1 | /*------------------------------------------------------------------------- |
2 | * |
3 | * gindatapage.c |
4 | * routines for handling GIN posting tree pages. |
5 | * |
6 | * |
7 | * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group |
8 | * Portions Copyright (c) 1994, Regents of the University of California |
9 | * |
10 | * IDENTIFICATION |
11 | * src/backend/access/gin/gindatapage.c |
12 | *------------------------------------------------------------------------- |
13 | */ |
14 | |
15 | #include "postgres.h" |
16 | |
17 | #include "access/gin_private.h" |
18 | #include "access/ginxlog.h" |
19 | #include "access/xloginsert.h" |
20 | #include "lib/ilist.h" |
21 | #include "miscadmin.h" |
22 | #include "storage/predicate.h" |
23 | #include "utils/rel.h" |
24 | |
25 | /* |
26 | * Min, Max and Target size of posting lists stored on leaf pages, in bytes. |
27 | * |
28 | * The code can deal with any size, but random access is more efficient when |
29 | * a number of smaller lists are stored, rather than one big list. If a |
30 | * posting list would become larger than Max size as a result of insertions, |
31 | * it is split into two. If a posting list would be smaller than minimum |
32 | * size, it is merged with the next posting list. |
33 | */ |
34 | #define GinPostingListSegmentMaxSize 384 |
35 | #define GinPostingListSegmentTargetSize 256 |
36 | #define GinPostingListSegmentMinSize 128 |
37 | |
38 | /* |
39 | * At least this many items fit in a GinPostingListSegmentMaxSize-bytes |
40 | * long segment. This is used when estimating how much space is required |
41 | * for N items, at minimum. |
42 | */ |
43 | #define MinTuplesPerSegment ((GinPostingListSegmentMaxSize - 2) / 6) |
44 | |
45 | /* |
46 | * A working struct for manipulating a posting tree leaf page. |
47 | */ |
48 | typedef struct |
49 | { |
50 | dlist_head segments; /* a list of leafSegmentInfos */ |
51 | |
52 | /* |
53 | * The following fields represent how the segments are split across pages, |
54 | * if a page split is required. Filled in by leafRepackItems. |
55 | */ |
56 | dlist_node *lastleft; /* last segment on left page */ |
57 | int lsize; /* total size on left page */ |
58 | int rsize; /* total size on right page */ |
59 | |
60 | bool oldformat; /* page is in pre-9.4 format on disk */ |
61 | |
62 | /* |
63 | * If we need WAL data representing the reconstructed leaf page, it's |
64 | * stored here by computeLeafRecompressWALData. |
65 | */ |
66 | char *walinfo; /* buffer start */ |
67 | int walinfolen; /* and length */ |
68 | } disassembledLeaf; |
69 | |
70 | typedef struct |
71 | { |
72 | dlist_node node; /* linked list pointers */ |
73 | |
74 | /*------------- |
75 | * 'action' indicates the status of this in-memory segment, compared to |
76 | * what's on disk. It is one of the GIN_SEGMENT_* action codes: |
77 | * |
78 | * UNMODIFIED no changes |
79 | * DELETE the segment is to be removed. 'seg' and 'items' are |
80 | * ignored |
81 | * INSERT this is a completely new segment |
82 | * REPLACE this replaces an existing segment with new content |
83 | * ADDITEMS like REPLACE, but no items have been removed, and we track |
84 | * in detail what items have been added to this segment, in |
85 | * 'modifieditems' |
86 | *------------- |
87 | */ |
88 | char action; |
89 | |
90 | ItemPointerData *modifieditems; |
91 | uint16 nmodifieditems; |
92 | |
93 | /* |
94 | * The following fields represent the items in this segment. If 'items' is |
95 | * not NULL, it contains a palloc'd array of the itemsin this segment. If |
96 | * 'seg' is not NULL, it contains the items in an already-compressed |
97 | * format. It can point to an on-disk page (!modified), or a palloc'd |
98 | * segment in memory. If both are set, they must represent the same items. |
99 | */ |
100 | GinPostingList *seg; |
101 | ItemPointer items; |
102 | int nitems; /* # of items in 'items', if items != NULL */ |
103 | } leafSegmentInfo; |
104 | |
105 | static ItemPointer dataLeafPageGetUncompressed(Page page, int *nitems); |
106 | static void dataSplitPageInternal(GinBtree btree, Buffer origbuf, |
107 | GinBtreeStack *stack, |
108 | void *insertdata, BlockNumber updateblkno, |
109 | Page *newlpage, Page *newrpage); |
110 | |
111 | static disassembledLeaf *disassembleLeaf(Page page); |
112 | static bool leafRepackItems(disassembledLeaf *leaf, ItemPointer remaining); |
113 | static bool addItemsToLeaf(disassembledLeaf *leaf, ItemPointer newItems, |
114 | int nNewItems); |
115 | |
116 | static void computeLeafRecompressWALData(disassembledLeaf *leaf); |
117 | static void dataPlaceToPageLeafRecompress(Buffer buf, disassembledLeaf *leaf); |
118 | static void dataPlaceToPageLeafSplit(disassembledLeaf *leaf, |
119 | ItemPointerData lbound, ItemPointerData rbound, |
120 | Page lpage, Page rpage); |
121 | |
122 | /* |
123 | * Read TIDs from leaf data page to single uncompressed array. The TIDs are |
124 | * returned in ascending order. |
125 | * |
126 | * advancePast is a hint, indicating that the caller is only interested in |
127 | * TIDs > advancePast. To return all items, use ItemPointerSetMin. |
128 | * |
129 | * Note: This function can still return items smaller than advancePast that |
130 | * are in the same posting list as the items of interest, so the caller must |
131 | * still check all the returned items. But passing it allows this function to |
132 | * skip whole posting lists. |
133 | */ |
134 | ItemPointer |
135 | GinDataLeafPageGetItems(Page page, int *nitems, ItemPointerData advancePast) |
136 | { |
137 | ItemPointer result; |
138 | |
139 | if (GinPageIsCompressed(page)) |
140 | { |
141 | GinPostingList *seg = GinDataLeafPageGetPostingList(page); |
142 | Size len = GinDataLeafPageGetPostingListSize(page); |
143 | Pointer endptr = ((Pointer) seg) + len; |
144 | GinPostingList *next; |
145 | |
146 | /* Skip to the segment containing advancePast+1 */ |
147 | if (ItemPointerIsValid(&advancePast)) |
148 | { |
149 | next = GinNextPostingListSegment(seg); |
150 | while ((Pointer) next < endptr && |
151 | ginCompareItemPointers(&next->first, &advancePast) <= 0) |
152 | { |
153 | seg = next; |
154 | next = GinNextPostingListSegment(seg); |
155 | } |
156 | len = endptr - (Pointer) seg; |
157 | } |
158 | |
159 | if (len > 0) |
160 | result = ginPostingListDecodeAllSegments(seg, len, nitems); |
161 | else |
162 | { |
163 | result = NULL; |
164 | *nitems = 0; |
165 | } |
166 | } |
167 | else |
168 | { |
169 | ItemPointer tmp = dataLeafPageGetUncompressed(page, nitems); |
170 | |
171 | result = palloc((*nitems) * sizeof(ItemPointerData)); |
172 | memcpy(result, tmp, (*nitems) * sizeof(ItemPointerData)); |
173 | } |
174 | |
175 | return result; |
176 | } |
177 | |
178 | /* |
179 | * Places all TIDs from leaf data page to bitmap. |
180 | */ |
181 | int |
182 | GinDataLeafPageGetItemsToTbm(Page page, TIDBitmap *tbm) |
183 | { |
184 | ItemPointer uncompressed; |
185 | int nitems; |
186 | |
187 | if (GinPageIsCompressed(page)) |
188 | { |
189 | GinPostingList *segment = GinDataLeafPageGetPostingList(page); |
190 | Size len = GinDataLeafPageGetPostingListSize(page); |
191 | |
192 | nitems = ginPostingListDecodeAllSegmentsToTbm(segment, len, tbm); |
193 | } |
194 | else |
195 | { |
196 | uncompressed = dataLeafPageGetUncompressed(page, &nitems); |
197 | |
198 | if (nitems > 0) |
199 | tbm_add_tuples(tbm, uncompressed, nitems, false); |
200 | } |
201 | |
202 | return nitems; |
203 | } |
204 | |
205 | /* |
206 | * Get pointer to the uncompressed array of items on a pre-9.4 format |
207 | * uncompressed leaf page. The number of items in the array is returned in |
208 | * *nitems. |
209 | */ |
210 | static ItemPointer |
211 | dataLeafPageGetUncompressed(Page page, int *nitems) |
212 | { |
213 | ItemPointer items; |
214 | |
215 | Assert(!GinPageIsCompressed(page)); |
216 | |
217 | /* |
218 | * In the old pre-9.4 page format, the whole page content is used for |
219 | * uncompressed items, and the number of items is stored in 'maxoff' |
220 | */ |
221 | items = (ItemPointer) GinDataPageGetData(page); |
222 | *nitems = GinPageGetOpaque(page)->maxoff; |
223 | |
224 | return items; |
225 | } |
226 | |
227 | /* |
228 | * Check if we should follow the right link to find the item we're searching |
