Warp/server/internal/roomserver/relations.go

package roomserver

import (
	"context"
	"encoding/json"
	"fmt"
	"slices"

	"github.com/dariuszjaniak/warp/server/internal/matrix/event"
	"github.com/dariuszjaniak/warp/server/internal/matrix/roomversion"
	"github.com/dariuszjaniak/warp/server/internal/storage"
)

// Relations: threads, edits, reactions and references.
//
// An event relates to another through content.m.relates_to, and nothing
// about that is checked by the authorisation rules — a relation is a
// client convention the server indexes, not part of the room's DAG. The
// server's part is three things: recording which event points at which
// (relationOf, written with the event), listing an event's children
// (/relations and /threads), and summarising them into the parent's
// unsigned.m.relations wherever the parent is served (the bundled
// aggregations), so that a client rendering a timeline knows a message
// has a thread under it or has been edited without asking about every
// event.

// relationOf reads content.m.relates_to. Both fields have to be strings,
// and an m.in_reply_to alone is not a relation in this sense: a reply
// points at an event without being one of its children.
func relationOf(content json.RawMessage) (eventID, relType string) {
	var probe struct {
		RelatesTo *struct {
			EventID any `json:"event_id"`
			RelType any `json:"rel_type"`
		} `json:"m.relates_to"`
	}
	if err := json.Unmarshal(content, &probe); err != nil || probe.RelatesTo == nil {
		return "", ""
	}
	id, ok1 := probe.RelatesTo.EventID.(string)
	rel, ok2 := probe.RelatesTo.RelType.(string)
	if !ok1 || !ok2 || id == "" || rel == "" {
		return "", ""
	}
	return id, rel
}

// ThreadOf is the thread an event is in: the root its m.relates_to names
// when the relation is m.thread, or "" for an event on the main timeline.
// The push server files each notification under it.
func ThreadOf(e *event.Event) string {
	if root, rel := relationOf(e.Content()); rel == storage.RelThread {
		return root
	}
	return ""
}

// searchableText is what the search index holds for an event: the body of
// a message, the name of a room or its topic, as the spec's three search
// keys. An encrypted message has no body the server can read and is never
// found, which is what end-to-end encryption means.
func searchableText(e *event.Event) (key, text string) {
	var field, k string
	switch e.Type() {
	case TypeMessage:
		field, k = "body", "content.body"
	case TypeName:
		field, k = "name", "content.name"
	case TypeTopic:
		field, k = "topic", "content.topic"
	default:
		return "", ""
	}
	var content map[string]any
	if err := json.Unmarshal(e.Content(), &content); err != nil {
		return "", ""
	}
	value, ok := content[field].(string)
	if !ok || value == "" {
		return "", ""
	}
	return k, value
}

// maxRecursionDepth is how far a recursive /relations walks: the spec's
// own suggestion, which covers a reaction to an edit in a thread.
const maxRecursionDepth = 3

// maxReferenceChunk caps the event IDs in an m.reference summary.
const maxReferenceChunk = 50

// RelationsRequest is one page of an event's children.
type RelationsRequest struct {
	RoomID  string
	EventID string

	// User is who is asking. The parent has to be visible to them, and
	// so does every child.
	User string

	// RelType and EventType narrow the children, when set.
	RelType   string
	EventType string

	// Low, High and Direction are a TimelineRequest's.
	Low       int64
	High      int64
	Direction storage.TimelineDirection

	Limit int

	// Recurse includes the children's children, down to
	// maxRecursionDepth levels.
	Recurse bool
}

// RelationsPage is one page of an event's children.
type RelationsPage struct {
	Timeline

	// Depth is how many levels the read went down: 1 without recursion.
	Depth int
}

// Relations reads one page of an event's children.
func (r *API) Relations(ctx context.Context, req RelationsRequest) (RelationsPage, error) {
	_, rules, err := r.roomAndRules(ctx, req.RoomID)
	if err != nil {
		return RelationsPage{}, err
	}
	visible, err := r.requireReadable(ctx, req.RoomID, req.User)
	if err != nil {
		return RelationsPage{}, err
	}
	parent, err := r.store.Event(ctx, req.RoomID, req.EventID)
	if err != nil {
		return RelationsPage{}, err
	}
	if parent.Outlier || !visible.Contains(parent.StreamOrdering) {
		return RelationsPage{}, fmt.Errorf("event %q: %w", req.EventID, ErrNoSuchEvent)
	}

