diff options
Diffstat (limited to 'vendor/google.golang.org/grpc/clientconn.go')
| -rw-r--r-- | vendor/google.golang.org/grpc/clientconn.go | 694 |
1 files changed, 511 insertions, 183 deletions
diff --git a/vendor/google.golang.org/grpc/clientconn.go b/vendor/google.golang.org/grpc/clientconn.go index 3a7614242..bfd7555a8 100644 --- a/vendor/google.golang.org/grpc/clientconn.go +++ b/vendor/google.golang.org/grpc/clientconn.go @@ -24,7 +24,6 @@ import ( "fmt" "math" "net/url" - "reflect" "strings" "sync" "sync/atomic" @@ -38,6 +37,7 @@ import ( "google.golang.org/grpc/internal/backoff" "google.golang.org/grpc/internal/channelz" "google.golang.org/grpc/internal/grpcsync" + "google.golang.org/grpc/internal/pretty" iresolver "google.golang.org/grpc/internal/resolver" "google.golang.org/grpc/internal/transport" "google.golang.org/grpc/keepalive" @@ -69,6 +69,9 @@ var ( errConnDrain = errors.New("grpc: the connection is drained") // errConnClosing indicates that the connection is closing. errConnClosing = errors.New("grpc: the connection is closing") + // errConnIdling indicates the the connection is being closed as the channel + // is moving to an idle mode due to inactivity. + errConnIdling = errors.New("grpc: the connection is closing due to channel idleness") // invalidDefaultServiceConfigErrPrefix is used to prefix the json parsing error for the default // service config. invalidDefaultServiceConfigErrPrefix = "grpc: the provided default service config is invalid" @@ -134,17 +137,29 @@ func (dcs *defaultConfigSelector) SelectConfig(rpcInfo iresolver.RPCInfo) (*ires // e.g. to use dns resolver, a "dns:///" prefix should be applied to the target. func DialContext(ctx context.Context, target string, opts ...DialOption) (conn *ClientConn, err error) { cc := &ClientConn{ - target: target, - csMgr: &connectivityStateManager{}, - conns: make(map[*addrConn]struct{}), - dopts: defaultDialOptions(), - blockingpicker: newPickerWrapper(), - czData: new(channelzData), - firstResolveEvent: grpcsync.NewEvent(), - } + target: target, + csMgr: &connectivityStateManager{}, + conns: make(map[*addrConn]struct{}), + dopts: defaultDialOptions(), + czData: new(channelzData), + } + + // We start the channel off in idle mode, but kick it out of idle at the end + // of this method, instead of waiting for the first RPC. Other gRPC + // implementations do wait for the first RPC to kick the channel out of + // idle. But doing so would be a major behavior change for our users who are + // used to seeing the channel active after Dial. + // + // Taking this approach of kicking it out of idle at the end of this method + // allows us to share the code between channel creation and exiting idle + // mode. This will also make it easy for us to switch to starting the + // channel off in idle, if at all we ever get to do that. + cc.idlenessState = ccIdlenessStateIdle + cc.retryThrottler.Store((*retryThrottler)(nil)) cc.safeConfigSelector.UpdateConfigSelector(&defaultConfigSelector{nil}) cc.ctx, cc.cancel = context.WithCancel(context.Background()) + cc.exitIdleCond = sync.NewCond(&cc.mu) disableGlobalOpts := false for _, opt := range opts { @@ -173,40 +188,11 @@ func DialContext(ctx context.Context, target string, opts ...DialOption) (conn * } }() - pid := cc.dopts.channelzParentID - cc.channelzID = channelz.RegisterChannel(&channelzChannel{cc}, pid, target) - ted := &channelz.TraceEventDesc{ - Desc: "Channel created", - Severity: channelz.CtInfo, - } - if cc.dopts.channelzParentID != nil { - ted.Parent = &channelz.TraceEventDesc{ - Desc: fmt.Sprintf("Nested Channel(id:%d) created", cc.channelzID.Int()), - Severity: channelz.CtInfo, - } - } - channelz.AddTraceEvent(logger, cc.channelzID, 1, ted) - cc.csMgr.channelzID = cc.channelzID + // Register ClientConn with channelz. + cc.channelzRegistration(target) - if cc.dopts.copts.TransportCredentials == nil && cc.dopts.copts.CredsBundle == nil { - return nil, errNoTransportSecurity - } - if cc.dopts.copts.TransportCredentials != nil && cc.dopts.copts.CredsBundle != nil { - return nil, errTransportCredsAndBundle - } - if cc.dopts.copts.CredsBundle != nil && cc.dopts.copts.CredsBundle.TransportCredentials() == nil { - return nil, errNoTransportCredsInBundle - } - transportCreds := cc.dopts.copts.TransportCredentials - if transportCreds == nil { - transportCreds = cc.dopts.copts.CredsBundle.TransportCredentials() - } - if transportCreds.Info().SecurityProtocol == "insecure" { - for _, cd := range cc.dopts.copts.PerRPCCredentials { - if cd.RequireTransportSecurity() { - return nil, errTransportCredentialsMissing - } - } + if err := cc.validateTransportCredentials(); err != nil { + return nil, err } if cc.dopts.defaultServiceConfigRawJSON != nil { @@ -249,15 +235,12 @@ func DialContext(ctx context.Context, target string, opts ...DialOption) (conn * } // Determine the resolver to use. - resolverBuilder, err := cc.parseTargetAndFindResolver() - if err != nil { + if err := cc.parseTargetAndFindResolver(); err != nil { return nil, err } - cc.authority, err = determineAuthority(cc.parsedTarget.Endpoint(), cc.target, cc.dopts) - if err != nil { + if err = cc.determineAuthority(); err != nil { return nil, err } - channelz.Infof(logger, cc.channelzID, "Channel authority set to %q", cc.authority) if cc.dopts.scChan != nil { // Blocking wait for the initial service config. @@ -275,57 +258,224 @@ func DialContext(ctx context.Context, target string, opts ...DialOption) (conn * go cc.scWatcher() } + // This creates the name resolver, load balancer, blocking picker etc. + if err := cc.exitIdleMode(); err != nil { + return nil, err + } + + // Configure idleness support with configured idle timeout or default idle + // timeout duration. Idleness can be explicitly disabled by the user, by + // setting the dial option to 0. + cc.idlenessMgr = newIdlenessManager(cc, cc.dopts.idleTimeout) + + // Return early for non-blocking dials. + if !cc.dopts.block { + return cc, nil + } + + // A blocking dial blocks until the clientConn is ready. + for { + s := cc.GetState() + if s == connectivity.Idle { + cc.Connect() + } + if s == connectivity.Ready { + return cc, nil + } else if cc.dopts.copts.FailOnNonTempDialError && s == connectivity.TransientFailure { + if err = cc.connectionError(); err != nil { + terr, ok := err.(interface { + Temporary() bool + }) + if ok && !terr.Temporary() { + return nil, err + } + } + } + if !cc.WaitForStateChange(ctx, s) { + // ctx got timeout or canceled. + if err = cc.connectionError(); err != nil && cc.dopts.returnLastError { + return nil, err + } + return nil, ctx.Err() + } + } +} + +// addTraceEvent is a helper method to add a trace event on the channel. If the +// channel is a nested one, the same event is also added on the parent channel. +func (cc *ClientConn) addTraceEvent(msg string) { + ted := &channelz.TraceEventDesc{ + Desc: fmt.Sprintf("Channel %s", msg), + Severity: channelz.CtInfo, + } + if cc.dopts.channelzParentID != nil { + ted.Parent = &channelz.TraceEventDesc{ + Desc: fmt.Sprintf("Nested channel(id:%d) %s", cc.channelzID.Int(), msg), + Severity: channelz.CtInfo, + } + } + channelz.AddTraceEvent(logger, cc.channelzID, 0, ted) +} + +// exitIdleMode moves the channel out of idle mode by recreating the name +// resolver and load balancer. +func (cc *ClientConn) exitIdleMode() error { + cc.mu.Lock() + if cc.conns == nil { + cc.mu.Unlock() + return errConnClosing + } + if cc.idlenessState != ccIdlenessStateIdle { + cc.mu.Unlock() + logger.Info("ClientConn asked to exit idle mode when not in idle mode") + return nil + } + + defer func() { + // When Close() and exitIdleMode() race against each other, one of the + // following two can happen: + // - Close() wins the race and runs first. exitIdleMode() runs after, and + // sees that the ClientConn is already closed and hence returns early. + // - exitIdleMode() wins the race and runs first and recreates the balancer + // and releases the lock before recreating the resolver. If Close() runs + // in this window, it will wait for exitIdleMode to complete. + // + // We achieve this synchronization using the below condition variable. + cc.mu.Lock() + cc.idlenessState = ccIdlenessStateActive + cc.exitIdleCond.Signal() + cc.mu.Unlock() + }() + + cc.idlenessState = ccIdlenessStateExitingIdle + exitedIdle := false + if cc.blockingpicker == nil { + cc.blockingpicker = newPickerWrapper() + } else { + cc.blockingpicker.exitIdleMode() + exitedIdle = true + } + var credsClone credentials.TransportCredentials if creds := cc.dopts.copts.TransportCredentials; creds != nil { credsClone = creds.Clone() } - cc.balancerWrapper = newCCBalancerWrapper(cc, balancer.BuildOptions{ - DialCreds: credsClone, - CredsBundle: cc.dopts.copts.CredsBundle, - Dialer: cc.dopts.copts.Dialer, - Authority: cc.authority, - CustomUserAgent: cc.dopts.copts.UserAgent, - ChannelzParentID: cc.channelzID, - Target: cc.parsedTarget, - }) + if cc.balancerWrapper == nil { + cc.balancerWrapper = newCCBalancerWrapper(cc, balancer.BuildOptions{ + DialCreds: credsClone, + CredsBundle: cc.dopts.copts.CredsBundle, + Dialer: cc.dopts.copts.Dialer, + Authority: cc.authority, + CustomUserAgent: cc.dopts.copts.UserAgent, + ChannelzParentID: