Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * tuplesort.c
4 : * Generalized tuple sorting routines.
5 : *
6 : * This module provides a generalized facility for tuple sorting, which can be
7 : * applied to different kinds of sortable objects. Implementation of
8 : * the particular sorting variants is given in tuplesortvariants.c.
9 : * This module works efficiently for both small and large amounts
10 : * of data. Small amounts are sorted in-memory using qsort(). Large
11 : * amounts are sorted using temporary files and a standard external sort
12 : * algorithm.
13 : *
14 : * See Knuth, volume 3, for more than you want to know about external
15 : * sorting algorithms. The algorithm we use is a balanced k-way merge.
16 : * Before PostgreSQL 15, we used the polyphase merge algorithm (Knuth's
17 : * Algorithm 5.4.2D), but with modern hardware, a straightforward balanced
18 : * merge is better. Knuth is assuming that tape drives are expensive
19 : * beasts, and in particular that there will always be many more runs than
20 : * tape drives. The polyphase merge algorithm was good at keeping all the
21 : * tape drives busy, but in our implementation a "tape drive" doesn't cost
22 : * much more than a few Kb of memory buffers, so we can afford to have
23 : * lots of them. In particular, if we can have as many tape drives as
24 : * sorted runs, we can eliminate any repeated I/O at all.
25 : *
26 : * Historically, we divided the input into sorted runs using replacement
27 : * selection, in the form of a priority tree implemented as a heap
28 : * (essentially Knuth's Algorithm 5.2.3H), but now we always use quicksort
29 : * for run generation.
30 : *
31 : * The approximate amount of memory allowed for any one sort operation
32 : * is specified in kilobytes by the caller (most pass work_mem). Initially,
33 : * we absorb tuples and simply store them in an unsorted array as long as
34 : * we haven't exceeded workMem. If we reach the end of the input without
35 : * exceeding workMem, we sort the array using qsort() and subsequently return
36 : * tuples just by scanning the tuple array sequentially. If we do exceed
37 : * workMem, we begin to emit tuples into sorted runs in temporary tapes.
38 : * When tuples are dumped in batch after quicksorting, we begin a new run
39 : * with a new output tape. If we reach the max number of tapes, we write
40 : * subsequent runs on the existing tapes in a round-robin fashion. We will
41 : * need multiple merge passes to finish the merge in that case. After the
42 : * end of the input is reached, we dump out remaining tuples in memory into
43 : * a final run, then merge the runs.
44 : *
45 : * When merging runs, we use a heap containing just the frontmost tuple from
46 : * each source run; we repeatedly output the smallest tuple and replace it
47 : * with the next tuple from its source tape (if any). When the heap empties,
48 : * the merge is complete. The basic merge algorithm thus needs very little
49 : * memory --- only M tuples for an M-way merge, and M is constrained to a
50 : * small number. However, we can still make good use of our full workMem
51 : * allocation by pre-reading additional blocks from each source tape. Without
52 : * prereading, our access pattern to the temporary file would be very erratic;
53 : * on average we'd read one block from each of M source tapes during the same
54 : * time that we're writing M blocks to the output tape, so there is no
55 : * sequentiality of access at all, defeating the read-ahead methods used by
56 : * most Unix kernels. Worse, the output tape gets written into a very random
57 : * sequence of blocks of the temp file, ensuring that things will be even
58 : * worse when it comes time to read that tape. A straightforward merge pass
59 : * thus ends up doing a lot of waiting for disk seeks. We can improve matters
60 : * by prereading from each source tape sequentially, loading about workMem/M
61 : * bytes from each tape in turn, and making the sequential blocks immediately
62 : * available for reuse. This approach helps to localize both read and write
63 : * accesses. The pre-reading is handled by logtape.c, we just tell it how
64 : * much memory to use for the buffers.
65 : *
66 : * In the current code we determine the number of input tapes M on the basis
67 : * of workMem: we want workMem/M to be large enough that we read a fair
68 : * amount of data each time we read from a tape, so as to maintain the
69 : * locality of access described above. Nonetheless, with large workMem we
70 : * can have many tapes. The logical "tapes" are implemented by logtape.c,
71 : * which avoids space wastage by recycling disk space as soon as each block
72 : * is read from its "tape".
73 : *
74 : * When the caller requests random access to the sort result, we form
75 : * the final sorted run on a logical tape which is then "frozen", so
76 : * that we can access it randomly. When the caller does not need random
77 : * access, we return from tuplesort_performsort() as soon as we are down
78 : * to one run per logical tape. The final merge is then performed
79 : * on-the-fly as the caller repeatedly calls tuplesort_getXXX; this
80 : * saves one cycle of writing all the data out to disk and reading it in.
81 : *
82 : * This module supports parallel sorting. Parallel sorts involve coordination
83 : * among one or more worker processes, and a leader process, each with its own
84 : * tuplesort state. The leader process (or, more accurately, the
85 : * Tuplesortstate associated with a leader process) creates a full tapeset
86 : * consisting of worker tapes with one run to merge; a run for every
87 : * worker process. This is then merged. Worker processes are guaranteed to
88 : * produce exactly one output run from their partial input.
89 : *
90 : *
91 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
92 : * Portions Copyright (c) 1994, Regents of the University of California
93 : *
94 : * IDENTIFICATION
95 : * src/backend/utils/sort/tuplesort.c
96 : *
97 : *-------------------------------------------------------------------------
98 : */
99 :
100 : #include "postgres.h"
101 :
102 : #include <limits.h>
103 :
104 : #include "commands/tablespace.h"
105 : #include "miscadmin.h"
106 : #include "pg_trace.h"
107 : #include "storage/shmem.h"
108 : #include "utils/memutils.h"
109 : #include "utils/pg_rusage.h"
110 : #include "utils/tuplesort.h"
111 :
112 : /*
113 : * Initial size of memtuples array. We're trying to select this size so that
114 : * array doesn't exceed ALLOCSET_SEPARATE_THRESHOLD and so that the overhead of
115 : * allocation might possibly be lowered. However, we don't consider array sizes
116 : * less than 1024.
117 : *
118 : */
119 : #define INITIAL_MEMTUPSIZE Max(1024, \
120 : ALLOCSET_SEPARATE_THRESHOLD / sizeof(SortTuple) + 1)
121 :
122 : /* GUC variables */
123 : #ifdef TRACE_SORT
124 : bool trace_sort = false;
125 : #endif
126 :
127 : #ifdef DEBUG_BOUNDED_SORT
128 : bool optimize_bounded_sort = true;
129 : #endif
130 :
131 :
132 : /*
133 : * During merge, we use a pre-allocated set of fixed-size slots to hold
134 : * tuples. To avoid palloc/pfree overhead.
135 : *
136 : * Merge doesn't require a lot of memory, so we can afford to waste some,
137 : * by using gratuitously-sized slots. If a tuple is larger than 1 kB, the
138 : * palloc() overhead is not significant anymore.
139 : *
140 : * 'nextfree' is valid when this chunk is in the free list. When in use, the
141 : * slot holds a tuple.
142 : */
143 : #define SLAB_SLOT_SIZE 1024
144 :
145 : typedef union SlabSlot
146 : {
147 : union SlabSlot *nextfree;
148 : char buffer[SLAB_SLOT_SIZE];
149 : } SlabSlot;
150 :
151 : /*
152 : * Possible states of a Tuplesort object. These denote the states that
153 : * persist between calls of Tuplesort routines.
154 : */
155 : typedef enum
156 : {
157 : TSS_INITIAL, /* Loading tuples; still within memory limit */
158 : TSS_BOUNDED, /* Loading tuples into bounded-size heap */
159 : TSS_BUILDRUNS, /* Loading tuples; writing to tape */
160 : TSS_SORTEDINMEM, /* Sort completed entirely in memory */
161 : TSS_SORTEDONTAPE, /* Sort completed, final run is on tape */
162 : TSS_FINALMERGE, /* Performing final merge on-the-fly */
163 : } TupSortStatus;
164 :
165 : /*
166 : * Parameters for calculation of number of tapes to use --- see inittapes()
167 : * and tuplesort_merge_order().
168 : *
169 : * In this calculation we assume that each tape will cost us about 1 blocks
170 : * worth of buffer space. This ignores the overhead of all the other data
171 : * structures needed for each tape, but it's probably close enough.
172 : *
173 : * MERGE_BUFFER_SIZE is how much buffer space we'd like to allocate for each
174 : * input tape, for pre-reading (see discussion at top of file). This is *in
175 : * addition to* the 1 block already included in TAPE_BUFFER_OVERHEAD.
176 : */
177 : #define MINORDER 6 /* minimum merge order */
178 : #define MAXORDER 500 /* maximum merge order */
179 : #define TAPE_BUFFER_OVERHEAD BLCKSZ
180 : #define MERGE_BUFFER_SIZE (BLCKSZ * 32)
181 :
182 :
183 : /*
184 : * Private state of a Tuplesort operation.
185 : */
186 : struct Tuplesortstate
187 : {
188 : TuplesortPublic base;
189 : TupSortStatus status; /* enumerated value as shown above */
190 : bool bounded; /* did caller specify a maximum number of
191 : * tuples to return? */
192 : bool boundUsed; /* true if we made use of a bounded heap */
193 : int bound; /* if bounded, the maximum number of tuples */
194 : int64 tupleMem; /* memory consumed by individual tuples.
195 : * storing this separately from what we track
196 : * in availMem allows us to subtract the
197 : * memory consumed by all tuples when dumping
198 : * tuples to tape */
199 : int64 availMem; /* remaining memory available, in bytes */
200 : int64 allowedMem; /* total memory allowed, in bytes */
201 : int maxTapes; /* max number of input tapes to merge in each
202 : * pass */
203 : int64 maxSpace; /* maximum amount of space occupied among sort
204 : * of groups, either in-memory or on-disk */
205 : bool isMaxSpaceDisk; /* true when maxSpace is value for on-disk
206 : * space, false when it's value for in-memory
207 : * space */
208 : TupSortStatus maxSpaceStatus; /* sort status when maxSpace was reached */
209 : LogicalTapeSet *tapeset; /* logtape.c object for tapes in a temp file */
210 :
211 : /*
212 : * This array holds the tuples now in sort memory. If we are in state
213 : * INITIAL, the tuples are in no particular order; if we are in state
214 : * SORTEDINMEM, the tuples are in final sorted order; in states BUILDRUNS
215 : * and FINALMERGE, the tuples are organized in "heap" order per Algorithm
216 : * H. In state SORTEDONTAPE, the array is not used.
217 : */
218 : SortTuple *memtuples; /* array of SortTuple structs */
219 : int memtupcount; /* number of tuples currently present */
220 : int memtupsize; /* allocated length of memtuples array */
221 : bool growmemtuples; /* memtuples' growth still underway? */
222 :
223 : /*
224 : * Memory for tuples is sometimes allocated using a simple slab allocator,
225 : * rather than with palloc(). Currently, we switch to slab allocation
226 : * when we start merging. Merging only needs to keep a small, fixed
227 : * number of tuples in memory at any time, so we can avoid the
228 : * palloc/pfree overhead by recycling a fixed number of fixed-size slots
229 : * to hold the tuples.
230 : *
231 : * For the slab, we use one large allocation, divided into SLAB_SLOT_SIZE
232 : * slots. The allocation is sized to have one slot per tape, plus one
233 : * additional slot. We need that many slots to hold all the tuples kept
234 : * in the heap during merge, plus the one we have last returned from the
235 : * sort, with tuplesort_gettuple.
236 : *
237 : * Initially, all the slots are kept in a linked list of free slots. When
238 : * a tuple is read from a tape, it is put to the next available slot, if
239 : * it fits. If the tuple is larger than SLAB_SLOT_SIZE, it is palloc'd
240 : * instead.
241 : *
242 : * When we're done processing a tuple, we return the slot back to the free
243 : * list, or pfree() if it was palloc'd. We know that a tuple was
244 : * allocated from the slab, if its pointer value is between
245 : * slabMemoryBegin and -End.
246 : *
247 : * When the slab allocator is used, the USEMEM/LACKMEM mechanism of
248 : * tracking memory usage is not used.
249 : */
250 : bool slabAllocatorUsed;
251 :
252 : char *slabMemoryBegin; /* beginning of slab memory arena */
253 : char *slabMemoryEnd; /* end of slab memory arena */
254 : SlabSlot *slabFreeHead; /* head of free list */
255 :
256 : /* Memory used for input and output tape buffers. */
257 : size_t tape_buffer_mem;
258 :
259 : /*
260 : * When we return a tuple to the caller in tuplesort_gettuple_XXX, that
261 : * came from a tape (that is, in TSS_SORTEDONTAPE or TSS_FINALMERGE
262 : * modes), we remember the tuple in 'lastReturnedTuple', so that we can
263 : * recycle the memory on next gettuple call.
264 : */
265 : void *lastReturnedTuple;
266 :
267 : /*
268 : * While building initial runs, this is the current output run number.
269 : * Afterwards, it is the number of initial runs we made.
270 : */
271 : int currentRun;
272 :
273 : /*
274 : * Logical tapes, for merging.
275 : *
276 : * The initial runs are written in the output tapes. In each merge pass,
277 : * the output tapes of the previous pass become the input tapes, and new
278 : * output tapes are created as needed. When nInputTapes equals
279 : * nInputRuns, there is only one merge pass left.
280 : */
281 : LogicalTape **inputTapes;
282 : int nInputTapes;
283 : int nInputRuns;
284 :
285 : LogicalTape **outputTapes;
286 : int nOutputTapes;
287 : int nOutputRuns;
288 :
289 : LogicalTape *destTape; /* current output tape */
290 :
291 : /*
292 : * These variables are used after completion of sorting to keep track of
293 : * the next tuple to return. (In the tape case, the tape's current read
294 : * position is also critical state.)
295 : */
296 : LogicalTape *result_tape; /* actual tape of finished output */
297 : int current; /* array index (only used if SORTEDINMEM) */
298 : bool eof_reached; /* reached EOF (needed for cursors) */
299 :
300 : /* markpos_xxx holds marked position for mark and restore */
301 : int64 markpos_block; /* tape block# (only used if SORTEDONTAPE) */
302 : int markpos_offset; /* saved "current", or offset in tape block */
303 : bool markpos_eof; /* saved "eof_reached" */
304 :
305 : /*
306 : * These variables are used during parallel sorting.
307 : *
308 : * worker is our worker identifier. Follows the general convention that
309 : * -1 value relates to a leader tuplesort, and values >= 0 worker
310 : * tuplesorts. (-1 can also be a serial tuplesort.)
311 : *
312 : * shared is mutable shared memory state, which is used to coordinate
313 : * parallel sorts.
314 : *
315 : * nParticipants is the number of worker Tuplesortstates known by the
316 : * leader to have actually been launched, which implies that they must
317 : * finish a run that the leader needs to merge. Typically includes a
318 : * worker state held by the leader process itself. Set in the leader
319 : * Tuplesortstate only.
320 : */
321 : int worker;
322 : Sharedsort *shared;
323 : int nParticipants;
324 :
325 : /*
326 : * Additional state for managing "abbreviated key" sortsupport routines
327 : * (which currently may be used by all cases except the hash index case).
328 : * Tracks the intervals at which the optimization's effectiveness is
329 : * tested.
330 : */
331 : int64 abbrevNext; /* Tuple # at which to next check
332 : * applicability */
333 :
334 : /*
335 : * Resource snapshot for time of sort start.
336 : */
337 : #ifdef TRACE_SORT
338 : PGRUsage ru_start;
339 : #endif
340 : };
341 :
342 : /*
343 : * Private mutable state of tuplesort-parallel-operation. This is allocated
344 : * in shared memory.
345 : */
346 : struct Sharedsort
347 : {
348 : /* mutex protects all fields prior to tapes */
349 : slock_t mutex;
350 :
351 : /*
352 : * currentWorker generates ordinal identifier numbers for parallel sort
353 : * workers. These start from 0, and are always gapless.
354 : *
355 : * Workers increment workersFinished to indicate having finished. If this
356 : * is equal to state.nParticipants within the leader, leader is ready to
357 : * merge worker runs.
358 : */
359 : int currentWorker;
360 : int workersFinished;
361 :
362 : /* Temporary file space */
363 : SharedFileSet fileset;
364 :
365 : /* Size of tapes flexible array */
366 : int nTapes;
367 :
368 : /*
369 : * Tapes array used by workers to report back information needed by the
370 : * leader to concatenate all worker tapes into one for merging
371 : */
372 : TapeShare tapes[FLEXIBLE_ARRAY_MEMBER];
373 : };
374 :
375 : /*
376 : * Is the given tuple allocated from the slab memory arena?
