// TestingSetupZoneConfigHook initializes the zone config hook
// to 'testingZoneConfigHook' which uses 'testingZoneConfig'.
// Settings go back to their previous values when the stopper runs our closer.
func TestingSetupZoneConfigHook(stopper *stop.Stopper) {
stopper.AddCloser(stop.CloserFn(testingResetZoneConfigHook))
testingLock.Lock()
defer testingLock.Unlock()
if testingHasHook {
panic("TestingSetupZoneConfigHook called without restoring state")
}
testingHasHook = true
testingZoneConfig = make(zoneConfigMap)
testingPreviousHook = ZoneConfigHook
ZoneConfigHook = testingZoneConfigHook
testingLargestIDHook = func(maxID uint32) (max uint32) {
testingLock.Lock()
defer testingLock.Unlock()
for id := range testingZoneConfig {
if maxID > 0 && id > maxID {
continue
}
if id > max {
max = id
}
}
return
}
}
// Start starts the TestServer by bootstrapping an in-memory store
// (defaults to maximum of 100M). The server is started, launching the
// node RPC server and all HTTP endpoints. Use the value of
// TestServer.ServingAddr() after Start() for client connections.
// Use TestServer.Stopper().Stop() to shutdown the server after the test
// completes.
func (ts *TestServer) Start(params base.TestServerArgs) error {
if ts.Cfg == nil {
panic("Cfg not set")
}
if params.Stopper == nil {
params.Stopper = stop.NewStopper()
}
if !params.PartOfCluster {
// Change the replication requirements so we don't get log spam about ranges
// not being replicated enough.
cfg := config.DefaultZoneConfig()
cfg.NumReplicas = 1
fn := config.TestingSetDefaultZoneConfig(cfg)
params.Stopper.AddCloser(stop.CloserFn(fn))
}
// Needs to be called before NewServer to ensure resolvers are initialized.
if err := ts.Cfg.InitNode(); err != nil {
return err
}
var err error
ts.Server, err = NewServer(*ts.Cfg, params.Stopper)
if err != nil {
return err
}
// Our context must be shared with our server.
ts.Cfg = &ts.Server.cfg
if err := ts.Server.Start(context.Background()); err != nil {
return err
}
// If enabled, wait for initial splits to complete before returning control.
// If initial splits do not complete, the server is stopped before
// returning.
if stk, ok := ts.cfg.TestingKnobs.Store.(*storage.StoreTestingKnobs); ok &&
stk.DisableSplitQueue {
return nil
}
if err := ts.WaitForInitialSplits(); err != nil {
ts.Stop()
return err
}
return nil
}
// New creates an instance of a gossip node.
// The higher level manages the NodeIDContainer instance (which can be shared by
// various server components). The ambient context is expected to already
// contain the node ID.
func New(
ambient log.AmbientContext,
nodeID *base.NodeIDContainer,
rpcContext *rpc.Context,
grpcServer *grpc.Server,
resolvers []resolver.Resolver,
stopper *stop.Stopper,
registry *metric.Registry,
) *Gossip {
ambient.SetEventLog("gossip", "gossip")
g := &Gossip{
server: newServer(ambient, nodeID, stopper, registry),
Connected: make(chan struct{}),
rpcContext: rpcContext,
outgoing: makeNodeSet(minPeers, metric.NewGauge(MetaConnectionsOutgoingGauge)),
bootstrapping: map[string]struct{}{},
disconnected: make(chan *client, 10),
stalledCh: make(chan struct{}, 1),
stallInterval: defaultStallInterval,
bootstrapInterval: defaultBootstrapInterval,
cullInterval: defaultCullInterval,
nodeDescs: map[roachpb.NodeID]*roachpb.NodeDescriptor{},
resolverAddrs: map[util.UnresolvedAddr]resolver.Resolver{},
bootstrapAddrs: map[util.UnresolvedAddr]roachpb.NodeID{},
}
stopper.AddCloser(stop.CloserFn(g.server.AmbientContext.FinishEventLog))
registry.AddMetric(g.outgoing.gauge)
g.clientsMu.breakers = map[string]*circuit.Breaker{}
resolverAddrs := make([]string, len(resolvers))
for i, resolver := range resolvers {
resolverAddrs[i] = resolver.Addr()
}
ctx := g.AnnotateCtx(context.Background())
if log.V(1) {
log.Infof(ctx, "initial resolvers: %v", resolverAddrs)
}
g.SetResolvers(resolvers)
g.mu.Lock()
// Add ourselves as a SystemConfig watcher.
