isle/go/daemon/network/network.go

//go:generate mockery --name Network --inpackage --filename network_mock.go

// Package network implements the Network type, which manages the daemon's
// membership in a single network.
package network

import (
	"bytes"
	"cmp"
	"context"
	"crypto/rand"
	"encoding/json"
	"errors"
	"fmt"
	"isle/bootstrap"
	"isle/daemon/children"
	"isle/daemon/daecommon"
	"isle/garage"
	"isle/jsonutil"
	"isle/nebula"
	"isle/secrets"
	"isle/toolkit"
	"net/netip"
	"slices"
	"sync"
	"time"

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

// GarageClientParams contains all the data needed to instantiate garage
// clients.
type GarageClientParams struct {
	Node                         garage.RemoteNode
	GlobalBucketS3APICredentials garage.S3APICredentials

	// RPCSecret may be empty, if the secret is not available on the host.
	//
	// TODO this shouldn't really be here I don't think, remove it?
	RPCSecret string
}

// GlobalBucketS3APIClient returns an S3 client pre-configured with access to
// the global bucket.
func (p GarageClientParams) GlobalBucketS3APIClient() *garage.S3APIClient {
	var (
		addr  = p.Node.S3APIAddr()
		creds = p.GlobalBucketS3APICredentials
	)
	return garage.NewS3APIClient(addr, creds)
}

// CreateHostOpts are optional parameters to the CreateHost method.
type CreateHostOpts struct {
	// IP address of the new host. An IP address will be randomly chosen if one
	// is not given here.
	IP netip.Addr

	// CanCreateHosts indicates that the bootstrap produced by CreateHost should
	// give the new host the ability to create new hosts as well.
	CanCreateHosts bool

	// TODO add nebula cert tags
}

// JoiningBootstrap wraps a normal Bootstrap to include extra data which a host
// might need while joining a network.
type JoiningBootstrap struct {
	Bootstrap bootstrap.Bootstrap
	Secrets   map[secrets.ID]json.RawMessage
}

// RPC defines the methods related to a single network which are available over
// the daemon's RPC interface.
type RPC interface {
	// GetHosts returns all hosts known to the network, sorted by their name.
	GetHosts(context.Context) ([]bootstrap.Host, error)

	// GetGarageClientParams returns a GarageClientParams for the current
	// network state.
	GetGarageClientParams(context.Context) (GarageClientParams, error)

	// GetNebulaCAPublicCredentials returns the CAPublicCredentials for the
	// network.
	GetNebulaCAPublicCredentials(
		context.Context,
	) (
		nebula.CAPublicCredentials, error,
	)

	// RemoveHost removes the host of the given name from the network.
	RemoveHost(ctx context.Context, hostName nebula.HostName) error

	// CreateHost creates a bootstrap for a new host with the given name and IP
	// address.
	CreateHost(
		context.Context, nebula.HostName, CreateHostOpts,
	) (
		JoiningBootstrap, error,
	)

	// CreateNebulaCertificate creates and signs a new nebula certficate for an
	// existing host, given the public key for that host. This is currently
	// mostly useful for creating certs for mobile devices.
	//
	// TODO replace this with CreateHostBootstrap, and the
	// CreateNebulaCertificate RPC method can just pull cert out of that.
	//
	// Errors:
	// - ErrHostNotFound
	CreateNebulaCertificate(
		context.Context, nebula.HostName, nebula.EncryptingPublicKey,
	) (
		nebula.Certificate, error,
	)

	// GetConfig returns the configuration currently in use.
	GetConfig(context.Context) (daecommon.NetworkConfig, error)

	// SetConfig overrides the current config with the given one, adjusting any
	// running child processes as needed.
	SetConfig(context.Context, daecommon.NetworkConfig) error
}

// Network manages membership in a single micropelago network. Each Network
// is comprised of a unique IP subnet, hosts connected together on that subnet
// via a VPN, an S3 storage layer only accessible to those hosts, plus other
// services built on this foundation.
//
// A single daemon (isle server) can manage multiple networks. Each network is
// expected to be independent of the others, ie they should not share any
// resources.
type Network interface {
	RPC

