will be released by transaction abort before _bt_end_vacuum gets called.
If either of these "can't happen" errors actually happened, we'd freeze up
trying to acquire an already-held lock. Latest word is that this does
not explain Martin Pitt's trouble report, but it still looks like a bug.
pointer" in every Snapshot struct. This allows removal of the case-by-case
tests in HeapTupleSatisfiesVisibility, which should make it a bit faster
(I didn't try any performance tests though). More importantly, we are no
longer violating portable C practices by assuming that small integers are
distinct from all pointer values, and HeapTupleSatisfiesDirty no longer
has a non-reentrant API involving side-effects on a global variable.
There were a couple of places calling HeapTupleSatisfiesXXX routines
directly rather than through the HeapTupleSatisfiesVisibility macro.
Since these places had to be changed anyway, I chose to make them go
through the macro for uniformity.
Along the way I renamed HeapTupleSatisfiesSnapshot to HeapTupleSatisfiesMVCC
to emphasize that it's only used with MVCC-type snapshots. I was sorely
tempted to rename HeapTupleSatisfiesVisibility to HeapTupleSatisfiesSnapshot,
but forebore for the moment to avoid confusion and reduce the likelihood that
this patch breaks some of the pending patches. Might want to reconsider
doing that later.
I refactored findsplitloc and checksplitloc so that the division of
labor is more clear IMO. I pushed all the space calculation inside the
loop to checksplitloc.
I also fixed the off by 4 in free space calculation caused by
PageGetFreeSpace subtracting sizeof(ItemIdData), even though it was
harmless, because it was distracting and I felt it might come back to
bite us in the future if we change the page layout or alignments.
There's now a new function PageGetExactFreeSpace that doesn't do the
subtraction.
findsplitloc now tries the "just the new item to right page" split as
well. If people don't like the refactoring, I can write a patch to just
add that.
Heikki Linnakangas
> Currently, an index split writes all the data on the split page to
> WAL. That's a lot of WAL traffic. The tuples that are copied to the
> right page need to be WAL logged, but the tuples that stay on the
> original page don't.
Heikki Linnakangas
Standard English uses "may", "can", and "might" in different ways:
may - permission, "You may borrow my rake."
can - ability, "I can lift that log."
might - possibility, "It might rain today."
Unfortunately, in conversational English, their use is often mixed, as
in, "You may use this variable to do X", when in fact, "can" is a better
choice. Similarly, "It may crash" is better stated, "It might crash".
exactly at the point where we need to insert a new item, the calculation used
the wrong size for the "high key" of the new left page. This could lead to
choosing an unworkable split, resulting in "PANIC: failed to add item to the
left sibling" (or "right sibling") failure. Although this bug has been there
a long time, it's very difficult to trigger a failure before 8.2, since there
was generally a lot of free space on both sides of a chosen split. In 8.2,
where the user-selected fill factor determines how much free space the code
tries to leave, an unworkable split is much more likely. Report by Joe
Conway, diagnosis and fix by Heikki Linnakangas.
hold true for operators in a btree operator family. This is mostly to
clarify my own thinking about what the planner can assume for optimization
purposes. (blowing dust off an old abstract-algebra textbook...)
per-column options for btree indexes. The planner's support for this is still
pretty rudimentary; it does not yet know how to plan mergejoins with
nondefault ordering options. The documentation is pretty rudimentary, too.
I'll work on improving that stuff later.
Note incompatible change from prior behavior: ORDER BY ... USING will now be
rejected if the operator is not a less-than or greater-than member of some
btree opclass. This prevents less-than-sane behavior if an operator that
doesn't actually define a proper sort ordering is selected.
having md.c return a success/failure boolean to smgr.c, which was just going
to elog anyway, let md.c issue the elog messages itself. This allows better
error reporting, particularly in cases such as "short read" or "short write"
which Peter was complaining of. Also, remove the kluge of allowing mdread()
to return zeroes from a read-beyond-EOF: this is now an error condition
except when InRecovery or zero_damaged_pages = true. (Hash indexes used to
require that behavior, but no more.) Also, enforce that mdwrite() is to be
used for rewriting existing blocks while mdextend() is to be used for
extending the relation EOF. This restriction lets us get rid of the old
ad-hoc defense against creating huge files by an accidental reference to
a bogus block number: we'll only create new segments in mdextend() not
mdwrite() or mdread(). (Again, when InRecovery we allow it anyway, since
we need to allow updates of blocks that were later truncated away.)