cc.channelzID, + Target: cc.parsedTarget, + }) + } else { + cc.balancerWrapper.exitIdleMode() + } + cc.firstResolveEvent = grpcsync.NewEvent() + cc.mu.Unlock() - // Build the resolver. - rWrapper, err := newCCResolverWrapper(cc, resolverBuilder) - if err != nil { - return nil, fmt.Errorf("failed to build resolver: %v", err) + // This needs to be called without cc.mu because this builds a new resolver + // which might update state or report error inline which needs to be handled + // by cc.updateResolverState() which also grabs cc.mu. + if err := cc.initResolverWrapper(credsClone); err != nil { + return err + } + + if exitedIdle { + cc.addTraceEvent("exiting idle mode") } + return nil +} + +// enterIdleMode puts the channel in idle mode, and as part of it shuts down the +// name resolver, load balancer and any subchannels. +func (cc *ClientConn) enterIdleMode() error { cc.mu.Lock() - cc.resolverWrapper = rWrapper + if cc.conns == nil { + cc.mu.Unlock() + return ErrClientConnClosing + } + if cc.idlenessState != ccIdlenessStateActive { + logger.Error("ClientConn asked to enter idle mode when not active") + return nil + } + + // cc.conns == nil is a proxy for the ClientConn being closed. So, instead + // of setting it to nil here, we recreate the map. This also means that we + // don't have to do this when exiting idle mode. + conns := cc.conns + cc.conns = make(map[*addrConn]struct{}) + + // TODO: Currently, we close the resolver wrapper upon entering idle mode + // and create a new one upon exiting idle mode. This means that the + // `cc.resolverWrapper` field would be overwritten everytime we exit idle + // mode. While this means that we need to hold `cc.mu` when accessing + // `cc.resolverWrapper`, it makes the code simpler in the wrapper. We should + // try to do the same for the balancer and picker wrappers too. + cc.resolverWrapper.close() + cc.blockingpicker.enterIdleMode() + cc.balancerWrapper.enterIdleMode() + cc.csMgr.updateState(connectivity.Idle) + cc.idlenessState = ccIdlenessStateIdle cc.mu.Unlock() - // A blocking dial blocks until the clientConn is ready. - if cc.dopts.block { - for { - cc.Connect() - s := cc.GetState() - if s == connectivity.Ready { - break - } else if cc.dopts.copts.FailOnNonTempDialError && s == connectivity.TransientFailure { - if err = cc.connectionError(); err != nil { - terr, ok := err.(interface { - Temporary() bool - }) - if ok && !terr.Temporary() { - return nil, err - } - } - } - if !cc.WaitForStateChange(ctx, s) { - // ctx got timeout or canceled. - if err = cc.connectionError(); err != nil && cc.dopts.returnLastError { - return nil, err - } - return nil, ctx.Err() + go func() { + cc.addTraceEvent("entering idle mode") + for ac := range conns { + ac.tearDown(errConnIdling) + } + }() + return nil +} + +// validateTransportCredentials performs a series of checks on the configured +// transport credentials. It returns a non-nil error if any of these conditions +// are met: +// - no transport creds and no creds bundle is configured +// - both transport creds and creds bundle are configured +// - creds bundle is configured, but it lacks a transport credentials +// - insecure transport creds configured alongside call creds that require +// transport level security +// +// If none of the above conditions are met, the configured credentials are +// deemed valid and a nil error is returned. +func (cc *ClientConn) validateTransportCredentials() error { + if cc.dopts.copts.TransportCredentials == nil && cc.dopts.copts.CredsBundle == nil { + return errNoTransportSecurity + } + if cc.dopts.copts.TransportCredentials != nil && cc.dopts.copts.CredsBundle != nil { + return errTransportCredsAndBundle + } + if cc.dopts.copts.CredsBundle != nil && cc.dopts.copts.CredsBundle.TransportCredentials() == nil { + return errNoTransportCredsInBundle + } + transportCreds := cc.dopts.copts.TransportCredentials + if transportCreds == nil { + transportCreds = cc.dopts.copts.CredsBundle.TransportCredentials() + } + if transportCreds.Info().SecurityProtocol == "insecure" { + for _, cd := range cc.dopts.copts.PerRPCCredentials { + if cd.RequireTransportSecurity() { + return errTransportCredentialsMissing } } } + return nil +} - return cc, nil +// channelzRegistration registers the newly created ClientConn with channelz and +// stores the returned identifier in `cc.channelzID` and `cc.csMgr.channelzID`. +// A channelz trace event is emitted for ClientConn creation. If the newly +// created ClientConn is a nested one, i.e a valid parent ClientConn ID is +// specified via a dial option, the