| 1 | //go:build linux |
| 2 | |
| 3 | package host |
| 4 | |
| 5 | import ( |
| 6 | "context" |
| 7 | "crypto/rand" |
| 8 | "encoding/hex" |
| 9 | "errors" |
| 10 | "io" |
| 11 | "os" |
| 12 | "reflect" |
| 13 | |
| 14 | "github.com/carlosframework/platform/internal/hostbox" |
| 15 | "golang.org/x/sys/unix" |
| 16 | ) |
| 17 | |
| 18 | // Retain one obstruction and, if creation loses its identity write, one |
| 19 | // unrecorded replacement. Further replacements require investigation. |
| 20 | type servingParkQuarantine struct { |
| 21 | BootID string `json:",omitempty"` |
| 22 | Version int |
| 23 | ID string |
| 24 | Parent, Obstruction servingArtifactDirectory |
| 25 | Previous *servingParkNamespace `json:",omitempty"` |
| 26 | Moved bool |
| 27 | Unrecorded servingArtifactDirectory `json:",omitempty"` |
| 28 | UnrecordedMoved bool `json:",omitempty"` |
| 29 | Fresh servingArtifactDirectory |
| 30 | Complete bool |
| 31 | } |
| 32 | |
| 33 | func (q servingParkQuarantine) name() string { |
| 34 | return ".serving-parked-quarantine-" + q.ID |
| 35 | } |
| 36 | |
| 37 | func (q servingParkQuarantine) unrecordedName() string { |
| 38 | return ".serving-parked-unrecorded-" + q.ID |
| 39 | } |
| 40 | |
| 41 | func (q servingParkQuarantine) validate() error { |
| 42 | id, err := hex.DecodeString(q.ID) |
| 43 | if err != nil || len(id) != 16 || hex.EncodeToString(id) != q.ID || q.Version != 1 || q.Parent.Inode == 0 || q.Parent.Device == 0 || q.Parent.Mount == 0 || q.Obstruction.Inode == 0 || q.Obstruction.Device != q.Parent.Device || q.Obstruction.Mount != q.Parent.Mount { |
| 44 | return hostbox.ErrChanged |
| 45 | } |
| 46 | if q.Previous != nil && (q.Previous.Version != 1 || q.Previous.BootID != q.BootID || q.Previous.Parent != q.Parent || q.Previous.Root.Inode == 0 || q.Previous.Root.Device != q.Parent.Device || q.Previous.Root.Mount != q.Parent.Mount || q.Previous.Root == q.Obstruction) { |
| 47 | return hostbox.ErrChanged |
| 48 | } |
| 49 | if q.Fresh != (servingArtifactDirectory{}) && (!q.Moved || q.Fresh.Inode == 0 || q.Fresh.Device != q.Parent.Device || q.Fresh.Mount != q.Parent.Mount || q.Fresh == q.Obstruction || q.Previous != nil && q.Fresh == q.Previous.Root) { |
| 50 | return hostbox.ErrChanged |
| 51 | } |
| 52 | if q.Unrecorded != (servingArtifactDirectory{}) && (!q.Moved || q.Unrecorded.Inode == 0 || q.Unrecorded.Device != q.Parent.Device || q.Unrecorded.Mount != q.Parent.Mount || q.Unrecorded == q.Obstruction || q.Unrecorded == q.Fresh || q.Previous != nil && q.Unrecorded == q.Previous.Root || q.Fresh.Inode != 0 && !q.UnrecordedMoved) { |
| 53 | return hostbox.ErrChanged |
| 54 | } |
| 55 | if q.UnrecordedMoved && q.Unrecorded.Inode == 0 { |
| 56 | return hostbox.ErrChanged |
| 57 | } |
| 58 | if q.Complete && (!q.Moved || q.Fresh.Inode == 0) { |
| 59 | return hostbox.ErrChanged |
| 60 | } |
| 61 | return nil |
| 62 | } |
| 63 | |
| 64 | // Publishing namespace.json can succeed before quarantine completion. Keep |
| 65 | // ordinary parking out until that final record is durable, including after a |
| 66 | // lost reply, so recovery can still require an empty replacement root. |
| 67 | func (f *servingParkFS) checkNamespaceQuarantine(n servingParkNamespace) error { |
| 68 | body, err := f.state.Read("namespace-quarantine.json") |
| 69 | if errors.Is(err, os.ErrNotExist) { |
| 70 | return nil |
| 71 | } |
| 72 | if err != nil { |
| 73 | return err |
| 74 | } |
| 75 | var q servingParkQuarantine |