229 | * for. |
230 | * |
231 | * Compares inserting item pointer with the right bound of the current page. |
232 | */ |
233 | static bool |
234 | dataIsMoveRight(GinBtree btree, Page page) |
235 | { |
236 | ItemPointer iptr = GinDataPageGetRightBound(page); |
237 | |
238 | if (GinPageRightMost(page)) |
239 | return false; |
240 | |
241 | return (ginCompareItemPointers(&btree->itemptr, iptr) > 0) ? true : false; |
242 | } |
243 | |
244 | /* |
245 | * Find correct PostingItem in non-leaf page. It is assumed that this is |
246 | * the correct page, and the searched value SHOULD be on the page. |
247 | */ |
248 | static BlockNumber |
249 | dataLocateItem(GinBtree btree, GinBtreeStack *stack) |
250 | { |
251 | OffsetNumber low, |
252 | high, |
253 | maxoff; |
254 | PostingItem *pitem = NULL; |
255 | int result; |
256 | Page page = BufferGetPage(stack->buffer); |
257 | |
258 | Assert(!GinPageIsLeaf(page)); |
259 | Assert(GinPageIsData(page)); |
260 | |
261 | if (btree->fullScan) |
262 | { |
263 | stack->off = FirstOffsetNumber; |
264 | stack->predictNumber *= GinPageGetOpaque(page)->maxoff; |
265 | return btree->getLeftMostChild(btree, page); |
266 | } |
267 | |
268 | low = FirstOffsetNumber; |
269 | maxoff = high = GinPageGetOpaque(page)->maxoff; |
270 | Assert(high >= low); |
271 | |
272 | high++; |
273 | |
274 | while (high > low) |
275 | { |
276 | OffsetNumber mid = low + ((high - low) / 2); |
277 | |
278 | pitem = GinDataPageGetPostingItem(page, mid); |
279 | |
280 | if (mid == maxoff) |
281 | { |
282 | /* |
283 | * Right infinity, page already correctly chosen with a help of |
284 | * dataIsMoveRight |
285 | */ |
286 | result = -1; |
287 | } |
288 | else |
289 | { |
290 | pitem = GinDataPageGetPostingItem(page, mid); |
291 | result = ginCompareItemPointers(&btree->itemptr, &(pitem->key)); |
292 | } |
293 | |
294 | if (result == 0) |
295 | { |
296 | stack->off = mid; |
297 | return PostingItemGetBlockNumber(pitem); |
298 | } |
299 | else if (result > 0) |
300 | low = mid + 1; |
301 | else |
302 | high = mid; |
303 | } |
304 | |
305 | Assert(high >= FirstOffsetNumber && high <= maxoff); |
306 | |
307 | stack->off = high; |
308 | pitem = GinDataPageGetPostingItem(page, high); |
309 | return PostingItemGetBlockNumber(pitem); |
310 | } |
311 | |
312 | /* |
313 | * Find link to blkno on non-leaf page, returns offset of PostingItem |
314 | */ |
315 | static OffsetNumber |
316 | dataFindChildPtr(GinBtree btree, Page page, BlockNumber blkno, OffsetNumber storedOff) |
317 | { |
318 | OffsetNumber i, |
319 | maxoff = GinPageGetOpaque(page)->maxoff; |
320 | PostingItem *pitem; |
321 | |
322 | Assert(!GinPageIsLeaf(page)); |
323 | Assert(GinPageIsData(page)); |
324 | |
325 | /* if page isn't changed, we return storedOff */ |
326 | if (storedOff >= FirstOffsetNumber && storedOff <= maxoff) |
327 | { |
328 | pitem = GinDataPageGetPostingItem(page, storedOff); |
329 | if (PostingItemGetBlockNumber(pitem) == blkno) |
330 | return storedOff; |
331 | |
332 | /* |
333 | * we hope, that needed pointer goes to right. It's true if there |
334 | * wasn't a deletion |
335 | */ |
336 | for (i = storedOff + 1; i <= maxoff; i++) |
337 | { |
338 | pitem = GinDataPageGetPostingItem(page, i); |
339 | if (PostingItemGetBlockNumber(pitem) == blkno) |
340 | return i; |
341 | } |
342 | |
343 | maxoff = storedOff - 1; |
344 | } |
345 | |
346 | /* last chance */ |
347 | for (i = FirstOffsetNumber; i <= maxoff; i++) |
348 | { |
349 | pitem = GinDataPageGetPostingItem(page, i); |
350 | if (PostingItemGetBlockNumber(pitem) == blkno) |
351 | return i; |
352 | } |
353 | |
354 | return InvalidOffsetNumber; |
355 | } |
356 | |
357 | /* |
358 | * Return blkno of leftmost child |
359 | */ |
360 | static BlockNumber |
361 | dataGetLeftMostPage(GinBtree btree, Page page) |
362 | { |
363 | PostingItem *pitem; |
364 | |
365 | Assert(!GinPageIsLeaf(page)); |
366 | Assert(GinPageIsData(page)); |
367 | Assert(GinPageGetOpaque(page)->maxoff >= FirstOffsetNumber); |
368 | |
369 | pitem = GinDataPageGetPostingItem(page, FirstOffsetNumber); |
370 | return PostingItemGetBlockNumber(pitem); |
371 | } |
372 | |
373 | /* |
374 | * Add PostingItem to a non-leaf page. |
375 | */ |
376 | void |
377 | GinDataPageAddPostingItem(Page page, PostingItem *data, OffsetNumber offset) |
378 | { |
379 | OffsetNumber maxoff = GinPageGetOpaque(page)->maxoff; |
380 | char *ptr; |
381 | |
382 | Assert(PostingItemGetBlockNumber(data) != InvalidBlockNumber); |
383 | Assert(!GinPageIsLeaf(page)); |
384 | |
385 | if (offset == InvalidOffsetNumber) |
386 | { |
387 | ptr = (char *) GinDataPageGetPostingItem(page, maxoff + 1); |
388 | } |
389 | else |
390 | { |
391 | ptr = (char *) GinDataPageGetPostingItem(page, offset); |
392 | if (offset != maxoff + 1) |
393 | memmove(ptr + sizeof(PostingItem), |
394 | ptr, |
395 | (maxoff - offset + 1) * sizeof(PostingItem)); |
396 | } |
397 | memcpy(ptr, data, sizeof(PostingItem)); |
398 | |
399 | maxoff++; |
400 | GinPageGetOpaque(page)->maxoff = maxoff; |
401 | |
402 | /* |
403 | * Also set pd_lower to the end of the posting items, to follow the |
404 | * "standard" page layout, so that we can squeeze out the unused space |
405 | * from full-page images. |
406 | */ |
407 | GinDataPageSetDataSize(page, maxoff * sizeof(PostingItem)); |
408 | } |
409 | |
410 | /* |
411 | * Delete posting item from non-leaf page |
412 | */ |
413 | void |
414 | GinPageDeletePostingItem(Page page, OffsetNumber offset) |
415 | { |
416 | OffsetNumber maxoff = GinPageGetOpaque(page)->maxoff; |
417 | |
418 | Assert(!GinPageIsLeaf(page)); |
419 | Assert(offset >= FirstOffsetNumber && offset <= maxoff); |
420 | |
421 | if (offset != maxoff) |
422 | memmove(GinDataPageGetPostingItem(page, offset), |
423 | GinDataPageGetPostingItem(page, offset + 1), |
424 | sizeof(PostingItem) * (maxoff - offset)); |
425 | |
426 | maxoff--; |
427 | GinPageGetOpaque(page)->maxoff = maxoff; |
428 | |
429 | GinDataPageSetDataSize(page, maxoff * sizeof(PostingItem)); |
430 | } |
431 | |
432 | /* |
433 | * Prepare to insert data on a leaf data page. |
434 | * |
435 | * If it will fit, return GPTP_INSERT after doing whatever setup is needed |
436 | * before we enter the insertion critical section. *ptp_workspace can be |
437 | * set to pass information along to the execPlaceToPage function. |
438 | * |
439 | * If it won't fit, perform a page split and return two temporary page |
440 | * images into *newlpage and *newrpage, with result GPTP_SPLIT. |
441 | * |
442 | * In neither case should the given page buffer be modified here. |
443 | */ |
444 | static GinPlaceToPageRC |
445 | dataBeginPlaceToPageLeaf(GinBtree btree, Buffer buf, GinBtreeStack *stack, |
446 | void *insertdata, |
447 | void **ptp_workspace, |
448 | Page *newlpage, Page *newrpage) |
449 | { |
450 | GinBtreeDataLeafInsertData *items = insertdata; |
451 | ItemPointer newItems = &items->items[items->curitem]; |
452 | int maxitems = items->nitem - items->curitem; |
453 | Page page = BufferGetPage(buf); |
454 | int i; |
455 | ItemPointerData rbound; |
456 | ItemPointerData lbound; |
457 | bool needsplit; |
458 | bool append; |
459 | int segsize; |
460 | Size freespace; |
461 | disassembledLeaf *leaf; |
462 | leafSegmentInfo *lastleftinfo; |
463 | ItemPointerData maxOldItem; |
464 | ItemPointerData remaining; |
465 | |
466 | rbound = *GinDataPageGetRightBound(page); |
467 | |
468 | /* |
469 | * Count how many of the new items belong to this page. |
470 | */ |
471 | if (!GinPageRightMost(page)) |
472 | { |
473 | for (i = 0; i < maxitems; i++) |
474 | { |
475 | if (ginCompareItemPointers(&newItems[i], &rbound) > 0) |
476 | { |
477 | /* |
478 | * This needs to go to some other location in the tree. (The |
479 | * caller should've chosen the insert location so that at |
480 | * least the first item goes here.) |
481 | */ |
482 | Assert(i > 0); |
483 | break; |
484 | } |
485 | } |
486 | maxitems = i; |
487 | } |
488 | |
489 | /* Disassemble the data on the page */ |
490 | leaf = disassembleLeaf(page); |
491 | |
492 | /* |
493 | * Are we appending to the end of the page? IOW, are all the new items |
494 | * larger than any of the existing items. |
495 | */ |
496 | if (!dlist_is_empty(&leaf->segments)) |
497 | { |