	// GetNetworkCreationParams returns the CreationParams that the Network was
	// originally created with.
	GetNetworkCreationParams(context.Context) (bootstrap.CreationParams, error)

	// Shutdown blocks until all resources held or created by the Network,
	// including child processes it has started, have been cleaned up.
	//
	// If this returns an error then it's possible that child processes are
	// still running and are no longer managed.
	Shutdown() error
}

////////////////////////////////////////////////////////////////////////////////
// Network implementation

// Opts are optional parameters which can be passed in when initializing a new
// Network instance. A nil Opts is equivalent to a zero value.
type Opts struct {
	ChildrenOpts *children.Opts

	GarageAdminToken string // Will be randomly generated if left unset.
}

func (o *Opts) withDefaults() *Opts {
	if o == nil {
		o = new(Opts)
	}

	if o.GarageAdminToken == "" {
		o.GarageAdminToken = toolkit.RandStr(32)
	}

	return o
}

type network struct {
	logger *mlog.Logger

	envBinDirPath string
	stateDir      toolkit.Dir
	runtimeDir    toolkit.Dir

	opts *Opts

	secretsStore secrets.Store

	l             sync.RWMutex
	children      *children.Children
	networkConfig daecommon.NetworkConfig
	currBootstrap bootstrap.Bootstrap

	workerCtx    context.Context
	workerCancel context.CancelFunc
	wg           sync.WaitGroup
}

// instatiateNetwork returns an instantiated *network instance which has not yet
// been initialized.
func instatiateNetwork(
	ctx context.Context,
	logger *mlog.Logger,
	networkConfig daecommon.NetworkConfig,
	envBinDirPath string,
	stateDir toolkit.Dir,
	runtimeDir toolkit.Dir,
	opts *Opts,
) *network {
	ctx = context.WithoutCancel(ctx)
	ctx, cancel := context.WithCancel(ctx)
	return &network{
		logger:        logger,
		networkConfig: networkConfig,
		envBinDirPath: envBinDirPath,
		stateDir:      stateDir,
		runtimeDir:    runtimeDir,
		opts:          opts.withDefaults(),
		workerCtx:     ctx,
		workerCancel:  cancel,
	}
}

// LoadCreationParams returns the CreationParams of a Network which was
// Created/Joined with the given state directory.
func LoadCreationParams(
	stateDir toolkit.Dir,
) (
	bootstrap.CreationParams, error,
) {
	var (
		// TODO store/load the creation params separately from the rest of
		// the bootstrap, since the bootstrap contains potentially the
		// entire host list of a network, which could be pretty bulky.
		bootstrapFilePath = bootstrap.StateDirPath(stateDir.Path)
		bs                bootstrap.Bootstrap
	)

	if err := jsonutil.LoadFile(&bs, bootstrapFilePath); err != nil {
		return bootstrap.CreationParams{}, fmt.Errorf(
			"loading bootstrap from %q: %w", bootstrapFilePath, err,
		)
	}

	return bs.NetworkCreationParams, nil
}

// Load initializes and returns a Network instance for a network which was
// previously joined or created, and which has the given ID.
func Load(
	ctx context.Context,
	logger *mlog.Logger,
	networkConfig daecommon.NetworkConfig,
	envBinDirPath string,
	stateDir toolkit.Dir,
	runtimeDir toolkit.Dir,
	opts *Opts,
) (
	Network, error,
) {
	n := instatiateNetwork(
		ctx,
		logger,
		networkConfig,
		envBinDirPath,
		stateDir,
		runtimeDir,
		opts,
	)

	if err := n.initializeDirs(true); err != nil {
		return nil, fmt.Errorf("initializing directories: %w", err)
	}

	var (
		currBootstrap     bootstrap.Bootstrap
		bootstrapFilePath = bootstrap.StateDirPath(n.stateDir.Path)
	)

	if err := jsonutil.LoadFile(&currBootstrap, bootstrapFilePath); err != nil {
		return nil, fmt.Errorf(
			"loading bootstrap from %q: %w", bootstrapFilePath, err,
		)
	} else if err := n.initialize(ctx, currBootstrap, false); err != nil {
		return nil, fmt.Errorf("initializing with bootstrap: %w", err)
	}

