| 1 | package edge |
| 2 | |
| 3 | // The mesh key — housekeeper-minted, edge-mirrored, HMAC hop vocabulary |
| 4 | // (spec docs/superpowers/specs/2026-08-22-multi-region-backhaul-design.md, |
| 5 | // Part 3, Task 2). |
| 6 | // |
| 7 | // A multi-region deployment has more than one edge; Tasks 3/4 have one edge |
| 8 | // forward a request to another (a "hop") when it does not own the region a |
| 9 | // request landed in. That hop has to prove it came from a box inside this |
| 10 | // deployment's own mesh — the writer map's answer is the same shape the |
| 11 | // housekeeper's other duties already use: exactly ONE box mints a shared |
| 12 | // secret, every box mirrors it, and no second election is invented. |
| 13 | // |
| 14 | // - mintMeshKey is the housekeeper duty (wired as its own loop, |
| 15 | // internal/host/housekeeper.go's mintMeshKeyLoop — see that func's |
| 16 | // comment for why it is NOT folded into retentionLoop's tick): 32 |
| 17 | // random bytes, hex-encoded, written to control/edge-mesh-key with an |
| 18 | // empty ifMatch — release.Store's create-only form. That IS the |
| 19 | // compare-and-swap: two housekeepers racing the same tick (the |
| 20 | // transient double-holder window housekeeper.go's poll cadence allows) |
| 21 | // both attempt the Put, exactly one wins, and the loser's |
| 22 | // release.ErrCASConflict is not a failure — it is proof a key already |
| 23 | // exists, which is the only thing this func promises. Every later tick |
| 24 | // calls it again and finds the same conflict, forever, for the |
| 25 | // deployment's life: minting is a one-time event, not a duty that ever |
| 26 | // "finishes" — and it stays SILENT after the first tick (see |
| 27 | // mintMeshKey's own comment): the steady state of every tick but one, |
| 28 | // forever, is not something worth a log line, exactly like the |
| 29 | // housekeeper election's own contention. |
| 30 | // |
| 31 | // - signHop / verifyHop are the hop vocabulary Task 3/4 wire onto the |
| 32 | // request path under the headers X-Carlos-Hop: 1 and X-Carlos-Hop-Auth: |
| 33 | // <unix>:<hex>. The MAC binds the target host into the signature so one |
| 34 | // hop's header cannot be replayed against another, and the timestamp |
| 35 | // bounds replay in time (±hopSkew) independent of the host check. The |
| 36 | // host is canonicalized (canonicalizeHopHost: lowercased, trailing dot |
| 37 | // stripped, :port stripped) before it ever reaches the MAC, and the MAC |
| 38 | // is always over the CONCRETE request host a hop is bound for — never a |
| 39 | // wildcard map key (e.g. "*.example.com") — because Task 3 verifies |
| 40 | // against the literal net/http request host and Task 4 signs the same; |
| 41 | // canonicalizing in one place (hopMAC) is what keeps a signer's and a |
| 42 | // verifier's idea of "host" from ever diverging on case, a trailing |
| 43 | // root-label dot, or a stray port. |
| 44 | // |
| 45 | // - Rotation's READ side ships in this task, ahead of any code that |
| 46 | // writes .next: control/edge-mesh-key.next mirrors a rotation |
| 47 | // successor while one is in flight, loadMeshKeys reads both, |
| 48 | // verifyHopEitherKey accepts a signature under either, and signingKey |
| 49 | // is what a signer consults to prefer .next the instant it exists — |
| 50 | // never waiting for a runbook step. |
| 51 | // |
| 52 | // What this removes is the PROMOTION window, not the PROPAGATION one. |
| 53 | // A signer starts using .next as soon as its OWN box has picked it up; |
| 54 | // a remote verifier accepts a .next-signed hop only once ITS OWN |
| 55 | // mirror has too (Task 3/4's poll of the control/ prefix, up to ~30s |
| 56 | // behind a write). Between those two moments a hop signed under .next |
