Code Coverage |
||||||||||
Lines |
Functions and Methods |
Classes and Traits |
||||||||
| Total | |
81.78% |
211 / 258 |
|
71.43% |
25 / 35 |
CRAP | |
0.00% |
0 / 1 |
| ContentCopyController | |
81.78% |
211 / 258 |
|
71.43% |
25 / 35 |
171.65 | |
0.00% |
0 / 1 |
| __construct | |
75.00% |
3 / 4 |
|
0.00% |
0 / 1 |
2.06 | |||
| jobDir | |
100.00% |
1 / 1 |
|
100.00% |
1 / 1 |
1 | |||
| statePath | |
100.00% |
1 / 1 |
|
100.00% |
1 / 1 |
1 | |||
| engineConfig | |
100.00% |
9 / 9 |
|
100.00% |
1 / 1 |
1 | |||
| engine | |
100.00% |
3 / 3 |
|
100.00% |
1 / 1 |
2 | |||
| lock | |
100.00% |
3 / 3 |
|
100.00% |
1 / 1 |
2 | |||
| jobToken | |
100.00% |
1 / 1 |
|
100.00% |
1 / 1 |
3 | |||
| controllerLockPath | |
100.00% |
1 / 1 |
|
100.00% |
1 / 1 |
1 | |||
| advance | |
46.51% |
20 / 43 |
|
0.00% |
0 / 1 |
49.43 | |||
| resetLocked | |
100.00% |
7 / 7 |
|
100.00% |
1 / 1 |
3 | |||
| withdrawCompatChannel | |
100.00% |
1 / 1 |
|
100.00% |
1 / 1 |
1 | |||
| status | |
100.00% |
8 / 8 |
|
100.00% |
1 / 1 |
2 | |||
| isInProgress | |
0.00% |
0 / 11 |
|
0.00% |
0 / 1 |
12 | |||
| isStale | |
100.00% |
2 / 2 |
|
100.00% |
1 / 1 |
2 | |||
| runSetup | |
91.67% |
11 / 12 |
|
0.00% |
0 / 1 |
3.01 | |||
| holdsLockOrFail | |
100.00% |
9 / 9 |
|
100.00% |
1 / 1 |
3 | |||
| runCopyBurst | |
90.91% |
10 / 11 |
|
0.00% |
0 / 1 |
6.03 | |||
| runCore | |
100.00% |
12 / 12 |
|
100.00% |
1 / 1 |
4 | |||
| runFinalize | |
93.75% |
15 / 16 |
|
0.00% |
0 / 1 |
7.01 | |||
| ownStepStillRunning | |
100.00% |
4 / 4 |
|
100.00% |
1 / 1 |
2 | |||
| swapAlreadyCompleted | |
100.00% |
7 / 7 |
|
100.00% |
1 / 1 |
4 | |||
| finishLocked | |
100.00% |
6 / 6 |
|
100.00% |
1 / 1 |
1 | |||
| callVerb | |
100.00% |
1 / 1 |
|
100.00% |
1 / 1 |
1 | |||
| verbFailed | |
100.00% |
3 / 3 |
|
100.00% |
1 / 1 |
4 | |||
| verbMessage | |
40.00% |
2 / 5 |
|
0.00% |
0 / 1 |
7.46 | |||
| fail | |
100.00% |
14 / 14 |
|
100.00% |
1 / 1 |
1 | |||
| writeRunning | |
100.00% |
9 / 9 |
|
100.00% |
1 / 1 |
2 | |||
| writeSuccess | |
100.00% |
9 / 9 |
|
100.00% |
1 / 1 |
2 | |||
| envelope | |
90.00% |
9 / 10 |
|
0.00% |
0 / 1 |
4.02 | |||
| phaseStatus | |
80.00% |
4 / 5 |
|
0.00% |
0 / 1 |
4.13 | |||
| jobExists | |
100.00% |
1 / 1 |
|
100.00% |
1 / 1 |
1 | |||
| loadState | |
100.00% |
10 / 10 |
|
100.00% |
1 / 1 |
5 | |||
| defaultState | |
100.00% |
7 / 7 |
|
100.00% |
1 / 1 |
1 | |||
| newLockToken | |
100.00% |
1 / 1 |
|
100.00% |
1 / 1 |
1 | |||
| saveState | |
63.64% |
7 / 11 |
|
0.00% |
0 / 1 |
7.73 | |||
| 1 | <?php |
| 2 | /** |
| 3 | * Orchestrates the stepped, chunked content-deploy (staging -> production) over many short requests. |
| 4 | * |
| 5 | * @package NewfoldLabs\WP\Module\Staging |
| 6 | */ |
| 7 | |
| 8 | namespace NewfoldLabs\WP\Module\Staging\Copy; |
| 9 | |
| 10 | use NewfoldLabs\WP\Module\Staging\Staging; |
| 11 | use NewfoldLabs\WP\Module\Staging\StagingCommandLock; |
| 12 | |
| 13 | /** |
| 14 | * Drives one deploy-files content copy through five phases -- setup -> copy -> core -> finalize -> |
| 15 | * done -- doing at most one bounded unit of work per advance() call, so a poll loop (the REST advance |
| 16 | * route or the WP-cron fallback) can finish an arbitrarily large deploy without any single request |
| 17 | * running long. |
| 18 | * |
| 19 | * The three bash bookends run through Staging's existing command path (token handshake, env, escaping): |
| 20 | * - setup -> `deploy_content_setup` (sweeps stale -new/-old; nothing else) |
| 21 | * - core -> `deploy_content_core` (WordPress core download + php-ai-client, deliberately |
| 22 | * after the copy so a long copy never leaves production |
| 23 | * running new core against old content) |
| 24 | * - finalize -> `deploy_content_finalize` (atomic promote of the finished -new tree over production) |
| 25 | * |
| 26 | * Between them, CopyEngine builds the `-new` hardlink-delta tree in bursts. Phase and progress are |
| 27 | * persisted under nfd-private so a crashed/closed request resumes instead of restarting, and every |
| 28 | * transition is mirrored into Staging's deploy-result file so the legacy `/staging/deploy` GET status |
| 29 | * poller and a recorded failure both keep working. |
| 30 | */ |
| 31 | class ContentCopyController { |
| 32 | |
| 33 | /** |
| 34 | * Running the bash `deploy_content_setup` bookend (core download, php-ai-client, stale sweep). |
| 35 | */ |
| 36 | const PHASE_SETUP = 'setup'; |
| 37 | |
| 38 | /** |
| 39 | * Driving CopyEngine to build the `-new` delta tree in bursts. |
| 40 | */ |
| 41 | const PHASE_COPY = 'copy'; |
| 42 | |
| 43 | /** |
| 44 | * Running the bash `deploy_content_core` bookend (WordPress core download + php-ai-client), after |
| 45 | * the copy finishes and immediately before the swap. Its own phase so finalize stays an instant, |
| 46 | * cannot-time-out rename -- see deploy_content_core in lib/.staging. |
| 47 | */ |
| 48 | const PHASE_CORE = 'core'; |
| 49 | |
| 50 | /** |
| 51 | * Running the bash `deploy_content_finalize` bookend (atomic promote). |
| 52 | */ |
| 53 | const PHASE_FINALIZE = 'finalize'; |
| 54 | |
| 55 | /** |
| 56 | * Advance() calls without a successful advance for this many seconds mark a job abandoned, so a |
| 57 | * crashed/closed deploy stops blocking later staging operations forever. Generous on purpose: an |
| 58 | * active deploy rewrites the phase-state file every burst, so only a truly stalled job trips it. |
| 59 | */ |
| 60 | const STALE_AFTER_SECONDS = 3600; |
| 61 | |
| 62 | /** |
