| 1 | /*------------------------------------------------------------------------- |
| 2 | * |
| 3 | * inval.c |
| 4 | * POSTGRES cache invalidation dispatcher code. |
| 5 | * |
| 6 | * This is subtle stuff, so pay attention: |
| 7 | * |
| 8 | * When a tuple is updated or deleted, our standard visibility rules |
| 9 | * consider that it is *still valid* so long as we are in the same command, |
| 10 | * ie, until the next CommandCounterIncrement() or transaction commit. |
| 11 | * (See access/heap/heapam_visibility.c, and note that system catalogs are |
| 12 | * generally scanned under the most current snapshot available, rather than |
| 13 | * the transaction snapshot.) At the command boundary, the old tuple stops |
| 14 | * being valid and the new version, if any, becomes valid. Therefore, |
| 15 | * we cannot simply flush a tuple from the system caches during heap_update() |
| 16 | * or heap_delete(). The tuple is still good at that point; what's more, |
| 17 | * even if we did flush it, it might be reloaded into the caches by a later |
| 18 | * request in the same command. So the correct behavior is to keep a list |
| 19 | * of outdated (updated/deleted) tuples and then do the required cache |
| 20 | * flushes at the next command boundary. We must also keep track of |
| 21 | * inserted tuples so that we can flush "negative" cache entries that match |
| 22 | * the new tuples; again, that mustn't happen until end of command. |
| 23 | * |
| 24 | * Once we have finished the command, we still need to remember inserted |
| 25 | * tuples (including new versions of updated tuples), so that we can flush |
| 26 | * them from the caches if we abort the transaction. Similarly, we'd better |
| 27 | * be able to flush "negative" cache entries that may have been loaded in |
| 28 | * place of deleted tuples, so we still need the deleted ones too. |
| 29 | * |
| 30 | * If we successfully complete the transaction, we have to broadcast all |
| 31 | * these invalidation events to other backends (via the SI message queue) |
| 32 | * so that they can flush obsolete entries from their caches. Note we have |
| 33 | * to record the transaction commit before sending SI messages, otherwise |
| 34 | * the other backends won't see our updated tuples as good. |
| 35 | * |
| 36 | * When a subtransaction aborts, we can process and discard any events |
| 37 | * it has queued. When a subtransaction commits, we just add its events |
| 38 | * to the pending lists of the parent transaction. |
| 39 | * |
| 40 | * In short, we need to remember until xact end every insert or delete |
| 41 | * of a tuple that might be in the system caches. Updates are treated as |
| 42 | * two events, delete + insert, for simplicity. (If the update doesn't |
| 43 | * change the tuple hash value, catcache.c optimizes this into one event.) |
| 44 | * |
| 45 | * We do not need to register EVERY tuple operation in this way, just those |
| 46 | * on tuples in relations that have associated catcaches. We do, however, |
| 47 | * have to register every operation on every tuple that *could* be in a |
| 48 | * catcache, whether or not it currently is in our cache. Also, if the |
| 49 | * tuple is in a relation that has multiple catcaches, we need to register |
| 50 | * an invalidation message for each such catcache. catcache.c's |
| 51 | * PrepareToInvalidateCacheTuple() routine provides the knowledge of which |
| 52 | * catcaches may need invalidation for a given tuple. |
| 53 | * |
| 54 | * Also, whenever we see an operation on a pg_class, pg_attribute, or |
| 55 | * pg_index tuple, we register a relcache flush operation for the relation |
| 56 | * described by that tuple (as specified in CacheInvalidateHeapTuple()). |
| 57 | * Likewise for pg_constraint tuples for foreign keys on relations. |
| 58 | * |
| 59 | * We keep the relcache flush requests in lists separate from the catcache |
| 60 | * tuple flush requests. This allows us to issue all the pending catcache |
| 61 | * flushes before we issue relcache flushes, which saves us from loading |
| 62 | * a catcache tuple during relcache load only to flush it again right away. |
| 63 | * Also, we avoid queuing multiple relcache flush requests for the same |
| 64 | * relation, since a relcache flush is relatively expensive to do. |
| 65 | * (XXX is it worth testing likewise for duplicate catcache flush entries? |
| 66 | * Probably not.) |
| 67 | * |
| 68 | * If a relcache flush is issued for a system relation that we preload |
| 69 | * from the relcache init file, we must also delete the init file so that |
| 70 | * it will be rebuilt during the next backend restart. The actual work of |
| 71 | * manipulating the init file is in relcache.c, but we keep track of the |
| 72 | * need for it here. |
| 73 | * |
| 74 | * The request lists proper are kept in CurTransactionContext of their |
| 75 | * creating (sub)transaction, since they can be forgotten on abort of that |
| 76 | * transaction but must be kept till top-level commit otherwise. For |
| 77 | * simplicity we keep the controlling list-of-lists in TopTransactionContext. |
| 78 | * |
| 79 | * Currently, inval messages are sent without regard for the possibility |
| 80 | * that the object described by the catalog tuple might be a session-local |
| 81 | * object such as a temporary table. This is because (1) this code has |
| 82 | * no practical way to tell the difference, and (2) it is not certain that |
| 83 | * other backends don't have catalog cache or even relcache entries for |
| 84 | * such tables, anyway; there is nothing that prevents that. It might be |
| 85 | * worth trying to avoid sending such inval traffic in the future, if those |
| 86 | * problems can be overcome cheaply. |
| 87 | * |
| 88 | * |
| 89 | * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group |
| 90 | * Portions Copyright (c) 1994, Regents of the University of California |
| 91 | * |
| 92 | * IDENTIFICATION |
| 93 | * src/backend/utils/cache/inval.c |
| 94 | * |
| 95 | *------------------------------------------------------------------------- |
| 96 | */ |
| 97 | #include "postgres.h" |
| 98 | |
| 99 | #include <limits.h> |
| 100 | |
| 101 | #include "access/htup_details.h" |
| 102 | #include "access/xact.h" |
| 103 | #include "catalog/catalog.h" |
| 104 | #include "catalog/pg_constraint.h" |
| 105 | #include "miscadmin.h" |
| 106 | #include "storage/sinval.h" |
| 107 | #include "storage/smgr.h" |
| 108 | #include "utils/catcache.h" |
| 109 | #include "utils/inval.h" |
| 110 | #include "utils/memdebug.h" |
| 111 | #include "utils/memutils.h" |
| 112 | #include "utils/rel.h" |
| 113 | #include "utils/relmapper.h" |
| 114 | #include "utils/snapmgr.h" |
| 115 | #include "utils/syscache.h" |
| 116 | |
| 117 | |
| 118 | /* |
| 119 | * To minimize palloc traffic, we keep pending requests in successively- |
| 120 | * larger chunks (a slightly more sophisticated version of an expansible |
| 121 | * array). All request types can be stored as SharedInvalidationMessage |
| 122 | * records. The ordering of requests within a list is never significant. |
| 123 | */ |
| 124 | typedef struct InvalidationChunk |
| 125 | { |
| 126 | struct InvalidationChunk *next; /* list link */ |
| 127 | int nitems; /* # items currently stored in chunk */ |
| 128 | int maxitems; /* size of allocated array in this chunk */ |
| 129 | SharedInvalidationMessage msgs[FLEXIBLE_ARRAY_MEMBER]; |
| 130 | } InvalidationChunk; |
| 131 | |
| 132 | typedef struct |
| 133 | { |
| 134 | InvalidationChunk *; /* list of chunks holding catcache msgs */ |