	return n, nil
}

// Join initializes and returns a Network instance for an existing network which
// was not previously joined to on this host. Once Join has been called for a
// particular network it will error on subsequent calls for that same network,
// Load should be used instead.
func Join(
	ctx context.Context,
	logger *mlog.Logger,
	networkConfig daecommon.NetworkConfig,
	joiningBootstrap JoiningBootstrap,
	envBinDirPath string,
	stateDir toolkit.Dir,
	runtimeDir toolkit.Dir,
	opts *Opts,
) (
	Network, error,
) {
	n := instatiateNetwork(
		ctx,
		logger,
		networkConfig,
		envBinDirPath,
		stateDir,
		runtimeDir,
		opts,
	)

	if err := n.initializeDirs(false); err != nil {
		return nil, fmt.Errorf("initializing directories: %w", err)
	}

	if err := secrets.Import(
		ctx, n.secretsStore, joiningBootstrap.Secrets,
	); err != nil {
		return nil, fmt.Errorf("importing secrets: %w", err)
	}

	if err := n.initialize(ctx, joiningBootstrap.Bootstrap, false); err != nil {
		return nil, fmt.Errorf("initializing with bootstrap: %w", err)
	}

	return n, nil
}

// Create initializes and returns a Network for a brand new network which uses
// the given creation parameters.
//
//   - name: Human-readable name of the network.
//   - domain: Primary domain name that network services are served under.
//   - ipNet: An IP subnet, in CIDR form, which will be the overall range of
//     possible IPs in the network. The first IP in this network range will
//     become this first host's IP.
//   - hostName: The name of this first host in the network.
//
// Errors:
//   - ErrInvalidConfig - if daemonConfig doesn't have 3 storage allocations
//     configured.
func Create(
	ctx context.Context,
	logger *mlog.Logger,
	networkConfig daecommon.NetworkConfig,
	envBinDirPath string,
	stateDir toolkit.Dir,
	runtimeDir toolkit.Dir,
	creationParams bootstrap.CreationParams,
	ipNet nebula.IPNet,
	hostName nebula.HostName,
	opts *Opts,
) (
	Network, error,
) {
	if len(networkConfig.Storage.Allocations) < 3 {
		return nil, ErrInvalidConfig.WithData(
			"At least three storage allocations are required.",
		)
	}

	nebulaCACreds, err := nebula.NewCACredentials(creationParams.Domain, ipNet)
	if err != nil {
		return nil, fmt.Errorf("creating nebula CA cert: %w", err)
	}

	garageRPCSecret := toolkit.RandStr(32)

	n := instatiateNetwork(
		ctx,
		logger,
		networkConfig,
		envBinDirPath,
		stateDir,
		runtimeDir,
		opts,
	)

	if err := n.initializeDirs(false); err != nil {
		return nil, fmt.Errorf("initializing directories: %w", err)
	}

	err = daecommon.SetGarageRPCSecret(ctx, n.secretsStore, garageRPCSecret)
	if err != nil {
		return nil, fmt.Errorf("setting garage RPC secret: %w", err)
	}

	err = daecommon.SetNebulaCASigningPrivateKey(
		ctx, n.secretsStore, nebulaCACreds.SigningPrivateKey,
	)
	if err != nil {
		return nil, fmt.Errorf("setting nebula CA signing key secret: %w", err)
	}

	hostBootstrap, err := bootstrap.New(
		nebulaCACreds,
		creationParams,
		map[nebula.HostName]bootstrap.Host{},
		hostName,
		ipNet.FirstAddr(),
	)
	if err != nil {
		return nil, fmt.Errorf("initializing bootstrap data: %w", err)
	}

	if err := n.initialize(ctx, hostBootstrap, true); err != nil {
		return nil, fmt.Errorf("initializing with bootstrap: %w", err)
	}

	return n, nil
}

func (n *network) initializeDirs(mayExist bool) error {
	secretsDir, err := n.stateDir.MkChildDir("secrets", mayExist)
	if err != nil {
		return fmt.Errorf("creating secrets dir: %w", err)
	}