| 63 | * Deploy finished successfully. |
| 64 | */ |
| 65 | const PHASE_DONE = 'done'; |
| 66 | |
| 67 | /** |
| 68 | * Deploy failed; the result file records why. |
| 69 | */ |
| 70 | const PHASE_ERROR = 'error'; |
| 71 | |
| 72 | /** |
| 73 | * The deploy command name this copy reports as, so it shares the legacy result/log plumbing. |
| 74 | */ |
| 75 | const DEPLOY_COMMAND = 'deploy_files'; |
| 76 | |
| 77 | /** |
| 78 | * The cross-version compat channel lib/.staging publishes the held mutex on (its LEGACY_LOCK_NAME |
| 79 | * transient). Kept in sync with that bash constant. |
| 80 | */ |
| 81 | const COMPAT_CHANNEL_TRANSIENT = 'nfd_staging_lock'; |
| 82 | |
| 83 | /** |
| 84 | * Staging instance used to resolve paths, run bash verbs, and write the result file. |
| 85 | * |
| 86 | * @var Staging |
| 87 | */ |
| 88 | private $staging; |
| 89 | |
| 90 | /** |
| 91 | * Callable that runs a bash verb and returns its array|WP_Error response. |
| 92 | * |
| 93 | * @var callable |
| 94 | */ |
| 95 | private $verb_runner; |
| 96 | |
| 97 | /** |
| 98 | * Lazily constructed copy engine. |
| 99 | * |
| 100 | * @var CopyEngine|null |
| 101 | */ |
| 102 | private $engine; |
| 103 | |
| 104 | /** |
| 105 | * Lazily constructed staging-lock accessor. |
| 106 | * |
| 107 | * @var StagingCommandLock|null |
| 108 | */ |
| 109 | private $lock; |
| 110 | |
| 111 | /** |
| 112 | * Construct a controller bound to a Staging instance. |
| 113 | * |
| 114 | * @param Staging $staging Staging instance (resolves dirs, runs verbs, writes result). |
| 115 | * @param callable|null $verb_runner Optional override for running a bash verb; defaults to |
| 116 | * Staging::runContentDeployVerb (used by tests to inject). |
| 117 | */ |
| 118 | public function __construct( Staging $staging, $verb_runner = null ) { |
| 119 | $this->staging = $staging; |
| 120 | $this->verb_runner = is_callable( $verb_runner ) |
| 121 | ? $verb_runner |
| 122 | : array( $staging, 'runContentDeployVerb' ); |
| 123 | } |
| 124 | |
| 125 | /** |
| 126 | * Absolute job directory for the engine's manifest/cursor/lock and this controller's phase file. |
| 127 | * |
| 128 | * @return string |
| 129 | */ |
| 130 | private function jobDir() { |
| 131 | return trailingslashit( $this->staging->getProductionDir() ) . 'nfd-private/copy-deploy-files'; |
| 132 | } |
| 133 | |
| 134 | /** |
| 135 | * Absolute path to the persisted phase-state file. |
| 136 | * |
| 137 | * @return string |
| 138 | */ |
| 139 | private function statePath() { |
| 140 | return $this->jobDir() . '/phase-state.json'; |
| 141 | } |
| 142 | |
| 143 | /** |
| 144 | * Resolve the CopyEngine config for a deploy-files content copy. |
| 145 | * |
| 146 | * @return array |
| 147 | */ |
| 148 | public function engineConfig() { |
| 149 | $production = untrailingslashit( $this->staging->getProductionDir() ); |
| 150 | $staging = untrailingslashit( $this->staging->getStagingDir() ); |
| 151 | |
| 152 | return array( |
| 153 | 'source_root' => $staging . '/wp-content', |
| 154 | 'dest_root' => $production . '/wp-content', |
| 155 | 'dest_suffix' => '-new', |
| 156 | 'link_dest_root' => $production . '/wp-content', |
| 157 | 'job_dir' => $this->jobDir(), |
| 158 | ); |
| 159 | } |
| 160 | |
| 161 | /** |
| 162 | * Lazily build (and reuse) the copy engine. |
| 163 | * |
| 164 | * @return CopyEngine |
| 165 | */ |
| 166 | private function engine() { |
| 167 | if ( null === $this->engine ) { |
| 168 | $this->engine = new CopyEngine( $this->engineConfig() ); |
| 169 | } |
| 170 | |
| 171 | return $this->engine; |
| 172 | } |
| 173 | |
| 174 | /** |
| 175 | * Lazily build (and reuse) a handle on the staging command mutex. |
| 176 | * |
| 177 | * Not the same lock as controllerLockPath(), which only serializes two concurrent advance() calls |
| 178 | * of THIS job. This one is the site-wide mutex shared with lib/.staging, and it is what keeps a |
| 179 | * clone, create, destroy or legacy deploy from running against the same trees while this job is |
| 180 | * mid-copy. |
| 181 | * |
| 182 | * @return StagingCommandLock |
| 183 | */ |
| 184 | private function lock() { |
| 185 | if ( null === $this->lock ) { |
| 186 | $this->lock = new StagingCommandLock( $this->staging->getProductionDir() ); |
| 187 | } |
| 188 | |
| 189 | return $this->lock; |
| 190 | } |
| 191 | |
| 192 | /** |
| 193 | * This job's staging-lock token: the identity its three bash bookends present to lib/.staging's |
| 194 | * lock_check so all of them, and the PHP copy bursts between them, share one hold on the mutex. |
| 195 | * |
| 196 | * @param array $state Current state. |
| 197 | * |
| 198 | * @return string |
| 199 | */ |
| 200 | private function jobToken( array $state ) { |
| 201 | // String-checked rather than cast: phase-state.json is read back with json_decode, so a |
| 202 | // truncated or hand-edited file can put an array or object here, and casting that would raise |
| 203 | // an "Array to string conversion" notice -- printed into the REST response body, where it |
| 204 | // breaks the JSON the poll loop parses. An unusable token reads as no token, which fails |
| 205 | // closed (every path that needs one refuses without it). |
| 206 | return isset( $state['lock_token'] ) && is_string( $state['lock_token'] ) ? $state['lock_token'] : ''; |
| 207 | } |
| 208 | |
| 209 | /** |
| 210 | * Absolute path to the controller's advisory lock file. |
| 211 | * |
| 212 | * @return string |
| 213 | */ |
| 214 | private function controllerLockPath() { |
