    if let Some(m) = m {
1157 
                        modules.entries.insert(path.clone(), m.clone());
1158 
                    }
1159 
                }
1160 
                if !modules.entries.is_empty() {
75167672 Andy McCall 1161 
                    entries.insert(
1162 
                        ".oxymodules".into(),
1163 
                        FsEntry::new(NodeKind::File, false, modules.to_text().into_bytes()),
1164 
                    );
4dee277c Andy McCall ↶ 1165 
                }
1166 
            }
958fb27d Andy McCall ↶ 1167 
            lap(3, &mut clock);
dfb23a3a Andy McCall ↶ 1168 
            // New nodes' identifiers derive from the commit, so importing the
1169 
            // same Git history twice gives the same changes (spec §23).
1170 
            let seed = *blake3::hash(format!("oxy git import\0{}", gc.sha).as_bytes()).as_bytes();
2306a4d2 Andy McCall ↶ 1171 
            let mut built = if partial {
1172 
                // Whether the new tree has anything under `dir`.
1173 
                let dir_needed = |dir: &str| {
1174 
                    let prefix = format!("{dir}/");
1175 
                    tree.range(prefix.clone()..)
1176 
                        .next()
1177 
                        .is_some_and(|(k, _)| k.starts_with(&prefix))
1178 
                };
1179 
                self.build_from_entries_seeded(
1180 
                    &base,
1181 
                    &base_snapshot,
1182 
                    &entries,
1183 
                    seed,
1184 
                    Some(&dir_needed),
1185 
                )?
1186 
            } else {
1187 
                self.build_from_entries_seeded(&base, &base_snapshot, &entries, seed, None)?
1188 
            };
958fb27d Andy McCall ↶ 1189 
            lap(4, &mut clock);
62794ecc Andy McCall ↶ 1190 
            // A file or submodule with several live blobs (branches that
1191 
            // replaced it independently) is resolved to the commit's blob:
1192 
            // Git's merge is the authority here, unlike a working copy,
1193 
            // where a candidate on disk is not yet a decision (spec §23).
1194 
            for (path, f) in &base_snapshot.files {
1195 
                let (FileContent::Opaque { blobs }, Some(e)) = (&f.content, entries.get(path))
1196 
                else {
1197 
                    continue;
1198 
                };
1199 
                if blobs.len() < 2 || f.kind != e.kind {
1200 
                    continue;
1201 
                }
1202 
                let chosen = self.store_blob(&e.content)?;
1203 
                for b in blobs {
1204 
                    built.edits.push(Edit::Replace {
1205 
                        node: f.node,
1206 
                        old: Some(*b),
1207 
                        new: Some(chosen),
1208 
                    });
1209 
                }
1210 
                built.resolves = true;
1211 
                built
1212 
                    .contributors
1213 
                    .extend(base.live_replaces(&f.node).into_iter().map(|(c, _)| c));
1214 
            }
e83354cd Andy McCall ↶ 1215 
            // It shares the base's tree view, which applying the change
1216 
            // then updates in place rather than copying.
1217 
            drop(base_snapshot);
1218 
            drop(entries);
c3069537 Andy McCall ↶ 1219 
            let mut deps: BTreeSet<ChangeHash> = built.contributors.iter().copied().collect();
1220 
            let mut message = gc.message.clone();
1221 
            if !message.ends_with('\n') {
1222 
                message.push('\n');
1223 
            }
75167672 Andy McCall 1224 
            message.push_str(&format!(
1225 
                "\nGit-Commit: {}\nGit-Author: {}\n",
1226 
                gc.sha, gc.author
1227 
            ));
c3069537 Andy McCall ↶ 1228 
            let mut change = Change {
6f8aeb99 Andy McCall ↶ 1229 
                version: crate::objects::CHANGE_VERSION,
c3069537 Andy McCall ↶ 1230 
                change_id: change_id_for(&gc.sha, &gc.message),
1231 
                deps: vec![],
1232 
                author: ObjectRef::ZERO,
75167672 Andy McCall 1233 
                hlc: Hlc {
1234 
                    wall_ms: gc.author_secs * 1000,
1235 
                    logical: 0,
1236 
                },
c3069537 Andy McCall ↶ 1237 
                message,
1238 
                tree_ops: built.tree_ops,
1239 
                edits: built.edits,
75167672 Andy McCall 1240 
                kind: if built.resolves {
1241 
                    ChangeKind::Resolution
1242 
                } else {
1243 
                    ChangeKind::Normal
1244 
                },
c3069537 Andy McCall ↶ 1245 
            };
2306a4d2 Andy McCall ↶ 1246 
            let t = std::time::Instant::now();
c3069537 Andy McCall ↶ 1247 
            deps.extend(referenced_changes(&change, &base));
2306a4d2 Andy McCall ↶ 1248 
            crate::timing::add("store: referenced", t.elapsed());
c3069537 Andy McCall ↶ 1249 
            change.deps = deps.into_iter().collect();
1250 
            change.normalize();
1251 
            let hash = self.store.put_change(&change)?;
dfb23a3a Andy McCall ↶ 1252 
            for d in &change.deps {
1253 
                heads_now.remove(d);
1254 
            }
1255 
            heads_now.insert(hash);
ab4d4e31 Andy McCall ↶ 1256 
            // Built graphs take its edits whole; files not built keep only
1257 
            // where its text is, read back from the store.
