wgengine/magicsock,types/logger: add latency logs for initial peer contacts (#19916)

In order to allow us to measure the performance effects of client-side netmap
caching, both with and without the feature enabled, add logs to record how long
it takes after a client restart or profile switch for the node to establish
contact with peers, relative to the first uncached netmap.

We do this by keeping track of a timestamp when the connection is constructed,
and logging a record for "new" peer contacts that records how long (in
microseconds) it took from the time the peer was recorded as a candidate.  The
message includes whether the contact was via DERP or direct, and whether a
cached netmap was in use at the time.

This builds on and extends the counters from #19699, but here we include new
contacts whether or not a cached netmap is in use, so that we can establish a
baseline for comparison.

Updates #12639
Updates tailscale/projects#27

Change-Id: I4f6d050e221f3881848d05a0425c4a5d1a59294c
Signed-off-by: M. J. Fromberger <fromberger@tailscale.com>

types/logger/logger.go

@@ -141,6 +141,7 @@ var rateFree = []string{
 	"magicsock: disco: ",
 	"magicsock: ParseEndpoint:",
 	// grinder stats lines
+	"magicsock: new contact: ",
 	"SetPrefs: %v",
 	"peer keys: %s",
 	"v%v peers: %v",

wgengine/magicsock/endpoint.go

@@ -139,11 +139,18 @@ func (de *endpoint) setBestAddrLocked(v addrQuality) {
 	if v.epAddr != de.bestAddr.epAddr {
 		de.probeUDPLifetime.resetCycleEndpointLocked()
 
-		// Reaching here, if we are using data from a cached netmap and we are
+		// Reaching here, if we are upgrading from an invalid (missing) address
-		// upgrading from an invalid (missing) address to a valid one, increment
+		// to a valid one, record metrics:
-		// the counter for peers established.
+		if !de.bestAddr.ap.IsValid() && v.ap.IsValid() {
-		if !de.bestAddr.ap.IsValid() && v.ap.IsValid() && de.c.usingCachedNetmap.Load() {
+			// If we are using data from a cached netmap, increment the counter for peers established.
-			metricCachedPeerContactDirect.Add(1)
+			isCached := de.c.usingCachedNetmap.Load()
+			if isCached {
+				metricCachedPeerContactDirect.Add(1)
+			}
+			// Regardless whether the netmap is cached, record how long it has
+			// been since the endpoint was initialized.
+			de.c.logf("magicsock: new contact: peer=%s usec=%d cached=%v via=direct",
+				de.publicKey.ShortString(), int64(mono.Since(de.c.initializedAt)/time.Microsecond), isCached)
 		}
 	}
 	de.bestAddr = v

wgengine/magicsock/magicsock.go

@@ -273,7 +273,7 @@ type Conn struct {
 	// discoAtomic is the current disco private and public keypair for this conn.
 	discoAtomic discoAtomic
 
-	// usingCacheNetmap is whether the latest update to self and peersByID are from a cached network map
+	// usingCacheNetmap is whether the latest update to self and peersByID are from a cached network map.
 	usingCachedNetmap atomic.Bool
 
 	// ============================================================
@@ -417,6 +417,11 @@ type Conn struct {
 	// metrics contains the metrics for the magicsock instance.
 	metrics *metrics
 
+	// initializedAt records the latest time at which the conn was
+	// (re)initialized for latency metrics. This is set at construction, and
+	// updated by netmap resets.
+	initializedAt mono.Time
+
 	// homeDERPGauge is the usermetric gauge for the home DERP region ID.
 	// This can be nil when [Options.Metrics] are not enabled.
 	homeDERPGauge *usermetric.Gauge
@@ -570,13 +575,14 @@ type UDPRelayAllocResp struct {
 func newConn(logf logger.Logf) *Conn {
 	discoPrivate := key.NewDisco()
 	c := &Conn{
-		logf:         logf,
+		logf:          logf,
-		derpRecvCh:   make(chan derpReadResult, 1), // must be buffered, see issue 3736
+		derpRecvCh:    make(chan derpReadResult, 1), // must be buffered, see issue 3736
-		derpStarted:  make(chan struct{}),
+		derpStarted:   make(chan struct{}),
-		peerLastDerp: make(map[key.NodePublic]int),
+		peerLastDerp:  make(map[key.NodePublic]int),
-		peerMap:      newPeerMap(),
+		peerMap:       newPeerMap(),
-		discoInfo:    make(map[key.DiscoPublic]*discoInfo),
+		discoInfo:     make(map[key.DiscoPublic]*discoInfo),
-		cloudInfo:    cloudinfo.New(logf),
+		cloudInfo:     cloudinfo.New(logf),
+		initializedAt: mono.Now(),
 	}
 	c.discoAtomic.Set(discoPrivate)
 	c.bind = &connBind{Conn: c, closed: true}
@@ -2362,9 +2368,16 @@ func (c *Conn) handleDiscoMessage(msg []byte, src epAddr, shouldBeRelayHandshake
 		// Reaching here, if we are using data from a cached network map the
 		// receipt of a CallMeMaybe from a peer indicates we have a sufficiently
 		// viable connection to that peer to count it as active while cached.
-		if c.usingCachedNetmap.Load() {
+		isCached := c.usingCachedNetmap.Load()
+		if isCached {
 			metricCachedPeerContactDERP.Add(1)
 		}
+		// If we did not already have an an endpoint for this peer, even a stale
+		// one, record how long it has been since the endpoint was initialized.
+		if !lastBest.ap.IsValid() {
+			c.logf("magicsock: new contact: peer=%s usec=%d cached=%v via=derp",
+				nodeKey.ShortString(), int64(mono.Since(c.initializedAt)/time.Microsecond), isCached)
+		}
 		if isVia {
 			c.dlogf("[v1] magicsock: disco: %v<-%v via %v (%v, %v)  got call-me-maybe-via, %d endpoints",
 				c.discoAtomic.Short(), epDisco.short, via.ServerDisco.ShortString(),
@@ -3025,7 +3038,11 @@ func (c *Conn) setNetworkMapInternal(self tailcfg.NodeView, peers []tailcfg.Node
 	selfView := c.self
 	peersSnap := c.peerSnapshotLocked()
 	isClosed := c.closed
-	c.usingCachedNetmap.Store(isCached)
+	wasCached := c.usingCachedNetmap.Swap(isCached)
+	if !self.Valid() {
+		c.initializedAt = mono.Now() // the netmap is being reset
+	}
+	initializedAt := c.initializedAt
 	if runtime.GOOS == "linux" && c.controlKnobs != nil {
 		curGRO = c.controlKnobs.DisableUDPGRO.Load()
 		curGSO = c.controlKnobs.DisableUDPGSO.Load()
@@ -3040,6 +3057,12 @@ func (c *Conn) setNetworkMapInternal(self tailcfg.NodeView, peers []tailcfg.Node
 	if isClosed {
 		return // nothing to do here, the conn is closed and the update is no longer relevant
 	}
+	if self.Valid() && wasCached && !isCached {
+		// Reaching here, we have a "real" netmap (self is valid), and we are
+		// transitioning from cached to non-cached, so record the latency from
+		// reset.
+		c.logf("magicsock: new contact: control-netmap usec=%d", int64(mono.Since(initializedAt)/time.Microsecond))
+	}
 
 	if udpOffloadKnobsChanged {
 		// A control-plane node attribute toggled UDP GRO or UDP GSO. Rebind