| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* nbtxlog.c |
| 4 |
* WAL replay logic for btrees. |
| 5 |
* |
| 6 |
* |
| 7 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 8 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 9 |
* |
| 10 |
* IDENTIFICATION |
| 11 |
* src/backend/access/nbtree/nbtxlog.c |
| 12 |
* |
| 13 |
*------------------------------------------------------------------------- |
| 14 |
*/ |
| 15 |
#include "postgres.h" |
| 16 |
|
| 17 |
#include "access/bufmask.h" |
| 18 |
#include "access/nbtree.h" |
| 19 |
#include "access/nbtxlog.h" |
| 20 |
#include "access/transam.h" |
| 21 |
#include "access/xlogutils.h" |
| 22 |
#include "storage/standby.h" |
| 23 |
#include "utils/memutils.h" |
| 24 |
|
| 25 |
static MemoryContext opCtx; /* working memory for operations */ |
| 26 |
|
| 27 |
/* |
| 28 |
* _bt_restore_page -- re-enter all the index tuples on a page |
| 29 |
* |
| 30 |
* The page is freshly init'd, and *from (length len) is a copy of what |
| 31 |
* had been its upper part (pd_upper to pd_special). We assume that the |
| 32 |
* tuples had been added to the page in item-number order, and therefore |
| 33 |
* the one with highest item number appears first (lowest on the page). |
| 34 |
*/ |
| 35 |
static void |
| 36 |
_bt_restore_page(Page page, char *from, int len) |
| 37 |
{ |
| 38 |
IndexTupleData itupdata; |
| 39 |
Size itemsz; |
| 40 |
char *end = from + len; |
| 41 |
Item items[MaxIndexTuplesPerPage]; |
| 42 |
uint16 itemsizes[MaxIndexTuplesPerPage]; |
| 43 |
int i; |
| 44 |
int nitems; |
| 45 |
|
| 46 |
/* |
| 47 |
* To get the items back in the original order, we add them to the page in |
| 48 |
* reverse. To figure out where one tuple ends and another begins, we |
| 49 |
* have to scan them in forward order first. |
| 50 |
*/ |
| 51 |
i = 0; |
| 52 |
while (from < end) |
| 53 |
{ |
| 54 |
/* |
| 55 |
* As we step through the items, 'from' won't always be properly |
| 56 |
* aligned, so we need to use memcpy(). Further, we use Item (which |
| 57 |
* is just a char*) here for our items array for the same reason; |
| 58 |
* wouldn't want the compiler or anyone thinking that an item is |
| 59 |
* aligned when it isn't. |
| 60 |
*/ |
| 61 |
memcpy(&itupdata, from, sizeof(IndexTupleData)); |
| 62 |
itemsz = IndexTupleSize(&itupdata); |
| 63 |
itemsz = MAXALIGN(itemsz); |
| 64 |
|
| 65 |
items[i] = (Item) from; |
| 66 |
itemsizes[i] = itemsz; |
| 67 |
i++; |
| 68 |
|
| 69 |
from += itemsz; |
| 70 |
} |
| 71 |
nitems = i; |
| 72 |
|
| 73 |
for (i = nitems - 1; i >= 0; i--) |
| 74 |
{ |
| 75 |
if (PageAddItem(page, items[i], itemsizes[i], nitems - i, |
| 76 |
false, false) == InvalidOffsetNumber) |
| 77 |
elog(PANIC, "_bt_restore_page: cannot add item to page"); |
| 78 |
} |
| 79 |
} |
| 80 |
|
| 81 |
static void |
| 82 |
_bt_restore_meta(XLogReaderState *record, uint8 block_id) |
| 83 |
{ |
| 84 |
XLogRecPtr lsn = record->EndRecPtr; |
| 85 |
Buffer metabuf; |
| 86 |
Page metapg; |
| 87 |
BTMetaPageData *md; |
| 88 |
BTPageOpaque pageop; |
| 89 |
xl_btree_metadata *xlrec; |
| 90 |
char *ptr; |
| 91 |
Size len; |
| 92 |
|
| 93 |
metabuf = XLogInitBufferForRedo(record, block_id); |
| 94 |
ptr = XLogRecGetBlockData(record, block_id, &len); |
| 95 |
|
| 96 |
Assert(len == sizeof(xl_btree_metadata)); |
| 97 |
Assert(BufferGetBlockNumber(metabuf) == BTREE_METAPAGE); |
| 98 |
xlrec = (xl_btree_metadata *) ptr; |
| 99 |
metapg = BufferGetPage(metabuf); |
| 100 |
|
| 101 |
_bt_pageinit(metapg, BufferGetPageSize(metabuf)); |
| 102 |
|
| 103 |
md = BTPageGetMeta(metapg); |
| 104 |
md->btm_magic = BTREE_MAGIC; |
| 105 |
md->btm_version = xlrec->version; |
| 106 |
md->btm_root = xlrec->root; |
| 107 |
md->btm_level = xlrec->level; |
| 108 |
md->btm_fastroot = xlrec->fastroot; |
| 109 |
md->btm_fastlevel = xlrec->fastlevel; |
| 110 |
/* Cannot log BTREE_MIN_VERSION index metapage without upgrade */ |
| 111 |
Assert(md->btm_version >= BTREE_NOVAC_VERSION); |
| 112 |
md->btm_last_cleanup_num_delpages = xlrec->last_cleanup_num_delpages; |
| 113 |
md->btm_last_cleanup_num_heap_tuples = -1.0; |
| 114 |
md->btm_allequalimage = xlrec->allequalimage; |
| 115 |
|
| 116 |
pageop = BTPageGetOpaque(metapg); |
| 117 |
pageop->btpo_flags = BTP_META; |
| 118 |
|
| 119 |
/* |
| 120 |
* Set pd_lower just past the end of the metadata. This is essential, |
| 121 |
* because without doing so, metadata will be lost if xlog.c compresses |
| 122 |
* the page. |
| 123 |
*/ |
| 124 |
((PageHeader) metapg)->pd_lower = |
| 125 |
((char *) md + sizeof(BTMetaPageData)) - (char *) metapg; |
| 126 |
|
| 127 |
PageSetLSN(metapg, lsn); |
| 128 |
MarkBufferDirty(metabuf); |
| 129 |
UnlockReleaseBuffer(metabuf); |
| 130 |
} |
| 131 |
|
| 132 |
/* |
| 133 |
* _bt_clear_incomplete_split -- clear INCOMPLETE_SPLIT flag on a page |
| 134 |
* |
| 135 |
* This is a common subroutine of the redo functions of all the WAL record |
| 136 |
* types that can insert a downlink: insert, split, and newroot. |
| 137 |
*/ |
| 138 |
static void |
| 139 |
_bt_clear_incomplete_split(XLogReaderState *record, uint8 block_id) |
| 140 |
{ |
| 141 |
XLogRecPtr lsn = record->EndRecPtr; |
| 142 |
Buffer buf; |
| 143 |
|
| 144 |
if (XLogReadBufferForRedo(record, block_id, &buf) == BLK_NEEDS_REDO) |
| 145 |
{ |
| 146 |
Page page = (Page) BufferGetPage(buf); |
