| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* nbtsplitloc.c |
| 4 |
* Choose split point code for Postgres btree implementation. |
| 5 |
* |
| 6 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 7 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 8 |
* |
| 9 |
* |
| 10 |
* IDENTIFICATION |
| 11 |
* src/backend/access/nbtree/nbtsplitloc.c |
| 12 |
* |
| 13 |
*------------------------------------------------------------------------- |
| 14 |
*/ |
| 15 |
#include "postgres.h" |
| 16 |
|
| 17 |
#include "access/nbtree.h" |
| 18 |
#include "common/int.h" |
| 19 |
|
| 20 |
typedef enum |
| 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 */ |
| 26 |
} FindSplitStrat; |
| 27 |
|
| 28 |
typedef struct |
| 29 |
{ |
| 30 |
/* details of free space left by split */ |
| 31 |
int16 curdelta; /* current leftfree/rightfree delta */ |
| 32 |
int16 leftfree; /* space left on left page post-split */ |
| 33 |
int16 rightfree; /* space left on right page post-split */ |
| 34 |
|
| 35 |
/* split point identifying fields (returned by _bt_findsplitloc) */ |
| 36 |
OffsetNumber firstrightoff; /* first origpage item on rightpage */ |
| 37 |
bool newitemonleft; /* new item goes on left, or right? */ |
| 38 |
} SplitPoint; |
| 39 |
|
| 40 |
typedef struct |
| 41 |
{ |
| 42 |
/* context data for _bt_recsplitloc */ |
| 43 |
Relation rel; /* index relation */ |
| 44 |
Page origpage; /* page undergoing split */ |
| 45 |
IndexTuple newitem; /* new item (cause of page split) */ |
| 46 |
Size newitemsz; /* size of newitem (includes line pointer) */ |
| 47 |
bool is_leaf; /* T if splitting a leaf page */ |
| 48 |
bool is_rightmost; /* T if splitting rightmost page on level */ |
| 49 |
OffsetNumber newitemoff; /* where the new item is to be inserted */ |
| 50 |
int leftspace; /* space available for items on left page */ |
| 51 |
int rightspace; /* space available for items on right page */ |
| 52 |
int olddataitemstotal; /* space taken by old items */ |
| 53 |
Size minfirstrightsz; /* smallest firstright size */ |
| 54 |
|
| 55 |
/* candidate split point data */ |
| 56 |
int maxsplits; /* maximum number of splits */ |
| 57 |
int nsplits; /* current number of splits */ |
| 58 |
SplitPoint *splits; /* all candidate split points for page */ |
| 59 |
int interval; /* current range of acceptable split points */ |
| 60 |
} FindSplitData; |
| 61 |
|
| 62 |
static void _bt_recsplitloc(FindSplitData *state, |
| 63 |
OffsetNumber firstrightoff, bool newitemonleft, |
| 64 |
int olddataitemstoleft, |
| 65 |
Size firstrightofforigpagetuplesz); |
| 66 |
static void _bt_deltasortsplits(FindSplitData *state, double fillfactormult, |
| 67 |
bool usemult); |
| 68 |
static int _bt_splitcmp(const void *arg1, const void *arg2); |
| 69 |
static bool _bt_afternewitemoff(FindSplitData *state, OffsetNumber maxoff, |
| 70 |
int leaffillfactor, bool *usemult); |
| 71 |
static bool _bt_adjacenthtid(ItemPointer lowhtid, ItemPointer highhtid); |
| 72 |
static OffsetNumber _bt_bestsplitloc(FindSplitData *state, int perfectpenalty, |
| 73 |
bool *newitemonleft, FindSplitStrat strategy); |
| 74 |
static int _bt_defaultinterval(FindSplitData *state); |
| 75 |
static int _bt_strategy(FindSplitData *state, SplitPoint *leftpage, |
| 76 |
SplitPoint *rightpage, FindSplitStrat *strategy); |
| 77 |
static void _bt_interval_edges(FindSplitData *state, |
| 78 |
SplitPoint **leftinterval, SplitPoint **rightinterval); |
| 79 |
static inline int _bt_split_penalty(FindSplitData *state, SplitPoint *split); |
| 80 |
static inline IndexTuple _bt_split_lastleft(FindSplitData *state, |
| 81 |
SplitPoint *split); |
| 82 |
static inline IndexTuple _bt_split_firstright(FindSplitData *state, |
| 83 |
SplitPoint *split); |
| 84 |
|
| 85 |
|
| 86 |
/* |
| 87 |