	// The parents to read the children of: the event, and with recursion
	// every descendant down to the depth limit. The filters apply at every
	// level, as the spec has it, so a walk for m.thread children only
	// follows threads.
	//
	// The children of one level are the next level, so the walk adds a
	// level to the parents only while it is above the limit: the events at
	// the limit are read as children and are nobody's parents. depth is
	// the deepest level that had anything in it.
	parents := []string{req.EventID}
	depth := 1
	if req.Recurse {
		depth = 0
		level := parents
		seen := map[string]bool{req.EventID: true}
		for depth < maxRecursionDepth {
			children, err := r.store.ChildEventIDs(ctx, req.RoomID, level, req.RelType, req.EventType)
			if err != nil {
				return RelationsPage{}, err
			}
			// A relation loop cannot be built — an event ID is a hash of
			// content naming its parent — but a repeat across levels can.
			children = slices.DeleteFunc(children, func(id string) bool {
				if seen[id] {
					return true
				}
				seen[id] = true
				return false
			})
			if len(children) == 0 {
				break
			}
			depth++
			if depth < maxRecursionDepth {
				parents = append(parents, children...)
			}
			level = children
		}
		depth = max(depth, 1)
	}

	empty := Timeline{Start: req.Low, End: req.Low}
	if req.Direction == storage.DirectionBackward && req.High != 0 {
		empty = Timeline{Start: req.High, End: req.High}
	}
	low, high, any := visible.Clamp(req.Low, req.High)
	if !any || req.Limit <= 0 {
		return RelationsPage{Timeline: empty, Depth: depth}, nil
	}

	stored, err := r.store.Relations(ctx, storage.RelationsRequest{
		RoomID:    req.RoomID,
		ParentIDs: parents,
		RelType:   req.RelType,
		EventType: req.EventType,
		Low:       low,
		High:      high,
		Direction: req.Direction,
		Limit:     req.Limit + 1,
	})
	if err != nil {
		return RelationsPage{}, err
	}

	out := RelationsPage{Timeline: empty, Depth: depth}
	if len(stored) > req.Limit {
		out.Limited = true
		stored = stored[:req.Limit]
	}
	if len(stored) == 0 {
		return out, nil
	}
	if req.Direction == storage.DirectionBackward {
		slices.Reverse(stored)
	}
	if out.Events, err = r.clientEvents(stored, rules, req.RoomID); err != nil {
		return RelationsPage{}, err
	}
	if err := r.bundleAggregations(ctx, req.RoomID, rules, req.User, visible, stored, out.Events); err != nil {
		return RelationsPage{}, err
	}
	out.Start = storage.PositionBefore(stored[0].StreamOrdering)
	out.End = stored[len(stored)-1].StreamOrdering
	return out, nil
}

// ThreadsRequest is one page of a room's threads.
type ThreadsRequest struct {
	RoomID string
	User   string

	// Participated keeps only the threads the user started or replied in.
	Participated bool

	// Before is where the previous page ended, zero for the first.
	Before int64

	Limit int
}

// ThreadsPage is one page of a room's threads: their root events, newest
// activity first, each with its m.thread summary.
type ThreadsPage struct {
	Roots []ClientEvent

	// Next is the Before of the next page, zero at the end.
	Next int64
}

// Threads reads one page of a room's threads.
func (r *API) Threads(ctx context.Context, req ThreadsRequest) (ThreadsPage, error) {
	_, rules, err := r.roomAndRules(ctx, req.RoomID)
	if err != nil {
		return ThreadsPage{}, err
	}
	visible, err := r.requireReadable(ctx, req.RoomID, req.User)
	if err != nil {
		return ThreadsPage{}, err
	}
	if req.Limit <= 0 {
		return ThreadsPage{}, nil
	}

	participant := ""
	if req.Participated {
		participant = req.User
	}
	threads, err := r.store.Threads(ctx, storage.ThreadsRequest{
		RoomID:         req.RoomID,
		ParticipatedBy: participant,
		Before:         req.Before,
		High:           visible.To,
		Limit:          req.Limit + 1,
	})
	if err != nil {
		return ThreadsPage{}, err
	}

	var out ThreadsPage
	if len(threads) > req.Limit {
		threads = threads[:req.Limit]
		out.Next = threads[len(threads)-1].LatestOrdering
	}

	ids := make([]string, 0, len(threads))
	for _, t := range threads {
		ids = append(ids, t.RootID)
	}
	stored, err := r.store.EventsByID(ctx, req.RoomID, ids)
	if err != nil {
		return ThreadsPage{}, err
	}
	byID := make(map[string]storage.StoredEvent, len(stored))
	for _, s := range stored {
		byID[s.EventID] = s
	}

	// In the order the store ranked them, leaving out any root the user may
	// not see: a thread started before a `joined` user's join is not
	// theirs to read, even if it is still going.
	roots := make([]storage.StoredEvent, 0, len(threads))
	for _, t := range threads {
		s, ok := byID[t.RootID]
		if !ok || s.Outlier || !visible.Contains(s.StreamOrdering) {
			continue
		}
		roots = append(roots, s)
	}
	if out.Roots, err = r.clientEvents(roots, rules, req.RoomID); err != nil {
		return ThreadsPage{}, err
	}
	if err := r.bundleAggregations(ctx, req.RoomID, rules, req.User, visible, roots, out.Roots); err != nil {
		return ThreadsPage{}, err
	}
	return out, nil
}

// threadSummary is unsigned.m.relations.m.thread.
type threadSummary struct {
	LatestEvent             ClientEvent `json:"latest_event"`
	Count                   int         `json:"count"`
	CurrentUserParticipated bool        `json:"current_user_participated"`
}

