isle/go/daemon/network/garage.go

package network

import (
	"bytes"
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"isle/bootstrap"
	"isle/daemon/daecommon"
	"isle/garage"
	"isle/nebula"
	"isle/secrets"
	"net"
	"net/netip"
	"path/filepath"
	"slices"
	"strconv"
	"time"

	"dev.mediocregopher.com/mediocre-go-lib.git/mctx"
	"dev.mediocregopher.com/mediocre-go-lib.git/mlog"
	"github.com/minio/minio-go/v7"
	"golang.org/x/exp/maps"
)

// Paths within garage's global bucket.
const (
	garageGlobalBucketBootstrapHostsDirPath = "bootstrap/hosts"
)

func getGarageClientParams(
	ctx context.Context,
	secretsStore secrets.Store,
	currBootstrap bootstrap.Bootstrap,
) (
	GarageClientParams, error,
) {
	creds, err := daecommon.GetGarageS3APIGlobalBucketCredentials(
		ctx, secretsStore,
	)
	if err != nil {
		return GarageClientParams{}, fmt.Errorf("getting garage global bucket creds: %w", err)
	}

	rpcSecret, err := daecommon.GetGarageRPCSecret(ctx, secretsStore)
	if err != nil && !errors.Is(err, secrets.ErrNotFound) {
		return GarageClientParams{}, fmt.Errorf("getting garage rpc secret: %w", err)
	}

	return GarageClientParams{
		Node:                         currBootstrap.ChooseGarageNode(),
		GlobalBucketS3APICredentials: creds,
		RPCSecret:                    rpcSecret,
	}, nil
}

func garageAdminClientLogger(logger *mlog.Logger) *mlog.Logger {
	return logger.WithNamespace("garageAdminClient")
}

// newGarageAdminClient will return an AdminClient for a local garage instance,
// or it will _panic_ if there is no local instance configured.
func newGarageAdminClient(
	logger *mlog.Logger,
	networkConfig daecommon.NetworkConfig,
	adminToken string,
	host bootstrap.Host,
) *garage.AdminClient {
	return garage.NewAdminClient(
		garageAdminClientLogger(logger),
		net.JoinHostPort(
			host.IP().String(),
			strconv.Itoa(networkConfig.Storage.Allocations[0].AdminPort),
		),
		adminToken,
	)
}

func garageApplyLayout(
	ctx context.Context,
	logger *mlog.Logger,
	networkConfig daecommon.NetworkConfig,
	adminToken string,
	prevHost, currHost bootstrap.Host,
) error {

	var (
		adminClient = newGarageAdminClient(
			logger, networkConfig, adminToken, currHost,
		)
		hostName = currHost.Name
		allocs   = networkConfig.Storage.Allocations
		roles    = make([]garage.Role, len(allocs))
		roleIDs  = map[string]struct{}{}

		idsToRemove = make([]string, 0, len(prevHost.Garage.Instances))
	)

	defer adminClient.Close()

	for i, alloc := range allocs {
		id := daecommon.BootstrapGarageHostForAlloc(currHost, alloc).ID
		roleIDs[id] = struct{}{}

		zone := string(hostName)
		if alloc.Zone != "" {
			zone = alloc.Zone
		}

		roles[i] = garage.Role{
			ID:       id,
			Capacity: alloc.Capacity * 1_000_000_000,
			Zone:     zone,
			Tags:     []string{},
		}
	}

	for _, prevInst := range prevHost.Garage.Instances {
		if _, ok := roleIDs[prevInst.ID]; !ok {
			idsToRemove = append(idsToRemove, prevInst.ID)
		}
	}

	return adminClient.ApplyLayout(ctx, roles, idsToRemove)
}

func garageInitializeGlobalBucket(
	ctx context.Context,
	logger *mlog.Logger,
	networkConfig daecommon.NetworkConfig,
	adminToken string,
	host bootstrap.Host,
) (
	garage.S3APICredentials, error,
) {
	adminClient := newGarageAdminClient(logger, networkConfig, adminToken, host)
	defer adminClient.Close()