* _bt_findsplitloc() -- find an appropriate place to split a page. |
| 88 |
* |
| 89 |
* The main goal here is to equalize the free space that will be on each |
| 90 |
* split page, *after accounting for the inserted tuple*. (If we fail to |
| 91 |
* account for it, we might find ourselves with too little room on the page |
| 92 |
* that it needs to go into!) |
| 93 |
* |
| 94 |
* If the page is the rightmost page on its level, we instead try to arrange |
| 95 |
* to leave the left split page fillfactor% full. In this way, when we are |
| 96 |
* inserting successively increasing keys (consider sequences, timestamps, |
| 97 |
* etc) we will end up with a tree whose pages are about fillfactor% full, |
| 98 |
* instead of the 50% full result that we'd get without this special case. |
| 99 |
* This is the same as nbtsort.c produces for a newly-created tree. Note |
| 100 |
* that leaf and nonleaf pages use different fillfactors. Note also that |
| 101 |
* there are a number of further special cases where fillfactor is not |
| 102 |
* applied in the standard way. |
| 103 |
* |
| 104 |
* We are passed the intended insert position of the new tuple, expressed as |
| 105 |
* the offsetnumber of the tuple it must go in front of (this could be |
| 106 |
* maxoff+1 if the tuple is to go at the end). The new tuple itself is also |
| 107 |
* passed, since it's needed to give some weight to how effective suffix |
| 108 |
* truncation will be. The implementation picks the split point that |
| 109 |
* maximizes the effectiveness of suffix truncation from a small list of |
| 110 |
* alternative candidate split points that leave each side of the split with |
| 111 |
* about the same share of free space. Suffix truncation is secondary to |
| 112 |
* equalizing free space, except in cases with large numbers of duplicates. |
| 113 |
* Note that it is always assumed that caller goes on to perform truncation, |
| 114 |
* even with pg_upgrade'd indexes where that isn't actually the case |
| 115 |
* (!heapkeyspace indexes). See nbtree/README for more information about |
| 116 |
* suffix truncation. |
| 117 |
* |
| 118 |
* We return the index of the first existing tuple that should go on the |
| 119 |
* righthand page (which is called firstrightoff), plus a boolean |
| 120 |
* indicating whether the new tuple goes on the left or right page. You |
| 121 |
* can think of the returned state as a point _between_ two adjacent data |
| 122 |
* items (lastleft and firstright data items) on an imaginary version of |
| 123 |
* origpage that already includes newitem. The bool is necessary to |
| 124 |
* disambiguate the case where firstrightoff == newitemoff (i.e. it is |
| 125 |
* sometimes needed to determine if the firstright tuple for the split is |
| 126 |
* newitem rather than the tuple from origpage at offset firstrightoff). |
| 127 |
*/ |
| 128 |
OffsetNumber |
| 129 |
_bt_findsplitloc(Relation rel, |
| 130 |
Page origpage, |
| 131 |
OffsetNumber newitemoff, |
| 132 |
Size newitemsz, |
| 133 |
IndexTuple newitem, |
| 134 |
bool *newitemonleft) |
| 135 |
{ |
| 136 |
BTPageOpaque opaque; |
| 137 |
int leftspace, |
| 138 |
rightspace, |
| 139 |
olddataitemstotal, |
| 140 |
olddataitemstoleft, |
| 141 |
perfectpenalty, |
| 142 |
leaffillfactor; |
| 143 |
FindSplitData state; |
| 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 = |
| 159 |
PageGetPageSize(origpage) - SizeOfPageHeaderData - |
| 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\"", |
| 261 |
RelationGetRelationName(rel)); |
| 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 == |
| 402 |
IndexRelationGetNumberOfKeyAttributes(state.rel)); |
| 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 |
} |
| 429 |
|
| 430 |
/* |
| 431 |
