/*
* Helper function to perform deletion of index entries from a bucket.
*
* This function expects that the caller has acquired a cleanup lock on the
* primary bucket page, and will return with a write lock again held on the
* primary bucket page. The lock won't necessarily be held continuously,
* though, because we'll release it when visiting overflow pages.
*
* It would be very bad if this function cleaned a page while some other
* backend was in the midst of scanning it, because hashgettuple assumes
* that the next valid TID will be greater than or equal to the current
* valid TID. There can't be any concurrent scans in progress when we first
* enter this function because of the cleanup lock we hold on the primary
* bucket page, but as soon as we release that lock, there might be. We
* handle that by conspiring to prevent those scans from passing our cleanup
* scan. To do that, we lock the next page in the bucket chain before
* releasing the lock on the previous page. (This type of lock chaining is
* not ideal, so we might want to look for a better solution at some point.)
*
* We need to retain a pin on the primary bucket to ensure that no concurrent
* split can start.
*/
void
hashbucketcleanup(Relation rel, Bucket cur_bucket, Buffer bucket_buf,
BlockNumber bucket_blkno, BufferAccessStrategy bstrategy,
uint32 maxbucket, uint32 highmask, uint32 lowmask,
double *tuples_removed, double *num_index_tuples,
bool split_cleanup,
IndexBulkDeleteCallback callback, void *callback_state)
{
BlockNumber blkno;
Buffer buf;
Bucket new_bucket PG_USED_FOR_ASSERTS_ONLY = InvalidBucket;
bool bucket_dirty = false;
blkno = bucket_blkno;
buf = bucket_buf;
if (split_cleanup)
new_bucket = _hash_get_newbucket_from_oldbucket(rel, cur_bucket,
lowmask, maxbucket);
/* Scan each page in bucket */
for (;;)
{
HashPageOpaque opaque;
OffsetNumber offno;
OffsetNumber maxoffno;
Buffer next_buf;
Page page;
OffsetNumber deletable[MaxOffsetNumber];
int ndeletable = 0;
bool retain_pin = false;
bool clear_dead_marking = false;
vacuum_delay_point();
page = BufferGetPage(buf);
opaque = (HashPageOpaque) PageGetSpecialPointer(page);
/* Scan each tuple in page */
maxoffno = PageGetMaxOffsetNumber(page);
for (offno = FirstOffsetNumber;
offno <= maxoffno;
offno = OffsetNumberNext(offno))
{
ItemPointer htup;
IndexTuple itup;
Bucket bucket;
bool kill_tuple = false;
itup = (IndexTuple) PageGetItem(page,
PageGetItemId(page, offno));
htup = &(itup->t_tid);
/*
* To remove the dead tuples, we strictly want to rely on results
* of callback function. refer btvacuumpage for detailed reason.
*/
if (callback && callback(htup, callback_state))
{
kill_tuple = true;
if (tuples_removed)
*tuples_removed += 1;
}
else if (split_cleanup)
{
/* delete the tuples that are moved by split. */
bucket = _hash_hashkey2bucket(_hash_get_indextuple_hashkey(itup),
maxbucket,
highmask,
lowmask);
/* mark the item for deletion */
if (bucket != cur_bucket)
{
/*
* We expect tuples to either belong to current bucket or
* new_bucket. This is ensured because we don't allow
* further splits from bucket that contains garbage. See
* comments in _hash_expandtable.
*/
Assert(bucket == new_bucket);
kill_tuple = true;
}
}
if (kill_tuple)
{
/* mark the item for deletion */
deletable[ndeletable++] = offno;
}
else
{
/* we're keeping it, so count it */
if (num_index_tuples)
*num_index_tuples += 1;
}
}
/* retain the pin on primary bucket page till end of bucket scan */
if (blkno == bucket_blkno)
retain_pin = true;
else
retain_pin = false;
blkno = opaque->hasho_nextblkno;
/*
* Apply deletions, advance to next page and write page if needed.
*/
if (ndeletable > 0)
{
/* No ereport(ERROR) until changes are logged */
START_CRIT_SECTION();
PageIndexMultiDelete(page, deletable, ndeletable);
bucket_dirty = true;
/*
* Let us mark the page as clean if vacuum removes the DEAD tuples
* from an index page. We do this by clearing
* LH_PAGE_HAS_DEAD_TUPLES flag.
*/
if (tuples_removed && *tuples_removed > 0 &&
H_HAS_DEAD_TUPLES(opaque))
{
opaque->hasho_flag &= ~LH_PAGE_HAS_DEAD_TUPLES;
clear_dead_marking = true;
}
MarkBufferDirty(buf);
/* XLOG stuff */
if (RelationNeedsWAL(rel))
{
xl_hash_delete xlrec;
XLogRecPtr recptr;
xlrec.clear_dead_marking = clear_dead_marking;
xlrec.is_primary_bucket_page = (buf == bucket_buf) ? true : false;
XLogBeginInsert();
XLogRegisterData((char *) &xlrec, SizeOfHashDelete);
/*
* bucket buffer needs to be registered to ensure that we can
* acquire a cleanup lock on it during replay.
*/
if (!xlrec.is_primary_bucket_page)
XLogRegisterBuffer(0, bucket_buf, REGBUF_STANDARD | REGBUF_NO_IMAGE);
XLogRegisterBuffer(1, buf, REGBUF_STANDARD);
XLogRegisterBufData(1, (char *) deletable,
ndeletable * sizeof(OffsetNumber));
recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_DELETE);
PageSetLSN(BufferGetPage(buf), recptr);
}
END_CRIT_SECTION();
}
/* bail out if there are no more pages to scan. */
if (!BlockNumberIsValid(blkno))
break;
next_buf = _hash_getbuf_with_strategy(rel, blkno, HASH_WRITE,
LH_OVERFLOW_PAGE,
bstrategy);
/*
* release the lock on previous page after acquiring the lock on next
* page
*/
if (retain_pin)
LockBuffer(buf, BUFFER_LOCK_UNLOCK);
else
_hash_relbuf(rel, buf);
buf = next_buf;
}
/*
* lock the bucket page to clear the garbage flag and squeeze the bucket.
