daemon.go

/*
 * Copyright (c) 2020. Ant Group. All rights reserved.
 * Copyright (c) 2022. Nydus Developers. All rights reserved.
 *
 * SPDX-License-Identifier: Apache-2.0
 */

package daemon

import (
	"os"
	"path/filepath"
	"sort"
	"sync"
	"sync/atomic"
	"syscall"
	"time"

	"github.com/pkg/errors"

	"github.com/containerd/containerd/log"

	"github.com/containerd/nydus-snapshotter/config"
	"github.com/containerd/nydus-snapshotter/config/daemonconfig"
	"github.com/containerd/nydus-snapshotter/pkg/daemon/types"
	"github.com/containerd/nydus-snapshotter/pkg/errdefs"
	"github.com/containerd/nydus-snapshotter/pkg/rafs"
	"github.com/containerd/nydus-snapshotter/pkg/supervisor"
	"github.com/containerd/nydus-snapshotter/pkg/utils/erofs"
	"github.com/containerd/nydus-snapshotter/pkg/utils/mount"
	"github.com/containerd/nydus-snapshotter/pkg/utils/retry"
)

const (
	APISocketFileName   = "api.sock"
	SharedNydusDaemonID = "shared_daemon"
)

type NewDaemonOpt func(d *Daemon) error

// Fields in this structure should be write-once, and caller should hold `Daemon.mu` when updating fields.
type ConfigState struct {
	// A unique ID generated by daemon manager to identify the nydusd instance.
	ID              string
	ProcessID       int
	APISocket       string
	DaemonMode      config.DaemonMode
	FsDriver        string
	LogDir          string
	LogLevel        string
	LogRotationSize int
	LogToStdout     bool
	Mountpoint      string
	SupervisorPath  string
	ThreadNum       int
	// Where the configuration file resides, all rafs instances share the same configuration template
	ConfigDir string
}

// TODO: Record queried nydusd state
type Daemon struct {
	States ConfigState

	mu sync.Mutex
	// Host all RAFS filesystems managed by this daemon:
	// fusedev dedicated mode: one and only one RAFS instance
	// fusedev shared mode: zero, one or more RAFS instances
	// fscache shared mode: zero, one or more RAFS instances
	RafsCache rafs.Cache

	// Protect nydusd http client
	cmu sync.Mutex
	// client will be rebuilt on Reconnect, skip marshal/unmarshal
	client NydusdClient

	// Nil means this daemon object has no supervisor
	Supervisor *supervisor.Supervisor
	Config     daemonconfig.DaemonConfig

	// How much CPU nydusd is utilizing when starts since full prefetch might
	// consume many CPU cycles
	StartupCPUUtilization float64
	Version               types.BuildTimeInfo

	ref int32
	// Cache the nydusd daemon state to avoid frequently querying nydusd by API.
	state types.DaemonState
}

func (d *Daemon) Lock() {
	d.mu.Lock()
}

func (d *Daemon) Unlock() {
	d.mu.Unlock()
}

func (d *Daemon) ID() string {
	return d.States.ID
}

func (d *Daemon) Pid() int {
	return d.States.ProcessID
}

func (d *Daemon) IncRef() {
	atomic.AddInt32(&d.ref, 1)
}

func (d *Daemon) DecRef() int32 {
	return atomic.AddInt32(&d.ref, -1)
}

func (d *Daemon) GetRef() int32 {
	return atomic.LoadInt32(&d.ref)
}

func (d *Daemon) HostMountpoint() (mnt string) {
	mnt = d.States.Mountpoint
	return
}

// Each nydusd daemon has a copy of configuration json file.
func (d *Daemon) ConfigFile(instanceID string) string {
	if instanceID == "" {
		return filepath.Join(d.States.ConfigDir, "config.json")
	}
	return filepath.Join(d.States.ConfigDir, instanceID, "config.json")
}