	n.secretsStore = secrets.NewFSStore(secretsDir.Path)
	return nil
}

func (n *network) periodically(
	label string,
	fn func(context.Context) error,
	period time.Duration,
) {
	n.wg.Add(1)
	go func() {
		defer n.wg.Done()

		ctx := mctx.Annotate(n.workerCtx, "workerLabel", label)

		ticker := time.NewTicker(period)
		defer ticker.Stop()

		n.logger.Info(ctx, "Starting background job runner")
		defer n.logger.Info(ctx, "Stopping background job runner")

		for {
			select {
			case <-ctx.Done():
				return

			case <-ticker.C:
				n.logger.Info(ctx, "Background job running")
				if err := fn(ctx); err != nil {
					n.logger.Error(ctx, "Background job failed", err)
				}
			}
		}
	}()
}

func (n *network) initialize(
	ctx context.Context,
	prevBootstrap bootstrap.Bootstrap,
	isCreate bool,
) error {
	prevThisHost := prevBootstrap.ThisHost()

	// we update this Host's data using whatever configuration has been provided
	// by the daemon config. This way the network has the most up-to-date
	// possible bootstrap. This updated bootstrap will later get updated in
	// garage as a background task, so other hosts will see it as well.
	currBootstrap, err := coalesceNetworkConfigAndBootstrap(
		n.networkConfig, prevBootstrap,
	)
	if err != nil {
		return fmt.Errorf("combining configuration into bootstrap: %w", err)
	}

	n.logger.Info(ctx, "Writing updated bootstrap to state dir")
	err = writeBootstrapToStateDir(n.stateDir.Path, currBootstrap)
	if err != nil {
		return fmt.Errorf("writing bootstrap to state dir: %w", err)
	}

	n.currBootstrap = currBootstrap

	n.logger.Info(ctx, "Creating child processes")
	n.children, err = children.New(
		ctx,
		n.logger.WithNamespace("children"),
		n.envBinDirPath,
		n.secretsStore,
		n.networkConfig,
		n.runtimeDir,
		n.opts.GarageAdminToken,
		n.currBootstrap,
		n.opts.ChildrenOpts,
	)
	if err != nil {
		return fmt.Errorf("creating child processes: %w", err)
	}

	n.logger.Info(ctx, "Child processes created")

	createGarageGlobalBucket := isCreate
	err = n.postChildrenInit(ctx, prevThisHost, createGarageGlobalBucket)
	if err != nil {
		n.logger.Error(ctx, "Post-initialization failed, stopping child processes", err)
		n.children.Shutdown()
		return fmt.Errorf("performing post-initialization: %w", err)
	}

	// Do this now so that everything is stable before returning. This also
	// serves a dual-purpose, as it makes sure that the PUT from the postChildrenInit
	// above has propagated from the local garage instance, if there is one.
	n.logger.Info(ctx, "Reloading hosts from network storage")
	if err = n.reloadHosts(ctx); err != nil {
		return fmt.Errorf("Reloading network bootstrap: %w", err)
	}

	n.periodically("reloadHosts", n.reloadHosts, 3*time.Minute)

	n.periodically(
		"removeOrphanGarageNodes", n.removeOrphanGarageNodes, 1*time.Minute,
	)

	return nil
}

// postChildrenInit performs steps which are required after children have been
// initialized.
func (n *network) postChildrenInit(
	ctx context.Context,
	prevThisHost bootstrap.Host,
	createGarageGlobalBucket bool,
) error {
	n.l.RLock()
	defer n.l.RUnlock()

	thisHost := n.currBootstrap.ThisHost()

	if len(thisHost.Garage.Instances) > 0 {
		n.logger.Info(ctx, "Applying garage layout")
		if err := garageApplyLayout(
			ctx,
			n.logger,
			n.networkConfig,
			n.opts.GarageAdminToken,
			prevThisHost, thisHost,
		); err != nil {
			return fmt.Errorf("applying garage layout: %w", err)
		}
	}

	if createGarageGlobalBucket {
		n.logger.Info(ctx, "Initializing garage shared global bucket")
		garageGlobalBucketCreds, err := garageInitializeGlobalBucket(
			ctx,
			n.logger,
			n.networkConfig,
			n.opts.GarageAdminToken,
			thisHost,
		)
		if err != nil {
			return fmt.Errorf("initializing global bucket: %w", err)
		}

