package storage

import (
	"bytes"
	"errors"
	"fmt"
	"io"
	"os"
	"path"
	"path/filepath"
	"reflect"
	"sort"
	"strings"
	"sync"
	"time"

	drivers "github.com/containers/storage/drivers"
	"github.com/containers/storage/pkg/archive"
	"github.com/containers/storage/pkg/idtools"
	"github.com/containers/storage/pkg/ioutils"
	"github.com/containers/storage/pkg/mount"
	"github.com/containers/storage/pkg/stringid"
	"github.com/containers/storage/pkg/system"
	"github.com/containers/storage/pkg/tarlog"
	"github.com/containers/storage/pkg/truncindex"
	multierror "github.com/hashicorp/go-multierror"
	"github.com/klauspost/pgzip"
	digest "github.com/opencontainers/go-digest"
	"github.com/opencontainers/selinux/go-selinux/label"
	"github.com/sirupsen/logrus"
	"github.com/vbatts/tar-split/archive/tar"
	"github.com/vbatts/tar-split/tar/asm"
	"github.com/vbatts/tar-split/tar/storage"
)

const (
	tarSplitSuffix = ".tar-split.gz"
	incompleteFlag = "incomplete"
)

// A Layer is a record of a copy-on-write layer that's stored by the lower
// level graph driver.
type Layer struct {
	// ID is either one which was specified at create-time, or a random
	// value which was generated by the library.
	ID string `json:"id"`

	// Names is an optional set of user-defined convenience values.  The
	// layer can be referred to by its ID or any of its names.  Names are
	// unique among layers.
	Names []string `json:"names,omitempty"`

	// Parent is the ID of a layer from which this layer inherits data.
	Parent string `json:"parent,omitempty"`

	// Metadata is data we keep for the convenience of the caller.  It is not
	// expected to be large, since it is kept in memory.
	Metadata string `json:"metadata,omitempty"`

	// MountLabel is an SELinux label which should be used when attempting to mount
	// the layer.
	MountLabel string `json:"mountlabel,omitempty"`

	// MountPoint is the path where the layer is mounted, or where it was most
	// recently mounted.  This can change between subsequent Unmount() and
	// Mount() calls, so the caller should consult this value after Mount()
	// succeeds to find the location of the container's root filesystem.
	MountPoint string `json:"-"`

	// MountCount is used as a reference count for the container's layer being
	// mounted at the mount point.
	MountCount int `json:"-"`

	// Created is the datestamp for when this layer was created.  Older
	// versions of the library did not track this information, so callers
	// will likely want to use the IsZero() method to verify that a value
	// is set before using it.
	Created time.Time `json:"created,omitempty"`

	// CompressedDigest is the digest of the blob that was last passed to
	// ApplyDiff() or Put(), as it was presented to us.
	CompressedDigest digest.Digest `json:"compressed-diff-digest,omitempty"`

	// CompressedSize is the length of the blob that was last passed to
	// ApplyDiff() or Put(), as it was presented to us.  If
	// CompressedDigest is not set, this should be treated as if it were an
	// uninitialized value.
	CompressedSize int64 `json:"compressed-size,omitempty"`

	// UncompressedDigest is the digest of the blob that was last passed to
	// ApplyDiff() or Put(), after we decompressed it.  Often referred to
	// as a DiffID.
	UncompressedDigest digest.Digest `json:"diff-digest,omitempty"`

	// UncompressedSize is the length of the blob that was last passed to
	// ApplyDiff() or Put(), after we decompressed it.  If
	// UncompressedDigest is not set, this should be treated as if it were
	// an uninitialized value.
	UncompressedSize int64 `json:"diff-size,omitempty"`

	// CompressionType is the type of compression which we detected on the blob
	// that was last passed to ApplyDiff() or Put().
	CompressionType archive.Compression `json:"compression,omitempty"`

	// UIDs and GIDs are lists of UIDs and GIDs used in the layer.  This
	// field is only populated (i.e., will only contain one or more
	// entries) if the layer was created using ApplyDiff() or Put().
	UIDs []uint32 `json:"uidset,omitempty"`
	GIDs []uint32 `json:"gidset,omitempty"`

	// Flags is arbitrary data about the layer.
	Flags map[string]interface{} `json:"flags,omitempty"`

	// UIDMap and GIDMap are used for setting up a layer's contents
	// for use inside of a user namespace where UID mapping is being used.
	UIDMap []idtools.IDMap `json:"uidmap,omitempty"`
	GIDMap []idtools.IDMap `json:"gidmap,omitempty"`

	// ReadOnly is true if this layer resides in a read-only layer store.
	ReadOnly bool `json:"-"`

	// BigDataNames is a list of names of data items that we keep for the
	// convenience of the caller.  They can be large, and are only in
	// memory when being read from or written to disk.
	BigDataNames []string `json:"big-data-names,omitempty"`
}

type layerMountPoint struct {
	ID         string `json:"id"`
	MountPoint string `json:"path"`
	MountCount int    `json:"count"`
}

// DiffOptions override the default behavior of Diff() methods.
type DiffOptions struct {
	// Compression, if set overrides the default compressor when generating a diff.
	Compression *archive.Compression
}

// ROLayerStore wraps a graph driver, adding the ability to refer to layers by
// name, and keeping track of parent-child relationships, along with a list of
// all known layers.
type ROLayerStore interface {
	ROFileBasedStore
	ROMetadataStore
	ROLayerBigDataStore

	// Exists checks if a layer with the specified name or ID is known.
	Exists(id string) bool

	// Get retrieves information about a layer given an ID or name.
	Get(id string) (*Layer, error)

	// Status returns an slice of key-value pairs, suitable for human consumption,
	// relaying whatever status information the underlying driver can share.
	Status() ([][2]string, error)

	// Changes returns a slice of Change structures, which contain a pathname
	// (Path) and a description of what sort of change (Kind) was made by the
	// layer (either ChangeModify, ChangeAdd, or ChangeDelete), relative to a
	// specified layer.  By default, the layer's parent is used as a reference.
	Changes(from, to string) ([]archive.Change, error)

	// Diff produces a tarstream which can be applied to a layer with the contents
	// of the first layer to produce a layer with the contents of the second layer.
	// By default, the parent of the second layer is used as the first
	// layer, so it need not be specified.  Options can be used to override
	// default behavior, but are also not required.
	Diff(from, to string, options *DiffOptions) (io.ReadCloser, error)

	// DiffSize produces an estimate of the length of the tarstream which would be
	// produced by Diff.
	DiffSize(from, to string) (int64, error)

	// Size produces a cached value for the uncompressed size of the layer,
	// if one is known, or -1 if it is not known.  If the layer can not be
	// found, it returns an error.
	Size(name string) (int64, error)

	// Lookup attempts to translate a name to an ID.  Most methods do this
	// implicitly.
	Lookup(name string) (string, error)

	// LayersByCompressedDigest returns a slice of the layers with the
	// specified compressed digest value recorded for them.
	LayersByCompressedDigest(d digest.Digest) ([]Layer, error)

	// LayersByUncompressedDigest returns a slice of the layers with the
	// specified uncompressed digest value recorded for them.
	LayersByUncompressedDigest(d digest.Digest) ([]Layer, error)

	// Layers returns a slice of the known layers.
	Layers() ([]Layer, error)
}

// LayerStore wraps a graph driver, adding the ability to refer to layers by
// name, and keeping track of parent-child relationships, along with a list of
// all known layers.
type LayerStore interface {
	ROLayerStore
	RWFileBasedStore
	RWMetadataStore
	FlaggableStore
	RWLayerBigDataStore

	// Create creates a new layer, optionally giving it a specified ID rather than
	// a randomly-generated one, either inheriting data from another specified
	// layer or the empty base layer.  The new layer can optionally be given names
	// and have an SELinux label specified for use when mounting it.  Some
	// underlying drivers can accept a "size" option.  At this time, most
	// underlying drivers do not themselves distinguish between writeable
	// and read-only layers.
	Create(id string, parent *Layer, names []string, mountLabel string, options map[string]string, moreOptions *LayerOptions, writeable bool) (*Layer, error)

	// CreateWithFlags combines the functions of Create and SetFlag.
	CreateWithFlags(id string, parent *Layer, names []string, mountLabel string, options map[string]string, moreOptions *LayerOptions, writeable bool, flags map[string]interface{}) (layer *Layer, err error)

	// Put combines the functions of CreateWithFlags and ApplyDiff.
	Put(id string, parent *Layer, names []string, mountLabel string, options map[string]string, moreOptions *LayerOptions, writeable bool, flags map[string]interface{}, diff io.Reader) (*Layer, int64, error)