// NydusdThreadNum returns how many working threads are needed of a single nydusd
func (d *Daemon) NydusdThreadNum() int {
	return d.States.ThreadNum
}

func (d *Daemon) GetAPISock() string {
	return d.States.APISocket
}

func (d *Daemon) LogFile() string {
	return filepath.Join(d.States.LogDir, "nydusd.log")
}

func (d *Daemon) AddRafsInstance(r *rafs.Rafs) {
	d.RafsCache.Add(r)
	d.IncRef()
	r.DaemonID = d.ID()
}

func (d *Daemon) RemoveRafsInstance(snapshotID string) {
	d.RafsCache.Remove(snapshotID)
	d.DecRef()
}

// Get and cache daemon current working state by querying nydusd:
// 1. INIT
// 2. READY: All needed resources are ready.
// 3. RUNNING
func (d *Daemon) GetState() (types.DaemonState, error) {
	c, err := d.GetClient()
	if err != nil {
		return types.DaemonStateUnknown, errors.Wrapf(err, "get daemon state")
	}
	info, err := c.GetDaemonInfo()
	if err != nil {
		return types.DaemonStateUnknown, err
	}

	st := info.DaemonState()

	d.Lock()
	d.state = st
	d.Version = info.DaemonVersion()
	d.Unlock()

	return st, nil
}

// Return the cached nydusd working status, no API is invoked.
func (d *Daemon) State() types.DaemonState {
	d.Lock()
	defer d.Unlock()
	return d.state
}

// Reset the cached nydusd working status
func (d *Daemon) ResetState() {
	d.Lock()
	defer d.Unlock()
	d.state = types.DaemonStateUnknown
}

// Wait for the nydusd daemon to reach specified state with timeout.
func (d *Daemon) WaitUntilState(expected types.DaemonState) error {
	return retry.Do(func() error {
		if expected == d.State() {
			return nil
		}

		state, err := d.GetState()
		if err != nil {
			return errors.Wrapf(err, "wait until daemon is %s", expected)
		}

		if state != expected {
			return errors.Errorf("daemon %s is not %s yet, current state %s",
				d.ID(), expected, state)
		}

		return nil
	},
		retry.LastErrorOnly(true),
		retry.Attempts(20), // totally wait for 2 seconds, should be enough
		retry.Delay(100*time.Millisecond),
	)
}

func (d *Daemon) IsSharedDaemon() bool {
	if d.States.DaemonMode != "" {
		return d.States.DaemonMode == config.DaemonModeShared
	}

	return d.HostMountpoint() == config.GetRootMountpoint()
}

func (d *Daemon) SharedMount(rafs *rafs.Rafs) error {
	defer d.SendStates()

	switch d.States.FsDriver {
	case config.FsDriverFscache:
		if err := d.sharedErofsMount(rafs); err != nil {
			return errors.Wrapf(err, "mount erofs")
		}
		return nil
	case config.FsDriverFusedev:
		return d.sharedFusedevMount(rafs)
	default:
		return errors.Errorf("unsupported fs driver %s", d.States.FsDriver)
	}
}

func (d *Daemon) sharedFusedevMount(rafs *rafs.Rafs) error {
	client, err := d.GetClient()
	if err != nil {
		return errors.Wrapf(err, "mount instance %s", rafs.SnapshotID)
	}

	bootstrap, err := rafs.BootstrapFile()
	if err != nil {
		return err
	}

	c, err := daemonconfig.NewDaemonConfig(d.States.FsDriver, d.ConfigFile(rafs.SnapshotID))
	if err != nil {
		return errors.Wrapf(err, "Failed to reload instance configuration %s",
			d.ConfigFile(rafs.SnapshotID))
	}

	cfg, err := c.DumpString()
	if err != nil {
		return errors.Wrap(err, "dump instance configuration")
	}

	err = client.Mount(rafs.RelaMountpoint(), bootstrap, cfg)
	if err != nil {
		return errors.Wrapf(err, "mount rafs instance")
	}

	return nil
}

func (d *Daemon) sharedErofsMount(ra *rafs.Rafs) error {
	client, err := d.GetClient()
	if err != nil {
		return errors.Wrapf(err, "bind blob %s", d.ID())
	}

