| 1 | /*------------------------------------------------------------------------- |
| 2 | * |
| 3 | * partcache.c |
| 4 | * Support routines for manipulating partition information cached in |
| 5 | * relcache |
| 6 | * |
| 7 | * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group |
| 8 | * Portions Copyright (c) 1994, Regents of the University of California |
| 9 | * |
| 10 | * IDENTIFICATION |
| 11 | * src/backend/utils/cache/partcache.c |
| 12 | * |
| 13 | *------------------------------------------------------------------------- |
| 14 | */ |
| 15 | #include "postgres.h" |
| 16 | |
| 17 | #include "access/hash.h" |
| 18 | #include "access/htup_details.h" |
| 19 | #include "access/nbtree.h" |
| 20 | #include "access/relation.h" |
| 21 | #include "catalog/partition.h" |
| 22 | #include "catalog/pg_inherits.h" |
| 23 | #include "catalog/pg_opclass.h" |
| 24 | #include "catalog/pg_partitioned_table.h" |
| 25 | #include "miscadmin.h" |
| 26 | #include "nodes/makefuncs.h" |
| 27 | #include "nodes/nodeFuncs.h" |
| 28 | #include "optimizer/optimizer.h" |
| 29 | #include "partitioning/partbounds.h" |
| 30 | #include "rewrite/rewriteHandler.h" |
| 31 | #include "utils/builtins.h" |
| 32 | #include "utils/datum.h" |
| 33 | #include "utils/lsyscache.h" |
| 34 | #include "utils/memutils.h" |
| 35 | #include "utils/partcache.h" |
| 36 | #include "utils/rel.h" |
| 37 | #include "utils/syscache.h" |
| 38 | |
| 39 | |
| 40 | static List *generate_partition_qual(Relation rel); |
| 41 | |
| 42 | /* |
| 43 | * RelationBuildPartitionKey |
| 44 | * Build partition key data of relation, and attach to relcache |
| 45 | * |
| 46 | * Partitioning key data is a complex structure; to avoid complicated logic to |
| 47 | * free individual elements whenever the relcache entry is flushed, we give it |
| 48 | * its own memory context, a child of CacheMemoryContext, which can easily be |
| 49 | * deleted on its own. To avoid leaking memory in that context in case of an |
| 50 | * error partway through this function, the context is initially created as a |
| 51 | * child of CurTransactionContext and only re-parented to CacheMemoryContext |
| 52 | * at the end, when no further errors are possible. Also, we don't make this |
| 53 | * context the current context except in very brief code sections, out of fear |
| 54 | * that some of our callees allocate memory on their own which would be leaked |
| 55 | * permanently. |
| 56 | */ |
| 57 | void |
| 58 | RelationBuildPartitionKey(Relation relation) |
| 59 | { |
| 60 | Form_pg_partitioned_table form; |
| 61 | HeapTuple tuple; |
| 62 | bool isnull; |
| 63 | int i; |
| 64 | PartitionKey key; |
| 65 | AttrNumber *attrs; |
| 66 | oidvector *opclass; |
| 67 | oidvector *collation; |
| 68 | ListCell *partexprs_item; |
| 69 | Datum datum; |
| 70 | MemoryContext partkeycxt, |
| 71 | oldcxt; |
| 72 | int16 procnum; |
| 73 | |
| 74 | tuple = SearchSysCache1(PARTRELID, |
| 75 | ObjectIdGetDatum(RelationGetRelid(relation))); |
| 76 | |
| 77 | /* |
| 78 | * The following happens when we have created our pg_class entry but not |
| 79 | * the pg_partitioned_table entry yet. |
| 80 | */ |
| 81 | if (!HeapTupleIsValid(tuple)) |
| 82 | return; |
| 83 | |
