Age Owner Branch data TLA 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. Large amounts are
11 : : * 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 : : * or radix sort 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 in memory 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 in-memory sorting, 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) 1996-2026, 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 "port/atomics.h"
108 : : #include "port/pg_bitutils.h"
109 : : #include "storage/shmem.h"
110 : : #include "utils/guc.h"
111 : : #include "utils/memutils.h"
112 : : #include "utils/pg_rusage.h"
113 : : #include "utils/tuplesort.h"
114 : :
115 : : /*
116 : : * Initial size of memtuples array. This must be more than
117 : : * ALLOCSET_SEPARATE_THRESHOLD; see comments in grow_memtuples(). Clamp at
118 : : * 1024 elements to avoid excessive reallocs.
119 : : */
120 : : #define INITIAL_MEMTUPSIZE Max(1024, \
121 : : ALLOCSET_SEPARATE_THRESHOLD / sizeof(SortTuple) + 1)
122 : :
123 : : /* GUC variables */
124 : : bool trace_sort = false;
125 : :
126 : : #ifdef DEBUG_BOUNDED_SORT
127 : : bool optimize_bounded_sort = true;
128 : : #endif
129 : :
130 : :
131 : : /*
132 : : * During merge, we use a pre-allocated set of fixed-size slots to hold
133 : : * tuples. To avoid palloc/pfree overhead.
134 : : *
135 : : * Merge doesn't require a lot of memory, so we can afford to waste some,
136 : : * by using gratuitously-sized slots. If a tuple is larger than 1 kB, the
137 : : * palloc() overhead is not significant anymore.
138 : : *
139 : : * 'nextfree' is valid when this chunk is in the free list. When in use, the
140 : : * slot holds a tuple.
141 : : */
142 : : #define SLAB_SLOT_SIZE 1024
143 : :
144 : : typedef union SlabSlot
145 : : {
146 : : union SlabSlot *nextfree;
147 : : char buffer[SLAB_SLOT_SIZE];
148 : : } SlabSlot;
149 : :
150 : : /*
151 : : * Possible states of a Tuplesort object. These denote the states that
152 : : * persist between calls of Tuplesort routines.
153 : : */
154 : : typedef enum
155 : : {
156 : : TSS_INITIAL, /* Loading tuples; still within memory limit */
157 : : TSS_BOUNDED, /* Loading tuples into bounded-size heap */
158 : : TSS_BUILDRUNS, /* Loading tuples; writing to tape */
159 : : TSS_SORTEDINMEM, /* Sort completed entirely in memory */
160 : : TSS_SORTEDONTAPE, /* Sort completed, final run is on tape */
161 : : TSS_FINALMERGE, /* Performing final merge on-the-fly */
162 : : } TupSortStatus;
163 : :
164 : : /*
165 : : * Parameters for calculation of number of tapes to use --- see inittapes()
166 : : * and tuplesort_merge_order().
167 : : *
168 : : * In this calculation we assume that each tape will cost us about 1 blocks
169 : : * worth of buffer space. This ignores the overhead of all the other data
170 : : * structures needed for each tape, but it's probably close enough.
171 : : *
172 : : * MERGE_BUFFER_SIZE is how much buffer space we'd like to allocate for each
173 : : * input tape, for pre-reading (see discussion at top of file). This is *in
174 : : * addition to* the 1 block already included in TAPE_BUFFER_OVERHEAD.
175 : : */
176 : : #define MINORDER 6 /* minimum merge order */
177 : : #define MAXORDER 500 /* maximum merge order */
178 : : #define TAPE_BUFFER_OVERHEAD BLCKSZ
179 : : #define MERGE_BUFFER_SIZE (BLCKSZ * 32)
180 : :
181 : :
182 : : /*
183 : : * Private state of a Tuplesort operation.
184 : : */
185 : : struct Tuplesortstate
186 : : {
187 : : TuplesortPublic base;
188 : : TupSortStatus status; /* enumerated value as shown above */
189 : : bool bounded; /* did caller specify a maximum number of
190 : : * tuples to return? */
191 : : bool boundUsed; /* true if we made use of a bounded heap */
192 : : int bound; /* if bounded, the maximum number of tuples */
193 : : int64 tupleMem; /* memory consumed by individual tuples.
194 : : * storing this separately from what we track
195 : : * in availMem allows us to subtract the
196 : : * memory consumed by all tuples when dumping
197 : : * tuples to tape */
198 : : int64 availMem; /* remaining memory available, in bytes */
199 : : int64 allowedMem; /* total memory allowed, in bytes */
200 : : int maxTapes; /* max number of input tapes to merge in each
201 : : * pass */
202 : : int64 maxSpace; /* maximum amount of space occupied among sort
203 : : * of groups, either in-memory or on-disk */
204 : : bool isMaxSpaceDisk; /* true when maxSpace tracks on-disk space,
205 : : * false means in-memory */
206 : : TupSortStatus maxSpaceStatus; /* sort status when maxSpace was reached */
207 : : LogicalTapeSet *tapeset; /* logtape.c object for tapes in a temp file */
208 : :
209 : : /*
210 : : * This array holds the tuples now in sort memory. If we are in state
211 : : * INITIAL, the tuples are in no particular order; if we are in state
212 : : * SORTEDINMEM, the tuples are in final sorted order; in states BUILDRUNS
213 : : * and FINALMERGE, the tuples are organized in "heap" order per Algorithm
214 : : * H. In state SORTEDONTAPE, the array is not used.
215 : : */
216 : : SortTuple *memtuples; /* array of SortTuple structs */
217 : : int memtupcount; /* number of tuples currently present */
218 : : int memtupsize; /* allocated length of memtuples array */
219 : : bool growmemtuples; /* memtuples' growth still underway? */
220 : :
221 : : /*
222 : : * Memory for tuples is sometimes allocated using a simple slab allocator,
223 : : * rather than with palloc(). Currently, we switch to slab allocation
224 : : * when we start merging. Merging only needs to keep a small, fixed
225 : : * number of tuples in memory at any time, so we can avoid the
226 : : * palloc/pfree overhead by recycling a fixed number of fixed-size slots
227 : : * to hold the tuples.
228 : : *
229 : : * For the slab, we use one large allocation, divided into SLAB_SLOT_SIZE
230 : : * slots. The allocation is sized to have one slot per tape, plus one
231 : : * additional slot. We need that many slots to hold all the tuples kept
232 : : * in the heap during merge, plus the one we have last returned from the
233 : : * sort, with tuplesort_gettuple.
234 : : *
235 : : * Initially, all the slots are kept in a linked list of free slots. When
236 : : * a tuple is read from a tape, it is put to the next available slot, if
237 : : * it fits. If the tuple is larger than SLAB_SLOT_SIZE, it is palloc'd
238 : : * instead.
239 : : *
240 : : * When we're done processing a tuple, we return the slot back to the free
241 : : * list, or pfree() if it was palloc'd. We know that a tuple was
242 : : * allocated from the slab, if its pointer value is between
243 : : * slabMemoryBegin and -End.
244 : : *
245 : : * When the slab allocator is used, the USEMEM/LACKMEM mechanism of
246 : : * tracking memory usage is not used.
247 : : */
248 : : bool slabAllocatorUsed;
249 : :
250 : : char *slabMemoryBegin; /* beginning of slab memory arena */
251 : : char *slabMemoryEnd; /* end of slab memory arena */
252 : : SlabSlot *slabFreeHead; /* head of free list */
253 : :
254 : : /* Memory used for input and output tape buffers. */
255 : : size_t tape_buffer_mem;
256 : :
257 : : /*
258 : : * When we return a tuple to the caller in tuplesort_gettuple_XXX, that
259 : : * came from a tape (that is, in TSS_SORTEDONTAPE or TSS_FINALMERGE
260 : : * modes), we remember the tuple in 'lastReturnedTuple', so that we can
261 : : * recycle the memory on next gettuple call.
262 : : */
263 : : void *lastReturnedTuple;
264 : :
265 : : /*
266 : : * While building initial runs, this is the current output run number.
267 : : * Afterwards, it is the number of initial runs we made.
268 : : */
269 : : int currentRun;
270 : :
271 : : /*
272 : : * Logical tapes, for merging.
273 : : *
274 : : * The initial runs are written in the output tapes. In each merge pass,
275 : : * the output tapes of the previous pass become the input tapes, and new
276 : : * output tapes are created as needed. When nInputTapes equals
277 : : * nInputRuns, there is only one merge pass left.
278 : : */
279 : : LogicalTape **inputTapes;
280 : : int nInputTapes;
281 : : int nInputRuns;
282 : :
283 : : LogicalTape **outputTapes;
284 : : int nOutputTapes;
285 : : int nOutputRuns;
286 : :
287 : : LogicalTape *destTape; /* current output tape */
288 : :
289 : : /*
290 : : * These variables are used after completion of sorting to keep track of
291 : : * the next tuple to return. (In the tape case, the tape's current read
292 : : * position is also critical state.)
293 : : */
294 : : LogicalTape *result_tape; /* actual tape of finished output */
295 : : int current; /* array index (only used if SORTEDINMEM) */
296 : : bool eof_reached; /* reached EOF (needed for cursors) */
297 : :
298 : : /* markpos_xxx holds marked position for mark and restore */
299 : : int64 markpos_block; /* tape block# (only used if SORTEDONTAPE) */
300 : : int markpos_offset; /* saved "current", or offset in tape block */
301 : : bool markpos_eof; /* saved "eof_reached" */
302 : :
303 : : /*
304 : : * These variables are used during parallel sorting.
305 : : *
306 : : * worker is our worker identifier. Follows the general convention that
307 : : * -1 value relates to a leader tuplesort, and values >= 0 worker
308 : : * tuplesorts. (-1 can also be a serial tuplesort.)
309 : : *
310 : : * shared is mutable shared memory state, which is used to coordinate
311 : : * parallel sorts.
312 : : *
313 : : * nParticipants is the number of worker Tuplesortstates known by the
314 : : * leader to have actually been launched, which implies that they must
315 : : * finish a run that the leader needs to merge. Typically includes a
316 : : * worker state held by the leader process itself. Set in the leader
317 : : * Tuplesortstate only.
318 : : */
319 : : int worker;
320 : : Sharedsort *shared;
321 : : int nParticipants;
322 : :
323 : : /*
324 : : * Additional state for managing "abbreviated key" sortsupport routines
325 : : * (which currently may be used by all cases except the hash index case).
326 : : * Tracks the intervals at which the optimization's effectiveness is
327 : : * tested.
328 : : */
329 : : int64 abbrevNext; /* Tuple # at which to next check
330 : : * applicability */
331 : :
332 : : /*
333 : : * Resource snapshot for time of sort start.
334 : : */
335 : : PGRUsage ru_start;
336 : : };
337 : :
338 : : /*
339 : : * Private mutable state of tuplesort-parallel-operation. This is allocated
340 : : * in shared memory.
341 : : */
342 : : struct Sharedsort
343 : : {
344 : : /*
345 : : * currentWorker generates ordinal identifier numbers for parallel sort
346 : : * workers. These start from 0, and are always gapless.
347 : : *
348 : : * Workers increment workersFinished to indicate having finished. If this
349 : : * is equal to state.nParticipants within the leader, leader is ready to
350 : : * merge worker runs.
351 : : */
352 : : pg_atomic_uint32 currentWorker;
353 : : pg_atomic_uint32 workersFinished;
354 : :
355 : : /* Temporary file space */
356 : : SharedFileSet fileset;
357 : :
358 : : /* Size of tapes flexible array */
359 : : int nTapes;
360 : :
361 : : /*
362 : : * Tapes array used by workers to report back information needed by the
363 : : * leader to concatenate all worker tapes into one for merging
364 : : */
365 : : TapeShare tapes[FLEXIBLE_ARRAY_MEMBER];
366 : : };
367 : :
368 : : /*
369 : : * Is the given tuple allocated from the slab memory arena?
370 : : */
371 : : #define IS_SLAB_SLOT(state, tuple) \
372 : : ((char *) (tuple) >= (state)->slabMemoryBegin && \
373 : : (char *) (tuple) < (state)->slabMemoryEnd)
374 : :
375 : : /*
376 : : * Return the given tuple to the slab memory free list, or free it
377 : : * if it was palloc'd.
378 : : */
379 : : #define RELEASE_SLAB_SLOT(state, tuple) \
380 : : do { \
381 : : SlabSlot *buf = (SlabSlot *) tuple; \
382 : : \
383 : : if (IS_SLAB_SLOT((state), buf)) \
384 : : { \
385 : : buf->nextfree = (state)->slabFreeHead; \
386 : : (state)->slabFreeHead = buf; \
387 : : } else \
388 : : pfree(buf); \
389 : : } while(0)
390 : :
391 : : #define REMOVEABBREV(state,stup,count) ((*(state)->base.removeabbrev) (state, stup, count))
392 : : #define COMPARETUP(state,a,b) ((*(state)->base.comparetup) (a, b, state))
393 : : #define WRITETUP(state,tape,stup) ((*(state)->base.writetup) (state, tape, stup))
394 : : #define READTUP(state,stup,tape,len) ((*(state)->base.readtup) (state, stup, tape, len))
395 : : #define FREESTATE(state) ((state)->base.freestate ? (*(state)->base.freestate) (state) : (void) 0)
396 : : #define LACKMEM(state) ((state)->availMem < 0 && !(state)->slabAllocatorUsed)
397 : : #define USEMEM(state,amt) ((state)->availMem -= (amt))
398 : : #define FREEMEM(state,amt) ((state)->availMem += (amt))
399 : : #define SERIAL(state) ((state)->shared == NULL)
400 : : #define WORKER(state) ((state)->shared && (state)->worker != -1)
401 : : #define LEADER(state) ((state)->shared && (state)->worker == -1)
402 : :
403 : : /*
404 : : * NOTES about on-tape representation of tuples:
405 : : *
406 : : * We require the first "unsigned int" of a stored tuple to be the total size
407 : : * on-tape of the tuple, including itself (so it is never zero; an all-zero
408 : : * unsigned int is used to delimit runs). The remainder of the stored tuple
409 : : * may or may not match the in-memory representation of the tuple ---
410 : : * any conversion needed is the job of the writetup and readtup routines.
411 : : *
412 : : * If state->sortopt contains TUPLESORT_RANDOMACCESS, then the stored
413 : : * representation of the tuple must be followed by another "unsigned int" that
414 : : * is a copy of the length --- so the total tape space used is actually
415 : : * sizeof(unsigned int) more than the stored length value. This allows
416 : : * read-backwards. When the random access flag was not specified, the
417 : : * write/read routines may omit the extra length word.
418 : : *
419 : : * writetup is expected to write both length words as well as the tuple
420 : : * data. When readtup is called, the tape is positioned just after the
421 : : * front length word; readtup must read the tuple data and advance past
422 : : * the back length word (if present).
423 : : *
424 : : * The write/read routines can make use of the tuple description data
425 : : * stored in the Tuplesortstate record, if needed. They are also expected
426 : : * to adjust state->availMem by the amount of memory space (not tape space!)
427 : : * released or consumed. There is no error return from either writetup
428 : : * or readtup; they should ereport() on failure.
429 : : *
430 : : *
431 : : * NOTES about memory consumption calculations:
432 : : *
433 : : * We count space allocated for tuples against the workMem limit, plus
434 : : * the space used by the variable-size memtuples array. Fixed-size space
435 : : * is not counted; it's small enough to not be interesting.
436 : : *
437 : : * Note that we count actual space used (as shown by GetMemoryChunkSpace)
438 : : * rather than the originally-requested size. This is important since
439 : : * palloc can add substantial overhead. It's not a complete answer since
440 : : * we won't count any wasted space in palloc allocation blocks, but it's
441 : : * a lot better than what we were doing before 7.3. As of 9.6, a
442 : : * separate memory context is used for caller passed tuples. Resetting
443 : : * it at certain key increments significantly ameliorates fragmentation.
444 : : * readtup routines use the slab allocator (they cannot use
445 : : * the reset context because it gets deleted at the point that merging
446 : : * begins).
447 : : */
448 : :
449 : :
450 : : static void tuplesort_begin_batch(Tuplesortstate *state);
451 : : static bool consider_abort_common(Tuplesortstate *state);
452 : : static void inittapes(Tuplesortstate *state, bool mergeruns);
453 : : static void inittapestate(Tuplesortstate *state, int maxTapes);
454 : : static void selectnewtape(Tuplesortstate *state);
455 : : static void init_slab_allocator(Tuplesortstate *state, int numSlots);
456 : : static void mergeruns(Tuplesortstate *state);
457 : : static void mergeonerun(Tuplesortstate *state);
458 : : static void beginmerge(Tuplesortstate *state);
459 : : static bool mergereadnext(Tuplesortstate *state, LogicalTape *srcTape, SortTuple *stup);
460 : : static void dumptuples(Tuplesortstate *state, bool alltuples);
461 : : static void make_bounded_heap(Tuplesortstate *state);
462 : : static void sort_bounded_heap(Tuplesortstate *state);
463 : : static void tuplesort_sort_memtuples(Tuplesortstate *state);
464 : : static void tuplesort_heap_insert(Tuplesortstate *state, SortTuple *tuple);
465 : : static void tuplesort_heap_replace_top(Tuplesortstate *state, SortTuple *tuple);
466 : : static void tuplesort_heap_delete_top(Tuplesortstate *state);
467 : : static void reversedirection(Tuplesortstate *state);
468 : : static unsigned int getlen(LogicalTape *tape, bool eofOK);
469 : : static void markrunend(LogicalTape *tape);
470 : : static int worker_get_identifier(Tuplesortstate *state);
471 : : static void worker_freeze_result_tape(Tuplesortstate *state);
472 : : static void worker_nomergeruns(Tuplesortstate *state);
473 : : static void leader_takeover_tapes(Tuplesortstate *state);
474 : : static void free_sort_tuple(Tuplesortstate *state, SortTuple *stup);
475 : : static void tuplesort_free(Tuplesortstate *state);
476 : : static void tuplesort_updatemax(Tuplesortstate *state);
477 : :
478 : :
479 : : /*
480 : : * Special versions of qsort just for SortTuple objects. qsort_tuple() sorts
481 : : * any variant of SortTuples, using the appropriate comparetup function.
482 : : * qsort_ssup() is specialized for the case where the comparetup function
483 : : * reduces to ApplySortComparator(), that is single-key MinimalTuple sorts
484 : : * and Datum sorts.
485 : : */
486 : :
487 : : #define ST_SORT qsort_tuple
488 : : #define ST_ELEMENT_TYPE SortTuple
489 : : #define ST_COMPARE_RUNTIME_POINTER
490 : : #define ST_COMPARE_ARG_TYPE Tuplesortstate
491 : : #define ST_CHECK_FOR_INTERRUPTS
492 : : #define ST_SCOPE static
493 : : #define ST_DECLARE
494 : : #define ST_DEFINE
495 : : #include "lib/sort_template.h"
496 : :
497 : : #define ST_SORT qsort_ssup
498 : : #define ST_ELEMENT_TYPE SortTuple
499 : : #define ST_COMPARE(a, b, ssup) \
500 : : ApplySortComparator((a)->datum1, (a)->isnull1, \
501 : : (b)->datum1, (b)->isnull1, (ssup))
502 : : #define ST_COMPARE_ARG_TYPE SortSupportData
503 : : #define ST_CHECK_FOR_INTERRUPTS
504 : : #define ST_SCOPE static
505 : : #define ST_DEFINE
506 : : #include "lib/sort_template.h"
507 : :
508 : : /* state for radix sort */
509 : : typedef struct RadixSortInfo
510 : : {
511 : : union
512 : : {
513 : : size_t count;
514 : : size_t offset;
515 : : };
516 : : size_t next_offset;
517 : : } RadixSortInfo;
518 : :
519 : : /*
520 : : * Threshold below which qsort_tuple() is generally faster than a radix sort.
521 : : */
522 : : #define QSORT_THRESHOLD 40
523 : :
524 : :
525 : : /*
526 : : * tuplesort_begin_xxx
527 : : *
528 : : * Initialize for a tuple sort operation.
529 : : *
530 : : * After calling tuplesort_begin, the caller should call tuplesort_putXXX
531 : : * zero or more times, then call tuplesort_performsort when all the tuples
532 : : * have been supplied. After performsort, retrieve the tuples in sorted
533 : : * order by calling tuplesort_getXXX until it returns false/NULL. (If random
534 : : * access was requested, rescan, markpos, and restorepos can also be called.)
535 : : * Call tuplesort_end to terminate the operation and release memory/disk space.
536 : : *
537 : : * Each variant of tuplesort_begin has a workMem parameter specifying the
538 : : * maximum number of kilobytes of RAM to use before spilling data to disk.
