Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * datachecksum_state.c
4 : * Background worker for enabling or disabling data checksums online as
5 : * well as functionality for manipulating data checksum state
6 : *
7 : * When enabling data checksums on a cluster at initdb time or when shut down
8 : * with pg_checksums, no extra process is required as each page is checksummed,
9 : * and verified, when accessed. When enabling checksums on an already running
10 : * cluster, this worker will ensure that all pages are checksummed before
11 : * verification of the checksums is turned on. In the case of disabling
12 : * checksums, the state transition is performed only in the control file, no
13 : * changes are performed on the data pages.
14 : *
15 : * Checksums can be either enabled or disabled cluster-wide, with on/off being
16 : * the end state for data_checksums.
17 : *
18 : * 1. Enabling checksums
19 : * ---------------------
20 : * When enabling checksums in an online cluster, data_checksums will be set to
21 : * "inprogress-on" which signals that write operations MUST compute and write
22 : * the checksum on the data page, but during reading the checksum SHALL NOT be
23 : * verified. This ensures that all objects created while checksums are being
24 : * enabled will have checksums set, but reads won't fail due to missing or
25 : * invalid checksums. Invalid checksums can be present in case the cluster had
26 : * checksums enabled, then disabled them and updated the page while they were
27 : * disabled.
28 : *
29 : * The DataChecksumsWorker will compile a list of all databases at the start,
30 : * any databases created concurrently will see the in-progress state and will
31 : * be checksummed automatically. All databases from the original list MUST BE
32 : * successfully processed in order for data checksums to be enabled, the only
33 : * exception are databases which are dropped before having been processed.
34 : *
35 : * For each database, all relations which have storage are read and every data
36 : * page is marked dirty to force a write with the checksum. This will generate
37 : * a lot of WAL as the entire database is read and written.
38 : *
39 : * If the processing is interrupted by a cluster crash or restart, it needs to
40 : * be restarted from the beginning again as state isn't persisted.
41 : *
42 : * 2. Disabling checksums
43 : * ----------------------
44 : * When disabling checksums, data_checksums will be set to "inprogress-off"
45 : * which signals that checksums are written but no longer need to be verified.
46 : * This ensures that backends which have not yet transitioned to the
47 : * "inprogress-off" state will still see valid checksums on pages.
48 : *
49 : * 3. Synchronization and Correctness
50 : * ----------------------------------
51 : * The processes involved in enabling or disabling data checksums in an
52 : * online cluster must be properly synchronized with the normal backends
53 : * serving concurrent queries to ensure correctness. Correctness is defined
54 : * as the following:
55 : *
56 : * - Backends SHALL NOT violate the data_checksums state they have agreed to
57 : * by acknowledging the procsignalbarrier: This means that all backends
58 : * MUST calculate and write data checksums during all states except off;
59 : * MUST validate checksums only in the 'on' state.
60 : * - Data checksums SHALL NOT be considered enabled cluster-wide until all
61 : * currently connected backends have state "on": This means that all
62 : * backends must wait on the procsignalbarrier to be acknowledged by all
63 : * before proceeding to validate data checksums.
64 : *
65 : * There are two steps of synchronization required for changing data_checksums
66 : * in an online cluster: (i) changing state in the active backends ("on",
67 : * "off", "inprogress-on" and "inprogress-off"), and (ii) ensuring no
68 : * incompatible objects and processes are left in a database when workers end.
69 : * The former deals with cluster-wide agreement on data checksum state and the
70 : * latter with ensuring that any concurrent activity cannot break the data
71 : * checksum contract during processing.
72 : *
73 : * Synchronizing the state change is done with procsignal barriers. Before
74 : * updating the data_checksums state in the control file, all other backends must absorb the
75 : * barrier. Barrier absorption will happen during interrupt processing, which
76 : * means that connected backends will change state at different times. If
77 : * waiting for a barrier is done during startup, for example during replay, it
78 : * is important to realize that any locks held by the startup process might
79 : * cause deadlocks if backends end up waiting for those locks while startup
80 : * is waiting for a procsignalbarrier.
81 : *
82 : * 3.1 When Enabling Data Checksums
83 : * --------------------------------
84 : * A process which fails to observe data checksums being enabled can induce two
85 : * types of errors: failing to write the checksum when modifying the page and
86 : * failing to validate the data checksum on the page when reading it.
87 : *
88 : * When processing starts all backends belong to one of the below sets, with
89 : * one of Bd and Bi being empty:
90 : *
91 : * Bg: Backend updating the global state and emitting the procsignalbarrier
92 : * Bd: Backends in "off" state
93 : * Bi: Backends in "inprogress-on" state
94 : *
95 : * If processing is started in an online cluster then all backends are in Bd.
96 : * If processing was halted by the cluster shutting down (due to a crash or
97 : * intentional restart), the controlfile state "inprogress-on" will be observed
98 : * on system startup and all backends will be placed in Bd. The controlfile
99 : * state will also be set to "off".
100 : *
101 : * Backends transition Bd -> Bi via a procsignalbarrier which is emitted by the
102 : * DataChecksumsWorkerLauncherMain. When all backends have acknowledged the
103 : * barrier then Bd will be empty and the next phase can begin: calculating and
104 : * writing data checksums with DataChecksumsWorkers. When the
105 : * DataChecksumsWorker processes have finished writing checksums on all pages,
106 : * data checksums are enabled cluster-wide via another procsignalbarrier.
107 : * There are four sets of backends where Bd shall be an empty set:
108 : *
109 : * Bg: Backend updating the global state and emitting the procsignalbarrier
110 : * Bd: Backends in "off" state
111 : * Be: Backends in "on" state
112 : * Bi: Backends in "inprogress-on" state
113 : *
114 : * Backends in Bi and Be will write checksums when modifying a page, but only
115 : * backends in Be will verify the checksum during reading. The Bg backend is
116 : * blocked waiting for all backends in Bi to process interrupts and move to
117 : * Be. Any backend starting while Bg is waiting on the procsignalbarrier will
118 : * observe the global state being "on" and will thus automatically belong to
119 : * Be. Checksums are enabled cluster-wide when Bi is an empty set. Bi and Be
120 : * are compatible sets while still operating based on their local state as
121 : * both write data checksums.
122 : *
123 : * 3.2 When Disabling Data Checksums
124 : * ---------------------------------
125 : * A process which fails to observe that data checksums have been disabled
126 : * can induce two types of errors: writing the checksum when modifying the
127 : * page and validating a data checksum which is no longer correct due to
128 : * modifications to the page. The former is not an error per se as data
129 : * integrity is maintained, but it is wasteful. The latter will cause errors
130 : * in user operations. Assuming the following sets of backends:
131 : *
132 : * Bg: Backend updating the global state and emitting the procsignalbarrier
133 : * Bd: Backends in "off" state
134 : * Be: Backends in "on" state
135 : * Bo: Backends in "inprogress-off" state
136 : * Bi: Backends in "inprogress-on" state
137 : *
138 : * Backends transition from the Be state to Bd like so: Be -> Bo -> Bd. From
139 : * all other states, the transition can be straight to Bd.
140 : *
141 : * The goal is to transition all backends to Bd making the others empty sets.
142 : * Backends in Bo write data checksums, but don't validate them, such that
143 : * backends still in Be can continue to validate pages until the barrier has
144 : * been absorbed such that they are in Bo. Once all backends are in Bo, the
145 : * barrier to transition to "off" can be raised and all backends can safely
146 : * stop writing data checksums as no backend is enforcing data checksum
147 : * validation any longer.
148 : *
149 : * 4. Future opportunities for optimizations
150 : * -----------------------------------------
151 : * Below are some potential optimizations and improvements which were brought
152 : * up during reviews of this feature, but which weren't implemented in the
153 : * initial version. These are ideas listed without any validation on their
154 : * feasibility or potential payoff. More discussion on (most of) these can be
155 : * found on the -hackers threads linked to in the commit message of this
156 : * feature.
157 : *
158 : * * Launching datachecksumsworker for resuming operation from the startup
159 : * process: Currently users have to restart processing manually after a
160 : * restart since dynamic background worker cannot be started from the
161 : * postmaster. Changing the startup process could make restarting the
162 : * processing automatic on cluster restart.
163 : * * Avoid dirtying the page when checksums already match: Iff the checksum
164 : * on the page happens to already match we still dirty the page. It should
165 : * be enough to only do the log_newpage_buffer() call in that case.
166 : * * Teach pg_checksums to avoid checksummed pages when pg_checksums is used
167 : * to enable checksums on a cluster which is in inprogress-on state and
168 : * may have checksummed pages (make pg_checksums be able to resume an
169 : * online operation). This should only be attempted for wal_level minimal.
170 : * * Restartability (not necessarily with page granularity).
171 : * * Avoid processing databases which were created during inprogress-on.
172 : * Right now all databases are processed regardless to be safe.
173 : * * Teach CREATE DATABASE to calculate checksums for databases created
174 : * during inprogress-on with a template database which has yet to be
175 : * processed.