2306a4d2 Andy McCall ↶ 1258 
            let t = std::time::Instant::now();
ab4d4e31 Andy McCall ↶ 1259 
            let (skeleton, texts) = self.store.get_change_lazy(&hash)?;
1260 
            base.add_known_mixed(hash, change, skeleton, texts);
2306a4d2 Andy McCall ↶ 1261 
            crate::timing::add("store: add to state", t.elapsed());
30ea9c1e Andy McCall ↶ 1262 
            known.imported.insert(gc.sha.clone(), hash);
958fb27d Andy McCall ↶ 1263 
            if waiting.get(gc.sha.as_str()).is_some_and(|n| *n > 0) {
fd280228 Andy McCall ↶ 1264 
                // A kept state holds its graphs until its children come:
1265 
                // past a few thousand, they go back into its history.
1266 
                if base.built_count() > KEPT_GRAPHS {
1267 
                    shed += base.shed_graphs(|b| self.store_blob(b))?;
1268 
                }
958fb27d Andy McCall ↶ 1269 
                kept.insert(gc.sha.clone(), (base, tree));
1270 
                kept_order.push_back(gc.sha.clone());
dfb23a3a Andy McCall ↶ 1271 
                if kept_order.len() > max_kept {
958fb27d Andy McCall ↶ 1272 
                    if let Some(old) = kept_order.pop_front() {
1273 
                        kept.remove(&old);
1274 
                    }
1275 
                }
1276 
            }
1277 
            last = Some(gc.sha.clone());
c3069537 Andy McCall ↶ 1278 
            report.commits += 1;
958fb27d Andy McCall ↶ 1279 
            lap(5, &mut clock);
fb5b6181 Andy McCall ↶ 1280 
            done += 1;
1281 
            if trace {
1282 
                let built = kept
1283 
                    .get(gc.sha.as_str())
1284 
                    .map(|(st, _)| st.built_count())
1285 
                    .unwrap_or(0);
1286 
                eprintln!(
fd280228 Andy McCall ↶ 1287 
                    "trace {done} {} parents={} base={base_how} merged={merged} built={built} kept={} shed={shed} rss_mb={}",
fb5b6181 Andy McCall ↶ 1288 
                    &gc.sha[..12.min(gc.sha.len())],
1289 
                    gc.parents.len(),
1290 
                    kept.len(),
1291 
                    rss_mb()
1292 
                );
1293 
            }
958fb27d Andy McCall ↶ 1294 
        }
1295 
        for (name, d) in [
1296 
            "git import: bases",
2306a4d2 Andy McCall ↶ 1297 
            "git import: tree",
958fb27d Andy McCall ↶ 1298 
            "git import: render base",
1299 
            "git import: read blobs",
1300 
            "git import: build change",
1301 
            "git import: store change",
1302 
        ]
1303 
        .iter()
1304 
        .zip(spent)
1305 
        {
1306 
            crate::timing::record(name, d);
2306a4d2 Andy McCall ↶ 1307 
        }
1308 
        crate::timing::report_totals();
1309 
        if crate::timing::enabled() {
fd280228 Andy McCall ↶ 1310 
            eprintln!(
1311 
                "timing git import: {rebuilt} bases rebuilt from history, {shed} graphs shed"
1312 
            );
c3069537 Andy McCall ↶ 1313 
        }
1314
30ea9c1e Andy McCall ↶ 1315 
        // The bookmark: heads of the old state plus every new change.
