Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * pgpa_trove.c
4 : * All of the advice given for a particular query, appropriately
5 : * organized for convenient access.
6 : *
7 : * This name comes from the English expression "trove of advice", which
8 : * means a collection of wisdom. This slightly unusual term is chosen
9 : * partly because it seems to fit and partly because it's not presently
10 : * used for anything else, making it easy to grep. Note that, while we
11 : * don't know whether the provided advice is actually wise, it's not our
12 : * job to question the user's choices.
13 : *
14 : * The goal of this module is to make it easy to locate the specific
15 : * bits of advice that pertain to any given part of a query, or to
16 : * determine that there are none.
17 : *
18 : * Copyright (c) 2016-2026, PostgreSQL Global Development Group
19 : *
20 : * contrib/pg_plan_advice/pgpa_trove.c
21 : *
22 : *-------------------------------------------------------------------------
23 : */
24 : #include "postgres.h"
25 :
26 : #include "pgpa_trove.h"
27 :
28 : #include "common/hashfn_unstable.h"
29 :
30 : /*
31 : * An advice trove is organized into a series of "slices", each of which
32 : * contains information about one topic e.g. scan methods. Each slice consists
33 : * of an array of trove entries plus a hash table that we can use to determine
34 : * which ones are relevant to a particular part of the query.
35 : */
36 : typedef struct pgpa_trove_slice
37 : {
38 : unsigned nallocated;
39 : unsigned nused;
40 : pgpa_trove_entry *entries;
41 : struct pgpa_trove_entry_hash *hash;
42 : } pgpa_trove_slice;
43 :
44 : /*
45 : * Scan advice is stored into 'scan'; join advice is stored into 'join'; and
46 : * advice that can apply to both cases is stored into 'rel'. This lets callers
47 : * ask just for what's relevant. These slices correspond to the possible values
48 : * of pgpa_trove_lookup_type.
49 : */
50 : struct pgpa_trove
51 : {
52 : pgpa_trove_slice join;
53 : pgpa_trove_slice rel;
54 : pgpa_trove_slice scan;
55 : };
56 :
57 : /*
58 : * We're going to build a hash table to allow clients of this module to find
59 : * relevant advice for a given part of the query quickly. However, we're going
60 : * to use only three of the five key fields as hash keys. There are two reasons
61 : * for this.
62 : *
63 : * First, it's allowable to set partition_schema to NULL to match a partition
64 : * with the correct name in any schema.
65 : *
66 : * Second, we expect the "occurrence" and "partition_schema" portions of the
67 : * relation identifiers to be mostly uninteresting. Most of the time, the
68 : * occurrence field will be 1 and the partition_schema values will all be the
69 : * same. Even when there is some variation, the absolute number of entries
70 : * that have the same values for all three of these key fields should be
71 : * quite small.
72 : */
73 : typedef struct
74 : {
75 : const char *alias_name;
76 : const char *partition_name;
77 : const char *plan_name;
78 : } pgpa_trove_entry_key;
79 :
80 : typedef struct
81 : {
82 : pgpa_trove_entry_key key;
83 : int status;
84 : Bitmapset *indexes;
85 : } pgpa_trove_entry_element;
86 :
87 : static uint32 pgpa_trove_entry_hash_key(pgpa_trove_entry_key key);
88 :
89 : static inline bool
90 310 : pgpa_trove_entry_compare_key(pgpa_trove_entry_key a, pgpa_trove_entry_key b)
91 : {
92 310 : if (strcmp(a.alias_name, b.alias_name) != 0)
93 40 : return false;
94 :
95 270 : if (!strings_equal_or_both_null(a.partition_name, b.partition_name))
96 0 : return false;
97 :
98 270 : if (!strings_equal_or_both_null(a.plan_name, b.plan_name))
99 0 : return false;
100 :
101 270 : return true;
102 : }
103 :
104 : #define SH_PREFIX pgpa_trove_entry
105 : #define SH_ELEMENT_TYPE pgpa_trove_entry_element
106 : #define SH_KEY_TYPE pgpa_trove_entry_key
107 : #define SH_KEY key
108 : #define SH_HASH_KEY(tb, key) pgpa_trove_entry_hash_key(key)
109 : #define SH_EQUAL(tb, a, b) pgpa_trove_entry_compare_key(a, b)
110 : #define SH_SCOPE static inline
111 : #define SH_DECLARE
112 : #define SH_DEFINE
113 : #include "lib/simplehash.h"
114 :
115 : static void pgpa_init_trove_slice(pgpa_trove_slice *tslice);
116 : static void pgpa_trove_add_to_slice(pgpa_trove_slice *tslice,
117 : pgpa_advice_tag_type tag,
118 : pgpa_advice_target *target);
119 : static void pgpa_trove_add_to_hash(pgpa_trove_entry_hash *hash,
120 : pgpa_advice_target *target,
121 : int index);
122 : static Bitmapset *pgpa_trove_slice_lookup(pgpa_trove_slice *tslice,
123 : pgpa_identifier *rid);
124 :
125 : /*
126 : * Build a trove of advice from a list of advice items.
