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) 1996-2020, 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 */
215 
216  page = BufferGetPage(buf);
217  if (PageIsNew(page))
218  PageInit(page, BLCKSZ, 0);
219 
220  if (fsm_set_avail(page, slot, new_cat))
221  MarkBufferDirtyHint(buf, false);
222  UnlockReleaseBuffer(buf);
223 }
224 
225 /*
226  * GetRecordedFreeSpace - return the amount of free space on a particular page,
227  * according to the FSM.
228  */
229 Size
231 {
232  FSMAddress addr;
233  uint16 slot;
234  Buffer buf;
235  uint8 cat;
236 
237  /* Get the location of the FSM byte representing the heap block */
238  addr = fsm_get_location(heapBlk, &slot);
239 
240  buf = fsm_readbuf(rel, addr, false);
241  if (!BufferIsValid(buf))
242  return 0;
243  cat = fsm_get_avail(BufferGetPage(buf), slot);
244  ReleaseBuffer(buf);
245 
246  return fsm_space_cat_to_avail(cat);
247 }
248 
249 /*
250  * FreeSpaceMapPrepareTruncateRel - prepare for truncation of a relation.
251  *
252  * nblocks is the new size of the heap.
253  *
254  * Return the number of blocks of new FSM.
255  * If it's InvalidBlockNumber, there is nothing to truncate;
256  * otherwise the caller is responsible for calling smgrtruncate()
257  * to truncate the FSM pages, and FreeSpaceMapVacuumRange()
258  * to update upper-level pages in the FSM.
259  */
262 {
263  BlockNumber new_nfsmblocks;
264  FSMAddress first_removed_address;
265  uint16 first_removed_slot;
266  Buffer buf;
267 
268  RelationOpenSmgr(rel);
269 
270  /*
271  * If no FSM has been created yet for this relation, there's nothing to
272  * truncate.
273  */
274  if (!smgrexists(rel->rd_smgr, FSM_FORKNUM))
275  return InvalidBlockNumber;
276 
277  /* Get the location in the FSM of the first removed heap block */
278  first_removed_address = fsm_get_location(nblocks, &first_removed_slot);
279 
280  /*
281  * Zero out the tail of the last remaining FSM page. If the slot
282  * representing the first removed heap block is at a page boundary, as the
283  * first slot on the FSM page that first_removed_address points to, we can
284  * just truncate that page altogether.
285  */
286  if (first_removed_slot > 0)
287  {
288  buf = fsm_readbuf(rel, first_removed_address, false);
289  if (!BufferIsValid(buf))
290  return InvalidBlockNumber; /* nothing to do; the FSM was already
291  * smaller */
293 
294  /* NO EREPORT(ERROR) from here till changes are logged */
296 
297  fsm_truncate_avail(BufferGetPage(buf), first_removed_slot);
298 
299  /*
300  * Truncation of a relation is WAL-logged at a higher-level, and we
301  * will be called at WAL replay. But if checksums are enabled, we need
302  * to still write a WAL record to protect against a torn page, if the
303  * page is flushed to disk before the truncation WAL record. We cannot
304  * use MarkBufferDirtyHint here, because that will not dirty the page
305  * during recovery.
306  */
307  MarkBufferDirty(buf);
309  log_newpage_buffer(buf, false);
310 
312 
313  UnlockReleaseBuffer(buf);
314 
315  new_nfsmblocks = fsm_logical_to_physical(first_removed_address) + 1;
316  }
317  else
318  {
319  new_nfsmblocks = fsm_logical_to_physical(first_removed_address);
320  if (smgrnblocks(rel->rd_smgr, FSM_FORKNUM) <= new_nfsmblocks)
321  return InvalidBlockNumber; /* nothing to do; the FSM was already
322  * smaller */
323  }
324 
325  return new_nfsmblocks;
326 }
327 
328 /*
329  * FreeSpaceMapVacuum - update upper-level pages in the rel's FSM
330  *
331  * We assume that the bottom-level pages have already been updated with
332  * new free-space information.