377 : */
378 : #define IS_SLAB_SLOT(state, tuple) \
379 : ((char *) (tuple) >= (state)->slabMemoryBegin && \
380 : (char *) (tuple) < (state)->slabMemoryEnd)
381 :
382 : /*
383 : * Return the given tuple to the slab memory free list, or free it
384 : * if it was palloc'd.
385 : */
386 : #define RELEASE_SLAB_SLOT(state, tuple) \
387 : do { \
388 : SlabSlot *buf = (SlabSlot *) tuple; \
389 : \
390 : if (IS_SLAB_SLOT((state), buf)) \
391 : { \
392 : buf->nextfree = (state)->slabFreeHead; \
393 : (state)->slabFreeHead = buf; \
394 : } else \
395 : pfree(buf); \
396 : } while(0)
397 :
398 : #define REMOVEABBREV(state,stup,count) ((*(state)->base.removeabbrev) (state, stup, count))
399 : #define COMPARETUP(state,a,b) ((*(state)->base.comparetup) (a, b, state))
400 : #define WRITETUP(state,tape,stup) ((*(state)->base.writetup) (state, tape, stup))
401 : #define READTUP(state,stup,tape,len) ((*(state)->base.readtup) (state, stup, tape, len))
402 : #define FREESTATE(state) ((state)->base.freestate ? (*(state)->base.freestate) (state) : (void) 0)
403 : #define LACKMEM(state) ((state)->availMem < 0 && !(state)->slabAllocatorUsed)
404 : #define USEMEM(state,amt) ((state)->availMem -= (amt))
405 : #define FREEMEM(state,amt) ((state)->availMem += (amt))
406 : #define SERIAL(state) ((state)->shared == NULL)
407 : #define WORKER(state) ((state)->shared && (state)->worker != -1)
408 : #define LEADER(state) ((state)->shared && (state)->worker == -1)
409 :
410 : /*
411 : * NOTES about on-tape representation of tuples:
412 : *
413 : * We require the first "unsigned int" of a stored tuple to be the total size
414 : * on-tape of the tuple, including itself (so it is never zero; an all-zero
415 : * unsigned int is used to delimit runs). The remainder of the stored tuple
416 : * may or may not match the in-memory representation of the tuple ---
417 : * any conversion needed is the job of the writetup and readtup routines.
418 : *
419 : * If state->sortopt contains TUPLESORT_RANDOMACCESS, then the stored
420 : * representation of the tuple must be followed by another "unsigned int" that
421 : * is a copy of the length --- so the total tape space used is actually
422 : * sizeof(unsigned int) more than the stored length value. This allows
423 : * read-backwards. When the random access flag was not specified, the
424 : * write/read routines may omit the extra length word.
425 : *
426 : * writetup is expected to write both length words as well as the tuple
427 : * data. When readtup is called, the tape is positioned just after the
428 : * front length word; readtup must read the tuple data and advance past
429 : * the back length word (if present).
430 : *
431 : * The write/read routines can make use of the tuple description data
432 : * stored in the Tuplesortstate record, if needed. They are also expected
433 : * to adjust state->availMem by the amount of memory space (not tape space!)
434 : * released or consumed. There is no error return from either writetup
435 : * or readtup; they should ereport() on failure.
436 : *
437 : *
438 : * NOTES about memory consumption calculations:
439 : *
440 : * We count space allocated for tuples against the workMem limit, plus
441 : * the space used by the variable-size memtuples array. Fixed-size space
442 : * is not counted; it's small enough to not be interesting.
443 : *
444 : * Note that we count actual space used (as shown by GetMemoryChunkSpace)
445 : * rather than the originally-requested size. This is important since
446 : * palloc can add substantial overhead. It's not a complete answer since
447 : * we won't count any wasted space in palloc allocation blocks, but it's
448 : * a lot better than what we were doing before 7.3. As of 9.6, a
449 : * separate memory context is used for caller passed tuples. Resetting
450 : * it at certain key increments significantly ameliorates fragmentation.
451 : * readtup routines use the slab allocator (they cannot use
452 : * the reset context because it gets deleted at the point that merging
453 : * begins).
454 : */
455 :
456 :
457 : static void tuplesort_begin_batch(Tuplesortstate *state);
458 : static bool consider_abort_common(Tuplesortstate *state);
459 : static void inittapes(Tuplesortstate *state, bool mergeruns);
460 : static void inittapestate(Tuplesortstate *state, int maxTapes);
461 : static void selectnewtape(Tuplesortstate *state);
462 : static void init_slab_allocator(Tuplesortstate *state, int numSlots);
463 : static void mergeruns(Tuplesortstate *state);
464 : static void mergeonerun(Tuplesortstate *state);
465 : static void beginmerge(Tuplesortstate *state);
466 : static bool mergereadnext(Tuplesortstate *state, LogicalTape *srcTape, SortTuple *stup);
467 : static void dumptuples(Tuplesortstate *state, bool alltuples);
468 : static void make_bounded_heap(Tuplesortstate *state);
469 : static void sort_bounded_heap(Tuplesortstate *state);
470 : static void tuplesort_sort_memtuples(Tuplesortstate *state);
471 : static void tuplesort_heap_insert(Tuplesortstate *state, SortTuple *tuple);
472 : static void tuplesort_heap_replace_top(Tuplesortstate *state, SortTuple *tuple);
473 : static void tuplesort_heap_delete_top(Tuplesortstate *state);
474 : static void reversedirection(Tuplesortstate *state);
475 : static unsigned int getlen(LogicalTape *tape, bool eofOK);
476 : static void markrunend(LogicalTape *tape);
477 : static int worker_get_identifier(Tuplesortstate *state);
478 : static void worker_freeze_result_tape(Tuplesortstate *state);
479 : static void worker_nomergeruns(Tuplesortstate *state);
480 : static void leader_takeover_tapes(Tuplesortstate *state);
481 : static void free_sort_tuple(Tuplesortstate *state, SortTuple *stup);
482 : static void tuplesort_free(Tuplesortstate *state);
483 : static void tuplesort_updatemax(Tuplesortstate *state);
484 :
485 : /*
486 : * Specialized comparators that we can inline into specialized sorts. The goal
487 : * is to try to sort two tuples without having to follow the pointers to the
488 : * comparator or the tuple.
489 : *
490 : * XXX: For now, there is no specialization for cases where datum1 is
491 : * authoritative and we don't even need to fall back to a callback at all (that
492 : * would be true for types like int4/int8/timestamp/date, but not true for
493 : * abbreviations of text or multi-key sorts. There could be! Is it worth it?
494 : */
495 :
496 : /* Used if first key's comparator is ssup_datum_unsigned_cmp */
497 : static pg_attribute_always_inline int
498 46409796 : qsort_tuple_unsigned_compare(SortTuple *a, SortTuple *b, Tuplesortstate *state)
499 : {
500 : int compare;
501 :
502 46409796 : compare = ApplyUnsignedSortComparator(a->datum1, a->isnull1,
503 46409796 : b->datum1, b->isnull1,
504 : &state->base.sortKeys[0]);
505 46409796 : if (compare != 0)
506 41876510 : return compare;
507 :
508 : /*
509 : * No need to waste effort calling the tiebreak function when there are no
510 : * other keys to sort on.
511 : */
512 4533286 : if (state->base.onlyKey != NULL)
513 0 : return 0;
514 :
515 4533286 : return state->base.comparetup_tiebreak(a, b, state);
516 : }
517 :
518 : #if SIZEOF_DATUM >= 8
519 : /* Used if first key's comparator is ssup_datum_signed_cmp */
520 : static pg_attribute_always_inline int
521 5627272 : qsort_tuple_signed_compare(SortTuple *a, SortTuple *b, Tuplesortstate *state)
522 : {
523 : int compare;
524 :
525 5627272 : compare = ApplySignedSortComparator(a->datum1, a->isnull1,
526 5627272 : b->datum1, b->isnull1,
527 : &state->base.sortKeys[0]);
528 :
529 5627272 : if (compare != 0)
530 5615326 : return compare;
531 :
532 : /*
533 : * No need to waste effort calling the tiebreak function when there are no
534 : * other keys to sort on.
535 : */
536 11946 : if (state->base.onlyKey != NULL)
537 1102 : return 0;
538 :
539 10844 : return state->base.comparetup_tiebreak(a, b, state);
540 : }
541 : #endif
542 :
543 : /* Used if first key's comparator is ssup_datum_int32_cmp */
544 : static pg_attribute_always_inline int
545 52075236 : qsort_tuple_int32_compare(SortTuple *a, SortTuple *b, Tuplesortstate *state)
546 : {
547 : int compare;
548 :
549 52075236 : compare = ApplyInt32SortComparator(a->datum1, a->isnull1,
550 52075236 : b->datum1, b->isnull1,
551 : &state->base.sortKeys[0]);
552 :
553 52075236 : if (compare != 0)
554 36763318 : return compare;
555 :
556 : /*
557 : * No need to waste effort calling the tiebreak function when there are no
558 : * other keys to sort on.
559 : */
560 15311918 : if (state->base.onlyKey != NULL)
561 1677194 : return 0;
562 :
563 13634724 : return state->base.comparetup_tiebreak(a, b, state);
564 : }
565 :
566 : /*
567 : * Special versions of qsort just for SortTuple objects. qsort_tuple() sorts
568 : * any variant of SortTuples, using the appropriate comparetup function.
569 : * qsort_ssup() is specialized for the case where the comparetup function
570 : * reduces to ApplySortComparator(), that is single-key MinimalTuple sorts
571 : * and Datum sorts. qsort_tuple_{unsigned,signed,int32} are specialized for
572 : * common comparison functions on pass-by-value leading datums.
573 : */
574 :
575 : #define ST_SORT qsort_tuple_unsigned
576 : #define ST_ELEMENT_TYPE SortTuple
577 : #define ST_COMPARE(a, b, state) qsort_tuple_unsigned_compare(a, b, state)
578 : #define ST_COMPARE_ARG_TYPE Tuplesortstate
579 : #define ST_CHECK_FOR_INTERRUPTS
580 : #define ST_SCOPE static
581 : #define ST_DEFINE
582 : #include "lib/sort_template.h"
583 :
584 : #if SIZEOF_DATUM >= 8
585 : #define ST_SORT qsort_tuple_signed
586 : #define ST_ELEMENT_TYPE SortTuple
587 : #define ST_COMPARE(a, b, state) qsort_tuple_signed_compare(a, b, state)
588 : #define ST_COMPARE_ARG_TYPE Tuplesortstate
589 : #define ST_CHECK_FOR_INTERRUPTS
590 : #define ST_SCOPE static
591 : #define ST_DEFINE
592 : #include "lib/sort_template.h"
593 : #endif
594 :
595 : #define ST_SORT qsort_tuple_int32
596 : #define ST_ELEMENT_TYPE SortTuple
597 : #define ST_COMPARE(a, b, state) qsort_tuple_int32_compare(a, b, state)
598 : #define ST_COMPARE_ARG_TYPE Tuplesortstate
599 : #define ST_CHECK_FOR_INTERRUPTS
600 : #define ST_SCOPE static
601 : #define ST_DEFINE
602 : #include "lib/sort_template.h"
603 :
604 : #define ST_SORT qsort_tuple
605 : #define ST_ELEMENT_TYPE SortTuple
606 : #define ST_COMPARE_RUNTIME_POINTER
607 : #define ST_COMPARE_ARG_TYPE Tuplesortstate
608 : #define ST_CHECK_FOR_INTERRUPTS
609 : #define ST_SCOPE static
610 : #define ST_DECLARE
611 : #define ST_DEFINE
612 : #include "lib/sort_template.h"
613 :
614 : #define ST_SORT qsort_ssup
615 : #define ST_ELEMENT_TYPE SortTuple
616 : #define ST_COMPARE(a, b, ssup) \
617 : ApplySortComparator((a)->datum1, (a)->isnull1, \
618 : (b)->datum1, (b)->isnull1, (ssup))
619 : #define ST_COMPARE_ARG_TYPE SortSupportData
620 : #define ST_CHECK_FOR_INTERRUPTS
621 : #define ST_SCOPE static
622 : #define ST_DEFINE
623 : #include "lib/sort_template.h"
624 :
625 : /*
626 : * tuplesort_begin_xxx
627 : *
628 : * Initialize for a tuple sort operation.
629 : *
630 : * After calling tuplesort_begin, the caller should call tuplesort_putXXX
631 : * zero or more times, then call tuplesort_performsort when all the tuples
632 : * have been supplied. After performsort, retrieve the tuples in sorted
633 : * order by calling tuplesort_getXXX until it returns false/NULL. (If random
634 : * access was requested, rescan, markpos, and restorepos can also be called.)
635 : * Call tuplesort_end to terminate the operation and release memory/disk space.
636 : *
637 : * Each variant of tuplesort_begin has a workMem parameter specifying the
638 : * maximum number of kilobytes of RAM to use before spilling data to disk.
639 : * (The normal value of this parameter is work_mem, but some callers use
640 : * other values.) Each variant also has a sortopt which is a bitmask of
641 : * sort options. See TUPLESORT_* definitions in tuplesort.h
642 : */
643 :
644 : Tuplesortstate *
645 234726 : tuplesort_begin_common(int workMem, SortCoordinate coordinate, int sortopt)
646 : {
647 : Tuplesortstate *state;
648 : MemoryContext maincontext;
649 : MemoryContext sortcontext;
650 : MemoryContext oldcontext;
651 :
652 : /* See leader_takeover_tapes() remarks on random access support */
653 234726 : if (coordinate && (sortopt & TUPLESORT_RANDOMACCESS))
654 0 : elog(ERROR, "random access disallowed under parallel sort");
655 :
656 : /*
657 : * Memory context surviving tuplesort_reset. This memory context holds
658 : * data which is useful to keep while sorting multiple similar batches.
659 : */
660 234726 : maincontext = AllocSetContextCreate(CurrentMemoryContext,
661 : "TupleSort main",
662 : ALLOCSET_DEFAULT_SIZES);
663 :
664 : /*
665 : * Create a working memory context for one sort operation. The content of
666 : * this context is deleted by tuplesort_reset.
667 : */
668 234726 : sortcontext = AllocSetContextCreate(maincontext,
669 : "TupleSort sort",
670 : ALLOCSET_DEFAULT_SIZES);
671 :
672 : /*
673 : * Additionally a working memory context for tuples is setup in
674 : * tuplesort_begin_batch.
675 : */
676 :
677 : /*
678 : * Make the Tuplesortstate within the per-sortstate context. This way, we
679 : * don't need a separate pfree() operation for it at shutdown.
680 : */
681 234726 : oldcontext = MemoryContextSwitchTo(maincontext);
682 :
683 234726 : state = (Tuplesortstate *) palloc0(sizeof(Tuplesortstate));
684 :
685 : #ifdef TRACE_SORT
686 234726 : if (trace_sort)
687 0 : pg_rusage_init(&state->ru_start);
688 : #endif
689 :
690 234726 : state->base.sortopt = sortopt;
691 234726 : state->base.tuples = true;
692 234726 : state->abbrevNext = 10;
693 :
694 : /*
695 : * workMem is forced to be at least 64KB, the current minimum valid value
696 : * for the work_mem GUC. This is a defense against parallel sort callers
697 : * that divide out memory among many workers in a way that leaves each
698 : * with very little memory.
699 : */
700 234726 : state->allowedMem = Max(workMem, 64) * (int64) 1024;
701 234726 : state->base.sortcontext = sortcontext;
702 234726 : state->base.maincontext = maincontext;
703 :
704 : /*
705 : * Initial size of array must be more than ALLOCSET_SEPARATE_THRESHOLD;
706 : * see comments in grow_memtuples().
707 : */
708 234726 : state->memtupsize = INITIAL_MEMTUPSIZE;
709 234726 : state->memtuples = NULL;
710 :
711 : /*
712 : * After all of the other non-parallel-related state, we setup all of the
713 : * state needed for each batch.