	// SetNames replaces the list of names associated with a layer with the
	// supplied values.
	// Deprecated: Prone to race conditions, suggested alternatives are `AddNames` and `RemoveNames`.
	SetNames(id string, names []string) error

	// AddNames adds the supplied values to the list of names associated with the layer with the
	// specified id.
	AddNames(id string, names []string) error

	// RemoveNames remove the supplied values from the list of names associated with the layer with the
	// specified id.
	RemoveNames(id string, names []string) error

	// Delete deletes a layer with the specified name or ID.
	Delete(id string) error

	// Wipe deletes all layers.
	Wipe() error

	// Mount mounts a layer for use.  If the specified layer is the parent of other
	// layers, it should not be written to.  An SELinux label to be applied to the
	// mount can be specified to override the one configured for the layer.
	// The mappings used by the container can be specified.
	Mount(id string, options drivers.MountOpts) (string, error)

	// Unmount unmounts a layer when it is no longer in use.
	Unmount(id string, force bool) (bool, error)

	// Mounted returns number of times the layer has been mounted.
	Mounted(id string) (int, error)

	// ParentOwners returns the UIDs and GIDs of parents of the layer's mountpoint
	// for which the layer's UID and GID maps don't contain corresponding entries.
	ParentOwners(id string) (uids, gids []int, err error)

	// ApplyDiff reads a tarstream which was created by a previous call to Diff and
	// applies its changes to a specified layer.
	ApplyDiff(to string, diff io.Reader) (int64, error)

	// ApplyDiffWithDiffer applies the changes through the differ callback function.
	// If to is the empty string, then a staging directory is created by the driver.
	ApplyDiffWithDiffer(to string, options *drivers.ApplyDiffOpts, differ drivers.Differ) (*drivers.DriverWithDifferOutput, error)

	// CleanupStagingDirectory cleanups the staging directory.  It can be used to cleanup the staging directory on errors
	CleanupStagingDirectory(stagingDirectory string) error

	// ApplyDiffFromStagingDirectory uses stagingDirectory to create the diff.
	ApplyDiffFromStagingDirectory(id, stagingDirectory string, diffOutput *drivers.DriverWithDifferOutput, options *drivers.ApplyDiffOpts) error

	// DifferTarget gets the location where files are stored for the layer.
	DifferTarget(id string) (string, error)

	// LoadLocked wraps Load in a locked state. This means it loads the store
	// and cleans-up invalid layers if needed.
	LoadLocked() error

	// PutAdditionalLayer creates a layer using the diff contained in the additional layer
	// store.
	// This API is experimental and can be changed without bumping the major version number.
	PutAdditionalLayer(id string, parentLayer *Layer, names []string, aLayer drivers.AdditionalLayer) (layer *Layer, err error)
}

type layerStore struct {
	lockfile           Locker
	mountsLockfile     Locker
	rundir             string
	driver             drivers.Driver
	layerdir           string
	layers             []*Layer
	idindex            *truncindex.TruncIndex
	byid               map[string]*Layer
	byname             map[string]*Layer
	bymount            map[string]*Layer
	bycompressedsum    map[digest.Digest][]string
	byuncompressedsum  map[digest.Digest][]string
	uidMap             []idtools.IDMap
	gidMap             []idtools.IDMap
	loadMut            sync.Mutex
	layerspathModified time.Time
}

func copyLayer(l *Layer) *Layer {
	return &Layer{
		ID:                 l.ID,
		Names:              copyStringSlice(l.Names),
		Parent:             l.Parent,
		Metadata:           l.Metadata,
		MountLabel:         l.MountLabel,
		MountPoint:         l.MountPoint,
		MountCount:         l.MountCount,
		Created:            l.Created,
		CompressedDigest:   l.CompressedDigest,
		CompressedSize:     l.CompressedSize,
		UncompressedDigest: l.UncompressedDigest,
		UncompressedSize:   l.UncompressedSize,
		CompressionType:    l.CompressionType,
		ReadOnly:           l.ReadOnly,
		BigDataNames:       copyStringSlice(l.BigDataNames),
		Flags:              copyStringInterfaceMap(l.Flags),
		UIDMap:             copyIDMap(l.UIDMap),
		GIDMap:             copyIDMap(l.GIDMap),
		UIDs:               copyUint32Slice(l.UIDs),
		GIDs:               copyUint32Slice(l.GIDs),
	}
}

func (r *layerStore) Layers() ([]Layer, error) {
	layers := make([]Layer, len(r.layers))
	for i := range r.layers {
		layers[i] = *copyLayer(r.layers[i])
	}
	return layers, nil
}

func (r *layerStore) mountspath() string {
	return filepath.Join(r.rundir, "mountpoints.json")
}

func (r *layerStore) layerspath() string {
	return filepath.Join(r.layerdir, "layers.json")
}

func (r *layerStore) Load() error {
	shouldSave := false
	rpath := r.layerspath()
	info, err := os.Stat(rpath)
	if err != nil {
		if !os.IsNotExist(err) {
			return err
		}
	} else {
		r.layerspathModified = info.ModTime()
	}
	data, err := os.ReadFile(rpath)
	if err != nil && !os.IsNotExist(err) {
		return err
	}
	layers := []*Layer{}
	idlist := []string{}
	ids := make(map[string]*Layer)
	names := make(map[string]*Layer)
	compressedsums := make(map[digest.Digest][]string)
	uncompressedsums := make(map[digest.Digest][]string)
	if r.IsReadWrite() {
		label.ClearLabels()
	}
	if err = json.Unmarshal(data, &layers); len(data) == 0 || err == nil {
		idlist = make([]string, 0, len(layers))
		for n, layer := range layers {
			ids[layer.ID] = layers[n]
			idlist = append(idlist, layer.ID)
			for _, name := range layer.Names {
				if conflict, ok := names[name]; ok {
					r.removeName(conflict, name)
					shouldSave = true
				}
				names[name] = layers[n]
			}
			if layer.CompressedDigest != "" {
				compressedsums[layer.CompressedDigest] = append(compressedsums[layer.CompressedDigest], layer.ID)
			}
			if layer.UncompressedDigest != "" {
				uncompressedsums[layer.UncompressedDigest] = append(uncompressedsums[layer.UncompressedDigest], layer.ID)
			}
			if layer.MountLabel != "" {
				label.ReserveLabel(layer.MountLabel)
			}
			layer.ReadOnly = !r.IsReadWrite()
		}
		err = nil
	}
	if shouldSave && (!r.IsReadWrite() || !r.Locked()) {
		return ErrDuplicateLayerNames
	}
	r.layers = layers
	r.idindex = truncindex.NewTruncIndex(idlist)
	r.byid = ids
	r.byname = names
	r.bycompressedsum = compressedsums
	r.byuncompressedsum = uncompressedsums

	// Load and merge information about which layers are mounted, and where.
	if r.IsReadWrite() {
		r.mountsLockfile.RLock()
		defer r.mountsLockfile.Unlock()
		if err = r.loadMounts(); err != nil {
			return err
		}

		// Last step: as we’re writable, try to remove anything that a previous
		// user of this storage area marked for deletion but didn't manage to
		// actually delete.
		if r.Locked() {
			for _, layer := range r.layers {
				if layer.Flags == nil {
					layer.Flags = make(map[string]interface{})
				}
				if layerHasIncompleteFlag(layer) {
					logrus.Warnf("Found incomplete layer %#v, deleting it", layer.ID)
					err = r.deleteInternal(layer.ID)
					if err != nil {
						break
					}
					shouldSave = true
				}
			}
		}
		if shouldSave {
			return r.saveLayers()
		}
	}

	return err
}

func (r *layerStore) LoadLocked() error {
	r.lockfile.Lock()
	defer r.lockfile.Unlock()
	return r.Load()
}

func (r *layerStore) loadMounts() error {
	mounts := make(map[string]*Layer)
	mpath := r.mountspath()
	data, err := os.ReadFile(mpath)
	if err != nil && !os.IsNotExist(err) {
		return err
	}
	layerMounts := []layerMountPoint{}
	if err = json.Unmarshal(data, &layerMounts); len(data) == 0 || err == nil {
		// Clear all of our mount information.  If another process
		// unmounted something, it (along with its zero count) won't
		// have been encoded into the version of mountpoints.json that
		// we're loading, so our count could fall out of sync with it
		// if we don't, and if we subsequently change something else,
		// we'd pass that error along to other process that reloaded
		// the data after we saved it.
		for _, layer := range r.layers {
			layer.MountPoint = ""
			layer.MountCount = 0
		}
		// All of the non-zero count values will have been encoded, so
		// we reset the still-mounted ones based on the contents.
		for _, mount := range layerMounts {
			if mount.MountPoint != "" {
				if layer, ok := r.lookup(mount.ID); ok {
					mounts[mount.MountPoint] = layer
					layer.MountPoint = mount.MountPoint
					layer.MountCount = mount.MountCount
				}
			}
		}
		err = nil
	}
	r.bymount = mounts
	return err
}

