0
0
Fork 0
mirror of https://github.com/matrix-org/dendrite synced 2024-07-07 05:28:38 +02:00
dendrite/syncapi/routing/messages.go
Till eb29a31550
Optimize /sync and history visibility (#2961)
Should fix the following issues or make a lot less worse when using
Postgres:

The main issue behind #2911: The client gives up after a certain time,
causing a cascade of context errors, because the response couldn't be
built up fast enough. This mostly happens on accounts with many rooms,
due to the inefficient way we're getting recent events and current state

For #2777: The queries for getting the membership events for history
visibility were being executed for each room (I think 185?), resulting
in a whooping 2k queries for membership events. (Getting the
statesnapshot -> block nids -> actual wanted membership event)

Both should now be better by:
- Using a LATERAL join to get all recent events for all joined rooms in
one go (TODO: maybe do the same for room summary and current state etc)
- If we're lazy loading on initial syncs, we're now not getting the
whole current state, just to drop the majority of it because we're lazy
loading members - we add a filter to exclude membership events on the
first call to `CurrentState`.
- Using an optimized query to get the membership events needed to
calculate history visibility

---------

Co-authored-by: kegsay <kegan@matrix.org>
2023-02-07 14:31:23 +01:00

556 lines
19 KiB
Go

// Copyright 2018 New Vector Ltd
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
package routing
import (
"context"
"fmt"
"math"
"net/http"
"sort"
"time"
"github.com/matrix-org/gomatrixserverlib"
"github.com/matrix-org/util"
"github.com/sirupsen/logrus"
"github.com/matrix-org/dendrite/clientapi/jsonerror"
"github.com/matrix-org/dendrite/internal/caching"
"github.com/matrix-org/dendrite/internal/sqlutil"
"github.com/matrix-org/dendrite/roomserver/api"
"github.com/matrix-org/dendrite/setup/config"
"github.com/matrix-org/dendrite/syncapi/internal"
"github.com/matrix-org/dendrite/syncapi/storage"
"github.com/matrix-org/dendrite/syncapi/sync"
"github.com/matrix-org/dendrite/syncapi/types"
userapi "github.com/matrix-org/dendrite/userapi/api"
)
type messagesReq struct {
ctx context.Context
db storage.Database
snapshot storage.DatabaseTransaction
rsAPI api.SyncRoomserverAPI
cfg *config.SyncAPI
roomID string
from *types.TopologyToken
to *types.TopologyToken
device *userapi.Device
wasToProvided bool
backwardOrdering bool
filter *gomatrixserverlib.RoomEventFilter
}
type messagesResp struct {
Start string `json:"start"`
StartStream string `json:"start_stream,omitempty"` // NOTSPEC: used by Cerulean, so clients can hit /messages then immediately /sync with a latest sync token
End string `json:"end,omitempty"`
Chunk []gomatrixserverlib.ClientEvent `json:"chunk"`
State []gomatrixserverlib.ClientEvent `json:"state,omitempty"`
}
// OnIncomingMessagesRequest implements the /messages endpoint from the
// client-server API.
// See: https://matrix.org/docs/spec/client_server/latest.html#get-matrix-client-r0-rooms-roomid-messages
// nolint:gocyclo
func OnIncomingMessagesRequest(
req *http.Request, db storage.Database, roomID string, device *userapi.Device,
rsAPI api.SyncRoomserverAPI,
cfg *config.SyncAPI,
srp *sync.RequestPool,
lazyLoadCache caching.LazyLoadCache,
) util.JSONResponse {
var err error
// NewDatabaseTransaction is used here instead of NewDatabaseSnapshot as we
// expect to be able to write to the database in response to a /messages
// request that requires backfilling from the roomserver or federation.
