Upgrade raft-autopilot and wait for autopilot it to stop when revoking leadership (#9644)

Fixes: 9626

.changelog/9626.txt

@@ -0,0 +1,3 @@
+```release-note:bug
+autopilot: Fixed a bug that would cause snapshot restoration to stop autopilot on the leader.
+```

agent/consul/leader.go

@@ -395,7 +395,9 @@ func (s *Server) revokeLeadership() {
 
 	s.resetConsistentReadReady()
 
-	s.autopilot.Stop()
+	// Stop returns a chan and we want to block until it is closed
+	// which indicates that autopilot is actually stopped.
+	<-s.autopilot.Stop()
 }
 
 // DEPRECATED (ACL-Legacy-Compat) - Remove once old ACL compatibility is removed

agent/consul/snapshot_endpoint_test.go

@@ -13,6 +13,8 @@ import (
 	"github.com/hashicorp/consul/sdk/testutil/retry"
 	"github.com/hashicorp/consul/testrpc"
 	msgpackrpc "github.com/hashicorp/net-rpc-msgpackrpc"
+	autopilot "github.com/hashicorp/raft-autopilot"
+	"github.com/stretchr/testify/require"
 )
 
 // verifySnapshot is a helper that does a snapshot and restore.
@@ -165,6 +167,11 @@ func TestSnapshot(t *testing.T) {
 
 	testrpc.WaitForLeader(t, s1.RPC, "dc1")
 	verifySnapshot(t, s1, "dc1", "")
+
+	// ensure autopilot is still running
+	// https://github.com/hashicorp/consul/issues/9626
+	apstatus, _ := s1.autopilot.IsRunning()
+	require.Equal(t, autopilot.Running, apstatus)
 }
 
 func TestSnapshot_LeaderState(t *testing.T) {

go.mod

@@ -53,7 +53,7 @@ require (
 	github.com/hashicorp/memberlist v0.2.2
 	github.com/hashicorp/net-rpc-msgpackrpc v0.0.0-20151116020338-a14192a58a69
 	github.com/hashicorp/raft v1.2.0
-	github.com/hashicorp/raft-autopilot v0.1.1
+	github.com/hashicorp/raft-autopilot v0.1.2
 	github.com/hashicorp/raft-boltdb v0.0.0-20171010151810-6e5ba93211ea
 	github.com/hashicorp/serf v0.9.5
 	github.com/hashicorp/vault/api v1.0.5-0.20200717191844-f687267c8086

go.sum

@@ -289,8 +289,8 @@ github.com/hashicorp/net-rpc-msgpackrpc v0.0.0-20151116020338-a14192a58a69/go.mo
 github.com/hashicorp/raft v1.1.1/go.mod h1:vPAJM8Asw6u8LxC3eJCUZmRP/E4QmUGE1R7g7k8sG/8=
 github.com/hashicorp/raft v1.2.0 h1:mHzHIrF0S91d3A7RPBvuqkgB4d/7oFJZyvf1Q4m7GA0=
 github.com/hashicorp/raft v1.2.0/go.mod h1:vPAJM8Asw6u8LxC3eJCUZmRP/E4QmUGE1R7g7k8sG/8=
-github.com/hashicorp/raft-autopilot v0.1.1 h1:f8Dv2y1Vq8ttuH2+oh5l87Paj/BINpMm5TBrMLx+qGQ=
-github.com/hashicorp/raft-autopilot v0.1.1/go.mod h1:HUBUSYtpQRVkgjvvoOgsZPvwe6b6FZJ1xXtaftRZvrA=
+github.com/hashicorp/raft-autopilot v0.1.2 h1:yeqdUjWLjVJkBM+mcVxqwxi+w+aHsb9cEON2dz69OCs=
+github.com/hashicorp/raft-autopilot v0.1.2/go.mod h1:Af4jZBwaNOI+tXfIqIdbcAnh/UyyqIMj/pOISIfhArw=
 github.com/hashicorp/raft-boltdb v0.0.0-20171010151810-6e5ba93211ea h1:xykPFhrBAS2J0VBzVa5e80b5ZtYuNQtgXjN40qBZlD4=
 github.com/hashicorp/raft-boltdb v0.0.0-20171010151810-6e5ba93211ea/go.mod h1:pNv7Wc3ycL6F5oOWn+tPGo2gWD4a5X+yp/ntwdKLjRk=
 github.com/hashicorp/serf v0.9.5 h1:EBWvyu9tcRszt3Bxp3KNssBMP1KuHWyO51lz9+786iM=