333  */
334 void
336 {
337  bool dummy;
338 
339  /* Recursively scan the tree, starting at the root */
340  (void) fsm_vacuum_page(rel, FSM_ROOT_ADDRESS,
342  &dummy);
343 }
344 
345 /*
346  * FreeSpaceMapVacuumRange - update upper-level pages in the rel's FSM
347  *
348  * As above, but assume that only heap pages between start and end-1 inclusive
349  * have new free-space information, so update only the upper-level slots
350  * covering that block range. end == InvalidBlockNumber is equivalent to
351  * "all the rest of the relation".
352  */
353 void
355 {
356  bool dummy;
357 
358  /* Recursively scan the tree, starting at the root */
359  if (end > start)
360  (void) fsm_vacuum_page(rel, FSM_ROOT_ADDRESS, start, end, &dummy);
361 }
362 
363 /******** Internal routines ********/
364 
365 /*
366  * Return category corresponding x bytes of free space
367  */
368 static uint8
370 {
371  int cat;
372 
373  Assert(avail < BLCKSZ);
374 
375  if (avail >= MaxFSMRequestSize)
376  return 255;
377 
378  cat = avail / FSM_CAT_STEP;
379 
380  /*
381  * The highest category, 255, is reserved for MaxFSMRequestSize bytes or
382  * more.
383  */
384  if (cat > 254)
385  cat = 254;
386 
387  return (uint8) cat;
388 }
389 
390 /*
391  * Return the lower bound of the range of free space represented by given
392  * category.
393  */
394 static Size
396 {
397  /* The highest category represents exactly MaxFSMRequestSize bytes. */
398  if (cat == 255)
399  return MaxFSMRequestSize;
400  else
401  return cat * FSM_CAT_STEP;
402 }
403 
404 /*
405  * Which category does a page need to have, to accommodate x bytes of data?
406  * While fsm_space_avail_to_cat() rounds down, this needs to round up.
407  */
408 static uint8
410 {
411  int cat;
412 
413  /* Can't ask for more space than the highest category represents */
414  if (needed > MaxFSMRequestSize)
415  elog(ERROR, "invalid FSM request size %zu", needed);
416 
417  if (needed == 0)
418  return 1;
419 
420  cat = (needed + FSM_CAT_STEP - 1) / FSM_CAT_STEP;
421 
422  if (cat > 255)
423  cat = 255;
424 
425  return (uint8) cat;
426 }
427 
428 /*
429  * Returns the physical block number of a FSM page
430  */
431 static BlockNumber
433 {
434  BlockNumber pages;
435  int leafno;
436  int l;
437 
438  /*
439  * Calculate the logical page number of the first leaf page below the
440  * given page.
441  */
442  leafno = addr.logpageno;
443  for (l = 0; l < addr.level; l++)
444  leafno *= SlotsPerFSMPage;
445 
446  /* Count upper level nodes required to address the leaf page */
447  pages = 0;
448  for (l = 0; l < FSM_TREE_DEPTH; l++)
449  {
450  pages += leafno + 1;
451  leafno /= SlotsPerFSMPage;
452  }
453 
454  /*
455  * If the page we were asked for wasn't at the bottom level, subtract the
456  * additional lower level pages we counted above.
457  */
458  pages -= addr.level;
459 
460  /* Turn the page count into 0-based block number */
461  return pages - 1;
462 }
463 
464 /*
465  * Return the FSM location corresponding to given heap block.
466  */
467 static FSMAddress
469 {
470  FSMAddress addr;
471 
472  addr.level = FSM_BOTTOM_LEVEL;
473  addr.logpageno = heapblk / SlotsPerFSMPage;
474  *slot = heapblk % SlotsPerFSMPage;
475 
476  return addr;
477 }
478 
479 /*
480  * Return the heap block number corresponding to given location in the FSM.
481  */
482 static BlockNumber
484 {
485  Assert(addr.level == FSM_BOTTOM_LEVEL);
486  return ((unsigned int) addr.logpageno) * SlotsPerFSMPage + slot;
487 }
488 
489 /*
490  * Given a logical address of a child page, get the logical page number of
491  * the parent, and the slot within the parent corresponding to the child.
492  */
493 static FSMAddress
495 {
496  FSMAddress parent;
497 
498  Assert(child.level < FSM_ROOT_LEVEL);
499 
500  parent.level = child.level + 1;
501  parent.logpageno = child.logpageno / SlotsPerFSMPage;
502  *slot = child.logpageno % SlotsPerFSMPage;
503 
504  return parent;
505 }
506 
507 /*
508  * Given a logical address of a parent page and a slot number, get the
509  * logical address of the corresponding child page.