1316 
        report.heads = old_heads.len();
1317 
        if last.is_none() {
1318 
            return Ok(report);
1319 
        }
dfb23a3a Andy McCall ↶ 1320 
        let heads: Vec<ChangeHash> = heads_now.into_iter().collect();
c3069537 Andy McCall ↶ 1321 
        report.heads = heads.len();
1322 
        let (op_ref, op) = self.head_op()?;
1323 
        let mut view = op.view.clone();
30ea9c1e Andy McCall ↶ 1324 
        let old_state_ref = view.bookmarks.get(&self.config.bookmark).copied();
1325 
        let state_ref = self.store.put_state(&StateObject::new(heads.clone()))?;
75167672 Andy McCall 1326 
        view.bookmarks
1327 
            .insert(self.config.bookmark.clone(), state_ref);
30ea9c1e Andy McCall ↶ 1328 
        self.record_op(&[op_ref], command, view)?;
1329 
        // The new state equals the tip's tree when the old one was itself a
1330 
        // Git commit's tree (fetched or exported), or there was none.
75167672 Andy McCall 1331 
        let old_known = old_heads.is_empty()
1332 
            || old_state_ref.is_some_and(|s| known.exported.values().any(|e| *e == s));
30ea9c1e Andy McCall ↶ 1333 
        if let (Some(tip), true) = (commits.last(), old_known) {
1334 
            self.record_git_state(git_dir, &state_ref, &tip.sha)?;
1335 
        }
1336 
        // The working copy follows when it was clean on the old heads.
1337 
        if old_heads.is_empty() {
1338 
            self.checkout(&heads, true, &format!("{command} (checkout)"))?;
1339 
        } else {
1340 
            let snap = self.snapshot(command)?;
1341 
            let clean = snap.wc.tree_ops.is_empty() && snap.wc.edits.is_empty();
1342 
            let mut base = snap.base.maximal(&snap.wc.deps);
1343 
            base.sort();
1344 
            let mut sorted_old = old_heads.clone();
1345 
            sorted_old.sort();
1346 
            if clean && base == sorted_old {
1347 
                self.checkout(&heads, false, &format!("{command} (checkout)"))?;
1348 
            }
1349 
        }
c3069537 Andy McCall ↶ 1350 
        Ok(report)
1351 
    }
1352
30ea9c1e Andy McCall ↶ 1353 
    /// Add to `st` the changes of every Git commit `rev-list args` lists:
1354 
    /// an imported commit's change, or the whole state an exported commit
1355 
    /// was made from (which may hold changes Git never saw).
75167672 Andy McCall 1356 
    fn add_git_ancestors(
958fb27d Andy McCall ↶ 1357 
        &self,
1358 
        st: &mut State,
1359 
        git_dir: &Path,
1360 
        args: &[&str],
1361 
        known: &GitKnown,
1362 
    ) -> Result<()> {
1363 
        self.add_git_ancestors_touching(st, git_dir, args, known, &mut HashSet::new())
1364 
    }
1365
1366 
    /// `add_git_ancestors`, also collecting in `touched` every node the
1367 
    /// changes it adds edit, place or remove.
1368 
    fn add_git_ancestors_touching(
75167672 Andy McCall 1369 
        &self,
1370 
        st: &mut State,
1371 
        git_dir: &Path,
1372 
        args: &[&str],
1373 
        known: &GitKnown,
958fb27d Andy McCall ↶ 1374 
        touched: &mut HashSet<NodeId>,
75167672 Andy McCall 1375 
    ) -> Result<()> {
30ea9c1e Andy McCall ↶ 1376 
        let mut full = vec!["rev-list"];
1377 
        full.extend(args);
1378 
        let out = git(git_dir, &full)?;
958fb27d Andy McCall ↶ 1379 
        let mut add = |st: &mut State, h: &ChangeHash| -> Result<()> {
1380 
            if st.contains(h) {
e83354cd Andy McCall ↶ 1381 
                return Ok(());
1382 
            }
1383 
            if !st.is_history_only() {
1384 
                let c = self.store.get_change(h)?;
1385 
                touched.extend(c.edits.iter().map(|e| e.node()));
1386 
                touched.extend(c.tree_ops.iter().map(|o| o.node()));
1387 
                st.add_unchecked(c);
958fb27d Andy McCall ↶ 1388 
                return Ok(());
e83354cd Andy McCall ↶ 1389 
            }
1390 
            // History only: the text of an edit waits unread unless its
1391 
            // file's graph is built.
