PostgreSQL Source Code  git master
tuplesort.c
Go to the documentation of this file.
1 /*-------------------------------------------------------------------------
2  *
3  * tuplesort.c
4  * Generalized tuple sorting routines.
5  *
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
11  * algorithm.
12  *
13  * See Knuth, volume 3, for more than you want to know about the external
14  * sorting algorithm. Historically, we divided the input into sorted runs
15  * using replacement selection, in the form of a priority tree implemented
16  * as a heap (essentially his Algorithm 5.2.3H), but now we always use
17  * quicksort for run generation. We merge the runs using polyphase merge,
18  * Knuth's Algorithm 5.4.2D. The logical "tapes" used by Algorithm D are
19  * implemented by logtape.c, which avoids space wastage by recycling disk
20  * space as soon as each block is read from its "tape".
21  *
22  * The approximate amount of memory allowed for any one sort operation
23  * is specified in kilobytes by the caller (most pass work_mem). Initially,
24  * we absorb tuples and simply store them in an unsorted array as long as
25  * we haven't exceeded workMem. If we reach the end of the input without
26  * exceeding workMem, we sort the array using qsort() and subsequently return
27  * tuples just by scanning the tuple array sequentially. If we do exceed
28  * workMem, we begin to emit tuples into sorted runs in temporary tapes.
29  * When tuples are dumped in batch after quicksorting, we begin a new run
30  * with a new output tape (selected per Algorithm D). After the end of the
31  * input is reached, we dump out remaining tuples in memory into a final run,
32  * then merge the runs using Algorithm D.
33  *
34  * When merging runs, we use a heap containing just the frontmost tuple from
35  * each source run; we repeatedly output the smallest tuple and replace it
36  * with the next tuple from its source tape (if any). When the heap empties,
37  * the merge is complete. The basic merge algorithm thus needs very little
38  * memory --- only M tuples for an M-way merge, and M is constrained to a
39  * small number. However, we can still make good use of our full workMem
40  * allocation by pre-reading additional blocks from each source tape. Without
41  * prereading, our access pattern to the temporary file would be very erratic;
42  * on average we'd read one block from each of M source tapes during the same
43  * time that we're writing M blocks to the output tape, so there is no
44  * sequentiality of access at all, defeating the read-ahead methods used by
45  * most Unix kernels. Worse, the output tape gets written into a very random
46  * sequence of blocks of the temp file, ensuring that things will be even
47  * worse when it comes time to read that tape. A straightforward merge pass
48  * thus ends up doing a lot of waiting for disk seeks. We can improve matters
49  * by prereading from each source tape sequentially, loading about workMem/M
50  * bytes from each tape in turn, and making the sequential blocks immediately
51  * available for reuse. This approach helps to localize both read and write
52  * accesses. The pre-reading is handled by logtape.c, we just tell it how
53  * much memory to use for the buffers.
54  *
55  * When the caller requests random access to the sort result, we form
56  * the final sorted run on a logical tape which is then "frozen", so
57  * that we can access it randomly. When the caller does not need random
58  * access, we return from tuplesort_performsort() as soon as we are down
59  * to one run per logical tape. The final merge is then performed
60  * on-the-fly as the caller repeatedly calls tuplesort_getXXX; this
61  * saves one cycle of writing all the data out to disk and reading it in.
62  *
63  * Before Postgres 8.2, we always used a seven-tape polyphase merge, on the
64  * grounds that 7 is the "sweet spot" on the tapes-to-passes curve according
65  * to Knuth's figure 70 (section 5.4.2). However, Knuth is assuming that
66  * tape drives are expensive beasts, and in particular that there will always
67  * be many more runs than tape drives. In our implementation a "tape drive"
68  * doesn't cost much more than a few Kb of memory buffers, so we can afford
69  * to have lots of them. In particular, if we can have as many tape drives
70  * as sorted runs, we can eliminate any repeated I/O at all. In the current
71  * code we determine the number of tapes M on the basis of workMem: we want
72  * workMem/M to be large enough that we read a fair amount of data each time
73  * we preread from a tape, so as to maintain the locality of access described
74  * above. Nonetheless, with large workMem we can have many tapes (but not
75  * too many -- see the comments in tuplesort_merge_order).
76  *
77  * This module supports parallel sorting. Parallel sorts involve coordination
78  * among one or more worker processes, and a leader process, each with its own
79  * tuplesort state. The leader process (or, more accurately, the
80  * Tuplesortstate associated with a leader process) creates a full tapeset
81  * consisting of worker tapes with one run to merge; a run for every
82  * worker process. This is then merged. Worker processes are guaranteed to
83  * produce exactly one output run from their partial input.
84  *
85  *
86  * Portions Copyright (c) 1996-2018, PostgreSQL Global Development Group
87  * Portions Copyright (c) 1994, Regents of the University of California
88  *
89  * IDENTIFICATION
90  * src/backend/utils/sort/tuplesort.c
91  *
92  *-------------------------------------------------------------------------
93  */
94 
95 #include "postgres.h"
96 
97 #include <limits.h>
98 
99 #include "access/htup_details.h"
100 #include "access/nbtree.h"
101 #include "access/hash.h"
102 #include "catalog/index.h"
103 #include "catalog/pg_am.h"
104 #include "commands/tablespace.h"
105 #include "executor/executor.h"
106 #include "miscadmin.h"
107 #include "pg_trace.h"
108 #include "utils/datum.h"
109 #include "utils/logtape.h"
110 #include "utils/lsyscache.h"
111 #include "utils/memutils.h"
112 #include "utils/pg_rusage.h"
113 #include "utils/rel.h"
114 #include "utils/sortsupport.h"
115 #include "utils/tuplesort.h"
116 
117 
118 /* sort-type codes for sort__start probes */
119 #define HEAP_SORT 0
120 #define INDEX_SORT 1
121 #define DATUM_SORT 2
122 #define CLUSTER_SORT 3
123 
124 /* Sort parallel code from state for sort__start probes */
125 #define PARALLEL_SORT(state) ((state)->shared == NULL ? 0 : \
126  (state)->worker >= 0 ? 1 : 2)
127 
128 /* GUC variables */
129 #ifdef TRACE_SORT
130 bool trace_sort = false;
131 #endif
132 
133 #ifdef DEBUG_BOUNDED_SORT
134 bool optimize_bounded_sort = true;
135 #endif
136 
137 
138 /*
139  * The objects we actually sort are SortTuple structs. These contain
140  * a pointer to the tuple proper (might be a MinimalTuple or IndexTuple),
141  * which is a separate palloc chunk --- we assume it is just one chunk and
142  * can be freed by a simple pfree() (except during merge, when we use a
143  * simple slab allocator). SortTuples also contain the tuple's first key
144  * column in Datum/nullflag format, and an index integer.
145  *
146  * Storing the first key column lets us save heap_getattr or index_getattr
147  * calls during tuple comparisons. We could extract and save all the key
148  * columns not just the first, but this would increase code complexity and
149  * overhead, and wouldn't actually save any comparison cycles in the common
150  * case where the first key determines the comparison result. Note that
151  * for a pass-by-reference datatype, datum1 points into the "tuple" storage.
152  *
153  * There is one special case: when the sort support infrastructure provides an
154  * "abbreviated key" representation, where the key is (typically) a pass by
155  * value proxy for a pass by reference type. In this case, the abbreviated key
156  * is stored in datum1 in place of the actual first key column.
157  *
158  * When sorting single Datums, the data value is represented directly by
159  * datum1/isnull1 for pass by value types (or null values). If the datatype is
160  * pass-by-reference and isnull1 is false, then "tuple" points to a separately
161  * palloc'd data value, otherwise "tuple" is NULL. The value of datum1 is then
162  * either the same pointer as "tuple", or is an abbreviated key value as
163  * described above. Accordingly, "tuple" is always used in preference to
164  * datum1 as the authoritative value for pass-by-reference cases.
165  *
166  * tupindex holds the input tape number that each tuple in the heap was read
167  * from during merge passes.
168  */
169 typedef struct
170 {
171  void *tuple; /* the tuple itself */
172  Datum datum1; /* value of first key column */
173  bool isnull1; /* is first key column NULL? */
174  int tupindex; /* see notes above */
175 } SortTuple;
176 
177 /*
178  * During merge, we use a pre-allocated set of fixed-size slots to hold
179  * tuples. To avoid palloc/pfree overhead.
180  *
181  * Merge doesn't require a lot of memory, so we can afford to waste some,
182  * by using gratuitously-sized slots. If a tuple is larger than 1 kB, the
183  * palloc() overhead is not significant anymore.
184  *
185  * 'nextfree' is valid when this chunk is in the free list. When in use, the
186  * slot holds a tuple.
187  */
188 #define SLAB_SLOT_SIZE 1024
189 
190 typedef union SlabSlot
191 {
194 } SlabSlot;
195 
196 /*
197  * Possible states of a Tuplesort object. These denote the states that
198  * persist between calls of Tuplesort routines.
199  */
200 typedef enum
201 {
202  TSS_INITIAL, /* Loading tuples; still within memory limit */
203  TSS_BOUNDED, /* Loading tuples into bounded-size heap */
204  TSS_BUILDRUNS, /* Loading tuples; writing to tape */
205  TSS_SORTEDINMEM, /* Sort completed entirely in memory */
206  TSS_SORTEDONTAPE, /* Sort completed, final run is on tape */
207  TSS_FINALMERGE /* Performing final merge on-the-fly */
208 } TupSortStatus;
209 
210 /*
211  * Parameters for calculation of number of tapes to use --- see inittapes()
212  * and tuplesort_merge_order().
213  *
214  * In this calculation we assume that each tape will cost us about 1 blocks
215  * worth of buffer space. This ignores the overhead of all the other data
216  * structures needed for each tape, but it's probably close enough.
217  *
218  * MERGE_BUFFER_SIZE is how much data we'd like to read from each input
219  * tape during a preread cycle (see discussion at top of file).
220  */
221 #define MINORDER 6 /* minimum merge order */
222 #define MAXORDER 500 /* maximum merge order */
223 #define TAPE_BUFFER_OVERHEAD BLCKSZ
224 #define MERGE_BUFFER_SIZE (BLCKSZ * 32)
225 
226 typedef int (*SortTupleComparator) (const SortTuple *a, const SortTuple *b,
228 
229 /*
230  * Private state of a Tuplesort operation.
231  */
233 {
234  TupSortStatus status; /* enumerated value as shown above */
235  int nKeys; /* number of columns in sort key */
236  bool randomAccess; /* did caller request random access? */
237  bool bounded; /* did caller specify a maximum number of
238  * tuples to return? */
239  bool boundUsed; /* true if we made use of a bounded heap */
240  int bound; /* if bounded, the maximum number of tuples */
241  bool tuples; /* Can SortTuple.tuple ever be set? */
242  int64 availMem; /* remaining memory available, in bytes */
243  int64 allowedMem; /* total memory allowed, in bytes */
244  int maxTapes; /* number of tapes (Knuth's T) */
245  int tapeRange; /* maxTapes-1 (Knuth's P) */
246  MemoryContext sortcontext; /* memory context holding most sort data */
247  MemoryContext tuplecontext; /* sub-context of sortcontext for tuple data */
248  LogicalTapeSet *tapeset; /* logtape.c object for tapes in a temp file */
249 
250  /*
251  * These function pointers decouple the routines that must know what kind
252  * of tuple we are sorting from the routines that don't need to know it.
253  * They are set up by the tuplesort_begin_xxx routines.
254  *
255  * Function to compare two tuples; result is per qsort() convention, ie:
256  * <0, 0, >0 according as a<b, a=b, a>b. The API must match
257  * qsort_arg_comparator.
258  */
260 
261  /*
262  * Function to copy a supplied input tuple into palloc'd space and set up
263  * its SortTuple representation (ie, set tuple/datum1/isnull1). Also,
264  * state->availMem must be decreased by the amount of space used for the
265  * tuple copy (note the SortTuple struct itself is not counted).
266  */
267  void (*copytup) (Tuplesortstate *state, SortTuple *stup, void *tup);
268 
269  /*
270  * Function to write a stored tuple onto tape. The representation of the
271  * tuple on tape need not be the same as it is in memory; requirements on
272  * the tape representation are given below. Unless the slab allocator is
273  * used, after writing the tuple, pfree() the out-of-line data (not the
274  * SortTuple struct!), and increase state->availMem by the amount of
275  * memory space thereby released.
276  */
277  void (*writetup) (Tuplesortstate *state, int tapenum,
278  SortTuple *stup);
279 
280  /*
281  * Function to read a stored tuple from tape back into memory. 'len' is
282  * the already-read length of the stored tuple. The tuple is allocated
283  * from the slab memory arena, or is palloc'd, see readtup_alloc().
284  */
285  void (*readtup) (Tuplesortstate *state, SortTuple *stup,
286  int tapenum, unsigned int len);
287 
288  /*
289  * This array holds the tuples now in sort memory. If we are in state
290  * INITIAL, the tuples are in no particular order; if we are in state
291  * SORTEDINMEM, the tuples are in final sorted order; in states BUILDRUNS
292  * and FINALMERGE, the tuples are organized in "heap" order per Algorithm
293  * H. In state SORTEDONTAPE, the array is not used.
294  */
295  SortTuple *memtuples; /* array of SortTuple structs */
296  int memtupcount; /* number of tuples currently present */
297  int memtupsize; /* allocated length of memtuples array */
298  bool growmemtuples; /* memtuples' growth still underway? */
299 
300  /*
301  * Memory for tuples is sometimes allocated using a simple slab allocator,
302  * rather than with palloc(). Currently, we switch to slab allocation
303  * when we start merging. Merging only needs to keep a small, fixed
304  * number of tuples in memory at any time, so we can avoid the
305  * palloc/pfree overhead by recycling a fixed number of fixed-size slots
306  * to hold the tuples.
307  *
308  * For the slab, we use one large allocation, divided into SLAB_SLOT_SIZE
309  * slots. The allocation is sized to have one slot per tape, plus one
310  * additional slot. We need that many slots to hold all the tuples kept
311  * in the heap during merge, plus the one we have last returned from the
312  * sort, with tuplesort_gettuple.
313  *
314  * Initially, all the slots are kept in a linked list of free slots. When
315  * a tuple is read from a tape, it is put to the next available slot, if
316  * it fits. If the tuple is larger than SLAB_SLOT_SIZE, it is palloc'd
317  * instead.
318  *
319  * When we're done processing a tuple, we return the slot back to the free
320  * list, or pfree() if it was palloc'd. We know that a tuple was
321  * allocated from the slab, if its pointer value is between
322  * slabMemoryBegin and -End.
323  *
324  * When the slab allocator is used, the USEMEM/LACKMEM mechanism of
325  * tracking memory usage is not used.
326  */
328 
329  char *slabMemoryBegin; /* beginning of slab memory arena */
330  char *slabMemoryEnd; /* end of slab memory arena */
331  SlabSlot *slabFreeHead; /* head of free list */
332 
333  /* Buffer size to use for reading input tapes, during merge. */
335 
336  /*
337  * When we return a tuple to the caller in tuplesort_gettuple_XXX, that
338  * came from a tape (that is, in TSS_SORTEDONTAPE or TSS_FINALMERGE
339  * modes), we remember the tuple in 'lastReturnedTuple', so that we can
340  * recycle the memory on next gettuple call.
341  */
343 
344  /*
345  * While building initial runs, this is the current output run number.
346  * Afterwards, it is the number of initial runs we made.
347  */
349 
350  /*
351  * Unless otherwise noted, all pointer variables below are pointers to
352  * arrays of length maxTapes, holding per-tape data.
353  */
354 
355  /*
356  * This variable is only used during merge passes. mergeactive[i] is true
357  * if we are reading an input run from (actual) tape number i and have not
358  * yet exhausted that run.
359  */
360  bool *mergeactive; /* active input run source? */
361 
362  /*
363  * Variables for Algorithm D. Note that destTape is a "logical" tape
364  * number, ie, an index into the tp_xxx[] arrays. Be careful to keep
365  * "logical" and "actual" tape numbers straight!
366  */
367  int Level; /* Knuth's l */
368  int destTape; /* current output tape (Knuth's j, less 1) */
369  int *tp_fib; /* Target Fibonacci run counts (A[]) */
370  int *tp_runs; /* # of real runs on each tape */
371  int *tp_dummy; /* # of dummy runs for each tape (D[]) */
372  int *tp_tapenum; /* Actual tape numbers (TAPE[]) */
373  int activeTapes; /* # of active input tapes in merge pass */
374 
375  /*
376  * These variables are used after completion of sorting to keep track of
377  * the next tuple to return. (In the tape case, the tape's current read
378  * position is also critical state.)
379  */
380  int result_tape; /* actual tape number of finished output */
381  int current; /* array index (only used if SORTEDINMEM) */
382  bool eof_reached; /* reached EOF (needed for cursors) */
383 
384  /* markpos_xxx holds marked position for mark and restore */
385  long markpos_block; /* tape block# (only used if SORTEDONTAPE) */
386  int markpos_offset; /* saved "current", or offset in tape block */
387  bool markpos_eof; /* saved "eof_reached" */
388 
389  /*
390  * These variables are used during parallel sorting.
391  *
392  * worker is our worker identifier. Follows the general convention that
393  * -1 value relates to a leader tuplesort, and values >= 0 worker
394  * tuplesorts. (-1 can also be a serial tuplesort.)
395  *
396  * shared is mutable shared memory state, which is used to coordinate
397  * parallel sorts.
398  *
399  * nParticipants is the number of worker Tuplesortstates known by the
400  * leader to have actually been launched, which implies that they must
401  * finish a run leader can merge. Typically includes a worker state held
402  * by the leader process itself. Set in the leader Tuplesortstate only.
403  */
404  int worker;
407 
408  /*
409  * The sortKeys variable is used by every case other than the hash index
410  * case; it is set by tuplesort_begin_xxx. tupDesc is only used by the
411  * MinimalTuple and CLUSTER routines, though.
412  */
414  SortSupport sortKeys; /* array of length nKeys */
415 
416  /*
417  * This variable is shared by the single-key MinimalTuple case and the
418  * Datum case (which both use qsort_ssup()). Otherwise it's NULL.
419  */
421 
422  /*
423  * Additional state for managing "abbreviated key" sortsupport routines
424  * (which currently may be used by all cases except the hash index case).
425  * Tracks the intervals at which the optimization's effectiveness is
426  * tested.
427  */
428  int64 abbrevNext; /* Tuple # at which to next check
429  * applicability */
430 
431  /*
432  * These variables are specific to the CLUSTER case; they are set by
433  * tuplesort_begin_cluster.
434  */
435  IndexInfo *indexInfo; /* info about index being used for reference */
436  EState *estate; /* for evaluating index expressions */
437 
438  /*
439  * These variables are specific to the IndexTuple case; they are set by
440  * tuplesort_begin_index_xxx and used only by the IndexTuple routines.
441  */
442  Relation heapRel; /* table the index is being built on */
443  Relation indexRel; /* index being built */
444 
445  /* These are specific to the index_btree subcase: */
446  bool enforceUnique; /* complain if we find duplicate tuples */
447 
448  /* These are specific to the index_hash subcase: */
449  uint32 high_mask; /* masks for sortable part of hash code */
452 
453  /*
454  * These variables are specific to the Datum case; they are set by
455  * tuplesort_begin_datum and used only by the DatumTuple routines.
456  */
458  /* we need typelen in order to know how to copy the Datums. */
460 
461  /*
462  * Resource snapshot for time of sort start.
463  */
464 #ifdef TRACE_SORT
466 #endif
467 };
468 
469 /*
470  * Private mutable state of tuplesort-parallel-operation. This is allocated
471  * in shared memory.
472  */
474 {
475  /* mutex protects all fields prior to tapes */
477 
478  /*
479  * currentWorker generates ordinal identifier numbers for parallel sort
480  * workers. These start from 0, and are always gapless.
481  *
482  * Workers increment workersFinished to indicate having finished. If this
483  * is equal to state.nParticipants within the leader, leader is ready to
484  * merge worker runs.
485  */
488 
489  /* Temporary file space */
491 
492  /* Size of tapes flexible array */
493  int nTapes;
494 
495  /*
496  * Tapes array used by workers to report back information needed by the
497  * leader to concatenate all worker tapes into one for merging
498  */
499  TapeShare tapes[FLEXIBLE_ARRAY_MEMBER];
500 };
501 
502 /*
503  * Is the given tuple allocated from the slab memory arena?
504  */
505 #define IS_SLAB_SLOT(state, tuple) \
506  ((char *) (tuple) >= (state)->slabMemoryBegin && \
507  (char *) (tuple) < (state)->slabMemoryEnd)
508 
509 /*
510  * Return the given tuple to the slab memory free list, or free it
511  * if it was palloc'd.
512  */
513 #define RELEASE_SLAB_SLOT(state, tuple) \
514  do { \
515  SlabSlot *buf = (SlabSlot *) tuple; \
516  \
517  if (IS_SLAB_SLOT((state), buf)) \
518  { \
519  buf->nextfree = (state)->slabFreeHead; \
520  (state)->slabFreeHead = buf; \
521  } else \
522  pfree(buf); \
523  } while(0)
524 
525 #define COMPARETUP(state,a,b) ((*(state)->comparetup) (a, b, state))
526 #define COPYTUP(state,stup,tup) ((*(state)->copytup) (state, stup, tup))
527 #define WRITETUP(state,tape,stup) ((*(state)->writetup) (state, tape, stup))
528 #define READTUP(state,stup,tape,len) ((*(state)->readtup) (state, stup, tape, len))
529 #define LACKMEM(state) ((state)->availMem < 0 && !(state)->slabAllocatorUsed)
530 #define USEMEM(state,amt) ((state)->availMem -= (amt))
531 #define FREEMEM(state,amt) ((state)->availMem += (amt))
532 #define SERIAL(state) ((state)->shared == NULL)
533 #define WORKER(state) ((state)->shared && (state)->worker != -1)
534 #define LEADER(state) ((state)->shared && (state)->worker == -1)
535 
536 /*
537  * NOTES about on-tape representation of tuples:
538  *
539  * We require the first "unsigned int" of a stored tuple to be the total size
540  * on-tape of the tuple, including itself (so it is never zero; an all-zero
541  * unsigned int is used to delimit runs). The remainder of the stored tuple
542  * may or may not match the in-memory representation of the tuple ---
543  * any conversion needed is the job of the writetup and readtup routines.
544  *
545  * If state->randomAccess is true, then the stored representation of the
546  * tuple must be followed by another "unsigned int" that is a copy of the
547  * length --- so the total tape space used is actually sizeof(unsigned int)
548  * more than the stored length value. This allows read-backwards. When
549  * randomAccess is not true, the write/read routines may omit the extra
550  * length word.
551  *
552  * writetup is expected to write both length words as well as the tuple
553  * data. When readtup is called, the tape is positioned just after the
554  * front length word; readtup must read the tuple data and advance past
555  * the back length word (if present).
556  *
557  * The write/read routines can make use of the tuple description data
558  * stored in the Tuplesortstate record, if needed. They are also expected
559  * to adjust state->availMem by the amount of memory space (not tape space!)
560  * released or consumed. There is no error return from either writetup
561  * or readtup; they should ereport() on failure.
562  *
563  *
564  * NOTES about memory consumption calculations:
565  *
566  * We count space allocated for tuples against the workMem limit, plus
567  * the space used by the variable-size memtuples array. Fixed-size space
568  * is not counted; it's small enough to not be interesting.
