implement ClusterMembershipStates for corrosion admin client

internal/corrosion/admin.go

@@ -1,10 +1,14 @@
 package corrosion
 
 import (
+	"encoding/json"
+	"errors"
 	"fmt"
 	"gvisor.dev/gvisor/pkg/binary"
 	"io"
 	"net"
+	"net/netip"
+	"time"
 )
 
 // AdminClient is a client for the Corrosion admin API.
@@ -24,28 +28,65 @@ func (c *AdminClient) Close() error {
 	return c.conn.Close()
 }
 
-func (c *AdminClient) SendCommand(cmd []byte) (<-chan []byte, error) {
+type Response struct {
+	JSON map[string]any
+	// Err is set if the response is an error or if an error occurred while processing the response.
+	Err error
+}
+
+// SendCommand sends a command to the Corrosion admin API and returns a channel that will receive responses.
+// The channel will be closed after sending the last or error response. The caller must read from the channel until
+// it is closed.
+func (c *AdminClient) SendCommand(cmd []byte) (<-chan Response, error) {
 	if _, err := c.conn.Write(encodeFrame(cmd)); err != nil {
 		return nil, fmt.Errorf("send command: %w", err)
 	}
 
-	// Read the first frame immediately to check for errors.
-	data, err := c.readFrame()
-	if err != nil {
-		return nil, fmt.Errorf("read response: %w", err)
-	}
-
-	ch := make(chan []byte)
+	ch := make(chan Response)
 	go func() {
 		defer close(ch)
-		ch <- data
 
 		for {
-			data, err = c.readFrame()
+			r := Response{}
+
+			data, err := c.readFrame()
 			if err != nil {
+				r.Err = err
+				ch <- r
 				return
 			}
-			ch <- data
+
+			var decoded any
+			if err = json.Unmarshal(data, &decoded); err != nil {
+				r.Err = fmt.Errorf("unmarshal response: %w", err)
+				ch <- r
+				// TODO: should we drain the connection here?
+				return
+			}
+
+			switch v := decoded.(type) {
+			case string:
+				if v == "Success" {
+					return
+				}
+				// Ignore other strings.
+			case map[string]any:
+				if errData, ok := v["Error"].(map[string]any); ok {
+					if errMsg, ok := errData["msg"].(string); ok {
+						r.Err = errors.New(errMsg)
+					} else {
+						r.Err = fmt.Errorf("invalid error response: %v", errData)
+					}
+					ch <- r
+					return
+				} else if jsonData, ok := v["Json"].(map[string]any); ok {
+					r.JSON = jsonData
+					ch <- r
+				}
+				// Ignore other maps.
+			default:
+				// Ignore other types.
+			}
 		}
 	}()
 
@@ -78,6 +119,141 @@ func (c *AdminClient) readFrame() ([]byte, error) {
 	return data, nil
 }
 