Also, clean up the original makeshift patch for bug #2737: move the
responsibility for padding relation segments to full length into md.c.
or contradictory keys even in cross-data-type scenarios. This is another
benefit of the opfamily rewrite: we can find the needed comparison
operators now.
cases. Operator classes now exist within "operator families". While most
families are equivalent to a single class, related classes can be grouped
into one family to represent the fact that they are semantically compatible.
Cross-type operators are now naturally adjunct parts of a family, without
having to wedge them into a particular opclass as we had done originally.
This commit restructures the catalogs and cleans up enough of the fallout so
that everything still works at least as well as before, but most of the work
needed to actually improve the planner's behavior will come later. Also,
there are not yet CREATE/DROP/ALTER OPERATOR FAMILY commands; the only way
to create a new family right now is to allow CREATE OPERATOR CLASS to make
one by default. I owe some more documentation work, too. But that can all
be done in smaller pieces once this infrastructure is in place.
deletion code to avoid the case where an upper-level btree page remains "half
dead" for a significant period of time, and to block insertions into a key
range that is in process of being re-assigned to the right sibling of the
deleted page's parent. This prevents the scenario reported by Ed L. wherein
index keys could become out-of-order in the grandparent index level.
Since this is a moderately invasive fix, I'm applying it only to HEAD.
The bug exists back to 7.4, but the back branches will get a different patch.
even when a single relation requires more than max_fsm_pages pages. Also,
make VACUUM emit a warning in this case, since it likely means that VACUUM
FULL or other drastic corrective measure is needed. Per reports from Jeff
Frost and others of unexpected changes in the claimed max_fsm_pages need.
on the same index page; we can avoid data copying as well as buffer refcount
manipulations in this common case. Makes for a small but noticeable
improvement in mergejoin speed.
Heikki Linnakangas
operation every so often. This improves the usefulness of PITR log
shipping for hot standby: formerly, if the standby server crashed, it
was necessary to restart it from the last base backup and replay all
the WAL since then. Now it will only need to reread about the same
amount of WAL as the master server would. The behavior might also
come in handy during a long PITR replay sequence. Simon Riggs,
with some editorialization by Tom Lane.
When we are about to split an index page to do an insertion, first look
to see if any entries marked LP_DELETE exist on the page, and if so remove
them to try to make enough space for the desired insert. This should reduce
index bloat in heavily-updated tables, although of course you still need
VACUUM eventually to clean up the heap.
Junji Teramoto
it can handle small fillfactors for ordinary-sized index entries without
failing on large ones; fix nbtinsert.c to distinguish leaf and nonleaf
pages; change the minimum fillfactor to 10% for all index types.
discussion (including making def_arg allow reserved words), add missed
opt_definition for UNIQUE case. Put the reloptions support code in a less
random place (I chose to make a new file access/common/reloptions.c).
Eliminate header inclusion creep. Make the index options functions safely
user-callable (seems like client apps might like to be able to test validity
of options before trying to make an index). Reduce overhead for normal case
with no options by allowing rd_options to be NULL. Fix some unmaintainably
klugy code, including getting rid of Natts_pg_class_fixed at long last.
Some stylistic cleanup too, and pay attention to keeping comments in sync
with code.
Documentation still needs work, though I did fix the omissions in
catalogs.sgml and indexam.sgml.
per-tuple space overhead for sorts in memory. I chose to replace the
previous patch that tried to write out the bare minimum amount of data
when sorting on disk; instead, just dump the MinimalTuples as-is. This
wastes 3 to 10 bytes per tuple depending on architecture and null-bitmap
length, but the simplification in the writetup/readtup routines seems
worth it.
(relpages/reltuples). To do this, create formal support in heapam.c for
"overwrite" tuple updates (including xlog replay capability) and use that
instead of the ad-hoc overwrites we'd been using in VACUUM and CREATE INDEX.