g.mu.is.registerCallback(KeySystemConfig, g.updateSystemConfig)
// Add ourselves as a node descriptor watcher.
g.mu.is.registerCallback(MakePrefixPattern(KeyNodeIDPrefix), g.updateNodeAddress)
g.mu.Unlock()
RegisterGossipServer(grpcServer, g.server)
return g
}
func (t *parallelTest) getClient(nodeIdx, clientIdx int) *gosql.DB {
for len(t.clients[nodeIdx]) <= clientIdx {
// Add a client.
pgURL, cleanupFunc := sqlutils.PGUrl(t.T,
t.cluster.Server(nodeIdx).ServingAddr(),
"TestParallel",
url.User(security.RootUser))
db, err := gosql.Open("postgres", pgURL.String())
if err != nil {
t.Fatal(err)
}
sqlutils.MakeSQLRunner(t, db).Exec("SET DATABASE = test")
t.cluster.Stopper().AddCloser(
stop.CloserFn(func() {
_ = db.Close()
cleanupFunc()
}))
t.clients[nodeIdx] = append(t.clients[nodeIdx], db)
}
return t.clients[nodeIdx][clientIdx]
}
func testRepairInner(ctx context.Context, t *testing.T, c cluster.Cluster, cfg cluster.TestConfig) {
dc := newDynamicClient(c, stopper)
stopper.AddCloser(stop.CloserFn(func() {
dc.Close(ctx)
}))
// Add some loads.
for i := 0; i < c.NumNodes()*2; i++ {
ID := i
stopper.RunWorker(func() {
insertLoad(ctx, t, dc, ID)
})
}
// TODO(bram): #5345 add repair mechanism.
select {
case <-stopper.ShouldStop():
case <-time.After(cfg.Duration):
}
}
// StartServer creates a test server and sets up a gosql DB connection.
// The server should be stopped by calling server.Stopper().Stop().
func StartServer(
t testing.TB, params base.TestServerArgs,
) (TestServerInterface, *gosql.DB, *client.DB) {
server, err := StartServerRaw(params)
if err != nil {
t.Fatal(err)
}
kvClient := server.KVClient().(*client.DB)
pgURL, cleanupGoDB := sqlutils.PGUrl(
t, server.ServingAddr(), "StartServer", url.User(security.RootUser))
pgURL.Path = params.UseDatabase
goDB, err := gosql.Open("postgres", pgURL.String())
if err != nil {
t.Fatal(err)
}
server.Stopper().AddCloser(
stop.CloserFn(func() {
_ = goDB.Close()
cleanupGoDB()
}))
return server, goDB, kvClient
}
func initBacktrace(logDir string) *stop.Stopper {
const ptracePath = "/opt/backtrace/bin/ptrace"
if _, err := os.Stat(ptracePath); err != nil {
log.Infof(context.TODO(), "backtrace disabled: %s", err)
return stop.NewStopper()
}
if err := bcd.EnableTracing(); err != nil {
log.Infof(context.TODO(), "unable to enable backtrace: %s", err)
return stop.NewStopper()
}
bcd.UpdateConfig(bcd.GlobalConfig{
PanicOnKillFailure: true,
ResendSignal: true,
RateLimit: time.Second * 3,
SynchronousPut: true,
})
// Use the default tracer implementation.
// false: Exclude system goroutines.
tracer := bcd.New(bcd.NewOptions{
IncludeSystemGs: false,
})
if err := tracer.SetOutputPath(logDir, 0755); err != nil {
log.Infof(context.TODO(), "unable to set output path: %s", err)
// Not a fatal error, continue.
