1 /*-------------------------------------------------------------------------
4 * Generalized tuple sorting routines.
6 * This module handles sorting of heap tuples, index tuples, or single
7 * Datums (and could easily support other kinds of sortable objects,
8 * if necessary). It works efficiently for both small and large amounts
9 * of data. Small amounts are sorted in-memory using qsort(). Large
10 * amounts are sorted using temporary files and a standard external sort
13 * See Knuth, volume 3, for more than you want to know about the external
14 * sorting algorithm. We divide the input into sorted runs using replacement
15 * selection, in the form of a priority tree implemented as a heap
16 * (essentially his Algorithm 5.2.3H), then merge the runs using polyphase
17 * merge, Knuth's Algorithm 5.4.2D. The logical "tapes" used by Algorithm D
18 * are implemented by logtape.c, which avoids space wastage by recycling
19 * disk space as soon as each block is read from its "tape".
21 * We do not form the initial runs using Knuth's recommended replacement
22 * selection data structure (Algorithm 5.4.1R), because it uses a fixed
23 * number of records in memory at all times. Since we are dealing with
24 * tuples that may vary considerably in size, we want to be able to vary
25 * the number of records kept in memory to ensure full utilization of the
26 * allowed sort memory space. So, we keep the tuples in a variable-size
27 * heap, with the next record to go out at the top of the heap. Like
28 * Algorithm 5.4.1R, each record is stored with the run number that it
29 * must go into, and we use (run number, key) as the ordering key for the
30 * heap. When the run number at the top of the heap changes, we know that
31 * no more records of the prior run are left in the heap.
33 * The approximate amount of memory allowed for any one sort operation
34 * is specified in kilobytes by the caller (most pass work_mem). Initially,
35 * we absorb tuples and simply store them in an unsorted array as long as
36 * we haven't exceeded workMem. If we reach the end of the input without
37 * exceeding workMem, we sort the array using qsort() and subsequently return
38 * tuples just by scanning the tuple array sequentially. If we do exceed
39 * workMem, we construct a heap using Algorithm H and begin to emit tuples
40 * into sorted runs in temporary tapes, emitting just enough tuples at each
41 * step to get back within the workMem limit. Whenever the run number at
42 * the top of the heap changes, we begin a new run with a new output tape
43 * (selected per Algorithm D). After the end of the input is reached,
44 * we dump out remaining tuples in memory into a final run (or two),
45 * then merge the runs using Algorithm D.
47 * When merging runs, we use a heap containing just the frontmost tuple from
48 * each source run; we repeatedly output the smallest tuple and insert the
49 * next tuple from its source tape (if any). When the heap empties, the merge
50 * is complete. The basic merge algorithm thus needs very little memory ---
51 * only M tuples for an M-way merge, and M is constrained to a small number.
52 * However, we can still make good use of our full workMem allocation by
53 * pre-reading additional tuples from each source tape. Without prereading,
54 * our access pattern to the temporary file would be very erratic; on average
55 * we'd read one block from each of M source tapes during the same time that
56 * we're writing M blocks to the output tape, so there is no sequentiality of
57 * access at all, defeating the read-ahead methods used by most Unix kernels.
58 * Worse, the output tape gets written into a very random sequence of blocks
59 * of the temp file, ensuring that things will be even worse when it comes
60 * time to read that tape. A straightforward merge pass thus ends up doing a
61 * lot of waiting for disk seeks. We can improve matters by prereading from
62 * each source tape sequentially, loading about workMem/M bytes from each tape
63 * in turn. Then we run the merge algorithm, writing but not reading until
64 * one of the preloaded tuple series runs out. Then we switch back to preread
65 * mode, fill memory again, and repeat. This approach helps to localize both
66 * read and write accesses.
68 * When the caller requests random access to the sort result, we form
69 * the final sorted run on a logical tape which is then "frozen", so
70 * that we can access it randomly. When the caller does not need random
71 * access, we return from tuplesort_performsort() as soon as we are down
72 * to one run per logical tape. The final merge is then performed
73 * on-the-fly as the caller repeatedly calls tuplesort_getXXX; this
74 * saves one cycle of writing all the data out to disk and reading it in.
76 * Before Postgres 8.2, we always used a seven-tape polyphase merge, on the
77 * grounds that 7 is the "sweet spot" on the tapes-to-passes curve according
78 * to Knuth's figure 70 (section 5.4.2). However, Knuth is assuming that
79 * tape drives are expensive beasts, and in particular that there will always
80 * be many more runs than tape drives. In our implementation a "tape drive"
81 * doesn't cost much more than a few Kb of memory buffers, so we can afford
82 * to have lots of them. In particular, if we can have as many tape drives
83 * as sorted runs, we can eliminate any repeated I/O at all. In the current
84 * code we determine the number of tapes M on the basis of workMem: we want
85 * workMem/M to be large enough that we read a fair amount of data each time
86 * we preread from a tape, so as to maintain the locality of access described
87 * above. Nonetheless, with large workMem we can have many tapes.
90 * Portions Copyright (c) 1996-2015, PostgreSQL Global Development Group
91 * Portions Copyright (c) 1994, Regents of the University of California
94 * src/backend/utils/sort/tuplesort.c
96 *-------------------------------------------------------------------------
103 #include "access/htup_details.h"
104 #include "access/nbtree.h"
105 #include "catalog/index.h"
106 #include "commands/tablespace.h"
107 #include "executor/executor.h"
108 #include "miscadmin.h"
109 #include "pg_trace.h"
110 #include "utils/datum.h"
111 #include "utils/logtape.h"
112 #include "utils/lsyscache.h"
113 #include "utils/memutils.h"
114 #include "utils/pg_rusage.h"
115 #include "utils/rel.h"
116 #include "utils/sortsupport.h"
117 #include "utils/tuplesort.h"
120 /* sort-type codes for sort__start probes */
124 #define CLUSTER_SORT 3
128 bool trace_sort = false;
131 #ifdef DEBUG_BOUNDED_SORT
132 bool optimize_bounded_sort = true;
137 * The objects we actually sort are SortTuple structs. These contain
138 * a pointer to the tuple proper (might be a MinimalTuple or IndexTuple),
139 * which is a separate palloc chunk --- we assume it is just one chunk and
140 * can be freed by a simple pfree(). SortTuples also contain the tuple's
141 * first key column in Datum/nullflag format, and an index integer.
143 * Storing the first key column lets us save heap_getattr or index_getattr
144 * calls during tuple comparisons. We could extract and save all the key
145 * columns not just the first, but this would increase code complexity and
146 * overhead, and wouldn't actually save any comparison cycles in the common
147 * case where the first key determines the comparison result. Note that
148 * for a pass-by-reference datatype, datum1 points into the "tuple" storage.
150 * When sorting single Datums, the data value is represented directly by
151 * datum1/isnull1. If the datatype is pass-by-reference and isnull1 is false,
152 * then datum1 points to a separately palloc'd data value that is also pointed
153 * to by the "tuple" pointer; otherwise "tuple" is NULL. There is one special
154 * case: when the sort support infrastructure provides an "abbreviated key"
155 * representation, where the key is (typically) a pass by value proxy for a
156 * pass by reference type.
158 * While building initial runs, tupindex holds the tuple's run number. During
159 * merge passes, we re-use it to hold the input tape number that each tuple in
160 * the heap was read from, or to hold the index of the next tuple pre-read
161 * from the same tape in the case of pre-read entries. tupindex goes unused
162 * if the sort occurs entirely in memory.
166 void *tuple; /* the tuple proper */
167 Datum datum1; /* value of first key column */
168 bool isnull1; /* is first key column NULL? */
169 int tupindex; /* see notes above */
174 * Possible states of a Tuplesort object. These denote the states that
175 * persist between calls of Tuplesort routines.
179 TSS_INITIAL, /* Loading tuples; still within memory limit */
180 TSS_BOUNDED, /* Loading tuples into bounded-size heap */
181 TSS_BUILDRUNS, /* Loading tuples; writing to tape */
182 TSS_SORTEDINMEM, /* Sort completed entirely in memory */
183 TSS_SORTEDONTAPE, /* Sort completed, final run is on tape */
184 TSS_FINALMERGE /* Performing final merge on-the-fly */
188 * Parameters for calculation of number of tapes to use --- see inittapes()
189 * and tuplesort_merge_order().
191 * In this calculation we assume that each tape will cost us about 3 blocks
192 * worth of buffer space (which is an underestimate for very large data
193 * volumes, but it's probably close enough --- see logtape.c).
195 * MERGE_BUFFER_SIZE is how much data we'd like to read from each input
196 * tape during a preread cycle (see discussion at top of file).
198 #define MINORDER 6 /* minimum merge order */
199 #define TAPE_BUFFER_OVERHEAD (BLCKSZ * 3)
200 #define MERGE_BUFFER_SIZE (BLCKSZ * 32)
202 typedef int (*SortTupleComparator) (const SortTuple *a, const SortTuple *b,
203 Tuplesortstate *state);
206 * Private state of a Tuplesort operation.
208 struct Tuplesortstate
210 TupSortStatus status; /* enumerated value as shown above */
211 int nKeys; /* number of columns in sort key */
212 bool randomAccess; /* did caller request random access? */
213 bool bounded; /* did caller specify a maximum number of
214 * tuples to return? */
215 bool boundUsed; /* true if we made use of a bounded heap */
216 int bound; /* if bounded, the maximum number of tuples */
217 int64 availMem; /* remaining memory available, in bytes */
218 int64 allowedMem; /* total memory allowed, in bytes */
219 int maxTapes; /* number of tapes (Knuth's T) */
220 int tapeRange; /* maxTapes-1 (Knuth's P) */
221 MemoryContext sortcontext; /* memory context holding all sort data */
222 LogicalTapeSet *tapeset; /* logtape.c object for tapes in a temp file */
225 * These function pointers decouple the routines that must know what kind
226 * of tuple we are sorting from the routines that don't need to know it.
227 * They are set up by the tuplesort_begin_xxx routines.
229 * Function to compare two tuples; result is per qsort() convention, ie:
230 * <0, 0, >0 according as a<b, a=b, a>b. The API must match
231 * qsort_arg_comparator.
233 SortTupleComparator comparetup;
236 * Function to copy a supplied input tuple into palloc'd space and set up
237 * its SortTuple representation (ie, set tuple/datum1/isnull1). Also,
238 * state->availMem must be decreased by the amount of space used for the
239 * tuple copy (note the SortTuple struct itself is not counted).
241 void (*copytup) (Tuplesortstate *state, SortTuple *stup, void *tup);
244 * Function to write a stored tuple onto tape. The representation of the
245 * tuple on tape need not be the same as it is in memory; requirements on
246 * the tape representation are given below. After writing the tuple,
247 * pfree() the out-of-line data (not the SortTuple struct!), and increase
248 * state->availMem by the amount of memory space thereby released.
250 void (*writetup) (Tuplesortstate *state, int tapenum,
254 * Function to read a stored tuple from tape back into memory. 'len' is
255 * the already-read length of the stored tuple. Create a palloc'd copy,
256 * initialize tuple/datum1/isnull1 in the target SortTuple struct, and
257 * decrease state->availMem by the amount of memory space consumed.
259 void (*readtup) (Tuplesortstate *state, SortTuple *stup,
260 int tapenum, unsigned int len);
263 * This array holds the tuples now in sort memory. If we are in state
264 * INITIAL, the tuples are in no particular order; if we are in state
265 * SORTEDINMEM, the tuples are in final sorted order; in states BUILDRUNS
266 * and FINALMERGE, the tuples are organized in "heap" order per Algorithm
267 * H. (Note that memtupcount only counts the tuples that are part of the
268 * heap --- during merge passes, memtuples[] entries beyond tapeRange are
269 * never in the heap and are used to hold pre-read tuples.) In state
270 * SORTEDONTAPE, the array is not used.
272 SortTuple *memtuples; /* array of SortTuple structs */
273 int memtupcount; /* number of tuples currently present */
274 int memtupsize; /* allocated length of memtuples array */
275 bool growmemtuples; /* memtuples' growth still underway? */
278 * While building initial runs, this is the current output run number
279 * (starting at 0). Afterwards, it is the number of initial runs we made.
284 * Unless otherwise noted, all pointer variables below are pointers to
285 * arrays of length maxTapes, holding per-tape data.
289 * These variables are only used during merge passes. mergeactive[i] is
290 * true if we are reading an input run from (actual) tape number i and
291 * have not yet exhausted that run. mergenext[i] is the memtuples index
292 * of the next pre-read tuple (next to be loaded into the heap) for tape
293 * i, or 0 if we are out of pre-read tuples. mergelast[i] similarly
294 * points to the last pre-read tuple from each tape. mergeavailslots[i]
295 * is the number of unused memtuples[] slots reserved for tape i, and
296 * mergeavailmem[i] is the amount of unused space allocated for tape i.
297 * mergefreelist and mergefirstfree keep track of unused locations in the
298 * memtuples[] array. The memtuples[].tupindex fields link together
299 * pre-read tuples for each tape as well as recycled locations in
300 * mergefreelist. It is OK to use 0 as a null link in these lists, because
301 * memtuples[0] is part of the merge heap and is never a pre-read tuple.
303 bool *mergeactive; /* active input run source? */
304 int *mergenext; /* first preread tuple for each source */
305 int *mergelast; /* last preread tuple for each source */
306 int *mergeavailslots; /* slots left for prereading each tape */
307 int64 *mergeavailmem; /* availMem for prereading each tape */
308 int mergefreelist; /* head of freelist of recycled slots */
309 int mergefirstfree; /* first slot never used in this merge */
312 * Variables for Algorithm D. Note that destTape is a "logical" tape
313 * number, ie, an index into the tp_xxx[] arrays. Be careful to keep
314 * "logical" and "actual" tape numbers straight!
316 int Level; /* Knuth's l */
317 int destTape; /* current output tape (Knuth's j, less 1) */
318 int *tp_fib; /* Target Fibonacci run counts (A[]) */
319 int *tp_runs; /* # of real runs on each tape */
320 int *tp_dummy; /* # of dummy runs for each tape (D[]) */
321 int *tp_tapenum; /* Actual tape numbers (TAPE[]) */
322 int activeTapes; /* # of active input tapes in merge pass */
325 * These variables are used after completion of sorting to keep track of
326 * the next tuple to return. (In the tape case, the tape's current read
327 * position is also critical state.)
329 int result_tape; /* actual tape number of finished output */
330 int current; /* array index (only used if SORTEDINMEM) */
331 bool eof_reached; /* reached EOF (needed for cursors) */
333 /* markpos_xxx holds marked position for mark and restore */
334 long markpos_block; /* tape block# (only used if SORTEDONTAPE) */
335 int markpos_offset; /* saved "current", or offset in tape block */
336 bool markpos_eof; /* saved "eof_reached" */
339 * The sortKeys variable is used by every case other than the datum and
340 * hash index cases; it is set by tuplesort_begin_xxx. tupDesc is only
341 * used by the MinimalTuple and CLUSTER routines, though.
