| 1 | package activator |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "crypto/sha256" |
| 6 | "encoding/hex" |
| 7 | "errors" |
| 8 | "fmt" |
| 9 | "io" |
| 10 | "log" |
| 11 | "os" |
| 12 | "path/filepath" |
| 13 | "strconv" |
| 14 | "strings" |
| 15 | |
| 16 | "github.com/carlosframework/platform/internal/instancefs" |
| 17 | "github.com/carlosframework/platform/internal/lease" |
| 18 | "github.com/carlosframework/platform/internal/registry" |
| 19 | "github.com/carlosframework/platform/internal/seeds" |
| 20 | ) |
| 21 | |
| 22 | // SeedObjects reads a seed's snapshot, seeds/<host>/<id>/snapshot.db, from |
| 23 | // this box's own region data bucket. |
| 24 | type SeedObjects interface { |
| 25 | Open(ctx context.Context, host, id string) (io.ReadCloser, error) |
| 26 | } |
| 27 | |
| 28 | // SeedMarkerSuffix names the installation evidence beside a placed |
| 29 | // database: <db>.seed holds the seed's id (spec §3.4 step 3, codex R2-1). |
| 30 | const SeedMarkerSuffix = ".seed" |
| 31 | |
| 32 | // seedMarkerMax bounds the marker read: it holds a 29-byte id and a newline, |
| 33 | // and it sits in tenant-writable ground. |
| 34 | const seedMarkerMax = 128 |
| 35 | |
| 36 | // hasOwnReplica and checkSeedDB are seams for the tests. |
| 37 | var ( |
| 38 | hasOwnReplica = lease.HasReplica |
| 39 | checkSeedDB = lease.CheckDBContext |
| 40 | ) |
| 41 | |
| 42 | // SeedPlacementError is a wake the seed fallback failed or refused. Its |
| 43 | // Sentence is the member's (SD-N14 to SD-N19): memberSafeWakeReason passes |
| 44 | // it to carlos logs whole. Err is the cause, for the box log. |
| 45 | type SeedPlacementError struct { |
| 46 | Sentence string |
| 47 | Err error |
| 48 | } |
| 49 | |
| 50 | func (e *SeedPlacementError) Error() string { return e.Sentence } |
| 51 | func (e *SeedPlacementError) Unwrap() error { return e.Err } |
| 52 | |
| 53 | func seedMismatch(host string) error { |
| 54 | return &SeedPlacementError{Sentence: fmt.Sprintf("%s could not start from its seed: the snapshot's length or sha256 does not match its record.", host)} |
| 55 | } |
| 56 | |
| 57 | func seedNotDB(host string, err error) error { |
| 58 | return &SeedPlacementError{Sentence: fmt.Sprintf("%s could not start from its seed: the snapshot is not a database that passes an integrity check.", host), Err: err} |
| 59 | } |
| 60 | |
| 61 | func seedUnread(host string, err error) error { |
| 62 | return &SeedPlacementError{Sentence: fmt.Sprintf("%s could not start from its seed: the snapshot could not be read. The next wake tries again.", host), Err: err} |
| 63 | } |
| 64 | |
| 65 | func seedRefusedFlat(host string) error { |
| 66 | return &SeedPlacementError{Sentence: fmt.Sprintf("%s could not start from its seed: this box does not run instances nested.", host)} |
| 67 | } |
| 68 | |
| 69 | func seedRefusedClaim(host string) error { |
| 70 | return &SeedPlacementError{Sentence: fmt.Sprintf("%s could not start from its seed: claim control is on for this box.", host)} |
| 71 | } |
| 72 | |
| 73 | func seedPlaceFailed(host string, err error) error { |
| 74 | return &SeedPlacementError{Sentence: fmt.Sprintf("%s could not start from its seed: it could not be put in place. The next wake tries again.", host), Err: err} |
| 75 | } |
| 76 | |
| 77 | // ownsReplica reports a seed whose host already has a replica of its own: |
| 78 | // placed is set only after the first upload, and committed only from |
| 79 | // placed (spec §5.4). Such a row restores its own lineage and never boots |