| 135 | InvalidationChunk *; /* list of chunks holding relcache msgs */ |
| 136 | } ; |
| 137 | |
| 138 | /*---------------- |
| 139 | * Invalidation info is divided into two lists: |
| 140 | * 1) events so far in current command, not yet reflected to caches. |
| 141 | * 2) events in previous commands of current transaction; these have |
| 142 | * been reflected to local caches, and must be either broadcast to |
| 143 | * other backends or rolled back from local cache when we commit |
| 144 | * or abort the transaction. |
| 145 | * Actually, we need two such lists for each level of nested transaction, |
| 146 | * so that we can discard events from an aborted subtransaction. When |
| 147 | * a subtransaction commits, we append its lists to the parent's lists. |
| 148 | * |
| 149 | * The relcache-file-invalidated flag can just be a simple boolean, |
| 150 | * since we only act on it at transaction commit; we don't care which |
| 151 | * command of the transaction set it. |
| 152 | *---------------- |
| 153 | */ |
| 154 | |
| 155 | typedef struct TransInvalidationInfo |
| 156 | { |
| 157 | /* Back link to parent transaction's info */ |
| 158 | struct TransInvalidationInfo *parent; |
| 159 | |
| 160 | /* Subtransaction nesting depth */ |
| 161 | int my_level; |
| 162 | |
| 163 | /* head of current-command event list */ |
| 164 | InvalidationListHeader CurrentCmdInvalidMsgs; |
| 165 | |
| 166 | /* head of previous-commands event list */ |
| 167 | InvalidationListHeader PriorCmdInvalidMsgs; |
| 168 | |
| 169 | /* init file must be invalidated? */ |
| 170 | bool RelcacheInitFileInval; |
| 171 | } TransInvalidationInfo; |
| 172 | |
| 173 | static TransInvalidationInfo *transInvalInfo = NULL; |
| 174 | |
| 175 | static SharedInvalidationMessage *SharedInvalidMessagesArray; |
| 176 | static int numSharedInvalidMessagesArray; |
| 177 | static int maxSharedInvalidMessagesArray; |
| 178 | |
| 179 | |
| 180 | /* |
| 181 | * Dynamically-registered callback functions. Current implementation |
| 182 | * assumes there won't be enough of these to justify a dynamically resizable |
| 183 | * array; it'd be easy to improve that if needed. |
| 184 | * |
| 185 | * To avoid searching in CallSyscacheCallbacks, all callbacks for a given |
| 186 | * syscache are linked into a list pointed to by syscache_callback_links[id]. |
| 187 | * The link values are syscache_callback_list[] index plus 1, or 0 for none. |
| 188 | */ |
| 189 | |
| 190 | #define MAX_SYSCACHE_CALLBACKS 64 |
| 191 | #define MAX_RELCACHE_CALLBACKS 10 |
| 192 | |
| 193 | static struct SYSCACHECALLBACK |
| 194 | { |
| 195 | int16 id; /* cache number */ |
| 196 | int16 link; /* next callback index+1 for same cache */ |
| 197 | SyscacheCallbackFunction function; |
| 198 | Datum arg; |
| 199 | } syscache_callback_list[MAX_SYSCACHE_CALLBACKS]; |
| 200 | |
| 201 | static int16 syscache_callback_links[SysCacheSize]; |
| 202 | |
| 203 | static int syscache_callback_count = 0; |
| 204 | |
| 205 | static struct RELCACHECALLBACK |
| 206 | { |
| 207 | RelcacheCallbackFunction function; |
| 208 | Datum arg; |
| 209 | } relcache_callback_list[MAX_RELCACHE_CALLBACKS]; |
| 210 | |
| 211 | static int relcache_callback_count = 0; |
| 212 | |
| 213 | /* ---------------------------------------------------------------- |
| 214 | * Invalidation list support functions |
| 215 | * |
| 216 | * These three routines encapsulate processing of the "chunked" |
| 217 | * representation of what is logically just a list of messages. |
| 218 | * ---------------------------------------------------------------- |
| 219 | */ |
| 220 | |
| 221 | /* |
| 222 | * AddInvalidationMessage |
| 223 | * Add an invalidation message to a list (of chunks). |
| 224 | * |
| 225 | * Note that we do not pay any great attention to maintaining the original |
| 226 | * ordering of the messages. |
| 227 | */ |
| 228 | static void |
| 229 | AddInvalidationMessage(InvalidationChunk **listHdr, |
| 230 | SharedInvalidationMessage *msg) |
| 231 | { |
| 232 | InvalidationChunk *chunk = *listHdr; |
| 233 | |
| 234 | if (chunk == NULL) |
| 235 | { |
| 236 | /* First time through; create initial chunk */ |
| 237 | #define FIRSTCHUNKSIZE 32 |
| 238 | chunk = (InvalidationChunk *) |
| 239 | MemoryContextAlloc(CurTransactionContext, |
| 240 | offsetof(InvalidationChunk, msgs) + |
| 241 | FIRSTCHUNKSIZE * sizeof(SharedInvalidationMessage)); |
| 242 | chunk->nitems = 0; |
| 243 | chunk->maxitems = FIRSTCHUNKSIZE; |
| 244 | chunk->next = *listHdr; |
| 245 | *listHdr = chunk; |
| 246 | } |
| 247 | else if (chunk->nitems >= chunk->maxitems) |
| 248 | { |
| 249 | /* Need another chunk; double size of last chunk */ |
| 250 | int chunksize = 2 * chunk->maxitems; |
| 251 | |
| 252 | chunk = (InvalidationChunk *) |
| 253 | MemoryContextAlloc(CurTransactionContext, |
| 254 | offsetof(InvalidationChunk, msgs) + |
| 255 | chunksize * sizeof(SharedInvalidationMessage)); |
| 256 | chunk->nitems = 0; |
| 257 | chunk->maxitems = chunksize; |
| 258 | chunk->next = *listHdr; |
| 259 | *listHdr = chunk; |
| 260 | } |
| 261 | /* Okay, add message to current chunk */ |
| 262 | chunk->msgs[chunk->nitems] = *msg; |
| 263 | chunk->nitems++; |
| 264 | } |
| 265 | |
| 266 | /* |
| 267 | * Append one list of invalidation message chunks to another, resetting |
| 268 | * the source chunk-list pointer to NULL. |
| 269 | */ |
| 270 | static void |
| 271 | AppendInvalidationMessageList(InvalidationChunk **destHdr, |
| 272 | InvalidationChunk **srcHdr) |
| 273 | { |
| 274 | InvalidationChunk *chunk = *srcHdr; |
| 275 | |
| 276 | if (chunk == NULL) |
| 277 | return; /* nothing to do */ |
| 278 | |
| 279 | while (chunk->next != NULL) |
| 280 | chunk = chunk->next; |
| 281 | |
| 282 | chunk->next = *destHdr; |
| 283 | |
| 284 | *destHdr = *srcHdr; |
| 285 | |
| 286 | *srcHdr = NULL; |
| 287 | } |
| 288 | |
| 289 | /* |
| 290 | * Process a list of invalidation messages. |
| 291 | * |
| 292 | * This is a macro that executes the given code fragment for each message in |
| 293 | * a message chunk list. The fragment should refer to the message as *msg. |
| 294 | */ |
| 295 | #define ProcessMessageList(listHdr, codeFragment) \ |
| 296 | do { \ |
| 297 | InvalidationChunk *_chunk; \ |
| 298 | for (_chunk = (listHdr); _chunk != NULL; _chunk = _chunk->next) \ |
| 299 | { \ |
| 300 | int _cindex; \ |
| 301 | for (_cindex = 0; _cindex < _chunk->nitems; _cindex++) \ |
| 302 | { \ |
| 303 | SharedInvalidationMessage *msg = &_chunk->msgs[_cindex]; \ |
| 304 | codeFragment; \ |
| 305 | } \ |
| 306 | } \ |
| 307 | } while (0) |
| 308 | |
| 309 | /* |
| 310 | * Process a list of invalidation messages group-wise. |
| 311 | * |
| 312 | * As above, but the code fragment can handle an array of messages. |
| 313 | * The fragment should refer to the messages as msgs[], with n entries. |
| 314 | */ |
| 315 | #define ProcessMessageListMulti(listHdr, codeFragment) \ |
| 316 | do { \ |
| 317 | InvalidationChunk *_chunk; \ |
| 318 | for (_chunk = (listHdr); _chunk != NULL; _chunk = _chunk->next) \ |
| 319 | { \ |
| 320 | SharedInvalidationMessage *msgs = _chunk->msgs; \ |
| 321 | int n = _chunk->nitems; \ |
| 322 | codeFragment; \ |
| 323 | } \ |
| 324 | } while (0) |
| 325 | |
| 326 | |
| 327 | /* ---------------------------------------------------------------- |
| 328 | * Invalidation set support functions |
| 329 | * |
| 330 | * These routines understand about the division of a logical invalidation |
| 331 | * list into separate physical lists for catcache and relcache entries. |
