Keep libraries under groups of their own, and give modules users in them (hq ADR 0328, issue 567)

ADR 0328 supersedes ADR 0325: shared data is a library with one home, kept usable by
its group rather than by ACL entries. The engine now makes each library's group and
each module's user in its libraries' groups; keeps a library's root and requested
sub-paths in its group with group write and setgid on its home, making missing
sub-paths as the root is; waits for the operator's word before touching a root that
holds data, and on the take act walks it by descriptor (no link, same filesystem,
bounded and resumed, files with other names left); holds the containers of a library
whose disk or mount is not there; and puts the tool runner's account in the group.
The ACL keeping, mask repair and watch are gone; an access is again as before ADR 0325.
This commit is contained in:
2026-10-12 22:11:09 +02:00
parent 303d4037e8
commit 2a2be687f3
35 changed files with 2583 additions and 5004 deletions
+170 -292
View File
@@ -2,7 +2,7 @@
package apply
// The descriptor half of walk.go and acl.go (novox/hq ADR 0325 §5, §8, §9, §10).
// The descriptor half of walk.go and library.go (novox/hq ADR 0328 §6, §7, §9).
import (
"errors"
@@ -10,6 +10,7 @@ import (
"io"
"os"
"sort"
"strings"
"time"
"golang.org/x/sys/unix"
@@ -29,29 +30,42 @@ var machineStats = walkStats{
fd: unix.Fstat,
}
// visitFunc is what a walk does to one folder or file, open on fd, whose stat is st: whether it changed it, and
// whether it found the entry narrowed and could not put it back. An error stops the walk.
type visitFunc func(fd int, st unix.Stat_t, folder bool) (changed, narrowed, unrepaired bool, err error)
// What a visit did to one entry.
const (
visitKept = iota // it had what it needs
visitChanged // it was changed
visitShared // a file with another name elsewhere: left as it is
)
// aclWalk is one walk under a kept folder, from the descriptor of the folder itself.
type aclWalk struct {
// visitFunc is what a walk does to one folder or file, open on fd, whose stat is st: what it did, or the error
// counted on that entry alone (errStopWalk stops the walk).
type visitFunc func(fd int, st unix.Stat_t, folder bool) (int, error)
// errStopWalk wraps an error that stops the whole walk, never one entry's alone.
type errStopWalk struct{ err error }
func (e errStopWalk) Error() string { return e.err.Error() }
func (e errStopWalk) Unwrap() error { return e.err }
// treeWalk is one walk under a library's root, from the descriptor of the root itself.
type treeWalk struct {
dev uint64
deadline time.Time
limit int
stats walkStats
visit visitFunc
counts WalkCounts
// at is the names of the levels below the folder down to the entry being walked; Cursor, once the walk met
// its bound, where the next goes on.
// at is the names of the levels below the root down to the entry being walked; Cursor, once the walk met its
// bound, where the next goes on.
at []string
Cursor []string
// Failed is the first failure on one entry, in words with no path.
Failed string
}
// openKept opens a kept folder as a mesh directory is opened (no link another account could make, the folder
// itself no link) and starts a walk from it on the folder's own device.
func openKept(path string, limits WalkLimits, stats walkStats, visit visitFunc) (*aclWalk, int, unix.Stat_t, error) {
// openRoot opens a library's root as a mesh directory is opened (no link another account could make, the root
// itself no link) and starts a walk from it on the root's own device.