| 84 | partkeycxt = AllocSetContextCreate(CurTransactionContext, |
| 85 | "partition key" , |
| 86 | ALLOCSET_SMALL_SIZES); |
| 87 | MemoryContextCopyAndSetIdentifier(partkeycxt, |
| 88 | RelationGetRelationName(relation)); |
| 89 | |
| 90 | key = (PartitionKey) MemoryContextAllocZero(partkeycxt, |
| 91 | sizeof(PartitionKeyData)); |
| 92 | |
| 93 | /* Fixed-length attributes */ |
| 94 | form = (Form_pg_partitioned_table) GETSTRUCT(tuple); |
| 95 | key->strategy = form->partstrat; |
| 96 | key->partnatts = form->partnatts; |
| 97 | |
| 98 | /* |
| 99 | * We can rely on the first variable-length attribute being mapped to the |
| 100 | * relevant field of the catalog's C struct, because all previous |
| 101 | * attributes are non-nullable and fixed-length. |
| 102 | */ |
| 103 | attrs = form->partattrs.values; |
| 104 | |
| 105 | /* But use the hard way to retrieve further variable-length attributes */ |
| 106 | /* Operator class */ |
| 107 | datum = SysCacheGetAttr(PARTRELID, tuple, |
| 108 | Anum_pg_partitioned_table_partclass, &isnull); |
| 109 | Assert(!isnull); |
| 110 | opclass = (oidvector *) DatumGetPointer(datum); |
| 111 | |
| 112 | /* Collation */ |
| 113 | datum = SysCacheGetAttr(PARTRELID, tuple, |
| 114 | Anum_pg_partitioned_table_partcollation, &isnull); |
| 115 | Assert(!isnull); |
| 116 | collation = (oidvector *) DatumGetPointer(datum); |
| 117 | |
| 118 | /* Expressions */ |
| 119 | datum = SysCacheGetAttr(PARTRELID, tuple, |
| 120 | Anum_pg_partitioned_table_partexprs, &isnull); |
| 121 | if (!isnull) |
| 122 | { |
| 123 | char *exprString; |
| 124 | Node *expr; |
| 125 | |
| 126 | exprString = TextDatumGetCString(datum); |
| 127 | expr = stringToNode(exprString); |
| 128 | pfree(exprString); |
| 129 | |
| 130 | /* |
| 131 | * Run the expressions through const-simplification since the planner |
| 132 | * will be comparing them to similarly-processed qual clause operands, |
| 133 | * and may fail to detect valid matches without this step; fix |
| 134 | * opfuncids while at it. We don't need to bother with |
| 135 | * canonicalize_qual() though, because partition expressions should be |
| 136 | * in canonical form already (ie, no need for OR-merging or constant |
| 137 | * elimination). |
| 138 | */ |
| 139 | expr = eval_const_expressions(NULL, expr); |
| 140 | fix_opfuncids(expr); |
| 141 | |
| 142 | oldcxt = MemoryContextSwitchTo(partkeycxt); |
| 143 | key->partexprs = (List *) copyObject(expr); |
| 144 | MemoryContextSwitchTo(oldcxt); |
| 145 | } |
| 146 | |
| 147 | /* Allocate assorted arrays in the partkeycxt, which we'll fill below */ |
| 148 | oldcxt = MemoryContextSwitchTo(partkeycxt); |
| 149 | key->partattrs = (AttrNumber *) palloc0(key->partnatts * sizeof(AttrNumber)); |
| 150 | key->partopfamily = (Oid *) palloc0(key->partnatts * sizeof(Oid)); |
| 151 | key->partopcintype = (Oid *) palloc0(key->partnatts * sizeof(Oid)); |
| 152 | key->partsupfunc = (FmgrInfo *) palloc0(key->partnatts * sizeof(FmgrInfo)); |
| 153 | |
| 154 | key->partcollation = (Oid *) palloc0(key->partnatts * sizeof(Oid)); |
| 155 | key->parttypid = (Oid *) palloc0(key->partnatts * sizeof(Oid)); |
| 156 | key->parttypmod = (int32 *) palloc0(key->partnatts * sizeof(int32)); |
