PostgreSQL Source Code  git master
freespace.c
Go to the documentation of this file.
1 /*-------------------------------------------------------------------------
2  *
3  * freespace.c
4  * POSTGRES free space map for quickly finding free space in relations
5  *
6  *
7  * Portions Copyright (c) 1996-2021, PostgreSQL Global Development Group
8  * Portions Copyright (c) 1994, Regents of the University of California
9  *
10  * IDENTIFICATION
11  * src/backend/storage/freespace/freespace.c
12  *
13  *
14  * NOTES:
15  *
16  * Free Space Map keeps track of the amount of free space on pages, and
17  * allows quickly searching for a page with enough free space. The FSM is
18  * stored in a dedicated relation fork of all heap relations, and those
19  * index access methods that need it (see also indexfsm.c). See README for
20  * more information.
21  *
22  *-------------------------------------------------------------------------
23  */
24 #include "postgres.h"
25 
26 #include "access/htup_details.h"
27 #include "access/xlogutils.h"
28 #include "miscadmin.h"
29 #include "storage/freespace.h"
30 #include "storage/fsm_internals.h"
31 #include "storage/lmgr.h"
32 #include "storage/smgr.h"
33 
34 
35 /*
36  * We use just one byte to store the amount of free space on a page, so we
37  * divide the amount of free space a page can have into 256 different
38  * categories. The highest category, 255, represents a page with at least
39  * MaxFSMRequestSize bytes of free space, and the second highest category
40  * represents the range from 254 * FSM_CAT_STEP, inclusive, to
41  * MaxFSMRequestSize, exclusive.
42  *
43  * MaxFSMRequestSize depends on the architecture and BLCKSZ, but assuming
44  * default 8k BLCKSZ, and that MaxFSMRequestSize is 8164 bytes, the
45  * categories look like this:
46  *
47  *
48  * Range Category
49  * 0 - 31 0
50  * 32 - 63 1
51  * ... ... ...
52  * 8096 - 8127 253
53  * 8128 - 8163 254
54  * 8164 - 8192 255
55  *
56  * The reason that MaxFSMRequestSize is special is that if MaxFSMRequestSize
57  * isn't equal to a range boundary, a page with exactly MaxFSMRequestSize
58  * bytes of free space wouldn't satisfy a request for MaxFSMRequestSize
59  * bytes. If there isn't more than MaxFSMRequestSize bytes of free space on a
60  * completely empty page, that would mean that we could never satisfy a
61  * request of exactly MaxFSMRequestSize bytes.
62  */
63 #define FSM_CATEGORIES 256
64 #define FSM_CAT_STEP (BLCKSZ / FSM_CATEGORIES)
65 #define MaxFSMRequestSize MaxHeapTupleSize
66 
67 /*
68  * Depth of the on-disk tree. We need to be able to address 2^32-1 blocks,
69  * and 1626 is the smallest number that satisfies X^3 >= 2^32-1. Likewise,
70  * 216 is the smallest number that satisfies X^4 >= 2^32-1. In practice,
71  * this means that 4096 bytes is the smallest BLCKSZ that we can get away
72  * with a 3-level tree, and 512 is the smallest we support.
73  */
74 #define FSM_TREE_DEPTH ((SlotsPerFSMPage >= 1626) ? 3 : 4)
75 
76 #define FSM_ROOT_LEVEL (FSM_TREE_DEPTH - 1)
77 #define FSM_BOTTOM_LEVEL 0
78 
79 /*
80  * The internal FSM routines work on a logical addressing scheme. Each
81  * level of the tree can be thought of as a separately addressable file.
82  */
83 typedef struct
84 {
85  int level; /* level */
86  int logpageno; /* page number within the level */
87 } FSMAddress;
88 
89 /* Address of the root page. */
91 
92 /* functions to navigate the tree */
93 static FSMAddress fsm_get_child(FSMAddress parent, uint16 slot);
94 static FSMAddress fsm_get_parent(FSMAddress child, uint16 *slot);
95 static FSMAddress fsm_get_location(BlockNumber heapblk, uint16 *slot);
98 
99 static Buffer fsm_readbuf(Relation rel, FSMAddress addr, bool extend);
100 static void fsm_extend(Relation rel, BlockNumber fsm_nblocks);
101 
102 /* functions to convert amount of free space to a FSM category */
103 static uint8 fsm_space_avail_to_cat(Size avail);
104 static uint8 fsm_space_needed_to_cat(Size needed);
105 static Size fsm_space_cat_to_avail(uint8 cat);
106 
107 /* workhorse functions for various operations */
108 static int fsm_set_and_search(Relation rel, FSMAddress addr, uint16 slot,
109  uint8 newValue, uint8 minValue);
110 static BlockNumber fsm_search(Relation rel, uint8 min_cat);
111 static uint8 fsm_vacuum_page(Relation rel, FSMAddress addr,
112  BlockNumber start, BlockNumber end,
113  bool *eof);
114 
115 
116 /******** Public API ********/
117 
118 /*
119  * GetPageWithFreeSpace - try to find a page in the given relation with
120  * at least the specified amount of free space.