	creds, err := adminClient.CreateS3APICredentials(
		ctx, garage.GlobalBucketS3APICredentialsName,
	)
	if err != nil {
		return creds, fmt.Errorf("creating global bucket credentials: %w", err)
	}

	bucketID, err := adminClient.CreateBucket(ctx, garage.GlobalBucket)
	if err != nil {
		return creds, fmt.Errorf("creating global bucket: %w", err)
	}

	if err := adminClient.GrantBucketPermissions(
		ctx,
		bucketID,
		creds.ID,
		garage.BucketPermissionRead,
		garage.BucketPermissionWrite,
	); err != nil {
		return creds, fmt.Errorf(
			"granting permissions to shared global bucket key: %w", err,
		)
	}

	return creds, nil
}

func getGarageBootstrapHosts(
	ctx context.Context,
	logger *mlog.Logger,
	secretsStore secrets.Store,
	currBootstrap bootstrap.Bootstrap,
) (
	map[nebula.HostName]bootstrap.Host, error,
) {
	garageClientParams, err := getGarageClientParams(
		ctx, secretsStore, currBootstrap,
	)
	if err != nil {
		return nil, fmt.Errorf("getting garage client params: %w", err)
	}

	var (
		client = garageClientParams.GlobalBucketS3APIClient()
		hosts  = map[nebula.HostName]bootstrap.Host{}

		objInfoCh = client.ListObjects(
			ctx, garage.GlobalBucket,
			minio.ListObjectsOptions{
				Prefix:    garageGlobalBucketBootstrapHostsDirPath,
				Recursive: true,
			},
		)
	)

	defer client.Close()

	for objInfo := range objInfoCh {

		ctx := mctx.Annotate(ctx, "objectKey", objInfo.Key)

		if objInfo.Err != nil {
			return nil, fmt.Errorf("listing objects: %w", objInfo.Err)
		}

		obj, err := client.GetObject(
			ctx, garage.GlobalBucket, objInfo.Key, minio.GetObjectOptions{},
		)

		if err != nil {
			return nil, fmt.Errorf("retrieving object %q: %w", objInfo.Key, err)
		}

		var authedHost bootstrap.AuthenticatedHost

		err = json.NewDecoder(obj).Decode(&authedHost)
		obj.Close()

		if err != nil {
			logger.Warn(ctx, "Object contains invalid json", err)
			continue
		}

		host, err := authedHost.Unwrap(currBootstrap.CAPublicCredentials)
		if err != nil {
			logger.Warn(ctx, "Host could not be authenticated", err)
		}

		hosts[host.Name] = host
	}

	return hosts, nil
}

// putGarageBoostrapHost places the <hostname>.json.signed file for this host
// into garage so that other hosts are able to see relevant configuration for
// it.
func putGarageBoostrapHost(
	ctx context.Context,
	secretsStore secrets.Store,
	currBootstrap bootstrap.Bootstrap,
) error {
	garageClientParams, err := getGarageClientParams(
		ctx, secretsStore, currBootstrap,
	)
	if err != nil {
		return fmt.Errorf("getting garage client params: %w", err)
	}

	var (
		host   = currBootstrap.ThisHost()
		client = garageClientParams.GlobalBucketS3APIClient()
	)

	defer client.Close()

	configured, err := nebula.Sign(
		host.HostConfigured, currBootstrap.PrivateCredentials.SigningPrivateKey,
	)
	if err != nil {
		return fmt.Errorf("signing host configured data: %w", err)
	}

	hostB, err := json.Marshal(bootstrap.AuthenticatedHost{
		Assigned:   currBootstrap.SignedHostAssigned,
		Configured: configured,
	})
	if err != nil {
		return fmt.Errorf("encoding host data: %w", err)
	}

	filePath := filepath.Join(
		garageGlobalBucketBootstrapHostsDirPath,
		string(host.Name)+".json.signed",
	)