539 : : * (The normal value of this parameter is work_mem, but some callers use
540 : : * other values.) Each variant also has a sortopt which is a bitmask of
541 : : * sort options. See TUPLESORT_* definitions in tuplesort.h
542 : : */
543 : :
544 : : Tuplesortstate *
1652 drowley@postgresql.o 545 :CBC 170552 : tuplesort_begin_common(int workMem, SortCoordinate coordinate, int sortopt)
546 : : {
547 : : Tuplesortstate *state;
548 : : MemoryContext maincontext;
549 : : MemoryContext sortcontext;
550 : : MemoryContext oldcontext;
551 : :
552 : : /* See leader_takeover_tapes() remarks on random access support */
553 [ + + - + ]: 170552 : if (coordinate && (sortopt & TUPLESORT_RANDOMACCESS))
3174 rhaas@postgresql.org 554 [ # # ]:UBC 0 : elog(ERROR, "random access disallowed under parallel sort");
555 : :
556 : : /*
557 : : * Memory context surviving tuplesort_reset. This memory context holds
558 : : * data which is useful to keep while sorting multiple similar batches.
559 : : */
2380 tomas.vondra@postgre 560 :CBC 170552 : maincontext = AllocSetContextCreate(CurrentMemoryContext,
561 : : "TupleSort main",
562 : : ALLOCSET_DEFAULT_SIZES);
563 : :
564 : : /*
565 : : * Create a working memory context for one sort operation. The content of
566 : : * this context is deleted by tuplesort_reset.
567 : : */
568 : 170552 : sortcontext = AllocSetContextCreate(maincontext,
569 : : "TupleSort sort",
570 : : ALLOCSET_DEFAULT_SIZES);
571 : :
572 : : /*
573 : : * Additionally a working memory context for tuples is setup in
574 : : * tuplesort_begin_batch.
575 : : */
576 : :
577 : : /*
578 : : * Make the Tuplesortstate within the per-sortstate context. This way, we
579 : : * don't need a separate pfree() operation for it at shutdown.
580 : : */
581 : 170552 : oldcontext = MemoryContextSwitchTo(maincontext);
582 : :
306 michael@paquier.xyz 583 : 170552 : state = palloc0_object(Tuplesortstate);
584 : :
7679 tgl@sss.pgh.pa.us 585 [ - + ]: 170552 : if (trace_sort)
7679 tgl@sss.pgh.pa.us 586 :UBC 0 : pg_rusage_init(&state->ru_start);
587 : :
1538 akorotkov@postgresql 588 :CBC 170552 : state->base.sortopt = sortopt;
589 : 170552 : state->base.tuples = true;
590 : 170552 : state->abbrevNext = 10;
591 : :
592 : : /*
593 : : * workMem is forced to be at least 64KB, the current minimum valid value
594 : : * for the work_mem GUC. This is a defense against parallel sort callers
595 : : * that divide out memory among many workers in a way that leaves each
596 : : * with very little memory.
597 : : */
3174 rhaas@postgresql.org 598 : 170552 : state->allowedMem = Max(workMem, 64) * (int64) 1024;
1538 akorotkov@postgresql 599 : 170552 : state->base.sortcontext = sortcontext;
600 : 170552 : state->base.maincontext = maincontext;
601 : :
2380 tomas.vondra@postgre 602 : 170552 : state->memtupsize = INITIAL_MEMTUPSIZE;
603 : 170552 : state->memtuples = NULL;
604 : :
605 : : /*
606 : : * After all of the other non-parallel-related state, we setup all of the
607 : : * state needed for each batch.
608 : : */
609 : 170552 : tuplesort_begin_batch(state);
610 : :
611 : : /*
612 : : * Initialize parallel-related state based on coordination information
613 : : * from caller
614 : : */
3174 rhaas@postgresql.org 615 [ + + ]: 170552 : if (!coordinate)
616 : : {
617 : : /* Serial sort */
618 : 169995 : state->shared = NULL;
619 : 169995 : state->worker = -1;
620 : 169995 : state->nParticipants = -1;
621 : : }
622 [ + + ]: 557 : else if (coordinate->isWorker)
623 : : {
624 : : /* Parallel worker produces exactly one final run from all input */
625 : 379 : state->shared = coordinate->sharedsort;
626 : 379 : state->worker = worker_get_identifier(state);
627 : 379 : state->nParticipants = -1;
628 : : }
629 : : else
630 : : {
631 : : /* Parallel leader state only used for final merge */
632 : 178 : state->shared = coordinate->sharedsort;
633 : 178 : state->worker = -1;
634 : 178 : state->nParticipants = coordinate->nParticipants;
635 [ - + ]: 178 : Assert(state->nParticipants >= 1);
636 : : }
637 : :
7533 tgl@sss.pgh.pa.us 638 : 170552 : MemoryContextSwitchTo(oldcontext);
639 : :
9857 640 : 170552 : return state;
641 : : }
642 : :
643 : : /*
644 : : * tuplesort_begin_batch
645 : : *
646 : : * Setup, or reset, all state need for processing a new set of tuples with this
647 : : * sort state. Called both from tuplesort_begin_common (the first time sorting
648 : : * with this sort state) and tuplesort_reset (for subsequent usages).
649 : : */
650 : : static void
2380 tomas.vondra@postgre 651 : 172679 : tuplesort_begin_batch(Tuplesortstate *state)
652 : : {
653 : : MemoryContext oldcontext;
654 : :
1538 akorotkov@postgresql 655 : 172679 : oldcontext = MemoryContextSwitchTo(state->base.maincontext);
656 : :
657 : : /*
658 : : * Caller tuple (e.g. IndexTuple) memory context.
659 : : *
660 : : * A dedicated child context used exclusively for caller passed tuples
661 : : * eases memory management. Resetting at key points reduces
662 : : * fragmentation. Note that the memtuples array of SortTuples is allocated
663 : : * in the parent context, not this context, because there is no need to
664 : : * free memtuples early. For bounded sorts, tuples may be pfreed in any
665 : : * order, so we use a regular aset.c context so that it can make use of
666 : : * free'd memory. When the sort is not bounded, we make use of a bump.c
667 : : * context as this keeps allocations more compact with less wastage.
668 : : * Allocations are also slightly more CPU efficient.
669 : : */
917 drowley@postgresql.o 670 [ + + ]: 172679 : if (TupleSortUseBumpTupleCxt(state->base.sortopt))
671 : 171785 : state->base.tuplecontext = BumpContextCreate(state->base.sortcontext,
672 : : "Caller tuples",
673 : : ALLOCSET_DEFAULT_SIZES);
674 : : else
1538 akorotkov@postgresql 675 : 894 : state->base.tuplecontext = AllocSetContextCreate(state->base.sortcontext,
676 : : "Caller tuples",
677 : : ALLOCSET_DEFAULT_SIZES);
678 : :
679 : :
2380 tomas.vondra@postgre 680 : 172679 : state->status = TSS_INITIAL;
681 : 172679 : state->bounded = false;
682 : 172679 : state->boundUsed = false;
683 : :
684 : 172679 : state->availMem = state->allowedMem;
685 : :
686 : 172679 : state->tapeset = NULL;
687 : :
688 : 172679 : state->memtupcount = 0;
689 : :
690 : 172679 : state->growmemtuples = true;
691 : 172679 : state->slabAllocatorUsed = false;
692 [ + + + + ]: 172679 : if (state->memtuples != NULL && state->memtupsize != INITIAL_MEMTUPSIZE)
693 : : {
694 : 96 : pfree(state->memtuples);
695 : 96 : state->memtuples = NULL;
696 : 96 : state->memtupsize = INITIAL_MEMTUPSIZE;
697 : : }
698 [ + + ]: 172679 : if (state->memtuples == NULL)
699 : : {
56 michael@paquier.xyz 700 :GNC 170648 : state->memtuples = palloc_array(SortTuple, state->memtupsize);
2380 tomas.vondra@postgre 701 :CBC 170648 : USEMEM(state, GetMemoryChunkSpace(state->memtuples));
702 : : }
703 : :
704 : : /* workMem must be large enough for the minimal memtuples array */
705 [ - + - - ]: 172679 : if (LACKMEM(state))
2380 tomas.vondra@postgre 706 [ # # ]:UBC 0 : elog(ERROR, "insufficient memory allowed for sort");
707 : :
2380 tomas.vondra@postgre 708 :CBC 172679 : state->currentRun = 0;
709 : :
710 : : /*
711 : : * Tape variables (inputTapes, outputTapes, etc.) will be initialized by
712 : : * inittapes(), if needed.
713 : : */
714 : :
1820 heikki.linnakangas@i 715 : 172679 : state->result_tape = NULL; /* flag that result tape has not been formed */
716 : :
2380 tomas.vondra@postgre 717 : 172679 : MemoryContextSwitchTo(oldcontext);
718 : 172679 : }
719 : :
720 : : /*
721 : : * tuplesort_set_bound
722 : : *
723 : : * Advise tuplesort that at most the first N result tuples are required.
724 : : *
725 : : * Must be called before inserting any tuples. (Actually, we could allow it
726 : : * as long as the sort hasn't spilled to disk, but there seems no need for
727 : : * delayed calls at the moment.)
728 : : *
729 : : * This is a hint only. The tuplesort may still return more tuples than
730 : : * requested. Parallel leader tuplesorts will always ignore the hint.
731 : : */
732 : : void
7101 tgl@sss.pgh.pa.us 733 : 805 : tuplesort_set_bound(Tuplesortstate *state, int64 bound)
734 : : {
735 : : /* Assert we're called before loading any tuples */
2587 alvherre@alvh.no-ip. 736 [ + - - + ]: 805 : Assert(state->status == TSS_INITIAL && state->memtupcount == 0);
737 : : /* Assert we allow bounded sorts */
1538 akorotkov@postgresql 738 [ - + ]: 805 : Assert(state->base.sortopt & TUPLESORT_ALLOWBOUNDED);
739 : : /* Can't set the bound twice, either */
7101 tgl@sss.pgh.pa.us 740 [ - + ]: 805 : Assert(!state->bounded);
741 : : /* Also, this shouldn't be called in a parallel worker */
3174 rhaas@postgresql.org 742 [ - + - - ]: 805 : Assert(!WORKER(state));
743 : :
744 : : /* Parallel leader allows but ignores hint */
2586 tgl@sss.pgh.pa.us 745 [ - + - - ]: 805 : if (LEADER(state))
2586 tgl@sss.pgh.pa.us 746 :UBC 0 : return;
747 : :
748 : : #ifdef DEBUG_BOUNDED_SORT
749 : : /* Honor GUC setting that disables the feature (for easy testing) */
750 : : if (!optimize_bounded_sort)
751 : : return;
752 : : #endif
753 : :
754 : : /* We want to be able to compute bound * 2, so limit the setting */
6906 bruce@momjian.us 755 [ - + ]:CBC 805 : if (bound > (int64) (INT_MAX / 2))
7101 tgl@sss.pgh.pa.us 756 :UBC 0 : return;
757 : :
7101 tgl@sss.pgh.pa.us 758 :CBC 805 : state->bounded = true;
759 : 805 : state->bound = (int) bound;
760 : :
761 : : /*
762 : : * Bounded sorts are not an effective target for abbreviated key
763 : : * optimization. Disable by setting state to be consistent with no
764 : : * abbreviation support.
765 : : */
1538 akorotkov@postgresql 766 : 805 : state->base.sortKeys->abbrev_converter = NULL;
767 [ + + ]: 805 : if (state->base.sortKeys->abbrev_full_comparator)
768 : 10 : state->base.sortKeys->comparator = state->base.sortKeys->abbrev_full_comparator;
769 : :
770 : : /* Not strictly necessary, but be tidy */
771 : 805 : state->base.sortKeys->abbrev_abort = NULL;
772 : 805 : state->base.sortKeys->abbrev_full_comparator = NULL;
773 : : }
774 : :
775 : : /*
776 : : * tuplesort_used_bound
777 : : *
778 : : * Allow callers to find out if the sort state was able to use a bound.
779 : : */
780 : : bool
2380 tomas.vondra@postgre 781 : 307 : tuplesort_used_bound(Tuplesortstate *state)
782 : : {
783 : 307 : return state->boundUsed;
784 : : }
785 : :
786 : : /*
787 : : * tuplesort_free
788 : : *
789 : : * Internal routine for freeing resources of tuplesort.
790 : : */
791 : : static void
792 : 172488 : tuplesort_free(Tuplesortstate *state)
793 : : {
794 : : /* context swap probably not needed, but let's be safe */
1538 akorotkov@postgresql 795 : 172488 : MemoryContext oldcontext = MemoryContextSwitchTo(state->base.sortcontext);
796 : : int64 spaceUsed;
797 : :
9857 tgl@sss.pgh.pa.us 798 [ + + ]: 172488 : if (state->tapeset)
7664 799 : 599 : spaceUsed = LogicalTapeSetBlocks(state->tapeset);
800 : : else
801 : 171889 : spaceUsed = (state->allowedMem - state->availMem + 1023) / 1024;
802 : :
803 : : /*
804 : : * Delete temporary "tape" files, if any.
805 : : *
806 : : * We don't bother to destroy the individual tapes here. They will go away
807 : : * with the sortcontext. (In TSS_FINALMERGE state, we have closed
808 : : * finished tapes already.)
809 : : */
7533 810 [ + + ]: 172488 : if (state->tapeset)
811 : 599 : LogicalTapeSetClose(state->tapeset);
812 : :
7679 813 [ - + ]: 172488 : if (trace_sort)
814 : : {
7664 tgl@sss.pgh.pa.us 815 [ # # ]:UBC 0 : if (state->tapeset)
562 peter@eisentraut.org 816 [ # # # # ]: 0 : elog(LOG, "%s of worker %d ended, %" PRId64 " disk blocks used: %s",
817 : : SERIAL(state) ? "external sort" : "parallel external sort",
818 : : state->worker, spaceUsed, pg_rusage_show(&state->ru_start));
819 : : else
820 [ # # # # ]: 0 : elog(LOG, "%s of worker %d ended, %" PRId64 " KB used: %s",
821 : : SERIAL(state) ? "internal sort" : "unperformed parallel sort",
822 : : state->worker, spaceUsed, pg_rusage_show(&state->ru_start));
823 : : }
824 : :
825 : : TRACE_POSTGRESQL_SORT_DONE(state->tapeset != NULL, spaceUsed);
826 : :
1538 akorotkov@postgresql 827 [ + + ]:CBC 172488 : FREESTATE(state);
7533 tgl@sss.pgh.pa.us 828 : 172488 : MemoryContextSwitchTo(oldcontext);
829 : :
830 : : /*
831 : : * Free the per-sort memory context, thereby releasing all working memory.
832 : : */
1538 akorotkov@postgresql 833 : 172488 : MemoryContextReset(state->base.sortcontext);
2380 tomas.vondra@postgre 834 : 172488 : }
835 : :
836 : : /*
837 : : * tuplesort_end
838 : : *
839 : : * Release resources and clean up.
840 : : *
841 : : * NOTE: after calling this, any pointers returned by tuplesort_getXXX are
842 : : * pointing to garbage. Be careful not to attempt to use or free such
843 : : * pointers afterwards!
844 : : */
845 : : void
846 : 170361 : tuplesort_end(Tuplesortstate *state)
847 : : {
848 : 170361 : tuplesort_free(state);
849 : :
850 : : /*
851 : : * Free the main memory context, including the Tuplesortstate struct
852 : : * itself.
853 : : */
1538 akorotkov@postgresql 854 : 170361 : MemoryContextDelete(state->base.maincontext);
2380 tomas.vondra@postgre 855 : 170361 : }
856 : :
857 : : /*
858 : : * tuplesort_updatemax
859 : : *
860 : : * Update maximum resource usage statistics.
861 : : */
862 : : static void
863 : 2391 : tuplesort_updatemax(Tuplesortstate *state)
864 : : {
865 : : int64 spaceUsed;
866 : : bool isSpaceDisk;
867 : :
868 : : /*
869 : : * Note: it might seem we should provide both memory and disk usage for a
870 : : * disk-based sort. However, the current code doesn't track memory space
871 : : * accurately once we have begun to return tuples to the caller (since we
872 : : * don't account for pfree's the caller is expected to do), so we cannot
873 : : * rely on availMem in a disk sort. This does not seem worth the overhead
874 : : * to fix. Is it worth creating an API for the memory context code to
875 : : * tell us how much is actually used in sortcontext?
876 : : */
877 [ + + ]: 2391 : if (state->tapeset)
878 : : {
879 : 4 : isSpaceDisk = true;
880 : 4 : spaceUsed = LogicalTapeSetBlocks(state->tapeset) * BLCKSZ;
881 : : }
882 : : else
883 : : {
884 : 2387 : isSpaceDisk = false;
885 : 2387 : spaceUsed = state->allowedMem - state->availMem;
886 : : }
887 : :
888 : : /*
889 : : * Sort evicts data to the disk when it wasn't able to fit that data into
890 : : * main memory. This is why we assume space used on the disk to be more
891 : : * important for tracking resource usage than space used in memory. Note
892 : : * that the amount of space occupied by some tupleset on the disk might be
893 : : * less than amount of space occupied by the same tupleset in memory due
894 : : * to more compact representation.
895 : : */
896 [ + + - + ]: 2391 : if ((isSpaceDisk && !state->isMaxSpaceDisk) ||
897 [ + - + + ]: 2387 : (isSpaceDisk == state->isMaxSpaceDisk && spaceUsed > state->maxSpace))
898 : : {
899 : 357 : state->maxSpace = spaceUsed;
900 : 357 : state->isMaxSpaceDisk = isSpaceDisk;
901 : 357 : state->maxSpaceStatus = state->status;
902 : : }
903 : 2391 : }
904 : :
905 : : /*
906 : : * tuplesort_reset
907 : : *
908 : : * Reset the tuplesort. Reset all the data in the tuplesort, but leave the
909 : : * meta-information in. After tuplesort_reset, tuplesort is ready to start
910 : : * a new sort. This allows avoiding recreation of tuple sort states (and
911 : : * save resources) when sorting multiple small batches.
912 : : */
913 : : void
914 : 2127 : tuplesort_reset(Tuplesortstate *state)
915 : : {
916 : 2127 : tuplesort_updatemax(state);
917 : 2127 : tuplesort_free(state);
918 : :
919 : : /*
920 : : * After we've freed up per-batch memory, re-setup all of the state common
921 : : * to both the first batch and any subsequent batch.
922 : : */
923 : 2127 : tuplesort_begin_batch(state);
924 : :
925 : 2127 : state->lastReturnedTuple = NULL;
926 : 2127 : state->slabMemoryBegin = NULL;
927 : 2127 : state->slabMemoryEnd = NULL;
928 : 2127 : state->slabFreeHead = NULL;
7533 tgl@sss.pgh.pa.us 929 : 2127 : }
930 : :
931 : : /*
932 : : * Grow the memtuples[] array, if possible within our memory constraint. We
933 : : * must not exceed INT_MAX tuples in memory or the caller-provided memory
934 : : * limit. Return true if we were able to enlarge the array, false if not.
935 : : *
936 : : * Normally, at each increment we double the size of the array. When doing
937 : : * that would exceed a limit, we attempt one last, smaller increase (and then
938 : : * clear the growmemtuples flag so we don't try any more). That allows us to
939 : : * use memory as fully as permitted; sticking to the pure doubling rule could
940 : : * result in almost half going unused. Because availMem moves around with
941 : : * tuple addition/removal, we need some rule to prevent making repeated small
942 : : * increases in memtupsize, which would just be useless thrashing. The
943 : : * growmemtuples flag accomplishes that and also prevents useless
944 : : * recalculations in this function.
945 : : */
946 : : static bool
947 : 5315 : grow_memtuples(Tuplesortstate *state)
948 : : {
949 : : int newmemtupsize;
5016 950 : 5315 : int memtupsize = state->memtupsize;
4848 noah@leadboat.com 951 : 5315 : int64 memNowUsed = state->allowedMem - state->availMem;
952 : :
953 : : /* Forget it if we've already maxed out memtuples, per comment above */
5016 tgl@sss.pgh.pa.us 954 [ + + ]: 5315 : if (!state->growmemtuples)
955 : 90 : return false;
956 : :
957 : : /* Select new value of memtupsize */
958 [ + + ]: 5225 : if (memNowUsed <= state->availMem)
959 : : {
960 : : /*
961 : : * We've used no more than half of allowedMem; double our usage,
962 : : * clamping at INT_MAX tuples.
963 : : */
4855 noah@leadboat.com 964 [ + - ]: 5133 : if (memtupsize < INT_MAX / 2)
965 : 5133 : newmemtupsize = memtupsize * 2;
966 : : else
967 : : {
4855 noah@leadboat.com 968 :UBC 0 : newmemtupsize = INT_MAX;
969 : 0 : state->growmemtuples = false;
970 : : }
971 : : }
972 : : else
973 : : {
974 : : /*
975 : : * This will be the last increment of memtupsize. Abandon doubling
976 : : * strategy and instead increase as much as we safely can.
977 : : *
978 : : * To stay within allowedMem, we can't increase memtupsize by more
979 : : * than availMem / sizeof(SortTuple) elements. In practice, we want
980 : : * to increase it by considerably less, because we need to leave some
981 : : * space for the tuples to which the new array slots will refer. We
982 : : * assume the new tuples will be about the same size as the tuples
983 : : * we've already seen, and thus we can extrapolate from the space
984 : : * consumption so far to estimate an appropriate new size for the
985 : : * memtuples array. The optimal value might be higher or lower than
986 : : * this estimate, but it's hard to know that in advance. We again
987 : : * clamp at INT_MAX tuples.