func (r *layerStore) Save() error {
	r.mountsLockfile.Lock()
	defer r.mountsLockfile.Unlock()
	defer r.mountsLockfile.Touch()
	if err := r.saveLayers(); err != nil {
		return err
	}
	return r.saveMounts()
}

func (r *layerStore) saveLayers() error {
	if !r.IsReadWrite() {
		return fmt.Errorf("not allowed to modify the layer store at %q: %w", r.layerspath(), ErrStoreIsReadOnly)
	}
	if !r.Locked() {
		return errors.New("layer store is not locked for writing")
	}
	rpath := r.layerspath()
	if err := os.MkdirAll(filepath.Dir(rpath), 0700); err != nil {
		return err
	}
	jldata, err := json.Marshal(&r.layers)
	if err != nil {
		return err
	}
	defer r.Touch()
	return ioutils.AtomicWriteFile(rpath, jldata, 0600)
}

func (r *layerStore) saveMounts() error {
	if !r.IsReadWrite() {
		return fmt.Errorf("not allowed to modify the layer store at %q: %w", r.layerspath(), ErrStoreIsReadOnly)
	}
	if !r.mountsLockfile.Locked() {
		return errors.New("layer store mount information is not locked for writing")
	}
	mpath := r.mountspath()
	if err := os.MkdirAll(filepath.Dir(mpath), 0700); err != nil {
		return err
	}
	mounts := make([]layerMountPoint, 0, len(r.layers))
	for _, layer := range r.layers {
		if layer.MountPoint != "" && layer.MountCount > 0 {
			mounts = append(mounts, layerMountPoint{
				ID:         layer.ID,
				MountPoint: layer.MountPoint,
				MountCount: layer.MountCount,
			})
		}
	}
	jmdata, err := json.Marshal(&mounts)
	if err != nil {
		return err
	}
	if err = ioutils.AtomicWriteFile(mpath, jmdata, 0600); err != nil {
		return err
	}
	return r.loadMounts()
}

func (s *store) newLayerStore(rundir string, layerdir string, driver drivers.Driver) (LayerStore, error) {
	if err := os.MkdirAll(rundir, 0700); err != nil {
		return nil, err
	}
	if err := os.MkdirAll(layerdir, 0700); err != nil {
		return nil, err
	}
	lockfile, err := GetLockfile(filepath.Join(layerdir, "layers.lock"))
	if err != nil {
		return nil, err
	}
	mountsLockfile, err := GetLockfile(filepath.Join(rundir, "mountpoints.lock"))
	if err != nil {
		return nil, err
	}
	rlstore := layerStore{
		lockfile:       lockfile,
		mountsLockfile: mountsLockfile,
		driver:         driver,
		rundir:         rundir,
		layerdir:       layerdir,
		byid:           make(map[string]*Layer),
		bymount:        make(map[string]*Layer),
		byname:         make(map[string]*Layer),
		uidMap:         copyIDMap(s.uidMap),
		gidMap:         copyIDMap(s.gidMap),
	}
	rlstore.Lock()
	defer rlstore.Unlock()
	if err := rlstore.Load(); err != nil {
		return nil, err
	}
	return &rlstore, nil
}

func newROLayerStore(rundir string, layerdir string, driver drivers.Driver) (ROLayerStore, error) {
	lockfile, err := GetROLockfile(filepath.Join(layerdir, "layers.lock"))
	if err != nil {
		return nil, err
	}
	rlstore := layerStore{
		lockfile:       lockfile,
		mountsLockfile: nil,
		driver:         driver,
		rundir:         rundir,
		layerdir:       layerdir,
		byid:           make(map[string]*Layer),
		bymount:        make(map[string]*Layer),
		byname:         make(map[string]*Layer),
	}
	rlstore.RLock()
	defer rlstore.Unlock()
	if err := rlstore.Load(); err != nil {
		return nil, err
	}
	return &rlstore, nil
}

func (r *layerStore) lookup(id string) (*Layer, bool) {
	if layer, ok := r.byid[id]; ok {
		return layer, ok
	} else if layer, ok := r.byname[id]; ok {
		return layer, ok
	} else if longid, err := r.idindex.Get(id); err == nil {
		layer, ok := r.byid[longid]
		return layer, ok
	}
	return nil, false
}

func (r *layerStore) Size(name string) (int64, error) {
	layer, ok := r.lookup(name)
	if !ok {
		return -1, ErrLayerUnknown
	}
	// We use the presence of a non-empty digest as an indicator that the size value was intentionally set, and that
	// a zero value is not just present because it was never set to anything else (which can happen if the layer was
	// created by a version of this library that didn't keep track of digest and size information).
	if layer.UncompressedDigest != "" {
		return layer.UncompressedSize, nil
	}
	return -1, nil
}

func (r *layerStore) ClearFlag(id string, flag string) error {
	if !r.IsReadWrite() {
		return fmt.Errorf("not allowed to clear flags on layers at %q: %w", r.layerspath(), ErrStoreIsReadOnly)
	}
	layer, ok := r.lookup(id)
	if !ok {
		return ErrLayerUnknown
	}
	delete(layer.Flags, flag)
	return r.Save()
}

func (r *layerStore) SetFlag(id string, flag string, value interface{}) error {
	if !r.IsReadWrite() {
		return fmt.Errorf("not allowed to set flags on layers at %q: %w", r.layerspath(), ErrStoreIsReadOnly)
	}
	layer, ok := r.lookup(id)
	if !ok {
		return ErrLayerUnknown
	}
	if layer.Flags == nil {
		layer.Flags = make(map[string]interface{})
	}
	layer.Flags[flag] = value
	return r.Save()
}

func (r *layerStore) Status() ([][2]string, error) {
	return r.driver.Status(), nil
}

func (r *layerStore) PutAdditionalLayer(id string, parentLayer *Layer, names []string, aLayer drivers.AdditionalLayer) (layer *Layer, err error) {
	if duplicateLayer, idInUse := r.byid[id]; idInUse {
		return duplicateLayer, ErrDuplicateID
	}
	for _, name := range names {
		if _, nameInUse := r.byname[name]; nameInUse {
			return nil, ErrDuplicateName
		}
	}

	parent := ""
	if parentLayer != nil {
		parent = parentLayer.ID
	}

	info, err := aLayer.Info()
	if err != nil {
		return nil, err
	}
	defer info.Close()
	layer = &Layer{}
	if err := json.NewDecoder(info).Decode(layer); err != nil {
		return nil, err
	}
	layer.ID = id
	layer.Parent = parent
	layer.Created = time.Now().UTC()

	if err := aLayer.CreateAs(id, parent); err != nil {
		return nil, err
	}

	// TODO: check if necessary fields are filled
	r.layers = append(r.layers, layer)
	r.idindex.Add(id)
	r.byid[id] = layer
	for _, name := range names { // names got from the additional layer store won't be used
		r.byname[name] = layer
	}
	if layer.CompressedDigest != "" {
		r.bycompressedsum[layer.CompressedDigest] = append(r.bycompressedsum[layer.CompressedDigest], layer.ID)
	}
	if layer.UncompressedDigest != "" {
		r.byuncompressedsum[layer.UncompressedDigest] = append(r.byuncompressedsum[layer.UncompressedDigest], layer.ID)
	}
	if err := r.Save(); err != nil {
		r.driver.Remove(id)
		return nil, err
	}
	return copyLayer(layer), nil
}