snapshot, err := db.NewDatabaseTransaction(req.Context())
if err != nil {
return jsonerror.InternalServerError()
}
var succeeded bool
defer sqlutil.EndTransactionWithCheck(snapshot, &succeeded, &err)
// check if the user has already forgotten about this room
membershipResp, err := getMembershipForUser(req.Context(), roomID, device.UserID, rsAPI)
if err != nil {
return jsonerror.InternalServerError()
}
if !membershipResp.RoomExists {
return util.JSONResponse{
Code: http.StatusForbidden,
JSON: jsonerror.Forbidden("room does not exist"),
}
}
if membershipResp.IsRoomForgotten {
return util.JSONResponse{
Code: http.StatusForbidden,
JSON: jsonerror.Forbidden("user already forgot about this room"),
}
}
filter, err := parseRoomEventFilter(req)
if err != nil {
return util.JSONResponse{
Code: http.StatusBadRequest,
JSON: jsonerror.InvalidArgumentValue("unable to parse filter"),
}
}
// Extract parameters from the request's URL.
// Pagination tokens.
var fromStream *types.StreamingToken
fromQuery := req.URL.Query().Get("from")
toQuery := req.URL.Query().Get("to")
emptyFromSupplied := fromQuery == ""
if emptyFromSupplied {
// NOTSPEC: We will pretend they used the latest sync token if no ?from= was provided.
// We do this to allow clients to get messages without having to call `/sync` e.g Cerulean
currPos := srp.Notifier.CurrentPosition()
fromQuery = currPos.String()
}
// Direction to return events from.
dir := req.URL.Query().Get("dir")
if dir != "b" && dir != "f" {
return util.JSONResponse{
Code: http.StatusBadRequest,
JSON: jsonerror.MissingArgument("Bad or missing dir query parameter (should be either 'b' or 'f')"),
}
}
// A boolean is easier to handle in this case, especially since dir is sure
// to have one of the two accepted values (so dir == "f" <=> !backwardOrdering).
backwardOrdering := (dir == "b")
from, err := types.NewTopologyTokenFromString(fromQuery)
if err != nil {
var streamToken types.StreamingToken
if streamToken, err = types.NewStreamTokenFromString(fromQuery); err != nil {
return util.JSONResponse{
Code: http.StatusBadRequest,
JSON: jsonerror.InvalidArgumentValue("Invalid from parameter: " + err.Error()),
}
} else {
fromStream = &streamToken
from, err = snapshot.StreamToTopologicalPosition(req.Context(), roomID, streamToken.PDUPosition, backwardOrdering)
if err != nil {
logrus.WithError(err).Errorf("Failed to get topological position for streaming token %v", streamToken)
return jsonerror.InternalServerError()
}
}
}
// Pagination tokens. To is optional, and its default value depends on the
// direction ("b" or "f").
var to types.TopologyToken
wasToProvided := true
if len(toQuery) > 0 {
to, err = types.NewTopologyTokenFromString(toQuery)
if err != nil {
var streamToken types.StreamingToken
if streamToken, err = types.NewStreamTokenFromString(toQuery); err != nil {
return util.JSONResponse{
Code: http.StatusBadRequest,
JSON: jsonerror.InvalidArgumentValue("Invalid to parameter: " + err.Error()),
}
} else {
to, err = snapshot.StreamToTopologicalPosition(req.Context(), roomID, streamToken.PDUPosition, !backwardOrdering)
if err != nil {
logrus.WithError(err).Errorf("Failed to get topological position for streaming token %v", streamToken)
return jsonerror.InternalServerError()
}
}
}
} else {
// If "to" isn't provided, it defaults to either the earliest stream
// position (if we're going backward) or to the latest one (if we're
// going forward).
to = types.TopologyToken{Depth: math.MaxInt64, PDUPosition: math.MaxInt64}
if backwardOrdering {
// go 1 earlier than the first event so we correctly fetch the earliest event
// this is because Database.GetEventsInTopologicalRange is exclusive of the lower-bound.
to = types.TopologyToken{}
}
wasToProvided = false
}
// TODO: Implement filtering (#587)
// Check the room ID's format.