121  *
122  * If successful, return the block number; if not, return InvalidBlockNumber.
123  *
124  * The caller must be prepared for the possibility that the returned page
125  * will turn out to have too little space available by the time the caller
126  * gets a lock on it. In that case, the caller should report the actual
127  * amount of free space available on that page and then try again (see
128  * RecordAndGetPageWithFreeSpace). If InvalidBlockNumber is returned,
129  * extend the relation.
130  */
133 {
134  uint8 min_cat = fsm_space_needed_to_cat(spaceNeeded);
135 
136  return fsm_search(rel, min_cat);
137 }
138 
139 /*
140  * RecordAndGetPageWithFreeSpace - update info about a page and try again.
141  *
142  * We provide this combo form to save some locking overhead, compared to
143  * separate RecordPageWithFreeSpace + GetPageWithFreeSpace calls. There's
144  * also some effort to return a page close to the old page; if there's a
145  * page with enough free space on the same FSM page where the old one page
146  * is located, it is preferred.
147  */
150  Size oldSpaceAvail, Size spaceNeeded)
151 {
152  int old_cat = fsm_space_avail_to_cat(oldSpaceAvail);
153  int search_cat = fsm_space_needed_to_cat(spaceNeeded);
154  FSMAddress addr;
155  uint16 slot;
156  int search_slot;
157 
158  /* Get the location of the FSM byte representing the heap block */
159  addr = fsm_get_location(oldPage, &slot);
160 
161  search_slot = fsm_set_and_search(rel, addr, slot, old_cat, search_cat);
162 
163  /*
164  * If fsm_set_and_search found a suitable new block, return that.
165  * Otherwise, search as usual.
166  */
167  if (search_slot != -1)
168  return fsm_get_heap_blk(addr, search_slot);
169  else
170  return fsm_search(rel, search_cat);
171 }
172 
173 /*
174  * RecordPageWithFreeSpace - update info about a page.
175  *
176  * Note that if the new spaceAvail value is higher than the old value stored
177  * in the FSM, the space might not become visible to searchers until the next
178  * FreeSpaceMapVacuum call, which updates the upper level pages.
179  */
180 void
182 {
183  int new_cat = fsm_space_avail_to_cat(spaceAvail);
184  FSMAddress addr;
185  uint16 slot;
186 
187  /* Get the location of the FSM byte representing the heap block */
188  addr = fsm_get_location(heapBlk, &slot);
189 
190  fsm_set_and_search(rel, addr, slot, new_cat, 0);
191 }
192 
193 /*
194  * XLogRecordPageWithFreeSpace - like RecordPageWithFreeSpace, for use in
195  * WAL replay
196  */
197 void
199  Size spaceAvail)
200 {
201  int new_cat = fsm_space_avail_to_cat(spaceAvail);
202  FSMAddress addr;
203  uint16 slot;
204  BlockNumber blkno;
205  Buffer buf;
206  Page page;
207 
208  /* Get the location of the FSM byte representing the heap block */
209  addr = fsm_get_location(heapBlk, &slot);
210  blkno = fsm_logical_to_physical(addr);
211 
212  /* If the page doesn't exist already, extend */
214  InvalidBuffer);
216 
217  page = BufferGetPage(buf);
218  if (PageIsNew(page))
219  PageInit(page, BLCKSZ, 0);
220 
221  if (fsm_set_avail(page, slot, new_cat))
222  MarkBufferDirtyHint(buf, false);
223  UnlockReleaseBuffer(buf);
224 }
225 
226 /*
227  * GetRecordedFreeSpace - return the amount of free space on a particular page,
228  * according to the FSM.
229  */
230 Size
232 {
233  FSMAddress addr;
234  uint16 slot;
235  Buffer buf;
236  uint8 cat;
237 
238  /* Get the location of the FSM byte representing the heap block */
239  addr = fsm_get_location(heapBlk, &slot);
240 
241  buf = fsm_readbuf(rel, addr, false);
242  if (!BufferIsValid(buf))
243  return 0;
244  cat = fsm_get_avail(BufferGetPage(buf), slot);
245  ReleaseBuffer(buf);
246 
247  return fsm_space_cat_to_avail(cat);
248 }
249 
250 /*
251  * FreeSpaceMapPrepareTruncateRel - prepare for truncation of a relation.
