Improve resilience (downtime history backfilling, better free disk space control)

config/config.go

@@ -49,6 +49,15 @@ type TelegramConfig struct {
 	Verbosity    uint8                 `yaml:":tdlib_verbosity"`
 	MAMThreshold uint32                `yaml:":mam_threshold"`
 	Tdlib        TelegramTdlibConfig   `yaml:":tdlib"`
+
+	// HistoryBackfillLimit caps how many messages a single chat's
+	// downtime-backfill sweep will pull on reconnect (see
+	// gateway.HistoryBackfillLimit). 0/unset falls back to a default of 1000.
+	HistoryBackfillLimit uint32 `yaml:":history_backfill_limit"`
+
+	// HistoryBackfillWorkers bounds how many chats backfill concurrently
+	// (see gateway.HistoryBackfillWorkers). 0/unset falls back to 4.
+	HistoryBackfillWorkers uint32 `yaml:":history_backfill_workers"`
 }
 
 // TelegramContentConfig is for :content: subtree
@@ -58,6 +67,12 @@ type TelegramContentConfig struct {
 	Upload string `yaml:":upload"`
 	User   string `yaml:":user"`
 	Quota  string `yaml:":quota"`
+
+	// MinFreeSpace is the minimum free disk space (e.g. "64M") required on
+	// the filesystem hosting TDlib's own files directory before a file
+	// download is allowed to proceed (see gateway.DownloadSpaceMargin).
+	// 0/unset falls back to a default of 64 MiB.
+	MinFreeSpace string `yaml:":min_free_space"`
 }
 
 // TelegramTdlibConfig is for :tdlib: subtree

config_schema.json

@@ -27,6 +27,9 @@
             },
             ":quota": {
               "type": "string"
+            },
+            ":min_free_space": {
+              "type": "string"
             }
           }
         },
@@ -36,6 +39,12 @@
         ":mam_threshold": {
           "type": "integer"
         },
+        ":history_backfill_limit": {
+          "type": "integer"
+        },
+        ":history_backfill_workers": {
+          "type": "integer"
+        },
         ":tdlib": {
           "required": [":client"],
           "type": "object",

go.mod

@@ -15,6 +15,7 @@ require (
 	github.com/xdg-go/stringprep v1.0.4
 	github.com/zelenin/go-tdlib v0.5.2
 	golang.org/x/crypto v0.48.0
+	golang.org/x/sys v0.41.0
 	gopkg.in/hraban/opus.v2 v2.0.0-20230925203106-0188a62cb302
 	gopkg.in/yaml.v2 v2.2.4
 	gosrc.io/xmpp v0.5.2-0.20211214110136-5f99e1cd06e1
@@ -63,7 +64,6 @@ require (
 	github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e // indirect
 	go.opencensus.io v0.22.5 // indirect
 	golang.org/x/net v0.50.0 // indirect
-	golang.org/x/sys v0.41.0 // indirect
 	golang.org/x/term v0.40.0 // indirect
 	golang.org/x/text v0.34.0 // indirect
 	golang.org/x/time v0.14.0 // indirect

persistence/sessions.go

@@ -58,6 +58,13 @@ type Session struct {
 	// (the zero value, e.g. for an account that predates this field)
 	// means "auto", matching e2ee.ModeAuto's own default.
 	OMEMO string `yaml:":omemo"`
+
+	// LastOnline is a Unix timestamp, refreshed periodically while the
+	// TDlib session is online and on graceful shutdown. It's used only to
+	// size a downtime-backfill window on reconnect (telegram.Client.Connect)
+	// - not user-configurable, so it's not part of ConfigKeys. Zero means
+	// "never seen online yet", which skips backfill entirely.
+	LastOnline int64 `yaml:":lastonline"`
 }
 
 const (

telegram/client.go

@@ -219,6 +219,12 @@ func buildCallAdapters(tdClient *client.Client, jid string, deps CallDeps) (
 
 type clientLocks struct {
 	authorizationReady sync.Mutex
+	// chatMessageLocksLock guards chatMessageLocks itself (the map, not its
+	// values): every existing caller of getChatMessageLock runs serially on
+	// the single updateHandler/dispatchUpdate goroutine, so this was
+	// previously safe by construction, but downtime backfill now populates
+	// it from an independent goroutine too
+	chatMessageLocksLock sync.Mutex
 	chatMessageLocks     map[int64]*sync.Mutex
 	resourcesLock        sync.Mutex
 	outboxLock         sync.Mutex
@@ -230,6 +236,13 @@ type clientLocks struct {
 	loginFinish        barrier
 	uploadingFilesLock sync.Mutex
 