569  *
570  * Note that we count actual space used (as shown by GetMemoryChunkSpace)
571  * rather than the originally-requested size. This is important since
572  * palloc can add substantial overhead. It's not a complete answer since
573  * we won't count any wasted space in palloc allocation blocks, but it's
574  * a lot better than what we were doing before 7.3. As of 9.6, a
575  * separate memory context is used for caller passed tuples. Resetting
576  * it at certain key increments significantly ameliorates fragmentation.
577  * Note that this places a responsibility on readtup and copytup routines
578  * to use the right memory context for these tuples (and to not use the
579  * reset context for anything whose lifetime needs to span multiple
580  * external sort runs).
581  */
582 
583 /* When using this macro, beware of double evaluation of len */
584 #define LogicalTapeReadExact(tapeset, tapenum, ptr, len) \
585  do { \
586  if (LogicalTapeRead(tapeset, tapenum, ptr, len) != (size_t) (len)) \
587  elog(ERROR, "unexpected end of data"); \
588  } while(0)
589 
590 
591 static Tuplesortstate *tuplesort_begin_common(int workMem,
592  SortCoordinate coordinate,
593  bool randomAccess);
594 static void puttuple_common(Tuplesortstate *state, SortTuple *tuple);
596 static void inittapes(Tuplesortstate *state, bool mergeruns);
597 static void inittapestate(Tuplesortstate *state, int maxTapes);
598 static void selectnewtape(Tuplesortstate *state);
599 static void init_slab_allocator(Tuplesortstate *state, int numSlots);
600 static void mergeruns(Tuplesortstate *state);
601 static void mergeonerun(Tuplesortstate *state);
602 static void beginmerge(Tuplesortstate *state);
603 static bool mergereadnext(Tuplesortstate *state, int srcTape, SortTuple *stup);
604 static void dumptuples(Tuplesortstate *state, bool alltuples);
612 static unsigned int getlen(Tuplesortstate *state, int tapenum, bool eofOK);
613 static void markrunend(Tuplesortstate *state, int tapenum);
614 static void *readtup_alloc(Tuplesortstate *state, Size tuplen);
615 static int comparetup_heap(const SortTuple *a, const SortTuple *b,
617 static void copytup_heap(Tuplesortstate *state, SortTuple *stup, void *tup);
618 static void writetup_heap(Tuplesortstate *state, int tapenum,
619  SortTuple *stup);
620 static void readtup_heap(Tuplesortstate *state, SortTuple *stup,
621  int tapenum, unsigned int len);
622 static int comparetup_cluster(const SortTuple *a, const SortTuple *b,
624 static void copytup_cluster(Tuplesortstate *state, SortTuple *stup, void *tup);
625 static void writetup_cluster(Tuplesortstate *state, int tapenum,
626  SortTuple *stup);
627 static void readtup_cluster(Tuplesortstate *state, SortTuple *stup,
628  int tapenum, unsigned int len);
629 static int comparetup_index_btree(const SortTuple *a, const SortTuple *b,
631 static int comparetup_index_hash(const SortTuple *a, const SortTuple *b,
633 static void copytup_index(Tuplesortstate *state, SortTuple *stup, void *tup);
634 static void writetup_index(Tuplesortstate *state, int tapenum,
635  SortTuple *stup);
636 static void readtup_index(Tuplesortstate *state, SortTuple *stup,
637  int tapenum, unsigned int len);
638 static int comparetup_datum(const SortTuple *a, const SortTuple *b,
640 static void copytup_datum(Tuplesortstate *state, SortTuple *stup, void *tup);
641 static void writetup_datum(Tuplesortstate *state, int tapenum,
642  SortTuple *stup);
643 static void readtup_datum(Tuplesortstate *state, SortTuple *stup,
644  int tapenum, unsigned int len);
649 static void free_sort_tuple(Tuplesortstate *state, SortTuple *stup);
650 
651 /*
652  * Special versions of qsort just for SortTuple objects. qsort_tuple() sorts
653  * any variant of SortTuples, using the appropriate comparetup function.
654  * qsort_ssup() is specialized for the case where the comparetup function
655  * reduces to ApplySortComparator(), that is single-key MinimalTuple sorts
656  * and Datum sorts.
657  */
658 #include "qsort_tuple.c"
659 
660 
661 /*
662  * tuplesort_begin_xxx
663  *
664  * Initialize for a tuple sort operation.
665  *
666  * After calling tuplesort_begin, the caller should call tuplesort_putXXX
667  * zero or more times, then call tuplesort_performsort when all the tuples
668  * have been supplied. After performsort, retrieve the tuples in sorted
669  * order by calling tuplesort_getXXX until it returns false/NULL. (If random
670  * access was requested, rescan, markpos, and restorepos can also be called.)
671  * Call tuplesort_end to terminate the operation and release memory/disk space.
672  *
673  * Each variant of tuplesort_begin has a workMem parameter specifying the
674  * maximum number of kilobytes of RAM to use before spilling data to disk.
675  * (The normal value of this parameter is work_mem, but some callers use
676  * other values.) Each variant also has a randomAccess parameter specifying
677  * whether the caller needs non-sequential access to the sort result.
678  */
679 
680 static Tuplesortstate *
681 tuplesort_begin_common(int workMem, SortCoordinate coordinate,
682  bool randomAccess)
683 {
685  MemoryContext sortcontext;
686  MemoryContext tuplecontext;
687  MemoryContext oldcontext;
688 
689  /* See leader_takeover_tapes() remarks on randomAccess support */
690  if (coordinate && randomAccess)
691  elog(ERROR, "random access disallowed under parallel sort");
692 
693  /*
694  * Create a working memory context for this sort operation. All data
695  * needed by the sort will live inside this context.
696  */
698  "TupleSort main",
700 
701  /*
702  * Caller tuple (e.g. IndexTuple) memory context.
703  *
704  * A dedicated child context used exclusively for caller passed tuples
705  * eases memory management. Resetting at key points reduces
706  * fragmentation. Note that the memtuples array of SortTuples is allocated
707  * in the parent context, not this context, because there is no need to
708  * free memtuples early.
709  */
710  tuplecontext = AllocSetContextCreate(sortcontext,
711  "Caller tuples",
713 
714  /*
715  * Make the Tuplesortstate within the per-sort context. This way, we
716  * don't need a separate pfree() operation for it at shutdown.
717  */
718  oldcontext = MemoryContextSwitchTo(sortcontext);
719 
720  state = (Tuplesortstate *) palloc0(sizeof(Tuplesortstate));
721 
722 #ifdef TRACE_SORT
723  if (trace_sort)
724  pg_rusage_init(&state->ru_start);
725 #endif
726 
727  state->status = TSS_INITIAL;
728  state->randomAccess = randomAccess;
729  state->bounded = false;
730  state->tuples = true;
731  state->boundUsed = false;
732 
733  /*
734  * workMem is forced to be at least 64KB, the current minimum valid value
735  * for the work_mem GUC. This is a defense against parallel sort callers
736  * that divide out memory among many workers in a way that leaves each
737  * with very little memory.
738  */
739  state->allowedMem = Max(workMem, 64) * (int64) 1024;
740  state->availMem = state->allowedMem;
741  state->sortcontext = sortcontext;
742  state->tuplecontext = tuplecontext;
743  state->tapeset = NULL;
744 
745  state->memtupcount = 0;
746 
747  /*
748  * Initial size of array must be more than ALLOCSET_SEPARATE_THRESHOLD;
749  * see comments in grow_memtuples().
750  */
751  state->memtupsize = Max(1024,
752  ALLOCSET_SEPARATE_THRESHOLD / sizeof(SortTuple) + 1);
753 
754  state->growmemtuples = true;
755  state->slabAllocatorUsed = false;
756  state->memtuples = (SortTuple *) palloc(state->memtupsize * sizeof(SortTuple));
757 
758  USEMEM(state, GetMemoryChunkSpace(state->memtuples));
759 
760  /* workMem must be large enough for the minimal memtuples array */
761  if (LACKMEM(state))
762  elog(ERROR, "insufficient memory allowed for sort");
763 
764  state->currentRun = 0;
765 
766  /*
767  * maxTapes, tapeRange, and Algorithm D variables will be initialized by
768  * inittapes(), if needed
769  */
770 
771  state->result_tape = -1; /* flag that result tape has not been formed */
772 
773  /*
774  * Initialize parallel-related state based on coordination information
775  * from caller
776  */
777  if (!coordinate)
778  {
779  /* Serial sort */
780  state->shared = NULL;
781  state->worker = -1;
782  state->nParticipants = -1;
783  }
784  else if (coordinate->isWorker)
785  {
786  /* Parallel worker produces exactly one final run from all input */
787  state->shared = coordinate->sharedsort;
788  state->worker = worker_get_identifier(state);
789  state->nParticipants = -1;
790  }
791  else
792  {
793  /* Parallel leader state only used for final merge */
794  state->shared = coordinate->sharedsort;
795  state->worker = -1;
796  state->nParticipants = coordinate->nParticipants;
797  Assert(state->nParticipants >= 1);
798  }
799 
800  MemoryContextSwitchTo(oldcontext);
801 
802  return state;
803 }
804 
807  int nkeys, AttrNumber *attNums,
808  Oid *sortOperators, Oid *sortCollations,
809  bool *nullsFirstFlags,
810  int workMem, SortCoordinate coordinate, bool randomAccess)
811 {
812  Tuplesortstate *state = tuplesort_begin_common(workMem, coordinate,
813  randomAccess);
814  MemoryContext oldcontext;
815  int i;
816 
817  oldcontext = MemoryContextSwitchTo(state->sortcontext);
818 
819  AssertArg(nkeys > 0);
820 
821 #ifdef TRACE_SORT
822  if (trace_sort)
823  elog(LOG,
824  "begin tuple sort: nkeys = %d, workMem = %d, randomAccess = %c",
825  nkeys, workMem, randomAccess ? 't' : 'f');
826 #endif
827 
828  state->nKeys = nkeys;
829 
830  TRACE_POSTGRESQL_SORT_START(HEAP_SORT,
831  false, /* no unique check */
832  nkeys,
833  workMem,
834  randomAccess,
835  PARALLEL_SORT(state));
836 
837  state->comparetup = comparetup_heap;
838  state->copytup = copytup_heap;
839  state->writetup = writetup_heap;
840  state->readtup = readtup_heap;
841 
842  state->tupDesc = tupDesc; /* assume we need not copy tupDesc */
843  state->abbrevNext = 10;
844 
845  /* Prepare SortSupport data for each column */
846  state->sortKeys = (SortSupport) palloc0(nkeys * sizeof(SortSupportData));
847 
848  for (i = 0; i < nkeys; i++)
849  {
850  SortSupport sortKey = state->sortKeys + i;
851 
852  AssertArg(attNums[i] != 0);
853  AssertArg(sortOperators[i] != 0);
854 
855  sortKey->ssup_cxt = CurrentMemoryContext;
856  sortKey->ssup_collation = sortCollations[i];
857  sortKey->ssup_nulls_first = nullsFirstFlags[i];
858  sortKey->ssup_attno = attNums[i];
859  /* Convey if abbreviation optimization is applicable in principle */
860  sortKey->abbreviate = (i == 0);
861 
862  PrepareSortSupportFromOrderingOp(sortOperators[i], sortKey);
863  }
864 
865  /*
866  * The "onlyKey" optimization cannot be used with abbreviated keys, since
867  * tie-breaker comparisons may be required. Typically, the optimization
868  * is only of value to pass-by-value types anyway, whereas abbreviated
869  * keys are typically only of value to pass-by-reference types.
870  */
871  if (nkeys == 1 && !state->sortKeys->abbrev_converter)
872  state->onlyKey = state->sortKeys;
873 
874  MemoryContextSwitchTo(oldcontext);
875 
876  return state;
877 }
878 
881  Relation indexRel,
882  int workMem,
883  SortCoordinate coordinate, bool randomAccess)
884 {
885  Tuplesortstate *state = tuplesort_begin_common(workMem, coordinate,
886  randomAccess);
887  ScanKey indexScanKey;
888  MemoryContext oldcontext;
889  int i;
890 
891  Assert(indexRel->rd_rel->relam == BTREE_AM_OID);
892 
893  oldcontext = MemoryContextSwitchTo(state->sortcontext);
894 
895 #ifdef TRACE_SORT
896  if (trace_sort)
897  elog(LOG,
898  "begin tuple sort: nkeys = %d, workMem = %d, randomAccess = %c",
900  workMem, randomAccess ? 't' : 'f');
901 #endif
902 
903  state->nKeys = RelationGetNumberOfAttributes(indexRel);
904 
905  TRACE_POSTGRESQL_SORT_START(CLUSTER_SORT,
906  false, /* no unique check */
907  state->nKeys,
908  workMem,
909  randomAccess,
910  PARALLEL_SORT(state));
911 
913  state->copytup = copytup_cluster;
914  state->writetup = writetup_cluster;
915  state->readtup = readtup_cluster;
916  state->abbrevNext = 10;
917 
918  state->indexInfo = BuildIndexInfo(indexRel);
919 
920  state->tupDesc = tupDesc; /* assume we need not copy tupDesc */
921 
922  indexScanKey = _bt_mkscankey_nodata(indexRel);
923 
924  if (state->indexInfo->ii_Expressions != NULL)
925  {
926  TupleTableSlot *slot;
927  ExprContext *econtext;
928 
929  /*
930  * We will need to use FormIndexDatum to evaluate the index
931  * expressions. To do that, we need an EState, as well as a
932  * TupleTableSlot to put the table tuples into. The econtext's
933  * scantuple has to point to that slot, too.
934  */
935  state->estate = CreateExecutorState();
936  slot = MakeSingleTupleTableSlot(tupDesc);
937  econtext = GetPerTupleExprContext(state->estate);
938  econtext->ecxt_scantuple = slot;
939  }
940 
941  /* Prepare SortSupport data for each column */
942  state->sortKeys = (SortSupport) palloc0(state->nKeys *
943  sizeof(SortSupportData));
944 
945  for (i = 0; i < state->nKeys; i++)
946  {
947  SortSupport sortKey = state->sortKeys + i;
948  ScanKey scanKey = indexScanKey + i;
949  int16 strategy;
950 
951  sortKey->ssup_cxt = CurrentMemoryContext;
952  sortKey->ssup_collation = scanKey->sk_collation;
953  sortKey->ssup_nulls_first =
954  (scanKey->sk_flags & SK_BT_NULLS_FIRST) != 0;
955  sortKey->ssup_attno = scanKey->sk_attno;
956  /* Convey if abbreviation optimization is applicable in principle */
957  sortKey->abbreviate = (i == 0);
958 
959  AssertState(sortKey->ssup_attno != 0);
960 
961  strategy = (scanKey->sk_flags & SK_BT_DESC) != 0 ?
963 
964  PrepareSortSupportFromIndexRel(indexRel, strategy, sortKey);
965  }
966 
967  _bt_freeskey(indexScanKey);
968 
969  MemoryContextSwitchTo(oldcontext);
970 
971  return state;
972 }
973 
976  Relation indexRel,
977  bool enforceUnique,
978  int workMem,
979  SortCoordinate coordinate,
980  bool randomAccess)
981 {
982  Tuplesortstate *state = tuplesort_begin_common(workMem, coordinate,
983  randomAccess);
984  ScanKey indexScanKey;
985  MemoryContext oldcontext;
986  int i;
987 
988  oldcontext = MemoryContextSwitchTo(state->sortcontext);
989 
990 #ifdef TRACE_SORT
991  if (trace_sort)
992  elog(LOG,
993  "begin index sort: unique = %c, workMem = %d, randomAccess = %c",
994  enforceUnique ? 't' : 'f',
995  workMem, randomAccess ? 't' : 'f');
996 #endif
997 
998  state->nKeys = RelationGetNumberOfAttributes(indexRel);
999 
1000  TRACE_POSTGRESQL_SORT_START(INDEX_SORT,
1001  enforceUnique,
1002  state->nKeys,
1003  workMem,
1004  randomAccess,
1005  PARALLEL_SORT(state));
1006 
1008  state->copytup = copytup_index;
1009  state->writetup = writetup_index;
1010  state->readtup = readtup_index;
1011  state->abbrevNext = 10;
1012 
1013  state->heapRel = heapRel;
1014  state->indexRel = indexRel;
1015  state->enforceUnique = enforceUnique;
1016 
1017  indexScanKey = _bt_mkscankey_nodata(indexRel);
1018  state->nKeys = RelationGetNumberOfAttributes(indexRel);
1019 
1020  /* Prepare SortSupport data for each column */
1021  state->sortKeys = (SortSupport) palloc0(state->nKeys *
1022  sizeof(SortSupportData));
1023 
1024  for (i = 0; i < state->nKeys; i++)
1025  {
1026  SortSupport sortKey = state->sortKeys + i;
1027  ScanKey scanKey = indexScanKey + i;
1028  int16 strategy;
1029 
1030  sortKey->ssup_cxt = CurrentMemoryContext;
1031  sortKey->ssup_collation = scanKey->sk_collation;
1032  sortKey->ssup_nulls_first =
1033  (scanKey->sk_flags & SK_BT_NULLS_FIRST) != 0;
1034  sortKey->ssup_attno = scanKey->sk_attno;
1035  /* Convey if abbreviation optimization is applicable in principle */
1036  sortKey->abbreviate = (i == 0);
1037 
1038  AssertState(sortKey->ssup_attno != 0);
1039 
1040  strategy = (scanKey->sk_flags & SK_BT_DESC) != 0 ?
1042 
1043  PrepareSortSupportFromIndexRel(indexRel, strategy, sortKey);
1044  }
1045 
1046  _bt_freeskey(indexScanKey);
1047 
1048  MemoryContextSwitchTo(oldcontext);
1049 
1050  return state;
1051 }
1052 
1055  Relation indexRel,
1056  uint32 high_mask,
1057  uint32 low_mask,
1058  uint32 max_buckets,
1059  int workMem,
1060  SortCoordinate coordinate,
1061  bool randomAccess)
1062 {
1063  Tuplesortstate *state = tuplesort_begin_common(workMem, coordinate,
1064  randomAccess);
1065  MemoryContext oldcontext;
1066 
1067  oldcontext = MemoryContextSwitchTo(state->sortcontext);
1068 
1069 #ifdef TRACE_SORT
1070  if (trace_sort)
1071  elog(LOG,
1072  "begin index sort: high_mask = 0x%x, low_mask = 0x%x, "
1073  "max_buckets = 0x%x, workMem = %d, randomAccess = %c",
1074  high_mask,
1075  low_mask,
1076  max_buckets,
1077  workMem, randomAccess ? 't' : 'f');
1078 #endif
1079 
1080  state->nKeys = 1; /* Only one sort column, the hash code */
1081 
1083  state->copytup = copytup_index;
1084  state->writetup = writetup_index;
1085  state->readtup = readtup_index;
1086 
1087  state->heapRel = heapRel;
1088  state->indexRel = indexRel;
1089 
1090  state->high_mask = high_mask;
1091  state->low_mask = low_mask;
1092  state->max_buckets = max_buckets;
1093 
1094  MemoryContextSwitchTo(oldcontext);
1095 
1096  return state;
1097 }
1098 
1100 tuplesort_begin_datum(Oid datumType, Oid sortOperator, Oid sortCollation,
1101  bool nullsFirstFlag, int workMem,
1102  SortCoordinate coordinate, bool randomAccess)
1103 {
1104  Tuplesortstate *state = tuplesort_begin_common(workMem, coordinate,
1105  randomAccess);
1106  MemoryContext oldcontext;
1107  int16 typlen;
1108  bool typbyval;
1109 
1110  oldcontext = MemoryContextSwitchTo(state->sortcontext);
1111 
1112 #ifdef TRACE_SORT
1113  if (trace_sort)
1114  elog(LOG,
1115  "begin datum sort: workMem = %d, randomAccess = %c",
1116  workMem, randomAccess ? 't' : 'f');
1117 #endif
1118 
1119  state->nKeys = 1; /* always a one-column sort */
1120 
1121  TRACE_POSTGRESQL_SORT_START(DATUM_SORT,
1122  false, /* no unique check */
1123  1,
1124  workMem,
1125  randomAccess,
1126  PARALLEL_SORT(state));
1127 
1128  state->comparetup = comparetup_datum;
1129  state->copytup = copytup_datum;
1130  state->writetup = writetup_datum;
1131  state->readtup = readtup_datum;
1132  state->abbrevNext = 10;
1133 
1134  state->datumType = datumType;
1135 
1136  /* lookup necessary attributes of the datum type */
1137  get_typlenbyval(datumType, &typlen, &typbyval);
1138  state->datumTypeLen = typlen;
1139  state->tuples = !typbyval;
1140 
1141  /* Prepare SortSupport data */
1142  state->sortKeys = (SortSupport) palloc0(sizeof(SortSupportData));
1143 
1145  state->sortKeys->ssup_collation = sortCollation;
1146  state->sortKeys->ssup_nulls_first = nullsFirstFlag;
1147 
1148  /*
1149  * Abbreviation is possible here only for by-reference types. In theory,
1150  * a pass-by-value datatype could have an abbreviated form that is cheaper
1151  * to compare. In a tuple sort, we could support that, because we can
1152  * always extract the original datum from the tuple is needed. Here, we
1153  * can't, because a datum sort only stores a single copy of the datum; the
1154  * "tuple" field of each sortTuple is NULL.
1155  */
1156  state->sortKeys->abbreviate = !typbyval;
1157 
1158  PrepareSortSupportFromOrderingOp(sortOperator, state->sortKeys);
1159 
1160  /*
1161  * The "onlyKey" optimization cannot be used with abbreviated keys, since
1162  * tie-breaker comparisons may be required. Typically, the optimization
1163  * is only of value to pass-by-value types anyway, whereas abbreviated
1164  * keys are typically only of value to pass-by-reference types.
1165  */
1166  if (!state->sortKeys->abbrev_converter)
1167  state->onlyKey = state->sortKeys;
1168 
1169  MemoryContextSwitchTo(oldcontext);
1170 
1171  return state;
1172 }
1173 
1174 /*
1175  * tuplesort_set_bound
1176  *
1177  * Advise tuplesort that at most the first N result tuples are required.
1178  *
1179  * Must be called before inserting any tuples. (Actually, we could allow it
1180  * as long as the sort hasn't spilled to disk, but there seems no need for
1181  * delayed calls at the moment.)
1182  *
1183  * This is a hint only. The tuplesort may still return more tuples than
1184  * requested. Parallel leader tuplesorts will always ignore the hint.
1185  */
1186 void
1188 {
1189  /* Assert we're called before loading any tuples */
1190  Assert(state->status == TSS_INITIAL);
1191  Assert(state->memtupcount == 0);
1192  Assert(!state->bounded);
1193  Assert(!WORKER(state));
1194 
1195 #ifdef DEBUG_BOUNDED_SORT
1196  /* Honor GUC setting that disables the feature (for easy testing) */
1197  if (!optimize_bounded_sort)
1198  return;
1199 #endif
1200 
1201  /* Parallel leader ignores hint */
1202  if (LEADER(state))
1203  return;
1204 
1205  /* We want to be able to compute bound * 2, so limit the setting */
1206  if (bound > (int64) (INT_MAX / 2))
1207  return;
1208 
1209  state->bounded = true;
1210  state->bound = (int) bound;
1211 
1212  /*
1213  * Bounded sorts are not an effective target for abbreviated key
1214  * optimization. Disable by setting state to be consistent with no
1215  * abbreviation support.