* if the current buffer is same as bucket buffer, then we already have
* lock on bucket page.
*/
if (buf != bucket_buf)
{
_hash_relbuf(rel, buf);
LockBuffer(bucket_buf, BUFFER_LOCK_EXCLUSIVE);
}
/*
* Clear the garbage flag from bucket after deleting the tuples that are
* moved by split. We purposefully clear the flag before squeeze bucket,
* so that after restart, vacuum shouldn't again try to delete the moved
* by split tuples.
*/
if (split_cleanup)
{
HashPageOpaque bucket_opaque;
Page page;
page = BufferGetPage(bucket_buf);
bucket_opaque = (HashPageOpaque) PageGetSpecialPointer(page);
/* No ereport(ERROR) until changes are logged */
START_CRIT_SECTION();
bucket_opaque->hasho_flag &= ~LH_BUCKET_NEEDS_SPLIT_CLEANUP;
MarkBufferDirty(bucket_buf);
/* XLOG stuff */
if (RelationNeedsWAL(rel))
{
XLogRecPtr recptr;
XLogBeginInsert();
XLogRegisterBuffer(0, bucket_buf, REGBUF_STANDARD);
recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_CLEANUP);
PageSetLSN(page, recptr);
}
END_CRIT_SECTION();
}
/*
* If we have deleted anything, try to compact free space. For squeezing
* the bucket, we must have a cleanup lock, else it can impact the
* ordering of tuples for a scan that has started before it.
*/
if (bucket_dirty && IsBufferCleanupOK(bucket_buf))
_hash_squeezebucket(rel, cur_bucket, bucket_blkno, bucket_buf,
bstrategy);
else
LockBuffer(bucket_buf, BUFFER_LOCK_UNLOCK);
}
/*
* replay allocation of page for split operation
*/
static void
hash_xlog_split_allocate_page(XLogReaderState *record)
{
XLogRecPtr lsn = record->EndRecPtr;
xl_hash_split_allocate_page *xlrec = (xl_hash_split_allocate_page *) XLogRecGetData(record);
Buffer oldbuf;
Buffer newbuf;
Buffer metabuf;
Size datalen PG_USED_FOR_ASSERTS_ONLY;
char *data;
XLogRedoAction action;
/*
* To be consistent with normal operation, here we take cleanup locks on
* both the old and new buckets even though there can't be any concurrent
* inserts.
*/
/* replay the record for old bucket */
action = XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true, &oldbuf);
/*
* Note that we still update the page even if it was restored from a full
* page image, because the special space is not included in the image.
*/
if (action == BLK_NEEDS_REDO || action == BLK_RESTORED)
{
Page oldpage;
HashPageOpaque oldopaque;
oldpage = BufferGetPage(oldbuf);
oldopaque = (HashPageOpaque) PageGetSpecialPointer(oldpage);
oldopaque->hasho_flag = xlrec->old_bucket_flag;
oldopaque->hasho_prevblkno = xlrec->new_bucket;
PageSetLSN(oldpage, lsn);
MarkBufferDirty(oldbuf);
}
/* replay the record for new bucket */
newbuf = XLogInitBufferForRedo(record, 1);
_hash_initbuf(newbuf, xlrec->new_bucket, xlrec->new_bucket,
xlrec->new_bucket_flag, true);
if (!IsBufferCleanupOK(newbuf))
elog(PANIC, "hash_xlog_split_allocate_page: failed to acquire cleanup lock");
MarkBufferDirty(newbuf);
PageSetLSN(BufferGetPage(newbuf), lsn);
/*
* We can release the lock on old bucket early as well but doing here to
* consistent with normal operation.
*/
if (BufferIsValid(oldbuf))
UnlockReleaseBuffer(oldbuf);
if (BufferIsValid(newbuf))
UnlockReleaseBuffer(newbuf);
/*
* Note: in normal operation, we'd update the meta page while still
* holding lock on the old and new bucket pages. But during replay it's
* not necessary to hold those locks, since no other bucket splits can be
* happening concurrently.
*/
/* replay the record for metapage changes */
if (XLogReadBufferForRedo(record, 2, &metabuf) == BLK_NEEDS_REDO)
{
Page page;
HashMetaPage metap;
page = BufferGetPage(metabuf);
metap = HashPageGetMeta(page);
metap->hashm_maxbucket = xlrec->new_bucket;
data = XLogRecGetBlockData(record, 2, &datalen);
if (xlrec->flags & XLH_SPLIT_META_UPDATE_MASKS)
{
uint32 lowmask;
uint32 *highmask;
/* extract low and high masks. */
memcpy(&lowmask, data, sizeof(uint32));
highmask = (uint32 *) ((char *) data + sizeof(uint32));
/* update metapage */
metap->hashm_lowmask = lowmask;
metap->hashm_highmask = *highmask;
data += sizeof(uint32) * 2;
}
if (xlrec->flags & XLH_SPLIT_META_UPDATE_SPLITPOINT)
{
uint32 ovflpoint;
uint32 *ovflpages;
/* extract information of overflow pages. */
memcpy(&ovflpoint, data, sizeof(uint32));
ovflpages = (uint32 *) ((char *) data + sizeof(uint32));
/* update metapage */
metap->hashm_spares[ovflpoint] = *ovflpages;
metap->hashm_ovflpoint = ovflpoint;
}
MarkBufferDirty(metabuf);
PageSetLSN(BufferGetPage(metabuf), lsn);
}
if (BufferIsValid(metabuf))
UnlockReleaseBuffer(metabuf);
}
/*
* Helper function to perform deletion of index entries from a bucket.