func (r *layerStore) Put(id string, parentLayer *Layer, names []string, mountLabel string, options map[string]string, moreOptions *LayerOptions, writeable bool, flags map[string]interface{}, diff io.Reader) (*Layer, int64, error) {
	if !r.IsReadWrite() {
		return nil, -1, fmt.Errorf("not allowed to create new layers at %q: %w", r.layerspath(), ErrStoreIsReadOnly)
	}
	if err := os.MkdirAll(r.rundir, 0700); err != nil {
		return nil, -1, err
	}
	if err := os.MkdirAll(r.layerdir, 0700); err != nil {
		return nil, -1, err
	}
	if id == "" {
		id = stringid.GenerateRandomID()
		_, idInUse := r.byid[id]
		for idInUse {
			id = stringid.GenerateRandomID()
			_, idInUse = r.byid[id]
		}
	}
	if duplicateLayer, idInUse := r.byid[id]; idInUse {
		return duplicateLayer, -1, ErrDuplicateID
	}
	names = dedupeNames(names)
	for _, name := range names {
		if _, nameInUse := r.byname[name]; nameInUse {
			return nil, -1, ErrDuplicateName
		}
	}
	parent := ""
	if parentLayer != nil {
		parent = parentLayer.ID
	}
	var parentMappings, templateIDMappings, oldMappings *idtools.IDMappings
	var (
		templateMetadata           string
		templateCompressedDigest   digest.Digest
		templateCompressedSize     int64
		templateUncompressedDigest digest.Digest
		templateUncompressedSize   int64
		templateCompressionType    archive.Compression
		templateUIDs, templateGIDs []uint32
		templateTSdata             []byte
	)
	if moreOptions.TemplateLayer != "" {
		var tserr error
		templateLayer, ok := r.lookup(moreOptions.TemplateLayer)
		if !ok {
			return nil, -1, ErrLayerUnknown
		}
		templateMetadata = templateLayer.Metadata
		templateIDMappings = idtools.NewIDMappingsFromMaps(templateLayer.UIDMap, templateLayer.GIDMap)
		templateCompressedDigest, templateCompressedSize = templateLayer.CompressedDigest, templateLayer.CompressedSize
		templateUncompressedDigest, templateUncompressedSize = templateLayer.UncompressedDigest, templateLayer.UncompressedSize
		templateCompressionType = templateLayer.CompressionType
		templateUIDs, templateGIDs = append([]uint32{}, templateLayer.UIDs...), append([]uint32{}, templateLayer.GIDs...)
		templateTSdata, tserr = os.ReadFile(r.tspath(templateLayer.ID))
		if tserr != nil && !os.IsNotExist(tserr) {
			return nil, -1, tserr
		}
	} else {
		templateIDMappings = &idtools.IDMappings{}
	}
	if parentLayer != nil {
		parentMappings = idtools.NewIDMappingsFromMaps(parentLayer.UIDMap, parentLayer.GIDMap)
	} else {
		parentMappings = &idtools.IDMappings{}
	}
	if mountLabel != "" {
		label.ReserveLabel(mountLabel)
	}

	// Before actually creating the layer, make a persistent record of it with incompleteFlag,
	// so that future processes have a chance to delete it.
	layer := &Layer{
		ID:                 id,
		Parent:             parent,
		Names:              names,
		MountLabel:         mountLabel,
		Metadata:           templateMetadata,
		Created:            time.Now().UTC(),
		CompressedDigest:   templateCompressedDigest,
		CompressedSize:     templateCompressedSize,
		UncompressedDigest: templateUncompressedDigest,
		UncompressedSize:   templateUncompressedSize,
		CompressionType:    templateCompressionType,
		UIDs:               templateUIDs,
		GIDs:               templateGIDs,
		Flags:              make(map[string]interface{}),
		UIDMap:             copyIDMap(moreOptions.UIDMap),
		GIDMap:             copyIDMap(moreOptions.GIDMap),
		BigDataNames:       []string{},
	}
	r.layers = append(r.layers, layer)
	r.idindex.Add(id)
	r.byid[id] = layer
	for _, name := range names {
		r.byname[name] = layer
	}
	for flag, value := range flags {
		layer.Flags[flag] = value
	}
	layer.Flags[incompleteFlag] = true

	succeeded := false
	cleanupFailureContext := ""
	defer func() {
		if !succeeded {
			// On any error, try both removing the driver's data as well
			// as the in-memory layer record.
			if err2 := r.Delete(layer.ID); err2 != nil {
				if cleanupFailureContext == "" {
					cleanupFailureContext = "unknown: cleanupFailureContext not set at the failure site"
				}
				logrus.Errorf("While recovering from a failure (%s), error deleting layer %#v: %v", cleanupFailureContext, layer.ID, err2)
			}
		}
	}()

	err := r.Save()
	if err != nil {
		cleanupFailureContext = "saving incomplete layer metadata"
		return nil, -1, err
	}

	idMappings := idtools.NewIDMappingsFromMaps(moreOptions.UIDMap, moreOptions.GIDMap)
	opts := drivers.CreateOpts{
		MountLabel: mountLabel,
		StorageOpt: options,
		IDMappings: idMappings,
	}
	if moreOptions.TemplateLayer != "" {
		if err := r.driver.CreateFromTemplate(id, moreOptions.TemplateLayer, templateIDMappings, parent, parentMappings, &opts, writeable); err != nil {
			cleanupFailureContext = "creating a layer from template"
			return nil, -1, fmt.Errorf("creating copy of template layer %q with ID %q: %w", moreOptions.TemplateLayer, id, err)
		}
		oldMappings = templateIDMappings
	} else {
		if writeable {
			if err := r.driver.CreateReadWrite(id, parent, &opts); err != nil {
				cleanupFailureContext = "creating a read-write layer"
				return nil, -1, fmt.Errorf("creating read-write layer with ID %q: %w", id, err)
			}
		} else {
			if err := r.driver.Create(id, parent, &opts); err != nil {
				cleanupFailureContext = "creating a read-only layer"
				return nil, -1, fmt.Errorf("creating layer with ID %q: %w", id, err)
			}
		}
		oldMappings = parentMappings
	}
	if !reflect.DeepEqual(oldMappings.UIDs(), idMappings.UIDs()) || !reflect.DeepEqual(oldMappings.GIDs(), idMappings.GIDs()) {
		if err := r.driver.UpdateLayerIDMap(id, oldMappings, idMappings, mountLabel); err != nil {
			cleanupFailureContext = "in UpdateLayerIDMap"
			return nil, -1, err
		}
	}
	if len(templateTSdata) > 0 {
		if err := os.MkdirAll(filepath.Dir(r.tspath(id)), 0o700); err != nil {
			cleanupFailureContext = "creating tar-split parent directory for a copy from template"
			return nil, -1, err
		}
		if err := ioutils.AtomicWriteFile(r.tspath(id), templateTSdata, 0o600); err != nil {
			cleanupFailureContext = "creating a tar-split copy from template"
			return nil, -1, err
		}
	}

	var size int64 = -1
	if diff != nil {
		size, err = r.applyDiffWithOptions(layer.ID, moreOptions, diff)
		if err != nil {
			cleanupFailureContext = "applying layer diff"
			return nil, -1, err
		}
	} else {
		// applyDiffWithOptions in the `diff != nil` case handles this bit for us
		if layer.CompressedDigest != "" {
			r.bycompressedsum[layer.CompressedDigest] = append(r.bycompressedsum[layer.CompressedDigest], layer.ID)
		}
		if layer.UncompressedDigest != "" {
			r.byuncompressedsum[layer.UncompressedDigest] = append(r.byuncompressedsum[layer.UncompressedDigest], layer.ID)
		}
	}
	delete(layer.Flags, incompleteFlag)
	err = r.Save()
	if err != nil {
		cleanupFailureContext = "saving finished layer metadata"
		return nil, -1, err
	}

	layer = copyLayer(layer)
	succeeded = true
	return layer, size, err
}

func (r *layerStore) CreateWithFlags(id string, parent *Layer, names []string, mountLabel string, options map[string]string, moreOptions *LayerOptions, writeable bool, flags map[string]interface{}) (layer *Layer, err error) {
	layer, _, err = r.Put(id, parent, names, mountLabel, options, moreOptions, writeable, flags, nil)
	return layer, err
}

func (r *layerStore) Create(id string, parent *Layer, names []string, mountLabel string, options map[string]string, moreOptions *LayerOptions, writeable bool) (layer *Layer, err error) {
	return r.CreateWithFlags(id, parent, names, mountLabel, options, moreOptions, writeable, nil)
}

func (r *layerStore) Mounted(id string) (int, error) {
	if !r.IsReadWrite() {
		return 0, fmt.Errorf("no mount information for layers at %q: %w", r.mountspath(), ErrStoreIsReadOnly)
	}
	r.mountsLockfile.RLock()
	defer r.mountsLockfile.Unlock()
	if modified, err := r.mountsLockfile.Modified(); modified || err != nil {
		if err = r.loadMounts(); err != nil {
			return 0, err
		}
	}
	layer, ok := r.lookup(id)
	if !ok {
		return 0, ErrLayerUnknown
	}
	return layer.MountCount, nil
}

func (r *layerStore) Mount(id string, options drivers.MountOpts) (string, error) {
	// check whether options include ro option
	hasReadOnlyOpt := func(opts []string) bool {
		for _, item := range opts {
			if item == "ro" {
				return true
			}
		}
		return false
	}

