PostgreSQL Source Code  git master
nbtsplitloc.c File Reference
#include "postgres.h"
#include "access/nbtree.h"
#include "common/int.h"
Include dependency graph for nbtsplitloc.c:

Go to the source code of this file.

Data Structures

struct  SplitPoint
 
struct  FindSplitData
 

Macros

#define LEAF_SPLIT_DISTANCE   0.050
 
#define INTERNAL_SPLIT_DISTANCE   0.075
 

Enumerations

enum  FindSplitStrat { SPLIT_DEFAULT , SPLIT_MANY_DUPLICATES , SPLIT_SINGLE_VALUE }
 

Functions

static void _bt_recsplitloc (FindSplitData *state, OffsetNumber firstrightoff, bool newitemonleft, int olddataitemstoleft, Size firstrightofforigpagetuplesz)
 
static void _bt_deltasortsplits (FindSplitData *state, double fillfactormult, bool usemult)
 
static int _bt_splitcmp (const void *arg1, const void *arg2)
 
static bool _bt_afternewitemoff (FindSplitData *state, OffsetNumber maxoff, int leaffillfactor, bool *usemult)
 
static bool _bt_adjacenthtid (ItemPointer lowhtid, ItemPointer highhtid)
 
static OffsetNumber _bt_bestsplitloc (FindSplitData *state, int perfectpenalty, bool *newitemonleft, FindSplitStrat strategy)
 
static int _bt_defaultinterval (FindSplitData *state)
 
static int _bt_strategy (FindSplitData *state, SplitPoint *leftpage, SplitPoint *rightpage, FindSplitStrat *strategy)
 
static void _bt_interval_edges (FindSplitData *state, SplitPoint **leftinterval, SplitPoint **rightinterval)
 
static int _bt_split_penalty (FindSplitData *state, SplitPoint *split)
 
static IndexTuple _bt_split_lastleft (FindSplitData *state, SplitPoint *split)
 
static IndexTuple _bt_split_firstright (FindSplitData *state, SplitPoint *split)
 
OffsetNumber _bt_findsplitloc (Relation rel, Page origpage, OffsetNumber newitemoff, Size newitemsz, IndexTuple newitem, bool *newitemonleft)
 

Macro Definition Documentation

◆ INTERNAL_SPLIT_DISTANCE

#define INTERNAL_SPLIT_DISTANCE   0.075

Definition at line 850 of file nbtsplitloc.c.

◆ LEAF_SPLIT_DISTANCE

#define LEAF_SPLIT_DISTANCE   0.050

Definition at line 849 of file nbtsplitloc.c.

Enumeration Type Documentation

◆ FindSplitStrat

Enumerator
SPLIT_DEFAULT 
SPLIT_MANY_DUPLICATES 
SPLIT_SINGLE_VALUE 

Definition at line 20 of file nbtsplitloc.c.

21 {
22  /* strategy for searching through materialized list of split points */
23  SPLIT_DEFAULT, /* give some weight to truncation */
24  SPLIT_MANY_DUPLICATES, /* find minimally distinguishing point */
25  SPLIT_SINGLE_VALUE, /* leave left page almost full */
FindSplitStrat
Definition: nbtsplitloc.c:21
@ SPLIT_DEFAULT
Definition: nbtsplitloc.c:23
@ SPLIT_MANY_DUPLICATES
Definition: nbtsplitloc.c:24
@ SPLIT_SINGLE_VALUE
Definition: nbtsplitloc.c:25

Function Documentation

◆ _bt_adjacenthtid()

static bool _bt_adjacenthtid ( ItemPointer  lowhtid,
ItemPointer  highhtid 
)
static

Definition at line 749 of file nbtsplitloc.c.

750 {
751  BlockNumber lowblk,
752  highblk;
753 
754  lowblk = ItemPointerGetBlockNumber(lowhtid);
755  highblk = ItemPointerGetBlockNumber(highhtid);
756 
757  /* Make optimistic assumption of adjacency when heap blocks match */
758  if (lowblk == highblk)
759  return true;
760 
761  /* When heap block one up, second offset should be FirstOffsetNumber */
762  if (lowblk + 1 == highblk &&
764  return true;
765 
766  return false;
767 }
uint32 BlockNumber
Definition: block.h:31
static OffsetNumber ItemPointerGetOffsetNumber(const ItemPointerData *pointer)
Definition: itemptr.h:124
static BlockNumber ItemPointerGetBlockNumber(const ItemPointerData *pointer)
Definition: itemptr.h:103
#define FirstOffsetNumber
Definition: off.h:27

References FirstOffsetNumber, ItemPointerGetBlockNumber(), and ItemPointerGetOffsetNumber().

Referenced by _bt_afternewitemoff().

◆ _bt_afternewitemoff()

static bool _bt_afternewitemoff ( FindSplitData state,
OffsetNumber  maxoff,
int  leaffillfactor,
bool usemult 
)
static

Definition at line 630 of file nbtsplitloc.c.

