1package store

3import (

4 "context"

5 "encoding/json"

6 "errors"

7 "fmt"

8 "io"

9 "log/slog"

10 "runtime"

11 "slices"

12 "sort"

13 "time"

15 "github.com/mjl-/bstore"

17 "github.com/mjl-/mox/message"

18 "github.com/mjl-/mox/mlog"

19 "github.com/mjl-/mox/moxio"

20)

22// Assign a new/incoming message to a thread. Message does not yet have an ID. If

23// this isn't a response, ThreadID should remain 0 (unless this is a message with

24// existing message-id) and the caller must set ThreadID to ID.

25// If the account is still busy upgrading messages with threadids in the background, parents

26// may have a threadid 0. That results in this message getting threadid 0, which

27// will handled by the background upgrade process assigning a threadid when it gets

28// to this message.

29func assignThread(log mlog.Log, tx *bstore.Tx, m *Message, part *message.Part) error {

30 if m.MessageID != "" {

31 // Match against existing different message with same Message-ID.

32 q := bstore.QueryTx[Message](tx)

33 q.FilterNonzero(Message{MessageID: m.MessageID})

34 q.FilterEqual("Expunged", false)

35 q.FilterNotEqual("ID", m.ID)

36 q.FilterNotEqual("ThreadID", int64(0))

37 q.SortAsc("ID")

38 q.Limit(1)

39 em, err := q.Get()

40 if err != nil && err != bstore.ErrAbsent {

41 return fmt.Errorf("looking up existing message with message-id: %v", err)

42 } else if err == nil {

43 assignParent(m, em, true)

44 return nil

45 }

46 }

48 h, err := part.Header()

49 if err != nil {

50 log.Errorx("assigning threads: parsing references/in-reply-to headers, not matching by message-id", err, slog.Int64("msgid", m.ID))

51 }

52 messageIDs, err := message.ReferencedIDs(h.Values("References"), h.Values("In-Reply-To"))

53 if err != nil {

54 log.Errorx("assigning threads: parsing references/in-reply-to headers, not matching by message-id", err, slog.Int64("msgid", m.ID))

55 }

56 for i := len(messageIDs) - 1; i >= 0; i-- {

57 messageID := messageIDs[i]

58 if messageID == m.MessageID {

59 continue

60 }

61 tm, _, err := lookupThreadMessage(tx, m.ID, messageID, m.SubjectBase, m.DSN)

62 if err != nil {

63 return fmt.Errorf("looking up thread message for new message: %v", err)

64 } else if tm != nil {

65 assignParent(m, *tm, true)

66 return nil

67 }

68 m.ThreadMissingLink = true

69 }

70 if len(messageIDs) > 0 {

71 return nil

72 }

74 var isResp bool

75 if part != nil && part.Envelope != nil {

76 m.SubjectBase, isResp = message.ThreadSubject(part.Envelope.Subject, false)

77 }

78 if !isResp || m.SubjectBase == "" {

79 return nil

80 }

81 m.ThreadMissingLink = true

82 tm, err := lookupThreadMessageSubject(tx, *m, m.SubjectBase)

83 if err != nil {

84 return fmt.Errorf("looking up thread message by subject: %v", err)

85 } else if tm != nil {

86 assignParent(m, *tm, true)

87 }

88 return nil

89}

91// assignParent assigns threading fields to m that make it a child of parent message pm.

92// updateSeen indicates if m.Seen should be cleared if pm is thread-muted.

93func assignParent(m *Message, pm Message, updateSeen bool) {

94 if pm.ThreadID == 0 {

95 panic(fmt.Sprintf("assigning message id %d/d%q to parent message id %d/%q which has threadid 0", m.ID, m.MessageID, pm.ID, pm.MessageID))

96 }

97 if m.ID == pm.ID {

98 panic(fmt.Sprintf("trying to make message id %d/%q its own parent", m.ID, m.MessageID))

99 }

100 m.ThreadID = pm.ThreadID

101 // Make sure we don't add cycles.

102 if !slices.Contains(pm.ThreadParentIDs, m.ID) {

103 m.ThreadParentIDs = append([]int64{pm.ID}, pm.ThreadParentIDs...)