176 : *
177 : *
178 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
179 : * Portions Copyright (c) 1994, Regents of the University of California
180 : *
181 : *
182 : * IDENTIFICATION
183 : * src/backend/postmaster/datachecksum_state.c
184 : *
185 : *-------------------------------------------------------------------------
186 : */
187 : #include "postgres.h"
188 :
189 : #include "access/genam.h"
190 : #include "access/heapam.h"
191 : #include "access/htup_details.h"
192 : #include "access/xact.h"
193 : #include "access/xlog.h"
194 : #include "access/xloginsert.h"
195 : #include "catalog/indexing.h"
196 : #include "catalog/pg_class.h"
197 : #include "catalog/pg_database.h"
198 : #include "commands/progress.h"
199 : #include "commands/vacuum.h"
200 : #include "common/relpath.h"
201 : #include "miscadmin.h"
202 : #include "pgstat.h"
203 : #include "postmaster/bgworker.h"
204 : #include "postmaster/bgwriter.h"
205 : #include "postmaster/datachecksum_state.h"
206 : #include "storage/bufmgr.h"
207 : #include "storage/checksum.h"
208 : #include "storage/ipc.h"
209 : #include "storage/latch.h"
210 : #include "storage/lmgr.h"
211 : #include "storage/lwlock.h"
212 : #include "storage/procarray.h"
213 : #include "storage/smgr.h"
214 : #include "storage/subsystems.h"
215 : #include "tcop/tcopprot.h"
216 : #include "utils/builtins.h"
217 : #include "utils/fmgroids.h"
218 : #include "utils/injection_point.h"
219 : #include "utils/lsyscache.h"
220 : #include "utils/ps_status.h"
221 : #include "utils/syscache.h"
222 : #include "utils/wait_event.h"
223 :
224 : /*
225 : * Configuration of conditions which must match when absorbing a procsignal
226 : * barrier during data checksum enable/disable operations. A single function
227 : * is used for absorbing all barriers, and the current and target states must
228 : * be defined as a from/to tuple in the checksum_barriers struct.
229 : */
230 : typedef struct ChecksumBarrierCondition
231 : {
232 : /* Current state of data checksums */
233 : int from;
234 : /* Target state for data checksums */
235 : int to;
236 : } ChecksumBarrierCondition;
237 :
238 : static const ChecksumBarrierCondition checksum_barriers[9] =
239 : {
240 : /*
241 : * Disabling checksums: If checksums are currently enabled, disabling must
242 : * go through the 'inprogress-off' state.
243 : */
244 : {PG_DATA_CHECKSUM_VERSION, PG_DATA_CHECKSUM_INPROGRESS_OFF},
245 : {PG_DATA_CHECKSUM_INPROGRESS_OFF, PG_DATA_CHECKSUM_OFF},
246 :
247 : /*
248 : * If checksums are in the process of being enabled, but are not yet being
249 : * verified, we can abort by going back to 'off' state.
250 : */
251 : {PG_DATA_CHECKSUM_INPROGRESS_ON, PG_DATA_CHECKSUM_OFF},
252 :
253 : /*
254 : * Enabling checksums must normally go through the 'inprogress-on' state.
255 : */
256 : {PG_DATA_CHECKSUM_OFF, PG_DATA_CHECKSUM_INPROGRESS_ON},
257 : {PG_DATA_CHECKSUM_INPROGRESS_ON, PG_DATA_CHECKSUM_VERSION},
258 :
259 : /*
260 : * If checksums are being disabled but all backends are still computing
261 : * checksums, we can go straight back to 'on'
262 : */
263 : {PG_DATA_CHECKSUM_INPROGRESS_OFF, PG_DATA_CHECKSUM_VERSION},
264 :
265 : /*
266 : * If checksums are being enabled when launcher_exit is executed, state is
267 : * set to off since we cannot reach on at that point.
268 : */
269 : {PG_DATA_CHECKSUM_INPROGRESS_ON, PG_DATA_CHECKSUM_INPROGRESS_OFF},
270 :
271 : /*
272 : * Transitions that can happen when a new request is made while another is
273 : * currently being processed.
274 : */
275 : {PG_DATA_CHECKSUM_INPROGRESS_OFF, PG_DATA_CHECKSUM_INPROGRESS_ON},
276 : {PG_DATA_CHECKSUM_OFF, PG_DATA_CHECKSUM_INPROGRESS_OFF},
277 : };
278 :
279 : /*
280 : * Signaling between backends calling pg_enable/disable_data_checksums, the
281 : * checksums launcher process, and the checksums worker process.
282 : *
283 : * This struct is protected by DataChecksumsWorkerLock
284 : */
285 : typedef struct DataChecksumsStateStruct
286 : {
287 : /*
288 : * These are set by pg_{enable|disable}_data_checksums, to tell the
289 : * launcher what the target state is.
290 : */
291 : DataChecksumsWorkerOperation launch_operation;
292 : int launch_cost_delay;
293 : int launch_cost_limit;
294 :
295 : /*
296 : * Is a launcher process currently running? This is set by the main
297 : * launcher process, after it has read the above launch_* parameters.
298 : */
299 : bool launcher_running;
300 :
301 : /*
302 : * PID of the worker process, if it's currently running, of InvalidPid if
303 : * none. This is set by the worker launcher when it starts waiting for a
304 : * worker process to finish.
305 : */
306 : pid_t worker_pid;
307 :
308 : /*
309 : * These fields indicate the target state that the launcher is currently
310 : * working towards. They can be different from the corresponding launch_*
311 : * fields, if a new pg_enable/disable_data_checksums() call was made while
312 : * the launcher/worker was already running.
313 : *
314 : * The below members are set when the launcher starts, and are only
315 : * accessed read-only by the single worker. Thus, we can access these
316 : * without a lock. If multiple workers, or dynamic cost parameters, are
317 : * supported at some point then this would need to be revisited.
318 : */
319 : DataChecksumsWorkerOperation operation;
320 : int cost_delay;
321 : int cost_limit;
322 :
323 : /*
324 : * Signaling between the launcher and the worker process. Protected by
325 : * DataChecksumsWorkerLock.
326 : */
327 :
328 : /* result, set by worker before exiting */
329 : DataChecksumsWorkerResult success;
330 :
331 : /*
332 : * Tells the worker process whether it should also process the shared
333 : * catalogs
334 : */
335 : bool process_shared_catalogs;
336 : } DataChecksumsStateStruct;
337 :
338 : /* Shared memory segment for datachecksumsworker */
339 : static DataChecksumsStateStruct *DataChecksumState;
340 :
341 : typedef struct DataChecksumsWorkerDatabase
342 : {
343 : Oid dboid;
344 : char *dbname;
345 : } DataChecksumsWorkerDatabase;
346 :
347 : /* Flag set by the interrupt handler */
348 : static volatile sig_atomic_t abort_requested = false;
349 :
350 : /*
351 : * Have we set the DataChecksumsStateStruct->launcher_running flag?
352 : * If we have, we need to clear it before exiting!
353 : */
354 : static volatile sig_atomic_t launcher_running = false;
355 :
356 : /* Are we enabling data checksums, or disabling them? */
357 : static DataChecksumsWorkerOperation operation;
358 :
359 : /* Prototypes */
360 : static void DataChecksumsShmemRequest(void *arg);
361 : static bool DatabaseExists(Oid dboid);
362 : static List *BuildDatabaseList(void);
363 : static List *BuildRelationList(bool temp_relations, bool include_shared);
364 : static void FreeDatabaseList(List *dblist);
365 : static DataChecksumsWorkerResult ProcessDatabase(DataChecksumsWorkerDatabase *db);
366 : static bool ProcessAllDatabases(void);
367 : static bool ProcessSingleRelationFork(Relation reln, ForkNumber forkNum, BufferAccessStrategy strategy);
368 : static void launcher_cancel_handler(SIGNAL_ARGS);
369 : static void WaitForAllTransactionsToFinish(void);
370 :
371 : const ShmemCallbacks DataChecksumsShmemCallbacks = {
372 : .request_fn = DataChecksumsShmemRequest,
373 : };
374 :
375 : #define CHECK_FOR_ABORT_REQUEST() \
376 : do { \
377 : LWLockAcquire(DataChecksumsWorkerLock, LW_SHARED); \
378 : if (DataChecksumState->launch_operation != operation) \
379 : abort_requested = true; \
380 : LWLockRelease(DataChecksumsWorkerLock); \
381 : } while (0)
382 :
383 :
384 : /*****************************************************************************
385 : * Functionality for manipulating the data checksum state in the cluster
386 : */
387 :
388 : void
389 8 : EmitAndWaitDataChecksumsBarrier(uint32 state)
390 : {
391 : uint64 barrier;
392 :
393 8 : switch (state)
394 : {
395 3 : case PG_DATA_CHECKSUM_INPROGRESS_ON:
396 3 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_ON);
397 3 : WaitForProcSignalBarrier(barrier);
398 3 : break;
399 :
400 1 : case PG_DATA_CHECKSUM_INPROGRESS_OFF:
401 1 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_OFF);
402 1 : WaitForProcSignalBarrier(barrier);
403 1 : break;
404 :
405 2 : case PG_DATA_CHECKSUM_VERSION:
406 2 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_ON);
407 2 : WaitForProcSignalBarrier(barrier);
408 2 : break;
409 :
410 2 : case PG_DATA_CHECKSUM_OFF:
411 2 : barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_OFF);
412 2 : WaitForProcSignalBarrier(barrier);
413 2 : break;
414 :
415 8 : default:
416 : Assert(false);
417 : }
418 8 : }
419 :
420 : /*
421 : * AbsorbDataChecksumsBarrier
422 : * Generic function for absorbing data checksum state changes
423 : *
424 : * All procsignalbarriers regarding data checksum state changes are absorbed
425 : * with this function. The set of conditions required for the state change to
426 : * be accepted are listed in the checksum_barriers struct, target_state is
427 : * used to look up the relevant entry.