| 76 | if servingParkDecode(body, &q) != nil || q.validate() != nil || !q.Complete || n.Parent != q.Parent || n.Root != q.Fresh || n.BootID != q.BootID || f.checkBoot(q.BootID) != nil { |
| 77 | return hostbox.ErrChanged |
| 78 | } |
| 79 | return nil |
| 80 | } |
| 81 | |
| 82 | // This is an internal mechanism, not a request handler. The caller supplies a |
| 83 | // parent identity from independent root installation evidence and holds the |
| 84 | // namespace lock. No caller path, child listing or quarantine deletion exists. |
| 85 | func (f *servingParkFS) recoverNamespace(ctx context.Context, expected servingArtifactDirectory) (servingParkQuarantine, error) { |
| 86 | var q servingParkQuarantine |
| 87 | if ctx == nil { |
| 88 | return q, hostbox.ErrUnsupported |
| 89 | } |
| 90 | if err := ctx.Err(); err != nil { |
| 91 | return q, err |
| 92 | } |
| 93 | boot, err := f.currentBoot() |
| 94 | if err != nil { |
| 95 | return q, err |
| 96 | } |
| 97 | parent, pid, err := f.parent() |
| 98 | if err != nil { |
| 99 | return q, err |
| 100 | } |
| 101 | defer parent.Close() |
| 102 | var st unix.Stat_t |
| 103 | if err = unix.Fstat(int(parent.Fd()), &st); err != nil { |
| 104 | return q, err |
| 105 | } |
| 106 | if pid != expected || st.Mode&0022 != 0 { |
| 107 | return q, hostbox.ErrChanged |
| 108 | } |
| 109 | ledger := servingParkLedger{state: f.state} |
| 110 | anchor, err := f.namespaceParent() |
| 111 | if err == nil && (anchor.Parent != expected || f.checkBoot(anchor.BootID) != nil) { |
| 112 | return q, hostbox.ErrChanged |
| 113 | } |
| 114 | if err != nil && !errors.Is(err, os.ErrNotExist) { |
| 115 | return q, err |
| 116 | } |
| 117 | save := func() error { |
| 118 | if err := ctx.Err(); err != nil { |
| 119 | return err |
| 120 | } |
| 121 | if err := q.validate(); err != nil { |
| 122 | return err |
| 123 | } |
| 124 | return ledger.write("namespace-quarantine.json", q) |
| 125 | } |
| 126 | body, err := f.state.Read("namespace-quarantine.json") |
| 127 | if err == nil { |
| 128 | if servingParkDecode(body, &q) != nil || q.validate() != nil || q.Parent != pid || f.checkBoot(q.BootID) != nil { |
| 129 | return q, hostbox.ErrChanged |
| 130 | } |
| 131 | } else if errors.Is(err, os.ErrNotExist) { |
| 132 | q.Version, q.Parent, q.BootID = 1, pid, boot |
| 133 | body, err = f.state.Read("namespace.json") |
| 134 | if err == nil { |
| 135 | var prior servingParkNamespace |
| 136 | if servingParkDecode(body, &prior) != nil || prior.Version != 1 || prior.Parent != pid || f.checkBoot(prior.BootID) != nil { |
| 137 | return q, hostbox.ErrChanged |
| 138 | } |
| 139 | q.Previous = &prior |
| 140 | } else if !errors.Is(err, os.ErrNotExist) { |
| 141 | return q, err |
| 142 | } |
| 143 | obstruction, err := servingParkOpenDir(parent, ".serving-parked") |
| 144 | if err != nil { |
| 145 | return q, err |
| 146 | } |
| 147 | q.Obstruction, err = servingParkDirID(obstruction) |
| 148 | obstruction.Close() |
| 149 | if err != nil { |
| 150 | return q, err |
| 151 | } |
| 152 | var nonce [16]byte |
| 153 | if _, err = rand.Read(nonce[:]); err != nil { |
| 154 | return q, err |
| 155 | } |
| 156 | q.ID = hex.EncodeToString(nonce[:]) |
| 157 | if err = save(); err != nil { |
| 158 | return q, err |
| 159 | } |
| 160 | } else { |
| 161 | return q, err |
| 162 | } |
| 163 | // A private record is necessary but not sufficient: it must still describe |
| 164 | // the installed namespace, or the exact replacement recorded by this run. |