1216  */
1217  state->sortKeys->abbrev_converter = NULL;
1218  if (state->sortKeys->abbrev_full_comparator)
1220 
1221  /* Not strictly necessary, but be tidy */
1222  state->sortKeys->abbrev_abort = NULL;
1223  state->sortKeys->abbrev_full_comparator = NULL;
1224 }
1225 
1226 /*
1227  * tuplesort_end
1228  *
1229  * Release resources and clean up.
1230  *
1231  * NOTE: after calling this, any pointers returned by tuplesort_getXXX are
1232  * pointing to garbage. Be careful not to attempt to use or free such
1233  * pointers afterwards!
1234  */
1235 void
1237 {
1238  /* context swap probably not needed, but let's be safe */
1239  MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
1240 
1241 #ifdef TRACE_SORT
1242  long spaceUsed;
1243 
1244  if (state->tapeset)
1245  spaceUsed = LogicalTapeSetBlocks(state->tapeset);
1246  else
1247  spaceUsed = (state->allowedMem - state->availMem + 1023) / 1024;
1248 #endif
1249 
1250  /*
1251  * Delete temporary "tape" files, if any.
1252  *
1253  * Note: want to include this in reported total cost of sort, hence need
1254  * for two #ifdef TRACE_SORT sections.
1255  */
1256  if (state->tapeset)
1257  LogicalTapeSetClose(state->tapeset);
1258 
1259 #ifdef TRACE_SORT
1260  if (trace_sort)
1261  {
1262  if (state->tapeset)
1263  elog(LOG, "%s of %d ended, %ld disk blocks used: %s",
1264  SERIAL(state) ? "external sort" : "parallel external sort",
1265  state->worker, spaceUsed, pg_rusage_show(&state->ru_start));
1266  else
1267  elog(LOG, "%s of %d ended, %ld KB used: %s",
1268  SERIAL(state) ? "internal sort" : "unperformed parallel sort",
1269  state->worker, spaceUsed, pg_rusage_show(&state->ru_start));
1270  }
1271 
1272  TRACE_POSTGRESQL_SORT_DONE(state->tapeset != NULL, spaceUsed);
1273 #else
1274 
1275  /*
1276  * If you disabled TRACE_SORT, you can still probe sort__done, but you
1277  * ain't getting space-used stats.
1278  */
1279  TRACE_POSTGRESQL_SORT_DONE(state->tapeset != NULL, 0L);
1280 #endif
1281 
1282  /* Free any execution state created for CLUSTER case */
1283  if (state->estate != NULL)
1284  {
1285  ExprContext *econtext = GetPerTupleExprContext(state->estate);
1286 
1288  FreeExecutorState(state->estate);
1289  }
1290 
1291  MemoryContextSwitchTo(oldcontext);
1292 
1293  /*
1294  * Free the per-sort memory context, thereby releasing all working memory,
1295  * including the Tuplesortstate struct itself.
1296  */
1298 }
1299 
1300 /*
1301  * Grow the memtuples[] array, if possible within our memory constraint. We
1302  * must not exceed INT_MAX tuples in memory or the caller-provided memory
1303  * limit. Return true if we were able to enlarge the array, false if not.
1304  *
1305  * Normally, at each increment we double the size of the array. When doing
1306  * that would exceed a limit, we attempt one last, smaller increase (and then
1307  * clear the growmemtuples flag so we don't try any more). That allows us to
1308  * use memory as fully as permitted; sticking to the pure doubling rule could
1309  * result in almost half going unused. Because availMem moves around with
1310  * tuple addition/removal, we need some rule to prevent making repeated small
1311  * increases in memtupsize, which would just be useless thrashing. The
1312  * growmemtuples flag accomplishes that and also prevents useless
1313  * recalculations in this function.
1314  */
1315 static bool
1317 {
1318  int newmemtupsize;
1319  int memtupsize = state->memtupsize;
1320  int64 memNowUsed = state->allowedMem - state->availMem;
1321 
1322  /* Forget it if we've already maxed out memtuples, per comment above */
1323  if (!state->growmemtuples)
1324  return false;
1325 
1326  /* Select new value of memtupsize */
1327  if (memNowUsed <= state->availMem)
1328  {
1329  /*
1330  * We've used no more than half of allowedMem; double our usage,
1331  * clamping at INT_MAX tuples.
1332  */
1333  if (memtupsize < INT_MAX / 2)
1334  newmemtupsize = memtupsize * 2;
1335  else
1336  {
1337  newmemtupsize = INT_MAX;
1338  state->growmemtuples = false;
1339  }
1340  }
1341  else
1342  {
1343  /*
1344  * This will be the last increment of memtupsize. Abandon doubling
1345  * strategy and instead increase as much as we safely can.
1346  *
1347  * To stay within allowedMem, we can't increase memtupsize by more
1348  * than availMem / sizeof(SortTuple) elements. In practice, we want
1349  * to increase it by considerably less, because we need to leave some
1350  * space for the tuples to which the new array slots will refer. We
1351  * assume the new tuples will be about the same size as the tuples
1352  * we've already seen, and thus we can extrapolate from the space
1353  * consumption so far to estimate an appropriate new size for the
1354  * memtuples array. The optimal value might be higher or lower than
1355  * this estimate, but it's hard to know that in advance. We again
1356  * clamp at INT_MAX tuples.
1357  *
1358  * This calculation is safe against enlarging the array so much that
1359  * LACKMEM becomes true, because the memory currently used includes
1360  * the present array; thus, there would be enough allowedMem for the
1361  * new array elements even if no other memory were currently used.
1362  *
1363  * We do the arithmetic in float8, because otherwise the product of
1364  * memtupsize and allowedMem could overflow. Any inaccuracy in the
1365  * result should be insignificant; but even if we computed a
1366  * completely insane result, the checks below will prevent anything
1367  * really bad from happening.
1368  */
1369  double grow_ratio;
1370 
1371  grow_ratio = (double) state->allowedMem / (double) memNowUsed;
1372  if (memtupsize * grow_ratio < INT_MAX)
1373  newmemtupsize = (int) (memtupsize * grow_ratio);
1374  else
1375  newmemtupsize = INT_MAX;
1376 
1377  /* We won't make any further enlargement attempts */
1378  state->growmemtuples = false;
1379  }
1380 
1381  /* Must enlarge array by at least one element, else report failure */
1382  if (newmemtupsize <= memtupsize)
1383  goto noalloc;
1384 
1385  /*
1386  * On a 32-bit machine, allowedMem could exceed MaxAllocHugeSize. Clamp
1387  * to ensure our request won't be rejected. Note that we can easily
1388  * exhaust address space before facing this outcome. (This is presently
1389  * impossible due to guc.c's MAX_KILOBYTES limitation on work_mem, but
1390  * don't rely on that at this distance.)
1391  */
1392  if ((Size) newmemtupsize >= MaxAllocHugeSize / sizeof(SortTuple))
1393  {
1394  newmemtupsize = (int) (MaxAllocHugeSize / sizeof(SortTuple));
1395  state->growmemtuples = false; /* can't grow any more */
1396  }
1397 
1398  /*
1399  * We need to be sure that we do not cause LACKMEM to become true, else
1400  * the space management algorithm will go nuts. The code above should
1401  * never generate a dangerous request, but to be safe, check explicitly
1402  * that the array growth fits within availMem. (We could still cause
1403  * LACKMEM if the memory chunk overhead associated with the memtuples
1404  * array were to increase. That shouldn't happen because we chose the
1405  * initial array size large enough to ensure that palloc will be treating
1406  * both old and new arrays as separate chunks. But we'll check LACKMEM
1407  * explicitly below just in case.)
1408  */
1409  if (state->availMem < (int64) ((newmemtupsize - memtupsize) * sizeof(SortTuple)))
1410  goto noalloc;
1411 
1412  /* OK, do it */
1413  FREEMEM(state, GetMemoryChunkSpace(state->memtuples));
1414  state->memtupsize = newmemtupsize;
1415  state->memtuples = (SortTuple *)
1416  repalloc_huge(state->memtuples,
1417  state->memtupsize * sizeof(SortTuple));
1418  USEMEM(state, GetMemoryChunkSpace(state->memtuples));
1419  if (LACKMEM(state))
1420  elog(ERROR, "unexpected out-of-memory situation in tuplesort");
1421  return true;
1422 
1423 noalloc:
1424  /* If for any reason we didn't realloc, shut off future attempts */
1425  state->growmemtuples = false;
1426  return false;
1427 }
1428 
1429 /*
1430  * Accept one tuple while collecting input data for sort.
1431  *
1432  * Note that the input data is always copied; the caller need not save it.
1433  */
1434 void
1436 {
1437  MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
1438  SortTuple stup;
1439 
1440  /*
1441  * Copy the given tuple into memory we control, and decrease availMem.
1442  * Then call the common code.
1443  */
1444  COPYTUP(state, &stup, (void *) slot);
1445 
1446  puttuple_common(state, &stup);
1447 
1448  MemoryContextSwitchTo(oldcontext);
1449 }
1450 
1451 /*
1452  * Accept one tuple while collecting input data for sort.
1453  *
1454  * Note that the input data is always copied; the caller need not save it.
1455  */
1456 void
1458 {
1459  MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
1460  SortTuple stup;
1461 
1462  /*
1463  * Copy the given tuple into memory we control, and decrease availMem.
1464  * Then call the common code.
1465  */
1466  COPYTUP(state, &stup, (void *) tup);
1467 
1468  puttuple_common(state, &stup);
1469 
1470  MemoryContextSwitchTo(oldcontext);
1471 }
1472 
1473 /*
1474  * Collect one index tuple while collecting input data for sort, building
1475  * it from caller-supplied values.
1476  */
1477 void
1479  ItemPointer self, Datum *values,
1480  bool *isnull)
1481 {
1482  MemoryContext oldcontext = MemoryContextSwitchTo(state->tuplecontext);
1483  SortTuple stup;
1484  Datum original;
1485  IndexTuple tuple;
1486 
1487  stup.tuple = index_form_tuple(RelationGetDescr(rel), values, isnull);
1488  tuple = ((IndexTuple) stup.tuple);
1489  tuple->t_tid = *self;
1490  USEMEM(state, GetMemoryChunkSpace(stup.tuple));
1491  /* set up first-column key value */
1492  original = index_getattr(tuple,
1493  1,
1494  RelationGetDescr(state->indexRel),
1495  &stup.isnull1);
1496 
1498 
1499  if (!state->sortKeys || !state->sortKeys->abbrev_converter || stup.isnull1)
1500  {
1501  /*
1502  * Store ordinary Datum representation, or NULL value. If there is a
1503  * converter it won't expect NULL values, and cost model is not
1504  * required to account for NULL, so in that case we avoid calling
1505  * converter and just set datum1 to zeroed representation (to be
1506  * consistent, and to support cheap inequality tests for NULL
1507  * abbreviated keys).
1508  */
1509  stup.datum1 = original;
1510  }
1511  else if (!consider_abort_common(state))
1512  {
1513  /* Store abbreviated key representation */
1514  stup.datum1 = state->sortKeys->abbrev_converter(original,
1515  state->sortKeys);
1516  }
1517  else
1518  {
1519  /* Abort abbreviation */
1520  int i;
1521 
1522  stup.datum1 = original;
1523 
1524  /*
1525  * Set state to be consistent with never trying abbreviation.
1526  *
1527  * Alter datum1 representation in already-copied tuples, so as to
1528  * ensure a consistent representation (current tuple was just
1529  * handled). It does not matter if some dumped tuples are already
1530  * sorted on tape, since serialized tuples lack abbreviated keys
1531  * (TSS_BUILDRUNS state prevents control reaching here in any case).
1532  */
1533  for (i = 0; i < state->memtupcount; i++)
1534  {
1535  SortTuple *mtup = &state->memtuples[i];
1536 
1537  tuple = mtup->tuple;
1538  mtup->datum1 = index_getattr(tuple,
1539  1,
1540  RelationGetDescr(state->indexRel),
1541  &mtup->isnull1);
1542  }
1543  }
1544 
1545  puttuple_common(state, &stup);
1546 
1547  MemoryContextSwitchTo(oldcontext);
1548 }
1549 
1550 /*
1551  * Accept one Datum while collecting input data for sort.
1552  *
1553  * If the Datum is pass-by-ref type, the value will be copied.
1554  */
1555 void
1557 {
1558  MemoryContext oldcontext = MemoryContextSwitchTo(state->tuplecontext);
1559  SortTuple stup;
1560 
1561  /*
1562  * Pass-by-value types or null values are just stored directly in
1563  * stup.datum1 (and stup.tuple is not used and set to NULL).
1564  *
1565  * Non-null pass-by-reference values need to be copied into memory we
1566  * control, and possibly abbreviated. The copied value is pointed to by
1567  * stup.tuple and is treated as the canonical copy (e.g. to return via
1568  * tuplesort_getdatum or when writing to tape); stup.datum1 gets the
1569  * abbreviated value if abbreviation is happening, otherwise it's
1570  * identical to stup.tuple.
1571  */
1572 
1573  if (isNull || !state->tuples)
1574  {
1575  /*
1576  * Set datum1 to zeroed representation for NULLs (to be consistent,
1577  * and to support cheap inequality tests for NULL abbreviated keys).
1578  */
1579  stup.datum1 = !isNull ? val : (Datum) 0;
1580  stup.isnull1 = isNull;
1581  stup.tuple = NULL; /* no separate storage */
1583  }
1584  else
1585  {
1586  Datum original = datumCopy(val, false, state->datumTypeLen);
1587 
1588  stup.isnull1 = false;
1589  stup.tuple = DatumGetPointer(original);
1590  USEMEM(state, GetMemoryChunkSpace(stup.tuple));
1592 
1593  if (!state->sortKeys->abbrev_converter)
1594  {
1595  stup.datum1 = original;
1596  }
1597  else if (!consider_abort_common(state))
1598  {
1599  /* Store abbreviated key representation */
1600  stup.datum1 = state->sortKeys->abbrev_converter(original,
1601  state->sortKeys);
1602  }
1603  else
1604  {
1605  /* Abort abbreviation */
1606  int i;
1607 
1608  stup.datum1 = original;
1609 
1610  /*
1611  * Set state to be consistent with never trying abbreviation.
1612  *
1613  * Alter datum1 representation in already-copied tuples, so as to
1614  * ensure a consistent representation (current tuple was just
1615  * handled). It does not matter if some dumped tuples are already
1616  * sorted on tape, since serialized tuples lack abbreviated keys
1617  * (TSS_BUILDRUNS state prevents control reaching here in any
1618  * case).
1619  */
1620  for (i = 0; i < state->memtupcount; i++)
1621  {
1622  SortTuple *mtup = &state->memtuples[i];
1623 
1624  mtup->datum1 = PointerGetDatum(mtup->tuple);
1625  }
1626  }
1627  }
1628 
1629  puttuple_common(state, &stup);
1630 
1631  MemoryContextSwitchTo(oldcontext);
1632 }
1633 
1634 /*
1635  * Shared code for tuple and datum cases.
1636  */
1637 static void
1639 {
1640  Assert(!LEADER(state));
1641 
1642  switch (state->status)
1643  {
1644  case TSS_INITIAL:
1645 
1646  /*
1647  * Save the tuple into the unsorted array. First, grow the array
1648  * as needed. Note that we try to grow the array when there is
1649  * still one free slot remaining --- if we fail, there'll still be
1650  * room to store the incoming tuple, and then we'll switch to
1651  * tape-based operation.
1652  */
1653  if (state->memtupcount >= state->memtupsize - 1)
1654  {
1655  (void) grow_memtuples(state);
1656  Assert(state->memtupcount < state->memtupsize);
1657  }
1658  state->memtuples[state->memtupcount++] = *tuple;
1659 
1660  /*
1661  * Check if it's time to switch over to a bounded heapsort. We do
1662  * so if the input tuple count exceeds twice the desired tuple
1663  * count (this is a heuristic for where heapsort becomes cheaper
1664  * than a quicksort), or if we've just filled workMem and have
1665  * enough tuples to meet the bound.
1666  *
1667  * Note that once we enter TSS_BOUNDED state we will always try to
1668  * complete the sort that way. In the worst case, if later input
1669  * tuples are larger than earlier ones, this might cause us to
1670  * exceed workMem significantly.
1671  */
1672  if (state->bounded &&
1673  (state->memtupcount > state->bound * 2 ||
1674  (state->memtupcount > state->bound && LACKMEM(state))))
1675  {
1676 #ifdef TRACE_SORT
1677  if (trace_sort)
1678  elog(LOG, "switching to bounded heapsort at %d tuples: %s",
1679  state->memtupcount,
1680  pg_rusage_show(&state->ru_start));
1681 #endif
1682  make_bounded_heap(state);
1683  return;
1684  }
1685 
1686  /*
1687  * Done if we still fit in available memory and have array slots.
1688  */
1689  if (state->memtupcount < state->memtupsize && !LACKMEM(state))
1690  return;
1691 
1692  /*
1693  * Nope; time to switch to tape-based operation.
1694  */
1695  inittapes(state, true);
1696 
1697  /*
1698  * Dump all tuples.
1699  */
1700  dumptuples(state, false);
1701  break;
1702 
1703  case TSS_BOUNDED:
1704 
1705  /*
1706  * We don't want to grow the array here, so check whether the new
1707  * tuple can be discarded before putting it in. This should be a
1708  * good speed optimization, too, since when there are many more
1709  * input tuples than the bound, most input tuples can be discarded
1710  * with just this one comparison. Note that because we currently
1711  * have the sort direction reversed, we must check for <= not >=.
1712  */
1713  if (COMPARETUP(state, tuple, &state->memtuples[0]) <= 0)
1714  {
1715  /* new tuple <= top of the heap, so we can discard it */
1716  free_sort_tuple(state, tuple);
1718  }
1719  else
1720  {
1721  /* discard top of heap, replacing it with the new tuple */
1722  free_sort_tuple(state, &state->memtuples[0]);
1723  tuplesort_heap_replace_top(state, tuple);
1724  }
1725  break;
1726 
1727  case TSS_BUILDRUNS:
1728 
1729  /*
1730  * Save the tuple into the unsorted array (there must be space)
1731  */
1732  state->memtuples[state->memtupcount++] = *tuple;
1733 
1734  /*
1735  * If we are over the memory limit, dump all tuples.
1736  */
1737  dumptuples(state, false);
1738  break;
1739 
1740  default:
1741  elog(ERROR, "invalid tuplesort state");
1742  break;
1743  }
1744 }
1745 
1746 static bool
1748 {
1749  Assert(state->sortKeys[0].abbrev_converter != NULL);
1750  Assert(state->sortKeys[0].abbrev_abort != NULL);
1751  Assert(state->sortKeys[0].abbrev_full_comparator != NULL);
1752 
1753  /*
1754  * Check effectiveness of abbreviation optimization. Consider aborting
1755  * when still within memory limit.
1756  */
1757  if (state->status == TSS_INITIAL &&
1758  state->memtupcount >= state->abbrevNext)
1759  {
1760  state->abbrevNext *= 2;
1761 
1762  /*
1763  * Check opclass-supplied abbreviation abort routine. It may indicate
1764  * that abbreviation should not proceed.
1765  */
1766  if (!state->sortKeys->abbrev_abort(state->memtupcount,
1767  state->sortKeys))
1768  return false;
1769 
1770  /*
1771  * Finally, restore authoritative comparator, and indicate that
1772  * abbreviation is not in play by setting abbrev_converter to NULL
1773  */
1774  state->sortKeys[0].comparator = state->sortKeys[0].abbrev_full_comparator;
1775  state->sortKeys[0].abbrev_converter = NULL;
1776  /* Not strictly necessary, but be tidy */
1777  state->sortKeys[0].abbrev_abort = NULL;
1778  state->sortKeys[0].abbrev_full_comparator = NULL;
1779 
1780  /* Give up - expect original pass-by-value representation */
1781  return true;
1782  }
1783 
1784  return false;
1785 }
1786 
1787 /*
1788  * All tuples have been provided; finish the sort.
1789  */
1790 void
1792 {
1793  MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
1794 
1795 #ifdef TRACE_SORT
1796  if (trace_sort)
1797  elog(LOG, "performsort of %d starting: %s",
1798  state->worker, pg_rusage_show(&state->ru_start));
1799 #endif
1800 
1801  switch (state->status)
1802  {
1803  case TSS_INITIAL:
1804 
1805  /*
1806  * We were able to accumulate all the tuples within the allowed
1807  * amount of memory, or leader to take over worker tapes
1808  */
1809  if (SERIAL(state))
1810  {
1811  /* Just qsort 'em and we're done */
1812  tuplesort_sort_memtuples(state);
1813  state->status = TSS_SORTEDINMEM;
1814  }
1815  else if (WORKER(state))
1816  {
1817  /*
1818  * Parallel workers must still dump out tuples to tape. No
1819  * merge is required to produce single output run, though.
1820  */
1821  inittapes(state, false);
1822  dumptuples(state, true);
1823  worker_nomergeruns(state);
1824  state->status = TSS_SORTEDONTAPE;
1825  }
1826  else
1827  {
1828  /*
1829  * Leader will take over worker tapes and merge worker runs.
1830  * Note that mergeruns sets the correct state->status.
1831  */
1832  leader_takeover_tapes(state);
1833  mergeruns(state);
1834  }
1835  state->current = 0;
1836  state->eof_reached = false;
1837  state->markpos_block = 0L;
1838  state->markpos_offset = 0;
1839  state->markpos_eof = false;
1840  break;
1841 
1842  case TSS_BOUNDED:
1843 
1844  /*
1845  * We were able to accumulate all the tuples required for output
1846  * in memory, using a heap to eliminate excess tuples. Now we
1847  * have to transform the heap to a properly-sorted array.
1848  */
1849  sort_bounded_heap(state);
1850  state->current = 0;
1851  state->eof_reached = false;
1852  state->markpos_offset = 0;
1853  state->markpos_eof = false;
1854  state->status = TSS_SORTEDINMEM;
1855  break;
1856 
1857  case TSS_BUILDRUNS:
1858 
1859  /*
1860  * Finish tape-based sort. First, flush all tuples remaining in
1861  * memory out to tape; then merge until we have a single remaining
1862  * run (or, if !randomAccess and !WORKER(), one run per tape).
1863  * Note that mergeruns sets the correct state->status.
1864  */
1865  dumptuples(state, true);
1866  mergeruns(state);
1867  state->eof_reached = false;
1868  state->markpos_block = 0L;
1869  state->markpos_offset = 0;
1870  state->markpos_eof = false;
1871  break;
1872 
1873  default:
1874  elog(ERROR, "invalid tuplesort state");
1875  break;
1876  }
1877 
1878 #ifdef TRACE_SORT
1879  if (trace_sort)
1880  {
1881  if (state->status == TSS_FINALMERGE)
1882  elog(LOG, "performsort of %d done (except %d-way final merge): %s",
1883  state->worker, state->activeTapes,
1884  pg_rusage_show(&state->ru_start));
1885  else
1886  elog(LOG, "performsort of %d done: %s",
1887  state->worker, pg_rusage_show(&state->ru_start));
1888  }
1889 #endif
1890 
1891  MemoryContextSwitchTo(oldcontext);
1892 }
1893 
1894 /*
1895  * Internal routine to fetch the next tuple in either forward or back
1896  * direction into *stup. Returns false if no more tuples.
1897  * Returned tuple belongs to tuplesort memory context, and must not be freed
1898  * by caller. Note that fetched tuple is stored in memory that may be
1899  * recycled by any future fetch.