| 147 |
BTPageOpaque pageop = BTPageGetOpaque(page); |
| 148 |
|
| 149 |
Assert(P_INCOMPLETE_SPLIT(pageop)); |
| 150 |
pageop->btpo_flags &= ~BTP_INCOMPLETE_SPLIT; |
| 151 |
|
| 152 |
PageSetLSN(page, lsn); |
| 153 |
MarkBufferDirty(buf); |
| 154 |
} |
| 155 |
if (BufferIsValid(buf)) |
| 156 |
UnlockReleaseBuffer(buf); |
| 157 |
} |
| 158 |
|
| 159 |
static void |
| 160 |
btree_xlog_insert(bool isleaf, bool ismeta, bool posting, |
| 161 |
XLogReaderState *record) |
| 162 |
{ |
| 163 |
XLogRecPtr lsn = record->EndRecPtr; |
| 164 |
xl_btree_insert *xlrec = (xl_btree_insert *) XLogRecGetData(record); |
| 165 |
Buffer buffer; |
| 166 |
Page page; |
| 167 |
|
| 168 |
/* |
| 169 |
* Insertion to an internal page finishes an incomplete split at the child |
| 170 |
* level. Clear the incomplete-split flag in the child. Note: during |
| 171 |
* normal operation, the child and parent pages are locked at the same |
| 172 |
* time (the locks are coupled), so that clearing the flag and inserting |
| 173 |
* the downlink appear atomic to other backends. We don't bother with |
| 174 |
* that during replay, because readers don't care about the |
| 175 |
* incomplete-split flag and there cannot be updates happening. |
| 176 |
*/ |
| 177 |
if (!isleaf) |
| 178 |
_bt_clear_incomplete_split(record, 1); |
| 179 |
if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO) |
| 180 |
{ |
| 181 |
Size datalen; |
| 182 |
char *datapos = XLogRecGetBlockData(record, 0, &datalen); |
| 183 |
|
| 184 |
page = BufferGetPage(buffer); |
| 185 |
|
| 186 |
if (!posting) |
| 187 |
{ |
| 188 |
/* Simple retail insertion */ |
| 189 |
if (PageAddItem(page, (Item) datapos, datalen, xlrec->offnum, |
| 190 |
false, false) == InvalidOffsetNumber) |
| 191 |
elog(PANIC, "failed to add new item"); |
| 192 |
} |
| 193 |
else |
| 194 |
{ |
| 195 |
ItemId itemid; |
| 196 |
IndexTuple oposting, |
| 197 |
newitem, |
| 198 |
nposting; |
| 199 |
uint16 postingoff; |
| 200 |
|
| 201 |
/* |
| 202 |
* A posting list split occurred during leaf page insertion. WAL |
| 203 |
* record data will start with an offset number representing the |
| 204 |
* point in an existing posting list that a split occurs at. |
| 205 |
* |
| 206 |
* Use _bt_swap_posting() to repeat posting list split steps from |
| 207 |
* primary. Note that newitem from WAL record is 'orignewitem', |
| 208 |
* not the final version of newitem that is actually inserted on |
| 209 |
* page. |
| 210 |
*/ |
| 211 |
postingoff = *((uint16 *) datapos); |
| 212 |
datapos += sizeof(uint16); |
| 213 |
datalen -= sizeof(uint16); |
| 214 |
|
| 215 |
itemid = PageGetItemId(page, OffsetNumberPrev(xlrec->offnum)); |
| 216 |
oposting = (IndexTuple) PageGetItem(page, itemid); |
| 217 |
|
| 218 |
/* Use mutable, aligned newitem copy in _bt_swap_posting() */ |
| 219 |
Assert(isleaf && postingoff > 0); |
| 220 |
newitem = CopyIndexTuple((IndexTuple) datapos); |
| 221 |
nposting = _bt_swap_posting(newitem, oposting, postingoff); |
| 222 |
|
| 223 |
/* Replace existing posting list with post-split version */ |
| 224 |
memcpy(oposting, nposting, MAXALIGN(IndexTupleSize(nposting))); |
| 225 |
|
| 226 |
/* Insert "final" new item (not orignewitem from WAL stream) */ |
| 227 |
Assert(IndexTupleSize(newitem) == datalen); |
| 228 |
if (PageAddItem(page, (Item) newitem, datalen, xlrec->offnum, |
| 229 |
false, false) == InvalidOffsetNumber) |
| 230 |
elog(PANIC, "failed to add posting split new item"); |
| 231 |
} |
| 232 |
|
| 233 |
PageSetLSN(page, lsn); |
| 234 |
MarkBufferDirty(buffer); |
| 235 |
} |
| 236 |
if (BufferIsValid(buffer)) |
| 237 |
UnlockReleaseBuffer(buffer); |
| 238 |
|
| 239 |
/* |
| 240 |
* Note: in normal operation, we'd update the metapage while still holding |
| 241 |
* lock on the page we inserted into. But during replay it's not |
| 242 |
* necessary to hold that lock, since no other index updates can be |
| 243 |
* happening concurrently, and readers will cope fine with following an |
| 244 |
* obsolete link from the metapage. |
| 245 |
*/ |
| 246 |
if (ismeta) |
| 247 |
_bt_restore_meta(record, 2); |
| 248 |
} |
| 249 |
|
| 250 |
static void |
| 251 |
btree_xlog_split(bool newitemonleft, XLogReaderState *record) |
| 252 |
{ |
| 253 |
XLogRecPtr lsn = record->EndRecPtr; |
| 254 |
xl_btree_split *xlrec = (xl_btree_split *) XLogRecGetData(record); |
| 255 |
bool isleaf = (xlrec->level == 0); |
| 256 |
Buffer buf; |
| 257 |
Buffer rbuf; |
| 258 |
Page rpage; |
| 259 |
BTPageOpaque ropaque; |
| 260 |
char *datapos; |
| 261 |
Size datalen; |
| 262 |
BlockNumber origpagenumber; |
| 263 |
BlockNumber rightpagenumber; |
| 264 |
BlockNumber spagenumber; |
| 265 |
|
| 266 |
XLogRecGetBlockTag(record, 0, NULL, NULL, &origpagenumber); |
| 267 |
XLogRecGetBlockTag(record, 1, NULL, NULL, &rightpagenumber); |
| 268 |
if (!XLogRecGetBlockTagExtended(record, 2, NULL, NULL, &spagenumber, NULL)) |
| 269 |
spagenumber = P_NONE; |
| 270 |
|
| 271 |
/* |
| 272 |
* Clear the incomplete split flag on the appropriate child page one level |
| 273 |
* down when origpage/buf is an internal page (there must have been |
| 274 |
* cascading page splits during original execution in the event of an |
| 275 |
* internal page split). This is like the corresponding btree_xlog_insert |
| 276 |