| 80 | // fresh: after commit the app retires its old process, so an empty boot |
| 81 | // would read as the import's data lost (preflight M11). |
| 82 | func ownsReplica(src *seeds.RowSeed) bool { |
| 83 | return src.State == seeds.StatePlaced || src.State == seeds.StateCommitted |
| 84 | } |
| 85 | |
| 86 | // seedRestores reports a row Restore sends through seededRestore: a seed |
| 87 | // still to place, or one whose own replica exists. |
| 88 | func seedRestores(src *seeds.RowSeed) bool { |
| 89 | return src != nil && (src.Placeable() || ownsReplica(src)) |
| 90 | } |
| 91 | |
| 92 | // seededRestore is the plain-path seed fallback (spec §3.4). It never |
| 93 | // reports fresh: a seeded row never runs the manifest's seed command, and |
| 94 | // every failure fails the wake instead of falling through to an empty |
| 95 | // database. Every file it touches in the instance directory goes through |
| 96 | // instancefs, so no symlink there is followed. |
| 97 | func (b *ExecBackend) seededRestore(ctx context.Context, rt registry.Route, src *seeds.RowSeed) (bool, error) { |
| 98 | host := rt.Host |
| 99 | fail := func(err error) (bool, error) { |
| 100 | var spe *SeedPlacementError |
| 101 | if errors.As(err, &spe) && spe.Err != nil { |
| 102 | log.Printf("activator: %s: seed %s: %v: %v", host, src.ID, err, spe.Err) |
| 103 | } else { |
| 104 | log.Printf("activator: %s: seed %s: %v", host, src.ID, err) |
| 105 | } |
| 106 | return false, err |
| 107 | } |
| 108 | if err := seeds.Check(host, src.ID); err != nil { |
| 109 | return fail(fmt.Errorf("seed fallback for %s: %w", host, err)) |
| 110 | } |
| 111 | // restoreOwn restores the host's own lineage, the staging fallback |
| 112 | // included, and never reports fresh: a host that has (or was listed as |
| 113 | // having) a replica of its own and finds none must not boot empty. The |
| 114 | // next wake asks again. |
| 115 | // |
| 116 | // The database it leaves is this seed's lineage, so its <db>.seed |
| 117 | // marker is written to name the seed (final review carry item): the |
| 118 | // rollback's missing-row purge and an aborted seed's local purge decide |
| 119 | // by the marker, and a database with none is left in place. |
| 120 | restoreOwn := func() (bool, error) { |
| 121 | fresh, err := b.restorePlain(ctx, rt) |
| 122 | if err != nil { |
| 123 | return false, err |
| 124 | } |
| 125 | if fresh { |
| 126 | return fail(fmt.Errorf("seed fallback for %s: found no replica of its own to restore", host)) |
| 127 | } |
| 128 | db := b.pathsFor(rt).dbPath |
| 129 | if err := instancefs.WriteAtomic(filepath.Dir(db), filepath.Base(db)+SeedMarkerSuffix, []byte(src.ID+"\n"), 0o600); err != nil { |
| 130 | return fail(seedPlaceFailed(host, err)) |
| 131 | } |
| 132 | return false, nil |
| 133 | } |
| 134 | // A placed seed stays on the row for life, and placed is set only once |
| 135 | // the host's own replica exists. So a placed row's wake is a plain |
| 136 | // restore on any box: no layout or claim refusal, no listing on every |
| 137 | // wake, and the staging fallback still promotes a moved host's replica |
| 138 | // (the snapshot was deleted at placement; it is never read again). A |
| 139 | // committed import was placed first, so it is restored the same way. |
| 140 | if ownsReplica(src) { |
| 141 | return restoreOwn() |
| 142 | } |
| 143 | if b.ClaimArmed { |
| 144 | return fail(seedRefusedClaim(host)) |
| 145 | } |