		err = daecommon.SetGarageS3APIGlobalBucketCredentials(
			ctx, n.secretsStore, garageGlobalBucketCreds,
		)
		if err != nil {
			return fmt.Errorf("storing global bucket creds: %w", err)
		}
	}

	for _, alloc := range n.networkConfig.Storage.Allocations {
		if err := garageWaitForAlloc(
			ctx, n.logger, alloc, n.opts.GarageAdminToken, thisHost,
		); err != nil {
			return fmt.Errorf(
				"waiting for alloc %+v to initialize: %w", alloc, err,
			)
		}
	}

	n.logger.Info(ctx, "Updating host info in garage")
	err := putGarageBoostrapHost(ctx, n.secretsStore, n.currBootstrap)
	if err != nil {
		return fmt.Errorf("updating host info in garage: %w", err)
	}

	return nil
}

func (n *network) reloadHosts(ctx context.Context) error {
	n.l.RLock()
	currBootstrap := n.currBootstrap
	n.l.RUnlock()

	n.logger.Info(ctx, "Checking for bootstrap changes")
	newHosts, err := getGarageBootstrapHosts(
		ctx, n.logger, n.secretsStore, currBootstrap,
	)
	if err != nil {
		return fmt.Errorf("getting hosts from garage: %w", err)
	}

	// TODO there's some potential race conditions here, where
	// CreateHost could be called at this point, write the new host to
	// garage and the bootstrap, but then this reload call removes the
	// host from this bootstrap/children until the next reload.

	newBootstrap := currBootstrap
	newBootstrap.Hosts = newHosts

	// the daemon's view of this host's bootstrap info takes precedence over
	// whatever is in garage. The garage version lacks the private credentials
	// which must be stored locally.
	thisHost := currBootstrap.ThisHost()
	newBootstrap.Hosts[thisHost.Name] = thisHost

	if _, err = n.reload(ctx, nil, &newBootstrap); err != nil {
		return fmt.Errorf("reloading with new host data: %w", err)
	}

	return nil
}

// In general each host will manage the garage cluster layout of its own storage
// allocations via garageApplyLayout. There are three situations which are
// handled here, rather than garageApplyLayout:
//
//   - A host removes all of its allocations via SetConfig.
//   - A host removes all of its allocations by calling Load with no allocations
//     in the provided daecommon.NetworkConfig.
//   - A host is removed from the network by another host.
//
// In all of these cases the host no longer has any garage instances running,
// and so can't call garageApplyLayout on itself. To combat this we have all
// hosts which do have garage instances running periodically check that there's
// not some garage nodes orphaned in the cluster layout, and remove them if so.
func (n *network) removeOrphanGarageNodes(ctx context.Context) error {
	n.l.RLock()
	defer n.l.RUnlock()

	thisHost := n.currBootstrap.ThisHost()
	if len(thisHost.Garage.Instances) == 0 {
		n.logger.Info(ctx, "No local garage instances, cannot remove orphans")
		return nil
	}

	adminClient := newGarageAdminClient(
		n.logger, n.networkConfig, n.opts.GarageAdminToken, thisHost,
	)
	defer adminClient.Close()

	clusterStatus, err := adminClient.Status(ctx)
	if err != nil {
		return fmt.Errorf("retrieving garage cluster status: %w", err)
	}

	buddyIP, buddyNodes := garageNodeBuddyPeers(clusterStatus, thisHost)
	if len(buddyNodes) == 0 {
		return nil
	}

	ctx = mctx.Annotate(ctx, "buddyIP", buddyIP)

	for _, host := range n.currBootstrap.Hosts {
		if host.IP() != buddyIP {
			continue
		} else if len(host.Garage.Instances) > 0 {
			n.logger.Info(ctx, "Buddy instance has garage nodes configured in its bootstrap, doing nothing")
			return nil
		}
		break
	}

	// Either the host is no longer in the network, or it no longer has any
	// garage instances set on it. Either way, remove its nodes from the cluster
	// layout.