632 {
633  int16 nkeyatts;
634  ItemId itemid;
635  IndexTuple tup;
636  int keepnatts;
637 
638  Assert(state->is_leaf && !state->is_rightmost);
639 
641 
642  /* Single key indexes not considered here */
643  if (nkeyatts == 1)
644  return false;
645 
646  /* Ascending insertion pattern never inferred when new item is first */
647  if (state->newitemoff == P_FIRSTKEY)
648  return false;
649 
650  /*
651  * Only apply optimization on pages with equisized tuples, since ordinal
652  * keys are likely to be fixed-width. Testing if the new tuple is
653  * variable width directly might also work, but that fails to apply the
654  * optimization to indexes with a numeric_ops attribute.
655  *
656  * Conclude that page has equisized tuples when the new item is the same
657  * width as the smallest item observed during pass over page, and other
658  * non-pivot tuples must be the same width as well. (Note that the
659  * possibly-truncated existing high key isn't counted in
660  * olddataitemstotal, and must be subtracted from maxoff.)
661  */
662  if (state->newitemsz != state->minfirstrightsz)
663  return false;
664  if (state->newitemsz * (maxoff - 1) != state->olddataitemstotal)
665  return false;
666 
667  /*
668  * Avoid applying optimization when tuples are wider than a tuple
669  * consisting of two non-NULL int8/int64 attributes (or four non-NULL
670  * int4/int32 attributes)
671  */
672  if (state->newitemsz >
673  MAXALIGN(sizeof(IndexTupleData) + sizeof(int64) * 2) +
674  sizeof(ItemIdData))
675  return false;
676 
677  /*
678  * At least the first attribute's value must be equal to the corresponding
679  * value in previous tuple to apply optimization. New item cannot be a
680  * duplicate, either.
681  *
682  * Handle case where new item is to the right of all items on the existing
683  * page. This is suggestive of monotonically increasing insertions in
684  * itself, so the "heap TID adjacency" test is not applied here.
685  */
686  if (state->newitemoff > maxoff)
687  {
688  itemid = PageGetItemId(state->origpage, maxoff);
689  tup = (IndexTuple) PageGetItem(state->origpage, itemid);
690  keepnatts = _bt_keep_natts_fast(state->rel, tup, state->newitem);
691 
692  if (keepnatts > 1 && keepnatts <= nkeyatts)
693  {
694  *usemult = true;
695  return true;
696  }
697 
698  return false;
699  }
700 
701  /*
702  * "Low cardinality leading column, high cardinality suffix column"
703  * indexes with a random insertion pattern (e.g., an index with a boolean
704  * column, such as an index on '(book_is_in_print, book_isbn)') present us
705  * with a risk of consistently misapplying the optimization. We're
706  * willing to accept very occasional misapplication of the optimization,
707  * provided the cases where we get it wrong are rare and self-limiting.
708  *
709  * Heap TID adjacency strongly suggests that the item just to the left was
710  * inserted very recently, which limits overapplication of the
711  * optimization. Besides, all inappropriate cases triggered here will
712  * still split in the middle of the page on average.
713  */
714  itemid = PageGetItemId(state->origpage, OffsetNumberPrev(state->newitemoff));
715  tup = (IndexTuple) PageGetItem(state->origpage, itemid);
716  /* Do cheaper test first */
717  if (BTreeTupleIsPosting(tup) ||
718  !_bt_adjacenthtid(&tup->t_tid, &state->newitem->t_tid))
719  return false;
720  /* Check same conditions as rightmost item case, too */
721  keepnatts = _bt_keep_natts_fast(state->rel, tup, state->newitem);
722 
723  if (keepnatts > 1 && keepnatts <= nkeyatts)
724  {
725  double interp = (double) state->newitemoff / ((double) maxoff + 1);
726  double leaffillfactormult = (double) leaffillfactor / 100.0;
727 
728  /*
729  * Don't allow caller to split after a new item when it will result in
730  * a split point to the right of the point that a leaf fillfactor
731  * split would use -- have caller apply leaf fillfactor instead
732  */
733  *usemult = interp > leaffillfactormult;
734 
735  return true;
736  }
737 
738  return false;
739 }
static Item PageGetItem(Page page, ItemId itemId)
Definition: bufpage.h:351
static ItemId PageGetItemId(Page page, OffsetNumber offsetNumber)
Definition: bufpage.h:240
signed short int16
Definition: c.h:493
#define MAXALIGN(LEN)
Definition: c.h:811
#define Assert(condition)
Definition: c.h:858
IndexTupleData * IndexTuple
Definition: itup.h:53
static bool BTreeTupleIsPosting(IndexTuple itup)
Definition: nbtree.h:492
#define P_FIRSTKEY
Definition: nbtree.h:368
static bool _bt_adjacenthtid(ItemPointer lowhtid, ItemPointer highhtid)
Definition: nbtsplitloc.c:749
int _bt_keep_natts_fast(Relation rel, IndexTuple lastleft, IndexTuple firstright)
Definition: nbtutils.c:4874
#define OffsetNumberPrev(offsetNumber)
Definition: off.h:54
#define IndexRelationGetNumberOfKeyAttributes(relation)
Definition: rel.h:524
ItemPointerData t_tid
Definition: itup.h:37
Definition: regguts.h:323

References _bt_adjacenthtid(), _bt_keep_natts_fast(), Assert, BTreeTupleIsPosting(), IndexRelationGetNumberOfKeyAttributes, MAXALIGN, OffsetNumberPrev, P_FIRSTKEY, PageGetItem(), PageGetItemId(), and IndexTupleData::t_tid.

Referenced by _bt_findsplitloc().

◆ _bt_bestsplitloc()

static OffsetNumber _bt_bestsplitloc ( FindSplitData state,
int  perfectpenalty,
bool newitemonleft,
FindSplitStrat  strategy 
)
static

Definition at line 788 of file nbtsplitloc.c.

790 {
791  int bestpenalty,
792  lowsplit;
793  int highsplit = Min(state->interval, state->nsplits);
794  SplitPoint *final;
795 
796  bestpenalty = INT_MAX;
797  lowsplit = 0;
798  for (int i = lowsplit; i < highsplit; i++)
799  {
800  int penalty;
801 
802  penalty = _bt_split_penalty(state, state->splits + i);
803 
804  if (penalty < bestpenalty)
805  {
806  bestpenalty = penalty;
807  lowsplit = i;
808  }
809 
810  if (penalty <= perfectpenalty)
811  break;
812  }
813 
814  final = &state->splits[lowsplit];
815 
816  /*
817  * There is a risk that the "many duplicates" strategy will repeatedly do
818  * the wrong thing when there are monotonically decreasing insertions to
819  * the right of a large group of duplicates. Repeated splits could leave
820  * a succession of right half pages with free space that can never be
821  * used. This must be avoided.
822  *
823  * Consider the example of the leftmost page in a single integer attribute
824  * NULLS FIRST index which is almost filled with NULLs. Monotonically
825  * decreasing integer insertions might cause the same leftmost page to
826  * split repeatedly at the same point. Each split derives its new high
827  * key from the lowest current value to the immediate right of the large
828  * group of NULLs, which will always be higher than all future integer
829  * insertions, directing all future integer insertions to the same
830  * leftmost page.
831  */
832  if (strategy == SPLIT_MANY_DUPLICATES && !state->is_rightmost &&
833  !final->newitemonleft && final->firstrightoff >= state->newitemoff &&
834  final->firstrightoff < state->newitemoff + 9)
835  {
836  /*
837  * Avoid the problem by performing a 50:50 split when the new item is
838  * just to the right of the would-be "many duplicates" split point.
839  * (Note that the test used for an insert that is "just to the right"
840  * of the split point is conservative.)
841  */
842  final = &state->splits[0];
843  }
844 
845  *newitemonleft = final->newitemonleft;
846  return final->firstrightoff;
847 }
#define Min(x, y)
Definition: c.h:1004
#define final(a, b, c)
Definition: hashfn.c:116
int i
Definition: isn.c:73
static int _bt_split_penalty(FindSplitData *state, SplitPoint *split)
Definition: nbtsplitloc.c:1131