1900  */
1901 static bool
1903  SortTuple *stup)
1904 {
1905  unsigned int tuplen;
1906  size_t nmoved;
1907 
1908  Assert(!WORKER(state));
1909 
1910  switch (state->status)
1911  {
1912  case TSS_SORTEDINMEM:
1913  Assert(forward || state->randomAccess);
1914  Assert(!state->slabAllocatorUsed);
1915  if (forward)
1916  {
1917  if (state->current < state->memtupcount)
1918  {
1919  *stup = state->memtuples[state->current++];
1920  return true;
1921  }
1922  state->eof_reached = true;
1923 
1924  /*
1925  * Complain if caller tries to retrieve more tuples than
1926  * originally asked for in a bounded sort. This is because
1927  * returning EOF here might be the wrong thing.
1928  */
1929  if (state->bounded && state->current >= state->bound)
1930  elog(ERROR, "retrieved too many tuples in a bounded sort");
1931 
1932  return false;
1933  }
1934  else
1935  {
1936  if (state->current <= 0)
1937  return false;
1938 
1939  /*
1940  * if all tuples are fetched already then we return last
1941  * tuple, else - tuple before last returned.
1942  */
1943  if (state->eof_reached)
1944  state->eof_reached = false;
1945  else
1946  {
1947  state->current--; /* last returned tuple */
1948  if (state->current <= 0)
1949  return false;
1950  }
1951  *stup = state->memtuples[state->current - 1];
1952  return true;
1953  }
1954  break;
1955 
1956  case TSS_SORTEDONTAPE:
1957  Assert(forward || state->randomAccess);
1958  Assert(state->slabAllocatorUsed);
1959 
1960  /*
1961  * The slot that held the tuple that we returned in previous
1962  * gettuple call can now be reused.
1963  */
1964  if (state->lastReturnedTuple)
1965  {
1966  RELEASE_SLAB_SLOT(state, state->lastReturnedTuple);
1967  state->lastReturnedTuple = NULL;
1968  }
1969 
1970  if (forward)
1971  {
1972  if (state->eof_reached)
1973  return false;
1974 
1975  if ((tuplen = getlen(state, state->result_tape, true)) != 0)
1976  {
1977  READTUP(state, stup, state->result_tape, tuplen);
1978 
1979  /*
1980  * Remember the tuple we return, so that we can recycle
1981  * its memory on next call. (This can be NULL, in the
1982  * !state->tuples case).
1983  */
1984  state->lastReturnedTuple = stup->tuple;
1985 
1986  return true;
1987  }
1988  else
1989  {
1990  state->eof_reached = true;
1991  return false;
1992  }
1993  }
1994 
1995  /*
1996  * Backward.
1997  *
1998  * if all tuples are fetched already then we return last tuple,
1999  * else - tuple before last returned.
2000  */
2001  if (state->eof_reached)
2002  {
2003  /*
2004  * Seek position is pointing just past the zero tuplen at the
2005  * end of file; back up to fetch last tuple's ending length
2006  * word. If seek fails we must have a completely empty file.
2007  */
2008  nmoved = LogicalTapeBackspace(state->tapeset,
2009  state->result_tape,
2010  2 * sizeof(unsigned int));
2011  if (nmoved == 0)
2012  return false;
2013  else if (nmoved != 2 * sizeof(unsigned int))
2014  elog(ERROR, "unexpected tape position");
2015  state->eof_reached = false;
2016  }
2017  else
2018  {
2019  /*
2020  * Back up and fetch previously-returned tuple's ending length
2021  * word. If seek fails, assume we are at start of file.
2022  */
2023  nmoved = LogicalTapeBackspace(state->tapeset,
2024  state->result_tape,
2025  sizeof(unsigned int));
2026  if (nmoved == 0)
2027  return false;
2028  else if (nmoved != sizeof(unsigned int))
2029  elog(ERROR, "unexpected tape position");
2030  tuplen = getlen(state, state->result_tape, false);
2031 
2032  /*
2033  * Back up to get ending length word of tuple before it.
2034  */
2035  nmoved = LogicalTapeBackspace(state->tapeset,
2036  state->result_tape,
2037  tuplen + 2 * sizeof(unsigned int));
2038  if (nmoved == tuplen + sizeof(unsigned int))
2039  {
2040  /*
2041  * We backed up over the previous tuple, but there was no
2042  * ending length word before it. That means that the prev
2043  * tuple is the first tuple in the file. It is now the
2044  * next to read in forward direction (not obviously right,
2045  * but that is what in-memory case does).
2046  */
2047  return false;
2048  }
2049  else if (nmoved != tuplen + 2 * sizeof(unsigned int))
2050  elog(ERROR, "bogus tuple length in backward scan");
2051  }
2052 
2053  tuplen = getlen(state, state->result_tape, false);
2054 
2055  /*
2056  * Now we have the length of the prior tuple, back up and read it.
2057  * Note: READTUP expects we are positioned after the initial
2058  * length word of the tuple, so back up to that point.
2059  */
2060  nmoved = LogicalTapeBackspace(state->tapeset,
2061  state->result_tape,
2062  tuplen);
2063  if (nmoved != tuplen)
2064  elog(ERROR, "bogus tuple length in backward scan");
2065  READTUP(state, stup, state->result_tape, tuplen);
2066 
2067  /*
2068  * Remember the tuple we return, so that we can recycle its memory
2069  * on next call. (This can be NULL, in the Datum case).
2070  */
2071  state->lastReturnedTuple = stup->tuple;
2072 
2073  return true;
2074 
2075  case TSS_FINALMERGE:
2076  Assert(forward);
2077  /* We are managing memory ourselves, with the slab allocator. */
2078  Assert(state->slabAllocatorUsed);
2079 
2080  /*
2081  * The slab slot holding the tuple that we returned in previous
2082  * gettuple call can now be reused.
2083  */
2084  if (state->lastReturnedTuple)
2085  {
2086  RELEASE_SLAB_SLOT(state, state->lastReturnedTuple);
2087  state->lastReturnedTuple = NULL;
2088  }
2089 
2090  /*
2091  * This code should match the inner loop of mergeonerun().
2092  */
2093  if (state->memtupcount > 0)
2094  {
2095  int srcTape = state->memtuples[0].tupindex;
2096  SortTuple newtup;
2097 
2098  *stup = state->memtuples[0];
2099 
2100  /*
2101  * Remember the tuple we return, so that we can recycle its
2102  * memory on next call. (This can be NULL, in the Datum case).
2103  */
2104  state->lastReturnedTuple = stup->tuple;
2105 
2106  /*
2107  * Pull next tuple from tape, and replace the returned tuple
2108  * at top of the heap with it.
2109  */
2110  if (!mergereadnext(state, srcTape, &newtup))
2111  {
2112  /*
2113  * If no more data, we've reached end of run on this tape.
2114  * Remove the top node from the heap.
2115  */
2117 
2118  /*
2119  * Rewind to free the read buffer. It'd go away at the
2120  * end of the sort anyway, but better to release the
2121  * memory early.
2122  */
2123  LogicalTapeRewindForWrite(state->tapeset, srcTape);
2124  return true;
2125  }
2126  newtup.tupindex = srcTape;
2127  tuplesort_heap_replace_top(state, &newtup);
2128  return true;
2129  }
2130  return false;
2131 
2132  default:
2133  elog(ERROR, "invalid tuplesort state");
2134  return false; /* keep compiler quiet */
2135  }
2136 }
2137 
2138 /*
2139  * Fetch the next tuple in either forward or back direction.
2140  * If successful, put tuple in slot and return true; else, clear the slot
2141  * and return false.
2142  *
2143  * Caller may optionally be passed back abbreviated value (on true return
2144  * value) when abbreviation was used, which can be used to cheaply avoid
2145  * equality checks that might otherwise be required. Caller can safely make a
2146  * determination of "non-equal tuple" based on simple binary inequality. A
2147  * NULL value in leading attribute will set abbreviated value to zeroed
2148  * representation, which caller may rely on in abbreviated inequality check.
2149  *
2150  * If copy is true, the slot receives a copied tuple that will stay valid
2151  * regardless of future manipulations of the tuplesort's state. Memory is
2152  * owned by the caller. If copy is false, the slot will just receive a
2153  * pointer to a tuple held within the tuplesort, which is more efficient, but
2154  * only safe for callers that are prepared to have any subsequent manipulation
2155  * of the tuplesort's state invalidate slot contents.
2156  */
2157 bool
2158 tuplesort_gettupleslot(Tuplesortstate *state, bool forward, bool copy,
2159  TupleTableSlot *slot, Datum *abbrev)
2160 {
2161  MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
2162  SortTuple stup;
2163 
2164  if (!tuplesort_gettuple_common(state, forward, &stup))
2165  stup.tuple = NULL;
2166 
2167  MemoryContextSwitchTo(oldcontext);
2168 
2169  if (stup.tuple)
2170  {
2171  /* Record abbreviated key for caller */
2172  if (state->sortKeys->abbrev_converter && abbrev)
2173  *abbrev = stup.datum1;
2174 
2175  if (copy)
2177 
2178  ExecStoreMinimalTuple((MinimalTuple) stup.tuple, slot, copy);
2179  return true;
2180  }
2181  else
2182  {
2183  ExecClearTuple(slot);
2184  return false;
2185  }
2186 }
2187 
2188 /*
2189  * Fetch the next tuple in either forward or back direction.
2190  * Returns NULL if no more tuples. Returned tuple belongs to tuplesort memory
2191  * context, and must not be freed by caller. Caller may not rely on tuple
2192  * remaining valid after any further manipulation of tuplesort.
2193  */
2194 HeapTuple
2196 {
2197  MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
2198  SortTuple stup;
2199 
2200  if (!tuplesort_gettuple_common(state, forward, &stup))
2201  stup.tuple = NULL;
2202 
2203  MemoryContextSwitchTo(oldcontext);
2204 
2205  return stup.tuple;
2206 }
2207 
2208 /*
2209  * Fetch the next index tuple in either forward or back direction.
2210  * Returns NULL if no more tuples. Returned tuple belongs to tuplesort memory
2211  * context, and must not be freed by caller. Caller may not rely on tuple
2212  * remaining valid after any further manipulation of tuplesort.
2213  */
2214 IndexTuple
2216 {
2217  MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
2218  SortTuple stup;
2219 
2220  if (!tuplesort_gettuple_common(state, forward, &stup))
2221  stup.tuple = NULL;
2222 
2223  MemoryContextSwitchTo(oldcontext);
2224 
2225  return (IndexTuple) stup.tuple;
2226 }
2227 
2228 /*
2229  * Fetch the next Datum in either forward or back direction.
2230  * Returns false if no more datums.
2231  *
2232  * If the Datum is pass-by-ref type, the returned value is freshly palloc'd
2233  * and is now owned by the caller (this differs from similar routines for
2234  * other types of tuplesorts).
2235  *
2236  * Caller may optionally be passed back abbreviated value (on true return
2237  * value) when abbreviation was used, which can be used to cheaply avoid
2238  * equality checks that might otherwise be required. Caller can safely make a
2239  * determination of "non-equal tuple" based on simple binary inequality. A
2240  * NULL value will have a zeroed abbreviated value representation, which caller
2241  * may rely on in abbreviated inequality check.
2242  */
2243 bool
2245  Datum *val, bool *isNull, Datum *abbrev)
2246 {
2247  MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
2248  SortTuple stup;
2249 
2250  if (!tuplesort_gettuple_common(state, forward, &stup))
2251  {
2252  MemoryContextSwitchTo(oldcontext);
2253  return false;
2254  }
2255 
2256  /* Record abbreviated key for caller */
2257  if (state->sortKeys->abbrev_converter && abbrev)
2258  *abbrev = stup.datum1;
2259 
2260  if (stup.isnull1 || !state->tuples)
2261  {
2262  *val = stup.datum1;
2263  *isNull = stup.isnull1;
2264  }
2265  else
2266  {
2267  /* use stup.tuple because stup.datum1 may be an abbreviation */
2268  *val = datumCopy(PointerGetDatum(stup.tuple), false, state->datumTypeLen);
2269  *isNull = false;
2270  }
2271 
2272  MemoryContextSwitchTo(oldcontext);
2273 
2274  return true;
2275 }
2276 
2277 /*
2278  * Advance over N tuples in either forward or back direction,
2279  * without returning any data. N==0 is a no-op.
2280  * Returns true if successful, false if ran out of tuples.
2281  */
2282 bool
2283 tuplesort_skiptuples(Tuplesortstate *state, int64 ntuples, bool forward)
2284 {
2285  MemoryContext oldcontext;
2286 
2287  /*
2288  * We don't actually support backwards skip yet, because no callers need
2289  * it. The API is designed to allow for that later, though.
2290  */
2291  Assert(forward);
2292  Assert(ntuples >= 0);
2293  Assert(!WORKER(state));
2294 
2295  switch (state->status)
2296  {
2297  case TSS_SORTEDINMEM:
2298  if (state->memtupcount - state->current >= ntuples)
2299  {
2300  state->current += ntuples;
2301  return true;
2302  }
2303  state->current = state->memtupcount;
2304  state->eof_reached = true;
2305 
2306  /*
2307  * Complain if caller tries to retrieve more tuples than
2308  * originally asked for in a bounded sort. This is because
2309  * returning EOF here might be the wrong thing.
2310  */
2311  if (state->bounded && state->current >= state->bound)
2312  elog(ERROR, "retrieved too many tuples in a bounded sort");
2313 
2314  return false;
2315 
2316  case TSS_SORTEDONTAPE:
2317  case TSS_FINALMERGE:
2318 
2319  /*
2320  * We could probably optimize these cases better, but for now it's
2321  * not worth the trouble.
2322  */
2323  oldcontext = MemoryContextSwitchTo(state->sortcontext);
2324  while (ntuples-- > 0)
2325  {
2326  SortTuple stup;
2327 
2328  if (!tuplesort_gettuple_common(state, forward, &stup))
2329  {
2330  MemoryContextSwitchTo(oldcontext);
2331  return false;
2332  }
2334  }
2335  MemoryContextSwitchTo(oldcontext);
2336  return true;
2337 
2338  default:
2339  elog(ERROR, "invalid tuplesort state");
2340  return false; /* keep compiler quiet */
2341  }
2342 }
2343 
2344 /*
2345  * tuplesort_merge_order - report merge order we'll use for given memory
2346  * (note: "merge order" just means the number of input tapes in the merge).
2347  *
2348  * This is exported for use by the planner. allowedMem is in bytes.
2349  */
2350 int
2351 tuplesort_merge_order(int64 allowedMem)
2352 {
2353  int mOrder;
2354 
2355  /*
2356  * We need one tape for each merge input, plus another one for the output,
2357  * and each of these tapes needs buffer space. In addition we want
2358  * MERGE_BUFFER_SIZE workspace per input tape (but the output tape doesn't
2359  * count).
2360  *
2361  * Note: you might be thinking we need to account for the memtuples[]
2362  * array in this calculation, but we effectively treat that as part of the
2363  * MERGE_BUFFER_SIZE workspace.
2364  */
2365  mOrder = (allowedMem - TAPE_BUFFER_OVERHEAD) /
2367 
2368  /*
2369  * Even in minimum memory, use at least a MINORDER merge. On the other
2370  * hand, even when we have lots of memory, do not use more than a MAXORDER
2371  * merge. Tapes are pretty cheap, but they're not entirely free. Each
2372  * additional tape reduces the amount of memory available to build runs,
2373  * which in turn can cause the same sort to need more runs, which makes
2374  * merging slower even if it can still be done in a single pass. Also,
2375  * high order merges are quite slow due to CPU cache effects; it can be
2376  * faster to pay the I/O cost of a polyphase merge than to perform a
2377  * single merge pass across many hundreds of tapes.
2378  */
2379  mOrder = Max(mOrder, MINORDER);
2380  mOrder = Min(mOrder, MAXORDER);
2381 
2382  return mOrder;
2383 }
2384 
2385 /*
2386  * inittapes - initialize for tape sorting.
2387  *
2388  * This is called only if we have found we won't sort in memory.
2389  */
2390 static void
2392 {
2393  int maxTapes,
2394  j;
2395 
2396  Assert(!LEADER(state));
2397 
2398  if (mergeruns)
2399  {
2400  /* Compute number of tapes to use: merge order plus 1 */
2401  maxTapes = tuplesort_merge_order(state->allowedMem) + 1;
2402  }
2403  else
2404  {
2405  /* Workers can sometimes produce single run, output without merge */
2406  Assert(WORKER(state));
2407  maxTapes = MINORDER + 1;
2408  }
2409 
2410 #ifdef TRACE_SORT
2411  if (trace_sort)
2412  elog(LOG, "%d switching to external sort with %d tapes: %s",
2413  state->worker, maxTapes, pg_rusage_show(&state->ru_start));
2414 #endif
2415 
2416  /* Create the tape set and allocate the per-tape data arrays */
2417  inittapestate(state, maxTapes);
2418  state->tapeset =
2419  LogicalTapeSetCreate(maxTapes, NULL,
2420  state->shared ? &state->shared->fileset : NULL,
2421  state->worker);
2422 
2423  state->currentRun = 0;
2424 
2425  /*
2426  * Initialize variables of Algorithm D (step D1).
2427  */
2428  for (j = 0; j < maxTapes; j++)
2429  {
2430  state->tp_fib[j] = 1;
2431  state->tp_runs[j] = 0;
2432  state->tp_dummy[j] = 1;
2433  state->tp_tapenum[j] = j;
2434  }
2435  state->tp_fib[state->tapeRange] = 0;
2436  state->tp_dummy[state->tapeRange] = 0;
2437 
2438  state->Level = 1;
2439  state->destTape = 0;
2440 
2441  state->status = TSS_BUILDRUNS;
2442 }
2443 
2444 /*
2445  * inittapestate - initialize generic tape management state
2446  */
2447 static void
2449 {
2450  int64 tapeSpace;
2451 
2452  /*
2453  * Decrease availMem to reflect the space needed for tape buffers; but
2454  * don't decrease it to the point that we have no room for tuples. (That
2455  * case is only likely to occur if sorting pass-by-value Datums; in all
2456  * other scenarios the memtuples[] array is unlikely to occupy more than
2457  * half of allowedMem. In the pass-by-value case it's not important to
2458  * account for tuple space, so we don't care if LACKMEM becomes
2459  * inaccurate.)
2460  */
2461  tapeSpace = (int64) maxTapes * TAPE_BUFFER_OVERHEAD;
2462 
2463  if (tapeSpace + GetMemoryChunkSpace(state->memtuples) < state->allowedMem)
2464  USEMEM(state, tapeSpace);
2465 
2466  /*
2467  * Make sure that the temp file(s) underlying the tape set are created in
2468  * suitable temp tablespaces. For parallel sorts, this should have been
2469  * called already, but it doesn't matter if it is called a second time.
2470  */
2472 
2473  state->mergeactive = (bool *) palloc0(maxTapes * sizeof(bool));
2474  state->tp_fib = (int *) palloc0(maxTapes * sizeof(int));
2475  state->tp_runs = (int *) palloc0(maxTapes * sizeof(int));
2476  state->tp_dummy = (int *) palloc0(maxTapes * sizeof(int));
2477  state->tp_tapenum = (int *) palloc0(maxTapes * sizeof(int));
2478 
2479  /* Record # of tapes allocated (for duration of sort) */
2480  state->maxTapes = maxTapes;
2481  /* Record maximum # of tapes usable as inputs when merging */
2482  state->tapeRange = maxTapes - 1;
2483 }
2484 
2485 /*
2486  * selectnewtape -- select new tape for new initial run.
2487  *
2488  * This is called after finishing a run when we know another run
2489  * must be started. This implements steps D3, D4 of Algorithm D.
2490  */
2491 static void
2493 {
2494  int j;
2495  int a;
2496 
2497  /* Step D3: advance j (destTape) */
2498  if (state->tp_dummy[state->destTape] < state->tp_dummy[state->destTape + 1])
2499  {
2500  state->destTape++;
2501  return;
2502  }
2503  if (state->tp_dummy[state->destTape] != 0)
2504  {
2505  state->destTape = 0;
2506  return;
2507  }
2508 
2509  /* Step D4: increase level */
2510  state->Level++;
2511  a = state->tp_fib[0];
2512  for (j = 0; j < state->tapeRange; j++)
2513  {
2514  state->tp_dummy[j] = a + state->tp_fib[j + 1] - state->tp_fib[j];
2515  state->tp_fib[j] = a + state->tp_fib[j + 1];
2516  }
2517  state->destTape = 0;
2518 }
2519 
2520 /*
2521  * Initialize the slab allocation arena, for the given number of slots.
2522  */
2523 static void
2525 {
2526  if (numSlots > 0)
2527  {
2528  char *p;
2529  int i;
2530 
2531  state->slabMemoryBegin = palloc(numSlots * SLAB_SLOT_SIZE);
2532  state->slabMemoryEnd = state->slabMemoryBegin +
2533  numSlots * SLAB_SLOT_SIZE;
2534  state->slabFreeHead = (SlabSlot *) state->slabMemoryBegin;
2535  USEMEM(state, numSlots * SLAB_SLOT_SIZE);
2536 
2537  p = state->slabMemoryBegin;
2538  for (i = 0; i < numSlots - 1; i++)
2539  {
2540  ((SlabSlot *) p)->nextfree = (SlabSlot *) (p + SLAB_SLOT_SIZE);
2541  p += SLAB_SLOT_SIZE;
2542  }
2543  ((SlabSlot *) p)->nextfree = NULL;
2544  }
2545  else
2546  {
2547  state->slabMemoryBegin = state->slabMemoryEnd = NULL;
2548  state->slabFreeHead = NULL;
2549  }
2550  state->slabAllocatorUsed = true;
2551 }
2552 
2553 /*
2554  * mergeruns -- merge all the completed initial runs.
2555  *
2556  * This implements steps D5, D6 of Algorithm D. All input data has
2557  * already been written to initial runs on tape (see dumptuples).
2558  */
2559 static void
2561 {
2562  int tapenum,
2563  svTape,
2564  svRuns,
2565  svDummy;
2566  int numTapes;
2567  int numInputTapes;
2568 
2569  Assert(state->status == TSS_BUILDRUNS);
2570  Assert(state->memtupcount == 0);
2571 
2572  if (state->sortKeys != NULL && state->sortKeys->abbrev_converter != NULL)
2573  {
2574  /*
2575  * If there are multiple runs to be merged, when we go to read back
2576  * tuples from disk, abbreviated keys will not have been stored, and
2577  * we don't care to regenerate them. Disable abbreviation from this
2578  * point on.
2579  */
2580  state->sortKeys->abbrev_converter = NULL;
2582 
2583  /* Not strictly necessary, but be tidy */
2584  state->sortKeys->abbrev_abort = NULL;
2585  state->sortKeys->abbrev_full_comparator = NULL;
2586  }
2587 
2588  /*
2589  * Reset tuple memory. We've freed all the tuples that we previously
2590  * allocated. We will use the slab allocator from now on.
2591  */
2593  state->tuplecontext = NULL;
2594 
2595  /*
2596  * We no longer need a large memtuples array. (We will allocate a smaller
2597  * one for the heap later.)
2598  */
2599  FREEMEM(state, GetMemoryChunkSpace(state->memtuples));
2600  pfree(state->memtuples);
2601  state->memtuples = NULL;
2602 
2603  /*
2604  * If we had fewer runs than tapes, refund the memory that we imagined we
2605  * would need for the tape buffers of the unused tapes.
