| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* nodeIncrementalSort.c |
| 4 |
* Routines to handle incremental sorting of relations. |
| 5 |
* |
| 6 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 7 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 8 |
* |
| 9 |
* IDENTIFICATION |
| 10 |
* src/backend/executor/nodeIncrementalSort.c |
| 11 |
* |
| 12 |
* DESCRIPTION |
| 13 |
* |
| 14 |
* Incremental sort is an optimized variant of multikey sort for cases |
| 15 |
* when the input is already sorted by a prefix of the sort keys. For |
| 16 |
* example when a sort by (key1, key2 ... keyN) is requested, and the |
| 17 |
* input is already sorted by (key1, key2 ... keyM), M < N, we can |
| 18 |
* divide the input into groups where keys (key1, ... keyM) are equal, |
| 19 |
* and only sort on the remaining columns. |
| 20 |
* |
| 21 |
* Consider the following example. We have input tuples consisting of |
| 22 |
* two integers (X, Y) already presorted by X, while it's required to |
| 23 |
* sort them by both X and Y. Let input tuples be following. |
| 24 |
* |
| 25 |
* (1, 5) |
| 26 |
* (1, 2) |
| 27 |
* (2, 9) |
| 28 |
* (2, 1) |
| 29 |
* (2, 5) |
| 30 |
* (3, 3) |
| 31 |
* (3, 7) |
| 32 |
* |
| 33 |
* An incremental sort algorithm would split the input into the following |
| 34 |
* groups, which have equal X, and then sort them by Y individually: |
| 35 |
* |
| 36 |
* (1, 5) (1, 2) |
| 37 |
* (2, 9) (2, 1) (2, 5) |
| 38 |
* (3, 3) (3, 7) |
| 39 |
* |
| 40 |
* After sorting these groups and putting them altogether, we would get |
| 41 |
* the following result which is sorted by X and Y, as requested: |
| 42 |
* |
| 43 |
* (1, 2) |
| 44 |
* (1, 5) |
| 45 |
* (2, 1) |
| 46 |
* (2, 5) |
| 47 |
* (2, 9) |
| 48 |
* (3, 3) |
| 49 |
* (3, 7) |
| 50 |
* |
| 51 |
* Incremental sort may be more efficient than plain sort, particularly |
| 52 |
* on large datasets, as it reduces the amount of data to sort at once, |
| 53 |
* making it more likely it fits into work_mem (eliminating the need to |
| 54 |
* spill to disk). But the main advantage of incremental sort is that |
| 55 |
* it can start producing rows early, before sorting the whole dataset, |
| 56 |
* which is a significant benefit especially for queries with LIMIT. |
| 57 |
* |
| 58 |
* The algorithm we've implemented here is modified from the theoretical |
| 59 |
* base described above by operating in two different modes: |
| 60 |
* - Fetching a minimum number of tuples without checking prefix key |
| 61 |
* group membership and sorting on all columns when safe. |
| 62 |
* - Fetching all tuples for a single prefix key group and sorting on |
| 63 |
* solely the unsorted columns. |
| 64 |
* We always begin in the first mode, and employ a heuristic to switch |
| 65 |
* into the second mode if we believe it's beneficial. |
| 66 |
* |
| 67 |
* Sorting incrementally can potentially use less memory, avoid fetching |
| 68 |
* and sorting all tuples in the dataset, and begin returning tuples before |
| 69 |
* the entire result set is available. |
| 70 |
* |
| 71 |
* The hybrid mode approach allows us to optimize for both very small |
| 72 |
* groups (where the overhead of a new tuplesort is high) and very large |
| 73 |
* groups (where we can lower cost by not having to sort on already sorted |
| 74 |
* columns), albeit at some extra cost while switching between modes. |
| 75 |
* |
| 76 |
*------------------------------------------------------------------------- |
| 77 |
*/ |
| 78 |
|
| 79 |
#include "postgres.h" |
| 80 |
|
| 81 |
#include "executor/execdebug.h" |
| 82 |
#include "executor/nodeIncrementalSort.h" |
| 83 |
#include "miscadmin.h" |
| 84 |
#include "utils/lsyscache.h" |
| 85 |
#include "utils/tuplesort.h" |
| 86 |
|
| 87 |
/* |
| 88 |
* We need to store the instrumentation information in either local node's sort |
| 89 |
* info or, for a parallel worker process, in the shared info (this avoids |
| 90 |
* having to additionally memcpy the info from local memory to shared memory |
| 91 |
* at each instrumentation call). This macro expands to choose the proper sort |
| 92 |
* state and group info. |
| 93 |
* |
| 94 |
* Arguments: |
| 95 |
* - node: type IncrementalSortState * |
| 96 |
* - groupName: the token fullsort or prefixsort |
| 97 |
*/ |
| 98 |
#define INSTRUMENT_SORT_GROUP(node, groupName) \ |
| 99 |
do { \ |
| 100 |
if ((node)->ss.ps.instrument != NULL) \ |
| 101 |
{ \ |
| 102 |
if ((node)->shared_info && (node)->am_worker) \ |
| 103 |
{ \ |
| 104 |
Assert(IsParallelWorker()); \ |
| 105 |
Assert(ParallelWorkerNumber < (node)->shared_info->num_workers); \ |
| 106 |
instrumentSortedGroup(&(node)->shared_info->sinfo[ParallelWorkerNumber].groupName##GroupInfo, \ |
| 107 |
(node)->groupName##_state); \ |
| 108 |
} \ |
| 109 |
else \ |
| 110 |
{ \ |
| 111 |
instrumentSortedGroup(&(node)->incsort_info.groupName##GroupInfo, \ |
| 112 |
(node)->groupName##_state); \ |
| 113 |
} \ |
| 114 |
} \ |
| 115 |
} while (0) |
| 116 |
|
| 117 |
|
| 118 |
/* ---------------------------------------------------------------- |
| 119 |
* instrumentSortedGroup |
| 120 |
* |
| 121 |
* Because incremental sort processes (potentially many) sort batches, we need |
| 122 |
* to capture tuplesort stats each time we finalize a sort state. This summary |
| 123 |
* data is later used for EXPLAIN ANALYZE output. |
| 124 |
* ---------------------------------------------------------------- |
| 125 |
*/ |
| 126 |
static void |
| 127 |
instrumentSortedGroup(IncrementalSortGroupInfo *groupInfo, |
| 128 |
Tuplesortstate *sortState) |
| 129 |
{ |
| 130 |
TuplesortInstrumentation sort_instr; |
| 131 |
|
| 132 |
groupInfo->groupCount++; |
| 133 |
|
| 134 |
tuplesort_get_stats(sortState, &sort_instr); |
| 135 |
|
| 136 |
/* Calculate total and maximum memory and disk space used. */ |
| 137 |
switch (sort_instr.spaceType) |
| 138 |
{ |
| 139 |
case SORT_SPACE_TYPE_DISK: |
| 140 |
groupInfo->totalDiskSpaceUsed += sort_instr.spaceUsed; |
| 141 |
if (sort_instr.spaceUsed > groupInfo->maxDiskSpaceUsed) |
| 142 |
groupInfo->maxDiskSpaceUsed = sort_instr.spaceUsed; |
| 143 |
|
| 144 |
break; |
| 145 |
case SORT_SPACE_TYPE_MEMORY: |
| 146 |
groupInfo->totalMemorySpaceUsed += sort_instr.spaceUsed; |
| 147 |
if (sort_instr.spaceUsed > groupInfo->maxMemorySpaceUsed) |
| 148 |
groupInfo->maxMemorySpaceUsed = sort_instr.spaceUsed; |
| 149 |
|
| 150 |
break; |
| 151 |
} |
| 152 |
|
| 153 |
/* Track each sort method we've used. */ |
| 154 |
groupInfo->sortMethods |= sort_instr.sortMethod; |
