Fix concurrency issues in controlclient, ipn, types/logger (#456)

Signed-Off-By: Dmytro Shynkevych <dmytro@tailscale.com>

control/controlclient/direct.go

@@ -92,6 +92,7 @@ type Direct struct {
 	authKey      string
 	tryingNewKey wgcfg.PrivateKey
 	expiry       *time.Time
+	// hostinfo is mutated in-place while mu is held.
 	hostinfo  *tailcfg.Hostinfo // always non-nil
 	endpoints []string
 	localPort uint16 // or zero to mean auto
@@ -262,6 +263,8 @@ func (c *Direct) doLogin(ctx context.Context, t *oauth2.Token, flags LoginFlags,
 	tryingNewKey := c.tryingNewKey
 	serverKey := c.serverKey
 	authKey := c.authKey
+	hostinfo := c.hostinfo
+	backendLogID := hostinfo.BackendLogID
 	expired := c.expiry != nil && !c.expiry.IsZero() && c.expiry.Before(c.timeNow())
 	c.mu.Unlock()
 
@@ -318,7 +321,7 @@ func (c *Direct) doLogin(ctx context.Context, t *oauth2.Token, flags LoginFlags,
 	if tryingNewKey == (wgcfg.PrivateKey{}) {
 		log.Fatalf("tryingNewKey is empty, give up")
 	}
-	if c.hostinfo.BackendLogID == "" {
+	if backendLogID == "" {
 		err = errors.New("hostinfo: BackendLogID missing")
 		return regen, url, err
 	}
@@ -326,7 +329,7 @@ func (c *Direct) doLogin(ctx context.Context, t *oauth2.Token, flags LoginFlags,
 		Version:    1,
 		OldNodeKey: tailcfg.NodeKey(oldNodeKey),
 		NodeKey:    tailcfg.NodeKey(tryingNewKey.Public()),
-		Hostinfo:   c.hostinfo,
+		Hostinfo:   hostinfo,
 		Followup:   url,
 	}
 	c.logf("RegisterReq: onode=%v node=%v fup=%v",
@@ -453,11 +456,12 @@ func (c *Direct) PollNetMap(ctx context.Context, maxPolls int, cb func(*NetworkM
 	serverURL := c.serverURL
 	serverKey := c.serverKey
 	hostinfo := c.hostinfo
+	backendLogID := hostinfo.BackendLogID
 	localPort := c.localPort
 	ep := append([]string(nil), c.endpoints...)
 	c.mu.Unlock()
 
-	if hostinfo.BackendLogID == "" {
+	if backendLogID == "" {
 		return errors.New("hostinfo: BackendLogID missing")
 	}
 

ipn/local.go

@@ -63,8 +63,10 @@ type LocalBackend struct {
 	stateKey StateKey
 	prefs    *Prefs
 	state    State
-	hiCache      *tailcfg.Hostinfo
-	netMapCache  *controlclient.NetworkMap
+	// hostinfo is mutated in-place while mu is held.
+	hostinfo *tailcfg.Hostinfo
+	// netMap is not mutated in-place once set.
+	netMap       *controlclient.NetworkMap
 	engineStatus EngineStatus
 	endpoints    []string
 	blocked      bool
@@ -106,11 +108,6 @@ func NewLocalBackend(logf logger.Logf, logid string, store StateStore, e wgengin
 	}
 	b.statusChanged = sync.NewCond(&b.statusLock)
 
-	if b.portpoll != nil {
-		go b.portpoll.Run(ctx)
-		go b.readPoller()
-	}
-
 	return b, nil
 }
 
@@ -146,11 +143,11 @@ func (b *LocalBackend) UpdateStatus(sb *ipnstate.StatusBuilder) {
 