2606  *
2607  * numTapes and numInputTapes reflect the actual number of tapes we will
2608  * use. Note that the output tape's tape number is maxTapes - 1, so the
2609  * tape numbers of the used tapes are not consecutive, and you cannot just
2610  * loop from 0 to numTapes to visit all used tapes!
2611  */
2612  if (state->Level == 1)
2613  {
2614  numInputTapes = state->currentRun;
2615  numTapes = numInputTapes + 1;
2616  FREEMEM(state, (state->maxTapes - numTapes) * TAPE_BUFFER_OVERHEAD);
2617  }
2618  else
2619  {
2620  numInputTapes = state->tapeRange;
2621  numTapes = state->maxTapes;
2622  }
2623 
2624  /*
2625  * Initialize the slab allocator. We need one slab slot per input tape,
2626  * for the tuples in the heap, plus one to hold the tuple last returned
2627  * from tuplesort_gettuple. (If we're sorting pass-by-val Datums,
2628  * however, we don't need to do allocate anything.)
2629  *
2630  * From this point on, we no longer use the USEMEM()/LACKMEM() mechanism
2631  * to track memory usage of individual tuples.
2632  */
2633  if (state->tuples)
2634  init_slab_allocator(state, numInputTapes + 1);
2635  else
2636  init_slab_allocator(state, 0);
2637 
2638  /*
2639  * Allocate a new 'memtuples' array, for the heap. It will hold one tuple
2640  * from each input tape.
2641  */
2642  state->memtupsize = numInputTapes;
2643  state->memtuples = (SortTuple *) palloc(numInputTapes * sizeof(SortTuple));
2644  USEMEM(state, GetMemoryChunkSpace(state->memtuples));
2645 
2646  /*
2647  * Use all the remaining memory we have available for read buffers among
2648  * the input tapes.
2649  *
2650  * We don't try to "rebalance" the memory among tapes, when we start a new
2651  * merge phase, even if some tapes are inactive in the new phase. That
2652  * would be hard, because logtape.c doesn't know where one run ends and
2653  * another begins. When a new merge phase begins, and a tape doesn't
2654  * participate in it, its buffer nevertheless already contains tuples from
2655  * the next run on same tape, so we cannot release the buffer. That's OK
2656  * in practice, merge performance isn't that sensitive to the amount of
2657  * buffers used, and most merge phases use all or almost all tapes,
2658  * anyway.
2659  */
2660 #ifdef TRACE_SORT
2661  if (trace_sort)
2662  elog(LOG, "%d using " INT64_FORMAT " KB of memory for read buffers among %d input tapes",
2663  state->worker, state->availMem / 1024, numInputTapes);
2664 #endif
2665 
2666  state->read_buffer_size = Max(state->availMem / numInputTapes, 0);
2667  USEMEM(state, state->read_buffer_size * numInputTapes);
2668 
2669  /* End of step D2: rewind all output tapes to prepare for merging */
2670  for (tapenum = 0; tapenum < state->tapeRange; tapenum++)
2671  LogicalTapeRewindForRead(state->tapeset, tapenum, state->read_buffer_size);
2672 
2673  for (;;)
2674  {
2675  /*
2676  * At this point we know that tape[T] is empty. If there's just one
2677  * (real or dummy) run left on each input tape, then only one merge
2678  * pass remains. If we don't have to produce a materialized sorted
2679  * tape, we can stop at this point and do the final merge on-the-fly.
2680  */
2681  if (!state->randomAccess && !WORKER(state))
2682  {
2683  bool allOneRun = true;
2684 
2685  Assert(state->tp_runs[state->tapeRange] == 0);
2686  for (tapenum = 0; tapenum < state->tapeRange; tapenum++)
2687  {
2688  if (state->tp_runs[tapenum] + state->tp_dummy[tapenum] != 1)
2689  {
2690  allOneRun = false;
2691  break;
2692  }
2693  }
2694  if (allOneRun)
2695  {
2696  /* Tell logtape.c we won't be writing anymore */
2698  /* Initialize for the final merge pass */
2699  beginmerge(state);
2700  state->status = TSS_FINALMERGE;
2701  return;
2702  }
2703  }
2704 
2705  /* Step D5: merge runs onto tape[T] until tape[P] is empty */
2706  while (state->tp_runs[state->tapeRange - 1] ||
2707  state->tp_dummy[state->tapeRange - 1])
2708  {
2709  bool allDummy = true;
2710 
2711  for (tapenum = 0; tapenum < state->tapeRange; tapenum++)
2712  {
2713  if (state->tp_dummy[tapenum] == 0)
2714  {
2715  allDummy = false;
2716  break;
2717  }
2718  }
2719 
2720  if (allDummy)
2721  {
2722  state->tp_dummy[state->tapeRange]++;
2723  for (tapenum = 0; tapenum < state->tapeRange; tapenum++)
2724  state->tp_dummy[tapenum]--;
2725  }
2726  else
2727  mergeonerun(state);
2728  }
2729 
2730  /* Step D6: decrease level */
2731  if (--state->Level == 0)
2732  break;
2733  /* rewind output tape T to use as new input */
2734  LogicalTapeRewindForRead(state->tapeset, state->tp_tapenum[state->tapeRange],
2735  state->read_buffer_size);
2736  /* rewind used-up input tape P, and prepare it for write pass */
2737  LogicalTapeRewindForWrite(state->tapeset, state->tp_tapenum[state->tapeRange - 1]);
2738  state->tp_runs[state->tapeRange - 1] = 0;
2739 
2740  /*
2741  * reassign tape units per step D6; note we no longer care about A[]
2742  */
2743  svTape = state->tp_tapenum[state->tapeRange];
2744  svDummy = state->tp_dummy[state->tapeRange];
2745  svRuns = state->tp_runs[state->tapeRange];
2746  for (tapenum = state->tapeRange; tapenum > 0; tapenum--)
2747  {
2748  state->tp_tapenum[tapenum] = state->tp_tapenum[tapenum - 1];
2749  state->tp_dummy[tapenum] = state->tp_dummy[tapenum - 1];
2750  state->tp_runs[tapenum] = state->tp_runs[tapenum - 1];
2751  }
2752  state->tp_tapenum[0] = svTape;
2753  state->tp_dummy[0] = svDummy;
2754  state->tp_runs[0] = svRuns;
2755  }
2756 
2757  /*
2758  * Done. Knuth says that the result is on TAPE[1], but since we exited
2759  * the loop without performing the last iteration of step D6, we have not
2760  * rearranged the tape unit assignment, and therefore the result is on
2761  * TAPE[T]. We need to do it this way so that we can freeze the final
2762  * output tape while rewinding it. The last iteration of step D6 would be
2763  * a waste of cycles anyway...
2764  */
2765  state->result_tape = state->tp_tapenum[state->tapeRange];
2766  if (!WORKER(state))
2767  LogicalTapeFreeze(state->tapeset, state->result_tape, NULL);
2768  else
2770  state->status = TSS_SORTEDONTAPE;
2771 
2772  /* Release the read buffers of all the other tapes, by rewinding them. */
2773  for (tapenum = 0; tapenum < state->maxTapes; tapenum++)
2774  {
2775  if (tapenum != state->result_tape)
2776  LogicalTapeRewindForWrite(state->tapeset, tapenum);
2777  }
2778 }
2779 
2780 /*
2781  * Merge one run from each input tape, except ones with dummy runs.
2782  *
2783  * This is the inner loop of Algorithm D step D5. We know that the
2784  * output tape is TAPE[T].
2785  */
2786 static void
2788 {
2789  int destTape = state->tp_tapenum[state->tapeRange];
2790  int srcTape;
2791 
2792  /*
2793  * Start the merge by loading one tuple from each active source tape into
2794  * the heap. We can also decrease the input run/dummy run counts.
2795  */
2796  beginmerge(state);
2797 
2798  /*
2799  * Execute merge by repeatedly extracting lowest tuple in heap, writing it
2800  * out, and replacing it with next tuple from same tape (if there is
2801  * another one).
2802  */
2803  while (state->memtupcount > 0)
2804  {
2805  SortTuple stup;
2806 
2807  /* write the tuple to destTape */
2808  srcTape = state->memtuples[0].tupindex;
2809  WRITETUP(state, destTape, &state->memtuples[0]);
2810 
2811  /* recycle the slot of the tuple we just wrote out, for the next read */
2812  if (state->memtuples[0].tuple)
2813  RELEASE_SLAB_SLOT(state, state->memtuples[0].tuple);
2814 
2815  /*
2816  * pull next tuple from the tape, and replace the written-out tuple in
2817  * the heap with it.
2818  */
2819  if (mergereadnext(state, srcTape, &stup))
2820  {
2821  stup.tupindex = srcTape;
2822  tuplesort_heap_replace_top(state, &stup);
2823 
2824  }
2825  else
2827  }
2828 
2829  /*
2830  * When the heap empties, we're done. Write an end-of-run marker on the
2831  * output tape, and increment its count of real runs.
2832  */
2833  markrunend(state, destTape);
2834  state->tp_runs[state->tapeRange]++;
2835 
2836 #ifdef TRACE_SORT
2837  if (trace_sort)
2838  elog(LOG, "%d finished %d-way merge step: %s", state->worker,
2839  state->activeTapes, pg_rusage_show(&state->ru_start));
2840 #endif
2841 }
2842 
2843 /*
2844  * beginmerge - initialize for a merge pass
2845  *
2846  * We decrease the counts of real and dummy runs for each tape, and mark
2847  * which tapes contain active input runs in mergeactive[]. Then, fill the
2848  * merge heap with the first tuple from each active tape.
2849  */
2850 static void
2852 {
2853  int activeTapes;
2854  int tapenum;
2855  int srcTape;
2856 
2857  /* Heap should be empty here */
2858  Assert(state->memtupcount == 0);
2859 
2860  /* Adjust run counts and mark the active tapes */
2861  memset(state->mergeactive, 0,
2862  state->maxTapes * sizeof(*state->mergeactive));
2863  activeTapes = 0;
2864  for (tapenum = 0; tapenum < state->tapeRange; tapenum++)
2865  {
2866  if (state->tp_dummy[tapenum] > 0)
2867  state->tp_dummy[tapenum]--;
2868  else
2869  {
2870  Assert(state->tp_runs[tapenum] > 0);
2871  state->tp_runs[tapenum]--;
2872  srcTape = state->tp_tapenum[tapenum];
2873  state->mergeactive[srcTape] = true;
2874  activeTapes++;
2875  }
2876  }
2877  Assert(activeTapes > 0);
2878  state->activeTapes = activeTapes;
2879 
2880  /* Load the merge heap with the first tuple from each input tape */
2881  for (srcTape = 0; srcTape < state->maxTapes; srcTape++)
2882  {
2883  SortTuple tup;
2884 
2885  if (mergereadnext(state, srcTape, &tup))
2886  {
2887  tup.tupindex = srcTape;
2888  tuplesort_heap_insert(state, &tup);
2889  }
2890  }
2891 }
2892 
2893 /*
2894  * mergereadnext - read next tuple from one merge input tape
2895  *
2896  * Returns false on EOF.
2897  */
2898 static bool
2900 {
2901  unsigned int tuplen;
2902 
2903  if (!state->mergeactive[srcTape])
2904  return false; /* tape's run is already exhausted */
2905 
2906  /* read next tuple, if any */
2907  if ((tuplen = getlen(state, srcTape, true)) == 0)
2908  {
2909  state->mergeactive[srcTape] = false;
2910  return false;
2911  }
2912  READTUP(state, stup, srcTape, tuplen);
2913 
2914  return true;
2915 }
2916 
2917 /*
2918  * dumptuples - remove tuples from memtuples and write initial run to tape
2919  *
2920  * When alltuples = true, dump everything currently in memory. (This case is
2921  * only used at end of input data.)
2922  */
2923 static void
2925 {
2926  int memtupwrite;
2927  int i;
2928 
2929  /*
2930  * Nothing to do if we still fit in available memory and have array slots,
2931  * unless this is the final call during initial run generation.
2932  */
2933  if (state->memtupcount < state->memtupsize && !LACKMEM(state) &&
2934  !alltuples)
2935  return;
2936 
2937  /*
2938  * Final call might require no sorting, in rare cases where we just so
2939  * happen to have previously LACKMEM()'d at the point where exactly all
2940  * remaining tuples are loaded into memory, just before input was
2941  * exhausted.
2942  *
2943  * In general, short final runs are quite possible. Rather than allowing
2944  * a special case where there was a superfluous selectnewtape() call (i.e.
2945  * a call with no subsequent run actually written to destTape), we prefer
2946  * to write out a 0 tuple run.
2947  *
2948  * mergereadnext() is prepared for 0 tuple runs, and will reliably mark
2949  * the tape inactive for the merge when called from beginmerge(). This
2950  * case is therefore similar to the case where mergeonerun() finds a dummy
2951  * run for the tape, and so doesn't need to merge a run from the tape (or
2952  * conceptually "merges" the dummy run, if you prefer). According to
2953  * Knuth, Algorithm D "isn't strictly optimal" in its method of
2954  * distribution and dummy run assignment; this edge case seems very
2955  * unlikely to make that appreciably worse.
2956  */
2957  Assert(state->status == TSS_BUILDRUNS);
2958 
2959  /*
2960  * It seems unlikely that this limit will ever be exceeded, but take no
2961  * chances
2962  */
2963  if (state->currentRun == INT_MAX)
2964  ereport(ERROR,
2965  (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
2966  errmsg("cannot have more than %d runs for an external sort",
2967  INT_MAX)));
2968 
2969  state->currentRun++;
2970 
2971 #ifdef TRACE_SORT
2972  if (trace_sort)
2973  elog(LOG, "%d starting quicksort of run %d: %s",
2974  state->worker, state->currentRun,
2975  pg_rusage_show(&state->ru_start));
2976 #endif
2977 
2978  /*
2979  * Sort all tuples accumulated within the allowed amount of memory for
2980  * this run using quicksort
2981  */
2982  tuplesort_sort_memtuples(state);
2983 
2984 #ifdef TRACE_SORT
2985  if (trace_sort)
2986  elog(LOG, "%d finished quicksort of run %d: %s",
2987  state->worker, state->currentRun,
2988  pg_rusage_show(&state->ru_start));
2989 #endif
2990 
2991  memtupwrite = state->memtupcount;
2992  for (i = 0; i < memtupwrite; i++)
2993  {
2994  WRITETUP(state, state->tp_tapenum[state->destTape],
2995  &state->memtuples[i]);
2996  state->memtupcount--;
2997  }
2998 
2999  /*
3000  * Reset tuple memory. We've freed all of the tuples that we previously
3001  * allocated. It's important to avoid fragmentation when there is a stark
3002  * change in the sizes of incoming tuples. Fragmentation due to
3003  * AllocSetFree's bucketing by size class might be particularly bad if
3004  * this step wasn't taken.
3005  */
3007 
3008  markrunend(state, state->tp_tapenum[state->destTape]);
3009  state->tp_runs[state->destTape]++;
3010  state->tp_dummy[state->destTape]--; /* per Alg D step D2 */
3011 
3012 #ifdef TRACE_SORT
3013  if (trace_sort)
3014  elog(LOG, "%d finished writing run %d to tape %d: %s",
3015  state->worker, state->currentRun, state->destTape,
3016  pg_rusage_show(&state->ru_start));
3017 #endif
3018 
3019  if (!alltuples)
3020  selectnewtape(state);
3021 }
3022 
3023 /*
3024  * tuplesort_rescan - rewind and replay the scan
3025  */
3026 void
3028 {
3029  MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
3030 
3031  Assert(state->randomAccess);
3032 
3033  switch (state->status)
3034  {
3035  case TSS_SORTEDINMEM:
3036  state->current = 0;
3037  state->eof_reached = false;
3038  state->markpos_offset = 0;
3039  state->markpos_eof = false;
3040  break;
3041  case TSS_SORTEDONTAPE:
3043  state->result_tape,
3044  0);
3045  state->eof_reached = false;
3046  state->markpos_block = 0L;
3047  state->markpos_offset = 0;
3048  state->markpos_eof = false;
3049  break;
3050  default:
3051  elog(ERROR, "invalid tuplesort state");
3052  break;
3053  }
3054 
3055  MemoryContextSwitchTo(oldcontext);
3056 }
3057 
3058 /*
3059  * tuplesort_markpos - saves current position in the merged sort file
3060  */
3061 void
3063 {
3064  MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
3065 
3066  Assert(state->randomAccess);
3067 
3068  switch (state->status)
3069  {
3070  case TSS_SORTEDINMEM:
3071  state->markpos_offset = state->current;
3072  state->markpos_eof = state->eof_reached;
3073  break;
3074  case TSS_SORTEDONTAPE:
3075  LogicalTapeTell(state->tapeset,
3076  state->result_tape,
3077  &state->markpos_block,
3078  &state->markpos_offset);
3079  state->markpos_eof = state->eof_reached;
3080  break;
3081  default:
3082  elog(ERROR, "invalid tuplesort state");
3083  break;
3084  }
3085 
3086  MemoryContextSwitchTo(oldcontext);
3087 }
3088 
3089 /*
3090  * tuplesort_restorepos - restores current position in merged sort file to
3091  * last saved position
3092  */
3093 void
3095 {
3096  MemoryContext oldcontext = MemoryContextSwitchTo(state->sortcontext);
3097 
3098  Assert(state->randomAccess);
3099 
3100  switch (state->status)
3101  {
3102  case TSS_SORTEDINMEM:
3103  state->current = state->markpos_offset;
3104  state->eof_reached = state->markpos_eof;
3105  break;
3106  case TSS_SORTEDONTAPE:
3107  LogicalTapeSeek(state->tapeset,
3108  state->result_tape,
3109  state->markpos_block,
3110  state->markpos_offset);
3111  state->eof_reached = state->markpos_eof;
3112  break;
3113  default:
3114  elog(ERROR, "invalid tuplesort state");
3115  break;
3116  }
3117 
3118  MemoryContextSwitchTo(oldcontext);
3119 }
3120 
3121 /*
3122  * tuplesort_get_stats - extract summary statistics
3123  *
3124  * This can be called after tuplesort_performsort() finishes to obtain
3125  * printable summary information about how the sort was performed.
3126  */
3127 void
3129  TuplesortInstrumentation *stats)
3130 {
3131  /*
3132  * Note: it might seem we should provide both memory and disk usage for a
3133  * disk-based sort. However, the current code doesn't track memory space
3134  * accurately once we have begun to return tuples to the caller (since we
3135  * don't account for pfree's the caller is expected to do), so we cannot
3136  * rely on availMem in a disk sort. This does not seem worth the overhead
3137  * to fix. Is it worth creating an API for the memory context code to
3138  * tell us how much is actually used in sortcontext?
3139  */
3140  if (state->tapeset)
3141  {
3143  stats->spaceUsed = LogicalTapeSetBlocks(state->tapeset) * (BLCKSZ / 1024);
3144  }
3145  else
3146  {
3148  stats->spaceUsed = (state->allowedMem - state->availMem + 1023) / 1024;
3149  }
3150 
3151  switch (state->status)
3152  {
3153  case TSS_SORTEDINMEM:
3154  if (state->boundUsed)
3156  else
3158  break;
3159  case TSS_SORTEDONTAPE:
3161  break;
3162  case TSS_FINALMERGE:
3164  break;
3165  default:
3167  break;
3168  }
3169 }
3170 
3171 /*
3172  * Convert TuplesortMethod to a string.
3173  */
3174 const char *
3176 {
3177  switch (m)
3178  {
3180  return "still in progress";
3182  return "top-N heapsort";
3183  case SORT_TYPE_QUICKSORT:
3184  return "quicksort";
3186  return "external sort";
3188  return "external merge";
3189  }
3190 
3191  return "unknown";
3192 }
3193 
3194 /*
3195  * Convert TuplesortSpaceType to a string.
3196  */
3197 const char *
3199 {
3201  return t == SORT_SPACE_TYPE_DISK ? "Disk" : "Memory";
3202 }
3203 
3204 
3205 /*
3206  * Heap manipulation routines, per Knuth's Algorithm 5.2.3H.
3207  */
3208 
3209 /*
3210  * Convert the existing unordered array of SortTuples to a bounded heap,
3211  * discarding all but the smallest "state->bound" tuples.
3212  *
3213  * When working with a bounded heap, we want to keep the largest entry
3214  * at the root (array entry zero), instead of the smallest as in the normal
3215  * sort case. This allows us to discard the largest entry cheaply.
3216  * Therefore, we temporarily reverse the sort direction.
3217  */
3218 static void
3220 {
3221  int tupcount = state->memtupcount;
3222  int i;
3223 
3224  Assert(state->status == TSS_INITIAL);
3225  Assert(state->bounded);
3226  Assert(tupcount >= state->bound);
3227  Assert(SERIAL(state));
3228 
3229  /* Reverse sort direction so largest entry will be at root */
3230  reversedirection(state);
3231 
3232  state->memtupcount = 0; /* make the heap empty */
3233  for (i = 0; i < tupcount; i++)
3234  {
3235  if (state->memtupcount < state->bound)
3236  {
3237  /* Insert next tuple into heap */
3238  /* Must copy source tuple to avoid possible overwrite */
3239  SortTuple stup = state->memtuples[i];
3240 
3241  tuplesort_heap_insert(state, &stup);
3242  }
3243  else
3244  {
3245  /*
3246  * The heap is full. Replace the largest entry with the new
3247  * tuple, or just discard it, if it's larger than anything already
3248  * in the heap.
3249  */
3250  if (COMPARETUP(state, &state->memtuples[i], &state->memtuples[0]) <= 0)
3251  {
3252  free_sort_tuple(state, &state->memtuples[i]);
3254  }
3255  else
3256  tuplesort_heap_replace_top(state, &state->memtuples[i]);
3257  }
3258  }
3259 
3260  Assert(state->memtupcount == state->bound);
3261  state->status = TSS_BOUNDED;
3262 }
3263 
3264 /*
3265  * Convert the bounded heap to a properly-sorted array
3266  */
3267 static void
3269 {
3270  int tupcount = state->memtupcount;
3271 
3272  Assert(state->status == TSS_BOUNDED);
3273  Assert(state->bounded);
3274  Assert(tupcount == state->bound);
3275  Assert(SERIAL(state));
3276 
3277  /*
3278  * We can unheapify in place because each delete-top call will remove the
3279  * largest entry, which we can promptly store in the newly freed slot at
3280  * the end. Once we're down to a single-entry heap, we're done.
3281  */
3282  while (state->memtupcount > 1)
3283  {
3284  SortTuple stup = state->memtuples[0];
3285 
3286  /* this sifts-up the next-largest entry and decreases memtupcount */
3288  state->memtuples[state->memtupcount] = stup;
3289  }
3290  state->memtupcount = tupcount;
3291 
3292  /*
3293  * Reverse sort direction back to the original state. This is not
3294  * actually necessary but seems like a good idea for tidiness.
3295  */
3296  reversedirection(state);
3297 
3298  state->status = TSS_SORTEDINMEM;
3299  state->boundUsed = true;
3300 }
3301 
3302 /*
3303  * Sort all memtuples using specialized qsort() routines.
3304  *
3305  * Quicksort is used for small in-memory sorts, and external sort runs.