| 146 | paths := b.pathsFor(rt) |
| 147 | if paths.layout != LayoutNested { |
| 148 | return fail(seedRefusedFlat(host)) |
| 149 | } |
| 150 | dir, base := filepath.Dir(paths.dbPath), filepath.Base(paths.dbPath) |
| 151 | // The instance directory may not exist yet on a first wake. Created as |
| 152 | // prepare and lease.Restore create it. |
| 153 | if err := os.MkdirAll(dir, 0o700); err != nil { |
| 154 | return fail(seedPlaceFailed(host, err)) |
| 155 | } |
| 156 | |
| 157 | // Step 1: a replica of its own means this host was placed already. A |
| 158 | // listing that cannot tell fails the wake: read as "no replica", it |
| 159 | // would place the snapshot over a lineage that exists. |
| 160 | has, err := hasOwnReplica(ctx, b.Cfg.Bucket, b.Cfg.hibernatePrefix(host), b.Cfg.Region) |
| 161 | if err != nil { |
| 162 | return fail(fmt.Errorf("seed fallback for %s: cannot tell whether it has a replica of its own: %w", host, err)) |
| 163 | } |
| 164 | if has { |
| 165 | if _, err := restoreOwn(); err != nil { |
| 166 | return false, err |
| 167 | } |
| 168 | b.finishPlacement(ctx, rt, src) |
| 169 | return false, nil |
| 170 | } |
| 171 | |
| 172 | // An earlier first-boot seed command at this path whose processes were |
| 173 | // not proven gone left an obligation beside the database |
| 174 | // (settleSeedObligation, as restorePlain answers it): it is answered |
| 175 | // before the file there is kept as this seed's, purged or replaced, so |
| 176 | // nothing is placed under a process that may still write the path. A |
| 177 | // seeded row runs no seed command itself, so this is another row's |
| 178 | // leftover. Unproven, the wake is refused and the lease left to its |
| 179 | // TTL, as on the plain path. |
| 180 | if err := b.settleSeedObligation(ctx, rt, paths.dbPath); err != nil { |
| 181 | return fail(err) |
| 182 | } |
| 183 | |
| 184 | // This seed, placed by an interrupted first wake: keep it. Replication |
| 185 | // starts in doWake, as generation one of the host's own lineage. |
| 186 | placed, err := seedPlacedAt(dir, base, src.ID) |
| 187 | if err != nil { |
| 188 | return fail(seedPlaceFailed(host, err)) |
| 189 | } |
| 190 | if placed { |
| 191 | log.Printf("activator: %s: keeping seed %s placed by an earlier wake", host, src.ID) |
| 192 | return false, nil |
| 193 | } |
| 194 | if b.Seeds == nil { |
| 195 | return fail(seedUnread(host, errors.New("no seed reader on this box"))) |
| 196 | } |
| 197 | // Anything else is purged before the fetch: another seed's database, |
| 198 | // one with no marker, their WAL and SHM, the marker, litestream's |
| 199 | // metadata, and an earlier fetch's leftovers. |
| 200 | if err := purgeSeedTarget(dir, base); err != nil { |
| 201 | return fail(seedPlaceFailed(host, err)) |
| 202 | } |
| 203 | if err := b.placeSeed(ctx, host, src, dir, base); err != nil { |
| 204 | return fail(err) |
| 205 | } |
| 206 | log.Printf("activator: %s: placed seed %s (%d bytes)", host, src.ID, src.Bytes) |
| 207 | return false, nil |
| 208 | } |
| 209 | |
| 210 | // reconcileSeedMarker runs first on the plain restore path, whether or not |
| 211 | // a database is at dbPath (final review F2b, re-review finding 8). A |
| 212 | // <db>.seed marker beside the path that names a seed other than the row's |
| 213 | // own current seed (the row has none, or another) is a leftover of that |
| 214 | // seed: an aborted --from seed whose purge ran on no box, a rolled-back |
| 215 | // import, or a placement that crashed between its marker and its rename |