1392 
            let (skeleton, texts) = self.store.get_change_lazy(h)?;
1393 
            touched.extend(skeleton.edits.iter().map(|e| e.node()));
1394 
            touched.extend(skeleton.tree_ops.iter().map(|o| o.node()));
1395 
            let built = skeleton
1396 
                .edits
1397 
                .iter()
1398 
                .any(|e| !matches!(e, Edit::Replace { .. }) && st.is_built(&e.node()));
1399 
            if built {
1400 
                let full = self.store.get_change(h)?;
1401 
                st.add_known_mixed(*h, full, skeleton, texts);
1402 
            } else {
1403 
                st.add_known(*h, skeleton, texts);
958fb27d Andy McCall ↶ 1404 
            }
1405 
            Ok(())
1406 
        };
c3069537 Andy McCall ↶ 1407 
        for sha in out.lines() {
30ea9c1e Andy McCall ↶ 1408 
            if let Some(h) = known.imported.get(sha) {
958fb27d Andy McCall ↶ 1409 
                add(st, h)?;
30ea9c1e Andy McCall ↶ 1410 
            } else if let Some(state_ref) = known.exported.get(sha) {
1411 
                let heads = self.store.get_state(state_ref)?.heads;
1412 
                for h in self.negotiate(&[], &heads)? {
958fb27d Andy McCall ↶ 1413 
                    add(st, &h)?;
30ea9c1e Andy McCall ↶ 1414 
                }
1415 
            }
1416 
        }
1417 
        Ok(())
1418 
    }
1419
1420 
    /// Commits this repository can relate to Oxygen history: imported
1421 
    /// ones by change, exported ones (through `git_dir`'s marks) by state.
1422 
    fn git_known(&self, git_dir: &Path) -> Result<GitKnown> {
1423 
        let imported = self.imported_commits()?;
1424 
        let marks_dir = self.git_marks_dir(git_dir)?;
1425 
        let oxy_marks = Self::read_oxy_marks(&marks_dir.join("oxy-marks"))?;
1426 
        let mut sha_of_mark: HashMap<u64, String> = HashMap::new();
1427 
        if let Ok(text) = fs::read_to_string(marks_dir.join("git-marks")) {
1428 
            for line in text.lines() {
1429 
                if let Some((m, sha)) = line.split_once(' ') {
1430 
                    if let Ok(m) = m.trim_start_matches(':').parse::<u64>() {
1431 
                        sha_of_mark.insert(m, sha.trim().to_string());
1432 
                    }
1433 
                }
1434 
            }
1435 
        }
1436 
        let mut exported = HashMap::new();
1437 
        for (state_ref, m) in oxy_marks {
1438 
            if let Some(sha) = sha_of_mark.get(&m) {
1439 
                exported.insert(sha.clone(), state_ref);
c3069537 Andy McCall ↶ 1440 
            }
1441 
        }
30ea9c1e Andy McCall ↶ 1442 
        for (state_ref, sha) in Self::read_git_states(&marks_dir.join("git-states"))? {
1443 
            exported.entry(sha).or_insert(state_ref);
1444 
        }
1445 
        Ok(GitKnown { imported, exported })
c3069537 Andy McCall ↶ 1446 
    }
1447
79ab8528 Andy McCall ↶ 1448 
    fn git_author(&self, c: &Change) -> String {
1449 
        match self.store.get_identity(&c.author) {
75167672 Andy McCall 1450 
            Ok(id) if !id.name.is_empty() => format!(
1451 
                "{} <{}>",
1452 
                id.name.replace(['<', '>', '\n'], " "),
1453 
                id.email.replace(['<', '>', '\n', ' '], "")
1454 
            ),
79ab8528 Andy McCall ↶ 1455 
            _ => "oxy <oxy@localhost>".to_string(),
1456 
        }
1457 
    }
1458
1459 
    #[allow(clippy::too_many_arguments)]
75167672 Andy McCall 1460 
    fn write_commit(
1461 
        &self,
1462 
        out: &mut Vec<u8>,
1463 
        reference: &str,
1464 
        mark: u64,
1465 
        author: &str,
1466 
        secs: u64,
1467 
        message: &str,
1468 
        parents: &[String],
1469 
        snapshot: &Snapshot,
1470 
    ) -> Result<()> {
79ab8528 Andy McCall ↶ 1471 
        out.extend_from_slice(format!("commit {reference}\nmark :{mark}\nauthor {author} {secs} +0000\ncommitter {author} {secs} +0000\ndata {}\n{message}\n", message.len()).as_bytes());
1472 
        for (i, p) in parents.iter().enumerate() {
75167672 Andy McCall 1473 
            out.extend_from_slice(
1474 
                format!("{} {p}\n", if i == 0 { "from" } else { "merge" }).as_bytes(),
1475 
            );
79ab8528 Andy McCall ↶ 1476 
        }
1477 
        out.extend_from_slice(b"deleteall\n");
4dee277c Andy McCall ↶ 1478 
        let has_gitmodules = snapshot.files.contains_key(".gitmodules");
79ab8528 Andy McCall ↶ 1479 
        for (path, f) in &snapshot.files {
4dee277c Andy McCall ↶ 1480 
            if path == ".oxymodules" {
1481 
                // Git reads `.gitmodules`; derive it unless the tree has one.