| 155 |
} |
| 156 |
|
| 157 |
/* ---------------------------------------------------------------- |
| 158 |
* preparePresortedCols |
| 159 |
* |
| 160 |
* Prepare information for presorted_keys comparisons. |
| 161 |
* ---------------------------------------------------------------- |
| 162 |
*/ |
| 163 |
static void |
| 164 |
preparePresortedCols(IncrementalSortState *node) |
| 165 |
{ |
| 166 |
IncrementalSort *plannode = castNode(IncrementalSort, node->ss.ps.plan); |
| 167 |
|
| 168 |
node->presorted_keys = |
| 169 |
(PresortedKeyData *) palloc(plannode->nPresortedCols * |
| 170 |
sizeof(PresortedKeyData)); |
| 171 |
|
| 172 |
/* Pre-cache comparison functions for each pre-sorted key. */ |
| 173 |
for (int i = 0; i < plannode->nPresortedCols; i++) |
| 174 |
{ |
| 175 |
Oid equalityOp, |
| 176 |
equalityFunc; |
| 177 |
PresortedKeyData *key; |
| 178 |
|
| 179 |
key = &node->presorted_keys[i]; |
| 180 |
key->attno = plannode->sort.sortColIdx[i]; |
| 181 |
|
| 182 |
equalityOp = get_equality_op_for_ordering_op(plannode->sort.sortOperators[i], |
| 183 |
NULL); |
| 184 |
if (!OidIsValid(equalityOp)) |
| 185 |
elog(ERROR, "missing equality operator for ordering operator %u", |
| 186 |
plannode->sort.sortOperators[i]); |
| 187 |
|
| 188 |
equalityFunc = get_opcode(equalityOp); |
| 189 |
if (!OidIsValid(equalityFunc)) |
| 190 |
elog(ERROR, "missing function for operator %u", equalityOp); |
| 191 |
|
| 192 |
/* Lookup the comparison function */ |
| 193 |
fmgr_info_cxt(equalityFunc, &key->flinfo, CurrentMemoryContext); |
| 194 |
|
| 195 |
/* We can initialize the callinfo just once and re-use it */ |
| 196 |
key->fcinfo = palloc0(SizeForFunctionCallInfo(2)); |
| 197 |
InitFunctionCallInfoData(*key->fcinfo, &key->flinfo, 2, |
| 198 |
plannode->sort.collations[i], NULL, NULL); |
| 199 |
key->fcinfo->args[0].isnull = false; |
| 200 |
key->fcinfo->args[1].isnull = false; |
| 201 |
} |
| 202 |
} |
| 203 |
|
| 204 |
/* ---------------------------------------------------------------- |
| 205 |
* isCurrentGroup |
| 206 |
* |
| 207 |
* Check whether a given tuple belongs to the current sort group by comparing |
| 208 |
* the presorted column values to the pivot tuple of the current group. |
| 209 |
* ---------------------------------------------------------------- |
| 210 |
*/ |
| 211 |
static bool |
| 212 |
isCurrentGroup(IncrementalSortState *node, TupleTableSlot *pivot, TupleTableSlot *tuple) |
| 213 |
{ |
| 214 |
int nPresortedCols; |
| 215 |
|
| 216 |
nPresortedCols = castNode(IncrementalSort, node->ss.ps.plan)->nPresortedCols; |
| 217 |
|
| 218 |
/* |
| 219 |
* That the input is sorted by keys * (0, ... n) implies that the tail |
| 220 |
* keys are more likely to change. Therefore we do our comparison starting |
| 221 |
* from the last pre-sorted column to optimize for early detection of |
| 222 |
* inequality and minimizing the number of function calls.. |
| 223 |
*/ |
| 224 |
for (int i = nPresortedCols - 1; i >= 0; i--) |
| 225 |
{ |
| 226 |
Datum datumA, |
| 227 |
datumB, |
| 228 |
result; |
| 229 |
bool isnullA, |
| 230 |
isnullB; |
| 231 |
AttrNumber attno = node->presorted_keys[i].attno; |
| 232 |
PresortedKeyData *key; |
| 233 |
|
| 234 |
datumA = slot_getattr(pivot, attno, &isnullA); |
| 235 |
datumB = slot_getattr(tuple, attno, &isnullB); |
| 236 |
|
| 237 |
/* Special case for NULL-vs-NULL, else use standard comparison */ |
| 238 |
if (isnullA || isnullB) |
| 239 |
{ |
| 240 |
if (isnullA == isnullB) |
| 241 |
continue; |
| 242 |
else |
| 243 |
return false; |
| 244 |
} |
| 245 |
|
| 246 |
key = &node->presorted_keys[i]; |
| 247 |
|
| 248 |
key->fcinfo->args[0].value = datumA; |
| 249 |
key->fcinfo->args[1].value = datumB; |
| 250 |
|
| 251 |
/* just for paranoia's sake, we reset isnull each time */ |
| 252 |
key->fcinfo->isnull = false; |
| 253 |
|
| 254 |
result = FunctionCallInvoke(key->fcinfo); |
| 255 |
|
| 256 |
/* Check for null result, since caller is clearly not expecting one */ |
| 257 |
if (key->fcinfo->isnull) |
| 258 |
elog(ERROR, "function %u returned NULL", key->flinfo.fn_oid); |
| 259 |
|
| 260 |
if (!DatumGetBool(result)) |
| 261 |
return false; |
| 262 |
} |
| 263 |
return true; |
| 264 |
} |
| 265 |
|
| 266 |
/* ---------------------------------------------------------------- |
| 267 |
* switchToPresortedPrefixMode |
| 268 |
* |
| 269 |
* When we determine that we've likely encountered a large batch of tuples all |
| 270 |
* having the same presorted prefix values, we want to optimize tuplesort by |
| 271 |
* only sorting on unsorted suffix keys. |
| 272 |
* |
| 273 |
* The problem is that we've already accumulated several tuples in another |
| 274 |
* tuplesort configured to sort by all columns (assuming that there may be |
| 275 |
* more than one prefix key group). So to switch to presorted prefix mode we |
| 276 |
* have to go back and look at all the tuples we've already accumulated to |
| 277 |
* verify they're all part of the same prefix key group before sorting them |
| 278 |
* solely by unsorted suffix keys. |
| 279 |
* |
| 280 |
* While it's likely that all tuples already fetched are all part of a single |
| 281 |
* prefix group, we also have to handle the possibility that there is at least |
| 282 |
* one different prefix key group before the large prefix key group. |
| 283 |
* ---------------------------------------------------------------- |
| 284 |
*/ |
| 285 |
static void |
| 286 |
switchToPresortedPrefixMode(PlanState *pstate) |
| 287 |
{ |
| 288 |
IncrementalSortState *node = castNode(IncrementalSortState, pstate); |
| 289 |
ScanDirection dir; |
| 290 |
int64 nTuples; |
| 291 |
TupleDesc tupDesc; |
| 292 |
PlanState *outerNode; |
| 293 |
IncrementalSort *plannode = castNode(IncrementalSort, node->ss.ps.plan); |
| 294 |
|
| 295 |
dir = node->ss.ps.state->es_direction; |
| 296 |
outerNode = outerPlanState(node); |
| 297 |
tupDesc = ExecGetResultType(outerNode); |
| 298 |
|
| 299 |
/* Configure the prefix sort state the first time around. */ |
| 300 |
if (node->prefixsort_state == NULL) |
| 301 |
{ |
| 302 |
Tuplesortstate *prefixsort_state; |
| 303 |
int nPresortedCols = plannode->nPresortedCols; |
| 304 |
|
| 305 |
/* |
| 306 |
* Optimize the sort by assuming the prefix columns are all equal and |
| 307 |
* thus we only need to sort by any remaining columns. |
| 308 |
*/ |
| 309 |
prefixsort_state = tuplesort_begin_heap(tupDesc, |
| 310 |
plannode->sort.numCols - nPresortedCols, |
| 311 |
&(plannode->sort.sortColIdx[nPresortedCols]), |
| 312 |