344 SortSupport sortKeys; /* array of length nKeys */
347 * This variable is shared by the single-key MinimalTuple case and the
348 * Datum case (which both use qsort_ssup()). Otherwise it's NULL.
353 * Additional state for managing "abbreviated key" sortsupport routines
354 * (which currently may be used by all cases except the Datum sort case and
355 * hash index case). Tracks the intervals at which the optimization's
356 * effectiveness is tested.
358 int64 abbrevNext; /* Tuple # at which to next check applicability */
361 * These variables are specific to the CLUSTER case; they are set by
362 * tuplesort_begin_cluster.
364 IndexInfo *indexInfo; /* info about index being used for reference */
365 EState *estate; /* for evaluating index expressions */
368 * These variables are specific to the IndexTuple case; they are set by
369 * tuplesort_begin_index_xxx and used only by the IndexTuple routines.
371 Relation heapRel; /* table the index is being built on */
372 Relation indexRel; /* index being built */
374 /* These are specific to the index_btree subcase: */
375 bool enforceUnique; /* complain if we find duplicate tuples */
377 /* These are specific to the index_hash subcase: */
378 uint32 hash_mask; /* mask for sortable part of hash code */
381 * These variables are specific to the Datum case; they are set by
382 * tuplesort_begin_datum and used only by the DatumTuple routines.
385 /* we need typelen and byval in order to know how to copy the Datums. */
390 * Resource snapshot for time of sort start.
397 #define COMPARETUP(state,a,b) ((*(state)->comparetup) (a, b, state))
398 #define COPYTUP(state,stup,tup) ((*(state)->copytup) (state, stup, tup))
399 #define WRITETUP(state,tape,stup) ((*(state)->writetup) (state, tape, stup))
400 #define READTUP(state,stup,tape,len) ((*(state)->readtup) (state, stup, tape, len))
401 #define LACKMEM(state) ((state)->availMem < 0)
402 #define USEMEM(state,amt) ((state)->availMem -= (amt))
403 #define FREEMEM(state,amt) ((state)->availMem += (amt))
406 * NOTES about on-tape representation of tuples:
408 * We require the first "unsigned int" of a stored tuple to be the total size
409 * on-tape of the tuple, including itself (so it is never zero; an all-zero
410 * unsigned int is used to delimit runs). The remainder of the stored tuple
411 * may or may not match the in-memory representation of the tuple ---
412 * any conversion needed is the job of the writetup and readtup routines.
414 * If state->randomAccess is true, then the stored representation of the
415 * tuple must be followed by another "unsigned int" that is a copy of the
416 * length --- so the total tape space used is actually sizeof(unsigned int)
417 * more than the stored length value. This allows read-backwards. When
418 * randomAccess is not true, the write/read routines may omit the extra
421 * writetup is expected to write both length words as well as the tuple
422 * data. When readtup is called, the tape is positioned just after the
423 * front length word; readtup must read the tuple data and advance past
424 * the back length word (if present).
426 * The write/read routines can make use of the tuple description data
427 * stored in the Tuplesortstate record, if needed. They are also expected
428 * to adjust state->availMem by the amount of memory space (not tape space!)
429 * released or consumed. There is no error return from either writetup
430 * or readtup; they should ereport() on failure.
433 * NOTES about memory consumption calculations:
435 * We count space allocated for tuples against the workMem limit, plus
436 * the space used by the variable-size memtuples array. Fixed-size space
437 * is not counted; it's small enough to not be interesting.
439 * Note that we count actual space used (as shown by GetMemoryChunkSpace)
440 * rather than the originally-requested size. This is important since
441 * palloc can add substantial overhead. It's not a complete answer since
442 * we won't count any wasted space in palloc allocation blocks, but it's
443 * a lot better than what we were doing before 7.3.
446 /* When using this macro, beware of double evaluation of len */
447 #define LogicalTapeReadExact(tapeset, tapenum, ptr, len) \
449 if (LogicalTapeRead(tapeset, tapenum, ptr, len) != (size_t) (len)) \
450 elog(ERROR, "unexpected end of data"); \
454 static Tuplesortstate *tuplesort_begin_common(int workMem, bool randomAccess);
455 static void puttuple_common(Tuplesortstate *state, SortTuple *tuple);
456 static bool consider_abort_common(Tuplesortstate *state);
457 static void inittapes(Tuplesortstate *state);
458 static void selectnewtape(Tuplesortstate *state);
459 static void mergeruns(Tuplesortstate *state);
460 static void mergeonerun(Tuplesortstate *state);
461 static void beginmerge(Tuplesortstate *state);
462 static void mergepreread(Tuplesortstate *state);
463 static void mergeprereadone(Tuplesortstate *state, int srcTape);
464 static void dumptuples(Tuplesortstate *state, bool alltuples);
465 static void make_bounded_heap(Tuplesortstate *state);
466 static void sort_bounded_heap(Tuplesortstate *state);
467 static void tuplesort_heap_insert(Tuplesortstate *state, SortTuple *tuple,
468 int tupleindex, bool checkIndex);
469 static void tuplesort_heap_siftup(Tuplesortstate *state, bool checkIndex);
470 static void reversedirection(Tuplesortstate *state);
471 static unsigned int getlen(Tuplesortstate *state, int tapenum, bool eofOK);
472 static void markrunend(Tuplesortstate *state, int tapenum);
473 static int comparetup_heap(const SortTuple *a, const SortTuple *b,
474 Tuplesortstate *state);
475 static void copytup_heap(Tuplesortstate *state, SortTuple *stup, void *tup);
476 static void writetup_heap(Tuplesortstate *state, int tapenum,
478 static void readtup_heap(Tuplesortstate *state, SortTuple *stup,
479 int tapenum, unsigned int len);
480 static int comparetup_cluster(const SortTuple *a, const SortTuple *b,
481 Tuplesortstate *state);
482 static void copytup_cluster(Tuplesortstate *state, SortTuple *stup, void *tup);
483 static void writetup_cluster(Tuplesortstate *state, int tapenum,
485 static void readtup_cluster(Tuplesortstate *state, SortTuple *stup,
486 int tapenum, unsigned int len);
487 static int comparetup_index_btree(const SortTuple *a, const SortTuple *b,
488 Tuplesortstate *state);
489 static int comparetup_index_hash(const SortTuple *a, const SortTuple *b,
490 Tuplesortstate *state);
491 static void copytup_index(Tuplesortstate *state, SortTuple *stup, void *tup);
492 static void writetup_index(Tuplesortstate *state, int tapenum,
494 static void readtup_index(Tuplesortstate *state, SortTuple *stup,
495 int tapenum, unsigned int len);
496 static int comparetup_datum(const SortTuple *a, const SortTuple *b,
497 Tuplesortstate *state);
498 static void copytup_datum(Tuplesortstate *state, SortTuple *stup, void *tup);
499 static void writetup_datum(Tuplesortstate *state, int tapenum,
501 static void readtup_datum(Tuplesortstate *state, SortTuple *stup,
502 int tapenum, unsigned int len);
503 static void free_sort_tuple(Tuplesortstate *state, SortTuple *stup);
506 * Special versions of qsort just for SortTuple objects. qsort_tuple() sorts
507 * any variant of SortTuples, using the appropriate comparetup function.
508 * qsort_ssup() is specialized for the case where the comparetup function
509 * reduces to ApplySortComparator(), that is single-key MinimalTuple sorts
512 #include "qsort_tuple.c"
516 * tuplesort_begin_xxx
518 * Initialize for a tuple sort operation.
520 * After calling tuplesort_begin, the caller should call tuplesort_putXXX
521 * zero or more times, then call tuplesort_performsort when all the tuples
522 * have been supplied. After performsort, retrieve the tuples in sorted
523 * order by calling tuplesort_getXXX until it returns false/NULL. (If random
524 * access was requested, rescan, markpos, and restorepos can also be called.)
525 * Call tuplesort_end to terminate the operation and release memory/disk space.
527 * Each variant of tuplesort_begin has a workMem parameter specifying the
528 * maximum number of kilobytes of RAM to use before spilling data to disk.
529 * (The normal value of this parameter is work_mem, but some callers use
530 * other values.) Each variant also has a randomAccess parameter specifying
531 * whether the caller needs non-sequential access to the sort result.
534 static Tuplesortstate *
535 tuplesort_begin_common(int workMem, bool randomAccess)
537 Tuplesortstate *state;
538 MemoryContext sortcontext;
539 MemoryContext oldcontext;
542 * Create a working memory context for this sort operation. All data
543 * needed by the sort will live inside this context.
545 sortcontext = AllocSetContextCreate(CurrentMemoryContext,
547 ALLOCSET_DEFAULT_MINSIZE,
548 ALLOCSET_DEFAULT_INITSIZE,
549 ALLOCSET_DEFAULT_MAXSIZE);
552 * Make the Tuplesortstate within the per-sort context. This way, we
553 * don't need a separate pfree() operation for it at shutdown.
555 oldcontext = MemoryContextSwitchTo(sortcontext);
557 state = (Tuplesortstate *) palloc0(sizeof(Tuplesortstate));
561 pg_rusage_init(&state->ru_start);
564 state->status = TSS_INITIAL;
565 state->randomAccess = randomAccess;
566 state->bounded = false;
567 state->boundUsed = false;
568 state->allowedMem = workMem * (int64) 1024;
569 state->availMem = state->allowedMem;
570 state->sortcontext = sortcontext;
571 state->tapeset = NULL;
573 state->memtupcount = 0;
574 state->memtupsize = 1024; /* initial guess */
575 state->growmemtuples = true;
576 state->memtuples = (SortTuple *) palloc(state->memtupsize * sizeof(SortTuple));
578 USEMEM(state, GetMemoryChunkSpace(state->memtuples));
580 /* workMem must be large enough for the minimal memtuples array */
582 elog(ERROR, "insufficient memory allowed for sort");
584 state->currentRun = 0;
587 * maxTapes, tapeRange, and Algorithm D variables will be initialized by
588 * inittapes(), if needed
591 state->result_tape = -1; /* flag that result tape has not been formed */
593 MemoryContextSwitchTo(oldcontext);
599 tuplesort_begin_heap(TupleDesc tupDesc,
600 int nkeys, AttrNumber *attNums,
601 Oid *sortOperators, Oid *sortCollations,
602 bool *nullsFirstFlags,
603 int workMem, bool randomAccess)
605 Tuplesortstate *state = tuplesort_begin_common(workMem, randomAccess);
606 MemoryContext oldcontext;
609 oldcontext = MemoryContextSwitchTo(state->sortcontext);
611 AssertArg(nkeys > 0);
616 "begin tuple sort: nkeys = %d, workMem = %d, randomAccess = %c",
617 nkeys, workMem, randomAccess ? 't' : 'f');
620 state->nKeys = nkeys;
622 TRACE_POSTGRESQL_SORT_START(HEAP_SORT,
623 false, /* no unique check */
628 state->comparetup = comparetup_heap;
629 state->copytup = copytup_heap;
630 state->writetup = writetup_heap;
631 state->readtup = readtup_heap;
633 state->tupDesc = tupDesc; /* assume we need not copy tupDesc */
634 state->abbrevNext = 10;
636 /* Prepare SortSupport data for each column */
637 state->sortKeys = (SortSupport) palloc0(nkeys * sizeof(SortSupportData));
639 for (i = 0; i < nkeys; i++)
641 SortSupport sortKey = state->sortKeys + i;
643 AssertArg(attNums[i] != 0);
644 AssertArg(sortOperators[i] != 0);
646 sortKey->ssup_cxt = CurrentMemoryContext;
647 sortKey->ssup_collation = sortCollations[i];
648 sortKey->ssup_nulls_first = nullsFirstFlags[i];
649 sortKey->ssup_attno = attNums[i];
650 /* Convey if abbreviation optimization is applicable in principle */
651 sortKey->abbreviate = (i == 0);
653 PrepareSortSupportFromOrderingOp(sortOperators[i], sortKey);
657 * The "onlyKey" optimization cannot be used with abbreviated keys, since
658 * tie-breaker comparisons may be required. Typically, the optimization is
659 * only of value to pass-by-value types anyway, whereas abbreviated keys
660 * are typically only of value to pass-by-reference types.
662 if (nkeys == 1 && !state->sortKeys->abbrev_converter)
663 state->onlyKey = state->sortKeys;
665 MemoryContextSwitchTo(oldcontext);
671 tuplesort_begin_cluster(TupleDesc tupDesc,
673 int workMem, bool randomAccess)
675 Tuplesortstate *state = tuplesort_begin_common(workMem, randomAccess);
676 ScanKey indexScanKey;
677 MemoryContext oldcontext;
680 Assert(indexRel->rd_rel->relam == BTREE_AM_OID);
682 oldcontext = MemoryContextSwitchTo(state->sortcontext);
687 "begin tuple sort: nkeys = %d, workMem = %d, randomAccess = %c",
688 RelationGetNumberOfAttributes(indexRel),
689 workMem, randomAccess ? 't' : 'f');
692 state->nKeys = RelationGetNumberOfAttributes(indexRel);
694 TRACE_POSTGRESQL_SORT_START(CLUSTER_SORT,
695 false, /* no unique check */
700 state->comparetup = comparetup_cluster;
701 state->copytup = copytup_cluster;
702 state->writetup = writetup_cluster;
703 state->readtup = readtup_cluster;
704 state->abbrevNext = 10;
706 state->indexInfo = BuildIndexInfo(indexRel);
708 state->tupDesc = tupDesc; /* assume we need not copy tupDesc */
710 indexScanKey = _bt_mkscankey_nodata(indexRel);
712 if (state->indexInfo->ii_Expressions != NULL)
714 TupleTableSlot *slot;
715 ExprContext *econtext;
718 * We will need to use FormIndexDatum to evaluate the index
719 * expressions. To do that, we need an EState, as well as a
720 * TupleTableSlot to put the table tuples into. The econtext's
721 * scantuple has to point to that slot, too.
723 state->estate = CreateExecutorState();
724 slot = MakeSingleTupleTableSlot(tupDesc);
725 econtext = GetPerTupleExprContext(state->estate);
726 econtext->ecxt_scantuple = slot;
729 /* Prepare SortSupport data for each column */
730 state->sortKeys = (SortSupport) palloc0(state->nKeys *
731 sizeof(SortSupportData));
733 for (i = 0; i < state->nKeys; i++)
735 SortSupport sortKey = state->sortKeys + i;
736 ScanKey scanKey = indexScanKey + i;
739 sortKey->ssup_cxt = CurrentMemoryContext;
740 sortKey->ssup_collation = scanKey->sk_collation;
741 sortKey->ssup_nulls_first =
742 (scanKey->sk_flags & SK_BT_NULLS_FIRST) != 0;
743 sortKey->ssup_attno = scanKey->sk_attno;
744 /* Convey if abbreviation optimization is applicable in principle */
745 sortKey->abbreviate = (i == 0);
747 AssertState(sortKey->ssup_attno != 0);
749 strategy = (scanKey->sk_flags & SK_BT_DESC) != 0 ?