*
* This function expects that the caller has acquired a cleanup lock on the
* primary bucket page, and will return with a write lock again held on the
* primary bucket page. The lock won't necessarily be held continuously,
* though, because we'll release it when visiting overflow pages.
*
* It would be very bad if this function cleaned a page while some other
* backend was in the midst of scanning it, because hashgettuple assumes
* that the next valid TID will be greater than or equal to the current
* valid TID. There can't be any concurrent scans in progress when we first
* enter this function because of the cleanup lock we hold on the primary
* bucket page, but as soon as we release that lock, there might be. We
* handle that by conspiring to prevent those scans from passing our cleanup
* scan. To do that, we lock the next page in the bucket chain before
* releasing the lock on the previous page. (This type of lock chaining is
* not ideal, so we might want to look for a better solution at some point.)
*
* We need to retain a pin on the primary bucket to ensure that no concurrent
* split can start.
*/
void
hashbucketcleanup(Relation rel, Bucket cur_bucket, Buffer bucket_buf,
BlockNumber bucket_blkno, BufferAccessStrategy bstrategy,
uint32 maxbucket, uint32 highmask, uint32 lowmask,
double *tuples_removed, double *num_index_tuples,
bool split_cleanup,
IndexBulkDeleteCallback callback, void *callback_state)
{
BlockNumber blkno;
Buffer buf;
Bucket new_bucket PG_USED_FOR_ASSERTS_ONLY = InvalidBucket;
bool bucket_dirty = false;
blkno = bucket_blkno;
buf = bucket_buf;
if (split_cleanup)
new_bucket = _hash_get_newbucket_from_oldbucket(rel, cur_bucket,
lowmask, maxbucket);
/* Scan each page in bucket */
for (;;)
{
HashPageOpaque opaque;
OffsetNumber offno;
OffsetNumber maxoffno;
Buffer next_buf;
Page page;
OffsetNumber deletable[MaxOffsetNumber];
int ndeletable = 0;
bool retain_pin = false;
bool curr_page_dirty = false;
vacuum_delay_point();
page = BufferGetPage(buf);
opaque = (HashPageOpaque) PageGetSpecialPointer(page);
/* Scan each tuple in page */
maxoffno = PageGetMaxOffsetNumber(page);
for (offno = FirstOffsetNumber;
offno <= maxoffno;
offno = OffsetNumberNext(offno))
{
ItemPointer htup;
IndexTuple itup;
Bucket bucket;
bool kill_tuple = false;
itup = (IndexTuple) PageGetItem(page,
PageGetItemId(page, offno));
htup = &(itup->t_tid);
/*
* To remove the dead tuples, we strictly want to rely on results
* of callback function. refer btvacuumpage for detailed reason.
*/
if (callback && callback(htup, callback_state))
{
kill_tuple = true;
if (tuples_removed)
*tuples_removed += 1;
}
else if (split_cleanup)
{
/* delete the tuples that are moved by split. */
bucket = _hash_hashkey2bucket(_hash_get_indextuple_hashkey(itup),
maxbucket,
highmask,
lowmask);
/* mark the item for deletion */
if (bucket != cur_bucket)
{
/*
* We expect tuples to either belong to curent bucket or
* new_bucket. This is ensured because we don't allow
* further splits from bucket that contains garbage. See
* comments in _hash_expandtable.
*/
Assert(bucket == new_bucket);
kill_tuple = true;
}
}
if (kill_tuple)
{
/* mark the item for deletion */
deletable[ndeletable++] = offno;
}
else
{
/* we're keeping it, so count it */
if (num_index_tuples)
*num_index_tuples += 1;
}
}
/* retain the pin on primary bucket page till end of bucket scan */
if (blkno == bucket_blkno)
retain_pin = true;
else
retain_pin = false;
blkno = opaque->hasho_nextblkno;
/*
* Apply deletions, advance to next page and write page if needed.
*/
if (ndeletable > 0)
{
PageIndexMultiDelete(page, deletable, ndeletable);
bucket_dirty = true;
curr_page_dirty = true;
}
/* bail out if there are no more pages to scan. */
if (!BlockNumberIsValid(blkno))
break;
next_buf = _hash_getbuf_with_strategy(rel, blkno, HASH_WRITE,
LH_OVERFLOW_PAGE,
bstrategy);
/*
* release the lock on previous page after acquiring the lock on next
* page
*/
if (curr_page_dirty)
{
if (retain_pin)
_hash_chgbufaccess(rel, buf, HASH_WRITE, HASH_NOLOCK);
else
_hash_wrtbuf(rel, buf);
curr_page_dirty = false;
}
else if (retain_pin)
_hash_chgbufaccess(rel, buf, HASH_READ, HASH_NOLOCK);
else
_hash_relbuf(rel, buf);
buf = next_buf;
}
/*
* lock the bucket page to clear the garbage flag and squeeze the bucket.
* if the current buffer is same as bucket buffer, then we already have
* lock on bucket page.
*/
if (buf != bucket_buf)
{
_hash_relbuf(rel, buf);
_hash_chgbufaccess(rel, bucket_buf, HASH_NOLOCK, HASH_WRITE);
}
/*
* Clear the garbage flag from bucket after deleting the tuples that are
* moved by split. We purposefully clear the flag before squeeze bucket,
* so that after restart, vacuum shouldn't again try to delete the moved
* by split tuples.
*/
if (split_cleanup)
{
HashPageOpaque bucket_opaque;
Page page;
page = BufferGetPage(bucket_buf);
bucket_opaque = (HashPageOpaque) PageGetSpecialPointer(page);
bucket_opaque->hasho_flag &= ~LH_BUCKET_NEEDS_SPLIT_CLEANUP;
}
/*
* If we have deleted anything, try to compact free space. For squeezing
* the bucket, we must have a cleanup lock, else it can impact the
* ordering of tuples for a scan that has started before it.