1482 
                if !has_gitmodules {
1483 
                    if let FileContent::Text { bytes, .. } = &f.content {
75167672 Andy McCall 1484 
                        let gm = crate::subrepo::Modules::parse(&String::from_utf8_lossy(bytes))
1485 
                            .to_gitmodules();
4dee277c Andy McCall ↶ 1486 
                        if !gm.is_empty() {
75167672 Andy McCall 1487 
                            out.extend_from_slice(
1488 
                                format!("M 100644 inline .gitmodules\ndata {}\n", gm.len())
1489 
                                    .as_bytes(),
1490 
                            );
4dee277c Andy McCall ↶ 1491 
                            out.extend_from_slice(gm.as_bytes());
1492 
                            out.push(b'\n');
1493 
                        }
1494 
                    }
1495 
                }
1496 
                continue;
1497 
            }
1498 
            if f.kind == NodeKind::Subrepo {
1499 
                if let FileContent::Opaque { blobs } = &f.content {
1500 
                    if let Some(b) = blobs.first() {
1501 
                        match crate::subrepo::Pin::decode(&self.read_blob(b)?)? {
75167672 Andy McCall 1502 
                            crate::subrepo::Pin::Git { sha, .. } => {
1503 
                                out.extend_from_slice(format!("M 160000 {sha} {path}\n").as_bytes())
1504 
                            }
1505 
                            crate::subrepo::Pin::Oxy { .. } => self
1506 
                                .git_export_skipped
1507 
                                .set(self.git_export_skipped.get() + 1),
4dee277c Andy McCall ↶ 1508 
                        }
1509 
                    }
1510 
                }
1511 
                continue;
1512 
            }
79ab8528 Andy McCall ↶ 1513 
            let (mode, bytes): (&str, Vec<u8>) = match (&f.kind, &f.content) {
1514 
                (NodeKind::Dir, _) => continue,
75167672 Andy McCall 1515 
                (NodeKind::Symlink, FileContent::Opaque { blobs }) => (
1516 
                    "120000",
1517 
                    blobs
1518 
                        .first()
1519 
                        .map(|b| self.read_blob(b))
1520 
                        .transpose()?
1521 
                        .unwrap_or_default(),
1522 
                ),
1523 
                (_, FileContent::Text { bytes, .. }) => {
1524 
                    (if f.mode.exec { "100755" } else { "100644" }, bytes.clone())
1525 
                }
1526 
                (_, FileContent::Opaque { blobs }) => (
1527 
                    if f.mode.exec { "100755" } else { "100644" },
1528 
                    blobs
1529 
                        .first()
1530 
                        .map(|b| self.read_blob(b))
1531 
                        .transpose()?
1532 
                        .unwrap_or_default(),
1533 
                ),
79ab8528 Andy McCall ↶ 1534 
                (_, FileContent::Dir) => continue,
1535 
            };
75167672 Andy McCall 1536 
            out.extend_from_slice(
1537 
                format!("M {mode} inline {path}\ndata {}\n", bytes.len()).as_bytes(),
1538 
            );
79ab8528 Andy McCall ↶ 1539 
            out.extend_from_slice(&bytes);
1540 
            out.push(b'\n');
1541 
        }
1542 
        out.push(b'\n');
1543 
        Ok(())
1544 
    }
1545 
}
62794ecc Andy McCall ↶ 1546
1547 
/// What `verify_git_import` found (spec §23: import must reproduce every tree).
1548 
#[derive(Debug, Clone, Default, PartialEq, Eq)]
1549 
pub struct GitVerifyReport {
1550 
    pub commits: usize,
1551 
    pub files_compared: usize,
1552 
    /// `(commit, what differs)`; empty when every commit's tree is reproduced.
1553 
    pub mismatches: Vec<(String, String)>,
1554 
}
1555
1556 
impl Repo {
1557 
    /// Check an import against its Git source: for every commit of
1558 
    /// `branch`, the state formed by the changes imported for it and all
1559 
    /// its ancestors must materialize to exactly that commit's tree, byte
1560 
    /// for byte, with the same executable bits, symlink targets and
1561 
    /// submodule pins. Reads only; stops listing after 50 differences.