3306  */
3307 static void
3309 {
3310  Assert(!LEADER(state));
3311 
3312  if (state->memtupcount > 1)
3313  {
3314  /* Can we use the single-key sort function? */
3315  if (state->onlyKey != NULL)
3316  qsort_ssup(state->memtuples, state->memtupcount,
3317  state->onlyKey);
3318  else
3319  qsort_tuple(state->memtuples,
3320  state->memtupcount,
3321  state->comparetup,
3322  state);
3323  }
3324 }
3325 
3326 /*
3327  * Insert a new tuple into an empty or existing heap, maintaining the
3328  * heap invariant. Caller is responsible for ensuring there's room.
3329  *
3330  * Note: For some callers, tuple points to a memtuples[] entry above the
3331  * end of the heap. This is safe as long as it's not immediately adjacent
3332  * to the end of the heap (ie, in the [memtupcount] array entry) --- if it
3333  * is, it might get overwritten before being moved into the heap!
3334  */
3335 static void
3337 {
3338  SortTuple *memtuples;
3339  int j;
3340 
3341  memtuples = state->memtuples;
3342  Assert(state->memtupcount < state->memtupsize);
3343 
3345 
3346  /*
3347  * Sift-up the new entry, per Knuth 5.2.3 exercise 16. Note that Knuth is
3348  * using 1-based array indexes, not 0-based.
3349  */
3350  j = state->memtupcount++;
3351  while (j > 0)
3352  {
3353  int i = (j - 1) >> 1;
3354 
3355  if (COMPARETUP(state, tuple, &memtuples[i]) >= 0)
3356  break;
3357  memtuples[j] = memtuples[i];
3358  j = i;
3359  }
3360  memtuples[j] = *tuple;
3361 }
3362 
3363 /*
3364  * Remove the tuple at state->memtuples[0] from the heap. Decrement
3365  * memtupcount, and sift up to maintain the heap invariant.
3366  *
3367  * The caller has already free'd the tuple the top node points to,
3368  * if necessary.
3369  */
3370 static void
3372 {
3373  SortTuple *memtuples = state->memtuples;
3374  SortTuple *tuple;
3375 
3376  if (--state->memtupcount <= 0)
3377  return;
3378 
3379  /*
3380  * Remove the last tuple in the heap, and re-insert it, by replacing the
3381  * current top node with it.
3382  */
3383  tuple = &memtuples[state->memtupcount];
3384  tuplesort_heap_replace_top(state, tuple);
3385 }
3386 
3387 /*
3388  * Replace the tuple at state->memtuples[0] with a new tuple. Sift up to
3389  * maintain the heap invariant.
3390  *
3391  * This corresponds to Knuth's "sift-up" algorithm (Algorithm 5.2.3H,
3392  * Heapsort, steps H3-H8).
3393  */
3394 static void
3396 {
3397  SortTuple *memtuples = state->memtuples;
3398  unsigned int i,
3399  n;
3400 
3401  Assert(state->memtupcount >= 1);
3402 
3404 
3405  /*
3406  * state->memtupcount is "int", but we use "unsigned int" for i, j, n.
3407  * This prevents overflow in the "2 * i + 1" calculation, since at the top
3408  * of the loop we must have i < n <= INT_MAX <= UINT_MAX/2.
3409  */
3410  n = state->memtupcount;
3411  i = 0; /* i is where the "hole" is */
3412  for (;;)
3413  {
3414  unsigned int j = 2 * i + 1;
3415 
3416  if (j >= n)
3417  break;
3418  if (j + 1 < n &&
3419  COMPARETUP(state, &memtuples[j], &memtuples[j + 1]) > 0)
3420  j++;
3421  if (COMPARETUP(state, tuple, &memtuples[j]) <= 0)
3422  break;
3423  memtuples[i] = memtuples[j];
3424  i = j;
3425  }
3426  memtuples[i] = *tuple;
3427 }
3428 
3429 /*
3430  * Function to reverse the sort direction from its current state
3431  *
3432  * It is not safe to call this when performing hash tuplesorts
3433  */
3434 static void
3436 {
3437  SortSupport sortKey = state->sortKeys;
3438  int nkey;
3439 
3440  for (nkey = 0; nkey < state->nKeys; nkey++, sortKey++)
3441  {
3442  sortKey->ssup_reverse = !sortKey->ssup_reverse;
3443  sortKey->ssup_nulls_first = !sortKey->ssup_nulls_first;
3444  }
3445 }
3446 
3447 
3448 /*
3449  * Tape interface routines
3450  */
3451 
3452 static unsigned int
3453 getlen(Tuplesortstate *state, int tapenum, bool eofOK)
3454 {
3455  unsigned int len;
3456 
3457  if (LogicalTapeRead(state->tapeset, tapenum,
3458  &len, sizeof(len)) != sizeof(len))
3459  elog(ERROR, "unexpected end of tape");
3460  if (len == 0 && !eofOK)
3461  elog(ERROR, "unexpected end of data");
3462  return len;
3463 }
3464 
3465 static void
3467 {
3468  unsigned int len = 0;
3469 
3470  LogicalTapeWrite(state->tapeset, tapenum, (void *) &len, sizeof(len));
3471 }
3472 
3473 /*
3474  * Get memory for tuple from within READTUP() routine.
3475  *
3476  * We use next free slot from the slab allocator, or palloc() if the tuple
3477  * is too large for that.
3478  */
3479 static void *
3481 {
3482  SlabSlot *buf;
3483 
3484  /*
3485  * We pre-allocate enough slots in the slab arena that we should never run
3486  * out.
3487  */
3488  Assert(state->slabFreeHead);
3489 
3490  if (tuplen > SLAB_SLOT_SIZE || !state->slabFreeHead)
3491  return MemoryContextAlloc(state->sortcontext, tuplen);
3492  else
3493  {
3494  buf = state->slabFreeHead;
3495  /* Reuse this slot */
3496  state->slabFreeHead = buf->nextfree;
3497 
3498  return buf;
3499  }
3500 }
3501 
3502 
3503 /*
3504  * Routines specialized for HeapTuple (actually MinimalTuple) case
3505  */
3506 
3507 static int
3509 {
3510  SortSupport sortKey = state->sortKeys;
3511  HeapTupleData ltup;
3512  HeapTupleData rtup;
3513  TupleDesc tupDesc;
3514  int nkey;
3515  int32 compare;
3516  AttrNumber attno;
3517  Datum datum1,
3518  datum2;
3519  bool isnull1,
3520  isnull2;
3521 
3522 
3523  /* Compare the leading sort key */
3524  compare = ApplySortComparator(a->datum1, a->isnull1,
3525  b->datum1, b->isnull1,
3526  sortKey);
3527  if (compare != 0)
3528  return compare;
3529 
3530  /* Compare additional sort keys */
3531  ltup.t_len = ((MinimalTuple) a->tuple)->t_len + MINIMAL_TUPLE_OFFSET;
3532  ltup.t_data = (HeapTupleHeader) ((char *) a->tuple - MINIMAL_TUPLE_OFFSET);
3533  rtup.t_len = ((MinimalTuple) b->tuple)->t_len + MINIMAL_TUPLE_OFFSET;
3534  rtup.t_data = (HeapTupleHeader) ((char *) b->tuple - MINIMAL_TUPLE_OFFSET);
3535  tupDesc = state->tupDesc;
3536 
3537  if (sortKey->abbrev_converter)
3538  {
3539  attno = sortKey->ssup_attno;
3540 
3541  datum1 = heap_getattr(&ltup, attno, tupDesc, &isnull1);
3542  datum2 = heap_getattr(&rtup, attno, tupDesc, &isnull2);
3543 
3544  compare = ApplySortAbbrevFullComparator(datum1, isnull1,
3545  datum2, isnull2,
3546  sortKey);
3547  if (compare != 0)
3548  return compare;
3549  }
3550 
3551  sortKey++;
3552  for (nkey = 1; nkey < state->nKeys; nkey++, sortKey++)
3553  {
3554  attno = sortKey->ssup_attno;
3555 
3556  datum1 = heap_getattr(&ltup, attno, tupDesc, &isnull1);
3557  datum2 = heap_getattr(&rtup, attno, tupDesc, &isnull2);
3558 
3559  compare = ApplySortComparator(datum1, isnull1,
3560  datum2, isnull2,
3561  sortKey);
3562  if (compare != 0)
3563  return compare;
3564  }
3565 
3566  return 0;
3567 }
3568 
3569 static void
3571 {
3572  /*
3573  * We expect the passed "tup" to be a TupleTableSlot, and form a
3574  * MinimalTuple using the exported interface for that.
3575  */
3576  TupleTableSlot *slot = (TupleTableSlot *) tup;
3577  Datum original;
3578  MinimalTuple tuple;
3579  HeapTupleData htup;
3580  MemoryContext oldcontext = MemoryContextSwitchTo(state->tuplecontext);
3581 
3582  /* copy the tuple into sort storage */
3583  tuple = ExecCopySlotMinimalTuple(slot);
3584  stup->tuple = (void *) tuple;
3585  USEMEM(state, GetMemoryChunkSpace(tuple));
3586  /* set up first-column key value */
3587  htup.t_len = tuple->t_len + MINIMAL_TUPLE_OFFSET;
3588  htup.t_data = (HeapTupleHeader) ((char *) tuple - MINIMAL_TUPLE_OFFSET);
3589  original = heap_getattr(&htup,
3590  state->sortKeys[0].ssup_attno,
3591  state->tupDesc,
3592  &stup->isnull1);
3593 
3594  MemoryContextSwitchTo(oldcontext);
3595 
3596  if (!state->sortKeys->abbrev_converter || stup->isnull1)
3597  {
3598  /*
3599  * Store ordinary Datum representation, or NULL value. If there is a
3600  * converter it won't expect NULL values, and cost model is not
3601  * required to account for NULL, so in that case we avoid calling
3602  * converter and just set datum1 to zeroed representation (to be
3603  * consistent, and to support cheap inequality tests for NULL
3604  * abbreviated keys).
3605  */
3606  stup->datum1 = original;
3607  }
3608  else if (!consider_abort_common(state))
3609  {
3610  /* Store abbreviated key representation */
3611  stup->datum1 = state->sortKeys->abbrev_converter(original,
3612  state->sortKeys);
3613  }
3614  else
3615  {
3616  /* Abort abbreviation */
3617  int i;
3618 
3619  stup->datum1 = original;
3620 
3621  /*
3622  * Set state to be consistent with never trying abbreviation.
3623  *
3624  * Alter datum1 representation in already-copied tuples, so as to
3625  * ensure a consistent representation (current tuple was just
3626  * handled). It does not matter if some dumped tuples are already
3627  * sorted on tape, since serialized tuples lack abbreviated keys
3628  * (TSS_BUILDRUNS state prevents control reaching here in any case).
3629  */
3630  for (i = 0; i < state->memtupcount; i++)
3631  {
3632  SortTuple *mtup = &state->memtuples[i];
3633 
3634  htup.t_len = ((MinimalTuple) mtup->tuple)->t_len +
3636  htup.t_data = (HeapTupleHeader) ((char *) mtup->tuple -
3638 
3639  mtup->datum1 = heap_getattr(&htup,
3640  state->sortKeys[0].ssup_attno,
3641  state->tupDesc,
3642  &mtup->isnull1);
3643  }
3644  }
3645 }
3646 
3647 static void
3649 {
3650  MinimalTuple tuple = (MinimalTuple) stup->tuple;
3651 
3652  /* the part of the MinimalTuple we'll write: */
3653  char *tupbody = (char *) tuple + MINIMAL_TUPLE_DATA_OFFSET;
3654  unsigned int tupbodylen = tuple->t_len - MINIMAL_TUPLE_DATA_OFFSET;
3655 
3656  /* total on-disk footprint: */
3657  unsigned int tuplen = tupbodylen + sizeof(int);
3658 
3659  LogicalTapeWrite(state->tapeset, tapenum,
3660  (void *) &tuplen, sizeof(tuplen));
3661  LogicalTapeWrite(state->tapeset, tapenum,
3662  (void *) tupbody, tupbodylen);
3663  if (state->randomAccess) /* need trailing length word? */
3664  LogicalTapeWrite(state->tapeset, tapenum,
3665  (void *) &tuplen, sizeof(tuplen));
3666 
3667  if (!state->slabAllocatorUsed)
3668  {
3669  FREEMEM(state, GetMemoryChunkSpace(tuple));
3670  heap_free_minimal_tuple(tuple);
3671  }
3672 }
3673 
3674 static void
3676  int tapenum, unsigned int len)
3677 {
3678  unsigned int tupbodylen = len - sizeof(int);
3679  unsigned int tuplen = tupbodylen + MINIMAL_TUPLE_DATA_OFFSET;
3680  MinimalTuple tuple = (MinimalTuple) readtup_alloc(state, tuplen);
3681  char *tupbody = (char *) tuple + MINIMAL_TUPLE_DATA_OFFSET;
3682  HeapTupleData htup;
3683 
3684  /* read in the tuple proper */
3685  tuple->t_len = tuplen;
3686  LogicalTapeReadExact(state->tapeset, tapenum,
3687  tupbody, tupbodylen);
3688  if (state->randomAccess) /* need trailing length word? */
3689  LogicalTapeReadExact(state->tapeset, tapenum,
3690  &tuplen, sizeof(tuplen));
3691  stup->tuple = (void *) tuple;
3692  /* set up first-column key value */
3693  htup.t_len = tuple->t_len + MINIMAL_TUPLE_OFFSET;
3694  htup.t_data = (HeapTupleHeader) ((char *) tuple - MINIMAL_TUPLE_OFFSET);
3695  stup->datum1 = heap_getattr(&htup,
3696  state->sortKeys[0].ssup_attno,
3697  state->tupDesc,
3698  &stup->isnull1);
3699 }
3700 
3701 /*
3702  * Routines specialized for the CLUSTER case (HeapTuple data, with
3703  * comparisons per a btree index definition)
3704  */
3705 
3706 static int
3709 {
3710  SortSupport sortKey = state->sortKeys;
3711  HeapTuple ltup;
3712  HeapTuple rtup;
3713  TupleDesc tupDesc;
3714  int nkey;
3715  int32 compare;
3716  Datum datum1,
3717  datum2;
3718  bool isnull1,
3719  isnull2;
3720  AttrNumber leading = state->indexInfo->ii_KeyAttrNumbers[0];
3721 
3722  /* Be prepared to compare additional sort keys */
3723  ltup = (HeapTuple) a->tuple;
3724  rtup = (HeapTuple) b->tuple;
3725  tupDesc = state->tupDesc;
3726 
3727  /* Compare the leading sort key, if it's simple */
3728  if (leading != 0)
3729  {
3730  compare = ApplySortComparator(a->datum1, a->isnull1,
3731  b->datum1, b->isnull1,
3732  sortKey);
3733  if (compare != 0)
3734  return compare;
3735 
3736  if (sortKey->abbrev_converter)
3737  {
3738  datum1 = heap_getattr(ltup, leading, tupDesc, &isnull1);
3739  datum2 = heap_getattr(rtup, leading, tupDesc, &isnull2);
3740 
3741  compare = ApplySortAbbrevFullComparator(datum1, isnull1,
3742  datum2, isnull2,
3743  sortKey);
3744  }
3745  if (compare != 0 || state->nKeys == 1)
3746  return compare;
3747  /* Compare additional columns the hard way */
3748  sortKey++;
3749  nkey = 1;
3750  }
3751  else
3752  {
3753  /* Must compare all keys the hard way */
3754  nkey = 0;
3755  }
3756 
3757  if (state->indexInfo->ii_Expressions == NULL)
3758  {
3759  /* If not expression index, just compare the proper heap attrs */
3760 
3761  for (; nkey < state->nKeys; nkey++, sortKey++)
3762  {
3763  AttrNumber attno = state->indexInfo->ii_KeyAttrNumbers[nkey];
3764 
3765  datum1 = heap_getattr(ltup, attno, tupDesc, &isnull1);
3766  datum2 = heap_getattr(rtup, attno, tupDesc, &isnull2);
3767 
3768  compare = ApplySortComparator(datum1, isnull1,
3769  datum2, isnull2,
3770  sortKey);
3771  if (compare != 0)
3772  return compare;
3773  }
3774  }
3775  else
3776  {
3777  /*
3778  * In the expression index case, compute the whole index tuple and
3779  * then compare values. It would perhaps be faster to compute only as
3780  * many columns as we need to compare, but that would require
3781  * duplicating all the logic in FormIndexDatum.
3782  */
3783  Datum l_index_values[INDEX_MAX_KEYS];
3784  bool l_index_isnull[INDEX_MAX_KEYS];
3785  Datum r_index_values[INDEX_MAX_KEYS];
3786  bool r_index_isnull[INDEX_MAX_KEYS];
3787  TupleTableSlot *ecxt_scantuple;
3788 
3789  /* Reset context each time to prevent memory leakage */
3791 
3792  ecxt_scantuple = GetPerTupleExprContext(state->estate)->ecxt_scantuple;
3793 
3794  ExecStoreTuple(ltup, ecxt_scantuple, InvalidBuffer, false);
3795  FormIndexDatum(state->indexInfo, ecxt_scantuple, state->estate,
3796  l_index_values, l_index_isnull);
3797 
3798  ExecStoreTuple(rtup, ecxt_scantuple, InvalidBuffer, false);
3799  FormIndexDatum(state->indexInfo, ecxt_scantuple, state->estate,
3800  r_index_values, r_index_isnull);
3801 
3802  for (; nkey < state->nKeys; nkey++, sortKey++)
3803  {
3804  compare = ApplySortComparator(l_index_values[nkey],
3805  l_index_isnull[nkey],
3806  r_index_values[nkey],
3807  r_index_isnull[nkey],
3808  sortKey);
3809  if (compare != 0)
3810  return compare;
3811  }
3812  }
3813 
3814  return 0;
3815 }
3816 
3817 static void
3819 {
3820  HeapTuple tuple = (HeapTuple) tup;
3821  Datum original;
3822  MemoryContext oldcontext = MemoryContextSwitchTo(state->tuplecontext);
3823 
3824  /* copy the tuple into sort storage */
3825  tuple = heap_copytuple(tuple);
3826  stup->tuple = (void *) tuple;
3827  USEMEM(state, GetMemoryChunkSpace(tuple));
3828 
3829  MemoryContextSwitchTo(oldcontext);
3830 
3831  /*
3832  * set up first-column key value, and potentially abbreviate, if it's a
3833  * simple column
3834  */
3835  if (state->indexInfo->ii_KeyAttrNumbers[0] == 0)
3836  return;
3837 
3838  original = heap_getattr(tuple,
3839  state->indexInfo->ii_KeyAttrNumbers[0],
3840  state->tupDesc,
3841  &stup->isnull1);
3842 
3843  if (!state->sortKeys->abbrev_converter || stup->isnull1)
3844  {
3845  /*
3846  * Store ordinary Datum representation, or NULL value. If there is a
3847  * converter it won't expect NULL values, and cost model is not
3848  * required to account for NULL, so in that case we avoid calling
3849  * converter and just set datum1 to zeroed representation (to be
3850  * consistent, and to support cheap inequality tests for NULL
3851  * abbreviated keys).
3852  */
3853  stup->datum1 = original;
3854  }
3855  else if (!consider_abort_common(state))
3856  {
3857  /* Store abbreviated key representation */
3858  stup->datum1 = state->sortKeys->abbrev_converter(original,
3859  state->sortKeys);
3860  }
3861  else
3862  {
3863  /* Abort abbreviation */
3864  int i;
3865 
3866  stup->datum1 = original;
3867 
3868  /*
3869  * Set state to be consistent with never trying abbreviation.
3870  *
3871  * Alter datum1 representation in already-copied tuples, so as to
3872  * ensure a consistent representation (current tuple was just
3873  * handled). It does not matter if some dumped tuples are already
3874  * sorted on tape, since serialized tuples lack abbreviated keys
3875  * (TSS_BUILDRUNS state prevents control reaching here in any case).
3876  */
3877  for (i = 0; i < state->memtupcount; i++)
3878  {
3879  SortTuple *mtup = &state->memtuples[i];
3880 
3881  tuple = (HeapTuple) mtup->tuple;
3882  mtup->datum1 = heap_getattr(tuple,
3883  state->indexInfo->ii_KeyAttrNumbers[0],
3884  state->tupDesc,
3885  &mtup->isnull1);
3886  }
3887  }
3888 }
3889 
3890 static void
3892 {
3893  HeapTuple tuple = (HeapTuple) stup->tuple;
3894  unsigned int tuplen = tuple->t_len + sizeof(ItemPointerData) + sizeof(int);
3895 
3896  /* We need to store t_self, but not other fields of HeapTupleData */
3897  LogicalTapeWrite(state->tapeset, tapenum,
3898  &tuplen, sizeof(tuplen));
3899  LogicalTapeWrite(state->tapeset, tapenum,
3900  &tuple->t_self, sizeof(ItemPointerData));
3901  LogicalTapeWrite(state->tapeset, tapenum,
3902  tuple->t_data, tuple->t_len);
3903  if (state->randomAccess) /* need trailing length word? */
3904  LogicalTapeWrite(state->tapeset, tapenum,
3905  &tuplen, sizeof(tuplen));
3906 
3907  if (!state->slabAllocatorUsed)
3908  {
3909  FREEMEM(state, GetMemoryChunkSpace(tuple));
3910  heap_freetuple(tuple);
3911  }
3912 }
3913 
3914 static void
3916  int tapenum, unsigned int tuplen)
3917 {
3918  unsigned int t_len = tuplen - sizeof(ItemPointerData) - sizeof(int);
3919  HeapTuple tuple = (HeapTuple) readtup_alloc(state,
3920  t_len + HEAPTUPLESIZE);
3921 
3922  /* Reconstruct the HeapTupleData header */
3923  tuple->t_data = (HeapTupleHeader) ((char *) tuple + HEAPTUPLESIZE);
3924  tuple->t_len = t_len;
3925  LogicalTapeReadExact(state->tapeset, tapenum,
3926  &tuple->t_self, sizeof(ItemPointerData));
3927  /* We don't currently bother to reconstruct t_tableOid */
3928  tuple->t_tableOid = InvalidOid;
3929  /* Read in the tuple body */
3930  LogicalTapeReadExact(state->tapeset, tapenum,
3931  tuple->t_data, tuple->t_len);
3932  if (state->randomAccess) /* need trailing length word? */
3933  LogicalTapeReadExact(state->tapeset, tapenum,
3934  &tuplen, sizeof(tuplen));
3935  stup->tuple = (void *) tuple;
3936  /* set up first-column key value, if it's a simple column */
3937  if (state->indexInfo->ii_KeyAttrNumbers[0] != 0)
3938  stup->datum1 = heap_getattr(tuple,
3939  state->indexInfo->ii_KeyAttrNumbers[0],
3940  state->tupDesc,
3941  &stup->isnull1);
3942 }
3943 
3944 /*
3945  * Routines specialized for IndexTuple case
3946  *
3947  * The btree and hash cases require separate comparison functions, but the
3948  * IndexTuple representation is the same so the copy/write/read support
3949  * functions can be shared.
3950  */
3951 
3952 static int
3955 {
3956  /*
3957  * This is similar to comparetup_heap(), but expects index tuples. There
3958  * is also special handling for enforcing uniqueness, and special
3959  * treatment for equal keys at the end.