+	// deliveredMessageIds tracks, per chat, the highest message id already
+	// handed to ProcessIncomingMessage - guards against double delivery
+	// between live updateNewMessage processing and downtime backfill
+	// (message ids are guaranteed strictly increasing per chat in TDlib)
+	deliveredMessageIds     map[int64]int64
+	deliveredMessageIdsLock sync.Mutex
+
 	authorizerReadLock  sync.Mutex
 	authorizerWriteLock sync.Mutex
 
@@ -308,6 +321,7 @@ func NewClient(conf config.TelegramConfig, jid string, component *xmpp.Component
 		MessageIdChanges:   make(map[int64]map[int64]*newId),
 		locks: clientLocks{
 			chatMessageLocks:    make(map[int64]*sync.Mutex),
+			deliveredMessageIds: make(map[int64]int64),
 		},
 	}, nil
 }

telegram/connect.go

@@ -2,6 +2,7 @@ package telegram
 
 import (
 	"github.com/pkg/errors"
+	"sync"
 	"time"
 
 	"dev.narayana.im/narayana/telegabber/xmpp/gateway"
@@ -156,12 +157,44 @@ func (c *Client) Connect(resource string, wasSessionLoginEmpty bool) error {
 	log.Debug("waiting for loginFinish")
 	c.locks.loginFinish.Wait()
 
+	// captured before we touch LastOnline again, to size the downtime
+	// backfill window below
+	lastOnline := c.Session.LastOnline
+
+	// Reserve chats needing a downtime backfill *before* live update
+	// processing starts, so a live message can never race ahead of that
+	// chat's older backfilled history (see reserveChatsForBackfill) - this
+	// means fetching the chat list here, synchronously, rather than in the
+	// background warm-up below. Only for KeepOnline sessions: KeepOnline=false
+	// already means "I'd rather lose a history gap than get flooded on
+	// reconnect" (relying on the XMPP server's own offline queue), and the
+	// same preference applies to a transport-wide downtime gap. lastOnline
+	// being zero means there's no prior watermark (first-ever connect), so
+	// there's nothing to catch up on either.
+	var backfillReserved map[int64]*sync.Mutex
+	if c.Session.KeepOnline && lastOnline != 0 {
+		chats, err := c.client.GetChats(&client.GetChatsRequest{
+			Limit: chatsLimit,
+		})
+		if err != nil {
+			log.Errorf("Could not retrieve chats for downtime backfill: %v", err)
+		} else {
+			backfillReserved = c.reserveChatsForBackfill(chats.ChatIds)
+		}
+	}
+
 	go c.updateHandler()
 	log.Warn("Going online")
 	c.online.Store(true)
 	c.locks.authorizationReady.Unlock()
 	c.addResource(resource)
 
+	if backfillReserved != nil {
+		// runs independently so a large backfill doesn't delay the "Logged
+		// in as" notification/subscribe below
+		go c.backfillDowntimeHistory(backfillReserved, lastOnline)
+	}
+
 	go func() {
 		chats, err := c.client.GetChats(&client.GetChatsRequest{
 			Limit: chatsLimit,

telegram/handlers.go

@@ -35,6 +35,9 @@ func int64SliceToStringSlice(ints []int64) []string {
 }
 
 func (c *Client) getChatMessageLock(chatID int64) *sync.Mutex {
+	c.locks.chatMessageLocksLock.Lock()
+	defer c.locks.chatMessageLocksLock.Unlock()
+
 	lock, ok := c.locks.chatMessageLocks[chatID]
 	if !ok {
 		lock = &sync.Mutex{}
@@ -44,6 +47,33 @@ func (c *Client) getChatMessageLock(chatID int64) *sync.Mutex {
 	return lock
 }
 