*/
if (bucket_dirty && IsBufferCleanupOK(bucket_buf))
_hash_squeezebucket(rel, cur_bucket, bucket_blkno, bucket_buf,
bstrategy);
else
_hash_chgbufaccess(rel, bucket_buf, HASH_WRITE, HASH_NOLOCK);
}
/*
* Attempt to expand the hash table by creating one new bucket.
*
* This will silently do nothing if we don't get cleanup lock on old or
* new bucket.
*
* Complete the pending splits and remove the tuples from old bucket,
* if there are any left over from the previous split.
*
* The caller must hold a pin, but no lock, on the metapage buffer.
* The buffer is returned in the same state.
*/
void
_hash_expandtable(Relation rel, Buffer metabuf)
{
HashMetaPage metap;
Bucket old_bucket;
Bucket new_bucket;
uint32 spare_ndx;
BlockNumber start_oblkno;
BlockNumber start_nblkno;
Buffer buf_nblkno;
Buffer buf_oblkno;
Page opage;
Page npage;
HashPageOpaque oopaque;
HashPageOpaque nopaque;
uint32 maxbucket;
uint32 highmask;
uint32 lowmask;
bool metap_update_masks = false;
bool metap_update_splitpoint = false;
restart_expand:
/*
* Write-lock the meta page. It used to be necessary to acquire a
* heavyweight lock to begin a split, but that is no longer required.
*/
LockBuffer(metabuf, BUFFER_LOCK_EXCLUSIVE);
_hash_checkpage(rel, metabuf, LH_META_PAGE);
metap = HashPageGetMeta(BufferGetPage(metabuf));
/*
* Check to see if split is still needed; someone else might have already
* done one while we waited for the lock.
*
* Make sure this stays in sync with _hash_doinsert()
*/
if (metap->hashm_ntuples <=
(double) metap->hashm_ffactor * (metap->hashm_maxbucket + 1))
goto fail;
/*
* Can't split anymore if maxbucket has reached its maximum possible
* value.
*
* Ideally we'd allow bucket numbers up to UINT_MAX-1 (no higher because
* the calculation maxbucket+1 mustn't overflow). Currently we restrict
* to half that because of overflow looping in _hash_log2() and
* insufficient space in hashm_spares[]. It's moot anyway because an
* index with 2^32 buckets would certainly overflow BlockNumber and hence
* _hash_alloc_buckets() would fail, but if we supported buckets smaller
* than a disk block then this would be an independent constraint.
*
* If you change this, see also the maximum initial number of buckets in
* _hash_init().
*/
if (metap->hashm_maxbucket >= (uint32) 0x7FFFFFFE)
goto fail;
/*
* Determine which bucket is to be split, and attempt to take cleanup lock
* on the old bucket. If we can't get the lock, give up.
*
* The cleanup lock protects us not only against other backends, but
* against our own backend as well.
*
* The cleanup lock is mainly to protect the split from concurrent
* inserts. See src/backend/access/hash/README, Lock Definitions for
* further details. Due to this locking restriction, if there is any
* pending scan, the split will give up which is not good, but harmless.
*/
new_bucket = metap->hashm_maxbucket + 1;
old_bucket = (new_bucket & metap->hashm_lowmask);
start_oblkno = BUCKET_TO_BLKNO(metap, old_bucket);
buf_oblkno = _hash_getbuf_with_condlock_cleanup(rel, start_oblkno, LH_BUCKET_PAGE);
if (!buf_oblkno)
goto fail;
opage = BufferGetPage(buf_oblkno);
oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
/*
* We want to finish the split from a bucket as there is no apparent
* benefit by not doing so and it will make the code complicated to finish
* the split that involves multiple buckets considering the case where new
* split also fails. We don't need to consider the new bucket for
* completing the split here as it is not possible that a re-split of new
* bucket starts when there is still a pending split from old bucket.
*/
if (H_BUCKET_BEING_SPLIT(oopaque))
{
/*
* Copy bucket mapping info now; refer the comment in code below where
* we copy this information before calling _hash_splitbucket to see
* why this is okay.
*/
maxbucket = metap->hashm_maxbucket;
highmask = metap->hashm_highmask;
lowmask = metap->hashm_lowmask;
/*
* Release the lock on metapage and old_bucket, before completing the
* split.
*/
LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
LockBuffer(buf_oblkno, BUFFER_LOCK_UNLOCK);
_hash_finish_split(rel, metabuf, buf_oblkno, old_bucket, maxbucket,
highmask, lowmask);
/* release the pin on old buffer and retry for expand. */
_hash_dropbuf(rel, buf_oblkno);
goto restart_expand;
}
/*
* Clean the tuples remained from the previous split. This operation
* requires cleanup lock and we already have one on the old bucket, so
* let's do it. We also don't want to allow further splits from the bucket
* till the garbage of previous split is cleaned. This has two
* advantages; first, it helps in avoiding the bloat due to garbage and
* second is, during cleanup of bucket, we are always sure that the
* garbage tuples belong to most recently split bucket. On the contrary,
* if we allow cleanup of bucket after meta page is updated to indicate
* the new split and before the actual split, the cleanup operation won't
* be able to decide whether the tuple has been moved to the newly created
* bucket and ended up deleting such tuples.
*/
if (H_NEEDS_SPLIT_CLEANUP(oopaque))
{
/*
* Copy bucket mapping info now; refer to the comment in code below
* where we copy this information before calling _hash_splitbucket to
* see why this is okay.
*/
maxbucket = metap->hashm_maxbucket;
highmask = metap->hashm_highmask;
lowmask = metap->hashm_lowmask;
/* Release the metapage lock. */
LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
hashbucketcleanup(rel, old_bucket, buf_oblkno, start_oblkno, NULL,
maxbucket, highmask, lowmask, NULL, NULL, true,
NULL, NULL);
_hash_dropbuf(rel, buf_oblkno);
goto restart_expand;
}
/*
* There shouldn't be any active scan on new bucket.