| 57 | // CAN be refused by a verifier that hasn't caught up — that is a real, |
| 58 | // if short, window, and per Task 3's rules the cost of that refusal is |
| 59 | // the request being served as an ordinary client (its X-Forwarded-For |
| 60 | // stripped), not dropped: a degraded hop, not an outage. What |
| 61 | // dual-accept actually buys is that the runbook's promotion step |
| 62 | // itself (copy .next over control/edge-mesh-key, delete .next) needs |
| 63 | // no synchronized cutover: because every verifier already accepts |
| 64 | // EITHER key well before promotion happens, that write can land as an |
| 65 | // ordinary, un-sequenced action with no box needing to be told "stop |
| 66 | // signing with the old key" first. |
| 67 | // |
| 68 | // Single-box inertness: on a deployment with one platform box, this box is |
| 69 | // always the housekeeper (housekeeper.go's housekeeperSole), so it mints its |
| 70 | // own mesh key and nothing else ever reads it — Tasks 3/4 are the first |
| 71 | // readers. Minting here is therefore inert today, by design: this task ships |
| 72 | // the vocabulary and the mint, not a consumer. |
| 73 | import ( |
| 74 | "context" |
| 75 | "crypto/hmac" |
| 76 | "crypto/rand" |
| 77 | "crypto/sha256" |
| 78 | "encoding/hex" |
| 79 | "errors" |
| 80 | "fmt" |
| 81 | "log" |
| 82 | "net" |
| 83 | "strconv" |
| 84 | "strings" |
| 85 | "time" |
| 86 | |
| 87 | "github.com/carlosframework/platform/internal/edgeproof" |
| 88 | "github.com/carlosframework/platform/internal/release" |
| 89 | ) |
| 90 | |
| 91 | // meshKeyKey is the control-plane object the housekeeper mints once per |
| 92 | // deployment. Named beside the storedoor ACL tripwire |
| 93 | // (TestDoorNeverServesCertMaterial), which refuses to serve it — a fleet |
| 94 | // box authenticating through the store door must never read this key. |
| 95 | const meshKeyKey = edgeproof.MeshKeyObject |
| 96 | |
| 97 | // meshKeyNextKey is the rotation successor, mirrored alongside meshKeyKey |
| 98 | // only while a rotation is in flight. Promoting it (copy over meshKeyKey, |
| 99 | // delete this object) is a runbook step, not code this task ships — see the |
| 100 | // file header. |
| 101 | const meshKeyNextKey = edgeproof.NextMeshKeyObject |
| 102 | |
| 103 | // meshKeyBytes is the random key length before hex encoding: 256 bits, |
| 104 | // matched to the HMAC-SHA256 hop signature below. |
| 105 | const meshKeyBytes = 32 |
| 106 | |
| 107 | // hopSkew bounds how far a hop signature's timestamp may drift from the |
| 108 | // verifier's clock, in either direction: generous enough for ordinary box |
| 109 | // clock skew and request latency, tight enough that a captured header stops |
| 110 | // working well within a minute. |
| 111 | const hopSkew = 30 * time.Second |
| 112 | |
| 113 | // mintMeshKey mints the deployment's mesh key if, and only if, nothing has |
| 114 | // minted one yet. crypto/rand is required — this key authenticates hops |
| 115 | // between platform boxes, so it gets the same random source every other |
| 116 | // credential in this codebase does, never math/rand. |
| 117 | // |
| 118 | // A lost race (release.ErrCASConflict — a key already exists, from this |
| 119 | // box's own earlier tick or another housekeeper that won the same race) is |
| 120 | // deliberately SILENT, not logged even at "already minted": that outcome is |
| 121 | // the steady state of every tick but the very first, forever, for the |
| 122 | // deployment's whole life — one line every tick would be ~264 lines/day of |
| 123 | // nothing happening, exactly the noise housekeeper.go's own election |
| 124 | // contention refuses to log for the same reason. Only an actual, novel mint |
| 125 | // gets a line. |