428 : */
429 : bool
430 278 : AbsorbDataChecksumsBarrier(ProcSignalBarrierType barrier)
431 : {
432 : uint32 target_state;
433 278 : int current = data_checksums;
434 278 : bool found = false;
435 :
436 : /*
437 : * Translate the barrier condition to the target state, doing it here
438 : * instead of in the procsignal code saves the latter from knowing about
439 : * checksum states.
440 : */
441 278 : switch (barrier)
442 : {
443 94 : case PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_ON:
444 94 : target_state = PG_DATA_CHECKSUM_INPROGRESS_ON;
445 94 : break;
446 71 : case PROCSIGNAL_BARRIER_CHECKSUM_ON:
447 71 : target_state = PG_DATA_CHECKSUM_VERSION;
448 71 : break;
449 54 : case PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_OFF:
450 54 : target_state = PG_DATA_CHECKSUM_INPROGRESS_OFF;
451 54 : break;
452 59 : case PROCSIGNAL_BARRIER_CHECKSUM_OFF:
453 59 : target_state = PG_DATA_CHECKSUM_OFF;
454 59 : break;
455 0 : default:
456 0 : elog(ERROR, "incorrect barrier \"%i\" received", barrier);
457 : }
458 :
459 : /*
460 : * If the target state matches the current state then the barrier has been
461 : * repeated.
462 : */
463 278 : if (current == target_state)
464 1 : return true;
465 :
466 : /*
467 : * If the cluster is in recovery we skip the validation of current state
468 : * since the replay is trusted.
469 : */
470 277 : if (RecoveryInProgress())
471 : {
472 48 : SetLocalDataChecksumState(target_state);
473 48 : return true;
474 : }
475 :
476 : /*
477 : * Find the barrier condition definition for the target state. Not finding
478 : * a condition would be a grave programmer error as the states are a
479 : * discrete set.
480 : */
481 1044 : for (int i = 0; i < lengthof(checksum_barriers) && !found; i++)
482 : {
483 815 : if (checksum_barriers[i].from == current && checksum_barriers[i].to == target_state)
484 229 : found = true;
485 : }
486 :
487 : /*
488 : * If the relevant state criteria aren't satisfied, throw an error which
489 : * will be caught by the procsignal machinery for a later retry.
490 : */
491 229 : if (!found)
492 0 : ereport(ERROR,
493 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
494 : errmsg("incorrect data checksum state %i for target state %i",
495 : current, target_state));
496 :
497 229 : SetLocalDataChecksumState(target_state);
498 229 : return true;
499 : }
500 :
501 :
502 : /*
503 : * Disables data checksums for the cluster, if applicable. Starts a background
504 : * worker which turns off the data checksums.
505 : */
506 : Datum
507 7 : disable_data_checksums(PG_FUNCTION_ARGS)
508 : {
509 7 : PreventCommandDuringRecovery("pg_disable_data_checksums()");
510 :
511 7 : if (!superuser())
512 0 : ereport(ERROR,
513 : errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
514 : errmsg("must be superuser to change data checksum state"));
515 :
516 7 : StartDataChecksumsWorkerLauncher(DISABLE_DATACHECKSUMS, 0, 0);
517 7 : PG_RETURN_VOID();
518 : }
519 :
520 : /*
521 : * Enables data checksums for the cluster, if applicable. Supports vacuum-
522 : * like cost based throttling to limit system load. Starts a background worker
523 : * which updates data checksums on existing data.
524 : */
525 : Datum
526 11 : enable_data_checksums(PG_FUNCTION_ARGS)
527 : {
528 11 : int cost_delay = PG_GETARG_INT32(0);
529 11 : int cost_limit = PG_GETARG_INT32(1);
530 :
531 11 : PreventCommandDuringRecovery("pg_enable_data_checksums()");
532 :
533 11 : if (!superuser())
534 0 : ereport(ERROR,
535 : errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
536 : errmsg("must be superuser to change data checksum state"));
537 :
538 11 : if (cost_delay < 0)
539 0 : ereport(ERROR,
540 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
541 : errmsg("cost delay cannot be a negative value"));
542 :
543 11 : if (cost_limit <= 0)
544 0 : ereport(ERROR,
545 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
546 : errmsg("cost limit must be greater than zero"));
547 :
548 11 : StartDataChecksumsWorkerLauncher(ENABLE_DATACHECKSUMS, cost_delay, cost_limit);
549 :
550 11 : PG_RETURN_VOID();
551 : }
552 :
553 :
554 : /*****************************************************************************
555 : * Functionality for running the datachecksumsworker and associated launcher
556 : */
557 :
558 : /*
559 : * StartDataChecksumsWorkerLauncher
560 : * Main entry point for datachecksumsworker launcher process
561 : *
562 : * The main entrypoint for starting data checksums processing for enabling as
563 : * well as disabling.
564 : */
565 : void
566 18 : StartDataChecksumsWorkerLauncher(DataChecksumsWorkerOperation op,
567 : int cost_delay,
568 : int cost_limit)
569 : {
570 : BackgroundWorker bgw;
571 : BackgroundWorkerHandle *bgw_handle;
572 : bool running;
573 :
574 : #ifdef USE_ASSERT_CHECKING
575 : /* The cost delay settings have no effect when disabling */
576 : if (op == DISABLE_DATACHECKSUMS)
577 : Assert(cost_delay == 0 && cost_limit == 0);
578 : #endif
579 :
580 18 : INJECTION_POINT("datachecksumsworker-startup-delay", NULL);
581 :
582 : /* Store the desired state in shared memory */
583 18 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
584 :
585 18 : DataChecksumState->launch_operation = op;
586 18 : DataChecksumState->launch_cost_delay = cost_delay;
587 18 : DataChecksumState->launch_cost_limit = cost_limit;
588 :
589 : /* Is the launcher already running? If so, what is it doing? */
590 18 : running = DataChecksumState->launcher_running;
591 :
592 18 : LWLockRelease(DataChecksumsWorkerLock);
593 :
594 : /*
595 : * Launch a new launcher process, if it's not running already.
596 : *
597 : * If the launcher is currently busy enabling the checksums, and we want
598 : * them disabled (or vice versa), the launcher will notice that at latest
599 : * when it's about to exit, and will loop back to process the new request.
600 : * So if the launcher is already running, we don't need to do anything
601 : * more here to abort it.
602 : *
603 : * If you call pg_enable/disable_data_checksums() twice in a row, before
604 : * the launcher has had a chance to start up, we still end up launching it
605 : * twice. That's OK, the second invocation will see that a launcher is
606 : * already running and exit quickly.
607 : */
608 18 : if (!running)
609 : {
610 18 : if ((op == ENABLE_DATACHECKSUMS && DataChecksumsOn()) ||
611 7 : (op == DISABLE_DATACHECKSUMS && DataChecksumsOff()))
612 : {
613 3 : ereport(LOG,
614 : errmsg("data checksums already in desired state, exiting"));
615 3 : return;
616 : }
617 :
618 : /*
619 : * Prepare the BackgroundWorker and launch it.
620 : */
621 15 : memset(&bgw, 0, sizeof(bgw));
622 15 : bgw.bgw_flags = BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
623 15 : bgw.bgw_start_time = BgWorkerStart_RecoveryFinished;
624 15 : snprintf(bgw.bgw_library_name, BGW_MAXLEN, "postgres");
625 15 : snprintf(bgw.bgw_function_name, BGW_MAXLEN, "DataChecksumsWorkerLauncherMain");
626 15 : snprintf(bgw.bgw_name, BGW_MAXLEN, "datachecksums launcher");
627 15 : snprintf(bgw.bgw_type, BGW_MAXLEN, "datachecksums launcher");
628 15 : bgw.bgw_restart_time = BGW_NEVER_RESTART;
629 15 : bgw.bgw_notify_pid = MyProcPid;
630 15 : bgw.bgw_main_arg = (Datum) 0;
631 :
632 15 : if (!RegisterDynamicBackgroundWorker(&bgw, &bgw_handle))
633 0 : ereport(ERROR,
634 : errcode(ERRCODE_INSUFFICIENT_RESOURCES),
635 : errmsg("failed to start background worker to process data checksums"));
636 : }
637 : else
638 : {
639 0 : ereport(LOG,
640 : errmsg("data checksum processing already running"));
641 : }
642 : }
643 :
644 : /*
645 : * ProcessSingleRelationFork
646 : * Enable data checksums in a single relation/fork.