988 : : *
989 : : * This calculation is safe against enlarging the array so much that
990 : : * LACKMEM becomes true, because the memory currently used includes
991 : : * the present array; thus, there would be enough allowedMem for the
992 : : * new array elements even if no other memory were currently used.
993 : : *
994 : : * We do the arithmetic in float8, because otherwise the product of
995 : : * memtupsize and allowedMem could overflow. Any inaccuracy in the
996 : : * result should be insignificant; but even if we computed a
997 : : * completely insane result, the checks below will prevent anything
998 : : * really bad from happening.
999 : : */
1000 : : double grow_ratio;
1001 : :
5016 tgl@sss.pgh.pa.us 1002 :CBC 92 : grow_ratio = (double) state->allowedMem / (double) memNowUsed;
4855 noah@leadboat.com 1003 [ + - ]: 92 : if (memtupsize * grow_ratio < INT_MAX)
1004 : 92 : newmemtupsize = (int) (memtupsize * grow_ratio);
1005 : : else
4855 noah@leadboat.com 1006 :UBC 0 : newmemtupsize = INT_MAX;
1007 : :
1008 : : /* We won't make any further enlargement attempts */
5016 tgl@sss.pgh.pa.us 1009 :CBC 92 : state->growmemtuples = false;
1010 : : }
1011 : :
1012 : : /* Must enlarge array by at least one element, else report failure */
1013 [ - + ]: 5225 : if (newmemtupsize <= memtupsize)
5016 tgl@sss.pgh.pa.us 1014 :UBC 0 : goto noalloc;
1015 : :
1016 : : /*
1017 : : * On a 32-bit machine, allowedMem could exceed MaxAllocHugeSize. Clamp
1018 : : * to ensure our request won't be rejected. Note that we can easily
1019 : : * exhaust address space before facing this outcome. (This is presently
1020 : : * impossible due to guc.c's MAX_KILOBYTES limitation on work_mem, but
1021 : : * don't rely on that at this distance.)
1022 : : */
4855 noah@leadboat.com 1023 [ - + ]:CBC 5225 : if ((Size) newmemtupsize >= MaxAllocHugeSize / sizeof(SortTuple))
1024 : : {
4855 noah@leadboat.com 1025 :UBC 0 : newmemtupsize = (int) (MaxAllocHugeSize / sizeof(SortTuple));
5016 tgl@sss.pgh.pa.us 1026 : 0 : state->growmemtuples = false; /* can't grow any more */
1027 : : }
1028 : :
1029 : : /*
1030 : : * We need to be sure that we do not cause LACKMEM to become true, else
1031 : : * the space management algorithm will go nuts. The code above should
1032 : : * never generate a dangerous request, but to be safe, check explicitly
1033 : : * that the array growth fits within availMem. (We could still cause
1034 : : * LACKMEM if the memory chunk overhead associated with the memtuples
1035 : : * array were to increase. That shouldn't happen because we chose the
1036 : : * initial array size large enough to ensure that palloc will be treating
1037 : : * both old and new arrays as separate chunks. But we'll check LACKMEM
1038 : : * explicitly below just in case.)
1039 : : */
4848 noah@leadboat.com 1040 [ - + ]:CBC 5225 : if (state->availMem < (int64) ((newmemtupsize - memtupsize) * sizeof(SortTuple)))
5016 tgl@sss.pgh.pa.us 1041 :UBC 0 : goto noalloc;
1042 : :
1043 : : /* OK, do it */
7533 tgl@sss.pgh.pa.us 1044 :CBC 5225 : FREEMEM(state, GetMemoryChunkSpace(state->memtuples));
5016 1045 : 5225 : state->memtupsize = newmemtupsize;
7533 1046 : 5225 : state->memtuples = (SortTuple *)
4855 noah@leadboat.com 1047 : 5225 : repalloc_huge(state->memtuples,
1048 : 5225 : state->memtupsize * sizeof(SortTuple));
7533 tgl@sss.pgh.pa.us 1049 : 5225 : USEMEM(state, GetMemoryChunkSpace(state->memtuples));
1050 [ - + - - ]: 5225 : if (LACKMEM(state))
4087 tgl@sss.pgh.pa.us 1051 [ # # ]:UBC 0 : elog(ERROR, "unexpected out-of-memory situation in tuplesort");
7533 tgl@sss.pgh.pa.us 1052 :CBC 5225 : return true;
1053 : :
5016 tgl@sss.pgh.pa.us 1054 :UBC 0 : noalloc:
1055 : : /* If for any reason we didn't realloc, shut off future attempts */
1056 : 0 : state->growmemtuples = false;
1057 : 0 : return false;
1058 : : }
1059 : :
1060 : : /*
1061 : : * Shared code for tuple and datum cases.
1062 : : */
1063 : : void
917 drowley@postgresql.o 1064 :CBC 19301191 : tuplesort_puttuple_common(Tuplesortstate *state, SortTuple *tuple,
1065 : : bool useAbbrev, Size tuplen)
1066 : : {
1538 akorotkov@postgresql 1067 : 19301191 : MemoryContext oldcontext = MemoryContextSwitchTo(state->base.sortcontext);
1068 : :
3174 rhaas@postgresql.org 1069 [ + + - + ]: 19301191 : Assert(!LEADER(state));
1070 : :
1071 : : /* account for the memory used for this tuple */
917 drowley@postgresql.o 1072 : 19301191 : USEMEM(state, tuplen);
1073 : 19301191 : state->tupleMem += tuplen;
1074 : :
1538 akorotkov@postgresql 1075 [ + + ]: 19301191 : if (!useAbbrev)
1076 : : {
1077 : : /*
1078 : : * Leave ordinary Datum representation, or NULL value. If there is a
1079 : : * converter it won't expect NULL values, and cost model is not
1080 : : * required to account for NULL, so in that case we avoid calling
1081 : : * converter and just set datum1 to zeroed representation (to be
1082 : : * consistent, and to support cheap inequality tests for NULL
1083 : : * abbreviated keys).
1084 : : */
1085 : : }
1086 [ + + ]: 2901290 : else if (!consider_abort_common(state))
1087 : : {
1088 : : /* Store abbreviated key representation */
1089 : 2901226 : tuple->datum1 = state->base.sortKeys->abbrev_converter(tuple->datum1,
1090 : : state->base.sortKeys);
1091 : : }
1092 : : else
1093 : : {
1094 : : /*
1095 : : * Set state to be consistent with never trying abbreviation.
1096 : : *
1097 : : * Alter datum1 representation in already-copied tuples, so as to
1098 : : * ensure a consistent representation (current tuple was just
1099 : : * handled). It does not matter if some dumped tuples are already
1100 : : * sorted on tape, since serialized tuples lack abbreviated keys
1101 : : * (TSS_BUILDRUNS state prevents control reaching here in any case).
1102 : : */
1103 : 64 : REMOVEABBREV(state, state->memtuples, state->memtupcount);
1104 : : }
1105 : :
9857 tgl@sss.pgh.pa.us 1106 [ + + + - ]: 19301191 : switch (state->status)
1107 : : {
9118 bruce@momjian.us 1108 : 16398132 : case TSS_INITIAL:
1109 : :
1110 : : /*
1111 : : * Save the tuple into the unsorted array. First, grow the array
1112 : : * as needed. Note that we try to grow the array when there is
1113 : : * still one free slot remaining --- if we fail, there'll still be
1114 : : * room to store the incoming tuple, and then we'll switch to
1115 : : * tape-based operation.
1116 : : */
7533 tgl@sss.pgh.pa.us 1117 [ + + ]: 16398132 : if (state->memtupcount >= state->memtupsize - 1)
1118 : : {
1119 : 5315 : (void) grow_memtuples(state);
1120 [ - + ]: 5315 : Assert(state->memtupcount < state->memtupsize);
1121 : : }
1122 : 16398132 : state->memtuples[state->memtupcount++] = *tuple;
1123 : :
1124 : : /*
1125 : : * Check if it's time to switch over to a bounded heapsort. We do
1126 : : * so if the input tuple count exceeds twice the desired tuple
1127 : : * count (this is a heuristic for where heapsort becomes cheaper
1128 : : * than a quicksort), or if we've just filled workMem and have
1129 : : * enough tuples to meet the bound.
1130 : : *
1131 : : * Note that once we enter TSS_BOUNDED state we will always try to
1132 : : * complete the sort that way. In the worst case, if later input
1133 : : * tuples are larger than earlier ones, this might cause us to
1134 : : * exceed workMem significantly.
1135 : : */
7101 1136 [ + + ]: 16398132 : if (state->bounded &&
1137 [ + + ]: 48379 : (state->memtupcount > state->bound * 2 ||
1138 [ + + - + : 48115 : (state->memtupcount > state->bound && LACKMEM(state))))
- - ]
1139 : : {
1140 [ - + ]: 264 : if (trace_sort)
7101 tgl@sss.pgh.pa.us 1141 [ # # ]:UBC 0 : elog(LOG, "switching to bounded heapsort at %d tuples: %s",
1142 : : state->memtupcount,
1143 : : pg_rusage_show(&state->ru_start));
7101 tgl@sss.pgh.pa.us 1144 :CBC 264 : make_bounded_heap(state);
1538 akorotkov@postgresql 1145 : 264 : MemoryContextSwitchTo(oldcontext);
7101 tgl@sss.pgh.pa.us 1146 : 264 : return;
1147 : : }
1148 : :
1149 : : /*
1150 : : * Done if we still fit in available memory and have array slots.
1151 : : */
7533 1152 [ + + - + : 16397868 : if (state->memtupcount < state->memtupsize && !LACKMEM(state))
- - ]
1153 : : {
1538 akorotkov@postgresql 1154 : 16397778 : MemoryContextSwitchTo(oldcontext);
9857 tgl@sss.pgh.pa.us 1155 : 16397778 : return;
1156 : : }
1157 : :
1158 : : /*
1159 : : * Nope; time to switch to tape-based operation.
1160 : : */
3174 rhaas@postgresql.org 1161 : 90 : inittapes(state, true);
1162 : :
1163 : : /*
1164 : : * Dump all tuples.
1165 : : */
9857 tgl@sss.pgh.pa.us 1166 : 90 : dumptuples(state, false);
1167 : 90 : break;
1168 : :
7101 1169 : 2132056 : case TSS_BOUNDED:
1170 : :
1171 : : /*
1172 : : * We don't want to grow the array here, so check whether the new
1173 : : * tuple can be discarded before putting it in. This should be a
1174 : : * good speed optimization, too, since when there are many more
1175 : : * input tuples than the bound, most input tuples can be discarded
1176 : : * with just this one comparison. Note that because we currently
1177 : : * have the sort direction reversed, we must check for <= not >=.
1178 : : */
1179 [ + + ]: 2132056 : if (COMPARETUP(state, tuple, &state->memtuples[0]) <= 0)
1180 : : {
1181 : : /* new tuple <= top of the heap, so we can discard it */
1182 : 1797356 : free_sort_tuple(state, tuple);
5319 rhaas@postgresql.org 1183 [ + + ]: 1797356 : CHECK_FOR_INTERRUPTS();
1184 : : }
1185 : : else
1186 : : {
1187 : : /* discard top of heap, replacing it with the new tuple */
7101 tgl@sss.pgh.pa.us 1188 : 334700 : free_sort_tuple(state, &state->memtuples[0]);
3300 rhaas@postgresql.org 1189 : 334700 : tuplesort_heap_replace_top(state, tuple);
1190 : : }
7101 tgl@sss.pgh.pa.us 1191 : 2132056 : break;
1192 : :
9857 1193 : 771003 : case TSS_BUILDRUNS:
1194 : :
1195 : : /*
1196 : : * Save the tuple into the unsorted array (there must be space)
1197 : : */
3300 rhaas@postgresql.org 1198 : 771003 : state->memtuples[state->memtupcount++] = *tuple;
1199 : :
1200 : : /*
1201 : : * If we are over the memory limit, dump all tuples.
1202 : : */
9857 tgl@sss.pgh.pa.us 1203 : 771003 : dumptuples(state, false);
1204 : 771003 : break;
1205 : :
9857 tgl@sss.pgh.pa.us 1206 :UBC 0 : default:
8480 1207 [ # # ]: 0 : elog(ERROR, "invalid tuplesort state");
1208 : : break;
1209 : : }
1538 akorotkov@postgresql 1210 :CBC 2903149 : MemoryContextSwitchTo(oldcontext);
1211 : : }
1212 : :
1213 : : static bool
4284 rhaas@postgresql.org 1214 : 2901290 : consider_abort_common(Tuplesortstate *state)
1215 : : {
1538 akorotkov@postgresql 1216 [ - + ]: 2901290 : Assert(state->base.sortKeys[0].abbrev_converter != NULL);
1217 [ - + ]: 2901290 : Assert(state->base.sortKeys[0].abbrev_abort != NULL);
1218 [ - + ]: 2901290 : Assert(state->base.sortKeys[0].abbrev_full_comparator != NULL);
1219 : :
1220 : : /*
1221 : : * Check effectiveness of abbreviation optimization. Consider aborting
1222 : : * when still within memory limit.
1223 : : */
4284 rhaas@postgresql.org 1224 [ + + ]: 2901290 : if (state->status == TSS_INITIAL &&
1225 [ + + ]: 2574675 : state->memtupcount >= state->abbrevNext)
1226 : : {
1227 : 3244 : state->abbrevNext *= 2;
1228 : :
1229 : : /*
1230 : : * Check opclass-supplied abbreviation abort routine. It may indicate
1231 : : * that abbreviation should not proceed.
1232 : : */
1538 akorotkov@postgresql 1233 [ + + ]: 3244 : if (!state->base.sortKeys->abbrev_abort(state->memtupcount,
1234 : : state->base.sortKeys))
4284 rhaas@postgresql.org 1235 : 3180 : return false;
1236 : :
1237 : : /*
1238 : : * Finally, restore authoritative comparator, and indicate that
1239 : : * abbreviation is not in play by setting abbrev_converter to NULL
1240 : : */
1538 akorotkov@postgresql 1241 : 64 : state->base.sortKeys[0].comparator = state->base.sortKeys[0].abbrev_full_comparator;
1242 : 64 : state->base.sortKeys[0].abbrev_converter = NULL;
1243 : : /* Not strictly necessary, but be tidy */
1244 : 64 : state->base.sortKeys[0].abbrev_abort = NULL;
1245 : 64 : state->base.sortKeys[0].abbrev_full_comparator = NULL;
1246 : :
1247 : : /* Give up - expect original pass-by-value representation */
4284 rhaas@postgresql.org 1248 : 64 : return true;
1249 : : }
1250 : :
1251 : 2898046 : return false;
1252 : : }
1253 : :
1254 : : /*
1255 : : * All tuples have been provided; finish the sort.
1256 : : */
1257 : : void
9857 tgl@sss.pgh.pa.us 1258 : 147119 : tuplesort_performsort(Tuplesortstate *state)
1259 : : {
1538 akorotkov@postgresql 1260 : 147119 : MemoryContext oldcontext = MemoryContextSwitchTo(state->base.sortcontext);
1261 : :
7679 tgl@sss.pgh.pa.us 1262 [ - + ]: 147119 : if (trace_sort)
2902 pg@bowt.ie 1263 [ # # ]:UBC 0 : elog(LOG, "performsort of worker %d starting: %s",
1264 : : state->worker, pg_rusage_show(&state->ru_start));
1265 : :
9857 tgl@sss.pgh.pa.us 1266 [ + + + - ]:CBC 147119 : switch (state->status)
1267 : : {
9118 bruce@momjian.us 1268 : 146765 : case TSS_INITIAL:
1269 : :
1270 : : /*
1271 : : * We were able to accumulate all the tuples within the allowed
1272 : : * amount of memory, or leader to take over worker tapes
1273 : : */
3174 rhaas@postgresql.org 1274 [ + + ]: 146765 : if (SERIAL(state))
1275 : : {
1276 : : /* Sort in memory and we're done */
1277 : 146256 : tuplesort_sort_memtuples(state);
1278 : 146195 : state->status = TSS_SORTEDINMEM;
1279 : : }
1280 [ + - + + ]: 509 : else if (WORKER(state))
1281 : : {
1282 : : /*
1283 : : * Parallel workers must still dump out tuples to tape. No
1284 : : * merge is required to produce single output run, though.
1285 : : */
1286 : 379 : inittapes(state, false);
1287 : 379 : dumptuples(state, true);
1288 : 379 : worker_nomergeruns(state);
1289 : 379 : state->status = TSS_SORTEDONTAPE;
1290 : : }
1291 : : else
1292 : : {
1293 : : /*
1294 : : * Leader will take over worker tapes and merge worker runs.
1295 : : * Note that mergeruns sets the correct state->status.
1296 : : */
1297 : 130 : leader_takeover_tapes(state);
1298 : 130 : mergeruns(state);
1299 : : }
9857 tgl@sss.pgh.pa.us 1300 : 146704 : state->current = 0;
1301 : 146704 : state->eof_reached = false;
3174 rhaas@postgresql.org 1302 : 146704 : state->markpos_block = 0L;
9857 tgl@sss.pgh.pa.us 1303 : 146704 : state->markpos_offset = 0;
1304 : 146704 : state->markpos_eof = false;
1305 : 146704 : break;
1306 : :
7101 1307 : 264 : case TSS_BOUNDED:
1308 : :
1309 : : /*
1310 : : * We were able to accumulate all the tuples required for output
1311 : : * in memory, using a heap to eliminate excess tuples. Now we
1312 : : * have to transform the heap to a properly-sorted array. Note
1313 : : * that sort_bounded_heap sets the correct state->status.
1314 : : */
6981 1315 : 264 : sort_bounded_heap(state);
7101 1316 : 264 : state->current = 0;
1317 : 264 : state->eof_reached = false;
1318 : 264 : state->markpos_offset = 0;
1319 : 264 : state->markpos_eof = false;
1320 : 264 : break;
1321 : :
9857 1322 : 90 : case TSS_BUILDRUNS:
1323 : :
1324 : : /*
1325 : : * Finish tape-based sort. First, flush all tuples remaining in
1326 : : * memory out to tape; then merge until we have a single remaining
1327 : : * run (or, if !randomAccess and !WORKER(), one run per tape).
1328 : : * Note that mergeruns sets the correct state->status.
1329 : : */
1330 : 90 : dumptuples(state, true);
1331 : 90 : mergeruns(state);
1332 : 90 : state->eof_reached = false;
1333 : 90 : state->markpos_block = 0L;
1334 : 90 : state->markpos_offset = 0;
1335 : 90 : state->markpos_eof = false;
1336 : 90 : break;
1337 : :
9857 tgl@sss.pgh.pa.us 1338 :UBC 0 : default:
8480 1339 [ # # ]: 0 : elog(ERROR, "invalid tuplesort state");
1340 : : break;
1341 : : }
1342 : :
7679 tgl@sss.pgh.pa.us 1343 [ - + ]:CBC 147058 : if (trace_sort)
1344 : : {
7523 tgl@sss.pgh.pa.us 1345 [ # # ]:UBC 0 : if (state->status == TSS_FINALMERGE)
2902 pg@bowt.ie 1346 [ # # ]: 0 : elog(LOG, "performsort of worker %d done (except %d-way final merge): %s",
1347 : : state->worker, state->nInputTapes,
1348 : : pg_rusage_show(&state->ru_start));
1349 : : else
1350 [ # # ]: 0 : elog(LOG, "performsort of worker %d done: %s",
1351 : : state->worker, pg_rusage_show(&state->ru_start));
1352 : : }
1353 : :
7533 tgl@sss.pgh.pa.us 1354 :CBC 147058 : MemoryContextSwitchTo(oldcontext);
9857 1355 : 147058 : }
1356 : :
1357 : : /*
1358 : : * Internal routine to fetch the next tuple in either forward or back
1359 : : * direction into *stup. Returns false if no more tuples.
1360 : : * Returned tuple belongs to tuplesort memory context, and must not be freed
1361 : : * by caller. Note that fetched tuple is stored in memory that may be
1362 : : * recycled by any future fetch.
1363 : : */
1364 : : bool
7533 1365 : 17891825 : tuplesort_gettuple_common(Tuplesortstate *state, bool forward,
1366 : : SortTuple *stup)
1367 : : {
1368 : : unsigned int tuplen;
1369 : : size_t nmoved;
1370 : :
3174 rhaas@postgresql.org 1371 [ + + - + ]: 17891825 : Assert(!WORKER(state));
1372 : :
9857 tgl@sss.pgh.pa.us 1373 [ + + + - ]: 17891825 : switch (state->status)
1374 : : {
1375 : 14948303 : case TSS_SORTEDINMEM:
1538 akorotkov@postgresql 1376 [ + + - + ]: 14948303 : Assert(forward || state->base.sortopt & TUPLESORT_RANDOMACCESS);
3661 heikki.linnakangas@i 1377 [ - + ]: 14948303 : Assert(!state->slabAllocatorUsed);
9857 tgl@sss.pgh.pa.us 1378 [ + + ]: 14948303 : if (forward)
1379 : : {
1380 [ + + ]: 14948259 : if (state->current < state->memtupcount)
1381 : : {
7533 1382 : 14802603 : *stup = state->memtuples[state->current++];
1383 : 14802603 : return true;
1384 : : }
9857 1385 : 145656 : state->eof_reached = true;
1386 : :
1387 : : /*
1388 : : * Complain if caller tries to retrieve more tuples than
1389 : : * originally asked for in a bounded sort. This is because
1390 : : * returning EOF here might be the wrong thing.