Take the responsibility for updating stats during CREATE INDEX out of the
individual index AMs, and do it where it belongs, in catalog/index.c. Aside
from being more modular, this avoids having to update the same tuple twice in
some paths through CREATE INDEX. It's probably not measurably faster, but
for sure it's a lot cleaner than before.
into a single mostly-physical-order scan of the index. This requires some
ticklish interlocking considerations, but should create no material
performance impact on normal index operations (at least given the
already-committed changes to make scans work a page at a time). VACUUM
itself should get significantly faster in any index that's degenerated to a
very nonlinear page order. Also, we save one pass over the index entirely,
except in the case where there were no deletions to do and so only one pass
happened anyway.
Original patch by Heikki Linnakangas, rework by Tom Lane.
btgettuple and btgetmulti). This eliminates the problem of "re-finding" the
exact stopping point, since the stopping point is effectively always a page
boundary, and index items are never moved across pre-existing page boundaries.
A small penalty is that the keys_are_unique optimization is effectively
disabled (and, therefore, is removed in this patch), causing us to apply
_bt_checkkeys() to at least one more tuple than necessary when looking up a
unique key. However, the advantages for non-unique cases seem great enough to
accept this tradeoff. Aside from simplifying and (sometimes) speeding up the
indexscan code, this will allow us to reimplement btbulkdelete as a largely
sequential scan instead of index-order traversal, thereby significantly
reducing the cost of VACUUM. Those changes will come in a separate patch.
Original patch by Heikki Linnakangas, rework by Tom Lane.
This formulation requires every AM to provide amvacuumcleanup, unlike before,
but it's surely a whole lot cleaner. Also, add an 'amstorage' column to
pg_am so that we can get rid of hardwired knowledge in DefineOpClass().
thereby saving a visit to the metapage in most index searches/updates.
This wouldn't actually save any I/O (since in the old regime the metapage
generally stayed in cache anyway), but it does provide a useful decrease
in bufmgr traffic in high-contention scenarios. Per my recent proposal.
upper-level insertion completes a previously-seen split, we cannot simply grab
the downlink block number out of the buffer, because the buffer could contain
a later state of the page --- or perhaps the page doesn't even exist at all
any more, due to relation truncation. These possibilities have been masked up
to now because the use of full_page_writes effectively ensured that no xlog
replay routine ever actually saw a page state newer than its own change.
Since we're deprecating full_page_writes in 8.1.*, there's no need to fix this
in existing release branches, but we need a fix in HEAD if we want to have any
hope of re-allowing full_page_writes. Accordingly, adjust the contents of
btree WAL records so that we can always get the downlink block number from the
WAL record rather than having to depend on buffer contents. Per report from
Kevin Grittner and Peter Brant.
Improve a few comments in related code while at it.
incrementally by successive inserts rather than by sorting the data.
We were only using the slow path during bootstrap, apparently because
when first written it failed during bootstrap --- but it works fine now
AFAICT. Removing it saves a hundred or so lines of code and produces
noticeably (~10%) smaller initial states of the system catalog indexes.
While that won't make much difference for heavily-modified catalogs,
for the more static ones there may be a useful long-term performance
improvement.
misleadingly-named WriteBuffer routine, and instead require routines that
change buffer pages to call MarkBufferDirty (which does exactly what it says).
We also require that they do so before calling XLogInsert; this takes care of
the synchronization requirement documented in SyncOneBuffer. Note that
because bufmgr takes the buffer content lock (in shared mode) while writing
out any buffer, it doesn't matter whether MarkBufferDirty is executed before
the buffer content change is complete, so long as the content change is
completed before releasing exclusive lock on the buffer. So it's OK to set
the dirtybit before we fill in the LSN.
This eliminates the former kluge of needing to set the dirtybit in LockBuffer.
Aside from making the code more transparent, we can also add some new
debugging assertions, in particular that the caller of MarkBufferDirty must
hold the buffer content lock, not merely a pin.
This commit doesn't make much functional change, but it does eliminate some
duplicated code --- for instance, PageIsNew tests are now done inside
XLogReadBuffer rather than by each caller.
The GIST xlog code still needs a lot of love, but I'll worry about that
separately.
passed extend = true whenever we are reading a page we intend to reinitialize
completely, even if we think the page "should exist". This is because it
might indeed not exist, if the relation got truncated sometime after the
current xlog record was made and before the crash we're trying to recover
from. These two thinkos appear to explain both of the old bug reports
discussed here:
http://archives.postgresql.org/pgsql-hackers/2005-05/msg01369.php
when an error occurs during xlog replay. Also, replace the former risky
'write into a fixed-size buffer with no overflow detection' API for XLOG
record description routines; use an expansible StringInfo instead. (The
latter accounts for most of the patch bulk.)