647 : *
648 : * Returns true if successful, and false if *aborted*. On error, an actual
649 : * error is raised in the lower levels.
650 : */
651 : static bool
652 7846 : ProcessSingleRelationFork(Relation reln, ForkNumber forkNum, BufferAccessStrategy strategy)
653 : {
654 7846 : BlockNumber numblocks = RelationGetNumberOfBlocksInFork(reln, forkNum);
655 : char activity[NAMEDATALEN * 2 + 128];
656 : char *relns;
657 :
658 7846 : relns = get_namespace_name(RelationGetNamespace(reln));
659 :
660 : /* Report the current relation to pg_stat_activity */
661 7846 : snprintf(activity, sizeof(activity) - 1, "processing: %s.%s (%s, %u blocks)",
662 7846 : (relns ? relns : ""), RelationGetRelationName(reln), forkNames[forkNum], numblocks);
663 7846 : pgstat_report_activity(STATE_RUNNING, activity);
664 7846 : pgstat_progress_update_param(PROGRESS_DATACHECKSUMS_BLOCKS_TOTAL, numblocks);
665 7846 : if (relns)
666 7846 : pfree(relns);
667 :
668 : /*
669 : * We are looping over the blocks which existed at the time of process
670 : * start, which is safe since new blocks are created with checksums set
671 : * already due to the state being "inprogress-on".
672 : */
673 48849 : for (BlockNumber blknum = 0; blknum < numblocks; blknum++)
674 : {
675 41003 : Buffer buf = ReadBufferExtended(reln, forkNum, blknum, RBM_NORMAL, strategy);
676 :
677 : /* Need to get an exclusive lock to mark the buffer as dirty */
678 41003 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
679 :
680 : /*
681 : * Mark the buffer as dirty and force a full page write. We have to
682 : * re-write the page to WAL even if the checksum hasn't changed,
683 : * because if there is a replica it might have a slightly different
684 : * version of the page with an invalid checksum, caused by unlogged
685 : * changes (e.g. hint bits) on the primary happening while checksums
686 : * were off. This can happen if there was a valid checksum on the page
687 : * at one point in the past, so only when checksums are first on, then
688 : * off, and then turned on again. TODO: investigate if this could be
689 : * avoided if the checksum is calculated to be correct and wal_level
690 : * is set to "minimal".
691 : *
692 : * Unlogged relations don't need WAL since they are reset to their
693 : * init fork on recovery. We still dirty the buffer so that the
694 : * checksum is written to disk at the next checkpoint.
695 : *
696 : * The init fork is an exception: it is WAL-logged so the standby can
697 : * materialize the relation after promotion (see
698 : * ResetUnloggedRelations()). Skipping it here would leave the
699 : * standby with a stale init fork that, once copied to the main fork
700 : * on promotion, would fail checksum verification on every read.
701 : */
702 41003 : START_CRIT_SECTION();
703 41003 : MarkBufferDirty(buf);
704 41003 : if (RelationNeedsWAL(reln) || forkNum == INIT_FORKNUM)
705 40969 : log_newpage_buffer(buf, false);
706 41003 : END_CRIT_SECTION();
707 :
708 41003 : UnlockReleaseBuffer(buf);
709 :
710 : /* Check if we are asked to abort, the abortion will bubble up. */
711 : Assert(operation == ENABLE_DATACHECKSUMS);
712 41003 : LWLockAcquire(DataChecksumsWorkerLock, LW_SHARED);
713 41003 : if (DataChecksumState->launch_operation == DISABLE_DATACHECKSUMS)
714 0 : abort_requested = true;
715 41003 : LWLockRelease(DataChecksumsWorkerLock);
716 :
717 41003 : if (abort_requested)
718 0 : return false;
719 :
720 : /* update the block counter */
721 41003 : pgstat_progress_update_param(PROGRESS_DATACHECKSUMS_BLOCKS_DONE,
722 41003 : (blknum + 1));
723 :
724 : /*
725 : * Processing is re-using the vacuum cost delay for process
726 : * throttling, hence why we call vacuum APIs here.
727 : */
728 41003 : vacuum_delay_point(false);
729 : }
730 :
731 7846 : return true;
732 : }
733 :
734 : /*
735 : * ProcessSingleRelationByOid
736 : * Process a single relation based on oid.
737 : *
738 : * Returns true if successful, and false if *aborted*. On error, an actual
739 : * error is raised in the lower levels.
740 : */
741 : static bool
742 6040 : ProcessSingleRelationByOid(Oid relationId, BufferAccessStrategy strategy)
743 : {
744 : Relation rel;
745 6040 : bool aborted = false;
746 :
747 6040 : StartTransactionCommand();
748 :
749 6040 : rel = try_relation_open(relationId, AccessShareLock);
750 6040 : if (rel == NULL)
751 : {
752 : /*
753 : * Relation no longer exists. We don't consider this an error since
754 : * there are no pages in it that need data checksums, and thus return
755 : * true. The worker operates off a list of relations generated at the
756 : * start of processing, so relations being dropped in the meantime is
757 : * to be expected.
758 : */
759 0 : CommitTransactionCommand();
760 0 : pgstat_report_activity(STATE_IDLE, NULL);
761 0 : return true;
762 : }
763 6040 : RelationGetSmgr(rel);
764 :
765 30200 : for (ForkNumber fnum = 0; fnum <= MAX_FORKNUM; fnum++)
766 : {
767 24160 : if (smgrexists(rel->rd_smgr, fnum))
768 : {
769 7846 : if (!ProcessSingleRelationFork(rel, fnum, strategy))
770 : {
771 0 : aborted = true;
772 0 : break;
773 : }
774 : }
775 : }
776 6040 : relation_close(rel, AccessShareLock);
777 :
778 6040 : CommitTransactionCommand();
779 6040 : pgstat_report_activity(STATE_IDLE, NULL);
780 :
781 6040 : return !aborted;
782 : }
783 :
784 : /*
785 : * ProcessDatabase
786 : * Enable data checksums in a single database.
787 : *
788 : * We do this by launching a dynamic background worker into this database, and
789 : * waiting for it to finish. We have to do this in a separate worker, since
790 : * each process can only be connected to one database during its lifetime.
791 : */
792 : static DataChecksumsWorkerResult
793 23 : ProcessDatabase(DataChecksumsWorkerDatabase *db)
794 : {
795 : BackgroundWorker bgw;
796 : BackgroundWorkerHandle *bgw_handle;
797 : BgwHandleStatus status;
798 : pid_t pid;
799 : char activity[NAMEDATALEN + 64];
800 :
801 23 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
802 23 : DataChecksumState->success = DATACHECKSUMSWORKER_FAILED;
803 23 : LWLockRelease(DataChecksumsWorkerLock);
804 :
805 23 : memset(&bgw, 0, sizeof(bgw));
806 23 : bgw.bgw_flags = BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
807 23 : bgw.bgw_start_time = BgWorkerStart_RecoveryFinished;
808 23 : snprintf(bgw.bgw_library_name, BGW_MAXLEN, "postgres");
809 23 : snprintf(bgw.bgw_function_name, BGW_MAXLEN, "%s", "DataChecksumsWorkerMain");
810 23 : snprintf(bgw.bgw_name, BGW_MAXLEN, "datachecksums worker");
811 23 : snprintf(bgw.bgw_type, BGW_MAXLEN, "datachecksums worker");
812 23 : bgw.bgw_restart_time = BGW_NEVER_RESTART;
813 23 : bgw.bgw_notify_pid = MyProcPid;
814 23 : bgw.bgw_main_arg = ObjectIdGetDatum(db->dboid);
815 :
816 : /*
817 : * If there are no worker slots available, there is little we can do. If
818 : * we retry in a bit it's still unlikely that the user has managed to
819 : * reconfigure in the meantime and we'd be run through retries fast.
820 : */
821 23 : if (!RegisterDynamicBackgroundWorker(&bgw, &bgw_handle))
822 : {
823 0 : ereport(WARNING,
824 : errmsg("could not start background worker for enabling data checksums in database \"%s\"",
825 : db->dbname),
826 : errhint("The \"%s\" setting might be too low.", "max_worker_processes"));
827 0 : return DATACHECKSUMSWORKER_FAILED;
828 : }
829 :
830 23 : status = WaitForBackgroundWorkerStartup(bgw_handle, &pid);
831 23 : if (status == BGWH_STOPPED)
832 : {
833 : /*
834 : * If the worker managed to start, and stop, before we got to waiting
835 : * for it we can see a STOPPED status here without it being a failure.