| 332 | * ---------------------------------------------------------------- |
| 333 | */ |
| 334 | |
| 335 | /* |
| 336 | * Add a catcache inval entry |
| 337 | */ |
| 338 | static void |
| 339 | AddCatcacheInvalidationMessage(InvalidationListHeader *hdr, |
| 340 | int id, uint32 hashValue, Oid dbId) |
| 341 | { |
| 342 | SharedInvalidationMessage msg; |
| 343 | |
| 344 | Assert(id < CHAR_MAX); |
| 345 | msg.cc.id = (int8) id; |
| 346 | msg.cc.dbId = dbId; |
| 347 | msg.cc.hashValue = hashValue; |
| 348 | |
| 349 | /* |
| 350 | * Define padding bytes in SharedInvalidationMessage structs to be |
| 351 | * defined. Otherwise the sinvaladt.c ringbuffer, which is accessed by |
| 352 | * multiple processes, will cause spurious valgrind warnings about |
| 353 | * undefined memory being used. That's because valgrind remembers the |
| 354 | * undefined bytes from the last local process's store, not realizing that |
| 355 | * another process has written since, filling the previously uninitialized |
| 356 | * bytes |
| 357 | */ |
| 358 | VALGRIND_MAKE_MEM_DEFINED(&msg, sizeof(msg)); |
| 359 | |
| 360 | AddInvalidationMessage(&hdr->cclist, &msg); |
| 361 | } |
| 362 | |
| 363 | /* |
| 364 | * Add a whole-catalog inval entry |
| 365 | */ |
| 366 | static void |
| 367 | AddCatalogInvalidationMessage(InvalidationListHeader *hdr, |
| 368 | Oid dbId, Oid catId) |
| 369 | { |
| 370 | SharedInvalidationMessage msg; |
| 371 | |
| 372 | msg.cat.id = SHAREDINVALCATALOG_ID; |
| 373 | msg.cat.dbId = dbId; |
| 374 | msg.cat.catId = catId; |
| 375 | /* check AddCatcacheInvalidationMessage() for an explanation */ |
| 376 | VALGRIND_MAKE_MEM_DEFINED(&msg, sizeof(msg)); |
| 377 | |
| 378 | AddInvalidationMessage(&hdr->cclist, &msg); |
| 379 | } |
| 380 | |
| 381 | /* |
| 382 | * Add a relcache inval entry |
| 383 | */ |
| 384 | static void |
| 385 | AddRelcacheInvalidationMessage(InvalidationListHeader *hdr, |
| 386 | Oid dbId, Oid relId) |
| 387 | { |
| 388 | SharedInvalidationMessage msg; |
| 389 | |
| 390 | /* |
| 391 | * Don't add a duplicate item. We assume dbId need not be checked because |
| 392 | * it will never change. InvalidOid for relId means all relations so we |
| 393 | * don't need to add individual ones when it is present. |
| 394 | */ |
| 395 | ProcessMessageList(hdr->rclist, |
| 396 | if (msg->rc.id == SHAREDINVALRELCACHE_ID && |
| 397 | (msg->rc.relId == relId || |
| 398 | msg->rc.relId == InvalidOid)) |
| 399 | return); |
| 400 | |
| 401 | /* OK, add the item */ |
| 402 | msg.rc.id = SHAREDINVALRELCACHE_ID; |
| 403 | msg.rc.dbId = dbId; |
| 404 | msg.rc.relId = relId; |
| 405 | /* check AddCatcacheInvalidationMessage() for an explanation */ |
| 406 | VALGRIND_MAKE_MEM_DEFINED(&msg, sizeof(msg)); |
| 407 | |
| 408 | AddInvalidationMessage(&hdr->rclist, &msg); |
| 409 | } |
| 410 | |
| 411 | /* |
| 412 | * Add a snapshot inval entry |
| 413 | */ |
| 414 | static void |
| 415 | AddSnapshotInvalidationMessage(InvalidationListHeader *hdr, |
| 416 | Oid dbId, Oid relId) |
| 417 | { |
| 418 | SharedInvalidationMessage msg; |
| 419 | |
| 420 | /* Don't add a duplicate item */ |
| 421 | /* We assume dbId need not be checked because it will never change */ |
| 422 | ProcessMessageList(hdr->rclist, |
| 423 | if (msg->sn.id == SHAREDINVALSNAPSHOT_ID && |
| 424 | msg->sn.relId == relId) |
| 425 | return); |
| 426 | |
| 427 | /* OK, add the item */ |
| 428 | msg.sn.id = SHAREDINVALSNAPSHOT_ID; |
| 429 | msg.sn.dbId = dbId; |
| 430 | msg.sn.relId = relId; |
| 431 | /* check AddCatcacheInvalidationMessage() for an explanation */ |
| 432 | VALGRIND_MAKE_MEM_DEFINED(&msg, sizeof(msg)); |
| 433 | |
| 434 | AddInvalidationMessage(&hdr->rclist, &msg); |
| 435 | } |
| 436 | |
| 437 | /* |
| 438 | * Append one list of invalidation messages to another, resetting |
| 439 | * the source list to empty. |
| 440 | */ |
| 441 | static void |
| 442 | AppendInvalidationMessages(InvalidationListHeader *dest, |
| 443 | InvalidationListHeader *src) |
| 444 | { |
| 445 | AppendInvalidationMessageList(&dest->cclist, &src->cclist); |
| 446 | AppendInvalidationMessageList(&dest->rclist, &src->rclist); |
| 447 | } |
| 448 | |
| 449 | /* |
| 450 | * Execute the given function for all the messages in an invalidation list. |
| 451 | * The list is not altered. |
| 452 | * |
| 453 | * catcache entries are processed first, for reasons mentioned above. |
| 454 | */ |
| 455 | static void |
| 456 | ProcessInvalidationMessages(InvalidationListHeader *hdr, |
| 457 | void (*func) (SharedInvalidationMessage *msg)) |
| 458 | { |
| 459 | ProcessMessageList(hdr->cclist, func(msg)); |
| 460 | ProcessMessageList(hdr->rclist, func(msg)); |
| 461 | } |
| 462 | |
| 463 | /* |
| 464 | * As above, but the function is able to process an array of messages |
| 465 | * rather than just one at a time. |
| 466 | */ |
| 467 | static void |
| 468 | ProcessInvalidationMessagesMulti(InvalidationListHeader *hdr, |
| 469 | void (*func) (const SharedInvalidationMessage *msgs, int n)) |
| 470 | { |
| 471 | ProcessMessageListMulti(hdr->cclist, func(msgs, n)); |
| 472 | ProcessMessageListMulti(hdr->rclist, func(msgs, n)); |
| 473 | } |
| 474 | |
| 475 | /* ---------------------------------------------------------------- |
| 476 | * private support functions |
| 477 | * ---------------------------------------------------------------- |
| 478 | */ |
| 479 | |
| 480 | /* |
| 481 | * RegisterCatcacheInvalidation |
| 482 | * |
| 483 | * Register an invalidation event for a catcache tuple entry. |
| 484 | */ |
| 485 | static void |
| 486 | RegisterCatcacheInvalidation(int cacheId, |
| 487 | uint32 hashValue, |
| 488 | Oid dbId) |
| 489 | { |
| 490 | AddCatcacheInvalidationMessage(&transInvalInfo->CurrentCmdInvalidMsgs, |
| 491 | cacheId, hashValue, dbId); |
| 492 | } |
| 493 | |
| 494 | /* |
| 495 | * RegisterCatalogInvalidation |
| 496 | * |
| 497 | * Register an invalidation event for all catcache entries from a catalog. |
| 498 | */ |
| 499 | static void |
| 500 | RegisterCatalogInvalidation(Oid dbId, Oid catId) |
| 501 | { |
| 502 | AddCatalogInvalidationMessage(&transInvalInfo->CurrentCmdInvalidMsgs, |
| 503 | dbId, catId); |
| 504 | } |
| 505 | |
| 506 | /* |
| 507 | * RegisterRelcacheInvalidation |
| 508 | * |
| 509 | * As above, but register a relcache invalidation event. |
| 510 | */ |
| 511 | static void |
| 512 | RegisterRelcacheInvalidation(Oid dbId, Oid relId) |
| 513 | { |
| 514 | AddRelcacheInvalidationMessage(&transInvalInfo->CurrentCmdInvalidMsgs, |
| 515 | dbId, relId); |
| 516 | |
| 517 | /* |
| 518 | * Most of the time, relcache invalidation is associated with system |
| 519 | * catalog updates, but there are a few cases where it isn't. Quick hack |
| 520 | * to ensure that the next CommandCounterIncrement() will think that we |
| 521 | * need to do CommandEndInvalidationMessages(). |
| 522 | */ |
| 523 | (void) GetCurrentCommandId(true); |
| 524 | |
| 525 | /* |
| 526 | * If the relation being invalidated is one of those cached in a relcache |
| 527 | * init file, mark that we need to zap that file at commit. For simplicity |
| 528 | * invalidations for a specific database always invalidate the shared file |
| 529 | * as well. Also zap when we are invalidating whole relcache. |
| 530 | */ |
| 531 | if (relId == InvalidOid || RelationIdIsInInitFile(relId)) |
| 532 | transInvalInfo->RelcacheInitFileInval = true; |
| 533 | } |
| 534 | |
| 535 | /* |
| 536 | * RegisterSnapshotInvalidation |
| 537 | * |