| 215 | return $this->jobDir() . '/controller.lock'; |
| 216 | } |
| 217 | |
| 218 | /** |
| 219 | * Do at most one bounded unit of work (one bash verb OR one engine burst) and return progress. |
| 220 | * |
| 221 | * The whole call runs under a non-blocking advisory lock, so a second caller -- a duplicate REST |
| 222 | * poll, the cron fallback, or a second browser tab -- can never run a phase transition concurrently |
| 223 | * (e.g. a setup that `rm -rf`s `-new` while another call's copy burst is filling it). A caller that |
| 224 | * cannot take the lock just reports current progress. |
| 225 | * |
| 226 | * @param bool $restart When true, discard any prior job (engine cursor/manifest + phase state) and |
| 227 | * start fresh. The UI passes this on the first advance of a new deploy so a |
| 228 | * previous run's terminal 'done'/'error' state does not short-circuit it. |
| 229 | * |
| 230 | * @return array { status: running|success|error, progress: array, phase: string, message?: string } |
| 231 | */ |
| 232 | public function advance( $restart = false ) { |
| 233 | $dir = $this->jobDir(); |
| 234 | if ( ! is_dir( $dir ) && ! wp_mkdir_p( $dir ) ) { |
| 235 | return array( |
| 236 | 'status' => 'error', |
| 237 | 'phase' => self::PHASE_ERROR, |
| 238 | 'progress' => array(), |
| 239 | 'message' => __( 'Could not create the content copy job directory.', 'wp-module-staging' ), |
| 240 | ); |
| 241 | } |
| 242 | |
| 243 | $lock = fopen( $this->controllerLockPath(), 'cb' ); // phpcs:ignore WordPress.WP.AlternativeFunctions.file_system_operations_fopen |
| 244 | if ( false === $lock ) { |
| 245 | return array( |
| 246 | 'status' => 'error', |
| 247 | 'phase' => self::PHASE_ERROR, |
| 248 | 'progress' => array(), |
| 249 | 'message' => __( 'Could not open the content copy lock.', 'wp-module-staging' ), |
| 250 | ); |
| 251 | } |
| 252 | if ( ! flock( $lock, LOCK_EX | LOCK_NB ) ) { |
| 253 | fclose( $lock ); |
| 254 | return $this->status(); // Another advance() is in flight; report without mutating. |
| 255 | } |
| 256 | |
| 257 | try { |
| 258 | if ( $restart ) { |
| 259 | $this->resetLocked(); |
| 260 | } elseif ( $this->jobExists() && $this->isStale() ) { |
| 261 | // A non-restart advance for a job that has not progressed in the stale window: it was |
| 262 | // abandoned (a tab asleep through a laptop suspend, say). Resuming it would run the copy |
| 263 | // against a tree another operation (clone/destroy) may have since mutated. Refuse and mark |
| 264 | // it failed; the client must start a fresh deploy (restart=true), which reseeds cleanly. |
| 265 | $stale = $this->loadState( false ); |
| 266 | $stale['phase'] = self::PHASE_ERROR; |
| 267 | $stale['message'] = __( 'The deploy was abandoned and timed out. Start it again.', 'wp-module-staging' ); |
| 268 | $this->saveState( $stale ); |
| 269 | return $this->envelope( 'error', $stale ); |
| 270 | } |
| 271 | |
| 272 | $state = $this->loadState( true ); |
| 273 | |
| 274 | switch ( $state['phase'] ) { |
| 275 | case self::PHASE_SETUP: |
| 276 | return $this->runSetup( $state ); |
| 277 | case self::PHASE_COPY: |
| 278 | return $this->runCopyBurst( $state ); |
| 279 | case self::PHASE_CORE: |
| 280 | return $this->runCore( $state ); |
| 281 | case self::PHASE_FINALIZE: |
| 282 | return $this->runFinalize( $state ); |
| 283 | case self::PHASE_ERROR: |
| 284 | return $this->envelope( 'error', $state ); |
| 285 | case self::PHASE_DONE: |
| 286 | default: |
| 287 | return $this->envelope( 'success', $state ); |
| 288 | } |
| 289 | } finally { |
| 290 | flock( $lock, LOCK_UN ); |
| 291 | fclose( $lock ); |
| 292 | } |
| 293 | } |
| 294 | |
| 295 | /** |
| 296 | * Discard any prior job so a new deploy starts clean. Clears the engine's cursor/manifest/lock and |
| 297 | * the phase-state file together, so the next setup sweeps stale `-new`/`-old` and the engine reseeds |
| 298 | * rather than resuming a previous run's cursor against a swept tree. Call only under the lock. |
| 299 | * |
| 300 | * @return void |
| 301 | */ |
| 302 | private function resetLocked() { |
| 303 | /* |
| 304 | * Release BEFORE the state file goes, because the state file holds the only record of the |
| 305 | * token. Deleting it first stranded the lock permanently: no later call could name the token, |
| 306 | * so nothing could ever release it and the site waited out the full TTL. That was reachable |
| 307 | * from an ordinary retry -- the UI resends `restart` after a gateway timeout, and a reload or |
| 308 | * a second tab does the same -- so the common case of "the first poll took too long" cost the |
| 309 | * user half an hour of refusals. |
| 310 | * |
| 311 | * A restart is also exactly when this is safe: the caller is abandoning the old job, so there |
| 312 | * is nothing of ours left that the lock protects. releaseIfOurs still declines a lock whose |
| 313 | * recorded holder is running, or cannot be shown not to be -- the one case where something of |
| 314 | * ours may still be writing, such as an orphaned setup part-way through its sweep. |
| 315 | */ |
| 316 | if ( is_file( $this->statePath() ) ) { |
| 317 | $old_token = $this->jobToken( $this->loadState( false ) ); |
| 318 | $this->lock()->releaseIfOurs( $old_token ); |
| 319 | $this->withdrawCompatChannel( $old_token ); |
| 320 | } |
| 321 | |
| 322 | $this->engine()->cleanup(); |
| 323 | if ( is_file( $this->statePath() ) ) { |
| 324 | wp_delete_file( $this->statePath() ); |
| 325 | } |
| 326 | } |
| 327 | |
| 328 | /** |