252  *
253  * nblocks is the new size of the heap.
254  *
255  * Return the number of blocks of new FSM.
256  * If it's InvalidBlockNumber, there is nothing to truncate;
257  * otherwise the caller is responsible for calling smgrtruncate()
258  * to truncate the FSM pages, and FreeSpaceMapVacuumRange()
259  * to update upper-level pages in the FSM.
260  */
263 {
264  BlockNumber new_nfsmblocks;
265  FSMAddress first_removed_address;
266  uint16 first_removed_slot;
267  Buffer buf;
268 
269  RelationOpenSmgr(rel);
270 
271  /*
272  * If no FSM has been created yet for this relation, there's nothing to
273  * truncate.
274  */
275  if (!smgrexists(rel->rd_smgr, FSM_FORKNUM))
276  return InvalidBlockNumber;
277 
278  /* Get the location in the FSM of the first removed heap block */
279  first_removed_address = fsm_get_location(nblocks, &first_removed_slot);
280 
281  /*
282  * Zero out the tail of the last remaining FSM page. If the slot
283  * representing the first removed heap block is at a page boundary, as the
284  * first slot on the FSM page that first_removed_address points to, we can
285  * just truncate that page altogether.
286  */
287  if (first_removed_slot > 0)
288  {
289  buf = fsm_readbuf(rel, first_removed_address, false);
290  if (!BufferIsValid(buf))
291  return InvalidBlockNumber; /* nothing to do; the FSM was already
292  * smaller */
294 
295  /* NO EREPORT(ERROR) from here till changes are logged */
297 
298  fsm_truncate_avail(BufferGetPage(buf), first_removed_slot);
299 
300  /*
301  * Truncation of a relation is WAL-logged at a higher-level, and we
302  * will be called at WAL replay. But if checksums are enabled, we need
303  * to still write a WAL record to protect against a torn page, if the
304  * page is flushed to disk before the truncation WAL record. We cannot
305  * use MarkBufferDirtyHint here, because that will not dirty the page
306  * during recovery.
307  */
308  MarkBufferDirty(buf);
310  log_newpage_buffer(buf, false);
311 
313 
314  UnlockReleaseBuffer(buf);
315 
316  new_nfsmblocks = fsm_logical_to_physical(first_removed_address) + 1;
317  }
318  else
319  {
320  new_nfsmblocks = fsm_logical_to_physical(first_removed_address);
321  if (smgrnblocks(rel->rd_smgr, FSM_FORKNUM) <= new_nfsmblocks)
322  return InvalidBlockNumber; /* nothing to do; the FSM was already
323  * smaller */
324  }
325 
326  return new_nfsmblocks;
327 }
328 
329 /*
330  * FreeSpaceMapVacuum - update upper-level pages in the rel's FSM
331  *
332  * We assume that the bottom-level pages have already been updated with
333  * new free-space information.
334  */
335 void
337 {
338  bool dummy;
339 
340  /* Recursively scan the tree, starting at the root */
341  (void) fsm_vacuum_page(rel, FSM_ROOT_ADDRESS,
343  &dummy);
344 }
345 
346 /*
347  * FreeSpaceMapVacuumRange - update upper-level pages in the rel's FSM
348  *
349  * As above, but assume that only heap pages between start and end-1 inclusive
350  * have new free-space information, so update only the upper-level slots
351  * covering that block range. end == InvalidBlockNumber is equivalent to
352  * "all the rest of the relation".
353  */
354 void
356 {
357  bool dummy;
358 
359  /* Recursively scan the tree, starting at the root */
360  if (end > start)
361  (void) fsm_vacuum_page(rel, FSM_ROOT_ADDRESS, start, end, &dummy);
362 }
363 
364 /******** Internal routines ********/
365 
366 /*
367  * Return category corresponding x bytes of free space
368  */
369 static uint8
371 {
372  int cat;
373 
374  Assert(avail < BLCKSZ);
375 
376  if (avail >= MaxFSMRequestSize)
377  return 255;
378 
379  cat = avail / FSM_CAT_STEP;
380 
381  /*
382  * The highest category, 255, is reserved for MaxFSMRequestSize bytes or
383  * more.
384  */
385  if (cat > 254)
386  cat = 254;
387 
388  return (uint8) cat;
389 }
390 
391 /*
392  * Return the lower bound of the range of free space represented by given
393  * category.