| 165 | body, err = f.state.Read("namespace.json") |
| 166 | if err == nil { |
| 167 | var current servingParkNamespace |
| 168 | if servingParkDecode(body, ¤t) != nil { |
| 169 | return q, hostbox.ErrChanged |
| 170 | } |
| 171 | old := q.Previous != nil && reflect.DeepEqual(current, *q.Previous) |
| 172 | fresh := q.Fresh.Inode != 0 && current.Version == 1 && current.Parent == q.Parent && current.Root == q.Fresh && current.BootID == q.BootID |
| 173 | if !old && !fresh || q.Complete && !fresh { |
| 174 | return q, hostbox.ErrChanged |
| 175 | } |
| 176 | } else if !errors.Is(err, os.ErrNotExist) || q.Previous != nil || q.Complete { |
| 177 | return q, err |
| 178 | } |
| 179 | if !q.Moved { |
| 180 | if err = f.moveQuarantine(ctx, parent, q); err != nil { |
| 181 | return q, err |
| 182 | } |
| 183 | q.Moved = true |
| 184 | if err = save(); err != nil { |
| 185 | return q, err |
| 186 | } |
| 187 | } |
| 188 | if err = f.checkQuarantine(parent, q); err != nil { |
| 189 | return q, err |
| 190 | } |
| 191 | if q.Fresh.Inode == 0 { |
| 192 | if err = f.createQuarantineReplacement(ctx, parent, &q, save); err != nil { |
| 193 | return q, err |
| 194 | } |
| 195 | } |
| 196 | if q.UnrecordedMoved { |
| 197 | if err = f.checkRetainedDirectory(parent, q.unrecordedName(), q.Unrecorded); err != nil { |
| 198 | return q, err |
| 199 | } |
| 200 | } |
| 201 | fresh, id, err := f.privateDir(parent, ".serving-parked") |
| 202 | if err != nil { |
| 203 | return q, err |
| 204 | } |
| 205 | defer fresh.Close() |
| 206 | if id != q.Fresh { |
| 207 | return q, hostbox.ErrChanged |
| 208 | } |
| 209 | // Recheck the effective mount/path before publishing the replacement. |
| 210 | check, current, err := f.parent() |
| 211 | if err != nil { |
| 212 | return q, err |
| 213 | } |
| 214 | check.Close() |
| 215 | if current != q.Parent { |
| 216 | return q, hostbox.ErrChanged |
| 217 | } |
| 218 | if q.Complete { |
| 219 | return q, ctx.Err() |
| 220 | } |
| 221 | // No retirement can use this root before publication. Even a directory |
| 222 | // with the right owner and mode is refused if it already contains data. |
| 223 | names, err := fresh.Readdirnames(1) |
| 224 | if len(names) != 0 || err != nil && !errors.Is(err, io.EOF) { |
| 225 | return q, hostbox.ErrChanged |
| 226 | } |
| 227 | if err = ctx.Err(); err != nil { |
| 228 | return q, err |
| 229 | } |
| 230 | if err = f.sync(int(parent.Fd())); err != nil { |
| 231 | return q, err |
| 232 | } |
| 233 | n := servingParkNamespace{Version: 1, Parent: q.Parent, Root: q.Fresh, BootID: q.BootID} |
| 234 | n.Topology, err = f.captureTopology(parent, fresh, n) |
| 235 | if err != nil { |
| 236 | return q, err |
| 237 | } |
| 238 | if err = ledger.write("namespace.json", n); err != nil { |
| 239 | return q, err |
| 240 | } |
| 241 | q.Complete = true |
| 242 | return q, save() |
| 243 | } |
| 244 | |
| 245 | func (f *servingParkFS) checkQuarantine(parent *os.File, q servingParkQuarantine) error { |
| 246 | return f.checkRetainedDirectory(parent, q.name(), q.Obstruction) |
| 247 | } |
| 248 | |
| 249 | func (f *servingParkFS) checkRetainedDirectory(parent *os.File, name string, expected servingArtifactDirectory) error { |
| 250 | d, err := servingParkOpenDir(parent, name) |
| 251 | if err != nil { |
| 252 | return err |
| 253 | } |
| 254 | defer d.Close() |
| 255 | id, err := servingParkDirID(d) |
| 256 | if err != nil || id != expected { |
| 257 | return hostbox.ErrChanged |
| 258 | } |