trace event is also added to the parent. +// +// Doesn't grab cc.mu as this method is expected to be called only at Dial time. +func (cc *ClientConn) channelzRegistration(target string) { + cc.channelzID = channelz.RegisterChannel(&channelzChannel{cc}, cc.dopts.channelzParentID, target) + cc.addTraceEvent("created") + cc.csMgr.channelzID = cc.channelzID } // chainUnaryClientInterceptors chains all unary client interceptors into one. @@ -471,7 +621,9 @@ type ClientConn struct { authority string // See determineAuthority(). dopts dialOptions // Default and user specified dial options. channelzID *channelz.Identifier // Channelz identifier for the channel. + resolverBuilder resolver.Builder // See parseTargetAndFindResolver(). balancerWrapper *ccBalancerWrapper // Uses gracefulswitch.balancer underneath. + idlenessMgr idlenessManager // The following provide their own synchronization, and therefore don't // require cc.mu to be held to access them. @@ -492,11 +644,31 @@ type ClientConn struct { sc *ServiceConfig // Latest service config received from the resolver. conns map[*addrConn]struct{} // Set to nil on close. mkp keepalive.ClientParameters // May be updated upon receipt of a GoAway. + idlenessState ccIdlenessState // Tracks idleness state of the channel. + exitIdleCond *sync.Cond // Signalled when channel exits idle. lceMu sync.Mutex // protects lastConnectionError lastConnectionError error } +// ccIdlenessState tracks the idleness state of the channel. +// +// Channels start off in `active` and move to `idle` after a period of +// inactivity. When moving back to `active` upon an incoming RPC, they +// transition through `exiting_idle`. This state is useful for synchronization +// with Close(). +// +// This state tracking is mostly for self-protection. The idlenessManager is +// expected to keep track of the state as well, and is expected not to call into +// the ClientConn unnecessarily. +type ccIdlenessState int8 + +const ( + ccIdlenessStateActive ccIdlenessState = iota + ccIdlenessStateIdle + ccIdlenessStateExitingIdle +) + // WaitForStateChange waits until the connectivity.State of ClientConn changes from sourceState or // ctx expires. A true value is returned in former case and false in latter. // @@ -536,7 +708,10 @@ func (cc *ClientConn) GetState() connectivity.State { // Notice: This API is EXPERIMENTAL and may be changed or removed in a later // release. func (cc *ClientConn) Connect() { - cc.balancerWrapper.exitIdle() + cc.exitIdleMode() + // If the ClientConn was not in idle mode, we need to call ExitIdle on the + // LB policy so that connections can be created. + cc.balancerWrapper.exitIdleMode() } func (cc *ClientConn) scWatcher() { @@ -693,6 +868,20 @@ func (cc *ClientConn) handleSubConnStateChange(sc balancer.SubConn, s connectivi cc.balancerWrapper.updateSubConnState(sc, s, err) } +// Makes a copy of the input addresses slice and clears out the balancer +// attributes field. Addresses are passed during subconn creation and address +// update operations. In both cases, we will clear the balancer attributes by +// calling this function, and therefore we will be able to use the Equal method +// provided by the resolver.Address type for comparison. +func copyAddressesWithoutBalancerAttributes(in []resolver.Address) []resolver.Address { + out := make([]resolver.Address, len(in)) + for i := range in { + out[i] = in[i] + out[i].BalancerAttributes = nil + } + return out +} + // newAddrConn creates an addrConn for addrs and adds it to cc.conns. // // Caller needs to make sure len(addrs) > 0. @@ -700,11 +889,12 @@ func (cc *ClientConn) newAddrConn(addrs []resolver.Address, opts balancer.NewSub ac := &addrConn{ state: connectivity.Idle, cc: cc, - addrs: addrs, + addrs: copyAddressesWithoutBalancerAttributes(addrs), scopts: opts, dopts: cc.dopts, czData: new(channelzData), resetBackoff: make(chan struct{}), + stateChan: make(chan struct{}), } ac.ctx, ac.cancel = context.WithCancel(cc.ctx) // Track ac in cc. This needs to be done before any getTransport(...) is called. @@ -798,9 +988,6 @@ func (ac *addrConn) connect() error { ac.mu.Unlock() return nil } - // Update connectivity state within the lock to prevent subsequent or - // concurrent calls from resetting the transport more than once. - ac.updateConnectivityState(connectivity.Connecting, nil) ac.mu.Unlock() ac.resetTransport() @@ -819,58 +1006,63 @@ func equalAddresses(a, b []resolver.Address) bool { return true } -// tryUpdateAddrs tries to update ac.addrs with the new addresses list. -// -// If ac is TransientFailure, it updates ac.addrs and returns