References _bt_split_penalty(), final, i, Min, and SPLIT_MANY_DUPLICATES.

Referenced by _bt_findsplitloc().

◆ _bt_defaultinterval()

static int _bt_defaultinterval ( FindSplitData state)
static

Definition at line 876 of file nbtsplitloc.c.

877 {
878  SplitPoint *spaceoptimal;
879  int16 tolerance,
880  lowleftfree,
881  lowrightfree,
882  highleftfree,
883  highrightfree;
884 
885  /*
886  * Determine leftfree and rightfree values that are higher and lower than
887  * we're willing to tolerate. Note that the final split interval will be
888  * about 10% of nsplits in the common case where all non-pivot tuples
889  * (data items) from a leaf page are uniformly sized. We're a bit more
890  * aggressive when splitting internal pages.
891  */
892  if (state->is_leaf)
893  tolerance = state->olddataitemstotal * LEAF_SPLIT_DISTANCE;
894  else
895  tolerance = state->olddataitemstotal * INTERNAL_SPLIT_DISTANCE;
896 
897  /* First candidate split point is the most evenly balanced */
898  spaceoptimal = state->splits;
899  lowleftfree = spaceoptimal->leftfree - tolerance;
900  lowrightfree = spaceoptimal->rightfree - tolerance;
901  highleftfree = spaceoptimal->leftfree + tolerance;
902  highrightfree = spaceoptimal->rightfree + tolerance;
903 
904  /*
905  * Iterate through split points, starting from the split immediately after
906  * 'spaceoptimal'. Find the first split point that divides free space so
907  * unevenly that including it in the split interval would be unacceptable.
908  */
909  for (int i = 1; i < state->nsplits; i++)
910  {
911  SplitPoint *split = state->splits + i;
912 
913  /* Cannot use curdelta here, since its value is often weighted */
914  if (split->leftfree < lowleftfree || split->rightfree < lowrightfree ||
915  split->leftfree > highleftfree || split->rightfree > highrightfree)
916  return i;
917  }
918 
919  return state->nsplits;
920 }
#define LEAF_SPLIT_DISTANCE
Definition: nbtsplitloc.c:849
#define INTERNAL_SPLIT_DISTANCE
Definition: nbtsplitloc.c:850
int16 leftfree
Definition: nbtsplitloc.c:32
int16 rightfree
Definition: nbtsplitloc.c:33

References i, INTERNAL_SPLIT_DISTANCE, LEAF_SPLIT_DISTANCE, SplitPoint::leftfree, and SplitPoint::rightfree.

Referenced by _bt_findsplitloc().

◆ _bt_deltasortsplits()

static void _bt_deltasortsplits ( FindSplitData state,
double  fillfactormult,
bool  usemult 
)
static

Definition at line 566 of file nbtsplitloc.c.

568 {
569  for (int i = 0; i < state->nsplits; i++)
570  {
571  SplitPoint *split = state->splits + i;
572  int16 delta;
573 
574  if (usemult)
575  delta = fillfactormult * split->leftfree -
576  (1.0 - fillfactormult) * split->rightfree;
577  else
578  delta = split->leftfree - split->rightfree;
579 
580  if (delta < 0)
581  delta = -delta;
582 
583  /* Save delta */
584  split->curdelta = delta;
585  }
586 
587  qsort(state->splits, state->nsplits, sizeof(SplitPoint), _bt_splitcmp);
588 }
if(TABLE==NULL||TABLE_index==NULL)
Definition: isn.c:77
static int _bt_splitcmp(const void *arg1, const void *arg2)
Definition: nbtsplitloc.c:594
#define qsort(a, b, c, d)
Definition: port.h:449
int16 curdelta
Definition: nbtsplitloc.c:31

References _bt_splitcmp(), SplitPoint::curdelta, i, if(), SplitPoint::leftfree, qsort, and SplitPoint::rightfree.

Referenced by _bt_findsplitloc().

◆ _bt_findsplitloc()

OffsetNumber _bt_findsplitloc ( Relation  rel,
Page  origpage,
OffsetNumber  newitemoff,
Size  newitemsz,
IndexTuple  newitem,
bool newitemonleft 
)

Definition at line 129 of file nbtsplitloc.c.