127 : *
128 : * Caller can obtain a list of advice items to pass to this function by
129 : * calling pgpa_parse().
130 : */
131 : pgpa_trove *
132 113 : pgpa_build_trove(List *advice_items)
133 : {
134 113 : pgpa_trove *trove = palloc_object(pgpa_trove);
135 :
136 113 : pgpa_init_trove_slice(&trove->join);
137 113 : pgpa_init_trove_slice(&trove->rel);
138 113 : pgpa_init_trove_slice(&trove->scan);
139 :
140 351 : foreach_ptr(pgpa_advice_item, item, advice_items)
141 : {
142 125 : switch (item->tag)
143 : {
144 19 : case PGPA_TAG_JOIN_ORDER:
145 : {
146 : pgpa_advice_target *target;
147 :
148 : /*
149 : * For most advice types, each element in the top-level
150 : * list is a separate target, but it's most convenient to
151 : * regard the entirety of a JOIN_ORDER specification as a
152 : * single target. Since it wasn't represented that way
153 : * during parsing, build a surrogate object now.
154 : */
155 19 : target = palloc0_object(pgpa_advice_target);
156 19 : target->ttype = PGPA_TARGET_ORDERED_LIST;
157 19 : target->children = item->targets;
158 :
159 19 : pgpa_trove_add_to_slice(&trove->join,
160 : item->tag, target);
161 : }
162 19 : break;
163 :
164 50 : case PGPA_TAG_BITMAP_HEAP_SCAN:
165 : case PGPA_TAG_INDEX_ONLY_SCAN:
166 : case PGPA_TAG_INDEX_SCAN:
167 : case PGPA_TAG_SEQ_SCAN:
168 : case PGPA_TAG_TID_SCAN:
169 :
170 : /*
171 : * Scan advice.
172 : */
173 151 : foreach_ptr(pgpa_advice_target, target, item->targets)
174 : {
175 : /*
176 : * For now, all of our scan types target single relations,
177 : * but in the future this might not be true, e.g. a custom
178 : * scan could replace a join.
179 : */
180 : Assert(target->ttype == PGPA_TARGET_IDENTIFIER);
181 51 : pgpa_trove_add_to_slice(&trove->scan,
182 : item->tag, target);
183 : }
184 50 : break;
185 :
186 34 : case PGPA_TAG_FOREIGN_JOIN:
187 : case PGPA_TAG_HASH_JOIN:
188 : case PGPA_TAG_MERGE_JOIN_MATERIALIZE:
189 : case PGPA_TAG_MERGE_JOIN_PLAIN:
190 : case PGPA_TAG_NESTED_LOOP_MATERIALIZE:
191 : case PGPA_TAG_NESTED_LOOP_MEMOIZE:
192 : case PGPA_TAG_NESTED_LOOP_PLAIN:
193 : case PGPA_TAG_SEMIJOIN_NON_UNIQUE:
194 : case PGPA_TAG_SEMIJOIN_UNIQUE:
195 :
196 : /*
197 : * Join strategy advice.