true. The updated -// addresses will be picked up by retry in the next iteration after backoff. -// -// If ac is Shutdown or Idle, it updates ac.addrs and returns true. -// -// If the addresses is the same as the old list, it does nothing and returns -// true. -// -// If ac is Connecting, it returns false. The caller should tear down the ac and -// create a new one. Note that the backoff will be reset when this happens. -// -// If ac is Ready, it checks whether current connected address of ac is in the -// new addrs list. -// - If true, it updates ac.addrs and returns true. The ac will keep using -// the existing connection. -// - If false, it does nothing and returns false. -func (ac *addrConn) tryUpdateAddrs(addrs []resolver.Address) bool { +// updateAddrs updates ac.addrs with the new addresses list and handles active +// connections or connection attempts. +func (ac *addrConn) updateAddrs(addrs []resolver.Address) { ac.mu.Lock() - defer ac.mu.Unlock() - channelz.Infof(logger, ac.channelzID, "addrConn: tryUpdateAddrs curAddr: %v, addrs: %v", ac.curAddr, addrs) + channelz.Infof(logger, ac.channelzID, "addrConn: updateAddrs curAddr: %v, addrs: %v", pretty.ToJSON(ac.curAddr), pretty.ToJSON(addrs)) + + addrs = copyAddressesWithoutBalancerAttributes(addrs) + if equalAddresses(ac.addrs, addrs) { + ac.mu.Unlock() + return + } + + ac.addrs = addrs + if ac.state == connectivity.Shutdown || ac.state == connectivity.TransientFailure || ac.state == connectivity.Idle { - ac.addrs = addrs - return true + // We were not connecting, so do nothing but update the addresses. + ac.mu.Unlock() + return } - if equalAddresses(ac.addrs, addrs) { - return true + if ac.state == connectivity.Ready { + // Try to find the connected address. + for _, a := range addrs { + a.ServerName = ac.cc.getServerName(a) + if a.Equal(ac.curAddr) { + // We are connected to a valid address, so do nothing but + // update the addresses. + ac.mu.Unlock() + return + } + } } - if ac.state == connectivity.Connecting { - return false - } + // We are either connected to the wrong address or currently connecting. + // Stop the current iteration and restart. - // ac.state is Ready, try to find the connected address. - var curAddrFound bool - for _, a := range addrs { - a.ServerName = ac.cc.getServerName(a) - if reflect.DeepEqual(ac.curAddr, a) { - curAddrFound = true - break - } + ac.cancel() + ac.ctx, ac.cancel = context.WithCancel(ac.cc.ctx) + + // We have to defer here because GracefulClose => Close => onClose, which + // requires locking ac.mu. + if ac.transport != nil { + defer ac.transport.GracefulClose() + ac.transport = nil } - channelz.Infof(logger, ac.channelzID, "addrConn: tryUpdateAddrs curAddrFound: %v", curAddrFound) - if curAddrFound { - ac.addrs = addrs + + if len(addrs) == 0 { + ac.updateConnectivityState(connectivity.Idle, nil) } - return curAddrFound + ac.mu.Unlock() + + // Since we were connecting/connected, we should start a new connection + // attempt. + go ac.resetTransport() } // getServerName determines the serverName to be used in the connection @@ -1023,39 +1215,40 @@ func (cc *ClientConn) Close() error { cc.mu.Unlock() return ErrClientConnClosing } + + for cc.idlenessState == ccIdlenessStateExitingIdle { + cc.exitIdleCond.Wait() + } + conns := cc.conns cc.conns = nil cc.csMgr.updateState(connectivity.Shutdown) + pWrapper := cc.blockingpicker rWrapper := cc.resolverWrapper - cc.resolverWrapper = nil bWrapper := cc.balancerWrapper + idlenessMgr := cc.idlenessMgr cc.mu.Unlock() // The order of closing matters here since the balancer wrapper assumes the // picker is closed before it is closed. - cc.blockingpicker.close() + if pWrapper != nil { + pWrapper.close() + } if bWrapper != nil { bWrapper.close() } if rWrapper != nil { rWrapper.close() } + if idlenessMgr != nil { + idlenessMgr.close() + } for ac := range conns { ac.tearDown(ErrClientConnClosing) } - ted := &channelz.TraceEventDesc{ - Desc: "Channel deleted", - Severity: channelz.CtInfo, - } - if cc.dopts.channelzParentID != nil { - ted.Parent = &channelz.TraceEventDesc{ - Desc: fmt.Sprintf("Nested channel(id:%d) deleted", cc.channelzID.Int()), - Severity: channelz.CtInfo, - } - } - channelz.AddTraceEvent(logger, cc.channelzID, 0, ted) + cc.addTraceEvent("deleted") // TraceEvent needs to be called before RemoveEntry, as TraceEvent may add // trace reference to the entity being deleted, and thus prevent it from being // deleted right away. @@ -1085,7 +1278,8 @@ type addrConn struct { addrs []resolver.Address // All addresses that the resolver resolved to. // Use updateConnectivityState