498 | lastleftinfo = dlist_container(leafSegmentInfo, node, |
499 | dlist_tail_node(&leaf->segments)); |
500 | if (!lastleftinfo->items) |
501 | lastleftinfo->items = ginPostingListDecode(lastleftinfo->seg, |
502 | &lastleftinfo->nitems); |
503 | maxOldItem = lastleftinfo->items[lastleftinfo->nitems - 1]; |
504 | if (ginCompareItemPointers(&newItems[0], &maxOldItem) >= 0) |
505 | append = true; |
506 | else |
507 | append = false; |
508 | } |
509 | else |
510 | { |
511 | ItemPointerSetMin(&maxOldItem); |
512 | append = true; |
513 | } |
514 | |
515 | /* |
516 | * If we're appending to the end of the page, we will append as many items |
517 | * as we can fit (after splitting), and stop when the pages becomes full. |
518 | * Otherwise we have to limit the number of new items to insert, because |
519 | * once we start packing we can't just stop when we run out of space, |
520 | * because we must make sure that all the old items still fit. |
521 | */ |
522 | if (GinPageIsCompressed(page)) |
523 | freespace = GinDataLeafPageGetFreeSpace(page); |
524 | else |
525 | freespace = 0; |
526 | if (append) |
527 | { |
528 | /* |
529 | * Even when appending, trying to append more items than will fit is |
530 | * not completely free, because we will merge the new items and old |
531 | * items into an array below. In the best case, every new item fits in |
532 | * a single byte, and we can use all the free space on the old page as |
533 | * well as the new page. For simplicity, ignore segment overhead etc. |
534 | */ |
535 | maxitems = Min(maxitems, freespace + GinDataPageMaxDataSize); |
536 | } |
537 | else |
538 | { |
539 | /* |
540 | * Calculate a conservative estimate of how many new items we can fit |
541 | * on the two pages after splitting. |
542 | * |
543 | * We can use any remaining free space on the old page to store full |
544 | * segments, as well as the new page. Each full-sized segment can hold |
545 | * at least MinTuplesPerSegment items |
546 | */ |
547 | int nnewsegments; |
548 | |
549 | nnewsegments = freespace / GinPostingListSegmentMaxSize; |
550 | nnewsegments += GinDataPageMaxDataSize / GinPostingListSegmentMaxSize; |
551 | maxitems = Min(maxitems, nnewsegments * MinTuplesPerSegment); |
552 | } |
553 | |
554 | /* Add the new items to the segment list */ |
555 | if (!addItemsToLeaf(leaf, newItems, maxitems)) |
556 | { |
557 | /* all items were duplicates, we have nothing to do */ |
558 | items->curitem += maxitems; |
559 | |
560 | return GPTP_NO_WORK; |
561 | } |
562 | |
563 | /* |
564 | * Pack the items back to compressed segments, ready for writing to disk. |
565 | */ |
566 | needsplit = leafRepackItems(leaf, &remaining); |
567 | |
568 | /* |
569 | * Did all the new items fit? |
570 | * |
571 | * If we're appending, it's OK if they didn't. But as a sanity check, |
572 | * verify that all the old items fit. |
573 | */ |
574 | if (ItemPointerIsValid(&remaining)) |
575 | { |
576 | if (!append || ItemPointerCompare(&maxOldItem, &remaining) >= 0) |
577 | elog(ERROR, "could not split GIN page; all old items didn't fit" ); |
578 | |
579 | /* Count how many of the new items did fit. */ |
580 | for (i = 0; i < maxitems; i++) |
581 | { |
582 | if (ginCompareItemPointers(&newItems[i], &remaining) >= 0) |
583 | break; |
584 | } |
585 | if (i == 0) |
586 | elog(ERROR, "could not split GIN page; no new items fit" ); |
587 | maxitems = i; |
588 | } |
589 | |
590 | if (!needsplit) |
591 | { |
592 | /* |
593 | * Great, all the items fit on a single page. If needed, prepare data |
594 | * for a WAL record describing the changes we'll make. |
595 | */ |
596 | if (RelationNeedsWAL(btree->index) && !btree->isBuild) |
597 | computeLeafRecompressWALData(leaf); |
598 | |
599 | /* |
600 | * We're ready to enter the critical section, but |
601 | * dataExecPlaceToPageLeaf will need access to the "leaf" data. |
602 | */ |
603 | *ptp_workspace = leaf; |
604 | |
605 | if (append) |
606 | elog(DEBUG2, "appended %d new items to block %u; %d bytes (%d to go)" , |
607 | maxitems, BufferGetBlockNumber(buf), (int) leaf->lsize, |
608 | items->nitem - items->curitem - maxitems); |
609 | else |
610 | elog(DEBUG2, "inserted %d new items to block %u; %d bytes (%d to go)" , |
611 | maxitems, BufferGetBlockNumber(buf), (int) leaf->lsize, |
612 | items->nitem - items->curitem - maxitems); |
613 | } |
614 | else |
615 | { |
616 | /* |
617 | * Have to split. |
618 | * |
619 | * leafRepackItems already divided the segments between the left and |
620 | * the right page. It filled the left page as full as possible, and |
621 | * put the rest to the right page. When building a new index, that's |
622 | * good, because the table is scanned from beginning to end and there |
623 | * won't be any more insertions to the left page during the build. |
624 | * This packs the index as tight as possible. But otherwise, split |
625 | * 50/50, by moving segments from the left page to the right page |
626 | * until they're balanced. |
627 | * |
628 | * As a further heuristic, when appending items to the end of the |
629 | * page, try to make the left page 75% full, on the assumption that |
630 | * subsequent insertions will probably also go to the end. This packs |
631 | * the index somewhat tighter when appending to a table, which is very |
632 | * common. |
633 | */ |
634 | if (!btree->isBuild) |
635 | { |
636 | while (dlist_has_prev(&leaf->segments, leaf->lastleft)) |
637 | { |
638 | lastleftinfo = dlist_container(leafSegmentInfo, node, leaf->lastleft); |
639 | |
640 | /* ignore deleted segments */ |
641 | if (lastleftinfo->action != GIN_SEGMENT_DELETE) |
642 | { |
643 | segsize = SizeOfGinPostingList(lastleftinfo->seg); |
644 | |
645 | /* |
646 | * Note that we check that the right page doesn't become |
647 | * more full than the left page even when appending. It's |
648 | * possible that we added enough items to make both pages |
649 | * more than 75% full. |
650 | */ |
651 | if ((leaf->lsize - segsize) - (leaf->rsize + segsize) < 0) |
652 | break; |
653 | if (append) |
654 | { |
655 | if ((leaf->lsize - segsize) < (BLCKSZ * 3) / 4) |
656 | break; |
657 | } |
658 | |
659 | leaf->lsize -= segsize; |
660 | leaf->rsize += segsize; |
661 | } |
662 | leaf->lastleft = dlist_prev_node(&leaf->segments, leaf->lastleft); |
663 | } |
664 | } |
665 | Assert(leaf->lsize <= GinDataPageMaxDataSize); |
666 | Assert(leaf->rsize <= GinDataPageMaxDataSize); |
667 | |
668 | /* |
669 | * Fetch the max item in the left page's last segment; it becomes the |
670 | * right bound of the page. |
671 | */ |
672 | lastleftinfo = dlist_container(leafSegmentInfo, node, leaf->lastleft); |
673 | if (!lastleftinfo->items) |
674 | lastleftinfo->items = ginPostingListDecode(lastleftinfo->seg, |
675 | &lastleftinfo->nitems); |
676 | lbound = lastleftinfo->items[lastleftinfo->nitems - 1]; |
677 | |
678 | /* |
679 | * Now allocate a couple of temporary page images, and fill them. |
680 | */ |
681 | *newlpage = palloc(BLCKSZ); |
682 | *newrpage = palloc(BLCKSZ); |
683 | |
684 | dataPlaceToPageLeafSplit(leaf, lbound, rbound, |
685 | *newlpage, *newrpage); |
686 | |
687 | Assert(GinPageRightMost(page) || |
688 | ginCompareItemPointers(GinDataPageGetRightBound(*newlpage), |
689 | GinDataPageGetRightBound(*newrpage)) < 0); |
690 | |
691 | if (append) |
692 | elog(DEBUG2, "appended %d items to block %u; split %d/%d (%d to go)" , |
693 | maxitems, BufferGetBlockNumber(buf), (int) leaf->lsize, (int) leaf->rsize, |
694 | items->nitem - items->curitem - maxitems); |
695 | else |
696 | elog(DEBUG2, "inserted %d items to block %u; split %d/%d (%d to go)" , |
697 | maxitems, BufferGetBlockNumber(buf), (int) leaf->lsize, (int) leaf->rsize, |
698 | items->nitem - items->curitem - maxitems); |
699 | } |
700 | |
701 | items->curitem += maxitems; |
702 | |
703 | return needsplit ? GPTP_SPLIT : GPTP_INSERT; |
704 | } |
705 | |
706 | /* |
707 | * Perform data insertion after beginPlaceToPage has decided it will fit. |
708 | * |
709 | * This is invoked within a critical section, and XLOG record creation (if |
710 | * needed) is already started. The target buffer is registered in slot 0. |
711 | */ |
712 | static void |
713 | dataExecPlaceToPageLeaf(GinBtree btree, Buffer buf, GinBtreeStack *stack, |
714 | void *insertdata, void *ptp_workspace) |