| 216 | // and left the marker alone. Whatever is there goes, through |
| 217 | // purgeSeedTarget (the database and its sidecars, the marker, any |
| 218 | // .seed-sd_*.tmp), before anything is restored or created, so the file is |
| 219 | // never served as this row's warm cache and no new database is ever born |
| 220 | // beside a foreign marker that a later wake would read as the old seed's. |
| 221 | // A marker naming the row's own seed, or none, changes nothing. A marker |
| 222 | // that cannot be read fails the wake, with nothing served or purged. |
| 223 | // |
| 224 | // A relative path is no seeded path (seeded rows are always given the |
| 225 | // nested, absolute one), so it is left to the restore as before. |
| 226 | func reconcileSeedMarker(rt registry.Route, dbPath string) error { |
| 227 | if !filepath.IsAbs(dbPath) { |
| 228 | return nil |
| 229 | } |
| 230 | dbPath = filepath.Clean(dbPath) |
| 231 | marker, err := ReadSeedMarker(dbPath) |
| 232 | if err != nil { |
| 233 | return fmt.Errorf("restore %s: the seed marker beside %s cannot be read: %w", rt.Host, dbPath, err) |
| 234 | } |
| 235 | own := rowSeedID(rt.SeedSource) |
| 236 | if marker == "" || marker == own { |
| 237 | return nil |
| 238 | } |
| 239 | dir, base, err := splitDBPath(dbPath) |
| 240 | if err != nil { |
| 241 | return err |
| 242 | } |
| 243 | if err := purgeSeedTarget(dir, base); err != nil { |
| 244 | return fmt.Errorf("restore %s: purge what seed %s left at %s: %w", rt.Host, marker, dbPath, err) |
| 245 | } |
| 246 | log.Printf("activator: %s: %s held seed %s's marker, not this route's (seed %q): purged before the restore", rt.Host, dbPath, marker, own) |
| 247 | return nil |
| 248 | } |
| 249 | |
| 250 | // seedPlacedAt reports whether dir/base is a regular file whose marker names |
| 251 | // id. A missing, symlinked or oversized marker or database is "not placed", |
| 252 | // and is purged; any other error is returned. |
| 253 | func seedPlacedAt(dir, base, id string) (bool, error) { |
| 254 | m, err := instancefs.ReadBeneathLimit(dir, base+SeedMarkerSuffix, seedMarkerMax) |
| 255 | if err != nil { |
| 256 | if notAPlacedFile(err) { |
| 257 | return false, nil |
| 258 | } |
| 259 | return false, err |
| 260 | } |
| 261 | if strings.TrimSpace(string(m)) != id { |
| 262 | return false, nil |
| 263 | } |
| 264 | f, err := instancefs.OpenBeneath(dir, base, os.O_RDONLY, 0) |
| 265 | if err != nil { |
| 266 | if notAPlacedFile(err) { |
| 267 | return false, nil |
| 268 | } |
| 269 | return false, err |
| 270 | } |
| 271 | f.Close() |
| 272 | return true, nil |
| 273 | } |
| 274 | |
| 275 | func notAPlacedFile(err error) bool { |
| 276 | return errors.Is(err, os.ErrNotExist) || errors.Is(err, instancefs.ErrSymlink) || |
| 277 | errors.Is(err, instancefs.ErrNotRegular) || errors.Is(err, instancefs.ErrTooLarge) |
| 278 | } |
| 279 | |
| 280 | // seedTempName is where a seed's snapshot is fetched to, in the instance |
| 281 | // directory. The id is a seed id (seeds.Check), one clean path element. |
| 282 | func seedTempName(id string) string { return ".seed-" + id + ".tmp" } |
| 283 | |
| 284 | // PurgeSeedLocal removes an instance's local database at dbPath and |
| 285 | // everything beside it that could stand in for one (purgeSeedTarget): the |
| 286 | // database, -wal, -shm, the <db>.seed marker, litestream's metadata and |
| 287 | // any seed fetch leftovers, through instancefs. It is the import |