135 {
136  BTPageOpaque opaque;
137  int leftspace,
138  rightspace,
139  olddataitemstotal,
140  olddataitemstoleft,
141  perfectpenalty,
142  leaffillfactor;
144  FindSplitStrat strategy;
145  ItemId itemid;
146  OffsetNumber offnum,
147  maxoff,
148  firstrightoff;
149  double fillfactormult;
150  bool usemult;
151  SplitPoint leftpage,
152  rightpage;
153 
154  opaque = BTPageGetOpaque(origpage);
155  maxoff = PageGetMaxOffsetNumber(origpage);
156 
157  /* Total free space available on a btree page, after fixed overhead */
158  leftspace = rightspace =
160  MAXALIGN(sizeof(BTPageOpaqueData));
161 
162  /* The right page will have the same high key as the old page */
163  if (!P_RIGHTMOST(opaque))
164  {
165  itemid = PageGetItemId(origpage, P_HIKEY);
166  rightspace -= (int) (MAXALIGN(ItemIdGetLength(itemid)) +
167  sizeof(ItemIdData));
168  }
169 
170  /* Count up total space in data items before actually scanning 'em */
171  olddataitemstotal = rightspace - (int) PageGetExactFreeSpace(origpage);
172  leaffillfactor = BTGetFillFactor(rel);
173 
174  /* Passed-in newitemsz is MAXALIGNED but does not include line pointer */
175  newitemsz += sizeof(ItemIdData);
176  state.rel = rel;
177  state.origpage = origpage;
178  state.newitem = newitem;
179  state.newitemsz = newitemsz;
180  state.is_leaf = P_ISLEAF(opaque);
181  state.is_rightmost = P_RIGHTMOST(opaque);
182  state.leftspace = leftspace;
183  state.rightspace = rightspace;
184  state.olddataitemstotal = olddataitemstotal;
185  state.minfirstrightsz = SIZE_MAX;
186  state.newitemoff = newitemoff;
187 
188  /* newitem cannot be a posting list item */
189  Assert(!BTreeTupleIsPosting(newitem));
190 
191  /*
192  * nsplits should never exceed maxoff because there will be at most as
193  * many candidate split points as there are points _between_ tuples, once
194  * you imagine that the new item is already on the original page (the
195  * final number of splits may be slightly lower because not all points
196  * between tuples will be legal).
197  */
198  state.maxsplits = maxoff;
199  state.splits = palloc(sizeof(SplitPoint) * state.maxsplits);
200  state.nsplits = 0;
201 
202  /*
203  * Scan through the data items and calculate space usage for a split at
204  * each possible position
205  */
206  olddataitemstoleft = 0;
207 
208  for (offnum = P_FIRSTDATAKEY(opaque);
209  offnum <= maxoff;
210  offnum = OffsetNumberNext(offnum))
211  {
212  Size itemsz;
213 
214  itemid = PageGetItemId(origpage, offnum);
215  itemsz = MAXALIGN(ItemIdGetLength(itemid)) + sizeof(ItemIdData);
216 
217  /*
218  * When item offset number is not newitemoff, neither side of the
219  * split can be newitem. Record a split after the previous data item
220  * from original page, but before the current data item from original
221  * page. (_bt_recsplitloc() will reject the split when there are no
222  * previous items, which we rely on.)
223  */
224  if (offnum < newitemoff)
225  _bt_recsplitloc(&state, offnum, false, olddataitemstoleft, itemsz);
226  else if (offnum > newitemoff)
227  _bt_recsplitloc(&state, offnum, true, olddataitemstoleft, itemsz);
228  else
229  {
230  /*
231  * Record a split after all "offnum < newitemoff" original page
232  * data items, but before newitem
233  */
234  _bt_recsplitloc(&state, offnum, false, olddataitemstoleft, itemsz);
235 
236  /*
237  * Record a split after newitem, but before data item from
238  * original page at offset newitemoff/current offset
239  */
240  _bt_recsplitloc(&state, offnum, true, olddataitemstoleft, itemsz);
241  }
242 
243  olddataitemstoleft += itemsz;
244  }
245 
246  /*
247  * Record a split after all original page data items, but before newitem.
248  * (Though only when it's possible that newitem will end up alone on new
249  * right page.)
250  */
251  Assert(olddataitemstoleft == olddataitemstotal);
252  if (newitemoff > maxoff)
253  _bt_recsplitloc(&state, newitemoff, false, olddataitemstotal, 0);
254 
255  /*
256  * I believe it is not possible to fail to find a feasible split, but just
257  * in case ...
258  */
259  if (state.nsplits == 0)
260  elog(ERROR, "could not find a feasible split point for index \"%s\"",
262 
263  /*
264  * Start search for a split point among list of legal split points. Give
265  * primary consideration to equalizing available free space in each half
266  * of the split initially (start with default strategy), while applying
267  * rightmost and split-after-new-item optimizations where appropriate.
268  * Either of the two other fallback strategies may be required for cases
269  * with a large number of duplicates around the original/space-optimal
270  * split point.
271  *
272  * Default strategy gives some weight to suffix truncation in deciding a
273  * split point on leaf pages. It attempts to select a split point where a
274  * distinguishing attribute appears earlier in the new high key for the
275  * left side of the split, in order to maximize the number of trailing
276  * attributes that can be truncated away. Only candidate split points
277  * that imply an acceptable balance of free space on each side are
278  * considered. See _bt_defaultinterval().
279  */
280  if (!state.is_leaf)
281  {
282  /* fillfactormult only used on rightmost page */
283  usemult = state.is_rightmost;
284  fillfactormult = BTREE_NONLEAF_FILLFACTOR / 100.0;
285  }
286  else if (state.is_rightmost)
287  {
288  /* Rightmost leaf page -- fillfactormult always used */
289  usemult = true;
290  fillfactormult = leaffillfactor / 100.0;
291  }
292  else if (_bt_afternewitemoff(&state, maxoff, leaffillfactor, &usemult))
293  {
294  /*
295  * New item inserted at rightmost point among a localized grouping on
296  * a leaf page -- apply "split after new item" optimization, either by
297  * applying leaf fillfactor multiplier, or by choosing the exact split
298  * point that leaves newitem as lastleft. (usemult is set for us.)
299  */
300  if (usemult)
301  {
302  /* fillfactormult should be set based on leaf fillfactor */
303  fillfactormult = leaffillfactor / 100.0;
304  }
305  else
306  {
307  /* find precise split point after newitemoff */
308  for (int i = 0; i < state.nsplits; i++)
309  {
310  SplitPoint *split = state.splits + i;
311 
312  if (split->newitemonleft &&
313  newitemoff == split->firstrightoff)
314  {
315  pfree(state.splits);
316  *newitemonleft = true;
317  return newitemoff;
318  }
319  }
320 
321  /*