*
* Note: it is safe to compute the new bucket's blkno here, even though we
* may still need to update the BUCKET_TO_BLKNO mapping. This is because
* the current value of hashm_spares[hashm_ovflpoint] correctly shows
* where we are going to put a new splitpoint's worth of buckets.
*/
start_nblkno = BUCKET_TO_BLKNO(metap, new_bucket);
/*
* If the split point is increasing we need to allocate a new batch of
* bucket pages.
*/
spare_ndx = _hash_spareindex(new_bucket + 1);
if (spare_ndx > metap->hashm_ovflpoint)
{
uint32 buckets_to_add;
Assert(spare_ndx == metap->hashm_ovflpoint + 1);
/*
* We treat allocation of buckets as a separate WAL-logged action.
* Even if we fail after this operation, won't leak bucket pages;
* rather, the next split will consume this space. In any case, even
* without failure we don't use all the space in one split operation.
*/
buckets_to_add = _hash_get_totalbuckets(spare_ndx) - new_bucket;
if (!_hash_alloc_buckets(rel, start_nblkno, buckets_to_add))
{
/* can't split due to BlockNumber overflow */
_hash_relbuf(rel, buf_oblkno);
goto fail;
}
}
/*
* Physically allocate the new bucket's primary page. We want to do this
* before changing the metapage's mapping info, in case we can't get the
* disk space. Ideally, we don't need to check for cleanup lock on new
* bucket as no other backend could find this bucket unless meta page is
* updated. However, it is good to be consistent with old bucket locking.
*/
buf_nblkno = _hash_getnewbuf(rel, start_nblkno, MAIN_FORKNUM);
if (!IsBufferCleanupOK(buf_nblkno))
{
_hash_relbuf(rel, buf_oblkno);
_hash_relbuf(rel, buf_nblkno);
goto fail;
}
/*
* Since we are scribbling on the pages in the shared buffers, establish a
* critical section. Any failure in this next code leaves us with a big
* problem: the metapage is effectively corrupt but could get written back
* to disk.
*/
START_CRIT_SECTION();
/*
* Okay to proceed with split. Update the metapage bucket mapping info.
*/
metap->hashm_maxbucket = new_bucket;
if (new_bucket > metap->hashm_highmask)
{
/* Starting a new doubling */
metap->hashm_lowmask = metap->hashm_highmask;
metap->hashm_highmask = new_bucket | metap->hashm_lowmask;
metap_update_masks = true;
}
/*
* If the split point is increasing we need to adjust the hashm_spares[]
* array and hashm_ovflpoint so that future overflow pages will be created
* beyond this new batch of bucket pages.
*/
if (spare_ndx > metap->hashm_ovflpoint)
{
metap->hashm_spares[spare_ndx] = metap->hashm_spares[metap->hashm_ovflpoint];
metap->hashm_ovflpoint = spare_ndx;
metap_update_splitpoint = true;
}
MarkBufferDirty(metabuf);
/*
* Copy bucket mapping info now; this saves re-accessing the meta page
* inside _hash_splitbucket's inner loop. Note that once we drop the
* split lock, other splits could begin, so these values might be out of
* date before _hash_splitbucket finishes. That's okay, since all it
* needs is to tell which of these two buckets to map hashkeys into.
*/
maxbucket = metap->hashm_maxbucket;
highmask = metap->hashm_highmask;
lowmask = metap->hashm_lowmask;
opage = BufferGetPage(buf_oblkno);
oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
/*
* Mark the old bucket to indicate that split is in progress. (At
* operation end, we will clear the split-in-progress flag.) Also, for a
* primary bucket page, hasho_prevblkno stores the number of buckets that
* existed as of the last split, so we must update that value here.
*/
oopaque->hasho_flag |= LH_BUCKET_BEING_SPLIT;
oopaque->hasho_prevblkno = maxbucket;
MarkBufferDirty(buf_oblkno);
npage = BufferGetPage(buf_nblkno);
/*
* initialize the new bucket's primary page and mark it to indicate that
* split is in progress.
*/
nopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
nopaque->hasho_prevblkno = maxbucket;
nopaque->hasho_nextblkno = InvalidBlockNumber;
nopaque->hasho_bucket = new_bucket;
nopaque->hasho_flag = LH_BUCKET_PAGE | LH_BUCKET_BEING_POPULATED;
nopaque->hasho_page_id = HASHO_PAGE_ID;
MarkBufferDirty(buf_nblkno);
/* XLOG stuff */
if (RelationNeedsWAL(rel))
{
xl_hash_split_allocate_page xlrec;
XLogRecPtr recptr;
xlrec.new_bucket = maxbucket;
xlrec.old_bucket_flag = oopaque->hasho_flag;
xlrec.new_bucket_flag = nopaque->hasho_flag;
xlrec.flags = 0;
XLogBeginInsert();
XLogRegisterBuffer(0, buf_oblkno, REGBUF_STANDARD);
XLogRegisterBuffer(1, buf_nblkno, REGBUF_WILL_INIT);
XLogRegisterBuffer(2, metabuf, REGBUF_STANDARD);
if (metap_update_masks)
{
xlrec.flags |= XLH_SPLIT_META_UPDATE_MASKS;
XLogRegisterBufData(2, (char *) &metap->hashm_lowmask, sizeof(uint32));
XLogRegisterBufData(2, (char *) &metap->hashm_highmask, sizeof(uint32));
}
if (metap_update_splitpoint)
{
xlrec.flags |= XLH_SPLIT_META_UPDATE_SPLITPOINT;
XLogRegisterBufData(2, (char *) &metap->hashm_ovflpoint,
sizeof(uint32));
XLogRegisterBufData(2,
(char *) &metap->hashm_spares[metap->hashm_ovflpoint],
sizeof(uint32));
}
XLogRegisterData((char *) &xlrec, SizeOfHashSplitAllocPage);
recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_ALLOCATE_PAGE);
PageSetLSN(BufferGetPage(buf_oblkno), recptr);
PageSetLSN(BufferGetPage(buf_nblkno), recptr);
PageSetLSN(BufferGetPage(metabuf), recptr);
}
END_CRIT_SECTION();
/* drop lock, but keep pin */
LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
/* Relocate records to the new bucket */
_hash_splitbucket(rel, metabuf,
old_bucket, new_bucket,
buf_oblkno, buf_nblkno, NULL,
maxbucket, highmask, lowmask);
/* all done, now release the pins on primary buckets. */
_hash_dropbuf(rel, buf_oblkno);
_hash_dropbuf(rel, buf_nblkno);
return;
/* Here if decide not to split or fail to acquire old bucket lock */
fail:
/* We didn't write the metapage, so just drop lock */
LockBuffer(metabuf, BUFFER_LOCK_UNLOCK);
}
/*
* _hash_splitbucket -- split 'obucket' into 'obucket' and 'nbucket'
*
* This routine is used to partition the tuples between old and new bucket and
* is used to finish the incomplete split operations. To finish the previously
* interrupted split operation, the caller needs to fill htab. If htab is set,
* then we skip the movement of tuples that exists in htab, otherwise NULL
* value of htab indicates movement of all the tuples that belong to the new
* bucket.