	_, err = client.PutObject(
		ctx,
		garage.GlobalBucket,
		filePath,
		bytes.NewReader(hostB),
		int64(len(hostB)),
		minio.PutObjectOptions{},
	)

	if err != nil {
		return fmt.Errorf("writing to %q in global bucket: %w", filePath, err)
	}

	return nil
}

func removeGarageBootstrapHost(
	ctx context.Context, client *garage.S3APIClient, hostName nebula.HostName,
) error {

	filePath := filepath.Join(
		garageGlobalBucketBootstrapHostsDirPath,
		string(hostName)+".json.signed",
	)

	return client.RemoveObject(
		ctx, garage.GlobalBucket, filePath, minio.RemoveObjectOptions{},
	)
}

// We can wait for the garage instance to appear healthy, but there are cases
// where they still haven't fully synced the list of buckets and bucket
// credentials. For those cases it's necessary to do this as an additional
// check.
func garageWaitForAlloc(
	ctx context.Context,
	logger *mlog.Logger,
	alloc daecommon.ConfigStorageAllocation,
	adminToken string,
	host bootstrap.Host,
) error {
	var (
		hostIP      = host.IP().String()
		adminClient = garage.NewAdminClient(
			garageAdminClientLogger(logger),
			net.JoinHostPort(hostIP, strconv.Itoa(alloc.AdminPort)),
			adminToken,
		)
	)

	defer adminClient.Close()
	ctx = mctx.WithAnnotator(ctx, alloc)

	for {
		logger.Info(ctx, "Checking if node has synced bucket list")
		buckets, err := adminClient.ListBuckets(ctx)
		if err != nil {
			return fmt.Errorf("listing buckets: %w", err)
		} else if len(buckets) == 0 {
			logger.WarnString(ctx, "No buckets found, will wait a bit and try again")
			select {
			case <-time.After(1 * time.Second):
				continue
			case <-ctx.Done():
				return ctx.Err()
			}
		}

		return nil
	}
}

// garageNodeBuddyPeers returns the "buddy" peers of the given host, based on
// the given garage cluster status. It will return zero values if the host has
// no buddy.
//
// For situations where we want one host to affect the cluster layout of another
// host's peers, we use a simple system to determine a single host which is
// responsible. The goal is not to be 100% race-proof (garage handles that), but
// rather to try to prevent all hosts from modifying the same host's layout at
// the same time.
//
// The system is to order all hosts by their IP, and say that each host is
// responsible for (aka the "buddy" of) the host immediately after their own in
// that list. The last host in that list is responsible for the first.
func garageNodeBuddyPeers(
	status garage.ClusterStatus, host bootstrap.Host,
) (
	netip.Addr, []garage.Role,
) {
	var (
		thisIP        = host.IP()
		nodeRolesByIP = map[netip.Addr][]garage.Role{}
	)

	for _, node := range status.Nodes {
		if node.Role == nil {
			continue
		}

		ip := node.Addr.Addr()
		nodeRolesByIP[ip] = append(nodeRolesByIP[ip], *node.Role)
	}

	// If there is only a single host in the cluster (or, somehow, none) then
	// that host has no buddy.
	if len(nodeRolesByIP) < 2 {
		return netip.Addr{}, nil
	}

	nodeIPs := maps.Keys(nodeRolesByIP)
	slices.SortFunc(nodeIPs, netip.Addr.Compare)

	for i, nodeIP := range nodeIPs {
		var buddyIP netip.Addr
		if i == len(nodeIPs)-1 {
			buddyIP = nodeIPs[0]
		} else if nodeIP == thisIP {
			buddyIP = nodeIPs[i+1]
		} else {
			continue
		}
		return buddyIP, nodeRolesByIP[buddyIP]
	}

	panic("Unreachable")
}