* call for internal pages. We're not clearing the incomplete split flag |
| 277 |
* for the current page split here (you can think of this as part of the |
| 278 |
* insert of newitem that the page split action needs to perform in |
| 279 |
* passing). |
| 280 |
* |
| 281 |
* Like in btree_xlog_insert, this can be done before locking other pages. |
| 282 |
* We never need to couple cross-level locks in REDO routines. |
| 283 |
*/ |
| 284 |
if (!isleaf) |
| 285 |
_bt_clear_incomplete_split(record, 3); |
| 286 |
|
| 287 |
/* Reconstruct right (new) sibling page from scratch */ |
| 288 |
rbuf = XLogInitBufferForRedo(record, 1); |
| 289 |
datapos = XLogRecGetBlockData(record, 1, &datalen); |
| 290 |
rpage = (Page) BufferGetPage(rbuf); |
| 291 |
|
| 292 |
_bt_pageinit(rpage, BufferGetPageSize(rbuf)); |
| 293 |
ropaque = BTPageGetOpaque(rpage); |
| 294 |
|
| 295 |
ropaque->btpo_prev = origpagenumber; |
| 296 |
ropaque->btpo_next = spagenumber; |
| 297 |
ropaque->btpo_level = xlrec->level; |
| 298 |
ropaque->btpo_flags = isleaf ? BTP_LEAF : 0; |
| 299 |
ropaque->btpo_cycleid = 0; |
| 300 |
|
| 301 |
_bt_restore_page(rpage, datapos, datalen); |
| 302 |
|
| 303 |
PageSetLSN(rpage, lsn); |
| 304 |
MarkBufferDirty(rbuf); |
| 305 |
|
| 306 |
/* Now reconstruct original page (left half of split) */ |
| 307 |
if (XLogReadBufferForRedo(record, 0, &buf) == BLK_NEEDS_REDO) |
| 308 |
{ |
| 309 |
/* |
| 310 |
* To retain the same physical order of the tuples that they had, we |
| 311 |
* initialize a temporary empty page for the left page and add all the |
| 312 |
* items to that in item number order. This mirrors how _bt_split() |
| 313 |
* works. Retaining the same physical order makes WAL consistency |
| 314 |
* checking possible. See also _bt_restore_page(), which does the |
| 315 |
* same for the right page. |
| 316 |
*/ |
| 317 |
Page origpage = (Page) BufferGetPage(buf); |
| 318 |
BTPageOpaque oopaque = BTPageGetOpaque(origpage); |
| 319 |
OffsetNumber off; |
| 320 |
IndexTuple newitem = NULL, |
| 321 |
left_hikey = NULL, |
| 322 |
nposting = NULL; |
| 323 |
Size newitemsz = 0, |
| 324 |
left_hikeysz = 0; |
| 325 |
Page leftpage; |
| 326 |
OffsetNumber leftoff, |
| 327 |
replacepostingoff = InvalidOffsetNumber; |
| 328 |
|
| 329 |
datapos = XLogRecGetBlockData(record, 0, &datalen); |
| 330 |
|
| 331 |
if (newitemonleft || xlrec->postingoff != 0) |
| 332 |
{ |
| 333 |
newitem = (IndexTuple) datapos; |
| 334 |
newitemsz = MAXALIGN(IndexTupleSize(newitem)); |
| 335 |
datapos += newitemsz; |
| 336 |
datalen -= newitemsz; |
| 337 |
|
| 338 |
if (xlrec->postingoff != 0) |
| 339 |
{ |
| 340 |
ItemId itemid; |
| 341 |
IndexTuple oposting; |
| 342 |
|
| 343 |
/* Posting list must be at offset number before new item's */ |
| 344 |
replacepostingoff = OffsetNumberPrev(xlrec->newitemoff); |
| 345 |
|
| 346 |
/* Use mutable, aligned newitem copy in _bt_swap_posting() */ |
| 347 |
newitem = CopyIndexTuple(newitem); |
| 348 |
itemid = PageGetItemId(origpage, replacepostingoff); |
| 349 |
oposting = (IndexTuple) PageGetItem(origpage, itemid); |
| 350 |
nposting = _bt_swap_posting(newitem, oposting, |
| 351 |
xlrec->postingoff); |
| 352 |
} |
| 353 |
} |
| 354 |
|
| 355 |
/* |
| 356 |
* Extract left hikey and its size. We assume that 16-bit alignment |
| 357 |
* is enough to apply IndexTupleSize (since it's fetching from a |
| 358 |
* uint16 field). |
| 359 |
*/ |
| 360 |
left_hikey = (IndexTuple) datapos; |
| 361 |
left_hikeysz = MAXALIGN(IndexTupleSize(left_hikey)); |
| 362 |
datapos += left_hikeysz; |
| 363 |
datalen -= left_hikeysz; |
| 364 |
|
| 365 |
Assert(datalen == 0); |
| 366 |
|
| 367 |
leftpage = PageGetTempPageCopySpecial(origpage); |
| 368 |
|
| 369 |
/* Add high key tuple from WAL record to temp page */ |
| 370 |
leftoff = P_HIKEY; |
| 371 |
if (PageAddItem(leftpage, (Item) left_hikey, left_hikeysz, P_HIKEY, |
| 372 |
false, false) == InvalidOffsetNumber) |
| 373 |
elog(ERROR, "failed to add high key to left page after split"); |
| 374 |
leftoff = OffsetNumberNext(leftoff); |
| 375 |
|
| 376 |
for (off = P_FIRSTDATAKEY(oopaque); off < xlrec->firstrightoff; off++) |
| 377 |
{ |
| 378 |
ItemId itemid; |
| 379 |
Size itemsz; |
| 380 |
IndexTuple item; |
| 381 |
|
| 382 |
/* Add replacement posting list when required */ |
| 383 |
if (off == replacepostingoff) |
| 384 |
{ |
| 385 |
Assert(newitemonleft || |
| 386 |
xlrec->firstrightoff == xlrec->newitemoff); |
| 387 |
if (PageAddItem(leftpage, (Item) nposting, |
| 388 |
MAXALIGN(IndexTupleSize(nposting)), leftoff, |
| 389 |
false, false) == InvalidOffsetNumber) |
| 390 |
elog(ERROR, "failed to add new posting list item to left page after split"); |
| 391 |
leftoff = OffsetNumberNext(leftoff); |
| 392 |
continue; /* don't insert oposting */ |
| 393 |
} |
| 394 |
|
| 395 |
/* add the new item if it was inserted on left page */ |
| 396 |
else if (newitemonleft && off == xlrec->newitemoff) |
| 397 |
{ |
| 398 |
if (PageAddItem(leftpage, (Item) newitem, newitemsz, leftoff, |
| 399 |
false, false) == InvalidOffsetNumber) |
| 400 |
elog(ERROR, "failed to add new item to left page after split"); |
| 401 |
leftoff = OffsetNumberNext(leftoff); |
| 402 |
} |
| 403 |
|
| 404 |
itemid = PageGetItemId(origpage, off); |
| 405 |
itemsz = ItemIdGetLength(itemid); |
| 406 |
item = (IndexTuple) PageGetItem(origpage, itemid); |