	buddyNodeIDs := make([]string, len(buddyNodes))
	for i, buddyNode := range buddyNodes {
		buddyNodeIDs[i] = buddyNode.ID
	}

	n.logger.Info(ctx, "Applying garage layout to remove orphaned garage nodes")
	if err := adminClient.ApplyLayout(ctx, nil, buddyNodeIDs); err != nil {
		return fmt.Errorf(
			"applying garage cluster layout, removing nodes %+v: %w",
			buddyNodes,
			err,
		)
	}

	return nil
}

// returns the bootstrap prior to the reload being applied.
func (n *network) reload(
	ctx context.Context,
	newNetworkConfig *daecommon.NetworkConfig,
	newBootstrap *bootstrap.Bootstrap,
) (
	bootstrap.Bootstrap, error,
) {
	n.l.Lock()
	defer n.l.Unlock()

	prevBootstrap := n.currBootstrap

	if newBootstrap != nil {
		n.currBootstrap = *newBootstrap
	}

	if newNetworkConfig != nil {
		n.networkConfig = *newNetworkConfig
	}

	var err error
	if n.currBootstrap, err = coalesceNetworkConfigAndBootstrap(
		n.networkConfig, n.currBootstrap,
	); err != nil {
		return bootstrap.Bootstrap{}, fmt.Errorf(
			"combining configuration into bootstrap: %w", err,
		)
	}

	n.logger.Info(ctx, "Writing updated bootstrap to state dir")
	err = writeBootstrapToStateDir(n.stateDir.Path, n.currBootstrap)
	if err != nil {
		return bootstrap.Bootstrap{}, fmt.Errorf(
			"writing bootstrap to state dir: %w", err,
		)
	}

	n.logger.Info(ctx, "Reloading child processes")
	err = n.children.Reload(ctx, n.networkConfig, n.currBootstrap)
	if err != nil {
		return bootstrap.Bootstrap{}, fmt.Errorf(
			"reloading child processes: %w", err,
		)
	}

	return prevBootstrap, nil
}

func withCurrBootstrap[Res any](
	n *network, fn func(bootstrap.Bootstrap) (Res, error),
) (Res, error) {
	n.l.RLock()
	defer n.l.RUnlock()

	currBootstrap := n.currBootstrap
	return fn(currBootstrap)
}

func (n *network) getBootstrap() (
	bootstrap.Bootstrap, error,
) {
	return withCurrBootstrap(n, func(
		currBootstrap bootstrap.Bootstrap,
	) (
		bootstrap.Bootstrap, error,
	) {
		return currBootstrap, nil
	})
}

func (n *network) GetHosts(ctx context.Context) ([]bootstrap.Host, error) {
	return withCurrBootstrap(n, func(
		currBootstrap bootstrap.Bootstrap,
	) (
		[]bootstrap.Host, error,
	) {
		hosts := maps.Values(currBootstrap.Hosts)
		slices.SortFunc(hosts, func(a, b bootstrap.Host) int {
			return cmp.Compare(a.Name, b.Name)
		})
		return hosts, nil
	})
}

func (n *network) GetGarageClientParams(
	ctx context.Context,
) (
	GarageClientParams, error,
) {
	return withCurrBootstrap(n, func(
		currBootstrap bootstrap.Bootstrap,
	) (
		GarageClientParams, error,
	) {
		return getGarageClientParams(ctx, n.secretsStore, currBootstrap)
	})
}

func (n *network) GetNebulaCAPublicCredentials(
	ctx context.Context,
) (
	nebula.CAPublicCredentials, error,
) {
	b, err := n.getBootstrap()
	if err != nil {
		return nebula.CAPublicCredentials{}, fmt.Errorf(
			"retrieving bootstrap: %w", err,
		)
	}

	return b.CAPublicCredentials, nil
}

func (n *network) RemoveHost(ctx context.Context, hostName nebula.HostName) error {
	// TODO RemoveHost should publish a certificate revocation for the host
	// being removed.
	_, err := withCurrBootstrap(n, func(
		currBootstrap bootstrap.Bootstrap,
	) (
		struct{}, error,
	) {
		garageClientParams, err := getGarageClientParams(
			ctx, n.secretsStore, currBootstrap,
		)
		if err != nil {
			return struct{}{}, fmt.Errorf("get garage client params: %w", err)
		}