| 157 | key->parttyplen = (int16 *) palloc0(key->partnatts * sizeof(int16)); |
| 158 | key->parttypbyval = (bool *) palloc0(key->partnatts * sizeof(bool)); |
| 159 | key->parttypalign = (char *) palloc0(key->partnatts * sizeof(char)); |
| 160 | key->parttypcoll = (Oid *) palloc0(key->partnatts * sizeof(Oid)); |
| 161 | MemoryContextSwitchTo(oldcxt); |
| 162 | |
| 163 | /* determine support function number to search for */ |
| 164 | procnum = (key->strategy == PARTITION_STRATEGY_HASH) ? |
| 165 | HASHEXTENDED_PROC : BTORDER_PROC; |
| 166 | |
| 167 | /* Copy partattrs and fill other per-attribute info */ |
| 168 | memcpy(key->partattrs, attrs, key->partnatts * sizeof(int16)); |
| 169 | partexprs_item = list_head(key->partexprs); |
| 170 | for (i = 0; i < key->partnatts; i++) |
| 171 | { |
| 172 | AttrNumber attno = key->partattrs[i]; |
| 173 | HeapTuple opclasstup; |
| 174 | Form_pg_opclass opclassform; |
| 175 | Oid funcid; |
| 176 | |
| 177 | /* Collect opfamily information */ |
| 178 | opclasstup = SearchSysCache1(CLAOID, |
| 179 | ObjectIdGetDatum(opclass->values[i])); |
| 180 | if (!HeapTupleIsValid(opclasstup)) |
| 181 | elog(ERROR, "cache lookup failed for opclass %u" , opclass->values[i]); |
| 182 | |
| 183 | opclassform = (Form_pg_opclass) GETSTRUCT(opclasstup); |
| 184 | key->partopfamily[i] = opclassform->opcfamily; |
| 185 | key->partopcintype[i] = opclassform->opcintype; |
| 186 | |
| 187 | /* Get a support function for the specified opfamily and datatypes */ |
| 188 | funcid = get_opfamily_proc(opclassform->opcfamily, |
| 189 | opclassform->opcintype, |
| 190 | opclassform->opcintype, |
| 191 | procnum); |
| 192 | if (!OidIsValid(funcid)) |
| 193 | ereport(ERROR, |
| 194 | (errcode(ERRCODE_INVALID_OBJECT_DEFINITION), |
| 195 | errmsg("operator class \"%s\" of access method %s is missing support function %d for type %s" , |
| 196 | NameStr(opclassform->opcname), |
| 197 | (key->strategy == PARTITION_STRATEGY_HASH) ? |
| 198 | "hash" : "btree" , |
| 199 | procnum, |
| 200 | format_type_be(opclassform->opcintype)))); |
| 201 | |
| 202 | fmgr_info_cxt(funcid, &key->partsupfunc[i], partkeycxt); |
| 203 | |
| 204 | /* Collation */ |
| 205 | key->partcollation[i] = collation->values[i]; |
| 206 | |
| 207 | /* Collect type information */ |
| 208 | if (attno != 0) |
| 209 | { |
| 210 | Form_pg_attribute att = TupleDescAttr(relation->rd_att, attno - 1); |
| 211 | |
| 212 | key->parttypid[i] = att->atttypid; |
| 213 | key->parttypmod[i] = att->atttypmod; |
| 214 | key->parttypcoll[i] = att->attcollation; |
| 215 | } |
| 216 | else |
| 217 | { |
| 218 | if (partexprs_item == NULL) |
| 219 | elog(ERROR, "wrong number of partition key expressions" ); |
| 220 | |
| 221 | key->parttypid[i] = exprType(lfirst(partexprs_item)); |
| 222 | key->parttypmod[i] = exprTypmod(lfirst(partexprs_item)); |
| 223 | key->parttypcoll[i] = exprCollation(lfirst(partexprs_item)); |
| 224 | |
| 225 | partexprs_item = lnext(partexprs_item); |
| 226 | } |
| 227 | get_typlenbyvalalign(key->parttypid[i], |
| 228 | &key->parttyplen[i], |
| 229 | &key->parttypbyval[i], |
| 230 | &key->parttypalign[i]); |
| 231 | |
| 232 | ReleaseSysCache(opclasstup); |
| 233 | } |
| 234 | |
| 235 | ReleaseSysCache(tuple); |
| 236 | |