714 : */
715 234726 : tuplesort_begin_batch(state);
716 :
717 : /*
718 : * Initialize parallel-related state based on coordination information
719 : * from caller
720 : */
721 234726 : if (!coordinate)
722 : {
723 : /* Serial sort */
724 234098 : state->shared = NULL;
725 234098 : state->worker = -1;
726 234098 : state->nParticipants = -1;
727 : }
728 628 : else if (coordinate->isWorker)
729 : {
730 : /* Parallel worker produces exactly one final run from all input */
731 420 : state->shared = coordinate->sharedsort;
732 420 : state->worker = worker_get_identifier(state);
733 420 : state->nParticipants = -1;
734 : }
735 : else
736 : {
737 : /* Parallel leader state only used for final merge */
738 208 : state->shared = coordinate->sharedsort;
739 208 : state->worker = -1;
740 208 : state->nParticipants = coordinate->nParticipants;
741 : Assert(state->nParticipants >= 1);
742 : }
743 :
744 234726 : MemoryContextSwitchTo(oldcontext);
745 :
746 234726 : return state;
747 : }
748 :
749 : /*
750 : * tuplesort_begin_batch
751 : *
752 : * Setup, or reset, all state need for processing a new set of tuples with this
753 : * sort state. Called both from tuplesort_begin_common (the first time sorting
754 : * with this sort state) and tuplesort_reset (for subsequent usages).
755 : */
756 : static void
757 236894 : tuplesort_begin_batch(Tuplesortstate *state)
758 : {
759 : MemoryContext oldcontext;
760 :
761 236894 : oldcontext = MemoryContextSwitchTo(state->base.maincontext);
762 :
763 : /*
764 : * Caller tuple (e.g. IndexTuple) memory context.
765 : *
766 : * A dedicated child context used exclusively for caller passed tuples
767 : * eases memory management. Resetting at key points reduces
768 : * fragmentation. Note that the memtuples array of SortTuples is allocated
769 : * in the parent context, not this context, because there is no need to
770 : * free memtuples early. For bounded sorts, tuples may be pfreed in any
771 : * order, so we use a regular aset.c context so that it can make use of
772 : * free'd memory. When the sort is not bounded, we make use of a bump.c
773 : * context as this keeps allocations more compact with less wastage.
774 : * Allocations are also slightly more CPU efficient.
775 : */
776 236894 : if (TupleSortUseBumpTupleCxt(state->base.sortopt))
777 235580 : state->base.tuplecontext = BumpContextCreate(state->base.sortcontext,
778 : "Caller tuples",
779 : ALLOCSET_DEFAULT_SIZES);
780 : else
781 1314 : state->base.tuplecontext = AllocSetContextCreate(state->base.sortcontext,
782 : "Caller tuples",
783 : ALLOCSET_DEFAULT_SIZES);
784 :
785 :
786 236894 : state->status = TSS_INITIAL;
787 236894 : state->bounded = false;
788 236894 : state->boundUsed = false;
789 :
790 236894 : state->availMem = state->allowedMem;
791 :
792 236894 : state->tapeset = NULL;
793 :
794 236894 : state->memtupcount = 0;
795 :
796 : /*
797 : * Initial size of array must be more than ALLOCSET_SEPARATE_THRESHOLD;
798 : * see comments in grow_memtuples().
799 : */
800 236894 : state->growmemtuples = true;
801 236894 : state->slabAllocatorUsed = false;
802 236894 : if (state->memtuples != NULL && state->memtupsize != INITIAL_MEMTUPSIZE)
803 : {
804 0 : pfree(state->memtuples);
805 0 : state->memtuples = NULL;
806 0 : state->memtupsize = INITIAL_MEMTUPSIZE;
807 : }
808 236894 : if (state->memtuples == NULL)
809 : {
810 234726 : state->memtuples = (SortTuple *) palloc(state->memtupsize * sizeof(SortTuple));
811 234726 : USEMEM(state, GetMemoryChunkSpace(state->memtuples));
812 : }
813 :
814 : /* workMem must be large enough for the minimal memtuples array */
815 236894 : if (LACKMEM(state))
816 0 : elog(ERROR, "insufficient memory allowed for sort");
817 :
818 236894 : state->currentRun = 0;
819 :
820 : /*
821 : * Tape variables (inputTapes, outputTapes, etc.) will be initialized by
822 : * inittapes(), if needed.
823 : */
824 :
825 236894 : state->result_tape = NULL; /* flag that result tape has not been formed */
826 :
827 236894 : MemoryContextSwitchTo(oldcontext);
828 236894 : }
829 :
830 : /*
831 : * tuplesort_set_bound
832 : *
833 : * Advise tuplesort that at most the first N result tuples are required.
834 : *
835 : * Must be called before inserting any tuples. (Actually, we could allow it
836 : * as long as the sort hasn't spilled to disk, but there seems no need for
837 : * delayed calls at the moment.)
838 : *
839 : * This is a hint only. The tuplesort may still return more tuples than
840 : * requested. Parallel leader tuplesorts will always ignore the hint.
841 : */
842 : void
843 1180 : tuplesort_set_bound(Tuplesortstate *state, int64 bound)
844 : {
845 : /* Assert we're called before loading any tuples */
846 : Assert(state->status == TSS_INITIAL && state->memtupcount == 0);
847 : /* Assert we allow bounded sorts */
848 : Assert(state->base.sortopt & TUPLESORT_ALLOWBOUNDED);
849 : /* Can't set the bound twice, either */
850 : Assert(!state->bounded);
851 : /* Also, this shouldn't be called in a parallel worker */
852 : Assert(!WORKER(state));
853 :
854 : /* Parallel leader allows but ignores hint */
855 1180 : if (LEADER(state))
856 0 : return;
857 :
858 : #ifdef DEBUG_BOUNDED_SORT
859 : /* Honor GUC setting that disables the feature (for easy testing) */
860 : if (!optimize_bounded_sort)
861 : return;
862 : #endif
863 :
864 : /* We want to be able to compute bound * 2, so limit the setting */
865 1180 : if (bound > (int64) (INT_MAX / 2))
866 0 : return;
867 :
868 1180 : state->bounded = true;
869 1180 : state->bound = (int) bound;
870 :
871 : /*
872 : * Bounded sorts are not an effective target for abbreviated key
873 : * optimization. Disable by setting state to be consistent with no
874 : * abbreviation support.
875 : */
876 1180 : state->base.sortKeys->abbrev_converter = NULL;
877 1180 : if (state->base.sortKeys->abbrev_full_comparator)
878 16 : state->base.sortKeys->comparator = state->base.sortKeys->abbrev_full_comparator;
879 :
880 : /* Not strictly necessary, but be tidy */
881 1180 : state->base.sortKeys->abbrev_abort = NULL;
882 1180 : state->base.sortKeys->abbrev_full_comparator = NULL;
883 : }
884 :
885 : /*
886 : * tuplesort_used_bound
887 : *
888 : * Allow callers to find out if the sort state was able to use a bound.
889 : */
890 : bool
891 94 : tuplesort_used_bound(Tuplesortstate *state)
892 : {
893 94 : return state->boundUsed;
894 : }
895 :
896 : /*
897 : * tuplesort_free
898 : *
899 : * Internal routine for freeing resources of tuplesort.
900 : */
901 : static void
902 236650 : tuplesort_free(Tuplesortstate *state)
903 : {
904 : /* context swap probably not needed, but let's be safe */
905 236650 : MemoryContext oldcontext = MemoryContextSwitchTo(state->base.sortcontext);
906 :
907 : #ifdef TRACE_SORT
908 : int64 spaceUsed;
909 :
910 236650 : if (state->tapeset)
911 678 : spaceUsed = LogicalTapeSetBlocks(state->tapeset);
912 : else
913 235972 : spaceUsed = (state->allowedMem - state->availMem + 1023) / 1024;
914 : #endif
915 :
916 : /*
917 : * Delete temporary "tape" files, if any.
918 : *
919 : * Note: want to include this in reported total cost of sort, hence need
920 : * for two #ifdef TRACE_SORT sections.
921 : *
922 : * We don't bother to destroy the individual tapes here. They will go away
923 : * with the sortcontext. (In TSS_FINALMERGE state, we have closed
924 : * finished tapes already.)
925 : */
926 236650 : if (state->tapeset)
927 678 : LogicalTapeSetClose(state->tapeset);
928 :
929 : #ifdef TRACE_SORT
930 236650 : if (trace_sort)
931 : {
932 0 : if (state->tapeset)
933 0 : elog(LOG, "%s of worker %d ended, %lld disk blocks used: %s",
934 : SERIAL(state) ? "external sort" : "parallel external sort",
935 : state->worker, (long long) spaceUsed, pg_rusage_show(&state->ru_start));
936 : else
937 0 : elog(LOG, "%s of worker %d ended, %lld KB used: %s",
938 : SERIAL(state) ? "internal sort" : "unperformed parallel sort",
939 : state->worker, (long long) spaceUsed, pg_rusage_show(&state->ru_start));
940 : }
941 :
942 : TRACE_POSTGRESQL_SORT_DONE(state->tapeset != NULL, spaceUsed);
943 : #else
944 :
945 : /*
946 : * If you disabled TRACE_SORT, you can still probe sort__done, but you
947 : * ain't getting space-used stats.
948 : */
949 : TRACE_POSTGRESQL_SORT_DONE(state->tapeset != NULL, 0L);
950 : #endif
951 :
952 236650 : FREESTATE(state);
953 236650 : MemoryContextSwitchTo(oldcontext);
954 :
955 : /*
956 : * Free the per-sort memory context, thereby releasing all working memory.
957 : */
958 236650 : MemoryContextReset(state->base.sortcontext);
959 236650 : }
960 :
961 : /*
962 : * tuplesort_end
963 : *
964 : * Release resources and clean up.
965 : *
966 : * NOTE: after calling this, any pointers returned by tuplesort_getXXX are
967 : * pointing to garbage. Be careful not to attempt to use or free such
968 : * pointers afterwards!
969 : */
970 : void
971 234482 : tuplesort_end(Tuplesortstate *state)
972 : {
973 234482 : tuplesort_free(state);
974 :
975 : /*
976 : * Free the main memory context, including the Tuplesortstate struct
977 : * itself.
978 : */
979 234482 : MemoryContextDelete(state->base.maincontext);
980 234482 : }
981 :
982 : /*
983 : * tuplesort_updatemax
984 : *
985 : * Update maximum resource usage statistics.
986 : */
987 : static void
988 2552 : tuplesort_updatemax(Tuplesortstate *state)
989 : {
990 : int64 spaceUsed;
991 : bool isSpaceDisk;
992 :
993 : /*
994 : * Note: it might seem we should provide both memory and disk usage for a
995 : * disk-based sort. However, the current code doesn't track memory space
996 : * accurately once we have begun to return tuples to the caller (since we
997 : * don't account for pfree's the caller is expected to do), so we cannot
998 : * rely on availMem in a disk sort. This does not seem worth the overhead
999 : * to fix. Is it worth creating an API for the memory context code to
1000 : * tell us how much is actually used in sortcontext?
1001 : */
1002 2552 : if (state->tapeset)
1003 : {
1004 0 : isSpaceDisk = true;
1005 0 : spaceUsed = LogicalTapeSetBlocks(state->tapeset) * BLCKSZ;
1006 : }
1007 : else
1008 : {
1009 2552 : isSpaceDisk = false;
1010 2552 : spaceUsed = state->allowedMem - state->availMem;
1011 : }
1012 :
1013 : /*
1014 : * Sort evicts data to the disk when it wasn't able to fit that data into
1015 : * main memory. This is why we assume space used on the disk to be more
1016 : * important for tracking resource usage than space used in memory. Note
1017 : * that the amount of space occupied by some tupleset on the disk might be
1018 : * less than amount of space occupied by the same tupleset in memory due
1019 : * to more compact representation.
1020 : */
1021 2552 : if ((isSpaceDisk && !state->isMaxSpaceDisk) ||
1022 2552 : (isSpaceDisk == state->isMaxSpaceDisk && spaceUsed > state->maxSpace))
1023 : {
1024 382 : state->maxSpace = spaceUsed;
1025 382 : state->isMaxSpaceDisk = isSpaceDisk;
1026 382 : state->maxSpaceStatus = state->status;
1027 : }
1028 2552 : }
1029 :
1030 : /*
1031 : * tuplesort_reset
1032 : *
1033 : * Reset the tuplesort. Reset all the data in the tuplesort, but leave the
1034 : * meta-information in. After tuplesort_reset, tuplesort is ready to start
1035 : * a new sort. This allows avoiding recreation of tuple sort states (and
1036 : * save resources) when sorting multiple small batches.
1037 : */
1038 : void
1039 2168 : tuplesort_reset(Tuplesortstate *state)
1040 : {
1041 2168 : tuplesort_updatemax(state);
1042 2168 : tuplesort_free(state);
1043 :
1044 : /*
1045 : * After we've freed up per-batch memory, re-setup all of the state common
1046 : * to both the first batch and any subsequent batch.
1047 : */
1048 2168 : tuplesort_begin_batch(state);
1049 :
1050 2168 : state->lastReturnedTuple = NULL;
1051 2168 : state->slabMemoryBegin = NULL;
1052 2168 : state->slabMemoryEnd = NULL;
1053 2168 : state->slabFreeHead = NULL;
1054 2168 : }
1055 :
1056 : /*
1057 : * Grow the memtuples[] array, if possible within our memory constraint. We
1058 : * must not exceed INT_MAX tuples in memory or the caller-provided memory
1059 : * limit. Return true if we were able to enlarge the array, false if not.
1060 : *
1061 : * Normally, at each increment we double the size of the array. When doing
1062 : * that would exceed a limit, we attempt one last, smaller increase (and then
1063 : * clear the growmemtuples flag so we don't try any more). That allows us to
1064 : * use memory as fully as permitted; sticking to the pure doubling rule could
1065 : * result in almost half going unused. Because availMem moves around with
1066 : * tuple addition/removal, we need some rule to prevent making repeated small
1067 : * increases in memtupsize, which would just be useless thrashing. The
1068 : * growmemtuples flag accomplishes that and also prevents useless
1069 : * recalculations in this function.
1070 : */
1071 : static bool
1072 6942 : grow_memtuples(Tuplesortstate *state)
1073 : {
1074 : int newmemtupsize;
1075 6942 : int memtupsize = state->memtupsize;
1076 6942 : int64 memNowUsed = state->allowedMem - state->availMem;
1077 :
1078 : /* Forget it if we've already maxed out memtuples, per comment above */
1079 6942 : if (!state->growmemtuples)
1080 114 : return false;
1081 :
1082 : /* Select new value of memtupsize */
1083 6828 : if (memNowUsed <= state->availMem)
1084 : {
1085 : /*
1086 : * We've used no more than half of allowedMem; double our usage,
1087 : * clamping at INT_MAX tuples.
1088 : */
1089 6706 : if (memtupsize < INT_MAX / 2)
1090 6706 : newmemtupsize = memtupsize * 2;
1091 : else
1092 : {
1093 0 : newmemtupsize = INT_MAX;
1094 0 : state->growmemtuples = false;
1095 : }
1096 : }
1097 : else
1098 : {
1099 : /*
1100 : * This will be the last increment of memtupsize. Abandon doubling
1101 : * strategy and instead increase as much as we safely can.
1102 : *
1103 : * To stay within allowedMem, we can't increase memtupsize by more
1104 : * than availMem / sizeof(SortTuple) elements. In practice, we want
1105 : * to increase it by considerably less, because we need to leave some
1106 : * space for the tuples to which the new array slots will refer. We
1107 : * assume the new tuples will be about the same size as the tuples
1108 : * we've already seen, and thus we can extrapolate from the space
1109 : * consumption so far to estimate an appropriate new size for the
1110 : * memtuples array. The optimal value might be higher or lower than
1111 : * this estimate, but it's hard to know that in advance. We again
1112 : * clamp at INT_MAX tuples.
1113 : *
1114 : * This calculation is safe against enlarging the array so much that
1115 : * LACKMEM becomes true, because the memory currently used includes
1116 : * the present array; thus, there would be enough allowedMem for the
1117 : * new array elements even if no other memory were currently used.
1118 : *
1119 : * We do the arithmetic in float8, because otherwise the product of
1120 : * memtupsize and allowedMem could overflow. Any inaccuracy in the
1121 : * result should be insignificant; but even if we computed a
1122 : * completely insane result, the checks below will prevent anything
1123 : * really bad from happening.