	// TODO: Why fs cache needing this work dir?
	if err := os.MkdirAll(ra.FscacheWorkDir(), 0755); err != nil {
		return errors.Wrapf(err, "failed to create fscache work dir %s", ra.FscacheWorkDir())
	}

	c, err := daemonconfig.NewDaemonConfig(d.States.FsDriver, d.ConfigFile(ra.SnapshotID))
	if err != nil {
		log.L.Errorf("Failed to reload daemon configuration %s, %s", d.ConfigFile(ra.SnapshotID), err)
		return err
	}

	cfgStr, err := c.DumpString()
	if err != nil {
		return err
	}

	if err := client.BindBlob(cfgStr); err != nil {
		return errors.Wrapf(err, "request to bind fscache blob")
	}

	mountPoint := ra.GetMountpoint()
	if err := os.MkdirAll(mountPoint, 0755); err != nil {
		return errors.Wrapf(err, "create mountpoint %s", mountPoint)
	}

	fscacheID := erofs.FscacheID(ra.SnapshotID)

	cfg := c.(*daemonconfig.FscacheDaemonConfig)
	ra.AddAnnotation(rafs.AnnoFsCacheDomainID, cfg.DomainID)
	ra.AddAnnotation(rafs.AnnoFsCacheID, fscacheID)

	if err := erofs.Mount(cfg.DomainID, fscacheID, mountPoint); err != nil {
		if !errdefs.IsErofsMounted(err) {
			return errors.Wrapf(err, "mount erofs to %s", mountPoint)
		}
		// When snapshotter exits (either normally or abnormally), it will not have a
		// chance to umount erofs mountpoint, so if snapshotter resumes running and mount
		// again (by a new request to create container), it will need to ignore the mount
		// error `device or resource busy`.
		log.L.Warnf("erofs mountpoint %s has been mounted", mountPoint)
	}

	return nil
}

func (d *Daemon) SharedUmount(rafs *rafs.Rafs) error {
	defer d.SendStates()

	switch d.States.FsDriver {
	case config.FsDriverFscache:
		if err := d.sharedErofsUmount(rafs); err != nil {
			return errors.Wrapf(err, "failed to erofs mount")
		}
		return nil
	case config.FsDriverFusedev:
		c, err := d.GetClient()
		if err != nil {
			return errors.Wrapf(err, "umount instance %s", rafs.SnapshotID)
		}
		return c.Umount(rafs.RelaMountpoint())
	default:
		return errors.Errorf("unsupported fs driver %s", d.States.FsDriver)
	}
}

func (d *Daemon) sharedErofsUmount(ra *rafs.Rafs) error {
	c, err := d.GetClient()
	if err != nil {
		return errors.Wrapf(err, "unbind blob %s", d.ID())
	}
	domainID := ra.Annotations[rafs.AnnoFsCacheDomainID]
	fscacheID := ra.Annotations[rafs.AnnoFsCacheID]

	if err := c.UnbindBlob(domainID, fscacheID); err != nil {
		return errors.Wrapf(err, "request to unbind fscache blob, domain %s, fscache %s", domainID, fscacheID)
	}

	mountpoint := ra.GetMountpoint()
	if err := erofs.Umount(mountpoint); err != nil {
		return errors.Wrapf(err, "umount erofs %s mountpoint, %s", err, mountpoint)
	}

	// delete fscache bootstrap cache file
	// erofs generate fscache cache file for bootstrap with fscacheID
	if err := c.UnbindBlob("", fscacheID); err != nil {
		log.L.Warnf("delete bootstrap %s err %s", fscacheID, err)
	}

	return nil
}

func (d *Daemon) UmountRafsInstance(r *rafs.Rafs) error {
	if d.IsSharedDaemon() {
		if err := d.SharedUmount(r); err != nil {
			return errors.Wrapf(err, "umount fs instance %s", r.SnapshotID)
		}
	}

	return nil
}

func (d *Daemon) UmountRafsInstances() error {
	if d.IsSharedDaemon() {
		d.RafsCache.Lock()
		defer d.RafsCache.Unlock()