*
* We are splitting a bucket that consists of a base bucket page and zero
* or more overflow (bucket chain) pages. We must relocate tuples that
* belong in the new bucket.
*
* The caller must hold cleanup locks on both buckets to ensure that
* no one else is trying to access them (see README).
*
* The caller must hold a pin, but no lock, on the metapage buffer.
* The buffer is returned in the same state. (The metapage is only
* touched if it becomes necessary to add or remove overflow pages.)
*
* Split needs to retain pin on primary bucket pages of both old and new
* buckets till end of operation. This is to prevent vacuum from starting
* while a split is in progress.
*
* In addition, the caller must have created the new bucket's base page,
* which is passed in buffer nbuf, pinned and write-locked. The lock will be
* released here and pin must be released by the caller. (The API is set up
* this way because we must do _hash_getnewbuf() before releasing the metapage
* write lock. So instead of passing the new bucket's start block number, we
* pass an actual buffer.)
*/
static void
_hash_splitbucket(Relation rel,
Buffer metabuf,
Bucket obucket,
Bucket nbucket,
Buffer obuf,
Buffer nbuf,
HTAB *htab,
uint32 maxbucket,
uint32 highmask,
uint32 lowmask)
{
Buffer bucket_obuf;
Buffer bucket_nbuf;
Page opage;
Page npage;
HashPageOpaque oopaque;
HashPageOpaque nopaque;
OffsetNumber itup_offsets[MaxIndexTuplesPerPage];
IndexTuple itups[MaxIndexTuplesPerPage];
Size all_tups_size = 0;
int i;
uint16 nitups = 0;
bucket_obuf = obuf;
opage = BufferGetPage(obuf);
oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
bucket_nbuf = nbuf;
npage = BufferGetPage(nbuf);
nopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
/*
* Partition the tuples in the old bucket between the old bucket and the
* new bucket, advancing along the old bucket's overflow bucket chain and
* adding overflow pages to the new bucket as needed. Outer loop iterates
* once per page in old bucket.
*/
for (;;)
{
BlockNumber oblkno;
OffsetNumber ooffnum;
OffsetNumber omaxoffnum;
/* Scan each tuple in old page */
omaxoffnum = PageGetMaxOffsetNumber(opage);
for (ooffnum = FirstOffsetNumber;
ooffnum <= omaxoffnum;
ooffnum = OffsetNumberNext(ooffnum))
{
IndexTuple itup;
Size itemsz;
Bucket bucket;
bool found = false;
/* skip dead tuples */
if (ItemIdIsDead(PageGetItemId(opage, ooffnum)))
continue;
/*
* Before inserting a tuple, probe the hash table containing TIDs
* of tuples belonging to new bucket, if we find a match, then
* skip that tuple, else fetch the item's hash key (conveniently
* stored in the item) and determine which bucket it now belongs
* in.
*/
itup = (IndexTuple) PageGetItem(opage,
PageGetItemId(opage, ooffnum));
if (htab)
(void) hash_search(htab, &itup->t_tid, HASH_FIND, &found);
if (found)
continue;
bucket = _hash_hashkey2bucket(_hash_get_indextuple_hashkey(itup),
maxbucket, highmask, lowmask);
if (bucket == nbucket)
{
IndexTuple new_itup;
/*
* make a copy of index tuple as we have to scribble on it.
*/
new_itup = CopyIndexTuple(itup);
/*
* mark the index tuple as moved by split, such tuples are
* skipped by scan if there is split in progress for a bucket.
*/
new_itup->t_info |= INDEX_MOVED_BY_SPLIT_MASK;
/*
* insert the tuple into the new bucket. if it doesn't fit on
* the current page in the new bucket, we must allocate a new
* overflow page and place the tuple on that page instead.
*/
itemsz = IndexTupleDSize(*new_itup);
itemsz = MAXALIGN(itemsz);
if (PageGetFreeSpaceForMultipleTuples(npage, nitups + 1) < (all_tups_size + itemsz))
{
/*
* Change the shared buffer state in critical section,
* otherwise any error could make it unrecoverable.
*/
START_CRIT_SECTION();
_hash_pgaddmultitup(rel, nbuf, itups, itup_offsets, nitups);
MarkBufferDirty(nbuf);
/* log the split operation before releasing the lock */
log_split_page(rel, nbuf);
END_CRIT_SECTION();
/* drop lock, but keep pin */
LockBuffer(nbuf, BUFFER_LOCK_UNLOCK);
/* be tidy */
for (i = 0; i < nitups; i++)
pfree(itups[i]);
nitups = 0;
all_tups_size = 0;
/* chain to a new overflow page */
nbuf = _hash_addovflpage(rel, metabuf, nbuf, (nbuf == bucket_nbuf) ? true : false);
npage = BufferGetPage(nbuf);
nopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
}
itups[nitups++] = new_itup;
all_tups_size += itemsz;
}
else
{
/*
* the tuple stays on this page, so nothing to do.