| 407 |
if (PageAddItem(leftpage, (Item) item, itemsz, leftoff, |
| 408 |
false, false) == InvalidOffsetNumber) |
| 409 |
elog(ERROR, "failed to add old item to left page after split"); |
| 410 |
leftoff = OffsetNumberNext(leftoff); |
| 411 |
} |
| 412 |
|
| 413 |
/* cope with possibility that newitem goes at the end */ |
| 414 |
if (newitemonleft && off == xlrec->newitemoff) |
| 415 |
{ |
| 416 |
if (PageAddItem(leftpage, (Item) newitem, newitemsz, leftoff, |
| 417 |
false, false) == InvalidOffsetNumber) |
| 418 |
elog(ERROR, "failed to add new item to left page after split"); |
| 419 |
leftoff = OffsetNumberNext(leftoff); |
| 420 |
} |
| 421 |
|
| 422 |
PageRestoreTempPage(leftpage, origpage); |
| 423 |
|
| 424 |
/* Fix opaque fields */ |
| 425 |
oopaque->btpo_flags = BTP_INCOMPLETE_SPLIT; |
| 426 |
if (isleaf) |
| 427 |
oopaque->btpo_flags |= BTP_LEAF; |
| 428 |
oopaque->btpo_next = rightpagenumber; |
| 429 |
oopaque->btpo_cycleid = 0; |
| 430 |
|
| 431 |
PageSetLSN(origpage, lsn); |
| 432 |
MarkBufferDirty(buf); |
| 433 |
} |
| 434 |
|
| 435 |
/* Fix left-link of the page to the right of the new right sibling */ |
| 436 |
if (spagenumber != P_NONE) |
| 437 |
{ |
| 438 |
Buffer sbuf; |
| 439 |
|
| 440 |
if (XLogReadBufferForRedo(record, 2, &sbuf) == BLK_NEEDS_REDO) |
| 441 |
{ |
| 442 |
Page spage = (Page) BufferGetPage(sbuf); |
| 443 |
BTPageOpaque spageop = BTPageGetOpaque(spage); |
| 444 |
|
| 445 |
spageop->btpo_prev = rightpagenumber; |
| 446 |
|
| 447 |
PageSetLSN(spage, lsn); |
| 448 |
MarkBufferDirty(sbuf); |
| 449 |
} |
| 450 |
if (BufferIsValid(sbuf)) |
| 451 |
UnlockReleaseBuffer(sbuf); |
| 452 |
} |
| 453 |
|
| 454 |
/* |
| 455 |
* Finally, release the remaining buffers. sbuf, rbuf, and buf must be |
| 456 |
* released together, so that readers cannot observe inconsistencies. |
| 457 |
*/ |
| 458 |
UnlockReleaseBuffer(rbuf); |
| 459 |
if (BufferIsValid(buf)) |
| 460 |
UnlockReleaseBuffer(buf); |
| 461 |
} |
| 462 |
|
| 463 |
static void |
| 464 |
btree_xlog_dedup(XLogReaderState *record) |
| 465 |
{ |
| 466 |
XLogRecPtr lsn = record->EndRecPtr; |
| 467 |
xl_btree_dedup *xlrec = (xl_btree_dedup *) XLogRecGetData(record); |
| 468 |
Buffer buf; |
| 469 |
|
| 470 |
if (XLogReadBufferForRedo(record, 0, &buf) == BLK_NEEDS_REDO) |
| 471 |
{ |
| 472 |
char *ptr = XLogRecGetBlockData(record, 0, NULL); |
| 473 |
Page page = (Page) BufferGetPage(buf); |
| 474 |
BTPageOpaque opaque = BTPageGetOpaque(page); |
| 475 |
OffsetNumber offnum, |
| 476 |
minoff, |
| 477 |
maxoff; |
| 478 |
BTDedupState state; |
| 479 |
BTDedupInterval *intervals; |
| 480 |
Page newpage; |
| 481 |
|
| 482 |
state = (BTDedupState) palloc(sizeof(BTDedupStateData)); |
| 483 |
state->deduplicate = true; /* unused */ |
| 484 |
state->nmaxitems = 0; /* unused */ |
| 485 |
/* Conservatively use larger maxpostingsize than primary */ |
| 486 |
state->maxpostingsize = BTMaxItemSize; |
| 487 |
state->base = NULL; |
| 488 |
state->baseoff = InvalidOffsetNumber; |
| 489 |
state->basetupsize = 0; |
| 490 |
state->htids = palloc(state->maxpostingsize); |
| 491 |
state->nhtids = 0; |
| 492 |
state->nitems = 0; |
| 493 |
state->phystupsize = 0; |
| 494 |
state->nintervals = 0; |
| 495 |
|
| 496 |
minoff = P_FIRSTDATAKEY(opaque); |
| 497 |
maxoff = PageGetMaxOffsetNumber(page); |
| 498 |
newpage = PageGetTempPageCopySpecial(page); |
| 499 |
|
| 500 |
if (!P_RIGHTMOST(opaque)) |
| 501 |
{ |
| 502 |
ItemId itemid = PageGetItemId(page, P_HIKEY); |
| 503 |
Size itemsz = ItemIdGetLength(itemid); |
| 504 |
IndexTuple item = (IndexTuple) PageGetItem(page, itemid); |
| 505 |
|
| 506 |
if (PageAddItem(newpage, (Item) item, itemsz, P_HIKEY, |
| 507 |
false, false) == InvalidOffsetNumber) |
| 508 |
elog(ERROR, "deduplication failed to add highkey"); |
| 509 |
} |
| 510 |
|
| 511 |
intervals = (BTDedupInterval *) ptr; |
| 512 |
for (offnum = minoff; |
| 513 |
offnum <= maxoff; |
| 514 |
offnum = OffsetNumberNext(offnum)) |
| 515 |
{ |
| 516 |
ItemId itemid = PageGetItemId(page, offnum); |
| 517 |
IndexTuple itup = (IndexTuple) PageGetItem(page, itemid); |
| 518 |
|
| 519 |
if (offnum == minoff) |
| 520 |
_bt_dedup_start_pending(state, itup, offnum); |
| 521 |
else if (state->nintervals < xlrec->nintervals && |
| 522 |
state->baseoff == intervals[state->nintervals].baseoff && |
| 523 |
state->nitems < intervals[state->nintervals].nitems) |
| 524 |
{ |
| 525 |
if (!_bt_dedup_save_htid(state, itup)) |
| 526 |
elog(ERROR, "deduplication failed to add heap tid to pending posting list"); |
| 527 |
} |
| 528 |
else |
| 529 |
{ |
| 530 |
_bt_dedup_finish_pending(newpage, state); |
| 531 |
_bt_dedup_start_pending(state, itup, offnum); |
| 532 |
} |
| 533 |
} |
| 534 |
|
| 535 |
_bt_dedup_finish_pending(newpage, state); |
| 536 |
Assert(state->nintervals == xlrec->nintervals); |
| 537 |
Assert(memcmp(state->intervals, intervals, |
| 538 |
state->nintervals * sizeof(BTDedupInterval)) == 0); |
| 539 |
|
| 540 |
if (P_HAS_GARBAGE(opaque)) |
| 541 |
{ |
| 542 |
BTPageOpaque nopaque = BTPageGetOpaque(newpage); |
| 543 |
|
| 544 |
nopaque->btpo_flags &= ~BTP_HAS_GARBAGE; |
| 545 |
} |
| 546 |
|
| 547 |
PageRestoreTempPage(newpage, page); |
| 548 |
PageSetLSN(page, lsn); |
| 549 |
MarkBufferDirty(buf); |
| 550 |
} |
| 551 |
|
| 552 |
if (BufferIsValid(buf)) |