1562 
    pub fn verify_git_import(&self, git_dir: &Path, branch: &str) -> Result<GitVerifyReport> {
1563 
        let commits = list_commits(git_dir, branch)?;
1564 
        let imported = self.imported_commits()?;
1565 
        let mut cat = CatFile::start(git_dir)?;
1566 
        let mut rep = GitVerifyReport::default();
1567 
        // Changes of each commit and all its ancestors, built up in topological order.
1568 
        let mut closure: HashMap<String, Arc<BTreeSet<ChangeHash>>> = HashMap::new();
a87c7b9d Andy McCall ↶ 1569 
        // Graphs (by key) and blobs already found equal to a Git blob.
1570 
        let mut verified_graphs: HashMap<[u8; 32], String> = HashMap::new();
1571 
        let mut verified_blobs: HashMap<ObjectRef, String> = HashMap::new();
57d8dc05 Andy McCall ↶ 1572 
        // The state of a commit whose first-parent children are still to
1573 
        // come, so a child adds its own changes instead of loading all its
1574 
        // history again (as the import does).
1575 
        let mut waiting: HashMap<&str, usize> = HashMap::new();
62794ecc Andy McCall ↶ 1576 
        for c in &commits {
57d8dc05 Andy McCall ↶ 1577 
            if let Some(p1) = c.parents.first() {
1578 
                *waiting.entry(p1.as_str()).or_default() += 1;
1579 
            }
1580 
        }
1581 
        let mut kept: HashMap<String, State> = HashMap::new();
1582 
        let mut kept_order: std::collections::VecDeque<String> = std::collections::VecDeque::new();
1583 
        for c in &commits {
62794ecc Andy McCall ↶ 1584 
            rep.commits += 1;
1585 
            let mut set: BTreeSet<ChangeHash> = BTreeSet::new();
1586 
            for p in &c.parents {
1587 
                if let Some(ps) = closure.get(p) {
1588 
                    set.extend(ps.iter().copied());
1589 
                }
1590 
            }
1591 
            match imported.get(&c.sha) {
1592 
                Some(h) => {
1593 
                    set.insert(*h);
1594 
                }
1595 
                None => {
1596 
                    rep.mismatches.push((
1597 
                        c.sha.clone(),
1598 
                        "no imported change carries this commit".into(),
1599 
                    ));
1600 
                }
1601 
            }
1602 
            let set = Arc::new(set);
1603 
            closure.insert(c.sha.clone(), set.clone());
1604 
            if rep.mismatches.len() >= 50 {
1605 
                break;
1606 
            }
1607 
            let heads: Vec<ChangeHash> = set.iter().copied().collect();
a87c7b9d Andy McCall ↶ 1608 
            // History only: a file whose graph (by key) or blob was already
1609 
            // found equal to this Git blob is not built, rendered or read
1610 
            // again; the rest are.
57d8dc05 Andy McCall ↶ 1611 
            let from_parent = c.parents.first().and_then(|p1| {
1612 
                let left = waiting.get_mut(p1.as_str()).map(|n| {
1613 
                    *n -= 1;
1614 
                    *n
1615 
                });
1616 
                if left == Some(0) {
1617 
                    kept_order.retain(|k| k != p1);
1618 
                    kept.remove(p1.as_str())
1619 
                } else {
1620 
                    kept.get(p1.as_str()).cloned()
1621 
                }
1622 
            });
1623 
            let state = match from_parent {
1624 
                Some(mut st) => {
1625 
                    // The changes the parent's state lacks. The state never
1626 
                    // builds a graph (they are built beside it), so they
1627 
                    // are read without their text.
1628 
                    for h in set.iter() {
1629 
                        if !st.contains(h) {
1630 
                            let (change, texts) = self.store.get_change_lazy(h)?;
1631 
                            st.add_known(*h, change, texts);
1632 
                        }
1633 
                    }
1634 
                    st
1635 
                }
1636 
                None => self.load_state_history(&heads)?,
1637 
            };
a87c7b9d Andy McCall ↶ 1638 
            let view = state.tree().view();
62794ecc Andy McCall ↶ 1639 
            let tree = list_tree(git_dir, &c.sha)?;
1640 
            let short = &c.sha[..12.min(c.sha.len())];
1641 
            let mut differ = |what: String| rep.mismatches.push((c.sha.clone(), what));
a87c7b9d Andy McCall ↶ 1642 
            let mut conflicted: Vec<String> =
1643 
                view.conflicts.iter().map(|c| format!("{c:?}")).collect();
1644 
            let mut need: Vec<NodeId> = Vec::new();
1645 
            for (path, (mode, sha)) in &tree {
1646 
                let Some(e) = view.entries.get(path) else {
1647 
                    continue;
1648 
                };
1649 
                if mode != "160000"
1650 
                    && mode != "120000"
1651 
                    && e.kind == NodeKind::File
1652 
                    && !e.mode.opaque
1653 
                {
1654 
                    let known = state
1655 
                        .graph_key(&e.node)
1656 
                        .is_some_and(|k| verified_graphs.get(&k) == Some(sha));
1657 
                    if !known {
1658 
                        need.push(e.node);
62794ecc Andy McCall ↶ 1659 
                    }
1660 
                }
1661 
            }
a87c7b9d Andy McCall ↶ 1662 
            // A graph's key is known only before it is built.