// referenceSummary is unsigned.m.relations.m.reference.
type referenceSummary struct {
	Chunk []referenceEntry `json:"chunk"`
}

type referenceEntry struct {
	EventID string `json:"event_id"`
}

// bundleAggregations adds unsigned.m.relations to the events of a read:
// the thread under each of them, its latest edit, and what references it.
//
// stored and events are the same events, in the same order — the rows and
// their client shape — because the summaries need what only the rows have
// (whether an event was redacted) and are written into the client shape.
// Annotations are not summarised: the spec dropped their aggregation, and a
// client counts reactions from the events themselves.
func (r *API) bundleAggregations(ctx context.Context, roomID string, rules roomversion.Rules, userID string, visible VisibleRange, stored []storage.StoredEvent, events []ClientEvent) error {
	if len(stored) == 0 {
		return nil
	}
	ids := make([]string, 0, len(stored))
	parents := make(map[string]storage.StoredEvent, len(stored))
	for _, s := range stored {
		ids = append(ids, s.EventID)
		parents[s.EventID] = s
	}

	aggregates, err := r.store.RelationAggregates(ctx, roomID, ids, visible.To)
	if err != nil {
		return err
	}
	if len(aggregates) == 0 {
		return nil
	}

	type thread struct {
		count        int
		latest       int64
		latestID     string
		participated bool
	}
	threads := map[string]*thread{}
	edits := map[string]storage.RelationAggregate{}
	referenced := map[string]bool{}
	for _, a := range aggregates {
		parent := parents[a.ParentID]
		switch a.RelType {
		case storage.RelThread:
			t := threads[a.ParentID]
			if t == nil {
				t = &thread{participated: parent.Sender == userID}
				threads[a.ParentID] = t
			}
			t.count += a.Count
			if a.LatestOrdering > t.latest {
				t.latest, t.latestID = a.LatestOrdering, a.LatestEventID
			}
			if a.Sender == userID {
				t.participated = true
			}
		case storage.RelReplace:
			// Only the original's sender may edit it, a state event cannot
			// be edited, and a redacted event's edits are moot: what they
			// replaced is gone.
			if a.Sender != parent.Sender || parent.IsState() || parent.RedactedBy != "" {
				continue
			}
			edits[a.ParentID] = a
		case storage.RelReference:
			referenced[a.ParentID] = true
		}
	}

	// The latest thread replies and edits, read in one go.
	latestIDs := make([]string, 0, len(threads)+len(edits))
	for _, t := range threads {
		latestIDs = append(latestIDs, t.latestID)
	}
	for _, a := range edits {
		latestIDs = append(latestIDs, a.LatestEventID)
	}
	latest, err := r.clientEventsByID(ctx, roomID, rules, latestIDs)
	if err != nil {
		return err
	}
	latestByID := make(map[string]ClientEvent, len(latest))
	for _, e := range latest {
		latestByID[e.EventID] = e
	}

	var references map[string][]string
	if len(referenced) > 0 {
		refParents := make([]string, 0, len(referenced))
		for id := range referenced {
			refParents = append(refParents, id)
		}
		if references, err = r.store.RelationEventIDs(ctx, roomID, refParents,
			storage.RelReference, visible.To, maxReferenceChunk); err != nil {
			return err
		}
	}

	for i := range events {
		id := events[i].EventID
		relations := map[string]any{}
		if t, ok := threads[id]; ok {
			if e, ok := latestByID[t.latestID]; ok {
				relations[storage.RelThread] = threadSummary{
					LatestEvent:             e,
					Count:                   t.count,
					CurrentUserParticipated: t.participated,
				}
			}
		}
		if a, ok := edits[id]; ok {
			// An edit has to keep the type of what it edits; one that does
			// not is not a replacement the client should render.
			if e, ok := latestByID[a.LatestEventID]; ok && e.Type == events[i].Type {
				relations[storage.RelReplace] = e
			}
		}
		if refs := references[id]; len(refs) > 0 {
			summary := referenceSummary{Chunk: make([]referenceEntry, 0, len(refs))}
			for _, ref := range refs {
				summary.Chunk = append(summary.Chunk, referenceEntry{EventID: ref})
			}
			relations[storage.RelReference] = summary
		}
		if len(relations) == 0 {
			continue
		}
		events[i].Unsigned = r.withUnsigned(events[i].Unsigned, "m.relations", relations)
	}
	return nil
}

// withUnsigned sets one key of an event's unsigned block.
func (r *API) withUnsigned(unsigned json.RawMessage, key string, value any) json.RawMessage {
	block := map[string]json.RawMessage{}
	if len(unsigned) > 0 {
		if err := json.Unmarshal(unsigned, &block); err != nil {
			r.log.Warn("an unsigned block this server built is unreadable", "error", err)
			block = map[string]json.RawMessage{}
		}
	}
	block[key] = mustMarshal(value)
	return mustMarshal(block)
}