510  */
511 static FSMAddress
513 {
514  FSMAddress child;
515 
516  Assert(parent.level > FSM_BOTTOM_LEVEL);
517 
518  child.level = parent.level - 1;
519  child.logpageno = parent.logpageno * SlotsPerFSMPage + slot;
520 
521  return child;
522 }
523 
524 /*
525  * Read a FSM page.
526  *
527  * If the page doesn't exist, InvalidBuffer is returned, or if 'extend' is
528  * true, the FSM file is extended.
529  */
530 static Buffer
531 fsm_readbuf(Relation rel, FSMAddress addr, bool extend)
532 {
533  BlockNumber blkno = fsm_logical_to_physical(addr);
534  Buffer buf;
535 
536  RelationOpenSmgr(rel);
537 
538  /*
539  * If we haven't cached the size of the FSM yet, check it first. Also
540  * recheck if the requested block seems to be past end, since our cached
541  * value might be stale. (We send smgr inval messages on truncation, but
542  * not on extension.)
543  */
545  blkno >= rel->rd_smgr->smgr_fsm_nblocks)
546  {
547  if (smgrexists(rel->rd_smgr, FSM_FORKNUM))
549  FSM_FORKNUM);
550  else
551  rel->rd_smgr->smgr_fsm_nblocks = 0;
552  }
553 
554  /* Handle requests beyond EOF */
555  if (blkno >= rel->rd_smgr->smgr_fsm_nblocks)
556  {
557  if (extend)
558  fsm_extend(rel, blkno + 1);
559  else
560  return InvalidBuffer;
561  }
562 
563  /*
564  * Use ZERO_ON_ERROR mode, and initialize the page if necessary. The FSM
565  * information is not accurate anyway, so it's better to clear corrupt
566  * pages than error out. Since the FSM changes are not WAL-logged, the
567  * so-called torn page problem on crash can lead to pages with corrupt
568  * headers, for example.
569  *
570  * The initialize-the-page part is trickier than it looks, because of the
571  * possibility of multiple backends doing this concurrently, and our
572  * desire to not uselessly take the buffer lock in the normal path where
573  * the page is OK. We must take the lock to initialize the page, so
574  * recheck page newness after we have the lock, in case someone else
575  * already did it. Also, because we initially check PageIsNew with no
576  * lock, it's possible to fall through and return the buffer while someone
577  * else is still initializing the page (i.e., we might see pd_upper as set
578  * but other page header fields are still zeroes). This is harmless for
579  * callers that will take a buffer lock themselves, but some callers
580  * inspect the page without any lock at all. The latter is OK only so
581  * long as it doesn't depend on the page header having correct contents.
582  * Current usage is safe because PageGetContents() does not require that.
583  */
584  buf = ReadBufferExtended(rel, FSM_FORKNUM, blkno, RBM_ZERO_ON_ERROR, NULL);
585  if (PageIsNew(BufferGetPage(buf)))
586  {
588  if (PageIsNew(BufferGetPage(buf)))
589  PageInit(BufferGetPage(buf), BLCKSZ, 0);
591  }
592  return buf;
593 }
594 
595 /*
596  * Ensure that the FSM fork is at least fsm_nblocks long, extending
597  * it if necessary with empty pages. And by empty, I mean pages filled
598  * with zeros, meaning there's no free space.
599  */
600 static void
601 fsm_extend(Relation rel, BlockNumber fsm_nblocks)
602 {
603  BlockNumber fsm_nblocks_now;
604  PGAlignedBlock pg;
605 
606  PageInit((Page) pg.data, BLCKSZ, 0);
607 
608  /*
609  * We use the relation extension lock to lock out other backends trying to
610  * extend the FSM at the same time. It also locks out extension of the
611  * main fork, unnecessarily, but extending the FSM happens seldom enough
612  * that it doesn't seem worthwhile to have a separate lock tag type for
613  * it.
614  *
615  * Note that another backend might have extended or created the relation
616  * by the time we get the lock.
