src/backend/access/heap/hio.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * hio.c
   4  *        POSTGRES heap access method input/output code.
   5  *
   6  * Portions Copyright (c) 1996-2011, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  *
  10  * IDENTIFICATION
  11  *        src/backend/access/heap/hio.c
  12  *
  13  *-------------------------------------------------------------------------
  14  */
  15
  16 #include "postgres.h"
  17
  18 #include "access/heapam.h"
  19 #include "access/hio.h"
  20 #include "storage/bufmgr.h"
  21 #include "storage/freespace.h"
  22 #include "storage/lmgr.h"
  23 #include "storage/smgr.h"
  24
  25
  26 /*
  27  * RelationPutHeapTuple - place tuple at specified page
  28  *
  29  * !!! EREPORT(ERROR) IS DISALLOWED HERE !!!  Must PANIC on failure!!!
  30  *
  31  * Note - caller must hold BUFFER_LOCK_EXCLUSIVE on the buffer.
  32  */
  33 void
  34 RelationPutHeapTuple(Relation relation,
  35                                          Buffer buffer,
  36                                          HeapTuple tuple)
  37 {
  38         Page            pageHeader;
  39         OffsetNumber offnum;
  40         ItemId          itemId;
  41         Item            item;
  42
  43         /* Add the tuple to the page */
  44         pageHeader = BufferGetPage(buffer);
  45
  46         offnum = PageAddItem(pageHeader, (Item) tuple->t_data,
  47                                                  tuple->t_len, InvalidOffsetNumber, false, true);
  48
  49         if (offnum == InvalidOffsetNumber)
  50                 elog(PANIC, "failed to add tuple to page");
  51
  52         /* Update tuple->t_self to the actual position where it was stored */
  53         ItemPointerSet(&(tuple->t_self), BufferGetBlockNumber(buffer), offnum);
  54
  55         /* Insert the correct position into CTID of the stored tuple, too */
  56         itemId = PageGetItemId(pageHeader, offnum);
  57         item = PageGetItem(pageHeader, itemId);
  58         ((HeapTupleHeader) item)->t_ctid = tuple->t_self;
  59 }
  60
  61 /*
  62  * Read in a buffer, using bulk-insert strategy if bistate isn't NULL.
  63  */
  64 static Buffer
  65 ReadBufferBI(Relation relation, BlockNumber targetBlock,
  66                          BulkInsertState bistate)
  67 {
  68         Buffer          buffer;
  69
  70         /* If not bulk-insert, exactly like ReadBuffer */
  71         if (!bistate)
  72                 return ReadBuffer(relation, targetBlock);
  73
  74         /* If we have the desired block already pinned, re-pin and return it */
  75         if (bistate->current_buf != InvalidBuffer)
  76         {
  77                 if (BufferGetBlockNumber(bistate->current_buf) == targetBlock)
  78                 {
  79                         IncrBufferRefCount(bistate->current_buf);
  80                         return bistate->current_buf;
  81                 }
  82                 /* ... else drop the old buffer */
  83                 ReleaseBuffer(bistate->current_buf);
  84                 bistate->current_buf = InvalidBuffer;
  85         }
  86
  87         /* Perform a read using the buffer strategy */
  88         buffer = ReadBufferExtended(relation, MAIN_FORKNUM, targetBlock,
  89                                                                 RBM_NORMAL, bistate->strategy);
  90
  91         /* Save the selected block as target for future inserts */
  92         IncrBufferRefCount(buffer);
  93         bistate->current_buf = buffer;
  94
  95         return buffer;
  96 }
  97
  98 /*
  99  * RelationGetBufferForTuple
 100  *
 101  *      Returns pinned and exclusive-locked buffer of a page in given relation
 102  *      with free space >= given len.
 103  *
 104  *      If otherBuffer is not InvalidBuffer, then it references a previously
 105  *      pinned buffer of another page in the same relation; on return, this
 106  *      buffer will also be exclusive-locked.  (This case is used by heap_update;
 107  *      the otherBuffer contains the tuple being updated.)