836 : */
837 0 : LWLockAcquire(DataChecksumsWorkerLock, LW_SHARED);
838 0 : if (DataChecksumState->success == DATACHECKSUMSWORKER_SUCCESSFUL)
839 : {
840 0 : LWLockRelease(DataChecksumsWorkerLock);
841 0 : pgstat_report_activity(STATE_IDLE, NULL);
842 0 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
843 0 : DataChecksumState->worker_pid = InvalidPid;
844 0 : LWLockRelease(DataChecksumsWorkerLock);
845 0 : return DataChecksumState->success;
846 : }
847 0 : LWLockRelease(DataChecksumsWorkerLock);
848 :
849 0 : ereport(WARNING,
850 : errmsg("could not start background worker for enabling data checksums in database \"%s\"",
851 : db->dbname),
852 : errhint("More details on the error might be found in the server log."));
853 :
854 : /*
855 : * Heuristic to see if the database was dropped, and if it was we can
856 : * treat it as not an error, else treat as fatal and error out.
857 : */
858 0 : if (DatabaseExists(db->dboid))
859 0 : return DATACHECKSUMSWORKER_FAILED;
860 : else
861 0 : return DATACHECKSUMSWORKER_DROPDB;
862 : }
863 :
864 : /*
865 : * If the postmaster crashed we cannot end up with a processed database so
866 : * we have no alternative other than exiting. When enabling checksums we
867 : * won't at this time have changed the data checksums state in pg_control
868 : * to enabled so when the cluster comes back up processing will have to be
869 : * restarted.
870 : */
871 23 : if (status == BGWH_POSTMASTER_DIED)
872 0 : ereport(FATAL,
873 : errcode(ERRCODE_ADMIN_SHUTDOWN),
874 : errmsg("cannot enable data checksums without the postmaster process"),
875 : errhint("Restart the database and restart data checksum processing by calling pg_enable_data_checksums()."));
876 :
877 : Assert(status == BGWH_STARTED);
878 23 : ereport(LOG,
879 : errmsg("initiating data checksum processing in database \"%s\"",
880 : db->dbname));
881 :
882 : /* Save the pid of the worker so we can signal it later */
883 23 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
884 23 : DataChecksumState->worker_pid = pid;
885 23 : LWLockRelease(DataChecksumsWorkerLock);
886 :
887 23 : snprintf(activity, sizeof(activity) - 1,
888 : "Waiting for worker in database %s (pid %ld)", db->dbname, (long) pid);
889 23 : pgstat_report_activity(STATE_RUNNING, activity);
890 :
891 23 : status = WaitForBackgroundWorkerShutdown(bgw_handle);
892 22 : if (status == BGWH_POSTMASTER_DIED)
893 0 : ereport(FATAL,
894 : errcode(ERRCODE_ADMIN_SHUTDOWN),
895 : errmsg("postmaster exited during data checksum processing in \"%s\"",
896 : db->dbname),
897 : errhint("Restart the database and restart data checksum processing by calling pg_enable_data_checksums()."));
898 :
899 22 : LWLockAcquire(DataChecksumsWorkerLock, LW_SHARED);
900 22 : if (DataChecksumState->success == DATACHECKSUMSWORKER_ABORTED)
901 0 : ereport(LOG,
902 : errmsg("data checksums processing was aborted in database \"%s\"",
903 : db->dbname));
904 22 : LWLockRelease(DataChecksumsWorkerLock);
905 :
906 22 : pgstat_report_activity(STATE_IDLE, NULL);
907 22 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
908 22 : DataChecksumState->worker_pid = InvalidPid;
909 22 : LWLockRelease(DataChecksumsWorkerLock);
910 :
911 22 : return DataChecksumState->success;
912 : }
913 :
914 : /*
915 : * launcher_exit
916 : *
917 : * Internal routine for cleaning up state when a launcher process which has
918 : * performed checksum operations exits. A launcher process which is exiting due
919 : * to a duplicate started launcher does not need to perform any cleanup and
920 : * this function should not be called. Otherwise, we need to clean up the abort
921 : * flag to ensure that processing can be started again if it was previously
922 : * aborted (note: started again, *not* restarted from where it left off).
923 : */
924 : static void
925 14 : launcher_exit(int code, Datum arg)
926 : {
927 14 : abort_requested = false;
928 :
929 14 : if (launcher_running)
930 : {
931 2 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
932 2 : if (DataChecksumState->worker_pid != InvalidPid)
933 : {
934 1 : ereport(LOG,
935 : errmsg("data checksums launcher exiting while worker is still running, signalling worker"));
936 1 : kill(DataChecksumState->worker_pid, SIGTERM);
937 : }
938 2 : LWLockRelease(DataChecksumsWorkerLock);
939 : }
940 :
941 : /*
942 : * If the launcher is exiting before data checksums are enabled then set
943 : * the state to off since processing cannot be resumed.
944 : */
945 14 : if (DataChecksumsInProgressOn())
946 1 : SetDataChecksumsOff();
947 :
948 14 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
949 14 : launcher_running = false;
950 14 : DataChecksumState->launcher_running = false;
951 14 : LWLockRelease(DataChecksumsWorkerLock);
952 14 : }
953 :
954 : /*
955 : * launcher_cancel_handler
956 : *
957 : * Internal routine for reacting to SIGINT and flagging the worker to abort.
958 : * The worker won't be interrupted immediately but will check for abort flag
959 : * between each block in a relation.
960 : */
961 : static void
962 0 : launcher_cancel_handler(SIGNAL_ARGS)
963 : {
964 0 : int save_errno = errno;
965 :
966 0 : abort_requested = true;
967 :
968 : /*
969 : * There is no sleeping in the main loop, the flag will be checked
970 : * periodically in ProcessSingleRelationFork. The worker does however
971 : * sleep when waiting for concurrent transactions to end so we still need
972 : * to set the latch.
973 : */
974 0 : SetLatch(MyLatch);
975 :
976 0 : errno = save_errno;
977 0 : }
978 :
979 : /*
980 : * WaitForAllTransactionsToFinish
981 : * Blocks awaiting all current transactions to finish
982 : *
983 : * Returns when all transactions which are active at the call of the function
984 : * have ended, or if the postmaster dies while waiting. If the postmaster dies
985 : * the abort flag will be set to indicate that the caller of this shouldn't
986 : * proceed.
987 : *
988 : * NB: this will return early, if aborted by SIGINT or if the target state
989 : * is changed while we're running.
990 : */
991 : static void
992 9 : WaitForAllTransactionsToFinish(void)
993 : {
994 : TransactionId waitforxid;
995 :
996 9 : LWLockAcquire(XidGenLock, LW_SHARED);
997 9 : waitforxid = XidFromFullTransactionId(TransamVariables->nextXid);
998 9 : LWLockRelease(XidGenLock);
999 :
1000 9 : while (TransactionIdPrecedes(GetOldestActiveTransactionId(false, true), waitforxid))
1001 : {
1002 : char activity[64];
1003 : int rc;
1004 :
1005 : /* Oldest running xid is older than us, so wait */
1006 0 : snprintf(activity,
1007 : sizeof(activity),
1008 : "Waiting for transactions older than %u to end",
1009 : waitforxid);
1010 0 : pgstat_report_activity(STATE_RUNNING, activity);
1011 :
1012 : /* Retry every 3 seconds */
1013 0 : ResetLatch(MyLatch);
1014 0 : rc = WaitLatch(MyLatch,
1015 : WL_LATCH_SET | WL_TIMEOUT | WL_POSTMASTER_DEATH,
1016 : 3000,
1017 : WAIT_EVENT_CHECKSUM_ENABLE_STARTCONDITION);
1018 :
1019 : /*
1020 : * If the postmaster died we won't be able to enable checksums
1021 : * cluster-wide so abort and hope to continue when restarted.
1022 : */
1023 0 : if (rc & WL_POSTMASTER_DEATH)
1024 0 : ereport(FATAL,
1025 : errcode(ERRCODE_ADMIN_SHUTDOWN),
1026 : errmsg("postmaster exited during data checksums processing"),
1027 : errhint("Data checksums processing must be restarted manually after cluster restart."));
1028 :
1029 0 : CHECK_FOR_INTERRUPTS();
1030 0 : CHECK_FOR_ABORT_REQUEST();
1031 :
1032 0 : if (abort_requested)
1033 0 : break;
1034 : }
1035 :
1036 9 : pgstat_report_activity(STATE_IDLE, NULL);
1037 9 : return;
1038 : }
1039 :
1040 : /*
1041 : * DataChecksumsWorkerLauncherMain
1042 : *
1043 : * Main function for launching dynamic background workers for processing data
1044 : * checksums in databases. This function has the bgworker management, with
1045 : * ProcessAllDatabases being responsible for looping over the databases and
1046 : * initiating processing.
