| 1 | package activator |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "errors" |
| 6 | "fmt" |
| 7 | "time" |
| 8 | |
| 9 | "github.com/carlosframework/platform/internal/registry" |
| 10 | "github.com/carlosframework/platform/internal/release" |
| 11 | ) |
| 12 | |
| 13 | // The held wake check's refusals (settings spec §4.3 rule 3). None of them |
| 14 | // ran anything of a build, so none is a failure of one. Members read them |
| 15 | // through memberSafeWakeReason (ruling R-WAKEWORDS), never as this text. |
| 16 | var ( |
| 17 | errBuildUnknown = errors.New("build unknown: use Update now or set a pin") |
| 18 | errStaleLink = errors.New("the installed build is not this instance's build, and root has not installed it yet") |
| 19 | errTargetFailed = errors.New("this instance's build failed on this box, so it is not started again") |
| 20 | ) |
| 21 | |
| 22 | // targetFailedError is errTargetFailed naming the build, so the member's |
| 23 | // sentence can name it too. |
| 24 | type targetFailedError struct{ host, build string } |
| 25 | |
| 26 | func (e *targetFailedError) Error() string { |
| 27 | return fmt.Sprintf("%v: %s on %s", errTargetFailed, e.build, e.host) |
| 28 | } |
| 29 | |
| 30 | func (e *targetFailedError) Unwrap() error { return errTargetFailed } |
| 31 | |
| 32 | // HeldWake is the agent's answer for a held instance's wake (settings spec |
| 33 | // §4.3 rule 3). |
| 34 | type HeldWake struct { |
| 35 | Target string // the build the instance must run |
| 36 | Held bool // false: root follows the channel for it; today's wake |
| 37 | Unknown bool // rule 3's "build unknown" |
| 38 | Failed bool // root reports Target failed on this box (its decision, or a refusal record) |
| 39 | // AskRoot: root has made no decision for the instance's current |
| 40 | // settings (ruling R-HELD-PURE), so whether it is held is not known |
| 41 | // here. The wake holds: it runs Target only if the live link already |
| 42 | // names it, and otherwise asks root for a pass and waits for its |
| 43 | // decision. Target is the build state's build ("" when there is none). |
| 44 | AskRoot bool |
| 45 | // Unread: the build state could not be read for this wake (its required |
| 46 | // refresh failed). Refused: never decided from a stale copy (Astra round |
| 47 | // 1 finding 1). |
| 48 | Unread bool |
| 49 | // Staged is a retained copy of Target on this box whose bytes match its |
| 50 | // staged manifest; nil when there is none. |
| 51 | Staged *release.StagedIdentity |
| 52 | } |
| 53 | |
| 54 | // heldPoll is how often a held wake re-reads the live link while it waits |
| 55 | // for root to install its target. A var so tests can shrink it. |
| 56 | var heldPoll = 250 * time.Millisecond |
| 57 | |
| 58 | // wakeCheck is settings spec §4.3 rule 3's wake half, with its own admission |
| 59 | // rule so it holds with CARLOS_SWAP off and no planner (Astra I3 #2, #5): |
| 60 | // |
| 61 | // - not held: "" and nothing to do, today's wake; |
| 62 | // - unread (the build state could not be read for this wake): |
| 63 | // errBuildStateUnwritten, nothing asked or run (Astra round 1 finding 1); |
| 64 | // - root has not decided for the current settings (AskRoot, ruling |
| 65 | // R-HELD-PURE): the build state's build runs if the live link already |
| 66 | // names it; otherwise root is asked for a pass (RequestStage naming no |
| 67 | // build) and the wake waits for its decision, within the same bound, |
| 68 | // then goes on below with that answer. It never runs a live link it |
| 69 | // does not know it may run; |
| 70 | // - unknown: errBuildUnknown, and root is not asked (it installs nothing); |
| 71 | // - a target root reports failed: refused (errTargetFailed) with no |
| 72 | // planner, even while it is still the live link; with a planner the |
| 73 | // wake is not held, and zdd-5's failure-aware plan chooses (recovery |
| 74 | // onto the good build, or a refusal); |
| 75 | // - the target installed: the target, and the caller checks the identity |
| 76 | // it captures against it (heldIdentityOK); |
| 77 | // - with no planner and direct set (the caller can exec a named file), a |
| 78 | // retained verified copy of the target, which the caller runs as its |
