control/controlclient: use lite map request handler to avoid aborting streams

Previously, any change to endpoints or hostinfo (or hostinfo's
netinfo) would result in the long-running map request HTTP stream
being torn down and restarted, losing all compression context along
with it.

This change makes us instead send a lite map request (OmitPeers: true,
Stream: false) that doesn't subscribe to anything, and then the
coordination server knows to not close other streams for that node
when it recives a lite request.

Fixes tailscale/corp#797

Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>

control/controlclient/auto.go

@@ -124,6 +124,7 @@ type Client struct {
 	synced          bool       // true if our netmap is up-to-date
 	hostinfo        *tailcfg.Hostinfo
 	inPollNetMap    bool // true if currently running a PollNetMap
+	inLiteMapUpdate bool // true if a lite (non-streaming) map request is outstanding
 	inSendStatus    int  // number of sendStatus calls currently in progress
 	state           State
 
@@ -201,6 +202,50 @@ func (c *Client) Start() {
 	go c.mapRoutine()
 }
 
+// sendNewMapRequest either sends a new OmitPeers, non-streaming map request
+// (to just send Hostinfo/Netinfo/Endpoints info, while keeping an existing
+// streaming response open), or start a new streaming one if necessary.
+//
+// It should be called whenever there's something new to tell the server.
+func (c *Client) sendNewMapRequest() {
+	c.mu.Lock()
+
+	// If we're not already streaming a netmap, or if we're already stuck
+	// in a lite update, then tear down everything and start a new stream
+	// (which starts by sending a new map request)
+	if !c.inPollNetMap || c.inLiteMapUpdate {
+		c.mu.Unlock()
+		c.cancelMapSafely()
+		return
+	}
+
+	// Otherwise, send a lite update that doesn't keep a
+	// long-running stream response.
+	defer c.mu.Unlock()
+	c.inLiteMapUpdate = true
+	ctx, cancel := context.WithTimeout(c.mapCtx, 10*time.Second)
+	go func() {
+		defer cancel()
+		t0 := time.Now()
+		err := c.direct.SendLiteMapUpdate(ctx)
+		d := time.Since(t0).Round(time.Millisecond)
+		c.mu.Lock()
+		c.inLiteMapUpdate = false
+		c.mu.Unlock()
+		if err == nil {
+			c.logf("[v1] successful lite map update in %v", d)
+			return
+		}
+		if ctx.Err() == nil {
+			c.logf("lite map update after %v: %v", d, err)
+		}
+		// Fall back to restarting the long-polling map
+		// request (the old heavy way) if the lite update
+		// failed for any reason.
+		c.cancelMapSafely()
+	}()
+}
+
 func (c *Client) cancelAuth() {
 	c.mu.Lock()
 	if c.authCancel != nil {
@@ -545,7 +590,7 @@ func (c *Client) SetHostinfo(hi *tailcfg.Hostinfo) {
 	c.logf("Hostinfo: %v", hi)
 
 	// Send new Hostinfo to server
-	c.cancelMapSafely()
+	c.sendNewMapRequest()
 }
 
 func (c *Client) SetNetInfo(ni *tailcfg.NetInfo) {
@@ -553,13 +598,12 @@ func (c *Client) SetNetInfo(ni *tailcfg.NetInfo) {
 		panic("nil NetInfo")
 	}
 	if !c.direct.SetNetInfo(ni) {
-		c.logf("[unexpected] duplicate NetInfo: %v", ni)
 		return
 	}
 	c.logf("NetInfo: %v", ni)
 
 	// Send new Hostinfo (which includes NetInfo) to server
-	c.cancelMapSafely()
+	c.sendNewMapRequest()
 }
 
 func (c *Client) sendStatus(who string, err error, url string, nm *NetworkMap) {
@@ -636,7 +680,7 @@ func (c *Client) Logout() {
 func (c *Client) UpdateEndpoints(localPort uint16, endpoints []string) {
 	changed := c.direct.SetEndpoints(localPort, endpoints)
 	if changed {
-		c.cancelMapSafely()
+		c.sendNewMapRequest()
 	}
 }
 

control/controlclient/direct.go

@@ -521,6 +521,18 @@ func inTest() bool { return flag.Lookup("test.v") != nil }
 // maxPolls is how many network maps to download; common values are 1
 // or -1 (to keep a long-poll query open to the server).
 func (c *Direct) PollNetMap(ctx context.Context, maxPolls int, cb func(*NetworkMap)) error {
+	return c.sendMapRequest(ctx, maxPolls, cb)
+}
+
+// SendLiteMapUpdate makes a /map request to update the server of our latest state,
+// but does not fetch anything. It returns an error if the server did not return a
+// successful 200 OK response.
+func (c *Direct) SendLiteMapUpdate(ctx context.Context) error {
+	return c.sendMapRequest(ctx, 1, nil)
+}
+
+// cb nil means to omit peers.
+func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*NetworkMap)) error {
 	c.mu.Lock()
 	persist := c.persist
 	serverURL := c.serverURL
@@ -555,6 +567,7 @@ func (c *Direct) PollNetMap(ctx context.Context, maxPolls int, cb func(*NetworkM
 		Stream:     allowStream,
 		Hostinfo:   hostinfo,
 		DebugFlags: c.debugFlags,
+		OmitPeers:  cb == nil,
 	}
 	if hostinfo != nil && ipForwardingBroken(hostinfo.RoutableIPs) {
 		old := request.DebugFlags
@@ -607,6 +620,11 @@ func (c *Direct) PollNetMap(ctx context.Context, maxPolls int, cb func(*NetworkM
 	}
 	defer res.Body.Close()
 
+	if cb == nil {
+		io.Copy(ioutil.Discard, res.Body)
+		return nil
+	}
+
 	// If we go more than pollTimeout without hearing from the server,
 	// end the long poll. We should be receiving a keep alive ping
 	// every minute.
@@ -724,6 +742,14 @@ func (c *Direct) PollNetMap(ctx context.Context, maxPolls int, cb func(*NetworkM
 			lastParsedPacketFilter = c.parsePacketFilter(pf)
 		}
 
+		// Get latest localPort. This might've changed if
+		// a lite map update occured meanwhile. This only affects
+		// the end-to-end test.
+		// TODO(bradfitz): remove the NetworkMap.LocalPort field entirely.
+		c.mu.Lock()
+		localPort = c.localPort
+		c.mu.Unlock()
+
 		nm := &NetworkMap{
 			NodeKey:      tailcfg.NodeKey(persist.PrivateNodeKey.Public()),
 			PrivateKey:   persist.PrivateNodeKey,