		client := garageClientParams.GlobalBucketS3APIClient()
		defer client.Close()

		return struct{}{}, removeGarageBootstrapHost(ctx, client, hostName)
	})
	return err
}

func makeCACreds(
	currBootstrap bootstrap.Bootstrap,
	caSigningPrivateKey nebula.SigningPrivateKey,
) nebula.CACredentials {
	return nebula.CACredentials{
		Public:            currBootstrap.CAPublicCredentials,
		SigningPrivateKey: caSigningPrivateKey,
	}
}

func chooseAvailableIP(b bootstrap.Bootstrap) (netip.Addr, error) {
	var (
		cidrIPNet = b.CAPublicCredentials.Cert.Unwrap().Details.Subnets[0]
		cidrMask  = cidrIPNet.Mask
		cidrIPB   = cidrIPNet.IP

		cidr           = netip.MustParsePrefix(cidrIPNet.String())
		cidrIP         = cidr.Addr()
		cidrSuffixBits = cidrIP.BitLen() - cidr.Bits()

		inUseIPs = make(map[netip.Addr]struct{}, len(b.Hosts))
	)

	for _, host := range b.Hosts {
		inUseIPs[host.IP()] = struct{}{}
	}

	// first check that there are any addresses at all. We can determine the
	// number of possible addresses using the network CIDR. The first IP in a
	// subnet is the network identifier, and is reserved. The last IP is the
	// broadcast IP, and is also reserved. Hence, the -2.
	usableIPs := (1 << cidrSuffixBits) - 2
	if len(inUseIPs) >= usableIPs {
		return netip.Addr{}, errors.New("no available IPs")
	}

	// We need to know the subnet broadcast address, so we don't accidentally
	// produce it.
	cidrBCastIPB := bytes.Clone(cidrIPB)
	for i := range cidrBCastIPB {
		cidrBCastIPB[i] |= ^cidrMask[i]
	}
	cidrBCastIP, ok := netip.AddrFromSlice(cidrBCastIPB)
	if !ok {
		panic(fmt.Sprintf("invalid broadcast ip calculated: %x", cidrBCastIP))
	}

	// Try a handful of times to pick an IP at random. This is preferred, as it
	// leaves less room for two different CreateHost calls to choose the same
	// IP.
	for range 20 {
		b := make([]byte, len(cidrIPB))
		if _, err := rand.Read(b); err != nil {
			return netip.Addr{}, fmt.Errorf("reading random bytes: %w", err)
		}

		for i := range b {
			b[i] = cidrIPB[i] | (b[i] & ^cidrMask[i])
		}

		ip, ok := netip.AddrFromSlice(b)
		if !ok {
			panic(fmt.Sprintf("generated invalid IP: %x", b))
		} else if !cidr.Contains(ip) {
			panic(fmt.Sprintf(
				"generated IP %v which is not in cidr %v", ip, cidr,
			))
		}

		if ip == cidrIP || ip == cidrBCastIP {
			continue
		}

		if _, inUse := inUseIPs[ip]; !inUse {
			return ip, nil
		}
	}

	// If randomly picking fails then just go through IPs one by one until the
	// free one is found.
	for ip := cidrIP.Next(); ip != cidrBCastIP; ip = ip.Next() {
		if _, inUse := inUseIPs[ip]; !inUse {
			return ip, nil
		}
	}

	panic("All ips are in-use, but somehow that wasn't determined earlier")
}

func (n *network) CreateHost(
	ctx context.Context,
	hostName nebula.HostName,
	opts CreateHostOpts,
) (
	JoiningBootstrap, error,
) {
	n.l.RLock()
	currBootstrap := n.currBootstrap
	n.l.RUnlock()

	ip := opts.IP
	if ip == (netip.Addr{}) {
		var err error
		if ip, err = chooseAvailableIP(currBootstrap); err != nil {
			return JoiningBootstrap{}, fmt.Errorf(
				"choosing available IP: %w", err,
			)
		}
	}
	// TODO if the ip is given, check that it's not already in use.