	// You are not allowed to mount layers from readonly stores if they
	// are not mounted read/only.
	if !r.IsReadWrite() && !hasReadOnlyOpt(options.Options) {
		return "", fmt.Errorf("not allowed to update mount locations for layers at %q: %w", r.mountspath(), ErrStoreIsReadOnly)
	}
	r.mountsLockfile.Lock()
	defer r.mountsLockfile.Unlock()
	if modified, err := r.mountsLockfile.Modified(); modified || err != nil {
		if err = r.loadMounts(); err != nil {
			return "", err
		}
	}
	defer r.mountsLockfile.Touch()
	layer, ok := r.lookup(id)
	if !ok {
		return "", ErrLayerUnknown
	}
	if layer.MountCount > 0 {
		mounted, err := mount.Mounted(layer.MountPoint)
		if err != nil {
			return "", err
		}
		// If the container is not mounted then we have a condition
		// where the kernel umounted the mount point. This means
		// that the mount count never got decremented.
		if mounted {
			layer.MountCount++
			return layer.MountPoint, r.saveMounts()
		}
	}
	if options.MountLabel == "" {
		options.MountLabel = layer.MountLabel
	}

	if (options.UidMaps != nil || options.GidMaps != nil) && !r.driver.SupportsShifting() {
		if !reflect.DeepEqual(options.UidMaps, layer.UIDMap) || !reflect.DeepEqual(options.GidMaps, layer.GIDMap) {
			return "", fmt.Errorf("cannot mount layer %v: shifting not enabled", layer.ID)
		}
	}
	mountpoint, err := r.driver.Get(id, options)
	if mountpoint != "" && err == nil {
		if layer.MountPoint != "" {
			delete(r.bymount, layer.MountPoint)
		}
		layer.MountPoint = filepath.Clean(mountpoint)
		layer.MountCount++
		r.bymount[layer.MountPoint] = layer
		err = r.saveMounts()
	}
	return mountpoint, err
}

func (r *layerStore) Unmount(id string, force bool) (bool, error) {
	if !r.IsReadWrite() {
		return false, fmt.Errorf("not allowed to update mount locations for layers at %q: %w", r.mountspath(), ErrStoreIsReadOnly)
	}
	r.mountsLockfile.Lock()
	defer r.mountsLockfile.Unlock()
	if modified, err := r.mountsLockfile.Modified(); modified || err != nil {
		if err = r.loadMounts(); err != nil {
			return false, err
		}
	}
	defer r.mountsLockfile.Touch()
	layer, ok := r.lookup(id)
	if !ok {
		layerByMount, ok := r.bymount[filepath.Clean(id)]
		if !ok {
			return false, ErrLayerUnknown
		}
		layer = layerByMount
	}
	if force {
		layer.MountCount = 1
	}
	if layer.MountCount > 1 {
		layer.MountCount--
		return true, r.saveMounts()
	}
	err := r.driver.Put(id)
	if err == nil || os.IsNotExist(err) {
		if layer.MountPoint != "" {
			delete(r.bymount, layer.MountPoint)
		}
		layer.MountCount--
		layer.MountPoint = ""
		return false, r.saveMounts()
	}
	return true, err
}

func (r *layerStore) ParentOwners(id string) (uids, gids []int, err error) {
	if !r.IsReadWrite() {
		return nil, nil, fmt.Errorf("no mount information for layers at %q: %w", r.mountspath(), ErrStoreIsReadOnly)
	}
	r.mountsLockfile.RLock()
	defer r.mountsLockfile.Unlock()
	if modified, err := r.mountsLockfile.Modified(); modified || err != nil {
		if err = r.loadMounts(); err != nil {
			return nil, nil, err
		}
	}
	layer, ok := r.lookup(id)
	if !ok {
		return nil, nil, ErrLayerUnknown
	}
	if len(layer.UIDMap) == 0 && len(layer.GIDMap) == 0 {
		// We're not using any mappings, so there aren't any unmapped IDs on parent directories.
		return nil, nil, nil
	}
	if layer.MountPoint == "" {
		// We don't know which directories to examine.
		return nil, nil, ErrLayerNotMounted
	}
	rootuid, rootgid, err := idtools.GetRootUIDGID(layer.UIDMap, layer.GIDMap)
	if err != nil {
		return nil, nil, fmt.Errorf("reading root ID values for layer %q: %w", layer.ID, err)
	}
	m := idtools.NewIDMappingsFromMaps(layer.UIDMap, layer.GIDMap)
	fsuids := make(map[int]struct{})
	fsgids := make(map[int]struct{})
	for dir := filepath.Dir(layer.MountPoint); dir != "" && dir != string(os.PathSeparator); dir = filepath.Dir(dir) {
		st, err := system.Stat(dir)
		if err != nil {
			return nil, nil, fmt.Errorf("read directory ownership: %w", err)
		}
		lst, err := system.Lstat(dir)
		if err != nil {
			return nil, nil, err
		}
		fsuid := int(st.UID())
		fsgid := int(st.GID())
		if _, _, err := m.ToContainer(idtools.IDPair{UID: fsuid, GID: rootgid}); err != nil {
			fsuids[fsuid] = struct{}{}
		}
		if _, _, err := m.ToContainer(idtools.IDPair{UID: rootuid, GID: fsgid}); err != nil {
			fsgids[fsgid] = struct{}{}
		}
		fsuid = int(lst.UID())
		fsgid = int(lst.GID())
		if _, _, err := m.ToContainer(idtools.IDPair{UID: fsuid, GID: rootgid}); err != nil {
			fsuids[fsuid] = struct{}{}
		}
		if _, _, err := m.ToContainer(idtools.IDPair{UID: rootuid, GID: fsgid}); err != nil {
			fsgids[fsgid] = struct{}{}
		}
	}
	for uid := range fsuids {
		uids = append(uids, uid)
	}
	for gid := range fsgids {
		gids = append(gids, gid)
	}
	if len(uids) > 1 {
		sort.Ints(uids)
	}
	if len(gids) > 1 {
		sort.Ints(gids)
	}
	return uids, gids, nil
}

func (r *layerStore) removeName(layer *Layer, name string) {
	layer.Names = stringSliceWithoutValue(layer.Names, name)
}

// Deprecated: Prone to race conditions, suggested alternatives are `AddNames` and `RemoveNames`.
func (r *layerStore) SetNames(id string, names []string) error {
	return r.updateNames(id, names, setNames)
}

func (r *layerStore) AddNames(id string, names []string) error {
	return r.updateNames(id, names, addNames)
}

func (r *layerStore) RemoveNames(id string, names []string) error {
	return r.updateNames(id, names, removeNames)
}

func (r *layerStore) updateNames(id string, names []string, op updateNameOperation) error {
	if !r.IsReadWrite() {
		return fmt.Errorf("not allowed to change layer name assignments at %q: %w", r.layerspath(), ErrStoreIsReadOnly)
	}
	layer, ok := r.lookup(id)
	if !ok {
		return ErrLayerUnknown
	}
	oldNames := layer.Names
	names, err := applyNameOperation(oldNames, names, op)
	if err != nil {
		return err
	}
	for _, name := range oldNames {
		delete(r.byname, name)
	}
	for _, name := range names {
		if otherLayer, ok := r.byname[name]; ok {
			r.removeName(otherLayer, name)
		}
		r.byname[name] = layer
	}
	layer.Names = names
	return r.Save()
}

func (r *layerStore) datadir(id string) string {
	return filepath.Join(r.layerdir, id)
}

func (r *layerStore) datapath(id, key string) string {
	return filepath.Join(r.datadir(id), makeBigDataBaseName(key))
}

func (r *layerStore) BigData(id, key string) (io.ReadCloser, error) {
	if key == "" {
		return nil, fmt.Errorf("can't retrieve layer big data value for empty name: %w", ErrInvalidBigDataName)
	}
	layer, ok := r.lookup(id)
	if !ok {
		return nil, fmt.Errorf("locating layer with ID %q: %w", id, ErrLayerUnknown)
	}
	return os.Open(r.datapath(layer.ID, key))
}

func (r *layerStore) SetBigData(id, key string, data io.Reader) error {
	if key == "" {
		return fmt.Errorf("can't set empty name for layer big data item: %w", ErrInvalidBigDataName)
	}
	if !r.IsReadWrite() {
		return fmt.Errorf("not allowed to save data items associated with layers at %q: %w", r.layerspath(), ErrStoreIsReadOnly)
	}
	layer, ok := r.lookup(id)
	if !ok {
		return fmt.Errorf("locating layer with ID %q to write bigdata: %w", id, ErrLayerUnknown)
	}
	err := os.MkdirAll(r.datadir(layer.ID), 0700)
	if err != nil {
		return err
	}