		instances := d.RafsCache.ListLocked()

		for _, r := range instances {
			if err := d.SharedUmount(r); err != nil {
				return errors.Wrapf(err, "umount fs instance %s", r.SnapshotID)
			}
		}
	}

	return nil
}

func (d *Daemon) SendStates() {
	su := d.Supervisor
	if su != nil {
		// TODO: This should be optional by checking snapshotter's configuration.
		// FIXME: Is it possible the states are overwritten during two API mounts.
		// FIXME: What if nydusd does not support sending states.
		err := su.FetchDaemonStates(func() error {
			if err := d.doSendStates(); err != nil {
				return errors.Wrapf(err, "send daemon %s states", d.ID())
			}
			return nil
		})
		if err != nil {
			log.L.Warnf("Daemon %s does not support sending states, %v", d.ID(), err)
		}
	}
}

func (d *Daemon) doSendStates() error {
	c, err := d.GetClient()
	if err != nil {
		return errors.Wrapf(err, "send states %s", d.ID())
	}

	if err := c.SendFd(); err != nil {
		return errors.Wrap(err, "request to send states")
	}

	return nil
}

func (d *Daemon) TakeOver() error {
	c, err := d.GetClient()
	if err != nil {
		return errors.Wrapf(err, "takeover daemon %s", d.ID())
	}

	if err := c.TakeOver(); err != nil {
		return errors.Wrap(err, "request to take over")
	}

	return nil
}

func (d *Daemon) Start() error {
	c, err := d.GetClient()
	if err != nil {
		return errors.Wrapf(err, "start service")
	}

	if err := c.Start(); err != nil {
		return errors.Wrap(err, "request to start service")
	}

	return nil
}

func (d *Daemon) Exit() error {
	c, err := d.GetClient()
	if err != nil {
		return errors.Wrapf(err, "start service")
	}

	if err := c.Exit(); err != nil {
		return errors.Wrap(err, "request to exit service")
	}

	return nil
}

func (d *Daemon) GetDaemonInfo() (*types.DaemonInfo, error) {
	c, err := d.GetClient()
	if err != nil {
		return nil, errors.Wrapf(err, "get daemon information")
	}

	return c.GetDaemonInfo()
}

func (d *Daemon) GetFsMetrics(sid string) (*types.FsMetrics, error) {
	c, err := d.GetClient()
	if err != nil {
		return nil, errors.Wrapf(err, "get fs metrics")
	}

	return c.GetFsMetrics(sid)
}

func (d *Daemon) GetInflightMetrics() (*types.InflightMetrics, error) {
	c, err := d.GetClient()
	if err != nil {
		return nil, errors.Wrapf(err, "get inflight metrics")
	}

	return c.GetInflightMetrics()
}

func (d *Daemon) GetCacheMetrics(sid string) (*types.CacheMetrics, error) {
	c, err := d.GetClient()
	if err != nil {
		return nil, errors.Wrapf(err, "get cache metrics")
	}
	return c.GetCacheMetrics(sid)
}

func (d *Daemon) GetClient() (NydusdClient, error) {
	d.cmu.Lock()
	defer d.cmu.Unlock()

	if d.client == nil {
		sock := d.GetAPISock()
		// The socket file may be residual from a dead nydusd
		err := WaitUntilSocketExisted(sock, d.Pid())
		if err != nil {
			return nil, errors.Wrapf(errdefs.ErrNotFound, "daemon socket %s", sock)
		}
		client, err := NewNydusClient(sock)
		if err != nil {
			return nil, errors.Wrapf(err, "create daemon %s client", d.ID())
		}
		d.client = client
	}

	return d.client, nil
}

func (d *Daemon) ResetClient() {
	d.cmu.Lock()
	d.client = nil
	d.cmu.Unlock()
}

func (d *Daemon) Terminate() error {
	// if we found pid here, we need to kill and wait process to exit, Pid=0 means somehow we lost
	// the daemon pid, so that we can't kill the process, just roughly umount the mountpoint
	d.Lock()
	defer d.Unlock()