1047 : */
1048 : void
1049 14 : DataChecksumsWorkerLauncherMain(Datum arg)
1050 : {
1051 :
1052 14 : ereport(DEBUG1,
1053 : errmsg("background worker \"datachecksums launcher\" started"));
1054 :
1055 14 : pqsignal(SIGTERM, die);
1056 14 : pqsignal(SIGINT, launcher_cancel_handler);
1057 14 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
1058 14 : pqsignal(SIGUSR2, PG_SIG_IGN);
1059 :
1060 14 : BackgroundWorkerUnblockSignals();
1061 :
1062 14 : MyBackendType = B_DATACHECKSUMSWORKER_LAUNCHER;
1063 14 : init_ps_display(NULL);
1064 :
1065 14 : INJECTION_POINT("datachecksumsworker-launcher-delay", NULL);
1066 :
1067 14 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
1068 :
1069 14 : if (DataChecksumState->launcher_running)
1070 : {
1071 0 : ereport(LOG,
1072 : errmsg("background worker \"datachecksums launcher\" already running, exiting"));
1073 : /* Launcher was already running, let it finish */
1074 0 : LWLockRelease(DataChecksumsWorkerLock);
1075 0 : return;
1076 : }
1077 :
1078 14 : on_shmem_exit(launcher_exit, 0);
1079 14 : launcher_running = true;
1080 :
1081 : /* Initialize a connection to shared catalogs only */
1082 14 : BackgroundWorkerInitializeConnectionByOid(InvalidOid, InvalidOid, 0);
1083 :
1084 14 : operation = DataChecksumState->launch_operation;
1085 14 : DataChecksumState->launcher_running = true;
1086 14 : DataChecksumState->operation = operation;
1087 14 : DataChecksumState->cost_delay = DataChecksumState->launch_cost_delay;
1088 14 : DataChecksumState->cost_limit = DataChecksumState->launch_cost_limit;
1089 14 : LWLockRelease(DataChecksumsWorkerLock);
1090 :
1091 : /*
1092 : * The target state can change while we are busy enabling/disabling
1093 : * checksums, if the user calls pg_disable/enable_data_checksums() before
1094 : * we are finished with the previous request. In that case, we will loop
1095 : * back here, to process the new request.
1096 : */
1097 14 : again:
1098 :
1099 14 : pgstat_progress_start_command(PROGRESS_COMMAND_DATACHECKSUMS,
1100 : InvalidOid);
1101 :
1102 14 : if (operation == ENABLE_DATACHECKSUMS)
1103 : {
1104 : /*
1105 : * If we are asked to enable checksums in a cluster which already has
1106 : * checksums enabled, exit immediately as there is nothing more to do.
1107 : */
1108 9 : if (DataChecksumsNeedVerify())
1109 0 : goto done;
1110 :
1111 9 : ereport(LOG,
1112 : errmsg("enabling data checksums requested, starting data checksum calculation"));
1113 :
1114 : /*
1115 : * Set the state to inprogress-on and wait on the procsignal barrier.
1116 : */
1117 9 : pgstat_progress_update_param(PROGRESS_DATACHECKSUMS_PHASE,
1118 : PROGRESS_DATACHECKSUMS_PHASE_ENABLING);
1119 9 : SetDataChecksumsOnInProgress();
1120 :
1121 : /*
1122 : * All backends are now in inprogress-on state and are writing data
1123 : * checksums. Start processing all data at rest.
1124 : */
1125 9 : if (!ProcessAllDatabases())
1126 : {
1127 : /*
1128 : * If the target state changed during processing then it's not a
1129 : * failure, so restart processing instead.
1130 : */
1131 0 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
1132 0 : if (DataChecksumState->launch_operation != operation)
1133 : {
1134 0 : LWLockRelease(DataChecksumsWorkerLock);
1135 0 : goto done;
1136 : }
1137 0 : LWLockRelease(DataChecksumsWorkerLock);
1138 0 : ereport(ERROR,
1139 : errcode(ERRCODE_INSUFFICIENT_RESOURCES),
1140 : errmsg("unable to enable data checksums in cluster"));
1141 : }
1142 :
1143 : /*
1144 : * Data checksums have been set on all pages, set the state to on in
1145 : * order to instruct backends to validate checksums on reading.
1146 : */
1147 7 : SetDataChecksumsOn();
1148 :
1149 7 : ereport(LOG,
1150 : errmsg("data checksums are now enabled"));
1151 : }
1152 5 : else if (operation == DISABLE_DATACHECKSUMS)
1153 : {
1154 5 : ereport(LOG,
1155 : errmsg("disabling data checksums requested"));
1156 :
1157 5 : pgstat_progress_update_param(PROGRESS_DATACHECKSUMS_PHASE,
1158 : PROGRESS_DATACHECKSUMS_PHASE_DISABLING);
1159 5 : SetDataChecksumsOff();
1160 5 : ereport(LOG,
1161 : errmsg("data checksums are now disabled"));
1162 : }
1163 : else
1164 : Assert(false);
1165 :
1166 0 : done:
1167 :
1168 : /*
1169 : * This state will only be displayed for a fleeting moment, but for the
1170 : * sake of correctness it is still added before ending the command.
1171 : */
1172 12 : pgstat_progress_update_param(PROGRESS_DATACHECKSUMS_PHASE,
1173 : PROGRESS_DATACHECKSUMS_PHASE_DONE);
1174 :
1175 : /*
1176 : * All done. But before we exit, check if the target state was changed
1177 : * while we were running. In that case we will have to start all over
1178 : * again.
1179 : */
1180 12 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
1181 12 : if (DataChecksumState->launch_operation != operation)
1182 : {
1183 0 : DataChecksumState->operation = DataChecksumState->launch_operation;
1184 0 : operation = DataChecksumState->launch_operation;
1185 0 : DataChecksumState->cost_delay = DataChecksumState->launch_cost_delay;
1186 0 : DataChecksumState->cost_limit = DataChecksumState->launch_cost_limit;
1187 0 : LWLockRelease(DataChecksumsWorkerLock);
1188 0 : goto again;
1189 : }
1190 :
1191 : /* Shut down progress reporting as we are done */
1192 12 : pgstat_progress_end_command();
1193 :
1194 12 : launcher_running = false;
1195 12 : DataChecksumState->launcher_running = false;
1196 12 : LWLockRelease(DataChecksumsWorkerLock);
1197 : }
1198 :
1199 : /*
1200 : * ProcessAllDatabases
1201 : * Compute the list of all databases and process checksums in each
1202 : *
1203 : * This will generate a list of databases to process for enabling checksums.
1204 : * If a database encounters a failure then processing will end immediately and
1205 : * return an error.
1206 : */
1207 : static bool
1208 9 : ProcessAllDatabases(void)
1209 : {
1210 : List *DatabaseList;
1211 9 : int cumulative_total = 0;
1212 :
1213 : /* Set up so first run processes shared catalogs, not once in every db */
1214 9 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
1215 9 : DataChecksumState->process_shared_catalogs = true;
1216 9 : LWLockRelease(DataChecksumsWorkerLock);
1217 :
1218 : /* Get a list of all databases to process */
1219 9 : WaitForAllTransactionsToFinish();
1220 9 : DatabaseList = BuildDatabaseList();
1221 :
1222 : /*
1223 : * Update progress reporting with the total number of databases we need to
1224 : * process. This number should not be changed during processing, the
1225 : * columns for processed databases is instead increased such that it can
1226 : * be compared against the total.
1227 : */
1228 : {
1229 9 : const int index[] = {
1230 : PROGRESS_DATACHECKSUMS_DBS_TOTAL,
1231 : PROGRESS_DATACHECKSUMS_DBS_DONE,
1232 : PROGRESS_DATACHECKSUMS_RELS_TOTAL,
1233 : PROGRESS_DATACHECKSUMS_RELS_DONE,
1234 : PROGRESS_DATACHECKSUMS_BLOCKS_TOTAL,
1235 : PROGRESS_DATACHECKSUMS_BLOCKS_DONE,
1236 : };
1237 :
1238 : int64 vals[6];
1239 :
1240 9 : vals[0] = list_length(DatabaseList);
1241 9 : vals[1] = 0;
1242 : /* translated to NULL */
1243 9 : vals[2] = -1;
1244 9 : vals[3] = -1;
1245 9 : vals[4] = -1;
1246 9 : vals[5] = -1;
1247 :
1248 9 : pgstat_progress_update_multi_param(6, index, vals);
1249 : }
1250 :
1251 37 : foreach_ptr(DataChecksumsWorkerDatabase, db, DatabaseList)
1252 : {
1253 : DataChecksumsWorkerResult result;
1254 :
1255 23 : result = ProcessDatabase(db);
1256 :
1257 : #ifdef USE_INJECTION_POINTS
1258 : /* Allow a test process to alter the result of the operation */
1259 22 : if (IS_INJECTION_POINT_ATTACHED("datachecksumsworker-fail-db-result"))
1260 : {
1261 1 : result = DATACHECKSUMSWORKER_FAILED;
1262 1 : INJECTION_POINT_CACHED("datachecksumsworker-fail-db-result",
1263 : db->dbname);
1264 : }
1265 : #endif
1266 :
1267 22 : pgstat_progress_update_param(PROGRESS_DATACHECKSUMS_DBS_DONE,
1268 : ++cumulative_total);
1269 :
1270 22 : if (result == DATACHECKSUMSWORKER_FAILED)
1271 : {
1272 : /*
1273 : * Disable checksums on cluster, because we failed one of the
1274 : * databases and this is an all or nothing process.