394  */
395 static Size
397 {
398  /* The highest category represents exactly MaxFSMRequestSize bytes. */
399  if (cat == 255)
400  return MaxFSMRequestSize;
401  else
402  return cat * FSM_CAT_STEP;
403 }
404 
405 /*
406  * Which category does a page need to have, to accommodate x bytes of data?
407  * While fsm_space_avail_to_cat() rounds down, this needs to round up.
408  */
409 static uint8
411 {
412  int cat;
413 
414  /* Can't ask for more space than the highest category represents */
415  if (needed > MaxFSMRequestSize)
416  elog(ERROR, "invalid FSM request size %zu", needed);
417 
418  if (needed == 0)
419  return 1;
420 
421  cat = (needed + FSM_CAT_STEP - 1) / FSM_CAT_STEP;
422 
423  if (cat > 255)
424  cat = 255;
425 
426  return (uint8) cat;
427 }
428 
429 /*
430  * Returns the physical block number of a FSM page
431  */
432 static BlockNumber
434 {
435  BlockNumber pages;
436  int leafno;
437  int l;
438 
439  /*
440  * Calculate the logical page number of the first leaf page below the
441  * given page.
442  */
443  leafno = addr.logpageno;
444  for (l = 0; l < addr.level; l++)
445  leafno *= SlotsPerFSMPage;
446 
447  /* Count upper level nodes required to address the leaf page */
448  pages = 0;
449  for (l = 0; l < FSM_TREE_DEPTH; l++)
450  {
451  pages += leafno + 1;
452  leafno /= SlotsPerFSMPage;
453  }
454 
455  /*
456  * If the page we were asked for wasn't at the bottom level, subtract the
457  * additional lower level pages we counted above.
458  */
459  pages -= addr.level;
460 
461  /* Turn the page count into 0-based block number */
462  return pages - 1;
463 }
464 
465 /*
466  * Return the FSM location corresponding to given heap block.
467  */
468 static FSMAddress
470 {
471  FSMAddress addr;
472 
473  addr.level = FSM_BOTTOM_LEVEL;
474  addr.logpageno = heapblk / SlotsPerFSMPage;
475  *slot = heapblk % SlotsPerFSMPage;
476 
477  return addr;
478 }
479 
480 /*
481  * Return the heap block number corresponding to given location in the FSM.
482  */
483 static BlockNumber
485 {
486  Assert(addr.level == FSM_BOTTOM_LEVEL);
487  return ((unsigned int) addr.logpageno) * SlotsPerFSMPage + slot;
488 }
489 
490 /*
491  * Given a logical address of a child page, get the logical page number of
492  * the parent, and the slot within the parent corresponding to the child.
493  */
494 static FSMAddress
496 {
497  FSMAddress parent;
498 
499  Assert(child.level < FSM_ROOT_LEVEL);
500 
501  parent.level = child.level + 1;
502  parent.logpageno = child.logpageno / SlotsPerFSMPage;
503  *slot = child.logpageno % SlotsPerFSMPage;
504 
505  return parent;
506 }
507 
508 /*
509  * Given a logical address of a parent page and a slot number, get the
510  * logical address of the corresponding child page.
511  */
512 static FSMAddress
514 {
515  FSMAddress child;
516 
517  Assert(parent.level > FSM_BOTTOM_LEVEL);
518 
519  child.level = parent.level - 1;
520  child.logpageno = parent.logpageno * SlotsPerFSMPage + slot;
521 
522  return child;
523 }
524 
525 /*
526  * Read a FSM page.
527  *
528  * If the page doesn't exist, InvalidBuffer is returned, or if 'extend' is
529  * true, the FSM file is extended.
530  */
531 static Buffer
532 fsm_readbuf(Relation rel, FSMAddress addr, bool extend)
533 {
534  BlockNumber blkno = fsm_logical_to_physical(addr);
535  Buffer buf;
536 
537  RelationOpenSmgr(rel);
538 
539  /*
540  * If we haven't cached the size of the FSM yet, check it first. Also
541  * recheck if the requested block seems to be past end, since our cached
542  * value might be stale. (We send smgr inval messages on truncation, but
543  * not on extension.)
544  */
546  blkno >= rel->rd_smgr->smgr_cached_nblocks[FSM_FORKNUM])
547  {
548  /* Invalidate the cache so smgrnblocks asks the kernel. */
550  if (smgrexists(rel->rd_smgr, FSM_FORKNUM))
552  else
554  }
555 
556  /* Handle requests beyond EOF */
557  if (blkno >= rel->rd_smgr->smgr_cached_nblocks[FSM_FORKNUM])
558  {
559  if (extend)
560  fsm_extend(rel, blkno + 1);
561  else
562  return InvalidBuffer;
563  }
564 
565  /*
566  * Use ZERO_ON_ERROR mode, and initialize the page if necessary. The FSM
567  * information is not accurate anyway, so it's better to clear corrupt
568  * pages than error out. Since the FSM changes are not WAL-logged, the
569  * so-called torn page problem on crash can lead to pages with corrupt
570  * headers, for example.