715 | { |
716 | disassembledLeaf *leaf = (disassembledLeaf *) ptp_workspace; |
717 | |
718 | /* Apply changes to page */ |
719 | dataPlaceToPageLeafRecompress(buf, leaf); |
720 | |
721 | /* If needed, register WAL data built by computeLeafRecompressWALData */ |
722 | if (RelationNeedsWAL(btree->index) && !btree->isBuild) |
723 | { |
724 | XLogRegisterBufData(0, leaf->walinfo, leaf->walinfolen); |
725 | } |
726 | } |
727 | |
728 | /* |
729 | * Vacuum a posting tree leaf page. |
730 | */ |
731 | void |
732 | ginVacuumPostingTreeLeaf(Relation indexrel, Buffer buffer, GinVacuumState *gvs) |
733 | { |
734 | Page page = BufferGetPage(buffer); |
735 | disassembledLeaf *leaf; |
736 | bool removedsomething = false; |
737 | dlist_iter iter; |
738 | |
739 | leaf = disassembleLeaf(page); |
740 | |
741 | /* Vacuum each segment. */ |
742 | dlist_foreach(iter, &leaf->segments) |
743 | { |
744 | leafSegmentInfo *seginfo = dlist_container(leafSegmentInfo, node, iter.cur); |
745 | int oldsegsize; |
746 | ItemPointer cleaned; |
747 | int ncleaned; |
748 | |
749 | if (!seginfo->items) |
750 | seginfo->items = ginPostingListDecode(seginfo->seg, |
751 | &seginfo->nitems); |
752 | if (seginfo->seg) |
753 | oldsegsize = SizeOfGinPostingList(seginfo->seg); |
754 | else |
755 | oldsegsize = GinDataPageMaxDataSize; |
756 | |
757 | cleaned = ginVacuumItemPointers(gvs, |
758 | seginfo->items, |
759 | seginfo->nitems, |
760 | &ncleaned); |
761 | pfree(seginfo->items); |
762 | seginfo->items = NULL; |
763 | seginfo->nitems = 0; |
764 | if (cleaned) |
765 | { |
766 | if (ncleaned > 0) |
767 | { |
768 | int npacked; |
769 | |
770 | seginfo->seg = ginCompressPostingList(cleaned, |
771 | ncleaned, |
772 | oldsegsize, |
773 | &npacked); |
774 | /* Removing an item never increases the size of the segment */ |
775 | if (npacked != ncleaned) |
776 | elog(ERROR, "could not fit vacuumed posting list" ); |
777 | seginfo->action = GIN_SEGMENT_REPLACE; |
778 | } |
779 | else |
780 | { |
781 | seginfo->seg = NULL; |
782 | seginfo->items = NULL; |
783 | seginfo->action = GIN_SEGMENT_DELETE; |
784 | } |
785 | seginfo->nitems = ncleaned; |
786 | |
787 | removedsomething = true; |
788 | } |
789 | } |
790 | |
791 | /* |
792 | * If we removed any items, reconstruct the page from the pieces. |
793 | * |
794 | * We don't try to re-encode the segments here, even though some of them |
795 | * might be really small now that we've removed some items from them. It |
796 | * seems like a waste of effort, as there isn't really any benefit from |
797 | * larger segments per se; larger segments only help to pack more items in |
798 | * the same space. We might as well delay doing that until the next |
799 | * insertion, which will need to re-encode at least part of the page |
800 | * anyway. |
801 | * |
802 | * Also note if the page was in uncompressed, pre-9.4 format before, it is |
803 | * now represented as one huge segment that contains all the items. It |
804 | * might make sense to split that, to speed up random access, but we don't |
805 | * bother. You'll have to REINDEX anyway if you want the full gain of the |
806 | * new tighter index format. |
807 | */ |
808 | if (removedsomething) |
809 | { |
810 | bool modified; |
811 | |
812 | /* |
813 | * Make sure we have a palloc'd copy of all segments, after the first |
814 | * segment that is modified. (dataPlaceToPageLeafRecompress requires |
815 | * this). |
816 | */ |
817 | modified = false; |
818 | dlist_foreach(iter, &leaf->segments) |
819 | { |
820 | leafSegmentInfo *seginfo = dlist_container(leafSegmentInfo, node, |
821 | iter.cur); |
822 | |
823 | if (seginfo->action != GIN_SEGMENT_UNMODIFIED) |
824 | modified = true; |
825 | if (modified && seginfo->action != GIN_SEGMENT_DELETE) |
826 | { |
827 | int segsize = SizeOfGinPostingList(seginfo->seg); |
828 | GinPostingList *tmp = (GinPostingList *) palloc(segsize); |
829 | |
830 | memcpy(tmp, seginfo->seg, segsize); |
831 | seginfo->seg = tmp; |
832 | } |
833 | } |
834 | |
835 | if (RelationNeedsWAL(indexrel)) |
836 | computeLeafRecompressWALData(leaf); |
837 | |
838 | /* Apply changes to page */ |
839 | START_CRIT_SECTION(); |
840 | |
841 | dataPlaceToPageLeafRecompress(buffer, leaf); |
842 | |
843 | MarkBufferDirty(buffer); |
844 | |
845 | if (RelationNeedsWAL(indexrel)) |
846 | { |
847 | XLogRecPtr recptr; |
848 | |
849 | XLogBeginInsert(); |
850 | XLogRegisterBuffer(0, buffer, REGBUF_STANDARD); |
851 | XLogRegisterBufData(0, leaf->walinfo, leaf->walinfolen); |
852 | recptr = XLogInsert(RM_GIN_ID, XLOG_GIN_VACUUM_DATA_LEAF_PAGE); |
853 | PageSetLSN(page, recptr); |
854 | } |
855 | |
856 | END_CRIT_SECTION(); |
857 | } |
858 | } |
859 | |
860 | /* |
861 | * Construct a ginxlogRecompressDataLeaf record representing the changes |
862 | * in *leaf. (Because this requires a palloc, we have to do it before |
863 | * we enter the critical section that actually updates the page.) |
864 | */ |
865 | static void |
866 | computeLeafRecompressWALData(disassembledLeaf *leaf) |
867 | { |
868 | int nmodified = 0; |
869 | char *walbufbegin; |
870 | char *walbufend; |
871 | dlist_iter iter; |
872 | int segno; |
873 | ginxlogRecompressDataLeaf *recompress_xlog; |
874 | |
875 | /* Count the modified segments */ |
876 | dlist_foreach(iter, &leaf->segments) |
877 | { |
878 | leafSegmentInfo *seginfo = dlist_container(leafSegmentInfo, node, |
879 | iter.cur); |
880 | |
881 | if (seginfo->action != GIN_SEGMENT_UNMODIFIED) |
882 | nmodified++; |
883 | } |
884 | |
885 | walbufbegin = |
886 | palloc(sizeof(ginxlogRecompressDataLeaf) + |
887 | BLCKSZ + /* max size needed to hold the segment data */ |
888 | nmodified * 2 /* (segno + action) per action */ |
889 | ); |
890 | walbufend = walbufbegin; |
891 | |
892 | recompress_xlog = (ginxlogRecompressDataLeaf *) walbufend; |
893 | walbufend += sizeof(ginxlogRecompressDataLeaf); |
894 | |
895 | recompress_xlog->nactions = nmodified; |
896 | |
897 | segno = 0; |
898 | dlist_foreach(iter, &leaf->segments) |
899 | { |
900 | leafSegmentInfo *seginfo = dlist_container(leafSegmentInfo, node, |
901 | iter.cur); |
902 | int segsize = 0; |
903 | int datalen; |
904 | uint8 action = seginfo->action; |
905 | |
906 | if (action == GIN_SEGMENT_UNMODIFIED) |
907 | { |
908 | segno++; |
909 | continue; |
910 | } |
911 | |
912 | if (action != GIN_SEGMENT_DELETE) |
913 | segsize = SizeOfGinPostingList(seginfo->seg); |
914 | |
915 | /* |
916 | * If storing the uncompressed list of added item pointers would take |
917 | * more space than storing the compressed segment as is, do that |
918 | * instead. |
919 | */ |
920 | if (action == GIN_SEGMENT_ADDITEMS && |
921 | seginfo->nmodifieditems * sizeof(ItemPointerData) > segsize) |
922 | { |
923 | action = GIN_SEGMENT_REPLACE; |
924 | } |
925 | |
926 | *((uint8 *) (walbufend++)) = segno; |
927 | *(walbufend++) = action; |
928 | |
929 | switch (action) |
930 | { |
931 | case GIN_SEGMENT_DELETE: |
932 | datalen = 0; |
933 | break; |
934 | |
935 | case GIN_SEGMENT_ADDITEMS: |
936 | datalen = seginfo->nmodifieditems * sizeof(ItemPointerData); |
937 | memcpy(walbufend, &seginfo->nmodifieditems, sizeof(uint16)); |
938 | memcpy(walbufend + sizeof(uint16), seginfo->modifieditems, datalen); |
939 | datalen += sizeof(uint16); |
940 | break; |
941 | |
942 | case GIN_SEGMENT_INSERT: |
943 | case GIN_SEGMENT_REPLACE: |
944 | datalen = SHORTALIGN(segsize); |
945 | memcpy(walbufend, seginfo->seg, segsize); |
946 | break; |
947 | |
948 | default: |
949 | elog(ERROR, "unexpected GIN leaf action %d" , action); |
950 | } |
951 | walbufend += datalen; |
952 | |
953 | if (action != GIN_SEGMENT_INSERT) |
954 | segno++; |
955 | } |
956 | |
957 | /* Pass back the constructed info via *leaf */ |
958 | leaf->walinfo = walbufbegin; |
959 | leaf->walinfolen = walbufend - walbufbegin; |
960 | } |
961 | |
962 | /* |
963 | * Assemble a disassembled posting tree leaf page back to a buffer. |
964 | * |
965 | * This just updates the target buffer; WAL stuff is caller's responsibility. |
966 | * |
967 | * NOTE: The segment pointers must not point directly to the same buffer, |
968 | * except for segments that have not been modified and whose preceding |
969 | * segments have not been modified either. |
970 | */ |
971 | static void |
