| 1 | //go:build linux |
| 2 | |
| 3 | package host |
| 4 | |
| 5 | import ( |
| 6 | "context" |
| 7 | "encoding/hex" |
| 8 | "errors" |
| 9 | "os" |
| 10 | |
| 11 | "github.com/carlosframework/platform/internal/hostbox" |
| 12 | "golang.org/x/sys/unix" |
| 13 | ) |
| 14 | |
| 15 | type servingParkTopology struct { |
| 16 | Parent, Root servingParkPersistentDirectory |
| 17 | } |
| 18 | |
| 19 | // This only obtains a handle for an already-open directory. It never opens a |
| 20 | // supplied handle, follows a request path or changes mount permissions. |
| 21 | func servingParkPersistentDir(d *os.File) (servingParkPersistentDirectory, error) { |
| 22 | var out servingParkPersistentDirectory |
| 23 | if d == nil { |
| 24 | return out, hostbox.ErrUnsupported |
| 25 | } |
| 26 | id, err := servingParkDirID(d) |
| 27 | if err != nil { |
| 28 | return out, err |
| 29 | } |
| 30 | var st unix.Stat_t |
| 31 | if err = unix.Fstat(int(d.Fd()), &st); err != nil { |
| 32 | return out, err |
| 33 | } |
| 34 | var fs unix.Statfs_t |
| 35 | if err = unix.Fstatfs(int(d.Fd()), &fs); err != nil { |
| 36 | return out, err |
| 37 | } |
| 38 | h, mount, err := unix.NameToHandleAt(int(d.Fd()), "", unix.AT_EMPTY_PATH) |
| 39 | if err != nil { |
| 40 | return out, err |
| 41 | } |
| 42 | if mount <= 0 || uint64(mount) != id.Mount { |
| 43 | return out, hostbox.ErrChanged |
| 44 | } |
| 45 | out = servingParkPersistentDirectory{Filesystem: fs.Fsid.Val, Type: h.Type(), Handle: hex.EncodeToString(h.Bytes()), Owner: st.Uid, Mode: st.Mode} |
| 46 | if !out.valid() { |
| 47 | return servingParkPersistentDirectory{}, hostbox.ErrChanged |
| 48 | } |
| 49 | return out, nil |
| 50 | } |
| 51 | |
| 52 | // Capture only while the recorded boot and both descriptor identities still |
| 53 | // match. Old namespaces are never upgraded by looking at their current paths. |
| 54 | func (f *servingParkFS) captureTopology(parent, root *os.File, n servingParkNamespace) (*servingParkTopology, error) { |
| 55 | if f.handle == nil { |
| 56 | return nil, hostbox.ErrUnsupported |
| 57 | } |
| 58 | if err := f.checkBoot(n.BootID); err != nil { |
| 59 | return nil, err |
| 60 | } |
| 61 | p, err := servingParkDirID(parent) |
| 62 | if err != nil || p != n.Parent { |
| 63 | return nil, hostbox.ErrChanged |
| 64 | } |
| 65 | r, err := servingParkDirID(root) |
| 66 | if err != nil || r != n.Root || r.Device != p.Device || r.Mount != p.Mount { |
| 67 | return nil, hostbox.ErrChanged |
| 68 | } |
| 69 | ph, err := f.handle(parent) |
| 70 | if err != nil { |
| 71 | return nil, err |
| 72 | } |
| 73 | rh, err := f.handle(root) |
| 74 | if err != nil { |
| 75 | return nil, err |
| 76 | } |
| 77 | if !ph.valid() || !rh.valid() || ph.Filesystem != rh.Filesystem || rh.Owner != f.owner || rh.Mode&0077 != 0 { |
| 78 | return nil, hostbox.ErrChanged |
| 79 | } |
| 80 | p, err = servingParkDirID(parent) |
| 81 | if err != nil || p != n.Parent { |
| 82 | return nil, hostbox.ErrChanged |
| 83 | } |
| 84 | r, err = servingParkDirID(root) |
| 85 | if err != nil || r != n.Root || f.checkBoot(n.BootID) != nil { |
| 86 | return nil, hostbox.ErrChanged |
| 87 | } |
| 88 | return &servingParkTopology{Parent: ph, Root: rh}, nil |
| 89 | } |
| 90 | |
| 91 | // This observation cannot be used by the park filesystem or sent to an agent. |
| 92 | // A later recovery transaction must preserve retirement history and validate |
| 93 | // every affected source/event before it can publish current mount authority. |
| 94 | type servingParkRebootObservation struct { |
| 95 | Volume string `json:",omitempty"` |
| 96 | Previous servingParkNamespace |
| 97 | Parent servingParkParent |
| 98 | BootID string |
| 99 | CurrentParent, CurrentRoot servingArtifactDirectory |
| 100 | } |
| 101 | |
| 102 | // Called under the namespace lock. This is a read-only prerequisite for reboot |
| 103 | // recovery, not permission to rewrite mount IDs or resume deletion. Quarantines |
| 104 | // require a separate transition covering the displaced root and retained data. |