+// reserveChatsForBackfill locks (creating if needed) the per-chat message
+// lock for every given chat id and returns them already held, keyed by
+// chat id. Call this before starting live update processing for a
+// downtime-backfill sweep: since updateNewMessage acquires the same lock
+// before delivering a live message, any live message for a reserved chat
+// will block until the backfill sweep delivers that chat's history and
+// releases it - guaranteeing backfilled (older) messages are always
+// delivered before a live (newer) one for the same chat, rather than
+// racing with it. Chats not in chatIds are left untouched and unaffected.
+func (c *Client) reserveChatsForBackfill(chatIds []int64) map[int64]*sync.Mutex {
+	c.locks.chatMessageLocksLock.Lock()
+	defer c.locks.chatMessageLocksLock.Unlock()
+
+	reserved := make(map[int64]*sync.Mutex, len(chatIds))
+	for _, chatId := range chatIds {
+		lock, ok := c.locks.chatMessageLocks[chatId]
+		if !ok {
+			lock = &sync.Mutex{}
+			c.locks.chatMessageLocks[chatId] = lock
+		}
+		lock.Lock()
+		reserved[chatId] = lock
+	}
+
+	return reserved
+}
+
 func (c *Client) cleanTempFile(path string) {
 	os.Remove(path)
 
@@ -84,6 +114,20 @@ func (c *Client) updateHandler() {
 	defer listener.Close()
 
 	for update := range listener.Updates {
+		c.dispatchUpdate(update)
+	}
+}
+
+// dispatchUpdate handles a single TDlib update, recovering from any panic so
+// that one malformed update (or a downstream failure like a disk-full file
+// write) can't take down the whole process for every hosted session.
+func (c *Client) dispatchUpdate(update client.Type) {
+	defer func() {
+		if r := recover(); r != nil {
+			log.Errorf("Recovered from panic while handling update %T: %v", update, r)
+		}
+	}()
+
 	if update.GetClass() == client.ClassUpdate {
 		// debug probe: did UpdateCall reach the listener at all?
 		if t := update.GetType(); t == client.TypeUpdateCall || t == client.TypeUpdateNewCallSignalingData {
@@ -153,13 +197,12 @@ func (c *Client) updateHandler() {
 			c.callDeps.TgManager.OnNewSignalingData(c.jid, typedUpdate)
 		default:
 			// log only handled types
-				continue
+			return
 		}
 
 		log.Debugf("%#v", update)
 	}
 }
-}
 
 // new user discovered
 func (c *Client) updateUser(update *client.UpdateUser) {
@@ -230,6 +273,11 @@ func (c *Client) updateNewMessage(update *client.UpdateNewMessage) {
 	go func() {
 		lock.Lock()
 		defer lock.Unlock()
+		defer func() {
+			if r := recover(); r != nil {
+				log.Errorf("Recovered from panic while processing message %v in chat %v: %v", update.Message.Id, chatId, r)
+			}
+		}()
 
 		var forceCmd bool
 		if c.LastBotCmdString != "" && update.Message.IsOutgoing {

telegram/utils.go

@@ -16,6 +16,8 @@ import (
 	"sort"
 	"strconv"
 	"strings"
+	"sync"
+	"syscall"
 	"time"
 	"unicode/utf8"
 
@@ -27,6 +29,7 @@ import (
 	log "github.com/sirupsen/logrus"
 	"github.com/soheilhy/args"
 	"github.com/zelenin/go-tdlib/client"
+	"golang.org/x/sys/unix"
 )
 
 type VCardInfo struct {
@@ -136,6 +139,14 @@ func NewMessageLimitSince(since int64) *MessageLimit {
 	return &limit
 }
 
+// NewMessageLimitSinceCapped is like NewMessageLimitSince, but overrides the
+// default 1000-message safety cap (see getNLastMessages) with maxMessages.
+func NewMessageLimitSinceCapped(since int64, maxMessages int32) *MessageLimit {
+	limit := NewMessageLimitSince(since)
+	limit.Messages = maxMessages
+	return limit
+}
+
 const AVATAR_SIZE_LIMIT int64 = 128 * 1024
 
 const (
@@ -1313,7 +1324,12 @@ func (c *Client) PermastoreFile(file *client.File, clone bool) (string, string)
 		}
 
 		size64 := uint64(file.Size)
-		c.prepareDiskSpace(size64)
+		if !c.prepareDiskSpace(size64) {
+			if !clone {
+				c.forgetTdlibFile(file.Id)
+			}
+			return "<ERROR>", ""
+		}
 
