| 1 | package instancectl |
| 2 | |
| 3 | import ( |
| 4 | "bufio" |
| 5 | "bytes" |
| 6 | "context" |
| 7 | "errors" |
| 8 | "fmt" |
| 9 | "io/fs" |
| 10 | "net/http" |
| 11 | "net/netip" |
| 12 | "os" |
| 13 | "path/filepath" |
| 14 | "sort" |
| 15 | "strings" |
| 16 | "time" |
| 17 | |
| 18 | "github.com/carlosframework/platform/internal/provision" |
| 19 | ) |
| 20 | |
| 21 | // What the control service does beyond start and stop, so the isolation |
| 22 | // the templates promise holds (#214, #178): it lays out each instance's |
| 23 | // view of its own build, re-opens the box's DNS resolver through the |
| 24 | // egress fence, and keeps tenants from filling the shared data disk. |
| 25 | |
| 26 | // Paths the instance units and the probe bind from (#214 item 2). The |
| 27 | // templates name them literally; the units test pins both sides. |
| 28 | const ( |
| 29 | // LiveDir is where the adopt pass keeps each host's live link, and |
| 30 | // where an instance finds its own build inside its namespace. |
| 31 | LiveDir = "/opt/carlos/live" |
| 32 | // SystemBinary is the platform binary every instance unit runs as |
| 33 | // ExecStart=; the probe's stand-in build. |
| 34 | SystemBinary = "/opt/carlos/carlos" |
| 35 | // InstancesSlice is the parent of every per-host slice and of the |
| 36 | // probe: an IPAddressAllow= on it reaches every instance unit. |
| 37 | InstancesSlice = "carlos-instances.slice" |
| 38 | ) |
| 39 | |
| 40 | // ensureExe writes the two sources an instance unit binds over its empty |
| 41 | // /opt/carlos (#214 item 2), under root/<iid>/: |
| 42 | // |
| 43 | // app -> target (the host's live link; the probe's: the platform binary) |
| 44 | // live/<name> -> ../app |
| 45 | // |
| 46 | // PID 1 chases app to the build when it binds it at /opt/carlos/app, and |
| 47 | // binds live/ at /opt/carlos/live, so inside the namespace |
| 48 | // /opt/carlos/live/<host> resolves to the build and nothing else of the |
| 49 | // version store exists. Both are links, never copies: a deploy repoints |
| 50 | // the host's live link and the next start binds the new build, exactly as |
| 51 | // before. Only root writes under root, so nothing here can be planted; |
| 52 | // anything else found in live/ is removed all the same, because live/ is |
| 53 | // bound into a tenant. |
| 54 | func ensureExe(root, iid, target, name string) error { |
| 55 | return ensureExeAs(root, iid, target, name, "app", "live") |
| 56 | } |
| 57 | |
| 58 | // ensureExeAs is ensureExe with the two sources named: app and live under |
| 59 | // root/<iid>/. The cmd template binds its own pair (CmdExeApp, CmdExeLive; |
| 60 | // merge ruling M4), so a command's start and an instance's or sidecar's |
| 61 | // start never write a source the other is about to bind. Inside the |
| 62 | // namespace both pairs are bound at the same places, so the link in live |
| 63 | // is ../app either way. |
| 64 | func ensureExeAs(root, iid, target, name, app, liveName string) error { |
| 65 | if root == "" { |
| 66 | return errors.New("no exe root") |
| 67 | } |
| 68 | if iid == "" || strings.ContainsAny(iid, "/\x00") || iid == "." || iid == ".." || |
| 69 | name == "" || strings.ContainsAny(name, "/\x00") || name == "." || name == ".." { |
| 70 | return fmt.Errorf("exe: %q/%q is not a name", iid, name) |
| 71 | } |
| 72 | dir := filepath.Join(root, iid) |
| 73 | live := filepath.Join(dir, liveName) |