| 329 | * Withdraw this job's publication from the cross-version compat channel (lib/.staging's |
| 330 | * LEGACY_LOCK_NAME transient) so a restart or retry is not blocked by a token a failed job left |
| 331 | * behind. |
| 332 | * |
| 333 | * A stepped deploy's setup announces `active:<ts>:<token>` on this transient, so a peer running an |
| 334 | * older copy of the script (which cannot see the lock FILE) still sees the mutex held. If the job |
| 335 | * then fails while holding the lock, nothing withdraws that announcement: the failing verb's |
| 336 | * release path leaves the held lock in place, and the transient ages out only at LOCK_TTL. A fresh |
| 337 | * deploy carries a new token, so lock_check's legacy_lock_held sees a foreign, still-fresh holder |
| 338 | * and refuses every verb, blocking the retry the user just asked for for up to half an hour. |
| 339 | * Clearing our own prior publication on restart closes that window. |
| 340 | * |
| 341 | * Cleared directly on production's options table via $wpdb (bypassing staging's prefix and any |
| 342 | * persistent object cache, matching lib/.staging's legacy_lock_announce). Guarded by token so a |
| 343 | * concurrent operation's publication is never touched. |
| 344 | * |
| 345 | * @param string $token This job's lock token (the one being discarded). |
| 346 | * |
| 347 | * @return void |
| 348 | */ |
| 349 | private function withdrawCompatChannel( $token ) { |
| 350 | $this->staging->withdrawCompatLockPublication( $token ); |
| 351 | } |
| 352 | |
| 353 | /** |
| 354 | * Report current phase/progress without advancing. |
| 355 | * |
| 356 | * @return array |
| 357 | */ |
| 358 | public function status() { |
| 359 | if ( ! $this->jobExists() ) { |
| 360 | return array( |
| 361 | 'status' => 'idle', |
| 362 | 'phase' => 'idle', |
| 363 | 'progress' => array(), |
| 364 | ); |
| 365 | } |
| 366 | |
| 367 | $state = $this->loadState( false ); |
| 368 | |
| 369 | return $this->envelope( $this->phaseStatus( $state['phase'] ), $state ); |
| 370 | } |
| 371 | |
| 372 | /** |
| 373 | * Whether a job exists and has not yet reached a terminal phase. |
| 374 | * |
| 375 | * @return bool |
| 376 | */ |
| 377 | public function isInProgress() { |
| 378 | if ( ! $this->jobExists() ) { |
| 379 | return false; |
| 380 | } |
| 381 | |
| 382 | $state = $this->loadState( false ); |
| 383 | |
| 384 | $active = in_array( |
| 385 | $state['phase'], |
| 386 | array( self::PHASE_SETUP, self::PHASE_COPY, self::PHASE_CORE, self::PHASE_FINALIZE ), |
| 387 | true |
| 388 | ); |
| 389 | if ( ! $active ) { |
| 390 | return false; |
| 391 | } |
| 392 | |
| 393 | // An abandoned job (crashed/closed before reaching a terminal phase) would otherwise block every |
| 394 | // later staging operation forever, since nothing ever flips its phase to done/error. Treat a job |
| 395 | // that has not advanced within the stale window as dead, not in flight. |
| 396 | return ! $this->isStale(); |
| 397 | } |
| 398 | |
| 399 | /** |
| 400 | * Whether a job's phase-state file has not been touched within the stale window. advance() rewrites |
| 401 | * it on every burst (writeRunning/saveState), so an older mtime means no progress for that long. |
| 402 | * Only meaningful when a job exists. |
| 403 | * |
| 404 | * @return bool |
| 405 | */ |
| 406 | private function isStale() { |
| 407 | $mtime = @filemtime( $this->statePath() ); |
| 408 | return false !== $mtime && $mtime < ( time() - self::STALE_AFTER_SECONDS ); |
| 409 | } |
| 410 | |
| 411 | /** |
| 412 | * Phase 1: run the bash setup bookend, then move to the copy phase. |
| 413 | * |
| 414 | * @param array $state Current state. |
| 415 | * |
| 416 | * @return array |
| 417 | */ |
| 418 | private function runSetup( $state ) { |
| 419 | // Start the engine from a clean slate every time setup runs. A support engineer who deletes the |
| 420 | // phase-state file (but not the engine's cursor) would otherwise fall back to the setup phase |
| 421 | // while a STALE cursor still points mid-tree; the copy phase would then resume from it and |
| 422 | // promote only the tail of the tree over production. cleanup() removes the cursor/manifest so the |
| 423 | // engine reseeds, matching the stale -new/-old sweep the bash setup bookend does on the files. |
| 424 | $this->engine()->cleanup(); |
| 425 | |
| 426 | $this->writeRunning( $state ); |
| 427 | |
| 428 | $response = $this->callVerb( 'deploy_content_setup', $state ); |
| 429 | if ( $this->verbFailed( $response ) ) { |
| 430 | // Setup is the verb most likely to outlive its request: its stale-`-new` sweep is one |
| 431 | // unlink per file. So it is also the one whose orphan the next poll runs into, and that |
| 432 | // refusal must not be read as a broken deploy. |
| 433 | $running = $this->ownStepStillRunning( $state ); |
| 434 | if ( null !== $running ) { |
| 435 | return $running; |
| 436 | } |
| 437 | |
| 438 | return $this->fail( $state, $this->verbMessage( $response, __( 'Content deploy setup failed.', 'wp-module-staging' ) ) ); |
| 439 | } |
| 440 | |
| 441 | $state['phase'] = self::PHASE_COPY; |
| 442 | $this->saveState( $state ); |
| 443 | $this->writeRunning( $state ); |
| 444 | |
| 445 | return $this->envelope( 'running', $state ); |
| 446 | } |
| 447 | |
| 448 | /** |
| 449 | * Confirm this job still holds the staging mutex, and build the failure envelope if it does not. |
| 450 | * |
| 451 | * Every phase that is about to act on the trees has to ask, not just the copy. lib/.staging's |