750 BTGreaterStrategyNumber : BTLessStrategyNumber;
752 PrepareSortSupportFromIndexRel(indexRel, strategy, sortKey);
755 _bt_freeskey(indexScanKey);
757 MemoryContextSwitchTo(oldcontext);
763 tuplesort_begin_index_btree(Relation heapRel,
766 int workMem, bool randomAccess)
768 Tuplesortstate *state = tuplesort_begin_common(workMem, randomAccess);
769 ScanKey indexScanKey;
770 MemoryContext oldcontext;
773 oldcontext = MemoryContextSwitchTo(state->sortcontext);
778 "begin index sort: unique = %c, workMem = %d, randomAccess = %c",
779 enforceUnique ? 't' : 'f',
780 workMem, randomAccess ? 't' : 'f');
783 state->nKeys = RelationGetNumberOfAttributes(indexRel);
785 TRACE_POSTGRESQL_SORT_START(INDEX_SORT,
791 state->comparetup = comparetup_index_btree;
792 state->copytup = copytup_index;
793 state->writetup = writetup_index;
794 state->readtup = readtup_index;
795 state->abbrevNext = 10;
797 state->heapRel = heapRel;
798 state->indexRel = indexRel;
799 state->enforceUnique = enforceUnique;
801 indexScanKey = _bt_mkscankey_nodata(indexRel);
802 state->nKeys = RelationGetNumberOfAttributes(indexRel);
804 /* Prepare SortSupport data for each column */
805 state->sortKeys = (SortSupport) palloc0(state->nKeys *
806 sizeof(SortSupportData));
808 for (i = 0; i < state->nKeys; i++)
810 SortSupport sortKey = state->sortKeys + i;
811 ScanKey scanKey = indexScanKey + i;
814 sortKey->ssup_cxt = CurrentMemoryContext;
815 sortKey->ssup_collation = scanKey->sk_collation;
816 sortKey->ssup_nulls_first =
817 (scanKey->sk_flags & SK_BT_NULLS_FIRST) != 0;
818 sortKey->ssup_attno = scanKey->sk_attno;
819 /* Convey if abbreviation optimization is applicable in principle */
820 sortKey->abbreviate = (i == 0);
822 AssertState(sortKey->ssup_attno != 0);
824 strategy = (scanKey->sk_flags & SK_BT_DESC) != 0 ?
825 BTGreaterStrategyNumber : BTLessStrategyNumber;
827 PrepareSortSupportFromIndexRel(indexRel, strategy, sortKey);
830 _bt_freeskey(indexScanKey);
832 MemoryContextSwitchTo(oldcontext);
838 tuplesort_begin_index_hash(Relation heapRel,
841 int workMem, bool randomAccess)
843 Tuplesortstate *state = tuplesort_begin_common(workMem, randomAccess);
844 MemoryContext oldcontext;
846 oldcontext = MemoryContextSwitchTo(state->sortcontext);
851 "begin index sort: hash_mask = 0x%x, workMem = %d, randomAccess = %c",
853 workMem, randomAccess ? 't' : 'f');
856 state->nKeys = 1; /* Only one sort column, the hash code */
858 state->comparetup = comparetup_index_hash;
859 state->copytup = copytup_index;
860 state->writetup = writetup_index;
861 state->readtup = readtup_index;
863 state->heapRel = heapRel;
864 state->indexRel = indexRel;
866 state->hash_mask = hash_mask;
868 MemoryContextSwitchTo(oldcontext);
874 tuplesort_begin_datum(Oid datumType, Oid sortOperator, Oid sortCollation,
876 int workMem, bool randomAccess)
878 Tuplesortstate *state = tuplesort_begin_common(workMem, randomAccess);
879 MemoryContext oldcontext;
883 oldcontext = MemoryContextSwitchTo(state->sortcontext);
888 "begin datum sort: workMem = %d, randomAccess = %c",
889 workMem, randomAccess ? 't' : 'f');
892 state->nKeys = 1; /* always a one-column sort */
894 TRACE_POSTGRESQL_SORT_START(DATUM_SORT,
895 false, /* no unique check */
900 state->comparetup = comparetup_datum;
901 state->copytup = copytup_datum;
902 state->writetup = writetup_datum;
903 state->readtup = readtup_datum;
905 state->datumType = datumType;
907 /* Prepare SortSupport data */
908 state->onlyKey = (SortSupport) palloc0(sizeof(SortSupportData));
910 state->onlyKey->ssup_cxt = CurrentMemoryContext;
911 state->onlyKey->ssup_collation = sortCollation;
912 state->onlyKey->ssup_nulls_first = nullsFirstFlag;
914 * Conversion to abbreviated representation infeasible in the Datum case.
915 * It must be possible to subsequently fetch original datum values within
916 * tuplesort_getdatum(), which would require special-case preservation of
919 state->onlyKey->abbreviate = false;
921 PrepareSortSupportFromOrderingOp(sortOperator, state->onlyKey);
923 /* lookup necessary attributes of the datum type */
924 get_typlenbyval(datumType, &typlen, &typbyval);
925 state->datumTypeLen = typlen;
926 state->datumTypeByVal = typbyval;
928 MemoryContextSwitchTo(oldcontext);
934 * tuplesort_set_bound
936 * Advise tuplesort that at most the first N result tuples are required.
938 * Must be called before inserting any tuples. (Actually, we could allow it
939 * as long as the sort hasn't spilled to disk, but there seems no need for
940 * delayed calls at the moment.)
942 * This is a hint only. The tuplesort may still return more tuples than
946 tuplesort_set_bound(Tuplesortstate *state, int64 bound)
948 /* Assert we're called before loading any tuples */
949 Assert(state->status == TSS_INITIAL);
950 Assert(state->memtupcount == 0);
951 Assert(!state->bounded);
953 #ifdef DEBUG_BOUNDED_SORT
954 /* Honor GUC setting that disables the feature (for easy testing) */
955 if (!optimize_bounded_sort)
959 /* We want to be able to compute bound * 2, so limit the setting */
960 if (bound > (int64) (INT_MAX / 2))
963 state->bounded = true;
964 state->bound = (int) bound;
967 * Bounded sorts are not an effective target for abbreviated key
968 * optimization. Disable by setting state to be consistent with no
969 * abbreviation support.
971 state->sortKeys->abbrev_converter = NULL;
972 if (state->sortKeys->abbrev_full_comparator)
973 state->sortKeys->comparator = state->sortKeys->abbrev_full_comparator;
975 /* Not strictly necessary, but be tidy */
976 state->sortKeys->abbrev_abort = NULL;
977 state->sortKeys->abbrev_full_comparator = NULL;
983 * Release resources and clean up.
985 * NOTE: after calling this, any pointers returned by tuplesort_getXXX are
986 * pointing to garbage. Be careful not to attempt to use or free such
987 * pointers afterwards!
990 tuplesort_end(Tuplesortstate *state)
992 /* context swap probably not needed, but let's be safe */
993 MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
999 spaceUsed = LogicalTapeSetBlocks(state->tapeset);
1001 spaceUsed = (state->allowedMem - state->availMem + 1023) / 1024;
1005 * Delete temporary "tape" files, if any.
1007 * Note: want to include this in reported total cost of sort, hence need
1008 * for two #ifdef TRACE_SORT sections.
1011 LogicalTapeSetClose(state->tapeset);
1017 elog(LOG, "external sort ended, %ld disk blocks used: %s",
1018 spaceUsed, pg_rusage_show(&state->ru_start));
1020 elog(LOG, "internal sort ended, %ld KB used: %s",
1021 spaceUsed, pg_rusage_show(&state->ru_start));
1024 TRACE_POSTGRESQL_SORT_DONE(state->tapeset != NULL, spaceUsed);
1028 * If you disabled TRACE_SORT, you can still probe sort__done, but you
1029 * ain't getting space-used stats.
1031 TRACE_POSTGRESQL_SORT_DONE(state->tapeset != NULL, 0L);
1034 /* Free any execution state created for CLUSTER case */
1035 if (state->estate != NULL)
1037 ExprContext *econtext = GetPerTupleExprContext(state->estate);
1039 ExecDropSingleTupleTableSlot(econtext->ecxt_scantuple);
1040 FreeExecutorState(state->estate);
1043 MemoryContextSwitchTo(oldcontext);
1046 * Free the per-sort memory context, thereby releasing all working memory,
1047 * including the Tuplesortstate struct itself.
1049 MemoryContextDelete(state->sortcontext);
1053 * Grow the memtuples[] array, if possible within our memory constraint. We
1054 * must not exceed INT_MAX tuples in memory or the caller-provided memory
1055 * limit. Return TRUE if we were able to enlarge the array, FALSE if not.
1057 * Normally, at each increment we double the size of the array. When doing
1058 * that would exceed a limit, we attempt one last, smaller increase (and then
1059 * clear the growmemtuples flag so we don't try any more). That allows us to
1060 * use memory as fully as permitted; sticking to the pure doubling rule could
1061 * result in almost half going unused. Because availMem moves around with
1062 * tuple addition/removal, we need some rule to prevent making repeated small
1063 * increases in memtupsize, which would just be useless thrashing. The
1064 * growmemtuples flag accomplishes that and also prevents useless
1065 * recalculations in this function.
1068 grow_memtuples(Tuplesortstate *state)
1071 int memtupsize = state->memtupsize;
1072 int64 memNowUsed = state->allowedMem - state->availMem;
1074 /* Forget it if we've already maxed out memtuples, per comment above */
1075 if (!state->growmemtuples)
1078 /* Select new value of memtupsize */
1079 if (memNowUsed <= state->availMem)
1082 * We've used no more than half of allowedMem; double our usage,
1083 * clamping at INT_MAX tuples.
1085 if (memtupsize < INT_MAX / 2)
1086 newmemtupsize = memtupsize * 2;
1089 newmemtupsize = INT_MAX;
1090 state->growmemtuples = false;
1096 * This will be the last increment of memtupsize. Abandon doubling
1097 * strategy and instead increase as much as we safely can.
1099 * To stay within allowedMem, we can't increase memtupsize by more
1100 * than availMem / sizeof(SortTuple) elements. In practice, we want
1101 * to increase it by considerably less, because we need to leave some
1102 * space for the tuples to which the new array slots will refer. We
1103 * assume the new tuples will be about the same size as the tuples
1104 * we've already seen, and thus we can extrapolate from the space
1105 * consumption so far to estimate an appropriate new size for the
1106 * memtuples array. The optimal value might be higher or lower than
1107 * this estimate, but it's hard to know that in advance. We again
1108 * clamp at INT_MAX tuples.
1110 * This calculation is safe against enlarging the array so much that
1111 * LACKMEM becomes true, because the memory currently used includes
1112 * the present array; thus, there would be enough allowedMem for the
1113 * new array elements even if no other memory were currently used.
1115 * We do the arithmetic in float8, because otherwise the product of
1116 * memtupsize and allowedMem could overflow. Any inaccuracy in the
1117 * result should be insignificant; but even if we computed a
1118 * completely insane result, the checks below will prevent anything
1119 * really bad from happening.
1123 grow_ratio = (double) state->allowedMem / (double) memNowUsed;
1124 if (memtupsize * grow_ratio < INT_MAX)
1125 newmemtupsize = (int) (memtupsize * grow_ratio);
1127 newmemtupsize = INT_MAX;
1129 /* We won't make any further enlargement attempts */
1130 state->growmemtuples = false;
1133 /* Must enlarge array by at least one element, else report failure */
1134 if (newmemtupsize <= memtupsize)
1138 * On a 32-bit machine, allowedMem could exceed MaxAllocHugeSize. Clamp
1139 * to ensure our request won't be rejected. Note that we can easily
1140 * exhaust address space before facing this outcome. (This is presently
1141 * impossible due to guc.c's MAX_KILOBYTES limitation on work_mem, but
1142 * don't rely on that at this distance.)
1144 if ((Size) newmemtupsize >= MaxAllocHugeSize / sizeof(SortTuple))
1146 newmemtupsize = (int) (MaxAllocHugeSize / sizeof(SortTuple));
1147 state->growmemtuples = false; /* can't grow any more */
1151 * We need to be sure that we do not cause LACKMEM to become true, else
1152 * the space management algorithm will go nuts. The code above should
1153 * never generate a dangerous request, but to be safe, check explicitly
1154 * that the array growth fits within availMem. (We could still cause
1155 * LACKMEM if the memory chunk overhead associated with the memtuples
1156 * array were to increase. That shouldn't happen with any sane value of
1157 * allowedMem, because at any array size large enough to risk LACKMEM,
1158 * palloc would be treating both old and new arrays as separate chunks.
1159 * But we'll check LACKMEM explicitly below just in case.)
1161 if (state->availMem < (int64) ((newmemtupsize - memtupsize) * sizeof(SortTuple)))
1165 FREEMEM(state, GetMemoryChunkSpace(state->memtuples));
1166 state->memtupsize = newmemtupsize;
1167 state->memtuples = (SortTuple *)
1168 repalloc_huge(state->memtuples,
1169 state->memtupsize * sizeof(SortTuple));
1170 USEMEM(state, GetMemoryChunkSpace(state->memtuples));
1172 elog(ERROR, "unexpected out-of-memory situation during sort");
1176 /* If for any reason we didn't realloc, shut off future attempts */
1177 state->growmemtuples = false;
1182 * Accept one tuple while collecting input data for sort.
1184 * Note that the input data is always copied; the caller need not save it.
1187 tuplesort_puttupleslot(Tuplesortstate *state, TupleTableSlot *slot)
1189 MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
1193 * Copy the given tuple into memory we control, and decrease availMem.
1194 * Then call the common code.
1196 COPYTUP(state, &stup, (void *) slot);
1198 puttuple_common(state, &stup);
1200 MemoryContextSwitchTo(oldcontext);
1204 * Accept one tuple while collecting input data for sort.
1206 * Note that the input data is always copied; the caller need not save it.
1209 tuplesort_putheaptuple(Tuplesortstate *state, HeapTuple tup)
1211 MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
1215 * Copy the given tuple into memory we control, and decrease availMem.
1216 * Then call the common code.
1218 COPYTUP(state, &stup, (void *) tup);
1220 puttuple_common(state, &stup);
1222 MemoryContextSwitchTo(oldcontext);
1226 * Collect one index tuple while collecting input data for sort, building
1227 * it from caller-supplied values.