| 288 | // rollback's local removal (seeded-instances spec §5.6 step 3), run after |
| 289 | // the host is fenced and stopped. A directory that is not there holds |
| 290 | // nothing, so a second run, or one for an instance never woken, is done. |
| 291 | func PurgeSeedLocal(dbPath string) error { |
| 292 | dir, base, err := splitDBPath(dbPath) |
| 293 | if err != nil { |
| 294 | return err |
| 295 | } |
| 296 | if _, err := os.Lstat(dir); errors.Is(err, os.ErrNotExist) { |
| 297 | return nil |
| 298 | } |
| 299 | return purgeSeedTarget(dir, base) |
| 300 | } |
| 301 | |
| 302 | // ReadSeedMarker is the seed id the <db>.seed marker beside dbPath names, |
| 303 | // read through instancefs: the evidence that the database there was placed |
| 304 | // for that seed (spec §3.4 step 3). No marker, or one that is a symlink, |
| 305 | // not a regular file or oversized, is "" (it names no seed); so is a |
| 306 | // directory that is not there. Any other failure to read it is an error: |
| 307 | // the caller cannot tell. |
| 308 | func ReadSeedMarker(dbPath string) (string, error) { |
| 309 | dir, base, err := splitDBPath(dbPath) |
| 310 | if err != nil { |
| 311 | return "", err |
| 312 | } |
| 313 | if _, err := os.Lstat(dir); errors.Is(err, os.ErrNotExist) { |
| 314 | return "", nil |
| 315 | } |
| 316 | m, err := instancefs.ReadBeneathLimit(dir, base+SeedMarkerSuffix, seedMarkerMax) |
| 317 | if err != nil { |
| 318 | if notAPlacedFile(err) { |
| 319 | return "", nil |
| 320 | } |
| 321 | return "", err |
| 322 | } |
| 323 | return strings.TrimSpace(string(m)), nil |
| 324 | } |
| 325 | |
| 326 | // splitDBPath is a clean absolute database path's directory and file name. |
| 327 | func splitDBPath(dbPath string) (dir, base string, err error) { |
| 328 | if !filepath.IsAbs(dbPath) || filepath.Clean(dbPath) != dbPath { |
| 329 | return "", "", fmt.Errorf("activator: %q is not a clean absolute database path", dbPath) |
| 330 | } |
| 331 | dir, base = filepath.Split(dbPath) |
| 332 | if base == "" { |
| 333 | return "", "", fmt.Errorf("activator: %q names no database", dbPath) |
| 334 | } |
| 335 | return filepath.Clean(dir), base, nil |
| 336 | } |
| 337 | |
| 338 | // purgeSeedTarget removes everything at and beside dir/base that could |
| 339 | // stand in for a seed's database: what PurgeLocalDBArtifacts removes in its |
| 340 | // full form, the marker, and any seed temp file, through instancefs. |
| 341 | func purgeSeedTarget(dir, base string) error { |
| 342 | names := []string{base, base + "-wal", base + "-shm", base + SeedMarkerSuffix, |
| 343 | base + "-litestream", ".activator-" + base + "-litestream"} |
| 344 | ents, err := os.ReadDir(dir) |
| 345 | if err != nil { |
| 346 | return err |
| 347 | } |
| 348 | for _, e := range ents { |
| 349 | if n := e.Name(); strings.HasPrefix(n, ".seed-sd_") && strings.Contains(n, ".tmp") { |
| 350 | names = append(names, n) |
| 351 | } |
| 352 | } |
| 353 | for _, n := range names { |
| 354 | if err := instancefs.RemoveAllBeneath(dir, n); err != nil { |
| 355 | return err |
| 356 | } |
| 357 | } |
| 358 | return nil |
| 359 | } |
| 360 | |
| 361 | // placeSeed fetches and checks the snapshot, then writes the marker and |
| 362 | // renames the snapshot onto dir/base. The marker goes first, so a database |
| 363 | // at that path always has one. The temp file and any sidecar the check left |
| 364 | // are removed on every path. |
| 365 | // |
| 366 | // The instance directory is the tenant's uid's to write, so the file is |