| 452 | * lock_check cannot answer it for us: presented with no lock file at all it simply acquires one, |
| 453 | * so a verb could not tell "we have held this all along" from "we lost it, and whatever took it |
| 454 | * has since finished and released". The difference matters. The failure that motivates this: the |
| 455 | * last copy burst moves the job to the core phase, the poll driver then stalls for longer than |
| 456 | * the lock TTL but less than the job's stale window (a suspended laptop, a starved cron), a clone |
| 457 | * takes the lock, rewrites staging from production, finishes and releases -- and the core and |
| 458 | * finalize phases would then promote a `-new` tree built from the pre-clone staging content over |
| 459 | * production, because the clone only ever touched staging and left the `-new` tree standing. |
| 460 | * |
| 461 | * Either operation answers it, because both only succeed on a lock that exists and still carries |
| 462 | * this job's token -- false while a thief holds it and false after a thief has released it. |
| 463 | * |
| 464 | * WHICH ONE MATTERS. $refresh also WRITES, and the value it writes records pid 0, meaning "no |
| 465 | * single process owns this right now". Only the copy phase may say that: it is the one phase with |
| 466 | * nothing of the job's running between calls, and the refresh is also what keeps the lock from |
| 467 | * ageing out across a long copy. The phases that invoke a bash verb must NOT write, because the |
| 468 | * pid already in the lock may belong to a verb of this same job that is still running -- a |
| 469 | * finalize whose PHP request was killed by the FPM timeout while the bash process carried on |
| 470 | * swapping directories. That pid is the only thing standing between the next poll and a SECOND |
| 471 | * concurrent finalize: lock_check refuses to reclaim a lock whose recorded process is alive, and |
| 472 | * zeroing it here would tell it there is nobody to wait for, mid-swap. So core and finalize check |
| 473 | * read-only and leave the pid alone; the verb's own lock_check refreshes the timestamp a moment |
| 474 | * later anyway. |
| 475 | * |
| 476 | * @param array $state Current state. |
| 477 | * @param array $failure Set to the failure envelope when this returns false. |
| 478 | * @param bool $refresh Whether to also push the lock's timestamp forward (copy phase only). |
| 479 | * |
| 480 | * @return bool |
| 481 | */ |
| 482 | private function holdsLockOrFail( $state, &$failure, $refresh ) { |
| 483 | $token = $this->jobToken( $state ); |
| 484 | $held = $refresh ? $this->lock()->refresh( $token ) : $this->lock()->holds( $token ); |
| 485 | |
| 486 | if ( $held ) { |
| 487 | return true; |
| 488 | } |
| 489 | |
| 490 | $failure = $this->fail( |
| 491 | $state, |
| 492 | __( 'The deploy lost its lock to another staging operation and was stopped before changing production. Start the deploy again.', 'wp-module-staging' ) |
| 493 | ); |
| 494 | |
| 495 | return false; |
| 496 | } |
| 497 | |
| 498 | /** |
| 499 | * Phase 2: run one CopyEngine burst. Still reports running when the copy completes, because the |
| 500 | * core-download and finalize bookends have not run yet, so the poll driver keeps calling advance(). |
| 501 | * |
| 502 | * @param array $state Current state. |
| 503 | * |
| 504 | * @return array |
| 505 | */ |
| 506 | private function runCopyBurst( $state ) { |
| 507 | // Every burst re-asserts the mutex, which is also what keeps it alive: nothing of ours runs |
| 508 | // between bursts, so without a refresh here the lock setup took would age past its TTL |
| 509 | // mid-copy and lib/.staging would hand the mutex to the next operation that asked -- which |
| 510 | // would then mutate the very trees this job is enumerating and writing into. |
| 511 | if ( ! $this->holdsLockOrFail( $state, $failure, true ) ) { |
| 512 | return $failure; |
| 513 | } |
| 514 | |
| 515 | $result = $this->engine()->advance(); |
| 516 | $state['progress'] = isset( $result['progress'] ) ? $result['progress'] : array(); |
| 517 | |
| 518 | if ( 'error' === $result['status'] ) { |
| 519 | return $this->fail( $state, isset( $result['message'] ) ? $result['message'] : __( 'Copying content files failed.', 'wp-module-staging' ) ); |
| 520 | } |
| 521 | |
| 522 | if ( 'success' === $result['status'] ) { |
| 523 | $state['phase'] = self::PHASE_CORE; |
| 524 | } |
| 525 | |
| 526 | $this->saveState( $state ); |
| 527 | $this->writeRunning( $state ); |
| 528 | |
| 529 | return $this->envelope( 'running', $state ); |
| 530 | } |
| 531 | |
| 532 | /** |
| 533 | * Phase 2.5: run the bash core-download bookend (WordPress core + php-ai-client), then move to |
| 534 | * finalize. Kept distinct from finalize so finalize is only the instant swap -- see |
| 535 | * deploy_content_core in lib/.staging. Still reports running so the poll driver continues. |
| 536 | * |
| 537 | * @param array $state Current state. |
| 538 | * |
| 539 | * @return array |
| 540 | */ |
| 541 | private function runCore( $state ) { |
| 542 | // Read-only: this phase invokes a bash verb, so it must not overwrite the pid in the |
| 543 | // lock. See holdsLockOrFail. |
| 544 | if ( ! $this->holdsLockOrFail( $state, $failure, false ) ) { |
| 545 | return $failure; |
| 546 | } |
| 547 | |
| 548 | $response = $this->callVerb( 'deploy_content_core', $state ); |
| 549 | if ( $this->verbFailed( $response ) ) { |
| 550 | $running = $this->ownStepStillRunning( $state ); |
| 551 | if ( null !== $running ) { |