func openRoot(path string, limits WalkLimits, stats walkStats, visit visitFunc) (*treeWalk, int, unix.Stat_t, error) {
var st unix.Stat_t
if err := keptPath(path); err != nil {
return nil, -1, st, err
@@ -59,36 +73,36 @@ func openKept(path string, limits WalkLimits, stats walkStats, visit visitFunc)
mw := &meshWalk{want: path}
fd, err := mw.open()
if errors.Is(err, errNotThereYet) {
return nil, -1, st, errors.New("the folder is not there; nothing was changed")
return nil, -1, st, errors.New("the root is not there; nothing was changed")
}
if err != nil {
return nil, -1, st, errors.New("the folder is not walked: " + pathless(err))
return nil, -1, st, errors.New("the root is not walked: " + pathless(err))
}
if mw.At != path {
unix.Close(fd)
return nil, -1, st, errors.New("the folder is reached through a link, and the engine walks no folder " +
"through one; nothing was changed")
return nil, -1, st, errors.New("the root is reached through a link, and the engine walks no root through " +
"one; nothing was changed")
}
if err := stats.fd(fd, &st); err != nil {
unix.Close(fd)
return nil, -1, st, fmt.Errorf("the folder cannot be read: %v; nothing was changed", err)
return nil, -1, st, fmt.Errorf("the root cannot be read: %v; nothing was changed", err)
}
return &aclWalk{dev: uint64(st.Dev), deadline: time.Now().Add(limits.Time), limit: limits.Entries,
return &treeWalk{dev: uint64(st.Dev), deadline: time.Now().Add(limits.Time), limit: limits.Entries,
stats: stats, visit: visit}, fd, st, nil
}
// walkKept walks path whole, or from resume, with visit: the walk, with what it did and the cursor where it
// stopped when it met its bound (with ErrWalkBound), and the walk's own error otherwise.
func walkKept(path string, limits WalkLimits, stats walkStats, resume []string, visit visitFunc) (*aclWalk, error) {
w, fd, st, err := openKept(path, limits, stats, visit)
// walkRoot walks path whole, or from resume, with visit: the walk, with what it did and the cursor where it stopped
// when it met its bound (with ErrWalkBound), and the walk's own error otherwise.
func walkRoot(path string, limits WalkLimits, stats walkStats, resume []string, visit visitFunc) (*treeWalk, error) {
w, fd, st, err := openRoot(path, limits, stats, visit)
if err != nil {
return &aclWalk{}, err
return &treeWalk{}, err
}
return w, w.dir(fd, st, resume)
}
// read counts one entry, and says the bound when it is met, keeping where the walk stood as the cursor.
func (w *aclWalk) read() error {
func (w *treeWalk) read() error {
if w.counts.Read >= w.limit || time.Now().After(w.deadline) {
w.Cursor = append([]string(nil), w.at...)
return ErrWalkBound
@@ -97,27 +111,9 @@ func (w *aclWalk) read() error {
return nil
}
// count adds one visit's result.
func (w *aclWalk) count(folder, changed, narrowed, unrepaired bool) {
switch {
case changed && folder:
w.counts.Folders++
case changed:
w.counts.Files++
default:
w.counts.Kept++
}
if narrowed {
w.counts.Narrowed++
}
if unrepaired {
w.counts.Unrepaired++
}
}
// dir is one folder, open on fd, whose stat is st: the folder itself (unless the walk resumes below it, which an
// earlier walk visited), then every entry in it in the order of their names. It closes fd.
func (w *aclWalk) dir(fd int, st unix.Stat_t, resume []string) error {
func (w *treeWalk) dir(fd int, st unix.Stat_t, resume []string) error {
f := os.NewFile(uintptr(fd), "")
defer f.Close()
if len(resume) == 0 {
@@ -154,9 +150,9 @@ func (w *aclWalk) dir(fd int, st unix.Stat_t, resume []string) error {
return nil
}
// entry is one name in the folder open on dirfd: a link passed by, a folder on this device walked, a file
// visited through its own descriptor, anything else left.
func (w *aclWalk) entry(dirfd int, name string, resume []string) error {
// entry is one name in the folder open on dirfd: a link passed by, a folder on this device walked, a file visited
// through its own descriptor, anything else left.
func (w *treeWalk) entry(dirfd int, name string, resume []string) error {
var st unix.Stat_t
if err := w.stats.at(dirfd, name, &st); err != nil {
if errors.Is(err, unix.ENOENT) {
@@ -212,192 +208,65 @@ func (w *aclWalk) entry(dirfd int, name string, resume []string) error {
return w.visited(fd, opened, false)
}
// visited visits one entry and counts what it did. A filesystem that allows no ACLs stops the walk; any other
// failure on one entry (an immutable file, say) is counted unrepaired, its words kept once, and the walk goes on.