| 74 | for _, d := range []string{root, dir, live} { |
| 75 | if err := os.Mkdir(d, 0o755); err != nil && !errors.Is(err, fs.ErrExist) { |
| 76 | return err |
| 77 | } |
| 78 | fi, err := os.Lstat(d) |
| 79 | if err != nil { |
| 80 | return err |
| 81 | } |
| 82 | if !fi.IsDir() { |
| 83 | return fmt.Errorf("exe: %s is not a directory", d) |
| 84 | } |
| 85 | } |
| 86 | if err := replaceLink(filepath.Join(dir, app), target); err != nil { |
| 87 | return err |
| 88 | } |
| 89 | if err := replaceLink(filepath.Join(live, name), "../app"); err != nil { |
| 90 | return err |
| 91 | } |
| 92 | entries, err := os.ReadDir(live) |
| 93 | if err != nil { |
| 94 | return err |
| 95 | } |
| 96 | for _, e := range entries { |
| 97 | if e.Name() != name { |
| 98 | if err := os.RemoveAll(filepath.Join(live, e.Name())); err != nil { |
| 99 | return err |
| 100 | } |
| 101 | } |
| 102 | } |
| 103 | return nil |
| 104 | } |
| 105 | |
| 106 | // replaceLink points path at target, atomically: a temp link renamed over |
| 107 | // whatever was there, so a unit starting at that moment binds the old |
| 108 | // target or the new one, never nothing. |
| 109 | func replaceLink(path, target string) error { |
| 110 | if cur, err := os.Readlink(path); err == nil && cur == target { |
| 111 | return nil |
| 112 | } |
| 113 | tmp := path + ".tmp" |
| 114 | _ = os.Remove(tmp) |
| 115 | if err := os.Symlink(target, tmp); err != nil { |
| 116 | return err |
| 117 | } |
| 118 | if err := os.Rename(tmp, path); err != nil { |
| 119 | os.Remove(tmp) |
| 120 | return err |
| 121 | } |
| 122 | return nil |
| 123 | } |
| 124 | |
| 125 | // ensureHostExe is ensureExe for a host about to start. |
| 126 | func (s *Server) ensureHostExe(host string) error { |
| 127 | return ensureExe(s.ExeRoot, provision.InstanceID(host), filepath.Join(LiveDir, host), host) |
| 128 | } |
| 129 | |
| 130 | // ParseResolvers returns the nameserver addresses of a resolv.conf, in |
| 131 | // order, zones dropped. Lines it cannot read are skipped: the resolver |
| 132 | // library skips them too. |
| 133 | func ParseResolvers(b []byte) []netip.Addr { |
| 134 | var out []netip.Addr |
| 135 | sc := bufio.NewScanner(bytes.NewReader(b)) |
| 136 | for sc.Scan() { |
| 137 | f := strings.Fields(sc.Text()) |
| 138 | if len(f) < 2 || f[0] != "nameserver" { |
| 139 | continue |
| 140 | } |
| 141 | a, err := netip.ParseAddr(f[1]) |
| 142 | if err != nil { |
| 143 | continue |
| 144 | } |
| 145 | out = append(out, a.WithZone("").Unmap()) |
| 146 | } |
| 147 | return out |
| 148 | } |
| 149 | |
| 150 | // ResolverAllowed says whether a nameserver may be re-opened through the |
| 151 | // instance egress fence (#214 item 3), and if not, why. IPAddressAllow= |
| 152 | // opens an address, not a port, and in both directions, so an allowed |
| 153 | // address must be one where nothing but a resolver can answer: |
| 154 | // |
| 155 | // - no loopback address, systemd-resolved's stubs (127.0.0.53/54) |
| 156 | // included: a tenant can bind the address itself and then reach every |
| 157 | // listener on the box bound to all addresses through it. Instances |
| 158 | // resolve through resolved's UPSTREAMS instead (the namespace binds |
| 159 | // /run/systemd/resolve/resolv.conf over /etc/resolv.conf); |
| 160 | // - no unspecified address, which means "this box"; |
| 161 | // - no metadata service; |
| 162 | // - no link-local address but Amazon's resolver, 169.254.169.253. An |