972 | dataPlaceToPageLeafRecompress(Buffer buf, disassembledLeaf *leaf) |
973 | { |
974 | Page page = BufferGetPage(buf); |
975 | char *ptr; |
976 | int newsize; |
977 | bool modified = false; |
978 | dlist_iter iter; |
979 | int segsize; |
980 | |
981 | /* |
982 | * If the page was in pre-9.4 format before, convert the header, and force |
983 | * all segments to be copied to the page whether they were modified or |
984 | * not. |
985 | */ |
986 | if (!GinPageIsCompressed(page)) |
987 | { |
988 | Assert(leaf->oldformat); |
989 | GinPageSetCompressed(page); |
990 | GinPageGetOpaque(page)->maxoff = InvalidOffsetNumber; |
991 | modified = true; |
992 | } |
993 | |
994 | ptr = (char *) GinDataLeafPageGetPostingList(page); |
995 | newsize = 0; |
996 | dlist_foreach(iter, &leaf->segments) |
997 | { |
998 | leafSegmentInfo *seginfo = dlist_container(leafSegmentInfo, node, iter.cur); |
999 | |
1000 | if (seginfo->action != GIN_SEGMENT_UNMODIFIED) |
1001 | modified = true; |
1002 | |
1003 | if (seginfo->action != GIN_SEGMENT_DELETE) |
1004 | { |
1005 | segsize = SizeOfGinPostingList(seginfo->seg); |
1006 | |
1007 | if (modified) |
1008 | memcpy(ptr, seginfo->seg, segsize); |
1009 | |
1010 | ptr += segsize; |
1011 | newsize += segsize; |
1012 | } |
1013 | } |
1014 | |
1015 | Assert(newsize <= GinDataPageMaxDataSize); |
1016 | GinDataPageSetDataSize(page, newsize); |
1017 | } |
1018 | |
1019 | /* |
1020 | * Like dataPlaceToPageLeafRecompress, but writes the disassembled leaf |
1021 | * segments to two pages instead of one. |
1022 | * |
1023 | * This is different from the non-split cases in that this does not modify |
1024 | * the original page directly, but writes to temporary in-memory copies of |
1025 | * the new left and right pages. |
1026 | */ |
1027 | static void |
1028 | dataPlaceToPageLeafSplit(disassembledLeaf *leaf, |
1029 | ItemPointerData lbound, ItemPointerData rbound, |
1030 | Page lpage, Page rpage) |
1031 | { |
1032 | char *ptr; |
1033 | int segsize; |
1034 | int lsize; |
1035 | int rsize; |
1036 | dlist_node *node; |
1037 | dlist_node *firstright; |
1038 | leafSegmentInfo *seginfo; |
1039 | |
1040 | /* Initialize temporary pages to hold the new left and right pages */ |
1041 | GinInitPage(lpage, GIN_DATA | GIN_LEAF | GIN_COMPRESSED, BLCKSZ); |
1042 | GinInitPage(rpage, GIN_DATA | GIN_LEAF | GIN_COMPRESSED, BLCKSZ); |
1043 | |
1044 | /* |
1045 | * Copy the segments that go to the left page. |
1046 | * |
1047 | * XXX: We should skip copying the unmodified part of the left page, like |
1048 | * we do when recompressing. |
1049 | */ |
1050 | lsize = 0; |
1051 | ptr = (char *) GinDataLeafPageGetPostingList(lpage); |
1052 | firstright = dlist_next_node(&leaf->segments, leaf->lastleft); |
1053 | for (node = dlist_head_node(&leaf->segments); |
1054 | node != firstright; |
1055 | node = dlist_next_node(&leaf->segments, node)) |
1056 | { |
1057 | seginfo = dlist_container(leafSegmentInfo, node, node); |
1058 | |
1059 | if (seginfo->action != GIN_SEGMENT_DELETE) |
1060 | { |
1061 | segsize = SizeOfGinPostingList(seginfo->seg); |
1062 | memcpy(ptr, seginfo->seg, segsize); |
1063 | ptr += segsize; |
1064 | lsize += segsize; |
1065 | } |
1066 | } |
1067 | Assert(lsize == leaf->lsize); |
1068 | GinDataPageSetDataSize(lpage, lsize); |
1069 | *GinDataPageGetRightBound(lpage) = lbound; |
1070 | |
1071 | /* Copy the segments that go to the right page */ |
1072 | ptr = (char *) GinDataLeafPageGetPostingList(rpage); |
1073 | rsize = 0; |
1074 | for (node = firstright; |
1075 | ; |
1076 | node = dlist_next_node(&leaf->segments, node)) |
1077 | { |
1078 | seginfo = dlist_container(leafSegmentInfo, node, node); |
1079 | |
1080 | if (seginfo->action != GIN_SEGMENT_DELETE) |
1081 | { |
1082 | segsize = SizeOfGinPostingList(seginfo->seg); |
1083 | memcpy(ptr, seginfo->seg, segsize); |
1084 | ptr += segsize; |
1085 | rsize += segsize; |
1086 | } |
1087 | |
1088 | if (!dlist_has_next(&leaf->segments, node)) |
1089 | break; |
1090 | } |
1091 | Assert(rsize == leaf->rsize); |
1092 | GinDataPageSetDataSize(rpage, rsize); |
1093 | *GinDataPageGetRightBound(rpage) = rbound; |
1094 | } |
1095 | |
1096 | /* |
1097 | * Prepare to insert data on an internal data page. |
1098 | * |
1099 | * If it will fit, return GPTP_INSERT after doing whatever setup is needed |
1100 | * before we enter the insertion critical section. *ptp_workspace can be |
1101 | * set to pass information along to the execPlaceToPage function. |
1102 | * |
1103 | * If it won't fit, perform a page split and return two temporary page |
1104 | * images into *newlpage and *newrpage, with result GPTP_SPLIT. |
1105 | * |
1106 | * In neither case should the given page buffer be modified here. |
1107 | * |
1108 | * Note: on insertion to an internal node, in addition to inserting the given |
1109 | * item, the downlink of the existing item at stack->off will be updated to |
1110 | * point to updateblkno. |
1111 | */ |
1112 | static GinPlaceToPageRC |
1113 | dataBeginPlaceToPageInternal(GinBtree btree, Buffer buf, GinBtreeStack *stack, |
1114 | void *insertdata, BlockNumber updateblkno, |
1115 | void **ptp_workspace, |
1116 | Page *newlpage, Page *newrpage) |
1117 | { |
1118 | Page page = BufferGetPage(buf); |
1119 | |
1120 | /* If it doesn't fit, deal with split case */ |
1121 | if (GinNonLeafDataPageGetFreeSpace(page) < sizeof(PostingItem)) |
1122 | { |
1123 | dataSplitPageInternal(btree, buf, stack, insertdata, updateblkno, |
1124 | newlpage, newrpage); |
1125 | return GPTP_SPLIT; |
1126 | } |
1127 | |
1128 | /* Else, we're ready to proceed with insertion */ |
1129 | return GPTP_INSERT; |
1130 | } |
1131 | |
1132 | /* |
1133 | * Perform data insertion after beginPlaceToPage has decided it will fit. |
1134 | * |
1135 | * This is invoked within a critical section, and XLOG record creation (if |
1136 | * needed) is already started. The target buffer is registered in slot 0. |
1137 | */ |
1138 | static void |
1139 | dataExecPlaceToPageInternal(GinBtree btree, Buffer buf, GinBtreeStack *stack, |
1140 | void *insertdata, BlockNumber updateblkno, |
1141 | void *ptp_workspace) |
1142 | { |
1143 | Page page = BufferGetPage(buf); |
1144 | OffsetNumber off = stack->off; |
1145 | PostingItem *pitem; |
1146 | |
1147 | /* Update existing downlink to point to next page (on internal page) */ |
1148 | pitem = GinDataPageGetPostingItem(page, off); |
1149 | PostingItemSetBlockNumber(pitem, updateblkno); |
1150 | |
1151 | /* Add new item */ |
1152 | pitem = (PostingItem *) insertdata; |
1153 | GinDataPageAddPostingItem(page, pitem, off); |
1154 | |
1155 | if (RelationNeedsWAL(btree->index) && !btree->isBuild) |
1156 | { |
1157 | /* |
1158 | * This must be static, because it has to survive until XLogInsert, |
1159 | * and we can't palloc here. Ugly, but the XLogInsert infrastructure |
1160 | * isn't reentrant anyway. |
1161 | */ |
1162 | static ginxlogInsertDataInternal data; |
1163 | |
1164 | data.offset = off; |
1165 | data.newitem = *pitem; |
1166 | |
1167 | XLogRegisterBufData(0, (char *) &data, |
1168 | sizeof(ginxlogInsertDataInternal)); |
1169 | } |
1170 | } |
1171 | |
1172 | /* |
1173 | * Prepare to insert data on a posting-tree data page. |
1174 | * |
1175 | * If it will fit, return GPTP_INSERT after doing whatever setup is needed |
1176 | * before we enter the insertion critical section. *ptp_workspace can be |
1177 | * set to pass information along to the execPlaceToPage function. |
1178 | * |
1179 | * If it won't fit, perform a page split and return two temporary page |
1180 | * images into *newlpage and *newrpage, with result GPTP_SPLIT. |
1181 | * |
1182 | * In neither case should the given page buffer be modified here. |
1183 | * |
1184 | * Note: on insertion to an internal node, in addition to inserting the given |
1185 | * item, the downlink of the existing item at stack->off will be updated to |
1186 | * point to updateblkno. |
1187 | * |
1188 | * Calls relevant function for internal or leaf page because they are handled |
1189 | * very differently. |
1190 | */ |
1191 | static GinPlaceToPageRC |
1192 | dataBeginPlaceToPage(GinBtree btree, Buffer buf, GinBtreeStack *stack, |
1193 | void *insertdata, BlockNumber updateblkno, |
1194 | void **ptp_workspace, |
1195 | Page *newlpage, Page *newrpage) |