| 105 | func (f *servingParkFS) inspectRebootTopology(ctx context.Context) (servingParkRebootObservation, error) { |
| 106 | var zero servingParkRebootObservation |
| 107 | if ctx == nil || f == nil || f.state == nil || f.handle == nil { |
| 108 | return zero, hostbox.ErrUnsupported |
| 109 | } |
| 110 | if err := ctx.Err(); err != nil { |
| 111 | return zero, err |
| 112 | } |
| 113 | boot, err := f.currentBoot() |
| 114 | if err != nil { |
| 115 | return zero, err |
| 116 | } |
| 117 | p, err := f.historicalNamespaceParent() |
| 118 | if err != nil { |
| 119 | return zero, err |
| 120 | } |
| 121 | var n servingParkNamespace |
| 122 | body, err := f.state.Read("namespace.json") |
| 123 | if err != nil { |
| 124 | return zero, err |
| 125 | } |
| 126 | if servingParkDecode(body, &n) != nil || n.Version != 1 || !validServingParkBootID(n.BootID) || n.BootID != p.BootID || n.Parent != p.Parent || n.Topology == nil || !n.Topology.Parent.valid() || !n.Topology.Root.valid() || n.Root.Device != n.Parent.Device || n.Root.Mount != n.Parent.Mount || n.Root.Inode == 0 { |
| 127 | return zero, hostbox.ErrChanged |
| 128 | } |
| 129 | if _, err := f.state.Read("namespace-quarantine.json"); !errors.Is(err, os.ErrNotExist) { |
| 130 | if err == nil { |
| 131 | err = hostbox.ErrChanged |
| 132 | } |
| 133 | return zero, err |
| 134 | } |
| 135 | parent, pid, err := f.parent() |
| 136 | if err != nil { |
| 137 | return zero, err |
| 138 | } |
| 139 | defer parent.Close() |
| 140 | root, rid, err := f.privateDir(parent, ".serving-parked") |
| 141 | if err != nil { |
| 142 | return zero, err |
| 143 | } |
| 144 | defer root.Close() |
| 145 | // Device and mount numbers are boot-local. Across boots, require the same |
| 146 | // filesystem and persistent handles below, plus unchanged inode numbers. |
| 147 | // Both directories must still share the current device and mount. |
| 148 | if n.BootID == boot && pid.Device != n.Parent.Device { |
| 149 | return zero, hostbox.ErrChanged |
| 150 | } |
| 151 | if n.BootID == boot && pid.Mount == n.Parent.Mount { |
| 152 | prior, err := servingParkReadBinding(f.state, "namespace") |
| 153 | if err != nil || prior == nil || prior.BeforeBoot != n.BootID || prior.BeforeParent != n.Parent { |
| 154 | return zero, hostbox.ErrChanged |
| 155 | } |
| 156 | } |
| 157 | if pid.Inode != n.Parent.Inode || rid.Device != pid.Device || rid.Inode != n.Root.Inode || rid.Mount != pid.Mount { |
| 158 | return zero, hostbox.ErrChanged |
| 159 | } |
| 160 | current := servingParkNamespace{Version: 1, Parent: pid, Root: rid, BootID: boot} |
| 161 | volume := "" |
| 162 | if f.volume != nil { |
| 163 | volume = f.volume(pid.Device) |
| 164 | } |
| 165 | topology, err := f.captureTopology(parent, root, current) |
| 166 | if err != nil { |
| 167 | return zero, err |
| 168 | } |
| 169 | if *topology != *n.Topology { |
| 170 | return zero, hostbox.ErrChanged |
| 171 | } |
| 172 | // The service user owns the mount parent. Reopen both names after reading |
| 173 | // the persistent handles; a detached original descriptor proves no path. |
| 174 | again, nextParent, err := f.parent() |
| 175 | if err != nil { |
| 176 | return zero, err |
| 177 | } |
| 178 | defer again.Close() |
| 179 | next, nextRoot, err := f.privateDir(again, ".serving-parked") |
| 180 | if err != nil { |
| 181 | return zero, err |
| 182 | } |
| 183 | defer next.Close() |
| 184 | if nextParent != pid || nextRoot != rid || f.checkBoot(boot) != nil { |
| 185 | return zero, hostbox.ErrChanged |
| 186 | } |
| 187 | againTopology, err := f.captureTopology(again, next, current) |
| 188 | if err != nil { |
| 189 | return zero, err |
| 190 | } |
| 191 | if *againTopology != *topology || f.volume != nil && f.volume(nextParent.Device) != volume { |
| 192 | return zero, hostbox.ErrChanged |
| 193 | } |
| 194 | if err := ctx.Err(); err != nil { |
| 195 | return zero, err |
| 196 | } |
| 197 | return servingParkRebootObservation{Volume: volume, Previous: n, Parent: p, BootID: boot, CurrentParent: pid, CurrentRoot: rid}, nil |
| 198 | } |
| 199 | |
amadan