Improve federation sender performance, implement backoff and blacklisting, fix up invites a bit (#1007)

* Improve federation sender performance and behaviour, add backoff

* Tweaks

* Tweaks

* Tweaks

* Take copies of events before passing to destination queues

* Don't accidentally drop queued messages

* Don't take copies again

* Tidy up a bit

* Break out statistics (tracked component-wide), report success and failures from Perform actions

* Fix comment, use atomic add

* Improve logic a bit, don't block on wakeup, move idle check

* Don't retry sucessful invites, don't dispatch sendEvent, sendInvite etc

* Dedupe destinations, fix other bug hopefully

* Dispatch sends again

* Federation sender to ignore invites that are destined locally

* Loopback invite events

* Remodel a bit with channels

* Linter

* Only loopback invite event if we know the room

* We should tell other resident servers about the invite if we know about the room

* Correct invite signing

* Fix invite loopback

* Check HTTP response codes, push new invites to front of queue

* Review comments

federationsender/types/statistics.go

@@ -0,0 +1,122 @@
+package types
+
+import (
+	"math"
+	"sync"
+	"time"
+
+	"github.com/matrix-org/gomatrixserverlib"
+	"go.uber.org/atomic"
+)
+
+const (
+	// How many times should we tolerate consecutive failures before we
+	// just blacklist the host altogether? Bear in mind that the backoff
+	// is exponential, so the max time here to attempt is 2**failures.
+	FailuresUntilBlacklist = 16 // 16 equates to roughly 18 hours.
+)
+
+// Statistics contains information about all of the remote federated
+// hosts that we have interacted with. It is basically a threadsafe
+// wrapper.
+type Statistics struct {
+	servers map[gomatrixserverlib.ServerName]*ServerStatistics
+	mutex   sync.RWMutex
+}
+
+// ForServer returns server statistics for the given server name. If it
+// does not exist, it will create empty statistics and return those.
+func (s *Statistics) ForServer(serverName gomatrixserverlib.ServerName) *ServerStatistics {
+	// If the map hasn't been initialised yet then do that.
+	if s.servers == nil {
+		s.mutex.Lock()
+		s.servers = make(map[gomatrixserverlib.ServerName]*ServerStatistics)
+		s.mutex.Unlock()
+	}
+	// Look up if we have statistics for this server already.
+	s.mutex.RLock()
+	server, found := s.servers[serverName]
+	s.mutex.RUnlock()
+	// If we don't, then make one.
+	if !found {
+		s.mutex.Lock()
+		server = &ServerStatistics{}
+		s.servers[serverName] = server
+		s.mutex.Unlock()
+	}
+	return server
+}
+
+// ServerStatistics contains information about our interactions with a
+// remote federated host, e.g. how many times we were successful, how
+// many times we failed etc. It also manages the backoff time and black-
+// listing a remote host if it remains uncooperative.
+type ServerStatistics struct {
+	blacklisted    atomic.Bool   // is the remote side dead?
+	backoffUntil   atomic.Value  // time.Time to wait until before sending requests
+	failCounter    atomic.Uint32 // how many times have we failed?
+	successCounter atomic.Uint32 // how many times have we succeeded?
+}
+
+// Success updates the server statistics with a new successful
+// attempt, which increases the sent counter and resets the idle and
+// failure counters. If a host was blacklisted at this point then
+// we will unblacklist it.
+func (s *ServerStatistics) Success() {
+	s.successCounter.Add(1)
+	s.failCounter.Store(0)
+	s.blacklisted.Store(false)
+}
+
+// Failure marks a failure and works out when to backoff until. It
+// returns true if the worker should give up altogether because of
+// too many consecutive failures. At this point the host is marked
+// as blacklisted.
+func (s *ServerStatistics) Failure() bool {
+	// Increase the fail counter.
+	failCounter := s.failCounter.Add(1)
+
+	// Check that we haven't failed more times than is acceptable.
+	if failCounter >= FailuresUntilBlacklist {
+		// We've exceeded the maximum amount of times we're willing
+		// to back off, which is probably in the region of hours by
+		// now. Mark the host as blacklisted and tell the caller to
+		// give up.
+		s.blacklisted.Store(true)
+		return true
+	}
+
+	// We're still under the threshold so work out the exponential
+	// backoff based on how many times we have failed already. The
+	// worker goroutine will wait until this time before processing
+	// anything from the queue.
+	backoffSeconds := time.Second * time.Duration(math.Exp2(float64(failCounter)))
+	s.backoffUntil.Store(
+		time.Now().Add(backoffSeconds),
+	)
+	return false
+}
+
+// BackoffDuration returns both a bool stating whether to wait,
+// and then if true, a duration to wait for.
+func (s *ServerStatistics) BackoffDuration() (bool, time.Duration) {
+	backoff, until := false, time.Second
+	if b, ok := s.backoffUntil.Load().(time.Time); ok {
+		if b.After(time.Now()) {
+			backoff, until = true, time.Until(b)
+		}
+	}
+	return backoff, until
+}
+
+// Blacklisted returns true if the server is blacklisted and false
+// otherwise.
+func (s *ServerStatistics) Blacklisted() bool {
+	return s.blacklisted.Load()
+}
+
+// SuccessCount returns the number of successful requests. This is
+// usually useful in constructing transaction IDs.
+func (s *ServerStatistics) SuccessCount() uint32 {
+	return s.successCounter.Load()
+}