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	store.commit() or { return err }123	report.head = final.id124	flags := events.latest_flags(store)!125	for record in fetched.saves {126		if flag := flags[record.id] {127			if flag.flagged {128				report.flagged << flag129			}130		}131	}132133	if head := head_before {134		if final.id == head.id {135			// The line did not move, so neither does the working tree.136			return report137		}138		files := history.files_of(store, final.id)!139		mut wanted := map[string]bool{}140		for file in files {141			wanted[file.path] = true142		}143		mut remove := []string{}144		for file in history.files_of(store, head.id)! {145			if file.path !in wanted {146				remove << file.path147			}148		}149		result := snapshot.restore_files(store, store.root, snapshot.load(store.root), files,150			remove)!151		report.removed = result.removed152	} else {153		files := history.files_of(store, final.id)!154		snapshot.restore_files(store, store.root, snapshot.load(store.root), files, [])!155	}156	history.refresh_log_md(store)157	return report158}159160// write_and_combine lands a fetch and, when the local line has saves the remote161// lacks, replays them on top of the remote head. It runs inside the caller's162// transaction and returns the new head.163fn write_and_combine(store &storage.Store, fetched Fetched, local_only []history.Save, clashes []Clash, side_of map[string]Side, mut report StealReport) !history.Save {164	write_fetched(store, fetched)!165	if fetched.saves.len == 0 {166		// The remote head is already on this line: level, or this machine is167		// ahead and the offer that follows a steal sends the rest.168		return history.head(store) or { return error('this repo has no head') }169	}170	if local_only.len == 0 {171		// Nothing of ours is missing from the remote: the line simply moves on.172		history.set_head(store, fetched.head)!173		report.forward = true174		return history.head(store) or { return error('the remote head did not land') }175	}176	remote_files := history.files_of(store, fetched.head)!177	mut mine_files := []snapshot.FileEntry{}178	if our_head := history.head(store) {179		mine_files = history.files_of(store, our_head.id)!180	}181182	// A file the remote changed is a clash, decided below, or one this machine183	// left as it was or changed the same way, so the remote's version stands.184	// Replaying this machine's steps on it could only conflict: an edit it185	// later undid would no longer apply.186	base_files := history.files_of(store, fetched.base)!187	settled := changed_paths(base_files, remote_files)188189	// Take the remote's line first, so the replayed saves chain onto it.190	history.set_head(store, fetched.head)!191	mut current := remote_files.clone()192	for save in local_only {193		parent_files := if save.parent != '' {194			history.files_of(store, save.parent)!195		} else {196			[]snapshot.FileEntry{}197		}198		target_files := history.files_of(store, save.id)!199		plan := apply.three_way(current, parent_files, target_files, [], settled)200		if apply.conflicts(plan) {201			return error('combine could not apply ${plan.conflict} — it is not a clash file')202		}203		if plan.applied.len == 0 {204			// Everything this save touched was a file the remote changed.205			continue206		}207		history.record_save(store, '', plan.files, history.SaveInput{208			message:       save.message209			created_at:    save.created_at210			machine:       save.machine211			agent:         save.agent212			session:       save.session213			is_checkpoint: save.is_checkpoint214			replayed_from: save.id215		})!216		current = plan.files.clone()217		report.replayed++218	}219220	mut mine_paths := []string{}221	for clash in clashes {222		if side_of[clash.path] == .mine {223			mine_paths << clash.path224		}225	}226	if mine_paths.len > 0 {227		// One last save puts my versions of the chosen files back.228		final_plan := apply.three_way(current, current, mine_files, mine_paths, [])229		if final_plan.applied.len > 0 {230			history.record_save(store, '', final_plan.files, history.SaveInput{231				message: 'combine: kept mine for ${mine_paths.len} ${if mine_paths.len == 1 {232					'file'233				} else {234					'files'235				}}'236			})!237			report.kept_mine = mine_paths.len238		}239	}240	return history.head(store) or { return error('the combined line has no head') }241}242243// find_clashes lists the files both sides changed, differently. Files changed244// identically on both sides are not clashes: there is nothing to decide.245pub fn find_clashes(base []snapshot.FileEntry, mine []snapshot.FileEntry, theirs []snapshot.FileEntry) []Clash {246	base_map := snapshot.index_files(base)247	mine_map := snapshot.index_files(mine)248	theirs_map := snapshot.index_files(theirs)249	mut paths := mine_map.keys()250	for path in theirs_map.keys() {251		if path !in mine_map {252			paths << path253		}254	}255	paths.sort()256	mut clashes := []Clash{}257	for path in paths {258		here := mine_map[path] or {259			snapshot.FileEntry{260				path: path261			}262		}263		there := theirs_map[path] or {264			snapshot.FileEntry{265				path: path266			}267		}268		was := base_map[path] or {269			snapshot.FileEntry{270				path: path271			}272		}273		if same(here, there) {274			continue275		}276		if same(here, was) || same(there, was) {277			continue278		}279		clashes << Clash{280			path:   path281			mine:   here282			theirs: there283			base:   was284		}285	}286	return clashes287}288289// changed_paths lists the paths whose entry differs between two trees.290fn changed_paths(before []snapshot.FileEntry, after []snapshot.FileEntry) []string {291	was := snapshot.index_files(before)292	now := snapshot.index_files(after)293	mut paths := []string{}294	for path, entry in now {295		if !same(entry, was[path] or { snapshot.FileEntry{} }) {296			paths << path297		}298	}299	for path, _ in was {300		if path !in now {301			paths << path302		}303	}304	return paths305}306307// combine_clashes