Qingqing Zhou
we are not holding a buffer content lock; where it was, InterruptHoldoffCount
is positive and so we'd not respond to cancel signals as intended. Also
add missing vacuum_delay_point() call in btvacuumcleanup. This should fix
complaint from Evgeny Gridasov about failure to respond to SIGINT/SIGTERM
in a timely fashion (bug #2257).
during the vacuumcleanup scan that we're going to do anyway. Should
save a few cycles (one calculation per page, not per tuple) as well
as not having to depend on assumptions about heap and index being
in step.
I think this could probably be made to work for GIST too, but that
code looks messy enough that I'm disinclined to try right now.
partial. None of the existing AMs do anything useful except counting
tuples when there's nothing to delete, and we can get a tuple count
from the heap as long as it's not a partial index. (hash actually can
skip anyway because it maintains a tuple count in the index metapage.)
GIST is not currently able to exploit this optimization because, due to
failure to index NULLs, GIST is always effectively partial. Possibly
we should fix that sometime.
Simon Riggs w/ some review by Tom Lane.
just refer to btree index entries as plain IndexTuples, which is what
they have been for a very long time. This is mostly just an exercise
in removing extraneous notation, but it does save a palloc/pfree cycle
per index insertion.
and non-required keys in a btree index scan, mark the required scankeys
with private flag bits SK_BT_REQFWD and/or SK_BT_REQBKWD. This seems
at least marginally clearer to me, and it eliminates a wired-into-the-
data-structure assumption that required keys are consecutive. Even though
that assumption will remain true for the foreseeable future, having it
in there makes the code seem more complex than necessary.
are two basically different kinds of scankeys, and we ought to try harder
to indicate which is used in each place in the code. I've chosen the names
"search scankey" and "insertion scankey", though you could make about
as good an argument for "operator scankey" and "comparison function
scankey".
rather than "return expr;" -- the latter style is used in most of the
tree. I kept the parentheses when they were necessary or useful because
the return expression was complex.
use it. While it normally has been opened earlier during btree index
build, testing shows that it's possible for the link to be closed again
if an sinval reset occurs while the index is being built.
_bt_checkkeys(), instead of checking it in the top-level nbtree.c routines
as formerly. This saves a little bit of loop overhead, but more importantly
it lets us skip performing the index key comparisons for dead tuples.
checks, which were once needed because PageGetMaxOffsetNumber would
fail on empty pages, but are now just redundant. Also, don't set up
local variables that aren't needed in the fast path --- most of the
time, we only need to advance offnum and not step across a page boundary.
Motivated by noticing _bt_step at the top of OProfile profile for a
pgbench run.
comment line where output as too long, and update typedefs for /lib
directory. Also fix case where identifiers were used as variable names
in the backend, but as typedefs in ecpg (favor the backend for
indenting).
Backpatch to 8.1.X.
on every index page they read; in particular to catch the case of an
all-zero page, which PageHeaderIsValid allows to pass. It turns out
hash already had this idea, but it was just Assert()ing things rather
than doing a straight error check, and the Asserts were partially
redundant with PageHeaderIsValid anyway. Per recent failure example
from Jim Nasby. (gist still needs the same treatment.)
the wrong buffer dirty when trying to kill a dead index entry that's on
a page after the one it started on. No risk of data corruption, just
inefficiency, but still a bug.
generated by bitmap index scans. Along the way, simplify and speed up
the code for counting sequential and index scans; it was both confusing
and inefficient to be taking care of that in the per-tuple loops, IMHO.
initdb forced because of internal changes in pg_stat view definitions.
on a page, as suggested by ITAGAKI Takahiro. Also, change a few places
that were using some other estimates of max-items-per-page to consistently
use MaxOffsetNumber. This is conservatively large --- we could have used
the new MaxHeapTuplesPerPage macro, or a similar one for index tuples ---
but those places are simply declaring a fixed-size buffer and assuming it
will work, rather than actively testing for overrun. It seems safer to
size these buffers in a way that can't overflow even if the page is
corrupt.
scankeys arrays that it needs can never have more than INDEX_MAX_KEYS
entries, so it's reasonable to just allocate them as fixed-size local
arrays, and save the cost of palloc/pfree. Not a huge savings, but
a cycle saved is a cycle earned ...
nonconsecutive columns of a multicolumn index, as per discussion around
mid-May (pghackers thread "Best way to scan on-disk bitmaps"). This
turns out to require only minimal changes in btree, and so far as I can
see none at all in GiST. btcostestimate did need some work, but its
original assumption that index selectivity == heap selectivity was
quite bogus even before this.
up have the standard layout with unused space between pd_lower and pd_upper.