| 367 | // pinned by identity: the one fetchSeed wrote is reopened beneath dir and |
| 368 | // must be the same file, the integrity check reads it through that |
| 369 | // descriptor, it must still be the entry at tmp after the check, and the |
| 370 | // entry renamed onto dir/base must still be it. Anything else is refused, |
| 371 | // and a refusal after the rename removes the marker and the database, so no |
| 372 | // unchecked file is ever kept as this seed. |
| 373 | func (b *ExecBackend) placeSeed(ctx context.Context, host string, src *seeds.RowSeed, dir, base string) error { |
| 374 | tmp := seedTempName(src.ID) |
| 375 | defer func() { |
| 376 | for _, n := range []string{tmp, tmp + "-wal", tmp + "-shm"} { |
| 377 | _ = instancefs.RemoveAllBeneath(dir, n) |
| 378 | } |
| 379 | }() |
| 380 | wrote, err := b.fetchSeed(ctx, host, src, dir, tmp) |
| 381 | if err != nil { |
| 382 | return err |
| 383 | } |
| 384 | f, err := openSameFile(dir, tmp, wrote) |
| 385 | if err != nil { |
| 386 | return seedPlaceFailed(host, err) |
| 387 | } |
| 388 | cerr := checkSeedDB(ctx, "/proc/self/fd/"+strconv.FormatUint(uint64(f.Fd()), 10)) |
| 389 | f.Close() |
| 390 | if cerr != nil { |
| 391 | return seedNotDB(host, cerr) |
| 392 | } |
| 393 | if err := sameFileAt(dir, tmp, wrote); err != nil { |
| 394 | return seedPlaceFailed(host, err) |
| 395 | } |
| 396 | if err := instancefs.WriteAtomic(dir, base+SeedMarkerSuffix, []byte(src.ID+"\n"), 0o600); err != nil { |
| 397 | return seedPlaceFailed(host, err) |
| 398 | } |
| 399 | if err := instancefs.RenameBetween(dir, tmp, dir, base); err != nil { |
| 400 | return seedPlaceFailed(host, err) |
| 401 | } |
| 402 | if err := sameFileAt(dir, base, wrote); err != nil { |
| 403 | _ = instancefs.RemoveAllBeneath(dir, base+SeedMarkerSuffix) |
| 404 | _ = instancefs.RemoveAllBeneath(dir, base) |
| 405 | return seedPlaceFailed(host, err) |
| 406 | } |
| 407 | return nil |
| 408 | } |
| 409 | |
| 410 | // errSwapped is a seed file that is no longer the one fetchSeed wrote. |
| 411 | var errSwapped = errors.New("the fetched snapshot was replaced in the instance directory") |
| 412 | |
| 413 | // openSameFile opens dir/name beneath dir, read-only, and refuses it unless |
| 414 | // it is the file want describes. |
| 415 | func openSameFile(dir, name string, want os.FileInfo) (*os.File, error) { |
| 416 | f, err := instancefs.OpenBeneath(dir, name, os.O_RDONLY, 0) |
| 417 | if err != nil { |
| 418 | return nil, err |
| 419 | } |
| 420 | fi, err := f.Stat() |
| 421 | if err != nil { |
| 422 | f.Close() |
| 423 | return nil, err |
| 424 | } |
| 425 | if !os.SameFile(fi, want) { |
| 426 | f.Close() |
| 427 | return nil, errSwapped |
| 428 | } |
| 429 | return f, nil |
| 430 | } |
| 431 | |
| 432 | // sameFileAt reports errSwapped unless dir/name is the file want describes. |
| 433 | func sameFileAt(dir, name string, want os.FileInfo) error { |
| 434 | f, err := openSameFile(dir, name, want) |
| 435 | if err != nil { |
| 436 | return err |
| 437 | } |
| 438 | return f.Close() |
| 439 | } |
| 440 | |
| 441 | // fetchSeed copies the snapshot into dir/tmp, checking its length and |
| 442 | // sha256 against the row as it writes, and returns the identity of the file |
| 443 | // it wrote. It reads at most one byte more than the row records, so an |
| 444 | // oversized object is refused without being read. |
| 445 | func (b *ExecBackend) fetchSeed(ctx context.Context, host string, src *seeds.RowSeed, dir, tmp string) (os.FileInfo, error) { |