198 : */
199 102 : foreach_ptr(pgpa_advice_target, target, item->targets)
200 : {
201 34 : pgpa_trove_add_to_slice(&trove->join,
202 : item->tag, target);
203 : }
204 34 : break;
205 :
206 22 : case PGPA_TAG_PARTITIONWISE:
207 : case PGPA_TAG_GATHER:
208 : case PGPA_TAG_GATHER_MERGE:
209 : case PGPA_TAG_NO_GATHER:
210 :
211 : /*
212 : * Advice about a RelOptInfo relevant to both scans and joins.
213 : */
214 73 : foreach_ptr(pgpa_advice_target, target, item->targets)
215 : {
216 29 : pgpa_trove_add_to_slice(&trove->rel,
217 : item->tag, target);
218 : }
219 22 : break;
220 : }
221 : }
222 :
223 113 : return trove;
224 : }
225 :
226 : /*
227 : * Search a trove of advice for relevant entries.
228 : *
229 : * All parameters are input parameters except for *result, which is an output
230 : * parameter used to return results to the caller.
231 : */
232 : void
233 860 : pgpa_trove_lookup(pgpa_trove *trove, pgpa_trove_lookup_type type,
234 : int nrids, pgpa_identifier *rids, pgpa_trove_result *result)
235 : {
236 : pgpa_trove_slice *tslice;
237 : Bitmapset *indexes;
238 :
239 : Assert(nrids > 0);
240 :
241 860 : if (type == PGPA_TROVE_LOOKUP_SCAN)
242 254 : tslice = &trove->scan;
243 606 : else if (type == PGPA_TROVE_LOOKUP_JOIN)
244 176 : tslice = &trove->join;
245 : else
246 430 : tslice = &trove->rel;
247 :
248 860 : indexes = pgpa_trove_slice_lookup(tslice, &rids[0]);
249 1286 : for (int i = 1; i < nrids; ++i)
250 : {
251 : Bitmapset *other_indexes;
252 :
253 : /*
254 : * If the caller is asking about two relations that aren't part of the
255 : * same subquery, they've messed up.
256 : */
257 : Assert(strings_equal_or_both_null(rids[0].plan_name,
258 : rids[i].plan_name));
259 :
260 426 : other_indexes = pgpa_trove_slice_lookup(tslice, &rids[i]);
261 426 : indexes = bms_union(indexes, other_indexes);
262 : }
263 :
264 860 : result->entries = tslice->entries;
265 860 : result->indexes = indexes;
266 860 : }
267 :
268 : /*
269 : * Return all entries in a trove slice to the caller.
270 : *
271 : * The first two arguments are input arguments, and the remainder are output
272 : * arguments.
273 : */
274 : void
275 339 : pgpa_trove_lookup_all(pgpa_trove *trove, pgpa_trove_lookup_type type,
276 : pgpa_trove_entry **entries, int *nentries)
277 : {
278 : pgpa_trove_slice *tslice;
279 :
280 339 : if (type == PGPA_TROVE_LOOKUP_SCAN)
281 113 : tslice = &trove->scan;
282 226 : else if (type == PGPA_TROVE_LOOKUP_JOIN)
283 113 : tslice = &trove->join;
284 : else
285 113 : tslice = &trove->rel;
286 :
287 339 : *entries = tslice->entries;
288 339 : *nentries = tslice->nused;
289 339 : }
290 :
291 : /*
292 : * Convert a trove entry to an item of plan advice that would produce it.
293 : */
294 : char *
295 133 : pgpa_cstring_trove_entry(pgpa_trove_entry *entry)
296 : {
297 : StringInfoData buf;
298 :
299 133 : initStringInfo(&buf);
300 133 : appendStringInfo(&buf, "%s", pgpa_cstring_advice_tag(entry->tag));
301 :
302 : /* JOIN_ORDER tags are transformed by pgpa_build_trove; undo that here */
303 133 : if (entry->tag != PGPA_TAG_JOIN_ORDER)
304 114 : appendStringInfoChar(&buf, '(');
305 : else
306 : Assert(entry->target->ttype == PGPA_TARGET_ORDERED_LIST);
307 :
308 133 : pgpa_format_advice_target(&buf, entry->target);
309 :
310 133 : if (entry->target->itarget != NULL)
311 : {
312 21 : appendStringInfoChar(&buf, ' ');
313 21 : pgpa_format_index_target(&buf, entry->target->itarget);
314 : }
315 :
316 133 : if (entry->tag != PGPA_TAG_JOIN_ORDER)
317 114 : appendStringInfoChar(&buf, ')');
318 :
319 133 : return buf.data;
320 : }
321 :
322 : /*
323 : * Set PGPA_TE_* flags on a set of trove entries.