1663 
            let keys: HashMap<NodeId, [u8; 32]> = need
1664 
                .iter()
1665 
                .filter_map(|n| Some((*n, state.graph_key(n)?)))
1666 
                .collect();
1667 
            let need_set: HashSet<NodeId> = need.iter().copied().collect();
57d8dc05 Andy McCall ↶ 1668 
            let graphs = self.build_graphs_detached(&state, &need)?;
62794ecc Andy McCall ↶ 1669 
            for (path, (mode, sha)) in &tree {
a87c7b9d Andy McCall ↶ 1670 
                let Some(e) = view.entries.get(path) else {
62794ecc Andy McCall ↶ 1671 
                    differ(format!("{short}: {path} is missing"));
1672 
                    continue;
1673 
                };
1674 
                rep.files_compared += 1;
1675 
                match mode.as_str() {
a87c7b9d Andy McCall ↶ 1676 
                    "160000" => {
1677 
                        let blobs = state.live_blobs(&e.node);
1678 
                        match (e.kind, blobs.as_slice()) {
1679 
                            (NodeKind::Subrepo, [b]) => {
1680 
                                let pin = crate::subrepo::Pin::decode(&self.read_blob(b)?)?;
1681 
                                if !matches!(&pin, crate::subrepo::Pin::Git { sha: s, .. } if s == sha)
1682 
                                {
1683 
                                    differ(format!("{short}: {path} pins another commit"));
1684 
                                }
62794ecc Andy McCall ↶ 1685 
                            }
a87c7b9d Andy McCall ↶ 1686 
                            (kind, _) => differ(format!(
1687 
                                "{short}: {path} is not one submodule pin: {kind:?} with {} pins",
1688 
                                blobs.len()
1689 
                            )),
62794ecc Andy McCall ↶ 1690 
                        }
a87c7b9d Andy McCall ↶ 1691 
                    }
62794ecc Andy McCall ↶ 1692 
                    "120000" => {
1693 
                        let target = cat.read(sha)?;
a87c7b9d Andy McCall ↶ 1694 
                        let blobs = state.live_blobs(&e.node);
1695 
                        match (e.kind, blobs.as_slice()) {
1696 
                            (NodeKind::Symlink, [b]) if self.read_blob(b)? == target => {}
62794ecc Andy McCall ↶ 1697 
                            _ => differ(format!("{short}: symlink {path} differs")),
1698 
                        }
1699 
                    }
1700 
                    _ => {
a87c7b9d Andy McCall ↶ 1701 
                        if e.mode.exec != (mode == "100755") {
1702 
                            differ(format!("{short}: {path} executable bit differs"));
1703 
                        }
1704 
                        if e.kind == NodeKind::File && !e.mode.opaque {
1705 
                            if !need_set.contains(&e.node) {
1706 
                                // Known equal: skipped when the graphs were chosen.
62794ecc Andy McCall ↶ 1707 
                                continue;
1708 
                            }
57d8dc05 Andy McCall ↶ 1709 
                            let (have, m) = match graphs.get(&e.node) {
1710 
                                Some(g) => state.render_graph(g),
1711 
                                None => (Vec::new(), Default::default()),
1712 
                            };
a87c7b9d Andy McCall ↶ 1713 
                            if m.is_conflicted() {
57d8dc05 Andy McCall ↶ 1714 
                                // Who conflicts: a copy of the state with this graph built.