3960  */
3961  SortSupport sortKey = state->sortKeys;
3962  IndexTuple tuple1;
3963  IndexTuple tuple2;
3964  int keysz;
3965  TupleDesc tupDes;
3966  bool equal_hasnull = false;
3967  int nkey;
3968  int32 compare;
3969  Datum datum1,
3970  datum2;
3971  bool isnull1,
3972  isnull2;
3973 
3974 
3975  /* Compare the leading sort key */
3976  compare = ApplySortComparator(a->datum1, a->isnull1,
3977  b->datum1, b->isnull1,
3978  sortKey);
3979  if (compare != 0)
3980  return compare;
3981 
3982  /* Compare additional sort keys */
3983  tuple1 = (IndexTuple) a->tuple;
3984  tuple2 = (IndexTuple) b->tuple;
3985  keysz = state->nKeys;
3986  tupDes = RelationGetDescr(state->indexRel);
3987 
3988  if (sortKey->abbrev_converter)
3989  {
3990  datum1 = index_getattr(tuple1, 1, tupDes, &isnull1);
3991  datum2 = index_getattr(tuple2, 1, tupDes, &isnull2);
3992 
3993  compare = ApplySortAbbrevFullComparator(datum1, isnull1,
3994  datum2, isnull2,
3995  sortKey);
3996  if (compare != 0)
3997  return compare;
3998  }
3999 
4000  /* they are equal, so we only need to examine one null flag */
4001  if (a->isnull1)
4002  equal_hasnull = true;
4003 
4004  sortKey++;
4005  for (nkey = 2; nkey <= keysz; nkey++, sortKey++)
4006  {
4007  datum1 = index_getattr(tuple1, nkey, tupDes, &isnull1);
4008  datum2 = index_getattr(tuple2, nkey, tupDes, &isnull2);
4009 
4010  compare = ApplySortComparator(datum1, isnull1,
4011  datum2, isnull2,
4012  sortKey);
4013  if (compare != 0)
4014  return compare; /* done when we find unequal attributes */
4015 
4016  /* they are equal, so we only need to examine one null flag */
4017  if (isnull1)
4018  equal_hasnull = true;
4019  }
4020 
4021  /*
4022  * If btree has asked us to enforce uniqueness, complain if two equal
4023  * tuples are detected (unless there was at least one NULL field).
4024  *
4025  * It is sufficient to make the test here, because if two tuples are equal
4026  * they *must* get compared at some stage of the sort --- otherwise the
4027  * sort algorithm wouldn't have checked whether one must appear before the
4028  * other.
4029  */
4030  if (state->enforceUnique && !equal_hasnull)
4031  {
4033  bool isnull[INDEX_MAX_KEYS];
4034  char *key_desc;
4035 
4036  /*
4037  * Some rather brain-dead implementations of qsort (such as the one in
4038  * QNX 4) will sometimes call the comparison routine to compare a
4039  * value to itself, but we always use our own implementation, which
4040  * does not.
4041  */
4042  Assert(tuple1 != tuple2);
4043 
4044  index_deform_tuple(tuple1, tupDes, values, isnull);
4045 
4046  key_desc = BuildIndexValueDescription(state->indexRel, values, isnull);
4047 
4048  ereport(ERROR,
4049  (errcode(ERRCODE_UNIQUE_VIOLATION),
4050  errmsg("could not create unique index \"%s\"",
4052  key_desc ? errdetail("Key %s is duplicated.", key_desc) :
4053  errdetail("Duplicate keys exist."),
4054  errtableconstraint(state->heapRel,
4055  RelationGetRelationName(state->indexRel))));
4056  }
4057 
4058  /*
4059  * If key values are equal, we sort on ItemPointer. This does not affect
4060  * validity of the finished index, but it may be useful to have index
4061  * scans in physical order.
4062  */
4063  {
4064  BlockNumber blk1 = ItemPointerGetBlockNumber(&tuple1->t_tid);
4065  BlockNumber blk2 = ItemPointerGetBlockNumber(&tuple2->t_tid);
4066 
4067  if (blk1 != blk2)
4068  return (blk1 < blk2) ? -1 : 1;
4069  }
4070  {
4073 
4074  if (pos1 != pos2)
4075  return (pos1 < pos2) ? -1 : 1;
4076  }
4077 
4078  return 0;
4079 }
4080 
4081 static int
4084 {
4085  Bucket bucket1;
4086  Bucket bucket2;
4087  IndexTuple tuple1;
4088  IndexTuple tuple2;
4089 
4090  /*
4091  * Fetch hash keys and mask off bits we don't want to sort by. We know
4092  * that the first column of the index tuple is the hash key.
4093  */
4094  Assert(!a->isnull1);
4096  state->max_buckets, state->high_mask,
4097  state->low_mask);
4098  Assert(!b->isnull1);
4100  state->max_buckets, state->high_mask,
4101  state->low_mask);
4102  if (bucket1 > bucket2)
4103  return 1;
4104  else if (bucket1 < bucket2)
4105  return -1;
4106 
4107  /*
4108  * If hash values are equal, we sort on ItemPointer. This does not affect
4109  * validity of the finished index, but it may be useful to have index
4110  * scans in physical order.
4111  */
4112  tuple1 = (IndexTuple) a->tuple;
4113  tuple2 = (IndexTuple) b->tuple;
4114 
4115  {
4116  BlockNumber blk1 = ItemPointerGetBlockNumber(&tuple1->t_tid);
4117  BlockNumber blk2 = ItemPointerGetBlockNumber(&tuple2->t_tid);
4118 
4119  if (blk1 != blk2)
4120  return (blk1 < blk2) ? -1 : 1;
4121  }
4122  {
4125 
4126  if (pos1 != pos2)
4127  return (pos1 < pos2) ? -1 : 1;
4128  }
4129 
4130  return 0;
4131 }
4132 
4133 static void
4135 {
4136  IndexTuple tuple = (IndexTuple) tup;
4137  unsigned int tuplen = IndexTupleSize(tuple);
4138  IndexTuple newtuple;
4139  Datum original;
4140 
4141  /* copy the tuple into sort storage */
4142  newtuple = (IndexTuple) MemoryContextAlloc(state->tuplecontext, tuplen);
4143  memcpy(newtuple, tuple, tuplen);
4144  USEMEM(state, GetMemoryChunkSpace(newtuple));
4145  stup->tuple = (void *) newtuple;
4146  /* set up first-column key value */
4147  original = index_getattr(newtuple,
4148  1,
4149  RelationGetDescr(state->indexRel),
4150  &stup->isnull1);
4151 
4152  if (!state->sortKeys->abbrev_converter || stup->isnull1)
4153  {
4154  /*
4155  * Store ordinary Datum representation, or NULL value. If there is a
4156  * converter it won't expect NULL values, and cost model is not
4157  * required to account for NULL, so in that case we avoid calling
4158  * converter and just set datum1 to zeroed representation (to be
4159  * consistent, and to support cheap inequality tests for NULL
4160  * abbreviated keys).
4161  */
4162  stup->datum1 = original;
4163  }
4164  else if (!consider_abort_common(state))
4165  {
4166  /* Store abbreviated key representation */
4167  stup->datum1 = state->sortKeys->abbrev_converter(original,
4168  state->sortKeys);
4169  }
4170  else
4171  {
4172  /* Abort abbreviation */
4173  int i;
4174 
4175  stup->datum1 = original;
4176 
4177  /*
4178  * Set state to be consistent with never trying abbreviation.
4179  *
4180  * Alter datum1 representation in already-copied tuples, so as to
4181  * ensure a consistent representation (current tuple was just
4182  * handled). It does not matter if some dumped tuples are already
4183  * sorted on tape, since serialized tuples lack abbreviated keys
4184  * (TSS_BUILDRUNS state prevents control reaching here in any case).
4185  */
4186  for (i = 0; i < state->memtupcount; i++)
4187  {
4188  SortTuple *mtup = &state->memtuples[i];
4189 
4190  tuple = (IndexTuple) mtup->tuple;
4191  mtup->datum1 = index_getattr(tuple,
4192  1,
4193  RelationGetDescr(state->indexRel),
4194  &mtup->isnull1);
4195  }
4196  }
4197 }
4198 
4199 static void
4201 {
4202  IndexTuple tuple = (IndexTuple) stup->tuple;
4203  unsigned int tuplen;
4204 
4205  tuplen = IndexTupleSize(tuple) + sizeof(tuplen);
4206  LogicalTapeWrite(state->tapeset, tapenum,
4207  (void *) &tuplen, sizeof(tuplen));
4208  LogicalTapeWrite(state->tapeset, tapenum,
4209  (void *) tuple, IndexTupleSize(tuple));
4210  if (state->randomAccess) /* need trailing length word? */
4211  LogicalTapeWrite(state->tapeset, tapenum,
4212  (void *) &tuplen, sizeof(tuplen));
4213 
4214  if (!state->slabAllocatorUsed)
4215  {
4216  FREEMEM(state, GetMemoryChunkSpace(tuple));
4217  pfree(tuple);
4218  }
4219 }
4220 
4221 static void
4223  int tapenum, unsigned int len)
4224 {
4225  unsigned int tuplen = len - sizeof(unsigned int);
4226  IndexTuple tuple = (IndexTuple) readtup_alloc(state, tuplen);
4227 
4228  LogicalTapeReadExact(state->tapeset, tapenum,
4229  tuple, tuplen);
4230  if (state->randomAccess) /* need trailing length word? */
4231  LogicalTapeReadExact(state->tapeset, tapenum,
4232  &tuplen, sizeof(tuplen));
4233  stup->tuple = (void *) tuple;
4234  /* set up first-column key value */
4235  stup->datum1 = index_getattr(tuple,
4236  1,
4237  RelationGetDescr(state->indexRel),
4238  &stup->isnull1);
4239 }
4240 
4241 /*
4242  * Routines specialized for DatumTuple case
4243  */
4244 
4245 static int
4247 {
4248  int compare;
4249 
4250  compare = ApplySortComparator(a->datum1, a->isnull1,
4251  b->datum1, b->isnull1,
4252  state->sortKeys);
4253  if (compare != 0)
4254  return compare;
4255 
4256  /* if we have abbreviations, then "tuple" has the original value */
4257 
4258  if (state->sortKeys->abbrev_converter)
4260  PointerGetDatum(b->tuple), b->isnull1,
4261  state->sortKeys);
4262 
4263  return compare;
4264 }
4265 
4266 static void
4268 {
4269  /* Not currently needed */
4270  elog(ERROR, "copytup_datum() should not be called");
4271 }
4272 
4273 static void
4275 {
4276  void *waddr;
4277  unsigned int tuplen;
4278  unsigned int writtenlen;
4279 
4280  if (stup->isnull1)
4281  {
4282  waddr = NULL;
4283  tuplen = 0;
4284  }
4285  else if (!state->tuples)
4286  {
4287  waddr = &stup->datum1;
4288  tuplen = sizeof(Datum);
4289  }
4290  else
4291  {
4292  waddr = stup->tuple;
4293  tuplen = datumGetSize(PointerGetDatum(stup->tuple), false, state->datumTypeLen);
4294  Assert(tuplen != 0);
4295  }
4296 
4297  writtenlen = tuplen + sizeof(unsigned int);
4298 
4299  LogicalTapeWrite(state->tapeset, tapenum,
4300  (void *) &writtenlen, sizeof(writtenlen));
4301  LogicalTapeWrite(state->tapeset, tapenum,
4302  waddr, tuplen);
4303  if (state->randomAccess) /* need trailing length word? */
4304  LogicalTapeWrite(state->tapeset, tapenum,
4305  (void *) &writtenlen, sizeof(writtenlen));
4306 
4307  if (!state->slabAllocatorUsed && stup->tuple)
4308  {
4309  FREEMEM(state, GetMemoryChunkSpace(stup->tuple));
4310  pfree(stup->tuple);
4311  }
4312 }
4313 
4314 static void
4316  int tapenum, unsigned int len)
4317 {
4318  unsigned int tuplen = len - sizeof(unsigned int);
4319 
4320  if (tuplen == 0)
4321  {
4322  /* it's NULL */
4323  stup->datum1 = (Datum) 0;
4324  stup->isnull1 = true;
4325  stup->tuple = NULL;
4326  }
4327  else if (!state->tuples)
4328  {
4329  Assert(tuplen == sizeof(Datum));
4330  LogicalTapeReadExact(state->tapeset, tapenum,
4331  &stup->datum1, tuplen);
4332  stup->isnull1 = false;
4333  stup->tuple = NULL;
4334  }
4335  else
4336  {
4337  void *raddr = readtup_alloc(state, tuplen);
4338 
4339  LogicalTapeReadExact(state->tapeset, tapenum,
4340  raddr, tuplen);
4341  stup->datum1 = PointerGetDatum(raddr);
4342  stup->isnull1 = false;
4343  stup->tuple = raddr;
4344  }
4345 
4346  if (state->randomAccess) /* need trailing length word? */
4347  LogicalTapeReadExact(state->tapeset, tapenum,
4348  &tuplen, sizeof(tuplen));
4349 }
4350 
4351 /*
4352  * Parallel sort routines
4353  */
4354 
4355 /*
4356  * tuplesort_estimate_shared - estimate required shared memory allocation
4357  *
4358  * nWorkers is an estimate of the number of workers (it's the number that
4359  * will be requested).
4360  */
4361 Size
4363 {
4364  Size tapesSize;
4365 
4366  Assert(nWorkers > 0);
4367 
4368  /* Make sure that BufFile shared state is MAXALIGN'd */
4369  tapesSize = mul_size(sizeof(TapeShare), nWorkers);
4370  tapesSize = MAXALIGN(add_size(tapesSize, offsetof(Sharedsort, tapes)));
4371 
4372  return tapesSize;
4373 }
4374 
4375 /*
4376  * tuplesort_initialize_shared - initialize shared tuplesort state
4377  *
4378  * Must be called from leader process before workers are launched, to
4379  * establish state needed up-front for worker tuplesortstates. nWorkers
4380  * should match the argument passed to tuplesort_estimate_shared().
4381  */
4382 void
4384 {
4385  int i;
4386 
4387  Assert(nWorkers > 0);
4388 
4389  SpinLockInit(&shared->mutex);
4390  shared->currentWorker = 0;
4391  shared->workersFinished = 0;
4392  SharedFileSetInit(&shared->fileset, seg);
4393  shared->nTapes = nWorkers;
4394  for (i = 0; i < nWorkers; i++)
4395  {
4396  shared->tapes[i].firstblocknumber = 0L;
4397  shared->tapes[i].buffilesize = 0;
4398  }
4399 }
4400 
4401 /*
4402  * tuplesort_attach_shared - attach to shared tuplesort state
4403  *
4404  * Must be called by all worker processes.
4405  */
4406 void
4408 {
4409  /* Attach to SharedFileSet */
4410  SharedFileSetAttach(&shared->fileset, seg);
4411 }
4412 
4413 /*
4414  * worker_get_identifier - Assign and return ordinal identifier for worker
4415  *
4416  * The order in which these are assigned is not well defined, and should not
4417  * matter; worker numbers across parallel sort participants need only be
4418  * distinct and gapless. logtape.c requires this.
4419  *
4420  * Note that the identifiers assigned from here have no relation to
4421  * ParallelWorkerNumber number, to avoid making any assumption about
4422  * caller's requirements. However, we do follow the ParallelWorkerNumber
4423  * convention of representing a non-worker with worker number -1. This
4424  * includes the leader, as well as serial Tuplesort processes.
4425  */
4426 static int
4428 {
4429  Sharedsort *shared = state->shared;
4430  int worker;
4431 
4432  Assert(WORKER(state));
4433 
4434  SpinLockAcquire(&shared->mutex);
4435  worker = shared->currentWorker++;
4436  SpinLockRelease(&shared->mutex);
4437 
4438  return worker;
4439 }
4440 
4441 /*
4442  * worker_freeze_result_tape - freeze worker's result tape for leader
4443  *
4444  * This is called by workers just after the result tape has been determined,
4445  * instead of calling LogicalTapeFreeze() directly. They do so because
4446  * workers require a few additional steps over similar serial
4447  * TSS_SORTEDONTAPE external sort cases, which also happen here. The extra
4448  * steps are around freeing now unneeded resources, and representing to
4449  * leader that worker's input run is available for its merge.
4450  *
4451  * There should only be one final output run for each worker, which consists
4452  * of all tuples that were originally input into worker.
4453  */
4454 static void
4456 {
4457  Sharedsort *shared = state->shared;
4458  TapeShare output;
4459 
4460  Assert(WORKER(state));
4461  Assert(state->result_tape != -1);
4462  Assert(state->memtupcount == 0);
4463 
4464  /*
4465  * Free most remaining memory, in case caller is sensitive to our holding
4466  * on to it. memtuples may not be a tiny merge heap at this point.
4467  */
4468  pfree(state->memtuples);
4469  /* Be tidy */
4470  state->memtuples = NULL;
4471  state->memtupsize = 0;
4472 
4473  /*
4474  * Parallel worker requires result tape metadata, which is to be stored in
4475  * shared memory for leader
4476  */
4477  LogicalTapeFreeze(state->tapeset, state->result_tape, &output);
4478 
4479  /* Store properties of output tape, and update finished worker count */
4480  SpinLockAcquire(&shared->mutex);
4481  shared->tapes[state->worker] = output;
4482  shared->workersFinished++;
4483  SpinLockRelease(&shared->mutex);
4484 }
4485 
4486 /*
4487  * worker_nomergeruns - dump memtuples in worker, without merging
4488  *
4489  * This called as an alternative to mergeruns() with a worker when no
4490  * merging is required.
4491  */
4492 static void
4494 {
4495  Assert(WORKER(state));
4496  Assert(state->result_tape == -1);
4497 
4498  state->result_tape = state->tp_tapenum[state->destTape];
4500 }
4501 
4502 /*
4503  * leader_takeover_tapes - create tapeset for leader from worker tapes
4504  *
4505  * So far, leader Tuplesortstate has performed no actual sorting. By now, all
4506  * sorting has occurred in workers, all of which must have already returned
4507  * from tuplesort_performsort().
4508  *
4509  * When this returns, leader process is left in a state that is virtually
4510  * indistinguishable from it having generated runs as a serial external sort
4511  * might have.
4512  */
4513 static void
4515 {
4516  Sharedsort *shared = state->shared;
4517  int nParticipants = state->nParticipants;
4518  int workersFinished;
4519  int j;
4520 
4521  Assert(LEADER(state));
4522  Assert(nParticipants >= 1);
4523 
4524  SpinLockAcquire(&shared->mutex);
4525  workersFinished = shared->workersFinished;
4526  SpinLockRelease(&shared->mutex);
4527 
4528  if (nParticipants != workersFinished)
4529  elog(ERROR, "cannot take over tapes before all workers finish");
4530 
4531  /*
4532  * Create the tapeset from worker tapes, including a leader-owned tape at
4533  * the end. Parallel workers are far more expensive than logical tapes,
4534  * so the number of tapes allocated here should never be excessive.
4535  *
4536  * We still have a leader tape, though it's not possible to write to it
4537  * due to restrictions in the shared fileset infrastructure used by
4538  * logtape.c. It will never be written to in practice because
4539  * randomAccess is disallowed for parallel sorts.
4540  */
4541  inittapestate(state, nParticipants + 1);
4542  state->tapeset = LogicalTapeSetCreate(nParticipants + 1, shared->tapes,
4543  &shared->fileset, state->worker);
4544 
4545  /* mergeruns() relies on currentRun for # of runs (in one-pass cases) */
4546  state->currentRun = nParticipants;
4547 
4548  /*
4549  * Initialize variables of Algorithm D to be consistent with runs from
4550  * workers having been generated in the leader.
4551  *
4552  * There will always be exactly 1 run per worker, and exactly one input
4553  * tape per run, because workers always output exactly 1 run, even when
4554  * there were no input tuples for workers to sort.
