steal.v

1//! Stealing: bringing another machine's saves in, and combining when both2//! machines moved.3//!4//! The order is always the same, and it is what makes an interrupted steal5//! harmless: read everything from the remote first, then land it in one6//! transaction, then write the working tree. Divergence is resolved per file —7//! a clash is never merged line by line, it is chosen.89module sync1011import compress.zlib12import os13import src.apply14import src.events15import src.history16import src.remote17import src.snapshot18import src.storage1920pub enum Side {21	mine22	theirs23}2425// Clash is one file both machines changed differently. `stat` is the short26// diff stat the terminal prompt shows, so the choice is informed.27pub struct Clash {28pub mut:29	path   string30	mine   snapshot.FileEntry31	theirs snapshot.FileEntry32	base   snapshot.FileEntry33	stat   string34}3536pub struct StealOptions {37pub mut:38	mine   bool // resolve every clash as mine39	theirs bool // resolve every clash as theirs40	// Asked once per clash when neither flag is set. Nil means "nobody to ask".41	ask fn (clash Clash) Side = unsafe { nil }42}4344pub struct StealReport {45pub mut:46	repo            string47	forward         bool48	fetched_saves   int49	fetched_objects int50	fetched_events  int51	clashes         []string52	kept_mine       int53	replayed        int54	removed         int55	head            string56	flagged         []events.Flag // incoming saves that are flagged57}5859pub const dirty_tree = 'save or checkpoint first — steal needs a clean tree'60pub const no_saves_here = 'this folder has files but no saves — steal a repo into an empty folder with "koh steal <name>"'6162// steal brings the remote's new saves in. Inside a repo with its own work this63// may have to combine; a repo with nothing of its own simply fast-forwards.64pub fn steal(store &storage.Store, mut backend remote.Backend, opts StealOptions) !StealReport {65	name := repo_name(store, mut backend)!66	head_before := history.head(store)67	// Combining rewrites the working tree, so unsaved work would be lost.68	if dirty(store, head_before)! {69		if head_before == none {70			return error(no_saves_here)71		}72		return error(dirty_tree)73	}74	fetched := fetch_back_to_known(store, mut backend, name)!75	mut report := StealReport{76		repo:            name77		fetched_saves:   fetched.saves.len78		fetched_objects: fetched.repo_objects.len79		fetched_events:  fetched.events_out.len80		head:            fetched.head81	}82	mut local_only := []history.Save{}83	if head := head_before {84		// The remote builds on `base`. If that save is here but no longer on85		// the line, neither line contains the other; jumping to the remote86		// would silently leave this line behind.87		local_only = saves_after(store, head.id, fetched.base) or {88			return error('the remote builds on save ${fetched.base}, which is not on this line — load the save you want, then offer')89		}90	}91	if fetched.base == '' && local_only.len > 0 {92		return error('the remote keeps a different history — it shares no save with this repo')93	}9495	strips_before := events.strip_events(store)!.len96	mut clashes := []Clash{}97	mut side_of := map[string]Side{}98	if fetched.saves.len > 0 && local_only.len > 0 {99		clashes = combine_clashes(store, fetched)!100		report.clashes = clashes.map(it.path)101		// Asked before the write lock: a person thinking must not stall every102		// agent's save on this repo.103		side_of = choose_sides(clashes, opts)!104	}105106	store.begin() or { return err }107	// The questions ran without the lock. If a save or an edit landed108	// meanwhile, the answers were about a line that no longer exists.109	if moved := changed_since(store, head_before, strips_before) {110		if moved {111			store.rollback()112			return error('this repo changed while steal was asking — run koh steal again')113		}114	} else {115		store.rollback()116		return err117	}118	final := (write_and_combine(store, fetched, local_only, clashes, side_of, mut report) or {119		store.rollback()120		return err121	})122	files := history.files_of(store, final.id) or {123		store.rollback()124		return err125	}126	snapshot.check_links(store.root, files) or {127		store.rollback()128		return err129	}130	store.commit() or { return err }131	report.head = final.id132	flags := events.latest_flags(store)!133	for record in fetched.saves {134		if flag := flags[record.id] {135			if flag.flagged {136				report.flagged << flag137			}138		}139	}140141	if head := head_before {142		if final.id == head.id {143			// The line did not move, so neither does the working tree.144			return report145		}146		mut wanted := map[string]bool{}147		for file in files {148			wanted[file.path] = true149		}150		mut remove := []string{}151		for file in history.files_of(store, head.id)! {152			if file.path !in wanted {153				remove << file.path154			}155		}156		result := snapshot.restore_files(store, store.root, snapshot.load(store.root), files,157			remove)!158		report.removed = result.removed159	} else {160		snapshot.restore_files(store, store.root, snapshot.load(store.root), files, [])!161	}162	history.refresh_log_md(store)163	return report164}165166// write_and_combine lands a fetch and, when the local line has saves the remote167// lacks, replays them on top of the remote head. It runs inside