When this is set, XLogInsert will omit the unused space without bothering
to scan it to see if it's zero. That saves time in XLogInsert, and also
allows reversion of my earlier patch to make PageRepairFragmentation et al
explicitly re-zero freed space. Per suggestion by Heikki Linnakangas.
Instead of a separate CRC on each backup block, include backup blocks
in their parent WAL record's CRC; this is important to ensure that the
backup block really goes with the WAL record, ie there was not a page
tear right at the start of the backup block. Implement a simple form
of compression of backup blocks: drop any run of zeroes starting at
pd_lower, so as not to store the unused 'hole' that commonly exists in
PG heap and index pages. Tweak PageRepairFragmentation and related
routines to ensure they keep the unused space zeroed, so that the above
compression method remains effective. All per recent discussions.
methods: they all invoke UpdateStats() since they have computed the
number of heap tuples, so I created a function in catalog/index.c that
each AM now calls.
and VACUUM: in the interval between adding a new page to the relation
and formatting it, it was possible for VACUUM to come along and decide
it should format the page too. Though not harmful in itself, this would
cause data loss if a third transaction were able to insert tuples into
the vacuumed page before the original extender got control back.
which is neither needed by nor related to that header. Remove the bogus
inclusion and instead include the header in those C files that actually
need it. Also fix unnecessary inclusions and bad inclusion order in
tsearch2 files.
Essentially, we shoehorn in a lockable-object-type field by taking
a byte away from the lockmethodid, which can surely fit in one byte
instead of two. This allows less artificial definitions of all the
other fields of LOCKTAG; we can get rid of the special pg_xactlock
pseudo-relation, and also support locks on individual tuples and
general database objects (including shared objects). None of those
possibilities are actually exploited just yet, however.
I removed pg_xactlock from pg_class, but did not force initdb for
that change. At this point, relkind 's' (SPECIAL) is unused and
could be removed entirely.
change saves a great deal of space in pg_proc and its primary index,
and it eliminates the former requirement that INDEX_MAX_KEYS and
FUNC_MAX_ARGS have the same value. INDEX_MAX_KEYS is still embedded
in the on-disk representation (because it affects index tuple header
size), but FUNC_MAX_ARGS is not. I believe it would now be possible
to increase FUNC_MAX_ARGS at little cost, but haven't experimented yet.
There are still a lot of vestigial references to FUNC_MAX_ARGS, which
I will clean up in a separate pass. However, getting rid of it
altogether would require changing the FunctionCallInfoData struct,
and I'm not sure I want to buy into that.
access: define new index access method functions 'amgetmulti' that can
fetch multiple TIDs per call. (The functions exist but are totally
untested as yet.) Since I was modifying pg_am anyway, remove the
no-longer-needed 'rel' parameter from amcostestimate functions, and
also remove the vestigial amowner column that was creating useless
work for Alvaro's shared-object-dependencies project.
Initdb forced due to changes in pg_am.
PageIndexTupleDelete() with a single pass of compactification ---
logic mostly lifted from PageRepairFragmentation. I noticed while
profiling that a VACUUM that's cleaning up a whole lot of deleted
tuples would spend as much as a third of its CPU time in
PageIndexTupleDelete; not too surprising considering the loop method
was roughly O(N^2) in the number of tuples involved.
convention for isnull flags. Also, remove the useless InsertIndexResult
return struct from index AM aminsert calls --- there is no reason for
the caller to know where in the index the tuple was inserted, and we
were wasting a palloc cycle per insert to deliver this uninteresting
value (plus nontrivial complexity in some AMs).