| 553 |
UnlockReleaseBuffer(buf); |
| 554 |
} |
| 555 |
|
| 556 |
static void |
| 557 |
btree_xlog_updates(Page page, OffsetNumber *updatedoffsets, |
| 558 |
xl_btree_update *updates, int nupdated) |
| 559 |
{ |
| 560 |
BTVacuumPosting vacposting; |
| 561 |
IndexTuple origtuple; |
| 562 |
ItemId itemid; |
| 563 |
Size itemsz; |
| 564 |
|
| 565 |
for (int i = 0; i < nupdated; i++) |
| 566 |
{ |
| 567 |
itemid = PageGetItemId(page, updatedoffsets[i]); |
| 568 |
origtuple = (IndexTuple) PageGetItem(page, itemid); |
| 569 |
|
| 570 |
vacposting = palloc(offsetof(BTVacuumPostingData, deletetids) + |
| 571 |
updates->ndeletedtids * sizeof(uint16)); |
| 572 |
vacposting->updatedoffset = updatedoffsets[i]; |
| 573 |
vacposting->itup = origtuple; |
| 574 |
vacposting->ndeletedtids = updates->ndeletedtids; |
| 575 |
memcpy(vacposting->deletetids, |
| 576 |
(char *) updates + SizeOfBtreeUpdate, |
| 577 |
updates->ndeletedtids * sizeof(uint16)); |
| 578 |
|
| 579 |
_bt_update_posting(vacposting); |
| 580 |
|
| 581 |
/* Overwrite updated version of tuple */ |
| 582 |
itemsz = MAXALIGN(IndexTupleSize(vacposting->itup)); |
| 583 |
if (!PageIndexTupleOverwrite(page, updatedoffsets[i], |
| 584 |
(Item) vacposting->itup, itemsz)) |
| 585 |
elog(PANIC, "failed to update partially dead item"); |
| 586 |
|
| 587 |
pfree(vacposting->itup); |
| 588 |
pfree(vacposting); |
| 589 |
|
| 590 |
/* advance to next xl_btree_update from array */ |
| 591 |
updates = (xl_btree_update *) |
| 592 |
((char *) updates + SizeOfBtreeUpdate + |
| 593 |
updates->ndeletedtids * sizeof(uint16)); |
| 594 |
} |
| 595 |
} |
| 596 |
|
| 597 |
static void |
| 598 |
btree_xlog_vacuum(XLogReaderState *record) |
| 599 |
{ |
| 600 |
XLogRecPtr lsn = record->EndRecPtr; |
| 601 |
xl_btree_vacuum *xlrec = (xl_btree_vacuum *) XLogRecGetData(record); |
| 602 |
Buffer buffer; |
| 603 |
Page page; |
| 604 |
BTPageOpaque opaque; |
| 605 |
|
| 606 |
/* |
| 607 |
* We need to take a cleanup lock here, just like btvacuumpage(). However, |
| 608 |
* it isn't necessary to exhaustively get a cleanup lock on every block in |
| 609 |
* the index during recovery (just getting a cleanup lock on pages with |
| 610 |
* items to kill suffices). See nbtree/README for details. |
| 611 |
*/ |
| 612 |
if (XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true, &buffer) |
| 613 |
== BLK_NEEDS_REDO) |
| 614 |
{ |
| 615 |
char *ptr = XLogRecGetBlockData(record, 0, NULL); |
| 616 |
|
| 617 |
page = (Page) BufferGetPage(buffer); |
| 618 |
|
| 619 |
if (xlrec->nupdated > 0) |
| 620 |
{ |
| 621 |
OffsetNumber *updatedoffsets; |
| 622 |
xl_btree_update *updates; |
| 623 |
|
| 624 |
updatedoffsets = (OffsetNumber *) |
| 625 |
(ptr + xlrec->ndeleted * sizeof(OffsetNumber)); |
| 626 |
updates = (xl_btree_update *) ((char *) updatedoffsets + |
| 627 |
xlrec->nupdated * |
| 628 |
sizeof(OffsetNumber)); |
| 629 |
|
| 630 |
btree_xlog_updates(page, updatedoffsets, updates, xlrec->nupdated); |
| 631 |
} |
| 632 |
|
| 633 |
if (xlrec->ndeleted > 0) |
| 634 |
PageIndexMultiDelete(page, (OffsetNumber *) ptr, xlrec->ndeleted); |
| 635 |
|
| 636 |
/* |
| 637 |
* Clear the vacuum cycle ID, and mark the page as not containing any |
| 638 |
* LP_DEAD items |
| 639 |
*/ |
| 640 |
opaque = BTPageGetOpaque(page); |
| 641 |
opaque->btpo_cycleid = 0; |
| 642 |
opaque->btpo_flags &= ~BTP_HAS_GARBAGE; |
| 643 |
|
| 644 |
PageSetLSN(page, lsn); |
| 645 |
MarkBufferDirty(buffer); |
| 646 |
} |
| 647 |
if (BufferIsValid(buffer)) |
| 648 |
UnlockReleaseBuffer(buffer); |
| 649 |
} |
| 650 |
|
| 651 |
static void |
| 652 |
btree_xlog_delete(XLogReaderState *record) |
| 653 |
{ |
| 654 |
XLogRecPtr lsn = record->EndRecPtr; |
| 655 |
xl_btree_delete *xlrec = (xl_btree_delete *) XLogRecGetData(record); |
| 656 |
Buffer buffer; |
| 657 |
Page page; |
| 658 |
BTPageOpaque opaque; |
| 659 |
|
| 660 |
/* |
| 661 |
* If we have any conflict processing to do, it must happen before we |
| 662 |
* update the page |
| 663 |
*/ |
| 664 |
if (InHotStandby) |
| 665 |
{ |
| 666 |
RelFileLocator rlocator; |
| 667 |
|
| 668 |
XLogRecGetBlockTag(record, 0, &rlocator, NULL, NULL); |
| 669 |
|
| 670 |
ResolveRecoveryConflictWithSnapshot(xlrec->snapshotConflictHorizon, |
| 671 |
xlrec->isCatalogRel, |
| 672 |
rlocator); |
| 673 |
} |
| 674 |
|
| 675 |
/* |
| 676 |
* We don't need to take a cleanup lock to apply these changes. See |
| 677 |
* nbtree/README for details. |
| 678 |
*/ |
| 679 |
if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO) |
| 680 |
{ |
| 681 |
char *ptr = XLogRecGetBlockData(record, 0, NULL); |
| 682 |
|
| 683 |
page = (Page) BufferGetPage(buffer); |
| 684 |
|
| 685 |
if (xlrec->nupdated > 0) |
| 686 |
{ |
| 687 |
OffsetNumber *updatedoffsets; |
| 688 |
xl_btree_update *updates; |
| 689 |
|
| 690 |
updatedoffsets = (OffsetNumber *) |
| 691 |
(ptr + xlrec->ndeleted * sizeof(OffsetNumber)); |
| 692 |
updates = (xl_btree_update *) ((char *) updatedoffsets + |
| 693 |
xlrec->nupdated * |
| 694 |
sizeof(OffsetNumber)); |
| 695 |
|
| 696 |
btree_xlog_updates(page, updatedoffsets, updates, xlrec->nupdated); |
| 697 |
} |
| 698 |
|
| 699 |
if (xlrec->ndeleted > 0) |
| 700 |
PageIndexMultiDelete(page, (OffsetNumber *) ptr, xlrec->ndeleted); |
| 701 |
|
| 702 |
/* |
| 703 |