| 446 | rc, err := b.Seeds.Open(ctx, host, src.ID) |
| 447 | if err != nil { |
| 448 | return nil, seedUnread(host, err) |
| 449 | } |
| 450 | defer rc.Close() |
| 451 | f, err := instancefs.OpenBeneath(dir, tmp, os.O_CREATE|os.O_EXCL|os.O_WRONLY, 0o600) |
| 452 | if err != nil { |
| 453 | return nil, seedPlaceFailed(host, err) |
| 454 | } |
| 455 | h := sha256.New() |
| 456 | r := &readErrs{r: io.LimitReader(rc, src.Bytes+1)} |
| 457 | n, cerr := io.Copy(io.MultiWriter(f, h), r) |
| 458 | serr := f.Sync() |
| 459 | fi, iderr := f.Stat() |
| 460 | ferr := f.Close() |
| 461 | if r.err != nil { |
| 462 | return nil, seedUnread(host, r.err) |
| 463 | } |
| 464 | if cerr != nil || serr != nil || iderr != nil || ferr != nil { |
| 465 | return nil, seedPlaceFailed(host, errors.Join(cerr, serr, iderr, ferr)) |
| 466 | } |
| 467 | if n != src.Bytes || hex.EncodeToString(h.Sum(nil)) != src.SHA256 { |
| 468 | return nil, seedMismatch(host) |
| 469 | } |
| 470 | return fi, nil |
| 471 | } |
| 472 | |
| 473 | // readErrs keeps the read side's error apart from the write side's, which |
| 474 | // io.Copy returns alike: a bucket that fails is SD-N16, a disk that fails is |
| 475 | // SD-N19. |
| 476 | type readErrs struct { |
| 477 | r io.Reader |
| 478 | err error |
| 479 | } |
| 480 | |
| 481 | func (e *readErrs) Read(p []byte) (int, error) { |
| 482 | n, err := e.r.Read(p) |
| 483 | if err != nil && err != io.EOF { |
| 484 | e.err = err |
| 485 | } |
| 486 | return n, err |
| 487 | } |
| 488 | |
| 489 | // SeedPlaced records a seed placed once the host's own replica holds it |
| 490 | // (spec §3.4 steps 4 and 5): it waits for the first upload, then |
| 491 | // OnSeedPlaced CASes the record to placed and deletes seeds/<host>/<id>/. |
| 492 | // The snapshot goes only after the replica exists, so a crash at any point |
| 493 | // retries from the placed local file or the snapshot. A replicator that |
| 494 | // cannot be asked for its first upload records nothing. |
| 495 | func (b *ExecBackend) SeedPlaced(ctx context.Context, rt registry.Route, rep Rep) error { |
| 496 | src, err := seeds.ParseRowSeed(rt.SeedSource) |
| 497 | if err != nil || src == nil || !src.Placeable() || src.State == seeds.StatePlaced { |
| 498 | return err |
| 499 | } |
| 500 | s, ok := rep.(interface{ SyncAndWait(context.Context) error }) |
| 501 | if !ok { |
| 502 | return fmt.Errorf("first upload of %s: the replicator cannot confirm it", rt.Host) |
| 503 | } |
| 504 | if err := s.SyncAndWait(ctx); err != nil { |
| 505 | return fmt.Errorf("first upload of %s: %w", rt.Host, err) |
| 506 | } |
| 507 | if b.OnSeedPlaced == nil { |
| 508 | return errors.New("no seed recorder on this box") |
| 509 | } |
| 510 | return b.OnSeedPlaced(ctx, rt.Account, rt.App, rt.Host, src.ID) |
| 511 | } |
| 512 | |
| 513 | // finishPlacement is SeedPlaced for a host whose own replica already |
| 514 | // exists (a placed CAS lost on an earlier wake). Best effort: a failure is |
| 515 | // logged, never the wake's, and the next wake asks again. |
| 516 | func (b *ExecBackend) finishPlacement(ctx context.Context, rt registry.Route, src *seeds.RowSeed) { |
| 517 | if b.OnSeedPlaced == nil { |
| 518 | return |
| 519 | } |
| 520 | if err := b.OnSeedPlaced(ctx, rt.Account, rt.App, rt.Host, src.ID); err != nil { |
| 521 | log.Printf("activator: %s: seed %s not recorded as placed yet: %v", rt.Host, src.ID, err) |
| 522 | } |
| 523 | } |
| 524 | |
amadan