1275 : */
1276 1 : SetDataChecksumsOff();
1277 1 : ereport(ERROR,
1278 : errcode(ERRCODE_INSUFFICIENT_RESOURCES),
1279 : errmsg("data checksums failed to get enabled in all databases, aborting"),
1280 : errhint("The server log might have more information on the cause of the error."));
1281 : }
1282 21 : else if (result == DATACHECKSUMSWORKER_ABORTED || abort_requested)
1283 : {
1284 : /* Abort flag set, so exit the whole process */
1285 0 : return false;
1286 : }
1287 :
1288 : /*
1289 : * When one database has completed, it will have done shared catalogs
1290 : * so we don't have to process them again.
1291 : */
1292 21 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
1293 21 : DataChecksumState->process_shared_catalogs = false;
1294 21 : LWLockRelease(DataChecksumsWorkerLock);
1295 : }
1296 :
1297 7 : FreeDatabaseList(DatabaseList);
1298 :
1299 7 : pgstat_progress_update_param(PROGRESS_DATACHECKSUMS_PHASE,
1300 : PROGRESS_DATACHECKSUMS_PHASE_WAITING_BARRIER);
1301 7 : return true;
1302 : }
1303 :
1304 : /*
1305 : * DataChecksumsShmemRequest
1306 : * Request datachecksumsworker-related shared memory
1307 : */
1308 : static void
1309 1255 : DataChecksumsShmemRequest(void *arg)
1310 : {
1311 1255 : ShmemRequestStruct(.name = "DataChecksumsWorker Data",
1312 : .size = sizeof(DataChecksumsStateStruct),
1313 : .ptr = (void **) &DataChecksumState,
1314 : );
1315 1255 : }
1316 :
1317 : /*
1318 : * DatabaseExists
1319 : *
1320 : * Scans the system catalog to check if a database with the given Oid exists
1321 : * and returns true if it is found and valid, else false. Note, we cannot use
1322 : * database_is_invalid_oid here as it will ERROR out, and we want to gracefully
1323 : * handle errors.
1324 : */
1325 : static bool
1326 0 : DatabaseExists(Oid dboid)
1327 : {
1328 : Relation rel;
1329 : ScanKeyData skey;
1330 : SysScanDesc scan;
1331 : bool found;
1332 : HeapTuple tuple;
1333 : Form_pg_database pg_database_tuple;
1334 :
1335 0 : StartTransactionCommand();
1336 :
1337 0 : rel = table_open(DatabaseRelationId, AccessShareLock);
1338 0 : ScanKeyInit(&skey,
1339 : Anum_pg_database_oid,
1340 : BTEqualStrategyNumber, F_OIDEQ,
1341 : ObjectIdGetDatum(dboid));
1342 0 : scan = systable_beginscan(rel, DatabaseOidIndexId, true, SnapshotSelf,
1343 : 1, &skey);
1344 0 : tuple = systable_getnext(scan);
1345 0 : found = HeapTupleIsValid(tuple);
1346 :
1347 : /* If the Oid exists, ensure that it's not partially dropped */
1348 0 : if (found)
1349 : {
1350 0 : pg_database_tuple = (Form_pg_database) GETSTRUCT(tuple);
1351 0 : if (database_is_invalid_form(pg_database_tuple))
1352 0 : found = false;
1353 : }
1354 :
1355 0 : systable_endscan(scan);
1356 0 : table_close(rel, AccessShareLock);
1357 :
1358 0 : CommitTransactionCommand();
1359 :
1360 0 : return found;
1361 : }
1362 :
1363 : /*
1364 : * BuildDatabaseList
1365 : * Compile a list of all currently available databases in the cluster
1366 : *
1367 : * This creates the list of databases for the datachecksumsworker workers to
1368 : * add checksums to. If the caller wants to ensure that no concurrently
1369 : * running CREATE DATABASE calls exist, this needs to be preceded by a call
1370 : * to WaitForAllTransactionsToFinish().
1371 : */
1372 : static List *
1373 9 : BuildDatabaseList(void)
1374 : {
1375 9 : List *DatabaseList = NIL;
1376 : Relation rel;
1377 : TableScanDesc scan;
1378 : HeapTuple tup;
1379 9 : MemoryContext ctx = CurrentMemoryContext;
1380 : MemoryContext oldctx;
1381 :
1382 9 : StartTransactionCommand();
1383 :
1384 9 : rel = table_open(DatabaseRelationId, AccessShareLock);
1385 9 : scan = table_beginscan_catalog(rel, 0, NULL);
1386 :
1387 36 : while (HeapTupleIsValid(tup = heap_getnext(scan, ForwardScanDirection)))
1388 : {
1389 27 : Form_pg_database pgdb = (Form_pg_database) GETSTRUCT(tup);
1390 : DataChecksumsWorkerDatabase *db;
1391 :
1392 27 : oldctx = MemoryContextSwitchTo(ctx);
1393 :
1394 27 : db = (DataChecksumsWorkerDatabase *) palloc0(sizeof(DataChecksumsWorkerDatabase));
1395 :
1396 27 : db->dboid = pgdb->oid;
1397 27 : db->dbname = pstrdup(NameStr(pgdb->datname));
1398 :
1399 27 : DatabaseList = lappend(DatabaseList, db);
1400 :
1401 27 : MemoryContextSwitchTo(oldctx);
1402 : }
1403 :
1404 9 : table_endscan(scan);
1405 9 : table_close(rel, AccessShareLock);
1406 :
1407 9 : CommitTransactionCommand();
1408 :
1409 9 : return DatabaseList;
1410 : }
1411 :
1412 : static void
1413 7 : FreeDatabaseList(List *dblist)
1414 : {
1415 7 : if (!dblist)
1416 0 : return;
1417 :
1418 35 : foreach_ptr(DataChecksumsWorkerDatabase, db, dblist)
1419 : {
1420 21 : if (db->dbname != NULL)
1421 21 : pfree(db->dbname);
1422 : }
1423 :
1424 7 : list_free_deep(dblist);
1425 : }
1426 :
1427 : /*
1428 : * BuildRelationList
1429 : * Compile a list of relations in the database
1430 : *
1431 : * Returns a list of OIDs for the requested relation types. If temp_relations
1432 : * is True then only temporary relations are returned. If temp_relations is
1433 : * False then non-temporary relations which have data checksums are returned.
1434 : * If include_shared is True then shared relations are included as well in a
1435 : * non-temporary list. include_shared has no relevance when building a list of
1436 : * temporary relations.
1437 : */
1438 : static List *
1439 69 : BuildRelationList(bool temp_relations, bool include_shared)
1440 : {
1441 69 : List *RelationList = NIL;
1442 : Relation rel;
1443 : TableScanDesc scan;
1444 : HeapTuple tup;
1445 69 : MemoryContext ctx = CurrentMemoryContext;
1446 : MemoryContext oldctx;
1447 :
1448 69 : StartTransactionCommand();
1449 :
1450 69 : rel = table_open(RelationRelationId, AccessShareLock);
1451 69 : scan = table_beginscan_catalog(rel, 0, NULL);
1452 :
1453 31302 : while (HeapTupleIsValid(tup = heap_getnext(scan, ForwardScanDirection)))
1454 : {
1455 31233 : Form_pg_class pgc = (Form_pg_class) GETSTRUCT(tup);
1456 :
1457 : /* Only include temporary relations when explicitly asked to */
1458 31233 : if (pgc->relpersistence == RELPERSISTENCE_TEMP)
1459 : {
1460 3 : if (!temp_relations)
1461 1 : continue;
1462 : }
1463 : else
1464 : {
1465 : /*
1466 : * If we are only interested in temp relations then continue
1467 : * immediately as the current relation isn't a temp relation.
1468 : */
1469 31230 : if (temp_relations)
1470 20820 : continue;
1471 :
1472 10410 : if (!RELKIND_HAS_STORAGE(pgc->relkind))
1473 3726 : continue;
1474 :
1475 6684 : if (pgc->relisshared && !include_shared)
1476 644 : continue;
1477 : }
1478 :
1479 6042 : oldctx = MemoryContextSwitchTo(ctx);
1480 6042 : RelationList = lappend_oid(RelationList, pgc->oid);
1481 6042 : MemoryContextSwitchTo(oldctx);
1482 : }
1483 :
1484 69 : table_endscan(scan);
1485 69 : table_close(rel, AccessShareLock);
1486 :
1487 69 : CommitTransactionCommand();
1488 :
1489 69 : return RelationList;
1490 : }
1491 :
1492 : /*
1493 : * DataChecksumsWorkerMain
1494 : *
1495 : * Main function for enabling checksums in a single database. This is the
1496 : * function set as the bgw_function_name in the dynamic background worker
1497 : * process initiated for each database by the worker launcher. After enabling
1498 : * data checksums in each applicable relation in the database, it will wait for
1499 : * all temporary relations that were present when the function started to
1500 : * disappear before returning. This is required since we cannot rewrite
1501 : * existing temporary relations with data checksums.