1196 | { |
1197 | Page page = BufferGetPage(buf); |
1198 | |
1199 | Assert(GinPageIsData(page)); |
1200 | |
1201 | if (GinPageIsLeaf(page)) |
1202 | return dataBeginPlaceToPageLeaf(btree, buf, stack, insertdata, |
1203 | ptp_workspace, |
1204 | newlpage, newrpage); |
1205 | else |
1206 | return dataBeginPlaceToPageInternal(btree, buf, stack, |
1207 | insertdata, updateblkno, |
1208 | ptp_workspace, |
1209 | newlpage, newrpage); |
1210 | } |
1211 | |
1212 | /* |
1213 | * Perform data insertion after beginPlaceToPage has decided it will fit. |
1214 | * |
1215 | * This is invoked within a critical section, and XLOG record creation (if |
1216 | * needed) is already started. The target buffer is registered in slot 0. |
1217 | * |
1218 | * Calls relevant function for internal or leaf page because they are handled |
1219 | * very differently. |
1220 | */ |
1221 | static void |
1222 | dataExecPlaceToPage(GinBtree btree, Buffer buf, GinBtreeStack *stack, |
1223 | void *insertdata, BlockNumber updateblkno, |
1224 | void *ptp_workspace) |
1225 | { |
1226 | Page page = BufferGetPage(buf); |
1227 | |
1228 | if (GinPageIsLeaf(page)) |
1229 | dataExecPlaceToPageLeaf(btree, buf, stack, insertdata, |
1230 | ptp_workspace); |
1231 | else |
1232 | dataExecPlaceToPageInternal(btree, buf, stack, insertdata, |
1233 | updateblkno, ptp_workspace); |
1234 | } |
1235 | |
1236 | /* |
1237 | * Split internal page and insert new data. |
1238 | * |
1239 | * Returns new temp pages to *newlpage and *newrpage. |
1240 | * The original buffer is left untouched. |
1241 | */ |
1242 | static void |
1243 | dataSplitPageInternal(GinBtree btree, Buffer origbuf, |
1244 | GinBtreeStack *stack, |
1245 | void *insertdata, BlockNumber updateblkno, |
1246 | Page *newlpage, Page *newrpage) |
1247 | { |
1248 | Page oldpage = BufferGetPage(origbuf); |
1249 | OffsetNumber off = stack->off; |
1250 | int nitems = GinPageGetOpaque(oldpage)->maxoff; |
1251 | int nleftitems; |
1252 | int nrightitems; |
1253 | Size pageSize = PageGetPageSize(oldpage); |
1254 | ItemPointerData oldbound = *GinDataPageGetRightBound(oldpage); |
1255 | ItemPointer bound; |
1256 | Page lpage; |
1257 | Page rpage; |
1258 | OffsetNumber separator; |
1259 | PostingItem allitems[(BLCKSZ / sizeof(PostingItem)) + 1]; |
1260 | |
1261 | lpage = PageGetTempPage(oldpage); |
1262 | rpage = PageGetTempPage(oldpage); |
1263 | GinInitPage(lpage, GinPageGetOpaque(oldpage)->flags, pageSize); |
1264 | GinInitPage(rpage, GinPageGetOpaque(oldpage)->flags, pageSize); |
1265 | |
1266 | /* |
1267 | * First construct a new list of PostingItems, which includes all the old |
1268 | * items, and the new item. |
1269 | */ |
1270 | memcpy(allitems, GinDataPageGetPostingItem(oldpage, FirstOffsetNumber), |
1271 | (off - 1) * sizeof(PostingItem)); |
1272 | |
1273 | allitems[off - 1] = *((PostingItem *) insertdata); |
1274 | memcpy(&allitems[off], GinDataPageGetPostingItem(oldpage, off), |
1275 | (nitems - (off - 1)) * sizeof(PostingItem)); |
1276 | nitems++; |
1277 | |
1278 | /* Update existing downlink to point to next page */ |
1279 | PostingItemSetBlockNumber(&allitems[off], updateblkno); |
1280 | |
1281 | /* |
1282 | * When creating a new index, fit as many tuples as possible on the left |
1283 | * page, on the assumption that the table is scanned from beginning to |
1284 | * end. This packs the index as tight as possible. |
1285 | */ |
1286 | if (btree->isBuild && GinPageRightMost(oldpage)) |
1287 | separator = GinNonLeafDataPageGetFreeSpace(rpage) / sizeof(PostingItem); |
1288 | else |
1289 | separator = nitems / 2; |
1290 | nleftitems = separator; |
1291 | nrightitems = nitems - separator; |
1292 | |
1293 | memcpy(GinDataPageGetPostingItem(lpage, FirstOffsetNumber), |
1294 | allitems, |
1295 | nleftitems * sizeof(PostingItem)); |
1296 | GinPageGetOpaque(lpage)->maxoff = nleftitems; |
1297 | memcpy(GinDataPageGetPostingItem(rpage, FirstOffsetNumber), |
1298 | &allitems[separator], |
1299 | nrightitems * sizeof(PostingItem)); |
1300 | GinPageGetOpaque(rpage)->maxoff = nrightitems; |
1301 | |
1302 | /* |
1303 | * Also set pd_lower for both pages, like GinDataPageAddPostingItem does. |
1304 | */ |
1305 | GinDataPageSetDataSize(lpage, nleftitems * sizeof(PostingItem)); |
1306 | GinDataPageSetDataSize(rpage, nrightitems * sizeof(PostingItem)); |
1307 | |
1308 | /* set up right bound for left page */ |
1309 | bound = GinDataPageGetRightBound(lpage); |
1310 | *bound = GinDataPageGetPostingItem(lpage, nleftitems)->key; |
1311 | |
1312 | /* set up right bound for right page */ |
1313 | *GinDataPageGetRightBound(rpage) = oldbound; |
1314 | |
1315 | /* return temp pages to caller */ |
1316 | *newlpage = lpage; |
1317 | *newrpage = rpage; |
1318 | } |
1319 | |
1320 | /* |
1321 | * Construct insertion payload for inserting the downlink for given buffer. |
1322 | */ |
1323 | static void * |
1324 | dataPrepareDownlink(GinBtree btree, Buffer lbuf) |
1325 | { |
1326 | PostingItem *pitem = palloc(sizeof(PostingItem)); |
1327 | Page lpage = BufferGetPage(lbuf); |
1328 | |
1329 | PostingItemSetBlockNumber(pitem, BufferGetBlockNumber(lbuf)); |
1330 | pitem->key = *GinDataPageGetRightBound(lpage); |
1331 | |
1332 | return pitem; |
1333 | } |
1334 | |
1335 | /* |
1336 | * Fills new root by right bound values from child. |
1337 | * Also called from ginxlog, should not use btree |
1338 | */ |
1339 | void |
1340 | ginDataFillRoot(GinBtree btree, Page root, BlockNumber lblkno, Page lpage, BlockNumber rblkno, Page rpage) |
1341 | { |
1342 | PostingItem li, |
1343 | ri; |
1344 | |
1345 | li.key = *GinDataPageGetRightBound(lpage); |
1346 | PostingItemSetBlockNumber(&li, lblkno); |
1347 | GinDataPageAddPostingItem(root, &li, InvalidOffsetNumber); |
1348 | |
1349 | ri.key = *GinDataPageGetRightBound(rpage); |
1350 | PostingItemSetBlockNumber(&ri, rblkno); |
1351 | GinDataPageAddPostingItem(root, &ri, InvalidOffsetNumber); |
1352 | } |
1353 | |
1354 | |
1355 | /*** Functions to work with disassembled leaf pages ***/ |
1356 | |
1357 | /* |
1358 | * Disassemble page into a disassembledLeaf struct. |
1359 | */ |
1360 | static disassembledLeaf * |
1361 | disassembleLeaf(Page page) |
1362 | { |
1363 | disassembledLeaf *leaf; |
1364 | GinPostingList *seg; |
1365 | Pointer segbegin; |
1366 | Pointer segend; |
1367 | |
1368 | leaf = palloc0(sizeof(disassembledLeaf)); |
1369 | dlist_init(&leaf->segments); |
1370 | |
1371 | if (GinPageIsCompressed(page)) |
1372 | { |
1373 | /* |
1374 | * Create a leafSegment entry for each segment. |
1375 | */ |
1376 | seg = GinDataLeafPageGetPostingList(page); |
1377 | segbegin = (Pointer) seg; |
1378 | segend = segbegin + GinDataLeafPageGetPostingListSize(page); |
1379 | while ((Pointer) seg < segend) |
1380 | { |
1381 | leafSegmentInfo *seginfo = palloc(sizeof(leafSegmentInfo)); |
1382 | |
1383 | seginfo->action = GIN_SEGMENT_UNMODIFIED; |
1384 | seginfo->seg = seg; |
1385 | seginfo->items = NULL; |
1386 | seginfo->nitems = 0; |
1387 | dlist_push_tail(&leaf->segments, &seginfo->node); |
1388 | |
1389 | seg = GinNextPostingListSegment(seg); |
1390 | } |
1391 | leaf->oldformat = false; |
1392 | } |
1393 | else |
1394 | { |
1395 | /* |
1396 | * A pre-9.4 format uncompressed page is represented by a single |
1397 | * segment, with an array of items. The corner case is uncompressed |
1398 | * page containing no items, which is represented as no segments. |
1399 | */ |
1400 | ItemPointer uncompressed; |
1401 | int nuncompressed; |
1402 | leafSegmentInfo *seginfo; |
1403 | |
1404 | uncompressed = dataLeafPageGetUncompressed(page, &nuncompressed); |
1405 | |
1406 | if (nuncompressed > 0) |
1407 | { |
1408 | seginfo = palloc(sizeof(leafSegmentInfo)); |
1409 | |
1410 | seginfo->action = GIN_SEGMENT_REPLACE; |
1411 | seginfo->seg = NULL; |
1412 | seginfo->items = palloc(nuncompressed * sizeof(ItemPointerData)); |
1413 | memcpy(seginfo->items, uncompressed, nuncompressed * sizeof(ItemPointerData)); |
1414 | seginfo->nitems = nuncompressed; |
1415 | |
1416 | dlist_push_tail(&leaf->segments, &seginfo->node); |
1417 | } |
1418 | |
1419 | leaf->oldformat = true; |
1420 | } |
1421 | |
1422 | return leaf; |
1423 | } |
1424 | |
1425 | /* |
1426 | * Distribute newItems to the segments. |
1427 | * |