| 163 | // allow cannot carry a zone, so fe80::1 would open that address on |
| 164 | // every link, and link-local v4 is where IMDS and the time service |
| 165 | // live. Boxes resolve through the VPC resolver; a link-local one is |
| 166 | // refused and the probe's dns line says so. |
| 167 | // |
| 168 | // Everything else is opened as a single address: the VPC resolver (VPC |
| 169 | // base + 2, inside RFC 1918), fd00:ec2::253, or a public one, unless it |
| 170 | // is one of the box's own addresses, which only the box can say: |
| 171 | // resolverAllow refuses those against localPrefixes, here, where the |
| 172 | // allow is written (#126). |
| 173 | func ResolverAllowed(a netip.Addr) (bool, string) { |
| 174 | switch { |
| 175 | case !a.IsValid(): |
| 176 | return false, "not an address" |
| 177 | case a == netip.MustParseAddr("127.0.0.53"), a == netip.MustParseAddr("127.0.0.54"): |
| 178 | return false, "systemd-resolved's stub is on loopback; instances use its upstreams" |
| 179 | case a.IsLoopback(): |
| 180 | return false, "a loopback resolver shares its address with services on this box" |
| 181 | case a.IsUnspecified(): |
| 182 | return false, "the unspecified address is this box" |
| 183 | case a == netip.MustParseAddr("169.254.169.254"), a == netip.MustParseAddr("fd00:ec2::254"): |
| 184 | return false, "the metadata service is not a resolver" |
| 185 | case a == netip.MustParseAddr("169.254.169.253"): |
| 186 | return true, "" |
| 187 | case a.IsLinkLocalUnicast(): |
| 188 | return false, "a link-local resolver cannot be opened for one link only" |
| 189 | } |
| 190 | return true, "" |
| 191 | } |
| 192 | |
| 193 | // resolverAllow is the IPAddressAllow= value for a resolv.conf: each |
| 194 | // allowed nameserver as a single-address prefix, sorted and deduplicated |
| 195 | // so the drop-in systemd writes does not churn. A nameserver inside one |
| 196 | // of the box's local prefixes is left out, and named in refused. |
| 197 | func resolverAllow(b []byte, local []netip.Prefix) (allow, refused []string) { |
| 198 | seen := map[string]bool{} |
| 199 | for _, a := range ParseResolvers(b) { |
| 200 | if ok, _ := ResolverAllowed(a); !ok { |
| 201 | continue |
| 202 | } |
| 203 | p := netip.PrefixFrom(a, a.BitLen()).String() |
| 204 | if seen[p] { |
| 205 | continue |
| 206 | } |
| 207 | seen[p] = true |
| 208 | if lp, ok := isLocal(a, local); ok { |
| 209 | refused = append(refused, fmt.Sprintf("%s is an address of this box (%s)", a, lp)) |
| 210 | continue |
| 211 | } |
| 212 | allow = append(allow, p) |
| 213 | } |
| 214 | sort.Strings(allow) |
| 215 | return allow, refused |
| 216 | } |
| 217 | |
| 218 | // ResolvedConf is systemd-resolved's list of its upstreams. When it |
| 219 | // exists, the namespace binds it over /etc/resolv.conf (the "-" bind in |
| 220 | // the templates), so it is what instances resolve through and what the |
| 221 | // allow must open. |
| 222 | const ResolvedConf = "/run/systemd/resolve/resolv.conf" |
| 223 | |
| 224 | // resolverFile is the resolv.conf an instance sees: resolved's upstream |
| 225 | // list when there is one, the box's own otherwise. Mirrors the template. |
| 226 | func (s *Server) resolverFile() string { |
| 227 | if s.ResolvedConf != "" { |
| 228 | if _, err := os.Stat(s.ResolvedConf); err == nil { |
| 229 | return s.ResolvedConf |
| 230 | } |
| 231 | } |
| 232 | return s.ResolvConf |
| 233 | } |