104 } else if pm.ID != m.ID {

105 m.ThreadParentIDs = []int64{pm.ID}

106 } else {

107 m.ThreadParentIDs = nil

108 }

109 if m.MessageID != "" && m.MessageID == pm.MessageID {

110 m.ThreadMissingLink = true

111 }

112 m.ThreadMuted = pm.ThreadMuted

113 m.ThreadCollapsed = pm.ThreadCollapsed

114 if updateSeen && m.ThreadMuted {

115 m.Seen = true

116 }

117}

118

119// ResetThreading resets the MessageID and SubjectBase fields for all messages in

120// the account. If clearIDs is true, all Thread* fields are also cleared. Changes

121// are made in transactions of batchSize changes. The total number of updated

122// messages is returned.

123//

124// ModSeq is not changed. Calles should bump the uid validity of the mailboxes

125// to propagate the changes to IMAP clients.

126func (a *Account) ResetThreading(ctx context.Context, log mlog.Log, batchSize int, clearIDs bool) (int, error) {

127 // todo: should this send Change events for ThreadMuted and ThreadCollapsed? worth it?

128

129 var lastID int64

130 total := 0

131 for {

132 n := 0

133

134 prepareMessages := func(in, out chan moxio.Work[Message, Message]) {

135 for {

136 w, ok := <-in

137 if !ok {

138 return

139 }

140

141 m := w.In

142

143 // We have the Message-ID and Subject headers in ParsedBuf. We use a partial part

144 // struct so we don't generate so much garbage for the garbage collector to sift

145 // through.

146 var part struct {

147 Envelope *message.Envelope

148 }

149 if err := json.Unmarshal(m.ParsedBuf, &part); err != nil {

150 log.Errorx("unmarshal json parsedbuf for setting message-id, skipping", err, slog.Int64("msgid", m.ID))

151 } else {

152 m.MessageID = ""

153 if part.Envelope != nil && part.Envelope.MessageID != "" {

154 s, _, err := message.MessageIDCanonical(part.Envelope.MessageID)

155 if err != nil {

156 log.Debugx("parsing message-id, skipping", err, slog.Int64("msgid", m.ID), slog.String("messageid", part.Envelope.MessageID))

157 }

158 m.MessageID = s

159 }

160 if part.Envelope != nil {

161 m.SubjectBase, _ = message.ThreadSubject(part.Envelope.Subject, false)

162 }

163 }

164 w.Out = m

165

166 out <- w

167 }

168 }

169

170 err := a.DB.Write(ctx, func(tx *bstore.Tx) error {

171 processMessage := func(in, m Message) error {

172 if clearIDs {

173 m.ThreadID = 0

174 m.ThreadParentIDs = nil

175 m.ThreadMissingLink = false

176 }

177 return tx.Update(&m)

178 }

179

180 // JSON parsing is relatively heavy, we benefit from multiple goroutines.

181 procs := runtime.GOMAXPROCS(0)

182 wq := moxio.NewWorkQueue[Message, Message](procs, 2*procs, prepareMessages, processMessage)

183

184 q := bstore.QueryTx[Message](tx)

185 q.FilterEqual("Expunged", false)

186 q.FilterGreater("ID", lastID)

187 q.SortAsc("ID")

188 err := q.ForEach(func(m Message) error {

189 // We process in batches so we don't block other operations for a long time.

190 if n >= batchSize {

191 return bstore.StopForEach

192 }

193 // Update starting point for next batch.

194 lastID = m.ID

195

196 n++

197 return wq.Add(m)

198 })

199 if err == nil {

200 err = wq.Finish()

201 }

202 wq.Stop()

203 return err

204 })

205 if err != nil {

206 return total, fmt.Errorf("upgrading account to threads storage, step 1/2: %w", err)

207 }

208 total += n

209 if n == 0 {

210 break

211 }

212 }

213 return total, nil

214}

215

216// AssignThreads assigns thread-related fields to messages with ID >=

217// startMessageID. Changes are committed each batchSize changes if txOpt is nil

218// (i.e. during automatic account upgrade, we don't want to block database access

219// for a long time). If txOpt is not nil, all changes are made in that

220// transaction.

221//

222// When resetting thread assignments, the caller must first clear the existing

223// thread fields.

224//

225// Messages are processed in order of ID, so when added to the account, not

226// necessarily by received/date. Most threaded messages can immediately be matched

227// to their parent message. If not, we keep track of the missing message-id and

228// resolve as soon as we encounter it. At the end, we resolve all remaining

229// messages, they start with a cycle.