-func (c *AdminClient) ClusterMembershipStates() (<-chan []byte, error) {
-	return c.SendCommand([]byte("{\"Cluster\":\"MembershipStates\"}"))
+type ClusterMembershipState struct {
+	ID        string
+	Addr      netip.AddrPort
+	State     string
+	Timestamp time.Time
+}
+
+var (
+	// MembershipStateAlive indicates that the member is active.
+	MembershipStateAlive string = "Alive"
+	// MembershipStateSuspect indicates that the member is active, but at least one cluster member suspects its down.
+	// For all purposes, a Suspect member is treated as if it were Alive until either it refutes the suspicion
+	// (becoming Alive) or fails to do so (being declared Down).
+	MembershipStateSuspect string = "Suspect"
+	// MembershipStateDown indicates that the member is confirmed Down. A member that reaches this state can't join
+	// the cluster with the same identity until the cluster forgets this knowledge.
+	MembershipStateDown string = "Down"
+)
+
+func parseClusterMembershipState(json map[string]any) (ClusterMembershipState, error) {
+	// Example JSON:
+	//  {
+	//    "id": {
+	//      "addr": "[fdcc:1d51:6bae:6bb2:53c0:8796:1be0:b783]:51001",
+	//      "cluster_id": 0,
+	//      "id": "10d69d6f-6578-4dcf-a285-e860e40c5f06",
+	//      "ts": 7435936225798880256
+	//    },
+	//    "incarnation": 0,
+	//    "state": "Down"
+	//  }
+
+	var state ClusterMembershipState
+	var err error
+
+	idObj, ok := json["id"].(map[string]any)
+	if !ok {
+		return state, fmt.Errorf("missing or invalid 'id' field")
+	}
+
+	// ID
+	if id, ok := idObj["id"].(string); ok {
+		state.ID = id
+	} else {
+		return state, fmt.Errorf("missing or invalid 'id' field")
+	}
+
+	// Addr
+	if addr, ok := idObj["addr"].(string); ok {
+		state.Addr, err = netip.ParseAddrPort(addr)
+		if err != nil {
+			return state, fmt.Errorf("parse 'addr' field: %w", err)
+		}
+	} else {
+		return state, fmt.Errorf("missing or invalid 'addr' field")
+	}
+
+	// State
+	if stateStr, ok := json["state"].(string); ok {
+		switch stateStr {
+		case MembershipStateAlive, MembershipStateSuspect, MembershipStateDown:
+			state.State = stateStr
+		default:
+			return state, fmt.Errorf("invalid 'state' field: %s", stateStr)
+		}
+	} else {
+		return state, fmt.Errorf("missing or invalid 'state' field")
+	}
+
+	// Timestamp
+	if ts, ok := idObj["ts"].(float64); ok {
+		state.Timestamp = ntp64ToTime(uint64(ts))
+	} else {
+		return state, fmt.Errorf("missing or invalid 'ts' field")
+	}
+
+	return state, nil
+}
+
+// nt64ToTime converts a 64-bit NTP timestamp relative to the Unix epoch (1st Jan 1970) to time.Time.
+// See for more details: https://datatracker.ietf.org/doc/html/rfc5905#section-6
+func ntp64ToTime(ntp uint64) time.Time {
+	// The NTP timestamp returned from Corrosion is relative to the Unix epoch (1st Jan 1970)
+	// so no need to subtract the 70 years offset.
+	secs := uint32(ntp >> 32)
+	frac := uint32(ntp)
+	// Convert the fraction to nanoseconds: frac * 1e9 / 2^32
+	nsecs := (uint64(frac) * 1000_000_000) >> 32
+
+	return time.Unix(int64(secs), int64(nsecs))
+}
+
+// ClusterMembershipStates returns the current membership SWIM states of all cluster members.
+// If latest is true, only the latest state of each member is returned.
+func (c *AdminClient) ClusterMembershipStates(latest bool) ([]ClusterMembershipState, error) {
+	respCh, err := c.SendCommand([]byte("{\"Cluster\":\"MembershipStates\"}"))
+	if err != nil {
+		return nil, err
+	}
+
+	var (
+		states       []ClusterMembershipState
+		latestStates map[string]ClusterMembershipState
+		parseErr     error
+	)
+	if latest {
+		latestStates = make(map[string]ClusterMembershipState)
+	}
+
+	for r := range respCh {
+		if r.Err != nil {
+			// It's safe to return here because the channel is closed after the first error response.
+			return nil, r.Err
+		}
+
+		s, err := parseClusterMembershipState(r.JSON)
+		if err != nil {
+			// Do not return early to drain the channel.
+			parseErr = errors.Join(parseErr, err)
+		} else {
+			if latest {
+				if existing, ok := latestStates[s.ID]; !ok || existing.Timestamp.Before(s.Timestamp) {
+					latestStates[s.ID] = s
+				}
+			} else {
+				states = append(states, s)
+			}
+		}
+	}
+
+	if latest {
+		states = make([]ClusterMembershipState, 0, len(latestStates))
+		for _, s := range latestStates {
+			states = append(states, s)
+		}
+	}
+	return states, parseErr
 }