 		// detect uploading files, there's no remote id for them yet
 		c.locks.uploadingFilesLock.Lock()
@@ -1329,6 +1345,8 @@ func (c *Client) PermastoreFile(file *client.File, clone bool) (string, string)
 		dest := c.content.Path + "/" + basename // destination path
 		link = c.content.Link + "/" + basename  // download link
 
+		var alreadyStored bool
+
 		if clone {
 			file, path, err := c.ForceOpenFile(file, 1)
 			if err == nil {
@@ -1344,20 +1362,27 @@ func (c *Client) PermastoreFile(file *client.File, clone bool) (string, string)
 				// create destination
 				tempFile, err := os.OpenFile(dest, os.O_CREATE|os.O_EXCL|os.O_WRONLY, mode)
 				if err != nil {
-					pathErr := err.(*os.PathError)
-					if pathErr.Err.Error() == "file exists" {
+					if os.IsExist(err) {
 						log.Warn(err.Error())
 						return src, link
+					}
+					if errors.Is(err, syscall.ENOSPC) {
+						log.Errorf("Not enough disk space to store %v, dropping it: %v", basename, err)
 					} else {
 						log.Errorf("File creation error: %v", err)
-						return "<ERROR>", ""
 					}
+					return "<ERROR>", ""
 				}
 				defer tempFile.Close()
 				// copy
 				_, err = io.Copy(tempFile, file)
 				if err != nil {
+					if errors.Is(err, syscall.ENOSPC) {
+						log.Errorf("Not enough disk space to store %v, dropping it: %v", basename, err)
+					} else {
 						log.Errorf("File copying error: %v", err)
+					}
+					os.Remove(dest) // don't leave a truncated file behind
 					return "<ERROR>", ""
 				}
 			} else if path != "" {
@@ -1371,9 +1396,19 @@ func (c *Client) PermastoreFile(file *client.File, clone bool) (string, string)
 			// move
 			err = os.Rename(src, dest)
 			if err != nil {
-				linkErr := err.(*os.LinkError)
-				if linkErr.Err.Error() == "file exists" {
+				// whatever the reason, our local copy is either redundant
+				// (dest already exists) or one we're giving up on - either
+				// way tell TDlib it can reclaim the space, rather than
+				// leaving a file it moved out from under it stuck in its
+				// own storage forever
+				defer c.forgetTdlibFile(file.Id)
+
+				if os.IsExist(err) {
 					log.Warn(err.Error())
+					alreadyStored = true
+				} else if errors.Is(err, syscall.ENOSPC) {
+					log.Errorf("Not enough disk space to move %v, dropping it: %v", basename, err)
+					return "<ERROR>", ""
 				} else {
 					log.Errorf("File moving error: %v", err)
 					return "<ERROR>", ""
@@ -1399,13 +1434,27 @@ func (c *Client) PermastoreFile(file *client.File, clone bool) (string, string)
 			}
 		}
 
-		// copy or move should have succeeded at this point
+		// copy or move should have succeeded at this point, unless the
+		// destination was already there (e.g. a prior partial move)
+		if !alreadyStored {
 			gateway.CachedStorageSize += size64
 		}
+	}
 
 	return src, link
 }
 
+// forgetTdlibFile tells TDlib it no longer needs to keep fileId in its own
+// local file cache. Used when PermastoreFile has decided a locally-moved
+// file is redundant or unrecoverable, so that a file we've stopped tracking
+// via gateway's own quota doesn't linger in TDlib's storage indefinitely.
+func (c *Client) forgetTdlibFile(fileId int32) {
+	_, err := c.client.DeleteFile(&client.DeleteFileRequest{FileId: fileId})
+	if err != nil {
+		log.Warnf("Could not tell TDlib to delete file %v from its own storage: %v", fileId, err)
+	}
+}
+
 func (c *Client) formatBantime(hours int64) int32 {
 	var until int32
 	if hours > 0 {
@@ -1799,7 +1848,12 @@ func (c *Client) ensureDownloadFile(file *client.File) *client.File {
 	defer gateway.StorageLock.Unlock()
 
 	if file != nil {
-		c.prepareDiskSpace(uint64(file.Size))
+		if !c.prepareDiskSpace(uint64(file.Size)) {
+			return file
+		}
+		if !hasFreeDiskSpace(c.parameters.FilesDirectory, uint64(file.Size)) {
+			return file
+		}
 
 		newFile, err := c.DownloadFile(file.Id, 1, true)
 		if err == nil {
@@ -1831,6 +1885,11 @@ func (c *Client) ProcessIncomingMessage(chatId int64, message *client.Message) {
 		return
 	}
 