| 126 | func mintMeshKey(ctx context.Context, s release.Store) error { |
| 127 | buf := make([]byte, meshKeyBytes) |
| 128 | if _, err := rand.Read(buf); err != nil { |
| 129 | return fmt.Errorf("mesh key: read random bytes: %w", err) |
| 130 | } |
| 131 | body := []byte(hex.EncodeToString(buf)) |
| 132 | if _, err := s.Put(ctx, meshKeyKey, body, ""); err != nil { |
| 133 | if errors.Is(err, release.ErrCASConflict) { |
| 134 | return nil |
| 135 | } |
| 136 | return fmt.Errorf("mesh key: put: %w", err) |
| 137 | } |
| 138 | log.Printf("edge: mesh key: minted") |
| 139 | return nil |
| 140 | } |
| 141 | |
| 142 | // MintMeshKey is mintMeshKey's exported wrapper — the seam |
| 143 | // internal/host's housekeeper duty (mintMeshKeyLoop, housekeeper.go) calls |
| 144 | // through, so the duty lives beside the housekeeper while the key's own |
| 145 | // shape (its object name, its random-byte discipline) stays edge's to own. |
| 146 | func MintMeshKey(ctx context.Context, s release.Store) error { |
| 147 | return mintMeshKey(ctx, s) |
| 148 | } |
| 149 | |
| 150 | // loadMeshKeys is the dual-accept read side's key material access: it reads |
| 151 | // BOTH control/edge-mesh-key and control/edge-mesh-key.next. current is nil |
| 152 | // only before the housekeeper has minted anything; next is nil the whole |
| 153 | // rest of the time — nearly always, since a rotation is a deliberate, |
| 154 | // short-lived operator action. |
| 155 | // |
| 156 | // The read itself lives in internal/edgeproof, because the console reads the |
| 157 | // same pair to check the edge's forwarded-for proof and must not import this |
| 158 | // package. Its failure rules are the ones this read always had: a bad .next |
| 159 | // is logged and treated as no rotation in flight (it is written by hand, and |
| 160 | // one bad upload must not take every hop down with the good current key); a |
| 161 | // bad current key fails the call, since there is no safe fallback for the |
| 162 | // live key being unreadable. |
| 163 | func loadMeshKeys(ctx context.Context, s release.Store) (current, next []byte, err error) { |
| 164 | return edgeproof.LoadKeys(ctx, s) |
| 165 | } |
| 166 | |
| 167 | // signingKey is which of the two mirrored keys a signer should use: .next |
| 168 | // the instant it exists, else current. A signer never waits for the |
| 169 | // runbook's promotion step — verifyHopEitherKey's dual-accept on the reader |
| 170 | // side is what makes that safe, so switching signers over early costs |
| 171 | // nothing and shortens how long a rotation needs both keys mirrored. |
| 172 | func signingKey(current, next []byte) []byte { |
| 173 | if len(next) > 0 { |
| 174 | return next |
| 175 | } |
| 176 | return current |
| 177 | } |
| 178 | |
| 179 | // signHop produces the value of the X-Carlos-Hop-Auth header: an |
| 180 | // HMAC-SHA256 over "host\nunix-timestamp", hex-encoded and prefixed with the |
| 181 | // timestamp so a verifier does not have to guess it. host is bound into the |
| 182 | // MAC so a signature minted for one hop cannot be replayed against another. |
| 183 | // host must be the CONCRETE host a hop is bound for (a net/http request |
| 184 | // host, e.g. from Task 4's forwarding decision) — never a wildcard home-map |
| 185 | // key like "*.example.com" — because hopMAC canonicalizes it the same way a |
| 186 | // verifier's literal request host is canonicalized, and a wildcard pattern |
| 187 | // is not a request host under any canonicalization. |
| 188 | func signHop(key []byte, host string, now time.Time) string { |
| 189 | ts := strconv.FormatInt(now.Unix(), 10) |
| 190 | return ts + ":" + hex.EncodeToString(hopMAC(key, host, ts)) |
| 191 | } |
| 192 | |
| 193 | // verifyHop checks an X-Carlos-Hop-Auth header against a single key: the |