1391 : : */
7101 1392 [ + + - + ]: 145656 : if (state->bounded && state->current >= state->bound)
7101 tgl@sss.pgh.pa.us 1393 [ # # ]:UBC 0 : elog(ERROR, "retrieved too many tuples in a bounded sort");
1394 : :
7533 tgl@sss.pgh.pa.us 1395 :CBC 145656 : return false;
1396 : : }
1397 : : else
1398 : : {
9857 1399 [ - + ]: 44 : if (state->current <= 0)
7533 tgl@sss.pgh.pa.us 1400 :UBC 0 : return false;
1401 : :
1402 : : /*
1403 : : * if all tuples are fetched already then we return last
1404 : : * tuple, else - tuple before last returned.
1405 : : */
9857 tgl@sss.pgh.pa.us 1406 [ + + ]:CBC 44 : if (state->eof_reached)
1407 : 8 : state->eof_reached = false;
1408 : : else
1409 : : {
9679 bruce@momjian.us 1410 : 36 : state->current--; /* last returned tuple */
9857 tgl@sss.pgh.pa.us 1411 [ + + ]: 36 : if (state->current <= 0)
7533 1412 : 4 : return false;
1413 : : }
1414 : 40 : *stup = state->memtuples[state->current - 1];
1415 : 40 : return true;
1416 : : }
1417 : : break;
1418 : :
9857 1419 : 196999 : case TSS_SORTEDONTAPE:
1538 akorotkov@postgresql 1420 [ + + - + ]: 196999 : Assert(forward || state->base.sortopt & TUPLESORT_RANDOMACCESS);
3661 heikki.linnakangas@i 1421 [ - + ]: 196999 : Assert(state->slabAllocatorUsed);
1422 : :
1423 : : /*
1424 : : * The slot that held the tuple that we returned in previous
1425 : : * gettuple call can now be reused.
1426 : : */
1427 [ + + ]: 196999 : if (state->lastReturnedTuple)
1428 : : {
1429 [ + - + - ]: 101900 : RELEASE_SLAB_SLOT(state, state->lastReturnedTuple);
1430 : 101900 : state->lastReturnedTuple = NULL;
1431 : : }
1432 : :
9857 tgl@sss.pgh.pa.us 1433 [ + + ]: 196999 : if (forward)
1434 : : {
1435 [ - + ]: 196979 : if (state->eof_reached)
7533 tgl@sss.pgh.pa.us 1436 :UBC 0 : return false;
1437 : :
1820 heikki.linnakangas@i 1438 [ + + ]:CBC 196979 : if ((tuplen = getlen(state->result_tape, true)) != 0)
1439 : : {
7533 tgl@sss.pgh.pa.us 1440 : 196960 : READTUP(state, stup, state->result_tape, tuplen);
1441 : :
1442 : : /*
1443 : : * Remember the tuple we return, so that we can recycle
1444 : : * its memory on next call. (This can be NULL, in the
1445 : : * !state->tuples case).
1446 : : */
3661 heikki.linnakangas@i 1447 : 196960 : state->lastReturnedTuple = stup->tuple;
1448 : :
7533 tgl@sss.pgh.pa.us 1449 : 196960 : return true;
1450 : : }
1451 : : else
1452 : : {
9857 1453 : 19 : state->eof_reached = true;
7533 1454 : 19 : return false;
1455 : : }
1456 : : }
1457 : :
1458 : : /*
1459 : : * Backward.
1460 : : *
1461 : : * if all tuples are fetched already then we return last tuple,
1462 : : * else - tuple before last returned.
1463 : : */
9857 1464 [ + + ]: 20 : if (state->eof_reached)
1465 : : {
1466 : : /*
1467 : : * Seek position is pointing just past the zero tuplen at the
1468 : : * end of file; back up to fetch last tuple's ending length
1469 : : * word. If seek fails we must have a completely empty file.
1470 : : */
1820 heikki.linnakangas@i 1471 : 8 : nmoved = LogicalTapeBackspace(state->result_tape,
1472 : : 2 * sizeof(unsigned int));
3581 1473 [ - + ]: 8 : if (nmoved == 0)
7533 tgl@sss.pgh.pa.us 1474 :UBC 0 : return false;
3581 heikki.linnakangas@i 1475 [ - + ]:CBC 8 : else if (nmoved != 2 * sizeof(unsigned int))
3581 heikki.linnakangas@i 1476 [ # # ]:UBC 0 : elog(ERROR, "unexpected tape position");
9857 tgl@sss.pgh.pa.us 1477 :CBC 8 : state->eof_reached = false;
1478 : : }
1479 : : else
1480 : : {
1481 : : /*
1482 : : * Back up and fetch previously-returned tuple's ending length
1483 : : * word. If seek fails, assume we are at start of file.
1484 : : */
1820 heikki.linnakangas@i 1485 : 12 : nmoved = LogicalTapeBackspace(state->result_tape,
1486 : : sizeof(unsigned int));
3581 1487 [ - + ]: 12 : if (nmoved == 0)
7533 tgl@sss.pgh.pa.us 1488 :UBC 0 : return false;
3581 heikki.linnakangas@i 1489 [ - + ]:CBC 12 : else if (nmoved != sizeof(unsigned int))
3581 heikki.linnakangas@i 1490 [ # # ]:UBC 0 : elog(ERROR, "unexpected tape position");
1820 heikki.linnakangas@i 1491 :CBC 12 : tuplen = getlen(state->result_tape, false);
1492 : :
1493 : : /*
1494 : : * Back up to get ending length word of tuple before it.
1495 : : */
1496 : 12 : nmoved = LogicalTapeBackspace(state->result_tape,
1497 : : tuplen + 2 * sizeof(unsigned int));
3581 1498 [ + + ]: 12 : if (nmoved == tuplen + sizeof(unsigned int))
1499 : : {
1500 : : /*
1501 : : * We backed up over the previous tuple, but there was no
1502 : : * ending length word before it. That means that the prev
1503 : : * tuple is the first tuple in the file. It is now the
1504 : : * next to read in forward direction (not obviously right,
1505 : : * but that is what in-memory case does).
1506 : : */
7533 tgl@sss.pgh.pa.us 1507 : 4 : return false;
1508 : : }
3581 heikki.linnakangas@i 1509 [ - + ]: 8 : else if (nmoved != tuplen + 2 * sizeof(unsigned int))
3581 heikki.linnakangas@i 1510 [ # # ]:UBC 0 : elog(ERROR, "bogus tuple length in backward scan");
1511 : : }
1512 : :
1820 heikki.linnakangas@i 1513 :CBC 16 : tuplen = getlen(state->result_tape, false);
1514 : :
1515 : : /*
1516 : : * Now we have the length of the prior tuple, back up and read it.
1517 : : * Note: READTUP expects we are positioned after the initial
1518 : : * length word of the tuple, so back up to that point.
1519 : : */
1520 : 16 : nmoved = LogicalTapeBackspace(state->result_tape,
1521 : : tuplen);
3581 1522 [ - + ]: 16 : if (nmoved != tuplen)
8480 tgl@sss.pgh.pa.us 1523 [ # # ]:UBC 0 : elog(ERROR, "bogus tuple length in backward scan");
7533 tgl@sss.pgh.pa.us 1524 :CBC 16 : READTUP(state, stup, state->result_tape, tuplen);
1525 : :
1526 : : /*
1527 : : * Remember the tuple we return, so that we can recycle its memory
1528 : : * on next call. (This can be NULL, in the Datum case).
1529 : : */
3661 heikki.linnakangas@i 1530 : 16 : state->lastReturnedTuple = stup->tuple;
1531 : :
7533 tgl@sss.pgh.pa.us 1532 : 16 : return true;
1533 : :
9857 1534 : 2746523 : case TSS_FINALMERGE:
1535 [ - + ]: 2746523 : Assert(forward);
1536 : : /* We are managing memory ourselves, with the slab allocator. */
3661 heikki.linnakangas@i 1537 [ - + ]: 2746523 : Assert(state->slabAllocatorUsed);
1538 : :
1539 : : /*
1540 : : * The slab slot holding the tuple that we returned in previous
1541 : : * gettuple call can now be reused.
1542 : : */
1543 [ + + ]: 2746523 : if (state->lastReturnedTuple)
1544 : : {
1545 [ + - + + ]: 2676297 : RELEASE_SLAB_SLOT(state, state->lastReturnedTuple);
1546 : 2676297 : state->lastReturnedTuple = NULL;
1547 : : }
1548 : :
1549 : : /*
1550 : : * This code should match the inner loop of mergeonerun().
1551 : : */
9844 tgl@sss.pgh.pa.us 1552 [ + + ]: 2746523 : if (state->memtupcount > 0)
1553 : : {
1820 heikki.linnakangas@i 1554 : 2746329 : int srcTapeIndex = state->memtuples[0].srctape;
1555 : 2746329 : LogicalTape *srcTape = state->inputTapes[srcTapeIndex];
1556 : : SortTuple newtup;
1557 : :
3661 1558 : 2746329 : *stup = state->memtuples[0];
1559 : :
1560 : : /*
1561 : : * Remember the tuple we return, so that we can recycle its
1562 : : * memory on next call. (This can be NULL, in the Datum case).
1563 : : */
1564 : 2746329 : state->lastReturnedTuple = stup->tuple;
1565 : :
1566 : : /*
1567 : : * Pull next tuple from tape, and replace the returned tuple
1568 : : * at top of the heap with it.
1569 : : */
1570 [ + + ]: 2746329 : if (!mergereadnext(state, srcTape, &newtup))
1571 : : {
1572 : : /*
1573 : : * If no more data, we've reached end of run on this tape.
1574 : : * Remove the top node from the heap.
1575 : : */
3300 rhaas@postgresql.org 1576 : 263 : tuplesort_heap_delete_top(state);
1820 heikki.linnakangas@i 1577 : 263 : state->nInputRuns--;
1578 : :
1579 : : /*
1580 : : * Close the tape. It'd go away at the end of the sort
1581 : : * anyway, but better to release the memory early.
1582 : : */
1538 akorotkov@postgresql 1583 : 263 : LogicalTapeClose(srcTape);
1584 : 263 : return true;
1585 : : }
1586 : 2746066 : newtup.srctape = srcTapeIndex;
1587 : 2746066 : tuplesort_heap_replace_top(state, &newtup);
1588 : 2746066 : return true;
1589 : : }
1590 : 194 : return false;
1591 : :
1538 akorotkov@postgresql 1592 :UBC 0 : default:
1593 [ # # ]: 0 : elog(ERROR, "invalid tuplesort state");
1594 : : return false; /* keep compiler quiet */
1595 : : }
1596 : : }
1597 : :
1598 : :
1599 : : /*
1600 : : * Advance over N tuples in either forward or back direction,
1601 : : * without returning any data. N==0 is a no-op.
1602 : : * Returns true if successful, false if ran out of tuples.
1603 : : */
1604 : : bool
4676 tgl@sss.pgh.pa.us 1605 :CBC 258 : tuplesort_skiptuples(Tuplesortstate *state, int64 ntuples, bool forward)
1606 : : {
1607 : : MemoryContext oldcontext;
1608 : :
1609 : : /*
1610 : : * We don't actually support backwards skip yet, because no callers need
1611 : : * it. The API is designed to allow for that later, though.
1612 : : */
1613 [ - + ]: 258 : Assert(forward);
1614 [ - + ]: 258 : Assert(ntuples >= 0);
3174 rhaas@postgresql.org 1615 [ - + - - ]: 258 : Assert(!WORKER(state));
1616 : :
4676 tgl@sss.pgh.pa.us 1617 [ + + - ]: 258 : switch (state->status)
1618 : : {
1619 : 242 : case TSS_SORTEDINMEM:
1620 [ + - ]: 242 : if (state->memtupcount - state->current >= ntuples)
1621 : : {
1622 : 242 : state->current += ntuples;
1623 : 242 : return true;
1624 : : }
4676 tgl@sss.pgh.pa.us 1625 :UBC 0 : state->current = state->memtupcount;
1626 : 0 : state->eof_reached = true;
1627 : :
1628 : : /*
1629 : : * Complain if caller tries to retrieve more tuples than
1630 : : * originally asked for in a bounded sort. This is because
1631 : : * returning EOF here might be the wrong thing.
1632 : : */
1633 [ # # # # ]: 0 : if (state->bounded && state->current >= state->bound)
1634 [ # # ]: 0 : elog(ERROR, "retrieved too many tuples in a bounded sort");
1635 : :
1636 : 0 : return false;
1637 : :
4676 tgl@sss.pgh.pa.us 1638 :CBC 16 : case TSS_SORTEDONTAPE:
1639 : : case TSS_FINALMERGE:
1640 : :
1641 : : /*
1642 : : * We could probably optimize these cases better, but for now it's
1643 : : * not worth the trouble.
1644 : : */
1538 akorotkov@postgresql 1645 : 16 : oldcontext = MemoryContextSwitchTo(state->base.sortcontext);
4676 tgl@sss.pgh.pa.us 1646 [ + + ]: 160088 : while (ntuples-- > 0)
1647 : : {
1648 : : SortTuple stup;
1649 : :
3591 rhaas@postgresql.org 1650 [ - + ]: 160072 : if (!tuplesort_gettuple_common(state, forward, &stup))
1651 : : {
4675 tgl@sss.pgh.pa.us 1652 :UBC 0 : MemoryContextSwitchTo(oldcontext);
4676 1653 : 0 : return false;
1654 : : }
4676 tgl@sss.pgh.pa.us 1655 [ - + ]:CBC 160072 : CHECK_FOR_INTERRUPTS();
1656 : : }
4675 1657 : 16 : MemoryContextSwitchTo(oldcontext);
4676 1658 : 16 : return true;
1659 : :
4676 tgl@sss.pgh.pa.us 1660 :UBC 0 : default:
1661 [ # # ]: 0 : elog(ERROR, "invalid tuplesort state");
1662 : : return false; /* keep compiler quiet */
1663 : : }
1664 : : }
1665 : :
1666 : : /*
1667 : : * tuplesort_merge_order - report merge order we'll use for given memory
1668 : : * (note: "merge order" just means the number of input tapes in the merge).
1669 : : *
1670 : : * This is exported for use by the planner. allowedMem is in bytes.
1671 : : */
1672 : : int
4848 noah@leadboat.com 1673 :CBC 10647 : tuplesort_merge_order(int64 allowedMem)
1674 : : {
1675 : : int mOrder;
1676 : :
1677 : : /*----------
1678 : : * In the merge phase, we need buffer space for each input and output tape.
1679 : : * Each pass in the balanced merge algorithm reads from M input tapes, and
1680 : : * writes to N output tapes. Each tape consumes TAPE_BUFFER_OVERHEAD bytes
1681 : : * of memory. In addition to that, we want MERGE_BUFFER_SIZE workspace per
1682 : : * input tape.
1683 : : *
1684 : : * totalMem = M * (TAPE_BUFFER_OVERHEAD + MERGE_BUFFER_SIZE) +
1685 : : * N * TAPE_BUFFER_OVERHEAD
1686 : : *
1687 : : * Except for the last and next-to-last merge passes, where there can be
1688 : : * fewer tapes left to process, M = N. We choose M so that we have the
1689 : : * desired amount of memory available for the input buffers
1690 : : * (TAPE_BUFFER_OVERHEAD + MERGE_BUFFER_SIZE), given the total memory
1691 : : * available for the tape buffers (allowedMem).
1692 : : *
1693 : : * Note: you might be thinking we need to account for the memtuples[]
1694 : : * array in this calculation, but we effectively treat that as part of the
1695 : : * MERGE_BUFFER_SIZE workspace.
1696 : : *----------
1697 : : */
1820 heikki.linnakangas@i 1698 : 10647 : mOrder = allowedMem /
1699 : : (2 * TAPE_BUFFER_OVERHEAD + MERGE_BUFFER_SIZE);
1700 : :
1701 : : /*
1702 : : * Even in minimum memory, use at least a MINORDER merge. On the other
1703 : : * hand, even when we have lots of memory, do not use more than a MAXORDER
1704 : : * merge. Tapes are pretty cheap, but they're not entirely free. Each
1705 : : * additional tape reduces the amount of memory available to build runs,
1706 : : * which in turn can cause the same sort to need more runs, which makes
1707 : : * merging slower even if it can still be done in a single pass. Also,
1708 : : * high order merges are quite slow due to CPU cache effects; it can be
1709 : : * faster to pay the I/O cost of a multi-pass merge than to perform a
1710 : : * single merge pass across many hundreds of tapes.
1711 : : */
7533 tgl@sss.pgh.pa.us 1712 : 10647 : mOrder = Max(mOrder, MINORDER);
3618 rhaas@postgresql.org 1713 : 10647 : mOrder = Min(mOrder, MAXORDER);
1714 : :
7533 tgl@sss.pgh.pa.us 1715 : 10647 : return mOrder;
1716 : : }
1717 : :
1718 : : /*
1719 : : * Helper function to calculate how much memory to allocate for the read buffer
1720 : : * of each input tape in a merge pass.
1721 : : *
1722 : : * 'avail_mem' is the amount of memory available for the buffers of all the
1723 : : * tapes, both input and output.
1724 : : * 'nInputTapes' and 'nInputRuns' are the number of input tapes and runs.
1725 : : * 'maxOutputTapes' is the max. number of output tapes we should produce.
1726 : : */
1727 : : static int64
1820 heikki.linnakangas@i 1728 : 260 : merge_read_buffer_size(int64 avail_mem, int nInputTapes, int nInputRuns,
1729 : : int maxOutputTapes)
1730 : : {
1731 : : int nOutputRuns;
1732 : : int nOutputTapes;
1733 : :
1734 : : /*
1735 : : * How many output tapes will we produce in this pass?
1736 : : *
1737 : : * This is nInputRuns / nInputTapes, rounded up.
1738 : : */
1739 : 260 : nOutputRuns = (nInputRuns + nInputTapes - 1) / nInputTapes;
1740 : :
1741 : 260 : nOutputTapes = Min(nOutputRuns, maxOutputTapes);
1742 : :
1743 : : /*
1744 : : * Each output tape consumes TAPE_BUFFER_OVERHEAD bytes of memory. All
1745 : : * remaining memory is divided evenly between the input tapes.
1746 : : *
1747 : : * This also follows from the formula in tuplesort_merge_order, but here
1748 : : * we derive the input buffer size from the amount of memory available,
1749 : : * and M and N.
1750 : : */
1751 : 260 : return Max((avail_mem - TAPE_BUFFER_OVERHEAD * nOutputTapes) / nInputTapes, 0);
1752 : : }
1753 : :
1754 : : /*
1755 : : * inittapes - initialize for tape sorting.
1756 : : *
1757 : : * This is called only if we have found we won't sort in memory.
1758 : : */
1759 : : static void
3174 rhaas@postgresql.org 1760 : 469 : inittapes(Tuplesortstate *state, bool mergeruns)
1761 : : {
1762 [ + + - + ]: 469 : Assert(!LEADER(state));
1763 : :
1764 [ + + ]: 469 : if (mergeruns)
1765 : : {
1766 : : /* Compute number of input tapes to use when merging */
1820 heikki.linnakangas@i 1767 : 90 : state->maxTapes = tuplesort_merge_order(state->allowedMem);
1768 : : }
1769 : : else
1770 : : {
1771 : : /* Workers can sometimes produce single run, output without merge */
3174 rhaas@postgresql.org 1772 [ + - - + ]: 379 : Assert(WORKER(state));
1820 heikki.linnakangas@i 1773 : 379 : state->maxTapes = MINORDER;
1774 : : }
1775 : :
7679 tgl@sss.pgh.pa.us 1776 [ - + ]: 469 : if (trace_sort)
2902 pg@bowt.ie 1777 [ # # ]:UBC 0 : elog(LOG, "worker %d switching to external sort with %d tapes: %s",
1778 : : state->worker, state->maxTapes, pg_rusage_show(&state->ru_start));
1779 : :
1780 : : /* Create the tape set */
1820 heikki.linnakangas@i 1781 :CBC 469 : inittapestate(state, state->maxTapes);
3174 rhaas@postgresql.org 1782 : 469 : state->tapeset =
1820 heikki.linnakangas@i 1783 : 469 : LogicalTapeSetCreate(false,
3174 rhaas@postgresql.org 1784 [ + + ]: 469 : state->shared ? &state->shared->fileset : NULL,
1785 : : state->worker);
1786 : :
3300 1787 : 469 : state->currentRun = 0;
1788 : :
1789 : : /*
1790 : : * Initialize logical tape arrays.