&(plannode->sort.sortOperators[nPresortedCols]), |
| 313 |
&(plannode->sort.collations[nPresortedCols]), |
| 314 |
&(plannode->sort.nullsFirst[nPresortedCols]), |
| 315 |
work_mem, |
| 316 |
NULL, |
| 317 |
node->bounded ? TUPLESORT_ALLOWBOUNDED : TUPLESORT_NONE); |
| 318 |
node->prefixsort_state = prefixsort_state; |
| 319 |
} |
| 320 |
else |
| 321 |
{ |
| 322 |
/* Next group of presorted data */ |
| 323 |
tuplesort_reset(node->prefixsort_state); |
| 324 |
} |
| 325 |
|
| 326 |
/* |
| 327 |
* If the current node has a bound, then it's reasonably likely that a |
| 328 |
* large prefix key group will benefit from bounded sort, so configure the |
| 329 |
* tuplesort to allow for that optimization. |
| 330 |
*/ |
| 331 |
if (node->bounded) |
| 332 |
{ |
| 333 |
SO1_printf("Setting bound on presorted prefix tuplesort to: " INT64_FORMAT "\n", |
| 334 |
node->bound - node->bound_Done); |
| 335 |
tuplesort_set_bound(node->prefixsort_state, |
| 336 |
node->bound - node->bound_Done); |
| 337 |
} |
| 338 |
|
| 339 |
/* |
| 340 |
* Copy as many tuples as we can (i.e., in the same prefix key group) from |
| 341 |
* the full sort state to the prefix sort state. |
| 342 |
*/ |
| 343 |
for (nTuples = 0; nTuples < node->n_fullsort_remaining; nTuples++) |
| 344 |
{ |
| 345 |
/* |
| 346 |
* When we encounter multiple prefix key groups inside the full sort |
| 347 |
* tuplesort we have to carry over the last read tuple into the next |
| 348 |
* batch. |
| 349 |
*/ |
| 350 |
if (nTuples == 0 && !TupIsNull(node->transfer_tuple)) |
| 351 |
{ |
| 352 |
tuplesort_puttupleslot(node->prefixsort_state, node->transfer_tuple); |
| 353 |
/* The carried over tuple is our new group pivot tuple. */ |
| 354 |
ExecCopySlot(node->group_pivot, node->transfer_tuple); |
| 355 |
} |
| 356 |
else |
| 357 |
{ |
| 358 |
tuplesort_gettupleslot(node->fullsort_state, |
| 359 |
ScanDirectionIsForward(dir), |
| 360 |
false, node->transfer_tuple, NULL); |
| 361 |
|
| 362 |
/* |
| 363 |
* If this is our first time through the loop, then we need to |
| 364 |
* save the first tuple we get as our new group pivot. |
| 365 |
*/ |
| 366 |
if (TupIsNull(node->group_pivot)) |
| 367 |
ExecCopySlot(node->group_pivot, node->transfer_tuple); |
| 368 |
|
| 369 |
if (isCurrentGroup(node, node->group_pivot, node->transfer_tuple)) |
| 370 |
{ |
| 371 |
tuplesort_puttupleslot(node->prefixsort_state, node->transfer_tuple); |
| 372 |
} |
| 373 |
else |
| 374 |
{ |
| 375 |
/* |
| 376 |
* The tuple isn't part of the current batch so we need to |
| 377 |
* carry it over into the next batch of tuples we transfer out |
| 378 |
* of the full sort tuplesort into the presorted prefix |
| 379 |
* tuplesort. We don't actually have to do anything special to |
| 380 |
* save the tuple since we've already loaded it into the |
| 381 |
* node->transfer_tuple slot, and, even though that slot |
| 382 |
* points to memory inside the full sort tuplesort, we can't |
| 383 |
* reset that tuplesort anyway until we've fully transferred |
| 384 |
* out its tuples, so this reference is safe. We do need to |
| 385 |
* reset the group pivot tuple though since we've finished the |
| 386 |
* current prefix key group. |
| 387 |
*/ |
| 388 |
ExecClearTuple(node->group_pivot); |
| 389 |
|
| 390 |
/* Break out of for-loop early */ |
| 391 |
break; |
| 392 |
} |
| 393 |
} |
| 394 |
} |
| 395 |
|
| 396 |
/* |
| 397 |
* Track how many tuples remain in the full sort batch so that we know if |
| 398 |
* we need to sort multiple prefix key groups before processing tuples |
| 399 |
* remaining in the large single prefix key group we think we've |
| 400 |
* encountered. |
| 401 |
*/ |
| 402 |
SO1_printf("Moving " INT64_FORMAT " tuples to presorted prefix tuplesort\n", nTuples); |
| 403 |
node->n_fullsort_remaining -= nTuples; |
| 404 |
SO1_printf("Setting n_fullsort_remaining to " INT64_FORMAT "\n", node->n_fullsort_remaining); |
| 405 |
|
| 406 |
if (node->n_fullsort_remaining == 0) |
| 407 |
{ |
| 408 |
/* |
| 409 |
* We've found that all tuples remaining in the full sort batch are in |
| 410 |
* the same prefix key group and moved all of those tuples into the |
| 411 |
* presorted prefix tuplesort. We don't know that we've yet found the |
| 412 |
* last tuple in the current prefix key group, so save our pivot |
| 413 |
* comparison tuple and continue fetching tuples from the outer |
| 414 |
* execution node to load into the presorted prefix tuplesort. |
| 415 |
*/ |
| 416 |
ExecCopySlot(node->group_pivot, node->transfer_tuple); |
| 417 |
SO_printf("Setting execution_status to INCSORT_LOADPREFIXSORT (switchToPresortedPrefixMode)\n"); |
| 418 |
node->execution_status = INCSORT_LOADPREFIXSORT; |
| 419 |
|
| 420 |
/* |
| 421 |
* Make sure we clear the transfer tuple slot so that next time we |
| 422 |
* encounter a large prefix key group we don't incorrectly assume we |
| 423 |
* have a tuple carried over from the previous group. |
| 424 |
*/ |
| 425 |
ExecClearTuple(node->transfer_tuple); |
| 426 |
} |
| 427 |
else |
| 428 |
{ |
| 429 |
/* |
| 430 |
* We finished a group but didn't consume all of the tuples from the |
| 431 |
* full sort state, so we'll sort this batch, let the outer node read |
| 432 |
* out all of those tuples, and then come back around to find another |
| 433 |
* batch. |
| 434 |
*/ |
| 435 |
SO1_printf("Sorting presorted prefix tuplesort with " INT64_FORMAT " tuples\n", nTuples); |
| 436 |
tuplesort_performsort(node->prefixsort_state); |
| 437 |
|
| 438 |
INSTRUMENT_SORT_GROUP(node, prefixsort); |
| 439 |
|
| 440 |
if (node->bounded) |
| 441 |
{ |
| 442 |
/* |
| 443 |
* If the current node has a bound and we've already sorted n |
| 444 |
* tuples, then the functional bound remaining is (original bound |
| 445 |
* - n), so store the current number of processed tuples for use |
| 446 |
* in configuring sorting bound. |
| 447 |
*/ |
| 448 |
SO2_printf("Changing bound_Done from " INT64_FORMAT " to " INT64_FORMAT "\n", |
| 449 |
Min(node->bound, node->bound_Done + nTuples), node->bound_Done); |
| 450 |
node->bound_Done = Min(node->bound, node->bound_Done + nTuples); |
| 451 |
} |
| 452 |
|
| 453 |
SO_printf("Setting execution_status to INCSORT_READPREFIXSORT (switchToPresortedPrefixMode)\n"); |
| 454 |
node->execution_status = INCSORT_READPREFIXSORT; |
| 455 |
} |
| 456 |
} |
| 457 |
|
| 458 |
/* |
| 459 |
* Sorting many small groups with tuplesort is inefficient. In order to |
| 460 |
* cope with this problem we don't start a new group until the current one |
| 461 |
* contains at least DEFAULT_MIN_GROUP_SIZE tuples (unfortunately this also |