| 237 | /* Assert that we're not leaking any old data during assignments below */ |
| 238 | Assert(relation->rd_partkeycxt == NULL); |
| 239 | Assert(relation->rd_partkey == NULL); |
| 240 | |
| 241 | /* |
| 242 | * Success --- reparent our context and make the relcache point to the |
| 243 | * newly constructed key |
| 244 | */ |
| 245 | MemoryContextSetParent(partkeycxt, CacheMemoryContext); |
| 246 | relation->rd_partkeycxt = partkeycxt; |
| 247 | relation->rd_partkey = key; |
| 248 | } |
| 249 | |
| 250 | /* |
| 251 | * RelationGetPartitionQual |
| 252 | * |
| 253 | * Returns a list of partition quals |
| 254 | */ |
| 255 | List * |
| 256 | RelationGetPartitionQual(Relation rel) |
| 257 | { |
| 258 | /* Quick exit */ |
| 259 | if (!rel->rd_rel->relispartition) |
| 260 | return NIL; |
| 261 | |
| 262 | return generate_partition_qual(rel); |
| 263 | } |
| 264 | |
| 265 | /* |
| 266 | * get_partition_qual_relid |
| 267 | * |
| 268 | * Returns an expression tree describing the passed-in relation's partition |
| 269 | * constraint. |
| 270 | * |
| 271 | * If the relation is not found, or is not a partition, or there is no |
| 272 | * partition constraint, return NULL. We must guard against the first two |
| 273 | * cases because this supports a SQL function that could be passed any OID. |
| 274 | * The last case can happen even if relispartition is true, when a default |
| 275 | * partition is the only partition. |
| 276 | */ |
| 277 | Expr * |
| 278 | get_partition_qual_relid(Oid relid) |
| 279 | { |
| 280 | Expr *result = NULL; |
| 281 | |
| 282 | /* Do the work only if this relation exists and is a partition. */ |
| 283 | if (get_rel_relispartition(relid)) |
| 284 | { |
| 285 | Relation rel = relation_open(relid, AccessShareLock); |
| 286 | List *and_args; |
| 287 | |
| 288 | and_args = generate_partition_qual(rel); |
| 289 | |
| 290 | /* Convert implicit-AND list format to boolean expression */ |
| 291 | if (and_args == NIL) |
| 292 | result = NULL; |
| 293 | else if (list_length(and_args) > 1) |
| 294 | result = makeBoolExpr(AND_EXPR, and_args, -1); |
| 295 | else |
| 296 | result = linitial(and_args); |
| 297 | |
| 298 | /* Keep the lock, to allow safe deparsing against the rel by caller. */ |
| 299 | relation_close(rel, NoLock); |
| 300 | } |
| 301 | |
| 302 | return result; |
| 303 | } |
| 304 | |
| 305 | /* |
| 306 | * generate_partition_qual |
| 307 | * |
| 308 | * Generate partition predicate from rel's partition bound expression. The |
| 309 | * function returns a NIL list if there is no predicate. |
| 310 | * |
| 311 | * We cache a copy of the result in the relcache entry, after constructing |
| 312 | * it using the caller's context. This approach avoids leaking any data |
| 313 | * into long-lived cache contexts, especially if we fail partway through. |
| 314 | */ |
| 315 | static List * |
| 316 | generate_partition_qual(Relation rel) |
| 317 | { |
| 318 | HeapTuple tuple; |
| 319 | MemoryContext oldcxt; |
| 320 | Datum boundDatum; |
| 321 | bool isnull; |
| 322 | List *my_qual = NIL, |
| 323 | *result = NIL; |
| 324 | Relation parent; |
| 325 | bool found_whole_row; |
| 326 | |
| 327 | /* Guard against stack overflow due to overly deep partition tree */ |