1230 tuplesort_putindextuplevalues(Tuplesortstate *state, Relation rel,
1231 ItemPointer self, Datum *values,
1234 MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
1239 stup.tuple = index_form_tuple(RelationGetDescr(rel), values, isnull);
1240 tuple = ((IndexTuple) stup.tuple);
1241 tuple->t_tid = *self;
1242 USEMEM(state, GetMemoryChunkSpace(stup.tuple));
1243 /* set up first-column key value */
1244 original = index_getattr(tuple,
1246 RelationGetDescr(state->indexRel),
1249 if (!state->sortKeys || !state->sortKeys->abbrev_converter || stup.isnull1)
1252 * Store ordinary Datum representation, or NULL value. If there is a
1253 * converter it won't expect NULL values, and cost model is not
1254 * required to account for NULL, so in that case we avoid calling
1255 * converter and just set datum1 to "void" representation (to be
1258 stup.datum1 = original;
1260 else if (!consider_abort_common(state))
1262 /* Store abbreviated key representation */
1263 stup.datum1 = state->sortKeys->abbrev_converter(original,
1268 /* Abort abbreviation */
1271 stup.datum1 = original;
1274 * Set state to be consistent with never trying abbreviation.
1276 * Alter datum1 representation in already-copied tuples, so as to
1277 * ensure a consistent representation (current tuple was just handled).
1278 * Note that we rely on all tuples copied so far actually being
1279 * contained within memtuples array.
1281 for (i = 0; i < state->memtupcount; i++)
1283 SortTuple *mtup = &state->memtuples[i];
1285 tuple = mtup->tuple;
1286 mtup->datum1 = index_getattr(tuple,
1288 RelationGetDescr(state->indexRel),
1293 puttuple_common(state, &stup);
1295 MemoryContextSwitchTo(oldcontext);
1299 * Accept one Datum while collecting input data for sort.
1301 * If the Datum is pass-by-ref type, the value will be copied.
1304 tuplesort_putdatum(Tuplesortstate *state, Datum val, bool isNull)
1306 MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
1310 * If it's a pass-by-reference value, copy it into memory we control, and
1311 * decrease availMem. Then call the common code.
1313 if (isNull || state->datumTypeByVal)
1316 stup.isnull1 = isNull;
1317 stup.tuple = NULL; /* no separate storage */
1321 stup.datum1 = datumCopy(val, false, state->datumTypeLen);
1322 stup.isnull1 = false;
1323 stup.tuple = DatumGetPointer(stup.datum1);
1324 USEMEM(state, GetMemoryChunkSpace(stup.tuple));
1327 puttuple_common(state, &stup);
1329 MemoryContextSwitchTo(oldcontext);
1333 * Shared code for tuple and datum cases.
1336 puttuple_common(Tuplesortstate *state, SortTuple *tuple)
1338 switch (state->status)
1343 * Save the tuple into the unsorted array. First, grow the array
1344 * as needed. Note that we try to grow the array when there is
1345 * still one free slot remaining --- if we fail, there'll still be
1346 * room to store the incoming tuple, and then we'll switch to
1347 * tape-based operation.
1349 if (state->memtupcount >= state->memtupsize - 1)
1351 (void) grow_memtuples(state);
1352 Assert(state->memtupcount < state->memtupsize);
1354 state->memtuples[state->memtupcount++] = *tuple;
1357 * Check if it's time to switch over to a bounded heapsort. We do
1358 * so if the input tuple count exceeds twice the desired tuple
1359 * count (this is a heuristic for where heapsort becomes cheaper
1360 * than a quicksort), or if we've just filled workMem and have
1361 * enough tuples to meet the bound.
1363 * Note that once we enter TSS_BOUNDED state we will always try to
1364 * complete the sort that way. In the worst case, if later input
1365 * tuples are larger than earlier ones, this might cause us to
1366 * exceed workMem significantly.
1368 if (state->bounded &&
1369 (state->memtupcount > state->bound * 2 ||
1370 (state->memtupcount > state->bound && LACKMEM(state))))
1374 elog(LOG, "switching to bounded heapsort at %d tuples: %s",
1376 pg_rusage_show(&state->ru_start));
1378 make_bounded_heap(state);
1383 * Done if we still fit in available memory and have array slots.
1385 if (state->memtupcount < state->memtupsize && !LACKMEM(state))
1389 * Nope; time to switch to tape-based operation.
1394 * Dump tuples until we are back under the limit.
1396 dumptuples(state, false);
1402 * We don't want to grow the array here, so check whether the new
1403 * tuple can be discarded before putting it in. This should be a
1404 * good speed optimization, too, since when there are many more
1405 * input tuples than the bound, most input tuples can be discarded
1406 * with just this one comparison. Note that because we currently
1407 * have the sort direction reversed, we must check for <= not >=.
1409 if (COMPARETUP(state, tuple, &state->memtuples[0]) <= 0)
1411 /* new tuple <= top of the heap, so we can discard it */
1412 free_sort_tuple(state, tuple);
1413 CHECK_FOR_INTERRUPTS();
1417 /* discard top of heap, sift up, insert new tuple */
1418 free_sort_tuple(state, &state->memtuples[0]);
1419 tuplesort_heap_siftup(state, false);
1420 tuplesort_heap_insert(state, tuple, 0, false);
1427 * Insert the tuple into the heap, with run number currentRun if
1428 * it can go into the current run, else run number currentRun+1.
1429 * The tuple can go into the current run if it is >= the first
1430 * not-yet-output tuple. (Actually, it could go into the current
1431 * run if it is >= the most recently output tuple ... but that
1432 * would require keeping around the tuple we last output, and it's
1433 * simplest to let writetup free each tuple as soon as it's
1436 * Note there will always be at least one tuple in the heap at
1437 * this point; see dumptuples.
1439 Assert(state->memtupcount > 0);
1440 if (COMPARETUP(state, tuple, &state->memtuples[0]) >= 0)
1441 tuplesort_heap_insert(state, tuple, state->currentRun, true);
1443 tuplesort_heap_insert(state, tuple, state->currentRun + 1, true);
1446 * If we are over the memory limit, dump tuples till we're under.
1448 dumptuples(state, false);
1452 elog(ERROR, "invalid tuplesort state");
1458 consider_abort_common(Tuplesortstate *state)
1460 Assert(state->sortKeys[0].abbrev_converter != NULL);
1461 Assert(state->sortKeys[0].abbrev_abort != NULL);
1462 Assert(state->sortKeys[0].abbrev_full_comparator != NULL);
1465 * Check effectiveness of abbreviation optimization. Consider aborting
1466 * when still within memory limit.
1468 if (state->status == TSS_INITIAL &&
1469 state->memtupcount >= state->abbrevNext)
1471 state->abbrevNext *= 2;
1474 * Check opclass-supplied abbreviation abort routine. It may
1475 * indicate that abbreviation should not proceed.
1477 if (!state->sortKeys->abbrev_abort(state->memtupcount,
1482 * Finally, restore authoritative comparator, and indicate that
1483 * abbreviation is not in play by setting abbrev_converter to NULL
1485 state->sortKeys[0].comparator = state->sortKeys[0].abbrev_full_comparator;
1486 state->sortKeys[0].abbrev_converter = NULL;
1487 /* Not strictly necessary, but be tidy */
1488 state->sortKeys[0].abbrev_abort = NULL;
1489 state->sortKeys[0].abbrev_full_comparator = NULL;
1491 /* Give up - expect original pass-by-value representation */
1499 * All tuples have been provided; finish the sort.
1502 tuplesort_performsort(Tuplesortstate *state)
1504 MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
1508 elog(LOG, "performsort starting: %s",
1509 pg_rusage_show(&state->ru_start));
1512 switch (state->status)
1517 * We were able to accumulate all the tuples within the allowed
1518 * amount of memory. Just qsort 'em and we're done.
1520 if (state->memtupcount > 1)
1522 /* Can we use the single-key sort function? */
1523 if (state->onlyKey != NULL)
1524 qsort_ssup(state->memtuples, state->memtupcount,
1527 qsort_tuple(state->memtuples,
1533 state->eof_reached = false;
1534 state->markpos_offset = 0;
1535 state->markpos_eof = false;
1536 state->status = TSS_SORTEDINMEM;
1542 * We were able to accumulate all the tuples required for output
1543 * in memory, using a heap to eliminate excess tuples. Now we
1544 * have to transform the heap to a properly-sorted array.
1546 sort_bounded_heap(state);
1548 state->eof_reached = false;
1549 state->markpos_offset = 0;
1550 state->markpos_eof = false;
1551 state->status = TSS_SORTEDINMEM;
1557 * Finish tape-based sort. First, flush all tuples remaining in
1558 * memory out to tape; then merge until we have a single remaining
1559 * run (or, if !randomAccess, one run per tape). Note that
1560 * mergeruns sets the correct state->status.
1562 dumptuples(state, true);
1564 state->eof_reached = false;
1565 state->markpos_block = 0L;
1566 state->markpos_offset = 0;
1567 state->markpos_eof = false;
1571 elog(ERROR, "invalid tuplesort state");
1578 if (state->status == TSS_FINALMERGE)
1579 elog(LOG, "performsort done (except %d-way final merge): %s",
1581 pg_rusage_show(&state->ru_start));
1583 elog(LOG, "performsort done: %s",
1584 pg_rusage_show(&state->ru_start));
1588 MemoryContextSwitchTo(oldcontext);
1592 * Internal routine to fetch the next tuple in either forward or back
1593 * direction into *stup. Returns FALSE if no more tuples.
1594 * If *should_free is set, the caller must pfree stup.tuple when done with it.
1597 tuplesort_gettuple_common(Tuplesortstate *state, bool forward,
1598 SortTuple *stup, bool *should_free)
1600 unsigned int tuplen;
1602 switch (state->status)
1604 case TSS_SORTEDINMEM:
1605 Assert(forward || state->randomAccess);
1606 *should_free = false;
1609 if (state->current < state->memtupcount)
1611 *stup = state->memtuples[state->current++];
1614 state->eof_reached = true;
1617 * Complain if caller tries to retrieve more tuples than
1618 * originally asked for in a bounded sort. This is because
1619 * returning EOF here might be the wrong thing.
1621 if (state->bounded && state->current >= state->bound)
1622 elog(ERROR, "retrieved too many tuples in a bounded sort");
1628 if (state->current <= 0)
1632 * if all tuples are fetched already then we return last
1633 * tuple, else - tuple before last returned.
1635 if (state->eof_reached)
1636 state->eof_reached = false;
1639 state->current--; /* last returned tuple */
1640 if (state->current <= 0)
1643 *stup = state->memtuples[state->current - 1];
1648 case TSS_SORTEDONTAPE:
1649 Assert(forward || state->randomAccess);
1650 *should_free = true;
1653 if (state->eof_reached)
1655 if ((tuplen = getlen(state, state->result_tape, true)) != 0)
1657 READTUP(state, stup, state->result_tape, tuplen);
1662 state->eof_reached = true;
1670 * if all tuples are fetched already then we return last tuple,
1671 * else - tuple before last returned.
1673 if (state->eof_reached)
1676 * Seek position is pointing just past the zero tuplen at the
1677 * end of file; back up to fetch last tuple's ending length
1678 * word. If seek fails we must have a completely empty file.
1680 if (!LogicalTapeBackspace(state->tapeset,
1682 2 * sizeof(unsigned int)))
1684 state->eof_reached = false;
1689 * Back up and fetch previously-returned tuple's ending length
1690 * word. If seek fails, assume we are at start of file.
1692 if (!LogicalTapeBackspace(state->tapeset,
1694 sizeof(unsigned int)))
1696 tuplen = getlen(state, state->result_tape, false);
1699 * Back up to get ending length word of tuple before it.
1701 if (!LogicalTapeBackspace(state->tapeset,
1703 tuplen + 2 * sizeof(unsigned int)))
1706 * If that fails, presumably the prev tuple is the first
1707 * in the file. Back up so that it becomes next to read
1708 * in forward direction (not obviously right, but that is
1709 * what in-memory case does).
1711 if (!LogicalTapeBackspace(state->tapeset,
1713 tuplen + sizeof(unsigned int)))
1714 elog(ERROR, "bogus tuple length in backward scan");
1719 tuplen = getlen(state, state->result_tape, false);
1722 * Now we have the length of the prior tuple, back up and read it.
1723 * Note: READTUP expects we are positioned after the initial
1724 * length word of the tuple, so back up to that point.
1726 if (!LogicalTapeBackspace(state->tapeset,
1729 elog(ERROR, "bogus tuple length in backward scan");
1730 READTUP(state, stup, state->result_tape, tuplen);
1733 case TSS_FINALMERGE:
1735 *should_free = true;
1738 * This code should match the inner loop of mergeonerun().
1740 if (state->memtupcount > 0)
1742 int srcTape = state->memtuples[0].tupindex;
1747 *stup = state->memtuples[0];
1748 /* returned tuple is no longer counted in our memory space */
1751 tuplen = GetMemoryChunkSpace(stup->tuple);
1752 state->availMem += tuplen;
1753 state->mergeavailmem[srcTape] += tuplen;
1755 tuplesort_heap_siftup(state, false);
1756 if ((tupIndex = state->mergenext[srcTape]) == 0)
1759 * out of preloaded data on this tape, try to read more
1761 * Unlike mergeonerun(), we only preload from the single
1762 * tape that's run dry. See mergepreread() comments.
1764 mergeprereadone(state, srcTape);
1767 * if still no data, we've reached end of run on this tape
1769 if ((tupIndex = state->mergenext[srcTape]) == 0)
1772 /* pull next preread tuple from list, insert in heap */
1773 newtup = &state->memtuples[tupIndex];
1774 state->mergenext[srcTape] = newtup->tupindex;
1775 if (state->mergenext[srcTape] == 0)
1776 state->mergelast[srcTape] = 0;
1777 tuplesort_heap_insert(state, newtup, srcTape, false);
1778 /* put the now-unused memtuples entry on the freelist */
1779 newtup->tupindex = state->mergefreelist;
1780 state->mergefreelist = tupIndex;
1781 state->mergeavailslots[srcTape]++;
1787 elog(ERROR, "invalid tuplesort state");
1788 return false; /* keep compiler quiet */
1793 * Fetch the next tuple in either forward or back direction.
1794 * If successful, put tuple in slot and return TRUE; else, clear the slot
1798 tuplesort_gettupleslot(Tuplesortstate *state, bool forward,
1799 TupleTableSlot *slot)
1801 MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
1805 if (!tuplesort_gettuple_common(state, forward, &stup, &should_free))
1808 MemoryContextSwitchTo(oldcontext);
1812 ExecStoreMinimalTuple((MinimalTuple) stup.tuple, slot, should_free);
1817 ExecClearTuple(slot);
1823 * Fetch the next tuple in either forward or back direction.