| 259 | return nil |
| 260 | } |
| 261 | |
| 262 | func (f *servingParkFS) moveQuarantine(ctx context.Context, parent *os.File, q servingParkQuarantine) error { |
| 263 | return f.moveRetainedDirectory(ctx, parent, q.name(), q.Obstruction) |
| 264 | } |
| 265 | |
| 266 | func (f *servingParkFS) moveRetainedDirectory(ctx context.Context, parent *os.File, name string, expected servingArtifactDirectory) error { |
| 267 | if err := ctx.Err(); err != nil { |
| 268 | return err |
| 269 | } |
| 270 | source, err := servingParkOpenDir(parent, ".serving-parked") |
| 271 | if errors.Is(err, os.ErrNotExist) { |
| 272 | // The rename may have succeeded before its reply or fsync was lost. |
| 273 | if err = f.checkRetainedDirectory(parent, name, expected); err != nil { |
| 274 | return err |
| 275 | } |
| 276 | return f.sync(int(parent.Fd())) |
| 277 | } |
| 278 | if err != nil { |
| 279 | return err |
| 280 | } |
| 281 | defer source.Close() |
| 282 | id, err := servingParkDirID(source) |
| 283 | if err != nil || id != expected { |
| 284 | return hostbox.ErrChanged |
| 285 | } |
| 286 | if err = f.rename(int(parent.Fd()), ".serving-parked", int(parent.Fd()), name, unix.RENAME_NOREPLACE); err != nil { |
| 287 | return err |
| 288 | } |
| 289 | if err = f.checkRetainedDirectory(parent, name, expected); err != nil { |
| 290 | return err |
| 291 | } |
| 292 | return f.sync(int(parent.Fd())) |
| 293 | } |
| 294 | |
| 295 | // EEXIST grants retention only. Persist the unexpected identity before moving |
| 296 | // it, and never inspect its contents. One spare name bounds crash recovery even |
| 297 | // if the service owner keeps replacing the root or private writes keep failing. |
| 298 | func (f *servingParkFS) createQuarantineReplacement(ctx context.Context, parent *os.File, q *servingParkQuarantine, save func() error) error { |
| 299 | for { |
| 300 | if err := ctx.Err(); err != nil { |
| 301 | return err |
| 302 | } |
| 303 | if q.Unrecorded.Inode != 0 { |
| 304 | if !q.UnrecordedMoved { |
| 305 | if err := f.moveRetainedDirectory(ctx, parent, q.unrecordedName(), q.Unrecorded); err != nil { |
| 306 | return err |
| 307 | } |
| 308 | q.UnrecordedMoved = true |
| 309 | if err := save(); err != nil { |
| 310 | return err |
| 311 | } |
| 312 | } |
| 313 | if err := f.checkRetainedDirectory(parent, q.unrecordedName(), q.Unrecorded); err != nil { |
| 314 | return err |
| 315 | } |
| 316 | } |
| 317 | err := unix.Mkdirat(int(parent.Fd()), ".serving-parked", 0700) |
| 318 | if errors.Is(err, unix.EEXIST) && q.Unrecorded.Inode == 0 { |
| 319 | d, err := servingParkOpenDir(parent, ".serving-parked") |
| 320 | if err != nil { |
| 321 | return err |
| 322 | } |
| 323 | id, err := servingParkDirID(d) |
| 324 | d.Close() |
| 325 | if err != nil { |
| 326 | return err |
| 327 | } |
| 328 | q.Unrecorded = id |
| 329 | if err := save(); err != nil { |
| 330 | return err |
| 331 | } |
| 332 | continue |
| 333 | } |
| 334 | if err != nil { |
| 335 | return err |
| 336 | } |
| 337 | fresh, id, err := f.privateDir(parent, ".serving-parked") |
| 338 | if err != nil { |
| 339 | return err |
| 340 | } |
| 341 | fresh.Close() |
| 342 | q.Fresh = id |
| 343 | // Directory durability precedes the private identity record. Otherwise |
| 344 | // a reboot could retain the record while losing its directory entry. |
| 345 | if err := f.sync(int(parent.Fd())); err != nil { |
| 346 | return err |
| 347 | } |
| 348 | return save() |
| 349 | } |
| 350 | } |
| 351 | |
amadan