1791 : : */
1820 heikki.linnakangas@i 1792 : 469 : state->inputTapes = NULL;
1793 : 469 : state->nInputTapes = 0;
1794 : 469 : state->nInputRuns = 0;
1795 : :
56 michael@paquier.xyz 1796 :GNC 469 : state->outputTapes = palloc0_array(LogicalTape *, state->maxTapes);
1820 heikki.linnakangas@i 1797 :CBC 469 : state->nOutputTapes = 0;
1798 : 469 : state->nOutputRuns = 0;
1799 : :
9857 tgl@sss.pgh.pa.us 1800 : 469 : state->status = TSS_BUILDRUNS;
1801 : :
1820 heikki.linnakangas@i 1802 : 469 : selectnewtape(state);
9857 tgl@sss.pgh.pa.us 1803 : 469 : }
1804 : :
1805 : : /*
1806 : : * inittapestate - initialize generic tape management state
1807 : : */
1808 : : static void
3174 rhaas@postgresql.org 1809 : 599 : inittapestate(Tuplesortstate *state, int maxTapes)
1810 : : {
1811 : : int64 tapeSpace;
1812 : :
1813 : : /*
1814 : : * Decrease availMem to reflect the space needed for tape buffers; but
1815 : : * don't decrease it to the point that we have no room for tuples. (That
1816 : : * case is only likely to occur if sorting pass-by-value Datums; in all
1817 : : * other scenarios the memtuples[] array is unlikely to occupy more than
1818 : : * half of allowedMem. In the pass-by-value case it's not important to
1819 : : * account for tuple space, so we don't care if LACKMEM becomes
1820 : : * inaccurate.)
1821 : : */
1822 : 599 : tapeSpace = (int64) maxTapes * TAPE_BUFFER_OVERHEAD;
1823 : :
1824 [ + + ]: 599 : if (tapeSpace + GetMemoryChunkSpace(state->memtuples) < state->allowedMem)
1825 : 521 : USEMEM(state, tapeSpace);
1826 : :
1827 : : /*
1828 : : * Make sure that the temp file(s) underlying the tape set are created in
1829 : : * suitable temp tablespaces. For parallel sorts, this should have been
1830 : : * called already, but it doesn't matter if it is called a second time.
1831 : : */
1832 : 599 : PrepareTempTablespaces();
1833 : 599 : }
1834 : :
1835 : : /*
1836 : : * selectnewtape -- select next tape to output to.
1837 : : *
1838 : : * This is called after finishing a run when we know another run
1839 : : * must be started. This is used both when building the initial
1840 : : * runs, and during merge passes.
1841 : : */
1842 : : static void
9857 tgl@sss.pgh.pa.us 1843 : 1304 : selectnewtape(Tuplesortstate *state)
1844 : : {
1845 : : /*
1846 : : * At the beginning of each merge pass, nOutputTapes and nOutputRuns are
1847 : : * both zero. On each call, we create a new output tape to hold the next
1848 : : * run, until maxTapes is reached. After that, we assign new runs to the
1849 : : * existing tapes in a round robin fashion.
1850 : : */
1813 heikki.linnakangas@i 1851 [ + + ]: 1304 : if (state->nOutputTapes < state->maxTapes)
1852 : : {
1853 : : /* Create a new tape to hold the next run */
1820 1854 [ - + ]: 876 : Assert(state->outputTapes[state->nOutputRuns] == NULL);
1855 [ - + ]: 876 : Assert(state->nOutputRuns == state->nOutputTapes);
1856 : 876 : state->destTape = LogicalTapeCreate(state->tapeset);
1813 1857 : 876 : state->outputTapes[state->nOutputTapes] = state->destTape;
1820 1858 : 876 : state->nOutputTapes++;
1859 : 876 : state->nOutputRuns++;
1860 : : }
1861 : : else
1862 : : {
1863 : : /*
1864 : : * We have reached the max number of tapes. Append to an existing
1865 : : * tape.
1866 : : */
1867 : 428 : state->destTape = state->outputTapes[state->nOutputRuns % state->nOutputTapes];
1868 : 428 : state->nOutputRuns++;
1869 : : }
9857 tgl@sss.pgh.pa.us 1870 : 1304 : }
1871 : :
1872 : : /*
1873 : : * Initialize the slab allocation arena, for the given number of slots.
1874 : : */
1875 : : static void
3661 heikki.linnakangas@i 1876 : 220 : init_slab_allocator(Tuplesortstate *state, int numSlots)
1877 : : {
1878 [ + + ]: 220 : if (numSlots > 0)
1879 : : {
1880 : : char *p;
1881 : : int i;
1882 : :
1883 : 204 : state->slabMemoryBegin = palloc(numSlots * SLAB_SLOT_SIZE);
1884 : 204 : state->slabMemoryEnd = state->slabMemoryBegin +
1885 : 204 : numSlots * SLAB_SLOT_SIZE;
1886 : 204 : state->slabFreeHead = (SlabSlot *) state->slabMemoryBegin;
1887 : 204 : USEMEM(state, numSlots * SLAB_SLOT_SIZE);
1888 : :
1889 : 204 : p = state->slabMemoryBegin;
1890 [ + + ]: 802 : for (i = 0; i < numSlots - 1; i++)
1891 : : {
1892 : 598 : ((SlabSlot *) p)->nextfree = (SlabSlot *) (p + SLAB_SLOT_SIZE);
1893 : 598 : p += SLAB_SLOT_SIZE;
1894 : : }
1895 : 204 : ((SlabSlot *) p)->nextfree = NULL;
1896 : : }
1897 : : else
1898 : : {
1899 : 16 : state->slabMemoryBegin = state->slabMemoryEnd = NULL;
1900 : 16 : state->slabFreeHead = NULL;
1901 : : }
1902 : 220 : state->slabAllocatorUsed = true;
1903 : 220 : }
1904 : :
1905 : : /*
1906 : : * mergeruns -- merge all the completed initial runs.
1907 : : *
1908 : : * This implements the Balanced k-Way Merge Algorithm. All input data has
1909 : : * already been written to initial runs on tape (see dumptuples).
1910 : : */
1911 : : static void
9857 tgl@sss.pgh.pa.us 1912 : 220 : mergeruns(Tuplesortstate *state)
1913 : : {
1914 : : int tapenum;
1915 : :
1916 [ - + ]: 220 : Assert(state->status == TSS_BUILDRUNS);
9844 1917 [ - + ]: 220 : Assert(state->memtupcount == 0);
1918 : :
1538 akorotkov@postgresql 1919 [ + + + + ]: 220 : if (state->base.sortKeys != NULL && state->base.sortKeys->abbrev_converter != NULL)
1920 : : {
1921 : : /*
1922 : : * If there are multiple runs to be merged, when we go to read back
1923 : : * tuples from disk, abbreviated keys will not have been stored, and
1924 : : * we don't care to regenerate them. Disable abbreviation from this
1925 : : * point on.
1926 : : */
1927 : 19 : state->base.sortKeys->abbrev_converter = NULL;
1928 : 19 : state->base.sortKeys->comparator = state->base.sortKeys->abbrev_full_comparator;
1929 : :
1930 : : /* Not strictly necessary, but be tidy */
1931 : 19 : state->base.sortKeys->abbrev_abort = NULL;
1932 : 19 : state->base.sortKeys->abbrev_full_comparator = NULL;
1933 : : }
1934 : :
1935 : : /*
1936 : : * Reset tuple memory. We've freed all the tuples that we previously
1937 : : * allocated. We will use the slab allocator from now on.
1938 : : */
1939 : 220 : MemoryContextResetOnly(state->base.tuplecontext);
1940 : :
1941 : : /*
1942 : : * We no longer need a large memtuples array. (We will allocate a smaller
1943 : : * one for the heap later.)
1944 : : */
3661 heikki.linnakangas@i 1945 : 220 : FREEMEM(state, GetMemoryChunkSpace(state->memtuples));
1946 : 220 : pfree(state->memtuples);
1947 : 220 : state->memtuples = NULL;
1948 : :
1949 : : /*
1950 : : * Initialize the slab allocator. We need one slab slot per input tape,
1951 : : * for the tuples in the heap, plus one to hold the tuple last returned
1952 : : * from tuplesort_gettuple. (If we're sorting pass-by-val Datums,
1953 : : * however, we don't need to do allocate anything.)
1954 : : *
1955 : : * In a multi-pass merge, we could shrink this allocation for the last
1956 : : * merge pass, if it has fewer tapes than previous passes, but we don't
1957 : : * bother.
1958 : : *
1959 : : * From this point on, we no longer use the USEMEM()/LACKMEM() mechanism
1960 : : * to track memory usage of individual tuples.
1961 : : */
1538 akorotkov@postgresql 1962 [ + + ]: 220 : if (state->base.tuples)
1820 heikki.linnakangas@i 1963 : 204 : init_slab_allocator(state, state->nOutputTapes + 1);
1964 : : else
3661 1965 : 16 : init_slab_allocator(state, 0);
1966 : :
1967 : : /*
1968 : : * Allocate a new 'memtuples' array, for the heap. It will hold one tuple
1969 : : * from each input tape.
1970 : : *
1971 : : * We could shrink this, too, between passes in a multi-pass merge, but we
1972 : : * don't bother. (The initial input tapes are still in outputTapes. The
1973 : : * number of input tapes will not increase between passes.)
1974 : : */
1820 1975 : 220 : state->memtupsize = state->nOutputTapes;
1538 akorotkov@postgresql 1976 : 440 : state->memtuples = (SortTuple *) MemoryContextAlloc(state->base.maincontext,
1820 heikki.linnakangas@i 1977 : 220 : state->nOutputTapes * sizeof(SortTuple));
3595 1978 : 220 : USEMEM(state, GetMemoryChunkSpace(state->memtuples));
1979 : :
1980 : : /*
1981 : : * Use all the remaining memory we have available for tape buffers among
1982 : : * all the input tapes. At the beginning of each merge pass, we will
1983 : : * divide this memory between the input and output tapes in the pass.
1984 : : */
1820 1985 : 220 : state->tape_buffer_mem = state->availMem;
1813 1986 : 220 : USEMEM(state, state->tape_buffer_mem);
3652 1987 [ - + ]: 220 : if (trace_sort)
1820 heikki.linnakangas@i 1988 [ # # ]:UBC 0 : elog(LOG, "worker %d using %zu KB of memory for tape buffers",
1989 : : state->worker, state->tape_buffer_mem / 1024);
1990 : :
1991 : : for (;;)
1992 : : {
1993 : : /*
1994 : : * On the first iteration, or if we have read all the runs from the
1995 : : * input tapes in a multi-pass merge, it's time to start a new pass.
1996 : : * Rewind all the output tapes, and make them inputs for the next
1997 : : * pass.
1998 : : */
1820 heikki.linnakangas@i 1999 [ + + ]:CBC 352 : if (state->nInputRuns == 0)
2000 : : {
2001 : : int64 input_buffer_size;
2002 : :
2003 : : /* Close the old, emptied, input tapes */
2004 [ + + ]: 260 : if (state->nInputTapes > 0)
2005 : : {
2006 [ + + ]: 280 : for (tapenum = 0; tapenum < state->nInputTapes; tapenum++)
2007 : 240 : LogicalTapeClose(state->inputTapes[tapenum]);
2008 : 40 : pfree(state->inputTapes);
2009 : : }
2010 : :
2011 : : /* Previous pass's outputs become next pass's inputs. */
2012 : 260 : state->inputTapes = state->outputTapes;
2013 : 260 : state->nInputTapes = state->nOutputTapes;
2014 : 260 : state->nInputRuns = state->nOutputRuns;
2015 : :
2016 : : /*
2017 : : * Reset output tape variables. The actual LogicalTapes will be
2018 : : * created as needed, here we only allocate the array to hold
2019 : : * them.
2020 : : */
56 michael@paquier.xyz 2021 :GNC 260 : state->outputTapes = palloc0_array(LogicalTape *, state->nInputTapes);
1820 heikki.linnakangas@i 2022 :CBC 260 : state->nOutputTapes = 0;
2023 : 260 : state->nOutputRuns = 0;
2024 : :
2025 : : /*
2026 : : * Redistribute the memory allocated for tape buffers, among the
2027 : : * new input and output tapes.
2028 : : */
2029 : 260 : input_buffer_size = merge_read_buffer_size(state->tape_buffer_mem,
2030 : : state->nInputTapes,
2031 : : state->nInputRuns,
2032 : : state->maxTapes);
2033 : :
2034 [ - + ]: 260 : if (trace_sort)
1820 heikki.linnakangas@i 2035 [ # # ]:UBC 0 : elog(LOG, "starting merge pass of %d input runs on %d tapes, " INT64_FORMAT " KB of memory for each input tape: %s",
2036 : : state->nInputRuns, state->nInputTapes, input_buffer_size / 1024,
2037 : : pg_rusage_show(&state->ru_start));
2038 : :
2039 : : /* Prepare the new input tapes for merge pass. */
1820 heikki.linnakangas@i 2040 [ + + ]:CBC 1026 : for (tapenum = 0; tapenum < state->nInputTapes; tapenum++)
2041 : 766 : LogicalTapeRewindForRead(state->inputTapes[tapenum], input_buffer_size);
2042 : :
2043 : : /*
2044 : : * If there's just one run left on each input tape, then only one
2045 : : * merge pass remains. If we don't have to produce a materialized
2046 : : * sorted tape, we can stop at this point and do the final merge
2047 : : * on-the-fly.
2048 : : */
1538 akorotkov@postgresql 2049 [ + + ]: 260 : if ((state->base.sortopt & TUPLESORT_RANDOMACCESS) == 0
1652 drowley@postgresql.o 2050 [ + + ]: 242 : && state->nInputRuns <= state->nInputTapes
1820 heikki.linnakangas@i 2051 [ + + + - ]: 206 : && !WORKER(state))
2052 : : {
2053 : : /* Tell logtape.c we won't be writing anymore */
7524 tgl@sss.pgh.pa.us 2054 : 206 : LogicalTapeSetForgetFreeSpace(state->tapeset);
2055 : : /* Initialize for the final merge pass */
3661 heikki.linnakangas@i 2056 : 206 : beginmerge(state);
9857 tgl@sss.pgh.pa.us 2057 : 206 : state->status = TSS_FINALMERGE;
2058 : 206 : return;
2059 : : }
2060 : : }
2061 : :
2062 : : /* Select an output tape */
1820 heikki.linnakangas@i 2063 : 146 : selectnewtape(state);
2064 : :
2065 : : /* Merge one run from each input tape. */
2066 : 146 : mergeonerun(state);
2067 : :
2068 : : /*
2069 : : * If the input tapes are empty, and we output only one output run,
2070 : : * we're done. The current output tape contains the final result.
2071 : : */
2072 [ + + + + ]: 146 : if (state->nInputRuns == 0 && state->nOutputRuns <= 1)
2073 : 14 : break;
2074 : : }
2075 : :
2076 : : /*
2077 : : * Done. The result is on a single run on a single tape.
2078 : : */
2079 : 14 : state->result_tape = state->outputTapes[0];
3174 rhaas@postgresql.org 2080 [ - + - - ]: 14 : if (!WORKER(state))
1820 heikki.linnakangas@i 2081 : 14 : LogicalTapeFreeze(state->result_tape, NULL);
2082 : : else
3174 rhaas@postgresql.org 2083 :UBC 0 : worker_freeze_result_tape(state);
9857 tgl@sss.pgh.pa.us 2084 :CBC 14 : state->status = TSS_SORTEDONTAPE;
2085 : :
2086 : : /* Close all the now-empty input tapes, to release their read buffers. */
1820 heikki.linnakangas@i 2087 [ + + ]: 62 : for (tapenum = 0; tapenum < state->nInputTapes; tapenum++)
2088 : 48 : LogicalTapeClose(state->inputTapes[tapenum]);
2089 : : }
2090 : :
2091 : : /*
2092 : : * Merge one run from each input tape.
2093 : : */
2094 : : static void
9857 tgl@sss.pgh.pa.us 2095 : 146 : mergeonerun(Tuplesortstate *state)
2096 : : {
2097 : : int srcTapeIndex;
2098 : : LogicalTape *srcTape;
2099 : :
2100 : : /*
2101 : : * Start the merge by loading one tuple from each active source tape into
2102 : : * the heap.
2103 : : */
3661 heikki.linnakangas@i 2104 : 146 : beginmerge(state);
2105 : :
1502 drowley@postgresql.o 2106 [ - + ]: 146 : Assert(state->slabAllocatorUsed);
2107 : :
2108 : : /*
2109 : : * Execute merge by repeatedly extracting lowest tuple in heap, writing it
2110 : : * out, and replacing it with next tuple from same tape (if there is
2111 : : * another one).
2112 : : */
9844 tgl@sss.pgh.pa.us 2113 [ + + ]: 980650 : while (state->memtupcount > 0)
2114 : : {
2115 : : SortTuple stup;
2116 : :
2117 : : /* write the tuple to destTape */
1820 heikki.linnakangas@i 2118 : 980504 : srcTapeIndex = state->memtuples[0].srctape;
2119 : 980504 : srcTape = state->inputTapes[srcTapeIndex];
2120 : 980504 : WRITETUP(state, state->destTape, &state->memtuples[0]);
2121 : :
2122 : : /* recycle the slot of the tuple we just wrote out, for the next read */
3556 tgl@sss.pgh.pa.us 2123 [ + + ]: 980504 : if (state->memtuples[0].tuple)
2124 [ + - + - ]: 890416 : RELEASE_SLAB_SLOT(state, state->memtuples[0].tuple);
2125 : :
2126 : : /*
2127 : : * pull next tuple from the tape, and replace the written-out tuple in
2128 : : * the heap with it.
2129 : : */
3661 heikki.linnakangas@i 2130 [ + + ]: 980504 : if (mergereadnext(state, srcTape, &stup))
2131 : : {
1820 2132 : 979788 : stup.srctape = srcTapeIndex;
3300 rhaas@postgresql.org 2133 : 979788 : tuplesort_heap_replace_top(state, &stup);
2134 : : }
2135 : : else
2136 : : {
2137 : 716 : tuplesort_heap_delete_top(state);
1820 heikki.linnakangas@i 2138 : 716 : state->nInputRuns--;
2139 : : }
2140 : : }
2141 : :
2142 : : /*
2143 : : * When the heap empties, we're done. Write an end-of-run marker on the
2144 : : * output tape.
2145 : : */
2146 : 146 : markrunend(state->destTape);
9857 tgl@sss.pgh.pa.us 2147 : 146 : }
2148 : :
2149 : : /*
2150 : : * beginmerge - initialize for a merge pass
2151 : : *
2152 : : * Fill the merge heap with the first tuple from each input tape.
2153 : : */
2154 : : static void
3661 heikki.linnakangas@i 2155 : 352 : beginmerge(Tuplesortstate *state)
2156 : : {
2157 : : int activeTapes;
2158 : : int srcTapeIndex;
2159 : :
2160 : : /* Heap should be empty here */
9844 tgl@sss.pgh.pa.us 2161 [ - + ]: 352 : Assert(state->memtupcount == 0);
2162 : :
1820 heikki.linnakangas@i 2163 : 352 : activeTapes = Min(state->nInputTapes, state->nInputRuns);
2164 : :
2165 [ + + ]: 1546 : for (srcTapeIndex = 0; srcTapeIndex < activeTapes; srcTapeIndex++)
2166 : : {
2167 : : SortTuple tup;
2168 : :
2169 [ + + ]: 1194 : if (mergereadnext(state, state->inputTapes[srcTapeIndex], &tup))
2170 : : {
2171 : 999 : tup.srctape = srcTapeIndex;
3300 rhaas@postgresql.org 2172 : 999 : tuplesort_heap_insert(state, &tup);
2173 : : }
2174 : : }
3861 2175 : 352 : }
2176 : :
2177 : : /*
2178 : : * mergereadnext - read next tuple from one merge input tape
2179 : : *
2180 : : * Returns false on EOF.
2181 : : */
2182 : : static bool
1820 heikki.linnakangas@i 2183 : 3728027 : mergereadnext(Tuplesortstate *state, LogicalTape *srcTape, SortTuple *stup)
2184 : : {
2185 : : unsigned int tuplen;
2186 : :
2187 : : /* read next tuple, if any */
2188 [ + + ]: 3728027 : if ((tuplen = getlen(srcTape, true)) == 0)
3661 2189 : 1174 : return false;
2190 : 3726853 : READTUP(state, stup, srcTape, tuplen);
2191 : :
2192 : 3726853 : return true;
2193 : : }
2194 : :
2195 : : /*
2196 : : * dumptuples - remove tuples from memtuples and write initial run to tape
2197 : : *
2198 : : * When alltuples = true, dump everything currently in memory. (This case is
2199 : : * only used at end of input data.)