| 552 | return $running; |
| 553 | } |
| 554 | |
| 555 | return $this->fail( $state, $this->verbMessage( $response, __( 'Preparing the WordPress core for deploy failed.', 'wp-module-staging' ) ) ); |
| 556 | } |
| 557 | |
| 558 | $state['phase'] = self::PHASE_FINALIZE; |
| 559 | $this->saveState( $state ); |
| 560 | $this->writeRunning( $state ); |
| 561 | |
| 562 | return $this->envelope( 'running', $state ); |
| 563 | } |
| 564 | |
| 565 | /** |
| 566 | * Phase 3: run the bash finalize bookend, clean up the engine, mark done. |
| 567 | * |
| 568 | * @param array $state Current state. |
| 569 | * |
| 570 | * @return array |
| 571 | */ |
| 572 | private function runFinalize( $state ) { |
| 573 | /* |
| 574 | * A finalize this job already ran may have SUCCEEDED without this phase ever being told. |
| 575 | * |
| 576 | * deploy_content_finalize swaps, releases the mutex, and only THEN deletes the `-old` asides, |
| 577 | * which on a normal deploy is the only step whose cost scales with the site's file count (the |
| 578 | * sweep in deploy_content_setup is the other, and only after an abandoned attempt). The point of that |
| 579 | * ordering is that the request blocking on it is free to die: by then production is live and |
| 580 | * correct. But the request dying is also what stops saveState(PHASE_DONE) below from running, |
| 581 | * so the job is still recorded as `finalize`, and the next poll re-enters here. |
| 582 | * |
| 583 | * Without this check, that poll found the lock gone (the verb released it) and reported "the |
| 584 | * deploy lost its lock and was stopped BEFORE changing production" -- on a deploy that had |
| 585 | * completed. That is the exact symptom the release-then-delete ordering was built to remove, |
| 586 | * reintroduced on the PHP side. |
| 587 | */ |
| 588 | if ( $this->swapAlreadyCompleted( $state ) ) { |
| 589 | return $this->finishLocked( $state ); |
| 590 | } |
| 591 | |
| 592 | // Read-only: this phase invokes a bash verb, so it must not overwrite the pid in the |
| 593 | // lock. See holdsLockOrFail. |
| 594 | if ( ! $this->holdsLockOrFail( $state, $failure, false ) ) { |
| 595 | return $failure; |
| 596 | } |
| 597 | |
| 598 | /* |
| 599 | * Recorded BEFORE the verb runs, and durably, because the process that learns the answer may |
| 600 | * not survive to report it. It is what lets the check above tell "a finalize of ours ran and |
| 601 | * the trees are gone because it promoted them" from "the copy phase produced nothing to |
| 602 | * promote in the first place". Both look identical on disk. |
| 603 | */ |
| 604 | if ( empty( $state['finalize_attempted'] ) ) { |
| 605 | $state['finalize_attempted'] = true; |
| 606 | $this->saveState( $state ); |
| 607 | } |
| 608 | |
| 609 | $response = $this->callVerb( 'deploy_content_finalize', $state ); |
| 610 | if ( $this->verbFailed( $response ) ) { |
| 611 | // An earlier finalize of this same job may have got further than the one we just tried. |
| 612 | if ( $this->swapAlreadyCompleted( $state ) ) { |
| 613 | return $this->finishLocked( $state ); |
| 614 | } |
| 615 | |
| 616 | $running = $this->ownStepStillRunning( $state ); |
| 617 | if ( null !== $running ) { |
| 618 | return $running; |
| 619 | } |
| 620 | |
| 621 | return $this->fail( $state, $this->verbMessage( $response, __( 'Finalizing the content deploy failed.', 'wp-module-staging' ) ) ); |
| 622 | } |
| 623 | |
| 624 | return $this->finishLocked( $state ); |
| 625 | } |
| 626 | |
| 627 | /** |
| 628 | * The `running` envelope to return when a verb was refused because an earlier verb of THIS job is |
| 629 | * still working, or null when that is not what happened. |
| 630 | * |
| 631 | * lock_check will not reclaim a lock whose recorded process is alive, and it reports that with the |
| 632 | * same customer-facing message as every other refusal, so the message cannot be used to tell them |
| 633 | * apart. When the live holder is one of our own verbs -- a bookend that outlived the request that |
| 634 | * started it and is still running -- the job is progressing, not broken, and the poll driver has |
| 635 | * to be told to come back rather than shown a failure. |
| 636 | * |
| 637 | * Deliberately does NOT save the phase state. advance()'s stale window is measured from the last |
| 638 | * state write, so leaving it alone is what bounds this: a verb that hangs for ever ends as a |
| 639 | * failed job instead of polling for ever. |
| 640 | * |
| 641 | * @param array $state Current state. |
| 642 | * |
| 643 | * @return array|null |
| 644 | */ |
| 645 | private function ownStepStillRunning( array $state ) { |
| 646 | if ( ! $this->lock()->heldByOwnLiveProcess( $this->jobToken( $state ) ) ) { |
| 647 | return null; |
| 648 | } |
| 649 | |
| 650 | $this->writeRunning( $state ); |
| 651 | |
| 652 | return $this->envelope( 'running', $state ); |
| 653 | } |
| 654 | |
| 655 | /** |
| 656 | * Whether the swap this job was running has already finished, leaving nothing to promote. |
| 657 | * |
| 658 | * Only ever true once a finalize has actually been invoked (see `finalize_attempted`): a job whose |
| 659 | * staging source had no content directories at all would otherwise have no `-new` trees from the |
| 660 | * start and would be reported as a completed swap without the verb ever running, which would also |
| 661 | * skip releasing the mutex. |
| 662 | * |
| 663 | * "Nothing to promote" is the same test bash itself uses to decide a directory is already done |