617  */
619 
620  /* Might have to re-open if a cache flush happened */
621  RelationOpenSmgr(rel);
622 
623  /*
624  * Create the FSM file first if it doesn't exist. If smgr_fsm_nblocks is
625  * positive then it must exist, no need for an smgrexists call.
626  */
627  if ((rel->rd_smgr->smgr_fsm_nblocks == 0 ||
630  smgrcreate(rel->rd_smgr, FSM_FORKNUM, false);
631 
632  fsm_nblocks_now = smgrnblocks(rel->rd_smgr, FSM_FORKNUM);
633 
634  while (fsm_nblocks_now < fsm_nblocks)
635  {
636  PageSetChecksumInplace((Page) pg.data, fsm_nblocks_now);
637 
638  smgrextend(rel->rd_smgr, FSM_FORKNUM, fsm_nblocks_now,
639  pg.data, false);
640  fsm_nblocks_now++;
641  }
642 
643  /* Update local cache with the up-to-date size */
644  rel->rd_smgr->smgr_fsm_nblocks = fsm_nblocks_now;
645 
647 }
648 
649 /*
650  * Set value in given FSM page and slot.
651  *
652  * If minValue > 0, the updated page is also searched for a page with at
653  * least minValue of free space. If one is found, its slot number is
654  * returned, -1 otherwise.
655  */
656 static int
658  uint8 newValue, uint8 minValue)
659 {
660  Buffer buf;
661  Page page;
662  int newslot = -1;
663 
664  buf = fsm_readbuf(rel, addr, true);
666 
667  page = BufferGetPage(buf);
668 
669  if (fsm_set_avail(page, slot, newValue))
670  MarkBufferDirtyHint(buf, false);
671 
672  if (minValue != 0)
673  {
674  /* Search while we still hold the lock */
675  newslot = fsm_search_avail(buf, minValue,
676  addr.level == FSM_BOTTOM_LEVEL,
677  true);
678  }
679 
680  UnlockReleaseBuffer(buf);
681 
682  return newslot;
683 }
684 
685 /*
686  * Search the tree for a heap page with at least min_cat of free space
687  */
688 static BlockNumber
689 fsm_search(Relation rel, uint8 min_cat)
690 {
691  int restarts = 0;
693 
694  for (;;)
695  {
696  int slot;
697  Buffer buf;
698  uint8 max_avail = 0;
699 
700  /* Read the FSM page. */
701  buf = fsm_readbuf(rel, addr, false);
702 
703  /* Search within the page */
704  if (BufferIsValid(buf))
705  {
707  slot = fsm_search_avail(buf, min_cat,
708  (addr.level == FSM_BOTTOM_LEVEL),
709  false);
710  if (slot == -1)
711  max_avail = fsm_get_max_avail(BufferGetPage(buf));
712  UnlockReleaseBuffer(buf);
713  }
714  else
715  slot = -1;
716 
717  if (slot != -1)
718  {
719  /*
720  * Descend the tree, or return the found block if we're at the
721  * bottom.
722  */
723  if (addr.level == FSM_BOTTOM_LEVEL)
724  return fsm_get_heap_blk(addr, slot);
725 
726  addr = fsm_get_child(addr, slot);
727  }
728  else if (addr.level == FSM_ROOT_LEVEL)
729  {
730  /*
731  * At the root, failure means there's no page with enough free
732  * space in the FSM. Give up.
733  */
734  return InvalidBlockNumber;
735  }
736  else
737  {
738  uint16 parentslot;
739  FSMAddress parent;
740 
741  /*
742  * At lower level, failure can happen if the value in the upper-
743  * level node didn't reflect the value on the lower page. Update
744  * the upper node, to avoid falling into the same trap again, and
745  * start over.
746  *
747  * There's a race condition here, if another backend updates this
748  * page right after we release it, and gets the lock on the parent
749  * page before us. We'll then update the parent page with the now
750  * stale information we had. It's OK, because it should happen
751  * rarely, and will be fixed by the next vacuum.
752  */
753  parent = fsm_get_parent(addr, &parentslot);
754  fsm_set_and_search(rel, parent, parentslot, max_avail, 0);
755 
756  /*
757  * If the upper pages are badly out of date, we might need to loop
758  * quite a few times, updating them as we go. Any inconsistencies
759  * should eventually be corrected and the loop should end. Looping
760  * indefinitely is nevertheless scary, so provide an emergency
761  * valve.