| 462 |
* means we can't assume small groups of tuples all have the same prefix keys.) |
| 463 |
* When we have a bound that's less than DEFAULT_MIN_GROUP_SIZE we start looking |
| 464 |
* for the new group as soon as we've met our bound to avoid fetching more |
| 465 |
* tuples than we absolutely have to fetch. |
| 466 |
*/ |
| 467 |
#define DEFAULT_MIN_GROUP_SIZE 32 |
| 468 |
|
| 469 |
/* |
| 470 |
* While we've optimized for small prefix key groups by not starting our prefix |
| 471 |
* key comparisons until we've reached a minimum number of tuples, we don't want |
| 472 |
* that optimization to cause us to lose out on the benefits of being able to |
| 473 |
* assume a large group of tuples is fully presorted by its prefix keys. |
| 474 |
* Therefore we use the DEFAULT_MAX_FULL_SORT_GROUP_SIZE cutoff as a heuristic |
| 475 |
* for determining when we believe we've encountered a large group, and, if we |
| 476 |
* get to that point without finding a new prefix key group we transition to |
| 477 |
* presorted prefix key mode. |
| 478 |
*/ |
| 479 |
#define DEFAULT_MAX_FULL_SORT_GROUP_SIZE (2 * DEFAULT_MIN_GROUP_SIZE) |
| 480 |
|
| 481 |
/* ---------------------------------------------------------------- |
| 482 |
* ExecIncrementalSort |
| 483 |
* |
| 484 |
* Assuming that outer subtree returns tuple presorted by some prefix |
| 485 |
* of target sort columns, performs incremental sort. |
| 486 |
* |
| 487 |
* Conditions: |
| 488 |
* -- none. |
| 489 |
* |
| 490 |
* Initial States: |
| 491 |
* -- the outer child is prepared to return the first tuple. |
| 492 |
* ---------------------------------------------------------------- |
| 493 |
*/ |
| 494 |
static TupleTableSlot * |
| 495 |
ExecIncrementalSort(PlanState *pstate) |
| 496 |
{ |
| 497 |
IncrementalSortState *node = castNode(IncrementalSortState, pstate); |
| 498 |
EState *estate; |
| 499 |
ScanDirection dir; |
| 500 |
Tuplesortstate *read_sortstate; |
| 501 |
Tuplesortstate *fullsort_state; |
| 502 |
TupleTableSlot *slot; |
| 503 |
IncrementalSort *plannode = (IncrementalSort *) node->ss.ps.plan; |
| 504 |
PlanState *outerNode; |
| 505 |
TupleDesc tupDesc; |
| 506 |
int64 nTuples = 0; |
| 507 |
int64 minGroupSize; |
| 508 |
|
| 509 |
CHECK_FOR_INTERRUPTS(); |
| 510 |
|
| 511 |
estate = node->ss.ps.state; |
| 512 |
dir = estate->es_direction; |
| 513 |
fullsort_state = node->fullsort_state; |
| 514 |
|
| 515 |
/* |
| 516 |
* If a previous iteration has sorted a batch, then we need to check to |
| 517 |
* see if there are any remaining tuples in that batch that we can return |
| 518 |
* before moving on to other execution states. |
| 519 |
*/ |
| 520 |
if (node->execution_status == INCSORT_READFULLSORT |
| 521 |
|| node->execution_status == INCSORT_READPREFIXSORT) |
| 522 |
{ |
| 523 |
/* |
| 524 |
* Return next tuple from the current sorted group set if available. |
| 525 |
*/ |
| 526 |
read_sortstate = node->execution_status == INCSORT_READFULLSORT ? |
| 527 |
fullsort_state : node->prefixsort_state; |
| 528 |
slot = node->ss.ps.ps_ResultTupleSlot; |
| 529 |
|
| 530 |
/* |
| 531 |
* We have to populate the slot from the tuplesort before checking |
| 532 |
* outerNodeDone because it will set the slot to NULL if no more |
| 533 |
* tuples remain. If the tuplesort is empty, but we don't have any |
| 534 |
* more tuples available for sort from the outer node, then |
| 535 |
* outerNodeDone will have been set so we'll return that now-empty |
| 536 |
* slot to the caller. |
| 537 |
*/ |
| 538 |
if (tuplesort_gettupleslot(read_sortstate, ScanDirectionIsForward(dir), |
| 539 |
false, slot, NULL) || node->outerNodeDone) |
| 540 |
|
| 541 |
/* |
| 542 |
* Note: there isn't a good test case for the node->outerNodeDone |
| 543 |
* check directly, but we need it for any plan where the outer |
| 544 |
* node will fail when trying to fetch too many tuples. |
| 545 |
*/ |
| 546 |
return slot; |
| 547 |
else if (node->n_fullsort_remaining > 0) |
| 548 |
{ |
| 549 |
/* |
| 550 |
* When we transition to presorted prefix mode, we might have |
| 551 |
* accumulated at least one additional prefix key group in the |
| 552 |
* full sort tuplesort. The first call to |
| 553 |
* switchToPresortedPrefixMode() will have pulled the first one of |
| 554 |
* those groups out, and we've returned those tuples to the parent |
| 555 |
* node, but if at this point we still have tuples remaining in |
| 556 |
* the full sort state (i.e., n_fullsort_remaining > 0), then we |
| 557 |
* need to re-execute the prefix mode transition function to pull |
| 558 |
* out the next prefix key group. |
| 559 |
*/ |
| 560 |
SO1_printf("Re-calling switchToPresortedPrefixMode() because n_fullsort_remaining is > 0 (" INT64_FORMAT ")\n", |
| 561 |
node->n_fullsort_remaining); |
| 562 |
switchToPresortedPrefixMode(pstate); |
| 563 |
} |
| 564 |
else |
| 565 |
{ |
| 566 |
/* |
| 567 |
* If we don't have any sorted tuples to read and we're not |
| 568 |
* currently transitioning into presorted prefix sort mode, then |
| 569 |
* it's time to start the process all over again by building a new |
| 570 |
* group in the full sort state. |
| 571 |
*/ |
| 572 |
SO_printf("Setting execution_status to INCSORT_LOADFULLSORT (n_fullsort_remaining > 0)\n"); |
| 573 |
node->execution_status = INCSORT_LOADFULLSORT; |
| 574 |
} |
| 575 |
} |
| 576 |
|
| 577 |
/* |
| 578 |
* Scan the subplan in the forward direction while creating the sorted |
| 579 |
* data. |
| 580 |
*/ |
| 581 |
estate->es_direction = ForwardScanDirection; |
| 582 |
|
| 583 |
outerNode = outerPlanState(node); |
| 584 |
tupDesc = ExecGetResultType(outerNode); |
| 585 |
|
| 586 |
/* Load tuples into the full sort state. */ |
| 587 |
if (node->execution_status == INCSORT_LOADFULLSORT) |
| 588 |
{ |
| 589 |
/* |
| 590 |
* Initialize sorting structures. |
| 591 |
*/ |
| 592 |
if (fullsort_state == NULL) |
| 593 |
{ |
| 594 |
/* |
| 595 |
* Initialize presorted column support structures for |
| 596 |
* isCurrentGroup(). It's correct to do this along with the |
| 597 |
* initial initialization for the full sort state (and not for the |
| 598 |
* prefix sort state) since we always load the full sort state |
| 599 |
* first. |
| 600 |
*/ |
| 601 |
preparePresortedCols(node); |
| 602 |
|
| 603 |
/* |
| 604 |
* Since we optimize small prefix key groups by accumulating a |
| 605 |
* minimum number of tuples before sorting, we can't assume that a |
| 606 |
* group of tuples all have the same prefix key values. Hence we |
| 607 |