func (w *aclWalk) visited(fd int, st unix.Stat_t, folder bool) error {
changed, narrowed, unrepaired, err := w.visit(fd, st, folder)
if errors.Is(err, ErrNoACL) {
return err
// visited visits one entry and counts what it did. An error that stops the walk stops it; any other failure on
// one entry (an immutable file, say) is counted failed, its words kept once, and the walk goes on.
func (w *treeWalk) visited(fd int, st unix.Stat_t, folder bool) error {
did, err := w.visit(fd, st, folder)
var stop errStopWalk
if errors.As(err, &stop) {
return stop.err
}
if err != nil {
unrepaired = true
w.counts.Failed++
if w.Failed == "" {
w.Failed = err.Error()
}
return nil
}
w.count(folder, changed, narrowed, unrepaired)
return nil
}
// getACL reads one of an entry's ACLs through its descriptor: the access ACL, or a folder's default. An access
// ACL that is not there is the one its mode stands for; a default that is not there is nil.
func getACL(fd int, name string, mode uint32) (acl, error) {
buf := make([]byte, 4+8*32)
for {
n, err := unix.Fgetxattr(fd, name, buf)
switch {
case errors.Is(err, unix.ENODATA):
if name == aclAccessXattr {
return minimalACL(mode), nil
}
return nil, nil
case errors.Is(err, unix.ERANGE):
buf = make([]byte, 2*len(buf))
continue
case errors.Is(err, unix.ENOTSUP), errors.Is(err, unix.EOPNOTSUPP):
return nil, ErrNoACL
case err != nil:
return nil, fmt.Errorf("an entry's ACL could not be read: %v", err)
}
return decodeACL(buf[:n])
}
}
// setACL writes one of an entry's ACLs through its descriptor, refusing one the kernel would.
func setACL(fd int, name string, a acl) error {
if err := a.check(); err != nil {
return err
}
err := unix.Fsetxattr(fd, name, a.encode(), 0)
switch {
case errors.Is(err, unix.ENOTSUP), errors.Is(err, unix.EOPNOTSUPP):
return ErrNoACL
case err != nil:
return fmt.Errorf("an entry's ACL could not be written: %v", err)
case did == visitShared:
w.counts.Shared++
case did == visitChanged && folder:
w.counts.Folders++
case did == visitChanged:
w.counts.Files++
default:
w.counts.Kept++
}
return nil
}
// sameACL is whether a and b hold the same entries, whatever their order.
func sameACL(a, b acl) bool {
return string(a.encode()) == string(b.encode())
}
// needs is, for a folder or a file, each kept uid's bits — except its owner's, which its owner's bits give and the
// mesh never changes.
func needs(users map[uint32]bool, owner uint32, folder bool) map[uint32]uint16 {
out := map[uint32]uint16{}
for uid, rw := range users {
if uid != owner {
out[uid] = needOn(rw, folder)
// groupVisit gives each entry the library's group gid, group read and write, and on a folder group enter and the
// setgid bit, so what is made beneath keeps the group (ADR 0328 §6, §7). The owner, the owner's bits and others'
// bits are never changed: a file's own setuid or setgid bit, which the kernel drops when the group changes, is put
// back as it was. A file with more than one name is left as it is — its other name may be outside the library, and
// giving it the group would give the group that other file.
func groupVisit(gid int) visitFunc {
return func(fd int, st unix.Stat_t, folder bool) (int, error) {
if !folder && st.Nlink > 1 {
return visitShared, nil
}
}
return out
}
// keepVisit makes users' entries on each entry (ADR 0325 §5): the access ACL with each uid's entry and a mask
// that covers it, and on a folder the same as its default, so what is made under it inherits them. An entry that
// had them is not written. A user that owns the entry is judged by its owner's bits, never given an entry: an
// owner whose bits fall short is unrepaired, said, and never changed.
func keepVisit(users map[uint32]bool) visitFunc {
return func(fd int, st unix.Stat_t, folder bool) (bool, bool, bool, error) {
mode := uint32(st.Mode) & 0o7777
need := needs(users, st.Uid, folder)
access, err := getACL(fd, aclAccessXattr, mode)
if err != nil {
return false, false, false, err
}
narrowed, unrepaired := judgeACL(access, users, st, folder)
changed := false
if kept := keptACL(access, need); !sameACL(kept, access) {
if err := setACL(fd, aclAccessXattr, kept); err != nil {
return false, narrowed, true, err
}
changed = true
}
want := mode | 0o060
if folder {
def, err := getACL(fd, aclDefaultXattr, mode)
if err != nil {
return changed, narrowed, unrepaired, err
}
base := def
if base == nil {
// No default yet: one whose owner's, owning group's and others' entries are the folder's own, as
// setfacl -d derives it.