* Subroutine to record a particular point between two tuples (possibly the |
| 432 |
* new item) on page (ie, combination of firstrightoff and newitemonleft |
| 433 |
* settings) in *state for later analysis. This is also a convenient point to |
| 434 |
* check if the split is legal (if it isn't, it won't be recorded). |
| 435 |
* |
| 436 |
* firstrightoff is the offset of the first item on the original page that |
| 437 |
* goes to the right page, and firstrightofforigpagetuplesz is the size of |
| 438 |
* that tuple. firstrightoff can be > max offset, which means that all the |
| 439 |
* old items go to the left page and only the new item goes to the right page. |
| 440 |
* We don't actually use firstrightofforigpagetuplesz in that case (actually, |
| 441 |
* we don't use it for _any_ split where the firstright tuple happens to be |
| 442 |
* newitem). |
| 443 |
* |
| 444 |
* olddataitemstoleft is the total size of all old items to the left of the |
| 445 |
* split point that is recorded here when legal. Should not include |
| 446 |
* newitemsz, since that is handled here. |
| 447 |
*/ |
| 448 |
static void |
| 449 |
_bt_recsplitloc(FindSplitData *state, |
| 450 |
OffsetNumber firstrightoff, |
| 451 |
bool newitemonleft, |
| 452 |
int olddataitemstoleft, |
| 453 |
Size firstrightofforigpagetuplesz) |
| 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 |
} |
| 560 |
|
| 561 |
/* |
| 562 |
* Subroutine to assign space deltas to materialized array of candidate split |
| 563 |
* points based on current fillfactor, and to sort array using that fillfactor |
| 564 |
*/ |
| 565 |
static void |
| 566 |
_bt_deltasortsplits(FindSplitData *state, double fillfactormult, |
| 567 |
bool usemult) |
| 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 |
} |
| 589 |
|
| 590 |
/* |
| 591 |
* qsort-style comparator used by _bt_deltasortsplits() |
| 592 |
*/ |
| 593 |
static int |
| 594 |
_bt_splitcmp(const void *arg1, const void *arg2) |
| 595 |
{ |
| 596 |
SplitPoint *split1 = (SplitPoint *) arg1; |
| 597 |
SplitPoint *split2 = (SplitPoint *) arg2; |
| 598 |
|
| 599 |
return pg_cmp_s16(split1->curdelta, split2->curdelta); |
| 600 |
} |
| 601 |
|
| 602 |
/* |
| 603 |
* Subroutine to determine whether or not a non-rightmost leaf page should be |
| 604 |
* split immediately after the would-be original page offset for the |
| 605 |
* new/incoming tuple (or should have leaf fillfactor applied when new item is |
| 606 |
* to the right on original page). This is appropriate when there is a |
| 607 |
* pattern of localized monotonically increasing insertions into a composite |
| 608 |
* index, where leading attribute values form local groupings, and we |
| 609 |
* anticipate further insertions of the same/current grouping (new item's |
| 610 |
* grouping) in the near future. This can be thought of as a variation on |
| 611 |
* applying leaf fillfactor during rightmost leaf page splits, since cases |
| 612 |
* that benefit will converge on packing leaf pages leaffillfactor% full over |
| 613 |
* time. |
| 614 |
* |
| 615 |
* We may leave extra free space remaining on the rightmost page of a "most |
| 616 |
* significant column" grouping of tuples if that grouping never ends up |
| 617 |
* having future insertions that use the free space. That effect is |
| 618 |
* self-limiting; a future grouping that becomes the "nearest on the right" |
| 619 |
* grouping of the affected grouping usually puts the extra free space to good |
| 620 |
* use. |
| 621 |
* |
| 622 |
* Caller uses optimization when routine returns true, though the exact action |
| 623 |
* taken by caller varies. Caller uses original leaf page fillfactor in |
| 624 |
* standard way rather than using the new item offset directly when *usemult |
| 625 |