| 664 | * (see finalize_new_content_dirs), and the states it could be confused with do not arise: a verb |
| 665 | * killed mid-swap leaves the later directories' `-new` in place AND leaves its lock behind, and a |
| 666 | * verb that failed and rolled back restores `-new` from each directory it had promoted. |
| 667 | * |
| 668 | * @param array $state Current state. |
| 669 | * |
| 670 | * @return bool |
| 671 | */ |
| 672 | private function swapAlreadyCompleted( array $state ) { |
| 673 | if ( empty( $state['finalize_attempted'] ) ) { |
| 674 | return false; |
| 675 | } |
| 676 | |
| 677 | $config = $this->engineConfig(); |
| 678 | foreach ( CopyEngine::DEFAULT_ROOTS as $dir ) { |
| 679 | if ( is_dir( $config['dest_root'] . '/' . $dir . $config['dest_suffix'] ) ) { |
| 680 | return false; |
| 681 | } |
| 682 | } |
| 683 | |
| 684 | return true; |
| 685 | } |
| 686 | |
| 687 | /** |
| 688 | * Record a completed deploy. Shared by the normal path and by the resumption check above, so a |
| 689 | * deploy finished by an orphaned verb is reported exactly like one finished in place. |
| 690 | * |
| 691 | * @param array $state Current state. |
| 692 | * |
| 693 | * @return array |
| 694 | */ |
| 695 | private function finishLocked( array $state ) { |
| 696 | $this->engine()->cleanup(); |
| 697 | |
| 698 | /* |
| 699 | * Normally the verb has already released the mutex by this point. It has not if the swap |
| 700 | * completed but the verb was killed before its release, so this is a best-effort tidy-up: |
| 701 | * releaseIfOurs only acts on a lock that is ours, records no bash pid, and is far enough |
| 702 | * inside the TTL that nobody may concurrently steal it. |
| 703 | */ |
| 704 | $this->lock()->releaseIfOurs( $this->jobToken( $state ) ); |
| 705 | |
| 706 | $state['phase'] = self::PHASE_DONE; |
| 707 | $this->saveState( $state ); |
| 708 | $this->writeSuccess( $state ); |
| 709 | |
| 710 | return $this->envelope( 'success', $state ); |
| 711 | } |
| 712 | |
| 713 | /** |
| 714 | * Invoke a bash verb through the configured runner, handing it this job's lock token. |
| 715 | * |
| 716 | * @param string $verb Verb name. |
| 717 | * @param array $state Current state, which carries this job's lock token. |
| 718 | * |
| 719 | * @return array|\WP_Error |
| 720 | */ |
| 721 | private function callVerb( $verb, array $state ) { |
| 722 | return call_user_func( $this->verb_runner, $verb, $this->jobToken( $state ) ); |
| 723 | } |
| 724 | |
| 725 | /** |
| 726 | * Whether a verb response indicates failure. |
| 727 | * |
| 728 | * @param mixed $response Verb response. |
| 729 | * |
| 730 | * @return bool |
| 731 | */ |
| 732 | private function verbFailed( $response ) { |
| 733 | if ( is_wp_error( $response ) ) { |
| 734 | return true; |
| 735 | } |
| 736 | |
| 737 | return is_array( $response ) && isset( $response['status'] ) && 'error' === $response['status']; |
| 738 | } |
| 739 | |
| 740 | /** |
| 741 | * Extract a human message from a failed verb response. |
| 742 | * |
| 743 | * @param mixed $response Verb response. |
| 744 | * @param string $default Fallback message. |
| 745 | * |
| 746 | * @return string |
| 747 | */ |
| 748 | private function verbMessage( $response, $default ) { |
| 749 | if ( is_wp_error( $response ) ) { |
| 750 | return $response->get_error_message(); |
| 751 | } |
| 752 | |
| 753 | if ( is_array( $response ) && ! empty( $response['message'] ) ) { |
| 754 | return (string) $response['message']; |
| 755 | } |
| 756 | |
| 757 | return $default; |
| 758 | } |
| 759 | |
| 760 | /** |
| 761 | * Record a failure: mark the error phase and persist it to the result file. |
| 762 | * |
| 763 | * @param array $state Current state. |
| 764 | * @param string $message Failure message. |
| 765 | * |
| 766 | * @return array |
| 767 | */ |
| 768 | private function fail( $state, $message ) { |
| 769 | $state['phase'] = self::PHASE_ERROR; |
| 770 | $state['message'] = $message; |
| 771 | $this->saveState( $state ); |
| 772 | |
| 773 | /* |
| 774 | * Give the mutex back. Every bash failure path releases through error(), but a failure on the |
| 775 | * PHP side -- an unreadable manifest, a full disk, a directory the account cannot enter -- |
| 776 | * had no release at all, so the lock the setup verb took stayed until its TTL ran out and |
| 777 | * every staging operation on the site was refused for up to half an hour over a deploy that |
| 778 | * was already over. |
| 779 | * |
| 780 | * releaseIfOurs is the right tool rather than an unlink because it refuses anything it cannot |
| 781 | * prove is free to delete: a lock carrying another token, one recording a bash pid (a verb may |
| 782 | * still be in flight under it, including an orphaned setup whose `rm -rf` is still running), |
| 783 | * or one close enough to the TTL that lib/.staging may concurrently decide it is abandoned. |
| 784 | */ |
| 785 | $this->lock()->releaseIfOurs( $this->jobToken( $state ) ); |
| 786 | $this->withdrawCompatChannel( $this->jobToken( $state ) ); |
| 787 | |
| 788 | $this->staging->recordDeployResult( |
| 789 | array( |
| 790 | 'status' => 'error', |
| 791 | 'command' => self::DEPLOY_COMMAND, |
| 792 | 'started_at' => $state['started_at'], |
| 793 | 'message' => $message, |
| 794 | ) |
| 795 | ); |
| 796 | |
| 797 | return $this->envelope( 'error', $state, $message ); |
| 798 | } |
| 799 | |
| 800 | /** |
| 801 | * Write a running snapshot to the shared deploy-result file. |
| 802 | * |
| 803 | * @param array $state Current state. |