1124 : */
1125 : double grow_ratio;
1126 :
1127 122 : grow_ratio = (double) state->allowedMem / (double) memNowUsed;
1128 122 : if (memtupsize * grow_ratio < INT_MAX)
1129 122 : newmemtupsize = (int) (memtupsize * grow_ratio);
1130 : else
1131 0 : newmemtupsize = INT_MAX;
1132 :
1133 : /* We won't make any further enlargement attempts */
1134 122 : state->growmemtuples = false;
1135 : }
1136 :
1137 : /* Must enlarge array by at least one element, else report failure */
1138 6828 : if (newmemtupsize <= memtupsize)
1139 0 : goto noalloc;
1140 :
1141 : /*
1142 : * On a 32-bit machine, allowedMem could exceed MaxAllocHugeSize. Clamp
1143 : * to ensure our request won't be rejected. Note that we can easily
1144 : * exhaust address space before facing this outcome. (This is presently
1145 : * impossible due to guc.c's MAX_KILOBYTES limitation on work_mem, but
1146 : * don't rely on that at this distance.)
1147 : */
1148 6828 : if ((Size) newmemtupsize >= MaxAllocHugeSize / sizeof(SortTuple))
1149 : {
1150 0 : newmemtupsize = (int) (MaxAllocHugeSize / sizeof(SortTuple));
1151 0 : state->growmemtuples = false; /* can't grow any more */
1152 : }
1153 :
1154 : /*
1155 : * We need to be sure that we do not cause LACKMEM to become true, else
1156 : * the space management algorithm will go nuts. The code above should
1157 : * never generate a dangerous request, but to be safe, check explicitly
1158 : * that the array growth fits within availMem. (We could still cause
1159 : * LACKMEM if the memory chunk overhead associated with the memtuples
1160 : * array were to increase. That shouldn't happen because we chose the
1161 : * initial array size large enough to ensure that palloc will be treating
1162 : * both old and new arrays as separate chunks. But we'll check LACKMEM
1163 : * explicitly below just in case.)
1164 : */
1165 6828 : if (state->availMem < (int64) ((newmemtupsize - memtupsize) * sizeof(SortTuple)))
1166 0 : goto noalloc;
1167 :
1168 : /* OK, do it */
1169 6828 : FREEMEM(state, GetMemoryChunkSpace(state->memtuples));
1170 6828 : state->memtupsize = newmemtupsize;
1171 6828 : state->memtuples = (SortTuple *)
1172 6828 : repalloc_huge(state->memtuples,
1173 6828 : state->memtupsize * sizeof(SortTuple));
1174 6828 : USEMEM(state, GetMemoryChunkSpace(state->memtuples));
1175 6828 : if (LACKMEM(state))
1176 0 : elog(ERROR, "unexpected out-of-memory situation in tuplesort");
1177 6828 : return true;
1178 :
1179 0 : noalloc:
1180 : /* If for any reason we didn't realloc, shut off future attempts */
1181 0 : state->growmemtuples = false;
1182 0 : return false;
1183 : }
1184 :
1185 : /*
1186 : * Shared code for tuple and datum cases.
1187 : */
1188 : void
1189 27816690 : tuplesort_puttuple_common(Tuplesortstate *state, SortTuple *tuple,
1190 : bool useAbbrev, Size tuplen)
1191 : {
1192 27816690 : MemoryContext oldcontext = MemoryContextSwitchTo(state->base.sortcontext);
1193 :
1194 : Assert(!LEADER(state));
1195 :
1196 : /* account for the memory used for this tuple */
1197 27816690 : USEMEM(state, tuplen);
1198 27816690 : state->tupleMem += tuplen;
1199 :
1200 27816690 : if (!useAbbrev)
1201 : {
1202 : /*
1203 : * Leave ordinary Datum representation, or NULL value. If there is a
1204 : * converter it won't expect NULL values, and cost model is not
1205 : * required to account for NULL, so in that case we avoid calling
1206 : * converter and just set datum1 to zeroed representation (to be
1207 : * consistent, and to support cheap inequality tests for NULL
1208 : * abbreviated keys).
1209 : */
1210 : }
1211 4419692 : else if (!consider_abort_common(state))
1212 : {
1213 : /* Store abbreviated key representation */
1214 4419596 : tuple->datum1 = state->base.sortKeys->abbrev_converter(tuple->datum1,
1215 : state->base.sortKeys);
1216 : }
1217 : else
1218 : {
1219 : /*
1220 : * Set state to be consistent with never trying abbreviation.
1221 : *
1222 : * Alter datum1 representation in already-copied tuples, so as to
1223 : * ensure a consistent representation (current tuple was just
1224 : * handled). It does not matter if some dumped tuples are already
1225 : * sorted on tape, since serialized tuples lack abbreviated keys
1226 : * (TSS_BUILDRUNS state prevents control reaching here in any case).
1227 : */
1228 96 : REMOVEABBREV(state, state->memtuples, state->memtupcount);
1229 : }
1230 :
1231 27816690 : switch (state->status)
1232 : {
1233 23021892 : case TSS_INITIAL:
1234 :
1235 : /*
1236 : * Save the tuple into the unsorted array. First, grow the array
1237 : * as needed. Note that we try to grow the array when there is
1238 : * still one free slot remaining --- if we fail, there'll still be
1239 : * room to store the incoming tuple, and then we'll switch to
1240 : * tape-based operation.
1241 : */
1242 23021892 : if (state->memtupcount >= state->memtupsize - 1)
1243 : {
1244 6942 : (void) grow_memtuples(state);
1245 : Assert(state->memtupcount < state->memtupsize);
1246 : }
1247 23021892 : state->memtuples[state->memtupcount++] = *tuple;
1248 :
1249 : /*
1250 : * Check if it's time to switch over to a bounded heapsort. We do
1251 : * so if the input tuple count exceeds twice the desired tuple
1252 : * count (this is a heuristic for where heapsort becomes cheaper
1253 : * than a quicksort), or if we've just filled workMem and have
1254 : * enough tuples to meet the bound.
1255 : *
1256 : * Note that once we enter TSS_BOUNDED state we will always try to
1257 : * complete the sort that way. In the worst case, if later input
1258 : * tuples are larger than earlier ones, this might cause us to
1259 : * exceed workMem significantly.
1260 : */
1261 23021892 : if (state->bounded &&
1262 43182 : (state->memtupcount > state->bound * 2 ||
1263 42784 : (state->memtupcount > state->bound && LACKMEM(state))))
1264 : {
1265 : #ifdef TRACE_SORT
1266 398 : if (trace_sort)
1267 0 : elog(LOG, "switching to bounded heapsort at %d tuples: %s",
1268 : state->memtupcount,
1269 : pg_rusage_show(&state->ru_start));
1270 : #endif
1271 398 : make_bounded_heap(state);
1272 398 : MemoryContextSwitchTo(oldcontext);
1273 398 : return;
1274 : }
1275 :
1276 : /*
1277 : * Done if we still fit in available memory and have array slots.
1278 : */
1279 23021494 : if (state->memtupcount < state->memtupsize && !LACKMEM(state))
1280 : {
1281 23021380 : MemoryContextSwitchTo(oldcontext);
1282 23021380 : return;
1283 : }
1284 :
1285 : /*
1286 : * Nope; time to switch to tape-based operation.
1287 : */
1288 114 : inittapes(state, true);
1289 :
1290 : /*
1291 : * Dump all tuples.
1292 : */
1293 114 : dumptuples(state, false);
1294 114 : break;
1295 :
1296 3734028 : case TSS_BOUNDED:
1297 :
1298 : /*
1299 : * We don't want to grow the array here, so check whether the new
1300 : * tuple can be discarded before putting it in. This should be a
1301 : * good speed optimization, too, since when there are many more
1302 : * input tuples than the bound, most input tuples can be discarded
1303 : * with just this one comparison. Note that because we currently
1304 : * have the sort direction reversed, we must check for <= not >=.
1305 : */
1306 3734028 : if (COMPARETUP(state, tuple, &state->memtuples[0]) <= 0)
1307 : {
1308 : /* new tuple <= top of the heap, so we can discard it */
1309 3231126 : free_sort_tuple(state, tuple);
1310 3231126 : CHECK_FOR_INTERRUPTS();
1311 : }
1312 : else
1313 : {
1314 : /* discard top of heap, replacing it with the new tuple */
1315 502902 : free_sort_tuple(state, &state->memtuples[0]);
1316 502902 : tuplesort_heap_replace_top(state, tuple);
1317 : }
1318 3734028 : break;
1319 :
1320 1060770 : case TSS_BUILDRUNS:
1321 :
1322 : /*
1323 : * Save the tuple into the unsorted array (there must be space)
1324 : */
1325 1060770 : state->memtuples[state->memtupcount++] = *tuple;
1326 :
1327 : /*
1328 : * If we are over the memory limit, dump all tuples.
1329 : */
1330 1060770 : dumptuples(state, false);
1331 1060770 : break;
1332 :
1333 0 : default:
1334 0 : elog(ERROR, "invalid tuplesort state");
1335 : break;
1336 : }
1337 4794912 : MemoryContextSwitchTo(oldcontext);
1338 : }
1339 :
1340 : static bool
1341 4419692 : consider_abort_common(Tuplesortstate *state)
1342 : {
1343 : Assert(state->base.sortKeys[0].abbrev_converter != NULL);
1344 : Assert(state->base.sortKeys[0].abbrev_abort != NULL);
1345 : Assert(state->base.sortKeys[0].abbrev_full_comparator != NULL);
1346 :
1347 : /*
1348 : * Check effectiveness of abbreviation optimization. Consider aborting
1349 : * when still within memory limit.
1350 : */
1351 4419692 : if (state->status == TSS_INITIAL &&
1352 3947262 : state->memtupcount >= state->abbrevNext)
1353 : {
1354 4904 : state->abbrevNext *= 2;
1355 :
1356 : /*
1357 : * Check opclass-supplied abbreviation abort routine. It may indicate
1358 : * that abbreviation should not proceed.
1359 : */
1360 4904 : if (!state->base.sortKeys->abbrev_abort(state->memtupcount,
1361 : state->base.sortKeys))
1362 4808 : return false;
1363 :
1364 : /*
1365 : * Finally, restore authoritative comparator, and indicate that
1366 : * abbreviation is not in play by setting abbrev_converter to NULL
1367 : */
1368 96 : state->base.sortKeys[0].comparator = state->base.sortKeys[0].abbrev_full_comparator;
1369 96 : state->base.sortKeys[0].abbrev_converter = NULL;
1370 : /* Not strictly necessary, but be tidy */
1371 96 : state->base.sortKeys[0].abbrev_abort = NULL;
1372 96 : state->base.sortKeys[0].abbrev_full_comparator = NULL;
1373 :
1374 : /* Give up - expect original pass-by-value representation */
1375 96 : return true;
1376 : }
1377 :
1378 4414788 : return false;
1379 : }
1380 :
1381 : /*
1382 : * All tuples have been provided; finish the sort.
1383 : */
1384 : void
1385 199386 : tuplesort_performsort(Tuplesortstate *state)
1386 : {
1387 199386 : MemoryContext oldcontext = MemoryContextSwitchTo(state->base.sortcontext);
1388 :
1389 : #ifdef TRACE_SORT
1390 199386 : if (trace_sort)
1391 0 : elog(LOG, "performsort of worker %d starting: %s",
1392 : state->worker, pg_rusage_show(&state->ru_start));
1393 : #endif
1394 :
1395 199386 : switch (state->status)
1396 : {
1397 198874 : case TSS_INITIAL:
1398 :
1399 : /*
1400 : * We were able to accumulate all the tuples within the allowed
1401 : * amount of memory, or leader to take over worker tapes
1402 : */
1403 198874 : if (SERIAL(state))
1404 : {
1405 : /* Just qsort 'em and we're done */
1406 198310 : tuplesort_sort_memtuples(state);
1407 198226 : state->status = TSS_SORTEDINMEM;
1408 : }
1409 564 : else if (WORKER(state))
1410 : {
1411 : /*
1412 : * Parallel workers must still dump out tuples to tape. No
1413 : * merge is required to produce single output run, though.
1414 : */
1415 420 : inittapes(state, false);
1416 420 : dumptuples(state, true);
1417 420 : worker_nomergeruns(state);
1418 420 : state->status = TSS_SORTEDONTAPE;
1419 : }
1420 : else
1421 : {
1422 : /*
1423 : * Leader will take over worker tapes and merge worker runs.
1424 : * Note that mergeruns sets the correct state->status.
1425 : */
1426 144 : leader_takeover_tapes(state);
1427 144 : mergeruns(state);
1428 : }
1429 198790 : state->current = 0;
1430 198790 : state->eof_reached = false;
1431 198790 : state->markpos_block = 0L;
1432 198790 : state->markpos_offset = 0;
1433 198790 : state->markpos_eof = false;
1434 198790 : break;
1435 :
1436 398 : case TSS_BOUNDED:
1437 :
1438 : /*
1439 : * We were able to accumulate all the tuples required for output
1440 : * in memory, using a heap to eliminate excess tuples. Now we
1441 : * have to transform the heap to a properly-sorted array. Note
1442 : * that sort_bounded_heap sets the correct state->status.
1443 : */
1444 398 : sort_bounded_heap(state);
1445 398 : state->current = 0;
1446 398 : state->eof_reached = false;
1447 398 : state->markpos_offset = 0;
1448 398 : state->markpos_eof = false;
1449 398 : break;
1450 :
1451 114 : case TSS_BUILDRUNS:
1452 :
1453 : /*
1454 : * Finish tape-based sort. First, flush all tuples remaining in
1455 : * memory out to tape; then merge until we have a single remaining
1456 : * run (or, if !randomAccess and !WORKER(), one run per tape).
1457 : * Note that mergeruns sets the correct state->status.
1458 : */
1459 114 : dumptuples(state, true);
1460 114 : mergeruns(state);
1461 114 : state->eof_reached = false;
1462 114 : state->markpos_block = 0L;
1463 114 : state->markpos_offset = 0;
1464 114 : state->markpos_eof = false;
1465 114 : break;
1466 :
1467 0 : default:
1468 0 : elog(ERROR, "invalid tuplesort state");
1469 : break;
1470 : }
1471 :
1472 : #ifdef TRACE_SORT
1473 199302 : if (trace_sort)
1474 : {
1475 0 : if (state->status == TSS_FINALMERGE)
1476 0 : elog(LOG, "performsort of worker %d done (except %d-way final merge): %s",
1477 : state->worker, state->nInputTapes,
1478 : pg_rusage_show(&state->ru_start));
1479 : else
1480 0 : elog(LOG, "performsort of worker %d done: %s",
1481 : state->worker, pg_rusage_show(&state->ru_start));
1482 : }
1483 : #endif
1484 :
1485 199302 : MemoryContextSwitchTo(oldcontext);
1486 199302 : }
1487 :
1488 : /*
1489 : * Internal routine to fetch the next tuple in either forward or back
1490 : * direction into *stup. Returns false if no more tuples.
1491 : * Returned tuple belongs to tuplesort memory context, and must not be freed
1492 : * by caller. Note that fetched tuple is stored in memory that may be
1493 : * recycled by any future fetch.
1494 : */
1495 : bool
1496 24974118 : tuplesort_gettuple_common(Tuplesortstate *state, bool forward,
1497 : SortTuple *stup)
1498 : {
1499 : unsigned int tuplen;
1500 : size_t nmoved;
1501 :
1502 : Assert(!WORKER(state));
1503 :
1504 24974118 : switch (state->status)
1505 : {
1506 20145572 : case TSS_SORTEDINMEM:
1507 : Assert(forward || state->base.sortopt & TUPLESORT_RANDOMACCESS);
1508 : Assert(!state->slabAllocatorUsed);
1509 20145572 : if (forward)
1510 : {
1511 20145506 : if (state->current < state->memtupcount)
1512 : {
1513 19948554 : *stup = state->memtuples[state->current++];
1514 19948554 : return true;
1515 : }
1516 196952 : state->eof_reached = true;
1517 :
1518 : /*
1519 : * Complain if caller tries to retrieve more tuples than
1520 : * originally asked for in a bounded sort. This is because
1521 : * returning EOF here might be the wrong thing.
1522 : */
1523 196952 : if (state->bounded && state->current >= state->bound)
1524 0 : elog(ERROR, "retrieved too many tuples in a bounded sort");
1525 :
1526 196952 : return false;
1527 : }
1528 : else
1529 : {
1530 66 : if (state->current <= 0)
1531 0 : return false;
1532 :
1533 : /*
1534 : * if all tuples are fetched already then we return last
1535 : * tuple, else - tuple before last returned.