* setup the full sort tuplesort to sort by all requested sort |
| 608 |
* keys. |
| 609 |
*/ |
| 610 |
fullsort_state = tuplesort_begin_heap(tupDesc, |
| 611 |
plannode->sort.numCols, |
| 612 |
plannode->sort.sortColIdx, |
| 613 |
plannode->sort.sortOperators, |
| 614 |
plannode->sort.collations, |
| 615 |
plannode->sort.nullsFirst, |
| 616 |
work_mem, |
| 617 |
NULL, |
| 618 |
node->bounded ? |
| 619 |
TUPLESORT_ALLOWBOUNDED : |
| 620 |
TUPLESORT_NONE); |
| 621 |
node->fullsort_state = fullsort_state; |
| 622 |
} |
| 623 |
else |
| 624 |
{ |
| 625 |
/* Reset sort for the next batch. */ |
| 626 |
tuplesort_reset(fullsort_state); |
| 627 |
} |
| 628 |
|
| 629 |
/* |
| 630 |
* Calculate the remaining tuples left if bounded and configure both |
| 631 |
* bounded sort and the minimum group size accordingly. |
| 632 |
*/ |
| 633 |
if (node->bounded) |
| 634 |
{ |
| 635 |
int64 currentBound = node->bound - node->bound_Done; |
| 636 |
|
| 637 |
/* |
| 638 |
* Bounded sort isn't likely to be a useful optimization for full |
| 639 |
* sort mode since we limit full sort mode to a relatively small |
| 640 |
* number of tuples and tuplesort doesn't switch over to top-n |
| 641 |
* heap sort anyway unless it hits (2 * bound) tuples. |
| 642 |
*/ |
| 643 |
if (currentBound < DEFAULT_MIN_GROUP_SIZE) |
| 644 |
tuplesort_set_bound(fullsort_state, currentBound); |
| 645 |
|
| 646 |
minGroupSize = Min(DEFAULT_MIN_GROUP_SIZE, currentBound); |
| 647 |
} |
| 648 |
else |
| 649 |
minGroupSize = DEFAULT_MIN_GROUP_SIZE; |
| 650 |
|
| 651 |
/* |
| 652 |
* Because we have to read the next tuple to find out that we've |
| 653 |
* encountered a new prefix key group, on subsequent groups we have to |
| 654 |
* carry over that extra tuple and add it to the new group's sort here |
| 655 |
* before we read any new tuples from the outer node. |
| 656 |
*/ |
| 657 |
if (!TupIsNull(node->group_pivot)) |
| 658 |
{ |
| 659 |
tuplesort_puttupleslot(fullsort_state, node->group_pivot); |
| 660 |
nTuples++; |
| 661 |
|
| 662 |
/* |
| 663 |
* We're in full sort mode accumulating a minimum number of tuples |
| 664 |
* and not checking for prefix key equality yet, so we can't |
| 665 |
* assume the group pivot tuple will remain the same -- unless |
| 666 |
* we're using a minimum group size of 1, in which case the pivot |
| 667 |
* is obviously still the pivot. |
| 668 |
*/ |
| 669 |
if (nTuples != minGroupSize) |
| 670 |
ExecClearTuple(node->group_pivot); |
| 671 |
} |
| 672 |
|
| 673 |
|
| 674 |
/* |
| 675 |
* Pull as many tuples from the outer node as possible given our |
| 676 |
* current operating mode. |
| 677 |
*/ |
| 678 |
for (;;) |
| 679 |
{ |
| 680 |
slot = ExecProcNode(outerNode); |
| 681 |
|
| 682 |
/* |
| 683 |
* If the outer node can't provide us any more tuples, then we can |
| 684 |
* sort the current group and return those tuples. |
| 685 |
*/ |
| 686 |
if (TupIsNull(slot)) |
| 687 |
{ |
| 688 |
/* |
| 689 |
* We need to know later if the outer node has completed to be |
| 690 |
* able to distinguish between being done with a batch and |
| 691 |
* being done with the whole node. |
| 692 |
*/ |
| 693 |
node->outerNodeDone = true; |
| 694 |
|
| 695 |
SO1_printf("Sorting fullsort with " INT64_FORMAT " tuples\n", nTuples); |
| 696 |
tuplesort_performsort(fullsort_state); |
| 697 |
|
| 698 |
INSTRUMENT_SORT_GROUP(node, fullsort); |
| 699 |
|
| 700 |
SO_printf("Setting execution_status to INCSORT_READFULLSORT (final tuple)\n"); |
| 701 |
node->execution_status = INCSORT_READFULLSORT; |
| 702 |
break; |
| 703 |
} |
| 704 |
|
| 705 |
/* Accumulate the next group of presorted tuples. */ |
| 706 |
if (nTuples < minGroupSize) |
| 707 |
{ |
| 708 |
/* |
| 709 |
* If we haven't yet hit our target minimum group size, then |
| 710 |
* we don't need to bother checking for inclusion in the |
| 711 |
* current prefix group since at this point we'll assume that |
| 712 |
* we'll full sort this batch to avoid a large number of very |
| 713 |
* tiny (and thus inefficient) sorts. |
| 714 |
*/ |
| 715 |
tuplesort_puttupleslot(fullsort_state, slot); |
| 716 |
nTuples++; |
| 717 |
|
| 718 |
/* |
| 719 |
* If we've reached our minimum group size, then we need to |
| 720 |
* store the most recent tuple as a pivot. |
| 721 |
*/ |
| 722 |
if (nTuples == minGroupSize) |
| 723 |
ExecCopySlot(node->group_pivot, slot); |
| 724 |
} |
| 725 |
else |
| 726 |
{ |
| 727 |
/* |
| 728 |
* If we've already accumulated enough tuples to reach our |
| 729 |
* minimum group size, then we need to compare any additional |
| 730 |
* tuples to our pivot tuple to see if we reach the end of |
| 731 |
* that prefix key group. Only after we find changed prefix |
| 732 |
* keys can we guarantee sort stability of the tuples we've |
| 733 |
* already accumulated. |
| 734 |
*/ |
| 735 |
if (isCurrentGroup(node, node->group_pivot, slot)) |
| 736 |
{ |
| 737 |
/* |
| 738 |
* As long as the prefix keys match the pivot tuple then |
| 739 |
* load the tuple into the tuplesort. |
| 740 |
*/ |
| 741 |
tuplesort_puttupleslot(fullsort_state, slot); |
| 742 |
nTuples++; |
| 743 |
} |
| 744 |
else |
| 745 |
{ |
| 746 |
/* |
| 747 |
* Since the tuple we fetched isn't part of the current |
| 748 |
* prefix key group we don't want to sort it as part of |
| 749 |
* the current batch. Instead we use the group_pivot slot |
| 750 |
* to carry it over to the next batch (even though we |
| 751 |
* won't actually treat it as a group pivot). |
| 752 |
*/ |
| 753 |
ExecCopySlot(node->group_pivot, slot); |
| 754 |
|
| 755 |
if (node->bounded) |
| 756 |
{ |
| 757 |
/* |
| 758 |
* If the current node has a bound, and we've already |
| 759 |
* sorted n tuples, then the functional bound |
| 760 |
* remaining is (original bound - n), so store the |
| 761 |
* current number of processed tuples for later use |
| 762 |
* configuring the sort state's bound. |
| 763 |
*/ |
| 764 |
SO2_printf("Changing bound_Done from " INT64_FORMAT " to " INT64_FORMAT "\n", |
| 765 |
node->bound_Done, |
| 766 |
Min(node->bound, node->bound_Done + nTuples)); |
| 767 |
node->bound_Done = Min(node->bound, node->bound_Done + nTuples); |
| 768 |
} |
| 769 |
|
| 770 |
/* |
| 771 |
* Once we find changed prefix keys we can complete the |
| 772 |
* sort and transition modes to reading out the sorted |
| 773 |
* tuples. |
| 774 |
*/ |
| 775 |
SO1_printf("Sorting fullsort tuplesort with " INT64_FORMAT " tuples\n", |
| 776 |
nTuples); |
| 777 |
tuplesort_performsort(fullsort_state); |
| 778 |
|
| 779 |
INSTRUMENT_SORT_GROUP(node, fullsort); |
| 780 |