 	// TODO: hostinfo, and its networkinfo
 	// TODO: EngineStatus copy (and deprecate it?)
-	if b.netMapCache != nil {
-		for id, up := range b.netMapCache.UserProfiles {
+	if b.netMap != nil {
+		for id, up := range b.netMap.UserProfiles {
 			sb.AddUser(id, up)
 		}
-		for _, p := range b.netMapCache.Peers {
+		for _, p := range b.netMap.Peers {
 			var lastSeen time.Time
 			if p.LastSeen != nil {
 				lastSeen = *p.LastSeen
@@ -184,104 +181,17 @@ func (b *LocalBackend) SetDecompressor(fn func() (controlclient.Decompressor, er
 	b.newDecompressor = fn
 }
 
-// Start applies the configuration specified in opts, and starts the
-// state machine.
-//
-// TODO(danderson): this function is trying to do too many things at
-// once: it loads state, or imports it, or updates prefs sometimes,
-// contains some settings that are one-shot things done by `tailscale
-// up` because we had nowhere else to put them, and there's no clear
-// guarantee that switching from one user's state to another is
-// actually a supported operation (it should be, but it's very unclear
-// from the following whether or not that is a safe transition).
-func (b *LocalBackend) Start(opts Options) error {
-	if opts.Prefs == nil && opts.StateKey == "" {
-		return errors.New("no state key or prefs provided")
-	}
-
-	if opts.Prefs != nil {
-		b.logf("Start: %v", opts.Prefs.Pretty())
-	} else {
-		b.logf("Start")
-	}
-
-	hi := controlclient.NewHostinfo()
-	hi.BackendLogID = b.backendLogID
-	hi.FrontendLogID = opts.FrontendLogID
-
-	b.mu.Lock()
-
-	if b.c != nil {
-		// TODO(apenwarr): avoid the need to reinit controlclient.
-		// This will trigger a full relogin/reconfigure cycle every
-		// time a Handle reconnects to the backend. Ideally, we
-		// would send the new Prefs and everything would get back
-		// into sync with the minimal changes. But that's not how it
-		// is right now, which is a sign that the code is still too
-		// complicated.
-		b.c.Shutdown()
-	}
-
-	if b.hiCache != nil {
-		hi.Services = b.hiCache.Services // keep any previous session and netinfo
-		hi.NetInfo = b.hiCache.NetInfo
-	}
-	b.hiCache = hi
-	b.state = NoState
-
-	if err := b.loadStateLocked(opts.StateKey, opts.Prefs, opts.LegacyConfigPath); err != nil {
-		b.mu.Unlock()
-		return fmt.Errorf("loading requested state: %v", err)
-	}
-
-	b.serverURL = b.prefs.ControlURL
-	hi.RoutableIPs = append(hi.RoutableIPs, b.prefs.AdvertiseRoutes...)
-	hi.RequestTags = append(hi.RequestTags, b.prefs.AdvertiseTags...)
-
-	b.notify = opts.Notify
-	b.netMapCache = nil
-	persist := b.prefs.Persist
-	b.mu.Unlock()
-
-	b.updateFilter(nil)
-
-	var err error
-	if persist == nil {
-		// let controlclient initialize it
-		persist = &controlclient.Persist{}
-	}
-	cli, err := controlclient.New(controlclient.Options{
-		Logf:            logger.WithPrefix(b.logf, "control: "),
-		Persist:         *persist,
-		ServerURL:       b.serverURL,
-		AuthKey:         opts.AuthKey,
-		Hostinfo:        hi,
-		KeepAlive:       true,
-		NewDecompressor: b.newDecompressor,
-		HTTPTestClient:  opts.HTTPTestClient,
-	})
-	if err != nil {
-		return err
-	}
-
-	b.mu.Lock()
-	b.c = cli
-	endpoints := b.endpoints
-	b.mu.Unlock()
-
-	if endpoints != nil {
-		cli.UpdateEndpoints(0, endpoints)
-	}
-
-	cli.SetStatusFunc(func(newSt controlclient.Status) {
-		if newSt.LoginFinished != nil {
+// setClientStatus is the callback invoked by the control client whenever it posts a new status.
+// Among other things, this is where we update the netmap, packet filters, DNS and DERP maps.
+func (b *LocalBackend) setClientStatus(st controlclient.Status) {
+	if st.LoginFinished != nil {
 		// Auth completed, unblock the engine
 		b.blockEngineUpdates(false)
 		b.authReconfig()
 		b.send(Notify{LoginFinished: &empty.Message{}})
 	}
-		if newSt.Persist != nil {
-			persist := *newSt.Persist // copy
+	if st.Persist != nil {
+		persist := *st.Persist // copy
 
