nfs/internal/nfs4server/layout.go

// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
// SPDX-License-Identifier: MIT

package nfs4server

import (
	"net"
	"sync"

	"sourcedock.dev/petrbalvin/nfs/internal/nfs4"
	"sourcedock.dev/petrbalvin/nfs/internal/nfsfs"
	"sourcedock.dev/petrbalvin/nfs/internal/xdr"
)

// layoutDeviceID names the one storage device behind every layout: the
// metadata server of this build is also its data server, so one identity
// serves both roles.
var layoutDeviceID = newDeviceID()

// newDeviceID builds the fixed device identity of this server.
func newDeviceID() [16]byte {
	var id [16]byte
	copy(id[:], "pnfs42mdsds00001")
	return id
}

// A layout is one granted pNFS layout: the session and client it belongs
// to, the file it covers, the byte range and the stateid the client
// addresses it by.
type layout struct {
	stateid  nfs4.Stateid
	sessID   nfs4.SessionID
	clientid uint64
	fh       nfsfs.Handle
	offset   uint64
	length   uint64
}

// layoutServer tracks the layouts the metadata server has granted. The
// store lives in memory: a server restart drops every layout and the
// clients re-request them through the grace window of the restart.
type layoutServer struct {
	mu      sync.Mutex
	next    uint64
	layouts map[string]*layout // by layout stateid other
}

func newLayoutServer() *layoutServer {
	return &layoutServer{next: randCounter(), layouts: make(map[string]*layout)}
}

// grant issues a layout over the requested range of the file handle,
// replacing the layout the same client already holds on the same file,
// so a re-request or a retry never piles entries up. The stateid other
// carries the LAYOUT mark and a counter; the sequence field starts at
// one, as RFC 8881 section 12.5.2 has it for layout stateids.
func (l *layoutServer) grant(sessID nfs4.SessionID, clientid uint64, fh nfsfs.Handle, offset, length uint64) *layout {
	l.mu.Lock()
	defer l.mu.Unlock()
	key := fileKey(fh)
	for _, lay := range l.layouts {
		if lay.clientid == clientid && fileKey(lay.fh) == key {
			lay.offset, lay.length, lay.sessID = offset, length, sessID
			return lay
		}
	}
	l.next++
	var st nfs4.Stateid
	setStateidSeq(&st, 1)
	copy(st[4:], "LAYOUT")
	for i := range 6 {
		st[15-i] = byte(l.next >> (8 * i))
	}
	lay := &layout{stateid: st, sessID: sessID, clientid: clientid, fh: fh, offset: offset, length: length}
	l.layouts[string(st[4:])] = lay
	return lay
}

// lookup finds a live layout by the stateid the client presents.
func (l *layoutServer) lookup(st nfs4.Stateid, fh nfsfs.Handle) (*layout, uint32) {
	l.mu.Lock()
	defer l.mu.Unlock()
	lay, ok := l.layouts[string(st[4:])]
	if !ok || fileKey(lay.fh) != fileKey(fh) {
		return nil, nfs4.ErrBadStateid
	}
	return lay, nfs4.ErrOK
}

// drop removes the layout identified by the stateid and answers the status
// of the removal.
func (l *layoutServer) drop(st nfs4.Stateid) uint32 {
	l.mu.Lock()
	defer l.mu.Unlock()
	if _, ok := l.layouts[string(st[4:])]; !ok {
		return nfs4.ErrBadStateid
	}
	delete(l.layouts, string(st[4:]))
	return nfs4.ErrOK
}

// dropSession drops every layout of one session, which DESTROY_SESSION
// requires: layouts are session bound state.
func (l *layoutServer) dropSession(sessID nfs4.SessionID) {
	l.mu.Lock()
	defer l.mu.Unlock()
	for other, lay := range l.layouts {
		if lay.sessID == sessID {
			delete(l.layouts, other)
		}
	}
}

