//! The tree graph: nodes, placements, and modes as atoms (spec §7.1, §7.2), //! with tree conflict detection and path assignment (spec §9.2). use crate::hash::ChangeHash; use crate::ids::NodeId; use crate::objects::{Mode, NodeKind, OpRef, TreeOp}; use std::collections::{BTreeMap, BTreeSet, HashMap, HashSet}; /// `(change, index into that change's tree_ops)`. pub type OpKey = (ChangeHash, u32); #[derive(Clone, Debug, PartialEq, Eq)] pub struct Placement { pub node: NodeId, pub parent: NodeId, pub name: String, } #[derive(Clone, Debug, Default)] pub struct TreeGraph { kinds: HashMap, introduced: HashMap, placements: HashMap, modes: HashMap, /// Placement atoms retired by `RemoveNode`/`Move`, and mode atoms retired /// by `SetMode`. Kept apart because `AddNode` introduces one of each /// under the same key. deleted_placements: HashMap>, deleted_modes: HashMap>, placements_of: HashMap>, modes_of: HashMap>, } #[derive(Clone, Debug, PartialEq, Eq)] pub enum TreeConflict { /// Two live nodes share `(parent, name)`. Name { parent: NodeId, name: String, nodes: Vec, }, /// One node has several live placements. MultiPlacement { node: NodeId, placements: Vec, }, /// A live node's parent is not live. Orphan { node: NodeId, parent: NodeId }, /// A directory is placed inside its own descendant. Cycle { node: NodeId }, /// Several live mode atoms. Mode { node: NodeId, modes: Vec }, } /// A node as it appears in the working copy. #[derive(Clone, Debug, PartialEq, Eq)] pub struct TreeEntry { pub path: String, pub node: NodeId, pub kind: NodeKind, pub mode: Mode, /// The placement this entry materializes (for building `RemoveNode`/`Move`). pub placement: OpKey, /// The live mode atom (for building `SetMode`). pub mode_atom: Option, } #[derive(Clone, Debug, Default, PartialEq, Eq)] pub struct TreeView { /// Path → entry, paths use `/` and have no leading slash; conflicted /// entries carry a `.oxy-conflict-` suffix. pub entries: BTreeMap, pub conflicts: Vec, } impl TreeView { pub fn is_conflicted(&self) -> bool { !self.conflicts.is_empty() } pub fn by_node(&self) -> HashMap { self.entries.values().map(|e| (e.node, e)).collect() } } impl TreeGraph { pub fn new() -> Self { Self::default() } /// Apply tree op `index` of `change`; `OpRef` self references must already be resolved. pub fn apply(&mut self, change: ChangeHash, index: u32, op: &TreeOp) { let key = (change, index); match op { TreeOp::AddNode { node, kind, mode, parent, name, } => { self.kinds.insert(*node, *kind); self.introduced.insert(*node, change); self.placements.insert( key, Placement { node: *node, parent: *parent, name: name.clone(), }, ); self.placements_of.entry(*node).or_default().insert(key); if *kind != NodeKind::Dir { self.modes.insert(key, (*node, *mode)); self.modes_of.entry(*node).or_default().insert(key); } } TreeOp::RemoveNode { placement, .. } => { self.deleted_placements .entry(op_key(*placement)) .or_default() .insert(change); } TreeOp::Move { node, old, parent, name, } => { self.deleted_placements .entry(op_key(*old)) .or_default() .insert(change); self.placements.insert( key, Placement { node: *node, parent: *parent, name: name.clone(), }, ); self.placements_of.entry(*node).or_default().insert(key); } TreeOp::SetMode { node, old, mode } => { self.deleted_modes .entry(op_key(*old)) .or_default() .insert(change); self.modes.insert(key, (*node, *mode)); self.modes_of.entry(*node).or_default().insert(key); } } } fn placement_live(&self, k: &OpKey) -> bool { self.deleted_placements.get(k).is_none_or(|d| d.is_empty()) } fn mode_live(&self, k: &OpKey) -> bool { self.deleted_modes.get(k).is_none_or(|d| d.is_empty()) } pub fn kind(&self, node: &NodeId) -> Option { if node.is_root() { Some(NodeKind::Dir) } else { self.kinds.get(node).copied() } } /// The change whose `AddNode` introduced `node`. pub fn introducer(&self, node: &NodeId) -> Option { self.introduced.get(node).copied() } pub fn placement(&self, k: &OpKey) -> Option<&Placement> { self.placements.get(k) } pub fn live_placements(&self, node: &NodeId) -> Vec<(OpKey, &Placement)> { self.placements_of .get(node) .into_iter() .flat_map(|ks| ks.iter()) .filter(|k| self.placement_live(k)) .filter_map(|k| self.placements.get(k).map(|p| (*k, p))) .collect() } pub fn live_modes(&self, node: &NodeId) -> Vec<(OpKey, Mode)> { self.modes_of .get(node) .into_iter() .flat_map(|ks| ks.iter()) .filter(|k| self.mode_live(k)) .filter_map(|k| self.modes.get(k).map(|(_, m)| (*k, *m))) .collect() } /// Every node that has at least one live placement. pub fn placed_nodes(&self) -> impl Iterator + '_ { self.placements_of .keys() .copied() .filter(|n| !self.live_placements(n).is_empty()) } /// Resolve the tree: assign paths, find conflicts. pub fn view(&self) -> TreeView { let mut view = TreeView::default(); // Live placements grouped by parent. let mut children: HashMap> = HashMap::new(); for (k, p) in &self.placements { if self.placement_live(k) { children.entry(p.parent).or_default().push((*k, p)); } } for c in children.values_mut() { c.sort_by(|a, b| a.1.name.cmp(&b.1.name).then(a.0.cmp(&b.0))); } // Multi-placement and mode conflicts. let mut nodes: Vec = self.placements_of.keys().copied().collect(); nodes.sort(); let mut multi: HashSet = HashSet::new(); // Each node's mode and the atom it comes from, for the entries. let mut modes: HashMap, Mode)> = HashMap::new(); for n in &nodes { let lp = self.live_placements(n); // One placement (nearly every node) cannot be in two places. if lp.len() > 1 { let distinct_places: BTreeSet<(&NodeId, &str)> = lp .iter() .map(|(_, p)| (&p.parent, p.name.as_str())) .collect(); if distinct_places.len() > 1 { multi.insert(*n); view.conflicts.push(TreeConflict::MultiPlacement { node: *n, placements: lp.iter().map(|(k, _)| *k).collect(), }); } } let lm = self.live_modes(n); if lm.len() > 1 { let distinct_modes: BTreeSet = lm.iter().map(|(_, m)| *m).collect(); if distinct_modes.len() > 1 { view.conflicts.push(TreeConflict::Mode { node: *n, modes: distinct_modes.into_iter().collect(), }); } } let mode_atom = lm.iter().map(|(k, _)| *k).max(); let mode = mode_atom .and_then(|k| self.modes.get(&k).map(|(_, m)| *m)) .unwrap_or_default(); modes.insert(*n, (mode_atom, mode)); } // BFS from the root, assigning paths. let mut reached: HashSet = HashSet::new(); reached.insert(NodeId::ROOT); let mut queue: Vec<(NodeId, String)> = vec![(NodeId::ROOT, String::new())]; let mut on_path: HashSet = HashSet::new(); while let Some((dir, dir_path)) = queue.pop() { let Some(kids) = children.get(&dir) else { continue; }; // Name conflicts within this directory: the children are sorted // by name, so each name's group is a run of them. let mut groups: Vec<(&str, &[(OpKey, &Placement)])> = Vec::new(); let mut start = 0; for i in 1..=kids.len() { if i == kids.len() || kids[i].1.name != kids[start].1.name { groups.push((kids[start].1.name.as_str(), &kids[start..i])); start = i; } } for (name, group) in groups { let distinct: BTreeSet = group.iter().map(|(_, p)| p.node).collect(); if distinct.len() > 1 { view.conflicts.push(TreeConflict::Name { parent: dir, name: name.to_string(), nodes: distinct.iter().copied().collect(), }); } let suffixed = distinct.len() > 1; for (i, (k, p)) in group.iter().enumerate() { let node = p.node; if node == dir || on_path.contains(&node) { view.conflicts.push(TreeConflict::Cycle { node }); continue; } let needs_suffix = suffixed || multi.contains(&node); let leaf = if needs_suffix { format!