322  * Cannot legally split after newitemoff; proceed with split
323  * without using fillfactor multiplier. This is defensive, and
324  * should never be needed in practice.
325  */
326  fillfactormult = 0.50;
327  }
328  }
329  else
330  {
331  /* Other leaf page. 50:50 page split. */
332  usemult = false;
333  /* fillfactormult not used, but be tidy */
334  fillfactormult = 0.50;
335  }
336 
337  /*
338  * Save leftmost and rightmost splits for page before original ordinal
339  * sort order is lost by delta/fillfactormult sort
340  */
341  leftpage = state.splits[0];
342  rightpage = state.splits[state.nsplits - 1];
343 
344  /* Give split points a fillfactormult-wise delta, and sort on deltas */
345  _bt_deltasortsplits(&state, fillfactormult, usemult);
346 
347  /* Determine split interval for default strategy */
348  state.interval = _bt_defaultinterval(&state);
349 
350  /*
351  * Determine if default strategy/split interval will produce a
352  * sufficiently distinguishing split, or if we should change strategies.
353  * Alternative strategies change the range of split points that are
354  * considered acceptable (split interval), and possibly change
355  * fillfactormult, in order to deal with pages with a large number of
356  * duplicates gracefully.
357  *
358  * Pass low and high splits for the entire page (actually, they're for an
359  * imaginary version of the page that includes newitem). These are used
360  * when the initial split interval encloses split points that are full of
361  * duplicates, and we need to consider if it's even possible to avoid
362  * appending a heap TID.
363  */
364  perfectpenalty = _bt_strategy(&state, &leftpage, &rightpage, &strategy);
365 
366  if (strategy == SPLIT_DEFAULT)
367  {
368  /*
369  * Default strategy worked out (always works out with internal page).
370  * Original split interval still stands.
371  */
372  }
373 
374  /*
375  * Many duplicates strategy is used when a heap TID would otherwise be
376  * appended, but the page isn't completely full of logical duplicates.
377  *
378  * The split interval is widened to include all legal candidate split
379  * points. There might be a few as two distinct values in the whole-page
380  * split interval, though it's also possible that most of the values on
381  * the page are unique. The final split point will either be to the
382  * immediate left or to the immediate right of the group of duplicate
383  * tuples that enclose the first/delta-optimal split point (perfect
384  * penalty was set so that the lowest delta split point that avoids
385  * appending a heap TID will be chosen). Maximizing the number of
386  * attributes that can be truncated away is not a goal of the many
387  * duplicates strategy.
388  *
389  * Single value strategy is used when it is impossible to avoid appending
390  * a heap TID. It arranges to leave the left page very full. This
391  * maximizes space utilization in cases where tuples with the same
392  * attribute values span many pages. Newly inserted duplicates will tend
393  * to have higher heap TID values, so we'll end up splitting to the right
394  * consistently. (Single value strategy is harmless though not
395  * particularly useful with !heapkeyspace indexes.)
396  */
397  else if (strategy == SPLIT_MANY_DUPLICATES)
398  {
399  Assert(state.is_leaf);
400  /* Shouldn't try to truncate away extra user attributes */
401  Assert(perfectpenalty ==
403  /* No need to resort splits -- no change in fillfactormult/deltas */
404  state.interval = state.nsplits;
405  }
406  else if (strategy == SPLIT_SINGLE_VALUE)
407  {
408  Assert(state.is_leaf);
409  /* Split near the end of the page */
410  usemult = true;
411  fillfactormult = BTREE_SINGLEVAL_FILLFACTOR / 100.0;
412  /* Resort split points with new delta */
413  _bt_deltasortsplits(&state, fillfactormult, usemult);
414  /* Appending a heap TID is unavoidable, so interval of 1 is fine */
415  state.interval = 1;
416  }
417 
418  /*
419  * Search among acceptable split points (using final split interval) for
420  * the entry that has the lowest penalty, and is therefore expected to
421  * maximize fan-out. Sets *newitemonleft for us.
422  */
423  firstrightoff = _bt_bestsplitloc(&state, perfectpenalty, newitemonleft,
424  strategy);
425  pfree(state.splits);
426 
427  return firstrightoff;
428 }
Size PageGetExactFreeSpace(Page page)
Definition: bufpage.c:958
static Size PageGetPageSize(Page page)
Definition: bufpage.h:273
#define SizeOfPageHeaderData
Definition: bufpage.h:213
static OffsetNumber PageGetMaxOffsetNumber(Page page)
Definition: bufpage.h:369
size_t Size
Definition: c.h:605
#define ERROR
Definition: elog.h:39
#define elog(elevel,...)
Definition: elog.h:224
#define ItemIdGetLength(itemId)
Definition: itemid.h:59
struct ItemIdData ItemIdData
void pfree(void *pointer)
Definition: mcxt.c:1520
void * palloc(Size size)
Definition: mcxt.c:1316
#define P_ISLEAF(opaque)
Definition: nbtree.h:220
#define BTREE_SINGLEVAL_FILLFACTOR
Definition: nbtree.h:202
#define P_HIKEY
Definition: nbtree.h:367
#define BTPageGetOpaque(page)
Definition: nbtree.h:73
#define P_FIRSTDATAKEY(opaque)
Definition: nbtree.h:369
#define P_RIGHTMOST(opaque)
Definition: nbtree.h:219
#define BTGetFillFactor(relation)
Definition: nbtree.h:1138
#define BTREE_NONLEAF_FILLFACTOR
Definition: nbtree.h:201
static void _bt_deltasortsplits(FindSplitData *state, double fillfactormult, bool usemult)
Definition: nbtsplitloc.c:566
static int _bt_strategy(FindSplitData *state, SplitPoint *leftpage, SplitPoint *rightpage, FindSplitStrat *strategy)
Definition: nbtsplitloc.c:934
static bool _bt_afternewitemoff(FindSplitData *state, OffsetNumber maxoff, int leaffillfactor, bool *usemult)
Definition: nbtsplitloc.c:630
static void _bt_recsplitloc(FindSplitData *state, OffsetNumber firstrightoff, bool newitemonleft, int olddataitemstoleft, Size firstrightofforigpagetuplesz)
Definition: nbtsplitloc.c:449
static OffsetNumber _bt_bestsplitloc(FindSplitData *state, int perfectpenalty, bool *newitemonleft, FindSplitStrat strategy)
Definition: nbtsplitloc.c:788
static int _bt_defaultinterval(FindSplitData *state)
Definition: nbtsplitloc.c:876
#define OffsetNumberNext(offsetNumber)
Definition: off.h:52
uint16 OffsetNumber
Definition: off.h:24
#define RelationGetRelationName(relation)
Definition: rel.h:539
bool newitemonleft
Definition: nbtsplitloc.c:37
OffsetNumber firstrightoff
Definition: nbtsplitloc.c:36