1536 : */
1537 66 : if (state->eof_reached)
1538 12 : state->eof_reached = false;
1539 : else
1540 : {
1541 54 : state->current--; /* last returned tuple */
1542 54 : if (state->current <= 0)
1543 6 : return false;
1544 : }
1545 60 : *stup = state->memtuples[state->current - 1];
1546 60 : return true;
1547 : }
1548 : break;
1549 :
1550 272994 : case TSS_SORTEDONTAPE:
1551 : Assert(forward || state->base.sortopt & TUPLESORT_RANDOMACCESS);
1552 : Assert(state->slabAllocatorUsed);
1553 :
1554 : /*
1555 : * The slot that held the tuple that we returned in previous
1556 : * gettuple call can now be reused.
1557 : */
1558 272994 : if (state->lastReturnedTuple)
1559 : {
1560 152850 : RELEASE_SLAB_SLOT(state, state->lastReturnedTuple);
1561 152850 : state->lastReturnedTuple = NULL;
1562 : }
1563 :
1564 272994 : if (forward)
1565 : {
1566 272964 : if (state->eof_reached)
1567 0 : return false;
1568 :
1569 272964 : if ((tuplen = getlen(state->result_tape, true)) != 0)
1570 : {
1571 272940 : READTUP(state, stup, state->result_tape, tuplen);
1572 :
1573 : /*
1574 : * Remember the tuple we return, so that we can recycle
1575 : * its memory on next call. (This can be NULL, in the
1576 : * !state->tuples case).
1577 : */
1578 272940 : state->lastReturnedTuple = stup->tuple;
1579 :
1580 272940 : return true;
1581 : }
1582 : else
1583 : {
1584 24 : state->eof_reached = true;
1585 24 : return false;
1586 : }
1587 : }
1588 :
1589 : /*
1590 : * Backward.
1591 : *
1592 : * if all tuples are fetched already then we return last tuple,
1593 : * else - tuple before last returned.
1594 : */
1595 30 : if (state->eof_reached)
1596 : {
1597 : /*
1598 : * Seek position is pointing just past the zero tuplen at the
1599 : * end of file; back up to fetch last tuple's ending length
1600 : * word. If seek fails we must have a completely empty file.
1601 : */
1602 12 : nmoved = LogicalTapeBackspace(state->result_tape,
1603 : 2 * sizeof(unsigned int));
1604 12 : if (nmoved == 0)
1605 0 : return false;
1606 12 : else if (nmoved != 2 * sizeof(unsigned int))
1607 0 : elog(ERROR, "unexpected tape position");
1608 12 : state->eof_reached = false;
1609 : }
1610 : else
1611 : {
1612 : /*
1613 : * Back up and fetch previously-returned tuple's ending length
1614 : * word. If seek fails, assume we are at start of file.
1615 : */
1616 18 : nmoved = LogicalTapeBackspace(state->result_tape,
1617 : sizeof(unsigned int));
1618 18 : if (nmoved == 0)
1619 0 : return false;
1620 18 : else if (nmoved != sizeof(unsigned int))
1621 0 : elog(ERROR, "unexpected tape position");
1622 18 : tuplen = getlen(state->result_tape, false);
1623 :
1624 : /*
1625 : * Back up to get ending length word of tuple before it.
1626 : */
1627 18 : nmoved = LogicalTapeBackspace(state->result_tape,
1628 : tuplen + 2 * sizeof(unsigned int));
1629 18 : if (nmoved == tuplen + sizeof(unsigned int))
1630 : {
1631 : /*
1632 : * We backed up over the previous tuple, but there was no
1633 : * ending length word before it. That means that the prev
1634 : * tuple is the first tuple in the file. It is now the
1635 : * next to read in forward direction (not obviously right,
1636 : * but that is what in-memory case does).
1637 : */
1638 6 : return false;
1639 : }
1640 12 : else if (nmoved != tuplen + 2 * sizeof(unsigned int))
1641 0 : elog(ERROR, "bogus tuple length in backward scan");
1642 : }
1643 :
1644 24 : tuplen = getlen(state->result_tape, false);
1645 :
1646 : /*
1647 : * Now we have the length of the prior tuple, back up and read it.
1648 : * Note: READTUP expects we are positioned after the initial
1649 : * length word of the tuple, so back up to that point.
1650 : */
1651 24 : nmoved = LogicalTapeBackspace(state->result_tape,
1652 : tuplen);
1653 24 : if (nmoved != tuplen)
1654 0 : elog(ERROR, "bogus tuple length in backward scan");
1655 24 : READTUP(state, stup, state->result_tape, tuplen);
1656 :
1657 : /*
1658 : * Remember the tuple we return, so that we can recycle its memory
1659 : * on next call. (This can be NULL, in the Datum case).
1660 : */
1661 24 : state->lastReturnedTuple = stup->tuple;
1662 :
1663 24 : return true;
1664 :
1665 4555552 : case TSS_FINALMERGE:
1666 : Assert(forward);
1667 : /* We are managing memory ourselves, with the slab allocator. */
1668 : Assert(state->slabAllocatorUsed);
1669 :
1670 : /*
1671 : * The slab slot holding the tuple that we returned in previous
1672 : * gettuple call can now be reused.
1673 : */
1674 4555552 : if (state->lastReturnedTuple)
1675 : {
1676 4495282 : RELEASE_SLAB_SLOT(state, state->lastReturnedTuple);
1677 4495282 : state->lastReturnedTuple = NULL;
1678 : }
1679 :
1680 : /*
1681 : * This code should match the inner loop of mergeonerun().
1682 : */
1683 4555552 : if (state->memtupcount > 0)
1684 : {
1685 4555330 : int srcTapeIndex = state->memtuples[0].srctape;
1686 4555330 : LogicalTape *srcTape = state->inputTapes[srcTapeIndex];
1687 : SortTuple newtup;
1688 :
1689 4555330 : *stup = state->memtuples[0];
1690 :
1691 : /*
1692 : * Remember the tuple we return, so that we can recycle its
1693 : * memory on next call. (This can be NULL, in the Datum case).
1694 : */
1695 4555330 : state->lastReturnedTuple = stup->tuple;
1696 :
1697 : /*
1698 : * Pull next tuple from tape, and replace the returned tuple
1699 : * at top of the heap with it.
1700 : */
1701 4555330 : if (!mergereadnext(state, srcTape, &newtup))
1702 : {
1703 : /*
1704 : * If no more data, we've reached end of run on this tape.
1705 : * Remove the top node from the heap.
1706 : */
1707 334 : tuplesort_heap_delete_top(state);
1708 334 : state->nInputRuns--;
1709 :
1710 : /*
1711 : * Close the tape. It'd go away at the end of the sort
1712 : * anyway, but better to release the memory early.
1713 : */
1714 334 : LogicalTapeClose(srcTape);
1715 334 : return true;
1716 : }
1717 4554996 : newtup.srctape = srcTapeIndex;
1718 4554996 : tuplesort_heap_replace_top(state, &newtup);
1719 4554996 : return true;
1720 : }
1721 222 : return false;
1722 :
1723 0 : default:
1724 0 : elog(ERROR, "invalid tuplesort state");
1725 : return false; /* keep compiler quiet */
1726 : }
1727 : }
1728 :
1729 :
1730 : /*
1731 : * Advance over N tuples in either forward or back direction,
1732 : * without returning any data. N==0 is a no-op.
1733 : * Returns true if successful, false if ran out of tuples.
1734 : */
1735 : bool
1736 392 : tuplesort_skiptuples(Tuplesortstate *state, int64 ntuples, bool forward)
1737 : {
1738 : MemoryContext oldcontext;
1739 :
1740 : /*
1741 : * We don't actually support backwards skip yet, because no callers need
1742 : * it. The API is designed to allow for that later, though.
1743 : */
1744 : Assert(forward);
1745 : Assert(ntuples >= 0);
1746 : Assert(!WORKER(state));
1747 :
1748 392 : switch (state->status)
1749 : {
1750 368 : case TSS_SORTEDINMEM:
1751 368 : if (state->memtupcount - state->current >= ntuples)
1752 : {
1753 368 : state->current += ntuples;
1754 368 : return true;
1755 : }
1756 0 : state->current = state->memtupcount;
1757 0 : state->eof_reached = true;
1758 :
1759 : /*
1760 : * Complain if caller tries to retrieve more tuples than
1761 : * originally asked for in a bounded sort. This is because
1762 : * returning EOF here might be the wrong thing.
1763 : */
1764 0 : if (state->bounded && state->current >= state->bound)
1765 0 : elog(ERROR, "retrieved too many tuples in a bounded sort");
1766 :
1767 0 : return false;
1768 :
1769 24 : case TSS_SORTEDONTAPE:
1770 : case TSS_FINALMERGE:
1771 :
1772 : /*
1773 : * We could probably optimize these cases better, but for now it's
1774 : * not worth the trouble.
1775 : */
1776 24 : oldcontext = MemoryContextSwitchTo(state->base.sortcontext);
1777 240132 : while (ntuples-- > 0)
1778 : {
1779 : SortTuple stup;
1780 :
1781 240108 : if (!tuplesort_gettuple_common(state, forward, &stup))
1782 : {
1783 0 : MemoryContextSwitchTo(oldcontext);
1784 0 : return false;
1785 : }
1786 240108 : CHECK_FOR_INTERRUPTS();
1787 : }
1788 24 : MemoryContextSwitchTo(oldcontext);
1789 24 : return true;
1790 :
1791 0 : default:
1792 0 : elog(ERROR, "invalid tuplesort state");
1793 : return false; /* keep compiler quiet */
1794 : }
1795 : }
1796 :
1797 : /*
1798 : * tuplesort_merge_order - report merge order we'll use for given memory
1799 : * (note: "merge order" just means the number of input tapes in the merge).
1800 : *
1801 : * This is exported for use by the planner. allowedMem is in bytes.
1802 : */
1803 : int
1804 16848 : tuplesort_merge_order(int64 allowedMem)
1805 : {
1806 : int mOrder;
1807 :
1808 : /*----------
1809 : * In the merge phase, we need buffer space for each input and output tape.
1810 : * Each pass in the balanced merge algorithm reads from M input tapes, and
1811 : * writes to N output tapes. Each tape consumes TAPE_BUFFER_OVERHEAD bytes
1812 : * of memory. In addition to that, we want MERGE_BUFFER_SIZE workspace per
1813 : * input tape.
1814 : *
1815 : * totalMem = M * (TAPE_BUFFER_OVERHEAD + MERGE_BUFFER_SIZE) +
1816 : * N * TAPE_BUFFER_OVERHEAD
1817 : *
1818 : * Except for the last and next-to-last merge passes, where there can be
1819 : * fewer tapes left to process, M = N. We choose M so that we have the
1820 : * desired amount of memory available for the input buffers
1821 : * (TAPE_BUFFER_OVERHEAD + MERGE_BUFFER_SIZE), given the total memory
1822 : * available for the tape buffers (allowedMem).
1823 : *
1824 : * Note: you might be thinking we need to account for the memtuples[]
1825 : * array in this calculation, but we effectively treat that as part of the
1826 : * MERGE_BUFFER_SIZE workspace.
1827 : *----------
1828 : */
1829 16848 : mOrder = allowedMem /
1830 : (2 * TAPE_BUFFER_OVERHEAD + MERGE_BUFFER_SIZE);
1831 :
1832 : /*
1833 : * Even in minimum memory, use at least a MINORDER merge. On the other
1834 : * hand, even when we have lots of memory, do not use more than a MAXORDER
1835 : * merge. Tapes are pretty cheap, but they're not entirely free. Each
1836 : * additional tape reduces the amount of memory available to build runs,
1837 : * which in turn can cause the same sort to need more runs, which makes
1838 : * merging slower even if it can still be done in a single pass. Also,
1839 : * high order merges are quite slow due to CPU cache effects; it can be
1840 : * faster to pay the I/O cost of a multi-pass merge than to perform a
1841 : * single merge pass across many hundreds of tapes.
1842 : */
1843 16848 : mOrder = Max(mOrder, MINORDER);
1844 16848 : mOrder = Min(mOrder, MAXORDER);
1845 :
1846 16848 : return mOrder;
1847 : }
1848 :
1849 : /*
1850 : * Helper function to calculate how much memory to allocate for the read buffer
1851 : * of each input tape in a merge pass.
1852 : *
1853 : * 'avail_mem' is the amount of memory available for the buffers of all the
1854 : * tapes, both input and output.
1855 : * 'nInputTapes' and 'nInputRuns' are the number of input tapes and runs.
1856 : * 'maxOutputTapes' is the max. number of output tapes we should produce.
1857 : */
1858 : static int64
1859 288 : merge_read_buffer_size(int64 avail_mem, int nInputTapes, int nInputRuns,
1860 : int maxOutputTapes)
1861 : {
1862 : int nOutputRuns;
1863 : int nOutputTapes;
1864 :
1865 : /*
1866 : * How many output tapes will we produce in this pass?
1867 : *
1868 : * This is nInputRuns / nInputTapes, rounded up.
1869 : */
1870 288 : nOutputRuns = (nInputRuns + nInputTapes - 1) / nInputTapes;
1871 :
1872 288 : nOutputTapes = Min(nOutputRuns, maxOutputTapes);
1873 :
1874 : /*
1875 : * Each output tape consumes TAPE_BUFFER_OVERHEAD bytes of memory. All
1876 : * remaining memory is divided evenly between the input tapes.
1877 : *
1878 : * This also follows from the formula in tuplesort_merge_order, but here
1879 : * we derive the input buffer size from the amount of memory available,
1880 : * and M and N.
1881 : */
1882 288 : return Max((avail_mem - TAPE_BUFFER_OVERHEAD * nOutputTapes) / nInputTapes, 0);
1883 : }
1884 :
1885 : /*
1886 : * inittapes - initialize for tape sorting.
1887 : *
1888 : * This is called only if we have found we won't sort in memory.
1889 : */
1890 : static void
1891 534 : inittapes(Tuplesortstate *state, bool mergeruns)
1892 : {
1893 : Assert(!LEADER(state));
1894 :
1895 534 : if (mergeruns)
1896 : {
1897 : /* Compute number of input tapes to use when merging */
1898 114 : state->maxTapes = tuplesort_merge_order(state->allowedMem);
1899 : }
1900 : else
1901 : {
1902 : /* Workers can sometimes produce single run, output without merge */
1903 : Assert(WORKER(state));
1904 420 : state->maxTapes = MINORDER;
1905 : }
1906 :
1907 : #ifdef TRACE_SORT
1908 534 : if (trace_sort)
1909 0 : elog(LOG, "worker %d switching to external sort with %d tapes: %s",
1910 : state->worker, state->maxTapes, pg_rusage_show(&state->ru_start));
1911 : #endif
1912 :
1913 : /* Create the tape set */
1914 534 : inittapestate(state, state->maxTapes);
1915 534 : state->tapeset =
1916 534 : LogicalTapeSetCreate(false,
1917 534 : state->shared ? &state->shared->fileset : NULL,
1918 : state->worker);
1919 :
1920 534 : state->currentRun = 0;
1921 :
1922 : /*
1923 : * Initialize logical tape arrays.
1924 : */
1925 534 : state->inputTapes = NULL;
1926 534 : state->nInputTapes = 0;
1927 534 : state->nInputRuns = 0;
1928 :
1929 534 : state->outputTapes = palloc0(state->maxTapes * sizeof(LogicalTape *));
1930 534 : state->nOutputTapes = 0;
1931 534 : state->nOutputRuns = 0;
1932 :
1933 534 : state->status = TSS_BUILDRUNS;
1934 :
1935 534 : selectnewtape(state);
1936 534 : }
1937 :
1938 : /*
1939 : * inittapestate - initialize generic tape management state
1940 : */
1941 : static void
1942 678 : inittapestate(Tuplesortstate *state, int maxTapes)
1943 : {
1944 : int64 tapeSpace;
1945 :
1946 : /*
1947 : * Decrease availMem to reflect the space needed for tape buffers; but
1948 : * don't decrease it to the point that we have no room for tuples. (That
1949 : * case is only likely to occur if sorting pass-by-value Datums; in all
1950 : * other scenarios the memtuples[] array is unlikely to occupy more than
1951 : * half of allowedMem. In the pass-by-value case it's not important to
1952 : * account for tuple space, so we don't care if LACKMEM becomes
1953 : * inaccurate.)
1954 : */
1955 678 : tapeSpace = (int64) maxTapes * TAPE_BUFFER_OVERHEAD;
1956 :
1957 678 : if (tapeSpace + GetMemoryChunkSpace(state->memtuples) < state->allowedMem)
1958 582 : USEMEM(state, tapeSpace);
1959 :
1960 : /*
1961 : * Make sure that the temp file(s) underlying the tape set are created in
1962 : * suitable temp tablespaces. For parallel sorts, this should have been
1963 : * called already, but it doesn't matter if it is called a second time.