// dropClient drops every layout of one client, which DESTROY_CLIENTID
// requires.
func (l *layoutServer) dropClient(clientid uint64) {
	l.mu.Lock()
	defer l.mu.Unlock()
	for other, lay := range l.layouts {
		if lay.clientid == clientid {
			delete(l.layouts, other)
		}
	}
}

// count reports how many layouts the server has granted.
func (l *layoutServer) count() int {
	l.mu.Lock()
	defer l.mu.Unlock()
	return len(l.layouts)
}

// layoutGetOp serves LAYOUTGET: it validates the open stateid the client
// presents, grants a flexfiles layout over the requested range and names
// this server as the data server the client reads and writes through.
func (h *Handler) layoutGetOp(d *xdr.Decoder, reg *fhreg, sessID nfs4.SessionID, clientid uint64) ([]byte, uint32, error) {
	if _, err := d.Bool(); err != nil { // signal_avail
		return nil, 0, err
	}
	layoutType, err := d.Uint32()
	if err != nil {
		return nil, 0, err
	}
	iomode, err := d.Uint32()
	if err != nil {
		return nil, 0, err
	}
	offset, err := d.Uint64()
	if err != nil {
		return nil, 0, err
	}
	length, err := d.Uint64()
	if err != nil {
		return nil, 0, err
	}
	if _, err = d.Uint64(); err != nil { // minlength
		return nil, 0, err
	}
	var openSt nfs4.Stateid
	raw, rerr := d.Raw(16)
	if rerr != nil {
		return nil, 0, rerr
	}
	copy(openSt[:], raw)
	maxcount, err := d.Uint32()
	if err != nil {
		return nil, 0, err
	}
	if layoutType != nfs4.LayoutTypeFlexfiles && layoutType != nfs4.LayoutTypeFlexFilesV2 &&
		layoutType != nfs4.LayoutTypeFiles && layoutType != nfs4.LayoutTypeBlock &&
		layoutType != nfs4.LayoutTypeObjects && layoutType != nfs4.LayoutTypeScsi {
		return nil, nfs4.ErrUnknownLayoutType, nil
	}
	if iomode == 0 || iomode > nfs4.IoModeRW {
		return nil, nfs4.ErrBadIOMode, nil
	}
	if !reg.haveCur {
		return nil, nfs4.ErrNoFileHandle, nil
	}
	// A layout hangs from a real open of the caller: the anonymous
	// stateid forms never qualify.
	if _, status := h.openStates().lookupOpen(openSt, reg.cur, clientid); status != nfs4.ErrOK {
		return nil, status, nil
	}
	// The body is built through one closure so the maxcount check can
	// run against a scratch stateid before anything is granted: a
	// request whose answer cannot fit leaves no layout behind, RFC 8881
	// section 18.43.
	build := func(st nfs4.Stateid) []byte {
		switch layoutType {
		case nfs4.LayoutTypeFiles:
			return nfs4.AppendFileLayoutBody(nil, layoutDeviceID,
				0 /* util: no striping */, 0, 0, [][]byte{reg.cur})
		case nfs4.LayoutTypeBlock:
			return nfs4.AppendBlockDeviceAddr(nil, nfs4.BlockVolume{
				DeviceID:   layoutDeviceID,
				BaseOffset: offset,
				BlockCount: length,
			})
		case nfs4.LayoutTypeObjects:
			return nfs4.AppendObjectLayoutBody(nil, layoutDeviceID, nfs4.ObjectLayout{
				NumComponents: 1, StripeUnit: 4096, GroupWidth: 1, GroupDepth: 1,
			})
		case nfs4.LayoutTypeScsi:
			return nfs4.AppendScsiLayoutBody(nil, layoutDeviceID, offset, length, offset)
		case nfs4.LayoutTypeFlexFilesV2:
			// The version two body of draft-haynes-nfsv4-flex-filesv2-00:
			// one stateid per version and the AUTH_NONE credential, which
			// the draft prescribes for tight coupling over synthetic
			// identities.
			return nfs4.AppendFlexFileLayoutBodyV2(nil, 0, 0, []nfs4.FlexMirrorV2{{
				DataServers: []nfs4.FlexDataServerV2{{
					DeviceID:   layoutDeviceID,
					Stateids:   []nfs4.Stateid{st},
					FHs:        [][]byte{reg.cur},
					AuthFlavor: 0, // AUTH_NONE
				}},
			}})
		default:
			return nfs4.AppendFlexFileLayoutBody(nil, 0, 0, []nfs4.FlexMirror{{
				DataServers: []nfs4.FlexDataServer{{
					DeviceID: layoutDeviceID,
					Stateid:  st,
					FHs:      [][]byte{reg.cur},
				}},
			}})
		}
	}
	appendRes := func(st nfs4.Stateid) []byte {
		return nfs4.AppendLayoutGetRes(nil, st, false, []nfs4.Layout4{{
			Offset: offset,
			Length: length,
			IoMode: iomode,
			Type:   layoutType,
			Body:   build(st),
		}})
	}
	if maxcount != 0 && uint32(len(appendRes(nfs4.Stateid{}))) > maxcount {
		return nil, nfs4.ErrTooSmall, nil
	}
	lay := h.layouts().grant(sessID, clientid, reg.cur, offset, length)
	return appendRes(lay.stateid), nfs4.ErrOK, nil
}