	// NewAtomicFileWriter doesn't overwrite/truncate the existing inode.
	// BigData() relies on this behaviour when opening the file for read
	// so that it is either accessing the old data or the new one.
	writer, err := ioutils.NewAtomicFileWriter(r.datapath(layer.ID, key), 0600)
	if err != nil {
		return fmt.Errorf("opening bigdata file: %w", err)
	}

	if _, err := io.Copy(writer, data); err != nil {
		writer.Close()
		return fmt.Errorf("copying bigdata for the layer: %w", err)

	}
	if err := writer.Close(); err != nil {
		return fmt.Errorf("closing bigdata file for the layer: %w", err)
	}

	addName := true
	for _, name := range layer.BigDataNames {
		if name == key {
			addName = false
			break
		}
	}
	if addName {
		layer.BigDataNames = append(layer.BigDataNames, key)
		return r.Save()
	}
	return nil
}

func (r *layerStore) BigDataNames(id string) ([]string, error) {
	layer, ok := r.lookup(id)
	if !ok {
		return nil, fmt.Errorf("locating layer with ID %q to retrieve bigdata names: %w", id, ErrImageUnknown)
	}
	return copyStringSlice(layer.BigDataNames), nil
}

func (r *layerStore) Metadata(id string) (string, error) {
	if layer, ok := r.lookup(id); ok {
		return layer.Metadata, nil
	}
	return "", ErrLayerUnknown
}

func (r *layerStore) SetMetadata(id, metadata string) error {
	if !r.IsReadWrite() {
		return fmt.Errorf("not allowed to modify layer metadata at %q: %w", r.layerspath(), ErrStoreIsReadOnly)
	}
	if layer, ok := r.lookup(id); ok {
		layer.Metadata = metadata
		return r.Save()
	}
	return ErrLayerUnknown
}

func (r *layerStore) tspath(id string) string {
	return filepath.Join(r.layerdir, id+tarSplitSuffix)
}

// layerHasIncompleteFlag returns true if layer.Flags contains an incompleteFlag set to true
func layerHasIncompleteFlag(layer *Layer) bool {
	// layer.Flags[…] is defined to succeed and return ok == false if Flags == nil
	if flagValue, ok := layer.Flags[incompleteFlag]; ok {
		if b, ok := flagValue.(bool); ok && b {
			return true
		}
	}
	return false
}

func (r *layerStore) deleteInternal(id string) error {
	if !r.IsReadWrite() {
		return fmt.Errorf("not allowed to delete layers at %q: %w", r.layerspath(), ErrStoreIsReadOnly)
	}
	layer, ok := r.lookup(id)
	if !ok {
		return ErrLayerUnknown
	}
	// Ensure that if we are interrupted, the layer will be cleaned up.
	if !layerHasIncompleteFlag(layer) {
		if layer.Flags == nil {
			layer.Flags = make(map[string]interface{})
		}
		layer.Flags[incompleteFlag] = true
		if err := r.Save(); err != nil {
			return err
		}
	}
	// We never unset incompleteFlag; below, we remove the entire object from r.layers.

	id = layer.ID
	err := r.driver.Remove(id)
	if err != nil {
		return err
	}

	os.Remove(r.tspath(id))
	os.RemoveAll(r.datadir(id))
	delete(r.byid, id)
	for _, name := range layer.Names {
		delete(r.byname, name)
	}
	r.idindex.Delete(id)
	mountLabel := layer.MountLabel
	if layer.MountPoint != "" {
		delete(r.bymount, layer.MountPoint)
	}
	r.deleteInDigestMap(id)
	toDeleteIndex := -1
	for i, candidate := range r.layers {
		if candidate.ID == id {
			toDeleteIndex = i
			break
		}
	}
	if toDeleteIndex != -1 {
		// delete the layer at toDeleteIndex
		if toDeleteIndex == len(r.layers)-1 {
			r.layers = r.layers[:len(r.layers)-1]
		} else {
			r.layers = append(r.layers[:toDeleteIndex], r.layers[toDeleteIndex+1:]...)
		}
	}
	if mountLabel != "" {
		var found bool
		for _, candidate := range r.layers {
			if candidate.MountLabel == mountLabel {
				found = true
				break
			}
		}
		if !found {
			label.ReleaseLabel(mountLabel)
		}
	}

	return nil
}

func (r *layerStore) deleteInDigestMap(id string) {
	for digest, layers := range r.bycompressedsum {
		for i, layerID := range layers {
			if layerID == id {
				layers = append(layers[:i], layers[i+1:]...)
				r.bycompressedsum[digest] = layers
				break
			}
		}
	}
	for digest, layers := range r.byuncompressedsum {
		for i, layerID := range layers {
			if layerID == id {
				layers = append(layers[:i], layers[i+1:]...)
				r.byuncompressedsum[digest] = layers
				break
			}
		}
	}
}

func (r *layerStore) Delete(id string) error {
	layer, ok := r.lookup(id)
	if !ok {
		return ErrLayerUnknown
	}
	id = layer.ID
	// The layer may already have been explicitly unmounted, but if not, we
	// should try to clean that up before we start deleting anything at the
	// driver level.
	mountCount, err := r.Mounted(id)
	if err != nil {
		return fmt.Errorf("checking if layer %q is still mounted: %w", id, err)
	}
	for mountCount > 0 {
		if _, err := r.Unmount(id, false); err != nil {
			return err
		}
		mountCount, err = r.Mounted(id)
		if err != nil {
			return fmt.Errorf("checking if layer %q is still mounted: %w", id, err)
		}
	}
	if err := r.deleteInternal(id); err != nil {
		return err
	}
	return r.Save()
}

func (r *layerStore) Lookup(name string) (id string, err error) {
	if layer, ok := r.lookup(name); ok {
		return layer.ID, nil
	}
	return "", ErrLayerUnknown
}

func (r *layerStore) Exists(id string) bool {
	_, ok := r.lookup(id)
	return ok
}

func (r *layerStore) Get(id string) (*Layer, error) {
	if layer, ok := r.lookup(id); ok {
		return copyLayer(layer), nil
	}
	return nil, ErrLayerUnknown
}

func (r *layerStore) Wipe() error {
	if !r.IsReadWrite() {
		return fmt.Errorf("not allowed to delete layers at %q: %w", r.layerspath(), ErrStoreIsReadOnly)
	}
	ids := make([]string, 0, len(r.byid))
	for id := range r.byid {
		ids = append(ids, id)
	}
	sort.Slice(ids, func(i, j int) bool {
		return r.byid[ids[i]].Created.After(r.byid[ids[j]].Created)
	})
	for _, id := range ids {
		if err := r.Delete(id); err != nil {
			return err
		}
	}
	return nil
}

func (r *layerStore) findParentAndLayer(from, to string) (fromID string, toID string, fromLayer, toLayer *Layer, err error) {
	var ok bool
	toLayer, ok = r.lookup(to)
	if !ok {
		return "", "", nil, nil, ErrLayerUnknown
	}
	to = toLayer.ID
	if from == "" {
		from = toLayer.Parent
	}
	if from != "" {
		fromLayer, ok = r.lookup(from)
		if ok {
			from = fromLayer.ID
		} else {
			fromLayer, ok = r.lookup(toLayer.Parent)
			if ok {
				from = fromLayer.ID
			}
		}
	}
	return from, to, fromLayer, toLayer, nil
}

func (r *layerStore) layerMappings(layer *Layer) *idtools.IDMappings {
	if layer == nil {
		return &idtools.IDMappings{}
	}
	return idtools.NewIDMappingsFromMaps(layer.UIDMap, layer.GIDMap)
}

func (r *layerStore) Changes(from, to string) ([]archive.Change, error) {
	from, to, fromLayer, toLayer, err := r.findParentAndLayer(from, to)
	if err != nil {
		return nil, ErrLayerUnknown
	}
	return r.driver.Changes(to, r.layerMappings(toLayer), from, r.layerMappings(fromLayer), toLayer.MountLabel)
}

type simpleGetCloser struct {
	r    *layerStore
	path string
	id   string
}

func (s *simpleGetCloser) Get(path string) (io.ReadCloser, error) {
	return os.Open(filepath.Join(s.path, path))
}

func (s *simpleGetCloser) Close() error {
	_, err := s.r.Unmount(s.id, false)
	return err
}

func (r *layerStore) newFileGetter(id string) (drivers.FileGetCloser, error) {
	if getter, ok := r.driver.(drivers.DiffGetterDriver); ok {
		return getter.DiffGetter(id)
	}
	path, err := r.Mount(id, drivers.MountOpts{})
	if err != nil {
		return nil, err
	}
	return &simpleGetCloser{
		r:    r,
		path: path,
		id:   id,
	}, nil
}