| 328 | check_stack_depth(); |
| 329 | |
| 330 | /* If we already cached the result, just return a copy */ |
| 331 | if (rel->rd_partcheckvalid) |
| 332 | return copyObject(rel->rd_partcheck); |
| 333 | |
| 334 | /* Grab at least an AccessShareLock on the parent table */ |
| 335 | parent = relation_open(get_partition_parent(RelationGetRelid(rel)), |
| 336 | AccessShareLock); |
| 337 | |
| 338 | /* Get pg_class.relpartbound */ |
| 339 | tuple = SearchSysCache1(RELOID, RelationGetRelid(rel)); |
| 340 | if (!HeapTupleIsValid(tuple)) |
| 341 | elog(ERROR, "cache lookup failed for relation %u" , |
| 342 | RelationGetRelid(rel)); |
| 343 | |
| 344 | boundDatum = SysCacheGetAttr(RELOID, tuple, |
| 345 | Anum_pg_class_relpartbound, |
| 346 | &isnull); |
| 347 | if (!isnull) |
| 348 | { |
| 349 | PartitionBoundSpec *bound; |
| 350 | |
| 351 | bound = castNode(PartitionBoundSpec, |
| 352 | stringToNode(TextDatumGetCString(boundDatum))); |
| 353 | |
| 354 | my_qual = get_qual_from_partbound(rel, parent, bound); |
| 355 | } |
| 356 | |
| 357 | ReleaseSysCache(tuple); |
| 358 | |
| 359 | /* Add the parent's quals to the list (if any) */ |
| 360 | if (parent->rd_rel->relispartition) |
| 361 | result = list_concat(generate_partition_qual(parent), my_qual); |
| 362 | else |
| 363 | result = my_qual; |
| 364 | |
| 365 | /* |
| 366 | * Change Vars to have partition's attnos instead of the parent's. We do |
| 367 | * this after we concatenate the parent's quals, because we want every Var |
| 368 | * in it to bear this relation's attnos. It's safe to assume varno = 1 |
| 369 | * here. |
| 370 | */ |
| 371 | result = map_partition_varattnos(result, 1, rel, parent, |
| 372 | &found_whole_row); |
| 373 | /* There can never be a whole-row reference here */ |
| 374 | if (found_whole_row) |
| 375 | elog(ERROR, "unexpected whole-row reference found in partition key" ); |
| 376 | |
| 377 | /* Assert that we're not leaking any old data during assignments below */ |
| 378 | Assert(rel->rd_partcheckcxt == NULL); |
| 379 | Assert(rel->rd_partcheck == NIL); |
| 380 | |
| 381 | /* |
| 382 | * Save a copy in the relcache. The order of these operations is fairly |
| 383 | * critical to avoid memory leaks and ensure that we don't leave a corrupt |
| 384 | * relcache entry if we fail partway through copyObject. |
| 385 | * |
| 386 | * If, as is definitely possible, the partcheck list is NIL, then we do |
| 387 | * not need to make a context to hold it. |
| 388 | */ |
| 389 | if (result != NIL) |
| 390 | { |
| 391 | rel->rd_partcheckcxt = AllocSetContextCreate(CacheMemoryContext, |
| 392 | "partition constraint" , |
| 393 | ALLOCSET_SMALL_SIZES); |
| 394 | MemoryContextCopyAndSetIdentifier(rel->rd_partcheckcxt, |
| 395 | RelationGetRelationName(rel)); |
| 396 | oldcxt = MemoryContextSwitchTo(rel->rd_partcheckcxt); |
| 397 | rel->rd_partcheck = copyObject(result); |
| 398 | MemoryContextSwitchTo(oldcxt); |
| 399 | } |
| 400 | else |
| 401 | rel->rd_partcheck = NIL; |
| 402 | rel->rd_partcheckvalid = true; |
| 403 | |
| 404 | /* Keep the parent locked until commit */ |
| 405 | relation_close(parent, NoLock); |
| 406 | |
| 407 | /* Return the working copy to the caller */ |
| 408 | return result; |
| 409 | } |
| 410 | |