1824 * Returns NULL if no more tuples. If *should_free is set, the
1825 * caller must pfree the returned tuple when done with it.
1828 tuplesort_getheaptuple(Tuplesortstate *state, bool forward, bool *should_free)
1830 MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
1833 if (!tuplesort_gettuple_common(state, forward, &stup, should_free))
1836 MemoryContextSwitchTo(oldcontext);
1842 * Fetch the next index tuple in either forward or back direction.
1843 * Returns NULL if no more tuples. If *should_free is set, the
1844 * caller must pfree the returned tuple when done with it.
1847 tuplesort_getindextuple(Tuplesortstate *state, bool forward,
1850 MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
1853 if (!tuplesort_gettuple_common(state, forward, &stup, should_free))
1856 MemoryContextSwitchTo(oldcontext);
1858 return (IndexTuple) stup.tuple;
1862 * Fetch the next Datum in either forward or back direction.
1863 * Returns FALSE if no more datums.
1865 * If the Datum is pass-by-ref type, the returned value is freshly palloc'd
1866 * and is now owned by the caller.
1869 tuplesort_getdatum(Tuplesortstate *state, bool forward,
1870 Datum *val, bool *isNull)
1872 MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
1876 if (!tuplesort_gettuple_common(state, forward, &stup, &should_free))
1878 MemoryContextSwitchTo(oldcontext);
1882 if (stup.isnull1 || state->datumTypeByVal)
1885 *isNull = stup.isnull1;
1892 *val = datumCopy(stup.datum1, false, state->datumTypeLen);
1896 MemoryContextSwitchTo(oldcontext);
1902 * Advance over N tuples in either forward or back direction,
1903 * without returning any data. N==0 is a no-op.
1904 * Returns TRUE if successful, FALSE if ran out of tuples.
1907 tuplesort_skiptuples(Tuplesortstate *state, int64 ntuples, bool forward)
1909 MemoryContext oldcontext;
1912 * We don't actually support backwards skip yet, because no callers need
1913 * it. The API is designed to allow for that later, though.
1916 Assert(ntuples >= 0);
1918 switch (state->status)
1920 case TSS_SORTEDINMEM:
1921 if (state->memtupcount - state->current >= ntuples)
1923 state->current += ntuples;
1926 state->current = state->memtupcount;
1927 state->eof_reached = true;
1930 * Complain if caller tries to retrieve more tuples than
1931 * originally asked for in a bounded sort. This is because
1932 * returning EOF here might be the wrong thing.
1934 if (state->bounded && state->current >= state->bound)
1935 elog(ERROR, "retrieved too many tuples in a bounded sort");
1939 case TSS_SORTEDONTAPE:
1940 case TSS_FINALMERGE:
1943 * We could probably optimize these cases better, but for now it's
1944 * not worth the trouble.
1946 oldcontext = MemoryContextSwitchTo(state->sortcontext);
1947 while (ntuples-- > 0)
1952 if (!tuplesort_gettuple_common(state, forward,
1953 &stup, &should_free))
1955 MemoryContextSwitchTo(oldcontext);
1958 if (should_free && stup.tuple)
1960 CHECK_FOR_INTERRUPTS();
1962 MemoryContextSwitchTo(oldcontext);
1966 elog(ERROR, "invalid tuplesort state");
1967 return false; /* keep compiler quiet */
1972 * tuplesort_merge_order - report merge order we'll use for given memory
1973 * (note: "merge order" just means the number of input tapes in the merge).
1975 * This is exported for use by the planner. allowedMem is in bytes.
1978 tuplesort_merge_order(int64 allowedMem)
1983 * We need one tape for each merge input, plus another one for the output,
1984 * and each of these tapes needs buffer space. In addition we want
1985 * MERGE_BUFFER_SIZE workspace per input tape (but the output tape doesn't
1988 * Note: you might be thinking we need to account for the memtuples[]
1989 * array in this calculation, but we effectively treat that as part of the
1990 * MERGE_BUFFER_SIZE workspace.
1992 mOrder = (allowedMem - TAPE_BUFFER_OVERHEAD) /
1993 (MERGE_BUFFER_SIZE + TAPE_BUFFER_OVERHEAD);
1995 /* Even in minimum memory, use at least a MINORDER merge */
1996 mOrder = Max(mOrder, MINORDER);
2002 * inittapes - initialize for tape sorting.
2004 * This is called only if we have found we don't have room to sort in memory.
2007 inittapes(Tuplesortstate *state)
2014 /* Compute number of tapes to use: merge order plus 1 */
2015 maxTapes = tuplesort_merge_order(state->allowedMem) + 1;
2018 * We must have at least 2*maxTapes slots in the memtuples[] array, else
2019 * we'd not have room for merge heap plus preread. It seems unlikely that
2020 * this case would ever occur, but be safe.
2022 maxTapes = Min(maxTapes, state->memtupsize / 2);
2024 state->maxTapes = maxTapes;
2025 state->tapeRange = maxTapes - 1;
2029 elog(LOG, "switching to external sort with %d tapes: %s",
2030 maxTapes, pg_rusage_show(&state->ru_start));
2034 * Decrease availMem to reflect the space needed for tape buffers; but
2035 * don't decrease it to the point that we have no room for tuples. (That
2036 * case is only likely to occur if sorting pass-by-value Datums; in all
2037 * other scenarios the memtuples[] array is unlikely to occupy more than
2038 * half of allowedMem. In the pass-by-value case it's not important to
2039 * account for tuple space, so we don't care if LACKMEM becomes
2042 tapeSpace = (int64) maxTapes *TAPE_BUFFER_OVERHEAD;
2044 if (tapeSpace + GetMemoryChunkSpace(state->memtuples) < state->allowedMem)
2045 USEMEM(state, tapeSpace);
2048 * Make sure that the temp file(s) underlying the tape set are created in
2049 * suitable temp tablespaces.
2051 PrepareTempTablespaces();
2054 * Create the tape set and allocate the per-tape data arrays.
2056 state->tapeset = LogicalTapeSetCreate(maxTapes);
2058 state->mergeactive = (bool *) palloc0(maxTapes * sizeof(bool));
2059 state->mergenext = (int *) palloc0(maxTapes * sizeof(int));
2060 state->mergelast = (int *) palloc0(maxTapes * sizeof(int));
2061 state->mergeavailslots = (int *) palloc0(maxTapes * sizeof(int));
2062 state->mergeavailmem = (int64 *) palloc0(maxTapes * sizeof(int64));
2063 state->tp_fib = (int *) palloc0(maxTapes * sizeof(int));
2064 state->tp_runs = (int *) palloc0(maxTapes * sizeof(int));
2065 state->tp_dummy = (int *) palloc0(maxTapes * sizeof(int));
2066 state->tp_tapenum = (int *) palloc0(maxTapes * sizeof(int));
2069 * Convert the unsorted contents of memtuples[] into a heap. Each tuple is
2070 * marked as belonging to run number zero.
2072 * NOTE: we pass false for checkIndex since there's no point in comparing
2073 * indexes in this step, even though we do intend the indexes to be part
2074 * of the sort key...
2076 ntuples = state->memtupcount;
2077 state->memtupcount = 0; /* make the heap empty */
2078 for (j = 0; j < ntuples; j++)
2080 /* Must copy source tuple to avoid possible overwrite */
2081 SortTuple stup = state->memtuples[j];
2083 tuplesort_heap_insert(state, &stup, 0, false);
2085 Assert(state->memtupcount == ntuples);
2087 state->currentRun = 0;
2090 * Initialize variables of Algorithm D (step D1).
2092 for (j = 0; j < maxTapes; j++)
2094 state->tp_fib[j] = 1;
2095 state->tp_runs[j] = 0;
2096 state->tp_dummy[j] = 1;
2097 state->tp_tapenum[j] = j;
2099 state->tp_fib[state->tapeRange] = 0;
2100 state->tp_dummy[state->tapeRange] = 0;
2103 state->destTape = 0;
2105 state->status = TSS_BUILDRUNS;
2109 * selectnewtape -- select new tape for new initial run.
2111 * This is called after finishing a run when we know another run
2112 * must be started. This implements steps D3, D4 of Algorithm D.
2115 selectnewtape(Tuplesortstate *state)
2120 /* Step D3: advance j (destTape) */
2121 if (state->tp_dummy[state->destTape] < state->tp_dummy[state->destTape + 1])
2126 if (state->tp_dummy[state->destTape] != 0)
2128 state->destTape = 0;
2132 /* Step D4: increase level */
2134 a = state->tp_fib[0];
2135 for (j = 0; j < state->tapeRange; j++)
2137 state->tp_dummy[j] = a + state->tp_fib[j + 1] - state->tp_fib[j];
2138 state->tp_fib[j] = a + state->tp_fib[j + 1];
2140 state->destTape = 0;
2144 * mergeruns -- merge all the completed initial runs.
2146 * This implements steps D5, D6 of Algorithm D. All input data has
2147 * already been written to initial runs on tape (see dumptuples).
2150 mergeruns(Tuplesortstate *state)
2157 Assert(state->status == TSS_BUILDRUNS);
2158 Assert(state->memtupcount == 0);
2161 * If we produced only one initial run (quite likely if the total data
2162 * volume is between 1X and 2X workMem), we can just use that tape as the
2163 * finished output, rather than doing a useless merge. (This obvious
2164 * optimization is not in Knuth's algorithm.)
2166 if (state->currentRun == 1)
2168 state->result_tape = state->tp_tapenum[state->destTape];
2169 /* must freeze and rewind the finished output tape */
2170 LogicalTapeFreeze(state->tapeset, state->result_tape);
2171 state->status = TSS_SORTEDONTAPE;
2175 if (state->sortKeys != NULL && state->sortKeys->abbrev_converter != NULL)
2178 * If there are multiple runs to be merged, when we go to read back
2179 * tuples from disk, abbreviated keys will not have been stored, and we
2180 * don't care to regenerate them. Disable abbreviation from this point
2183 state->sortKeys->abbrev_converter = NULL;
2184 state->sortKeys->comparator = state->sortKeys->abbrev_full_comparator;
2186 /* Not strictly necessary, but be tidy */
2187 state->sortKeys->abbrev_abort = NULL;
2188 state->sortKeys->abbrev_full_comparator = NULL;
2191 /* End of step D2: rewind all output tapes to prepare for merging */
2192 for (tapenum = 0; tapenum < state->tapeRange; tapenum++)
2193 LogicalTapeRewind(state->tapeset, tapenum, false);
2198 * At this point we know that tape[T] is empty. If there's just one
2199 * (real or dummy) run left on each input tape, then only one merge
2200 * pass remains. If we don't have to produce a materialized sorted
2201 * tape, we can stop at this point and do the final merge on-the-fly.
2203 if (!state->randomAccess)
2205 bool allOneRun = true;
2207 Assert(state->tp_runs[state->tapeRange] == 0);
2208 for (tapenum = 0; tapenum < state->tapeRange; tapenum++)
2210 if (state->tp_runs[tapenum] + state->tp_dummy[tapenum] != 1)
2218 /* Tell logtape.c we won't be writing anymore */
2219 LogicalTapeSetForgetFreeSpace(state->tapeset);
2220 /* Initialize for the final merge pass */
2222 state->status = TSS_FINALMERGE;
2227 /* Step D5: merge runs onto tape[T] until tape[P] is empty */
2228 while (state->tp_runs[state->tapeRange - 1] ||
2229 state->tp_dummy[state->tapeRange - 1])
2231 bool allDummy = true;
2233 for (tapenum = 0; tapenum < state->tapeRange; tapenum++)
2235 if (state->tp_dummy[tapenum] == 0)
2244 state->tp_dummy[state->tapeRange]++;
2245 for (tapenum = 0; tapenum < state->tapeRange; tapenum++)
2246 state->tp_dummy[tapenum]--;
2252 /* Step D6: decrease level */
2253 if (--state->Level == 0)
2255 /* rewind output tape T to use as new input */
2256 LogicalTapeRewind(state->tapeset, state->tp_tapenum[state->tapeRange],
2258 /* rewind used-up input tape P, and prepare it for write pass */
2259 LogicalTapeRewind(state->tapeset, state->tp_tapenum[state->tapeRange - 1],
2261 state->tp_runs[state->tapeRange - 1] = 0;
2264 * reassign tape units per step D6; note we no longer care about A[]
2266 svTape = state->tp_tapenum[state->tapeRange];
2267 svDummy = state->tp_dummy[state->tapeRange];
2268 svRuns = state->tp_runs[state->tapeRange];
2269 for (tapenum = state->tapeRange; tapenum > 0; tapenum--)
2271 state->tp_tapenum[tapenum] = state->tp_tapenum[tapenum - 1];
2272 state->tp_dummy[tapenum] = state->tp_dummy[tapenum - 1];
2273 state->tp_runs[tapenum] = state->tp_runs[tapenum - 1];
2275 state->tp_tapenum[0] = svTape;
2276 state->tp_dummy[0] = svDummy;
2277 state->tp_runs[0] = svRuns;
2281 * Done. Knuth says that the result is on TAPE[1], but since we exited
2282 * the loop without performing the last iteration of step D6, we have not
2283 * rearranged the tape unit assignment, and therefore the result is on
2284 * TAPE[T]. We need to do it this way so that we can freeze the final
2285 * output tape while rewinding it. The last iteration of step D6 would be
2286 * a waste of cycles anyway...
2288 state->result_tape = state->tp_tapenum[state->tapeRange];
2289 LogicalTapeFreeze(state->tapeset, state->result_tape);
2290 state->status = TSS_SORTEDONTAPE;
2294 * Merge one run from each input tape, except ones with dummy runs.
2296 * This is the inner loop of Algorithm D step D5. We know that the
2297 * output tape is TAPE[T].
2300 mergeonerun(Tuplesortstate *state)
2302 int destTape = state->tp_tapenum[state->tapeRange];
2310 * Start the merge by loading one tuple from each active source tape into
2311 * the heap. We can also decrease the input run/dummy run counts.