 108  *
 109  *      The reason for passing otherBuffer is that if two backends are doing
 110  *      concurrent heap_update operations, a deadlock could occur if they try
 111  *      to lock the same two buffers in opposite orders.  To ensure that this
 112  *      can't happen, we impose the rule that buffers of a relation must be
 113  *      locked in increasing page number order.  This is most conveniently done
 114  *      by having RelationGetBufferForTuple lock them both, with suitable care
 115  *      for ordering.
 116  *
 117  *      NOTE: it is unlikely, but not quite impossible, for otherBuffer to be the
 118  *      same buffer we select for insertion of the new tuple (this could only
 119  *      happen if space is freed in that page after heap_update finds there's not
 120  *      enough there).  In that case, the page will be pinned and locked only once.
 121  *
 122  *      We normally use FSM to help us find free space.  However,
 123  *      if HEAP_INSERT_SKIP_FSM is specified, we just append a new empty page to
 124  *      the end of the relation if the tuple won't fit on the current target page.
 125  *      This can save some cycles when we know the relation is new and doesn't
 126  *      contain useful amounts of free space.
 127  *
 128  *      HEAP_INSERT_SKIP_FSM is also useful for non-WAL-logged additions to a
 129  *      relation, if the caller holds exclusive lock and is careful to invalidate
 130  *      relation's smgr_targblock before the first insertion --- that ensures that
 131  *      all insertions will occur into newly added pages and not be intermixed
 132  *      with tuples from other transactions.  That way, a crash can't risk losing
 133  *      any committed data of other transactions.  (See heap_insert's comments
 134  *      for additional constraints needed for safe usage of this behavior.)
 135  *
 136  *      The caller can also provide a BulkInsertState object to optimize many
 137  *      insertions into the same relation.      This keeps a pin on the current
 138  *      insertion target page (to save pin/unpin cycles) and also passes a
 139  *      BULKWRITE buffer selection strategy object to the buffer manager.
 140  *      Passing NULL for bistate selects the default behavior.
 141  *
 142  *      We always try to avoid filling existing pages further than the fillfactor.
 143  *      This is OK since this routine is not consulted when updating a tuple and
 144  *      keeping it on the same page, which is the scenario fillfactor is meant
 145  *      to reserve space for.
 146  *
 147  *      ereport(ERROR) is allowed here, so this routine *must* be called
 148  *      before any (unlogged) changes are made in buffer pool.
 149  */
 150 Buffer
 151 RelationGetBufferForTuple(Relation relation, Size len,
 152                                                   Buffer otherBuffer, int options,
 153                                                   struct BulkInsertStateData * bistate)
 154 {
 155         bool            use_fsm = !(options & HEAP_INSERT_SKIP_FSM);
 156         Buffer          buffer = InvalidBuffer;
 157         Page            page;
 158         Size            pageFreeSpace,
 159                                 saveFreeSpace;
 160         BlockNumber targetBlock,
 161                                 otherBlock;
 162         bool            needLock;
 163
 164         len = MAXALIGN(len);            /* be conservative */
 165
 166         /* Bulk insert is not supported for updates, only inserts. */
 167         Assert(otherBuffer == InvalidBuffer || !bistate);
 168
 169         /*
 170          * If we're gonna fail for oversize tuple, do it right away
 171          */
 172         if (len > MaxHeapTupleSize)
 173                 ereport(ERROR,
 174                                 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
 175                                  errmsg("row is too big: size %lu, maximum size %lu",
 176                                                 (unsigned long) len,
 177                                                 (unsigned long) MaxHeapTupleSize)));
 178
 179         /* Compute desired extra freespace due to fillfactor option */
 180         saveFreeSpace = RelationGetTargetPageFreeSpace(relation,
 181                                                                                                    HEAP_DEFAULT_FILLFACTOR);
 182
 183         if (otherBuffer != InvalidBuffer)
 184                 otherBlock = BufferGetBlockNumber(otherBuffer);
 185         else
 186                 otherBlock = InvalidBlockNumber;                /* just to keep compiler quiet */
 187
 188         /*
 189          * We first try to put the tuple on the same page we last inserted a tuple
 190          * on, as cached in the BulkInsertState or relcache entry.      If that
 191          * doesn't work, we ask the Free Space Map to locate a suitable page.