| 538 | * Register an invalidation event for MVCC scans against a given catalog. |
| 539 | * Only needed for catalogs that don't have catcaches. |
| 540 | */ |
| 541 | static void |
| 542 | RegisterSnapshotInvalidation(Oid dbId, Oid relId) |
| 543 | { |
| 544 | AddSnapshotInvalidationMessage(&transInvalInfo->CurrentCmdInvalidMsgs, |
| 545 | dbId, relId); |
| 546 | } |
| 547 | |
| 548 | /* |
| 549 | * LocalExecuteInvalidationMessage |
| 550 | * |
| 551 | * Process a single invalidation message (which could be of any type). |
| 552 | * Only the local caches are flushed; this does not transmit the message |
| 553 | * to other backends. |
| 554 | */ |
| 555 | void |
| 556 | LocalExecuteInvalidationMessage(SharedInvalidationMessage *msg) |
| 557 | { |
| 558 | if (msg->id >= 0) |
| 559 | { |
| 560 | if (msg->cc.dbId == MyDatabaseId || msg->cc.dbId == InvalidOid) |
| 561 | { |
| 562 | InvalidateCatalogSnapshot(); |
| 563 | |
| 564 | SysCacheInvalidate(msg->cc.id, msg->cc.hashValue); |
| 565 | |
| 566 | CallSyscacheCallbacks(msg->cc.id, msg->cc.hashValue); |
| 567 | } |
| 568 | } |
| 569 | else if (msg->id == SHAREDINVALCATALOG_ID) |
| 570 | { |
| 571 | if (msg->cat.dbId == MyDatabaseId || msg->cat.dbId == InvalidOid) |
| 572 | { |
| 573 | InvalidateCatalogSnapshot(); |
| 574 | |
| 575 | CatalogCacheFlushCatalog(msg->cat.catId); |
| 576 | |
| 577 | /* CatalogCacheFlushCatalog calls CallSyscacheCallbacks as needed */ |
| 578 | } |
| 579 | } |
| 580 | else if (msg->id == SHAREDINVALRELCACHE_ID) |
| 581 | { |
| 582 | if (msg->rc.dbId == MyDatabaseId || msg->rc.dbId == InvalidOid) |
| 583 | { |
| 584 | int i; |
| 585 | |
| 586 | if (msg->rc.relId == InvalidOid) |
| 587 | RelationCacheInvalidate(); |
| 588 | else |
| 589 | RelationCacheInvalidateEntry(msg->rc.relId); |
| 590 | |
| 591 | for (i = 0; i < relcache_callback_count; i++) |
| 592 | { |
| 593 | struct RELCACHECALLBACK *ccitem = relcache_callback_list + i; |
| 594 | |
| 595 | ccitem->function(ccitem->arg, msg->rc.relId); |
| 596 | } |
| 597 | } |
| 598 | } |
| 599 | else if (msg->id == SHAREDINVALSMGR_ID) |
| 600 | { |
| 601 | /* |
| 602 | * We could have smgr entries for relations of other databases, so no |
| 603 | * short-circuit test is possible here. |
| 604 | */ |
| 605 | RelFileNodeBackend rnode; |
| 606 | |
| 607 | rnode.node = msg->sm.rnode; |
| 608 | rnode.backend = (msg->sm.backend_hi << 16) | (int) msg->sm.backend_lo; |
| 609 | smgrclosenode(rnode); |
| 610 | } |
| 611 | else if (msg->id == SHAREDINVALRELMAP_ID) |
| 612 | { |
| 613 | /* We only care about our own database and shared catalogs */ |
| 614 | if (msg->rm.dbId == InvalidOid) |
| 615 | RelationMapInvalidate(true); |
| 616 | else if (msg->rm.dbId == MyDatabaseId) |
| 617 | RelationMapInvalidate(false); |
| 618 | } |
| 619 | else if (msg->id == SHAREDINVALSNAPSHOT_ID) |
| 620 | { |
| 621 | /* We only care about our own database and shared catalogs */ |
| 622 | if (msg->rm.dbId == InvalidOid) |
| 623 | InvalidateCatalogSnapshot(); |
| 624 | else if (msg->rm.dbId == MyDatabaseId) |
| 625 | InvalidateCatalogSnapshot(); |
| 626 | } |
| 627 | else |
| 628 | elog(FATAL, "unrecognized SI message ID: %d" , msg->id); |
| 629 | } |
| 630 | |
| 631 | /* |
| 632 | * InvalidateSystemCaches |
| 633 | * |
| 634 | * This blows away all tuples in the system catalog caches and |
| 635 | * all the cached relation descriptors and smgr cache entries. |
| 636 | * Relation descriptors that have positive refcounts are then rebuilt. |
| 637 | * |
| 638 | * We call this when we see a shared-inval-queue overflow signal, |
| 639 | * since that tells us we've lost some shared-inval messages and hence |
| 640 | * don't know what needs to be invalidated. |
| 641 | */ |
| 642 | void |
| 643 | InvalidateSystemCaches(void) |
| 644 | { |
| 645 | int i; |
| 646 | |
| 647 | InvalidateCatalogSnapshot(); |
| 648 | ResetCatalogCaches(); |
| 649 | RelationCacheInvalidate(); /* gets smgr and relmap too */ |
| 650 | |
| 651 | for (i = 0; i < syscache_callback_count; i++) |
| 652 | { |
| 653 | struct SYSCACHECALLBACK *ccitem = syscache_callback_list + i; |
| 654 | |
| 655 | ccitem->function(ccitem->arg, ccitem->id, 0); |
| 656 | } |
| 657 | |
| 658 | for (i = 0; i < relcache_callback_count; i++) |
| 659 | { |
| 660 | struct RELCACHECALLBACK *ccitem = relcache_callback_list + i; |
| 661 | |
| 662 | ccitem->function(ccitem->arg, InvalidOid); |
| 663 | } |
| 664 | } |
| 665 | |
| 666 | |
| 667 | /* ---------------------------------------------------------------- |
| 668 | * public functions |
| 669 | * ---------------------------------------------------------------- |
| 670 | */ |
| 671 | |
| 672 | /* |
| 673 | * AcceptInvalidationMessages |
| 674 | * Read and process invalidation messages from the shared invalidation |
| 675 | * message queue. |
| 676 | * |
| 677 | * Note: |
| 678 | * This should be called as the first step in processing a transaction. |
| 679 | */ |
| 680 | void |
| 681 | AcceptInvalidationMessages(void) |
| 682 | { |
| 683 | ReceiveSharedInvalidMessages(LocalExecuteInvalidationMessage, |
| 684 | InvalidateSystemCaches); |
| 685 | |
| 686 | /* |
| 687 | * Test code to force cache flushes anytime a flush could happen. |
| 688 | * |
| 689 | * If used with CLOBBER_FREED_MEMORY, CLOBBER_CACHE_ALWAYS provides a |
| 690 | * fairly thorough test that the system contains no cache-flush hazards. |
| 691 | * However, it also makes the system unbelievably slow --- the regression |
| 692 | * tests take about 100 times longer than normal. |
| 693 | * |
| 694 | * If you're a glutton for punishment, try CLOBBER_CACHE_RECURSIVELY. This |
| 695 | * slows things by at least a factor of 10000, so I wouldn't suggest |
| 696 | * trying to run the entire regression tests that way. It's useful to try |
| 697 | * a few simple tests, to make sure that cache reload isn't subject to |
| 698 | * internal cache-flush hazards, but after you've done a few thousand |
| 699 | * recursive reloads it's unlikely you'll learn more. |
| 700 | */ |
| 701 | #if defined(CLOBBER_CACHE_ALWAYS) |
| 702 | { |
| 703 | static bool in_recursion = false; |
| 704 | |
| 705 | if (!in_recursion) |
| 706 | { |
| 707 | in_recursion = true; |
| 708 | InvalidateSystemCaches(); |
| 709 | in_recursion = false; |
| 710 | } |
| 711 | } |
| 712 | #elif defined(CLOBBER_CACHE_RECURSIVELY) |
| 713 | { |
| 714 | static int recursion_depth = 0; |
| 715 | |
| 716 | /* Maximum depth is arbitrary depending on your threshold of pain */ |
| 717 | if (recursion_depth < 3) |
| 718 | { |
| 719 | recursion_depth++; |
| 720 | InvalidateSystemCaches(); |
| 721 | recursion_depth--; |
| 722 | } |
| 723 | } |
| 724 | #endif |
| 725 | } |
| 726 | |
| 727 | /* |
| 728 | * PrepareInvalidationState |
| 729 | * Initialize inval lists for the current (sub)transaction. |
| 730 | */ |
| 731 | static void |
| 732 | PrepareInvalidationState(void) |
| 733 | { |
| 734 | TransInvalidationInfo *myInfo; |
| 735 | |
| 736 | if (transInvalInfo != NULL && |
| 737 | transInvalInfo->my_level == GetCurrentTransactionNestLevel()) |
| 738 | return; |
| 739 | |
| 740 | myInfo = (TransInvalidationInfo *) |
| 741 | MemoryContextAllocZero(TopTransactionContext, |
| 742 | sizeof(TransInvalidationInfo)); |
| 743 | myInfo->parent = transInvalInfo; |
| 744 | myInfo->my_level = GetCurrentTransactionNestLevel(); |
| 745 | |
| 746 | /* |
| 747 | * If there's any previous entry, this one should be for a deeper nesting |
| 748 | * level. |
| 749 | */ |
| 750 | Assert(transInvalInfo == NULL || |
| 751 | myInfo->my_level > transInvalInfo->my_level); |