762  */
763  if (restarts++ > 10000)
764  return InvalidBlockNumber;
765 
766  /* Start search all over from the root */
767  addr = FSM_ROOT_ADDRESS;
768  }
769  }
770 }
771 
772 
773 /*
774  * Recursive guts of FreeSpaceMapVacuum
775  *
776  * Examine the FSM page indicated by addr, as well as its children, updating
777  * upper-level nodes that cover the heap block range from start to end-1.
778  * (It's okay if end is beyond the actual end of the map.)
779  * Return the maximum freespace value on this page.
780  *
781  * If addr is past the end of the FSM, set *eof_p to true and return 0.
782  *
783  * This traverses the tree in depth-first order. The tree is stored
784  * physically in depth-first order, so this should be pretty I/O efficient.
785  */
786 static uint8
788  BlockNumber start, BlockNumber end,
789  bool *eof_p)
790 {
791  Buffer buf;
792  Page page;
793  uint8 max_avail;
794 
795  /* Read the page if it exists, or return EOF */
796  buf = fsm_readbuf(rel, addr, false);
797  if (!BufferIsValid(buf))
798  {
799  *eof_p = true;
800  return 0;
801  }
802  else
803  *eof_p = false;
804 
805  page = BufferGetPage(buf);
806 
807  /*
808  * If we're above the bottom level, recurse into children, and fix the
809  * information stored about them at this level.
810  */
811  if (addr.level > FSM_BOTTOM_LEVEL)
812  {
813  FSMAddress fsm_start,
814  fsm_end;
815  uint16 fsm_start_slot,
816  fsm_end_slot;
817  int slot,
818  start_slot,
819  end_slot;
820  bool eof = false;
821 
822  /*
823  * Compute the range of slots we need to update on this page, given
824  * the requested range of heap blocks to consider. The first slot to
825  * update is the one covering the "start" block, and the last slot is
826  * the one covering "end - 1". (Some of this work will be duplicated
827  * in each recursive call, but it's cheap enough to not worry about.)
828  */
829  fsm_start = fsm_get_location(start, &fsm_start_slot);
830  fsm_end = fsm_get_location(end - 1, &fsm_end_slot);
831 
832  while (fsm_start.level < addr.level)
833  {
834  fsm_start = fsm_get_parent(fsm_start, &fsm_start_slot);
835  fsm_end = fsm_get_parent(fsm_end, &fsm_end_slot);
836  }
837  Assert(fsm_start.level == addr.level);
838 
839  if (fsm_start.logpageno == addr.logpageno)
840  start_slot = fsm_start_slot;
841  else if (fsm_start.logpageno > addr.logpageno)
842  start_slot = SlotsPerFSMPage; /* shouldn't get here... */
843  else
844  start_slot = 0;
845 
846  if (fsm_end.logpageno == addr.logpageno)
847  end_slot = fsm_end_slot;
848  else if (fsm_end.logpageno > addr.logpageno)
849  end_slot = SlotsPerFSMPage - 1;
850  else
851  end_slot = -1; /* shouldn't get here... */
852 
853  for (slot = start_slot; slot <= end_slot; slot++)
854  {
855  int child_avail;
856 
858 
859  /* After we hit end-of-file, just clear the rest of the slots */
860  if (!eof)
861  child_avail = fsm_vacuum_page(rel, fsm_get_child(addr, slot),
862  start, end,
863  &eof);
864  else
865  child_avail = 0;
866 
867  /* Update information about the child */
868  if (fsm_get_avail(page, slot) != child_avail)
869  {
871  fsm_set_avail(page, slot, child_avail);
872  MarkBufferDirtyHint(buf, false);
874  }
875  }
876  }
877 
878  /* Now get the maximum value on the page, to return to caller */
879  max_avail = fsm_get_max_avail(page);
880 
881  /*
882  * Reset the next slot pointer. This encourages the use of low-numbered
883  * pages, increasing the chances that a later vacuum can truncate the
884  * relation. We don't bother with a lock here, nor with marking the page
885  * dirty if it wasn't already, since this is just a hint.