// layoutCommitOp serves LAYOUTCOMMIT: the layout must still be live and
// the client's last write offset may grow the file. The answer names the
// size the file is left at.
func (h *Handler) layoutCommitOp(d *xdr.Decoder, reg *fhreg) ([]byte, uint32, error) {
	if !reg.haveCur {
		return nil, nfs4.ErrNoFileHandle, nil
	}
	offset, err := d.Uint64()
	if err != nil {
		return nil, 0, err
	}
	length, err := d.Uint64()
	if err != nil {
		return nil, 0, err
	}
	if _, err = d.Bool(); err != nil { // reclaim
		return nil, 0, err
	}
	var st nfs4.Stateid
	raw, rerr := d.Raw(16)
	if rerr != nil {
		return nil, 0, rerr
	}
	copy(st[:], raw)
	lastWriteSet, err := d.Bool()
	if err != nil {
		return nil, 0, err
	}
	var lastWrite uint64
	if lastWriteSet {
		if lastWrite, err = d.Uint64(); err != nil {
			return nil, 0, err
		}
	}
	timeSet, err := d.Bool()
	if err != nil {
		return nil, 0, err
	}
	if timeSet {
		if _, err = d.Int64(); err != nil {
			return nil, 0, err
		}
		if _, err = d.Uint32(); err != nil {
			return nil, 0, err
		}
	}
	if _, err = d.Uint32(); err != nil { // layout update type
		return nil, 0, err
	}
	if _, err = d.VarOpaque(); err != nil { // layout update body
		return nil, 0, err
	}
	_ = offset
	_ = length
	if _, status := h.layouts().lookup(st, reg.cur); status != nfs4.ErrOK {
		return nil, status, nil
	}
	info, ferr := h.FS.Getattr(reg.cur)
	if ferr != nil {
		return nil, mapErr(ferr), nil
	}
	newSize := uint64(info.Size)
	if lastWriteSet && lastWrite+1 > newSize {
		w := h.writer()
		if w == nil {
			return nil, nfs4.ErrROFS, nil
		}
		grown := int64(lastWrite + 1)
		if err := w.Setattr(reg.cur, nfsfs.SetAttrs{Size: &grown}); err != nil {
			return nil, mapErr(err), nil
		}
		newSize = lastWrite + 1
	}
	return nfs4.AppendLayoutCommitRes(nil, newSize), nfs4.ErrOK, nil
}