230//

231// Does not set Seen flag for muted threads.

232//

233// Progress is written to progressWriter, every 100k messages.

234func (a *Account) AssignThreads(ctx context.Context, log mlog.Log, txOpt *bstore.Tx, startMessageID int64, batchSize int, progressWriter io.Writer) error {

235 // We use a more basic version of the thread-matching algorithm describe in:

236 // ../rfc/5256:443

237 // The algorithm assumes you'll select messages, then group into threads. We normally do

238 // thread-calculation when messages are delivered. Here, we assign threads as soon

239 // as we can, but will queue messages that reference known ancestors and resolve as

240 // soon as we process them. We can handle large number of messages, but not very

241 // quickly because we make lots of database queries.

242

243 type childMsg struct {

244 ID int64 // This message will be fetched and updated with the threading fields once the parent is resolved.

245 MessageID string // Of child message. Once child is resolved, its own children can be resolved too.

246 ThreadMissingLink bool

247 }

248 // Messages that have a References/In-Reply-To that we want to set as parent, but

249 // where the parent doesn't have a ThreadID yet are added to pending. The key is

250 // the normalized MessageID of the parent, and the value is a list of messages that

251 // can get resolved once the parent gets its ThreadID. The kids will get the same

252 // ThreadIDs, and they themselves may be parents to kids, and so on.

253 // For duplicate messages (messages with identical Message-ID), the second

254 // Message-ID to be added to pending is added under its own message-id, so it gets

255 // its original as parent.

256 pending := map[string][]childMsg{}

257

258 // Current tx. If not equal to txOpt, we clean it up before we leave.

259 var tx *bstore.Tx

260 defer func() {

261 if tx != nil && tx != txOpt {

262 err := tx.Rollback()

263 log.Check(err, "rolling back transaction")

264 }

265 }()

266

267 // Set thread-related fields for a single message. Caller must save the message,

268 // only if not an error and not added to the pending list.

269 assign := func(m *Message, references, inReplyTo []string, subject string) (pend bool, rerr error) {

270 if m.MessageID != "" {

271 // Attempt to match against existing different message with same Message-ID that

272 // already has a threadid.

273 // If there are multiple messages for a message-id a future call to assign may use

274 // its threadid, or it may end up in pending and we resolve it when we need to.

275 q := bstore.QueryTx[Message](tx)

276 q.FilterNonzero(Message{MessageID: m.MessageID})

277 q.FilterEqual("Expunged", false)

278 q.FilterLess("ID", m.ID)

279 q.SortAsc("ID")

280 q.Limit(1)

281 em, err := q.Get()

282 if err != nil && err != bstore.ErrAbsent {

283 return false, fmt.Errorf("looking up existing message with message-id: %v", err)

284 } else if err == nil {

285 if em.ThreadID == 0 {

286 pending[em.MessageID] = append(pending[em.MessageID], childMsg{m.ID, m.MessageID, true})

287 return true, nil

288 } else {

289 assignParent(m, em, false)

290 return false, nil

291 }

292 }

293 }

294

295 refids, err := message.ReferencedIDs(references, inReplyTo)

296 if err != nil {

297 log.Errorx("assigning threads: parsing references/in-reply-to headers, not matching by message-id", err, slog.Int64("msgid", m.ID))

298 }

299

300 for i := len(refids) - 1; i >= 0; i-- {

301 messageID := refids[i]

302 if messageID == m.MessageID {

303 continue

304 }

305 tm, exists, err := lookupThreadMessage(tx, m.ID, messageID, m.SubjectBase, m.DSN)

306 if err != nil {

307 return false, fmt.Errorf("lookup up thread by message-id %s for message id %d: %w", messageID, m.ID, err)

308 } else if tm != nil {

309 assignParent(m, *tm, false)

310 return false, nil

311 } else if exists {

312 pending[messageID] = append(pending[messageID], childMsg{m.ID, m.MessageID, i < len(refids)-1})

313 return true, nil

314 }

315 }

316

317 var subjectBase string

318 var isResp bool

319 if subject != "" {

320 subjectBase, isResp = message.ThreadSubject(subject, false)

321 }

322 if len(refids) > 0 || !isResp || subjectBase == "" {

323 m.ThreadID = m.ID

324 m.ThreadMissingLink = len(refids) > 0

325 return false, nil

326 }

327

328 // No references to use. If this is a reply/forward (based on subject), we'll match

329 // against base subject, at most 4 weeks back so we don't match against ancient

330 // messages and 1 day ahead so we can match against delayed deliveries.