vendor/github.com/hashicorp/raft-autopilot/autopilot.go

@@ -75,6 +75,31 @@ func WithPromoter(promoter Promoter) Option {
 	}
 }
 
+// ExecutionStatus represents the current status of the autopilot background go routines
+type ExecutionStatus string
+
+const (
+	NotRunning   ExecutionStatus = "not-running"
+	Running      ExecutionStatus = "running"
+	ShuttingDown ExecutionStatus = "shutting-down"
+)
+
+type execInfo struct {
+	// status is the current state of autopilot executation
+	status ExecutionStatus
+
+	// shutdown is a function that can be execute to shutdown a running
+	// autopilot's go routines.
+	shutdown context.CancelFunc
+
+	// done is a chan that will be closed when the running autopilot go
+	// routines have exited. Technically closing it is the very last
+	// thing done in the go routine but at that point enough state has
+	// been cleaned up that we would then allow it to be started
+	// immediately afterward
+	done chan struct{}
+}
+
 // Autopilot is the type to manage a running Raft instance.
 //
 // Each Raft node in the cluster will have a corresponding Autopilot instance but
@@ -132,10 +157,6 @@ type Autopilot struct {
 	// brought up.
 	startTime time.Time
 
-	// running is a simple bool to indicate whether the go routines to actually
-	// execute autopilot are currently running
-	running bool
-
 	// removeDeadCh is used to trigger the running autopilot go routines to
 	// find and remove any dead/failed servers
 	removeDeadCh chan struct{}
@@ -143,20 +164,20 @@ type Autopilot struct {
 	// reconcileCh is used to trigger an immediate round of reconciliation.
 	reconcileCh chan struct{}
 
-	// shutdown is a function that can be execute to shutdown a running
-	// autopilot's go routines.
-	shutdown context.CancelFunc
-	// done is a chan that will be closed when the running autopilot go
-	// routines have exited. Technically closing it is the very last
-	// thing done in the go routine but at that point enough state has
-	// been cleaned up that we would then allow it to be started
-	// immediately afterward
-	done chan struct{}
+	// leaderLock implements a cancellable mutex that will be used to ensure
+	// that only one autopilot go routine is the "leader". The leader is
+	// the go routine that is currently responsible for updating the
+	// autopilot state and performing raft promotions/demotions.
+	leaderLock *mutex
 
-	// runLock is meant to protect all of the fields regarding coordination
-	// of whether the autopilot go routines are running and
-	// starting/stopping them.
-	runLock sync.Mutex
+	// execution is the information about the most recent autopilot execution.
+	// Start will initialize this with the most recent execution and it will
+	// be updated by Stop and by the go routines being executed when they are
+	// finished.
+	execution *execInfo
+
+	// execLock protects access to the execution field
+	execLock sync.Mutex
 }
 
 // New will create a new Autopilot instance utilizing the given Raft and Delegate.
@@ -166,6 +187,7 @@ func New(raft Raft, delegate ApplicationIntegration, options ...Option) *Autopil
 	a := &Autopilot{
 		raft:     raft,
 		delegate: delegate,
+		state:    &State{},
 		promoter: DefaultPromoter(),
 		logger:   hclog.Default().Named("autopilot"),
 		// should this be buffered?
@@ -173,6 +195,7 @@ func New(raft Raft, delegate ApplicationIntegration, options ...Option) *Autopil
 		reconcileInterval: DefaultReconcileInterval,
 		updateInterval:    DefaultUpdateInterval,
 		time:              &runtimeTimeProvider{},
+		leaderLock:        newMutex(),
 	}
 
 	for _, opt := range options {

vendor/github.com/hashicorp/raft-autopilot/go.mod

@@ -7,4 +7,5 @@ require (
 	github.com/hashicorp/raft v1.2.0
 	github.com/stretchr/testify v1.6.1
 	go.uber.org/goleak v1.1.10
+	golang.org/x/sync v0.0.0-20190423024810-112230192c58
 )