// layoutReturnOp serves LAYOUTRETURN: the client gives the layout back.
// A file return drops the named layout, a whole client or file system
// return drops every layout of the session.
func (h *Handler) layoutReturnOp(d *xdr.Decoder, reg *fhreg, sessID nfs4.SessionID, clientid uint64) ([]byte, uint32, error) {
	reclaim, err := d.Bool()
	if err != nil {
		return nil, 0, err
	}
	layoutType, err := d.Uint32()
	if err != nil {
		return nil, 0, err
	}
	if _, err = d.Uint32(); err != nil { // iomode
		return nil, 0, err
	}
	kind, err := d.Uint32()
	if err != nil {
		return nil, 0, err
	}
	_ = reclaim
	if layoutType != nfs4.LayoutTypeFlexfiles && layoutType != nfs4.LayoutTypeFlexFilesV2 {
		return nil, nfs4.ErrUnknownLayoutType, nil
	}
	switch kind {
	case nfs4.ReturnFile:
		if _, err = d.Uint64(); err != nil { // offset
			return nil, 0, err
		}
		if _, err = d.Uint64(); err != nil { // length
			return nil, 0, err
		}
		var st nfs4.Stateid
		raw, rerr := d.Raw(16)
		if rerr != nil {
			return nil, 0, rerr
		}
		copy(st[:], raw)
		if _, err = d.Uint32(); err != nil { // ffsid_info_type
			return nil, 0, err
		}
		if !reg.haveCur {
			return nil, nfs4.ErrNoFileHandle, nil
		}
		if _, status := h.layouts().lookup(st, reg.cur); status != nfs4.ErrOK {
			return nil, status, nil
		}
		returned := st
		returned[0]++
		if status := h.layouts().drop(st); status != nfs4.ErrOK {
			return nil, status, nil
		}
		return nfs4.AppendLayoutReturnRes(nil, returned), nfs4.ErrOK, nil
	case nfs4.ReturnFsid, nfs4.ReturnAll:
		h.layouts().dropClient(clientid)
		h.layouts().dropSession(sessID)
		return nfs4.AppendLayoutReturnRes(nil, nfs4.Stateid{}), nfs4.ErrOK, nil
	default:
		return nil, nfs4.ErrBadLayout, nil
	}
}

// getDeviceInfoOp serves GETDEVICEINFO: the data server addresses the
// client needs to reach the storage behind the layout. The one device of
// this build is the metadata server itself.
func (h *Handler) getDeviceInfoOp(d *xdr.Decoder, ctx *connCB) ([]byte, uint32, error) {
	var deviceID [16]byte
	raw, err := d.Raw(16)
	if err != nil {
		return nil, 0, err
	}
	copy(deviceID[:], raw)
	layoutType, err := d.Uint32()
	if err != nil {
		return nil, 0, err
	}
	maxcount, err := d.Uint32()
	if err != nil {
		return nil, 0, err
	}
	if _, err = nfs4.ReadBitmap(d); err != nil { // notification types
		return nil, 0, err
	}
	if layoutType != nfs4.LayoutTypeFlexfiles && layoutType != nfs4.LayoutTypeFlexFilesV2 {
		return nil, nfs4.ErrUnknownLayoutType, nil
	}
	if deviceID != layoutDeviceID {
		return nil, nfs4.ErrNoEnt, nil
	}
	// The address body follows the layout type: flexfiles carries the
	// server list with versions, the files layout the stripe indices over
	// the multipath list, the others the emulated volume or component.
	var addr []byte
	switch layoutType {
	case nfs4.LayoutTypeFiles:
		addr = nfs4.AppendFileDeviceAddr(nil, []uint32{0},
			[]nfs4.NetAddr{{Netid: "tcp", Uaddr: h.deviceAddr(ctx)}})
	default:
		addr = nfs4.AppendFlexDeviceAddr(nil, nfs4.FlexDeviceAddr{
			NetAddrs: []nfs4.NetAddr{{Netid: "tcp", Uaddr: h.deviceAddr(ctx)}},
			Versions: []nfs4.FlexVersion{{
				Version:      4,
				MinorVersion: nfs4.MinorVersion,
				RSize:        uint32(nfs4.DefaultLimits.MaxRead),
				WSize:        uint32(nfs4.DefaultLimits.MaxWrite),
			}},
		})
	}
	res := nfs4.AppendGetDeviceInfoRes(nil, addr)
	if maxcount != 0 && uint32(len(res)) > maxcount {
		return nil, nfs4.ErrTooSmall, nil
	}
	return res, nfs4.ErrOK, nil
}