| 79 | // identity, recorded first; root is asked to install the target too, |
| 80 | // without waiting (R-RETAINED-NUDGE); |
| 81 | // - otherwise root is asked to install the target (RequestStage, the |
| 82 | // stage relay) and the wake waits for the live link to name it, within |
| 83 | // its own deadline and never past wakeWaitBound, the bound its waiters |
| 84 | // have. A wait that ends first is errStaleLink. The answer is asked |
| 85 | // again every heldReask polls, and a changed answer ends the wait at |
| 86 | // once with its own refusal (unknown, failed) or errStaleLink. |
| 87 | // |
| 88 | // It never sets e.pin (Astra I2 #4): with a planner, what boots after root's |
| 89 | // install is the plan's. The string is the target ("" when the wake is not |
| 90 | // held). |
| 91 | func (a *Service) wakeCheck(ctx context.Context, host string, direct bool) (string, *release.StagedIdentity, error) { |
| 92 | if a.HeldTarget == nil { |
| 93 | return "", nil, nil |
| 94 | } |
| 95 | h := a.HeldTarget(host) |
| 96 | if !h.Held { |
| 97 | return "", nil, nil |
| 98 | } |
| 99 | if h.Unread { |
| 100 | return "", nil, fmt.Errorf("%w: the build state of %s could not be read for its wake", errBuildStateUnwritten, host) |
| 101 | } |
| 102 | wait, cancel := context.WithTimeout(ctx, wakeWaitBound) |
| 103 | defer cancel() |
| 104 | if h.AskRoot { |
| 105 | var err error |
| 106 | if h, err = a.awaitRootDecision(wait, host, h); err != nil { |
| 107 | return "", nil, err |
| 108 | } |
| 109 | switch { |
| 110 | case !h.Held: |
| 111 | return "", nil, nil |
| 112 | case h.Unread: |
| 113 | return "", nil, fmt.Errorf("%w: the build state of %s could not be read for its wake", errBuildStateUnwritten, host) |
| 114 | } |
| 115 | } |
| 116 | if h.Unknown || h.Target == "" { |
| 117 | return "", nil, errBuildUnknown |
| 118 | } |
| 119 | if h.Failed { |
| 120 | if a.Planner != nil { |
| 121 | return "", nil, nil // zdd-5's decideWake: recovery onto the good build, or a refusal |
| 122 | } |
| 123 | return "", nil, &targetFailedError{host: host, build: h.Target} |
| 124 | } |
| 125 | installed := a.liveVersion(host) |
| 126 | if installed == h.Target { |
| 127 | return h.Target, nil, nil |
| 128 | } |
| 129 | if direct && a.Planner == nil && h.Staged != nil && h.Staged.Version == h.Target { |
| 130 | // Run the retained copy; step 2b records it first. Root is still |
| 131 | // asked to install the target, without waiting (ruling |
| 132 | // R-RETAINED-NUDGE): until it repoints, the live link names another |
| 133 | // build than the one running, and the sweep's convergence would |
| 134 | // cycle the instance after ConvergeGrace, waking it onto the same |
| 135 | // copy each time, until root's own pass. |
| 136 | if a.RequestStage != nil { |
| 137 | a.RequestStage(host, h.Target) |
| 138 | } |
| 139 | return h.Target, h.Staged, nil |
| 140 | } |
| 141 | if a.RequestStage != nil { |
| 142 | a.RequestStage(host, h.Target) |
| 143 | } |
| 144 | tick := time.NewTicker(heldPoll) |
| 145 | defer tick.Stop() |
| 146 | for n := 1; ; n++ { |
| 147 | if a.liveVersion(host) == h.Target { |
| 148 | return h.Target, nil, nil |
| 149 | } |
| 150 | if n%heldReask == 0 && wait.Err() == nil { |
| 151 | // The answer is asked again while the wake waits: a member's |
| 152 | // change (a pin, Update now, auto) or root's verdict (failed, |
| 153 | // unknown) ends this wait now, rather than at its bound onto a |
| 154 | // target that is no longer the instance's. The next wake checks |
| 155 | // afresh. The re-ask is bounded by the wait (Task 12a): it may |
| 156 | // read the bucket for up to its own budget, and an answer that |
| 157 | // comes back after the wait has ended is not waited for. |
| 158 | next, ok := a.reaskHeld(wait, host) |
| 159 | if !ok { |
| 160 | return "", nil, fmt.Errorf("%w: %s is on %q, wants %s", errStaleLink, host, installed, h.Target) |
| 161 | } |
| 162 | if !sameHeld(next, h) { |
| 163 | switch { |
| 164 | case next.Held && (next.Unknown || next.Target == ""): |
| 165 | return "", nil, errBuildUnknown |
| 166 | case next.Held && next.Failed && a.Planner == nil: |
| 167 | return "", nil, &targetFailedError{host: host, build: next.Target} |
| 168 | } |
| 169 | return "", nil, fmt.Errorf("%w: %s's target changed from %s while it waited", errStaleLink, host, h.Target) |