	caSigningPrivateKey, err := daecommon.GetNebulaCASigningPrivateKey(
		ctx, n.secretsStore,
	)
	if err != nil {
		return JoiningBootstrap{}, fmt.Errorf("getting CA signing key: %w", err)
	}

	var joiningBootstrap JoiningBootstrap
	joiningBootstrap.Bootstrap, err = bootstrap.New(
		makeCACreds(currBootstrap, caSigningPrivateKey),
		currBootstrap.NetworkCreationParams,
		currBootstrap.Hosts,
		hostName,
		ip,
	)
	if err != nil {
		return JoiningBootstrap{}, fmt.Errorf(
			"initializing bootstrap data: %w", err,
		)
	}

	secretsIDs := []secrets.ID{
		daecommon.GarageRPCSecretSecretID,
		daecommon.GarageS3APIGlobalBucketCredentialsSecretID,
	}

	if opts.CanCreateHosts {
		secretsIDs = append(
			secretsIDs, daecommon.NebulaCASigningPrivateKeySecretID,
		)
	}

	if joiningBootstrap.Secrets, err = secrets.Export(
		ctx, n.secretsStore, secretsIDs,
	); err != nil {
		return JoiningBootstrap{}, fmt.Errorf("exporting secrets: %w", err)
	}

	n.logger.Info(ctx, "Putting new host in garage")
	err = putGarageBoostrapHost(ctx, n.secretsStore, joiningBootstrap.Bootstrap)
	if err != nil {
		return JoiningBootstrap{}, fmt.Errorf("putting new host in garage: %w", err)
	}

	// the new bootstrap will have been initialized with both all existing hosts
	// (based on currBootstrap) and the host being created.
	newBootstrap := currBootstrap
	newBootstrap.Hosts = joiningBootstrap.Bootstrap.Hosts

	n.logger.Info(ctx, "Reloading local state with new host")
	if _, err = n.reload(ctx, nil, &newBootstrap); err != nil {
		return JoiningBootstrap{}, fmt.Errorf("reloading child processes: %w", err)
	}

	return joiningBootstrap, nil
}

func (n *network) CreateNebulaCertificate(
	ctx context.Context,
	hostName nebula.HostName,
	hostPubKey nebula.EncryptingPublicKey,
) (
	nebula.Certificate, error,
) {
	return withCurrBootstrap(n, func(
		currBootstrap bootstrap.Bootstrap,
	) (
		nebula.Certificate, error,
	) {
		host, ok := currBootstrap.Hosts[hostName]
		if !ok {
			return nebula.Certificate{}, ErrHostNotFound
		}
		ip := host.IP()

		caSigningPrivateKey, err := daecommon.GetNebulaCASigningPrivateKey(
			ctx, n.secretsStore,
		)
		if err != nil {
			return nebula.Certificate{}, fmt.Errorf("getting CA signing key: %w", err)
		}

		caCreds := makeCACreds(currBootstrap, caSigningPrivateKey)

		return nebula.NewHostCert(caCreds, hostPubKey, hostName, ip)
	})
}

func (n *network) GetConfig(context.Context) (daecommon.NetworkConfig, error) {
	n.l.RLock()
	defer n.l.RUnlock()
	return n.networkConfig, nil
}

func (n *network) SetConfig(
	ctx context.Context, config daecommon.NetworkConfig,
) error {
	prevBootstrap, err := n.reload(ctx, &config, nil)
	if err != nil {
		return fmt.Errorf("reloading config: %w", err)
	}

	if err := n.postChildrenInit(ctx, prevBootstrap.ThisHost(), false); err != nil {
		return fmt.Errorf("performing post-initialization: %w", err)
	}

	return nil
}

func (n *network) GetNetworkCreationParams(
	ctx context.Context,
) (
	bootstrap.CreationParams, error,
) {
	return withCurrBootstrap(n, func(
		currBootstrap bootstrap.Bootstrap,
	) (
		bootstrap.CreationParams, error,
	) {
		return currBootstrap.NetworkCreationParams, nil
	})
}

func (n *network) Shutdown() error {
	n.workerCancel()
	n.wg.Wait()

	if n.children != nil {
		n.logger.Info(context.Background(), "Shutting down children")
		n.children.Shutdown()
	}

	return nil
}