base = minimalACL(mode)
if g, ok := access.perm(aclGroupObj, 0); ok {
base[1].Perm = g
}
}
// The default is for what anyone makes under the folder, never its owner alone: every user's entry.
all := map[uint32]uint16{}
for uid, rw := range users {
all[uid] = needOn(rw, true)
}
if kept := keptACL(base, all); def == nil || !sameACL(kept, def) {
if err := setACL(fd, aclDefaultXattr, kept); err != nil {
return changed, narrowed, true, err
}
changed = true
want = mode | 0o070 | unix.S_ISGID
}
if int(st.Gid) == gid && want == mode {
return visitKept, nil
}
if int(st.Gid) != gid {
if err := unix.Fchown(fd, -1, gid); err != nil {
return 0, fmt.Errorf("an entry's group could not be set: %v", err)
}
}
return changed, narrowed, unrepaired, nil
if err := unix.Fchmod(fd, want); err != nil {
return 0, fmt.Errorf("an entry's mode could not be set: %v", err)
}
return visitChanged, nil
}
}
// judgeACL is whether an entry with access ACL a gives a user less than its right — by entry and mask together
// (ADR 0325 §6) — and whether that is one the mesh cannot put back: the owner's own bits.
func judgeACL(a acl, users map[uint32]bool, st unix.Stat_t, folder bool) (narrowed, unrepaired bool) {
mode := uint32(st.Mode) & 0o7777
for uid, rw := range users {
need := needOn(rw, folder)
if effective(a, uid, st.Uid, mode)&need == need {
continue
}
narrowed = true
if uid == st.Uid {
unrepaired = true
}
}
return narrowed, unrepaired
}
// dropVisit takes exactly the entries of each uid of drop out of each entry's access ACL and a folder's default
// (ADR 0325 §8): no other entry, and not the mask, so nobody else loses a right.
func dropVisit(drop map[uint32]bool) visitFunc {
return func(fd int, st unix.Stat_t, folder bool) (bool, bool, bool, error) {
mode := uint32(st.Mode) & 0o7777
changed := false
for _, name := range []string{aclAccessXattr, aclDefaultXattr} {
if name == aclDefaultXattr && !folder {
continue
}
a, err := getACL(fd, name, mode)
if errors.Is(err, ErrNoACL) {
continue
}
if err != nil {
// Counted unrepaired and said by the walk, which then walks it again: never read as "nothing to drop".
return changed, false, true, err
}
if a == nil {
continue
}
if left := withoutUsers(a, drop); len(left) != len(a) {
if err := setACL(fd, name, left); err != nil {
return changed, false, true, err
}
changed = true
}
}
return changed, false, false, nil
}
}
// dropThenKeep is drop then keep, on one walk: a user whose use changed, or one gone beside one come.
func dropThenKeep(drop map[uint32]bool, users map[uint32]bool) visitFunc {
d, k := dropVisit(drop), keepVisit(users)
return func(fd int, st unix.Stat_t, folder bool) (bool, bool, bool, error) {
c1, _, _, err := d(fd, st, folder)
if err != nil {
return c1, false, false, err
}
c2, n, u, err := k(fd, st, folder)
return c1 || c2, n, u, err
}
}
// folderEmpty is whether the folder open on fd holds nothing yet: the data the mesh may keep without asking.