1502 : */
1503 : void
1504 23 : DataChecksumsWorkerMain(Datum arg)
1505 : {
1506 23 : Oid dboid = DatumGetObjectId(arg);
1507 23 : List *RelationList = NIL;
1508 23 : List *InitialTempTableList = NIL;
1509 : BufferAccessStrategy strategy;
1510 23 : bool aborted = false;
1511 : int64 rels_done;
1512 : #ifdef USE_INJECTION_POINTS
1513 23 : bool retried = false;
1514 : #endif
1515 :
1516 23 : operation = ENABLE_DATACHECKSUMS;
1517 :
1518 23 : pqsignal(SIGTERM, die);
1519 23 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
1520 :
1521 23 : BackgroundWorkerUnblockSignals();
1522 :
1523 23 : MyBackendType = B_DATACHECKSUMSWORKER_WORKER;
1524 23 : init_ps_display(NULL);
1525 :
1526 23 : BackgroundWorkerInitializeConnectionByOid(dboid, InvalidOid,
1527 : BGWORKER_BYPASS_ALLOWCONN);
1528 :
1529 : /* worker will have a separate entry in pg_stat_progress_data_checksums */
1530 23 : pgstat_progress_start_command(PROGRESS_COMMAND_DATACHECKSUMS,
1531 : InvalidOid);
1532 :
1533 : /*
1534 : * Get a list of all temp tables present as we start in this database. We
1535 : * need to wait until they are all gone until we are done, since we cannot
1536 : * access these relations and modify them.
1537 : */
1538 23 : InitialTempTableList = BuildRelationList(true, false);
1539 :
1540 : /*
1541 : * Enable vacuum cost delay, if any. While this process isn't doing any
1542 : * vacuuming, we are re-using the infrastructure that vacuum cost delay
1543 : * provides rather than inventing something bespoke. This is an internal
1544 : * implementation detail and care should be taken to avoid it bleeding
1545 : * through to the user to avoid confusion.
1546 : *
1547 : * VacuumUpdateCosts() propagates the values to the variables actually
1548 : * read by vacuum_delay_point().
1549 : */
1550 23 : VacuumCostDelay = DataChecksumState->cost_delay;
1551 23 : VacuumCostLimit = DataChecksumState->cost_limit;
1552 23 : VacuumUpdateCosts();
1553 23 : VacuumCostBalance = 0;
1554 :
1555 : /*
1556 : * Create and set the vacuum strategy as our buffer strategy.
1557 : */
1558 23 : strategy = GetAccessStrategy(BAS_VACUUM);
1559 :
1560 23 : RelationList = BuildRelationList(false,
1561 23 : DataChecksumState->process_shared_catalogs);
1562 :
1563 : /* Update the total number of relations to be processed in this DB. */
1564 : {
1565 23 : const int index[] = {
1566 : PROGRESS_DATACHECKSUMS_RELS_TOTAL,
1567 : PROGRESS_DATACHECKSUMS_RELS_DONE
1568 : };
1569 :
1570 : int64 vals[2];
1571 :
1572 23 : vals[0] = list_length(RelationList);
1573 23 : vals[1] = 0;
1574 :
1575 23 : pgstat_progress_update_multi_param(2, index, vals);
1576 : }
1577 :
1578 : /* Process the relations */
1579 23 : rels_done = 0;
1580 6086 : foreach_oid(reloid, RelationList)
1581 : {
1582 6040 : bool costs_updated = false;
1583 :
1584 6040 : if (!ProcessSingleRelationByOid(reloid, strategy))
1585 : {
1586 0 : aborted = true;
1587 0 : break;
1588 : }
1589 :
1590 6040 : pgstat_progress_update_param(PROGRESS_DATACHECKSUMS_RELS_DONE,
1591 : ++rels_done);
1592 6040 : CHECK_FOR_INTERRUPTS();
1593 6040 : CHECK_FOR_ABORT_REQUEST();
1594 :
1595 6040 : if (abort_requested)
1596 0 : break;
1597 :
1598 : /*
1599 : * Check if the cost settings changed during runtime and if so, update
1600 : * to reflect the new values and signal that the access strategy needs
1601 : * to be refreshed.
1602 : */
1603 6040 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
1604 6040 : if ((DataChecksumState->launch_cost_delay != DataChecksumState->cost_delay)
1605 6040 : || (DataChecksumState->launch_cost_limit != DataChecksumState->cost_limit))
1606 : {
1607 0 : costs_updated = true;
1608 0 : VacuumCostDelay = DataChecksumState->launch_cost_delay;
1609 0 : VacuumCostLimit = DataChecksumState->launch_cost_limit;
1610 0 : VacuumUpdateCosts();
1611 :
1612 0 : DataChecksumState->cost_delay = DataChecksumState->launch_cost_delay;
1613 0 : DataChecksumState->cost_limit = DataChecksumState->launch_cost_limit;
1614 : }
1615 : else
1616 6040 : costs_updated = false;
1617 6040 : LWLockRelease(DataChecksumsWorkerLock);
1618 :
1619 6040 : if (costs_updated)
1620 : {
1621 0 : FreeAccessStrategy(strategy);
1622 0 : strategy = GetAccessStrategy(BAS_VACUUM);
1623 : }
1624 : }
1625 :
1626 23 : list_free(RelationList);
1627 23 : FreeAccessStrategy(strategy);
1628 :
1629 23 : if (aborted || abort_requested)
1630 : {
1631 0 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
1632 0 : DataChecksumState->success = DATACHECKSUMSWORKER_ABORTED;
1633 0 : LWLockRelease(DataChecksumsWorkerLock);
1634 0 : ereport(DEBUG1,
1635 : errmsg("data checksum processing aborted in database OID %u",
1636 : dboid));
1637 0 : return;
1638 : }
1639 :
1640 : /* The worker is about to wait for temporary tables to go away. */
1641 23 : pgstat_progress_update_param(PROGRESS_DATACHECKSUMS_PHASE,
1642 : PROGRESS_DATACHECKSUMS_PHASE_WAITING_TEMPREL);
1643 :
1644 : /*
1645 : * Wait for all temp tables that existed when we started to go away. This
1646 : * is necessary since we cannot "reach" them to enable checksums. Any temp
1647 : * tables created after we started will already have checksums in them
1648 : * (due to the "inprogress-on" state), so no need to wait for those.
1649 : */
1650 : for (;;)
1651 0 : {
1652 : List *CurrentTempTables;
1653 : int numleft;
1654 : char activity[64];
1655 :
1656 23 : CurrentTempTables = BuildRelationList(true, false);
1657 23 : numleft = 0;
1658 47 : foreach_oid(tmptbloid, InitialTempTableList)
1659 : {
1660 1 : if (list_member_oid(CurrentTempTables, tmptbloid))
1661 1 : numleft++;
1662 : }
1663 23 : list_free(CurrentTempTables);
1664 :
1665 : #ifdef USE_INJECTION_POINTS
1666 23 : if (IS_INJECTION_POINT_ATTACHED("datachecksumsworker-fake-temptable-wait"))
1667 : {
1668 : /* Make sure to just cause one retry */
1669 0 : if (!retried && numleft == 0)
1670 : {
1671 0 : numleft = 1;
1672 0 : retried = true;
1673 :
1674 0 : INJECTION_POINT_CACHED("datachecksumsworker-fake-temptable-wait", NULL);
1675 : }
1676 : }
1677 : #endif
1678 :
1679 23 : if (numleft == 0)
1680 22 : break;
1681 :
1682 : /*
1683 : * At least one temp table is left to wait for, indicate in pgstat
1684 : * activity and progress reporting.
1685 : */
1686 1 : snprintf(activity,
1687 : sizeof(activity),
1688 : "Waiting for %d temp tables to be removed", numleft);
1689 1 : pgstat_report_activity(STATE_RUNNING, activity);
1690 :
1691 : /* Retry every 3 seconds */
1692 1 : ResetLatch(MyLatch);
1693 1 : (void) WaitLatch(MyLatch,
1694 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
1695 : 3000,
1696 : WAIT_EVENT_CHECKSUM_ENABLE_TEMPTABLE_WAIT);
1697 :
1698 1 : CHECK_FOR_INTERRUPTS();
1699 0 : CHECK_FOR_ABORT_REQUEST();
1700 :
1701 0 : if (aborted || abort_requested)
1702 : {
1703 0 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
1704 0 : DataChecksumState->success = DATACHECKSUMSWORKER_ABORTED;
1705 0 : LWLockRelease(DataChecksumsWorkerLock);
1706 0 : ereport(LOG,
1707 : errmsg("data checksum processing aborted in database OID %u",
1708 : dboid));
1709 0 : return;
1710 : }
1711 : }
1712 :
1713 22 : list_free(InitialTempTableList);
1714 :
1715 : /* worker done */
1716 22 : pgstat_progress_end_command();
1717 :
1718 22 : LWLockAcquire(DataChecksumsWorkerLock, LW_EXCLUSIVE);
1719 22 : DataChecksumState->success = DATACHECKSUMSWORKER_SUCCESSFUL;
1720 22 : LWLockRelease(DataChecksumsWorkerLock);
1721 : }
|