func (r *layerStore) Diff(from, to string, options *DiffOptions) (io.ReadCloser, error) {
	var metadata storage.Unpacker

	from, to, fromLayer, toLayer, err := r.findParentAndLayer(from, to)
	if err != nil {
		return nil, ErrLayerUnknown
	}
	// Default to applying the type of compression that we noted was used
	// for the layerdiff when it was applied.
	compression := toLayer.CompressionType
	// If a particular compression type (or no compression) was selected,
	// use that instead.
	if options != nil && options.Compression != nil {
		compression = *options.Compression
	}
	maybeCompressReadCloser := func(rc io.ReadCloser) (io.ReadCloser, error) {
		// Depending on whether or not compression is desired, return either the
		// passed-in ReadCloser, or a new one that provides its readers with a
		// compressed version of the data that the original would have provided
		// to its readers.
		if compression == archive.Uncompressed {
			return rc, nil
		}
		preader, pwriter := io.Pipe()
		compressor, err := archive.CompressStream(pwriter, compression)
		if err != nil {
			rc.Close()
			pwriter.Close()
			preader.Close()
			return nil, err
		}
		go func() {
			defer pwriter.Close()
			defer compressor.Close()
			defer rc.Close()
			io.Copy(compressor, rc)
		}()
		return preader, nil
	}

	if from != toLayer.Parent {
		diff, err := r.driver.Diff(to, r.layerMappings(toLayer), from, r.layerMappings(fromLayer), toLayer.MountLabel)
		if err != nil {
			return nil, err
		}
		return maybeCompressReadCloser(diff)
	}

	if ad, ok := r.driver.(drivers.AdditionalLayerStoreDriver); ok {
		if aLayer, err := ad.LookupAdditionalLayerByID(to); err == nil {
			// This is an additional layer. We leverage blob API for acquiring the reproduced raw blob.
			info, err := aLayer.Info()
			if err != nil {
				aLayer.Release()
				return nil, err
			}
			defer info.Close()
			layer := &Layer{}
			if err := json.NewDecoder(info).Decode(layer); err != nil {
				aLayer.Release()
				return nil, err
			}
			blob, err := aLayer.Blob()
			if err != nil {
				aLayer.Release()
				return nil, err
			}
			// If layer compression type is different from the expected one, decompress and convert it.
			if compression != layer.CompressionType {
				diff, err := archive.DecompressStream(blob)
				if err != nil {
					if err2 := blob.Close(); err2 != nil {
						err = fmt.Errorf("failed to close blob file: %v: %w", err2, err)
					}
					aLayer.Release()
					return nil, err
				}
				rc, err := maybeCompressReadCloser(diff)
				if err != nil {
					if err2 := closeAll(blob.Close, diff.Close); err2 != nil {
						err = fmt.Errorf("failed to cleanup: %v: %w", err2, err)
					}
					aLayer.Release()
					return nil, err
				}
				return ioutils.NewReadCloserWrapper(rc, func() error {
					defer aLayer.Release()
					return closeAll(blob.Close, rc.Close)
				}), nil
			}
			return ioutils.NewReadCloserWrapper(blob, func() error { defer aLayer.Release(); return blob.Close() }), nil
		}
	}

	tsfile, err := os.Open(r.tspath(to))
	if err != nil {
		if !os.IsNotExist(err) {
			return nil, err
		}
		diff, err := r.driver.Diff(to, r.layerMappings(toLayer), from, r.layerMappings(fromLayer), toLayer.MountLabel)
		if err != nil {
			return nil, err
		}
		return maybeCompressReadCloser(diff)
	}

	decompressor, err := pgzip.NewReader(tsfile)
	if err != nil {
		if e := tsfile.Close(); e != nil {
			logrus.Debug(e)
		}
		return nil, err
	}

	metadata = storage.NewJSONUnpacker(decompressor)

	fgetter, err := r.newFileGetter(to)
	if err != nil {
		errs := multierror.Append(nil, fmt.Errorf("creating file-getter: %w", err))
		if err := decompressor.Close(); err != nil {
			errs = multierror.Append(errs, fmt.Errorf("closing decompressor: %w", err))
		}
		if err := tsfile.Close(); err != nil {
			errs = multierror.Append(errs, fmt.Errorf("closing tarstream headers: %w", err))
		}
		return nil, errs.ErrorOrNil()
	}

	tarstream := asm.NewOutputTarStream(fgetter, metadata)
	rc := ioutils.NewReadCloserWrapper(tarstream, func() error {
		var errs *multierror.Error
		if err := decompressor.Close(); err != nil {
			errs = multierror.Append(errs, fmt.Errorf("closing decompressor: %w", err))
		}
		if err := tsfile.Close(); err != nil {
			errs = multierror.Append(errs, fmt.Errorf("closing tarstream headers: %w", err))
		}
		if err := tarstream.Close(); err != nil {
			errs = multierror.Append(errs, fmt.Errorf("closing reconstructed tarstream: %w", err))
		}
		if err := fgetter.Close(); err != nil {
			errs = multierror.Append(errs, fmt.Errorf("closing file-getter: %w", err))
		}
		if errs != nil {
			return errs.ErrorOrNil()
		}
		return nil
	})
	return maybeCompressReadCloser(rc)
}

func (r *layerStore) DiffSize(from, to string) (size int64, err error) {
	var fromLayer, toLayer *Layer
	from, to, fromLayer, toLayer, err = r.findParentAndLayer(from, to)
	if err != nil {
		return -1, ErrLayerUnknown
	}
	return r.driver.DiffSize(to, r.layerMappings(toLayer), from, r.layerMappings(fromLayer), toLayer.MountLabel)
}

func (r *layerStore) ApplyDiff(to string, diff io.Reader) (size int64, err error) {
	return r.applyDiffWithOptions(to, nil, diff)
}

func (r *layerStore) applyDiffWithOptions(to string, layerOptions *LayerOptions, diff io.Reader) (size int64, err error) {
	if !r.IsReadWrite() {
		return -1, fmt.Errorf("not allowed to modify layer contents at %q: %w", r.layerspath(), ErrStoreIsReadOnly)
	}

	layer, ok := r.lookup(to)
	if !ok {
		return -1, ErrLayerUnknown
	}

	header := make([]byte, 10240)
	n, err := diff.Read(header)
	if err != nil && err != io.EOF {
		return -1, err
	}
	compression := archive.DetectCompression(header[:n])
	defragmented := io.MultiReader(bytes.NewBuffer(header[:n]), diff)

	// Decide if we need to compute digests
	var compressedDigest, uncompressedDigest digest.Digest       // = ""
	var compressedDigester, uncompressedDigester digest.Digester // = nil
	if layerOptions != nil && layerOptions.OriginalDigest != "" &&
		layerOptions.OriginalDigest.Algorithm() == digest.Canonical {
		compressedDigest = layerOptions.OriginalDigest
	} else {
		compressedDigester = digest.Canonical.Digester()
	}
	if layerOptions != nil && layerOptions.UncompressedDigest != "" &&
		layerOptions.UncompressedDigest.Algorithm() == digest.Canonical {
		uncompressedDigest = layerOptions.UncompressedDigest
	} else {
		uncompressedDigester = digest.Canonical.Digester()
	}

	var compressedWriter io.Writer
	if compressedDigester != nil {
		compressedWriter = compressedDigester.Hash()
	} else {
		compressedWriter = io.Discard
	}
	compressedCounter := ioutils.NewWriteCounter(compressedWriter)
	defragmented = io.TeeReader(defragmented, compressedCounter)