finds the files both lines changed, with the diff stat each308// prompt shows. It reads the remote head from the fetch, not the store, so it309// can run before anything is written.310fn combine_clashes(store &storage.Store, fetched Fetched) ![]Clash {311	base_files := if fetched.base != '' {312		history.files_of(store, fetched.base)!313	} else {314		[]snapshot.FileEntry{}315	}316	head := history.head(store) or { return []Clash{} }317	// The transaction applies the incoming strips before it replays, so the318	// clashes are found on the trees as they will be then, or the answers319	// would be about files that are no longer there.320	strips := fetched.events_out.filter(it.kind == events.kind_strip)321	// Base and mine are here already, so every strip reaches them (created 0).322	remote_head := fetched.saves.last()323	mut clashes := find_clashes(after_strips(base_files, fetched.base, 0, strips), after_strips(history.files_of(store,324		head.id)!, head.id, 0, strips), after_strips(to_files(remote_head), remote_head.id,325		remote_head.created_at, strips))326	for mut clash in clashes {327		mine := fetched_bytes(store, fetched, clash.mine)!328		theirs := fetched_bytes(store, fetched, clash.theirs)!329		delta := snapshot.diff_bytes(clash.path, 'modified', mine, theirs)!330		clash.stat = if delta.binary { 'binary' } else { '+${delta.added} -${delta.removed}' }331	}332	return clashes333}334335// after_strips is a save's file list once the given strip events are applied,336// by the same rule write_fetched uses.337fn after_strips(files []snapshot.FileEntry, save_id string, created_at i64, strips []events.Event) []snapshot.FileEntry {338	return files.filter(fn [strips, save_id, created_at] (file snapshot.FileEntry) bool {339		return !strips.any(reaches(it, save_id, created_at) && covers(it.path, file.path))340	})341}342343// fetched_bytes reads one side of a clash: from the fetch when it has just344// arrived, else from the store. An absent side is empty.345fn fetched_bytes(store &storage.Store, fetched Fetched, entry snapshot.FileEntry) ![]u8 {346	if entry.hash == '' {347		return []u8{}348	}349	if blob := fetched.repo_objects[entry.hash] {350		return zlib.decompress(blob)!351	}352	return store.get_object(entry.hash)!353}354355fn choose_sides(clashes []Clash, opts StealOptions) !map[string]Side {356	mut sides := map[string]Side{}357	if clashes.len == 0 {358		return sides359	}360	for clash in clashes {361		if opts.mine {362			sides[clash.path] = .mine363			continue364		}365		if opts.theirs {366			sides[clash.path] = .theirs367			continue368		}369		if opts.ask != unsafe { nil } {370			sides[clash.path] = opts.ask(clash)371			continue372		}373		return error('${clashes.len} files changed on both machines — run on a terminal to choose, or pass --mine or --theirs')374	}375	return sides376}377378fn same(a snapshot.FileEntry, b snapshot.FileEntry) bool {379	if a.hash == '' && b.hash == '' {380		return true381	}382	return a.hash == b.hash && a.executable == b.executable383}384385fn changed_since(store &storage.Store, before ?history.Save, strips_before int) !bool {386	if events.strip_events(store)!.len != strips_before {387		return true388	}389	now := history.head(store)390	if now or { history.Save{} }.id != before or { history.Save{} }.id {391		return true392	}393	return dirty(store, now)!394}395396// dirty reports whether the working tree holds anything a steal would397// overwrite. With no saves yet, that is every file but the untouched ignore398// list `koh init` wrote.399fn dirty(store &storage.Store, head ?history.Save) !bool {400	files := snapshot.walk(store, store.root, snapshot.load(store.root))!401	if saved := head {402		return !snapshot.same_tree(history.files_of(store, saved.id)!, files)403	}404	seed := storage.hash_hex(snapshot.default_ignore.bytes())405	return files.any(it.path != snapshot.ignore_file || it.hash != seed)406}407408// fetch_repo pulls a whole repo — head, saves, objects and events — for409// `koh steal <name>`, which lands a copy in a new folder.410pub fn fetch_repo(mut backend remote.Backend, name string) !Fetched {411	remote_head := backend.get(head_key(name))!412	if !remote_head.found {413		return error('the remote has no repo called "${name}"')414	}415	save_ids := backend.list('${name}/saves')!416	object_keys := backend.list('${name}/objects')!417	event_keys := backend.list('${name}/events')!418	mut fetched := Fetched{419		head: remote_head.body.bytestr().trim_space()420	}421	for key in object_keys {422		blob := backend.get(key)!423		if blob.found {424			fetched.repo_objects[os.base(key)] = blob.body425		}426	}427	for key in save_ids {428		body := backend.get(key)!429		if !body.found {430			continue431		}432		record := decode_save(body.body)!433		fetched.saves << record434	}435	fetched.saves = parent_first(fetched.saves)!436	for key in event_keys {437		body := backend.get(key)!438		if body.found {439			fetched.events_out << decode_event(body.body)!440		}441	}442	return fetched443}444445// parent_first orders saves so every parent lands before its children, as the446// foreign key on `parent` demands. Time cannot do it: a replayed save keeps its447// original, older time but sits on a newer parent.448fn parent_first(records []SaveRecord) ![]SaveRecord {449	mut by_id := map[string]SaveRecord{}450	for record in records {451		by_id[record.id] = record452	}453	mut placed := map[string]bool{}454	mut ordered := []SaveRecord{}455	for record in records {456		mut chain := []SaveRecord{}457		mut cursor := record.id458		for cursor != '' && cursor !in placed {459			if chain.len > records.len {460				return error('the save records on the remote form a loop')461			}462			current := by_id[cursor] or { return error('the remote is missing save ${cursor}') }463			chain << current464			cursor = current.parent465		}466		chain.reverse_in_place()467		for current in chain {468			placed[current.id] = true469			ordered << current470		}471	}472	return ordered473}