vendor/github.com/hashicorp/raft-autopilot/go.sum

@@ -67,6 +67,7 @@ golang.org/x/net v0.0.0-20190311183353-d8887717615a/go.mod h1:t9HGtf8HONx5eT2rtn
 golang.org/x/net v0.0.0-20190620200207-3b0461eec859/go.mod h1:z5CRVTTTmAJ677TzLLGU+0bjPO0LkuOLi4/5GtJWs/s=
 golang.org/x/sync v0.0.0-20181108010431-42b317875d0f h1:Bl/8QSvNqXvPGPGXa2z5xUTmV7VDcZyvRZ+QQXkXTZQ=
 golang.org/x/sync v0.0.0-20181108010431-42b317875d0f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
+golang.org/x/sync v0.0.0-20190423024810-112230192c58 h1:8gQV6CLnAEikrhgkHFbMAEhagSSnXWGV915qUMm9mrU=
 golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190222072716-a9d3bda3a223/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=

vendor/github.com/hashicorp/raft-autopilot/mutex.go

@@ -0,0 +1,35 @@
+/*
+This code was taken from the same implementation in a branch from Consul and then
+had the package updated and the mutex type unexported.
+*/
+package autopilot
+
+import (
+	"context"
+
+	"golang.org/x/sync/semaphore"
+)
+
+type mutex semaphore.Weighted
+
+// New returns a Mutex that is ready for use.
+func newMutex() *mutex {
+	return (*mutex)(semaphore.NewWeighted(1))
+}
+
+func (m *mutex) Lock() {
+	_ = (*semaphore.Weighted)(m).Acquire(context.Background(), 1)
+}
+
+func (m *mutex) Unlock() {
+	(*semaphore.Weighted)(m).Release(1)
+}
+
+// TryLock acquires the mutex, blocking until resources are available or ctx is
+// done. On success, returns nil. On failure, returns ctx.Err() and leaves the
+// semaphore unchanged.
+//
+// If ctx is already done, Acquire may still succeed without blocking.
+func (m *mutex) TryLock(ctx context.Context) error {
+	return (*semaphore.Weighted)(m).Acquire(ctx, 1)
+}

vendor/github.com/hashicorp/raft-autopilot/run.go

@@ -9,50 +9,100 @@ import (
 // When the context passed in is cancelled or the Stop method is called
 // then these routines will exit.
 func (a *Autopilot) Start(ctx context.Context) {
-	a.runLock.Lock()
-	defer a.runLock.Unlock()
+	a.execLock.Lock()
+	defer a.execLock.Unlock()
 
 	// already running so there is nothing to do
-	if a.running {
+	if a.execution != nil && a.execution.status == Running {
 		return
 	}
 
 	ctx, shutdown := context.WithCancel(ctx)
-	a.shutdown = shutdown
 	a.startTime = a.time.Now()
-	a.done = make(chan struct{})
 
-	// While a go routine executed by a.run below will periodically
-	// update the state, we want to go ahead and force updating it now
-	// so that during a leadership transfer we don't report an empty
-	// autopilot state. We put a pretty small timeout on this though
-	// so as to prevent leader establishment from taking too long
-	updateCtx, updateCancel := context.WithTimeout(ctx, time.Second)
-	defer updateCancel()
-	a.updateState(updateCtx)
+	exec := &execInfo{
+		status:   Running,
+		shutdown: shutdown,
+		done:     make(chan struct{}),
+	}
 
-	go a.run(ctx)
-	a.running = true
+	if a.execution == nil || a.execution.status == NotRunning {
+		// In theory with a nil execution or the current execution being in the not
+		// running state, we should be able to immediately gain the leader lock as
+		// nothing else should be running and holding the lock. While true we still
+		// gain the lock to ensure that only one thread may even attempt to be
+		// modifying the autopilot state at once.
+		ctx, cancel := context.WithTimeout(context.Background(), time.Second)
+		defer cancel()
+		if err := a.leaderLock.TryLock(ctx); err == nil {
+			a.updateState(ctx)
+			a.leaderLock.Unlock()
+		}
+	}
+
+	go a.beginExecution(ctx, exec)
+	a.execution = exec
+	return
 }
 
 // Stop will terminate the go routines being executed to perform autopilot.
 func (a *Autopilot) Stop() <-chan struct{} {
-	a.runLock.Lock()
-	defer a.runLock.Unlock()
+	a.execLock.Lock()
+	defer a.execLock.Unlock()
 
 	// Nothing to do
-	if !a.running {
+	if a.execution == nil || a.execution.status == NotRunning {
 		done := make(chan struct{})
 		close(done)
 		return done
 	}
 
-	a.shutdown()
-	return a.done
+	a.execution.shutdown()
+	a.execution.status = ShuttingDown
+	return a.execution.done
 }
 