// deviceAddr resolves the universal address the data server answers on:
// the configured value wins, then the local address of the connection the
// request rode in on, then the loopback default.
func (h *Handler) deviceAddr(ctx *connCB) string {
	if h.DeviceAddr != "" {
		return h.DeviceAddr
	}
	if ctx != nil && ctx.conn != nil {
		if u := uaddrOf(ctx.conn.LocalAddr().String()); u != "" {
			return u
		}
	}
	return "127.0.0.1.8.1"
}

// uaddrOf turns a host:port address into the universal address form of
// RFC 8435: decimal octets and port for IPv4, hex nibbles for IPv6.
func uaddrOf(hostPort string) string {
	host, portText, err := net.SplitHostPort(hostPort)
	if err != nil {
		return ""
	}
	ip := net.ParseIP(host)
	if ip == nil {
		return ""
	}
	port := 0
	if portText == "" {
		return ""
	}
	for _, r := range portText {
		if r < '0' || r > '9' {
			return ""
		}
		port = port*10 + int(r-'0')
		if port > 0xffff {
			return ""
		}
	}
	var v4 [4]byte
	if n := copy(v4[:], ip.To4()); n == 4 {
		return itoa(int(v4[0])) + "." + itoa(int(v4[1])) + "." + itoa(int(v4[2])) + "." +
			itoa(int(v4[3])) + "." + itoa(port>>8) + "." + itoa(port&0xff)
	}
	v6 := ip.To16()
	if v6 == nil {
		return ""
	}
	const hexDigits = "0123456789abcdef"
	out := make([]byte, 0, 16*3+8)
	for i, b := range v6 {
		if i > 0 {
			out = append(out, '.')
		}
		out = append(out, hexDigits[b>>4], hexDigits[b&0xf])
	}
	out = append(out, '.')
	out = append(out, itoa(port>>8)...)
	out = append(out, '.')
	out = append(out, itoa(port&0xff)...)
	return string(out)
}

// itoa renders a small non negative number in decimal.
func itoa(n int) string {
	if n == 0 {
		return "0"
	}
	var digits [8]byte
	i := len(digits)
	for n > 0 {
		i--
		digits[i] = byte('0' + n%10)
		n /= 10
	}
	return string(digits[i:])
}

// getDeviceListOp serves GETDEVICELIST: the one device of this build is
// the answer for every layout type it carries, RFC 5661 section 18.41.
func (h *Handler) getDeviceListOp(d *xdr.Decoder, ctx *connCB) ([]byte, uint32, error) {
	layoutType, err := d.Uint32()
	if err != nil {
		return nil, 0, err
	}
	if _, err = d.Uint32(); err != nil { // maxdevices
		return nil, 0, err
	}
	if _, err = d.Uint64(); err != nil { // cookie
		return nil, 0, err
	}
	if _, err = d.Raw(8); err != nil { // cookie verifier
		return nil, 0, err
	}
	switch layoutType {
	case nfs4.LayoutTypeFlexfiles, nfs4.LayoutTypeFlexFilesV2, nfs4.LayoutTypeFiles,
		nfs4.LayoutTypeBlock, nfs4.LayoutTypeObjects, nfs4.LayoutTypeScsi:
		return nfs4.AppendGetDeviceListRes(nil, 0, h.writeVerifier(),
			[][16]byte{layoutDeviceID}, true), nfs4.ErrOK, nil
	default:
		return nil, nfs4.ErrUnknownLayoutType, nil
	}
}