the caller's168// transaction and returns the new head.169fn write_and_combine(store &storage.Store, fetched Fetched, local_only []history.Save, clashes []Clash, side_of map[string]Side, mut report StealReport) !history.Save {170	write_fetched(store, fetched)!171	if fetched.saves.len == 0 {172		// The remote head is already on this line: level, or this machine is173		// ahead and the offer that follows a steal sends the rest.174		return history.head(store) or { return error('this repo has no head') }175	}176	if local_only.len == 0 {177		// Nothing of ours is missing from the remote: the line simply moves on.178		history.set_head(store, fetched.head)!179		report.forward = true180		return history.head(store) or { return error('the remote head did not land') }181	}182	remote_files := history.files_of(store, fetched.head)!183	mut mine_files := []snapshot.FileEntry{}184	if our_head := history.head(store) {185		mine_files = history.files_of(store, our_head.id)!186	}187188	// A file the remote changed is a clash, decided below, or one this machine189	// left as it was or changed the same way, so the remote's version stands.190	// Replaying this machine's steps on it could only conflict: an edit it191	// later undid would no longer apply.192	base_files := history.files_of(store, fetched.base)!193	settled := changed_paths(base_files, remote_files)194195	// Take the remote's line first, so the replayed saves chain onto it.196	history.set_head(store, fetched.head)!197	mut current := remote_files.clone()198	for save in local_only {199		parent_files := if save.parent != '' {200			history.files_of(store, save.parent)!201		} else {202			[]snapshot.FileEntry{}203		}204		target_files := history.files_of(store, save.id)!205		plan := apply.three_way(current, parent_files, target_files, [], settled)206		if apply.conflicts(plan) {207			return error('combine could not apply ${plan.conflict} — it is not a clash file')208		}209		if plan.applied.len == 0 {210			// Everything this save touched was a file the remote changed.211			continue212		}213		history.record_save(store, '', plan.files, history.SaveInput{214			message:       save.message215			created_at:    save.created_at216			machine:       save.machine217			agent:         save.agent218			session:       save.session219			is_checkpoint: save.is_checkpoint220			replayed_from: save.id221		})!222		current = plan.files.clone()223		report.replayed++224	}225226	mut mine_paths := []string{}227	for clash in clashes {228		if side_of[clash.path] == .mine {229			mine_paths << clash.path230		}231	}232	if mine_paths.len > 0 {233		// One last save puts my versions of the chosen files back.234		final_plan := apply.three_way(current, current, mine_files, mine_paths, [])235		if final_plan.applied.len > 0 {236			history.record_save(store, '', final_plan.files, history.SaveInput{237				message: 'combine: kept mine for ${mine_paths.len} ${if mine_paths.len == 1 {238					'file'239				} else {240					'files'241				}}'242			})!243			report.kept_mine = mine_paths.len244		}245	}246	return history.head(store) or { return error('the combined line has no head') }247}248249// find_clashes lists the files both sides changed, differently. Files changed250// identically on both sides are not clashes: there is nothing to decide.251pub fn find_clashes(base []snapshot.FileEntry, mine []snapshot.FileEntry, theirs []snapshot.FileEntry) []Clash {252	base_map := snapshot.index_files(base)253	mine_map := snapshot.index_files(mine)254	theirs_map := snapshot.index_files(theirs)255	mut paths := mine_map.keys()256	for path in theirs_map.keys() {257		if path !in mine_map {258			paths << path259		}260	}261	paths.sort()262	mut clashes := []Clash{}263	for path in paths {264		here := mine_map[path] or {265			snapshot.FileEntry{266				path: path267			}268		}269		there := theirs_map[path] or {270			snapshot.FileEntry{271				path: path272			}273		}274		was := base_map[path] or {275			snapshot.FileEntry{276				path: path277			}278		}279		if same(here, there) {280			continue281		}282		if same(here, was) || same(there, was) {283			continue284		}285		clashes << Clash{286			path:   path287			mine:   here288			theirs: there289			base:   was290		}291	}292	return clashes293}294295// changed_paths lists the paths whose entry differs between two trees.296fn changed_paths(before []snapshot.FileEntry, after []snapshot.FileEntry) []string {297	was := snapshot.index_files(before)298	now := snapshot.index_files(after)299	mut paths := []string{}300	for path, entry in now {301		if !same(entry, was[path] or { snapshot.FileEntry{} }) {302			paths << path303		}304	}305	for path, _ in was {306		if path !in now {307			paths << path308		}309	}310	return paths311}312313// combine_clashes finds the files both lines changed, with the diff stat each314// prompt shows. It reads the remote head from the fetch, not the store, so it315// can run before anything is written.316fn combine_clashes(store &storage.Store, fetched Fetched) ![]Clash {317	base_files := if fetched.base != '' {318		history.files_of(store, fetched.base)!319	} else {320		[]snapshot.FileEntry{}321	}322	head := history.head(store) or { return []Clash{} }323	// The transaction applies the incoming strips before it replays, so the324	// clashes are found on the trees as they will be then, or the answers325	// would be about files that