| 234 | |
| 235 | // allowResolvers re-opens the resolvers instances see on |
| 236 | // carlos-instances.slice. The empty assignment first resets the list, so |
| 237 | // a resolver that left the file is closed again. It runs before every |
| 238 | // probe, so the allow is in place before the first start and follows a |
| 239 | // resolver change within one probe period; the probe then checks the |
| 240 | // result from inside. It returns the nameservers it refused for being |
| 241 | // the box's own. When the box's addresses cannot be read, every resolver |
| 242 | // is closed and the error returned: an allow that might name the box |
| 243 | // itself is never written. |
| 244 | func (s *Server) allowResolvers(ctx context.Context) (refused []string, err error) { |
| 245 | b, err := os.ReadFile(s.resolverFile()) |
| 246 | if err != nil { |
| 247 | return nil, err |
| 248 | } |
| 249 | var allow []string |
| 250 | local, lerr := s.localPrefixes() |
| 251 | if lerr == nil { |
| 252 | allow, refused = resolverAllow(b, local) |
| 253 | } |
| 254 | args := []string{"set-property", "--runtime", InstancesSlice, "IPAddressAllow="} |
| 255 | if len(allow) > 0 { |
| 256 | args = append(args, "IPAddressAllow="+strings.Join(allow, " ")) |
| 257 | } |
| 258 | out, err := s.Systemctl(ctx, args...) |
| 259 | if err != nil { |
| 260 | return nil, fmt.Errorf("%s: %s", strings.TrimSpace(out), err) |
| 261 | } |
| 262 | if lerr != nil { |
| 263 | return nil, fmt.Errorf("this box's addresses: %v; every resolver is closed", lerr) |
| 264 | } |
| 265 | return refused, nil |
| 266 | } |
| 267 | |
| 268 | // localPrefixes is LocalAddrs, refusing when there is none. |
| 269 | func (s *Server) localPrefixes() ([]netip.Prefix, error) { |
| 270 | if s.LocalAddrs == nil { |
| 271 | return nil, errors.New("no way to list them") |
| 272 | } |
| 273 | return s.LocalAddrs() |
| 274 | } |
| 275 | |
| 276 | // resolverLocalLine is the verdict on allowResolvers' refusals, folded |
| 277 | // into the probe's report: the probe cannot make this check from inside |
| 278 | // its unit, which cannot list the box's addresses (#126). |
| 279 | func resolverLocalLine(refused []string) VerdictLine { |
| 280 | if len(refused) > 0 { |
| 281 | return VerdictLine{Check: "resolver-local", Verdict: "fail", Detail: strings.Join(refused, "; ")} |
| 282 | } |
| 283 | return VerdictLine{Check: "resolver-local", Verdict: "ok"} |
| 284 | } |
| 285 | |
| 286 | // The data disk (#214 item 1). Every instance's directory, the registry, |
| 287 | // the activator's state and litestream's shadow WAL share one filesystem, |
| 288 | // /data, on the root volume. A per-tenant quota is not achievable there |
| 289 | // today: ext4 project quotas need the "project" and "quota" features, |
| 290 | // which tune2fs sets only on an unmounted filesystem, plus a prjquota |
| 291 | // mount, and the root filesystem is never unmounted on a running box. |
| 292 | // Every process that writes /data is uid carlos, so reserved blocks |
| 293 | // cannot tell a tenant from the agent either. |
| 294 | // |
| 295 | // Nor can this service attribute usage well enough to stop a tenant for |
| 296 | // it: space held by unlinked-but-open files appears in no directory, so |
| 297 | // the guilty tenant can look small while honest ones look large, and a |
| 298 | // stop from here bypasses the activator's fence and holds (a host mid- |