* was also set to true here. Otherwise, caller applies optimization by |
| 626 |
* locating the legal split point that makes the new tuple the lastleft tuple |
| 627 |
* for the split. |
| 628 |
*/ |
| 629 |
static bool |
| 630 |
_bt_afternewitemoff(FindSplitData *state, OffsetNumber maxoff, |
| 631 |
int leaffillfactor, bool *usemult) |
| 632 |
{ |
| 633 |
int16 nkeyatts; |
| 634 |
ItemId itemid; |
| 635 |
IndexTuple tup; |
| 636 |
int keepnatts; |
| 637 |
|
| 638 |
Assert(state->is_leaf && !state->is_rightmost); |
| 639 |
|
| 640 |
nkeyatts = IndexRelationGetNumberOfKeyAttributes(state->rel); |
| 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 |
} |
| 740 |
|
| 741 |
/* |
| 742 |
* Subroutine for determining if two heap TIDS are "adjacent". |
| 743 |
* |
| 744 |
* Adjacent means that the high TID is very likely to have been inserted into |
| 745 |
* heap relation immediately after the low TID, probably during the current |
| 746 |
* transaction. |
| 747 |
*/ |
| 748 |
static bool |
| 749 |
_bt_adjacenthtid(ItemPointer lowhtid, ItemPointer highhtid) |
| 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 && |
| 763 |
ItemPointerGetOffsetNumber(highhtid) == FirstOffsetNumber) |
| 764 |
return true; |
| 765 |
|
| 766 |
return false; |
| 767 |
} |
| 768 |
|
| 769 |
/* |
| 770 |
* Subroutine to find the "best" split point among candidate split points. |
| 771 |
* The best split point is the split point with the lowest penalty among split |
| 772 |
* points that fall within current/final split interval. Penalty is an |
| 773 |
* abstract score, with a definition that varies depending on whether we're |
| 774 |
* splitting a leaf page or an internal page. See _bt_split_penalty() for |
| 775 |
* details. |
| 776 |
* |
| 777 |
* "perfectpenalty" is assumed to be the lowest possible penalty among |
| 778 |
* candidate split points. This allows us to return early without wasting |
| 779 |
* cycles on calculating the first differing attribute for all candidate |
| 780 |
* splits when that clearly cannot improve our choice (or when we only want a |
| 781 |
* minimally distinguishing split point, and don't want to make the split any |
| 782 |
* more unbalanced than is necessary). |
| 783 |
* |
| 784 |
* We return the index of the first existing tuple that should go on the right |
| 785 |
* page, plus a boolean indicating if new item is on left of split point. |
| 786 |
*/ |
| 787 |
static OffsetNumber |
| 788 |
_bt_bestsplitloc(FindSplitData *state, int perfectpenalty, |
| 789 |
bool *newitemonleft, FindSplitStrat strategy) |
| 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 |
} |
| 848 |
|
| 849 |
#define LEAF_SPLIT_DISTANCE 0.050 |
| 850 |
#define INTERNAL_SPLIT_DISTANCE 0.075 |
| 851 |
|
| 852 |
/* |
| 853 |
* Return a split interval to use for the default strategy. This is a limit |
| 854 |
* on the number of candidate split points to give further consideration to. |
| 855 |
* Only a fraction of all candidate splits points (those located at the start |
| 856 |
* of the now-sorted splits array) fall within the split interval. Split |
| 857 |
* interval is applied within _bt_bestsplitloc(). |
| 858 |
* |
| 859 |
* Split interval represents an acceptable range of split points -- those that |
| 860 |
* have leftfree and rightfree values that are acceptably balanced. The final |
| 861 |
* split point chosen is the split point with the lowest "penalty" among split |
| 862 |
* points in this split interval (unless we change our entire strategy, in |
| 863 |
* which case the interval also changes -- see _bt_strategy()). |
| 864 |
* |
| 865 |
* The "Prefix B-Trees" paper calls split interval sigma l for leaf splits, |
| 866 |
* and sigma b for internal ("branch") splits. It's hard to provide a |
| 867 |