2316 * Execute merge by repeatedly extracting lowest tuple in heap, writing it
2317 * out, and replacing it with next tuple from same tape (if there is
2320 while (state->memtupcount > 0)
2322 /* write the tuple to destTape */
2323 priorAvail = state->availMem;
2324 srcTape = state->memtuples[0].tupindex;
2325 WRITETUP(state, destTape, &state->memtuples[0]);
2326 /* writetup adjusted total free space, now fix per-tape space */
2327 spaceFreed = state->availMem - priorAvail;
2328 state->mergeavailmem[srcTape] += spaceFreed;
2329 /* compact the heap */
2330 tuplesort_heap_siftup(state, false);
2331 if ((tupIndex = state->mergenext[srcTape]) == 0)
2333 /* out of preloaded data on this tape, try to read more */
2334 mergepreread(state);
2335 /* if still no data, we've reached end of run on this tape */
2336 if ((tupIndex = state->mergenext[srcTape]) == 0)
2339 /* pull next preread tuple from list, insert in heap */
2340 tup = &state->memtuples[tupIndex];
2341 state->mergenext[srcTape] = tup->tupindex;
2342 if (state->mergenext[srcTape] == 0)
2343 state->mergelast[srcTape] = 0;
2344 tuplesort_heap_insert(state, tup, srcTape, false);
2345 /* put the now-unused memtuples entry on the freelist */
2346 tup->tupindex = state->mergefreelist;
2347 state->mergefreelist = tupIndex;
2348 state->mergeavailslots[srcTape]++;
2352 * When the heap empties, we're done. Write an end-of-run marker on the
2353 * output tape, and increment its count of real runs.
2355 markrunend(state, destTape);
2356 state->tp_runs[state->tapeRange]++;
2360 elog(LOG, "finished %d-way merge step: %s", state->activeTapes,
2361 pg_rusage_show(&state->ru_start));
2366 * beginmerge - initialize for a merge pass
2368 * We decrease the counts of real and dummy runs for each tape, and mark
2369 * which tapes contain active input runs in mergeactive[]. Then, load
2370 * as many tuples as we can from each active input tape, and finally
2371 * fill the merge heap with the first tuple from each active tape.
2374 beginmerge(Tuplesortstate *state)
2382 /* Heap should be empty here */
2383 Assert(state->memtupcount == 0);
2385 /* Adjust run counts and mark the active tapes */
2386 memset(state->mergeactive, 0,
2387 state->maxTapes * sizeof(*state->mergeactive));
2389 for (tapenum = 0; tapenum < state->tapeRange; tapenum++)
2391 if (state->tp_dummy[tapenum] > 0)
2392 state->tp_dummy[tapenum]--;
2395 Assert(state->tp_runs[tapenum] > 0);
2396 state->tp_runs[tapenum]--;
2397 srcTape = state->tp_tapenum[tapenum];
2398 state->mergeactive[srcTape] = true;
2402 state->activeTapes = activeTapes;
2404 /* Clear merge-pass state variables */
2405 memset(state->mergenext, 0,
2406 state->maxTapes * sizeof(*state->mergenext));
2407 memset(state->mergelast, 0,
2408 state->maxTapes * sizeof(*state->mergelast));
2409 state->mergefreelist = 0; /* nothing in the freelist */
2410 state->mergefirstfree = activeTapes; /* 1st slot avail for preread */
2413 * Initialize space allocation to let each active input tape have an equal
2414 * share of preread space.
2416 Assert(activeTapes > 0);
2417 slotsPerTape = (state->memtupsize - state->mergefirstfree) / activeTapes;
2418 Assert(slotsPerTape > 0);
2419 spacePerTape = state->availMem / activeTapes;
2420 for (srcTape = 0; srcTape < state->maxTapes; srcTape++)
2422 if (state->mergeactive[srcTape])
2424 state->mergeavailslots[srcTape] = slotsPerTape;
2425 state->mergeavailmem[srcTape] = spacePerTape;
2430 * Preread as many tuples as possible (and at least one) from each active
2433 mergepreread(state);
2435 /* Load the merge heap with the first tuple from each input tape */
2436 for (srcTape = 0; srcTape < state->maxTapes; srcTape++)
2438 int tupIndex = state->mergenext[srcTape];
2443 tup = &state->memtuples[tupIndex];
2444 state->mergenext[srcTape] = tup->tupindex;
2445 if (state->mergenext[srcTape] == 0)
2446 state->mergelast[srcTape] = 0;
2447 tuplesort_heap_insert(state, tup, srcTape, false);
2448 /* put the now-unused memtuples entry on the freelist */
2449 tup->tupindex = state->mergefreelist;
2450 state->mergefreelist = tupIndex;
2451 state->mergeavailslots[srcTape]++;
2457 * mergepreread - load tuples from merge input tapes
2459 * This routine exists to improve sequentiality of reads during a merge pass,
2460 * as explained in the header comments of this file. Load tuples from each
2461 * active source tape until the tape's run is exhausted or it has used up
2462 * its fair share of available memory. In any case, we guarantee that there
2463 * is at least one preread tuple available from each unexhausted input tape.
2465 * We invoke this routine at the start of a merge pass for initial load,
2466 * and then whenever any tape's preread data runs out. Note that we load
2467 * as much data as possible from all tapes, not just the one that ran out.
2468 * This is because logtape.c works best with a usage pattern that alternates
2469 * between reading a lot of data and writing a lot of data, so whenever we
2470 * are forced to read, we should fill working memory completely.
2472 * In FINALMERGE state, we *don't* use this routine, but instead just preread
2473 * from the single tape that ran dry. There's no read/write alternation in
2474 * that state and so no point in scanning through all the tapes to fix one.
2475 * (Moreover, there may be quite a lot of inactive tapes in that state, since
2476 * we might have had many fewer runs than tapes. In a regular tape-to-tape
2477 * merge we can expect most of the tapes to be active.)
2480 mergepreread(Tuplesortstate *state)
2484 for (srcTape = 0; srcTape < state->maxTapes; srcTape++)
2485 mergeprereadone(state, srcTape);
2489 * mergeprereadone - load tuples from one merge input tape
2491 * Read tuples from the specified tape until it has used up its free memory
2492 * or array slots; but ensure that we have at least one tuple, if any are
2496 mergeprereadone(Tuplesortstate *state, int srcTape)
2498 unsigned int tuplen;
2504 if (!state->mergeactive[srcTape])
2505 return; /* tape's run is already exhausted */
2506 priorAvail = state->availMem;
2507 state->availMem = state->mergeavailmem[srcTape];
2508 while ((state->mergeavailslots[srcTape] > 0 && !LACKMEM(state)) ||
2509 state->mergenext[srcTape] == 0)
2511 /* read next tuple, if any */
2512 if ((tuplen = getlen(state, srcTape, true)) == 0)
2514 state->mergeactive[srcTape] = false;
2517 READTUP(state, &stup, srcTape, tuplen);
2518 /* find a free slot in memtuples[] for it */
2519 tupIndex = state->mergefreelist;
2521 state->mergefreelist = state->memtuples[tupIndex].tupindex;
2524 tupIndex = state->mergefirstfree++;
2525 Assert(tupIndex < state->memtupsize);
2527 state->mergeavailslots[srcTape]--;
2528 /* store tuple, append to list for its tape */
2530 state->memtuples[tupIndex] = stup;
2531 if (state->mergelast[srcTape])
2532 state->memtuples[state->mergelast[srcTape]].tupindex = tupIndex;
2534 state->mergenext[srcTape] = tupIndex;
2535 state->mergelast[srcTape] = tupIndex;
2537 /* update per-tape and global availmem counts */
2538 spaceUsed = state->mergeavailmem[srcTape] - state->availMem;
2539 state->mergeavailmem[srcTape] = state->availMem;
2540 state->availMem = priorAvail - spaceUsed;
2544 * dumptuples - remove tuples from heap and write to tape
2546 * This is used during initial-run building, but not during merging.
2548 * When alltuples = false, dump only enough tuples to get under the
2549 * availMem limit (and leave at least one tuple in the heap in any case,
2550 * since puttuple assumes it always has a tuple to compare to). We also
2551 * insist there be at least one free slot in the memtuples[] array.
2553 * When alltuples = true, dump everything currently in memory.
2554 * (This case is only used at end of input data.)
2556 * If we empty the heap, close out the current run and return (this should
2557 * only happen at end of input data). If we see that the tuple run number
2558 * at the top of the heap has changed, start a new run.
2561 dumptuples(Tuplesortstate *state, bool alltuples)
2564 (LACKMEM(state) && state->memtupcount > 1) ||
2565 state->memtupcount >= state->memtupsize)
2568 * Dump the heap's frontmost entry, and sift up to remove it from the
2571 Assert(state->memtupcount > 0);
2572 WRITETUP(state, state->tp_tapenum[state->destTape],
2573 &state->memtuples[0]);
2574 tuplesort_heap_siftup(state, true);
2577 * If the heap is empty *or* top run number has changed, we've
2578 * finished the current run.
2580 if (state->memtupcount == 0 ||
2581 state->currentRun != state->memtuples[0].tupindex)
2583 markrunend(state, state->tp_tapenum[state->destTape]);
2584 state->currentRun++;
2585 state->tp_runs[state->destTape]++;
2586 state->tp_dummy[state->destTape]--; /* per Alg D step D2 */
2590 elog(LOG, "finished writing%s run %d to tape %d: %s",
2591 (state->memtupcount == 0) ? " final" : "",
2592 state->currentRun, state->destTape,
2593 pg_rusage_show(&state->ru_start));
2597 * Done if heap is empty, else prepare for new run.
2599 if (state->memtupcount == 0)
2601 Assert(state->currentRun == state->memtuples[0].tupindex);
2602 selectnewtape(state);
2608 * tuplesort_rescan - rewind and replay the scan
2611 tuplesort_rescan(Tuplesortstate *state)
2613 MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
2615 Assert(state->randomAccess);
2617 switch (state->status)
2619 case TSS_SORTEDINMEM:
2621 state->eof_reached = false;
2622 state->markpos_offset = 0;
2623 state->markpos_eof = false;
2625 case TSS_SORTEDONTAPE:
2626 LogicalTapeRewind(state->tapeset,
2629 state->eof_reached = false;
2630 state->markpos_block = 0L;
2631 state->markpos_offset = 0;
2632 state->markpos_eof = false;
2635 elog(ERROR, "invalid tuplesort state");
2639 MemoryContextSwitchTo(oldcontext);
2643 * tuplesort_markpos - saves current position in the merged sort file
2646 tuplesort_markpos(Tuplesortstate *state)
2648 MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
2650 Assert(state->randomAccess);
2652 switch (state->status)
2654 case TSS_SORTEDINMEM:
2655 state->markpos_offset = state->current;
2656 state->markpos_eof = state->eof_reached;
2658 case TSS_SORTEDONTAPE:
2659 LogicalTapeTell(state->tapeset,
2661 &state->markpos_block,
2662 &state->markpos_offset);
2663 state->markpos_eof = state->eof_reached;
2666 elog(ERROR, "invalid tuplesort state");
2670 MemoryContextSwitchTo(oldcontext);
2674 * tuplesort_restorepos - restores current position in merged sort file to
2675 * last saved position
2678 tuplesort_restorepos(Tuplesortstate *state)
2680 MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
2682 Assert(state->randomAccess);
2684 switch (state->status)
2686 case TSS_SORTEDINMEM:
2687 state->current = state->markpos_offset;
2688 state->eof_reached = state->markpos_eof;
2690 case TSS_SORTEDONTAPE:
2691 if (!LogicalTapeSeek(state->tapeset,
2693 state->markpos_block,
2694 state->markpos_offset))
2695 elog(ERROR, "tuplesort_restorepos failed");
2696 state->eof_reached = state->markpos_eof;
2699 elog(ERROR, "invalid tuplesort state");
2703 MemoryContextSwitchTo(oldcontext);
2707 * tuplesort_get_stats - extract summary statistics
2709 * This can be called after tuplesort_performsort() finishes to obtain
2710 * printable summary information about how the sort was performed.
2711 * spaceUsed is measured in kilobytes.
2714 tuplesort_get_stats(Tuplesortstate *state,
2715 const char **sortMethod,
2716 const char **spaceType,
2720 * Note: it might seem we should provide both memory and disk usage for a
2721 * disk-based sort. However, the current code doesn't track memory space
2722 * accurately once we have begun to return tuples to the caller (since we
2723 * don't account for pfree's the caller is expected to do), so we cannot
2724 * rely on availMem in a disk sort. This does not seem worth the overhead
2725 * to fix. Is it worth creating an API for the memory context code to
2726 * tell us how much is actually used in sortcontext?
2730 *spaceType = "Disk";
2731 *spaceUsed = LogicalTapeSetBlocks(state->tapeset) * (BLCKSZ / 1024);
2735 *spaceType = "Memory";
2736 *spaceUsed = (state->allowedMem - state->availMem + 1023) / 1024;
2739 switch (state->status)
2741 case TSS_SORTEDINMEM:
2742 if (state->boundUsed)
2743 *sortMethod = "top-N heapsort";
2745 *sortMethod = "quicksort";
2747 case TSS_SORTEDONTAPE:
2748 *sortMethod = "external sort";
2750 case TSS_FINALMERGE:
2751 *sortMethod = "external merge";
2754 *sortMethod = "still in progress";
2761 * Heap manipulation routines, per Knuth's Algorithm 5.2.3H.
2763 * Compare two SortTuples. If checkIndex is true, use the tuple index
2764 * as the front of the sort key; otherwise, no.
2767 #define HEAPCOMPARE(tup1,tup2) \
2768 (checkIndex && ((tup1)->tupindex != (tup2)->tupindex) ? \
2769 ((tup1)->tupindex) - ((tup2)->tupindex) : \
2770 COMPARETUP(state, tup1, tup2))
2773 * Convert the existing unordered array of SortTuples to a bounded heap,
2774 * discarding all but the smallest "state->bound" tuples.
2776 * When working with a bounded heap, we want to keep the largest entry
2777 * at the root (array entry zero), instead of the smallest as in the normal
2778 * sort case. This allows us to discard the largest entry cheaply.
2779 * Therefore, we temporarily reverse the sort direction.
2781 * We assume that all entries in a bounded heap will always have tupindex
2782 * zero; it therefore doesn't matter that HEAPCOMPARE() doesn't reverse
2783 * the direction of comparison for tupindexes.
2786 make_bounded_heap(Tuplesortstate *state)
2788 int tupcount = state->memtupcount;
2791 Assert(state->status == TSS_INITIAL);
2792 Assert(state->bounded);
2793 Assert(tupcount >= state->bound);
2795 /* Reverse sort direction so largest entry will be at root */
2796 reversedirection(state);
2798 state->memtupcount = 0; /* make the heap empty */
2799 for (i = 0; i < tupcount; i++)
2801 if (state->memtupcount >= state->bound &&
2802 COMPARETUP(state, &state->memtuples[i], &state->memtuples[0]) <= 0)
2804 /* New tuple would just get thrown out, so skip it */
2805 free_sort_tuple(state, &state->memtuples[i]);
2806 CHECK_FOR_INTERRUPTS();
2810 /* Insert next tuple into heap */
2811 /* Must copy source tuple to avoid possible overwrite */
2812 SortTuple stup = state->memtuples[i];
2814 tuplesort_heap_insert(state, &stup, 0, false);
2816 /* If heap too full, discard largest entry */
2817 if (state->memtupcount > state->bound)
2819 free_sort_tuple(state, &state->memtuples[0]);
2820 tuplesort_heap_siftup(state, false);
2825 Assert(state->memtupcount == state->bound);
2826 state->status = TSS_BOUNDED;
2830 * Convert the bounded heap to a properly-sorted array
2833 sort_bounded_heap(Tuplesortstate *state)
2835 int tupcount = state->memtupcount;
2837 Assert(state->status == TSS_BOUNDED);
2838 Assert(state->bounded);
2839 Assert(tupcount == state->bound);
2842 * We can unheapify in place because each sift-up will remove the largest
2843 * entry, which we can promptly store in the newly freed slot at the end.