+	if !c.markMessageDelivered(chatId, message.Id) {
+		log.Debugf("Skipping already-delivered message %v in chat %v", message.Id, chatId)
+		return
+	}
+
 	chat, _, _ := c.GetContactByID(chatId, nil, true)
 	safeToSend := true
 	groupChatFrom := ""
@@ -2458,6 +2517,9 @@ func (c *Client) getNLastMessages(chatID int64, limit *MessageLimit) ([]*client.
 		}
 	case MessageLimitSince:
 		safetyLimit = 1000
+		if limit.Messages > 0 {
+			safetyLimit = limit.Messages
+		}
 	}
 
 	safetyLoop:
@@ -2511,6 +2573,108 @@ func (c *Client) getNLastMessages(chatID int64, limit *MessageLimit) ([]*client.
 	return messages, nil
 }
 
+// backfillDowntimeHistory fetches, for every reserved chat, the messages
+// that arrived while the whole transport process (not just one XMPP
+// resource - see Client.Disconnect's KeepOnline handling for that case)
+// was down, and feeds them through the same pipeline live messages use so
+// attachments come through normally - unlike !history/!search, which never
+// download attachments. sinceUnix is 0 on a session's first-ever connect
+// (no LastOnline watermark persisted yet), which is deliberately left
+// alone here: there's no prior "downtime" to speak of, so no backfill
+// should run.
+//
+// reserved must be the result of reserveChatsForBackfill, called before
+// live update processing started for this connection: each chat's lock is
+// already held, and backfillChatHistory releases it once that chat's
+// history has been delivered, so a live message for the same chat that
+// arrives in the meantime blocks until backfill catches it up - rather
+// than racing ahead and being delivered before older backfilled messages.
+// Chats are processed concurrently (bounded by gateway.HistoryBackfillWorkers)
+// so one huge or FLOOD_WAIT-throttled chat doesn't hold up the others.
+func (c *Client) backfillDowntimeHistory(reserved map[int64]*sync.Mutex, sinceUnix int64) {
+	if sinceUnix == 0 {
+		// nothing to catch up on; still release every reservation
+		for _, lock := range reserved {
+			lock.Unlock()
+		}
+		return
+	}
+
+	log.Warnf("Transport was down for ≈%v, backfilling history for %v chats", time.Since(time.Unix(sinceUnix, 0)).Round(time.Second), len(reserved))
+
+	workers := int(gateway.HistoryBackfillWorkers)
+	if workers <= 0 {
+		workers = 1
+	}
+
+	jobs := make(chan int64, len(reserved))
+	for chatId := range reserved {
+		jobs <- chatId
+	}
+	close(jobs)
+
+	var wg sync.WaitGroup
+	for i := 0; i < workers; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			for chatId := range jobs {
+				c.backfillChatHistory(chatId, sinceUnix, reserved[chatId])
+			}
+		}()
+	}
+	wg.Wait()
+}
+
+// backfillChatHistory fetches and delivers one chat's history since the
+// given watermark, oldest first, then releases lock (already held by the
+// caller via reserveChatsForBackfill). If a chat accumulated more than
+// gateway.HistoryBackfillLimit messages during the downtime, only the most
+// recent ones are delivered - logged and skipped, rather than blocking
+// reconnect indefinitely trying to fetch an unbounded gap.
+func (c *Client) backfillChatHistory(chatId int64, since int64, lock *sync.Mutex) {
+	defer lock.Unlock()
+	defer func() {
+		if r := recover(); r != nil {
+			log.Errorf("Recovered from panic while backfilling chat %v: %v", chatId, r)
+		}
+	}()
+
+	limit := int32(gateway.HistoryBackfillLimit)
+	messages, err := c.getNLastMessages(chatId, NewMessageLimitSinceCapped(since, limit))
+	if err != nil {
+		log.Errorf("Failed to backfill history for chat %v, dropping this chat's gap: %v", chatId, err)
+		return
+	}
+	if len(messages) == 0 {
+		return
+	}
+	if limit > 0 && len(messages) >= int(limit) {
+		log.Warnf("Chat %v had ≥%v messages since downtime; delivering only the most recent ones", chatId, limit)
+	}
+
+	// oldest first, matching live delivery order
+	for i := len(messages) - 1; i >= 0; i-- {
+		c.ProcessIncomingMessage(chatId, messages[i])
+	}
+}
+
+// markMessageDelivered records that messageId is about to be delivered for
+// chatId, returning false if an equal-or-later message was already
+// delivered. This guards against double delivery between live
+// updateNewMessage processing and downtime backfill (message ids are
+// guaranteed strictly increasing per chat in TDlib).
+func (c *Client) markMessageDelivered(chatId, messageId int64) bool {
+	c.locks.deliveredMessageIdsLock.Lock()
+	defer c.locks.deliveredMessageIdsLock.Unlock()
+
+	if c.locks.deliveredMessageIds[chatId] >= messageId {
+		return false
+	}
+	c.locks.deliveredMessageIds[chatId] = messageId
+	return true
+}
+
 // GetMessagesBetween lazily fetches message history between given ids (from exclusive, last inclusive), also calculating completeness flag; negative limit means messages from the end
 func (c *Client) GetMessagesBetween(chatID, fromMessageId, lastMessageId int64, limit int32, reverse bool) (messages []*client.Message, complete bool, err error) {
 	log.WithFields(log.Fields{
@@ -2668,6 +2832,10 @@ func (c *Client) ForceOpenFile(tgFile *client.File, priority int32) (*os.File, s
 	} else
 	// obtain the photo right now if still not downloaded
 	if !tgFile.Local.IsDownloadingCompleted {
+		if !hasFreeDiskSpace(c.parameters.FilesDirectory, uint64(tgFile.Size)) {
+			return nil, path, errors.New("Not enough disk space to download file")
+		}
+
 		tdFile, tdErr := c.DownloadFile(tgFile.Id, priority, true)
 		if tdErr == nil {
 			path = tdFile.Local.Path
@@ -2818,17 +2986,54 @@ func (c *Client) sendPresence(args ...args.V) error {
 	return gateway.SendPresence(c.xmpp, c.jid, args...)
 }
 