I forced initdb because of the change in the signature of the aminsert
routines, even though nothing really looks at those pg_proc entries...
to write out data that we are about to tell the filesystem to drop.
smgr_internal_unlink already had a DropRelFileNodeBuffers call to
get rid of dead buffers without a write after it's no longer possible
to roll back the deleting transaction. Adding a similar call in
smgrtruncate simplifies callers and makes the overall division of
labor clearer. This patch removes the former behavior that VACUUM
would write all dirty buffers of a relation unconditionally.
Also performed an initial run through of upgrading our Copyright date to
extend to 2005 ... first run here was very simple ... change everything
where: grep 1996-2004 && the word 'Copyright' ... scanned through the
generated list with 'less' first, and after, to make sure that I only
picked up the right entries ...
a relation's number of blocks, rather than the possibly-obsolete value
in pg_class.relpages. Scale the value in pg_class.reltuples correspondingly
to arrive at a hopefully more accurate number of rows. When pg_class
contains 0/0, estimate a tuple width from the column datatypes and divide
that into current file size to estimate number of rows. This improved
methodology allows us to jettison the ancient hacks that put bogus default
values into pg_class when a table is first created. Also, per a suggestion
from Simon, make VACUUM (but not VACUUM FULL or ANALYZE) adjust the value
it puts into pg_class.reltuples to try to represent the mean tuple density
instead of the minimal density that actually prevails just after VACUUM.
These changes alter the plans selected for certain regression tests, so
update the expected files accordingly. (I removed join_1.out because
it's not clear if it still applies; we can add back any variant versions
as they are shown to be needed.)
Rather than using ReadBuffer() to increment the reference count on an
already-pinned buffer, we should use IncrBufferRefCount() as it is
faster and does not require acquiring the BufMgrLock.
in all cases when keep_buf = true. This allows ANALYZE's inner loop to
use heap_release_fetch, which saves multiple buffer lookups for the same
page and avoids overestimation of cost by the vacuum cost mechanism.
http://archives.postgresql.org/pgsql-hackers/2004-10/msg00464.php.
This fix is intended to be permanent: it moves the responsibility for
calling SetBufferCommitInfoNeedsSave() into the tqual.c routines,
eliminating the requirement for callers to test whether t_infomask changed.
Also, tighten validity checking on buffer IDs in bufmgr.c --- several
routines were paranoid about out-of-range shared buffer numbers but not
about out-of-range local ones, which seems a tad pointless.
relcache entries. Also, change TransactionIdIsCurrentTransactionId()
so that if consulted during transaction abort, it will not say that
the aborted xact is still current. (It would be better to ensure that
it's never called at all during abort, but I'm not sure we can easily
guarantee that.) In combination, these fix a crash we have seen
occasionally during parallel regression tests of 8.0.
value of 'start' could be past the end of the page, if the page was
split by some concurrent inserting process since we visited it. In
this situation the code could look at bogus entries and possibly find
a match (since after all those entries still contain what they had
before the split). This would lead to 'specified item offset is too large'
followed by 'PANIC: failed to add item to the page', as reported by Joe
Conway for scenarios involving heavy concurrent insertion activity.
of XLogInsert had the same sort of checkpoint interlock problem as
RecordTransactionCommit, and indeed I found some. Btree index build
and ALTER TABLE SET TABLESPACE write data outside the friendly confines
of the buffer manager, and therefore they have to take their own
responsibility for checkpoint interlock. The easiest solution seems to
be to force smgrimmedsync at the end of the index build or table copy,
even when the operation is being WAL-logged. This is sufficient since
the new index or table will be of interest to no one if we don't get
as far as committing the current transaction.
recovery more manageable. Also, undo recent change to add FILE_HEADER
and WASTED_SPACE records to XLOG; instead make the XLOG page header
variable-size with extra fields in the first page of an XLOG file.
This should fix the boundary-case bugs observed by Mark Kirkwood.
initdb forced due to change of XLOG representation.
keep track of portal-related resources separately from transaction-related
resources. This allows cursors to work in a somewhat sane fashion with
nested transactions. For now, cursor behavior is non-subtransactional,
that is a cursor's state does not roll back if you abort a subtransaction
that fetched from the cursor. We might want to change that later.
Tom Lane
Neil Conway
Bruce Momjian
Alvaro Herrera
Peter Eisentraut
PostgreSQL Daemon