* theoretical justification for the size of the split interval, though it's |
| 868 |
* clear that a small split interval can make tuples on level L+1 much smaller |
| 869 |
* on average, without noticeably affecting space utilization on level L. |
| 870 |
* (Note that the way that we calculate split interval might need to change if |
| 871 |
* suffix truncation is taught to truncate tuples "within" the last |
| 872 |
* attribute/datum for data types like text, which is more or less how it is |
| 873 |
* assumed to work in the paper.) |
| 874 |
*/ |
| 875 |
static int |
| 876 |
_bt_defaultinterval(FindSplitData *state) |
| 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 |
} |
| 921 |
|
| 922 |
/* |
| 923 |
* Subroutine to decide whether split should use default strategy/initial |
| 924 |
* split interval, or whether it should finish splitting the page using |
| 925 |
* alternative strategies (this is only possible with leaf pages). |
| 926 |
* |
| 927 |
* Caller uses alternative strategy (or sticks with default strategy) based |
| 928 |
* on how *strategy is set here. Return value is "perfect penalty", which is |
| 929 |
* passed to _bt_bestsplitloc() as a final constraint on how far caller is |
| 930 |
* willing to go to avoid appending a heap TID when using the many duplicates |
| 931 |
* strategy (it also saves _bt_bestsplitloc() useless cycles). |
| 932 |
*/ |
| 933 |
static int |
| 934 |
_bt_strategy(FindSplitData *state, SplitPoint *leftpage, |
| 935 |
SplitPoint *rightpage, FindSplitStrat *strategy) |
| 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 |
} |
| 1045 |
|
| 1046 |
/* |
| 1047 |
* Subroutine to locate leftmost and rightmost splits for current/default |
| 1048 |
* split interval. Note that it will be the same split iff there is only one |
| 1049 |
* split in interval. |
| 1050 |
*/ |
| 1051 |
static void |
| 1052 |
_bt_interval_edges(FindSplitData *state, SplitPoint **leftinterval, |
| 1053 |
SplitPoint **rightinterval) |
| 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 |
} |
| 1117 |
|
| 1118 |
/* |
| 1119 |
* Subroutine to find penalty for caller's candidate split point. |
| 1120 |
* |
| 1121 |
* On leaf pages, penalty is the attribute number that distinguishes each side |
| 1122 |
* of a split. It's the last attribute that needs to be included in new high |
| 1123 |
* key for left page. It can be greater than the number of key attributes in |
| 1124 |
* cases where a heap TID will need to be appended during truncation. |
| 1125 |
* |
| 1126 |
* On internal pages, penalty is simply the size of the firstright tuple for |
| 1127 |
* the split (including line pointer overhead). This tuple will become the |
| 1128 |
* new high key for the left page. |
| 1129 |
*/ |
| 1130 |
static inline int |
| 1131 |
_bt_split_penalty(FindSplitData *state, SplitPoint *split) |
| 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 |
} |
| 1154 |
|
| 1155 |
/* |
| 1156 |
* Subroutine to get a lastleft IndexTuple for a split point |
| 1157 |
*/ |
| 1158 |
static inline IndexTuple |
| 1159 |
_bt_split_lastleft(FindSplitData *state, SplitPoint *split) |
| 1160 |
{ |
| 1161 |
ItemId itemid; |
| 1162 |
|
| 1163 |
if (split->newitemonleft && split->firstrightoff == state->newitemoff) |
| 1164 |
return state->newitem; |
| 1165 |
|
| 1166 |
itemid = PageGetItemId(state->origpage, |
| 1167 |
OffsetNumberPrev(split->firstrightoff)); |
| 1168 |
return (IndexTuple) PageGetItem(state->origpage, itemid); |
| 1169 |
} |
| 1170 |
|
| 1171 |
/* |
| 1172 |
* Subroutine to get a firstright IndexTuple for a split point |
| 1173 |
*/ |
| 1174 |
static inline IndexTuple |
| 1175 |
_bt_split_firstright(FindSplitData *state, SplitPoint *split) |
| 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 |
} |