("{name}.oxy-conflict-{}", i + 1) } else { name.to_string() }; let path = if dir_path.is_empty() { leaf } else { format!("{dir_path}/{leaf}") }; let kind = self.kind(&node).unwrap_or(NodeKind::File); let (mode_atom, mode) = modes.get(&node).copied().unwrap_or_default(); view.entries.insert( path.clone(), TreeEntry { path: path.clone(), node, kind, mode, placement: *k, mode_atom, }, ); if kind == NodeKind::Dir && reached.insert(node) { on_path.insert(dir); queue.push((node, path)); } else { reached.insert(node); } } } } // Nodes with live placements that were never reached: orphans or cycles. for n in &nodes { if reached.contains(n) { continue; } let lp = self.live_placements(n); if lp.is_empty() { continue; } let mut seen = HashSet::new(); let mut cur = *n; let conflict = loop { if !seen.insert(cur) { break TreeConflict::Cycle { node: *n }; } let lp = self.live_placements(&cur); let Some((_, p)) = lp.first() else { break TreeConflict::Orphan { node: *n, parent: cur, }; }; if reached.contains(&p.parent) { // Parent reached but this node was not: the parent is not a // directory, so this is a name/kind problem surfaced as orphan. break TreeConflict::Orphan { node: *n, parent: p.parent, }; } cur = p.parent; }; if !view.conflicts.contains(&conflict) { view.conflicts.push(conflict); } } view } } impl TreeGraph { /// Serialize for a checkpoint. pub fn to_value(&self) -> crate::cbor::Value { use crate::cbor::Value; let opkey = |k: &OpKey| vec![Value::Bytes(k.0 .0.to_vec()), Value::Uint(k.1 as u64)]; let deleters = |m: &HashMap>| { let mut ks: Vec<&OpKey> = m.keys().collect(); ks.sort(); Value::Array( ks.into_iter() .map(|k| { let mut v = opkey(k); v.push(Value::Array( m[k].iter().map(|d| Value::Bytes(d.0.to_vec())).collect(), )); Value::Array(v) }) .collect(), ) }; let mut nodes: Vec<&NodeId> = self.kinds.keys().collect(); nodes.sort(); let mut pks: Vec<&OpKey> = self.placements.keys().collect(); pks.sort(); let mut mks: Vec<&OpKey> = self.modes.keys().collect(); mks.sort(); Value::map([ ( 0, Value::Array( nodes .into_iter() .map(|n| { Value::Array(vec![ Value::Bytes(n.0.to_vec()), Value::Uint(self.kinds[n].tag()), self.introduced .get(n) .map(|h| Value::Bytes(h.0.to_vec())) .unwrap_or(Value::Null), ]) }) .collect(), ), ), ( 1, Value::Array( pks.into_iter() .map(|k| { let p = &self.placements[k]; let mut v = opkey(k); v.extend([ Value::Bytes(p.node.0.to_vec()), Value::Bytes(p.parent.0.to_vec()), Value::Text(p.name.clone()), ]); Value::Array(v) }) .collect(), ), ), (2, deleters(&self.deleted_placements)), ( 3, Value::Array( mks.into_iter() .map(|k| { let (n, m) = &self.modes[k]; let mut v = opkey(k); v.extend([Value::Bytes(n.0.to_vec()), Value::Uint(m.bits())]); Value::Array(v) }) .collect(), ), ), (4, deleters(&self.deleted_modes)), ]) } pub fn from_value(v: &crate::cbor::Value) -> Result { use crate::objects::{array, bytes, field, map, oref, text, u32_, uint, ObjectError}; let m = map(v, "tree graph")?; let node = |v: &crate::cbor::Value| { NodeId::from_slice(bytes(v, "node")?) .ok_or_else(|| ObjectError::Schema("node: expected 16 bytes".into())) }; let opkey = |e: &[crate::cbor::Value]| -> Result { Ok((oref(&e[0], "op change")?, u32_(&e[1], "op index")?)) }; let mut g = TreeGraph::default(); for n in array(field(m, 0, "nodes")?, "nodes")? { let e = array(n, "node entry")?; if e.len() != 3 { return Err(ObjectError::Schema("node entry: expected 3 fields".into())); } let id = node(&e[0])?; g.kinds .insert(id, NodeKind::from_tag(uint(&e[1], "kind")?)