}
// Enable WARNING log output from the tracer.
tracer.AddOptions(nil, "-L", "WARNING")
info := build.GetInfo()
tracer.AddKV(nil, "cgo-compiler", info.CgoCompiler)
tracer.AddKV(nil, "go-version", info.GoVersion)
tracer.AddKV(nil, "platform", info.Platform)
tracer.AddKV(nil, "tag", info.Tag)
tracer.AddKV(nil, "time", info.Time)
// Register for traces on signal reception.
tracer.SetSigset(
[]os.Signal{
syscall.SIGABRT,
syscall.SIGFPE,
syscall.SIGSEGV,
syscall.SIGILL,
syscall.SIGBUS}...)
bcd.Register(tracer)
// Hook log.Fatal*.
log.SetExitFunc(func(code int) {
_ = bcd.Trace(tracer, fmt.Errorf("exit %d", code), nil)
os.Exit(code)
})
stopper := stop.NewStopper(stop.OnPanic(func(val interface{}) {
err, ok := val.(error)
if !ok {
err = fmt.Errorf("%v", val)
}
_ = bcd.Trace(tracer, err, nil)
panic(val)
}))
// Internally, backtrace uses an external program (/opt/backtrace/bin/ptrace)
// to generate traces. We direct the stdout for this program to a file for
// debugging our usage of backtrace.
if f, err := os.OpenFile(filepath.Join(logDir, "backtrace.out"),
os.O_RDWR|os.O_CREATE|os.O_APPEND, 0666); err != nil {
log.Infof(context.TODO(), "unable to open: %s", err)
} else {
stopper.AddCloser(stop.CloserFn(func() {
f.Close()
}))
tracer.SetPipes(nil, f)
}
tracer.SetLogLevel(bcd.LogMax)
log.Infof(context.TODO(), "backtrace enabled")
return stopper
}
// StartTestCluster starts up a TestCluster made up of `nodes` in-memory testing
// servers.
// The cluster should be stopped using cluster.Stop().
func StartTestCluster(t testing.TB, nodes int, args base.TestClusterArgs) *TestCluster {
if nodes < 1 {
t.Fatal("invalid cluster size: ", nodes)
}
if args.ServerArgs.JoinAddr != "" {
t.Fatal("can't specify a join addr when starting a cluster")
}
if args.ServerArgs.Stopper != nil {
t.Fatal("can't set individual server stoppers when starting a cluster")
}
storeKnobs := args.ServerArgs.Knobs.Store
if storeKnobs != nil &&
(storeKnobs.(*storage.StoreTestingKnobs).DisableSplitQueue ||
storeKnobs.(*storage.StoreTestingKnobs).DisableReplicateQueue) {
t.Fatal("can't disable an individual server's queues when starting a cluster; " +
"the cluster controls replication")
}
switch args.ReplicationMode {
case base.ReplicationAuto:
case base.ReplicationManual:
if args.ServerArgs.Knobs.Store == nil {
args.ServerArgs.Knobs.Store = &storage.StoreTestingKnobs{}
}
storeKnobs := args.ServerArgs.Knobs.Store.(*storage.StoreTestingKnobs)
storeKnobs.DisableSplitQueue = true
storeKnobs.DisableReplicateQueue = true
default:
t.Fatal("unexpected replication mode")
}
tc := &TestCluster{}
tc.stopper = stop.NewStopper()
for i := 0; i < nodes; i++ {
var serverArgs base.TestServerArgs
if perNodeServerArgs, ok := args.ServerArgsPerNode[i]; ok {
serverArgs = perNodeServerArgs
} else {
serverArgs = args.ServerArgs
}
serverArgs.PartOfCluster = true
if i > 0 {
serverArgs.JoinAddr = tc.Servers[0].ServingAddr()
}
tc.AddServer(t, serverArgs)
}
// Create a closer that will stop the individual server stoppers when the
// cluster stopper is stopped.
tc.stopper.AddCloser(stop.CloserFn(tc.stopServers))
tc.WaitForStores(t, tc.Servers[0].Gossip())
// TODO(peter): We should replace the hardcoded 3 with the default ZoneConfig
// replication factor.
if args.ReplicationMode == base.ReplicationAuto && nodes >= 3 {
if err := tc.waitForFullReplication(); err != nil {
t.Fatal(err)
}
}
return tc
}
// Start starts the server on the specified port, starts gossip and initializes
// the node using the engines from the server's context.