 		b.mu.Lock()
 		b.prefs.Persist = &persist
@@ -296,12 +206,12 @@ func (b *LocalBackend) Start(opts Options) error {
 		}
 		b.send(Notify{Prefs: prefs})
 	}
-		if newSt.NetMap != nil {
+	if st.NetMap != nil {
 		// Netmap is unchanged only when the diff is empty.
 		changed := true
 		b.mu.Lock()
-			if b.netMapCache != nil {
-				diff := newSt.NetMap.ConciseDiffFrom(b.netMapCache)
+		if b.netMap != nil {
+			diff := st.NetMap.ConciseDiffFrom(b.netMap)
 			if strings.TrimSpace(diff) == "" {
 				changed = false
 				b.logf("netmap diff: (none)")
@@ -310,39 +220,39 @@ func (b *LocalBackend) Start(opts Options) error {
 			}
 		}
 		disableDERP := b.prefs != nil && b.prefs.DisableDERP
-			b.netMapCache = newSt.NetMap
+		b.netMap = st.NetMap
 		b.mu.Unlock()
 
-			b.send(Notify{NetMap: newSt.NetMap})
+		b.send(Notify{NetMap: st.NetMap})
 		// There is nothing to update if the map hasn't changed.
 		if changed {
-				b.updateFilter(newSt.NetMap)
-				b.updateDNSMap(newSt.NetMap)
+			b.updateFilter(st.NetMap)
+			b.updateDNSMap(st.NetMap)
 		}
 		if disableDERP {
 			b.e.SetDERPMap(nil)
 		} else {
-				b.e.SetDERPMap(newSt.NetMap.DERPMap)
+			b.e.SetDERPMap(st.NetMap.DERPMap)
 		}
 	}
-		if newSt.URL != "" {
-			b.logf("Received auth URL: %.20v...", newSt.URL)
+	if st.URL != "" {
+		b.logf("Received auth URL: %.20v...", st.URL)
 
 		b.mu.Lock()
 		interact := b.interact
-			b.authURL = newSt.URL
+		b.authURL = st.URL
 		b.mu.Unlock()
 
 		if interact > 0 {
 			b.popBrowserAuthNow()
 		}
 	}
-		if newSt.Err != "" {
+	if st.Err != "" {
 		// TODO(crawshaw): display in the UI.
-			b.logf("Received error: %v", newSt.Err)
+		b.logf("Received error: %v", st.Err)
 		return
 	}
-		if newSt.NetMap != nil {
+	if st.NetMap != nil {
 		b.mu.Lock()
 		if b.state == NeedsLogin {
 			b.prefs.WantRunning = true
@@ -353,15 +263,17 @@ func (b *LocalBackend) Start(opts Options) error {
 		b.SetPrefs(prefs)
 	}
 	b.stateMachine()
-	})
+}
 
-	b.e.SetStatusCallback(func(s *wgengine.Status, err error) {
+// setWgengineStatus is the callback by the wireguard engine whenever it posts a new status.
+// This updates the endpoints both in the backend and in the control client.
+func (b *LocalBackend) setWgengineStatus(s *wgengine.Status, err error) {
 	if err != nil {
 		b.logf("wgengine status error: %#v", err)
 		return
 	}
 	if s == nil {
-			b.logf("weird: non-error wgengine update with status=nil: %v", s)
+		b.logf("[unexpected] non-error wgengine update with status=nil: %v", s)
 		return
 	}
 