571  *
572  * The initialize-the-page part is trickier than it looks, because of the
573  * possibility of multiple backends doing this concurrently, and our
574  * desire to not uselessly take the buffer lock in the normal path where
575  * the page is OK. We must take the lock to initialize the page, so
576  * recheck page newness after we have the lock, in case someone else
577  * already did it. Also, because we initially check PageIsNew with no
578  * lock, it's possible to fall through and return the buffer while someone
579  * else is still initializing the page (i.e., we might see pd_upper as set
580  * but other page header fields are still zeroes). This is harmless for
581  * callers that will take a buffer lock themselves, but some callers
582  * inspect the page without any lock at all. The latter is OK only so
583  * long as it doesn't depend on the page header having correct contents.
584  * Current usage is safe because PageGetContents() does not require that.
585  */
586  buf = ReadBufferExtended(rel, FSM_FORKNUM, blkno, RBM_ZERO_ON_ERROR, NULL);
587  if (PageIsNew(BufferGetPage(buf)))
588  {
590  if (PageIsNew(BufferGetPage(buf)))
591  PageInit(BufferGetPage(buf), BLCKSZ, 0);
593  }
594  return buf;
595 }
596 
597 /*
598  * Ensure that the FSM fork is at least fsm_nblocks long, extending
599  * it if necessary with empty pages. And by empty, I mean pages filled
600  * with zeros, meaning there's no free space.
601  */
602 static void
603 fsm_extend(Relation rel, BlockNumber fsm_nblocks)
604 {
605  BlockNumber fsm_nblocks_now;
606  PGAlignedBlock pg;
607 
608  PageInit((Page) pg.data, BLCKSZ, 0);
609 
610  /*
611  * We use the relation extension lock to lock out other backends trying to
612  * extend the FSM at the same time. It also locks out extension of the
613  * main fork, unnecessarily, but extending the FSM happens seldom enough
614  * that it doesn't seem worthwhile to have a separate lock tag type for
615  * it.
616  *
617  * Note that another backend might have extended or created the relation
618  * by the time we get the lock.
619  */
621 
622  /* Might have to re-open if a cache flush happened */
623  RelationOpenSmgr(rel);
624 
625  /*
626  * Create the FSM file first if it doesn't exist. If
627  * smgr_cached_nblocks[FSM_FORKNUM] is positive then it must exist, no
628  * need for an smgrexists call.
629  */
630  if ((rel->rd_smgr->smgr_cached_nblocks[FSM_FORKNUM] == 0 ||
633  smgrcreate(rel->rd_smgr, FSM_FORKNUM, false);
634 
635  /* Invalidate cache so that smgrnblocks() asks the kernel. */
637  fsm_nblocks_now = smgrnblocks(rel->rd_smgr, FSM_FORKNUM);
638 
639  while (fsm_nblocks_now < fsm_nblocks)
640  {
641  PageSetChecksumInplace((Page) pg.data, fsm_nblocks_now);
642 
643  smgrextend(rel->rd_smgr, FSM_FORKNUM, fsm_nblocks_now,
644  pg.data, false);
645  fsm_nblocks_now++;
646  }
647 
649 }
650 
651 /*
652  * Set value in given FSM page and slot.
653  *
654  * If minValue > 0, the updated page is also searched for a page with at
655  * least minValue of free space. If one is found, its slot number is
656  * returned, -1 otherwise.
657  */
658 static int
660  uint8 newValue, uint8 minValue)
661 {
662  Buffer buf;
663  Page page;
664  int newslot = -1;
665 
666  buf = fsm_readbuf(rel, addr, true);
668 
669  page = BufferGetPage(buf);
670 
671  if (fsm_set_avail(page, slot, newValue))
672  MarkBufferDirtyHint(buf, false);
673 
674  if (minValue != 0)
675  {
676  /* Search while we still hold the lock */
677  newslot = fsm_search_avail(buf, minValue,
678  addr.level == FSM_BOTTOM_LEVEL,
679  true);
680  }
681 
682  UnlockReleaseBuffer(buf);
683 
684  return newslot;
685 }
686 
687 /*
688  * Search the tree for a heap page with at least min_cat of free space
689  */
690 static BlockNumber
691 fsm_search(Relation rel, uint8 min_cat)
692 {
693  int restarts = 0;
695 
696  for (;;)
697  {
698  int slot;
699  Buffer buf;
700  uint8 max_avail = 0;
701 
702  /* Read the FSM page. */
703  buf = fsm_readbuf(rel, addr, false);
704 
705  /* Search within the page */
706  if (BufferIsValid(buf))
707  {
709  slot = fsm_search_avail(buf, min_cat,
710  (addr.level == FSM_BOTTOM_LEVEL),
711  false);
712  if (slot == -1)
713  max_avail = fsm_get_max_avail(BufferGetPage(buf));
714  UnlockReleaseBuffer(buf);
715  }
716  else
717  slot = -1;
718 
719  if (slot != -1)
720  {
721  /*
722  * Descend the tree, or return the found block if we're at the
723  * bottom.