-func (a *Autopilot) run(ctx context.Context) {
+// IsRunning returns the current execution status of the autopilot
+// go routines as well as a chan which will be closed when the
+// routines are no longer running
+func (a *Autopilot) IsRunning() (ExecutionStatus, <-chan struct{}) {
+	a.execLock.Lock()
+	defer a.execLock.Unlock()
+
+	if a.execution == nil || a.execution.status == NotRunning {
+		done := make(chan struct{})
+		close(done)
+		return NotRunning, done
+	}
+
+	return a.execution.status, a.execution.done
+}
+
+func (a *Autopilot) finishExecution(exec *execInfo) {
+	// need to gain the lock because if this was the active execution
+	// then these values may be read while they are updated.
+	a.execLock.Lock()
+	defer a.execLock.Unlock()
+
+	exec.shutdown = nil
+	exec.status = NotRunning
+	// this should be the final cleanup task as it is what notifies the rest
+	// of the world that we are now done
+	close(exec.done)
+	exec.done = nil
+}
+
+func (a *Autopilot) beginExecution(ctx context.Context, exec *execInfo) {
+	// This will wait for any other go routine to finish executing
+	// before running any code ourselves to prevent any conflicting
+	// activity between the two.
+	if err := a.leaderLock.TryLock(ctx); err != nil {
+		a.finishExecution(exec)
+		return
+	}
+
 	a.logger.Debug("autopilot is now running")
+
 	// autopilot needs to do 3 things
 	//
 	// 1. periodically update the cluster state
@@ -78,14 +128,8 @@ func (a *Autopilot) run(ctx context.Context) {
 
 		a.logger.Debug("autopilot is now stopped")
 
-		a.runLock.Lock()
-		a.shutdown = nil
-		a.running = false
-		// this should be the final cleanup task as it is what notifies the rest
-		// of the world that we are now done
-		close(a.done)
-		a.done = nil
-		a.runLock.Unlock()
+		a.finishExecution(exec)
+		a.leaderLock.Unlock()
 	}()
 
 	reconcileTicker := time.NewTicker(a.reconcileInterval)

vendor/github.com/hashicorp/raft-autopilot/state.go

@@ -30,10 +30,8 @@ type nextStateInputs struct {
 	Now          time.Time
 	StartTime    time.Time
 	Config       *Config
-	State        *State
 	RaftConfig   *raft.Configuration
 	KnownServers map[raft.ServerID]*Server
-	AliveServers map[raft.ServerID]*Server
 	LatestIndex  uint64
 	LastTerm     uint64
 	FetchedStats map[raft.ServerID]*ServerStats
@@ -48,16 +46,10 @@ type nextStateInputs struct {
 // - Current state
 // - Raft Configuration
 // - Known Servers
-// - Latest raft index (gatered right before the remote server stats so that they should
+// - Latest raft index (gathered right before the remote server stats so that they should
 //   be from about the same point in time)
 // - Stats for all non-left servers
 func (a *Autopilot) gatherNextStateInputs(ctx context.Context) (*nextStateInputs, error) {
-	// we are going to hold this lock for the entire function. In theory nothing should
-	// modify the state on any other go routine so this really shouldn't block anything
-	// else. However we want to ensure that the inputs are as consistent as possible.
-	a.stateLock.RLock()
-	defer a.stateLock.RUnlock()
-
 	// there are a lot of inputs to computing the next state so they get put into a
 	// struct so that we don't have to return 8 values.
 	inputs := &nextStateInputs{
@@ -72,12 +64,6 @@ func (a *Autopilot) gatherNextStateInputs(ctx context.Context) (*nextStateInputs
 	}
 	inputs.Config = config
 
-	// retrieve the current state
-	inputs.State = a.state
-	if inputs.State == nil {
-		inputs.State = &State{}
-	}
-
 	// retrieve the raft configuration
 	raftConfig, err := a.getRaftConfiguration()
 	if err != nil {
@@ -125,9 +111,6 @@ func (a *Autopilot) gatherNextStateInputs(ctx context.Context) (*nextStateInputs
 		return nil, ctx.Err()
 	}
 
-	// filter the known servers to have a map of just the alive servers
-	inputs.AliveServers = aliveServers(inputs.KnownServers)
-
 	// we only allow the fetch to take place for up to half the health interval
 	// the next health interval will attempt to fetch the stats again but if
 	// we do not see responses within this time then we can assume they are
@@ -136,7 +119,7 @@ func (a *Autopilot) gatherNextStateInputs(ctx context.Context) (*nextStateInputs
 	fetchCtx, cancel := context.WithDeadline(ctx, d)
 	defer cancel()
 