1428 | * Any segments that acquire new items are decoded, and the new items are |
1429 | * merged with the old items. |
1430 | * |
1431 | * Returns true if any new items were added. False means they were all |
1432 | * duplicates of existing items on the page. |
1433 | */ |
1434 | static bool |
1435 | addItemsToLeaf(disassembledLeaf *leaf, ItemPointer newItems, int nNewItems) |
1436 | { |
1437 | dlist_iter iter; |
1438 | ItemPointer nextnew = newItems; |
1439 | int newleft = nNewItems; |
1440 | bool modified = false; |
1441 | leafSegmentInfo *newseg; |
1442 | |
1443 | /* |
1444 | * If the page is completely empty, just construct one new segment to hold |
1445 | * all the new items. |
1446 | */ |
1447 | if (dlist_is_empty(&leaf->segments)) |
1448 | { |
1449 | newseg = palloc(sizeof(leafSegmentInfo)); |
1450 | newseg->seg = NULL; |
1451 | newseg->items = newItems; |
1452 | newseg->nitems = nNewItems; |
1453 | newseg->action = GIN_SEGMENT_INSERT; |
1454 | dlist_push_tail(&leaf->segments, &newseg->node); |
1455 | return true; |
1456 | } |
1457 | |
1458 | dlist_foreach(iter, &leaf->segments) |
1459 | { |
1460 | leafSegmentInfo *cur = (leafSegmentInfo *) dlist_container(leafSegmentInfo, node, iter.cur); |
1461 | int nthis; |
1462 | ItemPointer tmpitems; |
1463 | int ntmpitems; |
1464 | |
1465 | /* |
1466 | * How many of the new items fall into this segment? |
1467 | */ |
1468 | if (!dlist_has_next(&leaf->segments, iter.cur)) |
1469 | nthis = newleft; |
1470 | else |
1471 | { |
1472 | leafSegmentInfo *next; |
1473 | ItemPointerData next_first; |
1474 | |
1475 | next = (leafSegmentInfo *) dlist_container(leafSegmentInfo, node, |
1476 | dlist_next_node(&leaf->segments, iter.cur)); |
1477 | if (next->items) |
1478 | next_first = next->items[0]; |
1479 | else |
1480 | { |
1481 | Assert(next->seg != NULL); |
1482 | next_first = next->seg->first; |
1483 | } |
1484 | |
1485 | nthis = 0; |
1486 | while (nthis < newleft && ginCompareItemPointers(&nextnew[nthis], &next_first) < 0) |
1487 | nthis++; |
1488 | } |
1489 | if (nthis == 0) |
1490 | continue; |
1491 | |
1492 | /* Merge the new items with the existing items. */ |
1493 | if (!cur->items) |
1494 | cur->items = ginPostingListDecode(cur->seg, &cur->nitems); |
1495 | |
1496 | /* |
1497 | * Fast path for the important special case that we're appending to |
1498 | * the end of the page: don't let the last segment on the page grow |
1499 | * larger than the target, create a new segment before that happens. |
1500 | */ |
1501 | if (!dlist_has_next(&leaf->segments, iter.cur) && |
1502 | ginCompareItemPointers(&cur->items[cur->nitems - 1], &nextnew[0]) < 0 && |
1503 | cur->seg != NULL && |
1504 | SizeOfGinPostingList(cur->seg) >= GinPostingListSegmentTargetSize) |
1505 | { |
1506 | newseg = palloc(sizeof(leafSegmentInfo)); |
1507 | newseg->seg = NULL; |
1508 | newseg->items = nextnew; |
1509 | newseg->nitems = nthis; |
1510 | newseg->action = GIN_SEGMENT_INSERT; |
1511 | dlist_push_tail(&leaf->segments, &newseg->node); |
1512 | modified = true; |
1513 | break; |
1514 | } |
1515 | |
1516 | tmpitems = ginMergeItemPointers(cur->items, cur->nitems, |
1517 | nextnew, nthis, |
1518 | &ntmpitems); |
1519 | if (ntmpitems != cur->nitems) |
1520 | { |
1521 | /* |
1522 | * If there are no duplicates, track the added items so that we |
1523 | * can emit a compact ADDITEMS WAL record later on. (it doesn't |
1524 | * seem worth re-checking which items were duplicates, if there |
1525 | * were any) |
1526 | */ |
1527 | if (ntmpitems == nthis + cur->nitems && |
1528 | cur->action == GIN_SEGMENT_UNMODIFIED) |
1529 | { |
1530 | cur->action = GIN_SEGMENT_ADDITEMS; |
1531 | cur->modifieditems = nextnew; |
1532 | cur->nmodifieditems = nthis; |
1533 | } |
1534 | else |
1535 | cur->action = GIN_SEGMENT_REPLACE; |
1536 | |
1537 | cur->items = tmpitems; |
1538 | cur->nitems = ntmpitems; |
1539 | cur->seg = NULL; |
1540 | modified = true; |
1541 | } |
1542 | |
1543 | nextnew += nthis; |
1544 | newleft -= nthis; |
1545 | if (newleft == 0) |
1546 | break; |
1547 | } |
1548 | |
1549 | return modified; |
1550 | } |
1551 | |
1552 | /* |
1553 | * Recompresses all segments that have been modified. |
1554 | * |
1555 | * If not all the items fit on two pages (ie. after split), we store as |
1556 | * many items as fit, and set *remaining to the first item that didn't fit. |
1557 | * If all items fit, *remaining is set to invalid. |
1558 | * |
1559 | * Returns true if the page has to be split. |
1560 | */ |
1561 | static bool |
1562 | leafRepackItems(disassembledLeaf *leaf, ItemPointer remaining) |
1563 | { |
1564 | int pgused = 0; |
1565 | bool needsplit = false; |
1566 | dlist_iter iter; |
1567 | int segsize; |
1568 | leafSegmentInfo *nextseg; |
1569 | int npacked; |
1570 | bool modified; |
1571 | dlist_node *cur_node; |
1572 | dlist_node *next_node; |
1573 | |
1574 | ItemPointerSetInvalid(remaining); |
1575 | |
1576 | /* |
1577 | * cannot use dlist_foreach_modify here because we insert adjacent items |
1578 | * while iterating. |
1579 | */ |
1580 | for (cur_node = dlist_head_node(&leaf->segments); |
1581 | cur_node != NULL; |
1582 | cur_node = next_node) |
1583 | { |
1584 | leafSegmentInfo *seginfo = dlist_container(leafSegmentInfo, node, |
1585 | cur_node); |
1586 | |
1587 | if (dlist_has_next(&leaf->segments, cur_node)) |
1588 | next_node = dlist_next_node(&leaf->segments, cur_node); |
1589 | else |
1590 | next_node = NULL; |
1591 | |
1592 | /* Compress the posting list, if necessary */ |
1593 | if (seginfo->action != GIN_SEGMENT_DELETE) |
1594 | { |
1595 | if (seginfo->seg == NULL) |
1596 | { |
1597 | if (seginfo->nitems > GinPostingListSegmentMaxSize) |
1598 | npacked = 0; /* no chance that it would fit. */ |
1599 | else |
1600 | { |
1601 | seginfo->seg = ginCompressPostingList(seginfo->items, |
1602 | seginfo->nitems, |
1603 | GinPostingListSegmentMaxSize, |
1604 | &npacked); |
1605 | } |
1606 | if (npacked != seginfo->nitems) |
1607 | { |
1608 | /* |
1609 | * Too large. Compress again to the target size, and |
1610 | * create a new segment to represent the remaining items. |
1611 | * The new segment is inserted after this one, so it will |
1612 | * be processed in the next iteration of this loop. |
1613 | */ |
1614 | if (seginfo->seg) |
1615 | pfree(seginfo->seg); |
1616 | seginfo->seg = ginCompressPostingList(seginfo->items, |
1617 | seginfo->nitems, |
1618 | GinPostingListSegmentTargetSize, |
1619 | &npacked); |
1620 | if (seginfo->action != GIN_SEGMENT_INSERT) |
1621 | seginfo->action = GIN_SEGMENT_REPLACE; |
1622 | |
1623 | nextseg = palloc(sizeof(leafSegmentInfo)); |
1624 | nextseg->action = GIN_SEGMENT_INSERT; |
1625 | nextseg->seg = NULL; |
1626 | nextseg->items = &seginfo->items[npacked]; |
1627 | nextseg->nitems = seginfo->nitems - npacked; |
1628 | next_node = &nextseg->node; |
1629 | dlist_insert_after(cur_node, next_node); |
1630 | } |
1631 | } |
1632 | |
1633 | /* |
1634 | * If the segment is very small, merge it with the next segment. |
1635 | */ |
1636 | if (SizeOfGinPostingList(seginfo->seg) < GinPostingListSegmentMinSize && next_node) |
1637 | { |
1638 | int nmerged; |
1639 | |
1640 | nextseg = dlist_container(leafSegmentInfo, node, next_node); |
1641 | |
1642 | if (seginfo->items == NULL) |
1643 | seginfo->items = ginPostingListDecode(seginfo->seg, |
1644 | &seginfo->nitems); |
1645 | if (nextseg->items == NULL) |
1646 | nextseg->items = ginPostingListDecode(nextseg->seg, |
1647 | &nextseg->nitems); |
1648 | nextseg->items = |
1649 | ginMergeItemPointers(seginfo->items, seginfo->nitems, |
1650 | nextseg->items, nextseg->nitems, |
1651 | &nmerged); |
1652 | Assert(nmerged == seginfo->nitems + nextseg->nitems); |
1653 | nextseg->nitems = nmerged; |
1654 | nextseg->seg = NULL; |
1655 | |
1656 | nextseg->action = GIN_SEGMENT_REPLACE; |
1657 | nextseg->modifieditems = NULL; |
1658 | nextseg->nmodifieditems = 0; |
1659 | |
1660 | if (seginfo->action == GIN_SEGMENT_INSERT) |
1661 | { |
1662 | dlist_delete(cur_node); |
1663 | continue; |
1664 | } |
1665 | else |
1666 | { |
1667 | seginfo->action = GIN_SEGMENT_DELETE; |
1668 | seginfo->seg = NULL; |
1669 | } |
1670 | } |
1671 | |
1672 | seginfo->items = NULL; |
1673 | seginfo->nitems = 0; |
1674 | } |
1675 | |
1676 | if (seginfo->action == GIN_SEGMENT_DELETE) |