2200 : : */
2201 : : static void
9857 tgl@sss.pgh.pa.us 2202 : 771562 : dumptuples(Tuplesortstate *state, bool alltuples)
2203 : : {
2204 : : int memtupwrite;
2205 : : int i;
2206 : :
2207 : : /*
2208 : : * Nothing to do if we still fit in available memory and have array slots,
2209 : : * unless this is the final call during initial run generation.
2210 : : */
3300 rhaas@postgresql.org 2211 [ + + + + : 771562 : if (state->memtupcount < state->memtupsize && !LACKMEM(state) &&
- + ]
2212 [ + + ]: 770873 : !alltuples)
2213 : 770404 : return;
2214 : :
2215 : : /*
2216 : : * Final call might require no sorting, in rare cases where we just so
2217 : : * happen to have previously LACKMEM()'d at the point where exactly all
2218 : : * remaining tuples are loaded into memory, just before input was
2219 : : * exhausted. In general, short final runs are quite possible, but avoid
2220 : : * creating a completely empty run. In a worker, though, we must produce
2221 : : * at least one tape, even if it's empty.
2222 : : */
1820 heikki.linnakangas@i 2223 [ + + - + ]: 1158 : if (state->memtupcount == 0 && state->currentRun > 0)
1820 heikki.linnakangas@i 2224 :UBC 0 : return;
2225 : :
3839 rhaas@postgresql.org 2226 [ - + ]:CBC 1158 : Assert(state->status == TSS_BUILDRUNS);
2227 : :
2228 : : /*
2229 : : * It seems unlikely that this limit will ever be exceeded, but take no
2230 : : * chances
2231 : : */
2232 [ - + ]: 1158 : if (state->currentRun == INT_MAX)
3839 rhaas@postgresql.org 2233 [ # # ]:UBC 0 : ereport(ERROR,
2234 : : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
2235 : : errmsg("cannot have more than %d runs for an external sort",
2236 : : INT_MAX)));
2237 : :
1820 heikki.linnakangas@i 2238 [ + + ]:CBC 1158 : if (state->currentRun > 0)
2239 : 689 : selectnewtape(state);
2240 : :
3839 rhaas@postgresql.org 2241 : 1158 : state->currentRun++;
2242 : :
2243 [ - + ]: 1158 : if (trace_sort)
2902 pg@bowt.ie 2244 [ # # ]:UBC 0 : elog(LOG, "worker %d starting quicksort of run %d: %s",
2245 : : state->worker, state->currentRun,
2246 : : pg_rusage_show(&state->ru_start));
2247 : :
2248 : : /*
2249 : : * Sort all tuples accumulated within the allowed amount of memory for
2250 : : * this run.
2251 : : */
3839 rhaas@postgresql.org 2252 :CBC 1158 : tuplesort_sort_memtuples(state);
2253 : :
2254 [ - + ]: 1158 : if (trace_sort)
2902 pg@bowt.ie 2255 [ # # ]:UBC 0 : elog(LOG, "worker %d finished quicksort of run %d: %s",
2256 : : state->worker, state->currentRun,
2257 : : pg_rusage_show(&state->ru_start));
2258 : :
3839 rhaas@postgresql.org 2259 :CBC 1158 : memtupwrite = state->memtupcount;
2260 [ + + ]: 3087655 : for (i = 0; i < memtupwrite; i++)
2261 : : {
1502 drowley@postgresql.o 2262 : 3086497 : SortTuple *stup = &state->memtuples[i];
2263 : :
2264 : 3086497 : WRITETUP(state, state->destTape, stup);
2265 : : }
2266 : :
2267 : 1158 : state->memtupcount = 0;
2268 : :
2269 : : /*
2270 : : * Reset tuple memory. We've freed all of the tuples that we previously
2271 : : * allocated. It's important to avoid fragmentation when there is a stark
2272 : : * change in the sizes of incoming tuples. In bounded sorts,
2273 : : * fragmentation due to AllocSetFree's bucketing by size class might be
2274 : : * particularly bad if this step wasn't taken.
2275 : : */
1538 akorotkov@postgresql 2276 : 1158 : MemoryContextReset(state->base.tuplecontext);
2277 : :
2278 : : /*
2279 : : * Now update the memory accounting to subtract the memory used by the
2280 : : * tuple.
2281 : : */
917 drowley@postgresql.o 2282 : 1158 : FREEMEM(state, state->tupleMem);
2283 : 1158 : state->tupleMem = 0;
2284 : :
1820 heikki.linnakangas@i 2285 : 1158 : markrunend(state->destTape);
2286 : :
3839 rhaas@postgresql.org 2287 [ - + ]: 1158 : if (trace_sort)
2902 pg@bowt.ie 2288 [ # # ]:UBC 0 : elog(LOG, "worker %d finished writing run %d to tape %d: %s",
2289 : : state->worker, state->currentRun, (state->currentRun - 1) % state->nOutputTapes + 1,
2290 : : pg_rusage_show(&state->ru_start));
2291 : : }
2292 : :
2293 : : /*
2294 : : * tuplesort_rescan - rewind and replay the scan
2295 : : */
2296 : : void
9857 tgl@sss.pgh.pa.us 2297 :CBC 38 : tuplesort_rescan(Tuplesortstate *state)
2298 : : {
1538 akorotkov@postgresql 2299 : 38 : MemoryContext oldcontext = MemoryContextSwitchTo(state->base.sortcontext);
2300 : :
2301 [ - + ]: 38 : Assert(state->base.sortopt & TUPLESORT_RANDOMACCESS);
2302 : :
9857 tgl@sss.pgh.pa.us 2303 [ + + - ]: 38 : switch (state->status)
2304 : : {
2305 : 33 : case TSS_SORTEDINMEM:
2306 : 33 : state->current = 0;
2307 : 33 : state->eof_reached = false;
2308 : 33 : state->markpos_offset = 0;
2309 : 33 : state->markpos_eof = false;
2310 : 33 : break;
2311 : 5 : case TSS_SORTEDONTAPE:
1820 heikki.linnakangas@i 2312 : 5 : LogicalTapeRewindForRead(state->result_tape, 0);
9857 tgl@sss.pgh.pa.us 2313 : 5 : state->eof_reached = false;
2314 : 5 : state->markpos_block = 0L;
2315 : 5 : state->markpos_offset = 0;
2316 : 5 : state->markpos_eof = false;
2317 : 5 : break;
9857 tgl@sss.pgh.pa.us 2318 :UBC 0 : default:
8480 2319 [ # # ]: 0 : elog(ERROR, "invalid tuplesort state");
2320 : : break;
2321 : : }
2322 : :
7533 tgl@sss.pgh.pa.us 2323 :CBC 38 : MemoryContextSwitchTo(oldcontext);
9857 2324 : 38 : }
2325 : :
2326 : : /*
2327 : : * tuplesort_markpos - saves current position in the merged sort file
2328 : : */
2329 : : void
2330 : 357137 : tuplesort_markpos(Tuplesortstate *state)
2331 : : {
1538 akorotkov@postgresql 2332 : 357137 : MemoryContext oldcontext = MemoryContextSwitchTo(state->base.sortcontext);
2333 : :
2334 [ - + ]: 357137 : Assert(state->base.sortopt & TUPLESORT_RANDOMACCESS);
2335 : :
9857 tgl@sss.pgh.pa.us 2336 [ + + - ]: 357137 : switch (state->status)
2337 : : {
2338 : 351265 : case TSS_SORTEDINMEM:
2339 : 351265 : state->markpos_offset = state->current;
2340 : 351265 : state->markpos_eof = state->eof_reached;
2341 : 351265 : break;
2342 : 5872 : case TSS_SORTEDONTAPE:
1820 heikki.linnakangas@i 2343 : 5872 : LogicalTapeTell(state->result_tape,
2344 : : &state->markpos_block,
2345 : : &state->markpos_offset);
9857 tgl@sss.pgh.pa.us 2346 : 5872 : state->markpos_eof = state->eof_reached;
2347 : 5872 : break;
9857 tgl@sss.pgh.pa.us 2348 :UBC 0 : default:
8480 2349 [ # # ]: 0 : elog(ERROR, "invalid tuplesort state");
2350 : : break;
2351 : : }
2352 : :
7533 tgl@sss.pgh.pa.us 2353 :CBC 357137 : MemoryContextSwitchTo(oldcontext);
9857 2354 : 357137 : }
2355 : :
2356 : : /*
2357 : : * tuplesort_restorepos - restores current position in merged sort file to
2358 : : * last saved position
2359 : : */
2360 : : void
2361 : 28306 : tuplesort_restorepos(Tuplesortstate *state)
2362 : : {
1538 akorotkov@postgresql 2363 : 28306 : MemoryContext oldcontext = MemoryContextSwitchTo(state->base.sortcontext);
2364 : :
2365 [ - + ]: 28306 : Assert(state->base.sortopt & TUPLESORT_RANDOMACCESS);
2366 : :
9857 tgl@sss.pgh.pa.us 2367 [ + + - ]: 28306 : switch (state->status)
2368 : : {
2369 : 24178 : case TSS_SORTEDINMEM:
2370 : 24178 : state->current = state->markpos_offset;
2371 : 24178 : state->eof_reached = state->markpos_eof;
2372 : 24178 : break;
2373 : 4128 : case TSS_SORTEDONTAPE:
1820 heikki.linnakangas@i 2374 : 4128 : LogicalTapeSeek(state->result_tape,
2375 : : state->markpos_block,
2376 : : state->markpos_offset);
9857 tgl@sss.pgh.pa.us 2377 : 4128 : state->eof_reached = state->markpos_eof;
2378 : 4128 : break;
9857 tgl@sss.pgh.pa.us 2379 :UBC 0 : default:
8480 2380 [ # # ]: 0 : elog(ERROR, "invalid tuplesort state");
2381 : : break;
2382 : : }
2383 : :
7533 tgl@sss.pgh.pa.us 2384 :CBC 28306 : MemoryContextSwitchTo(oldcontext);
9857 2385 : 28306 : }
2386 : :
2387 : : /*
2388 : : * tuplesort_get_stats - extract summary statistics
2389 : : *
2390 : : * This can be called after tuplesort_performsort() finishes to obtain
2391 : : * printable summary information about how the sort was performed.
2392 : : */
2393 : : void
6272 2394 : 264 : tuplesort_get_stats(Tuplesortstate *state,
2395 : : TuplesortInstrumentation *stats)
2396 : : {
2397 : : /*
2398 : : * Note: it might seem we should provide both memory and disk usage for a
2399 : : * disk-based sort. However, the current code doesn't track memory space
2400 : : * accurately once we have begun to return tuples to the caller (since we
2401 : : * don't account for pfree's the caller is expected to do), so we cannot
2402 : : * rely on availMem in a disk sort. This does not seem worth the overhead
2403 : : * to fix. Is it worth creating an API for the memory context code to
2404 : : * tell us how much is actually used in sortcontext?
2405 : : */
2380 tomas.vondra@postgre 2406 : 264 : tuplesort_updatemax(state);
2407 : :
2408 [ + + ]: 264 : if (state->isMaxSpaceDisk)
3331 rhaas@postgresql.org 2409 : 4 : stats->spaceType = SORT_SPACE_TYPE_DISK;
2410 : : else
2411 : 260 : stats->spaceType = SORT_SPACE_TYPE_MEMORY;
2380 tomas.vondra@postgre 2412 : 264 : stats->spaceUsed = (state->maxSpace + 1023) / 1024;
2413 : :
2414 [ + - + - ]: 264 : switch (state->maxSpaceStatus)
2415 : : {
7101 tgl@sss.pgh.pa.us 2416 : 260 : case TSS_SORTEDINMEM:
2417 [ + + ]: 260 : if (state->boundUsed)
3331 rhaas@postgresql.org 2418 : 28 : stats->sortMethod = SORT_TYPE_TOP_N_HEAPSORT;
2419 : : else
2420 : 232 : stats->sortMethod = SORT_TYPE_QUICKSORT;
7101 tgl@sss.pgh.pa.us 2421 : 260 : break;
7101 tgl@sss.pgh.pa.us 2422 :UBC 0 : case TSS_SORTEDONTAPE:
3331 rhaas@postgresql.org 2423 : 0 : stats->sortMethod = SORT_TYPE_EXTERNAL_SORT;
7101 tgl@sss.pgh.pa.us 2424 : 0 : break;
7101 tgl@sss.pgh.pa.us 2425 :CBC 4 : case TSS_FINALMERGE:
3331 rhaas@postgresql.org 2426 : 4 : stats->sortMethod = SORT_TYPE_EXTERNAL_MERGE;
7101 tgl@sss.pgh.pa.us 2427 : 4 : break;
7101 tgl@sss.pgh.pa.us 2428 :UBC 0 : default:
3331 rhaas@postgresql.org 2429 : 0 : stats->sortMethod = SORT_TYPE_STILL_IN_PROGRESS;
7101 tgl@sss.pgh.pa.us 2430 : 0 : break;
2431 : : }
7101 tgl@sss.pgh.pa.us 2432 :CBC 264 : }
2433 : :
2434 : : /*
2435 : : * Convert TuplesortMethod to a string.
2436 : : */
2437 : : const char *
3331 rhaas@postgresql.org 2438 : 196 : tuplesort_method_name(TuplesortMethod m)
2439 : : {
2440 [ - + + - : 196 : switch (m)
+ - ]
2441 : : {
3331 rhaas@postgresql.org 2442 :UBC 0 : case SORT_TYPE_STILL_IN_PROGRESS:
2443 : 0 : return "still in progress";
3331 rhaas@postgresql.org 2444 :CBC 28 : case SORT_TYPE_TOP_N_HEAPSORT:
2445 : 28 : return "top-N heapsort";
2446 : 164 : case SORT_TYPE_QUICKSORT:
2447 : 164 : return "quicksort";
3331 rhaas@postgresql.org 2448 :UBC 0 : case SORT_TYPE_EXTERNAL_SORT:
2449 : 0 : return "external sort";
3331 rhaas@postgresql.org 2450 :CBC 4 : case SORT_TYPE_EXTERNAL_MERGE:
2451 : 4 : return "external merge";
2452 : : }
2453 : :
3331 rhaas@postgresql.org 2454 :UBC 0 : return "unknown";
2455 : : }
2456 : :
2457 : : /*
2458 : : * Convert TuplesortSpaceType to a string.
2459 : : */
2460 : : const char *
3331 rhaas@postgresql.org 2461 :CBC 172 : tuplesort_space_type_name(TuplesortSpaceType t)
2462 : : {
2463 [ + + - + ]: 172 : Assert(t == SORT_SPACE_TYPE_DISK || t == SORT_SPACE_TYPE_MEMORY);
2464 [ + + ]: 172 : return t == SORT_SPACE_TYPE_DISK ? "Disk" : "Memory";
2465 : : }
2466 : :
2467 : :
2468 : : /*
2469 : : * Heap manipulation routines, per Knuth's Algorithm 5.2.3H.
2470 : : */
2471 : :
2472 : : /*
2473 : : * Convert the existing unordered array of SortTuples to a bounded heap,
2474 : : * discarding all but the smallest "state->bound" tuples.
2475 : : *
2476 : : * When working with a bounded heap, we want to keep the largest entry
2477 : : * at the root (array entry zero), instead of the smallest as in the normal
2478 : : * sort case. This allows us to discard the largest entry cheaply.
2479 : : * Therefore, we temporarily reverse the sort direction.
2480 : : */
2481 : : static void
7101 tgl@sss.pgh.pa.us 2482 : 264 : make_bounded_heap(Tuplesortstate *state)
2483 : : {
6906 bruce@momjian.us 2484 : 264 : int tupcount = state->memtupcount;
2485 : : int i;
2486 : :
7101 tgl@sss.pgh.pa.us 2487 [ - + ]: 264 : Assert(state->status == TSS_INITIAL);
2488 [ - + ]: 264 : Assert(state->bounded);
2489 [ - + ]: 264 : Assert(tupcount >= state->bound);
3174 rhaas@postgresql.org 2490 [ - + ]: 264 : Assert(SERIAL(state));
2491 : :
2492 : : /* Reverse sort direction so largest entry will be at root */
4357 2493 : 264 : reversedirection(state);
2494 : :
7101 tgl@sss.pgh.pa.us 2495 : 264 : state->memtupcount = 0; /* make the heap empty */
6906 bruce@momjian.us 2496 [ + + ]: 34308 : for (i = 0; i < tupcount; i++)
2497 : : {
3683 heikki.linnakangas@i 2498 [ + + ]: 34044 : if (state->memtupcount < state->bound)
2499 : : {
2500 : : /* Insert next tuple into heap */
2501 : : /* Must copy source tuple to avoid possible overwrite */
6906 bruce@momjian.us 2502 : 16890 : SortTuple stup = state->memtuples[i];
2503 : :
3300 rhaas@postgresql.org 2504 : 16890 : tuplesort_heap_insert(state, &stup);
2505 : : }
2506 : : else
2507 : : {
2508 : : /*
2509 : : * The heap is full. Replace the largest entry with the new
2510 : : * tuple, or just discard it, if it's larger than anything already
2511 : : * in the heap.
2512 : : */
3683 heikki.linnakangas@i 2513 [ + + ]: 17154 : if (COMPARETUP(state, &state->memtuples[i], &state->memtuples[0]) <= 0)
2514 : : {
2515 : 7623 : free_sort_tuple(state, &state->memtuples[i]);
2516 [ - + ]: 7623 : CHECK_FOR_INTERRUPTS();
2517 : : }
2518 : : else
3300 rhaas@postgresql.org 2519 : 9531 : tuplesort_heap_replace_top(state, &state->memtuples[i]);
2520 : : }
2521 : : }
2522 : :
7101 tgl@sss.pgh.pa.us 2523 [ - + ]: 264 : Assert(state->memtupcount == state->bound);
2524 : 264 : state->status = TSS_BOUNDED;
2525 : 264 : }
2526 : :
2527 : : /*
2528 : : * Convert the bounded heap to a properly-sorted array
2529 : : */
2530 : : static void
2531 : 264 : sort_bounded_heap(Tuplesortstate *state)
2532 : : {
6906 bruce@momjian.us 2533 : 264 : int tupcount = state->memtupcount;
2534 : :
7101 tgl@sss.pgh.pa.us 2535 [ - + ]: 264 : Assert(state->status == TSS_BOUNDED);
2536 [ - + ]: 264 : Assert(state->bounded);
2537 [ - + ]: 264 : Assert(tupcount == state->bound);
3174 rhaas@postgresql.org 2538 [ - + ]: 264 : Assert(SERIAL(state));
2539 : :
2540 : : /*
2541 : : * We can unheapify in place because each delete-top call will remove the
2542 : : * largest entry, which we can promptly store in the newly freed slot at
2543 : : * the end. Once we're down to a single-entry heap, we're done.
2544 : : */
7101 tgl@sss.pgh.pa.us 2545 [ + + ]: 16890 : while (state->memtupcount > 1)
2546 : : {
6906 bruce@momjian.us 2547 : 16626 : SortTuple stup = state->memtuples[0];
2548 : :
2549 : : /* this sifts-up the next-largest entry and decreases memtupcount */
3300 rhaas@postgresql.org 2550 : 16626 : tuplesort_heap_delete_top(state);
7101 tgl@sss.pgh.pa.us 2551 : 16626 : state->memtuples[state->memtupcount] = stup;
2552 : : }
2553 : 264 : state->memtupcount = tupcount;
2554 : :
2555 : : /*
2556 : : * Reverse sort direction back to the original state. This is not
2557 : : * actually necessary but seems like a good idea for tidiness.
2558 : : */
4357 rhaas@postgresql.org 2559 : 264 : reversedirection(state);
2560 : :
7101 tgl@sss.pgh.pa.us 2561 : 264 : state->status = TSS_SORTEDINMEM;
2562 : 264 : state->boundUsed = true;
2563 : 264 : }
2564 : :
2565 : :
2566 : : /* radix sort routines */
2567 : :
2568 : : /*
2569 : : * Retrieve byte from datum, indexed by 'level': 0 for MSB, 7 for LSB
2570 : : */
2571 : : static inline uint8
240 john.naylor@postgres 2572 : 26428626 : current_byte(Datum key, int level)
2573 : : {
2574 : 26428626 : int shift = (sizeof(Datum) - 1 - level) * BITS_PER_BYTE;
2575 : :
2576 : 26428626 : return (key >> shift) & 0xFF;
2577 : : }
2578 : :
2579 : : /*
2580 : : * Normalize datum such that unsigned comparison is order-preserving,
2581 : : * taking ASC/DESC into account as well.
2582 : : */
2583 : : static inline Datum
2584 : 26476043 : normalize_datum(Datum orig, SortSupport ssup)
2585 : : {
2586 : : Datum norm_datum1;
2587 : :
63 john.naylor@postgres 2588 [ + + ]:GNC 26476043 : if (ssup->comparator == ssup_datum_int64_cmp)
166 john.naylor@postgres 2589 :CBC 1110020 : norm_datum1 = orig + (Int64GetDatum(PG_INT64_MAX)) + 1;
63 john.naylor@postgres 2590 [ + + ]:GNC 25366023 : else if (ssup->comparator == ssup_datum_uint64_cmp)
2591 : 6959343 : norm_datum1 = orig;
2592 : : else
2593 : : {
2594 : : /*
2595 : : * Truncate to uint32. For the int32 case, we don't need to do this,
2596 : : * but it forces the upper half of the datum to be zero regardless of
2597 : : * sign.