References _bt_afternewitemoff(), _bt_bestsplitloc(), _bt_defaultinterval(), _bt_deltasortsplits(), _bt_recsplitloc(), _bt_strategy(), Assert, BTGetFillFactor, BTPageGetOpaque, BTREE_NONLEAF_FILLFACTOR, BTREE_SINGLEVAL_FILLFACTOR, BTreeTupleIsPosting(), elog, ERROR, SplitPoint::firstrightoff, i, IndexRelationGetNumberOfKeyAttributes, ItemIdGetLength, MAXALIGN, SplitPoint::newitemonleft, OffsetNumberNext, P_FIRSTDATAKEY, P_HIKEY, P_ISLEAF, P_RIGHTMOST, PageGetExactFreeSpace(), PageGetItemId(), PageGetMaxOffsetNumber(), PageGetPageSize(), palloc(), pfree(), RelationGetRelationName, SizeOfPageHeaderData, SPLIT_DEFAULT, SPLIT_MANY_DUPLICATES, and SPLIT_SINGLE_VALUE.

Referenced by _bt_split().

◆ _bt_interval_edges()

static void _bt_interval_edges ( FindSplitData state,
SplitPoint **  leftinterval,
SplitPoint **  rightinterval 
)
static

Definition at line 1052 of file nbtsplitloc.c.

1054 {
1055  int highsplit = Min(state->interval, state->nsplits);
1056  SplitPoint *deltaoptimal;
1057 
1058  deltaoptimal = state->splits;
1059  *leftinterval = NULL;
1060  *rightinterval = NULL;
1061 
1062  /*
1063  * Delta is an absolute distance to optimal split point, so both the
1064  * leftmost and rightmost split point will usually be at the end of the
1065  * array
1066  */
1067  for (int i = highsplit - 1; i >= 0; i--)
1068  {
1069  SplitPoint *distant = state->splits + i;
1070 
1071  if (distant->firstrightoff < deltaoptimal->firstrightoff)
1072  {
1073  if (*leftinterval == NULL)
1074  *leftinterval = distant;
1075  }
1076  else if (distant->firstrightoff > deltaoptimal->firstrightoff)
1077  {
1078  if (*rightinterval == NULL)
1079  *rightinterval = distant;
1080  }
1081  else if (!distant->newitemonleft && deltaoptimal->newitemonleft)
1082  {
1083  /*
1084  * "incoming tuple will become firstright" (distant) is to the
1085  * left of "incoming tuple will become lastleft" (delta-optimal)
1086  */
1087  Assert(distant->firstrightoff == state->newitemoff);
1088  if (*leftinterval == NULL)
1089  *leftinterval = distant;
1090  }
1091  else if (distant->newitemonleft && !deltaoptimal->newitemonleft)
1092  {
1093  /*
1094  * "incoming tuple will become lastleft" (distant) is to the right
1095  * of "incoming tuple will become firstright" (delta-optimal)
1096  */
1097  Assert(distant->firstrightoff == state->newitemoff);
1098  if (*rightinterval == NULL)
1099  *rightinterval = distant;
1100  }
1101  else
1102  {
1103  /* There was only one or two splits in initial split interval */
1104  Assert(distant == deltaoptimal);
1105  if (*leftinterval == NULL)
1106  *leftinterval = distant;
1107  if (*rightinterval == NULL)
1108  *rightinterval = distant;
1109  }
1110 
1111  if (*leftinterval && *rightinterval)
1112  return;
1113  }
1114 
1115  Assert(false);
1116 }

References Assert, SplitPoint::firstrightoff, i, Min, and SplitPoint::newitemonleft.

Referenced by _bt_strategy().

◆ _bt_recsplitloc()

static void _bt_recsplitloc ( FindSplitData state,
OffsetNumber  firstrightoff,
bool  newitemonleft,
int  olddataitemstoleft,
Size  firstrightofforigpagetuplesz 
)
static

Definition at line 449 of file nbtsplitloc.c.