| 299 | // migration). So the guard refuses, it never stops: below the floor no |
| 300 | // instance starts or seeds, and a warning names the shortfall every |
| 301 | // DiskEvery while it lasts. Instances already awake keep running. |
| 302 | |
| 303 | const minStartFloor = 1 << 30 // 1 GiB |
| 304 | |
| 305 | // DiskFloor is the free space a start needs on a filesystem of total |
| 306 | // bytes: the larger of 1 GiB and 10%. |
| 307 | func DiskFloor(total uint64) uint64 { return max(minStartFloor, total/10) } |
| 308 | |
| 309 | // ErrDiskFull is a start refused for disk space. |
| 310 | var ErrDiskFull = errors.New("data disk too full") |
| 311 | |
| 312 | // diskRoom is nil when there is room to start an instance. |
| 313 | func (s *Server) diskRoom() error { |
| 314 | if s.Statfs == nil { |
| 315 | return nil |
| 316 | } |
| 317 | avail, total, err := s.Statfs(s.DataRoot) |
| 318 | if err != nil { |
| 319 | return fmt.Errorf("statfs %s: %w", s.DataRoot, err) |
| 320 | } |
| 321 | if floor := DiskFloor(total); avail < floor { |
| 322 | return fmt.Errorf("%w: %s free, below the %s floor", ErrDiskFull, humanBytes(avail), humanBytes(floor)) |
| 323 | } |
| 324 | return nil |
| 325 | } |
| 326 | |
| 327 | // refuseStart answers a start the disk or the exe layout cannot take. exe |
| 328 | // lays out the unit's view of its build (ensureHostExe, or ensureGenExe |
| 329 | // for a generation). It reports whether it answered. |
| 330 | func (s *Server) refuseStart(w http.ResponseWriter, host, unit string, exe func() error) bool { |
| 331 | if err := s.diskRoom(); err != nil { |
| 332 | s.logf("instance-control: %s: not starting %s: %v", host, unit, err) |
| 333 | status := http.StatusBadGateway |
| 334 | if errors.Is(err, ErrDiskFull) { |
| 335 | status = http.StatusInsufficientStorage |
| 336 | } |
| 337 | http.Error(w, "not starting "+unit+": "+err.Error(), status) |
| 338 | return true |
| 339 | } |
| 340 | if s.ExeRoot != "" && exe != nil { |
| 341 | if err := exe(); err != nil { |
| 342 | http.Error(w, "exe: "+err.Error(), http.StatusBadGateway) |
| 343 | return true |
| 344 | } |
| 345 | } |
| 346 | return false |
| 347 | } |
| 348 | |
| 349 | // warnIfFull logs the alert while the disk is below the floor. It reports |
| 350 | // whether it warned. |
| 351 | func (s *Server) warnIfFull() bool { |
| 352 | err := s.diskRoom() |
| 353 | if errors.Is(err, ErrDiskFull) { |
| 354 | s.logf("instance-control: disk: %v; refusing every instance start until there is room (awake instances keep running)", err) |
| 355 | return true |
| 356 | } |
| 357 | if err != nil { |
| 358 | s.logf("instance-control: disk: %v", err) |
| 359 | } |
| 360 | return false |
| 361 | } |
| 362 | |
| 363 | // diskLoop runs warnIfFull every DiskEvery. |
| 364 | func (s *Server) diskLoop(ctx context.Context) { |
| 365 | if s.DiskEvery <= 0 { |
| 366 | return |
| 367 | } |
| 368 | t := time.NewTicker(s.DiskEvery) |
| 369 | defer t.Stop() |
| 370 | for { |
| 371 | select { |
| 372 | case <-ctx.Done(): |
| 373 | return |
| 374 | case <-t.C: |
| 375 | s.warnIfFull() |
| 376 | } |
| 377 | } |
| 378 | } |
| 379 | |
| 380 | func humanBytes(n uint64) string { |
| 381 | switch { |
| 382 | case n >= 1<<30: |
| 383 | return fmt.Sprintf("%.1f GiB", float64(n)/(1<<30)) |
| 384 | case n >= 1<<20: |
| 385 | return fmt.Sprintf("%.0f MiB", float64(n)/(1<<20)) |
| 386 | } |
| 387 | return fmt.Sprintf("%d bytes", n) |
| 388 | } |
| 389 | |
amadan