|
| 781 |
SO_printf("Setting execution_status to INCSORT_READFULLSORT (found end of group)\n"); |
| 782 |
node->execution_status = INCSORT_READFULLSORT; |
| 783 |
break; |
| 784 |
} |
| 785 |
} |
| 786 |
|
| 787 |
/* |
| 788 |
* Unless we've already transitioned modes to reading from the |
| 789 |
* full sort state, then we assume that having read at least |
| 790 |
* DEFAULT_MAX_FULL_SORT_GROUP_SIZE tuples means it's likely we're |
| 791 |
* processing a large group of tuples all having equal prefix keys |
| 792 |
* (but haven't yet found the final tuple in that prefix key |
| 793 |
* group), so we need to transition into presorted prefix mode. |
| 794 |
*/ |
| 795 |
if (nTuples > DEFAULT_MAX_FULL_SORT_GROUP_SIZE && |
| 796 |
node->execution_status != INCSORT_READFULLSORT) |
| 797 |
{ |
| 798 |
/* |
| 799 |
* The group pivot we have stored has already been put into |
| 800 |
* the tuplesort; we don't want to carry it over. Since we |
| 801 |
* haven't yet found the end of the prefix key group, it might |
| 802 |
* seem like we should keep this, but we don't actually know |
| 803 |
* how many prefix key groups might be represented in the full |
| 804 |
* sort state, so we'll let the mode transition function |
| 805 |
* manage this state for us. |
| 806 |
*/ |
| 807 |
ExecClearTuple(node->group_pivot); |
| 808 |
|
| 809 |
/* |
| 810 |
* Unfortunately the tuplesort API doesn't include a way to |
| 811 |
* retrieve tuples unless a sort has been performed, so we |
| 812 |
* perform the sort even though we could just as easily rely |
| 813 |
* on FIFO retrieval semantics when transferring them to the |
| 814 |
* presorted prefix tuplesort. |
| 815 |
*/ |
| 816 |
SO1_printf("Sorting fullsort tuplesort with " INT64_FORMAT " tuples\n", nTuples); |
| 817 |
tuplesort_performsort(fullsort_state); |
| 818 |
|
| 819 |
INSTRUMENT_SORT_GROUP(node, fullsort); |
| 820 |
|
| 821 |
/* |
| 822 |
* If the full sort tuplesort happened to switch into top-n |
| 823 |
* heapsort mode then we will only be able to retrieve |
| 824 |
* currentBound tuples (since the tuplesort will have only |
| 825 |
* retained the top-n tuples). This is safe even though we |
| 826 |
* haven't yet completed fetching the current prefix key group |
| 827 |
* because the tuples we've "lost" already sorted "below" the |
| 828 |
* retained ones, and we're already contractually guaranteed |
| 829 |
* to not need any more than the currentBound tuples. |
| 830 |
*/ |
| 831 |
if (tuplesort_used_bound(node->fullsort_state)) |
| 832 |
{ |
| 833 |
int64 currentBound = node->bound - node->bound_Done; |
| 834 |
|
| 835 |
SO2_printf("Read " INT64_FORMAT " tuples, but setting to " INT64_FORMAT " because we used bounded sort\n", |
| 836 |
nTuples, Min(currentBound, nTuples)); |
| 837 |
nTuples = Min(currentBound, nTuples); |
| 838 |
} |
| 839 |
|
| 840 |
SO1_printf("Setting n_fullsort_remaining to " INT64_FORMAT " and calling switchToPresortedPrefixMode()\n", |
| 841 |
nTuples); |
| 842 |
|
| 843 |
/* |
| 844 |
* We might have multiple prefix key groups in the full sort |
| 845 |
* state, so the mode transition function needs to know that |
| 846 |
* it needs to move from the fullsort to presorted prefix |
| 847 |
* sort. |
| 848 |
*/ |
| 849 |
node->n_fullsort_remaining = nTuples; |
| 850 |
|
| 851 |
/* Transition the tuples to the presorted prefix tuplesort. */ |
| 852 |
switchToPresortedPrefixMode(pstate); |
| 853 |
|
| 854 |
/* |
| 855 |
* Since we know we had tuples to move to the presorted prefix |
| 856 |
* tuplesort, we know that unless that transition has verified |
| 857 |
* that all tuples belonged to the same prefix key group (in |
| 858 |
* which case we can go straight to continuing to load tuples |
| 859 |
* into that tuplesort), we should have a tuple to return |
| 860 |
* here. |
| 861 |
* |
| 862 |
* Either way, the appropriate execution status should have |
| 863 |
* been set by switchToPresortedPrefixMode(), so we can drop |
| 864 |
* out of the loop here and let the appropriate path kick in. |
| 865 |
*/ |
| 866 |
break; |
| 867 |
} |
| 868 |
} |
| 869 |
} |
| 870 |
|
| 871 |
if (node->execution_status == INCSORT_LOADPREFIXSORT) |
| 872 |
{ |
| 873 |
/* |
| 874 |
* We only enter this state after the mode transition function has |
| 875 |
* confirmed all remaining tuples from the full sort state have the |
| 876 |
* same prefix and moved those tuples to the prefix sort state. That |
| 877 |
* function has also set a group pivot tuple (which doesn't need to be |
| 878 |
* carried over; it's already been put into the prefix sort state). |
| 879 |
*/ |
| 880 |
Assert(!TupIsNull(node->group_pivot)); |
| 881 |
|
| 882 |
/* |
| 883 |
* Read tuples from the outer node and load them into the prefix sort |
| 884 |
* state until we encounter a tuple whose prefix keys don't match the |
| 885 |
* current group_pivot tuple, since we can't guarantee sort stability |
| 886 |
* until we have all tuples matching those prefix keys. |
| 887 |
*/ |
| 888 |
for (;;) |
| 889 |
{ |
| 890 |
slot = ExecProcNode(outerNode); |
| 891 |
|
| 892 |
/* |
| 893 |
* If we've exhausted tuples from the outer node we're done |
| 894 |
* loading the prefix sort state. |
| 895 |
*/ |
| 896 |
if (TupIsNull(slot)) |
| 897 |
{ |
| 898 |
/* |
| 899 |
* We need to know later if the outer node has completed to be |
| 900 |
* able to distinguish between being done with a batch and |
| 901 |
* being done with the whole node. |
| 902 |
*/ |
| 903 |
node->outerNodeDone = true; |
| 904 |
break; |
| 905 |
} |
| 906 |
|
| 907 |
/* |
| 908 |
* If the tuple's prefix keys match our pivot tuple, we're not |
| 909 |
* done yet and can load it into the prefix sort state. If not, we |
| 910 |
* don't want to sort it as part of the current batch. Instead we |
| 911 |
* use the group_pivot slot to carry it over to the next batch |
| 912 |
* (even though we won't actually treat it as a group pivot). |
| 913 |
*/ |
| 914 |
if (isCurrentGroup(node, node->group_pivot, slot)) |
| 915 |
{ |
| 916 |
tuplesort_puttupleslot(node->prefixsort_state, slot); |
| 917 |
nTuples++; |
| 918 |
} |
| 919 |
else |
| 920 |
{ |
| 921 |
ExecCopySlot(node->group_pivot, slot); |
| 922 |
break; |
| 923 |
} |
| 924 |
} |
| 925 |
|
| 926 |
/* |
| 927 |
* Perform the sort and begin returning the tuples to the parent plan |
| 928 |
* node. |
| 929 |
*/ |
| 930 |
SO1_printf("Sorting presorted prefix tuplesort with " INT64_FORMAT " tuples\n", nTuples); |
| 931 |
tuplesort_performsort(node->prefixsort_state); |
| 932 |
|
| 933 |