4555  */
4556  for (j = 0; j < state->maxTapes; j++)
4557  {
4558  /* One real run; no dummy runs for worker tapes */
4559  state->tp_fib[j] = 1;
4560  state->tp_runs[j] = 1;
4561  state->tp_dummy[j] = 0;
4562  state->tp_tapenum[j] = j;
4563  }
4564  /* Leader tape gets one dummy run, and no real runs */
4565  state->tp_fib[state->tapeRange] = 0;
4566  state->tp_runs[state->tapeRange] = 0;
4567  state->tp_dummy[state->tapeRange] = 1;
4568 
4569  state->Level = 1;
4570  state->destTape = 0;
4571 
4572  state->status = TSS_BUILDRUNS;
4573 }
4574 
4575 /*
4576  * Convenience routine to free a tuple previously loaded into sort memory
4577  */
4578 static void
4580 {
4581  FREEMEM(state, GetMemoryChunkSpace(stup->tuple));
4582  pfree(stup->tuple);
4583 }
struct SortSupportData * SortSupport
Definition: sortsupport.h:58
IndexTuple tuplesort_getindextuple(Tuplesortstate *state, bool forward)
Definition: tuplesort.c:2215
#define MINIMAL_TUPLE_DATA_OFFSET
Definition: htup_details.h:629
signed short int16
Definition: c.h:301
static int comparetup_index_hash(const SortTuple *a, const SortTuple *b, Tuplesortstate *state)
Definition: tuplesort.c:4082
void FormIndexDatum(IndexInfo *indexInfo, TupleTableSlot *slot, EState *estate, Datum *values, bool *isnull)
Definition: index.c:1941
HeapTuple heap_copytuple(HeapTuple tuple)
Definition: heaptuple.c:611
bool ssup_nulls_first
Definition: sortsupport.h:75
#define DatumGetUInt32(X)
Definition: postgres.h:469
int slock_t
Definition: s_lock.h:912
ScanKey _bt_mkscankey_nodata(Relation rel)
Definition: nbtutils.c:115
size_t LogicalTapeRead(LogicalTapeSet *lts, int tapenum, void *ptr, size_t size)
Definition: logtape.c:809
TupleTableSlot * ExecStoreTuple(HeapTuple tuple, TupleTableSlot *slot, Buffer buffer, bool shouldFree)
Definition: execTuples.c:356
static int comparetup_heap(const SortTuple *a, const SortTuple *b, Tuplesortstate *state)
Definition: tuplesort.c:3508
static void dumptuples(Tuplesortstate *state, bool alltuples)
Definition: tuplesort.c:2924
static void reversedirection(Tuplesortstate *state)
Definition: tuplesort.c:3435
bool tuplesort_getdatum(Tuplesortstate *state, bool forward, Datum *val, bool *isNull, Datum *abbrev)
Definition: tuplesort.c:2244
int64 availMem
Definition: tuplesort.c:242
size_t read_buffer_size
Definition: tuplesort.c:334
TupSortStatus status
Definition: tuplesort.c:234
static void writetup_cluster(Tuplesortstate *state, int tapenum, SortTuple *stup)
Definition: tuplesort.c:3891
Relation heapRel
Definition: tuplesort.c:442
void tuplesort_performsort(Tuplesortstate *state)
Definition: tuplesort.c:1791
static bool grow_memtuples(Tuplesortstate *state)
Definition: tuplesort.c:1316
void MemoryContextDelete(MemoryContext context)
Definition: mcxt.c:198
HeapTupleData * HeapTuple
Definition: htup.h:70
HeapTuple tuplesort_getheaptuple(Tuplesortstate *state, bool forward)
Definition: tuplesort.c:2195
TupleTableSlot * ExecStoreMinimalTuple(MinimalTuple mtup, TupleTableSlot *slot, bool shouldFree)
Definition: execTuples.c:420
#define BTGreaterStrategyNumber
Definition: stratnum.h:33
const char * tuplesort_space_type_name(TuplesortSpaceType t)
Definition: tuplesort.c:3198
void tuplesort_putdatum(Tuplesortstate *state, Datum val, bool isNull)
Definition: tuplesort.c:1556
#define AssertState(condition)
Definition: c.h:691
char * slabMemoryEnd
Definition: tuplesort.c:330
Tuplesortstate * tuplesort_begin_datum(Oid datumType, Oid sortOperator, Oid sortCollation, bool nullsFirstFlag, int workMem, SortCoordinate coordinate, bool randomAccess)
Definition: tuplesort.c:1100
void tuplesort_restorepos(Tuplesortstate *state)
Definition: tuplesort.c:3094
static void mergeonerun(Tuplesortstate *state)
Definition: tuplesort.c:2787
static void worker_freeze_result_tape(Tuplesortstate *state)
Definition: tuplesort.c:4455
PGRUsage ru_start
Definition: tuplesort.c:465
#define SERIAL(state)
Definition: tuplesort.c:532
slock_t mutex
Definition: tuplesort.c:476
static void sort_bounded_heap(Tuplesortstate *state)
Definition: tuplesort.c:3268
void SharedFileSetInit(SharedFileSet *fileset, dsm_segment *seg)
Definition: sharedfileset.c:47
TuplesortMethod
Definition: tuplesort.h:66
#define ResetPerTupleExprContext(estate)
Definition: executor.h:499
int64 abbrevNext
Definition: tuplesort.c:428
#define RelationGetDescr(relation)
Definition: rel.h:437
off_t buffilesize
Definition: logtape.h:53
Bucket _hash_hashkey2bucket(uint32 hashkey, uint32 maxbucket, uint32 highmask, uint32 lowmask)
Definition: hashutil.c:125
#define RelationGetNumberOfAttributes(relation)
Definition: rel.h:431
EState * estate
Definition: tuplesort.c:436
static void output(uint64 loop_count)
#define PointerGetDatum(X)
Definition: postgres.h:539
HeapTupleHeaderData * HeapTupleHeader
Definition: htup.h:23
static void copytup_index(Tuplesortstate *state, SortTuple *stup, void *tup)
Definition: tuplesort.c:4134
int(* SortTupleComparator)(const SortTuple *a, const SortTuple *b, Tuplesortstate *state)
Definition: tuplesort.c:226
SortSupport sortKeys
Definition: tuplesort.c:414
#define SpinLockInit(lock)
Definition: spin.h:60
void _bt_freeskey(ScanKey skey)
Definition: nbtutils.c:155
ItemPointerData t_tid
Definition: itup.h:37
void(* copytup)(Tuplesortstate *state, SortTuple *stup, void *tup)
Definition: tuplesort.c:267
#define BTREE_AM_OID
Definition: pg_am.h:70
#define Min(x, y)
Definition: c.h:846
char buffer[SLAB_SLOT_SIZE]
Definition: tuplesort.c:193
MinimalTuple ExecCopySlotMinimalTuple(TupleTableSlot *slot)
Definition: execTuples.c:613
void PrepareSortSupportFromOrderingOp(Oid orderingOp, SortSupport ssup)
Definition: sortsupport.c:133
bool randomAccess
Definition: tuplesort.c:236
TupleTableSlot * ExecClearTuple(TupleTableSlot *slot)
Definition: execTuples.c:475
static MemoryContext MemoryContextSwitchTo(MemoryContext context)
Definition: palloc.h:109
Datum datum1
Definition: tuplesort.c:172
Tuplesortstate * tuplesort_begin_cluster(TupleDesc tupDesc, Relation indexRel, int workMem, SortCoordinate coordinate, bool randomAccess)
Definition: tuplesort.c:880
#define InvalidBuffer
Definition: buf.h:25
SortTupleComparator comparetup
Definition: tuplesort.c:259
#define CLUSTER_SORT
Definition: tuplesort.c:122
Sharedsort * sharedsort
Definition: tuplesort.h:56
int errcode(int sqlerrcode)
Definition: elog.c:575
#define LogicalTapeReadExact(tapeset, tapenum, ptr, len)
Definition: tuplesort.c:584
bool growmemtuples
Definition: tuplesort.c:298
Tuplesortstate * tuplesort_begin_index_hash(Relation heapRel, Relation indexRel, uint32 high_mask, uint32 low_mask, uint32 max_buckets, int workMem, SortCoordinate coordinate, bool randomAccess)
Definition: tuplesort.c:1054
Size GetMemoryChunkSpace(void *pointer)
Definition: mcxt.c:390
long firstblocknumber
Definition: logtape.h:52
#define SLAB_SLOT_SIZE
Definition: tuplesort.c:188
static int comparetup_cluster(const SortTuple *a, const SortTuple *b, Tuplesortstate *state)
Definition: tuplesort.c:3707
void MemoryContextReset(MemoryContext context)
Definition: mcxt.c:134
uint32 BlockNumber
Definition: block.h:31
#define INDEX_SORT
Definition: tuplesort.c:120
IndexInfo * BuildIndexInfo(Relation index)
Definition: index.c:1701
static void inittapes(Tuplesortstate *state, bool mergeruns)
Definition: tuplesort.c:2391
void tuplesort_initialize_shared(Sharedsort *shared, int nWorkers, dsm_segment *seg)
Definition: tuplesort.c:4383
#define LOG
Definition: elog.h:26
Form_pg_class rd_rel
Definition: rel.h:114
void heap_freetuple(HeapTuple htup)
Definition: heaptuple.c:1373
unsigned int Oid
Definition: postgres_ext.h:31
#define HEAP_SORT
Definition: tuplesort.c:119
static int worker_get_identifier(Tuplesortstate *state)
Definition: tuplesort.c:4427
static void copytup_cluster(Tuplesortstate *state, SortTuple *stup, void *tup)
Definition: tuplesort.c:3818
static int comparetup_index_btree(const SortTuple *a, const SortTuple *b, Tuplesortstate *state)
Definition: tuplesort.c:3953
bool isnull1
Definition: tuplesort.c:173
void tuplesort_get_stats(Tuplesortstate *state, TuplesortInstrumentation *stats)
Definition: tuplesort.c:3128
const char * tuplesort_method_name(TuplesortMethod m)
Definition: tuplesort.c:3175
MinimalTuple heap_copy_minimal_tuple(MinimalTuple mtup)
Definition: heaptuple.c:1480
int markpos_offset
Definition: tuplesort.c:386
bool trace_sort
Definition: tuplesort.c:130
static void mergeruns(Tuplesortstate *state)
Definition: tuplesort.c:2560
void LogicalTapeRewindForWrite(LogicalTapeSet *lts, int tapenum)
Definition: logtape.c:783
static void markrunend(Tuplesortstate *state, int tapenum)
Definition: tuplesort.c:3466
static unsigned int getlen(Tuplesortstate *state, int tapenum, bool eofOK)
Definition: tuplesort.c:3453
#define PARALLEL_SORT(state)
Definition: tuplesort.c:125
#define MaxAllocHugeSize
Definition: memutils.h:44
signed int int32
Definition: c.h:302
static void init_slab_allocator(Tuplesortstate *state, int numSlots)
Definition: tuplesort.c:2524
#define DATUM_SORT
Definition: tuplesort.c:121
uint16 OffsetNumber
Definition: off.h:24
HeapTupleHeader t_data
Definition: htup.h:67
void pg_rusage_init(PGRUsage *ru0)
Definition: pg_rusage.c:27
Tuplesortstate * tuplesort_begin_heap(TupleDesc tupDesc, int nkeys, AttrNumber *attNums, Oid *sortOperators, Oid *sortCollations, bool *nullsFirstFlags, int workMem, SortCoordinate coordinate, bool randomAccess)
Definition: tuplesort.c:806
uint32 Bucket
Definition: hash.h:34
IndexTuple index_form_tuple(TupleDesc tupleDescriptor, Datum *values, bool *isnull)
Definition: indextuple.c:37
void FreeExecutorState(EState *estate)
Definition: execUtils.c:185
#define GetPerTupleExprContext(estate)
Definition: executor.h:490
void * tuple
Definition: tuplesort.c:171
int errtableconstraint(Relation rel, const char *conname)
Definition: relcache.c:5281
#define SpinLockAcquire(lock)
Definition: spin.h:62
#define TAPE_BUFFER_OVERHEAD
Definition: tuplesort.c:223
static bool consider_abort_common(Tuplesortstate *state)
Definition: tuplesort.c:1747
void pfree(void *pointer)
Definition: mcxt.c:936
union SlabSlot * nextfree
Definition: tuplesort.c:192
static int compare(const void *arg1, const void *arg2)
Definition: geqo_pool.c:145
void LogicalTapeWrite(LogicalTapeSet *lts, int tapenum, void *ptr, size_t size)
Definition: logtape.c:620
static void readtup_cluster(Tuplesortstate *state, SortTuple *stup, int tapenum, unsigned int len)
Definition: tuplesort.c:3915
#define ERROR
Definition: elog.h:43
void PrepareTempTablespaces(void)
Definition: tablespace.c:1287
void heap_free_minimal_tuple(MinimalTuple mtup)
Definition: heaptuple.c:1468
void * lastReturnedTuple
Definition: tuplesort.c:342
MemoryContext sortcontext
Definition: tuplesort.c:246
static void writetup_heap(Tuplesortstate *state, int tapenum, SortTuple *stup)
Definition: tuplesort.c:3648
static void writetup_index(Tuplesortstate *state, int tapenum, SortTuple *stup)
Definition: tuplesort.c:4200
MemoryContext ssup_cxt
Definition: sortsupport.h:66
uint32 high_mask
Definition: tuplesort.c:449
LogicalTapeSet * LogicalTapeSetCreate(int ntapes, TapeShare *shared, SharedFileSet *fileset, int worker)
Definition: logtape.c:509
#define MERGE_BUFFER_SIZE
Definition: tuplesort.c:224
ItemPointerData t_self
Definition: htup.h:65
static int comparetup_datum(const SortTuple *a, const SortTuple *b, Tuplesortstate *state)
Definition: tuplesort.c:4246
void tuplesort_rescan(Tuplesortstate *state)
Definition: tuplesort.c:3027
void(* readtup)(Tuplesortstate *state, SortTuple *stup, int tapenum, unsigned int len)
Definition: tuplesort.c:285
#define ALLOCSET_DEFAULT_SIZES
Definition: memutils.h:197
bool tuplesort_gettupleslot(Tuplesortstate *state, bool forward, bool copy, TupleTableSlot *slot, Datum *abbrev)
Definition: tuplesort.c:2158
int(* comparator)(Datum x, Datum y, SortSupport ssup)
Definition: sortsupport.h:107
static void free_sort_tuple(Tuplesortstate *state, SortTuple *stup)
Definition: tuplesort.c:4579
static void readtup_index(Tuplesortstate *state, SortTuple *stup, int tapenum, unsigned int len)
Definition: tuplesort.c:4222
uint32 t_len
Definition: htup.h:64
static void copytup_datum(Tuplesortstate *state, SortTuple *stup, void *tup)
Definition: tuplesort.c:4267
void tuplesort_set_bound(Tuplesortstate *state, int64 bound)
Definition: tuplesort.c:1187
int(* abbrev_full_comparator)(Datum x, Datum y, SortSupport ssup)
Definition: sortsupport.h:192
static char * buf
Definition: pg_test_fsync.c:67
Sharedsort * shared
Definition: tuplesort.c:405
TuplesortMethod sortMethod
Definition: tuplesort.h:83
static void readtup_datum(Tuplesortstate *state, SortTuple *stup, int tapenum, unsigned int len)
Definition: tuplesort.c:4315
void ExecDropSingleTupleTableSlot(TupleTableSlot *slot)
Definition: execTuples.c:247
MinimalTupleData * MinimalTuple
Definition: htup.h:27
IndexTupleData * IndexTuple
Definition: itup.h:53
#define COMPARETUP(state, a, b)
Definition: tuplesort.c:525
#define ALLOCSET_SEPARATE_THRESHOLD
Definition: memutils.h:224
int errdetail(const char *fmt,...)
Definition: elog.c:873
size_t LogicalTapeBackspace(LogicalTapeSet *lts, int tapenum, size_t size)
Definition: logtape.c:948
#define MINORDER
Definition: tuplesort.c:221
union SlabSlot SlabSlot
IndexInfo * indexInfo
Definition: tuplesort.c:435
static void tuplesort_heap_insert(Tuplesortstate *state, SortTuple *tuple)
Definition: tuplesort.c:3336
const char * pg_rusage_show(const PGRUsage *ru0)
Definition: pg_rusage.c:40
#define LEADER(state)
Definition: tuplesort.c:534
#define FREEMEM(state, amt)
Definition: tuplesort.c:531
#define RelationGetRelationName(relation)
Definition: rel.h:445
Datum(* abbrev_converter)(Datum original, SortSupport ssup)
Definition: sortsupport.h:173
static bool tuplesort_gettuple_common(Tuplesortstate *state, bool forward, SortTuple *stup)
Definition: tuplesort.c:1902
unsigned int uint32
Definition: c.h:314
LogicalTapeSet * tapeset
Definition: tuplesort.c:248
Oid t_tableOid
Definition: htup.h:66
void index_deform_tuple(IndexTuple tup, TupleDesc tupleDescriptor, Datum *values, bool *isnull)
Definition: indextuple.c:420
void LogicalTapeTell(LogicalTapeSet *lts, int tapenum, long *blocknum, int *offset)
Definition: logtape.c:1050
MemoryContext CurrentMemoryContext
Definition: mcxt.c:37
#define WORKER(state)
Definition: tuplesort.c:533
int64 allowedMem
Definition: tuplesort.c:243
TupleTableSlot * MakeSingleTupleTableSlot(TupleDesc tupdesc)
Definition: execTuples.c:232
int workersFinished
Definition: tuplesort.c:487
void PrepareSortSupportFromIndexRel(Relation indexRel, int16 strategy, SortSupport ssup)
Definition: sortsupport.c:160
int nTapes
Definition: tuplesort.c:493
#define ereport(elevel, rest)
Definition: elog.h:122
static void selectnewtape(Tuplesortstate *state)
Definition: tuplesort.c:2492
#define AssertArg(condition)
Definition: c.h:690
Datum datumCopy(Datum value, bool typByVal, int typLen)
Definition: datum.c:128
#define COPYTUP(state, stup, tup)
Definition: tuplesort.c:526
EState * CreateExecutorState(void)
Definition: execUtils.c:80
static void copytup_heap(Tuplesortstate *state, SortTuple *stup, void *tup)
Definition: tuplesort.c:3570
#define AllocSetContextCreate(parent, name, allocparams)
Definition: memutils.h:165
#define heap_getattr(tup, attnum, tupleDesc, isnull)
Definition: htup_details.h:774
#define READTUP(state, stup, tape, len)
Definition: tuplesort.c:528
#define SK_BT_NULLS_FIRST
Definition: nbtree.h:435
#define WRITETUP(state, tape, stup)
Definition: tuplesort.c:527
#define SpinLockRelease(lock)
Definition: spin.h:64
Size mul_size(Size s1, Size s2)
Definition: shmem.c:492
static void writetup_datum(Tuplesortstate *state, int tapenum, SortTuple *stup)
Definition: tuplesort.c:4274
#define RELEASE_SLAB_SLOT(state, tuple)
Definition: tuplesort.c:513
void * palloc0(Size size)
Definition: mcxt.c:864
Relation indexRel
Definition: tuplesort.c:443
char * slabMemoryBegin
Definition: tuplesort.c:329
uintptr_t Datum
Definition: postgres.h:365
Size add_size(Size s1, Size s2)
Definition: shmem.c:475
AttrNumber ssup_attno
Definition: sortsupport.h:81
static void puttuple_common(Tuplesortstate *state, SortTuple *tuple)
Definition: tuplesort.c:1638
void(* writetup)(Tuplesortstate *state, int tapenum, SortTuple *stup)
Definition: tuplesort.c:277
static void tuplesort_heap_delete_top(Tuplesortstate *state)
Definition: tuplesort.c:3371
#define InvalidOid
Definition: postgres_ext.h:36
void tuplesort_putindextuplevalues(Tuplesortstate *state, Relation rel, ItemPointer self, Datum *values, bool *isnull)
Definition: tuplesort.c:1478
static bool mergereadnext(Tuplesortstate *state, int srcTape, SortTuple *stup)
Definition: tuplesort.c:2899
#define Max(x, y)
Definition: c.h:840
void tuplesort_attach_shared(Sharedsort *shared, dsm_segment *seg)
Definition: tuplesort.c:4407
int sk_flags
Definition: skey.h:66
List * ii_Expressions
Definition: execnodes.h:148
#define Assert(condition)
Definition: c.h:688
#define SK_BT_DESC
Definition: nbtree.h:434
Definition: regguts.h:298
bool enforceUnique
Definition: tuplesort.c:446
TuplesortSpaceType
Definition: tuplesort.h:75
bool(* abbrev_abort)(int memtupcount, SortSupport ssup)
Definition: sortsupport.h:183
long markpos_block
Definition: tuplesort.c:385
struct ItemPointerData ItemPointerData
#define INDEX_MAX_KEYS
size_t Size
Definition: c.h:422
bool eof_reached
Definition: tuplesort.c:382
TupSortStatus
Definition: tuplesort.c:200
int tuplesort_merge_order(int64 allowedMem)
Definition: tuplesort.c:2351
void get_typlenbyval(Oid typid, int16 *typlen, bool *typbyval)
Definition: lsyscache.c:2005
TupleTableSlot * ecxt_scantuple
Definition: execnodes.h:211
#define MAXALIGN(LEN)
Definition: c.h:641
#define index_getattr(tup, attnum, tupleDesc, isnull)
Definition: itup.h:100
static void tuplesort_sort_memtuples(Tuplesortstate *state)
Definition: tuplesort.c:3308
AttrNumber ii_KeyAttrNumbers[INDEX_MAX_KEYS]
Definition: execnodes.h:147
#define ItemPointerGetOffsetNumber(pointer)
Definition: itemptr.h:95
int tupindex
Definition: tuplesort.c:174
Size tuplesort_estimate_shared(int nWorkers)
Definition: tuplesort.c:4362
uint32 max_buckets
Definition: tuplesort.c:451
#define MINIMAL_TUPLE_OFFSET
Definition: htup_details.h:625
#define INT64_FORMAT
Definition: c.h:356
void LogicalTapeRewindForRead(LogicalTapeSet *lts, int tapenum, size_t buffer_size)
Definition: logtape.c:708
int * tp_dummy
Definition: tuplesort.c:371
static void readtup_heap(Tuplesortstate *state, SortTuple *stup, int tapenum, unsigned int len)
Definition: tuplesort.c:3675
bool slabAllocatorUsed
Definition: tuplesort.c:327
static void leader_takeover_tapes(Tuplesortstate *state)
Definition: tuplesort.c:4514
#define DatumGetPointer(X)
Definition: postgres.h:532
void LogicalTapeFreeze(LogicalTapeSet *lts, int tapenum, TapeShare *share)
Definition: logtape.c:863
static Datum values[MAXATTR]
Definition: bootstrap.c:164
static Tuplesortstate * tuplesort_begin_common(int workMem, SortCoordinate coordinate, bool randomAccess)
Definition: tuplesort.c:681
static void * readtup_alloc(Tuplesortstate *state, Size tuplen)
Definition: tuplesort.c:3480
MemoryContext tuplecontext
Definition: tuplesort.c:247
void * repalloc_huge(void *pointer, Size size)
Definition: mcxt.c:1017
void * palloc(Size size)
Definition: mcxt.c:835
int errmsg(const char *fmt,...)
Definition: elog.c:797
void SharedFileSetAttach(SharedFileSet *fileset, dsm_segment *seg)
Definition: sharedfileset.c:76
int * tp_tapenum
Definition: tuplesort.c:372
#define USEMEM(state, amt)
Definition: tuplesort.c:530
bool markpos_eof
Definition: tuplesort.c:387
#define HEAPTUPLESIZE
Definition: htup.h:72
void * MemoryContextAlloc(MemoryContext context, Size size)
Definition: mcxt.c:693
Oid sk_collation
Definition: skey.h:70
int i
int currentWorker
Definition: tuplesort.c:486
uint32 low_mask
Definition: tuplesort.c:450
TapeShare tapes[FLEXIBLE_ARRAY_MEMBER]
Definition: tuplesort.c:499
void tuplesort_markpos(Tuplesortstate *state)
Definition: tuplesort.c:3062
void tuplesort_putheaptuple(Tuplesortstate *state, HeapTuple tup)
Definition: tuplesort.c:1457
void LogicalTapeSetClose(LogicalTapeSet *lts)
Definition: logtape.c:583
void LogicalTapeSeek(LogicalTapeSet *lts, int tapenum, long blocknum, int offset)
Definition: logtape.c:1019
TuplesortSpaceType spaceType
Definition: tuplesort.h:84
#define CHECK_FOR_INTERRUPTS()
Definition: miscadmin.h:98
SortSupport onlyKey
Definition: tuplesort.c:420
#define elog
Definition: elog.h:219
Tuplesortstate * tuplesort_begin_index_btree(Relation heapRel, Relation indexRel, bool enforceUnique, int workMem, SortCoordinate coordinate, bool randomAccess)
Definition: tuplesort.c:975
void LogicalTapeSetForgetFreeSpace(LogicalTapeSet *lts)
Definition: logtape.c:609
char * BuildIndexValueDescription(Relation indexRelation, Datum *values, bool *isnull)
Definition: genam.c:177
#define ItemPointerGetBlockNumber(pointer)
Definition: itemptr.h:76
SlabSlot * slabFreeHead
Definition: tuplesort.c:331
long LogicalTapeSetBlocks(LogicalTapeSet *lts)
Definition: logtape.c:1070
void tuplesort_end(Tuplesortstate *state)
Definition: tuplesort.c:1236
#define BTLessStrategyNumber
Definition: stratnum.h:29
static void worker_nomergeruns(Tuplesortstate *state)
Definition: tuplesort.c:4493
static int ApplySortComparator(Datum datum1, bool isNull1, Datum datum2, bool isNull2, SortSupport ssup)
Definition: sortsupport.h:201
Size datumGetSize(Datum value, bool typByVal, int typLen)
Definition: datum.c:61
bool tuplesort_skiptuples(Tuplesortstate *state, int64 ntuples, bool forward)
Definition: tuplesort.c:2283
static void make_bounded_heap(Tuplesortstate *state)
Definition: tuplesort.c:3219
static void tuplesort_heap_replace_top(Tuplesortstate *state, SortTuple *tuple)
Definition: tuplesort.c:3395
int16 AttrNumber
Definition: attnum.h:21
#define MAXORDER
Definition: tuplesort.c:222
static void beginmerge(Tuplesortstate *state)
Definition: tuplesort.c:2851
#define LACKMEM(state)
Definition: tuplesort.c:529
static void inittapestate(Tuplesortstate *state, int maxTapes)
Definition: tuplesort.c:2448
SharedFileSet fileset
Definition: tuplesort.c:490
long val
Definition: informix.c:689
#define offsetof(type, field)
Definition: c.h:611
bool * mergeactive
Definition: tuplesort.c:360
AttrNumber sk_attno
Definition: skey.h:67
TupleDesc tupDesc
Definition: tuplesort.c:413
#define IndexTupleSize(itup)
Definition: itup.h:70
void tuplesort_puttupleslot(Tuplesortstate *state, TupleTableSlot *slot)
Definition: tuplesort.c:1435
SortTuple * memtuples
Definition: tuplesort.c:295
static int ApplySortAbbrevFullComparator(Datum datum1, bool isNull1, Datum datum2, bool isNull2, SortSupport ssup)
Definition: sortsupport.h:239