| 170 | } |
| 171 | } |
| 172 | select { |
| 173 | case <-wait.Done(): |
| 174 | return "", nil, fmt.Errorf("%w: %s is on %q, wants %s", errStaleLink, host, installed, h.Target) |
| 175 | case <-tick.C: |
| 176 | } |
| 177 | } |
| 178 | } |
| 179 | |
| 180 | // heldReask is how many polls a held wake's wait makes between asking |
| 181 | // HeldTarget again (which may read the bucket). |
| 182 | const heldReask = 8 |
| 183 | |
| 184 | // awaitRootDecision is wakeCheck's wait for root when root has made no |
| 185 | // decision for the instance's current settings (h.AskRoot; ruling |
| 186 | // R-HELD-PURE). It returns h itself when the wake may go on now: its target |
| 187 | // failed (the failure rule decides, below) or the live link already names |
| 188 | // it, the build the data last ran. Otherwise it asks root for a pass, |
| 189 | // naming no build (it does not know which root will install), and asks |
| 190 | // HeldTarget again every heldReask polls until root's answer is no longer |
| 191 | // AskRoot, which it returns; or the live link names the target (root decided |
| 192 | // to hold it and installed it). A wait that ends first is errStaleLink. |
| 193 | func (a *Service) awaitRootDecision(wait context.Context, host string, h HeldWake) (HeldWake, error) { |
| 194 | if h.Failed || (h.Target != "" && a.liveVersion(host) == h.Target) { |
| 195 | return h, nil |
| 196 | } |
| 197 | if a.RequestStage != nil { |
| 198 | a.RequestStage(host, "") |
| 199 | } |
| 200 | tick := time.NewTicker(heldPoll) |
| 201 | defer tick.Stop() |
| 202 | for n := 1; ; n++ { |
| 203 | if h.Target != "" && a.liveVersion(host) == h.Target { |
| 204 | return h, nil |
| 205 | } |
| 206 | if n%heldReask == 0 && wait.Err() == nil { |
| 207 | next, ok := a.reaskHeld(wait, host) |
| 208 | if !ok { |
| 209 | break |
| 210 | } |
| 211 | if !next.AskRoot { |
| 212 | return next, nil |
| 213 | } |
| 214 | h = next |
| 215 | } |
| 216 | select { |
| 217 | case <-wait.Done(): |
| 218 | return HeldWake{}, fmt.Errorf("%w: root has not decided for %s's current update settings", errStaleLink, host) |
| 219 | case <-tick.C: |
| 220 | } |
| 221 | } |
| 222 | return HeldWake{}, fmt.Errorf("%w: root has not decided for %s's current update settings", errStaleLink, host) |
| 223 | } |
| 224 | |
| 225 | // reaskHeld is HeldTarget(host) bounded by wait: false when wait ends first. |
| 226 | // The call itself is not cancelled (HeldTarget takes no context); it finishes |
| 227 | // within its own budget into a buffered channel nobody reads. |
| 228 | func (a *Service) reaskHeld(wait context.Context, host string) (HeldWake, bool) { |
| 229 | got := make(chan HeldWake, 1) |
| 230 | go func() { got <- a.HeldTarget(host) }() |
| 231 | select { |
| 232 | case h := <-got: |
| 233 | return h, true |
| 234 | case <-wait.Done(): |
| 235 | return HeldWake{}, false |
| 236 | } |
| 237 | } |
| 238 | |
| 239 | // sameHeld reports whether two answers send the wake to the same place. |
| 240 | func sameHeld(x, y HeldWake) bool { |
| 241 | return x.Held == y.Held && x.Target == y.Target && x.Unknown == y.Unknown && x.Failed == y.Failed && |
| 242 | x.AskRoot == y.AskRoot && x.Unread == y.Unread |
| 243 | } |
| 244 | |
| 245 | // liveVersion is what live/<host> names now, "" when it cannot be read. |
| 246 | func (a *Service) liveVersion(host string) string { |
| 247 | if a.liveVer == nil { |
| 248 | return "" |
| 249 | } |
| 250 | v, _ := a.liveVer(host) |
| 251 | return v |
| 252 | } |
| 253 | |
| 254 | // heldIdentityOK reports whether a held wake may run id: its target, or the |
| 255 | // good build the planner chose as recovery because the failure book blocks |
| 256 | // the target (zero-downtime failure records outrank every rule, §4.3). An |
| 257 | // identity with no version (a live link outside the version store) is never |
| 258 | // the target (ruling R-EXEC-NOVERSION-2). |
| 259 | func heldIdentityOK(target string, id release.StagedIdentity, plan WakePlan, rt registry.Route) bool { |
| 260 | if target == "" || id.Version == target { |
| 261 | return true |
| 262 | } |
| 263 | return plan.planned() && !plan.FirstWake && id.Version != "" && id.Version == rt.GoodVersion |
| 264 | } |
| 265 | |
amadan