| 752 | |
| 753 | transInvalInfo = myInfo; |
| 754 | } |
| 755 | |
| 756 | /* |
| 757 | * PostPrepare_Inval |
| 758 | * Clean up after successful PREPARE. |
| 759 | * |
| 760 | * Here, we want to act as though the transaction aborted, so that we will |
| 761 | * undo any syscache changes it made, thereby bringing us into sync with the |
| 762 | * outside world, which doesn't believe the transaction committed yet. |
| 763 | * |
| 764 | * If the prepared transaction is later aborted, there is nothing more to |
| 765 | * do; if it commits, we will receive the consequent inval messages just |
| 766 | * like everyone else. |
| 767 | */ |
| 768 | void |
| 769 | PostPrepare_Inval(void) |
| 770 | { |
| 771 | AtEOXact_Inval(false); |
| 772 | } |
| 773 | |
| 774 | /* |
| 775 | * Collect invalidation messages into SharedInvalidMessagesArray array. |
| 776 | */ |
| 777 | static void |
| 778 | MakeSharedInvalidMessagesArray(const SharedInvalidationMessage *msgs, int n) |
| 779 | { |
| 780 | /* |
| 781 | * Initialise array first time through in each commit |
| 782 | */ |
| 783 | if (SharedInvalidMessagesArray == NULL) |
| 784 | { |
| 785 | maxSharedInvalidMessagesArray = FIRSTCHUNKSIZE; |
| 786 | numSharedInvalidMessagesArray = 0; |
| 787 | |
| 788 | /* |
| 789 | * Although this is being palloc'd we don't actually free it directly. |
| 790 | * We're so close to EOXact that we now we're going to lose it anyhow. |
| 791 | */ |
| 792 | SharedInvalidMessagesArray = palloc(maxSharedInvalidMessagesArray |
| 793 | * sizeof(SharedInvalidationMessage)); |
| 794 | } |
| 795 | |
| 796 | if ((numSharedInvalidMessagesArray + n) > maxSharedInvalidMessagesArray) |
| 797 | { |
| 798 | while ((numSharedInvalidMessagesArray + n) > maxSharedInvalidMessagesArray) |
| 799 | maxSharedInvalidMessagesArray *= 2; |
| 800 | |
| 801 | SharedInvalidMessagesArray = repalloc(SharedInvalidMessagesArray, |
| 802 | maxSharedInvalidMessagesArray |
| 803 | * sizeof(SharedInvalidationMessage)); |
| 804 | } |
| 805 | |
| 806 | /* |
| 807 | * Append the next chunk onto the array |
| 808 | */ |
| 809 | memcpy(SharedInvalidMessagesArray + numSharedInvalidMessagesArray, |
| 810 | msgs, n * sizeof(SharedInvalidationMessage)); |
| 811 | numSharedInvalidMessagesArray += n; |
| 812 | } |
| 813 | |
| 814 | /* |
| 815 | * xactGetCommittedInvalidationMessages() is executed by |
| 816 | * RecordTransactionCommit() to add invalidation messages onto the |
| 817 | * commit record. This applies only to commit message types, never to |
| 818 | * abort records. Must always run before AtEOXact_Inval(), since that |
| 819 | * removes the data we need to see. |
| 820 | * |
| 821 | * Remember that this runs before we have officially committed, so we |
| 822 | * must not do anything here to change what might occur *if* we should |
| 823 | * fail between here and the actual commit. |
| 824 | * |
| 825 | * see also xact_redo_commit() and xact_desc_commit() |
| 826 | */ |
| 827 | int |
| 828 | xactGetCommittedInvalidationMessages(SharedInvalidationMessage **msgs, |
| 829 | bool *RelcacheInitFileInval) |
| 830 | { |
| 831 | MemoryContext oldcontext; |
| 832 | |
| 833 | /* Quick exit if we haven't done anything with invalidation messages. */ |
| 834 | if (transInvalInfo == NULL) |
| 835 | { |
| 836 | *RelcacheInitFileInval = false; |
| 837 | *msgs = NULL; |
| 838 | return 0; |
| 839 | } |
| 840 | |
| 841 | /* Must be at top of stack */ |
| 842 | Assert(transInvalInfo->my_level == 1 && transInvalInfo->parent == NULL); |
| 843 | |
| 844 | /* |
| 845 | * Relcache init file invalidation requires processing both before and |
| 846 | * after we send the SI messages. However, we need not do anything unless |
| 847 | * we committed. |
| 848 | */ |
| 849 | *RelcacheInitFileInval = transInvalInfo->RelcacheInitFileInval; |
| 850 | |
| 851 | /* |
| 852 | * Walk through TransInvalidationInfo to collect all the messages into a |
| 853 | * single contiguous array of invalidation messages. It must be contiguous |
| 854 | * so we can copy directly into WAL message. Maintain the order that they |
| 855 | * would be processed in by AtEOXact_Inval(), to ensure emulated behaviour |
| 856 | * in redo is as similar as possible to original. We want the same bugs, |
| 857 | * if any, not new ones. |
| 858 | */ |
| 859 | oldcontext = MemoryContextSwitchTo(CurTransactionContext); |
| 860 | |
| 861 | ProcessInvalidationMessagesMulti(&transInvalInfo->CurrentCmdInvalidMsgs, |
| 862 | MakeSharedInvalidMessagesArray); |
| 863 | ProcessInvalidationMessagesMulti(&transInvalInfo->PriorCmdInvalidMsgs, |
| 864 | MakeSharedInvalidMessagesArray); |
| 865 | MemoryContextSwitchTo(oldcontext); |
| 866 | |
| 867 | Assert(!(numSharedInvalidMessagesArray > 0 && |
| 868 | SharedInvalidMessagesArray == NULL)); |
| 869 | |
| 870 | *msgs = SharedInvalidMessagesArray; |
| 871 | |
| 872 | return numSharedInvalidMessagesArray; |
| 873 | } |
| 874 | |
| 875 | /* |
| 876 | * ProcessCommittedInvalidationMessages is executed by xact_redo_commit() or |
| 877 | * standby_redo() to process invalidation messages. Currently that happens |
| 878 | * only at end-of-xact. |
| 879 | * |
| 880 | * Relcache init file invalidation requires processing both |
| 881 | * before and after we send the SI messages. See AtEOXact_Inval() |
| 882 | */ |
| 883 | void |
| 884 | ProcessCommittedInvalidationMessages(SharedInvalidationMessage *msgs, |
| 885 | int nmsgs, bool RelcacheInitFileInval, |
| 886 | Oid dbid, Oid tsid) |
| 887 | { |
| 888 | if (nmsgs <= 0) |
| 889 | return; |
| 890 | |
| 891 | elog(trace_recovery(DEBUG4), "replaying commit with %d messages%s" , nmsgs, |
| 892 | (RelcacheInitFileInval ? " and relcache file invalidation" : "" )); |
| 893 | |
| 894 | if (RelcacheInitFileInval) |
| 895 | { |
| 896 | elog(trace_recovery(DEBUG4), "removing relcache init files for database %u" , |
| 897 | dbid); |
| 898 | |
| 899 | /* |
| 900 | * RelationCacheInitFilePreInvalidate, when the invalidation message |
| 901 | * is for a specific database, requires DatabasePath to be set, but we |
| 902 | * should not use SetDatabasePath during recovery, since it is |
| 903 | * intended to be used only once by normal backends. Hence, a quick |
| 904 | * hack: set DatabasePath directly then unset after use. |
| 905 | */ |
| 906 | if (OidIsValid(dbid)) |
| 907 | DatabasePath = GetDatabasePath(dbid, tsid); |
| 908 | |
| 909 | RelationCacheInitFilePreInvalidate(); |
| 910 | |
| 911 | if (OidIsValid(dbid)) |
| 912 | { |
| 913 | pfree(DatabasePath); |
| 914 | DatabasePath = NULL; |
| 915 | } |
| 916 | } |
| 917 | |
| 918 | SendSharedInvalidMessages(msgs, nmsgs); |
| 919 | |
| 920 | if (RelcacheInitFileInval) |
| 921 | RelationCacheInitFilePostInvalidate(); |
| 922 | } |
| 923 | |
| 924 | /* |
| 925 | * AtEOXact_Inval |
| 926 | * Process queued-up invalidation messages at end of main transaction. |
| 927 | * |
| 928 | * If isCommit, we must send out the messages in our PriorCmdInvalidMsgs list |
| 929 | * to the shared invalidation message queue. Note that these will be read |
| 930 | * not only by other backends, but also by our own backend at the next |
| 931 | * transaction start (via AcceptInvalidationMessages). This means that |
| 932 | * we can skip immediate local processing of anything that's still in |
| 933 | * CurrentCmdInvalidMsgs, and just send that list out too. |
| 934 | * |
| 935 | * If not isCommit, we are aborting, and must locally process the messages |
| 936 | * in PriorCmdInvalidMsgs. No messages need be sent to other backends, |