//
// The passed context can be used to trace the server startup. The context
// should represent the general startup operation.
func (s *Server) Start(ctx context.Context) error {
ctx = s.AnnotateCtx(ctx)
startTime := timeutil.Now()
tlsConfig, err := s.cfg.GetServerTLSConfig()
if err != nil {
return err
}
httpServer := netutil.MakeServer(s.stopper, tlsConfig, s)
plainRedirectServer := netutil.MakeServer(s.stopper, tlsConfig, http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
http.Redirect(w, r, "https://"+r.Host+r.RequestURI, http.StatusPermanentRedirect)
}))
// The following code is a specialization of util/net.go's ListenAndServe
// which adds pgwire support. A single port is used to serve all protocols
// (pg, http, h2) via the following construction:
//
// non-TLS case:
// net.Listen -> cmux.New
// |
// - -> pgwire.Match -> pgwire.Server.ServeConn
// - -> cmux.Any -> grpc.(*Server).Serve
//
// TLS case:
// net.Listen -> cmux.New
// |
// - -> pgwire.Match -> pgwire.Server.ServeConn
// - -> cmux.Any -> grpc.(*Server).Serve
//
// Note that the difference between the TLS and non-TLS cases exists due to
// Go's lack of an h2c (HTTP2 Clear Text) implementation. See inline comments
// in util.ListenAndServe for an explanation of how h2c is implemented there
// and here.
ln, err := net.Listen("tcp", s.cfg.Addr)
if err != nil {
return err
}
log.Eventf(ctx, "listening on port %s", s.cfg.Addr)
unresolvedListenAddr, err := officialAddr(s.cfg.Addr, ln.Addr())
if err != nil {
return err
}
s.cfg.Addr = unresolvedListenAddr.String()
unresolvedAdvertAddr, err := officialAddr(s.cfg.AdvertiseAddr, ln.Addr())
if err != nil {
return err
}
s.cfg.AdvertiseAddr = unresolvedAdvertAddr.String()
s.rpcContext.SetLocalInternalServer(s.node)
m := cmux.New(ln)
pgL := m.Match(pgwire.Match)
anyL := m.Match(cmux.Any())
httpLn, err := net.Listen("tcp", s.cfg.HTTPAddr)
if err != nil {
return err
}
unresolvedHTTPAddr, err := officialAddr(s.cfg.HTTPAddr, httpLn.Addr())
if err != nil {
return err
}
s.cfg.HTTPAddr = unresolvedHTTPAddr.String()
workersCtx := s.AnnotateCtx(context.Background())
s.stopper.RunWorker(func() {
<-s.stopper.ShouldQuiesce()
if err := httpLn.Close(); err != nil {
log.Fatal(workersCtx, err)
}
})
if tlsConfig != nil {
httpMux := cmux.New(httpLn)
clearL := httpMux.Match(cmux.HTTP1())
tlsL := httpMux.Match(cmux.Any())
s.stopper.RunWorker(func() {
netutil.FatalIfUnexpected(httpMux.Serve())
})
s.stopper.RunWorker(func() {
netutil.FatalIfUnexpected(plainRedirectServer.Serve(clearL))
})
httpLn = tls.NewListener(tlsL, tlsConfig)
}
s.stopper.RunWorker(func() {
netutil.FatalIfUnexpected(httpServer.Serve(httpLn))
})
s.stopper.RunWorker(func() {
<-s.stopper.ShouldQuiesce()
netutil.FatalIfUnexpected(anyL.Close())
<-s.stopper.ShouldStop()
s.grpc.Stop()
})
s.stopper.RunWorker(func() {
netutil.FatalIfUnexpected(s.grpc.Serve(anyL))
})
s.stopper.RunWorker(func() {
pgCtx := s.pgServer.AmbientCtx.AnnotateCtx(context.Background())
netutil.FatalIfUnexpected(httpServer.ServeWith(s.stopper, pgL, func(conn net.Conn) {
connCtx := log.WithLogTagStr(pgCtx, "client", conn.RemoteAddr().String())
if err := s.pgServer.ServeConn(connCtx, conn); err != nil && !netutil.IsClosedConnection(err) {
// Report the error on this connection's context, so that we
// know which remote client caused the error when looking at
// the logs.