for updating addrConn's connectivity state. - state connectivity.State + state connectivity.State + stateChan chan struct{} // closed and recreated on every state change. backoffIdx int // Needs to be stateful for resetConnectBackoff. resetBackoff chan struct{} @@ -1099,6 +1293,9 @@ func (ac *addrConn) updateConnectivityState(s connectivity.State, lastErr error) if ac.state == s { return } + // When changing states, reset the state change channel. + close(ac.stateChan) + ac.stateChan = make(chan struct{}) ac.state = s if lastErr == nil { channelz.Infof(logger, ac.channelzID, "Subchannel Connectivity change to %v", s) @@ -1124,7 +1321,8 @@ func (ac *addrConn) adjustParams(r transport.GoAwayReason) { func (ac *addrConn) resetTransport() { ac.mu.Lock() - if ac.state == connectivity.Shutdown { + acCtx := ac.ctx + if acCtx.Err() != nil { ac.mu.Unlock() return } @@ -1152,15 +1350,14 @@ func (ac *addrConn) resetTransport() { ac.updateConnectivityState(connectivity.Connecting, nil) ac.mu.Unlock() - if err := ac.tryAllAddrs(addrs, connectDeadline); err != nil { + if err := ac.tryAllAddrs(acCtx, addrs, connectDeadline); err != nil { ac.cc.resolveNow(resolver.ResolveNowOptions{}) // After exhausting all addresses, the addrConn enters // TRANSIENT_FAILURE. - ac.mu.Lock() - if ac.state == connectivity.Shutdown { - ac.mu.Unlock() + if acCtx.Err() != nil { return } + ac.mu.Lock() ac.updateConnectivityState(connectivity.TransientFailure, err) // Backoff. @@ -1175,13 +1372,13 @@ func (ac *addrConn) resetTransport() { ac.mu.Unlock() case <-b: timer.Stop() - case <-ac.ctx.Done(): + case <-acCtx.Done(): timer.Stop() return } ac.mu.Lock() - if ac.state != connectivity.Shutdown { + if acCtx.Err() == nil { ac.updateConnectivityState(connectivity.Idle, err) } ac.mu.Unlock() @@ -1196,14 +1393,13 @@ func (ac *addrConn) resetTransport() { // tryAllAddrs tries to creates a connection to the addresses, and stop when at // the first successful one. It returns an error if no address was successfully // connected, or updates ac appropriately with the new transport. -func (ac *addrConn) tryAllAddrs(addrs []resolver.Address, connectDeadline time.Time) error { +func (ac *addrConn) tryAllAddrs(ctx context.Context, addrs []resolver.Address, connectDeadline time.Time) error { var firstConnErr error for _, addr := range addrs { - ac.mu.Lock() - if ac.state == connectivity.Shutdown { - ac.mu.Unlock() + if ctx.Err() != nil { return errConnClosing } + ac.mu.Lock() ac.cc.mu.RLock() ac.dopts.copts.KeepaliveParams = ac.cc.mkp @@ -1217,7 +1413,7 @@ func (ac *addrConn) tryAllAddrs(addrs []resolver.Address, connectDeadline time.T channelz.Infof(logger, ac.channelzID, "Subchannel picks a new address %q to connect", addr.Addr) - err := ac.createTransport(addr, copts, connectDeadline) + err := ac.createTransport(ctx, addr, copts, connectDeadline) if err == nil { return nil } @@ -1234,19 +1430,20 @@ func (ac *addrConn) tryAllAddrs(addrs []resolver.Address, connectDeadline time.T // createTransport creates a connection to addr. It returns an error if the // address was not successfully connected, or updates ac appropriately with the // new transport. -func (ac *addrConn) createTransport(addr resolver.Address, copts transport.ConnectOptions, connectDeadline time.Time) error { +func (ac *addrConn) createTransport(ctx context.Context, addr resolver.Address, copts transport.ConnectOptions, connectDeadline time.Time) error { addr.ServerName = ac.cc.getServerName(addr) - hctx, hcancel := context.WithCancel(ac.ctx) + hctx, hcancel := context.WithCancel(ctx) onClose := func(r transport.GoAwayReason) { ac.mu.Lock() defer ac.mu.Unlock() // adjust params based on GoAwayReason ac.adjustParams(r) - if ac.state == connectivity.Shutdown { - // Already shut down. tearDown() already cleared the transport and - // canceled hctx via ac.ctx, and we expected this connection to be - // closed, so do nothing here. + if ctx.Err() != nil { + // Already shut down or connection attempt canceled. tearDown() or + // updateAddrs() already cleared the transport and canceled hctx + // via ac.ctx, and we expected this connection to be closed, so do + // nothing here. return } hcancel() @@ -1265,7 +1462,7 @@ func (ac *addrConn) createTransport(addr resolver.Address, copts transport.Conne ac.updateConnectivityState(connectivity.Idle, nil) } - connectCtx, cancel := context.WithDeadline(ac.ctx, connectDeadline) + connectCtx, cancel := context.WithDeadline(ctx, connectDeadline) defer cancel() copts.ChannelzParentID = ac.channelzID @@ -1282,7 +1479,7 @@ func (ac *addrConn) createTransport(addr