| 937 | * since they'll not have seen our changed tuples anyway. We can forget |
| 938 | * about CurrentCmdInvalidMsgs too, since those changes haven't touched |
| 939 | * the caches yet. |
| 940 | * |
| 941 | * In any case, reset the various lists to empty. We need not physically |
| 942 | * free memory here, since TopTransactionContext is about to be emptied |
| 943 | * anyway. |
| 944 | * |
| 945 | * Note: |
| 946 | * This should be called as the last step in processing a transaction. |
| 947 | */ |
| 948 | void |
| 949 | AtEOXact_Inval(bool isCommit) |
| 950 | { |
| 951 | /* Quick exit if no messages */ |
| 952 | if (transInvalInfo == NULL) |
| 953 | return; |
| 954 | |
| 955 | /* Must be at top of stack */ |
| 956 | Assert(transInvalInfo->my_level == 1 && transInvalInfo->parent == NULL); |
| 957 | |
| 958 | if (isCommit) |
| 959 | { |
| 960 | /* |
| 961 | * Relcache init file invalidation requires processing both before and |
| 962 | * after we send the SI messages. However, we need not do anything |
| 963 | * unless we committed. |
| 964 | */ |
| 965 | if (transInvalInfo->RelcacheInitFileInval) |
| 966 | RelationCacheInitFilePreInvalidate(); |
| 967 | |
| 968 | AppendInvalidationMessages(&transInvalInfo->PriorCmdInvalidMsgs, |
| 969 | &transInvalInfo->CurrentCmdInvalidMsgs); |
| 970 | |
| 971 | ProcessInvalidationMessagesMulti(&transInvalInfo->PriorCmdInvalidMsgs, |
| 972 | SendSharedInvalidMessages); |
| 973 | |
| 974 | if (transInvalInfo->RelcacheInitFileInval) |
| 975 | RelationCacheInitFilePostInvalidate(); |
| 976 | } |
| 977 | else |
| 978 | { |
| 979 | ProcessInvalidationMessages(&transInvalInfo->PriorCmdInvalidMsgs, |
| 980 | LocalExecuteInvalidationMessage); |
| 981 | } |
| 982 | |
| 983 | /* Need not free anything explicitly */ |
| 984 | transInvalInfo = NULL; |
| 985 | SharedInvalidMessagesArray = NULL; |
| 986 | numSharedInvalidMessagesArray = 0; |
| 987 | } |
| 988 | |
| 989 | /* |
| 990 | * AtEOSubXact_Inval |
| 991 | * Process queued-up invalidation messages at end of subtransaction. |
| 992 | * |
| 993 | * If isCommit, process CurrentCmdInvalidMsgs if any (there probably aren't), |
| 994 | * and then attach both CurrentCmdInvalidMsgs and PriorCmdInvalidMsgs to the |
| 995 | * parent's PriorCmdInvalidMsgs list. |
| 996 | * |
| 997 | * If not isCommit, we are aborting, and must locally process the messages |
| 998 | * in PriorCmdInvalidMsgs. No messages need be sent to other backends. |
| 999 | * We can forget about CurrentCmdInvalidMsgs too, since those changes haven't |
| 1000 | * touched the caches yet. |
| 1001 | * |
| 1002 | * In any case, pop the transaction stack. We need not physically free memory |
| 1003 | * here, since CurTransactionContext is about to be emptied anyway |
| 1004 | * (if aborting). Beware of the possibility of aborting the same nesting |
| 1005 | * level twice, though. |
| 1006 | */ |
| 1007 | void |
| 1008 | AtEOSubXact_Inval(bool isCommit) |
| 1009 | { |
| 1010 | int my_level; |
| 1011 | TransInvalidationInfo *myInfo = transInvalInfo; |
| 1012 | |
| 1013 | /* Quick exit if no messages. */ |
| 1014 | if (myInfo == NULL) |
| 1015 | return; |
| 1016 | |
| 1017 | /* Also bail out quickly if messages are not for this level. */ |
| 1018 | my_level = GetCurrentTransactionNestLevel(); |
| 1019 | if (myInfo->my_level != my_level) |
| 1020 | { |
| 1021 | Assert(myInfo->my_level < my_level); |
| 1022 | return; |
| 1023 | } |
| 1024 | |
| 1025 | if (isCommit) |
| 1026 | { |
| 1027 | /* If CurrentCmdInvalidMsgs still has anything, fix it */ |
| 1028 | CommandEndInvalidationMessages(); |
| 1029 | |
| 1030 | /* |
| 1031 | * We create invalidation stack entries lazily, so the parent might |
| 1032 | * not have one. Instead of creating one, moving all the data over, |
| 1033 | * and then freeing our own, we can just adjust the level of our own |
| 1034 | * entry. |
| 1035 | */ |
| 1036 | if (myInfo->parent == NULL || myInfo->parent->my_level < my_level - 1) |
| 1037 | { |
| 1038 | myInfo->my_level--; |
| 1039 | return; |
| 1040 | } |
| 1041 | |
| 1042 | /* Pass up my inval messages to parent */ |
| 1043 | AppendInvalidationMessages(&myInfo->parent->PriorCmdInvalidMsgs, |
| 1044 | &myInfo->PriorCmdInvalidMsgs); |
| 1045 | |
| 1046 | /* Pending relcache inval becomes parent's problem too */ |
| 1047 | if (myInfo->RelcacheInitFileInval) |
| 1048 | myInfo->parent->RelcacheInitFileInval = true; |
| 1049 | |
| 1050 | /* Pop the transaction state stack */ |
| 1051 | transInvalInfo = myInfo->parent; |
| 1052 | |
| 1053 | /* Need not free anything else explicitly */ |
| 1054 | pfree(myInfo); |
| 1055 | } |
| 1056 | else |
| 1057 | { |
| 1058 | ProcessInvalidationMessages(&myInfo->PriorCmdInvalidMsgs, |
| 1059 | LocalExecuteInvalidationMessage); |
| 1060 | |
| 1061 | /* Pop the transaction state stack */ |
| 1062 | transInvalInfo = myInfo->parent; |
| 1063 | |
| 1064 | /* Need not free anything else explicitly */ |
| 1065 | pfree(myInfo); |
| 1066 | } |
| 1067 | } |
| 1068 | |
| 1069 | /* |
| 1070 | * CommandEndInvalidationMessages |
| 1071 | * Process queued-up invalidation messages at end of one command |
| 1072 | * in a transaction. |
| 1073 | * |
| 1074 | * Here, we send no messages to the shared queue, since we don't know yet if |
| 1075 | * we will commit. We do need to locally process the CurrentCmdInvalidMsgs |
| 1076 | * list, so as to flush our caches of any entries we have outdated in the |
| 1077 | * current command. We then move the current-cmd list over to become part |
| 1078 | * of the prior-cmds list. |
| 1079 | * |
| 1080 | * Note: |
| 1081 | * This should be called during CommandCounterIncrement(), |
| 1082 | * after we have advanced the command ID. |
| 1083 | */ |
| 1084 | void |
| 1085 | CommandEndInvalidationMessages(void) |
| 1086 | { |
| 1087 | /* |
| 1088 | * You might think this shouldn't be called outside any transaction, but |
| 1089 | * bootstrap does it, and also ABORT issued when not in a transaction. So |
| 1090 | * just quietly return if no state to work on. |
| 1091 | */ |
| 1092 | if (transInvalInfo == NULL) |
| 1093 | return; |
| 1094 | |
| 1095 | ProcessInvalidationMessages(&transInvalInfo->CurrentCmdInvalidMsgs, |
| 1096 | LocalExecuteInvalidationMessage); |
| 1097 | AppendInvalidationMessages(&transInvalInfo->PriorCmdInvalidMsgs, |
| 1098 | &transInvalInfo->CurrentCmdInvalidMsgs); |
| 1099 | } |
| 1100 | |
| 1101 | |
| 1102 | /* |
| 1103 | * CacheInvalidateHeapTuple |
| 1104 | * Register the given tuple for invalidation at end of command |
| 1105 | * (ie, current command is creating or outdating this tuple). |
| 1106 | * Also, detect whether a relcache invalidation is implied. |
| 1107 | * |
| 1108 | * For an insert or delete, tuple is the target tuple and newtuple is NULL. |
| 1109 | * For an update, we are called just once, with tuple being the old tuple |
| 1110 | * version and newtuple the new version. This allows avoidance of duplicate |
| 1111 | * effort during an update. |
| 1112 | */ |
| 1113 | void |
| 1114 | CacheInvalidateHeapTuple(Relation relation, |
| 1115 | HeapTuple tuple, |
| 1116 | HeapTuple newtuple) |
| 1117 | { |
| 1118 | Oid tupleRelId; |
| 1119 | Oid databaseId; |
| 1120 | Oid relationId; |
| 1121 | |
| 1122 | /* Do nothing during bootstrap */ |
| 1123 | if (IsBootstrapProcessingMode()) |
| 1124 | return; |
| 1125 | |
| 1126 | /* |
| 1127 | * We only need to worry about invalidation for tuples that are in system |
| 1128 | * catalogs; user-relation tuples are never in catcaches and can't affect |