886  */
887  ((FSMPage) PageGetContents(page))->fp_next_slot = 0;
888 
889  ReleaseBuffer(buf);
890 
891  return max_avail;
892 }
#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:1014
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:3553
static FSMAddress fsm_get_parent(FSMAddress child, uint16 *slot)
Definition: freespace.c:494
void MarkBufferDirty(Buffer buffer)
Definition: bufmgr.c:1468
#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:652
bool InRecovery
Definition: xlog.c:204
#define END_CRIT_SECTION()
Definition: miscadmin.h:134
Buffer XLogReadBufferExtended(RelFileNode rnode, ForkNumber forknum, BlockNumber blkno, ReadBufferMode mode)
Definition: xlogutils.c:442
unsigned char uint8
Definition: c.h:365
#define InvalidBuffer
Definition: buf.h:25
#define START_CRIT_SECTION()
Definition: miscadmin.h:132
static BlockNumber fsm_get_heap_blk(FSMAddress addr, uint16 slot)
Definition: freespace.c:483
uint32 BlockNumber
Definition: block.h:31
void ReleaseBuffer(Buffer buffer)
Definition: bufmgr.c:3483
static uint8 fsm_vacuum_page(Relation rel, FSMAddress addr, BlockNumber start, BlockNumber end, bool *eof)
Definition: freespace.c:787
BlockNumber smgr_fsm_nblocks
Definition: smgr.h:55
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
static FSMAddress fsm_get_child(FSMAddress parent, uint16 slot)
Definition: freespace.c:512
#define RelationOpenSmgr(relation)
Definition: rel.h:513
static const FSMAddress FSM_ROOT_ADDRESS
Definition: freespace.c:90
char data[BLCKSZ]
Definition: c.h:1104
unsigned short uint16
Definition: c.h:366
#define FSM_ROOT_LEVEL
Definition: freespace.c:76
static Size fsm_space_cat_to_avail(uint8 cat)
Definition: freespace.c:395
void UnlockReleaseBuffer(Buffer buffer)
Definition: bufmgr.c:3506
#define ERROR
Definition: elog.h:43
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:67
BlockNumber FreeSpaceMapPrepareTruncateRel(Relation rel, BlockNumber nblocks)
Definition: freespace.c:261
static FSMAddress fsm_get_location(BlockNumber heapblk, uint16 *slot)
Definition: freespace.c:468
#define BufferGetPage(buffer)
Definition: bufmgr.h:169
void FreeSpaceMapVacuum(Relation rel)
Definition: freespace.c:335
static uint8 fsm_space_avail_to_cat(Size avail)
Definition: freespace.c:369
void LockRelationForExtension(Relation relation, LOCKMODE lockmode)
Definition: lmgr.c:402
Size GetRecordedFreeSpace(Relation rel, BlockNumber heapBlk)
Definition: freespace.c:230
void UnlockRelationForExtension(Relation relation, LOCKMODE lockmode)
Definition: lmgr.c:452
#define PageGetContents(page)
Definition: bufpage.h:246
void LockBuffer(Buffer buffer, int mode)
Definition: bufmgr.c:3722
static void fsm_extend(Relation rel, BlockNumber fsm_nblocks)
Definition: freespace.c:601
static uint8 fsm_space_needed_to_cat(Size needed)
Definition: freespace.c:409
static int fsm_set_and_search(Relation rel, FSMAddress addr, uint16 slot, uint8 newValue, uint8 minValue)
Definition: freespace.c:657
BlockNumber smgrnblocks(SMgrRelation reln, ForkNumber forknum)
Definition: smgr.c:538
#define Assert(condition)
Definition: c.h:738
size_t Size
Definition: c.h:466
#define InvalidBlockNumber
Definition: block.h:33
void PageSetChecksumInplace(Page page, BlockNumber blkno)
Definition: bufpage.c:1194
static BlockNumber fsm_logical_to_physical(FSMAddress addr)
Definition: freespace.c:432
#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:562
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:531
#define elog(elevel,...)
Definition: elog.h:214
#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:99
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:354
#define XLogHintBitIsNeeded()
Definition: xlog.h:202
Pointer Page
Definition: bufpage.h:78
static BlockNumber fsm_search(Relation rel, uint8 min_cat)
Definition: freespace.c:689
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