-	inputs.FetchedStats = a.delegate.FetchServerStats(fetchCtx, inputs.AliveServers)
+	inputs.FetchedStats = a.delegate.FetchServerStats(fetchCtx, aliveServers(inputs.KnownServers))
 
 	// it might be nil but we propagate the ctx.Err just in case our context was
 	// cancelled since the last time we checked.
@@ -239,7 +222,7 @@ func (a *Autopilot) nextServers(inputs *nextStateInputs) map[raft.ServerID]*Serv
 	newServers := make(map[raft.ServerID]*ServerState)
 
 	for _, srv := range inputs.RaftConfig.Servers {
-		state := buildServerState(inputs, srv)
+		state := a.buildServerState(inputs, srv)
 
 		// update any promoter specific information. This isn't done within
 		// buildServerState to keep that function "pure" and not require
@@ -257,7 +240,7 @@ func (a *Autopilot) nextServers(inputs *nextStateInputs) map[raft.ServerID]*Serv
 // buildServerState takes all the nextStateInputs and builds out a ServerState
 // for the given Raft server. This will take into account the raft configuration
 // existing state, application known servers and recently fetched stats.
-func buildServerState(inputs *nextStateInputs, srv raft.Server) ServerState {
+func (a *Autopilot) buildServerState(inputs *nextStateInputs, srv raft.Server) ServerState {
 	// Note that the ordering of operations in this method are very important.
 	// We are building up the ServerState from the least important sources
 	// and overriding them with more up to date values.
@@ -292,23 +275,24 @@ func buildServerState(inputs *nextStateInputs, srv raft.Server) ServerState {
 		state.State = RaftLeader
 	}
 
-	var existingHealth *ServerHealth
+	var previousHealthy *bool
 
+	a.stateLock.RLock()
 	// copy some state from an existing server into the new state - most of this
 	// should be overridden soon but at this point we are just building the base.
-	if existing, found := inputs.State.Servers[srv.ID]; found {
+	if existing, found := a.state.Servers[srv.ID]; found {
 		state.Stats = existing.Stats
 		state.Health = existing.Health
-		existingHealth = &existing.Health
+		previousHealthy = &state.Health.Healthy
 
 		// it is is important to note that the map values we retrieved this from are
 		// stored by value. Therefore we are modifying a copy of what is in the existing
 		// state and not the actual state itself. We want to ensure that the Address
 		// is what Raft will know about.
-		existing.Server.Address = srv.Address
-
 		state.Server = existing.Server
+		state.Server.Address = srv.Address
 	}
+	a.stateLock.RUnlock()
 
 	// pull in the latest information from the applications knowledge of the
 	// server. Mainly we want the NodeStatus & Meta
@@ -317,8 +301,8 @@ func buildServerState(inputs *nextStateInputs, srv raft.Server) ServerState {
 		// map we retrieved this from has a non-pointer type value. We definitely
 		// do not want to modify the current known servers but we do want to ensure
 		// that we do not overwrite the Address
-		known.Address = srv.Address
 		state.Server = *known
+		state.Server.Address = srv.Address
 	} else {
 		// TODO (mkeeler) do we need a None state. In the previous autopilot code
 		// we would have set this to serf.StatusNone
@@ -336,7 +320,7 @@ func buildServerState(inputs *nextStateInputs, srv raft.Server) ServerState {
 	// the health status changes. No need for an else as we previously set
 	// it when we overwrote the whole Health structure when finding a
 	// server in the existing state
-	if existingHealth == nil || existingHealth.Healthy != state.Health.Healthy {
+	if previousHealthy == nil || *previousHealthy != state.Health.Healthy {
 		state.Health.StableSince = inputs.Now
 	}
 