log.Error(connCtx, err)
}
}))
})
if len(s.cfg.SocketFile) != 0 {
// Unix socket enabled: postgres protocol only.
unixLn, err := net.Listen("unix", s.cfg.SocketFile)
if err != nil {
return err
}
s.stopper.RunWorker(func() {
<-s.stopper.ShouldQuiesce()
if err := unixLn.Close(); err != nil {
log.Fatal(workersCtx, err)
}
})
s.stopper.RunWorker(func() {
pgCtx := s.pgServer.AmbientCtx.AnnotateCtx(context.Background())
netutil.FatalIfUnexpected(httpServer.ServeWith(s.stopper, unixLn, func(conn net.Conn) {
connCtx := log.WithLogTagStr(pgCtx, "client", conn.RemoteAddr().String())
if err := s.pgServer.ServeConn(connCtx, conn); err != nil && !netutil.IsClosedConnection(err) {
// Report the error on this connection's context, so that we
// know which remote client caused the error when looking at
// the logs.
log.Error(connCtx, err)
}
}))
})
}
// Enable the debug endpoints first to provide an earlier window
// into what's going on with the node in advance of exporting node
// functionality.
// TODO(marc): when cookie-based authentication exists,
// apply it for all web endpoints.
s.mux.HandleFunc(debugEndpoint, http.HandlerFunc(handleDebug))
s.gossip.Start(unresolvedAdvertAddr)
log.Event(ctx, "started gossip")
s.engines, err = s.cfg.CreateEngines()
if err != nil {
return errors.Wrap(err, "failed to create engines")
}
s.stopper.AddCloser(&s.engines)
// We might have to sleep a bit to protect against this node producing non-
// monotonic timestamps. Before restarting, its clock might have been driven
// by other nodes' fast clocks, but when we restarted, we lost all this
// information. For example, a client might have written a value at a
// timestamp that's in the future of the restarted node's clock, and if we
// don't do something, the same client's read would not return the written
// value. So, we wait up to MaxOffset; we couldn't have served timestamps more
// than MaxOffset in the future (assuming that MaxOffset was not changed, see
// #9733).
//
// As an optimization for tests, we don't sleep if all the stores are brand
// new. In this case, the node will not serve anything anyway until it
// synchronizes with other nodes.
{
anyStoreBootstrapped := false
for _, e := range s.engines {
if _, err := storage.ReadStoreIdent(ctx, e); err != nil {
// NotBootstrappedError is expected.
if _, ok := err.(*storage.NotBootstrappedError); !ok {
return err
}
} else {
anyStoreBootstrapped = true
break
}
}
if anyStoreBootstrapped {
sleepDuration := s.clock.MaxOffset() - timeutil.Since(startTime)
if sleepDuration > 0 {
log.Infof(ctx, "sleeping for %s to guarantee HLC monotonicity", sleepDuration)
time.Sleep(sleepDuration)
}
}
}
// Now that we have a monotonic HLC wrt previous incarnations of the process,
// init all the replicas.
err = s.node.start(
ctx,
unresolvedAdvertAddr,
s.engines,
s.cfg.NodeAttributes,
s.cfg.Locality,
)
if err != nil {
return err
}
log.Event(ctx, "started node")
s.nodeLiveness.StartHeartbeat(ctx, s.stopper)
// We can now add the node registry.
s.recorder.AddNode(s.registry, s.node.Descriptor, s.node.startedAt)
// Begin recording runtime statistics.
s.startSampleEnvironment(s.cfg.MetricsSampleInterval)
// Begin recording time series data collected by the status monitor.