324 : */
325 : void
326 729 : pgpa_trove_set_flags(pgpa_trove_entry *entries, Bitmapset *indexes, int flags)
327 : {
328 729 : int i = -1;
329 :
330 877 : while ((i = bms_next_member(indexes, i)) >= 0)
331 : {
332 148 : pgpa_trove_entry *entry = &entries[i];
333 :
334 148 : entry->flags |= flags;
335 : }
336 729 : }
337 :
338 : /*
339 : * Append a string representation of the specified PGPA_TE_* flags to the
340 : * given StringInfo.
341 : */
342 : void
343 133 : pgpa_trove_append_flags(StringInfo buf, int flags)
344 : {
345 133 : if ((flags & PGPA_TE_MATCH_FULL) != 0)
346 : {
347 : Assert((flags & PGPA_TE_MATCH_PARTIAL) != 0);
348 113 : appendStringInfo(buf, "matched");
349 : }
350 20 : else if ((flags & PGPA_TE_MATCH_PARTIAL) != 0)
351 4 : appendStringInfo(buf, "partially matched");
352 : else
353 16 : appendStringInfo(buf, "not matched");
354 133 : if ((flags & PGPA_TE_INAPPLICABLE) != 0)
355 4 : appendStringInfo(buf, ", inapplicable");
356 133 : if ((flags & PGPA_TE_CONFLICTING) != 0)
357 14 : appendStringInfo(buf, ", conflicting");
358 133 : if ((flags & PGPA_TE_FAILED) != 0)
359 30 : appendStringInfo(buf, ", failed");
360 133 : }
361 :
362 : /*
363 : * Add a new advice target to an existing pgpa_trove_slice object.
364 : */
365 : static void
366 133 : pgpa_trove_add_to_slice(pgpa_trove_slice *tslice,
367 : pgpa_advice_tag_type tag,
368 : pgpa_advice_target *target)
369 : {
370 : pgpa_trove_entry *entry;
371 :
372 133 : if (tslice->nused >= tslice->nallocated)
373 : {
374 : int new_allocated;
375 :
376 0 : new_allocated = tslice->nallocated * 2;
377 0 : tslice->entries = repalloc_array(tslice->entries, pgpa_trove_entry,
378 : new_allocated);
379 0 : tslice->nallocated = new_allocated;
380 : }
381 :
382 133 : entry = &tslice->entries[tslice->nused];
383 133 : entry->tag = tag;
384 133 : entry->target = target;
385 133 : entry->flags = 0;
386 :
387 133 : pgpa_trove_add_to_hash(tslice->hash, target, tslice->nused);
388 :
389 133 : tslice->nused++;
390 133 : }
391 :
392 : /*
393 : * Update the hash table for a newly-added advice target.
394 : */
395 : static void
396 202 : pgpa_trove_add_to_hash(pgpa_trove_entry_hash *hash, pgpa_advice_target *target,
397 : int index)
398 : {
399 : pgpa_trove_entry_key key;
400 : pgpa_trove_entry_element *element;
401 : bool found;
402 :
403 : /* For non-identifiers, add entries for all descendants. */
404 202 : if (target->ttype != PGPA_TARGET_IDENTIFIER)
405 : {
406 135 : foreach_ptr(pgpa_advice_target, child_target, target->children)
407 : {
408 69 : pgpa_trove_add_to_hash(hash, child_target, index);
409 : }
410 33 : return;
411 : }
412 :
413 : /* Sanity checks. */
414 : Assert(target->rid.occurrence > 0);
415 : Assert(target->rid.alias_name != NULL);
416 :
417 : /* Add an entry for this relation identifier. */
418 169 : key.alias_name = target->rid.alias_name;
419 169 : key.partition_name = target->rid.partrel;
420 169 : key.plan_name = target->rid.plan_name;
421 169 : element = pgpa_trove_entry_insert(hash, key, &found);
422 169 : if (!found)
423 159 : element->indexes = NULL;
424 169 : element->indexes = bms_add_member(element->indexes, index);
425 : }
426 :
427 : /*
428 : * Create and initialize a new pgpa_trove_slice object.
