Backport 1.13: Use our fork of bbolt to improve freelist performance (#…

…24010) (#24904)

changelog/24010.txt

@@ -0,0 +1,3 @@
+```release-note:improvement
+storage/raft: Upgrade to bbolt 1.3.8, along with an extra patch to reduce time scanning large freelist maps.
+```

command/agent/cache/cacheboltdb/bolt.go

@@ -12,10 +12,10 @@ import (
 	"time"
 
 	"github.com/golang/protobuf/proto"
+	bolt "github.com/hashicorp-forge/bbolt"
 	"github.com/hashicorp/go-hclog"
 	wrapping "github.com/hashicorp/go-kms-wrapping/v2"
 	"github.com/hashicorp/go-multierror"
-	bolt "go.etcd.io/bbolt"
 )
 
 const (

command/agent/cache/cacheboltdb/bolt_test.go

@@ -15,11 +15,11 @@ import (
 	"time"
 
 	"github.com/golang/protobuf/proto"
+	bolt "github.com/hashicorp-forge/bbolt"
 	"github.com/hashicorp/go-hclog"
 	"github.com/hashicorp/vault/command/agent/cache/keymanager"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
-	bolt "go.etcd.io/bbolt"
 )
 
 func getTestKeyManager(t *testing.T) keymanager.KeyManager {

go.mod

@@ -62,6 +62,7 @@ require (
 	github.com/google/go-github v17.0.0+incompatible
 	github.com/google/go-metrics-stackdriver v0.2.0
 	github.com/google/tink/go v1.6.1
+	github.com/hashicorp-forge/bbolt v1.3.8-hc3
 	github.com/hashicorp/cap v0.2.1-0.20230221194157-7894fed1633d
 	github.com/hashicorp/consul-template v0.29.5
 	github.com/hashicorp/consul/api v1.17.0

go.sum

@@ -1033,6 +1033,8 @@ github.com/gsterjov/go-libsecret v0.0.0-20161001094733-a6f4afe4910c h1:6rhixN/i8
 github.com/gsterjov/go-libsecret v0.0.0-20161001094733-a6f4afe4910c/go.mod h1:NMPJylDgVpX0MLRlPy15sqSwOFv/U1GZ2m21JhFfek0=
 github.com/hailocab/go-hostpool v0.0.0-20160125115350-e80d13ce29ed h1:5upAirOpQc1Q53c0bnx2ufif5kANL7bfZWcc6VJWJd8=
 github.com/hailocab/go-hostpool v0.0.0-20160125115350-e80d13ce29ed/go.mod h1:tMWxXQ9wFIaZeTI9F+hmhFiGpFmhOHzyShyFUhRm0H4=
+github.com/hashicorp-forge/bbolt v1.3.8-hc3 h1:iTWR3RDPj0TGChAvJ8QjHFcNFWAUVgNQV73IE6gAX4E=
+github.com/hashicorp-forge/bbolt v1.3.8-hc3/go.mod h1:sQBu5UIJ+rcUFU4Fo9rpTHNV935jwmGWS3dQ/MV8810=
 github.com/hashicorp/cap v0.2.1-0.20230221194157-7894fed1633d h1:29noMC2UssBX3F/BUmk0/j4PRUU4QvPTfyeOn3tmcOA=
 github.com/hashicorp/cap v0.2.1-0.20230221194157-7894fed1633d/go.mod h1:dHTmyMIVbzT981XxRoci5G//dfWmd/HhuNiCH6J5+IA=
 github.com/hashicorp/consul-template v0.29.5 h1:tzEo93RqODAX2cgOe/ke8xcpdPdxg5rxl6d22wE3f6c=

physical/raft/fsm.go

@@ -20,6 +20,7 @@ import (
 
 	"github.com/armon/go-metrics"
 	"github.com/golang/protobuf/proto"
+	bolt "github.com/hashicorp-forge/bbolt"
 	log "github.com/hashicorp/go-hclog"
 	"github.com/hashicorp/go-multierror"
 	"github.com/hashicorp/go-raftchunking"
@@ -28,7 +29,6 @@ import (
 	"github.com/hashicorp/vault/sdk/helper/jsonutil"
 	"github.com/hashicorp/vault/sdk/physical"
 	"github.com/hashicorp/vault/sdk/plugin/pb"
-	bolt "go.etcd.io/bbolt"
 )
 
 const (

physical/raft/raft.go

@@ -20,6 +20,7 @@ import (
 
 	"github.com/armon/go-metrics"
 	"github.com/golang/protobuf/proto"
+	bolt "github.com/hashicorp-forge/bbolt"
 	log "github.com/hashicorp/go-hclog"
 	wrapping "github.com/hashicorp/go-kms-wrapping/v2"
 	"github.com/hashicorp/go-raftchunking"
@@ -38,7 +39,7 @@ import (
 	"github.com/hashicorp/vault/vault/cluster"
 	"github.com/hashicorp/vault/vault/seal"
 	"github.com/hashicorp/vault/version"
-	bolt "go.etcd.io/bbolt"
+	etcdbolt "go.etcd.io/bbolt"
 )
 
 const (
@@ -401,7 +402,7 @@ func NewRaftBackend(conf map[string]string, logger log.Logger) (physical.Backend
 