| 1129 | * the relcache either. |
| 1130 | */ |
| 1131 | if (!IsCatalogRelation(relation)) |
| 1132 | return; |
| 1133 | |
| 1134 | /* |
| 1135 | * IsCatalogRelation() will return true for TOAST tables of system |
| 1136 | * catalogs, but we don't care about those, either. |
| 1137 | */ |
| 1138 | if (IsToastRelation(relation)) |
| 1139 | return; |
| 1140 | |
| 1141 | /* |
| 1142 | * If we're not prepared to queue invalidation messages for this |
| 1143 | * subtransaction level, get ready now. |
| 1144 | */ |
| 1145 | PrepareInvalidationState(); |
| 1146 | |
| 1147 | /* |
| 1148 | * First let the catcache do its thing |
| 1149 | */ |
| 1150 | tupleRelId = RelationGetRelid(relation); |
| 1151 | if (RelationInvalidatesSnapshotsOnly(tupleRelId)) |
| 1152 | { |
| 1153 | databaseId = IsSharedRelation(tupleRelId) ? InvalidOid : MyDatabaseId; |
| 1154 | RegisterSnapshotInvalidation(databaseId, tupleRelId); |
| 1155 | } |
| 1156 | else |
| 1157 | PrepareToInvalidateCacheTuple(relation, tuple, newtuple, |
| 1158 | RegisterCatcacheInvalidation); |
| 1159 | |
| 1160 | /* |
| 1161 | * Now, is this tuple one of the primary definers of a relcache entry? See |
| 1162 | * comments in file header for deeper explanation. |
| 1163 | * |
| 1164 | * Note we ignore newtuple here; we assume an update cannot move a tuple |
| 1165 | * from being part of one relcache entry to being part of another. |
| 1166 | */ |
| 1167 | if (tupleRelId == RelationRelationId) |
| 1168 | { |
| 1169 | Form_pg_class classtup = (Form_pg_class) GETSTRUCT(tuple); |
| 1170 | |
| 1171 | relationId = classtup->oid; |
| 1172 | if (classtup->relisshared) |
| 1173 | databaseId = InvalidOid; |
| 1174 | else |
| 1175 | databaseId = MyDatabaseId; |
| 1176 | } |
| 1177 | else if (tupleRelId == AttributeRelationId) |
| 1178 | { |
| 1179 | Form_pg_attribute atttup = (Form_pg_attribute) GETSTRUCT(tuple); |
| 1180 | |
| 1181 | relationId = atttup->attrelid; |
| 1182 | |
| 1183 | /* |
| 1184 | * KLUGE ALERT: we always send the relcache event with MyDatabaseId, |
| 1185 | * even if the rel in question is shared (which we can't easily tell). |
| 1186 | * This essentially means that only backends in this same database |
| 1187 | * will react to the relcache flush request. This is in fact |
| 1188 | * appropriate, since only those backends could see our pg_attribute |
| 1189 | * change anyway. It looks a bit ugly though. (In practice, shared |
| 1190 | * relations can't have schema changes after bootstrap, so we should |
| 1191 | * never come here for a shared rel anyway.) |
| 1192 | */ |
| 1193 | databaseId = MyDatabaseId; |
| 1194 | } |
| 1195 | else if (tupleRelId == IndexRelationId) |
| 1196 | { |
| 1197 | Form_pg_index indextup = (Form_pg_index) GETSTRUCT(tuple); |
| 1198 | |
| 1199 | /* |
| 1200 | * When a pg_index row is updated, we should send out a relcache inval |
| 1201 | * for the index relation. As above, we don't know the shared status |
| 1202 | * of the index, but in practice it doesn't matter since indexes of |
| 1203 | * shared catalogs can't have such updates. |
| 1204 | */ |
| 1205 | relationId = indextup->indexrelid; |
| 1206 | databaseId = MyDatabaseId; |
| 1207 | } |
| 1208 | else if (tupleRelId == ConstraintRelationId) |
| 1209 | { |
| 1210 | Form_pg_constraint constrtup = (Form_pg_constraint) GETSTRUCT(tuple); |
| 1211 | |
| 1212 | /* |
| 1213 | * Foreign keys are part of relcache entries, too, so send out an |
| 1214 | * inval for the table that the FK applies to. |
| 1215 | */ |
| 1216 | if (constrtup->contype == CONSTRAINT_FOREIGN && |
| 1217 | OidIsValid(constrtup->conrelid)) |
| 1218 | { |
| 1219 | relationId = constrtup->conrelid; |
| 1220 | databaseId = MyDatabaseId; |
| 1221 | } |
| 1222 | else |
| 1223 | return; |
| 1224 | } |
| 1225 | else |
| 1226 | return; |
| 1227 | |
| 1228 | /* |
| 1229 | * Yes. We need to register a relcache invalidation event. |
| 1230 | */ |
| 1231 | RegisterRelcacheInvalidation(databaseId, relationId); |
| 1232 | } |
| 1233 | |
| 1234 | /* |
| 1235 | * CacheInvalidateCatalog |
| 1236 | * Register invalidation of the whole content of a system catalog. |
| 1237 | * |
| 1238 | * This is normally used in VACUUM FULL/CLUSTER, where we haven't so much |
| 1239 | * changed any tuples as moved them around. Some uses of catcache entries |
| 1240 | * expect their TIDs to be correct, so we have to blow away the entries. |
| 1241 | * |
| 1242 | * Note: we expect caller to verify that the rel actually is a system |
| 1243 | * catalog. If it isn't, no great harm is done, just a wasted sinval message. |
| 1244 | */ |
| 1245 | void |
| 1246 | CacheInvalidateCatalog(Oid catalogId) |
| 1247 | { |
| 1248 | Oid databaseId; |
| 1249 | |
| 1250 | PrepareInvalidationState(); |
| 1251 | |
| 1252 | if (IsSharedRelation(catalogId)) |
| 1253 | databaseId = InvalidOid; |
| 1254 | else |
| 1255 | databaseId = MyDatabaseId; |
| 1256 | |
| 1257 | RegisterCatalogInvalidation(databaseId, catalogId); |
| 1258 | } |
| 1259 | |
| 1260 | /* |
| 1261 | * CacheInvalidateRelcache |
| 1262 | * Register invalidation of the specified relation's relcache entry |
| 1263 | * at end of command. |
| 1264 | * |
| 1265 | * This is used in places that need to force relcache rebuild but aren't |
| 1266 | * changing any of the tuples recognized as contributors to the relcache |
| 1267 | * entry by CacheInvalidateHeapTuple. (An example is dropping an index.) |
| 1268 | */ |
| 1269 | void |
| 1270 | CacheInvalidateRelcache(Relation relation) |
| 1271 | { |
| 1272 | Oid databaseId; |
| 1273 | Oid relationId; |
| 1274 | |
| 1275 | PrepareInvalidationState(); |
| 1276 | |
| 1277 | relationId = RelationGetRelid(relation); |
| 1278 | if (relation->rd_rel->relisshared) |
| 1279 | databaseId = InvalidOid; |
| 1280 | else |
| 1281 | databaseId = MyDatabaseId; |
| 1282 | |
| 1283 | RegisterRelcacheInvalidation(databaseId, relationId); |
| 1284 | } |
| 1285 | |
| 1286 | /* |
| 1287 | * CacheInvalidateRelcacheAll |
| 1288 | * Register invalidation of the whole relcache at the end of command. |
| 1289 | * |
| 1290 | * This is used by alter publication as changes in publications may affect |
| 1291 | * large number of tables. |
| 1292 | */ |
| 1293 | void |
| 1294 | CacheInvalidateRelcacheAll(void) |
| 1295 | { |
| 1296 | PrepareInvalidationState(); |
| 1297 | |
| 1298 | RegisterRelcacheInvalidation(InvalidOid, InvalidOid); |
| 1299 | } |
| 1300 | |
| 1301 | /* |
| 1302 | * CacheInvalidateRelcacheByTuple |
| 1303 | * As above, but relation is identified by passing its pg_class tuple. |
| 1304 | */ |
| 1305 | void |
| 1306 | CacheInvalidateRelcacheByTuple(HeapTuple classTuple) |
| 1307 | { |
| 1308 | Form_pg_class classtup = (Form_pg_class) GETSTRUCT(classTuple); |
| 1309 | Oid databaseId; |
| 1310 | Oid relationId; |
| 1311 | |
| 1312 | PrepareInvalidationState(); |
| 1313 | |
| 1314 | relationId = classtup->oid; |
| 1315 | if (classtup->relisshared) |
| 1316 | databaseId = InvalidOid; |
| 1317 | else |
| 1318 | databaseId = MyDatabaseId; |
| 1319 | RegisterRelcacheInvalidation(databaseId, relationId); |
| 1320 | } |
| 1321 | |
| 1322 | /* |
| 1323 | * CacheInvalidateRelcacheByRelid |
| 1324 | * As above, but relation is identified by passing its OID. |
| 1325 | * This is the least efficient of the three options; use one of |
| 1326 | * the above routines if you have a Relation or pg_class tuple. |
| 1327 | */ |
| 1328 | void |
| 1329 | CacheInvalidateRelcacheByRelid(Oid relid) |
| 1330 | { |
| 1331 | HeapTuple tup; |
| 1332 | |
| 1333 | PrepareInvalidationState(); |