454 {
455  int16 leftfree,
456  rightfree;
457  Size firstrightsz;
458  Size postingsz = 0;
459  bool newitemisfirstright;
460 
461  /* Is the new item going to be split point's firstright tuple? */
462  newitemisfirstright = (firstrightoff == state->newitemoff &&
463  !newitemonleft);
464 
465  if (newitemisfirstright)
466  firstrightsz = state->newitemsz;
467  else
468  {
469  firstrightsz = firstrightofforigpagetuplesz;
470 
471  /*
472  * Calculate suffix truncation space saving when firstright tuple is a
473  * posting list tuple, though only when the tuple is over 64 bytes
474  * including line pointer overhead (arbitrary). This avoids accessing
475  * the tuple in cases where its posting list must be very small (if
476  * tuple has one at all).
477  *
478  * Note: We don't do this in the case where firstright tuple is
479  * newitem, since newitem cannot have a posting list.
480  */
481  if (state->is_leaf && firstrightsz > 64)
482  {
483  ItemId itemid;
484  IndexTuple newhighkey;
485 
486  itemid = PageGetItemId(state->origpage, firstrightoff);
487  newhighkey = (IndexTuple) PageGetItem(state->origpage, itemid);
488 
489  if (BTreeTupleIsPosting(newhighkey))
490  postingsz = IndexTupleSize(newhighkey) -
491  BTreeTupleGetPostingOffset(newhighkey);
492  }
493  }
494 
495  /* Account for all the old tuples */
496  leftfree = state->leftspace - olddataitemstoleft;
497  rightfree = state->rightspace -
498  (state->olddataitemstotal - olddataitemstoleft);
499 
500  /*
501  * The first item on the right page becomes the high key of the left page;
502  * therefore it counts against left space as well as right space (we
503  * cannot assume that suffix truncation will make it any smaller). When
504  * index has included attributes, then those attributes of left page high
505  * key will be truncated leaving that page with slightly more free space.
506  * However, that shouldn't affect our ability to find valid split
507  * location, since we err in the direction of being pessimistic about free
508  * space on the left half. Besides, even when suffix truncation of
509  * non-TID attributes occurs, the new high key often won't even be a
510  * single MAXALIGN() quantum smaller than the firstright tuple it's based
511  * on.
512  *
513  * If we are on the leaf level, assume that suffix truncation cannot avoid
514  * adding a heap TID to the left half's new high key when splitting at the
515  * leaf level. In practice the new high key will often be smaller and
516  * will rarely be larger, but conservatively assume the worst case. We do
517  * go to the trouble of subtracting away posting list overhead, though
518  * only when it looks like it will make an appreciable difference.
519  * (Posting lists are the only case where truncation will typically make
520  * the final high key far smaller than firstright, so being a bit more
521  * precise there noticeably improves the balance of free space.)
522  */
523  if (state->is_leaf)
524  leftfree -= (int16) (firstrightsz +
525  MAXALIGN(sizeof(ItemPointerData)) -
526  postingsz);
527  else
528  leftfree -= (int16) firstrightsz;
529 
530  /* account for the new item */
531  if (newitemonleft)
532  leftfree -= (int16) state->newitemsz;
533  else
534  rightfree -= (int16) state->newitemsz;
535 
536  /*
537  * If we are not on the leaf level, we will be able to discard the key
538  * data from the first item that winds up on the right page.
539  */
540  if (!state->is_leaf)
541  rightfree += (int16) firstrightsz -
542  (int16) (MAXALIGN(sizeof(IndexTupleData)) + sizeof(ItemIdData));
543 
544  /* Record split if legal */
545  if (leftfree >= 0 && rightfree >= 0)
546  {
547  Assert(state->nsplits < state->maxsplits);
548 
549  /* Determine smallest firstright tuple size among legal splits */
550  state->minfirstrightsz = Min(state->minfirstrightsz, firstrightsz);
551 
552  state->splits[state->nsplits].curdelta = 0;
553  state->splits[state->nsplits].leftfree = leftfree;
554  state->splits[state->nsplits].rightfree = rightfree;
555  state->splits[state->nsplits].firstrightoff = firstrightoff;
556  state->splits[state->nsplits].newitemonleft = newitemonleft;
557  state->nsplits++;
558  }
559 }
#define IndexTupleSize(itup)
Definition: itup.h:70
static uint32 BTreeTupleGetPostingOffset(IndexTuple posting)
Definition: nbtree.h:529

References Assert, BTreeTupleGetPostingOffset(), BTreeTupleIsPosting(), IndexTupleSize, MAXALIGN, Min, PageGetItem(), and PageGetItemId().

Referenced by _bt_findsplitloc().

◆ _bt_split_firstright()

static IndexTuple _bt_split_firstright ( FindSplitData state,
SplitPoint split 
)
inlinestatic

Definition at line 1175 of file nbtsplitloc.c.

1176 {
1177  ItemId itemid;
1178 
1179  if (!split->newitemonleft && split->firstrightoff == state->newitemoff)
1180  return state->newitem;
1181 
1182  itemid = PageGetItemId(state->origpage, split->firstrightoff);
1183  return (IndexTuple) PageGetItem(state->origpage, itemid);
1184 }

References SplitPoint::firstrightoff, SplitPoint::newitemonleft, PageGetItem(), and PageGetItemId().

Referenced by _bt_split_penalty(), and _bt_strategy().

◆ _bt_split_lastleft()

static IndexTuple _bt_split_lastleft ( FindSplitData state,
SplitPoint split 
)
inlinestatic

Definition at line 1159 of file nbtsplitloc.c.

1160 {
1161  ItemId itemid;
1162 
1163  if (split->newitemonleft && split->firstrightoff == state->newitemoff)
1164  return state->newitem;
1165 
1166  itemid = PageGetItemId(state->origpage,
1168  return (IndexTuple) PageGetItem(state->origpage, itemid);
1169 }

References SplitPoint::firstrightoff, SplitPoint::newitemonleft, OffsetNumberPrev, PageGetItem(), and PageGetItemId().

Referenced by _bt_split_penalty(), and _bt_strategy().

◆ _bt_split_penalty()

static int _bt_split_penalty ( FindSplitData state,
SplitPoint split 
)
inlinestatic

Definition at line 1131 of file nbtsplitloc.c.

1132 {
1133  IndexTuple lastleft;
1134  IndexTuple firstright;
1135 
1136  if (!state->is_leaf)
1137  {
1138  ItemId itemid;
1139 
1140  if (!split->newitemonleft &&
1141  split->firstrightoff == state->newitemoff)
1142  return state->newitemsz;
1143 
1144  itemid = PageGetItemId(state->origpage, split->firstrightoff);
1145 
1146  return MAXALIGN(ItemIdGetLength(itemid)) + sizeof(ItemIdData);
1147  }
1148 
1149  lastleft = _bt_split_lastleft(state, split);
1150  firstright = _bt_split_firstright(state, split);
1151 
1152  return _bt_keep_natts_fast(state->rel, lastleft, firstright);
1153 }
static IndexTuple _bt_split_lastleft(FindSplitData *state, SplitPoint *split)
Definition: nbtsplitloc.c:1159
static IndexTuple _bt_split_firstright(FindSplitData *state, SplitPoint *split)
Definition: nbtsplitloc.c:1175

References _bt_keep_natts_fast(), _bt_split_firstright(), _bt_split_lastleft(), SplitPoint::firstrightoff, ItemIdGetLength, MAXALIGN, SplitPoint::newitemonleft, and PageGetItemId().