if _, _, err = gomatrixserverlib.SplitID('!', roomID); err != nil {
return util.JSONResponse{
Code: http.StatusBadRequest,
JSON: jsonerror.MissingArgument("Bad room ID: " + err.Error()),
}
}
// If the user already left the room, grep events from before that
if membershipResp.Membership == gomatrixserverlib.Leave {
var token types.TopologyToken
token, err = snapshot.EventPositionInTopology(req.Context(), membershipResp.EventID)
if err != nil {
return util.JSONResponse{
Code: http.StatusInternalServerError,
}
}
if backwardOrdering {
from = token
}
}
mReq := messagesReq{
ctx: req.Context(),
db: db,
snapshot: snapshot,
rsAPI: rsAPI,
cfg: cfg,
roomID: roomID,
from: &from,
to: &to,
wasToProvided: wasToProvided,
filter: filter,
backwardOrdering: backwardOrdering,
device: device,
}
clientEvents, start, end, err := mReq.retrieveEvents()
if err != nil {
util.GetLogger(req.Context()).WithError(err).Error("mreq.retrieveEvents failed")
return jsonerror.InternalServerError()
}
util.GetLogger(req.Context()).WithFields(logrus.Fields{
"from": from.String(),
"to": to.String(),
"limit": filter.Limit,
"backwards": backwardOrdering,
"return_start": start.String(),
"return_end": end.String(),
}).Info("Responding")
res := messagesResp{
Chunk: clientEvents,
Start: start.String(),
End: end.String(),
}
if filter.LazyLoadMembers {
membershipEvents, err := applyLazyLoadMembers(req.Context(), device, snapshot, roomID, clientEvents, lazyLoadCache)
if err != nil {
util.GetLogger(req.Context()).WithError(err).Error("failed to apply lazy loading")
return jsonerror.InternalServerError()
}
res.State = append(res.State, gomatrixserverlib.HeaderedToClientEvents(membershipEvents, gomatrixserverlib.FormatAll)...)
}
// If we didn't return any events, set the end to an empty string, so it will be omitted
// in the response JSON.
if len(res.Chunk) == 0 {
res.End = ""
}
if fromStream != nil {
res.StartStream = fromStream.String()
}
// Respond with the events.
succeeded = true
return util.JSONResponse{
Code: http.StatusOK,
JSON: res,
}
}
func getMembershipForUser(ctx context.Context, roomID, userID string, rsAPI api.SyncRoomserverAPI) (resp api.QueryMembershipForUserResponse, err error) {
req := api.QueryMembershipForUserRequest{
RoomID: roomID,
UserID: userID,
}
if err := rsAPI.QueryMembershipForUser(ctx, &req, &resp); err != nil {
return api.QueryMembershipForUserResponse{}, err
}
return resp, nil
}
// retrieveEvents retrieves events from the local database for a request on
// /messages. If there's not enough events to retrieve, it asks another
// homeserver in the room for older events.
// Returns an error if there was an issue talking to the database or with the
// remote homeserver.
func (r *messagesReq) retrieveEvents() (
clientEvents []gomatrixserverlib.ClientEvent, start,
end types.TopologyToken, err error,
) {
// Retrieve the events from the local database.
streamEvents, err := r.snapshot.GetEventsInTopologicalRange(r.ctx, r.from, r.to, r.roomID, r.filter, r.backwardOrdering)
if err != nil {
err = fmt.Errorf("GetEventsInRange: %w", err)
return
}
var events []*gomatrixserverlib.HeaderedEvent
util.GetLogger(r.ctx).WithFields(logrus.Fields{
"start": r.from,
"end": r.to,
"backwards": r.backwardOrdering,
}).Infof("Fetched %d events locally", len(streamEvents))
// There can be two reasons for streamEvents to be empty: either we've
// reached the oldest event in the room (or the most recent one, depending
// on the ordering), or we've reached a backward extremity.
if len(streamEvents) == 0 {
if events, err = r.handleEmptyEventsSlice(); err != nil {
return
}
} else {
if events, err = r.handleNonEmptyEventsSlice(streamEvents); err != nil {
return
}
}
// If we didn't get any event, we don't need to proceed any further.
if len(events) == 0 {
return []gomatrixserverlib.ClientEvent{}, *r.from, *r.to, nil
}
// Get the position of the first and the last event in the room's topology.