| 1334 | |
| 1335 | tup = SearchSysCache1(RELOID, ObjectIdGetDatum(relid)); |
| 1336 | if (!HeapTupleIsValid(tup)) |
| 1337 | elog(ERROR, "cache lookup failed for relation %u" , relid); |
| 1338 | CacheInvalidateRelcacheByTuple(tup); |
| 1339 | ReleaseSysCache(tup); |
| 1340 | } |
| 1341 | |
| 1342 | |
| 1343 | /* |
| 1344 | * CacheInvalidateSmgr |
| 1345 | * Register invalidation of smgr references to a physical relation. |
| 1346 | * |
| 1347 | * Sending this type of invalidation msg forces other backends to close open |
| 1348 | * smgr entries for the rel. This should be done to flush dangling open-file |
| 1349 | * references when the physical rel is being dropped or truncated. Because |
| 1350 | * these are nontransactional (i.e., not-rollback-able) operations, we just |
| 1351 | * send the inval message immediately without any queuing. |
| 1352 | * |
| 1353 | * Note: in most cases there will have been a relcache flush issued against |
| 1354 | * the rel at the logical level. We need a separate smgr-level flush because |
| 1355 | * it is possible for backends to have open smgr entries for rels they don't |
| 1356 | * have a relcache entry for, e.g. because the only thing they ever did with |
| 1357 | * the rel is write out dirty shared buffers. |
| 1358 | * |
| 1359 | * Note: because these messages are nontransactional, they won't be captured |
| 1360 | * in commit/abort WAL entries. Instead, calls to CacheInvalidateSmgr() |
| 1361 | * should happen in low-level smgr.c routines, which are executed while |
| 1362 | * replaying WAL as well as when creating it. |
| 1363 | * |
| 1364 | * Note: In order to avoid bloating SharedInvalidationMessage, we store only |
| 1365 | * three bytes of the backend ID using what would otherwise be padding space. |
| 1366 | * Thus, the maximum possible backend ID is 2^23-1. |
| 1367 | */ |
| 1368 | void |
| 1369 | CacheInvalidateSmgr(RelFileNodeBackend rnode) |
| 1370 | { |
| 1371 | SharedInvalidationMessage msg; |
| 1372 | |
| 1373 | msg.sm.id = SHAREDINVALSMGR_ID; |
| 1374 | msg.sm.backend_hi = rnode.backend >> 16; |
| 1375 | msg.sm.backend_lo = rnode.backend & 0xffff; |
| 1376 | msg.sm.rnode = rnode.node; |
| 1377 | /* check AddCatcacheInvalidationMessage() for an explanation */ |
| 1378 | VALGRIND_MAKE_MEM_DEFINED(&msg, sizeof(msg)); |
| 1379 | |
| 1380 | SendSharedInvalidMessages(&msg, 1); |
| 1381 | } |
| 1382 | |
| 1383 | /* |
| 1384 | * CacheInvalidateRelmap |
| 1385 | * Register invalidation of the relation mapping for a database, |
| 1386 | * or for the shared catalogs if databaseId is zero. |
| 1387 | * |
| 1388 | * Sending this type of invalidation msg forces other backends to re-read |
| 1389 | * the indicated relation mapping file. It is also necessary to send a |
| 1390 | * relcache inval for the specific relations whose mapping has been altered, |
| 1391 | * else the relcache won't get updated with the new filenode data. |
| 1392 | * |
| 1393 | * Note: because these messages are nontransactional, they won't be captured |
| 1394 | * in commit/abort WAL entries. Instead, calls to CacheInvalidateRelmap() |
| 1395 | * should happen in low-level relmapper.c routines, which are executed while |
| 1396 | * replaying WAL as well as when creating it. |
| 1397 | */ |
| 1398 | void |
| 1399 | CacheInvalidateRelmap(Oid databaseId) |
| 1400 | { |
| 1401 | SharedInvalidationMessage msg; |
| 1402 | |
| 1403 | msg.rm.id = SHAREDINVALRELMAP_ID; |
| 1404 | msg.rm.dbId = databaseId; |
| 1405 | /* check AddCatcacheInvalidationMessage() for an explanation */ |
| 1406 | VALGRIND_MAKE_MEM_DEFINED(&msg, sizeof(msg)); |
| 1407 | |
| 1408 | SendSharedInvalidMessages(&msg, 1); |
| 1409 | } |
| 1410 | |
| 1411 | |
| 1412 | /* |
| 1413 | * CacheRegisterSyscacheCallback |
| 1414 | * Register the specified function to be called for all future |
| 1415 | * invalidation events in the specified cache. The cache ID and the |
| 1416 | * hash value of the tuple being invalidated will be passed to the |
| 1417 | * function. |
| 1418 | * |
| 1419 | * NOTE: Hash value zero will be passed if a cache reset request is received. |
| 1420 | * In this case the called routines should flush all cached state. |
| 1421 | * Yes, there's a possibility of a false match to zero, but it doesn't seem |
| 1422 | * worth troubling over, especially since most of the current callees just |
| 1423 | * flush all cached state anyway. |
| 1424 | */ |
| 1425 | void |
| 1426 | CacheRegisterSyscacheCallback(int cacheid, |
| 1427 | SyscacheCallbackFunction func, |
| 1428 | Datum arg) |
| 1429 | { |
| 1430 | if (cacheid < 0 || cacheid >= SysCacheSize) |
| 1431 | elog(FATAL, "invalid cache ID: %d" , cacheid); |
| 1432 | if (syscache_callback_count >= MAX_SYSCACHE_CALLBACKS) |
| 1433 | elog(FATAL, "out of syscache_callback_list slots" ); |
| 1434 | |
| 1435 | if (syscache_callback_links[cacheid] == 0) |
| 1436 | { |
| 1437 | /* first callback for this cache */ |
| 1438 | syscache_callback_links[cacheid] = syscache_callback_count + 1; |
| 1439 | } |
| 1440 | else |
| 1441 | { |
| 1442 | /* add to end of chain, so that older callbacks are called first */ |
| 1443 | int i = syscache_callback_links[cacheid] - 1; |
| 1444 | |
| 1445 | while (syscache_callback_list[i].link > 0) |
| 1446 | i = syscache_callback_list[i].link - 1; |
| 1447 | syscache_callback_list[i].link = syscache_callback_count + 1; |
| 1448 | } |
| 1449 | |
| 1450 | syscache_callback_list[syscache_callback_count].id = cacheid; |
| 1451 | syscache_callback_list[syscache_callback_count].link = 0; |
| 1452 | syscache_callback_list[syscache_callback_count].function = func; |
| 1453 | syscache_callback_list[syscache_callback_count].arg = arg; |
| 1454 | |
| 1455 | ++syscache_callback_count; |
| 1456 | } |
| 1457 | |
| 1458 | /* |
| 1459 | * CacheRegisterRelcacheCallback |
| 1460 | * Register the specified function to be called for all future |
| 1461 | * relcache invalidation events. The OID of the relation being |
| 1462 | * invalidated will be passed to the function. |
| 1463 | * |
| 1464 | * NOTE: InvalidOid will be passed if a cache reset request is received. |
| 1465 | * In this case the called routines should flush all cached state. |
| 1466 | */ |
| 1467 | void |
| 1468 | CacheRegisterRelcacheCallback(RelcacheCallbackFunction func, |
| 1469 | Datum arg) |
| 1470 | { |
| 1471 | if (relcache_callback_count >= MAX_RELCACHE_CALLBACKS) |
| 1472 | elog(FATAL, "out of relcache_callback_list slots" ); |
| 1473 | |
| 1474 | relcache_callback_list[relcache_callback_count].function = func; |
| 1475 | relcache_callback_list[relcache_callback_count].arg = arg; |
| 1476 | |
| 1477 | ++relcache_callback_count; |
| 1478 | } |
| 1479 | |
| 1480 | /* |
| 1481 | * CallSyscacheCallbacks |
| 1482 | * |
| 1483 | * This is exported so that CatalogCacheFlushCatalog can call it, saving |
| 1484 | * this module from knowing which catcache IDs correspond to which catalogs. |
| 1485 | */ |
| 1486 | void |
| 1487 | CallSyscacheCallbacks(int cacheid, uint32 hashvalue) |
| 1488 | { |
| 1489 | int i; |
| 1490 | |
| 1491 | if (cacheid < 0 || cacheid >= SysCacheSize) |
| 1492 | elog(ERROR, "invalid cache ID: %d" , cacheid); |
| 1493 | |
| 1494 | i = syscache_callback_links[cacheid] - 1; |
| 1495 | while (i >= 0) |
| 1496 | { |
| 1497 | struct SYSCACHECALLBACK *ccitem = syscache_callback_list + i; |
| 1498 | |
| 1499 | Assert(ccitem->id == cacheid); |
| 1500 | ccitem->function(ccitem->arg, cacheid, hashvalue); |
| 1501 | i = ccitem->link - 1; |
| 1502 | } |
| 1503 | } |
| 1504 | |