 192          * Since the FSM's info might be out of date, we have to be prepared to
 193          * loop around and retry multiple times. (To insure this isn't an infinite
 194          * loop, we must update the FSM with the correct amount of free space on
 195          * each page that proves not to be suitable.)  If the FSM has no record of
 196          * a page with enough free space, we give up and extend the relation.
 197          *
 198          * When use_fsm is false, we either put the tuple onto the existing target
 199          * page or extend the relation.
 200          */
 201         if (len + saveFreeSpace > MaxHeapTupleSize)
 202         {
 203                 /* can't fit, don't bother asking FSM */
 204                 targetBlock = InvalidBlockNumber;
 205                 use_fsm = false;
 206         }
 207         else if (bistate && bistate->current_buf != InvalidBuffer)
 208                 targetBlock = BufferGetBlockNumber(bistate->current_buf);
 209         else
 210                 targetBlock = RelationGetTargetBlock(relation);
 211
 212         if (targetBlock == InvalidBlockNumber && use_fsm)
 213         {
 214                 /*
 215                  * We have no cached target page, so ask the FSM for an initial
 216                  * target.
 217                  */
 218                 targetBlock = GetPageWithFreeSpace(relation, len + saveFreeSpace);
 219
 220                 /*
 221                  * If the FSM knows nothing of the rel, try the last page before we
 222                  * give up and extend.  This avoids one-tuple-per-page syndrome during
 223                  * bootstrapping or in a recently-started system.
 224                  */
 225                 if (targetBlock == InvalidBlockNumber)
 226                 {
 227                         BlockNumber nblocks = RelationGetNumberOfBlocks(relation);
 228
 229                         if (nblocks > 0)
 230                                 targetBlock = nblocks - 1;
 231                 }
 232         }
 233
 234         while (targetBlock != InvalidBlockNumber)
 235         {
 236                 /*
 237                  * Read and exclusive-lock the target block, as well as the other
 238                  * block if one was given, taking suitable care with lock ordering and
 239                  * the possibility they are the same block.
 240                  */
 241                 if (otherBuffer == InvalidBuffer)
 242                 {
 243                         /* easy case */
 244                         buffer = ReadBufferBI(relation, targetBlock, bistate);
 245                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
 246                 }
 247                 else if (otherBlock == targetBlock)
 248                 {
 249                         /* also easy case */
 250                         buffer = otherBuffer;
 251                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
 252                 }
 253                 else if (otherBlock < targetBlock)
 254                 {
 255                         /* lock other buffer first */
 256                         buffer = ReadBuffer(relation, targetBlock);
 257                         LockBuffer(otherBuffer, BUFFER_LOCK_EXCLUSIVE);
 258                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
 259                 }
 260                 else
 261                 {
 262                         /* lock target buffer first */
 263                         buffer = ReadBuffer(relation, targetBlock);
 264                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
 265                         LockBuffer(otherBuffer, BUFFER_LOCK_EXCLUSIVE);
 266                 }
 267
 268                 /*
 269                  * Now we can check to see if there's enough free space here. If so,
 270                  * we're done.
 271                  */
 272                 page = BufferGetPage(buffer);
 273                 pageFreeSpace = PageGetHeapFreeSpace(page);
 274                 if (len + saveFreeSpace <= pageFreeSpace)
 275                 {
 276                         /* use this page as future insert target, too */
 277                         RelationSetTargetBlock(relation, targetBlock);
 278                         return buffer;
 279                 }
 280
 281                 /*
 282                  * Not enough space, so we must give up our page locks and pin (if
 283                  * any) and prepare to look elsewhere.  We don't care which order we
 284                  * unlock the two buffers in, so this can be slightly simpler than the
 285                  * code above.
