Merge pull request #2053 from filecoin-project/feat/worker-reconnect

seal-worker: Auto-restart if miner dies

api/api_common.go

@@ -38,6 +38,8 @@ type Common interface {
 
 	// trigger graceful shutdown
 	Shutdown(context.Context) error
+
+	Closing(context.Context) (<-chan struct{}, error)
 }
 
 // Version provides various build-time information

api/apistruct/struct.go

@@ -50,7 +50,8 @@ type CommonStruct struct {
 		LogList     func(context.Context) ([]string, error)     `perm:"write"`
 		LogSetLevel func(context.Context, string, string) error `perm:"write"`
 
-		Shutdown func(context.Context) error `perm:"admin"`
+		Shutdown func(context.Context) error                    `perm:"admin"`
+		Closing  func(context.Context) (<-chan struct{}, error) `perm:"read"`
 	}
 }
 
@@ -313,6 +314,10 @@ func (c *CommonStruct) Shutdown(ctx context.Context) error {
 	return c.Internal.Shutdown(ctx)
 }
 
+func (c *CommonStruct) Closing(ctx context.Context) (<-chan struct{}, error) {
+	return c.Internal.Closing(ctx)
+}
+
 // FullNodeStruct
 
 func (c *FullNodeStruct) ClientListImports(ctx context.Context) ([]api.Import, error) {

cmd/lotus-seal-worker/main.go

@@ -3,11 +3,14 @@ package main
 import (
 	"context"
 	"encoding/json"
+	"fmt"
 	"io/ioutil"
 	"net"
 	"net/http"
 	"os"
 	"path/filepath"
+	"syscall"
+	"time"
 
 	"github.com/google/uuid"
 	"github.com/gorilla/mux"
@@ -117,11 +120,19 @@ var runCmd = &cli.Command{
 		}
 
 		// Connect to storage-miner
-
-		nodeApi, closer, err := lcli.GetStorageMinerAPI(cctx)
-		if err != nil {
-			return xerrors.Errorf("getting miner api: %w", err)
+		var nodeApi api.StorageMiner
+		var closer func()
+		var err error
+		for {
+			nodeApi, closer, err = lcli.GetStorageMinerAPI(cctx)
+			if err == nil {
+				break
+			}
+			fmt.Printf("\r\x1b[0KConnecting to miner API... (%s)", err)
+			time.Sleep(time.Second)
+			continue
 		}
+
 		defer closer()
 		ctx := lcli.ReqContext(cctx)
 		ctx, cancel := context.WithCancel(ctx)
@@ -136,6 +147,8 @@ var runCmd = &cli.Command{
 		}
 		log.Infof("Remote version %s", v)
 
+		watchMinerConn(ctx, cctx, nodeApi)
+
 		// Check params
 
 		act, err := nodeApi.ActorAddress(ctx)
@@ -317,3 +330,42 @@ var runCmd = &cli.Command{
 		return srv.Serve(nl)
 	},
 }
+
+func watchMinerConn(ctx context.Context, cctx *cli.Context, nodeApi api.StorageMiner) {
+	go func() {
+		closing, err := nodeApi.Closing(ctx)
+		if err != nil {
+			log.Errorf("failed to get remote closing channel: %+v", err)
+		}
+
+		select {
+		case <-closing:
+		case <-ctx.Done():
+		}
+
+		if ctx.Err() != nil {
+			return // graceful shutdown
+		}
+
+		log.Warnf("Connection with miner node lost, restarting")
+
+		exe, err := os.Executable()
+		if err != nil {
+			log.Errorf("getting executable for auto-restart: %+v", err)
+		}
+
+		log.Sync()
+
+		// TODO: there are probably cleaner/more graceful ways to restart,
+		//  but this is good enough for now (FSM can recover from the mess this creates)
+		if err := syscall.Exec(exe, []string{exe, "run",
+			fmt.Sprintf("--address=%s", cctx.String("address")),
+			fmt.Sprintf("--no-local-storage=%t", cctx.Bool("no-local-storage")),
+			fmt.Sprintf("--precommit1=%t", cctx.Bool("precommit1")),
+			fmt.Sprintf("--precommit2=%t", cctx.Bool("precommit2")),
+			fmt.Sprintf("--commit=%t", cctx.Bool("commit")),
+		}, os.Environ()); err != nil {
+			fmt.Println(err)
+		}
+	}()
+}

node/impl/common/common.go

@@ -139,4 +139,8 @@ func (a *CommonAPI) Shutdown(ctx context.Context) error {
 	return nil
 }
 
+func (a *CommonAPI) Closing(ctx context.Context) (<-chan struct{}, error) {
+	return make(chan struct{}), nil // relies on jsonrpc closing
+}
+
 var _ api.Common = &CommonAPI{}