// This position is currently determined by the event's depth, so we could
// also use it instead of retrieving from the database. However, if we ever
// change the way topological positions are defined (as depth isn't the most
// reliable way to define it), it would be easier and less troublesome to
// only have to change it in one place, i.e. the database.
start, end, err = r.getStartEnd(events)
if err != nil {
return []gomatrixserverlib.ClientEvent{}, *r.from, *r.to, err
}
// Sort the events to ensure we send them in the right order.
if r.backwardOrdering {
// This reverses the array from old->new to new->old
reversed := func(in []*gomatrixserverlib.HeaderedEvent) []*gomatrixserverlib.HeaderedEvent {
out := make([]*gomatrixserverlib.HeaderedEvent, len(in))
for i := 0; i < len(in); i++ {
out[i] = in[len(in)-i-1]
}
return out
}
events = reversed(events)
}
if len(events) == 0 {
return []gomatrixserverlib.ClientEvent{}, *r.from, *r.to, nil
}
// Apply room history visibility filter
startTime := time.Now()
filteredEvents, err := internal.ApplyHistoryVisibilityFilter(r.ctx, r.snapshot, r.rsAPI, events, nil, r.device.UserID, "messages")
logrus.WithFields(logrus.Fields{
"duration": time.Since(startTime),
"room_id": r.roomID,
"events_before": len(events),
"events_after": len(filteredEvents),
}).Debug("applied history visibility (messages)")
return gomatrixserverlib.HeaderedToClientEvents(filteredEvents, gomatrixserverlib.FormatAll), start, end, err
}
func (r *messagesReq) getStartEnd(events []*gomatrixserverlib.HeaderedEvent) (start, end types.TopologyToken, err error) {
if r.backwardOrdering {
start = *r.from
if events[len(events)-1].Type() == gomatrixserverlib.MRoomCreate {
// NOTSPEC: We've hit the beginning of the room so there's really nowhere
// else to go. This seems to fix Element iOS from looping on /messages endlessly.
end = types.TopologyToken{}
} else {
end, err = r.snapshot.EventPositionInTopology(
r.ctx, events[0].EventID(),
)
// A stream/topological position is a cursor located between two events.
// While they are identified in the code by the event on their right (if
// we consider a left to right chronological order), tokens need to refer
// to them by the event on their left, therefore we need to decrement the
// end position we send in the response if we're going backward.
end.Decrement()
}
} else {
start = *r.from
end, err = r.snapshot.EventPositionInTopology(
r.ctx, events[len(events)-1].EventID(),
)
}
if err != nil {
err = fmt.Errorf("EventPositionInTopology: for end event %s: %w", events[len(events)-1].EventID(), err)
return
}
return
}
// handleEmptyEventsSlice handles the case where the initial request to the
// database returned an empty slice of events. It does so by checking whether
// the set is empty because we've reached a backward extremity, and if that is
// the case, by retrieving as much events as requested by backfilling from
// another homeserver.
// Returns an error if there was an issue talking with the database or
// backfilling.
func (r *messagesReq) handleEmptyEventsSlice() (
events []*gomatrixserverlib.HeaderedEvent, err error,
) {
backwardExtremities, err := r.snapshot.BackwardExtremitiesForRoom(r.ctx, r.roomID)
// Check if we have backward extremities for this room.
if len(backwardExtremities) > 0 {
// If so, retrieve as much events as needed through backfilling.
events, err = r.backfill(r.roomID, backwardExtremities, r.filter.Limit)
if err != nil {
return
}
} else {
// If not, it means the slice was empty because we reached the room's
// creation, so return an empty slice.
events = []*gomatrixserverlib.HeaderedEvent{}
}
return
}
// handleNonEmptyEventsSlice handles the case where the initial request to the
// database returned a non-empty slice of events. It does so by checking whether
// events are missing from the expected result, and retrieve missing events
// through backfilling if needed.