*/
Assert(bucket == obucket);
}
}
oblkno = oopaque->hasho_nextblkno;
/* retain the pin on the old primary bucket */
if (obuf == bucket_obuf)
LockBuffer(obuf, BUFFER_LOCK_UNLOCK);
else
_hash_relbuf(rel, obuf);
/* Exit loop if no more overflow pages in old bucket */
if (!BlockNumberIsValid(oblkno))
{
/*
* Change the shared buffer state in critical section, otherwise
* any error could make it unrecoverable.
*/
START_CRIT_SECTION();
_hash_pgaddmultitup(rel, nbuf, itups, itup_offsets, nitups);
MarkBufferDirty(nbuf);
/* log the split operation before releasing the lock */
log_split_page(rel, nbuf);
END_CRIT_SECTION();
if (nbuf == bucket_nbuf)
LockBuffer(nbuf, BUFFER_LOCK_UNLOCK);
else
_hash_relbuf(rel, nbuf);
/* be tidy */
for (i = 0; i < nitups; i++)
pfree(itups[i]);
break;
}
/* Else, advance to next old page */
obuf = _hash_getbuf(rel, oblkno, HASH_READ, LH_OVERFLOW_PAGE);
opage = BufferGetPage(obuf);
oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
}
/*
* We're at the end of the old bucket chain, so we're done partitioning
* the tuples. Mark the old and new buckets to indicate split is
* finished.
*
* To avoid deadlocks due to locking order of buckets, first lock the old
* bucket and then the new bucket.
*/
LockBuffer(bucket_obuf, BUFFER_LOCK_EXCLUSIVE);
opage = BufferGetPage(bucket_obuf);
oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
LockBuffer(bucket_nbuf, BUFFER_LOCK_EXCLUSIVE);
npage = BufferGetPage(bucket_nbuf);
nopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
START_CRIT_SECTION();
oopaque->hasho_flag &= ~LH_BUCKET_BEING_SPLIT;
nopaque->hasho_flag &= ~LH_BUCKET_BEING_POPULATED;
/*
* After the split is finished, mark the old bucket to indicate that it
* contains deletable tuples. We will clear split-cleanup flag after
* deleting such tuples either at the end of split or at the next split
* from old bucket or at the time of vacuum.
*/
oopaque->hasho_flag |= LH_BUCKET_NEEDS_SPLIT_CLEANUP;
/*
* now write the buffers, here we don't release the locks as caller is
* responsible to release locks.
*/
MarkBufferDirty(bucket_obuf);
MarkBufferDirty(bucket_nbuf);
if (RelationNeedsWAL(rel))
{
XLogRecPtr recptr;
xl_hash_split_complete xlrec;
xlrec.old_bucket_flag = oopaque->hasho_flag;
xlrec.new_bucket_flag = nopaque->hasho_flag;
XLogBeginInsert();
XLogRegisterData((char *) &xlrec, SizeOfHashSplitComplete);
XLogRegisterBuffer(0, bucket_obuf, REGBUF_STANDARD);
XLogRegisterBuffer(1, bucket_nbuf, REGBUF_STANDARD);
recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SPLIT_COMPLETE);
PageSetLSN(BufferGetPage(bucket_obuf), recptr);
PageSetLSN(BufferGetPage(bucket_nbuf), recptr);
}
END_CRIT_SECTION();
/*
* If possible, clean up the old bucket. We might not be able to do this
* if someone else has a pin on it, but if not then we can go ahead. This
* isn't absolutely necessary, but it reduces bloat; if we don't do it
* now, VACUUM will do it eventually, but maybe not until new overflow
* pages have been allocated. Note that there's no need to clean up the
* new bucket.
*/
if (IsBufferCleanupOK(bucket_obuf))
{
LockBuffer(bucket_nbuf, BUFFER_LOCK_UNLOCK);
hashbucketcleanup(rel, obucket, bucket_obuf,
BufferGetBlockNumber(bucket_obuf), NULL,
maxbucket, highmask, lowmask, NULL, NULL, true,
NULL, NULL);
}
else
{
LockBuffer(bucket_nbuf, BUFFER_LOCK_UNLOCK);
LockBuffer(bucket_obuf, BUFFER_LOCK_UNLOCK);
}
}
/*
* _hash_doinsert() -- Handle insertion of a single index tuple.
*
* This routine is called by the public interface routines, hashbuild
* and hashinsert. By here, itup is completely filled in.
*/
void
_hash_doinsert(Relation rel, IndexTuple itup)
{
Buffer buf = InvalidBuffer;
Buffer bucket_buf;
Buffer metabuf;
HashMetaPage metap;
BlockNumber blkno;
BlockNumber oldblkno;
bool retry;
Page page;
HashPageOpaque pageopaque;
Size itemsz;
bool do_expand;
uint32 hashkey;
Bucket bucket;
uint32 maxbucket;
uint32 highmask;
uint32 lowmask;
/*
* Get the hash key for the item (it's stored in the index tuple itself).
*/
hashkey = _hash_get_indextuple_hashkey(itup);
/* compute item size too */
itemsz = IndexTupleDSize(*itup);
itemsz = MAXALIGN(itemsz); /* be safe, PageAddItem will do this but we
* need to be consistent */
restart_insert:
/* Read the metapage */
metabuf = _hash_getbuf(rel, HASH_METAPAGE, HASH_READ, LH_META_PAGE);
metap = HashPageGetMeta(BufferGetPage(metabuf));
/*
* Check whether the item can fit on a hash page at all. (Eventually, we
* ought to try to apply TOAST methods if not.) Note that at this point,
* itemsz doesn't include the ItemId.