331 tm, err := lookupThreadMessageSubject(tx, *m, subjectBase)

332 if err != nil {

333 return false, fmt.Errorf("looking up recent messages by base subject %q: %w", subjectBase, err)

334 } else if tm != nil {

335 m.ThreadID = tm.ThreadID

336 m.ThreadParentIDs = []int64{tm.ThreadID} // Always under root message with subject-match.

337 m.ThreadMissingLink = true

338 m.ThreadMuted = tm.ThreadMuted

339 m.ThreadCollapsed = tm.ThreadCollapsed

340 } else {

341 m.ThreadID = m.ID

342 }

343 return false, nil

344 }

345

346 npendingResolved := 0

347

348 // Resolve pending messages that wait on m.MessageID to be resolved, recursively.

349 var resolvePending func(tm Message, cyclic bool) error

350 resolvePending = func(tm Message, cyclic bool) error {

351 if tm.MessageID == "" {

352 return nil

353 }

354 l := pending[tm.MessageID]

355 delete(pending, tm.MessageID)

356 for _, mi := range l {

357 m := Message{ID: mi.ID}

358 if err := tx.Get(&m); err != nil {

359 return fmt.Errorf("get message %d for resolving pending thread for message-id %s, %d: %w", mi.ID, tm.MessageID, tm.ID, err)

360 }

361 if m.ThreadID != 0 {

362 // ThreadID already set because this is a cyclic message. If we would assign a

363 // parent again, we would create a cycle.

364 if m.MessageID != tm.MessageID && !cyclic {

365 panic(fmt.Sprintf("threadid already set (%d) while handling non-cyclic message id %d/%q and with different message-id %q as parent message id %d", m.ThreadID, m.ID, m.MessageID, tm.MessageID, tm.ID))

366 }

367 continue

368 }

369 assignParent(&m, tm, false)

370 m.ThreadMissingLink = mi.ThreadMissingLink

371 if err := tx.Update(&m); err != nil {

372 return fmt.Errorf("update message %d for resolving pending thread for message-id %s, %d: %w", mi.ID, tm.MessageID, tm.ID, err)

373 }

374 if err := resolvePending(m, cyclic); err != nil {

375 return err

376 }

377 npendingResolved++

378 }

379 return nil

380 }

381

382 // Output of the worker goroutines.

383 type threadPrep struct {

384 references []string

385 inReplyTo []string

386 subject string

387 }

388

389 // Single allocation.

390 threadingFields := [][]byte{

391 []byte("references"),

392 []byte("in-reply-to"),

393 []byte("subject"),

394 }

395

396 // Worker goroutine function. We start with a reasonably large buffer for reading

397 // the header into. And we have scratch space to copy the needed headers into. That

398 // means we normally won't allocate any more buffers.

399 prepareMessages := func(in, out chan moxio.Work[Message, threadPrep]) {

400 headerbuf := make([]byte, 8*1024)

401 scratch := make([]byte, 4*1024)

402 for {

403 w, ok := <-in

404 if !ok {

405 return

406 }

407

408 m := w.In

409 var partialPart struct {

410 HeaderOffset int64

411 BodyOffset int64

412 }

413 if err := json.Unmarshal(m.ParsedBuf, &partialPart); err != nil {

414 w.Err = fmt.Errorf("unmarshal part: %v", err)

415 } else {

416 size := partialPart.BodyOffset - partialPart.HeaderOffset

417 if int(size) > len(headerbuf) {

418 headerbuf = make([]byte, size)

419 }

420 if size > 0 {

421 buf := headerbuf[:int(size)]

422 err := func() error {

423 mr := a.MessageReader(m)

424 defer mr.Close()

425

426 // ReadAt returns whole buffer or error. Single read should be fast.

427 n, err := mr.ReadAt(buf, partialPart.HeaderOffset)

428 if err != nil || n != len(buf) {

429 return fmt.Errorf("read header: %v", err)

430 }

431 return nil

432 }()

433 if err != nil {

434 w.Err = err

435 } else if h, err := message.ParseHeaderFields(buf, scratch, threadingFields); err != nil {

436 w.Err = err

437 } else {

438 w.Out.references = h["References"]

439 w.Out.inReplyTo = h["In-Reply-To"]

440 l := h["Subject"]

441 if len(l) > 0 {

442 w.Out.subject = l[0]

443 }

444 }

445 }

446 }

447

448 out <- w

449 }

450 }

451

452 // Assign threads to messages, possibly in batches.