?); if !e[2].is_null() { g.introduced.insert(id, oref(&e[2], "introducer")?); } } for p in array(field(m, 1, "placements")?, "placements")? { let e = array(p, "placement")?; if e.len() != 5 { return Err(ObjectError::Schema("placement: expected 5 fields".into())); } let k = opkey(e)?; let pl = Placement { node: node(&e[2])?, parent: node(&e[3])?, name: text(&e[4], "name")?.to_owned(), }; g.placements_of.entry(pl.node).or_default().insert(k); g.placements.insert(k, pl); } let deleters = |v: &crate::cbor::Value, what: &str| -> Result>, ObjectError> { let mut out = HashMap::new(); for d in array(v, what)? { let e = array(d, what)?; if e.len() != 3 { return Err(ObjectError::Schema(format!("{what}: expected 3 fields"))); } out.insert( opkey(e)?, array(&e[2], "deleters")? .iter() .map(|x| oref(x, "deleter")) .collect::, _>>()?, ); } Ok(out) }; g.deleted_placements = deleters(field(m, 2, "deleted placements")?, "deleted placement")?; for md in array(field(m, 3, "modes")?, "modes")? { let e = array(md, "mode")?; if e.len() != 4 { return Err(ObjectError::Schema("mode: expected 4 fields".into())); } let k = opkey(e)?; let n = node(&e[2])?; g.modes .insert(k, (n, Mode::from_bits(uint(&e[3], "mode bits")?)?)); g.modes_of.entry(n).or_default().insert(k); } g.deleted_modes = deleters(field(m, 4, "deleted modes")?, "deleted mode")?; Ok(g) } } pub fn op_key(r: OpRef) -> OpKey { (r.change, r.index) } #[cfg(test)] mod tests { use super::*; use crate::hash::ObjectRef; fn h(n: u8) -> ChangeHash { ObjectRef([n; 32]) } fn nid(n: u8) -> NodeId { NodeId([n; 16]) } fn add(g: &mut TreeGraph, c: u8, i: u32, node: u8, kind: NodeKind, parent: NodeId, name: &str) { g.apply( h(c), i, &TreeOp::AddNode { node: nid(node), kind, mode: Mode::TEXT, parent, name: name.into(), }, ); } fn paths(v: &TreeView) -> Vec { v.entries.keys().cloned().collect() } #[test] fn nested_paths_and_rename() { let mut g = TreeGraph::new(); add(&mut g, 1, 0, 1, NodeKind::Dir, NodeId::ROOT, "src"); add(&mut g, 1, 1, 2, NodeKind::File, nid(1), "main.rs"); add(&mut g, 1, 2, 3, NodeKind::File, NodeId::ROOT, "README"); let v = g.view(); assert_eq!(paths(&v), ["README", "src", "src/main.rs"]); assert!(!v.is_conflicted()); // Rename src/main.rs → src/lib.rs, move src → lib (subtree follows). g.apply( h(2), 0, &TreeOp::Move { node: nid(2), old: OpRef { change: h(1), index: 1, }, parent: nid(1), name: "lib.rs".into(), }, ); g.apply( h(2), 1, &TreeOp::Move { node: nid(1), old: OpRef { change: h(1), index: 0, }, parent: NodeId::ROOT, name: "lib".into(), }, ); let v = g.view(); assert_eq!(paths(&v), ["README", "lib", "lib/lib.rs"]); assert_eq!(v.entries["lib/lib.rs"].placement, (h(2), 0)); g.apply( h(3), 0, &TreeOp::RemoveNode { node: nid(3), placement: OpRef { change: h(1), index: 2, }, }, ); assert_eq!(paths(&g.view()), ["lib", "lib/lib.rs"]); } #[test] fn name_conflict_gets_suffixes() { let mut g = TreeGraph::new(); add(&mut g, 1, 0, 1, NodeKind::File, NodeId::ROOT, "a.txt"); add(&mut g, 2, 0, 2, NodeKind::File, NodeId::ROOT, "a.txt"); let v = g.view(); assert_eq!(paths(&v), ["a.txt.oxy-conflict-1", "a.txt.oxy-conflict-2"]); assert_eq!