| 194 | // timestamp must parse and fall within ±hopSkew of now, and the MAC must |
| 195 | // match under constant-time comparison (hmac.Equal) — a hop signature is |
| 196 | // exactly the kind of secret-derived value a timing side channel could leak, |
| 197 | // so no byte-wise or string-equality shortcut is taken anywhere in this |
| 198 | // path. |
| 199 | func verifyHop(key []byte, host, header string, now time.Time) bool { |
| 200 | ts, sig, ok := strings.Cut(header, ":") |
| 201 | if !ok || ts == "" || sig == "" { |
| 202 | return false |
| 203 | } |
| 204 | sec, err := strconv.ParseInt(ts, 10, 64) |
| 205 | if err != nil { |
| 206 | return false |
| 207 | } |
| 208 | skew := now.Sub(time.Unix(sec, 0)) |
| 209 | if skew < 0 { |
| 210 | skew = -skew |
| 211 | } |
| 212 | if skew > hopSkew { |
| 213 | return false |
| 214 | } |
| 215 | got, err := hex.DecodeString(sig) |
| 216 | if err != nil { |
| 217 | return false |
| 218 | } |
| 219 | return hmac.Equal(got, hopMAC(key, host, ts)) |
| 220 | } |
| 221 | |
| 222 | // verifyHopEitherKey is the dual-accept read side of a rotation: it accepts |
| 223 | // a signature made under either mirrored key, current or next, so neither a |
| 224 | // straggling signer still on the old key nor one that has already switched |
| 225 | // to .next (signingKey prefers it immediately) is ever refused while both |
| 226 | // are mirrored. Task 3/4's request path calls this once it has loaded both |
| 227 | // keys via loadMeshKeys. |
| 228 | func verifyHopEitherKey(current, next []byte, host, header string, now time.Time) bool { |
| 229 | if len(next) > 0 && verifyHop(next, host, header, now) { |
| 230 | return true |
| 231 | } |
| 232 | if len(current) > 0 && verifyHop(current, host, header, now) { |
| 233 | return true |
| 234 | } |
| 235 | return false |
| 236 | } |
| 237 | |
| 238 | // hopMAC is the MAC signHop and verifyHop both compute — one place so the |
| 239 | // two can never drift out of sync on what bytes are actually signed, and |
| 240 | // the one choke point that canonicalizes host before it enters the MAC. |
| 241 | func hopMAC(key []byte, host, ts string) []byte { |
| 242 | mac := hmac.New(sha256.New, key) |
| 243 | mac.Write([]byte(canonicalizeHopHost(host))) |
| 244 | mac.Write([]byte("\n")) |
| 245 | mac.Write([]byte(ts)) |
| 246 | return mac.Sum(nil) |
| 247 | } |
| 248 | |
| 249 | // canonicalizeHopHost normalizes a request host into the exact form the MAC |
| 250 | // is computed over, so a signer and a verifier can never disagree on what |
| 251 | // "host" means for the same request: |
| 252 | // |
| 253 | // - a :port is stripped via net.SplitHostPort when present — Task 3 |
| 254 | // verifies against a literal net/http request Host, which carries a |
| 255 | // port when the client set one (or the request came in over a |
| 256 | // non-default port), while a signer's own idea of the host (e.g. a |
| 257 | // home-map lookup key) never does; |
| 258 | // - lowercased — HTTP host headers are case-insensitive, but bytes |
| 259 | // entering an HMAC are not, so a client or intermediary that |
| 260 | // capitalizes a label must not be able to produce a request this MAC |
| 261 | // treats as a different host; |
| 262 | // - a single trailing root-label dot is stripped — "example.com." and |
| 263 | // "example.com" name the same host, and a client is free to send |
| 264 | // either. |
| 265 | // |
| 266 | // host is returned unchanged (net.SplitHostPort's error path) when it |
| 267 | // carries no port — the common case, and not a failure. |
| 268 | func canonicalizeHopHost(host string) string { |
| 269 | if h, _, err := net.SplitHostPort(host); err == nil { |
| 270 | host = h |
| 271 | } |
| 272 | return strings.TrimSuffix(strings.ToLower(host), ".") |
| 273 | } |
| 274 | |
amadan