resolver.Address, copts transport.Conne ac.mu.Lock() defer ac.mu.Unlock() - if ac.state == connectivity.Shutdown { + if ctx.Err() != nil { // This can happen if the subConn was removed while in `Connecting` // state. tearDown() would have set the state to `Shutdown`, but // would not have closed the transport since ac.transport would not @@ -1294,6 +1491,9 @@ func (ac *addrConn) createTransport(addr resolver.Address, copts transport.Conne // The error we pass to Close() is immaterial since there are no open // streams at this point, so no trailers with error details will be sent // out. We just need to pass a non-nil error. + // + // This can also happen when updateAddrs is called during a connection + // attempt. go newTr.Close(transport.ErrConnClosing) return nil } @@ -1401,6 +1601,29 @@ func (ac *addrConn) getReadyTransport() transport.ClientTransport { return nil } +// getTransport waits until the addrconn is ready and returns the transport. +// If the context expires first, returns an appropriate status. If the +// addrConn is stopped first, returns an Unavailable status error. +func (ac *addrConn) getTransport(ctx context.Context) (transport.ClientTransport, error) { + for ctx.Err() == nil { + ac.mu.Lock() + t, state, sc := ac.transport, ac.state, ac.stateChan + ac.mu.Unlock() + if state == connectivity.Ready { + return t, nil + } + if state == connectivity.Shutdown { + return nil, status.Errorf(codes.Unavailable, "SubConn shutting down") + } + + select { + case <-ctx.Done(): + case <-sc: + } + } + return nil, status.FromContextError(ctx.Err()).Err() +} + // tearDown starts to tear down the addrConn. // // Note that tearDown doesn't remove ac from ac.cc.conns, so the addrConn struct @@ -1552,7 +1775,14 @@ func (cc *ClientConn) connectionError() error { return cc.lastConnectionError } -func (cc *ClientConn) parseTargetAndFindResolver() (resolver.Builder, error) { +// parseTargetAndFindResolver parses the user's dial target and stores the +// parsed target in `cc.parsedTarget`. +// +// The resolver to use is determined based on the scheme in the parsed target +// and the same is stored in `cc.resolverBuilder`. +// +// Doesn't grab cc.mu as this method is expected to be called only at Dial time. +func (cc *ClientConn) parseTargetAndFindResolver() error { channelz.Infof(logger, cc.channelzID, "original dial target is: %q", cc.target) var rb resolver.Builder @@ -1564,7 +1794,8 @@ func (cc *ClientConn) parseTargetAndFindResolver() (resolver.Builder, error) { rb = cc.getResolver(parsedTarget.URL.Scheme) if rb != nil { cc.parsedTarget = parsedTarget - return rb, nil + cc.resolverBuilder = rb + return nil } } @@ -1579,38 +1810,98 @@ func (cc *ClientConn) parseTargetAndFindResolver() (resolver.Builder, error) { parsedTarget, err = parseTarget(canonicalTarget) if err != nil { channelz.Infof(logger, cc.channelzID, "dial target %q parse failed: %v", canonicalTarget, err) - return nil, err + return err } channelz.Infof(logger, cc.channelzID, "parsed dial target is: %+v", parsedTarget) rb = cc.getResolver(parsedTarget.URL.Scheme) if rb == nil { - return nil, fmt.Errorf("could not get resolver for default scheme: %q", parsedTarget.URL.Scheme) + return fmt.Errorf("could not get resolver for default scheme: %q", parsedTarget.URL.Scheme) } cc.parsedTarget = parsedTarget - return rb, nil + cc.resolverBuilder = rb + return nil } // parseTarget uses RFC 3986 semantics to parse the given target into a -// resolver.Target struct containing scheme, authority and url. Query -// params are stripped from the endpoint. +// resolver.Target struct containing url. Query params are stripped from the +// endpoint. func parseTarget(target string) (resolver.Target, error) { u, err := url.Parse(target) if err != nil { return resolver.Target{}, err } - return resolver.Target{ - Scheme: u.Scheme, - Authority: u.Host, - URL: *u, - }, nil + return resolver.Target{URL: *u}, nil +} + +func encodeAuthority(authority string) string { + const upperhex = "0123456789ABCDEF" + + // Return for characters that must be escaped as per + // Valid chars are mentioned here: + // https://datatracker.ietf.org/doc/html/rfc3986#section-3.2 + shouldEscape := func(c byte) bool { + // Alphanum are always allowed. + if 'a' <= c && c <= 'z' || 'A' <= c && c <= 'Z' || '0' <= c && c <= '9' { + return false + } + switch c { + case '-', '_', '.', '~': // Unreserved characters + return false + case '!', '$', '&', '\'', '(', ')', '*', '+', ',', ';', '=': // Subdelim characters + return false + case ':', '[', ']', '@': // Authority related delimeters + return false + } + // Everything else