1715 
                                let mut detail = state.clone();
1716 
                                self.load_graphs_of(&mut detail, &[e.node])?;
1717 
                                let who: Vec<String> = detail
a87c7b9d Andy McCall ↶ 1718 
                                    .conflict_regions(&e.node)
1719 
                                    .contributors
1720 
                                    .iter()
1721 
                                    .filter_map(|h| state.get(h))
1722 
                                    .map(|m| {
1723 
                                        let sha = m
1724 
                                            .message
1725 
                                            .lines()
1726 
                                            .rev()
1727 
                                            .find_map(|l| l.strip_prefix("Git-Commit: "))
1728 
                                            .unwrap_or("?");
1729 
                                        format!(
1730 
                                            "{} {:?}",
1731 
                                            &sha[..12.min(sha.len())],
1732 
                                            m.message.lines().next().unwrap_or("")
1733 
                                        )
1734 
                                    })
1735 
                                    .collect();
1736 
                                conflicted.push(format!("{path} between {}", who.join(" and ")));
1737 
                            }
1738 
                            let want = cat.read(sha)?;
1739 
                            if have != want {
1740 
                                differ(format!(
1741 
                                    "{short}: {path} differs: {}",
1742 
                                    first_difference(&have, &want)
1743 
                                ));
1744 
                            } else if !m.is_conflicted() {
1745 
                                if let Some(k) = keys.get(&e.node) {
1746 
                                    verified_graphs.insert(*k, sha.clone());
1747 
                                }
1748 
                            }
1749 
                            continue;
1750 
                        }
1751 
                        let blobs = state.live_blobs(&e.node);
1752 
                        let [b] = blobs.as_slice() else {
1753 
                            differ(format!("{short}: {path} is not one file"));
1754 
                            continue;
62794ecc Andy McCall ↶ 1755 
                        };
a87c7b9d Andy McCall ↶ 1756 
                        if verified_blobs.get(b) == Some(sha) {
1757 
                            continue;
62794ecc Andy McCall ↶ 1758 
                        }
a87c7b9d Andy McCall ↶ 1759 
                        if self.read_blob(b)? != cat.read(sha)? {
1760 
                            differ(format!("{short}: {path} differs"));
1761 
                        } else {
1762 
                            verified_blobs.insert(*b, sha.clone());
62794ecc Andy McCall ↶ 1763 
                        }
1764 
                    }
1765 
                }
1766 
            }
a87c7b9d Andy McCall ↶ 1767 
            if !conflicted.is_empty() {
1768 
                differ(format!(
1769 
                    "{short}: the state is conflicted: {}",
1770 
                    conflicted.join("; ")
1771 
                ));
1772 
            }
62794ecc Andy McCall ↶ 1773 
            // Import records submodules in `.oxymodules` as well (spec §23),
1774 
            // and export leaves it out again.
1775 
            let has_submodules = tree.values().any(|(m, _)| m == "160000");
e83354cd Andy McCall ↶ 1776 
            for (path, e) in view.entries.iter() {
a87c7b9d Andy McCall ↶ 1777 
                if e.kind != NodeKind::Dir
62794ecc Andy McCall ↶ 1778 
                    && !tree.contains_key(path)
1779 
                    && !(has_submodules && path == ".oxymodules")
1780 
                {
1781 
                    differ(format!("{short}: {path} is extra"));
57d8dc05 Andy McCall ↶ 1782 
                }
1783 
            }
1784 
            if waiting.get(c.sha.as_str()).is_some_and(|n| *n > 0) {
1785 
                kept.insert(c.sha.clone(), state);
1786 
                kept_order.push_back(c.sha.clone());
1787 
                if kept_order.len() > 8 {
1788 
                    if let Some(old) = kept_order.pop_front() {
1789 
                        kept.remove(&old);
1790 
                    }
62794ecc Andy McCall ↶ 1791 
                }
1792 
            }
1793 
            if rep.mismatches.len() >= 50 {
1794 
                break;
1795 
            }
1796 
        }
1797 
        Ok(rep)
1798 
    }
1799 
}
1800
1801 
/// Where two versions of a file first part, for a verification report.
1802 
fn first_difference(have: &[u8], want: &[u8]) -> String {
1803 
    let (h, w): (Vec<&[u8]>, Vec<&[u8]>) = (
1804 
        have.split(|b| *b == b'\n').collect(),
1805 
        want.split(|b| *b == b'\n').collect(),
1806 
    );
1807 
    let i = h
1808 
        .iter()
1809 
        .zip(&w)
1810 
        .position(|(a, b)| a != b)
1811 
        .unwrap_or(h.len().min(w.len()));
1812 
    let show = |l: Option<&&[u8]>| {
1813 
        l.map(|l| String::from_utf8_lossy(&l[..l.len().min(60)]).into_owned())
1814 
            .unwrap_or_else(|| "(end)".into())
1815 
    };
1816 
    format!(
1817 
        "line {}: has {:?}, Git has {:?} ({} lines, Git has {})",
1818 
        i + 1,
1819 
        show(h.get(i)),
1820 
        show(w.get(i)),
1821 
        h.len(),
1822 
        w.len()
1823 
    )
1824 
}