// Returns an error if there was an issue while backfilling.
func (r *messagesReq) handleNonEmptyEventsSlice(streamEvents []types.StreamEvent) (
events []*gomatrixserverlib.HeaderedEvent, err error,
) {
// Check if we have enough events.
isSetLargeEnough := len(streamEvents) >= r.filter.Limit
if !isSetLargeEnough {
// it might be fine we don't have up to 'limit' events, let's find out
if r.backwardOrdering {
if r.wasToProvided {
// The condition in the SQL query is a strict "greater than" so
// we need to check against to-1.
streamPos := types.StreamPosition(streamEvents[len(streamEvents)-1].StreamPosition)
isSetLargeEnough = (r.to.PDUPosition-1 == streamPos)
}
} else {
streamPos := types.StreamPosition(streamEvents[0].StreamPosition)
isSetLargeEnough = (r.from.PDUPosition-1 == streamPos)
}
}
// Check if the slice contains a backward extremity.
backwardExtremities, err := r.snapshot.BackwardExtremitiesForRoom(r.ctx, r.roomID)
if err != nil {
return
}
// Backfill is needed if we've reached a backward extremity and need more
// events. It's only needed if the direction is backward.
if len(backwardExtremities) > 0 && !isSetLargeEnough && r.backwardOrdering {
var pdus []*gomatrixserverlib.HeaderedEvent
// Only ask the remote server for enough events to reach the limit.
pdus, err = r.backfill(r.roomID, backwardExtremities, r.filter.Limit-len(streamEvents))
if err != nil {
return
}
// Append the PDUs to the list to send back to the client.
events = append(events, pdus...)
}
// Append the events ve previously retrieved locally.
events = append(events, r.snapshot.StreamEventsToEvents(nil, streamEvents)...)
sort.Sort(eventsByDepth(events))
return
}
type eventsByDepth []*gomatrixserverlib.HeaderedEvent
func (e eventsByDepth) Len() int {
return len(e)
}
func (e eventsByDepth) Swap(i, j int) {
e[i], e[j] = e[j], e[i]
}
func (e eventsByDepth) Less(i, j int) bool {
return e[i].Depth() < e[j].Depth()
}
// backfill performs a backfill request over the federation on another
// homeserver in the room.
// See: https://matrix.org/docs/spec/server_server/latest#get-matrix-federation-v1-backfill-roomid
// It also stores the PDUs retrieved from the remote homeserver's response to
// the database.
// Returns with an empty string if the remote homeserver didn't return with any
// event, or if there is no remote homeserver to contact.
// Returns an error if there was an issue with retrieving the list of servers in
// the room or sending the request.
func (r *messagesReq) backfill(roomID string, backwardsExtremities map[string][]string, limit int) ([]*gomatrixserverlib.HeaderedEvent, error) {
var res api.PerformBackfillResponse
err := r.rsAPI.PerformBackfill(context.Background(), &api.PerformBackfillRequest{
RoomID: roomID,
BackwardsExtremities: backwardsExtremities,
Limit: limit,
ServerName: r.cfg.Matrix.ServerName,
VirtualHost: r.device.UserDomain(),
}, &res)
if err != nil {
return nil, fmt.Errorf("PerformBackfill failed: %w", err)
}
util.GetLogger(r.ctx).WithField("new_events", len(res.Events)).Info("Storing new events from backfill")
// TODO: we should only be inserting events into the database from the roomserver's kafka output stream.
// Currently, this can race with live events for the room and cause problems. It's also just a bit unclear
// when you have multiple entry points to write events.
// we have to order these by depth, starting with the lowest because otherwise the topology tokens
// will skip over events that have the same depth but different stream positions due to the query which is:
// - anything less than the depth OR
// - anything with the same depth and a lower stream position.
sort.Sort(eventsByDepth(res.Events))
// Store the events in the database, while marking them as unfit to show
// up in responses to sync requests.
if res.HistoryVisibility == "" {
res.HistoryVisibility = gomatrixserverlib.HistoryVisibilityShared
}
for i := range res.Events {
_, err = r.db.WriteEvent(
context.Background(),
res.Events[i],
[]*gomatrixserverlib.HeaderedEvent{},
[]string{},
[]string{},
nil, true,
res.HistoryVisibility,
)
if err != nil {
return nil, err
}
}
// we may have got more than the requested limit so resize now
events := res.Events
if len(events) > limit {
// last `limit` events
events = events[len(events)-limit:]
}
for _, ev := range events {
ev.Visibility = res.HistoryVisibility
}
return events, nil
}