| 804 | * |
| 805 | * @return void |
| 806 | */ |
| 807 | private function writeRunning( $state ) { |
| 808 | $this->staging->recordDeployResult( |
| 809 | array( |
| 810 | 'status' => 'running', |
| 811 | 'command' => self::DEPLOY_COMMAND, |
| 812 | 'started_at' => $state['started_at'], |
| 813 | 'message' => __( 'Deployment in progress. This may take several minutes.', 'wp-module-staging' ), |
| 814 | 'progress' => isset( $state['progress'] ) ? $state['progress'] : array(), |
| 815 | ) |
| 816 | ); |
| 817 | } |
| 818 | |
| 819 | /** |
| 820 | * Write the final success result to the shared deploy-result file. |
| 821 | * |
| 822 | * @param array $state Current state. |
| 823 | * |
| 824 | * @return void |
| 825 | */ |
| 826 | private function writeSuccess( $state ) { |
| 827 | $this->staging->recordDeployResult( |
| 828 | array( |
| 829 | 'status' => 'success', |
| 830 | 'command' => self::DEPLOY_COMMAND, |
| 831 | 'started_at' => $state['started_at'], |
| 832 | 'message' => __( 'Files deployed successfully.', 'wp-module-staging' ), |
| 833 | 'progress' => isset( $state['progress'] ) ? $state['progress'] : array(), |
| 834 | ) |
| 835 | ); |
| 836 | } |
| 837 | |
| 838 | /** |
| 839 | * Build the standard return envelope from the current state. |
| 840 | * |
| 841 | * @param string $status running|success|error|idle. |
| 842 | * @param array $state Current state. |
| 843 | * @param string $message Optional message override. |
| 844 | * |
| 845 | * @return array |
| 846 | */ |
| 847 | private function envelope( $status, $state, $message = '' ) { |
| 848 | $out = array( |
| 849 | 'status' => $status, |
| 850 | 'phase' => $state['phase'], |
| 851 | 'progress' => isset( $state['progress'] ) ? $state['progress'] : array(), |
| 852 | ); |
| 853 | |
| 854 | if ( '' !== $message ) { |
| 855 | $out['message'] = $message; |
| 856 | } elseif ( ! empty( $state['message'] ) ) { |
| 857 | $out['message'] = $state['message']; |
| 858 | } |
| 859 | |
| 860 | return $out; |
| 861 | } |
| 862 | |
| 863 | /** |
| 864 | * Map a phase to its outward status. |
| 865 | * |
| 866 | * @param string $phase Phase name. |
| 867 | * |
| 868 | * @return string |
| 869 | */ |
| 870 | private function phaseStatus( $phase ) { |
| 871 | switch ( $phase ) { |
| 872 | case self::PHASE_DONE: |
| 873 | return 'success'; |
| 874 | case self::PHASE_ERROR: |
| 875 | return 'error'; |
| 876 | default: |
| 877 | return 'running'; |
| 878 | } |
| 879 | } |
| 880 | |
| 881 | /** |
| 882 | * Whether a persisted job-state file exists. |
| 883 | * |
| 884 | * @return bool |
| 885 | */ |
| 886 | private function jobExists() { |
| 887 | return is_file( $this->statePath() ); |
| 888 | } |
| 889 | |
| 890 | /** |
| 891 | * Load the phase state, optionally creating a fresh job file if none exists. |
| 892 | * |
| 893 | * @param bool $create_if_missing When true, seed and persist a fresh setup-phase state. |
| 894 | * |
| 895 | * @return array |
| 896 | */ |
| 897 | private function loadState( $create_if_missing ) { |
| 898 | $path = $this->statePath(); |
| 899 | if ( is_file( $path ) ) { |
| 900 | $raw = file_get_contents( $path ); // phpcs:ignore WordPress.WP.AlternativeFunctions.file_get_contents_file_get_contents |
| 901 | $decoded = $raw ? json_decode( $raw, true ) : null; |
| 902 | if ( is_array( $decoded ) ) { |
| 903 | return array_merge( $this->defaultState(), $decoded ); |
| 904 | } |
| 905 | } |
| 906 | |
| 907 | $state = $this->defaultState(); |
| 908 | if ( $create_if_missing ) { |
| 909 | $this->saveState( $state ); |
| 910 | } |
| 911 | |
| 912 | return $state; |
| 913 | } |
| 914 | |
| 915 | /** |
| 916 | * The initial phase state for a fresh job. |
| 917 | * |
| 918 | * @return array |
| 919 | */ |
| 920 | private function defaultState() { |
| 921 | return array( |
| 922 | 'phase' => self::PHASE_SETUP, |
| 923 | 'started_at' => time(), |
| 924 | 'progress' => array(), |
| 925 | 'message' => '', |
| 926 | 'lock_token' => self::newLockToken(), |
| 927 | ); |
| 928 | } |
| 929 | |
| 930 | /** |
| 931 | * Mint a staging-lock token: 32 hex characters, from a generator that never calls into |
| 932 | * WordPress. The token has to be alphanumeric because it travels as a shell argument and as a |
| 933 | * field in a colon-delimited lock value; see Staging::randomToken for why wp_generate_password |
| 934 | * is the wrong tool for it. |
| 935 | * |
| 936 | * @return string |
| 937 | */ |
| 938 | private static function newLockToken() { |
| 939 | return Staging::randomToken(); |
| 940 | } |
| 941 | |
| 942 | /** |
| 943 | * Persist the phase state atomically (temp file + rename). |
| 944 | * |
| 945 | * @param array $state State to persist. |
| 946 | * |
| 947 | * @return void |
| 948 | */ |
| 949 | private function saveState( array $state ) { |
| 950 | $dir = $this->jobDir(); |
| 951 | if ( ! is_dir( $dir ) && ! wp_mkdir_p( $dir ) ) { |
| 952 | return; |
| 953 | } |
| 954 | |
| 955 | $json = wp_json_encode( $state ); |
| 956 | if ( false === $json ) { |
| 957 | return; |
| 958 | } |
| 959 | |
| 960 | $tmp = $this->statePath() . '.tmp-' . getmypid(); |
| 961 | // phpcs:ignore WordPress.WP.AlternativeFunctions.file_system_operations_file_put_contents |
| 962 | if ( false === file_put_contents( $tmp, $json, LOCK_EX ) ) { |
| 963 | return; |
| 964 | } |
| 965 | if ( ! rename( $tmp, $this->statePath() ) ) { // phpcs:ignore WordPress.WP.AlternativeFunctions.file_system_operations_rename |
| 966 | wp_delete_file( $tmp ); |
| 967 | } |
| 968 | } |
| 969 | } |