@@ -383,8 +295,106 @@ func (b *LocalBackend) Start(opts Options) error {
 	b.statusLock.Unlock()
 
 	b.send(Notify{Engine: &es})
+}
+
+// Start applies the configuration specified in opts, and starts the
+// state machine.
+//
+// TODO(danderson): this function is trying to do too many things at
+// once: it loads state, or imports it, or updates prefs sometimes,
+// contains some settings that are one-shot things done by `tailscale
+// up` because we had nowhere else to put them, and there's no clear
+// guarantee that switching from one user's state to another is
+// actually a supported operation (it should be, but it's very unclear
+// from the following whether or not that is a safe transition).
+func (b *LocalBackend) Start(opts Options) error {
+	if opts.Prefs == nil && opts.StateKey == "" {
+		return errors.New("no state key or prefs provided")
+	}
+
+	if opts.Prefs != nil {
+		b.logf("Start: %v", opts.Prefs.Pretty())
+	} else {
+		b.logf("Start")
+	}
+
+	hostinfo := controlclient.NewHostinfo()
+	hostinfo.BackendLogID = b.backendLogID
+	hostinfo.FrontendLogID = opts.FrontendLogID
+
+	b.mu.Lock()
+
+	if b.c != nil {
+		// TODO(apenwarr): avoid the need to reinit controlclient.
+		// This will trigger a full relogin/reconfigure cycle every
+		// time a Handle reconnects to the backend. Ideally, we
+		// would send the new Prefs and everything would get back
+		// into sync with the minimal changes. But that's not how it
+		// is right now, which is a sign that the code is still too
+		// complicated.
+		b.c.Shutdown()
+	}
+
+	if b.hostinfo != nil {
+		hostinfo.Services = b.hostinfo.Services // keep any previous session and netinfo
+		hostinfo.NetInfo = b.hostinfo.NetInfo
+	}
+	b.hostinfo = hostinfo
+	b.state = NoState
+
+	if err := b.loadStateLocked(opts.StateKey, opts.Prefs, opts.LegacyConfigPath); err != nil {
+		b.mu.Unlock()
+		return fmt.Errorf("loading requested state: %v", err)
+	}
+
+	b.serverURL = b.prefs.ControlURL
+	hostinfo.RoutableIPs = append(hostinfo.RoutableIPs, b.prefs.AdvertiseRoutes...)
+	hostinfo.RequestTags = append(hostinfo.RequestTags, b.prefs.AdvertiseTags...)
+
+	b.notify = opts.Notify
+	b.netMap = nil
+	persist := b.prefs.Persist
+	b.mu.Unlock()
+
+	b.updateFilter(nil)
+
+	var err error
+	if persist == nil {
+		// let controlclient initialize it
+		persist = &controlclient.Persist{}
+	}
+	cli, err := controlclient.New(controlclient.Options{
+		Logf:            logger.WithPrefix(b.logf, "control: "),
+		Persist:         *persist,
+		ServerURL:       b.serverURL,
+		AuthKey:         opts.AuthKey,
+		Hostinfo:        hostinfo,
+		KeepAlive:       true,
+		NewDecompressor: b.newDecompressor,
+		HTTPTestClient:  opts.HTTPTestClient,
 	})
+	if err != nil {
+		return err
+	}
 