2844 * Once we're down to a single-entry heap, we're done.
2846 while (state->memtupcount > 1)
2848 SortTuple stup = state->memtuples[0];
2850 /* this sifts-up the next-largest entry and decreases memtupcount */
2851 tuplesort_heap_siftup(state, false);
2852 state->memtuples[state->memtupcount] = stup;
2854 state->memtupcount = tupcount;
2857 * Reverse sort direction back to the original state. This is not
2858 * actually necessary but seems like a good idea for tidiness.
2860 reversedirection(state);
2862 state->status = TSS_SORTEDINMEM;
2863 state->boundUsed = true;
2867 * Insert a new tuple into an empty or existing heap, maintaining the
2868 * heap invariant. Caller is responsible for ensuring there's room.
2870 * Note: we assume *tuple is a temporary variable that can be scribbled on.
2871 * For some callers, tuple actually points to a memtuples[] entry above the
2872 * end of the heap. This is safe as long as it's not immediately adjacent
2873 * to the end of the heap (ie, in the [memtupcount] array entry) --- if it
2874 * is, it might get overwritten before being moved into the heap!
2877 tuplesort_heap_insert(Tuplesortstate *state, SortTuple *tuple,
2878 int tupleindex, bool checkIndex)
2880 SortTuple *memtuples;
2884 * Save the tupleindex --- see notes above about writing on *tuple. It's a
2885 * historical artifact that tupleindex is passed as a separate argument
2886 * and not in *tuple, but it's notationally convenient so let's leave it
2889 tuple->tupindex = tupleindex;
2891 memtuples = state->memtuples;
2892 Assert(state->memtupcount < state->memtupsize);
2894 CHECK_FOR_INTERRUPTS();
2897 * Sift-up the new entry, per Knuth 5.2.3 exercise 16. Note that Knuth is
2898 * using 1-based array indexes, not 0-based.
2900 j = state->memtupcount++;
2903 int i = (j - 1) >> 1;
2905 if (HEAPCOMPARE(tuple, &memtuples[i]) >= 0)
2907 memtuples[j] = memtuples[i];
2910 memtuples[j] = *tuple;
2914 * The tuple at state->memtuples[0] has been removed from the heap.
2915 * Decrement memtupcount, and sift up to maintain the heap invariant.
2918 tuplesort_heap_siftup(Tuplesortstate *state, bool checkIndex)
2920 SortTuple *memtuples = state->memtuples;
2925 if (--state->memtupcount <= 0)
2928 CHECK_FOR_INTERRUPTS();
2930 n = state->memtupcount;
2931 tuple = &memtuples[n]; /* tuple that must be reinserted */
2932 i = 0; /* i is where the "hole" is */
2940 HEAPCOMPARE(&memtuples[j], &memtuples[j + 1]) > 0)
2942 if (HEAPCOMPARE(tuple, &memtuples[j]) <= 0)
2944 memtuples[i] = memtuples[j];
2947 memtuples[i] = *tuple;
2951 * Function to reverse the sort direction from its current state
2953 * It is not safe to call this when performing hash tuplesorts
2956 reversedirection(Tuplesortstate *state)
2958 SortSupport sortKey = state->sortKeys;
2961 for (nkey = 0; nkey < state->nKeys; nkey++, sortKey++)
2963 sortKey->ssup_reverse = !sortKey->ssup_reverse;
2964 sortKey->ssup_nulls_first = !sortKey->ssup_nulls_first;
2970 * Tape interface routines
2974 getlen(Tuplesortstate *state, int tapenum, bool eofOK)
2978 if (LogicalTapeRead(state->tapeset, tapenum,
2979 &len, sizeof(len)) != sizeof(len))
2980 elog(ERROR, "unexpected end of tape");
2981 if (len == 0 && !eofOK)
2982 elog(ERROR, "unexpected end of data");
2987 markrunend(Tuplesortstate *state, int tapenum)
2989 unsigned int len = 0;
2991 LogicalTapeWrite(state->tapeset, tapenum, (void *) &len, sizeof(len));
2996 * Routines specialized for HeapTuple (actually MinimalTuple) case
3000 comparetup_heap(const SortTuple *a, const SortTuple *b, Tuplesortstate *state)
3002 SortSupport sortKey = state->sortKeys;
3015 /* Compare the leading sort key */
3016 compare = ApplySortComparator(a->datum1, a->isnull1,
3017 b->datum1, b->isnull1,
3022 /* Compare additional sort keys */
3023 ltup.t_len = ((MinimalTuple) a->tuple)->t_len + MINIMAL_TUPLE_OFFSET;
3024 ltup.t_data = (HeapTupleHeader) ((char *) a->tuple - MINIMAL_TUPLE_OFFSET);
3025 rtup.t_len = ((MinimalTuple) b->tuple)->t_len + MINIMAL_TUPLE_OFFSET;
3026 rtup.t_data = (HeapTupleHeader) ((char *) b->tuple - MINIMAL_TUPLE_OFFSET);
3027 tupDesc = state->tupDesc;
3029 if (sortKey->abbrev_converter)
3031 attno = sortKey->ssup_attno;
3033 datum1 = heap_getattr(<up, attno, tupDesc, &isnull1);
3034 datum2 = heap_getattr(&rtup, attno, tupDesc, &isnull2);
3036 compare = ApplySortAbbrevFullComparator(datum1, isnull1,
3044 for (nkey = 1; nkey < state->nKeys; nkey++, sortKey++)
3046 attno = sortKey->ssup_attno;
3048 datum1 = heap_getattr(<up, attno, tupDesc, &isnull1);
3049 datum2 = heap_getattr(&rtup, attno, tupDesc, &isnull2);
3051 compare = ApplySortComparator(datum1, isnull1,
3062 copytup_heap(Tuplesortstate *state, SortTuple *stup, void *tup)
3065 * We expect the passed "tup" to be a TupleTableSlot, and form a
3066 * MinimalTuple using the exported interface for that.
3068 TupleTableSlot *slot = (TupleTableSlot *) tup;
3073 /* copy the tuple into sort storage */
3074 tuple = ExecCopySlotMinimalTuple(slot);
3075 stup->tuple = (void *) tuple;
3076 USEMEM(state, GetMemoryChunkSpace(tuple));
3077 /* set up first-column key value */
3078 htup.t_len = tuple->t_len + MINIMAL_TUPLE_OFFSET;
3079 htup.t_data = (HeapTupleHeader) ((char *) tuple - MINIMAL_TUPLE_OFFSET);
3080 original = heap_getattr(&htup,
3081 state->sortKeys[0].ssup_attno,
3085 if (!state->sortKeys->abbrev_converter || stup->isnull1)
3088 * Store ordinary Datum representation, or NULL value. If there is a
3089 * converter it won't expect NULL values, and cost model is not
3090 * required to account for NULL, so in that case we avoid calling
3091 * converter and just set datum1 to "void" representation (to be
3094 stup->datum1 = original;
3096 else if (!consider_abort_common(state))
3098 /* Store abbreviated key representation */
3099 stup->datum1 = state->sortKeys->abbrev_converter(original,
3104 /* Abort abbreviation */
3107 stup->datum1 = original;
3110 * Set state to be consistent with never trying abbreviation.
3112 * Alter datum1 representation in already-copied tuples, so as to
3113 * ensure a consistent representation (current tuple was just handled).
3114 * Note that we rely on all tuples copied so far actually being
3115 * contained within memtuples array.
3117 for (i = 0; i < state->memtupcount; i++)
3119 SortTuple *mtup = &state->memtuples[i];
3121 htup.t_len = ((MinimalTuple) mtup->tuple)->t_len +
3122 MINIMAL_TUPLE_OFFSET;
3123 htup.t_data = (HeapTupleHeader) ((char *) mtup->tuple -
3124 MINIMAL_TUPLE_OFFSET);
3126 mtup->datum1 = heap_getattr(&htup,
3127 state->sortKeys[0].ssup_attno,
3135 writetup_heap(Tuplesortstate *state, int tapenum, SortTuple *stup)
3137 MinimalTuple tuple = (MinimalTuple) stup->tuple;
3139 /* the part of the MinimalTuple we'll write: */
3140 char *tupbody = (char *) tuple + MINIMAL_TUPLE_DATA_OFFSET;
3141 unsigned int tupbodylen = tuple->t_len - MINIMAL_TUPLE_DATA_OFFSET;
3143 /* total on-disk footprint: */
3144 unsigned int tuplen = tupbodylen + sizeof(int);
3146 LogicalTapeWrite(state->tapeset, tapenum,
3147 (void *) &tuplen, sizeof(tuplen));
3148 LogicalTapeWrite(state->tapeset, tapenum,
3149 (void *) tupbody, tupbodylen);
3150 if (state->randomAccess) /* need trailing length word? */
3151 LogicalTapeWrite(state->tapeset, tapenum,
3152 (void *) &tuplen, sizeof(tuplen));
3154 FREEMEM(state, GetMemoryChunkSpace(tuple));
3155 heap_free_minimal_tuple(tuple);
3159 readtup_heap(Tuplesortstate *state, SortTuple *stup,
3160 int tapenum, unsigned int len)
3162 unsigned int tupbodylen = len - sizeof(int);
3163 unsigned int tuplen = tupbodylen + MINIMAL_TUPLE_DATA_OFFSET;
3164 MinimalTuple tuple = (MinimalTuple) palloc(tuplen);
3165 char *tupbody = (char *) tuple + MINIMAL_TUPLE_DATA_OFFSET;
3168 USEMEM(state, GetMemoryChunkSpace(tuple));
3169 /* read in the tuple proper */
3170 tuple->t_len = tuplen;
3171 LogicalTapeReadExact(state->tapeset, tapenum,
3172 tupbody, tupbodylen);
3173 if (state->randomAccess) /* need trailing length word? */
3174 LogicalTapeReadExact(state->tapeset, tapenum,
3175 &tuplen, sizeof(tuplen));
3176 stup->tuple = (void *) tuple;
3177 /* set up first-column key value */
3178 htup.t_len = tuple->t_len + MINIMAL_TUPLE_OFFSET;
3179 htup.t_data = (HeapTupleHeader) ((char *) tuple - MINIMAL_TUPLE_OFFSET);
3180 stup->datum1 = heap_getattr(&htup,
3181 state->sortKeys[0].ssup_attno,
3187 * Routines specialized for the CLUSTER case (HeapTuple data, with
3188 * comparisons per a btree index definition)
3192 comparetup_cluster(const SortTuple *a, const SortTuple *b,
3193 Tuplesortstate *state)
3195 SortSupport sortKey = state->sortKeys;
3205 AttrNumber leading = state->indexInfo->ii_KeyAttrNumbers[0];
3207 /* Be prepared to compare additional sort keys */
3208 ltup = (HeapTuple) a->tuple;
3209 rtup = (HeapTuple) b->tuple;
3210 tupDesc = state->tupDesc;
3212 /* Compare the leading sort key, if it's simple */
3215 compare = ApplySortComparator(a->datum1, a->isnull1,
3216 b->datum1, b->isnull1,
3221 if (sortKey->abbrev_converter)
3223 datum1 = heap_getattr(ltup, leading, tupDesc, &isnull1);
3224 datum2 = heap_getattr(rtup, leading, tupDesc, &isnull2);
3226 compare = ApplySortAbbrevFullComparator(datum1, isnull1,
3230 if (compare != 0 || state->nKeys == 1)
3232 /* Compare additional columns the hard way */
3238 /* Must compare all keys the hard way */
3242 if (state->indexInfo->ii_Expressions == NULL)
3244 /* If not expression index, just compare the proper heap attrs */
3246 for (; nkey < state->nKeys; nkey++, sortKey++)
3248 AttrNumber attno = state->indexInfo->ii_KeyAttrNumbers[nkey];
3250 datum1 = heap_getattr(ltup, attno, tupDesc, &isnull1);
3251 datum2 = heap_getattr(rtup, attno, tupDesc, &isnull2);
3253 compare = ApplySortComparator(datum1, isnull1,
3263 * In the expression index case, compute the whole index tuple and
3264 * then compare values. It would perhaps be faster to compute only as
3265 * many columns as we need to compare, but that would require
3266 * duplicating all the logic in FormIndexDatum.
3268 Datum l_index_values[INDEX_MAX_KEYS];
3269 bool l_index_isnull[INDEX_MAX_KEYS];
3270 Datum r_index_values[INDEX_MAX_KEYS];
3271 bool r_index_isnull[INDEX_MAX_KEYS];
3272 TupleTableSlot *ecxt_scantuple;
3274 /* Reset context each time to prevent memory leakage */
3275 ResetPerTupleExprContext(state->estate);
3277 ecxt_scantuple = GetPerTupleExprContext(state->estate)->ecxt_scantuple;
3279 ExecStoreTuple(ltup, ecxt_scantuple, InvalidBuffer, false);
3280 FormIndexDatum(state->indexInfo, ecxt_scantuple, state->estate,
3281 l_index_values, l_index_isnull);
3283 ExecStoreTuple(rtup, ecxt_scantuple, InvalidBuffer, false);
3284 FormIndexDatum(state->indexInfo, ecxt_scantuple, state->estate,
3285 r_index_values, r_index_isnull);
3287 for (; nkey < state->nKeys; nkey++, sortKey++)
3289 compare = ApplySortComparator(l_index_values[nkey],
3290 l_index_isnull[nkey],
3291 r_index_values[nkey],
3292 r_index_isnull[nkey],
3303 copytup_cluster(Tuplesortstate *state, SortTuple *stup, void *tup)
3305 HeapTuple tuple = (HeapTuple) tup;
3308 /* copy the tuple into sort storage */
3309 tuple = heap_copytuple(tuple);
3310 stup->tuple = (void *) tuple;
3311 USEMEM(state, GetMemoryChunkSpace(tuple));
3313 * set up first-column key value, and potentially abbreviate, if it's a
3316 if (state->indexInfo->ii_KeyAttrNumbers[0] == 0)
3319 original = heap_getattr(tuple,
3320 state->indexInfo->ii_KeyAttrNumbers[0],
3324 if (!state->sortKeys->abbrev_converter || stup->isnull1)
3327 * Store ordinary Datum representation, or NULL value. If there is a
3328 * converter it won't expect NULL values, and cost model is not
3329 * required to account for NULL, so in that case we avoid calling
3330 * converter and just set datum1 to "void" representation (to be
3333 stup->datum1 = original;
3335 else if (!consider_abort_common(state))
3337 /* Store abbreviated key representation */
3338 stup->datum1 = state->sortKeys->abbrev_converter(original,
3343 /* Abort abbreviation */
3346 stup->datum1 = original;
3349 * Set state to be consistent with never trying abbreviation.