INSTRUMENT_SORT_GROUP(node, prefixsort); |
| 934 |
|
| 935 |
SO_printf("Setting execution_status to INCSORT_READPREFIXSORT (found end of group)\n"); |
| 936 |
node->execution_status = INCSORT_READPREFIXSORT; |
| 937 |
|
| 938 |
if (node->bounded) |
| 939 |
{ |
| 940 |
/* |
| 941 |
* If the current node has a bound, and we've already sorted n |
| 942 |
* tuples, then the functional bound remaining is (original bound |
| 943 |
* - n), so store the current number of processed tuples for use |
| 944 |
* in configuring sorting bound. |
| 945 |
*/ |
| 946 |
SO2_printf("Changing bound_Done from " INT64_FORMAT " to " INT64_FORMAT "\n", |
| 947 |
node->bound_Done, |
| 948 |
Min(node->bound, node->bound_Done + nTuples)); |
| 949 |
node->bound_Done = Min(node->bound, node->bound_Done + nTuples); |
| 950 |
} |
| 951 |
} |
| 952 |
|
| 953 |
/* Restore to user specified direction. */ |
| 954 |
estate->es_direction = dir; |
| 955 |
|
| 956 |
/* |
| 957 |
* Get the first or next tuple from tuplesort. Returns NULL if no more |
| 958 |
* tuples. |
| 959 |
*/ |
| 960 |
read_sortstate = node->execution_status == INCSORT_READFULLSORT ? |
| 961 |
fullsort_state : node->prefixsort_state; |
| 962 |
slot = node->ss.ps.ps_ResultTupleSlot; |
| 963 |
(void) tuplesort_gettupleslot(read_sortstate, ScanDirectionIsForward(dir), |
| 964 |
false, slot, NULL); |
| 965 |
return slot; |
| 966 |
} |
| 967 |
|
| 968 |
/* ---------------------------------------------------------------- |
| 969 |
* ExecInitIncrementalSort |
| 970 |
* |
| 971 |
* Creates the run-time state information for the sort node |
| 972 |
* produced by the planner and initializes its outer subtree. |
| 973 |
* ---------------------------------------------------------------- |
| 974 |
*/ |
| 975 |
IncrementalSortState * |
| 976 |
ExecInitIncrementalSort(IncrementalSort *node, EState *estate, int eflags) |
| 977 |
{ |
| 978 |
IncrementalSortState *incrsortstate; |
| 979 |
|
| 980 |
SO_printf("ExecInitIncrementalSort: initializing sort node\n"); |
| 981 |
|
| 982 |
/* |
| 983 |
* Incremental sort can't be used with EXEC_FLAG_BACKWARD or |
| 984 |
* EXEC_FLAG_MARK, because the current sort state contains only one sort |
| 985 |
* batch rather than the full result set. |
| 986 |
*/ |
| 987 |
Assert((eflags & (EXEC_FLAG_BACKWARD | EXEC_FLAG_MARK)) == 0); |
| 988 |
|
| 989 |
/* Initialize state structure. */ |
| 990 |
incrsortstate = makeNode(IncrementalSortState); |
| 991 |
incrsortstate->ss.ps.plan = (Plan *) node; |
| 992 |
incrsortstate->ss.ps.state = estate; |
| 993 |
incrsortstate->ss.ps.ExecProcNode = ExecIncrementalSort; |
| 994 |
|
| 995 |
incrsortstate->execution_status = INCSORT_LOADFULLSORT; |
| 996 |
incrsortstate->bounded = false; |
| 997 |
incrsortstate->outerNodeDone = false; |
| 998 |
incrsortstate->bound_Done = 0; |
| 999 |
incrsortstate->fullsort_state = NULL; |
| 1000 |
incrsortstate->prefixsort_state = NULL; |
| 1001 |
incrsortstate->group_pivot = NULL; |
| 1002 |
incrsortstate->transfer_tuple = NULL; |
| 1003 |
incrsortstate->n_fullsort_remaining = 0; |
| 1004 |
incrsortstate->presorted_keys = NULL; |
| 1005 |
|
| 1006 |
if (incrsortstate->ss.ps.instrument != NULL) |
| 1007 |
{ |
| 1008 |
IncrementalSortGroupInfo *fullsortGroupInfo = |
| 1009 |
&incrsortstate->incsort_info.fullsortGroupInfo; |
| 1010 |
IncrementalSortGroupInfo *prefixsortGroupInfo = |
| 1011 |
&incrsortstate->incsort_info.prefixsortGroupInfo; |
| 1012 |
|
| 1013 |
fullsortGroupInfo->groupCount = 0; |
| 1014 |
fullsortGroupInfo->maxDiskSpaceUsed = 0; |
| 1015 |
fullsortGroupInfo->totalDiskSpaceUsed = 0; |
| 1016 |
fullsortGroupInfo->maxMemorySpaceUsed = 0; |
| 1017 |
fullsortGroupInfo->totalMemorySpaceUsed = 0; |
| 1018 |
fullsortGroupInfo->sortMethods = 0; |
| 1019 |
prefixsortGroupInfo->groupCount = 0; |
| 1020 |
prefixsortGroupInfo->maxDiskSpaceUsed = 0; |
| 1021 |
prefixsortGroupInfo->totalDiskSpaceUsed = 0; |
| 1022 |
prefixsortGroupInfo->maxMemorySpaceUsed = 0; |
| 1023 |
prefixsortGroupInfo->totalMemorySpaceUsed = 0; |
| 1024 |
prefixsortGroupInfo->sortMethods = 0; |
| 1025 |
} |
| 1026 |
|
| 1027 |
/* |
| 1028 |
* Miscellaneous initialization |
| 1029 |
* |
| 1030 |
* Sort nodes don't initialize their ExprContexts because they never call |
| 1031 |
* ExecQual or ExecProject. |
| 1032 |
*/ |
| 1033 |
|
| 1034 |
/* |
| 1035 |
* Initialize child nodes. |
| 1036 |
* |
| 1037 |
* Incremental sort does not support backwards scans and mark/restore, so |
| 1038 |
* we don't bother removing the flags from eflags here. We allow passing a |
| 1039 |
* REWIND flag, because although incremental sort can't use it, the child |
| 1040 |
* nodes may be able to do something more useful. |
| 1041 |
*/ |
| 1042 |
outerPlanState(incrsortstate) = ExecInitNode(outerPlan(node), estate, eflags); |
| 1043 |
|
| 1044 |
/* |
| 1045 |
* Initialize scan slot and type. |
| 1046 |
*/ |
| 1047 |
ExecCreateScanSlotFromOuterPlan(estate, &incrsortstate->ss, &TTSOpsMinimalTuple); |
| 1048 |
|
| 1049 |
/* |
| 1050 |
* Initialize return slot and type. No need to initialize projection info |
| 1051 |
* because we don't do any projections. |
| 1052 |
*/ |
| 1053 |
ExecInitResultTupleSlotTL(&incrsortstate->ss.ps, &TTSOpsMinimalTuple); |
| 1054 |
incrsortstate->ss.ps.ps_ProjInfo = NULL; |
| 1055 |
|
| 1056 |
/* |
| 1057 |
* Initialize standalone slots to store a tuple for pivot prefix keys and |
| 1058 |
* for carrying over a tuple from one batch to the next. |
| 1059 |
*/ |
| 1060 |
incrsortstate->group_pivot = |
| 1061 |
MakeSingleTupleTableSlot(ExecGetResultType(outerPlanState(incrsortstate)), |
| 1062 |
&TTSOpsMinimalTuple); |
| 1063 |
incrsortstate->transfer_tuple = |
| 1064 |
MakeSingleTupleTableSlot(ExecGetResultType(outerPlanState(incrsortstate)), |
| 1065 |
&TTSOpsMinimalTuple); |
| 1066 |
|
| 1067 |
SO_printf("ExecInitIncrementalSort: sort node initialized\n"); |
| 1068 |
|
| 1069 |
return incrsortstate; |
| 1070 |
} |
| 1071 |
|
| 1072 |
/* ---------------------------------------------------------------- |
| 1073 |
* ExecEndIncrementalSort(node) |
| 1074 |
* ---------------------------------------------------------------- |
| 1075 |
*/ |
| 1076 |
void |
| 1077 |
ExecEndIncrementalSort(IncrementalSortState *node) |
| 1078 |
{ |
| 1079 |
SO_printf("ExecEndIncrementalSort: shutting down sort node\n"); |
| 1080 |
|
| 1081 |
ExecDropSingleTupleTableSlot(node->group_pivot); |
| 1082 |
ExecDropSingleTupleTableSlot(node->transfer_tuple); |
| 1083 |
|
| 1084 |
/* |
| 1085 |
* Release tuplesort resources. |
| 1086 |
*/ |