Referenced by _bt_bestsplitloc().

◆ _bt_splitcmp()

static int _bt_splitcmp ( const void *  arg1,
const void *  arg2 
)
static

Definition at line 594 of file nbtsplitloc.c.

595 {
596  SplitPoint *split1 = (SplitPoint *) arg1;
597  SplitPoint *split2 = (SplitPoint *) arg2;
598 
599  return pg_cmp_s16(split1->curdelta, split2->curdelta);
600 }
static int pg_cmp_s16(int16 a, int16 b)
Definition: int.h:471

References SplitPoint::curdelta, and pg_cmp_s16().

Referenced by _bt_deltasortsplits().

◆ _bt_strategy()

static int _bt_strategy ( FindSplitData state,
SplitPoint leftpage,
SplitPoint rightpage,
FindSplitStrat strategy 
)
static

Definition at line 934 of file nbtsplitloc.c.

936 {
937  IndexTuple leftmost,
938  rightmost;
939  SplitPoint *leftinterval,
940  *rightinterval;
941  int perfectpenalty;
942  int indnkeyatts = IndexRelationGetNumberOfKeyAttributes(state->rel);
943 
944  /* Assume that alternative strategy won't be used for now */
945  *strategy = SPLIT_DEFAULT;
946 
947  /*
948  * Use smallest observed firstright item size for entire page (actually,
949  * entire imaginary version of page that includes newitem) as perfect
950  * penalty on internal pages. This can save cycles in the common case
951  * where most or all splits (not just splits within interval) have
952  * firstright tuples that are the same size.
953  */
954  if (!state->is_leaf)
955  return state->minfirstrightsz;
956 
957  /*
958  * Use leftmost and rightmost tuples from leftmost and rightmost splits in
959  * current split interval
960  */
961  _bt_interval_edges(state, &leftinterval, &rightinterval);
962  leftmost = _bt_split_lastleft(state, leftinterval);
963  rightmost = _bt_split_firstright(state, rightinterval);
964 
965  /*
966  * If initial split interval can produce a split point that will at least
967  * avoid appending a heap TID in new high key, we're done. Finish split
968  * with default strategy and initial split interval.
969  */
970  perfectpenalty = _bt_keep_natts_fast(state->rel, leftmost, rightmost);
971  if (perfectpenalty <= indnkeyatts)
972  return perfectpenalty;
973 
974  /*
975  * Work out how caller should finish split when even their "perfect"
976  * penalty for initial/default split interval indicates that the interval
977  * does not contain even a single split that avoids appending a heap TID.
978  *
979  * Use the leftmost split's lastleft tuple and the rightmost split's
980  * firstright tuple to assess every possible split.
981  */
982  leftmost = _bt_split_lastleft(state, leftpage);
983  rightmost = _bt_split_firstright(state, rightpage);
984 
985  /*
986  * If page (including new item) has many duplicates but is not entirely
987  * full of duplicates, a many duplicates strategy split will be performed.
988  * If page is entirely full of duplicates, a single value strategy split
989  * will be performed.
990  */
991  perfectpenalty = _bt_keep_natts_fast(state->rel, leftmost, rightmost);
992  if (perfectpenalty <= indnkeyatts)
993  {
994  *strategy = SPLIT_MANY_DUPLICATES;
995 
996  /*
997  * Many duplicates strategy should split at either side the group of
998  * duplicates that enclose the delta-optimal split point. Return
999  * indnkeyatts rather than the true perfect penalty to make that
1000  * happen. (If perfectpenalty was returned here then low cardinality
1001  * composite indexes could have continual unbalanced splits.)
1002  *
1003  * Note that caller won't go through with a many duplicates split in
1004  * rare cases where it looks like there are ever-decreasing insertions
1005  * to the immediate right of the split point. This must happen just
1006  * before a final decision is made, within _bt_bestsplitloc().
1007  */
1008  return indnkeyatts;
1009  }
1010 
1011  /*
1012  * Single value strategy is only appropriate with ever-increasing heap
1013  * TIDs; otherwise, original default strategy split should proceed to
1014  * avoid pathological performance. Use page high key to infer if this is
1015  * the rightmost page among pages that store the same duplicate value.
1016  * This should not prevent insertions of heap TIDs that are slightly out
1017  * of order from using single value strategy, since that's expected with
1018  * concurrent inserters of the same duplicate value.
1019  */
1020  else if (state->is_rightmost)
1021  *strategy = SPLIT_SINGLE_VALUE;
1022  else
1023  {
1024  ItemId itemid;
1025  IndexTuple hikey;
1026 
1027  itemid = PageGetItemId(state->origpage, P_HIKEY);
1028  hikey = (IndexTuple) PageGetItem(state->origpage, itemid);
1029  perfectpenalty = _bt_keep_natts_fast(state->rel, hikey,
1030  state->newitem);
1031  if (perfectpenalty <= indnkeyatts)
1032  *strategy = SPLIT_SINGLE_VALUE;
1033  else
1034  {
1035  /*
1036  * Have caller finish split using default strategy, since page
1037  * does not appear to be the rightmost page for duplicates of the
1038  * value the page is filled with
1039  */
1040  }
1041  }
1042 
1043  return perfectpenalty;
1044 }
static void _bt_interval_edges(FindSplitData *state, SplitPoint **leftinterval, SplitPoint **rightinterval)
Definition: nbtsplitloc.c:1052

References _bt_interval_edges(), _bt_keep_natts_fast(), _bt_split_firstright(), _bt_split_lastleft(), IndexRelationGetNumberOfKeyAttributes, P_HIKEY, PageGetItem(), PageGetItemId(), SPLIT_DEFAULT, SPLIT_MANY_DUPLICATES, and SPLIT_SINGLE_VALUE.

Referenced by _bt_findsplitloc().