-func (c *Client) prepareDiskSpace(size uint64) {
-	if gateway.StorageQuota > 0 && c.content.Path != "" {
-		var loweredQuota uint64
-		if gateway.StorageQuota >= size {
-			loweredQuota = gateway.StorageQuota - size
+// hasFreeDiskSpace reports whether the filesystem hosting path currently
+// has at least size + gateway.DownloadSpaceMargin bytes free. A statfs
+// failure is treated as "can't tell, proceed".
+func hasFreeDiskSpace(path string, size uint64) bool {
+	var stat unix.Statfs_t
+	if err := unix.Statfs(path, &stat); err != nil {
+		log.Warnf("Could not check free disk space at %v, proceeding anyway: %v", path, err)
+		return true
 	}
+
+	free := stat.Bavail * uint64(stat.Bsize)
+	needed := size + gateway.DownloadSpaceMargin
+
+	if free < needed {
+		log.Errorf("Not enough free disk space at %v: %v bytes free, need %v (%v file + %v margin)", path, free, needed, size, gateway.DownloadSpaceMargin)
+		return false
+	}
+
+	return true
+}
+
+// prepareDiskSpace makes room in c.content.Path for an incoming file of the
+// given size by evicting old files down to a lowered quota. Returns false
+// if the file should be discarded instead: either it exceeds the whole
+// quota by itself (cleaning everything still wouldn't fit it), or real
+// free disk space is still short after cleaning.
+func (c *Client) prepareDiskSpace(size uint64) bool {
+	if gateway.StorageQuota == 0 || c.content.Path == "" {
+		return true
+	}
+
+	if size > gateway.StorageQuota {
+		log.Errorf("File is %v bytes, exceeding the entire storage quota (%v bytes); discarding it rather than evicting everything else in a futile attempt to fit it", size, gateway.StorageQuota)
+		return false
+	}
+
+	loweredQuota := gateway.StorageQuota - size
 	if gateway.CachedStorageSize >= loweredQuota {
 		log.Warn("Storage is rapidly clogged")
 		gateway.CleanOldFiles(c.content.Path, loweredQuota)
 	}
+
+	if !hasFreeDiskSpace(c.content.Path, size) {
+		log.Errorf("Still not enough real disk space at %v after cleaning; discarding the file", c.content.Path)
+		return false
 	}
+
+	return true
 }
 