| 1087 |
if (node->fullsort_state != NULL) |
| 1088 |
{ |
| 1089 |
tuplesort_end(node->fullsort_state); |
| 1090 |
node->fullsort_state = NULL; |
| 1091 |
} |
| 1092 |
if (node->prefixsort_state != NULL) |
| 1093 |
{ |
| 1094 |
tuplesort_end(node->prefixsort_state); |
| 1095 |
node->prefixsort_state = NULL; |
| 1096 |
} |
| 1097 |
|
| 1098 |
/* |
| 1099 |
* Shut down the subplan. |
| 1100 |
*/ |
| 1101 |
ExecEndNode(outerPlanState(node)); |
| 1102 |
|
| 1103 |
SO_printf("ExecEndIncrementalSort: sort node shutdown\n"); |
| 1104 |
} |
| 1105 |
|
| 1106 |
void |
| 1107 |
ExecReScanIncrementalSort(IncrementalSortState *node) |
| 1108 |
{ |
| 1109 |
PlanState *outerPlan = outerPlanState(node); |
| 1110 |
|
| 1111 |
/* |
| 1112 |
* Incremental sort doesn't support efficient rescan even when parameters |
| 1113 |
* haven't changed (e.g., rewind) because unlike regular sort we don't |
| 1114 |
* store all tuples at once for the full sort. |
| 1115 |
* |
| 1116 |
* So even if EXEC_FLAG_REWIND is set we just reset all of our state and |
| 1117 |
* re-execute the sort along with the child node. Incremental sort itself |
| 1118 |
* can't do anything smarter, but maybe the child nodes can. |
| 1119 |
* |
| 1120 |
* In theory if we've only filled the full sort with one batch (and |
| 1121 |
* haven't reset it for a new batch yet) then we could efficiently rewind, |
| 1122 |
* but that seems a narrow enough case that it's not worth handling |
| 1123 |
* specially at this time. |
| 1124 |
*/ |
| 1125 |
|
| 1126 |
/* must drop pointer to sort result tuple */ |
| 1127 |
ExecClearTuple(node->ss.ps.ps_ResultTupleSlot); |
| 1128 |
|
| 1129 |
if (node->group_pivot != NULL) |
| 1130 |
ExecClearTuple(node->group_pivot); |
| 1131 |
if (node->transfer_tuple != NULL) |
| 1132 |
ExecClearTuple(node->transfer_tuple); |
| 1133 |
|
| 1134 |
node->outerNodeDone = false; |
| 1135 |
node->n_fullsort_remaining = 0; |
| 1136 |
node->bound_Done = 0; |
| 1137 |
|
| 1138 |
node->execution_status = INCSORT_LOADFULLSORT; |
| 1139 |
|
| 1140 |
/* |
| 1141 |
* If we've set up either of the sort states yet, we need to reset them. |
| 1142 |
* We could end them and null out the pointers, but there's no reason to |
| 1143 |
* repay the setup cost, and because ExecIncrementalSort guards presorted |
| 1144 |
* column functions by checking to see if the full sort state has been |
| 1145 |
* initialized yet, setting the sort states to null here might actually |
| 1146 |
* cause a leak. |
| 1147 |
*/ |
| 1148 |
if (node->fullsort_state != NULL) |
| 1149 |
tuplesort_reset(node->fullsort_state); |
| 1150 |
if (node->prefixsort_state != NULL) |
| 1151 |
tuplesort_reset(node->prefixsort_state); |
| 1152 |
|
| 1153 |
/* |
| 1154 |
* If chgParam of subnode is not null, then the plan will be re-scanned by |
| 1155 |
* the first ExecProcNode. |
| 1156 |
*/ |
| 1157 |
if (outerPlan->chgParam == NULL) |
| 1158 |
ExecReScan(outerPlan); |
| 1159 |
} |
| 1160 |
|
| 1161 |
/* ---------------------------------------------------------------- |
| 1162 |
* Parallel Query Support |
| 1163 |
* ---------------------------------------------------------------- |
| 1164 |
*/ |
| 1165 |
|
| 1166 |
/* ---------------------------------------------------------------- |
| 1167 |
* ExecSortEstimate |
| 1168 |
* |
| 1169 |
* Estimate space required to propagate sort statistics. |
| 1170 |
* ---------------------------------------------------------------- |
| 1171 |
*/ |
| 1172 |
void |
| 1173 |
ExecIncrementalSortEstimate(IncrementalSortState *node, ParallelContext *pcxt) |
| 1174 |
{ |
| 1175 |
Size size; |
| 1176 |
|
| 1177 |
/* don't need this if not instrumenting or no workers */ |
| 1178 |
if (!node->ss.ps.instrument || pcxt->nworkers == 0) |
| 1179 |
return; |
| 1180 |
|
| 1181 |
size = mul_size(pcxt->nworkers, sizeof(IncrementalSortInfo)); |
| 1182 |
size = add_size(size, offsetof(SharedIncrementalSortInfo, sinfo)); |
| 1183 |
shm_toc_estimate_chunk(&pcxt->estimator, size); |
| 1184 |
shm_toc_estimate_keys(&pcxt->estimator, 1); |
| 1185 |
} |
| 1186 |
|
| 1187 |
/* ---------------------------------------------------------------- |
| 1188 |
* ExecSortInitializeDSM |
| 1189 |
* |
| 1190 |
* Initialize DSM space for sort statistics. |
| 1191 |
* ---------------------------------------------------------------- |
| 1192 |
*/ |
| 1193 |
void |
| 1194 |
ExecIncrementalSortInitializeDSM(IncrementalSortState *node, ParallelContext *pcxt) |
| 1195 |
{ |
| 1196 |
Size size; |
| 1197 |
|
| 1198 |
/* don't need this if not instrumenting or no workers */ |
| 1199 |
if (!node->ss.ps.instrument || pcxt->nworkers == 0) |
| 1200 |
return; |
| 1201 |
|
| 1202 |
size = offsetof(SharedIncrementalSortInfo, sinfo) |
| 1203 |
+ pcxt->nworkers * sizeof(IncrementalSortInfo); |
| 1204 |
node->shared_info = shm_toc_allocate(pcxt->toc, size); |
| 1205 |
/* ensure any unfilled slots will contain zeroes */ |
| 1206 |
memset(node->shared_info, 0, size); |
| 1207 |
node->shared_info->num_workers = pcxt->nworkers; |
| 1208 |
shm_toc_insert(pcxt->toc, node->ss.ps.plan->plan_node_id, |
| 1209 |
node->shared_info); |
| 1210 |
} |
| 1211 |
|
| 1212 |
/* ---------------------------------------------------------------- |
| 1213 |
* ExecSortInitializeWorker |
| 1214 |
* |
| 1215 |
* Attach worker to DSM space for sort statistics. |
| 1216 |
* ---------------------------------------------------------------- |
| 1217 |
*/ |
| 1218 |
void |
| 1219 |
ExecIncrementalSortInitializeWorker(IncrementalSortState *node, ParallelWorkerContext *pwcxt) |
| 1220 |
{ |
| 1221 |
node->shared_info = |
| 1222 |
shm_toc_lookup(pwcxt->toc, node->ss.ps.plan->plan_node_id, true); |
| 1223 |
node->am_worker = true; |
| 1224 |
} |
| 1225 |
|
| 1226 |
/* ---------------------------------------------------------------- |
| 1227 |
* ExecSortRetrieveInstrumentation |
| 1228 |
* |
| 1229 |
* Transfer sort statistics from DSM to private memory. |
| 1230 |
* ---------------------------------------------------------------- |
| 1231 |
*/ |
| 1232 |
void |
| 1233 |
ExecIncrementalSortRetrieveInstrumentation(IncrementalSortState *node) |
| 1234 |
{ |
| 1235 |
Size size; |
| 1236 |
SharedIncrementalSortInfo *si; |
| 1237 |
|
| 1238 |
if (node->shared_info == NULL) |
| 1239 |
return; |
| 1240 |
|
| 1241 |
size = offsetof(SharedIncrementalSortInfo, sinfo) |
| 1242 |
+ node->shared_info->num_workers * sizeof(IncrementalSortInfo); |
| 1243 |
si = palloc(size); |
| 1244 |
memcpy(si, node->shared_info, size); |
| 1245 |
node->shared_info = si; |
| 1246 |
} |