* Do *not* clear the vacuum cycle ID, but do mark the page as not |
| 704 |
* containing any LP_DEAD items |
| 705 |
*/ |
| 706 |
opaque = BTPageGetOpaque(page); |
| 707 |
opaque->btpo_flags &= ~BTP_HAS_GARBAGE; |
| 708 |
|
| 709 |
PageSetLSN(page, lsn); |
| 710 |
MarkBufferDirty(buffer); |
| 711 |
} |
| 712 |
if (BufferIsValid(buffer)) |
| 713 |
UnlockReleaseBuffer(buffer); |
| 714 |
} |
| 715 |
|
| 716 |
static void |
| 717 |
btree_xlog_mark_page_halfdead(uint8 info, XLogReaderState *record) |
| 718 |
{ |
| 719 |
XLogRecPtr lsn = record->EndRecPtr; |
| 720 |
xl_btree_mark_page_halfdead *xlrec = (xl_btree_mark_page_halfdead *) XLogRecGetData(record); |
| 721 |
Buffer buffer; |
| 722 |
Page page; |
| 723 |
BTPageOpaque pageop; |
| 724 |
IndexTupleData trunctuple; |
| 725 |
|
| 726 |
/* |
| 727 |
* In normal operation, we would lock all the pages this WAL record |
| 728 |
* touches before changing any of them. In WAL replay, it should be okay |
| 729 |
* to lock just one page at a time, since no concurrent index updates can |
| 730 |
* be happening, and readers should not care whether they arrive at the |
| 731 |
* target page or not (since it's surely empty). |
| 732 |
*/ |
| 733 |
|
| 734 |
/* to-be-deleted subtree's parent page */ |
| 735 |
if (XLogReadBufferForRedo(record, 1, &buffer) == BLK_NEEDS_REDO) |
| 736 |
{ |
| 737 |
OffsetNumber poffset; |
| 738 |
ItemId itemid; |
| 739 |
IndexTuple itup; |
| 740 |
OffsetNumber nextoffset; |
| 741 |
BlockNumber rightsib; |
| 742 |
|
| 743 |
page = (Page) BufferGetPage(buffer); |
| 744 |
pageop = BTPageGetOpaque(page); |
| 745 |
|
| 746 |
poffset = xlrec->poffset; |
| 747 |
|
| 748 |
nextoffset = OffsetNumberNext(poffset); |
| 749 |
itemid = PageGetItemId(page, nextoffset); |
| 750 |
itup = (IndexTuple) PageGetItem(page, itemid); |
| 751 |
rightsib = BTreeTupleGetDownLink(itup); |
| 752 |
|
| 753 |
itemid = PageGetItemId(page, poffset); |
| 754 |
itup = (IndexTuple) PageGetItem(page, itemid); |
| 755 |
BTreeTupleSetDownLink(itup, rightsib); |
| 756 |
nextoffset = OffsetNumberNext(poffset); |
| 757 |
PageIndexTupleDelete(page, nextoffset); |
| 758 |
|
| 759 |
PageSetLSN(page, lsn); |
| 760 |
MarkBufferDirty(buffer); |
| 761 |
} |
| 762 |
|
| 763 |
/* |
| 764 |
* Don't need to couple cross-level locks in REDO routines, so release |
| 765 |
* lock on internal page immediately |
| 766 |
*/ |
| 767 |
if (BufferIsValid(buffer)) |
| 768 |
UnlockReleaseBuffer(buffer); |
| 769 |
|
| 770 |
/* Rewrite the leaf page as a halfdead page */ |
| 771 |
buffer = XLogInitBufferForRedo(record, 0); |
| 772 |
page = (Page) BufferGetPage(buffer); |
| 773 |
|
| 774 |
_bt_pageinit(page, BufferGetPageSize(buffer)); |
| 775 |
pageop = BTPageGetOpaque(page); |
| 776 |
|
| 777 |
pageop->btpo_prev = xlrec->leftblk; |
| 778 |
pageop->btpo_next = xlrec->rightblk; |
| 779 |
pageop->btpo_level = 0; |
| 780 |
pageop->btpo_flags = BTP_HALF_DEAD | BTP_LEAF; |
| 781 |
pageop->btpo_cycleid = 0; |
| 782 |
|
| 783 |
/* |
| 784 |
* Construct a dummy high key item that points to top parent page (value |
| 785 |
* is InvalidBlockNumber when the top parent page is the leaf page itself) |
| 786 |
*/ |
| 787 |
MemSet(&trunctuple, 0, sizeof(IndexTupleData)); |
| 788 |
trunctuple.t_info = sizeof(IndexTupleData); |
| 789 |
BTreeTupleSetTopParent(&trunctuple, xlrec->topparent); |
| 790 |
|
| 791 |
if (PageAddItem(page, (Item) &trunctuple, sizeof(IndexTupleData), P_HIKEY, |
| 792 |
false, false) == InvalidOffsetNumber) |
| 793 |
elog(ERROR, "could not add dummy high key to half-dead page"); |
| 794 |
|
| 795 |
PageSetLSN(page, lsn); |
| 796 |
MarkBufferDirty(buffer); |
| 797 |
UnlockReleaseBuffer(buffer); |
| 798 |
} |
| 799 |
|
| 800 |
|
| 801 |
static void |
| 802 |
btree_xlog_unlink_page(uint8 info, XLogReaderState *record) |
| 803 |
{ |
| 804 |
XLogRecPtr lsn = record->EndRecPtr; |
| 805 |
xl_btree_unlink_page *xlrec = (xl_btree_unlink_page *) XLogRecGetData(record); |
| 806 |
BlockNumber leftsib; |
| 807 |
BlockNumber rightsib; |
| 808 |
uint32 level; |
| 809 |
bool isleaf; |
| 810 |
FullTransactionId safexid; |
| 811 |
Buffer leftbuf; |
| 812 |
Buffer target; |
| 813 |
Buffer rightbuf; |
| 814 |
Page page; |
| 815 |
BTPageOpaque pageop; |
| 816 |
|
| 817 |
leftsib = xlrec->leftsib; |
| 818 |
rightsib = xlrec->rightsib; |
| 819 |
level = xlrec->level; |
| 820 |
isleaf = (level == 0); |
| 821 |
safexid = xlrec->safexid; |
| 822 |
|
| 823 |
/* No leaftopparent for level 0 (leaf page) or level 1 target */ |
| 824 |
Assert(!BlockNumberIsValid(xlrec->leaftopparent) || level > 1); |
| 825 |
|
| 826 |
/* |
| 827 |
* In normal operation, we would lock all the pages this WAL record |
| 828 |
* touches before changing any of them. In WAL replay, we at least lock |
| 829 |
* the pages in the same standard left-to-right order (leftsib, target, |
| 830 |
* rightsib), and don't release the sibling locks until the target is |
| 831 |
* marked deleted. |
| 832 |
*/ |
| 833 |
|
| 834 |
/* Fix right-link of left sibling, if any */ |
| 835 |
if (leftsib != P_NONE) |
| 836 |
{ |
| 837 |
if (XLogReadBufferForRedo(record, 1, &leftbuf) == BLK_NEEDS_REDO) |
| 838 |
{ |
| 839 |
page = (Page) BufferGetPage(leftbuf); |
| 840 |
pageop = BTPageGetOpaque(page); |
| 841 |
pageop->btpo_next = rightsib; |
| 842 |
|
| 843 |
PageSetLSN(page, lsn); |
| 844 |