3351 * Alter datum1 representation in already-copied tuples, so as to
3352 * ensure a consistent representation (current tuple was just handled).
3353 * Note that we rely on all tuples copied so far actually being
3354 * contained within memtuples array.
3356 for (i = 0; i < state->memtupcount; i++)
3358 SortTuple *mtup = &state->memtuples[i];
3360 tuple = (HeapTuple) mtup->tuple;
3361 mtup->datum1 = heap_getattr(tuple,
3362 state->indexInfo->ii_KeyAttrNumbers[0],
3370 writetup_cluster(Tuplesortstate *state, int tapenum, SortTuple *stup)
3372 HeapTuple tuple = (HeapTuple) stup->tuple;
3373 unsigned int tuplen = tuple->t_len + sizeof(ItemPointerData) + sizeof(int);
3375 /* We need to store t_self, but not other fields of HeapTupleData */
3376 LogicalTapeWrite(state->tapeset, tapenum,
3377 &tuplen, sizeof(tuplen));
3378 LogicalTapeWrite(state->tapeset, tapenum,
3379 &tuple->t_self, sizeof(ItemPointerData));
3380 LogicalTapeWrite(state->tapeset, tapenum,
3381 tuple->t_data, tuple->t_len);
3382 if (state->randomAccess) /* need trailing length word? */
3383 LogicalTapeWrite(state->tapeset, tapenum,
3384 &tuplen, sizeof(tuplen));
3386 FREEMEM(state, GetMemoryChunkSpace(tuple));
3387 heap_freetuple(tuple);
3391 readtup_cluster(Tuplesortstate *state, SortTuple *stup,
3392 int tapenum, unsigned int tuplen)
3394 unsigned int t_len = tuplen - sizeof(ItemPointerData) - sizeof(int);
3395 HeapTuple tuple = (HeapTuple) palloc(t_len + HEAPTUPLESIZE);
3397 USEMEM(state, GetMemoryChunkSpace(tuple));
3398 /* Reconstruct the HeapTupleData header */
3399 tuple->t_data = (HeapTupleHeader) ((char *) tuple + HEAPTUPLESIZE);
3400 tuple->t_len = t_len;
3401 LogicalTapeReadExact(state->tapeset, tapenum,
3402 &tuple->t_self, sizeof(ItemPointerData));
3403 /* We don't currently bother to reconstruct t_tableOid */
3404 tuple->t_tableOid = InvalidOid;
3405 /* Read in the tuple body */
3406 LogicalTapeReadExact(state->tapeset, tapenum,
3407 tuple->t_data, tuple->t_len);
3408 if (state->randomAccess) /* need trailing length word? */
3409 LogicalTapeReadExact(state->tapeset, tapenum,
3410 &tuplen, sizeof(tuplen));
3411 stup->tuple = (void *) tuple;
3412 /* set up first-column key value, if it's a simple column */
3413 if (state->indexInfo->ii_KeyAttrNumbers[0] != 0)
3414 stup->datum1 = heap_getattr(tuple,
3415 state->indexInfo->ii_KeyAttrNumbers[0],
3422 * Routines specialized for IndexTuple case
3424 * The btree and hash cases require separate comparison functions, but the
3425 * IndexTuple representation is the same so the copy/write/read support
3426 * functions can be shared.
3430 comparetup_index_btree(const SortTuple *a, const SortTuple *b,
3431 Tuplesortstate *state)
3434 * This is similar to comparetup_heap(), but expects index tuples. There
3435 * is also special handling for enforcing uniqueness, and special treatment
3436 * for equal keys at the end.
3438 SortSupport sortKey = state->sortKeys;
3443 bool equal_hasnull = false;
3452 /* Compare the leading sort key */
3453 compare = ApplySortComparator(a->datum1, a->isnull1,
3454 b->datum1, b->isnull1,
3459 /* Compare additional sort keys */
3460 tuple1 = (IndexTuple) a->tuple;
3461 tuple2 = (IndexTuple) b->tuple;
3462 keysz = state->nKeys;
3463 tupDes = RelationGetDescr(state->indexRel);
3465 if (sortKey->abbrev_converter)
3467 datum1 = index_getattr(tuple1, 1, tupDes, &isnull1);
3468 datum2 = index_getattr(tuple2, 1, tupDes, &isnull2);
3470 compare = ApplySortAbbrevFullComparator(datum1, isnull1,
3477 /* they are equal, so we only need to examine one null flag */
3479 equal_hasnull = true;
3482 for (nkey = 2; nkey <= keysz; nkey++, sortKey++)
3484 datum1 = index_getattr(tuple1, nkey, tupDes, &isnull1);
3485 datum2 = index_getattr(tuple2, nkey, tupDes, &isnull2);
3487 compare = ApplySortComparator(datum1, isnull1,
3491 return compare; /* done when we find unequal attributes */
3493 /* they are equal, so we only need to examine one null flag */
3495 equal_hasnull = true;
3499 * If btree has asked us to enforce uniqueness, complain if two equal
3500 * tuples are detected (unless there was at least one NULL field).
3502 * It is sufficient to make the test here, because if two tuples are equal
3503 * they *must* get compared at some stage of the sort --- otherwise the
3504 * sort algorithm wouldn't have checked whether one must appear before the
3507 if (state->enforceUnique && !equal_hasnull)
3509 Datum values[INDEX_MAX_KEYS];
3510 bool isnull[INDEX_MAX_KEYS];
3514 * Some rather brain-dead implementations of qsort (such as the one in
3515 * QNX 4) will sometimes call the comparison routine to compare a
3516 * value to itself, but we always use our own implementation, which
3519 Assert(tuple1 != tuple2);
3521 index_deform_tuple(tuple1, tupDes, values, isnull);
3523 key_desc = BuildIndexValueDescription(state->indexRel, values, isnull);
3526 (errcode(ERRCODE_UNIQUE_VIOLATION),
3527 errmsg("could not create unique index \"%s\"",
3528 RelationGetRelationName(state->indexRel)),
3529 key_desc ? errdetail("Key %s is duplicated.", key_desc) :
3530 errdetail("Duplicate keys exist."),
3531 errtableconstraint(state->heapRel,
3532 RelationGetRelationName(state->indexRel))));
3536 * If key values are equal, we sort on ItemPointer. This does not affect
3537 * validity of the finished index, but it may be useful to have index
3538 * scans in physical order.
3541 BlockNumber blk1 = ItemPointerGetBlockNumber(&tuple1->t_tid);
3542 BlockNumber blk2 = ItemPointerGetBlockNumber(&tuple2->t_tid);
3545 return (blk1 < blk2) ? -1 : 1;
3548 OffsetNumber pos1 = ItemPointerGetOffsetNumber(&tuple1->t_tid);
3549 OffsetNumber pos2 = ItemPointerGetOffsetNumber(&tuple2->t_tid);
3552 return (pos1 < pos2) ? -1 : 1;
3559 comparetup_index_hash(const SortTuple *a, const SortTuple *b,
3560 Tuplesortstate *state)
3568 * Fetch hash keys and mask off bits we don't want to sort by. We know
3569 * that the first column of the index tuple is the hash key.
3571 Assert(!a->isnull1);
3572 hash1 = DatumGetUInt32(a->datum1) & state->hash_mask;
3573 Assert(!b->isnull1);
3574 hash2 = DatumGetUInt32(b->datum1) & state->hash_mask;
3578 else if (hash1 < hash2)
3582 * If hash values are equal, we sort on ItemPointer. This does not affect
3583 * validity of the finished index, but it may be useful to have index
3584 * scans in physical order.
3586 tuple1 = (IndexTuple) a->tuple;
3587 tuple2 = (IndexTuple) b->tuple;
3590 BlockNumber blk1 = ItemPointerGetBlockNumber(&tuple1->t_tid);
3591 BlockNumber blk2 = ItemPointerGetBlockNumber(&tuple2->t_tid);
3594 return (blk1 < blk2) ? -1 : 1;
3597 OffsetNumber pos1 = ItemPointerGetOffsetNumber(&tuple1->t_tid);
3598 OffsetNumber pos2 = ItemPointerGetOffsetNumber(&tuple2->t_tid);
3601 return (pos1 < pos2) ? -1 : 1;
3608 copytup_index(Tuplesortstate *state, SortTuple *stup, void *tup)
3610 IndexTuple tuple = (IndexTuple) tup;
3611 unsigned int tuplen = IndexTupleSize(tuple);
3612 IndexTuple newtuple;
3615 /* copy the tuple into sort storage */
3616 newtuple = (IndexTuple) palloc(tuplen);
3617 memcpy(newtuple, tuple, tuplen);
3618 USEMEM(state, GetMemoryChunkSpace(newtuple));
3619 stup->tuple = (void *) newtuple;
3620 /* set up first-column key value */
3621 original = index_getattr(newtuple,
3623 RelationGetDescr(state->indexRel),
3626 if (!state->sortKeys->abbrev_converter || stup->isnull1)
3629 * Store ordinary Datum representation, or NULL value. If there is a
3630 * converter it won't expect NULL values, and cost model is not
3631 * required to account for NULL, so in that case we avoid calling
3632 * converter and just set datum1 to "void" representation (to be
3635 stup->datum1 = original;
3637 else if (!consider_abort_common(state))
3639 /* Store abbreviated key representation */
3640 stup->datum1 = state->sortKeys->abbrev_converter(original,
3645 /* Abort abbreviation */
3648 stup->datum1 = original;
3651 * Set state to be consistent with never trying abbreviation.
3653 * Alter datum1 representation in already-copied tuples, so as to
3654 * ensure a consistent representation (current tuple was just handled).
3655 * Note that we rely on all tuples copied so far actually being
3656 * contained within memtuples array.
3658 for (i = 0; i < state->memtupcount; i++)
3660 SortTuple *mtup = &state->memtuples[i];
3662 tuple = (IndexTuple) mtup->tuple;
3663 mtup->datum1 = index_getattr(tuple,
3665 RelationGetDescr(state->indexRel),
3672 writetup_index(Tuplesortstate *state, int tapenum, SortTuple *stup)
3674 IndexTuple tuple = (IndexTuple) stup->tuple;
3675 unsigned int tuplen;
3677 tuplen = IndexTupleSize(tuple) + sizeof(tuplen);
3678 LogicalTapeWrite(state->tapeset, tapenum,
3679 (void *) &tuplen, sizeof(tuplen));
3680 LogicalTapeWrite(state->tapeset, tapenum,
3681 (void *) tuple, IndexTupleSize(tuple));
3682 if (state->randomAccess) /* need trailing length word? */
3683 LogicalTapeWrite(state->tapeset, tapenum,
3684 (void *) &tuplen, sizeof(tuplen));
3686 FREEMEM(state, GetMemoryChunkSpace(tuple));
3691 readtup_index(Tuplesortstate *state, SortTuple *stup,
3692 int tapenum, unsigned int len)
3694 unsigned int tuplen = len - sizeof(unsigned int);
3695 IndexTuple tuple = (IndexTuple) palloc(tuplen);
3697 USEMEM(state, GetMemoryChunkSpace(tuple));
3698 LogicalTapeReadExact(state->tapeset, tapenum,
3700 if (state->randomAccess) /* need trailing length word? */
3701 LogicalTapeReadExact(state->tapeset, tapenum,
3702 &tuplen, sizeof(tuplen));
3703 stup->tuple = (void *) tuple;
3704 /* set up first-column key value */
3705 stup->datum1 = index_getattr(tuple,
3707 RelationGetDescr(state->indexRel),
3712 * Routines specialized for DatumTuple case
3716 comparetup_datum(const SortTuple *a, const SortTuple *b, Tuplesortstate *state)
3718 return ApplySortComparator(a->datum1, a->isnull1,
3719 b->datum1, b->isnull1,
3724 copytup_datum(Tuplesortstate *state, SortTuple *stup, void *tup)
3726 /* Not currently needed */
3727 elog(ERROR, "copytup_datum() should not be called");
3731 writetup_datum(Tuplesortstate *state, int tapenum, SortTuple *stup)
3734 unsigned int tuplen;
3735 unsigned int writtenlen;
3742 else if (state->datumTypeByVal)
3744 waddr = &stup->datum1;
3745 tuplen = sizeof(Datum);
3749 waddr = DatumGetPointer(stup->datum1);
3750 tuplen = datumGetSize(stup->datum1, false, state->datumTypeLen);
3751 Assert(tuplen != 0);
3754 writtenlen = tuplen + sizeof(unsigned int);
3756 LogicalTapeWrite(state->tapeset, tapenum,
3757 (void *) &writtenlen, sizeof(writtenlen));
3758 LogicalTapeWrite(state->tapeset, tapenum,
3760 if (state->randomAccess) /* need trailing length word? */
3761 LogicalTapeWrite(state->tapeset, tapenum,
3762 (void *) &writtenlen, sizeof(writtenlen));
3766 FREEMEM(state, GetMemoryChunkSpace(stup->tuple));
3772 readtup_datum(Tuplesortstate *state, SortTuple *stup,
3773 int tapenum, unsigned int len)
3775 unsigned int tuplen = len - sizeof(unsigned int);
3780 stup->datum1 = (Datum) 0;
3781 stup->isnull1 = true;
3784 else if (state->datumTypeByVal)
3786 Assert(tuplen == sizeof(Datum));
3787 LogicalTapeReadExact(state->tapeset, tapenum,
3788 &stup->datum1, tuplen);
3789 stup->isnull1 = false;
3794 void *raddr = palloc(tuplen);
3796 LogicalTapeReadExact(state->tapeset, tapenum,
3798 stup->datum1 = PointerGetDatum(raddr);
3799 stup->isnull1 = false;
3800 stup->tuple = raddr;
3801 USEMEM(state, GetMemoryChunkSpace(raddr));
3804 if (state->randomAccess) /* need trailing length word? */
3805 LogicalTapeReadExact(state->tapeset, tapenum,
3806 &tuplen, sizeof(tuplen));
3810 * Convenience routine to free a tuple previously loaded into sort memory
3813 free_sort_tuple(Tuplesortstate *state, SortTuple *stup)
3815 FREEMEM(state, GetMemoryChunkSpace(stup->tuple));