724  */
725  if (addr.level == FSM_BOTTOM_LEVEL)
726  return fsm_get_heap_blk(addr, slot);
727 
728  addr = fsm_get_child(addr, slot);
729  }
730  else if (addr.level == FSM_ROOT_LEVEL)
731  {
732  /*
733  * At the root, failure means there's no page with enough free
734  * space in the FSM. Give up.
735  */
736  return InvalidBlockNumber;
737  }
738  else
739  {
740  uint16 parentslot;
741  FSMAddress parent;
742 
743  /*
744  * At lower level, failure can happen if the value in the upper-
745  * level node didn't reflect the value on the lower page. Update
746  * the upper node, to avoid falling into the same trap again, and
747  * start over.
748  *
749  * There's a race condition here, if another backend updates this
750  * page right after we release it, and gets the lock on the parent
751  * page before us. We'll then update the parent page with the now
752  * stale information we had. It's OK, because it should happen
753  * rarely, and will be fixed by the next vacuum.
754  */
755  parent = fsm_get_parent(addr, &parentslot);
756  fsm_set_and_search(rel, parent, parentslot, max_avail, 0);
757 
758  /*
759  * If the upper pages are badly out of date, we might need to loop
760  * quite a few times, updating them as we go. Any inconsistencies
761  * should eventually be corrected and the loop should end. Looping
762  * indefinitely is nevertheless scary, so provide an emergency
763  * valve.
764  */
765  if (restarts++ > 10000)
766  return InvalidBlockNumber;
767 
768  /* Start search all over from the root */
769  addr = FSM_ROOT_ADDRESS;
770  }
771  }
772 }
773 
774 
775 /*
776  * Recursive guts of FreeSpaceMapVacuum
777  *
778  * Examine the FSM page indicated by addr, as well as its children, updating
779  * upper-level nodes that cover the heap block range from start to end-1.
780  * (It's okay if end is beyond the actual end of the map.)
781  * Return the maximum freespace value on this page.
782  *
783  * If addr is past the end of the FSM, set *eof_p to true and return 0.
784  *
785  * This traverses the tree in depth-first order. The tree is stored
786  * physically in depth-first order, so this should be pretty I/O efficient.
787  */
788 static uint8
790  BlockNumber start, BlockNumber end,
791  bool *eof_p)
792 {
793  Buffer buf;
794  Page page;
795  uint8 max_avail;
796 
797  /* Read the page if it exists, or return EOF */
798  buf = fsm_readbuf(rel, addr, false);
799  if (!BufferIsValid(buf))
800  {
801  *eof_p = true;
802  return 0;
803  }
804  else
805  *eof_p = false;
806 
807  page = BufferGetPage(buf);
808 
809  /*
810  * If we're above the bottom level, recurse into children, and fix the
811  * information stored about them at this level.
812  */
813  if (addr.level > FSM_BOTTOM_LEVEL)
814  {
815  FSMAddress fsm_start,
816  fsm_end;
817  uint16 fsm_start_slot,
818  fsm_end_slot;
819  int slot,
820  start_slot,
821  end_slot;
822  bool eof = false;
823 
824  /*
825  * Compute the range of slots we need to update on this page, given
826  * the requested range of heap blocks to consider. The first slot to
827  * update is the one covering the "start" block, and the last slot is
828  * the one covering "end - 1". (Some of this work will be duplicated
829  * in each recursive call, but it's cheap enough to not worry about.)