+	// At this point, we have finished using hostinfo without synchronization,
+	// so it is safe to start readPoller which concurrently writes to it.
+	if b.portpoll != nil {
+		go b.portpoll.Run(b.ctx)
+		go b.readPoller()
+	}
+
+	b.mu.Lock()
+	b.c = cli
+	endpoints := b.endpoints
+	b.mu.Unlock()
+
+	if endpoints != nil {
+		cli.UpdateEndpoints(0, endpoints)
+	}
+
+	cli.SetStatusFunc(b.setClientStatus)
+	b.e.SetStatusCallback(b.setWgengineStatus)
 	b.e.SetNetInfoCallback(b.setNetInfo)
 
 	b.mu.Lock()
@@ -477,13 +487,11 @@ func (b *LocalBackend) readPoller() {
 		}
 
 		b.mu.Lock()
-		if b.hiCache == nil {
-			// TODO(bradfitz): it's a little weird that this port poller
-			// is started (by NewLocalBackend) before the Start call.
-			b.hiCache = new(tailcfg.Hostinfo)
+		if b.hostinfo == nil {
+			b.hostinfo = new(tailcfg.Hostinfo)
 		}
-		b.hiCache.Services = sl
-		hi := b.hiCache
+		b.hostinfo.Services = sl
+		hi := b.hostinfo
 		b.mu.Unlock()
 
 		b.doSetHostinfoFilterServices(hi)
@@ -617,13 +625,23 @@ func (b *LocalBackend) StartLoginInteractive() {
 // FakeExpireAfter implements Backend.
 func (b *LocalBackend) FakeExpireAfter(x time.Duration) {
 	b.logf("FakeExpireAfter: %v", x)
-	if b.netMapCache != nil {
-		e := b.netMapCache.Expiry
-		if e.IsZero() || time.Until(e) > x {
-			b.netMapCache.Expiry = time.Now().Add(x)
+
+	b.mu.Lock()
+	defer b.mu.Unlock()
+
+	if b.netMap == nil {
+		return
 	}
-		b.send(Notify{NetMap: b.netMapCache})
+
+	// This function is called very rarely,
+	// so we prefer to fully copy the netmap over introducing in-place modification here.
+	mapCopy := *b.netMap
+	e := mapCopy.Expiry
+	if e.IsZero() || time.Until(e) > x {
+		mapCopy.Expiry = time.Now().Add(x)
 	}
+	b.netMap = &mapCopy
+	b.send(Notify{NetMap: b.netMap})
 }
 
 func (b *LocalBackend) parseWgStatus(s *wgengine.Status) (ret EngineStatus) {
@@ -680,13 +698,13 @@ func (b *LocalBackend) SetPrefs(new *Prefs) {
 			b.logf("Failed to save new controlclient state: %v", err)
 		}
 	}
-	oldHi := b.hiCache
+	oldHi := b.hostinfo
 	newHi := oldHi.Clone()
 	newHi.RoutableIPs = append([]wgcfg.CIDR(nil), b.prefs.AdvertiseRoutes...)
 	if h := new.Hostname; h != "" {
 		newHi.Hostname = h
 	}
-	b.hiCache = newHi
+	b.hostinfo = newHi
 	b.mu.Unlock()
 
 	b.logf("SetPrefs: %v", new.Pretty())
@@ -695,15 +713,15 @@ func (b *LocalBackend) SetPrefs(new *Prefs) {
 		b.doSetHostinfoFilterServices(newHi)
 	}
 
-	b.updateFilter(b.netMapCache)
+	b.updateFilter(b.netMap)
 	// TODO(dmytro): when Prefs gain an EnableTailscaleDNS toggle, updateDNSMap here.
 
 	turnDERPOff := new.DisableDERP && !old.DisableDERP
 	turnDERPOn := !new.DisableDERP && old.DisableDERP
 	if turnDERPOff {
 		b.e.SetDERPMap(nil)
-	} else if turnDERPOn && b.netMapCache != nil {
-		b.e.SetDERPMap(b.netMapCache.DERPMap)
+	} else if turnDERPOn && b.netMap != nil {
+		b.e.SetDERPMap(b.netMap.DERPMap)
 	}
 
 	if old.WantRunning != new.WantRunning {
@@ -741,7 +759,7 @@ func (b *LocalBackend) doSetHostinfoFilterServices(hi *tailcfg.Hostinfo) {
 // NetMap returns the latest cached network map received from
 // controlclient, or nil if no network map was received yet.
 func (b *LocalBackend) NetMap() *controlclient.NetworkMap {
-	return b.netMapCache
+	return b.netMap
 }
 