1964 : */
1965 678 : PrepareTempTablespaces();
1966 678 : }
1967 :
1968 : /*
1969 : * selectnewtape -- select next tape to output to.
1970 : *
1971 : * This is called after finishing a run when we know another run
1972 : * must be started. This is used both when building the initial
1973 : * runs, and during merge passes.
1974 : */
1975 : static void
1976 1602 : selectnewtape(Tuplesortstate *state)
1977 : {
1978 : /*
1979 : * At the beginning of each merge pass, nOutputTapes and nOutputRuns are
1980 : * both zero. On each call, we create a new output tape to hold the next
1981 : * run, until maxTapes is reached. After that, we assign new runs to the
1982 : * existing tapes in a round robin fashion.
1983 : */
1984 1602 : if (state->nOutputTapes < state->maxTapes)
1985 : {
1986 : /* Create a new tape to hold the next run */
1987 : Assert(state->outputTapes[state->nOutputRuns] == NULL);
1988 : Assert(state->nOutputRuns == state->nOutputTapes);
1989 1020 : state->destTape = LogicalTapeCreate(state->tapeset);
1990 1020 : state->outputTapes[state->nOutputTapes] = state->destTape;
1991 1020 : state->nOutputTapes++;
1992 1020 : state->nOutputRuns++;
1993 : }
1994 : else
1995 : {
1996 : /*
1997 : * We have reached the max number of tapes. Append to an existing
1998 : * tape.
1999 : */
2000 582 : state->destTape = state->outputTapes[state->nOutputRuns % state->nOutputTapes];
2001 582 : state->nOutputRuns++;
2002 : }
2003 1602 : }
2004 :
2005 : /*
2006 : * Initialize the slab allocation arena, for the given number of slots.
2007 : */
2008 : static void
2009 258 : init_slab_allocator(Tuplesortstate *state, int numSlots)
2010 : {
2011 258 : if (numSlots > 0)
2012 : {
2013 : char *p;
2014 : int i;
2015 :
2016 246 : state->slabMemoryBegin = palloc(numSlots * SLAB_SLOT_SIZE);
2017 246 : state->slabMemoryEnd = state->slabMemoryBegin +
2018 246 : numSlots * SLAB_SLOT_SIZE;
2019 246 : state->slabFreeHead = (SlabSlot *) state->slabMemoryBegin;
2020 246 : USEMEM(state, numSlots * SLAB_SLOT_SIZE);
2021 :
2022 246 : p = state->slabMemoryBegin;
2023 952 : for (i = 0; i < numSlots - 1; i++)
2024 : {
2025 706 : ((SlabSlot *) p)->nextfree = (SlabSlot *) (p + SLAB_SLOT_SIZE);
2026 706 : p += SLAB_SLOT_SIZE;
2027 : }
2028 246 : ((SlabSlot *) p)->nextfree = NULL;
2029 : }
2030 : else
2031 : {
2032 12 : state->slabMemoryBegin = state->slabMemoryEnd = NULL;
2033 12 : state->slabFreeHead = NULL;
2034 : }
2035 258 : state->slabAllocatorUsed = true;
2036 258 : }
2037 :
2038 : /*
2039 : * mergeruns -- merge all the completed initial runs.
2040 : *
2041 : * This implements the Balanced k-Way Merge Algorithm. All input data has
2042 : * already been written to initial runs on tape (see dumptuples).
2043 : */
2044 : static void
2045 258 : mergeruns(Tuplesortstate *state)
2046 : {
2047 : int tapenum;
2048 :
2049 : Assert(state->status == TSS_BUILDRUNS);
2050 : Assert(state->memtupcount == 0);
2051 :
2052 258 : if (state->base.sortKeys != NULL && state->base.sortKeys->abbrev_converter != NULL)
2053 : {
2054 : /*
2055 : * If there are multiple runs to be merged, when we go to read back
2056 : * tuples from disk, abbreviated keys will not have been stored, and
2057 : * we don't care to regenerate them. Disable abbreviation from this
2058 : * point on.
2059 : */
2060 30 : state->base.sortKeys->abbrev_converter = NULL;
2061 30 : state->base.sortKeys->comparator = state->base.sortKeys->abbrev_full_comparator;
2062 :
2063 : /* Not strictly necessary, but be tidy */
2064 30 : state->base.sortKeys->abbrev_abort = NULL;
2065 30 : state->base.sortKeys->abbrev_full_comparator = NULL;
2066 : }
2067 :
2068 : /*
2069 : * Reset tuple memory. We've freed all the tuples that we previously
2070 : * allocated. We will use the slab allocator from now on.
2071 : */
2072 258 : MemoryContextResetOnly(state->base.tuplecontext);
2073 :
2074 : /*
2075 : * We no longer need a large memtuples array. (We will allocate a smaller
2076 : * one for the heap later.)
2077 : */
2078 258 : FREEMEM(state, GetMemoryChunkSpace(state->memtuples));
2079 258 : pfree(state->memtuples);
2080 258 : state->memtuples = NULL;
2081 :
2082 : /*
2083 : * Initialize the slab allocator. We need one slab slot per input tape,
2084 : * for the tuples in the heap, plus one to hold the tuple last returned
2085 : * from tuplesort_gettuple. (If we're sorting pass-by-val Datums,
2086 : * however, we don't need to do allocate anything.)
2087 : *
2088 : * In a multi-pass merge, we could shrink this allocation for the last
2089 : * merge pass, if it has fewer tapes than previous passes, but we don't
2090 : * bother.
2091 : *
2092 : * From this point on, we no longer use the USEMEM()/LACKMEM() mechanism
2093 : * to track memory usage of individual tuples.
2094 : */
2095 258 : if (state->base.tuples)
2096 246 : init_slab_allocator(state, state->nOutputTapes + 1);
2097 : else
2098 12 : init_slab_allocator(state, 0);
2099 :
2100 : /*
2101 : * Allocate a new 'memtuples' array, for the heap. It will hold one tuple
2102 : * from each input tape.
2103 : *
2104 : * We could shrink this, too, between passes in a multi-pass merge, but we
2105 : * don't bother. (The initial input tapes are still in outputTapes. The
2106 : * number of input tapes will not increase between passes.)
2107 : */
2108 258 : state->memtupsize = state->nOutputTapes;
2109 516 : state->memtuples = (SortTuple *) MemoryContextAlloc(state->base.maincontext,
2110 258 : state->nOutputTapes * sizeof(SortTuple));
2111 258 : USEMEM(state, GetMemoryChunkSpace(state->memtuples));
2112 :
2113 : /*
2114 : * Use all the remaining memory we have available for tape buffers among
2115 : * all the input tapes. At the beginning of each merge pass, we will
2116 : * divide this memory between the input and output tapes in the pass.
2117 : */
2118 258 : state->tape_buffer_mem = state->availMem;
2119 258 : USEMEM(state, state->tape_buffer_mem);
2120 : #ifdef TRACE_SORT
2121 258 : if (trace_sort)
2122 0 : elog(LOG, "worker %d using %zu KB of memory for tape buffers",
2123 : state->worker, state->tape_buffer_mem / 1024);
2124 : #endif
2125 :
2126 : for (;;)
2127 : {
2128 : /*
2129 : * On the first iteration, or if we have read all the runs from the
2130 : * input tapes in a multi-pass merge, it's time to start a new pass.
2131 : * Rewind all the output tapes, and make them inputs for the next
2132 : * pass.
2133 : */
2134 396 : if (state->nInputRuns == 0)
2135 : {
2136 : int64 input_buffer_size;
2137 :
2138 : /* Close the old, emptied, input tapes */
2139 288 : if (state->nInputTapes > 0)
2140 : {
2141 210 : for (tapenum = 0; tapenum < state->nInputTapes; tapenum++)
2142 180 : LogicalTapeClose(state->inputTapes[tapenum]);
2143 30 : pfree(state->inputTapes);
2144 : }
2145 :
2146 : /* Previous pass's outputs become next pass's inputs. */
2147 288 : state->inputTapes = state->outputTapes;
2148 288 : state->nInputTapes = state->nOutputTapes;
2149 288 : state->nInputRuns = state->nOutputRuns;
2150 :
2151 : /*
2152 : * Reset output tape variables. The actual LogicalTapes will be
2153 : * created as needed, here we only allocate the array to hold
2154 : * them.
2155 : */
2156 288 : state->outputTapes = palloc0(state->nInputTapes * sizeof(LogicalTape *));
2157 288 : state->nOutputTapes = 0;
2158 288 : state->nOutputRuns = 0;
2159 :
2160 : /*
2161 : * Redistribute the memory allocated for tape buffers, among the
2162 : * new input and output tapes.
2163 : */
2164 288 : input_buffer_size = merge_read_buffer_size(state->tape_buffer_mem,
2165 : state->nInputTapes,
2166 : state->nInputRuns,
2167 : state->maxTapes);
2168 :
2169 : #ifdef TRACE_SORT
2170 288 : if (trace_sort)
2171 0 : elog(LOG, "starting merge pass of %d input runs on %d tapes, " INT64_FORMAT " KB of memory for each input tape: %s",
2172 : state->nInputRuns, state->nInputTapes, input_buffer_size / 1024,
2173 : pg_rusage_show(&state->ru_start));
2174 : #endif
2175 :
2176 : /* Prepare the new input tapes for merge pass. */
2177 1162 : for (tapenum = 0; tapenum < state->nInputTapes; tapenum++)
2178 874 : LogicalTapeRewindForRead(state->inputTapes[tapenum], input_buffer_size);
2179 :
2180 : /*
2181 : * If there's just one run left on each input tape, then only one
2182 : * merge pass remains. If we don't have to produce a materialized
2183 : * sorted tape, we can stop at this point and do the final merge
2184 : * on-the-fly.
2185 : */
2186 288 : if ((state->base.sortopt & TUPLESORT_RANDOMACCESS) == 0
2187 270 : && state->nInputRuns <= state->nInputTapes
2188 240 : && !WORKER(state))
2189 : {
2190 : /* Tell logtape.c we won't be writing anymore */
2191 240 : LogicalTapeSetForgetFreeSpace(state->tapeset);
2192 : /* Initialize for the final merge pass */
2193 240 : beginmerge(state);
2194 240 : state->status = TSS_FINALMERGE;
2195 240 : return;
2196 : }
2197 : }
2198 :
2199 : /* Select an output tape */
2200 156 : selectnewtape(state);
2201 :
2202 : /* Merge one run from each input tape. */
2203 156 : mergeonerun(state);
2204 :
2205 : /*
2206 : * If the input tapes are empty, and we output only one output run,
2207 : * we're done. The current output tape contains the final result.
2208 : */
2209 156 : if (state->nInputRuns == 0 && state->nOutputRuns <= 1)
2210 18 : break;
2211 : }
2212 :
2213 : /*
2214 : * Done. The result is on a single run on a single tape.
2215 : */
2216 18 : state->result_tape = state->outputTapes[0];
2217 18 : if (!WORKER(state))
2218 18 : LogicalTapeFreeze(state->result_tape, NULL);
2219 : else
2220 0 : worker_freeze_result_tape(state);
2221 18 : state->status = TSS_SORTEDONTAPE;
2222 :
2223 : /* Close all the now-empty input tapes, to release their read buffers. */
2224 102 : for (tapenum = 0; tapenum < state->nInputTapes; tapenum++)
2225 84 : LogicalTapeClose(state->inputTapes[tapenum]);
2226 : }
2227 :
2228 : /*
2229 : * Merge one run from each input tape.
2230 : */
2231 : static void
2232 156 : mergeonerun(Tuplesortstate *state)
2233 : {
2234 : int srcTapeIndex;
2235 : LogicalTape *srcTape;
2236 :
2237 : /*
2238 : * Start the merge by loading one tuple from each active source tape into
2239 : * the heap.
2240 : */
2241 156 : beginmerge(state);
2242 :
2243 : Assert(state->slabAllocatorUsed);
2244 :
2245 : /*
2246 : * Execute merge by repeatedly extracting lowest tuple in heap, writing it
2247 : * out, and replacing it with next tuple from same tape (if there is
2248 : * another one).
2249 : */
2250 855588 : while (state->memtupcount > 0)
2251 : {
2252 : SortTuple stup;
2253 :
2254 : /* write the tuple to destTape */
2255 855432 : srcTapeIndex = state->memtuples[0].srctape;
2256 855432 : srcTape = state->inputTapes[srcTapeIndex];
2257 855432 : WRITETUP(state, state->destTape, &state->memtuples[0]);
2258 :
2259 : /* recycle the slot of the tuple we just wrote out, for the next read */
2260 855432 : if (state->memtuples[0].tuple)
2261 735348 : RELEASE_SLAB_SLOT(state, state->memtuples[0].tuple);
2262 :
2263 : /*
2264 : * pull next tuple from the tape, and replace the written-out tuple in
2265 : * the heap with it.
2266 : */
2267 855432 : if (mergereadnext(state, srcTape, &stup))
2268 : {
2269 854586 : stup.srctape = srcTapeIndex;
2270 854586 : tuplesort_heap_replace_top(state, &stup);
2271 : }
2272 : else
2273 : {
2274 846 : tuplesort_heap_delete_top(state);
2275 846 : state->nInputRuns--;
2276 : }
2277 : }
2278 :
2279 : /*
2280 : * When the heap empties, we're done. Write an end-of-run marker on the
2281 : * output tape.
2282 : */
2283 156 : markrunend(state->destTape);
2284 156 : }
2285 :
2286 : /*
2287 : * beginmerge - initialize for a merge pass
2288 : *
2289 : * Fill the merge heap with the first tuple from each input tape.
2290 : */
2291 : static void
2292 396 : beginmerge(Tuplesortstate *state)
2293 : {
2294 : int activeTapes;
2295 : int srcTapeIndex;
2296 :
2297 : /* Heap should be empty here */
2298 : Assert(state->memtupcount == 0);
2299 :
2300 396 : activeTapes = Min(state->nInputTapes, state->nInputRuns);
2301 :
2302 1852 : for (srcTapeIndex = 0; srcTapeIndex < activeTapes; srcTapeIndex++)
2303 : {
2304 : SortTuple tup;
2305 :
2306 1456 : if (mergereadnext(state, state->inputTapes[srcTapeIndex], &tup))
2307 : {
2308 1228 : tup.srctape = srcTapeIndex;
2309 1228 : tuplesort_heap_insert(state, &tup);
2310 : }
2311 : }
2312 396 : }
2313 :
2314 : /*
2315 : * mergereadnext - read next tuple from one merge input tape
2316 : *
2317 : * Returns false on EOF.
2318 : */
2319 : static bool
2320 5412218 : mergereadnext(Tuplesortstate *state, LogicalTape *srcTape, SortTuple *stup)
2321 : {
2322 : unsigned int tuplen;
2323 :
2324 : /* read next tuple, if any */
2325 5412218 : if ((tuplen = getlen(srcTape, true)) == 0)
2326 1408 : return false;
2327 5410810 : READTUP(state, stup, srcTape, tuplen);
2328 :
2329 5410810 : return true;
2330 : }
2331 :
2332 : /*
2333 : * dumptuples - remove tuples from memtuples and write initial run to tape
2334 : *
2335 : * When alltuples = true, dump everything currently in memory. (This case is
2336 : * only used at end of input data.)
2337 : */
2338 : static void
2339 1061418 : dumptuples(Tuplesortstate *state, bool alltuples)
2340 : {
2341 : int memtupwrite;
2342 : int i;
2343 :
2344 : /*
2345 : * Nothing to do if we still fit in available memory and have array slots,
2346 : * unless this is the final call during initial run generation.
2347 : */
2348 1061418 : if (state->memtupcount < state->memtupsize && !LACKMEM(state) &&
2349 1060506 : !alltuples)
2350 1059972 : return;
2351 :
2352 : /*
2353 : * Final call might require no sorting, in rare cases where we just so
2354 : * happen to have previously LACKMEM()'d at the point where exactly all
2355 : * remaining tuples are loaded into memory, just before input was
2356 : * exhausted. In general, short final runs are quite possible, but avoid
2357 : * creating a completely empty run. In a worker, though, we must produce
2358 : * at least one tape, even if it's empty.