MarkBufferDirty(leftbuf); |
| 845 |
} |
| 846 |
} |
| 847 |
else |
| 848 |
leftbuf = InvalidBuffer; |
| 849 |
|
| 850 |
/* Rewrite target page as empty deleted page */ |
| 851 |
target = XLogInitBufferForRedo(record, 0); |
| 852 |
page = (Page) BufferGetPage(target); |
| 853 |
|
| 854 |
_bt_pageinit(page, BufferGetPageSize(target)); |
| 855 |
pageop = BTPageGetOpaque(page); |
| 856 |
|
| 857 |
pageop->btpo_prev = leftsib; |
| 858 |
pageop->btpo_next = rightsib; |
| 859 |
pageop->btpo_level = level; |
| 860 |
BTPageSetDeleted(page, safexid); |
| 861 |
if (isleaf) |
| 862 |
pageop->btpo_flags |= BTP_LEAF; |
| 863 |
pageop->btpo_cycleid = 0; |
| 864 |
|
| 865 |
PageSetLSN(page, lsn); |
| 866 |
MarkBufferDirty(target); |
| 867 |
|
| 868 |
/* Fix left-link of right sibling */ |
| 869 |
if (XLogReadBufferForRedo(record, 2, &rightbuf) == BLK_NEEDS_REDO) |
| 870 |
{ |
| 871 |
page = (Page) BufferGetPage(rightbuf); |
| 872 |
pageop = BTPageGetOpaque(page); |
| 873 |
pageop->btpo_prev = leftsib; |
| 874 |
|
| 875 |
PageSetLSN(page, lsn); |
| 876 |
MarkBufferDirty(rightbuf); |
| 877 |
} |
| 878 |
|
| 879 |
/* Release siblings */ |
| 880 |
if (BufferIsValid(leftbuf)) |
| 881 |
UnlockReleaseBuffer(leftbuf); |
| 882 |
if (BufferIsValid(rightbuf)) |
| 883 |
UnlockReleaseBuffer(rightbuf); |
| 884 |
|
| 885 |
/* Release target */ |
| 886 |
UnlockReleaseBuffer(target); |
| 887 |
|
| 888 |
/* |
| 889 |
* If we deleted a parent of the targeted leaf page, instead of the leaf |
| 890 |
* itself, update the leaf to point to the next remaining child in the |
| 891 |
* to-be-deleted subtree |
| 892 |
*/ |
| 893 |
if (XLogRecHasBlockRef(record, 3)) |
| 894 |
{ |
| 895 |
/* |
| 896 |
* There is no real data on the page, so we just re-create it from |
| 897 |
* scratch using the information from the WAL record. |
| 898 |
* |
| 899 |
* Note that we don't end up here when the target page is also the |
| 900 |
* leafbuf page. There is no need to add a dummy hikey item with a |
| 901 |
* top parent link when deleting leafbuf because it's the last page |
| 902 |
* we'll delete in the subtree undergoing deletion. |
| 903 |
*/ |
| 904 |
Buffer leafbuf; |
| 905 |
IndexTupleData trunctuple; |
| 906 |
|
| 907 |
Assert(!isleaf); |
| 908 |
|
| 909 |
leafbuf = XLogInitBufferForRedo(record, 3); |
| 910 |
page = (Page) BufferGetPage(leafbuf); |
| 911 |
|
| 912 |
_bt_pageinit(page, BufferGetPageSize(leafbuf)); |
| 913 |
pageop = BTPageGetOpaque(page); |
| 914 |
|
| 915 |
pageop->btpo_flags = BTP_HALF_DEAD | BTP_LEAF; |
| 916 |
pageop->btpo_prev = xlrec->leafleftsib; |
| 917 |
pageop->btpo_next = xlrec->leafrightsib; |
| 918 |
pageop->btpo_level = 0; |
| 919 |
pageop->btpo_cycleid = 0; |
| 920 |
|
| 921 |
/* Add a dummy hikey item */ |
| 922 |
MemSet(&trunctuple, 0, sizeof(IndexTupleData)); |
| 923 |
trunctuple.t_info = sizeof(IndexTupleData); |
| 924 |
BTreeTupleSetTopParent(&trunctuple, xlrec->leaftopparent); |
| 925 |
|
| 926 |
if (PageAddItem(page, (Item) &trunctuple, sizeof(IndexTupleData), P_HIKEY, |
| 927 |
false, false) == InvalidOffsetNumber) |
| 928 |
elog(ERROR, "could not add dummy high key to half-dead page"); |
| 929 |
|
| 930 |
PageSetLSN(page, lsn); |
| 931 |
MarkBufferDirty(leafbuf); |
| 932 |
UnlockReleaseBuffer(leafbuf); |
| 933 |
} |
| 934 |
|
| 935 |
/* Update metapage if needed */ |
| 936 |
if (info == XLOG_BTREE_UNLINK_PAGE_META) |
| 937 |
_bt_restore_meta(record, 4); |
| 938 |
} |
| 939 |
|
| 940 |
static void |
| 941 |
btree_xlog_newroot(XLogReaderState *record) |
| 942 |
{ |
| 943 |
XLogRecPtr lsn = record->EndRecPtr; |
| 944 |
xl_btree_newroot *xlrec = (xl_btree_newroot *) XLogRecGetData(record); |
| 945 |
Buffer buffer; |
| 946 |
Page page; |
| 947 |
BTPageOpaque pageop; |
| 948 |
char *ptr; |
| 949 |
Size len; |
| 950 |
|
| 951 |
buffer = XLogInitBufferForRedo(record, 0); |
| 952 |
page = (Page) BufferGetPage(buffer); |
| 953 |
|
| 954 |
_bt_pageinit(page, BufferGetPageSize(buffer)); |
| 955 |
pageop = BTPageGetOpaque(page); |
| 956 |
|
| 957 |
pageop->btpo_flags = BTP_ROOT; |
| 958 |
pageop->btpo_prev = pageop->btpo_next = P_NONE; |
| 959 |
pageop->btpo_level = xlrec->level; |
| 960 |
if (xlrec->level == 0) |
| 961 |
pageop->btpo_flags |= BTP_LEAF; |
| 962 |
pageop->btpo_cycleid = 0; |
| 963 |
|
| 964 |
if (xlrec->level > 0) |
| 965 |
{ |
| 966 |
ptr = XLogRecGetBlockData(record, 0, &len); |
| 967 |
_bt_restore_page(page, ptr, len); |
| 968 |
|
| 969 |
/* Clear the incomplete-split flag in left child */ |
| 970 |
_bt_clear_incomplete_split(record, 1); |
| 971 |
} |
| 972 |
|
| 973 |
PageSetLSN(page, lsn); |
| 974 |
MarkBufferDirty(buffer); |
| 975 |
UnlockReleaseBuffer(buffer); |
| 976 |
|
| 977 |
_bt_restore_meta(record, 2); |
| 978 |
} |
| 979 |
|
| 980 |
/* |
| 981 |
* In general VACUUM must defer recycling as a way of avoiding certain race |
| 982 |
* conditions. Deleted pages contain a safexid value that is used by VACUUM |
| 983 |
* to determine whether or not it's safe to place a page that was deleted by |
| 984 |
* VACUUM earlier into the FSM now. See nbtree/README. |
| 985 |
* |
| 986 |
* As far as any backend operating during original execution is concerned, the |
| 987 |
* FSM is a cache of recycle-safe pages; the mere presence of the page in the |
| 988 |
* FSM indicates that the page must already be safe to recycle (actually, |