2598 : : */
2599 : 18406680 : uint32 u32 = DatumGetUInt32(orig);
2600 : :
2601 [ + + ]: 18406680 : if (ssup->comparator == ssup_datum_int32_cmp)
2602 : 7799114 : norm_datum1 = UInt32GetDatum(u32 + ((uint32) PG_INT32_MAX) + 1);
2603 : : else
2604 : : {
2605 : 10607566 : norm_datum1 = UInt32GetDatum(u32);
2606 [ - + ]: 10607566 : Assert(ssup->comparator == ssup_datum_uint32_cmp);
2607 : : }
2608 : : }
2609 : :
240 john.naylor@postgres 2610 [ + + ]:CBC 26476043 : if (ssup->ssup_reverse)
2611 : 787998 : norm_datum1 = ~norm_datum1;
2612 : :
2613 : 26476043 : return norm_datum1;
2614 : : }
2615 : :
2616 : : /*
2617 : : * radix_sort_recursive
2618 : : *
2619 : : * Radix sort by (pass-by-value) datum1, diverting to qsort_tuple()
2620 : : * for tiebreaks.
2621 : : *
2622 : : * This is a modification of ska_byte_sort() from
2623 : : * https://github.com/skarupke/ska_sort
2624 : : * The original copyright notice follows:
2625 : : *
2626 : : * Copyright Malte Skarupke 2016.
2627 : : * Distributed under the Boost Software License, Version 1.0.
2628 : : *
2629 : : * Boost Software License - Version 1.0 - August 17th, 2003
2630 : : *
2631 : : * Permission is hereby granted, free of charge, to any person or organization
2632 : : * obtaining a copy of the software and accompanying documentation covered by
2633 : : * this license (the "Software") to use, reproduce, display, distribute,
2634 : : * execute, and transmit the Software, and to prepare derivative works of the
2635 : : * Software, and to permit third-parties to whom the Software is furnished to
2636 : : * do so, all subject to the following:
2637 : : *
2638 : : * The copyright notices in the Software and this entire statement, including
2639 : : * the above license grant, this restriction and the following disclaimer,
2640 : : * must be included in all copies of the Software, in whole or in part, and
2641 : : * all derivative works of the Software, unless such copies or derivative
2642 : : * works are solely in the form of machine-executable object code generated by
2643 : : * a source language processor.
2644 : : *
2645 : : * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
2646 : : * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
2647 : : * FITNESS FOR A PARTICULAR PURPOSE, TITLE AND NON-INFRINGEMENT. IN NO EVENT
2648 : : * SHALL THE COPYRIGHT HOLDERS OR ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE
2649 : : * FOR ANY DAMAGES OR OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE,
2650 : : * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
2651 : : * DEALINGS IN THE SOFTWARE.
2652 : : */
2653 : : static void
2654 : 47417 : radix_sort_recursive(SortTuple *begin, size_t n_elems, int level, Tuplesortstate *state)
2655 : : {
2656 : 47417 : RadixSortInfo partitions[256] = {0};
2657 : : uint8 remaining_partitions[256];
2658 : 47417 : size_t total = 0;
2659 : 47417 : int num_partitions = 0;
2660 : : int num_remaining;
2661 : 47417 : SortSupport ssup = &state->base.sortKeys[0];
2662 : : Datum ref_datum;
194 2663 : 47417 : Datum common_upper_bits = 0;
240 2664 : 47417 : size_t start_offset = 0;
2665 : 47417 : SortTuple *partition_begin = begin;
2666 : : int next_level;
2667 : :
2668 : : /* count number of occurrences of each byte */
194 2669 : 47417 : ref_datum = normalize_datum(begin[0].datum1, ssup);
240 2670 [ + + ]: 26476043 : for (SortTuple *st = begin; st < begin + n_elems; st++)
2671 : : {
2672 : : Datum this_datum;
2673 : : uint8 this_partition;
2674 : :
194 2675 : 26428626 : this_datum = normalize_datum(st->datum1, ssup);
2676 : : /* accumulate bits different from the reference datum */
2677 : 26428626 : common_upper_bits |= ref_datum ^ this_datum;
2678 : :
2679 : : /* extract the byte for this level from the normalized datum */
2680 : 26428626 : this_partition = current_byte(this_datum, level);
2681 : :
2682 : : /* save it for the permutation step */
240 2683 : 26428626 : st->curbyte = this_partition;
2684 : :
2685 : 26428626 : partitions[this_partition].count++;
2686 : :
2687 [ + + ]: 26428626 : CHECK_FOR_INTERRUPTS();
2688 : : }
2689 : :
2690 : : /* compute partition offsets */
2691 [ + + ]: 12186169 : for (int i = 0; i < 256; i++)
2692 : : {
2693 : 12138752 : size_t count = partitions[i].count;
2694 : :
2695 [ + + ]: 12138752 : if (count != 0)
2696 : : {
2697 : 2639308 : partitions[i].offset = total;
2698 : 2639308 : total += count;
2699 : 2639308 : remaining_partitions[num_partitions] = i;
2700 : 2639308 : num_partitions++;
2701 : : }
2702 : 12138752 : partitions[i].next_offset = total;
2703 : : }
2704 : :
2705 : : /*
2706 : : * Swap tuples to correct partition.
2707 : : *
2708 : : * In traditional American flag sort, a swap sends the current element to
2709 : : * the correct partition, but the array pointer only advances if the
2710 : : * partner of the swap happens to be an element that belongs in the
2711 : : * current partition. That only requires one pass through the array, but
2712 : : * the disadvantage is we don't know if the pointer can advance until the
2713 : : * swap completes. Here lies the most interesting innovation from the
2714 : : * upstream ska_byte_sort: After initiating the swap, we immediately
2715 : : * proceed to the next element. This makes better use of CPU pipelining,
2716 : : * but also means that we will often need multiple iterations of this
2717 : : * loop. ska_byte_sort() maintains a separate list of which partitions
2718 : : * haven't finished, which is updated every loop iteration. Here we simply
2719 : : * check each partition during every iteration.
2720 : : *
2721 : : * If we started with a single partition, there is nothing to do. If a
2722 : : * previous loop iteration results in only one partition that hasn't been
2723 : : * counted as sorted, we know it's actually sorted and can exit the loop.
2724 : : */
2725 : 47417 : num_remaining = num_partitions;
2726 [ + + ]: 195781 : while (num_remaining > 1)
2727 : : {
2728 : : /* start the count over */
2729 : 148364 : num_remaining = num_partitions;
2730 : :
2731 [ + + ]: 11721862 : for (int i = 0; i < num_partitions; i++)
2732 : : {
2733 : 11573498 : uint8 idx = remaining_partitions[i];
2734 : :
2735 : 11573498 : for (SortTuple *st = begin + partitions[idx].offset;
2736 [ + + ]: 28699642 : st < begin + partitions[idx].next_offset;
2737 : 17126144 : st++)
2738 : : {
2739 : 17126144 : size_t offset = partitions[st->curbyte].offset++;
2740 : : SortTuple tmp;
2741 : :
2742 : : /* swap current tuple with destination position */
2743 [ - + ]: 17126144 : Assert(offset < n_elems);
2744 : 17126144 : tmp = *st;
2745 : 17126144 : *st = begin[offset];
2746 : 17126144 : begin[offset] = tmp;
2747 : :
2748 [ + + ]: 17126144 : CHECK_FOR_INTERRUPTS();
2749 : : };
2750 : :
2751 : : /* Is this partition sorted? */
2752 [ + + ]: 11573498 : if (partitions[idx].offset == partitions[idx].next_offset)
2753 : 8935222 : num_remaining--;
2754 : : }
2755 : : }
2756 : :
2757 : : /* recurse */
2758 : :
194 2759 [ + + ]: 47417 : if (num_partitions == 1)
2760 : : {
2761 : : /*
2762 : : * There is only one distinct byte at the current level. It can happen
2763 : : * that some subsequent bytes are also the same for all input values,
2764 : : * such as the upper bytes of small integers. To skip unproductive
2765 : : * passes for that case, we compute the level where the input has more
2766 : : * than one distinct byte, so that the next recursion can start there.
2767 : : */
2768 [ + + ]: 9314 : if (common_upper_bits == 0)
2769 : 999 : next_level = sizeof(Datum);
2770 : : else
2771 : : {
2772 : : int diffpos;
2773 : :
2774 : : /*
2775 : : * The upper bits of common_upper_bits are zero where all datums
2776 : : * have the same bits.
2777 : : */
2778 : 8315 : diffpos = pg_leftmost_one_pos64(DatumGetUInt64(common_upper_bits));
2779 : 8315 : next_level = sizeof(Datum) - 1 - (diffpos / BITS_PER_BYTE);
2780 : : }
2781 : : }
2782 : : else
2783 : 38103 : next_level = level + 1;
2784 : :
166 2785 [ - + ]: 47417 : Assert(next_level > level);
2786 : :
240 2787 : 47417 : for (uint8 *rp = remaining_partitions;
2788 [ + + ]: 2686725 : rp < remaining_partitions + num_partitions;
2789 : 2639308 : rp++)
2790 : : {
2791 : 2639308 : size_t end_offset = partitions[*rp].next_offset;
2792 : 2639308 : SortTuple *partition_end = begin + end_offset;
2793 : 2639308 : size_t num_elements = end_offset - start_offset;
2794 : :
2795 [ + + ]: 2639308 : if (num_elements > 1)
2796 : : {
194 2797 [ + + ]: 870089 : if (next_level < sizeof(Datum))
2798 : : {
240 2799 [ + + ]: 682003 : if (num_elements < QSORT_THRESHOLD)
2800 : : {
2801 : 640078 : qsort_tuple(partition_begin,
2802 : : num_elements,
2803 : : state->base.comparetup,
2804 : : state);
2805 : : }
2806 : : else
2807 : : {
2808 : 41925 : radix_sort_recursive(partition_begin,
2809 : : num_elements,
2810 : : next_level,
2811 : : state);
2812 : : }
2813 : : }
2814 [ + + ]: 188086 : else if (state->base.onlyKey == NULL)
2815 : : {
2816 : : /*
2817 : : * We've finished radix sort on all bytes of the pass-by-value
2818 : : * datum (possibly abbreviated), now sort using the tiebreak
2819 : : * comparator.
2820 : : */
2821 : 92639 : qsort_tuple(partition_begin,
2822 : : num_elements,
2823 : : state->base.comparetup_tiebreak,
2824 : : state);
2825 : : }
2826 : : }
2827 : :
2828 : 2639308 : start_offset = end_offset;
2829 : 2639308 : partition_begin = partition_end;
2830 : : }
2831 : 47417 : }
2832 : :
2833 : : /*
2834 : : * Entry point for radix_sort_recursive
2835 : : *
2836 : : * Partition tuples by isnull1, then sort both partitions, using
2837 : : * radix sort on the NOT NULL partition if it's large enough.
2838 : : */
2839 : : static void
2840 : 7472 : radix_sort_tuple(SortTuple *data, size_t n, Tuplesortstate *state)
2841 : : {
2842 : 7472 : bool nulls_first = state->base.sortKeys[0].ssup_nulls_first;
2843 : : SortTuple *null_start;
2844 : : SortTuple *not_null_start;
2845 : 7472 : size_t d1 = 0,
2846 : : d2,
2847 : : null_count,
2848 : : not_null_count;
2849 : :
2850 : : /*
2851 : : * Find the first NOT NULL if NULLS FIRST, or first NULL if NULLS LAST.
2852 : : * This also serves as a quick check for the common case where all tuples
2853 : : * are NOT NULL in the first sort key with the default order ASC NULLS
2854 : : * LAST.
2855 : : */
2856 [ + + + + ]: 13317230 : while (d1 < n && data[d1].isnull1 == nulls_first)
2857 : : {
2858 : 13309758 : d1++;
2859 [ + + ]: 13309758 : CHECK_FOR_INTERRUPTS();
2860 : : }
2861 : :
2862 : : /*
2863 : : * If we have more than one tuple left after the quick check, partition
2864 : : * the remainder using branchless cyclic permutation, based on
2865 : : * https://orlp.net/blog/branchless-lomuto-partitioning/
2866 : : */
2867 [ - + ]: 7472 : Assert(n > 0);
2868 [ + + ]: 7472 : if (d1 < n - 1)
2869 : : {
2870 : 733 : size_t i = d1,
2871 : 733 : j = d1;
2872 : 733 : SortTuple tmp = data[d1]; /* create gap at front */
2873 : :
2874 [ + + ]: 433720 : while (j < n - 1)
2875 : : {
2876 : : /* gap is at j, move i's element to gap */
2877 : 432987 : data[j] = data[i];
2878 : : /* advance j to the first unknown element */
2879 : 432987 : j += 1;
2880 : : /* move the first unknown element back to i */
2881 : 432987 : data[i] = data[j];
2882 : : /* advance i if this element belongs in the left partition */
2883 : 432987 : i += (data[i].isnull1 == nulls_first);
2884 : :
2885 [ - + ]: 432987 : CHECK_FOR_INTERRUPTS();
2886 : : }
2887 : :
2888 : : /* place gap between left and right partitions */
2889 : 733 : data[j] = data[i];
2890 : : /* restore the saved element */
2891 : 733 : data[i] = tmp;
2892 : : /* assign it to the correct partition */
2893 : 733 : i += (data[i].isnull1 == nulls_first);
2894 : :
2895 : : /* d1 is now the number of elements in the left partition */
2896 : 733 : d1 = i;
2897 : : }
2898 : :
2899 : 7472 : d2 = n - d1;
2900 : :
2901 : : /* set pointers and counts for each partition */
2902 [ + + ]: 7472 : if (nulls_first)
2903 : : {
2904 : 628 : null_start = data;
2905 : 628 : null_count = d1;
2906 : 628 : not_null_start = data + d1;
2907 : 628 : not_null_count = d2;
2908 : : }
2909 : : else
2910 : : {
2911 : 6844 : not_null_start = data;
2912 : 6844 : not_null_count = d1;
2913 : 6844 : null_start = data + d1;
2914 : 6844 : null_count = d2;
2915 : : }
2916 : :
2917 : 7472 : for (SortTuple *st = null_start;
2918 [ + + ]: 9759 : st < null_start + null_count;
2919 : 2287 : st++)
2920 [ - + ]: 2287 : Assert(st->isnull1 == true);
2921 : 7472 : for (SortTuple *st = not_null_start;
2922 [ + + ]: 13748667 : st < not_null_start + not_null_count;
2923 : 13741195 : st++)
2924 [ - + ]: 13741195 : Assert(st->isnull1 == false);
2925 : :
2926 : : /*
2927 : : * Sort the NULL partition using tiebreak comparator, if necessary.
2928 : : */
2929 [ + + + + ]: 7472 : if (state->base.onlyKey == NULL && null_count > 1)
2930 : : {
2931 : 111 : qsort_tuple(null_start,
2932 : : null_count,
2933 : : state->base.comparetup_tiebreak,
2934 : : state);
2935 : : }
2936 : :
2937 : : /*
2938 : : * Sort the NOT NULL partition, using radix sort if large enough,
2939 : : * otherwise fall back to quicksort.
2940 : : */
2941 [ + + ]: 7472 : if (not_null_count < QSORT_THRESHOLD)
2942 : : {
2943 : 14 : qsort_tuple(not_null_start,
2944 : : not_null_count,
2945 : : state->base.comparetup,
2946 : : state);
2947 : : }
2948 : : else
2949 : : {
2950 : 7458 : bool presorted = true;
2951 : :
2952 : 7458 : for (SortTuple *st = not_null_start + 1;
2953 [ + + ]: 5076494 : st < not_null_start + not_null_count;
2954 : 5069036 : st++)
2955 : : {
2956 [ + + ]: 5074528 : if (COMPARETUP(state, st - 1, st) > 0)
2957 : : {
2958 : 5492 : presorted = false;
2959 : 5492 : break;
2960 : : }
2961 : :
2962 [ - + ]: 5069036 : CHECK_FOR_INTERRUPTS();
2963 : : }
2964 : :
2965 [ + + ]: 7458 : if (presorted)
2966 : 1966 : return;
2967 : : else
2968 : : {
2969 : 5492 : radix_sort_recursive(not_null_start,
2970 : : not_null_count,
2971 : : 0,
2972 : : state);
2973 : : }
2974 : : }
2975 : : }
2976 : :
2977 : : /* Verify in-memory sort using standard comparator. */
2978 : : static void
2979 : 7472 : verify_memtuples_sorted(Tuplesortstate *state)
2980 : : {
2981 : : #ifdef USE_ASSERT_CHECKING
2982 : 7472 : for (SortTuple *st = state->memtuples + 1;
2983 [ + + ]: 13743482 : st < state->memtuples + state->memtupcount;
2984 : 13736010 : st++)
2985 [ - + ]: 13736010 : Assert(COMPARETUP(state, st - 1, st) <= 0);
2986 : : #endif
2987 : 7472 : }
2988 : :
2989 : : /*
2990 : : * Sort all memtuples using specialized routines.
2991 : : *
2992 : : * Quicksort or radix sort is used for small in-memory sorts,
2993 : : * and external sort runs.
2994 : : */
2995 : : static void
3839 rhaas@postgresql.org 2996 : 147414 : tuplesort_sort_memtuples(Tuplesortstate *state)
2997 : : {
3174 2998 [ + + - + ]: 147414 : Assert(!LEADER(state));
2999 : :
3839 3000 [ + + ]: 147414 : if (state->memtupcount > 1)
3001 : : {
3002 : : /*
3003 : : * Do we have the leading column's value or abbreviation in datum1?
3004 : : */
1538 akorotkov@postgresql 3005 [ + + + + ]: 44271 : if (state->base.haveDatum1 && state->base.sortKeys)
3006 : : {
240 john.naylor@postgres 3007 : 44182 : SortSupport ssup = &state->base.sortKeys[0];
3008 : :
3009 : : /* Does it compare as an integer? */
3010 [ + + ]: 44182 : if (state->memtupcount >= QSORT_THRESHOLD &&
63 john.naylor@postgres 3011 [ + + ]:GNC 8585 : (ssup->comparator == ssup_datum_uint64_cmp ||
3012 [ + + ]: 8018 : ssup->comparator == ssup_datum_int64_cmp ||
3013 [ + + ]: 7831 : ssup->comparator == ssup_datum_uint32_cmp ||
240 john.naylor@postgres 3014 [ + + ]:CBC 4289 : ssup->comparator == ssup_datum_int32_cmp))
3015 : : {
3016 : 7472 : radix_sort_tuple(state->memtuples,
3017 : 7472 : state->memtupcount,
3018 : : state);
3019 : 7472 : verify_memtuples_sorted(state);
1652 tmunro@postgresql.or 3020 : 7472 : return;
3021 : : }
3022 : : }
3023 : :
3024 : : /* Can we use the single-key sort function? */
1538 akorotkov@postgresql 3025 [ + + ]: 36799 : if (state->base.onlyKey != NULL)
3026 : : {
3839 rhaas@postgresql.org 3027 : 26415 : qsort_ssup(state->memtuples, state->memtupcount,
1538 akorotkov@postgresql 3028 : 26415 : state->base.onlyKey);
3029 : : }
3030 : : else
3031 : : {
3839 rhaas@postgresql.org 3032 : 10384 : qsort_tuple(state->memtuples,
3033 : 10384 : state->memtupcount,
3034 : : state->base.comparetup,
3035 : : state);
3036 : : }
3037 : : }
3038 : : }
3039 : :
3040 : : /*
3041 : : * Insert a new tuple into an empty or existing heap, maintaining the
3042 : : * heap invariant. Caller is responsible for ensuring there's room.
3043 : : *
3044 : : * Note: For some callers, tuple points to a memtuples[] entry above the
3045 : : * end of the heap. This is safe as long as it's not immediately adjacent
3046 : : * to the end of the heap (ie, in the [memtupcount] array entry) --- if it
3047 : : * is, it might get overwritten before being moved into the heap!
3048 : : */
3049 : : static void
3300 3050 : 17889 : tuplesort_heap_insert(Tuplesortstate *state, SortTuple *tuple)
3051 : : {
3052 : : SortTuple *memtuples;
3053 : : int j;
3054 : :
9844 tgl@sss.pgh.pa.us 3055 : 17889 : memtuples = state->memtuples;
7533 3056 [ - + ]: 17889 : Assert(state->memtupcount < state->memtupsize);
3057 : :
5319 rhaas@postgresql.org 3058 [ + + ]: 17889 : CHECK_FOR_INTERRUPTS();
3059 : :
3060 : : /*
3061 : : * Sift-up the new entry, per Knuth 5.2.3 exercise 16. Note that Knuth is
3062 : : * using 1-based array indexes, not 0-based.