429 : */
430 : static void
431 339 : pgpa_init_trove_slice(pgpa_trove_slice *tslice)
432 : {
433 : /*
434 : * In an ideal world, we'll make tslice->nallocated big enough that the
435 : * array and hash table will be large enough to contain the number of
436 : * advice items in this trove slice, but a generous default value is not
437 : * good for performance, because pgpa_init_trove_slice() has to zero an
438 : * amount of memory proportional to tslice->nallocated. Hence, we keep the
439 : * starting value quite small, on the theory that advice strings will
440 : * often be relatively short.
441 : */
442 339 : tslice->nallocated = 16;
443 339 : tslice->nused = 0;
444 339 : tslice->entries = palloc_array(pgpa_trove_entry, tslice->nallocated);
445 339 : tslice->hash = pgpa_trove_entry_create(CurrentMemoryContext,
446 : tslice->nallocated, NULL);
447 339 : }
448 :
449 : /*
450 : * Fast hash function for a key consisting of alias_name, partition_name,
451 : * and plan_name.
452 : */
453 : static uint32
454 1457 : pgpa_trove_entry_hash_key(pgpa_trove_entry_key key)
455 : {
456 : fasthash_state hs;
457 : int sp_len;
458 :
459 1457 : fasthash_init(&hs, 0);
460 :
461 : /* alias_name may not be NULL */
462 1457 : sp_len = fasthash_accum_cstring(&hs, key.alias_name);
463 :
464 : /* partition_name and plan_name, however, can be NULL */
465 1457 : if (key.partition_name != NULL)
466 385 : sp_len += fasthash_accum_cstring(&hs, key.partition_name);
467 1457 : if (key.plan_name != NULL)
468 18 : sp_len += fasthash_accum_cstring(&hs, key.plan_name);
469 :
470 : /*
471 : * hashfn_unstable.h recommends using string length as tweak. It's not
472 : * clear to me what to do if there are multiple strings, so for now I'm
473 : * just using the total of all of the lengths.
474 : */
475 1457 : return fasthash_final32(&hs, sp_len);
476 : }
477 :
478 : /*
479 : * Look for matching entries.
480 : */
481 : static Bitmapset *
482 1286 : pgpa_trove_slice_lookup(pgpa_trove_slice *tslice, pgpa_identifier *rid)
483 : {
484 : pgpa_trove_entry_key key;
485 : pgpa_trove_entry_element *element;
486 1286 : Bitmapset *result = NULL;
487 :
488 : Assert(rid->occurrence >= 1);
489 :
490 1286 : key.alias_name = rid->alias_name;
491 1286 : key.partition_name = rid->partrel;
492 1286 : key.plan_name = rid->plan_name;
493 :
494 1286 : element = pgpa_trove_entry_lookup(tslice->hash, key);
495 :
496 1286 : if (element != NULL)
497 : {
498 260 : int i = -1;
499 :
500 535 : while ((i = bms_next_member(element->indexes, i)) >= 0)
501 : {
502 275 : pgpa_trove_entry *entry = &tslice->entries[i];
503 :
504 : /*
505 : * We know that this target or one of its descendants matches the
506 : * identifier on the three key fields above, but we don't know
507 : * which descendant or whether the occurrence and schema also
508 : * match.
509 : */
510 275 : if (pgpa_identifier_matches_target(rid, entry->target))
511 271 : result = bms_add_member(result, i);
512 : }
513 : }
514 :
515 1286 : return result;
516 : }
|