*
* XXX this is useless code if we are only storing hash keys.
*/
if (itemsz > HashMaxItemSize((Page) metap))
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
errmsg("index row size %zu exceeds hash maximum %zu",
itemsz, HashMaxItemSize((Page) metap)),
errhint("Values larger than a buffer page cannot be indexed.")));
oldblkno = InvalidBlockNumber;
retry = false;
/*
* Loop until we get a lock on the correct target bucket.
*/
for (;;)
{
/*
* Compute the target bucket number, and convert to block number.
*/
bucket = _hash_hashkey2bucket(hashkey,
metap->hashm_maxbucket,
metap->hashm_highmask,
metap->hashm_lowmask);
blkno = BUCKET_TO_BLKNO(metap, bucket);
/*
* Copy bucket mapping info now; refer the comment in
* _hash_expandtable where we copy this information before calling
* _hash_splitbucket to see why this is okay.
*/
maxbucket = metap->hashm_maxbucket;
highmask = metap->hashm_highmask;
lowmask = metap->hashm_lowmask;
/* Release metapage lock, but keep pin. */
_hash_chgbufaccess(rel, metabuf, HASH_READ, HASH_NOLOCK);
/*
* If the previous iteration of this loop locked the primary page of
* what is still the correct target bucket, we are done. Otherwise,
* drop any old lock before acquiring the new one.
*/
if (retry)
{
if (oldblkno == blkno)
break;
_hash_relbuf(rel, buf);
}
/* Fetch and lock the primary bucket page for the target bucket */
buf = _hash_getbuf(rel, blkno, HASH_WRITE, LH_BUCKET_PAGE);
/*
* Reacquire metapage lock and check that no bucket split has taken
* place while we were awaiting the bucket lock.
*/
_hash_chgbufaccess(rel, metabuf, HASH_NOLOCK, HASH_READ);
oldblkno = blkno;
retry = true;
}
/* remember the primary bucket buffer to release the pin on it at end. */
bucket_buf = buf;
page = BufferGetPage(buf);
pageopaque = (HashPageOpaque) PageGetSpecialPointer(page);
Assert(pageopaque->hasho_bucket == bucket);
/*
* If this bucket is in the process of being split, try to finish the
* split before inserting, because that might create room for the
* insertion to proceed without allocating an additional overflow page.
* It's only interesting to finish the split if we're trying to insert
* into the bucket from which we're removing tuples (the "old" bucket),
* not if we're trying to insert into the bucket into which tuples are
* being moved (the "new" bucket).
*/
if (H_BUCKET_BEING_SPLIT(pageopaque) && IsBufferCleanupOK(buf))
{
/* release the lock on bucket buffer, before completing the split. */
_hash_chgbufaccess(rel, buf, HASH_READ, HASH_NOLOCK);
_hash_finish_split(rel, metabuf, buf, pageopaque->hasho_bucket,
maxbucket, highmask, lowmask);
/* release the pin on old and meta buffer. retry for insert. */
_hash_dropbuf(rel, buf);
_hash_dropbuf(rel, metabuf);
goto restart_insert;
}
/* Do the insertion */
while (PageGetFreeSpace(page) < itemsz)
{
/*
* no space on this page; check for an overflow page
*/
BlockNumber nextblkno = pageopaque->hasho_nextblkno;
if (BlockNumberIsValid(nextblkno))
{
/*
* ovfl page exists; go get it. if it doesn't have room, we'll
* find out next pass through the loop test above. we always
* release both the lock and pin if this is an overflow page, but
* only the lock if this is the primary bucket page, since the pin
* on the primary bucket must be retained throughout the scan.
*/
if (buf != bucket_buf)
_hash_relbuf(rel, buf);
else
_hash_chgbufaccess(rel, buf, HASH_READ, HASH_NOLOCK);
buf = _hash_getbuf(rel, nextblkno, HASH_WRITE, LH_OVERFLOW_PAGE);
page = BufferGetPage(buf);
}
else
{
/*
* we're at the end of the bucket chain and we haven't found a
* page with enough room. allocate a new overflow page.
*/
/* release our write lock without modifying buffer */
_hash_chgbufaccess(rel, buf, HASH_READ, HASH_NOLOCK);
/* chain to a new overflow page */
buf = _hash_addovflpage(rel, metabuf, buf, (buf == bucket_buf) ? true : false);
page = BufferGetPage(buf);
/* should fit now, given test above */
Assert(PageGetFreeSpace(page) >= itemsz);
}
pageopaque = (HashPageOpaque) PageGetSpecialPointer(page);
Assert(pageopaque->hasho_flag == LH_OVERFLOW_PAGE);
Assert(pageopaque->hasho_bucket == bucket);
}
/* found page with enough space, so add the item here */
(void) _hash_pgaddtup(rel, buf, itemsz, itup);
/*
* dirty and release the modified page. if the page we modified was an
* overflow page, we also need to separately drop the pin we retained on
* the primary bucket page.
*/
MarkBufferDirty(buf);
_hash_relbuf(rel, buf);
if (buf != bucket_buf)
_hash_dropbuf(rel, bucket_buf);
/*
* Write-lock the metapage so we can increment the tuple count. After
* incrementing it, check to see if it's time for a split.
*/
_hash_chgbufaccess(rel, metabuf, HASH_NOLOCK, HASH_WRITE);
metap->hashm_ntuples += 1;
/* Make sure this stays in sync with _hash_expandtable() */
do_expand = metap->hashm_ntuples >
(double) metap->hashm_ffactor * (metap->hashm_maxbucket + 1);
/* Write out the metapage and drop lock, but keep pin */
_hash_chgbufaccess(rel, metabuf, HASH_WRITE, HASH_NOLOCK);
/* Attempt to split if a split is needed */
if (do_expand)
_hash_expandtable(rel, metabuf);
/* Finally drop our pin on the metapage */
_hash_dropbuf(rel, metabuf);
}