830  */
831  fsm_start = fsm_get_location(start, &fsm_start_slot);
832  fsm_end = fsm_get_location(end - 1, &fsm_end_slot);
833 
834  while (fsm_start.level < addr.level)
835  {
836  fsm_start = fsm_get_parent(fsm_start, &fsm_start_slot);
837  fsm_end = fsm_get_parent(fsm_end, &fsm_end_slot);
838  }
839  Assert(fsm_start.level == addr.level);
840 
841  if (fsm_start.logpageno == addr.logpageno)
842  start_slot = fsm_start_slot;
843  else if (fsm_start.logpageno > addr.logpageno)
844  start_slot = SlotsPerFSMPage; /* shouldn't get here... */
845  else
846  start_slot = 0;
847 
848  if (fsm_end.logpageno == addr.logpageno)
849  end_slot = fsm_end_slot;
850  else if (fsm_end.logpageno > addr.logpageno)
851  end_slot = SlotsPerFSMPage - 1;
852  else
853  end_slot = -1; /* shouldn't get here... */
854 
855  for (slot = start_slot; slot <= end_slot; slot++)
856  {
857  int child_avail;
858 
860 
861  /* After we hit end-of-file, just clear the rest of the slots */
862  if (!eof)
863  child_avail = fsm_vacuum_page(rel, fsm_get_child(addr, slot),
864  start, end,
865  &eof);
866  else
867  child_avail = 0;
868 
869  /* Update information about the child */
870  if (fsm_get_avail(page, slot) != child_avail)
871  {
873  fsm_set_avail(page, slot, child_avail);
874  MarkBufferDirtyHint(buf, false);
876  }
877  }
878  }
879 
880  /* Now get the maximum value on the page, to return to caller */
881  max_avail = fsm_get_max_avail(page);
882 
883  /*
884  * Reset the next slot pointer. This encourages the use of low-numbered
885  * pages, increasing the chances that a later vacuum can truncate the
886  * relation. We don't bother with a lock here, nor with marking the page
887  * dirty if it wasn't already, since this is just a hint.
888  */
889  ((FSMPage) PageGetContents(page))->fp_next_slot = 0;
890 
891  ReleaseBuffer(buf);
892 
893  return max_avail;
894 }
#define SlotsPerFSMPage
Definition: fsm_internals.h:61
int logpageno
Definition: freespace.c:86
#define BUFFER_LOCK_UNLOCK
Definition: bufmgr.h:96
uint8 fsm_get_max_avail(Page page)
Definition: fsmpage.c:138
bool fsm_truncate_avail(Page page, int nslots)
Definition: fsmpage.c:313
XLogRecPtr log_newpage_buffer(Buffer buffer, bool page_std)
Definition: xloginsert.c:1090
void smgrcreate(SMgrRelation reln, ForkNumber forknum, bool isRedo)
Definition: smgr.c:333
#define FSM_BOTTOM_LEVEL
Definition: freespace.c:77
void RecordPageWithFreeSpace(Relation rel, BlockNumber heapBlk, Size spaceAvail)
Definition: freespace.c:181
void MarkBufferDirtyHint(Buffer buffer, bool buffer_std)
Definition: bufmgr.c:3854
static FSMAddress fsm_get_parent(FSMAddress child, uint16 *slot)
Definition: freespace.c:495
void MarkBufferDirty(Buffer buffer)
Definition: bufmgr.c:1562
#define ExclusiveLock
Definition: lockdefs.h:44
bool fsm_set_avail(Page page, int slot, uint8 value)
Definition: fsmpage.c:63
struct SMgrRelationData * rd_smgr
Definition: rel.h:57
Buffer ReadBufferExtended(Relation reln, ForkNumber forkNum, BlockNumber blockNum, ReadBufferMode mode, BufferAccessStrategy strategy)
Definition: bufmgr.c:744
bool InRecovery
Definition: xlog.c:209
#define END_CRIT_SECTION()
Definition: miscadmin.h:137
unsigned char uint8
Definition: c.h:439
#define InvalidBuffer
Definition: buf.h:25
#define START_CRIT_SECTION()
Definition: miscadmin.h:135
static BlockNumber fsm_get_heap_blk(FSMAddress addr, uint16 slot)
Definition: freespace.c:484
uint32 BlockNumber
Definition: block.h:31
void ReleaseBuffer(Buffer buffer)
Definition: bufmgr.c:3784
static uint8 fsm_vacuum_page(Relation rel, FSMAddress addr, BlockNumber start, BlockNumber end, bool *eof)
Definition: freespace.c:789
bool smgrexists(SMgrRelation reln, ForkNumber forknum)
Definition: smgr.c:247
#define BUFFER_LOCK_EXCLUSIVE
Definition: bufmgr.h:98
#define FSM_CAT_STEP
Definition: freespace.c:64
BlockNumber smgr_cached_nblocks[MAX_FORKNUM+1]
Definition: smgr.h:54