 286                  */
 287                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
 288                 if (otherBuffer == InvalidBuffer)
 289                         ReleaseBuffer(buffer);
 290                 else if (otherBlock != targetBlock)
 291                 {
 292                         LockBuffer(otherBuffer, BUFFER_LOCK_UNLOCK);
 293                         ReleaseBuffer(buffer);
 294                 }
 295
 296                 /* Without FSM, always fall out of the loop and extend */
 297                 if (!use_fsm)
 298                         break;
 299
 300                 /*
 301                  * Update FSM as to condition of this page, and ask for another page
 302                  * to try.
 303                  */
 304                 targetBlock = RecordAndGetPageWithFreeSpace(relation,
 305                                                                                                         targetBlock,
 306                                                                                                         pageFreeSpace,
 307                                                                                                         len + saveFreeSpace);
 308         }
 309
 310         /*
 311          * Have to extend the relation.
 312          *
 313          * We have to use a lock to ensure no one else is extending the rel at the
 314          * same time, else we will both try to initialize the same new page.  We
 315          * can skip locking for new or temp relations, however, since no one else
 316          * could be accessing them.
 317          */
 318         needLock = !RELATION_IS_LOCAL(relation);
 319
 320         if (needLock)
 321                 LockRelationForExtension(relation, ExclusiveLock);
 322
 323         /*
 324          * XXX This does an lseek - rather expensive - but at the moment it is the
 325          * only way to accurately determine how many blocks are in a relation.  Is
 326          * it worth keeping an accurate file length in shared memory someplace,
 327          * rather than relying on the kernel to do it for us?
 328          */
 329         buffer = ReadBufferBI(relation, P_NEW, bistate);
 330
 331         /*
 332          * We can be certain that locking the otherBuffer first is OK, since it
 333          * must have a lower page number.
 334          */
 335         if (otherBuffer != InvalidBuffer)
 336                 LockBuffer(otherBuffer, BUFFER_LOCK_EXCLUSIVE);
 337
 338         /*
 339          * Now acquire lock on the new page.
 340          */
 341         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
 342
 343         /*
 344          * Release the file-extension lock; it's now OK for someone else to extend
 345          * the relation some more.      Note that we cannot release this lock before
 346          * we have buffer lock on the new page, or we risk a race condition
 347          * against vacuumlazy.c --- see comments therein.
 348          */
 349         if (needLock)
 350                 UnlockRelationForExtension(relation, ExclusiveLock);
 351
 352         /*
 353          * We need to initialize the empty new page.  Double-check that it really
 354          * is empty (this should never happen, but if it does we don't want to
 355          * risk wiping out valid data).
 356          */
 357         page = BufferGetPage(buffer);
 358
 359         if (!PageIsNew(page))
 360                 elog(ERROR, "page %u of relation \"%s\" should be empty but is not",
 361                          BufferGetBlockNumber(buffer),
 362                          RelationGetRelationName(relation));
 363
 364         PageInit(page, BufferGetPageSize(buffer), 0);
 365
 366         if (len > PageGetHeapFreeSpace(page))
 367         {
 368                 /* We should not get here given the test at the top */
 369                 elog(PANIC, "tuple is too big: size %lu", (unsigned long) len);
 370         }
 371
 372         /*
 373          * Remember the new page as our target for future insertions.
 374          *
 375          * XXX should we enter the new page into the free space map immediately,
 376          * or just keep it for this backend's exclusive use in the short run
 377          * (until VACUUM sees it)?      Seems to depend on whether you expect the
 378          * current backend to make more insertions or not, which is probably a
 379          * good bet most of the time.  So for now, don't add it to FSM yet.
 380          */
 381         RelationSetTargetBlock(relation, BufferGetBlockNumber(buffer));
 382
 383         return buffer;
 384 }