// folderEmpty is whether the folder open on fd holds nothing yet: a root the mesh may keep without asking.
func folderEmpty(fd int) (bool, error) {
dup, err := unix.Dup(fd)
if err != nil {
@@ -415,20 +284,11 @@ func folderEmpty(fd int) (bool, error) {
return false, fmt.Errorf("the folder could not be read: %v", err)
}
// keepTop keeps users' entries on the kept folder itself (ADR 0325 §5): what every apply does, beside the walks.
func keepTop(path string, users map[uint32]bool) (changed bool, err error) {
w, fd, st, err := openKept(path, WalkLimits{Entries: 1, Time: time.Minute}, machineStats, keepVisit(users))
if err != nil {
return false, err
}
defer unix.Close(fd)
changed, _, _, err = w.visit(fd, st, true)
return changed, err
}
// FolderFacts is what the engine reads of a folder before keeping it: its owner, group and whether it holds data.
// FolderFacts is what the engine reads of a library's root before keeping it: its owner, group, mode and whether
// it holds data.
type FolderFacts struct {
UID, GID uint32
Mode uint32
Empty bool
}
@@ -455,87 +315,105 @@ func readFolder(path string) (FolderFacts, error) {
if err != nil {
return FolderFacts{}, err
}
return FolderFacts{UID: st.Uid, GID: st.Gid, Empty: empty}, nil
return FolderFacts{UID: st.Uid, GID: st.Gid, Mode: uint32(st.Mode) & 0o7777, Empty: empty}, nil
}
// CreatePath makes a missing accessed folder (ADR 0325 §9): each missing level made from the nearest existing
// ancestor down, each with that ancestor's owner, group and mode. No link is followed (the ancestor is reached as
// a mesh directory is, and each level is opened with O_NOFOLLOW), no level is made on another filesystem, no file
// is made, and nothing that was there is changed: a level that appears meanwhile stops it. The caller has judged
// that the ancestor is on the placed filesystem and not root's. How many levels it made.
func CreatePath(path string) (int, error) {
return createPath(path, machineStats)
}
func createPath(path string, stats walkStats) (int, error) {
m, err := NearestAncestor(path)
// keepFolders keeps a library's root and each of its requested sub-paths — the folders themselves, never what is
// in them — in the group gid with group read, write and enter and the setgid bit (ADR 0328 §6). A sub-path that is
// missing is made, each level with the root's owner, the group, and the root's owner's and others' bits beside
// group rwx and setgid (§9). No link is followed (the root is reached as a mesh directory is, and every level below
// it by O_NOFOLLOW), nothing is made or changed on another filesystem, and the owner and the owner's and others'
// bits of a folder that was there are never changed. How many folders it changed and made.
func keepFolders(root string, paths []string, gid int) (changed, made int, err error) {
_, fd, rst, err := openRoot(root, WalkLimits{Entries: 1, Time: time.Minute}, machineStats, nil)
if err != nil {
return 0, fmt.Errorf("the folder cannot be made: %v; nothing was changed", err)
return 0, 0, err
}
if !m.Info.IsDir() {
return 0, errors.New("its nearest existing parent is not a folder (a file or a link), and a folder is " +
"made only below a folder; nothing was changed")
}
mw := &meshWalk{want: m.Ancestor}
fd, err := mw.open()
if err != nil {
return 0, fmt.Errorf("the nearest existing parent is not walked: %v; nothing was changed", err)
}
if mw.At != m.Ancestor {
unix.Close(fd)
return 0, errors.New("the nearest existing parent is reached through a link, and nothing is made " +
"through one; nothing was changed")
}
var anc unix.Stat_t
if err := stats.fd(fd, &anc); err != nil {
unix.Close(fd)
return 0, fmt.Errorf("the nearest existing parent cannot be read: %v; nothing was changed", err)
}
if why := rootOwned(anc.Uid, anc.Gid); why != "" {
unix.Close(fd)
return 0, fmt.Errorf("its nearest existing parent is %s, and nothing is made under root's data "+
"(novox/hq ADR 0325 §12); nothing was changed", why)
}
mode := uint32(anc.Mode) & 0o7777
made := 0
for _, name := range m.Levels {
// Made for its maker alone (the engine, as root), so it is nobody else's until it is the ancestor's.