1677 | continue; |
1678 | |
1679 | /* |
1680 | * OK, we now have a compressed version of this segment ready for |
1681 | * copying to the page. Did we exceed the size that fits on one page? |
1682 | */ |
1683 | segsize = SizeOfGinPostingList(seginfo->seg); |
1684 | if (pgused + segsize > GinDataPageMaxDataSize) |
1685 | { |
1686 | if (!needsplit) |
1687 | { |
1688 | /* switch to right page */ |
1689 | Assert(pgused > 0); |
1690 | leaf->lastleft = dlist_prev_node(&leaf->segments, cur_node); |
1691 | needsplit = true; |
1692 | leaf->lsize = pgused; |
1693 | pgused = 0; |
1694 | } |
1695 | else |
1696 | { |
1697 | /* |
1698 | * Filled both pages. The last segment we constructed did not |
1699 | * fit. |
1700 | */ |
1701 | *remaining = seginfo->seg->first; |
1702 | |
1703 | /* |
1704 | * remove all segments that did not fit from the list. |
1705 | */ |
1706 | while (dlist_has_next(&leaf->segments, cur_node)) |
1707 | dlist_delete(dlist_next_node(&leaf->segments, cur_node)); |
1708 | dlist_delete(cur_node); |
1709 | break; |
1710 | } |
1711 | } |
1712 | |
1713 | pgused += segsize; |
1714 | } |
1715 | |
1716 | if (!needsplit) |
1717 | { |
1718 | leaf->lsize = pgused; |
1719 | leaf->rsize = 0; |
1720 | } |
1721 | else |
1722 | leaf->rsize = pgused; |
1723 | |
1724 | Assert(leaf->lsize <= GinDataPageMaxDataSize); |
1725 | Assert(leaf->rsize <= GinDataPageMaxDataSize); |
1726 | |
1727 | /* |
1728 | * Make a palloc'd copy of every segment after the first modified one, |
1729 | * because as we start copying items to the original page, we might |
1730 | * overwrite an existing segment. |
1731 | */ |
1732 | modified = false; |
1733 | dlist_foreach(iter, &leaf->segments) |
1734 | { |
1735 | leafSegmentInfo *seginfo = dlist_container(leafSegmentInfo, node, |
1736 | iter.cur); |
1737 | |
1738 | if (!modified && seginfo->action != GIN_SEGMENT_UNMODIFIED) |
1739 | { |
1740 | modified = true; |
1741 | } |
1742 | else if (modified && seginfo->action == GIN_SEGMENT_UNMODIFIED) |
1743 | { |
1744 | GinPostingList *tmp; |
1745 | |
1746 | segsize = SizeOfGinPostingList(seginfo->seg); |
1747 | tmp = palloc(segsize); |
1748 | memcpy(tmp, seginfo->seg, segsize); |
1749 | seginfo->seg = tmp; |
1750 | } |
1751 | } |
1752 | |
1753 | return needsplit; |
1754 | } |
1755 | |
1756 | |
1757 | /*** Functions that are exported to the rest of the GIN code ***/ |
1758 | |
1759 | /* |
1760 | * Creates new posting tree containing the given TIDs. Returns the page |
1761 | * number of the root of the new posting tree. |
1762 | * |
1763 | * items[] must be in sorted order with no duplicates. |
1764 | */ |
1765 | BlockNumber |
1766 | createPostingTree(Relation index, ItemPointerData *items, uint32 nitems, |
1767 | GinStatsData *buildStats, Buffer entrybuffer) |
1768 | { |
1769 | BlockNumber blkno; |
1770 | Buffer buffer; |
1771 | Page tmppage; |
1772 | Page page; |
1773 | Pointer ptr; |
1774 | int nrootitems; |
1775 | int rootsize; |
1776 | bool is_build = (buildStats != NULL); |
1777 | |
1778 | /* Construct the new root page in memory first. */ |
1779 | tmppage = (Page) palloc(BLCKSZ); |
1780 | GinInitPage(tmppage, GIN_DATA | GIN_LEAF | GIN_COMPRESSED, BLCKSZ); |
1781 | GinPageGetOpaque(tmppage)->rightlink = InvalidBlockNumber; |
1782 | |
1783 | /* |
1784 | * Write as many of the items to the root page as fit. In segments of max |
1785 | * GinPostingListSegmentMaxSize bytes each. |
1786 | */ |
1787 | nrootitems = 0; |
1788 | rootsize = 0; |
1789 | ptr = (Pointer) GinDataLeafPageGetPostingList(tmppage); |
1790 | while (nrootitems < nitems) |
1791 | { |
1792 | GinPostingList *segment; |
1793 | int npacked; |
1794 | int segsize; |
1795 | |
1796 | segment = ginCompressPostingList(&items[nrootitems], |
1797 | nitems - nrootitems, |
1798 | GinPostingListSegmentMaxSize, |
1799 | &npacked); |
1800 | segsize = SizeOfGinPostingList(segment); |
1801 | if (rootsize + segsize > GinDataPageMaxDataSize) |
1802 | break; |
1803 | |
1804 | memcpy(ptr, segment, segsize); |
1805 | ptr += segsize; |
1806 | rootsize += segsize; |
1807 | nrootitems += npacked; |
1808 | pfree(segment); |
1809 | } |
1810 | GinDataPageSetDataSize(tmppage, rootsize); |
1811 | |
1812 | /* |
1813 | * All set. Get a new physical page, and copy the in-memory page to it. |
1814 | */ |
1815 | buffer = GinNewBuffer(index); |
1816 | page = BufferGetPage(buffer); |
1817 | blkno = BufferGetBlockNumber(buffer); |
1818 | |
1819 | /* |
1820 | * Copy any predicate locks from the entry tree leaf (containing posting |
1821 | * list) to the posting tree. |
1822 | */ |
1823 | PredicateLockPageSplit(index, BufferGetBlockNumber(entrybuffer), blkno); |
1824 | |
1825 | START_CRIT_SECTION(); |
1826 | |
1827 | PageRestoreTempPage(tmppage, page); |
1828 | MarkBufferDirty(buffer); |
1829 | |
1830 | if (RelationNeedsWAL(index) && !is_build) |
1831 | { |
1832 | XLogRecPtr recptr; |
1833 | ginxlogCreatePostingTree data; |
1834 | |
1835 | data.size = rootsize; |
1836 | |
1837 | XLogBeginInsert(); |
1838 | XLogRegisterData((char *) &data, sizeof(ginxlogCreatePostingTree)); |
1839 | |
1840 | XLogRegisterData((char *) GinDataLeafPageGetPostingList(page), |
1841 | rootsize); |
1842 | XLogRegisterBuffer(0, buffer, REGBUF_WILL_INIT); |
1843 | |
1844 | recptr = XLogInsert(RM_GIN_ID, XLOG_GIN_CREATE_PTREE); |
1845 | PageSetLSN(page, recptr); |
1846 | } |
1847 | |
1848 | UnlockReleaseBuffer(buffer); |
1849 | |
1850 | END_CRIT_SECTION(); |
1851 | |
1852 | /* During index build, count the newly-added data page */ |
1853 | if (buildStats) |
1854 | buildStats->nDataPages++; |
1855 | |
1856 | elog(DEBUG2, "created GIN posting tree with %d items" , nrootitems); |
1857 | |
1858 | /* |
1859 | * Add any remaining TIDs to the newly-created posting tree. |
1860 | */ |
1861 | if (nitems > nrootitems) |
1862 | { |
1863 | ginInsertItemPointers(index, blkno, |
1864 | items + nrootitems, |
1865 | nitems - nrootitems, |
1866 | buildStats); |
1867 | } |
1868 | |
1869 | return blkno; |
1870 | } |
1871 | |
1872 | static void |
1873 | ginPrepareDataScan(GinBtree btree, Relation index, BlockNumber rootBlkno) |
1874 | { |
1875 | memset(btree, 0, sizeof(GinBtreeData)); |
1876 | |
1877 | btree->index = index; |
1878 | btree->rootBlkno = rootBlkno; |
1879 | |
1880 | btree->findChildPage = dataLocateItem; |
1881 | btree->getLeftMostChild = dataGetLeftMostPage; |
1882 | btree->isMoveRight = dataIsMoveRight; |
1883 | btree->findItem = NULL; |
1884 | btree->findChildPtr = dataFindChildPtr; |
1885 | btree->beginPlaceToPage = dataBeginPlaceToPage; |
1886 | btree->execPlaceToPage = dataExecPlaceToPage; |
1887 | btree->fillRoot = ginDataFillRoot; |
1888 | btree->prepareDownlink = dataPrepareDownlink; |
1889 | |
1890 | btree->isData = true; |
1891 | btree->fullScan = false; |
1892 | btree->isBuild = false; |
1893 | } |
1894 | |
1895 | /* |
1896 | * Inserts array of item pointers, may execute several tree scan (very rare) |
1897 | */ |
1898 | void |
1899 | ginInsertItemPointers(Relation index, BlockNumber rootBlkno, |
1900 | ItemPointerData *items, uint32 nitem, |
1901 | GinStatsData *buildStats) |
1902 | { |
1903 | GinBtreeData btree; |
1904 | GinBtreeDataLeafInsertData insertdata; |
1905 | GinBtreeStack *stack; |
1906 | |
1907 | ginPrepareDataScan(&btree, index, rootBlkno); |
1908 | btree.isBuild = (buildStats != NULL); |
1909 | insertdata.items = items; |
1910 | insertdata.nitem = nitem; |
1911 | insertdata.curitem = 0; |
1912 | |
1913 | while (insertdata.curitem < insertdata.nitem) |
1914 | { |
1915 | /* search for the leaf page where the first item should go to */ |
1916 | btree.itemptr = insertdata.items[insertdata.curitem]; |
1917 | stack = ginFindLeafPage(&btree, false, true, NULL); |
1918 | |
1919 | ginInsertValue(&btree, stack, &insertdata, buildStats); |
1920 | } |
1921 | } |
1922 | |
1923 | /* |
1924 | * Starts a new scan on a posting tree. |
1925 | */ |
1926 | GinBtreeStack * |
1927 | ginScanBeginPostingTree(GinBtree btree, Relation index, BlockNumber rootBlkno, |
1928 | Snapshot snapshot) |
1929 | { |
1930 | GinBtreeStack *stack; |
1931 | |
1932 | ginPrepareDataScan(btree, index, rootBlkno); |
1933 | |
1934 | btree->fullScan = true; |
1935 | |
1936 | stack = ginFindLeafPage(btree, true, false, snapshot); |
1937 | |
1938 | return stack; |
1939 | } |
1940 | |