 		// Create the backend raft store for logs and stable storage.
 		dbPath := filepath.Join(path, "raft.db")
-		opts := boltOptions(dbPath)
+		opts := etcdboltOptions(dbPath)
 		raftOptions := raftboltdb.Options{
 			Path:        dbPath,
 			BoltOptions: opts,
@@ -632,7 +633,7 @@ func (b *RaftBackend) CollectMetrics(sink *metricsutil.ClusterMetricSink) {
 		stats = b.raft.Stats()
 	}
 	b.l.RUnlock()
-	b.collectMetricsWithStats(logstoreStats, sink, "logstore")
+	b.collectEtcdBoltMetricsWithStats(logstoreStats, sink, "logstore")
 	b.collectMetricsWithStats(fsmStats, sink, "fsm")
 	labels := []metrics.Label{
 		{
@@ -651,6 +652,29 @@ func (b *RaftBackend) CollectMetrics(sink *metricsutil.ClusterMetricSink) {
 }
 
 func (b *RaftBackend) collectMetricsWithStats(stats bolt.Stats, sink *metricsutil.ClusterMetricSink, database string) {
+	txstats := stats.TxStats
+	labels := []metricsutil.Label{{"database", database}}
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "freelist", "free_pages"}, float32(stats.FreePageN), labels)
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "freelist", "pending_pages"}, float32(stats.PendingPageN), labels)
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "freelist", "allocated_bytes"}, float32(stats.FreeAlloc), labels)
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "freelist", "used_bytes"}, float32(stats.FreelistInuse), labels)
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "transaction", "started_read_transactions"}, float32(stats.TxN), labels)
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "transaction", "currently_open_read_transactions"}, float32(stats.OpenTxN), labels)
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "page", "count"}, float32(txstats.GetPageCount()), labels)
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "page", "bytes_allocated"}, float32(txstats.GetPageAlloc()), labels)
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "cursor", "count"}, float32(txstats.GetCursorCount()), labels)
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "node", "count"}, float32(txstats.GetNodeCount()), labels)
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "node", "dereferences"}, float32(txstats.GetNodeDeref()), labels)
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "rebalance", "count"}, float32(txstats.GetRebalance()), labels)
+	sink.AddSampleWithLabels([]string{"raft_storage", "bolt", "rebalance", "time"}, float32(txstats.GetRebalanceTime().Milliseconds()), labels)
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "split", "count"}, float32(txstats.GetSplit()), labels)
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "spill", "count"}, float32(txstats.GetSpill()), labels)
+	sink.AddSampleWithLabels([]string{"raft_storage", "bolt", "spill", "time"}, float32(txstats.GetSpillTime().Milliseconds()), labels)
+	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "write", "count"}, float32(txstats.GetWrite()), labels)
+	sink.IncrCounterWithLabels([]string{"raft_storage", "bolt", "write", "time"}, float32(txstats.GetWriteTime().Milliseconds()), labels)
+}
+
+func (b *RaftBackend) collectEtcdBoltMetricsWithStats(stats etcdbolt.Stats, sink *metricsutil.ClusterMetricSink, database string) {
 	txstats := stats.TxStats
 	labels := []metricsutil.Label{{"database", database}}
 	sink.SetGaugeWithLabels([]string{"raft_storage", "bolt", "freelist", "free_pages"}, float32(stats.FreePageN), labels)
@@ -1967,3 +1991,39 @@ func boltOptions(path string) *bolt.Options {
 
 	return o
 }
+
+func etcdboltOptions(path string) *etcdbolt.Options {
+	o := &etcdbolt.Options{
+		Timeout:        1 * time.Second,
+		FreelistType:   etcdbolt.FreelistMapType,
+		NoFreelistSync: true,
+		MmapFlags:      getMmapFlags(path),
+	}
+
+	if os.Getenv("VAULT_RAFT_FREELIST_TYPE") == "array" {
+		o.FreelistType = etcdbolt.FreelistArrayType
+	}
+
+	if os.Getenv("VAULT_RAFT_FREELIST_SYNC") != "" {
+		o.NoFreelistSync = false
+	}
+
+	// By default, we want to set InitialMmapSize to 100GB, but only on 64bit platforms.
+	// Otherwise, we set it to whatever the value of VAULT_RAFT_INITIAL_MMAP_SIZE
+	// is, assuming it can be parsed as an int. Bolt itself sets this to 0 by default,
+	// so if users are wanting to turn this off, they can also set it to 0. Setting it
+	// to a negative value is the same as not setting it at all.
+	if os.Getenv("VAULT_RAFT_INITIAL_MMAP_SIZE") == "" {
+		o.InitialMmapSize = initialMmapSize
+	} else {
+		imms, err := strconv.Atoi(os.Getenv("VAULT_RAFT_INITIAL_MMAP_SIZE"))
+
+		// If there's an error here, it means they passed something that's not convertible to
+		// a number. Rather than fail startup, just ignore it.
+		if err == nil && imms > 0 {
+			o.InitialMmapSize = imms
+		}
+	}
+
+	return o
+}

physical/raft/raft_test.go

@@ -21,13 +21,13 @@ import (
 
 	"github.com/go-test/deep"
 	"github.com/golang/protobuf/proto"
+	bolt "github.com/hashicorp-forge/bbolt"
 	"github.com/hashicorp/go-hclog"
 	"github.com/hashicorp/go-secure-stdlib/base62"
 	"github.com/hashicorp/go-uuid"
 	"github.com/hashicorp/raft"
 	"github.com/hashicorp/vault/sdk/helper/jsonutil"
 	"github.com/hashicorp/vault/sdk/physical"
-	bolt "go.etcd.io/bbolt"
 )
 
 func connectPeers(nodes ...*RaftBackend) {

physical/raft/snapshot.go

@@ -18,10 +18,10 @@ import (
 	"time"
 
 	"github.com/golang/protobuf/proto"
+	bolt "github.com/hashicorp-forge/bbolt"
 	log "github.com/hashicorp/go-hclog"
 	"github.com/hashicorp/vault/sdk/plugin/pb"
 	"github.com/rboyer/safeio"
-	bolt "go.etcd.io/bbolt"
 	"go.uber.org/atomic"
 
 	"github.com/hashicorp/raft"