 // blockEngineUpdate sets b.blocked to block, while holding b.mu. Its
@@ -762,7 +780,7 @@ func (b *LocalBackend) authReconfig() {
 	b.mu.Lock()
 	blocked := b.blocked
 	uc := b.prefs
-	nm := b.netMapCache
+	nm := b.netMap
 	b.mu.Unlock()
 
 	if blocked {
@@ -939,7 +957,7 @@ func (b *LocalBackend) nextState() State {
 	b.assertClientLocked()
 	var (
 		c           = b.c
-		netMap      = b.netMapCache
+		netMap      = b.netMap
 		state       = b.state
 		wantRunning = b.prefs.WantRunning
 	)
@@ -1037,13 +1055,13 @@ func (b *LocalBackend) Logout() {
 	b.mu.Lock()
 	b.assertClientLocked()
 	c := b.c
-	b.netMapCache = nil
+	b.netMap = nil
 	b.mu.Unlock()
 
 	c.Logout()
 
 	b.mu.Lock()
-	b.netMapCache = nil
+	b.netMap = nil
 	b.mu.Unlock()
 
 	b.stateMachine()
@@ -1056,13 +1074,13 @@ func (b *LocalBackend) assertClientLocked() {
 	}
 }
 
-// setNetInfo sets b.hiCache.NetInfo to ni, and passes ni along to the
+// setNetInfo sets b.hostinfo.NetInfo to ni, and passes ni along to the
 // controlclient, if one exists.
 func (b *LocalBackend) setNetInfo(ni *tailcfg.NetInfo) {
 	b.mu.Lock()
 	c := b.c
-	if b.hiCache != nil {
-		b.hiCache.NetInfo = ni.Clone()
+	if b.hostinfo != nil {
+		b.hostinfo.NetInfo = ni.Clone()
 	}
 	b.mu.Unlock()
 

types/logger/logger.go

@@ -127,18 +127,23 @@ func RateLimitedFn(logf Logf, f time.Duration, burst int, maxCache int) Logf {
 // since the last time this identical line was logged.
 func LogOnChange(logf Logf, maxInterval time.Duration, timeNow func() time.Time) Logf {
 	var (
+		mu          sync.Mutex
 		sLastLogged string
 		tLastLogged = timeNow()
 	)
 
 	return func(format string, args ...interface{}) {
 		s := fmt.Sprintf(format, args...)
+
+		mu.Lock()
 		if s == sLastLogged && timeNow().Sub(tLastLogged) < maxInterval {
+			mu.Unlock()
 			return
 		}
-
 		sLastLogged = s
 		tLastLogged = timeNow()
+		mu.Unlock()
+
 		logf(s)
 	}
 

types/logger/logger_test.go

@@ -9,6 +9,7 @@ import (
 	"bytes"
 	"fmt"
 	"log"
+	"sync"
 	"testing"
 	"time"
 )
@@ -117,3 +118,31 @@ func TestArgWriter(t *testing.T) {
 		t.Errorf("got %q; want %q", got, want)
 	}
 }
+
+func TestSynchronization(t *testing.T) {
+	timeNow := testTimer(1 * time.Second)
+	tests := []struct {
+		name string
+		logf Logf
+	}{
+		{"RateLimitedFn", RateLimitedFn(t.Logf, 1*time.Minute, 2, 50)},
+		{"LogOnChange", LogOnChange(t.Logf, 5*time.Second, timeNow)},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var wg sync.WaitGroup
+			wg.Add(2)
+
+			f := func() {
+				tt.logf("1 2 3 4 5")
+				wg.Done()
+			}
+
+			go f()
+			go f()
+
+			wg.Wait()
+		})
+	}
+}