453 nassigned := 0

454 for {

455 n := 0

456 tx = txOpt

457 if tx == nil {

458 var err error

459 tx, err = a.DB.Begin(ctx, true)

460 if err != nil {

461 return fmt.Errorf("begin transaction: %w", err)

462 }

463 }

464

465 processMessage := func(m Message, prep threadPrep) error {

466 pend, err := assign(&m, prep.references, prep.inReplyTo, prep.subject)

467 if err != nil {

468 return fmt.Errorf("for msgid %d: %w", m.ID, err)

469 } else if pend {

470 return nil

471 }

472 if m.ThreadID == 0 {

473 panic(fmt.Sprintf("no threadid after assign of message id %d/%q", m.ID, m.MessageID))

474 }

475 // Fields have been set, store in database and resolve messages waiting for this MessageID.

476 if slices.Contains(m.ThreadParentIDs, m.ID) {

477 panic(fmt.Sprintf("message id %d/%q contains itself in parent ids %v", m.ID, m.MessageID, m.ThreadParentIDs))

478 }

479 if err := tx.Update(&m); err != nil {

480 return err

481 }

482 if err := resolvePending(m, false); err != nil {

483 return fmt.Errorf("resolving pending message-id: %v", err)

484 }

485 return nil

486 }

487

488 // Use multiple worker goroutines to parse headers from on-disk messages.

489 procs := runtime.GOMAXPROCS(0)

490 wq := moxio.NewWorkQueue[Message, threadPrep](2*procs, 4*procs, prepareMessages, processMessage)

491

492 // We assign threads in order by ID, so messages delivered in between our

493 // transaction will get assigned threads too: they'll have the highest id's.

494 q := bstore.QueryTx[Message](tx)

495 q.FilterGreaterEqual("ID", startMessageID)

496 q.FilterEqual("Expunged", false)

497 q.SortAsc("ID")

498 err := q.ForEach(func(m Message) error {

499 // Batch number of changes, so we give other users of account a change to run.

500 if txOpt == nil && n >= batchSize {

501 return bstore.StopForEach

502 }

503 // Starting point for next batch.

504 startMessageID = m.ID + 1

505 // Don't process again. Can happen when earlier upgrade was aborted.

506 if m.ThreadID != 0 {

507 return nil

508 }

509

510 n++

511 return wq.Add(m)

512 })

513 if err == nil {

514 err = wq.Finish()

515 }

516 wq.Stop()

517

518 if err == nil && txOpt == nil {

519 err = tx.Commit()

520 tx = nil

521 }

522 if err != nil {

523 return fmt.Errorf("assigning threads: %w", err)

524 }

525 if n == 0 {

526 break

527 }

528 nassigned += n

529 if nassigned%100000 == 0 {

530 log.Debug("assigning threads, progress", slog.Int("count", nassigned), slog.Int("unresolved", len(pending)))

531 if _, err := fmt.Fprintf(progressWriter, "assigning threads, progress: %d messages\n", nassigned); err != nil {

532 return fmt.Errorf("writing progress: %v", err)

533 }

534 }

535 }

536 if _, err := fmt.Fprintf(progressWriter, "assigning threads, done: %d messages\n", nassigned); err != nil {

537 return fmt.Errorf("writing progress: %v", err)

538 }

539

540 log.Debug("assigning threads, mostly done, finishing with resolving of cyclic messages", slog.Int("count", nassigned), slog.Int("unresolved", len(pending)))

541

542 if _, err := fmt.Fprintf(progressWriter, "assigning threads, resolving %d cyclic pending message-ids\n", len(pending)); err != nil {

543 return fmt.Errorf("writing progress: %v", err)

544 }

545

546 // Remaining messages in pending have cycles and possibly tails. The cycle is at

547 // the head of the thread. Once we resolve that, the rest of the thread can be

548 // resolved too. Ignoring self-references (duplicate messages), there can only be

549 // one cycle, and it is at the head. So we look for cycles, ignoring

550 // self-references, and resolve a message as soon as we see the cycle.

551

552 parent := map[string]string{} // Child Message-ID pointing to the parent Message-ID, excluding self-references.