s.tsDB.PollSource(
s.cfg.AmbientCtx, s.recorder, s.cfg.MetricsSampleInterval, ts.Resolution10s, s.stopper,
)
// Begin recording status summaries.
s.node.startWriteSummaries(s.cfg.MetricsSampleInterval)
// Create and start the schema change manager only after a NodeID
// has been assigned.
testingKnobs := &sql.SchemaChangerTestingKnobs{}
if s.cfg.TestingKnobs.SQLSchemaChanger != nil {
testingKnobs = s.cfg.TestingKnobs.SQLSchemaChanger.(*sql.SchemaChangerTestingKnobs)
}
sql.NewSchemaChangeManager(testingKnobs, *s.db, s.gossip, s.leaseMgr).Start(s.stopper)
s.distSQLServer.Start()
log.Infof(ctx, "starting %s server at %s", s.cfg.HTTPRequestScheme(), unresolvedHTTPAddr)
log.Infof(ctx, "starting grpc/postgres server at %s", unresolvedListenAddr)
log.Infof(ctx, "advertising CockroachDB node at %s", unresolvedAdvertAddr)
if len(s.cfg.SocketFile) != 0 {
log.Infof(ctx, "starting postgres server at unix:%s", s.cfg.SocketFile)
}
s.stopper.RunWorker(func() {
netutil.FatalIfUnexpected(m.Serve())
})
log.Event(ctx, "accepting connections")
// Initialize grpc-gateway mux and context.
jsonpb := &protoutil.JSONPb{
EnumsAsInts: true,
EmitDefaults: true,
Indent: " ",
}
protopb := new(protoutil.ProtoPb)
gwMux := gwruntime.NewServeMux(
gwruntime.WithMarshalerOption(gwruntime.MIMEWildcard, jsonpb),
gwruntime.WithMarshalerOption(httputil.JSONContentType, jsonpb),
gwruntime.WithMarshalerOption(httputil.AltJSONContentType, jsonpb),
gwruntime.WithMarshalerOption(httputil.ProtoContentType, protopb),
gwruntime.WithMarshalerOption(httputil.AltProtoContentType, protopb),
)
gwCtx, gwCancel := context.WithCancel(s.AnnotateCtx(context.Background()))
s.stopper.AddCloser(stop.CloserFn(gwCancel))
// Setup HTTP<->gRPC handlers.
conn, err := s.rpcContext.GRPCDial(s.cfg.Addr)
if err != nil {
return errors.Errorf("error constructing grpc-gateway: %s; are your certificates valid?", err)
}
for _, gw := range []grpcGatewayServer{s.admin, s.status, &s.tsServer} {
if err := gw.RegisterGateway(gwCtx, gwMux, conn); err != nil {
return err
}
}
var uiFileSystem http.FileSystem
uiDebug := envutil.EnvOrDefaultBool("COCKROACH_DEBUG_UI", false)
if uiDebug {
uiFileSystem = http.Dir("pkg/ui")
} else {
uiFileSystem = &assetfs.AssetFS{
Asset: ui.Asset,
AssetDir: ui.AssetDir,
AssetInfo: ui.AssetInfo,
}
}
uiFileServer := http.FileServer(uiFileSystem)
s.mux.HandleFunc("/", http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
if r.URL.Path == "/" {
if uiDebug {
r.URL.Path = "debug.html"
} else {
r.URL.Path = "release.html"
}
}
uiFileServer.ServeHTTP(w, r)
}))
// TODO(marc): when cookie-based authentication exists,
// apply it for all web endpoints.
s.mux.Handle(adminPrefix, gwMux)
s.mux.Handle(ts.URLPrefix, gwMux)
s.mux.Handle(statusPrefix, gwMux)
s.mux.Handle("/health", gwMux)
s.mux.Handle(statusVars, http.HandlerFunc(s.status.handleVars))
log.Event(ctx, "added http endpoints")
if err := sdnotify.Ready(); err != nil {
log.Errorf(ctx, "failed to signal readiness using systemd protocol: %s", err)
}
log.Event(ctx, "server ready")
return nil
}