( v.conflicts, vec![TreeConflict::Name { parent: NodeId::ROOT, name: "a.txt".into(), nodes: vec![nid(1), nid(2)] }] ); } #[test] fn concurrent_renames_are_multi_placement() { let mut g = TreeGraph::new(); add(&mut g, 1, 0, 1, NodeKind::File, NodeId::ROOT, "a"); g.apply( h(2), 0, &TreeOp::Move { node: nid(1), old: OpRef { change: h(1), index: 0, }, parent: NodeId::ROOT, name: "b".into(), }, ); g.apply( h(3), 0, &TreeOp::Move { node: nid(1), old: OpRef { change: h(1), index: 0, }, parent: NodeId::ROOT, name: "c".into(), }, ); let v = g.view(); assert_eq!(paths(&v), ["b.oxy-conflict-1", "c.oxy-conflict-1"]); assert!( matches!(v.conflicts[0], TreeConflict::MultiPlacement { node, .. } if node == nid(1)) ); } #[test] fn orphan_and_cycle() { let mut g = TreeGraph::new(); add(&mut g, 1, 0, 1, NodeKind::Dir, NodeId::ROOT, "d"); add(&mut g, 1, 1, 2, NodeKind::File, nid(1), "f"); // Concurrently: 2 removes d, 3 adds d/g. g.apply( h(2), 0, &TreeOp::RemoveNode { node: nid(1), placement: OpRef { change: h(1), index: 0, }, }, ); add(&mut g, 3, 0, 3, NodeKind::File, nid(1), "g"); let v = g.view(); assert_eq!(paths(&v), Vec::::new()); assert!(v.conflicts.contains(&TreeConflict::Orphan { node: nid(2), parent: nid(1) })); assert!(v.conflicts.contains(&TreeConflict::Orphan { node: nid(3), parent: nid(1) })); // Cycle: a/b, then concurrently move a under b and b under a's parent... simplest: move a into b. let mut g = TreeGraph::new(); add(&mut g, 1, 0, 1, NodeKind::Dir, NodeId::ROOT, "a"); add(&mut g, 1, 1, 2, NodeKind::Dir, nid(1), "b"); g.apply( h(2), 0, &TreeOp::Move { node: nid(1), old: OpRef { change: h(1), index: 0, }, parent: nid(2), name: "a".into(), }, ); let v = g.view(); assert!( v.conflicts .iter() .any(|c| matches!(c, TreeConflict::Cycle { .. })), "{:?}", v.conflicts ); } #[test] fn serialization_round_trip() { let mut g = TreeGraph::new(); add(&mut g, 1, 0, 1, NodeKind::Dir, NodeId::ROOT, "src"); add(&mut g, 1, 1, 2, NodeKind::File, nid(1), "main.rs"); g.apply( h(2), 0, &TreeOp::Move { node: nid(2), old: OpRef { change: h(1), index: 1, }, parent: nid(1), name: "lib.rs".into(), }, ); g.apply( h(3), 0, &TreeOp::SetMode { node: nid(2), old: OpRef { change: h(1), index: 1, }, mode: Mode { exec: true, opaque: false, }, }, ); let bytes = crate::cbor::encode(&g.to_value()); let back = TreeGraph::from_value(&crate::cbor::decode(&bytes).unwrap()).unwrap(); assert_eq!(back.view(), g.view()); assert_eq!(back.introducer(&nid(2)), Some(h(1))); assert_eq!(crate::cbor::encode(&back.to_value()), bytes); } #[test] fn mode_conflict_and_resolution() { let mut g = TreeGraph::new(); add(&mut g, 1, 0, 1, NodeKind::File, NodeId::ROOT, "run.sh"); g.apply( h(2), 0, &TreeOp::SetMode { node: nid(1), old: OpRef { change: h(1), index: 0, }, mode: Mode { exec: true, opaque: false, }, }, ); g.apply( h(3), 0, &TreeOp::SetMode { node: nid(1), old: OpRef { change: h(1), index: 0, }, mode: Mode { exec: false, opaque: true, }, }, ); let v = g.view(); assert!(matches!(v.conflicts[0], TreeConflict::Mode { .. })); // Materialized with the greatest mode atom. assert_eq!( v.entries["run.sh"].mode, Mode { exec: false, opaque: true } ); // The placement introduced by the same AddNode is unaffected. assert_eq!(v.entries["run.sh"].placement, (h(1), 0)); // Resolution: retire both live mode atoms; identical live values do not conflict. g.apply( h(4), 0, &TreeOp::SetMode { node: nid(1), old: OpRef { change: h(2), index: 0, }, mode: Mode { exec: true, opaque: false, }, }, ); g.apply( h(4), 1, &TreeOp::SetMode { node: nid(1), old: OpRef { change: h(3), index: 0, }, mode: Mode { exec: true, opaque: false, }, }, ); let v = g.view(); assert!(!v.is_conflicted(), "{:?}", v.conflicts); assert_eq!( v.entries["run.sh"].mode, Mode { exec: true, opaque: false } ); } }