	if d.Pid() > 0 {
		p, err := os.FindProcess(d.Pid())
		if err != nil {
			return errors.Wrapf(err, "find process %d", d.Pid())
		}
		if err = p.Signal(syscall.SIGTERM); err != nil {
			return errors.Wrapf(err, "send SIGTERM signal to process %d", d.Pid())
		}
	}

	return nil
}

func (d *Daemon) Wait() error {
	// if we found pid here, we need to kill and wait process to exit, Pid=0 means somehow we lost
	// the daemon pid, so that we can't kill the process, just roughly umount the mountpoint
	d.Lock()
	defer d.Unlock()

	if d.Pid() > 0 {
		p, err := os.FindProcess(d.Pid())
		if err != nil {
			return errors.Wrapf(err, "find process %d", d.Pid())
		}

		// if nydus-snapshotter restarts, it will break the relationship between nydusd and
		// nydus-snapshotter, p.Wait() will return err, so here should exclude this case
		if _, err = p.Wait(); err != nil && !errors.Is(err, syscall.ECHILD) {
			log.L.Errorf("failed to process wait, %v", err)
		} else if d.HostMountpoint() != "" && config.GetFsDriver() == config.FsDriverFusedev {
			// No need to umount if the nydusd never performs mount. In other word, it does not
			// associate with a host mountpoint.
			if err := mount.WaitUntilUnmounted(d.HostMountpoint()); err != nil {
				log.L.WithError(err).Errorf("umount %s", d.HostMountpoint())
			}
		}
	}

	return nil
}

// When daemon dies, clean up its vestige before start a new one.
func (d *Daemon) ClearVestige() {
	mounter := mount.Mounter{}
	if d.States.FsDriver == config.FsDriverFscache {
		instances := d.RafsCache.List()
		for _, i := range instances {
			if err := mounter.Umount(i.GetMountpoint()); err != nil {
				log.L.Warnf("Can't umount %s, %v", d.States.Mountpoint, err)
			}
		}
	} else {
		log.L.Infof("Unmounting %s when clear vestige", d.HostMountpoint())
		if err := mounter.Umount(d.HostMountpoint()); err != nil {
			log.L.Warnf("Can't umount %s, %v", d.States.Mountpoint, err)
		}
	}

	// Nydusd judges if it should enter failover phrase by checking
	// if unix socket is existed and it can't be connected.
	if err := os.Remove(d.GetAPISock()); err != nil {
		log.L.Warnf("Can't delete residual unix socket %s, %v", d.GetAPISock(), err)
	}

	// `CheckStatus->ensureClient` only checks if socket file is existed when building http client.
	// But the socket file may be residual and will be cleared before starting a new nydusd.
	// So clear the client by assigning nil
	d.ResetClient()
}

func (d *Daemon) CloneRafsInstances(src *Daemon) {
	instances := src.RafsCache.List()
	d.RafsCache.SetIntances(instances)
}

// Daemon must be started and reach RUNNING state before call this method
func (d *Daemon) RecoverRafsInstances() error {
	if d.IsSharedDaemon() {
		d.RafsCache.Lock()
		defer d.RafsCache.Unlock()

		instances := make([]*rafs.Rafs, 0, 16)
		for _, r := range d.RafsCache.ListLocked() {
			instances = append(instances, r)
		}

		sort.Slice(instances, func(i, j int) bool {
			return instances[i].Seq < instances[j].Seq
		})

		for _, i := range instances {
			if d.HostMountpoint() != i.GetMountpoint() {
				log.L.Infof("Recovered mount instance %s", i.SnapshotID)
				if err := d.SharedMount(i); err != nil {
					return err
				}
			}
		}
	}

	return nil
}

// Instantiate a daemon object
func NewDaemon(opt ...NewDaemonOpt) (*Daemon, error) {
	d := &Daemon{}
	d.States.ID = newID()
	d.States.DaemonMode = config.DaemonModeDedicated
	d.RafsCache = rafs.NewRafsCache()

	for _, o := range opt {
		err := o(d)
		if err != nil {
			return nil, err
		}
	}

	return d, nil
}