2359 : */
2360 1446 : if (state->memtupcount == 0 && state->currentRun > 0)
2361 0 : return;
2362 :
2363 : Assert(state->status == TSS_BUILDRUNS);
2364 :
2365 : /*
2366 : * It seems unlikely that this limit will ever be exceeded, but take no
2367 : * chances
2368 : */
2369 1446 : if (state->currentRun == INT_MAX)
2370 0 : ereport(ERROR,
2371 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
2372 : errmsg("cannot have more than %d runs for an external sort",
2373 : INT_MAX)));
2374 :
2375 1446 : if (state->currentRun > 0)
2376 912 : selectnewtape(state);
2377 :
2378 1446 : state->currentRun++;
2379 :
2380 : #ifdef TRACE_SORT
2381 1446 : if (trace_sort)
2382 0 : elog(LOG, "worker %d starting quicksort of run %d: %s",
2383 : state->worker, state->currentRun,
2384 : pg_rusage_show(&state->ru_start));
2385 : #endif
2386 :
2387 : /*
2388 : * Sort all tuples accumulated within the allowed amount of memory for
2389 : * this run using quicksort
2390 : */
2391 1446 : tuplesort_sort_memtuples(state);
2392 :
2393 : #ifdef TRACE_SORT
2394 1446 : if (trace_sort)
2395 0 : elog(LOG, "worker %d finished quicksort of run %d: %s",
2396 : state->worker, state->currentRun,
2397 : pg_rusage_show(&state->ru_start));
2398 : #endif
2399 :
2400 1446 : memtupwrite = state->memtupcount;
2401 5052028 : for (i = 0; i < memtupwrite; i++)
2402 : {
2403 5050582 : SortTuple *stup = &state->memtuples[i];
2404 :
2405 5050582 : WRITETUP(state, state->destTape, stup);
2406 : }
2407 :
2408 1446 : state->memtupcount = 0;
2409 :
2410 : /*
2411 : * Reset tuple memory. We've freed all of the tuples that we previously
2412 : * allocated. It's important to avoid fragmentation when there is a stark
2413 : * change in the sizes of incoming tuples. In bounded sorts,
2414 : * fragmentation due to AllocSetFree's bucketing by size class might be
2415 : * particularly bad if this step wasn't taken.
2416 : */
2417 1446 : MemoryContextReset(state->base.tuplecontext);
2418 :
2419 : /*
2420 : * Now update the memory accounting to subtract the memory used by the
2421 : * tuple.
2422 : */
2423 1446 : FREEMEM(state, state->tupleMem);
2424 1446 : state->tupleMem = 0;
2425 :
2426 1446 : markrunend(state->destTape);
2427 :
2428 : #ifdef TRACE_SORT
2429 1446 : if (trace_sort)
2430 0 : elog(LOG, "worker %d finished writing run %d to tape %d: %s",
2431 : state->worker, state->currentRun, (state->currentRun - 1) % state->nOutputTapes + 1,
2432 : pg_rusage_show(&state->ru_start));
2433 : #endif
2434 : }
2435 :
2436 : /*
2437 : * tuplesort_rescan - rewind and replay the scan
2438 : */
2439 : void
2440 46 : tuplesort_rescan(Tuplesortstate *state)
2441 : {
2442 46 : MemoryContext oldcontext = MemoryContextSwitchTo(state->base.sortcontext);
2443 :
2444 : Assert(state->base.sortopt & TUPLESORT_RANDOMACCESS);
2445 :
2446 46 : switch (state->status)
2447 : {
2448 40 : case TSS_SORTEDINMEM:
2449 40 : state->current = 0;
2450 40 : state->eof_reached = false;
2451 40 : state->markpos_offset = 0;
2452 40 : state->markpos_eof = false;
2453 40 : break;
2454 6 : case TSS_SORTEDONTAPE:
2455 6 : LogicalTapeRewindForRead(state->result_tape, 0);
2456 6 : state->eof_reached = false;
2457 6 : state->markpos_block = 0L;
2458 6 : state->markpos_offset = 0;
2459 6 : state->markpos_eof = false;
2460 6 : break;
2461 0 : default:
2462 0 : elog(ERROR, "invalid tuplesort state");
2463 : break;
2464 : }
2465 :
2466 46 : MemoryContextSwitchTo(oldcontext);
2467 46 : }
2468 :
2469 : /*
2470 : * tuplesort_markpos - saves current position in the merged sort file
2471 : */
2472 : void
2473 569360 : tuplesort_markpos(Tuplesortstate *state)
2474 : {
2475 569360 : MemoryContext oldcontext = MemoryContextSwitchTo(state->base.sortcontext);
2476 :
2477 : Assert(state->base.sortopt & TUPLESORT_RANDOMACCESS);
2478 :
2479 569360 : switch (state->status)
2480 : {
2481 560552 : case TSS_SORTEDINMEM:
2482 560552 : state->markpos_offset = state->current;
2483 560552 : state->markpos_eof = state->eof_reached;
2484 560552 : break;
2485 8808 : case TSS_SORTEDONTAPE:
2486 8808 : LogicalTapeTell(state->result_tape,
2487 : &state->markpos_block,
2488 : &state->markpos_offset);
2489 8808 : state->markpos_eof = state->eof_reached;
2490 8808 : break;
2491 0 : default:
2492 0 : elog(ERROR, "invalid tuplesort state");
2493 : break;
2494 : }
2495 :
2496 569360 : MemoryContextSwitchTo(oldcontext);
2497 569360 : }
2498 :
2499 : /*
2500 : * tuplesort_restorepos - restores current position in merged sort file to
2501 : * last saved position
2502 : */
2503 : void
2504 30062 : tuplesort_restorepos(Tuplesortstate *state)
2505 : {
2506 30062 : MemoryContext oldcontext = MemoryContextSwitchTo(state->base.sortcontext);
2507 :
2508 : Assert(state->base.sortopt & TUPLESORT_RANDOMACCESS);
2509 :
2510 30062 : switch (state->status)
2511 : {
2512 23870 : case TSS_SORTEDINMEM:
2513 23870 : state->current = state->markpos_offset;
2514 23870 : state->eof_reached = state->markpos_eof;
2515 23870 : break;
2516 6192 : case TSS_SORTEDONTAPE:
2517 6192 : LogicalTapeSeek(state->result_tape,
2518 : state->markpos_block,
2519 : state->markpos_offset);
2520 6192 : state->eof_reached = state->markpos_eof;
2521 6192 : break;
2522 0 : default:
2523 0 : elog(ERROR, "invalid tuplesort state");
2524 : break;
2525 : }
2526 :
2527 30062 : MemoryContextSwitchTo(oldcontext);
2528 30062 : }
2529 :
2530 : /*
2531 : * tuplesort_get_stats - extract summary statistics
2532 : *
2533 : * This can be called after tuplesort_performsort() finishes to obtain
2534 : * printable summary information about how the sort was performed.
2535 : */
2536 : void
2537 384 : tuplesort_get_stats(Tuplesortstate *state,
2538 : TuplesortInstrumentation *stats)
2539 : {
2540 : /*
2541 : * Note: it might seem we should provide both memory and disk usage for a
2542 : * disk-based sort. However, the current code doesn't track memory space
2543 : * accurately once we have begun to return tuples to the caller (since we
2544 : * don't account for pfree's the caller is expected to do), so we cannot
2545 : * rely on availMem in a disk sort. This does not seem worth the overhead
2546 : * to fix. Is it worth creating an API for the memory context code to
2547 : * tell us how much is actually used in sortcontext?
2548 : */
2549 384 : tuplesort_updatemax(state);
2550 :
2551 384 : if (state->isMaxSpaceDisk)
2552 0 : stats->spaceType = SORT_SPACE_TYPE_DISK;
2553 : else
2554 384 : stats->spaceType = SORT_SPACE_TYPE_MEMORY;
2555 384 : stats->spaceUsed = (state->maxSpace + 1023) / 1024;
2556 :
2557 384 : switch (state->maxSpaceStatus)
2558 : {
2559 384 : case TSS_SORTEDINMEM:
2560 384 : if (state->boundUsed)
2561 42 : stats->sortMethod = SORT_TYPE_TOP_N_HEAPSORT;
2562 : else
2563 342 : stats->sortMethod = SORT_TYPE_QUICKSORT;
2564 384 : break;
2565 0 : case TSS_SORTEDONTAPE:
2566 0 : stats->sortMethod = SORT_TYPE_EXTERNAL_SORT;
2567 0 : break;
2568 0 : case TSS_FINALMERGE:
2569 0 : stats->sortMethod = SORT_TYPE_EXTERNAL_MERGE;
2570 0 : break;
2571 0 : default:
2572 0 : stats->sortMethod = SORT_TYPE_STILL_IN_PROGRESS;
2573 0 : break;
2574 : }
2575 384 : }
2576 :
2577 : /*
2578 : * Convert TuplesortMethod to a string.
2579 : */
2580 : const char *
2581 282 : tuplesort_method_name(TuplesortMethod m)
2582 : {
2583 282 : switch (m)
2584 : {
2585 0 : case SORT_TYPE_STILL_IN_PROGRESS:
2586 0 : return "still in progress";
2587 42 : case SORT_TYPE_TOP_N_HEAPSORT:
2588 42 : return "top-N heapsort";
2589 240 : case SORT_TYPE_QUICKSORT:
2590 240 : return "quicksort";
2591 0 : case SORT_TYPE_EXTERNAL_SORT:
2592 0 : return "external sort";
2593 0 : case SORT_TYPE_EXTERNAL_MERGE:
2594 0 : return "external merge";
2595 : }
2596 :
2597 0 : return "unknown";
2598 : }
2599 :
2600 : /*
2601 : * Convert TuplesortSpaceType to a string.
2602 : */
2603 : const char *
2604 246 : tuplesort_space_type_name(TuplesortSpaceType t)
2605 : {
2606 : Assert(t == SORT_SPACE_TYPE_DISK || t == SORT_SPACE_TYPE_MEMORY);
2607 246 : return t == SORT_SPACE_TYPE_DISK ? "Disk" : "Memory";
2608 : }
2609 :
2610 :
2611 : /*
2612 : * Heap manipulation routines, per Knuth's Algorithm 5.2.3H.
2613 : */
2614 :
2615 : /*
2616 : * Convert the existing unordered array of SortTuples to a bounded heap,
2617 : * discarding all but the smallest "state->bound" tuples.
2618 : *
2619 : * When working with a bounded heap, we want to keep the largest entry
2620 : * at the root (array entry zero), instead of the smallest as in the normal
2621 : * sort case. This allows us to discard the largest entry cheaply.
2622 : * Therefore, we temporarily reverse the sort direction.
2623 : */
2624 : static void
2625 398 : make_bounded_heap(Tuplesortstate *state)
2626 : {
2627 398 : int tupcount = state->memtupcount;
2628 : int i;
2629 :
2630 : Assert(state->status == TSS_INITIAL);
2631 : Assert(state->bounded);
2632 : Assert(tupcount >= state->bound);
2633 : Assert(SERIAL(state));
2634 :
2635 : /* Reverse sort direction so largest entry will be at root */
2636 398 : reversedirection(state);
2637 :
2638 398 : state->memtupcount = 0; /* make the heap empty */
2639 37888 : for (i = 0; i < tupcount; i++)
2640 : {
2641 37490 : if (state->memtupcount < state->bound)
2642 : {
2643 : /* Insert next tuple into heap */
2644 : /* Must copy source tuple to avoid possible overwrite */
2645 18546 : SortTuple stup = state->memtuples[i];
2646 :
2647 18546 : tuplesort_heap_insert(state, &stup);
2648 : }
2649 : else
2650 : {
2651 : /*
2652 : * The heap is full. Replace the largest entry with the new
2653 : * tuple, or just discard it, if it's larger than anything already
2654 : * in the heap.
2655 : */
2656 18944 : if (COMPARETUP(state, &state->memtuples[i], &state->memtuples[0]) <= 0)
2657 : {
2658 9766 : free_sort_tuple(state, &state->memtuples[i]);
2659 9766 : CHECK_FOR_INTERRUPTS();
2660 : }
2661 : else
2662 9178 : tuplesort_heap_replace_top(state, &state->memtuples[i]);
2663 : }
2664 : }
2665 :
2666 : Assert(state->memtupcount == state->bound);
2667 398 : state->status = TSS_BOUNDED;
2668 398 : }
2669 :
2670 : /*
2671 : * Convert the bounded heap to a properly-sorted array
2672 : */
2673 : static void
2674 398 : sort_bounded_heap(Tuplesortstate *state)
2675 : {
2676 398 : int tupcount = state->memtupcount;
2677 :
2678 : Assert(state->status == TSS_BOUNDED);
2679 : Assert(state->bounded);
2680 : Assert(tupcount == state->bound);
2681 : Assert(SERIAL(state));
2682 :
2683 : /*
2684 : * We can unheapify in place because each delete-top call will remove the
2685 : * largest entry, which we can promptly store in the newly freed slot at
2686 : * the end. Once we're down to a single-entry heap, we're done.
2687 : */
2688 18546 : while (state->memtupcount > 1)
2689 : {
2690 18148 : SortTuple stup = state->memtuples[0];
2691 :
2692 : /* this sifts-up the next-largest entry and decreases memtupcount */
2693 18148 : tuplesort_heap_delete_top(state);
2694 18148 : state->memtuples[state->memtupcount] = stup;
2695 : }
2696 398 : state->memtupcount = tupcount;
2697 :
2698 : /*
2699 : * Reverse sort direction back to the original state. This is not
2700 : * actually necessary but seems like a good idea for tidiness.
2701 : */
2702 398 : reversedirection(state);
2703 :
2704 398 : state->status = TSS_SORTEDINMEM;
2705 398 : state->boundUsed = true;
2706 398 : }
2707 :
2708 : /*
2709 : * Sort all memtuples using specialized qsort() routines.
2710 : *
2711 : * Quicksort is used for small in-memory sorts, and external sort runs.
2712 : */
2713 : static void
2714 199756 : tuplesort_sort_memtuples(Tuplesortstate *state)
2715 : {
2716 : Assert(!LEADER(state));
2717 :
2718 199756 : if (state->memtupcount > 1)
2719 : {
2720 : /*
2721 : * Do we have the leading column's value or abbreviation in datum1,
2722 : * and is there a specialization for its comparator?
2723 : */
2724 60896 : if (state->base.haveDatum1 && state->base.sortKeys)
2725 : {
2726 60862 : if (state->base.sortKeys[0].comparator == ssup_datum_unsigned_cmp)
2727 : {
2728 2922 : qsort_tuple_unsigned(state->memtuples,
2729 2922 : state->memtupcount,
2730 : state);
2731 2906 : return;
2732 : }
2733 : #if SIZEOF_DATUM >= 8
2734 57940 : else if (state->base.sortKeys[0].comparator == ssup_datum_signed_cmp)
2735 : {
2736 1042 : qsort_tuple_signed(state->memtuples,
2737 1042 : state->memtupcount,
2738 : state);
2739 1042 : return;
2740 : }
2741 : #endif
2742 56898 : else if (state->base.sortKeys[0].comparator == ssup_datum_int32_cmp)
2743 : {
2744 38092 : qsort_tuple_int32(state->memtuples,
2745 38092 : state->memtupcount,
2746 : state);
2747 38032 : return;
2748 : }
2749 : }
2750 :
2751 : /* Can we use the single-key sort function? */
2752 18840 : if (state->base.onlyKey != NULL)
2753 : {
2754 8048 : qsort_ssup(state->memtuples, state->memtupcount,
2755 8048 : state->base.onlyKey);
2756 : }
2757 : else
2758 : {
2759 10792 : qsort_tuple(state->memtuples,
2760 10792 : state->memtupcount,
2761 : state->base.comparetup,
2762 : state);
2763 : }
2764 : }
2765 : }
2766 :
2767 : /*
2768 : * Insert a new tuple into an empty or existing heap, maintaining the
2769 : * heap invariant. Caller is responsible for ensuring there's room.
2770 : *
2771 : * Note: For some callers, tuple points to a memtuples[] entry above the
2772 : * end of the heap. This is safe as long as it's not immediately adjacent
2773 : * to the end of the heap (ie, in the [memtupcount] array entry) --- if it
2774 : * is, it might get overwritten before being moved into the heap!
2775 : */
2776 : static void
2777 19774 : tuplesort_heap_insert(Tuplesortstate *state, SortTuple *tuple)
2778 : {
2779 : SortTuple *memtuples;
2780 : int j;
2781 :
2782 19774 : memtuples = state->memtuples;
2783 : Assert(state->memtupcount < state->memtupsize);
2784 :
2785 19774 : CHECK_FOR_INTERRUPTS();
2786 :
2787 : /*
2788 : * Sift-up the new entry, per Knuth 5.2.3 exercise 16. Note that Knuth is
2789 : * using 1-based array indexes, not 0-based.