 func (c *Client) GetVcardInfo(toID int64) (VCardInfo, error) {

xmpp/component.go

@@ -94,8 +94,24 @@ func NewComponent(conf config.XMPPConfig, tc config.TelegramConfig, idsPath stri
 		}
 	}
 
+	if tc.Content.MinFreeSpace != "" {
+		gateway.DownloadSpaceMargin, err = parseSize(tc.Content.MinFreeSpace)
+		if err != nil {
+			log.Warnf("Error parsing min_free_space: %v; using the default", err)
+			gateway.DownloadSpaceMargin = 64 * 1024 * 1024
+		}
+	}
+
 	gateway.MAMThreshold = tc.MAMThreshold
 
+	if tc.HistoryBackfillLimit > 0 {
+		gateway.HistoryBackfillLimit = tc.HistoryBackfillLimit
+	}
+
+	if tc.HistoryBackfillWorkers > 0 {
+		gateway.HistoryBackfillWorkers = tc.HistoryBackfillWorkers
+	}
+
 	options := xmpp.ComponentOptions{
 		TransportConfiguration: xmpp.TransportConfiguration{
 			Address: conf.Host + ":" + conf.Port,
@@ -255,6 +271,11 @@ func heartbeat(component *xmpp.Component) {
 
 		sessionLock.Lock()
 		for _, session := range sessions {
+			if session.Online() {
+				session.Session.LastOnline = time.Now().Unix()
+				gateway.DirtySessions = true
+			}
+
 			session.DelayedStatusesLock.Lock()
 			for chatID, delayedStatus := range session.DelayedStatuses {
 				if delayedStatus.TimestampExpired <= now {
@@ -392,6 +413,12 @@ func Close(component *xmpp.Component) {
 	sessionLock.Lock()
 	// close all sessions
 	for _, session := range sessions {
+		if session.Online() {
+			// mark the exact moment we went offline, so a graceful
+			// shutdown doesn't cost the downtime-backfill window the
+			// up-to-60s staleness of the heartbeat-driven update above
+			session.Session.LastOnline = time.Now().Unix()
+		}
 		session.Disconnect("", true)
 	}
 	sessionLock.Unlock()

xmpp/gateway/gateway.go

@@ -98,6 +98,26 @@ var MessageOutgoingPermissionVersion = 0
 // MAMThreshold specifies a day limit behind which history should not be requested to avoid abuse detection and storage overload
 var MAMThreshold uint32
 
+// HistoryBackfillLimit caps how many messages a single chat's
+// downtime-backfill sweep will pull, regardless of gap size. 0 (unset)
+// means the default of 1000.
+var HistoryBackfillLimit uint32 = 1000
+
+// HistoryBackfillWorkers bounds how many chats backfill concurrently.
+// Chats are independent, so this is a concurrency cap, not a queue depth:
+// one huge or FLOOD_WAIT-throttled chat only occupies one worker slot,
+// leaving the rest free to make progress on other chats instead of
+// waiting behind it. 0 (unset) means the default of 4.
+var HistoryBackfillWorkers uint32 = 4
+
+// DownloadSpaceMargin is extra free-disk-space headroom, beyond a file's
+// own declared size, required before a TDlib download proceeds. Download
+// requests are already serialized system-wide via StorageLock, so this
+// only has to cover concurrent writes by other, unrelated processes on
+// the same machine - not a race against telegabber's own downloads. 0
+// (unset) means the default of 64 MiB.
+var DownloadSpaceMargin uint64 = 64 * 1024 * 1024
+
 // CHATNODE converts numeric id to node part of 1-1 chat JID
 func CHATNODE(chatId int64) string {
 	return strconv.FormatInt(chatId, 10)