| 989 |
* _bt_allocbuf() verifies it's safe using BTPageIsRecyclable(), but that's |
| 990 |
* just because it would be unwise to completely trust the FSM, given its |
| 991 |
* current limitations). |
| 992 |
* |
| 993 |
* This isn't sufficient to prevent similar concurrent recycling race |
| 994 |
* conditions during Hot Standby, though. For that we need to log a |
| 995 |
* xl_btree_reuse_page record at the point that a page is actually recycled |
| 996 |
* and reused for an entirely unrelated page inside _bt_split(). These |
| 997 |
* records include the same safexid value from the original deleted page, |
| 998 |
* stored in the record's snapshotConflictHorizon field. |
| 999 |
* |
| 1000 |
* The GlobalVisCheckRemovableFullXid() test in BTPageIsRecyclable() is used |
| 1001 |
* to determine if it's safe to recycle a page. This mirrors our own test: |
| 1002 |
* the PGPROC->xmin > limitXmin test inside GetConflictingVirtualXIDs(). |
| 1003 |
* Consequently, one XID value achieves the same exclusion effect on primary |
| 1004 |
* and standby. |
| 1005 |
*/ |
| 1006 |
static void |
| 1007 |
btree_xlog_reuse_page(XLogReaderState *record) |
| 1008 |
{ |
| 1009 |
xl_btree_reuse_page *xlrec = (xl_btree_reuse_page *) XLogRecGetData(record); |
| 1010 |
|
| 1011 |
if (InHotStandby) |
| 1012 |
ResolveRecoveryConflictWithSnapshotFullXid(xlrec->snapshotConflictHorizon, |
| 1013 |
xlrec->isCatalogRel, |
| 1014 |
xlrec->locator); |
| 1015 |
} |
| 1016 |
|
| 1017 |
void |
| 1018 |
btree_redo(XLogReaderState *record) |
| 1019 |
{ |
| 1020 |
uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK; |
| 1021 |
MemoryContext oldCtx; |
| 1022 |
|
| 1023 |
oldCtx = MemoryContextSwitchTo(opCtx); |
| 1024 |
switch (info) |
| 1025 |
{ |
| 1026 |
case XLOG_BTREE_INSERT_LEAF: |
| 1027 |
btree_xlog_insert(true, false, false, record); |
| 1028 |
break; |
| 1029 |
case XLOG_BTREE_INSERT_UPPER: |
| 1030 |
btree_xlog_insert(false, false, false, record); |
| 1031 |
break; |
| 1032 |
case XLOG_BTREE_INSERT_META: |
| 1033 |
btree_xlog_insert(false, true, false, record); |
| 1034 |
break; |
| 1035 |
case XLOG_BTREE_SPLIT_L: |
| 1036 |
btree_xlog_split(true, record); |
| 1037 |
break; |
| 1038 |
case XLOG_BTREE_SPLIT_R: |
| 1039 |
btree_xlog_split(false, record); |
| 1040 |
break; |
| 1041 |
case XLOG_BTREE_INSERT_POST: |
| 1042 |
btree_xlog_insert(true, false, true, record); |
| 1043 |
break; |
| 1044 |
case XLOG_BTREE_DEDUP: |
| 1045 |
btree_xlog_dedup(record); |
| 1046 |
break; |
| 1047 |
case XLOG_BTREE_VACUUM: |
| 1048 |
btree_xlog_vacuum(record); |
| 1049 |
break; |
| 1050 |
case XLOG_BTREE_DELETE: |
| 1051 |
btree_xlog_delete(record); |
| 1052 |
break; |
| 1053 |
case XLOG_BTREE_MARK_PAGE_HALFDEAD: |
| 1054 |
btree_xlog_mark_page_halfdead(info, record); |
| 1055 |
break; |
| 1056 |
case XLOG_BTREE_UNLINK_PAGE: |
| 1057 |
case XLOG_BTREE_UNLINK_PAGE_META: |
| 1058 |
btree_xlog_unlink_page(info, record); |
| 1059 |
break; |
| 1060 |
case XLOG_BTREE_NEWROOT: |
| 1061 |
btree_xlog_newroot(record); |
| 1062 |
break; |
| 1063 |
case XLOG_BTREE_REUSE_PAGE: |
| 1064 |
btree_xlog_reuse_page(record); |
| 1065 |
break; |
| 1066 |
case XLOG_BTREE_META_CLEANUP: |
| 1067 |
_bt_restore_meta(record, 0); |
| 1068 |
break; |
| 1069 |
default: |
| 1070 |
elog(PANIC, "btree_redo: unknown op code %u", info); |
| 1071 |
} |
| 1072 |
MemoryContextSwitchTo(oldCtx); |
| 1073 |
MemoryContextReset(opCtx); |
| 1074 |
} |
| 1075 |
|
| 1076 |
void |
| 1077 |
btree_xlog_startup(void) |
| 1078 |
{ |
| 1079 |
opCtx = AllocSetContextCreate(CurrentMemoryContext, |
| 1080 |
"Btree recovery temporary context", |
| 1081 |
ALLOCSET_DEFAULT_SIZES); |
| 1082 |
} |
| 1083 |
|
| 1084 |
void |
| 1085 |
btree_xlog_cleanup(void) |
| 1086 |
{ |
| 1087 |
MemoryContextDelete(opCtx); |
| 1088 |
opCtx = NULL; |
| 1089 |
} |
| 1090 |
|
| 1091 |
/* |
| 1092 |
* Mask a btree page before performing consistency checks on it. |
| 1093 |
*/ |
| 1094 |
void |
| 1095 |
btree_mask(char *pagedata, BlockNumber blkno) |
| 1096 |
{ |
| 1097 |
Page page = (Page) pagedata; |
| 1098 |
BTPageOpaque maskopaq; |
| 1099 |
|
| 1100 |
mask_page_lsn_and_checksum(page); |
| 1101 |
|
| 1102 |
mask_page_hint_bits(page); |
| 1103 |
mask_unused_space(page); |
| 1104 |
|
| 1105 |
maskopaq = BTPageGetOpaque(page); |
| 1106 |
|
| 1107 |
if (P_ISLEAF(maskopaq)) |
| 1108 |
{ |
| 1109 |
/* |
| 1110 |
* In btree leaf pages, it is possible to modify the LP_FLAGS without |
| 1111 |
* emitting any WAL record. Hence, mask the line pointer flags. See |
| 1112 |
* _bt_killitems(), _bt_check_unique() for details. |
| 1113 |
*/ |
| 1114 |
mask_lp_flags(page); |
| 1115 |
} |
| 1116 |
|
| 1117 |
/* |
| 1118 |
* BTP_HAS_GARBAGE is just an un-logged hint bit. So, mask it. See |
| 1119 |
* _bt_delete_or_dedup_one_page(), _bt_killitems(), and _bt_check_unique() |
| 1120 |
* for details. |
| 1121 |
*/ |
| 1122 |
maskopaq->btpo_flags &= ~BTP_HAS_GARBAGE; |
| 1123 |
|
| 1124 |
/* |
| 1125 |
* During replay of a btree page split, we don't set the BTP_SPLIT_END |
| 1126 |
* flag of the right sibling and initialize the cycle_id to 0 for the same |
| 1127 |
* page. See btree_xlog_split() for details. |
| 1128 |
*/ |
| 1129 |
maskopaq->btpo_flags &= ~BTP_SPLIT_END; |
| 1130 |
maskopaq->btpo_cycleid = 0; |
| 1131 |
} |