2790 : */
2791 19774 : j = state->memtupcount++;
2792 57602 : while (j > 0)
2793 : {
2794 50856 : int i = (j - 1) >> 1;
2795 :
2796 50856 : if (COMPARETUP(state, tuple, &memtuples[i]) >= 0)
2797 13028 : break;
2798 37828 : memtuples[j] = memtuples[i];
2799 37828 : j = i;
2800 : }
2801 19774 : memtuples[j] = *tuple;
2802 19774 : }
2803 :
2804 : /*
2805 : * Remove the tuple at state->memtuples[0] from the heap. Decrement
2806 : * memtupcount, and sift up to maintain the heap invariant.
2807 : *
2808 : * The caller has already free'd the tuple the top node points to,
2809 : * if necessary.
2810 : */
2811 : static void
2812 19328 : tuplesort_heap_delete_top(Tuplesortstate *state)
2813 : {
2814 19328 : SortTuple *memtuples = state->memtuples;
2815 : SortTuple *tuple;
2816 :
2817 19328 : if (--state->memtupcount <= 0)
2818 276 : return;
2819 :
2820 : /*
2821 : * Remove the last tuple in the heap, and re-insert it, by replacing the
2822 : * current top node with it.
2823 : */
2824 19052 : tuple = &memtuples[state->memtupcount];
2825 19052 : tuplesort_heap_replace_top(state, tuple);
2826 : }
2827 :
2828 : /*
2829 : * Replace the tuple at state->memtuples[0] with a new tuple. Sift up to
2830 : * maintain the heap invariant.
2831 : *
2832 : * This corresponds to Knuth's "sift-up" algorithm (Algorithm 5.2.3H,
2833 : * Heapsort, steps H3-H8).
2834 : */
2835 : static void
2836 5940714 : tuplesort_heap_replace_top(Tuplesortstate *state, SortTuple *tuple)
2837 : {
2838 5940714 : SortTuple *memtuples = state->memtuples;
2839 : unsigned int i,
2840 : n;
2841 :
2842 : Assert(state->memtupcount >= 1);
2843 :
2844 5940714 : CHECK_FOR_INTERRUPTS();
2845 :
2846 : /*
2847 : * state->memtupcount is "int", but we use "unsigned int" for i, j, n.
2848 : * This prevents overflow in the "2 * i + 1" calculation, since at the top
2849 : * of the loop we must have i < n <= INT_MAX <= UINT_MAX/2.
2850 : */
2851 5940714 : n = state->memtupcount;
2852 5940714 : i = 0; /* i is where the "hole" is */
2853 : for (;;)
2854 1679886 : {
2855 7620600 : unsigned int j = 2 * i + 1;
2856 :
2857 7620600 : if (j >= n)
2858 1063442 : break;
2859 8990398 : if (j + 1 < n &&
2860 2433240 : COMPARETUP(state, &memtuples[j], &memtuples[j + 1]) > 0)
2861 960578 : j++;
2862 6557158 : if (COMPARETUP(state, tuple, &memtuples[j]) <= 0)
2863 4877272 : break;
2864 1679886 : memtuples[i] = memtuples[j];
2865 1679886 : i = j;
2866 : }
2867 5940714 : memtuples[i] = *tuple;
2868 5940714 : }
2869 :
2870 : /*
2871 : * Function to reverse the sort direction from its current state
2872 : *
2873 : * It is not safe to call this when performing hash tuplesorts
2874 : */
2875 : static void
2876 796 : reversedirection(Tuplesortstate *state)
2877 : {
2878 796 : SortSupport sortKey = state->base.sortKeys;
2879 : int nkey;
2880 :
2881 1952 : for (nkey = 0; nkey < state->base.nKeys; nkey++, sortKey++)
2882 : {
2883 1156 : sortKey->ssup_reverse = !sortKey->ssup_reverse;
2884 1156 : sortKey->ssup_nulls_first = !sortKey->ssup_nulls_first;
2885 : }
2886 796 : }
2887 :
2888 :
2889 : /*
2890 : * Tape interface routines
2891 : */
2892 :
2893 : static unsigned int
2894 5685224 : getlen(LogicalTape *tape, bool eofOK)
2895 : {
2896 : unsigned int len;
2897 :
2898 5685224 : if (LogicalTapeRead(tape,
2899 : &len, sizeof(len)) != sizeof(len))
2900 0 : elog(ERROR, "unexpected end of tape");
2901 5685224 : if (len == 0 && !eofOK)
2902 0 : elog(ERROR, "unexpected end of data");
2903 5685224 : return len;
2904 : }
2905 :
2906 : static void
2907 1602 : markrunend(LogicalTape *tape)
2908 : {
2909 1602 : unsigned int len = 0;
2910 :
2911 1602 : LogicalTapeWrite(tape, &len, sizeof(len));
2912 1602 : }
2913 :
2914 : /*
2915 : * Get memory for tuple from within READTUP() routine.
2916 : *
2917 : * We use next free slot from the slab allocator, or palloc() if the tuple
2918 : * is too large for that.
2919 : */
2920 : void *
2921 5383546 : tuplesort_readtup_alloc(Tuplesortstate *state, Size tuplen)
2922 : {
2923 : SlabSlot *buf;
2924 :
2925 : /*
2926 : * We pre-allocate enough slots in the slab arena that we should never run
2927 : * out.
2928 : */
2929 : Assert(state->slabFreeHead);
2930 :
2931 5383546 : if (tuplen > SLAB_SLOT_SIZE || !state->slabFreeHead)
2932 0 : return MemoryContextAlloc(state->base.sortcontext, tuplen);
2933 : else
2934 : {
2935 5383546 : buf = state->slabFreeHead;
2936 : /* Reuse this slot */
2937 5383546 : state->slabFreeHead = buf->nextfree;
2938 :
2939 5383546 : return buf;
2940 : }
2941 : }
2942 :
2943 :
2944 : /*
2945 : * Parallel sort routines
2946 : */
2947 :
2948 : /*
2949 : * tuplesort_estimate_shared - estimate required shared memory allocation
2950 : *
2951 : * nWorkers is an estimate of the number of workers (it's the number that
2952 : * will be requested).
2953 : */
2954 : Size
2955 146 : tuplesort_estimate_shared(int nWorkers)
2956 : {
2957 : Size tapesSize;
2958 :
2959 : Assert(nWorkers > 0);
2960 :
2961 : /* Make sure that BufFile shared state is MAXALIGN'd */
2962 146 : tapesSize = mul_size(sizeof(TapeShare), nWorkers);
2963 146 : tapesSize = MAXALIGN(add_size(tapesSize, offsetof(Sharedsort, tapes)));
2964 :
2965 146 : return tapesSize;
2966 : }
2967 :
2968 : /*
2969 : * tuplesort_initialize_shared - initialize shared tuplesort state
2970 : *
2971 : * Must be called from leader process before workers are launched, to
2972 : * establish state needed up-front for worker tuplesortstates. nWorkers
2973 : * should match the argument passed to tuplesort_estimate_shared().
2974 : */
2975 : void
2976 210 : tuplesort_initialize_shared(Sharedsort *shared, int nWorkers, dsm_segment *seg)
2977 : {
2978 : int i;
2979 :
2980 : Assert(nWorkers > 0);
2981 :
2982 210 : SpinLockInit(&shared->mutex);
2983 210 : shared->currentWorker = 0;
2984 210 : shared->workersFinished = 0;
2985 210 : SharedFileSetInit(&shared->fileset, seg);
2986 210 : shared->nTapes = nWorkers;
2987 638 : for (i = 0; i < nWorkers; i++)
2988 : {
2989 428 : shared->tapes[i].firstblocknumber = 0L;
2990 : }
2991 210 : }
2992 :
2993 : /*
2994 : * tuplesort_attach_shared - attach to shared tuplesort state
2995 : *
2996 : * Must be called by all worker processes.
2997 : */
2998 : void
2999 212 : tuplesort_attach_shared(Sharedsort *shared, dsm_segment *seg)
3000 : {
3001 : /* Attach to SharedFileSet */
3002 212 : SharedFileSetAttach(&shared->fileset, seg);
3003 212 : }
3004 :
3005 : /*
3006 : * worker_get_identifier - Assign and return ordinal identifier for worker
3007 : *
3008 : * The order in which these are assigned is not well defined, and should not
3009 : * matter; worker numbers across parallel sort participants need only be
3010 : * distinct and gapless. logtape.c requires this.
3011 : *
3012 : * Note that the identifiers assigned from here have no relation to
3013 : * ParallelWorkerNumber number, to avoid making any assumption about
3014 : * caller's requirements. However, we do follow the ParallelWorkerNumber
3015 : * convention of representing a non-worker with worker number -1. This
3016 : * includes the leader, as well as serial Tuplesort processes.
3017 : */
3018 : static int
3019 420 : worker_get_identifier(Tuplesortstate *state)
3020 : {
3021 420 : Sharedsort *shared = state->shared;
3022 : int worker;
3023 :
3024 : Assert(WORKER(state));
3025 :
3026 420 : SpinLockAcquire(&shared->mutex);
3027 420 : worker = shared->currentWorker++;
3028 420 : SpinLockRelease(&shared->mutex);
3029 :
3030 420 : return worker;
3031 : }
3032 :
3033 : /*
3034 : * worker_freeze_result_tape - freeze worker's result tape for leader
3035 : *
3036 : * This is called by workers just after the result tape has been determined,
3037 : * instead of calling LogicalTapeFreeze() directly. They do so because
3038 : * workers require a few additional steps over similar serial
3039 : * TSS_SORTEDONTAPE external sort cases, which also happen here. The extra
3040 : * steps are around freeing now unneeded resources, and representing to
3041 : * leader that worker's input run is available for its merge.
3042 : *
3043 : * There should only be one final output run for each worker, which consists
3044 : * of all tuples that were originally input into worker.
3045 : */
3046 : static void
3047 420 : worker_freeze_result_tape(Tuplesortstate *state)
3048 : {
3049 420 : Sharedsort *shared = state->shared;
3050 : TapeShare output;
3051 :
3052 : Assert(WORKER(state));
3053 : Assert(state->result_tape != NULL);
3054 : Assert(state->memtupcount == 0);
3055 :
3056 : /*
3057 : * Free most remaining memory, in case caller is sensitive to our holding
3058 : * on to it. memtuples may not be a tiny merge heap at this point.
3059 : */
3060 420 : pfree(state->memtuples);
3061 : /* Be tidy */
3062 420 : state->memtuples = NULL;
3063 420 : state->memtupsize = 0;
3064 :
3065 : /*
3066 : * Parallel worker requires result tape metadata, which is to be stored in
3067 : * shared memory for leader
3068 : */
3069 420 : LogicalTapeFreeze(state->result_tape, &output);
3070 :
3071 : /* Store properties of output tape, and update finished worker count */
3072 420 : SpinLockAcquire(&shared->mutex);
3073 420 : shared->tapes[state->worker] = output;
3074 420 : shared->workersFinished++;
3075 420 : SpinLockRelease(&shared->mutex);
3076 420 : }
3077 :
3078 : /*
3079 : * worker_nomergeruns - dump memtuples in worker, without merging
3080 : *
3081 : * This called as an alternative to mergeruns() with a worker when no
3082 : * merging is required.
3083 : */
3084 : static void
3085 420 : worker_nomergeruns(Tuplesortstate *state)
3086 : {
3087 : Assert(WORKER(state));
3088 : Assert(state->result_tape == NULL);
3089 : Assert(state->nOutputRuns == 1);
3090 :
3091 420 : state->result_tape = state->destTape;
3092 420 : worker_freeze_result_tape(state);
3093 420 : }
3094 :
3095 : /*
3096 : * leader_takeover_tapes - create tapeset for leader from worker tapes
3097 : *
3098 : * So far, leader Tuplesortstate has performed no actual sorting. By now, all
3099 : * sorting has occurred in workers, all of which must have already returned
3100 : * from tuplesort_performsort().
3101 : *
3102 : * When this returns, leader process is left in a state that is virtually
3103 : * indistinguishable from it having generated runs as a serial external sort
3104 : * might have.
3105 : */
3106 : static void
3107 144 : leader_takeover_tapes(Tuplesortstate *state)
3108 : {
3109 144 : Sharedsort *shared = state->shared;
3110 144 : int nParticipants = state->nParticipants;
3111 : int workersFinished;
3112 : int j;
3113 :
3114 : Assert(LEADER(state));
3115 : Assert(nParticipants >= 1);
3116 :
3117 144 : SpinLockAcquire(&shared->mutex);
3118 144 : workersFinished = shared->workersFinished;
3119 144 : SpinLockRelease(&shared->mutex);
3120 :
3121 144 : if (nParticipants != workersFinished)
3122 0 : elog(ERROR, "cannot take over tapes before all workers finish");
3123 :
3124 : /*
3125 : * Create the tapeset from worker tapes, including a leader-owned tape at
3126 : * the end. Parallel workers are far more expensive than logical tapes,
3127 : * so the number of tapes allocated here should never be excessive.
3128 : */
3129 144 : inittapestate(state, nParticipants);
3130 144 : state->tapeset = LogicalTapeSetCreate(false, &shared->fileset, -1);
3131 :
3132 : /*
3133 : * Set currentRun to reflect the number of runs we will merge (it's not
3134 : * used for anything, this is just pro forma)
3135 : */
3136 144 : state->currentRun = nParticipants;
3137 :
3138 : /*
3139 : * Initialize the state to look the same as after building the initial
3140 : * runs.
3141 : *
3142 : * There will always be exactly 1 run per worker, and exactly one input
3143 : * tape per run, because workers always output exactly 1 run, even when
3144 : * there were no input tuples for workers to sort.
3145 : */
3146 144 : state->inputTapes = NULL;
3147 144 : state->nInputTapes = 0;
3148 144 : state->nInputRuns = 0;
3149 :
3150 144 : state->outputTapes = palloc0(nParticipants * sizeof(LogicalTape *));
3151 144 : state->nOutputTapes = nParticipants;
3152 144 : state->nOutputRuns = nParticipants;
3153 :
3154 436 : for (j = 0; j < nParticipants; j++)
3155 : {
3156 292 : state->outputTapes[j] = LogicalTapeImport(state->tapeset, j, &shared->tapes[j]);
3157 : }
3158 :
3159 144 : state->status = TSS_BUILDRUNS;
3160 144 : }
3161 :
3162 : /*
3163 : * Convenience routine to free a tuple previously loaded into sort memory
3164 : */
3165 : static void
3166 3743794 : free_sort_tuple(Tuplesortstate *state, SortTuple *stup)
3167 : {
3168 3743794 : if (stup->tuple)
3169 : {
3170 3583596 : FREEMEM(state, GetMemoryChunkSpace(stup->tuple));
3171 3583596 : pfree(stup->tuple);
3172 3583596 : stup->tuple = NULL;
3173 : }
3174 3743794 : }
3175 :
3176 : int
3177 0 : ssup_datum_unsigned_cmp(Datum x, Datum y, SortSupport ssup)
3178 : {
3179 0 : if (x < y)
3180 0 : return -1;
3181 0 : else if (x > y)
3182 0 : return 1;
3183 : else
3184 0 : return 0;
3185 : }
3186 :
3187 : #if SIZEOF_DATUM >= 8
3188 : int
3189 1045694 : ssup_datum_signed_cmp(Datum x, Datum y, SortSupport ssup)
3190 : {
3191 1045694 : int64 xx = DatumGetInt64(x);
3192 1045694 : int64 yy = DatumGetInt64(y);
3193 :
3194 1045694 : if (xx < yy)
3195 371156 : return -1;
3196 674538 : else if (xx > yy)
3197 347122 : return 1;
3198 : else
3199 327416 : return 0;
3200 : }
3201 : #endif
3202 :
3203 : int
3204 156924778 : ssup_datum_int32_cmp(Datum x, Datum y, SortSupport ssup)
3205 : {
3206 156924778 : int32 xx = DatumGetInt32(x);
3207 156924778 : int32 yy = DatumGetInt32(y);
3208 :
3209 156924778 : if (xx < yy)
3210 36851862 : return -1;
3211 120072916 : else if (xx > yy)
3212 35642168 : return 1;
3213 : else
3214 84430748 : return 0;
3215 : }
|