3063 : : */
9844 tgl@sss.pgh.pa.us 3064 : 17889 : j = state->memtupcount++;
3065 [ + + ]: 47677 : while (j > 0)
3066 : : {
9679 bruce@momjian.us 3067 : 43438 : int i = (j - 1) >> 1;
3068 : :
3300 rhaas@postgresql.org 3069 [ + + ]: 43438 : if (COMPARETUP(state, tuple, &memtuples[i]) >= 0)
9857 tgl@sss.pgh.pa.us 3070 : 13650 : break;
9844 3071 : 29788 : memtuples[j] = memtuples[i];
9857 3072 : 29788 : j = i;
3073 : : }
7533 3074 : 17889 : memtuples[j] = *tuple;
9857 3075 : 17889 : }
3076 : :
3077 : : /*
3078 : : * Remove the tuple at state->memtuples[0] from the heap. Decrement
3079 : : * memtupcount, and sift up to maintain the heap invariant.
3080 : : *
3081 : : * The caller has already free'd the tuple the top node points to,
3082 : : * if necessary.
3083 : : */
3084 : : static void
3300 rhaas@postgresql.org 3085 : 17605 : tuplesort_heap_delete_top(Tuplesortstate *state)
3086 : : {
7533 tgl@sss.pgh.pa.us 3087 : 17605 : SortTuple *memtuples = state->memtuples;
3088 : : SortTuple *tuple;
3089 : :
9844 3090 [ + + ]: 17605 : if (--state->memtupcount <= 0)
9857 3091 : 257 : return;
3092 : :
3093 : : /*
3094 : : * Remove the last tuple in the heap, and re-insert it, by replacing the
3095 : : * current top node with it.
3096 : : */
3683 heikki.linnakangas@i 3097 : 17348 : tuple = &memtuples[state->memtupcount];
3300 rhaas@postgresql.org 3098 : 17348 : tuplesort_heap_replace_top(state, tuple);
3099 : : }
3100 : :
3101 : : /*
3102 : : * Replace the tuple at state->memtuples[0] with a new tuple. Sift up to
3103 : : * maintain the heap invariant.
3104 : : *
3105 : : * This corresponds to Knuth's "sift-up" algorithm (Algorithm 5.2.3H,
3106 : : * Heapsort, steps H3-H8).
3107 : : */
3108 : : static void
3109 : 4087433 : tuplesort_heap_replace_top(Tuplesortstate *state, SortTuple *tuple)
3110 : : {
3683 heikki.linnakangas@i 3111 : 4087433 : SortTuple *memtuples = state->memtuples;
3112 : : unsigned int i,
3113 : : n;
3114 : :
3115 [ - + ]: 4087433 : Assert(state->memtupcount >= 1);
3116 : :
5319 rhaas@postgresql.org 3117 [ + + ]: 4087433 : CHECK_FOR_INTERRUPTS();
3118 : :
3119 : : /*
3120 : : * state->memtupcount is "int", but we use "unsigned int" for i, j, n.
3121 : : * This prevents overflow in the "2 * i + 1" calculation, since at the top
3122 : : * of the loop we must have i < n <= INT_MAX <= UINT_MAX/2.
3123 : : */
9844 tgl@sss.pgh.pa.us 3124 : 4087433 : n = state->memtupcount;
9857 3125 : 4087433 : i = 0; /* i is where the "hole" is */
3126 : : for (;;)
9844 3127 : 1326975 : {
3379 3128 : 5414408 : unsigned int j = 2 * i + 1;
3129 : :
9857 3130 [ + + ]: 5414408 : if (j >= n)
3131 : 969317 : break;
9679 bruce@momjian.us 3132 [ + + + + ]: 6084919 : if (j + 1 < n &&
3300 rhaas@postgresql.org 3133 : 1639828 : COMPARETUP(state, &memtuples[j], &memtuples[j + 1]) > 0)
9857 tgl@sss.pgh.pa.us 3134 : 663613 : j++;
3300 rhaas@postgresql.org 3135 [ + + ]: 4445091 : if (COMPARETUP(state, tuple, &memtuples[j]) <= 0)
9857 tgl@sss.pgh.pa.us 3136 : 3118116 : break;
9844 3137 : 1326975 : memtuples[i] = memtuples[j];
9857 3138 : 1326975 : i = j;
3139 : : }
7533 3140 : 4087433 : memtuples[i] = *tuple;
9857 3141 : 4087433 : }
3142 : :
3143 : : /*
3144 : : * Function to reverse the sort direction from its current state
3145 : : *
3146 : : * It is not safe to call this when performing hash tuplesorts
3147 : : */
3148 : : static void
4357 rhaas@postgresql.org 3149 : 528 : reversedirection(Tuplesortstate *state)
3150 : : {
1538 akorotkov@postgresql 3151 : 528 : SortSupport sortKey = state->base.sortKeys;
3152 : : int nkey;
3153 : :
3154 [ + + ]: 1280 : for (nkey = 0; nkey < state->base.nKeys; nkey++, sortKey++)
3155 : : {
4357 rhaas@postgresql.org 3156 : 752 : sortKey->ssup_reverse = !sortKey->ssup_reverse;
3157 : 752 : sortKey->ssup_nulls_first = !sortKey->ssup_nulls_first;
3158 : : }
3159 : 528 : }
3160 : :
3161 : :
3162 : : /*
3163 : : * Tape interface routines
3164 : : */
3165 : :
3166 : : static unsigned int
1820 heikki.linnakangas@i 3167 : 3925034 : getlen(LogicalTape *tape, bool eofOK)
3168 : : {
3169 : : unsigned int len;
3170 : :
3171 [ - + ]: 3925034 : if (LogicalTapeRead(tape,
3172 : : &len, sizeof(len)) != sizeof(len))
8480 tgl@sss.pgh.pa.us 3173 [ # # ]:UBC 0 : elog(ERROR, "unexpected end of tape");
9857 tgl@sss.pgh.pa.us 3174 [ + + - + ]:CBC 3925034 : if (len == 0 && !eofOK)
8480 tgl@sss.pgh.pa.us 3175 [ # # ]:UBC 0 : elog(ERROR, "unexpected end of data");
9857 tgl@sss.pgh.pa.us 3176 :CBC 3925034 : return len;
3177 : : }
3178 : :
3179 : : static void
1820 heikki.linnakangas@i 3180 : 1304 : markrunend(LogicalTape *tape)
3181 : : {
9679 bruce@momjian.us 3182 : 1304 : unsigned int len = 0;
3183 : :
1382 peter@eisentraut.org 3184 : 1304 : LogicalTapeWrite(tape, &len, sizeof(len));
9857 tgl@sss.pgh.pa.us 3185 : 1304 : }
3186 : :
3187 : : /*
3188 : : * Get memory for tuple from within READTUP() routine.
3189 : : *
3190 : : * We use next free slot from the slab allocator, or palloc() if the tuple
3191 : : * is too large for that.
3192 : : */
3193 : : void *
1538 akorotkov@postgresql 3194 : 3668645 : tuplesort_readtup_alloc(Tuplesortstate *state, Size tuplen)
3195 : : {
3196 : : SlabSlot *buf;
3197 : :
3198 : : /*
3199 : : * We pre-allocate enough slots in the slab arena that we should never run
3200 : : * out.
3201 : : */
3661 heikki.linnakangas@i 3202 [ - + ]: 3668645 : Assert(state->slabFreeHead);
3203 : :
3204 [ + + - + ]: 3668645 : if (tuplen > SLAB_SLOT_SIZE || !state->slabFreeHead)
1538 akorotkov@postgresql 3205 : 4 : return MemoryContextAlloc(state->base.sortcontext, tuplen);
3206 : : else
3207 : : {
3661 heikki.linnakangas@i 3208 : 3668641 : buf = state->slabFreeHead;
3209 : : /* Reuse this slot */
3210 : 3668641 : state->slabFreeHead = buf->nextfree;
3211 : :
3212 : 3668641 : return buf;
3213 : : }
3214 : : }
3215 : :
3216 : :
3217 : : /*
3218 : : * Parallel sort routines
3219 : : */
3220 : :
3221 : : /*
3222 : : * tuplesort_estimate_shared - estimate required shared memory allocation
3223 : : *
3224 : : * nWorkers is an estimate of the number of workers (it's the number that
3225 : : * will be requested).
3226 : : */
3227 : : Size
3174 rhaas@postgresql.org 3228 : 131 : tuplesort_estimate_shared(int nWorkers)
3229 : : {
3230 : : Size tapesSize;
3231 : :
3232 [ - + ]: 131 : Assert(nWorkers > 0);
3233 : :
3234 : : /* Make sure that BufFile shared state is MAXALIGN'd */
3235 : 131 : tapesSize = mul_size(sizeof(TapeShare), nWorkers);
3236 : 131 : tapesSize = MAXALIGN(add_size(tapesSize, offsetof(Sharedsort, tapes)));
3237 : :
3238 : 131 : return tapesSize;
3239 : : }
3240 : :
3241 : : /*
3242 : : * tuplesort_initialize_shared - initialize shared tuplesort state
3243 : : *
3244 : : * Must be called from leader process before workers are launched, to
3245 : : * establish state needed up-front for worker tuplesortstates. nWorkers
3246 : : * should match the argument passed to tuplesort_estimate_shared().
3247 : : */
3248 : : void
3249 : 179 : tuplesort_initialize_shared(Sharedsort *shared, int nWorkers, dsm_segment *seg)
3250 : : {
3251 : : int i;
3252 : :
3253 [ - + ]: 179 : Assert(nWorkers > 0);
3254 : :
20 nathan@postgresql.or 3255 :GNC 179 : pg_atomic_init_u32(&shared->currentWorker, 0);
3256 : 179 : pg_atomic_init_u32(&shared->workersFinished, 0);
3174 rhaas@postgresql.org 3257 :CBC 179 : SharedFileSetInit(&shared->fileset, seg);
3258 : 179 : shared->nTapes = nWorkers;
3259 [ + + ]: 562 : for (i = 0; i < nWorkers; i++)
3260 : : {
3261 : 383 : shared->tapes[i].firstblocknumber = 0L;
3262 : : }
3263 : 179 : }
3264 : :
3265 : : /*
3266 : : * tuplesort_attach_shared - attach to shared tuplesort state
3267 : : *
3268 : : * Must be called by all worker processes.
3269 : : */
3270 : : void
3271 : 201 : tuplesort_attach_shared(Sharedsort *shared, dsm_segment *seg)
3272 : : {
3273 : : /* Attach to SharedFileSet */
3274 : 201 : SharedFileSetAttach(&shared->fileset, seg);
3275 : 201 : }
3276 : :
3277 : : /*
3278 : : * worker_get_identifier - Assign and return ordinal identifier for worker
3279 : : *
3280 : : * The order in which these are assigned is not well defined, and should not
3281 : : * matter; worker numbers across parallel sort participants need only be
3282 : : * distinct and gapless. logtape.c requires this.
3283 : : *
3284 : : * Note that the identifiers assigned from here have no relation to
3285 : : * ParallelWorkerNumber number, to avoid making any assumption about
3286 : : * caller's requirements. However, we do follow the ParallelWorkerNumber
3287 : : * convention of representing a non-worker with worker number -1. This
3288 : : * includes the leader, as well as serial Tuplesort processes.
3289 : : */
3290 : : static int
3291 : 379 : worker_get_identifier(Tuplesortstate *state)
3292 : : {
3293 [ + - - + ]: 379 : Assert(WORKER(state));
3294 : :
20 nathan@postgresql.or 3295 :GNC 379 : return pg_atomic_fetch_add_u32(&state->shared->currentWorker, 1);
3296 : : }
3297 : :
3298 : : /*
3299 : : * worker_freeze_result_tape - freeze worker's result tape for leader
3300 : : *
3301 : : * This is called by workers just after the result tape has been determined,
3302 : : * instead of calling LogicalTapeFreeze() directly. They do so because
3303 : : * workers require a few additional steps over similar serial
3304 : : * TSS_SORTEDONTAPE external sort cases, which also happen here. The extra
3305 : : * steps are around freeing now unneeded resources, and representing to
3306 : : * leader that worker's input run is available for its merge.
3307 : : *
3308 : : * There should only be one final output run for each worker, which consists
3309 : : * of all tuples that were originally input into worker.
3310 : : */
3311 : : static void
3174 rhaas@postgresql.org 3312 :CBC 379 : worker_freeze_result_tape(Tuplesortstate *state)
3313 : : {
3314 : 379 : Sharedsort *shared = state->shared;
3315 : : TapeShare output;
3316 : :
3317 [ + - - + ]: 379 : Assert(WORKER(state));
1820 heikki.linnakangas@i 3318 [ - + ]: 379 : Assert(state->result_tape != NULL);
3174 rhaas@postgresql.org 3319 [ - + ]: 379 : Assert(state->memtupcount == 0);
3320 : :
3321 : : /*
3322 : : * Free most remaining memory, in case caller is sensitive to our holding
3323 : : * on to it. memtuples may not be a tiny merge heap at this point.
3324 : : */
3325 : 379 : pfree(state->memtuples);
3326 : : /* Be tidy */
3327 : 379 : state->memtuples = NULL;
3328 : 379 : state->memtupsize = 0;
3329 : :
3330 : : /*
3331 : : * Parallel worker requires result tape metadata, which is to be stored in
3332 : : * shared memory for leader
3333 : : */
1820 heikki.linnakangas@i 3334 : 379 : LogicalTapeFreeze(state->result_tape, &output);
3335 : :
3336 : : /* Store properties of output tape, and update finished worker count */
3174 rhaas@postgresql.org 3337 : 379 : shared->tapes[state->worker] = output;
20 nathan@postgresql.or 3338 :GNC 379 : pg_atomic_fetch_add_u32(&shared->workersFinished, 1);
3174 rhaas@postgresql.org 3339 :CBC 379 : }
3340 : :
3341 : : /*
3342 : : * worker_nomergeruns - dump memtuples in worker, without merging
3343 : : *
3344 : : * This called as an alternative to mergeruns() with a worker when no
3345 : : * merging is required.
3346 : : */
3347 : : static void
3348 : 379 : worker_nomergeruns(Tuplesortstate *state)
3349 : : {
3350 [ + - - + ]: 379 : Assert(WORKER(state));
1820 heikki.linnakangas@i 3351 [ - + ]: 379 : Assert(state->result_tape == NULL);
3352 [ - + ]: 379 : Assert(state->nOutputRuns == 1);
3353 : :
3354 : 379 : state->result_tape = state->destTape;
3174 rhaas@postgresql.org 3355 : 379 : worker_freeze_result_tape(state);
3356 : 379 : }
3357 : :
3358 : : /*
3359 : : * leader_takeover_tapes - create tapeset for leader from worker tapes
3360 : : *
3361 : : * So far, leader Tuplesortstate has performed no actual sorting. By now, all
3362 : : * sorting has occurred in workers, all of which must have already returned
3363 : : * from tuplesort_performsort().
3364 : : *
3365 : : * When this returns, leader process is left in a state that is virtually
3366 : : * indistinguishable from it having generated runs as a serial external sort
3367 : : * might have.
3368 : : */
3369 : : static void
3370 : 130 : leader_takeover_tapes(Tuplesortstate *state)
3371 : : {
3372 : 130 : Sharedsort *shared = state->shared;
3373 : 130 : int nParticipants = state->nParticipants;
3374 : : int workersFinished;
3375 : : int j;
3376 : :
3377 [ + - - + ]: 130 : Assert(LEADER(state));
3378 [ - + ]: 130 : Assert(nParticipants >= 1);
3379 : :
20 nathan@postgresql.or 3380 :GNC 130 : workersFinished = pg_atomic_read_membarrier_u32(&shared->workersFinished);
3381 : :
3174 rhaas@postgresql.org 3382 [ - + ]:CBC 130 : if (nParticipants != workersFinished)
3174 rhaas@postgresql.org 3383 [ # # ]:UBC 0 : elog(ERROR, "cannot take over tapes before all workers finish");
3384 : :
3385 : : /*
3386 : : * Create the tapeset from worker tapes, including a leader-owned tape at
3387 : : * the end. Parallel workers are far more expensive than logical tapes,
3388 : : * so the number of tapes allocated here should never be excessive.
3389 : : */
1820 heikki.linnakangas@i 3390 :CBC 130 : inittapestate(state, nParticipants);
3391 : 130 : state->tapeset = LogicalTapeSetCreate(false, &shared->fileset, -1);
3392 : :
3393 : : /*
3394 : : * Set currentRun to reflect the number of runs we will merge (it's not
3395 : : * used for anything, this is just pro forma)
3396 : : */
3174 rhaas@postgresql.org 3397 : 130 : state->currentRun = nParticipants;
3398 : :
3399 : : /*
3400 : : * Initialize the state to look the same as after building the initial
3401 : : * runs.
3402 : : *
3403 : : * There will always be exactly 1 run per worker, and exactly one input
3404 : : * tape per run, because workers always output exactly 1 run, even when
3405 : : * there were no input tuples for workers to sort.
3406 : : */
1820 heikki.linnakangas@i 3407 : 130 : state->inputTapes = NULL;
3408 : 130 : state->nInputTapes = 0;
3409 : 130 : state->nInputRuns = 0;
3410 : :
56 michael@paquier.xyz 3411 :GNC 130 : state->outputTapes = palloc0_array(LogicalTape *, nParticipants);
1820 heikki.linnakangas@i 3412 :CBC 130 : state->nOutputTapes = nParticipants;
3413 : 130 : state->nOutputRuns = nParticipants;
3414 : :
3415 [ + + ]: 413 : for (j = 0; j < nParticipants; j++)
3416 : : {
3417 : 283 : state->outputTapes[j] = LogicalTapeImport(state->tapeset, j, &shared->tapes[j]);
3418 : : }
3419 : :
3174 rhaas@postgresql.org 3420 : 130 : state->status = TSS_BUILDRUNS;
3421 : 130 : }
3422 : :
3423 : : /*
3424 : : * Convenience routine to free a tuple previously loaded into sort memory
3425 : : */
3426 : : static void
7101 tgl@sss.pgh.pa.us 3427 : 2139679 : free_sort_tuple(Tuplesortstate *state, SortTuple *stup)
3428 : : {
1917 drowley@postgresql.o 3429 [ + + ]: 2139679 : if (stup->tuple)
3430 : : {
3431 : 2044150 : FREEMEM(state, GetMemoryChunkSpace(stup->tuple));
3432 : 2044150 : pfree(stup->tuple);
3433 : 2044150 : stup->tuple = NULL;
3434 : : }
7101 tgl@sss.pgh.pa.us 3435 : 2139679 : }
3436 : :
3437 : : int
63 john.naylor@postgres 3438 :GNC 4774433 : ssup_datum_uint64_cmp(Datum x, Datum y, SortSupport ssup)
3439 : : {
1654 john.naylor@postgres 3440 [ + + ]:CBC 4774433 : if (x < y)
3441 : 2033283 : return -1;
3442 [ + + ]: 2741150 : else if (x > y)
3443 : 684235 : return 1;
3444 : : else
3445 : 2056915 : return 0;
3446 : : }
3447 : :
3448 : : int
63 john.naylor@postgres 3449 :GNC 3378217 : ssup_datum_int64_cmp(Datum x, Datum y, SortSupport ssup)
3450 : : {
1615 drowley@postgresql.o 3451 :CBC 3378217 : int64 xx = DatumGetInt64(x);
3452 : 3378217 : int64 yy = DatumGetInt64(y);
3453 : :
1654 john.naylor@postgres 3454 [ + + ]: 3378217 : if (xx < yy)
3455 : 2802388 : return -1;
3456 [ + + ]: 575829 : else if (xx > yy)
3457 : 247375 : return 1;
3458 : : else
3459 : 328454 : return 0;
3460 : : }
3461 : :
3462 : : int
63 john.naylor@postgres 3463 :GNC 140300646 : ssup_datum_uint32_cmp(Datum x, Datum y, SortSupport ssup)
3464 : : {
3465 : 140300646 : uint32 xx = DatumGetUInt32(x);
3466 : 140300646 : uint32 yy = DatumGetUInt32(y);
3467 : :
3468 [ + + ]: 140300646 : if (xx < yy)
3469 : 34303007 : return -1;
3470 [ + + ]: 105997639 : else if (xx > yy)
3471 : 31241737 : return 1;
3472 : : else
3473 : 74755902 : return 0;
3474 : : }
3475 : :
3476 : : int
1654 john.naylor@postgres 3477 :CBC 186965063 : ssup_datum_int32_cmp(Datum x, Datum y, SortSupport ssup)
3478 : : {
1615 drowley@postgresql.o 3479 : 186965063 : int32 xx = DatumGetInt32(x);
3480 : 186965063 : int32 yy = DatumGetInt32(y);
3481 : :
1654 john.naylor@postgres 3482 [ + + ]: 186965063 : if (xx < yy)
3483 : 40657308 : return -1;
3484 [ + + ]: 146307755 : else if (xx > yy)
3485 : 33568550 : return 1;
3486 : : else
3487 : 112739205 : return 0;
3488 : : }
|