are no longer there.326	strips := fetched.events_out.filter(it.kind == events.kind_strip)327	// Base and mine are here already, so every strip reaches them (created 0).328	remote_head := fetched.saves.last()329	mut clashes := find_clashes(after_strips(base_files, fetched.base, 0, strips), after_strips(history.files_of(store,330		head.id)!, head.id, 0, strips), after_strips(to_files(remote_head), remote_head.id,331		remote_head.created_at, strips))332	for mut clash in clashes {333		mine := fetched_bytes(store, fetched, clash.mine)!334		theirs := fetched_bytes(store, fetched, clash.theirs)!335		delta := snapshot.diff_bytes(clash.path, 'modified', mine, theirs)!336		clash.stat = if delta.binary { 'binary' } else { '+${delta.added} -${delta.removed}' }337	}338	return clashes339}340341// after_strips is a save's file list once the given strip events are applied,342// by the same rule write_fetched uses.343fn after_strips(files []snapshot.FileEntry, save_id string, created_at i64, strips []events.Event) []snapshot.FileEntry {344	return files.filter(fn [strips, save_id, created_at] (file snapshot.FileEntry) bool {345		return !strips.any(reaches(it, save_id, created_at) && covers(it.path, file.path))346	})347}348349// fetched_bytes reads one side of a clash: from the fetch when it has just350// arrived, else from the store. An absent side is empty.351fn fetched_bytes(store &storage.Store, fetched Fetched, entry snapshot.FileEntry) ![]u8 {352	if entry.hash == '' {353		return []u8{}354	}355	if blob := fetched.repo_objects[entry.hash] {356		return zlib.decompress(blob)!357	}358	return store.get_object(entry.hash)!359}360361fn choose_sides(clashes []Clash, opts StealOptions) !map[string]Side {362	mut sides := map[string]Side{}363	if clashes.len == 0 {364		return sides365	}366	for clash in clashes {367		if opts.mine {368			sides[clash.path] = .mine369			continue370		}371		if opts.theirs {372			sides[clash.path] = .theirs373			continue374		}375		if opts.ask != unsafe { nil } {376			sides[clash.path] = opts.ask(clash)377			continue378		}379		return error('${clashes.len} files changed on both machines — run on a terminal to choose, or pass --mine or --theirs')380	}381	return sides382}383384fn same(a snapshot.FileEntry, b snapshot.FileEntry) bool {385	if a.hash == '' && b.hash == '' {386		return true387	}388	return a.hash == b.hash && a.executable == b.executable389}390391fn changed_since(store &storage.Store, before ?history.Save, strips_before int) !bool {392	if events.strip_events(store)!.len != strips_before {393		return true394	}395	now := history.head(store)396	if now or { history.Save{} }.id != before or { history.Save{} }.id {397		return true398	}399	return dirty(store, now)!400}401402// dirty reports whether the working tree holds anything a steal would403// overwrite. With no saves yet, that is every file but the untouched ignore404// list `koh init` wrote.405fn dirty(store &storage.Store, head ?history.Save) !bool {406	files := snapshot.walk(store, store.root, snapshot.load(store.root))!407	if saved := head {408		return !snapshot.same_tree(history.files_of(store, saved.id)!, files)409	}410	seed := storage.hash_hex(snapshot.default_ignore.bytes())411	return files.any(it.path != snapshot.ignore_file || it.hash != seed)412}413414// fetch_repo pulls a whole repo — head, saves, objects and events — for415// `koh steal <name>`, which lands a copy in a new folder.416pub fn fetch_repo(mut backend remote.Backend, name string) !Fetched {417	remote_head := backend.get(head_key(name))!418	if !remote_head.found {419		return error('the remote has no repo called "${name}"')420	}421	save_ids := backend.list('${name}/saves')!422	object_keys := backend.list('${name}/objects')!423	event_keys := backend.list('${name}/events')!424	mut fetched := Fetched{425		head: remote_head.body.bytestr().trim_space()426	}427	for key in object_keys {428		blob := backend.get(key)!429		if blob.found {430			fetched.repo_objects[os.base(key)] = blob.body431		}432	}433	for key in save_ids {434		body := backend.get(key)!435		if !body.found {436			continue437		}438		record := decode_save(body.body)!439		fetched.saves << record440	}441	fetched.saves = parent_first(fetched.saves)!442	for key in event_keys {443		body := backend.get(key)!444		if body.found {445			fetched.events_out << decode_event(body.body)!446		}447	}448	return fetched449}450451// parent_first orders saves so every parent lands before its children, as the452// foreign key on `parent` demands. Time cannot do it: a replayed save keeps its453// original, older time but sits on a newer parent.454fn parent_first(records []SaveRecord) ![]SaveRecord {455	mut by_id := map[string]SaveRecord{}456	for record in records {457		by_id[record.id] = record458	}459	mut placed := map[string]bool{}460	mut ordered := []SaveRecord{}461	for record in records {462		mut chain := []SaveRecord{}463		mut cursor := record.id464		for cursor != '' && cursor !in placed {465			if chain.len > records.len {466				return error('the save records on the remote form a loop')467			}468			current := by_id[cursor] or { return error('the remote is missing save ${cursor}') }469			chain << current470			cursor = current.parent471		}472		chain.reverse_in_place()473		for current in chain {474			placed[current.id] = true475			ordered << current476		}477	}478	return ordered479}