must be escaped. + return true + } + + hexCount := 0 + for i := 0; i < len(authority); i++ { + c := authority[i] + if shouldEscape(c) { + hexCount++ + } + } + + if hexCount == 0 { + return authority + } + + required := len(authority) + 2*hexCount + t := make([]byte, required) + + j := 0 + // This logic is a barebones version of escape in the go net/url library. + for i := 0; i < len(authority); i++ { + switch c := authority[i]; { + case shouldEscape(c): + t[j] = '%' + t[j+1] = upperhex[c>>4] + t[j+2] = upperhex[c&15] + j += 3 + default: + t[j] = authority[i] + j++ + } + } + return string(t) } // Determine channel authority. The order of precedence is as follows: // - user specified authority override using `WithAuthority` dial option // - creds' notion of server name for the authentication handshake // - endpoint from dial target of the form "scheme://[authority]/endpoint" -func determineAuthority(endpoint, target string, dopts dialOptions) (string, error) { +// +// Stores the determined authority in `cc.authority`. +// +// Returns a non-nil error if the authority returned by the transport +// credentials do not match the authority configured through the dial option. +// +// Doesn't grab cc.mu as this method is expected to be called only at Dial time. +func (cc *ClientConn) determineAuthority() error { + dopts := cc.dopts // Historically, we had two options for users to specify the serverName or // authority for a channel. One was through the transport credentials // (either in its constructor, or through the OverrideServerName() method). @@ -1627,25 +1918,62 @@ func determineAuthority(endpoint, target string, dopts dialOptions) (string, err } authorityFromDialOption := dopts.authority if (authorityFromCreds != "" && authorityFromDialOption != "") && authorityFromCreds != authorityFromDialOption { - return "", fmt.Errorf("ClientConn's authority from transport creds %q and dial option %q don't match", authorityFromCreds, authorityFromDialOption) + return fmt.Errorf("ClientConn's authority from transport creds %q and dial option %q don't match", authorityFromCreds, authorityFromDialOption) } + endpoint := cc.parsedTarget.Endpoint() + target := cc.target switch { case authorityFromDialOption != "": - return authorityFromDialOption, nil + cc.authority = authorityFromDialOption case authorityFromCreds != "": - return authorityFromCreds, nil + cc.authority = authorityFromCreds case strings.HasPrefix(target, "unix:") || strings.HasPrefix(target, "unix-abstract:"): // TODO: remove when the unix resolver implements optional interface to // return channel authority. - return "localhost", nil + cc.authority = "localhost" case strings.HasPrefix(endpoint, ":"): - return "localhost" + endpoint, nil + cc.authority = "localhost" + endpoint default: // TODO: Define an optional interface on the resolver builder to return // the channel authority given the user's dial target. For resolvers // which don't implement this interface, we will use the endpoint from // "scheme://authority/endpoint" as the default authority. - return endpoint, nil + // Escape the endpoint to handle use cases where the endpoint + // might not be a valid authority by default. + // For example an endpoint which has multiple paths like + // 'a/b/c', which is not a valid authority by default. + cc.authority = encodeAuthority(endpoint) } + channelz.Infof(logger, cc.channelzID, "Channel authority set to %q", cc.authority) + return nil +} + +// initResolverWrapper creates a ccResolverWrapper, which builds the name +// resolver. This method grabs the lock to assign the newly built resolver +// wrapper to the cc.resolverWrapper field. +func (cc *ClientConn) initResolverWrapper(creds credentials.TransportCredentials) error { + rw, err := newCCResolverWrapper(cc, ccResolverWrapperOpts{ + target: cc.parsedTarget, + builder: cc.resolverBuilder, + bOpts: resolver.BuildOptions{ + DisableServiceConfig: cc.dopts.disableServiceConfig, + DialCreds: creds, + CredsBundle: cc.dopts.copts.CredsBundle, + Dialer: cc.dopts.copts.Dialer, + }, + channelzID: cc.channelzID, + }) + if err != nil { + return fmt.Errorf("failed to build resolver: %v", err) + } + // Resolver implementations may report state update or error inline when + // built (or right after), and this is handled in cc.updateResolverState. + // Also, an error from the resolver might lead to a re-resolution request + // from the balancer, which is handled in resolveNow() where + // `cc.resolverWrapper` is accessed. Hence, we need to hold the lock here. + cc.mu.Lock() + cc.resolverWrapper = rw + cc.mu.Unlock() + return nil } |