static FSMAddress fsm_get_child(FSMAddress parent, uint16 slot)
Definition: freespace.c:513
#define RelationOpenSmgr(relation)
Definition: rel.h:514
static const FSMAddress FSM_ROOT_ADDRESS
Definition: freespace.c:90
char data[BLCKSZ]
Definition: c.h:1141
unsigned short uint16
Definition: c.h:440
#define FSM_ROOT_LEVEL
Definition: freespace.c:76
static Size fsm_space_cat_to_avail(uint8 cat)
Definition: freespace.c:396
void UnlockReleaseBuffer(Buffer buffer)
Definition: bufmgr.c:3807
Buffer XLogReadBufferExtended(RelFileNode rnode, ForkNumber forknum, BlockNumber blkno, ReadBufferMode mode, Buffer recent_buffer)
Definition: xlogutils.c:443
#define ERROR
Definition: elog.h:46
int level
Definition: freespace.c:85
void XLogRecordPageWithFreeSpace(RelFileNode rnode, BlockNumber heapBlk, Size spaceAvail)
Definition: freespace.c:198
FSMPageData * FSMPage
Definition: fsm_internals.h:45
#define FSM_TREE_DEPTH
Definition: freespace.c:74
static char * buf
Definition: pg_test_fsync.c:68
BlockNumber FreeSpaceMapPrepareTruncateRel(Relation rel, BlockNumber nblocks)
Definition: freespace.c:262
static FSMAddress fsm_get_location(BlockNumber heapblk, uint16 *slot)
Definition: freespace.c:469
#define BufferGetPage(buffer)
Definition: bufmgr.h:169
void FreeSpaceMapVacuum(Relation rel)
Definition: freespace.c:336
static uint8 fsm_space_avail_to_cat(Size avail)
Definition: freespace.c:370
void LockRelationForExtension(Relation relation, LOCKMODE lockmode)
Definition: lmgr.c:403
Size GetRecordedFreeSpace(Relation rel, BlockNumber heapBlk)
Definition: freespace.c:231
void UnlockRelationForExtension(Relation relation, LOCKMODE lockmode)
Definition: lmgr.c:453
#define PageGetContents(page)
Definition: bufpage.h:246
void LockBuffer(Buffer buffer, int mode)
Definition: bufmgr.c:4023
static void fsm_extend(Relation rel, BlockNumber fsm_nblocks)
Definition: freespace.c:603
static uint8 fsm_space_needed_to_cat(Size needed)
Definition: freespace.c:410
static int fsm_set_and_search(Relation rel, FSMAddress addr, uint16 slot, uint8 newValue, uint8 minValue)
Definition: freespace.c:659
BlockNumber smgrnblocks(SMgrRelation reln, ForkNumber forknum)
Definition: smgr.c:548
#define Assert(condition)
Definition: c.h:804
size_t Size
Definition: c.h:540
#define InvalidBlockNumber
Definition: block.h:33
void PageSetChecksumInplace(Page page, BlockNumber blkno)
Definition: bufpage.c:1532
static BlockNumber fsm_logical_to_physical(FSMAddress addr)
Definition: freespace.c:433
#define BufferIsValid(bufnum)
Definition: bufmgr.h:123
uint8 fsm_get_avail(Page page, int slot)
Definition: fsmpage.c:122
#define RelationNeedsWAL(relation)
Definition: rel.h:570
void smgrextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, char *buffer, bool skipFsync)
Definition: smgr.c:462
#define PageIsNew(page)
Definition: bufpage.h:229
static Buffer fsm_readbuf(Relation rel, FSMAddress addr, bool extend)
Definition: freespace.c:532
#define elog(elevel,...)
Definition: elog.h:232
#define BUFFER_LOCK_SHARE
Definition: bufmgr.h:97
BlockNumber GetPageWithFreeSpace(Relation rel, Size spaceNeeded)
Definition: freespace.c:132
#define CHECK_FOR_INTERRUPTS()
Definition: miscadmin.h:102
BlockNumber RecordAndGetPageWithFreeSpace(Relation rel, BlockNumber oldPage, Size oldSpaceAvail, Size spaceNeeded)
Definition: freespace.c:149
#define MaxFSMRequestSize
Definition: freespace.c:65
int Buffer
Definition: buf.h:23
void FreeSpaceMapVacuumRange(Relation rel, BlockNumber start, BlockNumber end)
Definition: freespace.c:355
#define XLogHintBitIsNeeded()
Definition: xlog.h:212
Pointer Page
Definition: bufpage.h:78
static BlockNumber fsm_search(Relation rel, uint8 min_cat)
Definition: freespace.c:691
int fsm_search_avail(Buffer buf, uint8 minvalue, bool advancenext, bool exclusive_lock_held)
Definition: fsmpage.c:158
void PageInit(Page page, Size pageSize, Size specialSize)
Definition: bufpage.c:42