vendor/golang.org/x/sync/semaphore/semaphore.go

@@ -0,0 +1,136 @@
+// Copyright 2017 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package semaphore provides a weighted semaphore implementation.
+package semaphore // import "golang.org/x/sync/semaphore"
+
+import (
+	"container/list"
+	"context"
+	"sync"
+)
+
+type waiter struct {
+	n     int64
+	ready chan<- struct{} // Closed when semaphore acquired.
+}
+
+// NewWeighted creates a new weighted semaphore with the given
+// maximum combined weight for concurrent access.
+func NewWeighted(n int64) *Weighted {
+	w := &Weighted{size: n}
+	return w
+}
+
+// Weighted provides a way to bound concurrent access to a resource.
+// The callers can request access with a given weight.
+type Weighted struct {
+	size    int64
+	cur     int64
+	mu      sync.Mutex
+	waiters list.List
+}
+
+// Acquire acquires the semaphore with a weight of n, blocking until resources
+// are available or ctx is done. On success, returns nil. On failure, returns
+// ctx.Err() and leaves the semaphore unchanged.
+//
+// If ctx is already done, Acquire may still succeed without blocking.
+func (s *Weighted) Acquire(ctx context.Context, n int64) error {
+	s.mu.Lock()
+	if s.size-s.cur >= n && s.waiters.Len() == 0 {
+		s.cur += n
+		s.mu.Unlock()
+		return nil
+	}
+
+	if n > s.size {
+		// Don't make other Acquire calls block on one that's doomed to fail.
+		s.mu.Unlock()
+		<-ctx.Done()
+		return ctx.Err()
+	}
+
+	ready := make(chan struct{})
+	w := waiter{n: n, ready: ready}
+	elem := s.waiters.PushBack(w)
+	s.mu.Unlock()
+
+	select {
+	case <-ctx.Done():
+		err := ctx.Err()
+		s.mu.Lock()
+		select {
+		case <-ready:
+			// Acquired the semaphore after we were canceled.  Rather than trying to
+			// fix up the queue, just pretend we didn't notice the cancelation.
+			err = nil
+		default:
+			isFront := s.waiters.Front() == elem
+			s.waiters.Remove(elem)
+			// If we're at the front and there're extra tokens left, notify other waiters.
+			if isFront && s.size > s.cur {
+				s.notifyWaiters()
+			}
+		}
+		s.mu.Unlock()
+		return err
+
+	case <-ready:
+		return nil
+	}
+}
+
+// TryAcquire acquires the semaphore with a weight of n without blocking.
+// On success, returns true. On failure, returns false and leaves the semaphore unchanged.
+func (s *Weighted) TryAcquire(n int64) bool {
+	s.mu.Lock()
+	success := s.size-s.cur >= n && s.waiters.Len() == 0
+	if success {
+		s.cur += n
+	}
+	s.mu.Unlock()
+	return success
+}
+
+// Release releases the semaphore with a weight of n.
+func (s *Weighted) Release(n int64) {
+	s.mu.Lock()
+	s.cur -= n
+	if s.cur < 0 {
+		s.mu.Unlock()
+		panic("semaphore: released more than held")
+	}
+	s.notifyWaiters()
+	s.mu.Unlock()
+}
+
+func (s *Weighted) notifyWaiters() {
+	for {
+		next := s.waiters.Front()
+		if next == nil {
+			break // No more waiters blocked.
+		}
+
+		w := next.Value.(waiter)
+		if s.size-s.cur < w.n {
+			// Not enough tokens for the next waiter.  We could keep going (to try to
+			// find a waiter with a smaller request), but under load that could cause
+			// starvation for large requests; instead, we leave all remaining waiters
+			// blocked.
+			//
+			// Consider a semaphore used as a read-write lock, with N tokens, N
+			// readers, and one writer.  Each reader can Acquire(1) to obtain a read
+			// lock.  The writer can Acquire(N) to obtain a write lock, excluding all
+			// of the readers.  If we allow the readers to jump ahead in the queue,
+			// the writer will starve — there is always one token available for every
+			// reader.
+			break
+		}
+
+		s.cur += w.n
+		s.waiters.Remove(next)
+		close(w.ready)
+	}
+}

vendor/modules.txt

@@ -280,7 +280,7 @@ github.com/hashicorp/memberlist
 github.com/hashicorp/net-rpc-msgpackrpc
 # github.com/hashicorp/raft v1.2.0
 github.com/hashicorp/raft
-# github.com/hashicorp/raft-autopilot v0.1.1
+# github.com/hashicorp/raft-autopilot v0.1.2
 github.com/hashicorp/raft-autopilot
 # github.com/hashicorp/raft-boltdb v0.0.0-20171010151810-6e5ba93211ea
 github.com/hashicorp/raft-boltdb
@@ -507,6 +507,7 @@ golang.org/x/oauth2/jws
 golang.org/x/oauth2/jwt
 # golang.org/x/sync v0.0.0-20200317015054-43a5402ce75a
 golang.org/x/sync/errgroup
+golang.org/x/sync/semaphore
 golang.org/x/sync/singleflight
 # golang.org/x/sys v0.0.0-20201024232916-9f70ab9862d5
 golang.org/x/sys/cpu