if err := unix.Mkdirat(fd, name, 0o700); err != nil {
unix.Close(fd)
if errors.Is(err, unix.EEXIST) {
return made, errors.New("a level of the folder appeared while it was made, and nothing that is " +
"there is changed; it stopped there")
}
return made, fmt.Errorf("a level could not be made: %v", err)
defer unix.Close(fd)
keep := func(fd int, st unix.Stat_t) error {
mode := uint32(st.Mode) & 0o7777
want := mode | 0o070 | unix.S_ISGID
if int(st.Gid) == gid && want == mode {
return nil
}
next, err := unix.Openat(fd, name, unix.O_RDONLY|unix.O_DIRECTORY|unix.O_NOFOLLOW|unix.O_CLOEXEC, 0)
unix.Close(fd)
if int(st.Gid) != gid {
if err := unix.Fchown(fd, -1, gid); err != nil {
return fmt.Errorf("a folder's group could not be set: %v", err)
}
}
if err := unix.Fchmod(fd, want); err != nil {
return fmt.Errorf("a folder's mode could not be set: %v", err)
}
changed++
return nil
}
if err := keep(fd, rst); err != nil {
return changed, made, err
}
madeMode := uint32(rst.Mode)&0o0707 | 0o070 | unix.S_ISGID
for _, p := range paths {
n, err := keepPath(fd, rst, p, gid, madeMode, keep)
made += n
if err != nil {
return made, fmt.Errorf("a level made could not be opened: %v", err)
return changed, made, err
}
fd = next
made++
}
return changed, made, nil
}
// keepPath walks one sub-path down from the root open on rootFd, making each missing level, and keeps the last.
func keepPath(rootFd int, rst unix.Stat_t, p string, gid int, madeMode uint32,
keep func(int, unix.Stat_t) error) (int, error) {
made := 0
cur, err := unix.Dup(rootFd)
if err != nil {
return made, err
}
defer func() { unix.Close(cur) }()
const dirFlags = unix.O_RDONLY | unix.O_DIRECTORY | unix.O_NOFOLLOW | unix.O_CLOEXEC
levels := strings.Split(p, "/")
for i, name := range levels {
next, err := unix.Openat(cur, name, dirFlags, 0)
if errors.Is(err, unix.ENOENT) {
// Made for its maker alone (the engine, as root), so it is nobody else's until it is the root's.
if mkErr := unix.Mkdirat(cur, name, 0o700); mkErr != nil && !errors.Is(mkErr, unix.EEXIST) {
return made, fmt.Errorf("a level of the path %s could not be made: %v", p, mkErr)
}
if next, err = unix.Openat(cur, name, dirFlags, 0); err == nil {
if err = unix.Fchown(next, int(rst.Uid), gid); err == nil {
err = unix.Fchmod(next, madeMode)
}
if err != nil {
unix.Close(next)
return made, fmt.Errorf("a level of the path %s made could not be given the root's owner "+
"and the library's group: %v", p, err)
}
made++
}
}
if errors.Is(err, unix.ELOOP) || errors.Is(err, unix.ENOTDIR) {
return made, fmt.Errorf("the path %s is not a folder below the root (a link, or a file), and the "+
"engine keeps no path through one", p)
}
if err != nil {
return made, fmt.Errorf("the path %s could not be opened: %v", p, err)
}
unix.Close(cur)
cur = next
var st unix.Stat_t
if err := stats.fd(fd, &st); err != nil {
unix.Close(fd)
return made, fmt.Errorf("a level made could not be read: %v", err)
if err := unix.Fstat(cur, &st); err != nil {
return made, err
}
if uint64(st.Dev) != uint64(anc.Dev) {
unix.Close(fd)
return made, errors.New("a level made is on another filesystem than its parent, and it stopped there")
if st.Dev != rst.Dev {
return made, fmt.Errorf("the path %s lies on another filesystem than the root, and the engine keeps "+
"nothing there", p)
}
if st.Uid != anc.Uid || st.Gid != anc.Gid {
if err := unix.Fchown(fd, int(anc.Uid), int(anc.Gid)); err != nil {
unix.Close(fd)
return made, fmt.Errorf("a level made could not be given its parent's owner: %v", err)
if i == len(levels)-1 {
if err := keep(cur, st); err != nil {
return made, err
}
}
if err := unix.Fchmod(fd, mode); err != nil {
unix.Close(fd)
return made, fmt.Errorf("a level made could not be given its parent's mode: %v", err)
}
}
unix.Close(fd)
return made, nil
}