	tsdata := bytes.Buffer{}
	compressor, err := pgzip.NewWriterLevel(&tsdata, pgzip.BestSpeed)
	if err != nil {
		compressor = pgzip.NewWriter(&tsdata)
	}
	if err := compressor.SetConcurrency(1024*1024, 1); err != nil { // 1024*1024 is the hard-coded default; we're not changing that
		logrus.Infof("setting compression concurrency threads to 1: %v; ignoring", err)
	}
	metadata := storage.NewJSONPacker(compressor)
	uncompressed, err := archive.DecompressStream(defragmented)
	if err != nil {
		return -1, err
	}
	defer uncompressed.Close()
	uidLog := make(map[uint32]struct{})
	gidLog := make(map[uint32]struct{})
	idLogger, err := tarlog.NewLogger(func(h *tar.Header) {
		if !strings.HasPrefix(path.Base(h.Name), archive.WhiteoutPrefix) {
			uidLog[uint32(h.Uid)] = struct{}{}
			gidLog[uint32(h.Gid)] = struct{}{}
		}
	})
	if err != nil {
		return -1, err
	}
	defer idLogger.Close()
	uncompressedCounter := ioutils.NewWriteCounter(idLogger)
	uncompressedWriter := (io.Writer)(uncompressedCounter)
	if uncompressedDigester != nil {
		uncompressedWriter = io.MultiWriter(uncompressedWriter, uncompressedDigester.Hash())
	}
	payload, err := asm.NewInputTarStream(io.TeeReader(uncompressed, uncompressedWriter), metadata, storage.NewDiscardFilePutter())
	if err != nil {
		return -1, err
	}
	options := drivers.ApplyDiffOpts{
		Diff:       payload,
		Mappings:   r.layerMappings(layer),
		MountLabel: layer.MountLabel,
	}
	size, err = r.driver.ApplyDiff(layer.ID, layer.Parent, options)
	if err != nil {
		return -1, err
	}
	compressor.Close()
	if err == nil {
		if err := os.MkdirAll(filepath.Dir(r.tspath(layer.ID)), 0700); err != nil {
			return -1, err
		}
		if err := ioutils.AtomicWriteFile(r.tspath(layer.ID), tsdata.Bytes(), 0600); err != nil {
			return -1, err
		}
	}
	if compressedDigester != nil {
		compressedDigest = compressedDigester.Digest()
	}
	if uncompressedDigester != nil {
		uncompressedDigest = uncompressedDigester.Digest()
	}

	updateDigestMap := func(m *map[digest.Digest][]string, oldvalue, newvalue digest.Digest, id string) {
		var newList []string
		if oldvalue != "" {
			for _, value := range (*m)[oldvalue] {
				if value != id {
					newList = append(newList, value)
				}
			}
			if len(newList) > 0 {
				(*m)[oldvalue] = newList
			} else {
				delete(*m, oldvalue)
			}
		}
		if newvalue != "" {
			(*m)[newvalue] = append((*m)[newvalue], id)
		}
	}
	updateDigestMap(&r.bycompressedsum, layer.CompressedDigest, compressedDigest, layer.ID)
	layer.CompressedDigest = compressedDigest
	layer.CompressedSize = compressedCounter.Count
	updateDigestMap(&r.byuncompressedsum, layer.UncompressedDigest, uncompressedDigest, layer.ID)
	layer.UncompressedDigest = uncompressedDigest
	layer.UncompressedSize = uncompressedCounter.Count
	layer.CompressionType = compression
	layer.UIDs = make([]uint32, 0, len(uidLog))
	for uid := range uidLog {
		layer.UIDs = append(layer.UIDs, uid)
	}
	sort.Slice(layer.UIDs, func(i, j int) bool {
		return layer.UIDs[i] < layer.UIDs[j]
	})
	layer.GIDs = make([]uint32, 0, len(gidLog))
	for gid := range gidLog {
		layer.GIDs = append(layer.GIDs, gid)
	}
	sort.Slice(layer.GIDs, func(i, j int) bool {
		return layer.GIDs[i] < layer.GIDs[j]
	})

	err = r.Save()

	return size, err
}

func (r *layerStore) DifferTarget(id string) (string, error) {
	ddriver, ok := r.driver.(drivers.DriverWithDiffer)
	if !ok {
		return "", ErrNotSupported
	}
	layer, ok := r.lookup(id)
	if !ok {
		return "", ErrLayerUnknown
	}
	return ddriver.DifferTarget(layer.ID)
}

func (r *layerStore) ApplyDiffFromStagingDirectory(id, stagingDirectory string, diffOutput *drivers.DriverWithDifferOutput, options *drivers.ApplyDiffOpts) error {
	ddriver, ok := r.driver.(drivers.DriverWithDiffer)
	if !ok {
		return ErrNotSupported
	}
	layer, ok := r.lookup(id)
	if !ok {
		return ErrLayerUnknown
	}
	if options == nil {
		options = &drivers.ApplyDiffOpts{
			Mappings:   r.layerMappings(layer),
			MountLabel: layer.MountLabel,
		}
	}
	err := ddriver.ApplyDiffFromStagingDirectory(layer.ID, layer.Parent, stagingDirectory, diffOutput, options)
	if err != nil {
		return err
	}
	layer.UIDs = diffOutput.UIDs
	layer.GIDs = diffOutput.GIDs
	layer.UncompressedDigest = diffOutput.UncompressedDigest
	layer.UncompressedSize = diffOutput.Size
	layer.Metadata = diffOutput.Metadata
	if err = r.Save(); err != nil {
		return err
	}
	for k, v := range diffOutput.BigData {
		if err := r.SetBigData(id, k, bytes.NewReader(v)); err != nil {
			r.Delete(id)
			return err
		}
	}
	return err
}

func (r *layerStore) ApplyDiffWithDiffer(to string, options *drivers.ApplyDiffOpts, differ drivers.Differ) (*drivers.DriverWithDifferOutput, error) {
	ddriver, ok := r.driver.(drivers.DriverWithDiffer)
	if !ok {
		return nil, ErrNotSupported
	}

	if to == "" {
		output, err := ddriver.ApplyDiffWithDiffer("", "", options, differ)
		return &output, err
	}

	layer, ok := r.lookup(to)
	if !ok {
		return nil, ErrLayerUnknown
	}
	if options == nil {
		options = &drivers.ApplyDiffOpts{
			Mappings:   r.layerMappings(layer),
			MountLabel: layer.MountLabel,
		}
	}
	output, err := ddriver.ApplyDiffWithDiffer(layer.ID, layer.Parent, options, differ)
	if err != nil {
		return nil, err
	}
	layer.UIDs = output.UIDs
	layer.GIDs = output.GIDs
	err = r.Save()
	return &output, err
}

func (r *layerStore) CleanupStagingDirectory(stagingDirectory string) error {
	ddriver, ok := r.driver.(drivers.DriverWithDiffer)
	if !ok {
		return ErrNotSupported
	}
	return ddriver.CleanupStagingDirectory(stagingDirectory)
}

func (r *layerStore) layersByDigestMap(m map[digest.Digest][]string, d digest.Digest) ([]Layer, error) {
	var layers []Layer
	for _, layerID := range m[d] {
		layer, ok := r.lookup(layerID)
		if !ok {
			return nil, ErrLayerUnknown
		}
		layers = append(layers, *copyLayer(layer))
	}
	return layers, nil
}

func (r *layerStore) LayersByCompressedDigest(d digest.Digest) ([]Layer, error) {
	return r.layersByDigestMap(r.bycompressedsum, d)
}

func (r *layerStore) LayersByUncompressedDigest(d digest.Digest) ([]Layer, error) {
	return r.layersByDigestMap(r.byuncompressedsum, d)
}

func (r *layerStore) Lock() {
	r.lockfile.Lock()
}

func (r *layerStore) RecursiveLock() {
	r.lockfile.RecursiveLock()
}

func (r *layerStore) RLock() {
	r.lockfile.RLock()
}

func (r *layerStore) Unlock() {
	r.lockfile.Unlock()
}

func (r *layerStore) Touch() error {
	return r.lockfile.Touch()
}

func (r *layerStore) Modified() (bool, error) {
	var mmodified, tmodified bool
	lmodified, err := r.lockfile.Modified()
	if err != nil {
		return lmodified, err
	}
	if r.IsReadWrite() {
		r.mountsLockfile.RLock()
		defer r.mountsLockfile.Unlock()
		mmodified, err = r.mountsLockfile.Modified()
		if err != nil {
			return lmodified, err
		}
	}

	if lmodified || mmodified {
		return true, nil
	}

	// If the layers.json file has been modified manually, then we have to
	// reload the storage in any case.
	info, err := os.Stat(r.layerspath())
	if err != nil && !os.IsNotExist(err) {
		return false, fmt.Errorf("stat layers file: %w", err)
	}
	if info != nil {
		tmodified = info.ModTime() != r.layerspathModified
	}

	return tmodified, nil
}

func (r *layerStore) IsReadWrite() bool {
	return r.lockfile.IsReadWrite()
}

func (r *layerStore) TouchedSince(when time.Time) bool {
	return r.lockfile.TouchedSince(when)
}

func (r *layerStore) Locked() bool {
	return r.lockfile.Locked()
}

func (r *layerStore) ReloadIfChanged() error {
	r.loadMut.Lock()
	defer r.loadMut.Unlock()

	modified, err := r.Modified()
	if err == nil && modified {
		return r.Load()
	}
	return err
}

func closeAll(closes ...func() error) (rErr error) {
	for _, f := range closes {
		if err := f(); err != nil {
			if rErr == nil {
				rErr = fmt.Errorf("close error: %w", err)
				continue
			}
			rErr = fmt.Errorf("%v: %w", err, rErr)
		}
	}
	return
}